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
95 /* 512 entries per page on 64-bit archs, 64 pages max */
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 9
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_IOPOLL_BATCH 8
299 #define IO_COMPL_BATCH 32
300 #define IO_REQ_CACHE_SIZE 32
301 #define IO_REQ_ALLOC_BATCH 8
303 struct io_comp_state {
304 struct io_kiocb *reqs[IO_COMPL_BATCH];
306 /* inline/task_work completion list, under ->uring_lock */
307 struct list_head free_list;
310 struct io_submit_link {
311 struct io_kiocb *head;
312 struct io_kiocb *last;
315 struct io_submit_state {
316 struct blk_plug plug;
317 struct io_submit_link link;
320 * io_kiocb alloc cache
322 void *reqs[IO_REQ_CACHE_SIZE];
323 unsigned int free_reqs;
328 * Batch completion logic
330 struct io_comp_state comp;
333 * File reference cache
337 unsigned int file_refs;
338 unsigned int ios_left;
342 /* const or read-mostly hot data */
344 struct percpu_ref refs;
346 struct io_rings *rings;
348 unsigned int compat: 1;
349 unsigned int drain_next: 1;
350 unsigned int eventfd_async: 1;
351 unsigned int restricted: 1;
352 unsigned int off_timeout_used: 1;
353 unsigned int drain_active: 1;
354 } ____cacheline_aligned_in_smp;
356 /* submission data */
358 struct mutex uring_lock;
361 * Ring buffer of indices into array of io_uring_sqe, which is
362 * mmapped by the application using the IORING_OFF_SQES offset.
364 * This indirection could e.g. be used to assign fixed
365 * io_uring_sqe entries to operations and only submit them to
366 * the queue when needed.
368 * The kernel modifies neither the indices array nor the entries
372 struct io_uring_sqe *sq_sqes;
373 unsigned cached_sq_head;
375 struct list_head defer_list;
378 * Fixed resources fast path, should be accessed only under
379 * uring_lock, and updated through io_uring_register(2)
381 struct io_rsrc_node *rsrc_node;
382 struct io_file_table file_table;
383 unsigned nr_user_files;
384 unsigned nr_user_bufs;
385 struct io_mapped_ubuf **user_bufs;
387 struct io_submit_state submit_state;
388 struct list_head timeout_list;
389 struct list_head cq_overflow_list;
390 struct xarray io_buffers;
391 struct xarray personalities;
393 unsigned sq_thread_idle;
394 } ____cacheline_aligned_in_smp;
396 /* IRQ completion list, under ->completion_lock */
397 struct list_head locked_free_list;
398 unsigned int locked_free_nr;
400 const struct cred *sq_creds; /* cred used for __io_sq_thread() */
401 struct io_sq_data *sq_data; /* if using sq thread polling */
403 struct wait_queue_head sqo_sq_wait;
404 struct list_head sqd_list;
406 unsigned long check_cq_overflow;
409 unsigned cached_cq_tail;
411 struct eventfd_ctx *cq_ev_fd;
412 struct wait_queue_head poll_wait;
413 struct wait_queue_head cq_wait;
415 atomic_t cq_timeouts;
416 struct fasync_struct *cq_fasync;
417 unsigned cq_last_tm_flush;
418 } ____cacheline_aligned_in_smp;
421 spinlock_t completion_lock;
424 * ->iopoll_list is protected by the ctx->uring_lock for
425 * io_uring instances that don't use IORING_SETUP_SQPOLL.
426 * For SQPOLL, only the single threaded io_sq_thread() will
427 * manipulate the list, hence no extra locking is needed there.
429 struct list_head iopoll_list;
430 struct hlist_head *cancel_hash;
431 unsigned cancel_hash_bits;
432 bool poll_multi_queue;
433 } ____cacheline_aligned_in_smp;
435 struct io_restriction restrictions;
437 /* slow path rsrc auxilary data, used by update/register */
439 struct io_rsrc_node *rsrc_backup_node;
440 struct io_mapped_ubuf *dummy_ubuf;
441 struct io_rsrc_data *file_data;
442 struct io_rsrc_data *buf_data;
444 struct delayed_work rsrc_put_work;
445 struct llist_head rsrc_put_llist;
446 struct list_head rsrc_ref_list;
447 spinlock_t rsrc_ref_lock;
450 /* Keep this last, we don't need it for the fast path */
452 #if defined(CONFIG_UNIX)
453 struct socket *ring_sock;
455 /* hashed buffered write serialization */
456 struct io_wq_hash *hash_map;
458 /* Only used for accounting purposes */
459 struct user_struct *user;
460 struct mm_struct *mm_account;
462 /* ctx exit and cancelation */
463 struct llist_head fallback_llist;
464 struct delayed_work fallback_work;
465 struct work_struct exit_work;
466 struct list_head tctx_list;
467 struct completion ref_comp;
471 struct io_uring_task {
472 /* submission side */
475 struct wait_queue_head wait;
476 const struct io_ring_ctx *last;
478 struct percpu_counter inflight;
479 atomic_t inflight_tracked;
482 spinlock_t task_lock;
483 struct io_wq_work_list task_list;
484 unsigned long task_state;
485 struct callback_head task_work;
489 * First field must be the file pointer in all the
490 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
492 struct io_poll_iocb {
494 struct wait_queue_head *head;
498 struct wait_queue_entry wait;
501 struct io_poll_update {
507 bool update_user_data;
515 struct io_timeout_data {
516 struct io_kiocb *req;
517 struct hrtimer timer;
518 struct timespec64 ts;
519 enum hrtimer_mode mode;
524 struct sockaddr __user *addr;
525 int __user *addr_len;
527 unsigned long nofile;
547 struct list_head list;
548 /* head of the link, used by linked timeouts only */
549 struct io_kiocb *head;
552 struct io_timeout_rem {
557 struct timespec64 ts;
562 /* NOTE: kiocb has the file as the first member, so don't do it here */
570 struct sockaddr __user *addr;
577 struct compat_msghdr __user *umsg_compat;
578 struct user_msghdr __user *umsg;
584 struct io_buffer *kbuf;
590 struct filename *filename;
592 unsigned long nofile;
595 struct io_rsrc_update {
621 struct epoll_event event;
625 struct file *file_out;
626 struct file *file_in;
633 struct io_provide_buf {
647 const char __user *filename;
648 struct statx __user *buffer;
660 struct filename *oldpath;
661 struct filename *newpath;
669 struct filename *filename;
672 struct io_completion {
674 struct list_head list;
678 struct io_async_connect {
679 struct sockaddr_storage address;
682 struct io_async_msghdr {
683 struct iovec fast_iov[UIO_FASTIOV];
684 /* points to an allocated iov, if NULL we use fast_iov instead */
685 struct iovec *free_iov;
686 struct sockaddr __user *uaddr;
688 struct sockaddr_storage addr;
692 struct iovec fast_iov[UIO_FASTIOV];
693 const struct iovec *free_iovec;
694 struct iov_iter iter;
696 struct wait_page_queue wpq;
700 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
701 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
702 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
703 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
704 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
705 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
707 /* first byte is taken by user flags, shift it to not overlap */
712 REQ_F_LINK_TIMEOUT_BIT,
713 REQ_F_NEED_CLEANUP_BIT,
715 REQ_F_BUFFER_SELECTED_BIT,
716 REQ_F_LTIMEOUT_ACTIVE_BIT,
717 REQ_F_COMPLETE_INLINE_BIT,
719 REQ_F_DONT_REISSUE_BIT,
721 /* keep async read/write and isreg together and in order */
722 REQ_F_NOWAIT_READ_BIT,
723 REQ_F_NOWAIT_WRITE_BIT,
726 /* not a real bit, just to check we're not overflowing the space */
732 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
733 /* drain existing IO first */
734 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
736 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
737 /* doesn't sever on completion < 0 */
738 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
740 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
741 /* IOSQE_BUFFER_SELECT */
742 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
744 /* fail rest of links */
745 REQ_F_FAIL = BIT(REQ_F_FAIL_BIT),
746 /* on inflight list, should be cancelled and waited on exit reliably */
747 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
748 /* read/write uses file position */
749 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
750 /* must not punt to workers */
751 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
752 /* has or had linked timeout */
753 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
755 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
756 /* already went through poll handler */
757 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
758 /* buffer already selected */
759 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
760 /* linked timeout is active, i.e. prepared by link's head */
761 REQ_F_LTIMEOUT_ACTIVE = BIT(REQ_F_LTIMEOUT_ACTIVE_BIT),
762 /* completion is deferred through io_comp_state */
763 REQ_F_COMPLETE_INLINE = BIT(REQ_F_COMPLETE_INLINE_BIT),
764 /* caller should reissue async */
765 REQ_F_REISSUE = BIT(REQ_F_REISSUE_BIT),
766 /* don't attempt request reissue, see io_rw_reissue() */
767 REQ_F_DONT_REISSUE = BIT(REQ_F_DONT_REISSUE_BIT),
768 /* supports async reads */
769 REQ_F_NOWAIT_READ = BIT(REQ_F_NOWAIT_READ_BIT),
770 /* supports async writes */
771 REQ_F_NOWAIT_WRITE = BIT(REQ_F_NOWAIT_WRITE_BIT),
773 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
774 /* has creds assigned */
775 REQ_F_CREDS = BIT(REQ_F_CREDS_BIT),
779 struct io_poll_iocb poll;
780 struct io_poll_iocb *double_poll;
783 typedef void (*io_req_tw_func_t)(struct io_kiocb *req);
785 struct io_task_work {
787 struct io_wq_work_node node;
788 struct llist_node fallback_node;
790 io_req_tw_func_t func;
794 IORING_RSRC_FILE = 0,
795 IORING_RSRC_BUFFER = 1,
799 * NOTE! Each of the iocb union members has the file pointer
800 * as the first entry in their struct definition. So you can
801 * access the file pointer through any of the sub-structs,
802 * or directly as just 'ki_filp' in this struct.
808 struct io_poll_iocb poll;
809 struct io_poll_update poll_update;
810 struct io_accept accept;
812 struct io_cancel cancel;
813 struct io_timeout timeout;
814 struct io_timeout_rem timeout_rem;
815 struct io_connect connect;
816 struct io_sr_msg sr_msg;
818 struct io_close close;
819 struct io_rsrc_update rsrc_update;
820 struct io_fadvise fadvise;
821 struct io_madvise madvise;
822 struct io_epoll epoll;
823 struct io_splice splice;
824 struct io_provide_buf pbuf;
825 struct io_statx statx;
826 struct io_shutdown shutdown;
827 struct io_rename rename;
828 struct io_unlink unlink;
829 /* use only after cleaning per-op data, see io_clean_op() */
830 struct io_completion compl;
833 /* opcode allocated if it needs to store data for async defer */
836 /* polled IO has completed */
842 struct io_ring_ctx *ctx;
845 struct task_struct *task;
848 struct io_kiocb *link;
849 struct percpu_ref *fixed_rsrc_refs;
851 /* used with ctx->iopoll_list with reads/writes */
852 struct list_head inflight_entry;
853 struct io_task_work io_task_work;
854 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
855 struct hlist_node hash_node;
856 struct async_poll *apoll;
857 struct io_wq_work work;
858 const struct cred *creds;
860 /* store used ubuf, so we can prevent reloading */
861 struct io_mapped_ubuf *imu;
864 struct io_tctx_node {
865 struct list_head ctx_node;
866 struct task_struct *task;
867 struct io_ring_ctx *ctx;
870 struct io_defer_entry {
871 struct list_head list;
872 struct io_kiocb *req;
877 /* needs req->file assigned */
878 unsigned needs_file : 1;
879 /* hash wq insertion if file is a regular file */
880 unsigned hash_reg_file : 1;
881 /* unbound wq insertion if file is a non-regular file */
882 unsigned unbound_nonreg_file : 1;
883 /* opcode is not supported by this kernel */
884 unsigned not_supported : 1;
885 /* set if opcode supports polled "wait" */
887 unsigned pollout : 1;
888 /* op supports buffer selection */
889 unsigned buffer_select : 1;
890 /* do prep async if is going to be punted */
891 unsigned needs_async_setup : 1;
892 /* should block plug */
894 /* size of async data needed, if any */
895 unsigned short async_size;
898 static const struct io_op_def io_op_defs[] = {
899 [IORING_OP_NOP] = {},
900 [IORING_OP_READV] = {
902 .unbound_nonreg_file = 1,
905 .needs_async_setup = 1,
907 .async_size = sizeof(struct io_async_rw),
909 [IORING_OP_WRITEV] = {
912 .unbound_nonreg_file = 1,
914 .needs_async_setup = 1,
916 .async_size = sizeof(struct io_async_rw),
918 [IORING_OP_FSYNC] = {
921 [IORING_OP_READ_FIXED] = {
923 .unbound_nonreg_file = 1,
926 .async_size = sizeof(struct io_async_rw),
928 [IORING_OP_WRITE_FIXED] = {
931 .unbound_nonreg_file = 1,
934 .async_size = sizeof(struct io_async_rw),
936 [IORING_OP_POLL_ADD] = {
938 .unbound_nonreg_file = 1,
940 [IORING_OP_POLL_REMOVE] = {},
941 [IORING_OP_SYNC_FILE_RANGE] = {
944 [IORING_OP_SENDMSG] = {
946 .unbound_nonreg_file = 1,
948 .needs_async_setup = 1,
949 .async_size = sizeof(struct io_async_msghdr),
951 [IORING_OP_RECVMSG] = {
953 .unbound_nonreg_file = 1,
956 .needs_async_setup = 1,
957 .async_size = sizeof(struct io_async_msghdr),
959 [IORING_OP_TIMEOUT] = {
960 .async_size = sizeof(struct io_timeout_data),
962 [IORING_OP_TIMEOUT_REMOVE] = {
963 /* used by timeout updates' prep() */
965 [IORING_OP_ACCEPT] = {
967 .unbound_nonreg_file = 1,
970 [IORING_OP_ASYNC_CANCEL] = {},
971 [IORING_OP_LINK_TIMEOUT] = {
972 .async_size = sizeof(struct io_timeout_data),
974 [IORING_OP_CONNECT] = {
976 .unbound_nonreg_file = 1,
978 .needs_async_setup = 1,
979 .async_size = sizeof(struct io_async_connect),
981 [IORING_OP_FALLOCATE] = {
984 [IORING_OP_OPENAT] = {},
985 [IORING_OP_CLOSE] = {},
986 [IORING_OP_FILES_UPDATE] = {},
987 [IORING_OP_STATX] = {},
990 .unbound_nonreg_file = 1,
994 .async_size = sizeof(struct io_async_rw),
996 [IORING_OP_WRITE] = {
998 .unbound_nonreg_file = 1,
1001 .async_size = sizeof(struct io_async_rw),
1003 [IORING_OP_FADVISE] = {
1006 [IORING_OP_MADVISE] = {},
1007 [IORING_OP_SEND] = {
1009 .unbound_nonreg_file = 1,
1012 [IORING_OP_RECV] = {
1014 .unbound_nonreg_file = 1,
1018 [IORING_OP_OPENAT2] = {
1020 [IORING_OP_EPOLL_CTL] = {
1021 .unbound_nonreg_file = 1,
1023 [IORING_OP_SPLICE] = {
1026 .unbound_nonreg_file = 1,
1028 [IORING_OP_PROVIDE_BUFFERS] = {},
1029 [IORING_OP_REMOVE_BUFFERS] = {},
1033 .unbound_nonreg_file = 1,
1035 [IORING_OP_SHUTDOWN] = {
1038 [IORING_OP_RENAMEAT] = {},
1039 [IORING_OP_UNLINKAT] = {},
1042 static bool io_disarm_next(struct io_kiocb *req);
1043 static void io_uring_del_tctx_node(unsigned long index);
1044 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
1045 struct task_struct *task,
1047 static void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd);
1048 static struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx);
1050 static bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1051 long res, unsigned int cflags);
1052 static void io_put_req(struct io_kiocb *req);
1053 static void io_put_req_deferred(struct io_kiocb *req, int nr);
1054 static void io_dismantle_req(struct io_kiocb *req);
1055 static void io_put_task(struct task_struct *task, int nr);
1056 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
1057 static void io_queue_linked_timeout(struct io_kiocb *req);
1058 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
1059 struct io_uring_rsrc_update2 *up,
1061 static void io_clean_op(struct io_kiocb *req);
1062 static struct file *io_file_get(struct io_ring_ctx *ctx,
1063 struct io_submit_state *state,
1064 struct io_kiocb *req, int fd, bool fixed);
1065 static void __io_queue_sqe(struct io_kiocb *req);
1066 static void io_rsrc_put_work(struct work_struct *work);
1068 static void io_req_task_queue(struct io_kiocb *req);
1069 static void io_submit_flush_completions(struct io_ring_ctx *ctx);
1070 static bool io_poll_remove_waitqs(struct io_kiocb *req);
1071 static int io_req_prep_async(struct io_kiocb *req);
1073 static void io_fallback_req_func(struct work_struct *unused);
1075 static struct kmem_cache *req_cachep;
1077 static const struct file_operations io_uring_fops;
1079 struct sock *io_uring_get_socket(struct file *file)
1081 #if defined(CONFIG_UNIX)
1082 if (file->f_op == &io_uring_fops) {
1083 struct io_ring_ctx *ctx = file->private_data;
1085 return ctx->ring_sock->sk;
1090 EXPORT_SYMBOL(io_uring_get_socket);
1092 #define io_for_each_link(pos, head) \
1093 for (pos = (head); pos; pos = pos->link)
1095 static inline void io_req_set_rsrc_node(struct io_kiocb *req)
1097 struct io_ring_ctx *ctx = req->ctx;
1099 if (!req->fixed_rsrc_refs) {
1100 req->fixed_rsrc_refs = &ctx->rsrc_node->refs;
1101 percpu_ref_get(req->fixed_rsrc_refs);
1105 static void io_refs_resurrect(struct percpu_ref *ref, struct completion *compl)
1107 bool got = percpu_ref_tryget(ref);
1109 /* already at zero, wait for ->release() */
1111 wait_for_completion(compl);
1112 percpu_ref_resurrect(ref);
1114 percpu_ref_put(ref);
1117 static bool io_match_task(struct io_kiocb *head, struct task_struct *task,
1120 struct io_kiocb *req;
1122 if (task && head->task != task)
1127 io_for_each_link(req, head) {
1128 if (req->flags & REQ_F_INFLIGHT)
1134 static inline void req_set_fail(struct io_kiocb *req)
1136 req->flags |= REQ_F_FAIL;
1139 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1141 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1143 complete(&ctx->ref_comp);
1146 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1148 return !req->timeout.off;
1151 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1153 struct io_ring_ctx *ctx;
1156 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1161 * Use 5 bits less than the max cq entries, that should give us around
1162 * 32 entries per hash list if totally full and uniformly spread.
1164 hash_bits = ilog2(p->cq_entries);
1168 ctx->cancel_hash_bits = hash_bits;
1169 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1171 if (!ctx->cancel_hash)
1173 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1175 ctx->dummy_ubuf = kzalloc(sizeof(*ctx->dummy_ubuf), GFP_KERNEL);
1176 if (!ctx->dummy_ubuf)
1178 /* set invalid range, so io_import_fixed() fails meeting it */
1179 ctx->dummy_ubuf->ubuf = -1UL;
1181 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1182 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1185 ctx->flags = p->flags;
1186 init_waitqueue_head(&ctx->sqo_sq_wait);
1187 INIT_LIST_HEAD(&ctx->sqd_list);
1188 init_waitqueue_head(&ctx->poll_wait);
1189 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1190 init_completion(&ctx->ref_comp);
1191 xa_init_flags(&ctx->io_buffers, XA_FLAGS_ALLOC1);
1192 xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
1193 mutex_init(&ctx->uring_lock);
1194 init_waitqueue_head(&ctx->cq_wait);
1195 spin_lock_init(&ctx->completion_lock);
1196 INIT_LIST_HEAD(&ctx->iopoll_list);
1197 INIT_LIST_HEAD(&ctx->defer_list);
1198 INIT_LIST_HEAD(&ctx->timeout_list);
1199 spin_lock_init(&ctx->rsrc_ref_lock);
1200 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
1201 INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
1202 init_llist_head(&ctx->rsrc_put_llist);
1203 INIT_LIST_HEAD(&ctx->tctx_list);
1204 INIT_LIST_HEAD(&ctx->submit_state.comp.free_list);
1205 INIT_LIST_HEAD(&ctx->locked_free_list);
1206 INIT_DELAYED_WORK(&ctx->fallback_work, io_fallback_req_func);
1209 kfree(ctx->dummy_ubuf);
1210 kfree(ctx->cancel_hash);
1215 static void io_account_cq_overflow(struct io_ring_ctx *ctx)
1217 struct io_rings *r = ctx->rings;
1219 WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1);
1223 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1225 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1226 struct io_ring_ctx *ctx = req->ctx;
1228 return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail;
1234 #define FFS_ASYNC_READ 0x1UL
1235 #define FFS_ASYNC_WRITE 0x2UL
1237 #define FFS_ISREG 0x4UL
1239 #define FFS_ISREG 0x0UL
1241 #define FFS_MASK ~(FFS_ASYNC_READ|FFS_ASYNC_WRITE|FFS_ISREG)
1243 static inline bool io_req_ffs_set(struct io_kiocb *req)
1245 return IS_ENABLED(CONFIG_64BIT) && (req->flags & REQ_F_FIXED_FILE);
1248 static void io_req_track_inflight(struct io_kiocb *req)
1250 if (!(req->flags & REQ_F_INFLIGHT)) {
1251 req->flags |= REQ_F_INFLIGHT;
1252 atomic_inc(¤t->io_uring->inflight_tracked);
1256 static void io_prep_async_work(struct io_kiocb *req)
1258 const struct io_op_def *def = &io_op_defs[req->opcode];
1259 struct io_ring_ctx *ctx = req->ctx;
1261 if (!(req->flags & REQ_F_CREDS)) {
1262 req->flags |= REQ_F_CREDS;
1263 req->creds = get_current_cred();
1266 req->work.list.next = NULL;
1267 req->work.flags = 0;
1268 if (req->flags & REQ_F_FORCE_ASYNC)
1269 req->work.flags |= IO_WQ_WORK_CONCURRENT;
1271 if (req->flags & REQ_F_ISREG) {
1272 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1273 io_wq_hash_work(&req->work, file_inode(req->file));
1274 } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
1275 if (def->unbound_nonreg_file)
1276 req->work.flags |= IO_WQ_WORK_UNBOUND;
1279 switch (req->opcode) {
1280 case IORING_OP_SPLICE:
1282 if (!S_ISREG(file_inode(req->splice.file_in)->i_mode))
1283 req->work.flags |= IO_WQ_WORK_UNBOUND;
1288 static void io_prep_async_link(struct io_kiocb *req)
1290 struct io_kiocb *cur;
1292 if (req->flags & REQ_F_LINK_TIMEOUT) {
1293 struct io_ring_ctx *ctx = req->ctx;
1295 spin_lock_irq(&ctx->completion_lock);
1296 io_for_each_link(cur, req)
1297 io_prep_async_work(cur);
1298 spin_unlock_irq(&ctx->completion_lock);
1300 io_for_each_link(cur, req)
1301 io_prep_async_work(cur);
1305 static void io_queue_async_work(struct io_kiocb *req)
1307 struct io_ring_ctx *ctx = req->ctx;
1308 struct io_kiocb *link = io_prep_linked_timeout(req);
1309 struct io_uring_task *tctx = req->task->io_uring;
1312 BUG_ON(!tctx->io_wq);
1314 /* init ->work of the whole link before punting */
1315 io_prep_async_link(req);
1318 * Not expected to happen, but if we do have a bug where this _can_
1319 * happen, catch it here and ensure the request is marked as
1320 * canceled. That will make io-wq go through the usual work cancel
1321 * procedure rather than attempt to run this request (or create a new
1324 if (WARN_ON_ONCE(!same_thread_group(req->task, current)))
1325 req->work.flags |= IO_WQ_WORK_CANCEL;
1327 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1328 &req->work, req->flags);
1329 io_wq_enqueue(tctx->io_wq, &req->work);
1331 io_queue_linked_timeout(link);
1334 static void io_kill_timeout(struct io_kiocb *req, int status)
1335 __must_hold(&req->ctx->completion_lock)
1337 struct io_timeout_data *io = req->async_data;
1339 if (hrtimer_try_to_cancel(&io->timer) != -1) {
1340 atomic_set(&req->ctx->cq_timeouts,
1341 atomic_read(&req->ctx->cq_timeouts) + 1);
1342 list_del_init(&req->timeout.list);
1343 io_cqring_fill_event(req->ctx, req->user_data, status, 0);
1344 io_put_req_deferred(req, 1);
1348 static void io_queue_deferred(struct io_ring_ctx *ctx)
1350 while (!list_empty(&ctx->defer_list)) {
1351 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1352 struct io_defer_entry, list);
1354 if (req_need_defer(de->req, de->seq))
1356 list_del_init(&de->list);
1357 io_req_task_queue(de->req);
1362 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1364 u32 seq = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
1366 while (!list_empty(&ctx->timeout_list)) {
1367 u32 events_needed, events_got;
1368 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1369 struct io_kiocb, timeout.list);
1371 if (io_is_timeout_noseq(req))
1375 * Since seq can easily wrap around over time, subtract
1376 * the last seq at which timeouts were flushed before comparing.
1377 * Assuming not more than 2^31-1 events have happened since,
1378 * these subtractions won't have wrapped, so we can check if
1379 * target is in [last_seq, current_seq] by comparing the two.
1381 events_needed = req->timeout.target_seq - ctx->cq_last_tm_flush;
1382 events_got = seq - ctx->cq_last_tm_flush;
1383 if (events_got < events_needed)
1386 list_del_init(&req->timeout.list);
1387 io_kill_timeout(req, 0);
1389 ctx->cq_last_tm_flush = seq;
1392 static void __io_commit_cqring_flush(struct io_ring_ctx *ctx)
1394 if (ctx->off_timeout_used)
1395 io_flush_timeouts(ctx);
1396 if (ctx->drain_active)
1397 io_queue_deferred(ctx);
1400 static inline void io_commit_cqring(struct io_ring_ctx *ctx)
1402 if (unlikely(ctx->off_timeout_used || ctx->drain_active))
1403 __io_commit_cqring_flush(ctx);
1404 /* order cqe stores with ring update */
1405 smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
1408 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1410 struct io_rings *r = ctx->rings;
1412 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == ctx->sq_entries;
1415 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
1417 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
1420 static inline struct io_uring_cqe *io_get_cqe(struct io_ring_ctx *ctx)
1422 struct io_rings *rings = ctx->rings;
1423 unsigned tail, mask = ctx->cq_entries - 1;
1426 * writes to the cq entry need to come after reading head; the
1427 * control dependency is enough as we're using WRITE_ONCE to
1430 if (__io_cqring_events(ctx) == ctx->cq_entries)
1433 tail = ctx->cached_cq_tail++;
1434 return &rings->cqes[tail & mask];
1437 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1439 if (likely(!ctx->cq_ev_fd))
1441 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1443 return !ctx->eventfd_async || io_wq_current_is_worker();
1446 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1449 * wake_up_all() may seem excessive, but io_wake_function() and
1450 * io_should_wake() handle the termination of the loop and only
1451 * wake as many waiters as we need to.
1453 if (wq_has_sleeper(&ctx->cq_wait))
1454 wake_up_all(&ctx->cq_wait);
1455 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1456 wake_up(&ctx->sq_data->wait);
1457 if (io_should_trigger_evfd(ctx))
1458 eventfd_signal(ctx->cq_ev_fd, 1);
1459 if (waitqueue_active(&ctx->poll_wait)) {
1460 wake_up_interruptible(&ctx->poll_wait);
1461 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1465 static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx)
1467 if (ctx->flags & IORING_SETUP_SQPOLL) {
1468 if (wq_has_sleeper(&ctx->cq_wait))
1469 wake_up_all(&ctx->cq_wait);
1471 if (io_should_trigger_evfd(ctx))
1472 eventfd_signal(ctx->cq_ev_fd, 1);
1473 if (waitqueue_active(&ctx->poll_wait)) {
1474 wake_up_interruptible(&ctx->poll_wait);
1475 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1479 /* Returns true if there are no backlogged entries after the flush */
1480 static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1482 unsigned long flags;
1483 bool all_flushed, posted;
1485 if (!force && __io_cqring_events(ctx) == ctx->cq_entries)
1489 spin_lock_irqsave(&ctx->completion_lock, flags);
1490 while (!list_empty(&ctx->cq_overflow_list)) {
1491 struct io_uring_cqe *cqe = io_get_cqe(ctx);
1492 struct io_overflow_cqe *ocqe;
1496 ocqe = list_first_entry(&ctx->cq_overflow_list,
1497 struct io_overflow_cqe, list);
1499 memcpy(cqe, &ocqe->cqe, sizeof(*cqe));
1501 io_account_cq_overflow(ctx);
1504 list_del(&ocqe->list);
1508 all_flushed = list_empty(&ctx->cq_overflow_list);
1510 clear_bit(0, &ctx->check_cq_overflow);
1511 WRITE_ONCE(ctx->rings->sq_flags,
1512 ctx->rings->sq_flags & ~IORING_SQ_CQ_OVERFLOW);
1516 io_commit_cqring(ctx);
1517 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1519 io_cqring_ev_posted(ctx);
1523 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1527 if (test_bit(0, &ctx->check_cq_overflow)) {
1528 /* iopoll syncs against uring_lock, not completion_lock */
1529 if (ctx->flags & IORING_SETUP_IOPOLL)
1530 mutex_lock(&ctx->uring_lock);
1531 ret = __io_cqring_overflow_flush(ctx, force);
1532 if (ctx->flags & IORING_SETUP_IOPOLL)
1533 mutex_unlock(&ctx->uring_lock);
1540 * Shamelessly stolen from the mm implementation of page reference checking,
1541 * see commit f958d7b528b1 for details.
1543 #define req_ref_zero_or_close_to_overflow(req) \
1544 ((unsigned int) atomic_read(&(req->refs)) + 127u <= 127u)
1546 static inline bool req_ref_inc_not_zero(struct io_kiocb *req)
1548 return atomic_inc_not_zero(&req->refs);
1551 static inline bool req_ref_sub_and_test(struct io_kiocb *req, int refs)
1553 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1554 return atomic_sub_and_test(refs, &req->refs);
1557 static inline bool req_ref_put_and_test(struct io_kiocb *req)
1559 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1560 return atomic_dec_and_test(&req->refs);
1563 static inline void req_ref_put(struct io_kiocb *req)
1565 WARN_ON_ONCE(req_ref_put_and_test(req));
1568 static inline void req_ref_get(struct io_kiocb *req)
1570 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1571 atomic_inc(&req->refs);
1574 static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data,
1575 long res, unsigned int cflags)
1577 struct io_overflow_cqe *ocqe;
1579 ocqe = kmalloc(sizeof(*ocqe), GFP_ATOMIC | __GFP_ACCOUNT);
1582 * If we're in ring overflow flush mode, or in task cancel mode,
1583 * or cannot allocate an overflow entry, then we need to drop it
1586 io_account_cq_overflow(ctx);
1589 if (list_empty(&ctx->cq_overflow_list)) {
1590 set_bit(0, &ctx->check_cq_overflow);
1591 WRITE_ONCE(ctx->rings->sq_flags,
1592 ctx->rings->sq_flags | IORING_SQ_CQ_OVERFLOW);
1595 ocqe->cqe.user_data = user_data;
1596 ocqe->cqe.res = res;
1597 ocqe->cqe.flags = cflags;
1598 list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
1602 static inline bool __io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1603 long res, unsigned int cflags)
1605 struct io_uring_cqe *cqe;
1607 trace_io_uring_complete(ctx, user_data, res, cflags);
1610 * If we can't get a cq entry, userspace overflowed the
1611 * submission (by quite a lot). Increment the overflow count in
1614 cqe = io_get_cqe(ctx);
1616 WRITE_ONCE(cqe->user_data, user_data);
1617 WRITE_ONCE(cqe->res, res);
1618 WRITE_ONCE(cqe->flags, cflags);
1621 return io_cqring_event_overflow(ctx, user_data, res, cflags);
1624 /* not as hot to bloat with inlining */
1625 static noinline bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1626 long res, unsigned int cflags)
1628 return __io_cqring_fill_event(ctx, user_data, res, cflags);
1631 static void io_req_complete_post(struct io_kiocb *req, long res,
1632 unsigned int cflags)
1634 struct io_ring_ctx *ctx = req->ctx;
1635 unsigned long flags;
1637 spin_lock_irqsave(&ctx->completion_lock, flags);
1638 __io_cqring_fill_event(ctx, req->user_data, res, cflags);
1640 * If we're the last reference to this request, add to our locked
1643 if (req_ref_put_and_test(req)) {
1644 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
1645 if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL))
1646 io_disarm_next(req);
1648 io_req_task_queue(req->link);
1652 io_dismantle_req(req);
1653 io_put_task(req->task, 1);
1654 list_add(&req->compl.list, &ctx->locked_free_list);
1655 ctx->locked_free_nr++;
1657 if (!percpu_ref_tryget(&ctx->refs))
1660 io_commit_cqring(ctx);
1661 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1664 io_cqring_ev_posted(ctx);
1665 percpu_ref_put(&ctx->refs);
1669 static inline bool io_req_needs_clean(struct io_kiocb *req)
1671 return req->flags & IO_REQ_CLEAN_FLAGS;
1674 static void io_req_complete_state(struct io_kiocb *req, long res,
1675 unsigned int cflags)
1677 if (io_req_needs_clean(req))
1680 req->compl.cflags = cflags;
1681 req->flags |= REQ_F_COMPLETE_INLINE;
1684 static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags,
1685 long res, unsigned cflags)
1687 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1688 io_req_complete_state(req, res, cflags);
1690 io_req_complete_post(req, res, cflags);
1693 static inline void io_req_complete(struct io_kiocb *req, long res)
1695 __io_req_complete(req, 0, res, 0);
1698 static void io_req_complete_failed(struct io_kiocb *req, long res)
1702 io_req_complete_post(req, res, 0);
1706 * Don't initialise the fields below on every allocation, but do that in
1707 * advance and keep them valid across allocations.
1709 static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx)
1713 req->async_data = NULL;
1714 /* not necessary, but safer to zero */
1718 static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
1719 struct io_comp_state *cs)
1721 spin_lock_irq(&ctx->completion_lock);
1722 list_splice_init(&ctx->locked_free_list, &cs->free_list);
1723 ctx->locked_free_nr = 0;
1724 spin_unlock_irq(&ctx->completion_lock);
1727 /* Returns true IFF there are requests in the cache */
1728 static bool io_flush_cached_reqs(struct io_ring_ctx *ctx)
1730 struct io_submit_state *state = &ctx->submit_state;
1731 struct io_comp_state *cs = &state->comp;
1735 * If we have more than a batch's worth of requests in our IRQ side
1736 * locked cache, grab the lock and move them over to our submission
1739 if (READ_ONCE(ctx->locked_free_nr) > IO_COMPL_BATCH)
1740 io_flush_cached_locked_reqs(ctx, cs);
1742 nr = state->free_reqs;
1743 while (!list_empty(&cs->free_list)) {
1744 struct io_kiocb *req = list_first_entry(&cs->free_list,
1745 struct io_kiocb, compl.list);
1747 list_del(&req->compl.list);
1748 state->reqs[nr++] = req;
1749 if (nr == ARRAY_SIZE(state->reqs))
1753 state->free_reqs = nr;
1757 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
1759 struct io_submit_state *state = &ctx->submit_state;
1760 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1763 BUILD_BUG_ON(ARRAY_SIZE(state->reqs) < IO_REQ_ALLOC_BATCH);
1765 if (likely(state->free_reqs || io_flush_cached_reqs(ctx)))
1768 ret = kmem_cache_alloc_bulk(req_cachep, gfp, IO_REQ_ALLOC_BATCH,
1772 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1773 * retry single alloc to be on the safe side.
1775 if (unlikely(ret <= 0)) {
1776 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1777 if (!state->reqs[0])
1782 for (i = 0; i < ret; i++)
1783 io_preinit_req(state->reqs[i], ctx);
1784 state->free_reqs = ret;
1787 return state->reqs[state->free_reqs];
1790 static inline void io_put_file(struct file *file)
1796 static void io_dismantle_req(struct io_kiocb *req)
1798 unsigned int flags = req->flags;
1800 if (io_req_needs_clean(req))
1802 if (!(flags & REQ_F_FIXED_FILE))
1803 io_put_file(req->file);
1804 if (req->fixed_rsrc_refs)
1805 percpu_ref_put(req->fixed_rsrc_refs);
1806 if (req->async_data) {
1807 kfree(req->async_data);
1808 req->async_data = NULL;
1812 /* must to be called somewhat shortly after putting a request */
1813 static inline void io_put_task(struct task_struct *task, int nr)
1815 struct io_uring_task *tctx = task->io_uring;
1817 percpu_counter_sub(&tctx->inflight, nr);
1818 if (unlikely(atomic_read(&tctx->in_idle)))
1819 wake_up(&tctx->wait);
1820 put_task_struct_many(task, nr);
1823 static void __io_free_req(struct io_kiocb *req)
1825 struct io_ring_ctx *ctx = req->ctx;
1827 io_dismantle_req(req);
1828 io_put_task(req->task, 1);
1830 kmem_cache_free(req_cachep, req);
1831 percpu_ref_put(&ctx->refs);
1834 static inline void io_remove_next_linked(struct io_kiocb *req)
1836 struct io_kiocb *nxt = req->link;
1838 req->link = nxt->link;
1842 static bool io_kill_linked_timeout(struct io_kiocb *req)
1843 __must_hold(&req->ctx->completion_lock)
1845 struct io_kiocb *link = req->link;
1848 * Can happen if a linked timeout fired and link had been like
1849 * req -> link t-out -> link t-out [-> ...]
1851 if (link && (link->flags & REQ_F_LTIMEOUT_ACTIVE)) {
1852 struct io_timeout_data *io = link->async_data;
1854 io_remove_next_linked(req);
1855 link->timeout.head = NULL;
1856 if (hrtimer_try_to_cancel(&io->timer) != -1) {
1857 io_cqring_fill_event(link->ctx, link->user_data,
1859 io_put_req_deferred(link, 1);
1866 static void io_fail_links(struct io_kiocb *req)
1867 __must_hold(&req->ctx->completion_lock)
1869 struct io_kiocb *nxt, *link = req->link;
1876 trace_io_uring_fail_link(req, link);
1877 io_cqring_fill_event(link->ctx, link->user_data, -ECANCELED, 0);
1878 io_put_req_deferred(link, 2);
1883 static bool io_disarm_next(struct io_kiocb *req)
1884 __must_hold(&req->ctx->completion_lock)
1886 bool posted = false;
1888 if (likely(req->flags & REQ_F_LINK_TIMEOUT))
1889 posted = io_kill_linked_timeout(req);
1890 if (unlikely((req->flags & REQ_F_FAIL) &&
1891 !(req->flags & REQ_F_HARDLINK))) {
1892 posted |= (req->link != NULL);
1898 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1900 struct io_kiocb *nxt;
1903 * If LINK is set, we have dependent requests in this chain. If we
1904 * didn't fail this request, queue the first one up, moving any other
1905 * dependencies to the next request. In case of failure, fail the rest
1908 if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL)) {
1909 struct io_ring_ctx *ctx = req->ctx;
1910 unsigned long flags;
1913 spin_lock_irqsave(&ctx->completion_lock, flags);
1914 posted = io_disarm_next(req);
1916 io_commit_cqring(req->ctx);
1917 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1919 io_cqring_ev_posted(ctx);
1926 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1928 if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK))))
1930 return __io_req_find_next(req);
1933 static void ctx_flush_and_put(struct io_ring_ctx *ctx)
1937 if (ctx->submit_state.comp.nr) {
1938 mutex_lock(&ctx->uring_lock);
1939 io_submit_flush_completions(ctx);
1940 mutex_unlock(&ctx->uring_lock);
1942 percpu_ref_put(&ctx->refs);
1945 static void tctx_task_work(struct callback_head *cb)
1947 struct io_ring_ctx *ctx = NULL;
1948 struct io_uring_task *tctx = container_of(cb, struct io_uring_task,
1952 struct io_wq_work_node *node;
1954 spin_lock_irq(&tctx->task_lock);
1955 node = tctx->task_list.first;
1956 INIT_WQ_LIST(&tctx->task_list);
1957 spin_unlock_irq(&tctx->task_lock);
1960 struct io_wq_work_node *next = node->next;
1961 struct io_kiocb *req = container_of(node, struct io_kiocb,
1964 if (req->ctx != ctx) {
1965 ctx_flush_and_put(ctx);
1967 percpu_ref_get(&ctx->refs);
1969 req->io_task_work.func(req);
1972 if (wq_list_empty(&tctx->task_list)) {
1973 spin_lock_irq(&tctx->task_lock);
1974 clear_bit(0, &tctx->task_state);
1975 if (wq_list_empty(&tctx->task_list)) {
1976 spin_unlock_irq(&tctx->task_lock);
1979 spin_unlock_irq(&tctx->task_lock);
1980 /* another tctx_task_work() is enqueued, yield */
1981 if (test_and_set_bit(0, &tctx->task_state))
1987 ctx_flush_and_put(ctx);
1990 static void io_req_task_work_add(struct io_kiocb *req)
1992 struct task_struct *tsk = req->task;
1993 struct io_uring_task *tctx = tsk->io_uring;
1994 enum task_work_notify_mode notify;
1995 struct io_wq_work_node *node;
1996 unsigned long flags;
1998 WARN_ON_ONCE(!tctx);
2000 spin_lock_irqsave(&tctx->task_lock, flags);
2001 wq_list_add_tail(&req->io_task_work.node, &tctx->task_list);
2002 spin_unlock_irqrestore(&tctx->task_lock, flags);
2004 /* task_work already pending, we're done */
2005 if (test_bit(0, &tctx->task_state) ||
2006 test_and_set_bit(0, &tctx->task_state))
2010 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
2011 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
2012 * processing task_work. There's no reliable way to tell if TWA_RESUME
2015 notify = (req->ctx->flags & IORING_SETUP_SQPOLL) ? TWA_NONE : TWA_SIGNAL;
2016 if (!task_work_add(tsk, &tctx->task_work, notify)) {
2017 wake_up_process(tsk);
2021 clear_bit(0, &tctx->task_state);
2022 spin_lock_irqsave(&tctx->task_lock, flags);
2023 node = tctx->task_list.first;
2024 INIT_WQ_LIST(&tctx->task_list);
2025 spin_unlock_irqrestore(&tctx->task_lock, flags);
2028 req = container_of(node, struct io_kiocb, io_task_work.node);
2030 if (llist_add(&req->io_task_work.fallback_node,
2031 &req->ctx->fallback_llist))
2032 schedule_delayed_work(&req->ctx->fallback_work, 1);
2036 static void io_req_task_cancel(struct io_kiocb *req)
2038 struct io_ring_ctx *ctx = req->ctx;
2040 /* ctx is guaranteed to stay alive while we hold uring_lock */
2041 mutex_lock(&ctx->uring_lock);
2042 io_req_complete_failed(req, req->result);
2043 mutex_unlock(&ctx->uring_lock);
2046 static void io_req_task_submit(struct io_kiocb *req)
2048 struct io_ring_ctx *ctx = req->ctx;
2050 /* ctx stays valid until unlock, even if we drop all ours ctx->refs */
2051 mutex_lock(&ctx->uring_lock);
2052 if (!(req->task->flags & PF_EXITING) && !req->task->in_execve)
2053 __io_queue_sqe(req);
2055 io_req_complete_failed(req, -EFAULT);
2056 mutex_unlock(&ctx->uring_lock);
2059 static void io_req_task_queue_fail(struct io_kiocb *req, int ret)
2062 req->io_task_work.func = io_req_task_cancel;
2063 io_req_task_work_add(req);
2066 static void io_req_task_queue(struct io_kiocb *req)
2068 req->io_task_work.func = io_req_task_submit;
2069 io_req_task_work_add(req);
2072 static void io_req_task_queue_reissue(struct io_kiocb *req)
2074 req->io_task_work.func = io_queue_async_work;
2075 io_req_task_work_add(req);
2078 static inline void io_queue_next(struct io_kiocb *req)
2080 struct io_kiocb *nxt = io_req_find_next(req);
2083 io_req_task_queue(nxt);
2086 static void io_free_req(struct io_kiocb *req)
2093 struct task_struct *task;
2098 static inline void io_init_req_batch(struct req_batch *rb)
2105 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2106 struct req_batch *rb)
2109 io_put_task(rb->task, rb->task_refs);
2111 percpu_ref_put_many(&ctx->refs, rb->ctx_refs);
2114 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req,
2115 struct io_submit_state *state)
2118 io_dismantle_req(req);
2120 if (req->task != rb->task) {
2122 io_put_task(rb->task, rb->task_refs);
2123 rb->task = req->task;
2129 if (state->free_reqs != ARRAY_SIZE(state->reqs))
2130 state->reqs[state->free_reqs++] = req;
2132 list_add(&req->compl.list, &state->comp.free_list);
2135 static void io_submit_flush_completions(struct io_ring_ctx *ctx)
2136 __must_hold(&req->ctx->uring_lock)
2138 struct io_comp_state *cs = &ctx->submit_state.comp;
2140 struct req_batch rb;
2142 spin_lock_irq(&ctx->completion_lock);
2143 for (i = 0; i < nr; i++) {
2144 struct io_kiocb *req = cs->reqs[i];
2146 __io_cqring_fill_event(ctx, req->user_data, req->result,
2149 io_commit_cqring(ctx);
2150 spin_unlock_irq(&ctx->completion_lock);
2151 io_cqring_ev_posted(ctx);
2153 io_init_req_batch(&rb);
2154 for (i = 0; i < nr; i++) {
2155 struct io_kiocb *req = cs->reqs[i];
2157 /* submission and completion refs */
2158 if (req_ref_sub_and_test(req, 2))
2159 io_req_free_batch(&rb, req, &ctx->submit_state);
2162 io_req_free_batch_finish(ctx, &rb);
2167 * Drop reference to request, return next in chain (if there is one) if this
2168 * was the last reference to this request.
2170 static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2172 struct io_kiocb *nxt = NULL;
2174 if (req_ref_put_and_test(req)) {
2175 nxt = io_req_find_next(req);
2181 static inline void io_put_req(struct io_kiocb *req)
2183 if (req_ref_put_and_test(req))
2187 static void io_free_req_deferred(struct io_kiocb *req)
2189 req->io_task_work.func = io_free_req;
2190 io_req_task_work_add(req);
2193 static inline void io_put_req_deferred(struct io_kiocb *req, int refs)
2195 if (req_ref_sub_and_test(req, refs))
2196 io_free_req_deferred(req);
2199 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2201 /* See comment at the top of this file */
2203 return __io_cqring_events(ctx);
2206 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2208 struct io_rings *rings = ctx->rings;
2210 /* make sure SQ entry isn't read before tail */
2211 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2214 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2216 unsigned int cflags;
2218 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2219 cflags |= IORING_CQE_F_BUFFER;
2220 req->flags &= ~REQ_F_BUFFER_SELECTED;
2225 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2227 struct io_buffer *kbuf;
2229 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2230 return io_put_kbuf(req, kbuf);
2233 static inline bool io_run_task_work(void)
2235 if (test_thread_flag(TIF_NOTIFY_SIGNAL) || current->task_works) {
2236 __set_current_state(TASK_RUNNING);
2237 tracehook_notify_signal();
2245 * Find and free completed poll iocbs
2247 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2248 struct list_head *done, bool resubmit)
2250 struct req_batch rb;
2251 struct io_kiocb *req;
2253 /* order with ->result store in io_complete_rw_iopoll() */
2256 io_init_req_batch(&rb);
2257 while (!list_empty(done)) {
2260 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2261 list_del(&req->inflight_entry);
2263 if (READ_ONCE(req->result) == -EAGAIN && resubmit &&
2264 !(req->flags & REQ_F_DONT_REISSUE)) {
2265 req->iopoll_completed = 0;
2267 io_req_task_queue_reissue(req);
2271 if (req->flags & REQ_F_BUFFER_SELECTED)
2272 cflags = io_put_rw_kbuf(req);
2274 __io_cqring_fill_event(ctx, req->user_data, req->result, cflags);
2277 if (req_ref_put_and_test(req))
2278 io_req_free_batch(&rb, req, &ctx->submit_state);
2281 io_commit_cqring(ctx);
2282 io_cqring_ev_posted_iopoll(ctx);
2283 io_req_free_batch_finish(ctx, &rb);
2286 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2287 long min, bool resubmit)
2289 struct io_kiocb *req, *tmp;
2294 * Only spin for completions if we don't have multiple devices hanging
2295 * off our complete list, and we're under the requested amount.
2297 spin = !ctx->poll_multi_queue && *nr_events < min;
2299 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2300 struct kiocb *kiocb = &req->rw.kiocb;
2304 * Move completed and retryable entries to our local lists.
2305 * If we find a request that requires polling, break out
2306 * and complete those lists first, if we have entries there.
2308 if (READ_ONCE(req->iopoll_completed)) {
2309 list_move_tail(&req->inflight_entry, &done);
2312 if (!list_empty(&done))
2315 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2316 if (unlikely(ret < 0))
2321 /* iopoll may have completed current req */
2322 if (READ_ONCE(req->iopoll_completed))
2323 list_move_tail(&req->inflight_entry, &done);
2326 if (!list_empty(&done))
2327 io_iopoll_complete(ctx, nr_events, &done, resubmit);
2333 * We can't just wait for polled events to come to us, we have to actively
2334 * find and complete them.
2336 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2338 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2341 mutex_lock(&ctx->uring_lock);
2342 while (!list_empty(&ctx->iopoll_list)) {
2343 unsigned int nr_events = 0;
2345 io_do_iopoll(ctx, &nr_events, 0, false);
2347 /* let it sleep and repeat later if can't complete a request */
2351 * Ensure we allow local-to-the-cpu processing to take place,
2352 * in this case we need to ensure that we reap all events.
2353 * Also let task_work, etc. to progress by releasing the mutex
2355 if (need_resched()) {
2356 mutex_unlock(&ctx->uring_lock);
2358 mutex_lock(&ctx->uring_lock);
2361 mutex_unlock(&ctx->uring_lock);
2364 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2366 unsigned int nr_events = 0;
2370 * We disallow the app entering submit/complete with polling, but we
2371 * still need to lock the ring to prevent racing with polled issue
2372 * that got punted to a workqueue.
2374 mutex_lock(&ctx->uring_lock);
2376 * Don't enter poll loop if we already have events pending.
2377 * If we do, we can potentially be spinning for commands that
2378 * already triggered a CQE (eg in error).
2380 if (test_bit(0, &ctx->check_cq_overflow))
2381 __io_cqring_overflow_flush(ctx, false);
2382 if (io_cqring_events(ctx))
2386 * If a submit got punted to a workqueue, we can have the
2387 * application entering polling for a command before it gets
2388 * issued. That app will hold the uring_lock for the duration
2389 * of the poll right here, so we need to take a breather every
2390 * now and then to ensure that the issue has a chance to add
2391 * the poll to the issued list. Otherwise we can spin here
2392 * forever, while the workqueue is stuck trying to acquire the
2395 if (list_empty(&ctx->iopoll_list)) {
2396 u32 tail = ctx->cached_cq_tail;
2398 mutex_unlock(&ctx->uring_lock);
2400 mutex_lock(&ctx->uring_lock);
2402 /* some requests don't go through iopoll_list */
2403 if (tail != ctx->cached_cq_tail ||
2404 list_empty(&ctx->iopoll_list))
2407 ret = io_do_iopoll(ctx, &nr_events, min, true);
2408 } while (!ret && nr_events < min && !need_resched());
2410 mutex_unlock(&ctx->uring_lock);
2414 static void kiocb_end_write(struct io_kiocb *req)
2417 * Tell lockdep we inherited freeze protection from submission
2420 if (req->flags & REQ_F_ISREG) {
2421 struct super_block *sb = file_inode(req->file)->i_sb;
2423 __sb_writers_acquired(sb, SB_FREEZE_WRITE);
2429 static bool io_resubmit_prep(struct io_kiocb *req)
2431 struct io_async_rw *rw = req->async_data;
2434 return !io_req_prep_async(req);
2435 /* may have left rw->iter inconsistent on -EIOCBQUEUED */
2436 iov_iter_revert(&rw->iter, req->result - iov_iter_count(&rw->iter));
2440 static bool io_rw_should_reissue(struct io_kiocb *req)
2442 umode_t mode = file_inode(req->file)->i_mode;
2443 struct io_ring_ctx *ctx = req->ctx;
2445 if (!S_ISBLK(mode) && !S_ISREG(mode))
2447 if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
2448 !(ctx->flags & IORING_SETUP_IOPOLL)))
2451 * If ref is dying, we might be running poll reap from the exit work.
2452 * Don't attempt to reissue from that path, just let it fail with
2455 if (percpu_ref_is_dying(&ctx->refs))
2458 * Play it safe and assume not safe to re-import and reissue if we're
2459 * not in the original thread group (or in task context).
2461 if (!same_thread_group(req->task, current) || !in_task())
2466 static bool io_resubmit_prep(struct io_kiocb *req)
2470 static bool io_rw_should_reissue(struct io_kiocb *req)
2476 static void io_fallback_req_func(struct work_struct *work)
2478 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
2479 fallback_work.work);
2480 struct llist_node *node = llist_del_all(&ctx->fallback_llist);
2481 struct io_kiocb *req, *tmp;
2483 percpu_ref_get(&ctx->refs);
2484 llist_for_each_entry_safe(req, tmp, node, io_task_work.fallback_node)
2485 req->io_task_work.func(req);
2486 percpu_ref_put(&ctx->refs);
2489 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2490 unsigned int issue_flags)
2494 if (req->rw.kiocb.ki_flags & IOCB_WRITE)
2495 kiocb_end_write(req);
2496 if (res != req->result) {
2497 if ((res == -EAGAIN || res == -EOPNOTSUPP) &&
2498 io_rw_should_reissue(req)) {
2499 req->flags |= REQ_F_REISSUE;
2504 if (req->flags & REQ_F_BUFFER_SELECTED)
2505 cflags = io_put_rw_kbuf(req);
2506 __io_req_complete(req, issue_flags, res, cflags);
2509 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2511 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2513 __io_complete_rw(req, res, res2, 0);
2516 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2518 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2520 if (kiocb->ki_flags & IOCB_WRITE)
2521 kiocb_end_write(req);
2522 if (unlikely(res != req->result)) {
2523 if (!(res == -EAGAIN && io_rw_should_reissue(req) &&
2524 io_resubmit_prep(req))) {
2526 req->flags |= REQ_F_DONT_REISSUE;
2530 WRITE_ONCE(req->result, res);
2531 /* order with io_iopoll_complete() checking ->result */
2533 WRITE_ONCE(req->iopoll_completed, 1);
2537 * After the iocb has been issued, it's safe to be found on the poll list.
2538 * Adding the kiocb to the list AFTER submission ensures that we don't
2539 * find it from a io_do_iopoll() thread before the issuer is done
2540 * accessing the kiocb cookie.
2542 static void io_iopoll_req_issued(struct io_kiocb *req)
2544 struct io_ring_ctx *ctx = req->ctx;
2545 const bool in_async = io_wq_current_is_worker();
2547 /* workqueue context doesn't hold uring_lock, grab it now */
2548 if (unlikely(in_async))
2549 mutex_lock(&ctx->uring_lock);
2552 * Track whether we have multiple files in our lists. This will impact
2553 * how we do polling eventually, not spinning if we're on potentially
2554 * different devices.
2556 if (list_empty(&ctx->iopoll_list)) {
2557 ctx->poll_multi_queue = false;
2558 } else if (!ctx->poll_multi_queue) {
2559 struct io_kiocb *list_req;
2560 unsigned int queue_num0, queue_num1;
2562 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2565 if (list_req->file != req->file) {
2566 ctx->poll_multi_queue = true;
2568 queue_num0 = blk_qc_t_to_queue_num(list_req->rw.kiocb.ki_cookie);
2569 queue_num1 = blk_qc_t_to_queue_num(req->rw.kiocb.ki_cookie);
2570 if (queue_num0 != queue_num1)
2571 ctx->poll_multi_queue = true;
2576 * For fast devices, IO may have already completed. If it has, add
2577 * it to the front so we find it first.
2579 if (READ_ONCE(req->iopoll_completed))
2580 list_add(&req->inflight_entry, &ctx->iopoll_list);
2582 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2584 if (unlikely(in_async)) {
2586 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
2587 * in sq thread task context or in io worker task context. If
2588 * current task context is sq thread, we don't need to check
2589 * whether should wake up sq thread.
2591 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2592 wq_has_sleeper(&ctx->sq_data->wait))
2593 wake_up(&ctx->sq_data->wait);
2595 mutex_unlock(&ctx->uring_lock);
2599 static inline void io_state_file_put(struct io_submit_state *state)
2601 if (state->file_refs) {
2602 fput_many(state->file, state->file_refs);
2603 state->file_refs = 0;
2608 * Get as many references to a file as we have IOs left in this submission,
2609 * assuming most submissions are for one file, or at least that each file
2610 * has more than one submission.
2612 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2617 if (state->file_refs) {
2618 if (state->fd == fd) {
2622 io_state_file_put(state);
2624 state->file = fget_many(fd, state->ios_left);
2625 if (unlikely(!state->file))
2629 state->file_refs = state->ios_left - 1;
2633 static bool io_bdev_nowait(struct block_device *bdev)
2635 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2639 * If we tracked the file through the SCM inflight mechanism, we could support
2640 * any file. For now, just ensure that anything potentially problematic is done
2643 static bool __io_file_supports_nowait(struct file *file, int rw)
2645 umode_t mode = file_inode(file)->i_mode;
2647 if (S_ISBLK(mode)) {
2648 if (IS_ENABLED(CONFIG_BLOCK) &&
2649 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2655 if (S_ISREG(mode)) {
2656 if (IS_ENABLED(CONFIG_BLOCK) &&
2657 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2658 file->f_op != &io_uring_fops)
2663 /* any ->read/write should understand O_NONBLOCK */
2664 if (file->f_flags & O_NONBLOCK)
2667 if (!(file->f_mode & FMODE_NOWAIT))
2671 return file->f_op->read_iter != NULL;
2673 return file->f_op->write_iter != NULL;
2676 static bool io_file_supports_nowait(struct io_kiocb *req, int rw)
2678 if (rw == READ && (req->flags & REQ_F_NOWAIT_READ))
2680 else if (rw == WRITE && (req->flags & REQ_F_NOWAIT_WRITE))
2683 return __io_file_supports_nowait(req->file, rw);
2686 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2688 struct io_ring_ctx *ctx = req->ctx;
2689 struct kiocb *kiocb = &req->rw.kiocb;
2690 struct file *file = req->file;
2694 if (!io_req_ffs_set(req) && S_ISREG(file_inode(file)->i_mode))
2695 req->flags |= REQ_F_ISREG;
2697 kiocb->ki_pos = READ_ONCE(sqe->off);
2698 if (kiocb->ki_pos == -1 && !(file->f_mode & FMODE_STREAM)) {
2699 req->flags |= REQ_F_CUR_POS;
2700 kiocb->ki_pos = file->f_pos;
2702 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2703 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2704 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2708 /* don't allow async punt for O_NONBLOCK or RWF_NOWAIT */
2709 if ((kiocb->ki_flags & IOCB_NOWAIT) || (file->f_flags & O_NONBLOCK))
2710 req->flags |= REQ_F_NOWAIT;
2712 ioprio = READ_ONCE(sqe->ioprio);
2714 ret = ioprio_check_cap(ioprio);
2718 kiocb->ki_ioprio = ioprio;
2720 kiocb->ki_ioprio = get_current_ioprio();
2722 if (ctx->flags & IORING_SETUP_IOPOLL) {
2723 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2724 !kiocb->ki_filp->f_op->iopoll)
2727 kiocb->ki_flags |= IOCB_HIPRI;
2728 kiocb->ki_complete = io_complete_rw_iopoll;
2729 req->iopoll_completed = 0;
2731 if (kiocb->ki_flags & IOCB_HIPRI)
2733 kiocb->ki_complete = io_complete_rw;
2736 if (req->opcode == IORING_OP_READ_FIXED ||
2737 req->opcode == IORING_OP_WRITE_FIXED) {
2739 io_req_set_rsrc_node(req);
2742 req->rw.addr = READ_ONCE(sqe->addr);
2743 req->rw.len = READ_ONCE(sqe->len);
2744 req->buf_index = READ_ONCE(sqe->buf_index);
2748 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2754 case -ERESTARTNOINTR:
2755 case -ERESTARTNOHAND:
2756 case -ERESTART_RESTARTBLOCK:
2758 * We can't just restart the syscall, since previously
2759 * submitted sqes may already be in progress. Just fail this
2765 kiocb->ki_complete(kiocb, ret, 0);
2769 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2770 unsigned int issue_flags)
2772 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2773 struct io_async_rw *io = req->async_data;
2774 bool check_reissue = kiocb->ki_complete == io_complete_rw;
2776 /* add previously done IO, if any */
2777 if (io && io->bytes_done > 0) {
2779 ret = io->bytes_done;
2781 ret += io->bytes_done;
2784 if (req->flags & REQ_F_CUR_POS)
2785 req->file->f_pos = kiocb->ki_pos;
2786 if (ret >= 0 && check_reissue)
2787 __io_complete_rw(req, ret, 0, issue_flags);
2789 io_rw_done(kiocb, ret);
2791 if (check_reissue && (req->flags & REQ_F_REISSUE)) {
2792 req->flags &= ~REQ_F_REISSUE;
2793 if (io_resubmit_prep(req)) {
2795 io_req_task_queue_reissue(req);
2800 if (req->flags & REQ_F_BUFFER_SELECTED)
2801 cflags = io_put_rw_kbuf(req);
2802 __io_req_complete(req, issue_flags, ret, cflags);
2807 static int __io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter,
2808 struct io_mapped_ubuf *imu)
2810 size_t len = req->rw.len;
2811 u64 buf_end, buf_addr = req->rw.addr;
2814 if (unlikely(check_add_overflow(buf_addr, (u64)len, &buf_end)))
2816 /* not inside the mapped region */
2817 if (unlikely(buf_addr < imu->ubuf || buf_end > imu->ubuf_end))
2821 * May not be a start of buffer, set size appropriately
2822 * and advance us to the beginning.
2824 offset = buf_addr - imu->ubuf;
2825 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2829 * Don't use iov_iter_advance() here, as it's really slow for
2830 * using the latter parts of a big fixed buffer - it iterates
2831 * over each segment manually. We can cheat a bit here, because
2834 * 1) it's a BVEC iter, we set it up
2835 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2836 * first and last bvec
2838 * So just find our index, and adjust the iterator afterwards.
2839 * If the offset is within the first bvec (or the whole first
2840 * bvec, just use iov_iter_advance(). This makes it easier
2841 * since we can just skip the first segment, which may not
2842 * be PAGE_SIZE aligned.
2844 const struct bio_vec *bvec = imu->bvec;
2846 if (offset <= bvec->bv_len) {
2847 iov_iter_advance(iter, offset);
2849 unsigned long seg_skip;
2851 /* skip first vec */
2852 offset -= bvec->bv_len;
2853 seg_skip = 1 + (offset >> PAGE_SHIFT);
2855 iter->bvec = bvec + seg_skip;
2856 iter->nr_segs -= seg_skip;
2857 iter->count -= bvec->bv_len + offset;
2858 iter->iov_offset = offset & ~PAGE_MASK;
2865 static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
2867 struct io_ring_ctx *ctx = req->ctx;
2868 struct io_mapped_ubuf *imu = req->imu;
2869 u16 index, buf_index = req->buf_index;
2872 if (unlikely(buf_index >= ctx->nr_user_bufs))
2874 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2875 imu = READ_ONCE(ctx->user_bufs[index]);
2878 return __io_import_fixed(req, rw, iter, imu);
2881 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2884 mutex_unlock(&ctx->uring_lock);
2887 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2890 * "Normal" inline submissions always hold the uring_lock, since we
2891 * grab it from the system call. Same is true for the SQPOLL offload.
2892 * The only exception is when we've detached the request and issue it
2893 * from an async worker thread, grab the lock for that case.
2896 mutex_lock(&ctx->uring_lock);
2899 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2900 int bgid, struct io_buffer *kbuf,
2903 struct io_buffer *head;
2905 if (req->flags & REQ_F_BUFFER_SELECTED)
2908 io_ring_submit_lock(req->ctx, needs_lock);
2910 lockdep_assert_held(&req->ctx->uring_lock);
2912 head = xa_load(&req->ctx->io_buffers, bgid);
2914 if (!list_empty(&head->list)) {
2915 kbuf = list_last_entry(&head->list, struct io_buffer,
2917 list_del(&kbuf->list);
2920 xa_erase(&req->ctx->io_buffers, bgid);
2922 if (*len > kbuf->len)
2925 kbuf = ERR_PTR(-ENOBUFS);
2928 io_ring_submit_unlock(req->ctx, needs_lock);
2933 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2936 struct io_buffer *kbuf;
2939 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2940 bgid = req->buf_index;
2941 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2944 req->rw.addr = (u64) (unsigned long) kbuf;
2945 req->flags |= REQ_F_BUFFER_SELECTED;
2946 return u64_to_user_ptr(kbuf->addr);
2949 #ifdef CONFIG_COMPAT
2950 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2953 struct compat_iovec __user *uiov;
2954 compat_ssize_t clen;
2958 uiov = u64_to_user_ptr(req->rw.addr);
2959 if (!access_ok(uiov, sizeof(*uiov)))
2961 if (__get_user(clen, &uiov->iov_len))
2967 buf = io_rw_buffer_select(req, &len, needs_lock);
2969 return PTR_ERR(buf);
2970 iov[0].iov_base = buf;
2971 iov[0].iov_len = (compat_size_t) len;
2976 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2979 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2983 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2986 len = iov[0].iov_len;
2989 buf = io_rw_buffer_select(req, &len, needs_lock);
2991 return PTR_ERR(buf);
2992 iov[0].iov_base = buf;
2993 iov[0].iov_len = len;
2997 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3000 if (req->flags & REQ_F_BUFFER_SELECTED) {
3001 struct io_buffer *kbuf;
3003 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3004 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
3005 iov[0].iov_len = kbuf->len;
3008 if (req->rw.len != 1)
3011 #ifdef CONFIG_COMPAT
3012 if (req->ctx->compat)
3013 return io_compat_import(req, iov, needs_lock);
3016 return __io_iov_buffer_select(req, iov, needs_lock);
3019 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
3020 struct iov_iter *iter, bool needs_lock)
3022 void __user *buf = u64_to_user_ptr(req->rw.addr);
3023 size_t sqe_len = req->rw.len;
3024 u8 opcode = req->opcode;
3027 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3029 return io_import_fixed(req, rw, iter);
3032 /* buffer index only valid with fixed read/write, or buffer select */
3033 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3036 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3037 if (req->flags & REQ_F_BUFFER_SELECT) {
3038 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3040 return PTR_ERR(buf);
3041 req->rw.len = sqe_len;
3044 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3049 if (req->flags & REQ_F_BUFFER_SELECT) {
3050 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3052 iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
3057 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3061 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3063 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3067 * For files that don't have ->read_iter() and ->write_iter(), handle them
3068 * by looping over ->read() or ->write() manually.
3070 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3072 struct kiocb *kiocb = &req->rw.kiocb;
3073 struct file *file = req->file;
3077 * Don't support polled IO through this interface, and we can't
3078 * support non-blocking either. For the latter, this just causes
3079 * the kiocb to be handled from an async context.
3081 if (kiocb->ki_flags & IOCB_HIPRI)
3083 if (kiocb->ki_flags & IOCB_NOWAIT)
3086 while (iov_iter_count(iter)) {
3090 if (!iov_iter_is_bvec(iter)) {
3091 iovec = iov_iter_iovec(iter);
3093 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3094 iovec.iov_len = req->rw.len;
3098 nr = file->f_op->read(file, iovec.iov_base,
3099 iovec.iov_len, io_kiocb_ppos(kiocb));
3101 nr = file->f_op->write(file, iovec.iov_base,
3102 iovec.iov_len, io_kiocb_ppos(kiocb));
3111 if (nr != iovec.iov_len)
3115 iov_iter_advance(iter, nr);
3121 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3122 const struct iovec *fast_iov, struct iov_iter *iter)
3124 struct io_async_rw *rw = req->async_data;
3126 memcpy(&rw->iter, iter, sizeof(*iter));
3127 rw->free_iovec = iovec;
3129 /* can only be fixed buffers, no need to do anything */
3130 if (iov_iter_is_bvec(iter))
3133 unsigned iov_off = 0;
3135 rw->iter.iov = rw->fast_iov;
3136 if (iter->iov != fast_iov) {
3137 iov_off = iter->iov - fast_iov;
3138 rw->iter.iov += iov_off;
3140 if (rw->fast_iov != fast_iov)
3141 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3142 sizeof(struct iovec) * iter->nr_segs);
3144 req->flags |= REQ_F_NEED_CLEANUP;
3148 static inline int io_alloc_async_data(struct io_kiocb *req)
3150 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3151 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3152 return req->async_data == NULL;
3155 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3156 const struct iovec *fast_iov,
3157 struct iov_iter *iter, bool force)
3159 if (!force && !io_op_defs[req->opcode].needs_async_setup)
3161 if (!req->async_data) {
3162 if (io_alloc_async_data(req)) {
3167 io_req_map_rw(req, iovec, fast_iov, iter);
3172 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3174 struct io_async_rw *iorw = req->async_data;
3175 struct iovec *iov = iorw->fast_iov;
3178 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3179 if (unlikely(ret < 0))
3182 iorw->bytes_done = 0;
3183 iorw->free_iovec = iov;
3185 req->flags |= REQ_F_NEED_CLEANUP;
3189 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3191 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3193 return io_prep_rw(req, sqe);
3197 * This is our waitqueue callback handler, registered through lock_page_async()
3198 * when we initially tried to do the IO with the iocb armed our waitqueue.
3199 * This gets called when the page is unlocked, and we generally expect that to
3200 * happen when the page IO is completed and the page is now uptodate. This will
3201 * queue a task_work based retry of the operation, attempting to copy the data
3202 * again. If the latter fails because the page was NOT uptodate, then we will
3203 * do a thread based blocking retry of the operation. That's the unexpected
3206 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3207 int sync, void *arg)
3209 struct wait_page_queue *wpq;
3210 struct io_kiocb *req = wait->private;
3211 struct wait_page_key *key = arg;
3213 wpq = container_of(wait, struct wait_page_queue, wait);
3215 if (!wake_page_match(wpq, key))
3218 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3219 list_del_init(&wait->entry);
3221 /* submit ref gets dropped, acquire a new one */
3223 io_req_task_queue(req);
3228 * This controls whether a given IO request should be armed for async page
3229 * based retry. If we return false here, the request is handed to the async
3230 * worker threads for retry. If we're doing buffered reads on a regular file,
3231 * we prepare a private wait_page_queue entry and retry the operation. This
3232 * will either succeed because the page is now uptodate and unlocked, or it
3233 * will register a callback when the page is unlocked at IO completion. Through
3234 * that callback, io_uring uses task_work to setup a retry of the operation.
3235 * That retry will attempt the buffered read again. The retry will generally
3236 * succeed, or in rare cases where it fails, we then fall back to using the
3237 * async worker threads for a blocking retry.
3239 static bool io_rw_should_retry(struct io_kiocb *req)
3241 struct io_async_rw *rw = req->async_data;
3242 struct wait_page_queue *wait = &rw->wpq;
3243 struct kiocb *kiocb = &req->rw.kiocb;
3245 /* never retry for NOWAIT, we just complete with -EAGAIN */
3246 if (req->flags & REQ_F_NOWAIT)
3249 /* Only for buffered IO */
3250 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3254 * just use poll if we can, and don't attempt if the fs doesn't
3255 * support callback based unlocks
3257 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3260 wait->wait.func = io_async_buf_func;
3261 wait->wait.private = req;
3262 wait->wait.flags = 0;
3263 INIT_LIST_HEAD(&wait->wait.entry);
3264 kiocb->ki_flags |= IOCB_WAITQ;
3265 kiocb->ki_flags &= ~IOCB_NOWAIT;
3266 kiocb->ki_waitq = wait;
3270 static inline int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3272 if (req->file->f_op->read_iter)
3273 return call_read_iter(req->file, &req->rw.kiocb, iter);
3274 else if (req->file->f_op->read)
3275 return loop_rw_iter(READ, req, iter);
3280 static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3282 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3283 struct kiocb *kiocb = &req->rw.kiocb;
3284 struct iov_iter __iter, *iter = &__iter;
3285 struct io_async_rw *rw = req->async_data;
3286 ssize_t io_size, ret, ret2;
3287 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3293 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3297 io_size = iov_iter_count(iter);
3298 req->result = io_size;
3300 /* Ensure we clear previously set non-block flag */
3301 if (!force_nonblock)
3302 kiocb->ki_flags &= ~IOCB_NOWAIT;
3304 kiocb->ki_flags |= IOCB_NOWAIT;
3306 /* If the file doesn't support async, just async punt */
3307 if (force_nonblock && !io_file_supports_nowait(req, READ)) {
3308 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3309 return ret ?: -EAGAIN;
3312 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size);
3313 if (unlikely(ret)) {
3318 ret = io_iter_do_read(req, iter);
3320 if (ret == -EAGAIN || (req->flags & REQ_F_REISSUE)) {
3321 req->flags &= ~REQ_F_REISSUE;
3322 /* IOPOLL retry should happen for io-wq threads */
3323 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3325 /* no retry on NONBLOCK nor RWF_NOWAIT */
3326 if (req->flags & REQ_F_NOWAIT)
3328 /* some cases will consume bytes even on error returns */
3329 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3331 } else if (ret == -EIOCBQUEUED) {
3333 } else if (ret <= 0 || ret == io_size || !force_nonblock ||
3334 (req->flags & REQ_F_NOWAIT) || !(req->flags & REQ_F_ISREG)) {
3335 /* read all, failed, already did sync or don't want to retry */
3339 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3344 rw = req->async_data;
3345 /* now use our persistent iterator, if we aren't already */
3350 rw->bytes_done += ret;
3351 /* if we can retry, do so with the callbacks armed */
3352 if (!io_rw_should_retry(req)) {
3353 kiocb->ki_flags &= ~IOCB_WAITQ;
3358 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3359 * we get -EIOCBQUEUED, then we'll get a notification when the
3360 * desired page gets unlocked. We can also get a partial read
3361 * here, and if we do, then just retry at the new offset.
3363 ret = io_iter_do_read(req, iter);
3364 if (ret == -EIOCBQUEUED)
3366 /* we got some bytes, but not all. retry. */
3367 kiocb->ki_flags &= ~IOCB_WAITQ;
3368 } while (ret > 0 && ret < io_size);
3370 kiocb_done(kiocb, ret, issue_flags);
3372 /* it's faster to check here then delegate to kfree */
3378 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3380 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3382 return io_prep_rw(req, sqe);
3385 static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3387 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3388 struct kiocb *kiocb = &req->rw.kiocb;
3389 struct iov_iter __iter, *iter = &__iter;
3390 struct io_async_rw *rw = req->async_data;
3391 ssize_t ret, ret2, io_size;
3392 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3398 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3402 io_size = iov_iter_count(iter);
3403 req->result = io_size;
3405 /* Ensure we clear previously set non-block flag */
3406 if (!force_nonblock)
3407 kiocb->ki_flags &= ~IOCB_NOWAIT;
3409 kiocb->ki_flags |= IOCB_NOWAIT;
3411 /* If the file doesn't support async, just async punt */
3412 if (force_nonblock && !io_file_supports_nowait(req, WRITE))
3415 /* file path doesn't support NOWAIT for non-direct_IO */
3416 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3417 (req->flags & REQ_F_ISREG))
3420 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size);
3425 * Open-code file_start_write here to grab freeze protection,
3426 * which will be released by another thread in
3427 * io_complete_rw(). Fool lockdep by telling it the lock got
3428 * released so that it doesn't complain about the held lock when
3429 * we return to userspace.
3431 if (req->flags & REQ_F_ISREG) {
3432 sb_start_write(file_inode(req->file)->i_sb);
3433 __sb_writers_release(file_inode(req->file)->i_sb,
3436 kiocb->ki_flags |= IOCB_WRITE;
3438 if (req->file->f_op->write_iter)
3439 ret2 = call_write_iter(req->file, kiocb, iter);
3440 else if (req->file->f_op->write)
3441 ret2 = loop_rw_iter(WRITE, req, iter);
3445 if (req->flags & REQ_F_REISSUE) {
3446 req->flags &= ~REQ_F_REISSUE;
3451 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3452 * retry them without IOCB_NOWAIT.
3454 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3456 /* no retry on NONBLOCK nor RWF_NOWAIT */
3457 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3459 if (!force_nonblock || ret2 != -EAGAIN) {
3460 /* IOPOLL retry should happen for io-wq threads */
3461 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3464 kiocb_done(kiocb, ret2, issue_flags);
3467 /* some cases will consume bytes even on error returns */
3468 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3469 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3470 return ret ?: -EAGAIN;
3473 /* it's reportedly faster than delegating the null check to kfree() */
3479 static int io_renameat_prep(struct io_kiocb *req,
3480 const struct io_uring_sqe *sqe)
3482 struct io_rename *ren = &req->rename;
3483 const char __user *oldf, *newf;
3485 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3487 if (sqe->ioprio || sqe->buf_index)
3489 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3492 ren->old_dfd = READ_ONCE(sqe->fd);
3493 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3494 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3495 ren->new_dfd = READ_ONCE(sqe->len);
3496 ren->flags = READ_ONCE(sqe->rename_flags);
3498 ren->oldpath = getname(oldf);
3499 if (IS_ERR(ren->oldpath))
3500 return PTR_ERR(ren->oldpath);
3502 ren->newpath = getname(newf);
3503 if (IS_ERR(ren->newpath)) {
3504 putname(ren->oldpath);
3505 return PTR_ERR(ren->newpath);
3508 req->flags |= REQ_F_NEED_CLEANUP;
3512 static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3514 struct io_rename *ren = &req->rename;
3517 if (issue_flags & IO_URING_F_NONBLOCK)
3520 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3521 ren->newpath, ren->flags);
3523 req->flags &= ~REQ_F_NEED_CLEANUP;
3526 io_req_complete(req, ret);
3530 static int io_unlinkat_prep(struct io_kiocb *req,
3531 const struct io_uring_sqe *sqe)
3533 struct io_unlink *un = &req->unlink;
3534 const char __user *fname;
3536 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3538 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
3540 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3543 un->dfd = READ_ONCE(sqe->fd);
3545 un->flags = READ_ONCE(sqe->unlink_flags);
3546 if (un->flags & ~AT_REMOVEDIR)
3549 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3550 un->filename = getname(fname);
3551 if (IS_ERR(un->filename))
3552 return PTR_ERR(un->filename);
3554 req->flags |= REQ_F_NEED_CLEANUP;
3558 static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3560 struct io_unlink *un = &req->unlink;
3563 if (issue_flags & IO_URING_F_NONBLOCK)
3566 if (un->flags & AT_REMOVEDIR)
3567 ret = do_rmdir(un->dfd, un->filename);
3569 ret = do_unlinkat(un->dfd, un->filename);
3571 req->flags &= ~REQ_F_NEED_CLEANUP;
3574 io_req_complete(req, ret);
3578 static int io_shutdown_prep(struct io_kiocb *req,
3579 const struct io_uring_sqe *sqe)
3581 #if defined(CONFIG_NET)
3582 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3584 if (sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3588 req->shutdown.how = READ_ONCE(sqe->len);
3595 static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3597 #if defined(CONFIG_NET)
3598 struct socket *sock;
3601 if (issue_flags & IO_URING_F_NONBLOCK)
3604 sock = sock_from_file(req->file);
3605 if (unlikely(!sock))
3608 ret = __sys_shutdown_sock(sock, req->shutdown.how);
3611 io_req_complete(req, ret);
3618 static int __io_splice_prep(struct io_kiocb *req,
3619 const struct io_uring_sqe *sqe)
3621 struct io_splice *sp = &req->splice;
3622 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3624 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3628 sp->len = READ_ONCE(sqe->len);
3629 sp->flags = READ_ONCE(sqe->splice_flags);
3631 if (unlikely(sp->flags & ~valid_flags))
3634 sp->file_in = io_file_get(req->ctx, NULL, req,
3635 READ_ONCE(sqe->splice_fd_in),
3636 (sp->flags & SPLICE_F_FD_IN_FIXED));
3639 req->flags |= REQ_F_NEED_CLEANUP;
3643 static int io_tee_prep(struct io_kiocb *req,
3644 const struct io_uring_sqe *sqe)
3646 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3648 return __io_splice_prep(req, sqe);
3651 static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
3653 struct io_splice *sp = &req->splice;
3654 struct file *in = sp->file_in;
3655 struct file *out = sp->file_out;
3656 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3659 if (issue_flags & IO_URING_F_NONBLOCK)
3662 ret = do_tee(in, out, sp->len, flags);
3664 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3666 req->flags &= ~REQ_F_NEED_CLEANUP;
3670 io_req_complete(req, ret);
3674 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3676 struct io_splice *sp = &req->splice;
3678 sp->off_in = READ_ONCE(sqe->splice_off_in);
3679 sp->off_out = READ_ONCE(sqe->off);
3680 return __io_splice_prep(req, sqe);
3683 static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
3685 struct io_splice *sp = &req->splice;
3686 struct file *in = sp->file_in;
3687 struct file *out = sp->file_out;
3688 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3689 loff_t *poff_in, *poff_out;
3692 if (issue_flags & IO_URING_F_NONBLOCK)
3695 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3696 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3699 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3701 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3703 req->flags &= ~REQ_F_NEED_CLEANUP;
3707 io_req_complete(req, ret);
3712 * IORING_OP_NOP just posts a completion event, nothing else.
3714 static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
3716 struct io_ring_ctx *ctx = req->ctx;
3718 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3721 __io_req_complete(req, issue_flags, 0, 0);
3725 static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3727 struct io_ring_ctx *ctx = req->ctx;
3732 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3734 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3737 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3738 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3741 req->sync.off = READ_ONCE(sqe->off);
3742 req->sync.len = READ_ONCE(sqe->len);
3746 static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
3748 loff_t end = req->sync.off + req->sync.len;
3751 /* fsync always requires a blocking context */
3752 if (issue_flags & IO_URING_F_NONBLOCK)
3755 ret = vfs_fsync_range(req->file, req->sync.off,
3756 end > 0 ? end : LLONG_MAX,
3757 req->sync.flags & IORING_FSYNC_DATASYNC);
3760 io_req_complete(req, ret);
3764 static int io_fallocate_prep(struct io_kiocb *req,
3765 const struct io_uring_sqe *sqe)
3767 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3769 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3772 req->sync.off = READ_ONCE(sqe->off);
3773 req->sync.len = READ_ONCE(sqe->addr);
3774 req->sync.mode = READ_ONCE(sqe->len);
3778 static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
3782 /* fallocate always requiring blocking context */
3783 if (issue_flags & IO_URING_F_NONBLOCK)
3785 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3789 io_req_complete(req, ret);
3793 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3795 const char __user *fname;
3798 if (unlikely(sqe->ioprio || sqe->buf_index))
3800 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3803 /* open.how should be already initialised */
3804 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3805 req->open.how.flags |= O_LARGEFILE;
3807 req->open.dfd = READ_ONCE(sqe->fd);
3808 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3809 req->open.filename = getname(fname);
3810 if (IS_ERR(req->open.filename)) {
3811 ret = PTR_ERR(req->open.filename);
3812 req->open.filename = NULL;
3815 req->open.nofile = rlimit(RLIMIT_NOFILE);
3816 req->flags |= REQ_F_NEED_CLEANUP;
3820 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3824 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3826 mode = READ_ONCE(sqe->len);
3827 flags = READ_ONCE(sqe->open_flags);
3828 req->open.how = build_open_how(flags, mode);
3829 return __io_openat_prep(req, sqe);
3832 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3834 struct open_how __user *how;
3838 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3840 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3841 len = READ_ONCE(sqe->len);
3842 if (len < OPEN_HOW_SIZE_VER0)
3845 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3850 return __io_openat_prep(req, sqe);
3853 static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
3855 struct open_flags op;
3858 bool resolve_nonblock;
3861 ret = build_open_flags(&req->open.how, &op);
3864 nonblock_set = op.open_flag & O_NONBLOCK;
3865 resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
3866 if (issue_flags & IO_URING_F_NONBLOCK) {
3868 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
3869 * it'll always -EAGAIN
3871 if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
3873 op.lookup_flags |= LOOKUP_CACHED;
3874 op.open_flag |= O_NONBLOCK;
3877 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3881 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3884 * We could hang on to this 'fd' on retrying, but seems like
3885 * marginal gain for something that is now known to be a slower
3886 * path. So just put it, and we'll get a new one when we retry.
3890 ret = PTR_ERR(file);
3891 /* only retry if RESOLVE_CACHED wasn't already set by application */
3892 if (ret == -EAGAIN &&
3893 (!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)))
3898 if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
3899 file->f_flags &= ~O_NONBLOCK;
3900 fsnotify_open(file);
3901 fd_install(ret, file);
3903 putname(req->open.filename);
3904 req->flags &= ~REQ_F_NEED_CLEANUP;
3907 __io_req_complete(req, issue_flags, ret, 0);
3911 static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
3913 return io_openat2(req, issue_flags);
3916 static int io_remove_buffers_prep(struct io_kiocb *req,
3917 const struct io_uring_sqe *sqe)
3919 struct io_provide_buf *p = &req->pbuf;
3922 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3925 tmp = READ_ONCE(sqe->fd);
3926 if (!tmp || tmp > USHRT_MAX)
3929 memset(p, 0, sizeof(*p));
3931 p->bgid = READ_ONCE(sqe->buf_group);
3935 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3936 int bgid, unsigned nbufs)
3940 /* shouldn't happen */
3944 /* the head kbuf is the list itself */
3945 while (!list_empty(&buf->list)) {
3946 struct io_buffer *nxt;
3948 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3949 list_del(&nxt->list);
3956 xa_erase(&ctx->io_buffers, bgid);
3961 static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
3963 struct io_provide_buf *p = &req->pbuf;
3964 struct io_ring_ctx *ctx = req->ctx;
3965 struct io_buffer *head;
3967 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3969 io_ring_submit_lock(ctx, !force_nonblock);
3971 lockdep_assert_held(&ctx->uring_lock);
3974 head = xa_load(&ctx->io_buffers, p->bgid);
3976 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3980 /* complete before unlock, IOPOLL may need the lock */
3981 __io_req_complete(req, issue_flags, ret, 0);
3982 io_ring_submit_unlock(ctx, !force_nonblock);
3986 static int io_provide_buffers_prep(struct io_kiocb *req,
3987 const struct io_uring_sqe *sqe)
3989 unsigned long size, tmp_check;
3990 struct io_provide_buf *p = &req->pbuf;
3993 if (sqe->ioprio || sqe->rw_flags)
3996 tmp = READ_ONCE(sqe->fd);
3997 if (!tmp || tmp > USHRT_MAX)
4000 p->addr = READ_ONCE(sqe->addr);
4001 p->len = READ_ONCE(sqe->len);
4003 if (check_mul_overflow((unsigned long)p->len, (unsigned long)p->nbufs,
4006 if (check_add_overflow((unsigned long)p->addr, size, &tmp_check))
4009 size = (unsigned long)p->len * p->nbufs;
4010 if (!access_ok(u64_to_user_ptr(p->addr), size))
4013 p->bgid = READ_ONCE(sqe->buf_group);
4014 tmp = READ_ONCE(sqe->off);
4015 if (tmp > USHRT_MAX)
4021 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
4023 struct io_buffer *buf;
4024 u64 addr = pbuf->addr;
4025 int i, bid = pbuf->bid;
4027 for (i = 0; i < pbuf->nbufs; i++) {
4028 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
4033 buf->len = min_t(__u32, pbuf->len, MAX_RW_COUNT);
4038 INIT_LIST_HEAD(&buf->list);
4041 list_add_tail(&buf->list, &(*head)->list);
4045 return i ? i : -ENOMEM;
4048 static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
4050 struct io_provide_buf *p = &req->pbuf;
4051 struct io_ring_ctx *ctx = req->ctx;
4052 struct io_buffer *head, *list;
4054 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4056 io_ring_submit_lock(ctx, !force_nonblock);
4058 lockdep_assert_held(&ctx->uring_lock);
4060 list = head = xa_load(&ctx->io_buffers, p->bgid);
4062 ret = io_add_buffers(p, &head);
4063 if (ret >= 0 && !list) {
4064 ret = xa_insert(&ctx->io_buffers, p->bgid, head, GFP_KERNEL);
4066 __io_remove_buffers(ctx, head, p->bgid, -1U);
4070 /* complete before unlock, IOPOLL may need the lock */
4071 __io_req_complete(req, issue_flags, ret, 0);
4072 io_ring_submit_unlock(ctx, !force_nonblock);
4076 static int io_epoll_ctl_prep(struct io_kiocb *req,
4077 const struct io_uring_sqe *sqe)
4079 #if defined(CONFIG_EPOLL)
4080 if (sqe->ioprio || sqe->buf_index)
4082 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4085 req->epoll.epfd = READ_ONCE(sqe->fd);
4086 req->epoll.op = READ_ONCE(sqe->len);
4087 req->epoll.fd = READ_ONCE(sqe->off);
4089 if (ep_op_has_event(req->epoll.op)) {
4090 struct epoll_event __user *ev;
4092 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4093 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4103 static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4105 #if defined(CONFIG_EPOLL)
4106 struct io_epoll *ie = &req->epoll;
4108 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4110 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4111 if (force_nonblock && ret == -EAGAIN)
4116 __io_req_complete(req, issue_flags, ret, 0);
4123 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4125 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4126 if (sqe->ioprio || sqe->buf_index || sqe->off)
4128 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4131 req->madvise.addr = READ_ONCE(sqe->addr);
4132 req->madvise.len = READ_ONCE(sqe->len);
4133 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4140 static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4142 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4143 struct io_madvise *ma = &req->madvise;
4146 if (issue_flags & IO_URING_F_NONBLOCK)
4149 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4152 io_req_complete(req, ret);
4159 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4161 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4163 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4166 req->fadvise.offset = READ_ONCE(sqe->off);
4167 req->fadvise.len = READ_ONCE(sqe->len);
4168 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4172 static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4174 struct io_fadvise *fa = &req->fadvise;
4177 if (issue_flags & IO_URING_F_NONBLOCK) {
4178 switch (fa->advice) {
4179 case POSIX_FADV_NORMAL:
4180 case POSIX_FADV_RANDOM:
4181 case POSIX_FADV_SEQUENTIAL:
4188 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4191 __io_req_complete(req, issue_flags, ret, 0);
4195 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4197 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4199 if (sqe->ioprio || sqe->buf_index)
4201 if (req->flags & REQ_F_FIXED_FILE)
4204 req->statx.dfd = READ_ONCE(sqe->fd);
4205 req->statx.mask = READ_ONCE(sqe->len);
4206 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4207 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4208 req->statx.flags = READ_ONCE(sqe->statx_flags);
4213 static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4215 struct io_statx *ctx = &req->statx;
4218 if (issue_flags & IO_URING_F_NONBLOCK)
4221 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4226 io_req_complete(req, ret);
4230 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4232 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4234 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4235 sqe->rw_flags || sqe->buf_index)
4237 if (req->flags & REQ_F_FIXED_FILE)
4240 req->close.fd = READ_ONCE(sqe->fd);
4244 static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4246 struct files_struct *files = current->files;
4247 struct io_close *close = &req->close;
4248 struct fdtable *fdt;
4249 struct file *file = NULL;
4252 spin_lock(&files->file_lock);
4253 fdt = files_fdtable(files);
4254 if (close->fd >= fdt->max_fds) {
4255 spin_unlock(&files->file_lock);
4258 file = fdt->fd[close->fd];
4259 if (!file || file->f_op == &io_uring_fops) {
4260 spin_unlock(&files->file_lock);
4265 /* if the file has a flush method, be safe and punt to async */
4266 if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4267 spin_unlock(&files->file_lock);
4271 ret = __close_fd_get_file(close->fd, &file);
4272 spin_unlock(&files->file_lock);
4279 /* No ->flush() or already async, safely close from here */
4280 ret = filp_close(file, current->files);
4286 __io_req_complete(req, issue_flags, ret, 0);
4290 static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4292 struct io_ring_ctx *ctx = req->ctx;
4294 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4296 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4299 req->sync.off = READ_ONCE(sqe->off);
4300 req->sync.len = READ_ONCE(sqe->len);
4301 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4305 static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4309 /* sync_file_range always requires a blocking context */
4310 if (issue_flags & IO_URING_F_NONBLOCK)
4313 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4317 io_req_complete(req, ret);
4321 #if defined(CONFIG_NET)
4322 static int io_setup_async_msg(struct io_kiocb *req,
4323 struct io_async_msghdr *kmsg)
4325 struct io_async_msghdr *async_msg = req->async_data;
4329 if (io_alloc_async_data(req)) {
4330 kfree(kmsg->free_iov);
4333 async_msg = req->async_data;
4334 req->flags |= REQ_F_NEED_CLEANUP;
4335 memcpy(async_msg, kmsg, sizeof(*kmsg));
4336 async_msg->msg.msg_name = &async_msg->addr;
4337 /* if were using fast_iov, set it to the new one */
4338 if (!async_msg->free_iov)
4339 async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4344 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4345 struct io_async_msghdr *iomsg)
4347 iomsg->msg.msg_name = &iomsg->addr;
4348 iomsg->free_iov = iomsg->fast_iov;
4349 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4350 req->sr_msg.msg_flags, &iomsg->free_iov);
4353 static int io_sendmsg_prep_async(struct io_kiocb *req)
4357 ret = io_sendmsg_copy_hdr(req, req->async_data);
4359 req->flags |= REQ_F_NEED_CLEANUP;
4363 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4365 struct io_sr_msg *sr = &req->sr_msg;
4367 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4370 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4371 sr->len = READ_ONCE(sqe->len);
4372 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4373 if (sr->msg_flags & MSG_DONTWAIT)
4374 req->flags |= REQ_F_NOWAIT;
4376 #ifdef CONFIG_COMPAT
4377 if (req->ctx->compat)
4378 sr->msg_flags |= MSG_CMSG_COMPAT;
4383 static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4385 struct io_async_msghdr iomsg, *kmsg;
4386 struct socket *sock;
4391 sock = sock_from_file(req->file);
4392 if (unlikely(!sock))
4395 kmsg = req->async_data;
4397 ret = io_sendmsg_copy_hdr(req, &iomsg);
4403 flags = req->sr_msg.msg_flags;
4404 if (issue_flags & IO_URING_F_NONBLOCK)
4405 flags |= MSG_DONTWAIT;
4406 if (flags & MSG_WAITALL)
4407 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4409 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4410 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4411 return io_setup_async_msg(req, kmsg);
4412 if (ret == -ERESTARTSYS)
4415 /* fast path, check for non-NULL to avoid function call */
4417 kfree(kmsg->free_iov);
4418 req->flags &= ~REQ_F_NEED_CLEANUP;
4421 __io_req_complete(req, issue_flags, ret, 0);
4425 static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4427 struct io_sr_msg *sr = &req->sr_msg;
4430 struct socket *sock;
4435 sock = sock_from_file(req->file);
4436 if (unlikely(!sock))
4439 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4443 msg.msg_name = NULL;
4444 msg.msg_control = NULL;
4445 msg.msg_controllen = 0;
4446 msg.msg_namelen = 0;
4448 flags = req->sr_msg.msg_flags;
4449 if (issue_flags & IO_URING_F_NONBLOCK)
4450 flags |= MSG_DONTWAIT;
4451 if (flags & MSG_WAITALL)
4452 min_ret = iov_iter_count(&msg.msg_iter);
4454 msg.msg_flags = flags;
4455 ret = sock_sendmsg(sock, &msg);
4456 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4458 if (ret == -ERESTARTSYS)
4463 __io_req_complete(req, issue_flags, ret, 0);
4467 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4468 struct io_async_msghdr *iomsg)
4470 struct io_sr_msg *sr = &req->sr_msg;
4471 struct iovec __user *uiov;
4475 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4476 &iomsg->uaddr, &uiov, &iov_len);
4480 if (req->flags & REQ_F_BUFFER_SELECT) {
4483 if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4485 sr->len = iomsg->fast_iov[0].iov_len;
4486 iomsg->free_iov = NULL;
4488 iomsg->free_iov = iomsg->fast_iov;
4489 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4490 &iomsg->free_iov, &iomsg->msg.msg_iter,
4499 #ifdef CONFIG_COMPAT
4500 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4501 struct io_async_msghdr *iomsg)
4503 struct io_sr_msg *sr = &req->sr_msg;
4504 struct compat_iovec __user *uiov;
4509 ret = __get_compat_msghdr(&iomsg->msg, sr->umsg_compat, &iomsg->uaddr,
4514 uiov = compat_ptr(ptr);
4515 if (req->flags & REQ_F_BUFFER_SELECT) {
4516 compat_ssize_t clen;
4520 if (!access_ok(uiov, sizeof(*uiov)))
4522 if (__get_user(clen, &uiov->iov_len))
4527 iomsg->free_iov = NULL;
4529 iomsg->free_iov = iomsg->fast_iov;
4530 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4531 UIO_FASTIOV, &iomsg->free_iov,
4532 &iomsg->msg.msg_iter, true);
4541 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4542 struct io_async_msghdr *iomsg)
4544 iomsg->msg.msg_name = &iomsg->addr;
4546 #ifdef CONFIG_COMPAT
4547 if (req->ctx->compat)
4548 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4551 return __io_recvmsg_copy_hdr(req, iomsg);
4554 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4557 struct io_sr_msg *sr = &req->sr_msg;
4558 struct io_buffer *kbuf;
4560 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4565 req->flags |= REQ_F_BUFFER_SELECTED;
4569 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4571 return io_put_kbuf(req, req->sr_msg.kbuf);
4574 static int io_recvmsg_prep_async(struct io_kiocb *req)
4578 ret = io_recvmsg_copy_hdr(req, req->async_data);
4580 req->flags |= REQ_F_NEED_CLEANUP;
4584 static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4586 struct io_sr_msg *sr = &req->sr_msg;
4588 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4591 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4592 sr->len = READ_ONCE(sqe->len);
4593 sr->bgid = READ_ONCE(sqe->buf_group);
4594 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4595 if (sr->msg_flags & MSG_DONTWAIT)
4596 req->flags |= REQ_F_NOWAIT;
4598 #ifdef CONFIG_COMPAT
4599 if (req->ctx->compat)
4600 sr->msg_flags |= MSG_CMSG_COMPAT;
4605 static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
4607 struct io_async_msghdr iomsg, *kmsg;
4608 struct socket *sock;
4609 struct io_buffer *kbuf;
4612 int ret, cflags = 0;
4613 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4615 sock = sock_from_file(req->file);
4616 if (unlikely(!sock))
4619 kmsg = req->async_data;
4621 ret = io_recvmsg_copy_hdr(req, &iomsg);
4627 if (req->flags & REQ_F_BUFFER_SELECT) {
4628 kbuf = io_recv_buffer_select(req, !force_nonblock);
4630 return PTR_ERR(kbuf);
4631 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4632 kmsg->fast_iov[0].iov_len = req->sr_msg.len;
4633 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
4634 1, req->sr_msg.len);
4637 flags = req->sr_msg.msg_flags;
4639 flags |= MSG_DONTWAIT;
4640 if (flags & MSG_WAITALL)
4641 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4643 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4644 kmsg->uaddr, flags);
4645 if (force_nonblock && ret == -EAGAIN)
4646 return io_setup_async_msg(req, kmsg);
4647 if (ret == -ERESTARTSYS)
4650 if (req->flags & REQ_F_BUFFER_SELECTED)
4651 cflags = io_put_recv_kbuf(req);
4652 /* fast path, check for non-NULL to avoid function call */
4654 kfree(kmsg->free_iov);
4655 req->flags &= ~REQ_F_NEED_CLEANUP;
4656 if (ret < min_ret || ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4658 __io_req_complete(req, issue_flags, ret, cflags);
4662 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
4664 struct io_buffer *kbuf;
4665 struct io_sr_msg *sr = &req->sr_msg;
4667 void __user *buf = sr->buf;
4668 struct socket *sock;
4672 int ret, cflags = 0;
4673 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4675 sock = sock_from_file(req->file);
4676 if (unlikely(!sock))
4679 if (req->flags & REQ_F_BUFFER_SELECT) {
4680 kbuf = io_recv_buffer_select(req, !force_nonblock);
4682 return PTR_ERR(kbuf);
4683 buf = u64_to_user_ptr(kbuf->addr);
4686 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4690 msg.msg_name = NULL;
4691 msg.msg_control = NULL;
4692 msg.msg_controllen = 0;
4693 msg.msg_namelen = 0;
4694 msg.msg_iocb = NULL;
4697 flags = req->sr_msg.msg_flags;
4699 flags |= MSG_DONTWAIT;
4700 if (flags & MSG_WAITALL)
4701 min_ret = iov_iter_count(&msg.msg_iter);
4703 ret = sock_recvmsg(sock, &msg, flags);
4704 if (force_nonblock && ret == -EAGAIN)
4706 if (ret == -ERESTARTSYS)
4709 if (req->flags & REQ_F_BUFFER_SELECTED)
4710 cflags = io_put_recv_kbuf(req);
4711 if (ret < min_ret || ((flags & MSG_WAITALL) && (msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4713 __io_req_complete(req, issue_flags, ret, cflags);
4717 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4719 struct io_accept *accept = &req->accept;
4721 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4723 if (sqe->ioprio || sqe->len || sqe->buf_index)
4726 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4727 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4728 accept->flags = READ_ONCE(sqe->accept_flags);
4729 accept->nofile = rlimit(RLIMIT_NOFILE);
4733 static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
4735 struct io_accept *accept = &req->accept;
4736 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4737 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4740 if (req->file->f_flags & O_NONBLOCK)
4741 req->flags |= REQ_F_NOWAIT;
4743 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4744 accept->addr_len, accept->flags,
4746 if (ret == -EAGAIN && force_nonblock)
4749 if (ret == -ERESTARTSYS)
4753 __io_req_complete(req, issue_flags, ret, 0);
4757 static int io_connect_prep_async(struct io_kiocb *req)
4759 struct io_async_connect *io = req->async_data;
4760 struct io_connect *conn = &req->connect;
4762 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
4765 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4767 struct io_connect *conn = &req->connect;
4769 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4771 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4774 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4775 conn->addr_len = READ_ONCE(sqe->addr2);
4779 static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
4781 struct io_async_connect __io, *io;
4782 unsigned file_flags;
4784 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4786 if (req->async_data) {
4787 io = req->async_data;
4789 ret = move_addr_to_kernel(req->connect.addr,
4790 req->connect.addr_len,
4797 file_flags = force_nonblock ? O_NONBLOCK : 0;
4799 ret = __sys_connect_file(req->file, &io->address,
4800 req->connect.addr_len, file_flags);
4801 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4802 if (req->async_data)
4804 if (io_alloc_async_data(req)) {
4808 memcpy(req->async_data, &__io, sizeof(__io));
4811 if (ret == -ERESTARTSYS)
4816 __io_req_complete(req, issue_flags, ret, 0);
4819 #else /* !CONFIG_NET */
4820 #define IO_NETOP_FN(op) \
4821 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
4823 return -EOPNOTSUPP; \
4826 #define IO_NETOP_PREP(op) \
4828 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
4830 return -EOPNOTSUPP; \
4833 #define IO_NETOP_PREP_ASYNC(op) \
4835 static int io_##op##_prep_async(struct io_kiocb *req) \
4837 return -EOPNOTSUPP; \
4840 IO_NETOP_PREP_ASYNC(sendmsg);
4841 IO_NETOP_PREP_ASYNC(recvmsg);
4842 IO_NETOP_PREP_ASYNC(connect);
4843 IO_NETOP_PREP(accept);
4846 #endif /* CONFIG_NET */
4848 struct io_poll_table {
4849 struct poll_table_struct pt;
4850 struct io_kiocb *req;
4855 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4856 __poll_t mask, io_req_tw_func_t func)
4858 /* for instances that support it check for an event match first: */
4859 if (mask && !(mask & poll->events))
4862 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4864 list_del_init(&poll->wait.entry);
4867 req->io_task_work.func = func;
4870 * If this fails, then the task is exiting. When a task exits, the
4871 * work gets canceled, so just cancel this request as well instead
4872 * of executing it. We can't safely execute it anyway, as we may not
4873 * have the needed state needed for it anyway.
4875 io_req_task_work_add(req);
4879 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4880 __acquires(&req->ctx->completion_lock)
4882 struct io_ring_ctx *ctx = req->ctx;
4884 if (unlikely(req->task->flags & PF_EXITING))
4885 WRITE_ONCE(poll->canceled, true);
4887 if (!req->result && !READ_ONCE(poll->canceled)) {
4888 struct poll_table_struct pt = { ._key = poll->events };
4890 req->result = vfs_poll(req->file, &pt) & poll->events;
4893 spin_lock_irq(&ctx->completion_lock);
4894 if (!req->result && !READ_ONCE(poll->canceled)) {
4895 add_wait_queue(poll->head, &poll->wait);
4902 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4904 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4905 if (req->opcode == IORING_OP_POLL_ADD)
4906 return req->async_data;
4907 return req->apoll->double_poll;
4910 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4912 if (req->opcode == IORING_OP_POLL_ADD)
4914 return &req->apoll->poll;
4917 static void io_poll_remove_double(struct io_kiocb *req)
4918 __must_hold(&req->ctx->completion_lock)
4920 struct io_poll_iocb *poll = io_poll_get_double(req);
4922 lockdep_assert_held(&req->ctx->completion_lock);
4924 if (poll && poll->head) {
4925 struct wait_queue_head *head = poll->head;
4927 spin_lock(&head->lock);
4928 list_del_init(&poll->wait.entry);
4929 if (poll->wait.private)
4932 spin_unlock(&head->lock);
4936 static bool io_poll_complete(struct io_kiocb *req, __poll_t mask)
4937 __must_hold(&req->ctx->completion_lock)
4939 struct io_ring_ctx *ctx = req->ctx;
4940 unsigned flags = IORING_CQE_F_MORE;
4943 if (READ_ONCE(req->poll.canceled)) {
4945 req->poll.events |= EPOLLONESHOT;
4947 error = mangle_poll(mask);
4949 if (req->poll.events & EPOLLONESHOT)
4951 if (!io_cqring_fill_event(ctx, req->user_data, error, flags)) {
4952 req->poll.done = true;
4955 if (flags & IORING_CQE_F_MORE)
4958 io_commit_cqring(ctx);
4959 return !(flags & IORING_CQE_F_MORE);
4962 static void io_poll_task_func(struct io_kiocb *req)
4964 struct io_ring_ctx *ctx = req->ctx;
4965 struct io_kiocb *nxt;
4967 if (io_poll_rewait(req, &req->poll)) {
4968 spin_unlock_irq(&ctx->completion_lock);
4972 done = io_poll_complete(req, req->result);
4974 io_poll_remove_double(req);
4975 hash_del(&req->hash_node);
4978 add_wait_queue(req->poll.head, &req->poll.wait);
4980 spin_unlock_irq(&ctx->completion_lock);
4981 io_cqring_ev_posted(ctx);
4984 nxt = io_put_req_find_next(req);
4986 io_req_task_submit(nxt);
4991 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4992 int sync, void *key)
4994 struct io_kiocb *req = wait->private;
4995 struct io_poll_iocb *poll = io_poll_get_single(req);
4996 __poll_t mask = key_to_poll(key);
4998 /* for instances that support it check for an event match first: */
4999 if (mask && !(mask & poll->events))
5001 if (!(poll->events & EPOLLONESHOT))
5002 return poll->wait.func(&poll->wait, mode, sync, key);
5004 list_del_init(&wait->entry);
5009 spin_lock(&poll->head->lock);
5010 done = list_empty(&poll->wait.entry);
5012 list_del_init(&poll->wait.entry);
5013 /* make sure double remove sees this as being gone */
5014 wait->private = NULL;
5015 spin_unlock(&poll->head->lock);
5017 /* use wait func handler, so it matches the rq type */
5018 poll->wait.func(&poll->wait, mode, sync, key);
5025 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
5026 wait_queue_func_t wake_func)
5030 poll->canceled = false;
5031 #define IO_POLL_UNMASK (EPOLLERR|EPOLLHUP|EPOLLNVAL|EPOLLRDHUP)
5032 /* mask in events that we always want/need */
5033 poll->events = events | IO_POLL_UNMASK;
5034 INIT_LIST_HEAD(&poll->wait.entry);
5035 init_waitqueue_func_entry(&poll->wait, wake_func);
5038 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5039 struct wait_queue_head *head,
5040 struct io_poll_iocb **poll_ptr)
5042 struct io_kiocb *req = pt->req;
5045 * The file being polled uses multiple waitqueues for poll handling
5046 * (e.g. one for read, one for write). Setup a separate io_poll_iocb
5049 if (unlikely(pt->nr_entries)) {
5050 struct io_poll_iocb *poll_one = poll;
5052 /* already have a 2nd entry, fail a third attempt */
5054 pt->error = -EINVAL;
5058 * Can't handle multishot for double wait for now, turn it
5059 * into one-shot mode.
5061 if (!(poll_one->events & EPOLLONESHOT))
5062 poll_one->events |= EPOLLONESHOT;
5063 /* double add on the same waitqueue head, ignore */
5064 if (poll_one->head == head)
5066 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5068 pt->error = -ENOMEM;
5071 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5073 poll->wait.private = req;
5080 if (poll->events & EPOLLEXCLUSIVE)
5081 add_wait_queue_exclusive(head, &poll->wait);
5083 add_wait_queue(head, &poll->wait);
5086 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5087 struct poll_table_struct *p)
5089 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5090 struct async_poll *apoll = pt->req->apoll;
5092 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5095 static void io_async_task_func(struct io_kiocb *req)
5097 struct async_poll *apoll = req->apoll;
5098 struct io_ring_ctx *ctx = req->ctx;
5100 trace_io_uring_task_run(req->ctx, req, req->opcode, req->user_data);
5102 if (io_poll_rewait(req, &apoll->poll)) {
5103 spin_unlock_irq(&ctx->completion_lock);
5107 hash_del(&req->hash_node);
5108 io_poll_remove_double(req);
5109 spin_unlock_irq(&ctx->completion_lock);
5111 if (!READ_ONCE(apoll->poll.canceled))
5112 io_req_task_submit(req);
5114 io_req_complete_failed(req, -ECANCELED);
5117 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5120 struct io_kiocb *req = wait->private;
5121 struct io_poll_iocb *poll = &req->apoll->poll;
5123 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5126 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5129 static void io_poll_req_insert(struct io_kiocb *req)
5131 struct io_ring_ctx *ctx = req->ctx;
5132 struct hlist_head *list;
5134 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5135 hlist_add_head(&req->hash_node, list);
5138 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5139 struct io_poll_iocb *poll,
5140 struct io_poll_table *ipt, __poll_t mask,
5141 wait_queue_func_t wake_func)
5142 __acquires(&ctx->completion_lock)
5144 struct io_ring_ctx *ctx = req->ctx;
5145 bool cancel = false;
5147 INIT_HLIST_NODE(&req->hash_node);
5148 io_init_poll_iocb(poll, mask, wake_func);
5149 poll->file = req->file;
5150 poll->wait.private = req;
5152 ipt->pt._key = mask;
5155 ipt->nr_entries = 0;
5157 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5158 if (unlikely(!ipt->nr_entries) && !ipt->error)
5159 ipt->error = -EINVAL;
5161 spin_lock_irq(&ctx->completion_lock);
5162 if (ipt->error || (mask && (poll->events & EPOLLONESHOT)))
5163 io_poll_remove_double(req);
5164 if (likely(poll->head)) {
5165 spin_lock(&poll->head->lock);
5166 if (unlikely(list_empty(&poll->wait.entry))) {
5172 if ((mask && (poll->events & EPOLLONESHOT)) || ipt->error)
5173 list_del_init(&poll->wait.entry);
5175 WRITE_ONCE(poll->canceled, true);
5176 else if (!poll->done) /* actually waiting for an event */
5177 io_poll_req_insert(req);
5178 spin_unlock(&poll->head->lock);
5190 static int io_arm_poll_handler(struct io_kiocb *req)
5192 const struct io_op_def *def = &io_op_defs[req->opcode];
5193 struct io_ring_ctx *ctx = req->ctx;
5194 struct async_poll *apoll;
5195 struct io_poll_table ipt;
5196 __poll_t ret, mask = EPOLLONESHOT | POLLERR | POLLPRI;
5199 if (!req->file || !file_can_poll(req->file))
5200 return IO_APOLL_ABORTED;
5201 if (req->flags & REQ_F_POLLED)
5202 return IO_APOLL_ABORTED;
5203 if (!def->pollin && !def->pollout)
5204 return IO_APOLL_ABORTED;
5208 mask |= POLLIN | POLLRDNORM;
5210 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5211 if ((req->opcode == IORING_OP_RECVMSG) &&
5212 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5216 mask |= POLLOUT | POLLWRNORM;
5219 /* if we can't nonblock try, then no point in arming a poll handler */
5220 if (!io_file_supports_nowait(req, rw))
5221 return IO_APOLL_ABORTED;
5223 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5224 if (unlikely(!apoll))
5225 return IO_APOLL_ABORTED;
5226 apoll->double_poll = NULL;
5228 req->flags |= REQ_F_POLLED;
5229 ipt.pt._qproc = io_async_queue_proc;
5231 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5233 if (ret || ipt.error) {
5234 spin_unlock_irq(&ctx->completion_lock);
5236 return IO_APOLL_READY;
5237 return IO_APOLL_ABORTED;
5239 spin_unlock_irq(&ctx->completion_lock);
5240 trace_io_uring_poll_arm(ctx, req, req->opcode, req->user_data,
5241 mask, apoll->poll.events);
5245 static bool __io_poll_remove_one(struct io_kiocb *req,
5246 struct io_poll_iocb *poll, bool do_cancel)
5247 __must_hold(&req->ctx->completion_lock)
5249 bool do_complete = false;
5253 spin_lock(&poll->head->lock);
5255 WRITE_ONCE(poll->canceled, true);
5256 if (!list_empty(&poll->wait.entry)) {
5257 list_del_init(&poll->wait.entry);
5260 spin_unlock(&poll->head->lock);
5261 hash_del(&req->hash_node);
5265 static bool io_poll_remove_waitqs(struct io_kiocb *req)
5266 __must_hold(&req->ctx->completion_lock)
5270 io_poll_remove_double(req);
5271 do_complete = __io_poll_remove_one(req, io_poll_get_single(req), true);
5273 if (req->opcode != IORING_OP_POLL_ADD && do_complete) {
5274 /* non-poll requests have submit ref still */
5280 static bool io_poll_remove_one(struct io_kiocb *req)
5281 __must_hold(&req->ctx->completion_lock)
5285 do_complete = io_poll_remove_waitqs(req);
5287 io_cqring_fill_event(req->ctx, req->user_data, -ECANCELED, 0);
5288 io_commit_cqring(req->ctx);
5290 io_put_req_deferred(req, 1);
5297 * Returns true if we found and killed one or more poll requests
5299 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5302 struct hlist_node *tmp;
5303 struct io_kiocb *req;
5306 spin_lock_irq(&ctx->completion_lock);
5307 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5308 struct hlist_head *list;
5310 list = &ctx->cancel_hash[i];
5311 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5312 if (io_match_task(req, tsk, cancel_all))
5313 posted += io_poll_remove_one(req);
5316 spin_unlock_irq(&ctx->completion_lock);
5319 io_cqring_ev_posted(ctx);
5324 static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, __u64 sqe_addr,
5326 __must_hold(&ctx->completion_lock)
5328 struct hlist_head *list;
5329 struct io_kiocb *req;
5331 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5332 hlist_for_each_entry(req, list, hash_node) {
5333 if (sqe_addr != req->user_data)
5335 if (poll_only && req->opcode != IORING_OP_POLL_ADD)
5342 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr,
5344 __must_hold(&ctx->completion_lock)
5346 struct io_kiocb *req;
5348 req = io_poll_find(ctx, sqe_addr, poll_only);
5351 if (io_poll_remove_one(req))
5357 static __poll_t io_poll_parse_events(const struct io_uring_sqe *sqe,
5362 events = READ_ONCE(sqe->poll32_events);
5364 events = swahw32(events);
5366 if (!(flags & IORING_POLL_ADD_MULTI))
5367 events |= EPOLLONESHOT;
5368 return demangle_poll(events) | (events & (EPOLLEXCLUSIVE|EPOLLONESHOT));
5371 static int io_poll_update_prep(struct io_kiocb *req,
5372 const struct io_uring_sqe *sqe)
5374 struct io_poll_update *upd = &req->poll_update;
5377 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5379 if (sqe->ioprio || sqe->buf_index)
5381 flags = READ_ONCE(sqe->len);
5382 if (flags & ~(IORING_POLL_UPDATE_EVENTS | IORING_POLL_UPDATE_USER_DATA |
5383 IORING_POLL_ADD_MULTI))
5385 /* meaningless without update */
5386 if (flags == IORING_POLL_ADD_MULTI)
5389 upd->old_user_data = READ_ONCE(sqe->addr);
5390 upd->update_events = flags & IORING_POLL_UPDATE_EVENTS;
5391 upd->update_user_data = flags & IORING_POLL_UPDATE_USER_DATA;
5393 upd->new_user_data = READ_ONCE(sqe->off);
5394 if (!upd->update_user_data && upd->new_user_data)
5396 if (upd->update_events)
5397 upd->events = io_poll_parse_events(sqe, flags);
5398 else if (sqe->poll32_events)
5404 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5407 struct io_kiocb *req = wait->private;
5408 struct io_poll_iocb *poll = &req->poll;
5410 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5413 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5414 struct poll_table_struct *p)
5416 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5418 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5421 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5423 struct io_poll_iocb *poll = &req->poll;
5426 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5428 if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->addr)
5430 flags = READ_ONCE(sqe->len);
5431 if (flags & ~IORING_POLL_ADD_MULTI)
5434 poll->events = io_poll_parse_events(sqe, flags);
5438 static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5440 struct io_poll_iocb *poll = &req->poll;
5441 struct io_ring_ctx *ctx = req->ctx;
5442 struct io_poll_table ipt;
5445 ipt.pt._qproc = io_poll_queue_proc;
5447 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5450 if (mask) { /* no async, we'd stolen it */
5452 io_poll_complete(req, mask);
5454 spin_unlock_irq(&ctx->completion_lock);
5457 io_cqring_ev_posted(ctx);
5458 if (poll->events & EPOLLONESHOT)
5464 static int io_poll_update(struct io_kiocb *req, unsigned int issue_flags)
5466 struct io_ring_ctx *ctx = req->ctx;
5467 struct io_kiocb *preq;
5471 spin_lock_irq(&ctx->completion_lock);
5472 preq = io_poll_find(ctx, req->poll_update.old_user_data, true);
5478 if (!req->poll_update.update_events && !req->poll_update.update_user_data) {
5480 ret = io_poll_remove_one(preq) ? 0 : -EALREADY;
5485 * Don't allow racy completion with singleshot, as we cannot safely
5486 * update those. For multishot, if we're racing with completion, just
5487 * let completion re-add it.
5489 completing = !__io_poll_remove_one(preq, &preq->poll, false);
5490 if (completing && (preq->poll.events & EPOLLONESHOT)) {
5494 /* we now have a detached poll request. reissue. */
5498 spin_unlock_irq(&ctx->completion_lock);
5500 io_req_complete(req, ret);
5503 /* only mask one event flags, keep behavior flags */
5504 if (req->poll_update.update_events) {
5505 preq->poll.events &= ~0xffff;
5506 preq->poll.events |= req->poll_update.events & 0xffff;
5507 preq->poll.events |= IO_POLL_UNMASK;
5509 if (req->poll_update.update_user_data)
5510 preq->user_data = req->poll_update.new_user_data;
5511 spin_unlock_irq(&ctx->completion_lock);
5513 /* complete update request, we're done with it */
5514 io_req_complete(req, ret);
5517 ret = io_poll_add(preq, issue_flags);
5520 io_req_complete(preq, ret);
5526 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5528 struct io_timeout_data *data = container_of(timer,
5529 struct io_timeout_data, timer);
5530 struct io_kiocb *req = data->req;
5531 struct io_ring_ctx *ctx = req->ctx;
5532 unsigned long flags;
5534 spin_lock_irqsave(&ctx->completion_lock, flags);
5535 list_del_init(&req->timeout.list);
5536 atomic_set(&req->ctx->cq_timeouts,
5537 atomic_read(&req->ctx->cq_timeouts) + 1);
5539 io_cqring_fill_event(ctx, req->user_data, -ETIME, 0);
5540 io_commit_cqring(ctx);
5541 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5543 io_cqring_ev_posted(ctx);
5546 return HRTIMER_NORESTART;
5549 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5551 __must_hold(&ctx->completion_lock)
5553 struct io_timeout_data *io;
5554 struct io_kiocb *req;
5557 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5558 found = user_data == req->user_data;
5563 return ERR_PTR(-ENOENT);
5565 io = req->async_data;
5566 if (hrtimer_try_to_cancel(&io->timer) == -1)
5567 return ERR_PTR(-EALREADY);
5568 list_del_init(&req->timeout.list);
5572 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5573 __must_hold(&ctx->completion_lock)
5575 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5578 return PTR_ERR(req);
5581 io_cqring_fill_event(ctx, req->user_data, -ECANCELED, 0);
5582 io_put_req_deferred(req, 1);
5586 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5587 struct timespec64 *ts, enum hrtimer_mode mode)
5588 __must_hold(&ctx->completion_lock)
5590 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5591 struct io_timeout_data *data;
5594 return PTR_ERR(req);
5596 req->timeout.off = 0; /* noseq */
5597 data = req->async_data;
5598 list_add_tail(&req->timeout.list, &ctx->timeout_list);
5599 hrtimer_init(&data->timer, CLOCK_MONOTONIC, mode);
5600 data->timer.function = io_timeout_fn;
5601 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
5605 static int io_timeout_remove_prep(struct io_kiocb *req,
5606 const struct io_uring_sqe *sqe)
5608 struct io_timeout_rem *tr = &req->timeout_rem;
5610 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5612 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5614 if (sqe->ioprio || sqe->buf_index || sqe->len)
5617 tr->addr = READ_ONCE(sqe->addr);
5618 tr->flags = READ_ONCE(sqe->timeout_flags);
5619 if (tr->flags & IORING_TIMEOUT_UPDATE) {
5620 if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS))
5622 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
5624 } else if (tr->flags) {
5625 /* timeout removal doesn't support flags */
5632 static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
5634 return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
5639 * Remove or update an existing timeout command
5641 static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
5643 struct io_timeout_rem *tr = &req->timeout_rem;
5644 struct io_ring_ctx *ctx = req->ctx;
5647 spin_lock_irq(&ctx->completion_lock);
5648 if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE))
5649 ret = io_timeout_cancel(ctx, tr->addr);
5651 ret = io_timeout_update(ctx, tr->addr, &tr->ts,
5652 io_translate_timeout_mode(tr->flags));
5654 io_cqring_fill_event(ctx, req->user_data, ret, 0);
5655 io_commit_cqring(ctx);
5656 spin_unlock_irq(&ctx->completion_lock);
5657 io_cqring_ev_posted(ctx);
5664 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5665 bool is_timeout_link)
5667 struct io_timeout_data *data;
5669 u32 off = READ_ONCE(sqe->off);
5671 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5673 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5675 if (off && is_timeout_link)
5677 flags = READ_ONCE(sqe->timeout_flags);
5678 if (flags & ~IORING_TIMEOUT_ABS)
5681 req->timeout.off = off;
5682 if (unlikely(off && !req->ctx->off_timeout_used))
5683 req->ctx->off_timeout_used = true;
5685 if (!req->async_data && io_alloc_async_data(req))
5688 data = req->async_data;
5691 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5694 data->mode = io_translate_timeout_mode(flags);
5695 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5696 if (is_timeout_link)
5697 io_req_track_inflight(req);
5701 static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
5703 struct io_ring_ctx *ctx = req->ctx;
5704 struct io_timeout_data *data = req->async_data;
5705 struct list_head *entry;
5706 u32 tail, off = req->timeout.off;
5708 spin_lock_irq(&ctx->completion_lock);
5711 * sqe->off holds how many events that need to occur for this
5712 * timeout event to be satisfied. If it isn't set, then this is
5713 * a pure timeout request, sequence isn't used.
5715 if (io_is_timeout_noseq(req)) {
5716 entry = ctx->timeout_list.prev;
5720 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5721 req->timeout.target_seq = tail + off;
5723 /* Update the last seq here in case io_flush_timeouts() hasn't.
5724 * This is safe because ->completion_lock is held, and submissions
5725 * and completions are never mixed in the same ->completion_lock section.
5727 ctx->cq_last_tm_flush = tail;
5730 * Insertion sort, ensuring the first entry in the list is always
5731 * the one we need first.
5733 list_for_each_prev(entry, &ctx->timeout_list) {
5734 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5737 if (io_is_timeout_noseq(nxt))
5739 /* nxt.seq is behind @tail, otherwise would've been completed */
5740 if (off >= nxt->timeout.target_seq - tail)
5744 list_add(&req->timeout.list, entry);
5745 data->timer.function = io_timeout_fn;
5746 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5747 spin_unlock_irq(&ctx->completion_lock);
5751 struct io_cancel_data {
5752 struct io_ring_ctx *ctx;
5756 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5758 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5759 struct io_cancel_data *cd = data;
5761 return req->ctx == cd->ctx && req->user_data == cd->user_data;
5764 static int io_async_cancel_one(struct io_uring_task *tctx, u64 user_data,
5765 struct io_ring_ctx *ctx)
5767 struct io_cancel_data data = { .ctx = ctx, .user_data = user_data, };
5768 enum io_wq_cancel cancel_ret;
5771 if (!tctx || !tctx->io_wq)
5774 cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, &data, false);
5775 switch (cancel_ret) {
5776 case IO_WQ_CANCEL_OK:
5779 case IO_WQ_CANCEL_RUNNING:
5782 case IO_WQ_CANCEL_NOTFOUND:
5790 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5791 struct io_kiocb *req, __u64 sqe_addr,
5794 unsigned long flags;
5797 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5798 spin_lock_irqsave(&ctx->completion_lock, flags);
5801 ret = io_timeout_cancel(ctx, sqe_addr);
5804 ret = io_poll_cancel(ctx, sqe_addr, false);
5808 io_cqring_fill_event(ctx, req->user_data, ret, 0);
5809 io_commit_cqring(ctx);
5810 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5811 io_cqring_ev_posted(ctx);
5817 static int io_async_cancel_prep(struct io_kiocb *req,
5818 const struct io_uring_sqe *sqe)
5820 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5822 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5824 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5827 req->cancel.addr = READ_ONCE(sqe->addr);
5831 static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
5833 struct io_ring_ctx *ctx = req->ctx;
5834 u64 sqe_addr = req->cancel.addr;
5835 struct io_tctx_node *node;
5838 /* tasks should wait for their io-wq threads, so safe w/o sync */
5839 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5840 spin_lock_irq(&ctx->completion_lock);
5843 ret = io_timeout_cancel(ctx, sqe_addr);
5846 ret = io_poll_cancel(ctx, sqe_addr, false);
5849 spin_unlock_irq(&ctx->completion_lock);
5851 /* slow path, try all io-wq's */
5852 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5854 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
5855 struct io_uring_task *tctx = node->task->io_uring;
5857 ret = io_async_cancel_one(tctx, req->cancel.addr, ctx);
5861 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5863 spin_lock_irq(&ctx->completion_lock);
5865 io_cqring_fill_event(ctx, req->user_data, ret, 0);
5866 io_commit_cqring(ctx);
5867 spin_unlock_irq(&ctx->completion_lock);
5868 io_cqring_ev_posted(ctx);
5876 static int io_rsrc_update_prep(struct io_kiocb *req,
5877 const struct io_uring_sqe *sqe)
5879 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5881 if (sqe->ioprio || sqe->rw_flags)
5884 req->rsrc_update.offset = READ_ONCE(sqe->off);
5885 req->rsrc_update.nr_args = READ_ONCE(sqe->len);
5886 if (!req->rsrc_update.nr_args)
5888 req->rsrc_update.arg = READ_ONCE(sqe->addr);
5892 static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
5894 struct io_ring_ctx *ctx = req->ctx;
5895 struct io_uring_rsrc_update2 up;
5898 if (issue_flags & IO_URING_F_NONBLOCK)
5901 up.offset = req->rsrc_update.offset;
5902 up.data = req->rsrc_update.arg;
5907 mutex_lock(&ctx->uring_lock);
5908 ret = __io_register_rsrc_update(ctx, IORING_RSRC_FILE,
5909 &up, req->rsrc_update.nr_args);
5910 mutex_unlock(&ctx->uring_lock);
5914 __io_req_complete(req, issue_flags, ret, 0);
5918 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5920 switch (req->opcode) {
5923 case IORING_OP_READV:
5924 case IORING_OP_READ_FIXED:
5925 case IORING_OP_READ:
5926 return io_read_prep(req, sqe);
5927 case IORING_OP_WRITEV:
5928 case IORING_OP_WRITE_FIXED:
5929 case IORING_OP_WRITE:
5930 return io_write_prep(req, sqe);
5931 case IORING_OP_POLL_ADD:
5932 return io_poll_add_prep(req, sqe);
5933 case IORING_OP_POLL_REMOVE:
5934 return io_poll_update_prep(req, sqe);
5935 case IORING_OP_FSYNC:
5936 return io_fsync_prep(req, sqe);
5937 case IORING_OP_SYNC_FILE_RANGE:
5938 return io_sfr_prep(req, sqe);
5939 case IORING_OP_SENDMSG:
5940 case IORING_OP_SEND:
5941 return io_sendmsg_prep(req, sqe);
5942 case IORING_OP_RECVMSG:
5943 case IORING_OP_RECV:
5944 return io_recvmsg_prep(req, sqe);
5945 case IORING_OP_CONNECT:
5946 return io_connect_prep(req, sqe);
5947 case IORING_OP_TIMEOUT:
5948 return io_timeout_prep(req, sqe, false);
5949 case IORING_OP_TIMEOUT_REMOVE:
5950 return io_timeout_remove_prep(req, sqe);
5951 case IORING_OP_ASYNC_CANCEL:
5952 return io_async_cancel_prep(req, sqe);
5953 case IORING_OP_LINK_TIMEOUT:
5954 return io_timeout_prep(req, sqe, true);
5955 case IORING_OP_ACCEPT:
5956 return io_accept_prep(req, sqe);
5957 case IORING_OP_FALLOCATE:
5958 return io_fallocate_prep(req, sqe);
5959 case IORING_OP_OPENAT:
5960 return io_openat_prep(req, sqe);
5961 case IORING_OP_CLOSE:
5962 return io_close_prep(req, sqe);
5963 case IORING_OP_FILES_UPDATE:
5964 return io_rsrc_update_prep(req, sqe);
5965 case IORING_OP_STATX:
5966 return io_statx_prep(req, sqe);
5967 case IORING_OP_FADVISE:
5968 return io_fadvise_prep(req, sqe);
5969 case IORING_OP_MADVISE:
5970 return io_madvise_prep(req, sqe);
5971 case IORING_OP_OPENAT2:
5972 return io_openat2_prep(req, sqe);
5973 case IORING_OP_EPOLL_CTL:
5974 return io_epoll_ctl_prep(req, sqe);
5975 case IORING_OP_SPLICE:
5976 return io_splice_prep(req, sqe);
5977 case IORING_OP_PROVIDE_BUFFERS:
5978 return io_provide_buffers_prep(req, sqe);
5979 case IORING_OP_REMOVE_BUFFERS:
5980 return io_remove_buffers_prep(req, sqe);
5982 return io_tee_prep(req, sqe);
5983 case IORING_OP_SHUTDOWN:
5984 return io_shutdown_prep(req, sqe);
5985 case IORING_OP_RENAMEAT:
5986 return io_renameat_prep(req, sqe);
5987 case IORING_OP_UNLINKAT:
5988 return io_unlinkat_prep(req, sqe);
5991 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5996 static int io_req_prep_async(struct io_kiocb *req)
5998 if (!io_op_defs[req->opcode].needs_async_setup)
6000 if (WARN_ON_ONCE(req->async_data))
6002 if (io_alloc_async_data(req))
6005 switch (req->opcode) {
6006 case IORING_OP_READV:
6007 return io_rw_prep_async(req, READ);
6008 case IORING_OP_WRITEV:
6009 return io_rw_prep_async(req, WRITE);
6010 case IORING_OP_SENDMSG:
6011 return io_sendmsg_prep_async(req);
6012 case IORING_OP_RECVMSG:
6013 return io_recvmsg_prep_async(req);
6014 case IORING_OP_CONNECT:
6015 return io_connect_prep_async(req);
6017 printk_once(KERN_WARNING "io_uring: prep_async() bad opcode %d\n",
6022 static u32 io_get_sequence(struct io_kiocb *req)
6024 u32 seq = req->ctx->cached_sq_head;
6026 /* need original cached_sq_head, but it was increased for each req */
6027 io_for_each_link(req, req)
6032 static bool io_drain_req(struct io_kiocb *req)
6034 struct io_kiocb *pos;
6035 struct io_ring_ctx *ctx = req->ctx;
6036 struct io_defer_entry *de;
6041 * If we need to drain a request in the middle of a link, drain the
6042 * head request and the next request/link after the current link.
6043 * Considering sequential execution of links, IOSQE_IO_DRAIN will be
6044 * maintained for every request of our link.
6046 if (ctx->drain_next) {
6047 req->flags |= REQ_F_IO_DRAIN;
6048 ctx->drain_next = false;
6050 /* not interested in head, start from the first linked */
6051 io_for_each_link(pos, req->link) {
6052 if (pos->flags & REQ_F_IO_DRAIN) {
6053 ctx->drain_next = true;
6054 req->flags |= REQ_F_IO_DRAIN;
6059 /* Still need defer if there is pending req in defer list. */
6060 if (likely(list_empty_careful(&ctx->defer_list) &&
6061 !(req->flags & REQ_F_IO_DRAIN))) {
6062 ctx->drain_active = false;
6066 seq = io_get_sequence(req);
6067 /* Still a chance to pass the sequence check */
6068 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
6071 ret = io_req_prep_async(req);
6074 io_prep_async_link(req);
6075 de = kmalloc(sizeof(*de), GFP_KERNEL);
6079 io_req_complete_failed(req, ret);
6083 spin_lock_irq(&ctx->completion_lock);
6084 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
6085 spin_unlock_irq(&ctx->completion_lock);
6087 io_queue_async_work(req);
6091 trace_io_uring_defer(ctx, req, req->user_data);
6094 list_add_tail(&de->list, &ctx->defer_list);
6095 spin_unlock_irq(&ctx->completion_lock);
6099 static void io_clean_op(struct io_kiocb *req)
6101 if (req->flags & REQ_F_BUFFER_SELECTED) {
6102 switch (req->opcode) {
6103 case IORING_OP_READV:
6104 case IORING_OP_READ_FIXED:
6105 case IORING_OP_READ:
6106 kfree((void *)(unsigned long)req->rw.addr);
6108 case IORING_OP_RECVMSG:
6109 case IORING_OP_RECV:
6110 kfree(req->sr_msg.kbuf);
6115 if (req->flags & REQ_F_NEED_CLEANUP) {
6116 switch (req->opcode) {
6117 case IORING_OP_READV:
6118 case IORING_OP_READ_FIXED:
6119 case IORING_OP_READ:
6120 case IORING_OP_WRITEV:
6121 case IORING_OP_WRITE_FIXED:
6122 case IORING_OP_WRITE: {
6123 struct io_async_rw *io = req->async_data;
6125 kfree(io->free_iovec);
6128 case IORING_OP_RECVMSG:
6129 case IORING_OP_SENDMSG: {
6130 struct io_async_msghdr *io = req->async_data;
6132 kfree(io->free_iov);
6135 case IORING_OP_SPLICE:
6137 if (!(req->splice.flags & SPLICE_F_FD_IN_FIXED))
6138 io_put_file(req->splice.file_in);
6140 case IORING_OP_OPENAT:
6141 case IORING_OP_OPENAT2:
6142 if (req->open.filename)
6143 putname(req->open.filename);
6145 case IORING_OP_RENAMEAT:
6146 putname(req->rename.oldpath);
6147 putname(req->rename.newpath);
6149 case IORING_OP_UNLINKAT:
6150 putname(req->unlink.filename);
6154 if ((req->flags & REQ_F_POLLED) && req->apoll) {
6155 kfree(req->apoll->double_poll);
6159 if (req->flags & REQ_F_INFLIGHT) {
6160 struct io_uring_task *tctx = req->task->io_uring;
6162 atomic_dec(&tctx->inflight_tracked);
6164 if (req->flags & REQ_F_CREDS)
6165 put_cred(req->creds);
6167 req->flags &= ~IO_REQ_CLEAN_FLAGS;
6170 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
6172 struct io_ring_ctx *ctx = req->ctx;
6173 const struct cred *creds = NULL;
6176 if ((req->flags & REQ_F_CREDS) && req->creds != current_cred())
6177 creds = override_creds(req->creds);
6179 switch (req->opcode) {
6181 ret = io_nop(req, issue_flags);
6183 case IORING_OP_READV:
6184 case IORING_OP_READ_FIXED:
6185 case IORING_OP_READ:
6186 ret = io_read(req, issue_flags);
6188 case IORING_OP_WRITEV:
6189 case IORING_OP_WRITE_FIXED:
6190 case IORING_OP_WRITE:
6191 ret = io_write(req, issue_flags);
6193 case IORING_OP_FSYNC:
6194 ret = io_fsync(req, issue_flags);
6196 case IORING_OP_POLL_ADD:
6197 ret = io_poll_add(req, issue_flags);
6199 case IORING_OP_POLL_REMOVE:
6200 ret = io_poll_update(req, issue_flags);
6202 case IORING_OP_SYNC_FILE_RANGE:
6203 ret = io_sync_file_range(req, issue_flags);
6205 case IORING_OP_SENDMSG:
6206 ret = io_sendmsg(req, issue_flags);
6208 case IORING_OP_SEND:
6209 ret = io_send(req, issue_flags);
6211 case IORING_OP_RECVMSG:
6212 ret = io_recvmsg(req, issue_flags);
6214 case IORING_OP_RECV:
6215 ret = io_recv(req, issue_flags);
6217 case IORING_OP_TIMEOUT:
6218 ret = io_timeout(req, issue_flags);
6220 case IORING_OP_TIMEOUT_REMOVE:
6221 ret = io_timeout_remove(req, issue_flags);
6223 case IORING_OP_ACCEPT:
6224 ret = io_accept(req, issue_flags);
6226 case IORING_OP_CONNECT:
6227 ret = io_connect(req, issue_flags);
6229 case IORING_OP_ASYNC_CANCEL:
6230 ret = io_async_cancel(req, issue_flags);
6232 case IORING_OP_FALLOCATE:
6233 ret = io_fallocate(req, issue_flags);
6235 case IORING_OP_OPENAT:
6236 ret = io_openat(req, issue_flags);
6238 case IORING_OP_CLOSE:
6239 ret = io_close(req, issue_flags);
6241 case IORING_OP_FILES_UPDATE:
6242 ret = io_files_update(req, issue_flags);
6244 case IORING_OP_STATX:
6245 ret = io_statx(req, issue_flags);
6247 case IORING_OP_FADVISE:
6248 ret = io_fadvise(req, issue_flags);
6250 case IORING_OP_MADVISE:
6251 ret = io_madvise(req, issue_flags);
6253 case IORING_OP_OPENAT2:
6254 ret = io_openat2(req, issue_flags);
6256 case IORING_OP_EPOLL_CTL:
6257 ret = io_epoll_ctl(req, issue_flags);
6259 case IORING_OP_SPLICE:
6260 ret = io_splice(req, issue_flags);
6262 case IORING_OP_PROVIDE_BUFFERS:
6263 ret = io_provide_buffers(req, issue_flags);
6265 case IORING_OP_REMOVE_BUFFERS:
6266 ret = io_remove_buffers(req, issue_flags);
6269 ret = io_tee(req, issue_flags);
6271 case IORING_OP_SHUTDOWN:
6272 ret = io_shutdown(req, issue_flags);
6274 case IORING_OP_RENAMEAT:
6275 ret = io_renameat(req, issue_flags);
6277 case IORING_OP_UNLINKAT:
6278 ret = io_unlinkat(req, issue_flags);
6286 revert_creds(creds);
6289 /* If the op doesn't have a file, we're not polling for it */
6290 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file)
6291 io_iopoll_req_issued(req);
6296 static struct io_wq_work *io_wq_free_work(struct io_wq_work *work)
6298 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6300 req = io_put_req_find_next(req);
6301 return req ? &req->work : NULL;
6304 static void io_wq_submit_work(struct io_wq_work *work)
6306 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6307 struct io_kiocb *timeout;
6310 timeout = io_prep_linked_timeout(req);
6312 io_queue_linked_timeout(timeout);
6314 if (work->flags & IO_WQ_WORK_CANCEL)
6319 ret = io_issue_sqe(req, 0);
6321 * We can get EAGAIN for polled IO even though we're
6322 * forcing a sync submission from here, since we can't
6323 * wait for request slots on the block side.
6331 /* avoid locking problems by failing it from a clean context */
6333 /* io-wq is going to take one down */
6335 io_req_task_queue_fail(req, ret);
6339 static inline struct io_fixed_file *io_fixed_file_slot(struct io_file_table *table,
6342 return &table->files[i];
6345 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6348 struct io_fixed_file *slot = io_fixed_file_slot(&ctx->file_table, index);
6350 return (struct file *) (slot->file_ptr & FFS_MASK);
6353 static void io_fixed_file_set(struct io_fixed_file *file_slot, struct file *file)
6355 unsigned long file_ptr = (unsigned long) file;
6357 if (__io_file_supports_nowait(file, READ))
6358 file_ptr |= FFS_ASYNC_READ;
6359 if (__io_file_supports_nowait(file, WRITE))
6360 file_ptr |= FFS_ASYNC_WRITE;
6361 if (S_ISREG(file_inode(file)->i_mode))
6362 file_ptr |= FFS_ISREG;
6363 file_slot->file_ptr = file_ptr;
6366 static inline struct file *io_file_get_fixed(struct io_ring_ctx *ctx,
6367 struct io_kiocb *req, int fd)
6370 unsigned long file_ptr;
6372 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6374 fd = array_index_nospec(fd, ctx->nr_user_files);
6375 file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr;
6376 file = (struct file *) (file_ptr & FFS_MASK);
6377 file_ptr &= ~FFS_MASK;
6378 /* mask in overlapping REQ_F and FFS bits */
6379 req->flags |= (file_ptr << REQ_F_NOWAIT_READ_BIT);
6380 io_req_set_rsrc_node(req);
6384 static struct file *io_file_get_normal(struct io_ring_ctx *ctx,
6385 struct io_submit_state *state,
6386 struct io_kiocb *req, int fd)
6388 struct file *file = __io_file_get(state, fd);
6390 trace_io_uring_file_get(ctx, fd);
6392 /* we don't allow fixed io_uring files */
6393 if (file && unlikely(file->f_op == &io_uring_fops))
6394 io_req_track_inflight(req);
6398 static inline struct file *io_file_get(struct io_ring_ctx *ctx,
6399 struct io_submit_state *state,
6400 struct io_kiocb *req, int fd, bool fixed)
6403 return io_file_get_fixed(ctx, req, fd);
6405 return io_file_get_normal(ctx, state, req, fd);
6408 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6410 struct io_timeout_data *data = container_of(timer,
6411 struct io_timeout_data, timer);
6412 struct io_kiocb *prev, *req = data->req;
6413 struct io_ring_ctx *ctx = req->ctx;
6414 unsigned long flags;
6416 spin_lock_irqsave(&ctx->completion_lock, flags);
6417 prev = req->timeout.head;
6418 req->timeout.head = NULL;
6421 * We don't expect the list to be empty, that will only happen if we
6422 * race with the completion of the linked work.
6425 io_remove_next_linked(prev);
6426 if (!req_ref_inc_not_zero(prev))
6429 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6432 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6433 io_put_req_deferred(prev, 1);
6434 io_put_req_deferred(req, 1);
6436 io_req_complete_post(req, -ETIME, 0);
6438 return HRTIMER_NORESTART;
6441 static void io_queue_linked_timeout(struct io_kiocb *req)
6443 struct io_ring_ctx *ctx = req->ctx;
6445 spin_lock_irq(&ctx->completion_lock);
6447 * If the back reference is NULL, then our linked request finished
6448 * before we got a chance to setup the timer
6450 if (req->timeout.head) {
6451 struct io_timeout_data *data = req->async_data;
6453 data->timer.function = io_link_timeout_fn;
6454 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6457 spin_unlock_irq(&ctx->completion_lock);
6458 /* drop submission reference */
6462 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6464 struct io_kiocb *nxt = req->link;
6466 if (!nxt || (req->flags & REQ_F_LINK_TIMEOUT) ||
6467 nxt->opcode != IORING_OP_LINK_TIMEOUT)
6470 nxt->timeout.head = req;
6471 nxt->flags |= REQ_F_LTIMEOUT_ACTIVE;
6472 req->flags |= REQ_F_LINK_TIMEOUT;
6476 static void __io_queue_sqe(struct io_kiocb *req)
6477 __must_hold(&req->ctx->uring_lock)
6479 struct io_kiocb *linked_timeout = io_prep_linked_timeout(req);
6483 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
6486 * We async punt it if the file wasn't marked NOWAIT, or if the file
6487 * doesn't support non-blocking read/write attempts
6490 /* drop submission reference */
6491 if (req->flags & REQ_F_COMPLETE_INLINE) {
6492 struct io_ring_ctx *ctx = req->ctx;
6493 struct io_comp_state *cs = &ctx->submit_state.comp;
6495 cs->reqs[cs->nr++] = req;
6496 if (cs->nr == ARRAY_SIZE(cs->reqs))
6497 io_submit_flush_completions(ctx);
6501 } else if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6502 switch (io_arm_poll_handler(req)) {
6503 case IO_APOLL_READY:
6505 case IO_APOLL_ABORTED:
6507 * Queued up for async execution, worker will release
6508 * submit reference when the iocb is actually submitted.
6510 io_queue_async_work(req);
6514 io_req_complete_failed(req, ret);
6517 io_queue_linked_timeout(linked_timeout);
6520 static inline void io_queue_sqe(struct io_kiocb *req)
6521 __must_hold(&req->ctx->uring_lock)
6523 if (unlikely(req->ctx->drain_active) && io_drain_req(req))
6526 if (likely(!(req->flags & REQ_F_FORCE_ASYNC))) {
6527 __io_queue_sqe(req);
6529 int ret = io_req_prep_async(req);
6532 io_req_complete_failed(req, ret);
6534 io_queue_async_work(req);
6539 * Check SQE restrictions (opcode and flags).
6541 * Returns 'true' if SQE is allowed, 'false' otherwise.
6543 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6544 struct io_kiocb *req,
6545 unsigned int sqe_flags)
6547 if (likely(!ctx->restricted))
6550 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6553 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6554 ctx->restrictions.sqe_flags_required)
6557 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6558 ctx->restrictions.sqe_flags_required))
6564 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6565 const struct io_uring_sqe *sqe)
6566 __must_hold(&ctx->uring_lock)
6568 struct io_submit_state *state;
6569 unsigned int sqe_flags;
6570 int personality, ret = 0;
6572 /* req is partially pre-initialised, see io_preinit_req() */
6573 req->opcode = READ_ONCE(sqe->opcode);
6574 /* same numerical values with corresponding REQ_F_*, safe to copy */
6575 req->flags = sqe_flags = READ_ONCE(sqe->flags);
6576 req->user_data = READ_ONCE(sqe->user_data);
6578 req->fixed_rsrc_refs = NULL;
6579 /* one is dropped after submission, the other at completion */
6580 atomic_set(&req->refs, 2);
6581 req->task = current;
6583 /* enforce forwards compatibility on users */
6584 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6586 if (unlikely(req->opcode >= IORING_OP_LAST))
6588 if (!io_check_restriction(ctx, req, sqe_flags))
6591 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6592 !io_op_defs[req->opcode].buffer_select)
6594 if (unlikely(sqe_flags & IOSQE_IO_DRAIN))
6595 ctx->drain_active = true;
6597 personality = READ_ONCE(sqe->personality);
6599 req->creds = xa_load(&ctx->personalities, personality);
6602 get_cred(req->creds);
6603 req->flags |= REQ_F_CREDS;
6605 state = &ctx->submit_state;
6608 * Plug now if we have more than 1 IO left after this, and the target
6609 * is potentially a read/write to block based storage.
6611 if (!state->plug_started && state->ios_left > 1 &&
6612 io_op_defs[req->opcode].plug) {
6613 blk_start_plug(&state->plug);
6614 state->plug_started = true;
6617 if (io_op_defs[req->opcode].needs_file) {
6618 req->file = io_file_get(ctx, state, req, READ_ONCE(sqe->fd),
6619 (sqe_flags & IOSQE_FIXED_FILE));
6620 if (unlikely(!req->file))
6628 static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
6629 const struct io_uring_sqe *sqe)
6630 __must_hold(&ctx->uring_lock)
6632 struct io_submit_link *link = &ctx->submit_state.link;
6635 ret = io_init_req(ctx, req, sqe);
6636 if (unlikely(ret)) {
6639 /* fail even hard links since we don't submit */
6640 req_set_fail(link->head);
6641 io_req_complete_failed(link->head, -ECANCELED);
6644 io_req_complete_failed(req, ret);
6648 ret = io_req_prep(req, sqe);
6652 /* don't need @sqe from now on */
6653 trace_io_uring_submit_sqe(ctx, req, req->opcode, req->user_data,
6655 ctx->flags & IORING_SETUP_SQPOLL);
6658 * If we already have a head request, queue this one for async
6659 * submittal once the head completes. If we don't have a head but
6660 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6661 * submitted sync once the chain is complete. If none of those
6662 * conditions are true (normal request), then just queue it.
6665 struct io_kiocb *head = link->head;
6667 ret = io_req_prep_async(req);
6670 trace_io_uring_link(ctx, req, head);
6671 link->last->link = req;
6674 /* last request of a link, enqueue the link */
6675 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6680 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6692 * Batched submission is done, ensure local IO is flushed out.
6694 static void io_submit_state_end(struct io_submit_state *state,
6695 struct io_ring_ctx *ctx)
6697 if (state->link.head)
6698 io_queue_sqe(state->link.head);
6700 io_submit_flush_completions(ctx);
6701 if (state->plug_started)
6702 blk_finish_plug(&state->plug);
6703 io_state_file_put(state);
6707 * Start submission side cache.
6709 static void io_submit_state_start(struct io_submit_state *state,
6710 unsigned int max_ios)
6712 state->plug_started = false;
6713 state->ios_left = max_ios;
6714 /* set only head, no need to init link_last in advance */
6715 state->link.head = NULL;
6718 static void io_commit_sqring(struct io_ring_ctx *ctx)
6720 struct io_rings *rings = ctx->rings;
6723 * Ensure any loads from the SQEs are done at this point,
6724 * since once we write the new head, the application could
6725 * write new data to them.
6727 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6731 * Fetch an sqe, if one is available. Note this returns a pointer to memory
6732 * that is mapped by userspace. This means that care needs to be taken to
6733 * ensure that reads are stable, as we cannot rely on userspace always
6734 * being a good citizen. If members of the sqe are validated and then later
6735 * used, it's important that those reads are done through READ_ONCE() to
6736 * prevent a re-load down the line.
6738 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6740 unsigned head, mask = ctx->sq_entries - 1;
6741 unsigned sq_idx = ctx->cached_sq_head++ & mask;
6744 * The cached sq head (or cq tail) serves two purposes:
6746 * 1) allows us to batch the cost of updating the user visible
6748 * 2) allows the kernel side to track the head on its own, even
6749 * though the application is the one updating it.
6751 head = READ_ONCE(ctx->sq_array[sq_idx]);
6752 if (likely(head < ctx->sq_entries))
6753 return &ctx->sq_sqes[head];
6755 /* drop invalid entries */
6757 WRITE_ONCE(ctx->rings->sq_dropped,
6758 READ_ONCE(ctx->rings->sq_dropped) + 1);
6762 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6763 __must_hold(&ctx->uring_lock)
6765 struct io_uring_task *tctx;
6768 /* make sure SQ entry isn't read before tail */
6769 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6770 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6773 tctx = current->io_uring;
6774 tctx->cached_refs -= nr;
6775 if (unlikely(tctx->cached_refs < 0)) {
6776 unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR;
6778 percpu_counter_add(&tctx->inflight, refill);
6779 refcount_add(refill, ¤t->usage);
6780 tctx->cached_refs += refill;
6782 io_submit_state_start(&ctx->submit_state, nr);
6784 while (submitted < nr) {
6785 const struct io_uring_sqe *sqe;
6786 struct io_kiocb *req;
6788 req = io_alloc_req(ctx);
6789 if (unlikely(!req)) {
6791 submitted = -EAGAIN;
6794 sqe = io_get_sqe(ctx);
6795 if (unlikely(!sqe)) {
6796 kmem_cache_free(req_cachep, req);
6799 /* will complete beyond this point, count as submitted */
6801 if (io_submit_sqe(ctx, req, sqe))
6805 if (unlikely(submitted != nr)) {
6806 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6807 int unused = nr - ref_used;
6809 current->io_uring->cached_refs += unused;
6810 percpu_ref_put_many(&ctx->refs, unused);
6813 io_submit_state_end(&ctx->submit_state, ctx);
6814 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6815 io_commit_sqring(ctx);
6820 static inline bool io_sqd_events_pending(struct io_sq_data *sqd)
6822 return READ_ONCE(sqd->state);
6825 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6827 /* Tell userspace we may need a wakeup call */
6828 spin_lock_irq(&ctx->completion_lock);
6829 WRITE_ONCE(ctx->rings->sq_flags,
6830 ctx->rings->sq_flags | IORING_SQ_NEED_WAKEUP);
6831 spin_unlock_irq(&ctx->completion_lock);
6834 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6836 spin_lock_irq(&ctx->completion_lock);
6837 WRITE_ONCE(ctx->rings->sq_flags,
6838 ctx->rings->sq_flags & ~IORING_SQ_NEED_WAKEUP);
6839 spin_unlock_irq(&ctx->completion_lock);
6842 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
6844 unsigned int to_submit;
6847 to_submit = io_sqring_entries(ctx);
6848 /* if we're handling multiple rings, cap submit size for fairness */
6849 if (cap_entries && to_submit > IORING_SQPOLL_CAP_ENTRIES_VALUE)
6850 to_submit = IORING_SQPOLL_CAP_ENTRIES_VALUE;
6852 if (!list_empty(&ctx->iopoll_list) || to_submit) {
6853 unsigned nr_events = 0;
6854 const struct cred *creds = NULL;
6856 if (ctx->sq_creds != current_cred())
6857 creds = override_creds(ctx->sq_creds);
6859 mutex_lock(&ctx->uring_lock);
6860 if (!list_empty(&ctx->iopoll_list))
6861 io_do_iopoll(ctx, &nr_events, 0, true);
6864 * Don't submit if refs are dying, good for io_uring_register(),
6865 * but also it is relied upon by io_ring_exit_work()
6867 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)) &&
6868 !(ctx->flags & IORING_SETUP_R_DISABLED))
6869 ret = io_submit_sqes(ctx, to_submit);
6870 mutex_unlock(&ctx->uring_lock);
6872 if (to_submit && wq_has_sleeper(&ctx->sqo_sq_wait))
6873 wake_up(&ctx->sqo_sq_wait);
6875 revert_creds(creds);
6881 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
6883 struct io_ring_ctx *ctx;
6884 unsigned sq_thread_idle = 0;
6886 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6887 sq_thread_idle = max(sq_thread_idle, ctx->sq_thread_idle);
6888 sqd->sq_thread_idle = sq_thread_idle;
6891 static bool io_sqd_handle_event(struct io_sq_data *sqd)
6893 bool did_sig = false;
6894 struct ksignal ksig;
6896 if (test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state) ||
6897 signal_pending(current)) {
6898 mutex_unlock(&sqd->lock);
6899 if (signal_pending(current))
6900 did_sig = get_signal(&ksig);
6902 mutex_lock(&sqd->lock);
6904 return did_sig || test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
6907 static int io_sq_thread(void *data)
6909 struct io_sq_data *sqd = data;
6910 struct io_ring_ctx *ctx;
6911 unsigned long timeout = 0;
6912 char buf[TASK_COMM_LEN];
6915 snprintf(buf, sizeof(buf), "iou-sqp-%d", sqd->task_pid);
6916 set_task_comm(current, buf);
6918 if (sqd->sq_cpu != -1)
6919 set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
6921 set_cpus_allowed_ptr(current, cpu_online_mask);
6922 current->flags |= PF_NO_SETAFFINITY;
6924 mutex_lock(&sqd->lock);
6926 bool cap_entries, sqt_spin = false;
6928 if (io_sqd_events_pending(sqd) || signal_pending(current)) {
6929 if (io_sqd_handle_event(sqd))
6931 timeout = jiffies + sqd->sq_thread_idle;
6934 cap_entries = !list_is_singular(&sqd->ctx_list);
6935 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6936 int ret = __io_sq_thread(ctx, cap_entries);
6938 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
6941 if (io_run_task_work())
6944 if (sqt_spin || !time_after(jiffies, timeout)) {
6947 timeout = jiffies + sqd->sq_thread_idle;
6951 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
6952 if (!io_sqd_events_pending(sqd) && !current->task_works) {
6953 bool needs_sched = true;
6955 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6956 io_ring_set_wakeup_flag(ctx);
6958 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6959 !list_empty_careful(&ctx->iopoll_list)) {
6960 needs_sched = false;
6963 if (io_sqring_entries(ctx)) {
6964 needs_sched = false;
6970 mutex_unlock(&sqd->lock);
6972 mutex_lock(&sqd->lock);
6974 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6975 io_ring_clear_wakeup_flag(ctx);
6978 finish_wait(&sqd->wait, &wait);
6979 timeout = jiffies + sqd->sq_thread_idle;
6982 io_uring_cancel_generic(true, sqd);
6984 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6985 io_ring_set_wakeup_flag(ctx);
6987 mutex_unlock(&sqd->lock);
6989 complete(&sqd->exited);
6993 struct io_wait_queue {
6994 struct wait_queue_entry wq;
6995 struct io_ring_ctx *ctx;
6997 unsigned nr_timeouts;
7000 static inline bool io_should_wake(struct io_wait_queue *iowq)
7002 struct io_ring_ctx *ctx = iowq->ctx;
7003 int dist = ctx->cached_cq_tail - (int) iowq->cq_tail;
7006 * Wake up if we have enough events, or if a timeout occurred since we
7007 * started waiting. For timeouts, we always want to return to userspace,
7008 * regardless of event count.
7010 return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
7013 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
7014 int wake_flags, void *key)
7016 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
7020 * Cannot safely flush overflowed CQEs from here, ensure we wake up
7021 * the task, and the next invocation will do it.
7023 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->check_cq_overflow))
7024 return autoremove_wake_function(curr, mode, wake_flags, key);
7028 static int io_run_task_work_sig(void)
7030 if (io_run_task_work())
7032 if (!signal_pending(current))
7034 if (test_thread_flag(TIF_NOTIFY_SIGNAL))
7035 return -ERESTARTSYS;
7039 /* when returns >0, the caller should retry */
7040 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
7041 struct io_wait_queue *iowq,
7042 signed long *timeout)
7046 /* make sure we run task_work before checking for signals */
7047 ret = io_run_task_work_sig();
7048 if (ret || io_should_wake(iowq))
7050 /* let the caller flush overflows, retry */
7051 if (test_bit(0, &ctx->check_cq_overflow))
7054 *timeout = schedule_timeout(*timeout);
7055 return !*timeout ? -ETIME : 1;
7059 * Wait until events become available, if we don't already have some. The
7060 * application must reap them itself, as they reside on the shared cq ring.
7062 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
7063 const sigset_t __user *sig, size_t sigsz,
7064 struct __kernel_timespec __user *uts)
7066 struct io_wait_queue iowq;
7067 struct io_rings *rings = ctx->rings;
7068 signed long timeout = MAX_SCHEDULE_TIMEOUT;
7072 io_cqring_overflow_flush(ctx, false);
7073 if (io_cqring_events(ctx) >= min_events)
7075 if (!io_run_task_work())
7080 #ifdef CONFIG_COMPAT
7081 if (in_compat_syscall())
7082 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
7086 ret = set_user_sigmask(sig, sigsz);
7093 struct timespec64 ts;
7095 if (get_timespec64(&ts, uts))
7097 timeout = timespec64_to_jiffies(&ts);
7100 init_waitqueue_func_entry(&iowq.wq, io_wake_function);
7101 iowq.wq.private = current;
7102 INIT_LIST_HEAD(&iowq.wq.entry);
7104 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
7105 iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events;
7107 trace_io_uring_cqring_wait(ctx, min_events);
7109 /* if we can't even flush overflow, don't wait for more */
7110 if (!io_cqring_overflow_flush(ctx, false)) {
7114 prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq,
7115 TASK_INTERRUPTIBLE);
7116 ret = io_cqring_wait_schedule(ctx, &iowq, &timeout);
7117 finish_wait(&ctx->cq_wait, &iowq.wq);
7121 restore_saved_sigmask_unless(ret == -EINTR);
7123 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
7126 static void io_free_page_table(void **table, size_t size)
7128 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7130 for (i = 0; i < nr_tables; i++)
7135 static void **io_alloc_page_table(size_t size)
7137 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7138 size_t init_size = size;
7141 table = kcalloc(nr_tables, sizeof(*table), GFP_KERNEL);
7145 for (i = 0; i < nr_tables; i++) {
7146 unsigned int this_size = min_t(size_t, size, PAGE_SIZE);
7148 table[i] = kzalloc(this_size, GFP_KERNEL);
7150 io_free_page_table(table, init_size);
7158 static void io_rsrc_node_destroy(struct io_rsrc_node *ref_node)
7160 percpu_ref_exit(&ref_node->refs);
7164 static void io_rsrc_node_switch(struct io_ring_ctx *ctx,
7165 struct io_rsrc_data *data_to_kill)
7167 WARN_ON_ONCE(!ctx->rsrc_backup_node);
7168 WARN_ON_ONCE(data_to_kill && !ctx->rsrc_node);
7171 struct io_rsrc_node *rsrc_node = ctx->rsrc_node;
7173 rsrc_node->rsrc_data = data_to_kill;
7174 spin_lock_irq(&ctx->rsrc_ref_lock);
7175 list_add_tail(&rsrc_node->node, &ctx->rsrc_ref_list);
7176 spin_unlock_irq(&ctx->rsrc_ref_lock);
7178 atomic_inc(&data_to_kill->refs);
7179 percpu_ref_kill(&rsrc_node->refs);
7180 ctx->rsrc_node = NULL;
7183 if (!ctx->rsrc_node) {
7184 ctx->rsrc_node = ctx->rsrc_backup_node;
7185 ctx->rsrc_backup_node = NULL;
7189 static int io_rsrc_node_switch_start(struct io_ring_ctx *ctx)
7191 if (ctx->rsrc_backup_node)
7193 ctx->rsrc_backup_node = io_rsrc_node_alloc(ctx);
7194 return ctx->rsrc_backup_node ? 0 : -ENOMEM;
7197 static int io_rsrc_ref_quiesce(struct io_rsrc_data *data, struct io_ring_ctx *ctx)
7201 /* As we may drop ->uring_lock, other task may have started quiesce */
7205 data->quiesce = true;
7207 ret = io_rsrc_node_switch_start(ctx);
7210 io_rsrc_node_switch(ctx, data);
7212 /* kill initial ref, already quiesced if zero */
7213 if (atomic_dec_and_test(&data->refs))
7215 mutex_unlock(&ctx->uring_lock);
7216 flush_delayed_work(&ctx->rsrc_put_work);
7217 ret = wait_for_completion_interruptible(&data->done);
7219 mutex_lock(&ctx->uring_lock);
7223 atomic_inc(&data->refs);
7224 /* wait for all works potentially completing data->done */
7225 flush_delayed_work(&ctx->rsrc_put_work);
7226 reinit_completion(&data->done);
7228 ret = io_run_task_work_sig();
7229 mutex_lock(&ctx->uring_lock);
7231 data->quiesce = false;
7236 static u64 *io_get_tag_slot(struct io_rsrc_data *data, unsigned int idx)
7238 unsigned int off = idx & IO_RSRC_TAG_TABLE_MASK;
7239 unsigned int table_idx = idx >> IO_RSRC_TAG_TABLE_SHIFT;
7241 return &data->tags[table_idx][off];
7244 static void io_rsrc_data_free(struct io_rsrc_data *data)
7246 size_t size = data->nr * sizeof(data->tags[0][0]);
7249 io_free_page_table((void **)data->tags, size);
7253 static int io_rsrc_data_alloc(struct io_ring_ctx *ctx, rsrc_put_fn *do_put,
7254 u64 __user *utags, unsigned nr,
7255 struct io_rsrc_data **pdata)
7257 struct io_rsrc_data *data;
7261 data = kzalloc(sizeof(*data), GFP_KERNEL);
7264 data->tags = (u64 **)io_alloc_page_table(nr * sizeof(data->tags[0][0]));
7272 data->do_put = do_put;
7275 for (i = 0; i < nr; i++) {
7276 u64 *tag_slot = io_get_tag_slot(data, i);
7278 if (copy_from_user(tag_slot, &utags[i],
7284 atomic_set(&data->refs, 1);
7285 init_completion(&data->done);
7289 io_rsrc_data_free(data);
7293 static bool io_alloc_file_tables(struct io_file_table *table, unsigned nr_files)
7295 table->files = kvcalloc(nr_files, sizeof(table->files[0]), GFP_KERNEL);
7296 return !!table->files;
7299 static void io_free_file_tables(struct io_file_table *table)
7301 kvfree(table->files);
7302 table->files = NULL;
7305 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
7307 #if defined(CONFIG_UNIX)
7308 if (ctx->ring_sock) {
7309 struct sock *sock = ctx->ring_sock->sk;
7310 struct sk_buff *skb;
7312 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
7318 for (i = 0; i < ctx->nr_user_files; i++) {
7321 file = io_file_from_index(ctx, i);
7326 io_free_file_tables(&ctx->file_table);
7327 io_rsrc_data_free(ctx->file_data);
7328 ctx->file_data = NULL;
7329 ctx->nr_user_files = 0;
7332 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7336 if (!ctx->file_data)
7338 ret = io_rsrc_ref_quiesce(ctx->file_data, ctx);
7340 __io_sqe_files_unregister(ctx);
7344 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7345 __releases(&sqd->lock)
7347 WARN_ON_ONCE(sqd->thread == current);
7350 * Do the dance but not conditional clear_bit() because it'd race with
7351 * other threads incrementing park_pending and setting the bit.
7353 clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7354 if (atomic_dec_return(&sqd->park_pending))
7355 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7356 mutex_unlock(&sqd->lock);
7359 static void io_sq_thread_park(struct io_sq_data *sqd)
7360 __acquires(&sqd->lock)
7362 WARN_ON_ONCE(sqd->thread == current);
7364 atomic_inc(&sqd->park_pending);
7365 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7366 mutex_lock(&sqd->lock);
7368 wake_up_process(sqd->thread);
7371 static void io_sq_thread_stop(struct io_sq_data *sqd)
7373 WARN_ON_ONCE(sqd->thread == current);
7374 WARN_ON_ONCE(test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state));
7376 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7377 mutex_lock(&sqd->lock);
7379 wake_up_process(sqd->thread);
7380 mutex_unlock(&sqd->lock);
7381 wait_for_completion(&sqd->exited);
7384 static void io_put_sq_data(struct io_sq_data *sqd)
7386 if (refcount_dec_and_test(&sqd->refs)) {
7387 WARN_ON_ONCE(atomic_read(&sqd->park_pending));
7389 io_sq_thread_stop(sqd);
7394 static void io_sq_thread_finish(struct io_ring_ctx *ctx)
7396 struct io_sq_data *sqd = ctx->sq_data;
7399 io_sq_thread_park(sqd);
7400 list_del_init(&ctx->sqd_list);
7401 io_sqd_update_thread_idle(sqd);
7402 io_sq_thread_unpark(sqd);
7404 io_put_sq_data(sqd);
7405 ctx->sq_data = NULL;
7409 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7411 struct io_ring_ctx *ctx_attach;
7412 struct io_sq_data *sqd;
7415 f = fdget(p->wq_fd);
7417 return ERR_PTR(-ENXIO);
7418 if (f.file->f_op != &io_uring_fops) {
7420 return ERR_PTR(-EINVAL);
7423 ctx_attach = f.file->private_data;
7424 sqd = ctx_attach->sq_data;
7427 return ERR_PTR(-EINVAL);
7429 if (sqd->task_tgid != current->tgid) {
7431 return ERR_PTR(-EPERM);
7434 refcount_inc(&sqd->refs);
7439 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p,
7442 struct io_sq_data *sqd;
7445 if (p->flags & IORING_SETUP_ATTACH_WQ) {
7446 sqd = io_attach_sq_data(p);
7451 /* fall through for EPERM case, setup new sqd/task */
7452 if (PTR_ERR(sqd) != -EPERM)
7456 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7458 return ERR_PTR(-ENOMEM);
7460 atomic_set(&sqd->park_pending, 0);
7461 refcount_set(&sqd->refs, 1);
7462 INIT_LIST_HEAD(&sqd->ctx_list);
7463 mutex_init(&sqd->lock);
7464 init_waitqueue_head(&sqd->wait);
7465 init_completion(&sqd->exited);
7469 #if defined(CONFIG_UNIX)
7471 * Ensure the UNIX gc is aware of our file set, so we are certain that
7472 * the io_uring can be safely unregistered on process exit, even if we have
7473 * loops in the file referencing.
7475 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7477 struct sock *sk = ctx->ring_sock->sk;
7478 struct scm_fp_list *fpl;
7479 struct sk_buff *skb;
7482 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7486 skb = alloc_skb(0, GFP_KERNEL);
7495 fpl->user = get_uid(current_user());
7496 for (i = 0; i < nr; i++) {
7497 struct file *file = io_file_from_index(ctx, i + offset);
7501 fpl->fp[nr_files] = get_file(file);
7502 unix_inflight(fpl->user, fpl->fp[nr_files]);
7507 fpl->max = SCM_MAX_FD;
7508 fpl->count = nr_files;
7509 UNIXCB(skb).fp = fpl;
7510 skb->destructor = unix_destruct_scm;
7511 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7512 skb_queue_head(&sk->sk_receive_queue, skb);
7514 for (i = 0; i < nr_files; i++)
7525 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7526 * causes regular reference counting to break down. We rely on the UNIX
7527 * garbage collection to take care of this problem for us.
7529 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7531 unsigned left, total;
7535 left = ctx->nr_user_files;
7537 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7539 ret = __io_sqe_files_scm(ctx, this_files, total);
7543 total += this_files;
7549 while (total < ctx->nr_user_files) {
7550 struct file *file = io_file_from_index(ctx, total);
7560 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7566 static void io_rsrc_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
7568 struct file *file = prsrc->file;
7569 #if defined(CONFIG_UNIX)
7570 struct sock *sock = ctx->ring_sock->sk;
7571 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7572 struct sk_buff *skb;
7575 __skb_queue_head_init(&list);
7578 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7579 * remove this entry and rearrange the file array.
7581 skb = skb_dequeue(head);
7583 struct scm_fp_list *fp;
7585 fp = UNIXCB(skb).fp;
7586 for (i = 0; i < fp->count; i++) {
7589 if (fp->fp[i] != file)
7592 unix_notinflight(fp->user, fp->fp[i]);
7593 left = fp->count - 1 - i;
7595 memmove(&fp->fp[i], &fp->fp[i + 1],
7596 left * sizeof(struct file *));
7603 __skb_queue_tail(&list, skb);
7613 __skb_queue_tail(&list, skb);
7615 skb = skb_dequeue(head);
7618 if (skb_peek(&list)) {
7619 spin_lock_irq(&head->lock);
7620 while ((skb = __skb_dequeue(&list)) != NULL)
7621 __skb_queue_tail(head, skb);
7622 spin_unlock_irq(&head->lock);
7629 static void __io_rsrc_put_work(struct io_rsrc_node *ref_node)
7631 struct io_rsrc_data *rsrc_data = ref_node->rsrc_data;
7632 struct io_ring_ctx *ctx = rsrc_data->ctx;
7633 struct io_rsrc_put *prsrc, *tmp;
7635 list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
7636 list_del(&prsrc->list);
7639 bool lock_ring = ctx->flags & IORING_SETUP_IOPOLL;
7641 io_ring_submit_lock(ctx, lock_ring);
7642 spin_lock_irq(&ctx->completion_lock);
7643 io_cqring_fill_event(ctx, prsrc->tag, 0, 0);
7645 io_commit_cqring(ctx);
7646 spin_unlock_irq(&ctx->completion_lock);
7647 io_cqring_ev_posted(ctx);
7648 io_ring_submit_unlock(ctx, lock_ring);
7651 rsrc_data->do_put(ctx, prsrc);
7655 io_rsrc_node_destroy(ref_node);
7656 if (atomic_dec_and_test(&rsrc_data->refs))
7657 complete(&rsrc_data->done);
7660 static void io_rsrc_put_work(struct work_struct *work)
7662 struct io_ring_ctx *ctx;
7663 struct llist_node *node;
7665 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
7666 node = llist_del_all(&ctx->rsrc_put_llist);
7669 struct io_rsrc_node *ref_node;
7670 struct llist_node *next = node->next;
7672 ref_node = llist_entry(node, struct io_rsrc_node, llist);
7673 __io_rsrc_put_work(ref_node);
7678 static void io_rsrc_node_ref_zero(struct percpu_ref *ref)
7680 struct io_rsrc_node *node = container_of(ref, struct io_rsrc_node, refs);
7681 struct io_ring_ctx *ctx = node->rsrc_data->ctx;
7682 unsigned long flags;
7683 bool first_add = false;
7685 spin_lock_irqsave(&ctx->rsrc_ref_lock, flags);
7688 while (!list_empty(&ctx->rsrc_ref_list)) {
7689 node = list_first_entry(&ctx->rsrc_ref_list,
7690 struct io_rsrc_node, node);
7691 /* recycle ref nodes in order */
7694 list_del(&node->node);
7695 first_add |= llist_add(&node->llist, &ctx->rsrc_put_llist);
7697 spin_unlock_irqrestore(&ctx->rsrc_ref_lock, flags);
7700 mod_delayed_work(system_wq, &ctx->rsrc_put_work, HZ);
7703 static struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx)
7705 struct io_rsrc_node *ref_node;
7707 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7711 if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
7716 INIT_LIST_HEAD(&ref_node->node);
7717 INIT_LIST_HEAD(&ref_node->rsrc_list);
7718 ref_node->done = false;
7722 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7723 unsigned nr_args, u64 __user *tags)
7725 __s32 __user *fds = (__s32 __user *) arg;
7734 if (nr_args > IORING_MAX_FIXED_FILES)
7736 ret = io_rsrc_node_switch_start(ctx);
7739 ret = io_rsrc_data_alloc(ctx, io_rsrc_file_put, tags, nr_args,
7745 if (!io_alloc_file_tables(&ctx->file_table, nr_args))
7748 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7749 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7753 /* allow sparse sets */
7756 if (unlikely(*io_get_tag_slot(ctx->file_data, i)))
7763 if (unlikely(!file))
7767 * Don't allow io_uring instances to be registered. If UNIX
7768 * isn't enabled, then this causes a reference cycle and this
7769 * instance can never get freed. If UNIX is enabled we'll
7770 * handle it just fine, but there's still no point in allowing
7771 * a ring fd as it doesn't support regular read/write anyway.
7773 if (file->f_op == &io_uring_fops) {
7777 io_fixed_file_set(io_fixed_file_slot(&ctx->file_table, i), file);
7780 ret = io_sqe_files_scm(ctx);
7782 __io_sqe_files_unregister(ctx);
7786 io_rsrc_node_switch(ctx, NULL);
7789 for (i = 0; i < ctx->nr_user_files; i++) {
7790 file = io_file_from_index(ctx, i);
7794 io_free_file_tables(&ctx->file_table);
7795 ctx->nr_user_files = 0;
7797 io_rsrc_data_free(ctx->file_data);
7798 ctx->file_data = NULL;
7802 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7805 #if defined(CONFIG_UNIX)
7806 struct sock *sock = ctx->ring_sock->sk;
7807 struct sk_buff_head *head = &sock->sk_receive_queue;
7808 struct sk_buff *skb;
7811 * See if we can merge this file into an existing skb SCM_RIGHTS
7812 * file set. If there's no room, fall back to allocating a new skb
7813 * and filling it in.
7815 spin_lock_irq(&head->lock);
7816 skb = skb_peek(head);
7818 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7820 if (fpl->count < SCM_MAX_FD) {
7821 __skb_unlink(skb, head);
7822 spin_unlock_irq(&head->lock);
7823 fpl->fp[fpl->count] = get_file(file);
7824 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7826 spin_lock_irq(&head->lock);
7827 __skb_queue_head(head, skb);
7832 spin_unlock_irq(&head->lock);
7839 return __io_sqe_files_scm(ctx, 1, index);
7845 static int io_queue_rsrc_removal(struct io_rsrc_data *data, unsigned idx,
7846 struct io_rsrc_node *node, void *rsrc)
7848 struct io_rsrc_put *prsrc;
7850 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
7854 prsrc->tag = *io_get_tag_slot(data, idx);
7856 list_add(&prsrc->list, &node->rsrc_list);
7860 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7861 struct io_uring_rsrc_update2 *up,
7864 u64 __user *tags = u64_to_user_ptr(up->tags);
7865 __s32 __user *fds = u64_to_user_ptr(up->data);
7866 struct io_rsrc_data *data = ctx->file_data;
7867 struct io_fixed_file *file_slot;
7871 bool needs_switch = false;
7873 if (!ctx->file_data)
7875 if (up->offset + nr_args > ctx->nr_user_files)
7878 for (done = 0; done < nr_args; done++) {
7881 if ((tags && copy_from_user(&tag, &tags[done], sizeof(tag))) ||
7882 copy_from_user(&fd, &fds[done], sizeof(fd))) {
7886 if ((fd == IORING_REGISTER_FILES_SKIP || fd == -1) && tag) {
7890 if (fd == IORING_REGISTER_FILES_SKIP)
7893 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
7894 file_slot = io_fixed_file_slot(&ctx->file_table, i);
7896 if (file_slot->file_ptr) {
7897 file = (struct file *)(file_slot->file_ptr & FFS_MASK);
7898 err = io_queue_rsrc_removal(data, up->offset + done,
7899 ctx->rsrc_node, file);
7902 file_slot->file_ptr = 0;
7903 needs_switch = true;
7912 * Don't allow io_uring instances to be registered. If
7913 * UNIX isn't enabled, then this causes a reference
7914 * cycle and this instance can never get freed. If UNIX
7915 * is enabled we'll handle it just fine, but there's
7916 * still no point in allowing a ring fd as it doesn't
7917 * support regular read/write anyway.
7919 if (file->f_op == &io_uring_fops) {
7924 *io_get_tag_slot(data, up->offset + done) = tag;
7925 io_fixed_file_set(file_slot, file);
7926 err = io_sqe_file_register(ctx, file, i);
7928 file_slot->file_ptr = 0;
7936 io_rsrc_node_switch(ctx, data);
7937 return done ? done : err;
7940 static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx,
7941 struct task_struct *task)
7943 struct io_wq_hash *hash;
7944 struct io_wq_data data;
7945 unsigned int concurrency;
7947 mutex_lock(&ctx->uring_lock);
7948 hash = ctx->hash_map;
7950 hash = kzalloc(sizeof(*hash), GFP_KERNEL);
7952 mutex_unlock(&ctx->uring_lock);
7953 return ERR_PTR(-ENOMEM);
7955 refcount_set(&hash->refs, 1);
7956 init_waitqueue_head(&hash->wait);
7957 ctx->hash_map = hash;
7959 mutex_unlock(&ctx->uring_lock);
7963 data.free_work = io_wq_free_work;
7964 data.do_work = io_wq_submit_work;
7966 /* Do QD, or 4 * CPUS, whatever is smallest */
7967 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7969 return io_wq_create(concurrency, &data);
7972 static int io_uring_alloc_task_context(struct task_struct *task,
7973 struct io_ring_ctx *ctx)
7975 struct io_uring_task *tctx;
7978 tctx = kzalloc(sizeof(*tctx), GFP_KERNEL);
7979 if (unlikely(!tctx))
7982 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
7983 if (unlikely(ret)) {
7988 tctx->io_wq = io_init_wq_offload(ctx, task);
7989 if (IS_ERR(tctx->io_wq)) {
7990 ret = PTR_ERR(tctx->io_wq);
7991 percpu_counter_destroy(&tctx->inflight);
7997 init_waitqueue_head(&tctx->wait);
7998 atomic_set(&tctx->in_idle, 0);
7999 atomic_set(&tctx->inflight_tracked, 0);
8000 task->io_uring = tctx;
8001 spin_lock_init(&tctx->task_lock);
8002 INIT_WQ_LIST(&tctx->task_list);
8003 init_task_work(&tctx->task_work, tctx_task_work);
8007 void __io_uring_free(struct task_struct *tsk)
8009 struct io_uring_task *tctx = tsk->io_uring;
8011 WARN_ON_ONCE(!xa_empty(&tctx->xa));
8012 WARN_ON_ONCE(tctx->io_wq);
8013 WARN_ON_ONCE(tctx->cached_refs);
8015 percpu_counter_destroy(&tctx->inflight);
8017 tsk->io_uring = NULL;
8020 static int io_sq_offload_create(struct io_ring_ctx *ctx,
8021 struct io_uring_params *p)
8025 /* Retain compatibility with failing for an invalid attach attempt */
8026 if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
8027 IORING_SETUP_ATTACH_WQ) {
8030 f = fdget(p->wq_fd);
8033 if (f.file->f_op != &io_uring_fops) {
8039 if (ctx->flags & IORING_SETUP_SQPOLL) {
8040 struct task_struct *tsk;
8041 struct io_sq_data *sqd;
8044 sqd = io_get_sq_data(p, &attached);
8050 ctx->sq_creds = get_current_cred();
8052 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
8053 if (!ctx->sq_thread_idle)
8054 ctx->sq_thread_idle = HZ;
8056 io_sq_thread_park(sqd);
8057 list_add(&ctx->sqd_list, &sqd->ctx_list);
8058 io_sqd_update_thread_idle(sqd);
8059 /* don't attach to a dying SQPOLL thread, would be racy */
8060 ret = (attached && !sqd->thread) ? -ENXIO : 0;
8061 io_sq_thread_unpark(sqd);
8068 if (p->flags & IORING_SETUP_SQ_AFF) {
8069 int cpu = p->sq_thread_cpu;
8072 if (cpu >= nr_cpu_ids || !cpu_online(cpu))
8079 sqd->task_pid = current->pid;
8080 sqd->task_tgid = current->tgid;
8081 tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE);
8088 ret = io_uring_alloc_task_context(tsk, ctx);
8089 wake_up_new_task(tsk);
8092 } else if (p->flags & IORING_SETUP_SQ_AFF) {
8093 /* Can't have SQ_AFF without SQPOLL */
8100 complete(&ctx->sq_data->exited);
8102 io_sq_thread_finish(ctx);
8106 static inline void __io_unaccount_mem(struct user_struct *user,
8107 unsigned long nr_pages)
8109 atomic_long_sub(nr_pages, &user->locked_vm);
8112 static inline int __io_account_mem(struct user_struct *user,
8113 unsigned long nr_pages)
8115 unsigned long page_limit, cur_pages, new_pages;
8117 /* Don't allow more pages than we can safely lock */
8118 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
8121 cur_pages = atomic_long_read(&user->locked_vm);
8122 new_pages = cur_pages + nr_pages;
8123 if (new_pages > page_limit)
8125 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
8126 new_pages) != cur_pages);
8131 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8134 __io_unaccount_mem(ctx->user, nr_pages);
8136 if (ctx->mm_account)
8137 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
8140 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8145 ret = __io_account_mem(ctx->user, nr_pages);
8150 if (ctx->mm_account)
8151 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
8156 static void io_mem_free(void *ptr)
8163 page = virt_to_head_page(ptr);
8164 if (put_page_testzero(page))
8165 free_compound_page(page);
8168 static void *io_mem_alloc(size_t size)
8170 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
8171 __GFP_NORETRY | __GFP_ACCOUNT;
8173 return (void *) __get_free_pages(gfp_flags, get_order(size));
8176 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8179 struct io_rings *rings;
8180 size_t off, sq_array_size;
8182 off = struct_size(rings, cqes, cq_entries);
8183 if (off == SIZE_MAX)
8187 off = ALIGN(off, SMP_CACHE_BYTES);
8195 sq_array_size = array_size(sizeof(u32), sq_entries);
8196 if (sq_array_size == SIZE_MAX)
8199 if (check_add_overflow(off, sq_array_size, &off))
8205 static void io_buffer_unmap(struct io_ring_ctx *ctx, struct io_mapped_ubuf **slot)
8207 struct io_mapped_ubuf *imu = *slot;
8210 if (imu != ctx->dummy_ubuf) {
8211 for (i = 0; i < imu->nr_bvecs; i++)
8212 unpin_user_page(imu->bvec[i].bv_page);
8213 if (imu->acct_pages)
8214 io_unaccount_mem(ctx, imu->acct_pages);
8220 static void io_rsrc_buf_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
8222 io_buffer_unmap(ctx, &prsrc->buf);
8226 static void __io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8230 for (i = 0; i < ctx->nr_user_bufs; i++)
8231 io_buffer_unmap(ctx, &ctx->user_bufs[i]);
8232 kfree(ctx->user_bufs);
8233 io_rsrc_data_free(ctx->buf_data);
8234 ctx->user_bufs = NULL;
8235 ctx->buf_data = NULL;
8236 ctx->nr_user_bufs = 0;
8239 static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8246 ret = io_rsrc_ref_quiesce(ctx->buf_data, ctx);
8248 __io_sqe_buffers_unregister(ctx);
8252 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8253 void __user *arg, unsigned index)
8255 struct iovec __user *src;
8257 #ifdef CONFIG_COMPAT
8259 struct compat_iovec __user *ciovs;
8260 struct compat_iovec ciov;
8262 ciovs = (struct compat_iovec __user *) arg;
8263 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8266 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8267 dst->iov_len = ciov.iov_len;
8271 src = (struct iovec __user *) arg;
8272 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8278 * Not super efficient, but this is just a registration time. And we do cache
8279 * the last compound head, so generally we'll only do a full search if we don't
8282 * We check if the given compound head page has already been accounted, to
8283 * avoid double accounting it. This allows us to account the full size of the
8284 * page, not just the constituent pages of a huge page.
8286 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8287 int nr_pages, struct page *hpage)
8291 /* check current page array */
8292 for (i = 0; i < nr_pages; i++) {
8293 if (!PageCompound(pages[i]))
8295 if (compound_head(pages[i]) == hpage)
8299 /* check previously registered pages */
8300 for (i = 0; i < ctx->nr_user_bufs; i++) {
8301 struct io_mapped_ubuf *imu = ctx->user_bufs[i];
8303 for (j = 0; j < imu->nr_bvecs; j++) {
8304 if (!PageCompound(imu->bvec[j].bv_page))
8306 if (compound_head(imu->bvec[j].bv_page) == hpage)
8314 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8315 int nr_pages, struct io_mapped_ubuf *imu,
8316 struct page **last_hpage)
8320 imu->acct_pages = 0;
8321 for (i = 0; i < nr_pages; i++) {
8322 if (!PageCompound(pages[i])) {
8327 hpage = compound_head(pages[i]);
8328 if (hpage == *last_hpage)
8330 *last_hpage = hpage;
8331 if (headpage_already_acct(ctx, pages, i, hpage))
8333 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8337 if (!imu->acct_pages)
8340 ret = io_account_mem(ctx, imu->acct_pages);
8342 imu->acct_pages = 0;
8346 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
8347 struct io_mapped_ubuf **pimu,
8348 struct page **last_hpage)
8350 struct io_mapped_ubuf *imu = NULL;
8351 struct vm_area_struct **vmas = NULL;
8352 struct page **pages = NULL;
8353 unsigned long off, start, end, ubuf;
8355 int ret, pret, nr_pages, i;
8357 if (!iov->iov_base) {
8358 *pimu = ctx->dummy_ubuf;
8362 ubuf = (unsigned long) iov->iov_base;
8363 end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8364 start = ubuf >> PAGE_SHIFT;
8365 nr_pages = end - start;
8370 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
8374 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
8379 imu = kvmalloc(struct_size(imu, bvec, nr_pages), GFP_KERNEL);
8384 mmap_read_lock(current->mm);
8385 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
8387 if (pret == nr_pages) {
8388 /* don't support file backed memory */
8389 for (i = 0; i < nr_pages; i++) {
8390 struct vm_area_struct *vma = vmas[i];
8392 if (vma_is_shmem(vma))
8395 !is_file_hugepages(vma->vm_file)) {
8401 ret = pret < 0 ? pret : -EFAULT;
8403 mmap_read_unlock(current->mm);
8406 * if we did partial map, or found file backed vmas,
8407 * release any pages we did get
8410 unpin_user_pages(pages, pret);
8414 ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
8416 unpin_user_pages(pages, pret);
8420 off = ubuf & ~PAGE_MASK;
8421 size = iov->iov_len;
8422 for (i = 0; i < nr_pages; i++) {
8425 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8426 imu->bvec[i].bv_page = pages[i];
8427 imu->bvec[i].bv_len = vec_len;
8428 imu->bvec[i].bv_offset = off;
8432 /* store original address for later verification */
8434 imu->ubuf_end = ubuf + iov->iov_len;
8435 imu->nr_bvecs = nr_pages;
8446 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
8448 ctx->user_bufs = kcalloc(nr_args, sizeof(*ctx->user_bufs), GFP_KERNEL);
8449 return ctx->user_bufs ? 0 : -ENOMEM;
8452 static int io_buffer_validate(struct iovec *iov)
8454 unsigned long tmp, acct_len = iov->iov_len + (PAGE_SIZE - 1);
8457 * Don't impose further limits on the size and buffer
8458 * constraints here, we'll -EINVAL later when IO is
8459 * submitted if they are wrong.
8462 return iov->iov_len ? -EFAULT : 0;
8466 /* arbitrary limit, but we need something */
8467 if (iov->iov_len > SZ_1G)
8470 if (check_add_overflow((unsigned long)iov->iov_base, acct_len, &tmp))
8476 static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
8477 unsigned int nr_args, u64 __user *tags)
8479 struct page *last_hpage = NULL;
8480 struct io_rsrc_data *data;
8486 if (!nr_args || nr_args > IORING_MAX_REG_BUFFERS)
8488 ret = io_rsrc_node_switch_start(ctx);
8491 ret = io_rsrc_data_alloc(ctx, io_rsrc_buf_put, tags, nr_args, &data);
8494 ret = io_buffers_map_alloc(ctx, nr_args);
8496 io_rsrc_data_free(data);
8500 for (i = 0; i < nr_args; i++, ctx->nr_user_bufs++) {
8501 ret = io_copy_iov(ctx, &iov, arg, i);
8504 ret = io_buffer_validate(&iov);
8507 if (!iov.iov_base && *io_get_tag_slot(data, i)) {
8512 ret = io_sqe_buffer_register(ctx, &iov, &ctx->user_bufs[i],
8518 WARN_ON_ONCE(ctx->buf_data);
8520 ctx->buf_data = data;
8522 __io_sqe_buffers_unregister(ctx);
8524 io_rsrc_node_switch(ctx, NULL);
8528 static int __io_sqe_buffers_update(struct io_ring_ctx *ctx,
8529 struct io_uring_rsrc_update2 *up,
8530 unsigned int nr_args)
8532 u64 __user *tags = u64_to_user_ptr(up->tags);
8533 struct iovec iov, __user *iovs = u64_to_user_ptr(up->data);
8534 struct page *last_hpage = NULL;
8535 bool needs_switch = false;
8541 if (up->offset + nr_args > ctx->nr_user_bufs)
8544 for (done = 0; done < nr_args; done++) {
8545 struct io_mapped_ubuf *imu;
8546 int offset = up->offset + done;
8549 err = io_copy_iov(ctx, &iov, iovs, done);
8552 if (tags && copy_from_user(&tag, &tags[done], sizeof(tag))) {
8556 err = io_buffer_validate(&iov);
8559 if (!iov.iov_base && tag) {
8563 err = io_sqe_buffer_register(ctx, &iov, &imu, &last_hpage);
8567 i = array_index_nospec(offset, ctx->nr_user_bufs);
8568 if (ctx->user_bufs[i] != ctx->dummy_ubuf) {
8569 err = io_queue_rsrc_removal(ctx->buf_data, offset,
8570 ctx->rsrc_node, ctx->user_bufs[i]);
8571 if (unlikely(err)) {
8572 io_buffer_unmap(ctx, &imu);
8575 ctx->user_bufs[i] = NULL;
8576 needs_switch = true;
8579 ctx->user_bufs[i] = imu;
8580 *io_get_tag_slot(ctx->buf_data, offset) = tag;
8584 io_rsrc_node_switch(ctx, ctx->buf_data);
8585 return done ? done : err;
8588 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8590 __s32 __user *fds = arg;
8596 if (copy_from_user(&fd, fds, sizeof(*fds)))
8599 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8600 if (IS_ERR(ctx->cq_ev_fd)) {
8601 int ret = PTR_ERR(ctx->cq_ev_fd);
8603 ctx->cq_ev_fd = NULL;
8610 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8612 if (ctx->cq_ev_fd) {
8613 eventfd_ctx_put(ctx->cq_ev_fd);
8614 ctx->cq_ev_fd = NULL;
8621 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8623 struct io_buffer *buf;
8624 unsigned long index;
8626 xa_for_each(&ctx->io_buffers, index, buf)
8627 __io_remove_buffers(ctx, buf, index, -1U);
8630 static void io_req_cache_free(struct list_head *list, struct task_struct *tsk)
8632 struct io_kiocb *req, *nxt;
8634 list_for_each_entry_safe(req, nxt, list, compl.list) {
8635 if (tsk && req->task != tsk)
8637 list_del(&req->compl.list);
8638 kmem_cache_free(req_cachep, req);
8642 static void io_req_caches_free(struct io_ring_ctx *ctx)
8644 struct io_submit_state *submit_state = &ctx->submit_state;
8645 struct io_comp_state *cs = &ctx->submit_state.comp;
8647 mutex_lock(&ctx->uring_lock);
8649 if (submit_state->free_reqs) {
8650 kmem_cache_free_bulk(req_cachep, submit_state->free_reqs,
8651 submit_state->reqs);
8652 submit_state->free_reqs = 0;
8655 io_flush_cached_locked_reqs(ctx, cs);
8656 io_req_cache_free(&cs->free_list, NULL);
8657 mutex_unlock(&ctx->uring_lock);
8660 static void io_wait_rsrc_data(struct io_rsrc_data *data)
8662 if (data && !atomic_dec_and_test(&data->refs))
8663 wait_for_completion(&data->done);
8666 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8668 io_sq_thread_finish(ctx);
8670 if (ctx->mm_account) {
8671 mmdrop(ctx->mm_account);
8672 ctx->mm_account = NULL;
8675 /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */
8676 io_wait_rsrc_data(ctx->buf_data);
8677 io_wait_rsrc_data(ctx->file_data);
8679 mutex_lock(&ctx->uring_lock);
8681 __io_sqe_buffers_unregister(ctx);
8683 __io_sqe_files_unregister(ctx);
8685 __io_cqring_overflow_flush(ctx, true);
8686 mutex_unlock(&ctx->uring_lock);
8687 io_eventfd_unregister(ctx);
8688 io_destroy_buffers(ctx);
8690 put_cred(ctx->sq_creds);
8692 /* there are no registered resources left, nobody uses it */
8694 io_rsrc_node_destroy(ctx->rsrc_node);
8695 if (ctx->rsrc_backup_node)
8696 io_rsrc_node_destroy(ctx->rsrc_backup_node);
8697 flush_delayed_work(&ctx->rsrc_put_work);
8699 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
8700 WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist));
8702 #if defined(CONFIG_UNIX)
8703 if (ctx->ring_sock) {
8704 ctx->ring_sock->file = NULL; /* so that iput() is called */
8705 sock_release(ctx->ring_sock);
8709 io_mem_free(ctx->rings);
8710 io_mem_free(ctx->sq_sqes);
8712 percpu_ref_exit(&ctx->refs);
8713 free_uid(ctx->user);
8714 io_req_caches_free(ctx);
8716 io_wq_put_hash(ctx->hash_map);
8717 kfree(ctx->cancel_hash);
8718 kfree(ctx->dummy_ubuf);
8722 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8724 struct io_ring_ctx *ctx = file->private_data;
8727 poll_wait(file, &ctx->poll_wait, wait);
8729 * synchronizes with barrier from wq_has_sleeper call in
8733 if (!io_sqring_full(ctx))
8734 mask |= EPOLLOUT | EPOLLWRNORM;
8737 * Don't flush cqring overflow list here, just do a simple check.
8738 * Otherwise there could possible be ABBA deadlock:
8741 * lock(&ctx->uring_lock);
8743 * lock(&ctx->uring_lock);
8746 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
8747 * pushs them to do the flush.
8749 if (io_cqring_events(ctx) || test_bit(0, &ctx->check_cq_overflow))
8750 mask |= EPOLLIN | EPOLLRDNORM;
8755 static int io_uring_fasync(int fd, struct file *file, int on)
8757 struct io_ring_ctx *ctx = file->private_data;
8759 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8762 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8764 const struct cred *creds;
8766 creds = xa_erase(&ctx->personalities, id);
8775 struct io_tctx_exit {
8776 struct callback_head task_work;
8777 struct completion completion;
8778 struct io_ring_ctx *ctx;
8781 static void io_tctx_exit_cb(struct callback_head *cb)
8783 struct io_uring_task *tctx = current->io_uring;
8784 struct io_tctx_exit *work;
8786 work = container_of(cb, struct io_tctx_exit, task_work);
8788 * When @in_idle, we're in cancellation and it's racy to remove the
8789 * node. It'll be removed by the end of cancellation, just ignore it.
8791 if (!atomic_read(&tctx->in_idle))
8792 io_uring_del_tctx_node((unsigned long)work->ctx);
8793 complete(&work->completion);
8796 static bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
8798 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8800 return req->ctx == data;
8803 static void io_ring_exit_work(struct work_struct *work)
8805 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
8806 unsigned long timeout = jiffies + HZ * 60 * 5;
8807 struct io_tctx_exit exit;
8808 struct io_tctx_node *node;
8812 * If we're doing polled IO and end up having requests being
8813 * submitted async (out-of-line), then completions can come in while
8814 * we're waiting for refs to drop. We need to reap these manually,
8815 * as nobody else will be looking for them.
8818 io_uring_try_cancel_requests(ctx, NULL, true);
8820 struct io_sq_data *sqd = ctx->sq_data;
8821 struct task_struct *tsk;
8823 io_sq_thread_park(sqd);
8825 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
8826 io_wq_cancel_cb(tsk->io_uring->io_wq,
8827 io_cancel_ctx_cb, ctx, true);
8828 io_sq_thread_unpark(sqd);
8831 WARN_ON_ONCE(time_after(jiffies, timeout));
8832 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
8834 init_completion(&exit.completion);
8835 init_task_work(&exit.task_work, io_tctx_exit_cb);
8838 * Some may use context even when all refs and requests have been put,
8839 * and they are free to do so while still holding uring_lock or
8840 * completion_lock, see io_req_task_submit(). Apart from other work,
8841 * this lock/unlock section also waits them to finish.
8843 mutex_lock(&ctx->uring_lock);
8844 while (!list_empty(&ctx->tctx_list)) {
8845 WARN_ON_ONCE(time_after(jiffies, timeout));
8847 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
8849 /* don't spin on a single task if cancellation failed */
8850 list_rotate_left(&ctx->tctx_list);
8851 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
8852 if (WARN_ON_ONCE(ret))
8854 wake_up_process(node->task);
8856 mutex_unlock(&ctx->uring_lock);
8857 wait_for_completion(&exit.completion);
8858 mutex_lock(&ctx->uring_lock);
8860 mutex_unlock(&ctx->uring_lock);
8861 spin_lock_irq(&ctx->completion_lock);
8862 spin_unlock_irq(&ctx->completion_lock);
8864 io_ring_ctx_free(ctx);
8867 /* Returns true if we found and killed one or more timeouts */
8868 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
8871 struct io_kiocb *req, *tmp;
8874 spin_lock_irq(&ctx->completion_lock);
8875 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
8876 if (io_match_task(req, tsk, cancel_all)) {
8877 io_kill_timeout(req, -ECANCELED);
8882 io_commit_cqring(ctx);
8883 spin_unlock_irq(&ctx->completion_lock);
8885 io_cqring_ev_posted(ctx);
8886 return canceled != 0;
8889 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8891 unsigned long index;
8892 struct creds *creds;
8894 mutex_lock(&ctx->uring_lock);
8895 percpu_ref_kill(&ctx->refs);
8897 __io_cqring_overflow_flush(ctx, true);
8898 xa_for_each(&ctx->personalities, index, creds)
8899 io_unregister_personality(ctx, index);
8900 mutex_unlock(&ctx->uring_lock);
8902 io_kill_timeouts(ctx, NULL, true);
8903 io_poll_remove_all(ctx, NULL, true);
8905 /* if we failed setting up the ctx, we might not have any rings */
8906 io_iopoll_try_reap_events(ctx);
8908 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8910 * Use system_unbound_wq to avoid spawning tons of event kworkers
8911 * if we're exiting a ton of rings at the same time. It just adds
8912 * noise and overhead, there's no discernable change in runtime
8913 * over using system_wq.
8915 queue_work(system_unbound_wq, &ctx->exit_work);
8918 static int io_uring_release(struct inode *inode, struct file *file)
8920 struct io_ring_ctx *ctx = file->private_data;
8922 file->private_data = NULL;
8923 io_ring_ctx_wait_and_kill(ctx);
8927 struct io_task_cancel {
8928 struct task_struct *task;
8932 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8934 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8935 struct io_task_cancel *cancel = data;
8938 if (!cancel->all && (req->flags & REQ_F_LINK_TIMEOUT)) {
8939 unsigned long flags;
8940 struct io_ring_ctx *ctx = req->ctx;
8942 /* protect against races with linked timeouts */
8943 spin_lock_irqsave(&ctx->completion_lock, flags);
8944 ret = io_match_task(req, cancel->task, cancel->all);
8945 spin_unlock_irqrestore(&ctx->completion_lock, flags);
8947 ret = io_match_task(req, cancel->task, cancel->all);
8952 static bool io_cancel_defer_files(struct io_ring_ctx *ctx,
8953 struct task_struct *task, bool cancel_all)
8955 struct io_defer_entry *de;
8958 spin_lock_irq(&ctx->completion_lock);
8959 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8960 if (io_match_task(de->req, task, cancel_all)) {
8961 list_cut_position(&list, &ctx->defer_list, &de->list);
8965 spin_unlock_irq(&ctx->completion_lock);
8966 if (list_empty(&list))
8969 while (!list_empty(&list)) {
8970 de = list_first_entry(&list, struct io_defer_entry, list);
8971 list_del_init(&de->list);
8972 io_req_complete_failed(de->req, -ECANCELED);
8978 static bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
8980 struct io_tctx_node *node;
8981 enum io_wq_cancel cret;
8984 mutex_lock(&ctx->uring_lock);
8985 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
8986 struct io_uring_task *tctx = node->task->io_uring;
8989 * io_wq will stay alive while we hold uring_lock, because it's
8990 * killed after ctx nodes, which requires to take the lock.
8992 if (!tctx || !tctx->io_wq)
8994 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
8995 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8997 mutex_unlock(&ctx->uring_lock);
9002 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
9003 struct task_struct *task,
9006 struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
9007 struct io_uring_task *tctx = task ? task->io_uring : NULL;
9010 enum io_wq_cancel cret;
9014 ret |= io_uring_try_cancel_iowq(ctx);
9015 } else if (tctx && tctx->io_wq) {
9017 * Cancels requests of all rings, not only @ctx, but
9018 * it's fine as the task is in exit/exec.
9020 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
9022 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
9025 /* SQPOLL thread does its own polling */
9026 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
9027 (ctx->sq_data && ctx->sq_data->thread == current)) {
9028 while (!list_empty_careful(&ctx->iopoll_list)) {
9029 io_iopoll_try_reap_events(ctx);
9034 ret |= io_cancel_defer_files(ctx, task, cancel_all);
9035 ret |= io_poll_remove_all(ctx, task, cancel_all);
9036 ret |= io_kill_timeouts(ctx, task, cancel_all);
9038 ret |= io_run_task_work();
9045 static int __io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9047 struct io_uring_task *tctx = current->io_uring;
9048 struct io_tctx_node *node;
9051 if (unlikely(!tctx)) {
9052 ret = io_uring_alloc_task_context(current, ctx);
9055 tctx = current->io_uring;
9057 if (!xa_load(&tctx->xa, (unsigned long)ctx)) {
9058 node = kmalloc(sizeof(*node), GFP_KERNEL);
9062 node->task = current;
9064 ret = xa_err(xa_store(&tctx->xa, (unsigned long)ctx,
9071 mutex_lock(&ctx->uring_lock);
9072 list_add(&node->ctx_node, &ctx->tctx_list);
9073 mutex_unlock(&ctx->uring_lock);
9080 * Note that this task has used io_uring. We use it for cancelation purposes.
9082 static inline int io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9084 struct io_uring_task *tctx = current->io_uring;
9086 if (likely(tctx && tctx->last == ctx))
9088 return __io_uring_add_tctx_node(ctx);
9092 * Remove this io_uring_file -> task mapping.
9094 static void io_uring_del_tctx_node(unsigned long index)
9096 struct io_uring_task *tctx = current->io_uring;
9097 struct io_tctx_node *node;
9101 node = xa_erase(&tctx->xa, index);
9105 WARN_ON_ONCE(current != node->task);
9106 WARN_ON_ONCE(list_empty(&node->ctx_node));
9108 mutex_lock(&node->ctx->uring_lock);
9109 list_del(&node->ctx_node);
9110 mutex_unlock(&node->ctx->uring_lock);
9112 if (tctx->last == node->ctx)
9117 static void io_uring_clean_tctx(struct io_uring_task *tctx)
9119 struct io_wq *wq = tctx->io_wq;
9120 struct io_tctx_node *node;
9121 unsigned long index;
9123 xa_for_each(&tctx->xa, index, node)
9124 io_uring_del_tctx_node(index);
9127 * Must be after io_uring_del_task_file() (removes nodes under
9128 * uring_lock) to avoid race with io_uring_try_cancel_iowq().
9131 io_wq_put_and_exit(wq);
9135 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
9138 return atomic_read(&tctx->inflight_tracked);
9139 return percpu_counter_sum(&tctx->inflight);
9142 static void io_uring_drop_tctx_refs(struct task_struct *task)
9144 struct io_uring_task *tctx = task->io_uring;
9145 unsigned int refs = tctx->cached_refs;
9147 tctx->cached_refs = 0;
9148 percpu_counter_sub(&tctx->inflight, refs);
9149 put_task_struct_many(task, refs);
9153 * Find any io_uring ctx that this task has registered or done IO on, and cancel
9154 * requests. @sqd should be not-null IIF it's an SQPOLL thread cancellation.
9156 static void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
9158 struct io_uring_task *tctx = current->io_uring;
9159 struct io_ring_ctx *ctx;
9163 WARN_ON_ONCE(sqd && sqd->thread != current);
9165 if (!current->io_uring)
9168 io_wq_exit_start(tctx->io_wq);
9170 io_uring_drop_tctx_refs(current);
9171 atomic_inc(&tctx->in_idle);
9173 /* read completions before cancelations */
9174 inflight = tctx_inflight(tctx, !cancel_all);
9179 struct io_tctx_node *node;
9180 unsigned long index;
9182 xa_for_each(&tctx->xa, index, node) {
9183 /* sqpoll task will cancel all its requests */
9184 if (node->ctx->sq_data)
9186 io_uring_try_cancel_requests(node->ctx, current,
9190 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
9191 io_uring_try_cancel_requests(ctx, current,
9195 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
9197 * If we've seen completions, retry without waiting. This
9198 * avoids a race where a completion comes in before we did
9199 * prepare_to_wait().
9201 if (inflight == tctx_inflight(tctx, !cancel_all))
9203 finish_wait(&tctx->wait, &wait);
9205 atomic_dec(&tctx->in_idle);
9207 io_uring_clean_tctx(tctx);
9209 /* for exec all current's requests should be gone, kill tctx */
9210 __io_uring_free(current);
9214 void __io_uring_cancel(struct files_struct *files)
9216 io_uring_cancel_generic(!files, NULL);
9219 static void *io_uring_validate_mmap_request(struct file *file,
9220 loff_t pgoff, size_t sz)
9222 struct io_ring_ctx *ctx = file->private_data;
9223 loff_t offset = pgoff << PAGE_SHIFT;
9228 case IORING_OFF_SQ_RING:
9229 case IORING_OFF_CQ_RING:
9232 case IORING_OFF_SQES:
9236 return ERR_PTR(-EINVAL);
9239 page = virt_to_head_page(ptr);
9240 if (sz > page_size(page))
9241 return ERR_PTR(-EINVAL);
9248 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9250 size_t sz = vma->vm_end - vma->vm_start;
9254 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
9256 return PTR_ERR(ptr);
9258 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
9259 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
9262 #else /* !CONFIG_MMU */
9264 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9266 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
9269 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
9271 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
9274 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
9275 unsigned long addr, unsigned long len,
9276 unsigned long pgoff, unsigned long flags)
9280 ptr = io_uring_validate_mmap_request(file, pgoff, len);
9282 return PTR_ERR(ptr);
9284 return (unsigned long) ptr;
9287 #endif /* !CONFIG_MMU */
9289 static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
9294 if (!io_sqring_full(ctx))
9296 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9298 if (!io_sqring_full(ctx))
9301 } while (!signal_pending(current));
9303 finish_wait(&ctx->sqo_sq_wait, &wait);
9307 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9308 struct __kernel_timespec __user **ts,
9309 const sigset_t __user **sig)
9311 struct io_uring_getevents_arg arg;
9314 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9315 * is just a pointer to the sigset_t.
9317 if (!(flags & IORING_ENTER_EXT_ARG)) {
9318 *sig = (const sigset_t __user *) argp;
9324 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9325 * timespec and sigset_t pointers if good.
9327 if (*argsz != sizeof(arg))
9329 if (copy_from_user(&arg, argp, sizeof(arg)))
9331 *sig = u64_to_user_ptr(arg.sigmask);
9332 *argsz = arg.sigmask_sz;
9333 *ts = u64_to_user_ptr(arg.ts);
9337 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9338 u32, min_complete, u32, flags, const void __user *, argp,
9341 struct io_ring_ctx *ctx;
9348 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9349 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG)))
9353 if (unlikely(!f.file))
9357 if (unlikely(f.file->f_op != &io_uring_fops))
9361 ctx = f.file->private_data;
9362 if (unlikely(!percpu_ref_tryget(&ctx->refs)))
9366 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
9370 * For SQ polling, the thread will do all submissions and completions.
9371 * Just return the requested submit count, and wake the thread if
9375 if (ctx->flags & IORING_SETUP_SQPOLL) {
9376 io_cqring_overflow_flush(ctx, false);
9378 if (unlikely(ctx->sq_data->thread == NULL)) {
9382 if (flags & IORING_ENTER_SQ_WAKEUP)
9383 wake_up(&ctx->sq_data->wait);
9384 if (flags & IORING_ENTER_SQ_WAIT) {
9385 ret = io_sqpoll_wait_sq(ctx);
9389 submitted = to_submit;
9390 } else if (to_submit) {
9391 ret = io_uring_add_tctx_node(ctx);
9394 mutex_lock(&ctx->uring_lock);
9395 submitted = io_submit_sqes(ctx, to_submit);
9396 mutex_unlock(&ctx->uring_lock);
9398 if (submitted != to_submit)
9401 if (flags & IORING_ENTER_GETEVENTS) {
9402 const sigset_t __user *sig;
9403 struct __kernel_timespec __user *ts;
9405 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
9409 min_complete = min(min_complete, ctx->cq_entries);
9412 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9413 * space applications don't need to do io completion events
9414 * polling again, they can rely on io_sq_thread to do polling
9415 * work, which can reduce cpu usage and uring_lock contention.
9417 if (ctx->flags & IORING_SETUP_IOPOLL &&
9418 !(ctx->flags & IORING_SETUP_SQPOLL)) {
9419 ret = io_iopoll_check(ctx, min_complete);
9421 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
9426 percpu_ref_put(&ctx->refs);
9429 return submitted ? submitted : ret;
9432 #ifdef CONFIG_PROC_FS
9433 static int io_uring_show_cred(struct seq_file *m, unsigned int id,
9434 const struct cred *cred)
9436 struct user_namespace *uns = seq_user_ns(m);
9437 struct group_info *gi;
9442 seq_printf(m, "%5d\n", id);
9443 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
9444 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
9445 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
9446 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
9447 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
9448 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
9449 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
9450 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
9451 seq_puts(m, "\n\tGroups:\t");
9452 gi = cred->group_info;
9453 for (g = 0; g < gi->ngroups; g++) {
9454 seq_put_decimal_ull(m, g ? " " : "",
9455 from_kgid_munged(uns, gi->gid[g]));
9457 seq_puts(m, "\n\tCapEff:\t");
9458 cap = cred->cap_effective;
9459 CAP_FOR_EACH_U32(__capi)
9460 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
9465 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
9467 struct io_sq_data *sq = NULL;
9472 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9473 * since fdinfo case grabs it in the opposite direction of normal use
9474 * cases. If we fail to get the lock, we just don't iterate any
9475 * structures that could be going away outside the io_uring mutex.
9477 has_lock = mutex_trylock(&ctx->uring_lock);
9479 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) {
9485 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9486 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9487 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9488 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9489 struct file *f = io_file_from_index(ctx, i);
9492 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9494 seq_printf(m, "%5u: <none>\n", i);
9496 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9497 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9498 struct io_mapped_ubuf *buf = ctx->user_bufs[i];
9499 unsigned int len = buf->ubuf_end - buf->ubuf;
9501 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf, len);
9503 if (has_lock && !xa_empty(&ctx->personalities)) {
9504 unsigned long index;
9505 const struct cred *cred;
9507 seq_printf(m, "Personalities:\n");
9508 xa_for_each(&ctx->personalities, index, cred)
9509 io_uring_show_cred(m, index, cred);
9511 seq_printf(m, "PollList:\n");
9512 spin_lock_irq(&ctx->completion_lock);
9513 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9514 struct hlist_head *list = &ctx->cancel_hash[i];
9515 struct io_kiocb *req;
9517 hlist_for_each_entry(req, list, hash_node)
9518 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9519 req->task->task_works != NULL);
9521 spin_unlock_irq(&ctx->completion_lock);
9523 mutex_unlock(&ctx->uring_lock);
9526 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9528 struct io_ring_ctx *ctx = f->private_data;
9530 if (percpu_ref_tryget(&ctx->refs)) {
9531 __io_uring_show_fdinfo(ctx, m);
9532 percpu_ref_put(&ctx->refs);
9537 static const struct file_operations io_uring_fops = {
9538 .release = io_uring_release,
9539 .mmap = io_uring_mmap,
9541 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9542 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9544 .poll = io_uring_poll,
9545 .fasync = io_uring_fasync,
9546 #ifdef CONFIG_PROC_FS
9547 .show_fdinfo = io_uring_show_fdinfo,
9551 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9552 struct io_uring_params *p)
9554 struct io_rings *rings;
9555 size_t size, sq_array_offset;
9557 /* make sure these are sane, as we already accounted them */
9558 ctx->sq_entries = p->sq_entries;
9559 ctx->cq_entries = p->cq_entries;
9561 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9562 if (size == SIZE_MAX)
9565 rings = io_mem_alloc(size);
9570 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9571 rings->sq_ring_mask = p->sq_entries - 1;
9572 rings->cq_ring_mask = p->cq_entries - 1;
9573 rings->sq_ring_entries = p->sq_entries;
9574 rings->cq_ring_entries = p->cq_entries;
9576 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9577 if (size == SIZE_MAX) {
9578 io_mem_free(ctx->rings);
9583 ctx->sq_sqes = io_mem_alloc(size);
9584 if (!ctx->sq_sqes) {
9585 io_mem_free(ctx->rings);
9593 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
9597 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9601 ret = io_uring_add_tctx_node(ctx);
9606 fd_install(fd, file);
9611 * Allocate an anonymous fd, this is what constitutes the application
9612 * visible backing of an io_uring instance. The application mmaps this
9613 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9614 * we have to tie this fd to a socket for file garbage collection purposes.
9616 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
9619 #if defined(CONFIG_UNIX)
9622 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9625 return ERR_PTR(ret);
9628 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9629 O_RDWR | O_CLOEXEC);
9630 #if defined(CONFIG_UNIX)
9632 sock_release(ctx->ring_sock);
9633 ctx->ring_sock = NULL;
9635 ctx->ring_sock->file = file;
9641 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9642 struct io_uring_params __user *params)
9644 struct io_ring_ctx *ctx;
9650 if (entries > IORING_MAX_ENTRIES) {
9651 if (!(p->flags & IORING_SETUP_CLAMP))
9653 entries = IORING_MAX_ENTRIES;
9657 * Use twice as many entries for the CQ ring. It's possible for the
9658 * application to drive a higher depth than the size of the SQ ring,
9659 * since the sqes are only used at submission time. This allows for
9660 * some flexibility in overcommitting a bit. If the application has
9661 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9662 * of CQ ring entries manually.
9664 p->sq_entries = roundup_pow_of_two(entries);
9665 if (p->flags & IORING_SETUP_CQSIZE) {
9667 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9668 * to a power-of-two, if it isn't already. We do NOT impose
9669 * any cq vs sq ring sizing.
9673 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9674 if (!(p->flags & IORING_SETUP_CLAMP))
9676 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9678 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9679 if (p->cq_entries < p->sq_entries)
9682 p->cq_entries = 2 * p->sq_entries;
9685 ctx = io_ring_ctx_alloc(p);
9688 ctx->compat = in_compat_syscall();
9689 if (!capable(CAP_IPC_LOCK))
9690 ctx->user = get_uid(current_user());
9693 * This is just grabbed for accounting purposes. When a process exits,
9694 * the mm is exited and dropped before the files, hence we need to hang
9695 * on to this mm purely for the purposes of being able to unaccount
9696 * memory (locked/pinned vm). It's not used for anything else.
9698 mmgrab(current->mm);
9699 ctx->mm_account = current->mm;
9701 ret = io_allocate_scq_urings(ctx, p);
9705 ret = io_sq_offload_create(ctx, p);
9708 /* always set a rsrc node */
9709 ret = io_rsrc_node_switch_start(ctx);
9712 io_rsrc_node_switch(ctx, NULL);
9714 memset(&p->sq_off, 0, sizeof(p->sq_off));
9715 p->sq_off.head = offsetof(struct io_rings, sq.head);
9716 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9717 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9718 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9719 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9720 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9721 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9723 memset(&p->cq_off, 0, sizeof(p->cq_off));
9724 p->cq_off.head = offsetof(struct io_rings, cq.head);
9725 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9726 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9727 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9728 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9729 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9730 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9732 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9733 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9734 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9735 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
9736 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
9737 IORING_FEAT_RSRC_TAGS;
9739 if (copy_to_user(params, p, sizeof(*p))) {
9744 file = io_uring_get_file(ctx);
9746 ret = PTR_ERR(file);
9751 * Install ring fd as the very last thing, so we don't risk someone
9752 * having closed it before we finish setup
9754 ret = io_uring_install_fd(ctx, file);
9756 /* fput will clean it up */
9761 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9764 io_ring_ctx_wait_and_kill(ctx);
9769 * Sets up an aio uring context, and returns the fd. Applications asks for a
9770 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9771 * params structure passed in.
9773 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9775 struct io_uring_params p;
9778 if (copy_from_user(&p, params, sizeof(p)))
9780 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9785 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9786 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9787 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9788 IORING_SETUP_R_DISABLED))
9791 return io_uring_create(entries, &p, params);
9794 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9795 struct io_uring_params __user *, params)
9797 return io_uring_setup(entries, params);
9800 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9802 struct io_uring_probe *p;
9806 size = struct_size(p, ops, nr_args);
9807 if (size == SIZE_MAX)
9809 p = kzalloc(size, GFP_KERNEL);
9814 if (copy_from_user(p, arg, size))
9817 if (memchr_inv(p, 0, size))
9820 p->last_op = IORING_OP_LAST - 1;
9821 if (nr_args > IORING_OP_LAST)
9822 nr_args = IORING_OP_LAST;
9824 for (i = 0; i < nr_args; i++) {
9826 if (!io_op_defs[i].not_supported)
9827 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9832 if (copy_to_user(arg, p, size))
9839 static int io_register_personality(struct io_ring_ctx *ctx)
9841 const struct cred *creds;
9845 creds = get_current_cred();
9847 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
9848 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
9856 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9857 unsigned int nr_args)
9859 struct io_uring_restriction *res;
9863 /* Restrictions allowed only if rings started disabled */
9864 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9867 /* We allow only a single restrictions registration */
9868 if (ctx->restrictions.registered)
9871 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9874 size = array_size(nr_args, sizeof(*res));
9875 if (size == SIZE_MAX)
9878 res = memdup_user(arg, size);
9880 return PTR_ERR(res);
9884 for (i = 0; i < nr_args; i++) {
9885 switch (res[i].opcode) {
9886 case IORING_RESTRICTION_REGISTER_OP:
9887 if (res[i].register_op >= IORING_REGISTER_LAST) {
9892 __set_bit(res[i].register_op,
9893 ctx->restrictions.register_op);
9895 case IORING_RESTRICTION_SQE_OP:
9896 if (res[i].sqe_op >= IORING_OP_LAST) {
9901 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9903 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9904 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9906 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9907 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9916 /* Reset all restrictions if an error happened */
9918 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9920 ctx->restrictions.registered = true;
9926 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9928 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9931 if (ctx->restrictions.registered)
9932 ctx->restricted = 1;
9934 ctx->flags &= ~IORING_SETUP_R_DISABLED;
9935 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
9936 wake_up(&ctx->sq_data->wait);
9940 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
9941 struct io_uring_rsrc_update2 *up,
9949 if (check_add_overflow(up->offset, nr_args, &tmp))
9951 err = io_rsrc_node_switch_start(ctx);
9956 case IORING_RSRC_FILE:
9957 return __io_sqe_files_update(ctx, up, nr_args);
9958 case IORING_RSRC_BUFFER:
9959 return __io_sqe_buffers_update(ctx, up, nr_args);
9964 static int io_register_files_update(struct io_ring_ctx *ctx, void __user *arg,
9967 struct io_uring_rsrc_update2 up;
9971 memset(&up, 0, sizeof(up));
9972 if (copy_from_user(&up, arg, sizeof(struct io_uring_rsrc_update)))
9974 return __io_register_rsrc_update(ctx, IORING_RSRC_FILE, &up, nr_args);
9977 static int io_register_rsrc_update(struct io_ring_ctx *ctx, void __user *arg,
9978 unsigned size, unsigned type)
9980 struct io_uring_rsrc_update2 up;
9982 if (size != sizeof(up))
9984 if (copy_from_user(&up, arg, sizeof(up)))
9986 if (!up.nr || up.resv)
9988 return __io_register_rsrc_update(ctx, type, &up, up.nr);
9991 static int io_register_rsrc(struct io_ring_ctx *ctx, void __user *arg,
9992 unsigned int size, unsigned int type)
9994 struct io_uring_rsrc_register rr;
9996 /* keep it extendible */
9997 if (size != sizeof(rr))
10000 memset(&rr, 0, sizeof(rr));
10001 if (copy_from_user(&rr, arg, size))
10003 if (!rr.nr || rr.resv || rr.resv2)
10007 case IORING_RSRC_FILE:
10008 return io_sqe_files_register(ctx, u64_to_user_ptr(rr.data),
10009 rr.nr, u64_to_user_ptr(rr.tags));
10010 case IORING_RSRC_BUFFER:
10011 return io_sqe_buffers_register(ctx, u64_to_user_ptr(rr.data),
10012 rr.nr, u64_to_user_ptr(rr.tags));
10017 static int io_register_iowq_aff(struct io_ring_ctx *ctx, void __user *arg,
10020 struct io_uring_task *tctx = current->io_uring;
10021 cpumask_var_t new_mask;
10024 if (!tctx || !tctx->io_wq)
10027 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
10030 cpumask_clear(new_mask);
10031 if (len > cpumask_size())
10032 len = cpumask_size();
10034 if (copy_from_user(new_mask, arg, len)) {
10035 free_cpumask_var(new_mask);
10039 ret = io_wq_cpu_affinity(tctx->io_wq, new_mask);
10040 free_cpumask_var(new_mask);
10044 static int io_unregister_iowq_aff(struct io_ring_ctx *ctx)
10046 struct io_uring_task *tctx = current->io_uring;
10048 if (!tctx || !tctx->io_wq)
10051 return io_wq_cpu_affinity(tctx->io_wq, NULL);
10054 static bool io_register_op_must_quiesce(int op)
10057 case IORING_REGISTER_BUFFERS:
10058 case IORING_UNREGISTER_BUFFERS:
10059 case IORING_REGISTER_FILES:
10060 case IORING_UNREGISTER_FILES:
10061 case IORING_REGISTER_FILES_UPDATE:
10062 case IORING_REGISTER_PROBE:
10063 case IORING_REGISTER_PERSONALITY:
10064 case IORING_UNREGISTER_PERSONALITY:
10065 case IORING_REGISTER_FILES2:
10066 case IORING_REGISTER_FILES_UPDATE2:
10067 case IORING_REGISTER_BUFFERS2:
10068 case IORING_REGISTER_BUFFERS_UPDATE:
10069 case IORING_REGISTER_IOWQ_AFF:
10070 case IORING_UNREGISTER_IOWQ_AFF:
10077 static int io_ctx_quiesce(struct io_ring_ctx *ctx)
10081 percpu_ref_kill(&ctx->refs);
10084 * Drop uring mutex before waiting for references to exit. If another
10085 * thread is currently inside io_uring_enter() it might need to grab the
10086 * uring_lock to make progress. If we hold it here across the drain
10087 * wait, then we can deadlock. It's safe to drop the mutex here, since
10088 * no new references will come in after we've killed the percpu ref.
10090 mutex_unlock(&ctx->uring_lock);
10092 ret = wait_for_completion_interruptible(&ctx->ref_comp);
10095 ret = io_run_task_work_sig();
10096 } while (ret >= 0);
10097 mutex_lock(&ctx->uring_lock);
10100 io_refs_resurrect(&ctx->refs, &ctx->ref_comp);
10104 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
10105 void __user *arg, unsigned nr_args)
10106 __releases(ctx->uring_lock)
10107 __acquires(ctx->uring_lock)
10112 * We're inside the ring mutex, if the ref is already dying, then
10113 * someone else killed the ctx or is already going through
10114 * io_uring_register().
10116 if (percpu_ref_is_dying(&ctx->refs))
10119 if (ctx->restricted) {
10120 if (opcode >= IORING_REGISTER_LAST)
10122 opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
10123 if (!test_bit(opcode, ctx->restrictions.register_op))
10127 if (io_register_op_must_quiesce(opcode)) {
10128 ret = io_ctx_quiesce(ctx);
10134 case IORING_REGISTER_BUFFERS:
10135 ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
10137 case IORING_UNREGISTER_BUFFERS:
10139 if (arg || nr_args)
10141 ret = io_sqe_buffers_unregister(ctx);
10143 case IORING_REGISTER_FILES:
10144 ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
10146 case IORING_UNREGISTER_FILES:
10148 if (arg || nr_args)
10150 ret = io_sqe_files_unregister(ctx);
10152 case IORING_REGISTER_FILES_UPDATE:
10153 ret = io_register_files_update(ctx, arg, nr_args);
10155 case IORING_REGISTER_EVENTFD:
10156 case IORING_REGISTER_EVENTFD_ASYNC:
10160 ret = io_eventfd_register(ctx, arg);
10163 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
10164 ctx->eventfd_async = 1;
10166 ctx->eventfd_async = 0;
10168 case IORING_UNREGISTER_EVENTFD:
10170 if (arg || nr_args)
10172 ret = io_eventfd_unregister(ctx);
10174 case IORING_REGISTER_PROBE:
10176 if (!arg || nr_args > 256)
10178 ret = io_probe(ctx, arg, nr_args);
10180 case IORING_REGISTER_PERSONALITY:
10182 if (arg || nr_args)
10184 ret = io_register_personality(ctx);
10186 case IORING_UNREGISTER_PERSONALITY:
10190 ret = io_unregister_personality(ctx, nr_args);
10192 case IORING_REGISTER_ENABLE_RINGS:
10194 if (arg || nr_args)
10196 ret = io_register_enable_rings(ctx);
10198 case IORING_REGISTER_RESTRICTIONS:
10199 ret = io_register_restrictions(ctx, arg, nr_args);
10201 case IORING_REGISTER_FILES2:
10202 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE);
10204 case IORING_REGISTER_FILES_UPDATE2:
10205 ret = io_register_rsrc_update(ctx, arg, nr_args,
10208 case IORING_REGISTER_BUFFERS2:
10209 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER);
10211 case IORING_REGISTER_BUFFERS_UPDATE:
10212 ret = io_register_rsrc_update(ctx, arg, nr_args,
10213 IORING_RSRC_BUFFER);
10215 case IORING_REGISTER_IOWQ_AFF:
10217 if (!arg || !nr_args)
10219 ret = io_register_iowq_aff(ctx, arg, nr_args);
10221 case IORING_UNREGISTER_IOWQ_AFF:
10223 if (arg || nr_args)
10225 ret = io_unregister_iowq_aff(ctx);
10232 if (io_register_op_must_quiesce(opcode)) {
10233 /* bring the ctx back to life */
10234 percpu_ref_reinit(&ctx->refs);
10235 reinit_completion(&ctx->ref_comp);
10240 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
10241 void __user *, arg, unsigned int, nr_args)
10243 struct io_ring_ctx *ctx;
10252 if (f.file->f_op != &io_uring_fops)
10255 ctx = f.file->private_data;
10257 io_run_task_work();
10259 mutex_lock(&ctx->uring_lock);
10260 ret = __io_uring_register(ctx, opcode, arg, nr_args);
10261 mutex_unlock(&ctx->uring_lock);
10262 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
10263 ctx->cq_ev_fd != NULL, ret);
10269 static int __init io_uring_init(void)
10271 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
10272 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
10273 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
10276 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
10277 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
10278 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
10279 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
10280 BUILD_BUG_SQE_ELEM(1, __u8, flags);
10281 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
10282 BUILD_BUG_SQE_ELEM(4, __s32, fd);
10283 BUILD_BUG_SQE_ELEM(8, __u64, off);
10284 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
10285 BUILD_BUG_SQE_ELEM(16, __u64, addr);
10286 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
10287 BUILD_BUG_SQE_ELEM(24, __u32, len);
10288 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
10289 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
10290 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
10291 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
10292 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
10293 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
10294 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
10295 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
10296 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
10297 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
10298 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
10299 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
10300 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
10301 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
10302 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
10303 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
10304 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
10305 BUILD_BUG_SQE_ELEM(40, __u16, buf_group);
10306 BUILD_BUG_SQE_ELEM(42, __u16, personality);
10307 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
10309 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
10310 sizeof(struct io_uring_rsrc_update));
10311 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
10312 sizeof(struct io_uring_rsrc_update2));
10313 /* should fit into one byte */
10314 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
10316 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
10317 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
10319 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
10323 __initcall(io_uring_init);