1 // SPDX-License-Identifier: GPL-2.0
3 * Shared application/kernel submission and completion ring pairs, for
4 * supporting fast/efficient IO.
6 * A note on the read/write ordering memory barriers that are matched between
7 * the application and kernel side.
9 * After the application reads the CQ ring tail, it must use an
10 * appropriate smp_rmb() to pair with the smp_wmb() the kernel uses
11 * before writing the tail (using smp_load_acquire to read the tail will
12 * do). It also needs a smp_mb() before updating CQ head (ordering the
13 * entry load(s) with the head store), pairing with an implicit barrier
14 * through a control-dependency in io_get_cqe (smp_store_release to
15 * store head will do). Failure to do so could lead to reading invalid
18 * Likewise, the application must use an appropriate smp_wmb() before
19 * writing the SQ tail (ordering SQ entry stores with the tail store),
20 * which pairs with smp_load_acquire in io_get_sqring (smp_store_release
21 * to store the tail will do). And it needs a barrier ordering the SQ
22 * head load before writing new SQ entries (smp_load_acquire to read
25 * When using the SQ poll thread (IORING_SETUP_SQPOLL), the application
26 * needs to check the SQ flags for IORING_SQ_NEED_WAKEUP *after*
27 * updating the SQ tail; a full memory barrier smp_mb() is needed
30 * Also see the examples in the liburing library:
32 * git://git.kernel.dk/liburing
34 * io_uring also uses READ/WRITE_ONCE() for _any_ store or load that happens
35 * from data shared between the kernel and application. This is done both
36 * for ordering purposes, but also to ensure that once a value is loaded from
37 * data that the application could potentially modify, it remains stable.
39 * Copyright (C) 2018-2019 Jens Axboe
40 * Copyright (c) 2018-2019 Christoph Hellwig
42 #include <linux/kernel.h>
43 #include <linux/init.h>
44 #include <linux/errno.h>
45 #include <linux/syscalls.h>
46 #include <linux/compat.h>
47 #include <net/compat.h>
48 #include <linux/refcount.h>
49 #include <linux/uio.h>
50 #include <linux/bits.h>
52 #include <linux/sched/signal.h>
54 #include <linux/file.h>
55 #include <linux/fdtable.h>
57 #include <linux/mman.h>
58 #include <linux/percpu.h>
59 #include <linux/slab.h>
60 #include <linux/blkdev.h>
61 #include <linux/bvec.h>
62 #include <linux/net.h>
64 #include <net/af_unix.h>
66 #include <linux/anon_inodes.h>
67 #include <linux/sched/mm.h>
68 #include <linux/uaccess.h>
69 #include <linux/nospec.h>
70 #include <linux/sizes.h>
71 #include <linux/hugetlb.h>
72 #include <linux/highmem.h>
73 #include <linux/namei.h>
74 #include <linux/fsnotify.h>
75 #include <linux/fadvise.h>
76 #include <linux/eventpoll.h>
77 #include <linux/splice.h>
78 #include <linux/task_work.h>
79 #include <linux/pagemap.h>
80 #include <linux/io_uring.h>
82 #define CREATE_TRACE_POINTS
83 #include <trace/events/io_uring.h>
85 #include <uapi/linux/io_uring.h>
90 #define IORING_MAX_ENTRIES 32768
91 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
94 * Shift of 9 is 512 entries, or exactly one page on 64-bit archs
96 #define IORING_FILE_TABLE_SHIFT 9
97 #define IORING_MAX_FILES_TABLE (1U << IORING_FILE_TABLE_SHIFT)
98 #define IORING_FILE_TABLE_MASK (IORING_MAX_FILES_TABLE - 1)
99 #define IORING_MAX_FIXED_FILES (64 * IORING_MAX_FILES_TABLE)
100 #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
101 IORING_REGISTER_LAST + IORING_OP_LAST)
103 #define IO_RSRC_TAG_TABLE_SHIFT 9
104 #define IO_RSRC_TAG_TABLE_MAX (1U << IO_RSRC_TAG_TABLE_SHIFT)
105 #define IO_RSRC_TAG_TABLE_MASK (IO_RSRC_TAG_TABLE_MAX - 1)
107 #define IORING_MAX_REG_BUFFERS (1U << 14)
109 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
110 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
113 #define IO_TCTX_REFS_CACHE_NR (1U << 10)
116 u32 head ____cacheline_aligned_in_smp;
117 u32 tail ____cacheline_aligned_in_smp;
121 * This data is shared with the application through the mmap at offsets
122 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
124 * The offsets to the member fields are published through struct
125 * io_sqring_offsets when calling io_uring_setup.
129 * Head and tail offsets into the ring; the offsets need to be
130 * masked to get valid indices.
132 * The kernel controls head of the sq ring and the tail of the cq ring,
133 * and the application controls tail of the sq ring and the head of the
136 struct io_uring sq, cq;
138 * Bitmasks to apply to head and tail offsets (constant, equals
141 u32 sq_ring_mask, cq_ring_mask;
142 /* Ring sizes (constant, power of 2) */
143 u32 sq_ring_entries, cq_ring_entries;
145 * Number of invalid entries dropped by the kernel due to
146 * invalid index stored in array
148 * Written by the kernel, shouldn't be modified by the
149 * application (i.e. get number of "new events" by comparing to
152 * After a new SQ head value was read by the application this
153 * counter includes all submissions that were dropped reaching
154 * the new SQ head (and possibly more).
160 * Written by the kernel, shouldn't be modified by the
163 * The application needs a full memory barrier before checking
164 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
170 * Written by the application, shouldn't be modified by the
175 * Number of completion events lost because the queue was full;
176 * this should be avoided by the application by making sure
177 * there are not more requests pending than there is space in
178 * the completion queue.
180 * Written by the kernel, shouldn't be modified by the
181 * application (i.e. get number of "new events" by comparing to
184 * As completion events come in out of order this counter is not
185 * ordered with any other data.
189 * Ring buffer of completion events.
191 * The kernel writes completion events fresh every time they are
192 * produced, so the application is allowed to modify pending
195 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
198 enum io_uring_cmd_flags {
199 IO_URING_F_NONBLOCK = 1,
200 IO_URING_F_COMPLETE_DEFER = 2,
203 struct io_mapped_ubuf {
206 unsigned int nr_bvecs;
207 unsigned long acct_pages;
208 struct bio_vec bvec[];
213 struct io_overflow_cqe {
214 struct io_uring_cqe cqe;
215 struct list_head list;
218 struct io_fixed_file {
219 /* file * with additional FFS_* flags */
220 unsigned long file_ptr;
224 struct list_head list;
229 struct io_mapped_ubuf *buf;
233 struct io_file_table {
234 /* two level table */
235 struct io_fixed_file **files;
238 struct io_rsrc_node {
239 struct percpu_ref refs;
240 struct list_head node;
241 struct list_head rsrc_list;
242 struct io_rsrc_data *rsrc_data;
243 struct llist_node llist;
247 typedef void (rsrc_put_fn)(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc);
249 struct io_rsrc_data {
250 struct io_ring_ctx *ctx;
256 struct completion done;
261 struct list_head list;
267 struct io_restriction {
268 DECLARE_BITMAP(register_op, IORING_REGISTER_LAST);
269 DECLARE_BITMAP(sqe_op, IORING_OP_LAST);
270 u8 sqe_flags_allowed;
271 u8 sqe_flags_required;
276 IO_SQ_THREAD_SHOULD_STOP = 0,
277 IO_SQ_THREAD_SHOULD_PARK,
282 atomic_t park_pending;
285 /* ctx's that are using this sqd */
286 struct list_head ctx_list;
288 struct task_struct *thread;
289 struct wait_queue_head wait;
291 unsigned sq_thread_idle;
297 struct completion exited;
300 #define IO_IOPOLL_BATCH 8
301 #define IO_COMPL_BATCH 32
302 #define IO_REQ_CACHE_SIZE 32
303 #define IO_REQ_ALLOC_BATCH 8
305 struct io_comp_state {
306 struct io_kiocb *reqs[IO_COMPL_BATCH];
308 /* inline/task_work completion list, under ->uring_lock */
309 struct list_head free_list;
312 struct io_submit_link {
313 struct io_kiocb *head;
314 struct io_kiocb *last;
317 struct io_submit_state {
318 struct blk_plug plug;
319 struct io_submit_link link;
322 * io_kiocb alloc cache
324 void *reqs[IO_REQ_CACHE_SIZE];
325 unsigned int free_reqs;
330 * Batch completion logic
332 struct io_comp_state comp;
335 * File reference cache
339 unsigned int file_refs;
340 unsigned int ios_left;
344 /* const or read-mostly hot data */
346 struct percpu_ref refs;
348 struct io_rings *rings;
350 unsigned int compat: 1;
351 unsigned int drain_next: 1;
352 unsigned int eventfd_async: 1;
353 unsigned int restricted: 1;
354 unsigned int off_timeout_used: 1;
355 unsigned int drain_active: 1;
356 } ____cacheline_aligned_in_smp;
358 /* submission data */
360 struct mutex uring_lock;
363 * Ring buffer of indices into array of io_uring_sqe, which is
364 * mmapped by the application using the IORING_OFF_SQES offset.
366 * This indirection could e.g. be used to assign fixed
367 * io_uring_sqe entries to operations and only submit them to
368 * the queue when needed.
370 * The kernel modifies neither the indices array nor the entries
374 struct io_uring_sqe *sq_sqes;
375 unsigned cached_sq_head;
377 struct list_head defer_list;
380 * Fixed resources fast path, should be accessed only under
381 * uring_lock, and updated through io_uring_register(2)
383 struct io_rsrc_node *rsrc_node;
384 struct io_file_table file_table;
385 unsigned nr_user_files;
386 unsigned nr_user_bufs;
387 struct io_mapped_ubuf **user_bufs;
389 struct io_submit_state submit_state;
390 struct list_head timeout_list;
391 struct list_head cq_overflow_list;
392 struct xarray io_buffers;
393 struct xarray personalities;
395 unsigned sq_thread_idle;
396 } ____cacheline_aligned_in_smp;
398 /* IRQ completion list, under ->completion_lock */
399 struct list_head locked_free_list;
400 unsigned int locked_free_nr;
402 const struct cred *sq_creds; /* cred used for __io_sq_thread() */
403 struct io_sq_data *sq_data; /* if using sq thread polling */
405 struct wait_queue_head sqo_sq_wait;
406 struct list_head sqd_list;
408 unsigned long check_cq_overflow;
411 unsigned cached_cq_tail;
413 struct eventfd_ctx *cq_ev_fd;
414 struct wait_queue_head poll_wait;
415 struct wait_queue_head cq_wait;
417 atomic_t cq_timeouts;
418 struct fasync_struct *cq_fasync;
419 unsigned cq_last_tm_flush;
420 } ____cacheline_aligned_in_smp;
423 spinlock_t completion_lock;
426 * ->iopoll_list is protected by the ctx->uring_lock for
427 * io_uring instances that don't use IORING_SETUP_SQPOLL.
428 * For SQPOLL, only the single threaded io_sq_thread() will
429 * manipulate the list, hence no extra locking is needed there.
431 struct list_head iopoll_list;
432 struct hlist_head *cancel_hash;
433 unsigned cancel_hash_bits;
434 bool poll_multi_file;
435 } ____cacheline_aligned_in_smp;
437 struct io_restriction restrictions;
439 /* slow path rsrc auxilary data, used by update/register */
441 struct io_rsrc_node *rsrc_backup_node;
442 struct io_mapped_ubuf *dummy_ubuf;
443 struct io_rsrc_data *file_data;
444 struct io_rsrc_data *buf_data;
446 struct delayed_work rsrc_put_work;
447 struct llist_head rsrc_put_llist;
448 struct list_head rsrc_ref_list;
449 spinlock_t rsrc_ref_lock;
452 /* Keep this last, we don't need it for the fast path */
454 #if defined(CONFIG_UNIX)
455 struct socket *ring_sock;
457 /* hashed buffered write serialization */
458 struct io_wq_hash *hash_map;
460 /* Only used for accounting purposes */
461 struct user_struct *user;
462 struct mm_struct *mm_account;
464 /* ctx exit and cancelation */
465 struct callback_head *exit_task_work;
466 struct work_struct exit_work;
467 struct list_head tctx_list;
468 struct completion ref_comp;
472 struct io_uring_task {
473 /* submission side */
476 struct wait_queue_head wait;
477 const struct io_ring_ctx *last;
479 struct percpu_counter inflight;
480 atomic_t inflight_tracked;
483 spinlock_t task_lock;
484 struct io_wq_work_list task_list;
485 unsigned long task_state;
486 struct callback_head task_work;
490 * First field must be the file pointer in all the
491 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
493 struct io_poll_iocb {
495 struct wait_queue_head *head;
499 struct wait_queue_entry wait;
502 struct io_poll_update {
508 bool update_user_data;
516 struct io_timeout_data {
517 struct io_kiocb *req;
518 struct hrtimer timer;
519 struct timespec64 ts;
520 enum hrtimer_mode mode;
525 struct sockaddr __user *addr;
526 int __user *addr_len;
528 unsigned long nofile;
548 struct list_head list;
549 /* head of the link, used by linked timeouts only */
550 struct io_kiocb *head;
553 struct io_timeout_rem {
558 struct timespec64 ts;
563 /* NOTE: kiocb has the file as the first member, so don't do it here */
571 struct sockaddr __user *addr;
578 struct compat_msghdr __user *umsg_compat;
579 struct user_msghdr __user *umsg;
585 struct io_buffer *kbuf;
591 struct filename *filename;
593 unsigned long nofile;
596 struct io_rsrc_update {
622 struct epoll_event event;
626 struct file *file_out;
627 struct file *file_in;
634 struct io_provide_buf {
648 const char __user *filename;
649 struct statx __user *buffer;
661 struct filename *oldpath;
662 struct filename *newpath;
670 struct filename *filename;
673 struct io_completion {
675 struct list_head list;
679 struct io_async_connect {
680 struct sockaddr_storage address;
683 struct io_async_msghdr {
684 struct iovec fast_iov[UIO_FASTIOV];
685 /* points to an allocated iov, if NULL we use fast_iov instead */
686 struct iovec *free_iov;
687 struct sockaddr __user *uaddr;
689 struct sockaddr_storage addr;
693 struct iovec fast_iov[UIO_FASTIOV];
694 const struct iovec *free_iovec;
695 struct iov_iter iter;
697 struct wait_page_queue wpq;
701 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
702 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
703 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
704 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
705 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
706 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
708 /* first byte is taken by user flags, shift it to not overlap */
713 REQ_F_LINK_TIMEOUT_BIT,
714 REQ_F_NEED_CLEANUP_BIT,
716 REQ_F_BUFFER_SELECTED_BIT,
717 REQ_F_LTIMEOUT_ACTIVE_BIT,
718 REQ_F_COMPLETE_INLINE_BIT,
720 REQ_F_DONT_REISSUE_BIT,
722 /* keep async read/write and isreg together and in order */
723 REQ_F_ASYNC_READ_BIT,
724 REQ_F_ASYNC_WRITE_BIT,
727 /* not a real bit, just to check we're not overflowing the space */
733 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
734 /* drain existing IO first */
735 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
737 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
738 /* doesn't sever on completion < 0 */
739 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
741 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
742 /* IOSQE_BUFFER_SELECT */
743 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
745 /* fail rest of links */
746 REQ_F_FAIL = BIT(REQ_F_FAIL_BIT),
747 /* on inflight list, should be cancelled and waited on exit reliably */
748 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
749 /* read/write uses file position */
750 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
751 /* must not punt to workers */
752 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
753 /* has or had linked timeout */
754 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
756 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
757 /* already went through poll handler */
758 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
759 /* buffer already selected */
760 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
761 /* linked timeout is active, i.e. prepared by link's head */
762 REQ_F_LTIMEOUT_ACTIVE = BIT(REQ_F_LTIMEOUT_ACTIVE_BIT),
763 /* completion is deferred through io_comp_state */
764 REQ_F_COMPLETE_INLINE = BIT(REQ_F_COMPLETE_INLINE_BIT),
765 /* caller should reissue async */
766 REQ_F_REISSUE = BIT(REQ_F_REISSUE_BIT),
767 /* don't attempt request reissue, see io_rw_reissue() */
768 REQ_F_DONT_REISSUE = BIT(REQ_F_DONT_REISSUE_BIT),
769 /* supports async reads */
770 REQ_F_ASYNC_READ = BIT(REQ_F_ASYNC_READ_BIT),
771 /* supports async writes */
772 REQ_F_ASYNC_WRITE = BIT(REQ_F_ASYNC_WRITE_BIT),
774 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
775 /* has creds assigned */
776 REQ_F_CREDS = BIT(REQ_F_CREDS_BIT),
780 struct io_poll_iocb poll;
781 struct io_poll_iocb *double_poll;
784 struct io_task_work {
785 struct io_wq_work_node node;
786 task_work_func_t func;
790 IORING_RSRC_FILE = 0,
791 IORING_RSRC_BUFFER = 1,
795 * NOTE! Each of the iocb union members has the file pointer
796 * as the first entry in their struct definition. So you can
797 * access the file pointer through any of the sub-structs,
798 * or directly as just 'ki_filp' in this struct.
804 struct io_poll_iocb poll;
805 struct io_poll_update poll_update;
806 struct io_accept accept;
808 struct io_cancel cancel;
809 struct io_timeout timeout;
810 struct io_timeout_rem timeout_rem;
811 struct io_connect connect;
812 struct io_sr_msg sr_msg;
814 struct io_close close;
815 struct io_rsrc_update rsrc_update;
816 struct io_fadvise fadvise;
817 struct io_madvise madvise;
818 struct io_epoll epoll;
819 struct io_splice splice;
820 struct io_provide_buf pbuf;
821 struct io_statx statx;
822 struct io_shutdown shutdown;
823 struct io_rename rename;
824 struct io_unlink unlink;
825 /* use only after cleaning per-op data, see io_clean_op() */
826 struct io_completion compl;
829 /* opcode allocated if it needs to store data for async defer */
832 /* polled IO has completed */
838 struct io_ring_ctx *ctx;
841 struct task_struct *task;
844 struct io_kiocb *link;
845 struct percpu_ref *fixed_rsrc_refs;
847 /* used with ctx->iopoll_list with reads/writes */
848 struct list_head inflight_entry;
850 struct io_task_work io_task_work;
851 struct callback_head task_work;
853 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
854 struct hlist_node hash_node;
855 struct async_poll *apoll;
856 struct io_wq_work work;
857 const struct cred *creds;
859 /* store used ubuf, so we can prevent reloading */
860 struct io_mapped_ubuf *imu;
863 struct io_tctx_node {
864 struct list_head ctx_node;
865 struct task_struct *task;
866 struct io_ring_ctx *ctx;
869 struct io_defer_entry {
870 struct list_head list;
871 struct io_kiocb *req;
876 /* needs req->file assigned */
877 unsigned needs_file : 1;
878 /* hash wq insertion if file is a regular file */
879 unsigned hash_reg_file : 1;
880 /* unbound wq insertion if file is a non-regular file */
881 unsigned unbound_nonreg_file : 1;
882 /* opcode is not supported by this kernel */
883 unsigned not_supported : 1;
884 /* set if opcode supports polled "wait" */
886 unsigned pollout : 1;
887 /* op supports buffer selection */
888 unsigned buffer_select : 1;
889 /* do prep async if is going to be punted */
890 unsigned needs_async_setup : 1;
891 /* should block plug */
893 /* size of async data needed, if any */
894 unsigned short async_size;
897 static const struct io_op_def io_op_defs[] = {
898 [IORING_OP_NOP] = {},
899 [IORING_OP_READV] = {
901 .unbound_nonreg_file = 1,
904 .needs_async_setup = 1,
906 .async_size = sizeof(struct io_async_rw),
908 [IORING_OP_WRITEV] = {
911 .unbound_nonreg_file = 1,
913 .needs_async_setup = 1,
915 .async_size = sizeof(struct io_async_rw),
917 [IORING_OP_FSYNC] = {
920 [IORING_OP_READ_FIXED] = {
922 .unbound_nonreg_file = 1,
925 .async_size = sizeof(struct io_async_rw),
927 [IORING_OP_WRITE_FIXED] = {
930 .unbound_nonreg_file = 1,
933 .async_size = sizeof(struct io_async_rw),
935 [IORING_OP_POLL_ADD] = {
937 .unbound_nonreg_file = 1,
939 [IORING_OP_POLL_REMOVE] = {},
940 [IORING_OP_SYNC_FILE_RANGE] = {
943 [IORING_OP_SENDMSG] = {
945 .unbound_nonreg_file = 1,
947 .needs_async_setup = 1,
948 .async_size = sizeof(struct io_async_msghdr),
950 [IORING_OP_RECVMSG] = {
952 .unbound_nonreg_file = 1,
955 .needs_async_setup = 1,
956 .async_size = sizeof(struct io_async_msghdr),
958 [IORING_OP_TIMEOUT] = {
959 .async_size = sizeof(struct io_timeout_data),
961 [IORING_OP_TIMEOUT_REMOVE] = {
962 /* used by timeout updates' prep() */
964 [IORING_OP_ACCEPT] = {
966 .unbound_nonreg_file = 1,
969 [IORING_OP_ASYNC_CANCEL] = {},
970 [IORING_OP_LINK_TIMEOUT] = {
971 .async_size = sizeof(struct io_timeout_data),
973 [IORING_OP_CONNECT] = {
975 .unbound_nonreg_file = 1,
977 .needs_async_setup = 1,
978 .async_size = sizeof(struct io_async_connect),
980 [IORING_OP_FALLOCATE] = {
983 [IORING_OP_OPENAT] = {},
984 [IORING_OP_CLOSE] = {},
985 [IORING_OP_FILES_UPDATE] = {},
986 [IORING_OP_STATX] = {},
989 .unbound_nonreg_file = 1,
993 .async_size = sizeof(struct io_async_rw),
995 [IORING_OP_WRITE] = {
997 .unbound_nonreg_file = 1,
1000 .async_size = sizeof(struct io_async_rw),
1002 [IORING_OP_FADVISE] = {
1005 [IORING_OP_MADVISE] = {},
1006 [IORING_OP_SEND] = {
1008 .unbound_nonreg_file = 1,
1011 [IORING_OP_RECV] = {
1013 .unbound_nonreg_file = 1,
1017 [IORING_OP_OPENAT2] = {
1019 [IORING_OP_EPOLL_CTL] = {
1020 .unbound_nonreg_file = 1,
1022 [IORING_OP_SPLICE] = {
1025 .unbound_nonreg_file = 1,
1027 [IORING_OP_PROVIDE_BUFFERS] = {},
1028 [IORING_OP_REMOVE_BUFFERS] = {},
1032 .unbound_nonreg_file = 1,
1034 [IORING_OP_SHUTDOWN] = {
1037 [IORING_OP_RENAMEAT] = {},
1038 [IORING_OP_UNLINKAT] = {},
1041 static bool io_disarm_next(struct io_kiocb *req);
1042 static void io_uring_del_tctx_node(unsigned long index);
1043 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
1044 struct task_struct *task,
1046 static void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd);
1047 static struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx);
1049 static bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1050 long res, unsigned int cflags);
1051 static void io_put_req(struct io_kiocb *req);
1052 static void io_put_req_deferred(struct io_kiocb *req, int nr);
1053 static void io_dismantle_req(struct io_kiocb *req);
1054 static void io_put_task(struct task_struct *task, int nr);
1055 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
1056 static void io_queue_linked_timeout(struct io_kiocb *req);
1057 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
1058 struct io_uring_rsrc_update2 *up,
1060 static void io_clean_op(struct io_kiocb *req);
1061 static struct file *io_file_get(struct io_submit_state *state,
1062 struct io_kiocb *req, int fd, bool fixed);
1063 static void __io_queue_sqe(struct io_kiocb *req);
1064 static void io_rsrc_put_work(struct work_struct *work);
1066 static void io_req_task_queue(struct io_kiocb *req);
1067 static void io_submit_flush_completions(struct io_ring_ctx *ctx);
1068 static bool io_poll_remove_waitqs(struct io_kiocb *req);
1069 static int io_req_prep_async(struct io_kiocb *req);
1071 static struct kmem_cache *req_cachep;
1073 static const struct file_operations io_uring_fops;
1075 struct sock *io_uring_get_socket(struct file *file)
1077 #if defined(CONFIG_UNIX)
1078 if (file->f_op == &io_uring_fops) {
1079 struct io_ring_ctx *ctx = file->private_data;
1081 return ctx->ring_sock->sk;
1086 EXPORT_SYMBOL(io_uring_get_socket);
1088 #define io_for_each_link(pos, head) \
1089 for (pos = (head); pos; pos = pos->link)
1091 static inline void io_req_set_rsrc_node(struct io_kiocb *req)
1093 struct io_ring_ctx *ctx = req->ctx;
1095 if (!req->fixed_rsrc_refs) {
1096 req->fixed_rsrc_refs = &ctx->rsrc_node->refs;
1097 percpu_ref_get(req->fixed_rsrc_refs);
1101 static void io_refs_resurrect(struct percpu_ref *ref, struct completion *compl)
1103 bool got = percpu_ref_tryget(ref);
1105 /* already at zero, wait for ->release() */
1107 wait_for_completion(compl);
1108 percpu_ref_resurrect(ref);
1110 percpu_ref_put(ref);
1113 static bool io_match_task(struct io_kiocb *head, struct task_struct *task,
1116 struct io_kiocb *req;
1118 if (task && head->task != task)
1123 io_for_each_link(req, head) {
1124 if (req->flags & REQ_F_INFLIGHT)
1130 static inline void req_set_fail(struct io_kiocb *req)
1132 req->flags |= REQ_F_FAIL;
1135 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1137 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1139 complete(&ctx->ref_comp);
1142 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1144 return !req->timeout.off;
1147 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1149 struct io_ring_ctx *ctx;
1152 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1157 * Use 5 bits less than the max cq entries, that should give us around
1158 * 32 entries per hash list if totally full and uniformly spread.
1160 hash_bits = ilog2(p->cq_entries);
1164 ctx->cancel_hash_bits = hash_bits;
1165 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1167 if (!ctx->cancel_hash)
1169 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1171 ctx->dummy_ubuf = kzalloc(sizeof(*ctx->dummy_ubuf), GFP_KERNEL);
1172 if (!ctx->dummy_ubuf)
1174 /* set invalid range, so io_import_fixed() fails meeting it */
1175 ctx->dummy_ubuf->ubuf = -1UL;
1177 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1178 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1181 ctx->flags = p->flags;
1182 init_waitqueue_head(&ctx->sqo_sq_wait);
1183 INIT_LIST_HEAD(&ctx->sqd_list);
1184 init_waitqueue_head(&ctx->poll_wait);
1185 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1186 init_completion(&ctx->ref_comp);
1187 xa_init_flags(&ctx->io_buffers, XA_FLAGS_ALLOC1);
1188 xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
1189 mutex_init(&ctx->uring_lock);
1190 init_waitqueue_head(&ctx->cq_wait);
1191 spin_lock_init(&ctx->completion_lock);
1192 INIT_LIST_HEAD(&ctx->iopoll_list);
1193 INIT_LIST_HEAD(&ctx->defer_list);
1194 INIT_LIST_HEAD(&ctx->timeout_list);
1195 spin_lock_init(&ctx->rsrc_ref_lock);
1196 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
1197 INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
1198 init_llist_head(&ctx->rsrc_put_llist);
1199 INIT_LIST_HEAD(&ctx->tctx_list);
1200 INIT_LIST_HEAD(&ctx->submit_state.comp.free_list);
1201 INIT_LIST_HEAD(&ctx->locked_free_list);
1204 kfree(ctx->dummy_ubuf);
1205 kfree(ctx->cancel_hash);
1210 static void io_account_cq_overflow(struct io_ring_ctx *ctx)
1212 struct io_rings *r = ctx->rings;
1214 WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1);
1218 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1220 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1221 struct io_ring_ctx *ctx = req->ctx;
1223 return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail;
1229 static void io_req_track_inflight(struct io_kiocb *req)
1231 if (!(req->flags & REQ_F_INFLIGHT)) {
1232 req->flags |= REQ_F_INFLIGHT;
1233 atomic_inc(¤t->io_uring->inflight_tracked);
1237 static void io_prep_async_work(struct io_kiocb *req)
1239 const struct io_op_def *def = &io_op_defs[req->opcode];
1240 struct io_ring_ctx *ctx = req->ctx;
1242 if (!(req->flags & REQ_F_CREDS)) {
1243 req->flags |= REQ_F_CREDS;
1244 req->creds = get_current_cred();
1247 req->work.list.next = NULL;
1248 req->work.flags = 0;
1249 if (req->flags & REQ_F_FORCE_ASYNC)
1250 req->work.flags |= IO_WQ_WORK_CONCURRENT;
1252 if (req->flags & REQ_F_ISREG) {
1253 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1254 io_wq_hash_work(&req->work, file_inode(req->file));
1255 } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
1256 if (def->unbound_nonreg_file)
1257 req->work.flags |= IO_WQ_WORK_UNBOUND;
1260 switch (req->opcode) {
1261 case IORING_OP_SPLICE:
1263 if (!S_ISREG(file_inode(req->splice.file_in)->i_mode))
1264 req->work.flags |= IO_WQ_WORK_UNBOUND;
1269 static void io_prep_async_link(struct io_kiocb *req)
1271 struct io_kiocb *cur;
1273 io_for_each_link(cur, req)
1274 io_prep_async_work(cur);
1277 static void io_queue_async_work(struct io_kiocb *req)
1279 struct io_ring_ctx *ctx = req->ctx;
1280 struct io_kiocb *link = io_prep_linked_timeout(req);
1281 struct io_uring_task *tctx = req->task->io_uring;
1284 BUG_ON(!tctx->io_wq);
1286 /* init ->work of the whole link before punting */
1287 io_prep_async_link(req);
1288 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1289 &req->work, req->flags);
1290 io_wq_enqueue(tctx->io_wq, &req->work);
1292 io_queue_linked_timeout(link);
1295 static void io_kill_timeout(struct io_kiocb *req, int status)
1296 __must_hold(&req->ctx->completion_lock)
1298 struct io_timeout_data *io = req->async_data;
1300 if (hrtimer_try_to_cancel(&io->timer) != -1) {
1301 atomic_set(&req->ctx->cq_timeouts,
1302 atomic_read(&req->ctx->cq_timeouts) + 1);
1303 list_del_init(&req->timeout.list);
1304 io_cqring_fill_event(req->ctx, req->user_data, status, 0);
1305 io_put_req_deferred(req, 1);
1309 static void io_queue_deferred(struct io_ring_ctx *ctx)
1311 while (!list_empty(&ctx->defer_list)) {
1312 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1313 struct io_defer_entry, list);
1315 if (req_need_defer(de->req, de->seq))
1317 list_del_init(&de->list);
1318 io_req_task_queue(de->req);
1323 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1325 u32 seq = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
1327 while (!list_empty(&ctx->timeout_list)) {
1328 u32 events_needed, events_got;
1329 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1330 struct io_kiocb, timeout.list);
1332 if (io_is_timeout_noseq(req))
1336 * Since seq can easily wrap around over time, subtract
1337 * the last seq at which timeouts were flushed before comparing.
1338 * Assuming not more than 2^31-1 events have happened since,
1339 * these subtractions won't have wrapped, so we can check if
1340 * target is in [last_seq, current_seq] by comparing the two.
1342 events_needed = req->timeout.target_seq - ctx->cq_last_tm_flush;
1343 events_got = seq - ctx->cq_last_tm_flush;
1344 if (events_got < events_needed)
1347 list_del_init(&req->timeout.list);
1348 io_kill_timeout(req, 0);
1350 ctx->cq_last_tm_flush = seq;
1353 static void __io_commit_cqring_flush(struct io_ring_ctx *ctx)
1355 if (ctx->off_timeout_used)
1356 io_flush_timeouts(ctx);
1357 if (ctx->drain_active)
1358 io_queue_deferred(ctx);
1361 static inline void io_commit_cqring(struct io_ring_ctx *ctx)
1363 if (unlikely(ctx->off_timeout_used || ctx->drain_active))
1364 __io_commit_cqring_flush(ctx);
1365 /* order cqe stores with ring update */
1366 smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
1369 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1371 struct io_rings *r = ctx->rings;
1373 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == ctx->sq_entries;
1376 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
1378 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
1381 static inline struct io_uring_cqe *io_get_cqe(struct io_ring_ctx *ctx)
1383 struct io_rings *rings = ctx->rings;
1384 unsigned tail, mask = ctx->cq_entries - 1;
1387 * writes to the cq entry need to come after reading head; the
1388 * control dependency is enough as we're using WRITE_ONCE to
1391 if (__io_cqring_events(ctx) == ctx->cq_entries)
1394 tail = ctx->cached_cq_tail++;
1395 return &rings->cqes[tail & mask];
1398 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1400 if (likely(!ctx->cq_ev_fd))
1402 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1404 return !ctx->eventfd_async || io_wq_current_is_worker();
1407 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1409 /* see waitqueue_active() comment */
1412 if (waitqueue_active(&ctx->cq_wait))
1413 wake_up(&ctx->cq_wait);
1414 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1415 wake_up(&ctx->sq_data->wait);
1416 if (io_should_trigger_evfd(ctx))
1417 eventfd_signal(ctx->cq_ev_fd, 1);
1418 if (waitqueue_active(&ctx->poll_wait)) {
1419 wake_up_interruptible(&ctx->poll_wait);
1420 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1424 static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx)
1426 /* see waitqueue_active() comment */
1429 if (ctx->flags & IORING_SETUP_SQPOLL) {
1430 if (waitqueue_active(&ctx->cq_wait))
1431 wake_up(&ctx->cq_wait);
1433 if (io_should_trigger_evfd(ctx))
1434 eventfd_signal(ctx->cq_ev_fd, 1);
1435 if (waitqueue_active(&ctx->poll_wait)) {
1436 wake_up_interruptible(&ctx->poll_wait);
1437 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1441 /* Returns true if there are no backlogged entries after the flush */
1442 static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1444 unsigned long flags;
1445 bool all_flushed, posted;
1447 if (!force && __io_cqring_events(ctx) == ctx->cq_entries)
1451 spin_lock_irqsave(&ctx->completion_lock, flags);
1452 while (!list_empty(&ctx->cq_overflow_list)) {
1453 struct io_uring_cqe *cqe = io_get_cqe(ctx);
1454 struct io_overflow_cqe *ocqe;
1458 ocqe = list_first_entry(&ctx->cq_overflow_list,
1459 struct io_overflow_cqe, list);
1461 memcpy(cqe, &ocqe->cqe, sizeof(*cqe));
1463 io_account_cq_overflow(ctx);
1466 list_del(&ocqe->list);
1470 all_flushed = list_empty(&ctx->cq_overflow_list);
1472 clear_bit(0, &ctx->check_cq_overflow);
1473 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
1477 io_commit_cqring(ctx);
1478 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1480 io_cqring_ev_posted(ctx);
1484 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1488 if (test_bit(0, &ctx->check_cq_overflow)) {
1489 /* iopoll syncs against uring_lock, not completion_lock */
1490 if (ctx->flags & IORING_SETUP_IOPOLL)
1491 mutex_lock(&ctx->uring_lock);
1492 ret = __io_cqring_overflow_flush(ctx, force);
1493 if (ctx->flags & IORING_SETUP_IOPOLL)
1494 mutex_unlock(&ctx->uring_lock);
1501 * Shamelessly stolen from the mm implementation of page reference checking,
1502 * see commit f958d7b528b1 for details.
1504 #define req_ref_zero_or_close_to_overflow(req) \
1505 ((unsigned int) atomic_read(&(req->refs)) + 127u <= 127u)
1507 static inline bool req_ref_inc_not_zero(struct io_kiocb *req)
1509 return atomic_inc_not_zero(&req->refs);
1512 static inline bool req_ref_sub_and_test(struct io_kiocb *req, int refs)
1514 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1515 return atomic_sub_and_test(refs, &req->refs);
1518 static inline bool req_ref_put_and_test(struct io_kiocb *req)
1520 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1521 return atomic_dec_and_test(&req->refs);
1524 static inline void req_ref_put(struct io_kiocb *req)
1526 WARN_ON_ONCE(req_ref_put_and_test(req));
1529 static inline void req_ref_get(struct io_kiocb *req)
1531 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1532 atomic_inc(&req->refs);
1535 static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data,
1536 long res, unsigned int cflags)
1538 struct io_overflow_cqe *ocqe;
1540 ocqe = kmalloc(sizeof(*ocqe), GFP_ATOMIC | __GFP_ACCOUNT);
1543 * If we're in ring overflow flush mode, or in task cancel mode,
1544 * or cannot allocate an overflow entry, then we need to drop it
1547 io_account_cq_overflow(ctx);
1550 if (list_empty(&ctx->cq_overflow_list)) {
1551 set_bit(0, &ctx->check_cq_overflow);
1552 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
1554 ocqe->cqe.user_data = user_data;
1555 ocqe->cqe.res = res;
1556 ocqe->cqe.flags = cflags;
1557 list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
1561 static inline bool __io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1562 long res, unsigned int cflags)
1564 struct io_uring_cqe *cqe;
1566 trace_io_uring_complete(ctx, user_data, res, cflags);
1569 * If we can't get a cq entry, userspace overflowed the
1570 * submission (by quite a lot). Increment the overflow count in
1573 cqe = io_get_cqe(ctx);
1575 WRITE_ONCE(cqe->user_data, user_data);
1576 WRITE_ONCE(cqe->res, res);
1577 WRITE_ONCE(cqe->flags, cflags);
1580 return io_cqring_event_overflow(ctx, user_data, res, cflags);
1583 /* not as hot to bloat with inlining */
1584 static noinline bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1585 long res, unsigned int cflags)
1587 return __io_cqring_fill_event(ctx, user_data, res, cflags);
1590 static void io_req_complete_post(struct io_kiocb *req, long res,
1591 unsigned int cflags)
1593 struct io_ring_ctx *ctx = req->ctx;
1594 unsigned long flags;
1596 spin_lock_irqsave(&ctx->completion_lock, flags);
1597 __io_cqring_fill_event(ctx, req->user_data, res, cflags);
1599 * If we're the last reference to this request, add to our locked
1602 if (req_ref_put_and_test(req)) {
1603 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
1604 if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL))
1605 io_disarm_next(req);
1607 io_req_task_queue(req->link);
1611 io_dismantle_req(req);
1612 io_put_task(req->task, 1);
1613 list_add(&req->compl.list, &ctx->locked_free_list);
1614 ctx->locked_free_nr++;
1616 if (!percpu_ref_tryget(&ctx->refs))
1619 io_commit_cqring(ctx);
1620 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1623 io_cqring_ev_posted(ctx);
1624 percpu_ref_put(&ctx->refs);
1628 static inline bool io_req_needs_clean(struct io_kiocb *req)
1630 return req->flags & (REQ_F_BUFFER_SELECTED | REQ_F_NEED_CLEANUP |
1631 REQ_F_POLLED | REQ_F_INFLIGHT | REQ_F_CREDS);
1634 static void io_req_complete_state(struct io_kiocb *req, long res,
1635 unsigned int cflags)
1637 if (io_req_needs_clean(req))
1640 req->compl.cflags = cflags;
1641 req->flags |= REQ_F_COMPLETE_INLINE;
1644 static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags,
1645 long res, unsigned cflags)
1647 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1648 io_req_complete_state(req, res, cflags);
1650 io_req_complete_post(req, res, cflags);
1653 static inline void io_req_complete(struct io_kiocb *req, long res)
1655 __io_req_complete(req, 0, res, 0);
1658 static void io_req_complete_failed(struct io_kiocb *req, long res)
1662 io_req_complete_post(req, res, 0);
1665 static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
1666 struct io_comp_state *cs)
1668 spin_lock_irq(&ctx->completion_lock);
1669 list_splice_init(&ctx->locked_free_list, &cs->free_list);
1670 ctx->locked_free_nr = 0;
1671 spin_unlock_irq(&ctx->completion_lock);
1674 /* Returns true IFF there are requests in the cache */
1675 static bool io_flush_cached_reqs(struct io_ring_ctx *ctx)
1677 struct io_submit_state *state = &ctx->submit_state;
1678 struct io_comp_state *cs = &state->comp;
1682 * If we have more than a batch's worth of requests in our IRQ side
1683 * locked cache, grab the lock and move them over to our submission
1686 if (READ_ONCE(ctx->locked_free_nr) > IO_COMPL_BATCH)
1687 io_flush_cached_locked_reqs(ctx, cs);
1689 nr = state->free_reqs;
1690 while (!list_empty(&cs->free_list)) {
1691 struct io_kiocb *req = list_first_entry(&cs->free_list,
1692 struct io_kiocb, compl.list);
1694 list_del(&req->compl.list);
1695 state->reqs[nr++] = req;
1696 if (nr == ARRAY_SIZE(state->reqs))
1700 state->free_reqs = nr;
1704 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
1706 struct io_submit_state *state = &ctx->submit_state;
1708 BUILD_BUG_ON(IO_REQ_ALLOC_BATCH > ARRAY_SIZE(state->reqs));
1710 if (!state->free_reqs) {
1711 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1714 if (io_flush_cached_reqs(ctx))
1717 ret = kmem_cache_alloc_bulk(req_cachep, gfp, IO_REQ_ALLOC_BATCH,
1721 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1722 * retry single alloc to be on the safe side.
1724 if (unlikely(ret <= 0)) {
1725 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1726 if (!state->reqs[0])
1730 state->free_reqs = ret;
1734 return state->reqs[state->free_reqs];
1737 static inline void io_put_file(struct file *file)
1743 static void io_dismantle_req(struct io_kiocb *req)
1745 unsigned int flags = req->flags;
1747 if (io_req_needs_clean(req))
1749 if (!(flags & REQ_F_FIXED_FILE))
1750 io_put_file(req->file);
1751 if (req->fixed_rsrc_refs)
1752 percpu_ref_put(req->fixed_rsrc_refs);
1753 if (req->async_data)
1754 kfree(req->async_data);
1757 /* must to be called somewhat shortly after putting a request */
1758 static inline void io_put_task(struct task_struct *task, int nr)
1760 struct io_uring_task *tctx = task->io_uring;
1762 percpu_counter_sub(&tctx->inflight, nr);
1763 if (unlikely(atomic_read(&tctx->in_idle)))
1764 wake_up(&tctx->wait);
1765 put_task_struct_many(task, nr);
1768 static void __io_free_req(struct io_kiocb *req)
1770 struct io_ring_ctx *ctx = req->ctx;
1772 io_dismantle_req(req);
1773 io_put_task(req->task, 1);
1775 kmem_cache_free(req_cachep, req);
1776 percpu_ref_put(&ctx->refs);
1779 static inline void io_remove_next_linked(struct io_kiocb *req)
1781 struct io_kiocb *nxt = req->link;
1783 req->link = nxt->link;
1787 static bool io_kill_linked_timeout(struct io_kiocb *req)
1788 __must_hold(&req->ctx->completion_lock)
1790 struct io_kiocb *link = req->link;
1793 * Can happen if a linked timeout fired and link had been like
1794 * req -> link t-out -> link t-out [-> ...]
1796 if (link && (link->flags & REQ_F_LTIMEOUT_ACTIVE)) {
1797 struct io_timeout_data *io = link->async_data;
1799 io_remove_next_linked(req);
1800 link->timeout.head = NULL;
1801 if (hrtimer_try_to_cancel(&io->timer) != -1) {
1802 io_cqring_fill_event(link->ctx, link->user_data,
1804 io_put_req_deferred(link, 1);
1811 static void io_fail_links(struct io_kiocb *req)
1812 __must_hold(&req->ctx->completion_lock)
1814 struct io_kiocb *nxt, *link = req->link;
1821 trace_io_uring_fail_link(req, link);
1822 io_cqring_fill_event(link->ctx, link->user_data, -ECANCELED, 0);
1823 io_put_req_deferred(link, 2);
1828 static bool io_disarm_next(struct io_kiocb *req)
1829 __must_hold(&req->ctx->completion_lock)
1831 bool posted = false;
1833 if (likely(req->flags & REQ_F_LINK_TIMEOUT))
1834 posted = io_kill_linked_timeout(req);
1835 if (unlikely((req->flags & REQ_F_FAIL) &&
1836 !(req->flags & REQ_F_HARDLINK))) {
1837 posted |= (req->link != NULL);
1843 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1845 struct io_kiocb *nxt;
1848 * If LINK is set, we have dependent requests in this chain. If we
1849 * didn't fail this request, queue the first one up, moving any other
1850 * dependencies to the next request. In case of failure, fail the rest
1853 if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL)) {
1854 struct io_ring_ctx *ctx = req->ctx;
1855 unsigned long flags;
1858 spin_lock_irqsave(&ctx->completion_lock, flags);
1859 posted = io_disarm_next(req);
1861 io_commit_cqring(req->ctx);
1862 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1864 io_cqring_ev_posted(ctx);
1871 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1873 if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK))))
1875 return __io_req_find_next(req);
1878 static void ctx_flush_and_put(struct io_ring_ctx *ctx)
1882 if (ctx->submit_state.comp.nr) {
1883 mutex_lock(&ctx->uring_lock);
1884 io_submit_flush_completions(ctx);
1885 mutex_unlock(&ctx->uring_lock);
1887 percpu_ref_put(&ctx->refs);
1890 static bool __tctx_task_work(struct io_uring_task *tctx)
1892 struct io_ring_ctx *ctx = NULL;
1893 struct io_wq_work_list list;
1894 struct io_wq_work_node *node;
1896 if (wq_list_empty(&tctx->task_list))
1899 spin_lock_irq(&tctx->task_lock);
1900 list = tctx->task_list;
1901 INIT_WQ_LIST(&tctx->task_list);
1902 spin_unlock_irq(&tctx->task_lock);
1906 struct io_wq_work_node *next = node->next;
1907 struct io_kiocb *req;
1909 req = container_of(node, struct io_kiocb, io_task_work.node);
1910 if (req->ctx != ctx) {
1911 ctx_flush_and_put(ctx);
1913 percpu_ref_get(&ctx->refs);
1916 req->task_work.func(&req->task_work);
1920 ctx_flush_and_put(ctx);
1921 return list.first != NULL;
1924 static void tctx_task_work(struct callback_head *cb)
1926 struct io_uring_task *tctx = container_of(cb, struct io_uring_task, task_work);
1928 clear_bit(0, &tctx->task_state);
1930 while (__tctx_task_work(tctx))
1934 static int io_req_task_work_add(struct io_kiocb *req)
1936 struct task_struct *tsk = req->task;
1937 struct io_uring_task *tctx = tsk->io_uring;
1938 enum task_work_notify_mode notify;
1939 struct io_wq_work_node *node, *prev;
1940 unsigned long flags;
1943 if (unlikely(tsk->flags & PF_EXITING))
1946 WARN_ON_ONCE(!tctx);
1948 spin_lock_irqsave(&tctx->task_lock, flags);
1949 wq_list_add_tail(&req->io_task_work.node, &tctx->task_list);
1950 spin_unlock_irqrestore(&tctx->task_lock, flags);
1952 /* task_work already pending, we're done */
1953 if (test_bit(0, &tctx->task_state) ||
1954 test_and_set_bit(0, &tctx->task_state))
1958 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
1959 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
1960 * processing task_work. There's no reliable way to tell if TWA_RESUME
1963 notify = (req->ctx->flags & IORING_SETUP_SQPOLL) ? TWA_NONE : TWA_SIGNAL;
1965 if (!task_work_add(tsk, &tctx->task_work, notify)) {
1966 wake_up_process(tsk);
1971 * Slow path - we failed, find and delete work. if the work is not
1972 * in the list, it got run and we're fine.
1974 spin_lock_irqsave(&tctx->task_lock, flags);
1975 wq_list_for_each(node, prev, &tctx->task_list) {
1976 if (&req->io_task_work.node == node) {
1977 wq_list_del(&tctx->task_list, node, prev);
1982 spin_unlock_irqrestore(&tctx->task_lock, flags);
1983 clear_bit(0, &tctx->task_state);
1987 static bool io_run_task_work_head(struct callback_head **work_head)
1989 struct callback_head *work, *next;
1990 bool executed = false;
1993 work = xchg(work_head, NULL);
2009 static void io_task_work_add_head(struct callback_head **work_head,
2010 struct callback_head *task_work)
2012 struct callback_head *head;
2015 head = READ_ONCE(*work_head);
2016 task_work->next = head;
2017 } while (cmpxchg(work_head, head, task_work) != head);
2020 static void io_req_task_work_add_fallback(struct io_kiocb *req,
2021 task_work_func_t cb)
2023 init_task_work(&req->task_work, cb);
2024 io_task_work_add_head(&req->ctx->exit_task_work, &req->task_work);
2027 static void io_req_task_cancel(struct callback_head *cb)
2029 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2030 struct io_ring_ctx *ctx = req->ctx;
2032 /* ctx is guaranteed to stay alive while we hold uring_lock */
2033 mutex_lock(&ctx->uring_lock);
2034 io_req_complete_failed(req, req->result);
2035 mutex_unlock(&ctx->uring_lock);
2038 static void __io_req_task_submit(struct io_kiocb *req)
2040 struct io_ring_ctx *ctx = req->ctx;
2042 /* ctx stays valid until unlock, even if we drop all ours ctx->refs */
2043 mutex_lock(&ctx->uring_lock);
2044 if (!(current->flags & PF_EXITING) && !current->in_execve)
2045 __io_queue_sqe(req);
2047 io_req_complete_failed(req, -EFAULT);
2048 mutex_unlock(&ctx->uring_lock);
2051 static void io_req_task_submit(struct callback_head *cb)
2053 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2055 __io_req_task_submit(req);
2058 static void io_req_task_queue_fail(struct io_kiocb *req, int ret)
2061 req->task_work.func = io_req_task_cancel;
2063 if (unlikely(io_req_task_work_add(req)))
2064 io_req_task_work_add_fallback(req, io_req_task_cancel);
2067 static void io_req_task_queue(struct io_kiocb *req)
2069 req->task_work.func = io_req_task_submit;
2071 if (unlikely(io_req_task_work_add(req)))
2072 io_req_task_queue_fail(req, -ECANCELED);
2075 static inline void io_queue_next(struct io_kiocb *req)
2077 struct io_kiocb *nxt = io_req_find_next(req);
2080 io_req_task_queue(nxt);
2083 static void io_free_req(struct io_kiocb *req)
2090 struct task_struct *task;
2095 static inline void io_init_req_batch(struct req_batch *rb)
2102 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2103 struct req_batch *rb)
2106 io_put_task(rb->task, rb->task_refs);
2108 percpu_ref_put_many(&ctx->refs, rb->ctx_refs);
2111 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req,
2112 struct io_submit_state *state)
2115 io_dismantle_req(req);
2117 if (req->task != rb->task) {
2119 io_put_task(rb->task, rb->task_refs);
2120 rb->task = req->task;
2126 if (state->free_reqs != ARRAY_SIZE(state->reqs))
2127 state->reqs[state->free_reqs++] = req;
2129 list_add(&req->compl.list, &state->comp.free_list);
2132 static void io_submit_flush_completions(struct io_ring_ctx *ctx)
2134 struct io_comp_state *cs = &ctx->submit_state.comp;
2136 struct io_kiocb *req;
2137 struct req_batch rb;
2139 io_init_req_batch(&rb);
2140 spin_lock_irq(&ctx->completion_lock);
2141 for (i = 0; i < nr; i++) {
2143 __io_cqring_fill_event(ctx, req->user_data, req->result,
2146 io_commit_cqring(ctx);
2147 spin_unlock_irq(&ctx->completion_lock);
2149 io_cqring_ev_posted(ctx);
2150 for (i = 0; i < nr; i++) {
2153 /* submission and completion refs */
2154 if (req_ref_sub_and_test(req, 2))
2155 io_req_free_batch(&rb, req, &ctx->submit_state);
2158 io_req_free_batch_finish(ctx, &rb);
2163 * Drop reference to request, return next in chain (if there is one) if this
2164 * was the last reference to this request.
2166 static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2168 struct io_kiocb *nxt = NULL;
2170 if (req_ref_put_and_test(req)) {
2171 nxt = io_req_find_next(req);
2177 static inline void io_put_req(struct io_kiocb *req)
2179 if (req_ref_put_and_test(req))
2183 static void io_put_req_deferred_cb(struct callback_head *cb)
2185 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2190 static void io_free_req_deferred(struct io_kiocb *req)
2192 req->task_work.func = io_put_req_deferred_cb;
2193 if (unlikely(io_req_task_work_add(req)))
2194 io_req_task_work_add_fallback(req, io_put_req_deferred_cb);
2197 static inline void io_put_req_deferred(struct io_kiocb *req, int refs)
2199 if (req_ref_sub_and_test(req, refs))
2200 io_free_req_deferred(req);
2203 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2205 /* See comment at the top of this file */
2207 return __io_cqring_events(ctx);
2210 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2212 struct io_rings *rings = ctx->rings;
2214 /* make sure SQ entry isn't read before tail */
2215 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2218 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2220 unsigned int cflags;
2222 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2223 cflags |= IORING_CQE_F_BUFFER;
2224 req->flags &= ~REQ_F_BUFFER_SELECTED;
2229 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2231 struct io_buffer *kbuf;
2233 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2234 return io_put_kbuf(req, kbuf);
2237 static inline bool io_run_task_work(void)
2240 * Not safe to run on exiting task, and the task_work handling will
2241 * not add work to such a task.
2243 if (unlikely(current->flags & PF_EXITING))
2245 if (current->task_works) {
2246 __set_current_state(TASK_RUNNING);
2255 * Find and free completed poll iocbs
2257 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2258 struct list_head *done)
2260 struct req_batch rb;
2261 struct io_kiocb *req;
2263 /* order with ->result store in io_complete_rw_iopoll() */
2266 io_init_req_batch(&rb);
2267 while (!list_empty(done)) {
2270 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2271 list_del(&req->inflight_entry);
2273 if (READ_ONCE(req->result) == -EAGAIN &&
2274 !(req->flags & REQ_F_DONT_REISSUE)) {
2275 req->iopoll_completed = 0;
2277 io_queue_async_work(req);
2281 if (req->flags & REQ_F_BUFFER_SELECTED)
2282 cflags = io_put_rw_kbuf(req);
2284 __io_cqring_fill_event(ctx, req->user_data, req->result, cflags);
2287 if (req_ref_put_and_test(req))
2288 io_req_free_batch(&rb, req, &ctx->submit_state);
2291 io_commit_cqring(ctx);
2292 io_cqring_ev_posted_iopoll(ctx);
2293 io_req_free_batch_finish(ctx, &rb);
2296 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2299 struct io_kiocb *req, *tmp;
2305 * Only spin for completions if we don't have multiple devices hanging
2306 * off our complete list, and we're under the requested amount.
2308 spin = !ctx->poll_multi_file && *nr_events < min;
2311 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2312 struct kiocb *kiocb = &req->rw.kiocb;
2315 * Move completed and retryable entries to our local lists.
2316 * If we find a request that requires polling, break out
2317 * and complete those lists first, if we have entries there.
2319 if (READ_ONCE(req->iopoll_completed)) {
2320 list_move_tail(&req->inflight_entry, &done);
2323 if (!list_empty(&done))
2326 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2330 /* iopoll may have completed current req */
2331 if (READ_ONCE(req->iopoll_completed))
2332 list_move_tail(&req->inflight_entry, &done);
2339 if (!list_empty(&done))
2340 io_iopoll_complete(ctx, nr_events, &done);
2346 * We can't just wait for polled events to come to us, we have to actively
2347 * find and complete them.
2349 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2351 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2354 mutex_lock(&ctx->uring_lock);
2355 while (!list_empty(&ctx->iopoll_list)) {
2356 unsigned int nr_events = 0;
2358 io_do_iopoll(ctx, &nr_events, 0);
2360 /* let it sleep and repeat later if can't complete a request */
2364 * Ensure we allow local-to-the-cpu processing to take place,
2365 * in this case we need to ensure that we reap all events.
2366 * Also let task_work, etc. to progress by releasing the mutex
2368 if (need_resched()) {
2369 mutex_unlock(&ctx->uring_lock);
2371 mutex_lock(&ctx->uring_lock);
2374 mutex_unlock(&ctx->uring_lock);
2377 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2379 unsigned int nr_events = 0;
2383 * We disallow the app entering submit/complete with polling, but we
2384 * still need to lock the ring to prevent racing with polled issue
2385 * that got punted to a workqueue.
2387 mutex_lock(&ctx->uring_lock);
2389 * Don't enter poll loop if we already have events pending.
2390 * If we do, we can potentially be spinning for commands that
2391 * already triggered a CQE (eg in error).
2393 if (test_bit(0, &ctx->check_cq_overflow))
2394 __io_cqring_overflow_flush(ctx, false);
2395 if (io_cqring_events(ctx))
2399 * If a submit got punted to a workqueue, we can have the
2400 * application entering polling for a command before it gets
2401 * issued. That app will hold the uring_lock for the duration
2402 * of the poll right here, so we need to take a breather every
2403 * now and then to ensure that the issue has a chance to add
2404 * the poll to the issued list. Otherwise we can spin here
2405 * forever, while the workqueue is stuck trying to acquire the
2408 if (list_empty(&ctx->iopoll_list)) {
2409 mutex_unlock(&ctx->uring_lock);
2411 mutex_lock(&ctx->uring_lock);
2413 if (list_empty(&ctx->iopoll_list))
2416 ret = io_do_iopoll(ctx, &nr_events, min);
2417 } while (!ret && nr_events < min && !need_resched());
2419 mutex_unlock(&ctx->uring_lock);
2423 static void kiocb_end_write(struct io_kiocb *req)
2426 * Tell lockdep we inherited freeze protection from submission
2429 if (req->flags & REQ_F_ISREG) {
2430 struct super_block *sb = file_inode(req->file)->i_sb;
2432 __sb_writers_acquired(sb, SB_FREEZE_WRITE);
2438 static bool io_resubmit_prep(struct io_kiocb *req)
2440 struct io_async_rw *rw = req->async_data;
2443 return !io_req_prep_async(req);
2444 /* may have left rw->iter inconsistent on -EIOCBQUEUED */
2445 iov_iter_revert(&rw->iter, req->result - iov_iter_count(&rw->iter));
2449 static bool io_rw_should_reissue(struct io_kiocb *req)
2451 umode_t mode = file_inode(req->file)->i_mode;
2452 struct io_ring_ctx *ctx = req->ctx;
2454 if (!S_ISBLK(mode) && !S_ISREG(mode))
2456 if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
2457 !(ctx->flags & IORING_SETUP_IOPOLL)))
2460 * If ref is dying, we might be running poll reap from the exit work.
2461 * Don't attempt to reissue from that path, just let it fail with
2464 if (percpu_ref_is_dying(&ctx->refs))
2469 static bool io_resubmit_prep(struct io_kiocb *req)
2473 static bool io_rw_should_reissue(struct io_kiocb *req)
2479 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2480 unsigned int issue_flags)
2484 if (req->rw.kiocb.ki_flags & IOCB_WRITE)
2485 kiocb_end_write(req);
2486 if (res != req->result) {
2487 if ((res == -EAGAIN || res == -EOPNOTSUPP) &&
2488 io_rw_should_reissue(req)) {
2489 req->flags |= REQ_F_REISSUE;
2494 if (req->flags & REQ_F_BUFFER_SELECTED)
2495 cflags = io_put_rw_kbuf(req);
2496 __io_req_complete(req, issue_flags, res, cflags);
2499 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2501 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2503 __io_complete_rw(req, res, res2, 0);
2506 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2508 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2510 if (kiocb->ki_flags & IOCB_WRITE)
2511 kiocb_end_write(req);
2512 if (unlikely(res != req->result)) {
2513 if (!(res == -EAGAIN && io_rw_should_reissue(req) &&
2514 io_resubmit_prep(req))) {
2516 req->flags |= REQ_F_DONT_REISSUE;
2520 WRITE_ONCE(req->result, res);
2521 /* order with io_iopoll_complete() checking ->result */
2523 WRITE_ONCE(req->iopoll_completed, 1);
2527 * After the iocb has been issued, it's safe to be found on the poll list.
2528 * Adding the kiocb to the list AFTER submission ensures that we don't
2529 * find it from a io_do_iopoll() thread before the issuer is done
2530 * accessing the kiocb cookie.
2532 static void io_iopoll_req_issued(struct io_kiocb *req)
2534 struct io_ring_ctx *ctx = req->ctx;
2535 const bool in_async = io_wq_current_is_worker();
2537 /* workqueue context doesn't hold uring_lock, grab it now */
2538 if (unlikely(in_async))
2539 mutex_lock(&ctx->uring_lock);
2542 * Track whether we have multiple files in our lists. This will impact
2543 * how we do polling eventually, not spinning if we're on potentially
2544 * different devices.
2546 if (list_empty(&ctx->iopoll_list)) {
2547 ctx->poll_multi_file = false;
2548 } else if (!ctx->poll_multi_file) {
2549 struct io_kiocb *list_req;
2551 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2553 if (list_req->file != req->file)
2554 ctx->poll_multi_file = true;
2558 * For fast devices, IO may have already completed. If it has, add
2559 * it to the front so we find it first.
2561 if (READ_ONCE(req->iopoll_completed))
2562 list_add(&req->inflight_entry, &ctx->iopoll_list);
2564 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2566 if (unlikely(in_async)) {
2568 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
2569 * in sq thread task context or in io worker task context. If
2570 * current task context is sq thread, we don't need to check
2571 * whether should wake up sq thread.
2573 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2574 wq_has_sleeper(&ctx->sq_data->wait))
2575 wake_up(&ctx->sq_data->wait);
2577 mutex_unlock(&ctx->uring_lock);
2581 static inline void io_state_file_put(struct io_submit_state *state)
2583 if (state->file_refs) {
2584 fput_many(state->file, state->file_refs);
2585 state->file_refs = 0;
2590 * Get as many references to a file as we have IOs left in this submission,
2591 * assuming most submissions are for one file, or at least that each file
2592 * has more than one submission.
2594 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2599 if (state->file_refs) {
2600 if (state->fd == fd) {
2604 io_state_file_put(state);
2606 state->file = fget_many(fd, state->ios_left);
2607 if (unlikely(!state->file))
2611 state->file_refs = state->ios_left - 1;
2615 static bool io_bdev_nowait(struct block_device *bdev)
2617 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2621 * If we tracked the file through the SCM inflight mechanism, we could support
2622 * any file. For now, just ensure that anything potentially problematic is done
2625 static bool __io_file_supports_async(struct file *file, int rw)
2627 umode_t mode = file_inode(file)->i_mode;
2629 if (S_ISBLK(mode)) {
2630 if (IS_ENABLED(CONFIG_BLOCK) &&
2631 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2637 if (S_ISREG(mode)) {
2638 if (IS_ENABLED(CONFIG_BLOCK) &&
2639 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2640 file->f_op != &io_uring_fops)
2645 /* any ->read/write should understand O_NONBLOCK */
2646 if (file->f_flags & O_NONBLOCK)
2649 if (!(file->f_mode & FMODE_NOWAIT))
2653 return file->f_op->read_iter != NULL;
2655 return file->f_op->write_iter != NULL;
2658 static bool io_file_supports_async(struct io_kiocb *req, int rw)
2660 if (rw == READ && (req->flags & REQ_F_ASYNC_READ))
2662 else if (rw == WRITE && (req->flags & REQ_F_ASYNC_WRITE))
2665 return __io_file_supports_async(req->file, rw);
2668 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2670 struct io_ring_ctx *ctx = req->ctx;
2671 struct kiocb *kiocb = &req->rw.kiocb;
2672 struct file *file = req->file;
2676 if (!(req->flags & REQ_F_ISREG) && S_ISREG(file_inode(file)->i_mode))
2677 req->flags |= REQ_F_ISREG;
2679 kiocb->ki_pos = READ_ONCE(sqe->off);
2680 if (kiocb->ki_pos == -1 && !(file->f_mode & FMODE_STREAM)) {
2681 req->flags |= REQ_F_CUR_POS;
2682 kiocb->ki_pos = file->f_pos;
2684 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2685 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2686 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2690 /* don't allow async punt for O_NONBLOCK or RWF_NOWAIT */
2691 if ((kiocb->ki_flags & IOCB_NOWAIT) || (file->f_flags & O_NONBLOCK))
2692 req->flags |= REQ_F_NOWAIT;
2694 ioprio = READ_ONCE(sqe->ioprio);
2696 ret = ioprio_check_cap(ioprio);
2700 kiocb->ki_ioprio = ioprio;
2702 kiocb->ki_ioprio = get_current_ioprio();
2704 if (ctx->flags & IORING_SETUP_IOPOLL) {
2705 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2706 !kiocb->ki_filp->f_op->iopoll)
2709 kiocb->ki_flags |= IOCB_HIPRI;
2710 kiocb->ki_complete = io_complete_rw_iopoll;
2711 req->iopoll_completed = 0;
2713 if (kiocb->ki_flags & IOCB_HIPRI)
2715 kiocb->ki_complete = io_complete_rw;
2718 if (req->opcode == IORING_OP_READ_FIXED ||
2719 req->opcode == IORING_OP_WRITE_FIXED) {
2721 io_req_set_rsrc_node(req);
2724 req->rw.addr = READ_ONCE(sqe->addr);
2725 req->rw.len = READ_ONCE(sqe->len);
2726 req->buf_index = READ_ONCE(sqe->buf_index);
2730 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2736 case -ERESTARTNOINTR:
2737 case -ERESTARTNOHAND:
2738 case -ERESTART_RESTARTBLOCK:
2740 * We can't just restart the syscall, since previously
2741 * submitted sqes may already be in progress. Just fail this
2747 kiocb->ki_complete(kiocb, ret, 0);
2751 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2752 unsigned int issue_flags)
2754 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2755 struct io_async_rw *io = req->async_data;
2756 bool check_reissue = kiocb->ki_complete == io_complete_rw;
2758 /* add previously done IO, if any */
2759 if (io && io->bytes_done > 0) {
2761 ret = io->bytes_done;
2763 ret += io->bytes_done;
2766 if (req->flags & REQ_F_CUR_POS)
2767 req->file->f_pos = kiocb->ki_pos;
2768 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2769 __io_complete_rw(req, ret, 0, issue_flags);
2771 io_rw_done(kiocb, ret);
2773 if (check_reissue && req->flags & REQ_F_REISSUE) {
2774 req->flags &= ~REQ_F_REISSUE;
2775 if (io_resubmit_prep(req)) {
2777 io_queue_async_work(req);
2782 if (req->flags & REQ_F_BUFFER_SELECTED)
2783 cflags = io_put_rw_kbuf(req);
2784 __io_req_complete(req, issue_flags, ret, cflags);
2789 static int __io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter,
2790 struct io_mapped_ubuf *imu)
2792 size_t len = req->rw.len;
2793 u64 buf_end, buf_addr = req->rw.addr;
2796 if (unlikely(check_add_overflow(buf_addr, (u64)len, &buf_end)))
2798 /* not inside the mapped region */
2799 if (unlikely(buf_addr < imu->ubuf || buf_end > imu->ubuf_end))
2803 * May not be a start of buffer, set size appropriately
2804 * and advance us to the beginning.
2806 offset = buf_addr - imu->ubuf;
2807 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2811 * Don't use iov_iter_advance() here, as it's really slow for
2812 * using the latter parts of a big fixed buffer - it iterates
2813 * over each segment manually. We can cheat a bit here, because
2816 * 1) it's a BVEC iter, we set it up
2817 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2818 * first and last bvec
2820 * So just find our index, and adjust the iterator afterwards.
2821 * If the offset is within the first bvec (or the whole first
2822 * bvec, just use iov_iter_advance(). This makes it easier
2823 * since we can just skip the first segment, which may not
2824 * be PAGE_SIZE aligned.
2826 const struct bio_vec *bvec = imu->bvec;
2828 if (offset <= bvec->bv_len) {
2829 iov_iter_advance(iter, offset);
2831 unsigned long seg_skip;
2833 /* skip first vec */
2834 offset -= bvec->bv_len;
2835 seg_skip = 1 + (offset >> PAGE_SHIFT);
2837 iter->bvec = bvec + seg_skip;
2838 iter->nr_segs -= seg_skip;
2839 iter->count -= bvec->bv_len + offset;
2840 iter->iov_offset = offset & ~PAGE_MASK;
2847 static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
2849 struct io_ring_ctx *ctx = req->ctx;
2850 struct io_mapped_ubuf *imu = req->imu;
2851 u16 index, buf_index = req->buf_index;
2854 if (unlikely(buf_index >= ctx->nr_user_bufs))
2856 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2857 imu = READ_ONCE(ctx->user_bufs[index]);
2860 return __io_import_fixed(req, rw, iter, imu);
2863 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2866 mutex_unlock(&ctx->uring_lock);
2869 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2872 * "Normal" inline submissions always hold the uring_lock, since we
2873 * grab it from the system call. Same is true for the SQPOLL offload.
2874 * The only exception is when we've detached the request and issue it
2875 * from an async worker thread, grab the lock for that case.
2878 mutex_lock(&ctx->uring_lock);
2881 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2882 int bgid, struct io_buffer *kbuf,
2885 struct io_buffer *head;
2887 if (req->flags & REQ_F_BUFFER_SELECTED)
2890 io_ring_submit_lock(req->ctx, needs_lock);
2892 lockdep_assert_held(&req->ctx->uring_lock);
2894 head = xa_load(&req->ctx->io_buffers, bgid);
2896 if (!list_empty(&head->list)) {
2897 kbuf = list_last_entry(&head->list, struct io_buffer,
2899 list_del(&kbuf->list);
2902 xa_erase(&req->ctx->io_buffers, bgid);
2904 if (*len > kbuf->len)
2907 kbuf = ERR_PTR(-ENOBUFS);
2910 io_ring_submit_unlock(req->ctx, needs_lock);
2915 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2918 struct io_buffer *kbuf;
2921 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2922 bgid = req->buf_index;
2923 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2926 req->rw.addr = (u64) (unsigned long) kbuf;
2927 req->flags |= REQ_F_BUFFER_SELECTED;
2928 return u64_to_user_ptr(kbuf->addr);
2931 #ifdef CONFIG_COMPAT
2932 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2935 struct compat_iovec __user *uiov;
2936 compat_ssize_t clen;
2940 uiov = u64_to_user_ptr(req->rw.addr);
2941 if (!access_ok(uiov, sizeof(*uiov)))
2943 if (__get_user(clen, &uiov->iov_len))
2949 buf = io_rw_buffer_select(req, &len, needs_lock);
2951 return PTR_ERR(buf);
2952 iov[0].iov_base = buf;
2953 iov[0].iov_len = (compat_size_t) len;
2958 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2961 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2965 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2968 len = iov[0].iov_len;
2971 buf = io_rw_buffer_select(req, &len, needs_lock);
2973 return PTR_ERR(buf);
2974 iov[0].iov_base = buf;
2975 iov[0].iov_len = len;
2979 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2982 if (req->flags & REQ_F_BUFFER_SELECTED) {
2983 struct io_buffer *kbuf;
2985 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2986 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
2987 iov[0].iov_len = kbuf->len;
2990 if (req->rw.len != 1)
2993 #ifdef CONFIG_COMPAT
2994 if (req->ctx->compat)
2995 return io_compat_import(req, iov, needs_lock);
2998 return __io_iov_buffer_select(req, iov, needs_lock);
3001 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
3002 struct iov_iter *iter, bool needs_lock)
3004 void __user *buf = u64_to_user_ptr(req->rw.addr);
3005 size_t sqe_len = req->rw.len;
3006 u8 opcode = req->opcode;
3009 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3011 return io_import_fixed(req, rw, iter);
3014 /* buffer index only valid with fixed read/write, or buffer select */
3015 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3018 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3019 if (req->flags & REQ_F_BUFFER_SELECT) {
3020 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3022 return PTR_ERR(buf);
3023 req->rw.len = sqe_len;
3026 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3031 if (req->flags & REQ_F_BUFFER_SELECT) {
3032 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3034 iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
3039 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3043 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3045 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3049 * For files that don't have ->read_iter() and ->write_iter(), handle them
3050 * by looping over ->read() or ->write() manually.
3052 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3054 struct kiocb *kiocb = &req->rw.kiocb;
3055 struct file *file = req->file;
3059 * Don't support polled IO through this interface, and we can't
3060 * support non-blocking either. For the latter, this just causes
3061 * the kiocb to be handled from an async context.
3063 if (kiocb->ki_flags & IOCB_HIPRI)
3065 if (kiocb->ki_flags & IOCB_NOWAIT)
3068 while (iov_iter_count(iter)) {
3072 if (!iov_iter_is_bvec(iter)) {
3073 iovec = iov_iter_iovec(iter);
3075 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3076 iovec.iov_len = req->rw.len;
3080 nr = file->f_op->read(file, iovec.iov_base,
3081 iovec.iov_len, io_kiocb_ppos(kiocb));
3083 nr = file->f_op->write(file, iovec.iov_base,
3084 iovec.iov_len, io_kiocb_ppos(kiocb));
3093 if (nr != iovec.iov_len)
3097 iov_iter_advance(iter, nr);
3103 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3104 const struct iovec *fast_iov, struct iov_iter *iter)
3106 struct io_async_rw *rw = req->async_data;
3108 memcpy(&rw->iter, iter, sizeof(*iter));
3109 rw->free_iovec = iovec;
3111 /* can only be fixed buffers, no need to do anything */
3112 if (iov_iter_is_bvec(iter))
3115 unsigned iov_off = 0;
3117 rw->iter.iov = rw->fast_iov;
3118 if (iter->iov != fast_iov) {
3119 iov_off = iter->iov - fast_iov;
3120 rw->iter.iov += iov_off;
3122 if (rw->fast_iov != fast_iov)
3123 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3124 sizeof(struct iovec) * iter->nr_segs);
3126 req->flags |= REQ_F_NEED_CLEANUP;
3130 static inline int io_alloc_async_data(struct io_kiocb *req)
3132 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3133 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3134 return req->async_data == NULL;
3137 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3138 const struct iovec *fast_iov,
3139 struct iov_iter *iter, bool force)
3141 if (!force && !io_op_defs[req->opcode].needs_async_setup)
3143 if (!req->async_data) {
3144 if (io_alloc_async_data(req)) {
3149 io_req_map_rw(req, iovec, fast_iov, iter);
3154 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3156 struct io_async_rw *iorw = req->async_data;
3157 struct iovec *iov = iorw->fast_iov;
3160 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3161 if (unlikely(ret < 0))
3164 iorw->bytes_done = 0;
3165 iorw->free_iovec = iov;
3167 req->flags |= REQ_F_NEED_CLEANUP;
3171 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3173 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3175 return io_prep_rw(req, sqe);
3179 * This is our waitqueue callback handler, registered through lock_page_async()
3180 * when we initially tried to do the IO with the iocb armed our waitqueue.
3181 * This gets called when the page is unlocked, and we generally expect that to
3182 * happen when the page IO is completed and the page is now uptodate. This will
3183 * queue a task_work based retry of the operation, attempting to copy the data
3184 * again. If the latter fails because the page was NOT uptodate, then we will
3185 * do a thread based blocking retry of the operation. That's the unexpected
3188 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3189 int sync, void *arg)
3191 struct wait_page_queue *wpq;
3192 struct io_kiocb *req = wait->private;
3193 struct wait_page_key *key = arg;
3195 wpq = container_of(wait, struct wait_page_queue, wait);
3197 if (!wake_page_match(wpq, key))
3200 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3201 list_del_init(&wait->entry);
3203 /* submit ref gets dropped, acquire a new one */
3205 io_req_task_queue(req);
3210 * This controls whether a given IO request should be armed for async page
3211 * based retry. If we return false here, the request is handed to the async
3212 * worker threads for retry. If we're doing buffered reads on a regular file,
3213 * we prepare a private wait_page_queue entry and retry the operation. This
3214 * will either succeed because the page is now uptodate and unlocked, or it
3215 * will register a callback when the page is unlocked at IO completion. Through
3216 * that callback, io_uring uses task_work to setup a retry of the operation.
3217 * That retry will attempt the buffered read again. The retry will generally
3218 * succeed, or in rare cases where it fails, we then fall back to using the
3219 * async worker threads for a blocking retry.
3221 static bool io_rw_should_retry(struct io_kiocb *req)
3223 struct io_async_rw *rw = req->async_data;
3224 struct wait_page_queue *wait = &rw->wpq;
3225 struct kiocb *kiocb = &req->rw.kiocb;
3227 /* never retry for NOWAIT, we just complete with -EAGAIN */
3228 if (req->flags & REQ_F_NOWAIT)
3231 /* Only for buffered IO */
3232 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3236 * just use poll if we can, and don't attempt if the fs doesn't
3237 * support callback based unlocks
3239 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3242 wait->wait.func = io_async_buf_func;
3243 wait->wait.private = req;
3244 wait->wait.flags = 0;
3245 INIT_LIST_HEAD(&wait->wait.entry);
3246 kiocb->ki_flags |= IOCB_WAITQ;
3247 kiocb->ki_flags &= ~IOCB_NOWAIT;
3248 kiocb->ki_waitq = wait;
3252 static inline int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3254 if (req->file->f_op->read_iter)
3255 return call_read_iter(req->file, &req->rw.kiocb, iter);
3256 else if (req->file->f_op->read)
3257 return loop_rw_iter(READ, req, iter);
3262 static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3264 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3265 struct kiocb *kiocb = &req->rw.kiocb;
3266 struct iov_iter __iter, *iter = &__iter;
3267 struct io_async_rw *rw = req->async_data;
3268 ssize_t io_size, ret, ret2;
3269 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3275 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3279 io_size = iov_iter_count(iter);
3280 req->result = io_size;
3282 /* Ensure we clear previously set non-block flag */
3283 if (!force_nonblock)
3284 kiocb->ki_flags &= ~IOCB_NOWAIT;
3286 kiocb->ki_flags |= IOCB_NOWAIT;
3288 /* If the file doesn't support async, just async punt */
3289 if (force_nonblock && !io_file_supports_async(req, READ)) {
3290 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3291 return ret ?: -EAGAIN;
3294 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size);
3295 if (unlikely(ret)) {
3300 ret = io_iter_do_read(req, iter);
3302 if (ret == -EAGAIN || (req->flags & REQ_F_REISSUE)) {
3303 req->flags &= ~REQ_F_REISSUE;
3304 /* IOPOLL retry should happen for io-wq threads */
3305 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3307 /* no retry on NONBLOCK nor RWF_NOWAIT */
3308 if (req->flags & REQ_F_NOWAIT)
3310 /* some cases will consume bytes even on error returns */
3311 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3313 } else if (ret == -EIOCBQUEUED) {
3315 } else if (ret <= 0 || ret == io_size || !force_nonblock ||
3316 (req->flags & REQ_F_NOWAIT) || !(req->flags & REQ_F_ISREG)) {
3317 /* read all, failed, already did sync or don't want to retry */
3321 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3326 rw = req->async_data;
3327 /* now use our persistent iterator, if we aren't already */
3332 rw->bytes_done += ret;
3333 /* if we can retry, do so with the callbacks armed */
3334 if (!io_rw_should_retry(req)) {
3335 kiocb->ki_flags &= ~IOCB_WAITQ;
3340 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3341 * we get -EIOCBQUEUED, then we'll get a notification when the
3342 * desired page gets unlocked. We can also get a partial read
3343 * here, and if we do, then just retry at the new offset.
3345 ret = io_iter_do_read(req, iter);
3346 if (ret == -EIOCBQUEUED)
3348 /* we got some bytes, but not all. retry. */
3349 kiocb->ki_flags &= ~IOCB_WAITQ;
3350 } while (ret > 0 && ret < io_size);
3352 kiocb_done(kiocb, ret, issue_flags);
3354 /* it's faster to check here then delegate to kfree */
3360 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3362 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3364 return io_prep_rw(req, sqe);
3367 static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3369 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3370 struct kiocb *kiocb = &req->rw.kiocb;
3371 struct iov_iter __iter, *iter = &__iter;
3372 struct io_async_rw *rw = req->async_data;
3373 ssize_t ret, ret2, io_size;
3374 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3380 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3384 io_size = iov_iter_count(iter);
3385 req->result = io_size;
3387 /* Ensure we clear previously set non-block flag */
3388 if (!force_nonblock)
3389 kiocb->ki_flags &= ~IOCB_NOWAIT;
3391 kiocb->ki_flags |= IOCB_NOWAIT;
3393 /* If the file doesn't support async, just async punt */
3394 if (force_nonblock && !io_file_supports_async(req, WRITE))
3397 /* file path doesn't support NOWAIT for non-direct_IO */
3398 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3399 (req->flags & REQ_F_ISREG))
3402 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size);
3407 * Open-code file_start_write here to grab freeze protection,
3408 * which will be released by another thread in
3409 * io_complete_rw(). Fool lockdep by telling it the lock got
3410 * released so that it doesn't complain about the held lock when
3411 * we return to userspace.
3413 if (req->flags & REQ_F_ISREG) {
3414 sb_start_write(file_inode(req->file)->i_sb);
3415 __sb_writers_release(file_inode(req->file)->i_sb,
3418 kiocb->ki_flags |= IOCB_WRITE;
3420 if (req->file->f_op->write_iter)
3421 ret2 = call_write_iter(req->file, kiocb, iter);
3422 else if (req->file->f_op->write)
3423 ret2 = loop_rw_iter(WRITE, req, iter);
3427 if (req->flags & REQ_F_REISSUE) {
3428 req->flags &= ~REQ_F_REISSUE;
3433 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3434 * retry them without IOCB_NOWAIT.
3436 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3438 /* no retry on NONBLOCK nor RWF_NOWAIT */
3439 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3441 if (!force_nonblock || ret2 != -EAGAIN) {
3442 /* IOPOLL retry should happen for io-wq threads */
3443 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3446 kiocb_done(kiocb, ret2, issue_flags);
3449 /* some cases will consume bytes even on error returns */
3450 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3451 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3452 return ret ?: -EAGAIN;
3455 /* it's reportedly faster than delegating the null check to kfree() */
3461 static int io_renameat_prep(struct io_kiocb *req,
3462 const struct io_uring_sqe *sqe)
3464 struct io_rename *ren = &req->rename;
3465 const char __user *oldf, *newf;
3467 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3470 ren->old_dfd = READ_ONCE(sqe->fd);
3471 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3472 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3473 ren->new_dfd = READ_ONCE(sqe->len);
3474 ren->flags = READ_ONCE(sqe->rename_flags);
3476 ren->oldpath = getname(oldf);
3477 if (IS_ERR(ren->oldpath))
3478 return PTR_ERR(ren->oldpath);
3480 ren->newpath = getname(newf);
3481 if (IS_ERR(ren->newpath)) {
3482 putname(ren->oldpath);
3483 return PTR_ERR(ren->newpath);
3486 req->flags |= REQ_F_NEED_CLEANUP;
3490 static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3492 struct io_rename *ren = &req->rename;
3495 if (issue_flags & IO_URING_F_NONBLOCK)
3498 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3499 ren->newpath, ren->flags);
3501 req->flags &= ~REQ_F_NEED_CLEANUP;
3504 io_req_complete(req, ret);
3508 static int io_unlinkat_prep(struct io_kiocb *req,
3509 const struct io_uring_sqe *sqe)
3511 struct io_unlink *un = &req->unlink;
3512 const char __user *fname;
3514 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3517 un->dfd = READ_ONCE(sqe->fd);
3519 un->flags = READ_ONCE(sqe->unlink_flags);
3520 if (un->flags & ~AT_REMOVEDIR)
3523 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3524 un->filename = getname(fname);
3525 if (IS_ERR(un->filename))
3526 return PTR_ERR(un->filename);
3528 req->flags |= REQ_F_NEED_CLEANUP;
3532 static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3534 struct io_unlink *un = &req->unlink;
3537 if (issue_flags & IO_URING_F_NONBLOCK)
3540 if (un->flags & AT_REMOVEDIR)
3541 ret = do_rmdir(un->dfd, un->filename);
3543 ret = do_unlinkat(un->dfd, un->filename);
3545 req->flags &= ~REQ_F_NEED_CLEANUP;
3548 io_req_complete(req, ret);
3552 static int io_shutdown_prep(struct io_kiocb *req,
3553 const struct io_uring_sqe *sqe)
3555 #if defined(CONFIG_NET)
3556 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3558 if (sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3562 req->shutdown.how = READ_ONCE(sqe->len);
3569 static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3571 #if defined(CONFIG_NET)
3572 struct socket *sock;
3575 if (issue_flags & IO_URING_F_NONBLOCK)
3578 sock = sock_from_file(req->file);
3579 if (unlikely(!sock))
3582 ret = __sys_shutdown_sock(sock, req->shutdown.how);
3585 io_req_complete(req, ret);
3592 static int __io_splice_prep(struct io_kiocb *req,
3593 const struct io_uring_sqe *sqe)
3595 struct io_splice* sp = &req->splice;
3596 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3598 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3602 sp->len = READ_ONCE(sqe->len);
3603 sp->flags = READ_ONCE(sqe->splice_flags);
3605 if (unlikely(sp->flags & ~valid_flags))
3608 sp->file_in = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in),
3609 (sp->flags & SPLICE_F_FD_IN_FIXED));
3612 req->flags |= REQ_F_NEED_CLEANUP;
3616 static int io_tee_prep(struct io_kiocb *req,
3617 const struct io_uring_sqe *sqe)
3619 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3621 return __io_splice_prep(req, sqe);
3624 static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
3626 struct io_splice *sp = &req->splice;
3627 struct file *in = sp->file_in;
3628 struct file *out = sp->file_out;
3629 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3632 if (issue_flags & IO_URING_F_NONBLOCK)
3635 ret = do_tee(in, out, sp->len, flags);
3637 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3639 req->flags &= ~REQ_F_NEED_CLEANUP;
3643 io_req_complete(req, ret);
3647 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3649 struct io_splice* sp = &req->splice;
3651 sp->off_in = READ_ONCE(sqe->splice_off_in);
3652 sp->off_out = READ_ONCE(sqe->off);
3653 return __io_splice_prep(req, sqe);
3656 static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
3658 struct io_splice *sp = &req->splice;
3659 struct file *in = sp->file_in;
3660 struct file *out = sp->file_out;
3661 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3662 loff_t *poff_in, *poff_out;
3665 if (issue_flags & IO_URING_F_NONBLOCK)
3668 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3669 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3672 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3674 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3676 req->flags &= ~REQ_F_NEED_CLEANUP;
3680 io_req_complete(req, ret);
3685 * IORING_OP_NOP just posts a completion event, nothing else.
3687 static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
3689 struct io_ring_ctx *ctx = req->ctx;
3691 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3694 __io_req_complete(req, issue_flags, 0, 0);
3698 static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3700 struct io_ring_ctx *ctx = req->ctx;
3705 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3707 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3710 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3711 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3714 req->sync.off = READ_ONCE(sqe->off);
3715 req->sync.len = READ_ONCE(sqe->len);
3719 static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
3721 loff_t end = req->sync.off + req->sync.len;
3724 /* fsync always requires a blocking context */
3725 if (issue_flags & IO_URING_F_NONBLOCK)
3728 ret = vfs_fsync_range(req->file, req->sync.off,
3729 end > 0 ? end : LLONG_MAX,
3730 req->sync.flags & IORING_FSYNC_DATASYNC);
3733 io_req_complete(req, ret);
3737 static int io_fallocate_prep(struct io_kiocb *req,
3738 const struct io_uring_sqe *sqe)
3740 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3742 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3745 req->sync.off = READ_ONCE(sqe->off);
3746 req->sync.len = READ_ONCE(sqe->addr);
3747 req->sync.mode = READ_ONCE(sqe->len);
3751 static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
3755 /* fallocate always requiring blocking context */
3756 if (issue_flags & IO_URING_F_NONBLOCK)
3758 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3762 io_req_complete(req, ret);
3766 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3768 const char __user *fname;
3771 if (unlikely(sqe->ioprio || sqe->buf_index))
3773 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3776 /* open.how should be already initialised */
3777 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3778 req->open.how.flags |= O_LARGEFILE;
3780 req->open.dfd = READ_ONCE(sqe->fd);
3781 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3782 req->open.filename = getname(fname);
3783 if (IS_ERR(req->open.filename)) {
3784 ret = PTR_ERR(req->open.filename);
3785 req->open.filename = NULL;
3788 req->open.nofile = rlimit(RLIMIT_NOFILE);
3789 req->flags |= REQ_F_NEED_CLEANUP;
3793 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3797 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3799 mode = READ_ONCE(sqe->len);
3800 flags = READ_ONCE(sqe->open_flags);
3801 req->open.how = build_open_how(flags, mode);
3802 return __io_openat_prep(req, sqe);
3805 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3807 struct open_how __user *how;
3811 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3813 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3814 len = READ_ONCE(sqe->len);
3815 if (len < OPEN_HOW_SIZE_VER0)
3818 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3823 return __io_openat_prep(req, sqe);
3826 static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
3828 struct open_flags op;
3831 bool resolve_nonblock;
3834 ret = build_open_flags(&req->open.how, &op);
3837 nonblock_set = op.open_flag & O_NONBLOCK;
3838 resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
3839 if (issue_flags & IO_URING_F_NONBLOCK) {
3841 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
3842 * it'll always -EAGAIN
3844 if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
3846 op.lookup_flags |= LOOKUP_CACHED;
3847 op.open_flag |= O_NONBLOCK;
3850 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3854 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3855 /* only retry if RESOLVE_CACHED wasn't already set by application */
3856 if ((!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)) &&
3857 file == ERR_PTR(-EAGAIN)) {
3859 * We could hang on to this 'fd', but seems like marginal
3860 * gain for something that is now known to be a slower path.
3861 * So just put it, and we'll get a new one when we retry.
3869 ret = PTR_ERR(file);
3871 if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
3872 file->f_flags &= ~O_NONBLOCK;
3873 fsnotify_open(file);
3874 fd_install(ret, file);
3877 putname(req->open.filename);
3878 req->flags &= ~REQ_F_NEED_CLEANUP;
3881 __io_req_complete(req, issue_flags, ret, 0);
3885 static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
3887 return io_openat2(req, issue_flags);
3890 static int io_remove_buffers_prep(struct io_kiocb *req,
3891 const struct io_uring_sqe *sqe)
3893 struct io_provide_buf *p = &req->pbuf;
3896 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3899 tmp = READ_ONCE(sqe->fd);
3900 if (!tmp || tmp > USHRT_MAX)
3903 memset(p, 0, sizeof(*p));
3905 p->bgid = READ_ONCE(sqe->buf_group);
3909 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3910 int bgid, unsigned nbufs)
3914 /* shouldn't happen */
3918 /* the head kbuf is the list itself */
3919 while (!list_empty(&buf->list)) {
3920 struct io_buffer *nxt;
3922 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3923 list_del(&nxt->list);
3930 xa_erase(&ctx->io_buffers, bgid);
3935 static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
3937 struct io_provide_buf *p = &req->pbuf;
3938 struct io_ring_ctx *ctx = req->ctx;
3939 struct io_buffer *head;
3941 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3943 io_ring_submit_lock(ctx, !force_nonblock);
3945 lockdep_assert_held(&ctx->uring_lock);
3948 head = xa_load(&ctx->io_buffers, p->bgid);
3950 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3954 /* complete before unlock, IOPOLL may need the lock */
3955 __io_req_complete(req, issue_flags, ret, 0);
3956 io_ring_submit_unlock(ctx, !force_nonblock);
3960 static int io_provide_buffers_prep(struct io_kiocb *req,
3961 const struct io_uring_sqe *sqe)
3963 unsigned long size, tmp_check;
3964 struct io_provide_buf *p = &req->pbuf;
3967 if (sqe->ioprio || sqe->rw_flags)
3970 tmp = READ_ONCE(sqe->fd);
3971 if (!tmp || tmp > USHRT_MAX)
3974 p->addr = READ_ONCE(sqe->addr);
3975 p->len = READ_ONCE(sqe->len);
3977 if (check_mul_overflow((unsigned long)p->len, (unsigned long)p->nbufs,
3980 if (check_add_overflow((unsigned long)p->addr, size, &tmp_check))
3983 size = (unsigned long)p->len * p->nbufs;
3984 if (!access_ok(u64_to_user_ptr(p->addr), size))
3987 p->bgid = READ_ONCE(sqe->buf_group);
3988 tmp = READ_ONCE(sqe->off);
3989 if (tmp > USHRT_MAX)
3995 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3997 struct io_buffer *buf;
3998 u64 addr = pbuf->addr;
3999 int i, bid = pbuf->bid;
4001 for (i = 0; i < pbuf->nbufs; i++) {
4002 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
4007 buf->len = min_t(__u32, pbuf->len, MAX_RW_COUNT);
4012 INIT_LIST_HEAD(&buf->list);
4015 list_add_tail(&buf->list, &(*head)->list);
4019 return i ? i : -ENOMEM;
4022 static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
4024 struct io_provide_buf *p = &req->pbuf;
4025 struct io_ring_ctx *ctx = req->ctx;
4026 struct io_buffer *head, *list;
4028 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4030 io_ring_submit_lock(ctx, !force_nonblock);
4032 lockdep_assert_held(&ctx->uring_lock);
4034 list = head = xa_load(&ctx->io_buffers, p->bgid);
4036 ret = io_add_buffers(p, &head);
4037 if (ret >= 0 && !list) {
4038 ret = xa_insert(&ctx->io_buffers, p->bgid, head, GFP_KERNEL);
4040 __io_remove_buffers(ctx, head, p->bgid, -1U);
4044 /* complete before unlock, IOPOLL may need the lock */
4045 __io_req_complete(req, issue_flags, ret, 0);
4046 io_ring_submit_unlock(ctx, !force_nonblock);
4050 static int io_epoll_ctl_prep(struct io_kiocb *req,
4051 const struct io_uring_sqe *sqe)
4053 #if defined(CONFIG_EPOLL)
4054 if (sqe->ioprio || sqe->buf_index)
4056 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4059 req->epoll.epfd = READ_ONCE(sqe->fd);
4060 req->epoll.op = READ_ONCE(sqe->len);
4061 req->epoll.fd = READ_ONCE(sqe->off);
4063 if (ep_op_has_event(req->epoll.op)) {
4064 struct epoll_event __user *ev;
4066 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4067 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4077 static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4079 #if defined(CONFIG_EPOLL)
4080 struct io_epoll *ie = &req->epoll;
4082 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4084 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4085 if (force_nonblock && ret == -EAGAIN)
4090 __io_req_complete(req, issue_flags, ret, 0);
4097 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4099 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4100 if (sqe->ioprio || sqe->buf_index || sqe->off)
4102 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4105 req->madvise.addr = READ_ONCE(sqe->addr);
4106 req->madvise.len = READ_ONCE(sqe->len);
4107 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4114 static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4116 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4117 struct io_madvise *ma = &req->madvise;
4120 if (issue_flags & IO_URING_F_NONBLOCK)
4123 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4126 io_req_complete(req, ret);
4133 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4135 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4137 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4140 req->fadvise.offset = READ_ONCE(sqe->off);
4141 req->fadvise.len = READ_ONCE(sqe->len);
4142 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4146 static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4148 struct io_fadvise *fa = &req->fadvise;
4151 if (issue_flags & IO_URING_F_NONBLOCK) {
4152 switch (fa->advice) {
4153 case POSIX_FADV_NORMAL:
4154 case POSIX_FADV_RANDOM:
4155 case POSIX_FADV_SEQUENTIAL:
4162 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4165 __io_req_complete(req, issue_flags, ret, 0);
4169 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4171 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4173 if (sqe->ioprio || sqe->buf_index)
4175 if (req->flags & REQ_F_FIXED_FILE)
4178 req->statx.dfd = READ_ONCE(sqe->fd);
4179 req->statx.mask = READ_ONCE(sqe->len);
4180 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4181 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4182 req->statx.flags = READ_ONCE(sqe->statx_flags);
4187 static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4189 struct io_statx *ctx = &req->statx;
4192 if (issue_flags & IO_URING_F_NONBLOCK)
4195 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4200 io_req_complete(req, ret);
4204 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4206 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4208 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4209 sqe->rw_flags || sqe->buf_index)
4211 if (req->flags & REQ_F_FIXED_FILE)
4214 req->close.fd = READ_ONCE(sqe->fd);
4218 static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4220 struct files_struct *files = current->files;
4221 struct io_close *close = &req->close;
4222 struct fdtable *fdt;
4223 struct file *file = NULL;
4226 spin_lock(&files->file_lock);
4227 fdt = files_fdtable(files);
4228 if (close->fd >= fdt->max_fds) {
4229 spin_unlock(&files->file_lock);
4232 file = fdt->fd[close->fd];
4233 if (!file || file->f_op == &io_uring_fops) {
4234 spin_unlock(&files->file_lock);
4239 /* if the file has a flush method, be safe and punt to async */
4240 if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4241 spin_unlock(&files->file_lock);
4245 ret = __close_fd_get_file(close->fd, &file);
4246 spin_unlock(&files->file_lock);
4253 /* No ->flush() or already async, safely close from here */
4254 ret = filp_close(file, current->files);
4260 __io_req_complete(req, issue_flags, ret, 0);
4264 static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4266 struct io_ring_ctx *ctx = req->ctx;
4268 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4270 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4273 req->sync.off = READ_ONCE(sqe->off);
4274 req->sync.len = READ_ONCE(sqe->len);
4275 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4279 static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4283 /* sync_file_range always requires a blocking context */
4284 if (issue_flags & IO_URING_F_NONBLOCK)
4287 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4291 io_req_complete(req, ret);
4295 #if defined(CONFIG_NET)
4296 static int io_setup_async_msg(struct io_kiocb *req,
4297 struct io_async_msghdr *kmsg)
4299 struct io_async_msghdr *async_msg = req->async_data;
4303 if (io_alloc_async_data(req)) {
4304 kfree(kmsg->free_iov);
4307 async_msg = req->async_data;
4308 req->flags |= REQ_F_NEED_CLEANUP;
4309 memcpy(async_msg, kmsg, sizeof(*kmsg));
4310 async_msg->msg.msg_name = &async_msg->addr;
4311 /* if were using fast_iov, set it to the new one */
4312 if (!async_msg->free_iov)
4313 async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4318 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4319 struct io_async_msghdr *iomsg)
4321 iomsg->msg.msg_name = &iomsg->addr;
4322 iomsg->free_iov = iomsg->fast_iov;
4323 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4324 req->sr_msg.msg_flags, &iomsg->free_iov);
4327 static int io_sendmsg_prep_async(struct io_kiocb *req)
4331 ret = io_sendmsg_copy_hdr(req, req->async_data);
4333 req->flags |= REQ_F_NEED_CLEANUP;
4337 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4339 struct io_sr_msg *sr = &req->sr_msg;
4341 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4344 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4345 sr->len = READ_ONCE(sqe->len);
4346 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4347 if (sr->msg_flags & MSG_DONTWAIT)
4348 req->flags |= REQ_F_NOWAIT;
4350 #ifdef CONFIG_COMPAT
4351 if (req->ctx->compat)
4352 sr->msg_flags |= MSG_CMSG_COMPAT;
4357 static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4359 struct io_async_msghdr iomsg, *kmsg;
4360 struct socket *sock;
4365 sock = sock_from_file(req->file);
4366 if (unlikely(!sock))
4369 kmsg = req->async_data;
4371 ret = io_sendmsg_copy_hdr(req, &iomsg);
4377 flags = req->sr_msg.msg_flags;
4378 if (issue_flags & IO_URING_F_NONBLOCK)
4379 flags |= MSG_DONTWAIT;
4380 if (flags & MSG_WAITALL)
4381 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4383 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4384 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4385 return io_setup_async_msg(req, kmsg);
4386 if (ret == -ERESTARTSYS)
4389 /* fast path, check for non-NULL to avoid function call */
4391 kfree(kmsg->free_iov);
4392 req->flags &= ~REQ_F_NEED_CLEANUP;
4395 __io_req_complete(req, issue_flags, ret, 0);
4399 static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4401 struct io_sr_msg *sr = &req->sr_msg;
4404 struct socket *sock;
4409 sock = sock_from_file(req->file);
4410 if (unlikely(!sock))
4413 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4417 msg.msg_name = NULL;
4418 msg.msg_control = NULL;
4419 msg.msg_controllen = 0;
4420 msg.msg_namelen = 0;
4422 flags = req->sr_msg.msg_flags;
4423 if (issue_flags & IO_URING_F_NONBLOCK)
4424 flags |= MSG_DONTWAIT;
4425 if (flags & MSG_WAITALL)
4426 min_ret = iov_iter_count(&msg.msg_iter);
4428 msg.msg_flags = flags;
4429 ret = sock_sendmsg(sock, &msg);
4430 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4432 if (ret == -ERESTARTSYS)
4437 __io_req_complete(req, issue_flags, ret, 0);
4441 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4442 struct io_async_msghdr *iomsg)
4444 struct io_sr_msg *sr = &req->sr_msg;
4445 struct iovec __user *uiov;
4449 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4450 &iomsg->uaddr, &uiov, &iov_len);
4454 if (req->flags & REQ_F_BUFFER_SELECT) {
4457 if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4459 sr->len = iomsg->fast_iov[0].iov_len;
4460 iomsg->free_iov = NULL;
4462 iomsg->free_iov = iomsg->fast_iov;
4463 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4464 &iomsg->free_iov, &iomsg->msg.msg_iter,
4473 #ifdef CONFIG_COMPAT
4474 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4475 struct io_async_msghdr *iomsg)
4477 struct io_sr_msg *sr = &req->sr_msg;
4478 struct compat_iovec __user *uiov;
4483 ret = __get_compat_msghdr(&iomsg->msg, sr->umsg_compat, &iomsg->uaddr,
4488 uiov = compat_ptr(ptr);
4489 if (req->flags & REQ_F_BUFFER_SELECT) {
4490 compat_ssize_t clen;
4494 if (!access_ok(uiov, sizeof(*uiov)))
4496 if (__get_user(clen, &uiov->iov_len))
4501 iomsg->free_iov = NULL;
4503 iomsg->free_iov = iomsg->fast_iov;
4504 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4505 UIO_FASTIOV, &iomsg->free_iov,
4506 &iomsg->msg.msg_iter, true);
4515 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4516 struct io_async_msghdr *iomsg)
4518 iomsg->msg.msg_name = &iomsg->addr;
4520 #ifdef CONFIG_COMPAT
4521 if (req->ctx->compat)
4522 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4525 return __io_recvmsg_copy_hdr(req, iomsg);
4528 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4531 struct io_sr_msg *sr = &req->sr_msg;
4532 struct io_buffer *kbuf;
4534 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4539 req->flags |= REQ_F_BUFFER_SELECTED;
4543 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4545 return io_put_kbuf(req, req->sr_msg.kbuf);
4548 static int io_recvmsg_prep_async(struct io_kiocb *req)
4552 ret = io_recvmsg_copy_hdr(req, req->async_data);
4554 req->flags |= REQ_F_NEED_CLEANUP;
4558 static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4560 struct io_sr_msg *sr = &req->sr_msg;
4562 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4565 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4566 sr->len = READ_ONCE(sqe->len);
4567 sr->bgid = READ_ONCE(sqe->buf_group);
4568 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4569 if (sr->msg_flags & MSG_DONTWAIT)
4570 req->flags |= REQ_F_NOWAIT;
4572 #ifdef CONFIG_COMPAT
4573 if (req->ctx->compat)
4574 sr->msg_flags |= MSG_CMSG_COMPAT;
4579 static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
4581 struct io_async_msghdr iomsg, *kmsg;
4582 struct socket *sock;
4583 struct io_buffer *kbuf;
4586 int ret, cflags = 0;
4587 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4589 sock = sock_from_file(req->file);
4590 if (unlikely(!sock))
4593 kmsg = req->async_data;
4595 ret = io_recvmsg_copy_hdr(req, &iomsg);
4601 if (req->flags & REQ_F_BUFFER_SELECT) {
4602 kbuf = io_recv_buffer_select(req, !force_nonblock);
4604 return PTR_ERR(kbuf);
4605 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4606 kmsg->fast_iov[0].iov_len = req->sr_msg.len;
4607 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
4608 1, req->sr_msg.len);
4611 flags = req->sr_msg.msg_flags;
4613 flags |= MSG_DONTWAIT;
4614 if (flags & MSG_WAITALL)
4615 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4617 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4618 kmsg->uaddr, flags);
4619 if (force_nonblock && ret == -EAGAIN)
4620 return io_setup_async_msg(req, kmsg);
4621 if (ret == -ERESTARTSYS)
4624 if (req->flags & REQ_F_BUFFER_SELECTED)
4625 cflags = io_put_recv_kbuf(req);
4626 /* fast path, check for non-NULL to avoid function call */
4628 kfree(kmsg->free_iov);
4629 req->flags &= ~REQ_F_NEED_CLEANUP;
4630 if (ret < min_ret || ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4632 __io_req_complete(req, issue_flags, ret, cflags);
4636 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
4638 struct io_buffer *kbuf;
4639 struct io_sr_msg *sr = &req->sr_msg;
4641 void __user *buf = sr->buf;
4642 struct socket *sock;
4646 int ret, cflags = 0;
4647 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4649 sock = sock_from_file(req->file);
4650 if (unlikely(!sock))
4653 if (req->flags & REQ_F_BUFFER_SELECT) {
4654 kbuf = io_recv_buffer_select(req, !force_nonblock);
4656 return PTR_ERR(kbuf);
4657 buf = u64_to_user_ptr(kbuf->addr);
4660 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4664 msg.msg_name = NULL;
4665 msg.msg_control = NULL;
4666 msg.msg_controllen = 0;
4667 msg.msg_namelen = 0;
4668 msg.msg_iocb = NULL;
4671 flags = req->sr_msg.msg_flags;
4673 flags |= MSG_DONTWAIT;
4674 if (flags & MSG_WAITALL)
4675 min_ret = iov_iter_count(&msg.msg_iter);
4677 ret = sock_recvmsg(sock, &msg, flags);
4678 if (force_nonblock && ret == -EAGAIN)
4680 if (ret == -ERESTARTSYS)
4683 if (req->flags & REQ_F_BUFFER_SELECTED)
4684 cflags = io_put_recv_kbuf(req);
4685 if (ret < min_ret || ((flags & MSG_WAITALL) && (msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4687 __io_req_complete(req, issue_flags, ret, cflags);
4691 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4693 struct io_accept *accept = &req->accept;
4695 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4697 if (sqe->ioprio || sqe->len || sqe->buf_index)
4700 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4701 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4702 accept->flags = READ_ONCE(sqe->accept_flags);
4703 accept->nofile = rlimit(RLIMIT_NOFILE);
4707 static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
4709 struct io_accept *accept = &req->accept;
4710 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4711 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4714 if (req->file->f_flags & O_NONBLOCK)
4715 req->flags |= REQ_F_NOWAIT;
4717 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4718 accept->addr_len, accept->flags,
4720 if (ret == -EAGAIN && force_nonblock)
4723 if (ret == -ERESTARTSYS)
4727 __io_req_complete(req, issue_flags, ret, 0);
4731 static int io_connect_prep_async(struct io_kiocb *req)
4733 struct io_async_connect *io = req->async_data;
4734 struct io_connect *conn = &req->connect;
4736 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
4739 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4741 struct io_connect *conn = &req->connect;
4743 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4745 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4748 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4749 conn->addr_len = READ_ONCE(sqe->addr2);
4753 static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
4755 struct io_async_connect __io, *io;
4756 unsigned file_flags;
4758 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4760 if (req->async_data) {
4761 io = req->async_data;
4763 ret = move_addr_to_kernel(req->connect.addr,
4764 req->connect.addr_len,
4771 file_flags = force_nonblock ? O_NONBLOCK : 0;
4773 ret = __sys_connect_file(req->file, &io->address,
4774 req->connect.addr_len, file_flags);
4775 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4776 if (req->async_data)
4778 if (io_alloc_async_data(req)) {
4782 memcpy(req->async_data, &__io, sizeof(__io));
4785 if (ret == -ERESTARTSYS)
4790 __io_req_complete(req, issue_flags, ret, 0);
4793 #else /* !CONFIG_NET */
4794 #define IO_NETOP_FN(op) \
4795 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
4797 return -EOPNOTSUPP; \
4800 #define IO_NETOP_PREP(op) \
4802 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
4804 return -EOPNOTSUPP; \
4807 #define IO_NETOP_PREP_ASYNC(op) \
4809 static int io_##op##_prep_async(struct io_kiocb *req) \
4811 return -EOPNOTSUPP; \
4814 IO_NETOP_PREP_ASYNC(sendmsg);
4815 IO_NETOP_PREP_ASYNC(recvmsg);
4816 IO_NETOP_PREP_ASYNC(connect);
4817 IO_NETOP_PREP(accept);
4820 #endif /* CONFIG_NET */
4822 struct io_poll_table {
4823 struct poll_table_struct pt;
4824 struct io_kiocb *req;
4828 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4829 __poll_t mask, task_work_func_t func)
4833 /* for instances that support it check for an event match first: */
4834 if (mask && !(mask & poll->events))
4837 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4839 list_del_init(&poll->wait.entry);
4842 req->task_work.func = func;
4845 * If this fails, then the task is exiting. When a task exits, the
4846 * work gets canceled, so just cancel this request as well instead
4847 * of executing it. We can't safely execute it anyway, as we may not
4848 * have the needed state needed for it anyway.
4850 ret = io_req_task_work_add(req);
4851 if (unlikely(ret)) {
4852 WRITE_ONCE(poll->canceled, true);
4853 io_req_task_work_add_fallback(req, func);
4858 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4859 __acquires(&req->ctx->completion_lock)
4861 struct io_ring_ctx *ctx = req->ctx;
4863 if (!req->result && !READ_ONCE(poll->canceled)) {
4864 struct poll_table_struct pt = { ._key = poll->events };
4866 req->result = vfs_poll(req->file, &pt) & poll->events;
4869 spin_lock_irq(&ctx->completion_lock);
4870 if (!req->result && !READ_ONCE(poll->canceled)) {
4871 add_wait_queue(poll->head, &poll->wait);
4878 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4880 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4881 if (req->opcode == IORING_OP_POLL_ADD)
4882 return req->async_data;
4883 return req->apoll->double_poll;
4886 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4888 if (req->opcode == IORING_OP_POLL_ADD)
4890 return &req->apoll->poll;
4893 static void io_poll_remove_double(struct io_kiocb *req)
4894 __must_hold(&req->ctx->completion_lock)
4896 struct io_poll_iocb *poll = io_poll_get_double(req);
4898 lockdep_assert_held(&req->ctx->completion_lock);
4900 if (poll && poll->head) {
4901 struct wait_queue_head *head = poll->head;
4903 spin_lock(&head->lock);
4904 list_del_init(&poll->wait.entry);
4905 if (poll->wait.private)
4908 spin_unlock(&head->lock);
4912 static bool io_poll_complete(struct io_kiocb *req, __poll_t mask)
4913 __must_hold(&req->ctx->completion_lock)
4915 struct io_ring_ctx *ctx = req->ctx;
4916 unsigned flags = IORING_CQE_F_MORE;
4919 if (READ_ONCE(req->poll.canceled)) {
4921 req->poll.events |= EPOLLONESHOT;
4923 error = mangle_poll(mask);
4925 if (req->poll.events & EPOLLONESHOT)
4927 if (!io_cqring_fill_event(ctx, req->user_data, error, flags)) {
4928 io_poll_remove_waitqs(req);
4929 req->poll.done = true;
4932 if (flags & IORING_CQE_F_MORE)
4935 io_commit_cqring(ctx);
4936 return !(flags & IORING_CQE_F_MORE);
4939 static void io_poll_task_func(struct callback_head *cb)
4941 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4942 struct io_ring_ctx *ctx = req->ctx;
4943 struct io_kiocb *nxt;
4945 if (io_poll_rewait(req, &req->poll)) {
4946 spin_unlock_irq(&ctx->completion_lock);
4950 done = io_poll_complete(req, req->result);
4952 hash_del(&req->hash_node);
4955 add_wait_queue(req->poll.head, &req->poll.wait);
4957 spin_unlock_irq(&ctx->completion_lock);
4958 io_cqring_ev_posted(ctx);
4961 nxt = io_put_req_find_next(req);
4963 __io_req_task_submit(nxt);
4968 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4969 int sync, void *key)
4971 struct io_kiocb *req = wait->private;
4972 struct io_poll_iocb *poll = io_poll_get_single(req);
4973 __poll_t mask = key_to_poll(key);
4975 /* for instances that support it check for an event match first: */
4976 if (mask && !(mask & poll->events))
4978 if (!(poll->events & EPOLLONESHOT))
4979 return poll->wait.func(&poll->wait, mode, sync, key);
4981 list_del_init(&wait->entry);
4983 if (poll && poll->head) {
4986 spin_lock(&poll->head->lock);
4987 done = list_empty(&poll->wait.entry);
4989 list_del_init(&poll->wait.entry);
4990 /* make sure double remove sees this as being gone */
4991 wait->private = NULL;
4992 spin_unlock(&poll->head->lock);
4994 /* use wait func handler, so it matches the rq type */
4995 poll->wait.func(&poll->wait, mode, sync, key);
5002 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
5003 wait_queue_func_t wake_func)
5007 poll->canceled = false;
5008 #define IO_POLL_UNMASK (EPOLLERR|EPOLLHUP|EPOLLNVAL|EPOLLRDHUP)
5009 /* mask in events that we always want/need */
5010 poll->events = events | IO_POLL_UNMASK;
5011 INIT_LIST_HEAD(&poll->wait.entry);
5012 init_waitqueue_func_entry(&poll->wait, wake_func);
5015 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5016 struct wait_queue_head *head,
5017 struct io_poll_iocb **poll_ptr)
5019 struct io_kiocb *req = pt->req;
5022 * If poll->head is already set, it's because the file being polled
5023 * uses multiple waitqueues for poll handling (eg one for read, one
5024 * for write). Setup a separate io_poll_iocb if this happens.
5026 if (unlikely(poll->head)) {
5027 struct io_poll_iocb *poll_one = poll;
5029 /* already have a 2nd entry, fail a third attempt */
5031 pt->error = -EINVAL;
5035 * Can't handle multishot for double wait for now, turn it
5036 * into one-shot mode.
5038 if (!(poll_one->events & EPOLLONESHOT))
5039 poll_one->events |= EPOLLONESHOT;
5040 /* double add on the same waitqueue head, ignore */
5041 if (poll_one->head == head)
5043 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5045 pt->error = -ENOMEM;
5048 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5050 poll->wait.private = req;
5057 if (poll->events & EPOLLEXCLUSIVE)
5058 add_wait_queue_exclusive(head, &poll->wait);
5060 add_wait_queue(head, &poll->wait);
5063 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5064 struct poll_table_struct *p)
5066 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5067 struct async_poll *apoll = pt->req->apoll;
5069 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5072 static void io_async_task_func(struct callback_head *cb)
5074 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
5075 struct async_poll *apoll = req->apoll;
5076 struct io_ring_ctx *ctx = req->ctx;
5078 trace_io_uring_task_run(req->ctx, req, req->opcode, req->user_data);
5080 if (io_poll_rewait(req, &apoll->poll)) {
5081 spin_unlock_irq(&ctx->completion_lock);
5085 hash_del(&req->hash_node);
5086 io_poll_remove_double(req);
5087 spin_unlock_irq(&ctx->completion_lock);
5089 if (!READ_ONCE(apoll->poll.canceled))
5090 __io_req_task_submit(req);
5092 io_req_complete_failed(req, -ECANCELED);
5095 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5098 struct io_kiocb *req = wait->private;
5099 struct io_poll_iocb *poll = &req->apoll->poll;
5101 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5104 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5107 static void io_poll_req_insert(struct io_kiocb *req)
5109 struct io_ring_ctx *ctx = req->ctx;
5110 struct hlist_head *list;
5112 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5113 hlist_add_head(&req->hash_node, list);
5116 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5117 struct io_poll_iocb *poll,
5118 struct io_poll_table *ipt, __poll_t mask,
5119 wait_queue_func_t wake_func)
5120 __acquires(&ctx->completion_lock)
5122 struct io_ring_ctx *ctx = req->ctx;
5123 bool cancel = false;
5125 INIT_HLIST_NODE(&req->hash_node);
5126 io_init_poll_iocb(poll, mask, wake_func);
5127 poll->file = req->file;
5128 poll->wait.private = req;
5130 ipt->pt._key = mask;
5132 ipt->error = -EINVAL;
5134 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5136 spin_lock_irq(&ctx->completion_lock);
5137 if (likely(poll->head)) {
5138 spin_lock(&poll->head->lock);
5139 if (unlikely(list_empty(&poll->wait.entry))) {
5145 if ((mask && (poll->events & EPOLLONESHOT)) || ipt->error)
5146 list_del_init(&poll->wait.entry);
5148 WRITE_ONCE(poll->canceled, true);
5149 else if (!poll->done) /* actually waiting for an event */
5150 io_poll_req_insert(req);
5151 spin_unlock(&poll->head->lock);
5157 static bool io_arm_poll_handler(struct io_kiocb *req)
5159 const struct io_op_def *def = &io_op_defs[req->opcode];
5160 struct io_ring_ctx *ctx = req->ctx;
5161 struct async_poll *apoll;
5162 struct io_poll_table ipt;
5166 if (!req->file || !file_can_poll(req->file))
5168 if (req->flags & REQ_F_POLLED)
5172 else if (def->pollout)
5176 /* if we can't nonblock try, then no point in arming a poll handler */
5177 if (!io_file_supports_async(req, rw))
5180 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5181 if (unlikely(!apoll))
5183 apoll->double_poll = NULL;
5185 req->flags |= REQ_F_POLLED;
5188 mask = EPOLLONESHOT;
5190 mask |= POLLIN | POLLRDNORM;
5192 mask |= POLLOUT | POLLWRNORM;
5194 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5195 if ((req->opcode == IORING_OP_RECVMSG) &&
5196 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5199 mask |= POLLERR | POLLPRI;
5201 ipt.pt._qproc = io_async_queue_proc;
5203 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5205 if (ret || ipt.error) {
5206 io_poll_remove_double(req);
5207 spin_unlock_irq(&ctx->completion_lock);
5210 spin_unlock_irq(&ctx->completion_lock);
5211 trace_io_uring_poll_arm(ctx, req, req->opcode, req->user_data,
5212 mask, apoll->poll.events);
5216 static bool __io_poll_remove_one(struct io_kiocb *req,
5217 struct io_poll_iocb *poll, bool do_cancel)
5218 __must_hold(&req->ctx->completion_lock)
5220 bool do_complete = false;
5224 spin_lock(&poll->head->lock);
5226 WRITE_ONCE(poll->canceled, true);
5227 if (!list_empty(&poll->wait.entry)) {
5228 list_del_init(&poll->wait.entry);
5231 spin_unlock(&poll->head->lock);
5232 hash_del(&req->hash_node);
5236 static bool io_poll_remove_waitqs(struct io_kiocb *req)
5237 __must_hold(&req->ctx->completion_lock)
5241 io_poll_remove_double(req);
5242 do_complete = __io_poll_remove_one(req, io_poll_get_single(req), true);
5244 if (req->opcode != IORING_OP_POLL_ADD && do_complete) {
5245 /* non-poll requests have submit ref still */
5251 static bool io_poll_remove_one(struct io_kiocb *req)
5252 __must_hold(&req->ctx->completion_lock)
5256 do_complete = io_poll_remove_waitqs(req);
5258 io_cqring_fill_event(req->ctx, req->user_data, -ECANCELED, 0);
5259 io_commit_cqring(req->ctx);
5261 io_put_req_deferred(req, 1);
5268 * Returns true if we found and killed one or more poll requests
5270 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5273 struct hlist_node *tmp;
5274 struct io_kiocb *req;
5277 spin_lock_irq(&ctx->completion_lock);
5278 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5279 struct hlist_head *list;
5281 list = &ctx->cancel_hash[i];
5282 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5283 if (io_match_task(req, tsk, cancel_all))
5284 posted += io_poll_remove_one(req);
5287 spin_unlock_irq(&ctx->completion_lock);
5290 io_cqring_ev_posted(ctx);
5295 static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, __u64 sqe_addr,
5297 __must_hold(&ctx->completion_lock)
5299 struct hlist_head *list;
5300 struct io_kiocb *req;
5302 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5303 hlist_for_each_entry(req, list, hash_node) {
5304 if (sqe_addr != req->user_data)
5306 if (poll_only && req->opcode != IORING_OP_POLL_ADD)
5313 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr,
5315 __must_hold(&ctx->completion_lock)
5317 struct io_kiocb *req;
5319 req = io_poll_find(ctx, sqe_addr, poll_only);
5322 if (io_poll_remove_one(req))
5328 static __poll_t io_poll_parse_events(const struct io_uring_sqe *sqe,
5333 events = READ_ONCE(sqe->poll32_events);
5335 events = swahw32(events);
5337 if (!(flags & IORING_POLL_ADD_MULTI))
5338 events |= EPOLLONESHOT;
5339 return demangle_poll(events) | (events & (EPOLLEXCLUSIVE|EPOLLONESHOT));
5342 static int io_poll_update_prep(struct io_kiocb *req,
5343 const struct io_uring_sqe *sqe)
5345 struct io_poll_update *upd = &req->poll_update;
5348 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5350 if (sqe->ioprio || sqe->buf_index)
5352 flags = READ_ONCE(sqe->len);
5353 if (flags & ~(IORING_POLL_UPDATE_EVENTS | IORING_POLL_UPDATE_USER_DATA |
5354 IORING_POLL_ADD_MULTI))
5356 /* meaningless without update */
5357 if (flags == IORING_POLL_ADD_MULTI)
5360 upd->old_user_data = READ_ONCE(sqe->addr);
5361 upd->update_events = flags & IORING_POLL_UPDATE_EVENTS;
5362 upd->update_user_data = flags & IORING_POLL_UPDATE_USER_DATA;
5364 upd->new_user_data = READ_ONCE(sqe->off);
5365 if (!upd->update_user_data && upd->new_user_data)
5367 if (upd->update_events)
5368 upd->events = io_poll_parse_events(sqe, flags);
5369 else if (sqe->poll32_events)
5375 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5378 struct io_kiocb *req = wait->private;
5379 struct io_poll_iocb *poll = &req->poll;
5381 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5384 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5385 struct poll_table_struct *p)
5387 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5389 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5392 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5394 struct io_poll_iocb *poll = &req->poll;
5397 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5399 if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->addr)
5401 flags = READ_ONCE(sqe->len);
5402 if (flags & ~IORING_POLL_ADD_MULTI)
5405 poll->events = io_poll_parse_events(sqe, flags);
5409 static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5411 struct io_poll_iocb *poll = &req->poll;
5412 struct io_ring_ctx *ctx = req->ctx;
5413 struct io_poll_table ipt;
5416 ipt.pt._qproc = io_poll_queue_proc;
5418 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5421 if (mask) { /* no async, we'd stolen it */
5423 io_poll_complete(req, mask);
5425 spin_unlock_irq(&ctx->completion_lock);
5428 io_cqring_ev_posted(ctx);
5429 if (poll->events & EPOLLONESHOT)
5435 static int io_poll_update(struct io_kiocb *req, unsigned int issue_flags)
5437 struct io_ring_ctx *ctx = req->ctx;
5438 struct io_kiocb *preq;
5442 spin_lock_irq(&ctx->completion_lock);
5443 preq = io_poll_find(ctx, req->poll_update.old_user_data, true);
5449 if (!req->poll_update.update_events && !req->poll_update.update_user_data) {
5451 ret = io_poll_remove_one(preq) ? 0 : -EALREADY;
5456 * Don't allow racy completion with singleshot, as we cannot safely
5457 * update those. For multishot, if we're racing with completion, just
5458 * let completion re-add it.
5460 completing = !__io_poll_remove_one(preq, &preq->poll, false);
5461 if (completing && (preq->poll.events & EPOLLONESHOT)) {
5465 /* we now have a detached poll request. reissue. */
5469 spin_unlock_irq(&ctx->completion_lock);
5471 io_req_complete(req, ret);
5474 /* only mask one event flags, keep behavior flags */
5475 if (req->poll_update.update_events) {
5476 preq->poll.events &= ~0xffff;
5477 preq->poll.events |= req->poll_update.events & 0xffff;
5478 preq->poll.events |= IO_POLL_UNMASK;
5480 if (req->poll_update.update_user_data)
5481 preq->user_data = req->poll_update.new_user_data;
5482 spin_unlock_irq(&ctx->completion_lock);
5484 /* complete update request, we're done with it */
5485 io_req_complete(req, ret);
5488 ret = io_poll_add(preq, issue_flags);
5491 io_req_complete(preq, ret);
5497 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5499 struct io_timeout_data *data = container_of(timer,
5500 struct io_timeout_data, timer);
5501 struct io_kiocb *req = data->req;
5502 struct io_ring_ctx *ctx = req->ctx;
5503 unsigned long flags;
5505 spin_lock_irqsave(&ctx->completion_lock, flags);
5506 list_del_init(&req->timeout.list);
5507 atomic_set(&req->ctx->cq_timeouts,
5508 atomic_read(&req->ctx->cq_timeouts) + 1);
5510 io_cqring_fill_event(ctx, req->user_data, -ETIME, 0);
5511 io_commit_cqring(ctx);
5512 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5514 io_cqring_ev_posted(ctx);
5517 return HRTIMER_NORESTART;
5520 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5522 __must_hold(&ctx->completion_lock)
5524 struct io_timeout_data *io;
5525 struct io_kiocb *req;
5528 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5529 found = user_data == req->user_data;
5534 return ERR_PTR(-ENOENT);
5536 io = req->async_data;
5537 if (hrtimer_try_to_cancel(&io->timer) == -1)
5538 return ERR_PTR(-EALREADY);
5539 list_del_init(&req->timeout.list);
5543 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5544 __must_hold(&ctx->completion_lock)
5546 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5549 return PTR_ERR(req);
5552 io_cqring_fill_event(ctx, req->user_data, -ECANCELED, 0);
5553 io_put_req_deferred(req, 1);
5557 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5558 struct timespec64 *ts, enum hrtimer_mode mode)
5559 __must_hold(&ctx->completion_lock)
5561 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5562 struct io_timeout_data *data;
5565 return PTR_ERR(req);
5567 req->timeout.off = 0; /* noseq */
5568 data = req->async_data;
5569 list_add_tail(&req->timeout.list, &ctx->timeout_list);
5570 hrtimer_init(&data->timer, CLOCK_MONOTONIC, mode);
5571 data->timer.function = io_timeout_fn;
5572 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
5576 static int io_timeout_remove_prep(struct io_kiocb *req,
5577 const struct io_uring_sqe *sqe)
5579 struct io_timeout_rem *tr = &req->timeout_rem;
5581 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5583 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5585 if (sqe->ioprio || sqe->buf_index || sqe->len)
5588 tr->addr = READ_ONCE(sqe->addr);
5589 tr->flags = READ_ONCE(sqe->timeout_flags);
5590 if (tr->flags & IORING_TIMEOUT_UPDATE) {
5591 if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS))
5593 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
5595 } else if (tr->flags) {
5596 /* timeout removal doesn't support flags */
5603 static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
5605 return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
5610 * Remove or update an existing timeout command
5612 static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
5614 struct io_timeout_rem *tr = &req->timeout_rem;
5615 struct io_ring_ctx *ctx = req->ctx;
5618 spin_lock_irq(&ctx->completion_lock);
5619 if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE))
5620 ret = io_timeout_cancel(ctx, tr->addr);
5622 ret = io_timeout_update(ctx, tr->addr, &tr->ts,
5623 io_translate_timeout_mode(tr->flags));
5625 io_cqring_fill_event(ctx, req->user_data, ret, 0);
5626 io_commit_cqring(ctx);
5627 spin_unlock_irq(&ctx->completion_lock);
5628 io_cqring_ev_posted(ctx);
5635 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5636 bool is_timeout_link)
5638 struct io_timeout_data *data;
5640 u32 off = READ_ONCE(sqe->off);
5642 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5644 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5646 if (off && is_timeout_link)
5648 flags = READ_ONCE(sqe->timeout_flags);
5649 if (flags & ~IORING_TIMEOUT_ABS)
5652 req->timeout.off = off;
5653 if (unlikely(off && !req->ctx->off_timeout_used))
5654 req->ctx->off_timeout_used = true;
5656 if (!req->async_data && io_alloc_async_data(req))
5659 data = req->async_data;
5662 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5665 data->mode = io_translate_timeout_mode(flags);
5666 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5667 if (is_timeout_link)
5668 io_req_track_inflight(req);
5672 static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
5674 struct io_ring_ctx *ctx = req->ctx;
5675 struct io_timeout_data *data = req->async_data;
5676 struct list_head *entry;
5677 u32 tail, off = req->timeout.off;
5679 spin_lock_irq(&ctx->completion_lock);
5682 * sqe->off holds how many events that need to occur for this
5683 * timeout event to be satisfied. If it isn't set, then this is
5684 * a pure timeout request, sequence isn't used.
5686 if (io_is_timeout_noseq(req)) {
5687 entry = ctx->timeout_list.prev;
5691 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5692 req->timeout.target_seq = tail + off;
5694 /* Update the last seq here in case io_flush_timeouts() hasn't.
5695 * This is safe because ->completion_lock is held, and submissions
5696 * and completions are never mixed in the same ->completion_lock section.
5698 ctx->cq_last_tm_flush = tail;
5701 * Insertion sort, ensuring the first entry in the list is always
5702 * the one we need first.
5704 list_for_each_prev(entry, &ctx->timeout_list) {
5705 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5708 if (io_is_timeout_noseq(nxt))
5710 /* nxt.seq is behind @tail, otherwise would've been completed */
5711 if (off >= nxt->timeout.target_seq - tail)
5715 list_add(&req->timeout.list, entry);
5716 data->timer.function = io_timeout_fn;
5717 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5718 spin_unlock_irq(&ctx->completion_lock);
5722 struct io_cancel_data {
5723 struct io_ring_ctx *ctx;
5727 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5729 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5730 struct io_cancel_data *cd = data;
5732 return req->ctx == cd->ctx && req->user_data == cd->user_data;
5735 static int io_async_cancel_one(struct io_uring_task *tctx, u64 user_data,
5736 struct io_ring_ctx *ctx)
5738 struct io_cancel_data data = { .ctx = ctx, .user_data = user_data, };
5739 enum io_wq_cancel cancel_ret;
5742 if (!tctx || !tctx->io_wq)
5745 cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, &data, false);
5746 switch (cancel_ret) {
5747 case IO_WQ_CANCEL_OK:
5750 case IO_WQ_CANCEL_RUNNING:
5753 case IO_WQ_CANCEL_NOTFOUND:
5761 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5762 struct io_kiocb *req, __u64 sqe_addr,
5765 unsigned long flags;
5768 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5769 spin_lock_irqsave(&ctx->completion_lock, flags);
5772 ret = io_timeout_cancel(ctx, sqe_addr);
5775 ret = io_poll_cancel(ctx, sqe_addr, false);
5779 io_cqring_fill_event(ctx, req->user_data, ret, 0);
5780 io_commit_cqring(ctx);
5781 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5782 io_cqring_ev_posted(ctx);
5788 static int io_async_cancel_prep(struct io_kiocb *req,
5789 const struct io_uring_sqe *sqe)
5791 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5793 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5795 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5798 req->cancel.addr = READ_ONCE(sqe->addr);
5802 static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
5804 struct io_ring_ctx *ctx = req->ctx;
5805 u64 sqe_addr = req->cancel.addr;
5806 struct io_tctx_node *node;
5809 /* tasks should wait for their io-wq threads, so safe w/o sync */
5810 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5811 spin_lock_irq(&ctx->completion_lock);
5814 ret = io_timeout_cancel(ctx, sqe_addr);
5817 ret = io_poll_cancel(ctx, sqe_addr, false);
5820 spin_unlock_irq(&ctx->completion_lock);
5822 /* slow path, try all io-wq's */
5823 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5825 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
5826 struct io_uring_task *tctx = node->task->io_uring;
5828 ret = io_async_cancel_one(tctx, req->cancel.addr, ctx);
5832 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5834 spin_lock_irq(&ctx->completion_lock);
5836 io_cqring_fill_event(ctx, req->user_data, ret, 0);
5837 io_commit_cqring(ctx);
5838 spin_unlock_irq(&ctx->completion_lock);
5839 io_cqring_ev_posted(ctx);
5847 static int io_rsrc_update_prep(struct io_kiocb *req,
5848 const struct io_uring_sqe *sqe)
5850 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5852 if (sqe->ioprio || sqe->rw_flags)
5855 req->rsrc_update.offset = READ_ONCE(sqe->off);
5856 req->rsrc_update.nr_args = READ_ONCE(sqe->len);
5857 if (!req->rsrc_update.nr_args)
5859 req->rsrc_update.arg = READ_ONCE(sqe->addr);
5863 static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
5865 struct io_ring_ctx *ctx = req->ctx;
5866 struct io_uring_rsrc_update2 up;
5869 if (issue_flags & IO_URING_F_NONBLOCK)
5872 up.offset = req->rsrc_update.offset;
5873 up.data = req->rsrc_update.arg;
5878 mutex_lock(&ctx->uring_lock);
5879 ret = __io_register_rsrc_update(ctx, IORING_RSRC_FILE,
5880 &up, req->rsrc_update.nr_args);
5881 mutex_unlock(&ctx->uring_lock);
5885 __io_req_complete(req, issue_flags, ret, 0);
5889 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5891 switch (req->opcode) {
5894 case IORING_OP_READV:
5895 case IORING_OP_READ_FIXED:
5896 case IORING_OP_READ:
5897 return io_read_prep(req, sqe);
5898 case IORING_OP_WRITEV:
5899 case IORING_OP_WRITE_FIXED:
5900 case IORING_OP_WRITE:
5901 return io_write_prep(req, sqe);
5902 case IORING_OP_POLL_ADD:
5903 return io_poll_add_prep(req, sqe);
5904 case IORING_OP_POLL_REMOVE:
5905 return io_poll_update_prep(req, sqe);
5906 case IORING_OP_FSYNC:
5907 return io_fsync_prep(req, sqe);
5908 case IORING_OP_SYNC_FILE_RANGE:
5909 return io_sfr_prep(req, sqe);
5910 case IORING_OP_SENDMSG:
5911 case IORING_OP_SEND:
5912 return io_sendmsg_prep(req, sqe);
5913 case IORING_OP_RECVMSG:
5914 case IORING_OP_RECV:
5915 return io_recvmsg_prep(req, sqe);
5916 case IORING_OP_CONNECT:
5917 return io_connect_prep(req, sqe);
5918 case IORING_OP_TIMEOUT:
5919 return io_timeout_prep(req, sqe, false);
5920 case IORING_OP_TIMEOUT_REMOVE:
5921 return io_timeout_remove_prep(req, sqe);
5922 case IORING_OP_ASYNC_CANCEL:
5923 return io_async_cancel_prep(req, sqe);
5924 case IORING_OP_LINK_TIMEOUT:
5925 return io_timeout_prep(req, sqe, true);
5926 case IORING_OP_ACCEPT:
5927 return io_accept_prep(req, sqe);
5928 case IORING_OP_FALLOCATE:
5929 return io_fallocate_prep(req, sqe);
5930 case IORING_OP_OPENAT:
5931 return io_openat_prep(req, sqe);
5932 case IORING_OP_CLOSE:
5933 return io_close_prep(req, sqe);
5934 case IORING_OP_FILES_UPDATE:
5935 return io_rsrc_update_prep(req, sqe);
5936 case IORING_OP_STATX:
5937 return io_statx_prep(req, sqe);
5938 case IORING_OP_FADVISE:
5939 return io_fadvise_prep(req, sqe);
5940 case IORING_OP_MADVISE:
5941 return io_madvise_prep(req, sqe);
5942 case IORING_OP_OPENAT2:
5943 return io_openat2_prep(req, sqe);
5944 case IORING_OP_EPOLL_CTL:
5945 return io_epoll_ctl_prep(req, sqe);
5946 case IORING_OP_SPLICE:
5947 return io_splice_prep(req, sqe);
5948 case IORING_OP_PROVIDE_BUFFERS:
5949 return io_provide_buffers_prep(req, sqe);
5950 case IORING_OP_REMOVE_BUFFERS:
5951 return io_remove_buffers_prep(req, sqe);
5953 return io_tee_prep(req, sqe);
5954 case IORING_OP_SHUTDOWN:
5955 return io_shutdown_prep(req, sqe);
5956 case IORING_OP_RENAMEAT:
5957 return io_renameat_prep(req, sqe);
5958 case IORING_OP_UNLINKAT:
5959 return io_unlinkat_prep(req, sqe);
5962 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5967 static int io_req_prep_async(struct io_kiocb *req)
5969 if (!io_op_defs[req->opcode].needs_async_setup)
5971 if (WARN_ON_ONCE(req->async_data))
5973 if (io_alloc_async_data(req))
5976 switch (req->opcode) {
5977 case IORING_OP_READV:
5978 return io_rw_prep_async(req, READ);
5979 case IORING_OP_WRITEV:
5980 return io_rw_prep_async(req, WRITE);
5981 case IORING_OP_SENDMSG:
5982 return io_sendmsg_prep_async(req);
5983 case IORING_OP_RECVMSG:
5984 return io_recvmsg_prep_async(req);
5985 case IORING_OP_CONNECT:
5986 return io_connect_prep_async(req);
5988 printk_once(KERN_WARNING "io_uring: prep_async() bad opcode %d\n",
5993 static u32 io_get_sequence(struct io_kiocb *req)
5995 struct io_kiocb *pos;
5996 struct io_ring_ctx *ctx = req->ctx;
5999 io_for_each_link(pos, req)
6001 return ctx->cached_sq_head - nr_reqs;
6004 static bool io_drain_req(struct io_kiocb *req)
6006 struct io_kiocb *pos;
6007 struct io_ring_ctx *ctx = req->ctx;
6008 struct io_defer_entry *de;
6013 * If we need to drain a request in the middle of a link, drain the
6014 * head request and the next request/link after the current link.
6015 * Considering sequential execution of links, IOSQE_IO_DRAIN will be
6016 * maintained for every request of our link.
6018 if (ctx->drain_next) {
6019 req->flags |= REQ_F_IO_DRAIN;
6020 ctx->drain_next = false;
6022 /* not interested in head, start from the first linked */
6023 io_for_each_link(pos, req->link) {
6024 if (pos->flags & REQ_F_IO_DRAIN) {
6025 ctx->drain_next = true;
6026 req->flags |= REQ_F_IO_DRAIN;
6031 /* Still need defer if there is pending req in defer list. */
6032 if (likely(list_empty_careful(&ctx->defer_list) &&
6033 !(req->flags & REQ_F_IO_DRAIN))) {
6034 ctx->drain_active = false;
6038 seq = io_get_sequence(req);
6039 /* Still a chance to pass the sequence check */
6040 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
6043 ret = io_req_prep_async(req);
6046 io_prep_async_link(req);
6047 de = kmalloc(sizeof(*de), GFP_KERNEL);
6049 io_req_complete_failed(req, ret);
6053 spin_lock_irq(&ctx->completion_lock);
6054 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
6055 spin_unlock_irq(&ctx->completion_lock);
6057 io_queue_async_work(req);
6061 trace_io_uring_defer(ctx, req, req->user_data);
6064 list_add_tail(&de->list, &ctx->defer_list);
6065 spin_unlock_irq(&ctx->completion_lock);
6069 static void io_clean_op(struct io_kiocb *req)
6071 if (req->flags & REQ_F_BUFFER_SELECTED) {
6072 switch (req->opcode) {
6073 case IORING_OP_READV:
6074 case IORING_OP_READ_FIXED:
6075 case IORING_OP_READ:
6076 kfree((void *)(unsigned long)req->rw.addr);
6078 case IORING_OP_RECVMSG:
6079 case IORING_OP_RECV:
6080 kfree(req->sr_msg.kbuf);
6083 req->flags &= ~REQ_F_BUFFER_SELECTED;
6086 if (req->flags & REQ_F_NEED_CLEANUP) {
6087 switch (req->opcode) {
6088 case IORING_OP_READV:
6089 case IORING_OP_READ_FIXED:
6090 case IORING_OP_READ:
6091 case IORING_OP_WRITEV:
6092 case IORING_OP_WRITE_FIXED:
6093 case IORING_OP_WRITE: {
6094 struct io_async_rw *io = req->async_data;
6096 kfree(io->free_iovec);
6099 case IORING_OP_RECVMSG:
6100 case IORING_OP_SENDMSG: {
6101 struct io_async_msghdr *io = req->async_data;
6103 kfree(io->free_iov);
6106 case IORING_OP_SPLICE:
6108 if (!(req->splice.flags & SPLICE_F_FD_IN_FIXED))
6109 io_put_file(req->splice.file_in);
6111 case IORING_OP_OPENAT:
6112 case IORING_OP_OPENAT2:
6113 if (req->open.filename)
6114 putname(req->open.filename);
6116 case IORING_OP_RENAMEAT:
6117 putname(req->rename.oldpath);
6118 putname(req->rename.newpath);
6120 case IORING_OP_UNLINKAT:
6121 putname(req->unlink.filename);
6124 req->flags &= ~REQ_F_NEED_CLEANUP;
6126 if ((req->flags & REQ_F_POLLED) && req->apoll) {
6127 kfree(req->apoll->double_poll);
6131 if (req->flags & REQ_F_INFLIGHT) {
6132 struct io_uring_task *tctx = req->task->io_uring;
6134 atomic_dec(&tctx->inflight_tracked);
6135 req->flags &= ~REQ_F_INFLIGHT;
6137 if (req->flags & REQ_F_CREDS) {
6138 put_cred(req->creds);
6139 req->flags &= ~REQ_F_CREDS;
6143 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
6145 struct io_ring_ctx *ctx = req->ctx;
6146 const struct cred *creds = NULL;
6149 if ((req->flags & REQ_F_CREDS) && req->creds != current_cred())
6150 creds = override_creds(req->creds);
6152 switch (req->opcode) {
6154 ret = io_nop(req, issue_flags);
6156 case IORING_OP_READV:
6157 case IORING_OP_READ_FIXED:
6158 case IORING_OP_READ:
6159 ret = io_read(req, issue_flags);
6161 case IORING_OP_WRITEV:
6162 case IORING_OP_WRITE_FIXED:
6163 case IORING_OP_WRITE:
6164 ret = io_write(req, issue_flags);
6166 case IORING_OP_FSYNC:
6167 ret = io_fsync(req, issue_flags);
6169 case IORING_OP_POLL_ADD:
6170 ret = io_poll_add(req, issue_flags);
6172 case IORING_OP_POLL_REMOVE:
6173 ret = io_poll_update(req, issue_flags);
6175 case IORING_OP_SYNC_FILE_RANGE:
6176 ret = io_sync_file_range(req, issue_flags);
6178 case IORING_OP_SENDMSG:
6179 ret = io_sendmsg(req, issue_flags);
6181 case IORING_OP_SEND:
6182 ret = io_send(req, issue_flags);
6184 case IORING_OP_RECVMSG:
6185 ret = io_recvmsg(req, issue_flags);
6187 case IORING_OP_RECV:
6188 ret = io_recv(req, issue_flags);
6190 case IORING_OP_TIMEOUT:
6191 ret = io_timeout(req, issue_flags);
6193 case IORING_OP_TIMEOUT_REMOVE:
6194 ret = io_timeout_remove(req, issue_flags);
6196 case IORING_OP_ACCEPT:
6197 ret = io_accept(req, issue_flags);
6199 case IORING_OP_CONNECT:
6200 ret = io_connect(req, issue_flags);
6202 case IORING_OP_ASYNC_CANCEL:
6203 ret = io_async_cancel(req, issue_flags);
6205 case IORING_OP_FALLOCATE:
6206 ret = io_fallocate(req, issue_flags);
6208 case IORING_OP_OPENAT:
6209 ret = io_openat(req, issue_flags);
6211 case IORING_OP_CLOSE:
6212 ret = io_close(req, issue_flags);
6214 case IORING_OP_FILES_UPDATE:
6215 ret = io_files_update(req, issue_flags);
6217 case IORING_OP_STATX:
6218 ret = io_statx(req, issue_flags);
6220 case IORING_OP_FADVISE:
6221 ret = io_fadvise(req, issue_flags);
6223 case IORING_OP_MADVISE:
6224 ret = io_madvise(req, issue_flags);
6226 case IORING_OP_OPENAT2:
6227 ret = io_openat2(req, issue_flags);
6229 case IORING_OP_EPOLL_CTL:
6230 ret = io_epoll_ctl(req, issue_flags);
6232 case IORING_OP_SPLICE:
6233 ret = io_splice(req, issue_flags);
6235 case IORING_OP_PROVIDE_BUFFERS:
6236 ret = io_provide_buffers(req, issue_flags);
6238 case IORING_OP_REMOVE_BUFFERS:
6239 ret = io_remove_buffers(req, issue_flags);
6242 ret = io_tee(req, issue_flags);
6244 case IORING_OP_SHUTDOWN:
6245 ret = io_shutdown(req, issue_flags);
6247 case IORING_OP_RENAMEAT:
6248 ret = io_renameat(req, issue_flags);
6250 case IORING_OP_UNLINKAT:
6251 ret = io_unlinkat(req, issue_flags);
6259 revert_creds(creds);
6262 /* If the op doesn't have a file, we're not polling for it */
6263 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file)
6264 io_iopoll_req_issued(req);
6269 static void io_wq_submit_work(struct io_wq_work *work)
6271 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6272 struct io_kiocb *timeout;
6275 timeout = io_prep_linked_timeout(req);
6277 io_queue_linked_timeout(timeout);
6279 if (work->flags & IO_WQ_WORK_CANCEL)
6284 ret = io_issue_sqe(req, 0);
6286 * We can get EAGAIN for polled IO even though we're
6287 * forcing a sync submission from here, since we can't
6288 * wait for request slots on the block side.
6296 /* avoid locking problems by failing it from a clean context */
6298 /* io-wq is going to take one down */
6300 io_req_task_queue_fail(req, ret);
6304 #define FFS_ASYNC_READ 0x1UL
6305 #define FFS_ASYNC_WRITE 0x2UL
6307 #define FFS_ISREG 0x4UL
6309 #define FFS_ISREG 0x0UL
6311 #define FFS_MASK ~(FFS_ASYNC_READ|FFS_ASYNC_WRITE|FFS_ISREG)
6313 static inline struct io_fixed_file *io_fixed_file_slot(struct io_file_table *table,
6316 struct io_fixed_file *table_l2;
6318 table_l2 = table->files[i >> IORING_FILE_TABLE_SHIFT];
6319 return &table_l2[i & IORING_FILE_TABLE_MASK];
6322 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6325 struct io_fixed_file *slot = io_fixed_file_slot(&ctx->file_table, index);
6327 return (struct file *) (slot->file_ptr & FFS_MASK);
6330 static void io_fixed_file_set(struct io_fixed_file *file_slot, struct file *file)
6332 unsigned long file_ptr = (unsigned long) file;
6334 if (__io_file_supports_async(file, READ))
6335 file_ptr |= FFS_ASYNC_READ;
6336 if (__io_file_supports_async(file, WRITE))
6337 file_ptr |= FFS_ASYNC_WRITE;
6338 if (S_ISREG(file_inode(file)->i_mode))
6339 file_ptr |= FFS_ISREG;
6340 file_slot->file_ptr = file_ptr;
6343 static struct file *io_file_get(struct io_submit_state *state,
6344 struct io_kiocb *req, int fd, bool fixed)
6346 struct io_ring_ctx *ctx = req->ctx;
6350 unsigned long file_ptr;
6352 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6354 fd = array_index_nospec(fd, ctx->nr_user_files);
6355 file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr;
6356 file = (struct file *) (file_ptr & FFS_MASK);
6357 file_ptr &= ~FFS_MASK;
6358 /* mask in overlapping REQ_F and FFS bits */
6359 req->flags |= (file_ptr << REQ_F_ASYNC_READ_BIT);
6360 io_req_set_rsrc_node(req);
6362 trace_io_uring_file_get(ctx, fd);
6363 file = __io_file_get(state, fd);
6365 /* we don't allow fixed io_uring files */
6366 if (file && unlikely(file->f_op == &io_uring_fops))
6367 io_req_track_inflight(req);
6373 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6375 struct io_timeout_data *data = container_of(timer,
6376 struct io_timeout_data, timer);
6377 struct io_kiocb *prev, *req = data->req;
6378 struct io_ring_ctx *ctx = req->ctx;
6379 unsigned long flags;
6381 spin_lock_irqsave(&ctx->completion_lock, flags);
6382 prev = req->timeout.head;
6383 req->timeout.head = NULL;
6386 * We don't expect the list to be empty, that will only happen if we
6387 * race with the completion of the linked work.
6390 io_remove_next_linked(prev);
6391 if (!req_ref_inc_not_zero(prev))
6394 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6397 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6398 io_put_req_deferred(prev, 1);
6399 io_put_req_deferred(req, 1);
6401 io_req_complete_post(req, -ETIME, 0);
6403 return HRTIMER_NORESTART;
6406 static void io_queue_linked_timeout(struct io_kiocb *req)
6408 struct io_ring_ctx *ctx = req->ctx;
6410 spin_lock_irq(&ctx->completion_lock);
6412 * If the back reference is NULL, then our linked request finished
6413 * before we got a chance to setup the timer
6415 if (req->timeout.head) {
6416 struct io_timeout_data *data = req->async_data;
6418 data->timer.function = io_link_timeout_fn;
6419 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6422 spin_unlock_irq(&ctx->completion_lock);
6423 /* drop submission reference */
6427 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6429 struct io_kiocb *nxt = req->link;
6431 if (!nxt || (req->flags & REQ_F_LINK_TIMEOUT) ||
6432 nxt->opcode != IORING_OP_LINK_TIMEOUT)
6435 nxt->timeout.head = req;
6436 nxt->flags |= REQ_F_LTIMEOUT_ACTIVE;
6437 req->flags |= REQ_F_LINK_TIMEOUT;
6441 static void __io_queue_sqe(struct io_kiocb *req)
6443 struct io_kiocb *linked_timeout = io_prep_linked_timeout(req);
6446 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
6449 * We async punt it if the file wasn't marked NOWAIT, or if the file
6450 * doesn't support non-blocking read/write attempts
6453 /* drop submission reference */
6454 if (req->flags & REQ_F_COMPLETE_INLINE) {
6455 struct io_ring_ctx *ctx = req->ctx;
6456 struct io_comp_state *cs = &ctx->submit_state.comp;
6458 cs->reqs[cs->nr++] = req;
6459 if (cs->nr == ARRAY_SIZE(cs->reqs))
6460 io_submit_flush_completions(ctx);
6464 } else if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6465 if (!io_arm_poll_handler(req)) {
6467 * Queued up for async execution, worker will release
6468 * submit reference when the iocb is actually submitted.
6470 io_queue_async_work(req);
6473 io_req_complete_failed(req, ret);
6476 io_queue_linked_timeout(linked_timeout);
6479 static inline void io_queue_sqe(struct io_kiocb *req)
6481 if (unlikely(req->ctx->drain_active) && io_drain_req(req))
6484 if (likely(!(req->flags & REQ_F_FORCE_ASYNC))) {
6485 __io_queue_sqe(req);
6487 int ret = io_req_prep_async(req);
6490 io_req_complete_failed(req, ret);
6492 io_queue_async_work(req);
6497 * Check SQE restrictions (opcode and flags).
6499 * Returns 'true' if SQE is allowed, 'false' otherwise.
6501 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6502 struct io_kiocb *req,
6503 unsigned int sqe_flags)
6505 if (!ctx->restricted)
6508 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6511 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6512 ctx->restrictions.sqe_flags_required)
6515 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6516 ctx->restrictions.sqe_flags_required))
6522 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6523 const struct io_uring_sqe *sqe)
6525 struct io_submit_state *state;
6526 unsigned int sqe_flags;
6527 int personality, ret = 0;
6529 req->opcode = READ_ONCE(sqe->opcode);
6530 /* same numerical values with corresponding REQ_F_*, safe to copy */
6531 req->flags = sqe_flags = READ_ONCE(sqe->flags);
6532 req->user_data = READ_ONCE(sqe->user_data);
6533 req->async_data = NULL;
6537 req->fixed_rsrc_refs = NULL;
6538 /* one is dropped after submission, the other at completion */
6539 atomic_set(&req->refs, 2);
6540 req->task = current;
6543 /* enforce forwards compatibility on users */
6544 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6546 if (unlikely(req->opcode >= IORING_OP_LAST))
6548 if (unlikely(!io_check_restriction(ctx, req, sqe_flags)))
6551 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6552 !io_op_defs[req->opcode].buffer_select)
6554 if (unlikely(sqe_flags & IOSQE_IO_DRAIN))
6555 ctx->drain_active = true;
6557 personality = READ_ONCE(sqe->personality);
6559 req->creds = xa_load(&ctx->personalities, personality);
6562 get_cred(req->creds);
6563 req->flags |= REQ_F_CREDS;
6565 state = &ctx->submit_state;
6568 * Plug now if we have more than 1 IO left after this, and the target
6569 * is potentially a read/write to block based storage.
6571 if (!state->plug_started && state->ios_left > 1 &&
6572 io_op_defs[req->opcode].plug) {
6573 blk_start_plug(&state->plug);
6574 state->plug_started = true;
6577 if (io_op_defs[req->opcode].needs_file) {
6578 bool fixed = req->flags & REQ_F_FIXED_FILE;
6580 req->file = io_file_get(state, req, READ_ONCE(sqe->fd), fixed);
6581 if (unlikely(!req->file))
6589 static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
6590 const struct io_uring_sqe *sqe)
6592 struct io_submit_link *link = &ctx->submit_state.link;
6595 ret = io_init_req(ctx, req, sqe);
6596 if (unlikely(ret)) {
6599 /* fail even hard links since we don't submit */
6600 req_set_fail(link->head);
6601 io_req_complete_failed(link->head, -ECANCELED);
6604 io_req_complete_failed(req, ret);
6608 ret = io_req_prep(req, sqe);
6612 /* don't need @sqe from now on */
6613 trace_io_uring_submit_sqe(ctx, req, req->opcode, req->user_data,
6615 ctx->flags & IORING_SETUP_SQPOLL);
6618 * If we already have a head request, queue this one for async
6619 * submittal once the head completes. If we don't have a head but
6620 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6621 * submitted sync once the chain is complete. If none of those
6622 * conditions are true (normal request), then just queue it.
6625 struct io_kiocb *head = link->head;
6627 ret = io_req_prep_async(req);
6630 trace_io_uring_link(ctx, req, head);
6631 link->last->link = req;
6634 /* last request of a link, enqueue the link */
6635 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6640 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6652 * Batched submission is done, ensure local IO is flushed out.
6654 static void io_submit_state_end(struct io_submit_state *state,
6655 struct io_ring_ctx *ctx)
6657 if (state->link.head)
6658 io_queue_sqe(state->link.head);
6660 io_submit_flush_completions(ctx);
6661 if (state->plug_started)
6662 blk_finish_plug(&state->plug);
6663 io_state_file_put(state);
6667 * Start submission side cache.
6669 static void io_submit_state_start(struct io_submit_state *state,
6670 unsigned int max_ios)
6672 state->plug_started = false;
6673 state->ios_left = max_ios;
6674 /* set only head, no need to init link_last in advance */
6675 state->link.head = NULL;
6678 static void io_commit_sqring(struct io_ring_ctx *ctx)
6680 struct io_rings *rings = ctx->rings;
6683 * Ensure any loads from the SQEs are done at this point,
6684 * since once we write the new head, the application could
6685 * write new data to them.
6687 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6691 * Fetch an sqe, if one is available. Note this returns a pointer to memory
6692 * that is mapped by userspace. This means that care needs to be taken to
6693 * ensure that reads are stable, as we cannot rely on userspace always
6694 * being a good citizen. If members of the sqe are validated and then later
6695 * used, it's important that those reads are done through READ_ONCE() to
6696 * prevent a re-load down the line.
6698 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6700 unsigned head, mask = ctx->sq_entries - 1;
6701 unsigned sq_idx = ctx->cached_sq_head++ & mask;
6704 * The cached sq head (or cq tail) serves two purposes:
6706 * 1) allows us to batch the cost of updating the user visible
6708 * 2) allows the kernel side to track the head on its own, even
6709 * though the application is the one updating it.
6711 head = READ_ONCE(ctx->sq_array[sq_idx]);
6712 if (likely(head < ctx->sq_entries))
6713 return &ctx->sq_sqes[head];
6715 /* drop invalid entries */
6717 WRITE_ONCE(ctx->rings->sq_dropped,
6718 READ_ONCE(ctx->rings->sq_dropped) + 1);
6722 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6724 struct io_uring_task *tctx;
6727 /* make sure SQ entry isn't read before tail */
6728 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6729 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6732 tctx = current->io_uring;
6733 tctx->cached_refs -= nr;
6734 if (unlikely(tctx->cached_refs < 0)) {
6735 unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR;
6737 percpu_counter_add(&tctx->inflight, refill);
6738 refcount_add(refill, ¤t->usage);
6739 tctx->cached_refs += refill;
6741 io_submit_state_start(&ctx->submit_state, nr);
6743 while (submitted < nr) {
6744 const struct io_uring_sqe *sqe;
6745 struct io_kiocb *req;
6747 req = io_alloc_req(ctx);
6748 if (unlikely(!req)) {
6750 submitted = -EAGAIN;
6753 sqe = io_get_sqe(ctx);
6754 if (unlikely(!sqe)) {
6755 kmem_cache_free(req_cachep, req);
6758 /* will complete beyond this point, count as submitted */
6760 if (io_submit_sqe(ctx, req, sqe))
6764 if (unlikely(submitted != nr)) {
6765 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6766 int unused = nr - ref_used;
6768 current->io_uring->cached_refs += unused;
6769 percpu_ref_put_many(&ctx->refs, unused);
6772 io_submit_state_end(&ctx->submit_state, ctx);
6773 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6774 io_commit_sqring(ctx);
6779 static inline bool io_sqd_events_pending(struct io_sq_data *sqd)
6781 return READ_ONCE(sqd->state);
6784 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6786 /* Tell userspace we may need a wakeup call */
6787 spin_lock_irq(&ctx->completion_lock);
6788 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6789 spin_unlock_irq(&ctx->completion_lock);
6792 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6794 spin_lock_irq(&ctx->completion_lock);
6795 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6796 spin_unlock_irq(&ctx->completion_lock);
6799 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
6801 unsigned int to_submit;
6804 to_submit = io_sqring_entries(ctx);
6805 /* if we're handling multiple rings, cap submit size for fairness */
6806 if (cap_entries && to_submit > 8)
6809 if (!list_empty(&ctx->iopoll_list) || to_submit) {
6810 unsigned nr_events = 0;
6812 mutex_lock(&ctx->uring_lock);
6813 if (!list_empty(&ctx->iopoll_list))
6814 io_do_iopoll(ctx, &nr_events, 0);
6817 * Don't submit if refs are dying, good for io_uring_register(),
6818 * but also it is relied upon by io_ring_exit_work()
6820 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)) &&
6821 !(ctx->flags & IORING_SETUP_R_DISABLED))
6822 ret = io_submit_sqes(ctx, to_submit);
6823 mutex_unlock(&ctx->uring_lock);
6825 if (to_submit && wq_has_sleeper(&ctx->sqo_sq_wait))
6826 wake_up(&ctx->sqo_sq_wait);
6832 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
6834 struct io_ring_ctx *ctx;
6835 unsigned sq_thread_idle = 0;
6837 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6838 sq_thread_idle = max(sq_thread_idle, ctx->sq_thread_idle);
6839 sqd->sq_thread_idle = sq_thread_idle;
6842 static bool io_sqd_handle_event(struct io_sq_data *sqd)
6844 bool did_sig = false;
6845 struct ksignal ksig;
6847 if (test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state) ||
6848 signal_pending(current)) {
6849 mutex_unlock(&sqd->lock);
6850 if (signal_pending(current))
6851 did_sig = get_signal(&ksig);
6853 mutex_lock(&sqd->lock);
6856 return did_sig || test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
6859 static int io_sq_thread(void *data)
6861 struct io_sq_data *sqd = data;
6862 struct io_ring_ctx *ctx;
6863 unsigned long timeout = 0;
6864 char buf[TASK_COMM_LEN];
6867 snprintf(buf, sizeof(buf), "iou-sqp-%d", sqd->task_pid);
6868 set_task_comm(current, buf);
6870 if (sqd->sq_cpu != -1)
6871 set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
6873 set_cpus_allowed_ptr(current, cpu_online_mask);
6874 current->flags |= PF_NO_SETAFFINITY;
6876 mutex_lock(&sqd->lock);
6879 bool cap_entries, sqt_spin, needs_sched;
6881 if (io_sqd_events_pending(sqd) || signal_pending(current)) {
6882 if (io_sqd_handle_event(sqd))
6884 timeout = jiffies + sqd->sq_thread_idle;
6889 cap_entries = !list_is_singular(&sqd->ctx_list);
6890 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6891 const struct cred *creds = NULL;
6893 if (ctx->sq_creds != current_cred())
6894 creds = override_creds(ctx->sq_creds);
6895 ret = __io_sq_thread(ctx, cap_entries);
6897 revert_creds(creds);
6898 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
6902 if (sqt_spin || !time_after(jiffies, timeout)) {
6906 timeout = jiffies + sqd->sq_thread_idle;
6910 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
6911 if (!io_sqd_events_pending(sqd)) {
6913 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6914 io_ring_set_wakeup_flag(ctx);
6916 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6917 !list_empty_careful(&ctx->iopoll_list)) {
6918 needs_sched = false;
6921 if (io_sqring_entries(ctx)) {
6922 needs_sched = false;
6928 mutex_unlock(&sqd->lock);
6930 mutex_lock(&sqd->lock);
6932 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6933 io_ring_clear_wakeup_flag(ctx);
6936 finish_wait(&sqd->wait, &wait);
6937 timeout = jiffies + sqd->sq_thread_idle;
6940 io_uring_cancel_generic(true, sqd);
6942 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6943 io_ring_set_wakeup_flag(ctx);
6945 mutex_unlock(&sqd->lock);
6947 complete(&sqd->exited);
6951 struct io_wait_queue {
6952 struct wait_queue_entry wq;
6953 struct io_ring_ctx *ctx;
6955 unsigned nr_timeouts;
6958 static inline bool io_should_wake(struct io_wait_queue *iowq)
6960 struct io_ring_ctx *ctx = iowq->ctx;
6963 * Wake up if we have enough events, or if a timeout occurred since we
6964 * started waiting. For timeouts, we always want to return to userspace,
6965 * regardless of event count.
6967 return io_cqring_events(ctx) >= iowq->to_wait ||
6968 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6971 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6972 int wake_flags, void *key)
6974 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6978 * Cannot safely flush overflowed CQEs from here, ensure we wake up
6979 * the task, and the next invocation will do it.
6981 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->check_cq_overflow))
6982 return autoremove_wake_function(curr, mode, wake_flags, key);
6986 static int io_run_task_work_sig(void)
6988 if (io_run_task_work())
6990 if (!signal_pending(current))
6992 if (test_thread_flag(TIF_NOTIFY_SIGNAL))
6993 return -ERESTARTSYS;
6997 /* when returns >0, the caller should retry */
6998 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
6999 struct io_wait_queue *iowq,
7000 signed long *timeout)
7004 /* make sure we run task_work before checking for signals */
7005 ret = io_run_task_work_sig();
7006 if (ret || io_should_wake(iowq))
7008 /* let the caller flush overflows, retry */
7009 if (test_bit(0, &ctx->check_cq_overflow))
7012 *timeout = schedule_timeout(*timeout);
7013 return !*timeout ? -ETIME : 1;
7017 * Wait until events become available, if we don't already have some. The
7018 * application must reap them itself, as they reside on the shared cq ring.
7020 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
7021 const sigset_t __user *sig, size_t sigsz,
7022 struct __kernel_timespec __user *uts)
7024 struct io_wait_queue iowq = {
7027 .func = io_wake_function,
7028 .entry = LIST_HEAD_INIT(iowq.wq.entry),
7031 .to_wait = min_events,
7033 struct io_rings *rings = ctx->rings;
7034 signed long timeout = MAX_SCHEDULE_TIMEOUT;
7038 io_cqring_overflow_flush(ctx, false);
7039 if (io_cqring_events(ctx) >= min_events)
7041 if (!io_run_task_work())
7046 #ifdef CONFIG_COMPAT
7047 if (in_compat_syscall())
7048 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
7052 ret = set_user_sigmask(sig, sigsz);
7059 struct timespec64 ts;
7061 if (get_timespec64(&ts, uts))
7063 timeout = timespec64_to_jiffies(&ts);
7066 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
7067 trace_io_uring_cqring_wait(ctx, min_events);
7069 /* if we can't even flush overflow, don't wait for more */
7070 if (!io_cqring_overflow_flush(ctx, false)) {
7074 prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq,
7075 TASK_INTERRUPTIBLE);
7076 ret = io_cqring_wait_schedule(ctx, &iowq, &timeout);
7077 finish_wait(&ctx->cq_wait, &iowq.wq);
7081 restore_saved_sigmask_unless(ret == -EINTR);
7083 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
7086 static void io_free_page_table(void **table, size_t size)
7088 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7090 for (i = 0; i < nr_tables; i++)
7095 static void **io_alloc_page_table(size_t size)
7097 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7098 size_t init_size = size;
7101 table = kcalloc(nr_tables, sizeof(*table), GFP_KERNEL);
7105 for (i = 0; i < nr_tables; i++) {
7106 unsigned int this_size = min_t(size_t, size, PAGE_SIZE);
7108 table[i] = kzalloc(this_size, GFP_KERNEL);
7110 io_free_page_table(table, init_size);
7118 static inline void io_rsrc_ref_lock(struct io_ring_ctx *ctx)
7120 spin_lock_bh(&ctx->rsrc_ref_lock);
7123 static inline void io_rsrc_ref_unlock(struct io_ring_ctx *ctx)
7125 spin_unlock_bh(&ctx->rsrc_ref_lock);
7128 static void io_rsrc_node_destroy(struct io_rsrc_node *ref_node)
7130 percpu_ref_exit(&ref_node->refs);
7134 static void io_rsrc_node_switch(struct io_ring_ctx *ctx,
7135 struct io_rsrc_data *data_to_kill)
7137 WARN_ON_ONCE(!ctx->rsrc_backup_node);
7138 WARN_ON_ONCE(data_to_kill && !ctx->rsrc_node);
7141 struct io_rsrc_node *rsrc_node = ctx->rsrc_node;
7143 rsrc_node->rsrc_data = data_to_kill;
7144 io_rsrc_ref_lock(ctx);
7145 list_add_tail(&rsrc_node->node, &ctx->rsrc_ref_list);
7146 io_rsrc_ref_unlock(ctx);
7148 atomic_inc(&data_to_kill->refs);
7149 percpu_ref_kill(&rsrc_node->refs);
7150 ctx->rsrc_node = NULL;
7153 if (!ctx->rsrc_node) {
7154 ctx->rsrc_node = ctx->rsrc_backup_node;
7155 ctx->rsrc_backup_node = NULL;
7159 static int io_rsrc_node_switch_start(struct io_ring_ctx *ctx)
7161 if (ctx->rsrc_backup_node)
7163 ctx->rsrc_backup_node = io_rsrc_node_alloc(ctx);
7164 return ctx->rsrc_backup_node ? 0 : -ENOMEM;
7167 static int io_rsrc_ref_quiesce(struct io_rsrc_data *data, struct io_ring_ctx *ctx)
7171 /* As we may drop ->uring_lock, other task may have started quiesce */
7175 data->quiesce = true;
7177 ret = io_rsrc_node_switch_start(ctx);
7180 io_rsrc_node_switch(ctx, data);
7182 /* kill initial ref, already quiesced if zero */
7183 if (atomic_dec_and_test(&data->refs))
7185 flush_delayed_work(&ctx->rsrc_put_work);
7186 ret = wait_for_completion_interruptible(&data->done);
7190 atomic_inc(&data->refs);
7191 /* wait for all works potentially completing data->done */
7192 flush_delayed_work(&ctx->rsrc_put_work);
7193 reinit_completion(&data->done);
7195 mutex_unlock(&ctx->uring_lock);
7196 ret = io_run_task_work_sig();
7197 mutex_lock(&ctx->uring_lock);
7199 data->quiesce = false;
7204 static u64 *io_get_tag_slot(struct io_rsrc_data *data, unsigned int idx)
7206 unsigned int off = idx & IO_RSRC_TAG_TABLE_MASK;
7207 unsigned int table_idx = idx >> IO_RSRC_TAG_TABLE_SHIFT;
7209 return &data->tags[table_idx][off];
7212 static void io_rsrc_data_free(struct io_rsrc_data *data)
7214 size_t size = data->nr * sizeof(data->tags[0][0]);
7217 io_free_page_table((void **)data->tags, size);
7221 static int io_rsrc_data_alloc(struct io_ring_ctx *ctx, rsrc_put_fn *do_put,
7222 u64 __user *utags, unsigned nr,
7223 struct io_rsrc_data **pdata)
7225 struct io_rsrc_data *data;
7229 data = kzalloc(sizeof(*data), GFP_KERNEL);
7232 data->tags = (u64 **)io_alloc_page_table(nr * sizeof(data->tags[0][0]));
7240 data->do_put = do_put;
7243 for (i = 0; i < nr; i++) {
7244 u64 *tag_slot = io_get_tag_slot(data, i);
7246 if (copy_from_user(tag_slot, &utags[i],
7252 atomic_set(&data->refs, 1);
7253 init_completion(&data->done);
7257 io_rsrc_data_free(data);
7261 static bool io_alloc_file_tables(struct io_file_table *table, unsigned nr_files)
7263 size_t size = nr_files * sizeof(struct io_fixed_file);
7265 table->files = (struct io_fixed_file **)io_alloc_page_table(size);
7266 return !!table->files;
7269 static void io_free_file_tables(struct io_file_table *table, unsigned nr_files)
7271 size_t size = nr_files * sizeof(struct io_fixed_file);
7273 io_free_page_table((void **)table->files, size);
7274 table->files = NULL;
7277 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
7279 #if defined(CONFIG_UNIX)
7280 if (ctx->ring_sock) {
7281 struct sock *sock = ctx->ring_sock->sk;
7282 struct sk_buff *skb;
7284 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
7290 for (i = 0; i < ctx->nr_user_files; i++) {
7293 file = io_file_from_index(ctx, i);
7298 io_free_file_tables(&ctx->file_table, ctx->nr_user_files);
7299 io_rsrc_data_free(ctx->file_data);
7300 ctx->file_data = NULL;
7301 ctx->nr_user_files = 0;
7304 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7308 if (!ctx->file_data)
7310 ret = io_rsrc_ref_quiesce(ctx->file_data, ctx);
7312 __io_sqe_files_unregister(ctx);
7316 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7317 __releases(&sqd->lock)
7319 WARN_ON_ONCE(sqd->thread == current);
7322 * Do the dance but not conditional clear_bit() because it'd race with
7323 * other threads incrementing park_pending and setting the bit.
7325 clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7326 if (atomic_dec_return(&sqd->park_pending))
7327 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7328 mutex_unlock(&sqd->lock);
7331 static void io_sq_thread_park(struct io_sq_data *sqd)
7332 __acquires(&sqd->lock)
7334 WARN_ON_ONCE(sqd->thread == current);
7336 atomic_inc(&sqd->park_pending);
7337 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7338 mutex_lock(&sqd->lock);
7340 wake_up_process(sqd->thread);
7343 static void io_sq_thread_stop(struct io_sq_data *sqd)
7345 WARN_ON_ONCE(sqd->thread == current);
7346 WARN_ON_ONCE(test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state));
7348 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7349 mutex_lock(&sqd->lock);
7351 wake_up_process(sqd->thread);
7352 mutex_unlock(&sqd->lock);
7353 wait_for_completion(&sqd->exited);
7356 static void io_put_sq_data(struct io_sq_data *sqd)
7358 if (refcount_dec_and_test(&sqd->refs)) {
7359 WARN_ON_ONCE(atomic_read(&sqd->park_pending));
7361 io_sq_thread_stop(sqd);
7366 static void io_sq_thread_finish(struct io_ring_ctx *ctx)
7368 struct io_sq_data *sqd = ctx->sq_data;
7371 io_sq_thread_park(sqd);
7372 list_del_init(&ctx->sqd_list);
7373 io_sqd_update_thread_idle(sqd);
7374 io_sq_thread_unpark(sqd);
7376 io_put_sq_data(sqd);
7377 ctx->sq_data = NULL;
7381 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7383 struct io_ring_ctx *ctx_attach;
7384 struct io_sq_data *sqd;
7387 f = fdget(p->wq_fd);
7389 return ERR_PTR(-ENXIO);
7390 if (f.file->f_op != &io_uring_fops) {
7392 return ERR_PTR(-EINVAL);
7395 ctx_attach = f.file->private_data;
7396 sqd = ctx_attach->sq_data;
7399 return ERR_PTR(-EINVAL);
7401 if (sqd->task_tgid != current->tgid) {
7403 return ERR_PTR(-EPERM);
7406 refcount_inc(&sqd->refs);
7411 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p,
7414 struct io_sq_data *sqd;
7417 if (p->flags & IORING_SETUP_ATTACH_WQ) {
7418 sqd = io_attach_sq_data(p);
7423 /* fall through for EPERM case, setup new sqd/task */
7424 if (PTR_ERR(sqd) != -EPERM)
7428 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7430 return ERR_PTR(-ENOMEM);
7432 atomic_set(&sqd->park_pending, 0);
7433 refcount_set(&sqd->refs, 1);
7434 INIT_LIST_HEAD(&sqd->ctx_list);
7435 mutex_init(&sqd->lock);
7436 init_waitqueue_head(&sqd->wait);
7437 init_completion(&sqd->exited);
7441 #if defined(CONFIG_UNIX)
7443 * Ensure the UNIX gc is aware of our file set, so we are certain that
7444 * the io_uring can be safely unregistered on process exit, even if we have
7445 * loops in the file referencing.
7447 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7449 struct sock *sk = ctx->ring_sock->sk;
7450 struct scm_fp_list *fpl;
7451 struct sk_buff *skb;
7454 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7458 skb = alloc_skb(0, GFP_KERNEL);
7467 fpl->user = get_uid(current_user());
7468 for (i = 0; i < nr; i++) {
7469 struct file *file = io_file_from_index(ctx, i + offset);
7473 fpl->fp[nr_files] = get_file(file);
7474 unix_inflight(fpl->user, fpl->fp[nr_files]);
7479 fpl->max = SCM_MAX_FD;
7480 fpl->count = nr_files;
7481 UNIXCB(skb).fp = fpl;
7482 skb->destructor = unix_destruct_scm;
7483 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7484 skb_queue_head(&sk->sk_receive_queue, skb);
7486 for (i = 0; i < nr_files; i++)
7497 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7498 * causes regular reference counting to break down. We rely on the UNIX
7499 * garbage collection to take care of this problem for us.
7501 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7503 unsigned left, total;
7507 left = ctx->nr_user_files;
7509 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7511 ret = __io_sqe_files_scm(ctx, this_files, total);
7515 total += this_files;
7521 while (total < ctx->nr_user_files) {
7522 struct file *file = io_file_from_index(ctx, total);
7532 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7538 static void io_rsrc_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
7540 struct file *file = prsrc->file;
7541 #if defined(CONFIG_UNIX)
7542 struct sock *sock = ctx->ring_sock->sk;
7543 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7544 struct sk_buff *skb;
7547 __skb_queue_head_init(&list);
7550 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7551 * remove this entry and rearrange the file array.
7553 skb = skb_dequeue(head);
7555 struct scm_fp_list *fp;
7557 fp = UNIXCB(skb).fp;
7558 for (i = 0; i < fp->count; i++) {
7561 if (fp->fp[i] != file)
7564 unix_notinflight(fp->user, fp->fp[i]);
7565 left = fp->count - 1 - i;
7567 memmove(&fp->fp[i], &fp->fp[i + 1],
7568 left * sizeof(struct file *));
7575 __skb_queue_tail(&list, skb);
7585 __skb_queue_tail(&list, skb);
7587 skb = skb_dequeue(head);
7590 if (skb_peek(&list)) {
7591 spin_lock_irq(&head->lock);
7592 while ((skb = __skb_dequeue(&list)) != NULL)
7593 __skb_queue_tail(head, skb);
7594 spin_unlock_irq(&head->lock);
7601 static void __io_rsrc_put_work(struct io_rsrc_node *ref_node)
7603 struct io_rsrc_data *rsrc_data = ref_node->rsrc_data;
7604 struct io_ring_ctx *ctx = rsrc_data->ctx;
7605 struct io_rsrc_put *prsrc, *tmp;
7607 list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
7608 list_del(&prsrc->list);
7611 bool lock_ring = ctx->flags & IORING_SETUP_IOPOLL;
7613 io_ring_submit_lock(ctx, lock_ring);
7614 spin_lock_irq(&ctx->completion_lock);
7615 io_cqring_fill_event(ctx, prsrc->tag, 0, 0);
7617 io_commit_cqring(ctx);
7618 spin_unlock_irq(&ctx->completion_lock);
7619 io_cqring_ev_posted(ctx);
7620 io_ring_submit_unlock(ctx, lock_ring);
7623 rsrc_data->do_put(ctx, prsrc);
7627 io_rsrc_node_destroy(ref_node);
7628 if (atomic_dec_and_test(&rsrc_data->refs))
7629 complete(&rsrc_data->done);
7632 static void io_rsrc_put_work(struct work_struct *work)
7634 struct io_ring_ctx *ctx;
7635 struct llist_node *node;
7637 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
7638 node = llist_del_all(&ctx->rsrc_put_llist);
7641 struct io_rsrc_node *ref_node;
7642 struct llist_node *next = node->next;
7644 ref_node = llist_entry(node, struct io_rsrc_node, llist);
7645 __io_rsrc_put_work(ref_node);
7650 static void io_rsrc_node_ref_zero(struct percpu_ref *ref)
7652 struct io_rsrc_node *node = container_of(ref, struct io_rsrc_node, refs);
7653 struct io_ring_ctx *ctx = node->rsrc_data->ctx;
7654 bool first_add = false;
7656 io_rsrc_ref_lock(ctx);
7659 while (!list_empty(&ctx->rsrc_ref_list)) {
7660 node = list_first_entry(&ctx->rsrc_ref_list,
7661 struct io_rsrc_node, node);
7662 /* recycle ref nodes in order */
7665 list_del(&node->node);
7666 first_add |= llist_add(&node->llist, &ctx->rsrc_put_llist);
7668 io_rsrc_ref_unlock(ctx);
7671 mod_delayed_work(system_wq, &ctx->rsrc_put_work, HZ);
7674 static struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx)
7676 struct io_rsrc_node *ref_node;
7678 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7682 if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
7687 INIT_LIST_HEAD(&ref_node->node);
7688 INIT_LIST_HEAD(&ref_node->rsrc_list);
7689 ref_node->done = false;
7693 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7694 unsigned nr_args, u64 __user *tags)
7696 __s32 __user *fds = (__s32 __user *) arg;
7705 if (nr_args > IORING_MAX_FIXED_FILES)
7707 ret = io_rsrc_node_switch_start(ctx);
7710 ret = io_rsrc_data_alloc(ctx, io_rsrc_file_put, tags, nr_args,
7716 if (!io_alloc_file_tables(&ctx->file_table, nr_args))
7719 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7720 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7724 /* allow sparse sets */
7727 if (unlikely(*io_get_tag_slot(ctx->file_data, i)))
7734 if (unlikely(!file))
7738 * Don't allow io_uring instances to be registered. If UNIX
7739 * isn't enabled, then this causes a reference cycle and this
7740 * instance can never get freed. If UNIX is enabled we'll
7741 * handle it just fine, but there's still no point in allowing
7742 * a ring fd as it doesn't support regular read/write anyway.
7744 if (file->f_op == &io_uring_fops) {
7748 io_fixed_file_set(io_fixed_file_slot(&ctx->file_table, i), file);
7751 ret = io_sqe_files_scm(ctx);
7753 __io_sqe_files_unregister(ctx);
7757 io_rsrc_node_switch(ctx, NULL);
7760 for (i = 0; i < ctx->nr_user_files; i++) {
7761 file = io_file_from_index(ctx, i);
7765 io_free_file_tables(&ctx->file_table, nr_args);
7766 ctx->nr_user_files = 0;
7768 io_rsrc_data_free(ctx->file_data);
7769 ctx->file_data = NULL;
7773 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7776 #if defined(CONFIG_UNIX)
7777 struct sock *sock = ctx->ring_sock->sk;
7778 struct sk_buff_head *head = &sock->sk_receive_queue;
7779 struct sk_buff *skb;
7782 * See if we can merge this file into an existing skb SCM_RIGHTS
7783 * file set. If there's no room, fall back to allocating a new skb
7784 * and filling it in.
7786 spin_lock_irq(&head->lock);
7787 skb = skb_peek(head);
7789 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7791 if (fpl->count < SCM_MAX_FD) {
7792 __skb_unlink(skb, head);
7793 spin_unlock_irq(&head->lock);
7794 fpl->fp[fpl->count] = get_file(file);
7795 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7797 spin_lock_irq(&head->lock);
7798 __skb_queue_head(head, skb);
7803 spin_unlock_irq(&head->lock);
7810 return __io_sqe_files_scm(ctx, 1, index);
7816 static int io_queue_rsrc_removal(struct io_rsrc_data *data, unsigned idx,
7817 struct io_rsrc_node *node, void *rsrc)
7819 struct io_rsrc_put *prsrc;
7821 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
7825 prsrc->tag = *io_get_tag_slot(data, idx);
7827 list_add(&prsrc->list, &node->rsrc_list);
7831 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7832 struct io_uring_rsrc_update2 *up,
7835 u64 __user *tags = u64_to_user_ptr(up->tags);
7836 __s32 __user *fds = u64_to_user_ptr(up->data);
7837 struct io_rsrc_data *data = ctx->file_data;
7838 struct io_fixed_file *file_slot;
7842 bool needs_switch = false;
7844 if (!ctx->file_data)
7846 if (up->offset + nr_args > ctx->nr_user_files)
7849 for (done = 0; done < nr_args; done++) {
7852 if ((tags && copy_from_user(&tag, &tags[done], sizeof(tag))) ||
7853 copy_from_user(&fd, &fds[done], sizeof(fd))) {
7857 if ((fd == IORING_REGISTER_FILES_SKIP || fd == -1) && tag) {
7861 if (fd == IORING_REGISTER_FILES_SKIP)
7864 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
7865 file_slot = io_fixed_file_slot(&ctx->file_table, i);
7867 if (file_slot->file_ptr) {
7868 file = (struct file *)(file_slot->file_ptr & FFS_MASK);
7869 err = io_queue_rsrc_removal(data, up->offset + done,
7870 ctx->rsrc_node, file);
7873 file_slot->file_ptr = 0;
7874 needs_switch = true;
7883 * Don't allow io_uring instances to be registered. If
7884 * UNIX isn't enabled, then this causes a reference
7885 * cycle and this instance can never get freed. If UNIX
7886 * is enabled we'll handle it just fine, but there's
7887 * still no point in allowing a ring fd as it doesn't
7888 * support regular read/write anyway.
7890 if (file->f_op == &io_uring_fops) {
7895 *io_get_tag_slot(data, up->offset + done) = tag;
7896 io_fixed_file_set(file_slot, file);
7897 err = io_sqe_file_register(ctx, file, i);
7899 file_slot->file_ptr = 0;
7907 io_rsrc_node_switch(ctx, data);
7908 return done ? done : err;
7911 static struct io_wq_work *io_free_work(struct io_wq_work *work)
7913 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7915 req = io_put_req_find_next(req);
7916 return req ? &req->work : NULL;
7919 static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx,
7920 struct task_struct *task)
7922 struct io_wq_hash *hash;
7923 struct io_wq_data data;
7924 unsigned int concurrency;
7926 hash = ctx->hash_map;
7928 hash = kzalloc(sizeof(*hash), GFP_KERNEL);
7930 return ERR_PTR(-ENOMEM);
7931 refcount_set(&hash->refs, 1);
7932 init_waitqueue_head(&hash->wait);
7933 ctx->hash_map = hash;
7938 data.free_work = io_free_work;
7939 data.do_work = io_wq_submit_work;
7941 /* Do QD, or 4 * CPUS, whatever is smallest */
7942 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7944 return io_wq_create(concurrency, &data);
7947 static int io_uring_alloc_task_context(struct task_struct *task,
7948 struct io_ring_ctx *ctx)
7950 struct io_uring_task *tctx;
7953 tctx = kzalloc(sizeof(*tctx), GFP_KERNEL);
7954 if (unlikely(!tctx))
7957 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
7958 if (unlikely(ret)) {
7963 tctx->io_wq = io_init_wq_offload(ctx, task);
7964 if (IS_ERR(tctx->io_wq)) {
7965 ret = PTR_ERR(tctx->io_wq);
7966 percpu_counter_destroy(&tctx->inflight);
7972 init_waitqueue_head(&tctx->wait);
7973 atomic_set(&tctx->in_idle, 0);
7974 atomic_set(&tctx->inflight_tracked, 0);
7975 task->io_uring = tctx;
7976 spin_lock_init(&tctx->task_lock);
7977 INIT_WQ_LIST(&tctx->task_list);
7978 init_task_work(&tctx->task_work, tctx_task_work);
7982 void __io_uring_free(struct task_struct *tsk)
7984 struct io_uring_task *tctx = tsk->io_uring;
7986 WARN_ON_ONCE(!xa_empty(&tctx->xa));
7987 WARN_ON_ONCE(tctx->io_wq);
7988 WARN_ON_ONCE(tctx->cached_refs);
7990 percpu_counter_destroy(&tctx->inflight);
7992 tsk->io_uring = NULL;
7995 static int io_sq_offload_create(struct io_ring_ctx *ctx,
7996 struct io_uring_params *p)
8000 /* Retain compatibility with failing for an invalid attach attempt */
8001 if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
8002 IORING_SETUP_ATTACH_WQ) {
8005 f = fdget(p->wq_fd);
8009 if (f.file->f_op != &io_uring_fops)
8012 if (ctx->flags & IORING_SETUP_SQPOLL) {
8013 struct task_struct *tsk;
8014 struct io_sq_data *sqd;
8017 sqd = io_get_sq_data(p, &attached);
8023 ctx->sq_creds = get_current_cred();
8025 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
8026 if (!ctx->sq_thread_idle)
8027 ctx->sq_thread_idle = HZ;
8029 io_sq_thread_park(sqd);
8030 list_add(&ctx->sqd_list, &sqd->ctx_list);
8031 io_sqd_update_thread_idle(sqd);
8032 /* don't attach to a dying SQPOLL thread, would be racy */
8033 ret = (attached && !sqd->thread) ? -ENXIO : 0;
8034 io_sq_thread_unpark(sqd);
8041 if (p->flags & IORING_SETUP_SQ_AFF) {
8042 int cpu = p->sq_thread_cpu;
8045 if (cpu >= nr_cpu_ids || !cpu_online(cpu))
8052 sqd->task_pid = current->pid;
8053 sqd->task_tgid = current->tgid;
8054 tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE);
8061 ret = io_uring_alloc_task_context(tsk, ctx);
8062 wake_up_new_task(tsk);
8065 } else if (p->flags & IORING_SETUP_SQ_AFF) {
8066 /* Can't have SQ_AFF without SQPOLL */
8073 complete(&ctx->sq_data->exited);
8075 io_sq_thread_finish(ctx);
8079 static inline void __io_unaccount_mem(struct user_struct *user,
8080 unsigned long nr_pages)
8082 atomic_long_sub(nr_pages, &user->locked_vm);
8085 static inline int __io_account_mem(struct user_struct *user,
8086 unsigned long nr_pages)
8088 unsigned long page_limit, cur_pages, new_pages;
8090 /* Don't allow more pages than we can safely lock */
8091 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
8094 cur_pages = atomic_long_read(&user->locked_vm);
8095 new_pages = cur_pages + nr_pages;
8096 if (new_pages > page_limit)
8098 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
8099 new_pages) != cur_pages);
8104 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8107 __io_unaccount_mem(ctx->user, nr_pages);
8109 if (ctx->mm_account)
8110 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
8113 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8118 ret = __io_account_mem(ctx->user, nr_pages);
8123 if (ctx->mm_account)
8124 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
8129 static void io_mem_free(void *ptr)
8136 page = virt_to_head_page(ptr);
8137 if (put_page_testzero(page))
8138 free_compound_page(page);
8141 static void *io_mem_alloc(size_t size)
8143 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
8144 __GFP_NORETRY | __GFP_ACCOUNT;
8146 return (void *) __get_free_pages(gfp_flags, get_order(size));
8149 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8152 struct io_rings *rings;
8153 size_t off, sq_array_size;
8155 off = struct_size(rings, cqes, cq_entries);
8156 if (off == SIZE_MAX)
8160 off = ALIGN(off, SMP_CACHE_BYTES);
8168 sq_array_size = array_size(sizeof(u32), sq_entries);
8169 if (sq_array_size == SIZE_MAX)
8172 if (check_add_overflow(off, sq_array_size, &off))
8178 static void io_buffer_unmap(struct io_ring_ctx *ctx, struct io_mapped_ubuf **slot)
8180 struct io_mapped_ubuf *imu = *slot;
8183 if (imu != ctx->dummy_ubuf) {
8184 for (i = 0; i < imu->nr_bvecs; i++)
8185 unpin_user_page(imu->bvec[i].bv_page);
8186 if (imu->acct_pages)
8187 io_unaccount_mem(ctx, imu->acct_pages);
8193 static void io_rsrc_buf_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
8195 io_buffer_unmap(ctx, &prsrc->buf);
8199 static void __io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8203 for (i = 0; i < ctx->nr_user_bufs; i++)
8204 io_buffer_unmap(ctx, &ctx->user_bufs[i]);
8205 kfree(ctx->user_bufs);
8206 io_rsrc_data_free(ctx->buf_data);
8207 ctx->user_bufs = NULL;
8208 ctx->buf_data = NULL;
8209 ctx->nr_user_bufs = 0;
8212 static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8219 ret = io_rsrc_ref_quiesce(ctx->buf_data, ctx);
8221 __io_sqe_buffers_unregister(ctx);
8225 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8226 void __user *arg, unsigned index)
8228 struct iovec __user *src;
8230 #ifdef CONFIG_COMPAT
8232 struct compat_iovec __user *ciovs;
8233 struct compat_iovec ciov;
8235 ciovs = (struct compat_iovec __user *) arg;
8236 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8239 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8240 dst->iov_len = ciov.iov_len;
8244 src = (struct iovec __user *) arg;
8245 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8251 * Not super efficient, but this is just a registration time. And we do cache
8252 * the last compound head, so generally we'll only do a full search if we don't
8255 * We check if the given compound head page has already been accounted, to
8256 * avoid double accounting it. This allows us to account the full size of the
8257 * page, not just the constituent pages of a huge page.
8259 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8260 int nr_pages, struct page *hpage)
8264 /* check current page array */
8265 for (i = 0; i < nr_pages; i++) {
8266 if (!PageCompound(pages[i]))
8268 if (compound_head(pages[i]) == hpage)
8272 /* check previously registered pages */
8273 for (i = 0; i < ctx->nr_user_bufs; i++) {
8274 struct io_mapped_ubuf *imu = ctx->user_bufs[i];
8276 for (j = 0; j < imu->nr_bvecs; j++) {
8277 if (!PageCompound(imu->bvec[j].bv_page))
8279 if (compound_head(imu->bvec[j].bv_page) == hpage)
8287 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8288 int nr_pages, struct io_mapped_ubuf *imu,
8289 struct page **last_hpage)
8293 imu->acct_pages = 0;
8294 for (i = 0; i < nr_pages; i++) {
8295 if (!PageCompound(pages[i])) {
8300 hpage = compound_head(pages[i]);
8301 if (hpage == *last_hpage)
8303 *last_hpage = hpage;
8304 if (headpage_already_acct(ctx, pages, i, hpage))
8306 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8310 if (!imu->acct_pages)
8313 ret = io_account_mem(ctx, imu->acct_pages);
8315 imu->acct_pages = 0;
8319 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
8320 struct io_mapped_ubuf **pimu,
8321 struct page **last_hpage)
8323 struct io_mapped_ubuf *imu = NULL;
8324 struct vm_area_struct **vmas = NULL;
8325 struct page **pages = NULL;
8326 unsigned long off, start, end, ubuf;
8328 int ret, pret, nr_pages, i;
8330 if (!iov->iov_base) {
8331 *pimu = ctx->dummy_ubuf;
8335 ubuf = (unsigned long) iov->iov_base;
8336 end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8337 start = ubuf >> PAGE_SHIFT;
8338 nr_pages = end - start;
8343 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
8347 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
8352 imu = kvmalloc(struct_size(imu, bvec, nr_pages), GFP_KERNEL);
8357 mmap_read_lock(current->mm);
8358 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
8360 if (pret == nr_pages) {
8361 /* don't support file backed memory */
8362 for (i = 0; i < nr_pages; i++) {
8363 struct vm_area_struct *vma = vmas[i];
8365 if (vma_is_shmem(vma))
8368 !is_file_hugepages(vma->vm_file)) {
8374 ret = pret < 0 ? pret : -EFAULT;
8376 mmap_read_unlock(current->mm);
8379 * if we did partial map, or found file backed vmas,
8380 * release any pages we did get
8383 unpin_user_pages(pages, pret);
8387 ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
8389 unpin_user_pages(pages, pret);
8393 off = ubuf & ~PAGE_MASK;
8394 size = iov->iov_len;
8395 for (i = 0; i < nr_pages; i++) {
8398 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8399 imu->bvec[i].bv_page = pages[i];
8400 imu->bvec[i].bv_len = vec_len;
8401 imu->bvec[i].bv_offset = off;
8405 /* store original address for later verification */
8407 imu->ubuf_end = ubuf + iov->iov_len;
8408 imu->nr_bvecs = nr_pages;
8419 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
8421 ctx->user_bufs = kcalloc(nr_args, sizeof(*ctx->user_bufs), GFP_KERNEL);
8422 return ctx->user_bufs ? 0 : -ENOMEM;
8425 static int io_buffer_validate(struct iovec *iov)
8427 unsigned long tmp, acct_len = iov->iov_len + (PAGE_SIZE - 1);
8430 * Don't impose further limits on the size and buffer
8431 * constraints here, we'll -EINVAL later when IO is
8432 * submitted if they are wrong.
8435 return iov->iov_len ? -EFAULT : 0;
8439 /* arbitrary limit, but we need something */
8440 if (iov->iov_len > SZ_1G)
8443 if (check_add_overflow((unsigned long)iov->iov_base, acct_len, &tmp))
8449 static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
8450 unsigned int nr_args, u64 __user *tags)
8452 struct page *last_hpage = NULL;
8453 struct io_rsrc_data *data;
8459 if (!nr_args || nr_args > IORING_MAX_REG_BUFFERS)
8461 ret = io_rsrc_node_switch_start(ctx);
8464 ret = io_rsrc_data_alloc(ctx, io_rsrc_buf_put, tags, nr_args, &data);
8467 ret = io_buffers_map_alloc(ctx, nr_args);
8469 io_rsrc_data_free(data);
8473 for (i = 0; i < nr_args; i++, ctx->nr_user_bufs++) {
8474 ret = io_copy_iov(ctx, &iov, arg, i);
8477 ret = io_buffer_validate(&iov);
8480 if (!iov.iov_base && *io_get_tag_slot(data, i)) {
8485 ret = io_sqe_buffer_register(ctx, &iov, &ctx->user_bufs[i],
8491 WARN_ON_ONCE(ctx->buf_data);
8493 ctx->buf_data = data;
8495 __io_sqe_buffers_unregister(ctx);
8497 io_rsrc_node_switch(ctx, NULL);
8501 static int __io_sqe_buffers_update(struct io_ring_ctx *ctx,
8502 struct io_uring_rsrc_update2 *up,
8503 unsigned int nr_args)
8505 u64 __user *tags = u64_to_user_ptr(up->tags);
8506 struct iovec iov, __user *iovs = u64_to_user_ptr(up->data);
8507 struct page *last_hpage = NULL;
8508 bool needs_switch = false;
8514 if (up->offset + nr_args > ctx->nr_user_bufs)
8517 for (done = 0; done < nr_args; done++) {
8518 struct io_mapped_ubuf *imu;
8519 int offset = up->offset + done;
8522 err = io_copy_iov(ctx, &iov, iovs, done);
8525 if (tags && copy_from_user(&tag, &tags[done], sizeof(tag))) {
8529 err = io_buffer_validate(&iov);
8532 if (!iov.iov_base && tag) {
8536 err = io_sqe_buffer_register(ctx, &iov, &imu, &last_hpage);
8540 i = array_index_nospec(offset, ctx->nr_user_bufs);
8541 if (ctx->user_bufs[i] != ctx->dummy_ubuf) {
8542 err = io_queue_rsrc_removal(ctx->buf_data, offset,
8543 ctx->rsrc_node, ctx->user_bufs[i]);
8544 if (unlikely(err)) {
8545 io_buffer_unmap(ctx, &imu);
8548 ctx->user_bufs[i] = NULL;
8549 needs_switch = true;
8552 ctx->user_bufs[i] = imu;
8553 *io_get_tag_slot(ctx->buf_data, offset) = tag;
8557 io_rsrc_node_switch(ctx, ctx->buf_data);
8558 return done ? done : err;
8561 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8563 __s32 __user *fds = arg;
8569 if (copy_from_user(&fd, fds, sizeof(*fds)))
8572 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8573 if (IS_ERR(ctx->cq_ev_fd)) {
8574 int ret = PTR_ERR(ctx->cq_ev_fd);
8575 ctx->cq_ev_fd = NULL;
8582 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8584 if (ctx->cq_ev_fd) {
8585 eventfd_ctx_put(ctx->cq_ev_fd);
8586 ctx->cq_ev_fd = NULL;
8593 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8595 struct io_buffer *buf;
8596 unsigned long index;
8598 xa_for_each(&ctx->io_buffers, index, buf)
8599 __io_remove_buffers(ctx, buf, index, -1U);
8602 static void io_req_cache_free(struct list_head *list, struct task_struct *tsk)
8604 struct io_kiocb *req, *nxt;
8606 list_for_each_entry_safe(req, nxt, list, compl.list) {
8607 if (tsk && req->task != tsk)
8609 list_del(&req->compl.list);
8610 kmem_cache_free(req_cachep, req);
8614 static void io_req_caches_free(struct io_ring_ctx *ctx)
8616 struct io_submit_state *submit_state = &ctx->submit_state;
8617 struct io_comp_state *cs = &ctx->submit_state.comp;
8619 mutex_lock(&ctx->uring_lock);
8621 if (submit_state->free_reqs) {
8622 kmem_cache_free_bulk(req_cachep, submit_state->free_reqs,
8623 submit_state->reqs);
8624 submit_state->free_reqs = 0;
8627 io_flush_cached_locked_reqs(ctx, cs);
8628 io_req_cache_free(&cs->free_list, NULL);
8629 mutex_unlock(&ctx->uring_lock);
8632 static bool io_wait_rsrc_data(struct io_rsrc_data *data)
8636 if (!atomic_dec_and_test(&data->refs))
8637 wait_for_completion(&data->done);
8641 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8643 io_sq_thread_finish(ctx);
8645 if (ctx->mm_account) {
8646 mmdrop(ctx->mm_account);
8647 ctx->mm_account = NULL;
8650 mutex_lock(&ctx->uring_lock);
8651 if (io_wait_rsrc_data(ctx->buf_data))
8652 __io_sqe_buffers_unregister(ctx);
8653 if (io_wait_rsrc_data(ctx->file_data))
8654 __io_sqe_files_unregister(ctx);
8656 __io_cqring_overflow_flush(ctx, true);
8657 mutex_unlock(&ctx->uring_lock);
8658 io_eventfd_unregister(ctx);
8659 io_destroy_buffers(ctx);
8661 put_cred(ctx->sq_creds);
8663 /* there are no registered resources left, nobody uses it */
8665 io_rsrc_node_destroy(ctx->rsrc_node);
8666 if (ctx->rsrc_backup_node)
8667 io_rsrc_node_destroy(ctx->rsrc_backup_node);
8668 flush_delayed_work(&ctx->rsrc_put_work);
8670 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
8671 WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist));
8673 #if defined(CONFIG_UNIX)
8674 if (ctx->ring_sock) {
8675 ctx->ring_sock->file = NULL; /* so that iput() is called */
8676 sock_release(ctx->ring_sock);
8680 io_mem_free(ctx->rings);
8681 io_mem_free(ctx->sq_sqes);
8683 percpu_ref_exit(&ctx->refs);
8684 free_uid(ctx->user);
8685 io_req_caches_free(ctx);
8687 io_wq_put_hash(ctx->hash_map);
8688 kfree(ctx->cancel_hash);
8689 kfree(ctx->dummy_ubuf);
8693 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8695 struct io_ring_ctx *ctx = file->private_data;
8698 poll_wait(file, &ctx->poll_wait, wait);
8700 * synchronizes with barrier from wq_has_sleeper call in
8704 if (!io_sqring_full(ctx))
8705 mask |= EPOLLOUT | EPOLLWRNORM;
8708 * Don't flush cqring overflow list here, just do a simple check.
8709 * Otherwise there could possible be ABBA deadlock:
8712 * lock(&ctx->uring_lock);
8714 * lock(&ctx->uring_lock);
8717 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
8718 * pushs them to do the flush.
8720 if (io_cqring_events(ctx) || test_bit(0, &ctx->check_cq_overflow))
8721 mask |= EPOLLIN | EPOLLRDNORM;
8726 static int io_uring_fasync(int fd, struct file *file, int on)
8728 struct io_ring_ctx *ctx = file->private_data;
8730 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8733 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8735 const struct cred *creds;
8737 creds = xa_erase(&ctx->personalities, id);
8746 static inline bool io_run_ctx_fallback(struct io_ring_ctx *ctx)
8748 return io_run_task_work_head(&ctx->exit_task_work);
8751 struct io_tctx_exit {
8752 struct callback_head task_work;
8753 struct completion completion;
8754 struct io_ring_ctx *ctx;
8757 static void io_tctx_exit_cb(struct callback_head *cb)
8759 struct io_uring_task *tctx = current->io_uring;
8760 struct io_tctx_exit *work;
8762 work = container_of(cb, struct io_tctx_exit, task_work);
8764 * When @in_idle, we're in cancellation and it's racy to remove the
8765 * node. It'll be removed by the end of cancellation, just ignore it.
8767 if (!atomic_read(&tctx->in_idle))
8768 io_uring_del_tctx_node((unsigned long)work->ctx);
8769 complete(&work->completion);
8772 static bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
8774 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8776 return req->ctx == data;
8779 static void io_ring_exit_work(struct work_struct *work)
8781 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
8782 unsigned long timeout = jiffies + HZ * 60 * 5;
8783 struct io_tctx_exit exit;
8784 struct io_tctx_node *node;
8788 * If we're doing polled IO and end up having requests being
8789 * submitted async (out-of-line), then completions can come in while
8790 * we're waiting for refs to drop. We need to reap these manually,
8791 * as nobody else will be looking for them.
8794 io_uring_try_cancel_requests(ctx, NULL, true);
8796 struct io_sq_data *sqd = ctx->sq_data;
8797 struct task_struct *tsk;
8799 io_sq_thread_park(sqd);
8801 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
8802 io_wq_cancel_cb(tsk->io_uring->io_wq,
8803 io_cancel_ctx_cb, ctx, true);
8804 io_sq_thread_unpark(sqd);
8807 WARN_ON_ONCE(time_after(jiffies, timeout));
8808 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
8810 init_completion(&exit.completion);
8811 init_task_work(&exit.task_work, io_tctx_exit_cb);
8814 * Some may use context even when all refs and requests have been put,
8815 * and they are free to do so while still holding uring_lock or
8816 * completion_lock, see __io_req_task_submit(). Apart from other work,
8817 * this lock/unlock section also waits them to finish.
8819 mutex_lock(&ctx->uring_lock);
8820 while (!list_empty(&ctx->tctx_list)) {
8821 WARN_ON_ONCE(time_after(jiffies, timeout));
8823 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
8825 /* don't spin on a single task if cancellation failed */
8826 list_rotate_left(&ctx->tctx_list);
8827 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
8828 if (WARN_ON_ONCE(ret))
8830 wake_up_process(node->task);
8832 mutex_unlock(&ctx->uring_lock);
8833 wait_for_completion(&exit.completion);
8834 mutex_lock(&ctx->uring_lock);
8836 mutex_unlock(&ctx->uring_lock);
8837 spin_lock_irq(&ctx->completion_lock);
8838 spin_unlock_irq(&ctx->completion_lock);
8840 io_ring_ctx_free(ctx);
8843 /* Returns true if we found and killed one or more timeouts */
8844 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
8847 struct io_kiocb *req, *tmp;
8850 spin_lock_irq(&ctx->completion_lock);
8851 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
8852 if (io_match_task(req, tsk, cancel_all)) {
8853 io_kill_timeout(req, -ECANCELED);
8858 io_commit_cqring(ctx);
8859 spin_unlock_irq(&ctx->completion_lock);
8861 io_cqring_ev_posted(ctx);
8862 return canceled != 0;
8865 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8867 unsigned long index;
8868 struct creds *creds;
8870 mutex_lock(&ctx->uring_lock);
8871 percpu_ref_kill(&ctx->refs);
8873 __io_cqring_overflow_flush(ctx, true);
8874 xa_for_each(&ctx->personalities, index, creds)
8875 io_unregister_personality(ctx, index);
8876 mutex_unlock(&ctx->uring_lock);
8878 io_kill_timeouts(ctx, NULL, true);
8879 io_poll_remove_all(ctx, NULL, true);
8881 /* if we failed setting up the ctx, we might not have any rings */
8882 io_iopoll_try_reap_events(ctx);
8884 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8886 * Use system_unbound_wq to avoid spawning tons of event kworkers
8887 * if we're exiting a ton of rings at the same time. It just adds
8888 * noise and overhead, there's no discernable change in runtime
8889 * over using system_wq.
8891 queue_work(system_unbound_wq, &ctx->exit_work);
8894 static int io_uring_release(struct inode *inode, struct file *file)
8896 struct io_ring_ctx *ctx = file->private_data;
8898 file->private_data = NULL;
8899 io_ring_ctx_wait_and_kill(ctx);
8903 struct io_task_cancel {
8904 struct task_struct *task;
8908 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8910 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8911 struct io_task_cancel *cancel = data;
8914 if (!cancel->all && (req->flags & REQ_F_LINK_TIMEOUT)) {
8915 unsigned long flags;
8916 struct io_ring_ctx *ctx = req->ctx;
8918 /* protect against races with linked timeouts */
8919 spin_lock_irqsave(&ctx->completion_lock, flags);
8920 ret = io_match_task(req, cancel->task, cancel->all);
8921 spin_unlock_irqrestore(&ctx->completion_lock, flags);
8923 ret = io_match_task(req, cancel->task, cancel->all);
8928 static bool io_cancel_defer_files(struct io_ring_ctx *ctx,
8929 struct task_struct *task, bool cancel_all)
8931 struct io_defer_entry *de;
8934 spin_lock_irq(&ctx->completion_lock);
8935 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8936 if (io_match_task(de->req, task, cancel_all)) {
8937 list_cut_position(&list, &ctx->defer_list, &de->list);
8941 spin_unlock_irq(&ctx->completion_lock);
8942 if (list_empty(&list))
8945 while (!list_empty(&list)) {
8946 de = list_first_entry(&list, struct io_defer_entry, list);
8947 list_del_init(&de->list);
8948 io_req_complete_failed(de->req, -ECANCELED);
8954 static bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
8956 struct io_tctx_node *node;
8957 enum io_wq_cancel cret;
8960 mutex_lock(&ctx->uring_lock);
8961 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
8962 struct io_uring_task *tctx = node->task->io_uring;
8965 * io_wq will stay alive while we hold uring_lock, because it's
8966 * killed after ctx nodes, which requires to take the lock.
8968 if (!tctx || !tctx->io_wq)
8970 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
8971 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8973 mutex_unlock(&ctx->uring_lock);
8978 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
8979 struct task_struct *task,
8982 struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
8983 struct io_uring_task *tctx = task ? task->io_uring : NULL;
8986 enum io_wq_cancel cret;
8990 ret |= io_uring_try_cancel_iowq(ctx);
8991 } else if (tctx && tctx->io_wq) {
8993 * Cancels requests of all rings, not only @ctx, but
8994 * it's fine as the task is in exit/exec.
8996 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
8998 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
9001 /* SQPOLL thread does its own polling */
9002 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
9003 (ctx->sq_data && ctx->sq_data->thread == current)) {
9004 while (!list_empty_careful(&ctx->iopoll_list)) {
9005 io_iopoll_try_reap_events(ctx);
9010 ret |= io_cancel_defer_files(ctx, task, cancel_all);
9011 ret |= io_poll_remove_all(ctx, task, cancel_all);
9012 ret |= io_kill_timeouts(ctx, task, cancel_all);
9013 ret |= io_run_task_work();
9014 ret |= io_run_ctx_fallback(ctx);
9021 static int __io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9023 struct io_uring_task *tctx = current->io_uring;
9024 struct io_tctx_node *node;
9027 if (unlikely(!tctx)) {
9028 ret = io_uring_alloc_task_context(current, ctx);
9031 tctx = current->io_uring;
9033 if (!xa_load(&tctx->xa, (unsigned long)ctx)) {
9034 node = kmalloc(sizeof(*node), GFP_KERNEL);
9038 node->task = current;
9040 ret = xa_err(xa_store(&tctx->xa, (unsigned long)ctx,
9047 mutex_lock(&ctx->uring_lock);
9048 list_add(&node->ctx_node, &ctx->tctx_list);
9049 mutex_unlock(&ctx->uring_lock);
9056 * Note that this task has used io_uring. We use it for cancelation purposes.
9058 static inline int io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9060 struct io_uring_task *tctx = current->io_uring;
9062 if (likely(tctx && tctx->last == ctx))
9064 return __io_uring_add_tctx_node(ctx);
9068 * Remove this io_uring_file -> task mapping.
9070 static void io_uring_del_tctx_node(unsigned long index)
9072 struct io_uring_task *tctx = current->io_uring;
9073 struct io_tctx_node *node;
9077 node = xa_erase(&tctx->xa, index);
9081 WARN_ON_ONCE(current != node->task);
9082 WARN_ON_ONCE(list_empty(&node->ctx_node));
9084 mutex_lock(&node->ctx->uring_lock);
9085 list_del(&node->ctx_node);
9086 mutex_unlock(&node->ctx->uring_lock);
9088 if (tctx->last == node->ctx)
9093 static void io_uring_clean_tctx(struct io_uring_task *tctx)
9095 struct io_wq *wq = tctx->io_wq;
9096 struct io_tctx_node *node;
9097 unsigned long index;
9099 xa_for_each(&tctx->xa, index, node)
9100 io_uring_del_tctx_node(index);
9103 * Must be after io_uring_del_task_file() (removes nodes under
9104 * uring_lock) to avoid race with io_uring_try_cancel_iowq().
9107 io_wq_put_and_exit(wq);
9111 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
9114 return atomic_read(&tctx->inflight_tracked);
9115 return percpu_counter_sum(&tctx->inflight);
9118 static void io_uring_drop_tctx_refs(struct task_struct *task)
9120 struct io_uring_task *tctx = task->io_uring;
9121 unsigned int refs = tctx->cached_refs;
9123 tctx->cached_refs = 0;
9124 percpu_counter_sub(&tctx->inflight, refs);
9125 put_task_struct_many(task, refs);
9129 * Find any io_uring ctx that this task has registered or done IO on, and cancel
9130 * requests. @sqd should be not-null IIF it's an SQPOLL thread cancellation.
9132 static void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
9134 struct io_uring_task *tctx = current->io_uring;
9135 struct io_ring_ctx *ctx;
9139 WARN_ON_ONCE(sqd && sqd->thread != current);
9141 if (!current->io_uring)
9144 io_wq_exit_start(tctx->io_wq);
9146 io_uring_drop_tctx_refs(current);
9147 atomic_inc(&tctx->in_idle);
9149 /* read completions before cancelations */
9150 inflight = tctx_inflight(tctx, !cancel_all);
9155 struct io_tctx_node *node;
9156 unsigned long index;
9158 xa_for_each(&tctx->xa, index, node) {
9159 /* sqpoll task will cancel all its requests */
9160 if (node->ctx->sq_data)
9162 io_uring_try_cancel_requests(node->ctx, current,
9166 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
9167 io_uring_try_cancel_requests(ctx, current,
9171 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
9173 * If we've seen completions, retry without waiting. This
9174 * avoids a race where a completion comes in before we did
9175 * prepare_to_wait().
9177 if (inflight == tctx_inflight(tctx, !cancel_all))
9179 finish_wait(&tctx->wait, &wait);
9181 atomic_dec(&tctx->in_idle);
9183 io_uring_clean_tctx(tctx);
9185 /* for exec all current's requests should be gone, kill tctx */
9186 __io_uring_free(current);
9190 void __io_uring_cancel(struct files_struct *files)
9192 io_uring_cancel_generic(!files, NULL);
9195 static void *io_uring_validate_mmap_request(struct file *file,
9196 loff_t pgoff, size_t sz)
9198 struct io_ring_ctx *ctx = file->private_data;
9199 loff_t offset = pgoff << PAGE_SHIFT;
9204 case IORING_OFF_SQ_RING:
9205 case IORING_OFF_CQ_RING:
9208 case IORING_OFF_SQES:
9212 return ERR_PTR(-EINVAL);
9215 page = virt_to_head_page(ptr);
9216 if (sz > page_size(page))
9217 return ERR_PTR(-EINVAL);
9224 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9226 size_t sz = vma->vm_end - vma->vm_start;
9230 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
9232 return PTR_ERR(ptr);
9234 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
9235 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
9238 #else /* !CONFIG_MMU */
9240 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9242 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
9245 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
9247 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
9250 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
9251 unsigned long addr, unsigned long len,
9252 unsigned long pgoff, unsigned long flags)
9256 ptr = io_uring_validate_mmap_request(file, pgoff, len);
9258 return PTR_ERR(ptr);
9260 return (unsigned long) ptr;
9263 #endif /* !CONFIG_MMU */
9265 static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
9270 if (!io_sqring_full(ctx))
9272 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9274 if (!io_sqring_full(ctx))
9277 } while (!signal_pending(current));
9279 finish_wait(&ctx->sqo_sq_wait, &wait);
9283 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9284 struct __kernel_timespec __user **ts,
9285 const sigset_t __user **sig)
9287 struct io_uring_getevents_arg arg;
9290 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9291 * is just a pointer to the sigset_t.
9293 if (!(flags & IORING_ENTER_EXT_ARG)) {
9294 *sig = (const sigset_t __user *) argp;
9300 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9301 * timespec and sigset_t pointers if good.
9303 if (*argsz != sizeof(arg))
9305 if (copy_from_user(&arg, argp, sizeof(arg)))
9307 *sig = u64_to_user_ptr(arg.sigmask);
9308 *argsz = arg.sigmask_sz;
9309 *ts = u64_to_user_ptr(arg.ts);
9313 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9314 u32, min_complete, u32, flags, const void __user *, argp,
9317 struct io_ring_ctx *ctx;
9324 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9325 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG)))
9329 if (unlikely(!f.file))
9333 if (unlikely(f.file->f_op != &io_uring_fops))
9337 ctx = f.file->private_data;
9338 if (unlikely(!percpu_ref_tryget(&ctx->refs)))
9342 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
9346 * For SQ polling, the thread will do all submissions and completions.
9347 * Just return the requested submit count, and wake the thread if
9351 if (ctx->flags & IORING_SETUP_SQPOLL) {
9352 io_cqring_overflow_flush(ctx, false);
9355 if (unlikely(ctx->sq_data->thread == NULL)) {
9358 if (flags & IORING_ENTER_SQ_WAKEUP)
9359 wake_up(&ctx->sq_data->wait);
9360 if (flags & IORING_ENTER_SQ_WAIT) {
9361 ret = io_sqpoll_wait_sq(ctx);
9365 submitted = to_submit;
9366 } else if (to_submit) {
9367 ret = io_uring_add_tctx_node(ctx);
9370 mutex_lock(&ctx->uring_lock);
9371 submitted = io_submit_sqes(ctx, to_submit);
9372 mutex_unlock(&ctx->uring_lock);
9374 if (submitted != to_submit)
9377 if (flags & IORING_ENTER_GETEVENTS) {
9378 const sigset_t __user *sig;
9379 struct __kernel_timespec __user *ts;
9381 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
9385 min_complete = min(min_complete, ctx->cq_entries);
9388 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9389 * space applications don't need to do io completion events
9390 * polling again, they can rely on io_sq_thread to do polling
9391 * work, which can reduce cpu usage and uring_lock contention.
9393 if (ctx->flags & IORING_SETUP_IOPOLL &&
9394 !(ctx->flags & IORING_SETUP_SQPOLL)) {
9395 ret = io_iopoll_check(ctx, min_complete);
9397 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
9402 percpu_ref_put(&ctx->refs);
9405 return submitted ? submitted : ret;
9408 #ifdef CONFIG_PROC_FS
9409 static int io_uring_show_cred(struct seq_file *m, unsigned int id,
9410 const struct cred *cred)
9412 struct user_namespace *uns = seq_user_ns(m);
9413 struct group_info *gi;
9418 seq_printf(m, "%5d\n", id);
9419 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
9420 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
9421 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
9422 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
9423 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
9424 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
9425 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
9426 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
9427 seq_puts(m, "\n\tGroups:\t");
9428 gi = cred->group_info;
9429 for (g = 0; g < gi->ngroups; g++) {
9430 seq_put_decimal_ull(m, g ? " " : "",
9431 from_kgid_munged(uns, gi->gid[g]));
9433 seq_puts(m, "\n\tCapEff:\t");
9434 cap = cred->cap_effective;
9435 CAP_FOR_EACH_U32(__capi)
9436 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
9441 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
9443 struct io_sq_data *sq = NULL;
9448 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9449 * since fdinfo case grabs it in the opposite direction of normal use
9450 * cases. If we fail to get the lock, we just don't iterate any
9451 * structures that could be going away outside the io_uring mutex.
9453 has_lock = mutex_trylock(&ctx->uring_lock);
9455 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) {
9461 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9462 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9463 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9464 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9465 struct file *f = io_file_from_index(ctx, i);
9468 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9470 seq_printf(m, "%5u: <none>\n", i);
9472 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9473 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9474 struct io_mapped_ubuf *buf = ctx->user_bufs[i];
9475 unsigned int len = buf->ubuf_end - buf->ubuf;
9477 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf, len);
9479 if (has_lock && !xa_empty(&ctx->personalities)) {
9480 unsigned long index;
9481 const struct cred *cred;
9483 seq_printf(m, "Personalities:\n");
9484 xa_for_each(&ctx->personalities, index, cred)
9485 io_uring_show_cred(m, index, cred);
9487 seq_printf(m, "PollList:\n");
9488 spin_lock_irq(&ctx->completion_lock);
9489 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9490 struct hlist_head *list = &ctx->cancel_hash[i];
9491 struct io_kiocb *req;
9493 hlist_for_each_entry(req, list, hash_node)
9494 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9495 req->task->task_works != NULL);
9497 spin_unlock_irq(&ctx->completion_lock);
9499 mutex_unlock(&ctx->uring_lock);
9502 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9504 struct io_ring_ctx *ctx = f->private_data;
9506 if (percpu_ref_tryget(&ctx->refs)) {
9507 __io_uring_show_fdinfo(ctx, m);
9508 percpu_ref_put(&ctx->refs);
9513 static const struct file_operations io_uring_fops = {
9514 .release = io_uring_release,
9515 .mmap = io_uring_mmap,
9517 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9518 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9520 .poll = io_uring_poll,
9521 .fasync = io_uring_fasync,
9522 #ifdef CONFIG_PROC_FS
9523 .show_fdinfo = io_uring_show_fdinfo,
9527 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9528 struct io_uring_params *p)
9530 struct io_rings *rings;
9531 size_t size, sq_array_offset;
9533 /* make sure these are sane, as we already accounted them */
9534 ctx->sq_entries = p->sq_entries;
9535 ctx->cq_entries = p->cq_entries;
9537 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9538 if (size == SIZE_MAX)
9541 rings = io_mem_alloc(size);
9546 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9547 rings->sq_ring_mask = p->sq_entries - 1;
9548 rings->cq_ring_mask = p->cq_entries - 1;
9549 rings->sq_ring_entries = p->sq_entries;
9550 rings->cq_ring_entries = p->cq_entries;
9552 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9553 if (size == SIZE_MAX) {
9554 io_mem_free(ctx->rings);
9559 ctx->sq_sqes = io_mem_alloc(size);
9560 if (!ctx->sq_sqes) {
9561 io_mem_free(ctx->rings);
9569 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
9573 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9577 ret = io_uring_add_tctx_node(ctx);
9582 fd_install(fd, file);
9587 * Allocate an anonymous fd, this is what constitutes the application
9588 * visible backing of an io_uring instance. The application mmaps this
9589 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9590 * we have to tie this fd to a socket for file garbage collection purposes.
9592 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
9595 #if defined(CONFIG_UNIX)
9598 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9601 return ERR_PTR(ret);
9604 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9605 O_RDWR | O_CLOEXEC);
9606 #if defined(CONFIG_UNIX)
9608 sock_release(ctx->ring_sock);
9609 ctx->ring_sock = NULL;
9611 ctx->ring_sock->file = file;
9617 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9618 struct io_uring_params __user *params)
9620 struct io_ring_ctx *ctx;
9626 if (entries > IORING_MAX_ENTRIES) {
9627 if (!(p->flags & IORING_SETUP_CLAMP))
9629 entries = IORING_MAX_ENTRIES;
9633 * Use twice as many entries for the CQ ring. It's possible for the
9634 * application to drive a higher depth than the size of the SQ ring,
9635 * since the sqes are only used at submission time. This allows for
9636 * some flexibility in overcommitting a bit. If the application has
9637 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9638 * of CQ ring entries manually.
9640 p->sq_entries = roundup_pow_of_two(entries);
9641 if (p->flags & IORING_SETUP_CQSIZE) {
9643 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9644 * to a power-of-two, if it isn't already. We do NOT impose
9645 * any cq vs sq ring sizing.
9649 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9650 if (!(p->flags & IORING_SETUP_CLAMP))
9652 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9654 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9655 if (p->cq_entries < p->sq_entries)
9658 p->cq_entries = 2 * p->sq_entries;
9661 ctx = io_ring_ctx_alloc(p);
9664 ctx->compat = in_compat_syscall();
9665 if (!capable(CAP_IPC_LOCK))
9666 ctx->user = get_uid(current_user());
9669 * This is just grabbed for accounting purposes. When a process exits,
9670 * the mm is exited and dropped before the files, hence we need to hang
9671 * on to this mm purely for the purposes of being able to unaccount
9672 * memory (locked/pinned vm). It's not used for anything else.
9674 mmgrab(current->mm);
9675 ctx->mm_account = current->mm;
9677 ret = io_allocate_scq_urings(ctx, p);
9681 ret = io_sq_offload_create(ctx, p);
9684 /* always set a rsrc node */
9685 ret = io_rsrc_node_switch_start(ctx);
9688 io_rsrc_node_switch(ctx, NULL);
9690 memset(&p->sq_off, 0, sizeof(p->sq_off));
9691 p->sq_off.head = offsetof(struct io_rings, sq.head);
9692 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9693 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9694 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9695 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9696 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9697 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9699 memset(&p->cq_off, 0, sizeof(p->cq_off));
9700 p->cq_off.head = offsetof(struct io_rings, cq.head);
9701 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9702 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9703 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9704 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9705 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9706 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9708 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9709 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9710 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9711 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
9712 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
9713 IORING_FEAT_RSRC_TAGS;
9715 if (copy_to_user(params, p, sizeof(*p))) {
9720 file = io_uring_get_file(ctx);
9722 ret = PTR_ERR(file);
9727 * Install ring fd as the very last thing, so we don't risk someone
9728 * having closed it before we finish setup
9730 ret = io_uring_install_fd(ctx, file);
9732 /* fput will clean it up */
9737 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9740 io_ring_ctx_wait_and_kill(ctx);
9745 * Sets up an aio uring context, and returns the fd. Applications asks for a
9746 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9747 * params structure passed in.
9749 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9751 struct io_uring_params p;
9754 if (copy_from_user(&p, params, sizeof(p)))
9756 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9761 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9762 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9763 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9764 IORING_SETUP_R_DISABLED))
9767 return io_uring_create(entries, &p, params);
9770 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9771 struct io_uring_params __user *, params)
9773 return io_uring_setup(entries, params);
9776 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9778 struct io_uring_probe *p;
9782 size = struct_size(p, ops, nr_args);
9783 if (size == SIZE_MAX)
9785 p = kzalloc(size, GFP_KERNEL);
9790 if (copy_from_user(p, arg, size))
9793 if (memchr_inv(p, 0, size))
9796 p->last_op = IORING_OP_LAST - 1;
9797 if (nr_args > IORING_OP_LAST)
9798 nr_args = IORING_OP_LAST;
9800 for (i = 0; i < nr_args; i++) {
9802 if (!io_op_defs[i].not_supported)
9803 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9808 if (copy_to_user(arg, p, size))
9815 static int io_register_personality(struct io_ring_ctx *ctx)
9817 const struct cred *creds;
9821 creds = get_current_cred();
9823 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
9824 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
9831 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9832 unsigned int nr_args)
9834 struct io_uring_restriction *res;
9838 /* Restrictions allowed only if rings started disabled */
9839 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9842 /* We allow only a single restrictions registration */
9843 if (ctx->restrictions.registered)
9846 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9849 size = array_size(nr_args, sizeof(*res));
9850 if (size == SIZE_MAX)
9853 res = memdup_user(arg, size);
9855 return PTR_ERR(res);
9859 for (i = 0; i < nr_args; i++) {
9860 switch (res[i].opcode) {
9861 case IORING_RESTRICTION_REGISTER_OP:
9862 if (res[i].register_op >= IORING_REGISTER_LAST) {
9867 __set_bit(res[i].register_op,
9868 ctx->restrictions.register_op);
9870 case IORING_RESTRICTION_SQE_OP:
9871 if (res[i].sqe_op >= IORING_OP_LAST) {
9876 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9878 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9879 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9881 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9882 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9891 /* Reset all restrictions if an error happened */
9893 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9895 ctx->restrictions.registered = true;
9901 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9903 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9906 if (ctx->restrictions.registered)
9907 ctx->restricted = 1;
9909 ctx->flags &= ~IORING_SETUP_R_DISABLED;
9910 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
9911 wake_up(&ctx->sq_data->wait);
9915 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
9916 struct io_uring_rsrc_update2 *up,
9924 if (check_add_overflow(up->offset, nr_args, &tmp))
9926 err = io_rsrc_node_switch_start(ctx);
9931 case IORING_RSRC_FILE:
9932 return __io_sqe_files_update(ctx, up, nr_args);
9933 case IORING_RSRC_BUFFER:
9934 return __io_sqe_buffers_update(ctx, up, nr_args);
9939 static int io_register_files_update(struct io_ring_ctx *ctx, void __user *arg,
9942 struct io_uring_rsrc_update2 up;
9946 memset(&up, 0, sizeof(up));
9947 if (copy_from_user(&up, arg, sizeof(struct io_uring_rsrc_update)))
9949 return __io_register_rsrc_update(ctx, IORING_RSRC_FILE, &up, nr_args);
9952 static int io_register_rsrc_update(struct io_ring_ctx *ctx, void __user *arg,
9953 unsigned size, unsigned type)
9955 struct io_uring_rsrc_update2 up;
9957 if (size != sizeof(up))
9959 if (copy_from_user(&up, arg, sizeof(up)))
9961 if (!up.nr || up.resv)
9963 return __io_register_rsrc_update(ctx, type, &up, up.nr);
9966 static int io_register_rsrc(struct io_ring_ctx *ctx, void __user *arg,
9967 unsigned int size, unsigned int type)
9969 struct io_uring_rsrc_register rr;
9971 /* keep it extendible */
9972 if (size != sizeof(rr))
9975 memset(&rr, 0, sizeof(rr));
9976 if (copy_from_user(&rr, arg, size))
9978 if (!rr.nr || rr.resv || rr.resv2)
9982 case IORING_RSRC_FILE:
9983 return io_sqe_files_register(ctx, u64_to_user_ptr(rr.data),
9984 rr.nr, u64_to_user_ptr(rr.tags));
9985 case IORING_RSRC_BUFFER:
9986 return io_sqe_buffers_register(ctx, u64_to_user_ptr(rr.data),
9987 rr.nr, u64_to_user_ptr(rr.tags));
9992 static int io_register_iowq_aff(struct io_ring_ctx *ctx, void __user *arg,
9995 struct io_uring_task *tctx = current->io_uring;
9996 cpumask_var_t new_mask;
9999 if (!tctx || !tctx->io_wq)
10002 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
10005 cpumask_clear(new_mask);
10006 if (len > cpumask_size())
10007 len = cpumask_size();
10009 if (copy_from_user(new_mask, arg, len)) {
10010 free_cpumask_var(new_mask);
10014 ret = io_wq_cpu_affinity(tctx->io_wq, new_mask);
10015 free_cpumask_var(new_mask);
10019 static int io_unregister_iowq_aff(struct io_ring_ctx *ctx)
10021 struct io_uring_task *tctx = current->io_uring;
10023 if (!tctx || !tctx->io_wq)
10026 return io_wq_cpu_affinity(tctx->io_wq, NULL);
10029 static bool io_register_op_must_quiesce(int op)
10032 case IORING_REGISTER_BUFFERS:
10033 case IORING_UNREGISTER_BUFFERS:
10034 case IORING_REGISTER_FILES:
10035 case IORING_UNREGISTER_FILES:
10036 case IORING_REGISTER_FILES_UPDATE:
10037 case IORING_REGISTER_PROBE:
10038 case IORING_REGISTER_PERSONALITY:
10039 case IORING_UNREGISTER_PERSONALITY:
10040 case IORING_REGISTER_FILES2:
10041 case IORING_REGISTER_FILES_UPDATE2:
10042 case IORING_REGISTER_BUFFERS2:
10043 case IORING_REGISTER_BUFFERS_UPDATE:
10044 case IORING_REGISTER_IOWQ_AFF:
10045 case IORING_UNREGISTER_IOWQ_AFF:
10052 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
10053 void __user *arg, unsigned nr_args)
10054 __releases(ctx->uring_lock)
10055 __acquires(ctx->uring_lock)
10060 * We're inside the ring mutex, if the ref is already dying, then
10061 * someone else killed the ctx or is already going through
10062 * io_uring_register().
10064 if (percpu_ref_is_dying(&ctx->refs))
10067 if (ctx->restricted) {
10068 if (opcode >= IORING_REGISTER_LAST)
10070 opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
10071 if (!test_bit(opcode, ctx->restrictions.register_op))
10075 if (io_register_op_must_quiesce(opcode)) {
10076 percpu_ref_kill(&ctx->refs);
10079 * Drop uring mutex before waiting for references to exit. If
10080 * another thread is currently inside io_uring_enter() it might
10081 * need to grab the uring_lock to make progress. If we hold it
10082 * here across the drain wait, then we can deadlock. It's safe
10083 * to drop the mutex here, since no new references will come in
10084 * after we've killed the percpu ref.
10086 mutex_unlock(&ctx->uring_lock);
10088 ret = wait_for_completion_interruptible(&ctx->ref_comp);
10091 ret = io_run_task_work_sig();
10095 mutex_lock(&ctx->uring_lock);
10098 io_refs_resurrect(&ctx->refs, &ctx->ref_comp);
10104 case IORING_REGISTER_BUFFERS:
10105 ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
10107 case IORING_UNREGISTER_BUFFERS:
10109 if (arg || nr_args)
10111 ret = io_sqe_buffers_unregister(ctx);
10113 case IORING_REGISTER_FILES:
10114 ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
10116 case IORING_UNREGISTER_FILES:
10118 if (arg || nr_args)
10120 ret = io_sqe_files_unregister(ctx);
10122 case IORING_REGISTER_FILES_UPDATE:
10123 ret = io_register_files_update(ctx, arg, nr_args);
10125 case IORING_REGISTER_EVENTFD:
10126 case IORING_REGISTER_EVENTFD_ASYNC:
10130 ret = io_eventfd_register(ctx, arg);
10133 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
10134 ctx->eventfd_async = 1;
10136 ctx->eventfd_async = 0;
10138 case IORING_UNREGISTER_EVENTFD:
10140 if (arg || nr_args)
10142 ret = io_eventfd_unregister(ctx);
10144 case IORING_REGISTER_PROBE:
10146 if (!arg || nr_args > 256)
10148 ret = io_probe(ctx, arg, nr_args);
10150 case IORING_REGISTER_PERSONALITY:
10152 if (arg || nr_args)
10154 ret = io_register_personality(ctx);
10156 case IORING_UNREGISTER_PERSONALITY:
10160 ret = io_unregister_personality(ctx, nr_args);
10162 case IORING_REGISTER_ENABLE_RINGS:
10164 if (arg || nr_args)
10166 ret = io_register_enable_rings(ctx);
10168 case IORING_REGISTER_RESTRICTIONS:
10169 ret = io_register_restrictions(ctx, arg, nr_args);
10171 case IORING_REGISTER_FILES2:
10172 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE);
10174 case IORING_REGISTER_FILES_UPDATE2:
10175 ret = io_register_rsrc_update(ctx, arg, nr_args,
10178 case IORING_REGISTER_BUFFERS2:
10179 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER);
10181 case IORING_REGISTER_BUFFERS_UPDATE:
10182 ret = io_register_rsrc_update(ctx, arg, nr_args,
10183 IORING_RSRC_BUFFER);
10185 case IORING_REGISTER_IOWQ_AFF:
10187 if (!arg || !nr_args)
10189 ret = io_register_iowq_aff(ctx, arg, nr_args);
10191 case IORING_UNREGISTER_IOWQ_AFF:
10193 if (arg || nr_args)
10195 ret = io_unregister_iowq_aff(ctx);
10202 if (io_register_op_must_quiesce(opcode)) {
10203 /* bring the ctx back to life */
10204 percpu_ref_reinit(&ctx->refs);
10205 reinit_completion(&ctx->ref_comp);
10210 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
10211 void __user *, arg, unsigned int, nr_args)
10213 struct io_ring_ctx *ctx;
10222 if (f.file->f_op != &io_uring_fops)
10225 ctx = f.file->private_data;
10227 io_run_task_work();
10229 mutex_lock(&ctx->uring_lock);
10230 ret = __io_uring_register(ctx, opcode, arg, nr_args);
10231 mutex_unlock(&ctx->uring_lock);
10232 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
10233 ctx->cq_ev_fd != NULL, ret);
10239 static int __init io_uring_init(void)
10241 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
10242 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
10243 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
10246 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
10247 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
10248 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
10249 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
10250 BUILD_BUG_SQE_ELEM(1, __u8, flags);
10251 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
10252 BUILD_BUG_SQE_ELEM(4, __s32, fd);
10253 BUILD_BUG_SQE_ELEM(8, __u64, off);
10254 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
10255 BUILD_BUG_SQE_ELEM(16, __u64, addr);
10256 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
10257 BUILD_BUG_SQE_ELEM(24, __u32, len);
10258 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
10259 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
10260 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
10261 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
10262 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
10263 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
10264 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
10265 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
10266 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
10267 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
10268 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
10269 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
10270 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
10271 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
10272 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
10273 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
10274 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
10275 BUILD_BUG_SQE_ELEM(42, __u16, personality);
10276 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
10278 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
10279 sizeof(struct io_uring_rsrc_update));
10280 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
10281 sizeof(struct io_uring_rsrc_update2));
10282 /* should fit into one byte */
10283 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
10285 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
10286 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
10287 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
10291 __initcall(io_uring_init);