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])
1734 * Don't initialise the fields below on every allocation, but
1735 * do that in advance and keep valid on free.
1737 for (i = 0; i < ret; i++) {
1738 struct io_kiocb *req = state->reqs[i];
1742 req->async_data = NULL;
1743 /* not necessary, but safer to zero */
1746 state->free_reqs = ret;
1750 return state->reqs[state->free_reqs];
1753 static inline void io_put_file(struct file *file)
1759 static void io_dismantle_req(struct io_kiocb *req)
1761 unsigned int flags = req->flags;
1763 if (io_req_needs_clean(req))
1765 if (!(flags & REQ_F_FIXED_FILE))
1766 io_put_file(req->file);
1767 if (req->fixed_rsrc_refs)
1768 percpu_ref_put(req->fixed_rsrc_refs);
1769 if (req->async_data) {
1770 kfree(req->async_data);
1771 req->async_data = NULL;
1775 /* must to be called somewhat shortly after putting a request */
1776 static inline void io_put_task(struct task_struct *task, int nr)
1778 struct io_uring_task *tctx = task->io_uring;
1780 percpu_counter_sub(&tctx->inflight, nr);
1781 if (unlikely(atomic_read(&tctx->in_idle)))
1782 wake_up(&tctx->wait);
1783 put_task_struct_many(task, nr);
1786 static void __io_free_req(struct io_kiocb *req)
1788 struct io_ring_ctx *ctx = req->ctx;
1790 io_dismantle_req(req);
1791 io_put_task(req->task, 1);
1793 kmem_cache_free(req_cachep, req);
1794 percpu_ref_put(&ctx->refs);
1797 static inline void io_remove_next_linked(struct io_kiocb *req)
1799 struct io_kiocb *nxt = req->link;
1801 req->link = nxt->link;
1805 static bool io_kill_linked_timeout(struct io_kiocb *req)
1806 __must_hold(&req->ctx->completion_lock)
1808 struct io_kiocb *link = req->link;
1811 * Can happen if a linked timeout fired and link had been like
1812 * req -> link t-out -> link t-out [-> ...]
1814 if (link && (link->flags & REQ_F_LTIMEOUT_ACTIVE)) {
1815 struct io_timeout_data *io = link->async_data;
1817 io_remove_next_linked(req);
1818 link->timeout.head = NULL;
1819 if (hrtimer_try_to_cancel(&io->timer) != -1) {
1820 io_cqring_fill_event(link->ctx, link->user_data,
1822 io_put_req_deferred(link, 1);
1829 static void io_fail_links(struct io_kiocb *req)
1830 __must_hold(&req->ctx->completion_lock)
1832 struct io_kiocb *nxt, *link = req->link;
1839 trace_io_uring_fail_link(req, link);
1840 io_cqring_fill_event(link->ctx, link->user_data, -ECANCELED, 0);
1841 io_put_req_deferred(link, 2);
1846 static bool io_disarm_next(struct io_kiocb *req)
1847 __must_hold(&req->ctx->completion_lock)
1849 bool posted = false;
1851 if (likely(req->flags & REQ_F_LINK_TIMEOUT))
1852 posted = io_kill_linked_timeout(req);
1853 if (unlikely((req->flags & REQ_F_FAIL) &&
1854 !(req->flags & REQ_F_HARDLINK))) {
1855 posted |= (req->link != NULL);
1861 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1863 struct io_kiocb *nxt;
1866 * If LINK is set, we have dependent requests in this chain. If we
1867 * didn't fail this request, queue the first one up, moving any other
1868 * dependencies to the next request. In case of failure, fail the rest
1871 if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL)) {
1872 struct io_ring_ctx *ctx = req->ctx;
1873 unsigned long flags;
1876 spin_lock_irqsave(&ctx->completion_lock, flags);
1877 posted = io_disarm_next(req);
1879 io_commit_cqring(req->ctx);
1880 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1882 io_cqring_ev_posted(ctx);
1889 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1891 if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK))))
1893 return __io_req_find_next(req);
1896 static void ctx_flush_and_put(struct io_ring_ctx *ctx)
1900 if (ctx->submit_state.comp.nr) {
1901 mutex_lock(&ctx->uring_lock);
1902 io_submit_flush_completions(ctx);
1903 mutex_unlock(&ctx->uring_lock);
1905 percpu_ref_put(&ctx->refs);
1908 static void tctx_task_work(struct callback_head *cb)
1910 struct io_ring_ctx *ctx = NULL;
1911 struct io_uring_task *tctx = container_of(cb, struct io_uring_task,
1915 struct io_wq_work_node *node;
1917 spin_lock_irq(&tctx->task_lock);
1918 node = tctx->task_list.first;
1919 INIT_WQ_LIST(&tctx->task_list);
1920 spin_unlock_irq(&tctx->task_lock);
1923 struct io_wq_work_node *next = node->next;
1924 struct io_kiocb *req = container_of(node, struct io_kiocb,
1927 if (req->ctx != ctx) {
1928 ctx_flush_and_put(ctx);
1930 percpu_ref_get(&ctx->refs);
1932 req->task_work.func(&req->task_work);
1935 if (wq_list_empty(&tctx->task_list)) {
1936 clear_bit(0, &tctx->task_state);
1937 if (wq_list_empty(&tctx->task_list))
1939 /* another tctx_task_work() is enqueued, yield */
1940 if (test_and_set_bit(0, &tctx->task_state))
1946 ctx_flush_and_put(ctx);
1949 static int io_req_task_work_add(struct io_kiocb *req)
1951 struct task_struct *tsk = req->task;
1952 struct io_uring_task *tctx = tsk->io_uring;
1953 enum task_work_notify_mode notify;
1954 struct io_wq_work_node *node, *prev;
1955 unsigned long flags;
1958 if (unlikely(tsk->flags & PF_EXITING))
1961 WARN_ON_ONCE(!tctx);
1963 spin_lock_irqsave(&tctx->task_lock, flags);
1964 wq_list_add_tail(&req->io_task_work.node, &tctx->task_list);
1965 spin_unlock_irqrestore(&tctx->task_lock, flags);
1967 /* task_work already pending, we're done */
1968 if (test_bit(0, &tctx->task_state) ||
1969 test_and_set_bit(0, &tctx->task_state))
1973 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
1974 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
1975 * processing task_work. There's no reliable way to tell if TWA_RESUME
1978 notify = (req->ctx->flags & IORING_SETUP_SQPOLL) ? TWA_NONE : TWA_SIGNAL;
1980 if (!task_work_add(tsk, &tctx->task_work, notify)) {
1981 wake_up_process(tsk);
1986 * Slow path - we failed, find and delete work. if the work is not
1987 * in the list, it got run and we're fine.
1989 spin_lock_irqsave(&tctx->task_lock, flags);
1990 wq_list_for_each(node, prev, &tctx->task_list) {
1991 if (&req->io_task_work.node == node) {
1992 wq_list_del(&tctx->task_list, node, prev);
1997 spin_unlock_irqrestore(&tctx->task_lock, flags);
1998 clear_bit(0, &tctx->task_state);
2002 static bool io_run_task_work_head(struct callback_head **work_head)
2004 struct callback_head *work, *next;
2005 bool executed = false;
2008 work = xchg(work_head, NULL);
2024 static void io_task_work_add_head(struct callback_head **work_head,
2025 struct callback_head *task_work)
2027 struct callback_head *head;
2030 head = READ_ONCE(*work_head);
2031 task_work->next = head;
2032 } while (cmpxchg(work_head, head, task_work) != head);
2035 static void io_req_task_work_add_fallback(struct io_kiocb *req,
2036 task_work_func_t cb)
2038 init_task_work(&req->task_work, cb);
2039 io_task_work_add_head(&req->ctx->exit_task_work, &req->task_work);
2042 static void io_req_task_cancel(struct callback_head *cb)
2044 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2045 struct io_ring_ctx *ctx = req->ctx;
2047 /* ctx is guaranteed to stay alive while we hold uring_lock */
2048 mutex_lock(&ctx->uring_lock);
2049 io_req_complete_failed(req, req->result);
2050 mutex_unlock(&ctx->uring_lock);
2053 static void __io_req_task_submit(struct io_kiocb *req)
2055 struct io_ring_ctx *ctx = req->ctx;
2057 /* ctx stays valid until unlock, even if we drop all ours ctx->refs */
2058 mutex_lock(&ctx->uring_lock);
2059 if (!(current->flags & PF_EXITING) && !current->in_execve)
2060 __io_queue_sqe(req);
2062 io_req_complete_failed(req, -EFAULT);
2063 mutex_unlock(&ctx->uring_lock);
2066 static void io_req_task_submit(struct callback_head *cb)
2068 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2070 __io_req_task_submit(req);
2073 static void io_req_task_queue_fail(struct io_kiocb *req, int ret)
2076 req->task_work.func = io_req_task_cancel;
2078 if (unlikely(io_req_task_work_add(req)))
2079 io_req_task_work_add_fallback(req, io_req_task_cancel);
2082 static void io_req_task_queue(struct io_kiocb *req)
2084 req->task_work.func = io_req_task_submit;
2086 if (unlikely(io_req_task_work_add(req)))
2087 io_req_task_queue_fail(req, -ECANCELED);
2090 static inline void io_queue_next(struct io_kiocb *req)
2092 struct io_kiocb *nxt = io_req_find_next(req);
2095 io_req_task_queue(nxt);
2098 static void io_free_req(struct io_kiocb *req)
2105 struct task_struct *task;
2110 static inline void io_init_req_batch(struct req_batch *rb)
2117 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2118 struct req_batch *rb)
2121 io_put_task(rb->task, rb->task_refs);
2123 percpu_ref_put_many(&ctx->refs, rb->ctx_refs);
2126 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req,
2127 struct io_submit_state *state)
2130 io_dismantle_req(req);
2132 if (req->task != rb->task) {
2134 io_put_task(rb->task, rb->task_refs);
2135 rb->task = req->task;
2141 if (state->free_reqs != ARRAY_SIZE(state->reqs))
2142 state->reqs[state->free_reqs++] = req;
2144 list_add(&req->compl.list, &state->comp.free_list);
2147 static void io_submit_flush_completions(struct io_ring_ctx *ctx)
2149 struct io_comp_state *cs = &ctx->submit_state.comp;
2151 struct req_batch rb;
2153 spin_lock_irq(&ctx->completion_lock);
2154 for (i = 0; i < nr; i++) {
2155 struct io_kiocb *req = cs->reqs[i];
2157 __io_cqring_fill_event(ctx, req->user_data, req->result,
2160 io_commit_cqring(ctx);
2161 spin_unlock_irq(&ctx->completion_lock);
2162 io_cqring_ev_posted(ctx);
2164 io_init_req_batch(&rb);
2165 for (i = 0; i < nr; i++) {
2166 struct io_kiocb *req = cs->reqs[i];
2168 /* submission and completion refs */
2169 if (req_ref_sub_and_test(req, 2))
2170 io_req_free_batch(&rb, req, &ctx->submit_state);
2173 io_req_free_batch_finish(ctx, &rb);
2178 * Drop reference to request, return next in chain (if there is one) if this
2179 * was the last reference to this request.
2181 static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2183 struct io_kiocb *nxt = NULL;
2185 if (req_ref_put_and_test(req)) {
2186 nxt = io_req_find_next(req);
2192 static inline void io_put_req(struct io_kiocb *req)
2194 if (req_ref_put_and_test(req))
2198 static void io_put_req_deferred_cb(struct callback_head *cb)
2200 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2205 static void io_free_req_deferred(struct io_kiocb *req)
2207 req->task_work.func = io_put_req_deferred_cb;
2208 if (unlikely(io_req_task_work_add(req)))
2209 io_req_task_work_add_fallback(req, io_put_req_deferred_cb);
2212 static inline void io_put_req_deferred(struct io_kiocb *req, int refs)
2214 if (req_ref_sub_and_test(req, refs))
2215 io_free_req_deferred(req);
2218 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2220 /* See comment at the top of this file */
2222 return __io_cqring_events(ctx);
2225 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2227 struct io_rings *rings = ctx->rings;
2229 /* make sure SQ entry isn't read before tail */
2230 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2233 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2235 unsigned int cflags;
2237 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2238 cflags |= IORING_CQE_F_BUFFER;
2239 req->flags &= ~REQ_F_BUFFER_SELECTED;
2244 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2246 struct io_buffer *kbuf;
2248 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2249 return io_put_kbuf(req, kbuf);
2252 static inline bool io_run_task_work(void)
2254 if (current->task_works) {
2255 __set_current_state(TASK_RUNNING);
2264 * Find and free completed poll iocbs
2266 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2267 struct list_head *done)
2269 struct req_batch rb;
2270 struct io_kiocb *req;
2272 /* order with ->result store in io_complete_rw_iopoll() */
2275 io_init_req_batch(&rb);
2276 while (!list_empty(done)) {
2279 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2280 list_del(&req->inflight_entry);
2282 if (READ_ONCE(req->result) == -EAGAIN &&
2283 !(req->flags & REQ_F_DONT_REISSUE)) {
2284 req->iopoll_completed = 0;
2286 io_queue_async_work(req);
2290 if (req->flags & REQ_F_BUFFER_SELECTED)
2291 cflags = io_put_rw_kbuf(req);
2293 __io_cqring_fill_event(ctx, req->user_data, req->result, cflags);
2296 if (req_ref_put_and_test(req))
2297 io_req_free_batch(&rb, req, &ctx->submit_state);
2300 io_commit_cqring(ctx);
2301 io_cqring_ev_posted_iopoll(ctx);
2302 io_req_free_batch_finish(ctx, &rb);
2305 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2308 struct io_kiocb *req, *tmp;
2314 * Only spin for completions if we don't have multiple devices hanging
2315 * off our complete list, and we're under the requested amount.
2317 spin = !ctx->poll_multi_file && *nr_events < min;
2320 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2321 struct kiocb *kiocb = &req->rw.kiocb;
2324 * Move completed and retryable entries to our local lists.
2325 * If we find a request that requires polling, break out
2326 * and complete those lists first, if we have entries there.
2328 if (READ_ONCE(req->iopoll_completed)) {
2329 list_move_tail(&req->inflight_entry, &done);
2332 if (!list_empty(&done))
2335 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2339 /* iopoll may have completed current req */
2340 if (READ_ONCE(req->iopoll_completed))
2341 list_move_tail(&req->inflight_entry, &done);
2348 if (!list_empty(&done))
2349 io_iopoll_complete(ctx, nr_events, &done);
2355 * We can't just wait for polled events to come to us, we have to actively
2356 * find and complete them.
2358 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2360 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2363 mutex_lock(&ctx->uring_lock);
2364 while (!list_empty(&ctx->iopoll_list)) {
2365 unsigned int nr_events = 0;
2367 io_do_iopoll(ctx, &nr_events, 0);
2369 /* let it sleep and repeat later if can't complete a request */
2373 * Ensure we allow local-to-the-cpu processing to take place,
2374 * in this case we need to ensure that we reap all events.
2375 * Also let task_work, etc. to progress by releasing the mutex
2377 if (need_resched()) {
2378 mutex_unlock(&ctx->uring_lock);
2380 mutex_lock(&ctx->uring_lock);
2383 mutex_unlock(&ctx->uring_lock);
2386 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2388 unsigned int nr_events = 0;
2392 * We disallow the app entering submit/complete with polling, but we
2393 * still need to lock the ring to prevent racing with polled issue
2394 * that got punted to a workqueue.
2396 mutex_lock(&ctx->uring_lock);
2398 * Don't enter poll loop if we already have events pending.
2399 * If we do, we can potentially be spinning for commands that
2400 * already triggered a CQE (eg in error).
2402 if (test_bit(0, &ctx->check_cq_overflow))
2403 __io_cqring_overflow_flush(ctx, false);
2404 if (io_cqring_events(ctx))
2408 * If a submit got punted to a workqueue, we can have the
2409 * application entering polling for a command before it gets
2410 * issued. That app will hold the uring_lock for the duration
2411 * of the poll right here, so we need to take a breather every
2412 * now and then to ensure that the issue has a chance to add
2413 * the poll to the issued list. Otherwise we can spin here
2414 * forever, while the workqueue is stuck trying to acquire the
2417 if (list_empty(&ctx->iopoll_list)) {
2418 mutex_unlock(&ctx->uring_lock);
2420 mutex_lock(&ctx->uring_lock);
2422 if (list_empty(&ctx->iopoll_list))
2425 ret = io_do_iopoll(ctx, &nr_events, min);
2426 } while (!ret && nr_events < min && !need_resched());
2428 mutex_unlock(&ctx->uring_lock);
2432 static void kiocb_end_write(struct io_kiocb *req)
2435 * Tell lockdep we inherited freeze protection from submission
2438 if (req->flags & REQ_F_ISREG) {
2439 struct super_block *sb = file_inode(req->file)->i_sb;
2441 __sb_writers_acquired(sb, SB_FREEZE_WRITE);
2447 static bool io_resubmit_prep(struct io_kiocb *req)
2449 struct io_async_rw *rw = req->async_data;
2452 return !io_req_prep_async(req);
2453 /* may have left rw->iter inconsistent on -EIOCBQUEUED */
2454 iov_iter_revert(&rw->iter, req->result - iov_iter_count(&rw->iter));
2458 static bool io_rw_should_reissue(struct io_kiocb *req)
2460 umode_t mode = file_inode(req->file)->i_mode;
2461 struct io_ring_ctx *ctx = req->ctx;
2463 if (!S_ISBLK(mode) && !S_ISREG(mode))
2465 if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
2466 !(ctx->flags & IORING_SETUP_IOPOLL)))
2469 * If ref is dying, we might be running poll reap from the exit work.
2470 * Don't attempt to reissue from that path, just let it fail with
2473 if (percpu_ref_is_dying(&ctx->refs))
2478 static bool io_resubmit_prep(struct io_kiocb *req)
2482 static bool io_rw_should_reissue(struct io_kiocb *req)
2488 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2489 unsigned int issue_flags)
2493 if (req->rw.kiocb.ki_flags & IOCB_WRITE)
2494 kiocb_end_write(req);
2495 if (res != req->result) {
2496 if ((res == -EAGAIN || res == -EOPNOTSUPP) &&
2497 io_rw_should_reissue(req)) {
2498 req->flags |= REQ_F_REISSUE;
2503 if (req->flags & REQ_F_BUFFER_SELECTED)
2504 cflags = io_put_rw_kbuf(req);
2505 __io_req_complete(req, issue_flags, res, cflags);
2508 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2510 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2512 __io_complete_rw(req, res, res2, 0);
2515 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2517 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2519 if (kiocb->ki_flags & IOCB_WRITE)
2520 kiocb_end_write(req);
2521 if (unlikely(res != req->result)) {
2522 if (!(res == -EAGAIN && io_rw_should_reissue(req) &&
2523 io_resubmit_prep(req))) {
2525 req->flags |= REQ_F_DONT_REISSUE;
2529 WRITE_ONCE(req->result, res);
2530 /* order with io_iopoll_complete() checking ->result */
2532 WRITE_ONCE(req->iopoll_completed, 1);
2536 * After the iocb has been issued, it's safe to be found on the poll list.
2537 * Adding the kiocb to the list AFTER submission ensures that we don't
2538 * find it from a io_do_iopoll() thread before the issuer is done
2539 * accessing the kiocb cookie.
2541 static void io_iopoll_req_issued(struct io_kiocb *req)
2543 struct io_ring_ctx *ctx = req->ctx;
2544 const bool in_async = io_wq_current_is_worker();
2546 /* workqueue context doesn't hold uring_lock, grab it now */
2547 if (unlikely(in_async))
2548 mutex_lock(&ctx->uring_lock);
2551 * Track whether we have multiple files in our lists. This will impact
2552 * how we do polling eventually, not spinning if we're on potentially
2553 * different devices.
2555 if (list_empty(&ctx->iopoll_list)) {
2556 ctx->poll_multi_file = false;
2557 } else if (!ctx->poll_multi_file) {
2558 struct io_kiocb *list_req;
2560 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2562 if (list_req->file != req->file)
2563 ctx->poll_multi_file = true;
2567 * For fast devices, IO may have already completed. If it has, add
2568 * it to the front so we find it first.
2570 if (READ_ONCE(req->iopoll_completed))
2571 list_add(&req->inflight_entry, &ctx->iopoll_list);
2573 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2575 if (unlikely(in_async)) {
2577 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
2578 * in sq thread task context or in io worker task context. If
2579 * current task context is sq thread, we don't need to check
2580 * whether should wake up sq thread.
2582 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2583 wq_has_sleeper(&ctx->sq_data->wait))
2584 wake_up(&ctx->sq_data->wait);
2586 mutex_unlock(&ctx->uring_lock);
2590 static inline void io_state_file_put(struct io_submit_state *state)
2592 if (state->file_refs) {
2593 fput_many(state->file, state->file_refs);
2594 state->file_refs = 0;
2599 * Get as many references to a file as we have IOs left in this submission,
2600 * assuming most submissions are for one file, or at least that each file
2601 * has more than one submission.
2603 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2608 if (state->file_refs) {
2609 if (state->fd == fd) {
2613 io_state_file_put(state);
2615 state->file = fget_many(fd, state->ios_left);
2616 if (unlikely(!state->file))
2620 state->file_refs = state->ios_left - 1;
2624 static bool io_bdev_nowait(struct block_device *bdev)
2626 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2630 * If we tracked the file through the SCM inflight mechanism, we could support
2631 * any file. For now, just ensure that anything potentially problematic is done
2634 static bool __io_file_supports_async(struct file *file, int rw)
2636 umode_t mode = file_inode(file)->i_mode;
2638 if (S_ISBLK(mode)) {
2639 if (IS_ENABLED(CONFIG_BLOCK) &&
2640 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2646 if (S_ISREG(mode)) {
2647 if (IS_ENABLED(CONFIG_BLOCK) &&
2648 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2649 file->f_op != &io_uring_fops)
2654 /* any ->read/write should understand O_NONBLOCK */
2655 if (file->f_flags & O_NONBLOCK)
2658 if (!(file->f_mode & FMODE_NOWAIT))
2662 return file->f_op->read_iter != NULL;
2664 return file->f_op->write_iter != NULL;
2667 static bool io_file_supports_async(struct io_kiocb *req, int rw)
2669 if (rw == READ && (req->flags & REQ_F_ASYNC_READ))
2671 else if (rw == WRITE && (req->flags & REQ_F_ASYNC_WRITE))
2674 return __io_file_supports_async(req->file, rw);
2677 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2679 struct io_ring_ctx *ctx = req->ctx;
2680 struct kiocb *kiocb = &req->rw.kiocb;
2681 struct file *file = req->file;
2685 if (!(req->flags & REQ_F_ISREG) && S_ISREG(file_inode(file)->i_mode))
2686 req->flags |= REQ_F_ISREG;
2688 kiocb->ki_pos = READ_ONCE(sqe->off);
2689 if (kiocb->ki_pos == -1 && !(file->f_mode & FMODE_STREAM)) {
2690 req->flags |= REQ_F_CUR_POS;
2691 kiocb->ki_pos = file->f_pos;
2693 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2694 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2695 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2699 /* don't allow async punt for O_NONBLOCK or RWF_NOWAIT */
2700 if ((kiocb->ki_flags & IOCB_NOWAIT) || (file->f_flags & O_NONBLOCK))
2701 req->flags |= REQ_F_NOWAIT;
2703 ioprio = READ_ONCE(sqe->ioprio);
2705 ret = ioprio_check_cap(ioprio);
2709 kiocb->ki_ioprio = ioprio;
2711 kiocb->ki_ioprio = get_current_ioprio();
2713 if (ctx->flags & IORING_SETUP_IOPOLL) {
2714 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2715 !kiocb->ki_filp->f_op->iopoll)
2718 kiocb->ki_flags |= IOCB_HIPRI;
2719 kiocb->ki_complete = io_complete_rw_iopoll;
2720 req->iopoll_completed = 0;
2722 if (kiocb->ki_flags & IOCB_HIPRI)
2724 kiocb->ki_complete = io_complete_rw;
2727 if (req->opcode == IORING_OP_READ_FIXED ||
2728 req->opcode == IORING_OP_WRITE_FIXED) {
2730 io_req_set_rsrc_node(req);
2733 req->rw.addr = READ_ONCE(sqe->addr);
2734 req->rw.len = READ_ONCE(sqe->len);
2735 req->buf_index = READ_ONCE(sqe->buf_index);
2739 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2745 case -ERESTARTNOINTR:
2746 case -ERESTARTNOHAND:
2747 case -ERESTART_RESTARTBLOCK:
2749 * We can't just restart the syscall, since previously
2750 * submitted sqes may already be in progress. Just fail this
2756 kiocb->ki_complete(kiocb, ret, 0);
2760 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2761 unsigned int issue_flags)
2763 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2764 struct io_async_rw *io = req->async_data;
2765 bool check_reissue = kiocb->ki_complete == io_complete_rw;
2767 /* add previously done IO, if any */
2768 if (io && io->bytes_done > 0) {
2770 ret = io->bytes_done;
2772 ret += io->bytes_done;
2775 if (req->flags & REQ_F_CUR_POS)
2776 req->file->f_pos = kiocb->ki_pos;
2777 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2778 __io_complete_rw(req, ret, 0, issue_flags);
2780 io_rw_done(kiocb, ret);
2782 if (check_reissue && (req->flags & REQ_F_REISSUE)) {
2783 req->flags &= ~REQ_F_REISSUE;
2784 if (io_resubmit_prep(req)) {
2786 io_queue_async_work(req);
2791 if (req->flags & REQ_F_BUFFER_SELECTED)
2792 cflags = io_put_rw_kbuf(req);
2793 __io_req_complete(req, issue_flags, ret, cflags);
2798 static int __io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter,
2799 struct io_mapped_ubuf *imu)
2801 size_t len = req->rw.len;
2802 u64 buf_end, buf_addr = req->rw.addr;
2805 if (unlikely(check_add_overflow(buf_addr, (u64)len, &buf_end)))
2807 /* not inside the mapped region */
2808 if (unlikely(buf_addr < imu->ubuf || buf_end > imu->ubuf_end))
2812 * May not be a start of buffer, set size appropriately
2813 * and advance us to the beginning.
2815 offset = buf_addr - imu->ubuf;
2816 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2820 * Don't use iov_iter_advance() here, as it's really slow for
2821 * using the latter parts of a big fixed buffer - it iterates
2822 * over each segment manually. We can cheat a bit here, because
2825 * 1) it's a BVEC iter, we set it up
2826 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2827 * first and last bvec
2829 * So just find our index, and adjust the iterator afterwards.
2830 * If the offset is within the first bvec (or the whole first
2831 * bvec, just use iov_iter_advance(). This makes it easier
2832 * since we can just skip the first segment, which may not
2833 * be PAGE_SIZE aligned.
2835 const struct bio_vec *bvec = imu->bvec;
2837 if (offset <= bvec->bv_len) {
2838 iov_iter_advance(iter, offset);
2840 unsigned long seg_skip;
2842 /* skip first vec */
2843 offset -= bvec->bv_len;
2844 seg_skip = 1 + (offset >> PAGE_SHIFT);
2846 iter->bvec = bvec + seg_skip;
2847 iter->nr_segs -= seg_skip;
2848 iter->count -= bvec->bv_len + offset;
2849 iter->iov_offset = offset & ~PAGE_MASK;
2856 static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
2858 struct io_ring_ctx *ctx = req->ctx;
2859 struct io_mapped_ubuf *imu = req->imu;
2860 u16 index, buf_index = req->buf_index;
2863 if (unlikely(buf_index >= ctx->nr_user_bufs))
2865 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2866 imu = READ_ONCE(ctx->user_bufs[index]);
2869 return __io_import_fixed(req, rw, iter, imu);
2872 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2875 mutex_unlock(&ctx->uring_lock);
2878 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2881 * "Normal" inline submissions always hold the uring_lock, since we
2882 * grab it from the system call. Same is true for the SQPOLL offload.
2883 * The only exception is when we've detached the request and issue it
2884 * from an async worker thread, grab the lock for that case.
2887 mutex_lock(&ctx->uring_lock);
2890 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2891 int bgid, struct io_buffer *kbuf,
2894 struct io_buffer *head;
2896 if (req->flags & REQ_F_BUFFER_SELECTED)
2899 io_ring_submit_lock(req->ctx, needs_lock);
2901 lockdep_assert_held(&req->ctx->uring_lock);
2903 head = xa_load(&req->ctx->io_buffers, bgid);
2905 if (!list_empty(&head->list)) {
2906 kbuf = list_last_entry(&head->list, struct io_buffer,
2908 list_del(&kbuf->list);
2911 xa_erase(&req->ctx->io_buffers, bgid);
2913 if (*len > kbuf->len)
2916 kbuf = ERR_PTR(-ENOBUFS);
2919 io_ring_submit_unlock(req->ctx, needs_lock);
2924 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2927 struct io_buffer *kbuf;
2930 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2931 bgid = req->buf_index;
2932 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2935 req->rw.addr = (u64) (unsigned long) kbuf;
2936 req->flags |= REQ_F_BUFFER_SELECTED;
2937 return u64_to_user_ptr(kbuf->addr);
2940 #ifdef CONFIG_COMPAT
2941 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2944 struct compat_iovec __user *uiov;
2945 compat_ssize_t clen;
2949 uiov = u64_to_user_ptr(req->rw.addr);
2950 if (!access_ok(uiov, sizeof(*uiov)))
2952 if (__get_user(clen, &uiov->iov_len))
2958 buf = io_rw_buffer_select(req, &len, needs_lock);
2960 return PTR_ERR(buf);
2961 iov[0].iov_base = buf;
2962 iov[0].iov_len = (compat_size_t) len;
2967 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2970 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2974 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2977 len = iov[0].iov_len;
2980 buf = io_rw_buffer_select(req, &len, needs_lock);
2982 return PTR_ERR(buf);
2983 iov[0].iov_base = buf;
2984 iov[0].iov_len = len;
2988 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2991 if (req->flags & REQ_F_BUFFER_SELECTED) {
2992 struct io_buffer *kbuf;
2994 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2995 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
2996 iov[0].iov_len = kbuf->len;
2999 if (req->rw.len != 1)
3002 #ifdef CONFIG_COMPAT
3003 if (req->ctx->compat)
3004 return io_compat_import(req, iov, needs_lock);
3007 return __io_iov_buffer_select(req, iov, needs_lock);
3010 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
3011 struct iov_iter *iter, bool needs_lock)
3013 void __user *buf = u64_to_user_ptr(req->rw.addr);
3014 size_t sqe_len = req->rw.len;
3015 u8 opcode = req->opcode;
3018 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3020 return io_import_fixed(req, rw, iter);
3023 /* buffer index only valid with fixed read/write, or buffer select */
3024 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3027 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3028 if (req->flags & REQ_F_BUFFER_SELECT) {
3029 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3031 return PTR_ERR(buf);
3032 req->rw.len = sqe_len;
3035 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3040 if (req->flags & REQ_F_BUFFER_SELECT) {
3041 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3043 iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
3048 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3052 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3054 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3058 * For files that don't have ->read_iter() and ->write_iter(), handle them
3059 * by looping over ->read() or ->write() manually.
3061 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3063 struct kiocb *kiocb = &req->rw.kiocb;
3064 struct file *file = req->file;
3068 * Don't support polled IO through this interface, and we can't
3069 * support non-blocking either. For the latter, this just causes
3070 * the kiocb to be handled from an async context.
3072 if (kiocb->ki_flags & IOCB_HIPRI)
3074 if (kiocb->ki_flags & IOCB_NOWAIT)
3077 while (iov_iter_count(iter)) {
3081 if (!iov_iter_is_bvec(iter)) {
3082 iovec = iov_iter_iovec(iter);
3084 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3085 iovec.iov_len = req->rw.len;
3089 nr = file->f_op->read(file, iovec.iov_base,
3090 iovec.iov_len, io_kiocb_ppos(kiocb));
3092 nr = file->f_op->write(file, iovec.iov_base,
3093 iovec.iov_len, io_kiocb_ppos(kiocb));
3102 if (nr != iovec.iov_len)
3106 iov_iter_advance(iter, nr);
3112 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3113 const struct iovec *fast_iov, struct iov_iter *iter)
3115 struct io_async_rw *rw = req->async_data;
3117 memcpy(&rw->iter, iter, sizeof(*iter));
3118 rw->free_iovec = iovec;
3120 /* can only be fixed buffers, no need to do anything */
3121 if (iov_iter_is_bvec(iter))
3124 unsigned iov_off = 0;
3126 rw->iter.iov = rw->fast_iov;
3127 if (iter->iov != fast_iov) {
3128 iov_off = iter->iov - fast_iov;
3129 rw->iter.iov += iov_off;
3131 if (rw->fast_iov != fast_iov)
3132 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3133 sizeof(struct iovec) * iter->nr_segs);
3135 req->flags |= REQ_F_NEED_CLEANUP;
3139 static inline int io_alloc_async_data(struct io_kiocb *req)
3141 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3142 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3143 return req->async_data == NULL;
3146 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3147 const struct iovec *fast_iov,
3148 struct iov_iter *iter, bool force)
3150 if (!force && !io_op_defs[req->opcode].needs_async_setup)
3152 if (!req->async_data) {
3153 if (io_alloc_async_data(req)) {
3158 io_req_map_rw(req, iovec, fast_iov, iter);
3163 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3165 struct io_async_rw *iorw = req->async_data;
3166 struct iovec *iov = iorw->fast_iov;
3169 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3170 if (unlikely(ret < 0))
3173 iorw->bytes_done = 0;
3174 iorw->free_iovec = iov;
3176 req->flags |= REQ_F_NEED_CLEANUP;
3180 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3182 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3184 return io_prep_rw(req, sqe);
3188 * This is our waitqueue callback handler, registered through lock_page_async()
3189 * when we initially tried to do the IO with the iocb armed our waitqueue.
3190 * This gets called when the page is unlocked, and we generally expect that to
3191 * happen when the page IO is completed and the page is now uptodate. This will
3192 * queue a task_work based retry of the operation, attempting to copy the data
3193 * again. If the latter fails because the page was NOT uptodate, then we will
3194 * do a thread based blocking retry of the operation. That's the unexpected
3197 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3198 int sync, void *arg)
3200 struct wait_page_queue *wpq;
3201 struct io_kiocb *req = wait->private;
3202 struct wait_page_key *key = arg;
3204 wpq = container_of(wait, struct wait_page_queue, wait);
3206 if (!wake_page_match(wpq, key))
3209 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3210 list_del_init(&wait->entry);
3212 /* submit ref gets dropped, acquire a new one */
3214 io_req_task_queue(req);
3219 * This controls whether a given IO request should be armed for async page
3220 * based retry. If we return false here, the request is handed to the async
3221 * worker threads for retry. If we're doing buffered reads on a regular file,
3222 * we prepare a private wait_page_queue entry and retry the operation. This
3223 * will either succeed because the page is now uptodate and unlocked, or it
3224 * will register a callback when the page is unlocked at IO completion. Through
3225 * that callback, io_uring uses task_work to setup a retry of the operation.
3226 * That retry will attempt the buffered read again. The retry will generally
3227 * succeed, or in rare cases where it fails, we then fall back to using the
3228 * async worker threads for a blocking retry.
3230 static bool io_rw_should_retry(struct io_kiocb *req)
3232 struct io_async_rw *rw = req->async_data;
3233 struct wait_page_queue *wait = &rw->wpq;
3234 struct kiocb *kiocb = &req->rw.kiocb;
3236 /* never retry for NOWAIT, we just complete with -EAGAIN */
3237 if (req->flags & REQ_F_NOWAIT)
3240 /* Only for buffered IO */
3241 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3245 * just use poll if we can, and don't attempt if the fs doesn't
3246 * support callback based unlocks
3248 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3251 wait->wait.func = io_async_buf_func;
3252 wait->wait.private = req;
3253 wait->wait.flags = 0;
3254 INIT_LIST_HEAD(&wait->wait.entry);
3255 kiocb->ki_flags |= IOCB_WAITQ;
3256 kiocb->ki_flags &= ~IOCB_NOWAIT;
3257 kiocb->ki_waitq = wait;
3261 static inline int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3263 if (req->file->f_op->read_iter)
3264 return call_read_iter(req->file, &req->rw.kiocb, iter);
3265 else if (req->file->f_op->read)
3266 return loop_rw_iter(READ, req, iter);
3271 static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3273 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3274 struct kiocb *kiocb = &req->rw.kiocb;
3275 struct iov_iter __iter, *iter = &__iter;
3276 struct io_async_rw *rw = req->async_data;
3277 ssize_t io_size, ret, ret2;
3278 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3284 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3288 io_size = iov_iter_count(iter);
3289 req->result = io_size;
3291 /* Ensure we clear previously set non-block flag */
3292 if (!force_nonblock)
3293 kiocb->ki_flags &= ~IOCB_NOWAIT;
3295 kiocb->ki_flags |= IOCB_NOWAIT;
3297 /* If the file doesn't support async, just async punt */
3298 if (force_nonblock && !io_file_supports_async(req, READ)) {
3299 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3300 return ret ?: -EAGAIN;
3303 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size);
3304 if (unlikely(ret)) {
3309 ret = io_iter_do_read(req, iter);
3311 if (ret == -EAGAIN || (req->flags & REQ_F_REISSUE)) {
3312 req->flags &= ~REQ_F_REISSUE;
3313 /* IOPOLL retry should happen for io-wq threads */
3314 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3316 /* no retry on NONBLOCK nor RWF_NOWAIT */
3317 if (req->flags & REQ_F_NOWAIT)
3319 /* some cases will consume bytes even on error returns */
3320 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3322 } else if (ret == -EIOCBQUEUED) {
3324 } else if (ret <= 0 || ret == io_size || !force_nonblock ||
3325 (req->flags & REQ_F_NOWAIT) || !(req->flags & REQ_F_ISREG)) {
3326 /* read all, failed, already did sync or don't want to retry */
3330 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3335 rw = req->async_data;
3336 /* now use our persistent iterator, if we aren't already */
3341 rw->bytes_done += ret;
3342 /* if we can retry, do so with the callbacks armed */
3343 if (!io_rw_should_retry(req)) {
3344 kiocb->ki_flags &= ~IOCB_WAITQ;
3349 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3350 * we get -EIOCBQUEUED, then we'll get a notification when the
3351 * desired page gets unlocked. We can also get a partial read
3352 * here, and if we do, then just retry at the new offset.
3354 ret = io_iter_do_read(req, iter);
3355 if (ret == -EIOCBQUEUED)
3357 /* we got some bytes, but not all. retry. */
3358 kiocb->ki_flags &= ~IOCB_WAITQ;
3359 } while (ret > 0 && ret < io_size);
3361 kiocb_done(kiocb, ret, issue_flags);
3363 /* it's faster to check here then delegate to kfree */
3369 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3371 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3373 return io_prep_rw(req, sqe);
3376 static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3378 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3379 struct kiocb *kiocb = &req->rw.kiocb;
3380 struct iov_iter __iter, *iter = &__iter;
3381 struct io_async_rw *rw = req->async_data;
3382 ssize_t ret, ret2, io_size;
3383 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3389 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3393 io_size = iov_iter_count(iter);
3394 req->result = io_size;
3396 /* Ensure we clear previously set non-block flag */
3397 if (!force_nonblock)
3398 kiocb->ki_flags &= ~IOCB_NOWAIT;
3400 kiocb->ki_flags |= IOCB_NOWAIT;
3402 /* If the file doesn't support async, just async punt */
3403 if (force_nonblock && !io_file_supports_async(req, WRITE))
3406 /* file path doesn't support NOWAIT for non-direct_IO */
3407 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3408 (req->flags & REQ_F_ISREG))
3411 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size);
3416 * Open-code file_start_write here to grab freeze protection,
3417 * which will be released by another thread in
3418 * io_complete_rw(). Fool lockdep by telling it the lock got
3419 * released so that it doesn't complain about the held lock when
3420 * we return to userspace.
3422 if (req->flags & REQ_F_ISREG) {
3423 sb_start_write(file_inode(req->file)->i_sb);
3424 __sb_writers_release(file_inode(req->file)->i_sb,
3427 kiocb->ki_flags |= IOCB_WRITE;
3429 if (req->file->f_op->write_iter)
3430 ret2 = call_write_iter(req->file, kiocb, iter);
3431 else if (req->file->f_op->write)
3432 ret2 = loop_rw_iter(WRITE, req, iter);
3436 if (req->flags & REQ_F_REISSUE) {
3437 req->flags &= ~REQ_F_REISSUE;
3442 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3443 * retry them without IOCB_NOWAIT.
3445 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3447 /* no retry on NONBLOCK nor RWF_NOWAIT */
3448 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3450 if (!force_nonblock || ret2 != -EAGAIN) {
3451 /* IOPOLL retry should happen for io-wq threads */
3452 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3455 kiocb_done(kiocb, ret2, issue_flags);
3458 /* some cases will consume bytes even on error returns */
3459 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3460 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3461 return ret ?: -EAGAIN;
3464 /* it's reportedly faster than delegating the null check to kfree() */
3470 static int io_renameat_prep(struct io_kiocb *req,
3471 const struct io_uring_sqe *sqe)
3473 struct io_rename *ren = &req->rename;
3474 const char __user *oldf, *newf;
3476 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3478 if (sqe->ioprio || sqe->buf_index)
3480 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3483 ren->old_dfd = READ_ONCE(sqe->fd);
3484 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3485 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3486 ren->new_dfd = READ_ONCE(sqe->len);
3487 ren->flags = READ_ONCE(sqe->rename_flags);
3489 ren->oldpath = getname(oldf);
3490 if (IS_ERR(ren->oldpath))
3491 return PTR_ERR(ren->oldpath);
3493 ren->newpath = getname(newf);
3494 if (IS_ERR(ren->newpath)) {
3495 putname(ren->oldpath);
3496 return PTR_ERR(ren->newpath);
3499 req->flags |= REQ_F_NEED_CLEANUP;
3503 static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3505 struct io_rename *ren = &req->rename;
3508 if (issue_flags & IO_URING_F_NONBLOCK)
3511 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3512 ren->newpath, ren->flags);
3514 req->flags &= ~REQ_F_NEED_CLEANUP;
3517 io_req_complete(req, ret);
3521 static int io_unlinkat_prep(struct io_kiocb *req,
3522 const struct io_uring_sqe *sqe)
3524 struct io_unlink *un = &req->unlink;
3525 const char __user *fname;
3527 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3529 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
3531 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3534 un->dfd = READ_ONCE(sqe->fd);
3536 un->flags = READ_ONCE(sqe->unlink_flags);
3537 if (un->flags & ~AT_REMOVEDIR)
3540 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3541 un->filename = getname(fname);
3542 if (IS_ERR(un->filename))
3543 return PTR_ERR(un->filename);
3545 req->flags |= REQ_F_NEED_CLEANUP;
3549 static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3551 struct io_unlink *un = &req->unlink;
3554 if (issue_flags & IO_URING_F_NONBLOCK)
3557 if (un->flags & AT_REMOVEDIR)
3558 ret = do_rmdir(un->dfd, un->filename);
3560 ret = do_unlinkat(un->dfd, un->filename);
3562 req->flags &= ~REQ_F_NEED_CLEANUP;
3565 io_req_complete(req, ret);
3569 static int io_shutdown_prep(struct io_kiocb *req,
3570 const struct io_uring_sqe *sqe)
3572 #if defined(CONFIG_NET)
3573 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3575 if (sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3579 req->shutdown.how = READ_ONCE(sqe->len);
3586 static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3588 #if defined(CONFIG_NET)
3589 struct socket *sock;
3592 if (issue_flags & IO_URING_F_NONBLOCK)
3595 sock = sock_from_file(req->file);
3596 if (unlikely(!sock))
3599 ret = __sys_shutdown_sock(sock, req->shutdown.how);
3602 io_req_complete(req, ret);
3609 static int __io_splice_prep(struct io_kiocb *req,
3610 const struct io_uring_sqe *sqe)
3612 struct io_splice *sp = &req->splice;
3613 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3615 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3619 sp->len = READ_ONCE(sqe->len);
3620 sp->flags = READ_ONCE(sqe->splice_flags);
3622 if (unlikely(sp->flags & ~valid_flags))
3625 sp->file_in = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in),
3626 (sp->flags & SPLICE_F_FD_IN_FIXED));
3629 req->flags |= REQ_F_NEED_CLEANUP;
3633 static int io_tee_prep(struct io_kiocb *req,
3634 const struct io_uring_sqe *sqe)
3636 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3638 return __io_splice_prep(req, sqe);
3641 static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
3643 struct io_splice *sp = &req->splice;
3644 struct file *in = sp->file_in;
3645 struct file *out = sp->file_out;
3646 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3649 if (issue_flags & IO_URING_F_NONBLOCK)
3652 ret = do_tee(in, out, sp->len, flags);
3654 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3656 req->flags &= ~REQ_F_NEED_CLEANUP;
3660 io_req_complete(req, ret);
3664 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3666 struct io_splice *sp = &req->splice;
3668 sp->off_in = READ_ONCE(sqe->splice_off_in);
3669 sp->off_out = READ_ONCE(sqe->off);
3670 return __io_splice_prep(req, sqe);
3673 static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
3675 struct io_splice *sp = &req->splice;
3676 struct file *in = sp->file_in;
3677 struct file *out = sp->file_out;
3678 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3679 loff_t *poff_in, *poff_out;
3682 if (issue_flags & IO_URING_F_NONBLOCK)
3685 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3686 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3689 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3691 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3693 req->flags &= ~REQ_F_NEED_CLEANUP;
3697 io_req_complete(req, ret);
3702 * IORING_OP_NOP just posts a completion event, nothing else.
3704 static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
3706 struct io_ring_ctx *ctx = req->ctx;
3708 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3711 __io_req_complete(req, issue_flags, 0, 0);
3715 static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3717 struct io_ring_ctx *ctx = req->ctx;
3722 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3724 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3727 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3728 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3731 req->sync.off = READ_ONCE(sqe->off);
3732 req->sync.len = READ_ONCE(sqe->len);
3736 static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
3738 loff_t end = req->sync.off + req->sync.len;
3741 /* fsync always requires a blocking context */
3742 if (issue_flags & IO_URING_F_NONBLOCK)
3745 ret = vfs_fsync_range(req->file, req->sync.off,
3746 end > 0 ? end : LLONG_MAX,
3747 req->sync.flags & IORING_FSYNC_DATASYNC);
3750 io_req_complete(req, ret);
3754 static int io_fallocate_prep(struct io_kiocb *req,
3755 const struct io_uring_sqe *sqe)
3757 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3759 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3762 req->sync.off = READ_ONCE(sqe->off);
3763 req->sync.len = READ_ONCE(sqe->addr);
3764 req->sync.mode = READ_ONCE(sqe->len);
3768 static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
3772 /* fallocate always requiring blocking context */
3773 if (issue_flags & IO_URING_F_NONBLOCK)
3775 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3779 io_req_complete(req, ret);
3783 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3785 const char __user *fname;
3788 if (unlikely(sqe->ioprio || sqe->buf_index))
3790 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3793 /* open.how should be already initialised */
3794 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3795 req->open.how.flags |= O_LARGEFILE;
3797 req->open.dfd = READ_ONCE(sqe->fd);
3798 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3799 req->open.filename = getname(fname);
3800 if (IS_ERR(req->open.filename)) {
3801 ret = PTR_ERR(req->open.filename);
3802 req->open.filename = NULL;
3805 req->open.nofile = rlimit(RLIMIT_NOFILE);
3806 req->flags |= REQ_F_NEED_CLEANUP;
3810 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3814 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3816 mode = READ_ONCE(sqe->len);
3817 flags = READ_ONCE(sqe->open_flags);
3818 req->open.how = build_open_how(flags, mode);
3819 return __io_openat_prep(req, sqe);
3822 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3824 struct open_how __user *how;
3828 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3830 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3831 len = READ_ONCE(sqe->len);
3832 if (len < OPEN_HOW_SIZE_VER0)
3835 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3840 return __io_openat_prep(req, sqe);
3843 static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
3845 struct open_flags op;
3848 bool resolve_nonblock;
3851 ret = build_open_flags(&req->open.how, &op);
3854 nonblock_set = op.open_flag & O_NONBLOCK;
3855 resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
3856 if (issue_flags & IO_URING_F_NONBLOCK) {
3858 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
3859 * it'll always -EAGAIN
3861 if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
3863 op.lookup_flags |= LOOKUP_CACHED;
3864 op.open_flag |= O_NONBLOCK;
3867 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3871 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3874 * We could hang on to this 'fd' on retrying, but seems like
3875 * marginal gain for something that is now known to be a slower
3876 * path. So just put it, and we'll get a new one when we retry.
3880 ret = PTR_ERR(file);
3881 /* only retry if RESOLVE_CACHED wasn't already set by application */
3882 if (ret == -EAGAIN &&
3883 (!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)))
3888 if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
3889 file->f_flags &= ~O_NONBLOCK;
3890 fsnotify_open(file);
3891 fd_install(ret, file);
3893 putname(req->open.filename);
3894 req->flags &= ~REQ_F_NEED_CLEANUP;
3897 __io_req_complete(req, issue_flags, ret, 0);
3901 static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
3903 return io_openat2(req, issue_flags);
3906 static int io_remove_buffers_prep(struct io_kiocb *req,
3907 const struct io_uring_sqe *sqe)
3909 struct io_provide_buf *p = &req->pbuf;
3912 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3915 tmp = READ_ONCE(sqe->fd);
3916 if (!tmp || tmp > USHRT_MAX)
3919 memset(p, 0, sizeof(*p));
3921 p->bgid = READ_ONCE(sqe->buf_group);
3925 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3926 int bgid, unsigned nbufs)
3930 /* shouldn't happen */
3934 /* the head kbuf is the list itself */
3935 while (!list_empty(&buf->list)) {
3936 struct io_buffer *nxt;
3938 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3939 list_del(&nxt->list);
3946 xa_erase(&ctx->io_buffers, bgid);
3951 static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
3953 struct io_provide_buf *p = &req->pbuf;
3954 struct io_ring_ctx *ctx = req->ctx;
3955 struct io_buffer *head;
3957 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3959 io_ring_submit_lock(ctx, !force_nonblock);
3961 lockdep_assert_held(&ctx->uring_lock);
3964 head = xa_load(&ctx->io_buffers, p->bgid);
3966 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3970 /* complete before unlock, IOPOLL may need the lock */
3971 __io_req_complete(req, issue_flags, ret, 0);
3972 io_ring_submit_unlock(ctx, !force_nonblock);
3976 static int io_provide_buffers_prep(struct io_kiocb *req,
3977 const struct io_uring_sqe *sqe)
3979 unsigned long size, tmp_check;
3980 struct io_provide_buf *p = &req->pbuf;
3983 if (sqe->ioprio || sqe->rw_flags)
3986 tmp = READ_ONCE(sqe->fd);
3987 if (!tmp || tmp > USHRT_MAX)
3990 p->addr = READ_ONCE(sqe->addr);
3991 p->len = READ_ONCE(sqe->len);
3993 if (check_mul_overflow((unsigned long)p->len, (unsigned long)p->nbufs,
3996 if (check_add_overflow((unsigned long)p->addr, size, &tmp_check))
3999 size = (unsigned long)p->len * p->nbufs;
4000 if (!access_ok(u64_to_user_ptr(p->addr), size))
4003 p->bgid = READ_ONCE(sqe->buf_group);
4004 tmp = READ_ONCE(sqe->off);
4005 if (tmp > USHRT_MAX)
4011 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
4013 struct io_buffer *buf;
4014 u64 addr = pbuf->addr;
4015 int i, bid = pbuf->bid;
4017 for (i = 0; i < pbuf->nbufs; i++) {
4018 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
4023 buf->len = min_t(__u32, pbuf->len, MAX_RW_COUNT);
4028 INIT_LIST_HEAD(&buf->list);
4031 list_add_tail(&buf->list, &(*head)->list);
4035 return i ? i : -ENOMEM;
4038 static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
4040 struct io_provide_buf *p = &req->pbuf;
4041 struct io_ring_ctx *ctx = req->ctx;
4042 struct io_buffer *head, *list;
4044 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4046 io_ring_submit_lock(ctx, !force_nonblock);
4048 lockdep_assert_held(&ctx->uring_lock);
4050 list = head = xa_load(&ctx->io_buffers, p->bgid);
4052 ret = io_add_buffers(p, &head);
4053 if (ret >= 0 && !list) {
4054 ret = xa_insert(&ctx->io_buffers, p->bgid, head, GFP_KERNEL);
4056 __io_remove_buffers(ctx, head, p->bgid, -1U);
4060 /* complete before unlock, IOPOLL may need the lock */
4061 __io_req_complete(req, issue_flags, ret, 0);
4062 io_ring_submit_unlock(ctx, !force_nonblock);
4066 static int io_epoll_ctl_prep(struct io_kiocb *req,
4067 const struct io_uring_sqe *sqe)
4069 #if defined(CONFIG_EPOLL)
4070 if (sqe->ioprio || sqe->buf_index)
4072 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4075 req->epoll.epfd = READ_ONCE(sqe->fd);
4076 req->epoll.op = READ_ONCE(sqe->len);
4077 req->epoll.fd = READ_ONCE(sqe->off);
4079 if (ep_op_has_event(req->epoll.op)) {
4080 struct epoll_event __user *ev;
4082 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4083 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4093 static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4095 #if defined(CONFIG_EPOLL)
4096 struct io_epoll *ie = &req->epoll;
4098 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4100 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4101 if (force_nonblock && ret == -EAGAIN)
4106 __io_req_complete(req, issue_flags, ret, 0);
4113 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4115 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4116 if (sqe->ioprio || sqe->buf_index || sqe->off)
4118 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4121 req->madvise.addr = READ_ONCE(sqe->addr);
4122 req->madvise.len = READ_ONCE(sqe->len);
4123 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4130 static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4132 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4133 struct io_madvise *ma = &req->madvise;
4136 if (issue_flags & IO_URING_F_NONBLOCK)
4139 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4142 io_req_complete(req, ret);
4149 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4151 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4153 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4156 req->fadvise.offset = READ_ONCE(sqe->off);
4157 req->fadvise.len = READ_ONCE(sqe->len);
4158 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4162 static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4164 struct io_fadvise *fa = &req->fadvise;
4167 if (issue_flags & IO_URING_F_NONBLOCK) {
4168 switch (fa->advice) {
4169 case POSIX_FADV_NORMAL:
4170 case POSIX_FADV_RANDOM:
4171 case POSIX_FADV_SEQUENTIAL:
4178 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4181 __io_req_complete(req, issue_flags, ret, 0);
4185 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4187 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4189 if (sqe->ioprio || sqe->buf_index)
4191 if (req->flags & REQ_F_FIXED_FILE)
4194 req->statx.dfd = READ_ONCE(sqe->fd);
4195 req->statx.mask = READ_ONCE(sqe->len);
4196 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4197 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4198 req->statx.flags = READ_ONCE(sqe->statx_flags);
4203 static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4205 struct io_statx *ctx = &req->statx;
4208 if (issue_flags & IO_URING_F_NONBLOCK)
4211 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4216 io_req_complete(req, ret);
4220 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4222 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4224 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4225 sqe->rw_flags || sqe->buf_index)
4227 if (req->flags & REQ_F_FIXED_FILE)
4230 req->close.fd = READ_ONCE(sqe->fd);
4234 static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4236 struct files_struct *files = current->files;
4237 struct io_close *close = &req->close;
4238 struct fdtable *fdt;
4239 struct file *file = NULL;
4242 spin_lock(&files->file_lock);
4243 fdt = files_fdtable(files);
4244 if (close->fd >= fdt->max_fds) {
4245 spin_unlock(&files->file_lock);
4248 file = fdt->fd[close->fd];
4249 if (!file || file->f_op == &io_uring_fops) {
4250 spin_unlock(&files->file_lock);
4255 /* if the file has a flush method, be safe and punt to async */
4256 if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4257 spin_unlock(&files->file_lock);
4261 ret = __close_fd_get_file(close->fd, &file);
4262 spin_unlock(&files->file_lock);
4269 /* No ->flush() or already async, safely close from here */
4270 ret = filp_close(file, current->files);
4276 __io_req_complete(req, issue_flags, ret, 0);
4280 static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4282 struct io_ring_ctx *ctx = req->ctx;
4284 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4286 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4289 req->sync.off = READ_ONCE(sqe->off);
4290 req->sync.len = READ_ONCE(sqe->len);
4291 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4295 static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4299 /* sync_file_range always requires a blocking context */
4300 if (issue_flags & IO_URING_F_NONBLOCK)
4303 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4307 io_req_complete(req, ret);
4311 #if defined(CONFIG_NET)
4312 static int io_setup_async_msg(struct io_kiocb *req,
4313 struct io_async_msghdr *kmsg)
4315 struct io_async_msghdr *async_msg = req->async_data;
4319 if (io_alloc_async_data(req)) {
4320 kfree(kmsg->free_iov);
4323 async_msg = req->async_data;
4324 req->flags |= REQ_F_NEED_CLEANUP;
4325 memcpy(async_msg, kmsg, sizeof(*kmsg));
4326 async_msg->msg.msg_name = &async_msg->addr;
4327 /* if were using fast_iov, set it to the new one */
4328 if (!async_msg->free_iov)
4329 async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4334 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4335 struct io_async_msghdr *iomsg)
4337 iomsg->msg.msg_name = &iomsg->addr;
4338 iomsg->free_iov = iomsg->fast_iov;
4339 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4340 req->sr_msg.msg_flags, &iomsg->free_iov);
4343 static int io_sendmsg_prep_async(struct io_kiocb *req)
4347 ret = io_sendmsg_copy_hdr(req, req->async_data);
4349 req->flags |= REQ_F_NEED_CLEANUP;
4353 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4355 struct io_sr_msg *sr = &req->sr_msg;
4357 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4360 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4361 sr->len = READ_ONCE(sqe->len);
4362 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4363 if (sr->msg_flags & MSG_DONTWAIT)
4364 req->flags |= REQ_F_NOWAIT;
4366 #ifdef CONFIG_COMPAT
4367 if (req->ctx->compat)
4368 sr->msg_flags |= MSG_CMSG_COMPAT;
4373 static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4375 struct io_async_msghdr iomsg, *kmsg;
4376 struct socket *sock;
4381 sock = sock_from_file(req->file);
4382 if (unlikely(!sock))
4385 kmsg = req->async_data;
4387 ret = io_sendmsg_copy_hdr(req, &iomsg);
4393 flags = req->sr_msg.msg_flags;
4394 if (issue_flags & IO_URING_F_NONBLOCK)
4395 flags |= MSG_DONTWAIT;
4396 if (flags & MSG_WAITALL)
4397 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4399 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4400 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4401 return io_setup_async_msg(req, kmsg);
4402 if (ret == -ERESTARTSYS)
4405 /* fast path, check for non-NULL to avoid function call */
4407 kfree(kmsg->free_iov);
4408 req->flags &= ~REQ_F_NEED_CLEANUP;
4411 __io_req_complete(req, issue_flags, ret, 0);
4415 static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4417 struct io_sr_msg *sr = &req->sr_msg;
4420 struct socket *sock;
4425 sock = sock_from_file(req->file);
4426 if (unlikely(!sock))
4429 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4433 msg.msg_name = NULL;
4434 msg.msg_control = NULL;
4435 msg.msg_controllen = 0;
4436 msg.msg_namelen = 0;
4438 flags = req->sr_msg.msg_flags;
4439 if (issue_flags & IO_URING_F_NONBLOCK)
4440 flags |= MSG_DONTWAIT;
4441 if (flags & MSG_WAITALL)
4442 min_ret = iov_iter_count(&msg.msg_iter);
4444 msg.msg_flags = flags;
4445 ret = sock_sendmsg(sock, &msg);
4446 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4448 if (ret == -ERESTARTSYS)
4453 __io_req_complete(req, issue_flags, ret, 0);
4457 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4458 struct io_async_msghdr *iomsg)
4460 struct io_sr_msg *sr = &req->sr_msg;
4461 struct iovec __user *uiov;
4465 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4466 &iomsg->uaddr, &uiov, &iov_len);
4470 if (req->flags & REQ_F_BUFFER_SELECT) {
4473 if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4475 sr->len = iomsg->fast_iov[0].iov_len;
4476 iomsg->free_iov = NULL;
4478 iomsg->free_iov = iomsg->fast_iov;
4479 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4480 &iomsg->free_iov, &iomsg->msg.msg_iter,
4489 #ifdef CONFIG_COMPAT
4490 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4491 struct io_async_msghdr *iomsg)
4493 struct io_sr_msg *sr = &req->sr_msg;
4494 struct compat_iovec __user *uiov;
4499 ret = __get_compat_msghdr(&iomsg->msg, sr->umsg_compat, &iomsg->uaddr,
4504 uiov = compat_ptr(ptr);
4505 if (req->flags & REQ_F_BUFFER_SELECT) {
4506 compat_ssize_t clen;
4510 if (!access_ok(uiov, sizeof(*uiov)))
4512 if (__get_user(clen, &uiov->iov_len))
4517 iomsg->free_iov = NULL;
4519 iomsg->free_iov = iomsg->fast_iov;
4520 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4521 UIO_FASTIOV, &iomsg->free_iov,
4522 &iomsg->msg.msg_iter, true);
4531 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4532 struct io_async_msghdr *iomsg)
4534 iomsg->msg.msg_name = &iomsg->addr;
4536 #ifdef CONFIG_COMPAT
4537 if (req->ctx->compat)
4538 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4541 return __io_recvmsg_copy_hdr(req, iomsg);
4544 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4547 struct io_sr_msg *sr = &req->sr_msg;
4548 struct io_buffer *kbuf;
4550 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4555 req->flags |= REQ_F_BUFFER_SELECTED;
4559 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4561 return io_put_kbuf(req, req->sr_msg.kbuf);
4564 static int io_recvmsg_prep_async(struct io_kiocb *req)
4568 ret = io_recvmsg_copy_hdr(req, req->async_data);
4570 req->flags |= REQ_F_NEED_CLEANUP;
4574 static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4576 struct io_sr_msg *sr = &req->sr_msg;
4578 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4581 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4582 sr->len = READ_ONCE(sqe->len);
4583 sr->bgid = READ_ONCE(sqe->buf_group);
4584 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4585 if (sr->msg_flags & MSG_DONTWAIT)
4586 req->flags |= REQ_F_NOWAIT;
4588 #ifdef CONFIG_COMPAT
4589 if (req->ctx->compat)
4590 sr->msg_flags |= MSG_CMSG_COMPAT;
4595 static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
4597 struct io_async_msghdr iomsg, *kmsg;
4598 struct socket *sock;
4599 struct io_buffer *kbuf;
4602 int ret, cflags = 0;
4603 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4605 sock = sock_from_file(req->file);
4606 if (unlikely(!sock))
4609 kmsg = req->async_data;
4611 ret = io_recvmsg_copy_hdr(req, &iomsg);
4617 if (req->flags & REQ_F_BUFFER_SELECT) {
4618 kbuf = io_recv_buffer_select(req, !force_nonblock);
4620 return PTR_ERR(kbuf);
4621 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4622 kmsg->fast_iov[0].iov_len = req->sr_msg.len;
4623 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
4624 1, req->sr_msg.len);
4627 flags = req->sr_msg.msg_flags;
4629 flags |= MSG_DONTWAIT;
4630 if (flags & MSG_WAITALL)
4631 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4633 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4634 kmsg->uaddr, flags);
4635 if (force_nonblock && ret == -EAGAIN)
4636 return io_setup_async_msg(req, kmsg);
4637 if (ret == -ERESTARTSYS)
4640 if (req->flags & REQ_F_BUFFER_SELECTED)
4641 cflags = io_put_recv_kbuf(req);
4642 /* fast path, check for non-NULL to avoid function call */
4644 kfree(kmsg->free_iov);
4645 req->flags &= ~REQ_F_NEED_CLEANUP;
4646 if (ret < min_ret || ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4648 __io_req_complete(req, issue_flags, ret, cflags);
4652 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
4654 struct io_buffer *kbuf;
4655 struct io_sr_msg *sr = &req->sr_msg;
4657 void __user *buf = sr->buf;
4658 struct socket *sock;
4662 int ret, cflags = 0;
4663 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4665 sock = sock_from_file(req->file);
4666 if (unlikely(!sock))
4669 if (req->flags & REQ_F_BUFFER_SELECT) {
4670 kbuf = io_recv_buffer_select(req, !force_nonblock);
4672 return PTR_ERR(kbuf);
4673 buf = u64_to_user_ptr(kbuf->addr);
4676 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4680 msg.msg_name = NULL;
4681 msg.msg_control = NULL;
4682 msg.msg_controllen = 0;
4683 msg.msg_namelen = 0;
4684 msg.msg_iocb = NULL;
4687 flags = req->sr_msg.msg_flags;
4689 flags |= MSG_DONTWAIT;
4690 if (flags & MSG_WAITALL)
4691 min_ret = iov_iter_count(&msg.msg_iter);
4693 ret = sock_recvmsg(sock, &msg, flags);
4694 if (force_nonblock && ret == -EAGAIN)
4696 if (ret == -ERESTARTSYS)
4699 if (req->flags & REQ_F_BUFFER_SELECTED)
4700 cflags = io_put_recv_kbuf(req);
4701 if (ret < min_ret || ((flags & MSG_WAITALL) && (msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4703 __io_req_complete(req, issue_flags, ret, cflags);
4707 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4709 struct io_accept *accept = &req->accept;
4711 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4713 if (sqe->ioprio || sqe->len || sqe->buf_index)
4716 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4717 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4718 accept->flags = READ_ONCE(sqe->accept_flags);
4719 accept->nofile = rlimit(RLIMIT_NOFILE);
4723 static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
4725 struct io_accept *accept = &req->accept;
4726 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4727 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4730 if (req->file->f_flags & O_NONBLOCK)
4731 req->flags |= REQ_F_NOWAIT;
4733 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4734 accept->addr_len, accept->flags,
4736 if (ret == -EAGAIN && force_nonblock)
4739 if (ret == -ERESTARTSYS)
4743 __io_req_complete(req, issue_flags, ret, 0);
4747 static int io_connect_prep_async(struct io_kiocb *req)
4749 struct io_async_connect *io = req->async_data;
4750 struct io_connect *conn = &req->connect;
4752 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
4755 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4757 struct io_connect *conn = &req->connect;
4759 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4761 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4764 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4765 conn->addr_len = READ_ONCE(sqe->addr2);
4769 static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
4771 struct io_async_connect __io, *io;
4772 unsigned file_flags;
4774 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4776 if (req->async_data) {
4777 io = req->async_data;
4779 ret = move_addr_to_kernel(req->connect.addr,
4780 req->connect.addr_len,
4787 file_flags = force_nonblock ? O_NONBLOCK : 0;
4789 ret = __sys_connect_file(req->file, &io->address,
4790 req->connect.addr_len, file_flags);
4791 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4792 if (req->async_data)
4794 if (io_alloc_async_data(req)) {
4798 memcpy(req->async_data, &__io, sizeof(__io));
4801 if (ret == -ERESTARTSYS)
4806 __io_req_complete(req, issue_flags, ret, 0);
4809 #else /* !CONFIG_NET */
4810 #define IO_NETOP_FN(op) \
4811 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
4813 return -EOPNOTSUPP; \
4816 #define IO_NETOP_PREP(op) \
4818 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
4820 return -EOPNOTSUPP; \
4823 #define IO_NETOP_PREP_ASYNC(op) \
4825 static int io_##op##_prep_async(struct io_kiocb *req) \
4827 return -EOPNOTSUPP; \
4830 IO_NETOP_PREP_ASYNC(sendmsg);
4831 IO_NETOP_PREP_ASYNC(recvmsg);
4832 IO_NETOP_PREP_ASYNC(connect);
4833 IO_NETOP_PREP(accept);
4836 #endif /* CONFIG_NET */
4838 struct io_poll_table {
4839 struct poll_table_struct pt;
4840 struct io_kiocb *req;
4844 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4845 __poll_t mask, task_work_func_t func)
4849 /* for instances that support it check for an event match first: */
4850 if (mask && !(mask & poll->events))
4853 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4855 list_del_init(&poll->wait.entry);
4858 req->task_work.func = func;
4861 * If this fails, then the task is exiting. When a task exits, the
4862 * work gets canceled, so just cancel this request as well instead
4863 * of executing it. We can't safely execute it anyway, as we may not
4864 * have the needed state needed for it anyway.
4866 ret = io_req_task_work_add(req);
4867 if (unlikely(ret)) {
4868 WRITE_ONCE(poll->canceled, true);
4869 io_req_task_work_add_fallback(req, func);
4874 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4875 __acquires(&req->ctx->completion_lock)
4877 struct io_ring_ctx *ctx = req->ctx;
4879 if (!req->result && !READ_ONCE(poll->canceled)) {
4880 struct poll_table_struct pt = { ._key = poll->events };
4882 req->result = vfs_poll(req->file, &pt) & poll->events;
4885 spin_lock_irq(&ctx->completion_lock);
4886 if (!req->result && !READ_ONCE(poll->canceled)) {
4887 add_wait_queue(poll->head, &poll->wait);
4894 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4896 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4897 if (req->opcode == IORING_OP_POLL_ADD)
4898 return req->async_data;
4899 return req->apoll->double_poll;
4902 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4904 if (req->opcode == IORING_OP_POLL_ADD)
4906 return &req->apoll->poll;
4909 static void io_poll_remove_double(struct io_kiocb *req)
4910 __must_hold(&req->ctx->completion_lock)
4912 struct io_poll_iocb *poll = io_poll_get_double(req);
4914 lockdep_assert_held(&req->ctx->completion_lock);
4916 if (poll && poll->head) {
4917 struct wait_queue_head *head = poll->head;
4919 spin_lock(&head->lock);
4920 list_del_init(&poll->wait.entry);
4921 if (poll->wait.private)
4924 spin_unlock(&head->lock);
4928 static bool io_poll_complete(struct io_kiocb *req, __poll_t mask)
4929 __must_hold(&req->ctx->completion_lock)
4931 struct io_ring_ctx *ctx = req->ctx;
4932 unsigned flags = IORING_CQE_F_MORE;
4935 if (READ_ONCE(req->poll.canceled)) {
4937 req->poll.events |= EPOLLONESHOT;
4939 error = mangle_poll(mask);
4941 if (req->poll.events & EPOLLONESHOT)
4943 if (!io_cqring_fill_event(ctx, req->user_data, error, flags)) {
4944 io_poll_remove_waitqs(req);
4945 req->poll.done = true;
4948 if (flags & IORING_CQE_F_MORE)
4951 io_commit_cqring(ctx);
4952 return !(flags & IORING_CQE_F_MORE);
4955 static void io_poll_task_func(struct callback_head *cb)
4957 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4958 struct io_ring_ctx *ctx = req->ctx;
4959 struct io_kiocb *nxt;
4961 if (io_poll_rewait(req, &req->poll)) {
4962 spin_unlock_irq(&ctx->completion_lock);
4966 done = io_poll_complete(req, req->result);
4968 hash_del(&req->hash_node);
4971 add_wait_queue(req->poll.head, &req->poll.wait);
4973 spin_unlock_irq(&ctx->completion_lock);
4974 io_cqring_ev_posted(ctx);
4977 nxt = io_put_req_find_next(req);
4979 __io_req_task_submit(nxt);
4984 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4985 int sync, void *key)
4987 struct io_kiocb *req = wait->private;
4988 struct io_poll_iocb *poll = io_poll_get_single(req);
4989 __poll_t mask = key_to_poll(key);
4991 /* for instances that support it check for an event match first: */
4992 if (mask && !(mask & poll->events))
4994 if (!(poll->events & EPOLLONESHOT))
4995 return poll->wait.func(&poll->wait, mode, sync, key);
4997 list_del_init(&wait->entry);
4999 if (poll && poll->head) {
5002 spin_lock(&poll->head->lock);
5003 done = list_empty(&poll->wait.entry);
5005 list_del_init(&poll->wait.entry);
5006 /* make sure double remove sees this as being gone */
5007 wait->private = NULL;
5008 spin_unlock(&poll->head->lock);
5010 /* use wait func handler, so it matches the rq type */
5011 poll->wait.func(&poll->wait, mode, sync, key);
5018 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
5019 wait_queue_func_t wake_func)
5023 poll->canceled = false;
5024 #define IO_POLL_UNMASK (EPOLLERR|EPOLLHUP|EPOLLNVAL|EPOLLRDHUP)
5025 /* mask in events that we always want/need */
5026 poll->events = events | IO_POLL_UNMASK;
5027 INIT_LIST_HEAD(&poll->wait.entry);
5028 init_waitqueue_func_entry(&poll->wait, wake_func);
5031 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5032 struct wait_queue_head *head,
5033 struct io_poll_iocb **poll_ptr)
5035 struct io_kiocb *req = pt->req;
5038 * If poll->head is already set, it's because the file being polled
5039 * uses multiple waitqueues for poll handling (eg one for read, one
5040 * for write). Setup a separate io_poll_iocb if this happens.
5042 if (unlikely(poll->head)) {
5043 struct io_poll_iocb *poll_one = poll;
5045 /* already have a 2nd entry, fail a third attempt */
5047 pt->error = -EINVAL;
5051 * Can't handle multishot for double wait for now, turn it
5052 * into one-shot mode.
5054 if (!(poll_one->events & EPOLLONESHOT))
5055 poll_one->events |= EPOLLONESHOT;
5056 /* double add on the same waitqueue head, ignore */
5057 if (poll_one->head == head)
5059 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5061 pt->error = -ENOMEM;
5064 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5066 poll->wait.private = req;
5073 if (poll->events & EPOLLEXCLUSIVE)
5074 add_wait_queue_exclusive(head, &poll->wait);
5076 add_wait_queue(head, &poll->wait);
5079 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5080 struct poll_table_struct *p)
5082 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5083 struct async_poll *apoll = pt->req->apoll;
5085 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5088 static void io_async_task_func(struct callback_head *cb)
5090 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
5091 struct async_poll *apoll = req->apoll;
5092 struct io_ring_ctx *ctx = req->ctx;
5094 trace_io_uring_task_run(req->ctx, req, req->opcode, req->user_data);
5096 if (io_poll_rewait(req, &apoll->poll)) {
5097 spin_unlock_irq(&ctx->completion_lock);
5101 hash_del(&req->hash_node);
5102 io_poll_remove_double(req);
5103 spin_unlock_irq(&ctx->completion_lock);
5105 if (!READ_ONCE(apoll->poll.canceled))
5106 __io_req_task_submit(req);
5108 io_req_complete_failed(req, -ECANCELED);
5111 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5114 struct io_kiocb *req = wait->private;
5115 struct io_poll_iocb *poll = &req->apoll->poll;
5117 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5120 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5123 static void io_poll_req_insert(struct io_kiocb *req)
5125 struct io_ring_ctx *ctx = req->ctx;
5126 struct hlist_head *list;
5128 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5129 hlist_add_head(&req->hash_node, list);
5132 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5133 struct io_poll_iocb *poll,
5134 struct io_poll_table *ipt, __poll_t mask,
5135 wait_queue_func_t wake_func)
5136 __acquires(&ctx->completion_lock)
5138 struct io_ring_ctx *ctx = req->ctx;
5139 bool cancel = false;
5141 INIT_HLIST_NODE(&req->hash_node);
5142 io_init_poll_iocb(poll, mask, wake_func);
5143 poll->file = req->file;
5144 poll->wait.private = req;
5146 ipt->pt._key = mask;
5148 ipt->error = -EINVAL;
5150 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5152 spin_lock_irq(&ctx->completion_lock);
5153 if (likely(poll->head)) {
5154 spin_lock(&poll->head->lock);
5155 if (unlikely(list_empty(&poll->wait.entry))) {
5161 if ((mask && (poll->events & EPOLLONESHOT)) || ipt->error)
5162 list_del_init(&poll->wait.entry);
5164 WRITE_ONCE(poll->canceled, true);
5165 else if (!poll->done) /* actually waiting for an event */
5166 io_poll_req_insert(req);
5167 spin_unlock(&poll->head->lock);
5179 static int io_arm_poll_handler(struct io_kiocb *req)
5181 const struct io_op_def *def = &io_op_defs[req->opcode];
5182 struct io_ring_ctx *ctx = req->ctx;
5183 struct async_poll *apoll;
5184 struct io_poll_table ipt;
5185 __poll_t ret, mask = EPOLLONESHOT | POLLERR | POLLPRI;
5188 if (!req->file || !file_can_poll(req->file))
5189 return IO_APOLL_ABORTED;
5190 if (req->flags & REQ_F_POLLED)
5191 return IO_APOLL_ABORTED;
5192 if (!def->pollin && !def->pollout)
5193 return IO_APOLL_ABORTED;
5197 mask |= POLLIN | POLLRDNORM;
5199 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5200 if ((req->opcode == IORING_OP_RECVMSG) &&
5201 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5205 mask |= POLLOUT | POLLWRNORM;
5208 /* if we can't nonblock try, then no point in arming a poll handler */
5209 if (!io_file_supports_async(req, rw))
5210 return IO_APOLL_ABORTED;
5212 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5213 if (unlikely(!apoll))
5214 return IO_APOLL_ABORTED;
5215 apoll->double_poll = NULL;
5217 req->flags |= REQ_F_POLLED;
5218 ipt.pt._qproc = io_async_queue_proc;
5220 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5222 if (ret || ipt.error) {
5223 io_poll_remove_double(req);
5224 spin_unlock_irq(&ctx->completion_lock);
5226 return IO_APOLL_READY;
5227 return IO_APOLL_ABORTED;
5229 spin_unlock_irq(&ctx->completion_lock);
5230 trace_io_uring_poll_arm(ctx, req, req->opcode, req->user_data,
5231 mask, apoll->poll.events);
5235 static bool __io_poll_remove_one(struct io_kiocb *req,
5236 struct io_poll_iocb *poll, bool do_cancel)
5237 __must_hold(&req->ctx->completion_lock)
5239 bool do_complete = false;
5243 spin_lock(&poll->head->lock);
5245 WRITE_ONCE(poll->canceled, true);
5246 if (!list_empty(&poll->wait.entry)) {
5247 list_del_init(&poll->wait.entry);
5250 spin_unlock(&poll->head->lock);
5251 hash_del(&req->hash_node);
5255 static bool io_poll_remove_waitqs(struct io_kiocb *req)
5256 __must_hold(&req->ctx->completion_lock)
5260 io_poll_remove_double(req);
5261 do_complete = __io_poll_remove_one(req, io_poll_get_single(req), true);
5263 if (req->opcode != IORING_OP_POLL_ADD && do_complete) {
5264 /* non-poll requests have submit ref still */
5270 static bool io_poll_remove_one(struct io_kiocb *req)
5271 __must_hold(&req->ctx->completion_lock)
5275 do_complete = io_poll_remove_waitqs(req);
5277 io_cqring_fill_event(req->ctx, req->user_data, -ECANCELED, 0);
5278 io_commit_cqring(req->ctx);
5280 io_put_req_deferred(req, 1);
5287 * Returns true if we found and killed one or more poll requests
5289 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5292 struct hlist_node *tmp;
5293 struct io_kiocb *req;
5296 spin_lock_irq(&ctx->completion_lock);
5297 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5298 struct hlist_head *list;
5300 list = &ctx->cancel_hash[i];
5301 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5302 if (io_match_task(req, tsk, cancel_all))
5303 posted += io_poll_remove_one(req);
5306 spin_unlock_irq(&ctx->completion_lock);
5309 io_cqring_ev_posted(ctx);
5314 static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, __u64 sqe_addr,
5316 __must_hold(&ctx->completion_lock)
5318 struct hlist_head *list;
5319 struct io_kiocb *req;
5321 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5322 hlist_for_each_entry(req, list, hash_node) {
5323 if (sqe_addr != req->user_data)
5325 if (poll_only && req->opcode != IORING_OP_POLL_ADD)
5332 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr,
5334 __must_hold(&ctx->completion_lock)
5336 struct io_kiocb *req;
5338 req = io_poll_find(ctx, sqe_addr, poll_only);
5341 if (io_poll_remove_one(req))
5347 static __poll_t io_poll_parse_events(const struct io_uring_sqe *sqe,
5352 events = READ_ONCE(sqe->poll32_events);
5354 events = swahw32(events);
5356 if (!(flags & IORING_POLL_ADD_MULTI))
5357 events |= EPOLLONESHOT;
5358 return demangle_poll(events) | (events & (EPOLLEXCLUSIVE|EPOLLONESHOT));
5361 static int io_poll_update_prep(struct io_kiocb *req,
5362 const struct io_uring_sqe *sqe)
5364 struct io_poll_update *upd = &req->poll_update;
5367 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5369 if (sqe->ioprio || sqe->buf_index)
5371 flags = READ_ONCE(sqe->len);
5372 if (flags & ~(IORING_POLL_UPDATE_EVENTS | IORING_POLL_UPDATE_USER_DATA |
5373 IORING_POLL_ADD_MULTI))
5375 /* meaningless without update */
5376 if (flags == IORING_POLL_ADD_MULTI)
5379 upd->old_user_data = READ_ONCE(sqe->addr);
5380 upd->update_events = flags & IORING_POLL_UPDATE_EVENTS;
5381 upd->update_user_data = flags & IORING_POLL_UPDATE_USER_DATA;
5383 upd->new_user_data = READ_ONCE(sqe->off);
5384 if (!upd->update_user_data && upd->new_user_data)
5386 if (upd->update_events)
5387 upd->events = io_poll_parse_events(sqe, flags);
5388 else if (sqe->poll32_events)
5394 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5397 struct io_kiocb *req = wait->private;
5398 struct io_poll_iocb *poll = &req->poll;
5400 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5403 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5404 struct poll_table_struct *p)
5406 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5408 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5411 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5413 struct io_poll_iocb *poll = &req->poll;
5416 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5418 if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->addr)
5420 flags = READ_ONCE(sqe->len);
5421 if (flags & ~IORING_POLL_ADD_MULTI)
5424 poll->events = io_poll_parse_events(sqe, flags);
5428 static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5430 struct io_poll_iocb *poll = &req->poll;
5431 struct io_ring_ctx *ctx = req->ctx;
5432 struct io_poll_table ipt;
5435 ipt.pt._qproc = io_poll_queue_proc;
5437 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5440 if (mask) { /* no async, we'd stolen it */
5442 io_poll_complete(req, mask);
5444 spin_unlock_irq(&ctx->completion_lock);
5447 io_cqring_ev_posted(ctx);
5448 if (poll->events & EPOLLONESHOT)
5454 static int io_poll_update(struct io_kiocb *req, unsigned int issue_flags)
5456 struct io_ring_ctx *ctx = req->ctx;
5457 struct io_kiocb *preq;
5461 spin_lock_irq(&ctx->completion_lock);
5462 preq = io_poll_find(ctx, req->poll_update.old_user_data, true);
5468 if (!req->poll_update.update_events && !req->poll_update.update_user_data) {
5470 ret = io_poll_remove_one(preq) ? 0 : -EALREADY;
5475 * Don't allow racy completion with singleshot, as we cannot safely
5476 * update those. For multishot, if we're racing with completion, just
5477 * let completion re-add it.
5479 completing = !__io_poll_remove_one(preq, &preq->poll, false);
5480 if (completing && (preq->poll.events & EPOLLONESHOT)) {
5484 /* we now have a detached poll request. reissue. */
5488 spin_unlock_irq(&ctx->completion_lock);
5490 io_req_complete(req, ret);
5493 /* only mask one event flags, keep behavior flags */
5494 if (req->poll_update.update_events) {
5495 preq->poll.events &= ~0xffff;
5496 preq->poll.events |= req->poll_update.events & 0xffff;
5497 preq->poll.events |= IO_POLL_UNMASK;
5499 if (req->poll_update.update_user_data)
5500 preq->user_data = req->poll_update.new_user_data;
5501 spin_unlock_irq(&ctx->completion_lock);
5503 /* complete update request, we're done with it */
5504 io_req_complete(req, ret);
5507 ret = io_poll_add(preq, issue_flags);
5510 io_req_complete(preq, ret);
5516 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5518 struct io_timeout_data *data = container_of(timer,
5519 struct io_timeout_data, timer);
5520 struct io_kiocb *req = data->req;
5521 struct io_ring_ctx *ctx = req->ctx;
5522 unsigned long flags;
5524 spin_lock_irqsave(&ctx->completion_lock, flags);
5525 list_del_init(&req->timeout.list);
5526 atomic_set(&req->ctx->cq_timeouts,
5527 atomic_read(&req->ctx->cq_timeouts) + 1);
5529 io_cqring_fill_event(ctx, req->user_data, -ETIME, 0);
5530 io_commit_cqring(ctx);
5531 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5533 io_cqring_ev_posted(ctx);
5536 return HRTIMER_NORESTART;
5539 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5541 __must_hold(&ctx->completion_lock)
5543 struct io_timeout_data *io;
5544 struct io_kiocb *req;
5547 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5548 found = user_data == req->user_data;
5553 return ERR_PTR(-ENOENT);
5555 io = req->async_data;
5556 if (hrtimer_try_to_cancel(&io->timer) == -1)
5557 return ERR_PTR(-EALREADY);
5558 list_del_init(&req->timeout.list);
5562 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5563 __must_hold(&ctx->completion_lock)
5565 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5568 return PTR_ERR(req);
5571 io_cqring_fill_event(ctx, req->user_data, -ECANCELED, 0);
5572 io_put_req_deferred(req, 1);
5576 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5577 struct timespec64 *ts, enum hrtimer_mode mode)
5578 __must_hold(&ctx->completion_lock)
5580 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5581 struct io_timeout_data *data;
5584 return PTR_ERR(req);
5586 req->timeout.off = 0; /* noseq */
5587 data = req->async_data;
5588 list_add_tail(&req->timeout.list, &ctx->timeout_list);
5589 hrtimer_init(&data->timer, CLOCK_MONOTONIC, mode);
5590 data->timer.function = io_timeout_fn;
5591 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
5595 static int io_timeout_remove_prep(struct io_kiocb *req,
5596 const struct io_uring_sqe *sqe)
5598 struct io_timeout_rem *tr = &req->timeout_rem;
5600 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5602 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5604 if (sqe->ioprio || sqe->buf_index || sqe->len)
5607 tr->addr = READ_ONCE(sqe->addr);
5608 tr->flags = READ_ONCE(sqe->timeout_flags);
5609 if (tr->flags & IORING_TIMEOUT_UPDATE) {
5610 if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS))
5612 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
5614 } else if (tr->flags) {
5615 /* timeout removal doesn't support flags */
5622 static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
5624 return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
5629 * Remove or update an existing timeout command
5631 static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
5633 struct io_timeout_rem *tr = &req->timeout_rem;
5634 struct io_ring_ctx *ctx = req->ctx;
5637 spin_lock_irq(&ctx->completion_lock);
5638 if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE))
5639 ret = io_timeout_cancel(ctx, tr->addr);
5641 ret = io_timeout_update(ctx, tr->addr, &tr->ts,
5642 io_translate_timeout_mode(tr->flags));
5644 io_cqring_fill_event(ctx, req->user_data, ret, 0);
5645 io_commit_cqring(ctx);
5646 spin_unlock_irq(&ctx->completion_lock);
5647 io_cqring_ev_posted(ctx);
5654 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5655 bool is_timeout_link)
5657 struct io_timeout_data *data;
5659 u32 off = READ_ONCE(sqe->off);
5661 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5663 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5665 if (off && is_timeout_link)
5667 flags = READ_ONCE(sqe->timeout_flags);
5668 if (flags & ~IORING_TIMEOUT_ABS)
5671 req->timeout.off = off;
5672 if (unlikely(off && !req->ctx->off_timeout_used))
5673 req->ctx->off_timeout_used = true;
5675 if (!req->async_data && io_alloc_async_data(req))
5678 data = req->async_data;
5681 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5684 data->mode = io_translate_timeout_mode(flags);
5685 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5686 if (is_timeout_link)
5687 io_req_track_inflight(req);
5691 static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
5693 struct io_ring_ctx *ctx = req->ctx;
5694 struct io_timeout_data *data = req->async_data;
5695 struct list_head *entry;
5696 u32 tail, off = req->timeout.off;
5698 spin_lock_irq(&ctx->completion_lock);
5701 * sqe->off holds how many events that need to occur for this
5702 * timeout event to be satisfied. If it isn't set, then this is
5703 * a pure timeout request, sequence isn't used.
5705 if (io_is_timeout_noseq(req)) {
5706 entry = ctx->timeout_list.prev;
5710 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5711 req->timeout.target_seq = tail + off;
5713 /* Update the last seq here in case io_flush_timeouts() hasn't.
5714 * This is safe because ->completion_lock is held, and submissions
5715 * and completions are never mixed in the same ->completion_lock section.
5717 ctx->cq_last_tm_flush = tail;
5720 * Insertion sort, ensuring the first entry in the list is always
5721 * the one we need first.
5723 list_for_each_prev(entry, &ctx->timeout_list) {
5724 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5727 if (io_is_timeout_noseq(nxt))
5729 /* nxt.seq is behind @tail, otherwise would've been completed */
5730 if (off >= nxt->timeout.target_seq - tail)
5734 list_add(&req->timeout.list, entry);
5735 data->timer.function = io_timeout_fn;
5736 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5737 spin_unlock_irq(&ctx->completion_lock);
5741 struct io_cancel_data {
5742 struct io_ring_ctx *ctx;
5746 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5748 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5749 struct io_cancel_data *cd = data;
5751 return req->ctx == cd->ctx && req->user_data == cd->user_data;
5754 static int io_async_cancel_one(struct io_uring_task *tctx, u64 user_data,
5755 struct io_ring_ctx *ctx)
5757 struct io_cancel_data data = { .ctx = ctx, .user_data = user_data, };
5758 enum io_wq_cancel cancel_ret;
5761 if (!tctx || !tctx->io_wq)
5764 cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, &data, false);
5765 switch (cancel_ret) {
5766 case IO_WQ_CANCEL_OK:
5769 case IO_WQ_CANCEL_RUNNING:
5772 case IO_WQ_CANCEL_NOTFOUND:
5780 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5781 struct io_kiocb *req, __u64 sqe_addr,
5784 unsigned long flags;
5787 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5788 spin_lock_irqsave(&ctx->completion_lock, flags);
5791 ret = io_timeout_cancel(ctx, sqe_addr);
5794 ret = io_poll_cancel(ctx, sqe_addr, false);
5798 io_cqring_fill_event(ctx, req->user_data, ret, 0);
5799 io_commit_cqring(ctx);
5800 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5801 io_cqring_ev_posted(ctx);
5807 static int io_async_cancel_prep(struct io_kiocb *req,
5808 const struct io_uring_sqe *sqe)
5810 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5812 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5814 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5817 req->cancel.addr = READ_ONCE(sqe->addr);
5821 static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
5823 struct io_ring_ctx *ctx = req->ctx;
5824 u64 sqe_addr = req->cancel.addr;
5825 struct io_tctx_node *node;
5828 /* tasks should wait for their io-wq threads, so safe w/o sync */
5829 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5830 spin_lock_irq(&ctx->completion_lock);
5833 ret = io_timeout_cancel(ctx, sqe_addr);
5836 ret = io_poll_cancel(ctx, sqe_addr, false);
5839 spin_unlock_irq(&ctx->completion_lock);
5841 /* slow path, try all io-wq's */
5842 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5844 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
5845 struct io_uring_task *tctx = node->task->io_uring;
5847 ret = io_async_cancel_one(tctx, req->cancel.addr, ctx);
5851 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5853 spin_lock_irq(&ctx->completion_lock);
5855 io_cqring_fill_event(ctx, req->user_data, ret, 0);
5856 io_commit_cqring(ctx);
5857 spin_unlock_irq(&ctx->completion_lock);
5858 io_cqring_ev_posted(ctx);
5866 static int io_rsrc_update_prep(struct io_kiocb *req,
5867 const struct io_uring_sqe *sqe)
5869 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5871 if (sqe->ioprio || sqe->rw_flags)
5874 req->rsrc_update.offset = READ_ONCE(sqe->off);
5875 req->rsrc_update.nr_args = READ_ONCE(sqe->len);
5876 if (!req->rsrc_update.nr_args)
5878 req->rsrc_update.arg = READ_ONCE(sqe->addr);
5882 static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
5884 struct io_ring_ctx *ctx = req->ctx;
5885 struct io_uring_rsrc_update2 up;
5888 if (issue_flags & IO_URING_F_NONBLOCK)
5891 up.offset = req->rsrc_update.offset;
5892 up.data = req->rsrc_update.arg;
5897 mutex_lock(&ctx->uring_lock);
5898 ret = __io_register_rsrc_update(ctx, IORING_RSRC_FILE,
5899 &up, req->rsrc_update.nr_args);
5900 mutex_unlock(&ctx->uring_lock);
5904 __io_req_complete(req, issue_flags, ret, 0);
5908 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5910 switch (req->opcode) {
5913 case IORING_OP_READV:
5914 case IORING_OP_READ_FIXED:
5915 case IORING_OP_READ:
5916 return io_read_prep(req, sqe);
5917 case IORING_OP_WRITEV:
5918 case IORING_OP_WRITE_FIXED:
5919 case IORING_OP_WRITE:
5920 return io_write_prep(req, sqe);
5921 case IORING_OP_POLL_ADD:
5922 return io_poll_add_prep(req, sqe);
5923 case IORING_OP_POLL_REMOVE:
5924 return io_poll_update_prep(req, sqe);
5925 case IORING_OP_FSYNC:
5926 return io_fsync_prep(req, sqe);
5927 case IORING_OP_SYNC_FILE_RANGE:
5928 return io_sfr_prep(req, sqe);
5929 case IORING_OP_SENDMSG:
5930 case IORING_OP_SEND:
5931 return io_sendmsg_prep(req, sqe);
5932 case IORING_OP_RECVMSG:
5933 case IORING_OP_RECV:
5934 return io_recvmsg_prep(req, sqe);
5935 case IORING_OP_CONNECT:
5936 return io_connect_prep(req, sqe);
5937 case IORING_OP_TIMEOUT:
5938 return io_timeout_prep(req, sqe, false);
5939 case IORING_OP_TIMEOUT_REMOVE:
5940 return io_timeout_remove_prep(req, sqe);
5941 case IORING_OP_ASYNC_CANCEL:
5942 return io_async_cancel_prep(req, sqe);
5943 case IORING_OP_LINK_TIMEOUT:
5944 return io_timeout_prep(req, sqe, true);
5945 case IORING_OP_ACCEPT:
5946 return io_accept_prep(req, sqe);
5947 case IORING_OP_FALLOCATE:
5948 return io_fallocate_prep(req, sqe);
5949 case IORING_OP_OPENAT:
5950 return io_openat_prep(req, sqe);
5951 case IORING_OP_CLOSE:
5952 return io_close_prep(req, sqe);
5953 case IORING_OP_FILES_UPDATE:
5954 return io_rsrc_update_prep(req, sqe);
5955 case IORING_OP_STATX:
5956 return io_statx_prep(req, sqe);
5957 case IORING_OP_FADVISE:
5958 return io_fadvise_prep(req, sqe);
5959 case IORING_OP_MADVISE:
5960 return io_madvise_prep(req, sqe);
5961 case IORING_OP_OPENAT2:
5962 return io_openat2_prep(req, sqe);
5963 case IORING_OP_EPOLL_CTL:
5964 return io_epoll_ctl_prep(req, sqe);
5965 case IORING_OP_SPLICE:
5966 return io_splice_prep(req, sqe);
5967 case IORING_OP_PROVIDE_BUFFERS:
5968 return io_provide_buffers_prep(req, sqe);
5969 case IORING_OP_REMOVE_BUFFERS:
5970 return io_remove_buffers_prep(req, sqe);
5972 return io_tee_prep(req, sqe);
5973 case IORING_OP_SHUTDOWN:
5974 return io_shutdown_prep(req, sqe);
5975 case IORING_OP_RENAMEAT:
5976 return io_renameat_prep(req, sqe);
5977 case IORING_OP_UNLINKAT:
5978 return io_unlinkat_prep(req, sqe);
5981 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5986 static int io_req_prep_async(struct io_kiocb *req)
5988 if (!io_op_defs[req->opcode].needs_async_setup)
5990 if (WARN_ON_ONCE(req->async_data))
5992 if (io_alloc_async_data(req))
5995 switch (req->opcode) {
5996 case IORING_OP_READV:
5997 return io_rw_prep_async(req, READ);
5998 case IORING_OP_WRITEV:
5999 return io_rw_prep_async(req, WRITE);
6000 case IORING_OP_SENDMSG:
6001 return io_sendmsg_prep_async(req);
6002 case IORING_OP_RECVMSG:
6003 return io_recvmsg_prep_async(req);
6004 case IORING_OP_CONNECT:
6005 return io_connect_prep_async(req);
6007 printk_once(KERN_WARNING "io_uring: prep_async() bad opcode %d\n",
6012 static u32 io_get_sequence(struct io_kiocb *req)
6014 u32 seq = req->ctx->cached_sq_head;
6016 /* need original cached_sq_head, but it was increased for each req */
6017 io_for_each_link(req, req)
6022 static bool io_drain_req(struct io_kiocb *req)
6024 struct io_kiocb *pos;
6025 struct io_ring_ctx *ctx = req->ctx;
6026 struct io_defer_entry *de;
6031 * If we need to drain a request in the middle of a link, drain the
6032 * head request and the next request/link after the current link.
6033 * Considering sequential execution of links, IOSQE_IO_DRAIN will be
6034 * maintained for every request of our link.
6036 if (ctx->drain_next) {
6037 req->flags |= REQ_F_IO_DRAIN;
6038 ctx->drain_next = false;
6040 /* not interested in head, start from the first linked */
6041 io_for_each_link(pos, req->link) {
6042 if (pos->flags & REQ_F_IO_DRAIN) {
6043 ctx->drain_next = true;
6044 req->flags |= REQ_F_IO_DRAIN;
6049 /* Still need defer if there is pending req in defer list. */
6050 if (likely(list_empty_careful(&ctx->defer_list) &&
6051 !(req->flags & REQ_F_IO_DRAIN))) {
6052 ctx->drain_active = false;
6056 seq = io_get_sequence(req);
6057 /* Still a chance to pass the sequence check */
6058 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
6061 ret = io_req_prep_async(req);
6064 io_prep_async_link(req);
6065 de = kmalloc(sizeof(*de), GFP_KERNEL);
6067 io_req_complete_failed(req, ret);
6071 spin_lock_irq(&ctx->completion_lock);
6072 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
6073 spin_unlock_irq(&ctx->completion_lock);
6075 io_queue_async_work(req);
6079 trace_io_uring_defer(ctx, req, req->user_data);
6082 list_add_tail(&de->list, &ctx->defer_list);
6083 spin_unlock_irq(&ctx->completion_lock);
6087 static void io_clean_op(struct io_kiocb *req)
6089 if (req->flags & REQ_F_BUFFER_SELECTED) {
6090 switch (req->opcode) {
6091 case IORING_OP_READV:
6092 case IORING_OP_READ_FIXED:
6093 case IORING_OP_READ:
6094 kfree((void *)(unsigned long)req->rw.addr);
6096 case IORING_OP_RECVMSG:
6097 case IORING_OP_RECV:
6098 kfree(req->sr_msg.kbuf);
6103 if (req->flags & REQ_F_NEED_CLEANUP) {
6104 switch (req->opcode) {
6105 case IORING_OP_READV:
6106 case IORING_OP_READ_FIXED:
6107 case IORING_OP_READ:
6108 case IORING_OP_WRITEV:
6109 case IORING_OP_WRITE_FIXED:
6110 case IORING_OP_WRITE: {
6111 struct io_async_rw *io = req->async_data;
6113 kfree(io->free_iovec);
6116 case IORING_OP_RECVMSG:
6117 case IORING_OP_SENDMSG: {
6118 struct io_async_msghdr *io = req->async_data;
6120 kfree(io->free_iov);
6123 case IORING_OP_SPLICE:
6125 if (!(req->splice.flags & SPLICE_F_FD_IN_FIXED))
6126 io_put_file(req->splice.file_in);
6128 case IORING_OP_OPENAT:
6129 case IORING_OP_OPENAT2:
6130 if (req->open.filename)
6131 putname(req->open.filename);
6133 case IORING_OP_RENAMEAT:
6134 putname(req->rename.oldpath);
6135 putname(req->rename.newpath);
6137 case IORING_OP_UNLINKAT:
6138 putname(req->unlink.filename);
6142 if ((req->flags & REQ_F_POLLED) && req->apoll) {
6143 kfree(req->apoll->double_poll);
6147 if (req->flags & REQ_F_INFLIGHT) {
6148 struct io_uring_task *tctx = req->task->io_uring;
6150 atomic_dec(&tctx->inflight_tracked);
6152 if (req->flags & REQ_F_CREDS)
6153 put_cred(req->creds);
6155 req->flags &= ~IO_REQ_CLEAN_FLAGS;
6158 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
6160 struct io_ring_ctx *ctx = req->ctx;
6161 const struct cred *creds = NULL;
6164 if ((req->flags & REQ_F_CREDS) && req->creds != current_cred())
6165 creds = override_creds(req->creds);
6167 switch (req->opcode) {
6169 ret = io_nop(req, issue_flags);
6171 case IORING_OP_READV:
6172 case IORING_OP_READ_FIXED:
6173 case IORING_OP_READ:
6174 ret = io_read(req, issue_flags);
6176 case IORING_OP_WRITEV:
6177 case IORING_OP_WRITE_FIXED:
6178 case IORING_OP_WRITE:
6179 ret = io_write(req, issue_flags);
6181 case IORING_OP_FSYNC:
6182 ret = io_fsync(req, issue_flags);
6184 case IORING_OP_POLL_ADD:
6185 ret = io_poll_add(req, issue_flags);
6187 case IORING_OP_POLL_REMOVE:
6188 ret = io_poll_update(req, issue_flags);
6190 case IORING_OP_SYNC_FILE_RANGE:
6191 ret = io_sync_file_range(req, issue_flags);
6193 case IORING_OP_SENDMSG:
6194 ret = io_sendmsg(req, issue_flags);
6196 case IORING_OP_SEND:
6197 ret = io_send(req, issue_flags);
6199 case IORING_OP_RECVMSG:
6200 ret = io_recvmsg(req, issue_flags);
6202 case IORING_OP_RECV:
6203 ret = io_recv(req, issue_flags);
6205 case IORING_OP_TIMEOUT:
6206 ret = io_timeout(req, issue_flags);
6208 case IORING_OP_TIMEOUT_REMOVE:
6209 ret = io_timeout_remove(req, issue_flags);
6211 case IORING_OP_ACCEPT:
6212 ret = io_accept(req, issue_flags);
6214 case IORING_OP_CONNECT:
6215 ret = io_connect(req, issue_flags);
6217 case IORING_OP_ASYNC_CANCEL:
6218 ret = io_async_cancel(req, issue_flags);
6220 case IORING_OP_FALLOCATE:
6221 ret = io_fallocate(req, issue_flags);
6223 case IORING_OP_OPENAT:
6224 ret = io_openat(req, issue_flags);
6226 case IORING_OP_CLOSE:
6227 ret = io_close(req, issue_flags);
6229 case IORING_OP_FILES_UPDATE:
6230 ret = io_files_update(req, issue_flags);
6232 case IORING_OP_STATX:
6233 ret = io_statx(req, issue_flags);
6235 case IORING_OP_FADVISE:
6236 ret = io_fadvise(req, issue_flags);
6238 case IORING_OP_MADVISE:
6239 ret = io_madvise(req, issue_flags);
6241 case IORING_OP_OPENAT2:
6242 ret = io_openat2(req, issue_flags);
6244 case IORING_OP_EPOLL_CTL:
6245 ret = io_epoll_ctl(req, issue_flags);
6247 case IORING_OP_SPLICE:
6248 ret = io_splice(req, issue_flags);
6250 case IORING_OP_PROVIDE_BUFFERS:
6251 ret = io_provide_buffers(req, issue_flags);
6253 case IORING_OP_REMOVE_BUFFERS:
6254 ret = io_remove_buffers(req, issue_flags);
6257 ret = io_tee(req, issue_flags);
6259 case IORING_OP_SHUTDOWN:
6260 ret = io_shutdown(req, issue_flags);
6262 case IORING_OP_RENAMEAT:
6263 ret = io_renameat(req, issue_flags);
6265 case IORING_OP_UNLINKAT:
6266 ret = io_unlinkat(req, issue_flags);
6274 revert_creds(creds);
6277 /* If the op doesn't have a file, we're not polling for it */
6278 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file)
6279 io_iopoll_req_issued(req);
6284 static void io_wq_submit_work(struct io_wq_work *work)
6286 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6287 struct io_kiocb *timeout;
6290 timeout = io_prep_linked_timeout(req);
6292 io_queue_linked_timeout(timeout);
6294 if (work->flags & IO_WQ_WORK_CANCEL)
6299 ret = io_issue_sqe(req, 0);
6301 * We can get EAGAIN for polled IO even though we're
6302 * forcing a sync submission from here, since we can't
6303 * wait for request slots on the block side.
6311 /* avoid locking problems by failing it from a clean context */
6313 /* io-wq is going to take one down */
6315 io_req_task_queue_fail(req, ret);
6319 #define FFS_ASYNC_READ 0x1UL
6320 #define FFS_ASYNC_WRITE 0x2UL
6322 #define FFS_ISREG 0x4UL
6324 #define FFS_ISREG 0x0UL
6326 #define FFS_MASK ~(FFS_ASYNC_READ|FFS_ASYNC_WRITE|FFS_ISREG)
6328 static inline struct io_fixed_file *io_fixed_file_slot(struct io_file_table *table,
6331 struct io_fixed_file *table_l2;
6333 table_l2 = table->files[i >> IORING_FILE_TABLE_SHIFT];
6334 return &table_l2[i & IORING_FILE_TABLE_MASK];
6337 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6340 struct io_fixed_file *slot = io_fixed_file_slot(&ctx->file_table, index);
6342 return (struct file *) (slot->file_ptr & FFS_MASK);
6345 static void io_fixed_file_set(struct io_fixed_file *file_slot, struct file *file)
6347 unsigned long file_ptr = (unsigned long) file;
6349 if (__io_file_supports_async(file, READ))
6350 file_ptr |= FFS_ASYNC_READ;
6351 if (__io_file_supports_async(file, WRITE))
6352 file_ptr |= FFS_ASYNC_WRITE;
6353 if (S_ISREG(file_inode(file)->i_mode))
6354 file_ptr |= FFS_ISREG;
6355 file_slot->file_ptr = file_ptr;
6358 static struct file *io_file_get(struct io_submit_state *state,
6359 struct io_kiocb *req, int fd, bool fixed)
6361 struct io_ring_ctx *ctx = req->ctx;
6365 unsigned long file_ptr;
6367 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6369 fd = array_index_nospec(fd, ctx->nr_user_files);
6370 file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr;
6371 file = (struct file *) (file_ptr & FFS_MASK);
6372 file_ptr &= ~FFS_MASK;
6373 /* mask in overlapping REQ_F and FFS bits */
6374 req->flags |= (file_ptr << REQ_F_ASYNC_READ_BIT);
6375 io_req_set_rsrc_node(req);
6377 trace_io_uring_file_get(ctx, fd);
6378 file = __io_file_get(state, fd);
6380 /* we don't allow fixed io_uring files */
6381 if (file && unlikely(file->f_op == &io_uring_fops))
6382 io_req_track_inflight(req);
6388 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6390 struct io_timeout_data *data = container_of(timer,
6391 struct io_timeout_data, timer);
6392 struct io_kiocb *prev, *req = data->req;
6393 struct io_ring_ctx *ctx = req->ctx;
6394 unsigned long flags;
6396 spin_lock_irqsave(&ctx->completion_lock, flags);
6397 prev = req->timeout.head;
6398 req->timeout.head = NULL;
6401 * We don't expect the list to be empty, that will only happen if we
6402 * race with the completion of the linked work.
6405 io_remove_next_linked(prev);
6406 if (!req_ref_inc_not_zero(prev))
6409 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6412 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6413 io_put_req_deferred(prev, 1);
6414 io_put_req_deferred(req, 1);
6416 io_req_complete_post(req, -ETIME, 0);
6418 return HRTIMER_NORESTART;
6421 static void io_queue_linked_timeout(struct io_kiocb *req)
6423 struct io_ring_ctx *ctx = req->ctx;
6425 spin_lock_irq(&ctx->completion_lock);
6427 * If the back reference is NULL, then our linked request finished
6428 * before we got a chance to setup the timer
6430 if (req->timeout.head) {
6431 struct io_timeout_data *data = req->async_data;
6433 data->timer.function = io_link_timeout_fn;
6434 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6437 spin_unlock_irq(&ctx->completion_lock);
6438 /* drop submission reference */
6442 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6444 struct io_kiocb *nxt = req->link;
6446 if (!nxt || (req->flags & REQ_F_LINK_TIMEOUT) ||
6447 nxt->opcode != IORING_OP_LINK_TIMEOUT)
6450 nxt->timeout.head = req;
6451 nxt->flags |= REQ_F_LTIMEOUT_ACTIVE;
6452 req->flags |= REQ_F_LINK_TIMEOUT;
6456 static void __io_queue_sqe(struct io_kiocb *req)
6458 struct io_kiocb *linked_timeout = io_prep_linked_timeout(req);
6462 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
6465 * We async punt it if the file wasn't marked NOWAIT, or if the file
6466 * doesn't support non-blocking read/write attempts
6469 /* drop submission reference */
6470 if (req->flags & REQ_F_COMPLETE_INLINE) {
6471 struct io_ring_ctx *ctx = req->ctx;
6472 struct io_comp_state *cs = &ctx->submit_state.comp;
6474 cs->reqs[cs->nr++] = req;
6475 if (cs->nr == ARRAY_SIZE(cs->reqs))
6476 io_submit_flush_completions(ctx);
6480 } else if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6481 switch (io_arm_poll_handler(req)) {
6482 case IO_APOLL_READY:
6484 case IO_APOLL_ABORTED:
6486 * Queued up for async execution, worker will release
6487 * submit reference when the iocb is actually submitted.
6489 io_queue_async_work(req);
6493 io_req_complete_failed(req, ret);
6496 io_queue_linked_timeout(linked_timeout);
6499 static inline void io_queue_sqe(struct io_kiocb *req)
6501 if (unlikely(req->ctx->drain_active) && io_drain_req(req))
6504 if (likely(!(req->flags & REQ_F_FORCE_ASYNC))) {
6505 __io_queue_sqe(req);
6507 int ret = io_req_prep_async(req);
6510 io_req_complete_failed(req, ret);
6512 io_queue_async_work(req);
6517 * Check SQE restrictions (opcode and flags).
6519 * Returns 'true' if SQE is allowed, 'false' otherwise.
6521 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6522 struct io_kiocb *req,
6523 unsigned int sqe_flags)
6525 if (likely(!ctx->restricted))
6528 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6531 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6532 ctx->restrictions.sqe_flags_required)
6535 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6536 ctx->restrictions.sqe_flags_required))
6542 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6543 const struct io_uring_sqe *sqe)
6545 struct io_submit_state *state;
6546 unsigned int sqe_flags;
6547 int personality, ret = 0;
6549 req->opcode = READ_ONCE(sqe->opcode);
6550 /* same numerical values with corresponding REQ_F_*, safe to copy */
6551 req->flags = sqe_flags = READ_ONCE(sqe->flags);
6552 req->user_data = READ_ONCE(sqe->user_data);
6554 req->fixed_rsrc_refs = NULL;
6555 /* one is dropped after submission, the other at completion */
6556 atomic_set(&req->refs, 2);
6557 req->task = current;
6559 /* enforce forwards compatibility on users */
6560 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6562 if (unlikely(req->opcode >= IORING_OP_LAST))
6564 if (!io_check_restriction(ctx, req, sqe_flags))
6567 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6568 !io_op_defs[req->opcode].buffer_select)
6570 if (unlikely(sqe_flags & IOSQE_IO_DRAIN))
6571 ctx->drain_active = true;
6573 personality = READ_ONCE(sqe->personality);
6575 req->creds = xa_load(&ctx->personalities, personality);
6578 get_cred(req->creds);
6579 req->flags |= REQ_F_CREDS;
6581 state = &ctx->submit_state;
6584 * Plug now if we have more than 1 IO left after this, and the target
6585 * is potentially a read/write to block based storage.
6587 if (!state->plug_started && state->ios_left > 1 &&
6588 io_op_defs[req->opcode].plug) {
6589 blk_start_plug(&state->plug);
6590 state->plug_started = true;
6593 if (io_op_defs[req->opcode].needs_file) {
6594 bool fixed = req->flags & REQ_F_FIXED_FILE;
6596 req->file = io_file_get(state, req, READ_ONCE(sqe->fd), fixed);
6597 if (unlikely(!req->file))
6605 static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
6606 const struct io_uring_sqe *sqe)
6608 struct io_submit_link *link = &ctx->submit_state.link;
6611 ret = io_init_req(ctx, req, sqe);
6612 if (unlikely(ret)) {
6615 /* fail even hard links since we don't submit */
6616 req_set_fail(link->head);
6617 io_req_complete_failed(link->head, -ECANCELED);
6620 io_req_complete_failed(req, ret);
6624 ret = io_req_prep(req, sqe);
6628 /* don't need @sqe from now on */
6629 trace_io_uring_submit_sqe(ctx, req, req->opcode, req->user_data,
6631 ctx->flags & IORING_SETUP_SQPOLL);
6634 * If we already have a head request, queue this one for async
6635 * submittal once the head completes. If we don't have a head but
6636 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6637 * submitted sync once the chain is complete. If none of those
6638 * conditions are true (normal request), then just queue it.
6641 struct io_kiocb *head = link->head;
6643 ret = io_req_prep_async(req);
6646 trace_io_uring_link(ctx, req, head);
6647 link->last->link = req;
6650 /* last request of a link, enqueue the link */
6651 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6656 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6668 * Batched submission is done, ensure local IO is flushed out.
6670 static void io_submit_state_end(struct io_submit_state *state,
6671 struct io_ring_ctx *ctx)
6673 if (state->link.head)
6674 io_queue_sqe(state->link.head);
6676 io_submit_flush_completions(ctx);
6677 if (state->plug_started)
6678 blk_finish_plug(&state->plug);
6679 io_state_file_put(state);
6683 * Start submission side cache.
6685 static void io_submit_state_start(struct io_submit_state *state,
6686 unsigned int max_ios)
6688 state->plug_started = false;
6689 state->ios_left = max_ios;
6690 /* set only head, no need to init link_last in advance */
6691 state->link.head = NULL;
6694 static void io_commit_sqring(struct io_ring_ctx *ctx)
6696 struct io_rings *rings = ctx->rings;
6699 * Ensure any loads from the SQEs are done at this point,
6700 * since once we write the new head, the application could
6701 * write new data to them.
6703 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6707 * Fetch an sqe, if one is available. Note this returns a pointer to memory
6708 * that is mapped by userspace. This means that care needs to be taken to
6709 * ensure that reads are stable, as we cannot rely on userspace always
6710 * being a good citizen. If members of the sqe are validated and then later
6711 * used, it's important that those reads are done through READ_ONCE() to
6712 * prevent a re-load down the line.
6714 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6716 unsigned head, mask = ctx->sq_entries - 1;
6717 unsigned sq_idx = ctx->cached_sq_head++ & mask;
6720 * The cached sq head (or cq tail) serves two purposes:
6722 * 1) allows us to batch the cost of updating the user visible
6724 * 2) allows the kernel side to track the head on its own, even
6725 * though the application is the one updating it.
6727 head = READ_ONCE(ctx->sq_array[sq_idx]);
6728 if (likely(head < ctx->sq_entries))
6729 return &ctx->sq_sqes[head];
6731 /* drop invalid entries */
6733 WRITE_ONCE(ctx->rings->sq_dropped,
6734 READ_ONCE(ctx->rings->sq_dropped) + 1);
6738 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6740 struct io_uring_task *tctx;
6743 /* make sure SQ entry isn't read before tail */
6744 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6745 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6748 tctx = current->io_uring;
6749 tctx->cached_refs -= nr;
6750 if (unlikely(tctx->cached_refs < 0)) {
6751 unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR;
6753 percpu_counter_add(&tctx->inflight, refill);
6754 refcount_add(refill, ¤t->usage);
6755 tctx->cached_refs += refill;
6757 io_submit_state_start(&ctx->submit_state, nr);
6759 while (submitted < nr) {
6760 const struct io_uring_sqe *sqe;
6761 struct io_kiocb *req;
6763 req = io_alloc_req(ctx);
6764 if (unlikely(!req)) {
6766 submitted = -EAGAIN;
6769 sqe = io_get_sqe(ctx);
6770 if (unlikely(!sqe)) {
6771 kmem_cache_free(req_cachep, req);
6774 /* will complete beyond this point, count as submitted */
6776 if (io_submit_sqe(ctx, req, sqe))
6780 if (unlikely(submitted != nr)) {
6781 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6782 int unused = nr - ref_used;
6784 current->io_uring->cached_refs += unused;
6785 percpu_ref_put_many(&ctx->refs, unused);
6788 io_submit_state_end(&ctx->submit_state, ctx);
6789 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6790 io_commit_sqring(ctx);
6795 static inline bool io_sqd_events_pending(struct io_sq_data *sqd)
6797 return READ_ONCE(sqd->state);
6800 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6802 /* Tell userspace we may need a wakeup call */
6803 spin_lock_irq(&ctx->completion_lock);
6804 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6805 spin_unlock_irq(&ctx->completion_lock);
6808 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6810 spin_lock_irq(&ctx->completion_lock);
6811 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6812 spin_unlock_irq(&ctx->completion_lock);
6815 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
6817 unsigned int to_submit;
6820 to_submit = io_sqring_entries(ctx);
6821 /* if we're handling multiple rings, cap submit size for fairness */
6822 if (cap_entries && to_submit > IORING_SQPOLL_CAP_ENTRIES_VALUE)
6823 to_submit = IORING_SQPOLL_CAP_ENTRIES_VALUE;
6825 if (!list_empty(&ctx->iopoll_list) || to_submit) {
6826 unsigned nr_events = 0;
6827 const struct cred *creds = NULL;
6829 if (ctx->sq_creds != current_cred())
6830 creds = override_creds(ctx->sq_creds);
6832 mutex_lock(&ctx->uring_lock);
6833 if (!list_empty(&ctx->iopoll_list))
6834 io_do_iopoll(ctx, &nr_events, 0);
6837 * Don't submit if refs are dying, good for io_uring_register(),
6838 * but also it is relied upon by io_ring_exit_work()
6840 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)) &&
6841 !(ctx->flags & IORING_SETUP_R_DISABLED))
6842 ret = io_submit_sqes(ctx, to_submit);
6843 mutex_unlock(&ctx->uring_lock);
6845 if (to_submit && wq_has_sleeper(&ctx->sqo_sq_wait))
6846 wake_up(&ctx->sqo_sq_wait);
6848 revert_creds(creds);
6854 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
6856 struct io_ring_ctx *ctx;
6857 unsigned sq_thread_idle = 0;
6859 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6860 sq_thread_idle = max(sq_thread_idle, ctx->sq_thread_idle);
6861 sqd->sq_thread_idle = sq_thread_idle;
6864 static bool io_sqd_handle_event(struct io_sq_data *sqd)
6866 bool did_sig = false;
6867 struct ksignal ksig;
6869 if (test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state) ||
6870 signal_pending(current)) {
6871 mutex_unlock(&sqd->lock);
6872 if (signal_pending(current))
6873 did_sig = get_signal(&ksig);
6875 mutex_lock(&sqd->lock);
6877 return did_sig || test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
6880 static int io_sq_thread(void *data)
6882 struct io_sq_data *sqd = data;
6883 struct io_ring_ctx *ctx;
6884 unsigned long timeout = 0;
6885 char buf[TASK_COMM_LEN];
6888 snprintf(buf, sizeof(buf), "iou-sqp-%d", sqd->task_pid);
6889 set_task_comm(current, buf);
6891 if (sqd->sq_cpu != -1)
6892 set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
6894 set_cpus_allowed_ptr(current, cpu_online_mask);
6895 current->flags |= PF_NO_SETAFFINITY;
6897 mutex_lock(&sqd->lock);
6899 bool cap_entries, sqt_spin = false;
6901 if (io_sqd_events_pending(sqd) || signal_pending(current)) {
6902 if (io_sqd_handle_event(sqd))
6904 timeout = jiffies + sqd->sq_thread_idle;
6907 cap_entries = !list_is_singular(&sqd->ctx_list);
6908 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6909 int ret = __io_sq_thread(ctx, cap_entries);
6911 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
6914 if (io_run_task_work())
6917 if (sqt_spin || !time_after(jiffies, timeout)) {
6920 timeout = jiffies + sqd->sq_thread_idle;
6924 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
6925 if (!io_sqd_events_pending(sqd) && !current->task_works) {
6926 bool needs_sched = true;
6928 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6929 io_ring_set_wakeup_flag(ctx);
6931 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6932 !list_empty_careful(&ctx->iopoll_list)) {
6933 needs_sched = false;
6936 if (io_sqring_entries(ctx)) {
6937 needs_sched = false;
6943 mutex_unlock(&sqd->lock);
6945 mutex_lock(&sqd->lock);
6947 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6948 io_ring_clear_wakeup_flag(ctx);
6951 finish_wait(&sqd->wait, &wait);
6952 timeout = jiffies + sqd->sq_thread_idle;
6955 io_uring_cancel_generic(true, sqd);
6957 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6958 io_ring_set_wakeup_flag(ctx);
6960 mutex_unlock(&sqd->lock);
6962 complete(&sqd->exited);
6966 struct io_wait_queue {
6967 struct wait_queue_entry wq;
6968 struct io_ring_ctx *ctx;
6970 unsigned nr_timeouts;
6973 static inline bool io_should_wake(struct io_wait_queue *iowq)
6975 struct io_ring_ctx *ctx = iowq->ctx;
6978 * Wake up if we have enough events, or if a timeout occurred since we
6979 * started waiting. For timeouts, we always want to return to userspace,
6980 * regardless of event count.
6982 return io_cqring_events(ctx) >= iowq->to_wait ||
6983 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6986 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6987 int wake_flags, void *key)
6989 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6993 * Cannot safely flush overflowed CQEs from here, ensure we wake up
6994 * the task, and the next invocation will do it.
6996 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->check_cq_overflow))
6997 return autoremove_wake_function(curr, mode, wake_flags, key);
7001 static int io_run_task_work_sig(void)
7003 if (io_run_task_work())
7005 if (!signal_pending(current))
7007 if (test_thread_flag(TIF_NOTIFY_SIGNAL))
7008 return -ERESTARTSYS;
7012 /* when returns >0, the caller should retry */
7013 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
7014 struct io_wait_queue *iowq,
7015 signed long *timeout)
7019 /* make sure we run task_work before checking for signals */
7020 ret = io_run_task_work_sig();
7021 if (ret || io_should_wake(iowq))
7023 /* let the caller flush overflows, retry */
7024 if (test_bit(0, &ctx->check_cq_overflow))
7027 *timeout = schedule_timeout(*timeout);
7028 return !*timeout ? -ETIME : 1;
7032 * Wait until events become available, if we don't already have some. The
7033 * application must reap them itself, as they reside on the shared cq ring.
7035 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
7036 const sigset_t __user *sig, size_t sigsz,
7037 struct __kernel_timespec __user *uts)
7039 struct io_wait_queue iowq = {
7042 .func = io_wake_function,
7043 .entry = LIST_HEAD_INIT(iowq.wq.entry),
7046 .to_wait = min_events,
7048 struct io_rings *rings = ctx->rings;
7049 signed long timeout = MAX_SCHEDULE_TIMEOUT;
7053 io_cqring_overflow_flush(ctx, false);
7054 if (io_cqring_events(ctx) >= min_events)
7056 if (!io_run_task_work())
7061 #ifdef CONFIG_COMPAT
7062 if (in_compat_syscall())
7063 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
7067 ret = set_user_sigmask(sig, sigsz);
7074 struct timespec64 ts;
7076 if (get_timespec64(&ts, uts))
7078 timeout = timespec64_to_jiffies(&ts);
7081 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
7082 trace_io_uring_cqring_wait(ctx, min_events);
7084 /* if we can't even flush overflow, don't wait for more */
7085 if (!io_cqring_overflow_flush(ctx, false)) {
7089 prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq,
7090 TASK_INTERRUPTIBLE);
7091 ret = io_cqring_wait_schedule(ctx, &iowq, &timeout);
7092 finish_wait(&ctx->cq_wait, &iowq.wq);
7096 restore_saved_sigmask_unless(ret == -EINTR);
7098 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
7101 static void io_free_page_table(void **table, size_t size)
7103 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7105 for (i = 0; i < nr_tables; i++)
7110 static void **io_alloc_page_table(size_t size)
7112 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7113 size_t init_size = size;
7116 table = kcalloc(nr_tables, sizeof(*table), GFP_KERNEL);
7120 for (i = 0; i < nr_tables; i++) {
7121 unsigned int this_size = min_t(size_t, size, PAGE_SIZE);
7123 table[i] = kzalloc(this_size, GFP_KERNEL);
7125 io_free_page_table(table, init_size);
7133 static inline void io_rsrc_ref_lock(struct io_ring_ctx *ctx)
7135 spin_lock_bh(&ctx->rsrc_ref_lock);
7138 static inline void io_rsrc_ref_unlock(struct io_ring_ctx *ctx)
7140 spin_unlock_bh(&ctx->rsrc_ref_lock);
7143 static void io_rsrc_node_destroy(struct io_rsrc_node *ref_node)
7145 percpu_ref_exit(&ref_node->refs);
7149 static void io_rsrc_node_switch(struct io_ring_ctx *ctx,
7150 struct io_rsrc_data *data_to_kill)
7152 WARN_ON_ONCE(!ctx->rsrc_backup_node);
7153 WARN_ON_ONCE(data_to_kill && !ctx->rsrc_node);
7156 struct io_rsrc_node *rsrc_node = ctx->rsrc_node;
7158 rsrc_node->rsrc_data = data_to_kill;
7159 io_rsrc_ref_lock(ctx);
7160 list_add_tail(&rsrc_node->node, &ctx->rsrc_ref_list);
7161 io_rsrc_ref_unlock(ctx);
7163 atomic_inc(&data_to_kill->refs);
7164 percpu_ref_kill(&rsrc_node->refs);
7165 ctx->rsrc_node = NULL;
7168 if (!ctx->rsrc_node) {
7169 ctx->rsrc_node = ctx->rsrc_backup_node;
7170 ctx->rsrc_backup_node = NULL;
7174 static int io_rsrc_node_switch_start(struct io_ring_ctx *ctx)
7176 if (ctx->rsrc_backup_node)
7178 ctx->rsrc_backup_node = io_rsrc_node_alloc(ctx);
7179 return ctx->rsrc_backup_node ? 0 : -ENOMEM;
7182 static int io_rsrc_ref_quiesce(struct io_rsrc_data *data, struct io_ring_ctx *ctx)
7186 /* As we may drop ->uring_lock, other task may have started quiesce */
7190 data->quiesce = true;
7192 ret = io_rsrc_node_switch_start(ctx);
7195 io_rsrc_node_switch(ctx, data);
7197 /* kill initial ref, already quiesced if zero */
7198 if (atomic_dec_and_test(&data->refs))
7200 flush_delayed_work(&ctx->rsrc_put_work);
7201 ret = wait_for_completion_interruptible(&data->done);
7205 atomic_inc(&data->refs);
7206 /* wait for all works potentially completing data->done */
7207 flush_delayed_work(&ctx->rsrc_put_work);
7208 reinit_completion(&data->done);
7210 mutex_unlock(&ctx->uring_lock);
7211 ret = io_run_task_work_sig();
7212 mutex_lock(&ctx->uring_lock);
7214 data->quiesce = false;
7219 static u64 *io_get_tag_slot(struct io_rsrc_data *data, unsigned int idx)
7221 unsigned int off = idx & IO_RSRC_TAG_TABLE_MASK;
7222 unsigned int table_idx = idx >> IO_RSRC_TAG_TABLE_SHIFT;
7224 return &data->tags[table_idx][off];
7227 static void io_rsrc_data_free(struct io_rsrc_data *data)
7229 size_t size = data->nr * sizeof(data->tags[0][0]);
7232 io_free_page_table((void **)data->tags, size);
7236 static int io_rsrc_data_alloc(struct io_ring_ctx *ctx, rsrc_put_fn *do_put,
7237 u64 __user *utags, unsigned nr,
7238 struct io_rsrc_data **pdata)
7240 struct io_rsrc_data *data;
7244 data = kzalloc(sizeof(*data), GFP_KERNEL);
7247 data->tags = (u64 **)io_alloc_page_table(nr * sizeof(data->tags[0][0]));
7255 data->do_put = do_put;
7258 for (i = 0; i < nr; i++) {
7259 u64 *tag_slot = io_get_tag_slot(data, i);
7261 if (copy_from_user(tag_slot, &utags[i],
7267 atomic_set(&data->refs, 1);
7268 init_completion(&data->done);
7272 io_rsrc_data_free(data);
7276 static bool io_alloc_file_tables(struct io_file_table *table, unsigned nr_files)
7278 size_t size = nr_files * sizeof(struct io_fixed_file);
7280 table->files = (struct io_fixed_file **)io_alloc_page_table(size);
7281 return !!table->files;
7284 static void io_free_file_tables(struct io_file_table *table, unsigned nr_files)
7286 size_t size = nr_files * sizeof(struct io_fixed_file);
7288 io_free_page_table((void **)table->files, size);
7289 table->files = NULL;
7292 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
7294 #if defined(CONFIG_UNIX)
7295 if (ctx->ring_sock) {
7296 struct sock *sock = ctx->ring_sock->sk;
7297 struct sk_buff *skb;
7299 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
7305 for (i = 0; i < ctx->nr_user_files; i++) {
7308 file = io_file_from_index(ctx, i);
7313 io_free_file_tables(&ctx->file_table, ctx->nr_user_files);
7314 io_rsrc_data_free(ctx->file_data);
7315 ctx->file_data = NULL;
7316 ctx->nr_user_files = 0;
7319 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7323 if (!ctx->file_data)
7325 ret = io_rsrc_ref_quiesce(ctx->file_data, ctx);
7327 __io_sqe_files_unregister(ctx);
7331 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7332 __releases(&sqd->lock)
7334 WARN_ON_ONCE(sqd->thread == current);
7337 * Do the dance but not conditional clear_bit() because it'd race with
7338 * other threads incrementing park_pending and setting the bit.
7340 clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7341 if (atomic_dec_return(&sqd->park_pending))
7342 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7343 mutex_unlock(&sqd->lock);
7346 static void io_sq_thread_park(struct io_sq_data *sqd)
7347 __acquires(&sqd->lock)
7349 WARN_ON_ONCE(sqd->thread == current);
7351 atomic_inc(&sqd->park_pending);
7352 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7353 mutex_lock(&sqd->lock);
7355 wake_up_process(sqd->thread);
7358 static void io_sq_thread_stop(struct io_sq_data *sqd)
7360 WARN_ON_ONCE(sqd->thread == current);
7361 WARN_ON_ONCE(test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state));
7363 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7364 mutex_lock(&sqd->lock);
7366 wake_up_process(sqd->thread);
7367 mutex_unlock(&sqd->lock);
7368 wait_for_completion(&sqd->exited);
7371 static void io_put_sq_data(struct io_sq_data *sqd)
7373 if (refcount_dec_and_test(&sqd->refs)) {
7374 WARN_ON_ONCE(atomic_read(&sqd->park_pending));
7376 io_sq_thread_stop(sqd);
7381 static void io_sq_thread_finish(struct io_ring_ctx *ctx)
7383 struct io_sq_data *sqd = ctx->sq_data;
7386 io_sq_thread_park(sqd);
7387 list_del_init(&ctx->sqd_list);
7388 io_sqd_update_thread_idle(sqd);
7389 io_sq_thread_unpark(sqd);
7391 io_put_sq_data(sqd);
7392 ctx->sq_data = NULL;
7396 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7398 struct io_ring_ctx *ctx_attach;
7399 struct io_sq_data *sqd;
7402 f = fdget(p->wq_fd);
7404 return ERR_PTR(-ENXIO);
7405 if (f.file->f_op != &io_uring_fops) {
7407 return ERR_PTR(-EINVAL);
7410 ctx_attach = f.file->private_data;
7411 sqd = ctx_attach->sq_data;
7414 return ERR_PTR(-EINVAL);
7416 if (sqd->task_tgid != current->tgid) {
7418 return ERR_PTR(-EPERM);
7421 refcount_inc(&sqd->refs);
7426 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p,
7429 struct io_sq_data *sqd;
7432 if (p->flags & IORING_SETUP_ATTACH_WQ) {
7433 sqd = io_attach_sq_data(p);
7438 /* fall through for EPERM case, setup new sqd/task */
7439 if (PTR_ERR(sqd) != -EPERM)
7443 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7445 return ERR_PTR(-ENOMEM);
7447 atomic_set(&sqd->park_pending, 0);
7448 refcount_set(&sqd->refs, 1);
7449 INIT_LIST_HEAD(&sqd->ctx_list);
7450 mutex_init(&sqd->lock);
7451 init_waitqueue_head(&sqd->wait);
7452 init_completion(&sqd->exited);
7456 #if defined(CONFIG_UNIX)
7458 * Ensure the UNIX gc is aware of our file set, so we are certain that
7459 * the io_uring can be safely unregistered on process exit, even if we have
7460 * loops in the file referencing.
7462 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7464 struct sock *sk = ctx->ring_sock->sk;
7465 struct scm_fp_list *fpl;
7466 struct sk_buff *skb;
7469 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7473 skb = alloc_skb(0, GFP_KERNEL);
7482 fpl->user = get_uid(current_user());
7483 for (i = 0; i < nr; i++) {
7484 struct file *file = io_file_from_index(ctx, i + offset);
7488 fpl->fp[nr_files] = get_file(file);
7489 unix_inflight(fpl->user, fpl->fp[nr_files]);
7494 fpl->max = SCM_MAX_FD;
7495 fpl->count = nr_files;
7496 UNIXCB(skb).fp = fpl;
7497 skb->destructor = unix_destruct_scm;
7498 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7499 skb_queue_head(&sk->sk_receive_queue, skb);
7501 for (i = 0; i < nr_files; i++)
7512 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7513 * causes regular reference counting to break down. We rely on the UNIX
7514 * garbage collection to take care of this problem for us.
7516 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7518 unsigned left, total;
7522 left = ctx->nr_user_files;
7524 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7526 ret = __io_sqe_files_scm(ctx, this_files, total);
7530 total += this_files;
7536 while (total < ctx->nr_user_files) {
7537 struct file *file = io_file_from_index(ctx, total);
7547 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7553 static void io_rsrc_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
7555 struct file *file = prsrc->file;
7556 #if defined(CONFIG_UNIX)
7557 struct sock *sock = ctx->ring_sock->sk;
7558 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7559 struct sk_buff *skb;
7562 __skb_queue_head_init(&list);
7565 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7566 * remove this entry and rearrange the file array.
7568 skb = skb_dequeue(head);
7570 struct scm_fp_list *fp;
7572 fp = UNIXCB(skb).fp;
7573 for (i = 0; i < fp->count; i++) {
7576 if (fp->fp[i] != file)
7579 unix_notinflight(fp->user, fp->fp[i]);
7580 left = fp->count - 1 - i;
7582 memmove(&fp->fp[i], &fp->fp[i + 1],
7583 left * sizeof(struct file *));
7590 __skb_queue_tail(&list, skb);
7600 __skb_queue_tail(&list, skb);
7602 skb = skb_dequeue(head);
7605 if (skb_peek(&list)) {
7606 spin_lock_irq(&head->lock);
7607 while ((skb = __skb_dequeue(&list)) != NULL)
7608 __skb_queue_tail(head, skb);
7609 spin_unlock_irq(&head->lock);
7616 static void __io_rsrc_put_work(struct io_rsrc_node *ref_node)
7618 struct io_rsrc_data *rsrc_data = ref_node->rsrc_data;
7619 struct io_ring_ctx *ctx = rsrc_data->ctx;
7620 struct io_rsrc_put *prsrc, *tmp;
7622 list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
7623 list_del(&prsrc->list);
7626 bool lock_ring = ctx->flags & IORING_SETUP_IOPOLL;
7628 io_ring_submit_lock(ctx, lock_ring);
7629 spin_lock_irq(&ctx->completion_lock);
7630 io_cqring_fill_event(ctx, prsrc->tag, 0, 0);
7632 io_commit_cqring(ctx);
7633 spin_unlock_irq(&ctx->completion_lock);
7634 io_cqring_ev_posted(ctx);
7635 io_ring_submit_unlock(ctx, lock_ring);
7638 rsrc_data->do_put(ctx, prsrc);
7642 io_rsrc_node_destroy(ref_node);
7643 if (atomic_dec_and_test(&rsrc_data->refs))
7644 complete(&rsrc_data->done);
7647 static void io_rsrc_put_work(struct work_struct *work)
7649 struct io_ring_ctx *ctx;
7650 struct llist_node *node;
7652 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
7653 node = llist_del_all(&ctx->rsrc_put_llist);
7656 struct io_rsrc_node *ref_node;
7657 struct llist_node *next = node->next;
7659 ref_node = llist_entry(node, struct io_rsrc_node, llist);
7660 __io_rsrc_put_work(ref_node);
7665 static void io_rsrc_node_ref_zero(struct percpu_ref *ref)
7667 struct io_rsrc_node *node = container_of(ref, struct io_rsrc_node, refs);
7668 struct io_ring_ctx *ctx = node->rsrc_data->ctx;
7669 bool first_add = false;
7671 io_rsrc_ref_lock(ctx);
7674 while (!list_empty(&ctx->rsrc_ref_list)) {
7675 node = list_first_entry(&ctx->rsrc_ref_list,
7676 struct io_rsrc_node, node);
7677 /* recycle ref nodes in order */
7680 list_del(&node->node);
7681 first_add |= llist_add(&node->llist, &ctx->rsrc_put_llist);
7683 io_rsrc_ref_unlock(ctx);
7686 mod_delayed_work(system_wq, &ctx->rsrc_put_work, HZ);
7689 static struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx)
7691 struct io_rsrc_node *ref_node;
7693 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7697 if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
7702 INIT_LIST_HEAD(&ref_node->node);
7703 INIT_LIST_HEAD(&ref_node->rsrc_list);
7704 ref_node->done = false;
7708 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7709 unsigned nr_args, u64 __user *tags)
7711 __s32 __user *fds = (__s32 __user *) arg;
7720 if (nr_args > IORING_MAX_FIXED_FILES)
7722 ret = io_rsrc_node_switch_start(ctx);
7725 ret = io_rsrc_data_alloc(ctx, io_rsrc_file_put, tags, nr_args,
7731 if (!io_alloc_file_tables(&ctx->file_table, nr_args))
7734 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7735 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7739 /* allow sparse sets */
7742 if (unlikely(*io_get_tag_slot(ctx->file_data, i)))
7749 if (unlikely(!file))
7753 * Don't allow io_uring instances to be registered. If UNIX
7754 * isn't enabled, then this causes a reference cycle and this
7755 * instance can never get freed. If UNIX is enabled we'll
7756 * handle it just fine, but there's still no point in allowing
7757 * a ring fd as it doesn't support regular read/write anyway.
7759 if (file->f_op == &io_uring_fops) {
7763 io_fixed_file_set(io_fixed_file_slot(&ctx->file_table, i), file);
7766 ret = io_sqe_files_scm(ctx);
7768 __io_sqe_files_unregister(ctx);
7772 io_rsrc_node_switch(ctx, NULL);
7775 for (i = 0; i < ctx->nr_user_files; i++) {
7776 file = io_file_from_index(ctx, i);
7780 io_free_file_tables(&ctx->file_table, nr_args);
7781 ctx->nr_user_files = 0;
7783 io_rsrc_data_free(ctx->file_data);
7784 ctx->file_data = NULL;
7788 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7791 #if defined(CONFIG_UNIX)
7792 struct sock *sock = ctx->ring_sock->sk;
7793 struct sk_buff_head *head = &sock->sk_receive_queue;
7794 struct sk_buff *skb;
7797 * See if we can merge this file into an existing skb SCM_RIGHTS
7798 * file set. If there's no room, fall back to allocating a new skb
7799 * and filling it in.
7801 spin_lock_irq(&head->lock);
7802 skb = skb_peek(head);
7804 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7806 if (fpl->count < SCM_MAX_FD) {
7807 __skb_unlink(skb, head);
7808 spin_unlock_irq(&head->lock);
7809 fpl->fp[fpl->count] = get_file(file);
7810 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7812 spin_lock_irq(&head->lock);
7813 __skb_queue_head(head, skb);
7818 spin_unlock_irq(&head->lock);
7825 return __io_sqe_files_scm(ctx, 1, index);
7831 static int io_queue_rsrc_removal(struct io_rsrc_data *data, unsigned idx,
7832 struct io_rsrc_node *node, void *rsrc)
7834 struct io_rsrc_put *prsrc;
7836 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
7840 prsrc->tag = *io_get_tag_slot(data, idx);
7842 list_add(&prsrc->list, &node->rsrc_list);
7846 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7847 struct io_uring_rsrc_update2 *up,
7850 u64 __user *tags = u64_to_user_ptr(up->tags);
7851 __s32 __user *fds = u64_to_user_ptr(up->data);
7852 struct io_rsrc_data *data = ctx->file_data;
7853 struct io_fixed_file *file_slot;
7857 bool needs_switch = false;
7859 if (!ctx->file_data)
7861 if (up->offset + nr_args > ctx->nr_user_files)
7864 for (done = 0; done < nr_args; done++) {
7867 if ((tags && copy_from_user(&tag, &tags[done], sizeof(tag))) ||
7868 copy_from_user(&fd, &fds[done], sizeof(fd))) {
7872 if ((fd == IORING_REGISTER_FILES_SKIP || fd == -1) && tag) {
7876 if (fd == IORING_REGISTER_FILES_SKIP)
7879 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
7880 file_slot = io_fixed_file_slot(&ctx->file_table, i);
7882 if (file_slot->file_ptr) {
7883 file = (struct file *)(file_slot->file_ptr & FFS_MASK);
7884 err = io_queue_rsrc_removal(data, up->offset + done,
7885 ctx->rsrc_node, file);
7888 file_slot->file_ptr = 0;
7889 needs_switch = true;
7898 * Don't allow io_uring instances to be registered. If
7899 * UNIX isn't enabled, then this causes a reference
7900 * cycle and this instance can never get freed. If UNIX
7901 * is enabled we'll handle it just fine, but there's
7902 * still no point in allowing a ring fd as it doesn't
7903 * support regular read/write anyway.
7905 if (file->f_op == &io_uring_fops) {
7910 *io_get_tag_slot(data, up->offset + done) = tag;
7911 io_fixed_file_set(file_slot, file);
7912 err = io_sqe_file_register(ctx, file, i);
7914 file_slot->file_ptr = 0;
7922 io_rsrc_node_switch(ctx, data);
7923 return done ? done : err;
7926 static struct io_wq_work *io_free_work(struct io_wq_work *work)
7928 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7930 req = io_put_req_find_next(req);
7931 return req ? &req->work : NULL;
7934 static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx,
7935 struct task_struct *task)
7937 struct io_wq_hash *hash;
7938 struct io_wq_data data;
7939 unsigned int concurrency;
7941 hash = ctx->hash_map;
7943 hash = kzalloc(sizeof(*hash), GFP_KERNEL);
7945 return ERR_PTR(-ENOMEM);
7946 refcount_set(&hash->refs, 1);
7947 init_waitqueue_head(&hash->wait);
7948 ctx->hash_map = hash;
7953 data.free_work = io_free_work;
7954 data.do_work = io_wq_submit_work;
7956 /* Do QD, or 4 * CPUS, whatever is smallest */
7957 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7959 return io_wq_create(concurrency, &data);
7962 static int io_uring_alloc_task_context(struct task_struct *task,
7963 struct io_ring_ctx *ctx)
7965 struct io_uring_task *tctx;
7968 tctx = kzalloc(sizeof(*tctx), GFP_KERNEL);
7969 if (unlikely(!tctx))
7972 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
7973 if (unlikely(ret)) {
7978 tctx->io_wq = io_init_wq_offload(ctx, task);
7979 if (IS_ERR(tctx->io_wq)) {
7980 ret = PTR_ERR(tctx->io_wq);
7981 percpu_counter_destroy(&tctx->inflight);
7987 init_waitqueue_head(&tctx->wait);
7988 atomic_set(&tctx->in_idle, 0);
7989 atomic_set(&tctx->inflight_tracked, 0);
7990 task->io_uring = tctx;
7991 spin_lock_init(&tctx->task_lock);
7992 INIT_WQ_LIST(&tctx->task_list);
7993 init_task_work(&tctx->task_work, tctx_task_work);
7997 void __io_uring_free(struct task_struct *tsk)
7999 struct io_uring_task *tctx = tsk->io_uring;
8001 WARN_ON_ONCE(!xa_empty(&tctx->xa));
8002 WARN_ON_ONCE(tctx->io_wq);
8003 WARN_ON_ONCE(tctx->cached_refs);
8005 percpu_counter_destroy(&tctx->inflight);
8007 tsk->io_uring = NULL;
8010 static int io_sq_offload_create(struct io_ring_ctx *ctx,
8011 struct io_uring_params *p)
8015 /* Retain compatibility with failing for an invalid attach attempt */
8016 if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
8017 IORING_SETUP_ATTACH_WQ) {
8020 f = fdget(p->wq_fd);
8024 if (f.file->f_op != &io_uring_fops)
8027 if (ctx->flags & IORING_SETUP_SQPOLL) {
8028 struct task_struct *tsk;
8029 struct io_sq_data *sqd;
8032 sqd = io_get_sq_data(p, &attached);
8038 ctx->sq_creds = get_current_cred();
8040 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
8041 if (!ctx->sq_thread_idle)
8042 ctx->sq_thread_idle = HZ;
8044 io_sq_thread_park(sqd);
8045 list_add(&ctx->sqd_list, &sqd->ctx_list);
8046 io_sqd_update_thread_idle(sqd);
8047 /* don't attach to a dying SQPOLL thread, would be racy */
8048 ret = (attached && !sqd->thread) ? -ENXIO : 0;
8049 io_sq_thread_unpark(sqd);
8056 if (p->flags & IORING_SETUP_SQ_AFF) {
8057 int cpu = p->sq_thread_cpu;
8060 if (cpu >= nr_cpu_ids || !cpu_online(cpu))
8067 sqd->task_pid = current->pid;
8068 sqd->task_tgid = current->tgid;
8069 tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE);
8076 ret = io_uring_alloc_task_context(tsk, ctx);
8077 wake_up_new_task(tsk);
8080 } else if (p->flags & IORING_SETUP_SQ_AFF) {
8081 /* Can't have SQ_AFF without SQPOLL */
8088 complete(&ctx->sq_data->exited);
8090 io_sq_thread_finish(ctx);
8094 static inline void __io_unaccount_mem(struct user_struct *user,
8095 unsigned long nr_pages)
8097 atomic_long_sub(nr_pages, &user->locked_vm);
8100 static inline int __io_account_mem(struct user_struct *user,
8101 unsigned long nr_pages)
8103 unsigned long page_limit, cur_pages, new_pages;
8105 /* Don't allow more pages than we can safely lock */
8106 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
8109 cur_pages = atomic_long_read(&user->locked_vm);
8110 new_pages = cur_pages + nr_pages;
8111 if (new_pages > page_limit)
8113 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
8114 new_pages) != cur_pages);
8119 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8122 __io_unaccount_mem(ctx->user, nr_pages);
8124 if (ctx->mm_account)
8125 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
8128 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8133 ret = __io_account_mem(ctx->user, nr_pages);
8138 if (ctx->mm_account)
8139 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
8144 static void io_mem_free(void *ptr)
8151 page = virt_to_head_page(ptr);
8152 if (put_page_testzero(page))
8153 free_compound_page(page);
8156 static void *io_mem_alloc(size_t size)
8158 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
8159 __GFP_NORETRY | __GFP_ACCOUNT;
8161 return (void *) __get_free_pages(gfp_flags, get_order(size));
8164 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8167 struct io_rings *rings;
8168 size_t off, sq_array_size;
8170 off = struct_size(rings, cqes, cq_entries);
8171 if (off == SIZE_MAX)
8175 off = ALIGN(off, SMP_CACHE_BYTES);
8183 sq_array_size = array_size(sizeof(u32), sq_entries);
8184 if (sq_array_size == SIZE_MAX)
8187 if (check_add_overflow(off, sq_array_size, &off))
8193 static void io_buffer_unmap(struct io_ring_ctx *ctx, struct io_mapped_ubuf **slot)
8195 struct io_mapped_ubuf *imu = *slot;
8198 if (imu != ctx->dummy_ubuf) {
8199 for (i = 0; i < imu->nr_bvecs; i++)
8200 unpin_user_page(imu->bvec[i].bv_page);
8201 if (imu->acct_pages)
8202 io_unaccount_mem(ctx, imu->acct_pages);
8208 static void io_rsrc_buf_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
8210 io_buffer_unmap(ctx, &prsrc->buf);
8214 static void __io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8218 for (i = 0; i < ctx->nr_user_bufs; i++)
8219 io_buffer_unmap(ctx, &ctx->user_bufs[i]);
8220 kfree(ctx->user_bufs);
8221 io_rsrc_data_free(ctx->buf_data);
8222 ctx->user_bufs = NULL;
8223 ctx->buf_data = NULL;
8224 ctx->nr_user_bufs = 0;
8227 static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8234 ret = io_rsrc_ref_quiesce(ctx->buf_data, ctx);
8236 __io_sqe_buffers_unregister(ctx);
8240 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8241 void __user *arg, unsigned index)
8243 struct iovec __user *src;
8245 #ifdef CONFIG_COMPAT
8247 struct compat_iovec __user *ciovs;
8248 struct compat_iovec ciov;
8250 ciovs = (struct compat_iovec __user *) arg;
8251 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8254 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8255 dst->iov_len = ciov.iov_len;
8259 src = (struct iovec __user *) arg;
8260 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8266 * Not super efficient, but this is just a registration time. And we do cache
8267 * the last compound head, so generally we'll only do a full search if we don't
8270 * We check if the given compound head page has already been accounted, to
8271 * avoid double accounting it. This allows us to account the full size of the
8272 * page, not just the constituent pages of a huge page.
8274 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8275 int nr_pages, struct page *hpage)
8279 /* check current page array */
8280 for (i = 0; i < nr_pages; i++) {
8281 if (!PageCompound(pages[i]))
8283 if (compound_head(pages[i]) == hpage)
8287 /* check previously registered pages */
8288 for (i = 0; i < ctx->nr_user_bufs; i++) {
8289 struct io_mapped_ubuf *imu = ctx->user_bufs[i];
8291 for (j = 0; j < imu->nr_bvecs; j++) {
8292 if (!PageCompound(imu->bvec[j].bv_page))
8294 if (compound_head(imu->bvec[j].bv_page) == hpage)
8302 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8303 int nr_pages, struct io_mapped_ubuf *imu,
8304 struct page **last_hpage)
8308 imu->acct_pages = 0;
8309 for (i = 0; i < nr_pages; i++) {
8310 if (!PageCompound(pages[i])) {
8315 hpage = compound_head(pages[i]);
8316 if (hpage == *last_hpage)
8318 *last_hpage = hpage;
8319 if (headpage_already_acct(ctx, pages, i, hpage))
8321 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8325 if (!imu->acct_pages)
8328 ret = io_account_mem(ctx, imu->acct_pages);
8330 imu->acct_pages = 0;
8334 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
8335 struct io_mapped_ubuf **pimu,
8336 struct page **last_hpage)
8338 struct io_mapped_ubuf *imu = NULL;
8339 struct vm_area_struct **vmas = NULL;
8340 struct page **pages = NULL;
8341 unsigned long off, start, end, ubuf;
8343 int ret, pret, nr_pages, i;
8345 if (!iov->iov_base) {
8346 *pimu = ctx->dummy_ubuf;
8350 ubuf = (unsigned long) iov->iov_base;
8351 end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8352 start = ubuf >> PAGE_SHIFT;
8353 nr_pages = end - start;
8358 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
8362 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
8367 imu = kvmalloc(struct_size(imu, bvec, nr_pages), GFP_KERNEL);
8372 mmap_read_lock(current->mm);
8373 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
8375 if (pret == nr_pages) {
8376 /* don't support file backed memory */
8377 for (i = 0; i < nr_pages; i++) {
8378 struct vm_area_struct *vma = vmas[i];
8380 if (vma_is_shmem(vma))
8383 !is_file_hugepages(vma->vm_file)) {
8389 ret = pret < 0 ? pret : -EFAULT;
8391 mmap_read_unlock(current->mm);
8394 * if we did partial map, or found file backed vmas,
8395 * release any pages we did get
8398 unpin_user_pages(pages, pret);
8402 ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
8404 unpin_user_pages(pages, pret);
8408 off = ubuf & ~PAGE_MASK;
8409 size = iov->iov_len;
8410 for (i = 0; i < nr_pages; i++) {
8413 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8414 imu->bvec[i].bv_page = pages[i];
8415 imu->bvec[i].bv_len = vec_len;
8416 imu->bvec[i].bv_offset = off;
8420 /* store original address for later verification */
8422 imu->ubuf_end = ubuf + iov->iov_len;
8423 imu->nr_bvecs = nr_pages;
8434 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
8436 ctx->user_bufs = kcalloc(nr_args, sizeof(*ctx->user_bufs), GFP_KERNEL);
8437 return ctx->user_bufs ? 0 : -ENOMEM;
8440 static int io_buffer_validate(struct iovec *iov)
8442 unsigned long tmp, acct_len = iov->iov_len + (PAGE_SIZE - 1);
8445 * Don't impose further limits on the size and buffer
8446 * constraints here, we'll -EINVAL later when IO is
8447 * submitted if they are wrong.
8450 return iov->iov_len ? -EFAULT : 0;
8454 /* arbitrary limit, but we need something */
8455 if (iov->iov_len > SZ_1G)
8458 if (check_add_overflow((unsigned long)iov->iov_base, acct_len, &tmp))
8464 static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
8465 unsigned int nr_args, u64 __user *tags)
8467 struct page *last_hpage = NULL;
8468 struct io_rsrc_data *data;
8474 if (!nr_args || nr_args > IORING_MAX_REG_BUFFERS)
8476 ret = io_rsrc_node_switch_start(ctx);
8479 ret = io_rsrc_data_alloc(ctx, io_rsrc_buf_put, tags, nr_args, &data);
8482 ret = io_buffers_map_alloc(ctx, nr_args);
8484 io_rsrc_data_free(data);
8488 for (i = 0; i < nr_args; i++, ctx->nr_user_bufs++) {
8489 ret = io_copy_iov(ctx, &iov, arg, i);
8492 ret = io_buffer_validate(&iov);
8495 if (!iov.iov_base && *io_get_tag_slot(data, i)) {
8500 ret = io_sqe_buffer_register(ctx, &iov, &ctx->user_bufs[i],
8506 WARN_ON_ONCE(ctx->buf_data);
8508 ctx->buf_data = data;
8510 __io_sqe_buffers_unregister(ctx);
8512 io_rsrc_node_switch(ctx, NULL);
8516 static int __io_sqe_buffers_update(struct io_ring_ctx *ctx,
8517 struct io_uring_rsrc_update2 *up,
8518 unsigned int nr_args)
8520 u64 __user *tags = u64_to_user_ptr(up->tags);
8521 struct iovec iov, __user *iovs = u64_to_user_ptr(up->data);
8522 struct page *last_hpage = NULL;
8523 bool needs_switch = false;
8529 if (up->offset + nr_args > ctx->nr_user_bufs)
8532 for (done = 0; done < nr_args; done++) {
8533 struct io_mapped_ubuf *imu;
8534 int offset = up->offset + done;
8537 err = io_copy_iov(ctx, &iov, iovs, done);
8540 if (tags && copy_from_user(&tag, &tags[done], sizeof(tag))) {
8544 err = io_buffer_validate(&iov);
8547 if (!iov.iov_base && tag) {
8551 err = io_sqe_buffer_register(ctx, &iov, &imu, &last_hpage);
8555 i = array_index_nospec(offset, ctx->nr_user_bufs);
8556 if (ctx->user_bufs[i] != ctx->dummy_ubuf) {
8557 err = io_queue_rsrc_removal(ctx->buf_data, offset,
8558 ctx->rsrc_node, ctx->user_bufs[i]);
8559 if (unlikely(err)) {
8560 io_buffer_unmap(ctx, &imu);
8563 ctx->user_bufs[i] = NULL;
8564 needs_switch = true;
8567 ctx->user_bufs[i] = imu;
8568 *io_get_tag_slot(ctx->buf_data, offset) = tag;
8572 io_rsrc_node_switch(ctx, ctx->buf_data);
8573 return done ? done : err;
8576 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8578 __s32 __user *fds = arg;
8584 if (copy_from_user(&fd, fds, sizeof(*fds)))
8587 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8588 if (IS_ERR(ctx->cq_ev_fd)) {
8589 int ret = PTR_ERR(ctx->cq_ev_fd);
8591 ctx->cq_ev_fd = NULL;
8598 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8600 if (ctx->cq_ev_fd) {
8601 eventfd_ctx_put(ctx->cq_ev_fd);
8602 ctx->cq_ev_fd = NULL;
8609 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8611 struct io_buffer *buf;
8612 unsigned long index;
8614 xa_for_each(&ctx->io_buffers, index, buf)
8615 __io_remove_buffers(ctx, buf, index, -1U);
8618 static void io_req_cache_free(struct list_head *list, struct task_struct *tsk)
8620 struct io_kiocb *req, *nxt;
8622 list_for_each_entry_safe(req, nxt, list, compl.list) {
8623 if (tsk && req->task != tsk)
8625 list_del(&req->compl.list);
8626 kmem_cache_free(req_cachep, req);
8630 static void io_req_caches_free(struct io_ring_ctx *ctx)
8632 struct io_submit_state *submit_state = &ctx->submit_state;
8633 struct io_comp_state *cs = &ctx->submit_state.comp;
8635 mutex_lock(&ctx->uring_lock);
8637 if (submit_state->free_reqs) {
8638 kmem_cache_free_bulk(req_cachep, submit_state->free_reqs,
8639 submit_state->reqs);
8640 submit_state->free_reqs = 0;
8643 io_flush_cached_locked_reqs(ctx, cs);
8644 io_req_cache_free(&cs->free_list, NULL);
8645 mutex_unlock(&ctx->uring_lock);
8648 static bool io_wait_rsrc_data(struct io_rsrc_data *data)
8652 if (!atomic_dec_and_test(&data->refs))
8653 wait_for_completion(&data->done);
8657 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8659 io_sq_thread_finish(ctx);
8661 if (ctx->mm_account) {
8662 mmdrop(ctx->mm_account);
8663 ctx->mm_account = NULL;
8666 mutex_lock(&ctx->uring_lock);
8667 if (io_wait_rsrc_data(ctx->buf_data))
8668 __io_sqe_buffers_unregister(ctx);
8669 if (io_wait_rsrc_data(ctx->file_data))
8670 __io_sqe_files_unregister(ctx);
8672 __io_cqring_overflow_flush(ctx, true);
8673 mutex_unlock(&ctx->uring_lock);
8674 io_eventfd_unregister(ctx);
8675 io_destroy_buffers(ctx);
8677 put_cred(ctx->sq_creds);
8679 /* there are no registered resources left, nobody uses it */
8681 io_rsrc_node_destroy(ctx->rsrc_node);
8682 if (ctx->rsrc_backup_node)
8683 io_rsrc_node_destroy(ctx->rsrc_backup_node);
8684 flush_delayed_work(&ctx->rsrc_put_work);
8686 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
8687 WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist));
8689 #if defined(CONFIG_UNIX)
8690 if (ctx->ring_sock) {
8691 ctx->ring_sock->file = NULL; /* so that iput() is called */
8692 sock_release(ctx->ring_sock);
8696 io_mem_free(ctx->rings);
8697 io_mem_free(ctx->sq_sqes);
8699 percpu_ref_exit(&ctx->refs);
8700 free_uid(ctx->user);
8701 io_req_caches_free(ctx);
8703 io_wq_put_hash(ctx->hash_map);
8704 kfree(ctx->cancel_hash);
8705 kfree(ctx->dummy_ubuf);
8709 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8711 struct io_ring_ctx *ctx = file->private_data;
8714 poll_wait(file, &ctx->poll_wait, wait);
8716 * synchronizes with barrier from wq_has_sleeper call in
8720 if (!io_sqring_full(ctx))
8721 mask |= EPOLLOUT | EPOLLWRNORM;
8724 * Don't flush cqring overflow list here, just do a simple check.
8725 * Otherwise there could possible be ABBA deadlock:
8728 * lock(&ctx->uring_lock);
8730 * lock(&ctx->uring_lock);
8733 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
8734 * pushs them to do the flush.
8736 if (io_cqring_events(ctx) || test_bit(0, &ctx->check_cq_overflow))
8737 mask |= EPOLLIN | EPOLLRDNORM;
8742 static int io_uring_fasync(int fd, struct file *file, int on)
8744 struct io_ring_ctx *ctx = file->private_data;
8746 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8749 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8751 const struct cred *creds;
8753 creds = xa_erase(&ctx->personalities, id);
8762 static inline bool io_run_ctx_fallback(struct io_ring_ctx *ctx)
8764 return io_run_task_work_head(&ctx->exit_task_work);
8767 struct io_tctx_exit {
8768 struct callback_head task_work;
8769 struct completion completion;
8770 struct io_ring_ctx *ctx;
8773 static void io_tctx_exit_cb(struct callback_head *cb)
8775 struct io_uring_task *tctx = current->io_uring;
8776 struct io_tctx_exit *work;
8778 work = container_of(cb, struct io_tctx_exit, task_work);
8780 * When @in_idle, we're in cancellation and it's racy to remove the
8781 * node. It'll be removed by the end of cancellation, just ignore it.
8783 if (!atomic_read(&tctx->in_idle))
8784 io_uring_del_tctx_node((unsigned long)work->ctx);
8785 complete(&work->completion);
8788 static bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
8790 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8792 return req->ctx == data;
8795 static void io_ring_exit_work(struct work_struct *work)
8797 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
8798 unsigned long timeout = jiffies + HZ * 60 * 5;
8799 struct io_tctx_exit exit;
8800 struct io_tctx_node *node;
8804 * If we're doing polled IO and end up having requests being
8805 * submitted async (out-of-line), then completions can come in while
8806 * we're waiting for refs to drop. We need to reap these manually,
8807 * as nobody else will be looking for them.
8810 io_uring_try_cancel_requests(ctx, NULL, true);
8812 struct io_sq_data *sqd = ctx->sq_data;
8813 struct task_struct *tsk;
8815 io_sq_thread_park(sqd);
8817 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
8818 io_wq_cancel_cb(tsk->io_uring->io_wq,
8819 io_cancel_ctx_cb, ctx, true);
8820 io_sq_thread_unpark(sqd);
8823 WARN_ON_ONCE(time_after(jiffies, timeout));
8824 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
8826 init_completion(&exit.completion);
8827 init_task_work(&exit.task_work, io_tctx_exit_cb);
8830 * Some may use context even when all refs and requests have been put,
8831 * and they are free to do so while still holding uring_lock or
8832 * completion_lock, see __io_req_task_submit(). Apart from other work,
8833 * this lock/unlock section also waits them to finish.
8835 mutex_lock(&ctx->uring_lock);
8836 while (!list_empty(&ctx->tctx_list)) {
8837 WARN_ON_ONCE(time_after(jiffies, timeout));
8839 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
8841 /* don't spin on a single task if cancellation failed */
8842 list_rotate_left(&ctx->tctx_list);
8843 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
8844 if (WARN_ON_ONCE(ret))
8846 wake_up_process(node->task);
8848 mutex_unlock(&ctx->uring_lock);
8849 wait_for_completion(&exit.completion);
8850 mutex_lock(&ctx->uring_lock);
8852 mutex_unlock(&ctx->uring_lock);
8853 spin_lock_irq(&ctx->completion_lock);
8854 spin_unlock_irq(&ctx->completion_lock);
8856 io_ring_ctx_free(ctx);
8859 /* Returns true if we found and killed one or more timeouts */
8860 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
8863 struct io_kiocb *req, *tmp;
8866 spin_lock_irq(&ctx->completion_lock);
8867 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
8868 if (io_match_task(req, tsk, cancel_all)) {
8869 io_kill_timeout(req, -ECANCELED);
8874 io_commit_cqring(ctx);
8875 spin_unlock_irq(&ctx->completion_lock);
8877 io_cqring_ev_posted(ctx);
8878 return canceled != 0;
8881 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8883 unsigned long index;
8884 struct creds *creds;
8886 mutex_lock(&ctx->uring_lock);
8887 percpu_ref_kill(&ctx->refs);
8889 __io_cqring_overflow_flush(ctx, true);
8890 xa_for_each(&ctx->personalities, index, creds)
8891 io_unregister_personality(ctx, index);
8892 mutex_unlock(&ctx->uring_lock);
8894 io_kill_timeouts(ctx, NULL, true);
8895 io_poll_remove_all(ctx, NULL, true);
8897 /* if we failed setting up the ctx, we might not have any rings */
8898 io_iopoll_try_reap_events(ctx);
8900 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8902 * Use system_unbound_wq to avoid spawning tons of event kworkers
8903 * if we're exiting a ton of rings at the same time. It just adds
8904 * noise and overhead, there's no discernable change in runtime
8905 * over using system_wq.
8907 queue_work(system_unbound_wq, &ctx->exit_work);
8910 static int io_uring_release(struct inode *inode, struct file *file)
8912 struct io_ring_ctx *ctx = file->private_data;
8914 file->private_data = NULL;
8915 io_ring_ctx_wait_and_kill(ctx);
8919 struct io_task_cancel {
8920 struct task_struct *task;
8924 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8926 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8927 struct io_task_cancel *cancel = data;
8930 if (!cancel->all && (req->flags & REQ_F_LINK_TIMEOUT)) {
8931 unsigned long flags;
8932 struct io_ring_ctx *ctx = req->ctx;
8934 /* protect against races with linked timeouts */
8935 spin_lock_irqsave(&ctx->completion_lock, flags);
8936 ret = io_match_task(req, cancel->task, cancel->all);
8937 spin_unlock_irqrestore(&ctx->completion_lock, flags);
8939 ret = io_match_task(req, cancel->task, cancel->all);
8944 static bool io_cancel_defer_files(struct io_ring_ctx *ctx,
8945 struct task_struct *task, bool cancel_all)
8947 struct io_defer_entry *de;
8950 spin_lock_irq(&ctx->completion_lock);
8951 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8952 if (io_match_task(de->req, task, cancel_all)) {
8953 list_cut_position(&list, &ctx->defer_list, &de->list);
8957 spin_unlock_irq(&ctx->completion_lock);
8958 if (list_empty(&list))
8961 while (!list_empty(&list)) {
8962 de = list_first_entry(&list, struct io_defer_entry, list);
8963 list_del_init(&de->list);
8964 io_req_complete_failed(de->req, -ECANCELED);
8970 static bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
8972 struct io_tctx_node *node;
8973 enum io_wq_cancel cret;
8976 mutex_lock(&ctx->uring_lock);
8977 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
8978 struct io_uring_task *tctx = node->task->io_uring;
8981 * io_wq will stay alive while we hold uring_lock, because it's
8982 * killed after ctx nodes, which requires to take the lock.
8984 if (!tctx || !tctx->io_wq)
8986 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
8987 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8989 mutex_unlock(&ctx->uring_lock);
8994 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
8995 struct task_struct *task,
8998 struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
8999 struct io_uring_task *tctx = task ? task->io_uring : NULL;
9002 enum io_wq_cancel cret;
9006 ret |= io_uring_try_cancel_iowq(ctx);
9007 } else if (tctx && tctx->io_wq) {
9009 * Cancels requests of all rings, not only @ctx, but
9010 * it's fine as the task is in exit/exec.
9012 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
9014 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
9017 /* SQPOLL thread does its own polling */
9018 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
9019 (ctx->sq_data && ctx->sq_data->thread == current)) {
9020 while (!list_empty_careful(&ctx->iopoll_list)) {
9021 io_iopoll_try_reap_events(ctx);
9026 ret |= io_cancel_defer_files(ctx, task, cancel_all);
9027 ret |= io_poll_remove_all(ctx, task, cancel_all);
9028 ret |= io_kill_timeouts(ctx, task, cancel_all);
9030 ret |= io_run_task_work();
9031 ret |= io_run_ctx_fallback(ctx);
9038 static int __io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9040 struct io_uring_task *tctx = current->io_uring;
9041 struct io_tctx_node *node;
9044 if (unlikely(!tctx)) {
9045 ret = io_uring_alloc_task_context(current, ctx);
9048 tctx = current->io_uring;
9050 if (!xa_load(&tctx->xa, (unsigned long)ctx)) {
9051 node = kmalloc(sizeof(*node), GFP_KERNEL);
9055 node->task = current;
9057 ret = xa_err(xa_store(&tctx->xa, (unsigned long)ctx,
9064 mutex_lock(&ctx->uring_lock);
9065 list_add(&node->ctx_node, &ctx->tctx_list);
9066 mutex_unlock(&ctx->uring_lock);
9073 * Note that this task has used io_uring. We use it for cancelation purposes.
9075 static inline int io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9077 struct io_uring_task *tctx = current->io_uring;
9079 if (likely(tctx && tctx->last == ctx))
9081 return __io_uring_add_tctx_node(ctx);
9085 * Remove this io_uring_file -> task mapping.
9087 static void io_uring_del_tctx_node(unsigned long index)
9089 struct io_uring_task *tctx = current->io_uring;
9090 struct io_tctx_node *node;
9094 node = xa_erase(&tctx->xa, index);
9098 WARN_ON_ONCE(current != node->task);
9099 WARN_ON_ONCE(list_empty(&node->ctx_node));
9101 mutex_lock(&node->ctx->uring_lock);
9102 list_del(&node->ctx_node);
9103 mutex_unlock(&node->ctx->uring_lock);
9105 if (tctx->last == node->ctx)
9110 static void io_uring_clean_tctx(struct io_uring_task *tctx)
9112 struct io_wq *wq = tctx->io_wq;
9113 struct io_tctx_node *node;
9114 unsigned long index;
9116 xa_for_each(&tctx->xa, index, node)
9117 io_uring_del_tctx_node(index);
9120 * Must be after io_uring_del_task_file() (removes nodes under
9121 * uring_lock) to avoid race with io_uring_try_cancel_iowq().
9124 io_wq_put_and_exit(wq);
9128 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
9131 return atomic_read(&tctx->inflight_tracked);
9132 return percpu_counter_sum(&tctx->inflight);
9135 static void io_uring_drop_tctx_refs(struct task_struct *task)
9137 struct io_uring_task *tctx = task->io_uring;
9138 unsigned int refs = tctx->cached_refs;
9140 tctx->cached_refs = 0;
9141 percpu_counter_sub(&tctx->inflight, refs);
9142 put_task_struct_many(task, refs);
9146 * Find any io_uring ctx that this task has registered or done IO on, and cancel
9147 * requests. @sqd should be not-null IIF it's an SQPOLL thread cancellation.
9149 static void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
9151 struct io_uring_task *tctx = current->io_uring;
9152 struct io_ring_ctx *ctx;
9156 WARN_ON_ONCE(sqd && sqd->thread != current);
9158 if (!current->io_uring)
9161 io_wq_exit_start(tctx->io_wq);
9163 io_uring_drop_tctx_refs(current);
9164 atomic_inc(&tctx->in_idle);
9166 /* read completions before cancelations */
9167 inflight = tctx_inflight(tctx, !cancel_all);
9172 struct io_tctx_node *node;
9173 unsigned long index;
9175 xa_for_each(&tctx->xa, index, node) {
9176 /* sqpoll task will cancel all its requests */
9177 if (node->ctx->sq_data)
9179 io_uring_try_cancel_requests(node->ctx, current,
9183 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
9184 io_uring_try_cancel_requests(ctx, current,
9188 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
9190 * If we've seen completions, retry without waiting. This
9191 * avoids a race where a completion comes in before we did
9192 * prepare_to_wait().
9194 if (inflight == tctx_inflight(tctx, !cancel_all))
9196 finish_wait(&tctx->wait, &wait);
9198 atomic_dec(&tctx->in_idle);
9200 io_uring_clean_tctx(tctx);
9202 /* for exec all current's requests should be gone, kill tctx */
9203 __io_uring_free(current);
9207 void __io_uring_cancel(struct files_struct *files)
9209 io_uring_cancel_generic(!files, NULL);
9212 static void *io_uring_validate_mmap_request(struct file *file,
9213 loff_t pgoff, size_t sz)
9215 struct io_ring_ctx *ctx = file->private_data;
9216 loff_t offset = pgoff << PAGE_SHIFT;
9221 case IORING_OFF_SQ_RING:
9222 case IORING_OFF_CQ_RING:
9225 case IORING_OFF_SQES:
9229 return ERR_PTR(-EINVAL);
9232 page = virt_to_head_page(ptr);
9233 if (sz > page_size(page))
9234 return ERR_PTR(-EINVAL);
9241 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9243 size_t sz = vma->vm_end - vma->vm_start;
9247 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
9249 return PTR_ERR(ptr);
9251 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
9252 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
9255 #else /* !CONFIG_MMU */
9257 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9259 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
9262 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
9264 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
9267 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
9268 unsigned long addr, unsigned long len,
9269 unsigned long pgoff, unsigned long flags)
9273 ptr = io_uring_validate_mmap_request(file, pgoff, len);
9275 return PTR_ERR(ptr);
9277 return (unsigned long) ptr;
9280 #endif /* !CONFIG_MMU */
9282 static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
9287 if (!io_sqring_full(ctx))
9289 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9291 if (!io_sqring_full(ctx))
9294 } while (!signal_pending(current));
9296 finish_wait(&ctx->sqo_sq_wait, &wait);
9300 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9301 struct __kernel_timespec __user **ts,
9302 const sigset_t __user **sig)
9304 struct io_uring_getevents_arg arg;
9307 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9308 * is just a pointer to the sigset_t.
9310 if (!(flags & IORING_ENTER_EXT_ARG)) {
9311 *sig = (const sigset_t __user *) argp;
9317 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9318 * timespec and sigset_t pointers if good.
9320 if (*argsz != sizeof(arg))
9322 if (copy_from_user(&arg, argp, sizeof(arg)))
9324 *sig = u64_to_user_ptr(arg.sigmask);
9325 *argsz = arg.sigmask_sz;
9326 *ts = u64_to_user_ptr(arg.ts);
9330 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9331 u32, min_complete, u32, flags, const void __user *, argp,
9334 struct io_ring_ctx *ctx;
9341 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9342 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG)))
9346 if (unlikely(!f.file))
9350 if (unlikely(f.file->f_op != &io_uring_fops))
9354 ctx = f.file->private_data;
9355 if (unlikely(!percpu_ref_tryget(&ctx->refs)))
9359 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
9363 * For SQ polling, the thread will do all submissions and completions.
9364 * Just return the requested submit count, and wake the thread if
9368 if (ctx->flags & IORING_SETUP_SQPOLL) {
9369 io_cqring_overflow_flush(ctx, false);
9372 if (unlikely(ctx->sq_data->thread == NULL))
9374 if (flags & IORING_ENTER_SQ_WAKEUP)
9375 wake_up(&ctx->sq_data->wait);
9376 if (flags & IORING_ENTER_SQ_WAIT) {
9377 ret = io_sqpoll_wait_sq(ctx);
9381 submitted = to_submit;
9382 } else if (to_submit) {
9383 ret = io_uring_add_tctx_node(ctx);
9386 mutex_lock(&ctx->uring_lock);
9387 submitted = io_submit_sqes(ctx, to_submit);
9388 mutex_unlock(&ctx->uring_lock);
9390 if (submitted != to_submit)
9393 if (flags & IORING_ENTER_GETEVENTS) {
9394 const sigset_t __user *sig;
9395 struct __kernel_timespec __user *ts;
9397 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
9401 min_complete = min(min_complete, ctx->cq_entries);
9404 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9405 * space applications don't need to do io completion events
9406 * polling again, they can rely on io_sq_thread to do polling
9407 * work, which can reduce cpu usage and uring_lock contention.
9409 if (ctx->flags & IORING_SETUP_IOPOLL &&
9410 !(ctx->flags & IORING_SETUP_SQPOLL)) {
9411 ret = io_iopoll_check(ctx, min_complete);
9413 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
9418 percpu_ref_put(&ctx->refs);
9421 return submitted ? submitted : ret;
9424 #ifdef CONFIG_PROC_FS
9425 static int io_uring_show_cred(struct seq_file *m, unsigned int id,
9426 const struct cred *cred)
9428 struct user_namespace *uns = seq_user_ns(m);
9429 struct group_info *gi;
9434 seq_printf(m, "%5d\n", id);
9435 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
9436 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
9437 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
9438 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
9439 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
9440 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
9441 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
9442 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
9443 seq_puts(m, "\n\tGroups:\t");
9444 gi = cred->group_info;
9445 for (g = 0; g < gi->ngroups; g++) {
9446 seq_put_decimal_ull(m, g ? " " : "",
9447 from_kgid_munged(uns, gi->gid[g]));
9449 seq_puts(m, "\n\tCapEff:\t");
9450 cap = cred->cap_effective;
9451 CAP_FOR_EACH_U32(__capi)
9452 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
9457 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
9459 struct io_sq_data *sq = NULL;
9464 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9465 * since fdinfo case grabs it in the opposite direction of normal use
9466 * cases. If we fail to get the lock, we just don't iterate any
9467 * structures that could be going away outside the io_uring mutex.
9469 has_lock = mutex_trylock(&ctx->uring_lock);
9471 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) {
9477 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9478 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9479 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9480 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9481 struct file *f = io_file_from_index(ctx, i);
9484 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9486 seq_printf(m, "%5u: <none>\n", i);
9488 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9489 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9490 struct io_mapped_ubuf *buf = ctx->user_bufs[i];
9491 unsigned int len = buf->ubuf_end - buf->ubuf;
9493 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf, len);
9495 if (has_lock && !xa_empty(&ctx->personalities)) {
9496 unsigned long index;
9497 const struct cred *cred;
9499 seq_printf(m, "Personalities:\n");
9500 xa_for_each(&ctx->personalities, index, cred)
9501 io_uring_show_cred(m, index, cred);
9503 seq_printf(m, "PollList:\n");
9504 spin_lock_irq(&ctx->completion_lock);
9505 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9506 struct hlist_head *list = &ctx->cancel_hash[i];
9507 struct io_kiocb *req;
9509 hlist_for_each_entry(req, list, hash_node)
9510 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9511 req->task->task_works != NULL);
9513 spin_unlock_irq(&ctx->completion_lock);
9515 mutex_unlock(&ctx->uring_lock);
9518 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9520 struct io_ring_ctx *ctx = f->private_data;
9522 if (percpu_ref_tryget(&ctx->refs)) {
9523 __io_uring_show_fdinfo(ctx, m);
9524 percpu_ref_put(&ctx->refs);
9529 static const struct file_operations io_uring_fops = {
9530 .release = io_uring_release,
9531 .mmap = io_uring_mmap,
9533 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9534 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9536 .poll = io_uring_poll,
9537 .fasync = io_uring_fasync,
9538 #ifdef CONFIG_PROC_FS
9539 .show_fdinfo = io_uring_show_fdinfo,
9543 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9544 struct io_uring_params *p)
9546 struct io_rings *rings;
9547 size_t size, sq_array_offset;
9549 /* make sure these are sane, as we already accounted them */
9550 ctx->sq_entries = p->sq_entries;
9551 ctx->cq_entries = p->cq_entries;
9553 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9554 if (size == SIZE_MAX)
9557 rings = io_mem_alloc(size);
9562 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9563 rings->sq_ring_mask = p->sq_entries - 1;
9564 rings->cq_ring_mask = p->cq_entries - 1;
9565 rings->sq_ring_entries = p->sq_entries;
9566 rings->cq_ring_entries = p->cq_entries;
9568 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9569 if (size == SIZE_MAX) {
9570 io_mem_free(ctx->rings);
9575 ctx->sq_sqes = io_mem_alloc(size);
9576 if (!ctx->sq_sqes) {
9577 io_mem_free(ctx->rings);
9585 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
9589 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9593 ret = io_uring_add_tctx_node(ctx);
9598 fd_install(fd, file);
9603 * Allocate an anonymous fd, this is what constitutes the application
9604 * visible backing of an io_uring instance. The application mmaps this
9605 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9606 * we have to tie this fd to a socket for file garbage collection purposes.
9608 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
9611 #if defined(CONFIG_UNIX)
9614 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9617 return ERR_PTR(ret);
9620 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9621 O_RDWR | O_CLOEXEC);
9622 #if defined(CONFIG_UNIX)
9624 sock_release(ctx->ring_sock);
9625 ctx->ring_sock = NULL;
9627 ctx->ring_sock->file = file;
9633 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9634 struct io_uring_params __user *params)
9636 struct io_ring_ctx *ctx;
9642 if (entries > IORING_MAX_ENTRIES) {
9643 if (!(p->flags & IORING_SETUP_CLAMP))
9645 entries = IORING_MAX_ENTRIES;
9649 * Use twice as many entries for the CQ ring. It's possible for the
9650 * application to drive a higher depth than the size of the SQ ring,
9651 * since the sqes are only used at submission time. This allows for
9652 * some flexibility in overcommitting a bit. If the application has
9653 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9654 * of CQ ring entries manually.
9656 p->sq_entries = roundup_pow_of_two(entries);
9657 if (p->flags & IORING_SETUP_CQSIZE) {
9659 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9660 * to a power-of-two, if it isn't already. We do NOT impose
9661 * any cq vs sq ring sizing.
9665 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9666 if (!(p->flags & IORING_SETUP_CLAMP))
9668 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9670 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9671 if (p->cq_entries < p->sq_entries)
9674 p->cq_entries = 2 * p->sq_entries;
9677 ctx = io_ring_ctx_alloc(p);
9680 ctx->compat = in_compat_syscall();
9681 if (!capable(CAP_IPC_LOCK))
9682 ctx->user = get_uid(current_user());
9685 * This is just grabbed for accounting purposes. When a process exits,
9686 * the mm is exited and dropped before the files, hence we need to hang
9687 * on to this mm purely for the purposes of being able to unaccount
9688 * memory (locked/pinned vm). It's not used for anything else.
9690 mmgrab(current->mm);
9691 ctx->mm_account = current->mm;
9693 ret = io_allocate_scq_urings(ctx, p);
9697 ret = io_sq_offload_create(ctx, p);
9700 /* always set a rsrc node */
9701 ret = io_rsrc_node_switch_start(ctx);
9704 io_rsrc_node_switch(ctx, NULL);
9706 memset(&p->sq_off, 0, sizeof(p->sq_off));
9707 p->sq_off.head = offsetof(struct io_rings, sq.head);
9708 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9709 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9710 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9711 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9712 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9713 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9715 memset(&p->cq_off, 0, sizeof(p->cq_off));
9716 p->cq_off.head = offsetof(struct io_rings, cq.head);
9717 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9718 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9719 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9720 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9721 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9722 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9724 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9725 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9726 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9727 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
9728 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
9729 IORING_FEAT_RSRC_TAGS;
9731 if (copy_to_user(params, p, sizeof(*p))) {
9736 file = io_uring_get_file(ctx);
9738 ret = PTR_ERR(file);
9743 * Install ring fd as the very last thing, so we don't risk someone
9744 * having closed it before we finish setup
9746 ret = io_uring_install_fd(ctx, file);
9748 /* fput will clean it up */
9753 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9756 io_ring_ctx_wait_and_kill(ctx);
9761 * Sets up an aio uring context, and returns the fd. Applications asks for a
9762 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9763 * params structure passed in.
9765 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9767 struct io_uring_params p;
9770 if (copy_from_user(&p, params, sizeof(p)))
9772 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9777 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9778 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9779 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9780 IORING_SETUP_R_DISABLED))
9783 return io_uring_create(entries, &p, params);
9786 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9787 struct io_uring_params __user *, params)
9789 return io_uring_setup(entries, params);
9792 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9794 struct io_uring_probe *p;
9798 size = struct_size(p, ops, nr_args);
9799 if (size == SIZE_MAX)
9801 p = kzalloc(size, GFP_KERNEL);
9806 if (copy_from_user(p, arg, size))
9809 if (memchr_inv(p, 0, size))
9812 p->last_op = IORING_OP_LAST - 1;
9813 if (nr_args > IORING_OP_LAST)
9814 nr_args = IORING_OP_LAST;
9816 for (i = 0; i < nr_args; i++) {
9818 if (!io_op_defs[i].not_supported)
9819 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9824 if (copy_to_user(arg, p, size))
9831 static int io_register_personality(struct io_ring_ctx *ctx)
9833 const struct cred *creds;
9837 creds = get_current_cred();
9839 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
9840 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
9847 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9848 unsigned int nr_args)
9850 struct io_uring_restriction *res;
9854 /* Restrictions allowed only if rings started disabled */
9855 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9858 /* We allow only a single restrictions registration */
9859 if (ctx->restrictions.registered)
9862 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9865 size = array_size(nr_args, sizeof(*res));
9866 if (size == SIZE_MAX)
9869 res = memdup_user(arg, size);
9871 return PTR_ERR(res);
9875 for (i = 0; i < nr_args; i++) {
9876 switch (res[i].opcode) {
9877 case IORING_RESTRICTION_REGISTER_OP:
9878 if (res[i].register_op >= IORING_REGISTER_LAST) {
9883 __set_bit(res[i].register_op,
9884 ctx->restrictions.register_op);
9886 case IORING_RESTRICTION_SQE_OP:
9887 if (res[i].sqe_op >= IORING_OP_LAST) {
9892 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9894 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9895 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9897 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9898 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9907 /* Reset all restrictions if an error happened */
9909 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9911 ctx->restrictions.registered = true;
9917 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9919 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9922 if (ctx->restrictions.registered)
9923 ctx->restricted = 1;
9925 ctx->flags &= ~IORING_SETUP_R_DISABLED;
9926 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
9927 wake_up(&ctx->sq_data->wait);
9931 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
9932 struct io_uring_rsrc_update2 *up,
9940 if (check_add_overflow(up->offset, nr_args, &tmp))
9942 err = io_rsrc_node_switch_start(ctx);
9947 case IORING_RSRC_FILE:
9948 return __io_sqe_files_update(ctx, up, nr_args);
9949 case IORING_RSRC_BUFFER:
9950 return __io_sqe_buffers_update(ctx, up, nr_args);
9955 static int io_register_files_update(struct io_ring_ctx *ctx, void __user *arg,
9958 struct io_uring_rsrc_update2 up;
9962 memset(&up, 0, sizeof(up));
9963 if (copy_from_user(&up, arg, sizeof(struct io_uring_rsrc_update)))
9965 return __io_register_rsrc_update(ctx, IORING_RSRC_FILE, &up, nr_args);
9968 static int io_register_rsrc_update(struct io_ring_ctx *ctx, void __user *arg,
9969 unsigned size, unsigned type)
9971 struct io_uring_rsrc_update2 up;
9973 if (size != sizeof(up))
9975 if (copy_from_user(&up, arg, sizeof(up)))
9977 if (!up.nr || up.resv)
9979 return __io_register_rsrc_update(ctx, type, &up, up.nr);
9982 static int io_register_rsrc(struct io_ring_ctx *ctx, void __user *arg,
9983 unsigned int size, unsigned int type)
9985 struct io_uring_rsrc_register rr;
9987 /* keep it extendible */
9988 if (size != sizeof(rr))
9991 memset(&rr, 0, sizeof(rr));
9992 if (copy_from_user(&rr, arg, size))
9994 if (!rr.nr || rr.resv || rr.resv2)
9998 case IORING_RSRC_FILE:
9999 return io_sqe_files_register(ctx, u64_to_user_ptr(rr.data),
10000 rr.nr, u64_to_user_ptr(rr.tags));
10001 case IORING_RSRC_BUFFER:
10002 return io_sqe_buffers_register(ctx, u64_to_user_ptr(rr.data),
10003 rr.nr, u64_to_user_ptr(rr.tags));
10008 static int io_register_iowq_aff(struct io_ring_ctx *ctx, void __user *arg,
10011 struct io_uring_task *tctx = current->io_uring;
10012 cpumask_var_t new_mask;
10015 if (!tctx || !tctx->io_wq)
10018 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
10021 cpumask_clear(new_mask);
10022 if (len > cpumask_size())
10023 len = cpumask_size();
10025 if (copy_from_user(new_mask, arg, len)) {
10026 free_cpumask_var(new_mask);
10030 ret = io_wq_cpu_affinity(tctx->io_wq, new_mask);
10031 free_cpumask_var(new_mask);
10035 static int io_unregister_iowq_aff(struct io_ring_ctx *ctx)
10037 struct io_uring_task *tctx = current->io_uring;
10039 if (!tctx || !tctx->io_wq)
10042 return io_wq_cpu_affinity(tctx->io_wq, NULL);
10045 static bool io_register_op_must_quiesce(int op)
10048 case IORING_REGISTER_BUFFERS:
10049 case IORING_UNREGISTER_BUFFERS:
10050 case IORING_REGISTER_FILES:
10051 case IORING_UNREGISTER_FILES:
10052 case IORING_REGISTER_FILES_UPDATE:
10053 case IORING_REGISTER_PROBE:
10054 case IORING_REGISTER_PERSONALITY:
10055 case IORING_UNREGISTER_PERSONALITY:
10056 case IORING_REGISTER_FILES2:
10057 case IORING_REGISTER_FILES_UPDATE2:
10058 case IORING_REGISTER_BUFFERS2:
10059 case IORING_REGISTER_BUFFERS_UPDATE:
10060 case IORING_REGISTER_IOWQ_AFF:
10061 case IORING_UNREGISTER_IOWQ_AFF:
10068 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
10069 void __user *arg, unsigned nr_args)
10070 __releases(ctx->uring_lock)
10071 __acquires(ctx->uring_lock)
10076 * We're inside the ring mutex, if the ref is already dying, then
10077 * someone else killed the ctx or is already going through
10078 * io_uring_register().
10080 if (percpu_ref_is_dying(&ctx->refs))
10083 if (ctx->restricted) {
10084 if (opcode >= IORING_REGISTER_LAST)
10086 opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
10087 if (!test_bit(opcode, ctx->restrictions.register_op))
10091 if (io_register_op_must_quiesce(opcode)) {
10092 percpu_ref_kill(&ctx->refs);
10095 * Drop uring mutex before waiting for references to exit. If
10096 * another thread is currently inside io_uring_enter() it might
10097 * need to grab the uring_lock to make progress. If we hold it
10098 * here across the drain wait, then we can deadlock. It's safe
10099 * to drop the mutex here, since no new references will come in
10100 * after we've killed the percpu ref.
10102 mutex_unlock(&ctx->uring_lock);
10104 ret = wait_for_completion_interruptible(&ctx->ref_comp);
10107 ret = io_run_task_work_sig();
10111 mutex_lock(&ctx->uring_lock);
10114 io_refs_resurrect(&ctx->refs, &ctx->ref_comp);
10120 case IORING_REGISTER_BUFFERS:
10121 ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
10123 case IORING_UNREGISTER_BUFFERS:
10125 if (arg || nr_args)
10127 ret = io_sqe_buffers_unregister(ctx);
10129 case IORING_REGISTER_FILES:
10130 ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
10132 case IORING_UNREGISTER_FILES:
10134 if (arg || nr_args)
10136 ret = io_sqe_files_unregister(ctx);
10138 case IORING_REGISTER_FILES_UPDATE:
10139 ret = io_register_files_update(ctx, arg, nr_args);
10141 case IORING_REGISTER_EVENTFD:
10142 case IORING_REGISTER_EVENTFD_ASYNC:
10146 ret = io_eventfd_register(ctx, arg);
10149 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
10150 ctx->eventfd_async = 1;
10152 ctx->eventfd_async = 0;
10154 case IORING_UNREGISTER_EVENTFD:
10156 if (arg || nr_args)
10158 ret = io_eventfd_unregister(ctx);
10160 case IORING_REGISTER_PROBE:
10162 if (!arg || nr_args > 256)
10164 ret = io_probe(ctx, arg, nr_args);
10166 case IORING_REGISTER_PERSONALITY:
10168 if (arg || nr_args)
10170 ret = io_register_personality(ctx);
10172 case IORING_UNREGISTER_PERSONALITY:
10176 ret = io_unregister_personality(ctx, nr_args);
10178 case IORING_REGISTER_ENABLE_RINGS:
10180 if (arg || nr_args)
10182 ret = io_register_enable_rings(ctx);
10184 case IORING_REGISTER_RESTRICTIONS:
10185 ret = io_register_restrictions(ctx, arg, nr_args);
10187 case IORING_REGISTER_FILES2:
10188 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE);
10190 case IORING_REGISTER_FILES_UPDATE2:
10191 ret = io_register_rsrc_update(ctx, arg, nr_args,
10194 case IORING_REGISTER_BUFFERS2:
10195 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER);
10197 case IORING_REGISTER_BUFFERS_UPDATE:
10198 ret = io_register_rsrc_update(ctx, arg, nr_args,
10199 IORING_RSRC_BUFFER);
10201 case IORING_REGISTER_IOWQ_AFF:
10203 if (!arg || !nr_args)
10205 ret = io_register_iowq_aff(ctx, arg, nr_args);
10207 case IORING_UNREGISTER_IOWQ_AFF:
10209 if (arg || nr_args)
10211 ret = io_unregister_iowq_aff(ctx);
10218 if (io_register_op_must_quiesce(opcode)) {
10219 /* bring the ctx back to life */
10220 percpu_ref_reinit(&ctx->refs);
10221 reinit_completion(&ctx->ref_comp);
10226 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
10227 void __user *, arg, unsigned int, nr_args)
10229 struct io_ring_ctx *ctx;
10238 if (f.file->f_op != &io_uring_fops)
10241 ctx = f.file->private_data;
10243 io_run_task_work();
10245 mutex_lock(&ctx->uring_lock);
10246 ret = __io_uring_register(ctx, opcode, arg, nr_args);
10247 mutex_unlock(&ctx->uring_lock);
10248 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
10249 ctx->cq_ev_fd != NULL, ret);
10255 static int __init io_uring_init(void)
10257 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
10258 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
10259 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
10262 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
10263 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
10264 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
10265 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
10266 BUILD_BUG_SQE_ELEM(1, __u8, flags);
10267 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
10268 BUILD_BUG_SQE_ELEM(4, __s32, fd);
10269 BUILD_BUG_SQE_ELEM(8, __u64, off);
10270 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
10271 BUILD_BUG_SQE_ELEM(16, __u64, addr);
10272 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
10273 BUILD_BUG_SQE_ELEM(24, __u32, len);
10274 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
10275 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
10276 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
10277 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
10278 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
10279 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
10280 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
10281 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
10282 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
10283 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
10284 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
10285 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
10286 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
10287 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
10288 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
10289 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
10290 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
10291 BUILD_BUG_SQE_ELEM(40, __u16, buf_group);
10292 BUILD_BUG_SQE_ELEM(42, __u16, personality);
10293 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
10295 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
10296 sizeof(struct io_uring_rsrc_update));
10297 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
10298 sizeof(struct io_uring_rsrc_update2));
10299 /* should fit into one byte */
10300 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
10302 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
10303 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
10305 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
10309 __initcall(io_uring_init);