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;
2460 unsigned int poll_flags = BLK_POLL_NOSLEEP;
2461 DEFINE_IO_COMP_BATCH(iob);
2465 * Only spin for completions if we don't have multiple devices hanging
2466 * off our complete list, and we're under the requested amount.
2468 if (ctx->poll_multi_queue || *nr_events >= min)
2469 poll_flags |= BLK_POLL_ONESHOT;
2471 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2472 struct kiocb *kiocb = &req->rw.kiocb;
2476 * Move completed and retryable entries to our local lists.
2477 * If we find a request that requires polling, break out
2478 * and complete those lists first, if we have entries there.
2480 if (READ_ONCE(req->iopoll_completed)) {
2481 list_move_tail(&req->inflight_entry, &done);
2484 if (!list_empty(&done))
2487 ret = kiocb->ki_filp->f_op->iopoll(kiocb, &iob, poll_flags);
2488 if (unlikely(ret < 0))
2491 poll_flags |= BLK_POLL_ONESHOT;
2493 /* iopoll may have completed current req */
2494 if (!rq_list_empty(iob.req_list) ||
2495 READ_ONCE(req->iopoll_completed))
2496 list_move_tail(&req->inflight_entry, &done);
2499 if (!rq_list_empty(iob.req_list))
2501 if (!list_empty(&done))
2502 io_iopoll_complete(ctx, nr_events, &done);
2508 * We can't just wait for polled events to come to us, we have to actively
2509 * find and complete them.
2511 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2513 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2516 mutex_lock(&ctx->uring_lock);
2517 while (!list_empty(&ctx->iopoll_list)) {
2518 unsigned int nr_events = 0;
2520 io_do_iopoll(ctx, &nr_events, 0);
2522 /* let it sleep and repeat later if can't complete a request */
2526 * Ensure we allow local-to-the-cpu processing to take place,
2527 * in this case we need to ensure that we reap all events.
2528 * Also let task_work, etc. to progress by releasing the mutex
2530 if (need_resched()) {
2531 mutex_unlock(&ctx->uring_lock);
2533 mutex_lock(&ctx->uring_lock);
2536 mutex_unlock(&ctx->uring_lock);
2539 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2541 unsigned int nr_events = 0;
2545 * We disallow the app entering submit/complete with polling, but we
2546 * still need to lock the ring to prevent racing with polled issue
2547 * that got punted to a workqueue.
2549 mutex_lock(&ctx->uring_lock);
2551 * Don't enter poll loop if we already have events pending.
2552 * If we do, we can potentially be spinning for commands that
2553 * already triggered a CQE (eg in error).
2555 if (test_bit(0, &ctx->check_cq_overflow))
2556 __io_cqring_overflow_flush(ctx, false);
2557 if (io_cqring_events(ctx))
2561 * If a submit got punted to a workqueue, we can have the
2562 * application entering polling for a command before it gets
2563 * issued. That app will hold the uring_lock for the duration
2564 * of the poll right here, so we need to take a breather every
2565 * now and then to ensure that the issue has a chance to add
2566 * the poll to the issued list. Otherwise we can spin here
2567 * forever, while the workqueue is stuck trying to acquire the
2570 if (list_empty(&ctx->iopoll_list)) {
2571 u32 tail = ctx->cached_cq_tail;
2573 mutex_unlock(&ctx->uring_lock);
2575 mutex_lock(&ctx->uring_lock);
2577 /* some requests don't go through iopoll_list */
2578 if (tail != ctx->cached_cq_tail ||
2579 list_empty(&ctx->iopoll_list))
2582 ret = io_do_iopoll(ctx, &nr_events, min);
2583 } while (!ret && nr_events < min && !need_resched());
2585 mutex_unlock(&ctx->uring_lock);
2589 static void kiocb_end_write(struct io_kiocb *req)
2592 * Tell lockdep we inherited freeze protection from submission
2595 if (req->flags & REQ_F_ISREG) {
2596 struct super_block *sb = file_inode(req->file)->i_sb;
2598 __sb_writers_acquired(sb, SB_FREEZE_WRITE);
2604 static bool io_resubmit_prep(struct io_kiocb *req)
2606 struct io_async_rw *rw = req->async_data;
2609 return !io_req_prep_async(req);
2610 iov_iter_restore(&rw->iter, &rw->iter_state);
2614 static bool io_rw_should_reissue(struct io_kiocb *req)
2616 umode_t mode = file_inode(req->file)->i_mode;
2617 struct io_ring_ctx *ctx = req->ctx;
2619 if (!S_ISBLK(mode) && !S_ISREG(mode))
2621 if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
2622 !(ctx->flags & IORING_SETUP_IOPOLL)))
2625 * If ref is dying, we might be running poll reap from the exit work.
2626 * Don't attempt to reissue from that path, just let it fail with
2629 if (percpu_ref_is_dying(&ctx->refs))
2632 * Play it safe and assume not safe to re-import and reissue if we're
2633 * not in the original thread group (or in task context).
2635 if (!same_thread_group(req->task, current) || !in_task())
2640 static bool io_resubmit_prep(struct io_kiocb *req)
2644 static bool io_rw_should_reissue(struct io_kiocb *req)
2650 static bool __io_complete_rw_common(struct io_kiocb *req, long res)
2652 if (req->rw.kiocb.ki_flags & IOCB_WRITE)
2653 kiocb_end_write(req);
2654 if (res != req->result) {
2655 if ((res == -EAGAIN || res == -EOPNOTSUPP) &&
2656 io_rw_should_reissue(req)) {
2657 req->flags |= REQ_F_REISSUE;
2666 static void io_req_task_complete(struct io_kiocb *req, bool *locked)
2668 unsigned int cflags = io_put_rw_kbuf(req);
2669 long res = req->result;
2672 struct io_ring_ctx *ctx = req->ctx;
2673 struct io_submit_state *state = &ctx->submit_state;
2675 io_req_complete_state(req, res, cflags);
2676 state->compl_reqs[state->compl_nr++] = req;
2677 if (state->compl_nr == ARRAY_SIZE(state->compl_reqs))
2678 io_submit_flush_completions(ctx);
2680 io_req_complete_post(req, res, cflags);
2684 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2685 unsigned int issue_flags)
2687 if (__io_complete_rw_common(req, res))
2689 __io_req_complete(req, issue_flags, req->result, io_put_rw_kbuf(req));
2692 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2694 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2696 if (__io_complete_rw_common(req, res))
2699 req->io_task_work.func = io_req_task_complete;
2700 io_req_task_work_add(req);
2703 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2705 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2707 if (kiocb->ki_flags & IOCB_WRITE)
2708 kiocb_end_write(req);
2709 if (unlikely(res != req->result)) {
2710 if (res == -EAGAIN && io_rw_should_reissue(req)) {
2711 req->flags |= REQ_F_REISSUE;
2716 WRITE_ONCE(req->result, res);
2717 /* order with io_iopoll_complete() checking ->result */
2719 WRITE_ONCE(req->iopoll_completed, 1);
2723 * After the iocb has been issued, it's safe to be found on the poll list.
2724 * Adding the kiocb to the list AFTER submission ensures that we don't
2725 * find it from a io_do_iopoll() thread before the issuer is done
2726 * accessing the kiocb cookie.
2728 static void io_iopoll_req_issued(struct io_kiocb *req)
2730 struct io_ring_ctx *ctx = req->ctx;
2731 const bool in_async = io_wq_current_is_worker();
2733 /* workqueue context doesn't hold uring_lock, grab it now */
2734 if (unlikely(in_async))
2735 mutex_lock(&ctx->uring_lock);
2738 * Track whether we have multiple files in our lists. This will impact
2739 * how we do polling eventually, not spinning if we're on potentially
2740 * different devices.
2742 if (list_empty(&ctx->iopoll_list)) {
2743 ctx->poll_multi_queue = false;
2744 } else if (!ctx->poll_multi_queue) {
2745 struct io_kiocb *list_req;
2747 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2750 if (list_req->file != req->file)
2751 ctx->poll_multi_queue = true;
2755 * For fast devices, IO may have already completed. If it has, add
2756 * it to the front so we find it first.
2758 if (READ_ONCE(req->iopoll_completed))
2759 list_add(&req->inflight_entry, &ctx->iopoll_list);
2761 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2763 if (unlikely(in_async)) {
2765 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
2766 * in sq thread task context or in io worker task context. If
2767 * current task context is sq thread, we don't need to check
2768 * whether should wake up sq thread.
2770 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2771 wq_has_sleeper(&ctx->sq_data->wait))
2772 wake_up(&ctx->sq_data->wait);
2774 mutex_unlock(&ctx->uring_lock);
2778 static bool io_bdev_nowait(struct block_device *bdev)
2780 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2784 * If we tracked the file through the SCM inflight mechanism, we could support
2785 * any file. For now, just ensure that anything potentially problematic is done
2788 static bool __io_file_supports_nowait(struct file *file, int rw)
2790 umode_t mode = file_inode(file)->i_mode;
2792 if (S_ISBLK(mode)) {
2793 if (IS_ENABLED(CONFIG_BLOCK) &&
2794 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2800 if (S_ISREG(mode)) {
2801 if (IS_ENABLED(CONFIG_BLOCK) &&
2802 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2803 file->f_op != &io_uring_fops)
2808 /* any ->read/write should understand O_NONBLOCK */
2809 if (file->f_flags & O_NONBLOCK)
2812 if (!(file->f_mode & FMODE_NOWAIT))
2816 return file->f_op->read_iter != NULL;
2818 return file->f_op->write_iter != NULL;
2821 static bool io_file_supports_nowait(struct io_kiocb *req, int rw)
2823 if (rw == READ && (req->flags & REQ_F_NOWAIT_READ))
2825 else if (rw == WRITE && (req->flags & REQ_F_NOWAIT_WRITE))
2828 return __io_file_supports_nowait(req->file, rw);
2831 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2834 struct io_ring_ctx *ctx = req->ctx;
2835 struct kiocb *kiocb = &req->rw.kiocb;
2836 struct file *file = req->file;
2840 if (!io_req_ffs_set(req) && S_ISREG(file_inode(file)->i_mode))
2841 req->flags |= REQ_F_ISREG;
2843 kiocb->ki_pos = READ_ONCE(sqe->off);
2844 if (kiocb->ki_pos == -1 && !(file->f_mode & FMODE_STREAM)) {
2845 req->flags |= REQ_F_CUR_POS;
2846 kiocb->ki_pos = file->f_pos;
2848 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2849 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2850 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2855 * If the file is marked O_NONBLOCK, still allow retry for it if it
2856 * supports async. Otherwise it's impossible to use O_NONBLOCK files
2857 * reliably. If not, or it IOCB_NOWAIT is set, don't retry.
2859 if ((kiocb->ki_flags & IOCB_NOWAIT) ||
2860 ((file->f_flags & O_NONBLOCK) && !io_file_supports_nowait(req, rw)))
2861 req->flags |= REQ_F_NOWAIT;
2863 ioprio = READ_ONCE(sqe->ioprio);
2865 ret = ioprio_check_cap(ioprio);
2869 kiocb->ki_ioprio = ioprio;
2871 kiocb->ki_ioprio = get_current_ioprio();
2873 if (ctx->flags & IORING_SETUP_IOPOLL) {
2874 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2875 !kiocb->ki_filp->f_op->iopoll)
2878 kiocb->ki_flags |= IOCB_HIPRI | IOCB_ALLOC_CACHE;
2879 kiocb->ki_complete = io_complete_rw_iopoll;
2880 req->iopoll_completed = 0;
2882 if (kiocb->ki_flags & IOCB_HIPRI)
2884 kiocb->ki_complete = io_complete_rw;
2887 if (req->opcode == IORING_OP_READ_FIXED ||
2888 req->opcode == IORING_OP_WRITE_FIXED) {
2890 io_req_set_rsrc_node(req);
2893 req->rw.addr = READ_ONCE(sqe->addr);
2894 req->rw.len = READ_ONCE(sqe->len);
2895 req->buf_index = READ_ONCE(sqe->buf_index);
2899 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2905 case -ERESTARTNOINTR:
2906 case -ERESTARTNOHAND:
2907 case -ERESTART_RESTARTBLOCK:
2909 * We can't just restart the syscall, since previously
2910 * submitted sqes may already be in progress. Just fail this
2916 kiocb->ki_complete(kiocb, ret, 0);
2920 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2921 unsigned int issue_flags)
2923 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2924 struct io_async_rw *io = req->async_data;
2926 /* add previously done IO, if any */
2927 if (io && io->bytes_done > 0) {
2929 ret = io->bytes_done;
2931 ret += io->bytes_done;
2934 if (req->flags & REQ_F_CUR_POS)
2935 req->file->f_pos = kiocb->ki_pos;
2936 if (ret >= 0 && (kiocb->ki_complete == io_complete_rw))
2937 __io_complete_rw(req, ret, 0, issue_flags);
2939 io_rw_done(kiocb, ret);
2941 if (req->flags & REQ_F_REISSUE) {
2942 req->flags &= ~REQ_F_REISSUE;
2943 if (io_resubmit_prep(req)) {
2944 io_req_task_queue_reissue(req);
2946 unsigned int cflags = io_put_rw_kbuf(req);
2947 struct io_ring_ctx *ctx = req->ctx;
2950 if (!(issue_flags & IO_URING_F_NONBLOCK)) {
2951 mutex_lock(&ctx->uring_lock);
2952 __io_req_complete(req, issue_flags, ret, cflags);
2953 mutex_unlock(&ctx->uring_lock);
2955 __io_req_complete(req, issue_flags, ret, cflags);
2961 static int __io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter,
2962 struct io_mapped_ubuf *imu)
2964 size_t len = req->rw.len;
2965 u64 buf_end, buf_addr = req->rw.addr;
2968 if (unlikely(check_add_overflow(buf_addr, (u64)len, &buf_end)))
2970 /* not inside the mapped region */
2971 if (unlikely(buf_addr < imu->ubuf || buf_end > imu->ubuf_end))
2975 * May not be a start of buffer, set size appropriately
2976 * and advance us to the beginning.
2978 offset = buf_addr - imu->ubuf;
2979 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2983 * Don't use iov_iter_advance() here, as it's really slow for
2984 * using the latter parts of a big fixed buffer - it iterates
2985 * over each segment manually. We can cheat a bit here, because
2988 * 1) it's a BVEC iter, we set it up
2989 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2990 * first and last bvec
2992 * So just find our index, and adjust the iterator afterwards.
2993 * If the offset is within the first bvec (or the whole first
2994 * bvec, just use iov_iter_advance(). This makes it easier
2995 * since we can just skip the first segment, which may not
2996 * be PAGE_SIZE aligned.
2998 const struct bio_vec *bvec = imu->bvec;
3000 if (offset <= bvec->bv_len) {
3001 iov_iter_advance(iter, offset);
3003 unsigned long seg_skip;
3005 /* skip first vec */
3006 offset -= bvec->bv_len;
3007 seg_skip = 1 + (offset >> PAGE_SHIFT);
3009 iter->bvec = bvec + seg_skip;
3010 iter->nr_segs -= seg_skip;
3011 iter->count -= bvec->bv_len + offset;
3012 iter->iov_offset = offset & ~PAGE_MASK;
3019 static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
3021 struct io_ring_ctx *ctx = req->ctx;
3022 struct io_mapped_ubuf *imu = req->imu;
3023 u16 index, buf_index = req->buf_index;
3026 if (unlikely(buf_index >= ctx->nr_user_bufs))
3028 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
3029 imu = READ_ONCE(ctx->user_bufs[index]);
3032 return __io_import_fixed(req, rw, iter, imu);
3035 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
3038 mutex_unlock(&ctx->uring_lock);
3041 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
3044 * "Normal" inline submissions always hold the uring_lock, since we
3045 * grab it from the system call. Same is true for the SQPOLL offload.
3046 * The only exception is when we've detached the request and issue it
3047 * from an async worker thread, grab the lock for that case.
3050 mutex_lock(&ctx->uring_lock);
3053 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
3054 int bgid, struct io_buffer *kbuf,
3057 struct io_buffer *head;
3059 if (req->flags & REQ_F_BUFFER_SELECTED)
3062 io_ring_submit_lock(req->ctx, needs_lock);
3064 lockdep_assert_held(&req->ctx->uring_lock);
3066 head = xa_load(&req->ctx->io_buffers, bgid);
3068 if (!list_empty(&head->list)) {
3069 kbuf = list_last_entry(&head->list, struct io_buffer,
3071 list_del(&kbuf->list);
3074 xa_erase(&req->ctx->io_buffers, bgid);
3076 if (*len > kbuf->len)
3079 kbuf = ERR_PTR(-ENOBUFS);
3082 io_ring_submit_unlock(req->ctx, needs_lock);
3087 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
3090 struct io_buffer *kbuf;
3093 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3094 bgid = req->buf_index;
3095 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
3098 req->rw.addr = (u64) (unsigned long) kbuf;
3099 req->flags |= REQ_F_BUFFER_SELECTED;
3100 return u64_to_user_ptr(kbuf->addr);
3103 #ifdef CONFIG_COMPAT
3104 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
3107 struct compat_iovec __user *uiov;
3108 compat_ssize_t clen;
3112 uiov = u64_to_user_ptr(req->rw.addr);
3113 if (!access_ok(uiov, sizeof(*uiov)))
3115 if (__get_user(clen, &uiov->iov_len))
3121 buf = io_rw_buffer_select(req, &len, needs_lock);
3123 return PTR_ERR(buf);
3124 iov[0].iov_base = buf;
3125 iov[0].iov_len = (compat_size_t) len;
3130 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3133 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
3137 if (copy_from_user(iov, uiov, sizeof(*uiov)))
3140 len = iov[0].iov_len;
3143 buf = io_rw_buffer_select(req, &len, needs_lock);
3145 return PTR_ERR(buf);
3146 iov[0].iov_base = buf;
3147 iov[0].iov_len = len;
3151 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3154 if (req->flags & REQ_F_BUFFER_SELECTED) {
3155 struct io_buffer *kbuf;
3157 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3158 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
3159 iov[0].iov_len = kbuf->len;
3162 if (req->rw.len != 1)
3165 #ifdef CONFIG_COMPAT
3166 if (req->ctx->compat)
3167 return io_compat_import(req, iov, needs_lock);
3170 return __io_iov_buffer_select(req, iov, needs_lock);
3173 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
3174 struct iov_iter *iter, bool needs_lock)
3176 void __user *buf = u64_to_user_ptr(req->rw.addr);
3177 size_t sqe_len = req->rw.len;
3178 u8 opcode = req->opcode;
3181 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3183 return io_import_fixed(req, rw, iter);
3186 /* buffer index only valid with fixed read/write, or buffer select */
3187 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3190 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3191 if (req->flags & REQ_F_BUFFER_SELECT) {
3192 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3194 return PTR_ERR(buf);
3195 req->rw.len = sqe_len;
3198 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3203 if (req->flags & REQ_F_BUFFER_SELECT) {
3204 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3206 iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
3211 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3215 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3217 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3221 * For files that don't have ->read_iter() and ->write_iter(), handle them
3222 * by looping over ->read() or ->write() manually.
3224 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3226 struct kiocb *kiocb = &req->rw.kiocb;
3227 struct file *file = req->file;
3231 * Don't support polled IO through this interface, and we can't
3232 * support non-blocking either. For the latter, this just causes
3233 * the kiocb to be handled from an async context.
3235 if (kiocb->ki_flags & IOCB_HIPRI)
3237 if (kiocb->ki_flags & IOCB_NOWAIT)
3240 while (iov_iter_count(iter)) {
3244 if (!iov_iter_is_bvec(iter)) {
3245 iovec = iov_iter_iovec(iter);
3247 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3248 iovec.iov_len = req->rw.len;
3252 nr = file->f_op->read(file, iovec.iov_base,
3253 iovec.iov_len, io_kiocb_ppos(kiocb));
3255 nr = file->f_op->write(file, iovec.iov_base,
3256 iovec.iov_len, io_kiocb_ppos(kiocb));
3264 if (!iov_iter_is_bvec(iter)) {
3265 iov_iter_advance(iter, nr);
3271 if (nr != iovec.iov_len)
3278 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3279 const struct iovec *fast_iov, struct iov_iter *iter)
3281 struct io_async_rw *rw = req->async_data;
3283 memcpy(&rw->iter, iter, sizeof(*iter));
3284 rw->free_iovec = iovec;
3286 /* can only be fixed buffers, no need to do anything */
3287 if (iov_iter_is_bvec(iter))
3290 unsigned iov_off = 0;
3292 rw->iter.iov = rw->fast_iov;
3293 if (iter->iov != fast_iov) {
3294 iov_off = iter->iov - fast_iov;
3295 rw->iter.iov += iov_off;
3297 if (rw->fast_iov != fast_iov)
3298 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3299 sizeof(struct iovec) * iter->nr_segs);
3301 req->flags |= REQ_F_NEED_CLEANUP;
3305 static inline int io_alloc_async_data(struct io_kiocb *req)
3307 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3308 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3309 return req->async_data == NULL;
3312 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3313 const struct iovec *fast_iov,
3314 struct iov_iter *iter, bool force)
3316 if (!force && !io_op_defs[req->opcode].needs_async_setup)
3318 if (!req->async_data) {
3319 struct io_async_rw *iorw;
3321 if (io_alloc_async_data(req)) {
3326 io_req_map_rw(req, iovec, fast_iov, iter);
3327 iorw = req->async_data;
3328 /* we've copied and mapped the iter, ensure state is saved */
3329 iov_iter_save_state(&iorw->iter, &iorw->iter_state);
3334 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3336 struct io_async_rw *iorw = req->async_data;
3337 struct iovec *iov = iorw->fast_iov;
3340 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3341 if (unlikely(ret < 0))
3344 iorw->bytes_done = 0;
3345 iorw->free_iovec = iov;
3347 req->flags |= REQ_F_NEED_CLEANUP;
3348 iov_iter_save_state(&iorw->iter, &iorw->iter_state);
3352 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3354 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3356 return io_prep_rw(req, sqe, READ);
3360 * This is our waitqueue callback handler, registered through lock_page_async()
3361 * when we initially tried to do the IO with the iocb armed our waitqueue.
3362 * This gets called when the page is unlocked, and we generally expect that to
3363 * happen when the page IO is completed and the page is now uptodate. This will
3364 * queue a task_work based retry of the operation, attempting to copy the data
3365 * again. If the latter fails because the page was NOT uptodate, then we will
3366 * do a thread based blocking retry of the operation. That's the unexpected
3369 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3370 int sync, void *arg)
3372 struct wait_page_queue *wpq;
3373 struct io_kiocb *req = wait->private;
3374 struct wait_page_key *key = arg;
3376 wpq = container_of(wait, struct wait_page_queue, wait);
3378 if (!wake_page_match(wpq, key))
3381 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3382 list_del_init(&wait->entry);
3383 io_req_task_queue(req);
3388 * This controls whether a given IO request should be armed for async page
3389 * based retry. If we return false here, the request is handed to the async
3390 * worker threads for retry. If we're doing buffered reads on a regular file,
3391 * we prepare a private wait_page_queue entry and retry the operation. This
3392 * will either succeed because the page is now uptodate and unlocked, or it
3393 * will register a callback when the page is unlocked at IO completion. Through
3394 * that callback, io_uring uses task_work to setup a retry of the operation.
3395 * That retry will attempt the buffered read again. The retry will generally
3396 * succeed, or in rare cases where it fails, we then fall back to using the
3397 * async worker threads for a blocking retry.
3399 static bool io_rw_should_retry(struct io_kiocb *req)
3401 struct io_async_rw *rw = req->async_data;
3402 struct wait_page_queue *wait = &rw->wpq;
3403 struct kiocb *kiocb = &req->rw.kiocb;
3405 /* never retry for NOWAIT, we just complete with -EAGAIN */
3406 if (req->flags & REQ_F_NOWAIT)
3409 /* Only for buffered IO */
3410 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3414 * just use poll if we can, and don't attempt if the fs doesn't
3415 * support callback based unlocks
3417 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3420 wait->wait.func = io_async_buf_func;
3421 wait->wait.private = req;
3422 wait->wait.flags = 0;
3423 INIT_LIST_HEAD(&wait->wait.entry);
3424 kiocb->ki_flags |= IOCB_WAITQ;
3425 kiocb->ki_flags &= ~IOCB_NOWAIT;
3426 kiocb->ki_waitq = wait;
3430 static inline int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3432 if (req->file->f_op->read_iter)
3433 return call_read_iter(req->file, &req->rw.kiocb, iter);
3434 else if (req->file->f_op->read)
3435 return loop_rw_iter(READ, req, iter);
3440 static bool need_read_all(struct io_kiocb *req)
3442 return req->flags & REQ_F_ISREG ||
3443 S_ISBLK(file_inode(req->file)->i_mode);
3446 static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3448 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3449 struct kiocb *kiocb = &req->rw.kiocb;
3450 struct iov_iter __iter, *iter = &__iter;
3451 struct io_async_rw *rw = req->async_data;
3452 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3453 struct iov_iter_state __state, *state;
3458 state = &rw->iter_state;
3460 * We come here from an earlier attempt, restore our state to
3461 * match in case it doesn't. It's cheap enough that we don't
3462 * need to make this conditional.
3464 iov_iter_restore(iter, state);
3467 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3471 iov_iter_save_state(iter, state);
3473 req->result = iov_iter_count(iter);
3475 /* Ensure we clear previously set non-block flag */
3476 if (!force_nonblock)
3477 kiocb->ki_flags &= ~IOCB_NOWAIT;
3479 kiocb->ki_flags |= IOCB_NOWAIT;
3481 /* If the file doesn't support async, just async punt */
3482 if (force_nonblock && !io_file_supports_nowait(req, READ)) {
3483 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3484 return ret ?: -EAGAIN;
3487 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), req->result);
3488 if (unlikely(ret)) {
3493 ret = io_iter_do_read(req, iter);
3495 if (ret == -EAGAIN || (req->flags & REQ_F_REISSUE)) {
3496 req->flags &= ~REQ_F_REISSUE;
3497 /* IOPOLL retry should happen for io-wq threads */
3498 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3500 /* no retry on NONBLOCK nor RWF_NOWAIT */
3501 if (req->flags & REQ_F_NOWAIT)
3504 } else if (ret == -EIOCBQUEUED) {
3506 } else if (ret <= 0 || ret == req->result || !force_nonblock ||
3507 (req->flags & REQ_F_NOWAIT) || !need_read_all(req)) {
3508 /* read all, failed, already did sync or don't want to retry */
3513 * Don't depend on the iter state matching what was consumed, or being
3514 * untouched in case of error. Restore it and we'll advance it
3515 * manually if we need to.
3517 iov_iter_restore(iter, state);
3519 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3524 rw = req->async_data;
3526 * Now use our persistent iterator and state, if we aren't already.
3527 * We've restored and mapped the iter to match.
3529 if (iter != &rw->iter) {
3531 state = &rw->iter_state;
3536 * We end up here because of a partial read, either from
3537 * above or inside this loop. Advance the iter by the bytes
3538 * that were consumed.
3540 iov_iter_advance(iter, ret);
3541 if (!iov_iter_count(iter))
3543 rw->bytes_done += ret;
3544 iov_iter_save_state(iter, state);
3546 /* if we can retry, do so with the callbacks armed */
3547 if (!io_rw_should_retry(req)) {
3548 kiocb->ki_flags &= ~IOCB_WAITQ;
3553 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3554 * we get -EIOCBQUEUED, then we'll get a notification when the
3555 * desired page gets unlocked. We can also get a partial read
3556 * here, and if we do, then just retry at the new offset.
3558 ret = io_iter_do_read(req, iter);
3559 if (ret == -EIOCBQUEUED)
3561 /* we got some bytes, but not all. retry. */
3562 kiocb->ki_flags &= ~IOCB_WAITQ;
3563 iov_iter_restore(iter, state);
3566 kiocb_done(kiocb, ret, issue_flags);
3568 /* it's faster to check here then delegate to kfree */
3574 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3576 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3578 return io_prep_rw(req, sqe, WRITE);
3581 static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3583 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3584 struct kiocb *kiocb = &req->rw.kiocb;
3585 struct iov_iter __iter, *iter = &__iter;
3586 struct io_async_rw *rw = req->async_data;
3587 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3588 struct iov_iter_state __state, *state;
3593 state = &rw->iter_state;
3594 iov_iter_restore(iter, state);
3597 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3601 iov_iter_save_state(iter, state);
3603 req->result = iov_iter_count(iter);
3605 /* Ensure we clear previously set non-block flag */
3606 if (!force_nonblock)
3607 kiocb->ki_flags &= ~IOCB_NOWAIT;
3609 kiocb->ki_flags |= IOCB_NOWAIT;
3611 /* If the file doesn't support async, just async punt */
3612 if (force_nonblock && !io_file_supports_nowait(req, WRITE))
3615 /* file path doesn't support NOWAIT for non-direct_IO */
3616 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3617 (req->flags & REQ_F_ISREG))
3620 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), req->result);
3625 * Open-code file_start_write here to grab freeze protection,
3626 * which will be released by another thread in
3627 * io_complete_rw(). Fool lockdep by telling it the lock got
3628 * released so that it doesn't complain about the held lock when
3629 * we return to userspace.
3631 if (req->flags & REQ_F_ISREG) {
3632 sb_start_write(file_inode(req->file)->i_sb);
3633 __sb_writers_release(file_inode(req->file)->i_sb,
3636 kiocb->ki_flags |= IOCB_WRITE;
3638 if (req->file->f_op->write_iter)
3639 ret2 = call_write_iter(req->file, kiocb, iter);
3640 else if (req->file->f_op->write)
3641 ret2 = loop_rw_iter(WRITE, req, iter);
3645 if (req->flags & REQ_F_REISSUE) {
3646 req->flags &= ~REQ_F_REISSUE;
3651 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3652 * retry them without IOCB_NOWAIT.
3654 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3656 /* no retry on NONBLOCK nor RWF_NOWAIT */
3657 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3659 if (!force_nonblock || ret2 != -EAGAIN) {
3660 /* IOPOLL retry should happen for io-wq threads */
3661 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3664 kiocb_done(kiocb, ret2, issue_flags);
3667 iov_iter_restore(iter, state);
3668 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3669 return ret ?: -EAGAIN;
3672 /* it's reportedly faster than delegating the null check to kfree() */
3678 static int io_renameat_prep(struct io_kiocb *req,
3679 const struct io_uring_sqe *sqe)
3681 struct io_rename *ren = &req->rename;
3682 const char __user *oldf, *newf;
3684 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3686 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
3688 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3691 ren->old_dfd = READ_ONCE(sqe->fd);
3692 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3693 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3694 ren->new_dfd = READ_ONCE(sqe->len);
3695 ren->flags = READ_ONCE(sqe->rename_flags);
3697 ren->oldpath = getname(oldf);
3698 if (IS_ERR(ren->oldpath))
3699 return PTR_ERR(ren->oldpath);
3701 ren->newpath = getname(newf);
3702 if (IS_ERR(ren->newpath)) {
3703 putname(ren->oldpath);
3704 return PTR_ERR(ren->newpath);
3707 req->flags |= REQ_F_NEED_CLEANUP;
3711 static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3713 struct io_rename *ren = &req->rename;
3716 if (issue_flags & IO_URING_F_NONBLOCK)
3719 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3720 ren->newpath, ren->flags);
3722 req->flags &= ~REQ_F_NEED_CLEANUP;
3725 io_req_complete(req, ret);
3729 static int io_unlinkat_prep(struct io_kiocb *req,
3730 const struct io_uring_sqe *sqe)
3732 struct io_unlink *un = &req->unlink;
3733 const char __user *fname;
3735 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3737 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
3740 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3743 un->dfd = READ_ONCE(sqe->fd);
3745 un->flags = READ_ONCE(sqe->unlink_flags);
3746 if (un->flags & ~AT_REMOVEDIR)
3749 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3750 un->filename = getname(fname);
3751 if (IS_ERR(un->filename))
3752 return PTR_ERR(un->filename);
3754 req->flags |= REQ_F_NEED_CLEANUP;
3758 static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3760 struct io_unlink *un = &req->unlink;
3763 if (issue_flags & IO_URING_F_NONBLOCK)
3766 if (un->flags & AT_REMOVEDIR)
3767 ret = do_rmdir(un->dfd, un->filename);
3769 ret = do_unlinkat(un->dfd, un->filename);
3771 req->flags &= ~REQ_F_NEED_CLEANUP;
3774 io_req_complete(req, ret);
3778 static int io_mkdirat_prep(struct io_kiocb *req,
3779 const struct io_uring_sqe *sqe)
3781 struct io_mkdir *mkd = &req->mkdir;
3782 const char __user *fname;
3784 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3786 if (sqe->ioprio || sqe->off || sqe->rw_flags || sqe->buf_index ||
3789 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3792 mkd->dfd = READ_ONCE(sqe->fd);
3793 mkd->mode = READ_ONCE(sqe->len);
3795 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3796 mkd->filename = getname(fname);
3797 if (IS_ERR(mkd->filename))
3798 return PTR_ERR(mkd->filename);
3800 req->flags |= REQ_F_NEED_CLEANUP;
3804 static int io_mkdirat(struct io_kiocb *req, int issue_flags)
3806 struct io_mkdir *mkd = &req->mkdir;
3809 if (issue_flags & IO_URING_F_NONBLOCK)
3812 ret = do_mkdirat(mkd->dfd, mkd->filename, mkd->mode);
3814 req->flags &= ~REQ_F_NEED_CLEANUP;
3817 io_req_complete(req, ret);
3821 static int io_symlinkat_prep(struct io_kiocb *req,
3822 const struct io_uring_sqe *sqe)
3824 struct io_symlink *sl = &req->symlink;
3825 const char __user *oldpath, *newpath;
3827 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3829 if (sqe->ioprio || sqe->len || sqe->rw_flags || sqe->buf_index ||
3832 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3835 sl->new_dfd = READ_ONCE(sqe->fd);
3836 oldpath = u64_to_user_ptr(READ_ONCE(sqe->addr));
3837 newpath = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3839 sl->oldpath = getname(oldpath);
3840 if (IS_ERR(sl->oldpath))
3841 return PTR_ERR(sl->oldpath);
3843 sl->newpath = getname(newpath);
3844 if (IS_ERR(sl->newpath)) {
3845 putname(sl->oldpath);
3846 return PTR_ERR(sl->newpath);
3849 req->flags |= REQ_F_NEED_CLEANUP;
3853 static int io_symlinkat(struct io_kiocb *req, int issue_flags)
3855 struct io_symlink *sl = &req->symlink;
3858 if (issue_flags & IO_URING_F_NONBLOCK)
3861 ret = do_symlinkat(sl->oldpath, sl->new_dfd, sl->newpath);
3863 req->flags &= ~REQ_F_NEED_CLEANUP;
3866 io_req_complete(req, ret);
3870 static int io_linkat_prep(struct io_kiocb *req,
3871 const struct io_uring_sqe *sqe)
3873 struct io_hardlink *lnk = &req->hardlink;
3874 const char __user *oldf, *newf;
3876 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3878 if (sqe->ioprio || sqe->rw_flags || sqe->buf_index || sqe->splice_fd_in)
3880 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3883 lnk->old_dfd = READ_ONCE(sqe->fd);
3884 lnk->new_dfd = READ_ONCE(sqe->len);
3885 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3886 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3887 lnk->flags = READ_ONCE(sqe->hardlink_flags);
3889 lnk->oldpath = getname(oldf);
3890 if (IS_ERR(lnk->oldpath))
3891 return PTR_ERR(lnk->oldpath);
3893 lnk->newpath = getname(newf);
3894 if (IS_ERR(lnk->newpath)) {
3895 putname(lnk->oldpath);
3896 return PTR_ERR(lnk->newpath);
3899 req->flags |= REQ_F_NEED_CLEANUP;
3903 static int io_linkat(struct io_kiocb *req, int issue_flags)
3905 struct io_hardlink *lnk = &req->hardlink;
3908 if (issue_flags & IO_URING_F_NONBLOCK)
3911 ret = do_linkat(lnk->old_dfd, lnk->oldpath, lnk->new_dfd,
3912 lnk->newpath, lnk->flags);
3914 req->flags &= ~REQ_F_NEED_CLEANUP;
3917 io_req_complete(req, ret);
3921 static int io_shutdown_prep(struct io_kiocb *req,
3922 const struct io_uring_sqe *sqe)
3924 #if defined(CONFIG_NET)
3925 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3927 if (unlikely(sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3928 sqe->buf_index || sqe->splice_fd_in))
3931 req->shutdown.how = READ_ONCE(sqe->len);
3938 static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3940 #if defined(CONFIG_NET)
3941 struct socket *sock;
3944 if (issue_flags & IO_URING_F_NONBLOCK)
3947 sock = sock_from_file(req->file);
3948 if (unlikely(!sock))
3951 ret = __sys_shutdown_sock(sock, req->shutdown.how);
3954 io_req_complete(req, ret);
3961 static int __io_splice_prep(struct io_kiocb *req,
3962 const struct io_uring_sqe *sqe)
3964 struct io_splice *sp = &req->splice;
3965 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3967 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3971 sp->len = READ_ONCE(sqe->len);
3972 sp->flags = READ_ONCE(sqe->splice_flags);
3974 if (unlikely(sp->flags & ~valid_flags))
3977 sp->file_in = io_file_get(req->ctx, req, READ_ONCE(sqe->splice_fd_in),
3978 (sp->flags & SPLICE_F_FD_IN_FIXED));
3981 req->flags |= REQ_F_NEED_CLEANUP;
3985 static int io_tee_prep(struct io_kiocb *req,
3986 const struct io_uring_sqe *sqe)
3988 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3990 return __io_splice_prep(req, sqe);
3993 static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
3995 struct io_splice *sp = &req->splice;
3996 struct file *in = sp->file_in;
3997 struct file *out = sp->file_out;
3998 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
4001 if (issue_flags & IO_URING_F_NONBLOCK)
4004 ret = do_tee(in, out, sp->len, flags);
4006 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
4008 req->flags &= ~REQ_F_NEED_CLEANUP;
4012 io_req_complete(req, ret);
4016 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4018 struct io_splice *sp = &req->splice;
4020 sp->off_in = READ_ONCE(sqe->splice_off_in);
4021 sp->off_out = READ_ONCE(sqe->off);
4022 return __io_splice_prep(req, sqe);
4025 static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
4027 struct io_splice *sp = &req->splice;
4028 struct file *in = sp->file_in;
4029 struct file *out = sp->file_out;
4030 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
4031 loff_t *poff_in, *poff_out;
4034 if (issue_flags & IO_URING_F_NONBLOCK)
4037 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
4038 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
4041 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
4043 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
4045 req->flags &= ~REQ_F_NEED_CLEANUP;
4049 io_req_complete(req, ret);
4054 * IORING_OP_NOP just posts a completion event, nothing else.
4056 static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
4058 struct io_ring_ctx *ctx = req->ctx;
4060 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4063 __io_req_complete(req, issue_flags, 0, 0);
4067 static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4069 struct io_ring_ctx *ctx = req->ctx;
4074 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4076 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index ||
4080 req->sync.flags = READ_ONCE(sqe->fsync_flags);
4081 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
4084 req->sync.off = READ_ONCE(sqe->off);
4085 req->sync.len = READ_ONCE(sqe->len);
4089 static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
4091 loff_t end = req->sync.off + req->sync.len;
4094 /* fsync always requires a blocking context */
4095 if (issue_flags & IO_URING_F_NONBLOCK)
4098 ret = vfs_fsync_range(req->file, req->sync.off,
4099 end > 0 ? end : LLONG_MAX,
4100 req->sync.flags & IORING_FSYNC_DATASYNC);
4103 io_req_complete(req, ret);
4107 static int io_fallocate_prep(struct io_kiocb *req,
4108 const struct io_uring_sqe *sqe)
4110 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags ||
4113 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4116 req->sync.off = READ_ONCE(sqe->off);
4117 req->sync.len = READ_ONCE(sqe->addr);
4118 req->sync.mode = READ_ONCE(sqe->len);
4122 static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
4126 /* fallocate always requiring blocking context */
4127 if (issue_flags & IO_URING_F_NONBLOCK)
4129 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
4133 io_req_complete(req, ret);
4137 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4139 const char __user *fname;
4142 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4144 if (unlikely(sqe->ioprio || sqe->buf_index))
4146 if (unlikely(req->flags & REQ_F_FIXED_FILE))
4149 /* open.how should be already initialised */
4150 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
4151 req->open.how.flags |= O_LARGEFILE;
4153 req->open.dfd = READ_ONCE(sqe->fd);
4154 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
4155 req->open.filename = getname(fname);
4156 if (IS_ERR(req->open.filename)) {
4157 ret = PTR_ERR(req->open.filename);
4158 req->open.filename = NULL;
4162 req->open.file_slot = READ_ONCE(sqe->file_index);
4163 if (req->open.file_slot && (req->open.how.flags & O_CLOEXEC))
4166 req->open.nofile = rlimit(RLIMIT_NOFILE);
4167 req->flags |= REQ_F_NEED_CLEANUP;
4171 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4173 u64 mode = READ_ONCE(sqe->len);
4174 u64 flags = READ_ONCE(sqe->open_flags);
4176 req->open.how = build_open_how(flags, mode);
4177 return __io_openat_prep(req, sqe);
4180 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4182 struct open_how __user *how;
4186 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4187 len = READ_ONCE(sqe->len);
4188 if (len < OPEN_HOW_SIZE_VER0)
4191 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
4196 return __io_openat_prep(req, sqe);
4199 static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
4201 struct open_flags op;
4203 bool resolve_nonblock, nonblock_set;
4204 bool fixed = !!req->open.file_slot;
4207 ret = build_open_flags(&req->open.how, &op);
4210 nonblock_set = op.open_flag & O_NONBLOCK;
4211 resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
4212 if (issue_flags & IO_URING_F_NONBLOCK) {
4214 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
4215 * it'll always -EAGAIN
4217 if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
4219 op.lookup_flags |= LOOKUP_CACHED;
4220 op.open_flag |= O_NONBLOCK;
4224 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
4229 file = do_filp_open(req->open.dfd, req->open.filename, &op);
4232 * We could hang on to this 'fd' on retrying, but seems like
4233 * marginal gain for something that is now known to be a slower
4234 * path. So just put it, and we'll get a new one when we retry.
4239 ret = PTR_ERR(file);
4240 /* only retry if RESOLVE_CACHED wasn't already set by application */
4241 if (ret == -EAGAIN &&
4242 (!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)))
4247 if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
4248 file->f_flags &= ~O_NONBLOCK;
4249 fsnotify_open(file);
4252 fd_install(ret, file);
4254 ret = io_install_fixed_file(req, file, issue_flags,
4255 req->open.file_slot - 1);
4257 putname(req->open.filename);
4258 req->flags &= ~REQ_F_NEED_CLEANUP;
4261 __io_req_complete(req, issue_flags, ret, 0);
4265 static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
4267 return io_openat2(req, issue_flags);
4270 static int io_remove_buffers_prep(struct io_kiocb *req,
4271 const struct io_uring_sqe *sqe)
4273 struct io_provide_buf *p = &req->pbuf;
4276 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off ||
4280 tmp = READ_ONCE(sqe->fd);
4281 if (!tmp || tmp > USHRT_MAX)
4284 memset(p, 0, sizeof(*p));
4286 p->bgid = READ_ONCE(sqe->buf_group);
4290 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
4291 int bgid, unsigned nbufs)
4295 /* shouldn't happen */
4299 /* the head kbuf is the list itself */
4300 while (!list_empty(&buf->list)) {
4301 struct io_buffer *nxt;
4303 nxt = list_first_entry(&buf->list, struct io_buffer, list);
4304 list_del(&nxt->list);
4311 xa_erase(&ctx->io_buffers, bgid);
4316 static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
4318 struct io_provide_buf *p = &req->pbuf;
4319 struct io_ring_ctx *ctx = req->ctx;
4320 struct io_buffer *head;
4322 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4324 io_ring_submit_lock(ctx, !force_nonblock);
4326 lockdep_assert_held(&ctx->uring_lock);
4329 head = xa_load(&ctx->io_buffers, p->bgid);
4331 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
4335 /* complete before unlock, IOPOLL may need the lock */
4336 __io_req_complete(req, issue_flags, ret, 0);
4337 io_ring_submit_unlock(ctx, !force_nonblock);
4341 static int io_provide_buffers_prep(struct io_kiocb *req,
4342 const struct io_uring_sqe *sqe)
4344 unsigned long size, tmp_check;
4345 struct io_provide_buf *p = &req->pbuf;
4348 if (sqe->ioprio || sqe->rw_flags || sqe->splice_fd_in)
4351 tmp = READ_ONCE(sqe->fd);
4352 if (!tmp || tmp > USHRT_MAX)
4355 p->addr = READ_ONCE(sqe->addr);
4356 p->len = READ_ONCE(sqe->len);
4358 if (check_mul_overflow((unsigned long)p->len, (unsigned long)p->nbufs,
4361 if (check_add_overflow((unsigned long)p->addr, size, &tmp_check))
4364 size = (unsigned long)p->len * p->nbufs;
4365 if (!access_ok(u64_to_user_ptr(p->addr), size))
4368 p->bgid = READ_ONCE(sqe->buf_group);
4369 tmp = READ_ONCE(sqe->off);
4370 if (tmp > USHRT_MAX)
4376 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
4378 struct io_buffer *buf;
4379 u64 addr = pbuf->addr;
4380 int i, bid = pbuf->bid;
4382 for (i = 0; i < pbuf->nbufs; i++) {
4383 buf = kmalloc(sizeof(*buf), GFP_KERNEL_ACCOUNT);
4388 buf->len = min_t(__u32, pbuf->len, MAX_RW_COUNT);
4393 INIT_LIST_HEAD(&buf->list);
4396 list_add_tail(&buf->list, &(*head)->list);
4400 return i ? i : -ENOMEM;
4403 static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
4405 struct io_provide_buf *p = &req->pbuf;
4406 struct io_ring_ctx *ctx = req->ctx;
4407 struct io_buffer *head, *list;
4409 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4411 io_ring_submit_lock(ctx, !force_nonblock);
4413 lockdep_assert_held(&ctx->uring_lock);
4415 list = head = xa_load(&ctx->io_buffers, p->bgid);
4417 ret = io_add_buffers(p, &head);
4418 if (ret >= 0 && !list) {
4419 ret = xa_insert(&ctx->io_buffers, p->bgid, head, GFP_KERNEL);
4421 __io_remove_buffers(ctx, head, p->bgid, -1U);
4425 /* complete before unlock, IOPOLL may need the lock */
4426 __io_req_complete(req, issue_flags, ret, 0);
4427 io_ring_submit_unlock(ctx, !force_nonblock);
4431 static int io_epoll_ctl_prep(struct io_kiocb *req,
4432 const struct io_uring_sqe *sqe)
4434 #if defined(CONFIG_EPOLL)
4435 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
4437 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4440 req->epoll.epfd = READ_ONCE(sqe->fd);
4441 req->epoll.op = READ_ONCE(sqe->len);
4442 req->epoll.fd = READ_ONCE(sqe->off);
4444 if (ep_op_has_event(req->epoll.op)) {
4445 struct epoll_event __user *ev;
4447 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4448 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4458 static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4460 #if defined(CONFIG_EPOLL)
4461 struct io_epoll *ie = &req->epoll;
4463 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4465 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4466 if (force_nonblock && ret == -EAGAIN)
4471 __io_req_complete(req, issue_flags, ret, 0);
4478 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4480 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4481 if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->splice_fd_in)
4483 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4486 req->madvise.addr = READ_ONCE(sqe->addr);
4487 req->madvise.len = READ_ONCE(sqe->len);
4488 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4495 static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4497 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4498 struct io_madvise *ma = &req->madvise;
4501 if (issue_flags & IO_URING_F_NONBLOCK)
4504 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4507 io_req_complete(req, ret);
4514 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4516 if (sqe->ioprio || sqe->buf_index || sqe->addr || sqe->splice_fd_in)
4518 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4521 req->fadvise.offset = READ_ONCE(sqe->off);
4522 req->fadvise.len = READ_ONCE(sqe->len);
4523 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4527 static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4529 struct io_fadvise *fa = &req->fadvise;
4532 if (issue_flags & IO_URING_F_NONBLOCK) {
4533 switch (fa->advice) {
4534 case POSIX_FADV_NORMAL:
4535 case POSIX_FADV_RANDOM:
4536 case POSIX_FADV_SEQUENTIAL:
4543 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4546 __io_req_complete(req, issue_flags, ret, 0);
4550 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4552 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4554 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
4556 if (req->flags & REQ_F_FIXED_FILE)
4559 req->statx.dfd = READ_ONCE(sqe->fd);
4560 req->statx.mask = READ_ONCE(sqe->len);
4561 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4562 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4563 req->statx.flags = READ_ONCE(sqe->statx_flags);
4568 static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4570 struct io_statx *ctx = &req->statx;
4573 if (issue_flags & IO_URING_F_NONBLOCK)
4576 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4581 io_req_complete(req, ret);
4585 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4587 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4589 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4590 sqe->rw_flags || sqe->buf_index)
4592 if (req->flags & REQ_F_FIXED_FILE)
4595 req->close.fd = READ_ONCE(sqe->fd);
4596 req->close.file_slot = READ_ONCE(sqe->file_index);
4597 if (req->close.file_slot && req->close.fd)
4603 static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4605 struct files_struct *files = current->files;
4606 struct io_close *close = &req->close;
4607 struct fdtable *fdt;
4608 struct file *file = NULL;
4611 if (req->close.file_slot) {
4612 ret = io_close_fixed(req, issue_flags);
4616 spin_lock(&files->file_lock);
4617 fdt = files_fdtable(files);
4618 if (close->fd >= fdt->max_fds) {
4619 spin_unlock(&files->file_lock);
4622 file = fdt->fd[close->fd];
4623 if (!file || file->f_op == &io_uring_fops) {
4624 spin_unlock(&files->file_lock);
4629 /* if the file has a flush method, be safe and punt to async */
4630 if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4631 spin_unlock(&files->file_lock);
4635 ret = __close_fd_get_file(close->fd, &file);
4636 spin_unlock(&files->file_lock);
4643 /* No ->flush() or already async, safely close from here */
4644 ret = filp_close(file, current->files);
4650 __io_req_complete(req, issue_flags, ret, 0);
4654 static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4656 struct io_ring_ctx *ctx = req->ctx;
4658 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4660 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index ||
4664 req->sync.off = READ_ONCE(sqe->off);
4665 req->sync.len = READ_ONCE(sqe->len);
4666 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4670 static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4674 /* sync_file_range always requires a blocking context */
4675 if (issue_flags & IO_URING_F_NONBLOCK)
4678 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4682 io_req_complete(req, ret);
4686 #if defined(CONFIG_NET)
4687 static int io_setup_async_msg(struct io_kiocb *req,
4688 struct io_async_msghdr *kmsg)
4690 struct io_async_msghdr *async_msg = req->async_data;
4694 if (io_alloc_async_data(req)) {
4695 kfree(kmsg->free_iov);
4698 async_msg = req->async_data;
4699 req->flags |= REQ_F_NEED_CLEANUP;
4700 memcpy(async_msg, kmsg, sizeof(*kmsg));
4701 async_msg->msg.msg_name = &async_msg->addr;
4702 /* if were using fast_iov, set it to the new one */
4703 if (!async_msg->free_iov)
4704 async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4709 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4710 struct io_async_msghdr *iomsg)
4712 iomsg->msg.msg_name = &iomsg->addr;
4713 iomsg->free_iov = iomsg->fast_iov;
4714 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4715 req->sr_msg.msg_flags, &iomsg->free_iov);
4718 static int io_sendmsg_prep_async(struct io_kiocb *req)
4722 ret = io_sendmsg_copy_hdr(req, req->async_data);
4724 req->flags |= REQ_F_NEED_CLEANUP;
4728 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4730 struct io_sr_msg *sr = &req->sr_msg;
4732 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4735 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4736 sr->len = READ_ONCE(sqe->len);
4737 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4738 if (sr->msg_flags & MSG_DONTWAIT)
4739 req->flags |= REQ_F_NOWAIT;
4741 #ifdef CONFIG_COMPAT
4742 if (req->ctx->compat)
4743 sr->msg_flags |= MSG_CMSG_COMPAT;
4748 static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4750 struct io_async_msghdr iomsg, *kmsg;
4751 struct socket *sock;
4756 sock = sock_from_file(req->file);
4757 if (unlikely(!sock))
4760 kmsg = req->async_data;
4762 ret = io_sendmsg_copy_hdr(req, &iomsg);
4768 flags = req->sr_msg.msg_flags;
4769 if (issue_flags & IO_URING_F_NONBLOCK)
4770 flags |= MSG_DONTWAIT;
4771 if (flags & MSG_WAITALL)
4772 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4774 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4775 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4776 return io_setup_async_msg(req, kmsg);
4777 if (ret == -ERESTARTSYS)
4780 /* fast path, check for non-NULL to avoid function call */
4782 kfree(kmsg->free_iov);
4783 req->flags &= ~REQ_F_NEED_CLEANUP;
4786 __io_req_complete(req, issue_flags, ret, 0);
4790 static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4792 struct io_sr_msg *sr = &req->sr_msg;
4795 struct socket *sock;
4800 sock = sock_from_file(req->file);
4801 if (unlikely(!sock))
4804 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4808 msg.msg_name = NULL;
4809 msg.msg_control = NULL;
4810 msg.msg_controllen = 0;
4811 msg.msg_namelen = 0;
4813 flags = req->sr_msg.msg_flags;
4814 if (issue_flags & IO_URING_F_NONBLOCK)
4815 flags |= MSG_DONTWAIT;
4816 if (flags & MSG_WAITALL)
4817 min_ret = iov_iter_count(&msg.msg_iter);
4819 msg.msg_flags = flags;
4820 ret = sock_sendmsg(sock, &msg);
4821 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4823 if (ret == -ERESTARTSYS)
4828 __io_req_complete(req, issue_flags, ret, 0);
4832 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4833 struct io_async_msghdr *iomsg)
4835 struct io_sr_msg *sr = &req->sr_msg;
4836 struct iovec __user *uiov;
4840 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4841 &iomsg->uaddr, &uiov, &iov_len);
4845 if (req->flags & REQ_F_BUFFER_SELECT) {
4848 if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4850 sr->len = iomsg->fast_iov[0].iov_len;
4851 iomsg->free_iov = NULL;
4853 iomsg->free_iov = iomsg->fast_iov;
4854 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4855 &iomsg->free_iov, &iomsg->msg.msg_iter,
4864 #ifdef CONFIG_COMPAT
4865 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4866 struct io_async_msghdr *iomsg)
4868 struct io_sr_msg *sr = &req->sr_msg;
4869 struct compat_iovec __user *uiov;
4874 ret = __get_compat_msghdr(&iomsg->msg, sr->umsg_compat, &iomsg->uaddr,
4879 uiov = compat_ptr(ptr);
4880 if (req->flags & REQ_F_BUFFER_SELECT) {
4881 compat_ssize_t clen;
4885 if (!access_ok(uiov, sizeof(*uiov)))
4887 if (__get_user(clen, &uiov->iov_len))
4892 iomsg->free_iov = NULL;
4894 iomsg->free_iov = iomsg->fast_iov;
4895 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4896 UIO_FASTIOV, &iomsg->free_iov,
4897 &iomsg->msg.msg_iter, true);
4906 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4907 struct io_async_msghdr *iomsg)
4909 iomsg->msg.msg_name = &iomsg->addr;
4911 #ifdef CONFIG_COMPAT
4912 if (req->ctx->compat)
4913 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4916 return __io_recvmsg_copy_hdr(req, iomsg);
4919 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4922 struct io_sr_msg *sr = &req->sr_msg;
4923 struct io_buffer *kbuf;
4925 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4930 req->flags |= REQ_F_BUFFER_SELECTED;
4934 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4936 return io_put_kbuf(req, req->sr_msg.kbuf);
4939 static int io_recvmsg_prep_async(struct io_kiocb *req)
4943 ret = io_recvmsg_copy_hdr(req, req->async_data);
4945 req->flags |= REQ_F_NEED_CLEANUP;
4949 static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4951 struct io_sr_msg *sr = &req->sr_msg;
4953 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4956 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4957 sr->len = READ_ONCE(sqe->len);
4958 sr->bgid = READ_ONCE(sqe->buf_group);
4959 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4960 if (sr->msg_flags & MSG_DONTWAIT)
4961 req->flags |= REQ_F_NOWAIT;
4963 #ifdef CONFIG_COMPAT
4964 if (req->ctx->compat)
4965 sr->msg_flags |= MSG_CMSG_COMPAT;
4970 static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
4972 struct io_async_msghdr iomsg, *kmsg;
4973 struct socket *sock;
4974 struct io_buffer *kbuf;
4977 int ret, cflags = 0;
4978 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4980 sock = sock_from_file(req->file);
4981 if (unlikely(!sock))
4984 kmsg = req->async_data;
4986 ret = io_recvmsg_copy_hdr(req, &iomsg);
4992 if (req->flags & REQ_F_BUFFER_SELECT) {
4993 kbuf = io_recv_buffer_select(req, !force_nonblock);
4995 return PTR_ERR(kbuf);
4996 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4997 kmsg->fast_iov[0].iov_len = req->sr_msg.len;
4998 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
4999 1, req->sr_msg.len);
5002 flags = req->sr_msg.msg_flags;
5004 flags |= MSG_DONTWAIT;
5005 if (flags & MSG_WAITALL)
5006 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
5008 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
5009 kmsg->uaddr, flags);
5010 if (force_nonblock && ret == -EAGAIN)
5011 return io_setup_async_msg(req, kmsg);
5012 if (ret == -ERESTARTSYS)
5015 if (req->flags & REQ_F_BUFFER_SELECTED)
5016 cflags = io_put_recv_kbuf(req);
5017 /* fast path, check for non-NULL to avoid function call */
5019 kfree(kmsg->free_iov);
5020 req->flags &= ~REQ_F_NEED_CLEANUP;
5021 if (ret < min_ret || ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
5023 __io_req_complete(req, issue_flags, ret, cflags);
5027 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
5029 struct io_buffer *kbuf;
5030 struct io_sr_msg *sr = &req->sr_msg;
5032 void __user *buf = sr->buf;
5033 struct socket *sock;
5037 int ret, cflags = 0;
5038 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
5040 sock = sock_from_file(req->file);
5041 if (unlikely(!sock))
5044 if (req->flags & REQ_F_BUFFER_SELECT) {
5045 kbuf = io_recv_buffer_select(req, !force_nonblock);
5047 return PTR_ERR(kbuf);
5048 buf = u64_to_user_ptr(kbuf->addr);
5051 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
5055 msg.msg_name = NULL;
5056 msg.msg_control = NULL;
5057 msg.msg_controllen = 0;
5058 msg.msg_namelen = 0;
5059 msg.msg_iocb = NULL;
5062 flags = req->sr_msg.msg_flags;
5064 flags |= MSG_DONTWAIT;
5065 if (flags & MSG_WAITALL)
5066 min_ret = iov_iter_count(&msg.msg_iter);
5068 ret = sock_recvmsg(sock, &msg, flags);
5069 if (force_nonblock && ret == -EAGAIN)
5071 if (ret == -ERESTARTSYS)
5074 if (req->flags & REQ_F_BUFFER_SELECTED)
5075 cflags = io_put_recv_kbuf(req);
5076 if (ret < min_ret || ((flags & MSG_WAITALL) && (msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
5078 __io_req_complete(req, issue_flags, ret, cflags);
5082 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5084 struct io_accept *accept = &req->accept;
5086 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5088 if (sqe->ioprio || sqe->len || sqe->buf_index)
5091 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
5092 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
5093 accept->flags = READ_ONCE(sqe->accept_flags);
5094 accept->nofile = rlimit(RLIMIT_NOFILE);
5096 accept->file_slot = READ_ONCE(sqe->file_index);
5097 if (accept->file_slot && ((req->open.how.flags & O_CLOEXEC) ||
5098 (accept->flags & SOCK_CLOEXEC)))
5100 if (accept->flags & ~(SOCK_CLOEXEC | SOCK_NONBLOCK))
5102 if (SOCK_NONBLOCK != O_NONBLOCK && (accept->flags & SOCK_NONBLOCK))
5103 accept->flags = (accept->flags & ~SOCK_NONBLOCK) | O_NONBLOCK;
5107 static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
5109 struct io_accept *accept = &req->accept;
5110 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
5111 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
5112 bool fixed = !!accept->file_slot;
5116 if (req->file->f_flags & O_NONBLOCK)
5117 req->flags |= REQ_F_NOWAIT;
5120 fd = __get_unused_fd_flags(accept->flags, accept->nofile);
5121 if (unlikely(fd < 0))
5124 file = do_accept(req->file, file_flags, accept->addr, accept->addr_len,
5129 ret = PTR_ERR(file);
5130 if (ret == -EAGAIN && force_nonblock)
5132 if (ret == -ERESTARTSYS)
5135 } else if (!fixed) {
5136 fd_install(fd, file);
5139 ret = io_install_fixed_file(req, file, issue_flags,
5140 accept->file_slot - 1);
5142 __io_req_complete(req, issue_flags, ret, 0);
5146 static int io_connect_prep_async(struct io_kiocb *req)
5148 struct io_async_connect *io = req->async_data;
5149 struct io_connect *conn = &req->connect;
5151 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
5154 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5156 struct io_connect *conn = &req->connect;
5158 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5160 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags ||
5164 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
5165 conn->addr_len = READ_ONCE(sqe->addr2);
5169 static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
5171 struct io_async_connect __io, *io;
5172 unsigned file_flags;
5174 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
5176 if (req->async_data) {
5177 io = req->async_data;
5179 ret = move_addr_to_kernel(req->connect.addr,
5180 req->connect.addr_len,
5187 file_flags = force_nonblock ? O_NONBLOCK : 0;
5189 ret = __sys_connect_file(req->file, &io->address,
5190 req->connect.addr_len, file_flags);
5191 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
5192 if (req->async_data)
5194 if (io_alloc_async_data(req)) {
5198 memcpy(req->async_data, &__io, sizeof(__io));
5201 if (ret == -ERESTARTSYS)
5206 __io_req_complete(req, issue_flags, ret, 0);
5209 #else /* !CONFIG_NET */
5210 #define IO_NETOP_FN(op) \
5211 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
5213 return -EOPNOTSUPP; \
5216 #define IO_NETOP_PREP(op) \
5218 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
5220 return -EOPNOTSUPP; \
5223 #define IO_NETOP_PREP_ASYNC(op) \
5225 static int io_##op##_prep_async(struct io_kiocb *req) \
5227 return -EOPNOTSUPP; \
5230 IO_NETOP_PREP_ASYNC(sendmsg);
5231 IO_NETOP_PREP_ASYNC(recvmsg);
5232 IO_NETOP_PREP_ASYNC(connect);
5233 IO_NETOP_PREP(accept);
5236 #endif /* CONFIG_NET */
5238 struct io_poll_table {
5239 struct poll_table_struct pt;
5240 struct io_kiocb *req;
5245 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
5246 __poll_t mask, io_req_tw_func_t func)
5248 /* for instances that support it check for an event match first: */
5249 if (mask && !(mask & poll->events))
5252 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
5254 list_del_init(&poll->wait.entry);
5257 req->io_task_work.func = func;
5260 * If this fails, then the task is exiting. When a task exits, the
5261 * work gets canceled, so just cancel this request as well instead
5262 * of executing it. We can't safely execute it anyway, as we may not
5263 * have the needed state needed for it anyway.
5265 io_req_task_work_add(req);
5269 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
5270 __acquires(&req->ctx->completion_lock)
5272 struct io_ring_ctx *ctx = req->ctx;
5274 /* req->task == current here, checking PF_EXITING is safe */
5275 if (unlikely(req->task->flags & PF_EXITING))
5276 WRITE_ONCE(poll->canceled, true);
5278 if (!req->result && !READ_ONCE(poll->canceled)) {
5279 struct poll_table_struct pt = { ._key = poll->events };
5281 req->result = vfs_poll(req->file, &pt) & poll->events;
5284 spin_lock(&ctx->completion_lock);
5285 if (!req->result && !READ_ONCE(poll->canceled)) {
5286 add_wait_queue(poll->head, &poll->wait);
5293 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
5295 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
5296 if (req->opcode == IORING_OP_POLL_ADD)
5297 return req->async_data;
5298 return req->apoll->double_poll;
5301 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
5303 if (req->opcode == IORING_OP_POLL_ADD)
5305 return &req->apoll->poll;
5308 static void io_poll_remove_double(struct io_kiocb *req)
5309 __must_hold(&req->ctx->completion_lock)
5311 struct io_poll_iocb *poll = io_poll_get_double(req);
5313 lockdep_assert_held(&req->ctx->completion_lock);
5315 if (poll && poll->head) {
5316 struct wait_queue_head *head = poll->head;
5318 spin_lock_irq(&head->lock);
5319 list_del_init(&poll->wait.entry);
5320 if (poll->wait.private)
5323 spin_unlock_irq(&head->lock);
5327 static bool __io_poll_complete(struct io_kiocb *req, __poll_t mask)
5328 __must_hold(&req->ctx->completion_lock)
5330 struct io_ring_ctx *ctx = req->ctx;
5331 unsigned flags = IORING_CQE_F_MORE;
5334 if (READ_ONCE(req->poll.canceled)) {
5336 req->poll.events |= EPOLLONESHOT;
5338 error = mangle_poll(mask);
5340 if (req->poll.events & EPOLLONESHOT)
5342 if (!io_cqring_fill_event(ctx, req->user_data, error, flags)) {
5343 req->poll.events |= EPOLLONESHOT;
5346 if (flags & IORING_CQE_F_MORE)
5349 return !(flags & IORING_CQE_F_MORE);
5352 static inline bool io_poll_complete(struct io_kiocb *req, __poll_t mask)
5353 __must_hold(&req->ctx->completion_lock)
5357 done = __io_poll_complete(req, mask);
5358 io_commit_cqring(req->ctx);
5362 static void io_poll_task_func(struct io_kiocb *req, bool *locked)
5364 struct io_ring_ctx *ctx = req->ctx;
5365 struct io_kiocb *nxt;
5367 if (io_poll_rewait(req, &req->poll)) {
5368 spin_unlock(&ctx->completion_lock);
5372 if (req->poll.done) {
5373 spin_unlock(&ctx->completion_lock);
5376 done = __io_poll_complete(req, req->result);
5378 io_poll_remove_double(req);
5379 hash_del(&req->hash_node);
5380 req->poll.done = true;
5383 add_wait_queue(req->poll.head, &req->poll.wait);
5385 io_commit_cqring(ctx);
5386 spin_unlock(&ctx->completion_lock);
5387 io_cqring_ev_posted(ctx);
5390 nxt = io_put_req_find_next(req);
5392 io_req_task_submit(nxt, locked);
5397 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
5398 int sync, void *key)
5400 struct io_kiocb *req = wait->private;
5401 struct io_poll_iocb *poll = io_poll_get_single(req);
5402 __poll_t mask = key_to_poll(key);
5403 unsigned long flags;
5405 /* for instances that support it check for an event match first: */
5406 if (mask && !(mask & poll->events))
5408 if (!(poll->events & EPOLLONESHOT))
5409 return poll->wait.func(&poll->wait, mode, sync, key);
5411 list_del_init(&wait->entry);
5416 spin_lock_irqsave(&poll->head->lock, flags);
5417 done = list_empty(&poll->wait.entry);
5419 list_del_init(&poll->wait.entry);
5420 /* make sure double remove sees this as being gone */
5421 wait->private = NULL;
5422 spin_unlock_irqrestore(&poll->head->lock, flags);
5424 /* use wait func handler, so it matches the rq type */
5425 poll->wait.func(&poll->wait, mode, sync, key);
5432 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
5433 wait_queue_func_t wake_func)
5437 poll->canceled = false;
5438 #define IO_POLL_UNMASK (EPOLLERR|EPOLLHUP|EPOLLNVAL|EPOLLRDHUP)
5439 /* mask in events that we always want/need */
5440 poll->events = events | IO_POLL_UNMASK;
5441 INIT_LIST_HEAD(&poll->wait.entry);
5442 init_waitqueue_func_entry(&poll->wait, wake_func);
5445 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5446 struct wait_queue_head *head,
5447 struct io_poll_iocb **poll_ptr)
5449 struct io_kiocb *req = pt->req;
5452 * The file being polled uses multiple waitqueues for poll handling
5453 * (e.g. one for read, one for write). Setup a separate io_poll_iocb
5456 if (unlikely(pt->nr_entries)) {
5457 struct io_poll_iocb *poll_one = poll;
5459 /* double add on the same waitqueue head, ignore */
5460 if (poll_one->head == head)
5462 /* already have a 2nd entry, fail a third attempt */
5464 if ((*poll_ptr)->head == head)
5466 pt->error = -EINVAL;
5470 * Can't handle multishot for double wait for now, turn it
5471 * into one-shot mode.
5473 if (!(poll_one->events & EPOLLONESHOT))
5474 poll_one->events |= EPOLLONESHOT;
5475 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5477 pt->error = -ENOMEM;
5480 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5482 poll->wait.private = req;
5489 if (poll->events & EPOLLEXCLUSIVE)
5490 add_wait_queue_exclusive(head, &poll->wait);
5492 add_wait_queue(head, &poll->wait);
5495 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5496 struct poll_table_struct *p)
5498 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5499 struct async_poll *apoll = pt->req->apoll;
5501 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5504 static void io_async_task_func(struct io_kiocb *req, bool *locked)
5506 struct async_poll *apoll = req->apoll;
5507 struct io_ring_ctx *ctx = req->ctx;
5509 trace_io_uring_task_run(req->ctx, req, req->opcode, req->user_data);
5511 if (io_poll_rewait(req, &apoll->poll)) {
5512 spin_unlock(&ctx->completion_lock);
5516 hash_del(&req->hash_node);
5517 io_poll_remove_double(req);
5518 apoll->poll.done = true;
5519 spin_unlock(&ctx->completion_lock);
5521 if (!READ_ONCE(apoll->poll.canceled))
5522 io_req_task_submit(req, locked);
5524 io_req_complete_failed(req, -ECANCELED);
5527 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5530 struct io_kiocb *req = wait->private;
5531 struct io_poll_iocb *poll = &req->apoll->poll;
5533 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5536 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5539 static void io_poll_req_insert(struct io_kiocb *req)
5541 struct io_ring_ctx *ctx = req->ctx;
5542 struct hlist_head *list;
5544 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5545 hlist_add_head(&req->hash_node, list);
5548 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5549 struct io_poll_iocb *poll,
5550 struct io_poll_table *ipt, __poll_t mask,
5551 wait_queue_func_t wake_func)
5552 __acquires(&ctx->completion_lock)
5554 struct io_ring_ctx *ctx = req->ctx;
5555 bool cancel = false;
5557 INIT_HLIST_NODE(&req->hash_node);
5558 io_init_poll_iocb(poll, mask, wake_func);
5559 poll->file = req->file;
5560 poll->wait.private = req;
5562 ipt->pt._key = mask;
5565 ipt->nr_entries = 0;
5567 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5568 if (unlikely(!ipt->nr_entries) && !ipt->error)
5569 ipt->error = -EINVAL;
5571 spin_lock(&ctx->completion_lock);
5572 if (ipt->error || (mask && (poll->events & EPOLLONESHOT)))
5573 io_poll_remove_double(req);
5574 if (likely(poll->head)) {
5575 spin_lock_irq(&poll->head->lock);
5576 if (unlikely(list_empty(&poll->wait.entry))) {
5582 if ((mask && (poll->events & EPOLLONESHOT)) || ipt->error)
5583 list_del_init(&poll->wait.entry);
5585 WRITE_ONCE(poll->canceled, true);
5586 else if (!poll->done) /* actually waiting for an event */
5587 io_poll_req_insert(req);
5588 spin_unlock_irq(&poll->head->lock);
5600 static int io_arm_poll_handler(struct io_kiocb *req)
5602 const struct io_op_def *def = &io_op_defs[req->opcode];
5603 struct io_ring_ctx *ctx = req->ctx;
5604 struct async_poll *apoll;
5605 struct io_poll_table ipt;
5606 __poll_t ret, mask = EPOLLONESHOT | POLLERR | POLLPRI;
5609 if (!req->file || !file_can_poll(req->file))
5610 return IO_APOLL_ABORTED;
5611 if (req->flags & REQ_F_POLLED)
5612 return IO_APOLL_ABORTED;
5613 if (!def->pollin && !def->pollout)
5614 return IO_APOLL_ABORTED;
5618 mask |= POLLIN | POLLRDNORM;
5620 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5621 if ((req->opcode == IORING_OP_RECVMSG) &&
5622 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5626 mask |= POLLOUT | POLLWRNORM;
5629 /* if we can't nonblock try, then no point in arming a poll handler */
5630 if (!io_file_supports_nowait(req, rw))
5631 return IO_APOLL_ABORTED;
5633 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5634 if (unlikely(!apoll))
5635 return IO_APOLL_ABORTED;
5636 apoll->double_poll = NULL;
5638 req->flags |= REQ_F_POLLED;
5639 ipt.pt._qproc = io_async_queue_proc;
5640 io_req_set_refcount(req);
5642 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5644 spin_unlock(&ctx->completion_lock);
5645 if (ret || ipt.error)
5646 return ret ? IO_APOLL_READY : IO_APOLL_ABORTED;
5648 trace_io_uring_poll_arm(ctx, req, req->opcode, req->user_data,
5649 mask, apoll->poll.events);
5653 static bool __io_poll_remove_one(struct io_kiocb *req,
5654 struct io_poll_iocb *poll, bool do_cancel)
5655 __must_hold(&req->ctx->completion_lock)
5657 bool do_complete = false;
5661 spin_lock_irq(&poll->head->lock);
5663 WRITE_ONCE(poll->canceled, true);
5664 if (!list_empty(&poll->wait.entry)) {
5665 list_del_init(&poll->wait.entry);
5668 spin_unlock_irq(&poll->head->lock);
5669 hash_del(&req->hash_node);
5673 static bool io_poll_remove_one(struct io_kiocb *req)
5674 __must_hold(&req->ctx->completion_lock)
5678 io_poll_remove_double(req);
5679 do_complete = __io_poll_remove_one(req, io_poll_get_single(req), true);
5682 io_cqring_fill_event(req->ctx, req->user_data, -ECANCELED, 0);
5683 io_commit_cqring(req->ctx);
5685 io_put_req_deferred(req);
5691 * Returns true if we found and killed one or more poll requests
5693 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5696 struct hlist_node *tmp;
5697 struct io_kiocb *req;
5700 spin_lock(&ctx->completion_lock);
5701 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5702 struct hlist_head *list;
5704 list = &ctx->cancel_hash[i];
5705 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5706 if (io_match_task(req, tsk, cancel_all))
5707 posted += io_poll_remove_one(req);
5710 spin_unlock(&ctx->completion_lock);
5713 io_cqring_ev_posted(ctx);
5718 static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, __u64 sqe_addr,
5720 __must_hold(&ctx->completion_lock)
5722 struct hlist_head *list;
5723 struct io_kiocb *req;
5725 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5726 hlist_for_each_entry(req, list, hash_node) {
5727 if (sqe_addr != req->user_data)
5729 if (poll_only && req->opcode != IORING_OP_POLL_ADD)
5736 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr,
5738 __must_hold(&ctx->completion_lock)
5740 struct io_kiocb *req;
5742 req = io_poll_find(ctx, sqe_addr, poll_only);
5745 if (io_poll_remove_one(req))
5751 static __poll_t io_poll_parse_events(const struct io_uring_sqe *sqe,
5756 events = READ_ONCE(sqe->poll32_events);
5758 events = swahw32(events);
5760 if (!(flags & IORING_POLL_ADD_MULTI))
5761 events |= EPOLLONESHOT;
5762 return demangle_poll(events) | (events & (EPOLLEXCLUSIVE|EPOLLONESHOT));
5765 static int io_poll_update_prep(struct io_kiocb *req,
5766 const struct io_uring_sqe *sqe)
5768 struct io_poll_update *upd = &req->poll_update;
5771 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5773 if (sqe->ioprio || sqe->buf_index || sqe->splice_fd_in)
5775 flags = READ_ONCE(sqe->len);
5776 if (flags & ~(IORING_POLL_UPDATE_EVENTS | IORING_POLL_UPDATE_USER_DATA |
5777 IORING_POLL_ADD_MULTI))
5779 /* meaningless without update */
5780 if (flags == IORING_POLL_ADD_MULTI)
5783 upd->old_user_data = READ_ONCE(sqe->addr);
5784 upd->update_events = flags & IORING_POLL_UPDATE_EVENTS;
5785 upd->update_user_data = flags & IORING_POLL_UPDATE_USER_DATA;
5787 upd->new_user_data = READ_ONCE(sqe->off);
5788 if (!upd->update_user_data && upd->new_user_data)
5790 if (upd->update_events)
5791 upd->events = io_poll_parse_events(sqe, flags);
5792 else if (sqe->poll32_events)
5798 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5801 struct io_kiocb *req = wait->private;
5802 struct io_poll_iocb *poll = &req->poll;
5804 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5807 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5808 struct poll_table_struct *p)
5810 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5812 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5815 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5817 struct io_poll_iocb *poll = &req->poll;
5820 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5822 if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->addr)
5824 flags = READ_ONCE(sqe->len);
5825 if (flags & ~IORING_POLL_ADD_MULTI)
5828 io_req_set_refcount(req);
5829 poll->events = io_poll_parse_events(sqe, flags);
5833 static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5835 struct io_poll_iocb *poll = &req->poll;
5836 struct io_ring_ctx *ctx = req->ctx;
5837 struct io_poll_table ipt;
5841 ipt.pt._qproc = io_poll_queue_proc;
5843 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5846 if (mask) { /* no async, we'd stolen it */
5848 done = io_poll_complete(req, mask);
5850 spin_unlock(&ctx->completion_lock);
5853 io_cqring_ev_posted(ctx);
5860 static int io_poll_update(struct io_kiocb *req, unsigned int issue_flags)
5862 struct io_ring_ctx *ctx = req->ctx;
5863 struct io_kiocb *preq;
5867 spin_lock(&ctx->completion_lock);
5868 preq = io_poll_find(ctx, req->poll_update.old_user_data, true);
5874 if (!req->poll_update.update_events && !req->poll_update.update_user_data) {
5876 ret = io_poll_remove_one(preq) ? 0 : -EALREADY;
5881 * Don't allow racy completion with singleshot, as we cannot safely
5882 * update those. For multishot, if we're racing with completion, just
5883 * let completion re-add it.
5885 completing = !__io_poll_remove_one(preq, &preq->poll, false);
5886 if (completing && (preq->poll.events & EPOLLONESHOT)) {
5890 /* we now have a detached poll request. reissue. */
5894 spin_unlock(&ctx->completion_lock);
5896 io_req_complete(req, ret);
5899 /* only mask one event flags, keep behavior flags */
5900 if (req->poll_update.update_events) {
5901 preq->poll.events &= ~0xffff;
5902 preq->poll.events |= req->poll_update.events & 0xffff;
5903 preq->poll.events |= IO_POLL_UNMASK;
5905 if (req->poll_update.update_user_data)
5906 preq->user_data = req->poll_update.new_user_data;
5907 spin_unlock(&ctx->completion_lock);
5909 /* complete update request, we're done with it */
5910 io_req_complete(req, ret);
5913 ret = io_poll_add(preq, issue_flags);
5916 io_req_complete(preq, ret);
5922 static void io_req_task_timeout(struct io_kiocb *req, bool *locked)
5925 io_req_complete_post(req, -ETIME, 0);
5928 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5930 struct io_timeout_data *data = container_of(timer,
5931 struct io_timeout_data, timer);
5932 struct io_kiocb *req = data->req;
5933 struct io_ring_ctx *ctx = req->ctx;
5934 unsigned long flags;
5936 spin_lock_irqsave(&ctx->timeout_lock, flags);
5937 list_del_init(&req->timeout.list);
5938 atomic_set(&req->ctx->cq_timeouts,
5939 atomic_read(&req->ctx->cq_timeouts) + 1);
5940 spin_unlock_irqrestore(&ctx->timeout_lock, flags);
5942 req->io_task_work.func = io_req_task_timeout;
5943 io_req_task_work_add(req);
5944 return HRTIMER_NORESTART;
5947 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5949 __must_hold(&ctx->timeout_lock)
5951 struct io_timeout_data *io;
5952 struct io_kiocb *req;
5955 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5956 found = user_data == req->user_data;
5961 return ERR_PTR(-ENOENT);
5963 io = req->async_data;
5964 if (hrtimer_try_to_cancel(&io->timer) == -1)
5965 return ERR_PTR(-EALREADY);
5966 list_del_init(&req->timeout.list);
5970 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5971 __must_hold(&ctx->completion_lock)
5972 __must_hold(&ctx->timeout_lock)
5974 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5977 return PTR_ERR(req);
5980 io_cqring_fill_event(ctx, req->user_data, -ECANCELED, 0);
5981 io_put_req_deferred(req);
5985 static clockid_t io_timeout_get_clock(struct io_timeout_data *data)
5987 switch (data->flags & IORING_TIMEOUT_CLOCK_MASK) {
5988 case IORING_TIMEOUT_BOOTTIME:
5989 return CLOCK_BOOTTIME;
5990 case IORING_TIMEOUT_REALTIME:
5991 return CLOCK_REALTIME;
5993 /* can't happen, vetted at prep time */
5997 return CLOCK_MONOTONIC;
6001 static int io_linked_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
6002 struct timespec64 *ts, enum hrtimer_mode mode)
6003 __must_hold(&ctx->timeout_lock)
6005 struct io_timeout_data *io;
6006 struct io_kiocb *req;
6009 list_for_each_entry(req, &ctx->ltimeout_list, timeout.list) {
6010 found = user_data == req->user_data;
6017 io = req->async_data;
6018 if (hrtimer_try_to_cancel(&io->timer) == -1)
6020 hrtimer_init(&io->timer, io_timeout_get_clock(io), mode);
6021 io->timer.function = io_link_timeout_fn;
6022 hrtimer_start(&io->timer, timespec64_to_ktime(*ts), mode);
6026 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
6027 struct timespec64 *ts, enum hrtimer_mode mode)
6028 __must_hold(&ctx->timeout_lock)
6030 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
6031 struct io_timeout_data *data;
6034 return PTR_ERR(req);
6036 req->timeout.off = 0; /* noseq */
6037 data = req->async_data;
6038 list_add_tail(&req->timeout.list, &ctx->timeout_list);
6039 hrtimer_init(&data->timer, io_timeout_get_clock(data), mode);
6040 data->timer.function = io_timeout_fn;
6041 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
6045 static int io_timeout_remove_prep(struct io_kiocb *req,
6046 const struct io_uring_sqe *sqe)
6048 struct io_timeout_rem *tr = &req->timeout_rem;
6050 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
6052 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
6054 if (sqe->ioprio || sqe->buf_index || sqe->len || sqe->splice_fd_in)
6057 tr->ltimeout = false;
6058 tr->addr = READ_ONCE(sqe->addr);
6059 tr->flags = READ_ONCE(sqe->timeout_flags);
6060 if (tr->flags & IORING_TIMEOUT_UPDATE_MASK) {
6061 if (hweight32(tr->flags & IORING_TIMEOUT_CLOCK_MASK) > 1)
6063 if (tr->flags & IORING_LINK_TIMEOUT_UPDATE)
6064 tr->ltimeout = true;
6065 if (tr->flags & ~(IORING_TIMEOUT_UPDATE_MASK|IORING_TIMEOUT_ABS))
6067 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
6069 } else if (tr->flags) {
6070 /* timeout removal doesn't support flags */
6077 static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
6079 return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
6084 * Remove or update an existing timeout command
6086 static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
6088 struct io_timeout_rem *tr = &req->timeout_rem;
6089 struct io_ring_ctx *ctx = req->ctx;
6092 if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE)) {
6093 spin_lock(&ctx->completion_lock);
6094 spin_lock_irq(&ctx->timeout_lock);
6095 ret = io_timeout_cancel(ctx, tr->addr);
6096 spin_unlock_irq(&ctx->timeout_lock);
6097 spin_unlock(&ctx->completion_lock);
6099 enum hrtimer_mode mode = io_translate_timeout_mode(tr->flags);
6101 spin_lock_irq(&ctx->timeout_lock);
6103 ret = io_linked_timeout_update(ctx, tr->addr, &tr->ts, mode);
6105 ret = io_timeout_update(ctx, tr->addr, &tr->ts, mode);
6106 spin_unlock_irq(&ctx->timeout_lock);
6111 io_req_complete_post(req, ret, 0);
6115 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6116 bool is_timeout_link)
6118 struct io_timeout_data *data;
6120 u32 off = READ_ONCE(sqe->off);
6122 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
6124 if (sqe->ioprio || sqe->buf_index || sqe->len != 1 ||
6127 if (off && is_timeout_link)
6129 flags = READ_ONCE(sqe->timeout_flags);
6130 if (flags & ~(IORING_TIMEOUT_ABS | IORING_TIMEOUT_CLOCK_MASK))
6132 /* more than one clock specified is invalid, obviously */
6133 if (hweight32(flags & IORING_TIMEOUT_CLOCK_MASK) > 1)
6136 INIT_LIST_HEAD(&req->timeout.list);
6137 req->timeout.off = off;
6138 if (unlikely(off && !req->ctx->off_timeout_used))
6139 req->ctx->off_timeout_used = true;
6141 if (!req->async_data && io_alloc_async_data(req))
6144 data = req->async_data;
6146 data->flags = flags;
6148 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
6151 data->mode = io_translate_timeout_mode(flags);
6152 hrtimer_init(&data->timer, io_timeout_get_clock(data), data->mode);
6154 if (is_timeout_link) {
6155 struct io_submit_link *link = &req->ctx->submit_state.link;
6159 if (link->last->opcode == IORING_OP_LINK_TIMEOUT)
6161 req->timeout.head = link->last;
6162 link->last->flags |= REQ_F_ARM_LTIMEOUT;
6167 static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
6169 struct io_ring_ctx *ctx = req->ctx;
6170 struct io_timeout_data *data = req->async_data;
6171 struct list_head *entry;
6172 u32 tail, off = req->timeout.off;
6174 spin_lock_irq(&ctx->timeout_lock);
6177 * sqe->off holds how many events that need to occur for this
6178 * timeout event to be satisfied. If it isn't set, then this is
6179 * a pure timeout request, sequence isn't used.
6181 if (io_is_timeout_noseq(req)) {
6182 entry = ctx->timeout_list.prev;
6186 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
6187 req->timeout.target_seq = tail + off;
6189 /* Update the last seq here in case io_flush_timeouts() hasn't.
6190 * This is safe because ->completion_lock is held, and submissions
6191 * and completions are never mixed in the same ->completion_lock section.
6193 ctx->cq_last_tm_flush = tail;
6196 * Insertion sort, ensuring the first entry in the list is always
6197 * the one we need first.
6199 list_for_each_prev(entry, &ctx->timeout_list) {
6200 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
6203 if (io_is_timeout_noseq(nxt))
6205 /* nxt.seq is behind @tail, otherwise would've been completed */
6206 if (off >= nxt->timeout.target_seq - tail)
6210 list_add(&req->timeout.list, entry);
6211 data->timer.function = io_timeout_fn;
6212 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
6213 spin_unlock_irq(&ctx->timeout_lock);
6217 struct io_cancel_data {
6218 struct io_ring_ctx *ctx;
6222 static bool io_cancel_cb(struct io_wq_work *work, void *data)
6224 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6225 struct io_cancel_data *cd = data;
6227 return req->ctx == cd->ctx && req->user_data == cd->user_data;
6230 static int io_async_cancel_one(struct io_uring_task *tctx, u64 user_data,
6231 struct io_ring_ctx *ctx)
6233 struct io_cancel_data data = { .ctx = ctx, .user_data = user_data, };
6234 enum io_wq_cancel cancel_ret;
6237 if (!tctx || !tctx->io_wq)
6240 cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, &data, false);
6241 switch (cancel_ret) {
6242 case IO_WQ_CANCEL_OK:
6245 case IO_WQ_CANCEL_RUNNING:
6248 case IO_WQ_CANCEL_NOTFOUND:
6256 static int io_try_cancel_userdata(struct io_kiocb *req, u64 sqe_addr)
6258 struct io_ring_ctx *ctx = req->ctx;
6261 WARN_ON_ONCE(!io_wq_current_is_worker() && req->task != current);
6263 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
6267 spin_lock(&ctx->completion_lock);
6268 spin_lock_irq(&ctx->timeout_lock);
6269 ret = io_timeout_cancel(ctx, sqe_addr);
6270 spin_unlock_irq(&ctx->timeout_lock);
6273 ret = io_poll_cancel(ctx, sqe_addr, false);
6275 spin_unlock(&ctx->completion_lock);
6279 static int io_async_cancel_prep(struct io_kiocb *req,
6280 const struct io_uring_sqe *sqe)
6282 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
6284 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
6286 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags ||
6290 req->cancel.addr = READ_ONCE(sqe->addr);
6294 static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
6296 struct io_ring_ctx *ctx = req->ctx;
6297 u64 sqe_addr = req->cancel.addr;
6298 struct io_tctx_node *node;
6301 ret = io_try_cancel_userdata(req, sqe_addr);
6305 /* slow path, try all io-wq's */
6306 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
6308 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
6309 struct io_uring_task *tctx = node->task->io_uring;
6311 ret = io_async_cancel_one(tctx, req->cancel.addr, ctx);
6315 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
6319 io_req_complete_post(req, ret, 0);
6323 static int io_rsrc_update_prep(struct io_kiocb *req,
6324 const struct io_uring_sqe *sqe)
6326 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
6328 if (sqe->ioprio || sqe->rw_flags || sqe->splice_fd_in)
6331 req->rsrc_update.offset = READ_ONCE(sqe->off);
6332 req->rsrc_update.nr_args = READ_ONCE(sqe->len);
6333 if (!req->rsrc_update.nr_args)
6335 req->rsrc_update.arg = READ_ONCE(sqe->addr);
6339 static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
6341 struct io_ring_ctx *ctx = req->ctx;
6342 struct io_uring_rsrc_update2 up;
6345 up.offset = req->rsrc_update.offset;
6346 up.data = req->rsrc_update.arg;
6351 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
6352 ret = __io_register_rsrc_update(ctx, IORING_RSRC_FILE,
6353 &up, req->rsrc_update.nr_args);
6354 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
6358 __io_req_complete(req, issue_flags, ret, 0);
6362 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
6364 switch (req->opcode) {
6367 case IORING_OP_READV:
6368 case IORING_OP_READ_FIXED:
6369 case IORING_OP_READ:
6370 return io_read_prep(req, sqe);
6371 case IORING_OP_WRITEV:
6372 case IORING_OP_WRITE_FIXED:
6373 case IORING_OP_WRITE:
6374 return io_write_prep(req, sqe);
6375 case IORING_OP_POLL_ADD:
6376 return io_poll_add_prep(req, sqe);
6377 case IORING_OP_POLL_REMOVE:
6378 return io_poll_update_prep(req, sqe);
6379 case IORING_OP_FSYNC:
6380 return io_fsync_prep(req, sqe);
6381 case IORING_OP_SYNC_FILE_RANGE:
6382 return io_sfr_prep(req, sqe);
6383 case IORING_OP_SENDMSG:
6384 case IORING_OP_SEND:
6385 return io_sendmsg_prep(req, sqe);
6386 case IORING_OP_RECVMSG:
6387 case IORING_OP_RECV:
6388 return io_recvmsg_prep(req, sqe);
6389 case IORING_OP_CONNECT:
6390 return io_connect_prep(req, sqe);
6391 case IORING_OP_TIMEOUT:
6392 return io_timeout_prep(req, sqe, false);
6393 case IORING_OP_TIMEOUT_REMOVE:
6394 return io_timeout_remove_prep(req, sqe);
6395 case IORING_OP_ASYNC_CANCEL:
6396 return io_async_cancel_prep(req, sqe);
6397 case IORING_OP_LINK_TIMEOUT:
6398 return io_timeout_prep(req, sqe, true);
6399 case IORING_OP_ACCEPT:
6400 return io_accept_prep(req, sqe);
6401 case IORING_OP_FALLOCATE:
6402 return io_fallocate_prep(req, sqe);
6403 case IORING_OP_OPENAT:
6404 return io_openat_prep(req, sqe);
6405 case IORING_OP_CLOSE:
6406 return io_close_prep(req, sqe);
6407 case IORING_OP_FILES_UPDATE:
6408 return io_rsrc_update_prep(req, sqe);
6409 case IORING_OP_STATX:
6410 return io_statx_prep(req, sqe);
6411 case IORING_OP_FADVISE:
6412 return io_fadvise_prep(req, sqe);
6413 case IORING_OP_MADVISE:
6414 return io_madvise_prep(req, sqe);
6415 case IORING_OP_OPENAT2:
6416 return io_openat2_prep(req, sqe);
6417 case IORING_OP_EPOLL_CTL:
6418 return io_epoll_ctl_prep(req, sqe);
6419 case IORING_OP_SPLICE:
6420 return io_splice_prep(req, sqe);
6421 case IORING_OP_PROVIDE_BUFFERS:
6422 return io_provide_buffers_prep(req, sqe);
6423 case IORING_OP_REMOVE_BUFFERS:
6424 return io_remove_buffers_prep(req, sqe);
6426 return io_tee_prep(req, sqe);
6427 case IORING_OP_SHUTDOWN:
6428 return io_shutdown_prep(req, sqe);
6429 case IORING_OP_RENAMEAT:
6430 return io_renameat_prep(req, sqe);
6431 case IORING_OP_UNLINKAT:
6432 return io_unlinkat_prep(req, sqe);
6433 case IORING_OP_MKDIRAT:
6434 return io_mkdirat_prep(req, sqe);
6435 case IORING_OP_SYMLINKAT:
6436 return io_symlinkat_prep(req, sqe);
6437 case IORING_OP_LINKAT:
6438 return io_linkat_prep(req, sqe);
6441 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
6446 static int io_req_prep_async(struct io_kiocb *req)
6448 if (!io_op_defs[req->opcode].needs_async_setup)
6450 if (WARN_ON_ONCE(req->async_data))
6452 if (io_alloc_async_data(req))
6455 switch (req->opcode) {
6456 case IORING_OP_READV:
6457 return io_rw_prep_async(req, READ);
6458 case IORING_OP_WRITEV:
6459 return io_rw_prep_async(req, WRITE);
6460 case IORING_OP_SENDMSG:
6461 return io_sendmsg_prep_async(req);
6462 case IORING_OP_RECVMSG:
6463 return io_recvmsg_prep_async(req);
6464 case IORING_OP_CONNECT:
6465 return io_connect_prep_async(req);
6467 printk_once(KERN_WARNING "io_uring: prep_async() bad opcode %d\n",
6472 static u32 io_get_sequence(struct io_kiocb *req)
6474 u32 seq = req->ctx->cached_sq_head;
6476 /* need original cached_sq_head, but it was increased for each req */
6477 io_for_each_link(req, req)
6482 static bool io_drain_req(struct io_kiocb *req)
6484 struct io_kiocb *pos;
6485 struct io_ring_ctx *ctx = req->ctx;
6486 struct io_defer_entry *de;
6490 if (req->flags & REQ_F_FAIL) {
6491 io_req_complete_fail_submit(req);
6496 * If we need to drain a request in the middle of a link, drain the
6497 * head request and the next request/link after the current link.
6498 * Considering sequential execution of links, IOSQE_IO_DRAIN will be
6499 * maintained for every request of our link.
6501 if (ctx->drain_next) {
6502 req->flags |= REQ_F_IO_DRAIN;
6503 ctx->drain_next = false;
6505 /* not interested in head, start from the first linked */
6506 io_for_each_link(pos, req->link) {
6507 if (pos->flags & REQ_F_IO_DRAIN) {
6508 ctx->drain_next = true;
6509 req->flags |= REQ_F_IO_DRAIN;
6514 /* Still need defer if there is pending req in defer list. */
6515 if (likely(list_empty_careful(&ctx->defer_list) &&
6516 !(req->flags & REQ_F_IO_DRAIN))) {
6517 ctx->drain_active = false;
6521 seq = io_get_sequence(req);
6522 /* Still a chance to pass the sequence check */
6523 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
6526 ret = io_req_prep_async(req);
6529 io_prep_async_link(req);
6530 de = kmalloc(sizeof(*de), GFP_KERNEL);
6534 io_req_complete_failed(req, ret);
6538 spin_lock(&ctx->completion_lock);
6539 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
6540 spin_unlock(&ctx->completion_lock);
6542 io_queue_async_work(req, NULL);
6546 trace_io_uring_defer(ctx, req, req->user_data);
6549 list_add_tail(&de->list, &ctx->defer_list);
6550 spin_unlock(&ctx->completion_lock);
6554 static void io_clean_op(struct io_kiocb *req)
6556 if (req->flags & REQ_F_BUFFER_SELECTED) {
6557 switch (req->opcode) {
6558 case IORING_OP_READV:
6559 case IORING_OP_READ_FIXED:
6560 case IORING_OP_READ:
6561 kfree((void *)(unsigned long)req->rw.addr);
6563 case IORING_OP_RECVMSG:
6564 case IORING_OP_RECV:
6565 kfree(req->sr_msg.kbuf);
6570 if (req->flags & REQ_F_NEED_CLEANUP) {
6571 switch (req->opcode) {
6572 case IORING_OP_READV:
6573 case IORING_OP_READ_FIXED:
6574 case IORING_OP_READ:
6575 case IORING_OP_WRITEV:
6576 case IORING_OP_WRITE_FIXED:
6577 case IORING_OP_WRITE: {
6578 struct io_async_rw *io = req->async_data;
6580 kfree(io->free_iovec);
6583 case IORING_OP_RECVMSG:
6584 case IORING_OP_SENDMSG: {
6585 struct io_async_msghdr *io = req->async_data;
6587 kfree(io->free_iov);
6590 case IORING_OP_SPLICE:
6592 if (!(req->splice.flags & SPLICE_F_FD_IN_FIXED))
6593 io_put_file(req->splice.file_in);
6595 case IORING_OP_OPENAT:
6596 case IORING_OP_OPENAT2:
6597 if (req->open.filename)
6598 putname(req->open.filename);
6600 case IORING_OP_RENAMEAT:
6601 putname(req->rename.oldpath);
6602 putname(req->rename.newpath);
6604 case IORING_OP_UNLINKAT:
6605 putname(req->unlink.filename);
6607 case IORING_OP_MKDIRAT:
6608 putname(req->mkdir.filename);
6610 case IORING_OP_SYMLINKAT:
6611 putname(req->symlink.oldpath);
6612 putname(req->symlink.newpath);
6614 case IORING_OP_LINKAT:
6615 putname(req->hardlink.oldpath);
6616 putname(req->hardlink.newpath);
6620 if ((req->flags & REQ_F_POLLED) && req->apoll) {
6621 kfree(req->apoll->double_poll);
6625 if (req->flags & REQ_F_INFLIGHT) {
6626 struct io_uring_task *tctx = req->task->io_uring;
6628 atomic_dec(&tctx->inflight_tracked);
6630 if (req->flags & REQ_F_CREDS)
6631 put_cred(req->creds);
6633 req->flags &= ~IO_REQ_CLEAN_FLAGS;
6636 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
6638 struct io_ring_ctx *ctx = req->ctx;
6639 const struct cred *creds = NULL;
6642 if ((req->flags & REQ_F_CREDS) && req->creds != current_cred())
6643 creds = override_creds(req->creds);
6645 switch (req->opcode) {
6647 ret = io_nop(req, issue_flags);
6649 case IORING_OP_READV:
6650 case IORING_OP_READ_FIXED:
6651 case IORING_OP_READ:
6652 ret = io_read(req, issue_flags);
6654 case IORING_OP_WRITEV:
6655 case IORING_OP_WRITE_FIXED:
6656 case IORING_OP_WRITE:
6657 ret = io_write(req, issue_flags);
6659 case IORING_OP_FSYNC:
6660 ret = io_fsync(req, issue_flags);
6662 case IORING_OP_POLL_ADD:
6663 ret = io_poll_add(req, issue_flags);
6665 case IORING_OP_POLL_REMOVE:
6666 ret = io_poll_update(req, issue_flags);
6668 case IORING_OP_SYNC_FILE_RANGE:
6669 ret = io_sync_file_range(req, issue_flags);
6671 case IORING_OP_SENDMSG:
6672 ret = io_sendmsg(req, issue_flags);
6674 case IORING_OP_SEND:
6675 ret = io_send(req, issue_flags);
6677 case IORING_OP_RECVMSG:
6678 ret = io_recvmsg(req, issue_flags);
6680 case IORING_OP_RECV:
6681 ret = io_recv(req, issue_flags);
6683 case IORING_OP_TIMEOUT:
6684 ret = io_timeout(req, issue_flags);
6686 case IORING_OP_TIMEOUT_REMOVE:
6687 ret = io_timeout_remove(req, issue_flags);
6689 case IORING_OP_ACCEPT:
6690 ret = io_accept(req, issue_flags);
6692 case IORING_OP_CONNECT:
6693 ret = io_connect(req, issue_flags);
6695 case IORING_OP_ASYNC_CANCEL:
6696 ret = io_async_cancel(req, issue_flags);
6698 case IORING_OP_FALLOCATE:
6699 ret = io_fallocate(req, issue_flags);
6701 case IORING_OP_OPENAT:
6702 ret = io_openat(req, issue_flags);
6704 case IORING_OP_CLOSE:
6705 ret = io_close(req, issue_flags);
6707 case IORING_OP_FILES_UPDATE:
6708 ret = io_files_update(req, issue_flags);
6710 case IORING_OP_STATX:
6711 ret = io_statx(req, issue_flags);
6713 case IORING_OP_FADVISE:
6714 ret = io_fadvise(req, issue_flags);
6716 case IORING_OP_MADVISE:
6717 ret = io_madvise(req, issue_flags);
6719 case IORING_OP_OPENAT2:
6720 ret = io_openat2(req, issue_flags);
6722 case IORING_OP_EPOLL_CTL:
6723 ret = io_epoll_ctl(req, issue_flags);
6725 case IORING_OP_SPLICE:
6726 ret = io_splice(req, issue_flags);
6728 case IORING_OP_PROVIDE_BUFFERS:
6729 ret = io_provide_buffers(req, issue_flags);
6731 case IORING_OP_REMOVE_BUFFERS:
6732 ret = io_remove_buffers(req, issue_flags);
6735 ret = io_tee(req, issue_flags);
6737 case IORING_OP_SHUTDOWN:
6738 ret = io_shutdown(req, issue_flags);
6740 case IORING_OP_RENAMEAT:
6741 ret = io_renameat(req, issue_flags);
6743 case IORING_OP_UNLINKAT:
6744 ret = io_unlinkat(req, issue_flags);
6746 case IORING_OP_MKDIRAT:
6747 ret = io_mkdirat(req, issue_flags);
6749 case IORING_OP_SYMLINKAT:
6750 ret = io_symlinkat(req, issue_flags);
6752 case IORING_OP_LINKAT:
6753 ret = io_linkat(req, issue_flags);
6761 revert_creds(creds);
6764 /* If the op doesn't have a file, we're not polling for it */
6765 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file)
6766 io_iopoll_req_issued(req);
6771 static struct io_wq_work *io_wq_free_work(struct io_wq_work *work)
6773 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6775 req = io_put_req_find_next(req);
6776 return req ? &req->work : NULL;
6779 static void io_wq_submit_work(struct io_wq_work *work)
6781 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6782 struct io_kiocb *timeout;
6785 /* one will be dropped by ->io_free_work() after returning to io-wq */
6786 if (!(req->flags & REQ_F_REFCOUNT))
6787 __io_req_set_refcount(req, 2);
6791 timeout = io_prep_linked_timeout(req);
6793 io_queue_linked_timeout(timeout);
6795 /* either cancelled or io-wq is dying, so don't touch tctx->iowq */
6796 if (work->flags & IO_WQ_WORK_CANCEL)
6801 ret = io_issue_sqe(req, 0);
6803 * We can get EAGAIN for polled IO even though we're
6804 * forcing a sync submission from here, since we can't
6805 * wait for request slots on the block side.
6813 /* avoid locking problems by failing it from a clean context */
6815 io_req_task_queue_fail(req, ret);
6818 static inline struct io_fixed_file *io_fixed_file_slot(struct io_file_table *table,
6821 return &table->files[i];
6824 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6827 struct io_fixed_file *slot = io_fixed_file_slot(&ctx->file_table, index);
6829 return (struct file *) (slot->file_ptr & FFS_MASK);
6832 static void io_fixed_file_set(struct io_fixed_file *file_slot, struct file *file)
6834 unsigned long file_ptr = (unsigned long) file;
6836 if (__io_file_supports_nowait(file, READ))
6837 file_ptr |= FFS_ASYNC_READ;
6838 if (__io_file_supports_nowait(file, WRITE))
6839 file_ptr |= FFS_ASYNC_WRITE;
6840 if (S_ISREG(file_inode(file)->i_mode))
6841 file_ptr |= FFS_ISREG;
6842 file_slot->file_ptr = file_ptr;
6845 static inline struct file *io_file_get_fixed(struct io_ring_ctx *ctx,
6846 struct io_kiocb *req, int fd)
6849 unsigned long file_ptr;
6851 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6853 fd = array_index_nospec(fd, ctx->nr_user_files);
6854 file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr;
6855 file = (struct file *) (file_ptr & FFS_MASK);
6856 file_ptr &= ~FFS_MASK;
6857 /* mask in overlapping REQ_F and FFS bits */
6858 req->flags |= (file_ptr << REQ_F_NOWAIT_READ_BIT);
6859 io_req_set_rsrc_node(req);
6863 static struct file *io_file_get_normal(struct io_ring_ctx *ctx,
6864 struct io_kiocb *req, int fd)
6866 struct file *file = fget(fd);
6868 trace_io_uring_file_get(ctx, fd);
6870 /* we don't allow fixed io_uring files */
6871 if (file && unlikely(file->f_op == &io_uring_fops))
6872 io_req_track_inflight(req);
6876 static inline struct file *io_file_get(struct io_ring_ctx *ctx,
6877 struct io_kiocb *req, int fd, bool fixed)
6880 return io_file_get_fixed(ctx, req, fd);
6882 return io_file_get_normal(ctx, req, fd);
6885 static void io_req_task_link_timeout(struct io_kiocb *req, bool *locked)
6887 struct io_kiocb *prev = req->timeout.prev;
6891 ret = io_try_cancel_userdata(req, prev->user_data);
6892 io_req_complete_post(req, ret ?: -ETIME, 0);
6895 io_req_complete_post(req, -ETIME, 0);
6899 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6901 struct io_timeout_data *data = container_of(timer,
6902 struct io_timeout_data, timer);
6903 struct io_kiocb *prev, *req = data->req;
6904 struct io_ring_ctx *ctx = req->ctx;
6905 unsigned long flags;
6907 spin_lock_irqsave(&ctx->timeout_lock, flags);
6908 prev = req->timeout.head;
6909 req->timeout.head = NULL;
6912 * We don't expect the list to be empty, that will only happen if we
6913 * race with the completion of the linked work.
6916 io_remove_next_linked(prev);
6917 if (!req_ref_inc_not_zero(prev))
6920 list_del(&req->timeout.list);
6921 req->timeout.prev = prev;
6922 spin_unlock_irqrestore(&ctx->timeout_lock, flags);
6924 req->io_task_work.func = io_req_task_link_timeout;
6925 io_req_task_work_add(req);
6926 return HRTIMER_NORESTART;
6929 static void io_queue_linked_timeout(struct io_kiocb *req)
6931 struct io_ring_ctx *ctx = req->ctx;
6933 spin_lock_irq(&ctx->timeout_lock);
6935 * If the back reference is NULL, then our linked request finished
6936 * before we got a chance to setup the timer
6938 if (req->timeout.head) {
6939 struct io_timeout_data *data = req->async_data;
6941 data->timer.function = io_link_timeout_fn;
6942 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6944 list_add_tail(&req->timeout.list, &ctx->ltimeout_list);
6946 spin_unlock_irq(&ctx->timeout_lock);
6947 /* drop submission reference */
6951 static void __io_queue_sqe(struct io_kiocb *req)
6952 __must_hold(&req->ctx->uring_lock)
6954 struct io_kiocb *linked_timeout;
6958 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
6961 * We async punt it if the file wasn't marked NOWAIT, or if the file
6962 * doesn't support non-blocking read/write attempts
6965 if (req->flags & REQ_F_COMPLETE_INLINE) {
6966 struct io_ring_ctx *ctx = req->ctx;
6967 struct io_submit_state *state = &ctx->submit_state;
6969 state->compl_reqs[state->compl_nr++] = req;
6970 if (state->compl_nr == ARRAY_SIZE(state->compl_reqs))
6971 io_submit_flush_completions(ctx);
6975 linked_timeout = io_prep_linked_timeout(req);
6977 io_queue_linked_timeout(linked_timeout);
6978 } else if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6979 linked_timeout = io_prep_linked_timeout(req);
6981 switch (io_arm_poll_handler(req)) {
6982 case IO_APOLL_READY:
6984 io_unprep_linked_timeout(req);
6986 case IO_APOLL_ABORTED:
6988 * Queued up for async execution, worker will release
6989 * submit reference when the iocb is actually submitted.
6991 io_queue_async_work(req, NULL);
6996 io_queue_linked_timeout(linked_timeout);
6998 io_req_complete_failed(req, ret);
7002 static inline void io_queue_sqe(struct io_kiocb *req)
7003 __must_hold(&req->ctx->uring_lock)
7005 if (unlikely(req->ctx->drain_active) && io_drain_req(req))
7008 if (likely(!(req->flags & (REQ_F_FORCE_ASYNC | REQ_F_FAIL)))) {
7009 __io_queue_sqe(req);
7010 } else if (req->flags & REQ_F_FAIL) {
7011 io_req_complete_fail_submit(req);
7013 int ret = io_req_prep_async(req);
7016 io_req_complete_failed(req, ret);
7018 io_queue_async_work(req, NULL);
7023 * Check SQE restrictions (opcode and flags).
7025 * Returns 'true' if SQE is allowed, 'false' otherwise.
7027 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
7028 struct io_kiocb *req,
7029 unsigned int sqe_flags)
7031 if (likely(!ctx->restricted))
7034 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
7037 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
7038 ctx->restrictions.sqe_flags_required)
7041 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
7042 ctx->restrictions.sqe_flags_required))
7048 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
7049 const struct io_uring_sqe *sqe)
7050 __must_hold(&ctx->uring_lock)
7052 struct io_submit_state *state;
7053 unsigned int sqe_flags;
7054 int personality, ret = 0;
7056 /* req is partially pre-initialised, see io_preinit_req() */
7057 req->opcode = READ_ONCE(sqe->opcode);
7058 /* same numerical values with corresponding REQ_F_*, safe to copy */
7059 req->flags = sqe_flags = READ_ONCE(sqe->flags);
7060 req->user_data = READ_ONCE(sqe->user_data);
7062 req->fixed_rsrc_refs = NULL;
7063 req->task = current;
7065 /* enforce forwards compatibility on users */
7066 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
7068 if (unlikely(req->opcode >= IORING_OP_LAST))
7070 if (!io_check_restriction(ctx, req, sqe_flags))
7073 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
7074 !io_op_defs[req->opcode].buffer_select)
7076 if (unlikely(sqe_flags & IOSQE_IO_DRAIN))
7077 ctx->drain_active = true;
7079 personality = READ_ONCE(sqe->personality);
7081 req->creds = xa_load(&ctx->personalities, personality);
7084 get_cred(req->creds);
7085 req->flags |= REQ_F_CREDS;
7087 state = &ctx->submit_state;
7090 * Plug now if we have more than 1 IO left after this, and the target
7091 * is potentially a read/write to block based storage.
7093 if (!state->plug_started && state->ios_left > 1 &&
7094 io_op_defs[req->opcode].plug) {
7095 blk_start_plug(&state->plug);
7096 state->plug_started = true;
7099 if (io_op_defs[req->opcode].needs_file) {
7100 req->file = io_file_get(ctx, req, READ_ONCE(sqe->fd),
7101 (sqe_flags & IOSQE_FIXED_FILE));
7102 if (unlikely(!req->file))
7110 static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
7111 const struct io_uring_sqe *sqe)
7112 __must_hold(&ctx->uring_lock)
7114 struct io_submit_link *link = &ctx->submit_state.link;
7117 ret = io_init_req(ctx, req, sqe);
7118 if (unlikely(ret)) {
7120 /* fail even hard links since we don't submit */
7123 * we can judge a link req is failed or cancelled by if
7124 * REQ_F_FAIL is set, but the head is an exception since
7125 * it may be set REQ_F_FAIL because of other req's failure
7126 * so let's leverage req->result to distinguish if a head
7127 * is set REQ_F_FAIL because of its failure or other req's
7128 * failure so that we can set the correct ret code for it.
7129 * init result here to avoid affecting the normal path.
7131 if (!(link->head->flags & REQ_F_FAIL))
7132 req_fail_link_node(link->head, -ECANCELED);
7133 } else if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
7135 * the current req is a normal req, we should return
7136 * error and thus break the submittion loop.
7138 io_req_complete_failed(req, ret);
7141 req_fail_link_node(req, ret);
7143 ret = io_req_prep(req, sqe);
7148 /* don't need @sqe from now on */
7149 trace_io_uring_submit_sqe(ctx, req, req->opcode, req->user_data,
7151 ctx->flags & IORING_SETUP_SQPOLL);
7154 * If we already have a head request, queue this one for async
7155 * submittal once the head completes. If we don't have a head but
7156 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
7157 * submitted sync once the chain is complete. If none of those
7158 * conditions are true (normal request), then just queue it.
7161 struct io_kiocb *head = link->head;
7163 if (!(req->flags & REQ_F_FAIL)) {
7164 ret = io_req_prep_async(req);
7165 if (unlikely(ret)) {
7166 req_fail_link_node(req, ret);
7167 if (!(head->flags & REQ_F_FAIL))
7168 req_fail_link_node(head, -ECANCELED);
7171 trace_io_uring_link(ctx, req, head);
7172 link->last->link = req;
7175 /* last request of a link, enqueue the link */
7176 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
7181 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
7193 * Batched submission is done, ensure local IO is flushed out.
7195 static void io_submit_state_end(struct io_submit_state *state,
7196 struct io_ring_ctx *ctx)
7198 if (state->link.head)
7199 io_queue_sqe(state->link.head);
7200 if (state->compl_nr)
7201 io_submit_flush_completions(ctx);
7202 if (state->plug_started)
7203 blk_finish_plug(&state->plug);
7207 * Start submission side cache.
7209 static void io_submit_state_start(struct io_submit_state *state,
7210 unsigned int max_ios)
7212 state->plug_started = false;
7213 state->ios_left = max_ios;
7214 /* set only head, no need to init link_last in advance */
7215 state->link.head = NULL;
7218 static void io_commit_sqring(struct io_ring_ctx *ctx)
7220 struct io_rings *rings = ctx->rings;
7223 * Ensure any loads from the SQEs are done at this point,
7224 * since once we write the new head, the application could
7225 * write new data to them.
7227 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
7231 * Fetch an sqe, if one is available. Note this returns a pointer to memory
7232 * that is mapped by userspace. This means that care needs to be taken to
7233 * ensure that reads are stable, as we cannot rely on userspace always
7234 * being a good citizen. If members of the sqe are validated and then later
7235 * used, it's important that those reads are done through READ_ONCE() to
7236 * prevent a re-load down the line.
7238 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
7240 unsigned head, mask = ctx->sq_entries - 1;
7241 unsigned sq_idx = ctx->cached_sq_head++ & mask;
7244 * The cached sq head (or cq tail) serves two purposes:
7246 * 1) allows us to batch the cost of updating the user visible
7248 * 2) allows the kernel side to track the head on its own, even
7249 * though the application is the one updating it.
7251 head = READ_ONCE(ctx->sq_array[sq_idx]);
7252 if (likely(head < ctx->sq_entries))
7253 return &ctx->sq_sqes[head];
7255 /* drop invalid entries */
7257 WRITE_ONCE(ctx->rings->sq_dropped,
7258 READ_ONCE(ctx->rings->sq_dropped) + 1);
7262 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
7263 __must_hold(&ctx->uring_lock)
7267 /* make sure SQ entry isn't read before tail */
7268 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
7269 if (!percpu_ref_tryget_many(&ctx->refs, nr))
7271 io_get_task_refs(nr);
7273 io_submit_state_start(&ctx->submit_state, nr);
7274 while (submitted < nr) {
7275 const struct io_uring_sqe *sqe;
7276 struct io_kiocb *req;
7278 req = io_alloc_req(ctx);
7279 if (unlikely(!req)) {
7281 submitted = -EAGAIN;
7284 sqe = io_get_sqe(ctx);
7285 if (unlikely(!sqe)) {
7286 list_add(&req->inflight_entry, &ctx->submit_state.free_list);
7289 /* will complete beyond this point, count as submitted */
7291 if (io_submit_sqe(ctx, req, sqe))
7295 if (unlikely(submitted != nr)) {
7296 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
7297 int unused = nr - ref_used;
7299 current->io_uring->cached_refs += unused;
7300 percpu_ref_put_many(&ctx->refs, unused);
7303 io_submit_state_end(&ctx->submit_state, ctx);
7304 /* Commit SQ ring head once we've consumed and submitted all SQEs */
7305 io_commit_sqring(ctx);
7310 static inline bool io_sqd_events_pending(struct io_sq_data *sqd)
7312 return READ_ONCE(sqd->state);
7315 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
7317 /* Tell userspace we may need a wakeup call */
7318 spin_lock(&ctx->completion_lock);
7319 WRITE_ONCE(ctx->rings->sq_flags,
7320 ctx->rings->sq_flags | IORING_SQ_NEED_WAKEUP);
7321 spin_unlock(&ctx->completion_lock);
7324 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
7326 spin_lock(&ctx->completion_lock);
7327 WRITE_ONCE(ctx->rings->sq_flags,
7328 ctx->rings->sq_flags & ~IORING_SQ_NEED_WAKEUP);
7329 spin_unlock(&ctx->completion_lock);
7332 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
7334 unsigned int to_submit;
7337 to_submit = io_sqring_entries(ctx);
7338 /* if we're handling multiple rings, cap submit size for fairness */
7339 if (cap_entries && to_submit > IORING_SQPOLL_CAP_ENTRIES_VALUE)
7340 to_submit = IORING_SQPOLL_CAP_ENTRIES_VALUE;
7342 if (!list_empty(&ctx->iopoll_list) || to_submit) {
7343 unsigned nr_events = 0;
7344 const struct cred *creds = NULL;
7346 if (ctx->sq_creds != current_cred())
7347 creds = override_creds(ctx->sq_creds);
7349 mutex_lock(&ctx->uring_lock);
7350 if (!list_empty(&ctx->iopoll_list))
7351 io_do_iopoll(ctx, &nr_events, 0);
7354 * Don't submit if refs are dying, good for io_uring_register(),
7355 * but also it is relied upon by io_ring_exit_work()
7357 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)) &&
7358 !(ctx->flags & IORING_SETUP_R_DISABLED))
7359 ret = io_submit_sqes(ctx, to_submit);
7360 mutex_unlock(&ctx->uring_lock);
7362 if (to_submit && wq_has_sleeper(&ctx->sqo_sq_wait))
7363 wake_up(&ctx->sqo_sq_wait);
7365 revert_creds(creds);
7371 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
7373 struct io_ring_ctx *ctx;
7374 unsigned sq_thread_idle = 0;
7376 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7377 sq_thread_idle = max(sq_thread_idle, ctx->sq_thread_idle);
7378 sqd->sq_thread_idle = sq_thread_idle;
7381 static bool io_sqd_handle_event(struct io_sq_data *sqd)
7383 bool did_sig = false;
7384 struct ksignal ksig;
7386 if (test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state) ||
7387 signal_pending(current)) {
7388 mutex_unlock(&sqd->lock);
7389 if (signal_pending(current))
7390 did_sig = get_signal(&ksig);
7392 mutex_lock(&sqd->lock);
7394 return did_sig || test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7397 static int io_sq_thread(void *data)
7399 struct io_sq_data *sqd = data;
7400 struct io_ring_ctx *ctx;
7401 unsigned long timeout = 0;
7402 char buf[TASK_COMM_LEN];
7405 snprintf(buf, sizeof(buf), "iou-sqp-%d", sqd->task_pid);
7406 set_task_comm(current, buf);
7408 if (sqd->sq_cpu != -1)
7409 set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
7411 set_cpus_allowed_ptr(current, cpu_online_mask);
7412 current->flags |= PF_NO_SETAFFINITY;
7414 mutex_lock(&sqd->lock);
7416 bool cap_entries, sqt_spin = false;
7418 if (io_sqd_events_pending(sqd) || signal_pending(current)) {
7419 if (io_sqd_handle_event(sqd))
7421 timeout = jiffies + sqd->sq_thread_idle;
7424 cap_entries = !list_is_singular(&sqd->ctx_list);
7425 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
7426 int ret = __io_sq_thread(ctx, cap_entries);
7428 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
7431 if (io_run_task_work())
7434 if (sqt_spin || !time_after(jiffies, timeout)) {
7437 timeout = jiffies + sqd->sq_thread_idle;
7441 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
7442 if (!io_sqd_events_pending(sqd) && !current->task_works) {
7443 bool needs_sched = true;
7445 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
7446 io_ring_set_wakeup_flag(ctx);
7448 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
7449 !list_empty_careful(&ctx->iopoll_list)) {
7450 needs_sched = false;
7453 if (io_sqring_entries(ctx)) {
7454 needs_sched = false;
7460 mutex_unlock(&sqd->lock);
7462 mutex_lock(&sqd->lock);
7464 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7465 io_ring_clear_wakeup_flag(ctx);
7468 finish_wait(&sqd->wait, &wait);
7469 timeout = jiffies + sqd->sq_thread_idle;
7472 io_uring_cancel_generic(true, sqd);
7474 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7475 io_ring_set_wakeup_flag(ctx);
7477 mutex_unlock(&sqd->lock);
7479 complete(&sqd->exited);
7483 struct io_wait_queue {
7484 struct wait_queue_entry wq;
7485 struct io_ring_ctx *ctx;
7487 unsigned nr_timeouts;
7490 static inline bool io_should_wake(struct io_wait_queue *iowq)
7492 struct io_ring_ctx *ctx = iowq->ctx;
7493 int dist = ctx->cached_cq_tail - (int) iowq->cq_tail;
7496 * Wake up if we have enough events, or if a timeout occurred since we
7497 * started waiting. For timeouts, we always want to return to userspace,
7498 * regardless of event count.
7500 return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
7503 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
7504 int wake_flags, void *key)
7506 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
7510 * Cannot safely flush overflowed CQEs from here, ensure we wake up
7511 * the task, and the next invocation will do it.
7513 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->check_cq_overflow))
7514 return autoremove_wake_function(curr, mode, wake_flags, key);
7518 static int io_run_task_work_sig(void)
7520 if (io_run_task_work())
7522 if (!signal_pending(current))
7524 if (test_thread_flag(TIF_NOTIFY_SIGNAL))
7525 return -ERESTARTSYS;
7529 /* when returns >0, the caller should retry */
7530 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
7531 struct io_wait_queue *iowq,
7532 signed long *timeout)
7536 /* make sure we run task_work before checking for signals */
7537 ret = io_run_task_work_sig();
7538 if (ret || io_should_wake(iowq))
7540 /* let the caller flush overflows, retry */
7541 if (test_bit(0, &ctx->check_cq_overflow))
7544 *timeout = schedule_timeout(*timeout);
7545 return !*timeout ? -ETIME : 1;
7549 * Wait until events become available, if we don't already have some. The
7550 * application must reap them itself, as they reside on the shared cq ring.
7552 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
7553 const sigset_t __user *sig, size_t sigsz,
7554 struct __kernel_timespec __user *uts)
7556 struct io_wait_queue iowq;
7557 struct io_rings *rings = ctx->rings;
7558 signed long timeout = MAX_SCHEDULE_TIMEOUT;
7562 io_cqring_overflow_flush(ctx);
7563 if (io_cqring_events(ctx) >= min_events)
7565 if (!io_run_task_work())
7570 struct timespec64 ts;
7572 if (get_timespec64(&ts, uts))
7574 timeout = timespec64_to_jiffies(&ts);
7578 #ifdef CONFIG_COMPAT
7579 if (in_compat_syscall())
7580 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
7584 ret = set_user_sigmask(sig, sigsz);
7590 init_waitqueue_func_entry(&iowq.wq, io_wake_function);
7591 iowq.wq.private = current;
7592 INIT_LIST_HEAD(&iowq.wq.entry);
7594 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
7595 iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events;
7597 trace_io_uring_cqring_wait(ctx, min_events);
7599 /* if we can't even flush overflow, don't wait for more */
7600 if (!io_cqring_overflow_flush(ctx)) {
7604 prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq,
7605 TASK_INTERRUPTIBLE);
7606 ret = io_cqring_wait_schedule(ctx, &iowq, &timeout);
7607 finish_wait(&ctx->cq_wait, &iowq.wq);
7611 restore_saved_sigmask_unless(ret == -EINTR);
7613 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
7616 static void io_free_page_table(void **table, size_t size)
7618 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7620 for (i = 0; i < nr_tables; i++)
7625 static void **io_alloc_page_table(size_t size)
7627 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7628 size_t init_size = size;
7631 table = kcalloc(nr_tables, sizeof(*table), GFP_KERNEL_ACCOUNT);
7635 for (i = 0; i < nr_tables; i++) {
7636 unsigned int this_size = min_t(size_t, size, PAGE_SIZE);
7638 table[i] = kzalloc(this_size, GFP_KERNEL_ACCOUNT);
7640 io_free_page_table(table, init_size);
7648 static void io_rsrc_node_destroy(struct io_rsrc_node *ref_node)
7650 percpu_ref_exit(&ref_node->refs);
7654 static void io_rsrc_node_ref_zero(struct percpu_ref *ref)
7656 struct io_rsrc_node *node = container_of(ref, struct io_rsrc_node, refs);
7657 struct io_ring_ctx *ctx = node->rsrc_data->ctx;
7658 unsigned long flags;
7659 bool first_add = false;
7661 spin_lock_irqsave(&ctx->rsrc_ref_lock, flags);
7664 while (!list_empty(&ctx->rsrc_ref_list)) {
7665 node = list_first_entry(&ctx->rsrc_ref_list,
7666 struct io_rsrc_node, node);
7667 /* recycle ref nodes in order */
7670 list_del(&node->node);
7671 first_add |= llist_add(&node->llist, &ctx->rsrc_put_llist);
7673 spin_unlock_irqrestore(&ctx->rsrc_ref_lock, flags);
7676 mod_delayed_work(system_wq, &ctx->rsrc_put_work, HZ);
7679 static struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx)
7681 struct io_rsrc_node *ref_node;
7683 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7687 if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
7692 INIT_LIST_HEAD(&ref_node->node);
7693 INIT_LIST_HEAD(&ref_node->rsrc_list);
7694 ref_node->done = false;
7698 static void io_rsrc_node_switch(struct io_ring_ctx *ctx,
7699 struct io_rsrc_data *data_to_kill)
7701 WARN_ON_ONCE(!ctx->rsrc_backup_node);
7702 WARN_ON_ONCE(data_to_kill && !ctx->rsrc_node);
7705 struct io_rsrc_node *rsrc_node = ctx->rsrc_node;
7707 rsrc_node->rsrc_data = data_to_kill;
7708 spin_lock_irq(&ctx->rsrc_ref_lock);
7709 list_add_tail(&rsrc_node->node, &ctx->rsrc_ref_list);
7710 spin_unlock_irq(&ctx->rsrc_ref_lock);
7712 atomic_inc(&data_to_kill->refs);
7713 percpu_ref_kill(&rsrc_node->refs);
7714 ctx->rsrc_node = NULL;
7717 if (!ctx->rsrc_node) {
7718 ctx->rsrc_node = ctx->rsrc_backup_node;
7719 ctx->rsrc_backup_node = NULL;
7723 static int io_rsrc_node_switch_start(struct io_ring_ctx *ctx)
7725 if (ctx->rsrc_backup_node)
7727 ctx->rsrc_backup_node = io_rsrc_node_alloc(ctx);
7728 return ctx->rsrc_backup_node ? 0 : -ENOMEM;
7731 static int io_rsrc_ref_quiesce(struct io_rsrc_data *data, struct io_ring_ctx *ctx)
7735 /* As we may drop ->uring_lock, other task may have started quiesce */
7739 data->quiesce = true;
7741 ret = io_rsrc_node_switch_start(ctx);
7744 io_rsrc_node_switch(ctx, data);
7746 /* kill initial ref, already quiesced if zero */
7747 if (atomic_dec_and_test(&data->refs))
7749 mutex_unlock(&ctx->uring_lock);
7750 flush_delayed_work(&ctx->rsrc_put_work);
7751 ret = wait_for_completion_interruptible(&data->done);
7753 mutex_lock(&ctx->uring_lock);
7757 atomic_inc(&data->refs);
7758 /* wait for all works potentially completing data->done */
7759 flush_delayed_work(&ctx->rsrc_put_work);
7760 reinit_completion(&data->done);
7762 ret = io_run_task_work_sig();
7763 mutex_lock(&ctx->uring_lock);
7765 data->quiesce = false;
7770 static u64 *io_get_tag_slot(struct io_rsrc_data *data, unsigned int idx)
7772 unsigned int off = idx & IO_RSRC_TAG_TABLE_MASK;
7773 unsigned int table_idx = idx >> IO_RSRC_TAG_TABLE_SHIFT;
7775 return &data->tags[table_idx][off];
7778 static void io_rsrc_data_free(struct io_rsrc_data *data)
7780 size_t size = data->nr * sizeof(data->tags[0][0]);
7783 io_free_page_table((void **)data->tags, size);
7787 static int io_rsrc_data_alloc(struct io_ring_ctx *ctx, rsrc_put_fn *do_put,
7788 u64 __user *utags, unsigned nr,
7789 struct io_rsrc_data **pdata)
7791 struct io_rsrc_data *data;
7795 data = kzalloc(sizeof(*data), GFP_KERNEL);
7798 data->tags = (u64 **)io_alloc_page_table(nr * sizeof(data->tags[0][0]));
7806 data->do_put = do_put;
7809 for (i = 0; i < nr; i++) {
7810 u64 *tag_slot = io_get_tag_slot(data, i);
7812 if (copy_from_user(tag_slot, &utags[i],
7818 atomic_set(&data->refs, 1);
7819 init_completion(&data->done);
7823 io_rsrc_data_free(data);
7827 static bool io_alloc_file_tables(struct io_file_table *table, unsigned nr_files)
7829 table->files = kvcalloc(nr_files, sizeof(table->files[0]),
7830 GFP_KERNEL_ACCOUNT);
7831 return !!table->files;
7834 static void io_free_file_tables(struct io_file_table *table)
7836 kvfree(table->files);
7837 table->files = NULL;
7840 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
7842 #if defined(CONFIG_UNIX)
7843 if (ctx->ring_sock) {
7844 struct sock *sock = ctx->ring_sock->sk;
7845 struct sk_buff *skb;
7847 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
7853 for (i = 0; i < ctx->nr_user_files; i++) {
7856 file = io_file_from_index(ctx, i);
7861 io_free_file_tables(&ctx->file_table);
7862 io_rsrc_data_free(ctx->file_data);
7863 ctx->file_data = NULL;
7864 ctx->nr_user_files = 0;
7867 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7871 if (!ctx->file_data)
7873 ret = io_rsrc_ref_quiesce(ctx->file_data, ctx);
7875 __io_sqe_files_unregister(ctx);
7879 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7880 __releases(&sqd->lock)
7882 WARN_ON_ONCE(sqd->thread == current);
7885 * Do the dance but not conditional clear_bit() because it'd race with
7886 * other threads incrementing park_pending and setting the bit.
7888 clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7889 if (atomic_dec_return(&sqd->park_pending))
7890 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7891 mutex_unlock(&sqd->lock);
7894 static void io_sq_thread_park(struct io_sq_data *sqd)
7895 __acquires(&sqd->lock)
7897 WARN_ON_ONCE(sqd->thread == current);
7899 atomic_inc(&sqd->park_pending);
7900 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7901 mutex_lock(&sqd->lock);
7903 wake_up_process(sqd->thread);
7906 static void io_sq_thread_stop(struct io_sq_data *sqd)
7908 WARN_ON_ONCE(sqd->thread == current);
7909 WARN_ON_ONCE(test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state));
7911 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7912 mutex_lock(&sqd->lock);
7914 wake_up_process(sqd->thread);
7915 mutex_unlock(&sqd->lock);
7916 wait_for_completion(&sqd->exited);
7919 static void io_put_sq_data(struct io_sq_data *sqd)
7921 if (refcount_dec_and_test(&sqd->refs)) {
7922 WARN_ON_ONCE(atomic_read(&sqd->park_pending));
7924 io_sq_thread_stop(sqd);
7929 static void io_sq_thread_finish(struct io_ring_ctx *ctx)
7931 struct io_sq_data *sqd = ctx->sq_data;
7934 io_sq_thread_park(sqd);
7935 list_del_init(&ctx->sqd_list);
7936 io_sqd_update_thread_idle(sqd);
7937 io_sq_thread_unpark(sqd);
7939 io_put_sq_data(sqd);
7940 ctx->sq_data = NULL;
7944 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7946 struct io_ring_ctx *ctx_attach;
7947 struct io_sq_data *sqd;
7950 f = fdget(p->wq_fd);
7952 return ERR_PTR(-ENXIO);
7953 if (f.file->f_op != &io_uring_fops) {
7955 return ERR_PTR(-EINVAL);
7958 ctx_attach = f.file->private_data;
7959 sqd = ctx_attach->sq_data;
7962 return ERR_PTR(-EINVAL);
7964 if (sqd->task_tgid != current->tgid) {
7966 return ERR_PTR(-EPERM);
7969 refcount_inc(&sqd->refs);
7974 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p,
7977 struct io_sq_data *sqd;
7980 if (p->flags & IORING_SETUP_ATTACH_WQ) {
7981 sqd = io_attach_sq_data(p);
7986 /* fall through for EPERM case, setup new sqd/task */
7987 if (PTR_ERR(sqd) != -EPERM)
7991 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7993 return ERR_PTR(-ENOMEM);
7995 atomic_set(&sqd->park_pending, 0);
7996 refcount_set(&sqd->refs, 1);
7997 INIT_LIST_HEAD(&sqd->ctx_list);
7998 mutex_init(&sqd->lock);
7999 init_waitqueue_head(&sqd->wait);
8000 init_completion(&sqd->exited);
8004 #if defined(CONFIG_UNIX)
8006 * Ensure the UNIX gc is aware of our file set, so we are certain that
8007 * the io_uring can be safely unregistered on process exit, even if we have
8008 * loops in the file referencing.
8010 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
8012 struct sock *sk = ctx->ring_sock->sk;
8013 struct scm_fp_list *fpl;
8014 struct sk_buff *skb;
8017 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
8021 skb = alloc_skb(0, GFP_KERNEL);
8030 fpl->user = get_uid(current_user());
8031 for (i = 0; i < nr; i++) {
8032 struct file *file = io_file_from_index(ctx, i + offset);
8036 fpl->fp[nr_files] = get_file(file);
8037 unix_inflight(fpl->user, fpl->fp[nr_files]);
8042 fpl->max = SCM_MAX_FD;
8043 fpl->count = nr_files;
8044 UNIXCB(skb).fp = fpl;
8045 skb->destructor = unix_destruct_scm;
8046 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
8047 skb_queue_head(&sk->sk_receive_queue, skb);
8049 for (i = 0; i < nr_files; i++)
8060 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
8061 * causes regular reference counting to break down. We rely on the UNIX
8062 * garbage collection to take care of this problem for us.
8064 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
8066 unsigned left, total;
8070 left = ctx->nr_user_files;
8072 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
8074 ret = __io_sqe_files_scm(ctx, this_files, total);
8078 total += this_files;
8084 while (total < ctx->nr_user_files) {
8085 struct file *file = io_file_from_index(ctx, total);
8095 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
8101 static void io_rsrc_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
8103 struct file *file = prsrc->file;
8104 #if defined(CONFIG_UNIX)
8105 struct sock *sock = ctx->ring_sock->sk;
8106 struct sk_buff_head list, *head = &sock->sk_receive_queue;
8107 struct sk_buff *skb;
8110 __skb_queue_head_init(&list);
8113 * Find the skb that holds this file in its SCM_RIGHTS. When found,
8114 * remove this entry and rearrange the file array.
8116 skb = skb_dequeue(head);
8118 struct scm_fp_list *fp;
8120 fp = UNIXCB(skb).fp;
8121 for (i = 0; i < fp->count; i++) {
8124 if (fp->fp[i] != file)
8127 unix_notinflight(fp->user, fp->fp[i]);
8128 left = fp->count - 1 - i;
8130 memmove(&fp->fp[i], &fp->fp[i + 1],
8131 left * sizeof(struct file *));
8138 __skb_queue_tail(&list, skb);
8148 __skb_queue_tail(&list, skb);
8150 skb = skb_dequeue(head);
8153 if (skb_peek(&list)) {
8154 spin_lock_irq(&head->lock);
8155 while ((skb = __skb_dequeue(&list)) != NULL)
8156 __skb_queue_tail(head, skb);
8157 spin_unlock_irq(&head->lock);
8164 static void __io_rsrc_put_work(struct io_rsrc_node *ref_node)
8166 struct io_rsrc_data *rsrc_data = ref_node->rsrc_data;
8167 struct io_ring_ctx *ctx = rsrc_data->ctx;
8168 struct io_rsrc_put *prsrc, *tmp;
8170 list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
8171 list_del(&prsrc->list);
8174 bool lock_ring = ctx->flags & IORING_SETUP_IOPOLL;
8176 io_ring_submit_lock(ctx, lock_ring);
8177 spin_lock(&ctx->completion_lock);
8178 io_cqring_fill_event(ctx, prsrc->tag, 0, 0);
8180 io_commit_cqring(ctx);
8181 spin_unlock(&ctx->completion_lock);
8182 io_cqring_ev_posted(ctx);
8183 io_ring_submit_unlock(ctx, lock_ring);
8186 rsrc_data->do_put(ctx, prsrc);
8190 io_rsrc_node_destroy(ref_node);
8191 if (atomic_dec_and_test(&rsrc_data->refs))
8192 complete(&rsrc_data->done);
8195 static void io_rsrc_put_work(struct work_struct *work)
8197 struct io_ring_ctx *ctx;
8198 struct llist_node *node;
8200 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
8201 node = llist_del_all(&ctx->rsrc_put_llist);
8204 struct io_rsrc_node *ref_node;
8205 struct llist_node *next = node->next;
8207 ref_node = llist_entry(node, struct io_rsrc_node, llist);
8208 __io_rsrc_put_work(ref_node);
8213 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
8214 unsigned nr_args, u64 __user *tags)
8216 __s32 __user *fds = (__s32 __user *) arg;
8225 if (nr_args > IORING_MAX_FIXED_FILES)
8227 if (nr_args > rlimit(RLIMIT_NOFILE))
8229 ret = io_rsrc_node_switch_start(ctx);
8232 ret = io_rsrc_data_alloc(ctx, io_rsrc_file_put, tags, nr_args,
8238 if (!io_alloc_file_tables(&ctx->file_table, nr_args))
8241 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
8242 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
8246 /* allow sparse sets */
8249 if (unlikely(*io_get_tag_slot(ctx->file_data, i)))
8256 if (unlikely(!file))
8260 * Don't allow io_uring instances to be registered. If UNIX
8261 * isn't enabled, then this causes a reference cycle and this
8262 * instance can never get freed. If UNIX is enabled we'll
8263 * handle it just fine, but there's still no point in allowing
8264 * a ring fd as it doesn't support regular read/write anyway.
8266 if (file->f_op == &io_uring_fops) {
8270 io_fixed_file_set(io_fixed_file_slot(&ctx->file_table, i), file);
8273 ret = io_sqe_files_scm(ctx);
8275 __io_sqe_files_unregister(ctx);
8279 io_rsrc_node_switch(ctx, NULL);
8282 for (i = 0; i < ctx->nr_user_files; i++) {
8283 file = io_file_from_index(ctx, i);
8287 io_free_file_tables(&ctx->file_table);
8288 ctx->nr_user_files = 0;
8290 io_rsrc_data_free(ctx->file_data);
8291 ctx->file_data = NULL;
8295 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
8298 #if defined(CONFIG_UNIX)
8299 struct sock *sock = ctx->ring_sock->sk;
8300 struct sk_buff_head *head = &sock->sk_receive_queue;
8301 struct sk_buff *skb;
8304 * See if we can merge this file into an existing skb SCM_RIGHTS
8305 * file set. If there's no room, fall back to allocating a new skb
8306 * and filling it in.
8308 spin_lock_irq(&head->lock);
8309 skb = skb_peek(head);
8311 struct scm_fp_list *fpl = UNIXCB(skb).fp;
8313 if (fpl->count < SCM_MAX_FD) {
8314 __skb_unlink(skb, head);
8315 spin_unlock_irq(&head->lock);
8316 fpl->fp[fpl->count] = get_file(file);
8317 unix_inflight(fpl->user, fpl->fp[fpl->count]);
8319 spin_lock_irq(&head->lock);
8320 __skb_queue_head(head, skb);
8325 spin_unlock_irq(&head->lock);
8332 return __io_sqe_files_scm(ctx, 1, index);
8338 static int io_queue_rsrc_removal(struct io_rsrc_data *data, unsigned idx,
8339 struct io_rsrc_node *node, void *rsrc)
8341 struct io_rsrc_put *prsrc;
8343 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
8347 prsrc->tag = *io_get_tag_slot(data, idx);
8349 list_add(&prsrc->list, &node->rsrc_list);
8353 static int io_install_fixed_file(struct io_kiocb *req, struct file *file,
8354 unsigned int issue_flags, u32 slot_index)
8356 struct io_ring_ctx *ctx = req->ctx;
8357 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
8358 bool needs_switch = false;
8359 struct io_fixed_file *file_slot;
8362 io_ring_submit_lock(ctx, !force_nonblock);
8363 if (file->f_op == &io_uring_fops)
8366 if (!ctx->file_data)
8369 if (slot_index >= ctx->nr_user_files)
8372 slot_index = array_index_nospec(slot_index, ctx->nr_user_files);
8373 file_slot = io_fixed_file_slot(&ctx->file_table, slot_index);
8375 if (file_slot->file_ptr) {
8376 struct file *old_file;
8378 ret = io_rsrc_node_switch_start(ctx);
8382 old_file = (struct file *)(file_slot->file_ptr & FFS_MASK);
8383 ret = io_queue_rsrc_removal(ctx->file_data, slot_index,
8384 ctx->rsrc_node, old_file);
8387 file_slot->file_ptr = 0;
8388 needs_switch = true;
8391 *io_get_tag_slot(ctx->file_data, slot_index) = 0;
8392 io_fixed_file_set(file_slot, file);
8393 ret = io_sqe_file_register(ctx, file, slot_index);
8395 file_slot->file_ptr = 0;
8402 io_rsrc_node_switch(ctx, ctx->file_data);
8403 io_ring_submit_unlock(ctx, !force_nonblock);
8409 static int io_close_fixed(struct io_kiocb *req, unsigned int issue_flags)
8411 unsigned int offset = req->close.file_slot - 1;
8412 struct io_ring_ctx *ctx = req->ctx;
8413 struct io_fixed_file *file_slot;
8417 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
8419 if (unlikely(!ctx->file_data))
8422 if (offset >= ctx->nr_user_files)
8424 ret = io_rsrc_node_switch_start(ctx);
8428 i = array_index_nospec(offset, ctx->nr_user_files);
8429 file_slot = io_fixed_file_slot(&ctx->file_table, i);
8431 if (!file_slot->file_ptr)
8434 file = (struct file *)(file_slot->file_ptr & FFS_MASK);
8435 ret = io_queue_rsrc_removal(ctx->file_data, offset, ctx->rsrc_node, file);
8439 file_slot->file_ptr = 0;
8440 io_rsrc_node_switch(ctx, ctx->file_data);
8443 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
8447 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
8448 struct io_uring_rsrc_update2 *up,
8451 u64 __user *tags = u64_to_user_ptr(up->tags);
8452 __s32 __user *fds = u64_to_user_ptr(up->data);
8453 struct io_rsrc_data *data = ctx->file_data;
8454 struct io_fixed_file *file_slot;
8458 bool needs_switch = false;
8460 if (!ctx->file_data)
8462 if (up->offset + nr_args > ctx->nr_user_files)
8465 for (done = 0; done < nr_args; done++) {
8468 if ((tags && copy_from_user(&tag, &tags[done], sizeof(tag))) ||
8469 copy_from_user(&fd, &fds[done], sizeof(fd))) {
8473 if ((fd == IORING_REGISTER_FILES_SKIP || fd == -1) && tag) {
8477 if (fd == IORING_REGISTER_FILES_SKIP)
8480 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
8481 file_slot = io_fixed_file_slot(&ctx->file_table, i);
8483 if (file_slot->file_ptr) {
8484 file = (struct file *)(file_slot->file_ptr & FFS_MASK);
8485 err = io_queue_rsrc_removal(data, up->offset + done,
8486 ctx->rsrc_node, file);
8489 file_slot->file_ptr = 0;
8490 needs_switch = true;
8499 * Don't allow io_uring instances to be registered. If
8500 * UNIX isn't enabled, then this causes a reference
8501 * cycle and this instance can never get freed. If UNIX
8502 * is enabled we'll handle it just fine, but there's
8503 * still no point in allowing a ring fd as it doesn't
8504 * support regular read/write anyway.
8506 if (file->f_op == &io_uring_fops) {
8511 *io_get_tag_slot(data, up->offset + done) = tag;
8512 io_fixed_file_set(file_slot, file);
8513 err = io_sqe_file_register(ctx, file, i);
8515 file_slot->file_ptr = 0;
8523 io_rsrc_node_switch(ctx, data);
8524 return done ? done : err;
8527 static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx,
8528 struct task_struct *task)
8530 struct io_wq_hash *hash;
8531 struct io_wq_data data;
8532 unsigned int concurrency;
8534 mutex_lock(&ctx->uring_lock);
8535 hash = ctx->hash_map;
8537 hash = kzalloc(sizeof(*hash), GFP_KERNEL);
8539 mutex_unlock(&ctx->uring_lock);
8540 return ERR_PTR(-ENOMEM);
8542 refcount_set(&hash->refs, 1);
8543 init_waitqueue_head(&hash->wait);
8544 ctx->hash_map = hash;
8546 mutex_unlock(&ctx->uring_lock);
8550 data.free_work = io_wq_free_work;
8551 data.do_work = io_wq_submit_work;
8553 /* Do QD, or 4 * CPUS, whatever is smallest */
8554 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
8556 return io_wq_create(concurrency, &data);
8559 static int io_uring_alloc_task_context(struct task_struct *task,
8560 struct io_ring_ctx *ctx)
8562 struct io_uring_task *tctx;
8565 tctx = kzalloc(sizeof(*tctx), GFP_KERNEL);
8566 if (unlikely(!tctx))
8569 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
8570 if (unlikely(ret)) {
8575 tctx->io_wq = io_init_wq_offload(ctx, task);
8576 if (IS_ERR(tctx->io_wq)) {
8577 ret = PTR_ERR(tctx->io_wq);
8578 percpu_counter_destroy(&tctx->inflight);
8584 init_waitqueue_head(&tctx->wait);
8585 atomic_set(&tctx->in_idle, 0);
8586 atomic_set(&tctx->inflight_tracked, 0);
8587 task->io_uring = tctx;
8588 spin_lock_init(&tctx->task_lock);
8589 INIT_WQ_LIST(&tctx->task_list);
8590 init_task_work(&tctx->task_work, tctx_task_work);
8594 void __io_uring_free(struct task_struct *tsk)
8596 struct io_uring_task *tctx = tsk->io_uring;
8598 WARN_ON_ONCE(!xa_empty(&tctx->xa));
8599 WARN_ON_ONCE(tctx->io_wq);
8600 WARN_ON_ONCE(tctx->cached_refs);
8602 percpu_counter_destroy(&tctx->inflight);
8604 tsk->io_uring = NULL;
8607 static int io_sq_offload_create(struct io_ring_ctx *ctx,
8608 struct io_uring_params *p)
8612 /* Retain compatibility with failing for an invalid attach attempt */
8613 if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
8614 IORING_SETUP_ATTACH_WQ) {
8617 f = fdget(p->wq_fd);
8620 if (f.file->f_op != &io_uring_fops) {
8626 if (ctx->flags & IORING_SETUP_SQPOLL) {
8627 struct task_struct *tsk;
8628 struct io_sq_data *sqd;
8631 sqd = io_get_sq_data(p, &attached);
8637 ctx->sq_creds = get_current_cred();
8639 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
8640 if (!ctx->sq_thread_idle)
8641 ctx->sq_thread_idle = HZ;
8643 io_sq_thread_park(sqd);
8644 list_add(&ctx->sqd_list, &sqd->ctx_list);
8645 io_sqd_update_thread_idle(sqd);
8646 /* don't attach to a dying SQPOLL thread, would be racy */
8647 ret = (attached && !sqd->thread) ? -ENXIO : 0;
8648 io_sq_thread_unpark(sqd);
8655 if (p->flags & IORING_SETUP_SQ_AFF) {
8656 int cpu = p->sq_thread_cpu;
8659 if (cpu >= nr_cpu_ids || !cpu_online(cpu))
8666 sqd->task_pid = current->pid;
8667 sqd->task_tgid = current->tgid;
8668 tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE);
8675 ret = io_uring_alloc_task_context(tsk, ctx);
8676 wake_up_new_task(tsk);
8679 } else if (p->flags & IORING_SETUP_SQ_AFF) {
8680 /* Can't have SQ_AFF without SQPOLL */
8687 complete(&ctx->sq_data->exited);
8689 io_sq_thread_finish(ctx);
8693 static inline void __io_unaccount_mem(struct user_struct *user,
8694 unsigned long nr_pages)
8696 atomic_long_sub(nr_pages, &user->locked_vm);
8699 static inline int __io_account_mem(struct user_struct *user,
8700 unsigned long nr_pages)
8702 unsigned long page_limit, cur_pages, new_pages;
8704 /* Don't allow more pages than we can safely lock */
8705 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
8708 cur_pages = atomic_long_read(&user->locked_vm);
8709 new_pages = cur_pages + nr_pages;
8710 if (new_pages > page_limit)
8712 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
8713 new_pages) != cur_pages);
8718 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8721 __io_unaccount_mem(ctx->user, nr_pages);
8723 if (ctx->mm_account)
8724 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
8727 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8732 ret = __io_account_mem(ctx->user, nr_pages);
8737 if (ctx->mm_account)
8738 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
8743 static void io_mem_free(void *ptr)
8750 page = virt_to_head_page(ptr);
8751 if (put_page_testzero(page))
8752 free_compound_page(page);
8755 static void *io_mem_alloc(size_t size)
8757 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
8758 __GFP_NORETRY | __GFP_ACCOUNT;
8760 return (void *) __get_free_pages(gfp_flags, get_order(size));
8763 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8766 struct io_rings *rings;
8767 size_t off, sq_array_size;
8769 off = struct_size(rings, cqes, cq_entries);
8770 if (off == SIZE_MAX)
8774 off = ALIGN(off, SMP_CACHE_BYTES);
8782 sq_array_size = array_size(sizeof(u32), sq_entries);
8783 if (sq_array_size == SIZE_MAX)
8786 if (check_add_overflow(off, sq_array_size, &off))
8792 static void io_buffer_unmap(struct io_ring_ctx *ctx, struct io_mapped_ubuf **slot)
8794 struct io_mapped_ubuf *imu = *slot;
8797 if (imu != ctx->dummy_ubuf) {
8798 for (i = 0; i < imu->nr_bvecs; i++)
8799 unpin_user_page(imu->bvec[i].bv_page);
8800 if (imu->acct_pages)
8801 io_unaccount_mem(ctx, imu->acct_pages);
8807 static void io_rsrc_buf_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
8809 io_buffer_unmap(ctx, &prsrc->buf);
8813 static void __io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8817 for (i = 0; i < ctx->nr_user_bufs; i++)
8818 io_buffer_unmap(ctx, &ctx->user_bufs[i]);
8819 kfree(ctx->user_bufs);
8820 io_rsrc_data_free(ctx->buf_data);
8821 ctx->user_bufs = NULL;
8822 ctx->buf_data = NULL;
8823 ctx->nr_user_bufs = 0;
8826 static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8833 ret = io_rsrc_ref_quiesce(ctx->buf_data, ctx);
8835 __io_sqe_buffers_unregister(ctx);
8839 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8840 void __user *arg, unsigned index)
8842 struct iovec __user *src;
8844 #ifdef CONFIG_COMPAT
8846 struct compat_iovec __user *ciovs;
8847 struct compat_iovec ciov;
8849 ciovs = (struct compat_iovec __user *) arg;
8850 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8853 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8854 dst->iov_len = ciov.iov_len;
8858 src = (struct iovec __user *) arg;
8859 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8865 * Not super efficient, but this is just a registration time. And we do cache
8866 * the last compound head, so generally we'll only do a full search if we don't
8869 * We check if the given compound head page has already been accounted, to
8870 * avoid double accounting it. This allows us to account the full size of the
8871 * page, not just the constituent pages of a huge page.
8873 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8874 int nr_pages, struct page *hpage)
8878 /* check current page array */
8879 for (i = 0; i < nr_pages; i++) {
8880 if (!PageCompound(pages[i]))
8882 if (compound_head(pages[i]) == hpage)
8886 /* check previously registered pages */
8887 for (i = 0; i < ctx->nr_user_bufs; i++) {
8888 struct io_mapped_ubuf *imu = ctx->user_bufs[i];
8890 for (j = 0; j < imu->nr_bvecs; j++) {
8891 if (!PageCompound(imu->bvec[j].bv_page))
8893 if (compound_head(imu->bvec[j].bv_page) == hpage)
8901 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8902 int nr_pages, struct io_mapped_ubuf *imu,
8903 struct page **last_hpage)
8907 imu->acct_pages = 0;
8908 for (i = 0; i < nr_pages; i++) {
8909 if (!PageCompound(pages[i])) {
8914 hpage = compound_head(pages[i]);
8915 if (hpage == *last_hpage)
8917 *last_hpage = hpage;
8918 if (headpage_already_acct(ctx, pages, i, hpage))
8920 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8924 if (!imu->acct_pages)
8927 ret = io_account_mem(ctx, imu->acct_pages);
8929 imu->acct_pages = 0;
8933 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
8934 struct io_mapped_ubuf **pimu,
8935 struct page **last_hpage)
8937 struct io_mapped_ubuf *imu = NULL;
8938 struct vm_area_struct **vmas = NULL;
8939 struct page **pages = NULL;
8940 unsigned long off, start, end, ubuf;
8942 int ret, pret, nr_pages, i;
8944 if (!iov->iov_base) {
8945 *pimu = ctx->dummy_ubuf;
8949 ubuf = (unsigned long) iov->iov_base;
8950 end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8951 start = ubuf >> PAGE_SHIFT;
8952 nr_pages = end - start;
8957 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
8961 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
8966 imu = kvmalloc(struct_size(imu, bvec, nr_pages), GFP_KERNEL);
8971 mmap_read_lock(current->mm);
8972 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
8974 if (pret == nr_pages) {
8975 /* don't support file backed memory */
8976 for (i = 0; i < nr_pages; i++) {
8977 struct vm_area_struct *vma = vmas[i];
8979 if (vma_is_shmem(vma))
8982 !is_file_hugepages(vma->vm_file)) {
8988 ret = pret < 0 ? pret : -EFAULT;
8990 mmap_read_unlock(current->mm);
8993 * if we did partial map, or found file backed vmas,
8994 * release any pages we did get
8997 unpin_user_pages(pages, pret);
9001 ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
9003 unpin_user_pages(pages, pret);
9007 off = ubuf & ~PAGE_MASK;
9008 size = iov->iov_len;
9009 for (i = 0; i < nr_pages; i++) {
9012 vec_len = min_t(size_t, size, PAGE_SIZE - off);
9013 imu->bvec[i].bv_page = pages[i];
9014 imu->bvec[i].bv_len = vec_len;
9015 imu->bvec[i].bv_offset = off;
9019 /* store original address for later verification */
9021 imu->ubuf_end = ubuf + iov->iov_len;
9022 imu->nr_bvecs = nr_pages;
9033 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
9035 ctx->user_bufs = kcalloc(nr_args, sizeof(*ctx->user_bufs), GFP_KERNEL);
9036 return ctx->user_bufs ? 0 : -ENOMEM;
9039 static int io_buffer_validate(struct iovec *iov)
9041 unsigned long tmp, acct_len = iov->iov_len + (PAGE_SIZE - 1);
9044 * Don't impose further limits on the size and buffer
9045 * constraints here, we'll -EINVAL later when IO is
9046 * submitted if they are wrong.
9049 return iov->iov_len ? -EFAULT : 0;
9053 /* arbitrary limit, but we need something */
9054 if (iov->iov_len > SZ_1G)
9057 if (check_add_overflow((unsigned long)iov->iov_base, acct_len, &tmp))
9063 static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
9064 unsigned int nr_args, u64 __user *tags)
9066 struct page *last_hpage = NULL;
9067 struct io_rsrc_data *data;
9073 if (!nr_args || nr_args > IORING_MAX_REG_BUFFERS)
9075 ret = io_rsrc_node_switch_start(ctx);
9078 ret = io_rsrc_data_alloc(ctx, io_rsrc_buf_put, tags, nr_args, &data);
9081 ret = io_buffers_map_alloc(ctx, nr_args);
9083 io_rsrc_data_free(data);
9087 for (i = 0; i < nr_args; i++, ctx->nr_user_bufs++) {
9088 ret = io_copy_iov(ctx, &iov, arg, i);
9091 ret = io_buffer_validate(&iov);
9094 if (!iov.iov_base && *io_get_tag_slot(data, i)) {
9099 ret = io_sqe_buffer_register(ctx, &iov, &ctx->user_bufs[i],
9105 WARN_ON_ONCE(ctx->buf_data);
9107 ctx->buf_data = data;
9109 __io_sqe_buffers_unregister(ctx);
9111 io_rsrc_node_switch(ctx, NULL);
9115 static int __io_sqe_buffers_update(struct io_ring_ctx *ctx,
9116 struct io_uring_rsrc_update2 *up,
9117 unsigned int nr_args)
9119 u64 __user *tags = u64_to_user_ptr(up->tags);
9120 struct iovec iov, __user *iovs = u64_to_user_ptr(up->data);
9121 struct page *last_hpage = NULL;
9122 bool needs_switch = false;
9128 if (up->offset + nr_args > ctx->nr_user_bufs)
9131 for (done = 0; done < nr_args; done++) {
9132 struct io_mapped_ubuf *imu;
9133 int offset = up->offset + done;
9136 err = io_copy_iov(ctx, &iov, iovs, done);
9139 if (tags && copy_from_user(&tag, &tags[done], sizeof(tag))) {
9143 err = io_buffer_validate(&iov);
9146 if (!iov.iov_base && tag) {
9150 err = io_sqe_buffer_register(ctx, &iov, &imu, &last_hpage);
9154 i = array_index_nospec(offset, ctx->nr_user_bufs);
9155 if (ctx->user_bufs[i] != ctx->dummy_ubuf) {
9156 err = io_queue_rsrc_removal(ctx->buf_data, offset,
9157 ctx->rsrc_node, ctx->user_bufs[i]);
9158 if (unlikely(err)) {
9159 io_buffer_unmap(ctx, &imu);
9162 ctx->user_bufs[i] = NULL;
9163 needs_switch = true;
9166 ctx->user_bufs[i] = imu;
9167 *io_get_tag_slot(ctx->buf_data, offset) = tag;
9171 io_rsrc_node_switch(ctx, ctx->buf_data);
9172 return done ? done : err;
9175 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
9177 __s32 __user *fds = arg;
9183 if (copy_from_user(&fd, fds, sizeof(*fds)))
9186 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
9187 if (IS_ERR(ctx->cq_ev_fd)) {
9188 int ret = PTR_ERR(ctx->cq_ev_fd);
9190 ctx->cq_ev_fd = NULL;
9197 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
9199 if (ctx->cq_ev_fd) {
9200 eventfd_ctx_put(ctx->cq_ev_fd);
9201 ctx->cq_ev_fd = NULL;
9208 static void io_destroy_buffers(struct io_ring_ctx *ctx)
9210 struct io_buffer *buf;
9211 unsigned long index;
9213 xa_for_each(&ctx->io_buffers, index, buf) {
9214 __io_remove_buffers(ctx, buf, index, -1U);
9219 static void io_req_cache_free(struct list_head *list)
9221 struct io_kiocb *req, *nxt;
9223 list_for_each_entry_safe(req, nxt, list, inflight_entry) {
9224 list_del(&req->inflight_entry);
9225 kmem_cache_free(req_cachep, req);
9229 static void io_req_caches_free(struct io_ring_ctx *ctx)
9231 struct io_submit_state *state = &ctx->submit_state;
9233 mutex_lock(&ctx->uring_lock);
9235 if (state->free_reqs) {
9236 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
9237 state->free_reqs = 0;
9240 io_flush_cached_locked_reqs(ctx, state);
9241 io_req_cache_free(&state->free_list);
9242 mutex_unlock(&ctx->uring_lock);
9245 static void io_wait_rsrc_data(struct io_rsrc_data *data)
9247 if (data && !atomic_dec_and_test(&data->refs))
9248 wait_for_completion(&data->done);
9251 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
9253 io_sq_thread_finish(ctx);
9255 if (ctx->mm_account) {
9256 mmdrop(ctx->mm_account);
9257 ctx->mm_account = NULL;
9260 /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */
9261 io_wait_rsrc_data(ctx->buf_data);
9262 io_wait_rsrc_data(ctx->file_data);
9264 mutex_lock(&ctx->uring_lock);
9266 __io_sqe_buffers_unregister(ctx);
9268 __io_sqe_files_unregister(ctx);
9270 __io_cqring_overflow_flush(ctx, true);
9271 mutex_unlock(&ctx->uring_lock);
9272 io_eventfd_unregister(ctx);
9273 io_destroy_buffers(ctx);
9275 put_cred(ctx->sq_creds);
9277 /* there are no registered resources left, nobody uses it */
9279 io_rsrc_node_destroy(ctx->rsrc_node);
9280 if (ctx->rsrc_backup_node)
9281 io_rsrc_node_destroy(ctx->rsrc_backup_node);
9282 flush_delayed_work(&ctx->rsrc_put_work);
9284 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
9285 WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist));
9287 #if defined(CONFIG_UNIX)
9288 if (ctx->ring_sock) {
9289 ctx->ring_sock->file = NULL; /* so that iput() is called */
9290 sock_release(ctx->ring_sock);
9293 WARN_ON_ONCE(!list_empty(&ctx->ltimeout_list));
9295 io_mem_free(ctx->rings);
9296 io_mem_free(ctx->sq_sqes);
9298 percpu_ref_exit(&ctx->refs);
9299 free_uid(ctx->user);
9300 io_req_caches_free(ctx);
9302 io_wq_put_hash(ctx->hash_map);
9303 kfree(ctx->cancel_hash);
9304 kfree(ctx->dummy_ubuf);
9308 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
9310 struct io_ring_ctx *ctx = file->private_data;
9313 poll_wait(file, &ctx->poll_wait, wait);
9315 * synchronizes with barrier from wq_has_sleeper call in
9319 if (!io_sqring_full(ctx))
9320 mask |= EPOLLOUT | EPOLLWRNORM;
9323 * Don't flush cqring overflow list here, just do a simple check.
9324 * Otherwise there could possible be ABBA deadlock:
9327 * lock(&ctx->uring_lock);
9329 * lock(&ctx->uring_lock);
9332 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
9333 * pushs them to do the flush.
9335 if (io_cqring_events(ctx) || test_bit(0, &ctx->check_cq_overflow))
9336 mask |= EPOLLIN | EPOLLRDNORM;
9341 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
9343 const struct cred *creds;
9345 creds = xa_erase(&ctx->personalities, id);
9354 struct io_tctx_exit {
9355 struct callback_head task_work;
9356 struct completion completion;
9357 struct io_ring_ctx *ctx;
9360 static void io_tctx_exit_cb(struct callback_head *cb)
9362 struct io_uring_task *tctx = current->io_uring;
9363 struct io_tctx_exit *work;
9365 work = container_of(cb, struct io_tctx_exit, task_work);
9367 * When @in_idle, we're in cancellation and it's racy to remove the
9368 * node. It'll be removed by the end of cancellation, just ignore it.
9370 if (!atomic_read(&tctx->in_idle))
9371 io_uring_del_tctx_node((unsigned long)work->ctx);
9372 complete(&work->completion);
9375 static bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
9377 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
9379 return req->ctx == data;
9382 static void io_ring_exit_work(struct work_struct *work)
9384 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
9385 unsigned long timeout = jiffies + HZ * 60 * 5;
9386 unsigned long interval = HZ / 20;
9387 struct io_tctx_exit exit;
9388 struct io_tctx_node *node;
9392 * If we're doing polled IO and end up having requests being
9393 * submitted async (out-of-line), then completions can come in while
9394 * we're waiting for refs to drop. We need to reap these manually,
9395 * as nobody else will be looking for them.
9398 io_uring_try_cancel_requests(ctx, NULL, true);
9400 struct io_sq_data *sqd = ctx->sq_data;
9401 struct task_struct *tsk;
9403 io_sq_thread_park(sqd);
9405 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
9406 io_wq_cancel_cb(tsk->io_uring->io_wq,
9407 io_cancel_ctx_cb, ctx, true);
9408 io_sq_thread_unpark(sqd);
9411 if (WARN_ON_ONCE(time_after(jiffies, timeout))) {
9412 /* there is little hope left, don't run it too often */
9415 } while (!wait_for_completion_timeout(&ctx->ref_comp, interval));
9417 init_completion(&exit.completion);
9418 init_task_work(&exit.task_work, io_tctx_exit_cb);
9421 * Some may use context even when all refs and requests have been put,
9422 * and they are free to do so while still holding uring_lock or
9423 * completion_lock, see io_req_task_submit(). Apart from other work,
9424 * this lock/unlock section also waits them to finish.
9426 mutex_lock(&ctx->uring_lock);
9427 while (!list_empty(&ctx->tctx_list)) {
9428 WARN_ON_ONCE(time_after(jiffies, timeout));
9430 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
9432 /* don't spin on a single task if cancellation failed */
9433 list_rotate_left(&ctx->tctx_list);
9434 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
9435 if (WARN_ON_ONCE(ret))
9437 wake_up_process(node->task);
9439 mutex_unlock(&ctx->uring_lock);
9440 wait_for_completion(&exit.completion);
9441 mutex_lock(&ctx->uring_lock);
9443 mutex_unlock(&ctx->uring_lock);
9444 spin_lock(&ctx->completion_lock);
9445 spin_unlock(&ctx->completion_lock);
9447 io_ring_ctx_free(ctx);
9450 /* Returns true if we found and killed one or more timeouts */
9451 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
9454 struct io_kiocb *req, *tmp;
9457 spin_lock(&ctx->completion_lock);
9458 spin_lock_irq(&ctx->timeout_lock);
9459 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
9460 if (io_match_task(req, tsk, cancel_all)) {
9461 io_kill_timeout(req, -ECANCELED);
9465 spin_unlock_irq(&ctx->timeout_lock);
9467 io_commit_cqring(ctx);
9468 spin_unlock(&ctx->completion_lock);
9470 io_cqring_ev_posted(ctx);
9471 return canceled != 0;
9474 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
9476 unsigned long index;
9477 struct creds *creds;
9479 mutex_lock(&ctx->uring_lock);
9480 percpu_ref_kill(&ctx->refs);
9482 __io_cqring_overflow_flush(ctx, true);
9483 xa_for_each(&ctx->personalities, index, creds)
9484 io_unregister_personality(ctx, index);
9485 mutex_unlock(&ctx->uring_lock);
9487 io_kill_timeouts(ctx, NULL, true);
9488 io_poll_remove_all(ctx, NULL, true);
9490 /* if we failed setting up the ctx, we might not have any rings */
9491 io_iopoll_try_reap_events(ctx);
9493 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
9495 * Use system_unbound_wq to avoid spawning tons of event kworkers
9496 * if we're exiting a ton of rings at the same time. It just adds
9497 * noise and overhead, there's no discernable change in runtime
9498 * over using system_wq.
9500 queue_work(system_unbound_wq, &ctx->exit_work);
9503 static int io_uring_release(struct inode *inode, struct file *file)
9505 struct io_ring_ctx *ctx = file->private_data;
9507 file->private_data = NULL;
9508 io_ring_ctx_wait_and_kill(ctx);
9512 struct io_task_cancel {
9513 struct task_struct *task;
9517 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
9519 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
9520 struct io_task_cancel *cancel = data;
9523 if (!cancel->all && (req->flags & REQ_F_LINK_TIMEOUT)) {
9524 struct io_ring_ctx *ctx = req->ctx;
9526 /* protect against races with linked timeouts */
9527 spin_lock(&ctx->completion_lock);
9528 ret = io_match_task(req, cancel->task, cancel->all);
9529 spin_unlock(&ctx->completion_lock);
9531 ret = io_match_task(req, cancel->task, cancel->all);
9536 static bool io_cancel_defer_files(struct io_ring_ctx *ctx,
9537 struct task_struct *task, bool cancel_all)
9539 struct io_defer_entry *de;
9542 spin_lock(&ctx->completion_lock);
9543 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
9544 if (io_match_task(de->req, task, cancel_all)) {
9545 list_cut_position(&list, &ctx->defer_list, &de->list);
9549 spin_unlock(&ctx->completion_lock);
9550 if (list_empty(&list))
9553 while (!list_empty(&list)) {
9554 de = list_first_entry(&list, struct io_defer_entry, list);
9555 list_del_init(&de->list);
9556 io_req_complete_failed(de->req, -ECANCELED);
9562 static bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
9564 struct io_tctx_node *node;
9565 enum io_wq_cancel cret;
9568 mutex_lock(&ctx->uring_lock);
9569 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
9570 struct io_uring_task *tctx = node->task->io_uring;
9573 * io_wq will stay alive while we hold uring_lock, because it's
9574 * killed after ctx nodes, which requires to take the lock.
9576 if (!tctx || !tctx->io_wq)
9578 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
9579 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
9581 mutex_unlock(&ctx->uring_lock);
9586 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
9587 struct task_struct *task,
9590 struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
9591 struct io_uring_task *tctx = task ? task->io_uring : NULL;
9594 enum io_wq_cancel cret;
9598 ret |= io_uring_try_cancel_iowq(ctx);
9599 } else if (tctx && tctx->io_wq) {
9601 * Cancels requests of all rings, not only @ctx, but
9602 * it's fine as the task is in exit/exec.
9604 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
9606 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
9609 /* SQPOLL thread does its own polling */
9610 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
9611 (ctx->sq_data && ctx->sq_data->thread == current)) {
9612 while (!list_empty_careful(&ctx->iopoll_list)) {
9613 io_iopoll_try_reap_events(ctx);
9618 ret |= io_cancel_defer_files(ctx, task, cancel_all);
9619 ret |= io_poll_remove_all(ctx, task, cancel_all);
9620 ret |= io_kill_timeouts(ctx, task, cancel_all);
9622 ret |= io_run_task_work();
9629 static int __io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9631 struct io_uring_task *tctx = current->io_uring;
9632 struct io_tctx_node *node;
9635 if (unlikely(!tctx)) {
9636 ret = io_uring_alloc_task_context(current, ctx);
9639 tctx = current->io_uring;
9641 if (!xa_load(&tctx->xa, (unsigned long)ctx)) {
9642 node = kmalloc(sizeof(*node), GFP_KERNEL);
9646 node->task = current;
9648 ret = xa_err(xa_store(&tctx->xa, (unsigned long)ctx,
9655 mutex_lock(&ctx->uring_lock);
9656 list_add(&node->ctx_node, &ctx->tctx_list);
9657 mutex_unlock(&ctx->uring_lock);
9664 * Note that this task has used io_uring. We use it for cancelation purposes.
9666 static inline int io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9668 struct io_uring_task *tctx = current->io_uring;
9670 if (likely(tctx && tctx->last == ctx))
9672 return __io_uring_add_tctx_node(ctx);
9676 * Remove this io_uring_file -> task mapping.
9678 static void io_uring_del_tctx_node(unsigned long index)
9680 struct io_uring_task *tctx = current->io_uring;
9681 struct io_tctx_node *node;
9685 node = xa_erase(&tctx->xa, index);
9689 WARN_ON_ONCE(current != node->task);
9690 WARN_ON_ONCE(list_empty(&node->ctx_node));
9692 mutex_lock(&node->ctx->uring_lock);
9693 list_del(&node->ctx_node);
9694 mutex_unlock(&node->ctx->uring_lock);
9696 if (tctx->last == node->ctx)
9701 static void io_uring_clean_tctx(struct io_uring_task *tctx)
9703 struct io_wq *wq = tctx->io_wq;
9704 struct io_tctx_node *node;
9705 unsigned long index;
9707 xa_for_each(&tctx->xa, index, node) {
9708 io_uring_del_tctx_node(index);
9713 * Must be after io_uring_del_task_file() (removes nodes under
9714 * uring_lock) to avoid race with io_uring_try_cancel_iowq().
9716 io_wq_put_and_exit(wq);
9721 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
9724 return atomic_read(&tctx->inflight_tracked);
9725 return percpu_counter_sum(&tctx->inflight);
9728 static void io_uring_drop_tctx_refs(struct task_struct *task)
9730 struct io_uring_task *tctx = task->io_uring;
9731 unsigned int refs = tctx->cached_refs;
9734 tctx->cached_refs = 0;
9735 percpu_counter_sub(&tctx->inflight, refs);
9736 put_task_struct_many(task, refs);
9741 * Find any io_uring ctx that this task has registered or done IO on, and cancel
9742 * requests. @sqd should be not-null IIF it's an SQPOLL thread cancellation.
9744 static void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
9746 struct io_uring_task *tctx = current->io_uring;
9747 struct io_ring_ctx *ctx;
9751 WARN_ON_ONCE(sqd && sqd->thread != current);
9753 if (!current->io_uring)
9756 io_wq_exit_start(tctx->io_wq);
9758 atomic_inc(&tctx->in_idle);
9760 io_uring_drop_tctx_refs(current);
9761 /* read completions before cancelations */
9762 inflight = tctx_inflight(tctx, !cancel_all);
9767 struct io_tctx_node *node;
9768 unsigned long index;
9770 xa_for_each(&tctx->xa, index, node) {
9771 /* sqpoll task will cancel all its requests */
9772 if (node->ctx->sq_data)
9774 io_uring_try_cancel_requests(node->ctx, current,
9778 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
9779 io_uring_try_cancel_requests(ctx, current,
9783 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
9784 io_uring_drop_tctx_refs(current);
9786 * If we've seen completions, retry without waiting. This
9787 * avoids a race where a completion comes in before we did
9788 * prepare_to_wait().
9790 if (inflight == tctx_inflight(tctx, !cancel_all))
9792 finish_wait(&tctx->wait, &wait);
9794 atomic_dec(&tctx->in_idle);
9796 io_uring_clean_tctx(tctx);
9798 /* for exec all current's requests should be gone, kill tctx */
9799 __io_uring_free(current);
9803 void __io_uring_cancel(bool cancel_all)
9805 io_uring_cancel_generic(cancel_all, NULL);
9808 static void *io_uring_validate_mmap_request(struct file *file,
9809 loff_t pgoff, size_t sz)
9811 struct io_ring_ctx *ctx = file->private_data;
9812 loff_t offset = pgoff << PAGE_SHIFT;
9817 case IORING_OFF_SQ_RING:
9818 case IORING_OFF_CQ_RING:
9821 case IORING_OFF_SQES:
9825 return ERR_PTR(-EINVAL);
9828 page = virt_to_head_page(ptr);
9829 if (sz > page_size(page))
9830 return ERR_PTR(-EINVAL);
9837 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9839 size_t sz = vma->vm_end - vma->vm_start;
9843 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
9845 return PTR_ERR(ptr);
9847 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
9848 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
9851 #else /* !CONFIG_MMU */
9853 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9855 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
9858 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
9860 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
9863 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
9864 unsigned long addr, unsigned long len,
9865 unsigned long pgoff, unsigned long flags)
9869 ptr = io_uring_validate_mmap_request(file, pgoff, len);
9871 return PTR_ERR(ptr);
9873 return (unsigned long) ptr;
9876 #endif /* !CONFIG_MMU */
9878 static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
9883 if (!io_sqring_full(ctx))
9885 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9887 if (!io_sqring_full(ctx))
9890 } while (!signal_pending(current));
9892 finish_wait(&ctx->sqo_sq_wait, &wait);
9896 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9897 struct __kernel_timespec __user **ts,
9898 const sigset_t __user **sig)
9900 struct io_uring_getevents_arg arg;
9903 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9904 * is just a pointer to the sigset_t.
9906 if (!(flags & IORING_ENTER_EXT_ARG)) {
9907 *sig = (const sigset_t __user *) argp;
9913 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9914 * timespec and sigset_t pointers if good.
9916 if (*argsz != sizeof(arg))
9918 if (copy_from_user(&arg, argp, sizeof(arg)))
9920 *sig = u64_to_user_ptr(arg.sigmask);
9921 *argsz = arg.sigmask_sz;
9922 *ts = u64_to_user_ptr(arg.ts);
9926 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9927 u32, min_complete, u32, flags, const void __user *, argp,
9930 struct io_ring_ctx *ctx;
9937 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9938 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG)))
9942 if (unlikely(!f.file))
9946 if (unlikely(f.file->f_op != &io_uring_fops))
9950 ctx = f.file->private_data;
9951 if (unlikely(!percpu_ref_tryget(&ctx->refs)))
9955 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
9959 * For SQ polling, the thread will do all submissions and completions.
9960 * Just return the requested submit count, and wake the thread if
9964 if (ctx->flags & IORING_SETUP_SQPOLL) {
9965 io_cqring_overflow_flush(ctx);
9967 if (unlikely(ctx->sq_data->thread == NULL)) {
9971 if (flags & IORING_ENTER_SQ_WAKEUP)
9972 wake_up(&ctx->sq_data->wait);
9973 if (flags & IORING_ENTER_SQ_WAIT) {
9974 ret = io_sqpoll_wait_sq(ctx);
9978 submitted = to_submit;
9979 } else if (to_submit) {
9980 ret = io_uring_add_tctx_node(ctx);
9983 mutex_lock(&ctx->uring_lock);
9984 submitted = io_submit_sqes(ctx, to_submit);
9985 mutex_unlock(&ctx->uring_lock);
9987 if (submitted != to_submit)
9990 if (flags & IORING_ENTER_GETEVENTS) {
9991 const sigset_t __user *sig;
9992 struct __kernel_timespec __user *ts;
9994 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
9998 min_complete = min(min_complete, ctx->cq_entries);
10001 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
10002 * space applications don't need to do io completion events
10003 * polling again, they can rely on io_sq_thread to do polling
10004 * work, which can reduce cpu usage and uring_lock contention.
10006 if (ctx->flags & IORING_SETUP_IOPOLL &&
10007 !(ctx->flags & IORING_SETUP_SQPOLL)) {
10008 ret = io_iopoll_check(ctx, min_complete);
10010 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
10015 percpu_ref_put(&ctx->refs);
10018 return submitted ? submitted : ret;
10021 #ifdef CONFIG_PROC_FS
10022 static int io_uring_show_cred(struct seq_file *m, unsigned int id,
10023 const struct cred *cred)
10025 struct user_namespace *uns = seq_user_ns(m);
10026 struct group_info *gi;
10031 seq_printf(m, "%5d\n", id);
10032 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
10033 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
10034 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
10035 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
10036 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
10037 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
10038 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
10039 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
10040 seq_puts(m, "\n\tGroups:\t");
10041 gi = cred->group_info;
10042 for (g = 0; g < gi->ngroups; g++) {
10043 seq_put_decimal_ull(m, g ? " " : "",
10044 from_kgid_munged(uns, gi->gid[g]));
10046 seq_puts(m, "\n\tCapEff:\t");
10047 cap = cred->cap_effective;
10048 CAP_FOR_EACH_U32(__capi)
10049 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
10054 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
10056 struct io_sq_data *sq = NULL;
10061 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
10062 * since fdinfo case grabs it in the opposite direction of normal use
10063 * cases. If we fail to get the lock, we just don't iterate any
10064 * structures that could be going away outside the io_uring mutex.
10066 has_lock = mutex_trylock(&ctx->uring_lock);
10068 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) {
10074 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
10075 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
10076 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
10077 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
10078 struct file *f = io_file_from_index(ctx, i);
10081 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
10083 seq_printf(m, "%5u: <none>\n", i);
10085 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
10086 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
10087 struct io_mapped_ubuf *buf = ctx->user_bufs[i];
10088 unsigned int len = buf->ubuf_end - buf->ubuf;
10090 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf, len);
10092 if (has_lock && !xa_empty(&ctx->personalities)) {
10093 unsigned long index;
10094 const struct cred *cred;
10096 seq_printf(m, "Personalities:\n");
10097 xa_for_each(&ctx->personalities, index, cred)
10098 io_uring_show_cred(m, index, cred);
10100 seq_printf(m, "PollList:\n");
10101 spin_lock(&ctx->completion_lock);
10102 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
10103 struct hlist_head *list = &ctx->cancel_hash[i];
10104 struct io_kiocb *req;
10106 hlist_for_each_entry(req, list, hash_node)
10107 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
10108 req->task->task_works != NULL);
10110 spin_unlock(&ctx->completion_lock);
10112 mutex_unlock(&ctx->uring_lock);
10115 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
10117 struct io_ring_ctx *ctx = f->private_data;
10119 if (percpu_ref_tryget(&ctx->refs)) {
10120 __io_uring_show_fdinfo(ctx, m);
10121 percpu_ref_put(&ctx->refs);
10126 static const struct file_operations io_uring_fops = {
10127 .release = io_uring_release,
10128 .mmap = io_uring_mmap,
10130 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
10131 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
10133 .poll = io_uring_poll,
10134 #ifdef CONFIG_PROC_FS
10135 .show_fdinfo = io_uring_show_fdinfo,
10139 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
10140 struct io_uring_params *p)
10142 struct io_rings *rings;
10143 size_t size, sq_array_offset;
10145 /* make sure these are sane, as we already accounted them */
10146 ctx->sq_entries = p->sq_entries;
10147 ctx->cq_entries = p->cq_entries;
10149 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
10150 if (size == SIZE_MAX)
10153 rings = io_mem_alloc(size);
10157 ctx->rings = rings;
10158 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
10159 rings->sq_ring_mask = p->sq_entries - 1;
10160 rings->cq_ring_mask = p->cq_entries - 1;
10161 rings->sq_ring_entries = p->sq_entries;
10162 rings->cq_ring_entries = p->cq_entries;
10164 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
10165 if (size == SIZE_MAX) {
10166 io_mem_free(ctx->rings);
10171 ctx->sq_sqes = io_mem_alloc(size);
10172 if (!ctx->sq_sqes) {
10173 io_mem_free(ctx->rings);
10181 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
10185 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
10189 ret = io_uring_add_tctx_node(ctx);
10194 fd_install(fd, file);
10199 * Allocate an anonymous fd, this is what constitutes the application
10200 * visible backing of an io_uring instance. The application mmaps this
10201 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
10202 * we have to tie this fd to a socket for file garbage collection purposes.
10204 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
10207 #if defined(CONFIG_UNIX)
10210 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
10213 return ERR_PTR(ret);
10216 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
10217 O_RDWR | O_CLOEXEC);
10218 #if defined(CONFIG_UNIX)
10219 if (IS_ERR(file)) {
10220 sock_release(ctx->ring_sock);
10221 ctx->ring_sock = NULL;
10223 ctx->ring_sock->file = file;
10229 static int io_uring_create(unsigned entries, struct io_uring_params *p,
10230 struct io_uring_params __user *params)
10232 struct io_ring_ctx *ctx;
10238 if (entries > IORING_MAX_ENTRIES) {
10239 if (!(p->flags & IORING_SETUP_CLAMP))
10241 entries = IORING_MAX_ENTRIES;
10245 * Use twice as many entries for the CQ ring. It's possible for the
10246 * application to drive a higher depth than the size of the SQ ring,
10247 * since the sqes are only used at submission time. This allows for
10248 * some flexibility in overcommitting a bit. If the application has
10249 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
10250 * of CQ ring entries manually.
10252 p->sq_entries = roundup_pow_of_two(entries);
10253 if (p->flags & IORING_SETUP_CQSIZE) {
10255 * If IORING_SETUP_CQSIZE is set, we do the same roundup
10256 * to a power-of-two, if it isn't already. We do NOT impose
10257 * any cq vs sq ring sizing.
10259 if (!p->cq_entries)
10261 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
10262 if (!(p->flags & IORING_SETUP_CLAMP))
10264 p->cq_entries = IORING_MAX_CQ_ENTRIES;
10266 p->cq_entries = roundup_pow_of_two(p->cq_entries);
10267 if (p->cq_entries < p->sq_entries)
10270 p->cq_entries = 2 * p->sq_entries;
10273 ctx = io_ring_ctx_alloc(p);
10276 ctx->compat = in_compat_syscall();
10277 if (!capable(CAP_IPC_LOCK))
10278 ctx->user = get_uid(current_user());
10281 * This is just grabbed for accounting purposes. When a process exits,
10282 * the mm is exited and dropped before the files, hence we need to hang
10283 * on to this mm purely for the purposes of being able to unaccount
10284 * memory (locked/pinned vm). It's not used for anything else.
10286 mmgrab(current->mm);
10287 ctx->mm_account = current->mm;
10289 ret = io_allocate_scq_urings(ctx, p);
10293 ret = io_sq_offload_create(ctx, p);
10296 /* always set a rsrc node */
10297 ret = io_rsrc_node_switch_start(ctx);
10300 io_rsrc_node_switch(ctx, NULL);
10302 memset(&p->sq_off, 0, sizeof(p->sq_off));
10303 p->sq_off.head = offsetof(struct io_rings, sq.head);
10304 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
10305 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
10306 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
10307 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
10308 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
10309 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
10311 memset(&p->cq_off, 0, sizeof(p->cq_off));
10312 p->cq_off.head = offsetof(struct io_rings, cq.head);
10313 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
10314 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
10315 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
10316 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
10317 p->cq_off.cqes = offsetof(struct io_rings, cqes);
10318 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
10320 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
10321 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
10322 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
10323 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
10324 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
10325 IORING_FEAT_RSRC_TAGS;
10327 if (copy_to_user(params, p, sizeof(*p))) {
10332 file = io_uring_get_file(ctx);
10333 if (IS_ERR(file)) {
10334 ret = PTR_ERR(file);
10339 * Install ring fd as the very last thing, so we don't risk someone
10340 * having closed it before we finish setup
10342 ret = io_uring_install_fd(ctx, file);
10344 /* fput will clean it up */
10349 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
10352 io_ring_ctx_wait_and_kill(ctx);
10357 * Sets up an aio uring context, and returns the fd. Applications asks for a
10358 * ring size, we return the actual sq/cq ring sizes (among other things) in the
10359 * params structure passed in.
10361 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
10363 struct io_uring_params p;
10366 if (copy_from_user(&p, params, sizeof(p)))
10368 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
10373 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
10374 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
10375 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
10376 IORING_SETUP_R_DISABLED))
10379 return io_uring_create(entries, &p, params);
10382 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
10383 struct io_uring_params __user *, params)
10385 return io_uring_setup(entries, params);
10388 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
10390 struct io_uring_probe *p;
10394 size = struct_size(p, ops, nr_args);
10395 if (size == SIZE_MAX)
10397 p = kzalloc(size, GFP_KERNEL);
10402 if (copy_from_user(p, arg, size))
10405 if (memchr_inv(p, 0, size))
10408 p->last_op = IORING_OP_LAST - 1;
10409 if (nr_args > IORING_OP_LAST)
10410 nr_args = IORING_OP_LAST;
10412 for (i = 0; i < nr_args; i++) {
10414 if (!io_op_defs[i].not_supported)
10415 p->ops[i].flags = IO_URING_OP_SUPPORTED;
10420 if (copy_to_user(arg, p, size))
10427 static int io_register_personality(struct io_ring_ctx *ctx)
10429 const struct cred *creds;
10433 creds = get_current_cred();
10435 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
10436 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
10444 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
10445 unsigned int nr_args)
10447 struct io_uring_restriction *res;
10451 /* Restrictions allowed only if rings started disabled */
10452 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
10455 /* We allow only a single restrictions registration */
10456 if (ctx->restrictions.registered)
10459 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
10462 size = array_size(nr_args, sizeof(*res));
10463 if (size == SIZE_MAX)
10466 res = memdup_user(arg, size);
10468 return PTR_ERR(res);
10472 for (i = 0; i < nr_args; i++) {
10473 switch (res[i].opcode) {
10474 case IORING_RESTRICTION_REGISTER_OP:
10475 if (res[i].register_op >= IORING_REGISTER_LAST) {
10480 __set_bit(res[i].register_op,
10481 ctx->restrictions.register_op);
10483 case IORING_RESTRICTION_SQE_OP:
10484 if (res[i].sqe_op >= IORING_OP_LAST) {
10489 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
10491 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
10492 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
10494 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
10495 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
10504 /* Reset all restrictions if an error happened */
10506 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
10508 ctx->restrictions.registered = true;
10514 static int io_register_enable_rings(struct io_ring_ctx *ctx)
10516 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
10519 if (ctx->restrictions.registered)
10520 ctx->restricted = 1;
10522 ctx->flags &= ~IORING_SETUP_R_DISABLED;
10523 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
10524 wake_up(&ctx->sq_data->wait);
10528 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
10529 struct io_uring_rsrc_update2 *up,
10537 if (check_add_overflow(up->offset, nr_args, &tmp))
10539 err = io_rsrc_node_switch_start(ctx);
10544 case IORING_RSRC_FILE:
10545 return __io_sqe_files_update(ctx, up, nr_args);
10546 case IORING_RSRC_BUFFER:
10547 return __io_sqe_buffers_update(ctx, up, nr_args);
10552 static int io_register_files_update(struct io_ring_ctx *ctx, void __user *arg,
10555 struct io_uring_rsrc_update2 up;
10559 memset(&up, 0, sizeof(up));
10560 if (copy_from_user(&up, arg, sizeof(struct io_uring_rsrc_update)))
10562 return __io_register_rsrc_update(ctx, IORING_RSRC_FILE, &up, nr_args);
10565 static int io_register_rsrc_update(struct io_ring_ctx *ctx, void __user *arg,
10566 unsigned size, unsigned type)
10568 struct io_uring_rsrc_update2 up;
10570 if (size != sizeof(up))
10572 if (copy_from_user(&up, arg, sizeof(up)))
10574 if (!up.nr || up.resv)
10576 return __io_register_rsrc_update(ctx, type, &up, up.nr);
10579 static int io_register_rsrc(struct io_ring_ctx *ctx, void __user *arg,
10580 unsigned int size, unsigned int type)
10582 struct io_uring_rsrc_register rr;
10584 /* keep it extendible */
10585 if (size != sizeof(rr))
10588 memset(&rr, 0, sizeof(rr));
10589 if (copy_from_user(&rr, arg, size))
10591 if (!rr.nr || rr.resv || rr.resv2)
10595 case IORING_RSRC_FILE:
10596 return io_sqe_files_register(ctx, u64_to_user_ptr(rr.data),
10597 rr.nr, u64_to_user_ptr(rr.tags));
10598 case IORING_RSRC_BUFFER:
10599 return io_sqe_buffers_register(ctx, u64_to_user_ptr(rr.data),
10600 rr.nr, u64_to_user_ptr(rr.tags));
10605 static int io_register_iowq_aff(struct io_ring_ctx *ctx, void __user *arg,
10608 struct io_uring_task *tctx = current->io_uring;
10609 cpumask_var_t new_mask;
10612 if (!tctx || !tctx->io_wq)
10615 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
10618 cpumask_clear(new_mask);
10619 if (len > cpumask_size())
10620 len = cpumask_size();
10622 if (copy_from_user(new_mask, arg, len)) {
10623 free_cpumask_var(new_mask);
10627 ret = io_wq_cpu_affinity(tctx->io_wq, new_mask);
10628 free_cpumask_var(new_mask);
10632 static int io_unregister_iowq_aff(struct io_ring_ctx *ctx)
10634 struct io_uring_task *tctx = current->io_uring;
10636 if (!tctx || !tctx->io_wq)
10639 return io_wq_cpu_affinity(tctx->io_wq, NULL);
10642 static int io_register_iowq_max_workers(struct io_ring_ctx *ctx,
10645 struct io_uring_task *tctx = NULL;
10646 struct io_sq_data *sqd = NULL;
10647 __u32 new_count[2];
10650 if (copy_from_user(new_count, arg, sizeof(new_count)))
10652 for (i = 0; i < ARRAY_SIZE(new_count); i++)
10653 if (new_count[i] > INT_MAX)
10656 if (ctx->flags & IORING_SETUP_SQPOLL) {
10657 sqd = ctx->sq_data;
10660 * Observe the correct sqd->lock -> ctx->uring_lock
10661 * ordering. Fine to drop uring_lock here, we hold
10662 * a ref to the ctx.
10664 refcount_inc(&sqd->refs);
10665 mutex_unlock(&ctx->uring_lock);
10666 mutex_lock(&sqd->lock);
10667 mutex_lock(&ctx->uring_lock);
10669 tctx = sqd->thread->io_uring;
10672 tctx = current->io_uring;
10676 if (!tctx || !tctx->io_wq)
10679 ret = io_wq_max_workers(tctx->io_wq, new_count);
10684 mutex_unlock(&sqd->lock);
10685 io_put_sq_data(sqd);
10688 if (copy_to_user(arg, new_count, sizeof(new_count)))
10694 mutex_unlock(&sqd->lock);
10695 io_put_sq_data(sqd);
10700 static bool io_register_op_must_quiesce(int op)
10703 case IORING_REGISTER_BUFFERS:
10704 case IORING_UNREGISTER_BUFFERS:
10705 case IORING_REGISTER_FILES:
10706 case IORING_UNREGISTER_FILES:
10707 case IORING_REGISTER_FILES_UPDATE:
10708 case IORING_REGISTER_PROBE:
10709 case IORING_REGISTER_PERSONALITY:
10710 case IORING_UNREGISTER_PERSONALITY:
10711 case IORING_REGISTER_FILES2:
10712 case IORING_REGISTER_FILES_UPDATE2:
10713 case IORING_REGISTER_BUFFERS2:
10714 case IORING_REGISTER_BUFFERS_UPDATE:
10715 case IORING_REGISTER_IOWQ_AFF:
10716 case IORING_UNREGISTER_IOWQ_AFF:
10717 case IORING_REGISTER_IOWQ_MAX_WORKERS:
10724 static int io_ctx_quiesce(struct io_ring_ctx *ctx)
10728 percpu_ref_kill(&ctx->refs);
10731 * Drop uring mutex before waiting for references to exit. If another
10732 * thread is currently inside io_uring_enter() it might need to grab the
10733 * uring_lock to make progress. If we hold it here across the drain
10734 * wait, then we can deadlock. It's safe to drop the mutex here, since
10735 * no new references will come in after we've killed the percpu ref.
10737 mutex_unlock(&ctx->uring_lock);
10739 ret = wait_for_completion_interruptible(&ctx->ref_comp);
10742 ret = io_run_task_work_sig();
10743 } while (ret >= 0);
10744 mutex_lock(&ctx->uring_lock);
10747 io_refs_resurrect(&ctx->refs, &ctx->ref_comp);
10751 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
10752 void __user *arg, unsigned nr_args)
10753 __releases(ctx->uring_lock)
10754 __acquires(ctx->uring_lock)
10759 * We're inside the ring mutex, if the ref is already dying, then
10760 * someone else killed the ctx or is already going through
10761 * io_uring_register().
10763 if (percpu_ref_is_dying(&ctx->refs))
10766 if (ctx->restricted) {
10767 if (opcode >= IORING_REGISTER_LAST)
10769 opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
10770 if (!test_bit(opcode, ctx->restrictions.register_op))
10774 if (io_register_op_must_quiesce(opcode)) {
10775 ret = io_ctx_quiesce(ctx);
10781 case IORING_REGISTER_BUFFERS:
10782 ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
10784 case IORING_UNREGISTER_BUFFERS:
10786 if (arg || nr_args)
10788 ret = io_sqe_buffers_unregister(ctx);
10790 case IORING_REGISTER_FILES:
10791 ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
10793 case IORING_UNREGISTER_FILES:
10795 if (arg || nr_args)
10797 ret = io_sqe_files_unregister(ctx);
10799 case IORING_REGISTER_FILES_UPDATE:
10800 ret = io_register_files_update(ctx, arg, nr_args);
10802 case IORING_REGISTER_EVENTFD:
10803 case IORING_REGISTER_EVENTFD_ASYNC:
10807 ret = io_eventfd_register(ctx, arg);
10810 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
10811 ctx->eventfd_async = 1;
10813 ctx->eventfd_async = 0;
10815 case IORING_UNREGISTER_EVENTFD:
10817 if (arg || nr_args)
10819 ret = io_eventfd_unregister(ctx);
10821 case IORING_REGISTER_PROBE:
10823 if (!arg || nr_args > 256)
10825 ret = io_probe(ctx, arg, nr_args);
10827 case IORING_REGISTER_PERSONALITY:
10829 if (arg || nr_args)
10831 ret = io_register_personality(ctx);
10833 case IORING_UNREGISTER_PERSONALITY:
10837 ret = io_unregister_personality(ctx, nr_args);
10839 case IORING_REGISTER_ENABLE_RINGS:
10841 if (arg || nr_args)
10843 ret = io_register_enable_rings(ctx);
10845 case IORING_REGISTER_RESTRICTIONS:
10846 ret = io_register_restrictions(ctx, arg, nr_args);
10848 case IORING_REGISTER_FILES2:
10849 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE);
10851 case IORING_REGISTER_FILES_UPDATE2:
10852 ret = io_register_rsrc_update(ctx, arg, nr_args,
10855 case IORING_REGISTER_BUFFERS2:
10856 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER);
10858 case IORING_REGISTER_BUFFERS_UPDATE:
10859 ret = io_register_rsrc_update(ctx, arg, nr_args,
10860 IORING_RSRC_BUFFER);
10862 case IORING_REGISTER_IOWQ_AFF:
10864 if (!arg || !nr_args)
10866 ret = io_register_iowq_aff(ctx, arg, nr_args);
10868 case IORING_UNREGISTER_IOWQ_AFF:
10870 if (arg || nr_args)
10872 ret = io_unregister_iowq_aff(ctx);
10874 case IORING_REGISTER_IOWQ_MAX_WORKERS:
10876 if (!arg || nr_args != 2)
10878 ret = io_register_iowq_max_workers(ctx, arg);
10885 if (io_register_op_must_quiesce(opcode)) {
10886 /* bring the ctx back to life */
10887 percpu_ref_reinit(&ctx->refs);
10888 reinit_completion(&ctx->ref_comp);
10893 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
10894 void __user *, arg, unsigned int, nr_args)
10896 struct io_ring_ctx *ctx;
10905 if (f.file->f_op != &io_uring_fops)
10908 ctx = f.file->private_data;
10910 io_run_task_work();
10912 mutex_lock(&ctx->uring_lock);
10913 ret = __io_uring_register(ctx, opcode, arg, nr_args);
10914 mutex_unlock(&ctx->uring_lock);
10915 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
10916 ctx->cq_ev_fd != NULL, ret);
10922 static int __init io_uring_init(void)
10924 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
10925 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
10926 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
10929 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
10930 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
10931 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
10932 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
10933 BUILD_BUG_SQE_ELEM(1, __u8, flags);
10934 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
10935 BUILD_BUG_SQE_ELEM(4, __s32, fd);
10936 BUILD_BUG_SQE_ELEM(8, __u64, off);
10937 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
10938 BUILD_BUG_SQE_ELEM(16, __u64, addr);
10939 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
10940 BUILD_BUG_SQE_ELEM(24, __u32, len);
10941 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
10942 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
10943 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
10944 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
10945 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
10946 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
10947 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
10948 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
10949 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
10950 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
10951 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
10952 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
10953 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
10954 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
10955 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
10956 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
10957 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
10958 BUILD_BUG_SQE_ELEM(40, __u16, buf_group);
10959 BUILD_BUG_SQE_ELEM(42, __u16, personality);
10960 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
10961 BUILD_BUG_SQE_ELEM(44, __u32, file_index);
10963 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
10964 sizeof(struct io_uring_rsrc_update));
10965 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
10966 sizeof(struct io_uring_rsrc_update2));
10968 /* ->buf_index is u16 */
10969 BUILD_BUG_ON(IORING_MAX_REG_BUFFERS >= (1u << 16));
10971 /* should fit into one byte */
10972 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
10974 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
10975 BUILD_BUG_ON(__REQ_F_LAST_BIT > 8 * sizeof(int));
10977 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
10981 __initcall(io_uring_init);