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 unsigned cq_last_tm_flush;
407 } ____cacheline_aligned_in_smp;
410 spinlock_t completion_lock;
412 spinlock_t timeout_lock;
415 * ->iopoll_list is protected by the ctx->uring_lock for
416 * io_uring instances that don't use IORING_SETUP_SQPOLL.
417 * For SQPOLL, only the single threaded io_sq_thread() will
418 * manipulate the list, hence no extra locking is needed there.
420 struct list_head iopoll_list;
421 struct hlist_head *cancel_hash;
422 unsigned cancel_hash_bits;
423 bool poll_multi_queue;
424 } ____cacheline_aligned_in_smp;
426 struct io_restriction restrictions;
428 /* slow path rsrc auxilary data, used by update/register */
430 struct io_rsrc_node *rsrc_backup_node;
431 struct io_mapped_ubuf *dummy_ubuf;
432 struct io_rsrc_data *file_data;
433 struct io_rsrc_data *buf_data;
435 struct delayed_work rsrc_put_work;
436 struct llist_head rsrc_put_llist;
437 struct list_head rsrc_ref_list;
438 spinlock_t rsrc_ref_lock;
441 /* Keep this last, we don't need it for the fast path */
443 #if defined(CONFIG_UNIX)
444 struct socket *ring_sock;
446 /* hashed buffered write serialization */
447 struct io_wq_hash *hash_map;
449 /* Only used for accounting purposes */
450 struct user_struct *user;
451 struct mm_struct *mm_account;
453 /* ctx exit and cancelation */
454 struct llist_head fallback_llist;
455 struct delayed_work fallback_work;
456 struct work_struct exit_work;
457 struct list_head tctx_list;
458 struct completion ref_comp;
462 struct io_uring_task {
463 /* submission side */
466 struct wait_queue_head wait;
467 const struct io_ring_ctx *last;
469 struct percpu_counter inflight;
470 atomic_t inflight_tracked;
473 spinlock_t task_lock;
474 struct io_wq_work_list task_list;
475 struct callback_head task_work;
480 * First field must be the file pointer in all the
481 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
483 struct io_poll_iocb {
485 struct wait_queue_head *head;
489 struct wait_queue_entry wait;
492 struct io_poll_update {
498 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;
715 struct iov_iter_state iter_state;
717 struct wait_page_queue wpq;
721 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
722 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
723 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
724 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
725 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
726 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
728 /* first byte is taken by user flags, shift it to not overlap */
733 REQ_F_LINK_TIMEOUT_BIT,
734 REQ_F_NEED_CLEANUP_BIT,
736 REQ_F_BUFFER_SELECTED_BIT,
737 REQ_F_COMPLETE_INLINE_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 /* supports async reads */
786 REQ_F_NOWAIT_READ = BIT(REQ_F_NOWAIT_READ_BIT),
787 /* supports async writes */
788 REQ_F_NOWAIT_WRITE = BIT(REQ_F_NOWAIT_WRITE_BIT),
790 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
791 /* has creds assigned */
792 REQ_F_CREDS = BIT(REQ_F_CREDS_BIT),
793 /* skip refcounting if not set */
794 REQ_F_REFCOUNT = BIT(REQ_F_REFCOUNT_BIT),
795 /* there is a linked timeout that has to be armed */
796 REQ_F_ARM_LTIMEOUT = BIT(REQ_F_ARM_LTIMEOUT_BIT),
800 struct io_poll_iocb poll;
801 struct io_poll_iocb *double_poll;
804 typedef void (*io_req_tw_func_t)(struct io_kiocb *req, bool *locked);
806 struct io_task_work {
808 struct io_wq_work_node node;
809 struct llist_node fallback_node;
811 io_req_tw_func_t func;
815 IORING_RSRC_FILE = 0,
816 IORING_RSRC_BUFFER = 1,
820 * NOTE! Each of the iocb union members has the file pointer
821 * as the first entry in their struct definition. So you can
822 * access the file pointer through any of the sub-structs,
823 * or directly as just 'ki_filp' in this struct.
829 struct io_poll_iocb poll;
830 struct io_poll_update poll_update;
831 struct io_accept accept;
833 struct io_cancel cancel;
834 struct io_timeout timeout;
835 struct io_timeout_rem timeout_rem;
836 struct io_connect connect;
837 struct io_sr_msg sr_msg;
839 struct io_close close;
840 struct io_rsrc_update rsrc_update;
841 struct io_fadvise fadvise;
842 struct io_madvise madvise;
843 struct io_epoll epoll;
844 struct io_splice splice;
845 struct io_provide_buf pbuf;
846 struct io_statx statx;
847 struct io_shutdown shutdown;
848 struct io_rename rename;
849 struct io_unlink unlink;
850 struct io_mkdir mkdir;
851 struct io_symlink symlink;
852 struct io_hardlink hardlink;
853 /* use only after cleaning per-op data, see io_clean_op() */
854 struct io_completion compl;
857 /* opcode allocated if it needs to store data for async defer */
860 /* polled IO has completed */
866 struct io_ring_ctx *ctx;
869 struct task_struct *task;
872 struct io_kiocb *link;
873 struct percpu_ref *fixed_rsrc_refs;
875 /* used with ctx->iopoll_list with reads/writes */
876 struct list_head inflight_entry;
877 struct io_task_work io_task_work;
878 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
879 struct hlist_node hash_node;
880 struct async_poll *apoll;
881 struct io_wq_work work;
882 const struct cred *creds;
884 /* store used ubuf, so we can prevent reloading */
885 struct io_mapped_ubuf *imu;
888 struct io_tctx_node {
889 struct list_head ctx_node;
890 struct task_struct *task;
891 struct io_ring_ctx *ctx;
894 struct io_defer_entry {
895 struct list_head list;
896 struct io_kiocb *req;
901 /* needs req->file assigned */
902 unsigned needs_file : 1;
903 /* hash wq insertion if file is a regular file */
904 unsigned hash_reg_file : 1;
905 /* unbound wq insertion if file is a non-regular file */
906 unsigned unbound_nonreg_file : 1;
907 /* opcode is not supported by this kernel */
908 unsigned not_supported : 1;
909 /* set if opcode supports polled "wait" */
911 unsigned pollout : 1;
912 /* op supports buffer selection */
913 unsigned buffer_select : 1;
914 /* do prep async if is going to be punted */
915 unsigned needs_async_setup : 1;
916 /* should block plug */
918 /* size of async data needed, if any */
919 unsigned short async_size;
922 static const struct io_op_def io_op_defs[] = {
923 [IORING_OP_NOP] = {},
924 [IORING_OP_READV] = {
926 .unbound_nonreg_file = 1,
929 .needs_async_setup = 1,
931 .async_size = sizeof(struct io_async_rw),
933 [IORING_OP_WRITEV] = {
936 .unbound_nonreg_file = 1,
938 .needs_async_setup = 1,
940 .async_size = sizeof(struct io_async_rw),
942 [IORING_OP_FSYNC] = {
945 [IORING_OP_READ_FIXED] = {
947 .unbound_nonreg_file = 1,
950 .async_size = sizeof(struct io_async_rw),
952 [IORING_OP_WRITE_FIXED] = {
955 .unbound_nonreg_file = 1,
958 .async_size = sizeof(struct io_async_rw),
960 [IORING_OP_POLL_ADD] = {
962 .unbound_nonreg_file = 1,
964 [IORING_OP_POLL_REMOVE] = {},
965 [IORING_OP_SYNC_FILE_RANGE] = {
968 [IORING_OP_SENDMSG] = {
970 .unbound_nonreg_file = 1,
972 .needs_async_setup = 1,
973 .async_size = sizeof(struct io_async_msghdr),
975 [IORING_OP_RECVMSG] = {
977 .unbound_nonreg_file = 1,
980 .needs_async_setup = 1,
981 .async_size = sizeof(struct io_async_msghdr),
983 [IORING_OP_TIMEOUT] = {
984 .async_size = sizeof(struct io_timeout_data),
986 [IORING_OP_TIMEOUT_REMOVE] = {
987 /* used by timeout updates' prep() */
989 [IORING_OP_ACCEPT] = {
991 .unbound_nonreg_file = 1,
994 [IORING_OP_ASYNC_CANCEL] = {},
995 [IORING_OP_LINK_TIMEOUT] = {
996 .async_size = sizeof(struct io_timeout_data),
998 [IORING_OP_CONNECT] = {
1000 .unbound_nonreg_file = 1,
1002 .needs_async_setup = 1,
1003 .async_size = sizeof(struct io_async_connect),
1005 [IORING_OP_FALLOCATE] = {
1008 [IORING_OP_OPENAT] = {},
1009 [IORING_OP_CLOSE] = {},
1010 [IORING_OP_FILES_UPDATE] = {},
1011 [IORING_OP_STATX] = {},
1012 [IORING_OP_READ] = {
1014 .unbound_nonreg_file = 1,
1018 .async_size = sizeof(struct io_async_rw),
1020 [IORING_OP_WRITE] = {
1023 .unbound_nonreg_file = 1,
1026 .async_size = sizeof(struct io_async_rw),
1028 [IORING_OP_FADVISE] = {
1031 [IORING_OP_MADVISE] = {},
1032 [IORING_OP_SEND] = {
1034 .unbound_nonreg_file = 1,
1037 [IORING_OP_RECV] = {
1039 .unbound_nonreg_file = 1,
1043 [IORING_OP_OPENAT2] = {
1045 [IORING_OP_EPOLL_CTL] = {
1046 .unbound_nonreg_file = 1,
1048 [IORING_OP_SPLICE] = {
1051 .unbound_nonreg_file = 1,
1053 [IORING_OP_PROVIDE_BUFFERS] = {},
1054 [IORING_OP_REMOVE_BUFFERS] = {},
1058 .unbound_nonreg_file = 1,
1060 [IORING_OP_SHUTDOWN] = {
1063 [IORING_OP_RENAMEAT] = {},
1064 [IORING_OP_UNLINKAT] = {},
1065 [IORING_OP_MKDIRAT] = {},
1066 [IORING_OP_SYMLINKAT] = {},
1067 [IORING_OP_LINKAT] = {},
1070 /* requests with any of those set should undergo io_disarm_next() */
1071 #define IO_DISARM_MASK (REQ_F_ARM_LTIMEOUT | REQ_F_LINK_TIMEOUT | REQ_F_FAIL)
1073 static bool io_disarm_next(struct io_kiocb *req);
1074 static void io_uring_del_tctx_node(unsigned long index);
1075 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
1076 struct task_struct *task,
1078 static void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd);
1080 static bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1081 long res, unsigned int cflags);
1082 static void io_put_req(struct io_kiocb *req);
1083 static void io_put_req_deferred(struct io_kiocb *req);
1084 static void io_dismantle_req(struct io_kiocb *req);
1085 static void io_queue_linked_timeout(struct io_kiocb *req);
1086 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
1087 struct io_uring_rsrc_update2 *up,
1089 static void io_clean_op(struct io_kiocb *req);
1090 static struct file *io_file_get(struct io_ring_ctx *ctx,
1091 struct io_kiocb *req, int fd, bool fixed);
1092 static void __io_queue_sqe(struct io_kiocb *req);
1093 static void io_rsrc_put_work(struct work_struct *work);
1095 static void io_req_task_queue(struct io_kiocb *req);
1096 static void io_submit_flush_completions(struct io_ring_ctx *ctx);
1097 static int io_req_prep_async(struct io_kiocb *req);
1099 static int io_install_fixed_file(struct io_kiocb *req, struct file *file,
1100 unsigned int issue_flags, u32 slot_index);
1101 static int io_close_fixed(struct io_kiocb *req, unsigned int issue_flags);
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);
1620 static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx)
1622 /* see waitqueue_active() comment */
1625 if (ctx->flags & IORING_SETUP_SQPOLL) {
1626 if (waitqueue_active(&ctx->cq_wait))
1627 wake_up_all(&ctx->cq_wait);
1629 if (io_should_trigger_evfd(ctx))
1630 eventfd_signal(ctx->cq_ev_fd, 1);
1631 if (waitqueue_active(&ctx->poll_wait))
1632 wake_up_interruptible(&ctx->poll_wait);
1635 /* Returns true if there are no backlogged entries after the flush */
1636 static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1638 bool all_flushed, posted;
1640 if (!force && __io_cqring_events(ctx) == ctx->cq_entries)
1644 spin_lock(&ctx->completion_lock);
1645 while (!list_empty(&ctx->cq_overflow_list)) {
1646 struct io_uring_cqe *cqe = io_get_cqe(ctx);
1647 struct io_overflow_cqe *ocqe;
1651 ocqe = list_first_entry(&ctx->cq_overflow_list,
1652 struct io_overflow_cqe, list);
1654 memcpy(cqe, &ocqe->cqe, sizeof(*cqe));
1656 io_account_cq_overflow(ctx);
1659 list_del(&ocqe->list);
1663 all_flushed = list_empty(&ctx->cq_overflow_list);
1665 clear_bit(0, &ctx->check_cq_overflow);
1666 WRITE_ONCE(ctx->rings->sq_flags,
1667 ctx->rings->sq_flags & ~IORING_SQ_CQ_OVERFLOW);
1671 io_commit_cqring(ctx);
1672 spin_unlock(&ctx->completion_lock);
1674 io_cqring_ev_posted(ctx);
1678 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx)
1682 if (test_bit(0, &ctx->check_cq_overflow)) {
1683 /* iopoll syncs against uring_lock, not completion_lock */
1684 if (ctx->flags & IORING_SETUP_IOPOLL)
1685 mutex_lock(&ctx->uring_lock);
1686 ret = __io_cqring_overflow_flush(ctx, false);
1687 if (ctx->flags & IORING_SETUP_IOPOLL)
1688 mutex_unlock(&ctx->uring_lock);
1694 /* must to be called somewhat shortly after putting a request */
1695 static inline void io_put_task(struct task_struct *task, int nr)
1697 struct io_uring_task *tctx = task->io_uring;
1699 if (likely(task == current)) {
1700 tctx->cached_refs += nr;
1702 percpu_counter_sub(&tctx->inflight, nr);
1703 if (unlikely(atomic_read(&tctx->in_idle)))
1704 wake_up(&tctx->wait);
1705 put_task_struct_many(task, nr);
1709 static void io_task_refs_refill(struct io_uring_task *tctx)
1711 unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR;
1713 percpu_counter_add(&tctx->inflight, refill);
1714 refcount_add(refill, ¤t->usage);
1715 tctx->cached_refs += refill;
1718 static inline void io_get_task_refs(int nr)
1720 struct io_uring_task *tctx = current->io_uring;
1722 tctx->cached_refs -= nr;
1723 if (unlikely(tctx->cached_refs < 0))
1724 io_task_refs_refill(tctx);
1727 static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data,
1728 long res, unsigned int cflags)
1730 struct io_overflow_cqe *ocqe;
1732 ocqe = kmalloc(sizeof(*ocqe), GFP_ATOMIC | __GFP_ACCOUNT);
1735 * If we're in ring overflow flush mode, or in task cancel mode,
1736 * or cannot allocate an overflow entry, then we need to drop it
1739 io_account_cq_overflow(ctx);
1742 if (list_empty(&ctx->cq_overflow_list)) {
1743 set_bit(0, &ctx->check_cq_overflow);
1744 WRITE_ONCE(ctx->rings->sq_flags,
1745 ctx->rings->sq_flags | IORING_SQ_CQ_OVERFLOW);
1748 ocqe->cqe.user_data = user_data;
1749 ocqe->cqe.res = res;
1750 ocqe->cqe.flags = cflags;
1751 list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
1755 static inline bool __io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1756 long res, unsigned int cflags)
1758 struct io_uring_cqe *cqe;
1760 trace_io_uring_complete(ctx, user_data, res, cflags);
1763 * If we can't get a cq entry, userspace overflowed the
1764 * submission (by quite a lot). Increment the overflow count in
1767 cqe = io_get_cqe(ctx);
1769 WRITE_ONCE(cqe->user_data, user_data);
1770 WRITE_ONCE(cqe->res, res);
1771 WRITE_ONCE(cqe->flags, cflags);
1774 return io_cqring_event_overflow(ctx, user_data, res, cflags);
1777 /* not as hot to bloat with inlining */
1778 static noinline bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1779 long res, unsigned int cflags)
1781 return __io_cqring_fill_event(ctx, user_data, res, cflags);
1784 static void io_req_complete_post(struct io_kiocb *req, long res,
1785 unsigned int cflags)
1787 struct io_ring_ctx *ctx = req->ctx;
1789 spin_lock(&ctx->completion_lock);
1790 __io_cqring_fill_event(ctx, req->user_data, res, cflags);
1792 * If we're the last reference to this request, add to our locked
1795 if (req_ref_put_and_test(req)) {
1796 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
1797 if (req->flags & IO_DISARM_MASK)
1798 io_disarm_next(req);
1800 io_req_task_queue(req->link);
1804 io_dismantle_req(req);
1805 io_put_task(req->task, 1);
1806 list_add(&req->inflight_entry, &ctx->locked_free_list);
1807 ctx->locked_free_nr++;
1809 if (!percpu_ref_tryget(&ctx->refs))
1812 io_commit_cqring(ctx);
1813 spin_unlock(&ctx->completion_lock);
1816 io_cqring_ev_posted(ctx);
1817 percpu_ref_put(&ctx->refs);
1821 static inline bool io_req_needs_clean(struct io_kiocb *req)
1823 return req->flags & IO_REQ_CLEAN_FLAGS;
1826 static void io_req_complete_state(struct io_kiocb *req, long res,
1827 unsigned int cflags)
1829 if (io_req_needs_clean(req))
1832 req->compl.cflags = cflags;
1833 req->flags |= REQ_F_COMPLETE_INLINE;
1836 static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags,
1837 long res, unsigned cflags)
1839 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1840 io_req_complete_state(req, res, cflags);
1842 io_req_complete_post(req, res, cflags);
1845 static inline void io_req_complete(struct io_kiocb *req, long res)
1847 __io_req_complete(req, 0, res, 0);
1850 static void io_req_complete_failed(struct io_kiocb *req, long res)
1853 io_req_complete_post(req, res, 0);
1856 static void io_req_complete_fail_submit(struct io_kiocb *req)
1859 * We don't submit, fail them all, for that replace hardlinks with
1860 * normal links. Extra REQ_F_LINK is tolerated.
1862 req->flags &= ~REQ_F_HARDLINK;
1863 req->flags |= REQ_F_LINK;
1864 io_req_complete_failed(req, req->result);
1868 * Don't initialise the fields below on every allocation, but do that in
1869 * advance and keep them valid across allocations.
1871 static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx)
1875 req->async_data = NULL;
1876 /* not necessary, but safer to zero */
1880 static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
1881 struct io_submit_state *state)
1883 spin_lock(&ctx->completion_lock);
1884 list_splice_init(&ctx->locked_free_list, &state->free_list);
1885 ctx->locked_free_nr = 0;
1886 spin_unlock(&ctx->completion_lock);
1889 /* Returns true IFF there are requests in the cache */
1890 static bool io_flush_cached_reqs(struct io_ring_ctx *ctx)
1892 struct io_submit_state *state = &ctx->submit_state;
1896 * If we have more than a batch's worth of requests in our IRQ side
1897 * locked cache, grab the lock and move them over to our submission
1900 if (READ_ONCE(ctx->locked_free_nr) > IO_COMPL_BATCH)
1901 io_flush_cached_locked_reqs(ctx, state);
1903 nr = state->free_reqs;
1904 while (!list_empty(&state->free_list)) {
1905 struct io_kiocb *req = list_first_entry(&state->free_list,
1906 struct io_kiocb, inflight_entry);
1908 list_del(&req->inflight_entry);
1909 state->reqs[nr++] = req;
1910 if (nr == ARRAY_SIZE(state->reqs))
1914 state->free_reqs = nr;
1919 * A request might get retired back into the request caches even before opcode
1920 * handlers and io_issue_sqe() are done with it, e.g. inline completion path.
1921 * Because of that, io_alloc_req() should be called only under ->uring_lock
1922 * and with extra caution to not get a request that is still worked on.
1924 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
1925 __must_hold(&ctx->uring_lock)
1927 struct io_submit_state *state = &ctx->submit_state;
1928 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1931 BUILD_BUG_ON(ARRAY_SIZE(state->reqs) < IO_REQ_ALLOC_BATCH);
1933 if (likely(state->free_reqs || io_flush_cached_reqs(ctx)))
1936 ret = kmem_cache_alloc_bulk(req_cachep, gfp, IO_REQ_ALLOC_BATCH,
1940 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1941 * retry single alloc to be on the safe side.
1943 if (unlikely(ret <= 0)) {
1944 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1945 if (!state->reqs[0])
1950 for (i = 0; i < ret; i++)
1951 io_preinit_req(state->reqs[i], ctx);
1952 state->free_reqs = ret;
1955 return state->reqs[state->free_reqs];
1958 static inline void io_put_file(struct file *file)
1964 static void io_dismantle_req(struct io_kiocb *req)
1966 unsigned int flags = req->flags;
1968 if (io_req_needs_clean(req))
1970 if (!(flags & REQ_F_FIXED_FILE))
1971 io_put_file(req->file);
1972 if (req->fixed_rsrc_refs)
1973 percpu_ref_put(req->fixed_rsrc_refs);
1974 if (req->async_data) {
1975 kfree(req->async_data);
1976 req->async_data = NULL;
1980 static void __io_free_req(struct io_kiocb *req)
1982 struct io_ring_ctx *ctx = req->ctx;
1984 io_dismantle_req(req);
1985 io_put_task(req->task, 1);
1987 spin_lock(&ctx->completion_lock);
1988 list_add(&req->inflight_entry, &ctx->locked_free_list);
1989 ctx->locked_free_nr++;
1990 spin_unlock(&ctx->completion_lock);
1992 percpu_ref_put(&ctx->refs);
1995 static inline void io_remove_next_linked(struct io_kiocb *req)
1997 struct io_kiocb *nxt = req->link;
1999 req->link = nxt->link;
2003 static bool io_kill_linked_timeout(struct io_kiocb *req)
2004 __must_hold(&req->ctx->completion_lock)
2005 __must_hold(&req->ctx->timeout_lock)
2007 struct io_kiocb *link = req->link;
2009 if (link && link->opcode == IORING_OP_LINK_TIMEOUT) {
2010 struct io_timeout_data *io = link->async_data;
2012 io_remove_next_linked(req);
2013 link->timeout.head = NULL;
2014 if (hrtimer_try_to_cancel(&io->timer) != -1) {
2015 list_del(&link->timeout.list);
2016 io_cqring_fill_event(link->ctx, link->user_data,
2018 io_put_req_deferred(link);
2025 static void io_fail_links(struct io_kiocb *req)
2026 __must_hold(&req->ctx->completion_lock)
2028 struct io_kiocb *nxt, *link = req->link;
2032 long res = -ECANCELED;
2034 if (link->flags & REQ_F_FAIL)
2040 trace_io_uring_fail_link(req, link);
2041 io_cqring_fill_event(link->ctx, link->user_data, res, 0);
2042 io_put_req_deferred(link);
2047 static bool io_disarm_next(struct io_kiocb *req)
2048 __must_hold(&req->ctx->completion_lock)
2050 bool posted = false;
2052 if (req->flags & REQ_F_ARM_LTIMEOUT) {
2053 struct io_kiocb *link = req->link;
2055 req->flags &= ~REQ_F_ARM_LTIMEOUT;
2056 if (link && link->opcode == IORING_OP_LINK_TIMEOUT) {
2057 io_remove_next_linked(req);
2058 io_cqring_fill_event(link->ctx, link->user_data,
2060 io_put_req_deferred(link);
2063 } else if (req->flags & REQ_F_LINK_TIMEOUT) {
2064 struct io_ring_ctx *ctx = req->ctx;
2066 spin_lock_irq(&ctx->timeout_lock);
2067 posted = io_kill_linked_timeout(req);
2068 spin_unlock_irq(&ctx->timeout_lock);
2070 if (unlikely((req->flags & REQ_F_FAIL) &&
2071 !(req->flags & REQ_F_HARDLINK))) {
2072 posted |= (req->link != NULL);
2078 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
2080 struct io_kiocb *nxt;
2083 * If LINK is set, we have dependent requests in this chain. If we
2084 * didn't fail this request, queue the first one up, moving any other
2085 * dependencies to the next request. In case of failure, fail the rest
2088 if (req->flags & IO_DISARM_MASK) {
2089 struct io_ring_ctx *ctx = req->ctx;
2092 spin_lock(&ctx->completion_lock);
2093 posted = io_disarm_next(req);
2095 io_commit_cqring(req->ctx);
2096 spin_unlock(&ctx->completion_lock);
2098 io_cqring_ev_posted(ctx);
2105 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
2107 if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK))))
2109 return __io_req_find_next(req);
2112 static void ctx_flush_and_put(struct io_ring_ctx *ctx, bool *locked)
2117 if (ctx->submit_state.compl_nr)
2118 io_submit_flush_completions(ctx);
2119 mutex_unlock(&ctx->uring_lock);
2122 percpu_ref_put(&ctx->refs);
2125 static void tctx_task_work(struct callback_head *cb)
2127 bool locked = false;
2128 struct io_ring_ctx *ctx = NULL;
2129 struct io_uring_task *tctx = container_of(cb, struct io_uring_task,
2133 struct io_wq_work_node *node;
2135 if (!tctx->task_list.first && locked && ctx->submit_state.compl_nr)
2136 io_submit_flush_completions(ctx);
2138 spin_lock_irq(&tctx->task_lock);
2139 node = tctx->task_list.first;
2140 INIT_WQ_LIST(&tctx->task_list);
2142 tctx->task_running = false;
2143 spin_unlock_irq(&tctx->task_lock);
2148 struct io_wq_work_node *next = node->next;
2149 struct io_kiocb *req = container_of(node, struct io_kiocb,
2152 if (req->ctx != ctx) {
2153 ctx_flush_and_put(ctx, &locked);
2155 /* if not contended, grab and improve batching */
2156 locked = mutex_trylock(&ctx->uring_lock);
2157 percpu_ref_get(&ctx->refs);
2159 req->io_task_work.func(req, &locked);
2166 ctx_flush_and_put(ctx, &locked);
2169 static void io_req_task_work_add(struct io_kiocb *req)
2171 struct task_struct *tsk = req->task;
2172 struct io_uring_task *tctx = tsk->io_uring;
2173 enum task_work_notify_mode notify;
2174 struct io_wq_work_node *node;
2175 unsigned long flags;
2178 WARN_ON_ONCE(!tctx);
2180 spin_lock_irqsave(&tctx->task_lock, flags);
2181 wq_list_add_tail(&req->io_task_work.node, &tctx->task_list);
2182 running = tctx->task_running;
2184 tctx->task_running = true;
2185 spin_unlock_irqrestore(&tctx->task_lock, flags);
2187 /* task_work already pending, we're done */
2192 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
2193 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
2194 * processing task_work. There's no reliable way to tell if TWA_RESUME
2197 notify = (req->ctx->flags & IORING_SETUP_SQPOLL) ? TWA_NONE : TWA_SIGNAL;
2198 if (!task_work_add(tsk, &tctx->task_work, notify)) {
2199 wake_up_process(tsk);
2203 spin_lock_irqsave(&tctx->task_lock, flags);
2204 tctx->task_running = false;
2205 node = tctx->task_list.first;
2206 INIT_WQ_LIST(&tctx->task_list);
2207 spin_unlock_irqrestore(&tctx->task_lock, flags);
2210 req = container_of(node, struct io_kiocb, io_task_work.node);
2212 if (llist_add(&req->io_task_work.fallback_node,
2213 &req->ctx->fallback_llist))
2214 schedule_delayed_work(&req->ctx->fallback_work, 1);
2218 static void io_req_task_cancel(struct io_kiocb *req, bool *locked)
2220 struct io_ring_ctx *ctx = req->ctx;
2222 /* not needed for normal modes, but SQPOLL depends on it */
2223 io_tw_lock(ctx, locked);
2224 io_req_complete_failed(req, req->result);
2227 static void io_req_task_submit(struct io_kiocb *req, bool *locked)
2229 struct io_ring_ctx *ctx = req->ctx;
2231 io_tw_lock(ctx, locked);
2232 /* req->task == current here, checking PF_EXITING is safe */
2233 if (likely(!(req->task->flags & PF_EXITING)))
2234 __io_queue_sqe(req);
2236 io_req_complete_failed(req, -EFAULT);
2239 static void io_req_task_queue_fail(struct io_kiocb *req, int ret)
2242 req->io_task_work.func = io_req_task_cancel;
2243 io_req_task_work_add(req);
2246 static void io_req_task_queue(struct io_kiocb *req)
2248 req->io_task_work.func = io_req_task_submit;
2249 io_req_task_work_add(req);
2252 static void io_req_task_queue_reissue(struct io_kiocb *req)
2254 req->io_task_work.func = io_queue_async_work;
2255 io_req_task_work_add(req);
2258 static inline void io_queue_next(struct io_kiocb *req)
2260 struct io_kiocb *nxt = io_req_find_next(req);
2263 io_req_task_queue(nxt);
2266 static void io_free_req(struct io_kiocb *req)
2272 static void io_free_req_work(struct io_kiocb *req, bool *locked)
2278 struct task_struct *task;
2283 static inline void io_init_req_batch(struct req_batch *rb)
2290 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2291 struct req_batch *rb)
2294 percpu_ref_put_many(&ctx->refs, rb->ctx_refs);
2296 io_put_task(rb->task, rb->task_refs);
2299 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req,
2300 struct io_submit_state *state)
2303 io_dismantle_req(req);
2305 if (req->task != rb->task) {
2307 io_put_task(rb->task, rb->task_refs);
2308 rb->task = req->task;
2314 if (state->free_reqs != ARRAY_SIZE(state->reqs))
2315 state->reqs[state->free_reqs++] = req;
2317 list_add(&req->inflight_entry, &state->free_list);
2320 static void io_submit_flush_completions(struct io_ring_ctx *ctx)
2321 __must_hold(&ctx->uring_lock)
2323 struct io_submit_state *state = &ctx->submit_state;
2324 int i, nr = state->compl_nr;
2325 struct req_batch rb;
2327 spin_lock(&ctx->completion_lock);
2328 for (i = 0; i < nr; i++) {
2329 struct io_kiocb *req = state->compl_reqs[i];
2331 __io_cqring_fill_event(ctx, req->user_data, req->result,
2334 io_commit_cqring(ctx);
2335 spin_unlock(&ctx->completion_lock);
2336 io_cqring_ev_posted(ctx);
2338 io_init_req_batch(&rb);
2339 for (i = 0; i < nr; i++) {
2340 struct io_kiocb *req = state->compl_reqs[i];
2342 if (req_ref_put_and_test(req))
2343 io_req_free_batch(&rb, req, &ctx->submit_state);
2346 io_req_free_batch_finish(ctx, &rb);
2347 state->compl_nr = 0;
2351 * Drop reference to request, return next in chain (if there is one) if this
2352 * was the last reference to this request.
2354 static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2356 struct io_kiocb *nxt = NULL;
2358 if (req_ref_put_and_test(req)) {
2359 nxt = io_req_find_next(req);
2365 static inline void io_put_req(struct io_kiocb *req)
2367 if (req_ref_put_and_test(req))
2371 static inline void io_put_req_deferred(struct io_kiocb *req)
2373 if (req_ref_put_and_test(req)) {
2374 req->io_task_work.func = io_free_req_work;
2375 io_req_task_work_add(req);
2379 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2381 /* See comment at the top of this file */
2383 return __io_cqring_events(ctx);
2386 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2388 struct io_rings *rings = ctx->rings;
2390 /* make sure SQ entry isn't read before tail */
2391 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2394 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2396 unsigned int cflags;
2398 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2399 cflags |= IORING_CQE_F_BUFFER;
2400 req->flags &= ~REQ_F_BUFFER_SELECTED;
2405 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2407 struct io_buffer *kbuf;
2409 if (likely(!(req->flags & REQ_F_BUFFER_SELECTED)))
2411 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2412 return io_put_kbuf(req, kbuf);
2415 static inline bool io_run_task_work(void)
2417 if (test_thread_flag(TIF_NOTIFY_SIGNAL) || current->task_works) {
2418 __set_current_state(TASK_RUNNING);
2419 tracehook_notify_signal();
2427 * Find and free completed poll iocbs
2429 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2430 struct list_head *done)
2432 struct req_batch rb;
2433 struct io_kiocb *req;
2435 /* order with ->result store in io_complete_rw_iopoll() */
2438 io_init_req_batch(&rb);
2439 while (!list_empty(done)) {
2440 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2441 list_del(&req->inflight_entry);
2443 __io_cqring_fill_event(ctx, req->user_data, req->result,
2444 io_put_rw_kbuf(req));
2447 if (req_ref_put_and_test(req))
2448 io_req_free_batch(&rb, req, &ctx->submit_state);
2451 io_commit_cqring(ctx);
2452 io_cqring_ev_posted_iopoll(ctx);
2453 io_req_free_batch_finish(ctx, &rb);
2456 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2459 struct io_kiocb *req, *tmp;
2464 * Only spin for completions if we don't have multiple devices hanging
2465 * off our complete list, and we're under the requested amount.
2467 spin = !ctx->poll_multi_queue && *nr_events < min;
2469 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2470 struct kiocb *kiocb = &req->rw.kiocb;
2474 * Move completed and retryable entries to our local lists.
2475 * If we find a request that requires polling, break out
2476 * and complete those lists first, if we have entries there.
2478 if (READ_ONCE(req->iopoll_completed)) {
2479 list_move_tail(&req->inflight_entry, &done);
2482 if (!list_empty(&done))
2485 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2486 if (unlikely(ret < 0))
2491 /* iopoll may have completed current req */
2492 if (READ_ONCE(req->iopoll_completed))
2493 list_move_tail(&req->inflight_entry, &done);
2496 if (!list_empty(&done))
2497 io_iopoll_complete(ctx, nr_events, &done);
2503 * We can't just wait for polled events to come to us, we have to actively
2504 * find and complete them.
2506 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2508 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2511 mutex_lock(&ctx->uring_lock);
2512 while (!list_empty(&ctx->iopoll_list)) {
2513 unsigned int nr_events = 0;
2515 io_do_iopoll(ctx, &nr_events, 0);
2517 /* let it sleep and repeat later if can't complete a request */
2521 * Ensure we allow local-to-the-cpu processing to take place,
2522 * in this case we need to ensure that we reap all events.
2523 * Also let task_work, etc. to progress by releasing the mutex
2525 if (need_resched()) {
2526 mutex_unlock(&ctx->uring_lock);
2528 mutex_lock(&ctx->uring_lock);
2531 mutex_unlock(&ctx->uring_lock);
2534 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2536 unsigned int nr_events = 0;
2540 * We disallow the app entering submit/complete with polling, but we
2541 * still need to lock the ring to prevent racing with polled issue
2542 * that got punted to a workqueue.
2544 mutex_lock(&ctx->uring_lock);
2546 * Don't enter poll loop if we already have events pending.
2547 * If we do, we can potentially be spinning for commands that
2548 * already triggered a CQE (eg in error).
2550 if (test_bit(0, &ctx->check_cq_overflow))
2551 __io_cqring_overflow_flush(ctx, false);
2552 if (io_cqring_events(ctx))
2556 * If a submit got punted to a workqueue, we can have the
2557 * application entering polling for a command before it gets
2558 * issued. That app will hold the uring_lock for the duration
2559 * of the poll right here, so we need to take a breather every
2560 * now and then to ensure that the issue has a chance to add
2561 * the poll to the issued list. Otherwise we can spin here
2562 * forever, while the workqueue is stuck trying to acquire the
2565 if (list_empty(&ctx->iopoll_list)) {
2566 u32 tail = ctx->cached_cq_tail;
2568 mutex_unlock(&ctx->uring_lock);
2570 mutex_lock(&ctx->uring_lock);
2572 /* some requests don't go through iopoll_list */
2573 if (tail != ctx->cached_cq_tail ||
2574 list_empty(&ctx->iopoll_list))
2577 ret = io_do_iopoll(ctx, &nr_events, min);
2578 } while (!ret && nr_events < min && !need_resched());
2580 mutex_unlock(&ctx->uring_lock);
2584 static void kiocb_end_write(struct io_kiocb *req)
2587 * Tell lockdep we inherited freeze protection from submission
2590 if (req->flags & REQ_F_ISREG) {
2591 struct super_block *sb = file_inode(req->file)->i_sb;
2593 __sb_writers_acquired(sb, SB_FREEZE_WRITE);
2599 static bool io_resubmit_prep(struct io_kiocb *req)
2601 struct io_async_rw *rw = req->async_data;
2604 return !io_req_prep_async(req);
2605 iov_iter_restore(&rw->iter, &rw->iter_state);
2609 static bool io_rw_should_reissue(struct io_kiocb *req)
2611 umode_t mode = file_inode(req->file)->i_mode;
2612 struct io_ring_ctx *ctx = req->ctx;
2614 if (!S_ISBLK(mode) && !S_ISREG(mode))
2616 if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
2617 !(ctx->flags & IORING_SETUP_IOPOLL)))
2620 * If ref is dying, we might be running poll reap from the exit work.
2621 * Don't attempt to reissue from that path, just let it fail with
2624 if (percpu_ref_is_dying(&ctx->refs))
2627 * Play it safe and assume not safe to re-import and reissue if we're
2628 * not in the original thread group (or in task context).
2630 if (!same_thread_group(req->task, current) || !in_task())
2635 static bool io_resubmit_prep(struct io_kiocb *req)
2639 static bool io_rw_should_reissue(struct io_kiocb *req)
2645 static bool __io_complete_rw_common(struct io_kiocb *req, long res)
2647 if (req->rw.kiocb.ki_flags & IOCB_WRITE)
2648 kiocb_end_write(req);
2649 if (res != req->result) {
2650 if ((res == -EAGAIN || res == -EOPNOTSUPP) &&
2651 io_rw_should_reissue(req)) {
2652 req->flags |= REQ_F_REISSUE;
2661 static void io_req_task_complete(struct io_kiocb *req, bool *locked)
2663 unsigned int cflags = io_put_rw_kbuf(req);
2664 long res = req->result;
2667 struct io_ring_ctx *ctx = req->ctx;
2668 struct io_submit_state *state = &ctx->submit_state;
2670 io_req_complete_state(req, res, cflags);
2671 state->compl_reqs[state->compl_nr++] = req;
2672 if (state->compl_nr == ARRAY_SIZE(state->compl_reqs))
2673 io_submit_flush_completions(ctx);
2675 io_req_complete_post(req, res, cflags);
2679 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2680 unsigned int issue_flags)
2682 if (__io_complete_rw_common(req, res))
2684 __io_req_complete(req, issue_flags, req->result, io_put_rw_kbuf(req));
2687 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2689 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2691 if (__io_complete_rw_common(req, res))
2694 req->io_task_work.func = io_req_task_complete;
2695 io_req_task_work_add(req);
2698 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2700 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2702 if (kiocb->ki_flags & IOCB_WRITE)
2703 kiocb_end_write(req);
2704 if (unlikely(res != req->result)) {
2705 if (res == -EAGAIN && io_rw_should_reissue(req)) {
2706 req->flags |= REQ_F_REISSUE;
2711 WRITE_ONCE(req->result, res);
2712 /* order with io_iopoll_complete() checking ->result */
2714 WRITE_ONCE(req->iopoll_completed, 1);
2718 * After the iocb has been issued, it's safe to be found on the poll list.
2719 * Adding the kiocb to the list AFTER submission ensures that we don't
2720 * find it from a io_do_iopoll() thread before the issuer is done
2721 * accessing the kiocb cookie.
2723 static void io_iopoll_req_issued(struct io_kiocb *req)
2725 struct io_ring_ctx *ctx = req->ctx;
2726 const bool in_async = io_wq_current_is_worker();
2728 /* workqueue context doesn't hold uring_lock, grab it now */
2729 if (unlikely(in_async))
2730 mutex_lock(&ctx->uring_lock);
2733 * Track whether we have multiple files in our lists. This will impact
2734 * how we do polling eventually, not spinning if we're on potentially
2735 * different devices.
2737 if (list_empty(&ctx->iopoll_list)) {
2738 ctx->poll_multi_queue = false;
2739 } else if (!ctx->poll_multi_queue) {
2740 struct io_kiocb *list_req;
2741 unsigned int queue_num0, queue_num1;
2743 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2746 if (list_req->file != req->file) {
2747 ctx->poll_multi_queue = true;
2749 queue_num0 = blk_qc_t_to_queue_num(list_req->rw.kiocb.ki_cookie);
2750 queue_num1 = blk_qc_t_to_queue_num(req->rw.kiocb.ki_cookie);
2751 if (queue_num0 != queue_num1)
2752 ctx->poll_multi_queue = true;
2757 * For fast devices, IO may have already completed. If it has, add
2758 * it to the front so we find it first.
2760 if (READ_ONCE(req->iopoll_completed))
2761 list_add(&req->inflight_entry, &ctx->iopoll_list);
2763 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2765 if (unlikely(in_async)) {
2767 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
2768 * in sq thread task context or in io worker task context. If
2769 * current task context is sq thread, we don't need to check
2770 * whether should wake up sq thread.
2772 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2773 wq_has_sleeper(&ctx->sq_data->wait))
2774 wake_up(&ctx->sq_data->wait);
2776 mutex_unlock(&ctx->uring_lock);
2780 static bool io_bdev_nowait(struct block_device *bdev)
2782 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2786 * If we tracked the file through the SCM inflight mechanism, we could support
2787 * any file. For now, just ensure that anything potentially problematic is done
2790 static bool __io_file_supports_nowait(struct file *file, int rw)
2792 umode_t mode = file_inode(file)->i_mode;
2794 if (S_ISBLK(mode)) {
2795 if (IS_ENABLED(CONFIG_BLOCK) &&
2796 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2802 if (S_ISREG(mode)) {
2803 if (IS_ENABLED(CONFIG_BLOCK) &&
2804 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2805 file->f_op != &io_uring_fops)
2810 /* any ->read/write should understand O_NONBLOCK */
2811 if (file->f_flags & O_NONBLOCK)
2814 if (!(file->f_mode & FMODE_NOWAIT))
2818 return file->f_op->read_iter != NULL;
2820 return file->f_op->write_iter != NULL;
2823 static bool io_file_supports_nowait(struct io_kiocb *req, int rw)
2825 if (rw == READ && (req->flags & REQ_F_NOWAIT_READ))
2827 else if (rw == WRITE && (req->flags & REQ_F_NOWAIT_WRITE))
2830 return __io_file_supports_nowait(req->file, rw);
2833 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2836 struct io_ring_ctx *ctx = req->ctx;
2837 struct kiocb *kiocb = &req->rw.kiocb;
2838 struct file *file = req->file;
2842 if (!io_req_ffs_set(req) && S_ISREG(file_inode(file)->i_mode))
2843 req->flags |= REQ_F_ISREG;
2845 kiocb->ki_pos = READ_ONCE(sqe->off);
2846 if (kiocb->ki_pos == -1 && !(file->f_mode & FMODE_STREAM)) {
2847 req->flags |= REQ_F_CUR_POS;
2848 kiocb->ki_pos = file->f_pos;
2850 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2851 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2852 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2857 * If the file is marked O_NONBLOCK, still allow retry for it if it
2858 * supports async. Otherwise it's impossible to use O_NONBLOCK files
2859 * reliably. If not, or it IOCB_NOWAIT is set, don't retry.
2861 if ((kiocb->ki_flags & IOCB_NOWAIT) ||
2862 ((file->f_flags & O_NONBLOCK) && !io_file_supports_nowait(req, rw)))
2863 req->flags |= REQ_F_NOWAIT;
2865 ioprio = READ_ONCE(sqe->ioprio);
2867 ret = ioprio_check_cap(ioprio);
2871 kiocb->ki_ioprio = ioprio;
2873 kiocb->ki_ioprio = get_current_ioprio();
2875 if (ctx->flags & IORING_SETUP_IOPOLL) {
2876 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2877 !kiocb->ki_filp->f_op->iopoll)
2880 kiocb->ki_flags |= IOCB_HIPRI | IOCB_ALLOC_CACHE;
2881 kiocb->ki_complete = io_complete_rw_iopoll;
2882 req->iopoll_completed = 0;
2884 if (kiocb->ki_flags & IOCB_HIPRI)
2886 kiocb->ki_complete = io_complete_rw;
2889 if (req->opcode == IORING_OP_READ_FIXED ||
2890 req->opcode == IORING_OP_WRITE_FIXED) {
2892 io_req_set_rsrc_node(req);
2895 req->rw.addr = READ_ONCE(sqe->addr);
2896 req->rw.len = READ_ONCE(sqe->len);
2897 req->buf_index = READ_ONCE(sqe->buf_index);
2901 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2907 case -ERESTARTNOINTR:
2908 case -ERESTARTNOHAND:
2909 case -ERESTART_RESTARTBLOCK:
2911 * We can't just restart the syscall, since previously
2912 * submitted sqes may already be in progress. Just fail this
2918 kiocb->ki_complete(kiocb, ret, 0);
2922 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2923 unsigned int issue_flags)
2925 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2926 struct io_async_rw *io = req->async_data;
2928 /* add previously done IO, if any */
2929 if (io && io->bytes_done > 0) {
2931 ret = io->bytes_done;
2933 ret += io->bytes_done;
2936 if (req->flags & REQ_F_CUR_POS)
2937 req->file->f_pos = kiocb->ki_pos;
2938 if (ret >= 0 && (kiocb->ki_complete == io_complete_rw))
2939 __io_complete_rw(req, ret, 0, issue_flags);
2941 io_rw_done(kiocb, ret);
2943 if (req->flags & REQ_F_REISSUE) {
2944 req->flags &= ~REQ_F_REISSUE;
2945 if (io_resubmit_prep(req)) {
2946 io_req_task_queue_reissue(req);
2948 unsigned int cflags = io_put_rw_kbuf(req);
2949 struct io_ring_ctx *ctx = req->ctx;
2952 if (!(issue_flags & IO_URING_F_NONBLOCK)) {
2953 mutex_lock(&ctx->uring_lock);
2954 __io_req_complete(req, issue_flags, ret, cflags);
2955 mutex_unlock(&ctx->uring_lock);
2957 __io_req_complete(req, issue_flags, ret, cflags);
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));
3266 if (!iov_iter_is_bvec(iter)) {
3267 iov_iter_advance(iter, nr);
3273 if (nr != iovec.iov_len)
3280 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3281 const struct iovec *fast_iov, struct iov_iter *iter)
3283 struct io_async_rw *rw = req->async_data;
3285 memcpy(&rw->iter, iter, sizeof(*iter));
3286 rw->free_iovec = iovec;
3288 /* can only be fixed buffers, no need to do anything */
3289 if (iov_iter_is_bvec(iter))
3292 unsigned iov_off = 0;
3294 rw->iter.iov = rw->fast_iov;
3295 if (iter->iov != fast_iov) {
3296 iov_off = iter->iov - fast_iov;
3297 rw->iter.iov += iov_off;
3299 if (rw->fast_iov != fast_iov)
3300 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3301 sizeof(struct iovec) * iter->nr_segs);
3303 req->flags |= REQ_F_NEED_CLEANUP;
3307 static inline int io_alloc_async_data(struct io_kiocb *req)
3309 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3310 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3311 return req->async_data == NULL;
3314 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3315 const struct iovec *fast_iov,
3316 struct iov_iter *iter, bool force)
3318 if (!force && !io_op_defs[req->opcode].needs_async_setup)
3320 if (!req->async_data) {
3321 struct io_async_rw *iorw;
3323 if (io_alloc_async_data(req)) {
3328 io_req_map_rw(req, iovec, fast_iov, iter);
3329 iorw = req->async_data;
3330 /* we've copied and mapped the iter, ensure state is saved */
3331 iov_iter_save_state(&iorw->iter, &iorw->iter_state);
3336 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3338 struct io_async_rw *iorw = req->async_data;
3339 struct iovec *iov = iorw->fast_iov;
3342 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3343 if (unlikely(ret < 0))
3346 iorw->bytes_done = 0;
3347 iorw->free_iovec = iov;
3349 req->flags |= REQ_F_NEED_CLEANUP;
3350 iov_iter_save_state(&iorw->iter, &iorw->iter_state);
3354 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3356 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3358 return io_prep_rw(req, sqe, READ);
3362 * This is our waitqueue callback handler, registered through lock_page_async()
3363 * when we initially tried to do the IO with the iocb armed our waitqueue.
3364 * This gets called when the page is unlocked, and we generally expect that to
3365 * happen when the page IO is completed and the page is now uptodate. This will
3366 * queue a task_work based retry of the operation, attempting to copy the data
3367 * again. If the latter fails because the page was NOT uptodate, then we will
3368 * do a thread based blocking retry of the operation. That's the unexpected
3371 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3372 int sync, void *arg)
3374 struct wait_page_queue *wpq;
3375 struct io_kiocb *req = wait->private;
3376 struct wait_page_key *key = arg;
3378 wpq = container_of(wait, struct wait_page_queue, wait);
3380 if (!wake_page_match(wpq, key))
3383 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3384 list_del_init(&wait->entry);
3385 io_req_task_queue(req);
3390 * This controls whether a given IO request should be armed for async page
3391 * based retry. If we return false here, the request is handed to the async
3392 * worker threads for retry. If we're doing buffered reads on a regular file,
3393 * we prepare a private wait_page_queue entry and retry the operation. This
3394 * will either succeed because the page is now uptodate and unlocked, or it
3395 * will register a callback when the page is unlocked at IO completion. Through
3396 * that callback, io_uring uses task_work to setup a retry of the operation.
3397 * That retry will attempt the buffered read again. The retry will generally
3398 * succeed, or in rare cases where it fails, we then fall back to using the
3399 * async worker threads for a blocking retry.
3401 static bool io_rw_should_retry(struct io_kiocb *req)
3403 struct io_async_rw *rw = req->async_data;
3404 struct wait_page_queue *wait = &rw->wpq;
3405 struct kiocb *kiocb = &req->rw.kiocb;
3407 /* never retry for NOWAIT, we just complete with -EAGAIN */
3408 if (req->flags & REQ_F_NOWAIT)
3411 /* Only for buffered IO */
3412 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3416 * just use poll if we can, and don't attempt if the fs doesn't
3417 * support callback based unlocks
3419 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3422 wait->wait.func = io_async_buf_func;
3423 wait->wait.private = req;
3424 wait->wait.flags = 0;
3425 INIT_LIST_HEAD(&wait->wait.entry);
3426 kiocb->ki_flags |= IOCB_WAITQ;
3427 kiocb->ki_flags &= ~IOCB_NOWAIT;
3428 kiocb->ki_waitq = wait;
3432 static inline int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3434 if (req->file->f_op->read_iter)
3435 return call_read_iter(req->file, &req->rw.kiocb, iter);
3436 else if (req->file->f_op->read)
3437 return loop_rw_iter(READ, req, iter);
3442 static bool need_read_all(struct io_kiocb *req)
3444 return req->flags & REQ_F_ISREG ||
3445 S_ISBLK(file_inode(req->file)->i_mode);
3448 static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3450 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3451 struct kiocb *kiocb = &req->rw.kiocb;
3452 struct iov_iter __iter, *iter = &__iter;
3453 struct io_async_rw *rw = req->async_data;
3454 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3455 struct iov_iter_state __state, *state;
3460 state = &rw->iter_state;
3462 * We come here from an earlier attempt, restore our state to
3463 * match in case it doesn't. It's cheap enough that we don't
3464 * need to make this conditional.
3466 iov_iter_restore(iter, state);
3469 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3473 iov_iter_save_state(iter, state);
3475 req->result = iov_iter_count(iter);
3477 /* Ensure we clear previously set non-block flag */
3478 if (!force_nonblock)
3479 kiocb->ki_flags &= ~IOCB_NOWAIT;
3481 kiocb->ki_flags |= IOCB_NOWAIT;
3483 /* If the file doesn't support async, just async punt */
3484 if (force_nonblock && !io_file_supports_nowait(req, READ)) {
3485 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3486 return ret ?: -EAGAIN;
3489 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), req->result);
3490 if (unlikely(ret)) {
3495 ret = io_iter_do_read(req, iter);
3497 if (ret == -EAGAIN || (req->flags & REQ_F_REISSUE)) {
3498 req->flags &= ~REQ_F_REISSUE;
3499 /* IOPOLL retry should happen for io-wq threads */
3500 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3502 /* no retry on NONBLOCK nor RWF_NOWAIT */
3503 if (req->flags & REQ_F_NOWAIT)
3506 } else if (ret == -EIOCBQUEUED) {
3508 } else if (ret <= 0 || ret == req->result || !force_nonblock ||
3509 (req->flags & REQ_F_NOWAIT) || !need_read_all(req)) {
3510 /* read all, failed, already did sync or don't want to retry */
3515 * Don't depend on the iter state matching what was consumed, or being
3516 * untouched in case of error. Restore it and we'll advance it
3517 * manually if we need to.
3519 iov_iter_restore(iter, state);
3521 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3526 rw = req->async_data;
3528 * Now use our persistent iterator and state, if we aren't already.
3529 * We've restored and mapped the iter to match.
3531 if (iter != &rw->iter) {
3533 state = &rw->iter_state;
3538 * We end up here because of a partial read, either from
3539 * above or inside this loop. Advance the iter by the bytes
3540 * that were consumed.
3542 iov_iter_advance(iter, ret);
3543 if (!iov_iter_count(iter))
3545 rw->bytes_done += ret;
3546 iov_iter_save_state(iter, state);
3548 /* if we can retry, do so with the callbacks armed */
3549 if (!io_rw_should_retry(req)) {
3550 kiocb->ki_flags &= ~IOCB_WAITQ;
3555 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3556 * we get -EIOCBQUEUED, then we'll get a notification when the
3557 * desired page gets unlocked. We can also get a partial read
3558 * here, and if we do, then just retry at the new offset.
3560 ret = io_iter_do_read(req, iter);
3561 if (ret == -EIOCBQUEUED)
3563 /* we got some bytes, but not all. retry. */
3564 kiocb->ki_flags &= ~IOCB_WAITQ;
3565 iov_iter_restore(iter, state);
3568 kiocb_done(kiocb, ret, issue_flags);
3570 /* it's faster to check here then delegate to kfree */
3576 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3578 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3580 return io_prep_rw(req, sqe, WRITE);
3583 static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3585 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3586 struct kiocb *kiocb = &req->rw.kiocb;
3587 struct iov_iter __iter, *iter = &__iter;
3588 struct io_async_rw *rw = req->async_data;
3589 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3590 struct iov_iter_state __state, *state;
3595 state = &rw->iter_state;
3596 iov_iter_restore(iter, state);
3599 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3603 iov_iter_save_state(iter, state);
3605 req->result = iov_iter_count(iter);
3607 /* Ensure we clear previously set non-block flag */
3608 if (!force_nonblock)
3609 kiocb->ki_flags &= ~IOCB_NOWAIT;
3611 kiocb->ki_flags |= IOCB_NOWAIT;
3613 /* If the file doesn't support async, just async punt */
3614 if (force_nonblock && !io_file_supports_nowait(req, WRITE))
3617 /* file path doesn't support NOWAIT for non-direct_IO */
3618 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3619 (req->flags & REQ_F_ISREG))
3622 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), req->result);
3627 * Open-code file_start_write here to grab freeze protection,
3628 * which will be released by another thread in
3629 * io_complete_rw(). Fool lockdep by telling it the lock got
3630 * released so that it doesn't complain about the held lock when
3631 * we return to userspace.
3633 if (req->flags & REQ_F_ISREG) {
3634 sb_start_write(file_inode(req->file)->i_sb);
3635 __sb_writers_release(file_inode(req->file)->i_sb,
3638 kiocb->ki_flags |= IOCB_WRITE;
3640 if (req->file->f_op->write_iter)
3641 ret2 = call_write_iter(req->file, kiocb, iter);
3642 else if (req->file->f_op->write)
3643 ret2 = loop_rw_iter(WRITE, req, iter);
3647 if (req->flags & REQ_F_REISSUE) {
3648 req->flags &= ~REQ_F_REISSUE;
3653 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3654 * retry them without IOCB_NOWAIT.
3656 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3658 /* no retry on NONBLOCK nor RWF_NOWAIT */
3659 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3661 if (!force_nonblock || ret2 != -EAGAIN) {
3662 /* IOPOLL retry should happen for io-wq threads */
3663 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3666 kiocb_done(kiocb, ret2, issue_flags);
3669 iov_iter_restore(iter, state);
3670 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3671 return ret ?: -EAGAIN;
3674 /* it's reportedly faster than delegating the null check to kfree() */
3680 static int io_renameat_prep(struct io_kiocb *req,
3681 const struct io_uring_sqe *sqe)
3683 struct io_rename *ren = &req->rename;
3684 const char __user *oldf, *newf;
3686 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3688 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
3690 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3693 ren->old_dfd = READ_ONCE(sqe->fd);
3694 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3695 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3696 ren->new_dfd = READ_ONCE(sqe->len);
3697 ren->flags = READ_ONCE(sqe->rename_flags);
3699 ren->oldpath = getname(oldf);
3700 if (IS_ERR(ren->oldpath))
3701 return PTR_ERR(ren->oldpath);
3703 ren->newpath = getname(newf);
3704 if (IS_ERR(ren->newpath)) {
3705 putname(ren->oldpath);
3706 return PTR_ERR(ren->newpath);
3709 req->flags |= REQ_F_NEED_CLEANUP;
3713 static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3715 struct io_rename *ren = &req->rename;
3718 if (issue_flags & IO_URING_F_NONBLOCK)
3721 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3722 ren->newpath, ren->flags);
3724 req->flags &= ~REQ_F_NEED_CLEANUP;
3727 io_req_complete(req, ret);
3731 static int io_unlinkat_prep(struct io_kiocb *req,
3732 const struct io_uring_sqe *sqe)
3734 struct io_unlink *un = &req->unlink;
3735 const char __user *fname;
3737 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3739 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
3742 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3745 un->dfd = READ_ONCE(sqe->fd);
3747 un->flags = READ_ONCE(sqe->unlink_flags);
3748 if (un->flags & ~AT_REMOVEDIR)
3751 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3752 un->filename = getname(fname);
3753 if (IS_ERR(un->filename))
3754 return PTR_ERR(un->filename);
3756 req->flags |= REQ_F_NEED_CLEANUP;
3760 static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3762 struct io_unlink *un = &req->unlink;
3765 if (issue_flags & IO_URING_F_NONBLOCK)
3768 if (un->flags & AT_REMOVEDIR)
3769 ret = do_rmdir(un->dfd, un->filename);
3771 ret = do_unlinkat(un->dfd, un->filename);
3773 req->flags &= ~REQ_F_NEED_CLEANUP;
3776 io_req_complete(req, ret);
3780 static int io_mkdirat_prep(struct io_kiocb *req,
3781 const struct io_uring_sqe *sqe)
3783 struct io_mkdir *mkd = &req->mkdir;
3784 const char __user *fname;
3786 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3788 if (sqe->ioprio || sqe->off || sqe->rw_flags || sqe->buf_index ||
3791 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3794 mkd->dfd = READ_ONCE(sqe->fd);
3795 mkd->mode = READ_ONCE(sqe->len);
3797 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3798 mkd->filename = getname(fname);
3799 if (IS_ERR(mkd->filename))
3800 return PTR_ERR(mkd->filename);
3802 req->flags |= REQ_F_NEED_CLEANUP;
3806 static int io_mkdirat(struct io_kiocb *req, int issue_flags)
3808 struct io_mkdir *mkd = &req->mkdir;
3811 if (issue_flags & IO_URING_F_NONBLOCK)
3814 ret = do_mkdirat(mkd->dfd, mkd->filename, mkd->mode);
3816 req->flags &= ~REQ_F_NEED_CLEANUP;
3819 io_req_complete(req, ret);
3823 static int io_symlinkat_prep(struct io_kiocb *req,
3824 const struct io_uring_sqe *sqe)
3826 struct io_symlink *sl = &req->symlink;
3827 const char __user *oldpath, *newpath;
3829 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3831 if (sqe->ioprio || sqe->len || sqe->rw_flags || sqe->buf_index ||
3834 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3837 sl->new_dfd = READ_ONCE(sqe->fd);
3838 oldpath = u64_to_user_ptr(READ_ONCE(sqe->addr));
3839 newpath = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3841 sl->oldpath = getname(oldpath);
3842 if (IS_ERR(sl->oldpath))
3843 return PTR_ERR(sl->oldpath);
3845 sl->newpath = getname(newpath);
3846 if (IS_ERR(sl->newpath)) {
3847 putname(sl->oldpath);
3848 return PTR_ERR(sl->newpath);
3851 req->flags |= REQ_F_NEED_CLEANUP;
3855 static int io_symlinkat(struct io_kiocb *req, int issue_flags)
3857 struct io_symlink *sl = &req->symlink;
3860 if (issue_flags & IO_URING_F_NONBLOCK)
3863 ret = do_symlinkat(sl->oldpath, sl->new_dfd, sl->newpath);
3865 req->flags &= ~REQ_F_NEED_CLEANUP;
3868 io_req_complete(req, ret);
3872 static int io_linkat_prep(struct io_kiocb *req,
3873 const struct io_uring_sqe *sqe)
3875 struct io_hardlink *lnk = &req->hardlink;
3876 const char __user *oldf, *newf;
3878 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3880 if (sqe->ioprio || sqe->rw_flags || sqe->buf_index || sqe->splice_fd_in)
3882 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3885 lnk->old_dfd = READ_ONCE(sqe->fd);
3886 lnk->new_dfd = READ_ONCE(sqe->len);
3887 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3888 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3889 lnk->flags = READ_ONCE(sqe->hardlink_flags);
3891 lnk->oldpath = getname(oldf);
3892 if (IS_ERR(lnk->oldpath))
3893 return PTR_ERR(lnk->oldpath);
3895 lnk->newpath = getname(newf);
3896 if (IS_ERR(lnk->newpath)) {
3897 putname(lnk->oldpath);
3898 return PTR_ERR(lnk->newpath);
3901 req->flags |= REQ_F_NEED_CLEANUP;
3905 static int io_linkat(struct io_kiocb *req, int issue_flags)
3907 struct io_hardlink *lnk = &req->hardlink;
3910 if (issue_flags & IO_URING_F_NONBLOCK)
3913 ret = do_linkat(lnk->old_dfd, lnk->oldpath, lnk->new_dfd,
3914 lnk->newpath, lnk->flags);
3916 req->flags &= ~REQ_F_NEED_CLEANUP;
3919 io_req_complete(req, ret);
3923 static int io_shutdown_prep(struct io_kiocb *req,
3924 const struct io_uring_sqe *sqe)
3926 #if defined(CONFIG_NET)
3927 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3929 if (unlikely(sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3930 sqe->buf_index || sqe->splice_fd_in))
3933 req->shutdown.how = READ_ONCE(sqe->len);
3940 static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3942 #if defined(CONFIG_NET)
3943 struct socket *sock;
3946 if (issue_flags & IO_URING_F_NONBLOCK)
3949 sock = sock_from_file(req->file);
3950 if (unlikely(!sock))
3953 ret = __sys_shutdown_sock(sock, req->shutdown.how);
3956 io_req_complete(req, ret);
3963 static int __io_splice_prep(struct io_kiocb *req,
3964 const struct io_uring_sqe *sqe)
3966 struct io_splice *sp = &req->splice;
3967 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3969 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3973 sp->len = READ_ONCE(sqe->len);
3974 sp->flags = READ_ONCE(sqe->splice_flags);
3976 if (unlikely(sp->flags & ~valid_flags))
3979 sp->file_in = io_file_get(req->ctx, req, READ_ONCE(sqe->splice_fd_in),
3980 (sp->flags & SPLICE_F_FD_IN_FIXED));
3983 req->flags |= REQ_F_NEED_CLEANUP;
3987 static int io_tee_prep(struct io_kiocb *req,
3988 const struct io_uring_sqe *sqe)
3990 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3992 return __io_splice_prep(req, sqe);
3995 static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
3997 struct io_splice *sp = &req->splice;
3998 struct file *in = sp->file_in;
3999 struct file *out = sp->file_out;
4000 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
4003 if (issue_flags & IO_URING_F_NONBLOCK)
4006 ret = do_tee(in, out, sp->len, flags);
4008 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
4010 req->flags &= ~REQ_F_NEED_CLEANUP;
4014 io_req_complete(req, ret);
4018 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4020 struct io_splice *sp = &req->splice;
4022 sp->off_in = READ_ONCE(sqe->splice_off_in);
4023 sp->off_out = READ_ONCE(sqe->off);
4024 return __io_splice_prep(req, sqe);
4027 static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
4029 struct io_splice *sp = &req->splice;
4030 struct file *in = sp->file_in;
4031 struct file *out = sp->file_out;
4032 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
4033 loff_t *poff_in, *poff_out;
4036 if (issue_flags & IO_URING_F_NONBLOCK)
4039 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
4040 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
4043 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
4045 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
4047 req->flags &= ~REQ_F_NEED_CLEANUP;
4051 io_req_complete(req, ret);
4056 * IORING_OP_NOP just posts a completion event, nothing else.
4058 static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
4060 struct io_ring_ctx *ctx = req->ctx;
4062 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4065 __io_req_complete(req, issue_flags, 0, 0);
4069 static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4071 struct io_ring_ctx *ctx = req->ctx;
4076 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4078 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index ||
4082 req->sync.flags = READ_ONCE(sqe->fsync_flags);
4083 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
4086 req->sync.off = READ_ONCE(sqe->off);
4087 req->sync.len = READ_ONCE(sqe->len);
4091 static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
4093 loff_t end = req->sync.off + req->sync.len;
4096 /* fsync always requires a blocking context */
4097 if (issue_flags & IO_URING_F_NONBLOCK)
4100 ret = vfs_fsync_range(req->file, req->sync.off,
4101 end > 0 ? end : LLONG_MAX,
4102 req->sync.flags & IORING_FSYNC_DATASYNC);
4105 io_req_complete(req, ret);
4109 static int io_fallocate_prep(struct io_kiocb *req,
4110 const struct io_uring_sqe *sqe)
4112 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags ||
4115 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4118 req->sync.off = READ_ONCE(sqe->off);
4119 req->sync.len = READ_ONCE(sqe->addr);
4120 req->sync.mode = READ_ONCE(sqe->len);
4124 static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
4128 /* fallocate always requiring blocking context */
4129 if (issue_flags & IO_URING_F_NONBLOCK)
4131 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
4135 io_req_complete(req, ret);
4139 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4141 const char __user *fname;
4144 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4146 if (unlikely(sqe->ioprio || sqe->buf_index))
4148 if (unlikely(req->flags & REQ_F_FIXED_FILE))
4151 /* open.how should be already initialised */
4152 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
4153 req->open.how.flags |= O_LARGEFILE;
4155 req->open.dfd = READ_ONCE(sqe->fd);
4156 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
4157 req->open.filename = getname(fname);
4158 if (IS_ERR(req->open.filename)) {
4159 ret = PTR_ERR(req->open.filename);
4160 req->open.filename = NULL;
4164 req->open.file_slot = READ_ONCE(sqe->file_index);
4165 if (req->open.file_slot && (req->open.how.flags & O_CLOEXEC))
4168 req->open.nofile = rlimit(RLIMIT_NOFILE);
4169 req->flags |= REQ_F_NEED_CLEANUP;
4173 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4175 u64 mode = READ_ONCE(sqe->len);
4176 u64 flags = READ_ONCE(sqe->open_flags);
4178 req->open.how = build_open_how(flags, mode);
4179 return __io_openat_prep(req, sqe);
4182 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4184 struct open_how __user *how;
4188 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4189 len = READ_ONCE(sqe->len);
4190 if (len < OPEN_HOW_SIZE_VER0)
4193 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
4198 return __io_openat_prep(req, sqe);
4201 static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
4203 struct open_flags op;
4205 bool resolve_nonblock, nonblock_set;
4206 bool fixed = !!req->open.file_slot;
4209 ret = build_open_flags(&req->open.how, &op);
4212 nonblock_set = op.open_flag & O_NONBLOCK;
4213 resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
4214 if (issue_flags & IO_URING_F_NONBLOCK) {
4216 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
4217 * it'll always -EAGAIN
4219 if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
4221 op.lookup_flags |= LOOKUP_CACHED;
4222 op.open_flag |= O_NONBLOCK;
4226 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
4231 file = do_filp_open(req->open.dfd, req->open.filename, &op);
4234 * We could hang on to this 'fd' on retrying, but seems like
4235 * marginal gain for something that is now known to be a slower
4236 * path. So just put it, and we'll get a new one when we retry.
4241 ret = PTR_ERR(file);
4242 /* only retry if RESOLVE_CACHED wasn't already set by application */
4243 if (ret == -EAGAIN &&
4244 (!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)))
4249 if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
4250 file->f_flags &= ~O_NONBLOCK;
4251 fsnotify_open(file);
4254 fd_install(ret, file);
4256 ret = io_install_fixed_file(req, file, issue_flags,
4257 req->open.file_slot - 1);
4259 putname(req->open.filename);
4260 req->flags &= ~REQ_F_NEED_CLEANUP;
4263 __io_req_complete(req, issue_flags, ret, 0);
4267 static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
4269 return io_openat2(req, issue_flags);
4272 static int io_remove_buffers_prep(struct io_kiocb *req,
4273 const struct io_uring_sqe *sqe)
4275 struct io_provide_buf *p = &req->pbuf;
4278 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off ||
4282 tmp = READ_ONCE(sqe->fd);
4283 if (!tmp || tmp > USHRT_MAX)
4286 memset(p, 0, sizeof(*p));
4288 p->bgid = READ_ONCE(sqe->buf_group);
4292 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
4293 int bgid, unsigned nbufs)
4297 /* shouldn't happen */
4301 /* the head kbuf is the list itself */
4302 while (!list_empty(&buf->list)) {
4303 struct io_buffer *nxt;
4305 nxt = list_first_entry(&buf->list, struct io_buffer, list);
4306 list_del(&nxt->list);
4313 xa_erase(&ctx->io_buffers, bgid);
4318 static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
4320 struct io_provide_buf *p = &req->pbuf;
4321 struct io_ring_ctx *ctx = req->ctx;
4322 struct io_buffer *head;
4324 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4326 io_ring_submit_lock(ctx, !force_nonblock);
4328 lockdep_assert_held(&ctx->uring_lock);
4331 head = xa_load(&ctx->io_buffers, p->bgid);
4333 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
4337 /* complete before unlock, IOPOLL may need the lock */
4338 __io_req_complete(req, issue_flags, ret, 0);
4339 io_ring_submit_unlock(ctx, !force_nonblock);
4343 static int io_provide_buffers_prep(struct io_kiocb *req,
4344 const struct io_uring_sqe *sqe)
4346 unsigned long size, tmp_check;
4347 struct io_provide_buf *p = &req->pbuf;
4350 if (sqe->ioprio || sqe->rw_flags || sqe->splice_fd_in)
4353 tmp = READ_ONCE(sqe->fd);
4354 if (!tmp || tmp > USHRT_MAX)
4357 p->addr = READ_ONCE(sqe->addr);
4358 p->len = READ_ONCE(sqe->len);
4360 if (check_mul_overflow((unsigned long)p->len, (unsigned long)p->nbufs,
4363 if (check_add_overflow((unsigned long)p->addr, size, &tmp_check))
4366 size = (unsigned long)p->len * p->nbufs;
4367 if (!access_ok(u64_to_user_ptr(p->addr), size))
4370 p->bgid = READ_ONCE(sqe->buf_group);
4371 tmp = READ_ONCE(sqe->off);
4372 if (tmp > USHRT_MAX)
4378 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
4380 struct io_buffer *buf;
4381 u64 addr = pbuf->addr;
4382 int i, bid = pbuf->bid;
4384 for (i = 0; i < pbuf->nbufs; i++) {
4385 buf = kmalloc(sizeof(*buf), GFP_KERNEL_ACCOUNT);
4390 buf->len = min_t(__u32, pbuf->len, MAX_RW_COUNT);
4395 INIT_LIST_HEAD(&buf->list);
4398 list_add_tail(&buf->list, &(*head)->list);
4402 return i ? i : -ENOMEM;
4405 static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
4407 struct io_provide_buf *p = &req->pbuf;
4408 struct io_ring_ctx *ctx = req->ctx;
4409 struct io_buffer *head, *list;
4411 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4413 io_ring_submit_lock(ctx, !force_nonblock);
4415 lockdep_assert_held(&ctx->uring_lock);
4417 list = head = xa_load(&ctx->io_buffers, p->bgid);
4419 ret = io_add_buffers(p, &head);
4420 if (ret >= 0 && !list) {
4421 ret = xa_insert(&ctx->io_buffers, p->bgid, head, GFP_KERNEL);
4423 __io_remove_buffers(ctx, head, p->bgid, -1U);
4427 /* complete before unlock, IOPOLL may need the lock */
4428 __io_req_complete(req, issue_flags, ret, 0);
4429 io_ring_submit_unlock(ctx, !force_nonblock);
4433 static int io_epoll_ctl_prep(struct io_kiocb *req,
4434 const struct io_uring_sqe *sqe)
4436 #if defined(CONFIG_EPOLL)
4437 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
4439 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4442 req->epoll.epfd = READ_ONCE(sqe->fd);
4443 req->epoll.op = READ_ONCE(sqe->len);
4444 req->epoll.fd = READ_ONCE(sqe->off);
4446 if (ep_op_has_event(req->epoll.op)) {
4447 struct epoll_event __user *ev;
4449 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4450 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4460 static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4462 #if defined(CONFIG_EPOLL)
4463 struct io_epoll *ie = &req->epoll;
4465 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4467 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4468 if (force_nonblock && ret == -EAGAIN)
4473 __io_req_complete(req, issue_flags, ret, 0);
4480 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4482 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4483 if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->splice_fd_in)
4485 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4488 req->madvise.addr = READ_ONCE(sqe->addr);
4489 req->madvise.len = READ_ONCE(sqe->len);
4490 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4497 static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4499 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4500 struct io_madvise *ma = &req->madvise;
4503 if (issue_flags & IO_URING_F_NONBLOCK)
4506 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4509 io_req_complete(req, ret);
4516 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4518 if (sqe->ioprio || sqe->buf_index || sqe->addr || sqe->splice_fd_in)
4520 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4523 req->fadvise.offset = READ_ONCE(sqe->off);
4524 req->fadvise.len = READ_ONCE(sqe->len);
4525 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4529 static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4531 struct io_fadvise *fa = &req->fadvise;
4534 if (issue_flags & IO_URING_F_NONBLOCK) {
4535 switch (fa->advice) {
4536 case POSIX_FADV_NORMAL:
4537 case POSIX_FADV_RANDOM:
4538 case POSIX_FADV_SEQUENTIAL:
4545 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4548 __io_req_complete(req, issue_flags, ret, 0);
4552 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4554 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4556 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
4558 if (req->flags & REQ_F_FIXED_FILE)
4561 req->statx.dfd = READ_ONCE(sqe->fd);
4562 req->statx.mask = READ_ONCE(sqe->len);
4563 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4564 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4565 req->statx.flags = READ_ONCE(sqe->statx_flags);
4570 static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4572 struct io_statx *ctx = &req->statx;
4575 if (issue_flags & IO_URING_F_NONBLOCK)
4578 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4583 io_req_complete(req, ret);
4587 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4589 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4591 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4592 sqe->rw_flags || sqe->buf_index)
4594 if (req->flags & REQ_F_FIXED_FILE)
4597 req->close.fd = READ_ONCE(sqe->fd);
4598 req->close.file_slot = READ_ONCE(sqe->file_index);
4599 if (req->close.file_slot && req->close.fd)
4605 static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4607 struct files_struct *files = current->files;
4608 struct io_close *close = &req->close;
4609 struct fdtable *fdt;
4610 struct file *file = NULL;
4613 if (req->close.file_slot) {
4614 ret = io_close_fixed(req, issue_flags);
4618 spin_lock(&files->file_lock);
4619 fdt = files_fdtable(files);
4620 if (close->fd >= fdt->max_fds) {
4621 spin_unlock(&files->file_lock);
4624 file = fdt->fd[close->fd];
4625 if (!file || file->f_op == &io_uring_fops) {
4626 spin_unlock(&files->file_lock);
4631 /* if the file has a flush method, be safe and punt to async */
4632 if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4633 spin_unlock(&files->file_lock);
4637 ret = __close_fd_get_file(close->fd, &file);
4638 spin_unlock(&files->file_lock);
4645 /* No ->flush() or already async, safely close from here */
4646 ret = filp_close(file, current->files);
4652 __io_req_complete(req, issue_flags, ret, 0);
4656 static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4658 struct io_ring_ctx *ctx = req->ctx;
4660 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4662 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index ||
4666 req->sync.off = READ_ONCE(sqe->off);
4667 req->sync.len = READ_ONCE(sqe->len);
4668 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4672 static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4676 /* sync_file_range always requires a blocking context */
4677 if (issue_flags & IO_URING_F_NONBLOCK)
4680 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4684 io_req_complete(req, ret);
4688 #if defined(CONFIG_NET)
4689 static int io_setup_async_msg(struct io_kiocb *req,
4690 struct io_async_msghdr *kmsg)
4692 struct io_async_msghdr *async_msg = req->async_data;
4696 if (io_alloc_async_data(req)) {
4697 kfree(kmsg->free_iov);
4700 async_msg = req->async_data;
4701 req->flags |= REQ_F_NEED_CLEANUP;
4702 memcpy(async_msg, kmsg, sizeof(*kmsg));
4703 async_msg->msg.msg_name = &async_msg->addr;
4704 /* if were using fast_iov, set it to the new one */
4705 if (!async_msg->free_iov)
4706 async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4711 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4712 struct io_async_msghdr *iomsg)
4714 iomsg->msg.msg_name = &iomsg->addr;
4715 iomsg->free_iov = iomsg->fast_iov;
4716 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4717 req->sr_msg.msg_flags, &iomsg->free_iov);
4720 static int io_sendmsg_prep_async(struct io_kiocb *req)
4724 ret = io_sendmsg_copy_hdr(req, req->async_data);
4726 req->flags |= REQ_F_NEED_CLEANUP;
4730 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4732 struct io_sr_msg *sr = &req->sr_msg;
4734 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4737 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4738 sr->len = READ_ONCE(sqe->len);
4739 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4740 if (sr->msg_flags & MSG_DONTWAIT)
4741 req->flags |= REQ_F_NOWAIT;
4743 #ifdef CONFIG_COMPAT
4744 if (req->ctx->compat)
4745 sr->msg_flags |= MSG_CMSG_COMPAT;
4750 static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4752 struct io_async_msghdr iomsg, *kmsg;
4753 struct socket *sock;
4758 sock = sock_from_file(req->file);
4759 if (unlikely(!sock))
4762 kmsg = req->async_data;
4764 ret = io_sendmsg_copy_hdr(req, &iomsg);
4770 flags = req->sr_msg.msg_flags;
4771 if (issue_flags & IO_URING_F_NONBLOCK)
4772 flags |= MSG_DONTWAIT;
4773 if (flags & MSG_WAITALL)
4774 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4776 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4777 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4778 return io_setup_async_msg(req, kmsg);
4779 if (ret == -ERESTARTSYS)
4782 /* fast path, check for non-NULL to avoid function call */
4784 kfree(kmsg->free_iov);
4785 req->flags &= ~REQ_F_NEED_CLEANUP;
4788 __io_req_complete(req, issue_flags, ret, 0);
4792 static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4794 struct io_sr_msg *sr = &req->sr_msg;
4797 struct socket *sock;
4802 sock = sock_from_file(req->file);
4803 if (unlikely(!sock))
4806 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4810 msg.msg_name = NULL;
4811 msg.msg_control = NULL;
4812 msg.msg_controllen = 0;
4813 msg.msg_namelen = 0;
4815 flags = req->sr_msg.msg_flags;
4816 if (issue_flags & IO_URING_F_NONBLOCK)
4817 flags |= MSG_DONTWAIT;
4818 if (flags & MSG_WAITALL)
4819 min_ret = iov_iter_count(&msg.msg_iter);
4821 msg.msg_flags = flags;
4822 ret = sock_sendmsg(sock, &msg);
4823 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4825 if (ret == -ERESTARTSYS)
4830 __io_req_complete(req, issue_flags, ret, 0);
4834 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4835 struct io_async_msghdr *iomsg)
4837 struct io_sr_msg *sr = &req->sr_msg;
4838 struct iovec __user *uiov;
4842 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4843 &iomsg->uaddr, &uiov, &iov_len);
4847 if (req->flags & REQ_F_BUFFER_SELECT) {
4850 if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4852 sr->len = iomsg->fast_iov[0].iov_len;
4853 iomsg->free_iov = NULL;
4855 iomsg->free_iov = iomsg->fast_iov;
4856 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4857 &iomsg->free_iov, &iomsg->msg.msg_iter,
4866 #ifdef CONFIG_COMPAT
4867 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4868 struct io_async_msghdr *iomsg)
4870 struct io_sr_msg *sr = &req->sr_msg;
4871 struct compat_iovec __user *uiov;
4876 ret = __get_compat_msghdr(&iomsg->msg, sr->umsg_compat, &iomsg->uaddr,
4881 uiov = compat_ptr(ptr);
4882 if (req->flags & REQ_F_BUFFER_SELECT) {
4883 compat_ssize_t clen;
4887 if (!access_ok(uiov, sizeof(*uiov)))
4889 if (__get_user(clen, &uiov->iov_len))
4894 iomsg->free_iov = NULL;
4896 iomsg->free_iov = iomsg->fast_iov;
4897 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4898 UIO_FASTIOV, &iomsg->free_iov,
4899 &iomsg->msg.msg_iter, true);
4908 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4909 struct io_async_msghdr *iomsg)
4911 iomsg->msg.msg_name = &iomsg->addr;
4913 #ifdef CONFIG_COMPAT
4914 if (req->ctx->compat)
4915 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4918 return __io_recvmsg_copy_hdr(req, iomsg);
4921 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4924 struct io_sr_msg *sr = &req->sr_msg;
4925 struct io_buffer *kbuf;
4927 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4932 req->flags |= REQ_F_BUFFER_SELECTED;
4936 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4938 return io_put_kbuf(req, req->sr_msg.kbuf);
4941 static int io_recvmsg_prep_async(struct io_kiocb *req)
4945 ret = io_recvmsg_copy_hdr(req, req->async_data);
4947 req->flags |= REQ_F_NEED_CLEANUP;
4951 static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4953 struct io_sr_msg *sr = &req->sr_msg;
4955 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4958 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4959 sr->len = READ_ONCE(sqe->len);
4960 sr->bgid = READ_ONCE(sqe->buf_group);
4961 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4962 if (sr->msg_flags & MSG_DONTWAIT)
4963 req->flags |= REQ_F_NOWAIT;
4965 #ifdef CONFIG_COMPAT
4966 if (req->ctx->compat)
4967 sr->msg_flags |= MSG_CMSG_COMPAT;
4972 static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
4974 struct io_async_msghdr iomsg, *kmsg;
4975 struct socket *sock;
4976 struct io_buffer *kbuf;
4979 int ret, cflags = 0;
4980 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4982 sock = sock_from_file(req->file);
4983 if (unlikely(!sock))
4986 kmsg = req->async_data;
4988 ret = io_recvmsg_copy_hdr(req, &iomsg);
4994 if (req->flags & REQ_F_BUFFER_SELECT) {
4995 kbuf = io_recv_buffer_select(req, !force_nonblock);
4997 return PTR_ERR(kbuf);
4998 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4999 kmsg->fast_iov[0].iov_len = req->sr_msg.len;
5000 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
5001 1, req->sr_msg.len);
5004 flags = req->sr_msg.msg_flags;
5006 flags |= MSG_DONTWAIT;
5007 if (flags & MSG_WAITALL)
5008 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
5010 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
5011 kmsg->uaddr, flags);
5012 if (force_nonblock && ret == -EAGAIN)
5013 return io_setup_async_msg(req, kmsg);
5014 if (ret == -ERESTARTSYS)
5017 if (req->flags & REQ_F_BUFFER_SELECTED)
5018 cflags = io_put_recv_kbuf(req);
5019 /* fast path, check for non-NULL to avoid function call */
5021 kfree(kmsg->free_iov);
5022 req->flags &= ~REQ_F_NEED_CLEANUP;
5023 if (ret < min_ret || ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
5025 __io_req_complete(req, issue_flags, ret, cflags);
5029 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
5031 struct io_buffer *kbuf;
5032 struct io_sr_msg *sr = &req->sr_msg;
5034 void __user *buf = sr->buf;
5035 struct socket *sock;
5039 int ret, cflags = 0;
5040 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
5042 sock = sock_from_file(req->file);
5043 if (unlikely(!sock))
5046 if (req->flags & REQ_F_BUFFER_SELECT) {
5047 kbuf = io_recv_buffer_select(req, !force_nonblock);
5049 return PTR_ERR(kbuf);
5050 buf = u64_to_user_ptr(kbuf->addr);
5053 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
5057 msg.msg_name = NULL;
5058 msg.msg_control = NULL;
5059 msg.msg_controllen = 0;
5060 msg.msg_namelen = 0;
5061 msg.msg_iocb = NULL;
5064 flags = req->sr_msg.msg_flags;
5066 flags |= MSG_DONTWAIT;
5067 if (flags & MSG_WAITALL)
5068 min_ret = iov_iter_count(&msg.msg_iter);
5070 ret = sock_recvmsg(sock, &msg, flags);
5071 if (force_nonblock && ret == -EAGAIN)
5073 if (ret == -ERESTARTSYS)
5076 if (req->flags & REQ_F_BUFFER_SELECTED)
5077 cflags = io_put_recv_kbuf(req);
5078 if (ret < min_ret || ((flags & MSG_WAITALL) && (msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
5080 __io_req_complete(req, issue_flags, ret, cflags);
5084 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5086 struct io_accept *accept = &req->accept;
5088 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5090 if (sqe->ioprio || sqe->len || sqe->buf_index)
5093 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
5094 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
5095 accept->flags = READ_ONCE(sqe->accept_flags);
5096 accept->nofile = rlimit(RLIMIT_NOFILE);
5098 accept->file_slot = READ_ONCE(sqe->file_index);
5099 if (accept->file_slot && ((req->open.how.flags & O_CLOEXEC) ||
5100 (accept->flags & SOCK_CLOEXEC)))
5102 if (accept->flags & ~(SOCK_CLOEXEC | SOCK_NONBLOCK))
5104 if (SOCK_NONBLOCK != O_NONBLOCK && (accept->flags & SOCK_NONBLOCK))
5105 accept->flags = (accept->flags & ~SOCK_NONBLOCK) | O_NONBLOCK;
5109 static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
5111 struct io_accept *accept = &req->accept;
5112 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
5113 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
5114 bool fixed = !!accept->file_slot;
5118 if (req->file->f_flags & O_NONBLOCK)
5119 req->flags |= REQ_F_NOWAIT;
5122 fd = __get_unused_fd_flags(accept->flags, accept->nofile);
5123 if (unlikely(fd < 0))
5126 file = do_accept(req->file, file_flags, accept->addr, accept->addr_len,
5131 ret = PTR_ERR(file);
5132 if (ret == -EAGAIN && force_nonblock)
5134 if (ret == -ERESTARTSYS)
5137 } else if (!fixed) {
5138 fd_install(fd, file);
5141 ret = io_install_fixed_file(req, file, issue_flags,
5142 accept->file_slot - 1);
5144 __io_req_complete(req, issue_flags, ret, 0);
5148 static int io_connect_prep_async(struct io_kiocb *req)
5150 struct io_async_connect *io = req->async_data;
5151 struct io_connect *conn = &req->connect;
5153 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
5156 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5158 struct io_connect *conn = &req->connect;
5160 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5162 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags ||
5166 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
5167 conn->addr_len = READ_ONCE(sqe->addr2);
5171 static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
5173 struct io_async_connect __io, *io;
5174 unsigned file_flags;
5176 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
5178 if (req->async_data) {
5179 io = req->async_data;
5181 ret = move_addr_to_kernel(req->connect.addr,
5182 req->connect.addr_len,
5189 file_flags = force_nonblock ? O_NONBLOCK : 0;
5191 ret = __sys_connect_file(req->file, &io->address,
5192 req->connect.addr_len, file_flags);
5193 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
5194 if (req->async_data)
5196 if (io_alloc_async_data(req)) {
5200 memcpy(req->async_data, &__io, sizeof(__io));
5203 if (ret == -ERESTARTSYS)
5208 __io_req_complete(req, issue_flags, ret, 0);
5211 #else /* !CONFIG_NET */
5212 #define IO_NETOP_FN(op) \
5213 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
5215 return -EOPNOTSUPP; \
5218 #define IO_NETOP_PREP(op) \
5220 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
5222 return -EOPNOTSUPP; \
5225 #define IO_NETOP_PREP_ASYNC(op) \
5227 static int io_##op##_prep_async(struct io_kiocb *req) \
5229 return -EOPNOTSUPP; \
5232 IO_NETOP_PREP_ASYNC(sendmsg);
5233 IO_NETOP_PREP_ASYNC(recvmsg);
5234 IO_NETOP_PREP_ASYNC(connect);
5235 IO_NETOP_PREP(accept);
5238 #endif /* CONFIG_NET */
5240 struct io_poll_table {
5241 struct poll_table_struct pt;
5242 struct io_kiocb *req;
5247 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
5248 __poll_t mask, io_req_tw_func_t func)
5250 /* for instances that support it check for an event match first: */
5251 if (mask && !(mask & poll->events))
5254 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
5256 list_del_init(&poll->wait.entry);
5259 req->io_task_work.func = func;
5262 * If this fails, then the task is exiting. When a task exits, the
5263 * work gets canceled, so just cancel this request as well instead
5264 * of executing it. We can't safely execute it anyway, as we may not
5265 * have the needed state needed for it anyway.
5267 io_req_task_work_add(req);
5271 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
5272 __acquires(&req->ctx->completion_lock)
5274 struct io_ring_ctx *ctx = req->ctx;
5276 /* req->task == current here, checking PF_EXITING is safe */
5277 if (unlikely(req->task->flags & PF_EXITING))
5278 WRITE_ONCE(poll->canceled, true);
5280 if (!req->result && !READ_ONCE(poll->canceled)) {
5281 struct poll_table_struct pt = { ._key = poll->events };
5283 req->result = vfs_poll(req->file, &pt) & poll->events;
5286 spin_lock(&ctx->completion_lock);
5287 if (!req->result && !READ_ONCE(poll->canceled)) {
5288 add_wait_queue(poll->head, &poll->wait);
5295 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
5297 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
5298 if (req->opcode == IORING_OP_POLL_ADD)
5299 return req->async_data;
5300 return req->apoll->double_poll;
5303 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
5305 if (req->opcode == IORING_OP_POLL_ADD)
5307 return &req->apoll->poll;
5310 static void io_poll_remove_double(struct io_kiocb *req)
5311 __must_hold(&req->ctx->completion_lock)
5313 struct io_poll_iocb *poll = io_poll_get_double(req);
5315 lockdep_assert_held(&req->ctx->completion_lock);
5317 if (poll && poll->head) {
5318 struct wait_queue_head *head = poll->head;
5320 spin_lock_irq(&head->lock);
5321 list_del_init(&poll->wait.entry);
5322 if (poll->wait.private)
5325 spin_unlock_irq(&head->lock);
5329 static bool __io_poll_complete(struct io_kiocb *req, __poll_t mask)
5330 __must_hold(&req->ctx->completion_lock)
5332 struct io_ring_ctx *ctx = req->ctx;
5333 unsigned flags = IORING_CQE_F_MORE;
5336 if (READ_ONCE(req->poll.canceled)) {
5338 req->poll.events |= EPOLLONESHOT;
5340 error = mangle_poll(mask);
5342 if (req->poll.events & EPOLLONESHOT)
5344 if (!io_cqring_fill_event(ctx, req->user_data, error, flags)) {
5345 req->poll.events |= EPOLLONESHOT;
5348 if (flags & IORING_CQE_F_MORE)
5351 return !(flags & IORING_CQE_F_MORE);
5354 static inline bool io_poll_complete(struct io_kiocb *req, __poll_t mask)
5355 __must_hold(&req->ctx->completion_lock)
5359 done = __io_poll_complete(req, mask);
5360 io_commit_cqring(req->ctx);
5364 static void io_poll_task_func(struct io_kiocb *req, bool *locked)
5366 struct io_ring_ctx *ctx = req->ctx;
5367 struct io_kiocb *nxt;
5369 if (io_poll_rewait(req, &req->poll)) {
5370 spin_unlock(&ctx->completion_lock);
5374 if (req->poll.done) {
5375 spin_unlock(&ctx->completion_lock);
5378 done = __io_poll_complete(req, req->result);
5380 io_poll_remove_double(req);
5381 hash_del(&req->hash_node);
5382 req->poll.done = true;
5385 add_wait_queue(req->poll.head, &req->poll.wait);
5387 io_commit_cqring(ctx);
5388 spin_unlock(&ctx->completion_lock);
5389 io_cqring_ev_posted(ctx);
5392 nxt = io_put_req_find_next(req);
5394 io_req_task_submit(nxt, locked);
5399 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
5400 int sync, void *key)
5402 struct io_kiocb *req = wait->private;
5403 struct io_poll_iocb *poll = io_poll_get_single(req);
5404 __poll_t mask = key_to_poll(key);
5405 unsigned long flags;
5407 /* for instances that support it check for an event match first: */
5408 if (mask && !(mask & poll->events))
5410 if (!(poll->events & EPOLLONESHOT))
5411 return poll->wait.func(&poll->wait, mode, sync, key);
5413 list_del_init(&wait->entry);
5418 spin_lock_irqsave(&poll->head->lock, flags);
5419 done = list_empty(&poll->wait.entry);
5421 list_del_init(&poll->wait.entry);
5422 /* make sure double remove sees this as being gone */
5423 wait->private = NULL;
5424 spin_unlock_irqrestore(&poll->head->lock, flags);
5426 /* use wait func handler, so it matches the rq type */
5427 poll->wait.func(&poll->wait, mode, sync, key);
5434 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
5435 wait_queue_func_t wake_func)
5439 poll->canceled = false;
5440 #define IO_POLL_UNMASK (EPOLLERR|EPOLLHUP|EPOLLNVAL|EPOLLRDHUP)
5441 /* mask in events that we always want/need */
5442 poll->events = events | IO_POLL_UNMASK;
5443 INIT_LIST_HEAD(&poll->wait.entry);
5444 init_waitqueue_func_entry(&poll->wait, wake_func);
5447 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5448 struct wait_queue_head *head,
5449 struct io_poll_iocb **poll_ptr)
5451 struct io_kiocb *req = pt->req;
5454 * The file being polled uses multiple waitqueues for poll handling
5455 * (e.g. one for read, one for write). Setup a separate io_poll_iocb
5458 if (unlikely(pt->nr_entries)) {
5459 struct io_poll_iocb *poll_one = poll;
5461 /* double add on the same waitqueue head, ignore */
5462 if (poll_one->head == head)
5464 /* already have a 2nd entry, fail a third attempt */
5466 if ((*poll_ptr)->head == head)
5468 pt->error = -EINVAL;
5472 * Can't handle multishot for double wait for now, turn it
5473 * into one-shot mode.
5475 if (!(poll_one->events & EPOLLONESHOT))
5476 poll_one->events |= EPOLLONESHOT;
5477 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5479 pt->error = -ENOMEM;
5482 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5484 poll->wait.private = req;
5491 if (poll->events & EPOLLEXCLUSIVE)
5492 add_wait_queue_exclusive(head, &poll->wait);
5494 add_wait_queue(head, &poll->wait);
5497 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5498 struct poll_table_struct *p)
5500 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5501 struct async_poll *apoll = pt->req->apoll;
5503 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5506 static void io_async_task_func(struct io_kiocb *req, bool *locked)
5508 struct async_poll *apoll = req->apoll;
5509 struct io_ring_ctx *ctx = req->ctx;
5511 trace_io_uring_task_run(req->ctx, req, req->opcode, req->user_data);
5513 if (io_poll_rewait(req, &apoll->poll)) {
5514 spin_unlock(&ctx->completion_lock);
5518 hash_del(&req->hash_node);
5519 io_poll_remove_double(req);
5520 apoll->poll.done = true;
5521 spin_unlock(&ctx->completion_lock);
5523 if (!READ_ONCE(apoll->poll.canceled))
5524 io_req_task_submit(req, locked);
5526 io_req_complete_failed(req, -ECANCELED);
5529 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5532 struct io_kiocb *req = wait->private;
5533 struct io_poll_iocb *poll = &req->apoll->poll;
5535 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5538 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5541 static void io_poll_req_insert(struct io_kiocb *req)
5543 struct io_ring_ctx *ctx = req->ctx;
5544 struct hlist_head *list;
5546 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5547 hlist_add_head(&req->hash_node, list);
5550 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5551 struct io_poll_iocb *poll,
5552 struct io_poll_table *ipt, __poll_t mask,
5553 wait_queue_func_t wake_func)
5554 __acquires(&ctx->completion_lock)
5556 struct io_ring_ctx *ctx = req->ctx;
5557 bool cancel = false;
5559 INIT_HLIST_NODE(&req->hash_node);
5560 io_init_poll_iocb(poll, mask, wake_func);
5561 poll->file = req->file;
5562 poll->wait.private = req;
5564 ipt->pt._key = mask;
5567 ipt->nr_entries = 0;
5569 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5570 if (unlikely(!ipt->nr_entries) && !ipt->error)
5571 ipt->error = -EINVAL;
5573 spin_lock(&ctx->completion_lock);
5574 if (ipt->error || (mask && (poll->events & EPOLLONESHOT)))
5575 io_poll_remove_double(req);
5576 if (likely(poll->head)) {
5577 spin_lock_irq(&poll->head->lock);
5578 if (unlikely(list_empty(&poll->wait.entry))) {
5584 if ((mask && (poll->events & EPOLLONESHOT)) || ipt->error)
5585 list_del_init(&poll->wait.entry);
5587 WRITE_ONCE(poll->canceled, true);
5588 else if (!poll->done) /* actually waiting for an event */
5589 io_poll_req_insert(req);
5590 spin_unlock_irq(&poll->head->lock);
5602 static int io_arm_poll_handler(struct io_kiocb *req)
5604 const struct io_op_def *def = &io_op_defs[req->opcode];
5605 struct io_ring_ctx *ctx = req->ctx;
5606 struct async_poll *apoll;
5607 struct io_poll_table ipt;
5608 __poll_t ret, mask = EPOLLONESHOT | POLLERR | POLLPRI;
5611 if (!req->file || !file_can_poll(req->file))
5612 return IO_APOLL_ABORTED;
5613 if (req->flags & REQ_F_POLLED)
5614 return IO_APOLL_ABORTED;
5615 if (!def->pollin && !def->pollout)
5616 return IO_APOLL_ABORTED;
5620 mask |= POLLIN | POLLRDNORM;
5622 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5623 if ((req->opcode == IORING_OP_RECVMSG) &&
5624 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5628 mask |= POLLOUT | POLLWRNORM;
5631 /* if we can't nonblock try, then no point in arming a poll handler */
5632 if (!io_file_supports_nowait(req, rw))
5633 return IO_APOLL_ABORTED;
5635 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5636 if (unlikely(!apoll))
5637 return IO_APOLL_ABORTED;
5638 apoll->double_poll = NULL;
5640 req->flags |= REQ_F_POLLED;
5641 ipt.pt._qproc = io_async_queue_proc;
5642 io_req_set_refcount(req);
5644 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5646 spin_unlock(&ctx->completion_lock);
5647 if (ret || ipt.error)
5648 return ret ? IO_APOLL_READY : IO_APOLL_ABORTED;
5650 trace_io_uring_poll_arm(ctx, req, req->opcode, req->user_data,
5651 mask, apoll->poll.events);
5655 static bool __io_poll_remove_one(struct io_kiocb *req,
5656 struct io_poll_iocb *poll, bool do_cancel)
5657 __must_hold(&req->ctx->completion_lock)
5659 bool do_complete = false;
5663 spin_lock_irq(&poll->head->lock);
5665 WRITE_ONCE(poll->canceled, true);
5666 if (!list_empty(&poll->wait.entry)) {
5667 list_del_init(&poll->wait.entry);
5670 spin_unlock_irq(&poll->head->lock);
5671 hash_del(&req->hash_node);
5675 static bool io_poll_remove_one(struct io_kiocb *req)
5676 __must_hold(&req->ctx->completion_lock)
5680 io_poll_remove_double(req);
5681 do_complete = __io_poll_remove_one(req, io_poll_get_single(req), true);
5684 io_cqring_fill_event(req->ctx, req->user_data, -ECANCELED, 0);
5685 io_commit_cqring(req->ctx);
5687 io_put_req_deferred(req);
5693 * Returns true if we found and killed one or more poll requests
5695 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5698 struct hlist_node *tmp;
5699 struct io_kiocb *req;
5702 spin_lock(&ctx->completion_lock);
5703 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5704 struct hlist_head *list;
5706 list = &ctx->cancel_hash[i];
5707 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5708 if (io_match_task(req, tsk, cancel_all))
5709 posted += io_poll_remove_one(req);
5712 spin_unlock(&ctx->completion_lock);
5715 io_cqring_ev_posted(ctx);
5720 static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, __u64 sqe_addr,
5722 __must_hold(&ctx->completion_lock)
5724 struct hlist_head *list;
5725 struct io_kiocb *req;
5727 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5728 hlist_for_each_entry(req, list, hash_node) {
5729 if (sqe_addr != req->user_data)
5731 if (poll_only && req->opcode != IORING_OP_POLL_ADD)
5738 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr,
5740 __must_hold(&ctx->completion_lock)
5742 struct io_kiocb *req;
5744 req = io_poll_find(ctx, sqe_addr, poll_only);
5747 if (io_poll_remove_one(req))
5753 static __poll_t io_poll_parse_events(const struct io_uring_sqe *sqe,
5758 events = READ_ONCE(sqe->poll32_events);
5760 events = swahw32(events);
5762 if (!(flags & IORING_POLL_ADD_MULTI))
5763 events |= EPOLLONESHOT;
5764 return demangle_poll(events) | (events & (EPOLLEXCLUSIVE|EPOLLONESHOT));
5767 static int io_poll_update_prep(struct io_kiocb *req,
5768 const struct io_uring_sqe *sqe)
5770 struct io_poll_update *upd = &req->poll_update;
5773 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5775 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
5777 flags = READ_ONCE(sqe->len);
5778 if (flags & ~(IORING_POLL_UPDATE_EVENTS | IORING_POLL_UPDATE_USER_DATA |
5779 IORING_POLL_ADD_MULTI))
5781 /* meaningless without update */
5782 if (flags == IORING_POLL_ADD_MULTI)
5785 upd->old_user_data = READ_ONCE(sqe->addr);
5786 upd->update_events = flags & IORING_POLL_UPDATE_EVENTS;
5787 upd->update_user_data = flags & IORING_POLL_UPDATE_USER_DATA;
5789 upd->new_user_data = READ_ONCE(sqe->off);
5790 if (!upd->update_user_data && upd->new_user_data)
5792 if (upd->update_events)
5793 upd->events = io_poll_parse_events(sqe, flags);
5794 else if (sqe->poll32_events)
5800 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5803 struct io_kiocb *req = wait->private;
5804 struct io_poll_iocb *poll = &req->poll;
5806 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5809 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5810 struct poll_table_struct *p)
5812 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5814 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5817 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5819 struct io_poll_iocb *poll = &req->poll;
5822 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5824 if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->addr)
5826 flags = READ_ONCE(sqe->len);
5827 if (flags & ~IORING_POLL_ADD_MULTI)
5830 io_req_set_refcount(req);
5831 poll->events = io_poll_parse_events(sqe, flags);
5835 static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5837 struct io_poll_iocb *poll = &req->poll;
5838 struct io_ring_ctx *ctx = req->ctx;
5839 struct io_poll_table ipt;
5843 ipt.pt._qproc = io_poll_queue_proc;
5845 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5848 if (mask) { /* no async, we'd stolen it */
5850 done = io_poll_complete(req, mask);
5852 spin_unlock(&ctx->completion_lock);
5855 io_cqring_ev_posted(ctx);
5862 static int io_poll_update(struct io_kiocb *req, unsigned int issue_flags)
5864 struct io_ring_ctx *ctx = req->ctx;
5865 struct io_kiocb *preq;
5869 spin_lock(&ctx->completion_lock);
5870 preq = io_poll_find(ctx, req->poll_update.old_user_data, true);
5876 if (!req->poll_update.update_events && !req->poll_update.update_user_data) {
5878 ret = io_poll_remove_one(preq) ? 0 : -EALREADY;
5883 * Don't allow racy completion with singleshot, as we cannot safely
5884 * update those. For multishot, if we're racing with completion, just
5885 * let completion re-add it.
5887 completing = !__io_poll_remove_one(preq, &preq->poll, false);
5888 if (completing && (preq->poll.events & EPOLLONESHOT)) {
5892 /* we now have a detached poll request. reissue. */
5896 spin_unlock(&ctx->completion_lock);
5898 io_req_complete(req, ret);
5901 /* only mask one event flags, keep behavior flags */
5902 if (req->poll_update.update_events) {
5903 preq->poll.events &= ~0xffff;
5904 preq->poll.events |= req->poll_update.events & 0xffff;
5905 preq->poll.events |= IO_POLL_UNMASK;
5907 if (req->poll_update.update_user_data)
5908 preq->user_data = req->poll_update.new_user_data;
5909 spin_unlock(&ctx->completion_lock);
5911 /* complete update request, we're done with it */
5912 io_req_complete(req, ret);
5915 ret = io_poll_add(preq, issue_flags);
5918 io_req_complete(preq, ret);
5924 static void io_req_task_timeout(struct io_kiocb *req, bool *locked)
5927 io_req_complete_post(req, -ETIME, 0);
5930 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5932 struct io_timeout_data *data = container_of(timer,
5933 struct io_timeout_data, timer);
5934 struct io_kiocb *req = data->req;
5935 struct io_ring_ctx *ctx = req->ctx;
5936 unsigned long flags;
5938 spin_lock_irqsave(&ctx->timeout_lock, flags);
5939 list_del_init(&req->timeout.list);
5940 atomic_set(&req->ctx->cq_timeouts,
5941 atomic_read(&req->ctx->cq_timeouts) + 1);
5942 spin_unlock_irqrestore(&ctx->timeout_lock, flags);
5944 req->io_task_work.func = io_req_task_timeout;
5945 io_req_task_work_add(req);
5946 return HRTIMER_NORESTART;
5949 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5951 __must_hold(&ctx->timeout_lock)
5953 struct io_timeout_data *io;
5954 struct io_kiocb *req;
5957 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5958 found = user_data == req->user_data;
5963 return ERR_PTR(-ENOENT);
5965 io = req->async_data;
5966 if (hrtimer_try_to_cancel(&io->timer) == -1)
5967 return ERR_PTR(-EALREADY);
5968 list_del_init(&req->timeout.list);
5972 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5973 __must_hold(&ctx->completion_lock)
5974 __must_hold(&ctx->timeout_lock)
5976 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5979 return PTR_ERR(req);
5982 io_cqring_fill_event(ctx, req->user_data, -ECANCELED, 0);
5983 io_put_req_deferred(req);
5987 static clockid_t io_timeout_get_clock(struct io_timeout_data *data)
5989 switch (data->flags & IORING_TIMEOUT_CLOCK_MASK) {
5990 case IORING_TIMEOUT_BOOTTIME:
5991 return CLOCK_BOOTTIME;
5992 case IORING_TIMEOUT_REALTIME:
5993 return CLOCK_REALTIME;
5995 /* can't happen, vetted at prep time */
5999 return CLOCK_MONOTONIC;
6003 static int io_linked_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
6004 struct timespec64 *ts, enum hrtimer_mode mode)
6005 __must_hold(&ctx->timeout_lock)
6007 struct io_timeout_data *io;
6008 struct io_kiocb *req;
6011 list_for_each_entry(req, &ctx->ltimeout_list, timeout.list) {
6012 found = user_data == req->user_data;
6019 io = req->async_data;
6020 if (hrtimer_try_to_cancel(&io->timer) == -1)
6022 hrtimer_init(&io->timer, io_timeout_get_clock(io), mode);
6023 io->timer.function = io_link_timeout_fn;
6024 hrtimer_start(&io->timer, timespec64_to_ktime(*ts), mode);
6028 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
6029 struct timespec64 *ts, enum hrtimer_mode mode)
6030 __must_hold(&ctx->timeout_lock)
6032 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
6033 struct io_timeout_data *data;
6036 return PTR_ERR(req);
6038 req->timeout.off = 0; /* noseq */
6039 data = req->async_data;
6040 list_add_tail(&req->timeout.list, &ctx->timeout_list);
6041 hrtimer_init(&data->timer, io_timeout_get_clock(data), mode);
6042 data->timer.function = io_timeout_fn;
6043 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
6047 static int io_timeout_remove_prep(struct io_kiocb *req,
6048 const struct io_uring_sqe *sqe)
6050 struct io_timeout_rem *tr = &req->timeout_rem;
6052 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
6054 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
6056 if (sqe->ioprio || sqe->buf_index || sqe->len || sqe->splice_fd_in)
6059 tr->ltimeout = false;
6060 tr->addr = READ_ONCE(sqe->addr);
6061 tr->flags = READ_ONCE(sqe->timeout_flags);
6062 if (tr->flags & IORING_TIMEOUT_UPDATE_MASK) {
6063 if (hweight32(tr->flags & IORING_TIMEOUT_CLOCK_MASK) > 1)
6065 if (tr->flags & IORING_LINK_TIMEOUT_UPDATE)
6066 tr->ltimeout = true;
6067 if (tr->flags & ~(IORING_TIMEOUT_UPDATE_MASK|IORING_TIMEOUT_ABS))
6069 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
6071 } else if (tr->flags) {
6072 /* timeout removal doesn't support flags */
6079 static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
6081 return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
6086 * Remove or update an existing timeout command
6088 static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
6090 struct io_timeout_rem *tr = &req->timeout_rem;
6091 struct io_ring_ctx *ctx = req->ctx;
6094 if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE)) {
6095 spin_lock(&ctx->completion_lock);
6096 spin_lock_irq(&ctx->timeout_lock);
6097 ret = io_timeout_cancel(ctx, tr->addr);
6098 spin_unlock_irq(&ctx->timeout_lock);
6099 spin_unlock(&ctx->completion_lock);
6101 enum hrtimer_mode mode = io_translate_timeout_mode(tr->flags);
6103 spin_lock_irq(&ctx->timeout_lock);
6105 ret = io_linked_timeout_update(ctx, tr->addr, &tr->ts, mode);
6107 ret = io_timeout_update(ctx, tr->addr, &tr->ts, mode);
6108 spin_unlock_irq(&ctx->timeout_lock);
6113 io_req_complete_post(req, ret, 0);
6117 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6118 bool is_timeout_link)
6120 struct io_timeout_data *data;
6122 u32 off = READ_ONCE(sqe->off);
6124 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
6126 if (sqe->ioprio || sqe->buf_index || sqe->len != 1 ||
6129 if (off && is_timeout_link)
6131 flags = READ_ONCE(sqe->timeout_flags);
6132 if (flags & ~(IORING_TIMEOUT_ABS | IORING_TIMEOUT_CLOCK_MASK))
6134 /* more than one clock specified is invalid, obviously */
6135 if (hweight32(flags & IORING_TIMEOUT_CLOCK_MASK) > 1)
6138 INIT_LIST_HEAD(&req->timeout.list);
6139 req->timeout.off = off;
6140 if (unlikely(off && !req->ctx->off_timeout_used))
6141 req->ctx->off_timeout_used = true;
6143 if (!req->async_data && io_alloc_async_data(req))
6146 data = req->async_data;
6148 data->flags = flags;
6150 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
6153 data->mode = io_translate_timeout_mode(flags);
6154 hrtimer_init(&data->timer, io_timeout_get_clock(data), data->mode);
6156 if (is_timeout_link) {
6157 struct io_submit_link *link = &req->ctx->submit_state.link;
6161 if (link->last->opcode == IORING_OP_LINK_TIMEOUT)
6163 req->timeout.head = link->last;
6164 link->last->flags |= REQ_F_ARM_LTIMEOUT;
6169 static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
6171 struct io_ring_ctx *ctx = req->ctx;
6172 struct io_timeout_data *data = req->async_data;
6173 struct list_head *entry;
6174 u32 tail, off = req->timeout.off;
6176 spin_lock_irq(&ctx->timeout_lock);
6179 * sqe->off holds how many events that need to occur for this
6180 * timeout event to be satisfied. If it isn't set, then this is
6181 * a pure timeout request, sequence isn't used.
6183 if (io_is_timeout_noseq(req)) {
6184 entry = ctx->timeout_list.prev;
6188 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
6189 req->timeout.target_seq = tail + off;
6191 /* Update the last seq here in case io_flush_timeouts() hasn't.
6192 * This is safe because ->completion_lock is held, and submissions
6193 * and completions are never mixed in the same ->completion_lock section.
6195 ctx->cq_last_tm_flush = tail;
6198 * Insertion sort, ensuring the first entry in the list is always
6199 * the one we need first.
6201 list_for_each_prev(entry, &ctx->timeout_list) {
6202 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
6205 if (io_is_timeout_noseq(nxt))
6207 /* nxt.seq is behind @tail, otherwise would've been completed */
6208 if (off >= nxt->timeout.target_seq - tail)
6212 list_add(&req->timeout.list, entry);
6213 data->timer.function = io_timeout_fn;
6214 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
6215 spin_unlock_irq(&ctx->timeout_lock);
6219 struct io_cancel_data {
6220 struct io_ring_ctx *ctx;
6224 static bool io_cancel_cb(struct io_wq_work *work, void *data)
6226 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6227 struct io_cancel_data *cd = data;
6229 return req->ctx == cd->ctx && req->user_data == cd->user_data;
6232 static int io_async_cancel_one(struct io_uring_task *tctx, u64 user_data,
6233 struct io_ring_ctx *ctx)
6235 struct io_cancel_data data = { .ctx = ctx, .user_data = user_data, };
6236 enum io_wq_cancel cancel_ret;
6239 if (!tctx || !tctx->io_wq)
6242 cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, &data, false);
6243 switch (cancel_ret) {
6244 case IO_WQ_CANCEL_OK:
6247 case IO_WQ_CANCEL_RUNNING:
6250 case IO_WQ_CANCEL_NOTFOUND:
6258 static int io_try_cancel_userdata(struct io_kiocb *req, u64 sqe_addr)
6260 struct io_ring_ctx *ctx = req->ctx;
6263 WARN_ON_ONCE(!io_wq_current_is_worker() && req->task != current);
6265 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
6269 spin_lock(&ctx->completion_lock);
6270 spin_lock_irq(&ctx->timeout_lock);
6271 ret = io_timeout_cancel(ctx, sqe_addr);
6272 spin_unlock_irq(&ctx->timeout_lock);
6275 ret = io_poll_cancel(ctx, sqe_addr, false);
6277 spin_unlock(&ctx->completion_lock);
6281 static int io_async_cancel_prep(struct io_kiocb *req,
6282 const struct io_uring_sqe *sqe)
6284 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
6286 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
6288 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags ||
6292 req->cancel.addr = READ_ONCE(sqe->addr);
6296 static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
6298 struct io_ring_ctx *ctx = req->ctx;
6299 u64 sqe_addr = req->cancel.addr;
6300 struct io_tctx_node *node;
6303 ret = io_try_cancel_userdata(req, sqe_addr);
6307 /* slow path, try all io-wq's */
6308 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
6310 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
6311 struct io_uring_task *tctx = node->task->io_uring;
6313 ret = io_async_cancel_one(tctx, req->cancel.addr, ctx);
6317 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
6321 io_req_complete_post(req, ret, 0);
6325 static int io_rsrc_update_prep(struct io_kiocb *req,
6326 const struct io_uring_sqe *sqe)
6328 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
6330 if (sqe->ioprio || sqe->rw_flags || sqe->splice_fd_in)
6333 req->rsrc_update.offset = READ_ONCE(sqe->off);
6334 req->rsrc_update.nr_args = READ_ONCE(sqe->len);
6335 if (!req->rsrc_update.nr_args)
6337 req->rsrc_update.arg = READ_ONCE(sqe->addr);
6341 static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
6343 struct io_ring_ctx *ctx = req->ctx;
6344 struct io_uring_rsrc_update2 up;
6347 up.offset = req->rsrc_update.offset;
6348 up.data = req->rsrc_update.arg;
6353 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
6354 ret = __io_register_rsrc_update(ctx, IORING_RSRC_FILE,
6355 &up, req->rsrc_update.nr_args);
6356 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
6360 __io_req_complete(req, issue_flags, ret, 0);
6364 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
6366 switch (req->opcode) {
6369 case IORING_OP_READV:
6370 case IORING_OP_READ_FIXED:
6371 case IORING_OP_READ:
6372 return io_read_prep(req, sqe);
6373 case IORING_OP_WRITEV:
6374 case IORING_OP_WRITE_FIXED:
6375 case IORING_OP_WRITE:
6376 return io_write_prep(req, sqe);
6377 case IORING_OP_POLL_ADD:
6378 return io_poll_add_prep(req, sqe);
6379 case IORING_OP_POLL_REMOVE:
6380 return io_poll_update_prep(req, sqe);
6381 case IORING_OP_FSYNC:
6382 return io_fsync_prep(req, sqe);
6383 case IORING_OP_SYNC_FILE_RANGE:
6384 return io_sfr_prep(req, sqe);
6385 case IORING_OP_SENDMSG:
6386 case IORING_OP_SEND:
6387 return io_sendmsg_prep(req, sqe);
6388 case IORING_OP_RECVMSG:
6389 case IORING_OP_RECV:
6390 return io_recvmsg_prep(req, sqe);
6391 case IORING_OP_CONNECT:
6392 return io_connect_prep(req, sqe);
6393 case IORING_OP_TIMEOUT:
6394 return io_timeout_prep(req, sqe, false);
6395 case IORING_OP_TIMEOUT_REMOVE:
6396 return io_timeout_remove_prep(req, sqe);
6397 case IORING_OP_ASYNC_CANCEL:
6398 return io_async_cancel_prep(req, sqe);
6399 case IORING_OP_LINK_TIMEOUT:
6400 return io_timeout_prep(req, sqe, true);
6401 case IORING_OP_ACCEPT:
6402 return io_accept_prep(req, sqe);
6403 case IORING_OP_FALLOCATE:
6404 return io_fallocate_prep(req, sqe);
6405 case IORING_OP_OPENAT:
6406 return io_openat_prep(req, sqe);
6407 case IORING_OP_CLOSE:
6408 return io_close_prep(req, sqe);
6409 case IORING_OP_FILES_UPDATE:
6410 return io_rsrc_update_prep(req, sqe);
6411 case IORING_OP_STATX:
6412 return io_statx_prep(req, sqe);
6413 case IORING_OP_FADVISE:
6414 return io_fadvise_prep(req, sqe);
6415 case IORING_OP_MADVISE:
6416 return io_madvise_prep(req, sqe);
6417 case IORING_OP_OPENAT2:
6418 return io_openat2_prep(req, sqe);
6419 case IORING_OP_EPOLL_CTL:
6420 return io_epoll_ctl_prep(req, sqe);
6421 case IORING_OP_SPLICE:
6422 return io_splice_prep(req, sqe);
6423 case IORING_OP_PROVIDE_BUFFERS:
6424 return io_provide_buffers_prep(req, sqe);
6425 case IORING_OP_REMOVE_BUFFERS:
6426 return io_remove_buffers_prep(req, sqe);
6428 return io_tee_prep(req, sqe);
6429 case IORING_OP_SHUTDOWN:
6430 return io_shutdown_prep(req, sqe);
6431 case IORING_OP_RENAMEAT:
6432 return io_renameat_prep(req, sqe);
6433 case IORING_OP_UNLINKAT:
6434 return io_unlinkat_prep(req, sqe);
6435 case IORING_OP_MKDIRAT:
6436 return io_mkdirat_prep(req, sqe);
6437 case IORING_OP_SYMLINKAT:
6438 return io_symlinkat_prep(req, sqe);
6439 case IORING_OP_LINKAT:
6440 return io_linkat_prep(req, sqe);
6443 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
6448 static int io_req_prep_async(struct io_kiocb *req)
6450 if (!io_op_defs[req->opcode].needs_async_setup)
6452 if (WARN_ON_ONCE(req->async_data))
6454 if (io_alloc_async_data(req))
6457 switch (req->opcode) {
6458 case IORING_OP_READV:
6459 return io_rw_prep_async(req, READ);
6460 case IORING_OP_WRITEV:
6461 return io_rw_prep_async(req, WRITE);
6462 case IORING_OP_SENDMSG:
6463 return io_sendmsg_prep_async(req);
6464 case IORING_OP_RECVMSG:
6465 return io_recvmsg_prep_async(req);
6466 case IORING_OP_CONNECT:
6467 return io_connect_prep_async(req);
6469 printk_once(KERN_WARNING "io_uring: prep_async() bad opcode %d\n",
6474 static u32 io_get_sequence(struct io_kiocb *req)
6476 u32 seq = req->ctx->cached_sq_head;
6478 /* need original cached_sq_head, but it was increased for each req */
6479 io_for_each_link(req, req)
6484 static bool io_drain_req(struct io_kiocb *req)
6486 struct io_kiocb *pos;
6487 struct io_ring_ctx *ctx = req->ctx;
6488 struct io_defer_entry *de;
6492 if (req->flags & REQ_F_FAIL) {
6493 io_req_complete_fail_submit(req);
6498 * If we need to drain a request in the middle of a link, drain the
6499 * head request and the next request/link after the current link.
6500 * Considering sequential execution of links, IOSQE_IO_DRAIN will be
6501 * maintained for every request of our link.
6503 if (ctx->drain_next) {
6504 req->flags |= REQ_F_IO_DRAIN;
6505 ctx->drain_next = false;
6507 /* not interested in head, start from the first linked */
6508 io_for_each_link(pos, req->link) {
6509 if (pos->flags & REQ_F_IO_DRAIN) {
6510 ctx->drain_next = true;
6511 req->flags |= REQ_F_IO_DRAIN;
6516 /* Still need defer if there is pending req in defer list. */
6517 if (likely(list_empty_careful(&ctx->defer_list) &&
6518 !(req->flags & REQ_F_IO_DRAIN))) {
6519 ctx->drain_active = false;
6523 seq = io_get_sequence(req);
6524 /* Still a chance to pass the sequence check */
6525 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
6528 ret = io_req_prep_async(req);
6531 io_prep_async_link(req);
6532 de = kmalloc(sizeof(*de), GFP_KERNEL);
6536 io_req_complete_failed(req, ret);
6540 spin_lock(&ctx->completion_lock);
6541 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
6542 spin_unlock(&ctx->completion_lock);
6544 io_queue_async_work(req, NULL);
6548 trace_io_uring_defer(ctx, req, req->user_data);
6551 list_add_tail(&de->list, &ctx->defer_list);
6552 spin_unlock(&ctx->completion_lock);
6556 static void io_clean_op(struct io_kiocb *req)
6558 if (req->flags & REQ_F_BUFFER_SELECTED) {
6559 switch (req->opcode) {
6560 case IORING_OP_READV:
6561 case IORING_OP_READ_FIXED:
6562 case IORING_OP_READ:
6563 kfree((void *)(unsigned long)req->rw.addr);
6565 case IORING_OP_RECVMSG:
6566 case IORING_OP_RECV:
6567 kfree(req->sr_msg.kbuf);
6572 if (req->flags & REQ_F_NEED_CLEANUP) {
6573 switch (req->opcode) {
6574 case IORING_OP_READV:
6575 case IORING_OP_READ_FIXED:
6576 case IORING_OP_READ:
6577 case IORING_OP_WRITEV:
6578 case IORING_OP_WRITE_FIXED:
6579 case IORING_OP_WRITE: {
6580 struct io_async_rw *io = req->async_data;
6582 kfree(io->free_iovec);
6585 case IORING_OP_RECVMSG:
6586 case IORING_OP_SENDMSG: {
6587 struct io_async_msghdr *io = req->async_data;
6589 kfree(io->free_iov);
6592 case IORING_OP_SPLICE:
6594 if (!(req->splice.flags & SPLICE_F_FD_IN_FIXED))
6595 io_put_file(req->splice.file_in);
6597 case IORING_OP_OPENAT:
6598 case IORING_OP_OPENAT2:
6599 if (req->open.filename)
6600 putname(req->open.filename);
6602 case IORING_OP_RENAMEAT:
6603 putname(req->rename.oldpath);
6604 putname(req->rename.newpath);
6606 case IORING_OP_UNLINKAT:
6607 putname(req->unlink.filename);
6609 case IORING_OP_MKDIRAT:
6610 putname(req->mkdir.filename);
6612 case IORING_OP_SYMLINKAT:
6613 putname(req->symlink.oldpath);
6614 putname(req->symlink.newpath);
6616 case IORING_OP_LINKAT:
6617 putname(req->hardlink.oldpath);
6618 putname(req->hardlink.newpath);
6622 if ((req->flags & REQ_F_POLLED) && req->apoll) {
6623 kfree(req->apoll->double_poll);
6627 if (req->flags & REQ_F_INFLIGHT) {
6628 struct io_uring_task *tctx = req->task->io_uring;
6630 atomic_dec(&tctx->inflight_tracked);
6632 if (req->flags & REQ_F_CREDS)
6633 put_cred(req->creds);
6635 req->flags &= ~IO_REQ_CLEAN_FLAGS;
6638 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
6640 struct io_ring_ctx *ctx = req->ctx;
6641 const struct cred *creds = NULL;
6644 if ((req->flags & REQ_F_CREDS) && req->creds != current_cred())
6645 creds = override_creds(req->creds);
6647 switch (req->opcode) {
6649 ret = io_nop(req, issue_flags);
6651 case IORING_OP_READV:
6652 case IORING_OP_READ_FIXED:
6653 case IORING_OP_READ:
6654 ret = io_read(req, issue_flags);
6656 case IORING_OP_WRITEV:
6657 case IORING_OP_WRITE_FIXED:
6658 case IORING_OP_WRITE:
6659 ret = io_write(req, issue_flags);
6661 case IORING_OP_FSYNC:
6662 ret = io_fsync(req, issue_flags);
6664 case IORING_OP_POLL_ADD:
6665 ret = io_poll_add(req, issue_flags);
6667 case IORING_OP_POLL_REMOVE:
6668 ret = io_poll_update(req, issue_flags);
6670 case IORING_OP_SYNC_FILE_RANGE:
6671 ret = io_sync_file_range(req, issue_flags);
6673 case IORING_OP_SENDMSG:
6674 ret = io_sendmsg(req, issue_flags);
6676 case IORING_OP_SEND:
6677 ret = io_send(req, issue_flags);
6679 case IORING_OP_RECVMSG:
6680 ret = io_recvmsg(req, issue_flags);
6682 case IORING_OP_RECV:
6683 ret = io_recv(req, issue_flags);
6685 case IORING_OP_TIMEOUT:
6686 ret = io_timeout(req, issue_flags);
6688 case IORING_OP_TIMEOUT_REMOVE:
6689 ret = io_timeout_remove(req, issue_flags);
6691 case IORING_OP_ACCEPT:
6692 ret = io_accept(req, issue_flags);
6694 case IORING_OP_CONNECT:
6695 ret = io_connect(req, issue_flags);
6697 case IORING_OP_ASYNC_CANCEL:
6698 ret = io_async_cancel(req, issue_flags);
6700 case IORING_OP_FALLOCATE:
6701 ret = io_fallocate(req, issue_flags);
6703 case IORING_OP_OPENAT:
6704 ret = io_openat(req, issue_flags);
6706 case IORING_OP_CLOSE:
6707 ret = io_close(req, issue_flags);
6709 case IORING_OP_FILES_UPDATE:
6710 ret = io_files_update(req, issue_flags);
6712 case IORING_OP_STATX:
6713 ret = io_statx(req, issue_flags);
6715 case IORING_OP_FADVISE:
6716 ret = io_fadvise(req, issue_flags);
6718 case IORING_OP_MADVISE:
6719 ret = io_madvise(req, issue_flags);
6721 case IORING_OP_OPENAT2:
6722 ret = io_openat2(req, issue_flags);
6724 case IORING_OP_EPOLL_CTL:
6725 ret = io_epoll_ctl(req, issue_flags);
6727 case IORING_OP_SPLICE:
6728 ret = io_splice(req, issue_flags);
6730 case IORING_OP_PROVIDE_BUFFERS:
6731 ret = io_provide_buffers(req, issue_flags);
6733 case IORING_OP_REMOVE_BUFFERS:
6734 ret = io_remove_buffers(req, issue_flags);
6737 ret = io_tee(req, issue_flags);
6739 case IORING_OP_SHUTDOWN:
6740 ret = io_shutdown(req, issue_flags);
6742 case IORING_OP_RENAMEAT:
6743 ret = io_renameat(req, issue_flags);
6745 case IORING_OP_UNLINKAT:
6746 ret = io_unlinkat(req, issue_flags);
6748 case IORING_OP_MKDIRAT:
6749 ret = io_mkdirat(req, issue_flags);
6751 case IORING_OP_SYMLINKAT:
6752 ret = io_symlinkat(req, issue_flags);
6754 case IORING_OP_LINKAT:
6755 ret = io_linkat(req, issue_flags);
6763 revert_creds(creds);
6766 /* If the op doesn't have a file, we're not polling for it */
6767 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file)
6768 io_iopoll_req_issued(req);
6773 static struct io_wq_work *io_wq_free_work(struct io_wq_work *work)
6775 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6777 req = io_put_req_find_next(req);
6778 return req ? &req->work : NULL;
6781 static void io_wq_submit_work(struct io_wq_work *work)
6783 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6784 struct io_kiocb *timeout;
6787 /* one will be dropped by ->io_free_work() after returning to io-wq */
6788 if (!(req->flags & REQ_F_REFCOUNT))
6789 __io_req_set_refcount(req, 2);
6793 timeout = io_prep_linked_timeout(req);
6795 io_queue_linked_timeout(timeout);
6797 /* either cancelled or io-wq is dying, so don't touch tctx->iowq */
6798 if (work->flags & IO_WQ_WORK_CANCEL)
6803 ret = io_issue_sqe(req, 0);
6805 * We can get EAGAIN for polled IO even though we're
6806 * forcing a sync submission from here, since we can't
6807 * wait for request slots on the block side.
6815 /* avoid locking problems by failing it from a clean context */
6817 io_req_task_queue_fail(req, ret);
6820 static inline struct io_fixed_file *io_fixed_file_slot(struct io_file_table *table,
6823 return &table->files[i];
6826 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6829 struct io_fixed_file *slot = io_fixed_file_slot(&ctx->file_table, index);
6831 return (struct file *) (slot->file_ptr & FFS_MASK);
6834 static void io_fixed_file_set(struct io_fixed_file *file_slot, struct file *file)
6836 unsigned long file_ptr = (unsigned long) file;
6838 if (__io_file_supports_nowait(file, READ))
6839 file_ptr |= FFS_ASYNC_READ;
6840 if (__io_file_supports_nowait(file, WRITE))
6841 file_ptr |= FFS_ASYNC_WRITE;
6842 if (S_ISREG(file_inode(file)->i_mode))
6843 file_ptr |= FFS_ISREG;
6844 file_slot->file_ptr = file_ptr;
6847 static inline struct file *io_file_get_fixed(struct io_ring_ctx *ctx,
6848 struct io_kiocb *req, int fd)
6851 unsigned long file_ptr;
6853 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6855 fd = array_index_nospec(fd, ctx->nr_user_files);
6856 file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr;
6857 file = (struct file *) (file_ptr & FFS_MASK);
6858 file_ptr &= ~FFS_MASK;
6859 /* mask in overlapping REQ_F and FFS bits */
6860 req->flags |= (file_ptr << REQ_F_NOWAIT_READ_BIT);
6861 io_req_set_rsrc_node(req);
6865 static struct file *io_file_get_normal(struct io_ring_ctx *ctx,
6866 struct io_kiocb *req, int fd)
6868 struct file *file = fget(fd);
6870 trace_io_uring_file_get(ctx, fd);
6872 /* we don't allow fixed io_uring files */
6873 if (file && unlikely(file->f_op == &io_uring_fops))
6874 io_req_track_inflight(req);
6878 static inline struct file *io_file_get(struct io_ring_ctx *ctx,
6879 struct io_kiocb *req, int fd, bool fixed)
6882 return io_file_get_fixed(ctx, req, fd);
6884 return io_file_get_normal(ctx, req, fd);
6887 static void io_req_task_link_timeout(struct io_kiocb *req, bool *locked)
6889 struct io_kiocb *prev = req->timeout.prev;
6893 ret = io_try_cancel_userdata(req, prev->user_data);
6894 io_req_complete_post(req, ret ?: -ETIME, 0);
6897 io_req_complete_post(req, -ETIME, 0);
6901 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6903 struct io_timeout_data *data = container_of(timer,
6904 struct io_timeout_data, timer);
6905 struct io_kiocb *prev, *req = data->req;
6906 struct io_ring_ctx *ctx = req->ctx;
6907 unsigned long flags;
6909 spin_lock_irqsave(&ctx->timeout_lock, flags);
6910 prev = req->timeout.head;
6911 req->timeout.head = NULL;
6914 * We don't expect the list to be empty, that will only happen if we
6915 * race with the completion of the linked work.
6918 io_remove_next_linked(prev);
6919 if (!req_ref_inc_not_zero(prev))
6922 list_del(&req->timeout.list);
6923 req->timeout.prev = prev;
6924 spin_unlock_irqrestore(&ctx->timeout_lock, flags);
6926 req->io_task_work.func = io_req_task_link_timeout;
6927 io_req_task_work_add(req);
6928 return HRTIMER_NORESTART;
6931 static void io_queue_linked_timeout(struct io_kiocb *req)
6933 struct io_ring_ctx *ctx = req->ctx;
6935 spin_lock_irq(&ctx->timeout_lock);
6937 * If the back reference is NULL, then our linked request finished
6938 * before we got a chance to setup the timer
6940 if (req->timeout.head) {
6941 struct io_timeout_data *data = req->async_data;
6943 data->timer.function = io_link_timeout_fn;
6944 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6946 list_add_tail(&req->timeout.list, &ctx->ltimeout_list);
6948 spin_unlock_irq(&ctx->timeout_lock);
6949 /* drop submission reference */
6953 static void __io_queue_sqe(struct io_kiocb *req)
6954 __must_hold(&req->ctx->uring_lock)
6956 struct io_kiocb *linked_timeout;
6960 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
6963 * We async punt it if the file wasn't marked NOWAIT, or if the file
6964 * doesn't support non-blocking read/write attempts
6967 if (req->flags & REQ_F_COMPLETE_INLINE) {
6968 struct io_ring_ctx *ctx = req->ctx;
6969 struct io_submit_state *state = &ctx->submit_state;
6971 state->compl_reqs[state->compl_nr++] = req;
6972 if (state->compl_nr == ARRAY_SIZE(state->compl_reqs))
6973 io_submit_flush_completions(ctx);
6977 linked_timeout = io_prep_linked_timeout(req);
6979 io_queue_linked_timeout(linked_timeout);
6980 } else if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6981 linked_timeout = io_prep_linked_timeout(req);
6983 switch (io_arm_poll_handler(req)) {
6984 case IO_APOLL_READY:
6986 io_unprep_linked_timeout(req);
6988 case IO_APOLL_ABORTED:
6990 * Queued up for async execution, worker will release
6991 * submit reference when the iocb is actually submitted.
6993 io_queue_async_work(req, NULL);
6998 io_queue_linked_timeout(linked_timeout);
7000 io_req_complete_failed(req, ret);
7004 static inline void io_queue_sqe(struct io_kiocb *req)
7005 __must_hold(&req->ctx->uring_lock)
7007 if (unlikely(req->ctx->drain_active) && io_drain_req(req))
7010 if (likely(!(req->flags & (REQ_F_FORCE_ASYNC | REQ_F_FAIL)))) {
7011 __io_queue_sqe(req);
7012 } else if (req->flags & REQ_F_FAIL) {
7013 io_req_complete_fail_submit(req);
7015 int ret = io_req_prep_async(req);
7018 io_req_complete_failed(req, ret);
7020 io_queue_async_work(req, NULL);
7025 * Check SQE restrictions (opcode and flags).
7027 * Returns 'true' if SQE is allowed, 'false' otherwise.
7029 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
7030 struct io_kiocb *req,
7031 unsigned int sqe_flags)
7033 if (likely(!ctx->restricted))
7036 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
7039 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
7040 ctx->restrictions.sqe_flags_required)
7043 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
7044 ctx->restrictions.sqe_flags_required))
7050 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
7051 const struct io_uring_sqe *sqe)
7052 __must_hold(&ctx->uring_lock)
7054 struct io_submit_state *state;
7055 unsigned int sqe_flags;
7056 int personality, ret = 0;
7058 /* req is partially pre-initialised, see io_preinit_req() */
7059 req->opcode = READ_ONCE(sqe->opcode);
7060 /* same numerical values with corresponding REQ_F_*, safe to copy */
7061 req->flags = sqe_flags = READ_ONCE(sqe->flags);
7062 req->user_data = READ_ONCE(sqe->user_data);
7064 req->fixed_rsrc_refs = NULL;
7065 req->task = current;
7067 /* enforce forwards compatibility on users */
7068 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
7070 if (unlikely(req->opcode >= IORING_OP_LAST))
7072 if (!io_check_restriction(ctx, req, sqe_flags))
7075 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
7076 !io_op_defs[req->opcode].buffer_select)
7078 if (unlikely(sqe_flags & IOSQE_IO_DRAIN))
7079 ctx->drain_active = true;
7081 personality = READ_ONCE(sqe->personality);
7083 req->creds = xa_load(&ctx->personalities, personality);
7086 get_cred(req->creds);
7087 req->flags |= REQ_F_CREDS;
7089 state = &ctx->submit_state;
7092 * Plug now if we have more than 1 IO left after this, and the target
7093 * is potentially a read/write to block based storage.
7095 if (!state->plug_started && state->ios_left > 1 &&
7096 io_op_defs[req->opcode].plug) {
7097 blk_start_plug(&state->plug);
7098 state->plug_started = true;
7101 if (io_op_defs[req->opcode].needs_file) {
7102 req->file = io_file_get(ctx, req, READ_ONCE(sqe->fd),
7103 (sqe_flags & IOSQE_FIXED_FILE));
7104 if (unlikely(!req->file))
7112 static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
7113 const struct io_uring_sqe *sqe)
7114 __must_hold(&ctx->uring_lock)
7116 struct io_submit_link *link = &ctx->submit_state.link;
7119 ret = io_init_req(ctx, req, sqe);
7120 if (unlikely(ret)) {
7122 /* fail even hard links since we don't submit */
7125 * we can judge a link req is failed or cancelled by if
7126 * REQ_F_FAIL is set, but the head is an exception since
7127 * it may be set REQ_F_FAIL because of other req's failure
7128 * so let's leverage req->result to distinguish if a head
7129 * is set REQ_F_FAIL because of its failure or other req's
7130 * failure so that we can set the correct ret code for it.
7131 * init result here to avoid affecting the normal path.
7133 if (!(link->head->flags & REQ_F_FAIL))
7134 req_fail_link_node(link->head, -ECANCELED);
7135 } else if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
7137 * the current req is a normal req, we should return
7138 * error and thus break the submittion loop.
7140 io_req_complete_failed(req, ret);
7143 req_fail_link_node(req, ret);
7145 ret = io_req_prep(req, sqe);
7150 /* don't need @sqe from now on */
7151 trace_io_uring_submit_sqe(ctx, req, req->opcode, req->user_data,
7153 ctx->flags & IORING_SETUP_SQPOLL);
7156 * If we already have a head request, queue this one for async
7157 * submittal once the head completes. If we don't have a head but
7158 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
7159 * submitted sync once the chain is complete. If none of those
7160 * conditions are true (normal request), then just queue it.
7163 struct io_kiocb *head = link->head;
7165 if (!(req->flags & REQ_F_FAIL)) {
7166 ret = io_req_prep_async(req);
7167 if (unlikely(ret)) {
7168 req_fail_link_node(req, ret);
7169 if (!(head->flags & REQ_F_FAIL))
7170 req_fail_link_node(head, -ECANCELED);
7173 trace_io_uring_link(ctx, req, head);
7174 link->last->link = req;
7177 /* last request of a link, enqueue the link */
7178 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
7183 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
7195 * Batched submission is done, ensure local IO is flushed out.
7197 static void io_submit_state_end(struct io_submit_state *state,
7198 struct io_ring_ctx *ctx)
7200 if (state->link.head)
7201 io_queue_sqe(state->link.head);
7202 if (state->compl_nr)
7203 io_submit_flush_completions(ctx);
7204 if (state->plug_started)
7205 blk_finish_plug(&state->plug);
7209 * Start submission side cache.
7211 static void io_submit_state_start(struct io_submit_state *state,
7212 unsigned int max_ios)
7214 state->plug_started = false;
7215 state->ios_left = max_ios;
7216 /* set only head, no need to init link_last in advance */
7217 state->link.head = NULL;
7220 static void io_commit_sqring(struct io_ring_ctx *ctx)
7222 struct io_rings *rings = ctx->rings;
7225 * Ensure any loads from the SQEs are done at this point,
7226 * since once we write the new head, the application could
7227 * write new data to them.
7229 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
7233 * Fetch an sqe, if one is available. Note this returns a pointer to memory
7234 * that is mapped by userspace. This means that care needs to be taken to
7235 * ensure that reads are stable, as we cannot rely on userspace always
7236 * being a good citizen. If members of the sqe are validated and then later
7237 * used, it's important that those reads are done through READ_ONCE() to
7238 * prevent a re-load down the line.
7240 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
7242 unsigned head, mask = ctx->sq_entries - 1;
7243 unsigned sq_idx = ctx->cached_sq_head++ & mask;
7246 * The cached sq head (or cq tail) serves two purposes:
7248 * 1) allows us to batch the cost of updating the user visible
7250 * 2) allows the kernel side to track the head on its own, even
7251 * though the application is the one updating it.
7253 head = READ_ONCE(ctx->sq_array[sq_idx]);
7254 if (likely(head < ctx->sq_entries))
7255 return &ctx->sq_sqes[head];
7257 /* drop invalid entries */
7259 WRITE_ONCE(ctx->rings->sq_dropped,
7260 READ_ONCE(ctx->rings->sq_dropped) + 1);
7264 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
7265 __must_hold(&ctx->uring_lock)
7269 /* make sure SQ entry isn't read before tail */
7270 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
7271 if (!percpu_ref_tryget_many(&ctx->refs, nr))
7273 io_get_task_refs(nr);
7275 io_submit_state_start(&ctx->submit_state, nr);
7276 while (submitted < nr) {
7277 const struct io_uring_sqe *sqe;
7278 struct io_kiocb *req;
7280 req = io_alloc_req(ctx);
7281 if (unlikely(!req)) {
7283 submitted = -EAGAIN;
7286 sqe = io_get_sqe(ctx);
7287 if (unlikely(!sqe)) {
7288 list_add(&req->inflight_entry, &ctx->submit_state.free_list);
7291 /* will complete beyond this point, count as submitted */
7293 if (io_submit_sqe(ctx, req, sqe))
7297 if (unlikely(submitted != nr)) {
7298 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
7299 int unused = nr - ref_used;
7301 current->io_uring->cached_refs += unused;
7302 percpu_ref_put_many(&ctx->refs, unused);
7305 io_submit_state_end(&ctx->submit_state, ctx);
7306 /* Commit SQ ring head once we've consumed and submitted all SQEs */
7307 io_commit_sqring(ctx);
7312 static inline bool io_sqd_events_pending(struct io_sq_data *sqd)
7314 return READ_ONCE(sqd->state);
7317 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
7319 /* Tell userspace we may need a wakeup call */
7320 spin_lock(&ctx->completion_lock);
7321 WRITE_ONCE(ctx->rings->sq_flags,
7322 ctx->rings->sq_flags | IORING_SQ_NEED_WAKEUP);
7323 spin_unlock(&ctx->completion_lock);
7326 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
7328 spin_lock(&ctx->completion_lock);
7329 WRITE_ONCE(ctx->rings->sq_flags,
7330 ctx->rings->sq_flags & ~IORING_SQ_NEED_WAKEUP);
7331 spin_unlock(&ctx->completion_lock);
7334 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
7336 unsigned int to_submit;
7339 to_submit = io_sqring_entries(ctx);
7340 /* if we're handling multiple rings, cap submit size for fairness */
7341 if (cap_entries && to_submit > IORING_SQPOLL_CAP_ENTRIES_VALUE)
7342 to_submit = IORING_SQPOLL_CAP_ENTRIES_VALUE;
7344 if (!list_empty(&ctx->iopoll_list) || to_submit) {
7345 unsigned nr_events = 0;
7346 const struct cred *creds = NULL;
7348 if (ctx->sq_creds != current_cred())
7349 creds = override_creds(ctx->sq_creds);
7351 mutex_lock(&ctx->uring_lock);
7352 if (!list_empty(&ctx->iopoll_list))
7353 io_do_iopoll(ctx, &nr_events, 0);
7356 * Don't submit if refs are dying, good for io_uring_register(),
7357 * but also it is relied upon by io_ring_exit_work()
7359 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)) &&
7360 !(ctx->flags & IORING_SETUP_R_DISABLED))
7361 ret = io_submit_sqes(ctx, to_submit);
7362 mutex_unlock(&ctx->uring_lock);
7364 if (to_submit && wq_has_sleeper(&ctx->sqo_sq_wait))
7365 wake_up(&ctx->sqo_sq_wait);
7367 revert_creds(creds);
7373 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
7375 struct io_ring_ctx *ctx;
7376 unsigned sq_thread_idle = 0;
7378 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7379 sq_thread_idle = max(sq_thread_idle, ctx->sq_thread_idle);
7380 sqd->sq_thread_idle = sq_thread_idle;
7383 static bool io_sqd_handle_event(struct io_sq_data *sqd)
7385 bool did_sig = false;
7386 struct ksignal ksig;
7388 if (test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state) ||
7389 signal_pending(current)) {
7390 mutex_unlock(&sqd->lock);
7391 if (signal_pending(current))
7392 did_sig = get_signal(&ksig);
7394 mutex_lock(&sqd->lock);
7396 return did_sig || test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7399 static int io_sq_thread(void *data)
7401 struct io_sq_data *sqd = data;
7402 struct io_ring_ctx *ctx;
7403 unsigned long timeout = 0;
7404 char buf[TASK_COMM_LEN];
7407 snprintf(buf, sizeof(buf), "iou-sqp-%d", sqd->task_pid);
7408 set_task_comm(current, buf);
7410 if (sqd->sq_cpu != -1)
7411 set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
7413 set_cpus_allowed_ptr(current, cpu_online_mask);
7414 current->flags |= PF_NO_SETAFFINITY;
7416 mutex_lock(&sqd->lock);
7418 bool cap_entries, sqt_spin = false;
7420 if (io_sqd_events_pending(sqd) || signal_pending(current)) {
7421 if (io_sqd_handle_event(sqd))
7423 timeout = jiffies + sqd->sq_thread_idle;
7426 cap_entries = !list_is_singular(&sqd->ctx_list);
7427 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
7428 int ret = __io_sq_thread(ctx, cap_entries);
7430 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
7433 if (io_run_task_work())
7436 if (sqt_spin || !time_after(jiffies, timeout)) {
7439 timeout = jiffies + sqd->sq_thread_idle;
7443 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
7444 if (!io_sqd_events_pending(sqd) && !current->task_works) {
7445 bool needs_sched = true;
7447 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
7448 io_ring_set_wakeup_flag(ctx);
7450 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
7451 !list_empty_careful(&ctx->iopoll_list)) {
7452 needs_sched = false;
7455 if (io_sqring_entries(ctx)) {
7456 needs_sched = false;
7462 mutex_unlock(&sqd->lock);
7464 mutex_lock(&sqd->lock);
7466 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7467 io_ring_clear_wakeup_flag(ctx);
7470 finish_wait(&sqd->wait, &wait);
7471 timeout = jiffies + sqd->sq_thread_idle;
7474 io_uring_cancel_generic(true, sqd);
7476 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7477 io_ring_set_wakeup_flag(ctx);
7479 mutex_unlock(&sqd->lock);
7481 complete(&sqd->exited);
7485 struct io_wait_queue {
7486 struct wait_queue_entry wq;
7487 struct io_ring_ctx *ctx;
7489 unsigned nr_timeouts;
7492 static inline bool io_should_wake(struct io_wait_queue *iowq)
7494 struct io_ring_ctx *ctx = iowq->ctx;
7495 int dist = ctx->cached_cq_tail - (int) iowq->cq_tail;
7498 * Wake up if we have enough events, or if a timeout occurred since we
7499 * started waiting. For timeouts, we always want to return to userspace,
7500 * regardless of event count.
7502 return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
7505 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
7506 int wake_flags, void *key)
7508 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
7512 * Cannot safely flush overflowed CQEs from here, ensure we wake up
7513 * the task, and the next invocation will do it.
7515 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->check_cq_overflow))
7516 return autoremove_wake_function(curr, mode, wake_flags, key);
7520 static int io_run_task_work_sig(void)
7522 if (io_run_task_work())
7524 if (!signal_pending(current))
7526 if (test_thread_flag(TIF_NOTIFY_SIGNAL))
7527 return -ERESTARTSYS;
7531 /* when returns >0, the caller should retry */
7532 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
7533 struct io_wait_queue *iowq,
7534 signed long *timeout)
7538 /* make sure we run task_work before checking for signals */
7539 ret = io_run_task_work_sig();
7540 if (ret || io_should_wake(iowq))
7542 /* let the caller flush overflows, retry */
7543 if (test_bit(0, &ctx->check_cq_overflow))
7546 *timeout = schedule_timeout(*timeout);
7547 return !*timeout ? -ETIME : 1;
7551 * Wait until events become available, if we don't already have some. The
7552 * application must reap them itself, as they reside on the shared cq ring.
7554 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
7555 const sigset_t __user *sig, size_t sigsz,
7556 struct __kernel_timespec __user *uts)
7558 struct io_wait_queue iowq;
7559 struct io_rings *rings = ctx->rings;
7560 signed long timeout = MAX_SCHEDULE_TIMEOUT;
7564 io_cqring_overflow_flush(ctx);
7565 if (io_cqring_events(ctx) >= min_events)
7567 if (!io_run_task_work())
7572 struct timespec64 ts;
7574 if (get_timespec64(&ts, uts))
7576 timeout = timespec64_to_jiffies(&ts);
7580 #ifdef CONFIG_COMPAT
7581 if (in_compat_syscall())
7582 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
7586 ret = set_user_sigmask(sig, sigsz);
7592 init_waitqueue_func_entry(&iowq.wq, io_wake_function);
7593 iowq.wq.private = current;
7594 INIT_LIST_HEAD(&iowq.wq.entry);
7596 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
7597 iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events;
7599 trace_io_uring_cqring_wait(ctx, min_events);
7601 /* if we can't even flush overflow, don't wait for more */
7602 if (!io_cqring_overflow_flush(ctx)) {
7606 prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq,
7607 TASK_INTERRUPTIBLE);
7608 ret = io_cqring_wait_schedule(ctx, &iowq, &timeout);
7609 finish_wait(&ctx->cq_wait, &iowq.wq);
7613 restore_saved_sigmask_unless(ret == -EINTR);
7615 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
7618 static void io_free_page_table(void **table, size_t size)
7620 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7622 for (i = 0; i < nr_tables; i++)
7627 static void **io_alloc_page_table(size_t size)
7629 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7630 size_t init_size = size;
7633 table = kcalloc(nr_tables, sizeof(*table), GFP_KERNEL_ACCOUNT);
7637 for (i = 0; i < nr_tables; i++) {
7638 unsigned int this_size = min_t(size_t, size, PAGE_SIZE);
7640 table[i] = kzalloc(this_size, GFP_KERNEL_ACCOUNT);
7642 io_free_page_table(table, init_size);
7650 static void io_rsrc_node_destroy(struct io_rsrc_node *ref_node)
7652 percpu_ref_exit(&ref_node->refs);
7656 static void io_rsrc_node_ref_zero(struct percpu_ref *ref)
7658 struct io_rsrc_node *node = container_of(ref, struct io_rsrc_node, refs);
7659 struct io_ring_ctx *ctx = node->rsrc_data->ctx;
7660 unsigned long flags;
7661 bool first_add = false;
7663 spin_lock_irqsave(&ctx->rsrc_ref_lock, flags);
7666 while (!list_empty(&ctx->rsrc_ref_list)) {
7667 node = list_first_entry(&ctx->rsrc_ref_list,
7668 struct io_rsrc_node, node);
7669 /* recycle ref nodes in order */
7672 list_del(&node->node);
7673 first_add |= llist_add(&node->llist, &ctx->rsrc_put_llist);
7675 spin_unlock_irqrestore(&ctx->rsrc_ref_lock, flags);
7678 mod_delayed_work(system_wq, &ctx->rsrc_put_work, HZ);
7681 static struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx)
7683 struct io_rsrc_node *ref_node;
7685 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7689 if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
7694 INIT_LIST_HEAD(&ref_node->node);
7695 INIT_LIST_HEAD(&ref_node->rsrc_list);
7696 ref_node->done = false;
7700 static void io_rsrc_node_switch(struct io_ring_ctx *ctx,
7701 struct io_rsrc_data *data_to_kill)
7703 WARN_ON_ONCE(!ctx->rsrc_backup_node);
7704 WARN_ON_ONCE(data_to_kill && !ctx->rsrc_node);
7707 struct io_rsrc_node *rsrc_node = ctx->rsrc_node;
7709 rsrc_node->rsrc_data = data_to_kill;
7710 spin_lock_irq(&ctx->rsrc_ref_lock);
7711 list_add_tail(&rsrc_node->node, &ctx->rsrc_ref_list);
7712 spin_unlock_irq(&ctx->rsrc_ref_lock);
7714 atomic_inc(&data_to_kill->refs);
7715 percpu_ref_kill(&rsrc_node->refs);
7716 ctx->rsrc_node = NULL;
7719 if (!ctx->rsrc_node) {
7720 ctx->rsrc_node = ctx->rsrc_backup_node;
7721 ctx->rsrc_backup_node = NULL;
7725 static int io_rsrc_node_switch_start(struct io_ring_ctx *ctx)
7727 if (ctx->rsrc_backup_node)
7729 ctx->rsrc_backup_node = io_rsrc_node_alloc(ctx);
7730 return ctx->rsrc_backup_node ? 0 : -ENOMEM;
7733 static int io_rsrc_ref_quiesce(struct io_rsrc_data *data, struct io_ring_ctx *ctx)
7737 /* As we may drop ->uring_lock, other task may have started quiesce */
7741 data->quiesce = true;
7743 ret = io_rsrc_node_switch_start(ctx);
7746 io_rsrc_node_switch(ctx, data);
7748 /* kill initial ref, already quiesced if zero */
7749 if (atomic_dec_and_test(&data->refs))
7751 mutex_unlock(&ctx->uring_lock);
7752 flush_delayed_work(&ctx->rsrc_put_work);
7753 ret = wait_for_completion_interruptible(&data->done);
7755 mutex_lock(&ctx->uring_lock);
7759 atomic_inc(&data->refs);
7760 /* wait for all works potentially completing data->done */
7761 flush_delayed_work(&ctx->rsrc_put_work);
7762 reinit_completion(&data->done);
7764 ret = io_run_task_work_sig();
7765 mutex_lock(&ctx->uring_lock);
7767 data->quiesce = false;
7772 static u64 *io_get_tag_slot(struct io_rsrc_data *data, unsigned int idx)
7774 unsigned int off = idx & IO_RSRC_TAG_TABLE_MASK;
7775 unsigned int table_idx = idx >> IO_RSRC_TAG_TABLE_SHIFT;
7777 return &data->tags[table_idx][off];
7780 static void io_rsrc_data_free(struct io_rsrc_data *data)
7782 size_t size = data->nr * sizeof(data->tags[0][0]);
7785 io_free_page_table((void **)data->tags, size);
7789 static int io_rsrc_data_alloc(struct io_ring_ctx *ctx, rsrc_put_fn *do_put,
7790 u64 __user *utags, unsigned nr,
7791 struct io_rsrc_data **pdata)
7793 struct io_rsrc_data *data;
7797 data = kzalloc(sizeof(*data), GFP_KERNEL);
7800 data->tags = (u64 **)io_alloc_page_table(nr * sizeof(data->tags[0][0]));
7808 data->do_put = do_put;
7811 for (i = 0; i < nr; i++) {
7812 u64 *tag_slot = io_get_tag_slot(data, i);
7814 if (copy_from_user(tag_slot, &utags[i],
7820 atomic_set(&data->refs, 1);
7821 init_completion(&data->done);
7825 io_rsrc_data_free(data);
7829 static bool io_alloc_file_tables(struct io_file_table *table, unsigned nr_files)
7831 table->files = kvcalloc(nr_files, sizeof(table->files[0]),
7832 GFP_KERNEL_ACCOUNT);
7833 return !!table->files;
7836 static void io_free_file_tables(struct io_file_table *table)
7838 kvfree(table->files);
7839 table->files = NULL;
7842 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
7844 #if defined(CONFIG_UNIX)
7845 if (ctx->ring_sock) {
7846 struct sock *sock = ctx->ring_sock->sk;
7847 struct sk_buff *skb;
7849 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
7855 for (i = 0; i < ctx->nr_user_files; i++) {
7858 file = io_file_from_index(ctx, i);
7863 io_free_file_tables(&ctx->file_table);
7864 io_rsrc_data_free(ctx->file_data);
7865 ctx->file_data = NULL;
7866 ctx->nr_user_files = 0;
7869 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7873 if (!ctx->file_data)
7875 ret = io_rsrc_ref_quiesce(ctx->file_data, ctx);
7877 __io_sqe_files_unregister(ctx);
7881 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7882 __releases(&sqd->lock)
7884 WARN_ON_ONCE(sqd->thread == current);
7887 * Do the dance but not conditional clear_bit() because it'd race with
7888 * other threads incrementing park_pending and setting the bit.
7890 clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7891 if (atomic_dec_return(&sqd->park_pending))
7892 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7893 mutex_unlock(&sqd->lock);
7896 static void io_sq_thread_park(struct io_sq_data *sqd)
7897 __acquires(&sqd->lock)
7899 WARN_ON_ONCE(sqd->thread == current);
7901 atomic_inc(&sqd->park_pending);
7902 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7903 mutex_lock(&sqd->lock);
7905 wake_up_process(sqd->thread);
7908 static void io_sq_thread_stop(struct io_sq_data *sqd)
7910 WARN_ON_ONCE(sqd->thread == current);
7911 WARN_ON_ONCE(test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state));
7913 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7914 mutex_lock(&sqd->lock);
7916 wake_up_process(sqd->thread);
7917 mutex_unlock(&sqd->lock);
7918 wait_for_completion(&sqd->exited);
7921 static void io_put_sq_data(struct io_sq_data *sqd)
7923 if (refcount_dec_and_test(&sqd->refs)) {
7924 WARN_ON_ONCE(atomic_read(&sqd->park_pending));
7926 io_sq_thread_stop(sqd);
7931 static void io_sq_thread_finish(struct io_ring_ctx *ctx)
7933 struct io_sq_data *sqd = ctx->sq_data;
7936 io_sq_thread_park(sqd);
7937 list_del_init(&ctx->sqd_list);
7938 io_sqd_update_thread_idle(sqd);
7939 io_sq_thread_unpark(sqd);
7941 io_put_sq_data(sqd);
7942 ctx->sq_data = NULL;
7946 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7948 struct io_ring_ctx *ctx_attach;
7949 struct io_sq_data *sqd;
7952 f = fdget(p->wq_fd);
7954 return ERR_PTR(-ENXIO);
7955 if (f.file->f_op != &io_uring_fops) {
7957 return ERR_PTR(-EINVAL);
7960 ctx_attach = f.file->private_data;
7961 sqd = ctx_attach->sq_data;
7964 return ERR_PTR(-EINVAL);
7966 if (sqd->task_tgid != current->tgid) {
7968 return ERR_PTR(-EPERM);
7971 refcount_inc(&sqd->refs);
7976 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p,
7979 struct io_sq_data *sqd;
7982 if (p->flags & IORING_SETUP_ATTACH_WQ) {
7983 sqd = io_attach_sq_data(p);
7988 /* fall through for EPERM case, setup new sqd/task */
7989 if (PTR_ERR(sqd) != -EPERM)
7993 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7995 return ERR_PTR(-ENOMEM);
7997 atomic_set(&sqd->park_pending, 0);
7998 refcount_set(&sqd->refs, 1);
7999 INIT_LIST_HEAD(&sqd->ctx_list);
8000 mutex_init(&sqd->lock);
8001 init_waitqueue_head(&sqd->wait);
8002 init_completion(&sqd->exited);
8006 #if defined(CONFIG_UNIX)
8008 * Ensure the UNIX gc is aware of our file set, so we are certain that
8009 * the io_uring can be safely unregistered on process exit, even if we have
8010 * loops in the file referencing.
8012 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
8014 struct sock *sk = ctx->ring_sock->sk;
8015 struct scm_fp_list *fpl;
8016 struct sk_buff *skb;
8019 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
8023 skb = alloc_skb(0, GFP_KERNEL);
8032 fpl->user = get_uid(current_user());
8033 for (i = 0; i < nr; i++) {
8034 struct file *file = io_file_from_index(ctx, i + offset);
8038 fpl->fp[nr_files] = get_file(file);
8039 unix_inflight(fpl->user, fpl->fp[nr_files]);
8044 fpl->max = SCM_MAX_FD;
8045 fpl->count = nr_files;
8046 UNIXCB(skb).fp = fpl;
8047 skb->destructor = unix_destruct_scm;
8048 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
8049 skb_queue_head(&sk->sk_receive_queue, skb);
8051 for (i = 0; i < nr_files; i++)
8062 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
8063 * causes regular reference counting to break down. We rely on the UNIX
8064 * garbage collection to take care of this problem for us.
8066 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
8068 unsigned left, total;
8072 left = ctx->nr_user_files;
8074 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
8076 ret = __io_sqe_files_scm(ctx, this_files, total);
8080 total += this_files;
8086 while (total < ctx->nr_user_files) {
8087 struct file *file = io_file_from_index(ctx, total);
8097 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
8103 static void io_rsrc_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
8105 struct file *file = prsrc->file;
8106 #if defined(CONFIG_UNIX)
8107 struct sock *sock = ctx->ring_sock->sk;
8108 struct sk_buff_head list, *head = &sock->sk_receive_queue;
8109 struct sk_buff *skb;
8112 __skb_queue_head_init(&list);
8115 * Find the skb that holds this file in its SCM_RIGHTS. When found,
8116 * remove this entry and rearrange the file array.
8118 skb = skb_dequeue(head);
8120 struct scm_fp_list *fp;
8122 fp = UNIXCB(skb).fp;
8123 for (i = 0; i < fp->count; i++) {
8126 if (fp->fp[i] != file)
8129 unix_notinflight(fp->user, fp->fp[i]);
8130 left = fp->count - 1 - i;
8132 memmove(&fp->fp[i], &fp->fp[i + 1],
8133 left * sizeof(struct file *));
8140 __skb_queue_tail(&list, skb);
8150 __skb_queue_tail(&list, skb);
8152 skb = skb_dequeue(head);
8155 if (skb_peek(&list)) {
8156 spin_lock_irq(&head->lock);
8157 while ((skb = __skb_dequeue(&list)) != NULL)
8158 __skb_queue_tail(head, skb);
8159 spin_unlock_irq(&head->lock);
8166 static void __io_rsrc_put_work(struct io_rsrc_node *ref_node)
8168 struct io_rsrc_data *rsrc_data = ref_node->rsrc_data;
8169 struct io_ring_ctx *ctx = rsrc_data->ctx;
8170 struct io_rsrc_put *prsrc, *tmp;
8172 list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
8173 list_del(&prsrc->list);
8176 bool lock_ring = ctx->flags & IORING_SETUP_IOPOLL;
8178 io_ring_submit_lock(ctx, lock_ring);
8179 spin_lock(&ctx->completion_lock);
8180 io_cqring_fill_event(ctx, prsrc->tag, 0, 0);
8182 io_commit_cqring(ctx);
8183 spin_unlock(&ctx->completion_lock);
8184 io_cqring_ev_posted(ctx);
8185 io_ring_submit_unlock(ctx, lock_ring);
8188 rsrc_data->do_put(ctx, prsrc);
8192 io_rsrc_node_destroy(ref_node);
8193 if (atomic_dec_and_test(&rsrc_data->refs))
8194 complete(&rsrc_data->done);
8197 static void io_rsrc_put_work(struct work_struct *work)
8199 struct io_ring_ctx *ctx;
8200 struct llist_node *node;
8202 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
8203 node = llist_del_all(&ctx->rsrc_put_llist);
8206 struct io_rsrc_node *ref_node;
8207 struct llist_node *next = node->next;
8209 ref_node = llist_entry(node, struct io_rsrc_node, llist);
8210 __io_rsrc_put_work(ref_node);
8215 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
8216 unsigned nr_args, u64 __user *tags)
8218 __s32 __user *fds = (__s32 __user *) arg;
8227 if (nr_args > IORING_MAX_FIXED_FILES)
8229 if (nr_args > rlimit(RLIMIT_NOFILE))
8231 ret = io_rsrc_node_switch_start(ctx);
8234 ret = io_rsrc_data_alloc(ctx, io_rsrc_file_put, tags, nr_args,
8240 if (!io_alloc_file_tables(&ctx->file_table, nr_args))
8243 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
8244 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
8248 /* allow sparse sets */
8251 if (unlikely(*io_get_tag_slot(ctx->file_data, i)))
8258 if (unlikely(!file))
8262 * Don't allow io_uring instances to be registered. If UNIX
8263 * isn't enabled, then this causes a reference cycle and this
8264 * instance can never get freed. If UNIX is enabled we'll
8265 * handle it just fine, but there's still no point in allowing
8266 * a ring fd as it doesn't support regular read/write anyway.
8268 if (file->f_op == &io_uring_fops) {
8272 io_fixed_file_set(io_fixed_file_slot(&ctx->file_table, i), file);
8275 ret = io_sqe_files_scm(ctx);
8277 __io_sqe_files_unregister(ctx);
8281 io_rsrc_node_switch(ctx, NULL);
8284 for (i = 0; i < ctx->nr_user_files; i++) {
8285 file = io_file_from_index(ctx, i);
8289 io_free_file_tables(&ctx->file_table);
8290 ctx->nr_user_files = 0;
8292 io_rsrc_data_free(ctx->file_data);
8293 ctx->file_data = NULL;
8297 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
8300 #if defined(CONFIG_UNIX)
8301 struct sock *sock = ctx->ring_sock->sk;
8302 struct sk_buff_head *head = &sock->sk_receive_queue;
8303 struct sk_buff *skb;
8306 * See if we can merge this file into an existing skb SCM_RIGHTS
8307 * file set. If there's no room, fall back to allocating a new skb
8308 * and filling it in.
8310 spin_lock_irq(&head->lock);
8311 skb = skb_peek(head);
8313 struct scm_fp_list *fpl = UNIXCB(skb).fp;
8315 if (fpl->count < SCM_MAX_FD) {
8316 __skb_unlink(skb, head);
8317 spin_unlock_irq(&head->lock);
8318 fpl->fp[fpl->count] = get_file(file);
8319 unix_inflight(fpl->user, fpl->fp[fpl->count]);
8321 spin_lock_irq(&head->lock);
8322 __skb_queue_head(head, skb);
8327 spin_unlock_irq(&head->lock);
8334 return __io_sqe_files_scm(ctx, 1, index);
8340 static int io_queue_rsrc_removal(struct io_rsrc_data *data, unsigned idx,
8341 struct io_rsrc_node *node, void *rsrc)
8343 struct io_rsrc_put *prsrc;
8345 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
8349 prsrc->tag = *io_get_tag_slot(data, idx);
8351 list_add(&prsrc->list, &node->rsrc_list);
8355 static int io_install_fixed_file(struct io_kiocb *req, struct file *file,
8356 unsigned int issue_flags, u32 slot_index)
8358 struct io_ring_ctx *ctx = req->ctx;
8359 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
8360 bool needs_switch = false;
8361 struct io_fixed_file *file_slot;
8364 io_ring_submit_lock(ctx, !force_nonblock);
8365 if (file->f_op == &io_uring_fops)
8368 if (!ctx->file_data)
8371 if (slot_index >= ctx->nr_user_files)
8374 slot_index = array_index_nospec(slot_index, ctx->nr_user_files);
8375 file_slot = io_fixed_file_slot(&ctx->file_table, slot_index);
8377 if (file_slot->file_ptr) {
8378 struct file *old_file;
8380 ret = io_rsrc_node_switch_start(ctx);
8384 old_file = (struct file *)(file_slot->file_ptr & FFS_MASK);
8385 ret = io_queue_rsrc_removal(ctx->file_data, slot_index,
8386 ctx->rsrc_node, old_file);
8389 file_slot->file_ptr = 0;
8390 needs_switch = true;
8393 *io_get_tag_slot(ctx->file_data, slot_index) = 0;
8394 io_fixed_file_set(file_slot, file);
8395 ret = io_sqe_file_register(ctx, file, slot_index);
8397 file_slot->file_ptr = 0;
8404 io_rsrc_node_switch(ctx, ctx->file_data);
8405 io_ring_submit_unlock(ctx, !force_nonblock);
8411 static int io_close_fixed(struct io_kiocb *req, unsigned int issue_flags)
8413 unsigned int offset = req->close.file_slot - 1;
8414 struct io_ring_ctx *ctx = req->ctx;
8415 struct io_fixed_file *file_slot;
8419 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
8421 if (unlikely(!ctx->file_data))
8424 if (offset >= ctx->nr_user_files)
8426 ret = io_rsrc_node_switch_start(ctx);
8430 i = array_index_nospec(offset, ctx->nr_user_files);
8431 file_slot = io_fixed_file_slot(&ctx->file_table, i);
8433 if (!file_slot->file_ptr)
8436 file = (struct file *)(file_slot->file_ptr & FFS_MASK);
8437 ret = io_queue_rsrc_removal(ctx->file_data, offset, ctx->rsrc_node, file);
8441 file_slot->file_ptr = 0;
8442 io_rsrc_node_switch(ctx, ctx->file_data);
8445 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
8449 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
8450 struct io_uring_rsrc_update2 *up,
8453 u64 __user *tags = u64_to_user_ptr(up->tags);
8454 __s32 __user *fds = u64_to_user_ptr(up->data);
8455 struct io_rsrc_data *data = ctx->file_data;
8456 struct io_fixed_file *file_slot;
8460 bool needs_switch = false;
8462 if (!ctx->file_data)
8464 if (up->offset + nr_args > ctx->nr_user_files)
8467 for (done = 0; done < nr_args; done++) {
8470 if ((tags && copy_from_user(&tag, &tags[done], sizeof(tag))) ||
8471 copy_from_user(&fd, &fds[done], sizeof(fd))) {
8475 if ((fd == IORING_REGISTER_FILES_SKIP || fd == -1) && tag) {
8479 if (fd == IORING_REGISTER_FILES_SKIP)
8482 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
8483 file_slot = io_fixed_file_slot(&ctx->file_table, i);
8485 if (file_slot->file_ptr) {
8486 file = (struct file *)(file_slot->file_ptr & FFS_MASK);
8487 err = io_queue_rsrc_removal(data, up->offset + done,
8488 ctx->rsrc_node, file);
8491 file_slot->file_ptr = 0;
8492 needs_switch = true;
8501 * Don't allow io_uring instances to be registered. If
8502 * UNIX isn't enabled, then this causes a reference
8503 * cycle and this instance can never get freed. If UNIX
8504 * is enabled we'll handle it just fine, but there's
8505 * still no point in allowing a ring fd as it doesn't
8506 * support regular read/write anyway.
8508 if (file->f_op == &io_uring_fops) {
8513 *io_get_tag_slot(data, up->offset + done) = tag;
8514 io_fixed_file_set(file_slot, file);
8515 err = io_sqe_file_register(ctx, file, i);
8517 file_slot->file_ptr = 0;
8525 io_rsrc_node_switch(ctx, data);
8526 return done ? done : err;
8529 static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx,
8530 struct task_struct *task)
8532 struct io_wq_hash *hash;
8533 struct io_wq_data data;
8534 unsigned int concurrency;
8536 mutex_lock(&ctx->uring_lock);
8537 hash = ctx->hash_map;
8539 hash = kzalloc(sizeof(*hash), GFP_KERNEL);
8541 mutex_unlock(&ctx->uring_lock);
8542 return ERR_PTR(-ENOMEM);
8544 refcount_set(&hash->refs, 1);
8545 init_waitqueue_head(&hash->wait);
8546 ctx->hash_map = hash;
8548 mutex_unlock(&ctx->uring_lock);
8552 data.free_work = io_wq_free_work;
8553 data.do_work = io_wq_submit_work;
8555 /* Do QD, or 4 * CPUS, whatever is smallest */
8556 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
8558 return io_wq_create(concurrency, &data);
8561 static int io_uring_alloc_task_context(struct task_struct *task,
8562 struct io_ring_ctx *ctx)
8564 struct io_uring_task *tctx;
8567 tctx = kzalloc(sizeof(*tctx), GFP_KERNEL);
8568 if (unlikely(!tctx))
8571 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
8572 if (unlikely(ret)) {
8577 tctx->io_wq = io_init_wq_offload(ctx, task);
8578 if (IS_ERR(tctx->io_wq)) {
8579 ret = PTR_ERR(tctx->io_wq);
8580 percpu_counter_destroy(&tctx->inflight);
8586 init_waitqueue_head(&tctx->wait);
8587 atomic_set(&tctx->in_idle, 0);
8588 atomic_set(&tctx->inflight_tracked, 0);
8589 task->io_uring = tctx;
8590 spin_lock_init(&tctx->task_lock);
8591 INIT_WQ_LIST(&tctx->task_list);
8592 init_task_work(&tctx->task_work, tctx_task_work);
8596 void __io_uring_free(struct task_struct *tsk)
8598 struct io_uring_task *tctx = tsk->io_uring;
8600 WARN_ON_ONCE(!xa_empty(&tctx->xa));
8601 WARN_ON_ONCE(tctx->io_wq);
8602 WARN_ON_ONCE(tctx->cached_refs);
8604 percpu_counter_destroy(&tctx->inflight);
8606 tsk->io_uring = NULL;
8609 static int io_sq_offload_create(struct io_ring_ctx *ctx,
8610 struct io_uring_params *p)
8614 /* Retain compatibility with failing for an invalid attach attempt */
8615 if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
8616 IORING_SETUP_ATTACH_WQ) {
8619 f = fdget(p->wq_fd);
8622 if (f.file->f_op != &io_uring_fops) {
8628 if (ctx->flags & IORING_SETUP_SQPOLL) {
8629 struct task_struct *tsk;
8630 struct io_sq_data *sqd;
8633 sqd = io_get_sq_data(p, &attached);
8639 ctx->sq_creds = get_current_cred();
8641 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
8642 if (!ctx->sq_thread_idle)
8643 ctx->sq_thread_idle = HZ;
8645 io_sq_thread_park(sqd);
8646 list_add(&ctx->sqd_list, &sqd->ctx_list);
8647 io_sqd_update_thread_idle(sqd);
8648 /* don't attach to a dying SQPOLL thread, would be racy */
8649 ret = (attached && !sqd->thread) ? -ENXIO : 0;
8650 io_sq_thread_unpark(sqd);
8657 if (p->flags & IORING_SETUP_SQ_AFF) {
8658 int cpu = p->sq_thread_cpu;
8661 if (cpu >= nr_cpu_ids || !cpu_online(cpu))
8668 sqd->task_pid = current->pid;
8669 sqd->task_tgid = current->tgid;
8670 tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE);
8677 ret = io_uring_alloc_task_context(tsk, ctx);
8678 wake_up_new_task(tsk);
8681 } else if (p->flags & IORING_SETUP_SQ_AFF) {
8682 /* Can't have SQ_AFF without SQPOLL */
8689 complete(&ctx->sq_data->exited);
8691 io_sq_thread_finish(ctx);
8695 static inline void __io_unaccount_mem(struct user_struct *user,
8696 unsigned long nr_pages)
8698 atomic_long_sub(nr_pages, &user->locked_vm);
8701 static inline int __io_account_mem(struct user_struct *user,
8702 unsigned long nr_pages)
8704 unsigned long page_limit, cur_pages, new_pages;
8706 /* Don't allow more pages than we can safely lock */
8707 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
8710 cur_pages = atomic_long_read(&user->locked_vm);
8711 new_pages = cur_pages + nr_pages;
8712 if (new_pages > page_limit)
8714 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
8715 new_pages) != cur_pages);
8720 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8723 __io_unaccount_mem(ctx->user, nr_pages);
8725 if (ctx->mm_account)
8726 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
8729 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8734 ret = __io_account_mem(ctx->user, nr_pages);
8739 if (ctx->mm_account)
8740 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
8745 static void io_mem_free(void *ptr)
8752 page = virt_to_head_page(ptr);
8753 if (put_page_testzero(page))
8754 free_compound_page(page);
8757 static void *io_mem_alloc(size_t size)
8759 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
8760 __GFP_NORETRY | __GFP_ACCOUNT;
8762 return (void *) __get_free_pages(gfp_flags, get_order(size));
8765 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8768 struct io_rings *rings;
8769 size_t off, sq_array_size;
8771 off = struct_size(rings, cqes, cq_entries);
8772 if (off == SIZE_MAX)
8776 off = ALIGN(off, SMP_CACHE_BYTES);
8784 sq_array_size = array_size(sizeof(u32), sq_entries);
8785 if (sq_array_size == SIZE_MAX)
8788 if (check_add_overflow(off, sq_array_size, &off))
8794 static void io_buffer_unmap(struct io_ring_ctx *ctx, struct io_mapped_ubuf **slot)
8796 struct io_mapped_ubuf *imu = *slot;
8799 if (imu != ctx->dummy_ubuf) {
8800 for (i = 0; i < imu->nr_bvecs; i++)
8801 unpin_user_page(imu->bvec[i].bv_page);
8802 if (imu->acct_pages)
8803 io_unaccount_mem(ctx, imu->acct_pages);
8809 static void io_rsrc_buf_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
8811 io_buffer_unmap(ctx, &prsrc->buf);
8815 static void __io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8819 for (i = 0; i < ctx->nr_user_bufs; i++)
8820 io_buffer_unmap(ctx, &ctx->user_bufs[i]);
8821 kfree(ctx->user_bufs);
8822 io_rsrc_data_free(ctx->buf_data);
8823 ctx->user_bufs = NULL;
8824 ctx->buf_data = NULL;
8825 ctx->nr_user_bufs = 0;
8828 static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8835 ret = io_rsrc_ref_quiesce(ctx->buf_data, ctx);
8837 __io_sqe_buffers_unregister(ctx);
8841 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8842 void __user *arg, unsigned index)
8844 struct iovec __user *src;
8846 #ifdef CONFIG_COMPAT
8848 struct compat_iovec __user *ciovs;
8849 struct compat_iovec ciov;
8851 ciovs = (struct compat_iovec __user *) arg;
8852 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8855 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8856 dst->iov_len = ciov.iov_len;
8860 src = (struct iovec __user *) arg;
8861 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8867 * Not super efficient, but this is just a registration time. And we do cache
8868 * the last compound head, so generally we'll only do a full search if we don't
8871 * We check if the given compound head page has already been accounted, to
8872 * avoid double accounting it. This allows us to account the full size of the
8873 * page, not just the constituent pages of a huge page.
8875 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8876 int nr_pages, struct page *hpage)
8880 /* check current page array */
8881 for (i = 0; i < nr_pages; i++) {
8882 if (!PageCompound(pages[i]))
8884 if (compound_head(pages[i]) == hpage)
8888 /* check previously registered pages */
8889 for (i = 0; i < ctx->nr_user_bufs; i++) {
8890 struct io_mapped_ubuf *imu = ctx->user_bufs[i];
8892 for (j = 0; j < imu->nr_bvecs; j++) {
8893 if (!PageCompound(imu->bvec[j].bv_page))
8895 if (compound_head(imu->bvec[j].bv_page) == hpage)
8903 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8904 int nr_pages, struct io_mapped_ubuf *imu,
8905 struct page **last_hpage)
8909 imu->acct_pages = 0;
8910 for (i = 0; i < nr_pages; i++) {
8911 if (!PageCompound(pages[i])) {
8916 hpage = compound_head(pages[i]);
8917 if (hpage == *last_hpage)
8919 *last_hpage = hpage;
8920 if (headpage_already_acct(ctx, pages, i, hpage))
8922 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8926 if (!imu->acct_pages)
8929 ret = io_account_mem(ctx, imu->acct_pages);
8931 imu->acct_pages = 0;
8935 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
8936 struct io_mapped_ubuf **pimu,
8937 struct page **last_hpage)
8939 struct io_mapped_ubuf *imu = NULL;
8940 struct vm_area_struct **vmas = NULL;
8941 struct page **pages = NULL;
8942 unsigned long off, start, end, ubuf;
8944 int ret, pret, nr_pages, i;
8946 if (!iov->iov_base) {
8947 *pimu = ctx->dummy_ubuf;
8951 ubuf = (unsigned long) iov->iov_base;
8952 end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8953 start = ubuf >> PAGE_SHIFT;
8954 nr_pages = end - start;
8959 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
8963 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
8968 imu = kvmalloc(struct_size(imu, bvec, nr_pages), GFP_KERNEL);
8973 mmap_read_lock(current->mm);
8974 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
8976 if (pret == nr_pages) {
8977 /* don't support file backed memory */
8978 for (i = 0; i < nr_pages; i++) {
8979 struct vm_area_struct *vma = vmas[i];
8981 if (vma_is_shmem(vma))
8984 !is_file_hugepages(vma->vm_file)) {
8990 ret = pret < 0 ? pret : -EFAULT;
8992 mmap_read_unlock(current->mm);
8995 * if we did partial map, or found file backed vmas,
8996 * release any pages we did get
8999 unpin_user_pages(pages, pret);
9003 ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
9005 unpin_user_pages(pages, pret);
9009 off = ubuf & ~PAGE_MASK;
9010 size = iov->iov_len;
9011 for (i = 0; i < nr_pages; i++) {
9014 vec_len = min_t(size_t, size, PAGE_SIZE - off);
9015 imu->bvec[i].bv_page = pages[i];
9016 imu->bvec[i].bv_len = vec_len;
9017 imu->bvec[i].bv_offset = off;
9021 /* store original address for later verification */
9023 imu->ubuf_end = ubuf + iov->iov_len;
9024 imu->nr_bvecs = nr_pages;
9035 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
9037 ctx->user_bufs = kcalloc(nr_args, sizeof(*ctx->user_bufs), GFP_KERNEL);
9038 return ctx->user_bufs ? 0 : -ENOMEM;
9041 static int io_buffer_validate(struct iovec *iov)
9043 unsigned long tmp, acct_len = iov->iov_len + (PAGE_SIZE - 1);
9046 * Don't impose further limits on the size and buffer
9047 * constraints here, we'll -EINVAL later when IO is
9048 * submitted if they are wrong.
9051 return iov->iov_len ? -EFAULT : 0;
9055 /* arbitrary limit, but we need something */
9056 if (iov->iov_len > SZ_1G)
9059 if (check_add_overflow((unsigned long)iov->iov_base, acct_len, &tmp))
9065 static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
9066 unsigned int nr_args, u64 __user *tags)
9068 struct page *last_hpage = NULL;
9069 struct io_rsrc_data *data;
9075 if (!nr_args || nr_args > IORING_MAX_REG_BUFFERS)
9077 ret = io_rsrc_node_switch_start(ctx);
9080 ret = io_rsrc_data_alloc(ctx, io_rsrc_buf_put, tags, nr_args, &data);
9083 ret = io_buffers_map_alloc(ctx, nr_args);
9085 io_rsrc_data_free(data);
9089 for (i = 0; i < nr_args; i++, ctx->nr_user_bufs++) {
9090 ret = io_copy_iov(ctx, &iov, arg, i);
9093 ret = io_buffer_validate(&iov);
9096 if (!iov.iov_base && *io_get_tag_slot(data, i)) {
9101 ret = io_sqe_buffer_register(ctx, &iov, &ctx->user_bufs[i],
9107 WARN_ON_ONCE(ctx->buf_data);
9109 ctx->buf_data = data;
9111 __io_sqe_buffers_unregister(ctx);
9113 io_rsrc_node_switch(ctx, NULL);
9117 static int __io_sqe_buffers_update(struct io_ring_ctx *ctx,
9118 struct io_uring_rsrc_update2 *up,
9119 unsigned int nr_args)
9121 u64 __user *tags = u64_to_user_ptr(up->tags);
9122 struct iovec iov, __user *iovs = u64_to_user_ptr(up->data);
9123 struct page *last_hpage = NULL;
9124 bool needs_switch = false;
9130 if (up->offset + nr_args > ctx->nr_user_bufs)
9133 for (done = 0; done < nr_args; done++) {
9134 struct io_mapped_ubuf *imu;
9135 int offset = up->offset + done;
9138 err = io_copy_iov(ctx, &iov, iovs, done);
9141 if (tags && copy_from_user(&tag, &tags[done], sizeof(tag))) {
9145 err = io_buffer_validate(&iov);
9148 if (!iov.iov_base && tag) {
9152 err = io_sqe_buffer_register(ctx, &iov, &imu, &last_hpage);
9156 i = array_index_nospec(offset, ctx->nr_user_bufs);
9157 if (ctx->user_bufs[i] != ctx->dummy_ubuf) {
9158 err = io_queue_rsrc_removal(ctx->buf_data, offset,
9159 ctx->rsrc_node, ctx->user_bufs[i]);
9160 if (unlikely(err)) {
9161 io_buffer_unmap(ctx, &imu);
9164 ctx->user_bufs[i] = NULL;
9165 needs_switch = true;
9168 ctx->user_bufs[i] = imu;
9169 *io_get_tag_slot(ctx->buf_data, offset) = tag;
9173 io_rsrc_node_switch(ctx, ctx->buf_data);
9174 return done ? done : err;
9177 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
9179 __s32 __user *fds = arg;
9185 if (copy_from_user(&fd, fds, sizeof(*fds)))
9188 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
9189 if (IS_ERR(ctx->cq_ev_fd)) {
9190 int ret = PTR_ERR(ctx->cq_ev_fd);
9192 ctx->cq_ev_fd = NULL;
9199 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
9201 if (ctx->cq_ev_fd) {
9202 eventfd_ctx_put(ctx->cq_ev_fd);
9203 ctx->cq_ev_fd = NULL;
9210 static void io_destroy_buffers(struct io_ring_ctx *ctx)
9212 struct io_buffer *buf;
9213 unsigned long index;
9215 xa_for_each(&ctx->io_buffers, index, buf) {
9216 __io_remove_buffers(ctx, buf, index, -1U);
9221 static void io_req_cache_free(struct list_head *list)
9223 struct io_kiocb *req, *nxt;
9225 list_for_each_entry_safe(req, nxt, list, inflight_entry) {
9226 list_del(&req->inflight_entry);
9227 kmem_cache_free(req_cachep, req);
9231 static void io_req_caches_free(struct io_ring_ctx *ctx)
9233 struct io_submit_state *state = &ctx->submit_state;
9235 mutex_lock(&ctx->uring_lock);
9237 if (state->free_reqs) {
9238 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
9239 state->free_reqs = 0;
9242 io_flush_cached_locked_reqs(ctx, state);
9243 io_req_cache_free(&state->free_list);
9244 mutex_unlock(&ctx->uring_lock);
9247 static void io_wait_rsrc_data(struct io_rsrc_data *data)
9249 if (data && !atomic_dec_and_test(&data->refs))
9250 wait_for_completion(&data->done);
9253 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
9255 io_sq_thread_finish(ctx);
9257 if (ctx->mm_account) {
9258 mmdrop(ctx->mm_account);
9259 ctx->mm_account = NULL;
9262 /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */
9263 io_wait_rsrc_data(ctx->buf_data);
9264 io_wait_rsrc_data(ctx->file_data);
9266 mutex_lock(&ctx->uring_lock);
9268 __io_sqe_buffers_unregister(ctx);
9270 __io_sqe_files_unregister(ctx);
9272 __io_cqring_overflow_flush(ctx, true);
9273 mutex_unlock(&ctx->uring_lock);
9274 io_eventfd_unregister(ctx);
9275 io_destroy_buffers(ctx);
9277 put_cred(ctx->sq_creds);
9279 /* there are no registered resources left, nobody uses it */
9281 io_rsrc_node_destroy(ctx->rsrc_node);
9282 if (ctx->rsrc_backup_node)
9283 io_rsrc_node_destroy(ctx->rsrc_backup_node);
9284 flush_delayed_work(&ctx->rsrc_put_work);
9286 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
9287 WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist));
9289 #if defined(CONFIG_UNIX)
9290 if (ctx->ring_sock) {
9291 ctx->ring_sock->file = NULL; /* so that iput() is called */
9292 sock_release(ctx->ring_sock);
9295 WARN_ON_ONCE(!list_empty(&ctx->ltimeout_list));
9297 io_mem_free(ctx->rings);
9298 io_mem_free(ctx->sq_sqes);
9300 percpu_ref_exit(&ctx->refs);
9301 free_uid(ctx->user);
9302 io_req_caches_free(ctx);
9304 io_wq_put_hash(ctx->hash_map);
9305 kfree(ctx->cancel_hash);
9306 kfree(ctx->dummy_ubuf);
9310 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
9312 struct io_ring_ctx *ctx = file->private_data;
9315 poll_wait(file, &ctx->poll_wait, wait);
9317 * synchronizes with barrier from wq_has_sleeper call in
9321 if (!io_sqring_full(ctx))
9322 mask |= EPOLLOUT | EPOLLWRNORM;
9325 * Don't flush cqring overflow list here, just do a simple check.
9326 * Otherwise there could possible be ABBA deadlock:
9329 * lock(&ctx->uring_lock);
9331 * lock(&ctx->uring_lock);
9334 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
9335 * pushs them to do the flush.
9337 if (io_cqring_events(ctx) || test_bit(0, &ctx->check_cq_overflow))
9338 mask |= EPOLLIN | EPOLLRDNORM;
9343 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
9345 const struct cred *creds;
9347 creds = xa_erase(&ctx->personalities, id);
9356 struct io_tctx_exit {
9357 struct callback_head task_work;
9358 struct completion completion;
9359 struct io_ring_ctx *ctx;
9362 static void io_tctx_exit_cb(struct callback_head *cb)
9364 struct io_uring_task *tctx = current->io_uring;
9365 struct io_tctx_exit *work;
9367 work = container_of(cb, struct io_tctx_exit, task_work);
9369 * When @in_idle, we're in cancellation and it's racy to remove the
9370 * node. It'll be removed by the end of cancellation, just ignore it.
9372 if (!atomic_read(&tctx->in_idle))
9373 io_uring_del_tctx_node((unsigned long)work->ctx);
9374 complete(&work->completion);
9377 static bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
9379 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
9381 return req->ctx == data;
9384 static void io_ring_exit_work(struct work_struct *work)
9386 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
9387 unsigned long timeout = jiffies + HZ * 60 * 5;
9388 unsigned long interval = HZ / 20;
9389 struct io_tctx_exit exit;
9390 struct io_tctx_node *node;
9394 * If we're doing polled IO and end up having requests being
9395 * submitted async (out-of-line), then completions can come in while
9396 * we're waiting for refs to drop. We need to reap these manually,
9397 * as nobody else will be looking for them.
9400 io_uring_try_cancel_requests(ctx, NULL, true);
9402 struct io_sq_data *sqd = ctx->sq_data;
9403 struct task_struct *tsk;
9405 io_sq_thread_park(sqd);
9407 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
9408 io_wq_cancel_cb(tsk->io_uring->io_wq,
9409 io_cancel_ctx_cb, ctx, true);
9410 io_sq_thread_unpark(sqd);
9413 if (WARN_ON_ONCE(time_after(jiffies, timeout))) {
9414 /* there is little hope left, don't run it too often */
9417 } while (!wait_for_completion_timeout(&ctx->ref_comp, interval));
9419 init_completion(&exit.completion);
9420 init_task_work(&exit.task_work, io_tctx_exit_cb);
9423 * Some may use context even when all refs and requests have been put,
9424 * and they are free to do so while still holding uring_lock or
9425 * completion_lock, see io_req_task_submit(). Apart from other work,
9426 * this lock/unlock section also waits them to finish.
9428 mutex_lock(&ctx->uring_lock);
9429 while (!list_empty(&ctx->tctx_list)) {
9430 WARN_ON_ONCE(time_after(jiffies, timeout));
9432 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
9434 /* don't spin on a single task if cancellation failed */
9435 list_rotate_left(&ctx->tctx_list);
9436 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
9437 if (WARN_ON_ONCE(ret))
9439 wake_up_process(node->task);
9441 mutex_unlock(&ctx->uring_lock);
9442 wait_for_completion(&exit.completion);
9443 mutex_lock(&ctx->uring_lock);
9445 mutex_unlock(&ctx->uring_lock);
9446 spin_lock(&ctx->completion_lock);
9447 spin_unlock(&ctx->completion_lock);
9449 io_ring_ctx_free(ctx);
9452 /* Returns true if we found and killed one or more timeouts */
9453 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
9456 struct io_kiocb *req, *tmp;
9459 spin_lock(&ctx->completion_lock);
9460 spin_lock_irq(&ctx->timeout_lock);
9461 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
9462 if (io_match_task(req, tsk, cancel_all)) {
9463 io_kill_timeout(req, -ECANCELED);
9467 spin_unlock_irq(&ctx->timeout_lock);
9469 io_commit_cqring(ctx);
9470 spin_unlock(&ctx->completion_lock);
9472 io_cqring_ev_posted(ctx);
9473 return canceled != 0;
9476 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
9478 unsigned long index;
9479 struct creds *creds;
9481 mutex_lock(&ctx->uring_lock);
9482 percpu_ref_kill(&ctx->refs);
9484 __io_cqring_overflow_flush(ctx, true);
9485 xa_for_each(&ctx->personalities, index, creds)
9486 io_unregister_personality(ctx, index);
9487 mutex_unlock(&ctx->uring_lock);
9489 io_kill_timeouts(ctx, NULL, true);
9490 io_poll_remove_all(ctx, NULL, true);
9492 /* if we failed setting up the ctx, we might not have any rings */
9493 io_iopoll_try_reap_events(ctx);
9495 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
9497 * Use system_unbound_wq to avoid spawning tons of event kworkers
9498 * if we're exiting a ton of rings at the same time. It just adds
9499 * noise and overhead, there's no discernable change in runtime
9500 * over using system_wq.
9502 queue_work(system_unbound_wq, &ctx->exit_work);
9505 static int io_uring_release(struct inode *inode, struct file *file)
9507 struct io_ring_ctx *ctx = file->private_data;
9509 file->private_data = NULL;
9510 io_ring_ctx_wait_and_kill(ctx);
9514 struct io_task_cancel {
9515 struct task_struct *task;
9519 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
9521 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
9522 struct io_task_cancel *cancel = data;
9525 if (!cancel->all && (req->flags & REQ_F_LINK_TIMEOUT)) {
9526 struct io_ring_ctx *ctx = req->ctx;
9528 /* protect against races with linked timeouts */
9529 spin_lock(&ctx->completion_lock);
9530 ret = io_match_task(req, cancel->task, cancel->all);
9531 spin_unlock(&ctx->completion_lock);
9533 ret = io_match_task(req, cancel->task, cancel->all);
9538 static bool io_cancel_defer_files(struct io_ring_ctx *ctx,
9539 struct task_struct *task, bool cancel_all)
9541 struct io_defer_entry *de;
9544 spin_lock(&ctx->completion_lock);
9545 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
9546 if (io_match_task(de->req, task, cancel_all)) {
9547 list_cut_position(&list, &ctx->defer_list, &de->list);
9551 spin_unlock(&ctx->completion_lock);
9552 if (list_empty(&list))
9555 while (!list_empty(&list)) {
9556 de = list_first_entry(&list, struct io_defer_entry, list);
9557 list_del_init(&de->list);
9558 io_req_complete_failed(de->req, -ECANCELED);
9564 static bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
9566 struct io_tctx_node *node;
9567 enum io_wq_cancel cret;
9570 mutex_lock(&ctx->uring_lock);
9571 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
9572 struct io_uring_task *tctx = node->task->io_uring;
9575 * io_wq will stay alive while we hold uring_lock, because it's
9576 * killed after ctx nodes, which requires to take the lock.
9578 if (!tctx || !tctx->io_wq)
9580 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
9581 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
9583 mutex_unlock(&ctx->uring_lock);
9588 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
9589 struct task_struct *task,
9592 struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
9593 struct io_uring_task *tctx = task ? task->io_uring : NULL;
9596 enum io_wq_cancel cret;
9600 ret |= io_uring_try_cancel_iowq(ctx);
9601 } else if (tctx && tctx->io_wq) {
9603 * Cancels requests of all rings, not only @ctx, but
9604 * it's fine as the task is in exit/exec.
9606 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
9608 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
9611 /* SQPOLL thread does its own polling */
9612 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
9613 (ctx->sq_data && ctx->sq_data->thread == current)) {
9614 while (!list_empty_careful(&ctx->iopoll_list)) {
9615 io_iopoll_try_reap_events(ctx);
9620 ret |= io_cancel_defer_files(ctx, task, cancel_all);
9621 ret |= io_poll_remove_all(ctx, task, cancel_all);
9622 ret |= io_kill_timeouts(ctx, task, cancel_all);
9624 ret |= io_run_task_work();
9631 static int __io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9633 struct io_uring_task *tctx = current->io_uring;
9634 struct io_tctx_node *node;
9637 if (unlikely(!tctx)) {
9638 ret = io_uring_alloc_task_context(current, ctx);
9641 tctx = current->io_uring;
9643 if (!xa_load(&tctx->xa, (unsigned long)ctx)) {
9644 node = kmalloc(sizeof(*node), GFP_KERNEL);
9648 node->task = current;
9650 ret = xa_err(xa_store(&tctx->xa, (unsigned long)ctx,
9657 mutex_lock(&ctx->uring_lock);
9658 list_add(&node->ctx_node, &ctx->tctx_list);
9659 mutex_unlock(&ctx->uring_lock);
9666 * Note that this task has used io_uring. We use it for cancelation purposes.
9668 static inline int io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9670 struct io_uring_task *tctx = current->io_uring;
9672 if (likely(tctx && tctx->last == ctx))
9674 return __io_uring_add_tctx_node(ctx);
9678 * Remove this io_uring_file -> task mapping.
9680 static void io_uring_del_tctx_node(unsigned long index)
9682 struct io_uring_task *tctx = current->io_uring;
9683 struct io_tctx_node *node;
9687 node = xa_erase(&tctx->xa, index);
9691 WARN_ON_ONCE(current != node->task);
9692 WARN_ON_ONCE(list_empty(&node->ctx_node));
9694 mutex_lock(&node->ctx->uring_lock);
9695 list_del(&node->ctx_node);
9696 mutex_unlock(&node->ctx->uring_lock);
9698 if (tctx->last == node->ctx)
9703 static void io_uring_clean_tctx(struct io_uring_task *tctx)
9705 struct io_wq *wq = tctx->io_wq;
9706 struct io_tctx_node *node;
9707 unsigned long index;
9709 xa_for_each(&tctx->xa, index, node) {
9710 io_uring_del_tctx_node(index);
9715 * Must be after io_uring_del_task_file() (removes nodes under
9716 * uring_lock) to avoid race with io_uring_try_cancel_iowq().
9718 io_wq_put_and_exit(wq);
9723 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
9726 return atomic_read(&tctx->inflight_tracked);
9727 return percpu_counter_sum(&tctx->inflight);
9730 static void io_uring_drop_tctx_refs(struct task_struct *task)
9732 struct io_uring_task *tctx = task->io_uring;
9733 unsigned int refs = tctx->cached_refs;
9736 tctx->cached_refs = 0;
9737 percpu_counter_sub(&tctx->inflight, refs);
9738 put_task_struct_many(task, refs);
9743 * Find any io_uring ctx that this task has registered or done IO on, and cancel
9744 * requests. @sqd should be not-null IIF it's an SQPOLL thread cancellation.
9746 static void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
9748 struct io_uring_task *tctx = current->io_uring;
9749 struct io_ring_ctx *ctx;
9753 WARN_ON_ONCE(sqd && sqd->thread != current);
9755 if (!current->io_uring)
9758 io_wq_exit_start(tctx->io_wq);
9760 atomic_inc(&tctx->in_idle);
9762 io_uring_drop_tctx_refs(current);
9763 /* read completions before cancelations */
9764 inflight = tctx_inflight(tctx, !cancel_all);
9769 struct io_tctx_node *node;
9770 unsigned long index;
9772 xa_for_each(&tctx->xa, index, node) {
9773 /* sqpoll task will cancel all its requests */
9774 if (node->ctx->sq_data)
9776 io_uring_try_cancel_requests(node->ctx, current,
9780 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
9781 io_uring_try_cancel_requests(ctx, current,
9785 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
9786 io_uring_drop_tctx_refs(current);
9788 * If we've seen completions, retry without waiting. This
9789 * avoids a race where a completion comes in before we did
9790 * prepare_to_wait().
9792 if (inflight == tctx_inflight(tctx, !cancel_all))
9794 finish_wait(&tctx->wait, &wait);
9796 atomic_dec(&tctx->in_idle);
9798 io_uring_clean_tctx(tctx);
9800 /* for exec all current's requests should be gone, kill tctx */
9801 __io_uring_free(current);
9805 void __io_uring_cancel(bool cancel_all)
9807 io_uring_cancel_generic(cancel_all, NULL);
9810 static void *io_uring_validate_mmap_request(struct file *file,
9811 loff_t pgoff, size_t sz)
9813 struct io_ring_ctx *ctx = file->private_data;
9814 loff_t offset = pgoff << PAGE_SHIFT;
9819 case IORING_OFF_SQ_RING:
9820 case IORING_OFF_CQ_RING:
9823 case IORING_OFF_SQES:
9827 return ERR_PTR(-EINVAL);
9830 page = virt_to_head_page(ptr);
9831 if (sz > page_size(page))
9832 return ERR_PTR(-EINVAL);
9839 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9841 size_t sz = vma->vm_end - vma->vm_start;
9845 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
9847 return PTR_ERR(ptr);
9849 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
9850 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
9853 #else /* !CONFIG_MMU */
9855 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9857 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
9860 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
9862 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
9865 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
9866 unsigned long addr, unsigned long len,
9867 unsigned long pgoff, unsigned long flags)
9871 ptr = io_uring_validate_mmap_request(file, pgoff, len);
9873 return PTR_ERR(ptr);
9875 return (unsigned long) ptr;
9878 #endif /* !CONFIG_MMU */
9880 static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
9885 if (!io_sqring_full(ctx))
9887 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9889 if (!io_sqring_full(ctx))
9892 } while (!signal_pending(current));
9894 finish_wait(&ctx->sqo_sq_wait, &wait);
9898 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9899 struct __kernel_timespec __user **ts,
9900 const sigset_t __user **sig)
9902 struct io_uring_getevents_arg arg;
9905 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9906 * is just a pointer to the sigset_t.
9908 if (!(flags & IORING_ENTER_EXT_ARG)) {
9909 *sig = (const sigset_t __user *) argp;
9915 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9916 * timespec and sigset_t pointers if good.
9918 if (*argsz != sizeof(arg))
9920 if (copy_from_user(&arg, argp, sizeof(arg)))
9922 *sig = u64_to_user_ptr(arg.sigmask);
9923 *argsz = arg.sigmask_sz;
9924 *ts = u64_to_user_ptr(arg.ts);
9928 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9929 u32, min_complete, u32, flags, const void __user *, argp,
9932 struct io_ring_ctx *ctx;
9939 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9940 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG)))
9944 if (unlikely(!f.file))
9948 if (unlikely(f.file->f_op != &io_uring_fops))
9952 ctx = f.file->private_data;
9953 if (unlikely(!percpu_ref_tryget(&ctx->refs)))
9957 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
9961 * For SQ polling, the thread will do all submissions and completions.
9962 * Just return the requested submit count, and wake the thread if
9966 if (ctx->flags & IORING_SETUP_SQPOLL) {
9967 io_cqring_overflow_flush(ctx);
9969 if (unlikely(ctx->sq_data->thread == NULL)) {
9973 if (flags & IORING_ENTER_SQ_WAKEUP)
9974 wake_up(&ctx->sq_data->wait);
9975 if (flags & IORING_ENTER_SQ_WAIT) {
9976 ret = io_sqpoll_wait_sq(ctx);
9980 submitted = to_submit;
9981 } else if (to_submit) {
9982 ret = io_uring_add_tctx_node(ctx);
9985 mutex_lock(&ctx->uring_lock);
9986 submitted = io_submit_sqes(ctx, to_submit);
9987 mutex_unlock(&ctx->uring_lock);
9989 if (submitted != to_submit)
9992 if (flags & IORING_ENTER_GETEVENTS) {
9993 const sigset_t __user *sig;
9994 struct __kernel_timespec __user *ts;
9996 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
10000 min_complete = min(min_complete, ctx->cq_entries);
10003 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
10004 * space applications don't need to do io completion events
10005 * polling again, they can rely on io_sq_thread to do polling
10006 * work, which can reduce cpu usage and uring_lock contention.
10008 if (ctx->flags & IORING_SETUP_IOPOLL &&
10009 !(ctx->flags & IORING_SETUP_SQPOLL)) {
10010 ret = io_iopoll_check(ctx, min_complete);
10012 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
10017 percpu_ref_put(&ctx->refs);
10020 return submitted ? submitted : ret;
10023 #ifdef CONFIG_PROC_FS
10024 static int io_uring_show_cred(struct seq_file *m, unsigned int id,
10025 const struct cred *cred)
10027 struct user_namespace *uns = seq_user_ns(m);
10028 struct group_info *gi;
10033 seq_printf(m, "%5d\n", id);
10034 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
10035 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
10036 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
10037 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
10038 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
10039 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
10040 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
10041 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
10042 seq_puts(m, "\n\tGroups:\t");
10043 gi = cred->group_info;
10044 for (g = 0; g < gi->ngroups; g++) {
10045 seq_put_decimal_ull(m, g ? " " : "",
10046 from_kgid_munged(uns, gi->gid[g]));
10048 seq_puts(m, "\n\tCapEff:\t");
10049 cap = cred->cap_effective;
10050 CAP_FOR_EACH_U32(__capi)
10051 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
10056 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
10058 struct io_sq_data *sq = NULL;
10063 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
10064 * since fdinfo case grabs it in the opposite direction of normal use
10065 * cases. If we fail to get the lock, we just don't iterate any
10066 * structures that could be going away outside the io_uring mutex.
10068 has_lock = mutex_trylock(&ctx->uring_lock);
10070 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) {
10076 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
10077 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
10078 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
10079 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
10080 struct file *f = io_file_from_index(ctx, i);
10083 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
10085 seq_printf(m, "%5u: <none>\n", i);
10087 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
10088 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
10089 struct io_mapped_ubuf *buf = ctx->user_bufs[i];
10090 unsigned int len = buf->ubuf_end - buf->ubuf;
10092 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf, len);
10094 if (has_lock && !xa_empty(&ctx->personalities)) {
10095 unsigned long index;
10096 const struct cred *cred;
10098 seq_printf(m, "Personalities:\n");
10099 xa_for_each(&ctx->personalities, index, cred)
10100 io_uring_show_cred(m, index, cred);
10102 seq_printf(m, "PollList:\n");
10103 spin_lock(&ctx->completion_lock);
10104 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
10105 struct hlist_head *list = &ctx->cancel_hash[i];
10106 struct io_kiocb *req;
10108 hlist_for_each_entry(req, list, hash_node)
10109 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
10110 req->task->task_works != NULL);
10112 spin_unlock(&ctx->completion_lock);
10114 mutex_unlock(&ctx->uring_lock);
10117 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
10119 struct io_ring_ctx *ctx = f->private_data;
10121 if (percpu_ref_tryget(&ctx->refs)) {
10122 __io_uring_show_fdinfo(ctx, m);
10123 percpu_ref_put(&ctx->refs);
10128 static const struct file_operations io_uring_fops = {
10129 .release = io_uring_release,
10130 .mmap = io_uring_mmap,
10132 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
10133 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
10135 .poll = io_uring_poll,
10136 #ifdef CONFIG_PROC_FS
10137 .show_fdinfo = io_uring_show_fdinfo,
10141 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
10142 struct io_uring_params *p)
10144 struct io_rings *rings;
10145 size_t size, sq_array_offset;
10147 /* make sure these are sane, as we already accounted them */
10148 ctx->sq_entries = p->sq_entries;
10149 ctx->cq_entries = p->cq_entries;
10151 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
10152 if (size == SIZE_MAX)
10155 rings = io_mem_alloc(size);
10159 ctx->rings = rings;
10160 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
10161 rings->sq_ring_mask = p->sq_entries - 1;
10162 rings->cq_ring_mask = p->cq_entries - 1;
10163 rings->sq_ring_entries = p->sq_entries;
10164 rings->cq_ring_entries = p->cq_entries;
10166 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
10167 if (size == SIZE_MAX) {
10168 io_mem_free(ctx->rings);
10173 ctx->sq_sqes = io_mem_alloc(size);
10174 if (!ctx->sq_sqes) {
10175 io_mem_free(ctx->rings);
10183 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
10187 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
10191 ret = io_uring_add_tctx_node(ctx);
10196 fd_install(fd, file);
10201 * Allocate an anonymous fd, this is what constitutes the application
10202 * visible backing of an io_uring instance. The application mmaps this
10203 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
10204 * we have to tie this fd to a socket for file garbage collection purposes.
10206 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
10209 #if defined(CONFIG_UNIX)
10212 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
10215 return ERR_PTR(ret);
10218 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
10219 O_RDWR | O_CLOEXEC);
10220 #if defined(CONFIG_UNIX)
10221 if (IS_ERR(file)) {
10222 sock_release(ctx->ring_sock);
10223 ctx->ring_sock = NULL;
10225 ctx->ring_sock->file = file;
10231 static int io_uring_create(unsigned entries, struct io_uring_params *p,
10232 struct io_uring_params __user *params)
10234 struct io_ring_ctx *ctx;
10240 if (entries > IORING_MAX_ENTRIES) {
10241 if (!(p->flags & IORING_SETUP_CLAMP))
10243 entries = IORING_MAX_ENTRIES;
10247 * Use twice as many entries for the CQ ring. It's possible for the
10248 * application to drive a higher depth than the size of the SQ ring,
10249 * since the sqes are only used at submission time. This allows for
10250 * some flexibility in overcommitting a bit. If the application has
10251 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
10252 * of CQ ring entries manually.
10254 p->sq_entries = roundup_pow_of_two(entries);
10255 if (p->flags & IORING_SETUP_CQSIZE) {
10257 * If IORING_SETUP_CQSIZE is set, we do the same roundup
10258 * to a power-of-two, if it isn't already. We do NOT impose
10259 * any cq vs sq ring sizing.
10261 if (!p->cq_entries)
10263 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
10264 if (!(p->flags & IORING_SETUP_CLAMP))
10266 p->cq_entries = IORING_MAX_CQ_ENTRIES;
10268 p->cq_entries = roundup_pow_of_two(p->cq_entries);
10269 if (p->cq_entries < p->sq_entries)
10272 p->cq_entries = 2 * p->sq_entries;
10275 ctx = io_ring_ctx_alloc(p);
10278 ctx->compat = in_compat_syscall();
10279 if (!capable(CAP_IPC_LOCK))
10280 ctx->user = get_uid(current_user());
10283 * This is just grabbed for accounting purposes. When a process exits,
10284 * the mm is exited and dropped before the files, hence we need to hang
10285 * on to this mm purely for the purposes of being able to unaccount
10286 * memory (locked/pinned vm). It's not used for anything else.
10288 mmgrab(current->mm);
10289 ctx->mm_account = current->mm;
10291 ret = io_allocate_scq_urings(ctx, p);
10295 ret = io_sq_offload_create(ctx, p);
10298 /* always set a rsrc node */
10299 ret = io_rsrc_node_switch_start(ctx);
10302 io_rsrc_node_switch(ctx, NULL);
10304 memset(&p->sq_off, 0, sizeof(p->sq_off));
10305 p->sq_off.head = offsetof(struct io_rings, sq.head);
10306 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
10307 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
10308 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
10309 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
10310 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
10311 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
10313 memset(&p->cq_off, 0, sizeof(p->cq_off));
10314 p->cq_off.head = offsetof(struct io_rings, cq.head);
10315 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
10316 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
10317 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
10318 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
10319 p->cq_off.cqes = offsetof(struct io_rings, cqes);
10320 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
10322 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
10323 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
10324 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
10325 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
10326 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
10327 IORING_FEAT_RSRC_TAGS;
10329 if (copy_to_user(params, p, sizeof(*p))) {
10334 file = io_uring_get_file(ctx);
10335 if (IS_ERR(file)) {
10336 ret = PTR_ERR(file);
10341 * Install ring fd as the very last thing, so we don't risk someone
10342 * having closed it before we finish setup
10344 ret = io_uring_install_fd(ctx, file);
10346 /* fput will clean it up */
10351 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
10354 io_ring_ctx_wait_and_kill(ctx);
10359 * Sets up an aio uring context, and returns the fd. Applications asks for a
10360 * ring size, we return the actual sq/cq ring sizes (among other things) in the
10361 * params structure passed in.
10363 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
10365 struct io_uring_params p;
10368 if (copy_from_user(&p, params, sizeof(p)))
10370 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
10375 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
10376 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
10377 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
10378 IORING_SETUP_R_DISABLED))
10381 return io_uring_create(entries, &p, params);
10384 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
10385 struct io_uring_params __user *, params)
10387 return io_uring_setup(entries, params);
10390 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
10392 struct io_uring_probe *p;
10396 size = struct_size(p, ops, nr_args);
10397 if (size == SIZE_MAX)
10399 p = kzalloc(size, GFP_KERNEL);
10404 if (copy_from_user(p, arg, size))
10407 if (memchr_inv(p, 0, size))
10410 p->last_op = IORING_OP_LAST - 1;
10411 if (nr_args > IORING_OP_LAST)
10412 nr_args = IORING_OP_LAST;
10414 for (i = 0; i < nr_args; i++) {
10416 if (!io_op_defs[i].not_supported)
10417 p->ops[i].flags = IO_URING_OP_SUPPORTED;
10422 if (copy_to_user(arg, p, size))
10429 static int io_register_personality(struct io_ring_ctx *ctx)
10431 const struct cred *creds;
10435 creds = get_current_cred();
10437 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
10438 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
10446 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
10447 unsigned int nr_args)
10449 struct io_uring_restriction *res;
10453 /* Restrictions allowed only if rings started disabled */
10454 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
10457 /* We allow only a single restrictions registration */
10458 if (ctx->restrictions.registered)
10461 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
10464 size = array_size(nr_args, sizeof(*res));
10465 if (size == SIZE_MAX)
10468 res = memdup_user(arg, size);
10470 return PTR_ERR(res);
10474 for (i = 0; i < nr_args; i++) {
10475 switch (res[i].opcode) {
10476 case IORING_RESTRICTION_REGISTER_OP:
10477 if (res[i].register_op >= IORING_REGISTER_LAST) {
10482 __set_bit(res[i].register_op,
10483 ctx->restrictions.register_op);
10485 case IORING_RESTRICTION_SQE_OP:
10486 if (res[i].sqe_op >= IORING_OP_LAST) {
10491 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
10493 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
10494 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
10496 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
10497 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
10506 /* Reset all restrictions if an error happened */
10508 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
10510 ctx->restrictions.registered = true;
10516 static int io_register_enable_rings(struct io_ring_ctx *ctx)
10518 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
10521 if (ctx->restrictions.registered)
10522 ctx->restricted = 1;
10524 ctx->flags &= ~IORING_SETUP_R_DISABLED;
10525 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
10526 wake_up(&ctx->sq_data->wait);
10530 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
10531 struct io_uring_rsrc_update2 *up,
10539 if (check_add_overflow(up->offset, nr_args, &tmp))
10541 err = io_rsrc_node_switch_start(ctx);
10546 case IORING_RSRC_FILE:
10547 return __io_sqe_files_update(ctx, up, nr_args);
10548 case IORING_RSRC_BUFFER:
10549 return __io_sqe_buffers_update(ctx, up, nr_args);
10554 static int io_register_files_update(struct io_ring_ctx *ctx, void __user *arg,
10557 struct io_uring_rsrc_update2 up;
10561 memset(&up, 0, sizeof(up));
10562 if (copy_from_user(&up, arg, sizeof(struct io_uring_rsrc_update)))
10564 return __io_register_rsrc_update(ctx, IORING_RSRC_FILE, &up, nr_args);
10567 static int io_register_rsrc_update(struct io_ring_ctx *ctx, void __user *arg,
10568 unsigned size, unsigned type)
10570 struct io_uring_rsrc_update2 up;
10572 if (size != sizeof(up))
10574 if (copy_from_user(&up, arg, sizeof(up)))
10576 if (!up.nr || up.resv)
10578 return __io_register_rsrc_update(ctx, type, &up, up.nr);
10581 static int io_register_rsrc(struct io_ring_ctx *ctx, void __user *arg,
10582 unsigned int size, unsigned int type)
10584 struct io_uring_rsrc_register rr;
10586 /* keep it extendible */
10587 if (size != sizeof(rr))
10590 memset(&rr, 0, sizeof(rr));
10591 if (copy_from_user(&rr, arg, size))
10593 if (!rr.nr || rr.resv || rr.resv2)
10597 case IORING_RSRC_FILE:
10598 return io_sqe_files_register(ctx, u64_to_user_ptr(rr.data),
10599 rr.nr, u64_to_user_ptr(rr.tags));
10600 case IORING_RSRC_BUFFER:
10601 return io_sqe_buffers_register(ctx, u64_to_user_ptr(rr.data),
10602 rr.nr, u64_to_user_ptr(rr.tags));
10607 static int io_register_iowq_aff(struct io_ring_ctx *ctx, void __user *arg,
10610 struct io_uring_task *tctx = current->io_uring;
10611 cpumask_var_t new_mask;
10614 if (!tctx || !tctx->io_wq)
10617 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
10620 cpumask_clear(new_mask);
10621 if (len > cpumask_size())
10622 len = cpumask_size();
10624 if (copy_from_user(new_mask, arg, len)) {
10625 free_cpumask_var(new_mask);
10629 ret = io_wq_cpu_affinity(tctx->io_wq, new_mask);
10630 free_cpumask_var(new_mask);
10634 static int io_unregister_iowq_aff(struct io_ring_ctx *ctx)
10636 struct io_uring_task *tctx = current->io_uring;
10638 if (!tctx || !tctx->io_wq)
10641 return io_wq_cpu_affinity(tctx->io_wq, NULL);
10644 static int io_register_iowq_max_workers(struct io_ring_ctx *ctx,
10647 struct io_uring_task *tctx = NULL;
10648 struct io_sq_data *sqd = NULL;
10649 __u32 new_count[2];
10652 if (copy_from_user(new_count, arg, sizeof(new_count)))
10654 for (i = 0; i < ARRAY_SIZE(new_count); i++)
10655 if (new_count[i] > INT_MAX)
10658 if (ctx->flags & IORING_SETUP_SQPOLL) {
10659 sqd = ctx->sq_data;
10662 * Observe the correct sqd->lock -> ctx->uring_lock
10663 * ordering. Fine to drop uring_lock here, we hold
10664 * a ref to the ctx.
10666 refcount_inc(&sqd->refs);
10667 mutex_unlock(&ctx->uring_lock);
10668 mutex_lock(&sqd->lock);
10669 mutex_lock(&ctx->uring_lock);
10671 tctx = sqd->thread->io_uring;
10674 tctx = current->io_uring;
10678 if (!tctx || !tctx->io_wq)
10681 ret = io_wq_max_workers(tctx->io_wq, new_count);
10686 mutex_unlock(&sqd->lock);
10687 io_put_sq_data(sqd);
10690 if (copy_to_user(arg, new_count, sizeof(new_count)))
10696 mutex_unlock(&sqd->lock);
10697 io_put_sq_data(sqd);
10702 static bool io_register_op_must_quiesce(int op)
10705 case IORING_REGISTER_BUFFERS:
10706 case IORING_UNREGISTER_BUFFERS:
10707 case IORING_REGISTER_FILES:
10708 case IORING_UNREGISTER_FILES:
10709 case IORING_REGISTER_FILES_UPDATE:
10710 case IORING_REGISTER_PROBE:
10711 case IORING_REGISTER_PERSONALITY:
10712 case IORING_UNREGISTER_PERSONALITY:
10713 case IORING_REGISTER_FILES2:
10714 case IORING_REGISTER_FILES_UPDATE2:
10715 case IORING_REGISTER_BUFFERS2:
10716 case IORING_REGISTER_BUFFERS_UPDATE:
10717 case IORING_REGISTER_IOWQ_AFF:
10718 case IORING_UNREGISTER_IOWQ_AFF:
10719 case IORING_REGISTER_IOWQ_MAX_WORKERS:
10726 static int io_ctx_quiesce(struct io_ring_ctx *ctx)
10730 percpu_ref_kill(&ctx->refs);
10733 * Drop uring mutex before waiting for references to exit. If another
10734 * thread is currently inside io_uring_enter() it might need to grab the
10735 * uring_lock to make progress. If we hold it here across the drain
10736 * wait, then we can deadlock. It's safe to drop the mutex here, since
10737 * no new references will come in after we've killed the percpu ref.
10739 mutex_unlock(&ctx->uring_lock);
10741 ret = wait_for_completion_interruptible(&ctx->ref_comp);
10744 ret = io_run_task_work_sig();
10745 } while (ret >= 0);
10746 mutex_lock(&ctx->uring_lock);
10749 io_refs_resurrect(&ctx->refs, &ctx->ref_comp);
10753 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
10754 void __user *arg, unsigned nr_args)
10755 __releases(ctx->uring_lock)
10756 __acquires(ctx->uring_lock)
10761 * We're inside the ring mutex, if the ref is already dying, then
10762 * someone else killed the ctx or is already going through
10763 * io_uring_register().
10765 if (percpu_ref_is_dying(&ctx->refs))
10768 if (ctx->restricted) {
10769 if (opcode >= IORING_REGISTER_LAST)
10771 opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
10772 if (!test_bit(opcode, ctx->restrictions.register_op))
10776 if (io_register_op_must_quiesce(opcode)) {
10777 ret = io_ctx_quiesce(ctx);
10783 case IORING_REGISTER_BUFFERS:
10784 ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
10786 case IORING_UNREGISTER_BUFFERS:
10788 if (arg || nr_args)
10790 ret = io_sqe_buffers_unregister(ctx);
10792 case IORING_REGISTER_FILES:
10793 ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
10795 case IORING_UNREGISTER_FILES:
10797 if (arg || nr_args)
10799 ret = io_sqe_files_unregister(ctx);
10801 case IORING_REGISTER_FILES_UPDATE:
10802 ret = io_register_files_update(ctx, arg, nr_args);
10804 case IORING_REGISTER_EVENTFD:
10805 case IORING_REGISTER_EVENTFD_ASYNC:
10809 ret = io_eventfd_register(ctx, arg);
10812 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
10813 ctx->eventfd_async = 1;
10815 ctx->eventfd_async = 0;
10817 case IORING_UNREGISTER_EVENTFD:
10819 if (arg || nr_args)
10821 ret = io_eventfd_unregister(ctx);
10823 case IORING_REGISTER_PROBE:
10825 if (!arg || nr_args > 256)
10827 ret = io_probe(ctx, arg, nr_args);
10829 case IORING_REGISTER_PERSONALITY:
10831 if (arg || nr_args)
10833 ret = io_register_personality(ctx);
10835 case IORING_UNREGISTER_PERSONALITY:
10839 ret = io_unregister_personality(ctx, nr_args);
10841 case IORING_REGISTER_ENABLE_RINGS:
10843 if (arg || nr_args)
10845 ret = io_register_enable_rings(ctx);
10847 case IORING_REGISTER_RESTRICTIONS:
10848 ret = io_register_restrictions(ctx, arg, nr_args);
10850 case IORING_REGISTER_FILES2:
10851 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE);
10853 case IORING_REGISTER_FILES_UPDATE2:
10854 ret = io_register_rsrc_update(ctx, arg, nr_args,
10857 case IORING_REGISTER_BUFFERS2:
10858 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER);
10860 case IORING_REGISTER_BUFFERS_UPDATE:
10861 ret = io_register_rsrc_update(ctx, arg, nr_args,
10862 IORING_RSRC_BUFFER);
10864 case IORING_REGISTER_IOWQ_AFF:
10866 if (!arg || !nr_args)
10868 ret = io_register_iowq_aff(ctx, arg, nr_args);
10870 case IORING_UNREGISTER_IOWQ_AFF:
10872 if (arg || nr_args)
10874 ret = io_unregister_iowq_aff(ctx);
10876 case IORING_REGISTER_IOWQ_MAX_WORKERS:
10878 if (!arg || nr_args != 2)
10880 ret = io_register_iowq_max_workers(ctx, arg);
10887 if (io_register_op_must_quiesce(opcode)) {
10888 /* bring the ctx back to life */
10889 percpu_ref_reinit(&ctx->refs);
10890 reinit_completion(&ctx->ref_comp);
10895 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
10896 void __user *, arg, unsigned int, nr_args)
10898 struct io_ring_ctx *ctx;
10907 if (f.file->f_op != &io_uring_fops)
10910 ctx = f.file->private_data;
10912 io_run_task_work();
10914 mutex_lock(&ctx->uring_lock);
10915 ret = __io_uring_register(ctx, opcode, arg, nr_args);
10916 mutex_unlock(&ctx->uring_lock);
10917 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
10918 ctx->cq_ev_fd != NULL, ret);
10924 static int __init io_uring_init(void)
10926 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
10927 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
10928 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
10931 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
10932 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
10933 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
10934 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
10935 BUILD_BUG_SQE_ELEM(1, __u8, flags);
10936 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
10937 BUILD_BUG_SQE_ELEM(4, __s32, fd);
10938 BUILD_BUG_SQE_ELEM(8, __u64, off);
10939 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
10940 BUILD_BUG_SQE_ELEM(16, __u64, addr);
10941 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
10942 BUILD_BUG_SQE_ELEM(24, __u32, len);
10943 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
10944 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
10945 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
10946 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
10947 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
10948 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
10949 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
10950 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
10951 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
10952 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
10953 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
10954 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
10955 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
10956 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
10957 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
10958 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
10959 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
10960 BUILD_BUG_SQE_ELEM(40, __u16, buf_group);
10961 BUILD_BUG_SQE_ELEM(42, __u16, personality);
10962 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
10963 BUILD_BUG_SQE_ELEM(44, __u32, file_index);
10965 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
10966 sizeof(struct io_uring_rsrc_update));
10967 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
10968 sizeof(struct io_uring_rsrc_update2));
10970 /* ->buf_index is u16 */
10971 BUILD_BUG_ON(IORING_MAX_REG_BUFFERS >= (1u << 16));
10973 /* should fit into one byte */
10974 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
10976 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
10977 BUILD_BUG_ON(__REQ_F_LAST_BIT > 8 * sizeof(int));
10979 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
10983 __initcall(io_uring_init);