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_used: 1;
356 } ____cacheline_aligned_in_smp;
358 /* submission data */
360 struct mutex uring_lock;
363 * Ring buffer of indices into array of io_uring_sqe, which is
364 * mmapped by the application using the IORING_OFF_SQES offset.
366 * This indirection could e.g. be used to assign fixed
367 * io_uring_sqe entries to operations and only submit them to
368 * the queue when needed.
370 * The kernel modifies neither the indices array nor the entries
374 struct io_uring_sqe *sq_sqes;
375 unsigned cached_sq_head;
377 struct list_head defer_list;
380 * Fixed resources fast path, should be accessed only under
381 * uring_lock, and updated through io_uring_register(2)
383 struct io_rsrc_node *rsrc_node;
384 struct io_file_table file_table;
385 unsigned nr_user_files;
386 unsigned nr_user_bufs;
387 struct io_mapped_ubuf **user_bufs;
389 struct io_submit_state submit_state;
390 struct list_head timeout_list;
391 struct list_head cq_overflow_list;
392 struct xarray io_buffers;
393 struct xarray personalities;
395 unsigned sq_thread_idle;
396 } ____cacheline_aligned_in_smp;
398 /* IRQ completion list, under ->completion_lock */
399 struct list_head locked_free_list;
400 unsigned int locked_free_nr;
402 const struct cred *sq_creds; /* cred used for __io_sq_thread() */
403 struct io_sq_data *sq_data; /* if using sq thread polling */
405 struct wait_queue_head sqo_sq_wait;
406 struct list_head sqd_list;
408 unsigned long check_cq_overflow;
411 unsigned cached_cq_tail;
413 struct eventfd_ctx *cq_ev_fd;
414 struct wait_queue_head poll_wait;
415 struct wait_queue_head cq_wait;
417 atomic_t cq_timeouts;
418 struct fasync_struct *cq_fasync;
419 unsigned cq_last_tm_flush;
420 } ____cacheline_aligned_in_smp;
423 spinlock_t completion_lock;
426 * ->iopoll_list is protected by the ctx->uring_lock for
427 * io_uring instances that don't use IORING_SETUP_SQPOLL.
428 * For SQPOLL, only the single threaded io_sq_thread() will
429 * manipulate the list, hence no extra locking is needed there.
431 struct list_head iopoll_list;
432 struct hlist_head *cancel_hash;
433 unsigned cancel_hash_bits;
434 bool poll_multi_file;
435 } ____cacheline_aligned_in_smp;
437 struct io_restriction restrictions;
439 /* slow path rsrc auxilary data, used by update/register */
441 struct io_rsrc_node *rsrc_backup_node;
442 struct io_mapped_ubuf *dummy_ubuf;
443 struct io_rsrc_data *file_data;
444 struct io_rsrc_data *buf_data;
446 struct delayed_work rsrc_put_work;
447 struct llist_head rsrc_put_llist;
448 struct list_head rsrc_ref_list;
449 spinlock_t rsrc_ref_lock;
452 /* Keep this last, we don't need it for the fast path */
454 #if defined(CONFIG_UNIX)
455 struct socket *ring_sock;
457 /* hashed buffered write serialization */
458 struct io_wq_hash *hash_map;
460 /* Only used for accounting purposes */
461 struct user_struct *user;
462 struct mm_struct *mm_account;
464 /* ctx exit and cancelation */
465 struct callback_head *exit_task_work;
466 struct work_struct exit_work;
467 struct list_head tctx_list;
468 struct completion ref_comp;
472 struct io_uring_task {
473 /* submission side */
476 struct wait_queue_head wait;
477 const struct io_ring_ctx *last;
479 struct percpu_counter inflight;
480 atomic_t inflight_tracked;
483 spinlock_t task_lock;
484 struct io_wq_work_list task_list;
485 unsigned long task_state;
486 struct callback_head task_work;
490 * First field must be the file pointer in all the
491 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
493 struct io_poll_iocb {
495 struct wait_queue_head *head;
499 struct wait_queue_entry wait;
502 struct io_poll_update {
508 bool update_user_data;
516 struct io_timeout_data {
517 struct io_kiocb *req;
518 struct hrtimer timer;
519 struct timespec64 ts;
520 enum hrtimer_mode mode;
525 struct sockaddr __user *addr;
526 int __user *addr_len;
528 unsigned long nofile;
548 struct list_head list;
549 /* head of the link, used by linked timeouts only */
550 struct io_kiocb *head;
553 struct io_timeout_rem {
558 struct timespec64 ts;
563 /* NOTE: kiocb has the file as the first member, so don't do it here */
571 struct sockaddr __user *addr;
578 struct compat_msghdr __user *umsg_compat;
579 struct user_msghdr __user *umsg;
585 struct io_buffer *kbuf;
591 struct filename *filename;
593 unsigned long nofile;
596 struct io_rsrc_update {
622 struct epoll_event event;
626 struct file *file_out;
627 struct file *file_in;
634 struct io_provide_buf {
648 const char __user *filename;
649 struct statx __user *buffer;
661 struct filename *oldpath;
662 struct filename *newpath;
670 struct filename *filename;
673 struct io_completion {
675 struct list_head list;
679 struct io_async_connect {
680 struct sockaddr_storage address;
683 struct io_async_msghdr {
684 struct iovec fast_iov[UIO_FASTIOV];
685 /* points to an allocated iov, if NULL we use fast_iov instead */
686 struct iovec *free_iov;
687 struct sockaddr __user *uaddr;
689 struct sockaddr_storage addr;
693 struct iovec fast_iov[UIO_FASTIOV];
694 const struct iovec *free_iovec;
695 struct iov_iter iter;
697 struct wait_page_queue wpq;
701 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
702 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
703 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
704 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
705 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
706 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
708 /* first byte is taken by user flags, shift it to not overlap */
713 REQ_F_LINK_TIMEOUT_BIT,
714 REQ_F_NEED_CLEANUP_BIT,
716 REQ_F_BUFFER_SELECTED_BIT,
717 REQ_F_LTIMEOUT_ACTIVE_BIT,
718 REQ_F_COMPLETE_INLINE_BIT,
720 REQ_F_DONT_REISSUE_BIT,
721 /* keep async read/write and isreg together and in order */
722 REQ_F_ASYNC_READ_BIT,
723 REQ_F_ASYNC_WRITE_BIT,
726 /* not a real bit, just to check we're not overflowing the space */
732 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
733 /* drain existing IO first */
734 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
736 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
737 /* doesn't sever on completion < 0 */
738 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
740 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
741 /* IOSQE_BUFFER_SELECT */
742 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
744 /* fail rest of links */
745 REQ_F_FAIL = BIT(REQ_F_FAIL_BIT),
746 /* on inflight list, should be cancelled and waited on exit reliably */
747 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
748 /* read/write uses file position */
749 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
750 /* must not punt to workers */
751 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
752 /* has or had linked timeout */
753 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
755 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
756 /* already went through poll handler */
757 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
758 /* buffer already selected */
759 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
760 /* linked timeout is active, i.e. prepared by link's head */
761 REQ_F_LTIMEOUT_ACTIVE = BIT(REQ_F_LTIMEOUT_ACTIVE_BIT),
762 /* completion is deferred through io_comp_state */
763 REQ_F_COMPLETE_INLINE = BIT(REQ_F_COMPLETE_INLINE_BIT),
764 /* caller should reissue async */
765 REQ_F_REISSUE = BIT(REQ_F_REISSUE_BIT),
766 /* don't attempt request reissue, see io_rw_reissue() */
767 REQ_F_DONT_REISSUE = BIT(REQ_F_DONT_REISSUE_BIT),
768 /* supports async reads */
769 REQ_F_ASYNC_READ = BIT(REQ_F_ASYNC_READ_BIT),
770 /* supports async writes */
771 REQ_F_ASYNC_WRITE = BIT(REQ_F_ASYNC_WRITE_BIT),
773 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
777 struct io_poll_iocb poll;
778 struct io_poll_iocb *double_poll;
781 struct io_task_work {
782 struct io_wq_work_node node;
783 task_work_func_t func;
787 IORING_RSRC_FILE = 0,
788 IORING_RSRC_BUFFER = 1,
792 * NOTE! Each of the iocb union members has the file pointer
793 * as the first entry in their struct definition. So you can
794 * access the file pointer through any of the sub-structs,
795 * or directly as just 'ki_filp' in this struct.
801 struct io_poll_iocb poll;
802 struct io_poll_update poll_update;
803 struct io_accept accept;
805 struct io_cancel cancel;
806 struct io_timeout timeout;
807 struct io_timeout_rem timeout_rem;
808 struct io_connect connect;
809 struct io_sr_msg sr_msg;
811 struct io_close close;
812 struct io_rsrc_update rsrc_update;
813 struct io_fadvise fadvise;
814 struct io_madvise madvise;
815 struct io_epoll epoll;
816 struct io_splice splice;
817 struct io_provide_buf pbuf;
818 struct io_statx statx;
819 struct io_shutdown shutdown;
820 struct io_rename rename;
821 struct io_unlink unlink;
822 /* use only after cleaning per-op data, see io_clean_op() */
823 struct io_completion compl;
826 /* opcode allocated if it needs to store data for async defer */
829 /* polled IO has completed */
835 struct io_ring_ctx *ctx;
838 struct task_struct *task;
841 struct io_kiocb *link;
842 struct percpu_ref *fixed_rsrc_refs;
844 /* used with ctx->iopoll_list with reads/writes */
845 struct list_head inflight_entry;
847 struct io_task_work io_task_work;
848 struct callback_head task_work;
850 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
851 struct hlist_node hash_node;
852 struct async_poll *apoll;
853 struct io_wq_work work;
854 /* store used ubuf, so we can prevent reloading */
855 struct io_mapped_ubuf *imu;
858 struct io_tctx_node {
859 struct list_head ctx_node;
860 struct task_struct *task;
861 struct io_ring_ctx *ctx;
864 struct io_defer_entry {
865 struct list_head list;
866 struct io_kiocb *req;
871 /* needs req->file assigned */
872 unsigned needs_file : 1;
873 /* hash wq insertion if file is a regular file */
874 unsigned hash_reg_file : 1;
875 /* unbound wq insertion if file is a non-regular file */
876 unsigned unbound_nonreg_file : 1;
877 /* opcode is not supported by this kernel */
878 unsigned not_supported : 1;
879 /* set if opcode supports polled "wait" */
881 unsigned pollout : 1;
882 /* op supports buffer selection */
883 unsigned buffer_select : 1;
884 /* do prep async if is going to be punted */
885 unsigned needs_async_setup : 1;
886 /* should block plug */
888 /* size of async data needed, if any */
889 unsigned short async_size;
892 static const struct io_op_def io_op_defs[] = {
893 [IORING_OP_NOP] = {},
894 [IORING_OP_READV] = {
896 .unbound_nonreg_file = 1,
899 .needs_async_setup = 1,
901 .async_size = sizeof(struct io_async_rw),
903 [IORING_OP_WRITEV] = {
906 .unbound_nonreg_file = 1,
908 .needs_async_setup = 1,
910 .async_size = sizeof(struct io_async_rw),
912 [IORING_OP_FSYNC] = {
915 [IORING_OP_READ_FIXED] = {
917 .unbound_nonreg_file = 1,
920 .async_size = sizeof(struct io_async_rw),
922 [IORING_OP_WRITE_FIXED] = {
925 .unbound_nonreg_file = 1,
928 .async_size = sizeof(struct io_async_rw),
930 [IORING_OP_POLL_ADD] = {
932 .unbound_nonreg_file = 1,
934 [IORING_OP_POLL_REMOVE] = {},
935 [IORING_OP_SYNC_FILE_RANGE] = {
938 [IORING_OP_SENDMSG] = {
940 .unbound_nonreg_file = 1,
942 .needs_async_setup = 1,
943 .async_size = sizeof(struct io_async_msghdr),
945 [IORING_OP_RECVMSG] = {
947 .unbound_nonreg_file = 1,
950 .needs_async_setup = 1,
951 .async_size = sizeof(struct io_async_msghdr),
953 [IORING_OP_TIMEOUT] = {
954 .async_size = sizeof(struct io_timeout_data),
956 [IORING_OP_TIMEOUT_REMOVE] = {
957 /* used by timeout updates' prep() */
959 [IORING_OP_ACCEPT] = {
961 .unbound_nonreg_file = 1,
964 [IORING_OP_ASYNC_CANCEL] = {},
965 [IORING_OP_LINK_TIMEOUT] = {
966 .async_size = sizeof(struct io_timeout_data),
968 [IORING_OP_CONNECT] = {
970 .unbound_nonreg_file = 1,
972 .needs_async_setup = 1,
973 .async_size = sizeof(struct io_async_connect),
975 [IORING_OP_FALLOCATE] = {
978 [IORING_OP_OPENAT] = {},
979 [IORING_OP_CLOSE] = {},
980 [IORING_OP_FILES_UPDATE] = {},
981 [IORING_OP_STATX] = {},
984 .unbound_nonreg_file = 1,
988 .async_size = sizeof(struct io_async_rw),
990 [IORING_OP_WRITE] = {
992 .unbound_nonreg_file = 1,
995 .async_size = sizeof(struct io_async_rw),
997 [IORING_OP_FADVISE] = {
1000 [IORING_OP_MADVISE] = {},
1001 [IORING_OP_SEND] = {
1003 .unbound_nonreg_file = 1,
1006 [IORING_OP_RECV] = {
1008 .unbound_nonreg_file = 1,
1012 [IORING_OP_OPENAT2] = {
1014 [IORING_OP_EPOLL_CTL] = {
1015 .unbound_nonreg_file = 1,
1017 [IORING_OP_SPLICE] = {
1020 .unbound_nonreg_file = 1,
1022 [IORING_OP_PROVIDE_BUFFERS] = {},
1023 [IORING_OP_REMOVE_BUFFERS] = {},
1027 .unbound_nonreg_file = 1,
1029 [IORING_OP_SHUTDOWN] = {
1032 [IORING_OP_RENAMEAT] = {},
1033 [IORING_OP_UNLINKAT] = {},
1036 static bool io_disarm_next(struct io_kiocb *req);
1037 static void io_uring_del_tctx_node(unsigned long index);
1038 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
1039 struct task_struct *task,
1041 static void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd);
1042 static struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx);
1044 static bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1045 long res, unsigned int cflags);
1046 static void io_put_req(struct io_kiocb *req);
1047 static void io_put_req_deferred(struct io_kiocb *req, int nr);
1048 static void io_dismantle_req(struct io_kiocb *req);
1049 static void io_put_task(struct task_struct *task, int nr);
1050 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
1051 static void io_queue_linked_timeout(struct io_kiocb *req);
1052 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
1053 struct io_uring_rsrc_update2 *up,
1055 static void io_clean_op(struct io_kiocb *req);
1056 static struct file *io_file_get(struct io_submit_state *state,
1057 struct io_kiocb *req, int fd, bool fixed);
1058 static void __io_queue_sqe(struct io_kiocb *req);
1059 static void io_rsrc_put_work(struct work_struct *work);
1061 static void io_req_task_queue(struct io_kiocb *req);
1062 static void io_submit_flush_completions(struct io_comp_state *cs,
1063 struct io_ring_ctx *ctx);
1064 static bool io_poll_remove_waitqs(struct io_kiocb *req);
1065 static int io_req_prep_async(struct io_kiocb *req);
1067 static struct kmem_cache *req_cachep;
1069 static const struct file_operations io_uring_fops;
1071 struct sock *io_uring_get_socket(struct file *file)
1073 #if defined(CONFIG_UNIX)
1074 if (file->f_op == &io_uring_fops) {
1075 struct io_ring_ctx *ctx = file->private_data;
1077 return ctx->ring_sock->sk;
1082 EXPORT_SYMBOL(io_uring_get_socket);
1084 #define io_for_each_link(pos, head) \
1085 for (pos = (head); pos; pos = pos->link)
1087 static inline void io_req_set_rsrc_node(struct io_kiocb *req)
1089 struct io_ring_ctx *ctx = req->ctx;
1091 if (!req->fixed_rsrc_refs) {
1092 req->fixed_rsrc_refs = &ctx->rsrc_node->refs;
1093 percpu_ref_get(req->fixed_rsrc_refs);
1097 static void io_refs_resurrect(struct percpu_ref *ref, struct completion *compl)
1099 bool got = percpu_ref_tryget(ref);
1101 /* already at zero, wait for ->release() */
1103 wait_for_completion(compl);
1104 percpu_ref_resurrect(ref);
1106 percpu_ref_put(ref);
1109 static bool io_match_task(struct io_kiocb *head, struct task_struct *task,
1112 struct io_kiocb *req;
1114 if (task && head->task != task)
1119 io_for_each_link(req, head) {
1120 if (req->flags & REQ_F_INFLIGHT)
1126 static inline void req_set_fail(struct io_kiocb *req)
1128 req->flags |= REQ_F_FAIL;
1131 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1133 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1135 complete(&ctx->ref_comp);
1138 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1140 return !req->timeout.off;
1143 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1145 struct io_ring_ctx *ctx;
1148 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1153 * Use 5 bits less than the max cq entries, that should give us around
1154 * 32 entries per hash list if totally full and uniformly spread.
1156 hash_bits = ilog2(p->cq_entries);
1160 ctx->cancel_hash_bits = hash_bits;
1161 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1163 if (!ctx->cancel_hash)
1165 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1167 ctx->dummy_ubuf = kzalloc(sizeof(*ctx->dummy_ubuf), GFP_KERNEL);
1168 if (!ctx->dummy_ubuf)
1170 /* set invalid range, so io_import_fixed() fails meeting it */
1171 ctx->dummy_ubuf->ubuf = -1UL;
1173 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1174 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1177 ctx->flags = p->flags;
1178 init_waitqueue_head(&ctx->sqo_sq_wait);
1179 INIT_LIST_HEAD(&ctx->sqd_list);
1180 init_waitqueue_head(&ctx->poll_wait);
1181 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1182 init_completion(&ctx->ref_comp);
1183 xa_init_flags(&ctx->io_buffers, XA_FLAGS_ALLOC1);
1184 xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
1185 mutex_init(&ctx->uring_lock);
1186 init_waitqueue_head(&ctx->cq_wait);
1187 spin_lock_init(&ctx->completion_lock);
1188 INIT_LIST_HEAD(&ctx->iopoll_list);
1189 INIT_LIST_HEAD(&ctx->defer_list);
1190 INIT_LIST_HEAD(&ctx->timeout_list);
1191 spin_lock_init(&ctx->rsrc_ref_lock);
1192 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
1193 INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
1194 init_llist_head(&ctx->rsrc_put_llist);
1195 INIT_LIST_HEAD(&ctx->tctx_list);
1196 INIT_LIST_HEAD(&ctx->submit_state.comp.free_list);
1197 INIT_LIST_HEAD(&ctx->locked_free_list);
1200 kfree(ctx->dummy_ubuf);
1201 kfree(ctx->cancel_hash);
1206 static void io_account_cq_overflow(struct io_ring_ctx *ctx)
1208 struct io_rings *r = ctx->rings;
1210 WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1);
1214 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1216 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1217 struct io_ring_ctx *ctx = req->ctx;
1219 return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail;
1225 static void io_req_track_inflight(struct io_kiocb *req)
1227 if (!(req->flags & REQ_F_INFLIGHT)) {
1228 req->flags |= REQ_F_INFLIGHT;
1229 atomic_inc(¤t->io_uring->inflight_tracked);
1233 static void io_prep_async_work(struct io_kiocb *req)
1235 const struct io_op_def *def = &io_op_defs[req->opcode];
1236 struct io_ring_ctx *ctx = req->ctx;
1238 if (!req->work.creds)
1239 req->work.creds = get_current_cred();
1241 req->work.list.next = NULL;
1242 req->work.flags = 0;
1243 if (req->flags & REQ_F_FORCE_ASYNC)
1244 req->work.flags |= IO_WQ_WORK_CONCURRENT;
1246 if (req->flags & REQ_F_ISREG) {
1247 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1248 io_wq_hash_work(&req->work, file_inode(req->file));
1249 } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
1250 if (def->unbound_nonreg_file)
1251 req->work.flags |= IO_WQ_WORK_UNBOUND;
1254 switch (req->opcode) {
1255 case IORING_OP_SPLICE:
1257 if (!S_ISREG(file_inode(req->splice.file_in)->i_mode))
1258 req->work.flags |= IO_WQ_WORK_UNBOUND;
1263 static void io_prep_async_link(struct io_kiocb *req)
1265 struct io_kiocb *cur;
1267 io_for_each_link(cur, req)
1268 io_prep_async_work(cur);
1271 static void io_queue_async_work(struct io_kiocb *req)
1273 struct io_ring_ctx *ctx = req->ctx;
1274 struct io_kiocb *link = io_prep_linked_timeout(req);
1275 struct io_uring_task *tctx = req->task->io_uring;
1278 BUG_ON(!tctx->io_wq);
1280 /* init ->work of the whole link before punting */
1281 io_prep_async_link(req);
1282 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1283 &req->work, req->flags);
1284 io_wq_enqueue(tctx->io_wq, &req->work);
1286 io_queue_linked_timeout(link);
1289 static void io_kill_timeout(struct io_kiocb *req, int status)
1290 __must_hold(&req->ctx->completion_lock)
1292 struct io_timeout_data *io = req->async_data;
1294 if (hrtimer_try_to_cancel(&io->timer) != -1) {
1295 atomic_set(&req->ctx->cq_timeouts,
1296 atomic_read(&req->ctx->cq_timeouts) + 1);
1297 list_del_init(&req->timeout.list);
1298 io_cqring_fill_event(req->ctx, req->user_data, status, 0);
1299 io_put_req_deferred(req, 1);
1303 static void io_queue_deferred(struct io_ring_ctx *ctx)
1305 while (!list_empty(&ctx->defer_list)) {
1306 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1307 struct io_defer_entry, list);
1309 if (req_need_defer(de->req, de->seq))
1311 list_del_init(&de->list);
1312 io_req_task_queue(de->req);
1317 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1319 u32 seq = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
1321 while (!list_empty(&ctx->timeout_list)) {
1322 u32 events_needed, events_got;
1323 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1324 struct io_kiocb, timeout.list);
1326 if (io_is_timeout_noseq(req))
1330 * Since seq can easily wrap around over time, subtract
1331 * the last seq at which timeouts were flushed before comparing.
1332 * Assuming not more than 2^31-1 events have happened since,
1333 * these subtractions won't have wrapped, so we can check if
1334 * target is in [last_seq, current_seq] by comparing the two.
1336 events_needed = req->timeout.target_seq - ctx->cq_last_tm_flush;
1337 events_got = seq - ctx->cq_last_tm_flush;
1338 if (events_got < events_needed)
1341 list_del_init(&req->timeout.list);
1342 io_kill_timeout(req, 0);
1344 ctx->cq_last_tm_flush = seq;
1347 static void io_commit_cqring(struct io_ring_ctx *ctx)
1349 if (unlikely(ctx->off_timeout_used || ctx->drain_used)) {
1350 if (ctx->off_timeout_used)
1351 io_flush_timeouts(ctx);
1352 if (ctx->drain_used)
1353 io_queue_deferred(ctx);
1355 /* order cqe stores with ring update */
1356 smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
1359 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1361 struct io_rings *r = ctx->rings;
1363 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == ctx->sq_entries;
1366 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
1368 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
1371 static inline struct io_uring_cqe *io_get_cqe(struct io_ring_ctx *ctx)
1373 struct io_rings *rings = ctx->rings;
1374 unsigned tail, mask = ctx->cq_entries - 1;
1377 * writes to the cq entry need to come after reading head; the
1378 * control dependency is enough as we're using WRITE_ONCE to
1381 if (__io_cqring_events(ctx) == ctx->cq_entries)
1384 tail = ctx->cached_cq_tail++;
1385 return &rings->cqes[tail & mask];
1388 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1390 if (likely(!ctx->cq_ev_fd))
1392 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1394 return !ctx->eventfd_async || io_wq_current_is_worker();
1397 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1399 /* see waitqueue_active() comment */
1402 if (waitqueue_active(&ctx->cq_wait))
1403 wake_up(&ctx->cq_wait);
1404 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1405 wake_up(&ctx->sq_data->wait);
1406 if (io_should_trigger_evfd(ctx))
1407 eventfd_signal(ctx->cq_ev_fd, 1);
1408 if (waitqueue_active(&ctx->poll_wait)) {
1409 wake_up_interruptible(&ctx->poll_wait);
1410 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1414 static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx)
1416 /* see waitqueue_active() comment */
1419 if (ctx->flags & IORING_SETUP_SQPOLL) {
1420 if (waitqueue_active(&ctx->cq_wait))
1421 wake_up(&ctx->cq_wait);
1423 if (io_should_trigger_evfd(ctx))
1424 eventfd_signal(ctx->cq_ev_fd, 1);
1425 if (waitqueue_active(&ctx->poll_wait)) {
1426 wake_up_interruptible(&ctx->poll_wait);
1427 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1431 /* Returns true if there are no backlogged entries after the flush */
1432 static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1434 unsigned long flags;
1435 bool all_flushed, posted;
1437 if (!force && __io_cqring_events(ctx) == ctx->cq_entries)
1441 spin_lock_irqsave(&ctx->completion_lock, flags);
1442 while (!list_empty(&ctx->cq_overflow_list)) {
1443 struct io_uring_cqe *cqe = io_get_cqe(ctx);
1444 struct io_overflow_cqe *ocqe;
1448 ocqe = list_first_entry(&ctx->cq_overflow_list,
1449 struct io_overflow_cqe, list);
1451 memcpy(cqe, &ocqe->cqe, sizeof(*cqe));
1453 io_account_cq_overflow(ctx);
1456 list_del(&ocqe->list);
1460 all_flushed = list_empty(&ctx->cq_overflow_list);
1462 clear_bit(0, &ctx->check_cq_overflow);
1463 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
1467 io_commit_cqring(ctx);
1468 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1470 io_cqring_ev_posted(ctx);
1474 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1478 if (test_bit(0, &ctx->check_cq_overflow)) {
1479 /* iopoll syncs against uring_lock, not completion_lock */
1480 if (ctx->flags & IORING_SETUP_IOPOLL)
1481 mutex_lock(&ctx->uring_lock);
1482 ret = __io_cqring_overflow_flush(ctx, force);
1483 if (ctx->flags & IORING_SETUP_IOPOLL)
1484 mutex_unlock(&ctx->uring_lock);
1491 * Shamelessly stolen from the mm implementation of page reference checking,
1492 * see commit f958d7b528b1 for details.
1494 #define req_ref_zero_or_close_to_overflow(req) \
1495 ((unsigned int) atomic_read(&(req->refs)) + 127u <= 127u)
1497 static inline bool req_ref_inc_not_zero(struct io_kiocb *req)
1499 return atomic_inc_not_zero(&req->refs);
1502 static inline bool req_ref_sub_and_test(struct io_kiocb *req, int refs)
1504 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1505 return atomic_sub_and_test(refs, &req->refs);
1508 static inline bool req_ref_put_and_test(struct io_kiocb *req)
1510 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1511 return atomic_dec_and_test(&req->refs);
1514 static inline void req_ref_put(struct io_kiocb *req)
1516 WARN_ON_ONCE(req_ref_put_and_test(req));
1519 static inline void req_ref_get(struct io_kiocb *req)
1521 WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
1522 atomic_inc(&req->refs);
1525 static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data,
1526 long res, unsigned int cflags)
1528 struct io_overflow_cqe *ocqe;
1530 ocqe = kmalloc(sizeof(*ocqe), GFP_ATOMIC | __GFP_ACCOUNT);
1533 * If we're in ring overflow flush mode, or in task cancel mode,
1534 * or cannot allocate an overflow entry, then we need to drop it
1537 io_account_cq_overflow(ctx);
1540 if (list_empty(&ctx->cq_overflow_list)) {
1541 set_bit(0, &ctx->check_cq_overflow);
1542 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
1544 ocqe->cqe.user_data = user_data;
1545 ocqe->cqe.res = res;
1546 ocqe->cqe.flags = cflags;
1547 list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
1551 static inline bool __io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1552 long res, unsigned int cflags)
1554 struct io_uring_cqe *cqe;
1556 trace_io_uring_complete(ctx, user_data, res, cflags);
1559 * If we can't get a cq entry, userspace overflowed the
1560 * submission (by quite a lot). Increment the overflow count in
1563 cqe = io_get_cqe(ctx);
1565 WRITE_ONCE(cqe->user_data, user_data);
1566 WRITE_ONCE(cqe->res, res);
1567 WRITE_ONCE(cqe->flags, cflags);
1570 return io_cqring_event_overflow(ctx, user_data, res, cflags);
1573 /* not as hot to bloat with inlining */
1574 static noinline bool io_cqring_fill_event(struct io_ring_ctx *ctx, u64 user_data,
1575 long res, unsigned int cflags)
1577 return __io_cqring_fill_event(ctx, user_data, res, cflags);
1580 static void io_req_complete_post(struct io_kiocb *req, long res,
1581 unsigned int cflags)
1583 struct io_ring_ctx *ctx = req->ctx;
1584 unsigned long flags;
1586 spin_lock_irqsave(&ctx->completion_lock, flags);
1587 __io_cqring_fill_event(ctx, req->user_data, res, cflags);
1589 * If we're the last reference to this request, add to our locked
1592 if (req_ref_put_and_test(req)) {
1593 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
1594 if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL))
1595 io_disarm_next(req);
1597 io_req_task_queue(req->link);
1601 io_dismantle_req(req);
1602 io_put_task(req->task, 1);
1603 list_add(&req->compl.list, &ctx->locked_free_list);
1604 ctx->locked_free_nr++;
1606 if (!percpu_ref_tryget(&ctx->refs))
1609 io_commit_cqring(ctx);
1610 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1613 io_cqring_ev_posted(ctx);
1614 percpu_ref_put(&ctx->refs);
1618 static inline bool io_req_needs_clean(struct io_kiocb *req)
1620 return req->flags & (REQ_F_BUFFER_SELECTED | REQ_F_NEED_CLEANUP |
1621 REQ_F_POLLED | REQ_F_INFLIGHT);
1624 static void io_req_complete_state(struct io_kiocb *req, long res,
1625 unsigned int cflags)
1627 if (io_req_needs_clean(req))
1630 req->compl.cflags = cflags;
1631 req->flags |= REQ_F_COMPLETE_INLINE;
1634 static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags,
1635 long res, unsigned cflags)
1637 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1638 io_req_complete_state(req, res, cflags);
1640 io_req_complete_post(req, res, cflags);
1643 static inline void io_req_complete(struct io_kiocb *req, long res)
1645 __io_req_complete(req, 0, res, 0);
1648 static void io_req_complete_failed(struct io_kiocb *req, long res)
1652 io_req_complete_post(req, res, 0);
1655 static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
1656 struct io_comp_state *cs)
1658 spin_lock_irq(&ctx->completion_lock);
1659 list_splice_init(&ctx->locked_free_list, &cs->free_list);
1660 ctx->locked_free_nr = 0;
1661 spin_unlock_irq(&ctx->completion_lock);
1664 /* Returns true IFF there are requests in the cache */
1665 static bool io_flush_cached_reqs(struct io_ring_ctx *ctx)
1667 struct io_submit_state *state = &ctx->submit_state;
1668 struct io_comp_state *cs = &state->comp;
1672 * If we have more than a batch's worth of requests in our IRQ side
1673 * locked cache, grab the lock and move them over to our submission
1676 if (READ_ONCE(ctx->locked_free_nr) > IO_COMPL_BATCH)
1677 io_flush_cached_locked_reqs(ctx, cs);
1679 nr = state->free_reqs;
1680 while (!list_empty(&cs->free_list)) {
1681 struct io_kiocb *req = list_first_entry(&cs->free_list,
1682 struct io_kiocb, compl.list);
1684 list_del(&req->compl.list);
1685 state->reqs[nr++] = req;
1686 if (nr == ARRAY_SIZE(state->reqs))
1690 state->free_reqs = nr;
1694 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
1696 struct io_submit_state *state = &ctx->submit_state;
1698 BUILD_BUG_ON(IO_REQ_ALLOC_BATCH > ARRAY_SIZE(state->reqs));
1700 if (!state->free_reqs) {
1701 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1704 if (io_flush_cached_reqs(ctx))
1707 ret = kmem_cache_alloc_bulk(req_cachep, gfp, IO_REQ_ALLOC_BATCH,
1711 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1712 * retry single alloc to be on the safe side.
1714 if (unlikely(ret <= 0)) {
1715 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1716 if (!state->reqs[0])
1720 state->free_reqs = ret;
1724 return state->reqs[state->free_reqs];
1727 static inline void io_put_file(struct file *file)
1733 static void io_dismantle_req(struct io_kiocb *req)
1735 unsigned int flags = req->flags;
1737 if (io_req_needs_clean(req))
1739 if (!(flags & REQ_F_FIXED_FILE))
1740 io_put_file(req->file);
1741 if (req->fixed_rsrc_refs)
1742 percpu_ref_put(req->fixed_rsrc_refs);
1743 if (req->async_data)
1744 kfree(req->async_data);
1745 if (req->work.creds) {
1746 put_cred(req->work.creds);
1747 req->work.creds = NULL;
1751 /* must to be called somewhat shortly after putting a request */
1752 static inline void io_put_task(struct task_struct *task, int nr)
1754 struct io_uring_task *tctx = task->io_uring;
1756 percpu_counter_sub(&tctx->inflight, nr);
1757 if (unlikely(atomic_read(&tctx->in_idle)))
1758 wake_up(&tctx->wait);
1759 put_task_struct_many(task, nr);
1762 static void __io_free_req(struct io_kiocb *req)
1764 struct io_ring_ctx *ctx = req->ctx;
1766 io_dismantle_req(req);
1767 io_put_task(req->task, 1);
1769 kmem_cache_free(req_cachep, req);
1770 percpu_ref_put(&ctx->refs);
1773 static inline void io_remove_next_linked(struct io_kiocb *req)
1775 struct io_kiocb *nxt = req->link;
1777 req->link = nxt->link;
1781 static bool io_kill_linked_timeout(struct io_kiocb *req)
1782 __must_hold(&req->ctx->completion_lock)
1784 struct io_kiocb *link = req->link;
1787 * Can happen if a linked timeout fired and link had been like
1788 * req -> link t-out -> link t-out [-> ...]
1790 if (link && (link->flags & REQ_F_LTIMEOUT_ACTIVE)) {
1791 struct io_timeout_data *io = link->async_data;
1793 io_remove_next_linked(req);
1794 link->timeout.head = NULL;
1795 if (hrtimer_try_to_cancel(&io->timer) != -1) {
1796 io_cqring_fill_event(link->ctx, link->user_data,
1798 io_put_req_deferred(link, 1);
1805 static void io_fail_links(struct io_kiocb *req)
1806 __must_hold(&req->ctx->completion_lock)
1808 struct io_kiocb *nxt, *link = req->link;
1815 trace_io_uring_fail_link(req, link);
1816 io_cqring_fill_event(link->ctx, link->user_data, -ECANCELED, 0);
1817 io_put_req_deferred(link, 2);
1822 static bool io_disarm_next(struct io_kiocb *req)
1823 __must_hold(&req->ctx->completion_lock)
1825 bool posted = false;
1827 if (likely(req->flags & REQ_F_LINK_TIMEOUT))
1828 posted = io_kill_linked_timeout(req);
1829 if (unlikely((req->flags & REQ_F_FAIL) &&
1830 !(req->flags & REQ_F_HARDLINK))) {
1831 posted |= (req->link != NULL);
1837 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1839 struct io_kiocb *nxt;
1842 * If LINK is set, we have dependent requests in this chain. If we
1843 * didn't fail this request, queue the first one up, moving any other
1844 * dependencies to the next request. In case of failure, fail the rest
1847 if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL)) {
1848 struct io_ring_ctx *ctx = req->ctx;
1849 unsigned long flags;
1852 spin_lock_irqsave(&ctx->completion_lock, flags);
1853 posted = io_disarm_next(req);
1855 io_commit_cqring(req->ctx);
1856 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1858 io_cqring_ev_posted(ctx);
1865 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1867 if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK))))
1869 return __io_req_find_next(req);
1872 static void ctx_flush_and_put(struct io_ring_ctx *ctx)
1876 if (ctx->submit_state.comp.nr) {
1877 mutex_lock(&ctx->uring_lock);
1878 io_submit_flush_completions(&ctx->submit_state.comp, ctx);
1879 mutex_unlock(&ctx->uring_lock);
1881 percpu_ref_put(&ctx->refs);
1884 static bool __tctx_task_work(struct io_uring_task *tctx)
1886 struct io_ring_ctx *ctx = NULL;
1887 struct io_wq_work_list list;
1888 struct io_wq_work_node *node;
1890 if (wq_list_empty(&tctx->task_list))
1893 spin_lock_irq(&tctx->task_lock);
1894 list = tctx->task_list;
1895 INIT_WQ_LIST(&tctx->task_list);
1896 spin_unlock_irq(&tctx->task_lock);
1900 struct io_wq_work_node *next = node->next;
1901 struct io_kiocb *req;
1903 req = container_of(node, struct io_kiocb, io_task_work.node);
1904 if (req->ctx != ctx) {
1905 ctx_flush_and_put(ctx);
1907 percpu_ref_get(&ctx->refs);
1910 req->task_work.func(&req->task_work);
1914 ctx_flush_and_put(ctx);
1915 return list.first != NULL;
1918 static void tctx_task_work(struct callback_head *cb)
1920 struct io_uring_task *tctx = container_of(cb, struct io_uring_task, task_work);
1922 clear_bit(0, &tctx->task_state);
1924 while (__tctx_task_work(tctx))
1928 static int io_req_task_work_add(struct io_kiocb *req)
1930 struct task_struct *tsk = req->task;
1931 struct io_uring_task *tctx = tsk->io_uring;
1932 enum task_work_notify_mode notify;
1933 struct io_wq_work_node *node, *prev;
1934 unsigned long flags;
1937 if (unlikely(tsk->flags & PF_EXITING))
1940 WARN_ON_ONCE(!tctx);
1942 spin_lock_irqsave(&tctx->task_lock, flags);
1943 wq_list_add_tail(&req->io_task_work.node, &tctx->task_list);
1944 spin_unlock_irqrestore(&tctx->task_lock, flags);
1946 /* task_work already pending, we're done */
1947 if (test_bit(0, &tctx->task_state) ||
1948 test_and_set_bit(0, &tctx->task_state))
1952 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
1953 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
1954 * processing task_work. There's no reliable way to tell if TWA_RESUME
1957 notify = (req->ctx->flags & IORING_SETUP_SQPOLL) ? TWA_NONE : TWA_SIGNAL;
1959 if (!task_work_add(tsk, &tctx->task_work, notify)) {
1960 wake_up_process(tsk);
1965 * Slow path - we failed, find and delete work. if the work is not
1966 * in the list, it got run and we're fine.
1968 spin_lock_irqsave(&tctx->task_lock, flags);
1969 wq_list_for_each(node, prev, &tctx->task_list) {
1970 if (&req->io_task_work.node == node) {
1971 wq_list_del(&tctx->task_list, node, prev);
1976 spin_unlock_irqrestore(&tctx->task_lock, flags);
1977 clear_bit(0, &tctx->task_state);
1981 static bool io_run_task_work_head(struct callback_head **work_head)
1983 struct callback_head *work, *next;
1984 bool executed = false;
1987 work = xchg(work_head, NULL);
2003 static void io_task_work_add_head(struct callback_head **work_head,
2004 struct callback_head *task_work)
2006 struct callback_head *head;
2009 head = READ_ONCE(*work_head);
2010 task_work->next = head;
2011 } while (cmpxchg(work_head, head, task_work) != head);
2014 static void io_req_task_work_add_fallback(struct io_kiocb *req,
2015 task_work_func_t cb)
2017 init_task_work(&req->task_work, cb);
2018 io_task_work_add_head(&req->ctx->exit_task_work, &req->task_work);
2021 static void io_req_task_cancel(struct callback_head *cb)
2023 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2024 struct io_ring_ctx *ctx = req->ctx;
2026 /* ctx is guaranteed to stay alive while we hold uring_lock */
2027 mutex_lock(&ctx->uring_lock);
2028 io_req_complete_failed(req, req->result);
2029 mutex_unlock(&ctx->uring_lock);
2032 static void __io_req_task_submit(struct io_kiocb *req)
2034 struct io_ring_ctx *ctx = req->ctx;
2036 /* ctx stays valid until unlock, even if we drop all ours ctx->refs */
2037 mutex_lock(&ctx->uring_lock);
2038 if (!(current->flags & PF_EXITING) && !current->in_execve)
2039 __io_queue_sqe(req);
2041 io_req_complete_failed(req, -EFAULT);
2042 mutex_unlock(&ctx->uring_lock);
2045 static void io_req_task_submit(struct callback_head *cb)
2047 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2049 __io_req_task_submit(req);
2052 static void io_req_task_queue_fail(struct io_kiocb *req, int ret)
2055 req->task_work.func = io_req_task_cancel;
2057 if (unlikely(io_req_task_work_add(req)))
2058 io_req_task_work_add_fallback(req, io_req_task_cancel);
2061 static void io_req_task_queue(struct io_kiocb *req)
2063 req->task_work.func = io_req_task_submit;
2065 if (unlikely(io_req_task_work_add(req)))
2066 io_req_task_queue_fail(req, -ECANCELED);
2069 static inline void io_queue_next(struct io_kiocb *req)
2071 struct io_kiocb *nxt = io_req_find_next(req);
2074 io_req_task_queue(nxt);
2077 static void io_free_req(struct io_kiocb *req)
2084 struct task_struct *task;
2089 static inline void io_init_req_batch(struct req_batch *rb)
2096 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2097 struct req_batch *rb)
2100 io_put_task(rb->task, rb->task_refs);
2102 percpu_ref_put_many(&ctx->refs, rb->ctx_refs);
2105 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req,
2106 struct io_submit_state *state)
2109 io_dismantle_req(req);
2111 if (req->task != rb->task) {
2113 io_put_task(rb->task, rb->task_refs);
2114 rb->task = req->task;
2120 if (state->free_reqs != ARRAY_SIZE(state->reqs))
2121 state->reqs[state->free_reqs++] = req;
2123 list_add(&req->compl.list, &state->comp.free_list);
2126 static void io_submit_flush_completions(struct io_comp_state *cs,
2127 struct io_ring_ctx *ctx)
2130 struct io_kiocb *req;
2131 struct req_batch rb;
2133 io_init_req_batch(&rb);
2134 spin_lock_irq(&ctx->completion_lock);
2135 for (i = 0; i < nr; i++) {
2137 __io_cqring_fill_event(ctx, req->user_data, req->result,
2140 io_commit_cqring(ctx);
2141 spin_unlock_irq(&ctx->completion_lock);
2143 io_cqring_ev_posted(ctx);
2144 for (i = 0; i < nr; i++) {
2147 /* submission and completion refs */
2148 if (req_ref_sub_and_test(req, 2))
2149 io_req_free_batch(&rb, req, &ctx->submit_state);
2152 io_req_free_batch_finish(ctx, &rb);
2157 * Drop reference to request, return next in chain (if there is one) if this
2158 * was the last reference to this request.
2160 static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2162 struct io_kiocb *nxt = NULL;
2164 if (req_ref_put_and_test(req)) {
2165 nxt = io_req_find_next(req);
2171 static inline void io_put_req(struct io_kiocb *req)
2173 if (req_ref_put_and_test(req))
2177 static void io_put_req_deferred_cb(struct callback_head *cb)
2179 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2184 static void io_free_req_deferred(struct io_kiocb *req)
2186 req->task_work.func = io_put_req_deferred_cb;
2187 if (unlikely(io_req_task_work_add(req)))
2188 io_req_task_work_add_fallback(req, io_put_req_deferred_cb);
2191 static inline void io_put_req_deferred(struct io_kiocb *req, int refs)
2193 if (req_ref_sub_and_test(req, refs))
2194 io_free_req_deferred(req);
2197 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2199 /* See comment at the top of this file */
2201 return __io_cqring_events(ctx);
2204 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2206 struct io_rings *rings = ctx->rings;
2208 /* make sure SQ entry isn't read before tail */
2209 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2212 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2214 unsigned int cflags;
2216 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2217 cflags |= IORING_CQE_F_BUFFER;
2218 req->flags &= ~REQ_F_BUFFER_SELECTED;
2223 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2225 struct io_buffer *kbuf;
2227 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2228 return io_put_kbuf(req, kbuf);
2231 static inline bool io_run_task_work(void)
2234 * Not safe to run on exiting task, and the task_work handling will
2235 * not add work to such a task.
2237 if (unlikely(current->flags & PF_EXITING))
2239 if (current->task_works) {
2240 __set_current_state(TASK_RUNNING);
2249 * Find and free completed poll iocbs
2251 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2252 struct list_head *done)
2254 struct req_batch rb;
2255 struct io_kiocb *req;
2257 /* order with ->result store in io_complete_rw_iopoll() */
2260 io_init_req_batch(&rb);
2261 while (!list_empty(done)) {
2264 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2265 list_del(&req->inflight_entry);
2267 if (READ_ONCE(req->result) == -EAGAIN &&
2268 !(req->flags & REQ_F_DONT_REISSUE)) {
2269 req->iopoll_completed = 0;
2271 io_queue_async_work(req);
2275 if (req->flags & REQ_F_BUFFER_SELECTED)
2276 cflags = io_put_rw_kbuf(req);
2278 __io_cqring_fill_event(ctx, req->user_data, req->result, cflags);
2281 if (req_ref_put_and_test(req))
2282 io_req_free_batch(&rb, req, &ctx->submit_state);
2285 io_commit_cqring(ctx);
2286 io_cqring_ev_posted_iopoll(ctx);
2287 io_req_free_batch_finish(ctx, &rb);
2290 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2293 struct io_kiocb *req, *tmp;
2299 * Only spin for completions if we don't have multiple devices hanging
2300 * off our complete list, and we're under the requested amount.
2302 spin = !ctx->poll_multi_file && *nr_events < min;
2305 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2306 struct kiocb *kiocb = &req->rw.kiocb;
2309 * Move completed and retryable entries to our local lists.
2310 * If we find a request that requires polling, break out
2311 * and complete those lists first, if we have entries there.
2313 if (READ_ONCE(req->iopoll_completed)) {
2314 list_move_tail(&req->inflight_entry, &done);
2317 if (!list_empty(&done))
2320 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2324 /* iopoll may have completed current req */
2325 if (READ_ONCE(req->iopoll_completed))
2326 list_move_tail(&req->inflight_entry, &done);
2333 if (!list_empty(&done))
2334 io_iopoll_complete(ctx, nr_events, &done);
2340 * We can't just wait for polled events to come to us, we have to actively
2341 * find and complete them.
2343 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2345 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2348 mutex_lock(&ctx->uring_lock);
2349 while (!list_empty(&ctx->iopoll_list)) {
2350 unsigned int nr_events = 0;
2352 io_do_iopoll(ctx, &nr_events, 0);
2354 /* let it sleep and repeat later if can't complete a request */
2358 * Ensure we allow local-to-the-cpu processing to take place,
2359 * in this case we need to ensure that we reap all events.
2360 * Also let task_work, etc. to progress by releasing the mutex
2362 if (need_resched()) {
2363 mutex_unlock(&ctx->uring_lock);
2365 mutex_lock(&ctx->uring_lock);
2368 mutex_unlock(&ctx->uring_lock);
2371 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2373 unsigned int nr_events = 0;
2377 * We disallow the app entering submit/complete with polling, but we
2378 * still need to lock the ring to prevent racing with polled issue
2379 * that got punted to a workqueue.
2381 mutex_lock(&ctx->uring_lock);
2383 * Don't enter poll loop if we already have events pending.
2384 * If we do, we can potentially be spinning for commands that
2385 * already triggered a CQE (eg in error).
2387 if (test_bit(0, &ctx->check_cq_overflow))
2388 __io_cqring_overflow_flush(ctx, false);
2389 if (io_cqring_events(ctx))
2393 * If a submit got punted to a workqueue, we can have the
2394 * application entering polling for a command before it gets
2395 * issued. That app will hold the uring_lock for the duration
2396 * of the poll right here, so we need to take a breather every
2397 * now and then to ensure that the issue has a chance to add
2398 * the poll to the issued list. Otherwise we can spin here
2399 * forever, while the workqueue is stuck trying to acquire the
2402 if (list_empty(&ctx->iopoll_list)) {
2403 mutex_unlock(&ctx->uring_lock);
2405 mutex_lock(&ctx->uring_lock);
2407 if (list_empty(&ctx->iopoll_list))
2410 ret = io_do_iopoll(ctx, &nr_events, min);
2411 } while (!ret && nr_events < min && !need_resched());
2413 mutex_unlock(&ctx->uring_lock);
2417 static void kiocb_end_write(struct io_kiocb *req)
2420 * Tell lockdep we inherited freeze protection from submission
2423 if (req->flags & REQ_F_ISREG) {
2424 struct super_block *sb = file_inode(req->file)->i_sb;
2426 __sb_writers_acquired(sb, SB_FREEZE_WRITE);
2432 static bool io_resubmit_prep(struct io_kiocb *req)
2434 struct io_async_rw *rw = req->async_data;
2437 return !io_req_prep_async(req);
2438 /* may have left rw->iter inconsistent on -EIOCBQUEUED */
2439 iov_iter_revert(&rw->iter, req->result - iov_iter_count(&rw->iter));
2443 static bool io_rw_should_reissue(struct io_kiocb *req)
2445 umode_t mode = file_inode(req->file)->i_mode;
2446 struct io_ring_ctx *ctx = req->ctx;
2448 if (!S_ISBLK(mode) && !S_ISREG(mode))
2450 if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
2451 !(ctx->flags & IORING_SETUP_IOPOLL)))
2454 * If ref is dying, we might be running poll reap from the exit work.
2455 * Don't attempt to reissue from that path, just let it fail with
2458 if (percpu_ref_is_dying(&ctx->refs))
2463 static bool io_resubmit_prep(struct io_kiocb *req)
2467 static bool io_rw_should_reissue(struct io_kiocb *req)
2473 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2474 unsigned int issue_flags)
2478 if (req->rw.kiocb.ki_flags & IOCB_WRITE)
2479 kiocb_end_write(req);
2480 if (res != req->result) {
2481 if ((res == -EAGAIN || res == -EOPNOTSUPP) &&
2482 io_rw_should_reissue(req)) {
2483 req->flags |= REQ_F_REISSUE;
2488 if (req->flags & REQ_F_BUFFER_SELECTED)
2489 cflags = io_put_rw_kbuf(req);
2490 __io_req_complete(req, issue_flags, res, cflags);
2493 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2495 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2497 __io_complete_rw(req, res, res2, 0);
2500 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2502 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2504 if (kiocb->ki_flags & IOCB_WRITE)
2505 kiocb_end_write(req);
2506 if (unlikely(res != req->result)) {
2507 if (!(res == -EAGAIN && io_rw_should_reissue(req) &&
2508 io_resubmit_prep(req))) {
2510 req->flags |= REQ_F_DONT_REISSUE;
2514 WRITE_ONCE(req->result, res);
2515 /* order with io_iopoll_complete() checking ->result */
2517 WRITE_ONCE(req->iopoll_completed, 1);
2521 * After the iocb has been issued, it's safe to be found on the poll list.
2522 * Adding the kiocb to the list AFTER submission ensures that we don't
2523 * find it from a io_do_iopoll() thread before the issuer is done
2524 * accessing the kiocb cookie.
2526 static void io_iopoll_req_issued(struct io_kiocb *req)
2528 struct io_ring_ctx *ctx = req->ctx;
2529 const bool in_async = io_wq_current_is_worker();
2531 /* workqueue context doesn't hold uring_lock, grab it now */
2532 if (unlikely(in_async))
2533 mutex_lock(&ctx->uring_lock);
2536 * Track whether we have multiple files in our lists. This will impact
2537 * how we do polling eventually, not spinning if we're on potentially
2538 * different devices.
2540 if (list_empty(&ctx->iopoll_list)) {
2541 ctx->poll_multi_file = false;
2542 } else if (!ctx->poll_multi_file) {
2543 struct io_kiocb *list_req;
2545 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2547 if (list_req->file != req->file)
2548 ctx->poll_multi_file = true;
2552 * For fast devices, IO may have already completed. If it has, add
2553 * it to the front so we find it first.
2555 if (READ_ONCE(req->iopoll_completed))
2556 list_add(&req->inflight_entry, &ctx->iopoll_list);
2558 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2560 if (unlikely(in_async)) {
2562 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
2563 * in sq thread task context or in io worker task context. If
2564 * current task context is sq thread, we don't need to check
2565 * whether should wake up sq thread.
2567 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2568 wq_has_sleeper(&ctx->sq_data->wait))
2569 wake_up(&ctx->sq_data->wait);
2571 mutex_unlock(&ctx->uring_lock);
2575 static inline void io_state_file_put(struct io_submit_state *state)
2577 if (state->file_refs) {
2578 fput_many(state->file, state->file_refs);
2579 state->file_refs = 0;
2584 * Get as many references to a file as we have IOs left in this submission,
2585 * assuming most submissions are for one file, or at least that each file
2586 * has more than one submission.
2588 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2593 if (state->file_refs) {
2594 if (state->fd == fd) {
2598 io_state_file_put(state);
2600 state->file = fget_many(fd, state->ios_left);
2601 if (unlikely(!state->file))
2605 state->file_refs = state->ios_left - 1;
2609 static bool io_bdev_nowait(struct block_device *bdev)
2611 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2615 * If we tracked the file through the SCM inflight mechanism, we could support
2616 * any file. For now, just ensure that anything potentially problematic is done
2619 static bool __io_file_supports_async(struct file *file, int rw)
2621 umode_t mode = file_inode(file)->i_mode;
2623 if (S_ISBLK(mode)) {
2624 if (IS_ENABLED(CONFIG_BLOCK) &&
2625 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2631 if (S_ISREG(mode)) {
2632 if (IS_ENABLED(CONFIG_BLOCK) &&
2633 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2634 file->f_op != &io_uring_fops)
2639 /* any ->read/write should understand O_NONBLOCK */
2640 if (file->f_flags & O_NONBLOCK)
2643 if (!(file->f_mode & FMODE_NOWAIT))
2647 return file->f_op->read_iter != NULL;
2649 return file->f_op->write_iter != NULL;
2652 static bool io_file_supports_async(struct io_kiocb *req, int rw)
2654 if (rw == READ && (req->flags & REQ_F_ASYNC_READ))
2656 else if (rw == WRITE && (req->flags & REQ_F_ASYNC_WRITE))
2659 return __io_file_supports_async(req->file, rw);
2662 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2664 struct io_ring_ctx *ctx = req->ctx;
2665 struct kiocb *kiocb = &req->rw.kiocb;
2666 struct file *file = req->file;
2670 if (!(req->flags & REQ_F_ISREG) && S_ISREG(file_inode(file)->i_mode))
2671 req->flags |= REQ_F_ISREG;
2673 kiocb->ki_pos = READ_ONCE(sqe->off);
2674 if (kiocb->ki_pos == -1 && !(file->f_mode & FMODE_STREAM)) {
2675 req->flags |= REQ_F_CUR_POS;
2676 kiocb->ki_pos = file->f_pos;
2678 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2679 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2680 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2684 /* don't allow async punt for O_NONBLOCK or RWF_NOWAIT */
2685 if ((kiocb->ki_flags & IOCB_NOWAIT) || (file->f_flags & O_NONBLOCK))
2686 req->flags |= REQ_F_NOWAIT;
2688 ioprio = READ_ONCE(sqe->ioprio);
2690 ret = ioprio_check_cap(ioprio);
2694 kiocb->ki_ioprio = ioprio;
2696 kiocb->ki_ioprio = get_current_ioprio();
2698 if (ctx->flags & IORING_SETUP_IOPOLL) {
2699 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2700 !kiocb->ki_filp->f_op->iopoll)
2703 kiocb->ki_flags |= IOCB_HIPRI;
2704 kiocb->ki_complete = io_complete_rw_iopoll;
2705 req->iopoll_completed = 0;
2707 if (kiocb->ki_flags & IOCB_HIPRI)
2709 kiocb->ki_complete = io_complete_rw;
2712 if (req->opcode == IORING_OP_READ_FIXED ||
2713 req->opcode == IORING_OP_WRITE_FIXED) {
2715 io_req_set_rsrc_node(req);
2718 req->rw.addr = READ_ONCE(sqe->addr);
2719 req->rw.len = READ_ONCE(sqe->len);
2720 req->buf_index = READ_ONCE(sqe->buf_index);
2724 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2730 case -ERESTARTNOINTR:
2731 case -ERESTARTNOHAND:
2732 case -ERESTART_RESTARTBLOCK:
2734 * We can't just restart the syscall, since previously
2735 * submitted sqes may already be in progress. Just fail this
2741 kiocb->ki_complete(kiocb, ret, 0);
2745 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2746 unsigned int issue_flags)
2748 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2749 struct io_async_rw *io = req->async_data;
2750 bool check_reissue = kiocb->ki_complete == io_complete_rw;
2752 /* add previously done IO, if any */
2753 if (io && io->bytes_done > 0) {
2755 ret = io->bytes_done;
2757 ret += io->bytes_done;
2760 if (req->flags & REQ_F_CUR_POS)
2761 req->file->f_pos = kiocb->ki_pos;
2762 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2763 __io_complete_rw(req, ret, 0, issue_flags);
2765 io_rw_done(kiocb, ret);
2767 if (check_reissue && req->flags & REQ_F_REISSUE) {
2768 req->flags &= ~REQ_F_REISSUE;
2769 if (io_resubmit_prep(req)) {
2771 io_queue_async_work(req);
2776 if (req->flags & REQ_F_BUFFER_SELECTED)
2777 cflags = io_put_rw_kbuf(req);
2778 __io_req_complete(req, issue_flags, ret, cflags);
2783 static int __io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter,
2784 struct io_mapped_ubuf *imu)
2786 size_t len = req->rw.len;
2787 u64 buf_end, buf_addr = req->rw.addr;
2790 if (unlikely(check_add_overflow(buf_addr, (u64)len, &buf_end)))
2792 /* not inside the mapped region */
2793 if (unlikely(buf_addr < imu->ubuf || buf_end > imu->ubuf_end))
2797 * May not be a start of buffer, set size appropriately
2798 * and advance us to the beginning.
2800 offset = buf_addr - imu->ubuf;
2801 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2805 * Don't use iov_iter_advance() here, as it's really slow for
2806 * using the latter parts of a big fixed buffer - it iterates
2807 * over each segment manually. We can cheat a bit here, because
2810 * 1) it's a BVEC iter, we set it up
2811 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2812 * first and last bvec
2814 * So just find our index, and adjust the iterator afterwards.
2815 * If the offset is within the first bvec (or the whole first
2816 * bvec, just use iov_iter_advance(). This makes it easier
2817 * since we can just skip the first segment, which may not
2818 * be PAGE_SIZE aligned.
2820 const struct bio_vec *bvec = imu->bvec;
2822 if (offset <= bvec->bv_len) {
2823 iov_iter_advance(iter, offset);
2825 unsigned long seg_skip;
2827 /* skip first vec */
2828 offset -= bvec->bv_len;
2829 seg_skip = 1 + (offset >> PAGE_SHIFT);
2831 iter->bvec = bvec + seg_skip;
2832 iter->nr_segs -= seg_skip;
2833 iter->count -= bvec->bv_len + offset;
2834 iter->iov_offset = offset & ~PAGE_MASK;
2841 static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
2843 struct io_ring_ctx *ctx = req->ctx;
2844 struct io_mapped_ubuf *imu = req->imu;
2845 u16 index, buf_index = req->buf_index;
2848 if (unlikely(buf_index >= ctx->nr_user_bufs))
2850 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2851 imu = READ_ONCE(ctx->user_bufs[index]);
2854 return __io_import_fixed(req, rw, iter, imu);
2857 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2860 mutex_unlock(&ctx->uring_lock);
2863 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2866 * "Normal" inline submissions always hold the uring_lock, since we
2867 * grab it from the system call. Same is true for the SQPOLL offload.
2868 * The only exception is when we've detached the request and issue it
2869 * from an async worker thread, grab the lock for that case.
2872 mutex_lock(&ctx->uring_lock);
2875 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2876 int bgid, struct io_buffer *kbuf,
2879 struct io_buffer *head;
2881 if (req->flags & REQ_F_BUFFER_SELECTED)
2884 io_ring_submit_lock(req->ctx, needs_lock);
2886 lockdep_assert_held(&req->ctx->uring_lock);
2888 head = xa_load(&req->ctx->io_buffers, bgid);
2890 if (!list_empty(&head->list)) {
2891 kbuf = list_last_entry(&head->list, struct io_buffer,
2893 list_del(&kbuf->list);
2896 xa_erase(&req->ctx->io_buffers, bgid);
2898 if (*len > kbuf->len)
2901 kbuf = ERR_PTR(-ENOBUFS);
2904 io_ring_submit_unlock(req->ctx, needs_lock);
2909 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2912 struct io_buffer *kbuf;
2915 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2916 bgid = req->buf_index;
2917 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2920 req->rw.addr = (u64) (unsigned long) kbuf;
2921 req->flags |= REQ_F_BUFFER_SELECTED;
2922 return u64_to_user_ptr(kbuf->addr);
2925 #ifdef CONFIG_COMPAT
2926 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2929 struct compat_iovec __user *uiov;
2930 compat_ssize_t clen;
2934 uiov = u64_to_user_ptr(req->rw.addr);
2935 if (!access_ok(uiov, sizeof(*uiov)))
2937 if (__get_user(clen, &uiov->iov_len))
2943 buf = io_rw_buffer_select(req, &len, needs_lock);
2945 return PTR_ERR(buf);
2946 iov[0].iov_base = buf;
2947 iov[0].iov_len = (compat_size_t) len;
2952 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2955 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2959 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2962 len = iov[0].iov_len;
2965 buf = io_rw_buffer_select(req, &len, needs_lock);
2967 return PTR_ERR(buf);
2968 iov[0].iov_base = buf;
2969 iov[0].iov_len = len;
2973 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2976 if (req->flags & REQ_F_BUFFER_SELECTED) {
2977 struct io_buffer *kbuf;
2979 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2980 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
2981 iov[0].iov_len = kbuf->len;
2984 if (req->rw.len != 1)
2987 #ifdef CONFIG_COMPAT
2988 if (req->ctx->compat)
2989 return io_compat_import(req, iov, needs_lock);
2992 return __io_iov_buffer_select(req, iov, needs_lock);
2995 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
2996 struct iov_iter *iter, bool needs_lock)
2998 void __user *buf = u64_to_user_ptr(req->rw.addr);
2999 size_t sqe_len = req->rw.len;
3000 u8 opcode = req->opcode;
3003 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3005 return io_import_fixed(req, rw, iter);
3008 /* buffer index only valid with fixed read/write, or buffer select */
3009 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3012 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3013 if (req->flags & REQ_F_BUFFER_SELECT) {
3014 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3016 return PTR_ERR(buf);
3017 req->rw.len = sqe_len;
3020 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3025 if (req->flags & REQ_F_BUFFER_SELECT) {
3026 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3028 iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
3033 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3037 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3039 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3043 * For files that don't have ->read_iter() and ->write_iter(), handle them
3044 * by looping over ->read() or ->write() manually.
3046 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3048 struct kiocb *kiocb = &req->rw.kiocb;
3049 struct file *file = req->file;
3053 * Don't support polled IO through this interface, and we can't
3054 * support non-blocking either. For the latter, this just causes
3055 * the kiocb to be handled from an async context.
3057 if (kiocb->ki_flags & IOCB_HIPRI)
3059 if (kiocb->ki_flags & IOCB_NOWAIT)
3062 while (iov_iter_count(iter)) {
3066 if (!iov_iter_is_bvec(iter)) {
3067 iovec = iov_iter_iovec(iter);
3069 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3070 iovec.iov_len = req->rw.len;
3074 nr = file->f_op->read(file, iovec.iov_base,
3075 iovec.iov_len, io_kiocb_ppos(kiocb));
3077 nr = file->f_op->write(file, iovec.iov_base,
3078 iovec.iov_len, io_kiocb_ppos(kiocb));
3087 if (nr != iovec.iov_len)
3091 iov_iter_advance(iter, nr);
3097 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3098 const struct iovec *fast_iov, struct iov_iter *iter)
3100 struct io_async_rw *rw = req->async_data;
3102 memcpy(&rw->iter, iter, sizeof(*iter));
3103 rw->free_iovec = iovec;
3105 /* can only be fixed buffers, no need to do anything */
3106 if (iov_iter_is_bvec(iter))
3109 unsigned iov_off = 0;
3111 rw->iter.iov = rw->fast_iov;
3112 if (iter->iov != fast_iov) {
3113 iov_off = iter->iov - fast_iov;
3114 rw->iter.iov += iov_off;
3116 if (rw->fast_iov != fast_iov)
3117 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3118 sizeof(struct iovec) * iter->nr_segs);
3120 req->flags |= REQ_F_NEED_CLEANUP;
3124 static inline int io_alloc_async_data(struct io_kiocb *req)
3126 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3127 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3128 return req->async_data == NULL;
3131 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3132 const struct iovec *fast_iov,
3133 struct iov_iter *iter, bool force)
3135 if (!force && !io_op_defs[req->opcode].needs_async_setup)
3137 if (!req->async_data) {
3138 if (io_alloc_async_data(req)) {
3143 io_req_map_rw(req, iovec, fast_iov, iter);
3148 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3150 struct io_async_rw *iorw = req->async_data;
3151 struct iovec *iov = iorw->fast_iov;
3154 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3155 if (unlikely(ret < 0))
3158 iorw->bytes_done = 0;
3159 iorw->free_iovec = iov;
3161 req->flags |= REQ_F_NEED_CLEANUP;
3165 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3167 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3169 return io_prep_rw(req, sqe);
3173 * This is our waitqueue callback handler, registered through lock_page_async()
3174 * when we initially tried to do the IO with the iocb armed our waitqueue.
3175 * This gets called when the page is unlocked, and we generally expect that to
3176 * happen when the page IO is completed and the page is now uptodate. This will
3177 * queue a task_work based retry of the operation, attempting to copy the data
3178 * again. If the latter fails because the page was NOT uptodate, then we will
3179 * do a thread based blocking retry of the operation. That's the unexpected
3182 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3183 int sync, void *arg)
3185 struct wait_page_queue *wpq;
3186 struct io_kiocb *req = wait->private;
3187 struct wait_page_key *key = arg;
3189 wpq = container_of(wait, struct wait_page_queue, wait);
3191 if (!wake_page_match(wpq, key))
3194 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3195 list_del_init(&wait->entry);
3197 /* submit ref gets dropped, acquire a new one */
3199 io_req_task_queue(req);
3204 * This controls whether a given IO request should be armed for async page
3205 * based retry. If we return false here, the request is handed to the async
3206 * worker threads for retry. If we're doing buffered reads on a regular file,
3207 * we prepare a private wait_page_queue entry and retry the operation. This
3208 * will either succeed because the page is now uptodate and unlocked, or it
3209 * will register a callback when the page is unlocked at IO completion. Through
3210 * that callback, io_uring uses task_work to setup a retry of the operation.
3211 * That retry will attempt the buffered read again. The retry will generally
3212 * succeed, or in rare cases where it fails, we then fall back to using the
3213 * async worker threads for a blocking retry.
3215 static bool io_rw_should_retry(struct io_kiocb *req)
3217 struct io_async_rw *rw = req->async_data;
3218 struct wait_page_queue *wait = &rw->wpq;
3219 struct kiocb *kiocb = &req->rw.kiocb;
3221 /* never retry for NOWAIT, we just complete with -EAGAIN */
3222 if (req->flags & REQ_F_NOWAIT)
3225 /* Only for buffered IO */
3226 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3230 * just use poll if we can, and don't attempt if the fs doesn't
3231 * support callback based unlocks
3233 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3236 wait->wait.func = io_async_buf_func;
3237 wait->wait.private = req;
3238 wait->wait.flags = 0;
3239 INIT_LIST_HEAD(&wait->wait.entry);
3240 kiocb->ki_flags |= IOCB_WAITQ;
3241 kiocb->ki_flags &= ~IOCB_NOWAIT;
3242 kiocb->ki_waitq = wait;
3246 static inline int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3248 if (req->file->f_op->read_iter)
3249 return call_read_iter(req->file, &req->rw.kiocb, iter);
3250 else if (req->file->f_op->read)
3251 return loop_rw_iter(READ, req, iter);
3256 static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3258 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3259 struct kiocb *kiocb = &req->rw.kiocb;
3260 struct iov_iter __iter, *iter = &__iter;
3261 struct io_async_rw *rw = req->async_data;
3262 ssize_t io_size, ret, ret2;
3263 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3269 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3273 io_size = iov_iter_count(iter);
3274 req->result = io_size;
3276 /* Ensure we clear previously set non-block flag */
3277 if (!force_nonblock)
3278 kiocb->ki_flags &= ~IOCB_NOWAIT;
3280 kiocb->ki_flags |= IOCB_NOWAIT;
3282 /* If the file doesn't support async, just async punt */
3283 if (force_nonblock && !io_file_supports_async(req, READ)) {
3284 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3285 return ret ?: -EAGAIN;
3288 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size);
3289 if (unlikely(ret)) {
3294 ret = io_iter_do_read(req, iter);
3296 if (ret == -EAGAIN || (req->flags & REQ_F_REISSUE)) {
3297 req->flags &= ~REQ_F_REISSUE;
3298 /* IOPOLL retry should happen for io-wq threads */
3299 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3301 /* no retry on NONBLOCK nor RWF_NOWAIT */
3302 if (req->flags & REQ_F_NOWAIT)
3304 /* some cases will consume bytes even on error returns */
3305 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3307 } else if (ret == -EIOCBQUEUED) {
3309 } else if (ret <= 0 || ret == io_size || !force_nonblock ||
3310 (req->flags & REQ_F_NOWAIT) || !(req->flags & REQ_F_ISREG)) {
3311 /* read all, failed, already did sync or don't want to retry */
3315 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3320 rw = req->async_data;
3321 /* now use our persistent iterator, if we aren't already */
3326 rw->bytes_done += ret;
3327 /* if we can retry, do so with the callbacks armed */
3328 if (!io_rw_should_retry(req)) {
3329 kiocb->ki_flags &= ~IOCB_WAITQ;
3334 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3335 * we get -EIOCBQUEUED, then we'll get a notification when the
3336 * desired page gets unlocked. We can also get a partial read
3337 * here, and if we do, then just retry at the new offset.
3339 ret = io_iter_do_read(req, iter);
3340 if (ret == -EIOCBQUEUED)
3342 /* we got some bytes, but not all. retry. */
3343 kiocb->ki_flags &= ~IOCB_WAITQ;
3344 } while (ret > 0 && ret < io_size);
3346 kiocb_done(kiocb, ret, issue_flags);
3348 /* it's faster to check here then delegate to kfree */
3354 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3356 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3358 return io_prep_rw(req, sqe);
3361 static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3363 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3364 struct kiocb *kiocb = &req->rw.kiocb;
3365 struct iov_iter __iter, *iter = &__iter;
3366 struct io_async_rw *rw = req->async_data;
3367 ssize_t ret, ret2, io_size;
3368 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3374 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3378 io_size = iov_iter_count(iter);
3379 req->result = io_size;
3381 /* Ensure we clear previously set non-block flag */
3382 if (!force_nonblock)
3383 kiocb->ki_flags &= ~IOCB_NOWAIT;
3385 kiocb->ki_flags |= IOCB_NOWAIT;
3387 /* If the file doesn't support async, just async punt */
3388 if (force_nonblock && !io_file_supports_async(req, WRITE))
3391 /* file path doesn't support NOWAIT for non-direct_IO */
3392 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3393 (req->flags & REQ_F_ISREG))
3396 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size);
3401 * Open-code file_start_write here to grab freeze protection,
3402 * which will be released by another thread in
3403 * io_complete_rw(). Fool lockdep by telling it the lock got
3404 * released so that it doesn't complain about the held lock when
3405 * we return to userspace.
3407 if (req->flags & REQ_F_ISREG) {
3408 sb_start_write(file_inode(req->file)->i_sb);
3409 __sb_writers_release(file_inode(req->file)->i_sb,
3412 kiocb->ki_flags |= IOCB_WRITE;
3414 if (req->file->f_op->write_iter)
3415 ret2 = call_write_iter(req->file, kiocb, iter);
3416 else if (req->file->f_op->write)
3417 ret2 = loop_rw_iter(WRITE, req, iter);
3421 if (req->flags & REQ_F_REISSUE) {
3422 req->flags &= ~REQ_F_REISSUE;
3427 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3428 * retry them without IOCB_NOWAIT.
3430 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3432 /* no retry on NONBLOCK nor RWF_NOWAIT */
3433 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3435 if (!force_nonblock || ret2 != -EAGAIN) {
3436 /* IOPOLL retry should happen for io-wq threads */
3437 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3440 kiocb_done(kiocb, ret2, issue_flags);
3443 /* some cases will consume bytes even on error returns */
3444 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3445 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3446 return ret ?: -EAGAIN;
3449 /* it's reportedly faster than delegating the null check to kfree() */
3455 static int io_renameat_prep(struct io_kiocb *req,
3456 const struct io_uring_sqe *sqe)
3458 struct io_rename *ren = &req->rename;
3459 const char __user *oldf, *newf;
3461 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3464 ren->old_dfd = READ_ONCE(sqe->fd);
3465 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3466 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3467 ren->new_dfd = READ_ONCE(sqe->len);
3468 ren->flags = READ_ONCE(sqe->rename_flags);
3470 ren->oldpath = getname(oldf);
3471 if (IS_ERR(ren->oldpath))
3472 return PTR_ERR(ren->oldpath);
3474 ren->newpath = getname(newf);
3475 if (IS_ERR(ren->newpath)) {
3476 putname(ren->oldpath);
3477 return PTR_ERR(ren->newpath);
3480 req->flags |= REQ_F_NEED_CLEANUP;
3484 static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3486 struct io_rename *ren = &req->rename;
3489 if (issue_flags & IO_URING_F_NONBLOCK)
3492 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3493 ren->newpath, ren->flags);
3495 req->flags &= ~REQ_F_NEED_CLEANUP;
3498 io_req_complete(req, ret);
3502 static int io_unlinkat_prep(struct io_kiocb *req,
3503 const struct io_uring_sqe *sqe)
3505 struct io_unlink *un = &req->unlink;
3506 const char __user *fname;
3508 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3511 un->dfd = READ_ONCE(sqe->fd);
3513 un->flags = READ_ONCE(sqe->unlink_flags);
3514 if (un->flags & ~AT_REMOVEDIR)
3517 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3518 un->filename = getname(fname);
3519 if (IS_ERR(un->filename))
3520 return PTR_ERR(un->filename);
3522 req->flags |= REQ_F_NEED_CLEANUP;
3526 static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3528 struct io_unlink *un = &req->unlink;
3531 if (issue_flags & IO_URING_F_NONBLOCK)
3534 if (un->flags & AT_REMOVEDIR)
3535 ret = do_rmdir(un->dfd, un->filename);
3537 ret = do_unlinkat(un->dfd, un->filename);
3539 req->flags &= ~REQ_F_NEED_CLEANUP;
3542 io_req_complete(req, ret);
3546 static int io_shutdown_prep(struct io_kiocb *req,
3547 const struct io_uring_sqe *sqe)
3549 #if defined(CONFIG_NET)
3550 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3552 if (sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3556 req->shutdown.how = READ_ONCE(sqe->len);
3563 static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3565 #if defined(CONFIG_NET)
3566 struct socket *sock;
3569 if (issue_flags & IO_URING_F_NONBLOCK)
3572 sock = sock_from_file(req->file);
3573 if (unlikely(!sock))
3576 ret = __sys_shutdown_sock(sock, req->shutdown.how);
3579 io_req_complete(req, ret);
3586 static int __io_splice_prep(struct io_kiocb *req,
3587 const struct io_uring_sqe *sqe)
3589 struct io_splice* sp = &req->splice;
3590 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3592 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3596 sp->len = READ_ONCE(sqe->len);
3597 sp->flags = READ_ONCE(sqe->splice_flags);
3599 if (unlikely(sp->flags & ~valid_flags))
3602 sp->file_in = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in),
3603 (sp->flags & SPLICE_F_FD_IN_FIXED));
3606 req->flags |= REQ_F_NEED_CLEANUP;
3610 static int io_tee_prep(struct io_kiocb *req,
3611 const struct io_uring_sqe *sqe)
3613 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3615 return __io_splice_prep(req, sqe);
3618 static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
3620 struct io_splice *sp = &req->splice;
3621 struct file *in = sp->file_in;
3622 struct file *out = sp->file_out;
3623 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3626 if (issue_flags & IO_URING_F_NONBLOCK)
3629 ret = do_tee(in, out, sp->len, flags);
3631 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3633 req->flags &= ~REQ_F_NEED_CLEANUP;
3637 io_req_complete(req, ret);
3641 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3643 struct io_splice* sp = &req->splice;
3645 sp->off_in = READ_ONCE(sqe->splice_off_in);
3646 sp->off_out = READ_ONCE(sqe->off);
3647 return __io_splice_prep(req, sqe);
3650 static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
3652 struct io_splice *sp = &req->splice;
3653 struct file *in = sp->file_in;
3654 struct file *out = sp->file_out;
3655 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3656 loff_t *poff_in, *poff_out;
3659 if (issue_flags & IO_URING_F_NONBLOCK)
3662 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3663 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3666 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3668 if (!(sp->flags & SPLICE_F_FD_IN_FIXED))
3670 req->flags &= ~REQ_F_NEED_CLEANUP;
3674 io_req_complete(req, ret);
3679 * IORING_OP_NOP just posts a completion event, nothing else.
3681 static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
3683 struct io_ring_ctx *ctx = req->ctx;
3685 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3688 __io_req_complete(req, issue_flags, 0, 0);
3692 static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3694 struct io_ring_ctx *ctx = req->ctx;
3699 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3701 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3704 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3705 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3708 req->sync.off = READ_ONCE(sqe->off);
3709 req->sync.len = READ_ONCE(sqe->len);
3713 static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
3715 loff_t end = req->sync.off + req->sync.len;
3718 /* fsync always requires a blocking context */
3719 if (issue_flags & IO_URING_F_NONBLOCK)
3722 ret = vfs_fsync_range(req->file, req->sync.off,
3723 end > 0 ? end : LLONG_MAX,
3724 req->sync.flags & IORING_FSYNC_DATASYNC);
3727 io_req_complete(req, ret);
3731 static int io_fallocate_prep(struct io_kiocb *req,
3732 const struct io_uring_sqe *sqe)
3734 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3736 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3739 req->sync.off = READ_ONCE(sqe->off);
3740 req->sync.len = READ_ONCE(sqe->addr);
3741 req->sync.mode = READ_ONCE(sqe->len);
3745 static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
3749 /* fallocate always requiring blocking context */
3750 if (issue_flags & IO_URING_F_NONBLOCK)
3752 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3756 io_req_complete(req, ret);
3760 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3762 const char __user *fname;
3765 if (unlikely(sqe->ioprio || sqe->buf_index))
3767 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3770 /* open.how should be already initialised */
3771 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3772 req->open.how.flags |= O_LARGEFILE;
3774 req->open.dfd = READ_ONCE(sqe->fd);
3775 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3776 req->open.filename = getname(fname);
3777 if (IS_ERR(req->open.filename)) {
3778 ret = PTR_ERR(req->open.filename);
3779 req->open.filename = NULL;
3782 req->open.nofile = rlimit(RLIMIT_NOFILE);
3783 req->flags |= REQ_F_NEED_CLEANUP;
3787 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3791 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3793 mode = READ_ONCE(sqe->len);
3794 flags = READ_ONCE(sqe->open_flags);
3795 req->open.how = build_open_how(flags, mode);
3796 return __io_openat_prep(req, sqe);
3799 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3801 struct open_how __user *how;
3805 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3807 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3808 len = READ_ONCE(sqe->len);
3809 if (len < OPEN_HOW_SIZE_VER0)
3812 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3817 return __io_openat_prep(req, sqe);
3820 static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
3822 struct open_flags op;
3825 bool resolve_nonblock;
3828 ret = build_open_flags(&req->open.how, &op);
3831 nonblock_set = op.open_flag & O_NONBLOCK;
3832 resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
3833 if (issue_flags & IO_URING_F_NONBLOCK) {
3835 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
3836 * it'll always -EAGAIN
3838 if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
3840 op.lookup_flags |= LOOKUP_CACHED;
3841 op.open_flag |= O_NONBLOCK;
3844 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3848 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3849 /* only retry if RESOLVE_CACHED wasn't already set by application */
3850 if ((!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)) &&
3851 file == ERR_PTR(-EAGAIN)) {
3853 * We could hang on to this 'fd', but seems like marginal
3854 * gain for something that is now known to be a slower path.
3855 * So just put it, and we'll get a new one when we retry.
3863 ret = PTR_ERR(file);
3865 if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
3866 file->f_flags &= ~O_NONBLOCK;
3867 fsnotify_open(file);
3868 fd_install(ret, file);
3871 putname(req->open.filename);
3872 req->flags &= ~REQ_F_NEED_CLEANUP;
3875 __io_req_complete(req, issue_flags, ret, 0);
3879 static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
3881 return io_openat2(req, issue_flags);
3884 static int io_remove_buffers_prep(struct io_kiocb *req,
3885 const struct io_uring_sqe *sqe)
3887 struct io_provide_buf *p = &req->pbuf;
3890 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3893 tmp = READ_ONCE(sqe->fd);
3894 if (!tmp || tmp > USHRT_MAX)
3897 memset(p, 0, sizeof(*p));
3899 p->bgid = READ_ONCE(sqe->buf_group);
3903 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3904 int bgid, unsigned nbufs)
3908 /* shouldn't happen */
3912 /* the head kbuf is the list itself */
3913 while (!list_empty(&buf->list)) {
3914 struct io_buffer *nxt;
3916 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3917 list_del(&nxt->list);
3924 xa_erase(&ctx->io_buffers, bgid);
3929 static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
3931 struct io_provide_buf *p = &req->pbuf;
3932 struct io_ring_ctx *ctx = req->ctx;
3933 struct io_buffer *head;
3935 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3937 io_ring_submit_lock(ctx, !force_nonblock);
3939 lockdep_assert_held(&ctx->uring_lock);
3942 head = xa_load(&ctx->io_buffers, p->bgid);
3944 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3948 /* complete before unlock, IOPOLL may need the lock */
3949 __io_req_complete(req, issue_flags, ret, 0);
3950 io_ring_submit_unlock(ctx, !force_nonblock);
3954 static int io_provide_buffers_prep(struct io_kiocb *req,
3955 const struct io_uring_sqe *sqe)
3957 unsigned long size, tmp_check;
3958 struct io_provide_buf *p = &req->pbuf;
3961 if (sqe->ioprio || sqe->rw_flags)
3964 tmp = READ_ONCE(sqe->fd);
3965 if (!tmp || tmp > USHRT_MAX)
3968 p->addr = READ_ONCE(sqe->addr);
3969 p->len = READ_ONCE(sqe->len);
3971 if (check_mul_overflow((unsigned long)p->len, (unsigned long)p->nbufs,
3974 if (check_add_overflow((unsigned long)p->addr, size, &tmp_check))
3977 size = (unsigned long)p->len * p->nbufs;
3978 if (!access_ok(u64_to_user_ptr(p->addr), size))
3981 p->bgid = READ_ONCE(sqe->buf_group);
3982 tmp = READ_ONCE(sqe->off);
3983 if (tmp > USHRT_MAX)
3989 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3991 struct io_buffer *buf;
3992 u64 addr = pbuf->addr;
3993 int i, bid = pbuf->bid;
3995 for (i = 0; i < pbuf->nbufs; i++) {
3996 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
4001 buf->len = min_t(__u32, pbuf->len, MAX_RW_COUNT);
4006 INIT_LIST_HEAD(&buf->list);
4009 list_add_tail(&buf->list, &(*head)->list);
4013 return i ? i : -ENOMEM;
4016 static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
4018 struct io_provide_buf *p = &req->pbuf;
4019 struct io_ring_ctx *ctx = req->ctx;
4020 struct io_buffer *head, *list;
4022 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4024 io_ring_submit_lock(ctx, !force_nonblock);
4026 lockdep_assert_held(&ctx->uring_lock);
4028 list = head = xa_load(&ctx->io_buffers, p->bgid);
4030 ret = io_add_buffers(p, &head);
4031 if (ret >= 0 && !list) {
4032 ret = xa_insert(&ctx->io_buffers, p->bgid, head, GFP_KERNEL);
4034 __io_remove_buffers(ctx, head, p->bgid, -1U);
4038 /* complete before unlock, IOPOLL may need the lock */
4039 __io_req_complete(req, issue_flags, ret, 0);
4040 io_ring_submit_unlock(ctx, !force_nonblock);
4044 static int io_epoll_ctl_prep(struct io_kiocb *req,
4045 const struct io_uring_sqe *sqe)
4047 #if defined(CONFIG_EPOLL)
4048 if (sqe->ioprio || sqe->buf_index)
4050 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4053 req->epoll.epfd = READ_ONCE(sqe->fd);
4054 req->epoll.op = READ_ONCE(sqe->len);
4055 req->epoll.fd = READ_ONCE(sqe->off);
4057 if (ep_op_has_event(req->epoll.op)) {
4058 struct epoll_event __user *ev;
4060 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4061 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4071 static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4073 #if defined(CONFIG_EPOLL)
4074 struct io_epoll *ie = &req->epoll;
4076 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4078 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4079 if (force_nonblock && ret == -EAGAIN)
4084 __io_req_complete(req, issue_flags, ret, 0);
4091 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4093 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4094 if (sqe->ioprio || sqe->buf_index || sqe->off)
4096 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4099 req->madvise.addr = READ_ONCE(sqe->addr);
4100 req->madvise.len = READ_ONCE(sqe->len);
4101 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4108 static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4110 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4111 struct io_madvise *ma = &req->madvise;
4114 if (issue_flags & IO_URING_F_NONBLOCK)
4117 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4120 io_req_complete(req, ret);
4127 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4129 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4131 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4134 req->fadvise.offset = READ_ONCE(sqe->off);
4135 req->fadvise.len = READ_ONCE(sqe->len);
4136 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4140 static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4142 struct io_fadvise *fa = &req->fadvise;
4145 if (issue_flags & IO_URING_F_NONBLOCK) {
4146 switch (fa->advice) {
4147 case POSIX_FADV_NORMAL:
4148 case POSIX_FADV_RANDOM:
4149 case POSIX_FADV_SEQUENTIAL:
4156 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4159 __io_req_complete(req, issue_flags, ret, 0);
4163 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4165 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4167 if (sqe->ioprio || sqe->buf_index)
4169 if (req->flags & REQ_F_FIXED_FILE)
4172 req->statx.dfd = READ_ONCE(sqe->fd);
4173 req->statx.mask = READ_ONCE(sqe->len);
4174 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4175 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4176 req->statx.flags = READ_ONCE(sqe->statx_flags);
4181 static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4183 struct io_statx *ctx = &req->statx;
4186 if (issue_flags & IO_URING_F_NONBLOCK)
4189 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4194 io_req_complete(req, ret);
4198 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4200 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4202 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4203 sqe->rw_flags || sqe->buf_index)
4205 if (req->flags & REQ_F_FIXED_FILE)
4208 req->close.fd = READ_ONCE(sqe->fd);
4212 static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4214 struct files_struct *files = current->files;
4215 struct io_close *close = &req->close;
4216 struct fdtable *fdt;
4217 struct file *file = NULL;
4220 spin_lock(&files->file_lock);
4221 fdt = files_fdtable(files);
4222 if (close->fd >= fdt->max_fds) {
4223 spin_unlock(&files->file_lock);
4226 file = fdt->fd[close->fd];
4227 if (!file || file->f_op == &io_uring_fops) {
4228 spin_unlock(&files->file_lock);
4233 /* if the file has a flush method, be safe and punt to async */
4234 if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4235 spin_unlock(&files->file_lock);
4239 ret = __close_fd_get_file(close->fd, &file);
4240 spin_unlock(&files->file_lock);
4247 /* No ->flush() or already async, safely close from here */
4248 ret = filp_close(file, current->files);
4254 __io_req_complete(req, issue_flags, ret, 0);
4258 static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4260 struct io_ring_ctx *ctx = req->ctx;
4262 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4264 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4267 req->sync.off = READ_ONCE(sqe->off);
4268 req->sync.len = READ_ONCE(sqe->len);
4269 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4273 static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4277 /* sync_file_range always requires a blocking context */
4278 if (issue_flags & IO_URING_F_NONBLOCK)
4281 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4285 io_req_complete(req, ret);
4289 #if defined(CONFIG_NET)
4290 static int io_setup_async_msg(struct io_kiocb *req,
4291 struct io_async_msghdr *kmsg)
4293 struct io_async_msghdr *async_msg = req->async_data;
4297 if (io_alloc_async_data(req)) {
4298 kfree(kmsg->free_iov);
4301 async_msg = req->async_data;
4302 req->flags |= REQ_F_NEED_CLEANUP;
4303 memcpy(async_msg, kmsg, sizeof(*kmsg));
4304 async_msg->msg.msg_name = &async_msg->addr;
4305 /* if were using fast_iov, set it to the new one */
4306 if (!async_msg->free_iov)
4307 async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4312 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4313 struct io_async_msghdr *iomsg)
4315 iomsg->msg.msg_name = &iomsg->addr;
4316 iomsg->free_iov = iomsg->fast_iov;
4317 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4318 req->sr_msg.msg_flags, &iomsg->free_iov);
4321 static int io_sendmsg_prep_async(struct io_kiocb *req)
4325 ret = io_sendmsg_copy_hdr(req, req->async_data);
4327 req->flags |= REQ_F_NEED_CLEANUP;
4331 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4333 struct io_sr_msg *sr = &req->sr_msg;
4335 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4338 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4339 sr->len = READ_ONCE(sqe->len);
4340 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4341 if (sr->msg_flags & MSG_DONTWAIT)
4342 req->flags |= REQ_F_NOWAIT;
4344 #ifdef CONFIG_COMPAT
4345 if (req->ctx->compat)
4346 sr->msg_flags |= MSG_CMSG_COMPAT;
4351 static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4353 struct io_async_msghdr iomsg, *kmsg;
4354 struct socket *sock;
4359 sock = sock_from_file(req->file);
4360 if (unlikely(!sock))
4363 kmsg = req->async_data;
4365 ret = io_sendmsg_copy_hdr(req, &iomsg);
4371 flags = req->sr_msg.msg_flags;
4372 if (issue_flags & IO_URING_F_NONBLOCK)
4373 flags |= MSG_DONTWAIT;
4374 if (flags & MSG_WAITALL)
4375 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4377 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4378 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4379 return io_setup_async_msg(req, kmsg);
4380 if (ret == -ERESTARTSYS)
4383 /* fast path, check for non-NULL to avoid function call */
4385 kfree(kmsg->free_iov);
4386 req->flags &= ~REQ_F_NEED_CLEANUP;
4389 __io_req_complete(req, issue_flags, ret, 0);
4393 static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4395 struct io_sr_msg *sr = &req->sr_msg;
4398 struct socket *sock;
4403 sock = sock_from_file(req->file);
4404 if (unlikely(!sock))
4407 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4411 msg.msg_name = NULL;
4412 msg.msg_control = NULL;
4413 msg.msg_controllen = 0;
4414 msg.msg_namelen = 0;
4416 flags = req->sr_msg.msg_flags;
4417 if (issue_flags & IO_URING_F_NONBLOCK)
4418 flags |= MSG_DONTWAIT;
4419 if (flags & MSG_WAITALL)
4420 min_ret = iov_iter_count(&msg.msg_iter);
4422 msg.msg_flags = flags;
4423 ret = sock_sendmsg(sock, &msg);
4424 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4426 if (ret == -ERESTARTSYS)
4431 __io_req_complete(req, issue_flags, ret, 0);
4435 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4436 struct io_async_msghdr *iomsg)
4438 struct io_sr_msg *sr = &req->sr_msg;
4439 struct iovec __user *uiov;
4443 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4444 &iomsg->uaddr, &uiov, &iov_len);
4448 if (req->flags & REQ_F_BUFFER_SELECT) {
4451 if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4453 sr->len = iomsg->fast_iov[0].iov_len;
4454 iomsg->free_iov = NULL;
4456 iomsg->free_iov = iomsg->fast_iov;
4457 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4458 &iomsg->free_iov, &iomsg->msg.msg_iter,
4467 #ifdef CONFIG_COMPAT
4468 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4469 struct io_async_msghdr *iomsg)
4471 struct io_sr_msg *sr = &req->sr_msg;
4472 struct compat_iovec __user *uiov;
4477 ret = __get_compat_msghdr(&iomsg->msg, sr->umsg_compat, &iomsg->uaddr,
4482 uiov = compat_ptr(ptr);
4483 if (req->flags & REQ_F_BUFFER_SELECT) {
4484 compat_ssize_t clen;
4488 if (!access_ok(uiov, sizeof(*uiov)))
4490 if (__get_user(clen, &uiov->iov_len))
4495 iomsg->free_iov = NULL;
4497 iomsg->free_iov = iomsg->fast_iov;
4498 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4499 UIO_FASTIOV, &iomsg->free_iov,
4500 &iomsg->msg.msg_iter, true);
4509 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4510 struct io_async_msghdr *iomsg)
4512 iomsg->msg.msg_name = &iomsg->addr;
4514 #ifdef CONFIG_COMPAT
4515 if (req->ctx->compat)
4516 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4519 return __io_recvmsg_copy_hdr(req, iomsg);
4522 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4525 struct io_sr_msg *sr = &req->sr_msg;
4526 struct io_buffer *kbuf;
4528 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4533 req->flags |= REQ_F_BUFFER_SELECTED;
4537 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4539 return io_put_kbuf(req, req->sr_msg.kbuf);
4542 static int io_recvmsg_prep_async(struct io_kiocb *req)
4546 ret = io_recvmsg_copy_hdr(req, req->async_data);
4548 req->flags |= REQ_F_NEED_CLEANUP;
4552 static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4554 struct io_sr_msg *sr = &req->sr_msg;
4556 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4559 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4560 sr->len = READ_ONCE(sqe->len);
4561 sr->bgid = READ_ONCE(sqe->buf_group);
4562 sr->msg_flags = READ_ONCE(sqe->msg_flags) | MSG_NOSIGNAL;
4563 if (sr->msg_flags & MSG_DONTWAIT)
4564 req->flags |= REQ_F_NOWAIT;
4566 #ifdef CONFIG_COMPAT
4567 if (req->ctx->compat)
4568 sr->msg_flags |= MSG_CMSG_COMPAT;
4573 static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
4575 struct io_async_msghdr iomsg, *kmsg;
4576 struct socket *sock;
4577 struct io_buffer *kbuf;
4580 int ret, cflags = 0;
4581 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4583 sock = sock_from_file(req->file);
4584 if (unlikely(!sock))
4587 kmsg = req->async_data;
4589 ret = io_recvmsg_copy_hdr(req, &iomsg);
4595 if (req->flags & REQ_F_BUFFER_SELECT) {
4596 kbuf = io_recv_buffer_select(req, !force_nonblock);
4598 return PTR_ERR(kbuf);
4599 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4600 kmsg->fast_iov[0].iov_len = req->sr_msg.len;
4601 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
4602 1, req->sr_msg.len);
4605 flags = req->sr_msg.msg_flags;
4607 flags |= MSG_DONTWAIT;
4608 if (flags & MSG_WAITALL)
4609 min_ret = iov_iter_count(&kmsg->msg.msg_iter);
4611 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4612 kmsg->uaddr, flags);
4613 if (force_nonblock && ret == -EAGAIN)
4614 return io_setup_async_msg(req, kmsg);
4615 if (ret == -ERESTARTSYS)
4618 if (req->flags & REQ_F_BUFFER_SELECTED)
4619 cflags = io_put_recv_kbuf(req);
4620 /* fast path, check for non-NULL to avoid function call */
4622 kfree(kmsg->free_iov);
4623 req->flags &= ~REQ_F_NEED_CLEANUP;
4624 if (ret < min_ret || ((flags & MSG_WAITALL) && (kmsg->msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4626 __io_req_complete(req, issue_flags, ret, cflags);
4630 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
4632 struct io_buffer *kbuf;
4633 struct io_sr_msg *sr = &req->sr_msg;
4635 void __user *buf = sr->buf;
4636 struct socket *sock;
4640 int ret, cflags = 0;
4641 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4643 sock = sock_from_file(req->file);
4644 if (unlikely(!sock))
4647 if (req->flags & REQ_F_BUFFER_SELECT) {
4648 kbuf = io_recv_buffer_select(req, !force_nonblock);
4650 return PTR_ERR(kbuf);
4651 buf = u64_to_user_ptr(kbuf->addr);
4654 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4658 msg.msg_name = NULL;
4659 msg.msg_control = NULL;
4660 msg.msg_controllen = 0;
4661 msg.msg_namelen = 0;
4662 msg.msg_iocb = NULL;
4665 flags = req->sr_msg.msg_flags;
4667 flags |= MSG_DONTWAIT;
4668 if (flags & MSG_WAITALL)
4669 min_ret = iov_iter_count(&msg.msg_iter);
4671 ret = sock_recvmsg(sock, &msg, flags);
4672 if (force_nonblock && ret == -EAGAIN)
4674 if (ret == -ERESTARTSYS)
4677 if (req->flags & REQ_F_BUFFER_SELECTED)
4678 cflags = io_put_recv_kbuf(req);
4679 if (ret < min_ret || ((flags & MSG_WAITALL) && (msg.msg_flags & (MSG_TRUNC | MSG_CTRUNC))))
4681 __io_req_complete(req, issue_flags, ret, cflags);
4685 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4687 struct io_accept *accept = &req->accept;
4689 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4691 if (sqe->ioprio || sqe->len || sqe->buf_index)
4694 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4695 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4696 accept->flags = READ_ONCE(sqe->accept_flags);
4697 accept->nofile = rlimit(RLIMIT_NOFILE);
4701 static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
4703 struct io_accept *accept = &req->accept;
4704 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4705 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4708 if (req->file->f_flags & O_NONBLOCK)
4709 req->flags |= REQ_F_NOWAIT;
4711 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4712 accept->addr_len, accept->flags,
4714 if (ret == -EAGAIN && force_nonblock)
4717 if (ret == -ERESTARTSYS)
4721 __io_req_complete(req, issue_flags, ret, 0);
4725 static int io_connect_prep_async(struct io_kiocb *req)
4727 struct io_async_connect *io = req->async_data;
4728 struct io_connect *conn = &req->connect;
4730 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
4733 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4735 struct io_connect *conn = &req->connect;
4737 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4739 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4742 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4743 conn->addr_len = READ_ONCE(sqe->addr2);
4747 static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
4749 struct io_async_connect __io, *io;
4750 unsigned file_flags;
4752 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4754 if (req->async_data) {
4755 io = req->async_data;
4757 ret = move_addr_to_kernel(req->connect.addr,
4758 req->connect.addr_len,
4765 file_flags = force_nonblock ? O_NONBLOCK : 0;
4767 ret = __sys_connect_file(req->file, &io->address,
4768 req->connect.addr_len, file_flags);
4769 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4770 if (req->async_data)
4772 if (io_alloc_async_data(req)) {
4776 memcpy(req->async_data, &__io, sizeof(__io));
4779 if (ret == -ERESTARTSYS)
4784 __io_req_complete(req, issue_flags, ret, 0);
4787 #else /* !CONFIG_NET */
4788 #define IO_NETOP_FN(op) \
4789 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
4791 return -EOPNOTSUPP; \
4794 #define IO_NETOP_PREP(op) \
4796 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
4798 return -EOPNOTSUPP; \
4801 #define IO_NETOP_PREP_ASYNC(op) \
4803 static int io_##op##_prep_async(struct io_kiocb *req) \
4805 return -EOPNOTSUPP; \
4808 IO_NETOP_PREP_ASYNC(sendmsg);
4809 IO_NETOP_PREP_ASYNC(recvmsg);
4810 IO_NETOP_PREP_ASYNC(connect);
4811 IO_NETOP_PREP(accept);
4814 #endif /* CONFIG_NET */
4816 struct io_poll_table {
4817 struct poll_table_struct pt;
4818 struct io_kiocb *req;
4822 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4823 __poll_t mask, task_work_func_t func)
4827 /* for instances that support it check for an event match first: */
4828 if (mask && !(mask & poll->events))
4831 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4833 list_del_init(&poll->wait.entry);
4836 req->task_work.func = func;
4839 * If this fails, then the task is exiting. When a task exits, the
4840 * work gets canceled, so just cancel this request as well instead
4841 * of executing it. We can't safely execute it anyway, as we may not
4842 * have the needed state needed for it anyway.
4844 ret = io_req_task_work_add(req);
4845 if (unlikely(ret)) {
4846 WRITE_ONCE(poll->canceled, true);
4847 io_req_task_work_add_fallback(req, func);
4852 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4853 __acquires(&req->ctx->completion_lock)
4855 struct io_ring_ctx *ctx = req->ctx;
4857 if (!req->result && !READ_ONCE(poll->canceled)) {
4858 struct poll_table_struct pt = { ._key = poll->events };
4860 req->result = vfs_poll(req->file, &pt) & poll->events;
4863 spin_lock_irq(&ctx->completion_lock);
4864 if (!req->result && !READ_ONCE(poll->canceled)) {
4865 add_wait_queue(poll->head, &poll->wait);
4872 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4874 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4875 if (req->opcode == IORING_OP_POLL_ADD)
4876 return req->async_data;
4877 return req->apoll->double_poll;
4880 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4882 if (req->opcode == IORING_OP_POLL_ADD)
4884 return &req->apoll->poll;
4887 static void io_poll_remove_double(struct io_kiocb *req)
4888 __must_hold(&req->ctx->completion_lock)
4890 struct io_poll_iocb *poll = io_poll_get_double(req);
4892 lockdep_assert_held(&req->ctx->completion_lock);
4894 if (poll && poll->head) {
4895 struct wait_queue_head *head = poll->head;
4897 spin_lock(&head->lock);
4898 list_del_init(&poll->wait.entry);
4899 if (poll->wait.private)
4902 spin_unlock(&head->lock);
4906 static bool io_poll_complete(struct io_kiocb *req, __poll_t mask)
4907 __must_hold(&req->ctx->completion_lock)
4909 struct io_ring_ctx *ctx = req->ctx;
4910 unsigned flags = IORING_CQE_F_MORE;
4913 if (READ_ONCE(req->poll.canceled)) {
4915 req->poll.events |= EPOLLONESHOT;
4917 error = mangle_poll(mask);
4919 if (req->poll.events & EPOLLONESHOT)
4921 if (!io_cqring_fill_event(ctx, req->user_data, error, flags)) {
4922 io_poll_remove_waitqs(req);
4923 req->poll.done = true;
4926 if (flags & IORING_CQE_F_MORE)
4929 io_commit_cqring(ctx);
4930 return !(flags & IORING_CQE_F_MORE);
4933 static void io_poll_task_func(struct callback_head *cb)
4935 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4936 struct io_ring_ctx *ctx = req->ctx;
4937 struct io_kiocb *nxt;
4939 if (io_poll_rewait(req, &req->poll)) {
4940 spin_unlock_irq(&ctx->completion_lock);
4944 done = io_poll_complete(req, req->result);
4946 hash_del(&req->hash_node);
4949 add_wait_queue(req->poll.head, &req->poll.wait);
4951 spin_unlock_irq(&ctx->completion_lock);
4952 io_cqring_ev_posted(ctx);
4955 nxt = io_put_req_find_next(req);
4957 __io_req_task_submit(nxt);
4962 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4963 int sync, void *key)
4965 struct io_kiocb *req = wait->private;
4966 struct io_poll_iocb *poll = io_poll_get_single(req);
4967 __poll_t mask = key_to_poll(key);
4969 /* for instances that support it check for an event match first: */
4970 if (mask && !(mask & poll->events))
4972 if (!(poll->events & EPOLLONESHOT))
4973 return poll->wait.func(&poll->wait, mode, sync, key);
4975 list_del_init(&wait->entry);
4977 if (poll && poll->head) {
4980 spin_lock(&poll->head->lock);
4981 done = list_empty(&poll->wait.entry);
4983 list_del_init(&poll->wait.entry);
4984 /* make sure double remove sees this as being gone */
4985 wait->private = NULL;
4986 spin_unlock(&poll->head->lock);
4988 /* use wait func handler, so it matches the rq type */
4989 poll->wait.func(&poll->wait, mode, sync, key);
4996 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
4997 wait_queue_func_t wake_func)
5001 poll->canceled = false;
5002 #define IO_POLL_UNMASK (EPOLLERR|EPOLLHUP|EPOLLNVAL|EPOLLRDHUP)
5003 /* mask in events that we always want/need */
5004 poll->events = events | IO_POLL_UNMASK;
5005 INIT_LIST_HEAD(&poll->wait.entry);
5006 init_waitqueue_func_entry(&poll->wait, wake_func);
5009 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5010 struct wait_queue_head *head,
5011 struct io_poll_iocb **poll_ptr)
5013 struct io_kiocb *req = pt->req;
5016 * If poll->head is already set, it's because the file being polled
5017 * uses multiple waitqueues for poll handling (eg one for read, one
5018 * for write). Setup a separate io_poll_iocb if this happens.
5020 if (unlikely(poll->head)) {
5021 struct io_poll_iocb *poll_one = poll;
5023 /* already have a 2nd entry, fail a third attempt */
5025 pt->error = -EINVAL;
5029 * Can't handle multishot for double wait for now, turn it
5030 * into one-shot mode.
5032 if (!(poll_one->events & EPOLLONESHOT))
5033 poll_one->events |= EPOLLONESHOT;
5034 /* double add on the same waitqueue head, ignore */
5035 if (poll_one->head == head)
5037 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5039 pt->error = -ENOMEM;
5042 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5044 poll->wait.private = req;
5051 if (poll->events & EPOLLEXCLUSIVE)
5052 add_wait_queue_exclusive(head, &poll->wait);
5054 add_wait_queue(head, &poll->wait);
5057 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5058 struct poll_table_struct *p)
5060 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5061 struct async_poll *apoll = pt->req->apoll;
5063 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5066 static void io_async_task_func(struct callback_head *cb)
5068 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
5069 struct async_poll *apoll = req->apoll;
5070 struct io_ring_ctx *ctx = req->ctx;
5072 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
5074 if (io_poll_rewait(req, &apoll->poll)) {
5075 spin_unlock_irq(&ctx->completion_lock);
5079 hash_del(&req->hash_node);
5080 io_poll_remove_double(req);
5081 spin_unlock_irq(&ctx->completion_lock);
5083 if (!READ_ONCE(apoll->poll.canceled))
5084 __io_req_task_submit(req);
5086 io_req_complete_failed(req, -ECANCELED);
5089 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5092 struct io_kiocb *req = wait->private;
5093 struct io_poll_iocb *poll = &req->apoll->poll;
5095 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5098 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5101 static void io_poll_req_insert(struct io_kiocb *req)
5103 struct io_ring_ctx *ctx = req->ctx;
5104 struct hlist_head *list;
5106 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5107 hlist_add_head(&req->hash_node, list);
5110 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5111 struct io_poll_iocb *poll,
5112 struct io_poll_table *ipt, __poll_t mask,
5113 wait_queue_func_t wake_func)
5114 __acquires(&ctx->completion_lock)
5116 struct io_ring_ctx *ctx = req->ctx;
5117 bool cancel = false;
5119 INIT_HLIST_NODE(&req->hash_node);
5120 io_init_poll_iocb(poll, mask, wake_func);
5121 poll->file = req->file;
5122 poll->wait.private = req;
5124 ipt->pt._key = mask;
5126 ipt->error = -EINVAL;
5128 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5130 spin_lock_irq(&ctx->completion_lock);
5131 if (likely(poll->head)) {
5132 spin_lock(&poll->head->lock);
5133 if (unlikely(list_empty(&poll->wait.entry))) {
5139 if ((mask && (poll->events & EPOLLONESHOT)) || ipt->error)
5140 list_del_init(&poll->wait.entry);
5142 WRITE_ONCE(poll->canceled, true);
5143 else if (!poll->done) /* actually waiting for an event */
5144 io_poll_req_insert(req);
5145 spin_unlock(&poll->head->lock);
5151 static bool io_arm_poll_handler(struct io_kiocb *req)
5153 const struct io_op_def *def = &io_op_defs[req->opcode];
5154 struct io_ring_ctx *ctx = req->ctx;
5155 struct async_poll *apoll;
5156 struct io_poll_table ipt;
5160 if (!req->file || !file_can_poll(req->file))
5162 if (req->flags & REQ_F_POLLED)
5166 else if (def->pollout)
5170 /* if we can't nonblock try, then no point in arming a poll handler */
5171 if (!io_file_supports_async(req, rw))
5174 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5175 if (unlikely(!apoll))
5177 apoll->double_poll = NULL;
5179 req->flags |= REQ_F_POLLED;
5182 mask = EPOLLONESHOT;
5184 mask |= POLLIN | POLLRDNORM;
5186 mask |= POLLOUT | POLLWRNORM;
5188 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5189 if ((req->opcode == IORING_OP_RECVMSG) &&
5190 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5193 mask |= POLLERR | POLLPRI;
5195 ipt.pt._qproc = io_async_queue_proc;
5197 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5199 if (ret || ipt.error) {
5200 io_poll_remove_double(req);
5201 spin_unlock_irq(&ctx->completion_lock);
5204 spin_unlock_irq(&ctx->completion_lock);
5205 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
5206 apoll->poll.events);
5210 static bool __io_poll_remove_one(struct io_kiocb *req,
5211 struct io_poll_iocb *poll, bool do_cancel)
5212 __must_hold(&req->ctx->completion_lock)
5214 bool do_complete = false;
5218 spin_lock(&poll->head->lock);
5220 WRITE_ONCE(poll->canceled, true);
5221 if (!list_empty(&poll->wait.entry)) {
5222 list_del_init(&poll->wait.entry);
5225 spin_unlock(&poll->head->lock);
5226 hash_del(&req->hash_node);
5230 static bool io_poll_remove_waitqs(struct io_kiocb *req)
5231 __must_hold(&req->ctx->completion_lock)
5235 io_poll_remove_double(req);
5236 do_complete = __io_poll_remove_one(req, io_poll_get_single(req), true);
5238 if (req->opcode != IORING_OP_POLL_ADD && do_complete) {
5239 /* non-poll requests have submit ref still */
5245 static bool io_poll_remove_one(struct io_kiocb *req)
5246 __must_hold(&req->ctx->completion_lock)
5250 do_complete = io_poll_remove_waitqs(req);
5252 io_cqring_fill_event(req->ctx, req->user_data, -ECANCELED, 0);
5253 io_commit_cqring(req->ctx);
5255 io_put_req_deferred(req, 1);
5262 * Returns true if we found and killed one or more poll requests
5264 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5267 struct hlist_node *tmp;
5268 struct io_kiocb *req;
5271 spin_lock_irq(&ctx->completion_lock);
5272 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5273 struct hlist_head *list;
5275 list = &ctx->cancel_hash[i];
5276 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5277 if (io_match_task(req, tsk, cancel_all))
5278 posted += io_poll_remove_one(req);
5281 spin_unlock_irq(&ctx->completion_lock);
5284 io_cqring_ev_posted(ctx);
5289 static struct io_kiocb *io_poll_find(struct io_ring_ctx *ctx, __u64 sqe_addr,
5291 __must_hold(&ctx->completion_lock)
5293 struct hlist_head *list;
5294 struct io_kiocb *req;
5296 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5297 hlist_for_each_entry(req, list, hash_node) {
5298 if (sqe_addr != req->user_data)
5300 if (poll_only && req->opcode != IORING_OP_POLL_ADD)
5307 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr,
5309 __must_hold(&ctx->completion_lock)
5311 struct io_kiocb *req;
5313 req = io_poll_find(ctx, sqe_addr, poll_only);
5316 if (io_poll_remove_one(req))
5322 static __poll_t io_poll_parse_events(const struct io_uring_sqe *sqe,
5327 events = READ_ONCE(sqe->poll32_events);
5329 events = swahw32(events);
5331 if (!(flags & IORING_POLL_ADD_MULTI))
5332 events |= EPOLLONESHOT;
5333 return demangle_poll(events) | (events & (EPOLLEXCLUSIVE|EPOLLONESHOT));
5336 static int io_poll_update_prep(struct io_kiocb *req,
5337 const struct io_uring_sqe *sqe)
5339 struct io_poll_update *upd = &req->poll_update;
5342 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5344 if (sqe->ioprio || sqe->buf_index)
5346 flags = READ_ONCE(sqe->len);
5347 if (flags & ~(IORING_POLL_UPDATE_EVENTS | IORING_POLL_UPDATE_USER_DATA |
5348 IORING_POLL_ADD_MULTI))
5350 /* meaningless without update */
5351 if (flags == IORING_POLL_ADD_MULTI)
5354 upd->old_user_data = READ_ONCE(sqe->addr);
5355 upd->update_events = flags & IORING_POLL_UPDATE_EVENTS;
5356 upd->update_user_data = flags & IORING_POLL_UPDATE_USER_DATA;
5358 upd->new_user_data = READ_ONCE(sqe->off);
5359 if (!upd->update_user_data && upd->new_user_data)
5361 if (upd->update_events)
5362 upd->events = io_poll_parse_events(sqe, flags);
5363 else if (sqe->poll32_events)
5369 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5372 struct io_kiocb *req = wait->private;
5373 struct io_poll_iocb *poll = &req->poll;
5375 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5378 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5379 struct poll_table_struct *p)
5381 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5383 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5386 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5388 struct io_poll_iocb *poll = &req->poll;
5391 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5393 if (sqe->ioprio || sqe->buf_index || sqe->off || sqe->addr)
5395 flags = READ_ONCE(sqe->len);
5396 if (flags & ~IORING_POLL_ADD_MULTI)
5399 poll->events = io_poll_parse_events(sqe, flags);
5403 static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5405 struct io_poll_iocb *poll = &req->poll;
5406 struct io_ring_ctx *ctx = req->ctx;
5407 struct io_poll_table ipt;
5410 ipt.pt._qproc = io_poll_queue_proc;
5412 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5415 if (mask) { /* no async, we'd stolen it */
5417 io_poll_complete(req, mask);
5419 spin_unlock_irq(&ctx->completion_lock);
5422 io_cqring_ev_posted(ctx);
5423 if (poll->events & EPOLLONESHOT)
5429 static int io_poll_update(struct io_kiocb *req, unsigned int issue_flags)
5431 struct io_ring_ctx *ctx = req->ctx;
5432 struct io_kiocb *preq;
5436 spin_lock_irq(&ctx->completion_lock);
5437 preq = io_poll_find(ctx, req->poll_update.old_user_data, true);
5443 if (!req->poll_update.update_events && !req->poll_update.update_user_data) {
5445 ret = io_poll_remove_one(preq) ? 0 : -EALREADY;
5450 * Don't allow racy completion with singleshot, as we cannot safely
5451 * update those. For multishot, if we're racing with completion, just
5452 * let completion re-add it.
5454 completing = !__io_poll_remove_one(preq, &preq->poll, false);
5455 if (completing && (preq->poll.events & EPOLLONESHOT)) {
5459 /* we now have a detached poll request. reissue. */
5463 spin_unlock_irq(&ctx->completion_lock);
5465 io_req_complete(req, ret);
5468 /* only mask one event flags, keep behavior flags */
5469 if (req->poll_update.update_events) {
5470 preq->poll.events &= ~0xffff;
5471 preq->poll.events |= req->poll_update.events & 0xffff;
5472 preq->poll.events |= IO_POLL_UNMASK;
5474 if (req->poll_update.update_user_data)
5475 preq->user_data = req->poll_update.new_user_data;
5476 spin_unlock_irq(&ctx->completion_lock);
5478 /* complete update request, we're done with it */
5479 io_req_complete(req, ret);
5482 ret = io_poll_add(preq, issue_flags);
5485 io_req_complete(preq, ret);
5491 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5493 struct io_timeout_data *data = container_of(timer,
5494 struct io_timeout_data, timer);
5495 struct io_kiocb *req = data->req;
5496 struct io_ring_ctx *ctx = req->ctx;
5497 unsigned long flags;
5499 spin_lock_irqsave(&ctx->completion_lock, flags);
5500 list_del_init(&req->timeout.list);
5501 atomic_set(&req->ctx->cq_timeouts,
5502 atomic_read(&req->ctx->cq_timeouts) + 1);
5504 io_cqring_fill_event(ctx, req->user_data, -ETIME, 0);
5505 io_commit_cqring(ctx);
5506 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5508 io_cqring_ev_posted(ctx);
5511 return HRTIMER_NORESTART;
5514 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5516 __must_hold(&ctx->completion_lock)
5518 struct io_timeout_data *io;
5519 struct io_kiocb *req;
5522 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5523 found = user_data == req->user_data;
5528 return ERR_PTR(-ENOENT);
5530 io = req->async_data;
5531 if (hrtimer_try_to_cancel(&io->timer) == -1)
5532 return ERR_PTR(-EALREADY);
5533 list_del_init(&req->timeout.list);
5537 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5538 __must_hold(&ctx->completion_lock)
5540 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5543 return PTR_ERR(req);
5546 io_cqring_fill_event(ctx, req->user_data, -ECANCELED, 0);
5547 io_put_req_deferred(req, 1);
5551 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5552 struct timespec64 *ts, enum hrtimer_mode mode)
5553 __must_hold(&ctx->completion_lock)
5555 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5556 struct io_timeout_data *data;
5559 return PTR_ERR(req);
5561 req->timeout.off = 0; /* noseq */
5562 data = req->async_data;
5563 list_add_tail(&req->timeout.list, &ctx->timeout_list);
5564 hrtimer_init(&data->timer, CLOCK_MONOTONIC, mode);
5565 data->timer.function = io_timeout_fn;
5566 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
5570 static int io_timeout_remove_prep(struct io_kiocb *req,
5571 const struct io_uring_sqe *sqe)
5573 struct io_timeout_rem *tr = &req->timeout_rem;
5575 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5577 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5579 if (sqe->ioprio || sqe->buf_index || sqe->len)
5582 tr->addr = READ_ONCE(sqe->addr);
5583 tr->flags = READ_ONCE(sqe->timeout_flags);
5584 if (tr->flags & IORING_TIMEOUT_UPDATE) {
5585 if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS))
5587 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
5589 } else if (tr->flags) {
5590 /* timeout removal doesn't support flags */
5597 static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
5599 return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
5604 * Remove or update an existing timeout command
5606 static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
5608 struct io_timeout_rem *tr = &req->timeout_rem;
5609 struct io_ring_ctx *ctx = req->ctx;
5612 spin_lock_irq(&ctx->completion_lock);
5613 if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE))
5614 ret = io_timeout_cancel(ctx, tr->addr);
5616 ret = io_timeout_update(ctx, tr->addr, &tr->ts,
5617 io_translate_timeout_mode(tr->flags));
5619 io_cqring_fill_event(ctx, req->user_data, ret, 0);
5620 io_commit_cqring(ctx);
5621 spin_unlock_irq(&ctx->completion_lock);
5622 io_cqring_ev_posted(ctx);
5629 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5630 bool is_timeout_link)
5632 struct io_timeout_data *data;
5634 u32 off = READ_ONCE(sqe->off);
5636 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5638 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5640 if (off && is_timeout_link)
5642 flags = READ_ONCE(sqe->timeout_flags);
5643 if (flags & ~IORING_TIMEOUT_ABS)
5646 req->timeout.off = off;
5647 if (unlikely(off && !req->ctx->off_timeout_used))
5648 req->ctx->off_timeout_used = true;
5650 if (!req->async_data && io_alloc_async_data(req))
5653 data = req->async_data;
5656 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5659 data->mode = io_translate_timeout_mode(flags);
5660 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5661 if (is_timeout_link)
5662 io_req_track_inflight(req);
5666 static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
5668 struct io_ring_ctx *ctx = req->ctx;
5669 struct io_timeout_data *data = req->async_data;
5670 struct list_head *entry;
5671 u32 tail, off = req->timeout.off;
5673 spin_lock_irq(&ctx->completion_lock);
5676 * sqe->off holds how many events that need to occur for this
5677 * timeout event to be satisfied. If it isn't set, then this is
5678 * a pure timeout request, sequence isn't used.
5680 if (io_is_timeout_noseq(req)) {
5681 entry = ctx->timeout_list.prev;
5685 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5686 req->timeout.target_seq = tail + off;
5688 /* Update the last seq here in case io_flush_timeouts() hasn't.
5689 * This is safe because ->completion_lock is held, and submissions
5690 * and completions are never mixed in the same ->completion_lock section.
5692 ctx->cq_last_tm_flush = tail;
5695 * Insertion sort, ensuring the first entry in the list is always
5696 * the one we need first.
5698 list_for_each_prev(entry, &ctx->timeout_list) {
5699 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5702 if (io_is_timeout_noseq(nxt))
5704 /* nxt.seq is behind @tail, otherwise would've been completed */
5705 if (off >= nxt->timeout.target_seq - tail)
5709 list_add(&req->timeout.list, entry);
5710 data->timer.function = io_timeout_fn;
5711 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5712 spin_unlock_irq(&ctx->completion_lock);
5716 struct io_cancel_data {
5717 struct io_ring_ctx *ctx;
5721 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5723 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5724 struct io_cancel_data *cd = data;
5726 return req->ctx == cd->ctx && req->user_data == cd->user_data;
5729 static int io_async_cancel_one(struct io_uring_task *tctx, u64 user_data,
5730 struct io_ring_ctx *ctx)
5732 struct io_cancel_data data = { .ctx = ctx, .user_data = user_data, };
5733 enum io_wq_cancel cancel_ret;
5736 if (!tctx || !tctx->io_wq)
5739 cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, &data, false);
5740 switch (cancel_ret) {
5741 case IO_WQ_CANCEL_OK:
5744 case IO_WQ_CANCEL_RUNNING:
5747 case IO_WQ_CANCEL_NOTFOUND:
5755 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5756 struct io_kiocb *req, __u64 sqe_addr,
5759 unsigned long flags;
5762 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5763 spin_lock_irqsave(&ctx->completion_lock, flags);
5766 ret = io_timeout_cancel(ctx, sqe_addr);
5769 ret = io_poll_cancel(ctx, sqe_addr, false);
5773 io_cqring_fill_event(ctx, req->user_data, ret, 0);
5774 io_commit_cqring(ctx);
5775 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5776 io_cqring_ev_posted(ctx);
5782 static int io_async_cancel_prep(struct io_kiocb *req,
5783 const struct io_uring_sqe *sqe)
5785 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5787 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5789 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5792 req->cancel.addr = READ_ONCE(sqe->addr);
5796 static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
5798 struct io_ring_ctx *ctx = req->ctx;
5799 u64 sqe_addr = req->cancel.addr;
5800 struct io_tctx_node *node;
5803 /* tasks should wait for their io-wq threads, so safe w/o sync */
5804 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5805 spin_lock_irq(&ctx->completion_lock);
5808 ret = io_timeout_cancel(ctx, sqe_addr);
5811 ret = io_poll_cancel(ctx, sqe_addr, false);
5814 spin_unlock_irq(&ctx->completion_lock);
5816 /* slow path, try all io-wq's */
5817 io_ring_submit_lock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5819 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
5820 struct io_uring_task *tctx = node->task->io_uring;
5822 ret = io_async_cancel_one(tctx, req->cancel.addr, ctx);
5826 io_ring_submit_unlock(ctx, !(issue_flags & IO_URING_F_NONBLOCK));
5828 spin_lock_irq(&ctx->completion_lock);
5830 io_cqring_fill_event(ctx, req->user_data, ret, 0);
5831 io_commit_cqring(ctx);
5832 spin_unlock_irq(&ctx->completion_lock);
5833 io_cqring_ev_posted(ctx);
5841 static int io_rsrc_update_prep(struct io_kiocb *req,
5842 const struct io_uring_sqe *sqe)
5844 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5846 if (sqe->ioprio || sqe->rw_flags)
5849 req->rsrc_update.offset = READ_ONCE(sqe->off);
5850 req->rsrc_update.nr_args = READ_ONCE(sqe->len);
5851 if (!req->rsrc_update.nr_args)
5853 req->rsrc_update.arg = READ_ONCE(sqe->addr);
5857 static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
5859 struct io_ring_ctx *ctx = req->ctx;
5860 struct io_uring_rsrc_update2 up;
5863 if (issue_flags & IO_URING_F_NONBLOCK)
5866 up.offset = req->rsrc_update.offset;
5867 up.data = req->rsrc_update.arg;
5872 mutex_lock(&ctx->uring_lock);
5873 ret = __io_register_rsrc_update(ctx, IORING_RSRC_FILE,
5874 &up, req->rsrc_update.nr_args);
5875 mutex_unlock(&ctx->uring_lock);
5879 __io_req_complete(req, issue_flags, ret, 0);
5883 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5885 switch (req->opcode) {
5888 case IORING_OP_READV:
5889 case IORING_OP_READ_FIXED:
5890 case IORING_OP_READ:
5891 return io_read_prep(req, sqe);
5892 case IORING_OP_WRITEV:
5893 case IORING_OP_WRITE_FIXED:
5894 case IORING_OP_WRITE:
5895 return io_write_prep(req, sqe);
5896 case IORING_OP_POLL_ADD:
5897 return io_poll_add_prep(req, sqe);
5898 case IORING_OP_POLL_REMOVE:
5899 return io_poll_update_prep(req, sqe);
5900 case IORING_OP_FSYNC:
5901 return io_fsync_prep(req, sqe);
5902 case IORING_OP_SYNC_FILE_RANGE:
5903 return io_sfr_prep(req, sqe);
5904 case IORING_OP_SENDMSG:
5905 case IORING_OP_SEND:
5906 return io_sendmsg_prep(req, sqe);
5907 case IORING_OP_RECVMSG:
5908 case IORING_OP_RECV:
5909 return io_recvmsg_prep(req, sqe);
5910 case IORING_OP_CONNECT:
5911 return io_connect_prep(req, sqe);
5912 case IORING_OP_TIMEOUT:
5913 return io_timeout_prep(req, sqe, false);
5914 case IORING_OP_TIMEOUT_REMOVE:
5915 return io_timeout_remove_prep(req, sqe);
5916 case IORING_OP_ASYNC_CANCEL:
5917 return io_async_cancel_prep(req, sqe);
5918 case IORING_OP_LINK_TIMEOUT:
5919 return io_timeout_prep(req, sqe, true);
5920 case IORING_OP_ACCEPT:
5921 return io_accept_prep(req, sqe);
5922 case IORING_OP_FALLOCATE:
5923 return io_fallocate_prep(req, sqe);
5924 case IORING_OP_OPENAT:
5925 return io_openat_prep(req, sqe);
5926 case IORING_OP_CLOSE:
5927 return io_close_prep(req, sqe);
5928 case IORING_OP_FILES_UPDATE:
5929 return io_rsrc_update_prep(req, sqe);
5930 case IORING_OP_STATX:
5931 return io_statx_prep(req, sqe);
5932 case IORING_OP_FADVISE:
5933 return io_fadvise_prep(req, sqe);
5934 case IORING_OP_MADVISE:
5935 return io_madvise_prep(req, sqe);
5936 case IORING_OP_OPENAT2:
5937 return io_openat2_prep(req, sqe);
5938 case IORING_OP_EPOLL_CTL:
5939 return io_epoll_ctl_prep(req, sqe);
5940 case IORING_OP_SPLICE:
5941 return io_splice_prep(req, sqe);
5942 case IORING_OP_PROVIDE_BUFFERS:
5943 return io_provide_buffers_prep(req, sqe);
5944 case IORING_OP_REMOVE_BUFFERS:
5945 return io_remove_buffers_prep(req, sqe);
5947 return io_tee_prep(req, sqe);
5948 case IORING_OP_SHUTDOWN:
5949 return io_shutdown_prep(req, sqe);
5950 case IORING_OP_RENAMEAT:
5951 return io_renameat_prep(req, sqe);
5952 case IORING_OP_UNLINKAT:
5953 return io_unlinkat_prep(req, sqe);
5956 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5961 static int io_req_prep_async(struct io_kiocb *req)
5963 if (!io_op_defs[req->opcode].needs_async_setup)
5965 if (WARN_ON_ONCE(req->async_data))
5967 if (io_alloc_async_data(req))
5970 switch (req->opcode) {
5971 case IORING_OP_READV:
5972 return io_rw_prep_async(req, READ);
5973 case IORING_OP_WRITEV:
5974 return io_rw_prep_async(req, WRITE);
5975 case IORING_OP_SENDMSG:
5976 return io_sendmsg_prep_async(req);
5977 case IORING_OP_RECVMSG:
5978 return io_recvmsg_prep_async(req);
5979 case IORING_OP_CONNECT:
5980 return io_connect_prep_async(req);
5982 printk_once(KERN_WARNING "io_uring: prep_async() bad opcode %d\n",
5987 static u32 io_get_sequence(struct io_kiocb *req)
5989 struct io_kiocb *pos;
5990 struct io_ring_ctx *ctx = req->ctx;
5993 io_for_each_link(pos, req)
5995 return ctx->cached_sq_head - nr_reqs;
5998 static bool io_drain_req(struct io_kiocb *req)
6000 struct io_ring_ctx *ctx = req->ctx;
6001 struct io_defer_entry *de;
6005 /* Still need defer if there is pending req in defer list. */
6006 if (likely(list_empty_careful(&ctx->defer_list) &&
6007 !(req->flags & REQ_F_IO_DRAIN)))
6010 seq = io_get_sequence(req);
6011 /* Still a chance to pass the sequence check */
6012 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
6015 ret = io_req_prep_async(req);
6018 io_prep_async_link(req);
6019 de = kmalloc(sizeof(*de), GFP_KERNEL);
6021 io_req_complete_failed(req, ret);
6025 spin_lock_irq(&ctx->completion_lock);
6026 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
6027 spin_unlock_irq(&ctx->completion_lock);
6029 io_queue_async_work(req);
6033 trace_io_uring_defer(ctx, req, req->user_data);
6036 list_add_tail(&de->list, &ctx->defer_list);
6037 spin_unlock_irq(&ctx->completion_lock);
6041 static void io_clean_op(struct io_kiocb *req)
6043 if (req->flags & REQ_F_BUFFER_SELECTED) {
6044 switch (req->opcode) {
6045 case IORING_OP_READV:
6046 case IORING_OP_READ_FIXED:
6047 case IORING_OP_READ:
6048 kfree((void *)(unsigned long)req->rw.addr);
6050 case IORING_OP_RECVMSG:
6051 case IORING_OP_RECV:
6052 kfree(req->sr_msg.kbuf);
6055 req->flags &= ~REQ_F_BUFFER_SELECTED;
6058 if (req->flags & REQ_F_NEED_CLEANUP) {
6059 switch (req->opcode) {
6060 case IORING_OP_READV:
6061 case IORING_OP_READ_FIXED:
6062 case IORING_OP_READ:
6063 case IORING_OP_WRITEV:
6064 case IORING_OP_WRITE_FIXED:
6065 case IORING_OP_WRITE: {
6066 struct io_async_rw *io = req->async_data;
6068 kfree(io->free_iovec);
6071 case IORING_OP_RECVMSG:
6072 case IORING_OP_SENDMSG: {
6073 struct io_async_msghdr *io = req->async_data;
6075 kfree(io->free_iov);
6078 case IORING_OP_SPLICE:
6080 if (!(req->splice.flags & SPLICE_F_FD_IN_FIXED))
6081 io_put_file(req->splice.file_in);
6083 case IORING_OP_OPENAT:
6084 case IORING_OP_OPENAT2:
6085 if (req->open.filename)
6086 putname(req->open.filename);
6088 case IORING_OP_RENAMEAT:
6089 putname(req->rename.oldpath);
6090 putname(req->rename.newpath);
6092 case IORING_OP_UNLINKAT:
6093 putname(req->unlink.filename);
6096 req->flags &= ~REQ_F_NEED_CLEANUP;
6098 if ((req->flags & REQ_F_POLLED) && req->apoll) {
6099 kfree(req->apoll->double_poll);
6103 if (req->flags & REQ_F_INFLIGHT) {
6104 struct io_uring_task *tctx = req->task->io_uring;
6106 atomic_dec(&tctx->inflight_tracked);
6107 req->flags &= ~REQ_F_INFLIGHT;
6111 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
6113 struct io_ring_ctx *ctx = req->ctx;
6114 const struct cred *creds = NULL;
6117 if (req->work.creds && req->work.creds != current_cred())
6118 creds = override_creds(req->work.creds);
6120 switch (req->opcode) {
6122 ret = io_nop(req, issue_flags);
6124 case IORING_OP_READV:
6125 case IORING_OP_READ_FIXED:
6126 case IORING_OP_READ:
6127 ret = io_read(req, issue_flags);
6129 case IORING_OP_WRITEV:
6130 case IORING_OP_WRITE_FIXED:
6131 case IORING_OP_WRITE:
6132 ret = io_write(req, issue_flags);
6134 case IORING_OP_FSYNC:
6135 ret = io_fsync(req, issue_flags);
6137 case IORING_OP_POLL_ADD:
6138 ret = io_poll_add(req, issue_flags);
6140 case IORING_OP_POLL_REMOVE:
6141 ret = io_poll_update(req, issue_flags);
6143 case IORING_OP_SYNC_FILE_RANGE:
6144 ret = io_sync_file_range(req, issue_flags);
6146 case IORING_OP_SENDMSG:
6147 ret = io_sendmsg(req, issue_flags);
6149 case IORING_OP_SEND:
6150 ret = io_send(req, issue_flags);
6152 case IORING_OP_RECVMSG:
6153 ret = io_recvmsg(req, issue_flags);
6155 case IORING_OP_RECV:
6156 ret = io_recv(req, issue_flags);
6158 case IORING_OP_TIMEOUT:
6159 ret = io_timeout(req, issue_flags);
6161 case IORING_OP_TIMEOUT_REMOVE:
6162 ret = io_timeout_remove(req, issue_flags);
6164 case IORING_OP_ACCEPT:
6165 ret = io_accept(req, issue_flags);
6167 case IORING_OP_CONNECT:
6168 ret = io_connect(req, issue_flags);
6170 case IORING_OP_ASYNC_CANCEL:
6171 ret = io_async_cancel(req, issue_flags);
6173 case IORING_OP_FALLOCATE:
6174 ret = io_fallocate(req, issue_flags);
6176 case IORING_OP_OPENAT:
6177 ret = io_openat(req, issue_flags);
6179 case IORING_OP_CLOSE:
6180 ret = io_close(req, issue_flags);
6182 case IORING_OP_FILES_UPDATE:
6183 ret = io_files_update(req, issue_flags);
6185 case IORING_OP_STATX:
6186 ret = io_statx(req, issue_flags);
6188 case IORING_OP_FADVISE:
6189 ret = io_fadvise(req, issue_flags);
6191 case IORING_OP_MADVISE:
6192 ret = io_madvise(req, issue_flags);
6194 case IORING_OP_OPENAT2:
6195 ret = io_openat2(req, issue_flags);
6197 case IORING_OP_EPOLL_CTL:
6198 ret = io_epoll_ctl(req, issue_flags);
6200 case IORING_OP_SPLICE:
6201 ret = io_splice(req, issue_flags);
6203 case IORING_OP_PROVIDE_BUFFERS:
6204 ret = io_provide_buffers(req, issue_flags);
6206 case IORING_OP_REMOVE_BUFFERS:
6207 ret = io_remove_buffers(req, issue_flags);
6210 ret = io_tee(req, issue_flags);
6212 case IORING_OP_SHUTDOWN:
6213 ret = io_shutdown(req, issue_flags);
6215 case IORING_OP_RENAMEAT:
6216 ret = io_renameat(req, issue_flags);
6218 case IORING_OP_UNLINKAT:
6219 ret = io_unlinkat(req, issue_flags);
6227 revert_creds(creds);
6230 /* If the op doesn't have a file, we're not polling for it */
6231 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file)
6232 io_iopoll_req_issued(req);
6237 static void io_wq_submit_work(struct io_wq_work *work)
6239 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6240 struct io_kiocb *timeout;
6243 timeout = io_prep_linked_timeout(req);
6245 io_queue_linked_timeout(timeout);
6247 if (work->flags & IO_WQ_WORK_CANCEL)
6252 ret = io_issue_sqe(req, 0);
6254 * We can get EAGAIN for polled IO even though we're
6255 * forcing a sync submission from here, since we can't
6256 * wait for request slots on the block side.
6264 /* avoid locking problems by failing it from a clean context */
6266 /* io-wq is going to take one down */
6268 io_req_task_queue_fail(req, ret);
6272 #define FFS_ASYNC_READ 0x1UL
6273 #define FFS_ASYNC_WRITE 0x2UL
6275 #define FFS_ISREG 0x4UL
6277 #define FFS_ISREG 0x0UL
6279 #define FFS_MASK ~(FFS_ASYNC_READ|FFS_ASYNC_WRITE|FFS_ISREG)
6281 static inline struct io_fixed_file *io_fixed_file_slot(struct io_file_table *table,
6284 struct io_fixed_file *table_l2;
6286 table_l2 = table->files[i >> IORING_FILE_TABLE_SHIFT];
6287 return &table_l2[i & IORING_FILE_TABLE_MASK];
6290 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6293 struct io_fixed_file *slot = io_fixed_file_slot(&ctx->file_table, index);
6295 return (struct file *) (slot->file_ptr & FFS_MASK);
6298 static void io_fixed_file_set(struct io_fixed_file *file_slot, struct file *file)
6300 unsigned long file_ptr = (unsigned long) file;
6302 if (__io_file_supports_async(file, READ))
6303 file_ptr |= FFS_ASYNC_READ;
6304 if (__io_file_supports_async(file, WRITE))
6305 file_ptr |= FFS_ASYNC_WRITE;
6306 if (S_ISREG(file_inode(file)->i_mode))
6307 file_ptr |= FFS_ISREG;
6308 file_slot->file_ptr = file_ptr;
6311 static struct file *io_file_get(struct io_submit_state *state,
6312 struct io_kiocb *req, int fd, bool fixed)
6314 struct io_ring_ctx *ctx = req->ctx;
6318 unsigned long file_ptr;
6320 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6322 fd = array_index_nospec(fd, ctx->nr_user_files);
6323 file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr;
6324 file = (struct file *) (file_ptr & FFS_MASK);
6325 file_ptr &= ~FFS_MASK;
6326 /* mask in overlapping REQ_F and FFS bits */
6327 req->flags |= (file_ptr << REQ_F_ASYNC_READ_BIT);
6328 io_req_set_rsrc_node(req);
6330 trace_io_uring_file_get(ctx, fd);
6331 file = __io_file_get(state, fd);
6333 /* we don't allow fixed io_uring files */
6334 if (file && unlikely(file->f_op == &io_uring_fops))
6335 io_req_track_inflight(req);
6341 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6343 struct io_timeout_data *data = container_of(timer,
6344 struct io_timeout_data, timer);
6345 struct io_kiocb *prev, *req = data->req;
6346 struct io_ring_ctx *ctx = req->ctx;
6347 unsigned long flags;
6349 spin_lock_irqsave(&ctx->completion_lock, flags);
6350 prev = req->timeout.head;
6351 req->timeout.head = NULL;
6354 * We don't expect the list to be empty, that will only happen if we
6355 * race with the completion of the linked work.
6358 io_remove_next_linked(prev);
6359 if (!req_ref_inc_not_zero(prev))
6362 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6365 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6366 io_put_req_deferred(prev, 1);
6367 io_put_req_deferred(req, 1);
6369 io_req_complete_post(req, -ETIME, 0);
6371 return HRTIMER_NORESTART;
6374 static void io_queue_linked_timeout(struct io_kiocb *req)
6376 struct io_ring_ctx *ctx = req->ctx;
6378 spin_lock_irq(&ctx->completion_lock);
6380 * If the back reference is NULL, then our linked request finished
6381 * before we got a chance to setup the timer
6383 if (req->timeout.head) {
6384 struct io_timeout_data *data = req->async_data;
6386 data->timer.function = io_link_timeout_fn;
6387 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6390 spin_unlock_irq(&ctx->completion_lock);
6391 /* drop submission reference */
6395 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6397 struct io_kiocb *nxt = req->link;
6399 if (!nxt || (req->flags & REQ_F_LINK_TIMEOUT) ||
6400 nxt->opcode != IORING_OP_LINK_TIMEOUT)
6403 nxt->timeout.head = req;
6404 nxt->flags |= REQ_F_LTIMEOUT_ACTIVE;
6405 req->flags |= REQ_F_LINK_TIMEOUT;
6409 static void __io_queue_sqe(struct io_kiocb *req)
6411 struct io_kiocb *linked_timeout = io_prep_linked_timeout(req);
6414 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
6417 * We async punt it if the file wasn't marked NOWAIT, or if the file
6418 * doesn't support non-blocking read/write attempts
6421 /* drop submission reference */
6422 if (req->flags & REQ_F_COMPLETE_INLINE) {
6423 struct io_ring_ctx *ctx = req->ctx;
6424 struct io_comp_state *cs = &ctx->submit_state.comp;
6426 cs->reqs[cs->nr++] = req;
6427 if (cs->nr == ARRAY_SIZE(cs->reqs))
6428 io_submit_flush_completions(cs, ctx);
6432 } else if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6433 if (!io_arm_poll_handler(req)) {
6435 * Queued up for async execution, worker will release
6436 * submit reference when the iocb is actually submitted.
6438 io_queue_async_work(req);
6441 io_req_complete_failed(req, ret);
6444 io_queue_linked_timeout(linked_timeout);
6447 static inline void io_queue_sqe(struct io_kiocb *req)
6449 if (unlikely(req->ctx->drain_used) && io_drain_req(req))
6452 if (likely(!(req->flags & REQ_F_FORCE_ASYNC))) {
6453 __io_queue_sqe(req);
6455 int ret = io_req_prep_async(req);
6458 io_req_complete_failed(req, ret);
6460 io_queue_async_work(req);
6465 * Check SQE restrictions (opcode and flags).
6467 * Returns 'true' if SQE is allowed, 'false' otherwise.
6469 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6470 struct io_kiocb *req,
6471 unsigned int sqe_flags)
6473 if (!ctx->restricted)
6476 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6479 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6480 ctx->restrictions.sqe_flags_required)
6483 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6484 ctx->restrictions.sqe_flags_required))
6490 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6491 const struct io_uring_sqe *sqe)
6493 struct io_submit_state *state;
6494 unsigned int sqe_flags;
6495 int personality, ret = 0;
6497 req->opcode = READ_ONCE(sqe->opcode);
6498 /* same numerical values with corresponding REQ_F_*, safe to copy */
6499 req->flags = sqe_flags = READ_ONCE(sqe->flags);
6500 req->user_data = READ_ONCE(sqe->user_data);
6501 req->async_data = NULL;
6505 req->fixed_rsrc_refs = NULL;
6506 /* one is dropped after submission, the other at completion */
6507 atomic_set(&req->refs, 2);
6508 req->task = current;
6510 req->work.creds = NULL;
6512 /* enforce forwards compatibility on users */
6513 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6515 if (unlikely(req->opcode >= IORING_OP_LAST))
6517 if (unlikely(!io_check_restriction(ctx, req, sqe_flags)))
6520 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6521 !io_op_defs[req->opcode].buffer_select)
6524 personality = READ_ONCE(sqe->personality);
6526 req->work.creds = xa_load(&ctx->personalities, personality);
6527 if (!req->work.creds)
6529 get_cred(req->work.creds);
6531 state = &ctx->submit_state;
6534 * Plug now if we have more than 1 IO left after this, and the target
6535 * is potentially a read/write to block based storage.
6537 if (!state->plug_started && state->ios_left > 1 &&
6538 io_op_defs[req->opcode].plug) {
6539 blk_start_plug(&state->plug);
6540 state->plug_started = true;
6543 if (io_op_defs[req->opcode].needs_file) {
6544 bool fixed = req->flags & REQ_F_FIXED_FILE;
6546 req->file = io_file_get(state, req, READ_ONCE(sqe->fd), fixed);
6547 if (unlikely(!req->file))
6555 static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
6556 const struct io_uring_sqe *sqe)
6558 struct io_submit_link *link = &ctx->submit_state.link;
6561 ret = io_init_req(ctx, req, sqe);
6562 if (unlikely(ret)) {
6565 /* fail even hard links since we don't submit */
6566 req_set_fail(link->head);
6567 io_req_complete_failed(link->head, -ECANCELED);
6570 io_req_complete_failed(req, ret);
6574 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
6575 ctx->drain_used = true;
6578 * Taking sequential execution of a link, draining both sides
6579 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6580 * requests in the link. So, it drains the head and the
6581 * next after the link request. The last one is done via
6582 * drain_next flag to persist the effect across calls.
6585 link->head->flags |= REQ_F_IO_DRAIN;
6586 ctx->drain_next = 1;
6590 ret = io_req_prep(req, sqe);
6594 /* don't need @sqe from now on */
6595 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6596 true, ctx->flags & IORING_SETUP_SQPOLL);
6599 * If we already have a head request, queue this one for async
6600 * submittal once the head completes. If we don't have a head but
6601 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6602 * submitted sync once the chain is complete. If none of those
6603 * conditions are true (normal request), then just queue it.
6606 struct io_kiocb *head = link->head;
6608 ret = io_req_prep_async(req);
6611 trace_io_uring_link(ctx, req, head);
6612 link->last->link = req;
6615 /* last request of a link, enqueue the link */
6616 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6621 if (unlikely(ctx->drain_next)) {
6622 req->flags |= REQ_F_IO_DRAIN;
6623 ctx->drain_next = 0;
6625 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6637 * Batched submission is done, ensure local IO is flushed out.
6639 static void io_submit_state_end(struct io_submit_state *state,
6640 struct io_ring_ctx *ctx)
6642 if (state->link.head)
6643 io_queue_sqe(state->link.head);
6645 io_submit_flush_completions(&state->comp, ctx);
6646 if (state->plug_started)
6647 blk_finish_plug(&state->plug);
6648 io_state_file_put(state);
6652 * Start submission side cache.
6654 static void io_submit_state_start(struct io_submit_state *state,
6655 unsigned int max_ios)
6657 state->plug_started = false;
6658 state->ios_left = max_ios;
6659 /* set only head, no need to init link_last in advance */
6660 state->link.head = NULL;
6663 static void io_commit_sqring(struct io_ring_ctx *ctx)
6665 struct io_rings *rings = ctx->rings;
6668 * Ensure any loads from the SQEs are done at this point,
6669 * since once we write the new head, the application could
6670 * write new data to them.
6672 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6676 * Fetch an sqe, if one is available. Note this returns a pointer to memory
6677 * that is mapped by userspace. This means that care needs to be taken to
6678 * ensure that reads are stable, as we cannot rely on userspace always
6679 * being a good citizen. If members of the sqe are validated and then later
6680 * used, it's important that those reads are done through READ_ONCE() to
6681 * prevent a re-load down the line.
6683 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6685 unsigned head, mask = ctx->sq_entries - 1;
6686 unsigned sq_idx = ctx->cached_sq_head++ & mask;
6689 * The cached sq head (or cq tail) serves two purposes:
6691 * 1) allows us to batch the cost of updating the user visible
6693 * 2) allows the kernel side to track the head on its own, even
6694 * though the application is the one updating it.
6696 head = READ_ONCE(ctx->sq_array[sq_idx]);
6697 if (likely(head < ctx->sq_entries))
6698 return &ctx->sq_sqes[head];
6700 /* drop invalid entries */
6702 WRITE_ONCE(ctx->rings->sq_dropped,
6703 READ_ONCE(ctx->rings->sq_dropped) + 1);
6707 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6709 struct io_uring_task *tctx;
6712 /* make sure SQ entry isn't read before tail */
6713 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6714 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6717 tctx = current->io_uring;
6718 tctx->cached_refs -= nr;
6719 if (unlikely(tctx->cached_refs < 0)) {
6720 unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR;
6722 percpu_counter_add(&tctx->inflight, refill);
6723 refcount_add(refill, ¤t->usage);
6724 tctx->cached_refs += refill;
6726 io_submit_state_start(&ctx->submit_state, nr);
6728 while (submitted < nr) {
6729 const struct io_uring_sqe *sqe;
6730 struct io_kiocb *req;
6732 req = io_alloc_req(ctx);
6733 if (unlikely(!req)) {
6735 submitted = -EAGAIN;
6738 sqe = io_get_sqe(ctx);
6739 if (unlikely(!sqe)) {
6740 kmem_cache_free(req_cachep, req);
6743 /* will complete beyond this point, count as submitted */
6745 if (io_submit_sqe(ctx, req, sqe))
6749 if (unlikely(submitted != nr)) {
6750 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6751 int unused = nr - ref_used;
6753 current->io_uring->cached_refs += unused;
6754 percpu_ref_put_many(&ctx->refs, unused);
6757 io_submit_state_end(&ctx->submit_state, ctx);
6758 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6759 io_commit_sqring(ctx);
6764 static inline bool io_sqd_events_pending(struct io_sq_data *sqd)
6766 return READ_ONCE(sqd->state);
6769 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6771 /* Tell userspace we may need a wakeup call */
6772 spin_lock_irq(&ctx->completion_lock);
6773 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6774 spin_unlock_irq(&ctx->completion_lock);
6777 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6779 spin_lock_irq(&ctx->completion_lock);
6780 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6781 spin_unlock_irq(&ctx->completion_lock);
6784 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
6786 unsigned int to_submit;
6789 to_submit = io_sqring_entries(ctx);
6790 /* if we're handling multiple rings, cap submit size for fairness */
6791 if (cap_entries && to_submit > 8)
6794 if (!list_empty(&ctx->iopoll_list) || to_submit) {
6795 unsigned nr_events = 0;
6797 mutex_lock(&ctx->uring_lock);
6798 if (!list_empty(&ctx->iopoll_list))
6799 io_do_iopoll(ctx, &nr_events, 0);
6802 * Don't submit if refs are dying, good for io_uring_register(),
6803 * but also it is relied upon by io_ring_exit_work()
6805 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)) &&
6806 !(ctx->flags & IORING_SETUP_R_DISABLED))
6807 ret = io_submit_sqes(ctx, to_submit);
6808 mutex_unlock(&ctx->uring_lock);
6810 if (to_submit && wq_has_sleeper(&ctx->sqo_sq_wait))
6811 wake_up(&ctx->sqo_sq_wait);
6817 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
6819 struct io_ring_ctx *ctx;
6820 unsigned sq_thread_idle = 0;
6822 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6823 sq_thread_idle = max(sq_thread_idle, ctx->sq_thread_idle);
6824 sqd->sq_thread_idle = sq_thread_idle;
6827 static bool io_sqd_handle_event(struct io_sq_data *sqd)
6829 bool did_sig = false;
6830 struct ksignal ksig;
6832 if (test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state) ||
6833 signal_pending(current)) {
6834 mutex_unlock(&sqd->lock);
6835 if (signal_pending(current))
6836 did_sig = get_signal(&ksig);
6838 mutex_lock(&sqd->lock);
6841 return did_sig || test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
6844 static int io_sq_thread(void *data)
6846 struct io_sq_data *sqd = data;
6847 struct io_ring_ctx *ctx;
6848 unsigned long timeout = 0;
6849 char buf[TASK_COMM_LEN];
6852 snprintf(buf, sizeof(buf), "iou-sqp-%d", sqd->task_pid);
6853 set_task_comm(current, buf);
6855 if (sqd->sq_cpu != -1)
6856 set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
6858 set_cpus_allowed_ptr(current, cpu_online_mask);
6859 current->flags |= PF_NO_SETAFFINITY;
6861 mutex_lock(&sqd->lock);
6864 bool cap_entries, sqt_spin, needs_sched;
6866 if (io_sqd_events_pending(sqd) || signal_pending(current)) {
6867 if (io_sqd_handle_event(sqd))
6869 timeout = jiffies + sqd->sq_thread_idle;
6874 cap_entries = !list_is_singular(&sqd->ctx_list);
6875 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6876 const struct cred *creds = NULL;
6878 if (ctx->sq_creds != current_cred())
6879 creds = override_creds(ctx->sq_creds);
6880 ret = __io_sq_thread(ctx, cap_entries);
6882 revert_creds(creds);
6883 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
6887 if (sqt_spin || !time_after(jiffies, timeout)) {
6891 timeout = jiffies + sqd->sq_thread_idle;
6895 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
6896 if (!io_sqd_events_pending(sqd)) {
6898 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6899 io_ring_set_wakeup_flag(ctx);
6901 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6902 !list_empty_careful(&ctx->iopoll_list)) {
6903 needs_sched = false;
6906 if (io_sqring_entries(ctx)) {
6907 needs_sched = false;
6913 mutex_unlock(&sqd->lock);
6915 mutex_lock(&sqd->lock);
6917 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6918 io_ring_clear_wakeup_flag(ctx);
6921 finish_wait(&sqd->wait, &wait);
6922 timeout = jiffies + sqd->sq_thread_idle;
6925 io_uring_cancel_generic(true, sqd);
6927 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6928 io_ring_set_wakeup_flag(ctx);
6930 mutex_unlock(&sqd->lock);
6932 complete(&sqd->exited);
6936 struct io_wait_queue {
6937 struct wait_queue_entry wq;
6938 struct io_ring_ctx *ctx;
6940 unsigned nr_timeouts;
6943 static inline bool io_should_wake(struct io_wait_queue *iowq)
6945 struct io_ring_ctx *ctx = iowq->ctx;
6948 * Wake up if we have enough events, or if a timeout occurred since we
6949 * started waiting. For timeouts, we always want to return to userspace,
6950 * regardless of event count.
6952 return io_cqring_events(ctx) >= iowq->to_wait ||
6953 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6956 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6957 int wake_flags, void *key)
6959 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6963 * Cannot safely flush overflowed CQEs from here, ensure we wake up
6964 * the task, and the next invocation will do it.
6966 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->check_cq_overflow))
6967 return autoremove_wake_function(curr, mode, wake_flags, key);
6971 static int io_run_task_work_sig(void)
6973 if (io_run_task_work())
6975 if (!signal_pending(current))
6977 if (test_thread_flag(TIF_NOTIFY_SIGNAL))
6978 return -ERESTARTSYS;
6982 /* when returns >0, the caller should retry */
6983 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
6984 struct io_wait_queue *iowq,
6985 signed long *timeout)
6989 /* make sure we run task_work before checking for signals */
6990 ret = io_run_task_work_sig();
6991 if (ret || io_should_wake(iowq))
6993 /* let the caller flush overflows, retry */
6994 if (test_bit(0, &ctx->check_cq_overflow))
6997 *timeout = schedule_timeout(*timeout);
6998 return !*timeout ? -ETIME : 1;
7002 * Wait until events become available, if we don't already have some. The
7003 * application must reap them itself, as they reside on the shared cq ring.
7005 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
7006 const sigset_t __user *sig, size_t sigsz,
7007 struct __kernel_timespec __user *uts)
7009 struct io_wait_queue iowq = {
7012 .func = io_wake_function,
7013 .entry = LIST_HEAD_INIT(iowq.wq.entry),
7016 .to_wait = min_events,
7018 struct io_rings *rings = ctx->rings;
7019 signed long timeout = MAX_SCHEDULE_TIMEOUT;
7023 io_cqring_overflow_flush(ctx, false);
7024 if (io_cqring_events(ctx) >= min_events)
7026 if (!io_run_task_work())
7031 #ifdef CONFIG_COMPAT
7032 if (in_compat_syscall())
7033 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
7037 ret = set_user_sigmask(sig, sigsz);
7044 struct timespec64 ts;
7046 if (get_timespec64(&ts, uts))
7048 timeout = timespec64_to_jiffies(&ts);
7051 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
7052 trace_io_uring_cqring_wait(ctx, min_events);
7054 /* if we can't even flush overflow, don't wait for more */
7055 if (!io_cqring_overflow_flush(ctx, false)) {
7059 prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq,
7060 TASK_INTERRUPTIBLE);
7061 ret = io_cqring_wait_schedule(ctx, &iowq, &timeout);
7062 finish_wait(&ctx->cq_wait, &iowq.wq);
7066 restore_saved_sigmask_unless(ret == -EINTR);
7068 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
7071 static void io_free_page_table(void **table, size_t size)
7073 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7075 for (i = 0; i < nr_tables; i++)
7080 static void **io_alloc_page_table(size_t size)
7082 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE);
7083 size_t init_size = size;
7086 table = kcalloc(nr_tables, sizeof(*table), GFP_KERNEL);
7090 for (i = 0; i < nr_tables; i++) {
7091 unsigned int this_size = min(size, PAGE_SIZE);
7093 table[i] = kzalloc(this_size, GFP_KERNEL);
7095 io_free_page_table(table, init_size);
7103 static inline void io_rsrc_ref_lock(struct io_ring_ctx *ctx)
7105 spin_lock_bh(&ctx->rsrc_ref_lock);
7108 static inline void io_rsrc_ref_unlock(struct io_ring_ctx *ctx)
7110 spin_unlock_bh(&ctx->rsrc_ref_lock);
7113 static void io_rsrc_node_destroy(struct io_rsrc_node *ref_node)
7115 percpu_ref_exit(&ref_node->refs);
7119 static void io_rsrc_node_switch(struct io_ring_ctx *ctx,
7120 struct io_rsrc_data *data_to_kill)
7122 WARN_ON_ONCE(!ctx->rsrc_backup_node);
7123 WARN_ON_ONCE(data_to_kill && !ctx->rsrc_node);
7126 struct io_rsrc_node *rsrc_node = ctx->rsrc_node;
7128 rsrc_node->rsrc_data = data_to_kill;
7129 io_rsrc_ref_lock(ctx);
7130 list_add_tail(&rsrc_node->node, &ctx->rsrc_ref_list);
7131 io_rsrc_ref_unlock(ctx);
7133 atomic_inc(&data_to_kill->refs);
7134 percpu_ref_kill(&rsrc_node->refs);
7135 ctx->rsrc_node = NULL;
7138 if (!ctx->rsrc_node) {
7139 ctx->rsrc_node = ctx->rsrc_backup_node;
7140 ctx->rsrc_backup_node = NULL;
7144 static int io_rsrc_node_switch_start(struct io_ring_ctx *ctx)
7146 if (ctx->rsrc_backup_node)
7148 ctx->rsrc_backup_node = io_rsrc_node_alloc(ctx);
7149 return ctx->rsrc_backup_node ? 0 : -ENOMEM;
7152 static int io_rsrc_ref_quiesce(struct io_rsrc_data *data, struct io_ring_ctx *ctx)
7156 /* As we may drop ->uring_lock, other task may have started quiesce */
7160 data->quiesce = true;
7162 ret = io_rsrc_node_switch_start(ctx);
7165 io_rsrc_node_switch(ctx, data);
7167 /* kill initial ref, already quiesced if zero */
7168 if (atomic_dec_and_test(&data->refs))
7170 flush_delayed_work(&ctx->rsrc_put_work);
7171 ret = wait_for_completion_interruptible(&data->done);
7175 atomic_inc(&data->refs);
7176 /* wait for all works potentially completing data->done */
7177 flush_delayed_work(&ctx->rsrc_put_work);
7178 reinit_completion(&data->done);
7180 mutex_unlock(&ctx->uring_lock);
7181 ret = io_run_task_work_sig();
7182 mutex_lock(&ctx->uring_lock);
7184 data->quiesce = false;
7189 static u64 *io_get_tag_slot(struct io_rsrc_data *data, unsigned int idx)
7191 unsigned int off = idx & IO_RSRC_TAG_TABLE_MASK;
7192 unsigned int table_idx = idx >> IO_RSRC_TAG_TABLE_SHIFT;
7194 return &data->tags[table_idx][off];
7197 static void io_rsrc_data_free(struct io_rsrc_data *data)
7199 size_t size = data->nr * sizeof(data->tags[0][0]);
7202 io_free_page_table((void **)data->tags, size);
7206 static int io_rsrc_data_alloc(struct io_ring_ctx *ctx, rsrc_put_fn *do_put,
7207 u64 __user *utags, unsigned nr,
7208 struct io_rsrc_data **pdata)
7210 struct io_rsrc_data *data;
7214 data = kzalloc(sizeof(*data), GFP_KERNEL);
7217 data->tags = (u64 **)io_alloc_page_table(nr * sizeof(data->tags[0][0]));
7225 data->do_put = do_put;
7228 for (i = 0; i < nr; i++) {
7229 u64 *tag_slot = io_get_tag_slot(data, i);
7231 if (copy_from_user(tag_slot, &utags[i],
7237 atomic_set(&data->refs, 1);
7238 init_completion(&data->done);
7242 io_rsrc_data_free(data);
7246 static bool io_alloc_file_tables(struct io_file_table *table, unsigned nr_files)
7248 size_t size = nr_files * sizeof(struct io_fixed_file);
7250 table->files = (struct io_fixed_file **)io_alloc_page_table(size);
7251 return !!table->files;
7254 static void io_free_file_tables(struct io_file_table *table, unsigned nr_files)
7256 size_t size = nr_files * sizeof(struct io_fixed_file);
7258 io_free_page_table((void **)table->files, size);
7259 table->files = NULL;
7262 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
7264 #if defined(CONFIG_UNIX)
7265 if (ctx->ring_sock) {
7266 struct sock *sock = ctx->ring_sock->sk;
7267 struct sk_buff *skb;
7269 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
7275 for (i = 0; i < ctx->nr_user_files; i++) {
7278 file = io_file_from_index(ctx, i);
7283 io_free_file_tables(&ctx->file_table, ctx->nr_user_files);
7284 io_rsrc_data_free(ctx->file_data);
7285 ctx->file_data = NULL;
7286 ctx->nr_user_files = 0;
7289 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7293 if (!ctx->file_data)
7295 ret = io_rsrc_ref_quiesce(ctx->file_data, ctx);
7297 __io_sqe_files_unregister(ctx);
7301 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7302 __releases(&sqd->lock)
7304 WARN_ON_ONCE(sqd->thread == current);
7307 * Do the dance but not conditional clear_bit() because it'd race with
7308 * other threads incrementing park_pending and setting the bit.
7310 clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7311 if (atomic_dec_return(&sqd->park_pending))
7312 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7313 mutex_unlock(&sqd->lock);
7316 static void io_sq_thread_park(struct io_sq_data *sqd)
7317 __acquires(&sqd->lock)
7319 WARN_ON_ONCE(sqd->thread == current);
7321 atomic_inc(&sqd->park_pending);
7322 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7323 mutex_lock(&sqd->lock);
7325 wake_up_process(sqd->thread);
7328 static void io_sq_thread_stop(struct io_sq_data *sqd)
7330 WARN_ON_ONCE(sqd->thread == current);
7331 WARN_ON_ONCE(test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state));
7333 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7334 mutex_lock(&sqd->lock);
7336 wake_up_process(sqd->thread);
7337 mutex_unlock(&sqd->lock);
7338 wait_for_completion(&sqd->exited);
7341 static void io_put_sq_data(struct io_sq_data *sqd)
7343 if (refcount_dec_and_test(&sqd->refs)) {
7344 WARN_ON_ONCE(atomic_read(&sqd->park_pending));
7346 io_sq_thread_stop(sqd);
7351 static void io_sq_thread_finish(struct io_ring_ctx *ctx)
7353 struct io_sq_data *sqd = ctx->sq_data;
7356 io_sq_thread_park(sqd);
7357 list_del_init(&ctx->sqd_list);
7358 io_sqd_update_thread_idle(sqd);
7359 io_sq_thread_unpark(sqd);
7361 io_put_sq_data(sqd);
7362 ctx->sq_data = NULL;
7366 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7368 struct io_ring_ctx *ctx_attach;
7369 struct io_sq_data *sqd;
7372 f = fdget(p->wq_fd);
7374 return ERR_PTR(-ENXIO);
7375 if (f.file->f_op != &io_uring_fops) {
7377 return ERR_PTR(-EINVAL);
7380 ctx_attach = f.file->private_data;
7381 sqd = ctx_attach->sq_data;
7384 return ERR_PTR(-EINVAL);
7386 if (sqd->task_tgid != current->tgid) {
7388 return ERR_PTR(-EPERM);
7391 refcount_inc(&sqd->refs);
7396 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p,
7399 struct io_sq_data *sqd;
7402 if (p->flags & IORING_SETUP_ATTACH_WQ) {
7403 sqd = io_attach_sq_data(p);
7408 /* fall through for EPERM case, setup new sqd/task */
7409 if (PTR_ERR(sqd) != -EPERM)
7413 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7415 return ERR_PTR(-ENOMEM);
7417 atomic_set(&sqd->park_pending, 0);
7418 refcount_set(&sqd->refs, 1);
7419 INIT_LIST_HEAD(&sqd->ctx_list);
7420 mutex_init(&sqd->lock);
7421 init_waitqueue_head(&sqd->wait);
7422 init_completion(&sqd->exited);
7426 #if defined(CONFIG_UNIX)
7428 * Ensure the UNIX gc is aware of our file set, so we are certain that
7429 * the io_uring can be safely unregistered on process exit, even if we have
7430 * loops in the file referencing.
7432 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7434 struct sock *sk = ctx->ring_sock->sk;
7435 struct scm_fp_list *fpl;
7436 struct sk_buff *skb;
7439 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7443 skb = alloc_skb(0, GFP_KERNEL);
7452 fpl->user = get_uid(current_user());
7453 for (i = 0; i < nr; i++) {
7454 struct file *file = io_file_from_index(ctx, i + offset);
7458 fpl->fp[nr_files] = get_file(file);
7459 unix_inflight(fpl->user, fpl->fp[nr_files]);
7464 fpl->max = SCM_MAX_FD;
7465 fpl->count = nr_files;
7466 UNIXCB(skb).fp = fpl;
7467 skb->destructor = unix_destruct_scm;
7468 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7469 skb_queue_head(&sk->sk_receive_queue, skb);
7471 for (i = 0; i < nr_files; i++)
7482 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7483 * causes regular reference counting to break down. We rely on the UNIX
7484 * garbage collection to take care of this problem for us.
7486 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7488 unsigned left, total;
7492 left = ctx->nr_user_files;
7494 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7496 ret = __io_sqe_files_scm(ctx, this_files, total);
7500 total += this_files;
7506 while (total < ctx->nr_user_files) {
7507 struct file *file = io_file_from_index(ctx, total);
7517 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7523 static void io_rsrc_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
7525 struct file *file = prsrc->file;
7526 #if defined(CONFIG_UNIX)
7527 struct sock *sock = ctx->ring_sock->sk;
7528 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7529 struct sk_buff *skb;
7532 __skb_queue_head_init(&list);
7535 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7536 * remove this entry and rearrange the file array.
7538 skb = skb_dequeue(head);
7540 struct scm_fp_list *fp;
7542 fp = UNIXCB(skb).fp;
7543 for (i = 0; i < fp->count; i++) {
7546 if (fp->fp[i] != file)
7549 unix_notinflight(fp->user, fp->fp[i]);
7550 left = fp->count - 1 - i;
7552 memmove(&fp->fp[i], &fp->fp[i + 1],
7553 left * sizeof(struct file *));
7560 __skb_queue_tail(&list, skb);
7570 __skb_queue_tail(&list, skb);
7572 skb = skb_dequeue(head);
7575 if (skb_peek(&list)) {
7576 spin_lock_irq(&head->lock);
7577 while ((skb = __skb_dequeue(&list)) != NULL)
7578 __skb_queue_tail(head, skb);
7579 spin_unlock_irq(&head->lock);
7586 static void __io_rsrc_put_work(struct io_rsrc_node *ref_node)
7588 struct io_rsrc_data *rsrc_data = ref_node->rsrc_data;
7589 struct io_ring_ctx *ctx = rsrc_data->ctx;
7590 struct io_rsrc_put *prsrc, *tmp;
7592 list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
7593 list_del(&prsrc->list);
7596 bool lock_ring = ctx->flags & IORING_SETUP_IOPOLL;
7598 io_ring_submit_lock(ctx, lock_ring);
7599 spin_lock_irq(&ctx->completion_lock);
7600 io_cqring_fill_event(ctx, prsrc->tag, 0, 0);
7602 io_commit_cqring(ctx);
7603 spin_unlock_irq(&ctx->completion_lock);
7604 io_cqring_ev_posted(ctx);
7605 io_ring_submit_unlock(ctx, lock_ring);
7608 rsrc_data->do_put(ctx, prsrc);
7612 io_rsrc_node_destroy(ref_node);
7613 if (atomic_dec_and_test(&rsrc_data->refs))
7614 complete(&rsrc_data->done);
7617 static void io_rsrc_put_work(struct work_struct *work)
7619 struct io_ring_ctx *ctx;
7620 struct llist_node *node;
7622 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
7623 node = llist_del_all(&ctx->rsrc_put_llist);
7626 struct io_rsrc_node *ref_node;
7627 struct llist_node *next = node->next;
7629 ref_node = llist_entry(node, struct io_rsrc_node, llist);
7630 __io_rsrc_put_work(ref_node);
7635 static void io_rsrc_node_ref_zero(struct percpu_ref *ref)
7637 struct io_rsrc_node *node = container_of(ref, struct io_rsrc_node, refs);
7638 struct io_ring_ctx *ctx = node->rsrc_data->ctx;
7639 bool first_add = false;
7641 io_rsrc_ref_lock(ctx);
7644 while (!list_empty(&ctx->rsrc_ref_list)) {
7645 node = list_first_entry(&ctx->rsrc_ref_list,
7646 struct io_rsrc_node, node);
7647 /* recycle ref nodes in order */
7650 list_del(&node->node);
7651 first_add |= llist_add(&node->llist, &ctx->rsrc_put_llist);
7653 io_rsrc_ref_unlock(ctx);
7656 mod_delayed_work(system_wq, &ctx->rsrc_put_work, HZ);
7659 static struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx)
7661 struct io_rsrc_node *ref_node;
7663 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7667 if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
7672 INIT_LIST_HEAD(&ref_node->node);
7673 INIT_LIST_HEAD(&ref_node->rsrc_list);
7674 ref_node->done = false;
7678 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7679 unsigned nr_args, u64 __user *tags)
7681 __s32 __user *fds = (__s32 __user *) arg;
7690 if (nr_args > IORING_MAX_FIXED_FILES)
7692 ret = io_rsrc_node_switch_start(ctx);
7695 ret = io_rsrc_data_alloc(ctx, io_rsrc_file_put, tags, nr_args,
7701 if (!io_alloc_file_tables(&ctx->file_table, nr_args))
7704 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7705 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7709 /* allow sparse sets */
7712 if (unlikely(*io_get_tag_slot(ctx->file_data, i)))
7719 if (unlikely(!file))
7723 * Don't allow io_uring instances to be registered. If UNIX
7724 * isn't enabled, then this causes a reference cycle and this
7725 * instance can never get freed. If UNIX is enabled we'll
7726 * handle it just fine, but there's still no point in allowing
7727 * a ring fd as it doesn't support regular read/write anyway.
7729 if (file->f_op == &io_uring_fops) {
7733 io_fixed_file_set(io_fixed_file_slot(&ctx->file_table, i), file);
7736 ret = io_sqe_files_scm(ctx);
7738 __io_sqe_files_unregister(ctx);
7742 io_rsrc_node_switch(ctx, NULL);
7745 for (i = 0; i < ctx->nr_user_files; i++) {
7746 file = io_file_from_index(ctx, i);
7750 io_free_file_tables(&ctx->file_table, nr_args);
7751 ctx->nr_user_files = 0;
7753 io_rsrc_data_free(ctx->file_data);
7754 ctx->file_data = NULL;
7758 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7761 #if defined(CONFIG_UNIX)
7762 struct sock *sock = ctx->ring_sock->sk;
7763 struct sk_buff_head *head = &sock->sk_receive_queue;
7764 struct sk_buff *skb;
7767 * See if we can merge this file into an existing skb SCM_RIGHTS
7768 * file set. If there's no room, fall back to allocating a new skb
7769 * and filling it in.
7771 spin_lock_irq(&head->lock);
7772 skb = skb_peek(head);
7774 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7776 if (fpl->count < SCM_MAX_FD) {
7777 __skb_unlink(skb, head);
7778 spin_unlock_irq(&head->lock);
7779 fpl->fp[fpl->count] = get_file(file);
7780 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7782 spin_lock_irq(&head->lock);
7783 __skb_queue_head(head, skb);
7788 spin_unlock_irq(&head->lock);
7795 return __io_sqe_files_scm(ctx, 1, index);
7801 static int io_queue_rsrc_removal(struct io_rsrc_data *data, unsigned idx,
7802 struct io_rsrc_node *node, void *rsrc)
7804 struct io_rsrc_put *prsrc;
7806 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
7810 prsrc->tag = *io_get_tag_slot(data, idx);
7812 list_add(&prsrc->list, &node->rsrc_list);
7816 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7817 struct io_uring_rsrc_update2 *up,
7820 u64 __user *tags = u64_to_user_ptr(up->tags);
7821 __s32 __user *fds = u64_to_user_ptr(up->data);
7822 struct io_rsrc_data *data = ctx->file_data;
7823 struct io_fixed_file *file_slot;
7827 bool needs_switch = false;
7829 if (!ctx->file_data)
7831 if (up->offset + nr_args > ctx->nr_user_files)
7834 for (done = 0; done < nr_args; done++) {
7837 if ((tags && copy_from_user(&tag, &tags[done], sizeof(tag))) ||
7838 copy_from_user(&fd, &fds[done], sizeof(fd))) {
7842 if ((fd == IORING_REGISTER_FILES_SKIP || fd == -1) && tag) {
7846 if (fd == IORING_REGISTER_FILES_SKIP)
7849 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
7850 file_slot = io_fixed_file_slot(&ctx->file_table, i);
7852 if (file_slot->file_ptr) {
7853 file = (struct file *)(file_slot->file_ptr & FFS_MASK);
7854 err = io_queue_rsrc_removal(data, up->offset + done,
7855 ctx->rsrc_node, file);
7858 file_slot->file_ptr = 0;
7859 needs_switch = true;
7868 * Don't allow io_uring instances to be registered. If
7869 * UNIX isn't enabled, then this causes a reference
7870 * cycle and this instance can never get freed. If UNIX
7871 * is enabled we'll handle it just fine, but there's
7872 * still no point in allowing a ring fd as it doesn't
7873 * support regular read/write anyway.
7875 if (file->f_op == &io_uring_fops) {
7880 *io_get_tag_slot(data, up->offset + done) = tag;
7881 io_fixed_file_set(file_slot, file);
7882 err = io_sqe_file_register(ctx, file, i);
7884 file_slot->file_ptr = 0;
7892 io_rsrc_node_switch(ctx, data);
7893 return done ? done : err;
7896 static struct io_wq_work *io_free_work(struct io_wq_work *work)
7898 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7900 req = io_put_req_find_next(req);
7901 return req ? &req->work : NULL;
7904 static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx,
7905 struct task_struct *task)
7907 struct io_wq_hash *hash;
7908 struct io_wq_data data;
7909 unsigned int concurrency;
7911 hash = ctx->hash_map;
7913 hash = kzalloc(sizeof(*hash), GFP_KERNEL);
7915 return ERR_PTR(-ENOMEM);
7916 refcount_set(&hash->refs, 1);
7917 init_waitqueue_head(&hash->wait);
7918 ctx->hash_map = hash;
7923 data.free_work = io_free_work;
7924 data.do_work = io_wq_submit_work;
7926 /* Do QD, or 4 * CPUS, whatever is smallest */
7927 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7929 return io_wq_create(concurrency, &data);
7932 static int io_uring_alloc_task_context(struct task_struct *task,
7933 struct io_ring_ctx *ctx)
7935 struct io_uring_task *tctx;
7938 tctx = kzalloc(sizeof(*tctx), GFP_KERNEL);
7939 if (unlikely(!tctx))
7942 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
7943 if (unlikely(ret)) {
7948 tctx->io_wq = io_init_wq_offload(ctx, task);
7949 if (IS_ERR(tctx->io_wq)) {
7950 ret = PTR_ERR(tctx->io_wq);
7951 percpu_counter_destroy(&tctx->inflight);
7957 init_waitqueue_head(&tctx->wait);
7958 atomic_set(&tctx->in_idle, 0);
7959 atomic_set(&tctx->inflight_tracked, 0);
7960 task->io_uring = tctx;
7961 spin_lock_init(&tctx->task_lock);
7962 INIT_WQ_LIST(&tctx->task_list);
7963 init_task_work(&tctx->task_work, tctx_task_work);
7967 void __io_uring_free(struct task_struct *tsk)
7969 struct io_uring_task *tctx = tsk->io_uring;
7971 WARN_ON_ONCE(!xa_empty(&tctx->xa));
7972 WARN_ON_ONCE(tctx->io_wq);
7973 WARN_ON_ONCE(tctx->cached_refs);
7975 percpu_counter_destroy(&tctx->inflight);
7977 tsk->io_uring = NULL;
7980 static int io_sq_offload_create(struct io_ring_ctx *ctx,
7981 struct io_uring_params *p)
7985 /* Retain compatibility with failing for an invalid attach attempt */
7986 if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
7987 IORING_SETUP_ATTACH_WQ) {
7990 f = fdget(p->wq_fd);
7994 if (f.file->f_op != &io_uring_fops)
7997 if (ctx->flags & IORING_SETUP_SQPOLL) {
7998 struct task_struct *tsk;
7999 struct io_sq_data *sqd;
8002 sqd = io_get_sq_data(p, &attached);
8008 ctx->sq_creds = get_current_cred();
8010 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
8011 if (!ctx->sq_thread_idle)
8012 ctx->sq_thread_idle = HZ;
8014 io_sq_thread_park(sqd);
8015 list_add(&ctx->sqd_list, &sqd->ctx_list);
8016 io_sqd_update_thread_idle(sqd);
8017 /* don't attach to a dying SQPOLL thread, would be racy */
8018 ret = (attached && !sqd->thread) ? -ENXIO : 0;
8019 io_sq_thread_unpark(sqd);
8026 if (p->flags & IORING_SETUP_SQ_AFF) {
8027 int cpu = p->sq_thread_cpu;
8030 if (cpu >= nr_cpu_ids || !cpu_online(cpu))
8037 sqd->task_pid = current->pid;
8038 sqd->task_tgid = current->tgid;
8039 tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE);
8046 ret = io_uring_alloc_task_context(tsk, ctx);
8047 wake_up_new_task(tsk);
8050 } else if (p->flags & IORING_SETUP_SQ_AFF) {
8051 /* Can't have SQ_AFF without SQPOLL */
8058 complete(&ctx->sq_data->exited);
8060 io_sq_thread_finish(ctx);
8064 static inline void __io_unaccount_mem(struct user_struct *user,
8065 unsigned long nr_pages)
8067 atomic_long_sub(nr_pages, &user->locked_vm);
8070 static inline int __io_account_mem(struct user_struct *user,
8071 unsigned long nr_pages)
8073 unsigned long page_limit, cur_pages, new_pages;
8075 /* Don't allow more pages than we can safely lock */
8076 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
8079 cur_pages = atomic_long_read(&user->locked_vm);
8080 new_pages = cur_pages + nr_pages;
8081 if (new_pages > page_limit)
8083 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
8084 new_pages) != cur_pages);
8089 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8092 __io_unaccount_mem(ctx->user, nr_pages);
8094 if (ctx->mm_account)
8095 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
8098 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8103 ret = __io_account_mem(ctx->user, nr_pages);
8108 if (ctx->mm_account)
8109 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
8114 static void io_mem_free(void *ptr)
8121 page = virt_to_head_page(ptr);
8122 if (put_page_testzero(page))
8123 free_compound_page(page);
8126 static void *io_mem_alloc(size_t size)
8128 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
8129 __GFP_NORETRY | __GFP_ACCOUNT;
8131 return (void *) __get_free_pages(gfp_flags, get_order(size));
8134 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8137 struct io_rings *rings;
8138 size_t off, sq_array_size;
8140 off = struct_size(rings, cqes, cq_entries);
8141 if (off == SIZE_MAX)
8145 off = ALIGN(off, SMP_CACHE_BYTES);
8153 sq_array_size = array_size(sizeof(u32), sq_entries);
8154 if (sq_array_size == SIZE_MAX)
8157 if (check_add_overflow(off, sq_array_size, &off))
8163 static void io_buffer_unmap(struct io_ring_ctx *ctx, struct io_mapped_ubuf **slot)
8165 struct io_mapped_ubuf *imu = *slot;
8168 if (imu != ctx->dummy_ubuf) {
8169 for (i = 0; i < imu->nr_bvecs; i++)
8170 unpin_user_page(imu->bvec[i].bv_page);
8171 if (imu->acct_pages)
8172 io_unaccount_mem(ctx, imu->acct_pages);
8178 static void io_rsrc_buf_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
8180 io_buffer_unmap(ctx, &prsrc->buf);
8184 static void __io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8188 for (i = 0; i < ctx->nr_user_bufs; i++)
8189 io_buffer_unmap(ctx, &ctx->user_bufs[i]);
8190 kfree(ctx->user_bufs);
8191 io_rsrc_data_free(ctx->buf_data);
8192 ctx->user_bufs = NULL;
8193 ctx->buf_data = NULL;
8194 ctx->nr_user_bufs = 0;
8197 static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8204 ret = io_rsrc_ref_quiesce(ctx->buf_data, ctx);
8206 __io_sqe_buffers_unregister(ctx);
8210 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8211 void __user *arg, unsigned index)
8213 struct iovec __user *src;
8215 #ifdef CONFIG_COMPAT
8217 struct compat_iovec __user *ciovs;
8218 struct compat_iovec ciov;
8220 ciovs = (struct compat_iovec __user *) arg;
8221 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8224 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8225 dst->iov_len = ciov.iov_len;
8229 src = (struct iovec __user *) arg;
8230 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8236 * Not super efficient, but this is just a registration time. And we do cache
8237 * the last compound head, so generally we'll only do a full search if we don't
8240 * We check if the given compound head page has already been accounted, to
8241 * avoid double accounting it. This allows us to account the full size of the
8242 * page, not just the constituent pages of a huge page.
8244 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8245 int nr_pages, struct page *hpage)
8249 /* check current page array */
8250 for (i = 0; i < nr_pages; i++) {
8251 if (!PageCompound(pages[i]))
8253 if (compound_head(pages[i]) == hpage)
8257 /* check previously registered pages */
8258 for (i = 0; i < ctx->nr_user_bufs; i++) {
8259 struct io_mapped_ubuf *imu = ctx->user_bufs[i];
8261 for (j = 0; j < imu->nr_bvecs; j++) {
8262 if (!PageCompound(imu->bvec[j].bv_page))
8264 if (compound_head(imu->bvec[j].bv_page) == hpage)
8272 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8273 int nr_pages, struct io_mapped_ubuf *imu,
8274 struct page **last_hpage)
8278 imu->acct_pages = 0;
8279 for (i = 0; i < nr_pages; i++) {
8280 if (!PageCompound(pages[i])) {
8285 hpage = compound_head(pages[i]);
8286 if (hpage == *last_hpage)
8288 *last_hpage = hpage;
8289 if (headpage_already_acct(ctx, pages, i, hpage))
8291 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8295 if (!imu->acct_pages)
8298 ret = io_account_mem(ctx, imu->acct_pages);
8300 imu->acct_pages = 0;
8304 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
8305 struct io_mapped_ubuf **pimu,
8306 struct page **last_hpage)
8308 struct io_mapped_ubuf *imu = NULL;
8309 struct vm_area_struct **vmas = NULL;
8310 struct page **pages = NULL;
8311 unsigned long off, start, end, ubuf;
8313 int ret, pret, nr_pages, i;
8315 if (!iov->iov_base) {
8316 *pimu = ctx->dummy_ubuf;
8320 ubuf = (unsigned long) iov->iov_base;
8321 end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8322 start = ubuf >> PAGE_SHIFT;
8323 nr_pages = end - start;
8328 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
8332 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
8337 imu = kvmalloc(struct_size(imu, bvec, nr_pages), GFP_KERNEL);
8342 mmap_read_lock(current->mm);
8343 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
8345 if (pret == nr_pages) {
8346 /* don't support file backed memory */
8347 for (i = 0; i < nr_pages; i++) {
8348 struct vm_area_struct *vma = vmas[i];
8350 if (vma_is_shmem(vma))
8353 !is_file_hugepages(vma->vm_file)) {
8359 ret = pret < 0 ? pret : -EFAULT;
8361 mmap_read_unlock(current->mm);
8364 * if we did partial map, or found file backed vmas,
8365 * release any pages we did get
8368 unpin_user_pages(pages, pret);
8372 ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
8374 unpin_user_pages(pages, pret);
8378 off = ubuf & ~PAGE_MASK;
8379 size = iov->iov_len;
8380 for (i = 0; i < nr_pages; i++) {
8383 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8384 imu->bvec[i].bv_page = pages[i];
8385 imu->bvec[i].bv_len = vec_len;
8386 imu->bvec[i].bv_offset = off;
8390 /* store original address for later verification */
8392 imu->ubuf_end = ubuf + iov->iov_len;
8393 imu->nr_bvecs = nr_pages;
8404 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
8406 ctx->user_bufs = kcalloc(nr_args, sizeof(*ctx->user_bufs), GFP_KERNEL);
8407 return ctx->user_bufs ? 0 : -ENOMEM;
8410 static int io_buffer_validate(struct iovec *iov)
8412 unsigned long tmp, acct_len = iov->iov_len + (PAGE_SIZE - 1);
8415 * Don't impose further limits on the size and buffer
8416 * constraints here, we'll -EINVAL later when IO is
8417 * submitted if they are wrong.
8420 return iov->iov_len ? -EFAULT : 0;
8424 /* arbitrary limit, but we need something */
8425 if (iov->iov_len > SZ_1G)
8428 if (check_add_overflow((unsigned long)iov->iov_base, acct_len, &tmp))
8434 static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
8435 unsigned int nr_args, u64 __user *tags)
8437 struct page *last_hpage = NULL;
8438 struct io_rsrc_data *data;
8444 if (!nr_args || nr_args > IORING_MAX_REG_BUFFERS)
8446 ret = io_rsrc_node_switch_start(ctx);
8449 ret = io_rsrc_data_alloc(ctx, io_rsrc_buf_put, tags, nr_args, &data);
8452 ret = io_buffers_map_alloc(ctx, nr_args);
8454 io_rsrc_data_free(data);
8458 for (i = 0; i < nr_args; i++, ctx->nr_user_bufs++) {
8459 ret = io_copy_iov(ctx, &iov, arg, i);
8462 ret = io_buffer_validate(&iov);
8465 if (!iov.iov_base && *io_get_tag_slot(data, i)) {
8470 ret = io_sqe_buffer_register(ctx, &iov, &ctx->user_bufs[i],
8476 WARN_ON_ONCE(ctx->buf_data);
8478 ctx->buf_data = data;
8480 __io_sqe_buffers_unregister(ctx);
8482 io_rsrc_node_switch(ctx, NULL);
8486 static int __io_sqe_buffers_update(struct io_ring_ctx *ctx,
8487 struct io_uring_rsrc_update2 *up,
8488 unsigned int nr_args)
8490 u64 __user *tags = u64_to_user_ptr(up->tags);
8491 struct iovec iov, __user *iovs = u64_to_user_ptr(up->data);
8492 struct page *last_hpage = NULL;
8493 bool needs_switch = false;
8499 if (up->offset + nr_args > ctx->nr_user_bufs)
8502 for (done = 0; done < nr_args; done++) {
8503 struct io_mapped_ubuf *imu;
8504 int offset = up->offset + done;
8507 err = io_copy_iov(ctx, &iov, iovs, done);
8510 if (tags && copy_from_user(&tag, &tags[done], sizeof(tag))) {
8514 err = io_buffer_validate(&iov);
8517 if (!iov.iov_base && tag) {
8521 err = io_sqe_buffer_register(ctx, &iov, &imu, &last_hpage);
8525 i = array_index_nospec(offset, ctx->nr_user_bufs);
8526 if (ctx->user_bufs[i] != ctx->dummy_ubuf) {
8527 err = io_queue_rsrc_removal(ctx->buf_data, offset,
8528 ctx->rsrc_node, ctx->user_bufs[i]);
8529 if (unlikely(err)) {
8530 io_buffer_unmap(ctx, &imu);
8533 ctx->user_bufs[i] = NULL;
8534 needs_switch = true;
8537 ctx->user_bufs[i] = imu;
8538 *io_get_tag_slot(ctx->buf_data, offset) = tag;
8542 io_rsrc_node_switch(ctx, ctx->buf_data);
8543 return done ? done : err;
8546 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8548 __s32 __user *fds = arg;
8554 if (copy_from_user(&fd, fds, sizeof(*fds)))
8557 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8558 if (IS_ERR(ctx->cq_ev_fd)) {
8559 int ret = PTR_ERR(ctx->cq_ev_fd);
8560 ctx->cq_ev_fd = NULL;
8567 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8569 if (ctx->cq_ev_fd) {
8570 eventfd_ctx_put(ctx->cq_ev_fd);
8571 ctx->cq_ev_fd = NULL;
8578 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8580 struct io_buffer *buf;
8581 unsigned long index;
8583 xa_for_each(&ctx->io_buffers, index, buf)
8584 __io_remove_buffers(ctx, buf, index, -1U);
8587 static void io_req_cache_free(struct list_head *list, struct task_struct *tsk)
8589 struct io_kiocb *req, *nxt;
8591 list_for_each_entry_safe(req, nxt, list, compl.list) {
8592 if (tsk && req->task != tsk)
8594 list_del(&req->compl.list);
8595 kmem_cache_free(req_cachep, req);
8599 static void io_req_caches_free(struct io_ring_ctx *ctx)
8601 struct io_submit_state *submit_state = &ctx->submit_state;
8602 struct io_comp_state *cs = &ctx->submit_state.comp;
8604 mutex_lock(&ctx->uring_lock);
8606 if (submit_state->free_reqs) {
8607 kmem_cache_free_bulk(req_cachep, submit_state->free_reqs,
8608 submit_state->reqs);
8609 submit_state->free_reqs = 0;
8612 io_flush_cached_locked_reqs(ctx, cs);
8613 io_req_cache_free(&cs->free_list, NULL);
8614 mutex_unlock(&ctx->uring_lock);
8617 static bool io_wait_rsrc_data(struct io_rsrc_data *data)
8621 if (!atomic_dec_and_test(&data->refs))
8622 wait_for_completion(&data->done);
8626 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8628 io_sq_thread_finish(ctx);
8630 if (ctx->mm_account) {
8631 mmdrop(ctx->mm_account);
8632 ctx->mm_account = NULL;
8635 mutex_lock(&ctx->uring_lock);
8636 if (io_wait_rsrc_data(ctx->buf_data))
8637 __io_sqe_buffers_unregister(ctx);
8638 if (io_wait_rsrc_data(ctx->file_data))
8639 __io_sqe_files_unregister(ctx);
8641 __io_cqring_overflow_flush(ctx, true);
8642 mutex_unlock(&ctx->uring_lock);
8643 io_eventfd_unregister(ctx);
8644 io_destroy_buffers(ctx);
8646 put_cred(ctx->sq_creds);
8648 /* there are no registered resources left, nobody uses it */
8650 io_rsrc_node_destroy(ctx->rsrc_node);
8651 if (ctx->rsrc_backup_node)
8652 io_rsrc_node_destroy(ctx->rsrc_backup_node);
8653 flush_delayed_work(&ctx->rsrc_put_work);
8655 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
8656 WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist));
8658 #if defined(CONFIG_UNIX)
8659 if (ctx->ring_sock) {
8660 ctx->ring_sock->file = NULL; /* so that iput() is called */
8661 sock_release(ctx->ring_sock);
8665 io_mem_free(ctx->rings);
8666 io_mem_free(ctx->sq_sqes);
8668 percpu_ref_exit(&ctx->refs);
8669 free_uid(ctx->user);
8670 io_req_caches_free(ctx);
8672 io_wq_put_hash(ctx->hash_map);
8673 kfree(ctx->cancel_hash);
8674 kfree(ctx->dummy_ubuf);
8678 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8680 struct io_ring_ctx *ctx = file->private_data;
8683 poll_wait(file, &ctx->poll_wait, wait);
8685 * synchronizes with barrier from wq_has_sleeper call in
8689 if (!io_sqring_full(ctx))
8690 mask |= EPOLLOUT | EPOLLWRNORM;
8693 * Don't flush cqring overflow list here, just do a simple check.
8694 * Otherwise there could possible be ABBA deadlock:
8697 * lock(&ctx->uring_lock);
8699 * lock(&ctx->uring_lock);
8702 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
8703 * pushs them to do the flush.
8705 if (io_cqring_events(ctx) || test_bit(0, &ctx->check_cq_overflow))
8706 mask |= EPOLLIN | EPOLLRDNORM;
8711 static int io_uring_fasync(int fd, struct file *file, int on)
8713 struct io_ring_ctx *ctx = file->private_data;
8715 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8718 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8720 const struct cred *creds;
8722 creds = xa_erase(&ctx->personalities, id);
8731 static inline bool io_run_ctx_fallback(struct io_ring_ctx *ctx)
8733 return io_run_task_work_head(&ctx->exit_task_work);
8736 struct io_tctx_exit {
8737 struct callback_head task_work;
8738 struct completion completion;
8739 struct io_ring_ctx *ctx;
8742 static void io_tctx_exit_cb(struct callback_head *cb)
8744 struct io_uring_task *tctx = current->io_uring;
8745 struct io_tctx_exit *work;
8747 work = container_of(cb, struct io_tctx_exit, task_work);
8749 * When @in_idle, we're in cancellation and it's racy to remove the
8750 * node. It'll be removed by the end of cancellation, just ignore it.
8752 if (!atomic_read(&tctx->in_idle))
8753 io_uring_del_tctx_node((unsigned long)work->ctx);
8754 complete(&work->completion);
8757 static bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
8759 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8761 return req->ctx == data;
8764 static void io_ring_exit_work(struct work_struct *work)
8766 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
8767 unsigned long timeout = jiffies + HZ * 60 * 5;
8768 struct io_tctx_exit exit;
8769 struct io_tctx_node *node;
8773 * If we're doing polled IO and end up having requests being
8774 * submitted async (out-of-line), then completions can come in while
8775 * we're waiting for refs to drop. We need to reap these manually,
8776 * as nobody else will be looking for them.
8779 io_uring_try_cancel_requests(ctx, NULL, true);
8781 struct io_sq_data *sqd = ctx->sq_data;
8782 struct task_struct *tsk;
8784 io_sq_thread_park(sqd);
8786 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
8787 io_wq_cancel_cb(tsk->io_uring->io_wq,
8788 io_cancel_ctx_cb, ctx, true);
8789 io_sq_thread_unpark(sqd);
8792 WARN_ON_ONCE(time_after(jiffies, timeout));
8793 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
8795 init_completion(&exit.completion);
8796 init_task_work(&exit.task_work, io_tctx_exit_cb);
8799 * Some may use context even when all refs and requests have been put,
8800 * and they are free to do so while still holding uring_lock or
8801 * completion_lock, see __io_req_task_submit(). Apart from other work,
8802 * this lock/unlock section also waits them to finish.
8804 mutex_lock(&ctx->uring_lock);
8805 while (!list_empty(&ctx->tctx_list)) {
8806 WARN_ON_ONCE(time_after(jiffies, timeout));
8808 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
8810 /* don't spin on a single task if cancellation failed */
8811 list_rotate_left(&ctx->tctx_list);
8812 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
8813 if (WARN_ON_ONCE(ret))
8815 wake_up_process(node->task);
8817 mutex_unlock(&ctx->uring_lock);
8818 wait_for_completion(&exit.completion);
8819 mutex_lock(&ctx->uring_lock);
8821 mutex_unlock(&ctx->uring_lock);
8822 spin_lock_irq(&ctx->completion_lock);
8823 spin_unlock_irq(&ctx->completion_lock);
8825 io_ring_ctx_free(ctx);
8828 /* Returns true if we found and killed one or more timeouts */
8829 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
8832 struct io_kiocb *req, *tmp;
8835 spin_lock_irq(&ctx->completion_lock);
8836 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
8837 if (io_match_task(req, tsk, cancel_all)) {
8838 io_kill_timeout(req, -ECANCELED);
8843 io_commit_cqring(ctx);
8844 spin_unlock_irq(&ctx->completion_lock);
8846 io_cqring_ev_posted(ctx);
8847 return canceled != 0;
8850 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8852 unsigned long index;
8853 struct creds *creds;
8855 mutex_lock(&ctx->uring_lock);
8856 percpu_ref_kill(&ctx->refs);
8858 __io_cqring_overflow_flush(ctx, true);
8859 xa_for_each(&ctx->personalities, index, creds)
8860 io_unregister_personality(ctx, index);
8861 mutex_unlock(&ctx->uring_lock);
8863 io_kill_timeouts(ctx, NULL, true);
8864 io_poll_remove_all(ctx, NULL, true);
8866 /* if we failed setting up the ctx, we might not have any rings */
8867 io_iopoll_try_reap_events(ctx);
8869 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8871 * Use system_unbound_wq to avoid spawning tons of event kworkers
8872 * if we're exiting a ton of rings at the same time. It just adds
8873 * noise and overhead, there's no discernable change in runtime
8874 * over using system_wq.
8876 queue_work(system_unbound_wq, &ctx->exit_work);
8879 static int io_uring_release(struct inode *inode, struct file *file)
8881 struct io_ring_ctx *ctx = file->private_data;
8883 file->private_data = NULL;
8884 io_ring_ctx_wait_and_kill(ctx);
8888 struct io_task_cancel {
8889 struct task_struct *task;
8893 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8895 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8896 struct io_task_cancel *cancel = data;
8899 if (!cancel->all && (req->flags & REQ_F_LINK_TIMEOUT)) {
8900 unsigned long flags;
8901 struct io_ring_ctx *ctx = req->ctx;
8903 /* protect against races with linked timeouts */
8904 spin_lock_irqsave(&ctx->completion_lock, flags);
8905 ret = io_match_task(req, cancel->task, cancel->all);
8906 spin_unlock_irqrestore(&ctx->completion_lock, flags);
8908 ret = io_match_task(req, cancel->task, cancel->all);
8913 static bool io_cancel_defer_files(struct io_ring_ctx *ctx,
8914 struct task_struct *task, bool cancel_all)
8916 struct io_defer_entry *de;
8919 spin_lock_irq(&ctx->completion_lock);
8920 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8921 if (io_match_task(de->req, task, cancel_all)) {
8922 list_cut_position(&list, &ctx->defer_list, &de->list);
8926 spin_unlock_irq(&ctx->completion_lock);
8927 if (list_empty(&list))
8930 while (!list_empty(&list)) {
8931 de = list_first_entry(&list, struct io_defer_entry, list);
8932 list_del_init(&de->list);
8933 io_req_complete_failed(de->req, -ECANCELED);
8939 static bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
8941 struct io_tctx_node *node;
8942 enum io_wq_cancel cret;
8945 mutex_lock(&ctx->uring_lock);
8946 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
8947 struct io_uring_task *tctx = node->task->io_uring;
8950 * io_wq will stay alive while we hold uring_lock, because it's
8951 * killed after ctx nodes, which requires to take the lock.
8953 if (!tctx || !tctx->io_wq)
8955 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
8956 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8958 mutex_unlock(&ctx->uring_lock);
8963 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
8964 struct task_struct *task,
8967 struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
8968 struct io_uring_task *tctx = task ? task->io_uring : NULL;
8971 enum io_wq_cancel cret;
8975 ret |= io_uring_try_cancel_iowq(ctx);
8976 } else if (tctx && tctx->io_wq) {
8978 * Cancels requests of all rings, not only @ctx, but
8979 * it's fine as the task is in exit/exec.
8981 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
8983 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8986 /* SQPOLL thread does its own polling */
8987 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
8988 (ctx->sq_data && ctx->sq_data->thread == current)) {
8989 while (!list_empty_careful(&ctx->iopoll_list)) {
8990 io_iopoll_try_reap_events(ctx);
8995 ret |= io_cancel_defer_files(ctx, task, cancel_all);
8996 ret |= io_poll_remove_all(ctx, task, cancel_all);
8997 ret |= io_kill_timeouts(ctx, task, cancel_all);
8998 ret |= io_run_task_work();
8999 ret |= io_run_ctx_fallback(ctx);
9006 static int __io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9008 struct io_uring_task *tctx = current->io_uring;
9009 struct io_tctx_node *node;
9012 if (unlikely(!tctx)) {
9013 ret = io_uring_alloc_task_context(current, ctx);
9016 tctx = current->io_uring;
9018 if (!xa_load(&tctx->xa, (unsigned long)ctx)) {
9019 node = kmalloc(sizeof(*node), GFP_KERNEL);
9023 node->task = current;
9025 ret = xa_err(xa_store(&tctx->xa, (unsigned long)ctx,
9032 mutex_lock(&ctx->uring_lock);
9033 list_add(&node->ctx_node, &ctx->tctx_list);
9034 mutex_unlock(&ctx->uring_lock);
9041 * Note that this task has used io_uring. We use it for cancelation purposes.
9043 static inline int io_uring_add_tctx_node(struct io_ring_ctx *ctx)
9045 struct io_uring_task *tctx = current->io_uring;
9047 if (likely(tctx && tctx->last == ctx))
9049 return __io_uring_add_tctx_node(ctx);
9053 * Remove this io_uring_file -> task mapping.
9055 static void io_uring_del_tctx_node(unsigned long index)
9057 struct io_uring_task *tctx = current->io_uring;
9058 struct io_tctx_node *node;
9062 node = xa_erase(&tctx->xa, index);
9066 WARN_ON_ONCE(current != node->task);
9067 WARN_ON_ONCE(list_empty(&node->ctx_node));
9069 mutex_lock(&node->ctx->uring_lock);
9070 list_del(&node->ctx_node);
9071 mutex_unlock(&node->ctx->uring_lock);
9073 if (tctx->last == node->ctx)
9078 static void io_uring_clean_tctx(struct io_uring_task *tctx)
9080 struct io_wq *wq = tctx->io_wq;
9081 struct io_tctx_node *node;
9082 unsigned long index;
9084 xa_for_each(&tctx->xa, index, node)
9085 io_uring_del_tctx_node(index);
9088 * Must be after io_uring_del_task_file() (removes nodes under
9089 * uring_lock) to avoid race with io_uring_try_cancel_iowq().
9092 io_wq_put_and_exit(wq);
9096 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
9099 return atomic_read(&tctx->inflight_tracked);
9100 return percpu_counter_sum(&tctx->inflight);
9103 static void io_uring_drop_tctx_refs(struct task_struct *task)
9105 struct io_uring_task *tctx = task->io_uring;
9106 unsigned int refs = tctx->cached_refs;
9108 tctx->cached_refs = 0;
9109 percpu_counter_sub(&tctx->inflight, refs);
9110 put_task_struct_many(task, refs);
9114 * Find any io_uring ctx that this task has registered or done IO on, and cancel
9115 * requests. @sqd should be not-null IIF it's an SQPOLL thread cancellation.
9117 static void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
9119 struct io_uring_task *tctx = current->io_uring;
9120 struct io_ring_ctx *ctx;
9124 WARN_ON_ONCE(sqd && sqd->thread != current);
9126 if (!current->io_uring)
9129 io_wq_exit_start(tctx->io_wq);
9131 io_uring_drop_tctx_refs(current);
9132 atomic_inc(&tctx->in_idle);
9134 /* read completions before cancelations */
9135 inflight = tctx_inflight(tctx, !cancel_all);
9140 struct io_tctx_node *node;
9141 unsigned long index;
9143 xa_for_each(&tctx->xa, index, node) {
9144 /* sqpoll task will cancel all its requests */
9145 if (node->ctx->sq_data)
9147 io_uring_try_cancel_requests(node->ctx, current,
9151 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
9152 io_uring_try_cancel_requests(ctx, current,
9156 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
9158 * If we've seen completions, retry without waiting. This
9159 * avoids a race where a completion comes in before we did
9160 * prepare_to_wait().
9162 if (inflight == tctx_inflight(tctx, !cancel_all))
9164 finish_wait(&tctx->wait, &wait);
9166 atomic_dec(&tctx->in_idle);
9168 io_uring_clean_tctx(tctx);
9170 /* for exec all current's requests should be gone, kill tctx */
9171 __io_uring_free(current);
9175 void __io_uring_cancel(struct files_struct *files)
9177 io_uring_cancel_generic(!files, NULL);
9180 static void *io_uring_validate_mmap_request(struct file *file,
9181 loff_t pgoff, size_t sz)
9183 struct io_ring_ctx *ctx = file->private_data;
9184 loff_t offset = pgoff << PAGE_SHIFT;
9189 case IORING_OFF_SQ_RING:
9190 case IORING_OFF_CQ_RING:
9193 case IORING_OFF_SQES:
9197 return ERR_PTR(-EINVAL);
9200 page = virt_to_head_page(ptr);
9201 if (sz > page_size(page))
9202 return ERR_PTR(-EINVAL);
9209 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9211 size_t sz = vma->vm_end - vma->vm_start;
9215 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
9217 return PTR_ERR(ptr);
9219 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
9220 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
9223 #else /* !CONFIG_MMU */
9225 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9227 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
9230 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
9232 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
9235 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
9236 unsigned long addr, unsigned long len,
9237 unsigned long pgoff, unsigned long flags)
9241 ptr = io_uring_validate_mmap_request(file, pgoff, len);
9243 return PTR_ERR(ptr);
9245 return (unsigned long) ptr;
9248 #endif /* !CONFIG_MMU */
9250 static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
9255 if (!io_sqring_full(ctx))
9257 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9259 if (!io_sqring_full(ctx))
9262 } while (!signal_pending(current));
9264 finish_wait(&ctx->sqo_sq_wait, &wait);
9268 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9269 struct __kernel_timespec __user **ts,
9270 const sigset_t __user **sig)
9272 struct io_uring_getevents_arg arg;
9275 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9276 * is just a pointer to the sigset_t.
9278 if (!(flags & IORING_ENTER_EXT_ARG)) {
9279 *sig = (const sigset_t __user *) argp;
9285 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9286 * timespec and sigset_t pointers if good.
9288 if (*argsz != sizeof(arg))
9290 if (copy_from_user(&arg, argp, sizeof(arg)))
9292 *sig = u64_to_user_ptr(arg.sigmask);
9293 *argsz = arg.sigmask_sz;
9294 *ts = u64_to_user_ptr(arg.ts);
9298 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9299 u32, min_complete, u32, flags, const void __user *, argp,
9302 struct io_ring_ctx *ctx;
9309 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9310 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG)))
9314 if (unlikely(!f.file))
9318 if (unlikely(f.file->f_op != &io_uring_fops))
9322 ctx = f.file->private_data;
9323 if (unlikely(!percpu_ref_tryget(&ctx->refs)))
9327 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
9331 * For SQ polling, the thread will do all submissions and completions.
9332 * Just return the requested submit count, and wake the thread if
9336 if (ctx->flags & IORING_SETUP_SQPOLL) {
9337 io_cqring_overflow_flush(ctx, false);
9340 if (unlikely(ctx->sq_data->thread == NULL)) {
9343 if (flags & IORING_ENTER_SQ_WAKEUP)
9344 wake_up(&ctx->sq_data->wait);
9345 if (flags & IORING_ENTER_SQ_WAIT) {
9346 ret = io_sqpoll_wait_sq(ctx);
9350 submitted = to_submit;
9351 } else if (to_submit) {
9352 ret = io_uring_add_tctx_node(ctx);
9355 mutex_lock(&ctx->uring_lock);
9356 submitted = io_submit_sqes(ctx, to_submit);
9357 mutex_unlock(&ctx->uring_lock);
9359 if (submitted != to_submit)
9362 if (flags & IORING_ENTER_GETEVENTS) {
9363 const sigset_t __user *sig;
9364 struct __kernel_timespec __user *ts;
9366 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
9370 min_complete = min(min_complete, ctx->cq_entries);
9373 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9374 * space applications don't need to do io completion events
9375 * polling again, they can rely on io_sq_thread to do polling
9376 * work, which can reduce cpu usage and uring_lock contention.
9378 if (ctx->flags & IORING_SETUP_IOPOLL &&
9379 !(ctx->flags & IORING_SETUP_SQPOLL)) {
9380 ret = io_iopoll_check(ctx, min_complete);
9382 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
9387 percpu_ref_put(&ctx->refs);
9390 return submitted ? submitted : ret;
9393 #ifdef CONFIG_PROC_FS
9394 static int io_uring_show_cred(struct seq_file *m, unsigned int id,
9395 const struct cred *cred)
9397 struct user_namespace *uns = seq_user_ns(m);
9398 struct group_info *gi;
9403 seq_printf(m, "%5d\n", id);
9404 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
9405 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
9406 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
9407 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
9408 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
9409 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
9410 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
9411 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
9412 seq_puts(m, "\n\tGroups:\t");
9413 gi = cred->group_info;
9414 for (g = 0; g < gi->ngroups; g++) {
9415 seq_put_decimal_ull(m, g ? " " : "",
9416 from_kgid_munged(uns, gi->gid[g]));
9418 seq_puts(m, "\n\tCapEff:\t");
9419 cap = cred->cap_effective;
9420 CAP_FOR_EACH_U32(__capi)
9421 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
9426 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
9428 struct io_sq_data *sq = NULL;
9433 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9434 * since fdinfo case grabs it in the opposite direction of normal use
9435 * cases. If we fail to get the lock, we just don't iterate any
9436 * structures that could be going away outside the io_uring mutex.
9438 has_lock = mutex_trylock(&ctx->uring_lock);
9440 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) {
9446 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9447 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9448 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9449 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9450 struct file *f = io_file_from_index(ctx, i);
9453 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9455 seq_printf(m, "%5u: <none>\n", i);
9457 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9458 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9459 struct io_mapped_ubuf *buf = ctx->user_bufs[i];
9460 unsigned int len = buf->ubuf_end - buf->ubuf;
9462 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf, len);
9464 if (has_lock && !xa_empty(&ctx->personalities)) {
9465 unsigned long index;
9466 const struct cred *cred;
9468 seq_printf(m, "Personalities:\n");
9469 xa_for_each(&ctx->personalities, index, cred)
9470 io_uring_show_cred(m, index, cred);
9472 seq_printf(m, "PollList:\n");
9473 spin_lock_irq(&ctx->completion_lock);
9474 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9475 struct hlist_head *list = &ctx->cancel_hash[i];
9476 struct io_kiocb *req;
9478 hlist_for_each_entry(req, list, hash_node)
9479 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9480 req->task->task_works != NULL);
9482 spin_unlock_irq(&ctx->completion_lock);
9484 mutex_unlock(&ctx->uring_lock);
9487 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9489 struct io_ring_ctx *ctx = f->private_data;
9491 if (percpu_ref_tryget(&ctx->refs)) {
9492 __io_uring_show_fdinfo(ctx, m);
9493 percpu_ref_put(&ctx->refs);
9498 static const struct file_operations io_uring_fops = {
9499 .release = io_uring_release,
9500 .mmap = io_uring_mmap,
9502 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9503 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9505 .poll = io_uring_poll,
9506 .fasync = io_uring_fasync,
9507 #ifdef CONFIG_PROC_FS
9508 .show_fdinfo = io_uring_show_fdinfo,
9512 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9513 struct io_uring_params *p)
9515 struct io_rings *rings;
9516 size_t size, sq_array_offset;
9518 /* make sure these are sane, as we already accounted them */
9519 ctx->sq_entries = p->sq_entries;
9520 ctx->cq_entries = p->cq_entries;
9522 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9523 if (size == SIZE_MAX)
9526 rings = io_mem_alloc(size);
9531 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9532 rings->sq_ring_mask = p->sq_entries - 1;
9533 rings->cq_ring_mask = p->cq_entries - 1;
9534 rings->sq_ring_entries = p->sq_entries;
9535 rings->cq_ring_entries = p->cq_entries;
9537 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9538 if (size == SIZE_MAX) {
9539 io_mem_free(ctx->rings);
9544 ctx->sq_sqes = io_mem_alloc(size);
9545 if (!ctx->sq_sqes) {
9546 io_mem_free(ctx->rings);
9554 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
9558 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9562 ret = io_uring_add_tctx_node(ctx);
9567 fd_install(fd, file);
9572 * Allocate an anonymous fd, this is what constitutes the application
9573 * visible backing of an io_uring instance. The application mmaps this
9574 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9575 * we have to tie this fd to a socket for file garbage collection purposes.
9577 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
9580 #if defined(CONFIG_UNIX)
9583 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9586 return ERR_PTR(ret);
9589 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9590 O_RDWR | O_CLOEXEC);
9591 #if defined(CONFIG_UNIX)
9593 sock_release(ctx->ring_sock);
9594 ctx->ring_sock = NULL;
9596 ctx->ring_sock->file = file;
9602 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9603 struct io_uring_params __user *params)
9605 struct io_ring_ctx *ctx;
9611 if (entries > IORING_MAX_ENTRIES) {
9612 if (!(p->flags & IORING_SETUP_CLAMP))
9614 entries = IORING_MAX_ENTRIES;
9618 * Use twice as many entries for the CQ ring. It's possible for the
9619 * application to drive a higher depth than the size of the SQ ring,
9620 * since the sqes are only used at submission time. This allows for
9621 * some flexibility in overcommitting a bit. If the application has
9622 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9623 * of CQ ring entries manually.
9625 p->sq_entries = roundup_pow_of_two(entries);
9626 if (p->flags & IORING_SETUP_CQSIZE) {
9628 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9629 * to a power-of-two, if it isn't already. We do NOT impose
9630 * any cq vs sq ring sizing.
9634 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9635 if (!(p->flags & IORING_SETUP_CLAMP))
9637 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9639 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9640 if (p->cq_entries < p->sq_entries)
9643 p->cq_entries = 2 * p->sq_entries;
9646 ctx = io_ring_ctx_alloc(p);
9649 ctx->compat = in_compat_syscall();
9650 if (!capable(CAP_IPC_LOCK))
9651 ctx->user = get_uid(current_user());
9654 * This is just grabbed for accounting purposes. When a process exits,
9655 * the mm is exited and dropped before the files, hence we need to hang
9656 * on to this mm purely for the purposes of being able to unaccount
9657 * memory (locked/pinned vm). It's not used for anything else.
9659 mmgrab(current->mm);
9660 ctx->mm_account = current->mm;
9662 ret = io_allocate_scq_urings(ctx, p);
9666 ret = io_sq_offload_create(ctx, p);
9669 /* always set a rsrc node */
9670 ret = io_rsrc_node_switch_start(ctx);
9673 io_rsrc_node_switch(ctx, NULL);
9675 memset(&p->sq_off, 0, sizeof(p->sq_off));
9676 p->sq_off.head = offsetof(struct io_rings, sq.head);
9677 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9678 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9679 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9680 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9681 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9682 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9684 memset(&p->cq_off, 0, sizeof(p->cq_off));
9685 p->cq_off.head = offsetof(struct io_rings, cq.head);
9686 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9687 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9688 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9689 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9690 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9691 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9693 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9694 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9695 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9696 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
9697 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
9698 IORING_FEAT_RSRC_TAGS;
9700 if (copy_to_user(params, p, sizeof(*p))) {
9705 file = io_uring_get_file(ctx);
9707 ret = PTR_ERR(file);
9712 * Install ring fd as the very last thing, so we don't risk someone
9713 * having closed it before we finish setup
9715 ret = io_uring_install_fd(ctx, file);
9717 /* fput will clean it up */
9722 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9725 io_ring_ctx_wait_and_kill(ctx);
9730 * Sets up an aio uring context, and returns the fd. Applications asks for a
9731 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9732 * params structure passed in.
9734 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9736 struct io_uring_params p;
9739 if (copy_from_user(&p, params, sizeof(p)))
9741 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9746 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9747 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9748 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9749 IORING_SETUP_R_DISABLED))
9752 return io_uring_create(entries, &p, params);
9755 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9756 struct io_uring_params __user *, params)
9758 return io_uring_setup(entries, params);
9761 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9763 struct io_uring_probe *p;
9767 size = struct_size(p, ops, nr_args);
9768 if (size == SIZE_MAX)
9770 p = kzalloc(size, GFP_KERNEL);
9775 if (copy_from_user(p, arg, size))
9778 if (memchr_inv(p, 0, size))
9781 p->last_op = IORING_OP_LAST - 1;
9782 if (nr_args > IORING_OP_LAST)
9783 nr_args = IORING_OP_LAST;
9785 for (i = 0; i < nr_args; i++) {
9787 if (!io_op_defs[i].not_supported)
9788 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9793 if (copy_to_user(arg, p, size))
9800 static int io_register_personality(struct io_ring_ctx *ctx)
9802 const struct cred *creds;
9806 creds = get_current_cred();
9808 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
9809 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
9816 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9817 unsigned int nr_args)
9819 struct io_uring_restriction *res;
9823 /* Restrictions allowed only if rings started disabled */
9824 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9827 /* We allow only a single restrictions registration */
9828 if (ctx->restrictions.registered)
9831 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9834 size = array_size(nr_args, sizeof(*res));
9835 if (size == SIZE_MAX)
9838 res = memdup_user(arg, size);
9840 return PTR_ERR(res);
9844 for (i = 0; i < nr_args; i++) {
9845 switch (res[i].opcode) {
9846 case IORING_RESTRICTION_REGISTER_OP:
9847 if (res[i].register_op >= IORING_REGISTER_LAST) {
9852 __set_bit(res[i].register_op,
9853 ctx->restrictions.register_op);
9855 case IORING_RESTRICTION_SQE_OP:
9856 if (res[i].sqe_op >= IORING_OP_LAST) {
9861 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9863 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9864 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9866 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9867 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9876 /* Reset all restrictions if an error happened */
9878 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9880 ctx->restrictions.registered = true;
9886 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9888 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9891 if (ctx->restrictions.registered)
9892 ctx->restricted = 1;
9894 ctx->flags &= ~IORING_SETUP_R_DISABLED;
9895 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
9896 wake_up(&ctx->sq_data->wait);
9900 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type,
9901 struct io_uring_rsrc_update2 *up,
9909 if (check_add_overflow(up->offset, nr_args, &tmp))
9911 err = io_rsrc_node_switch_start(ctx);
9916 case IORING_RSRC_FILE:
9917 return __io_sqe_files_update(ctx, up, nr_args);
9918 case IORING_RSRC_BUFFER:
9919 return __io_sqe_buffers_update(ctx, up, nr_args);
9924 static int io_register_files_update(struct io_ring_ctx *ctx, void __user *arg,
9927 struct io_uring_rsrc_update2 up;
9931 memset(&up, 0, sizeof(up));
9932 if (copy_from_user(&up, arg, sizeof(struct io_uring_rsrc_update)))
9934 return __io_register_rsrc_update(ctx, IORING_RSRC_FILE, &up, nr_args);
9937 static int io_register_rsrc_update(struct io_ring_ctx *ctx, void __user *arg,
9938 unsigned size, unsigned type)
9940 struct io_uring_rsrc_update2 up;
9942 if (size != sizeof(up))
9944 if (copy_from_user(&up, arg, sizeof(up)))
9946 if (!up.nr || up.resv)
9948 return __io_register_rsrc_update(ctx, type, &up, up.nr);
9951 static int io_register_rsrc(struct io_ring_ctx *ctx, void __user *arg,
9952 unsigned int size, unsigned int type)
9954 struct io_uring_rsrc_register rr;
9956 /* keep it extendible */
9957 if (size != sizeof(rr))
9960 memset(&rr, 0, sizeof(rr));
9961 if (copy_from_user(&rr, arg, size))
9963 if (!rr.nr || rr.resv || rr.resv2)
9967 case IORING_RSRC_FILE:
9968 return io_sqe_files_register(ctx, u64_to_user_ptr(rr.data),
9969 rr.nr, u64_to_user_ptr(rr.tags));
9970 case IORING_RSRC_BUFFER:
9971 return io_sqe_buffers_register(ctx, u64_to_user_ptr(rr.data),
9972 rr.nr, u64_to_user_ptr(rr.tags));
9977 static bool io_register_op_must_quiesce(int op)
9980 case IORING_REGISTER_BUFFERS:
9981 case IORING_UNREGISTER_BUFFERS:
9982 case IORING_REGISTER_FILES:
9983 case IORING_UNREGISTER_FILES:
9984 case IORING_REGISTER_FILES_UPDATE:
9985 case IORING_REGISTER_PROBE:
9986 case IORING_REGISTER_PERSONALITY:
9987 case IORING_UNREGISTER_PERSONALITY:
9988 case IORING_REGISTER_FILES2:
9989 case IORING_REGISTER_FILES_UPDATE2:
9990 case IORING_REGISTER_BUFFERS2:
9991 case IORING_REGISTER_BUFFERS_UPDATE:
9998 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
9999 void __user *arg, unsigned nr_args)
10000 __releases(ctx->uring_lock)
10001 __acquires(ctx->uring_lock)
10006 * We're inside the ring mutex, if the ref is already dying, then
10007 * someone else killed the ctx or is already going through
10008 * io_uring_register().
10010 if (percpu_ref_is_dying(&ctx->refs))
10013 if (ctx->restricted) {
10014 if (opcode >= IORING_REGISTER_LAST)
10016 opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
10017 if (!test_bit(opcode, ctx->restrictions.register_op))
10021 if (io_register_op_must_quiesce(opcode)) {
10022 percpu_ref_kill(&ctx->refs);
10025 * Drop uring mutex before waiting for references to exit. If
10026 * another thread is currently inside io_uring_enter() it might
10027 * need to grab the uring_lock to make progress. If we hold it
10028 * here across the drain wait, then we can deadlock. It's safe
10029 * to drop the mutex here, since no new references will come in
10030 * after we've killed the percpu ref.
10032 mutex_unlock(&ctx->uring_lock);
10034 ret = wait_for_completion_interruptible(&ctx->ref_comp);
10037 ret = io_run_task_work_sig();
10041 mutex_lock(&ctx->uring_lock);
10044 io_refs_resurrect(&ctx->refs, &ctx->ref_comp);
10050 case IORING_REGISTER_BUFFERS:
10051 ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
10053 case IORING_UNREGISTER_BUFFERS:
10055 if (arg || nr_args)
10057 ret = io_sqe_buffers_unregister(ctx);
10059 case IORING_REGISTER_FILES:
10060 ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
10062 case IORING_UNREGISTER_FILES:
10064 if (arg || nr_args)
10066 ret = io_sqe_files_unregister(ctx);
10068 case IORING_REGISTER_FILES_UPDATE:
10069 ret = io_register_files_update(ctx, arg, nr_args);
10071 case IORING_REGISTER_EVENTFD:
10072 case IORING_REGISTER_EVENTFD_ASYNC:
10076 ret = io_eventfd_register(ctx, arg);
10079 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
10080 ctx->eventfd_async = 1;
10082 ctx->eventfd_async = 0;
10084 case IORING_UNREGISTER_EVENTFD:
10086 if (arg || nr_args)
10088 ret = io_eventfd_unregister(ctx);
10090 case IORING_REGISTER_PROBE:
10092 if (!arg || nr_args > 256)
10094 ret = io_probe(ctx, arg, nr_args);
10096 case IORING_REGISTER_PERSONALITY:
10098 if (arg || nr_args)
10100 ret = io_register_personality(ctx);
10102 case IORING_UNREGISTER_PERSONALITY:
10106 ret = io_unregister_personality(ctx, nr_args);
10108 case IORING_REGISTER_ENABLE_RINGS:
10110 if (arg || nr_args)
10112 ret = io_register_enable_rings(ctx);
10114 case IORING_REGISTER_RESTRICTIONS:
10115 ret = io_register_restrictions(ctx, arg, nr_args);
10117 case IORING_REGISTER_FILES2:
10118 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE);
10120 case IORING_REGISTER_FILES_UPDATE2:
10121 ret = io_register_rsrc_update(ctx, arg, nr_args,
10124 case IORING_REGISTER_BUFFERS2:
10125 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER);
10127 case IORING_REGISTER_BUFFERS_UPDATE:
10128 ret = io_register_rsrc_update(ctx, arg, nr_args,
10129 IORING_RSRC_BUFFER);
10136 if (io_register_op_must_quiesce(opcode)) {
10137 /* bring the ctx back to life */
10138 percpu_ref_reinit(&ctx->refs);
10139 reinit_completion(&ctx->ref_comp);
10144 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
10145 void __user *, arg, unsigned int, nr_args)
10147 struct io_ring_ctx *ctx;
10156 if (f.file->f_op != &io_uring_fops)
10159 ctx = f.file->private_data;
10161 io_run_task_work();
10163 mutex_lock(&ctx->uring_lock);
10164 ret = __io_uring_register(ctx, opcode, arg, nr_args);
10165 mutex_unlock(&ctx->uring_lock);
10166 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
10167 ctx->cq_ev_fd != NULL, ret);
10173 static int __init io_uring_init(void)
10175 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
10176 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
10177 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
10180 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
10181 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
10182 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
10183 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
10184 BUILD_BUG_SQE_ELEM(1, __u8, flags);
10185 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
10186 BUILD_BUG_SQE_ELEM(4, __s32, fd);
10187 BUILD_BUG_SQE_ELEM(8, __u64, off);
10188 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
10189 BUILD_BUG_SQE_ELEM(16, __u64, addr);
10190 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
10191 BUILD_BUG_SQE_ELEM(24, __u32, len);
10192 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
10193 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
10194 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
10195 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
10196 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
10197 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
10198 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
10199 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
10200 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
10201 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
10202 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
10203 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
10204 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
10205 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
10206 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
10207 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
10208 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
10209 BUILD_BUG_SQE_ELEM(42, __u16, personality);
10210 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
10212 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
10213 sizeof(struct io_uring_rsrc_update));
10214 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
10215 sizeof(struct io_uring_rsrc_update2));
10216 /* should fit into one byte */
10217 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
10219 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
10220 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
10221 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
10225 __initcall(io_uring_init);