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)
92 #define IORING_SQPOLL_CAP_ENTRIES_VALUE 8
95 * Shift of 9 is 512 entries, or exactly one page on 64-bit archs
97 #define IORING_FILE_TABLE_SHIFT 9
98 #define IORING_MAX_FILES_TABLE (1U << IORING_FILE_TABLE_SHIFT)
99 #define IORING_FILE_TABLE_MASK (IORING_MAX_FILES_TABLE - 1)
100 #define IORING_MAX_FIXED_FILES (64 * IORING_MAX_FILES_TABLE)
101 #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
102 IORING_REGISTER_LAST + IORING_OP_LAST)
104 #define IO_RSRC_TAG_TABLE_SHIFT 9
105 #define IO_RSRC_TAG_TABLE_MAX (1U << IO_RSRC_TAG_TABLE_SHIFT)
106 #define IO_RSRC_TAG_TABLE_MASK (IO_RSRC_TAG_TABLE_MAX - 1)
108 #define IORING_MAX_REG_BUFFERS (1U << 14)
110 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
111 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
113 #define IO_REQ_CLEAN_FLAGS (REQ_F_BUFFER_SELECTED | REQ_F_NEED_CLEANUP | \
114 REQ_F_POLLED | REQ_F_INFLIGHT | REQ_F_CREDS)
116 #define IO_TCTX_REFS_CACHE_NR (1U << 10)
119 u32 head ____cacheline_aligned_in_smp;
120 u32 tail ____cacheline_aligned_in_smp;
124 * This data is shared with the application through the mmap at offsets
125 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
127 * The offsets to the member fields are published through struct
128 * io_sqring_offsets when calling io_uring_setup.
132 * Head and tail offsets into the ring; the offsets need to be
133 * masked to get valid indices.
135 * The kernel controls head of the sq ring and the tail of the cq ring,
136 * and the application controls tail of the sq ring and the head of the
139 struct io_uring sq, cq;
141 * Bitmasks to apply to head and tail offsets (constant, equals
144 u32 sq_ring_mask, cq_ring_mask;
145 /* Ring sizes (constant, power of 2) */
146 u32 sq_ring_entries, cq_ring_entries;
148 * Number of invalid entries dropped by the kernel due to
149 * invalid index stored in array
151 * Written by the kernel, shouldn't be modified by the
152 * application (i.e. get number of "new events" by comparing to
155 * After a new SQ head value was read by the application this
156 * counter includes all submissions that were dropped reaching
157 * the new SQ head (and possibly more).
163 * Written by the kernel, shouldn't be modified by the
166 * The application needs a full memory barrier before checking
167 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
173 * Written by the application, shouldn't be modified by the
178 * Number of completion events lost because the queue was full;
179 * this should be avoided by the application by making sure
180 * there are not more requests pending than there is space in
181 * the completion queue.
183 * Written by the kernel, shouldn't be modified by the
184 * application (i.e. get number of "new events" by comparing to
187 * As completion events come in out of order this counter is not
188 * ordered with any other data.
192 * Ring buffer of completion events.
194 * The kernel writes completion events fresh every time they are
195 * produced, so the application is allowed to modify pending
198 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
201 enum io_uring_cmd_flags {
202 IO_URING_F_NONBLOCK = 1,
203 IO_URING_F_COMPLETE_DEFER = 2,
206 struct io_mapped_ubuf {
209 unsigned int nr_bvecs;
210 unsigned long acct_pages;
211 struct bio_vec bvec[];
216 struct io_overflow_cqe {
217 struct io_uring_cqe cqe;
218 struct list_head list;
221 struct io_fixed_file {
222 /* file * with additional FFS_* flags */
223 unsigned long file_ptr;
227 struct list_head list;
232 struct io_mapped_ubuf *buf;
236 struct io_file_table {
237 /* two level table */
238 struct io_fixed_file **files;
241 struct io_rsrc_node {
242 struct percpu_ref refs;
243 struct list_head node;
244 struct list_head rsrc_list;
245 struct io_rsrc_data *rsrc_data;
246 struct llist_node llist;
250 typedef void (rsrc_put_fn)(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc);
252 struct io_rsrc_data {
253 struct io_ring_ctx *ctx;
259 struct completion done;
264 struct list_head list;
270 struct io_restriction {
271 DECLARE_BITMAP(register_op, IORING_REGISTER_LAST);
272 DECLARE_BITMAP(sqe_op, IORING_OP_LAST);
273 u8 sqe_flags_allowed;
274 u8 sqe_flags_required;
279 IO_SQ_THREAD_SHOULD_STOP = 0,
280 IO_SQ_THREAD_SHOULD_PARK,
285 atomic_t park_pending;
288 /* ctx's that are using this sqd */
289 struct list_head ctx_list;
291 struct task_struct *thread;
292 struct wait_queue_head wait;
294 unsigned sq_thread_idle;
300 struct completion exited;
303 #define IO_IOPOLL_BATCH 8
304 #define IO_COMPL_BATCH 32
305 #define IO_REQ_CACHE_SIZE 32
306 #define IO_REQ_ALLOC_BATCH 8
308 struct io_comp_state {
309 struct io_kiocb *reqs[IO_COMPL_BATCH];
311 /* inline/task_work completion list, under ->uring_lock */
312 struct list_head free_list;
315 struct io_submit_link {
316 struct io_kiocb *head;
317 struct io_kiocb *last;
320 struct io_submit_state {
321 struct blk_plug plug;
322 struct io_submit_link link;
325 * io_kiocb alloc cache
327 void *reqs[IO_REQ_CACHE_SIZE];
328 unsigned int free_reqs;
333 * Batch completion logic
335 struct io_comp_state comp;
338 * File reference cache
342 unsigned int file_refs;
343 unsigned int ios_left;
347 /* const or read-mostly hot data */
349 struct percpu_ref refs;
351 struct io_rings *rings;
353 unsigned int compat: 1;
354 unsigned int drain_next: 1;
355 unsigned int eventfd_async: 1;
356 unsigned int restricted: 1;
357 unsigned int off_timeout_used: 1;
358 unsigned int drain_active: 1;
359 } ____cacheline_aligned_in_smp;
361 /* submission data */
363 struct mutex uring_lock;
366 * Ring buffer of indices into array of io_uring_sqe, which is
367 * mmapped by the application using the IORING_OFF_SQES offset.
369 * This indirection could e.g. be used to assign fixed
370 * io_uring_sqe entries to operations and only submit them to
371 * the queue when needed.
373 * The kernel modifies neither the indices array nor the entries
377 struct io_uring_sqe *sq_sqes;
378 unsigned cached_sq_head;
380 struct list_head defer_list;
383 * Fixed resources fast path, should be accessed only under
384 * uring_lock, and updated through io_uring_register(2)
386 struct io_rsrc_node *rsrc_node;
387 struct io_file_table file_table;
388 unsigned nr_user_files;
389 unsigned nr_user_bufs;
390 struct io_mapped_ubuf **user_bufs;
392 struct io_submit_state submit_state;
393 struct list_head timeout_list;
394 struct list_head cq_overflow_list;
395 struct xarray io_buffers;
396 struct xarray personalities;
398 unsigned sq_thread_idle;
399 } ____cacheline_aligned_in_smp;
401 /* IRQ completion list, under ->completion_lock */
402 struct list_head locked_free_list;
403 unsigned int locked_free_nr;
405 const struct cred *sq_creds; /* cred used for __io_sq_thread() */
406 struct io_sq_data *sq_data; /* if using sq thread polling */
408 struct wait_queue_head sqo_sq_wait;
409 struct list_head sqd_list;
411 unsigned long check_cq_overflow;
414 unsigned cached_cq_tail;
416 struct eventfd_ctx *cq_ev_fd;
417 struct wait_queue_head poll_wait;
418 struct wait_queue_head cq_wait;
420 atomic_t cq_timeouts;
421 struct fasync_struct *cq_fasync;
422 unsigned cq_last_tm_flush;
423 } ____cacheline_aligned_in_smp;
426 spinlock_t completion_lock;
429 * ->iopoll_list is protected by the ctx->uring_lock for
430 * io_uring instances that don't use IORING_SETUP_SQPOLL.
431 * For SQPOLL, only the single threaded io_sq_thread() will
432 * manipulate the list, hence no extra locking is needed there.
434 struct list_head iopoll_list;
435 struct hlist_head *cancel_hash;
436 unsigned cancel_hash_bits;
437 bool poll_multi_file;
438 } ____cacheline_aligned_in_smp;
440 struct io_restriction restrictions;
442 /* slow path rsrc auxilary data, used by update/register */
444 struct io_rsrc_node *rsrc_backup_node;
445 struct io_mapped_ubuf *dummy_ubuf;
446 struct io_rsrc_data *file_data;
447 struct io_rsrc_data *buf_data;
449 struct delayed_work rsrc_put_work;
450 struct llist_head rsrc_put_llist;
451 struct list_head rsrc_ref_list;
452 spinlock_t rsrc_ref_lock;
455 /* Keep this last, we don't need it for the fast path */
457 #if defined(CONFIG_UNIX)
458 struct socket *ring_sock;
460 /* hashed buffered write serialization */
461 struct io_wq_hash *hash_map;
463 /* Only used for accounting purposes */
464 struct user_struct *user;
465 struct mm_struct *mm_account;
467 /* ctx exit and cancelation */
468 struct callback_head *exit_task_work;
469 struct work_struct exit_work;
470 struct list_head tctx_list;
471 struct completion ref_comp;
475 struct io_uring_task {
476 /* submission side */
479 struct wait_queue_head wait;
480 const struct io_ring_ctx *last;
482 struct percpu_counter inflight;
483 atomic_t inflight_tracked;
486 spinlock_t task_lock;
487 struct io_wq_work_list task_list;
488 unsigned long task_state;
489 struct callback_head task_work;
493 * First field must be the file pointer in all the
494 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
496 struct io_poll_iocb {
498 struct wait_queue_head *head;
502 struct wait_queue_entry wait;
505 struct io_poll_update {
511 bool update_user_data;
519 struct io_timeout_data {
520 struct io_kiocb *req;
521 struct hrtimer timer;
522 struct timespec64 ts;
523 enum hrtimer_mode mode;
528 struct sockaddr __user *addr;
529 int __user *addr_len;
531 unsigned long nofile;
551 struct list_head list;
552 /* head of the link, used by linked timeouts only */
553 struct io_kiocb *head;
556 struct io_timeout_rem {
561 struct timespec64 ts;
566 /* NOTE: kiocb has the file as the first member, so don't do it here */
574 struct sockaddr __user *addr;
581 struct compat_msghdr __user *umsg_compat;
582 struct user_msghdr __user *umsg;
588 struct io_buffer *kbuf;
594 struct filename *filename;
596 unsigned long nofile;
599 struct io_rsrc_update {
625 struct epoll_event event;
629 struct file *file_out;
630 struct file *file_in;
637 struct io_provide_buf {
651 const char __user *filename;
652 struct statx __user *buffer;
664 struct filename *oldpath;
665 struct filename *newpath;
673 struct filename *filename;
676 struct io_completion {
678 struct list_head list;
682 struct io_async_connect {
683 struct sockaddr_storage address;
686 struct io_async_msghdr {
687 struct iovec fast_iov[UIO_FASTIOV];
688 /* points to an allocated iov, if NULL we use fast_iov instead */
689 struct iovec *free_iov;
690 struct sockaddr __user *uaddr;
692 struct sockaddr_storage addr;
696 struct iovec fast_iov[UIO_FASTIOV];
697 const struct iovec *free_iovec;
698 struct iov_iter iter;
700 struct wait_page_queue wpq;
704 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
705 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
706 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
707 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
708 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
709 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
711 /* first byte is taken by user flags, shift it to not overlap */
716 REQ_F_LINK_TIMEOUT_BIT,
717 REQ_F_NEED_CLEANUP_BIT,
719 REQ_F_BUFFER_SELECTED_BIT,
720 REQ_F_LTIMEOUT_ACTIVE_BIT,
721 REQ_F_COMPLETE_INLINE_BIT,
723 REQ_F_DONT_REISSUE_BIT,
725 /* keep async read/write and isreg together and in order */
726 REQ_F_ASYNC_READ_BIT,
727 REQ_F_ASYNC_WRITE_BIT,
730 /* not a real bit, just to check we're not overflowing the space */
736 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
737 /* drain existing IO first */
738 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
740 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
741 /* doesn't sever on completion < 0 */
742 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
744 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
745 /* IOSQE_BUFFER_SELECT */
746 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
748 /* fail rest of links */
749 REQ_F_FAIL = BIT(REQ_F_FAIL_BIT),
750 /* on inflight list, should be cancelled and waited on exit reliably */
751 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
752 /* read/write uses file position */
753 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
754 /* must not punt to workers */
755 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
756 /* has or had linked timeout */
757 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
759 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
760 /* already went through poll handler */
761 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
762 /* buffer already selected */
763 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
764 /* linked timeout is active, i.e. prepared by link's head */
765 REQ_F_LTIMEOUT_ACTIVE = BIT(REQ_F_LTIMEOUT_ACTIVE_BIT),
766 /* completion is deferred through io_comp_state */
767 REQ_F_COMPLETE_INLINE = BIT(REQ_F_COMPLETE_INLINE_BIT),
768 /* caller should reissue async */
769 REQ_F_REISSUE = BIT(REQ_F_REISSUE_BIT),
770 /* don't attempt request reissue, see io_rw_reissue() */
771 REQ_F_DONT_REISSUE = BIT(REQ_F_DONT_REISSUE_BIT),
772 /* supports async reads */
773 REQ_F_ASYNC_READ = BIT(REQ_F_ASYNC_READ_BIT),
774 /* supports async writes */
775 REQ_F_ASYNC_WRITE = BIT(REQ_F_ASYNC_WRITE_BIT),
777 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
778 /* has creds assigned */
779 REQ_F_CREDS = BIT(REQ_F_CREDS_BIT),
783 struct io_poll_iocb poll;
784 struct io_poll_iocb *double_poll;
787 struct io_task_work {
788 struct io_wq_work_node node;
789 task_work_func_t func;
793 IORING_RSRC_FILE = 0,
794 IORING_RSRC_BUFFER = 1,
798 * NOTE! Each of the iocb union members has the file pointer
799 * as the first entry in their struct definition. So you can
800 * access the file pointer through any of the sub-structs,
801 * or directly as just 'ki_filp' in this struct.
807 struct io_poll_iocb poll;
808 struct io_poll_update poll_update;
809 struct io_accept accept;
811 struct io_cancel cancel;
812 struct io_timeout timeout;
813 struct io_timeout_rem timeout_rem;
814 struct io_connect connect;
815 struct io_sr_msg sr_msg;
817 struct io_close close;
818 struct io_rsrc_update rsrc_update;
819 struct io_fadvise fadvise;
820 struct io_madvise madvise;
821 struct io_epoll epoll;
822 struct io_splice splice;
823 struct io_provide_buf pbuf;
824 struct io_statx statx;
825 struct io_shutdown shutdown;
826 struct io_rename rename;
827 struct io_unlink unlink;
828 /* use only after cleaning per-op data, see io_clean_op() */
829 struct io_completion compl;
832 /* opcode allocated if it needs to store data for async defer */
835 /* polled IO has completed */
841 struct io_ring_ctx *ctx;
844 struct task_struct *task;
847 struct io_kiocb *link;
848 struct percpu_ref *fixed_rsrc_refs;
850 /* used with ctx->iopoll_list with reads/writes */
851 struct list_head inflight_entry;
853 struct io_task_work io_task_work;
854 struct callback_head task_work;
856 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
857 struct hlist_node hash_node;
858 struct async_poll *apoll;
859 struct io_wq_work work;
860 const struct cred *creds;
862 /* store used ubuf, so we can prevent reloading */
863 struct io_mapped_ubuf *imu;
866 struct io_tctx_node {
867 struct list_head ctx_node;
868 struct task_struct *task;
869 struct io_ring_ctx *ctx;
872 struct io_defer_entry {
873 struct list_head list;
874 struct io_kiocb *req;
879 /* needs req->file assigned */
880 unsigned needs_file : 1;
881 /* hash wq insertion if file is a regular file */
882 unsigned hash_reg_file : 1;
883 /* unbound wq insertion if file is a non-regular file */
884 unsigned unbound_nonreg_file : 1;
885 /* opcode is not supported by this kernel */
886 unsigned not_supported : 1;
887 /* set if opcode supports polled "wait" */
889 unsigned pollout : 1;
890 /* op supports buffer selection */
891 unsigned buffer_select : 1;
892 /* do prep async if is going to be punted */
893 unsigned needs_async_setup : 1;
894 /* should block plug */
896 /* size of async data needed, if any */
897 unsigned short async_size;
900 static const struct io_op_def io_op_defs[] = {
901 [IORING_OP_NOP] = {},
902 [IORING_OP_READV] = {
904 .unbound_nonreg_file = 1,
907 .needs_async_setup = 1,
909 .async_size = sizeof(struct io_async_rw),
911 [IORING_OP_WRITEV] = {
914 .unbound_nonreg_file = 1,
916 .needs_async_setup = 1,
918 .async_size = sizeof(struct io_async_rw),
920 [IORING_OP_FSYNC] = {
923 [IORING_OP_READ_FIXED] = {
925 .unbound_nonreg_file = 1,
928 .async_size = sizeof(struct io_async_rw),
930 [IORING_OP_WRITE_FIXED] = {
933 .unbound_nonreg_file = 1,
936 .async_size = sizeof(struct io_async_rw),
938 [IORING_OP_POLL_ADD] = {
940 .unbound_nonreg_file = 1,
942 [IORING_OP_POLL_REMOVE] = {},
943 [IORING_OP_SYNC_FILE_RANGE] = {
946 [IORING_OP_SENDMSG] = {
948 .unbound_nonreg_file = 1,
950 .needs_async_setup = 1,
951 .async_size = sizeof(struct io_async_msghdr),
953 [IORING_OP_RECVMSG] = {
955 .unbound_nonreg_file = 1,
958 .needs_async_setup = 1,
959 .async_size = sizeof(struct io_async_msghdr),
961 [IORING_OP_TIMEOUT] = {
962 .async_size = sizeof(struct io_timeout_data),
964 [IORING_OP_TIMEOUT_REMOVE] = {
965 /* used by timeout updates' prep() */
967 [IORING_OP_ACCEPT] = {
969 .unbound_nonreg_file = 1,
972 [IORING_OP_ASYNC_CANCEL] = {},
973 [IORING_OP_LINK_TIMEOUT] = {
974 .async_size = sizeof(struct io_timeout_data),
976 [IORING_OP_CONNECT] = {
978 .unbound_nonreg_file = 1,
980 .needs_async_setup = 1,
981 .async_size = sizeof(struct io_async_connect),
983 [IORING_OP_FALLOCATE] = {
986 [IORING_OP_OPENAT] = {},
987 [IORING_OP_CLOSE] = {},
988 [IORING_OP_FILES_UPDATE] = {},
989 [IORING_OP_STATX] = {},
992 .unbound_nonreg_file = 1,
996 .async_size = sizeof(struct io_async_rw),
998 [IORING_OP_WRITE] = {
1000 .unbound_nonreg_file = 1,
1003 .async_size = sizeof(struct io_async_rw),
1005 [IORING_OP_FADVISE] = {
1008 [IORING_OP_MADVISE] = {},
1009 [IORING_OP_SEND] = {
1011 .unbound_nonreg_file = 1,
1014 [IORING_OP_RECV] = {
1016 .unbound_nonreg_file = 1,
1020 [IORING_OP_OPENAT2] = {
1022 [IORING_OP_EPOLL_CTL] = {
1023 .unbound_nonreg_file = 1,
1025 [IORING_OP_SPLICE] = {
1028 .unbound_nonreg_file = 1,
1030 [IORING_OP_PROVIDE_BUFFERS] = {},
1031 [IORING_OP_REMOVE_BUFFERS] = {},
1035 .unbound_nonreg_file = 1,
1037 [IORING_OP_SHUTDOWN] = {
1040 [IORING_OP_RENAMEAT] = {},
1041 [IORING_OP_UNLINKAT] = {},
1044 static bool io_disarm_next(struct io_kiocb *req);
1045 static void io_uring_del_tctx_node(unsigned long index);
1046 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
1047 struct task_struct *task,
1049 static void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd);
1050 static struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx);
1052 static bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1053 long res, unsigned int cflags);
1054 static void io_put_req(struct io_kiocb *req);
1055 static void io_put_req_deferred(struct io_kiocb *req, int nr);
1056 static void io_dismantle_req(struct io_kiocb *req);
1057 static void io_put_task(struct task_struct *task, int nr);
1058 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
1059 static void io_queue_linked_timeout(struct io_kiocb *req);
1060 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
1061 struct io_uring_rsrc_update2 *up,
1063 static void io_clean_op(struct io_kiocb *req);
1064 static struct file *io_file_get(struct io_submit_state *state,
1065 struct io_kiocb *req, int fd, bool fixed);
1066 static void __io_queue_sqe(struct io_kiocb *req);
1067 static void io_rsrc_put_work(struct work_struct *work);
1069 static void io_req_task_queue(struct io_kiocb *req);
1070 static void io_submit_flush_completions(struct io_ring_ctx *ctx);
1071 static bool io_poll_remove_waitqs(struct io_kiocb *req);
1072 static int io_req_prep_async(struct io_kiocb *req);
1074 static struct kmem_cache *req_cachep;
1076 static const struct file_operations io_uring_fops;
1078 struct sock *io_uring_get_socket(struct file *file)
1080 #if defined(CONFIG_UNIX)
1081 if (file->f_op == &io_uring_fops) {
1082 struct io_ring_ctx *ctx = file->private_data;
1084 return ctx->ring_sock->sk;
1089 EXPORT_SYMBOL(io_uring_get_socket);
1091 #define io_for_each_link(pos, head) \
1092 for (pos = (head); pos; pos = pos->link)
1094 static inline void io_req_set_rsrc_node(struct io_kiocb *req)
1096 struct io_ring_ctx *ctx = req->ctx;
1098 if (!req->fixed_rsrc_refs) {
1099 req->fixed_rsrc_refs = &ctx->rsrc_node->refs;
1100 percpu_ref_get(req->fixed_rsrc_refs);
1104 static void io_refs_resurrect(struct percpu_ref *ref, struct completion *compl)
1106 bool got = percpu_ref_tryget(ref);
1108 /* already at zero, wait for ->release() */
1110 wait_for_completion(compl);
1111 percpu_ref_resurrect(ref);
1113 percpu_ref_put(ref);
1116 static bool io_match_task(struct io_kiocb *head, struct task_struct *task,
1119 struct io_kiocb *req;
1121 if (task && head->task != task)
1126 io_for_each_link(req, head) {
1127 if (req->flags & REQ_F_INFLIGHT)
1133 static inline void req_set_fail(struct io_kiocb *req)
1135 req->flags |= REQ_F_FAIL;
1138 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1140 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1142 complete(&ctx->ref_comp);
1145 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1147 return !req->timeout.off;
1150 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1152 struct io_ring_ctx *ctx;
1155 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1160 * Use 5 bits less than the max cq entries, that should give us around
1161 * 32 entries per hash list if totally full and uniformly spread.
1163 hash_bits = ilog2(p->cq_entries);
1167 ctx->cancel_hash_bits = hash_bits;
1168 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1170 if (!ctx->cancel_hash)
1172 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1174 ctx->dummy_ubuf = kzalloc(sizeof(*ctx->dummy_ubuf), GFP_KERNEL);
1175 if (!ctx->dummy_ubuf)
1177 /* set invalid range, so io_import_fixed() fails meeting it */
1178 ctx->dummy_ubuf->ubuf = -1UL;
1180 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1181 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1184 ctx->flags = p->flags;
1185 init_waitqueue_head(&ctx->sqo_sq_wait);
1186 INIT_LIST_HEAD(&ctx->sqd_list);
1187 init_waitqueue_head(&ctx->poll_wait);
1188 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1189 init_completion(&ctx->ref_comp);
1190 xa_init_flags(&ctx->io_buffers, XA_FLAGS_ALLOC1);
1191 xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
1192 mutex_init(&ctx->uring_lock);
1193 init_waitqueue_head(&ctx->cq_wait);
1194 spin_lock_init(&ctx->completion_lock);
1195 INIT_LIST_HEAD(&ctx->iopoll_list);
1196 INIT_LIST_HEAD(&ctx->defer_list);
1197 INIT_LIST_HEAD(&ctx->timeout_list);
1198 spin_lock_init(&ctx->rsrc_ref_lock);
1199 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
1200 INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
1201 init_llist_head(&ctx->rsrc_put_llist);
1202 INIT_LIST_HEAD(&ctx->tctx_list);
1203 INIT_LIST_HEAD(&ctx->submit_state.comp.free_list);
1204 INIT_LIST_HEAD(&ctx->locked_free_list);
1207 kfree(ctx->dummy_ubuf);
1208 kfree(ctx->cancel_hash);
1213 static void io_account_cq_overflow(struct io_ring_ctx *ctx)
1215 struct io_rings *r = ctx->rings;
1217 WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1);
1221 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1223 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1224 struct io_ring_ctx *ctx = req->ctx;
1226 return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail;
1232 static void io_req_track_inflight(struct io_kiocb *req)
1234 if (!(req->flags & REQ_F_INFLIGHT)) {
1235 req->flags |= REQ_F_INFLIGHT;
1236 atomic_inc(¤t->io_uring->inflight_tracked);
1240 static void io_prep_async_work(struct io_kiocb *req)
1242 const struct io_op_def *def = &io_op_defs[req->opcode];
1243 struct io_ring_ctx *ctx = req->ctx;
1245 if (!(req->flags & REQ_F_CREDS)) {
1246 req->flags |= REQ_F_CREDS;
1247 req->creds = get_current_cred();
1250 req->work.list.next = NULL;
1251 req->work.flags = 0;
1252 if (req->flags & REQ_F_FORCE_ASYNC)
1253 req->work.flags |= IO_WQ_WORK_CONCURRENT;
1255 if (req->flags & REQ_F_ISREG) {
1256 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1257 io_wq_hash_work(&req->work, file_inode(req->file));
1258 } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
1259 if (def->unbound_nonreg_file)
1260 req->work.flags |= IO_WQ_WORK_UNBOUND;
1263 switch (req->opcode) {
1264 case IORING_OP_SPLICE:
1266 if (!S_ISREG(file_inode(req->splice.file_in)->i_mode))
1267 req->work.flags |= IO_WQ_WORK_UNBOUND;
1272 static void io_prep_async_link(struct io_kiocb *req)
1274 struct io_kiocb *cur;
1276 io_for_each_link(cur, req)
1277 io_prep_async_work(cur);
1280 static void io_queue_async_work(struct io_kiocb *req)
1282 struct io_ring_ctx *ctx = req->ctx;
1283 struct io_kiocb *link = io_prep_linked_timeout(req);
1284 struct io_uring_task *tctx = req->task->io_uring;
1287 BUG_ON(!tctx->io_wq);
1289 /* init ->work of the whole link before punting */
1290 io_prep_async_link(req);
1291 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1292 &req->work, req->flags);
1293 io_wq_enqueue(tctx->io_wq, &req->work);
1295 io_queue_linked_timeout(link);
1298 static void io_kill_timeout(struct io_kiocb *req, int status)
1299 __must_hold(&req->ctx->completion_lock)
1301 struct io_timeout_data *io = req->async_data;
1303 if (hrtimer_try_to_cancel(&io->timer) != -1) {
1304 atomic_set(&req->ctx->cq_timeouts,
1305 atomic_read(&req->ctx->cq_timeouts) + 1);
1306 list_del_init(&req->timeout.list);
1307 io_cqring_fill_event(req->ctx, req->user_data, status, 0);
1308 io_put_req_deferred(req, 1);
1312 static void io_queue_deferred(struct io_ring_ctx *ctx)
1314 while (!list_empty(&ctx->defer_list)) {
1315 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1316 struct io_defer_entry, list);
1318 if (req_need_defer(de->req, de->seq))
1320 list_del_init(&de->list);
1321 io_req_task_queue(de->req);
1326 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1328 u32 seq = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
1330 while (!list_empty(&ctx->timeout_list)) {
1331 u32 events_needed, events_got;
1332 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1333 struct io_kiocb, timeout.list);
1335 if (io_is_timeout_noseq(req))
1339 * Since seq can easily wrap around over time, subtract
1340 * the last seq at which timeouts were flushed before comparing.
1341 * Assuming not more than 2^31-1 events have happened since,
1342 * these subtractions won't have wrapped, so we can check if
1343 * target is in [last_seq, current_seq] by comparing the two.
1345 events_needed = req->timeout.target_seq - ctx->cq_last_tm_flush;
1346 events_got = seq - ctx->cq_last_tm_flush;
1347 if (events_got < events_needed)
1350 list_del_init(&req->timeout.list);
1351 io_kill_timeout(req, 0);
1353 ctx->cq_last_tm_flush = seq;
1356 static void __io_commit_cqring_flush(struct io_ring_ctx *ctx)
1358 if (ctx->off_timeout_used)
1359 io_flush_timeouts(ctx);
1360 if (ctx->drain_active)
1361 io_queue_deferred(ctx);
1364 static inline void io_commit_cqring(struct io_ring_ctx *ctx)
1366 if (unlikely(ctx->off_timeout_used || ctx->drain_active))
1367 __io_commit_cqring_flush(ctx);
1368 /* order cqe stores with ring update */
1369 smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
1372 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1374 struct io_rings *r = ctx->rings;
1376 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == ctx->sq_entries;
1379 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
1381 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
1384 static inline struct io_uring_cqe *io_get_cqe(struct io_ring_ctx *ctx)
1386 struct io_rings *rings = ctx->rings;
1387 unsigned tail, mask = ctx->cq_entries - 1;
1390 * writes to the cq entry need to come after reading head; the
1391 * control dependency is enough as we're using WRITE_ONCE to
1394 if (__io_cqring_events(ctx) == ctx->cq_entries)
1397 tail = ctx->cached_cq_tail++;
1398 return &rings->cqes[tail & mask];
1401 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1403 if (likely(!ctx->cq_ev_fd))
1405 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1407 return !ctx->eventfd_async || io_wq_current_is_worker();
1410 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1412 /* see waitqueue_active() comment */
1415 if (waitqueue_active(&ctx->cq_wait))
1416 wake_up(&ctx->cq_wait);
1417 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1418 wake_up(&ctx->sq_data->wait);
1419 if (io_should_trigger_evfd(ctx))
1420 eventfd_signal(ctx->cq_ev_fd, 1);
1421 if (waitqueue_active(&ctx->poll_wait)) {
1422 wake_up_interruptible(&ctx->poll_wait);
1423 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1427 static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx)
1429 /* see waitqueue_active() comment */
1432 if (ctx->flags & IORING_SETUP_SQPOLL) {
1433 if (waitqueue_active(&ctx->cq_wait))
1434 wake_up(&ctx->cq_wait);
1436 if (io_should_trigger_evfd(ctx))
1437 eventfd_signal(ctx->cq_ev_fd, 1);
1438 if (waitqueue_active(&ctx->poll_wait)) {
1439 wake_up_interruptible(&ctx->poll_wait);
1440 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1444 /* Returns true if there are no backlogged entries after the flush */
1445 static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1447 unsigned long flags;
1448 bool all_flushed, posted;
1450 if (!force && __io_cqring_events(ctx) == ctx->cq_entries)
1454 spin_lock_irqsave(&ctx->completion_lock, flags);
1455 while (!list_empty(&ctx->cq_overflow_list)) {
1456 struct io_uring_cqe *cqe = io_get_cqe(ctx);
1457 struct io_overflow_cqe *ocqe;
1461 ocqe = list_first_entry(&ctx->cq_overflow_list,
1462 struct io_overflow_cqe, list);
1464 memcpy(cqe, &ocqe->cqe, sizeof(*cqe));
1466 io_account_cq_overflow(ctx);
1469 list_del(&ocqe->list);
1473 all_flushed = list_empty(&ctx->cq_overflow_list);
1475 clear_bit(0, &ctx->check_cq_overflow);
1476 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
1480 io_commit_cqring(ctx);
1481 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1483 io_cqring_ev_posted(ctx);
1487 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1491 if (test_bit(0, &ctx->check_cq_overflow)) {
1492 /* iopoll syncs against uring_lock, not completion_lock */
1493 if (ctx->flags & IORING_SETUP_IOPOLL)
1494 mutex_lock(&ctx->uring_lock);
1495 ret = __io_cqring_overflow_flush(ctx, force);
1496 if (ctx->flags & IORING_SETUP_IOPOLL)
1497 mutex_unlock(&ctx->uring_lock);
1504 * Shamelessly stolen from the mm implementation of page reference checking,
1505 * see commit f958d7b528b1 for details.
1507 #define req_ref_zero_or_close_to_overflow(req) \
1508 ((unsigned int) atomic_read(&(req->refs)) + 127u <= 127u)
1510 static inline bool req_ref_inc_not_zero(struct io_kiocb *req)
1512 return atomic_inc_not_zero(&req->refs);
1515 static inline bool req_ref_sub_and_test(struct io_kiocb *req, int refs)
1517 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1518 return atomic_sub_and_test(refs, &req->refs);
1521 static inline bool req_ref_put_and_test(struct io_kiocb *req)
1523 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1524 return atomic_dec_and_test(&req->refs);
1527 static inline void req_ref_put(struct io_kiocb *req)
1529 WARN_ON_ONCE(req_ref_put_and_test(req));
1532 static inline void req_ref_get(struct io_kiocb *req)
1534 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1535 atomic_inc(&req->refs);
1538 static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data,
1539 long res, unsigned int cflags)
1541 struct io_overflow_cqe *ocqe;
1543 ocqe = kmalloc(sizeof(*ocqe), GFP_ATOMIC | __GFP_ACCOUNT);
1546 * If we're in ring overflow flush mode, or in task cancel mode,
1547 * or cannot allocate an overflow entry, then we need to drop it
1550 io_account_cq_overflow(ctx);
1553 if (list_empty(&ctx->cq_overflow_list)) {
1554 set_bit(0, &ctx->check_cq_overflow);
1555 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
1557 ocqe->cqe.user_data = user_data;
1558 ocqe->cqe.res = res;
1559 ocqe->cqe.flags = cflags;
1560 list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
1564 static inline bool __io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1565 long res, unsigned int cflags)
1567 struct io_uring_cqe *cqe;
1569 trace_io_uring_complete(ctx, user_data, res, cflags);
1572 * If we can't get a cq entry, userspace overflowed the
1573 * submission (by quite a lot). Increment the overflow count in
1576 cqe = io_get_cqe(ctx);
1578 WRITE_ONCE(cqe->user_data, user_data);
1579 WRITE_ONCE(cqe->res, res);
1580 WRITE_ONCE(cqe->flags, cflags);
1583 return io_cqring_event_overflow(ctx, user_data, res, cflags);
1586 /* not as hot to bloat with inlining */
1587 static noinline bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1588 long res, unsigned int cflags)
1590 return __io_cqring_fill_event(ctx, user_data, res, cflags);
1593 static void io_req_complete_post(struct io_kiocb *req, long res,
1594 unsigned int cflags)
1596 struct io_ring_ctx *ctx = req->ctx;
1597 unsigned long flags;
1599 spin_lock_irqsave(&ctx->completion_lock, flags);
1600 __io_cqring_fill_event(ctx, req->user_data, res, cflags);
1602 * If we're the last reference to this request, add to our locked
1605 if (req_ref_put_and_test(req)) {
1606 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
1607 if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL))
1608 io_disarm_next(req);
1610 io_req_task_queue(req->link);
1614 io_dismantle_req(req);
1615 io_put_task(req->task, 1);
1616 list_add(&req->compl.list, &ctx->locked_free_list);
1617 ctx->locked_free_nr++;
1619 if (!percpu_ref_tryget(&ctx->refs))
1622 io_commit_cqring(ctx);
1623 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1626 io_cqring_ev_posted(ctx);
1627 percpu_ref_put(&ctx->refs);
1631 static inline bool io_req_needs_clean(struct io_kiocb *req)
1633 return req->flags & IO_REQ_CLEAN_FLAGS;
1636 static void io_req_complete_state(struct io_kiocb *req, long res,
1637 unsigned int cflags)
1639 if (io_req_needs_clean(req))
1642 req->compl.cflags = cflags;
1643 req->flags |= REQ_F_COMPLETE_INLINE;
1646 static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags,
1647 long res, unsigned cflags)
1649 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1650 io_req_complete_state(req, res, cflags);
1652 io_req_complete_post(req, res, cflags);
1655 static inline void io_req_complete(struct io_kiocb *req, long res)
1657 __io_req_complete(req, 0, res, 0);
1660 static void io_req_complete_failed(struct io_kiocb *req, long res)
1664 io_req_complete_post(req, res, 0);
1667 static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
1668 struct io_comp_state *cs)
1670 spin_lock_irq(&ctx->completion_lock);
1671 list_splice_init(&ctx->locked_free_list, &cs->free_list);
1672 ctx->locked_free_nr = 0;
1673 spin_unlock_irq(&ctx->completion_lock);
1676 /* Returns true IFF there are requests in the cache */
1677 static bool io_flush_cached_reqs(struct io_ring_ctx *ctx)
1679 struct io_submit_state *state = &ctx->submit_state;
1680 struct io_comp_state *cs = &state->comp;
1684 * If we have more than a batch's worth of requests in our IRQ side
1685 * locked cache, grab the lock and move them over to our submission
1688 if (READ_ONCE(ctx->locked_free_nr) > IO_COMPL_BATCH)
1689 io_flush_cached_locked_reqs(ctx, cs);
1691 nr = state->free_reqs;
1692 while (!list_empty(&cs->free_list)) {
1693 struct io_kiocb *req = list_first_entry(&cs->free_list,
1694 struct io_kiocb, compl.list);
1696 list_del(&req->compl.list);
1697 state->reqs[nr++] = req;
1698 if (nr == ARRAY_SIZE(state->reqs))
1702 state->free_reqs = nr;
1706 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
1708 struct io_submit_state *state = &ctx->submit_state;
1710 BUILD_BUG_ON(ARRAY_SIZE(state->reqs) < IO_REQ_ALLOC_BATCH);
1712 if (!state->free_reqs) {
1713 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1716 if (io_flush_cached_reqs(ctx))
1719 ret = kmem_cache_alloc_bulk(req_cachep, gfp, IO_REQ_ALLOC_BATCH,
1723 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1724 * retry single alloc to be on the safe side.
1726 if (unlikely(ret <= 0)) {
1727 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1728 if (!state->reqs[0])
1732 state->free_reqs = ret;
1736 return state->reqs[state->free_reqs];
1739 static inline void io_put_file(struct file *file)
1745 static void io_dismantle_req(struct io_kiocb *req)
1747 unsigned int flags = req->flags;
1749 if (io_req_needs_clean(req))
1751 if (!(flags & REQ_F_FIXED_FILE))
1752 io_put_file(req->file);
1753 if (req->fixed_rsrc_refs)
1754 percpu_ref_put(req->fixed_rsrc_refs);
1755 if (req->async_data)
1756 kfree(req->async_data);
1759 /* must to be called somewhat shortly after putting a request */
1760 static inline void io_put_task(struct task_struct *task, int nr)
1762 struct io_uring_task *tctx = task->io_uring;
1764 percpu_counter_sub(&tctx->inflight, nr);
1765 if (unlikely(atomic_read(&tctx->in_idle)))
1766 wake_up(&tctx->wait);
1767 put_task_struct_many(task, nr);
1770 static void __io_free_req(struct io_kiocb *req)
1772 struct io_ring_ctx *ctx = req->ctx;
1774 io_dismantle_req(req);
1775 io_put_task(req->task, 1);
1777 kmem_cache_free(req_cachep, req);
1778 percpu_ref_put(&ctx->refs);
1781 static inline void io_remove_next_linked(struct io_kiocb *req)
1783 struct io_kiocb *nxt = req->link;
1785 req->link = nxt->link;
1789 static bool io_kill_linked_timeout(struct io_kiocb *req)
1790 __must_hold(&req->ctx->completion_lock)
1792 struct io_kiocb *link = req->link;
1795 * Can happen if a linked timeout fired and link had been like
1796 * req -> link t-out -> link t-out [-> ...]
1798 if (link && (link->flags & REQ_F_LTIMEOUT_ACTIVE)) {
1799 struct io_timeout_data *io = link->async_data;
1801 io_remove_next_linked(req);
1802 link->timeout.head = NULL;
1803 if (hrtimer_try_to_cancel(&io->timer) != -1) {
1804 io_cqring_fill_event(link->ctx, link->user_data,
1806 io_put_req_deferred(link, 1);
1813 static void io_fail_links(struct io_kiocb *req)
1814 __must_hold(&req->ctx->completion_lock)
1816 struct io_kiocb *nxt, *link = req->link;
1823 trace_io_uring_fail_link(req, link);
1824 io_cqring_fill_event(link->ctx, link->user_data, -ECANCELED, 0);
1825 io_put_req_deferred(link, 2);
1830 static bool io_disarm_next(struct io_kiocb *req)
1831 __must_hold(&req->ctx->completion_lock)
1833 bool posted = false;
1835 if (likely(req->flags & REQ_F_LINK_TIMEOUT))
1836 posted = io_kill_linked_timeout(req);
1837 if (unlikely((req->flags & REQ_F_FAIL) &&
1838 !(req->flags & REQ_F_HARDLINK))) {
1839 posted |= (req->link != NULL);
1845 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1847 struct io_kiocb *nxt;
1850 * If LINK is set, we have dependent requests in this chain. If we
1851 * didn't fail this request, queue the first one up, moving any other
1852 * dependencies to the next request. In case of failure, fail the rest
1855 if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL)) {
1856 struct io_ring_ctx *ctx = req->ctx;
1857 unsigned long flags;
1860 spin_lock_irqsave(&ctx->completion_lock, flags);
1861 posted = io_disarm_next(req);
1863 io_commit_cqring(req->ctx);
1864 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1866 io_cqring_ev_posted(ctx);
1873 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1875 if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK))))
1877 return __io_req_find_next(req);
1880 static void ctx_flush_and_put(struct io_ring_ctx *ctx)
1884 if (ctx->submit_state.comp.nr) {
1885 mutex_lock(&ctx->uring_lock);
1886 io_submit_flush_completions(ctx);
1887 mutex_unlock(&ctx->uring_lock);
1889 percpu_ref_put(&ctx->refs);
1892 static void tctx_task_work(struct callback_head *cb)
1894 struct io_ring_ctx *ctx = NULL;
1895 struct io_uring_task *tctx = container_of(cb, struct io_uring_task,
1899 struct io_wq_work_node *node;
1901 spin_lock_irq(&tctx->task_lock);
1902 node = tctx->task_list.first;
1903 INIT_WQ_LIST(&tctx->task_list);
1904 spin_unlock_irq(&tctx->task_lock);
1907 struct io_wq_work_node *next = node->next;
1908 struct io_kiocb *req = container_of(node, struct io_kiocb,
1911 if (req->ctx != ctx) {
1912 ctx_flush_and_put(ctx);
1914 percpu_ref_get(&ctx->refs);
1916 req->task_work.func(&req->task_work);
1919 if (wq_list_empty(&tctx->task_list)) {
1920 clear_bit(0, &tctx->task_state);
1921 if (wq_list_empty(&tctx->task_list))
1923 /* another tctx_task_work() is enqueued, yield */
1924 if (test_and_set_bit(0, &tctx->task_state))
1930 ctx_flush_and_put(ctx);
1933 static int io_req_task_work_add(struct io_kiocb *req)
1935 struct task_struct *tsk = req->task;
1936 struct io_uring_task *tctx = tsk->io_uring;
1937 enum task_work_notify_mode notify;
1938 struct io_wq_work_node *node, *prev;
1939 unsigned long flags;
1942 if (unlikely(tsk->flags & PF_EXITING))
1945 WARN_ON_ONCE(!tctx);
1947 spin_lock_irqsave(&tctx->task_lock, flags);
1948 wq_list_add_tail(&req->io_task_work.node, &tctx->task_list);
1949 spin_unlock_irqrestore(&tctx->task_lock, flags);
1951 /* task_work already pending, we're done */
1952 if (test_bit(0, &tctx->task_state) ||
1953 test_and_set_bit(0, &tctx->task_state))
1957 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
1958 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
1959 * processing task_work. There's no reliable way to tell if TWA_RESUME
1962 notify = (req->ctx->flags & IORING_SETUP_SQPOLL) ? TWA_NONE : TWA_SIGNAL;
1964 if (!task_work_add(tsk, &tctx->task_work, notify)) {
1965 wake_up_process(tsk);
1970 * Slow path - we failed, find and delete work. if the work is not
1971 * in the list, it got run and we're fine.
1973 spin_lock_irqsave(&tctx->task_lock, flags);
1974 wq_list_for_each(node, prev, &tctx->task_list) {
1975 if (&req->io_task_work.node == node) {
1976 wq_list_del(&tctx->task_list, node, prev);
1981 spin_unlock_irqrestore(&tctx->task_lock, flags);
1982 clear_bit(0, &tctx->task_state);
1986 static bool io_run_task_work_head(struct callback_head **work_head)
1988 struct callback_head *work, *next;
1989 bool executed = false;
1992 work = xchg(work_head, NULL);
2008 static void io_task_work_add_head(struct callback_head **work_head,
2009 struct callback_head *task_work)
2011 struct callback_head *head;
2014 head = READ_ONCE(*work_head);
2015 task_work->next = head;
2016 } while (cmpxchg(work_head, head, task_work) != head);
2019 static void io_req_task_work_add_fallback(struct io_kiocb *req,
2020 task_work_func_t cb)
2022 init_task_work(&req->task_work, cb);
2023 io_task_work_add_head(&req->ctx->exit_task_work, &req->task_work);
2026 static void io_req_task_cancel(struct callback_head *cb)
2028 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2029 struct io_ring_ctx *ctx = req->ctx;
2031 /* ctx is guaranteed to stay alive while we hold uring_lock */
2032 mutex_lock(&ctx->uring_lock);
2033 io_req_complete_failed(req, req->result);
2034 mutex_unlock(&ctx->uring_lock);
2037 static void __io_req_task_submit(struct io_kiocb *req)
2039 struct io_ring_ctx *ctx = req->ctx;
2041 /* ctx stays valid until unlock, even if we drop all ours ctx->refs */
2042 mutex_lock(&ctx->uring_lock);
2043 if (!(current->flags & PF_EXITING) && !current->in_execve)
2044 __io_queue_sqe(req);
2046 io_req_complete_failed(req, -EFAULT);
2047 mutex_unlock(&ctx->uring_lock);
2050 static void io_req_task_submit(struct callback_head *cb)
2052 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2054 __io_req_task_submit(req);
2057 static void io_req_task_queue_fail(struct io_kiocb *req, int ret)
2060 req->task_work.func = io_req_task_cancel;
2062 if (unlikely(io_req_task_work_add(req)))
2063 io_req_task_work_add_fallback(req, io_req_task_cancel);
2066 static void io_req_task_queue(struct io_kiocb *req)
2068 req->task_work.func = io_req_task_submit;
2070 if (unlikely(io_req_task_work_add(req)))
2071 io_req_task_queue_fail(req, -ECANCELED);
2074 static inline void io_queue_next(struct io_kiocb *req)
2076 struct io_kiocb *nxt = io_req_find_next(req);
2079 io_req_task_queue(nxt);
2082 static void io_free_req(struct io_kiocb *req)
2089 struct task_struct *task;
2094 static inline void io_init_req_batch(struct req_batch *rb)
2101 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2102 struct req_batch *rb)
2105 io_put_task(rb->task, rb->task_refs);
2107 percpu_ref_put_many(&ctx->refs, rb->ctx_refs);
2110 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req,
2111 struct io_submit_state *state)
2114 io_dismantle_req(req);
2116 if (req->task != rb->task) {
2118 io_put_task(rb->task, rb->task_refs);
2119 rb->task = req->task;
2125 if (state->free_reqs != ARRAY_SIZE(state->reqs))
2126 state->reqs[state->free_reqs++] = req;
2128 list_add(&req->compl.list, &state->comp.free_list);
2131 static void io_submit_flush_completions(struct io_ring_ctx *ctx)
2133 struct io_comp_state *cs = &ctx->submit_state.comp;
2135 struct io_kiocb *req;
2136 struct req_batch rb;
2138 io_init_req_batch(&rb);
2139 spin_lock_irq(&ctx->completion_lock);
2140 for (i = 0; i < nr; i++) {
2142 __io_cqring_fill_event(ctx, req->user_data, req->result,
2145 io_commit_cqring(ctx);
2146 spin_unlock_irq(&ctx->completion_lock);
2148 io_cqring_ev_posted(ctx);
2149 for (i = 0; i < nr; i++) {
2152 /* submission and completion refs */
2153 if (req_ref_sub_and_test(req, 2))
2154 io_req_free_batch(&rb, req, &ctx->submit_state);
2157 io_req_free_batch_finish(ctx, &rb);
2162 * Drop reference to request, return next in chain (if there is one) if this
2163 * was the last reference to this request.
2165 static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2167 struct io_kiocb *nxt = NULL;
2169 if (req_ref_put_and_test(req)) {
2170 nxt = io_req_find_next(req);
2176 static inline void io_put_req(struct io_kiocb *req)
2178 if (req_ref_put_and_test(req))
2182 static void io_put_req_deferred_cb(struct callback_head *cb)
2184 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2189 static void io_free_req_deferred(struct io_kiocb *req)
2191 req->task_work.func = io_put_req_deferred_cb;
2192 if (unlikely(io_req_task_work_add(req)))
2193 io_req_task_work_add_fallback(req, io_put_req_deferred_cb);
2196 static inline void io_put_req_deferred(struct io_kiocb *req, int refs)
2198 if (req_ref_sub_and_test(req, refs))
2199 io_free_req_deferred(req);
2202 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2204 /* See comment at the top of this file */
2206 return __io_cqring_events(ctx);
2209 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2211 struct io_rings *rings = ctx->rings;
2213 /* make sure SQ entry isn't read before tail */
2214 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2217 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2219 unsigned int cflags;
2221 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2222 cflags |= IORING_CQE_F_BUFFER;
2223 req->flags &= ~REQ_F_BUFFER_SELECTED;
2228 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2230 struct io_buffer *kbuf;
2232 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2233 return io_put_kbuf(req, kbuf);
2236 static inline bool io_run_task_work(void)
2239 * Not safe to run on exiting task, and the task_work handling will
2240 * not add work to such a task.
2242 if (unlikely(current->flags & PF_EXITING))
2244 if (current->task_works) {
2245 __set_current_state(TASK_RUNNING);
2254 * Find and free completed poll iocbs
2256 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2257 struct list_head *done)
2259 struct req_batch rb;
2260 struct io_kiocb *req;
2262 /* order with ->result store in io_complete_rw_iopoll() */
2265 io_init_req_batch(&rb);
2266 while (!list_empty(done)) {
2269 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2270 list_del(&req->inflight_entry);
2272 if (READ_ONCE(req->result) == -EAGAIN &&
2273 !(req->flags & REQ_F_DONT_REISSUE)) {
2274 req->iopoll_completed = 0;
2276 io_queue_async_work(req);
2280 if (req->flags & REQ_F_BUFFER_SELECTED)
2281 cflags = io_put_rw_kbuf(req);
2283 __io_cqring_fill_event(ctx, req->user_data, req->result, cflags);
2286 if (req_ref_put_and_test(req))
2287 io_req_free_batch(&rb, req, &ctx->submit_state);
2290 io_commit_cqring(ctx);
2291 io_cqring_ev_posted_iopoll(ctx);
2292 io_req_free_batch_finish(ctx, &rb);
2295 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2298 struct io_kiocb *req, *tmp;
2304 * Only spin for completions if we don't have multiple devices hanging
2305 * off our complete list, and we're under the requested amount.
2307 spin = !ctx->poll_multi_file && *nr_events < min;
2310 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2311 struct kiocb *kiocb = &req->rw.kiocb;
2314 * Move completed and retryable entries to our local lists.
2315 * If we find a request that requires polling, break out
2316 * and complete those lists first, if we have entries there.
2318 if (READ_ONCE(req->iopoll_completed)) {
2319 list_move_tail(&req->inflight_entry, &done);
2322 if (!list_empty(&done))
2325 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2329 /* iopoll may have completed current req */
2330 if (READ_ONCE(req->iopoll_completed))
2331 list_move_tail(&req->inflight_entry, &done);
2338 if (!list_empty(&done))
2339 io_iopoll_complete(ctx, nr_events, &done);
2345 * We can't just wait for polled events to come to us, we have to actively
2346 * find and complete them.
2348 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2350 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2353 mutex_lock(&ctx->uring_lock);
2354 while (!list_empty(&ctx->iopoll_list)) {
2355 unsigned int nr_events = 0;
2357 io_do_iopoll(ctx, &nr_events, 0);
2359 /* let it sleep and repeat later if can't complete a request */
2363 * Ensure we allow local-to-the-cpu processing to take place,
2364 * in this case we need to ensure that we reap all events.
2365 * Also let task_work, etc. to progress by releasing the mutex
2367 if (need_resched()) {
2368 mutex_unlock(&ctx->uring_lock);
2370 mutex_lock(&ctx->uring_lock);
2373 mutex_unlock(&ctx->uring_lock);
2376 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2378 unsigned int nr_events = 0;
2382 * We disallow the app entering submit/complete with polling, but we
2383 * still need to lock the ring to prevent racing with polled issue
2384 * that got punted to a workqueue.
2386 mutex_lock(&ctx->uring_lock);
2388 * Don't enter poll loop if we already have events pending.
2389 * If we do, we can potentially be spinning for commands that
2390 * already triggered a CQE (eg in error).
2392 if (test_bit(0, &ctx->check_cq_overflow))
2393 __io_cqring_overflow_flush(ctx, false);
2394 if (io_cqring_events(ctx))
2398 * If a submit got punted to a workqueue, we can have the
2399 * application entering polling for a command before it gets
2400 * issued. That app will hold the uring_lock for the duration
2401 * of the poll right here, so we need to take a breather every
2402 * now and then to ensure that the issue has a chance to add
2403 * the poll to the issued list. Otherwise we can spin here
2404 * forever, while the workqueue is stuck trying to acquire the
2407 if (list_empty(&ctx->iopoll_list)) {
2408 mutex_unlock(&ctx->uring_lock);
2410 mutex_lock(&ctx->uring_lock);
2412 if (list_empty(&ctx->iopoll_list))
2415 ret = io_do_iopoll(ctx, &nr_events, min);
2416 } while (!ret && nr_events < min && !need_resched());
2418 mutex_unlock(&ctx->uring_lock);
2422 static void kiocb_end_write(struct io_kiocb *req)
2425 * Tell lockdep we inherited freeze protection from submission
2428 if (req->flags & REQ_F_ISREG) {
2429 struct super_block *sb = file_inode(req->file)->i_sb;
2431 __sb_writers_acquired(sb, SB_FREEZE_WRITE);
2437 static bool io_resubmit_prep(struct io_kiocb *req)
2439 struct io_async_rw *rw = req->async_data;
2442 return !io_req_prep_async(req);
2443 /* may have left rw->iter inconsistent on -EIOCBQUEUED */
2444 iov_iter_revert(&rw->iter, req->result - iov_iter_count(&rw->iter));
2448 static bool io_rw_should_reissue(struct io_kiocb *req)
2450 umode_t mode = file_inode(req->file)->i_mode;
2451 struct io_ring_ctx *ctx = req->ctx;
2453 if (!S_ISBLK(mode) && !S_ISREG(mode))
2455 if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
2456 !(ctx->flags & IORING_SETUP_IOPOLL)))
2459 * If ref is dying, we might be running poll reap from the exit work.
2460 * Don't attempt to reissue from that path, just let it fail with
2463 if (percpu_ref_is_dying(&ctx->refs))
2468 static bool io_resubmit_prep(struct io_kiocb *req)
2472 static bool io_rw_should_reissue(struct io_kiocb *req)
2478 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2479 unsigned int issue_flags)
2483 if (req->rw.kiocb.ki_flags & IOCB_WRITE)
2484 kiocb_end_write(req);
2485 if (res != req->result) {
2486 if ((res == -EAGAIN || res == -EOPNOTSUPP) &&
2487 io_rw_should_reissue(req)) {
2488 req->flags |= REQ_F_REISSUE;
2493 if (req->flags & REQ_F_BUFFER_SELECTED)
2494 cflags = io_put_rw_kbuf(req);
2495 __io_req_complete(req, issue_flags, res, cflags);
2498 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2500 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2502 __io_complete_rw(req, res, res2, 0);
2505 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2507 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2509 if (kiocb->ki_flags & IOCB_WRITE)
2510 kiocb_end_write(req);
2511 if (unlikely(res != req->result)) {
2512 if (!(res == -EAGAIN && io_rw_should_reissue(req) &&
2513 io_resubmit_prep(req))) {
2515 req->flags |= REQ_F_DONT_REISSUE;
2519 WRITE_ONCE(req->result, res);
2520 /* order with io_iopoll_complete() checking ->result */
2522 WRITE_ONCE(req->iopoll_completed, 1);
2526 * After the iocb has been issued, it's safe to be found on the poll list.
2527 * Adding the kiocb to the list AFTER submission ensures that we don't
2528 * find it from a io_do_iopoll() thread before the issuer is done
2529 * accessing the kiocb cookie.
2531 static void io_iopoll_req_issued(struct io_kiocb *req)
2533 struct io_ring_ctx *ctx = req->ctx;
2534 const bool in_async = io_wq_current_is_worker();
2536 /* workqueue context doesn't hold uring_lock, grab it now */
2537 if (unlikely(in_async))
2538 mutex_lock(&ctx->uring_lock);
2541 * Track whether we have multiple files in our lists. This will impact
2542 * how we do polling eventually, not spinning if we're on potentially
2543 * different devices.
2545 if (list_empty(&ctx->iopoll_list)) {
2546 ctx->poll_multi_file = false;
2547 } else if (!ctx->poll_multi_file) {
2548 struct io_kiocb *list_req;
2550 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2552 if (list_req->file != req->file)
2553 ctx->poll_multi_file = true;
2557 * For fast devices, IO may have already completed. If it has, add
2558 * it to the front so we find it first.
2560 if (READ_ONCE(req->iopoll_completed))
2561 list_add(&req->inflight_entry, &ctx->iopoll_list);
2563 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2565 if (unlikely(in_async)) {
2567 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
2568 * in sq thread task context or in io worker task context. If
2569 * current task context is sq thread, we don't need to check
2570 * whether should wake up sq thread.
2572 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2573 wq_has_sleeper(&ctx->sq_data->wait))
2574 wake_up(&ctx->sq_data->wait);
2576 mutex_unlock(&ctx->uring_lock);
2580 static inline void io_state_file_put(struct io_submit_state *state)
2582 if (state->file_refs) {
2583 fput_many(state->file, state->file_refs);
2584 state->file_refs = 0;
2589 * Get as many references to a file as we have IOs left in this submission,
2590 * assuming most submissions are for one file, or at least that each file
2591 * has more than one submission.
2593 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2598 if (state->file_refs) {
2599 if (state->fd == fd) {
2603 io_state_file_put(state);
2605 state->file = fget_many(fd, state->ios_left);
2606 if (unlikely(!state->file))
2610 state->file_refs = state->ios_left - 1;
2614 static bool io_bdev_nowait(struct block_device *bdev)
2616 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2620 * If we tracked the file through the SCM inflight mechanism, we could support
2621 * any file. For now, just ensure that anything potentially problematic is done
2624 static bool __io_file_supports_async(struct file *file, int rw)
2626 umode_t mode = file_inode(file)->i_mode;
2628 if (S_ISBLK(mode)) {
2629 if (IS_ENABLED(CONFIG_BLOCK) &&
2630 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2636 if (S_ISREG(mode)) {
2637 if (IS_ENABLED(CONFIG_BLOCK) &&
2638 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2639 file->f_op != &io_uring_fops)
2644 /* any ->read/write should understand O_NONBLOCK */
2645 if (file->f_flags & O_NONBLOCK)
2648 if (!(file->f_mode & FMODE_NOWAIT))
2652 return file->f_op->read_iter != NULL;
2654 return file->f_op->write_iter != NULL;
2657 static bool io_file_supports_async(struct io_kiocb *req, int rw)
2659 if (rw == READ && (req->flags & REQ_F_ASYNC_READ))
2661 else if (rw == WRITE && (req->flags & REQ_F_ASYNC_WRITE))
2664 return __io_file_supports_async(req->file, rw);
2667 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2669 struct io_ring_ctx *ctx = req->ctx;
2670 struct kiocb *kiocb = &req->rw.kiocb;
2671 struct file *file = req->file;
2675 if (!(req->flags & REQ_F_ISREG) && S_ISREG(file_inode(file)->i_mode))
2676 req->flags |= REQ_F_ISREG;
2678 kiocb->ki_pos = READ_ONCE(sqe->off);
2679 if (kiocb->ki_pos == -1 && !(file->f_mode & FMODE_STREAM)) {
2680 req->flags |= REQ_F_CUR_POS;
2681 kiocb->ki_pos = file->f_pos;
2683 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2684 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2685 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2689 /* don't allow async punt for O_NONBLOCK or RWF_NOWAIT */
2690 if ((kiocb->ki_flags & IOCB_NOWAIT) || (file->f_flags & O_NONBLOCK))
2691 req->flags |= REQ_F_NOWAIT;
2693 ioprio = READ_ONCE(sqe->ioprio);
2695 ret = ioprio_check_cap(ioprio);
2699 kiocb->ki_ioprio = ioprio;
2701 kiocb->ki_ioprio = get_current_ioprio();
2703 if (ctx->flags & IORING_SETUP_IOPOLL) {
2704 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2705 !kiocb->ki_filp->f_op->iopoll)
2708 kiocb->ki_flags |= IOCB_HIPRI;
2709 kiocb->ki_complete = io_complete_rw_iopoll;
2710 req->iopoll_completed = 0;
2712 if (kiocb->ki_flags & IOCB_HIPRI)
2714 kiocb->ki_complete = io_complete_rw;
2717 if (req->opcode == IORING_OP_READ_FIXED ||
2718 req->opcode == IORING_OP_WRITE_FIXED) {
2720 io_req_set_rsrc_node(req);
2723 req->rw.addr = READ_ONCE(sqe->addr);
2724 req->rw.len = READ_ONCE(sqe->len);
2725 req->buf_index = READ_ONCE(sqe->buf_index);
2729 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2735 case -ERESTARTNOINTR:
2736 case -ERESTARTNOHAND:
2737 case -ERESTART_RESTARTBLOCK:
2739 * We can't just restart the syscall, since previously
2740 * submitted sqes may already be in progress. Just fail this
2746 kiocb->ki_complete(kiocb, ret, 0);
2750 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2751 unsigned int issue_flags)
2753 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2754 struct io_async_rw *io = req->async_data;
2755 bool check_reissue = kiocb->ki_complete == io_complete_rw;
2757 /* add previously done IO, if any */
2758 if (io && io->bytes_done > 0) {
2760 ret = io->bytes_done;
2762 ret += io->bytes_done;
2765 if (req->flags & REQ_F_CUR_POS)
2766 req->file->f_pos = kiocb->ki_pos;
2767 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2768 __io_complete_rw(req, ret, 0, issue_flags);
2770 io_rw_done(kiocb, ret);
2772 if (check_reissue && (req->flags & REQ_F_REISSUE)) {
2773 req->flags &= ~REQ_F_REISSUE;
2774 if (io_resubmit_prep(req)) {
2776 io_queue_async_work(req);
2781 if (req->flags & REQ_F_BUFFER_SELECTED)
2782 cflags = io_put_rw_kbuf(req);
2783 __io_req_complete(req, issue_flags, ret, cflags);
2788 static int __io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter,
2789 struct io_mapped_ubuf *imu)
2791 size_t len = req->rw.len;
2792 u64 buf_end, buf_addr = req->rw.addr;
2795 if (unlikely(check_add_overflow(buf_addr, (u64)len, &buf_end)))
2797 /* not inside the mapped region */
2798 if (unlikely(buf_addr < imu->ubuf || buf_end > imu->ubuf_end))
2802 * May not be a start of buffer, set size appropriately
2803 * and advance us to the beginning.
2805 offset = buf_addr - imu->ubuf;
2806 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2810 * Don't use iov_iter_advance() here, as it's really slow for
2811 * using the latter parts of a big fixed buffer - it iterates
2812 * over each segment manually. We can cheat a bit here, because
2815 * 1) it's a BVEC iter, we set it up
2816 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2817 * first and last bvec
2819 * So just find our index, and adjust the iterator afterwards.
2820 * If the offset is within the first bvec (or the whole first
2821 * bvec, just use iov_iter_advance(). This makes it easier
2822 * since we can just skip the first segment, which may not
2823 * be PAGE_SIZE aligned.
2825 const struct bio_vec *bvec = imu->bvec;
2827 if (offset <= bvec->bv_len) {
2828 iov_iter_advance(iter, offset);
2830 unsigned long seg_skip;
2832 /* skip first vec */
2833 offset -= bvec->bv_len;
2834 seg_skip = 1 + (offset >> PAGE_SHIFT);
2836 iter->bvec = bvec + seg_skip;
2837 iter->nr_segs -= seg_skip;
2838 iter->count -= bvec->bv_len + offset;
2839 iter->iov_offset = offset & ~PAGE_MASK;
2846 static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
2848 struct io_ring_ctx *ctx = req->ctx;
2849 struct io_mapped_ubuf *imu = req->imu;
2850 u16 index, buf_index = req->buf_index;
2853 if (unlikely(buf_index >= ctx->nr_user_bufs))
2855 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2856 imu = READ_ONCE(ctx->user_bufs[index]);
2859 return __io_import_fixed(req, rw, iter, imu);
2862 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2865 mutex_unlock(&ctx->uring_lock);
2868 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2871 * "Normal" inline submissions always hold the uring_lock, since we
2872 * grab it from the system call. Same is true for the SQPOLL offload.
2873 * The only exception is when we've detached the request and issue it
2874 * from an async worker thread, grab the lock for that case.
2877 mutex_lock(&ctx->uring_lock);
2880 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2881 int bgid, struct io_buffer *kbuf,
2884 struct io_buffer *head;
2886 if (req->flags & REQ_F_BUFFER_SELECTED)
2889 io_ring_submit_lock(req->ctx, needs_lock);
2891 lockdep_assert_held(&req->ctx->uring_lock);
2893 head = xa_load(&req->ctx->io_buffers, bgid);
2895 if (!list_empty(&head->list)) {
2896 kbuf = list_last_entry(&head->list, struct io_buffer,
2898 list_del(&kbuf->list);
2901 xa_erase(&req->ctx->io_buffers, bgid);
2903 if (*len > kbuf->len)
2906 kbuf = ERR_PTR(-ENOBUFS);
2909 io_ring_submit_unlock(req->ctx, needs_lock);
2914 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2917 struct io_buffer *kbuf;
2920 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2921 bgid = req->buf_index;
2922 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2925 req->rw.addr = (u64) (unsigned long) kbuf;
2926 req->flags |= REQ_F_BUFFER_SELECTED;
2927 return u64_to_user_ptr(kbuf->addr);
2930 #ifdef CONFIG_COMPAT
2931 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2934 struct compat_iovec __user *uiov;
2935 compat_ssize_t clen;
2939 uiov = u64_to_user_ptr(req->rw.addr);
2940 if (!access_ok(uiov, sizeof(*uiov)))
2942 if (__get_user(clen, &uiov->iov_len))
2948 buf = io_rw_buffer_select(req, &len, needs_lock);
2950 return PTR_ERR(buf);
2951 iov[0].iov_base = buf;
2952 iov[0].iov_len = (compat_size_t) len;
2957 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2960 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2964 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2967 len = iov[0].iov_len;
2970 buf = io_rw_buffer_select(req, &len, needs_lock);
2972 return PTR_ERR(buf);
2973 iov[0].iov_base = buf;
2974 iov[0].iov_len = len;
2978 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2981 if (req->flags & REQ_F_BUFFER_SELECTED) {
2982 struct io_buffer *kbuf;
2984 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2985 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
2986 iov[0].iov_len = kbuf->len;
2989 if (req->rw.len != 1)
2992 #ifdef CONFIG_COMPAT
2993 if (req->ctx->compat)
2994 return io_compat_import(req, iov, needs_lock);
2997 return __io_iov_buffer_select(req, iov, needs_lock);
3000 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
3001 struct iov_iter *iter, bool needs_lock)
3003 void __user *buf = u64_to_user_ptr(req->rw.addr);
3004 size_t sqe_len = req->rw.len;
3005 u8 opcode = req->opcode;
3008 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3010 return io_import_fixed(req, rw, iter);
3013 /* buffer index only valid with fixed read/write, or buffer select */
3014 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3017 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3018 if (req->flags & REQ_F_BUFFER_SELECT) {
3019 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3021 return PTR_ERR(buf);
3022 req->rw.len = sqe_len;
3025 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3030 if (req->flags & REQ_F_BUFFER_SELECT) {
3031 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3033 iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
3038 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3042 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3044 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3048 * For files that don't have ->read_iter() and ->write_iter(), handle them
3049 * by looping over ->read() or ->write() manually.
3051 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3053 struct kiocb *kiocb = &req->rw.kiocb;
3054 struct file *file = req->file;
3058 * Don't support polled IO through this interface, and we can't
3059 * support non-blocking either. For the latter, this just causes
3060 * the kiocb to be handled from an async context.
3062 if (kiocb->ki_flags & IOCB_HIPRI)
3064 if (kiocb->ki_flags & IOCB_NOWAIT)
3067 while (iov_iter_count(iter)) {
3071 if (!iov_iter_is_bvec(iter)) {
3072 iovec = iov_iter_iovec(iter);
3074 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3075 iovec.iov_len = req->rw.len;
3079 nr = file->f_op->read(file, iovec.iov_base,
3080 iovec.iov_len, io_kiocb_ppos(kiocb));
3082 nr = file->f_op->write(file, iovec.iov_base,
3083 iovec.iov_len, io_kiocb_ppos(kiocb));
3092 if (nr != iovec.iov_len)
3096 iov_iter_advance(iter, nr);
3102 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3103 const struct iovec *fast_iov, struct iov_iter *iter)
3105 struct io_async_rw *rw = req->async_data;
3107 memcpy(&rw->iter, iter, sizeof(*iter));
3108 rw->free_iovec = iovec;
3110 /* can only be fixed buffers, no need to do anything */
3111 if (iov_iter_is_bvec(iter))
3114 unsigned iov_off = 0;
3116 rw->iter.iov = rw->fast_iov;
3117 if (iter->iov != fast_iov) {
3118 iov_off = iter->iov - fast_iov;
3119 rw->iter.iov += iov_off;
3121 if (rw->fast_iov != fast_iov)
3122 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3123 sizeof(struct iovec) * iter->nr_segs);
3125 req->flags |= REQ_F_NEED_CLEANUP;
3129 static inline int io_alloc_async_data(struct io_kiocb *req)
3131 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3132 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3133 return req->async_data == NULL;
3136 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3137 const struct iovec *fast_iov,
3138 struct iov_iter *iter, bool force)
3140 if (!force && !io_op_defs[req->opcode].needs_async_setup)
3142 if (!req->async_data) {
3143 if (io_alloc_async_data(req)) {
3148 io_req_map_rw(req, iovec, fast_iov, iter);
3153 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3155 struct io_async_rw *iorw = req->async_data;
3156 struct iovec *iov = iorw->fast_iov;
3159 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3160 if (unlikely(ret < 0))
3163 iorw->bytes_done = 0;
3164 iorw->free_iovec = iov;
3166 req->flags |= REQ_F_NEED_CLEANUP;
3170 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3172 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3174 return io_prep_rw(req, sqe);
3178 * This is our waitqueue callback handler, registered through lock_page_async()
3179 * when we initially tried to do the IO with the iocb armed our waitqueue.
3180 * This gets called when the page is unlocked, and we generally expect that to
3181 * happen when the page IO is completed and the page is now uptodate. This will
3182 * queue a task_work based retry of the operation, attempting to copy the data
3183 * again. If the latter fails because the page was NOT uptodate, then we will
3184 * do a thread based blocking retry of the operation. That's the unexpected
3187 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3188 int sync, void *arg)
3190 struct wait_page_queue *wpq;
3191 struct io_kiocb *req = wait->private;
3192 struct wait_page_key *key = arg;
3194 wpq = container_of(wait, struct wait_page_queue, wait);
3196 if (!wake_page_match(wpq, key))
3199 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3200 list_del_init(&wait->entry);
3202 /* submit ref gets dropped, acquire a new one */
3204 io_req_task_queue(req);
3209 * This controls whether a given IO request should be armed for async page
3210 * based retry. If we return false here, the request is handed to the async
3211 * worker threads for retry. If we're doing buffered reads on a regular file,
3212 * we prepare a private wait_page_queue entry and retry the operation. This
3213 * will either succeed because the page is now uptodate and unlocked, or it
3214 * will register a callback when the page is unlocked at IO completion. Through
3215 * that callback, io_uring uses task_work to setup a retry of the operation.
3216 * That retry will attempt the buffered read again. The retry will generally
3217 * succeed, or in rare cases where it fails, we then fall back to using the
3218 * async worker threads for a blocking retry.
3220 static bool io_rw_should_retry(struct io_kiocb *req)
3222 struct io_async_rw *rw = req->async_data;
3223 struct wait_page_queue *wait = &rw->wpq;
3224 struct kiocb *kiocb = &req->rw.kiocb;
3226 /* never retry for NOWAIT, we just complete with -EAGAIN */
3227 if (req->flags & REQ_F_NOWAIT)
3230 /* Only for buffered IO */
3231 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3235 * just use poll if we can, and don't attempt if the fs doesn't
3236 * support callback based unlocks
3238 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3241 wait->wait.func = io_async_buf_func;
3242 wait->wait.private = req;
3243 wait->wait.flags = 0;
3244 INIT_LIST_HEAD(&wait->wait.entry);
3245 kiocb->ki_flags |= IOCB_WAITQ;
3246 kiocb->ki_flags &= ~IOCB_NOWAIT;
3247 kiocb->ki_waitq = wait;
3251 static inline int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3253 if (req->file->f_op->read_iter)
3254 return call_read_iter(req->file, &req->rw.kiocb, iter);
3255 else if (req->file->f_op->read)
3256 return loop_rw_iter(READ, req, iter);
3261 static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3263 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3264 struct kiocb *kiocb = &req->rw.kiocb;
3265 struct iov_iter __iter, *iter = &__iter;
3266 struct io_async_rw *rw = req->async_data;
3267 ssize_t io_size, ret, ret2;
3268 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3274 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3278 io_size = iov_iter_count(iter);
3279 req->result = io_size;
3281 /* Ensure we clear previously set non-block flag */
3282 if (!force_nonblock)
3283 kiocb->ki_flags &= ~IOCB_NOWAIT;
3285 kiocb->ki_flags |= IOCB_NOWAIT;
3287 /* If the file doesn't support async, just async punt */
3288 if (force_nonblock && !io_file_supports_async(req, READ)) {
3289 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3290 return ret ?: -EAGAIN;
3293 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size);
3294 if (unlikely(ret)) {
3299 ret = io_iter_do_read(req, iter);
3301 if (ret == -EAGAIN || (req->flags & REQ_F_REISSUE)) {
3302 req->flags &= ~REQ_F_REISSUE;
3303 /* IOPOLL retry should happen for io-wq threads */
3304 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3306 /* no retry on NONBLOCK nor RWF_NOWAIT */
3307 if (req->flags & REQ_F_NOWAIT)
3309 /* some cases will consume bytes even on error returns */
3310 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3312 } else if (ret == -EIOCBQUEUED) {
3314 } else if (ret <= 0 || ret == io_size || !force_nonblock ||
3315 (req->flags & REQ_F_NOWAIT) || !(req->flags & REQ_F_ISREG)) {
3316 /* read all, failed, already did sync or don't want to retry */
3320 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3325 rw = req->async_data;
3326 /* now use our persistent iterator, if we aren't already */
3331 rw->bytes_done += ret;
3332 /* if we can retry, do so with the callbacks armed */
3333 if (!io_rw_should_retry(req)) {
3334 kiocb->ki_flags &= ~IOCB_WAITQ;
3339 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3340 * we get -EIOCBQUEUED, then we'll get a notification when the
3341 * desired page gets unlocked. We can also get a partial read
3342 * here, and if we do, then just retry at the new offset.
3344 ret = io_iter_do_read(req, iter);
3345 if (ret == -EIOCBQUEUED)
3347 /* we got some bytes, but not all. retry. */
3348 kiocb->ki_flags &= ~IOCB_WAITQ;
3349 } while (ret > 0 && ret < io_size);
3351 kiocb_done(kiocb, ret, issue_flags);
3353 /* it's faster to check here then delegate to kfree */
3359 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3361 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3363 return io_prep_rw(req, sqe);
3366 static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3368 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3369 struct kiocb *kiocb = &req->rw.kiocb;
3370 struct iov_iter __iter, *iter = &__iter;
3371 struct io_async_rw *rw = req->async_data;
3372 ssize_t ret, ret2, io_size;
3373 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3379 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3383 io_size = iov_iter_count(iter);
3384 req->result = io_size;
3386 /* Ensure we clear previously set non-block flag */
3387 if (!force_nonblock)
3388 kiocb->ki_flags &= ~IOCB_NOWAIT;
3390 kiocb->ki_flags |= IOCB_NOWAIT;
3392 /* If the file doesn't support async, just async punt */
3393 if (force_nonblock && !io_file_supports_async(req, WRITE))
3396 /* file path doesn't support NOWAIT for non-direct_IO */
3397 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3398 (req->flags & REQ_F_ISREG))
3401 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size);
3406 * Open-code file_start_write here to grab freeze protection,
3407 * which will be released by another thread in
3408 * io_complete_rw(). Fool lockdep by telling it the lock got
3409 * released so that it doesn't complain about the held lock when
3410 * we return to userspace.
3412 if (req->flags & REQ_F_ISREG) {
3413 sb_start_write(file_inode(req->file)->i_sb);
3414 __sb_writers_release(file_inode(req->file)->i_sb,
3417 kiocb->ki_flags |= IOCB_WRITE;
3419 if (req->file->f_op->write_iter)
3420 ret2 = call_write_iter(req->file, kiocb, iter);
3421 else if (req->file->f_op->write)
3422 ret2 = loop_rw_iter(WRITE, req, iter);
3426 if (req->flags & REQ_F_REISSUE) {
3427 req->flags &= ~REQ_F_REISSUE;
3432 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3433 * retry them without IOCB_NOWAIT.
3435 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3437 /* no retry on NONBLOCK nor RWF_NOWAIT */
3438 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3440 if (!force_nonblock || ret2 != -EAGAIN) {
3441 /* IOPOLL retry should happen for io-wq threads */
3442 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3445 kiocb_done(kiocb, ret2, issue_flags);
3448 /* some cases will consume bytes even on error returns */
3449 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3450 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3451 return ret ?: -EAGAIN;
3454 /* it's reportedly faster than delegating the null check to kfree() */
3460 static int io_renameat_prep(struct io_kiocb *req,
3461 const struct io_uring_sqe *sqe)
3463 struct io_rename *ren = &req->rename;
3464 const char __user *oldf, *newf;
3466 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3469 ren->old_dfd = READ_ONCE(sqe->fd);
3470 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3471 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3472 ren->new_dfd = READ_ONCE(sqe->len);
3473 ren->flags = READ_ONCE(sqe->rename_flags);
3475 ren->oldpath = getname(oldf);
3476 if (IS_ERR(ren->oldpath))
3477 return PTR_ERR(ren->oldpath);
3479 ren->newpath = getname(newf);
3480 if (IS_ERR(ren->newpath)) {
3481 putname(ren->oldpath);
3482 return PTR_ERR(ren->newpath);
3485 req->flags |= REQ_F_NEED_CLEANUP;
3489 static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3491 struct io_rename *ren = &req->rename;
3494 if (issue_flags & IO_URING_F_NONBLOCK)
3497 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3498 ren->newpath, ren->flags);
3500 req->flags &= ~REQ_F_NEED_CLEANUP;
3503 io_req_complete(req, ret);
3507 static int io_unlinkat_prep(struct io_kiocb *req,
3508 const struct io_uring_sqe *sqe)
3510 struct io_unlink *un = &req->unlink;
3511 const char __user *fname;
3513 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3516 un->dfd = READ_ONCE(sqe->fd);
3518 un->flags = READ_ONCE(sqe->unlink_flags);
3519 if (un->flags & ~AT_REMOVEDIR)
3522 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3523 un->filename = getname(fname);
3524 if (IS_ERR(un->filename))
3525 return PTR_ERR(un->filename);
3527 req->flags |= REQ_F_NEED_CLEANUP;
3531 static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3533 struct io_unlink *un = &req->unlink;
3536 if (issue_flags & IO_URING_F_NONBLOCK)
3539 if (un->flags & AT_REMOVEDIR)
3540 ret = do_rmdir(un->dfd, un->filename);
3542 ret = do_unlinkat(un->dfd, un->filename);
3544 req->flags &= ~REQ_F_NEED_CLEANUP;
3547 io_req_complete(req, ret);
3551 static int io_shutdown_prep(struct io_kiocb *req,
3552 const struct io_uring_sqe *sqe)
3554 #if defined(CONFIG_NET)
3555 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3557 if (sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3561 req->shutdown.how = READ_ONCE(sqe->len);
3568 static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3570 #if defined(CONFIG_NET)
3571 struct socket *sock;
3574 if (issue_flags & IO_URING_F_NONBLOCK)
3577 sock = sock_from_file(req->file);
3578 if (unlikely(!sock))
3581 ret = __sys_shutdown_sock(sock, req->shutdown.how);
3584 io_req_complete(req, ret);
3591 static int __io_splice_prep(struct io_kiocb *req,
3592 const struct io_uring_sqe *sqe)
3594 struct io_splice *sp = &req->splice;
3595 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3597 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3601 sp->len = READ_ONCE(sqe->len);
3602 sp->flags = READ_ONCE(sqe->splice_flags);
3604 if (unlikely(sp->flags & ~valid_flags))
3607 sp->file_in = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in),
3608 (sp->flags & SPLICE_F_FD_IN_FIXED));
3611 req->flags |= REQ_F_NEED_CLEANUP;
3615 static int io_tee_prep(struct io_kiocb *req,
3616 const struct io_uring_sqe *sqe)
3618 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3620 return __io_splice_prep(req, sqe);
3623 static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
3625 struct io_splice *sp = &req->splice;
3626 struct file *in = sp->file_in;
3627 struct file *out = sp->file_out;
3628 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3631 if (issue_flags & IO_URING_F_NONBLOCK)
3634 ret = do_tee(in, out, sp->len, flags);
3636 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3638 req->flags &= ~REQ_F_NEED_CLEANUP;
3642 io_req_complete(req, ret);
3646 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3648 struct io_splice *sp = &req->splice;
3650 sp->off_in = READ_ONCE(sqe->splice_off_in);
3651 sp->off_out = READ_ONCE(sqe->off);
3652 return __io_splice_prep(req, sqe);
3655 static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
3657 struct io_splice *sp = &req->splice;
3658 struct file *in = sp->file_in;
3659 struct file *out = sp->file_out;
3660 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3661 loff_t *poff_in, *poff_out;
3664 if (issue_flags & IO_URING_F_NONBLOCK)
3667 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3668 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3671 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3673 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3675 req->flags &= ~REQ_F_NEED_CLEANUP;
3679 io_req_complete(req, ret);
3684 * IORING_OP_NOP just posts a completion event, nothing else.
3686 static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
3688 struct io_ring_ctx *ctx = req->ctx;
3690 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3693 __io_req_complete(req, issue_flags, 0, 0);
3697 static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3699 struct io_ring_ctx *ctx = req->ctx;
3704 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3706 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3709 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3710 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3713 req->sync.off = READ_ONCE(sqe->off);
3714 req->sync.len = READ_ONCE(sqe->len);
3718 static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
3720 loff_t end = req->sync.off + req->sync.len;
3723 /* fsync always requires a blocking context */
3724 if (issue_flags & IO_URING_F_NONBLOCK)
3727 ret = vfs_fsync_range(req->file, req->sync.off,
3728 end > 0 ? end : LLONG_MAX,
3729 req->sync.flags & IORING_FSYNC_DATASYNC);
3732 io_req_complete(req, ret);
3736 static int io_fallocate_prep(struct io_kiocb *req,
3737 const struct io_uring_sqe *sqe)
3739 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3741 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3744 req->sync.off = READ_ONCE(sqe->off);
3745 req->sync.len = READ_ONCE(sqe->addr);
3746 req->sync.mode = READ_ONCE(sqe->len);
3750 static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
3754 /* fallocate always requiring blocking context */
3755 if (issue_flags & IO_URING_F_NONBLOCK)
3757 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3761 io_req_complete(req, ret);
3765 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3767 const char __user *fname;
3770 if (unlikely(sqe->ioprio || sqe->buf_index))
3772 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3775 /* open.how should be already initialised */
3776 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3777 req->open.how.flags |= O_LARGEFILE;
3779 req->open.dfd = READ_ONCE(sqe->fd);
3780 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3781 req->open.filename = getname(fname);
3782 if (IS_ERR(req->open.filename)) {
3783 ret = PTR_ERR(req->open.filename);
3784 req->open.filename = NULL;
3787 req->open.nofile = rlimit(RLIMIT_NOFILE);
3788 req->flags |= REQ_F_NEED_CLEANUP;
3792 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3796 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3798 mode = READ_ONCE(sqe->len);
3799 flags = READ_ONCE(sqe->open_flags);
3800 req->open.how = build_open_how(flags, mode);
3801 return __io_openat_prep(req, sqe);
3804 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3806 struct open_how __user *how;
3810 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3812 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3813 len = READ_ONCE(sqe->len);
3814 if (len < OPEN_HOW_SIZE_VER0)
3817 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3822 return __io_openat_prep(req, sqe);
3825 static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
3827 struct open_flags op;
3830 bool resolve_nonblock;
3833 ret = build_open_flags(&req->open.how, &op);
3836 nonblock_set = op.open_flag & O_NONBLOCK;
3837 resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
3838 if (issue_flags & IO_URING_F_NONBLOCK) {
3840 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
3841 * it'll always -EAGAIN
3843 if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
3845 op.lookup_flags |= LOOKUP_CACHED;
3846 op.open_flag |= O_NONBLOCK;
3849 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3853 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3854 /* only retry if RESOLVE_CACHED wasn't already set by application */
3855 if ((!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)) &&
3856 file == ERR_PTR(-EAGAIN)) {
3858 * We could hang on to this 'fd', but seems like marginal
3859 * gain for something that is now known to be a slower path.
3860 * So just put it, and we'll get a new one when we retry.
3868 ret = PTR_ERR(file);
3870 if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
3871 file->f_flags &= ~O_NONBLOCK;
3872 fsnotify_open(file);
3873 fd_install(ret, file);
3876 putname(req->open.filename);
3877 req->flags &= ~REQ_F_NEED_CLEANUP;
3880 __io_req_complete(req, issue_flags, ret, 0);
3884 static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
3886 return io_openat2(req, issue_flags);
3889 static int io_remove_buffers_prep(struct io_kiocb *req,
3890 const struct io_uring_sqe *sqe)
3892 struct io_provide_buf *p = &req->pbuf;
3895 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3898 tmp = READ_ONCE(sqe->fd);
3899 if (!tmp || tmp > USHRT_MAX)
3902 memset(p, 0, sizeof(*p));
3904 p->bgid = READ_ONCE(sqe->buf_group);
3908 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3909 int bgid, unsigned nbufs)
3913 /* shouldn't happen */
3917 /* the head kbuf is the list itself */
3918 while (!list_empty(&buf->list)) {
3919 struct io_buffer *nxt;
3921 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3922 list_del(&nxt->list);
3929 xa_erase(&ctx->io_buffers, bgid);
3934 static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
3936 struct io_provide_buf *p = &req->pbuf;
3937 struct io_ring_ctx *ctx = req->ctx;
3938 struct io_buffer *head;
3940 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3942 io_ring_submit_lock(ctx, !force_nonblock);
3944 lockdep_assert_held(&ctx->uring_lock);
3947 head = xa_load(&ctx->io_buffers, p->bgid);
3949 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3953 /* complete before unlock, IOPOLL may need the lock */
3954 __io_req_complete(req, issue_flags, ret, 0);
3955 io_ring_submit_unlock(ctx, !force_nonblock);
3959 static int io_provide_buffers_prep(struct io_kiocb *req,
3960 const struct io_uring_sqe *sqe)
3962 unsigned long size, tmp_check;
3963 struct io_provide_buf *p = &req->pbuf;
3966 if (sqe->ioprio || sqe->rw_flags)
3969 tmp = READ_ONCE(sqe->fd);
3970 if (!tmp || tmp > USHRT_MAX)
3973 p->addr = READ_ONCE(sqe->addr);
3974 p->len = READ_ONCE(sqe->len);
3976 if (check_mul_overflow((unsigned long)p->len, (unsigned long)p->nbufs,
3979 if (check_add_overflow((unsigned long)p->addr, size, &tmp_check))
3982 size = (unsigned long)p->len * p->nbufs;
3983 if (!access_ok(u64_to_user_ptr(p->addr), size))
3986 p->bgid = READ_ONCE(sqe->buf_group);
3987 tmp = READ_ONCE(sqe->off);
3988 if (tmp > USHRT_MAX)
3994 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3996 struct io_buffer *buf;
3997 u64 addr = pbuf->addr;
3998 int i, bid = pbuf->bid;
4000 for (i = 0; i < pbuf->nbufs; i++) {
4001 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
4006 buf->len = min_t(__u32, pbuf->len, MAX_RW_COUNT);
4011 INIT_LIST_HEAD(&buf->list);
4014 list_add_tail(&buf->list, &(*head)->list);
4018 return i ? i : -ENOMEM;
4021 static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
4023 struct io_provide_buf *p = &req->pbuf;
4024 struct io_ring_ctx *ctx = req->ctx;
4025 struct io_buffer *head, *list;
4027 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4029 io_ring_submit_lock(ctx, !force_nonblock);
4031 lockdep_assert_held(&ctx->uring_lock);
4033 list = head = xa_load(&ctx->io_buffers, p->bgid);
4035 ret = io_add_buffers(p, &head);
4036 if (ret >= 0 && !list) {
4037 ret = xa_insert(&ctx->io_buffers, p->bgid, head, GFP_KERNEL);
4039 __io_remove_buffers(ctx, head, p->bgid, -1U);
4043 /* complete before unlock, IOPOLL may need the lock */
4044 __io_req_complete(req, issue_flags, ret, 0);
4045 io_ring_submit_unlock(ctx, !force_nonblock);
4049 static int io_epoll_ctl_prep(struct io_kiocb *req,
4050 const struct io_uring_sqe *sqe)
4052 #if defined(CONFIG_EPOLL)
4053 if (sqe->ioprio || sqe->buf_index)
4055 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4058 req->epoll.epfd = READ_ONCE(sqe->fd);
4059 req->epoll.op = READ_ONCE(sqe->len);
4060 req->epoll.fd = READ_ONCE(sqe->off);
4062 if (ep_op_has_event(req->epoll.op)) {
4063 struct epoll_event __user *ev;
4065 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4066 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4076 static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4078 #if defined(CONFIG_EPOLL)
4079 struct io_epoll *ie = &req->epoll;
4081 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4083 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4084 if (force_nonblock && ret == -EAGAIN)
4089 __io_req_complete(req, issue_flags, ret, 0);
4096 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4098 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4099 if (sqe->ioprio || sqe->buf_index || sqe->off)
4101 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4104 req->madvise.addr = READ_ONCE(sqe->addr);
4105 req->madvise.len = READ_ONCE(sqe->len);
4106 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4113 static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4115 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4116 struct io_madvise *ma = &req->madvise;
4119 if (issue_flags & IO_URING_F_NONBLOCK)
4122 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4125 io_req_complete(req, ret);
4132 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4134 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4136 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4139 req->fadvise.offset = READ_ONCE(sqe->off);
4140 req->fadvise.len = READ_ONCE(sqe->len);
4141 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4145 static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4147 struct io_fadvise *fa = &req->fadvise;
4150 if (issue_flags & IO_URING_F_NONBLOCK) {
4151 switch (fa->advice) {
4152 case POSIX_FADV_NORMAL:
4153 case POSIX_FADV_RANDOM:
4154 case POSIX_FADV_SEQUENTIAL:
4161 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4164 __io_req_complete(req, issue_flags, ret, 0);
4168 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4170 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4172 if (sqe->ioprio || sqe->buf_index)
4174 if (req->flags & REQ_F_FIXED_FILE)
4177 req->statx.dfd = READ_ONCE(sqe->fd);
4178 req->statx.mask = READ_ONCE(sqe->len);
4179 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4180 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4181 req->statx.flags = READ_ONCE(sqe->statx_flags);
4186 static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4188 struct io_statx *ctx = &req->statx;
4191 if (issue_flags & IO_URING_F_NONBLOCK)
4194 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4199 io_req_complete(req, ret);
4203 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4205 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4207 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4208 sqe->rw_flags || sqe->buf_index)
4210 if (req->flags & REQ_F_FIXED_FILE)
4213 req->close.fd = READ_ONCE(sqe->fd);
4217 static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4219 struct files_struct *files = current->files;
4220 struct io_close *close = &req->close;
4221 struct fdtable *fdt;
4222 struct file *file = NULL;
4225 spin_lock(&files->file_lock);
4226 fdt = files_fdtable(files);
4227 if (close->fd >= fdt->max_fds) {
4228 spin_unlock(&files->file_lock);
4231 file = fdt->fd[close->fd];
4232 if (!file || file->f_op == &io_uring_fops) {
4233 spin_unlock(&files->file_lock);
4238 /* if the file has a flush method, be safe and punt to async */
4239 if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4240 spin_unlock(&files->file_lock);
4244 ret = __close_fd_get_file(close->fd, &file);
4245 spin_unlock(&files->file_lock);
4252 /* No ->flush() or already async, safely close from here */
4253 ret = filp_close(file, current->files);
4259 __io_req_complete(req, issue_flags, ret, 0);
4263 static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4265 struct io_ring_ctx *ctx = req->ctx;
4267 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4269 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4272 req->sync.off = READ_ONCE(sqe->off);
4273 req->sync.len = READ_ONCE(sqe->len);
4274 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4278 static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4282 /* sync_file_range always requires a blocking context */
4283 if (issue_flags & IO_URING_F_NONBLOCK)
4286 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4290 io_req_complete(req, ret);
4294 #if defined(CONFIG_NET)
4295 static int io_setup_async_msg(struct io_kiocb *req,
4296 struct io_async_msghdr *kmsg)
4298 struct io_async_msghdr *async_msg = req->async_data;
4302 if (io_alloc_async_data(req)) {
4303 kfree(kmsg->free_iov);
4306 async_msg = req->async_data;
4307 req->flags |= REQ_F_NEED_CLEANUP;
4308 memcpy(async_msg, kmsg, sizeof(*kmsg));
4309 async_msg->msg.msg_name = &async_msg->addr;
4310 /* if were using fast_iov, set it to the new one */
4311 if (!async_msg->free_iov)
4312 async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4317 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4318 struct io_async_msghdr *iomsg)
4320 iomsg->msg.msg_name = &iomsg->addr;
4321 iomsg->free_iov = iomsg->fast_iov;
4322 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4323 req->sr_msg.msg_flags, &iomsg->free_iov);
4326 static int io_sendmsg_prep_async(struct io_kiocb *req)
4330 ret = io_sendmsg_copy_hdr(req, req->async_data);
4332 req->flags |= REQ_F_NEED_CLEANUP;
4336 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4338 struct io_sr_msg *sr = &req->sr_msg;
4340 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4343 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4344 sr->len = READ_ONCE(sqe->len);
4345 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4346 if (sr->msg_flags & MSG_DONTWAIT)
4347 req->flags |= REQ_F_NOWAIT;
4349 #ifdef CONFIG_COMPAT
4350 if (req->ctx->compat)
4351 sr->msg_flags |= MSG_CMSG_COMPAT;
4356 static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4358 struct io_async_msghdr iomsg, *kmsg;
4359 struct socket *sock;
4364 sock = sock_from_file(req->file);
4365 if (unlikely(!sock))
4368 kmsg = req->async_data;
4370 ret = io_sendmsg_copy_hdr(req, &iomsg);
4376 flags = req->sr_msg.msg_flags;
4377 if (issue_flags & IO_URING_F_NONBLOCK)
4378 flags |= MSG_DONTWAIT;
4379 if (flags & MSG_WAITALL)
4380 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4382 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4383 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4384 return io_setup_async_msg(req, kmsg);
4385 if (ret == -ERESTARTSYS)
4388 /* fast path, check for non-NULL to avoid function call */
4390 kfree(kmsg->free_iov);
4391 req->flags &= ~REQ_F_NEED_CLEANUP;
4394 __io_req_complete(req, issue_flags, ret, 0);
4398 static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4400 struct io_sr_msg *sr = &req->sr_msg;
4403 struct socket *sock;
4408 sock = sock_from_file(req->file);
4409 if (unlikely(!sock))
4412 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4416 msg.msg_name = NULL;
4417 msg.msg_control = NULL;
4418 msg.msg_controllen = 0;
4419 msg.msg_namelen = 0;
4421 flags = req->sr_msg.msg_flags;
4422 if (issue_flags & IO_URING_F_NONBLOCK)
4423 flags |= MSG_DONTWAIT;
4424 if (flags & MSG_WAITALL)
4425 min_ret = iov_iter_count(&msg.msg_iter);
4427 msg.msg_flags = flags;
4428 ret = sock_sendmsg(sock, &msg);
4429 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4431 if (ret == -ERESTARTSYS)
4436 __io_req_complete(req, issue_flags, ret, 0);
4440 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4441 struct io_async_msghdr *iomsg)
4443 struct io_sr_msg *sr = &req->sr_msg;
4444 struct iovec __user *uiov;
4448 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4449 &iomsg->uaddr, &uiov, &iov_len);
4453 if (req->flags & REQ_F_BUFFER_SELECT) {
4456 if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4458 sr->len = iomsg->fast_iov[0].iov_len;
4459 iomsg->free_iov = NULL;
4461 iomsg->free_iov = iomsg->fast_iov;
4462 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4463 &iomsg->free_iov, &iomsg->msg.msg_iter,
4472 #ifdef CONFIG_COMPAT
4473 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4474 struct io_async_msghdr *iomsg)
4476 struct io_sr_msg *sr = &req->sr_msg;
4477 struct compat_iovec __user *uiov;
4482 ret = __get_compat_msghdr(&iomsg->msg, sr->umsg_compat, &iomsg->uaddr,
4487 uiov = compat_ptr(ptr);
4488 if (req->flags & REQ_F_BUFFER_SELECT) {
4489 compat_ssize_t clen;
4493 if (!access_ok(uiov, sizeof(*uiov)))
4495 if (__get_user(clen, &uiov->iov_len))
4500 iomsg->free_iov = NULL;
4502 iomsg->free_iov = iomsg->fast_iov;
4503 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4504 UIO_FASTIOV, &iomsg->free_iov,
4505 &iomsg->msg.msg_iter, true);
4514 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4515 struct io_async_msghdr *iomsg)
4517 iomsg->msg.msg_name = &iomsg->addr;
4519 #ifdef CONFIG_COMPAT
4520 if (req->ctx->compat)
4521 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4524 return __io_recvmsg_copy_hdr(req, iomsg);
4527 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4530 struct io_sr_msg *sr = &req->sr_msg;
4531 struct io_buffer *kbuf;
4533 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4538 req->flags |= REQ_F_BUFFER_SELECTED;
4542 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4544 return io_put_kbuf(req, req->sr_msg.kbuf);
4547 static int io_recvmsg_prep_async(struct io_kiocb *req)
4551 ret = io_recvmsg_copy_hdr(req, req->async_data);
4553 req->flags |= REQ_F_NEED_CLEANUP;
4557 static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4559 struct io_sr_msg *sr = &req->sr_msg;
4561 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4564 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4565 sr->len = READ_ONCE(sqe->len);
4566 sr->bgid = READ_ONCE(sqe->buf_group);
4567 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4568 if (sr->msg_flags & MSG_DONTWAIT)
4569 req->flags |= REQ_F_NOWAIT;
4571 #ifdef CONFIG_COMPAT
4572 if (req->ctx->compat)
4573 sr->msg_flags |= MSG_CMSG_COMPAT;
4578 static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
4580 struct io_async_msghdr iomsg, *kmsg;
4581 struct socket *sock;
4582 struct io_buffer *kbuf;
4585 int ret, cflags = 0;
4586 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4588 sock = sock_from_file(req->file);
4589 if (unlikely(!sock))
4592 kmsg = req->async_data;
4594 ret = io_recvmsg_copy_hdr(req, &iomsg);
4600 if (req->flags & REQ_F_BUFFER_SELECT) {
4601 kbuf = io_recv_buffer_select(req, !force_nonblock);
4603 return PTR_ERR(kbuf);
4604 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4605 kmsg->fast_iov[0].iov_len = req->sr_msg.len;
4606 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
4607 1, req->sr_msg.len);
4610 flags = req->sr_msg.msg_flags;
4612 flags |= MSG_DONTWAIT;
4613 if (flags & MSG_WAITALL)
4614 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4616 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4617 kmsg->uaddr, flags);
4618 if (force_nonblock && ret == -EAGAIN)
4619 return io_setup_async_msg(req, kmsg);
4620 if (ret == -ERESTARTSYS)
4623 if (req->flags & REQ_F_BUFFER_SELECTED)
4624 cflags = io_put_recv_kbuf(req);
4625 /* fast path, check for non-NULL to avoid function call */
4627 kfree(kmsg->free_iov);
4628 req->flags &= ~REQ_F_NEED_CLEANUP;
4629 if (ret < min_ret || ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4631 __io_req_complete(req, issue_flags, ret, cflags);
4635 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
4637 struct io_buffer *kbuf;
4638 struct io_sr_msg *sr = &req->sr_msg;
4640 void __user *buf = sr->buf;
4641 struct socket *sock;
4645 int ret, cflags = 0;
4646 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4648 sock = sock_from_file(req->file);
4649 if (unlikely(!sock))
4652 if (req->flags & REQ_F_BUFFER_SELECT) {
4653 kbuf = io_recv_buffer_select(req, !force_nonblock);
4655 return PTR_ERR(kbuf);
4656 buf = u64_to_user_ptr(kbuf->addr);
4659 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4663 msg.msg_name = NULL;
4664 msg.msg_control = NULL;
4665 msg.msg_controllen = 0;
4666 msg.msg_namelen = 0;
4667 msg.msg_iocb = NULL;
4670 flags = req->sr_msg.msg_flags;
4672 flags |= MSG_DONTWAIT;
4673 if (flags & MSG_WAITALL)
4674 min_ret = iov_iter_count(&msg.msg_iter);
4676 ret = sock_recvmsg(sock, &msg, flags);
4677 if (force_nonblock && ret == -EAGAIN)
4679 if (ret == -ERESTARTSYS)
4682 if (req->flags & REQ_F_BUFFER_SELECTED)
4683 cflags = io_put_recv_kbuf(req);
4684 if (ret < min_ret || ((flags & MSG_WAITALL) && (msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4686 __io_req_complete(req, issue_flags, ret, cflags);
4690 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4692 struct io_accept *accept = &req->accept;
4694 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4696 if (sqe->ioprio || sqe->len || sqe->buf_index)
4699 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4700 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4701 accept->flags = READ_ONCE(sqe->accept_flags);
4702 accept->nofile = rlimit(RLIMIT_NOFILE);
4706 static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
4708 struct io_accept *accept = &req->accept;
4709 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4710 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4713 if (req->file->f_flags & O_NONBLOCK)
4714 req->flags |= REQ_F_NOWAIT;
4716 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4717 accept->addr_len, accept->flags,
4719 if (ret == -EAGAIN && force_nonblock)
4722 if (ret == -ERESTARTSYS)
4726 __io_req_complete(req, issue_flags, ret, 0);
4730 static int io_connect_prep_async(struct io_kiocb *req)
4732 struct io_async_connect *io = req->async_data;
4733 struct io_connect *conn = &req->connect;
4735 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
4738 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4740 struct io_connect *conn = &req->connect;
4742 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4744 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4747 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4748 conn->addr_len = READ_ONCE(sqe->addr2);
4752 static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
4754 struct io_async_connect __io, *io;
4755 unsigned file_flags;
4757 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4759 if (req->async_data) {
4760 io = req->async_data;
4762 ret = move_addr_to_kernel(req->connect.addr,
4763 req->connect.addr_len,
4770 file_flags = force_nonblock ? O_NONBLOCK : 0;
4772 ret = __sys_connect_file(req->file, &io->address,
4773 req->connect.addr_len, file_flags);
4774 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4775 if (req->async_data)
4777 if (io_alloc_async_data(req)) {
4781 memcpy(req->async_data, &__io, sizeof(__io));
4784 if (ret == -ERESTARTSYS)
4789 __io_req_complete(req, issue_flags, ret, 0);
4792 #else /* !CONFIG_NET */
4793 #define IO_NETOP_FN(op) \
4794 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
4796 return -EOPNOTSUPP; \
4799 #define IO_NETOP_PREP(op) \
4801 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
4803 return -EOPNOTSUPP; \
4806 #define IO_NETOP_PREP_ASYNC(op) \
4808 static int io_##op##_prep_async(struct io_kiocb *req) \
4810 return -EOPNOTSUPP; \
4813 IO_NETOP_PREP_ASYNC(sendmsg);
4814 IO_NETOP_PREP_ASYNC(recvmsg);
4815 IO_NETOP_PREP_ASYNC(connect);
4816 IO_NETOP_PREP(accept);
4819 #endif /* CONFIG_NET */
4821 struct io_poll_table {
4822 struct poll_table_struct pt;
4823 struct io_kiocb *req;
4827 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4828 __poll_t mask, task_work_func_t func)
4832 /* for instances that support it check for an event match first: */
4833 if (mask && !(mask & poll->events))
4836 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4838 list_del_init(&poll->wait.entry);
4841 req->task_work.func = func;
4844 * If this fails, then the task is exiting. When a task exits, the
4845 * work gets canceled, so just cancel this request as well instead
4846 * of executing it. We can't safely execute it anyway, as we may not
4847 * have the needed state needed for it anyway.
4849 ret = io_req_task_work_add(req);
4850 if (unlikely(ret)) {
4851 WRITE_ONCE(poll->canceled, true);
4852 io_req_task_work_add_fallback(req, func);
4857 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4858 __acquires(&req->ctx->completion_lock)
4860 struct io_ring_ctx *ctx = req->ctx;
4862 if (!req->result && !READ_ONCE(poll->canceled)) {
4863 struct poll_table_struct pt = { ._key = poll->events };
4865 req->result = vfs_poll(req->file, &pt) & poll->events;
4868 spin_lock_irq(&ctx->completion_lock);
4869 if (!req->result && !READ_ONCE(poll->canceled)) {
4870 add_wait_queue(poll->head, &poll->wait);
4877 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4879 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4880 if (req->opcode == IORING_OP_POLL_ADD)
4881 return req->async_data;
4882 return req->apoll->double_poll;
4885 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4887 if (req->opcode == IORING_OP_POLL_ADD)
4889 return &req->apoll->poll;
4892 static void io_poll_remove_double(struct io_kiocb *req)
4893 __must_hold(&req->ctx->completion_lock)
4895 struct io_poll_iocb *poll = io_poll_get_double(req);
4897 lockdep_assert_held(&req->ctx->completion_lock);
4899 if (poll && poll->head) {
4900 struct wait_queue_head *head = poll->head;
4902 spin_lock(&head->lock);
4903 list_del_init(&poll->wait.entry);
4904 if (poll->wait.private)
4907 spin_unlock(&head->lock);
4911 static bool io_poll_complete(struct io_kiocb *req, __poll_t mask)
4912 __must_hold(&req->ctx->completion_lock)
4914 struct io_ring_ctx *ctx = req->ctx;
4915 unsigned flags = IORING_CQE_F_MORE;
4918 if (READ_ONCE(req->poll.canceled)) {
4920 req->poll.events |= EPOLLONESHOT;
4922 error = mangle_poll(mask);
4924 if (req->poll.events & EPOLLONESHOT)
4926 if (!io_cqring_fill_event(ctx, req->user_data, error, flags)) {
4927 io_poll_remove_waitqs(req);
4928 req->poll.done = true;
4931 if (flags & IORING_CQE_F_MORE)
4934 io_commit_cqring(ctx);
4935 return !(flags & IORING_CQE_F_MORE);
4938 static void io_poll_task_func(struct callback_head *cb)
4940 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4941 struct io_ring_ctx *ctx = req->ctx;
4942 struct io_kiocb *nxt;
4944 if (io_poll_rewait(req, &req->poll)) {
4945 spin_unlock_irq(&ctx->completion_lock);
4949 done = io_poll_complete(req, req->result);
4951 hash_del(&req->hash_node);
4954 add_wait_queue(req->poll.head, &req->poll.wait);
4956 spin_unlock_irq(&ctx->completion_lock);
4957 io_cqring_ev_posted(ctx);
4960 nxt = io_put_req_find_next(req);
4962 __io_req_task_submit(nxt);
4967 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4968 int sync, void *key)
4970 struct io_kiocb *req = wait->private;
4971 struct io_poll_iocb *poll = io_poll_get_single(req);
4972 __poll_t mask = key_to_poll(key);
4974 /* for instances that support it check for an event match first: */
4975 if (mask && !(mask & poll->events))
4977 if (!(poll->events & EPOLLONESHOT))
4978 return poll->wait.func(&poll->wait, mode, sync, key);
4980 list_del_init(&wait->entry);
4982 if (poll && poll->head) {
4985 spin_lock(&poll->head->lock);
4986 done = list_empty(&poll->wait.entry);
4988 list_del_init(&poll->wait.entry);
4989 /* make sure double remove sees this as being gone */
4990 wait->private = NULL;
4991 spin_unlock(&poll->head->lock);
4993 /* use wait func handler, so it matches the rq type */
4994 poll->wait.func(&poll->wait, mode, sync, key);
5001 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
5002 wait_queue_func_t wake_func)
5006 poll->canceled = false;
5007 #define IO_POLL_UNMASK (EPOLLERR|EPOLLHUP|EPOLLNVAL|EPOLLRDHUP)
5008 /* mask in events that we always want/need */
5009 poll->events = events | IO_POLL_UNMASK;
5010 INIT_LIST_HEAD(&poll->wait.entry);
5011 init_waitqueue_func_entry(&poll->wait, wake_func);
5014 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5015 struct wait_queue_head *head,
5016 struct io_poll_iocb **poll_ptr)
5018 struct io_kiocb *req = pt->req;
5021 * If poll->head is already set, it's because the file being polled
5022 * uses multiple waitqueues for poll handling (eg one for read, one
5023 * for write). Setup a separate io_poll_iocb if this happens.
5025 if (unlikely(poll->head)) {
5026 struct io_poll_iocb *poll_one = poll;
5028 /* already have a 2nd entry, fail a third attempt */
5030 pt->error = -EINVAL;
5034 * Can't handle multishot for double wait for now, turn it
5035 * into one-shot mode.
5037 if (!(poll_one->events & EPOLLONESHOT))
5038 poll_one->events |= EPOLLONESHOT;
5039 /* double add on the same waitqueue head, ignore */
5040 if (poll_one->head == head)
5042 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5044 pt->error = -ENOMEM;
5047 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5049 poll->wait.private = req;
5056 if (poll->events & EPOLLEXCLUSIVE)
5057 add_wait_queue_exclusive(head, &poll->wait);
5059 add_wait_queue(head, &poll->wait);
5062 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5063 struct poll_table_struct *p)
5065 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5066 struct async_poll *apoll = pt->req->apoll;
5068 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5071 static void io_async_task_func(struct callback_head *cb)
5073 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
5074 struct async_poll *apoll = req->apoll;
5075 struct io_ring_ctx *ctx = req->ctx;
5077 trace_io_uring_task_run(req->ctx, req, req->opcode, req->user_data);
5079 if (io_poll_rewait(req, &apoll->poll)) {
5080 spin_unlock_irq(&ctx->completion_lock);
5084 hash_del(&req->hash_node);
5085 io_poll_remove_double(req);
5086 spin_unlock_irq(&ctx->completion_lock);
5088 if (!READ_ONCE(apoll->poll.canceled))
5089 __io_req_task_submit(req);
5091 io_req_complete_failed(req, -ECANCELED);
5094 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5097 struct io_kiocb *req = wait->private;
5098 struct io_poll_iocb *poll = &req->apoll->poll;
5100 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5103 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5106 static void io_poll_req_insert(struct io_kiocb *req)
5108 struct io_ring_ctx *ctx = req->ctx;
5109 struct hlist_head *list;
5111 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5112 hlist_add_head(&req->hash_node, list);
5115 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5116 struct io_poll_iocb *poll,
5117 struct io_poll_table *ipt, __poll_t mask,
5118 wait_queue_func_t wake_func)
5119 __acquires(&ctx->completion_lock)
5121 struct io_ring_ctx *ctx = req->ctx;
5122 bool cancel = false;
5124 INIT_HLIST_NODE(&req->hash_node);
5125 io_init_poll_iocb(poll, mask, wake_func);
5126 poll->file = req->file;
5127 poll->wait.private = req;
5129 ipt->pt._key = mask;
5131 ipt->error = -EINVAL;
5133 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5135 spin_lock_irq(&ctx->completion_lock);
5136 if (likely(poll->head)) {
5137 spin_lock(&poll->head->lock);
5138 if (unlikely(list_empty(&poll->wait.entry))) {
5144 if ((mask && (poll->events & EPOLLONESHOT)) || ipt->error)
5145 list_del_init(&poll->wait.entry);
5147 WRITE_ONCE(poll->canceled, true);
5148 else if (!poll->done) /* actually waiting for an event */
5149 io_poll_req_insert(req);
5150 spin_unlock(&poll->head->lock);
5156 static bool io_arm_poll_handler(struct io_kiocb *req)
5158 const struct io_op_def *def = &io_op_defs[req->opcode];
5159 struct io_ring_ctx *ctx = req->ctx;
5160 struct async_poll *apoll;
5161 struct io_poll_table ipt;
5165 if (!req->file || !file_can_poll(req->file))
5167 if (req->flags & REQ_F_POLLED)
5171 else if (def->pollout)
5175 /* if we can't nonblock try, then no point in arming a poll handler */
5176 if (!io_file_supports_async(req, rw))
5179 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5180 if (unlikely(!apoll))
5182 apoll->double_poll = NULL;
5184 req->flags |= REQ_F_POLLED;
5187 mask = EPOLLONESHOT;
5189 mask |= POLLIN | POLLRDNORM;
5191 mask |= POLLOUT | POLLWRNORM;
5193 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5194 if ((req->opcode == IORING_OP_RECVMSG) &&
5195 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5198 mask |= POLLERR | POLLPRI;
5200 ipt.pt._qproc = io_async_queue_proc;
5202 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5204 if (ret || ipt.error) {
5205 io_poll_remove_double(req);
5206 spin_unlock_irq(&ctx->completion_lock);
5209 spin_unlock_irq(&ctx->completion_lock);
5210 trace_io_uring_poll_arm(ctx, req, req->opcode, req->user_data,
5211 mask, apoll->poll.events);
5215 static bool __io_poll_remove_one(struct io_kiocb *req,
5216 struct io_poll_iocb *poll, bool do_cancel)
5217 __must_hold(&req->ctx->completion_lock)
5219 bool do_complete = false;
5223 spin_lock(&poll->head->lock);
5225 WRITE_ONCE(poll->canceled, true);
5226 if (!list_empty(&poll->wait.entry)) {
5227 list_del_init(&poll->wait.entry);
5230 spin_unlock(&poll->head->lock);
5231 hash_del(&req->hash_node);
5235 static bool io_poll_remove_waitqs(struct io_kiocb *req)
5236 __must_hold(&req->ctx->completion_lock)
5240 io_poll_remove_double(req);
5241 do_complete = __io_poll_remove_one(req, io_poll_get_single(req), true);
5243 if (req->opcode != IORING_OP_POLL_ADD && do_complete) {
5244 /* non-poll requests have submit ref still */
5250 static bool io_poll_remove_one(struct io_kiocb *req)
5251 __must_hold(&req->ctx->completion_lock)
5255 do_complete = io_poll_remove_waitqs(req);
5257 io_cqring_fill_event(req->ctx, req->user_data, -ECANCELED, 0);
5258 io_commit_cqring(req->ctx);
5260 io_put_req_deferred(req, 1);
5267 * Returns true if we found and killed one or more poll requests
5269 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5272 struct hlist_node *tmp;
5273 struct io_kiocb *req;
5276 spin_lock_irq(&ctx->completion_lock);
5277 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5278 struct hlist_head *list;
5280 list = &ctx->cancel_hash[i];
5281 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5282 if (io_match_task(req, tsk, cancel_all))
5283 posted += io_poll_remove_one(req);
5286 spin_unlock_irq(&ctx->completion_lock);
5289 io_cqring_ev_posted(ctx);
5294 static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, __u64 sqe_addr,
5296 __must_hold(&ctx->completion_lock)
5298 struct hlist_head *list;
5299 struct io_kiocb *req;
5301 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5302 hlist_for_each_entry(req, list, hash_node) {
5303 if (sqe_addr != req->user_data)
5305 if (poll_only && req->opcode != IORING_OP_POLL_ADD)
5312 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr,
5314 __must_hold(&ctx->completion_lock)
5316 struct io_kiocb *req;
5318 req = io_poll_find(ctx, sqe_addr, poll_only);
5321 if (io_poll_remove_one(req))
5327 static __poll_t io_poll_parse_events(const struct io_uring_sqe *sqe,
5332 events = READ_ONCE(sqe->poll32_events);
5334 events = swahw32(events);
5336 if (!(flags & IORING_POLL_ADD_MULTI))
5337 events |= EPOLLONESHOT;
5338 return demangle_poll(events) | (events & (EPOLLEXCLUSIVE|EPOLLONESHOT));
5341 static int io_poll_update_prep(struct io_kiocb *req,
5342 const struct io_uring_sqe *sqe)
5344 struct io_poll_update *upd = &req->poll_update;
5347 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5349 if (sqe->ioprio || sqe->buf_index)
5351 flags = READ_ONCE(sqe->len);
5352 if (flags & ~(IORING_POLL_UPDATE_EVENTS | IORING_POLL_UPDATE_USER_DATA |
5353 IORING_POLL_ADD_MULTI))
5355 /* meaningless without update */
5356 if (flags == IORING_POLL_ADD_MULTI)
5359 upd->old_user_data = READ_ONCE(sqe->addr);
5360 upd->update_events = flags & IORING_POLL_UPDATE_EVENTS;
5361 upd->update_user_data = flags & IORING_POLL_UPDATE_USER_DATA;
5363 upd->new_user_data = READ_ONCE(sqe->off);
5364 if (!upd->update_user_data && upd->new_user_data)
5366 if (upd->update_events)
5367 upd->events = io_poll_parse_events(sqe, flags);
5368 else if (sqe->poll32_events)
5374 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5377 struct io_kiocb *req = wait->private;
5378 struct io_poll_iocb *poll = &req->poll;
5380 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5383 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5384 struct poll_table_struct *p)
5386 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5388 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5391 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5393 struct io_poll_iocb *poll = &req->poll;
5396 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5398 if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->addr)
5400 flags = READ_ONCE(sqe->len);
5401 if (flags & ~IORING_POLL_ADD_MULTI)
5404 poll->events = io_poll_parse_events(sqe, flags);
5408 static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5410 struct io_poll_iocb *poll = &req->poll;
5411 struct io_ring_ctx *ctx = req->ctx;
5412 struct io_poll_table ipt;
5415 ipt.pt._qproc = io_poll_queue_proc;
5417 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5420 if (mask) { /* no async, we'd stolen it */
5422 io_poll_complete(req, mask);
5424 spin_unlock_irq(&ctx->completion_lock);
5427 io_cqring_ev_posted(ctx);
5428 if (poll->events & EPOLLONESHOT)
5434 static int io_poll_update(struct io_kiocb *req, unsigned int issue_flags)
5436 struct io_ring_ctx *ctx = req->ctx;
5437 struct io_kiocb *preq;
5441 spin_lock_irq(&ctx->completion_lock);
5442 preq = io_poll_find(ctx, req->poll_update.old_user_data, true);
5448 if (!req->poll_update.update_events && !req->poll_update.update_user_data) {
5450 ret = io_poll_remove_one(preq) ? 0 : -EALREADY;
5455 * Don't allow racy completion with singleshot, as we cannot safely
5456 * update those. For multishot, if we're racing with completion, just
5457 * let completion re-add it.
5459 completing = !__io_poll_remove_one(preq, &preq->poll, false);
5460 if (completing && (preq->poll.events & EPOLLONESHOT)) {
5464 /* we now have a detached poll request. reissue. */
5468 spin_unlock_irq(&ctx->completion_lock);
5470 io_req_complete(req, ret);
5473 /* only mask one event flags, keep behavior flags */
5474 if (req->poll_update.update_events) {
5475 preq->poll.events &= ~0xffff;
5476 preq->poll.events |= req->poll_update.events & 0xffff;
5477 preq->poll.events |= IO_POLL_UNMASK;
5479 if (req->poll_update.update_user_data)
5480 preq->user_data = req->poll_update.new_user_data;
5481 spin_unlock_irq(&ctx->completion_lock);
5483 /* complete update request, we're done with it */
5484 io_req_complete(req, ret);
5487 ret = io_poll_add(preq, issue_flags);
5490 io_req_complete(preq, ret);
5496 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5498 struct io_timeout_data *data = container_of(timer,
5499 struct io_timeout_data, timer);
5500 struct io_kiocb *req = data->req;
5501 struct io_ring_ctx *ctx = req->ctx;
5502 unsigned long flags;
5504 spin_lock_irqsave(&ctx->completion_lock, flags);
5505 list_del_init(&req->timeout.list);
5506 atomic_set(&req->ctx->cq_timeouts,
5507 atomic_read(&req->ctx->cq_timeouts) + 1);
5509 io_cqring_fill_event(ctx, req->user_data, -ETIME, 0);
5510 io_commit_cqring(ctx);
5511 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5513 io_cqring_ev_posted(ctx);
5516 return HRTIMER_NORESTART;
5519 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5521 __must_hold(&ctx->completion_lock)
5523 struct io_timeout_data *io;
5524 struct io_kiocb *req;
5527 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5528 found = user_data == req->user_data;
5533 return ERR_PTR(-ENOENT);
5535 io = req->async_data;
5536 if (hrtimer_try_to_cancel(&io->timer) == -1)
5537 return ERR_PTR(-EALREADY);
5538 list_del_init(&req->timeout.list);
5542 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5543 __must_hold(&ctx->completion_lock)
5545 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5548 return PTR_ERR(req);
5551 io_cqring_fill_event(ctx, req->user_data, -ECANCELED, 0);
5552 io_put_req_deferred(req, 1);
5556 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5557 struct timespec64 *ts, enum hrtimer_mode mode)
5558 __must_hold(&ctx->completion_lock)
5560 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5561 struct io_timeout_data *data;
5564 return PTR_ERR(req);
5566 req->timeout.off = 0; /* noseq */
5567 data = req->async_data;
5568 list_add_tail(&req->timeout.list, &ctx->timeout_list);
5569 hrtimer_init(&data->timer, CLOCK_MONOTONIC, mode);
5570 data->timer.function = io_timeout_fn;
5571 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
5575 static int io_timeout_remove_prep(struct io_kiocb *req,
5576 const struct io_uring_sqe *sqe)
5578 struct io_timeout_rem *tr = &req->timeout_rem;
5580 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5582 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5584 if (sqe->ioprio || sqe->buf_index || sqe->len)
5587 tr->addr = READ_ONCE(sqe->addr);
5588 tr->flags = READ_ONCE(sqe->timeout_flags);
5589 if (tr->flags & IORING_TIMEOUT_UPDATE) {
5590 if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS))
5592 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
5594 } else if (tr->flags) {
5595 /* timeout removal doesn't support flags */
5602 static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
5604 return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
5609 * Remove or update an existing timeout command
5611 static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
5613 struct io_timeout_rem *tr = &req->timeout_rem;
5614 struct io_ring_ctx *ctx = req->ctx;
5617 spin_lock_irq(&ctx->completion_lock);
5618 if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE))
5619 ret = io_timeout_cancel(ctx, tr->addr);
5621 ret = io_timeout_update(ctx, tr->addr, &tr->ts,
5622 io_translate_timeout_mode(tr->flags));
5624 io_cqring_fill_event(ctx, req->user_data, ret, 0);
5625 io_commit_cqring(ctx);
5626 spin_unlock_irq(&ctx->completion_lock);
5627 io_cqring_ev_posted(ctx);
5634 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5635 bool is_timeout_link)
5637 struct io_timeout_data *data;
5639 u32 off = READ_ONCE(sqe->off);
5641 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5643 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5645 if (off && is_timeout_link)
5647 flags = READ_ONCE(sqe->timeout_flags);
5648 if (flags & ~IORING_TIMEOUT_ABS)
5651 req->timeout.off = off;
5652 if (unlikely(off && !req->ctx->off_timeout_used))
5653 req->ctx->off_timeout_used = true;
5655 if (!req->async_data && io_alloc_async_data(req))
5658 data = req->async_data;
5661 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5664 data->mode = io_translate_timeout_mode(flags);
5665 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5666 if (is_timeout_link)
5667 io_req_track_inflight(req);
5671 static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
5673 struct io_ring_ctx *ctx = req->ctx;
5674 struct io_timeout_data *data = req->async_data;
5675 struct list_head *entry;
5676 u32 tail, off = req->timeout.off;
5678 spin_lock_irq(&ctx->completion_lock);
5681 * sqe->off holds how many events that need to occur for this
5682 * timeout event to be satisfied. If it isn't set, then this is
5683 * a pure timeout request, sequence isn't used.
5685 if (io_is_timeout_noseq(req)) {
5686 entry = ctx->timeout_list.prev;
5690 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5691 req->timeout.target_seq = tail + off;
5693 /* Update the last seq here in case io_flush_timeouts() hasn't.
5694 * This is safe because ->completion_lock is held, and submissions
5695 * and completions are never mixed in the same ->completion_lock section.
5697 ctx->cq_last_tm_flush = tail;
5700 * Insertion sort, ensuring the first entry in the list is always
5701 * the one we need first.
5703 list_for_each_prev(entry, &ctx->timeout_list) {
5704 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5707 if (io_is_timeout_noseq(nxt))
5709 /* nxt.seq is behind @tail, otherwise would've been completed */
5710 if (off >= nxt->timeout.target_seq - tail)
5714 list_add(&req->timeout.list, entry);
5715 data->timer.function = io_timeout_fn;
5716 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5717 spin_unlock_irq(&ctx->completion_lock);
5721 struct io_cancel_data {
5722 struct io_ring_ctx *ctx;
5726 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5728 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5729 struct io_cancel_data *cd = data;
5731 return req->ctx == cd->ctx && req->user_data == cd->user_data;
5734 static int io_async_cancel_one(struct io_uring_task *tctx, u64 user_data,
5735 struct io_ring_ctx *ctx)
5737 struct io_cancel_data data = { .ctx = ctx, .user_data = user_data, };
5738 enum io_wq_cancel cancel_ret;
5741 if (!tctx || !tctx->io_wq)
5744 cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, &data, false);
5745 switch (cancel_ret) {
5746 case IO_WQ_CANCEL_OK:
5749 case IO_WQ_CANCEL_RUNNING:
5752 case IO_WQ_CANCEL_NOTFOUND:
5760 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5761 struct io_kiocb *req, __u64 sqe_addr,
5764 unsigned long flags;
5767 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5768 spin_lock_irqsave(&ctx->completion_lock, flags);
5771 ret = io_timeout_cancel(ctx, sqe_addr);
5774 ret = io_poll_cancel(ctx, sqe_addr, false);
5778 io_cqring_fill_event(ctx, req->user_data, ret, 0);
5779 io_commit_cqring(ctx);
5780 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5781 io_cqring_ev_posted(ctx);
5787 static int io_async_cancel_prep(struct io_kiocb *req,
5788 const struct io_uring_sqe *sqe)
5790 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5792 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5794 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5797 req->cancel.addr = READ_ONCE(sqe->addr);
5801 static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
5803 struct io_ring_ctx *ctx = req->ctx;
5804 u64 sqe_addr = req->cancel.addr;
5805 struct io_tctx_node *node;
5808 /* tasks should wait for their io-wq threads, so safe w/o sync */
5809 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5810 spin_lock_irq(&ctx->completion_lock);
5813 ret = io_timeout_cancel(ctx, sqe_addr);
5816 ret = io_poll_cancel(ctx, sqe_addr, false);
5819 spin_unlock_irq(&ctx->completion_lock);
5821 /* slow path, try all io-wq's */
5822 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5824 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
5825 struct io_uring_task *tctx = node->task->io_uring;
5827 ret = io_async_cancel_one(tctx, req->cancel.addr, ctx);
5831 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5833 spin_lock_irq(&ctx->completion_lock);
5835 io_cqring_fill_event(ctx, req->user_data, ret, 0);
5836 io_commit_cqring(ctx);
5837 spin_unlock_irq(&ctx->completion_lock);
5838 io_cqring_ev_posted(ctx);
5846 static int io_rsrc_update_prep(struct io_kiocb *req,
5847 const struct io_uring_sqe *sqe)
5849 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5851 if (sqe->ioprio || sqe->rw_flags)
5854 req->rsrc_update.offset = READ_ONCE(sqe->off);
5855 req->rsrc_update.nr_args = READ_ONCE(sqe->len);
5856 if (!req->rsrc_update.nr_args)
5858 req->rsrc_update.arg = READ_ONCE(sqe->addr);
5862 static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
5864 struct io_ring_ctx *ctx = req->ctx;
5865 struct io_uring_rsrc_update2 up;
5868 if (issue_flags & IO_URING_F_NONBLOCK)
5871 up.offset = req->rsrc_update.offset;
5872 up.data = req->rsrc_update.arg;
5877 mutex_lock(&ctx->uring_lock);
5878 ret = __io_register_rsrc_update(ctx, IORING_RSRC_FILE,
5879 &up, req->rsrc_update.nr_args);
5880 mutex_unlock(&ctx->uring_lock);
5884 __io_req_complete(req, issue_flags, ret, 0);
5888 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5890 switch (req->opcode) {
5893 case IORING_OP_READV:
5894 case IORING_OP_READ_FIXED:
5895 case IORING_OP_READ:
5896 return io_read_prep(req, sqe);
5897 case IORING_OP_WRITEV:
5898 case IORING_OP_WRITE_FIXED:
5899 case IORING_OP_WRITE:
5900 return io_write_prep(req, sqe);
5901 case IORING_OP_POLL_ADD:
5902 return io_poll_add_prep(req, sqe);
5903 case IORING_OP_POLL_REMOVE:
5904 return io_poll_update_prep(req, sqe);
5905 case IORING_OP_FSYNC:
5906 return io_fsync_prep(req, sqe);
5907 case IORING_OP_SYNC_FILE_RANGE:
5908 return io_sfr_prep(req, sqe);
5909 case IORING_OP_SENDMSG:
5910 case IORING_OP_SEND:
5911 return io_sendmsg_prep(req, sqe);
5912 case IORING_OP_RECVMSG:
5913 case IORING_OP_RECV:
5914 return io_recvmsg_prep(req, sqe);
5915 case IORING_OP_CONNECT:
5916 return io_connect_prep(req, sqe);
5917 case IORING_OP_TIMEOUT:
5918 return io_timeout_prep(req, sqe, false);
5919 case IORING_OP_TIMEOUT_REMOVE:
5920 return io_timeout_remove_prep(req, sqe);
5921 case IORING_OP_ASYNC_CANCEL:
5922 return io_async_cancel_prep(req, sqe);
5923 case IORING_OP_LINK_TIMEOUT:
5924 return io_timeout_prep(req, sqe, true);
5925 case IORING_OP_ACCEPT:
5926 return io_accept_prep(req, sqe);
5927 case IORING_OP_FALLOCATE:
5928 return io_fallocate_prep(req, sqe);
5929 case IORING_OP_OPENAT:
5930 return io_openat_prep(req, sqe);
5931 case IORING_OP_CLOSE:
5932 return io_close_prep(req, sqe);
5933 case IORING_OP_FILES_UPDATE:
5934 return io_rsrc_update_prep(req, sqe);
5935 case IORING_OP_STATX:
5936 return io_statx_prep(req, sqe);
5937 case IORING_OP_FADVISE:
5938 return io_fadvise_prep(req, sqe);
5939 case IORING_OP_MADVISE:
5940 return io_madvise_prep(req, sqe);
5941 case IORING_OP_OPENAT2:
5942 return io_openat2_prep(req, sqe);
5943 case IORING_OP_EPOLL_CTL:
5944 return io_epoll_ctl_prep(req, sqe);
5945 case IORING_OP_SPLICE:
5946 return io_splice_prep(req, sqe);
5947 case IORING_OP_PROVIDE_BUFFERS:
5948 return io_provide_buffers_prep(req, sqe);
5949 case IORING_OP_REMOVE_BUFFERS:
5950 return io_remove_buffers_prep(req, sqe);
5952 return io_tee_prep(req, sqe);
5953 case IORING_OP_SHUTDOWN:
5954 return io_shutdown_prep(req, sqe);
5955 case IORING_OP_RENAMEAT:
5956 return io_renameat_prep(req, sqe);
5957 case IORING_OP_UNLINKAT:
5958 return io_unlinkat_prep(req, sqe);
5961 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5966 static int io_req_prep_async(struct io_kiocb *req)
5968 if (!io_op_defs[req->opcode].needs_async_setup)
5970 if (WARN_ON_ONCE(req->async_data))
5972 if (io_alloc_async_data(req))
5975 switch (req->opcode) {
5976 case IORING_OP_READV:
5977 return io_rw_prep_async(req, READ);
5978 case IORING_OP_WRITEV:
5979 return io_rw_prep_async(req, WRITE);
5980 case IORING_OP_SENDMSG:
5981 return io_sendmsg_prep_async(req);
5982 case IORING_OP_RECVMSG:
5983 return io_recvmsg_prep_async(req);
5984 case IORING_OP_CONNECT:
5985 return io_connect_prep_async(req);
5987 printk_once(KERN_WARNING "io_uring: prep_async() bad opcode %d\n",
5992 static u32 io_get_sequence(struct io_kiocb *req)
5994 u32 seq = req->ctx->cached_sq_head;
5996 /* need original cached_sq_head, but it was increased for each req */
5997 io_for_each_link(req, req)
6002 static bool io_drain_req(struct io_kiocb *req)
6004 struct io_kiocb *pos;
6005 struct io_ring_ctx *ctx = req->ctx;
6006 struct io_defer_entry *de;
6011 * If we need to drain a request in the middle of a link, drain the
6012 * head request and the next request/link after the current link.
6013 * Considering sequential execution of links, IOSQE_IO_DRAIN will be
6014 * maintained for every request of our link.
6016 if (ctx->drain_next) {
6017 req->flags |= REQ_F_IO_DRAIN;
6018 ctx->drain_next = false;
6020 /* not interested in head, start from the first linked */
6021 io_for_each_link(pos, req->link) {
6022 if (pos->flags & REQ_F_IO_DRAIN) {
6023 ctx->drain_next = true;
6024 req->flags |= REQ_F_IO_DRAIN;
6029 /* Still need defer if there is pending req in defer list. */
6030 if (likely(list_empty_careful(&ctx->defer_list) &&
6031 !(req->flags & REQ_F_IO_DRAIN))) {
6032 ctx->drain_active = false;
6036 seq = io_get_sequence(req);
6037 /* Still a chance to pass the sequence check */
6038 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
6041 ret = io_req_prep_async(req);
6044 io_prep_async_link(req);
6045 de = kmalloc(sizeof(*de), GFP_KERNEL);
6047 io_req_complete_failed(req, ret);
6051 spin_lock_irq(&ctx->completion_lock);
6052 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
6053 spin_unlock_irq(&ctx->completion_lock);
6055 io_queue_async_work(req);
6059 trace_io_uring_defer(ctx, req, req->user_data);
6062 list_add_tail(&de->list, &ctx->defer_list);
6063 spin_unlock_irq(&ctx->completion_lock);
6067 static void io_clean_op(struct io_kiocb *req)
6069 if (req->flags & REQ_F_BUFFER_SELECTED) {
6070 switch (req->opcode) {
6071 case IORING_OP_READV:
6072 case IORING_OP_READ_FIXED:
6073 case IORING_OP_READ:
6074 kfree((void *)(unsigned long)req->rw.addr);
6076 case IORING_OP_RECVMSG:
6077 case IORING_OP_RECV:
6078 kfree(req->sr_msg.kbuf);
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);
6122 if ((req->flags & REQ_F_POLLED) && req->apoll) {
6123 kfree(req->apoll->double_poll);
6127 if (req->flags & REQ_F_INFLIGHT) {
6128 struct io_uring_task *tctx = req->task->io_uring;
6130 atomic_dec(&tctx->inflight_tracked);
6132 if (req->flags & REQ_F_CREDS)
6133 put_cred(req->creds);
6135 req->flags &= ~IO_REQ_CLEAN_FLAGS;
6138 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
6140 struct io_ring_ctx *ctx = req->ctx;
6141 const struct cred *creds = NULL;
6144 if ((req->flags & REQ_F_CREDS) && req->creds != current_cred())
6145 creds = override_creds(req->creds);
6147 switch (req->opcode) {
6149 ret = io_nop(req, issue_flags);
6151 case IORING_OP_READV:
6152 case IORING_OP_READ_FIXED:
6153 case IORING_OP_READ:
6154 ret = io_read(req, issue_flags);
6156 case IORING_OP_WRITEV:
6157 case IORING_OP_WRITE_FIXED:
6158 case IORING_OP_WRITE:
6159 ret = io_write(req, issue_flags);
6161 case IORING_OP_FSYNC:
6162 ret = io_fsync(req, issue_flags);
6164 case IORING_OP_POLL_ADD:
6165 ret = io_poll_add(req, issue_flags);
6167 case IORING_OP_POLL_REMOVE:
6168 ret = io_poll_update(req, issue_flags);
6170 case IORING_OP_SYNC_FILE_RANGE:
6171 ret = io_sync_file_range(req, issue_flags);
6173 case IORING_OP_SENDMSG:
6174 ret = io_sendmsg(req, issue_flags);
6176 case IORING_OP_SEND:
6177 ret = io_send(req, issue_flags);
6179 case IORING_OP_RECVMSG:
6180 ret = io_recvmsg(req, issue_flags);
6182 case IORING_OP_RECV:
6183 ret = io_recv(req, issue_flags);
6185 case IORING_OP_TIMEOUT:
6186 ret = io_timeout(req, issue_flags);
6188 case IORING_OP_TIMEOUT_REMOVE:
6189 ret = io_timeout_remove(req, issue_flags);
6191 case IORING_OP_ACCEPT:
6192 ret = io_accept(req, issue_flags);
6194 case IORING_OP_CONNECT:
6195 ret = io_connect(req, issue_flags);
6197 case IORING_OP_ASYNC_CANCEL:
6198 ret = io_async_cancel(req, issue_flags);
6200 case IORING_OP_FALLOCATE:
6201 ret = io_fallocate(req, issue_flags);
6203 case IORING_OP_OPENAT:
6204 ret = io_openat(req, issue_flags);
6206 case IORING_OP_CLOSE:
6207 ret = io_close(req, issue_flags);
6209 case IORING_OP_FILES_UPDATE:
6210 ret = io_files_update(req, issue_flags);
6212 case IORING_OP_STATX:
6213 ret = io_statx(req, issue_flags);
6215 case IORING_OP_FADVISE:
6216 ret = io_fadvise(req, issue_flags);
6218 case IORING_OP_MADVISE:
6219 ret = io_madvise(req, issue_flags);
6221 case IORING_OP_OPENAT2:
6222 ret = io_openat2(req, issue_flags);
6224 case IORING_OP_EPOLL_CTL:
6225 ret = io_epoll_ctl(req, issue_flags);
6227 case IORING_OP_SPLICE:
6228 ret = io_splice(req, issue_flags);
6230 case IORING_OP_PROVIDE_BUFFERS:
6231 ret = io_provide_buffers(req, issue_flags);
6233 case IORING_OP_REMOVE_BUFFERS:
6234 ret = io_remove_buffers(req, issue_flags);
6237 ret = io_tee(req, issue_flags);
6239 case IORING_OP_SHUTDOWN:
6240 ret = io_shutdown(req, issue_flags);
6242 case IORING_OP_RENAMEAT:
6243 ret = io_renameat(req, issue_flags);
6245 case IORING_OP_UNLINKAT:
6246 ret = io_unlinkat(req, issue_flags);
6254 revert_creds(creds);
6257 /* If the op doesn't have a file, we're not polling for it */
6258 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file)
6259 io_iopoll_req_issued(req);
6264 static void io_wq_submit_work(struct io_wq_work *work)
6266 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6267 struct io_kiocb *timeout;
6270 timeout = io_prep_linked_timeout(req);
6272 io_queue_linked_timeout(timeout);
6274 if (work->flags & IO_WQ_WORK_CANCEL)
6279 ret = io_issue_sqe(req, 0);
6281 * We can get EAGAIN for polled IO even though we're
6282 * forcing a sync submission from here, since we can't
6283 * wait for request slots on the block side.
6291 /* avoid locking problems by failing it from a clean context */
6293 /* io-wq is going to take one down */
6295 io_req_task_queue_fail(req, ret);
6299 #define FFS_ASYNC_READ 0x1UL
6300 #define FFS_ASYNC_WRITE 0x2UL
6302 #define FFS_ISREG 0x4UL
6304 #define FFS_ISREG 0x0UL
6306 #define FFS_MASK ~(FFS_ASYNC_READ|FFS_ASYNC_WRITE|FFS_ISREG)
6308 static inline struct io_fixed_file *io_fixed_file_slot(struct io_file_table *table,
6311 struct io_fixed_file *table_l2;
6313 table_l2 = table->files[i >> IORING_FILE_TABLE_SHIFT];
6314 return &table_l2[i & IORING_FILE_TABLE_MASK];
6317 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6320 struct io_fixed_file *slot = io_fixed_file_slot(&ctx->file_table, index);
6322 return (struct file *) (slot->file_ptr & FFS_MASK);
6325 static void io_fixed_file_set(struct io_fixed_file *file_slot, struct file *file)
6327 unsigned long file_ptr = (unsigned long) file;
6329 if (__io_file_supports_async(file, READ))
6330 file_ptr |= FFS_ASYNC_READ;
6331 if (__io_file_supports_async(file, WRITE))
6332 file_ptr |= FFS_ASYNC_WRITE;
6333 if (S_ISREG(file_inode(file)->i_mode))
6334 file_ptr |= FFS_ISREG;
6335 file_slot->file_ptr = file_ptr;
6338 static struct file *io_file_get(struct io_submit_state *state,
6339 struct io_kiocb *req, int fd, bool fixed)
6341 struct io_ring_ctx *ctx = req->ctx;
6345 unsigned long file_ptr;
6347 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6349 fd = array_index_nospec(fd, ctx->nr_user_files);
6350 file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr;
6351 file = (struct file *) (file_ptr & FFS_MASK);
6352 file_ptr &= ~FFS_MASK;
6353 /* mask in overlapping REQ_F and FFS bits */
6354 req->flags |= (file_ptr << REQ_F_ASYNC_READ_BIT);
6355 io_req_set_rsrc_node(req);
6357 trace_io_uring_file_get(ctx, fd);
6358 file = __io_file_get(state, fd);
6360 /* we don't allow fixed io_uring files */
6361 if (file && unlikely(file->f_op == &io_uring_fops))
6362 io_req_track_inflight(req);
6368 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6370 struct io_timeout_data *data = container_of(timer,
6371 struct io_timeout_data, timer);
6372 struct io_kiocb *prev, *req = data->req;
6373 struct io_ring_ctx *ctx = req->ctx;
6374 unsigned long flags;
6376 spin_lock_irqsave(&ctx->completion_lock, flags);
6377 prev = req->timeout.head;
6378 req->timeout.head = NULL;
6381 * We don't expect the list to be empty, that will only happen if we
6382 * race with the completion of the linked work.
6385 io_remove_next_linked(prev);
6386 if (!req_ref_inc_not_zero(prev))
6389 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6392 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6393 io_put_req_deferred(prev, 1);
6394 io_put_req_deferred(req, 1);
6396 io_req_complete_post(req, -ETIME, 0);
6398 return HRTIMER_NORESTART;
6401 static void io_queue_linked_timeout(struct io_kiocb *req)
6403 struct io_ring_ctx *ctx = req->ctx;
6405 spin_lock_irq(&ctx->completion_lock);
6407 * If the back reference is NULL, then our linked request finished
6408 * before we got a chance to setup the timer
6410 if (req->timeout.head) {
6411 struct io_timeout_data *data = req->async_data;
6413 data->timer.function = io_link_timeout_fn;
6414 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6417 spin_unlock_irq(&ctx->completion_lock);
6418 /* drop submission reference */
6422 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6424 struct io_kiocb *nxt = req->link;
6426 if (!nxt || (req->flags & REQ_F_LINK_TIMEOUT) ||
6427 nxt->opcode != IORING_OP_LINK_TIMEOUT)
6430 nxt->timeout.head = req;
6431 nxt->flags |= REQ_F_LTIMEOUT_ACTIVE;
6432 req->flags |= REQ_F_LINK_TIMEOUT;
6436 static void __io_queue_sqe(struct io_kiocb *req)
6438 struct io_kiocb *linked_timeout = io_prep_linked_timeout(req);
6441 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
6444 * We async punt it if the file wasn't marked NOWAIT, or if the file
6445 * doesn't support non-blocking read/write attempts
6448 /* drop submission reference */
6449 if (req->flags & REQ_F_COMPLETE_INLINE) {
6450 struct io_ring_ctx *ctx = req->ctx;
6451 struct io_comp_state *cs = &ctx->submit_state.comp;
6453 cs->reqs[cs->nr++] = req;
6454 if (cs->nr == ARRAY_SIZE(cs->reqs))
6455 io_submit_flush_completions(ctx);
6459 } else if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6460 if (!io_arm_poll_handler(req)) {
6462 * Queued up for async execution, worker will release
6463 * submit reference when the iocb is actually submitted.
6465 io_queue_async_work(req);
6468 io_req_complete_failed(req, ret);
6471 io_queue_linked_timeout(linked_timeout);
6474 static inline void io_queue_sqe(struct io_kiocb *req)
6476 if (unlikely(req->ctx->drain_active) && io_drain_req(req))
6479 if (likely(!(req->flags & REQ_F_FORCE_ASYNC))) {
6480 __io_queue_sqe(req);
6482 int ret = io_req_prep_async(req);
6485 io_req_complete_failed(req, ret);
6487 io_queue_async_work(req);
6492 * Check SQE restrictions (opcode and flags).
6494 * Returns 'true' if SQE is allowed, 'false' otherwise.
6496 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6497 struct io_kiocb *req,
6498 unsigned int sqe_flags)
6500 if (!ctx->restricted)
6503 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6506 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6507 ctx->restrictions.sqe_flags_required)
6510 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6511 ctx->restrictions.sqe_flags_required))
6517 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6518 const struct io_uring_sqe *sqe)
6520 struct io_submit_state *state;
6521 unsigned int sqe_flags;
6522 int personality, ret = 0;
6524 req->opcode = READ_ONCE(sqe->opcode);
6525 /* same numerical values with corresponding REQ_F_*, safe to copy */
6526 req->flags = sqe_flags = READ_ONCE(sqe->flags);
6527 req->user_data = READ_ONCE(sqe->user_data);
6528 req->async_data = NULL;
6532 req->fixed_rsrc_refs = NULL;
6533 /* one is dropped after submission, the other at completion */
6534 atomic_set(&req->refs, 2);
6535 req->task = current;
6538 /* enforce forwards compatibility on users */
6539 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6541 if (unlikely(req->opcode >= IORING_OP_LAST))
6543 if (unlikely(!io_check_restriction(ctx, req, sqe_flags)))
6546 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6547 !io_op_defs[req->opcode].buffer_select)
6549 if (unlikely(sqe_flags & IOSQE_IO_DRAIN))
6550 ctx->drain_active = true;
6552 personality = READ_ONCE(sqe->personality);
6554 req->creds = xa_load(&ctx->personalities, personality);
6557 get_cred(req->creds);
6558 req->flags |= REQ_F_CREDS;
6560 state = &ctx->submit_state;
6563 * Plug now if we have more than 1 IO left after this, and the target
6564 * is potentially a read/write to block based storage.
6566 if (!state->plug_started && state->ios_left > 1 &&
6567 io_op_defs[req->opcode].plug) {
6568 blk_start_plug(&state->plug);
6569 state->plug_started = true;
6572 if (io_op_defs[req->opcode].needs_file) {
6573 bool fixed = req->flags & REQ_F_FIXED_FILE;
6575 req->file = io_file_get(state, req, READ_ONCE(sqe->fd), fixed);
6576 if (unlikely(!req->file))
6584 static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
6585 const struct io_uring_sqe *sqe)
6587 struct io_submit_link *link = &ctx->submit_state.link;
6590 ret = io_init_req(ctx, req, sqe);
6591 if (unlikely(ret)) {
6594 /* fail even hard links since we don't submit */
6595 req_set_fail(link->head);
6596 io_req_complete_failed(link->head, -ECANCELED);
6599 io_req_complete_failed(req, ret);
6603 ret = io_req_prep(req, sqe);
6607 /* don't need @sqe from now on */
6608 trace_io_uring_submit_sqe(ctx, req, req->opcode, req->user_data,
6610 ctx->flags & IORING_SETUP_SQPOLL);
6613 * If we already have a head request, queue this one for async
6614 * submittal once the head completes. If we don't have a head but
6615 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6616 * submitted sync once the chain is complete. If none of those
6617 * conditions are true (normal request), then just queue it.
6620 struct io_kiocb *head = link->head;
6622 ret = io_req_prep_async(req);
6625 trace_io_uring_link(ctx, req, head);
6626 link->last->link = req;
6629 /* last request of a link, enqueue the link */
6630 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6635 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6647 * Batched submission is done, ensure local IO is flushed out.
6649 static void io_submit_state_end(struct io_submit_state *state,
6650 struct io_ring_ctx *ctx)
6652 if (state->link.head)
6653 io_queue_sqe(state->link.head);
6655 io_submit_flush_completions(ctx);
6656 if (state->plug_started)
6657 blk_finish_plug(&state->plug);
6658 io_state_file_put(state);
6662 * Start submission side cache.
6664 static void io_submit_state_start(struct io_submit_state *state,
6665 unsigned int max_ios)
6667 state->plug_started = false;
6668 state->ios_left = max_ios;
6669 /* set only head, no need to init link_last in advance */
6670 state->link.head = NULL;
6673 static void io_commit_sqring(struct io_ring_ctx *ctx)
6675 struct io_rings *rings = ctx->rings;
6678 * Ensure any loads from the SQEs are done at this point,
6679 * since once we write the new head, the application could
6680 * write new data to them.
6682 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6686 * Fetch an sqe, if one is available. Note this returns a pointer to memory
6687 * that is mapped by userspace. This means that care needs to be taken to
6688 * ensure that reads are stable, as we cannot rely on userspace always
6689 * being a good citizen. If members of the sqe are validated and then later
6690 * used, it's important that those reads are done through READ_ONCE() to
6691 * prevent a re-load down the line.
6693 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6695 unsigned head, mask = ctx->sq_entries - 1;
6696 unsigned sq_idx = ctx->cached_sq_head++ & mask;
6699 * The cached sq head (or cq tail) serves two purposes:
6701 * 1) allows us to batch the cost of updating the user visible
6703 * 2) allows the kernel side to track the head on its own, even
6704 * though the application is the one updating it.
6706 head = READ_ONCE(ctx->sq_array[sq_idx]);
6707 if (likely(head < ctx->sq_entries))
6708 return &ctx->sq_sqes[head];
6710 /* drop invalid entries */
6712 WRITE_ONCE(ctx->rings->sq_dropped,
6713 READ_ONCE(ctx->rings->sq_dropped) + 1);
6717 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6719 struct io_uring_task *tctx;
6722 /* make sure SQ entry isn't read before tail */
6723 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6724 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6727 tctx = current->io_uring;
6728 tctx->cached_refs -= nr;
6729 if (unlikely(tctx->cached_refs < 0)) {
6730 unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR;
6732 percpu_counter_add(&tctx->inflight, refill);
6733 refcount_add(refill, ¤t->usage);
6734 tctx->cached_refs += refill;
6736 io_submit_state_start(&ctx->submit_state, nr);
6738 while (submitted < nr) {
6739 const struct io_uring_sqe *sqe;
6740 struct io_kiocb *req;
6742 req = io_alloc_req(ctx);
6743 if (unlikely(!req)) {
6745 submitted = -EAGAIN;
6748 sqe = io_get_sqe(ctx);
6749 if (unlikely(!sqe)) {
6750 kmem_cache_free(req_cachep, req);
6753 /* will complete beyond this point, count as submitted */
6755 if (io_submit_sqe(ctx, req, sqe))
6759 if (unlikely(submitted != nr)) {
6760 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6761 int unused = nr - ref_used;
6763 current->io_uring->cached_refs += unused;
6764 percpu_ref_put_many(&ctx->refs, unused);
6767 io_submit_state_end(&ctx->submit_state, ctx);
6768 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6769 io_commit_sqring(ctx);
6774 static inline bool io_sqd_events_pending(struct io_sq_data *sqd)
6776 return READ_ONCE(sqd->state);
6779 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6781 /* Tell userspace we may need a wakeup call */
6782 spin_lock_irq(&ctx->completion_lock);
6783 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6784 spin_unlock_irq(&ctx->completion_lock);
6787 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6789 spin_lock_irq(&ctx->completion_lock);
6790 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6791 spin_unlock_irq(&ctx->completion_lock);
6794 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
6796 unsigned int to_submit;
6799 to_submit = io_sqring_entries(ctx);
6800 /* if we're handling multiple rings, cap submit size for fairness */
6801 if (cap_entries && to_submit > IORING_SQPOLL_CAP_ENTRIES_VALUE)
6802 to_submit = IORING_SQPOLL_CAP_ENTRIES_VALUE;
6804 if (!list_empty(&ctx->iopoll_list) || to_submit) {
6805 unsigned nr_events = 0;
6806 const struct cred *creds = NULL;
6808 if (ctx->sq_creds != current_cred())
6809 creds = override_creds(ctx->sq_creds);
6811 mutex_lock(&ctx->uring_lock);
6812 if (!list_empty(&ctx->iopoll_list))
6813 io_do_iopoll(ctx, &nr_events, 0);
6816 * Don't submit if refs are dying, good for io_uring_register(),
6817 * but also it is relied upon by io_ring_exit_work()
6819 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)) &&
6820 !(ctx->flags & IORING_SETUP_R_DISABLED))
6821 ret = io_submit_sqes(ctx, to_submit);
6822 mutex_unlock(&ctx->uring_lock);
6824 if (to_submit && wq_has_sleeper(&ctx->sqo_sq_wait))
6825 wake_up(&ctx->sqo_sq_wait);
6827 revert_creds(creds);
6833 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
6835 struct io_ring_ctx *ctx;
6836 unsigned sq_thread_idle = 0;
6838 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6839 sq_thread_idle = max(sq_thread_idle, ctx->sq_thread_idle);
6840 sqd->sq_thread_idle = sq_thread_idle;
6843 static bool io_sqd_handle_event(struct io_sq_data *sqd)
6845 bool did_sig = false;
6846 struct ksignal ksig;
6848 if (test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state) ||
6849 signal_pending(current)) {
6850 mutex_unlock(&sqd->lock);
6851 if (signal_pending(current))
6852 did_sig = get_signal(&ksig);
6854 mutex_lock(&sqd->lock);
6857 return did_sig || test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
6860 static int io_sq_thread(void *data)
6862 struct io_sq_data *sqd = data;
6863 struct io_ring_ctx *ctx;
6864 unsigned long timeout = 0;
6865 char buf[TASK_COMM_LEN];
6868 snprintf(buf, sizeof(buf), "iou-sqp-%d", sqd->task_pid);
6869 set_task_comm(current, buf);
6871 if (sqd->sq_cpu != -1)
6872 set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
6874 set_cpus_allowed_ptr(current, cpu_online_mask);
6875 current->flags |= PF_NO_SETAFFINITY;
6877 mutex_lock(&sqd->lock);
6879 bool cap_entries, sqt_spin = false;
6881 if (io_sqd_events_pending(sqd) || signal_pending(current)) {
6882 if (io_sqd_handle_event(sqd))
6884 timeout = jiffies + sqd->sq_thread_idle;
6888 cap_entries = !list_is_singular(&sqd->ctx_list);
6889 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6890 int ret = __io_sq_thread(ctx, cap_entries);
6892 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
6896 if (sqt_spin || !time_after(jiffies, timeout)) {
6900 timeout = jiffies + sqd->sq_thread_idle;
6904 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
6905 if (!io_sqd_events_pending(sqd) && !io_run_task_work()) {
6906 bool needs_sched = true;
6908 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6909 io_ring_set_wakeup_flag(ctx);
6911 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6912 !list_empty_careful(&ctx->iopoll_list)) {
6913 needs_sched = false;
6916 if (io_sqring_entries(ctx)) {
6917 needs_sched = false;
6923 mutex_unlock(&sqd->lock);
6925 mutex_lock(&sqd->lock);
6927 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6928 io_ring_clear_wakeup_flag(ctx);
6931 finish_wait(&sqd->wait, &wait);
6932 timeout = jiffies + sqd->sq_thread_idle;
6935 io_uring_cancel_generic(true, sqd);
6937 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6938 io_ring_set_wakeup_flag(ctx);
6940 mutex_unlock(&sqd->lock);
6942 complete(&sqd->exited);
6946 struct io_wait_queue {
6947 struct wait_queue_entry wq;
6948 struct io_ring_ctx *ctx;
6950 unsigned nr_timeouts;
6953 static inline bool io_should_wake(struct io_wait_queue *iowq)
6955 struct io_ring_ctx *ctx = iowq->ctx;
6958 * Wake up if we have enough events, or if a timeout occurred since we
6959 * started waiting. For timeouts, we always want to return to userspace,
6960 * regardless of event count.
6962 return io_cqring_events(ctx) >= iowq->to_wait ||
6963 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6966 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6967 int wake_flags, void *key)
6969 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6973 * Cannot safely flush overflowed CQEs from here, ensure we wake up
6974 * the task, and the next invocation will do it.
6976 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->check_cq_overflow))
6977 return autoremove_wake_function(curr, mode, wake_flags, key);
6981 static int io_run_task_work_sig(void)
6983 if (io_run_task_work())
6985 if (!signal_pending(current))
6987 if (test_thread_flag(TIF_NOTIFY_SIGNAL))
6988 return -ERESTARTSYS;
6992 /* when returns >0, the caller should retry */
6993 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
6994 struct io_wait_queue *iowq,
6995 signed long *timeout)
6999 /* make sure we run task_work before checking for signals */
7000 ret = io_run_task_work_sig();
7001 if (ret || io_should_wake(iowq))
7003 /* let the caller flush overflows, retry */
7004 if (test_bit(0, &ctx->check_cq_overflow))
7007 *timeout = schedule_timeout(*timeout);
7008 return !*timeout ? -ETIME : 1;
7012 * Wait until events become available, if we don't already have some. The
7013 * application must reap them itself, as they reside on the shared cq ring.
7015 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
7016 const sigset_t __user *sig, size_t sigsz,
7017 struct __kernel_timespec __user *uts)
7019 struct io_wait_queue iowq = {
7022 .func = io_wake_function,
7023 .entry = LIST_HEAD_INIT(iowq.wq.entry),
7026 .to_wait = min_events,
7028 struct io_rings *rings = ctx->rings;
7029 signed long timeout = MAX_SCHEDULE_TIMEOUT;
7033 io_cqring_overflow_flush(ctx, false);
7034 if (io_cqring_events(ctx) >= min_events)
7036 if (!io_run_task_work())
7041 #ifdef CONFIG_COMPAT
7042 if (in_compat_syscall())
7043 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
7047 ret = set_user_sigmask(sig, sigsz);
7054 struct timespec64 ts;
7056 if (get_timespec64(&ts, uts))
7058 timeout = timespec64_to_jiffies(&ts);
7061 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
7062 trace_io_uring_cqring_wait(ctx, min_events);
7064 /* if we can't even flush overflow, don't wait for more */
7065 if (!io_cqring_overflow_flush(ctx, false)) {
7069 prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq,
7070 TASK_INTERRUPTIBLE);
7071 ret = io_cqring_wait_schedule(ctx, &iowq, &timeout);
7072 finish_wait(&ctx->cq_wait, &iowq.wq);
7076 restore_saved_sigmask_unless(ret == -EINTR);
7078 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
7081 static void io_free_page_table(void **table, size_t size)
7083 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7085 for (i = 0; i < nr_tables; i++)
7090 static void **io_alloc_page_table(size_t size)
7092 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7093 size_t init_size = size;
7096 table = kcalloc(nr_tables, sizeof(*table), GFP_KERNEL);
7100 for (i = 0; i < nr_tables; i++) {
7101 unsigned int this_size = min_t(size_t, size, PAGE_SIZE);
7103 table[i] = kzalloc(this_size, GFP_KERNEL);
7105 io_free_page_table(table, init_size);
7113 static inline void io_rsrc_ref_lock(struct io_ring_ctx *ctx)
7115 spin_lock_bh(&ctx->rsrc_ref_lock);
7118 static inline void io_rsrc_ref_unlock(struct io_ring_ctx *ctx)
7120 spin_unlock_bh(&ctx->rsrc_ref_lock);
7123 static void io_rsrc_node_destroy(struct io_rsrc_node *ref_node)
7125 percpu_ref_exit(&ref_node->refs);
7129 static void io_rsrc_node_switch(struct io_ring_ctx *ctx,
7130 struct io_rsrc_data *data_to_kill)
7132 WARN_ON_ONCE(!ctx->rsrc_backup_node);
7133 WARN_ON_ONCE(data_to_kill && !ctx->rsrc_node);
7136 struct io_rsrc_node *rsrc_node = ctx->rsrc_node;
7138 rsrc_node->rsrc_data = data_to_kill;
7139 io_rsrc_ref_lock(ctx);
7140 list_add_tail(&rsrc_node->node, &ctx->rsrc_ref_list);
7141 io_rsrc_ref_unlock(ctx);
7143 atomic_inc(&data_to_kill->refs);
7144 percpu_ref_kill(&rsrc_node->refs);
7145 ctx->rsrc_node = NULL;
7148 if (!ctx->rsrc_node) {
7149 ctx->rsrc_node = ctx->rsrc_backup_node;
7150 ctx->rsrc_backup_node = NULL;
7154 static int io_rsrc_node_switch_start(struct io_ring_ctx *ctx)
7156 if (ctx->rsrc_backup_node)
7158 ctx->rsrc_backup_node = io_rsrc_node_alloc(ctx);
7159 return ctx->rsrc_backup_node ? 0 : -ENOMEM;
7162 static int io_rsrc_ref_quiesce(struct io_rsrc_data *data, struct io_ring_ctx *ctx)
7166 /* As we may drop ->uring_lock, other task may have started quiesce */
7170 data->quiesce = true;
7172 ret = io_rsrc_node_switch_start(ctx);
7175 io_rsrc_node_switch(ctx, data);
7177 /* kill initial ref, already quiesced if zero */
7178 if (atomic_dec_and_test(&data->refs))
7180 flush_delayed_work(&ctx->rsrc_put_work);
7181 ret = wait_for_completion_interruptible(&data->done);
7185 atomic_inc(&data->refs);
7186 /* wait for all works potentially completing data->done */
7187 flush_delayed_work(&ctx->rsrc_put_work);
7188 reinit_completion(&data->done);
7190 mutex_unlock(&ctx->uring_lock);
7191 ret = io_run_task_work_sig();
7192 mutex_lock(&ctx->uring_lock);
7194 data->quiesce = false;
7199 static u64 *io_get_tag_slot(struct io_rsrc_data *data, unsigned int idx)
7201 unsigned int off = idx & IO_RSRC_TAG_TABLE_MASK;
7202 unsigned int table_idx = idx >> IO_RSRC_TAG_TABLE_SHIFT;
7204 return &data->tags[table_idx][off];
7207 static void io_rsrc_data_free(struct io_rsrc_data *data)
7209 size_t size = data->nr * sizeof(data->tags[0][0]);
7212 io_free_page_table((void **)data->tags, size);
7216 static int io_rsrc_data_alloc(struct io_ring_ctx *ctx, rsrc_put_fn *do_put,
7217 u64 __user *utags, unsigned nr,
7218 struct io_rsrc_data **pdata)
7220 struct io_rsrc_data *data;
7224 data = kzalloc(sizeof(*data), GFP_KERNEL);
7227 data->tags = (u64 **)io_alloc_page_table(nr * sizeof(data->tags[0][0]));
7235 data->do_put = do_put;
7238 for (i = 0; i < nr; i++) {
7239 u64 *tag_slot = io_get_tag_slot(data, i);
7241 if (copy_from_user(tag_slot, &utags[i],
7247 atomic_set(&data->refs, 1);
7248 init_completion(&data->done);
7252 io_rsrc_data_free(data);
7256 static bool io_alloc_file_tables(struct io_file_table *table, unsigned nr_files)
7258 size_t size = nr_files * sizeof(struct io_fixed_file);
7260 table->files = (struct io_fixed_file **)io_alloc_page_table(size);
7261 return !!table->files;
7264 static void io_free_file_tables(struct io_file_table *table, unsigned nr_files)
7266 size_t size = nr_files * sizeof(struct io_fixed_file);
7268 io_free_page_table((void **)table->files, size);
7269 table->files = NULL;
7272 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
7274 #if defined(CONFIG_UNIX)
7275 if (ctx->ring_sock) {
7276 struct sock *sock = ctx->ring_sock->sk;
7277 struct sk_buff *skb;
7279 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
7285 for (i = 0; i < ctx->nr_user_files; i++) {
7288 file = io_file_from_index(ctx, i);
7293 io_free_file_tables(&ctx->file_table, ctx->nr_user_files);
7294 io_rsrc_data_free(ctx->file_data);
7295 ctx->file_data = NULL;
7296 ctx->nr_user_files = 0;
7299 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7303 if (!ctx->file_data)
7305 ret = io_rsrc_ref_quiesce(ctx->file_data, ctx);
7307 __io_sqe_files_unregister(ctx);
7311 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7312 __releases(&sqd->lock)
7314 WARN_ON_ONCE(sqd->thread == current);
7317 * Do the dance but not conditional clear_bit() because it'd race with
7318 * other threads incrementing park_pending and setting the bit.
7320 clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7321 if (atomic_dec_return(&sqd->park_pending))
7322 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7323 mutex_unlock(&sqd->lock);
7326 static void io_sq_thread_park(struct io_sq_data *sqd)
7327 __acquires(&sqd->lock)
7329 WARN_ON_ONCE(sqd->thread == current);
7331 atomic_inc(&sqd->park_pending);
7332 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7333 mutex_lock(&sqd->lock);
7335 wake_up_process(sqd->thread);
7338 static void io_sq_thread_stop(struct io_sq_data *sqd)
7340 WARN_ON_ONCE(sqd->thread == current);
7341 WARN_ON_ONCE(test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state));
7343 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7344 mutex_lock(&sqd->lock);
7346 wake_up_process(sqd->thread);
7347 mutex_unlock(&sqd->lock);
7348 wait_for_completion(&sqd->exited);
7351 static void io_put_sq_data(struct io_sq_data *sqd)
7353 if (refcount_dec_and_test(&sqd->refs)) {
7354 WARN_ON_ONCE(atomic_read(&sqd->park_pending));
7356 io_sq_thread_stop(sqd);
7361 static void io_sq_thread_finish(struct io_ring_ctx *ctx)
7363 struct io_sq_data *sqd = ctx->sq_data;
7366 io_sq_thread_park(sqd);
7367 list_del_init(&ctx->sqd_list);
7368 io_sqd_update_thread_idle(sqd);
7369 io_sq_thread_unpark(sqd);
7371 io_put_sq_data(sqd);
7372 ctx->sq_data = NULL;
7376 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7378 struct io_ring_ctx *ctx_attach;
7379 struct io_sq_data *sqd;
7382 f = fdget(p->wq_fd);
7384 return ERR_PTR(-ENXIO);
7385 if (f.file->f_op != &io_uring_fops) {
7387 return ERR_PTR(-EINVAL);
7390 ctx_attach = f.file->private_data;
7391 sqd = ctx_attach->sq_data;
7394 return ERR_PTR(-EINVAL);
7396 if (sqd->task_tgid != current->tgid) {
7398 return ERR_PTR(-EPERM);
7401 refcount_inc(&sqd->refs);
7406 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p,
7409 struct io_sq_data *sqd;
7412 if (p->flags & IORING_SETUP_ATTACH_WQ) {
7413 sqd = io_attach_sq_data(p);
7418 /* fall through for EPERM case, setup new sqd/task */
7419 if (PTR_ERR(sqd) != -EPERM)
7423 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7425 return ERR_PTR(-ENOMEM);
7427 atomic_set(&sqd->park_pending, 0);
7428 refcount_set(&sqd->refs, 1);
7429 INIT_LIST_HEAD(&sqd->ctx_list);
7430 mutex_init(&sqd->lock);
7431 init_waitqueue_head(&sqd->wait);
7432 init_completion(&sqd->exited);
7436 #if defined(CONFIG_UNIX)
7438 * Ensure the UNIX gc is aware of our file set, so we are certain that
7439 * the io_uring can be safely unregistered on process exit, even if we have
7440 * loops in the file referencing.
7442 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7444 struct sock *sk = ctx->ring_sock->sk;
7445 struct scm_fp_list *fpl;
7446 struct sk_buff *skb;
7449 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7453 skb = alloc_skb(0, GFP_KERNEL);
7462 fpl->user = get_uid(current_user());
7463 for (i = 0; i < nr; i++) {
7464 struct file *file = io_file_from_index(ctx, i + offset);
7468 fpl->fp[nr_files] = get_file(file);
7469 unix_inflight(fpl->user, fpl->fp[nr_files]);
7474 fpl->max = SCM_MAX_FD;
7475 fpl->count = nr_files;
7476 UNIXCB(skb).fp = fpl;
7477 skb->destructor = unix_destruct_scm;
7478 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7479 skb_queue_head(&sk->sk_receive_queue, skb);
7481 for (i = 0; i < nr_files; i++)
7492 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7493 * causes regular reference counting to break down. We rely on the UNIX
7494 * garbage collection to take care of this problem for us.
7496 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7498 unsigned left, total;
7502 left = ctx->nr_user_files;
7504 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7506 ret = __io_sqe_files_scm(ctx, this_files, total);
7510 total += this_files;
7516 while (total < ctx->nr_user_files) {
7517 struct file *file = io_file_from_index(ctx, total);
7527 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7533 static void io_rsrc_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
7535 struct file *file = prsrc->file;
7536 #if defined(CONFIG_UNIX)
7537 struct sock *sock = ctx->ring_sock->sk;
7538 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7539 struct sk_buff *skb;
7542 __skb_queue_head_init(&list);
7545 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7546 * remove this entry and rearrange the file array.
7548 skb = skb_dequeue(head);
7550 struct scm_fp_list *fp;
7552 fp = UNIXCB(skb).fp;
7553 for (i = 0; i < fp->count; i++) {
7556 if (fp->fp[i] != file)
7559 unix_notinflight(fp->user, fp->fp[i]);
7560 left = fp->count - 1 - i;
7562 memmove(&fp->fp[i], &fp->fp[i + 1],
7563 left * sizeof(struct file *));
7570 __skb_queue_tail(&list, skb);
7580 __skb_queue_tail(&list, skb);
7582 skb = skb_dequeue(head);
7585 if (skb_peek(&list)) {
7586 spin_lock_irq(&head->lock);
7587 while ((skb = __skb_dequeue(&list)) != NULL)
7588 __skb_queue_tail(head, skb);
7589 spin_unlock_irq(&head->lock);
7596 static void __io_rsrc_put_work(struct io_rsrc_node *ref_node)
7598 struct io_rsrc_data *rsrc_data = ref_node->rsrc_data;
7599 struct io_ring_ctx *ctx = rsrc_data->ctx;
7600 struct io_rsrc_put *prsrc, *tmp;
7602 list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
7603 list_del(&prsrc->list);
7606 bool lock_ring = ctx->flags & IORING_SETUP_IOPOLL;
7608 io_ring_submit_lock(ctx, lock_ring);
7609 spin_lock_irq(&ctx->completion_lock);
7610 io_cqring_fill_event(ctx, prsrc->tag, 0, 0);
7612 io_commit_cqring(ctx);
7613 spin_unlock_irq(&ctx->completion_lock);
7614 io_cqring_ev_posted(ctx);
7615 io_ring_submit_unlock(ctx, lock_ring);
7618 rsrc_data->do_put(ctx, prsrc);
7622 io_rsrc_node_destroy(ref_node);
7623 if (atomic_dec_and_test(&rsrc_data->refs))
7624 complete(&rsrc_data->done);
7627 static void io_rsrc_put_work(struct work_struct *work)
7629 struct io_ring_ctx *ctx;
7630 struct llist_node *node;
7632 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
7633 node = llist_del_all(&ctx->rsrc_put_llist);
7636 struct io_rsrc_node *ref_node;
7637 struct llist_node *next = node->next;
7639 ref_node = llist_entry(node, struct io_rsrc_node, llist);
7640 __io_rsrc_put_work(ref_node);
7645 static void io_rsrc_node_ref_zero(struct percpu_ref *ref)
7647 struct io_rsrc_node *node = container_of(ref, struct io_rsrc_node, refs);
7648 struct io_ring_ctx *ctx = node->rsrc_data->ctx;
7649 bool first_add = false;
7651 io_rsrc_ref_lock(ctx);
7654 while (!list_empty(&ctx->rsrc_ref_list)) {
7655 node = list_first_entry(&ctx->rsrc_ref_list,
7656 struct io_rsrc_node, node);
7657 /* recycle ref nodes in order */
7660 list_del(&node->node);
7661 first_add |= llist_add(&node->llist, &ctx->rsrc_put_llist);
7663 io_rsrc_ref_unlock(ctx);
7666 mod_delayed_work(system_wq, &ctx->rsrc_put_work, HZ);
7669 static struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx)
7671 struct io_rsrc_node *ref_node;
7673 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7677 if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
7682 INIT_LIST_HEAD(&ref_node->node);
7683 INIT_LIST_HEAD(&ref_node->rsrc_list);
7684 ref_node->done = false;
7688 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7689 unsigned nr_args, u64 __user *tags)
7691 __s32 __user *fds = (__s32 __user *) arg;
7700 if (nr_args > IORING_MAX_FIXED_FILES)
7702 ret = io_rsrc_node_switch_start(ctx);
7705 ret = io_rsrc_data_alloc(ctx, io_rsrc_file_put, tags, nr_args,
7711 if (!io_alloc_file_tables(&ctx->file_table, nr_args))
7714 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7715 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7719 /* allow sparse sets */
7722 if (unlikely(*io_get_tag_slot(ctx->file_data, i)))
7729 if (unlikely(!file))
7733 * Don't allow io_uring instances to be registered. If UNIX
7734 * isn't enabled, then this causes a reference cycle and this
7735 * instance can never get freed. If UNIX is enabled we'll
7736 * handle it just fine, but there's still no point in allowing
7737 * a ring fd as it doesn't support regular read/write anyway.
7739 if (file->f_op == &io_uring_fops) {
7743 io_fixed_file_set(io_fixed_file_slot(&ctx->file_table, i), file);
7746 ret = io_sqe_files_scm(ctx);
7748 __io_sqe_files_unregister(ctx);
7752 io_rsrc_node_switch(ctx, NULL);
7755 for (i = 0; i < ctx->nr_user_files; i++) {
7756 file = io_file_from_index(ctx, i);
7760 io_free_file_tables(&ctx->file_table, nr_args);
7761 ctx->nr_user_files = 0;
7763 io_rsrc_data_free(ctx->file_data);
7764 ctx->file_data = NULL;
7768 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7771 #if defined(CONFIG_UNIX)
7772 struct sock *sock = ctx->ring_sock->sk;
7773 struct sk_buff_head *head = &sock->sk_receive_queue;
7774 struct sk_buff *skb;
7777 * See if we can merge this file into an existing skb SCM_RIGHTS
7778 * file set. If there's no room, fall back to allocating a new skb
7779 * and filling it in.
7781 spin_lock_irq(&head->lock);
7782 skb = skb_peek(head);
7784 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7786 if (fpl->count < SCM_MAX_FD) {
7787 __skb_unlink(skb, head);
7788 spin_unlock_irq(&head->lock);
7789 fpl->fp[fpl->count] = get_file(file);
7790 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7792 spin_lock_irq(&head->lock);
7793 __skb_queue_head(head, skb);
7798 spin_unlock_irq(&head->lock);
7805 return __io_sqe_files_scm(ctx, 1, index);
7811 static int io_queue_rsrc_removal(struct io_rsrc_data *data, unsigned idx,
7812 struct io_rsrc_node *node, void *rsrc)
7814 struct io_rsrc_put *prsrc;
7816 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
7820 prsrc->tag = *io_get_tag_slot(data, idx);
7822 list_add(&prsrc->list, &node->rsrc_list);
7826 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7827 struct io_uring_rsrc_update2 *up,
7830 u64 __user *tags = u64_to_user_ptr(up->tags);
7831 __s32 __user *fds = u64_to_user_ptr(up->data);
7832 struct io_rsrc_data *data = ctx->file_data;
7833 struct io_fixed_file *file_slot;
7837 bool needs_switch = false;
7839 if (!ctx->file_data)
7841 if (up->offset + nr_args > ctx->nr_user_files)
7844 for (done = 0; done < nr_args; done++) {
7847 if ((tags && copy_from_user(&tag, &tags[done], sizeof(tag))) ||
7848 copy_from_user(&fd, &fds[done], sizeof(fd))) {
7852 if ((fd == IORING_REGISTER_FILES_SKIP || fd == -1) && tag) {
7856 if (fd == IORING_REGISTER_FILES_SKIP)
7859 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
7860 file_slot = io_fixed_file_slot(&ctx->file_table, i);
7862 if (file_slot->file_ptr) {
7863 file = (struct file *)(file_slot->file_ptr & FFS_MASK);
7864 err = io_queue_rsrc_removal(data, up->offset + done,
7865 ctx->rsrc_node, file);
7868 file_slot->file_ptr = 0;
7869 needs_switch = true;
7878 * Don't allow io_uring instances to be registered. If
7879 * UNIX isn't enabled, then this causes a reference
7880 * cycle and this instance can never get freed. If UNIX
7881 * is enabled we'll handle it just fine, but there's
7882 * still no point in allowing a ring fd as it doesn't
7883 * support regular read/write anyway.
7885 if (file->f_op == &io_uring_fops) {
7890 *io_get_tag_slot(data, up->offset + done) = tag;
7891 io_fixed_file_set(file_slot, file);
7892 err = io_sqe_file_register(ctx, file, i);
7894 file_slot->file_ptr = 0;
7902 io_rsrc_node_switch(ctx, data);
7903 return done ? done : err;
7906 static struct io_wq_work *io_free_work(struct io_wq_work *work)
7908 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7910 req = io_put_req_find_next(req);
7911 return req ? &req->work : NULL;
7914 static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx,
7915 struct task_struct *task)
7917 struct io_wq_hash *hash;
7918 struct io_wq_data data;
7919 unsigned int concurrency;
7921 hash = ctx->hash_map;
7923 hash = kzalloc(sizeof(*hash), GFP_KERNEL);
7925 return ERR_PTR(-ENOMEM);
7926 refcount_set(&hash->refs, 1);
7927 init_waitqueue_head(&hash->wait);
7928 ctx->hash_map = hash;
7933 data.free_work = io_free_work;
7934 data.do_work = io_wq_submit_work;
7936 /* Do QD, or 4 * CPUS, whatever is smallest */
7937 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7939 return io_wq_create(concurrency, &data);
7942 static int io_uring_alloc_task_context(struct task_struct *task,
7943 struct io_ring_ctx *ctx)
7945 struct io_uring_task *tctx;
7948 tctx = kzalloc(sizeof(*tctx), GFP_KERNEL);
7949 if (unlikely(!tctx))
7952 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
7953 if (unlikely(ret)) {
7958 tctx->io_wq = io_init_wq_offload(ctx, task);
7959 if (IS_ERR(tctx->io_wq)) {
7960 ret = PTR_ERR(tctx->io_wq);
7961 percpu_counter_destroy(&tctx->inflight);
7967 init_waitqueue_head(&tctx->wait);
7968 atomic_set(&tctx->in_idle, 0);
7969 atomic_set(&tctx->inflight_tracked, 0);
7970 task->io_uring = tctx;
7971 spin_lock_init(&tctx->task_lock);
7972 INIT_WQ_LIST(&tctx->task_list);
7973 init_task_work(&tctx->task_work, tctx_task_work);
7977 void __io_uring_free(struct task_struct *tsk)
7979 struct io_uring_task *tctx = tsk->io_uring;
7981 WARN_ON_ONCE(!xa_empty(&tctx->xa));
7982 WARN_ON_ONCE(tctx->io_wq);
7983 WARN_ON_ONCE(tctx->cached_refs);
7985 percpu_counter_destroy(&tctx->inflight);
7987 tsk->io_uring = NULL;
7990 static int io_sq_offload_create(struct io_ring_ctx *ctx,
7991 struct io_uring_params *p)
7995 /* Retain compatibility with failing for an invalid attach attempt */
7996 if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
7997 IORING_SETUP_ATTACH_WQ) {
8000 f = fdget(p->wq_fd);
8004 if (f.file->f_op != &io_uring_fops)
8007 if (ctx->flags & IORING_SETUP_SQPOLL) {
8008 struct task_struct *tsk;
8009 struct io_sq_data *sqd;
8012 sqd = io_get_sq_data(p, &attached);
8018 ctx->sq_creds = get_current_cred();
8020 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
8021 if (!ctx->sq_thread_idle)
8022 ctx->sq_thread_idle = HZ;
8024 io_sq_thread_park(sqd);
8025 list_add(&ctx->sqd_list, &sqd->ctx_list);
8026 io_sqd_update_thread_idle(sqd);
8027 /* don't attach to a dying SQPOLL thread, would be racy */
8028 ret = (attached && !sqd->thread) ? -ENXIO : 0;
8029 io_sq_thread_unpark(sqd);
8036 if (p->flags & IORING_SETUP_SQ_AFF) {
8037 int cpu = p->sq_thread_cpu;
8040 if (cpu >= nr_cpu_ids || !cpu_online(cpu))
8047 sqd->task_pid = current->pid;
8048 sqd->task_tgid = current->tgid;
8049 tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE);
8056 ret = io_uring_alloc_task_context(tsk, ctx);
8057 wake_up_new_task(tsk);
8060 } else if (p->flags & IORING_SETUP_SQ_AFF) {
8061 /* Can't have SQ_AFF without SQPOLL */
8068 complete(&ctx->sq_data->exited);
8070 io_sq_thread_finish(ctx);
8074 static inline void __io_unaccount_mem(struct user_struct *user,
8075 unsigned long nr_pages)
8077 atomic_long_sub(nr_pages, &user->locked_vm);
8080 static inline int __io_account_mem(struct user_struct *user,
8081 unsigned long nr_pages)
8083 unsigned long page_limit, cur_pages, new_pages;
8085 /* Don't allow more pages than we can safely lock */
8086 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
8089 cur_pages = atomic_long_read(&user->locked_vm);
8090 new_pages = cur_pages + nr_pages;
8091 if (new_pages > page_limit)
8093 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
8094 new_pages) != cur_pages);
8099 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8102 __io_unaccount_mem(ctx->user, nr_pages);
8104 if (ctx->mm_account)
8105 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
8108 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8113 ret = __io_account_mem(ctx->user, nr_pages);
8118 if (ctx->mm_account)
8119 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
8124 static void io_mem_free(void *ptr)
8131 page = virt_to_head_page(ptr);
8132 if (put_page_testzero(page))
8133 free_compound_page(page);
8136 static void *io_mem_alloc(size_t size)
8138 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
8139 __GFP_NORETRY | __GFP_ACCOUNT;
8141 return (void *) __get_free_pages(gfp_flags, get_order(size));
8144 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8147 struct io_rings *rings;
8148 size_t off, sq_array_size;
8150 off = struct_size(rings, cqes, cq_entries);
8151 if (off == SIZE_MAX)
8155 off = ALIGN(off, SMP_CACHE_BYTES);
8163 sq_array_size = array_size(sizeof(u32), sq_entries);
8164 if (sq_array_size == SIZE_MAX)
8167 if (check_add_overflow(off, sq_array_size, &off))
8173 static void io_buffer_unmap(struct io_ring_ctx *ctx, struct io_mapped_ubuf **slot)
8175 struct io_mapped_ubuf *imu = *slot;
8178 if (imu != ctx->dummy_ubuf) {
8179 for (i = 0; i < imu->nr_bvecs; i++)
8180 unpin_user_page(imu->bvec[i].bv_page);
8181 if (imu->acct_pages)
8182 io_unaccount_mem(ctx, imu->acct_pages);
8188 static void io_rsrc_buf_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
8190 io_buffer_unmap(ctx, &prsrc->buf);
8194 static void __io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8198 for (i = 0; i < ctx->nr_user_bufs; i++)
8199 io_buffer_unmap(ctx, &ctx->user_bufs[i]);
8200 kfree(ctx->user_bufs);
8201 io_rsrc_data_free(ctx->buf_data);
8202 ctx->user_bufs = NULL;
8203 ctx->buf_data = NULL;
8204 ctx->nr_user_bufs = 0;
8207 static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8214 ret = io_rsrc_ref_quiesce(ctx->buf_data, ctx);
8216 __io_sqe_buffers_unregister(ctx);
8220 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8221 void __user *arg, unsigned index)
8223 struct iovec __user *src;
8225 #ifdef CONFIG_COMPAT
8227 struct compat_iovec __user *ciovs;
8228 struct compat_iovec ciov;
8230 ciovs = (struct compat_iovec __user *) arg;
8231 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8234 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8235 dst->iov_len = ciov.iov_len;
8239 src = (struct iovec __user *) arg;
8240 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8246 * Not super efficient, but this is just a registration time. And we do cache
8247 * the last compound head, so generally we'll only do a full search if we don't
8250 * We check if the given compound head page has already been accounted, to
8251 * avoid double accounting it. This allows us to account the full size of the
8252 * page, not just the constituent pages of a huge page.
8254 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8255 int nr_pages, struct page *hpage)
8259 /* check current page array */
8260 for (i = 0; i < nr_pages; i++) {
8261 if (!PageCompound(pages[i]))
8263 if (compound_head(pages[i]) == hpage)
8267 /* check previously registered pages */
8268 for (i = 0; i < ctx->nr_user_bufs; i++) {
8269 struct io_mapped_ubuf *imu = ctx->user_bufs[i];
8271 for (j = 0; j < imu->nr_bvecs; j++) {
8272 if (!PageCompound(imu->bvec[j].bv_page))
8274 if (compound_head(imu->bvec[j].bv_page) == hpage)
8282 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8283 int nr_pages, struct io_mapped_ubuf *imu,
8284 struct page **last_hpage)
8288 imu->acct_pages = 0;
8289 for (i = 0; i < nr_pages; i++) {
8290 if (!PageCompound(pages[i])) {
8295 hpage = compound_head(pages[i]);
8296 if (hpage == *last_hpage)
8298 *last_hpage = hpage;
8299 if (headpage_already_acct(ctx, pages, i, hpage))
8301 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8305 if (!imu->acct_pages)
8308 ret = io_account_mem(ctx, imu->acct_pages);
8310 imu->acct_pages = 0;
8314 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
8315 struct io_mapped_ubuf **pimu,
8316 struct page **last_hpage)
8318 struct io_mapped_ubuf *imu = NULL;
8319 struct vm_area_struct **vmas = NULL;
8320 struct page **pages = NULL;
8321 unsigned long off, start, end, ubuf;
8323 int ret, pret, nr_pages, i;
8325 if (!iov->iov_base) {
8326 *pimu = ctx->dummy_ubuf;
8330 ubuf = (unsigned long) iov->iov_base;
8331 end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8332 start = ubuf >> PAGE_SHIFT;
8333 nr_pages = end - start;
8338 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
8342 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
8347 imu = kvmalloc(struct_size(imu, bvec, nr_pages), GFP_KERNEL);
8352 mmap_read_lock(current->mm);
8353 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
8355 if (pret == nr_pages) {
8356 /* don't support file backed memory */
8357 for (i = 0; i < nr_pages; i++) {
8358 struct vm_area_struct *vma = vmas[i];
8360 if (vma_is_shmem(vma))
8363 !is_file_hugepages(vma->vm_file)) {
8369 ret = pret < 0 ? pret : -EFAULT;
8371 mmap_read_unlock(current->mm);
8374 * if we did partial map, or found file backed vmas,
8375 * release any pages we did get
8378 unpin_user_pages(pages, pret);
8382 ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
8384 unpin_user_pages(pages, pret);
8388 off = ubuf & ~PAGE_MASK;
8389 size = iov->iov_len;
8390 for (i = 0; i < nr_pages; i++) {
8393 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8394 imu->bvec[i].bv_page = pages[i];
8395 imu->bvec[i].bv_len = vec_len;
8396 imu->bvec[i].bv_offset = off;
8400 /* store original address for later verification */
8402 imu->ubuf_end = ubuf + iov->iov_len;
8403 imu->nr_bvecs = nr_pages;
8414 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
8416 ctx->user_bufs = kcalloc(nr_args, sizeof(*ctx->user_bufs), GFP_KERNEL);
8417 return ctx->user_bufs ? 0 : -ENOMEM;
8420 static int io_buffer_validate(struct iovec *iov)
8422 unsigned long tmp, acct_len = iov->iov_len + (PAGE_SIZE - 1);
8425 * Don't impose further limits on the size and buffer
8426 * constraints here, we'll -EINVAL later when IO is
8427 * submitted if they are wrong.
8430 return iov->iov_len ? -EFAULT : 0;
8434 /* arbitrary limit, but we need something */
8435 if (iov->iov_len > SZ_1G)
8438 if (check_add_overflow((unsigned long)iov->iov_base, acct_len, &tmp))
8444 static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
8445 unsigned int nr_args, u64 __user *tags)
8447 struct page *last_hpage = NULL;
8448 struct io_rsrc_data *data;
8454 if (!nr_args || nr_args > IORING_MAX_REG_BUFFERS)
8456 ret = io_rsrc_node_switch_start(ctx);
8459 ret = io_rsrc_data_alloc(ctx, io_rsrc_buf_put, tags, nr_args, &data);
8462 ret = io_buffers_map_alloc(ctx, nr_args);
8464 io_rsrc_data_free(data);
8468 for (i = 0; i < nr_args; i++, ctx->nr_user_bufs++) {
8469 ret = io_copy_iov(ctx, &iov, arg, i);
8472 ret = io_buffer_validate(&iov);
8475 if (!iov.iov_base && *io_get_tag_slot(data, i)) {
8480 ret = io_sqe_buffer_register(ctx, &iov, &ctx->user_bufs[i],
8486 WARN_ON_ONCE(ctx->buf_data);
8488 ctx->buf_data = data;
8490 __io_sqe_buffers_unregister(ctx);
8492 io_rsrc_node_switch(ctx, NULL);
8496 static int __io_sqe_buffers_update(struct io_ring_ctx *ctx,
8497 struct io_uring_rsrc_update2 *up,
8498 unsigned int nr_args)
8500 u64 __user *tags = u64_to_user_ptr(up->tags);
8501 struct iovec iov, __user *iovs = u64_to_user_ptr(up->data);
8502 struct page *last_hpage = NULL;
8503 bool needs_switch = false;
8509 if (up->offset + nr_args > ctx->nr_user_bufs)
8512 for (done = 0; done < nr_args; done++) {
8513 struct io_mapped_ubuf *imu;
8514 int offset = up->offset + done;
8517 err = io_copy_iov(ctx, &iov, iovs, done);
8520 if (tags && copy_from_user(&tag, &tags[done], sizeof(tag))) {
8524 err = io_buffer_validate(&iov);
8527 if (!iov.iov_base && tag) {
8531 err = io_sqe_buffer_register(ctx, &iov, &imu, &last_hpage);
8535 i = array_index_nospec(offset, ctx->nr_user_bufs);
8536 if (ctx->user_bufs[i] != ctx->dummy_ubuf) {
8537 err = io_queue_rsrc_removal(ctx->buf_data, offset,
8538 ctx->rsrc_node, ctx->user_bufs[i]);
8539 if (unlikely(err)) {
8540 io_buffer_unmap(ctx, &imu);
8543 ctx->user_bufs[i] = NULL;
8544 needs_switch = true;
8547 ctx->user_bufs[i] = imu;
8548 *io_get_tag_slot(ctx->buf_data, offset) = tag;
8552 io_rsrc_node_switch(ctx, ctx->buf_data);
8553 return done ? done : err;
8556 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8558 __s32 __user *fds = arg;
8564 if (copy_from_user(&fd, fds, sizeof(*fds)))
8567 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8568 if (IS_ERR(ctx->cq_ev_fd)) {
8569 int ret = PTR_ERR(ctx->cq_ev_fd);
8571 ctx->cq_ev_fd = NULL;
8578 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8580 if (ctx->cq_ev_fd) {
8581 eventfd_ctx_put(ctx->cq_ev_fd);
8582 ctx->cq_ev_fd = NULL;
8589 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8591 struct io_buffer *buf;
8592 unsigned long index;
8594 xa_for_each(&ctx->io_buffers, index, buf)
8595 __io_remove_buffers(ctx, buf, index, -1U);
8598 static void io_req_cache_free(struct list_head *list, struct task_struct *tsk)
8600 struct io_kiocb *req, *nxt;
8602 list_for_each_entry_safe(req, nxt, list, compl.list) {
8603 if (tsk && req->task != tsk)
8605 list_del(&req->compl.list);
8606 kmem_cache_free(req_cachep, req);
8610 static void io_req_caches_free(struct io_ring_ctx *ctx)
8612 struct io_submit_state *submit_state = &ctx->submit_state;
8613 struct io_comp_state *cs = &ctx->submit_state.comp;
8615 mutex_lock(&ctx->uring_lock);
8617 if (submit_state->free_reqs) {
8618 kmem_cache_free_bulk(req_cachep, submit_state->free_reqs,
8619 submit_state->reqs);
8620 submit_state->free_reqs = 0;
8623 io_flush_cached_locked_reqs(ctx, cs);
8624 io_req_cache_free(&cs->free_list, NULL);
8625 mutex_unlock(&ctx->uring_lock);
8628 static bool io_wait_rsrc_data(struct io_rsrc_data *data)
8632 if (!atomic_dec_and_test(&data->refs))
8633 wait_for_completion(&data->done);
8637 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8639 io_sq_thread_finish(ctx);
8641 if (ctx->mm_account) {
8642 mmdrop(ctx->mm_account);
8643 ctx->mm_account = NULL;
8646 mutex_lock(&ctx->uring_lock);
8647 if (io_wait_rsrc_data(ctx->buf_data))
8648 __io_sqe_buffers_unregister(ctx);
8649 if (io_wait_rsrc_data(ctx->file_data))
8650 __io_sqe_files_unregister(ctx);
8652 __io_cqring_overflow_flush(ctx, true);
8653 mutex_unlock(&ctx->uring_lock);
8654 io_eventfd_unregister(ctx);
8655 io_destroy_buffers(ctx);
8657 put_cred(ctx->sq_creds);
8659 /* there are no registered resources left, nobody uses it */
8661 io_rsrc_node_destroy(ctx->rsrc_node);
8662 if (ctx->rsrc_backup_node)
8663 io_rsrc_node_destroy(ctx->rsrc_backup_node);
8664 flush_delayed_work(&ctx->rsrc_put_work);
8666 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
8667 WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist));
8669 #if defined(CONFIG_UNIX)
8670 if (ctx->ring_sock) {
8671 ctx->ring_sock->file = NULL; /* so that iput() is called */
8672 sock_release(ctx->ring_sock);
8676 io_mem_free(ctx->rings);
8677 io_mem_free(ctx->sq_sqes);
8679 percpu_ref_exit(&ctx->refs);
8680 free_uid(ctx->user);
8681 io_req_caches_free(ctx);
8683 io_wq_put_hash(ctx->hash_map);
8684 kfree(ctx->cancel_hash);
8685 kfree(ctx->dummy_ubuf);
8689 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8691 struct io_ring_ctx *ctx = file->private_data;
8694 poll_wait(file, &ctx->poll_wait, wait);
8696 * synchronizes with barrier from wq_has_sleeper call in
8700 if (!io_sqring_full(ctx))
8701 mask |= EPOLLOUT | EPOLLWRNORM;
8704 * Don't flush cqring overflow list here, just do a simple check.
8705 * Otherwise there could possible be ABBA deadlock:
8708 * lock(&ctx->uring_lock);
8710 * lock(&ctx->uring_lock);
8713 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
8714 * pushs them to do the flush.
8716 if (io_cqring_events(ctx) || test_bit(0, &ctx->check_cq_overflow))
8717 mask |= EPOLLIN | EPOLLRDNORM;
8722 static int io_uring_fasync(int fd, struct file *file, int on)
8724 struct io_ring_ctx *ctx = file->private_data;
8726 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8729 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8731 const struct cred *creds;
8733 creds = xa_erase(&ctx->personalities, id);
8742 static inline bool io_run_ctx_fallback(struct io_ring_ctx *ctx)
8744 return io_run_task_work_head(&ctx->exit_task_work);
8747 struct io_tctx_exit {
8748 struct callback_head task_work;
8749 struct completion completion;
8750 struct io_ring_ctx *ctx;
8753 static void io_tctx_exit_cb(struct callback_head *cb)
8755 struct io_uring_task *tctx = current->io_uring;
8756 struct io_tctx_exit *work;
8758 work = container_of(cb, struct io_tctx_exit, task_work);
8760 * When @in_idle, we're in cancellation and it's racy to remove the
8761 * node. It'll be removed by the end of cancellation, just ignore it.
8763 if (!atomic_read(&tctx->in_idle))
8764 io_uring_del_tctx_node((unsigned long)work->ctx);
8765 complete(&work->completion);
8768 static bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
8770 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8772 return req->ctx == data;
8775 static void io_ring_exit_work(struct work_struct *work)
8777 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
8778 unsigned long timeout = jiffies + HZ * 60 * 5;
8779 struct io_tctx_exit exit;
8780 struct io_tctx_node *node;
8784 * If we're doing polled IO and end up having requests being
8785 * submitted async (out-of-line), then completions can come in while
8786 * we're waiting for refs to drop. We need to reap these manually,
8787 * as nobody else will be looking for them.
8790 io_uring_try_cancel_requests(ctx, NULL, true);
8792 struct io_sq_data *sqd = ctx->sq_data;
8793 struct task_struct *tsk;
8795 io_sq_thread_park(sqd);
8797 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
8798 io_wq_cancel_cb(tsk->io_uring->io_wq,
8799 io_cancel_ctx_cb, ctx, true);
8800 io_sq_thread_unpark(sqd);
8803 WARN_ON_ONCE(time_after(jiffies, timeout));
8804 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
8806 init_completion(&exit.completion);
8807 init_task_work(&exit.task_work, io_tctx_exit_cb);
8810 * Some may use context even when all refs and requests have been put,
8811 * and they are free to do so while still holding uring_lock or
8812 * completion_lock, see __io_req_task_submit(). Apart from other work,
8813 * this lock/unlock section also waits them to finish.
8815 mutex_lock(&ctx->uring_lock);
8816 while (!list_empty(&ctx->tctx_list)) {
8817 WARN_ON_ONCE(time_after(jiffies, timeout));
8819 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
8821 /* don't spin on a single task if cancellation failed */
8822 list_rotate_left(&ctx->tctx_list);
8823 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
8824 if (WARN_ON_ONCE(ret))
8826 wake_up_process(node->task);
8828 mutex_unlock(&ctx->uring_lock);
8829 wait_for_completion(&exit.completion);
8830 mutex_lock(&ctx->uring_lock);
8832 mutex_unlock(&ctx->uring_lock);
8833 spin_lock_irq(&ctx->completion_lock);
8834 spin_unlock_irq(&ctx->completion_lock);
8836 io_ring_ctx_free(ctx);
8839 /* Returns true if we found and killed one or more timeouts */
8840 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
8843 struct io_kiocb *req, *tmp;
8846 spin_lock_irq(&ctx->completion_lock);
8847 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
8848 if (io_match_task(req, tsk, cancel_all)) {
8849 io_kill_timeout(req, -ECANCELED);
8854 io_commit_cqring(ctx);
8855 spin_unlock_irq(&ctx->completion_lock);
8857 io_cqring_ev_posted(ctx);
8858 return canceled != 0;
8861 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8863 unsigned long index;
8864 struct creds *creds;
8866 mutex_lock(&ctx->uring_lock);
8867 percpu_ref_kill(&ctx->refs);
8869 __io_cqring_overflow_flush(ctx, true);
8870 xa_for_each(&ctx->personalities, index, creds)
8871 io_unregister_personality(ctx, index);
8872 mutex_unlock(&ctx->uring_lock);
8874 io_kill_timeouts(ctx, NULL, true);
8875 io_poll_remove_all(ctx, NULL, true);
8877 /* if we failed setting up the ctx, we might not have any rings */
8878 io_iopoll_try_reap_events(ctx);
8880 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8882 * Use system_unbound_wq to avoid spawning tons of event kworkers
8883 * if we're exiting a ton of rings at the same time. It just adds
8884 * noise and overhead, there's no discernable change in runtime
8885 * over using system_wq.
8887 queue_work(system_unbound_wq, &ctx->exit_work);
8890 static int io_uring_release(struct inode *inode, struct file *file)
8892 struct io_ring_ctx *ctx = file->private_data;
8894 file->private_data = NULL;
8895 io_ring_ctx_wait_and_kill(ctx);
8899 struct io_task_cancel {
8900 struct task_struct *task;
8904 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8906 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8907 struct io_task_cancel *cancel = data;
8910 if (!cancel->all && (req->flags & REQ_F_LINK_TIMEOUT)) {
8911 unsigned long flags;
8912 struct io_ring_ctx *ctx = req->ctx;
8914 /* protect against races with linked timeouts */
8915 spin_lock_irqsave(&ctx->completion_lock, flags);
8916 ret = io_match_task(req, cancel->task, cancel->all);
8917 spin_unlock_irqrestore(&ctx->completion_lock, flags);
8919 ret = io_match_task(req, cancel->task, cancel->all);
8924 static bool io_cancel_defer_files(struct io_ring_ctx *ctx,
8925 struct task_struct *task, bool cancel_all)
8927 struct io_defer_entry *de;
8930 spin_lock_irq(&ctx->completion_lock);
8931 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8932 if (io_match_task(de->req, task, cancel_all)) {
8933 list_cut_position(&list, &ctx->defer_list, &de->list);
8937 spin_unlock_irq(&ctx->completion_lock);
8938 if (list_empty(&list))
8941 while (!list_empty(&list)) {
8942 de = list_first_entry(&list, struct io_defer_entry, list);
8943 list_del_init(&de->list);
8944 io_req_complete_failed(de->req, -ECANCELED);
8950 static bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
8952 struct io_tctx_node *node;
8953 enum io_wq_cancel cret;
8956 mutex_lock(&ctx->uring_lock);
8957 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
8958 struct io_uring_task *tctx = node->task->io_uring;
8961 * io_wq will stay alive while we hold uring_lock, because it's
8962 * killed after ctx nodes, which requires to take the lock.
8964 if (!tctx || !tctx->io_wq)
8966 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
8967 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8969 mutex_unlock(&ctx->uring_lock);
8974 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
8975 struct task_struct *task,
8978 struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
8979 struct io_uring_task *tctx = task ? task->io_uring : NULL;
8982 enum io_wq_cancel cret;
8986 ret |= io_uring_try_cancel_iowq(ctx);
8987 } else if (tctx && tctx->io_wq) {
8989 * Cancels requests of all rings, not only @ctx, but
8990 * it's fine as the task is in exit/exec.
8992 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
8994 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8997 /* SQPOLL thread does its own polling */
8998 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
8999 (ctx->sq_data && ctx->sq_data->thread == current)) {
9000 while (!list_empty_careful(&ctx->iopoll_list)) {
9001 io_iopoll_try_reap_events(ctx);
9006 ret |= io_cancel_defer_files(ctx, task, cancel_all);
9007 ret |= io_poll_remove_all(ctx, task, cancel_all);
9008 ret |= io_kill_timeouts(ctx, task, cancel_all);
9009 ret |= io_run_task_work();
9010 ret |= io_run_ctx_fallback(ctx);
9017 static int __io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9019 struct io_uring_task *tctx = current->io_uring;
9020 struct io_tctx_node *node;
9023 if (unlikely(!tctx)) {
9024 ret = io_uring_alloc_task_context(current, ctx);
9027 tctx = current->io_uring;
9029 if (!xa_load(&tctx->xa, (unsigned long)ctx)) {
9030 node = kmalloc(sizeof(*node), GFP_KERNEL);
9034 node->task = current;
9036 ret = xa_err(xa_store(&tctx->xa, (unsigned long)ctx,
9043 mutex_lock(&ctx->uring_lock);
9044 list_add(&node->ctx_node, &ctx->tctx_list);
9045 mutex_unlock(&ctx->uring_lock);
9052 * Note that this task has used io_uring. We use it for cancelation purposes.
9054 static inline int io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9056 struct io_uring_task *tctx = current->io_uring;
9058 if (likely(tctx && tctx->last == ctx))
9060 return __io_uring_add_tctx_node(ctx);
9064 * Remove this io_uring_file -> task mapping.
9066 static void io_uring_del_tctx_node(unsigned long index)
9068 struct io_uring_task *tctx = current->io_uring;
9069 struct io_tctx_node *node;
9073 node = xa_erase(&tctx->xa, index);
9077 WARN_ON_ONCE(current != node->task);
9078 WARN_ON_ONCE(list_empty(&node->ctx_node));
9080 mutex_lock(&node->ctx->uring_lock);
9081 list_del(&node->ctx_node);
9082 mutex_unlock(&node->ctx->uring_lock);
9084 if (tctx->last == node->ctx)
9089 static void io_uring_clean_tctx(struct io_uring_task *tctx)
9091 struct io_wq *wq = tctx->io_wq;
9092 struct io_tctx_node *node;
9093 unsigned long index;
9095 xa_for_each(&tctx->xa, index, node)
9096 io_uring_del_tctx_node(index);
9099 * Must be after io_uring_del_task_file() (removes nodes under
9100 * uring_lock) to avoid race with io_uring_try_cancel_iowq().
9103 io_wq_put_and_exit(wq);
9107 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
9110 return atomic_read(&tctx->inflight_tracked);
9111 return percpu_counter_sum(&tctx->inflight);
9114 static void io_uring_drop_tctx_refs(struct task_struct *task)
9116 struct io_uring_task *tctx = task->io_uring;
9117 unsigned int refs = tctx->cached_refs;
9119 tctx->cached_refs = 0;
9120 percpu_counter_sub(&tctx->inflight, refs);
9121 put_task_struct_many(task, refs);
9125 * Find any io_uring ctx that this task has registered or done IO on, and cancel
9126 * requests. @sqd should be not-null IIF it's an SQPOLL thread cancellation.
9128 static void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
9130 struct io_uring_task *tctx = current->io_uring;
9131 struct io_ring_ctx *ctx;
9135 WARN_ON_ONCE(sqd && sqd->thread != current);
9137 if (!current->io_uring)
9140 io_wq_exit_start(tctx->io_wq);
9142 io_uring_drop_tctx_refs(current);
9143 atomic_inc(&tctx->in_idle);
9145 /* read completions before cancelations */
9146 inflight = tctx_inflight(tctx, !cancel_all);
9151 struct io_tctx_node *node;
9152 unsigned long index;
9154 xa_for_each(&tctx->xa, index, node) {
9155 /* sqpoll task will cancel all its requests */
9156 if (node->ctx->sq_data)
9158 io_uring_try_cancel_requests(node->ctx, current,
9162 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
9163 io_uring_try_cancel_requests(ctx, current,
9167 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
9169 * If we've seen completions, retry without waiting. This
9170 * avoids a race where a completion comes in before we did
9171 * prepare_to_wait().
9173 if (inflight == tctx_inflight(tctx, !cancel_all))
9175 finish_wait(&tctx->wait, &wait);
9177 atomic_dec(&tctx->in_idle);
9179 io_uring_clean_tctx(tctx);
9181 /* for exec all current's requests should be gone, kill tctx */
9182 __io_uring_free(current);
9186 void __io_uring_cancel(struct files_struct *files)
9188 io_uring_cancel_generic(!files, NULL);
9191 static void *io_uring_validate_mmap_request(struct file *file,
9192 loff_t pgoff, size_t sz)
9194 struct io_ring_ctx *ctx = file->private_data;
9195 loff_t offset = pgoff << PAGE_SHIFT;
9200 case IORING_OFF_SQ_RING:
9201 case IORING_OFF_CQ_RING:
9204 case IORING_OFF_SQES:
9208 return ERR_PTR(-EINVAL);
9211 page = virt_to_head_page(ptr);
9212 if (sz > page_size(page))
9213 return ERR_PTR(-EINVAL);
9220 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9222 size_t sz = vma->vm_end - vma->vm_start;
9226 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
9228 return PTR_ERR(ptr);
9230 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
9231 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
9234 #else /* !CONFIG_MMU */
9236 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9238 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
9241 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
9243 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
9246 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
9247 unsigned long addr, unsigned long len,
9248 unsigned long pgoff, unsigned long flags)
9252 ptr = io_uring_validate_mmap_request(file, pgoff, len);
9254 return PTR_ERR(ptr);
9256 return (unsigned long) ptr;
9259 #endif /* !CONFIG_MMU */
9261 static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
9266 if (!io_sqring_full(ctx))
9268 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9270 if (!io_sqring_full(ctx))
9273 } while (!signal_pending(current));
9275 finish_wait(&ctx->sqo_sq_wait, &wait);
9279 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9280 struct __kernel_timespec __user **ts,
9281 const sigset_t __user **sig)
9283 struct io_uring_getevents_arg arg;
9286 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9287 * is just a pointer to the sigset_t.
9289 if (!(flags & IORING_ENTER_EXT_ARG)) {
9290 *sig = (const sigset_t __user *) argp;
9296 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9297 * timespec and sigset_t pointers if good.
9299 if (*argsz != sizeof(arg))
9301 if (copy_from_user(&arg, argp, sizeof(arg)))
9303 *sig = u64_to_user_ptr(arg.sigmask);
9304 *argsz = arg.sigmask_sz;
9305 *ts = u64_to_user_ptr(arg.ts);
9309 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9310 u32, min_complete, u32, flags, const void __user *, argp,
9313 struct io_ring_ctx *ctx;
9320 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9321 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG)))
9325 if (unlikely(!f.file))
9329 if (unlikely(f.file->f_op != &io_uring_fops))
9333 ctx = f.file->private_data;
9334 if (unlikely(!percpu_ref_tryget(&ctx->refs)))
9338 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
9342 * For SQ polling, the thread will do all submissions and completions.
9343 * Just return the requested submit count, and wake the thread if
9347 if (ctx->flags & IORING_SETUP_SQPOLL) {
9348 io_cqring_overflow_flush(ctx, false);
9351 if (unlikely(ctx->sq_data->thread == NULL))
9353 if (flags & IORING_ENTER_SQ_WAKEUP)
9354 wake_up(&ctx->sq_data->wait);
9355 if (flags & IORING_ENTER_SQ_WAIT) {
9356 ret = io_sqpoll_wait_sq(ctx);
9360 submitted = to_submit;
9361 } else if (to_submit) {
9362 ret = io_uring_add_tctx_node(ctx);
9365 mutex_lock(&ctx->uring_lock);
9366 submitted = io_submit_sqes(ctx, to_submit);
9367 mutex_unlock(&ctx->uring_lock);
9369 if (submitted != to_submit)
9372 if (flags & IORING_ENTER_GETEVENTS) {
9373 const sigset_t __user *sig;
9374 struct __kernel_timespec __user *ts;
9376 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
9380 min_complete = min(min_complete, ctx->cq_entries);
9383 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9384 * space applications don't need to do io completion events
9385 * polling again, they can rely on io_sq_thread to do polling
9386 * work, which can reduce cpu usage and uring_lock contention.
9388 if (ctx->flags & IORING_SETUP_IOPOLL &&
9389 !(ctx->flags & IORING_SETUP_SQPOLL)) {
9390 ret = io_iopoll_check(ctx, min_complete);
9392 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
9397 percpu_ref_put(&ctx->refs);
9400 return submitted ? submitted : ret;
9403 #ifdef CONFIG_PROC_FS
9404 static int io_uring_show_cred(struct seq_file *m, unsigned int id,
9405 const struct cred *cred)
9407 struct user_namespace *uns = seq_user_ns(m);
9408 struct group_info *gi;
9413 seq_printf(m, "%5d\n", id);
9414 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
9415 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
9416 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
9417 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
9418 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
9419 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
9420 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
9421 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
9422 seq_puts(m, "\n\tGroups:\t");
9423 gi = cred->group_info;
9424 for (g = 0; g < gi->ngroups; g++) {
9425 seq_put_decimal_ull(m, g ? " " : "",
9426 from_kgid_munged(uns, gi->gid[g]));
9428 seq_puts(m, "\n\tCapEff:\t");
9429 cap = cred->cap_effective;
9430 CAP_FOR_EACH_U32(__capi)
9431 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
9436 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
9438 struct io_sq_data *sq = NULL;
9443 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9444 * since fdinfo case grabs it in the opposite direction of normal use
9445 * cases. If we fail to get the lock, we just don't iterate any
9446 * structures that could be going away outside the io_uring mutex.
9448 has_lock = mutex_trylock(&ctx->uring_lock);
9450 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) {
9456 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9457 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9458 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9459 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9460 struct file *f = io_file_from_index(ctx, i);
9463 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9465 seq_printf(m, "%5u: <none>\n", i);
9467 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9468 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9469 struct io_mapped_ubuf *buf = ctx->user_bufs[i];
9470 unsigned int len = buf->ubuf_end - buf->ubuf;
9472 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf, len);
9474 if (has_lock && !xa_empty(&ctx->personalities)) {
9475 unsigned long index;
9476 const struct cred *cred;
9478 seq_printf(m, "Personalities:\n");
9479 xa_for_each(&ctx->personalities, index, cred)
9480 io_uring_show_cred(m, index, cred);
9482 seq_printf(m, "PollList:\n");
9483 spin_lock_irq(&ctx->completion_lock);
9484 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9485 struct hlist_head *list = &ctx->cancel_hash[i];
9486 struct io_kiocb *req;
9488 hlist_for_each_entry(req, list, hash_node)
9489 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9490 req->task->task_works != NULL);
9492 spin_unlock_irq(&ctx->completion_lock);
9494 mutex_unlock(&ctx->uring_lock);
9497 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9499 struct io_ring_ctx *ctx = f->private_data;
9501 if (percpu_ref_tryget(&ctx->refs)) {
9502 __io_uring_show_fdinfo(ctx, m);
9503 percpu_ref_put(&ctx->refs);
9508 static const struct file_operations io_uring_fops = {
9509 .release = io_uring_release,
9510 .mmap = io_uring_mmap,
9512 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9513 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9515 .poll = io_uring_poll,
9516 .fasync = io_uring_fasync,
9517 #ifdef CONFIG_PROC_FS
9518 .show_fdinfo = io_uring_show_fdinfo,
9522 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9523 struct io_uring_params *p)
9525 struct io_rings *rings;
9526 size_t size, sq_array_offset;
9528 /* make sure these are sane, as we already accounted them */
9529 ctx->sq_entries = p->sq_entries;
9530 ctx->cq_entries = p->cq_entries;
9532 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9533 if (size == SIZE_MAX)
9536 rings = io_mem_alloc(size);
9541 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9542 rings->sq_ring_mask = p->sq_entries - 1;
9543 rings->cq_ring_mask = p->cq_entries - 1;
9544 rings->sq_ring_entries = p->sq_entries;
9545 rings->cq_ring_entries = p->cq_entries;
9547 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9548 if (size == SIZE_MAX) {
9549 io_mem_free(ctx->rings);
9554 ctx->sq_sqes = io_mem_alloc(size);
9555 if (!ctx->sq_sqes) {
9556 io_mem_free(ctx->rings);
9564 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
9568 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9572 ret = io_uring_add_tctx_node(ctx);
9577 fd_install(fd, file);
9582 * Allocate an anonymous fd, this is what constitutes the application
9583 * visible backing of an io_uring instance. The application mmaps this
9584 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9585 * we have to tie this fd to a socket for file garbage collection purposes.
9587 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
9590 #if defined(CONFIG_UNIX)
9593 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9596 return ERR_PTR(ret);
9599 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9600 O_RDWR | O_CLOEXEC);
9601 #if defined(CONFIG_UNIX)
9603 sock_release(ctx->ring_sock);
9604 ctx->ring_sock = NULL;
9606 ctx->ring_sock->file = file;
9612 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9613 struct io_uring_params __user *params)
9615 struct io_ring_ctx *ctx;
9621 if (entries > IORING_MAX_ENTRIES) {
9622 if (!(p->flags & IORING_SETUP_CLAMP))
9624 entries = IORING_MAX_ENTRIES;
9628 * Use twice as many entries for the CQ ring. It's possible for the
9629 * application to drive a higher depth than the size of the SQ ring,
9630 * since the sqes are only used at submission time. This allows for
9631 * some flexibility in overcommitting a bit. If the application has
9632 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9633 * of CQ ring entries manually.
9635 p->sq_entries = roundup_pow_of_two(entries);
9636 if (p->flags & IORING_SETUP_CQSIZE) {
9638 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9639 * to a power-of-two, if it isn't already. We do NOT impose
9640 * any cq vs sq ring sizing.
9644 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9645 if (!(p->flags & IORING_SETUP_CLAMP))
9647 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9649 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9650 if (p->cq_entries < p->sq_entries)
9653 p->cq_entries = 2 * p->sq_entries;
9656 ctx = io_ring_ctx_alloc(p);
9659 ctx->compat = in_compat_syscall();
9660 if (!capable(CAP_IPC_LOCK))
9661 ctx->user = get_uid(current_user());
9664 * This is just grabbed for accounting purposes. When a process exits,
9665 * the mm is exited and dropped before the files, hence we need to hang
9666 * on to this mm purely for the purposes of being able to unaccount
9667 * memory (locked/pinned vm). It's not used for anything else.
9669 mmgrab(current->mm);
9670 ctx->mm_account = current->mm;
9672 ret = io_allocate_scq_urings(ctx, p);
9676 ret = io_sq_offload_create(ctx, p);
9679 /* always set a rsrc node */
9680 ret = io_rsrc_node_switch_start(ctx);
9683 io_rsrc_node_switch(ctx, NULL);
9685 memset(&p->sq_off, 0, sizeof(p->sq_off));
9686 p->sq_off.head = offsetof(struct io_rings, sq.head);
9687 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9688 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9689 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9690 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9691 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9692 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9694 memset(&p->cq_off, 0, sizeof(p->cq_off));
9695 p->cq_off.head = offsetof(struct io_rings, cq.head);
9696 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9697 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9698 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9699 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9700 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9701 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9703 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9704 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9705 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9706 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
9707 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
9708 IORING_FEAT_RSRC_TAGS;
9710 if (copy_to_user(params, p, sizeof(*p))) {
9715 file = io_uring_get_file(ctx);
9717 ret = PTR_ERR(file);
9722 * Install ring fd as the very last thing, so we don't risk someone
9723 * having closed it before we finish setup
9725 ret = io_uring_install_fd(ctx, file);
9727 /* fput will clean it up */
9732 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9735 io_ring_ctx_wait_and_kill(ctx);
9740 * Sets up an aio uring context, and returns the fd. Applications asks for a
9741 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9742 * params structure passed in.
9744 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9746 struct io_uring_params p;
9749 if (copy_from_user(&p, params, sizeof(p)))
9751 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9756 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9757 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9758 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9759 IORING_SETUP_R_DISABLED))
9762 return io_uring_create(entries, &p, params);
9765 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9766 struct io_uring_params __user *, params)
9768 return io_uring_setup(entries, params);
9771 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9773 struct io_uring_probe *p;
9777 size = struct_size(p, ops, nr_args);
9778 if (size == SIZE_MAX)
9780 p = kzalloc(size, GFP_KERNEL);
9785 if (copy_from_user(p, arg, size))
9788 if (memchr_inv(p, 0, size))
9791 p->last_op = IORING_OP_LAST - 1;
9792 if (nr_args > IORING_OP_LAST)
9793 nr_args = IORING_OP_LAST;
9795 for (i = 0; i < nr_args; i++) {
9797 if (!io_op_defs[i].not_supported)
9798 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9803 if (copy_to_user(arg, p, size))
9810 static int io_register_personality(struct io_ring_ctx *ctx)
9812 const struct cred *creds;
9816 creds = get_current_cred();
9818 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
9819 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
9826 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9827 unsigned int nr_args)
9829 struct io_uring_restriction *res;
9833 /* Restrictions allowed only if rings started disabled */
9834 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9837 /* We allow only a single restrictions registration */
9838 if (ctx->restrictions.registered)
9841 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9844 size = array_size(nr_args, sizeof(*res));
9845 if (size == SIZE_MAX)
9848 res = memdup_user(arg, size);
9850 return PTR_ERR(res);
9854 for (i = 0; i < nr_args; i++) {
9855 switch (res[i].opcode) {
9856 case IORING_RESTRICTION_REGISTER_OP:
9857 if (res[i].register_op >= IORING_REGISTER_LAST) {
9862 __set_bit(res[i].register_op,
9863 ctx->restrictions.register_op);
9865 case IORING_RESTRICTION_SQE_OP:
9866 if (res[i].sqe_op >= IORING_OP_LAST) {
9871 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9873 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9874 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9876 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9877 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9886 /* Reset all restrictions if an error happened */
9888 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9890 ctx->restrictions.registered = true;
9896 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9898 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9901 if (ctx->restrictions.registered)
9902 ctx->restricted = 1;
9904 ctx->flags &= ~IORING_SETUP_R_DISABLED;
9905 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
9906 wake_up(&ctx->sq_data->wait);
9910 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
9911 struct io_uring_rsrc_update2 *up,
9919 if (check_add_overflow(up->offset, nr_args, &tmp))
9921 err = io_rsrc_node_switch_start(ctx);
9926 case IORING_RSRC_FILE:
9927 return __io_sqe_files_update(ctx, up, nr_args);
9928 case IORING_RSRC_BUFFER:
9929 return __io_sqe_buffers_update(ctx, up, nr_args);
9934 static int io_register_files_update(struct io_ring_ctx *ctx, void __user *arg,
9937 struct io_uring_rsrc_update2 up;
9941 memset(&up, 0, sizeof(up));
9942 if (copy_from_user(&up, arg, sizeof(struct io_uring_rsrc_update)))
9944 return __io_register_rsrc_update(ctx, IORING_RSRC_FILE, &up, nr_args);
9947 static int io_register_rsrc_update(struct io_ring_ctx *ctx, void __user *arg,
9948 unsigned size, unsigned type)
9950 struct io_uring_rsrc_update2 up;
9952 if (size != sizeof(up))
9954 if (copy_from_user(&up, arg, sizeof(up)))
9956 if (!up.nr || up.resv)
9958 return __io_register_rsrc_update(ctx, type, &up, up.nr);
9961 static int io_register_rsrc(struct io_ring_ctx *ctx, void __user *arg,
9962 unsigned int size, unsigned int type)
9964 struct io_uring_rsrc_register rr;
9966 /* keep it extendible */
9967 if (size != sizeof(rr))
9970 memset(&rr, 0, sizeof(rr));
9971 if (copy_from_user(&rr, arg, size))
9973 if (!rr.nr || rr.resv || rr.resv2)
9977 case IORING_RSRC_FILE:
9978 return io_sqe_files_register(ctx, u64_to_user_ptr(rr.data),
9979 rr.nr, u64_to_user_ptr(rr.tags));
9980 case IORING_RSRC_BUFFER:
9981 return io_sqe_buffers_register(ctx, u64_to_user_ptr(rr.data),
9982 rr.nr, u64_to_user_ptr(rr.tags));
9987 static int io_register_iowq_aff(struct io_ring_ctx *ctx, void __user *arg,
9990 struct io_uring_task *tctx = current->io_uring;
9991 cpumask_var_t new_mask;
9994 if (!tctx || !tctx->io_wq)
9997 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
10000 cpumask_clear(new_mask);
10001 if (len > cpumask_size())
10002 len = cpumask_size();
10004 if (copy_from_user(new_mask, arg, len)) {
10005 free_cpumask_var(new_mask);
10009 ret = io_wq_cpu_affinity(tctx->io_wq, new_mask);
10010 free_cpumask_var(new_mask);
10014 static int io_unregister_iowq_aff(struct io_ring_ctx *ctx)
10016 struct io_uring_task *tctx = current->io_uring;
10018 if (!tctx || !tctx->io_wq)
10021 return io_wq_cpu_affinity(tctx->io_wq, NULL);
10024 static bool io_register_op_must_quiesce(int op)
10027 case IORING_REGISTER_BUFFERS:
10028 case IORING_UNREGISTER_BUFFERS:
10029 case IORING_REGISTER_FILES:
10030 case IORING_UNREGISTER_FILES:
10031 case IORING_REGISTER_FILES_UPDATE:
10032 case IORING_REGISTER_PROBE:
10033 case IORING_REGISTER_PERSONALITY:
10034 case IORING_UNREGISTER_PERSONALITY:
10035 case IORING_REGISTER_FILES2:
10036 case IORING_REGISTER_FILES_UPDATE2:
10037 case IORING_REGISTER_BUFFERS2:
10038 case IORING_REGISTER_BUFFERS_UPDATE:
10039 case IORING_REGISTER_IOWQ_AFF:
10040 case IORING_UNREGISTER_IOWQ_AFF:
10047 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
10048 void __user *arg, unsigned nr_args)
10049 __releases(ctx->uring_lock)
10050 __acquires(ctx->uring_lock)
10055 * We're inside the ring mutex, if the ref is already dying, then
10056 * someone else killed the ctx or is already going through
10057 * io_uring_register().
10059 if (percpu_ref_is_dying(&ctx->refs))
10062 if (ctx->restricted) {
10063 if (opcode >= IORING_REGISTER_LAST)
10065 opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
10066 if (!test_bit(opcode, ctx->restrictions.register_op))
10070 if (io_register_op_must_quiesce(opcode)) {
10071 percpu_ref_kill(&ctx->refs);
10074 * Drop uring mutex before waiting for references to exit. If
10075 * another thread is currently inside io_uring_enter() it might
10076 * need to grab the uring_lock to make progress. If we hold it
10077 * here across the drain wait, then we can deadlock. It's safe
10078 * to drop the mutex here, since no new references will come in
10079 * after we've killed the percpu ref.
10081 mutex_unlock(&ctx->uring_lock);
10083 ret = wait_for_completion_interruptible(&ctx->ref_comp);
10086 ret = io_run_task_work_sig();
10090 mutex_lock(&ctx->uring_lock);
10093 io_refs_resurrect(&ctx->refs, &ctx->ref_comp);
10099 case IORING_REGISTER_BUFFERS:
10100 ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
10102 case IORING_UNREGISTER_BUFFERS:
10104 if (arg || nr_args)
10106 ret = io_sqe_buffers_unregister(ctx);
10108 case IORING_REGISTER_FILES:
10109 ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
10111 case IORING_UNREGISTER_FILES:
10113 if (arg || nr_args)
10115 ret = io_sqe_files_unregister(ctx);
10117 case IORING_REGISTER_FILES_UPDATE:
10118 ret = io_register_files_update(ctx, arg, nr_args);
10120 case IORING_REGISTER_EVENTFD:
10121 case IORING_REGISTER_EVENTFD_ASYNC:
10125 ret = io_eventfd_register(ctx, arg);
10128 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
10129 ctx->eventfd_async = 1;
10131 ctx->eventfd_async = 0;
10133 case IORING_UNREGISTER_EVENTFD:
10135 if (arg || nr_args)
10137 ret = io_eventfd_unregister(ctx);
10139 case IORING_REGISTER_PROBE:
10141 if (!arg || nr_args > 256)
10143 ret = io_probe(ctx, arg, nr_args);
10145 case IORING_REGISTER_PERSONALITY:
10147 if (arg || nr_args)
10149 ret = io_register_personality(ctx);
10151 case IORING_UNREGISTER_PERSONALITY:
10155 ret = io_unregister_personality(ctx, nr_args);
10157 case IORING_REGISTER_ENABLE_RINGS:
10159 if (arg || nr_args)
10161 ret = io_register_enable_rings(ctx);
10163 case IORING_REGISTER_RESTRICTIONS:
10164 ret = io_register_restrictions(ctx, arg, nr_args);
10166 case IORING_REGISTER_FILES2:
10167 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE);
10169 case IORING_REGISTER_FILES_UPDATE2:
10170 ret = io_register_rsrc_update(ctx, arg, nr_args,
10173 case IORING_REGISTER_BUFFERS2:
10174 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER);
10176 case IORING_REGISTER_BUFFERS_UPDATE:
10177 ret = io_register_rsrc_update(ctx, arg, nr_args,
10178 IORING_RSRC_BUFFER);
10180 case IORING_REGISTER_IOWQ_AFF:
10182 if (!arg || !nr_args)
10184 ret = io_register_iowq_aff(ctx, arg, nr_args);
10186 case IORING_UNREGISTER_IOWQ_AFF:
10188 if (arg || nr_args)
10190 ret = io_unregister_iowq_aff(ctx);
10197 if (io_register_op_must_quiesce(opcode)) {
10198 /* bring the ctx back to life */
10199 percpu_ref_reinit(&ctx->refs);
10200 reinit_completion(&ctx->ref_comp);
10205 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
10206 void __user *, arg, unsigned int, nr_args)
10208 struct io_ring_ctx *ctx;
10217 if (f.file->f_op != &io_uring_fops)
10220 ctx = f.file->private_data;
10222 io_run_task_work();
10224 mutex_lock(&ctx->uring_lock);
10225 ret = __io_uring_register(ctx, opcode, arg, nr_args);
10226 mutex_unlock(&ctx->uring_lock);
10227 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
10228 ctx->cq_ev_fd != NULL, ret);
10234 static int __init io_uring_init(void)
10236 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
10237 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
10238 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
10241 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
10242 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
10243 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
10244 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
10245 BUILD_BUG_SQE_ELEM(1, __u8, flags);
10246 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
10247 BUILD_BUG_SQE_ELEM(4, __s32, fd);
10248 BUILD_BUG_SQE_ELEM(8, __u64, off);
10249 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
10250 BUILD_BUG_SQE_ELEM(16, __u64, addr);
10251 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
10252 BUILD_BUG_SQE_ELEM(24, __u32, len);
10253 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
10254 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
10255 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
10256 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
10257 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
10258 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
10259 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
10260 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
10261 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
10262 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
10263 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
10264 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
10265 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
10266 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
10267 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
10268 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
10269 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
10270 BUILD_BUG_SQE_ELEM(42, __u16, personality);
10271 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
10273 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
10274 sizeof(struct io_uring_rsrc_update));
10275 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
10276 sizeof(struct io_uring_rsrc_update2));
10277 /* should fit into one byte */
10278 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
10280 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
10281 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
10282 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
10286 __initcall(io_uring_init);