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>
82 #define CREATE_TRACE_POINTS
83 #include <trace/events/io_uring.h>
85 #include <uapi/linux/io_uring.h>
90 #define IORING_MAX_ENTRIES 32768
91 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
94 * Shift of 9 is 512 entries, or exactly one page on 64-bit archs
96 #define IORING_FILE_TABLE_SHIFT 9
97 #define IORING_MAX_FILES_TABLE (1U << IORING_FILE_TABLE_SHIFT)
98 #define IORING_FILE_TABLE_MASK (IORING_MAX_FILES_TABLE - 1)
99 #define IORING_MAX_FIXED_FILES (64 * IORING_MAX_FILES_TABLE)
100 #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
101 IORING_REGISTER_LAST + IORING_OP_LAST)
103 #define IO_RSRC_TAG_TABLE_SHIFT 9
104 #define IO_RSRC_TAG_TABLE_MAX (1U << IO_RSRC_TAG_TABLE_SHIFT)
105 #define IO_RSRC_TAG_TABLE_MASK (IO_RSRC_TAG_TABLE_MAX - 1)
107 #define IORING_MAX_REG_BUFFERS (1U << 14)
109 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
110 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
113 #define IO_TCTX_REFS_CACHE_NR (1U << 10)
116 u32 head ____cacheline_aligned_in_smp;
117 u32 tail ____cacheline_aligned_in_smp;
121 * This data is shared with the application through the mmap at offsets
122 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
124 * The offsets to the member fields are published through struct
125 * io_sqring_offsets when calling io_uring_setup.
129 * Head and tail offsets into the ring; the offsets need to be
130 * masked to get valid indices.
132 * The kernel controls head of the sq ring and the tail of the cq ring,
133 * and the application controls tail of the sq ring and the head of the
136 struct io_uring sq, cq;
138 * Bitmasks to apply to head and tail offsets (constant, equals
141 u32 sq_ring_mask, cq_ring_mask;
142 /* Ring sizes (constant, power of 2) */
143 u32 sq_ring_entries, cq_ring_entries;
145 * Number of invalid entries dropped by the kernel due to
146 * invalid index stored in array
148 * Written by the kernel, shouldn't be modified by the
149 * application (i.e. get number of "new events" by comparing to
152 * After a new SQ head value was read by the application this
153 * counter includes all submissions that were dropped reaching
154 * the new SQ head (and possibly more).
160 * Written by the kernel, shouldn't be modified by the
163 * The application needs a full memory barrier before checking
164 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
170 * Written by the application, shouldn't be modified by the
175 * Number of completion events lost because the queue was full;
176 * this should be avoided by the application by making sure
177 * there are not more requests pending than there is space in
178 * the completion queue.
180 * Written by the kernel, shouldn't be modified by the
181 * application (i.e. get number of "new events" by comparing to
184 * As completion events come in out of order this counter is not
185 * ordered with any other data.
189 * Ring buffer of completion events.
191 * The kernel writes completion events fresh every time they are
192 * produced, so the application is allowed to modify pending
195 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
198 enum io_uring_cmd_flags {
199 IO_URING_F_NONBLOCK = 1,
200 IO_URING_F_COMPLETE_DEFER = 2,
203 struct io_mapped_ubuf {
206 unsigned int nr_bvecs;
207 unsigned long acct_pages;
208 struct bio_vec bvec[];
213 struct io_overflow_cqe {
214 struct io_uring_cqe cqe;
215 struct list_head list;
218 struct io_fixed_file {
219 /* file * with additional FFS_* flags */
220 unsigned long file_ptr;
224 struct list_head list;
229 struct io_mapped_ubuf *buf;
233 struct io_file_table {
234 /* two level table */
235 struct io_fixed_file **files;
238 struct io_rsrc_node {
239 struct percpu_ref refs;
240 struct list_head node;
241 struct list_head rsrc_list;
242 struct io_rsrc_data *rsrc_data;
243 struct llist_node llist;
247 typedef void (rsrc_put_fn)(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc);
249 struct io_rsrc_data {
250 struct io_ring_ctx *ctx;
256 struct completion done;
261 struct list_head list;
267 struct io_restriction {
268 DECLARE_BITMAP(register_op, IORING_REGISTER_LAST);
269 DECLARE_BITMAP(sqe_op, IORING_OP_LAST);
270 u8 sqe_flags_allowed;
271 u8 sqe_flags_required;
276 IO_SQ_THREAD_SHOULD_STOP = 0,
277 IO_SQ_THREAD_SHOULD_PARK,
282 atomic_t park_pending;
285 /* ctx's that are using this sqd */
286 struct list_head ctx_list;
288 struct task_struct *thread;
289 struct wait_queue_head wait;
291 unsigned sq_thread_idle;
297 struct completion exited;
300 #define IO_IOPOLL_BATCH 8
301 #define IO_COMPL_BATCH 32
302 #define IO_REQ_CACHE_SIZE 32
303 #define IO_REQ_ALLOC_BATCH 8
305 struct io_comp_state {
306 struct io_kiocb *reqs[IO_COMPL_BATCH];
308 /* inline/task_work completion list, under ->uring_lock */
309 struct list_head free_list;
312 struct io_submit_link {
313 struct io_kiocb *head;
314 struct io_kiocb *last;
317 struct io_submit_state {
318 struct blk_plug plug;
319 struct io_submit_link link;
322 * io_kiocb alloc cache
324 void *reqs[IO_REQ_CACHE_SIZE];
325 unsigned int free_reqs;
330 * Batch completion logic
332 struct io_comp_state comp;
335 * File reference cache
339 unsigned int file_refs;
340 unsigned int ios_left;
344 /* const or read-mostly hot data */
346 struct percpu_ref refs;
348 struct io_rings *rings;
350 unsigned int compat: 1;
351 unsigned int drain_next: 1;
352 unsigned int eventfd_async: 1;
353 unsigned int restricted: 1;
354 unsigned int off_timeout_used: 1;
355 unsigned int drain_active: 1;
356 } ____cacheline_aligned_in_smp;
358 /* submission data */
360 struct mutex uring_lock;
363 * Ring buffer of indices into array of io_uring_sqe, which is
364 * mmapped by the application using the IORING_OFF_SQES offset.
366 * This indirection could e.g. be used to assign fixed
367 * io_uring_sqe entries to operations and only submit them to
368 * the queue when needed.
370 * The kernel modifies neither the indices array nor the entries
374 struct io_uring_sqe *sq_sqes;
375 unsigned cached_sq_head;
377 struct list_head defer_list;
380 * Fixed resources fast path, should be accessed only under
381 * uring_lock, and updated through io_uring_register(2)
383 struct io_rsrc_node *rsrc_node;
384 struct io_file_table file_table;
385 unsigned nr_user_files;
386 unsigned nr_user_bufs;
387 struct io_mapped_ubuf **user_bufs;
389 struct io_submit_state submit_state;
390 struct list_head timeout_list;
391 struct list_head cq_overflow_list;
392 struct xarray io_buffers;
393 struct xarray personalities;
395 unsigned sq_thread_idle;
396 } ____cacheline_aligned_in_smp;
398 /* IRQ completion list, under ->completion_lock */
399 struct list_head locked_free_list;
400 unsigned int locked_free_nr;
402 const struct cred *sq_creds; /* cred used for __io_sq_thread() */
403 struct io_sq_data *sq_data; /* if using sq thread polling */
405 struct wait_queue_head sqo_sq_wait;
406 struct list_head sqd_list;
408 unsigned long check_cq_overflow;
411 unsigned cached_cq_tail;
413 struct eventfd_ctx *cq_ev_fd;
414 struct wait_queue_head poll_wait;
415 struct wait_queue_head cq_wait;
417 atomic_t cq_timeouts;
418 struct fasync_struct *cq_fasync;
419 unsigned cq_last_tm_flush;
420 } ____cacheline_aligned_in_smp;
423 spinlock_t completion_lock;
426 * ->iopoll_list is protected by the ctx->uring_lock for
427 * io_uring instances that don't use IORING_SETUP_SQPOLL.
428 * For SQPOLL, only the single threaded io_sq_thread() will
429 * manipulate the list, hence no extra locking is needed there.
431 struct list_head iopoll_list;
432 struct hlist_head *cancel_hash;
433 unsigned cancel_hash_bits;
434 bool poll_multi_file;
435 } ____cacheline_aligned_in_smp;
437 struct io_restriction restrictions;
439 /* slow path rsrc auxilary data, used by update/register */
441 struct io_rsrc_node *rsrc_backup_node;
442 struct io_mapped_ubuf *dummy_ubuf;
443 struct io_rsrc_data *file_data;
444 struct io_rsrc_data *buf_data;
446 struct delayed_work rsrc_put_work;
447 struct llist_head rsrc_put_llist;
448 struct list_head rsrc_ref_list;
449 spinlock_t rsrc_ref_lock;
452 /* Keep this last, we don't need it for the fast path */
454 #if defined(CONFIG_UNIX)
455 struct socket *ring_sock;
457 /* hashed buffered write serialization */
458 struct io_wq_hash *hash_map;
460 /* Only used for accounting purposes */
461 struct user_struct *user;
462 struct mm_struct *mm_account;
464 /* ctx exit and cancelation */
465 struct callback_head *exit_task_work;
466 struct work_struct exit_work;
467 struct list_head tctx_list;
468 struct completion ref_comp;
472 struct io_uring_task {
473 /* submission side */
476 struct wait_queue_head wait;
477 const struct io_ring_ctx *last;
479 struct percpu_counter inflight;
480 atomic_t inflight_tracked;
483 spinlock_t task_lock;
484 struct io_wq_work_list task_list;
485 unsigned long task_state;
486 struct callback_head task_work;
490 * First field must be the file pointer in all the
491 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
493 struct io_poll_iocb {
495 struct wait_queue_head *head;
499 struct wait_queue_entry wait;
502 struct io_poll_update {
508 bool update_user_data;
516 struct io_timeout_data {
517 struct io_kiocb *req;
518 struct hrtimer timer;
519 struct timespec64 ts;
520 enum hrtimer_mode mode;
525 struct sockaddr __user *addr;
526 int __user *addr_len;
528 unsigned long nofile;
548 struct list_head list;
549 /* head of the link, used by linked timeouts only */
550 struct io_kiocb *head;
553 struct io_timeout_rem {
558 struct timespec64 ts;
563 /* NOTE: kiocb has the file as the first member, so don't do it here */
571 struct sockaddr __user *addr;
578 struct compat_msghdr __user *umsg_compat;
579 struct user_msghdr __user *umsg;
585 struct io_buffer *kbuf;
591 struct filename *filename;
593 unsigned long nofile;
596 struct io_rsrc_update {
622 struct epoll_event event;
626 struct file *file_out;
627 struct file *file_in;
634 struct io_provide_buf {
648 const char __user *filename;
649 struct statx __user *buffer;
661 struct filename *oldpath;
662 struct filename *newpath;
670 struct filename *filename;
673 struct io_completion {
675 struct list_head list;
679 struct io_async_connect {
680 struct sockaddr_storage address;
683 struct io_async_msghdr {
684 struct iovec fast_iov[UIO_FASTIOV];
685 /* points to an allocated iov, if NULL we use fast_iov instead */
686 struct iovec *free_iov;
687 struct sockaddr __user *uaddr;
689 struct sockaddr_storage addr;
693 struct iovec fast_iov[UIO_FASTIOV];
694 const struct iovec *free_iovec;
695 struct iov_iter iter;
697 struct wait_page_queue wpq;
701 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
702 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
703 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
704 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
705 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
706 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
708 /* first byte is taken by user flags, shift it to not overlap */
713 REQ_F_LINK_TIMEOUT_BIT,
714 REQ_F_NEED_CLEANUP_BIT,
716 REQ_F_BUFFER_SELECTED_BIT,
717 REQ_F_LTIMEOUT_ACTIVE_BIT,
718 REQ_F_COMPLETE_INLINE_BIT,
720 REQ_F_DONT_REISSUE_BIT,
721 /* keep async read/write and isreg together and in order */
722 REQ_F_ASYNC_READ_BIT,
723 REQ_F_ASYNC_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_ASYNC_READ = BIT(REQ_F_ASYNC_READ_BIT),
770 /* supports async writes */
771 REQ_F_ASYNC_WRITE = BIT(REQ_F_ASYNC_WRITE_BIT),
773 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
777 struct io_poll_iocb poll;
778 struct io_poll_iocb *double_poll;
781 struct io_task_work {
782 struct io_wq_work_node node;
783 task_work_func_t func;
787 IORING_RSRC_FILE = 0,
788 IORING_RSRC_BUFFER = 1,
792 * NOTE! Each of the iocb union members has the file pointer
793 * as the first entry in their struct definition. So you can
794 * access the file pointer through any of the sub-structs,
795 * or directly as just 'ki_filp' in this struct.
801 struct io_poll_iocb poll;
802 struct io_poll_update poll_update;
803 struct io_accept accept;
805 struct io_cancel cancel;
806 struct io_timeout timeout;
807 struct io_timeout_rem timeout_rem;
808 struct io_connect connect;
809 struct io_sr_msg sr_msg;
811 struct io_close close;
812 struct io_rsrc_update rsrc_update;
813 struct io_fadvise fadvise;
814 struct io_madvise madvise;
815 struct io_epoll epoll;
816 struct io_splice splice;
817 struct io_provide_buf pbuf;
818 struct io_statx statx;
819 struct io_shutdown shutdown;
820 struct io_rename rename;
821 struct io_unlink unlink;
822 /* use only after cleaning per-op data, see io_clean_op() */
823 struct io_completion compl;
826 /* opcode allocated if it needs to store data for async defer */
829 /* polled IO has completed */
835 struct io_ring_ctx *ctx;
838 struct task_struct *task;
841 struct io_kiocb *link;
842 struct percpu_ref *fixed_rsrc_refs;
844 /* used with ctx->iopoll_list with reads/writes */
845 struct list_head inflight_entry;
847 struct io_task_work io_task_work;
848 struct callback_head task_work;
850 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
851 struct hlist_node hash_node;
852 struct async_poll *apoll;
853 struct io_wq_work work;
854 /* store used ubuf, so we can prevent reloading */
855 struct io_mapped_ubuf *imu;
858 struct io_tctx_node {
859 struct list_head ctx_node;
860 struct task_struct *task;
861 struct io_ring_ctx *ctx;
864 struct io_defer_entry {
865 struct list_head list;
866 struct io_kiocb *req;
871 /* needs req->file assigned */
872 unsigned needs_file : 1;
873 /* hash wq insertion if file is a regular file */
874 unsigned hash_reg_file : 1;
875 /* unbound wq insertion if file is a non-regular file */
876 unsigned unbound_nonreg_file : 1;
877 /* opcode is not supported by this kernel */
878 unsigned not_supported : 1;
879 /* set if opcode supports polled "wait" */
881 unsigned pollout : 1;
882 /* op supports buffer selection */
883 unsigned buffer_select : 1;
884 /* do prep async if is going to be punted */
885 unsigned needs_async_setup : 1;
886 /* should block plug */
888 /* size of async data needed, if any */
889 unsigned short async_size;
892 static const struct io_op_def io_op_defs[] = {
893 [IORING_OP_NOP] = {},
894 [IORING_OP_READV] = {
896 .unbound_nonreg_file = 1,
899 .needs_async_setup = 1,
901 .async_size = sizeof(struct io_async_rw),
903 [IORING_OP_WRITEV] = {
906 .unbound_nonreg_file = 1,
908 .needs_async_setup = 1,
910 .async_size = sizeof(struct io_async_rw),
912 [IORING_OP_FSYNC] = {
915 [IORING_OP_READ_FIXED] = {
917 .unbound_nonreg_file = 1,
920 .async_size = sizeof(struct io_async_rw),
922 [IORING_OP_WRITE_FIXED] = {
925 .unbound_nonreg_file = 1,
928 .async_size = sizeof(struct io_async_rw),
930 [IORING_OP_POLL_ADD] = {
932 .unbound_nonreg_file = 1,
934 [IORING_OP_POLL_REMOVE] = {},
935 [IORING_OP_SYNC_FILE_RANGE] = {
938 [IORING_OP_SENDMSG] = {
940 .unbound_nonreg_file = 1,
942 .needs_async_setup = 1,
943 .async_size = sizeof(struct io_async_msghdr),
945 [IORING_OP_RECVMSG] = {
947 .unbound_nonreg_file = 1,
950 .needs_async_setup = 1,
951 .async_size = sizeof(struct io_async_msghdr),
953 [IORING_OP_TIMEOUT] = {
954 .async_size = sizeof(struct io_timeout_data),
956 [IORING_OP_TIMEOUT_REMOVE] = {
957 /* used by timeout updates' prep() */
959 [IORING_OP_ACCEPT] = {
961 .unbound_nonreg_file = 1,
964 [IORING_OP_ASYNC_CANCEL] = {},
965 [IORING_OP_LINK_TIMEOUT] = {
966 .async_size = sizeof(struct io_timeout_data),
968 [IORING_OP_CONNECT] = {
970 .unbound_nonreg_file = 1,
972 .needs_async_setup = 1,
973 .async_size = sizeof(struct io_async_connect),
975 [IORING_OP_FALLOCATE] = {
978 [IORING_OP_OPENAT] = {},
979 [IORING_OP_CLOSE] = {},
980 [IORING_OP_FILES_UPDATE] = {},
981 [IORING_OP_STATX] = {},
984 .unbound_nonreg_file = 1,
988 .async_size = sizeof(struct io_async_rw),
990 [IORING_OP_WRITE] = {
992 .unbound_nonreg_file = 1,
995 .async_size = sizeof(struct io_async_rw),
997 [IORING_OP_FADVISE] = {
1000 [IORING_OP_MADVISE] = {},
1001 [IORING_OP_SEND] = {
1003 .unbound_nonreg_file = 1,
1006 [IORING_OP_RECV] = {
1008 .unbound_nonreg_file = 1,
1012 [IORING_OP_OPENAT2] = {
1014 [IORING_OP_EPOLL_CTL] = {
1015 .unbound_nonreg_file = 1,
1017 [IORING_OP_SPLICE] = {
1020 .unbound_nonreg_file = 1,
1022 [IORING_OP_PROVIDE_BUFFERS] = {},
1023 [IORING_OP_REMOVE_BUFFERS] = {},
1027 .unbound_nonreg_file = 1,
1029 [IORING_OP_SHUTDOWN] = {
1032 [IORING_OP_RENAMEAT] = {},
1033 [IORING_OP_UNLINKAT] = {},
1036 static bool io_disarm_next(struct io_kiocb *req);
1037 static void io_uring_del_tctx_node(unsigned long index);
1038 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
1039 struct task_struct *task,
1041 static void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd);
1042 static struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx);
1044 static bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1045 long res, unsigned int cflags);
1046 static void io_put_req(struct io_kiocb *req);
1047 static void io_put_req_deferred(struct io_kiocb *req, int nr);
1048 static void io_dismantle_req(struct io_kiocb *req);
1049 static void io_put_task(struct task_struct *task, int nr);
1050 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
1051 static void io_queue_linked_timeout(struct io_kiocb *req);
1052 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
1053 struct io_uring_rsrc_update2 *up,
1055 static void io_clean_op(struct io_kiocb *req);
1056 static struct file *io_file_get(struct io_submit_state *state,
1057 struct io_kiocb *req, int fd, bool fixed);
1058 static void __io_queue_sqe(struct io_kiocb *req);
1059 static void io_rsrc_put_work(struct work_struct *work);
1061 static void io_req_task_queue(struct io_kiocb *req);
1062 static void io_submit_flush_completions(struct io_ring_ctx *ctx);
1063 static bool io_poll_remove_waitqs(struct io_kiocb *req);
1064 static int io_req_prep_async(struct io_kiocb *req);
1066 static struct kmem_cache *req_cachep;
1068 static const struct file_operations io_uring_fops;
1070 struct sock *io_uring_get_socket(struct file *file)
1072 #if defined(CONFIG_UNIX)
1073 if (file->f_op == &io_uring_fops) {
1074 struct io_ring_ctx *ctx = file->private_data;
1076 return ctx->ring_sock->sk;
1081 EXPORT_SYMBOL(io_uring_get_socket);
1083 #define io_for_each_link(pos, head) \
1084 for (pos = (head); pos; pos = pos->link)
1086 static inline void io_req_set_rsrc_node(struct io_kiocb *req)
1088 struct io_ring_ctx *ctx = req->ctx;
1090 if (!req->fixed_rsrc_refs) {
1091 req->fixed_rsrc_refs = &ctx->rsrc_node->refs;
1092 percpu_ref_get(req->fixed_rsrc_refs);
1096 static void io_refs_resurrect(struct percpu_ref *ref, struct completion *compl)
1098 bool got = percpu_ref_tryget(ref);
1100 /* already at zero, wait for ->release() */
1102 wait_for_completion(compl);
1103 percpu_ref_resurrect(ref);
1105 percpu_ref_put(ref);
1108 static bool io_match_task(struct io_kiocb *head, struct task_struct *task,
1111 struct io_kiocb *req;
1113 if (task && head->task != task)
1118 io_for_each_link(req, head) {
1119 if (req->flags & REQ_F_INFLIGHT)
1125 static inline void req_set_fail(struct io_kiocb *req)
1127 req->flags |= REQ_F_FAIL;
1130 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1132 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1134 complete(&ctx->ref_comp);
1137 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1139 return !req->timeout.off;
1142 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1144 struct io_ring_ctx *ctx;
1147 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1152 * Use 5 bits less than the max cq entries, that should give us around
1153 * 32 entries per hash list if totally full and uniformly spread.
1155 hash_bits = ilog2(p->cq_entries);
1159 ctx->cancel_hash_bits = hash_bits;
1160 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1162 if (!ctx->cancel_hash)
1164 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1166 ctx->dummy_ubuf = kzalloc(sizeof(*ctx->dummy_ubuf), GFP_KERNEL);
1167 if (!ctx->dummy_ubuf)
1169 /* set invalid range, so io_import_fixed() fails meeting it */
1170 ctx->dummy_ubuf->ubuf = -1UL;
1172 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1173 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1176 ctx->flags = p->flags;
1177 init_waitqueue_head(&ctx->sqo_sq_wait);
1178 INIT_LIST_HEAD(&ctx->sqd_list);
1179 init_waitqueue_head(&ctx->poll_wait);
1180 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1181 init_completion(&ctx->ref_comp);
1182 xa_init_flags(&ctx->io_buffers, XA_FLAGS_ALLOC1);
1183 xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
1184 mutex_init(&ctx->uring_lock);
1185 init_waitqueue_head(&ctx->cq_wait);
1186 spin_lock_init(&ctx->completion_lock);
1187 INIT_LIST_HEAD(&ctx->iopoll_list);
1188 INIT_LIST_HEAD(&ctx->defer_list);
1189 INIT_LIST_HEAD(&ctx->timeout_list);
1190 spin_lock_init(&ctx->rsrc_ref_lock);
1191 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
1192 INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
1193 init_llist_head(&ctx->rsrc_put_llist);
1194 INIT_LIST_HEAD(&ctx->tctx_list);
1195 INIT_LIST_HEAD(&ctx->submit_state.comp.free_list);
1196 INIT_LIST_HEAD(&ctx->locked_free_list);
1199 kfree(ctx->dummy_ubuf);
1200 kfree(ctx->cancel_hash);
1205 static void io_account_cq_overflow(struct io_ring_ctx *ctx)
1207 struct io_rings *r = ctx->rings;
1209 WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1);
1213 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1215 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1216 struct io_ring_ctx *ctx = req->ctx;
1218 return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail;
1224 static void io_req_track_inflight(struct io_kiocb *req)
1226 if (!(req->flags & REQ_F_INFLIGHT)) {
1227 req->flags |= REQ_F_INFLIGHT;
1228 atomic_inc(¤t->io_uring->inflight_tracked);
1232 static void io_prep_async_work(struct io_kiocb *req)
1234 const struct io_op_def *def = &io_op_defs[req->opcode];
1235 struct io_ring_ctx *ctx = req->ctx;
1237 if (!req->work.creds)
1238 req->work.creds = get_current_cred();
1240 req->work.list.next = NULL;
1241 req->work.flags = 0;
1242 if (req->flags & REQ_F_FORCE_ASYNC)
1243 req->work.flags |= IO_WQ_WORK_CONCURRENT;
1245 if (req->flags & REQ_F_ISREG) {
1246 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1247 io_wq_hash_work(&req->work, file_inode(req->file));
1248 } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
1249 if (def->unbound_nonreg_file)
1250 req->work.flags |= IO_WQ_WORK_UNBOUND;
1253 switch (req->opcode) {
1254 case IORING_OP_SPLICE:
1256 if (!S_ISREG(file_inode(req->splice.file_in)->i_mode))
1257 req->work.flags |= IO_WQ_WORK_UNBOUND;
1262 static void io_prep_async_link(struct io_kiocb *req)
1264 struct io_kiocb *cur;
1266 io_for_each_link(cur, req)
1267 io_prep_async_work(cur);
1270 static void io_queue_async_work(struct io_kiocb *req)
1272 struct io_ring_ctx *ctx = req->ctx;
1273 struct io_kiocb *link = io_prep_linked_timeout(req);
1274 struct io_uring_task *tctx = req->task->io_uring;
1277 BUG_ON(!tctx->io_wq);
1279 /* init ->work of the whole link before punting */
1280 io_prep_async_link(req);
1281 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1282 &req->work, req->flags);
1283 io_wq_enqueue(tctx->io_wq, &req->work);
1285 io_queue_linked_timeout(link);
1288 static void io_kill_timeout(struct io_kiocb *req, int status)
1289 __must_hold(&req->ctx->completion_lock)
1291 struct io_timeout_data *io = req->async_data;
1293 if (hrtimer_try_to_cancel(&io->timer) != -1) {
1294 atomic_set(&req->ctx->cq_timeouts,
1295 atomic_read(&req->ctx->cq_timeouts) + 1);
1296 list_del_init(&req->timeout.list);
1297 io_cqring_fill_event(req->ctx, req->user_data, status, 0);
1298 io_put_req_deferred(req, 1);
1302 static void io_queue_deferred(struct io_ring_ctx *ctx)
1304 while (!list_empty(&ctx->defer_list)) {
1305 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1306 struct io_defer_entry, list);
1308 if (req_need_defer(de->req, de->seq))
1310 list_del_init(&de->list);
1311 io_req_task_queue(de->req);
1316 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1318 u32 seq = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
1320 while (!list_empty(&ctx->timeout_list)) {
1321 u32 events_needed, events_got;
1322 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1323 struct io_kiocb, timeout.list);
1325 if (io_is_timeout_noseq(req))
1329 * Since seq can easily wrap around over time, subtract
1330 * the last seq at which timeouts were flushed before comparing.
1331 * Assuming not more than 2^31-1 events have happened since,
1332 * these subtractions won't have wrapped, so we can check if
1333 * target is in [last_seq, current_seq] by comparing the two.
1335 events_needed = req->timeout.target_seq - ctx->cq_last_tm_flush;
1336 events_got = seq - ctx->cq_last_tm_flush;
1337 if (events_got < events_needed)
1340 list_del_init(&req->timeout.list);
1341 io_kill_timeout(req, 0);
1343 ctx->cq_last_tm_flush = seq;
1346 static void __io_commit_cqring_flush(struct io_ring_ctx *ctx)
1348 if (ctx->off_timeout_used)
1349 io_flush_timeouts(ctx);
1350 if (ctx->drain_active)
1351 io_queue_deferred(ctx);
1354 static inline void io_commit_cqring(struct io_ring_ctx *ctx)
1356 if (unlikely(ctx->off_timeout_used || ctx->drain_active))
1357 __io_commit_cqring_flush(ctx);
1358 /* order cqe stores with ring update */
1359 smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
1362 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1364 struct io_rings *r = ctx->rings;
1366 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == ctx->sq_entries;
1369 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
1371 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
1374 static inline struct io_uring_cqe *io_get_cqe(struct io_ring_ctx *ctx)
1376 struct io_rings *rings = ctx->rings;
1377 unsigned tail, mask = ctx->cq_entries - 1;
1380 * writes to the cq entry need to come after reading head; the
1381 * control dependency is enough as we're using WRITE_ONCE to
1384 if (__io_cqring_events(ctx) == ctx->cq_entries)
1387 tail = ctx->cached_cq_tail++;
1388 return &rings->cqes[tail & mask];
1391 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1393 if (likely(!ctx->cq_ev_fd))
1395 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1397 return !ctx->eventfd_async || io_wq_current_is_worker();
1400 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1402 /* see waitqueue_active() comment */
1405 if (waitqueue_active(&ctx->cq_wait))
1406 wake_up(&ctx->cq_wait);
1407 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1408 wake_up(&ctx->sq_data->wait);
1409 if (io_should_trigger_evfd(ctx))
1410 eventfd_signal(ctx->cq_ev_fd, 1);
1411 if (waitqueue_active(&ctx->poll_wait)) {
1412 wake_up_interruptible(&ctx->poll_wait);
1413 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1417 static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx)
1419 /* see waitqueue_active() comment */
1422 if (ctx->flags & IORING_SETUP_SQPOLL) {
1423 if (waitqueue_active(&ctx->cq_wait))
1424 wake_up(&ctx->cq_wait);
1426 if (io_should_trigger_evfd(ctx))
1427 eventfd_signal(ctx->cq_ev_fd, 1);
1428 if (waitqueue_active(&ctx->poll_wait)) {
1429 wake_up_interruptible(&ctx->poll_wait);
1430 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1434 /* Returns true if there are no backlogged entries after the flush */
1435 static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1437 unsigned long flags;
1438 bool all_flushed, posted;
1440 if (!force && __io_cqring_events(ctx) == ctx->cq_entries)
1444 spin_lock_irqsave(&ctx->completion_lock, flags);
1445 while (!list_empty(&ctx->cq_overflow_list)) {
1446 struct io_uring_cqe *cqe = io_get_cqe(ctx);
1447 struct io_overflow_cqe *ocqe;
1451 ocqe = list_first_entry(&ctx->cq_overflow_list,
1452 struct io_overflow_cqe, list);
1454 memcpy(cqe, &ocqe->cqe, sizeof(*cqe));
1456 io_account_cq_overflow(ctx);
1459 list_del(&ocqe->list);
1463 all_flushed = list_empty(&ctx->cq_overflow_list);
1465 clear_bit(0, &ctx->check_cq_overflow);
1466 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
1470 io_commit_cqring(ctx);
1471 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1473 io_cqring_ev_posted(ctx);
1477 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1481 if (test_bit(0, &ctx->check_cq_overflow)) {
1482 /* iopoll syncs against uring_lock, not completion_lock */
1483 if (ctx->flags & IORING_SETUP_IOPOLL)
1484 mutex_lock(&ctx->uring_lock);
1485 ret = __io_cqring_overflow_flush(ctx, force);
1486 if (ctx->flags & IORING_SETUP_IOPOLL)
1487 mutex_unlock(&ctx->uring_lock);
1494 * Shamelessly stolen from the mm implementation of page reference checking,
1495 * see commit f958d7b528b1 for details.
1497 #define req_ref_zero_or_close_to_overflow(req) \
1498 ((unsigned int) atomic_read(&(req->refs)) + 127u <= 127u)
1500 static inline bool req_ref_inc_not_zero(struct io_kiocb *req)
1502 return atomic_inc_not_zero(&req->refs);
1505 static inline bool req_ref_sub_and_test(struct io_kiocb *req, int refs)
1507 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1508 return atomic_sub_and_test(refs, &req->refs);
1511 static inline bool req_ref_put_and_test(struct io_kiocb *req)
1513 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1514 return atomic_dec_and_test(&req->refs);
1517 static inline void req_ref_put(struct io_kiocb *req)
1519 WARN_ON_ONCE(req_ref_put_and_test(req));
1522 static inline void req_ref_get(struct io_kiocb *req)
1524 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1525 atomic_inc(&req->refs);
1528 static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data,
1529 long res, unsigned int cflags)
1531 struct io_overflow_cqe *ocqe;
1533 ocqe = kmalloc(sizeof(*ocqe), GFP_ATOMIC | __GFP_ACCOUNT);
1536 * If we're in ring overflow flush mode, or in task cancel mode,
1537 * or cannot allocate an overflow entry, then we need to drop it
1540 io_account_cq_overflow(ctx);
1543 if (list_empty(&ctx->cq_overflow_list)) {
1544 set_bit(0, &ctx->check_cq_overflow);
1545 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
1547 ocqe->cqe.user_data = user_data;
1548 ocqe->cqe.res = res;
1549 ocqe->cqe.flags = cflags;
1550 list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
1554 static inline bool __io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1555 long res, unsigned int cflags)
1557 struct io_uring_cqe *cqe;
1559 trace_io_uring_complete(ctx, user_data, res, cflags);
1562 * If we can't get a cq entry, userspace overflowed the
1563 * submission (by quite a lot). Increment the overflow count in
1566 cqe = io_get_cqe(ctx);
1568 WRITE_ONCE(cqe->user_data, user_data);
1569 WRITE_ONCE(cqe->res, res);
1570 WRITE_ONCE(cqe->flags, cflags);
1573 return io_cqring_event_overflow(ctx, user_data, res, cflags);
1576 /* not as hot to bloat with inlining */
1577 static noinline bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1578 long res, unsigned int cflags)
1580 return __io_cqring_fill_event(ctx, user_data, res, cflags);
1583 static void io_req_complete_post(struct io_kiocb *req, long res,
1584 unsigned int cflags)
1586 struct io_ring_ctx *ctx = req->ctx;
1587 unsigned long flags;
1589 spin_lock_irqsave(&ctx->completion_lock, flags);
1590 __io_cqring_fill_event(ctx, req->user_data, res, cflags);
1592 * If we're the last reference to this request, add to our locked
1595 if (req_ref_put_and_test(req)) {
1596 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
1597 if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL))
1598 io_disarm_next(req);
1600 io_req_task_queue(req->link);
1604 io_dismantle_req(req);
1605 io_put_task(req->task, 1);
1606 list_add(&req->compl.list, &ctx->locked_free_list);
1607 ctx->locked_free_nr++;
1609 if (!percpu_ref_tryget(&ctx->refs))
1612 io_commit_cqring(ctx);
1613 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1616 io_cqring_ev_posted(ctx);
1617 percpu_ref_put(&ctx->refs);
1621 static inline bool io_req_needs_clean(struct io_kiocb *req)
1623 return req->flags & (REQ_F_BUFFER_SELECTED | REQ_F_NEED_CLEANUP |
1624 REQ_F_POLLED | REQ_F_INFLIGHT);
1627 static void io_req_complete_state(struct io_kiocb *req, long res,
1628 unsigned int cflags)
1630 if (io_req_needs_clean(req))
1633 req->compl.cflags = cflags;
1634 req->flags |= REQ_F_COMPLETE_INLINE;
1637 static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags,
1638 long res, unsigned cflags)
1640 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1641 io_req_complete_state(req, res, cflags);
1643 io_req_complete_post(req, res, cflags);
1646 static inline void io_req_complete(struct io_kiocb *req, long res)
1648 __io_req_complete(req, 0, res, 0);
1651 static void io_req_complete_failed(struct io_kiocb *req, long res)
1655 io_req_complete_post(req, res, 0);
1658 static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
1659 struct io_comp_state *cs)
1661 spin_lock_irq(&ctx->completion_lock);
1662 list_splice_init(&ctx->locked_free_list, &cs->free_list);
1663 ctx->locked_free_nr = 0;
1664 spin_unlock_irq(&ctx->completion_lock);
1667 /* Returns true IFF there are requests in the cache */
1668 static bool io_flush_cached_reqs(struct io_ring_ctx *ctx)
1670 struct io_submit_state *state = &ctx->submit_state;
1671 struct io_comp_state *cs = &state->comp;
1675 * If we have more than a batch's worth of requests in our IRQ side
1676 * locked cache, grab the lock and move them over to our submission
1679 if (READ_ONCE(ctx->locked_free_nr) > IO_COMPL_BATCH)
1680 io_flush_cached_locked_reqs(ctx, cs);
1682 nr = state->free_reqs;
1683 while (!list_empty(&cs->free_list)) {
1684 struct io_kiocb *req = list_first_entry(&cs->free_list,
1685 struct io_kiocb, compl.list);
1687 list_del(&req->compl.list);
1688 state->reqs[nr++] = req;
1689 if (nr == ARRAY_SIZE(state->reqs))
1693 state->free_reqs = nr;
1697 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
1699 struct io_submit_state *state = &ctx->submit_state;
1701 BUILD_BUG_ON(IO_REQ_ALLOC_BATCH > ARRAY_SIZE(state->reqs));
1703 if (!state->free_reqs) {
1704 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1707 if (io_flush_cached_reqs(ctx))
1710 ret = kmem_cache_alloc_bulk(req_cachep, gfp, IO_REQ_ALLOC_BATCH,
1714 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1715 * retry single alloc to be on the safe side.
1717 if (unlikely(ret <= 0)) {
1718 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1719 if (!state->reqs[0])
1723 state->free_reqs = ret;
1727 return state->reqs[state->free_reqs];
1730 static inline void io_put_file(struct file *file)
1736 static void io_dismantle_req(struct io_kiocb *req)
1738 unsigned int flags = req->flags;
1740 if (io_req_needs_clean(req))
1742 if (!(flags & REQ_F_FIXED_FILE))
1743 io_put_file(req->file);
1744 if (req->fixed_rsrc_refs)
1745 percpu_ref_put(req->fixed_rsrc_refs);
1746 if (req->async_data)
1747 kfree(req->async_data);
1748 if (req->work.creds) {
1749 put_cred(req->work.creds);
1750 req->work.creds = NULL;
1754 /* must to be called somewhat shortly after putting a request */
1755 static inline void io_put_task(struct task_struct *task, int nr)
1757 struct io_uring_task *tctx = task->io_uring;
1759 percpu_counter_sub(&tctx->inflight, nr);
1760 if (unlikely(atomic_read(&tctx->in_idle)))
1761 wake_up(&tctx->wait);
1762 put_task_struct_many(task, nr);
1765 static void __io_free_req(struct io_kiocb *req)
1767 struct io_ring_ctx *ctx = req->ctx;
1769 io_dismantle_req(req);
1770 io_put_task(req->task, 1);
1772 kmem_cache_free(req_cachep, req);
1773 percpu_ref_put(&ctx->refs);
1776 static inline void io_remove_next_linked(struct io_kiocb *req)
1778 struct io_kiocb *nxt = req->link;
1780 req->link = nxt->link;
1784 static bool io_kill_linked_timeout(struct io_kiocb *req)
1785 __must_hold(&req->ctx->completion_lock)
1787 struct io_kiocb *link = req->link;
1790 * Can happen if a linked timeout fired and link had been like
1791 * req -> link t-out -> link t-out [-> ...]
1793 if (link && (link->flags & REQ_F_LTIMEOUT_ACTIVE)) {
1794 struct io_timeout_data *io = link->async_data;
1796 io_remove_next_linked(req);
1797 link->timeout.head = NULL;
1798 if (hrtimer_try_to_cancel(&io->timer) != -1) {
1799 io_cqring_fill_event(link->ctx, link->user_data,
1801 io_put_req_deferred(link, 1);
1808 static void io_fail_links(struct io_kiocb *req)
1809 __must_hold(&req->ctx->completion_lock)
1811 struct io_kiocb *nxt, *link = req->link;
1818 trace_io_uring_fail_link(req, link);
1819 io_cqring_fill_event(link->ctx, link->user_data, -ECANCELED, 0);
1820 io_put_req_deferred(link, 2);
1825 static bool io_disarm_next(struct io_kiocb *req)
1826 __must_hold(&req->ctx->completion_lock)
1828 bool posted = false;
1830 if (likely(req->flags & REQ_F_LINK_TIMEOUT))
1831 posted = io_kill_linked_timeout(req);
1832 if (unlikely((req->flags & REQ_F_FAIL) &&
1833 !(req->flags & REQ_F_HARDLINK))) {
1834 posted |= (req->link != NULL);
1840 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1842 struct io_kiocb *nxt;
1845 * If LINK is set, we have dependent requests in this chain. If we
1846 * didn't fail this request, queue the first one up, moving any other
1847 * dependencies to the next request. In case of failure, fail the rest
1850 if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL)) {
1851 struct io_ring_ctx *ctx = req->ctx;
1852 unsigned long flags;
1855 spin_lock_irqsave(&ctx->completion_lock, flags);
1856 posted = io_disarm_next(req);
1858 io_commit_cqring(req->ctx);
1859 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1861 io_cqring_ev_posted(ctx);
1868 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1870 if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK))))
1872 return __io_req_find_next(req);
1875 static void ctx_flush_and_put(struct io_ring_ctx *ctx)
1879 if (ctx->submit_state.comp.nr) {
1880 mutex_lock(&ctx->uring_lock);
1881 io_submit_flush_completions(ctx);
1882 mutex_unlock(&ctx->uring_lock);
1884 percpu_ref_put(&ctx->refs);
1887 static bool __tctx_task_work(struct io_uring_task *tctx)
1889 struct io_ring_ctx *ctx = NULL;
1890 struct io_wq_work_list list;
1891 struct io_wq_work_node *node;
1893 if (wq_list_empty(&tctx->task_list))
1896 spin_lock_irq(&tctx->task_lock);
1897 list = tctx->task_list;
1898 INIT_WQ_LIST(&tctx->task_list);
1899 spin_unlock_irq(&tctx->task_lock);
1903 struct io_wq_work_node *next = node->next;
1904 struct io_kiocb *req;
1906 req = container_of(node, struct io_kiocb, io_task_work.node);
1907 if (req->ctx != ctx) {
1908 ctx_flush_and_put(ctx);
1910 percpu_ref_get(&ctx->refs);
1913 req->task_work.func(&req->task_work);
1917 ctx_flush_and_put(ctx);
1918 return list.first != NULL;
1921 static void tctx_task_work(struct callback_head *cb)
1923 struct io_uring_task *tctx = container_of(cb, struct io_uring_task, task_work);
1925 clear_bit(0, &tctx->task_state);
1927 while (__tctx_task_work(tctx))
1931 static int io_req_task_work_add(struct io_kiocb *req)
1933 struct task_struct *tsk = req->task;
1934 struct io_uring_task *tctx = tsk->io_uring;
1935 enum task_work_notify_mode notify;
1936 struct io_wq_work_node *node, *prev;
1937 unsigned long flags;
1940 if (unlikely(tsk->flags & PF_EXITING))
1943 WARN_ON_ONCE(!tctx);
1945 spin_lock_irqsave(&tctx->task_lock, flags);
1946 wq_list_add_tail(&req->io_task_work.node, &tctx->task_list);
1947 spin_unlock_irqrestore(&tctx->task_lock, flags);
1949 /* task_work already pending, we're done */
1950 if (test_bit(0, &tctx->task_state) ||
1951 test_and_set_bit(0, &tctx->task_state))
1955 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
1956 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
1957 * processing task_work. There's no reliable way to tell if TWA_RESUME
1960 notify = (req->ctx->flags & IORING_SETUP_SQPOLL) ? TWA_NONE : TWA_SIGNAL;
1962 if (!task_work_add(tsk, &tctx->task_work, notify)) {
1963 wake_up_process(tsk);
1968 * Slow path - we failed, find and delete work. if the work is not
1969 * in the list, it got run and we're fine.
1971 spin_lock_irqsave(&tctx->task_lock, flags);
1972 wq_list_for_each(node, prev, &tctx->task_list) {
1973 if (&req->io_task_work.node == node) {
1974 wq_list_del(&tctx->task_list, node, prev);
1979 spin_unlock_irqrestore(&tctx->task_lock, flags);
1980 clear_bit(0, &tctx->task_state);
1984 static bool io_run_task_work_head(struct callback_head **work_head)
1986 struct callback_head *work, *next;
1987 bool executed = false;
1990 work = xchg(work_head, NULL);
2006 static void io_task_work_add_head(struct callback_head **work_head,
2007 struct callback_head *task_work)
2009 struct callback_head *head;
2012 head = READ_ONCE(*work_head);
2013 task_work->next = head;
2014 } while (cmpxchg(work_head, head, task_work) != head);
2017 static void io_req_task_work_add_fallback(struct io_kiocb *req,
2018 task_work_func_t cb)
2020 init_task_work(&req->task_work, cb);
2021 io_task_work_add_head(&req->ctx->exit_task_work, &req->task_work);
2024 static void io_req_task_cancel(struct callback_head *cb)
2026 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2027 struct io_ring_ctx *ctx = req->ctx;
2029 /* ctx is guaranteed to stay alive while we hold uring_lock */
2030 mutex_lock(&ctx->uring_lock);
2031 io_req_complete_failed(req, req->result);
2032 mutex_unlock(&ctx->uring_lock);
2035 static void __io_req_task_submit(struct io_kiocb *req)
2037 struct io_ring_ctx *ctx = req->ctx;
2039 /* ctx stays valid until unlock, even if we drop all ours ctx->refs */
2040 mutex_lock(&ctx->uring_lock);
2041 if (!(current->flags & PF_EXITING) && !current->in_execve)
2042 __io_queue_sqe(req);
2044 io_req_complete_failed(req, -EFAULT);
2045 mutex_unlock(&ctx->uring_lock);
2048 static void io_req_task_submit(struct callback_head *cb)
2050 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2052 __io_req_task_submit(req);
2055 static void io_req_task_queue_fail(struct io_kiocb *req, int ret)
2058 req->task_work.func = io_req_task_cancel;
2060 if (unlikely(io_req_task_work_add(req)))
2061 io_req_task_work_add_fallback(req, io_req_task_cancel);
2064 static void io_req_task_queue(struct io_kiocb *req)
2066 req->task_work.func = io_req_task_submit;
2068 if (unlikely(io_req_task_work_add(req)))
2069 io_req_task_queue_fail(req, -ECANCELED);
2072 static inline void io_queue_next(struct io_kiocb *req)
2074 struct io_kiocb *nxt = io_req_find_next(req);
2077 io_req_task_queue(nxt);
2080 static void io_free_req(struct io_kiocb *req)
2087 struct task_struct *task;
2092 static inline void io_init_req_batch(struct req_batch *rb)
2099 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2100 struct req_batch *rb)
2103 io_put_task(rb->task, rb->task_refs);
2105 percpu_ref_put_many(&ctx->refs, rb->ctx_refs);
2108 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req,
2109 struct io_submit_state *state)
2112 io_dismantle_req(req);
2114 if (req->task != rb->task) {
2116 io_put_task(rb->task, rb->task_refs);
2117 rb->task = req->task;
2123 if (state->free_reqs != ARRAY_SIZE(state->reqs))
2124 state->reqs[state->free_reqs++] = req;
2126 list_add(&req->compl.list, &state->comp.free_list);
2129 static void io_submit_flush_completions(struct io_ring_ctx *ctx)
2131 struct io_comp_state *cs = &ctx->submit_state.comp;
2133 struct io_kiocb *req;
2134 struct req_batch rb;
2136 io_init_req_batch(&rb);
2137 spin_lock_irq(&ctx->completion_lock);
2138 for (i = 0; i < nr; i++) {
2140 __io_cqring_fill_event(ctx, req->user_data, req->result,
2143 io_commit_cqring(ctx);
2144 spin_unlock_irq(&ctx->completion_lock);
2146 io_cqring_ev_posted(ctx);
2147 for (i = 0; i < nr; i++) {
2150 /* submission and completion refs */
2151 if (req_ref_sub_and_test(req, 2))
2152 io_req_free_batch(&rb, req, &ctx->submit_state);
2155 io_req_free_batch_finish(ctx, &rb);
2160 * Drop reference to request, return next in chain (if there is one) if this
2161 * was the last reference to this request.
2163 static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2165 struct io_kiocb *nxt = NULL;
2167 if (req_ref_put_and_test(req)) {
2168 nxt = io_req_find_next(req);
2174 static inline void io_put_req(struct io_kiocb *req)
2176 if (req_ref_put_and_test(req))
2180 static void io_put_req_deferred_cb(struct callback_head *cb)
2182 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2187 static void io_free_req_deferred(struct io_kiocb *req)
2189 req->task_work.func = io_put_req_deferred_cb;
2190 if (unlikely(io_req_task_work_add(req)))
2191 io_req_task_work_add_fallback(req, io_put_req_deferred_cb);
2194 static inline void io_put_req_deferred(struct io_kiocb *req, int refs)
2196 if (req_ref_sub_and_test(req, refs))
2197 io_free_req_deferred(req);
2200 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2202 /* See comment at the top of this file */
2204 return __io_cqring_events(ctx);
2207 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2209 struct io_rings *rings = ctx->rings;
2211 /* make sure SQ entry isn't read before tail */
2212 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2215 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2217 unsigned int cflags;
2219 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2220 cflags |= IORING_CQE_F_BUFFER;
2221 req->flags &= ~REQ_F_BUFFER_SELECTED;
2226 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2228 struct io_buffer *kbuf;
2230 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2231 return io_put_kbuf(req, kbuf);
2234 static inline bool io_run_task_work(void)
2237 * Not safe to run on exiting task, and the task_work handling will
2238 * not add work to such a task.
2240 if (unlikely(current->flags & PF_EXITING))
2242 if (current->task_works) {
2243 __set_current_state(TASK_RUNNING);
2252 * Find and free completed poll iocbs
2254 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2255 struct list_head *done)
2257 struct req_batch rb;
2258 struct io_kiocb *req;
2260 /* order with ->result store in io_complete_rw_iopoll() */
2263 io_init_req_batch(&rb);
2264 while (!list_empty(done)) {
2267 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2268 list_del(&req->inflight_entry);
2270 if (READ_ONCE(req->result) == -EAGAIN &&
2271 !(req->flags & REQ_F_DONT_REISSUE)) {
2272 req->iopoll_completed = 0;
2274 io_queue_async_work(req);
2278 if (req->flags & REQ_F_BUFFER_SELECTED)
2279 cflags = io_put_rw_kbuf(req);
2281 __io_cqring_fill_event(ctx, req->user_data, req->result, cflags);
2284 if (req_ref_put_and_test(req))
2285 io_req_free_batch(&rb, req, &ctx->submit_state);
2288 io_commit_cqring(ctx);
2289 io_cqring_ev_posted_iopoll(ctx);
2290 io_req_free_batch_finish(ctx, &rb);
2293 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2296 struct io_kiocb *req, *tmp;
2302 * Only spin for completions if we don't have multiple devices hanging
2303 * off our complete list, and we're under the requested amount.
2305 spin = !ctx->poll_multi_file && *nr_events < min;
2308 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2309 struct kiocb *kiocb = &req->rw.kiocb;
2312 * Move completed and retryable entries to our local lists.
2313 * If we find a request that requires polling, break out
2314 * and complete those lists first, if we have entries there.
2316 if (READ_ONCE(req->iopoll_completed)) {
2317 list_move_tail(&req->inflight_entry, &done);
2320 if (!list_empty(&done))
2323 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2327 /* iopoll may have completed current req */
2328 if (READ_ONCE(req->iopoll_completed))
2329 list_move_tail(&req->inflight_entry, &done);
2336 if (!list_empty(&done))
2337 io_iopoll_complete(ctx, nr_events, &done);
2343 * We can't just wait for polled events to come to us, we have to actively
2344 * find and complete them.
2346 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2348 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2351 mutex_lock(&ctx->uring_lock);
2352 while (!list_empty(&ctx->iopoll_list)) {
2353 unsigned int nr_events = 0;
2355 io_do_iopoll(ctx, &nr_events, 0);
2357 /* let it sleep and repeat later if can't complete a request */
2361 * Ensure we allow local-to-the-cpu processing to take place,
2362 * in this case we need to ensure that we reap all events.
2363 * Also let task_work, etc. to progress by releasing the mutex
2365 if (need_resched()) {
2366 mutex_unlock(&ctx->uring_lock);
2368 mutex_lock(&ctx->uring_lock);
2371 mutex_unlock(&ctx->uring_lock);
2374 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2376 unsigned int nr_events = 0;
2380 * We disallow the app entering submit/complete with polling, but we
2381 * still need to lock the ring to prevent racing with polled issue
2382 * that got punted to a workqueue.
2384 mutex_lock(&ctx->uring_lock);
2386 * Don't enter poll loop if we already have events pending.
2387 * If we do, we can potentially be spinning for commands that
2388 * already triggered a CQE (eg in error).
2390 if (test_bit(0, &ctx->check_cq_overflow))
2391 __io_cqring_overflow_flush(ctx, false);
2392 if (io_cqring_events(ctx))
2396 * If a submit got punted to a workqueue, we can have the
2397 * application entering polling for a command before it gets
2398 * issued. That app will hold the uring_lock for the duration
2399 * of the poll right here, so we need to take a breather every
2400 * now and then to ensure that the issue has a chance to add
2401 * the poll to the issued list. Otherwise we can spin here
2402 * forever, while the workqueue is stuck trying to acquire the
2405 if (list_empty(&ctx->iopoll_list)) {
2406 mutex_unlock(&ctx->uring_lock);
2408 mutex_lock(&ctx->uring_lock);
2410 if (list_empty(&ctx->iopoll_list))
2413 ret = io_do_iopoll(ctx, &nr_events, min);
2414 } while (!ret && nr_events < min && !need_resched());
2416 mutex_unlock(&ctx->uring_lock);
2420 static void kiocb_end_write(struct io_kiocb *req)
2423 * Tell lockdep we inherited freeze protection from submission
2426 if (req->flags & REQ_F_ISREG) {
2427 struct super_block *sb = file_inode(req->file)->i_sb;
2429 __sb_writers_acquired(sb, SB_FREEZE_WRITE);
2435 static bool io_resubmit_prep(struct io_kiocb *req)
2437 struct io_async_rw *rw = req->async_data;
2440 return !io_req_prep_async(req);
2441 /* may have left rw->iter inconsistent on -EIOCBQUEUED */
2442 iov_iter_revert(&rw->iter, req->result - iov_iter_count(&rw->iter));
2446 static bool io_rw_should_reissue(struct io_kiocb *req)
2448 umode_t mode = file_inode(req->file)->i_mode;
2449 struct io_ring_ctx *ctx = req->ctx;
2451 if (!S_ISBLK(mode) && !S_ISREG(mode))
2453 if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
2454 !(ctx->flags & IORING_SETUP_IOPOLL)))
2457 * If ref is dying, we might be running poll reap from the exit work.
2458 * Don't attempt to reissue from that path, just let it fail with
2461 if (percpu_ref_is_dying(&ctx->refs))
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_complete_rw(struct io_kiocb *req, long res, long res2,
2477 unsigned int issue_flags)
2481 if (req->rw.kiocb.ki_flags & IOCB_WRITE)
2482 kiocb_end_write(req);
2483 if (res != req->result) {
2484 if ((res == -EAGAIN || res == -EOPNOTSUPP) &&
2485 io_rw_should_reissue(req)) {
2486 req->flags |= REQ_F_REISSUE;
2491 if (req->flags & REQ_F_BUFFER_SELECTED)
2492 cflags = io_put_rw_kbuf(req);
2493 __io_req_complete(req, issue_flags, res, cflags);
2496 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2498 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2500 __io_complete_rw(req, res, res2, 0);
2503 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2505 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2507 if (kiocb->ki_flags & IOCB_WRITE)
2508 kiocb_end_write(req);
2509 if (unlikely(res != req->result)) {
2510 if (!(res == -EAGAIN && io_rw_should_reissue(req) &&
2511 io_resubmit_prep(req))) {
2513 req->flags |= REQ_F_DONT_REISSUE;
2517 WRITE_ONCE(req->result, res);
2518 /* order with io_iopoll_complete() checking ->result */
2520 WRITE_ONCE(req->iopoll_completed, 1);
2524 * After the iocb has been issued, it's safe to be found on the poll list.
2525 * Adding the kiocb to the list AFTER submission ensures that we don't
2526 * find it from a io_do_iopoll() thread before the issuer is done
2527 * accessing the kiocb cookie.
2529 static void io_iopoll_req_issued(struct io_kiocb *req)
2531 struct io_ring_ctx *ctx = req->ctx;
2532 const bool in_async = io_wq_current_is_worker();
2534 /* workqueue context doesn't hold uring_lock, grab it now */
2535 if (unlikely(in_async))
2536 mutex_lock(&ctx->uring_lock);
2539 * Track whether we have multiple files in our lists. This will impact
2540 * how we do polling eventually, not spinning if we're on potentially
2541 * different devices.
2543 if (list_empty(&ctx->iopoll_list)) {
2544 ctx->poll_multi_file = false;
2545 } else if (!ctx->poll_multi_file) {
2546 struct io_kiocb *list_req;
2548 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2550 if (list_req->file != req->file)
2551 ctx->poll_multi_file = true;
2555 * For fast devices, IO may have already completed. If it has, add
2556 * it to the front so we find it first.
2558 if (READ_ONCE(req->iopoll_completed))
2559 list_add(&req->inflight_entry, &ctx->iopoll_list);
2561 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2563 if (unlikely(in_async)) {
2565 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
2566 * in sq thread task context or in io worker task context. If
2567 * current task context is sq thread, we don't need to check
2568 * whether should wake up sq thread.
2570 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2571 wq_has_sleeper(&ctx->sq_data->wait))
2572 wake_up(&ctx->sq_data->wait);
2574 mutex_unlock(&ctx->uring_lock);
2578 static inline void io_state_file_put(struct io_submit_state *state)
2580 if (state->file_refs) {
2581 fput_many(state->file, state->file_refs);
2582 state->file_refs = 0;
2587 * Get as many references to a file as we have IOs left in this submission,
2588 * assuming most submissions are for one file, or at least that each file
2589 * has more than one submission.
2591 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2596 if (state->file_refs) {
2597 if (state->fd == fd) {
2601 io_state_file_put(state);
2603 state->file = fget_many(fd, state->ios_left);
2604 if (unlikely(!state->file))
2608 state->file_refs = state->ios_left - 1;
2612 static bool io_bdev_nowait(struct block_device *bdev)
2614 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2618 * If we tracked the file through the SCM inflight mechanism, we could support
2619 * any file. For now, just ensure that anything potentially problematic is done
2622 static bool __io_file_supports_async(struct file *file, int rw)
2624 umode_t mode = file_inode(file)->i_mode;
2626 if (S_ISBLK(mode)) {
2627 if (IS_ENABLED(CONFIG_BLOCK) &&
2628 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2634 if (S_ISREG(mode)) {
2635 if (IS_ENABLED(CONFIG_BLOCK) &&
2636 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2637 file->f_op != &io_uring_fops)
2642 /* any ->read/write should understand O_NONBLOCK */
2643 if (file->f_flags & O_NONBLOCK)
2646 if (!(file->f_mode & FMODE_NOWAIT))
2650 return file->f_op->read_iter != NULL;
2652 return file->f_op->write_iter != NULL;
2655 static bool io_file_supports_async(struct io_kiocb *req, int rw)
2657 if (rw == READ && (req->flags & REQ_F_ASYNC_READ))
2659 else if (rw == WRITE && (req->flags & REQ_F_ASYNC_WRITE))
2662 return __io_file_supports_async(req->file, rw);
2665 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2667 struct io_ring_ctx *ctx = req->ctx;
2668 struct kiocb *kiocb = &req->rw.kiocb;
2669 struct file *file = req->file;
2673 if (!(req->flags & REQ_F_ISREG) && S_ISREG(file_inode(file)->i_mode))
2674 req->flags |= REQ_F_ISREG;
2676 kiocb->ki_pos = READ_ONCE(sqe->off);
2677 if (kiocb->ki_pos == -1 && !(file->f_mode & FMODE_STREAM)) {
2678 req->flags |= REQ_F_CUR_POS;
2679 kiocb->ki_pos = file->f_pos;
2681 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2682 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2683 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2687 /* don't allow async punt for O_NONBLOCK or RWF_NOWAIT */
2688 if ((kiocb->ki_flags & IOCB_NOWAIT) || (file->f_flags & O_NONBLOCK))
2689 req->flags |= REQ_F_NOWAIT;
2691 ioprio = READ_ONCE(sqe->ioprio);
2693 ret = ioprio_check_cap(ioprio);
2697 kiocb->ki_ioprio = ioprio;
2699 kiocb->ki_ioprio = get_current_ioprio();
2701 if (ctx->flags & IORING_SETUP_IOPOLL) {
2702 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2703 !kiocb->ki_filp->f_op->iopoll)
2706 kiocb->ki_flags |= IOCB_HIPRI;
2707 kiocb->ki_complete = io_complete_rw_iopoll;
2708 req->iopoll_completed = 0;
2710 if (kiocb->ki_flags & IOCB_HIPRI)
2712 kiocb->ki_complete = io_complete_rw;
2715 if (req->opcode == IORING_OP_READ_FIXED ||
2716 req->opcode == IORING_OP_WRITE_FIXED) {
2718 io_req_set_rsrc_node(req);
2721 req->rw.addr = READ_ONCE(sqe->addr);
2722 req->rw.len = READ_ONCE(sqe->len);
2723 req->buf_index = READ_ONCE(sqe->buf_index);
2727 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2733 case -ERESTARTNOINTR:
2734 case -ERESTARTNOHAND:
2735 case -ERESTART_RESTARTBLOCK:
2737 * We can't just restart the syscall, since previously
2738 * submitted sqes may already be in progress. Just fail this
2744 kiocb->ki_complete(kiocb, ret, 0);
2748 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2749 unsigned int issue_flags)
2751 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2752 struct io_async_rw *io = req->async_data;
2753 bool check_reissue = kiocb->ki_complete == io_complete_rw;
2755 /* add previously done IO, if any */
2756 if (io && io->bytes_done > 0) {
2758 ret = io->bytes_done;
2760 ret += io->bytes_done;
2763 if (req->flags & REQ_F_CUR_POS)
2764 req->file->f_pos = kiocb->ki_pos;
2765 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2766 __io_complete_rw(req, ret, 0, issue_flags);
2768 io_rw_done(kiocb, ret);
2770 if (check_reissue && req->flags & REQ_F_REISSUE) {
2771 req->flags &= ~REQ_F_REISSUE;
2772 if (io_resubmit_prep(req)) {
2774 io_queue_async_work(req);
2779 if (req->flags & REQ_F_BUFFER_SELECTED)
2780 cflags = io_put_rw_kbuf(req);
2781 __io_req_complete(req, issue_flags, ret, cflags);
2786 static int __io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter,
2787 struct io_mapped_ubuf *imu)
2789 size_t len = req->rw.len;
2790 u64 buf_end, buf_addr = req->rw.addr;
2793 if (unlikely(check_add_overflow(buf_addr, (u64)len, &buf_end)))
2795 /* not inside the mapped region */
2796 if (unlikely(buf_addr < imu->ubuf || buf_end > imu->ubuf_end))
2800 * May not be a start of buffer, set size appropriately
2801 * and advance us to the beginning.
2803 offset = buf_addr - imu->ubuf;
2804 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2808 * Don't use iov_iter_advance() here, as it's really slow for
2809 * using the latter parts of a big fixed buffer - it iterates
2810 * over each segment manually. We can cheat a bit here, because
2813 * 1) it's a BVEC iter, we set it up
2814 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2815 * first and last bvec
2817 * So just find our index, and adjust the iterator afterwards.
2818 * If the offset is within the first bvec (or the whole first
2819 * bvec, just use iov_iter_advance(). This makes it easier
2820 * since we can just skip the first segment, which may not
2821 * be PAGE_SIZE aligned.
2823 const struct bio_vec *bvec = imu->bvec;
2825 if (offset <= bvec->bv_len) {
2826 iov_iter_advance(iter, offset);
2828 unsigned long seg_skip;
2830 /* skip first vec */
2831 offset -= bvec->bv_len;
2832 seg_skip = 1 + (offset >> PAGE_SHIFT);
2834 iter->bvec = bvec + seg_skip;
2835 iter->nr_segs -= seg_skip;
2836 iter->count -= bvec->bv_len + offset;
2837 iter->iov_offset = offset & ~PAGE_MASK;
2844 static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
2846 struct io_ring_ctx *ctx = req->ctx;
2847 struct io_mapped_ubuf *imu = req->imu;
2848 u16 index, buf_index = req->buf_index;
2851 if (unlikely(buf_index >= ctx->nr_user_bufs))
2853 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2854 imu = READ_ONCE(ctx->user_bufs[index]);
2857 return __io_import_fixed(req, rw, iter, imu);
2860 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2863 mutex_unlock(&ctx->uring_lock);
2866 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2869 * "Normal" inline submissions always hold the uring_lock, since we
2870 * grab it from the system call. Same is true for the SQPOLL offload.
2871 * The only exception is when we've detached the request and issue it
2872 * from an async worker thread, grab the lock for that case.
2875 mutex_lock(&ctx->uring_lock);
2878 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2879 int bgid, struct io_buffer *kbuf,
2882 struct io_buffer *head;
2884 if (req->flags & REQ_F_BUFFER_SELECTED)
2887 io_ring_submit_lock(req->ctx, needs_lock);
2889 lockdep_assert_held(&req->ctx->uring_lock);
2891 head = xa_load(&req->ctx->io_buffers, bgid);
2893 if (!list_empty(&head->list)) {
2894 kbuf = list_last_entry(&head->list, struct io_buffer,
2896 list_del(&kbuf->list);
2899 xa_erase(&req->ctx->io_buffers, bgid);
2901 if (*len > kbuf->len)
2904 kbuf = ERR_PTR(-ENOBUFS);
2907 io_ring_submit_unlock(req->ctx, needs_lock);
2912 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2915 struct io_buffer *kbuf;
2918 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2919 bgid = req->buf_index;
2920 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2923 req->rw.addr = (u64) (unsigned long) kbuf;
2924 req->flags |= REQ_F_BUFFER_SELECTED;
2925 return u64_to_user_ptr(kbuf->addr);
2928 #ifdef CONFIG_COMPAT
2929 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2932 struct compat_iovec __user *uiov;
2933 compat_ssize_t clen;
2937 uiov = u64_to_user_ptr(req->rw.addr);
2938 if (!access_ok(uiov, sizeof(*uiov)))
2940 if (__get_user(clen, &uiov->iov_len))
2946 buf = io_rw_buffer_select(req, &len, needs_lock);
2948 return PTR_ERR(buf);
2949 iov[0].iov_base = buf;
2950 iov[0].iov_len = (compat_size_t) len;
2955 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2958 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2962 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2965 len = iov[0].iov_len;
2968 buf = io_rw_buffer_select(req, &len, needs_lock);
2970 return PTR_ERR(buf);
2971 iov[0].iov_base = buf;
2972 iov[0].iov_len = len;
2976 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2979 if (req->flags & REQ_F_BUFFER_SELECTED) {
2980 struct io_buffer *kbuf;
2982 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2983 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
2984 iov[0].iov_len = kbuf->len;
2987 if (req->rw.len != 1)
2990 #ifdef CONFIG_COMPAT
2991 if (req->ctx->compat)
2992 return io_compat_import(req, iov, needs_lock);
2995 return __io_iov_buffer_select(req, iov, needs_lock);
2998 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
2999 struct iov_iter *iter, bool needs_lock)
3001 void __user *buf = u64_to_user_ptr(req->rw.addr);
3002 size_t sqe_len = req->rw.len;
3003 u8 opcode = req->opcode;
3006 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3008 return io_import_fixed(req, rw, iter);
3011 /* buffer index only valid with fixed read/write, or buffer select */
3012 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3015 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3016 if (req->flags & REQ_F_BUFFER_SELECT) {
3017 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3019 return PTR_ERR(buf);
3020 req->rw.len = sqe_len;
3023 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3028 if (req->flags & REQ_F_BUFFER_SELECT) {
3029 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3031 iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
3036 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3040 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3042 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3046 * For files that don't have ->read_iter() and ->write_iter(), handle them
3047 * by looping over ->read() or ->write() manually.
3049 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3051 struct kiocb *kiocb = &req->rw.kiocb;
3052 struct file *file = req->file;
3056 * Don't support polled IO through this interface, and we can't
3057 * support non-blocking either. For the latter, this just causes
3058 * the kiocb to be handled from an async context.
3060 if (kiocb->ki_flags & IOCB_HIPRI)
3062 if (kiocb->ki_flags & IOCB_NOWAIT)
3065 while (iov_iter_count(iter)) {
3069 if (!iov_iter_is_bvec(iter)) {
3070 iovec = iov_iter_iovec(iter);
3072 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3073 iovec.iov_len = req->rw.len;
3077 nr = file->f_op->read(file, iovec.iov_base,
3078 iovec.iov_len, io_kiocb_ppos(kiocb));
3080 nr = file->f_op->write(file, iovec.iov_base,
3081 iovec.iov_len, io_kiocb_ppos(kiocb));
3090 if (nr != iovec.iov_len)
3094 iov_iter_advance(iter, nr);
3100 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3101 const struct iovec *fast_iov, struct iov_iter *iter)
3103 struct io_async_rw *rw = req->async_data;
3105 memcpy(&rw->iter, iter, sizeof(*iter));
3106 rw->free_iovec = iovec;
3108 /* can only be fixed buffers, no need to do anything */
3109 if (iov_iter_is_bvec(iter))
3112 unsigned iov_off = 0;
3114 rw->iter.iov = rw->fast_iov;
3115 if (iter->iov != fast_iov) {
3116 iov_off = iter->iov - fast_iov;
3117 rw->iter.iov += iov_off;
3119 if (rw->fast_iov != fast_iov)
3120 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3121 sizeof(struct iovec) * iter->nr_segs);
3123 req->flags |= REQ_F_NEED_CLEANUP;
3127 static inline int io_alloc_async_data(struct io_kiocb *req)
3129 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3130 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3131 return req->async_data == NULL;
3134 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3135 const struct iovec *fast_iov,
3136 struct iov_iter *iter, bool force)
3138 if (!force && !io_op_defs[req->opcode].needs_async_setup)
3140 if (!req->async_data) {
3141 if (io_alloc_async_data(req)) {
3146 io_req_map_rw(req, iovec, fast_iov, iter);
3151 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3153 struct io_async_rw *iorw = req->async_data;
3154 struct iovec *iov = iorw->fast_iov;
3157 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3158 if (unlikely(ret < 0))
3161 iorw->bytes_done = 0;
3162 iorw->free_iovec = iov;
3164 req->flags |= REQ_F_NEED_CLEANUP;
3168 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3170 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3172 return io_prep_rw(req, sqe);
3176 * This is our waitqueue callback handler, registered through lock_page_async()
3177 * when we initially tried to do the IO with the iocb armed our waitqueue.
3178 * This gets called when the page is unlocked, and we generally expect that to
3179 * happen when the page IO is completed and the page is now uptodate. This will
3180 * queue a task_work based retry of the operation, attempting to copy the data
3181 * again. If the latter fails because the page was NOT uptodate, then we will
3182 * do a thread based blocking retry of the operation. That's the unexpected
3185 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3186 int sync, void *arg)
3188 struct wait_page_queue *wpq;
3189 struct io_kiocb *req = wait->private;
3190 struct wait_page_key *key = arg;
3192 wpq = container_of(wait, struct wait_page_queue, wait);
3194 if (!wake_page_match(wpq, key))
3197 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3198 list_del_init(&wait->entry);
3200 /* submit ref gets dropped, acquire a new one */
3202 io_req_task_queue(req);
3207 * This controls whether a given IO request should be armed for async page
3208 * based retry. If we return false here, the request is handed to the async
3209 * worker threads for retry. If we're doing buffered reads on a regular file,
3210 * we prepare a private wait_page_queue entry and retry the operation. This
3211 * will either succeed because the page is now uptodate and unlocked, or it
3212 * will register a callback when the page is unlocked at IO completion. Through
3213 * that callback, io_uring uses task_work to setup a retry of the operation.
3214 * That retry will attempt the buffered read again. The retry will generally
3215 * succeed, or in rare cases where it fails, we then fall back to using the
3216 * async worker threads for a blocking retry.
3218 static bool io_rw_should_retry(struct io_kiocb *req)
3220 struct io_async_rw *rw = req->async_data;
3221 struct wait_page_queue *wait = &rw->wpq;
3222 struct kiocb *kiocb = &req->rw.kiocb;
3224 /* never retry for NOWAIT, we just complete with -EAGAIN */
3225 if (req->flags & REQ_F_NOWAIT)
3228 /* Only for buffered IO */
3229 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3233 * just use poll if we can, and don't attempt if the fs doesn't
3234 * support callback based unlocks
3236 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3239 wait->wait.func = io_async_buf_func;
3240 wait->wait.private = req;
3241 wait->wait.flags = 0;
3242 INIT_LIST_HEAD(&wait->wait.entry);
3243 kiocb->ki_flags |= IOCB_WAITQ;
3244 kiocb->ki_flags &= ~IOCB_NOWAIT;
3245 kiocb->ki_waitq = wait;
3249 static inline int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3251 if (req->file->f_op->read_iter)
3252 return call_read_iter(req->file, &req->rw.kiocb, iter);
3253 else if (req->file->f_op->read)
3254 return loop_rw_iter(READ, req, iter);
3259 static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3261 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3262 struct kiocb *kiocb = &req->rw.kiocb;
3263 struct iov_iter __iter, *iter = &__iter;
3264 struct io_async_rw *rw = req->async_data;
3265 ssize_t io_size, ret, ret2;
3266 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3272 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3276 io_size = iov_iter_count(iter);
3277 req->result = io_size;
3279 /* Ensure we clear previously set non-block flag */
3280 if (!force_nonblock)
3281 kiocb->ki_flags &= ~IOCB_NOWAIT;
3283 kiocb->ki_flags |= IOCB_NOWAIT;
3285 /* If the file doesn't support async, just async punt */
3286 if (force_nonblock && !io_file_supports_async(req, READ)) {
3287 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3288 return ret ?: -EAGAIN;
3291 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size);
3292 if (unlikely(ret)) {
3297 ret = io_iter_do_read(req, iter);
3299 if (ret == -EAGAIN || (req->flags & REQ_F_REISSUE)) {
3300 req->flags &= ~REQ_F_REISSUE;
3301 /* IOPOLL retry should happen for io-wq threads */
3302 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3304 /* no retry on NONBLOCK nor RWF_NOWAIT */
3305 if (req->flags & REQ_F_NOWAIT)
3307 /* some cases will consume bytes even on error returns */
3308 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3310 } else if (ret == -EIOCBQUEUED) {
3312 } else if (ret <= 0 || ret == io_size || !force_nonblock ||
3313 (req->flags & REQ_F_NOWAIT) || !(req->flags & REQ_F_ISREG)) {
3314 /* read all, failed, already did sync or don't want to retry */
3318 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3323 rw = req->async_data;
3324 /* now use our persistent iterator, if we aren't already */
3329 rw->bytes_done += ret;
3330 /* if we can retry, do so with the callbacks armed */
3331 if (!io_rw_should_retry(req)) {
3332 kiocb->ki_flags &= ~IOCB_WAITQ;
3337 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3338 * we get -EIOCBQUEUED, then we'll get a notification when the
3339 * desired page gets unlocked. We can also get a partial read
3340 * here, and if we do, then just retry at the new offset.
3342 ret = io_iter_do_read(req, iter);
3343 if (ret == -EIOCBQUEUED)
3345 /* we got some bytes, but not all. retry. */
3346 kiocb->ki_flags &= ~IOCB_WAITQ;
3347 } while (ret > 0 && ret < io_size);
3349 kiocb_done(kiocb, ret, issue_flags);
3351 /* it's faster to check here then delegate to kfree */
3357 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3359 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3361 return io_prep_rw(req, sqe);
3364 static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3366 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3367 struct kiocb *kiocb = &req->rw.kiocb;
3368 struct iov_iter __iter, *iter = &__iter;
3369 struct io_async_rw *rw = req->async_data;
3370 ssize_t ret, ret2, io_size;
3371 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3377 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3381 io_size = iov_iter_count(iter);
3382 req->result = io_size;
3384 /* Ensure we clear previously set non-block flag */
3385 if (!force_nonblock)
3386 kiocb->ki_flags &= ~IOCB_NOWAIT;
3388 kiocb->ki_flags |= IOCB_NOWAIT;
3390 /* If the file doesn't support async, just async punt */
3391 if (force_nonblock && !io_file_supports_async(req, WRITE))
3394 /* file path doesn't support NOWAIT for non-direct_IO */
3395 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3396 (req->flags & REQ_F_ISREG))
3399 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size);
3404 * Open-code file_start_write here to grab freeze protection,
3405 * which will be released by another thread in
3406 * io_complete_rw(). Fool lockdep by telling it the lock got
3407 * released so that it doesn't complain about the held lock when
3408 * we return to userspace.
3410 if (req->flags & REQ_F_ISREG) {
3411 sb_start_write(file_inode(req->file)->i_sb);
3412 __sb_writers_release(file_inode(req->file)->i_sb,
3415 kiocb->ki_flags |= IOCB_WRITE;
3417 if (req->file->f_op->write_iter)
3418 ret2 = call_write_iter(req->file, kiocb, iter);
3419 else if (req->file->f_op->write)
3420 ret2 = loop_rw_iter(WRITE, req, iter);
3424 if (req->flags & REQ_F_REISSUE) {
3425 req->flags &= ~REQ_F_REISSUE;
3430 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3431 * retry them without IOCB_NOWAIT.
3433 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3435 /* no retry on NONBLOCK nor RWF_NOWAIT */
3436 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3438 if (!force_nonblock || ret2 != -EAGAIN) {
3439 /* IOPOLL retry should happen for io-wq threads */
3440 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3443 kiocb_done(kiocb, ret2, issue_flags);
3446 /* some cases will consume bytes even on error returns */
3447 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3448 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3449 return ret ?: -EAGAIN;
3452 /* it's reportedly faster than delegating the null check to kfree() */
3458 static int io_renameat_prep(struct io_kiocb *req,
3459 const struct io_uring_sqe *sqe)
3461 struct io_rename *ren = &req->rename;
3462 const char __user *oldf, *newf;
3464 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3467 ren->old_dfd = READ_ONCE(sqe->fd);
3468 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3469 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3470 ren->new_dfd = READ_ONCE(sqe->len);
3471 ren->flags = READ_ONCE(sqe->rename_flags);
3473 ren->oldpath = getname(oldf);
3474 if (IS_ERR(ren->oldpath))
3475 return PTR_ERR(ren->oldpath);
3477 ren->newpath = getname(newf);
3478 if (IS_ERR(ren->newpath)) {
3479 putname(ren->oldpath);
3480 return PTR_ERR(ren->newpath);
3483 req->flags |= REQ_F_NEED_CLEANUP;
3487 static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3489 struct io_rename *ren = &req->rename;
3492 if (issue_flags & IO_URING_F_NONBLOCK)
3495 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3496 ren->newpath, ren->flags);
3498 req->flags &= ~REQ_F_NEED_CLEANUP;
3501 io_req_complete(req, ret);
3505 static int io_unlinkat_prep(struct io_kiocb *req,
3506 const struct io_uring_sqe *sqe)
3508 struct io_unlink *un = &req->unlink;
3509 const char __user *fname;
3511 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3514 un->dfd = READ_ONCE(sqe->fd);
3516 un->flags = READ_ONCE(sqe->unlink_flags);
3517 if (un->flags & ~AT_REMOVEDIR)
3520 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3521 un->filename = getname(fname);
3522 if (IS_ERR(un->filename))
3523 return PTR_ERR(un->filename);
3525 req->flags |= REQ_F_NEED_CLEANUP;
3529 static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3531 struct io_unlink *un = &req->unlink;
3534 if (issue_flags & IO_URING_F_NONBLOCK)
3537 if (un->flags & AT_REMOVEDIR)
3538 ret = do_rmdir(un->dfd, un->filename);
3540 ret = do_unlinkat(un->dfd, un->filename);
3542 req->flags &= ~REQ_F_NEED_CLEANUP;
3545 io_req_complete(req, ret);
3549 static int io_shutdown_prep(struct io_kiocb *req,
3550 const struct io_uring_sqe *sqe)
3552 #if defined(CONFIG_NET)
3553 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3555 if (sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3559 req->shutdown.how = READ_ONCE(sqe->len);
3566 static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3568 #if defined(CONFIG_NET)
3569 struct socket *sock;
3572 if (issue_flags & IO_URING_F_NONBLOCK)
3575 sock = sock_from_file(req->file);
3576 if (unlikely(!sock))
3579 ret = __sys_shutdown_sock(sock, req->shutdown.how);
3582 io_req_complete(req, ret);
3589 static int __io_splice_prep(struct io_kiocb *req,
3590 const struct io_uring_sqe *sqe)
3592 struct io_splice* sp = &req->splice;
3593 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3595 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3599 sp->len = READ_ONCE(sqe->len);
3600 sp->flags = READ_ONCE(sqe->splice_flags);
3602 if (unlikely(sp->flags & ~valid_flags))
3605 sp->file_in = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in),
3606 (sp->flags & SPLICE_F_FD_IN_FIXED));
3609 req->flags |= REQ_F_NEED_CLEANUP;
3613 static int io_tee_prep(struct io_kiocb *req,
3614 const struct io_uring_sqe *sqe)
3616 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3618 return __io_splice_prep(req, sqe);
3621 static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
3623 struct io_splice *sp = &req->splice;
3624 struct file *in = sp->file_in;
3625 struct file *out = sp->file_out;
3626 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3629 if (issue_flags & IO_URING_F_NONBLOCK)
3632 ret = do_tee(in, out, sp->len, flags);
3634 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3636 req->flags &= ~REQ_F_NEED_CLEANUP;
3640 io_req_complete(req, ret);
3644 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3646 struct io_splice* sp = &req->splice;
3648 sp->off_in = READ_ONCE(sqe->splice_off_in);
3649 sp->off_out = READ_ONCE(sqe->off);
3650 return __io_splice_prep(req, sqe);
3653 static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
3655 struct io_splice *sp = &req->splice;
3656 struct file *in = sp->file_in;
3657 struct file *out = sp->file_out;
3658 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3659 loff_t *poff_in, *poff_out;
3662 if (issue_flags & IO_URING_F_NONBLOCK)
3665 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3666 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3669 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3671 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3673 req->flags &= ~REQ_F_NEED_CLEANUP;
3677 io_req_complete(req, ret);
3682 * IORING_OP_NOP just posts a completion event, nothing else.
3684 static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
3686 struct io_ring_ctx *ctx = req->ctx;
3688 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3691 __io_req_complete(req, issue_flags, 0, 0);
3695 static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3697 struct io_ring_ctx *ctx = req->ctx;
3702 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3704 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3707 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3708 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3711 req->sync.off = READ_ONCE(sqe->off);
3712 req->sync.len = READ_ONCE(sqe->len);
3716 static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
3718 loff_t end = req->sync.off + req->sync.len;
3721 /* fsync always requires a blocking context */
3722 if (issue_flags & IO_URING_F_NONBLOCK)
3725 ret = vfs_fsync_range(req->file, req->sync.off,
3726 end > 0 ? end : LLONG_MAX,
3727 req->sync.flags & IORING_FSYNC_DATASYNC);
3730 io_req_complete(req, ret);
3734 static int io_fallocate_prep(struct io_kiocb *req,
3735 const struct io_uring_sqe *sqe)
3737 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3739 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3742 req->sync.off = READ_ONCE(sqe->off);
3743 req->sync.len = READ_ONCE(sqe->addr);
3744 req->sync.mode = READ_ONCE(sqe->len);
3748 static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
3752 /* fallocate always requiring blocking context */
3753 if (issue_flags & IO_URING_F_NONBLOCK)
3755 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3759 io_req_complete(req, ret);
3763 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3765 const char __user *fname;
3768 if (unlikely(sqe->ioprio || sqe->buf_index))
3770 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3773 /* open.how should be already initialised */
3774 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3775 req->open.how.flags |= O_LARGEFILE;
3777 req->open.dfd = READ_ONCE(sqe->fd);
3778 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3779 req->open.filename = getname(fname);
3780 if (IS_ERR(req->open.filename)) {
3781 ret = PTR_ERR(req->open.filename);
3782 req->open.filename = NULL;
3785 req->open.nofile = rlimit(RLIMIT_NOFILE);
3786 req->flags |= REQ_F_NEED_CLEANUP;
3790 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3794 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3796 mode = READ_ONCE(sqe->len);
3797 flags = READ_ONCE(sqe->open_flags);
3798 req->open.how = build_open_how(flags, mode);
3799 return __io_openat_prep(req, sqe);
3802 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3804 struct open_how __user *how;
3808 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3810 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3811 len = READ_ONCE(sqe->len);
3812 if (len < OPEN_HOW_SIZE_VER0)
3815 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3820 return __io_openat_prep(req, sqe);
3823 static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
3825 struct open_flags op;
3828 bool resolve_nonblock;
3831 ret = build_open_flags(&req->open.how, &op);
3834 nonblock_set = op.open_flag & O_NONBLOCK;
3835 resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
3836 if (issue_flags & IO_URING_F_NONBLOCK) {
3838 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
3839 * it'll always -EAGAIN
3841 if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
3843 op.lookup_flags |= LOOKUP_CACHED;
3844 op.open_flag |= O_NONBLOCK;
3847 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3851 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3852 /* only retry if RESOLVE_CACHED wasn't already set by application */
3853 if ((!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)) &&
3854 file == ERR_PTR(-EAGAIN)) {
3856 * We could hang on to this 'fd', but seems like marginal
3857 * gain for something that is now known to be a slower path.
3858 * So just put it, and we'll get a new one when we retry.
3866 ret = PTR_ERR(file);
3868 if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
3869 file->f_flags &= ~O_NONBLOCK;
3870 fsnotify_open(file);
3871 fd_install(ret, file);
3874 putname(req->open.filename);
3875 req->flags &= ~REQ_F_NEED_CLEANUP;
3878 __io_req_complete(req, issue_flags, ret, 0);
3882 static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
3884 return io_openat2(req, issue_flags);
3887 static int io_remove_buffers_prep(struct io_kiocb *req,
3888 const struct io_uring_sqe *sqe)
3890 struct io_provide_buf *p = &req->pbuf;
3893 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3896 tmp = READ_ONCE(sqe->fd);
3897 if (!tmp || tmp > USHRT_MAX)
3900 memset(p, 0, sizeof(*p));
3902 p->bgid = READ_ONCE(sqe->buf_group);
3906 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3907 int bgid, unsigned nbufs)
3911 /* shouldn't happen */
3915 /* the head kbuf is the list itself */
3916 while (!list_empty(&buf->list)) {
3917 struct io_buffer *nxt;
3919 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3920 list_del(&nxt->list);
3927 xa_erase(&ctx->io_buffers, bgid);
3932 static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
3934 struct io_provide_buf *p = &req->pbuf;
3935 struct io_ring_ctx *ctx = req->ctx;
3936 struct io_buffer *head;
3938 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3940 io_ring_submit_lock(ctx, !force_nonblock);
3942 lockdep_assert_held(&ctx->uring_lock);
3945 head = xa_load(&ctx->io_buffers, p->bgid);
3947 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3951 /* complete before unlock, IOPOLL may need the lock */
3952 __io_req_complete(req, issue_flags, ret, 0);
3953 io_ring_submit_unlock(ctx, !force_nonblock);
3957 static int io_provide_buffers_prep(struct io_kiocb *req,
3958 const struct io_uring_sqe *sqe)
3960 unsigned long size, tmp_check;
3961 struct io_provide_buf *p = &req->pbuf;
3964 if (sqe->ioprio || sqe->rw_flags)
3967 tmp = READ_ONCE(sqe->fd);
3968 if (!tmp || tmp > USHRT_MAX)
3971 p->addr = READ_ONCE(sqe->addr);
3972 p->len = READ_ONCE(sqe->len);
3974 if (check_mul_overflow((unsigned long)p->len, (unsigned long)p->nbufs,
3977 if (check_add_overflow((unsigned long)p->addr, size, &tmp_check))
3980 size = (unsigned long)p->len * p->nbufs;
3981 if (!access_ok(u64_to_user_ptr(p->addr), size))
3984 p->bgid = READ_ONCE(sqe->buf_group);
3985 tmp = READ_ONCE(sqe->off);
3986 if (tmp > USHRT_MAX)
3992 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3994 struct io_buffer *buf;
3995 u64 addr = pbuf->addr;
3996 int i, bid = pbuf->bid;
3998 for (i = 0; i < pbuf->nbufs; i++) {
3999 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
4004 buf->len = min_t(__u32, pbuf->len, MAX_RW_COUNT);
4009 INIT_LIST_HEAD(&buf->list);
4012 list_add_tail(&buf->list, &(*head)->list);
4016 return i ? i : -ENOMEM;
4019 static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
4021 struct io_provide_buf *p = &req->pbuf;
4022 struct io_ring_ctx *ctx = req->ctx;
4023 struct io_buffer *head, *list;
4025 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4027 io_ring_submit_lock(ctx, !force_nonblock);
4029 lockdep_assert_held(&ctx->uring_lock);
4031 list = head = xa_load(&ctx->io_buffers, p->bgid);
4033 ret = io_add_buffers(p, &head);
4034 if (ret >= 0 && !list) {
4035 ret = xa_insert(&ctx->io_buffers, p->bgid, head, GFP_KERNEL);
4037 __io_remove_buffers(ctx, head, p->bgid, -1U);
4041 /* complete before unlock, IOPOLL may need the lock */
4042 __io_req_complete(req, issue_flags, ret, 0);
4043 io_ring_submit_unlock(ctx, !force_nonblock);
4047 static int io_epoll_ctl_prep(struct io_kiocb *req,
4048 const struct io_uring_sqe *sqe)
4050 #if defined(CONFIG_EPOLL)
4051 if (sqe->ioprio || sqe->buf_index)
4053 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4056 req->epoll.epfd = READ_ONCE(sqe->fd);
4057 req->epoll.op = READ_ONCE(sqe->len);
4058 req->epoll.fd = READ_ONCE(sqe->off);
4060 if (ep_op_has_event(req->epoll.op)) {
4061 struct epoll_event __user *ev;
4063 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4064 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4074 static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4076 #if defined(CONFIG_EPOLL)
4077 struct io_epoll *ie = &req->epoll;
4079 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4081 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4082 if (force_nonblock && ret == -EAGAIN)
4087 __io_req_complete(req, issue_flags, ret, 0);
4094 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4096 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4097 if (sqe->ioprio || sqe->buf_index || sqe->off)
4099 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4102 req->madvise.addr = READ_ONCE(sqe->addr);
4103 req->madvise.len = READ_ONCE(sqe->len);
4104 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4111 static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4113 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4114 struct io_madvise *ma = &req->madvise;
4117 if (issue_flags & IO_URING_F_NONBLOCK)
4120 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4123 io_req_complete(req, ret);
4130 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4132 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4134 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4137 req->fadvise.offset = READ_ONCE(sqe->off);
4138 req->fadvise.len = READ_ONCE(sqe->len);
4139 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4143 static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4145 struct io_fadvise *fa = &req->fadvise;
4148 if (issue_flags & IO_URING_F_NONBLOCK) {
4149 switch (fa->advice) {
4150 case POSIX_FADV_NORMAL:
4151 case POSIX_FADV_RANDOM:
4152 case POSIX_FADV_SEQUENTIAL:
4159 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4162 __io_req_complete(req, issue_flags, ret, 0);
4166 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4168 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4170 if (sqe->ioprio || sqe->buf_index)
4172 if (req->flags & REQ_F_FIXED_FILE)
4175 req->statx.dfd = READ_ONCE(sqe->fd);
4176 req->statx.mask = READ_ONCE(sqe->len);
4177 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4178 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4179 req->statx.flags = READ_ONCE(sqe->statx_flags);
4184 static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4186 struct io_statx *ctx = &req->statx;
4189 if (issue_flags & IO_URING_F_NONBLOCK)
4192 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4197 io_req_complete(req, ret);
4201 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4203 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4205 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4206 sqe->rw_flags || sqe->buf_index)
4208 if (req->flags & REQ_F_FIXED_FILE)
4211 req->close.fd = READ_ONCE(sqe->fd);
4215 static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4217 struct files_struct *files = current->files;
4218 struct io_close *close = &req->close;
4219 struct fdtable *fdt;
4220 struct file *file = NULL;
4223 spin_lock(&files->file_lock);
4224 fdt = files_fdtable(files);
4225 if (close->fd >= fdt->max_fds) {
4226 spin_unlock(&files->file_lock);
4229 file = fdt->fd[close->fd];
4230 if (!file || file->f_op == &io_uring_fops) {
4231 spin_unlock(&files->file_lock);
4236 /* if the file has a flush method, be safe and punt to async */
4237 if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4238 spin_unlock(&files->file_lock);
4242 ret = __close_fd_get_file(close->fd, &file);
4243 spin_unlock(&files->file_lock);
4250 /* No ->flush() or already async, safely close from here */
4251 ret = filp_close(file, current->files);
4257 __io_req_complete(req, issue_flags, ret, 0);
4261 static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4263 struct io_ring_ctx *ctx = req->ctx;
4265 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4267 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4270 req->sync.off = READ_ONCE(sqe->off);
4271 req->sync.len = READ_ONCE(sqe->len);
4272 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4276 static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4280 /* sync_file_range always requires a blocking context */
4281 if (issue_flags & IO_URING_F_NONBLOCK)
4284 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4288 io_req_complete(req, ret);
4292 #if defined(CONFIG_NET)
4293 static int io_setup_async_msg(struct io_kiocb *req,
4294 struct io_async_msghdr *kmsg)
4296 struct io_async_msghdr *async_msg = req->async_data;
4300 if (io_alloc_async_data(req)) {
4301 kfree(kmsg->free_iov);
4304 async_msg = req->async_data;
4305 req->flags |= REQ_F_NEED_CLEANUP;
4306 memcpy(async_msg, kmsg, sizeof(*kmsg));
4307 async_msg->msg.msg_name = &async_msg->addr;
4308 /* if were using fast_iov, set it to the new one */
4309 if (!async_msg->free_iov)
4310 async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4315 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4316 struct io_async_msghdr *iomsg)
4318 iomsg->msg.msg_name = &iomsg->addr;
4319 iomsg->free_iov = iomsg->fast_iov;
4320 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4321 req->sr_msg.msg_flags, &iomsg->free_iov);
4324 static int io_sendmsg_prep_async(struct io_kiocb *req)
4328 ret = io_sendmsg_copy_hdr(req, req->async_data);
4330 req->flags |= REQ_F_NEED_CLEANUP;
4334 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4336 struct io_sr_msg *sr = &req->sr_msg;
4338 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4341 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4342 sr->len = READ_ONCE(sqe->len);
4343 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4344 if (sr->msg_flags & MSG_DONTWAIT)
4345 req->flags |= REQ_F_NOWAIT;
4347 #ifdef CONFIG_COMPAT
4348 if (req->ctx->compat)
4349 sr->msg_flags |= MSG_CMSG_COMPAT;
4354 static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4356 struct io_async_msghdr iomsg, *kmsg;
4357 struct socket *sock;
4362 sock = sock_from_file(req->file);
4363 if (unlikely(!sock))
4366 kmsg = req->async_data;
4368 ret = io_sendmsg_copy_hdr(req, &iomsg);
4374 flags = req->sr_msg.msg_flags;
4375 if (issue_flags & IO_URING_F_NONBLOCK)
4376 flags |= MSG_DONTWAIT;
4377 if (flags & MSG_WAITALL)
4378 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4380 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4381 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4382 return io_setup_async_msg(req, kmsg);
4383 if (ret == -ERESTARTSYS)
4386 /* fast path, check for non-NULL to avoid function call */
4388 kfree(kmsg->free_iov);
4389 req->flags &= ~REQ_F_NEED_CLEANUP;
4392 __io_req_complete(req, issue_flags, ret, 0);
4396 static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4398 struct io_sr_msg *sr = &req->sr_msg;
4401 struct socket *sock;
4406 sock = sock_from_file(req->file);
4407 if (unlikely(!sock))
4410 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4414 msg.msg_name = NULL;
4415 msg.msg_control = NULL;
4416 msg.msg_controllen = 0;
4417 msg.msg_namelen = 0;
4419 flags = req->sr_msg.msg_flags;
4420 if (issue_flags & IO_URING_F_NONBLOCK)
4421 flags |= MSG_DONTWAIT;
4422 if (flags & MSG_WAITALL)
4423 min_ret = iov_iter_count(&msg.msg_iter);
4425 msg.msg_flags = flags;
4426 ret = sock_sendmsg(sock, &msg);
4427 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4429 if (ret == -ERESTARTSYS)
4434 __io_req_complete(req, issue_flags, ret, 0);
4438 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4439 struct io_async_msghdr *iomsg)
4441 struct io_sr_msg *sr = &req->sr_msg;
4442 struct iovec __user *uiov;
4446 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4447 &iomsg->uaddr, &uiov, &iov_len);
4451 if (req->flags & REQ_F_BUFFER_SELECT) {
4454 if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4456 sr->len = iomsg->fast_iov[0].iov_len;
4457 iomsg->free_iov = NULL;
4459 iomsg->free_iov = iomsg->fast_iov;
4460 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4461 &iomsg->free_iov, &iomsg->msg.msg_iter,
4470 #ifdef CONFIG_COMPAT
4471 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4472 struct io_async_msghdr *iomsg)
4474 struct io_sr_msg *sr = &req->sr_msg;
4475 struct compat_iovec __user *uiov;
4480 ret = __get_compat_msghdr(&iomsg->msg, sr->umsg_compat, &iomsg->uaddr,
4485 uiov = compat_ptr(ptr);
4486 if (req->flags & REQ_F_BUFFER_SELECT) {
4487 compat_ssize_t clen;
4491 if (!access_ok(uiov, sizeof(*uiov)))
4493 if (__get_user(clen, &uiov->iov_len))
4498 iomsg->free_iov = NULL;
4500 iomsg->free_iov = iomsg->fast_iov;
4501 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4502 UIO_FASTIOV, &iomsg->free_iov,
4503 &iomsg->msg.msg_iter, true);
4512 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4513 struct io_async_msghdr *iomsg)
4515 iomsg->msg.msg_name = &iomsg->addr;
4517 #ifdef CONFIG_COMPAT
4518 if (req->ctx->compat)
4519 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4522 return __io_recvmsg_copy_hdr(req, iomsg);
4525 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4528 struct io_sr_msg *sr = &req->sr_msg;
4529 struct io_buffer *kbuf;
4531 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4536 req->flags |= REQ_F_BUFFER_SELECTED;
4540 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4542 return io_put_kbuf(req, req->sr_msg.kbuf);
4545 static int io_recvmsg_prep_async(struct io_kiocb *req)
4549 ret = io_recvmsg_copy_hdr(req, req->async_data);
4551 req->flags |= REQ_F_NEED_CLEANUP;
4555 static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4557 struct io_sr_msg *sr = &req->sr_msg;
4559 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4562 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4563 sr->len = READ_ONCE(sqe->len);
4564 sr->bgid = READ_ONCE(sqe->buf_group);
4565 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4566 if (sr->msg_flags & MSG_DONTWAIT)
4567 req->flags |= REQ_F_NOWAIT;
4569 #ifdef CONFIG_COMPAT
4570 if (req->ctx->compat)
4571 sr->msg_flags |= MSG_CMSG_COMPAT;
4576 static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
4578 struct io_async_msghdr iomsg, *kmsg;
4579 struct socket *sock;
4580 struct io_buffer *kbuf;
4583 int ret, cflags = 0;
4584 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4586 sock = sock_from_file(req->file);
4587 if (unlikely(!sock))
4590 kmsg = req->async_data;
4592 ret = io_recvmsg_copy_hdr(req, &iomsg);
4598 if (req->flags & REQ_F_BUFFER_SELECT) {
4599 kbuf = io_recv_buffer_select(req, !force_nonblock);
4601 return PTR_ERR(kbuf);
4602 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4603 kmsg->fast_iov[0].iov_len = req->sr_msg.len;
4604 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
4605 1, req->sr_msg.len);
4608 flags = req->sr_msg.msg_flags;
4610 flags |= MSG_DONTWAIT;
4611 if (flags & MSG_WAITALL)
4612 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4614 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4615 kmsg->uaddr, flags);
4616 if (force_nonblock && ret == -EAGAIN)
4617 return io_setup_async_msg(req, kmsg);
4618 if (ret == -ERESTARTSYS)
4621 if (req->flags & REQ_F_BUFFER_SELECTED)
4622 cflags = io_put_recv_kbuf(req);
4623 /* fast path, check for non-NULL to avoid function call */
4625 kfree(kmsg->free_iov);
4626 req->flags &= ~REQ_F_NEED_CLEANUP;
4627 if (ret < min_ret || ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4629 __io_req_complete(req, issue_flags, ret, cflags);
4633 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
4635 struct io_buffer *kbuf;
4636 struct io_sr_msg *sr = &req->sr_msg;
4638 void __user *buf = sr->buf;
4639 struct socket *sock;
4643 int ret, cflags = 0;
4644 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4646 sock = sock_from_file(req->file);
4647 if (unlikely(!sock))
4650 if (req->flags & REQ_F_BUFFER_SELECT) {
4651 kbuf = io_recv_buffer_select(req, !force_nonblock);
4653 return PTR_ERR(kbuf);
4654 buf = u64_to_user_ptr(kbuf->addr);
4657 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4661 msg.msg_name = NULL;
4662 msg.msg_control = NULL;
4663 msg.msg_controllen = 0;
4664 msg.msg_namelen = 0;
4665 msg.msg_iocb = NULL;
4668 flags = req->sr_msg.msg_flags;
4670 flags |= MSG_DONTWAIT;
4671 if (flags & MSG_WAITALL)
4672 min_ret = iov_iter_count(&msg.msg_iter);
4674 ret = sock_recvmsg(sock, &msg, flags);
4675 if (force_nonblock && ret == -EAGAIN)
4677 if (ret == -ERESTARTSYS)
4680 if (req->flags & REQ_F_BUFFER_SELECTED)
4681 cflags = io_put_recv_kbuf(req);
4682 if (ret < min_ret || ((flags & MSG_WAITALL) && (msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4684 __io_req_complete(req, issue_flags, ret, cflags);
4688 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4690 struct io_accept *accept = &req->accept;
4692 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4694 if (sqe->ioprio || sqe->len || sqe->buf_index)
4697 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4698 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4699 accept->flags = READ_ONCE(sqe->accept_flags);
4700 accept->nofile = rlimit(RLIMIT_NOFILE);
4704 static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
4706 struct io_accept *accept = &req->accept;
4707 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4708 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4711 if (req->file->f_flags & O_NONBLOCK)
4712 req->flags |= REQ_F_NOWAIT;
4714 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4715 accept->addr_len, accept->flags,
4717 if (ret == -EAGAIN && force_nonblock)
4720 if (ret == -ERESTARTSYS)
4724 __io_req_complete(req, issue_flags, ret, 0);
4728 static int io_connect_prep_async(struct io_kiocb *req)
4730 struct io_async_connect *io = req->async_data;
4731 struct io_connect *conn = &req->connect;
4733 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
4736 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4738 struct io_connect *conn = &req->connect;
4740 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4742 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4745 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4746 conn->addr_len = READ_ONCE(sqe->addr2);
4750 static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
4752 struct io_async_connect __io, *io;
4753 unsigned file_flags;
4755 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4757 if (req->async_data) {
4758 io = req->async_data;
4760 ret = move_addr_to_kernel(req->connect.addr,
4761 req->connect.addr_len,
4768 file_flags = force_nonblock ? O_NONBLOCK : 0;
4770 ret = __sys_connect_file(req->file, &io->address,
4771 req->connect.addr_len, file_flags);
4772 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4773 if (req->async_data)
4775 if (io_alloc_async_data(req)) {
4779 memcpy(req->async_data, &__io, sizeof(__io));
4782 if (ret == -ERESTARTSYS)
4787 __io_req_complete(req, issue_flags, ret, 0);
4790 #else /* !CONFIG_NET */
4791 #define IO_NETOP_FN(op) \
4792 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
4794 return -EOPNOTSUPP; \
4797 #define IO_NETOP_PREP(op) \
4799 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
4801 return -EOPNOTSUPP; \
4804 #define IO_NETOP_PREP_ASYNC(op) \
4806 static int io_##op##_prep_async(struct io_kiocb *req) \
4808 return -EOPNOTSUPP; \
4811 IO_NETOP_PREP_ASYNC(sendmsg);
4812 IO_NETOP_PREP_ASYNC(recvmsg);
4813 IO_NETOP_PREP_ASYNC(connect);
4814 IO_NETOP_PREP(accept);
4817 #endif /* CONFIG_NET */
4819 struct io_poll_table {
4820 struct poll_table_struct pt;
4821 struct io_kiocb *req;
4825 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4826 __poll_t mask, task_work_func_t func)
4830 /* for instances that support it check for an event match first: */
4831 if (mask && !(mask & poll->events))
4834 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4836 list_del_init(&poll->wait.entry);
4839 req->task_work.func = func;
4842 * If this fails, then the task is exiting. When a task exits, the
4843 * work gets canceled, so just cancel this request as well instead
4844 * of executing it. We can't safely execute it anyway, as we may not
4845 * have the needed state needed for it anyway.
4847 ret = io_req_task_work_add(req);
4848 if (unlikely(ret)) {
4849 WRITE_ONCE(poll->canceled, true);
4850 io_req_task_work_add_fallback(req, func);
4855 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4856 __acquires(&req->ctx->completion_lock)
4858 struct io_ring_ctx *ctx = req->ctx;
4860 if (!req->result && !READ_ONCE(poll->canceled)) {
4861 struct poll_table_struct pt = { ._key = poll->events };
4863 req->result = vfs_poll(req->file, &pt) & poll->events;
4866 spin_lock_irq(&ctx->completion_lock);
4867 if (!req->result && !READ_ONCE(poll->canceled)) {
4868 add_wait_queue(poll->head, &poll->wait);
4875 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4877 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4878 if (req->opcode == IORING_OP_POLL_ADD)
4879 return req->async_data;
4880 return req->apoll->double_poll;
4883 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4885 if (req->opcode == IORING_OP_POLL_ADD)
4887 return &req->apoll->poll;
4890 static void io_poll_remove_double(struct io_kiocb *req)
4891 __must_hold(&req->ctx->completion_lock)
4893 struct io_poll_iocb *poll = io_poll_get_double(req);
4895 lockdep_assert_held(&req->ctx->completion_lock);
4897 if (poll && poll->head) {
4898 struct wait_queue_head *head = poll->head;
4900 spin_lock(&head->lock);
4901 list_del_init(&poll->wait.entry);
4902 if (poll->wait.private)
4905 spin_unlock(&head->lock);
4909 static bool io_poll_complete(struct io_kiocb *req, __poll_t mask)
4910 __must_hold(&req->ctx->completion_lock)
4912 struct io_ring_ctx *ctx = req->ctx;
4913 unsigned flags = IORING_CQE_F_MORE;
4916 if (READ_ONCE(req->poll.canceled)) {
4918 req->poll.events |= EPOLLONESHOT;
4920 error = mangle_poll(mask);
4922 if (req->poll.events & EPOLLONESHOT)
4924 if (!io_cqring_fill_event(ctx, req->user_data, error, flags)) {
4925 io_poll_remove_waitqs(req);
4926 req->poll.done = true;
4929 if (flags & IORING_CQE_F_MORE)
4932 io_commit_cqring(ctx);
4933 return !(flags & IORING_CQE_F_MORE);
4936 static void io_poll_task_func(struct callback_head *cb)
4938 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4939 struct io_ring_ctx *ctx = req->ctx;
4940 struct io_kiocb *nxt;
4942 if (io_poll_rewait(req, &req->poll)) {
4943 spin_unlock_irq(&ctx->completion_lock);
4947 done = io_poll_complete(req, req->result);
4949 hash_del(&req->hash_node);
4952 add_wait_queue(req->poll.head, &req->poll.wait);
4954 spin_unlock_irq(&ctx->completion_lock);
4955 io_cqring_ev_posted(ctx);
4958 nxt = io_put_req_find_next(req);
4960 __io_req_task_submit(nxt);
4965 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4966 int sync, void *key)
4968 struct io_kiocb *req = wait->private;
4969 struct io_poll_iocb *poll = io_poll_get_single(req);
4970 __poll_t mask = key_to_poll(key);
4972 /* for instances that support it check for an event match first: */
4973 if (mask && !(mask & poll->events))
4975 if (!(poll->events & EPOLLONESHOT))
4976 return poll->wait.func(&poll->wait, mode, sync, key);
4978 list_del_init(&wait->entry);
4980 if (poll && poll->head) {
4983 spin_lock(&poll->head->lock);
4984 done = list_empty(&poll->wait.entry);
4986 list_del_init(&poll->wait.entry);
4987 /* make sure double remove sees this as being gone */
4988 wait->private = NULL;
4989 spin_unlock(&poll->head->lock);
4991 /* use wait func handler, so it matches the rq type */
4992 poll->wait.func(&poll->wait, mode, sync, key);
4999 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
5000 wait_queue_func_t wake_func)
5004 poll->canceled = false;
5005 #define IO_POLL_UNMASK (EPOLLERR|EPOLLHUP|EPOLLNVAL|EPOLLRDHUP)
5006 /* mask in events that we always want/need */
5007 poll->events = events | IO_POLL_UNMASK;
5008 INIT_LIST_HEAD(&poll->wait.entry);
5009 init_waitqueue_func_entry(&poll->wait, wake_func);
5012 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5013 struct wait_queue_head *head,
5014 struct io_poll_iocb **poll_ptr)
5016 struct io_kiocb *req = pt->req;
5019 * If poll->head is already set, it's because the file being polled
5020 * uses multiple waitqueues for poll handling (eg one for read, one
5021 * for write). Setup a separate io_poll_iocb if this happens.
5023 if (unlikely(poll->head)) {
5024 struct io_poll_iocb *poll_one = poll;
5026 /* already have a 2nd entry, fail a third attempt */
5028 pt->error = -EINVAL;
5032 * Can't handle multishot for double wait for now, turn it
5033 * into one-shot mode.
5035 if (!(poll_one->events & EPOLLONESHOT))
5036 poll_one->events |= EPOLLONESHOT;
5037 /* double add on the same waitqueue head, ignore */
5038 if (poll_one->head == head)
5040 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5042 pt->error = -ENOMEM;
5045 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5047 poll->wait.private = req;
5054 if (poll->events & EPOLLEXCLUSIVE)
5055 add_wait_queue_exclusive(head, &poll->wait);
5057 add_wait_queue(head, &poll->wait);
5060 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5061 struct poll_table_struct *p)
5063 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5064 struct async_poll *apoll = pt->req->apoll;
5066 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5069 static void io_async_task_func(struct callback_head *cb)
5071 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
5072 struct async_poll *apoll = req->apoll;
5073 struct io_ring_ctx *ctx = req->ctx;
5075 trace_io_uring_task_run(req->ctx, req, req->opcode, req->user_data);
5077 if (io_poll_rewait(req, &apoll->poll)) {
5078 spin_unlock_irq(&ctx->completion_lock);
5082 hash_del(&req->hash_node);
5083 io_poll_remove_double(req);
5084 spin_unlock_irq(&ctx->completion_lock);
5086 if (!READ_ONCE(apoll->poll.canceled))
5087 __io_req_task_submit(req);
5089 io_req_complete_failed(req, -ECANCELED);
5092 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5095 struct io_kiocb *req = wait->private;
5096 struct io_poll_iocb *poll = &req->apoll->poll;
5098 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5101 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5104 static void io_poll_req_insert(struct io_kiocb *req)
5106 struct io_ring_ctx *ctx = req->ctx;
5107 struct hlist_head *list;
5109 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5110 hlist_add_head(&req->hash_node, list);
5113 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5114 struct io_poll_iocb *poll,
5115 struct io_poll_table *ipt, __poll_t mask,
5116 wait_queue_func_t wake_func)
5117 __acquires(&ctx->completion_lock)
5119 struct io_ring_ctx *ctx = req->ctx;
5120 bool cancel = false;
5122 INIT_HLIST_NODE(&req->hash_node);
5123 io_init_poll_iocb(poll, mask, wake_func);
5124 poll->file = req->file;
5125 poll->wait.private = req;
5127 ipt->pt._key = mask;
5129 ipt->error = -EINVAL;
5131 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5133 spin_lock_irq(&ctx->completion_lock);
5134 if (likely(poll->head)) {
5135 spin_lock(&poll->head->lock);
5136 if (unlikely(list_empty(&poll->wait.entry))) {
5142 if ((mask && (poll->events & EPOLLONESHOT)) || ipt->error)
5143 list_del_init(&poll->wait.entry);
5145 WRITE_ONCE(poll->canceled, true);
5146 else if (!poll->done) /* actually waiting for an event */
5147 io_poll_req_insert(req);
5148 spin_unlock(&poll->head->lock);
5154 static bool io_arm_poll_handler(struct io_kiocb *req)
5156 const struct io_op_def *def = &io_op_defs[req->opcode];
5157 struct io_ring_ctx *ctx = req->ctx;
5158 struct async_poll *apoll;
5159 struct io_poll_table ipt;
5163 if (!req->file || !file_can_poll(req->file))
5165 if (req->flags & REQ_F_POLLED)
5169 else if (def->pollout)
5173 /* if we can't nonblock try, then no point in arming a poll handler */
5174 if (!io_file_supports_async(req, rw))
5177 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5178 if (unlikely(!apoll))
5180 apoll->double_poll = NULL;
5182 req->flags |= REQ_F_POLLED;
5185 mask = EPOLLONESHOT;
5187 mask |= POLLIN | POLLRDNORM;
5189 mask |= POLLOUT | POLLWRNORM;
5191 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5192 if ((req->opcode == IORING_OP_RECVMSG) &&
5193 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5196 mask |= POLLERR | POLLPRI;
5198 ipt.pt._qproc = io_async_queue_proc;
5200 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5202 if (ret || ipt.error) {
5203 io_poll_remove_double(req);
5204 spin_unlock_irq(&ctx->completion_lock);
5207 spin_unlock_irq(&ctx->completion_lock);
5208 trace_io_uring_poll_arm(ctx, req, req->opcode, req->user_data,
5209 mask, apoll->poll.events);
5213 static bool __io_poll_remove_one(struct io_kiocb *req,
5214 struct io_poll_iocb *poll, bool do_cancel)
5215 __must_hold(&req->ctx->completion_lock)
5217 bool do_complete = false;
5221 spin_lock(&poll->head->lock);
5223 WRITE_ONCE(poll->canceled, true);
5224 if (!list_empty(&poll->wait.entry)) {
5225 list_del_init(&poll->wait.entry);
5228 spin_unlock(&poll->head->lock);
5229 hash_del(&req->hash_node);
5233 static bool io_poll_remove_waitqs(struct io_kiocb *req)
5234 __must_hold(&req->ctx->completion_lock)
5238 io_poll_remove_double(req);
5239 do_complete = __io_poll_remove_one(req, io_poll_get_single(req), true);
5241 if (req->opcode != IORING_OP_POLL_ADD && do_complete) {
5242 /* non-poll requests have submit ref still */
5248 static bool io_poll_remove_one(struct io_kiocb *req)
5249 __must_hold(&req->ctx->completion_lock)
5253 do_complete = io_poll_remove_waitqs(req);
5255 io_cqring_fill_event(req->ctx, req->user_data, -ECANCELED, 0);
5256 io_commit_cqring(req->ctx);
5258 io_put_req_deferred(req, 1);
5265 * Returns true if we found and killed one or more poll requests
5267 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5270 struct hlist_node *tmp;
5271 struct io_kiocb *req;
5274 spin_lock_irq(&ctx->completion_lock);
5275 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5276 struct hlist_head *list;
5278 list = &ctx->cancel_hash[i];
5279 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5280 if (io_match_task(req, tsk, cancel_all))
5281 posted += io_poll_remove_one(req);
5284 spin_unlock_irq(&ctx->completion_lock);
5287 io_cqring_ev_posted(ctx);
5292 static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, __u64 sqe_addr,
5294 __must_hold(&ctx->completion_lock)
5296 struct hlist_head *list;
5297 struct io_kiocb *req;
5299 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5300 hlist_for_each_entry(req, list, hash_node) {
5301 if (sqe_addr != req->user_data)
5303 if (poll_only && req->opcode != IORING_OP_POLL_ADD)
5310 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr,
5312 __must_hold(&ctx->completion_lock)
5314 struct io_kiocb *req;
5316 req = io_poll_find(ctx, sqe_addr, poll_only);
5319 if (io_poll_remove_one(req))
5325 static __poll_t io_poll_parse_events(const struct io_uring_sqe *sqe,
5330 events = READ_ONCE(sqe->poll32_events);
5332 events = swahw32(events);
5334 if (!(flags & IORING_POLL_ADD_MULTI))
5335 events |= EPOLLONESHOT;
5336 return demangle_poll(events) | (events & (EPOLLEXCLUSIVE|EPOLLONESHOT));
5339 static int io_poll_update_prep(struct io_kiocb *req,
5340 const struct io_uring_sqe *sqe)
5342 struct io_poll_update *upd = &req->poll_update;
5345 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5347 if (sqe->ioprio || sqe->buf_index)
5349 flags = READ_ONCE(sqe->len);
5350 if (flags & ~(IORING_POLL_UPDATE_EVENTS | IORING_POLL_UPDATE_USER_DATA |
5351 IORING_POLL_ADD_MULTI))
5353 /* meaningless without update */
5354 if (flags == IORING_POLL_ADD_MULTI)
5357 upd->old_user_data = READ_ONCE(sqe->addr);
5358 upd->update_events = flags & IORING_POLL_UPDATE_EVENTS;
5359 upd->update_user_data = flags & IORING_POLL_UPDATE_USER_DATA;
5361 upd->new_user_data = READ_ONCE(sqe->off);
5362 if (!upd->update_user_data && upd->new_user_data)
5364 if (upd->update_events)
5365 upd->events = io_poll_parse_events(sqe, flags);
5366 else if (sqe->poll32_events)
5372 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5375 struct io_kiocb *req = wait->private;
5376 struct io_poll_iocb *poll = &req->poll;
5378 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5381 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5382 struct poll_table_struct *p)
5384 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5386 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5389 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5391 struct io_poll_iocb *poll = &req->poll;
5394 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5396 if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->addr)
5398 flags = READ_ONCE(sqe->len);
5399 if (flags & ~IORING_POLL_ADD_MULTI)
5402 poll->events = io_poll_parse_events(sqe, flags);
5406 static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5408 struct io_poll_iocb *poll = &req->poll;
5409 struct io_ring_ctx *ctx = req->ctx;
5410 struct io_poll_table ipt;
5413 ipt.pt._qproc = io_poll_queue_proc;
5415 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5418 if (mask) { /* no async, we'd stolen it */
5420 io_poll_complete(req, mask);
5422 spin_unlock_irq(&ctx->completion_lock);
5425 io_cqring_ev_posted(ctx);
5426 if (poll->events & EPOLLONESHOT)
5432 static int io_poll_update(struct io_kiocb *req, unsigned int issue_flags)
5434 struct io_ring_ctx *ctx = req->ctx;
5435 struct io_kiocb *preq;
5439 spin_lock_irq(&ctx->completion_lock);
5440 preq = io_poll_find(ctx, req->poll_update.old_user_data, true);
5446 if (!req->poll_update.update_events && !req->poll_update.update_user_data) {
5448 ret = io_poll_remove_one(preq) ? 0 : -EALREADY;
5453 * Don't allow racy completion with singleshot, as we cannot safely
5454 * update those. For multishot, if we're racing with completion, just
5455 * let completion re-add it.
5457 completing = !__io_poll_remove_one(preq, &preq->poll, false);
5458 if (completing && (preq->poll.events & EPOLLONESHOT)) {
5462 /* we now have a detached poll request. reissue. */
5466 spin_unlock_irq(&ctx->completion_lock);
5468 io_req_complete(req, ret);
5471 /* only mask one event flags, keep behavior flags */
5472 if (req->poll_update.update_events) {
5473 preq->poll.events &= ~0xffff;
5474 preq->poll.events |= req->poll_update.events & 0xffff;
5475 preq->poll.events |= IO_POLL_UNMASK;
5477 if (req->poll_update.update_user_data)
5478 preq->user_data = req->poll_update.new_user_data;
5479 spin_unlock_irq(&ctx->completion_lock);
5481 /* complete update request, we're done with it */
5482 io_req_complete(req, ret);
5485 ret = io_poll_add(preq, issue_flags);
5488 io_req_complete(preq, ret);
5494 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5496 struct io_timeout_data *data = container_of(timer,
5497 struct io_timeout_data, timer);
5498 struct io_kiocb *req = data->req;
5499 struct io_ring_ctx *ctx = req->ctx;
5500 unsigned long flags;
5502 spin_lock_irqsave(&ctx->completion_lock, flags);
5503 list_del_init(&req->timeout.list);
5504 atomic_set(&req->ctx->cq_timeouts,
5505 atomic_read(&req->ctx->cq_timeouts) + 1);
5507 io_cqring_fill_event(ctx, req->user_data, -ETIME, 0);
5508 io_commit_cqring(ctx);
5509 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5511 io_cqring_ev_posted(ctx);
5514 return HRTIMER_NORESTART;
5517 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5519 __must_hold(&ctx->completion_lock)
5521 struct io_timeout_data *io;
5522 struct io_kiocb *req;
5525 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5526 found = user_data == req->user_data;
5531 return ERR_PTR(-ENOENT);
5533 io = req->async_data;
5534 if (hrtimer_try_to_cancel(&io->timer) == -1)
5535 return ERR_PTR(-EALREADY);
5536 list_del_init(&req->timeout.list);
5540 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5541 __must_hold(&ctx->completion_lock)
5543 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5546 return PTR_ERR(req);
5549 io_cqring_fill_event(ctx, req->user_data, -ECANCELED, 0);
5550 io_put_req_deferred(req, 1);
5554 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5555 struct timespec64 *ts, enum hrtimer_mode mode)
5556 __must_hold(&ctx->completion_lock)
5558 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5559 struct io_timeout_data *data;
5562 return PTR_ERR(req);
5564 req->timeout.off = 0; /* noseq */
5565 data = req->async_data;
5566 list_add_tail(&req->timeout.list, &ctx->timeout_list);
5567 hrtimer_init(&data->timer, CLOCK_MONOTONIC, mode);
5568 data->timer.function = io_timeout_fn;
5569 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
5573 static int io_timeout_remove_prep(struct io_kiocb *req,
5574 const struct io_uring_sqe *sqe)
5576 struct io_timeout_rem *tr = &req->timeout_rem;
5578 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5580 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5582 if (sqe->ioprio || sqe->buf_index || sqe->len)
5585 tr->addr = READ_ONCE(sqe->addr);
5586 tr->flags = READ_ONCE(sqe->timeout_flags);
5587 if (tr->flags & IORING_TIMEOUT_UPDATE) {
5588 if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS))
5590 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
5592 } else if (tr->flags) {
5593 /* timeout removal doesn't support flags */
5600 static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
5602 return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
5607 * Remove or update an existing timeout command
5609 static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
5611 struct io_timeout_rem *tr = &req->timeout_rem;
5612 struct io_ring_ctx *ctx = req->ctx;
5615 spin_lock_irq(&ctx->completion_lock);
5616 if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE))
5617 ret = io_timeout_cancel(ctx, tr->addr);
5619 ret = io_timeout_update(ctx, tr->addr, &tr->ts,
5620 io_translate_timeout_mode(tr->flags));
5622 io_cqring_fill_event(ctx, req->user_data, ret, 0);
5623 io_commit_cqring(ctx);
5624 spin_unlock_irq(&ctx->completion_lock);
5625 io_cqring_ev_posted(ctx);
5632 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5633 bool is_timeout_link)
5635 struct io_timeout_data *data;
5637 u32 off = READ_ONCE(sqe->off);
5639 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5641 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5643 if (off && is_timeout_link)
5645 flags = READ_ONCE(sqe->timeout_flags);
5646 if (flags & ~IORING_TIMEOUT_ABS)
5649 req->timeout.off = off;
5650 if (unlikely(off && !req->ctx->off_timeout_used))
5651 req->ctx->off_timeout_used = true;
5653 if (!req->async_data && io_alloc_async_data(req))
5656 data = req->async_data;
5659 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5662 data->mode = io_translate_timeout_mode(flags);
5663 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5664 if (is_timeout_link)
5665 io_req_track_inflight(req);
5669 static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
5671 struct io_ring_ctx *ctx = req->ctx;
5672 struct io_timeout_data *data = req->async_data;
5673 struct list_head *entry;
5674 u32 tail, off = req->timeout.off;
5676 spin_lock_irq(&ctx->completion_lock);
5679 * sqe->off holds how many events that need to occur for this
5680 * timeout event to be satisfied. If it isn't set, then this is
5681 * a pure timeout request, sequence isn't used.
5683 if (io_is_timeout_noseq(req)) {
5684 entry = ctx->timeout_list.prev;
5688 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5689 req->timeout.target_seq = tail + off;
5691 /* Update the last seq here in case io_flush_timeouts() hasn't.
5692 * This is safe because ->completion_lock is held, and submissions
5693 * and completions are never mixed in the same ->completion_lock section.
5695 ctx->cq_last_tm_flush = tail;
5698 * Insertion sort, ensuring the first entry in the list is always
5699 * the one we need first.
5701 list_for_each_prev(entry, &ctx->timeout_list) {
5702 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5705 if (io_is_timeout_noseq(nxt))
5707 /* nxt.seq is behind @tail, otherwise would've been completed */
5708 if (off >= nxt->timeout.target_seq - tail)
5712 list_add(&req->timeout.list, entry);
5713 data->timer.function = io_timeout_fn;
5714 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5715 spin_unlock_irq(&ctx->completion_lock);
5719 struct io_cancel_data {
5720 struct io_ring_ctx *ctx;
5724 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5726 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5727 struct io_cancel_data *cd = data;
5729 return req->ctx == cd->ctx && req->user_data == cd->user_data;
5732 static int io_async_cancel_one(struct io_uring_task *tctx, u64 user_data,
5733 struct io_ring_ctx *ctx)
5735 struct io_cancel_data data = { .ctx = ctx, .user_data = user_data, };
5736 enum io_wq_cancel cancel_ret;
5739 if (!tctx || !tctx->io_wq)
5742 cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, &data, false);
5743 switch (cancel_ret) {
5744 case IO_WQ_CANCEL_OK:
5747 case IO_WQ_CANCEL_RUNNING:
5750 case IO_WQ_CANCEL_NOTFOUND:
5758 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5759 struct io_kiocb *req, __u64 sqe_addr,
5762 unsigned long flags;
5765 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5766 spin_lock_irqsave(&ctx->completion_lock, flags);
5769 ret = io_timeout_cancel(ctx, sqe_addr);
5772 ret = io_poll_cancel(ctx, sqe_addr, false);
5776 io_cqring_fill_event(ctx, req->user_data, ret, 0);
5777 io_commit_cqring(ctx);
5778 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5779 io_cqring_ev_posted(ctx);
5785 static int io_async_cancel_prep(struct io_kiocb *req,
5786 const struct io_uring_sqe *sqe)
5788 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5790 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5792 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5795 req->cancel.addr = READ_ONCE(sqe->addr);
5799 static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
5801 struct io_ring_ctx *ctx = req->ctx;
5802 u64 sqe_addr = req->cancel.addr;
5803 struct io_tctx_node *node;
5806 /* tasks should wait for their io-wq threads, so safe w/o sync */
5807 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5808 spin_lock_irq(&ctx->completion_lock);
5811 ret = io_timeout_cancel(ctx, sqe_addr);
5814 ret = io_poll_cancel(ctx, sqe_addr, false);
5817 spin_unlock_irq(&ctx->completion_lock);
5819 /* slow path, try all io-wq's */
5820 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5822 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
5823 struct io_uring_task *tctx = node->task->io_uring;
5825 ret = io_async_cancel_one(tctx, req->cancel.addr, ctx);
5829 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5831 spin_lock_irq(&ctx->completion_lock);
5833 io_cqring_fill_event(ctx, req->user_data, ret, 0);
5834 io_commit_cqring(ctx);
5835 spin_unlock_irq(&ctx->completion_lock);
5836 io_cqring_ev_posted(ctx);
5844 static int io_rsrc_update_prep(struct io_kiocb *req,
5845 const struct io_uring_sqe *sqe)
5847 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5849 if (sqe->ioprio || sqe->rw_flags)
5852 req->rsrc_update.offset = READ_ONCE(sqe->off);
5853 req->rsrc_update.nr_args = READ_ONCE(sqe->len);
5854 if (!req->rsrc_update.nr_args)
5856 req->rsrc_update.arg = READ_ONCE(sqe->addr);
5860 static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
5862 struct io_ring_ctx *ctx = req->ctx;
5863 struct io_uring_rsrc_update2 up;
5866 if (issue_flags & IO_URING_F_NONBLOCK)
5869 up.offset = req->rsrc_update.offset;
5870 up.data = req->rsrc_update.arg;
5875 mutex_lock(&ctx->uring_lock);
5876 ret = __io_register_rsrc_update(ctx, IORING_RSRC_FILE,
5877 &up, req->rsrc_update.nr_args);
5878 mutex_unlock(&ctx->uring_lock);
5882 __io_req_complete(req, issue_flags, ret, 0);
5886 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5888 switch (req->opcode) {
5891 case IORING_OP_READV:
5892 case IORING_OP_READ_FIXED:
5893 case IORING_OP_READ:
5894 return io_read_prep(req, sqe);
5895 case IORING_OP_WRITEV:
5896 case IORING_OP_WRITE_FIXED:
5897 case IORING_OP_WRITE:
5898 return io_write_prep(req, sqe);
5899 case IORING_OP_POLL_ADD:
5900 return io_poll_add_prep(req, sqe);
5901 case IORING_OP_POLL_REMOVE:
5902 return io_poll_update_prep(req, sqe);
5903 case IORING_OP_FSYNC:
5904 return io_fsync_prep(req, sqe);
5905 case IORING_OP_SYNC_FILE_RANGE:
5906 return io_sfr_prep(req, sqe);
5907 case IORING_OP_SENDMSG:
5908 case IORING_OP_SEND:
5909 return io_sendmsg_prep(req, sqe);
5910 case IORING_OP_RECVMSG:
5911 case IORING_OP_RECV:
5912 return io_recvmsg_prep(req, sqe);
5913 case IORING_OP_CONNECT:
5914 return io_connect_prep(req, sqe);
5915 case IORING_OP_TIMEOUT:
5916 return io_timeout_prep(req, sqe, false);
5917 case IORING_OP_TIMEOUT_REMOVE:
5918 return io_timeout_remove_prep(req, sqe);
5919 case IORING_OP_ASYNC_CANCEL:
5920 return io_async_cancel_prep(req, sqe);
5921 case IORING_OP_LINK_TIMEOUT:
5922 return io_timeout_prep(req, sqe, true);
5923 case IORING_OP_ACCEPT:
5924 return io_accept_prep(req, sqe);
5925 case IORING_OP_FALLOCATE:
5926 return io_fallocate_prep(req, sqe);
5927 case IORING_OP_OPENAT:
5928 return io_openat_prep(req, sqe);
5929 case IORING_OP_CLOSE:
5930 return io_close_prep(req, sqe);
5931 case IORING_OP_FILES_UPDATE:
5932 return io_rsrc_update_prep(req, sqe);
5933 case IORING_OP_STATX:
5934 return io_statx_prep(req, sqe);
5935 case IORING_OP_FADVISE:
5936 return io_fadvise_prep(req, sqe);
5937 case IORING_OP_MADVISE:
5938 return io_madvise_prep(req, sqe);
5939 case IORING_OP_OPENAT2:
5940 return io_openat2_prep(req, sqe);
5941 case IORING_OP_EPOLL_CTL:
5942 return io_epoll_ctl_prep(req, sqe);
5943 case IORING_OP_SPLICE:
5944 return io_splice_prep(req, sqe);
5945 case IORING_OP_PROVIDE_BUFFERS:
5946 return io_provide_buffers_prep(req, sqe);
5947 case IORING_OP_REMOVE_BUFFERS:
5948 return io_remove_buffers_prep(req, sqe);
5950 return io_tee_prep(req, sqe);
5951 case IORING_OP_SHUTDOWN:
5952 return io_shutdown_prep(req, sqe);
5953 case IORING_OP_RENAMEAT:
5954 return io_renameat_prep(req, sqe);
5955 case IORING_OP_UNLINKAT:
5956 return io_unlinkat_prep(req, sqe);
5959 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5964 static int io_req_prep_async(struct io_kiocb *req)
5966 if (!io_op_defs[req->opcode].needs_async_setup)
5968 if (WARN_ON_ONCE(req->async_data))
5970 if (io_alloc_async_data(req))
5973 switch (req->opcode) {
5974 case IORING_OP_READV:
5975 return io_rw_prep_async(req, READ);
5976 case IORING_OP_WRITEV:
5977 return io_rw_prep_async(req, WRITE);
5978 case IORING_OP_SENDMSG:
5979 return io_sendmsg_prep_async(req);
5980 case IORING_OP_RECVMSG:
5981 return io_recvmsg_prep_async(req);
5982 case IORING_OP_CONNECT:
5983 return io_connect_prep_async(req);
5985 printk_once(KERN_WARNING "io_uring: prep_async() bad opcode %d\n",
5990 static u32 io_get_sequence(struct io_kiocb *req)
5992 struct io_kiocb *pos;
5993 struct io_ring_ctx *ctx = req->ctx;
5996 io_for_each_link(pos, req)
5998 return ctx->cached_sq_head - nr_reqs;
6001 static bool io_drain_req(struct io_kiocb *req)
6003 struct io_kiocb *pos;
6004 struct io_ring_ctx *ctx = req->ctx;
6005 struct io_defer_entry *de;
6010 * If we need to drain a request in the middle of a link, drain the
6011 * head request and the next request/link after the current link.
6012 * Considering sequential execution of links, IOSQE_IO_DRAIN will be
6013 * maintained for every request of our link.
6015 if (ctx->drain_next) {
6016 req->flags |= REQ_F_IO_DRAIN;
6017 ctx->drain_next = false;
6019 /* not interested in head, start from the first linked */
6020 io_for_each_link(pos, req->link) {
6021 if (pos->flags & REQ_F_IO_DRAIN) {
6022 ctx->drain_next = true;
6023 req->flags |= REQ_F_IO_DRAIN;
6028 /* Still need defer if there is pending req in defer list. */
6029 if (likely(list_empty_careful(&ctx->defer_list) &&
6030 !(req->flags & REQ_F_IO_DRAIN))) {
6031 ctx->drain_active = false;
6035 seq = io_get_sequence(req);
6036 /* Still a chance to pass the sequence check */
6037 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
6040 ret = io_req_prep_async(req);
6043 io_prep_async_link(req);
6044 de = kmalloc(sizeof(*de), GFP_KERNEL);
6046 io_req_complete_failed(req, ret);
6050 spin_lock_irq(&ctx->completion_lock);
6051 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
6052 spin_unlock_irq(&ctx->completion_lock);
6054 io_queue_async_work(req);
6058 trace_io_uring_defer(ctx, req, req->user_data);
6061 list_add_tail(&de->list, &ctx->defer_list);
6062 spin_unlock_irq(&ctx->completion_lock);
6066 static void io_clean_op(struct io_kiocb *req)
6068 if (req->flags & REQ_F_BUFFER_SELECTED) {
6069 switch (req->opcode) {
6070 case IORING_OP_READV:
6071 case IORING_OP_READ_FIXED:
6072 case IORING_OP_READ:
6073 kfree((void *)(unsigned long)req->rw.addr);
6075 case IORING_OP_RECVMSG:
6076 case IORING_OP_RECV:
6077 kfree(req->sr_msg.kbuf);
6080 req->flags &= ~REQ_F_BUFFER_SELECTED;
6083 if (req->flags & REQ_F_NEED_CLEANUP) {
6084 switch (req->opcode) {
6085 case IORING_OP_READV:
6086 case IORING_OP_READ_FIXED:
6087 case IORING_OP_READ:
6088 case IORING_OP_WRITEV:
6089 case IORING_OP_WRITE_FIXED:
6090 case IORING_OP_WRITE: {
6091 struct io_async_rw *io = req->async_data;
6093 kfree(io->free_iovec);
6096 case IORING_OP_RECVMSG:
6097 case IORING_OP_SENDMSG: {
6098 struct io_async_msghdr *io = req->async_data;
6100 kfree(io->free_iov);
6103 case IORING_OP_SPLICE:
6105 if (!(req->splice.flags & SPLICE_F_FD_IN_FIXED))
6106 io_put_file(req->splice.file_in);
6108 case IORING_OP_OPENAT:
6109 case IORING_OP_OPENAT2:
6110 if (req->open.filename)
6111 putname(req->open.filename);
6113 case IORING_OP_RENAMEAT:
6114 putname(req->rename.oldpath);
6115 putname(req->rename.newpath);
6117 case IORING_OP_UNLINKAT:
6118 putname(req->unlink.filename);
6121 req->flags &= ~REQ_F_NEED_CLEANUP;
6123 if ((req->flags & REQ_F_POLLED) && req->apoll) {
6124 kfree(req->apoll->double_poll);
6128 if (req->flags & REQ_F_INFLIGHT) {
6129 struct io_uring_task *tctx = req->task->io_uring;
6131 atomic_dec(&tctx->inflight_tracked);
6132 req->flags &= ~REQ_F_INFLIGHT;
6136 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
6138 struct io_ring_ctx *ctx = req->ctx;
6139 const struct cred *creds = NULL;
6142 if (req->work.creds && req->work.creds != current_cred())
6143 creds = override_creds(req->work.creds);
6145 switch (req->opcode) {
6147 ret = io_nop(req, issue_flags);
6149 case IORING_OP_READV:
6150 case IORING_OP_READ_FIXED:
6151 case IORING_OP_READ:
6152 ret = io_read(req, issue_flags);
6154 case IORING_OP_WRITEV:
6155 case IORING_OP_WRITE_FIXED:
6156 case IORING_OP_WRITE:
6157 ret = io_write(req, issue_flags);
6159 case IORING_OP_FSYNC:
6160 ret = io_fsync(req, issue_flags);
6162 case IORING_OP_POLL_ADD:
6163 ret = io_poll_add(req, issue_flags);
6165 case IORING_OP_POLL_REMOVE:
6166 ret = io_poll_update(req, issue_flags);
6168 case IORING_OP_SYNC_FILE_RANGE:
6169 ret = io_sync_file_range(req, issue_flags);
6171 case IORING_OP_SENDMSG:
6172 ret = io_sendmsg(req, issue_flags);
6174 case IORING_OP_SEND:
6175 ret = io_send(req, issue_flags);
6177 case IORING_OP_RECVMSG:
6178 ret = io_recvmsg(req, issue_flags);
6180 case IORING_OP_RECV:
6181 ret = io_recv(req, issue_flags);
6183 case IORING_OP_TIMEOUT:
6184 ret = io_timeout(req, issue_flags);
6186 case IORING_OP_TIMEOUT_REMOVE:
6187 ret = io_timeout_remove(req, issue_flags);
6189 case IORING_OP_ACCEPT:
6190 ret = io_accept(req, issue_flags);
6192 case IORING_OP_CONNECT:
6193 ret = io_connect(req, issue_flags);
6195 case IORING_OP_ASYNC_CANCEL:
6196 ret = io_async_cancel(req, issue_flags);
6198 case IORING_OP_FALLOCATE:
6199 ret = io_fallocate(req, issue_flags);
6201 case IORING_OP_OPENAT:
6202 ret = io_openat(req, issue_flags);
6204 case IORING_OP_CLOSE:
6205 ret = io_close(req, issue_flags);
6207 case IORING_OP_FILES_UPDATE:
6208 ret = io_files_update(req, issue_flags);
6210 case IORING_OP_STATX:
6211 ret = io_statx(req, issue_flags);
6213 case IORING_OP_FADVISE:
6214 ret = io_fadvise(req, issue_flags);
6216 case IORING_OP_MADVISE:
6217 ret = io_madvise(req, issue_flags);
6219 case IORING_OP_OPENAT2:
6220 ret = io_openat2(req, issue_flags);
6222 case IORING_OP_EPOLL_CTL:
6223 ret = io_epoll_ctl(req, issue_flags);
6225 case IORING_OP_SPLICE:
6226 ret = io_splice(req, issue_flags);
6228 case IORING_OP_PROVIDE_BUFFERS:
6229 ret = io_provide_buffers(req, issue_flags);
6231 case IORING_OP_REMOVE_BUFFERS:
6232 ret = io_remove_buffers(req, issue_flags);
6235 ret = io_tee(req, issue_flags);
6237 case IORING_OP_SHUTDOWN:
6238 ret = io_shutdown(req, issue_flags);
6240 case IORING_OP_RENAMEAT:
6241 ret = io_renameat(req, issue_flags);
6243 case IORING_OP_UNLINKAT:
6244 ret = io_unlinkat(req, issue_flags);
6252 revert_creds(creds);
6255 /* If the op doesn't have a file, we're not polling for it */
6256 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file)
6257 io_iopoll_req_issued(req);
6262 static void io_wq_submit_work(struct io_wq_work *work)
6264 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6265 struct io_kiocb *timeout;
6268 timeout = io_prep_linked_timeout(req);
6270 io_queue_linked_timeout(timeout);
6272 if (work->flags & IO_WQ_WORK_CANCEL)
6277 ret = io_issue_sqe(req, 0);
6279 * We can get EAGAIN for polled IO even though we're
6280 * forcing a sync submission from here, since we can't
6281 * wait for request slots on the block side.
6289 /* avoid locking problems by failing it from a clean context */
6291 /* io-wq is going to take one down */
6293 io_req_task_queue_fail(req, ret);
6297 #define FFS_ASYNC_READ 0x1UL
6298 #define FFS_ASYNC_WRITE 0x2UL
6300 #define FFS_ISREG 0x4UL
6302 #define FFS_ISREG 0x0UL
6304 #define FFS_MASK ~(FFS_ASYNC_READ|FFS_ASYNC_WRITE|FFS_ISREG)
6306 static inline struct io_fixed_file *io_fixed_file_slot(struct io_file_table *table,
6309 struct io_fixed_file *table_l2;
6311 table_l2 = table->files[i >> IORING_FILE_TABLE_SHIFT];
6312 return &table_l2[i & IORING_FILE_TABLE_MASK];
6315 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6318 struct io_fixed_file *slot = io_fixed_file_slot(&ctx->file_table, index);
6320 return (struct file *) (slot->file_ptr & FFS_MASK);
6323 static void io_fixed_file_set(struct io_fixed_file *file_slot, struct file *file)
6325 unsigned long file_ptr = (unsigned long) file;
6327 if (__io_file_supports_async(file, READ))
6328 file_ptr |= FFS_ASYNC_READ;
6329 if (__io_file_supports_async(file, WRITE))
6330 file_ptr |= FFS_ASYNC_WRITE;
6331 if (S_ISREG(file_inode(file)->i_mode))
6332 file_ptr |= FFS_ISREG;
6333 file_slot->file_ptr = file_ptr;
6336 static struct file *io_file_get(struct io_submit_state *state,
6337 struct io_kiocb *req, int fd, bool fixed)
6339 struct io_ring_ctx *ctx = req->ctx;
6343 unsigned long file_ptr;
6345 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6347 fd = array_index_nospec(fd, ctx->nr_user_files);
6348 file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr;
6349 file = (struct file *) (file_ptr & FFS_MASK);
6350 file_ptr &= ~FFS_MASK;
6351 /* mask in overlapping REQ_F and FFS bits */
6352 req->flags |= (file_ptr << REQ_F_ASYNC_READ_BIT);
6353 io_req_set_rsrc_node(req);
6355 trace_io_uring_file_get(ctx, fd);
6356 file = __io_file_get(state, fd);
6358 /* we don't allow fixed io_uring files */
6359 if (file && unlikely(file->f_op == &io_uring_fops))
6360 io_req_track_inflight(req);
6366 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6368 struct io_timeout_data *data = container_of(timer,
6369 struct io_timeout_data, timer);
6370 struct io_kiocb *prev, *req = data->req;
6371 struct io_ring_ctx *ctx = req->ctx;
6372 unsigned long flags;
6374 spin_lock_irqsave(&ctx->completion_lock, flags);
6375 prev = req->timeout.head;
6376 req->timeout.head = NULL;
6379 * We don't expect the list to be empty, that will only happen if we
6380 * race with the completion of the linked work.
6383 io_remove_next_linked(prev);
6384 if (!req_ref_inc_not_zero(prev))
6387 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6390 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6391 io_put_req_deferred(prev, 1);
6392 io_put_req_deferred(req, 1);
6394 io_req_complete_post(req, -ETIME, 0);
6396 return HRTIMER_NORESTART;
6399 static void io_queue_linked_timeout(struct io_kiocb *req)
6401 struct io_ring_ctx *ctx = req->ctx;
6403 spin_lock_irq(&ctx->completion_lock);
6405 * If the back reference is NULL, then our linked request finished
6406 * before we got a chance to setup the timer
6408 if (req->timeout.head) {
6409 struct io_timeout_data *data = req->async_data;
6411 data->timer.function = io_link_timeout_fn;
6412 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6415 spin_unlock_irq(&ctx->completion_lock);
6416 /* drop submission reference */
6420 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6422 struct io_kiocb *nxt = req->link;
6424 if (!nxt || (req->flags & REQ_F_LINK_TIMEOUT) ||
6425 nxt->opcode != IORING_OP_LINK_TIMEOUT)
6428 nxt->timeout.head = req;
6429 nxt->flags |= REQ_F_LTIMEOUT_ACTIVE;
6430 req->flags |= REQ_F_LINK_TIMEOUT;
6434 static void __io_queue_sqe(struct io_kiocb *req)
6436 struct io_kiocb *linked_timeout = io_prep_linked_timeout(req);
6439 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
6442 * We async punt it if the file wasn't marked NOWAIT, or if the file
6443 * doesn't support non-blocking read/write attempts
6446 /* drop submission reference */
6447 if (req->flags & REQ_F_COMPLETE_INLINE) {
6448 struct io_ring_ctx *ctx = req->ctx;
6449 struct io_comp_state *cs = &ctx->submit_state.comp;
6451 cs->reqs[cs->nr++] = req;
6452 if (cs->nr == ARRAY_SIZE(cs->reqs))
6453 io_submit_flush_completions(ctx);
6457 } else if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6458 if (!io_arm_poll_handler(req)) {
6460 * Queued up for async execution, worker will release
6461 * submit reference when the iocb is actually submitted.
6463 io_queue_async_work(req);
6466 io_req_complete_failed(req, ret);
6469 io_queue_linked_timeout(linked_timeout);
6472 static inline void io_queue_sqe(struct io_kiocb *req)
6474 if (unlikely(req->ctx->drain_active) && io_drain_req(req))
6477 if (likely(!(req->flags & REQ_F_FORCE_ASYNC))) {
6478 __io_queue_sqe(req);
6480 int ret = io_req_prep_async(req);
6483 io_req_complete_failed(req, ret);
6485 io_queue_async_work(req);
6490 * Check SQE restrictions (opcode and flags).
6492 * Returns 'true' if SQE is allowed, 'false' otherwise.
6494 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6495 struct io_kiocb *req,
6496 unsigned int sqe_flags)
6498 if (!ctx->restricted)
6501 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6504 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6505 ctx->restrictions.sqe_flags_required)
6508 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6509 ctx->restrictions.sqe_flags_required))
6515 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6516 const struct io_uring_sqe *sqe)
6518 struct io_submit_state *state;
6519 unsigned int sqe_flags;
6520 int personality, ret = 0;
6522 req->opcode = READ_ONCE(sqe->opcode);
6523 /* same numerical values with corresponding REQ_F_*, safe to copy */
6524 req->flags = sqe_flags = READ_ONCE(sqe->flags);
6525 req->user_data = READ_ONCE(sqe->user_data);
6526 req->async_data = NULL;
6530 req->fixed_rsrc_refs = NULL;
6531 /* one is dropped after submission, the other at completion */
6532 atomic_set(&req->refs, 2);
6533 req->task = current;
6535 req->work.creds = NULL;
6537 /* enforce forwards compatibility on users */
6538 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6540 if (unlikely(req->opcode >= IORING_OP_LAST))
6542 if (unlikely(!io_check_restriction(ctx, req, sqe_flags)))
6545 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6546 !io_op_defs[req->opcode].buffer_select)
6548 if (unlikely(sqe_flags & IOSQE_IO_DRAIN))
6549 ctx->drain_active = true;
6551 personality = READ_ONCE(sqe->personality);
6553 req->work.creds = xa_load(&ctx->personalities, personality);
6554 if (!req->work.creds)
6556 get_cred(req->work.creds);
6558 state = &ctx->submit_state;
6561 * Plug now if we have more than 1 IO left after this, and the target
6562 * is potentially a read/write to block based storage.
6564 if (!state->plug_started && state->ios_left > 1 &&
6565 io_op_defs[req->opcode].plug) {
6566 blk_start_plug(&state->plug);
6567 state->plug_started = true;
6570 if (io_op_defs[req->opcode].needs_file) {
6571 bool fixed = req->flags & REQ_F_FIXED_FILE;
6573 req->file = io_file_get(state, req, READ_ONCE(sqe->fd), fixed);
6574 if (unlikely(!req->file))
6582 static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
6583 const struct io_uring_sqe *sqe)
6585 struct io_submit_link *link = &ctx->submit_state.link;
6588 ret = io_init_req(ctx, req, sqe);
6589 if (unlikely(ret)) {
6592 /* fail even hard links since we don't submit */
6593 req_set_fail(link->head);
6594 io_req_complete_failed(link->head, -ECANCELED);
6597 io_req_complete_failed(req, ret);
6601 ret = io_req_prep(req, sqe);
6605 /* don't need @sqe from now on */
6606 trace_io_uring_submit_sqe(ctx, req, req->opcode, req->user_data,
6608 ctx->flags & IORING_SETUP_SQPOLL);
6611 * If we already have a head request, queue this one for async
6612 * submittal once the head completes. If we don't have a head but
6613 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6614 * submitted sync once the chain is complete. If none of those
6615 * conditions are true (normal request), then just queue it.
6618 struct io_kiocb *head = link->head;
6620 ret = io_req_prep_async(req);
6623 trace_io_uring_link(ctx, req, head);
6624 link->last->link = req;
6627 /* last request of a link, enqueue the link */
6628 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6633 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6645 * Batched submission is done, ensure local IO is flushed out.
6647 static void io_submit_state_end(struct io_submit_state *state,
6648 struct io_ring_ctx *ctx)
6650 if (state->link.head)
6651 io_queue_sqe(state->link.head);
6653 io_submit_flush_completions(ctx);
6654 if (state->plug_started)
6655 blk_finish_plug(&state->plug);
6656 io_state_file_put(state);
6660 * Start submission side cache.
6662 static void io_submit_state_start(struct io_submit_state *state,
6663 unsigned int max_ios)
6665 state->plug_started = false;
6666 state->ios_left = max_ios;
6667 /* set only head, no need to init link_last in advance */
6668 state->link.head = NULL;
6671 static void io_commit_sqring(struct io_ring_ctx *ctx)
6673 struct io_rings *rings = ctx->rings;
6676 * Ensure any loads from the SQEs are done at this point,
6677 * since once we write the new head, the application could
6678 * write new data to them.
6680 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6684 * Fetch an sqe, if one is available. Note this returns a pointer to memory
6685 * that is mapped by userspace. This means that care needs to be taken to
6686 * ensure that reads are stable, as we cannot rely on userspace always
6687 * being a good citizen. If members of the sqe are validated and then later
6688 * used, it's important that those reads are done through READ_ONCE() to
6689 * prevent a re-load down the line.
6691 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6693 unsigned head, mask = ctx->sq_entries - 1;
6694 unsigned sq_idx = ctx->cached_sq_head++ & mask;
6697 * The cached sq head (or cq tail) serves two purposes:
6699 * 1) allows us to batch the cost of updating the user visible
6701 * 2) allows the kernel side to track the head on its own, even
6702 * though the application is the one updating it.
6704 head = READ_ONCE(ctx->sq_array[sq_idx]);
6705 if (likely(head < ctx->sq_entries))
6706 return &ctx->sq_sqes[head];
6708 /* drop invalid entries */
6710 WRITE_ONCE(ctx->rings->sq_dropped,
6711 READ_ONCE(ctx->rings->sq_dropped) + 1);
6715 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6717 struct io_uring_task *tctx;
6720 /* make sure SQ entry isn't read before tail */
6721 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6722 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6725 tctx = current->io_uring;
6726 tctx->cached_refs -= nr;
6727 if (unlikely(tctx->cached_refs < 0)) {
6728 unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR;
6730 percpu_counter_add(&tctx->inflight, refill);
6731 refcount_add(refill, ¤t->usage);
6732 tctx->cached_refs += refill;
6734 io_submit_state_start(&ctx->submit_state, nr);
6736 while (submitted < nr) {
6737 const struct io_uring_sqe *sqe;
6738 struct io_kiocb *req;
6740 req = io_alloc_req(ctx);
6741 if (unlikely(!req)) {
6743 submitted = -EAGAIN;
6746 sqe = io_get_sqe(ctx);
6747 if (unlikely(!sqe)) {
6748 kmem_cache_free(req_cachep, req);
6751 /* will complete beyond this point, count as submitted */
6753 if (io_submit_sqe(ctx, req, sqe))
6757 if (unlikely(submitted != nr)) {
6758 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6759 int unused = nr - ref_used;
6761 current->io_uring->cached_refs += unused;
6762 percpu_ref_put_many(&ctx->refs, unused);
6765 io_submit_state_end(&ctx->submit_state, ctx);
6766 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6767 io_commit_sqring(ctx);
6772 static inline bool io_sqd_events_pending(struct io_sq_data *sqd)
6774 return READ_ONCE(sqd->state);
6777 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6779 /* Tell userspace we may need a wakeup call */
6780 spin_lock_irq(&ctx->completion_lock);
6781 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6782 spin_unlock_irq(&ctx->completion_lock);
6785 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6787 spin_lock_irq(&ctx->completion_lock);
6788 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6789 spin_unlock_irq(&ctx->completion_lock);
6792 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
6794 unsigned int to_submit;
6797 to_submit = io_sqring_entries(ctx);
6798 /* if we're handling multiple rings, cap submit size for fairness */
6799 if (cap_entries && to_submit > 8)
6802 if (!list_empty(&ctx->iopoll_list) || to_submit) {
6803 unsigned nr_events = 0;
6805 mutex_lock(&ctx->uring_lock);
6806 if (!list_empty(&ctx->iopoll_list))
6807 io_do_iopoll(ctx, &nr_events, 0);
6810 * Don't submit if refs are dying, good for io_uring_register(),
6811 * but also it is relied upon by io_ring_exit_work()
6813 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)) &&
6814 !(ctx->flags & IORING_SETUP_R_DISABLED))
6815 ret = io_submit_sqes(ctx, to_submit);
6816 mutex_unlock(&ctx->uring_lock);
6818 if (to_submit && wq_has_sleeper(&ctx->sqo_sq_wait))
6819 wake_up(&ctx->sqo_sq_wait);
6825 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
6827 struct io_ring_ctx *ctx;
6828 unsigned sq_thread_idle = 0;
6830 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6831 sq_thread_idle = max(sq_thread_idle, ctx->sq_thread_idle);
6832 sqd->sq_thread_idle = sq_thread_idle;
6835 static bool io_sqd_handle_event(struct io_sq_data *sqd)
6837 bool did_sig = false;
6838 struct ksignal ksig;
6840 if (test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state) ||
6841 signal_pending(current)) {
6842 mutex_unlock(&sqd->lock);
6843 if (signal_pending(current))
6844 did_sig = get_signal(&ksig);
6846 mutex_lock(&sqd->lock);
6849 return did_sig || test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
6852 static int io_sq_thread(void *data)
6854 struct io_sq_data *sqd = data;
6855 struct io_ring_ctx *ctx;
6856 unsigned long timeout = 0;
6857 char buf[TASK_COMM_LEN];
6860 snprintf(buf, sizeof(buf), "iou-sqp-%d", sqd->task_pid);
6861 set_task_comm(current, buf);
6863 if (sqd->sq_cpu != -1)
6864 set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
6866 set_cpus_allowed_ptr(current, cpu_online_mask);
6867 current->flags |= PF_NO_SETAFFINITY;
6869 mutex_lock(&sqd->lock);
6872 bool cap_entries, sqt_spin, needs_sched;
6874 if (io_sqd_events_pending(sqd) || signal_pending(current)) {
6875 if (io_sqd_handle_event(sqd))
6877 timeout = jiffies + sqd->sq_thread_idle;
6882 cap_entries = !list_is_singular(&sqd->ctx_list);
6883 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6884 const struct cred *creds = NULL;
6886 if (ctx->sq_creds != current_cred())
6887 creds = override_creds(ctx->sq_creds);
6888 ret = __io_sq_thread(ctx, cap_entries);
6890 revert_creds(creds);
6891 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
6895 if (sqt_spin || !time_after(jiffies, timeout)) {
6899 timeout = jiffies + sqd->sq_thread_idle;
6903 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
6904 if (!io_sqd_events_pending(sqd)) {
6906 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6907 io_ring_set_wakeup_flag(ctx);
6909 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6910 !list_empty_careful(&ctx->iopoll_list)) {
6911 needs_sched = false;
6914 if (io_sqring_entries(ctx)) {
6915 needs_sched = false;
6921 mutex_unlock(&sqd->lock);
6923 mutex_lock(&sqd->lock);
6925 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6926 io_ring_clear_wakeup_flag(ctx);
6929 finish_wait(&sqd->wait, &wait);
6930 timeout = jiffies + sqd->sq_thread_idle;
6933 io_uring_cancel_generic(true, sqd);
6935 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6936 io_ring_set_wakeup_flag(ctx);
6938 mutex_unlock(&sqd->lock);
6940 complete(&sqd->exited);
6944 struct io_wait_queue {
6945 struct wait_queue_entry wq;
6946 struct io_ring_ctx *ctx;
6948 unsigned nr_timeouts;
6951 static inline bool io_should_wake(struct io_wait_queue *iowq)
6953 struct io_ring_ctx *ctx = iowq->ctx;
6956 * Wake up if we have enough events, or if a timeout occurred since we
6957 * started waiting. For timeouts, we always want to return to userspace,
6958 * regardless of event count.
6960 return io_cqring_events(ctx) >= iowq->to_wait ||
6961 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6964 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6965 int wake_flags, void *key)
6967 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6971 * Cannot safely flush overflowed CQEs from here, ensure we wake up
6972 * the task, and the next invocation will do it.
6974 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->check_cq_overflow))
6975 return autoremove_wake_function(curr, mode, wake_flags, key);
6979 static int io_run_task_work_sig(void)
6981 if (io_run_task_work())
6983 if (!signal_pending(current))
6985 if (test_thread_flag(TIF_NOTIFY_SIGNAL))
6986 return -ERESTARTSYS;
6990 /* when returns >0, the caller should retry */
6991 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
6992 struct io_wait_queue *iowq,
6993 signed long *timeout)
6997 /* make sure we run task_work before checking for signals */
6998 ret = io_run_task_work_sig();
6999 if (ret || io_should_wake(iowq))
7001 /* let the caller flush overflows, retry */
7002 if (test_bit(0, &ctx->check_cq_overflow))
7005 *timeout = schedule_timeout(*timeout);
7006 return !*timeout ? -ETIME : 1;
7010 * Wait until events become available, if we don't already have some. The
7011 * application must reap them itself, as they reside on the shared cq ring.
7013 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
7014 const sigset_t __user *sig, size_t sigsz,
7015 struct __kernel_timespec __user *uts)
7017 struct io_wait_queue iowq = {
7020 .func = io_wake_function,
7021 .entry = LIST_HEAD_INIT(iowq.wq.entry),
7024 .to_wait = min_events,
7026 struct io_rings *rings = ctx->rings;
7027 signed long timeout = MAX_SCHEDULE_TIMEOUT;
7031 io_cqring_overflow_flush(ctx, false);
7032 if (io_cqring_events(ctx) >= min_events)
7034 if (!io_run_task_work())
7039 #ifdef CONFIG_COMPAT
7040 if (in_compat_syscall())
7041 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
7045 ret = set_user_sigmask(sig, sigsz);
7052 struct timespec64 ts;
7054 if (get_timespec64(&ts, uts))
7056 timeout = timespec64_to_jiffies(&ts);
7059 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
7060 trace_io_uring_cqring_wait(ctx, min_events);
7062 /* if we can't even flush overflow, don't wait for more */
7063 if (!io_cqring_overflow_flush(ctx, false)) {
7067 prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq,
7068 TASK_INTERRUPTIBLE);
7069 ret = io_cqring_wait_schedule(ctx, &iowq, &timeout);
7070 finish_wait(&ctx->cq_wait, &iowq.wq);
7074 restore_saved_sigmask_unless(ret == -EINTR);
7076 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
7079 static void io_free_page_table(void **table, size_t size)
7081 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7083 for (i = 0; i < nr_tables; i++)
7088 static void **io_alloc_page_table(size_t size)
7090 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7091 size_t init_size = size;
7094 table = kcalloc(nr_tables, sizeof(*table), GFP_KERNEL);
7098 for (i = 0; i < nr_tables; i++) {
7099 unsigned int this_size = min_t(size_t, size, PAGE_SIZE);
7101 table[i] = kzalloc(this_size, GFP_KERNEL);
7103 io_free_page_table(table, init_size);
7111 static inline void io_rsrc_ref_lock(struct io_ring_ctx *ctx)
7113 spin_lock_bh(&ctx->rsrc_ref_lock);
7116 static inline void io_rsrc_ref_unlock(struct io_ring_ctx *ctx)
7118 spin_unlock_bh(&ctx->rsrc_ref_lock);
7121 static void io_rsrc_node_destroy(struct io_rsrc_node *ref_node)
7123 percpu_ref_exit(&ref_node->refs);
7127 static void io_rsrc_node_switch(struct io_ring_ctx *ctx,
7128 struct io_rsrc_data *data_to_kill)
7130 WARN_ON_ONCE(!ctx->rsrc_backup_node);
7131 WARN_ON_ONCE(data_to_kill && !ctx->rsrc_node);
7134 struct io_rsrc_node *rsrc_node = ctx->rsrc_node;
7136 rsrc_node->rsrc_data = data_to_kill;
7137 io_rsrc_ref_lock(ctx);
7138 list_add_tail(&rsrc_node->node, &ctx->rsrc_ref_list);
7139 io_rsrc_ref_unlock(ctx);
7141 atomic_inc(&data_to_kill->refs);
7142 percpu_ref_kill(&rsrc_node->refs);
7143 ctx->rsrc_node = NULL;
7146 if (!ctx->rsrc_node) {
7147 ctx->rsrc_node = ctx->rsrc_backup_node;
7148 ctx->rsrc_backup_node = NULL;
7152 static int io_rsrc_node_switch_start(struct io_ring_ctx *ctx)
7154 if (ctx->rsrc_backup_node)
7156 ctx->rsrc_backup_node = io_rsrc_node_alloc(ctx);
7157 return ctx->rsrc_backup_node ? 0 : -ENOMEM;
7160 static int io_rsrc_ref_quiesce(struct io_rsrc_data *data, struct io_ring_ctx *ctx)
7164 /* As we may drop ->uring_lock, other task may have started quiesce */
7168 data->quiesce = true;
7170 ret = io_rsrc_node_switch_start(ctx);
7173 io_rsrc_node_switch(ctx, data);
7175 /* kill initial ref, already quiesced if zero */
7176 if (atomic_dec_and_test(&data->refs))
7178 flush_delayed_work(&ctx->rsrc_put_work);
7179 ret = wait_for_completion_interruptible(&data->done);
7183 atomic_inc(&data->refs);
7184 /* wait for all works potentially completing data->done */
7185 flush_delayed_work(&ctx->rsrc_put_work);
7186 reinit_completion(&data->done);
7188 mutex_unlock(&ctx->uring_lock);
7189 ret = io_run_task_work_sig();
7190 mutex_lock(&ctx->uring_lock);
7192 data->quiesce = false;
7197 static u64 *io_get_tag_slot(struct io_rsrc_data *data, unsigned int idx)
7199 unsigned int off = idx & IO_RSRC_TAG_TABLE_MASK;
7200 unsigned int table_idx = idx >> IO_RSRC_TAG_TABLE_SHIFT;
7202 return &data->tags[table_idx][off];
7205 static void io_rsrc_data_free(struct io_rsrc_data *data)
7207 size_t size = data->nr * sizeof(data->tags[0][0]);
7210 io_free_page_table((void **)data->tags, size);
7214 static int io_rsrc_data_alloc(struct io_ring_ctx *ctx, rsrc_put_fn *do_put,
7215 u64 __user *utags, unsigned nr,
7216 struct io_rsrc_data **pdata)
7218 struct io_rsrc_data *data;
7222 data = kzalloc(sizeof(*data), GFP_KERNEL);
7225 data->tags = (u64 **)io_alloc_page_table(nr * sizeof(data->tags[0][0]));
7233 data->do_put = do_put;
7236 for (i = 0; i < nr; i++) {
7237 u64 *tag_slot = io_get_tag_slot(data, i);
7239 if (copy_from_user(tag_slot, &utags[i],
7245 atomic_set(&data->refs, 1);
7246 init_completion(&data->done);
7250 io_rsrc_data_free(data);
7254 static bool io_alloc_file_tables(struct io_file_table *table, unsigned nr_files)
7256 size_t size = nr_files * sizeof(struct io_fixed_file);
7258 table->files = (struct io_fixed_file **)io_alloc_page_table(size);
7259 return !!table->files;
7262 static void io_free_file_tables(struct io_file_table *table, unsigned nr_files)
7264 size_t size = nr_files * sizeof(struct io_fixed_file);
7266 io_free_page_table((void **)table->files, size);
7267 table->files = NULL;
7270 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
7272 #if defined(CONFIG_UNIX)
7273 if (ctx->ring_sock) {
7274 struct sock *sock = ctx->ring_sock->sk;
7275 struct sk_buff *skb;
7277 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
7283 for (i = 0; i < ctx->nr_user_files; i++) {
7286 file = io_file_from_index(ctx, i);
7291 io_free_file_tables(&ctx->file_table, ctx->nr_user_files);
7292 io_rsrc_data_free(ctx->file_data);
7293 ctx->file_data = NULL;
7294 ctx->nr_user_files = 0;
7297 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7301 if (!ctx->file_data)
7303 ret = io_rsrc_ref_quiesce(ctx->file_data, ctx);
7305 __io_sqe_files_unregister(ctx);
7309 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7310 __releases(&sqd->lock)
7312 WARN_ON_ONCE(sqd->thread == current);
7315 * Do the dance but not conditional clear_bit() because it'd race with
7316 * other threads incrementing park_pending and setting the bit.
7318 clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7319 if (atomic_dec_return(&sqd->park_pending))
7320 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7321 mutex_unlock(&sqd->lock);
7324 static void io_sq_thread_park(struct io_sq_data *sqd)
7325 __acquires(&sqd->lock)
7327 WARN_ON_ONCE(sqd->thread == current);
7329 atomic_inc(&sqd->park_pending);
7330 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7331 mutex_lock(&sqd->lock);
7333 wake_up_process(sqd->thread);
7336 static void io_sq_thread_stop(struct io_sq_data *sqd)
7338 WARN_ON_ONCE(sqd->thread == current);
7339 WARN_ON_ONCE(test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state));
7341 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7342 mutex_lock(&sqd->lock);
7344 wake_up_process(sqd->thread);
7345 mutex_unlock(&sqd->lock);
7346 wait_for_completion(&sqd->exited);
7349 static void io_put_sq_data(struct io_sq_data *sqd)
7351 if (refcount_dec_and_test(&sqd->refs)) {
7352 WARN_ON_ONCE(atomic_read(&sqd->park_pending));
7354 io_sq_thread_stop(sqd);
7359 static void io_sq_thread_finish(struct io_ring_ctx *ctx)
7361 struct io_sq_data *sqd = ctx->sq_data;
7364 io_sq_thread_park(sqd);
7365 list_del_init(&ctx->sqd_list);
7366 io_sqd_update_thread_idle(sqd);
7367 io_sq_thread_unpark(sqd);
7369 io_put_sq_data(sqd);
7370 ctx->sq_data = NULL;
7374 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7376 struct io_ring_ctx *ctx_attach;
7377 struct io_sq_data *sqd;
7380 f = fdget(p->wq_fd);
7382 return ERR_PTR(-ENXIO);
7383 if (f.file->f_op != &io_uring_fops) {
7385 return ERR_PTR(-EINVAL);
7388 ctx_attach = f.file->private_data;
7389 sqd = ctx_attach->sq_data;
7392 return ERR_PTR(-EINVAL);
7394 if (sqd->task_tgid != current->tgid) {
7396 return ERR_PTR(-EPERM);
7399 refcount_inc(&sqd->refs);
7404 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p,
7407 struct io_sq_data *sqd;
7410 if (p->flags & IORING_SETUP_ATTACH_WQ) {
7411 sqd = io_attach_sq_data(p);
7416 /* fall through for EPERM case, setup new sqd/task */
7417 if (PTR_ERR(sqd) != -EPERM)
7421 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7423 return ERR_PTR(-ENOMEM);
7425 atomic_set(&sqd->park_pending, 0);
7426 refcount_set(&sqd->refs, 1);
7427 INIT_LIST_HEAD(&sqd->ctx_list);
7428 mutex_init(&sqd->lock);
7429 init_waitqueue_head(&sqd->wait);
7430 init_completion(&sqd->exited);
7434 #if defined(CONFIG_UNIX)
7436 * Ensure the UNIX gc is aware of our file set, so we are certain that
7437 * the io_uring can be safely unregistered on process exit, even if we have
7438 * loops in the file referencing.
7440 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7442 struct sock *sk = ctx->ring_sock->sk;
7443 struct scm_fp_list *fpl;
7444 struct sk_buff *skb;
7447 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7451 skb = alloc_skb(0, GFP_KERNEL);
7460 fpl->user = get_uid(current_user());
7461 for (i = 0; i < nr; i++) {
7462 struct file *file = io_file_from_index(ctx, i + offset);
7466 fpl->fp[nr_files] = get_file(file);
7467 unix_inflight(fpl->user, fpl->fp[nr_files]);
7472 fpl->max = SCM_MAX_FD;
7473 fpl->count = nr_files;
7474 UNIXCB(skb).fp = fpl;
7475 skb->destructor = unix_destruct_scm;
7476 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7477 skb_queue_head(&sk->sk_receive_queue, skb);
7479 for (i = 0; i < nr_files; i++)
7490 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7491 * causes regular reference counting to break down. We rely on the UNIX
7492 * garbage collection to take care of this problem for us.
7494 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7496 unsigned left, total;
7500 left = ctx->nr_user_files;
7502 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7504 ret = __io_sqe_files_scm(ctx, this_files, total);
7508 total += this_files;
7514 while (total < ctx->nr_user_files) {
7515 struct file *file = io_file_from_index(ctx, total);
7525 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7531 static void io_rsrc_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
7533 struct file *file = prsrc->file;
7534 #if defined(CONFIG_UNIX)
7535 struct sock *sock = ctx->ring_sock->sk;
7536 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7537 struct sk_buff *skb;
7540 __skb_queue_head_init(&list);
7543 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7544 * remove this entry and rearrange the file array.
7546 skb = skb_dequeue(head);
7548 struct scm_fp_list *fp;
7550 fp = UNIXCB(skb).fp;
7551 for (i = 0; i < fp->count; i++) {
7554 if (fp->fp[i] != file)
7557 unix_notinflight(fp->user, fp->fp[i]);
7558 left = fp->count - 1 - i;
7560 memmove(&fp->fp[i], &fp->fp[i + 1],
7561 left * sizeof(struct file *));
7568 __skb_queue_tail(&list, skb);
7578 __skb_queue_tail(&list, skb);
7580 skb = skb_dequeue(head);
7583 if (skb_peek(&list)) {
7584 spin_lock_irq(&head->lock);
7585 while ((skb = __skb_dequeue(&list)) != NULL)
7586 __skb_queue_tail(head, skb);
7587 spin_unlock_irq(&head->lock);
7594 static void __io_rsrc_put_work(struct io_rsrc_node *ref_node)
7596 struct io_rsrc_data *rsrc_data = ref_node->rsrc_data;
7597 struct io_ring_ctx *ctx = rsrc_data->ctx;
7598 struct io_rsrc_put *prsrc, *tmp;
7600 list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
7601 list_del(&prsrc->list);
7604 bool lock_ring = ctx->flags & IORING_SETUP_IOPOLL;
7606 io_ring_submit_lock(ctx, lock_ring);
7607 spin_lock_irq(&ctx->completion_lock);
7608 io_cqring_fill_event(ctx, prsrc->tag, 0, 0);
7610 io_commit_cqring(ctx);
7611 spin_unlock_irq(&ctx->completion_lock);
7612 io_cqring_ev_posted(ctx);
7613 io_ring_submit_unlock(ctx, lock_ring);
7616 rsrc_data->do_put(ctx, prsrc);
7620 io_rsrc_node_destroy(ref_node);
7621 if (atomic_dec_and_test(&rsrc_data->refs))
7622 complete(&rsrc_data->done);
7625 static void io_rsrc_put_work(struct work_struct *work)
7627 struct io_ring_ctx *ctx;
7628 struct llist_node *node;
7630 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
7631 node = llist_del_all(&ctx->rsrc_put_llist);
7634 struct io_rsrc_node *ref_node;
7635 struct llist_node *next = node->next;
7637 ref_node = llist_entry(node, struct io_rsrc_node, llist);
7638 __io_rsrc_put_work(ref_node);
7643 static void io_rsrc_node_ref_zero(struct percpu_ref *ref)
7645 struct io_rsrc_node *node = container_of(ref, struct io_rsrc_node, refs);
7646 struct io_ring_ctx *ctx = node->rsrc_data->ctx;
7647 bool first_add = false;
7649 io_rsrc_ref_lock(ctx);
7652 while (!list_empty(&ctx->rsrc_ref_list)) {
7653 node = list_first_entry(&ctx->rsrc_ref_list,
7654 struct io_rsrc_node, node);
7655 /* recycle ref nodes in order */
7658 list_del(&node->node);
7659 first_add |= llist_add(&node->llist, &ctx->rsrc_put_llist);
7661 io_rsrc_ref_unlock(ctx);
7664 mod_delayed_work(system_wq, &ctx->rsrc_put_work, HZ);
7667 static struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx)
7669 struct io_rsrc_node *ref_node;
7671 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7675 if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
7680 INIT_LIST_HEAD(&ref_node->node);
7681 INIT_LIST_HEAD(&ref_node->rsrc_list);
7682 ref_node->done = false;
7686 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7687 unsigned nr_args, u64 __user *tags)
7689 __s32 __user *fds = (__s32 __user *) arg;
7698 if (nr_args > IORING_MAX_FIXED_FILES)
7700 ret = io_rsrc_node_switch_start(ctx);
7703 ret = io_rsrc_data_alloc(ctx, io_rsrc_file_put, tags, nr_args,
7709 if (!io_alloc_file_tables(&ctx->file_table, nr_args))
7712 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7713 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7717 /* allow sparse sets */
7720 if (unlikely(*io_get_tag_slot(ctx->file_data, i)))
7727 if (unlikely(!file))
7731 * Don't allow io_uring instances to be registered. If UNIX
7732 * isn't enabled, then this causes a reference cycle and this
7733 * instance can never get freed. If UNIX is enabled we'll
7734 * handle it just fine, but there's still no point in allowing
7735 * a ring fd as it doesn't support regular read/write anyway.
7737 if (file->f_op == &io_uring_fops) {
7741 io_fixed_file_set(io_fixed_file_slot(&ctx->file_table, i), file);
7744 ret = io_sqe_files_scm(ctx);
7746 __io_sqe_files_unregister(ctx);
7750 io_rsrc_node_switch(ctx, NULL);
7753 for (i = 0; i < ctx->nr_user_files; i++) {
7754 file = io_file_from_index(ctx, i);
7758 io_free_file_tables(&ctx->file_table, nr_args);
7759 ctx->nr_user_files = 0;
7761 io_rsrc_data_free(ctx->file_data);
7762 ctx->file_data = NULL;
7766 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7769 #if defined(CONFIG_UNIX)
7770 struct sock *sock = ctx->ring_sock->sk;
7771 struct sk_buff_head *head = &sock->sk_receive_queue;
7772 struct sk_buff *skb;
7775 * See if we can merge this file into an existing skb SCM_RIGHTS
7776 * file set. If there's no room, fall back to allocating a new skb
7777 * and filling it in.
7779 spin_lock_irq(&head->lock);
7780 skb = skb_peek(head);
7782 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7784 if (fpl->count < SCM_MAX_FD) {
7785 __skb_unlink(skb, head);
7786 spin_unlock_irq(&head->lock);
7787 fpl->fp[fpl->count] = get_file(file);
7788 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7790 spin_lock_irq(&head->lock);
7791 __skb_queue_head(head, skb);
7796 spin_unlock_irq(&head->lock);
7803 return __io_sqe_files_scm(ctx, 1, index);
7809 static int io_queue_rsrc_removal(struct io_rsrc_data *data, unsigned idx,
7810 struct io_rsrc_node *node, void *rsrc)
7812 struct io_rsrc_put *prsrc;
7814 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
7818 prsrc->tag = *io_get_tag_slot(data, idx);
7820 list_add(&prsrc->list, &node->rsrc_list);
7824 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7825 struct io_uring_rsrc_update2 *up,
7828 u64 __user *tags = u64_to_user_ptr(up->tags);
7829 __s32 __user *fds = u64_to_user_ptr(up->data);
7830 struct io_rsrc_data *data = ctx->file_data;
7831 struct io_fixed_file *file_slot;
7835 bool needs_switch = false;
7837 if (!ctx->file_data)
7839 if (up->offset + nr_args > ctx->nr_user_files)
7842 for (done = 0; done < nr_args; done++) {
7845 if ((tags && copy_from_user(&tag, &tags[done], sizeof(tag))) ||
7846 copy_from_user(&fd, &fds[done], sizeof(fd))) {
7850 if ((fd == IORING_REGISTER_FILES_SKIP || fd == -1) && tag) {
7854 if (fd == IORING_REGISTER_FILES_SKIP)
7857 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
7858 file_slot = io_fixed_file_slot(&ctx->file_table, i);
7860 if (file_slot->file_ptr) {
7861 file = (struct file *)(file_slot->file_ptr & FFS_MASK);
7862 err = io_queue_rsrc_removal(data, up->offset + done,
7863 ctx->rsrc_node, file);
7866 file_slot->file_ptr = 0;
7867 needs_switch = true;
7876 * Don't allow io_uring instances to be registered. If
7877 * UNIX isn't enabled, then this causes a reference
7878 * cycle and this instance can never get freed. If UNIX
7879 * is enabled we'll handle it just fine, but there's
7880 * still no point in allowing a ring fd as it doesn't
7881 * support regular read/write anyway.
7883 if (file->f_op == &io_uring_fops) {
7888 *io_get_tag_slot(data, up->offset + done) = tag;
7889 io_fixed_file_set(file_slot, file);
7890 err = io_sqe_file_register(ctx, file, i);
7892 file_slot->file_ptr = 0;
7900 io_rsrc_node_switch(ctx, data);
7901 return done ? done : err;
7904 static struct io_wq_work *io_free_work(struct io_wq_work *work)
7906 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7908 req = io_put_req_find_next(req);
7909 return req ? &req->work : NULL;
7912 static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx,
7913 struct task_struct *task)
7915 struct io_wq_hash *hash;
7916 struct io_wq_data data;
7917 unsigned int concurrency;
7919 hash = ctx->hash_map;
7921 hash = kzalloc(sizeof(*hash), GFP_KERNEL);
7923 return ERR_PTR(-ENOMEM);
7924 refcount_set(&hash->refs, 1);
7925 init_waitqueue_head(&hash->wait);
7926 ctx->hash_map = hash;
7931 data.free_work = io_free_work;
7932 data.do_work = io_wq_submit_work;
7934 /* Do QD, or 4 * CPUS, whatever is smallest */
7935 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7937 return io_wq_create(concurrency, &data);
7940 static int io_uring_alloc_task_context(struct task_struct *task,
7941 struct io_ring_ctx *ctx)
7943 struct io_uring_task *tctx;
7946 tctx = kzalloc(sizeof(*tctx), GFP_KERNEL);
7947 if (unlikely(!tctx))
7950 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
7951 if (unlikely(ret)) {
7956 tctx->io_wq = io_init_wq_offload(ctx, task);
7957 if (IS_ERR(tctx->io_wq)) {
7958 ret = PTR_ERR(tctx->io_wq);
7959 percpu_counter_destroy(&tctx->inflight);
7965 init_waitqueue_head(&tctx->wait);
7966 atomic_set(&tctx->in_idle, 0);
7967 atomic_set(&tctx->inflight_tracked, 0);
7968 task->io_uring = tctx;
7969 spin_lock_init(&tctx->task_lock);
7970 INIT_WQ_LIST(&tctx->task_list);
7971 init_task_work(&tctx->task_work, tctx_task_work);
7975 void __io_uring_free(struct task_struct *tsk)
7977 struct io_uring_task *tctx = tsk->io_uring;
7979 WARN_ON_ONCE(!xa_empty(&tctx->xa));
7980 WARN_ON_ONCE(tctx->io_wq);
7981 WARN_ON_ONCE(tctx->cached_refs);
7983 percpu_counter_destroy(&tctx->inflight);
7985 tsk->io_uring = NULL;
7988 static int io_sq_offload_create(struct io_ring_ctx *ctx,
7989 struct io_uring_params *p)
7993 /* Retain compatibility with failing for an invalid attach attempt */
7994 if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
7995 IORING_SETUP_ATTACH_WQ) {
7998 f = fdget(p->wq_fd);
8002 if (f.file->f_op != &io_uring_fops)
8005 if (ctx->flags & IORING_SETUP_SQPOLL) {
8006 struct task_struct *tsk;
8007 struct io_sq_data *sqd;
8010 sqd = io_get_sq_data(p, &attached);
8016 ctx->sq_creds = get_current_cred();
8018 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
8019 if (!ctx->sq_thread_idle)
8020 ctx->sq_thread_idle = HZ;
8022 io_sq_thread_park(sqd);
8023 list_add(&ctx->sqd_list, &sqd->ctx_list);
8024 io_sqd_update_thread_idle(sqd);
8025 /* don't attach to a dying SQPOLL thread, would be racy */
8026 ret = (attached && !sqd->thread) ? -ENXIO : 0;
8027 io_sq_thread_unpark(sqd);
8034 if (p->flags & IORING_SETUP_SQ_AFF) {
8035 int cpu = p->sq_thread_cpu;
8038 if (cpu >= nr_cpu_ids || !cpu_online(cpu))
8045 sqd->task_pid = current->pid;
8046 sqd->task_tgid = current->tgid;
8047 tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE);
8054 ret = io_uring_alloc_task_context(tsk, ctx);
8055 wake_up_new_task(tsk);
8058 } else if (p->flags & IORING_SETUP_SQ_AFF) {
8059 /* Can't have SQ_AFF without SQPOLL */
8066 complete(&ctx->sq_data->exited);
8068 io_sq_thread_finish(ctx);
8072 static inline void __io_unaccount_mem(struct user_struct *user,
8073 unsigned long nr_pages)
8075 atomic_long_sub(nr_pages, &user->locked_vm);
8078 static inline int __io_account_mem(struct user_struct *user,
8079 unsigned long nr_pages)
8081 unsigned long page_limit, cur_pages, new_pages;
8083 /* Don't allow more pages than we can safely lock */
8084 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
8087 cur_pages = atomic_long_read(&user->locked_vm);
8088 new_pages = cur_pages + nr_pages;
8089 if (new_pages > page_limit)
8091 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
8092 new_pages) != cur_pages);
8097 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8100 __io_unaccount_mem(ctx->user, nr_pages);
8102 if (ctx->mm_account)
8103 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
8106 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8111 ret = __io_account_mem(ctx->user, nr_pages);
8116 if (ctx->mm_account)
8117 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
8122 static void io_mem_free(void *ptr)
8129 page = virt_to_head_page(ptr);
8130 if (put_page_testzero(page))
8131 free_compound_page(page);
8134 static void *io_mem_alloc(size_t size)
8136 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
8137 __GFP_NORETRY | __GFP_ACCOUNT;
8139 return (void *) __get_free_pages(gfp_flags, get_order(size));
8142 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8145 struct io_rings *rings;
8146 size_t off, sq_array_size;
8148 off = struct_size(rings, cqes, cq_entries);
8149 if (off == SIZE_MAX)
8153 off = ALIGN(off, SMP_CACHE_BYTES);
8161 sq_array_size = array_size(sizeof(u32), sq_entries);
8162 if (sq_array_size == SIZE_MAX)
8165 if (check_add_overflow(off, sq_array_size, &off))
8171 static void io_buffer_unmap(struct io_ring_ctx *ctx, struct io_mapped_ubuf **slot)
8173 struct io_mapped_ubuf *imu = *slot;
8176 if (imu != ctx->dummy_ubuf) {
8177 for (i = 0; i < imu->nr_bvecs; i++)
8178 unpin_user_page(imu->bvec[i].bv_page);
8179 if (imu->acct_pages)
8180 io_unaccount_mem(ctx, imu->acct_pages);
8186 static void io_rsrc_buf_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
8188 io_buffer_unmap(ctx, &prsrc->buf);
8192 static void __io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8196 for (i = 0; i < ctx->nr_user_bufs; i++)
8197 io_buffer_unmap(ctx, &ctx->user_bufs[i]);
8198 kfree(ctx->user_bufs);
8199 io_rsrc_data_free(ctx->buf_data);
8200 ctx->user_bufs = NULL;
8201 ctx->buf_data = NULL;
8202 ctx->nr_user_bufs = 0;
8205 static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8212 ret = io_rsrc_ref_quiesce(ctx->buf_data, ctx);
8214 __io_sqe_buffers_unregister(ctx);
8218 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8219 void __user *arg, unsigned index)
8221 struct iovec __user *src;
8223 #ifdef CONFIG_COMPAT
8225 struct compat_iovec __user *ciovs;
8226 struct compat_iovec ciov;
8228 ciovs = (struct compat_iovec __user *) arg;
8229 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8232 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8233 dst->iov_len = ciov.iov_len;
8237 src = (struct iovec __user *) arg;
8238 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8244 * Not super efficient, but this is just a registration time. And we do cache
8245 * the last compound head, so generally we'll only do a full search if we don't
8248 * We check if the given compound head page has already been accounted, to
8249 * avoid double accounting it. This allows us to account the full size of the
8250 * page, not just the constituent pages of a huge page.
8252 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8253 int nr_pages, struct page *hpage)
8257 /* check current page array */
8258 for (i = 0; i < nr_pages; i++) {
8259 if (!PageCompound(pages[i]))
8261 if (compound_head(pages[i]) == hpage)
8265 /* check previously registered pages */
8266 for (i = 0; i < ctx->nr_user_bufs; i++) {
8267 struct io_mapped_ubuf *imu = ctx->user_bufs[i];
8269 for (j = 0; j < imu->nr_bvecs; j++) {
8270 if (!PageCompound(imu->bvec[j].bv_page))
8272 if (compound_head(imu->bvec[j].bv_page) == hpage)
8280 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8281 int nr_pages, struct io_mapped_ubuf *imu,
8282 struct page **last_hpage)
8286 imu->acct_pages = 0;
8287 for (i = 0; i < nr_pages; i++) {
8288 if (!PageCompound(pages[i])) {
8293 hpage = compound_head(pages[i]);
8294 if (hpage == *last_hpage)
8296 *last_hpage = hpage;
8297 if (headpage_already_acct(ctx, pages, i, hpage))
8299 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8303 if (!imu->acct_pages)
8306 ret = io_account_mem(ctx, imu->acct_pages);
8308 imu->acct_pages = 0;
8312 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
8313 struct io_mapped_ubuf **pimu,
8314 struct page **last_hpage)
8316 struct io_mapped_ubuf *imu = NULL;
8317 struct vm_area_struct **vmas = NULL;
8318 struct page **pages = NULL;
8319 unsigned long off, start, end, ubuf;
8321 int ret, pret, nr_pages, i;
8323 if (!iov->iov_base) {
8324 *pimu = ctx->dummy_ubuf;
8328 ubuf = (unsigned long) iov->iov_base;
8329 end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8330 start = ubuf >> PAGE_SHIFT;
8331 nr_pages = end - start;
8336 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
8340 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
8345 imu = kvmalloc(struct_size(imu, bvec, nr_pages), GFP_KERNEL);
8350 mmap_read_lock(current->mm);
8351 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
8353 if (pret == nr_pages) {
8354 /* don't support file backed memory */
8355 for (i = 0; i < nr_pages; i++) {
8356 struct vm_area_struct *vma = vmas[i];
8358 if (vma_is_shmem(vma))
8361 !is_file_hugepages(vma->vm_file)) {
8367 ret = pret < 0 ? pret : -EFAULT;
8369 mmap_read_unlock(current->mm);
8372 * if we did partial map, or found file backed vmas,
8373 * release any pages we did get
8376 unpin_user_pages(pages, pret);
8380 ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
8382 unpin_user_pages(pages, pret);
8386 off = ubuf & ~PAGE_MASK;
8387 size = iov->iov_len;
8388 for (i = 0; i < nr_pages; i++) {
8391 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8392 imu->bvec[i].bv_page = pages[i];
8393 imu->bvec[i].bv_len = vec_len;
8394 imu->bvec[i].bv_offset = off;
8398 /* store original address for later verification */
8400 imu->ubuf_end = ubuf + iov->iov_len;
8401 imu->nr_bvecs = nr_pages;
8412 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
8414 ctx->user_bufs = kcalloc(nr_args, sizeof(*ctx->user_bufs), GFP_KERNEL);
8415 return ctx->user_bufs ? 0 : -ENOMEM;
8418 static int io_buffer_validate(struct iovec *iov)
8420 unsigned long tmp, acct_len = iov->iov_len + (PAGE_SIZE - 1);
8423 * Don't impose further limits on the size and buffer
8424 * constraints here, we'll -EINVAL later when IO is
8425 * submitted if they are wrong.
8428 return iov->iov_len ? -EFAULT : 0;
8432 /* arbitrary limit, but we need something */
8433 if (iov->iov_len > SZ_1G)
8436 if (check_add_overflow((unsigned long)iov->iov_base, acct_len, &tmp))
8442 static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
8443 unsigned int nr_args, u64 __user *tags)
8445 struct page *last_hpage = NULL;
8446 struct io_rsrc_data *data;
8452 if (!nr_args || nr_args > IORING_MAX_REG_BUFFERS)
8454 ret = io_rsrc_node_switch_start(ctx);
8457 ret = io_rsrc_data_alloc(ctx, io_rsrc_buf_put, tags, nr_args, &data);
8460 ret = io_buffers_map_alloc(ctx, nr_args);
8462 io_rsrc_data_free(data);
8466 for (i = 0; i < nr_args; i++, ctx->nr_user_bufs++) {
8467 ret = io_copy_iov(ctx, &iov, arg, i);
8470 ret = io_buffer_validate(&iov);
8473 if (!iov.iov_base && *io_get_tag_slot(data, i)) {
8478 ret = io_sqe_buffer_register(ctx, &iov, &ctx->user_bufs[i],
8484 WARN_ON_ONCE(ctx->buf_data);
8486 ctx->buf_data = data;
8488 __io_sqe_buffers_unregister(ctx);
8490 io_rsrc_node_switch(ctx, NULL);
8494 static int __io_sqe_buffers_update(struct io_ring_ctx *ctx,
8495 struct io_uring_rsrc_update2 *up,
8496 unsigned int nr_args)
8498 u64 __user *tags = u64_to_user_ptr(up->tags);
8499 struct iovec iov, __user *iovs = u64_to_user_ptr(up->data);
8500 struct page *last_hpage = NULL;
8501 bool needs_switch = false;
8507 if (up->offset + nr_args > ctx->nr_user_bufs)
8510 for (done = 0; done < nr_args; done++) {
8511 struct io_mapped_ubuf *imu;
8512 int offset = up->offset + done;
8515 err = io_copy_iov(ctx, &iov, iovs, done);
8518 if (tags && copy_from_user(&tag, &tags[done], sizeof(tag))) {
8522 err = io_buffer_validate(&iov);
8525 if (!iov.iov_base && tag) {
8529 err = io_sqe_buffer_register(ctx, &iov, &imu, &last_hpage);
8533 i = array_index_nospec(offset, ctx->nr_user_bufs);
8534 if (ctx->user_bufs[i] != ctx->dummy_ubuf) {
8535 err = io_queue_rsrc_removal(ctx->buf_data, offset,
8536 ctx->rsrc_node, ctx->user_bufs[i]);
8537 if (unlikely(err)) {
8538 io_buffer_unmap(ctx, &imu);
8541 ctx->user_bufs[i] = NULL;
8542 needs_switch = true;
8545 ctx->user_bufs[i] = imu;
8546 *io_get_tag_slot(ctx->buf_data, offset) = tag;
8550 io_rsrc_node_switch(ctx, ctx->buf_data);
8551 return done ? done : err;
8554 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8556 __s32 __user *fds = arg;
8562 if (copy_from_user(&fd, fds, sizeof(*fds)))
8565 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8566 if (IS_ERR(ctx->cq_ev_fd)) {
8567 int ret = PTR_ERR(ctx->cq_ev_fd);
8568 ctx->cq_ev_fd = NULL;
8575 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8577 if (ctx->cq_ev_fd) {
8578 eventfd_ctx_put(ctx->cq_ev_fd);
8579 ctx->cq_ev_fd = NULL;
8586 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8588 struct io_buffer *buf;
8589 unsigned long index;
8591 xa_for_each(&ctx->io_buffers, index, buf)
8592 __io_remove_buffers(ctx, buf, index, -1U);
8595 static void io_req_cache_free(struct list_head *list, struct task_struct *tsk)
8597 struct io_kiocb *req, *nxt;
8599 list_for_each_entry_safe(req, nxt, list, compl.list) {
8600 if (tsk && req->task != tsk)
8602 list_del(&req->compl.list);
8603 kmem_cache_free(req_cachep, req);
8607 static void io_req_caches_free(struct io_ring_ctx *ctx)
8609 struct io_submit_state *submit_state = &ctx->submit_state;
8610 struct io_comp_state *cs = &ctx->submit_state.comp;
8612 mutex_lock(&ctx->uring_lock);
8614 if (submit_state->free_reqs) {
8615 kmem_cache_free_bulk(req_cachep, submit_state->free_reqs,
8616 submit_state->reqs);
8617 submit_state->free_reqs = 0;
8620 io_flush_cached_locked_reqs(ctx, cs);
8621 io_req_cache_free(&cs->free_list, NULL);
8622 mutex_unlock(&ctx->uring_lock);
8625 static bool io_wait_rsrc_data(struct io_rsrc_data *data)
8629 if (!atomic_dec_and_test(&data->refs))
8630 wait_for_completion(&data->done);
8634 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8636 io_sq_thread_finish(ctx);
8638 if (ctx->mm_account) {
8639 mmdrop(ctx->mm_account);
8640 ctx->mm_account = NULL;
8643 mutex_lock(&ctx->uring_lock);
8644 if (io_wait_rsrc_data(ctx->buf_data))
8645 __io_sqe_buffers_unregister(ctx);
8646 if (io_wait_rsrc_data(ctx->file_data))
8647 __io_sqe_files_unregister(ctx);
8649 __io_cqring_overflow_flush(ctx, true);
8650 mutex_unlock(&ctx->uring_lock);
8651 io_eventfd_unregister(ctx);
8652 io_destroy_buffers(ctx);
8654 put_cred(ctx->sq_creds);
8656 /* there are no registered resources left, nobody uses it */
8658 io_rsrc_node_destroy(ctx->rsrc_node);
8659 if (ctx->rsrc_backup_node)
8660 io_rsrc_node_destroy(ctx->rsrc_backup_node);
8661 flush_delayed_work(&ctx->rsrc_put_work);
8663 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
8664 WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist));
8666 #if defined(CONFIG_UNIX)
8667 if (ctx->ring_sock) {
8668 ctx->ring_sock->file = NULL; /* so that iput() is called */
8669 sock_release(ctx->ring_sock);
8673 io_mem_free(ctx->rings);
8674 io_mem_free(ctx->sq_sqes);
8676 percpu_ref_exit(&ctx->refs);
8677 free_uid(ctx->user);
8678 io_req_caches_free(ctx);
8680 io_wq_put_hash(ctx->hash_map);
8681 kfree(ctx->cancel_hash);
8682 kfree(ctx->dummy_ubuf);
8686 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8688 struct io_ring_ctx *ctx = file->private_data;
8691 poll_wait(file, &ctx->poll_wait, wait);
8693 * synchronizes with barrier from wq_has_sleeper call in
8697 if (!io_sqring_full(ctx))
8698 mask |= EPOLLOUT | EPOLLWRNORM;
8701 * Don't flush cqring overflow list here, just do a simple check.
8702 * Otherwise there could possible be ABBA deadlock:
8705 * lock(&ctx->uring_lock);
8707 * lock(&ctx->uring_lock);
8710 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
8711 * pushs them to do the flush.
8713 if (io_cqring_events(ctx) || test_bit(0, &ctx->check_cq_overflow))
8714 mask |= EPOLLIN | EPOLLRDNORM;
8719 static int io_uring_fasync(int fd, struct file *file, int on)
8721 struct io_ring_ctx *ctx = file->private_data;
8723 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8726 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8728 const struct cred *creds;
8730 creds = xa_erase(&ctx->personalities, id);
8739 static inline bool io_run_ctx_fallback(struct io_ring_ctx *ctx)
8741 return io_run_task_work_head(&ctx->exit_task_work);
8744 struct io_tctx_exit {
8745 struct callback_head task_work;
8746 struct completion completion;
8747 struct io_ring_ctx *ctx;
8750 static void io_tctx_exit_cb(struct callback_head *cb)
8752 struct io_uring_task *tctx = current->io_uring;
8753 struct io_tctx_exit *work;
8755 work = container_of(cb, struct io_tctx_exit, task_work);
8757 * When @in_idle, we're in cancellation and it's racy to remove the
8758 * node. It'll be removed by the end of cancellation, just ignore it.
8760 if (!atomic_read(&tctx->in_idle))
8761 io_uring_del_tctx_node((unsigned long)work->ctx);
8762 complete(&work->completion);
8765 static bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
8767 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8769 return req->ctx == data;
8772 static void io_ring_exit_work(struct work_struct *work)
8774 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
8775 unsigned long timeout = jiffies + HZ * 60 * 5;
8776 struct io_tctx_exit exit;
8777 struct io_tctx_node *node;
8781 * If we're doing polled IO and end up having requests being
8782 * submitted async (out-of-line), then completions can come in while
8783 * we're waiting for refs to drop. We need to reap these manually,
8784 * as nobody else will be looking for them.
8787 io_uring_try_cancel_requests(ctx, NULL, true);
8789 struct io_sq_data *sqd = ctx->sq_data;
8790 struct task_struct *tsk;
8792 io_sq_thread_park(sqd);
8794 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
8795 io_wq_cancel_cb(tsk->io_uring->io_wq,
8796 io_cancel_ctx_cb, ctx, true);
8797 io_sq_thread_unpark(sqd);
8800 WARN_ON_ONCE(time_after(jiffies, timeout));
8801 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
8803 init_completion(&exit.completion);
8804 init_task_work(&exit.task_work, io_tctx_exit_cb);
8807 * Some may use context even when all refs and requests have been put,
8808 * and they are free to do so while still holding uring_lock or
8809 * completion_lock, see __io_req_task_submit(). Apart from other work,
8810 * this lock/unlock section also waits them to finish.
8812 mutex_lock(&ctx->uring_lock);
8813 while (!list_empty(&ctx->tctx_list)) {
8814 WARN_ON_ONCE(time_after(jiffies, timeout));
8816 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
8818 /* don't spin on a single task if cancellation failed */
8819 list_rotate_left(&ctx->tctx_list);
8820 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
8821 if (WARN_ON_ONCE(ret))
8823 wake_up_process(node->task);
8825 mutex_unlock(&ctx->uring_lock);
8826 wait_for_completion(&exit.completion);
8827 mutex_lock(&ctx->uring_lock);
8829 mutex_unlock(&ctx->uring_lock);
8830 spin_lock_irq(&ctx->completion_lock);
8831 spin_unlock_irq(&ctx->completion_lock);
8833 io_ring_ctx_free(ctx);
8836 /* Returns true if we found and killed one or more timeouts */
8837 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
8840 struct io_kiocb *req, *tmp;
8843 spin_lock_irq(&ctx->completion_lock);
8844 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
8845 if (io_match_task(req, tsk, cancel_all)) {
8846 io_kill_timeout(req, -ECANCELED);
8851 io_commit_cqring(ctx);
8852 spin_unlock_irq(&ctx->completion_lock);
8854 io_cqring_ev_posted(ctx);
8855 return canceled != 0;
8858 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8860 unsigned long index;
8861 struct creds *creds;
8863 mutex_lock(&ctx->uring_lock);
8864 percpu_ref_kill(&ctx->refs);
8866 __io_cqring_overflow_flush(ctx, true);
8867 xa_for_each(&ctx->personalities, index, creds)
8868 io_unregister_personality(ctx, index);
8869 mutex_unlock(&ctx->uring_lock);
8871 io_kill_timeouts(ctx, NULL, true);
8872 io_poll_remove_all(ctx, NULL, true);
8874 /* if we failed setting up the ctx, we might not have any rings */
8875 io_iopoll_try_reap_events(ctx);
8877 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8879 * Use system_unbound_wq to avoid spawning tons of event kworkers
8880 * if we're exiting a ton of rings at the same time. It just adds
8881 * noise and overhead, there's no discernable change in runtime
8882 * over using system_wq.
8884 queue_work(system_unbound_wq, &ctx->exit_work);
8887 static int io_uring_release(struct inode *inode, struct file *file)
8889 struct io_ring_ctx *ctx = file->private_data;
8891 file->private_data = NULL;
8892 io_ring_ctx_wait_and_kill(ctx);
8896 struct io_task_cancel {
8897 struct task_struct *task;
8901 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8903 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8904 struct io_task_cancel *cancel = data;
8907 if (!cancel->all && (req->flags & REQ_F_LINK_TIMEOUT)) {
8908 unsigned long flags;
8909 struct io_ring_ctx *ctx = req->ctx;
8911 /* protect against races with linked timeouts */
8912 spin_lock_irqsave(&ctx->completion_lock, flags);
8913 ret = io_match_task(req, cancel->task, cancel->all);
8914 spin_unlock_irqrestore(&ctx->completion_lock, flags);
8916 ret = io_match_task(req, cancel->task, cancel->all);
8921 static bool io_cancel_defer_files(struct io_ring_ctx *ctx,
8922 struct task_struct *task, bool cancel_all)
8924 struct io_defer_entry *de;
8927 spin_lock_irq(&ctx->completion_lock);
8928 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8929 if (io_match_task(de->req, task, cancel_all)) {
8930 list_cut_position(&list, &ctx->defer_list, &de->list);
8934 spin_unlock_irq(&ctx->completion_lock);
8935 if (list_empty(&list))
8938 while (!list_empty(&list)) {
8939 de = list_first_entry(&list, struct io_defer_entry, list);
8940 list_del_init(&de->list);
8941 io_req_complete_failed(de->req, -ECANCELED);
8947 static bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
8949 struct io_tctx_node *node;
8950 enum io_wq_cancel cret;
8953 mutex_lock(&ctx->uring_lock);
8954 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
8955 struct io_uring_task *tctx = node->task->io_uring;
8958 * io_wq will stay alive while we hold uring_lock, because it's
8959 * killed after ctx nodes, which requires to take the lock.
8961 if (!tctx || !tctx->io_wq)
8963 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
8964 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8966 mutex_unlock(&ctx->uring_lock);
8971 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
8972 struct task_struct *task,
8975 struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
8976 struct io_uring_task *tctx = task ? task->io_uring : NULL;
8979 enum io_wq_cancel cret;
8983 ret |= io_uring_try_cancel_iowq(ctx);
8984 } else if (tctx && tctx->io_wq) {
8986 * Cancels requests of all rings, not only @ctx, but
8987 * it's fine as the task is in exit/exec.
8989 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
8991 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8994 /* SQPOLL thread does its own polling */
8995 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
8996 (ctx->sq_data && ctx->sq_data->thread == current)) {
8997 while (!list_empty_careful(&ctx->iopoll_list)) {
8998 io_iopoll_try_reap_events(ctx);
9003 ret |= io_cancel_defer_files(ctx, task, cancel_all);
9004 ret |= io_poll_remove_all(ctx, task, cancel_all);
9005 ret |= io_kill_timeouts(ctx, task, cancel_all);
9006 ret |= io_run_task_work();
9007 ret |= io_run_ctx_fallback(ctx);
9014 static int __io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9016 struct io_uring_task *tctx = current->io_uring;
9017 struct io_tctx_node *node;
9020 if (unlikely(!tctx)) {
9021 ret = io_uring_alloc_task_context(current, ctx);
9024 tctx = current->io_uring;
9026 if (!xa_load(&tctx->xa, (unsigned long)ctx)) {
9027 node = kmalloc(sizeof(*node), GFP_KERNEL);
9031 node->task = current;
9033 ret = xa_err(xa_store(&tctx->xa, (unsigned long)ctx,
9040 mutex_lock(&ctx->uring_lock);
9041 list_add(&node->ctx_node, &ctx->tctx_list);
9042 mutex_unlock(&ctx->uring_lock);
9049 * Note that this task has used io_uring. We use it for cancelation purposes.
9051 static inline int io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9053 struct io_uring_task *tctx = current->io_uring;
9055 if (likely(tctx && tctx->last == ctx))
9057 return __io_uring_add_tctx_node(ctx);
9061 * Remove this io_uring_file -> task mapping.
9063 static void io_uring_del_tctx_node(unsigned long index)
9065 struct io_uring_task *tctx = current->io_uring;
9066 struct io_tctx_node *node;
9070 node = xa_erase(&tctx->xa, index);
9074 WARN_ON_ONCE(current != node->task);
9075 WARN_ON_ONCE(list_empty(&node->ctx_node));
9077 mutex_lock(&node->ctx->uring_lock);
9078 list_del(&node->ctx_node);
9079 mutex_unlock(&node->ctx->uring_lock);
9081 if (tctx->last == node->ctx)
9086 static void io_uring_clean_tctx(struct io_uring_task *tctx)
9088 struct io_wq *wq = tctx->io_wq;
9089 struct io_tctx_node *node;
9090 unsigned long index;
9092 xa_for_each(&tctx->xa, index, node)
9093 io_uring_del_tctx_node(index);
9096 * Must be after io_uring_del_task_file() (removes nodes under
9097 * uring_lock) to avoid race with io_uring_try_cancel_iowq().
9100 io_wq_put_and_exit(wq);
9104 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
9107 return atomic_read(&tctx->inflight_tracked);
9108 return percpu_counter_sum(&tctx->inflight);
9111 static void io_uring_drop_tctx_refs(struct task_struct *task)
9113 struct io_uring_task *tctx = task->io_uring;
9114 unsigned int refs = tctx->cached_refs;
9116 tctx->cached_refs = 0;
9117 percpu_counter_sub(&tctx->inflight, refs);
9118 put_task_struct_many(task, refs);
9122 * Find any io_uring ctx that this task has registered or done IO on, and cancel
9123 * requests. @sqd should be not-null IIF it's an SQPOLL thread cancellation.
9125 static void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
9127 struct io_uring_task *tctx = current->io_uring;
9128 struct io_ring_ctx *ctx;
9132 WARN_ON_ONCE(sqd && sqd->thread != current);
9134 if (!current->io_uring)
9137 io_wq_exit_start(tctx->io_wq);
9139 io_uring_drop_tctx_refs(current);
9140 atomic_inc(&tctx->in_idle);
9142 /* read completions before cancelations */
9143 inflight = tctx_inflight(tctx, !cancel_all);
9148 struct io_tctx_node *node;
9149 unsigned long index;
9151 xa_for_each(&tctx->xa, index, node) {
9152 /* sqpoll task will cancel all its requests */
9153 if (node->ctx->sq_data)
9155 io_uring_try_cancel_requests(node->ctx, current,
9159 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
9160 io_uring_try_cancel_requests(ctx, current,
9164 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
9166 * If we've seen completions, retry without waiting. This
9167 * avoids a race where a completion comes in before we did
9168 * prepare_to_wait().
9170 if (inflight == tctx_inflight(tctx, !cancel_all))
9172 finish_wait(&tctx->wait, &wait);
9174 atomic_dec(&tctx->in_idle);
9176 io_uring_clean_tctx(tctx);
9178 /* for exec all current's requests should be gone, kill tctx */
9179 __io_uring_free(current);
9183 void __io_uring_cancel(struct files_struct *files)
9185 io_uring_cancel_generic(!files, NULL);
9188 static void *io_uring_validate_mmap_request(struct file *file,
9189 loff_t pgoff, size_t sz)
9191 struct io_ring_ctx *ctx = file->private_data;
9192 loff_t offset = pgoff << PAGE_SHIFT;
9197 case IORING_OFF_SQ_RING:
9198 case IORING_OFF_CQ_RING:
9201 case IORING_OFF_SQES:
9205 return ERR_PTR(-EINVAL);
9208 page = virt_to_head_page(ptr);
9209 if (sz > page_size(page))
9210 return ERR_PTR(-EINVAL);
9217 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9219 size_t sz = vma->vm_end - vma->vm_start;
9223 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
9225 return PTR_ERR(ptr);
9227 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
9228 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
9231 #else /* !CONFIG_MMU */
9233 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9235 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
9238 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
9240 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
9243 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
9244 unsigned long addr, unsigned long len,
9245 unsigned long pgoff, unsigned long flags)
9249 ptr = io_uring_validate_mmap_request(file, pgoff, len);
9251 return PTR_ERR(ptr);
9253 return (unsigned long) ptr;
9256 #endif /* !CONFIG_MMU */
9258 static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
9263 if (!io_sqring_full(ctx))
9265 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9267 if (!io_sqring_full(ctx))
9270 } while (!signal_pending(current));
9272 finish_wait(&ctx->sqo_sq_wait, &wait);
9276 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9277 struct __kernel_timespec __user **ts,
9278 const sigset_t __user **sig)
9280 struct io_uring_getevents_arg arg;
9283 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9284 * is just a pointer to the sigset_t.
9286 if (!(flags & IORING_ENTER_EXT_ARG)) {
9287 *sig = (const sigset_t __user *) argp;
9293 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9294 * timespec and sigset_t pointers if good.
9296 if (*argsz != sizeof(arg))
9298 if (copy_from_user(&arg, argp, sizeof(arg)))
9300 *sig = u64_to_user_ptr(arg.sigmask);
9301 *argsz = arg.sigmask_sz;
9302 *ts = u64_to_user_ptr(arg.ts);
9306 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9307 u32, min_complete, u32, flags, const void __user *, argp,
9310 struct io_ring_ctx *ctx;
9317 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9318 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG)))
9322 if (unlikely(!f.file))
9326 if (unlikely(f.file->f_op != &io_uring_fops))
9330 ctx = f.file->private_data;
9331 if (unlikely(!percpu_ref_tryget(&ctx->refs)))
9335 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
9339 * For SQ polling, the thread will do all submissions and completions.
9340 * Just return the requested submit count, and wake the thread if
9344 if (ctx->flags & IORING_SETUP_SQPOLL) {
9345 io_cqring_overflow_flush(ctx, false);
9348 if (unlikely(ctx->sq_data->thread == NULL)) {
9351 if (flags & IORING_ENTER_SQ_WAKEUP)
9352 wake_up(&ctx->sq_data->wait);
9353 if (flags & IORING_ENTER_SQ_WAIT) {
9354 ret = io_sqpoll_wait_sq(ctx);
9358 submitted = to_submit;
9359 } else if (to_submit) {
9360 ret = io_uring_add_tctx_node(ctx);
9363 mutex_lock(&ctx->uring_lock);
9364 submitted = io_submit_sqes(ctx, to_submit);
9365 mutex_unlock(&ctx->uring_lock);
9367 if (submitted != to_submit)
9370 if (flags & IORING_ENTER_GETEVENTS) {
9371 const sigset_t __user *sig;
9372 struct __kernel_timespec __user *ts;
9374 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
9378 min_complete = min(min_complete, ctx->cq_entries);
9381 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9382 * space applications don't need to do io completion events
9383 * polling again, they can rely on io_sq_thread to do polling
9384 * work, which can reduce cpu usage and uring_lock contention.
9386 if (ctx->flags & IORING_SETUP_IOPOLL &&
9387 !(ctx->flags & IORING_SETUP_SQPOLL)) {
9388 ret = io_iopoll_check(ctx, min_complete);
9390 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
9395 percpu_ref_put(&ctx->refs);
9398 return submitted ? submitted : ret;
9401 #ifdef CONFIG_PROC_FS
9402 static int io_uring_show_cred(struct seq_file *m, unsigned int id,
9403 const struct cred *cred)
9405 struct user_namespace *uns = seq_user_ns(m);
9406 struct group_info *gi;
9411 seq_printf(m, "%5d\n", id);
9412 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
9413 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
9414 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
9415 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
9416 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
9417 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
9418 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
9419 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
9420 seq_puts(m, "\n\tGroups:\t");
9421 gi = cred->group_info;
9422 for (g = 0; g < gi->ngroups; g++) {
9423 seq_put_decimal_ull(m, g ? " " : "",
9424 from_kgid_munged(uns, gi->gid[g]));
9426 seq_puts(m, "\n\tCapEff:\t");
9427 cap = cred->cap_effective;
9428 CAP_FOR_EACH_U32(__capi)
9429 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
9434 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
9436 struct io_sq_data *sq = NULL;
9441 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9442 * since fdinfo case grabs it in the opposite direction of normal use
9443 * cases. If we fail to get the lock, we just don't iterate any
9444 * structures that could be going away outside the io_uring mutex.
9446 has_lock = mutex_trylock(&ctx->uring_lock);
9448 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) {
9454 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9455 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9456 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9457 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9458 struct file *f = io_file_from_index(ctx, i);
9461 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9463 seq_printf(m, "%5u: <none>\n", i);
9465 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9466 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9467 struct io_mapped_ubuf *buf = ctx->user_bufs[i];
9468 unsigned int len = buf->ubuf_end - buf->ubuf;
9470 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf, len);
9472 if (has_lock && !xa_empty(&ctx->personalities)) {
9473 unsigned long index;
9474 const struct cred *cred;
9476 seq_printf(m, "Personalities:\n");
9477 xa_for_each(&ctx->personalities, index, cred)
9478 io_uring_show_cred(m, index, cred);
9480 seq_printf(m, "PollList:\n");
9481 spin_lock_irq(&ctx->completion_lock);
9482 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9483 struct hlist_head *list = &ctx->cancel_hash[i];
9484 struct io_kiocb *req;
9486 hlist_for_each_entry(req, list, hash_node)
9487 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9488 req->task->task_works != NULL);
9490 spin_unlock_irq(&ctx->completion_lock);
9492 mutex_unlock(&ctx->uring_lock);
9495 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9497 struct io_ring_ctx *ctx = f->private_data;
9499 if (percpu_ref_tryget(&ctx->refs)) {
9500 __io_uring_show_fdinfo(ctx, m);
9501 percpu_ref_put(&ctx->refs);
9506 static const struct file_operations io_uring_fops = {
9507 .release = io_uring_release,
9508 .mmap = io_uring_mmap,
9510 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9511 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9513 .poll = io_uring_poll,
9514 .fasync = io_uring_fasync,
9515 #ifdef CONFIG_PROC_FS
9516 .show_fdinfo = io_uring_show_fdinfo,
9520 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9521 struct io_uring_params *p)
9523 struct io_rings *rings;
9524 size_t size, sq_array_offset;
9526 /* make sure these are sane, as we already accounted them */
9527 ctx->sq_entries = p->sq_entries;
9528 ctx->cq_entries = p->cq_entries;
9530 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9531 if (size == SIZE_MAX)
9534 rings = io_mem_alloc(size);
9539 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9540 rings->sq_ring_mask = p->sq_entries - 1;
9541 rings->cq_ring_mask = p->cq_entries - 1;
9542 rings->sq_ring_entries = p->sq_entries;
9543 rings->cq_ring_entries = p->cq_entries;
9545 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9546 if (size == SIZE_MAX) {
9547 io_mem_free(ctx->rings);
9552 ctx->sq_sqes = io_mem_alloc(size);
9553 if (!ctx->sq_sqes) {
9554 io_mem_free(ctx->rings);
9562 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
9566 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9570 ret = io_uring_add_tctx_node(ctx);
9575 fd_install(fd, file);
9580 * Allocate an anonymous fd, this is what constitutes the application
9581 * visible backing of an io_uring instance. The application mmaps this
9582 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9583 * we have to tie this fd to a socket for file garbage collection purposes.
9585 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
9588 #if defined(CONFIG_UNIX)
9591 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9594 return ERR_PTR(ret);
9597 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9598 O_RDWR | O_CLOEXEC);
9599 #if defined(CONFIG_UNIX)
9601 sock_release(ctx->ring_sock);
9602 ctx->ring_sock = NULL;
9604 ctx->ring_sock->file = file;
9610 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9611 struct io_uring_params __user *params)
9613 struct io_ring_ctx *ctx;
9619 if (entries > IORING_MAX_ENTRIES) {
9620 if (!(p->flags & IORING_SETUP_CLAMP))
9622 entries = IORING_MAX_ENTRIES;
9626 * Use twice as many entries for the CQ ring. It's possible for the
9627 * application to drive a higher depth than the size of the SQ ring,
9628 * since the sqes are only used at submission time. This allows for
9629 * some flexibility in overcommitting a bit. If the application has
9630 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9631 * of CQ ring entries manually.
9633 p->sq_entries = roundup_pow_of_two(entries);
9634 if (p->flags & IORING_SETUP_CQSIZE) {
9636 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9637 * to a power-of-two, if it isn't already. We do NOT impose
9638 * any cq vs sq ring sizing.
9642 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9643 if (!(p->flags & IORING_SETUP_CLAMP))
9645 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9647 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9648 if (p->cq_entries < p->sq_entries)
9651 p->cq_entries = 2 * p->sq_entries;
9654 ctx = io_ring_ctx_alloc(p);
9657 ctx->compat = in_compat_syscall();
9658 if (!capable(CAP_IPC_LOCK))
9659 ctx->user = get_uid(current_user());
9662 * This is just grabbed for accounting purposes. When a process exits,
9663 * the mm is exited and dropped before the files, hence we need to hang
9664 * on to this mm purely for the purposes of being able to unaccount
9665 * memory (locked/pinned vm). It's not used for anything else.
9667 mmgrab(current->mm);
9668 ctx->mm_account = current->mm;
9670 ret = io_allocate_scq_urings(ctx, p);
9674 ret = io_sq_offload_create(ctx, p);
9677 /* always set a rsrc node */
9678 ret = io_rsrc_node_switch_start(ctx);
9681 io_rsrc_node_switch(ctx, NULL);
9683 memset(&p->sq_off, 0, sizeof(p->sq_off));
9684 p->sq_off.head = offsetof(struct io_rings, sq.head);
9685 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9686 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9687 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9688 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9689 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9690 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9692 memset(&p->cq_off, 0, sizeof(p->cq_off));
9693 p->cq_off.head = offsetof(struct io_rings, cq.head);
9694 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9695 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9696 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9697 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9698 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9699 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9701 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9702 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9703 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9704 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
9705 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
9706 IORING_FEAT_RSRC_TAGS;
9708 if (copy_to_user(params, p, sizeof(*p))) {
9713 file = io_uring_get_file(ctx);
9715 ret = PTR_ERR(file);
9720 * Install ring fd as the very last thing, so we don't risk someone
9721 * having closed it before we finish setup
9723 ret = io_uring_install_fd(ctx, file);
9725 /* fput will clean it up */
9730 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9733 io_ring_ctx_wait_and_kill(ctx);
9738 * Sets up an aio uring context, and returns the fd. Applications asks for a
9739 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9740 * params structure passed in.
9742 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9744 struct io_uring_params p;
9747 if (copy_from_user(&p, params, sizeof(p)))
9749 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9754 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9755 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9756 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9757 IORING_SETUP_R_DISABLED))
9760 return io_uring_create(entries, &p, params);
9763 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9764 struct io_uring_params __user *, params)
9766 return io_uring_setup(entries, params);
9769 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9771 struct io_uring_probe *p;
9775 size = struct_size(p, ops, nr_args);
9776 if (size == SIZE_MAX)
9778 p = kzalloc(size, GFP_KERNEL);
9783 if (copy_from_user(p, arg, size))
9786 if (memchr_inv(p, 0, size))
9789 p->last_op = IORING_OP_LAST - 1;
9790 if (nr_args > IORING_OP_LAST)
9791 nr_args = IORING_OP_LAST;
9793 for (i = 0; i < nr_args; i++) {
9795 if (!io_op_defs[i].not_supported)
9796 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9801 if (copy_to_user(arg, p, size))
9808 static int io_register_personality(struct io_ring_ctx *ctx)
9810 const struct cred *creds;
9814 creds = get_current_cred();
9816 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
9817 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
9824 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9825 unsigned int nr_args)
9827 struct io_uring_restriction *res;
9831 /* Restrictions allowed only if rings started disabled */
9832 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9835 /* We allow only a single restrictions registration */
9836 if (ctx->restrictions.registered)
9839 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9842 size = array_size(nr_args, sizeof(*res));
9843 if (size == SIZE_MAX)
9846 res = memdup_user(arg, size);
9848 return PTR_ERR(res);
9852 for (i = 0; i < nr_args; i++) {
9853 switch (res[i].opcode) {
9854 case IORING_RESTRICTION_REGISTER_OP:
9855 if (res[i].register_op >= IORING_REGISTER_LAST) {
9860 __set_bit(res[i].register_op,
9861 ctx->restrictions.register_op);
9863 case IORING_RESTRICTION_SQE_OP:
9864 if (res[i].sqe_op >= IORING_OP_LAST) {
9869 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9871 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9872 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9874 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9875 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9884 /* Reset all restrictions if an error happened */
9886 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9888 ctx->restrictions.registered = true;
9894 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9896 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9899 if (ctx->restrictions.registered)
9900 ctx->restricted = 1;
9902 ctx->flags &= ~IORING_SETUP_R_DISABLED;
9903 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
9904 wake_up(&ctx->sq_data->wait);
9908 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
9909 struct io_uring_rsrc_update2 *up,
9917 if (check_add_overflow(up->offset, nr_args, &tmp))
9919 err = io_rsrc_node_switch_start(ctx);
9924 case IORING_RSRC_FILE:
9925 return __io_sqe_files_update(ctx, up, nr_args);
9926 case IORING_RSRC_BUFFER:
9927 return __io_sqe_buffers_update(ctx, up, nr_args);
9932 static int io_register_files_update(struct io_ring_ctx *ctx, void __user *arg,
9935 struct io_uring_rsrc_update2 up;
9939 memset(&up, 0, sizeof(up));
9940 if (copy_from_user(&up, arg, sizeof(struct io_uring_rsrc_update)))
9942 return __io_register_rsrc_update(ctx, IORING_RSRC_FILE, &up, nr_args);
9945 static int io_register_rsrc_update(struct io_ring_ctx *ctx, void __user *arg,
9946 unsigned size, unsigned type)
9948 struct io_uring_rsrc_update2 up;
9950 if (size != sizeof(up))
9952 if (copy_from_user(&up, arg, sizeof(up)))
9954 if (!up.nr || up.resv)
9956 return __io_register_rsrc_update(ctx, type, &up, up.nr);
9959 static int io_register_rsrc(struct io_ring_ctx *ctx, void __user *arg,
9960 unsigned int size, unsigned int type)
9962 struct io_uring_rsrc_register rr;
9964 /* keep it extendible */
9965 if (size != sizeof(rr))
9968 memset(&rr, 0, sizeof(rr));
9969 if (copy_from_user(&rr, arg, size))
9971 if (!rr.nr || rr.resv || rr.resv2)
9975 case IORING_RSRC_FILE:
9976 return io_sqe_files_register(ctx, u64_to_user_ptr(rr.data),
9977 rr.nr, u64_to_user_ptr(rr.tags));
9978 case IORING_RSRC_BUFFER:
9979 return io_sqe_buffers_register(ctx, u64_to_user_ptr(rr.data),
9980 rr.nr, u64_to_user_ptr(rr.tags));
9985 static int io_register_iowq_aff(struct io_ring_ctx *ctx, void __user *arg,
9988 struct io_uring_task *tctx = current->io_uring;
9989 cpumask_var_t new_mask;
9992 if (!tctx || !tctx->io_wq)
9995 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
9998 cpumask_clear(new_mask);
9999 if (len > cpumask_size())
10000 len = cpumask_size();
10002 if (copy_from_user(new_mask, arg, len)) {
10003 free_cpumask_var(new_mask);
10007 ret = io_wq_cpu_affinity(tctx->io_wq, new_mask);
10008 free_cpumask_var(new_mask);
10012 static int io_unregister_iowq_aff(struct io_ring_ctx *ctx)
10014 struct io_uring_task *tctx = current->io_uring;
10016 if (!tctx || !tctx->io_wq)
10019 return io_wq_cpu_affinity(tctx->io_wq, NULL);
10022 static bool io_register_op_must_quiesce(int op)
10025 case IORING_REGISTER_BUFFERS:
10026 case IORING_UNREGISTER_BUFFERS:
10027 case IORING_REGISTER_FILES:
10028 case IORING_UNREGISTER_FILES:
10029 case IORING_REGISTER_FILES_UPDATE:
10030 case IORING_REGISTER_PROBE:
10031 case IORING_REGISTER_PERSONALITY:
10032 case IORING_UNREGISTER_PERSONALITY:
10033 case IORING_REGISTER_FILES2:
10034 case IORING_REGISTER_FILES_UPDATE2:
10035 case IORING_REGISTER_BUFFERS2:
10036 case IORING_REGISTER_BUFFERS_UPDATE:
10037 case IORING_REGISTER_IOWQ_AFF:
10038 case IORING_UNREGISTER_IOWQ_AFF:
10045 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
10046 void __user *arg, unsigned nr_args)
10047 __releases(ctx->uring_lock)
10048 __acquires(ctx->uring_lock)
10053 * We're inside the ring mutex, if the ref is already dying, then
10054 * someone else killed the ctx or is already going through
10055 * io_uring_register().
10057 if (percpu_ref_is_dying(&ctx->refs))
10060 if (ctx->restricted) {
10061 if (opcode >= IORING_REGISTER_LAST)
10063 opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
10064 if (!test_bit(opcode, ctx->restrictions.register_op))
10068 if (io_register_op_must_quiesce(opcode)) {
10069 percpu_ref_kill(&ctx->refs);
10072 * Drop uring mutex before waiting for references to exit. If
10073 * another thread is currently inside io_uring_enter() it might
10074 * need to grab the uring_lock to make progress. If we hold it
10075 * here across the drain wait, then we can deadlock. It's safe
10076 * to drop the mutex here, since no new references will come in
10077 * after we've killed the percpu ref.
10079 mutex_unlock(&ctx->uring_lock);
10081 ret = wait_for_completion_interruptible(&ctx->ref_comp);
10084 ret = io_run_task_work_sig();
10088 mutex_lock(&ctx->uring_lock);
10091 io_refs_resurrect(&ctx->refs, &ctx->ref_comp);
10097 case IORING_REGISTER_BUFFERS:
10098 ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
10100 case IORING_UNREGISTER_BUFFERS:
10102 if (arg || nr_args)
10104 ret = io_sqe_buffers_unregister(ctx);
10106 case IORING_REGISTER_FILES:
10107 ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
10109 case IORING_UNREGISTER_FILES:
10111 if (arg || nr_args)
10113 ret = io_sqe_files_unregister(ctx);
10115 case IORING_REGISTER_FILES_UPDATE:
10116 ret = io_register_files_update(ctx, arg, nr_args);
10118 case IORING_REGISTER_EVENTFD:
10119 case IORING_REGISTER_EVENTFD_ASYNC:
10123 ret = io_eventfd_register(ctx, arg);
10126 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
10127 ctx->eventfd_async = 1;
10129 ctx->eventfd_async = 0;
10131 case IORING_UNREGISTER_EVENTFD:
10133 if (arg || nr_args)
10135 ret = io_eventfd_unregister(ctx);
10137 case IORING_REGISTER_PROBE:
10139 if (!arg || nr_args > 256)
10141 ret = io_probe(ctx, arg, nr_args);
10143 case IORING_REGISTER_PERSONALITY:
10145 if (arg || nr_args)
10147 ret = io_register_personality(ctx);
10149 case IORING_UNREGISTER_PERSONALITY:
10153 ret = io_unregister_personality(ctx, nr_args);
10155 case IORING_REGISTER_ENABLE_RINGS:
10157 if (arg || nr_args)
10159 ret = io_register_enable_rings(ctx);
10161 case IORING_REGISTER_RESTRICTIONS:
10162 ret = io_register_restrictions(ctx, arg, nr_args);
10164 case IORING_REGISTER_FILES2:
10165 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE);
10167 case IORING_REGISTER_FILES_UPDATE2:
10168 ret = io_register_rsrc_update(ctx, arg, nr_args,
10171 case IORING_REGISTER_BUFFERS2:
10172 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER);
10174 case IORING_REGISTER_BUFFERS_UPDATE:
10175 ret = io_register_rsrc_update(ctx, arg, nr_args,
10176 IORING_RSRC_BUFFER);
10178 case IORING_REGISTER_IOWQ_AFF:
10180 if (!arg || !nr_args)
10182 ret = io_register_iowq_aff(ctx, arg, nr_args);
10184 case IORING_UNREGISTER_IOWQ_AFF:
10186 if (arg || nr_args)
10188 ret = io_unregister_iowq_aff(ctx);
10195 if (io_register_op_must_quiesce(opcode)) {
10196 /* bring the ctx back to life */
10197 percpu_ref_reinit(&ctx->refs);
10198 reinit_completion(&ctx->ref_comp);
10203 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
10204 void __user *, arg, unsigned int, nr_args)
10206 struct io_ring_ctx *ctx;
10215 if (f.file->f_op != &io_uring_fops)
10218 ctx = f.file->private_data;
10220 io_run_task_work();
10222 mutex_lock(&ctx->uring_lock);
10223 ret = __io_uring_register(ctx, opcode, arg, nr_args);
10224 mutex_unlock(&ctx->uring_lock);
10225 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
10226 ctx->cq_ev_fd != NULL, ret);
10232 static int __init io_uring_init(void)
10234 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
10235 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
10236 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
10239 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
10240 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
10241 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
10242 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
10243 BUILD_BUG_SQE_ELEM(1, __u8, flags);
10244 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
10245 BUILD_BUG_SQE_ELEM(4, __s32, fd);
10246 BUILD_BUG_SQE_ELEM(8, __u64, off);
10247 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
10248 BUILD_BUG_SQE_ELEM(16, __u64, addr);
10249 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
10250 BUILD_BUG_SQE_ELEM(24, __u32, len);
10251 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
10252 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
10253 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
10254 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
10255 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
10256 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
10257 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
10258 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
10259 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
10260 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
10261 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
10262 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
10263 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
10264 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
10265 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
10266 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
10267 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
10268 BUILD_BUG_SQE_ELEM(42, __u16, personality);
10269 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
10271 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
10272 sizeof(struct io_uring_rsrc_update));
10273 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
10274 sizeof(struct io_uring_rsrc_update2));
10275 /* should fit into one byte */
10276 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
10278 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
10279 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
10280 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
10284 __initcall(io_uring_init);