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_queue;
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_queue && *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_queue = false;
2557 } else if (!ctx->poll_multi_queue) {
2558 struct io_kiocb *list_req;
2559 unsigned int queue_num0, queue_num1;
2561 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2564 if (list_req->file != req->file) {
2565 ctx->poll_multi_queue = true;
2567 queue_num0 = blk_qc_t_to_queue_num(list_req->rw.kiocb.ki_cookie);
2568 queue_num1 = blk_qc_t_to_queue_num(req->rw.kiocb.ki_cookie);
2569 if (queue_num0 != queue_num1)
2570 ctx->poll_multi_queue = true;
2575 * For fast devices, IO may have already completed. If it has, add
2576 * it to the front so we find it first.
2578 if (READ_ONCE(req->iopoll_completed))
2579 list_add(&req->inflight_entry, &ctx->iopoll_list);
2581 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2583 if (unlikely(in_async)) {
2585 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
2586 * in sq thread task context or in io worker task context. If
2587 * current task context is sq thread, we don't need to check
2588 * whether should wake up sq thread.
2590 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2591 wq_has_sleeper(&ctx->sq_data->wait))
2592 wake_up(&ctx->sq_data->wait);
2594 mutex_unlock(&ctx->uring_lock);
2598 static inline void io_state_file_put(struct io_submit_state *state)
2600 if (state->file_refs) {
2601 fput_many(state->file, state->file_refs);
2602 state->file_refs = 0;
2607 * Get as many references to a file as we have IOs left in this submission,
2608 * assuming most submissions are for one file, or at least that each file
2609 * has more than one submission.
2611 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2616 if (state->file_refs) {
2617 if (state->fd == fd) {
2621 io_state_file_put(state);
2623 state->file = fget_many(fd, state->ios_left);
2624 if (unlikely(!state->file))
2628 state->file_refs = state->ios_left - 1;
2632 static bool io_bdev_nowait(struct block_device *bdev)
2634 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2638 * If we tracked the file through the SCM inflight mechanism, we could support
2639 * any file. For now, just ensure that anything potentially problematic is done
2642 static bool __io_file_supports_async(struct file *file, int rw)
2644 umode_t mode = file_inode(file)->i_mode;
2646 if (S_ISBLK(mode)) {
2647 if (IS_ENABLED(CONFIG_BLOCK) &&
2648 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2654 if (S_ISREG(mode)) {
2655 if (IS_ENABLED(CONFIG_BLOCK) &&
2656 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2657 file->f_op != &io_uring_fops)
2662 /* any ->read/write should understand O_NONBLOCK */
2663 if (file->f_flags & O_NONBLOCK)
2666 if (!(file->f_mode & FMODE_NOWAIT))
2670 return file->f_op->read_iter != NULL;
2672 return file->f_op->write_iter != NULL;
2675 static bool io_file_supports_async(struct io_kiocb *req, int rw)
2677 if (rw == READ && (req->flags & REQ_F_ASYNC_READ))
2679 else if (rw == WRITE && (req->flags & REQ_F_ASYNC_WRITE))
2682 return __io_file_supports_async(req->file, rw);
2685 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2687 struct io_ring_ctx *ctx = req->ctx;
2688 struct kiocb *kiocb = &req->rw.kiocb;
2689 struct file *file = req->file;
2693 if (!(req->flags & REQ_F_ISREG) && S_ISREG(file_inode(file)->i_mode))
2694 req->flags |= REQ_F_ISREG;
2696 kiocb->ki_pos = READ_ONCE(sqe->off);
2697 if (kiocb->ki_pos == -1 && !(file->f_mode & FMODE_STREAM)) {
2698 req->flags |= REQ_F_CUR_POS;
2699 kiocb->ki_pos = file->f_pos;
2701 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2702 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2703 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2707 /* don't allow async punt for O_NONBLOCK or RWF_NOWAIT */
2708 if ((kiocb->ki_flags & IOCB_NOWAIT) || (file->f_flags & O_NONBLOCK))
2709 req->flags |= REQ_F_NOWAIT;
2711 ioprio = READ_ONCE(sqe->ioprio);
2713 ret = ioprio_check_cap(ioprio);
2717 kiocb->ki_ioprio = ioprio;
2719 kiocb->ki_ioprio = get_current_ioprio();
2721 if (ctx->flags & IORING_SETUP_IOPOLL) {
2722 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2723 !kiocb->ki_filp->f_op->iopoll)
2726 kiocb->ki_flags |= IOCB_HIPRI;
2727 kiocb->ki_complete = io_complete_rw_iopoll;
2728 req->iopoll_completed = 0;
2730 if (kiocb->ki_flags & IOCB_HIPRI)
2732 kiocb->ki_complete = io_complete_rw;
2735 if (req->opcode == IORING_OP_READ_FIXED ||
2736 req->opcode == IORING_OP_WRITE_FIXED) {
2738 io_req_set_rsrc_node(req);
2741 req->rw.addr = READ_ONCE(sqe->addr);
2742 req->rw.len = READ_ONCE(sqe->len);
2743 req->buf_index = READ_ONCE(sqe->buf_index);
2747 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2753 case -ERESTARTNOINTR:
2754 case -ERESTARTNOHAND:
2755 case -ERESTART_RESTARTBLOCK:
2757 * We can't just restart the syscall, since previously
2758 * submitted sqes may already be in progress. Just fail this
2764 kiocb->ki_complete(kiocb, ret, 0);
2768 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2769 unsigned int issue_flags)
2771 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2772 struct io_async_rw *io = req->async_data;
2773 bool check_reissue = kiocb->ki_complete == io_complete_rw;
2775 /* add previously done IO, if any */
2776 if (io && io->bytes_done > 0) {
2778 ret = io->bytes_done;
2780 ret += io->bytes_done;
2783 if (req->flags & REQ_F_CUR_POS)
2784 req->file->f_pos = kiocb->ki_pos;
2785 if (ret >= 0 && check_reissue)
2786 __io_complete_rw(req, ret, 0, issue_flags);
2788 io_rw_done(kiocb, ret);
2790 if (check_reissue && (req->flags & REQ_F_REISSUE)) {
2791 req->flags &= ~REQ_F_REISSUE;
2792 if (io_resubmit_prep(req)) {
2794 io_queue_async_work(req);
2799 if (req->flags & REQ_F_BUFFER_SELECTED)
2800 cflags = io_put_rw_kbuf(req);
2801 __io_req_complete(req, issue_flags, ret, cflags);
2806 static int __io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter,
2807 struct io_mapped_ubuf *imu)
2809 size_t len = req->rw.len;
2810 u64 buf_end, buf_addr = req->rw.addr;
2813 if (unlikely(check_add_overflow(buf_addr, (u64)len, &buf_end)))
2815 /* not inside the mapped region */
2816 if (unlikely(buf_addr < imu->ubuf || buf_end > imu->ubuf_end))
2820 * May not be a start of buffer, set size appropriately
2821 * and advance us to the beginning.
2823 offset = buf_addr - imu->ubuf;
2824 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2828 * Don't use iov_iter_advance() here, as it's really slow for
2829 * using the latter parts of a big fixed buffer - it iterates
2830 * over each segment manually. We can cheat a bit here, because
2833 * 1) it's a BVEC iter, we set it up
2834 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2835 * first and last bvec
2837 * So just find our index, and adjust the iterator afterwards.
2838 * If the offset is within the first bvec (or the whole first
2839 * bvec, just use iov_iter_advance(). This makes it easier
2840 * since we can just skip the first segment, which may not
2841 * be PAGE_SIZE aligned.
2843 const struct bio_vec *bvec = imu->bvec;
2845 if (offset <= bvec->bv_len) {
2846 iov_iter_advance(iter, offset);
2848 unsigned long seg_skip;
2850 /* skip first vec */
2851 offset -= bvec->bv_len;
2852 seg_skip = 1 + (offset >> PAGE_SHIFT);
2854 iter->bvec = bvec + seg_skip;
2855 iter->nr_segs -= seg_skip;
2856 iter->count -= bvec->bv_len + offset;
2857 iter->iov_offset = offset & ~PAGE_MASK;
2864 static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
2866 struct io_ring_ctx *ctx = req->ctx;
2867 struct io_mapped_ubuf *imu = req->imu;
2868 u16 index, buf_index = req->buf_index;
2871 if (unlikely(buf_index >= ctx->nr_user_bufs))
2873 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2874 imu = READ_ONCE(ctx->user_bufs[index]);
2877 return __io_import_fixed(req, rw, iter, imu);
2880 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2883 mutex_unlock(&ctx->uring_lock);
2886 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2889 * "Normal" inline submissions always hold the uring_lock, since we
2890 * grab it from the system call. Same is true for the SQPOLL offload.
2891 * The only exception is when we've detached the request and issue it
2892 * from an async worker thread, grab the lock for that case.
2895 mutex_lock(&ctx->uring_lock);
2898 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2899 int bgid, struct io_buffer *kbuf,
2902 struct io_buffer *head;
2904 if (req->flags & REQ_F_BUFFER_SELECTED)
2907 io_ring_submit_lock(req->ctx, needs_lock);
2909 lockdep_assert_held(&req->ctx->uring_lock);
2911 head = xa_load(&req->ctx->io_buffers, bgid);
2913 if (!list_empty(&head->list)) {
2914 kbuf = list_last_entry(&head->list, struct io_buffer,
2916 list_del(&kbuf->list);
2919 xa_erase(&req->ctx->io_buffers, bgid);
2921 if (*len > kbuf->len)
2924 kbuf = ERR_PTR(-ENOBUFS);
2927 io_ring_submit_unlock(req->ctx, needs_lock);
2932 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2935 struct io_buffer *kbuf;
2938 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2939 bgid = req->buf_index;
2940 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2943 req->rw.addr = (u64) (unsigned long) kbuf;
2944 req->flags |= REQ_F_BUFFER_SELECTED;
2945 return u64_to_user_ptr(kbuf->addr);
2948 #ifdef CONFIG_COMPAT
2949 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2952 struct compat_iovec __user *uiov;
2953 compat_ssize_t clen;
2957 uiov = u64_to_user_ptr(req->rw.addr);
2958 if (!access_ok(uiov, sizeof(*uiov)))
2960 if (__get_user(clen, &uiov->iov_len))
2966 buf = io_rw_buffer_select(req, &len, needs_lock);
2968 return PTR_ERR(buf);
2969 iov[0].iov_base = buf;
2970 iov[0].iov_len = (compat_size_t) len;
2975 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2978 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2982 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2985 len = iov[0].iov_len;
2988 buf = io_rw_buffer_select(req, &len, needs_lock);
2990 return PTR_ERR(buf);
2991 iov[0].iov_base = buf;
2992 iov[0].iov_len = len;
2996 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2999 if (req->flags & REQ_F_BUFFER_SELECTED) {
3000 struct io_buffer *kbuf;
3002 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3003 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
3004 iov[0].iov_len = kbuf->len;
3007 if (req->rw.len != 1)
3010 #ifdef CONFIG_COMPAT
3011 if (req->ctx->compat)
3012 return io_compat_import(req, iov, needs_lock);
3015 return __io_iov_buffer_select(req, iov, needs_lock);
3018 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
3019 struct iov_iter *iter, bool needs_lock)
3021 void __user *buf = u64_to_user_ptr(req->rw.addr);
3022 size_t sqe_len = req->rw.len;
3023 u8 opcode = req->opcode;
3026 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3028 return io_import_fixed(req, rw, iter);
3031 /* buffer index only valid with fixed read/write, or buffer select */
3032 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3035 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3036 if (req->flags & REQ_F_BUFFER_SELECT) {
3037 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3039 return PTR_ERR(buf);
3040 req->rw.len = sqe_len;
3043 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3048 if (req->flags & REQ_F_BUFFER_SELECT) {
3049 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3051 iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
3056 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3060 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3062 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3066 * For files that don't have ->read_iter() and ->write_iter(), handle them
3067 * by looping over ->read() or ->write() manually.
3069 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3071 struct kiocb *kiocb = &req->rw.kiocb;
3072 struct file *file = req->file;
3076 * Don't support polled IO through this interface, and we can't
3077 * support non-blocking either. For the latter, this just causes
3078 * the kiocb to be handled from an async context.
3080 if (kiocb->ki_flags & IOCB_HIPRI)
3082 if (kiocb->ki_flags & IOCB_NOWAIT)
3085 while (iov_iter_count(iter)) {
3089 if (!iov_iter_is_bvec(iter)) {
3090 iovec = iov_iter_iovec(iter);
3092 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3093 iovec.iov_len = req->rw.len;
3097 nr = file->f_op->read(file, iovec.iov_base,
3098 iovec.iov_len, io_kiocb_ppos(kiocb));
3100 nr = file->f_op->write(file, iovec.iov_base,
3101 iovec.iov_len, io_kiocb_ppos(kiocb));
3110 if (nr != iovec.iov_len)
3114 iov_iter_advance(iter, nr);
3120 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3121 const struct iovec *fast_iov, struct iov_iter *iter)
3123 struct io_async_rw *rw = req->async_data;
3125 memcpy(&rw->iter, iter, sizeof(*iter));
3126 rw->free_iovec = iovec;
3128 /* can only be fixed buffers, no need to do anything */
3129 if (iov_iter_is_bvec(iter))
3132 unsigned iov_off = 0;
3134 rw->iter.iov = rw->fast_iov;
3135 if (iter->iov != fast_iov) {
3136 iov_off = iter->iov - fast_iov;
3137 rw->iter.iov += iov_off;
3139 if (rw->fast_iov != fast_iov)
3140 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3141 sizeof(struct iovec) * iter->nr_segs);
3143 req->flags |= REQ_F_NEED_CLEANUP;
3147 static inline int io_alloc_async_data(struct io_kiocb *req)
3149 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3150 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3151 return req->async_data == NULL;
3154 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3155 const struct iovec *fast_iov,
3156 struct iov_iter *iter, bool force)
3158 if (!force && !io_op_defs[req->opcode].needs_async_setup)
3160 if (!req->async_data) {
3161 if (io_alloc_async_data(req)) {
3166 io_req_map_rw(req, iovec, fast_iov, iter);
3171 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3173 struct io_async_rw *iorw = req->async_data;
3174 struct iovec *iov = iorw->fast_iov;
3177 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3178 if (unlikely(ret < 0))
3181 iorw->bytes_done = 0;
3182 iorw->free_iovec = iov;
3184 req->flags |= REQ_F_NEED_CLEANUP;
3188 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3190 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3192 return io_prep_rw(req, sqe);
3196 * This is our waitqueue callback handler, registered through lock_page_async()
3197 * when we initially tried to do the IO with the iocb armed our waitqueue.
3198 * This gets called when the page is unlocked, and we generally expect that to
3199 * happen when the page IO is completed and the page is now uptodate. This will
3200 * queue a task_work based retry of the operation, attempting to copy the data
3201 * again. If the latter fails because the page was NOT uptodate, then we will
3202 * do a thread based blocking retry of the operation. That's the unexpected
3205 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3206 int sync, void *arg)
3208 struct wait_page_queue *wpq;
3209 struct io_kiocb *req = wait->private;
3210 struct wait_page_key *key = arg;
3212 wpq = container_of(wait, struct wait_page_queue, wait);
3214 if (!wake_page_match(wpq, key))
3217 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3218 list_del_init(&wait->entry);
3220 /* submit ref gets dropped, acquire a new one */
3222 io_req_task_queue(req);
3227 * This controls whether a given IO request should be armed for async page
3228 * based retry. If we return false here, the request is handed to the async
3229 * worker threads for retry. If we're doing buffered reads on a regular file,
3230 * we prepare a private wait_page_queue entry and retry the operation. This
3231 * will either succeed because the page is now uptodate and unlocked, or it
3232 * will register a callback when the page is unlocked at IO completion. Through
3233 * that callback, io_uring uses task_work to setup a retry of the operation.
3234 * That retry will attempt the buffered read again. The retry will generally
3235 * succeed, or in rare cases where it fails, we then fall back to using the
3236 * async worker threads for a blocking retry.
3238 static bool io_rw_should_retry(struct io_kiocb *req)
3240 struct io_async_rw *rw = req->async_data;
3241 struct wait_page_queue *wait = &rw->wpq;
3242 struct kiocb *kiocb = &req->rw.kiocb;
3244 /* never retry for NOWAIT, we just complete with -EAGAIN */
3245 if (req->flags & REQ_F_NOWAIT)
3248 /* Only for buffered IO */
3249 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3253 * just use poll if we can, and don't attempt if the fs doesn't
3254 * support callback based unlocks
3256 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3259 wait->wait.func = io_async_buf_func;
3260 wait->wait.private = req;
3261 wait->wait.flags = 0;
3262 INIT_LIST_HEAD(&wait->wait.entry);
3263 kiocb->ki_flags |= IOCB_WAITQ;
3264 kiocb->ki_flags &= ~IOCB_NOWAIT;
3265 kiocb->ki_waitq = wait;
3269 static inline int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3271 if (req->file->f_op->read_iter)
3272 return call_read_iter(req->file, &req->rw.kiocb, iter);
3273 else if (req->file->f_op->read)
3274 return loop_rw_iter(READ, req, iter);
3279 static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3281 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3282 struct kiocb *kiocb = &req->rw.kiocb;
3283 struct iov_iter __iter, *iter = &__iter;
3284 struct io_async_rw *rw = req->async_data;
3285 ssize_t io_size, ret, ret2;
3286 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3292 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3296 io_size = iov_iter_count(iter);
3297 req->result = io_size;
3299 /* Ensure we clear previously set non-block flag */
3300 if (!force_nonblock)
3301 kiocb->ki_flags &= ~IOCB_NOWAIT;
3303 kiocb->ki_flags |= IOCB_NOWAIT;
3305 /* If the file doesn't support async, just async punt */
3306 if (force_nonblock && !io_file_supports_async(req, READ)) {
3307 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3308 return ret ?: -EAGAIN;
3311 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size);
3312 if (unlikely(ret)) {
3317 ret = io_iter_do_read(req, iter);
3319 if (ret == -EAGAIN || (req->flags & REQ_F_REISSUE)) {
3320 req->flags &= ~REQ_F_REISSUE;
3321 /* IOPOLL retry should happen for io-wq threads */
3322 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3324 /* no retry on NONBLOCK nor RWF_NOWAIT */
3325 if (req->flags & REQ_F_NOWAIT)
3327 /* some cases will consume bytes even on error returns */
3328 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3330 } else if (ret == -EIOCBQUEUED) {
3332 } else if (ret <= 0 || ret == io_size || !force_nonblock ||
3333 (req->flags & REQ_F_NOWAIT) || !(req->flags & REQ_F_ISREG)) {
3334 /* read all, failed, already did sync or don't want to retry */
3338 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3343 rw = req->async_data;
3344 /* now use our persistent iterator, if we aren't already */
3349 rw->bytes_done += ret;
3350 /* if we can retry, do so with the callbacks armed */
3351 if (!io_rw_should_retry(req)) {
3352 kiocb->ki_flags &= ~IOCB_WAITQ;
3357 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3358 * we get -EIOCBQUEUED, then we'll get a notification when the
3359 * desired page gets unlocked. We can also get a partial read
3360 * here, and if we do, then just retry at the new offset.
3362 ret = io_iter_do_read(req, iter);
3363 if (ret == -EIOCBQUEUED)
3365 /* we got some bytes, but not all. retry. */
3366 kiocb->ki_flags &= ~IOCB_WAITQ;
3367 } while (ret > 0 && ret < io_size);
3369 kiocb_done(kiocb, ret, issue_flags);
3371 /* it's faster to check here then delegate to kfree */
3377 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3379 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3381 return io_prep_rw(req, sqe);
3384 static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3386 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3387 struct kiocb *kiocb = &req->rw.kiocb;
3388 struct iov_iter __iter, *iter = &__iter;
3389 struct io_async_rw *rw = req->async_data;
3390 ssize_t ret, ret2, io_size;
3391 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3397 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3401 io_size = iov_iter_count(iter);
3402 req->result = io_size;
3404 /* Ensure we clear previously set non-block flag */
3405 if (!force_nonblock)
3406 kiocb->ki_flags &= ~IOCB_NOWAIT;
3408 kiocb->ki_flags |= IOCB_NOWAIT;
3410 /* If the file doesn't support async, just async punt */
3411 if (force_nonblock && !io_file_supports_async(req, WRITE))
3414 /* file path doesn't support NOWAIT for non-direct_IO */
3415 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3416 (req->flags & REQ_F_ISREG))
3419 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size);
3424 * Open-code file_start_write here to grab freeze protection,
3425 * which will be released by another thread in
3426 * io_complete_rw(). Fool lockdep by telling it the lock got
3427 * released so that it doesn't complain about the held lock when
3428 * we return to userspace.
3430 if (req->flags & REQ_F_ISREG) {
3431 sb_start_write(file_inode(req->file)->i_sb);
3432 __sb_writers_release(file_inode(req->file)->i_sb,
3435 kiocb->ki_flags |= IOCB_WRITE;
3437 if (req->file->f_op->write_iter)
3438 ret2 = call_write_iter(req->file, kiocb, iter);
3439 else if (req->file->f_op->write)
3440 ret2 = loop_rw_iter(WRITE, req, iter);
3444 if (req->flags & REQ_F_REISSUE) {
3445 req->flags &= ~REQ_F_REISSUE;
3450 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3451 * retry them without IOCB_NOWAIT.
3453 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3455 /* no retry on NONBLOCK nor RWF_NOWAIT */
3456 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3458 if (!force_nonblock || ret2 != -EAGAIN) {
3459 /* IOPOLL retry should happen for io-wq threads */
3460 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3463 kiocb_done(kiocb, ret2, issue_flags);
3466 /* some cases will consume bytes even on error returns */
3467 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3468 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3469 return ret ?: -EAGAIN;
3472 /* it's reportedly faster than delegating the null check to kfree() */
3478 static int io_renameat_prep(struct io_kiocb *req,
3479 const struct io_uring_sqe *sqe)
3481 struct io_rename *ren = &req->rename;
3482 const char __user *oldf, *newf;
3484 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3486 if (sqe->ioprio || sqe->buf_index)
3488 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3491 ren->old_dfd = READ_ONCE(sqe->fd);
3492 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3493 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3494 ren->new_dfd = READ_ONCE(sqe->len);
3495 ren->flags = READ_ONCE(sqe->rename_flags);
3497 ren->oldpath = getname(oldf);
3498 if (IS_ERR(ren->oldpath))
3499 return PTR_ERR(ren->oldpath);
3501 ren->newpath = getname(newf);
3502 if (IS_ERR(ren->newpath)) {
3503 putname(ren->oldpath);
3504 return PTR_ERR(ren->newpath);
3507 req->flags |= REQ_F_NEED_CLEANUP;
3511 static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3513 struct io_rename *ren = &req->rename;
3516 if (issue_flags & IO_URING_F_NONBLOCK)
3519 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3520 ren->newpath, ren->flags);
3522 req->flags &= ~REQ_F_NEED_CLEANUP;
3525 io_req_complete(req, ret);
3529 static int io_unlinkat_prep(struct io_kiocb *req,
3530 const struct io_uring_sqe *sqe)
3532 struct io_unlink *un = &req->unlink;
3533 const char __user *fname;
3535 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3537 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
3539 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3542 un->dfd = READ_ONCE(sqe->fd);
3544 un->flags = READ_ONCE(sqe->unlink_flags);
3545 if (un->flags & ~AT_REMOVEDIR)
3548 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3549 un->filename = getname(fname);
3550 if (IS_ERR(un->filename))
3551 return PTR_ERR(un->filename);
3553 req->flags |= REQ_F_NEED_CLEANUP;
3557 static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3559 struct io_unlink *un = &req->unlink;
3562 if (issue_flags & IO_URING_F_NONBLOCK)
3565 if (un->flags & AT_REMOVEDIR)
3566 ret = do_rmdir(un->dfd, un->filename);
3568 ret = do_unlinkat(un->dfd, un->filename);
3570 req->flags &= ~REQ_F_NEED_CLEANUP;
3573 io_req_complete(req, ret);
3577 static int io_shutdown_prep(struct io_kiocb *req,
3578 const struct io_uring_sqe *sqe)
3580 #if defined(CONFIG_NET)
3581 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3583 if (sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3587 req->shutdown.how = READ_ONCE(sqe->len);
3594 static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3596 #if defined(CONFIG_NET)
3597 struct socket *sock;
3600 if (issue_flags & IO_URING_F_NONBLOCK)
3603 sock = sock_from_file(req->file);
3604 if (unlikely(!sock))
3607 ret = __sys_shutdown_sock(sock, req->shutdown.how);
3610 io_req_complete(req, ret);
3617 static int __io_splice_prep(struct io_kiocb *req,
3618 const struct io_uring_sqe *sqe)
3620 struct io_splice *sp = &req->splice;
3621 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3623 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3627 sp->len = READ_ONCE(sqe->len);
3628 sp->flags = READ_ONCE(sqe->splice_flags);
3630 if (unlikely(sp->flags & ~valid_flags))
3633 sp->file_in = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in),
3634 (sp->flags & SPLICE_F_FD_IN_FIXED));
3637 req->flags |= REQ_F_NEED_CLEANUP;
3641 static int io_tee_prep(struct io_kiocb *req,
3642 const struct io_uring_sqe *sqe)
3644 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3646 return __io_splice_prep(req, sqe);
3649 static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
3651 struct io_splice *sp = &req->splice;
3652 struct file *in = sp->file_in;
3653 struct file *out = sp->file_out;
3654 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3657 if (issue_flags & IO_URING_F_NONBLOCK)
3660 ret = do_tee(in, out, sp->len, flags);
3662 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3664 req->flags &= ~REQ_F_NEED_CLEANUP;
3668 io_req_complete(req, ret);
3672 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3674 struct io_splice *sp = &req->splice;
3676 sp->off_in = READ_ONCE(sqe->splice_off_in);
3677 sp->off_out = READ_ONCE(sqe->off);
3678 return __io_splice_prep(req, sqe);
3681 static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
3683 struct io_splice *sp = &req->splice;
3684 struct file *in = sp->file_in;
3685 struct file *out = sp->file_out;
3686 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3687 loff_t *poff_in, *poff_out;
3690 if (issue_flags & IO_URING_F_NONBLOCK)
3693 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3694 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3697 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3699 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3701 req->flags &= ~REQ_F_NEED_CLEANUP;
3705 io_req_complete(req, ret);
3710 * IORING_OP_NOP just posts a completion event, nothing else.
3712 static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
3714 struct io_ring_ctx *ctx = req->ctx;
3716 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3719 __io_req_complete(req, issue_flags, 0, 0);
3723 static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3725 struct io_ring_ctx *ctx = req->ctx;
3730 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3732 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3735 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3736 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3739 req->sync.off = READ_ONCE(sqe->off);
3740 req->sync.len = READ_ONCE(sqe->len);
3744 static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
3746 loff_t end = req->sync.off + req->sync.len;
3749 /* fsync always requires a blocking context */
3750 if (issue_flags & IO_URING_F_NONBLOCK)
3753 ret = vfs_fsync_range(req->file, req->sync.off,
3754 end > 0 ? end : LLONG_MAX,
3755 req->sync.flags & IORING_FSYNC_DATASYNC);
3758 io_req_complete(req, ret);
3762 static int io_fallocate_prep(struct io_kiocb *req,
3763 const struct io_uring_sqe *sqe)
3765 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3767 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3770 req->sync.off = READ_ONCE(sqe->off);
3771 req->sync.len = READ_ONCE(sqe->addr);
3772 req->sync.mode = READ_ONCE(sqe->len);
3776 static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
3780 /* fallocate always requiring blocking context */
3781 if (issue_flags & IO_URING_F_NONBLOCK)
3783 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3787 io_req_complete(req, ret);
3791 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3793 const char __user *fname;
3796 if (unlikely(sqe->ioprio || sqe->buf_index))
3798 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3801 /* open.how should be already initialised */
3802 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3803 req->open.how.flags |= O_LARGEFILE;
3805 req->open.dfd = READ_ONCE(sqe->fd);
3806 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3807 req->open.filename = getname(fname);
3808 if (IS_ERR(req->open.filename)) {
3809 ret = PTR_ERR(req->open.filename);
3810 req->open.filename = NULL;
3813 req->open.nofile = rlimit(RLIMIT_NOFILE);
3814 req->flags |= REQ_F_NEED_CLEANUP;
3818 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3822 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3824 mode = READ_ONCE(sqe->len);
3825 flags = READ_ONCE(sqe->open_flags);
3826 req->open.how = build_open_how(flags, mode);
3827 return __io_openat_prep(req, sqe);
3830 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3832 struct open_how __user *how;
3836 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3838 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3839 len = READ_ONCE(sqe->len);
3840 if (len < OPEN_HOW_SIZE_VER0)
3843 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3848 return __io_openat_prep(req, sqe);
3851 static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
3853 struct open_flags op;
3856 bool resolve_nonblock;
3859 ret = build_open_flags(&req->open.how, &op);
3862 nonblock_set = op.open_flag & O_NONBLOCK;
3863 resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
3864 if (issue_flags & IO_URING_F_NONBLOCK) {
3866 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
3867 * it'll always -EAGAIN
3869 if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
3871 op.lookup_flags |= LOOKUP_CACHED;
3872 op.open_flag |= O_NONBLOCK;
3875 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3879 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3882 * We could hang on to this 'fd' on retrying, but seems like
3883 * marginal gain for something that is now known to be a slower
3884 * path. So just put it, and we'll get a new one when we retry.
3888 ret = PTR_ERR(file);
3889 /* only retry if RESOLVE_CACHED wasn't already set by application */
3890 if (ret == -EAGAIN &&
3891 (!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)))
3896 if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
3897 file->f_flags &= ~O_NONBLOCK;
3898 fsnotify_open(file);
3899 fd_install(ret, file);
3901 putname(req->open.filename);
3902 req->flags &= ~REQ_F_NEED_CLEANUP;
3905 __io_req_complete(req, issue_flags, ret, 0);
3909 static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
3911 return io_openat2(req, issue_flags);
3914 static int io_remove_buffers_prep(struct io_kiocb *req,
3915 const struct io_uring_sqe *sqe)
3917 struct io_provide_buf *p = &req->pbuf;
3920 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3923 tmp = READ_ONCE(sqe->fd);
3924 if (!tmp || tmp > USHRT_MAX)
3927 memset(p, 0, sizeof(*p));
3929 p->bgid = READ_ONCE(sqe->buf_group);
3933 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3934 int bgid, unsigned nbufs)
3938 /* shouldn't happen */
3942 /* the head kbuf is the list itself */
3943 while (!list_empty(&buf->list)) {
3944 struct io_buffer *nxt;
3946 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3947 list_del(&nxt->list);
3954 xa_erase(&ctx->io_buffers, bgid);
3959 static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
3961 struct io_provide_buf *p = &req->pbuf;
3962 struct io_ring_ctx *ctx = req->ctx;
3963 struct io_buffer *head;
3965 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3967 io_ring_submit_lock(ctx, !force_nonblock);
3969 lockdep_assert_held(&ctx->uring_lock);
3972 head = xa_load(&ctx->io_buffers, p->bgid);
3974 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3978 /* complete before unlock, IOPOLL may need the lock */
3979 __io_req_complete(req, issue_flags, ret, 0);
3980 io_ring_submit_unlock(ctx, !force_nonblock);
3984 static int io_provide_buffers_prep(struct io_kiocb *req,
3985 const struct io_uring_sqe *sqe)
3987 unsigned long size, tmp_check;
3988 struct io_provide_buf *p = &req->pbuf;
3991 if (sqe->ioprio || sqe->rw_flags)
3994 tmp = READ_ONCE(sqe->fd);
3995 if (!tmp || tmp > USHRT_MAX)
3998 p->addr = READ_ONCE(sqe->addr);
3999 p->len = READ_ONCE(sqe->len);
4001 if (check_mul_overflow((unsigned long)p->len, (unsigned long)p->nbufs,
4004 if (check_add_overflow((unsigned long)p->addr, size, &tmp_check))
4007 size = (unsigned long)p->len * p->nbufs;
4008 if (!access_ok(u64_to_user_ptr(p->addr), size))
4011 p->bgid = READ_ONCE(sqe->buf_group);
4012 tmp = READ_ONCE(sqe->off);
4013 if (tmp > USHRT_MAX)
4019 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
4021 struct io_buffer *buf;
4022 u64 addr = pbuf->addr;
4023 int i, bid = pbuf->bid;
4025 for (i = 0; i < pbuf->nbufs; i++) {
4026 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
4031 buf->len = min_t(__u32, pbuf->len, MAX_RW_COUNT);
4036 INIT_LIST_HEAD(&buf->list);
4039 list_add_tail(&buf->list, &(*head)->list);
4043 return i ? i : -ENOMEM;
4046 static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
4048 struct io_provide_buf *p = &req->pbuf;
4049 struct io_ring_ctx *ctx = req->ctx;
4050 struct io_buffer *head, *list;
4052 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4054 io_ring_submit_lock(ctx, !force_nonblock);
4056 lockdep_assert_held(&ctx->uring_lock);
4058 list = head = xa_load(&ctx->io_buffers, p->bgid);
4060 ret = io_add_buffers(p, &head);
4061 if (ret >= 0 && !list) {
4062 ret = xa_insert(&ctx->io_buffers, p->bgid, head, GFP_KERNEL);
4064 __io_remove_buffers(ctx, head, p->bgid, -1U);
4068 /* complete before unlock, IOPOLL may need the lock */
4069 __io_req_complete(req, issue_flags, ret, 0);
4070 io_ring_submit_unlock(ctx, !force_nonblock);
4074 static int io_epoll_ctl_prep(struct io_kiocb *req,
4075 const struct io_uring_sqe *sqe)
4077 #if defined(CONFIG_EPOLL)
4078 if (sqe->ioprio || sqe->buf_index)
4080 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4083 req->epoll.epfd = READ_ONCE(sqe->fd);
4084 req->epoll.op = READ_ONCE(sqe->len);
4085 req->epoll.fd = READ_ONCE(sqe->off);
4087 if (ep_op_has_event(req->epoll.op)) {
4088 struct epoll_event __user *ev;
4090 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4091 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4101 static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4103 #if defined(CONFIG_EPOLL)
4104 struct io_epoll *ie = &req->epoll;
4106 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4108 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4109 if (force_nonblock && ret == -EAGAIN)
4114 __io_req_complete(req, issue_flags, ret, 0);
4121 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4123 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4124 if (sqe->ioprio || sqe->buf_index || sqe->off)
4126 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4129 req->madvise.addr = READ_ONCE(sqe->addr);
4130 req->madvise.len = READ_ONCE(sqe->len);
4131 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4138 static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4140 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4141 struct io_madvise *ma = &req->madvise;
4144 if (issue_flags & IO_URING_F_NONBLOCK)
4147 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4150 io_req_complete(req, ret);
4157 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4159 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4161 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4164 req->fadvise.offset = READ_ONCE(sqe->off);
4165 req->fadvise.len = READ_ONCE(sqe->len);
4166 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4170 static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4172 struct io_fadvise *fa = &req->fadvise;
4175 if (issue_flags & IO_URING_F_NONBLOCK) {
4176 switch (fa->advice) {
4177 case POSIX_FADV_NORMAL:
4178 case POSIX_FADV_RANDOM:
4179 case POSIX_FADV_SEQUENTIAL:
4186 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4189 __io_req_complete(req, issue_flags, ret, 0);
4193 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4195 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4197 if (sqe->ioprio || sqe->buf_index)
4199 if (req->flags & REQ_F_FIXED_FILE)
4202 req->statx.dfd = READ_ONCE(sqe->fd);
4203 req->statx.mask = READ_ONCE(sqe->len);
4204 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4205 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4206 req->statx.flags = READ_ONCE(sqe->statx_flags);
4211 static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4213 struct io_statx *ctx = &req->statx;
4216 if (issue_flags & IO_URING_F_NONBLOCK)
4219 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4224 io_req_complete(req, ret);
4228 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4230 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4232 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4233 sqe->rw_flags || sqe->buf_index)
4235 if (req->flags & REQ_F_FIXED_FILE)
4238 req->close.fd = READ_ONCE(sqe->fd);
4242 static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4244 struct files_struct *files = current->files;
4245 struct io_close *close = &req->close;
4246 struct fdtable *fdt;
4247 struct file *file = NULL;
4250 spin_lock(&files->file_lock);
4251 fdt = files_fdtable(files);
4252 if (close->fd >= fdt->max_fds) {
4253 spin_unlock(&files->file_lock);
4256 file = fdt->fd[close->fd];
4257 if (!file || file->f_op == &io_uring_fops) {
4258 spin_unlock(&files->file_lock);
4263 /* if the file has a flush method, be safe and punt to async */
4264 if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4265 spin_unlock(&files->file_lock);
4269 ret = __close_fd_get_file(close->fd, &file);
4270 spin_unlock(&files->file_lock);
4277 /* No ->flush() or already async, safely close from here */
4278 ret = filp_close(file, current->files);
4284 __io_req_complete(req, issue_flags, ret, 0);
4288 static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4290 struct io_ring_ctx *ctx = req->ctx;
4292 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4294 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4297 req->sync.off = READ_ONCE(sqe->off);
4298 req->sync.len = READ_ONCE(sqe->len);
4299 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4303 static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4307 /* sync_file_range always requires a blocking context */
4308 if (issue_flags & IO_URING_F_NONBLOCK)
4311 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4315 io_req_complete(req, ret);
4319 #if defined(CONFIG_NET)
4320 static int io_setup_async_msg(struct io_kiocb *req,
4321 struct io_async_msghdr *kmsg)
4323 struct io_async_msghdr *async_msg = req->async_data;
4327 if (io_alloc_async_data(req)) {
4328 kfree(kmsg->free_iov);
4331 async_msg = req->async_data;
4332 req->flags |= REQ_F_NEED_CLEANUP;
4333 memcpy(async_msg, kmsg, sizeof(*kmsg));
4334 async_msg->msg.msg_name = &async_msg->addr;
4335 /* if were using fast_iov, set it to the new one */
4336 if (!async_msg->free_iov)
4337 async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4342 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4343 struct io_async_msghdr *iomsg)
4345 iomsg->msg.msg_name = &iomsg->addr;
4346 iomsg->free_iov = iomsg->fast_iov;
4347 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4348 req->sr_msg.msg_flags, &iomsg->free_iov);
4351 static int io_sendmsg_prep_async(struct io_kiocb *req)
4355 ret = io_sendmsg_copy_hdr(req, req->async_data);
4357 req->flags |= REQ_F_NEED_CLEANUP;
4361 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4363 struct io_sr_msg *sr = &req->sr_msg;
4365 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4368 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4369 sr->len = READ_ONCE(sqe->len);
4370 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4371 if (sr->msg_flags & MSG_DONTWAIT)
4372 req->flags |= REQ_F_NOWAIT;
4374 #ifdef CONFIG_COMPAT
4375 if (req->ctx->compat)
4376 sr->msg_flags |= MSG_CMSG_COMPAT;
4381 static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4383 struct io_async_msghdr iomsg, *kmsg;
4384 struct socket *sock;
4389 sock = sock_from_file(req->file);
4390 if (unlikely(!sock))
4393 kmsg = req->async_data;
4395 ret = io_sendmsg_copy_hdr(req, &iomsg);
4401 flags = req->sr_msg.msg_flags;
4402 if (issue_flags & IO_URING_F_NONBLOCK)
4403 flags |= MSG_DONTWAIT;
4404 if (flags & MSG_WAITALL)
4405 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4407 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4408 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4409 return io_setup_async_msg(req, kmsg);
4410 if (ret == -ERESTARTSYS)
4413 /* fast path, check for non-NULL to avoid function call */
4415 kfree(kmsg->free_iov);
4416 req->flags &= ~REQ_F_NEED_CLEANUP;
4419 __io_req_complete(req, issue_flags, ret, 0);
4423 static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4425 struct io_sr_msg *sr = &req->sr_msg;
4428 struct socket *sock;
4433 sock = sock_from_file(req->file);
4434 if (unlikely(!sock))
4437 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4441 msg.msg_name = NULL;
4442 msg.msg_control = NULL;
4443 msg.msg_controllen = 0;
4444 msg.msg_namelen = 0;
4446 flags = req->sr_msg.msg_flags;
4447 if (issue_flags & IO_URING_F_NONBLOCK)
4448 flags |= MSG_DONTWAIT;
4449 if (flags & MSG_WAITALL)
4450 min_ret = iov_iter_count(&msg.msg_iter);
4452 msg.msg_flags = flags;
4453 ret = sock_sendmsg(sock, &msg);
4454 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4456 if (ret == -ERESTARTSYS)
4461 __io_req_complete(req, issue_flags, ret, 0);
4465 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4466 struct io_async_msghdr *iomsg)
4468 struct io_sr_msg *sr = &req->sr_msg;
4469 struct iovec __user *uiov;
4473 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4474 &iomsg->uaddr, &uiov, &iov_len);
4478 if (req->flags & REQ_F_BUFFER_SELECT) {
4481 if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4483 sr->len = iomsg->fast_iov[0].iov_len;
4484 iomsg->free_iov = NULL;
4486 iomsg->free_iov = iomsg->fast_iov;
4487 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4488 &iomsg->free_iov, &iomsg->msg.msg_iter,
4497 #ifdef CONFIG_COMPAT
4498 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4499 struct io_async_msghdr *iomsg)
4501 struct io_sr_msg *sr = &req->sr_msg;
4502 struct compat_iovec __user *uiov;
4507 ret = __get_compat_msghdr(&iomsg->msg, sr->umsg_compat, &iomsg->uaddr,
4512 uiov = compat_ptr(ptr);
4513 if (req->flags & REQ_F_BUFFER_SELECT) {
4514 compat_ssize_t clen;
4518 if (!access_ok(uiov, sizeof(*uiov)))
4520 if (__get_user(clen, &uiov->iov_len))
4525 iomsg->free_iov = NULL;
4527 iomsg->free_iov = iomsg->fast_iov;
4528 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4529 UIO_FASTIOV, &iomsg->free_iov,
4530 &iomsg->msg.msg_iter, true);
4539 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4540 struct io_async_msghdr *iomsg)
4542 iomsg->msg.msg_name = &iomsg->addr;
4544 #ifdef CONFIG_COMPAT
4545 if (req->ctx->compat)
4546 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4549 return __io_recvmsg_copy_hdr(req, iomsg);
4552 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4555 struct io_sr_msg *sr = &req->sr_msg;
4556 struct io_buffer *kbuf;
4558 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4563 req->flags |= REQ_F_BUFFER_SELECTED;
4567 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4569 return io_put_kbuf(req, req->sr_msg.kbuf);
4572 static int io_recvmsg_prep_async(struct io_kiocb *req)
4576 ret = io_recvmsg_copy_hdr(req, req->async_data);
4578 req->flags |= REQ_F_NEED_CLEANUP;
4582 static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4584 struct io_sr_msg *sr = &req->sr_msg;
4586 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4589 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4590 sr->len = READ_ONCE(sqe->len);
4591 sr->bgid = READ_ONCE(sqe->buf_group);
4592 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4593 if (sr->msg_flags & MSG_DONTWAIT)
4594 req->flags |= REQ_F_NOWAIT;
4596 #ifdef CONFIG_COMPAT
4597 if (req->ctx->compat)
4598 sr->msg_flags |= MSG_CMSG_COMPAT;
4603 static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
4605 struct io_async_msghdr iomsg, *kmsg;
4606 struct socket *sock;
4607 struct io_buffer *kbuf;
4610 int ret, cflags = 0;
4611 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4613 sock = sock_from_file(req->file);
4614 if (unlikely(!sock))
4617 kmsg = req->async_data;
4619 ret = io_recvmsg_copy_hdr(req, &iomsg);
4625 if (req->flags & REQ_F_BUFFER_SELECT) {
4626 kbuf = io_recv_buffer_select(req, !force_nonblock);
4628 return PTR_ERR(kbuf);
4629 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4630 kmsg->fast_iov[0].iov_len = req->sr_msg.len;
4631 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
4632 1, req->sr_msg.len);
4635 flags = req->sr_msg.msg_flags;
4637 flags |= MSG_DONTWAIT;
4638 if (flags & MSG_WAITALL)
4639 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4641 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4642 kmsg->uaddr, flags);
4643 if (force_nonblock && ret == -EAGAIN)
4644 return io_setup_async_msg(req, kmsg);
4645 if (ret == -ERESTARTSYS)
4648 if (req->flags & REQ_F_BUFFER_SELECTED)
4649 cflags = io_put_recv_kbuf(req);
4650 /* fast path, check for non-NULL to avoid function call */
4652 kfree(kmsg->free_iov);
4653 req->flags &= ~REQ_F_NEED_CLEANUP;
4654 if (ret < min_ret || ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4656 __io_req_complete(req, issue_flags, ret, cflags);
4660 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
4662 struct io_buffer *kbuf;
4663 struct io_sr_msg *sr = &req->sr_msg;
4665 void __user *buf = sr->buf;
4666 struct socket *sock;
4670 int ret, cflags = 0;
4671 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4673 sock = sock_from_file(req->file);
4674 if (unlikely(!sock))
4677 if (req->flags & REQ_F_BUFFER_SELECT) {
4678 kbuf = io_recv_buffer_select(req, !force_nonblock);
4680 return PTR_ERR(kbuf);
4681 buf = u64_to_user_ptr(kbuf->addr);
4684 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4688 msg.msg_name = NULL;
4689 msg.msg_control = NULL;
4690 msg.msg_controllen = 0;
4691 msg.msg_namelen = 0;
4692 msg.msg_iocb = NULL;
4695 flags = req->sr_msg.msg_flags;
4697 flags |= MSG_DONTWAIT;
4698 if (flags & MSG_WAITALL)
4699 min_ret = iov_iter_count(&msg.msg_iter);
4701 ret = sock_recvmsg(sock, &msg, flags);
4702 if (force_nonblock && ret == -EAGAIN)
4704 if (ret == -ERESTARTSYS)
4707 if (req->flags & REQ_F_BUFFER_SELECTED)
4708 cflags = io_put_recv_kbuf(req);
4709 if (ret < min_ret || ((flags & MSG_WAITALL) && (msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4711 __io_req_complete(req, issue_flags, ret, cflags);
4715 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4717 struct io_accept *accept = &req->accept;
4719 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4721 if (sqe->ioprio || sqe->len || sqe->buf_index)
4724 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4725 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4726 accept->flags = READ_ONCE(sqe->accept_flags);
4727 accept->nofile = rlimit(RLIMIT_NOFILE);
4731 static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
4733 struct io_accept *accept = &req->accept;
4734 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4735 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4738 if (req->file->f_flags & O_NONBLOCK)
4739 req->flags |= REQ_F_NOWAIT;
4741 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4742 accept->addr_len, accept->flags,
4744 if (ret == -EAGAIN && force_nonblock)
4747 if (ret == -ERESTARTSYS)
4751 __io_req_complete(req, issue_flags, ret, 0);
4755 static int io_connect_prep_async(struct io_kiocb *req)
4757 struct io_async_connect *io = req->async_data;
4758 struct io_connect *conn = &req->connect;
4760 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
4763 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4765 struct io_connect *conn = &req->connect;
4767 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4769 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4772 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4773 conn->addr_len = READ_ONCE(sqe->addr2);
4777 static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
4779 struct io_async_connect __io, *io;
4780 unsigned file_flags;
4782 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4784 if (req->async_data) {
4785 io = req->async_data;
4787 ret = move_addr_to_kernel(req->connect.addr,
4788 req->connect.addr_len,
4795 file_flags = force_nonblock ? O_NONBLOCK : 0;
4797 ret = __sys_connect_file(req->file, &io->address,
4798 req->connect.addr_len, file_flags);
4799 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4800 if (req->async_data)
4802 if (io_alloc_async_data(req)) {
4806 memcpy(req->async_data, &__io, sizeof(__io));
4809 if (ret == -ERESTARTSYS)
4814 __io_req_complete(req, issue_flags, ret, 0);
4817 #else /* !CONFIG_NET */
4818 #define IO_NETOP_FN(op) \
4819 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
4821 return -EOPNOTSUPP; \
4824 #define IO_NETOP_PREP(op) \
4826 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
4828 return -EOPNOTSUPP; \
4831 #define IO_NETOP_PREP_ASYNC(op) \
4833 static int io_##op##_prep_async(struct io_kiocb *req) \
4835 return -EOPNOTSUPP; \
4838 IO_NETOP_PREP_ASYNC(sendmsg);
4839 IO_NETOP_PREP_ASYNC(recvmsg);
4840 IO_NETOP_PREP_ASYNC(connect);
4841 IO_NETOP_PREP(accept);
4844 #endif /* CONFIG_NET */
4846 struct io_poll_table {
4847 struct poll_table_struct pt;
4848 struct io_kiocb *req;
4852 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4853 __poll_t mask, task_work_func_t func)
4857 /* for instances that support it check for an event match first: */
4858 if (mask && !(mask & poll->events))
4861 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4863 list_del_init(&poll->wait.entry);
4866 req->task_work.func = func;
4869 * If this fails, then the task is exiting. When a task exits, the
4870 * work gets canceled, so just cancel this request as well instead
4871 * of executing it. We can't safely execute it anyway, as we may not
4872 * have the needed state needed for it anyway.
4874 ret = io_req_task_work_add(req);
4875 if (unlikely(ret)) {
4876 WRITE_ONCE(poll->canceled, true);
4877 io_req_task_work_add_fallback(req, func);
4882 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4883 __acquires(&req->ctx->completion_lock)
4885 struct io_ring_ctx *ctx = req->ctx;
4887 if (!req->result && !READ_ONCE(poll->canceled)) {
4888 struct poll_table_struct pt = { ._key = poll->events };
4890 req->result = vfs_poll(req->file, &pt) & poll->events;
4893 spin_lock_irq(&ctx->completion_lock);
4894 if (!req->result && !READ_ONCE(poll->canceled)) {
4895 add_wait_queue(poll->head, &poll->wait);
4902 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4904 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4905 if (req->opcode == IORING_OP_POLL_ADD)
4906 return req->async_data;
4907 return req->apoll->double_poll;
4910 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4912 if (req->opcode == IORING_OP_POLL_ADD)
4914 return &req->apoll->poll;
4917 static void io_poll_remove_double(struct io_kiocb *req)
4918 __must_hold(&req->ctx->completion_lock)
4920 struct io_poll_iocb *poll = io_poll_get_double(req);
4922 lockdep_assert_held(&req->ctx->completion_lock);
4924 if (poll && poll->head) {
4925 struct wait_queue_head *head = poll->head;
4927 spin_lock(&head->lock);
4928 list_del_init(&poll->wait.entry);
4929 if (poll->wait.private)
4932 spin_unlock(&head->lock);
4936 static bool io_poll_complete(struct io_kiocb *req, __poll_t mask)
4937 __must_hold(&req->ctx->completion_lock)
4939 struct io_ring_ctx *ctx = req->ctx;
4940 unsigned flags = IORING_CQE_F_MORE;
4943 if (READ_ONCE(req->poll.canceled)) {
4945 req->poll.events |= EPOLLONESHOT;
4947 error = mangle_poll(mask);
4949 if (req->poll.events & EPOLLONESHOT)
4951 if (!io_cqring_fill_event(ctx, req->user_data, error, flags)) {
4952 io_poll_remove_waitqs(req);
4953 req->poll.done = true;
4956 if (flags & IORING_CQE_F_MORE)
4959 io_commit_cqring(ctx);
4960 return !(flags & IORING_CQE_F_MORE);
4963 static void io_poll_task_func(struct callback_head *cb)
4965 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4966 struct io_ring_ctx *ctx = req->ctx;
4967 struct io_kiocb *nxt;
4969 if (io_poll_rewait(req, &req->poll)) {
4970 spin_unlock_irq(&ctx->completion_lock);
4974 done = io_poll_complete(req, req->result);
4976 hash_del(&req->hash_node);
4979 add_wait_queue(req->poll.head, &req->poll.wait);
4981 spin_unlock_irq(&ctx->completion_lock);
4982 io_cqring_ev_posted(ctx);
4985 nxt = io_put_req_find_next(req);
4987 __io_req_task_submit(nxt);
4992 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4993 int sync, void *key)
4995 struct io_kiocb *req = wait->private;
4996 struct io_poll_iocb *poll = io_poll_get_single(req);
4997 __poll_t mask = key_to_poll(key);
4999 /* for instances that support it check for an event match first: */
5000 if (mask && !(mask & poll->events))
5002 if (!(poll->events & EPOLLONESHOT))
5003 return poll->wait.func(&poll->wait, mode, sync, key);
5005 list_del_init(&wait->entry);
5007 if (poll && poll->head) {
5010 spin_lock(&poll->head->lock);
5011 done = list_empty(&poll->wait.entry);
5013 list_del_init(&poll->wait.entry);
5014 /* make sure double remove sees this as being gone */
5015 wait->private = NULL;
5016 spin_unlock(&poll->head->lock);
5018 /* use wait func handler, so it matches the rq type */
5019 poll->wait.func(&poll->wait, mode, sync, key);
5026 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
5027 wait_queue_func_t wake_func)
5031 poll->canceled = false;
5032 #define IO_POLL_UNMASK (EPOLLERR|EPOLLHUP|EPOLLNVAL|EPOLLRDHUP)
5033 /* mask in events that we always want/need */
5034 poll->events = events | IO_POLL_UNMASK;
5035 INIT_LIST_HEAD(&poll->wait.entry);
5036 init_waitqueue_func_entry(&poll->wait, wake_func);
5039 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5040 struct wait_queue_head *head,
5041 struct io_poll_iocb **poll_ptr)
5043 struct io_kiocb *req = pt->req;
5046 * If poll->head is already set, it's because the file being polled
5047 * uses multiple waitqueues for poll handling (eg one for read, one
5048 * for write). Setup a separate io_poll_iocb if this happens.
5050 if (unlikely(poll->head)) {
5051 struct io_poll_iocb *poll_one = poll;
5053 /* already have a 2nd entry, fail a third attempt */
5055 pt->error = -EINVAL;
5059 * Can't handle multishot for double wait for now, turn it
5060 * into one-shot mode.
5062 if (!(poll_one->events & EPOLLONESHOT))
5063 poll_one->events |= EPOLLONESHOT;
5064 /* double add on the same waitqueue head, ignore */
5065 if (poll_one->head == head)
5067 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5069 pt->error = -ENOMEM;
5072 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5074 poll->wait.private = req;
5081 if (poll->events & EPOLLEXCLUSIVE)
5082 add_wait_queue_exclusive(head, &poll->wait);
5084 add_wait_queue(head, &poll->wait);
5087 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5088 struct poll_table_struct *p)
5090 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5091 struct async_poll *apoll = pt->req->apoll;
5093 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5096 static void io_async_task_func(struct callback_head *cb)
5098 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
5099 struct async_poll *apoll = req->apoll;
5100 struct io_ring_ctx *ctx = req->ctx;
5102 trace_io_uring_task_run(req->ctx, req, req->opcode, req->user_data);
5104 if (io_poll_rewait(req, &apoll->poll)) {
5105 spin_unlock_irq(&ctx->completion_lock);
5109 hash_del(&req->hash_node);
5110 io_poll_remove_double(req);
5111 spin_unlock_irq(&ctx->completion_lock);
5113 if (!READ_ONCE(apoll->poll.canceled))
5114 __io_req_task_submit(req);
5116 io_req_complete_failed(req, -ECANCELED);
5119 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5122 struct io_kiocb *req = wait->private;
5123 struct io_poll_iocb *poll = &req->apoll->poll;
5125 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5128 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5131 static void io_poll_req_insert(struct io_kiocb *req)
5133 struct io_ring_ctx *ctx = req->ctx;
5134 struct hlist_head *list;
5136 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5137 hlist_add_head(&req->hash_node, list);
5140 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5141 struct io_poll_iocb *poll,
5142 struct io_poll_table *ipt, __poll_t mask,
5143 wait_queue_func_t wake_func)
5144 __acquires(&ctx->completion_lock)
5146 struct io_ring_ctx *ctx = req->ctx;
5147 bool cancel = false;
5149 INIT_HLIST_NODE(&req->hash_node);
5150 io_init_poll_iocb(poll, mask, wake_func);
5151 poll->file = req->file;
5152 poll->wait.private = req;
5154 ipt->pt._key = mask;
5156 ipt->error = -EINVAL;
5158 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5160 spin_lock_irq(&ctx->completion_lock);
5161 if (likely(poll->head)) {
5162 spin_lock(&poll->head->lock);
5163 if (unlikely(list_empty(&poll->wait.entry))) {
5169 if ((mask && (poll->events & EPOLLONESHOT)) || ipt->error)
5170 list_del_init(&poll->wait.entry);
5172 WRITE_ONCE(poll->canceled, true);
5173 else if (!poll->done) /* actually waiting for an event */
5174 io_poll_req_insert(req);
5175 spin_unlock(&poll->head->lock);
5187 static int io_arm_poll_handler(struct io_kiocb *req)
5189 const struct io_op_def *def = &io_op_defs[req->opcode];
5190 struct io_ring_ctx *ctx = req->ctx;
5191 struct async_poll *apoll;
5192 struct io_poll_table ipt;
5193 __poll_t ret, mask = EPOLLONESHOT | POLLERR | POLLPRI;
5196 if (!req->file || !file_can_poll(req->file))
5197 return IO_APOLL_ABORTED;
5198 if (req->flags & REQ_F_POLLED)
5199 return IO_APOLL_ABORTED;
5200 if (!def->pollin && !def->pollout)
5201 return IO_APOLL_ABORTED;
5205 mask |= POLLIN | POLLRDNORM;
5207 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5208 if ((req->opcode == IORING_OP_RECVMSG) &&
5209 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5213 mask |= POLLOUT | POLLWRNORM;
5216 /* if we can't nonblock try, then no point in arming a poll handler */
5217 if (!io_file_supports_async(req, rw))
5218 return IO_APOLL_ABORTED;
5220 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5221 if (unlikely(!apoll))
5222 return IO_APOLL_ABORTED;
5223 apoll->double_poll = NULL;
5225 req->flags |= REQ_F_POLLED;
5226 ipt.pt._qproc = io_async_queue_proc;
5228 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5230 if (ret || ipt.error) {
5231 io_poll_remove_double(req);
5232 spin_unlock_irq(&ctx->completion_lock);
5234 return IO_APOLL_READY;
5235 return IO_APOLL_ABORTED;
5237 spin_unlock_irq(&ctx->completion_lock);
5238 trace_io_uring_poll_arm(ctx, req, req->opcode, req->user_data,
5239 mask, apoll->poll.events);
5243 static bool __io_poll_remove_one(struct io_kiocb *req,
5244 struct io_poll_iocb *poll, bool do_cancel)
5245 __must_hold(&req->ctx->completion_lock)
5247 bool do_complete = false;
5251 spin_lock(&poll->head->lock);
5253 WRITE_ONCE(poll->canceled, true);
5254 if (!list_empty(&poll->wait.entry)) {
5255 list_del_init(&poll->wait.entry);
5258 spin_unlock(&poll->head->lock);
5259 hash_del(&req->hash_node);
5263 static bool io_poll_remove_waitqs(struct io_kiocb *req)
5264 __must_hold(&req->ctx->completion_lock)
5268 io_poll_remove_double(req);
5269 do_complete = __io_poll_remove_one(req, io_poll_get_single(req), true);
5271 if (req->opcode != IORING_OP_POLL_ADD && do_complete) {
5272 /* non-poll requests have submit ref still */
5278 static bool io_poll_remove_one(struct io_kiocb *req)
5279 __must_hold(&req->ctx->completion_lock)
5283 do_complete = io_poll_remove_waitqs(req);
5285 io_cqring_fill_event(req->ctx, req->user_data, -ECANCELED, 0);
5286 io_commit_cqring(req->ctx);
5288 io_put_req_deferred(req, 1);
5295 * Returns true if we found and killed one or more poll requests
5297 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5300 struct hlist_node *tmp;
5301 struct io_kiocb *req;
5304 spin_lock_irq(&ctx->completion_lock);
5305 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5306 struct hlist_head *list;
5308 list = &ctx->cancel_hash[i];
5309 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5310 if (io_match_task(req, tsk, cancel_all))
5311 posted += io_poll_remove_one(req);
5314 spin_unlock_irq(&ctx->completion_lock);
5317 io_cqring_ev_posted(ctx);
5322 static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, __u64 sqe_addr,
5324 __must_hold(&ctx->completion_lock)
5326 struct hlist_head *list;
5327 struct io_kiocb *req;
5329 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5330 hlist_for_each_entry(req, list, hash_node) {
5331 if (sqe_addr != req->user_data)
5333 if (poll_only && req->opcode != IORING_OP_POLL_ADD)
5340 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr,
5342 __must_hold(&ctx->completion_lock)
5344 struct io_kiocb *req;
5346 req = io_poll_find(ctx, sqe_addr, poll_only);
5349 if (io_poll_remove_one(req))
5355 static __poll_t io_poll_parse_events(const struct io_uring_sqe *sqe,
5360 events = READ_ONCE(sqe->poll32_events);
5362 events = swahw32(events);
5364 if (!(flags & IORING_POLL_ADD_MULTI))
5365 events |= EPOLLONESHOT;
5366 return demangle_poll(events) | (events & (EPOLLEXCLUSIVE|EPOLLONESHOT));
5369 static int io_poll_update_prep(struct io_kiocb *req,
5370 const struct io_uring_sqe *sqe)
5372 struct io_poll_update *upd = &req->poll_update;
5375 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5377 if (sqe->ioprio || sqe->buf_index)
5379 flags = READ_ONCE(sqe->len);
5380 if (flags & ~(IORING_POLL_UPDATE_EVENTS | IORING_POLL_UPDATE_USER_DATA |
5381 IORING_POLL_ADD_MULTI))
5383 /* meaningless without update */
5384 if (flags == IORING_POLL_ADD_MULTI)
5387 upd->old_user_data = READ_ONCE(sqe->addr);
5388 upd->update_events = flags & IORING_POLL_UPDATE_EVENTS;
5389 upd->update_user_data = flags & IORING_POLL_UPDATE_USER_DATA;
5391 upd->new_user_data = READ_ONCE(sqe->off);
5392 if (!upd->update_user_data && upd->new_user_data)
5394 if (upd->update_events)
5395 upd->events = io_poll_parse_events(sqe, flags);
5396 else if (sqe->poll32_events)
5402 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5405 struct io_kiocb *req = wait->private;
5406 struct io_poll_iocb *poll = &req->poll;
5408 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5411 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5412 struct poll_table_struct *p)
5414 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5416 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5419 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5421 struct io_poll_iocb *poll = &req->poll;
5424 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5426 if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->addr)
5428 flags = READ_ONCE(sqe->len);
5429 if (flags & ~IORING_POLL_ADD_MULTI)
5432 poll->events = io_poll_parse_events(sqe, flags);
5436 static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5438 struct io_poll_iocb *poll = &req->poll;
5439 struct io_ring_ctx *ctx = req->ctx;
5440 struct io_poll_table ipt;
5443 ipt.pt._qproc = io_poll_queue_proc;
5445 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5448 if (mask) { /* no async, we'd stolen it */
5450 io_poll_complete(req, mask);
5452 spin_unlock_irq(&ctx->completion_lock);
5455 io_cqring_ev_posted(ctx);
5456 if (poll->events & EPOLLONESHOT)
5462 static int io_poll_update(struct io_kiocb *req, unsigned int issue_flags)
5464 struct io_ring_ctx *ctx = req->ctx;
5465 struct io_kiocb *preq;
5469 spin_lock_irq(&ctx->completion_lock);
5470 preq = io_poll_find(ctx, req->poll_update.old_user_data, true);
5476 if (!req->poll_update.update_events && !req->poll_update.update_user_data) {
5478 ret = io_poll_remove_one(preq) ? 0 : -EALREADY;
5483 * Don't allow racy completion with singleshot, as we cannot safely
5484 * update those. For multishot, if we're racing with completion, just
5485 * let completion re-add it.
5487 completing = !__io_poll_remove_one(preq, &preq->poll, false);
5488 if (completing && (preq->poll.events & EPOLLONESHOT)) {
5492 /* we now have a detached poll request. reissue. */
5496 spin_unlock_irq(&ctx->completion_lock);
5498 io_req_complete(req, ret);
5501 /* only mask one event flags, keep behavior flags */
5502 if (req->poll_update.update_events) {
5503 preq->poll.events &= ~0xffff;
5504 preq->poll.events |= req->poll_update.events & 0xffff;
5505 preq->poll.events |= IO_POLL_UNMASK;
5507 if (req->poll_update.update_user_data)
5508 preq->user_data = req->poll_update.new_user_data;
5509 spin_unlock_irq(&ctx->completion_lock);
5511 /* complete update request, we're done with it */
5512 io_req_complete(req, ret);
5515 ret = io_poll_add(preq, issue_flags);
5518 io_req_complete(preq, ret);
5524 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5526 struct io_timeout_data *data = container_of(timer,
5527 struct io_timeout_data, timer);
5528 struct io_kiocb *req = data->req;
5529 struct io_ring_ctx *ctx = req->ctx;
5530 unsigned long flags;
5532 spin_lock_irqsave(&ctx->completion_lock, flags);
5533 list_del_init(&req->timeout.list);
5534 atomic_set(&req->ctx->cq_timeouts,
5535 atomic_read(&req->ctx->cq_timeouts) + 1);
5537 io_cqring_fill_event(ctx, req->user_data, -ETIME, 0);
5538 io_commit_cqring(ctx);
5539 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5541 io_cqring_ev_posted(ctx);
5544 return HRTIMER_NORESTART;
5547 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5549 __must_hold(&ctx->completion_lock)
5551 struct io_timeout_data *io;
5552 struct io_kiocb *req;
5555 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5556 found = user_data == req->user_data;
5561 return ERR_PTR(-ENOENT);
5563 io = req->async_data;
5564 if (hrtimer_try_to_cancel(&io->timer) == -1)
5565 return ERR_PTR(-EALREADY);
5566 list_del_init(&req->timeout.list);
5570 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5571 __must_hold(&ctx->completion_lock)
5573 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5576 return PTR_ERR(req);
5579 io_cqring_fill_event(ctx, req->user_data, -ECANCELED, 0);
5580 io_put_req_deferred(req, 1);
5584 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5585 struct timespec64 *ts, enum hrtimer_mode mode)
5586 __must_hold(&ctx->completion_lock)
5588 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5589 struct io_timeout_data *data;
5592 return PTR_ERR(req);
5594 req->timeout.off = 0; /* noseq */
5595 data = req->async_data;
5596 list_add_tail(&req->timeout.list, &ctx->timeout_list);
5597 hrtimer_init(&data->timer, CLOCK_MONOTONIC, mode);
5598 data->timer.function = io_timeout_fn;
5599 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
5603 static int io_timeout_remove_prep(struct io_kiocb *req,
5604 const struct io_uring_sqe *sqe)
5606 struct io_timeout_rem *tr = &req->timeout_rem;
5608 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5610 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5612 if (sqe->ioprio || sqe->buf_index || sqe->len)
5615 tr->addr = READ_ONCE(sqe->addr);
5616 tr->flags = READ_ONCE(sqe->timeout_flags);
5617 if (tr->flags & IORING_TIMEOUT_UPDATE) {
5618 if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS))
5620 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
5622 } else if (tr->flags) {
5623 /* timeout removal doesn't support flags */
5630 static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
5632 return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
5637 * Remove or update an existing timeout command
5639 static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
5641 struct io_timeout_rem *tr = &req->timeout_rem;
5642 struct io_ring_ctx *ctx = req->ctx;
5645 spin_lock_irq(&ctx->completion_lock);
5646 if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE))
5647 ret = io_timeout_cancel(ctx, tr->addr);
5649 ret = io_timeout_update(ctx, tr->addr, &tr->ts,
5650 io_translate_timeout_mode(tr->flags));
5652 io_cqring_fill_event(ctx, req->user_data, ret, 0);
5653 io_commit_cqring(ctx);
5654 spin_unlock_irq(&ctx->completion_lock);
5655 io_cqring_ev_posted(ctx);
5662 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5663 bool is_timeout_link)
5665 struct io_timeout_data *data;
5667 u32 off = READ_ONCE(sqe->off);
5669 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5671 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5673 if (off && is_timeout_link)
5675 flags = READ_ONCE(sqe->timeout_flags);
5676 if (flags & ~IORING_TIMEOUT_ABS)
5679 req->timeout.off = off;
5680 if (unlikely(off && !req->ctx->off_timeout_used))
5681 req->ctx->off_timeout_used = true;
5683 if (!req->async_data && io_alloc_async_data(req))
5686 data = req->async_data;
5689 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5692 data->mode = io_translate_timeout_mode(flags);
5693 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5694 if (is_timeout_link)
5695 io_req_track_inflight(req);
5699 static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
5701 struct io_ring_ctx *ctx = req->ctx;
5702 struct io_timeout_data *data = req->async_data;
5703 struct list_head *entry;
5704 u32 tail, off = req->timeout.off;
5706 spin_lock_irq(&ctx->completion_lock);
5709 * sqe->off holds how many events that need to occur for this
5710 * timeout event to be satisfied. If it isn't set, then this is
5711 * a pure timeout request, sequence isn't used.
5713 if (io_is_timeout_noseq(req)) {
5714 entry = ctx->timeout_list.prev;
5718 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5719 req->timeout.target_seq = tail + off;
5721 /* Update the last seq here in case io_flush_timeouts() hasn't.
5722 * This is safe because ->completion_lock is held, and submissions
5723 * and completions are never mixed in the same ->completion_lock section.
5725 ctx->cq_last_tm_flush = tail;
5728 * Insertion sort, ensuring the first entry in the list is always
5729 * the one we need first.
5731 list_for_each_prev(entry, &ctx->timeout_list) {
5732 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5735 if (io_is_timeout_noseq(nxt))
5737 /* nxt.seq is behind @tail, otherwise would've been completed */
5738 if (off >= nxt->timeout.target_seq - tail)
5742 list_add(&req->timeout.list, entry);
5743 data->timer.function = io_timeout_fn;
5744 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5745 spin_unlock_irq(&ctx->completion_lock);
5749 struct io_cancel_data {
5750 struct io_ring_ctx *ctx;
5754 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5756 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5757 struct io_cancel_data *cd = data;
5759 return req->ctx == cd->ctx && req->user_data == cd->user_data;
5762 static int io_async_cancel_one(struct io_uring_task *tctx, u64 user_data,
5763 struct io_ring_ctx *ctx)
5765 struct io_cancel_data data = { .ctx = ctx, .user_data = user_data, };
5766 enum io_wq_cancel cancel_ret;
5769 if (!tctx || !tctx->io_wq)
5772 cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, &data, false);
5773 switch (cancel_ret) {
5774 case IO_WQ_CANCEL_OK:
5777 case IO_WQ_CANCEL_RUNNING:
5780 case IO_WQ_CANCEL_NOTFOUND:
5788 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5789 struct io_kiocb *req, __u64 sqe_addr,
5792 unsigned long flags;
5795 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5796 spin_lock_irqsave(&ctx->completion_lock, flags);
5799 ret = io_timeout_cancel(ctx, sqe_addr);
5802 ret = io_poll_cancel(ctx, sqe_addr, false);
5806 io_cqring_fill_event(ctx, req->user_data, ret, 0);
5807 io_commit_cqring(ctx);
5808 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5809 io_cqring_ev_posted(ctx);
5815 static int io_async_cancel_prep(struct io_kiocb *req,
5816 const struct io_uring_sqe *sqe)
5818 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5820 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5822 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5825 req->cancel.addr = READ_ONCE(sqe->addr);
5829 static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
5831 struct io_ring_ctx *ctx = req->ctx;
5832 u64 sqe_addr = req->cancel.addr;
5833 struct io_tctx_node *node;
5836 /* tasks should wait for their io-wq threads, so safe w/o sync */
5837 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5838 spin_lock_irq(&ctx->completion_lock);
5841 ret = io_timeout_cancel(ctx, sqe_addr);
5844 ret = io_poll_cancel(ctx, sqe_addr, false);
5847 spin_unlock_irq(&ctx->completion_lock);
5849 /* slow path, try all io-wq's */
5850 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5852 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
5853 struct io_uring_task *tctx = node->task->io_uring;
5855 ret = io_async_cancel_one(tctx, req->cancel.addr, ctx);
5859 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5861 spin_lock_irq(&ctx->completion_lock);
5863 io_cqring_fill_event(ctx, req->user_data, ret, 0);
5864 io_commit_cqring(ctx);
5865 spin_unlock_irq(&ctx->completion_lock);
5866 io_cqring_ev_posted(ctx);
5874 static int io_rsrc_update_prep(struct io_kiocb *req,
5875 const struct io_uring_sqe *sqe)
5877 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5879 if (sqe->ioprio || sqe->rw_flags)
5882 req->rsrc_update.offset = READ_ONCE(sqe->off);
5883 req->rsrc_update.nr_args = READ_ONCE(sqe->len);
5884 if (!req->rsrc_update.nr_args)
5886 req->rsrc_update.arg = READ_ONCE(sqe->addr);
5890 static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
5892 struct io_ring_ctx *ctx = req->ctx;
5893 struct io_uring_rsrc_update2 up;
5896 if (issue_flags & IO_URING_F_NONBLOCK)
5899 up.offset = req->rsrc_update.offset;
5900 up.data = req->rsrc_update.arg;
5905 mutex_lock(&ctx->uring_lock);
5906 ret = __io_register_rsrc_update(ctx, IORING_RSRC_FILE,
5907 &up, req->rsrc_update.nr_args);
5908 mutex_unlock(&ctx->uring_lock);
5912 __io_req_complete(req, issue_flags, ret, 0);
5916 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5918 switch (req->opcode) {
5921 case IORING_OP_READV:
5922 case IORING_OP_READ_FIXED:
5923 case IORING_OP_READ:
5924 return io_read_prep(req, sqe);
5925 case IORING_OP_WRITEV:
5926 case IORING_OP_WRITE_FIXED:
5927 case IORING_OP_WRITE:
5928 return io_write_prep(req, sqe);
5929 case IORING_OP_POLL_ADD:
5930 return io_poll_add_prep(req, sqe);
5931 case IORING_OP_POLL_REMOVE:
5932 return io_poll_update_prep(req, sqe);
5933 case IORING_OP_FSYNC:
5934 return io_fsync_prep(req, sqe);
5935 case IORING_OP_SYNC_FILE_RANGE:
5936 return io_sfr_prep(req, sqe);
5937 case IORING_OP_SENDMSG:
5938 case IORING_OP_SEND:
5939 return io_sendmsg_prep(req, sqe);
5940 case IORING_OP_RECVMSG:
5941 case IORING_OP_RECV:
5942 return io_recvmsg_prep(req, sqe);
5943 case IORING_OP_CONNECT:
5944 return io_connect_prep(req, sqe);
5945 case IORING_OP_TIMEOUT:
5946 return io_timeout_prep(req, sqe, false);
5947 case IORING_OP_TIMEOUT_REMOVE:
5948 return io_timeout_remove_prep(req, sqe);
5949 case IORING_OP_ASYNC_CANCEL:
5950 return io_async_cancel_prep(req, sqe);
5951 case IORING_OP_LINK_TIMEOUT:
5952 return io_timeout_prep(req, sqe, true);
5953 case IORING_OP_ACCEPT:
5954 return io_accept_prep(req, sqe);
5955 case IORING_OP_FALLOCATE:
5956 return io_fallocate_prep(req, sqe);
5957 case IORING_OP_OPENAT:
5958 return io_openat_prep(req, sqe);
5959 case IORING_OP_CLOSE:
5960 return io_close_prep(req, sqe);
5961 case IORING_OP_FILES_UPDATE:
5962 return io_rsrc_update_prep(req, sqe);
5963 case IORING_OP_STATX:
5964 return io_statx_prep(req, sqe);
5965 case IORING_OP_FADVISE:
5966 return io_fadvise_prep(req, sqe);
5967 case IORING_OP_MADVISE:
5968 return io_madvise_prep(req, sqe);
5969 case IORING_OP_OPENAT2:
5970 return io_openat2_prep(req, sqe);
5971 case IORING_OP_EPOLL_CTL:
5972 return io_epoll_ctl_prep(req, sqe);
5973 case IORING_OP_SPLICE:
5974 return io_splice_prep(req, sqe);
5975 case IORING_OP_PROVIDE_BUFFERS:
5976 return io_provide_buffers_prep(req, sqe);
5977 case IORING_OP_REMOVE_BUFFERS:
5978 return io_remove_buffers_prep(req, sqe);
5980 return io_tee_prep(req, sqe);
5981 case IORING_OP_SHUTDOWN:
5982 return io_shutdown_prep(req, sqe);
5983 case IORING_OP_RENAMEAT:
5984 return io_renameat_prep(req, sqe);
5985 case IORING_OP_UNLINKAT:
5986 return io_unlinkat_prep(req, sqe);
5989 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5994 static int io_req_prep_async(struct io_kiocb *req)
5996 if (!io_op_defs[req->opcode].needs_async_setup)
5998 if (WARN_ON_ONCE(req->async_data))
6000 if (io_alloc_async_data(req))
6003 switch (req->opcode) {
6004 case IORING_OP_READV:
6005 return io_rw_prep_async(req, READ);
6006 case IORING_OP_WRITEV:
6007 return io_rw_prep_async(req, WRITE);
6008 case IORING_OP_SENDMSG:
6009 return io_sendmsg_prep_async(req);
6010 case IORING_OP_RECVMSG:
6011 return io_recvmsg_prep_async(req);
6012 case IORING_OP_CONNECT:
6013 return io_connect_prep_async(req);
6015 printk_once(KERN_WARNING "io_uring: prep_async() bad opcode %d\n",
6020 static u32 io_get_sequence(struct io_kiocb *req)
6022 u32 seq = req->ctx->cached_sq_head;
6024 /* need original cached_sq_head, but it was increased for each req */
6025 io_for_each_link(req, req)
6030 static bool io_drain_req(struct io_kiocb *req)
6032 struct io_kiocb *pos;
6033 struct io_ring_ctx *ctx = req->ctx;
6034 struct io_defer_entry *de;
6039 * If we need to drain a request in the middle of a link, drain the
6040 * head request and the next request/link after the current link.
6041 * Considering sequential execution of links, IOSQE_IO_DRAIN will be
6042 * maintained for every request of our link.
6044 if (ctx->drain_next) {
6045 req->flags |= REQ_F_IO_DRAIN;
6046 ctx->drain_next = false;
6048 /* not interested in head, start from the first linked */
6049 io_for_each_link(pos, req->link) {
6050 if (pos->flags & REQ_F_IO_DRAIN) {
6051 ctx->drain_next = true;
6052 req->flags |= REQ_F_IO_DRAIN;
6057 /* Still need defer if there is pending req in defer list. */
6058 if (likely(list_empty_careful(&ctx->defer_list) &&
6059 !(req->flags & REQ_F_IO_DRAIN))) {
6060 ctx->drain_active = false;
6064 seq = io_get_sequence(req);
6065 /* Still a chance to pass the sequence check */
6066 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
6069 ret = io_req_prep_async(req);
6072 io_prep_async_link(req);
6073 de = kmalloc(sizeof(*de), GFP_KERNEL);
6075 io_req_complete_failed(req, ret);
6079 spin_lock_irq(&ctx->completion_lock);
6080 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
6081 spin_unlock_irq(&ctx->completion_lock);
6083 io_queue_async_work(req);
6087 trace_io_uring_defer(ctx, req, req->user_data);
6090 list_add_tail(&de->list, &ctx->defer_list);
6091 spin_unlock_irq(&ctx->completion_lock);
6095 static void io_clean_op(struct io_kiocb *req)
6097 if (req->flags & REQ_F_BUFFER_SELECTED) {
6098 switch (req->opcode) {
6099 case IORING_OP_READV:
6100 case IORING_OP_READ_FIXED:
6101 case IORING_OP_READ:
6102 kfree((void *)(unsigned long)req->rw.addr);
6104 case IORING_OP_RECVMSG:
6105 case IORING_OP_RECV:
6106 kfree(req->sr_msg.kbuf);
6111 if (req->flags & REQ_F_NEED_CLEANUP) {
6112 switch (req->opcode) {
6113 case IORING_OP_READV:
6114 case IORING_OP_READ_FIXED:
6115 case IORING_OP_READ:
6116 case IORING_OP_WRITEV:
6117 case IORING_OP_WRITE_FIXED:
6118 case IORING_OP_WRITE: {
6119 struct io_async_rw *io = req->async_data;
6121 kfree(io->free_iovec);
6124 case IORING_OP_RECVMSG:
6125 case IORING_OP_SENDMSG: {
6126 struct io_async_msghdr *io = req->async_data;
6128 kfree(io->free_iov);
6131 case IORING_OP_SPLICE:
6133 if (!(req->splice.flags & SPLICE_F_FD_IN_FIXED))
6134 io_put_file(req->splice.file_in);
6136 case IORING_OP_OPENAT:
6137 case IORING_OP_OPENAT2:
6138 if (req->open.filename)
6139 putname(req->open.filename);
6141 case IORING_OP_RENAMEAT:
6142 putname(req->rename.oldpath);
6143 putname(req->rename.newpath);
6145 case IORING_OP_UNLINKAT:
6146 putname(req->unlink.filename);
6150 if ((req->flags & REQ_F_POLLED) && req->apoll) {
6151 kfree(req->apoll->double_poll);
6155 if (req->flags & REQ_F_INFLIGHT) {
6156 struct io_uring_task *tctx = req->task->io_uring;
6158 atomic_dec(&tctx->inflight_tracked);
6160 if (req->flags & REQ_F_CREDS)
6161 put_cred(req->creds);
6163 req->flags &= ~IO_REQ_CLEAN_FLAGS;
6166 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
6168 struct io_ring_ctx *ctx = req->ctx;
6169 const struct cred *creds = NULL;
6172 if ((req->flags & REQ_F_CREDS) && req->creds != current_cred())
6173 creds = override_creds(req->creds);
6175 switch (req->opcode) {
6177 ret = io_nop(req, issue_flags);
6179 case IORING_OP_READV:
6180 case IORING_OP_READ_FIXED:
6181 case IORING_OP_READ:
6182 ret = io_read(req, issue_flags);
6184 case IORING_OP_WRITEV:
6185 case IORING_OP_WRITE_FIXED:
6186 case IORING_OP_WRITE:
6187 ret = io_write(req, issue_flags);
6189 case IORING_OP_FSYNC:
6190 ret = io_fsync(req, issue_flags);
6192 case IORING_OP_POLL_ADD:
6193 ret = io_poll_add(req, issue_flags);
6195 case IORING_OP_POLL_REMOVE:
6196 ret = io_poll_update(req, issue_flags);
6198 case IORING_OP_SYNC_FILE_RANGE:
6199 ret = io_sync_file_range(req, issue_flags);
6201 case IORING_OP_SENDMSG:
6202 ret = io_sendmsg(req, issue_flags);
6204 case IORING_OP_SEND:
6205 ret = io_send(req, issue_flags);
6207 case IORING_OP_RECVMSG:
6208 ret = io_recvmsg(req, issue_flags);
6210 case IORING_OP_RECV:
6211 ret = io_recv(req, issue_flags);
6213 case IORING_OP_TIMEOUT:
6214 ret = io_timeout(req, issue_flags);
6216 case IORING_OP_TIMEOUT_REMOVE:
6217 ret = io_timeout_remove(req, issue_flags);
6219 case IORING_OP_ACCEPT:
6220 ret = io_accept(req, issue_flags);
6222 case IORING_OP_CONNECT:
6223 ret = io_connect(req, issue_flags);
6225 case IORING_OP_ASYNC_CANCEL:
6226 ret = io_async_cancel(req, issue_flags);
6228 case IORING_OP_FALLOCATE:
6229 ret = io_fallocate(req, issue_flags);
6231 case IORING_OP_OPENAT:
6232 ret = io_openat(req, issue_flags);
6234 case IORING_OP_CLOSE:
6235 ret = io_close(req, issue_flags);
6237 case IORING_OP_FILES_UPDATE:
6238 ret = io_files_update(req, issue_flags);
6240 case IORING_OP_STATX:
6241 ret = io_statx(req, issue_flags);
6243 case IORING_OP_FADVISE:
6244 ret = io_fadvise(req, issue_flags);
6246 case IORING_OP_MADVISE:
6247 ret = io_madvise(req, issue_flags);
6249 case IORING_OP_OPENAT2:
6250 ret = io_openat2(req, issue_flags);
6252 case IORING_OP_EPOLL_CTL:
6253 ret = io_epoll_ctl(req, issue_flags);
6255 case IORING_OP_SPLICE:
6256 ret = io_splice(req, issue_flags);
6258 case IORING_OP_PROVIDE_BUFFERS:
6259 ret = io_provide_buffers(req, issue_flags);
6261 case IORING_OP_REMOVE_BUFFERS:
6262 ret = io_remove_buffers(req, issue_flags);
6265 ret = io_tee(req, issue_flags);
6267 case IORING_OP_SHUTDOWN:
6268 ret = io_shutdown(req, issue_flags);
6270 case IORING_OP_RENAMEAT:
6271 ret = io_renameat(req, issue_flags);
6273 case IORING_OP_UNLINKAT:
6274 ret = io_unlinkat(req, issue_flags);
6282 revert_creds(creds);
6285 /* If the op doesn't have a file, we're not polling for it */
6286 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file)
6287 io_iopoll_req_issued(req);
6292 static void io_wq_submit_work(struct io_wq_work *work)
6294 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6295 struct io_kiocb *timeout;
6298 timeout = io_prep_linked_timeout(req);
6300 io_queue_linked_timeout(timeout);
6302 if (work->flags & IO_WQ_WORK_CANCEL)
6307 ret = io_issue_sqe(req, 0);
6309 * We can get EAGAIN for polled IO even though we're
6310 * forcing a sync submission from here, since we can't
6311 * wait for request slots on the block side.
6319 /* avoid locking problems by failing it from a clean context */
6321 /* io-wq is going to take one down */
6323 io_req_task_queue_fail(req, ret);
6327 #define FFS_ASYNC_READ 0x1UL
6328 #define FFS_ASYNC_WRITE 0x2UL
6330 #define FFS_ISREG 0x4UL
6332 #define FFS_ISREG 0x0UL
6334 #define FFS_MASK ~(FFS_ASYNC_READ|FFS_ASYNC_WRITE|FFS_ISREG)
6336 static inline struct io_fixed_file *io_fixed_file_slot(struct io_file_table *table,
6339 struct io_fixed_file *table_l2;
6341 table_l2 = table->files[i >> IORING_FILE_TABLE_SHIFT];
6342 return &table_l2[i & IORING_FILE_TABLE_MASK];
6345 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6348 struct io_fixed_file *slot = io_fixed_file_slot(&ctx->file_table, index);
6350 return (struct file *) (slot->file_ptr & FFS_MASK);
6353 static void io_fixed_file_set(struct io_fixed_file *file_slot, struct file *file)
6355 unsigned long file_ptr = (unsigned long) file;
6357 if (__io_file_supports_async(file, READ))
6358 file_ptr |= FFS_ASYNC_READ;
6359 if (__io_file_supports_async(file, WRITE))
6360 file_ptr |= FFS_ASYNC_WRITE;
6361 if (S_ISREG(file_inode(file)->i_mode))
6362 file_ptr |= FFS_ISREG;
6363 file_slot->file_ptr = file_ptr;
6366 static struct file *io_file_get(struct io_submit_state *state,
6367 struct io_kiocb *req, int fd, bool fixed)
6369 struct io_ring_ctx *ctx = req->ctx;
6373 unsigned long file_ptr;
6375 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6377 fd = array_index_nospec(fd, ctx->nr_user_files);
6378 file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr;
6379 file = (struct file *) (file_ptr & FFS_MASK);
6380 file_ptr &= ~FFS_MASK;
6381 /* mask in overlapping REQ_F and FFS bits */
6382 req->flags |= (file_ptr << REQ_F_ASYNC_READ_BIT);
6383 io_req_set_rsrc_node(req);
6385 trace_io_uring_file_get(ctx, fd);
6386 file = __io_file_get(state, fd);
6388 /* we don't allow fixed io_uring files */
6389 if (file && unlikely(file->f_op == &io_uring_fops))
6390 io_req_track_inflight(req);
6396 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6398 struct io_timeout_data *data = container_of(timer,
6399 struct io_timeout_data, timer);
6400 struct io_kiocb *prev, *req = data->req;
6401 struct io_ring_ctx *ctx = req->ctx;
6402 unsigned long flags;
6404 spin_lock_irqsave(&ctx->completion_lock, flags);
6405 prev = req->timeout.head;
6406 req->timeout.head = NULL;
6409 * We don't expect the list to be empty, that will only happen if we
6410 * race with the completion of the linked work.
6413 io_remove_next_linked(prev);
6414 if (!req_ref_inc_not_zero(prev))
6417 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6420 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6421 io_put_req_deferred(prev, 1);
6422 io_put_req_deferred(req, 1);
6424 io_req_complete_post(req, -ETIME, 0);
6426 return HRTIMER_NORESTART;
6429 static void io_queue_linked_timeout(struct io_kiocb *req)
6431 struct io_ring_ctx *ctx = req->ctx;
6433 spin_lock_irq(&ctx->completion_lock);
6435 * If the back reference is NULL, then our linked request finished
6436 * before we got a chance to setup the timer
6438 if (req->timeout.head) {
6439 struct io_timeout_data *data = req->async_data;
6441 data->timer.function = io_link_timeout_fn;
6442 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6445 spin_unlock_irq(&ctx->completion_lock);
6446 /* drop submission reference */
6450 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6452 struct io_kiocb *nxt = req->link;
6454 if (!nxt || (req->flags & REQ_F_LINK_TIMEOUT) ||
6455 nxt->opcode != IORING_OP_LINK_TIMEOUT)
6458 nxt->timeout.head = req;
6459 nxt->flags |= REQ_F_LTIMEOUT_ACTIVE;
6460 req->flags |= REQ_F_LINK_TIMEOUT;
6464 static void __io_queue_sqe(struct io_kiocb *req)
6466 struct io_kiocb *linked_timeout = io_prep_linked_timeout(req);
6470 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
6473 * We async punt it if the file wasn't marked NOWAIT, or if the file
6474 * doesn't support non-blocking read/write attempts
6477 /* drop submission reference */
6478 if (req->flags & REQ_F_COMPLETE_INLINE) {
6479 struct io_ring_ctx *ctx = req->ctx;
6480 struct io_comp_state *cs = &ctx->submit_state.comp;
6482 cs->reqs[cs->nr++] = req;
6483 if (cs->nr == ARRAY_SIZE(cs->reqs))
6484 io_submit_flush_completions(ctx);
6488 } else if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6489 switch (io_arm_poll_handler(req)) {
6490 case IO_APOLL_READY:
6492 case IO_APOLL_ABORTED:
6494 * Queued up for async execution, worker will release
6495 * submit reference when the iocb is actually submitted.
6497 io_queue_async_work(req);
6501 io_req_complete_failed(req, ret);
6504 io_queue_linked_timeout(linked_timeout);
6507 static inline void io_queue_sqe(struct io_kiocb *req)
6509 if (unlikely(req->ctx->drain_active) && io_drain_req(req))
6512 if (likely(!(req->flags & REQ_F_FORCE_ASYNC))) {
6513 __io_queue_sqe(req);
6515 int ret = io_req_prep_async(req);
6518 io_req_complete_failed(req, ret);
6520 io_queue_async_work(req);
6525 * Check SQE restrictions (opcode and flags).
6527 * Returns 'true' if SQE is allowed, 'false' otherwise.
6529 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6530 struct io_kiocb *req,
6531 unsigned int sqe_flags)
6533 if (likely(!ctx->restricted))
6536 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6539 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6540 ctx->restrictions.sqe_flags_required)
6543 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6544 ctx->restrictions.sqe_flags_required))
6550 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6551 const struct io_uring_sqe *sqe)
6553 struct io_submit_state *state;
6554 unsigned int sqe_flags;
6555 int personality, ret = 0;
6557 req->opcode = READ_ONCE(sqe->opcode);
6558 /* same numerical values with corresponding REQ_F_*, safe to copy */
6559 req->flags = sqe_flags = READ_ONCE(sqe->flags);
6560 req->user_data = READ_ONCE(sqe->user_data);
6562 req->fixed_rsrc_refs = NULL;
6563 /* one is dropped after submission, the other at completion */
6564 atomic_set(&req->refs, 2);
6565 req->task = current;
6567 /* enforce forwards compatibility on users */
6568 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6570 if (unlikely(req->opcode >= IORING_OP_LAST))
6572 if (!io_check_restriction(ctx, req, sqe_flags))
6575 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6576 !io_op_defs[req->opcode].buffer_select)
6578 if (unlikely(sqe_flags & IOSQE_IO_DRAIN))
6579 ctx->drain_active = true;
6581 personality = READ_ONCE(sqe->personality);
6583 req->creds = xa_load(&ctx->personalities, personality);
6586 get_cred(req->creds);
6587 req->flags |= REQ_F_CREDS;
6589 state = &ctx->submit_state;
6592 * Plug now if we have more than 1 IO left after this, and the target
6593 * is potentially a read/write to block based storage.
6595 if (!state->plug_started && state->ios_left > 1 &&
6596 io_op_defs[req->opcode].plug) {
6597 blk_start_plug(&state->plug);
6598 state->plug_started = true;
6601 if (io_op_defs[req->opcode].needs_file) {
6602 bool fixed = req->flags & REQ_F_FIXED_FILE;
6604 req->file = io_file_get(state, req, READ_ONCE(sqe->fd), fixed);
6605 if (unlikely(!req->file))
6613 static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
6614 const struct io_uring_sqe *sqe)
6616 struct io_submit_link *link = &ctx->submit_state.link;
6619 ret = io_init_req(ctx, req, sqe);
6620 if (unlikely(ret)) {
6623 /* fail even hard links since we don't submit */
6624 req_set_fail(link->head);
6625 io_req_complete_failed(link->head, -ECANCELED);
6628 io_req_complete_failed(req, ret);
6632 ret = io_req_prep(req, sqe);
6636 /* don't need @sqe from now on */
6637 trace_io_uring_submit_sqe(ctx, req, req->opcode, req->user_data,
6639 ctx->flags & IORING_SETUP_SQPOLL);
6642 * If we already have a head request, queue this one for async
6643 * submittal once the head completes. If we don't have a head but
6644 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6645 * submitted sync once the chain is complete. If none of those
6646 * conditions are true (normal request), then just queue it.
6649 struct io_kiocb *head = link->head;
6651 ret = io_req_prep_async(req);
6654 trace_io_uring_link(ctx, req, head);
6655 link->last->link = req;
6658 /* last request of a link, enqueue the link */
6659 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6664 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6676 * Batched submission is done, ensure local IO is flushed out.
6678 static void io_submit_state_end(struct io_submit_state *state,
6679 struct io_ring_ctx *ctx)
6681 if (state->link.head)
6682 io_queue_sqe(state->link.head);
6684 io_submit_flush_completions(ctx);
6685 if (state->plug_started)
6686 blk_finish_plug(&state->plug);
6687 io_state_file_put(state);
6691 * Start submission side cache.
6693 static void io_submit_state_start(struct io_submit_state *state,
6694 unsigned int max_ios)
6696 state->plug_started = false;
6697 state->ios_left = max_ios;
6698 /* set only head, no need to init link_last in advance */
6699 state->link.head = NULL;
6702 static void io_commit_sqring(struct io_ring_ctx *ctx)
6704 struct io_rings *rings = ctx->rings;
6707 * Ensure any loads from the SQEs are done at this point,
6708 * since once we write the new head, the application could
6709 * write new data to them.
6711 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6715 * Fetch an sqe, if one is available. Note this returns a pointer to memory
6716 * that is mapped by userspace. This means that care needs to be taken to
6717 * ensure that reads are stable, as we cannot rely on userspace always
6718 * being a good citizen. If members of the sqe are validated and then later
6719 * used, it's important that those reads are done through READ_ONCE() to
6720 * prevent a re-load down the line.
6722 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6724 unsigned head, mask = ctx->sq_entries - 1;
6725 unsigned sq_idx = ctx->cached_sq_head++ & mask;
6728 * The cached sq head (or cq tail) serves two purposes:
6730 * 1) allows us to batch the cost of updating the user visible
6732 * 2) allows the kernel side to track the head on its own, even
6733 * though the application is the one updating it.
6735 head = READ_ONCE(ctx->sq_array[sq_idx]);
6736 if (likely(head < ctx->sq_entries))
6737 return &ctx->sq_sqes[head];
6739 /* drop invalid entries */
6741 WRITE_ONCE(ctx->rings->sq_dropped,
6742 READ_ONCE(ctx->rings->sq_dropped) + 1);
6746 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6748 struct io_uring_task *tctx;
6751 /* make sure SQ entry isn't read before tail */
6752 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6753 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6756 tctx = current->io_uring;
6757 tctx->cached_refs -= nr;
6758 if (unlikely(tctx->cached_refs < 0)) {
6759 unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR;
6761 percpu_counter_add(&tctx->inflight, refill);
6762 refcount_add(refill, ¤t->usage);
6763 tctx->cached_refs += refill;
6765 io_submit_state_start(&ctx->submit_state, nr);
6767 while (submitted < nr) {
6768 const struct io_uring_sqe *sqe;
6769 struct io_kiocb *req;
6771 req = io_alloc_req(ctx);
6772 if (unlikely(!req)) {
6774 submitted = -EAGAIN;
6777 sqe = io_get_sqe(ctx);
6778 if (unlikely(!sqe)) {
6779 kmem_cache_free(req_cachep, req);
6782 /* will complete beyond this point, count as submitted */
6784 if (io_submit_sqe(ctx, req, sqe))
6788 if (unlikely(submitted != nr)) {
6789 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6790 int unused = nr - ref_used;
6792 current->io_uring->cached_refs += unused;
6793 percpu_ref_put_many(&ctx->refs, unused);
6796 io_submit_state_end(&ctx->submit_state, ctx);
6797 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6798 io_commit_sqring(ctx);
6803 static inline bool io_sqd_events_pending(struct io_sq_data *sqd)
6805 return READ_ONCE(sqd->state);
6808 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6810 /* Tell userspace we may need a wakeup call */
6811 spin_lock_irq(&ctx->completion_lock);
6812 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6813 spin_unlock_irq(&ctx->completion_lock);
6816 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6818 spin_lock_irq(&ctx->completion_lock);
6819 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6820 spin_unlock_irq(&ctx->completion_lock);
6823 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
6825 unsigned int to_submit;
6828 to_submit = io_sqring_entries(ctx);
6829 /* if we're handling multiple rings, cap submit size for fairness */
6830 if (cap_entries && to_submit > IORING_SQPOLL_CAP_ENTRIES_VALUE)
6831 to_submit = IORING_SQPOLL_CAP_ENTRIES_VALUE;
6833 if (!list_empty(&ctx->iopoll_list) || to_submit) {
6834 unsigned nr_events = 0;
6835 const struct cred *creds = NULL;
6837 if (ctx->sq_creds != current_cred())
6838 creds = override_creds(ctx->sq_creds);
6840 mutex_lock(&ctx->uring_lock);
6841 if (!list_empty(&ctx->iopoll_list))
6842 io_do_iopoll(ctx, &nr_events, 0);
6845 * Don't submit if refs are dying, good for io_uring_register(),
6846 * but also it is relied upon by io_ring_exit_work()
6848 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)) &&
6849 !(ctx->flags & IORING_SETUP_R_DISABLED))
6850 ret = io_submit_sqes(ctx, to_submit);
6851 mutex_unlock(&ctx->uring_lock);
6853 if (to_submit && wq_has_sleeper(&ctx->sqo_sq_wait))
6854 wake_up(&ctx->sqo_sq_wait);
6856 revert_creds(creds);
6862 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
6864 struct io_ring_ctx *ctx;
6865 unsigned sq_thread_idle = 0;
6867 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6868 sq_thread_idle = max(sq_thread_idle, ctx->sq_thread_idle);
6869 sqd->sq_thread_idle = sq_thread_idle;
6872 static bool io_sqd_handle_event(struct io_sq_data *sqd)
6874 bool did_sig = false;
6875 struct ksignal ksig;
6877 if (test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state) ||
6878 signal_pending(current)) {
6879 mutex_unlock(&sqd->lock);
6880 if (signal_pending(current))
6881 did_sig = get_signal(&ksig);
6883 mutex_lock(&sqd->lock);
6885 return did_sig || test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
6888 static int io_sq_thread(void *data)
6890 struct io_sq_data *sqd = data;
6891 struct io_ring_ctx *ctx;
6892 unsigned long timeout = 0;
6893 char buf[TASK_COMM_LEN];
6896 snprintf(buf, sizeof(buf), "iou-sqp-%d", sqd->task_pid);
6897 set_task_comm(current, buf);
6899 if (sqd->sq_cpu != -1)
6900 set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
6902 set_cpus_allowed_ptr(current, cpu_online_mask);
6903 current->flags |= PF_NO_SETAFFINITY;
6905 mutex_lock(&sqd->lock);
6907 bool cap_entries, sqt_spin = false;
6909 if (io_sqd_events_pending(sqd) || signal_pending(current)) {
6910 if (io_sqd_handle_event(sqd))
6912 timeout = jiffies + sqd->sq_thread_idle;
6915 cap_entries = !list_is_singular(&sqd->ctx_list);
6916 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6917 int ret = __io_sq_thread(ctx, cap_entries);
6919 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
6922 if (io_run_task_work())
6925 if (sqt_spin || !time_after(jiffies, timeout)) {
6928 timeout = jiffies + sqd->sq_thread_idle;
6932 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
6933 if (!io_sqd_events_pending(sqd) && !current->task_works) {
6934 bool needs_sched = true;
6936 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6937 io_ring_set_wakeup_flag(ctx);
6939 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6940 !list_empty_careful(&ctx->iopoll_list)) {
6941 needs_sched = false;
6944 if (io_sqring_entries(ctx)) {
6945 needs_sched = false;
6951 mutex_unlock(&sqd->lock);
6953 mutex_lock(&sqd->lock);
6955 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6956 io_ring_clear_wakeup_flag(ctx);
6959 finish_wait(&sqd->wait, &wait);
6960 timeout = jiffies + sqd->sq_thread_idle;
6963 io_uring_cancel_generic(true, sqd);
6965 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6966 io_ring_set_wakeup_flag(ctx);
6968 mutex_unlock(&sqd->lock);
6970 complete(&sqd->exited);
6974 struct io_wait_queue {
6975 struct wait_queue_entry wq;
6976 struct io_ring_ctx *ctx;
6978 unsigned nr_timeouts;
6981 static inline bool io_should_wake(struct io_wait_queue *iowq)
6983 struct io_ring_ctx *ctx = iowq->ctx;
6986 * Wake up if we have enough events, or if a timeout occurred since we
6987 * started waiting. For timeouts, we always want to return to userspace,
6988 * regardless of event count.
6990 return io_cqring_events(ctx) >= iowq->to_wait ||
6991 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6994 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6995 int wake_flags, void *key)
6997 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
7001 * Cannot safely flush overflowed CQEs from here, ensure we wake up
7002 * the task, and the next invocation will do it.
7004 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->check_cq_overflow))
7005 return autoremove_wake_function(curr, mode, wake_flags, key);
7009 static int io_run_task_work_sig(void)
7011 if (io_run_task_work())
7013 if (!signal_pending(current))
7015 if (test_thread_flag(TIF_NOTIFY_SIGNAL))
7016 return -ERESTARTSYS;
7020 /* when returns >0, the caller should retry */
7021 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
7022 struct io_wait_queue *iowq,
7023 signed long *timeout)
7027 /* make sure we run task_work before checking for signals */
7028 ret = io_run_task_work_sig();
7029 if (ret || io_should_wake(iowq))
7031 /* let the caller flush overflows, retry */
7032 if (test_bit(0, &ctx->check_cq_overflow))
7035 *timeout = schedule_timeout(*timeout);
7036 return !*timeout ? -ETIME : 1;
7040 * Wait until events become available, if we don't already have some. The
7041 * application must reap them itself, as they reside on the shared cq ring.
7043 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
7044 const sigset_t __user *sig, size_t sigsz,
7045 struct __kernel_timespec __user *uts)
7047 struct io_wait_queue iowq = {
7050 .func = io_wake_function,
7051 .entry = LIST_HEAD_INIT(iowq.wq.entry),
7054 .to_wait = min_events,
7056 struct io_rings *rings = ctx->rings;
7057 signed long timeout = MAX_SCHEDULE_TIMEOUT;
7061 io_cqring_overflow_flush(ctx, false);
7062 if (io_cqring_events(ctx) >= min_events)
7064 if (!io_run_task_work())
7069 #ifdef CONFIG_COMPAT
7070 if (in_compat_syscall())
7071 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
7075 ret = set_user_sigmask(sig, sigsz);
7082 struct timespec64 ts;
7084 if (get_timespec64(&ts, uts))
7086 timeout = timespec64_to_jiffies(&ts);
7089 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
7090 trace_io_uring_cqring_wait(ctx, min_events);
7092 /* if we can't even flush overflow, don't wait for more */
7093 if (!io_cqring_overflow_flush(ctx, false)) {
7097 prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq,
7098 TASK_INTERRUPTIBLE);
7099 ret = io_cqring_wait_schedule(ctx, &iowq, &timeout);
7100 finish_wait(&ctx->cq_wait, &iowq.wq);
7104 restore_saved_sigmask_unless(ret == -EINTR);
7106 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
7109 static void io_free_page_table(void **table, size_t size)
7111 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7113 for (i = 0; i < nr_tables; i++)
7118 static void **io_alloc_page_table(size_t size)
7120 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7121 size_t init_size = size;
7124 table = kcalloc(nr_tables, sizeof(*table), GFP_KERNEL);
7128 for (i = 0; i < nr_tables; i++) {
7129 unsigned int this_size = min_t(size_t, size, PAGE_SIZE);
7131 table[i] = kzalloc(this_size, GFP_KERNEL);
7133 io_free_page_table(table, init_size);
7141 static inline void io_rsrc_ref_lock(struct io_ring_ctx *ctx)
7143 spin_lock_bh(&ctx->rsrc_ref_lock);
7146 static inline void io_rsrc_ref_unlock(struct io_ring_ctx *ctx)
7148 spin_unlock_bh(&ctx->rsrc_ref_lock);
7151 static void io_rsrc_node_destroy(struct io_rsrc_node *ref_node)
7153 percpu_ref_exit(&ref_node->refs);
7157 static void io_rsrc_node_switch(struct io_ring_ctx *ctx,
7158 struct io_rsrc_data *data_to_kill)
7160 WARN_ON_ONCE(!ctx->rsrc_backup_node);
7161 WARN_ON_ONCE(data_to_kill && !ctx->rsrc_node);
7164 struct io_rsrc_node *rsrc_node = ctx->rsrc_node;
7166 rsrc_node->rsrc_data = data_to_kill;
7167 io_rsrc_ref_lock(ctx);
7168 list_add_tail(&rsrc_node->node, &ctx->rsrc_ref_list);
7169 io_rsrc_ref_unlock(ctx);
7171 atomic_inc(&data_to_kill->refs);
7172 percpu_ref_kill(&rsrc_node->refs);
7173 ctx->rsrc_node = NULL;
7176 if (!ctx->rsrc_node) {
7177 ctx->rsrc_node = ctx->rsrc_backup_node;
7178 ctx->rsrc_backup_node = NULL;
7182 static int io_rsrc_node_switch_start(struct io_ring_ctx *ctx)
7184 if (ctx->rsrc_backup_node)
7186 ctx->rsrc_backup_node = io_rsrc_node_alloc(ctx);
7187 return ctx->rsrc_backup_node ? 0 : -ENOMEM;
7190 static int io_rsrc_ref_quiesce(struct io_rsrc_data *data, struct io_ring_ctx *ctx)
7194 /* As we may drop ->uring_lock, other task may have started quiesce */
7198 data->quiesce = true;
7200 ret = io_rsrc_node_switch_start(ctx);
7203 io_rsrc_node_switch(ctx, data);
7205 /* kill initial ref, already quiesced if zero */
7206 if (atomic_dec_and_test(&data->refs))
7208 flush_delayed_work(&ctx->rsrc_put_work);
7209 ret = wait_for_completion_interruptible(&data->done);
7213 atomic_inc(&data->refs);
7214 /* wait for all works potentially completing data->done */
7215 flush_delayed_work(&ctx->rsrc_put_work);
7216 reinit_completion(&data->done);
7218 mutex_unlock(&ctx->uring_lock);
7219 ret = io_run_task_work_sig();
7220 mutex_lock(&ctx->uring_lock);
7222 data->quiesce = false;
7227 static u64 *io_get_tag_slot(struct io_rsrc_data *data, unsigned int idx)
7229 unsigned int off = idx & IO_RSRC_TAG_TABLE_MASK;
7230 unsigned int table_idx = idx >> IO_RSRC_TAG_TABLE_SHIFT;
7232 return &data->tags[table_idx][off];
7235 static void io_rsrc_data_free(struct io_rsrc_data *data)
7237 size_t size = data->nr * sizeof(data->tags[0][0]);
7240 io_free_page_table((void **)data->tags, size);
7244 static int io_rsrc_data_alloc(struct io_ring_ctx *ctx, rsrc_put_fn *do_put,
7245 u64 __user *utags, unsigned nr,
7246 struct io_rsrc_data **pdata)
7248 struct io_rsrc_data *data;
7252 data = kzalloc(sizeof(*data), GFP_KERNEL);
7255 data->tags = (u64 **)io_alloc_page_table(nr * sizeof(data->tags[0][0]));
7263 data->do_put = do_put;
7266 for (i = 0; i < nr; i++) {
7267 u64 *tag_slot = io_get_tag_slot(data, i);
7269 if (copy_from_user(tag_slot, &utags[i],
7275 atomic_set(&data->refs, 1);
7276 init_completion(&data->done);
7280 io_rsrc_data_free(data);
7284 static bool io_alloc_file_tables(struct io_file_table *table, unsigned nr_files)
7286 size_t size = nr_files * sizeof(struct io_fixed_file);
7288 table->files = (struct io_fixed_file **)io_alloc_page_table(size);
7289 return !!table->files;
7292 static void io_free_file_tables(struct io_file_table *table, unsigned nr_files)
7294 size_t size = nr_files * sizeof(struct io_fixed_file);
7296 io_free_page_table((void **)table->files, size);
7297 table->files = NULL;
7300 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
7302 #if defined(CONFIG_UNIX)
7303 if (ctx->ring_sock) {
7304 struct sock *sock = ctx->ring_sock->sk;
7305 struct sk_buff *skb;
7307 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
7313 for (i = 0; i < ctx->nr_user_files; i++) {
7316 file = io_file_from_index(ctx, i);
7321 io_free_file_tables(&ctx->file_table, ctx->nr_user_files);
7322 io_rsrc_data_free(ctx->file_data);
7323 ctx->file_data = NULL;
7324 ctx->nr_user_files = 0;
7327 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7331 if (!ctx->file_data)
7333 ret = io_rsrc_ref_quiesce(ctx->file_data, ctx);
7335 __io_sqe_files_unregister(ctx);
7339 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7340 __releases(&sqd->lock)
7342 WARN_ON_ONCE(sqd->thread == current);
7345 * Do the dance but not conditional clear_bit() because it'd race with
7346 * other threads incrementing park_pending and setting the bit.
7348 clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7349 if (atomic_dec_return(&sqd->park_pending))
7350 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7351 mutex_unlock(&sqd->lock);
7354 static void io_sq_thread_park(struct io_sq_data *sqd)
7355 __acquires(&sqd->lock)
7357 WARN_ON_ONCE(sqd->thread == current);
7359 atomic_inc(&sqd->park_pending);
7360 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7361 mutex_lock(&sqd->lock);
7363 wake_up_process(sqd->thread);
7366 static void io_sq_thread_stop(struct io_sq_data *sqd)
7368 WARN_ON_ONCE(sqd->thread == current);
7369 WARN_ON_ONCE(test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state));
7371 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7372 mutex_lock(&sqd->lock);
7374 wake_up_process(sqd->thread);
7375 mutex_unlock(&sqd->lock);
7376 wait_for_completion(&sqd->exited);
7379 static void io_put_sq_data(struct io_sq_data *sqd)
7381 if (refcount_dec_and_test(&sqd->refs)) {
7382 WARN_ON_ONCE(atomic_read(&sqd->park_pending));
7384 io_sq_thread_stop(sqd);
7389 static void io_sq_thread_finish(struct io_ring_ctx *ctx)
7391 struct io_sq_data *sqd = ctx->sq_data;
7394 io_sq_thread_park(sqd);
7395 list_del_init(&ctx->sqd_list);
7396 io_sqd_update_thread_idle(sqd);
7397 io_sq_thread_unpark(sqd);
7399 io_put_sq_data(sqd);
7400 ctx->sq_data = NULL;
7404 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7406 struct io_ring_ctx *ctx_attach;
7407 struct io_sq_data *sqd;
7410 f = fdget(p->wq_fd);
7412 return ERR_PTR(-ENXIO);
7413 if (f.file->f_op != &io_uring_fops) {
7415 return ERR_PTR(-EINVAL);
7418 ctx_attach = f.file->private_data;
7419 sqd = ctx_attach->sq_data;
7422 return ERR_PTR(-EINVAL);
7424 if (sqd->task_tgid != current->tgid) {
7426 return ERR_PTR(-EPERM);
7429 refcount_inc(&sqd->refs);
7434 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p,
7437 struct io_sq_data *sqd;
7440 if (p->flags & IORING_SETUP_ATTACH_WQ) {
7441 sqd = io_attach_sq_data(p);
7446 /* fall through for EPERM case, setup new sqd/task */
7447 if (PTR_ERR(sqd) != -EPERM)
7451 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7453 return ERR_PTR(-ENOMEM);
7455 atomic_set(&sqd->park_pending, 0);
7456 refcount_set(&sqd->refs, 1);
7457 INIT_LIST_HEAD(&sqd->ctx_list);
7458 mutex_init(&sqd->lock);
7459 init_waitqueue_head(&sqd->wait);
7460 init_completion(&sqd->exited);
7464 #if defined(CONFIG_UNIX)
7466 * Ensure the UNIX gc is aware of our file set, so we are certain that
7467 * the io_uring can be safely unregistered on process exit, even if we have
7468 * loops in the file referencing.
7470 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7472 struct sock *sk = ctx->ring_sock->sk;
7473 struct scm_fp_list *fpl;
7474 struct sk_buff *skb;
7477 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7481 skb = alloc_skb(0, GFP_KERNEL);
7490 fpl->user = get_uid(current_user());
7491 for (i = 0; i < nr; i++) {
7492 struct file *file = io_file_from_index(ctx, i + offset);
7496 fpl->fp[nr_files] = get_file(file);
7497 unix_inflight(fpl->user, fpl->fp[nr_files]);
7502 fpl->max = SCM_MAX_FD;
7503 fpl->count = nr_files;
7504 UNIXCB(skb).fp = fpl;
7505 skb->destructor = unix_destruct_scm;
7506 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7507 skb_queue_head(&sk->sk_receive_queue, skb);
7509 for (i = 0; i < nr_files; i++)
7520 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7521 * causes regular reference counting to break down. We rely on the UNIX
7522 * garbage collection to take care of this problem for us.
7524 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7526 unsigned left, total;
7530 left = ctx->nr_user_files;
7532 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7534 ret = __io_sqe_files_scm(ctx, this_files, total);
7538 total += this_files;
7544 while (total < ctx->nr_user_files) {
7545 struct file *file = io_file_from_index(ctx, total);
7555 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7561 static void io_rsrc_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
7563 struct file *file = prsrc->file;
7564 #if defined(CONFIG_UNIX)
7565 struct sock *sock = ctx->ring_sock->sk;
7566 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7567 struct sk_buff *skb;
7570 __skb_queue_head_init(&list);
7573 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7574 * remove this entry and rearrange the file array.
7576 skb = skb_dequeue(head);
7578 struct scm_fp_list *fp;
7580 fp = UNIXCB(skb).fp;
7581 for (i = 0; i < fp->count; i++) {
7584 if (fp->fp[i] != file)
7587 unix_notinflight(fp->user, fp->fp[i]);
7588 left = fp->count - 1 - i;
7590 memmove(&fp->fp[i], &fp->fp[i + 1],
7591 left * sizeof(struct file *));
7598 __skb_queue_tail(&list, skb);
7608 __skb_queue_tail(&list, skb);
7610 skb = skb_dequeue(head);
7613 if (skb_peek(&list)) {
7614 spin_lock_irq(&head->lock);
7615 while ((skb = __skb_dequeue(&list)) != NULL)
7616 __skb_queue_tail(head, skb);
7617 spin_unlock_irq(&head->lock);
7624 static void __io_rsrc_put_work(struct io_rsrc_node *ref_node)
7626 struct io_rsrc_data *rsrc_data = ref_node->rsrc_data;
7627 struct io_ring_ctx *ctx = rsrc_data->ctx;
7628 struct io_rsrc_put *prsrc, *tmp;
7630 list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
7631 list_del(&prsrc->list);
7634 bool lock_ring = ctx->flags & IORING_SETUP_IOPOLL;
7636 io_ring_submit_lock(ctx, lock_ring);
7637 spin_lock_irq(&ctx->completion_lock);
7638 io_cqring_fill_event(ctx, prsrc->tag, 0, 0);
7640 io_commit_cqring(ctx);
7641 spin_unlock_irq(&ctx->completion_lock);
7642 io_cqring_ev_posted(ctx);
7643 io_ring_submit_unlock(ctx, lock_ring);
7646 rsrc_data->do_put(ctx, prsrc);
7650 io_rsrc_node_destroy(ref_node);
7651 if (atomic_dec_and_test(&rsrc_data->refs))
7652 complete(&rsrc_data->done);
7655 static void io_rsrc_put_work(struct work_struct *work)
7657 struct io_ring_ctx *ctx;
7658 struct llist_node *node;
7660 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
7661 node = llist_del_all(&ctx->rsrc_put_llist);
7664 struct io_rsrc_node *ref_node;
7665 struct llist_node *next = node->next;
7667 ref_node = llist_entry(node, struct io_rsrc_node, llist);
7668 __io_rsrc_put_work(ref_node);
7673 static void io_rsrc_node_ref_zero(struct percpu_ref *ref)
7675 struct io_rsrc_node *node = container_of(ref, struct io_rsrc_node, refs);
7676 struct io_ring_ctx *ctx = node->rsrc_data->ctx;
7677 bool first_add = false;
7679 io_rsrc_ref_lock(ctx);
7682 while (!list_empty(&ctx->rsrc_ref_list)) {
7683 node = list_first_entry(&ctx->rsrc_ref_list,
7684 struct io_rsrc_node, node);
7685 /* recycle ref nodes in order */
7688 list_del(&node->node);
7689 first_add |= llist_add(&node->llist, &ctx->rsrc_put_llist);
7691 io_rsrc_ref_unlock(ctx);
7694 mod_delayed_work(system_wq, &ctx->rsrc_put_work, HZ);
7697 static struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx)
7699 struct io_rsrc_node *ref_node;
7701 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7705 if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
7710 INIT_LIST_HEAD(&ref_node->node);
7711 INIT_LIST_HEAD(&ref_node->rsrc_list);
7712 ref_node->done = false;
7716 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7717 unsigned nr_args, u64 __user *tags)
7719 __s32 __user *fds = (__s32 __user *) arg;
7728 if (nr_args > IORING_MAX_FIXED_FILES)
7730 ret = io_rsrc_node_switch_start(ctx);
7733 ret = io_rsrc_data_alloc(ctx, io_rsrc_file_put, tags, nr_args,
7739 if (!io_alloc_file_tables(&ctx->file_table, nr_args))
7742 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7743 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7747 /* allow sparse sets */
7750 if (unlikely(*io_get_tag_slot(ctx->file_data, i)))
7757 if (unlikely(!file))
7761 * Don't allow io_uring instances to be registered. If UNIX
7762 * isn't enabled, then this causes a reference cycle and this
7763 * instance can never get freed. If UNIX is enabled we'll
7764 * handle it just fine, but there's still no point in allowing
7765 * a ring fd as it doesn't support regular read/write anyway.
7767 if (file->f_op == &io_uring_fops) {
7771 io_fixed_file_set(io_fixed_file_slot(&ctx->file_table, i), file);
7774 ret = io_sqe_files_scm(ctx);
7776 __io_sqe_files_unregister(ctx);
7780 io_rsrc_node_switch(ctx, NULL);
7783 for (i = 0; i < ctx->nr_user_files; i++) {
7784 file = io_file_from_index(ctx, i);
7788 io_free_file_tables(&ctx->file_table, nr_args);
7789 ctx->nr_user_files = 0;
7791 io_rsrc_data_free(ctx->file_data);
7792 ctx->file_data = NULL;
7796 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7799 #if defined(CONFIG_UNIX)
7800 struct sock *sock = ctx->ring_sock->sk;
7801 struct sk_buff_head *head = &sock->sk_receive_queue;
7802 struct sk_buff *skb;
7805 * See if we can merge this file into an existing skb SCM_RIGHTS
7806 * file set. If there's no room, fall back to allocating a new skb
7807 * and filling it in.
7809 spin_lock_irq(&head->lock);
7810 skb = skb_peek(head);
7812 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7814 if (fpl->count < SCM_MAX_FD) {
7815 __skb_unlink(skb, head);
7816 spin_unlock_irq(&head->lock);
7817 fpl->fp[fpl->count] = get_file(file);
7818 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7820 spin_lock_irq(&head->lock);
7821 __skb_queue_head(head, skb);
7826 spin_unlock_irq(&head->lock);
7833 return __io_sqe_files_scm(ctx, 1, index);
7839 static int io_queue_rsrc_removal(struct io_rsrc_data *data, unsigned idx,
7840 struct io_rsrc_node *node, void *rsrc)
7842 struct io_rsrc_put *prsrc;
7844 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
7848 prsrc->tag = *io_get_tag_slot(data, idx);
7850 list_add(&prsrc->list, &node->rsrc_list);
7854 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7855 struct io_uring_rsrc_update2 *up,
7858 u64 __user *tags = u64_to_user_ptr(up->tags);
7859 __s32 __user *fds = u64_to_user_ptr(up->data);
7860 struct io_rsrc_data *data = ctx->file_data;
7861 struct io_fixed_file *file_slot;
7865 bool needs_switch = false;
7867 if (!ctx->file_data)
7869 if (up->offset + nr_args > ctx->nr_user_files)
7872 for (done = 0; done < nr_args; done++) {
7875 if ((tags && copy_from_user(&tag, &tags[done], sizeof(tag))) ||
7876 copy_from_user(&fd, &fds[done], sizeof(fd))) {
7880 if ((fd == IORING_REGISTER_FILES_SKIP || fd == -1) && tag) {
7884 if (fd == IORING_REGISTER_FILES_SKIP)
7887 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
7888 file_slot = io_fixed_file_slot(&ctx->file_table, i);
7890 if (file_slot->file_ptr) {
7891 file = (struct file *)(file_slot->file_ptr & FFS_MASK);
7892 err = io_queue_rsrc_removal(data, up->offset + done,
7893 ctx->rsrc_node, file);
7896 file_slot->file_ptr = 0;
7897 needs_switch = true;
7906 * Don't allow io_uring instances to be registered. If
7907 * UNIX isn't enabled, then this causes a reference
7908 * cycle and this instance can never get freed. If UNIX
7909 * is enabled we'll handle it just fine, but there's
7910 * still no point in allowing a ring fd as it doesn't
7911 * support regular read/write anyway.
7913 if (file->f_op == &io_uring_fops) {
7918 *io_get_tag_slot(data, up->offset + done) = tag;
7919 io_fixed_file_set(file_slot, file);
7920 err = io_sqe_file_register(ctx, file, i);
7922 file_slot->file_ptr = 0;
7930 io_rsrc_node_switch(ctx, data);
7931 return done ? done : err;
7934 static struct io_wq_work *io_free_work(struct io_wq_work *work)
7936 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7938 req = io_put_req_find_next(req);
7939 return req ? &req->work : NULL;
7942 static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx,
7943 struct task_struct *task)
7945 struct io_wq_hash *hash;
7946 struct io_wq_data data;
7947 unsigned int concurrency;
7949 hash = ctx->hash_map;
7951 hash = kzalloc(sizeof(*hash), GFP_KERNEL);
7953 return ERR_PTR(-ENOMEM);
7954 refcount_set(&hash->refs, 1);
7955 init_waitqueue_head(&hash->wait);
7956 ctx->hash_map = hash;
7961 data.free_work = io_free_work;
7962 data.do_work = io_wq_submit_work;
7964 /* Do QD, or 4 * CPUS, whatever is smallest */
7965 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7967 return io_wq_create(concurrency, &data);
7970 static int io_uring_alloc_task_context(struct task_struct *task,
7971 struct io_ring_ctx *ctx)
7973 struct io_uring_task *tctx;
7976 tctx = kzalloc(sizeof(*tctx), GFP_KERNEL);
7977 if (unlikely(!tctx))
7980 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
7981 if (unlikely(ret)) {
7986 tctx->io_wq = io_init_wq_offload(ctx, task);
7987 if (IS_ERR(tctx->io_wq)) {
7988 ret = PTR_ERR(tctx->io_wq);
7989 percpu_counter_destroy(&tctx->inflight);
7995 init_waitqueue_head(&tctx->wait);
7996 atomic_set(&tctx->in_idle, 0);
7997 atomic_set(&tctx->inflight_tracked, 0);
7998 task->io_uring = tctx;
7999 spin_lock_init(&tctx->task_lock);
8000 INIT_WQ_LIST(&tctx->task_list);
8001 init_task_work(&tctx->task_work, tctx_task_work);
8005 void __io_uring_free(struct task_struct *tsk)
8007 struct io_uring_task *tctx = tsk->io_uring;
8009 WARN_ON_ONCE(!xa_empty(&tctx->xa));
8010 WARN_ON_ONCE(tctx->io_wq);
8011 WARN_ON_ONCE(tctx->cached_refs);
8013 percpu_counter_destroy(&tctx->inflight);
8015 tsk->io_uring = NULL;
8018 static int io_sq_offload_create(struct io_ring_ctx *ctx,
8019 struct io_uring_params *p)
8023 /* Retain compatibility with failing for an invalid attach attempt */
8024 if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
8025 IORING_SETUP_ATTACH_WQ) {
8028 f = fdget(p->wq_fd);
8032 if (f.file->f_op != &io_uring_fops)
8035 if (ctx->flags & IORING_SETUP_SQPOLL) {
8036 struct task_struct *tsk;
8037 struct io_sq_data *sqd;
8040 sqd = io_get_sq_data(p, &attached);
8046 ctx->sq_creds = get_current_cred();
8048 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
8049 if (!ctx->sq_thread_idle)
8050 ctx->sq_thread_idle = HZ;
8052 io_sq_thread_park(sqd);
8053 list_add(&ctx->sqd_list, &sqd->ctx_list);
8054 io_sqd_update_thread_idle(sqd);
8055 /* don't attach to a dying SQPOLL thread, would be racy */
8056 ret = (attached && !sqd->thread) ? -ENXIO : 0;
8057 io_sq_thread_unpark(sqd);
8064 if (p->flags & IORING_SETUP_SQ_AFF) {
8065 int cpu = p->sq_thread_cpu;
8068 if (cpu >= nr_cpu_ids || !cpu_online(cpu))
8075 sqd->task_pid = current->pid;
8076 sqd->task_tgid = current->tgid;
8077 tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE);
8084 ret = io_uring_alloc_task_context(tsk, ctx);
8085 wake_up_new_task(tsk);
8088 } else if (p->flags & IORING_SETUP_SQ_AFF) {
8089 /* Can't have SQ_AFF without SQPOLL */
8096 complete(&ctx->sq_data->exited);
8098 io_sq_thread_finish(ctx);
8102 static inline void __io_unaccount_mem(struct user_struct *user,
8103 unsigned long nr_pages)
8105 atomic_long_sub(nr_pages, &user->locked_vm);
8108 static inline int __io_account_mem(struct user_struct *user,
8109 unsigned long nr_pages)
8111 unsigned long page_limit, cur_pages, new_pages;
8113 /* Don't allow more pages than we can safely lock */
8114 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
8117 cur_pages = atomic_long_read(&user->locked_vm);
8118 new_pages = cur_pages + nr_pages;
8119 if (new_pages > page_limit)
8121 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
8122 new_pages) != cur_pages);
8127 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8130 __io_unaccount_mem(ctx->user, nr_pages);
8132 if (ctx->mm_account)
8133 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
8136 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8141 ret = __io_account_mem(ctx->user, nr_pages);
8146 if (ctx->mm_account)
8147 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
8152 static void io_mem_free(void *ptr)
8159 page = virt_to_head_page(ptr);
8160 if (put_page_testzero(page))
8161 free_compound_page(page);
8164 static void *io_mem_alloc(size_t size)
8166 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
8167 __GFP_NORETRY | __GFP_ACCOUNT;
8169 return (void *) __get_free_pages(gfp_flags, get_order(size));
8172 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8175 struct io_rings *rings;
8176 size_t off, sq_array_size;
8178 off = struct_size(rings, cqes, cq_entries);
8179 if (off == SIZE_MAX)
8183 off = ALIGN(off, SMP_CACHE_BYTES);
8191 sq_array_size = array_size(sizeof(u32), sq_entries);
8192 if (sq_array_size == SIZE_MAX)
8195 if (check_add_overflow(off, sq_array_size, &off))
8201 static void io_buffer_unmap(struct io_ring_ctx *ctx, struct io_mapped_ubuf **slot)
8203 struct io_mapped_ubuf *imu = *slot;
8206 if (imu != ctx->dummy_ubuf) {
8207 for (i = 0; i < imu->nr_bvecs; i++)
8208 unpin_user_page(imu->bvec[i].bv_page);
8209 if (imu->acct_pages)
8210 io_unaccount_mem(ctx, imu->acct_pages);
8216 static void io_rsrc_buf_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
8218 io_buffer_unmap(ctx, &prsrc->buf);
8222 static void __io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8226 for (i = 0; i < ctx->nr_user_bufs; i++)
8227 io_buffer_unmap(ctx, &ctx->user_bufs[i]);
8228 kfree(ctx->user_bufs);
8229 io_rsrc_data_free(ctx->buf_data);
8230 ctx->user_bufs = NULL;
8231 ctx->buf_data = NULL;
8232 ctx->nr_user_bufs = 0;
8235 static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8242 ret = io_rsrc_ref_quiesce(ctx->buf_data, ctx);
8244 __io_sqe_buffers_unregister(ctx);
8248 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8249 void __user *arg, unsigned index)
8251 struct iovec __user *src;
8253 #ifdef CONFIG_COMPAT
8255 struct compat_iovec __user *ciovs;
8256 struct compat_iovec ciov;
8258 ciovs = (struct compat_iovec __user *) arg;
8259 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8262 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8263 dst->iov_len = ciov.iov_len;
8267 src = (struct iovec __user *) arg;
8268 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8274 * Not super efficient, but this is just a registration time. And we do cache
8275 * the last compound head, so generally we'll only do a full search if we don't
8278 * We check if the given compound head page has already been accounted, to
8279 * avoid double accounting it. This allows us to account the full size of the
8280 * page, not just the constituent pages of a huge page.
8282 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8283 int nr_pages, struct page *hpage)
8287 /* check current page array */
8288 for (i = 0; i < nr_pages; i++) {
8289 if (!PageCompound(pages[i]))
8291 if (compound_head(pages[i]) == hpage)
8295 /* check previously registered pages */
8296 for (i = 0; i < ctx->nr_user_bufs; i++) {
8297 struct io_mapped_ubuf *imu = ctx->user_bufs[i];
8299 for (j = 0; j < imu->nr_bvecs; j++) {
8300 if (!PageCompound(imu->bvec[j].bv_page))
8302 if (compound_head(imu->bvec[j].bv_page) == hpage)
8310 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8311 int nr_pages, struct io_mapped_ubuf *imu,
8312 struct page **last_hpage)
8316 imu->acct_pages = 0;
8317 for (i = 0; i < nr_pages; i++) {
8318 if (!PageCompound(pages[i])) {
8323 hpage = compound_head(pages[i]);
8324 if (hpage == *last_hpage)
8326 *last_hpage = hpage;
8327 if (headpage_already_acct(ctx, pages, i, hpage))
8329 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8333 if (!imu->acct_pages)
8336 ret = io_account_mem(ctx, imu->acct_pages);
8338 imu->acct_pages = 0;
8342 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
8343 struct io_mapped_ubuf **pimu,
8344 struct page **last_hpage)
8346 struct io_mapped_ubuf *imu = NULL;
8347 struct vm_area_struct **vmas = NULL;
8348 struct page **pages = NULL;
8349 unsigned long off, start, end, ubuf;
8351 int ret, pret, nr_pages, i;
8353 if (!iov->iov_base) {
8354 *pimu = ctx->dummy_ubuf;
8358 ubuf = (unsigned long) iov->iov_base;
8359 end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8360 start = ubuf >> PAGE_SHIFT;
8361 nr_pages = end - start;
8366 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
8370 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
8375 imu = kvmalloc(struct_size(imu, bvec, nr_pages), GFP_KERNEL);
8380 mmap_read_lock(current->mm);
8381 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
8383 if (pret == nr_pages) {
8384 /* don't support file backed memory */
8385 for (i = 0; i < nr_pages; i++) {
8386 struct vm_area_struct *vma = vmas[i];
8388 if (vma_is_shmem(vma))
8391 !is_file_hugepages(vma->vm_file)) {
8397 ret = pret < 0 ? pret : -EFAULT;
8399 mmap_read_unlock(current->mm);
8402 * if we did partial map, or found file backed vmas,
8403 * release any pages we did get
8406 unpin_user_pages(pages, pret);
8410 ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
8412 unpin_user_pages(pages, pret);
8416 off = ubuf & ~PAGE_MASK;
8417 size = iov->iov_len;
8418 for (i = 0; i < nr_pages; i++) {
8421 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8422 imu->bvec[i].bv_page = pages[i];
8423 imu->bvec[i].bv_len = vec_len;
8424 imu->bvec[i].bv_offset = off;
8428 /* store original address for later verification */
8430 imu->ubuf_end = ubuf + iov->iov_len;
8431 imu->nr_bvecs = nr_pages;
8442 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
8444 ctx->user_bufs = kcalloc(nr_args, sizeof(*ctx->user_bufs), GFP_KERNEL);
8445 return ctx->user_bufs ? 0 : -ENOMEM;
8448 static int io_buffer_validate(struct iovec *iov)
8450 unsigned long tmp, acct_len = iov->iov_len + (PAGE_SIZE - 1);
8453 * Don't impose further limits on the size and buffer
8454 * constraints here, we'll -EINVAL later when IO is
8455 * submitted if they are wrong.
8458 return iov->iov_len ? -EFAULT : 0;
8462 /* arbitrary limit, but we need something */
8463 if (iov->iov_len > SZ_1G)
8466 if (check_add_overflow((unsigned long)iov->iov_base, acct_len, &tmp))
8472 static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
8473 unsigned int nr_args, u64 __user *tags)
8475 struct page *last_hpage = NULL;
8476 struct io_rsrc_data *data;
8482 if (!nr_args || nr_args > IORING_MAX_REG_BUFFERS)
8484 ret = io_rsrc_node_switch_start(ctx);
8487 ret = io_rsrc_data_alloc(ctx, io_rsrc_buf_put, tags, nr_args, &data);
8490 ret = io_buffers_map_alloc(ctx, nr_args);
8492 io_rsrc_data_free(data);
8496 for (i = 0; i < nr_args; i++, ctx->nr_user_bufs++) {
8497 ret = io_copy_iov(ctx, &iov, arg, i);
8500 ret = io_buffer_validate(&iov);
8503 if (!iov.iov_base && *io_get_tag_slot(data, i)) {
8508 ret = io_sqe_buffer_register(ctx, &iov, &ctx->user_bufs[i],
8514 WARN_ON_ONCE(ctx->buf_data);
8516 ctx->buf_data = data;
8518 __io_sqe_buffers_unregister(ctx);
8520 io_rsrc_node_switch(ctx, NULL);
8524 static int __io_sqe_buffers_update(struct io_ring_ctx *ctx,
8525 struct io_uring_rsrc_update2 *up,
8526 unsigned int nr_args)
8528 u64 __user *tags = u64_to_user_ptr(up->tags);
8529 struct iovec iov, __user *iovs = u64_to_user_ptr(up->data);
8530 struct page *last_hpage = NULL;
8531 bool needs_switch = false;
8537 if (up->offset + nr_args > ctx->nr_user_bufs)
8540 for (done = 0; done < nr_args; done++) {
8541 struct io_mapped_ubuf *imu;
8542 int offset = up->offset + done;
8545 err = io_copy_iov(ctx, &iov, iovs, done);
8548 if (tags && copy_from_user(&tag, &tags[done], sizeof(tag))) {
8552 err = io_buffer_validate(&iov);
8555 if (!iov.iov_base && tag) {
8559 err = io_sqe_buffer_register(ctx, &iov, &imu, &last_hpage);
8563 i = array_index_nospec(offset, ctx->nr_user_bufs);
8564 if (ctx->user_bufs[i] != ctx->dummy_ubuf) {
8565 err = io_queue_rsrc_removal(ctx->buf_data, offset,
8566 ctx->rsrc_node, ctx->user_bufs[i]);
8567 if (unlikely(err)) {
8568 io_buffer_unmap(ctx, &imu);
8571 ctx->user_bufs[i] = NULL;
8572 needs_switch = true;
8575 ctx->user_bufs[i] = imu;
8576 *io_get_tag_slot(ctx->buf_data, offset) = tag;
8580 io_rsrc_node_switch(ctx, ctx->buf_data);
8581 return done ? done : err;
8584 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8586 __s32 __user *fds = arg;
8592 if (copy_from_user(&fd, fds, sizeof(*fds)))
8595 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8596 if (IS_ERR(ctx->cq_ev_fd)) {
8597 int ret = PTR_ERR(ctx->cq_ev_fd);
8599 ctx->cq_ev_fd = NULL;
8606 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8608 if (ctx->cq_ev_fd) {
8609 eventfd_ctx_put(ctx->cq_ev_fd);
8610 ctx->cq_ev_fd = NULL;
8617 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8619 struct io_buffer *buf;
8620 unsigned long index;
8622 xa_for_each(&ctx->io_buffers, index, buf)
8623 __io_remove_buffers(ctx, buf, index, -1U);
8626 static void io_req_cache_free(struct list_head *list, struct task_struct *tsk)
8628 struct io_kiocb *req, *nxt;
8630 list_for_each_entry_safe(req, nxt, list, compl.list) {
8631 if (tsk && req->task != tsk)
8633 list_del(&req->compl.list);
8634 kmem_cache_free(req_cachep, req);
8638 static void io_req_caches_free(struct io_ring_ctx *ctx)
8640 struct io_submit_state *submit_state = &ctx->submit_state;
8641 struct io_comp_state *cs = &ctx->submit_state.comp;
8643 mutex_lock(&ctx->uring_lock);
8645 if (submit_state->free_reqs) {
8646 kmem_cache_free_bulk(req_cachep, submit_state->free_reqs,
8647 submit_state->reqs);
8648 submit_state->free_reqs = 0;
8651 io_flush_cached_locked_reqs(ctx, cs);
8652 io_req_cache_free(&cs->free_list, NULL);
8653 mutex_unlock(&ctx->uring_lock);
8656 static bool io_wait_rsrc_data(struct io_rsrc_data *data)
8660 if (!atomic_dec_and_test(&data->refs))
8661 wait_for_completion(&data->done);
8665 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8667 io_sq_thread_finish(ctx);
8669 if (ctx->mm_account) {
8670 mmdrop(ctx->mm_account);
8671 ctx->mm_account = NULL;
8674 mutex_lock(&ctx->uring_lock);
8675 if (io_wait_rsrc_data(ctx->buf_data))
8676 __io_sqe_buffers_unregister(ctx);
8677 if (io_wait_rsrc_data(ctx->file_data))
8678 __io_sqe_files_unregister(ctx);
8680 __io_cqring_overflow_flush(ctx, true);
8681 mutex_unlock(&ctx->uring_lock);
8682 io_eventfd_unregister(ctx);
8683 io_destroy_buffers(ctx);
8685 put_cred(ctx->sq_creds);
8687 /* there are no registered resources left, nobody uses it */
8689 io_rsrc_node_destroy(ctx->rsrc_node);
8690 if (ctx->rsrc_backup_node)
8691 io_rsrc_node_destroy(ctx->rsrc_backup_node);
8692 flush_delayed_work(&ctx->rsrc_put_work);
8694 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
8695 WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist));
8697 #if defined(CONFIG_UNIX)
8698 if (ctx->ring_sock) {
8699 ctx->ring_sock->file = NULL; /* so that iput() is called */
8700 sock_release(ctx->ring_sock);
8704 io_mem_free(ctx->rings);
8705 io_mem_free(ctx->sq_sqes);
8707 percpu_ref_exit(&ctx->refs);
8708 free_uid(ctx->user);
8709 io_req_caches_free(ctx);
8711 io_wq_put_hash(ctx->hash_map);
8712 kfree(ctx->cancel_hash);
8713 kfree(ctx->dummy_ubuf);
8717 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8719 struct io_ring_ctx *ctx = file->private_data;
8722 poll_wait(file, &ctx->poll_wait, wait);
8724 * synchronizes with barrier from wq_has_sleeper call in
8728 if (!io_sqring_full(ctx))
8729 mask |= EPOLLOUT | EPOLLWRNORM;
8732 * Don't flush cqring overflow list here, just do a simple check.
8733 * Otherwise there could possible be ABBA deadlock:
8736 * lock(&ctx->uring_lock);
8738 * lock(&ctx->uring_lock);
8741 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
8742 * pushs them to do the flush.
8744 if (io_cqring_events(ctx) || test_bit(0, &ctx->check_cq_overflow))
8745 mask |= EPOLLIN | EPOLLRDNORM;
8750 static int io_uring_fasync(int fd, struct file *file, int on)
8752 struct io_ring_ctx *ctx = file->private_data;
8754 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8757 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8759 const struct cred *creds;
8761 creds = xa_erase(&ctx->personalities, id);
8770 static inline bool io_run_ctx_fallback(struct io_ring_ctx *ctx)
8772 return io_run_task_work_head(&ctx->exit_task_work);
8775 struct io_tctx_exit {
8776 struct callback_head task_work;
8777 struct completion completion;
8778 struct io_ring_ctx *ctx;
8781 static void io_tctx_exit_cb(struct callback_head *cb)
8783 struct io_uring_task *tctx = current->io_uring;
8784 struct io_tctx_exit *work;
8786 work = container_of(cb, struct io_tctx_exit, task_work);
8788 * When @in_idle, we're in cancellation and it's racy to remove the
8789 * node. It'll be removed by the end of cancellation, just ignore it.
8791 if (!atomic_read(&tctx->in_idle))
8792 io_uring_del_tctx_node((unsigned long)work->ctx);
8793 complete(&work->completion);
8796 static bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
8798 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8800 return req->ctx == data;
8803 static void io_ring_exit_work(struct work_struct *work)
8805 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
8806 unsigned long timeout = jiffies + HZ * 60 * 5;
8807 struct io_tctx_exit exit;
8808 struct io_tctx_node *node;
8812 * If we're doing polled IO and end up having requests being
8813 * submitted async (out-of-line), then completions can come in while
8814 * we're waiting for refs to drop. We need to reap these manually,
8815 * as nobody else will be looking for them.
8818 io_uring_try_cancel_requests(ctx, NULL, true);
8820 struct io_sq_data *sqd = ctx->sq_data;
8821 struct task_struct *tsk;
8823 io_sq_thread_park(sqd);
8825 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
8826 io_wq_cancel_cb(tsk->io_uring->io_wq,
8827 io_cancel_ctx_cb, ctx, true);
8828 io_sq_thread_unpark(sqd);
8831 WARN_ON_ONCE(time_after(jiffies, timeout));
8832 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
8834 init_completion(&exit.completion);
8835 init_task_work(&exit.task_work, io_tctx_exit_cb);
8838 * Some may use context even when all refs and requests have been put,
8839 * and they are free to do so while still holding uring_lock or
8840 * completion_lock, see __io_req_task_submit(). Apart from other work,
8841 * this lock/unlock section also waits them to finish.
8843 mutex_lock(&ctx->uring_lock);
8844 while (!list_empty(&ctx->tctx_list)) {
8845 WARN_ON_ONCE(time_after(jiffies, timeout));
8847 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
8849 /* don't spin on a single task if cancellation failed */
8850 list_rotate_left(&ctx->tctx_list);
8851 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
8852 if (WARN_ON_ONCE(ret))
8854 wake_up_process(node->task);
8856 mutex_unlock(&ctx->uring_lock);
8857 wait_for_completion(&exit.completion);
8858 mutex_lock(&ctx->uring_lock);
8860 mutex_unlock(&ctx->uring_lock);
8861 spin_lock_irq(&ctx->completion_lock);
8862 spin_unlock_irq(&ctx->completion_lock);
8864 io_ring_ctx_free(ctx);
8867 /* Returns true if we found and killed one or more timeouts */
8868 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
8871 struct io_kiocb *req, *tmp;
8874 spin_lock_irq(&ctx->completion_lock);
8875 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
8876 if (io_match_task(req, tsk, cancel_all)) {
8877 io_kill_timeout(req, -ECANCELED);
8882 io_commit_cqring(ctx);
8883 spin_unlock_irq(&ctx->completion_lock);
8885 io_cqring_ev_posted(ctx);
8886 return canceled != 0;
8889 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8891 unsigned long index;
8892 struct creds *creds;
8894 mutex_lock(&ctx->uring_lock);
8895 percpu_ref_kill(&ctx->refs);
8897 __io_cqring_overflow_flush(ctx, true);
8898 xa_for_each(&ctx->personalities, index, creds)
8899 io_unregister_personality(ctx, index);
8900 mutex_unlock(&ctx->uring_lock);
8902 io_kill_timeouts(ctx, NULL, true);
8903 io_poll_remove_all(ctx, NULL, true);
8905 /* if we failed setting up the ctx, we might not have any rings */
8906 io_iopoll_try_reap_events(ctx);
8908 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8910 * Use system_unbound_wq to avoid spawning tons of event kworkers
8911 * if we're exiting a ton of rings at the same time. It just adds
8912 * noise and overhead, there's no discernable change in runtime
8913 * over using system_wq.
8915 queue_work(system_unbound_wq, &ctx->exit_work);
8918 static int io_uring_release(struct inode *inode, struct file *file)
8920 struct io_ring_ctx *ctx = file->private_data;
8922 file->private_data = NULL;
8923 io_ring_ctx_wait_and_kill(ctx);
8927 struct io_task_cancel {
8928 struct task_struct *task;
8932 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8934 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8935 struct io_task_cancel *cancel = data;
8938 if (!cancel->all && (req->flags & REQ_F_LINK_TIMEOUT)) {
8939 unsigned long flags;
8940 struct io_ring_ctx *ctx = req->ctx;
8942 /* protect against races with linked timeouts */
8943 spin_lock_irqsave(&ctx->completion_lock, flags);
8944 ret = io_match_task(req, cancel->task, cancel->all);
8945 spin_unlock_irqrestore(&ctx->completion_lock, flags);
8947 ret = io_match_task(req, cancel->task, cancel->all);
8952 static bool io_cancel_defer_files(struct io_ring_ctx *ctx,
8953 struct task_struct *task, bool cancel_all)
8955 struct io_defer_entry *de;
8958 spin_lock_irq(&ctx->completion_lock);
8959 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8960 if (io_match_task(de->req, task, cancel_all)) {
8961 list_cut_position(&list, &ctx->defer_list, &de->list);
8965 spin_unlock_irq(&ctx->completion_lock);
8966 if (list_empty(&list))
8969 while (!list_empty(&list)) {
8970 de = list_first_entry(&list, struct io_defer_entry, list);
8971 list_del_init(&de->list);
8972 io_req_complete_failed(de->req, -ECANCELED);
8978 static bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
8980 struct io_tctx_node *node;
8981 enum io_wq_cancel cret;
8984 mutex_lock(&ctx->uring_lock);
8985 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
8986 struct io_uring_task *tctx = node->task->io_uring;
8989 * io_wq will stay alive while we hold uring_lock, because it's
8990 * killed after ctx nodes, which requires to take the lock.
8992 if (!tctx || !tctx->io_wq)
8994 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
8995 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8997 mutex_unlock(&ctx->uring_lock);
9002 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
9003 struct task_struct *task,
9006 struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
9007 struct io_uring_task *tctx = task ? task->io_uring : NULL;
9010 enum io_wq_cancel cret;
9014 ret |= io_uring_try_cancel_iowq(ctx);
9015 } else if (tctx && tctx->io_wq) {
9017 * Cancels requests of all rings, not only @ctx, but
9018 * it's fine as the task is in exit/exec.
9020 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
9022 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
9025 /* SQPOLL thread does its own polling */
9026 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
9027 (ctx->sq_data && ctx->sq_data->thread == current)) {
9028 while (!list_empty_careful(&ctx->iopoll_list)) {
9029 io_iopoll_try_reap_events(ctx);
9034 ret |= io_cancel_defer_files(ctx, task, cancel_all);
9035 ret |= io_poll_remove_all(ctx, task, cancel_all);
9036 ret |= io_kill_timeouts(ctx, task, cancel_all);
9038 ret |= io_run_task_work();
9039 ret |= io_run_ctx_fallback(ctx);
9046 static int __io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9048 struct io_uring_task *tctx = current->io_uring;
9049 struct io_tctx_node *node;
9052 if (unlikely(!tctx)) {
9053 ret = io_uring_alloc_task_context(current, ctx);
9056 tctx = current->io_uring;
9058 if (!xa_load(&tctx->xa, (unsigned long)ctx)) {
9059 node = kmalloc(sizeof(*node), GFP_KERNEL);
9063 node->task = current;
9065 ret = xa_err(xa_store(&tctx->xa, (unsigned long)ctx,
9072 mutex_lock(&ctx->uring_lock);
9073 list_add(&node->ctx_node, &ctx->tctx_list);
9074 mutex_unlock(&ctx->uring_lock);
9081 * Note that this task has used io_uring. We use it for cancelation purposes.
9083 static inline int io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9085 struct io_uring_task *tctx = current->io_uring;
9087 if (likely(tctx && tctx->last == ctx))
9089 return __io_uring_add_tctx_node(ctx);
9093 * Remove this io_uring_file -> task mapping.
9095 static void io_uring_del_tctx_node(unsigned long index)
9097 struct io_uring_task *tctx = current->io_uring;
9098 struct io_tctx_node *node;
9102 node = xa_erase(&tctx->xa, index);
9106 WARN_ON_ONCE(current != node->task);
9107 WARN_ON_ONCE(list_empty(&node->ctx_node));
9109 mutex_lock(&node->ctx->uring_lock);
9110 list_del(&node->ctx_node);
9111 mutex_unlock(&node->ctx->uring_lock);
9113 if (tctx->last == node->ctx)
9118 static void io_uring_clean_tctx(struct io_uring_task *tctx)
9120 struct io_wq *wq = tctx->io_wq;
9121 struct io_tctx_node *node;
9122 unsigned long index;
9124 xa_for_each(&tctx->xa, index, node)
9125 io_uring_del_tctx_node(index);
9128 * Must be after io_uring_del_task_file() (removes nodes under
9129 * uring_lock) to avoid race with io_uring_try_cancel_iowq().
9132 io_wq_put_and_exit(wq);
9136 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
9139 return atomic_read(&tctx->inflight_tracked);
9140 return percpu_counter_sum(&tctx->inflight);
9143 static void io_uring_drop_tctx_refs(struct task_struct *task)
9145 struct io_uring_task *tctx = task->io_uring;
9146 unsigned int refs = tctx->cached_refs;
9148 tctx->cached_refs = 0;
9149 percpu_counter_sub(&tctx->inflight, refs);
9150 put_task_struct_many(task, refs);
9154 * Find any io_uring ctx that this task has registered or done IO on, and cancel
9155 * requests. @sqd should be not-null IIF it's an SQPOLL thread cancellation.
9157 static void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
9159 struct io_uring_task *tctx = current->io_uring;
9160 struct io_ring_ctx *ctx;
9164 WARN_ON_ONCE(sqd && sqd->thread != current);
9166 if (!current->io_uring)
9169 io_wq_exit_start(tctx->io_wq);
9171 io_uring_drop_tctx_refs(current);
9172 atomic_inc(&tctx->in_idle);
9174 /* read completions before cancelations */
9175 inflight = tctx_inflight(tctx, !cancel_all);
9180 struct io_tctx_node *node;
9181 unsigned long index;
9183 xa_for_each(&tctx->xa, index, node) {
9184 /* sqpoll task will cancel all its requests */
9185 if (node->ctx->sq_data)
9187 io_uring_try_cancel_requests(node->ctx, current,
9191 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
9192 io_uring_try_cancel_requests(ctx, current,
9196 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
9198 * If we've seen completions, retry without waiting. This
9199 * avoids a race where a completion comes in before we did
9200 * prepare_to_wait().
9202 if (inflight == tctx_inflight(tctx, !cancel_all))
9204 finish_wait(&tctx->wait, &wait);
9206 atomic_dec(&tctx->in_idle);
9208 io_uring_clean_tctx(tctx);
9210 /* for exec all current's requests should be gone, kill tctx */
9211 __io_uring_free(current);
9215 void __io_uring_cancel(struct files_struct *files)
9217 io_uring_cancel_generic(!files, NULL);
9220 static void *io_uring_validate_mmap_request(struct file *file,
9221 loff_t pgoff, size_t sz)
9223 struct io_ring_ctx *ctx = file->private_data;
9224 loff_t offset = pgoff << PAGE_SHIFT;
9229 case IORING_OFF_SQ_RING:
9230 case IORING_OFF_CQ_RING:
9233 case IORING_OFF_SQES:
9237 return ERR_PTR(-EINVAL);
9240 page = virt_to_head_page(ptr);
9241 if (sz > page_size(page))
9242 return ERR_PTR(-EINVAL);
9249 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9251 size_t sz = vma->vm_end - vma->vm_start;
9255 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
9257 return PTR_ERR(ptr);
9259 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
9260 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
9263 #else /* !CONFIG_MMU */
9265 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9267 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
9270 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
9272 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
9275 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
9276 unsigned long addr, unsigned long len,
9277 unsigned long pgoff, unsigned long flags)
9281 ptr = io_uring_validate_mmap_request(file, pgoff, len);
9283 return PTR_ERR(ptr);
9285 return (unsigned long) ptr;
9288 #endif /* !CONFIG_MMU */
9290 static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
9295 if (!io_sqring_full(ctx))
9297 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9299 if (!io_sqring_full(ctx))
9302 } while (!signal_pending(current));
9304 finish_wait(&ctx->sqo_sq_wait, &wait);
9308 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9309 struct __kernel_timespec __user **ts,
9310 const sigset_t __user **sig)
9312 struct io_uring_getevents_arg arg;
9315 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9316 * is just a pointer to the sigset_t.
9318 if (!(flags & IORING_ENTER_EXT_ARG)) {
9319 *sig = (const sigset_t __user *) argp;
9325 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9326 * timespec and sigset_t pointers if good.
9328 if (*argsz != sizeof(arg))
9330 if (copy_from_user(&arg, argp, sizeof(arg)))
9332 *sig = u64_to_user_ptr(arg.sigmask);
9333 *argsz = arg.sigmask_sz;
9334 *ts = u64_to_user_ptr(arg.ts);
9338 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9339 u32, min_complete, u32, flags, const void __user *, argp,
9342 struct io_ring_ctx *ctx;
9349 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9350 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG)))
9354 if (unlikely(!f.file))
9358 if (unlikely(f.file->f_op != &io_uring_fops))
9362 ctx = f.file->private_data;
9363 if (unlikely(!percpu_ref_tryget(&ctx->refs)))
9367 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
9371 * For SQ polling, the thread will do all submissions and completions.
9372 * Just return the requested submit count, and wake the thread if
9376 if (ctx->flags & IORING_SETUP_SQPOLL) {
9377 io_cqring_overflow_flush(ctx, false);
9380 if (unlikely(ctx->sq_data->thread == NULL))
9382 if (flags & IORING_ENTER_SQ_WAKEUP)
9383 wake_up(&ctx->sq_data->wait);
9384 if (flags & IORING_ENTER_SQ_WAIT) {
9385 ret = io_sqpoll_wait_sq(ctx);
9389 submitted = to_submit;
9390 } else if (to_submit) {
9391 ret = io_uring_add_tctx_node(ctx);
9394 mutex_lock(&ctx->uring_lock);
9395 submitted = io_submit_sqes(ctx, to_submit);
9396 mutex_unlock(&ctx->uring_lock);
9398 if (submitted != to_submit)
9401 if (flags & IORING_ENTER_GETEVENTS) {
9402 const sigset_t __user *sig;
9403 struct __kernel_timespec __user *ts;
9405 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
9409 min_complete = min(min_complete, ctx->cq_entries);
9412 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9413 * space applications don't need to do io completion events
9414 * polling again, they can rely on io_sq_thread to do polling
9415 * work, which can reduce cpu usage and uring_lock contention.
9417 if (ctx->flags & IORING_SETUP_IOPOLL &&
9418 !(ctx->flags & IORING_SETUP_SQPOLL)) {
9419 ret = io_iopoll_check(ctx, min_complete);
9421 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
9426 percpu_ref_put(&ctx->refs);
9429 return submitted ? submitted : ret;
9432 #ifdef CONFIG_PROC_FS
9433 static int io_uring_show_cred(struct seq_file *m, unsigned int id,
9434 const struct cred *cred)
9436 struct user_namespace *uns = seq_user_ns(m);
9437 struct group_info *gi;
9442 seq_printf(m, "%5d\n", id);
9443 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
9444 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
9445 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
9446 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
9447 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
9448 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
9449 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
9450 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
9451 seq_puts(m, "\n\tGroups:\t");
9452 gi = cred->group_info;
9453 for (g = 0; g < gi->ngroups; g++) {
9454 seq_put_decimal_ull(m, g ? " " : "",
9455 from_kgid_munged(uns, gi->gid[g]));
9457 seq_puts(m, "\n\tCapEff:\t");
9458 cap = cred->cap_effective;
9459 CAP_FOR_EACH_U32(__capi)
9460 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
9465 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
9467 struct io_sq_data *sq = NULL;
9472 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9473 * since fdinfo case grabs it in the opposite direction of normal use
9474 * cases. If we fail to get the lock, we just don't iterate any
9475 * structures that could be going away outside the io_uring mutex.
9477 has_lock = mutex_trylock(&ctx->uring_lock);
9479 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) {
9485 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9486 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9487 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9488 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9489 struct file *f = io_file_from_index(ctx, i);
9492 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9494 seq_printf(m, "%5u: <none>\n", i);
9496 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9497 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9498 struct io_mapped_ubuf *buf = ctx->user_bufs[i];
9499 unsigned int len = buf->ubuf_end - buf->ubuf;
9501 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf, len);
9503 if (has_lock && !xa_empty(&ctx->personalities)) {
9504 unsigned long index;
9505 const struct cred *cred;
9507 seq_printf(m, "Personalities:\n");
9508 xa_for_each(&ctx->personalities, index, cred)
9509 io_uring_show_cred(m, index, cred);
9511 seq_printf(m, "PollList:\n");
9512 spin_lock_irq(&ctx->completion_lock);
9513 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9514 struct hlist_head *list = &ctx->cancel_hash[i];
9515 struct io_kiocb *req;
9517 hlist_for_each_entry(req, list, hash_node)
9518 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9519 req->task->task_works != NULL);
9521 spin_unlock_irq(&ctx->completion_lock);
9523 mutex_unlock(&ctx->uring_lock);
9526 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9528 struct io_ring_ctx *ctx = f->private_data;
9530 if (percpu_ref_tryget(&ctx->refs)) {
9531 __io_uring_show_fdinfo(ctx, m);
9532 percpu_ref_put(&ctx->refs);
9537 static const struct file_operations io_uring_fops = {
9538 .release = io_uring_release,
9539 .mmap = io_uring_mmap,
9541 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9542 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9544 .poll = io_uring_poll,
9545 .fasync = io_uring_fasync,
9546 #ifdef CONFIG_PROC_FS
9547 .show_fdinfo = io_uring_show_fdinfo,
9551 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9552 struct io_uring_params *p)
9554 struct io_rings *rings;
9555 size_t size, sq_array_offset;
9557 /* make sure these are sane, as we already accounted them */
9558 ctx->sq_entries = p->sq_entries;
9559 ctx->cq_entries = p->cq_entries;
9561 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9562 if (size == SIZE_MAX)
9565 rings = io_mem_alloc(size);
9570 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9571 rings->sq_ring_mask = p->sq_entries - 1;
9572 rings->cq_ring_mask = p->cq_entries - 1;
9573 rings->sq_ring_entries = p->sq_entries;
9574 rings->cq_ring_entries = p->cq_entries;
9576 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9577 if (size == SIZE_MAX) {
9578 io_mem_free(ctx->rings);
9583 ctx->sq_sqes = io_mem_alloc(size);
9584 if (!ctx->sq_sqes) {
9585 io_mem_free(ctx->rings);
9593 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
9597 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9601 ret = io_uring_add_tctx_node(ctx);
9606 fd_install(fd, file);
9611 * Allocate an anonymous fd, this is what constitutes the application
9612 * visible backing of an io_uring instance. The application mmaps this
9613 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9614 * we have to tie this fd to a socket for file garbage collection purposes.
9616 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
9619 #if defined(CONFIG_UNIX)
9622 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9625 return ERR_PTR(ret);
9628 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9629 O_RDWR | O_CLOEXEC);
9630 #if defined(CONFIG_UNIX)
9632 sock_release(ctx->ring_sock);
9633 ctx->ring_sock = NULL;
9635 ctx->ring_sock->file = file;
9641 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9642 struct io_uring_params __user *params)
9644 struct io_ring_ctx *ctx;
9650 if (entries > IORING_MAX_ENTRIES) {
9651 if (!(p->flags & IORING_SETUP_CLAMP))
9653 entries = IORING_MAX_ENTRIES;
9657 * Use twice as many entries for the CQ ring. It's possible for the
9658 * application to drive a higher depth than the size of the SQ ring,
9659 * since the sqes are only used at submission time. This allows for
9660 * some flexibility in overcommitting a bit. If the application has
9661 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9662 * of CQ ring entries manually.
9664 p->sq_entries = roundup_pow_of_two(entries);
9665 if (p->flags & IORING_SETUP_CQSIZE) {
9667 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9668 * to a power-of-two, if it isn't already. We do NOT impose
9669 * any cq vs sq ring sizing.
9673 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9674 if (!(p->flags & IORING_SETUP_CLAMP))
9676 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9678 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9679 if (p->cq_entries < p->sq_entries)
9682 p->cq_entries = 2 * p->sq_entries;
9685 ctx = io_ring_ctx_alloc(p);
9688 ctx->compat = in_compat_syscall();
9689 if (!capable(CAP_IPC_LOCK))
9690 ctx->user = get_uid(current_user());
9693 * This is just grabbed for accounting purposes. When a process exits,
9694 * the mm is exited and dropped before the files, hence we need to hang
9695 * on to this mm purely for the purposes of being able to unaccount
9696 * memory (locked/pinned vm). It's not used for anything else.
9698 mmgrab(current->mm);
9699 ctx->mm_account = current->mm;
9701 ret = io_allocate_scq_urings(ctx, p);
9705 ret = io_sq_offload_create(ctx, p);
9708 /* always set a rsrc node */
9709 ret = io_rsrc_node_switch_start(ctx);
9712 io_rsrc_node_switch(ctx, NULL);
9714 memset(&p->sq_off, 0, sizeof(p->sq_off));
9715 p->sq_off.head = offsetof(struct io_rings, sq.head);
9716 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9717 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9718 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9719 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9720 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9721 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9723 memset(&p->cq_off, 0, sizeof(p->cq_off));
9724 p->cq_off.head = offsetof(struct io_rings, cq.head);
9725 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9726 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9727 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9728 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9729 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9730 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9732 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9733 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9734 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9735 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
9736 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
9737 IORING_FEAT_RSRC_TAGS;
9739 if (copy_to_user(params, p, sizeof(*p))) {
9744 file = io_uring_get_file(ctx);
9746 ret = PTR_ERR(file);
9751 * Install ring fd as the very last thing, so we don't risk someone
9752 * having closed it before we finish setup
9754 ret = io_uring_install_fd(ctx, file);
9756 /* fput will clean it up */
9761 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9764 io_ring_ctx_wait_and_kill(ctx);
9769 * Sets up an aio uring context, and returns the fd. Applications asks for a
9770 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9771 * params structure passed in.
9773 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9775 struct io_uring_params p;
9778 if (copy_from_user(&p, params, sizeof(p)))
9780 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9785 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9786 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9787 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9788 IORING_SETUP_R_DISABLED))
9791 return io_uring_create(entries, &p, params);
9794 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9795 struct io_uring_params __user *, params)
9797 return io_uring_setup(entries, params);
9800 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9802 struct io_uring_probe *p;
9806 size = struct_size(p, ops, nr_args);
9807 if (size == SIZE_MAX)
9809 p = kzalloc(size, GFP_KERNEL);
9814 if (copy_from_user(p, arg, size))
9817 if (memchr_inv(p, 0, size))
9820 p->last_op = IORING_OP_LAST - 1;
9821 if (nr_args > IORING_OP_LAST)
9822 nr_args = IORING_OP_LAST;
9824 for (i = 0; i < nr_args; i++) {
9826 if (!io_op_defs[i].not_supported)
9827 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9832 if (copy_to_user(arg, p, size))
9839 static int io_register_personality(struct io_ring_ctx *ctx)
9841 const struct cred *creds;
9845 creds = get_current_cred();
9847 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
9848 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
9855 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9856 unsigned int nr_args)
9858 struct io_uring_restriction *res;
9862 /* Restrictions allowed only if rings started disabled */
9863 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9866 /* We allow only a single restrictions registration */
9867 if (ctx->restrictions.registered)
9870 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9873 size = array_size(nr_args, sizeof(*res));
9874 if (size == SIZE_MAX)
9877 res = memdup_user(arg, size);
9879 return PTR_ERR(res);
9883 for (i = 0; i < nr_args; i++) {
9884 switch (res[i].opcode) {
9885 case IORING_RESTRICTION_REGISTER_OP:
9886 if (res[i].register_op >= IORING_REGISTER_LAST) {
9891 __set_bit(res[i].register_op,
9892 ctx->restrictions.register_op);
9894 case IORING_RESTRICTION_SQE_OP:
9895 if (res[i].sqe_op >= IORING_OP_LAST) {
9900 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9902 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9903 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9905 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9906 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9915 /* Reset all restrictions if an error happened */
9917 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9919 ctx->restrictions.registered = true;
9925 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9927 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9930 if (ctx->restrictions.registered)
9931 ctx->restricted = 1;
9933 ctx->flags &= ~IORING_SETUP_R_DISABLED;
9934 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
9935 wake_up(&ctx->sq_data->wait);
9939 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
9940 struct io_uring_rsrc_update2 *up,
9948 if (check_add_overflow(up->offset, nr_args, &tmp))
9950 err = io_rsrc_node_switch_start(ctx);
9955 case IORING_RSRC_FILE:
9956 return __io_sqe_files_update(ctx, up, nr_args);
9957 case IORING_RSRC_BUFFER:
9958 return __io_sqe_buffers_update(ctx, up, nr_args);
9963 static int io_register_files_update(struct io_ring_ctx *ctx, void __user *arg,
9966 struct io_uring_rsrc_update2 up;
9970 memset(&up, 0, sizeof(up));
9971 if (copy_from_user(&up, arg, sizeof(struct io_uring_rsrc_update)))
9973 return __io_register_rsrc_update(ctx, IORING_RSRC_FILE, &up, nr_args);
9976 static int io_register_rsrc_update(struct io_ring_ctx *ctx, void __user *arg,
9977 unsigned size, unsigned type)
9979 struct io_uring_rsrc_update2 up;
9981 if (size != sizeof(up))
9983 if (copy_from_user(&up, arg, sizeof(up)))
9985 if (!up.nr || up.resv)
9987 return __io_register_rsrc_update(ctx, type, &up, up.nr);
9990 static int io_register_rsrc(struct io_ring_ctx *ctx, void __user *arg,
9991 unsigned int size, unsigned int type)
9993 struct io_uring_rsrc_register rr;
9995 /* keep it extendible */
9996 if (size != sizeof(rr))
9999 memset(&rr, 0, sizeof(rr));
10000 if (copy_from_user(&rr, arg, size))
10002 if (!rr.nr || rr.resv || rr.resv2)
10006 case IORING_RSRC_FILE:
10007 return io_sqe_files_register(ctx, u64_to_user_ptr(rr.data),
10008 rr.nr, u64_to_user_ptr(rr.tags));
10009 case IORING_RSRC_BUFFER:
10010 return io_sqe_buffers_register(ctx, u64_to_user_ptr(rr.data),
10011 rr.nr, u64_to_user_ptr(rr.tags));
10016 static int io_register_iowq_aff(struct io_ring_ctx *ctx, void __user *arg,
10019 struct io_uring_task *tctx = current->io_uring;
10020 cpumask_var_t new_mask;
10023 if (!tctx || !tctx->io_wq)
10026 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
10029 cpumask_clear(new_mask);
10030 if (len > cpumask_size())
10031 len = cpumask_size();
10033 if (copy_from_user(new_mask, arg, len)) {
10034 free_cpumask_var(new_mask);
10038 ret = io_wq_cpu_affinity(tctx->io_wq, new_mask);
10039 free_cpumask_var(new_mask);
10043 static int io_unregister_iowq_aff(struct io_ring_ctx *ctx)
10045 struct io_uring_task *tctx = current->io_uring;
10047 if (!tctx || !tctx->io_wq)
10050 return io_wq_cpu_affinity(tctx->io_wq, NULL);
10053 static bool io_register_op_must_quiesce(int op)
10056 case IORING_REGISTER_BUFFERS:
10057 case IORING_UNREGISTER_BUFFERS:
10058 case IORING_REGISTER_FILES:
10059 case IORING_UNREGISTER_FILES:
10060 case IORING_REGISTER_FILES_UPDATE:
10061 case IORING_REGISTER_PROBE:
10062 case IORING_REGISTER_PERSONALITY:
10063 case IORING_UNREGISTER_PERSONALITY:
10064 case IORING_REGISTER_FILES2:
10065 case IORING_REGISTER_FILES_UPDATE2:
10066 case IORING_REGISTER_BUFFERS2:
10067 case IORING_REGISTER_BUFFERS_UPDATE:
10068 case IORING_REGISTER_IOWQ_AFF:
10069 case IORING_UNREGISTER_IOWQ_AFF:
10076 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
10077 void __user *arg, unsigned nr_args)
10078 __releases(ctx->uring_lock)
10079 __acquires(ctx->uring_lock)
10084 * We're inside the ring mutex, if the ref is already dying, then
10085 * someone else killed the ctx or is already going through
10086 * io_uring_register().
10088 if (percpu_ref_is_dying(&ctx->refs))
10091 if (ctx->restricted) {
10092 if (opcode >= IORING_REGISTER_LAST)
10094 opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
10095 if (!test_bit(opcode, ctx->restrictions.register_op))
10099 if (io_register_op_must_quiesce(opcode)) {
10100 percpu_ref_kill(&ctx->refs);
10103 * Drop uring mutex before waiting for references to exit. If
10104 * another thread is currently inside io_uring_enter() it might
10105 * need to grab the uring_lock to make progress. If we hold it
10106 * here across the drain wait, then we can deadlock. It's safe
10107 * to drop the mutex here, since no new references will come in
10108 * after we've killed the percpu ref.
10110 mutex_unlock(&ctx->uring_lock);
10112 ret = wait_for_completion_interruptible(&ctx->ref_comp);
10115 ret = io_run_task_work_sig();
10119 mutex_lock(&ctx->uring_lock);
10122 io_refs_resurrect(&ctx->refs, &ctx->ref_comp);
10128 case IORING_REGISTER_BUFFERS:
10129 ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
10131 case IORING_UNREGISTER_BUFFERS:
10133 if (arg || nr_args)
10135 ret = io_sqe_buffers_unregister(ctx);
10137 case IORING_REGISTER_FILES:
10138 ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
10140 case IORING_UNREGISTER_FILES:
10142 if (arg || nr_args)
10144 ret = io_sqe_files_unregister(ctx);
10146 case IORING_REGISTER_FILES_UPDATE:
10147 ret = io_register_files_update(ctx, arg, nr_args);
10149 case IORING_REGISTER_EVENTFD:
10150 case IORING_REGISTER_EVENTFD_ASYNC:
10154 ret = io_eventfd_register(ctx, arg);
10157 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
10158 ctx->eventfd_async = 1;
10160 ctx->eventfd_async = 0;
10162 case IORING_UNREGISTER_EVENTFD:
10164 if (arg || nr_args)
10166 ret = io_eventfd_unregister(ctx);
10168 case IORING_REGISTER_PROBE:
10170 if (!arg || nr_args > 256)
10172 ret = io_probe(ctx, arg, nr_args);
10174 case IORING_REGISTER_PERSONALITY:
10176 if (arg || nr_args)
10178 ret = io_register_personality(ctx);
10180 case IORING_UNREGISTER_PERSONALITY:
10184 ret = io_unregister_personality(ctx, nr_args);
10186 case IORING_REGISTER_ENABLE_RINGS:
10188 if (arg || nr_args)
10190 ret = io_register_enable_rings(ctx);
10192 case IORING_REGISTER_RESTRICTIONS:
10193 ret = io_register_restrictions(ctx, arg, nr_args);
10195 case IORING_REGISTER_FILES2:
10196 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE);
10198 case IORING_REGISTER_FILES_UPDATE2:
10199 ret = io_register_rsrc_update(ctx, arg, nr_args,
10202 case IORING_REGISTER_BUFFERS2:
10203 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER);
10205 case IORING_REGISTER_BUFFERS_UPDATE:
10206 ret = io_register_rsrc_update(ctx, arg, nr_args,
10207 IORING_RSRC_BUFFER);
10209 case IORING_REGISTER_IOWQ_AFF:
10211 if (!arg || !nr_args)
10213 ret = io_register_iowq_aff(ctx, arg, nr_args);
10215 case IORING_UNREGISTER_IOWQ_AFF:
10217 if (arg || nr_args)
10219 ret = io_unregister_iowq_aff(ctx);
10226 if (io_register_op_must_quiesce(opcode)) {
10227 /* bring the ctx back to life */
10228 percpu_ref_reinit(&ctx->refs);
10229 reinit_completion(&ctx->ref_comp);
10234 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
10235 void __user *, arg, unsigned int, nr_args)
10237 struct io_ring_ctx *ctx;
10246 if (f.file->f_op != &io_uring_fops)
10249 ctx = f.file->private_data;
10251 io_run_task_work();
10253 mutex_lock(&ctx->uring_lock);
10254 ret = __io_uring_register(ctx, opcode, arg, nr_args);
10255 mutex_unlock(&ctx->uring_lock);
10256 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
10257 ctx->cq_ev_fd != NULL, ret);
10263 static int __init io_uring_init(void)
10265 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
10266 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
10267 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
10270 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
10271 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
10272 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
10273 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
10274 BUILD_BUG_SQE_ELEM(1, __u8, flags);
10275 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
10276 BUILD_BUG_SQE_ELEM(4, __s32, fd);
10277 BUILD_BUG_SQE_ELEM(8, __u64, off);
10278 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
10279 BUILD_BUG_SQE_ELEM(16, __u64, addr);
10280 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
10281 BUILD_BUG_SQE_ELEM(24, __u32, len);
10282 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
10283 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
10284 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
10285 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
10286 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
10287 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
10288 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
10289 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
10290 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
10291 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
10292 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
10293 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
10294 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
10295 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
10296 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
10297 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
10298 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
10299 BUILD_BUG_SQE_ELEM(40, __u16, buf_group);
10300 BUILD_BUG_SQE_ELEM(42, __u16, personality);
10301 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
10303 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
10304 sizeof(struct io_uring_rsrc_update));
10305 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
10306 sizeof(struct io_uring_rsrc_update2));
10307 /* should fit into one byte */
10308 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
10310 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
10311 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
10313 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
10317 __initcall(io_uring_init);