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_cqring (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>
81 #include <linux/freezer.h>
83 #define CREATE_TRACE_POINTS
84 #include <trace/events/io_uring.h>
86 #include <uapi/linux/io_uring.h>
91 #define IORING_MAX_ENTRIES 32768
92 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
95 * Shift of 9 is 512 entries, or exactly one page on 64-bit archs
97 #define IORING_FILE_TABLE_SHIFT 9
98 #define IORING_MAX_FILES_TABLE (1U << IORING_FILE_TABLE_SHIFT)
99 #define IORING_FILE_TABLE_MASK (IORING_MAX_FILES_TABLE - 1)
100 #define IORING_MAX_FIXED_FILES (64 * IORING_MAX_FILES_TABLE)
101 #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
102 IORING_REGISTER_LAST + IORING_OP_LAST)
104 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
105 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
109 u32 head ____cacheline_aligned_in_smp;
110 u32 tail ____cacheline_aligned_in_smp;
114 * This data is shared with the application through the mmap at offsets
115 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
117 * The offsets to the member fields are published through struct
118 * io_sqring_offsets when calling io_uring_setup.
122 * Head and tail offsets into the ring; the offsets need to be
123 * masked to get valid indices.
125 * The kernel controls head of the sq ring and the tail of the cq ring,
126 * and the application controls tail of the sq ring and the head of the
129 struct io_uring sq, cq;
131 * Bitmasks to apply to head and tail offsets (constant, equals
134 u32 sq_ring_mask, cq_ring_mask;
135 /* Ring sizes (constant, power of 2) */
136 u32 sq_ring_entries, cq_ring_entries;
138 * Number of invalid entries dropped by the kernel due to
139 * invalid index stored in array
141 * Written by the kernel, shouldn't be modified by the
142 * application (i.e. get number of "new events" by comparing to
145 * After a new SQ head value was read by the application this
146 * counter includes all submissions that were dropped reaching
147 * the new SQ head (and possibly more).
153 * Written by the kernel, shouldn't be modified by the
156 * The application needs a full memory barrier before checking
157 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
163 * Written by the application, shouldn't be modified by the
168 * Number of completion events lost because the queue was full;
169 * this should be avoided by the application by making sure
170 * there are not more requests pending than there is space in
171 * the completion queue.
173 * Written by the kernel, shouldn't be modified by the
174 * application (i.e. get number of "new events" by comparing to
177 * As completion events come in out of order this counter is not
178 * ordered with any other data.
182 * Ring buffer of completion events.
184 * The kernel writes completion events fresh every time they are
185 * produced, so the application is allowed to modify pending
188 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
191 enum io_uring_cmd_flags {
192 IO_URING_F_NONBLOCK = 1,
193 IO_URING_F_COMPLETE_DEFER = 2,
196 struct io_mapped_ubuf {
199 struct bio_vec *bvec;
200 unsigned int nr_bvecs;
201 unsigned long acct_pages;
207 struct list_head list;
214 struct fixed_rsrc_table {
218 struct fixed_rsrc_ref_node {
219 struct percpu_ref refs;
220 struct list_head node;
221 struct list_head rsrc_list;
222 struct fixed_rsrc_data *rsrc_data;
223 void (*rsrc_put)(struct io_ring_ctx *ctx,
224 struct io_rsrc_put *prsrc);
225 struct llist_node llist;
229 struct fixed_rsrc_data {
230 struct fixed_rsrc_table *table;
231 struct io_ring_ctx *ctx;
233 struct fixed_rsrc_ref_node *node;
234 struct percpu_ref refs;
235 struct completion done;
240 struct list_head list;
246 struct io_restriction {
247 DECLARE_BITMAP(register_op, IORING_REGISTER_LAST);
248 DECLARE_BITMAP(sqe_op, IORING_OP_LAST);
249 u8 sqe_flags_allowed;
250 u8 sqe_flags_required;
255 IO_SQ_THREAD_SHOULD_STOP = 0,
256 IO_SQ_THREAD_SHOULD_PARK,
261 struct rw_semaphore rw_lock;
263 /* ctx's that are using this sqd */
264 struct list_head ctx_list;
265 struct list_head ctx_new_list;
267 struct task_struct *thread;
268 struct wait_queue_head wait;
270 unsigned sq_thread_idle;
275 struct completion startup;
276 struct completion exited;
279 #define IO_IOPOLL_BATCH 8
280 #define IO_COMPL_BATCH 32
281 #define IO_REQ_CACHE_SIZE 32
282 #define IO_REQ_ALLOC_BATCH 8
284 struct io_comp_state {
285 struct io_kiocb *reqs[IO_COMPL_BATCH];
287 unsigned int locked_free_nr;
288 /* inline/task_work completion list, under ->uring_lock */
289 struct list_head free_list;
290 /* IRQ completion list, under ->completion_lock */
291 struct list_head locked_free_list;
294 struct io_submit_link {
295 struct io_kiocb *head;
296 struct io_kiocb *last;
299 struct io_submit_state {
300 struct blk_plug plug;
301 struct io_submit_link link;
304 * io_kiocb alloc cache
306 void *reqs[IO_REQ_CACHE_SIZE];
307 unsigned int free_reqs;
312 * Batch completion logic
314 struct io_comp_state comp;
317 * File reference cache
321 unsigned int file_refs;
322 unsigned int ios_left;
327 struct percpu_ref refs;
328 } ____cacheline_aligned_in_smp;
332 unsigned int compat: 1;
333 unsigned int cq_overflow_flushed: 1;
334 unsigned int drain_next: 1;
335 unsigned int eventfd_async: 1;
336 unsigned int restricted: 1;
339 * Ring buffer of indices into array of io_uring_sqe, which is
340 * mmapped by the application using the IORING_OFF_SQES offset.
342 * This indirection could e.g. be used to assign fixed
343 * io_uring_sqe entries to operations and only submit them to
344 * the queue when needed.
346 * The kernel modifies neither the indices array nor the entries
350 unsigned cached_sq_head;
353 unsigned sq_thread_idle;
354 unsigned cached_sq_dropped;
355 unsigned cached_cq_overflow;
356 unsigned long sq_check_overflow;
358 /* hashed buffered write serialization */
359 struct io_wq_hash *hash_map;
361 struct list_head defer_list;
362 struct list_head timeout_list;
363 struct list_head cq_overflow_list;
365 struct io_uring_sqe *sq_sqes;
366 } ____cacheline_aligned_in_smp;
369 struct mutex uring_lock;
370 wait_queue_head_t wait;
371 } ____cacheline_aligned_in_smp;
373 struct io_submit_state submit_state;
375 struct io_rings *rings;
377 /* Only used for accounting purposes */
378 struct mm_struct *mm_account;
380 const struct cred *sq_creds; /* cred used for __io_sq_thread() */
381 struct io_sq_data *sq_data; /* if using sq thread polling */
383 struct wait_queue_head sqo_sq_wait;
384 struct list_head sqd_list;
387 * If used, fixed file set. Writers must ensure that ->refs is dead,
388 * readers must ensure that ->refs is alive as long as the file* is
389 * used. Only updated through io_uring_register(2).
391 struct fixed_rsrc_data *file_data;
392 unsigned nr_user_files;
394 /* if used, fixed mapped user buffers */
395 unsigned nr_user_bufs;
396 struct io_mapped_ubuf *user_bufs;
398 struct user_struct *user;
400 struct completion ref_comp;
401 struct completion sq_thread_comp;
403 #if defined(CONFIG_UNIX)
404 struct socket *ring_sock;
407 struct idr io_buffer_idr;
409 struct xarray personalities;
413 unsigned cached_cq_tail;
416 atomic_t cq_timeouts;
417 unsigned cq_last_tm_flush;
418 unsigned long cq_check_overflow;
419 struct wait_queue_head cq_wait;
420 struct fasync_struct *cq_fasync;
421 struct eventfd_ctx *cq_ev_fd;
422 } ____cacheline_aligned_in_smp;
425 spinlock_t completion_lock;
428 * ->iopoll_list is protected by the ctx->uring_lock for
429 * io_uring instances that don't use IORING_SETUP_SQPOLL.
430 * For SQPOLL, only the single threaded io_sq_thread() will
431 * manipulate the list, hence no extra locking is needed there.
433 struct list_head iopoll_list;
434 struct hlist_head *cancel_hash;
435 unsigned cancel_hash_bits;
436 bool poll_multi_file;
438 spinlock_t inflight_lock;
439 struct list_head inflight_list;
440 } ____cacheline_aligned_in_smp;
442 struct delayed_work rsrc_put_work;
443 struct llist_head rsrc_put_llist;
444 struct list_head rsrc_ref_list;
445 spinlock_t rsrc_ref_lock;
447 struct io_restriction restrictions;
450 struct callback_head *exit_task_work;
452 struct wait_queue_head hash_wait;
454 /* Keep this last, we don't need it for the fast path */
455 struct work_struct exit_work;
456 struct list_head tctx_list;
460 * First field must be the file pointer in all the
461 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
463 struct io_poll_iocb {
465 struct wait_queue_head *head;
469 struct wait_queue_entry wait;
472 struct io_poll_remove {
482 struct io_timeout_data {
483 struct io_kiocb *req;
484 struct hrtimer timer;
485 struct timespec64 ts;
486 enum hrtimer_mode mode;
491 struct sockaddr __user *addr;
492 int __user *addr_len;
494 unsigned long nofile;
514 struct list_head list;
515 /* head of the link, used by linked timeouts only */
516 struct io_kiocb *head;
519 struct io_timeout_rem {
524 struct timespec64 ts;
529 /* NOTE: kiocb has the file as the first member, so don't do it here */
537 struct sockaddr __user *addr;
544 struct user_msghdr __user *umsg;
550 struct io_buffer *kbuf;
556 struct filename *filename;
558 unsigned long nofile;
561 struct io_rsrc_update {
587 struct epoll_event event;
591 struct file *file_out;
592 struct file *file_in;
599 struct io_provide_buf {
613 const char __user *filename;
614 struct statx __user *buffer;
626 struct filename *oldpath;
627 struct filename *newpath;
635 struct filename *filename;
638 struct io_completion {
640 struct list_head list;
644 struct io_async_connect {
645 struct sockaddr_storage address;
648 struct io_async_msghdr {
649 struct iovec fast_iov[UIO_FASTIOV];
650 /* points to an allocated iov, if NULL we use fast_iov instead */
651 struct iovec *free_iov;
652 struct sockaddr __user *uaddr;
654 struct sockaddr_storage addr;
658 struct iovec fast_iov[UIO_FASTIOV];
659 const struct iovec *free_iovec;
660 struct iov_iter iter;
662 struct wait_page_queue wpq;
666 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
667 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
668 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
669 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
670 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
671 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
677 REQ_F_LINK_TIMEOUT_BIT,
679 REQ_F_NEED_CLEANUP_BIT,
681 REQ_F_BUFFER_SELECTED_BIT,
682 REQ_F_NO_FILE_TABLE_BIT,
683 REQ_F_LTIMEOUT_ACTIVE_BIT,
684 REQ_F_COMPLETE_INLINE_BIT,
686 /* not a real bit, just to check we're not overflowing the space */
692 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
693 /* drain existing IO first */
694 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
696 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
697 /* doesn't sever on completion < 0 */
698 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
700 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
701 /* IOSQE_BUFFER_SELECT */
702 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
704 /* fail rest of links */
705 REQ_F_FAIL_LINK = BIT(REQ_F_FAIL_LINK_BIT),
706 /* on inflight list, should be cancelled and waited on exit reliably */
707 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
708 /* read/write uses file position */
709 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
710 /* must not punt to workers */
711 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
712 /* has or had linked timeout */
713 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
715 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
717 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
718 /* already went through poll handler */
719 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
720 /* buffer already selected */
721 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
722 /* doesn't need file table for this request */
723 REQ_F_NO_FILE_TABLE = BIT(REQ_F_NO_FILE_TABLE_BIT),
724 /* linked timeout is active, i.e. prepared by link's head */
725 REQ_F_LTIMEOUT_ACTIVE = BIT(REQ_F_LTIMEOUT_ACTIVE_BIT),
726 /* completion is deferred through io_comp_state */
727 REQ_F_COMPLETE_INLINE = BIT(REQ_F_COMPLETE_INLINE_BIT),
731 struct io_poll_iocb poll;
732 struct io_poll_iocb *double_poll;
735 struct io_task_work {
736 struct io_wq_work_node node;
737 task_work_func_t func;
741 * NOTE! Each of the iocb union members has the file pointer
742 * as the first entry in their struct definition. So you can
743 * access the file pointer through any of the sub-structs,
744 * or directly as just 'ki_filp' in this struct.
750 struct io_poll_iocb poll;
751 struct io_poll_remove poll_remove;
752 struct io_accept accept;
754 struct io_cancel cancel;
755 struct io_timeout timeout;
756 struct io_timeout_rem timeout_rem;
757 struct io_connect connect;
758 struct io_sr_msg sr_msg;
760 struct io_close close;
761 struct io_rsrc_update rsrc_update;
762 struct io_fadvise fadvise;
763 struct io_madvise madvise;
764 struct io_epoll epoll;
765 struct io_splice splice;
766 struct io_provide_buf pbuf;
767 struct io_statx statx;
768 struct io_shutdown shutdown;
769 struct io_rename rename;
770 struct io_unlink unlink;
771 /* use only after cleaning per-op data, see io_clean_op() */
772 struct io_completion compl;
775 /* opcode allocated if it needs to store data for async defer */
778 /* polled IO has completed */
784 struct io_ring_ctx *ctx;
787 struct task_struct *task;
790 struct io_kiocb *link;
791 struct percpu_ref *fixed_rsrc_refs;
794 * 1. used with ctx->iopoll_list with reads/writes
795 * 2. to track reqs with ->files (see io_op_def::file_table)
797 struct list_head inflight_entry;
799 struct io_task_work io_task_work;
800 struct callback_head task_work;
802 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
803 struct hlist_node hash_node;
804 struct async_poll *apoll;
805 struct io_wq_work work;
808 struct io_tctx_node {
809 struct list_head ctx_node;
810 struct task_struct *task;
811 struct io_ring_ctx *ctx;
814 struct io_defer_entry {
815 struct list_head list;
816 struct io_kiocb *req;
821 /* needs req->file assigned */
822 unsigned needs_file : 1;
823 /* hash wq insertion if file is a regular file */
824 unsigned hash_reg_file : 1;
825 /* unbound wq insertion if file is a non-regular file */
826 unsigned unbound_nonreg_file : 1;
827 /* opcode is not supported by this kernel */
828 unsigned not_supported : 1;
829 /* set if opcode supports polled "wait" */
831 unsigned pollout : 1;
832 /* op supports buffer selection */
833 unsigned buffer_select : 1;
834 /* must always have async data allocated */
835 unsigned needs_async_data : 1;
836 /* should block plug */
838 /* size of async data needed, if any */
839 unsigned short async_size;
842 static const struct io_op_def io_op_defs[] = {
843 [IORING_OP_NOP] = {},
844 [IORING_OP_READV] = {
846 .unbound_nonreg_file = 1,
849 .needs_async_data = 1,
851 .async_size = sizeof(struct io_async_rw),
853 [IORING_OP_WRITEV] = {
856 .unbound_nonreg_file = 1,
858 .needs_async_data = 1,
860 .async_size = sizeof(struct io_async_rw),
862 [IORING_OP_FSYNC] = {
865 [IORING_OP_READ_FIXED] = {
867 .unbound_nonreg_file = 1,
870 .async_size = sizeof(struct io_async_rw),
872 [IORING_OP_WRITE_FIXED] = {
875 .unbound_nonreg_file = 1,
878 .async_size = sizeof(struct io_async_rw),
880 [IORING_OP_POLL_ADD] = {
882 .unbound_nonreg_file = 1,
884 [IORING_OP_POLL_REMOVE] = {},
885 [IORING_OP_SYNC_FILE_RANGE] = {
888 [IORING_OP_SENDMSG] = {
890 .unbound_nonreg_file = 1,
892 .needs_async_data = 1,
893 .async_size = sizeof(struct io_async_msghdr),
895 [IORING_OP_RECVMSG] = {
897 .unbound_nonreg_file = 1,
900 .needs_async_data = 1,
901 .async_size = sizeof(struct io_async_msghdr),
903 [IORING_OP_TIMEOUT] = {
904 .needs_async_data = 1,
905 .async_size = sizeof(struct io_timeout_data),
907 [IORING_OP_TIMEOUT_REMOVE] = {
908 /* used by timeout updates' prep() */
910 [IORING_OP_ACCEPT] = {
912 .unbound_nonreg_file = 1,
915 [IORING_OP_ASYNC_CANCEL] = {},
916 [IORING_OP_LINK_TIMEOUT] = {
917 .needs_async_data = 1,
918 .async_size = sizeof(struct io_timeout_data),
920 [IORING_OP_CONNECT] = {
922 .unbound_nonreg_file = 1,
924 .needs_async_data = 1,
925 .async_size = sizeof(struct io_async_connect),
927 [IORING_OP_FALLOCATE] = {
930 [IORING_OP_OPENAT] = {},
931 [IORING_OP_CLOSE] = {},
932 [IORING_OP_FILES_UPDATE] = {},
933 [IORING_OP_STATX] = {},
936 .unbound_nonreg_file = 1,
940 .async_size = sizeof(struct io_async_rw),
942 [IORING_OP_WRITE] = {
944 .unbound_nonreg_file = 1,
947 .async_size = sizeof(struct io_async_rw),
949 [IORING_OP_FADVISE] = {
952 [IORING_OP_MADVISE] = {},
955 .unbound_nonreg_file = 1,
960 .unbound_nonreg_file = 1,
964 [IORING_OP_OPENAT2] = {
966 [IORING_OP_EPOLL_CTL] = {
967 .unbound_nonreg_file = 1,
969 [IORING_OP_SPLICE] = {
972 .unbound_nonreg_file = 1,
974 [IORING_OP_PROVIDE_BUFFERS] = {},
975 [IORING_OP_REMOVE_BUFFERS] = {},
979 .unbound_nonreg_file = 1,
981 [IORING_OP_SHUTDOWN] = {
984 [IORING_OP_RENAMEAT] = {},
985 [IORING_OP_UNLINKAT] = {},
988 static bool io_disarm_next(struct io_kiocb *req);
989 static void io_uring_del_task_file(unsigned long index);
990 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
991 struct task_struct *task,
992 struct files_struct *files);
993 static void io_uring_cancel_sqpoll(struct io_ring_ctx *ctx);
994 static void destroy_fixed_rsrc_ref_node(struct fixed_rsrc_ref_node *ref_node);
995 static struct fixed_rsrc_ref_node *alloc_fixed_rsrc_ref_node(
996 struct io_ring_ctx *ctx);
997 static void io_ring_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc);
999 static bool io_rw_reissue(struct io_kiocb *req);
1000 static void io_cqring_fill_event(struct io_kiocb *req, long res);
1001 static void io_put_req(struct io_kiocb *req);
1002 static void io_put_req_deferred(struct io_kiocb *req, int nr);
1003 static void io_double_put_req(struct io_kiocb *req);
1004 static void io_dismantle_req(struct io_kiocb *req);
1005 static void io_put_task(struct task_struct *task, int nr);
1006 static void io_queue_next(struct io_kiocb *req);
1007 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
1008 static void __io_queue_linked_timeout(struct io_kiocb *req);
1009 static void io_queue_linked_timeout(struct io_kiocb *req);
1010 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
1011 struct io_uring_rsrc_update *ip,
1013 static void __io_clean_op(struct io_kiocb *req);
1014 static struct file *io_file_get(struct io_submit_state *state,
1015 struct io_kiocb *req, int fd, bool fixed);
1016 static void __io_queue_sqe(struct io_kiocb *req);
1017 static void io_rsrc_put_work(struct work_struct *work);
1019 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
1020 struct iov_iter *iter, bool needs_lock);
1021 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
1022 const struct iovec *fast_iov,
1023 struct iov_iter *iter, bool force);
1024 static void io_req_task_queue(struct io_kiocb *req);
1025 static void io_submit_flush_completions(struct io_comp_state *cs,
1026 struct io_ring_ctx *ctx);
1028 static struct kmem_cache *req_cachep;
1030 static const struct file_operations io_uring_fops;
1032 struct sock *io_uring_get_socket(struct file *file)
1034 #if defined(CONFIG_UNIX)
1035 if (file->f_op == &io_uring_fops) {
1036 struct io_ring_ctx *ctx = file->private_data;
1038 return ctx->ring_sock->sk;
1043 EXPORT_SYMBOL(io_uring_get_socket);
1045 #define io_for_each_link(pos, head) \
1046 for (pos = (head); pos; pos = pos->link)
1048 static inline void io_clean_op(struct io_kiocb *req)
1050 if (req->flags & (REQ_F_NEED_CLEANUP | REQ_F_BUFFER_SELECTED))
1054 static inline void io_set_resource_node(struct io_kiocb *req)
1056 struct io_ring_ctx *ctx = req->ctx;
1058 if (!req->fixed_rsrc_refs) {
1059 req->fixed_rsrc_refs = &ctx->file_data->node->refs;
1060 percpu_ref_get(req->fixed_rsrc_refs);
1064 static bool io_match_task(struct io_kiocb *head,
1065 struct task_struct *task,
1066 struct files_struct *files)
1068 struct io_kiocb *req;
1070 if (task && head->task != task) {
1071 /* in terms of cancelation, always match if req task is dead */
1072 if (head->task->flags & PF_EXITING)
1079 io_for_each_link(req, head) {
1080 if (req->flags & REQ_F_INFLIGHT)
1082 if (req->task->files == files)
1088 static inline void req_set_fail_links(struct io_kiocb *req)
1090 if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
1091 req->flags |= REQ_F_FAIL_LINK;
1094 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1096 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1098 complete(&ctx->ref_comp);
1101 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1103 return !req->timeout.off;
1106 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1108 struct io_ring_ctx *ctx;
1111 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1116 * Use 5 bits less than the max cq entries, that should give us around
1117 * 32 entries per hash list if totally full and uniformly spread.
1119 hash_bits = ilog2(p->cq_entries);
1123 ctx->cancel_hash_bits = hash_bits;
1124 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1126 if (!ctx->cancel_hash)
1128 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1130 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1131 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1134 ctx->flags = p->flags;
1135 init_waitqueue_head(&ctx->sqo_sq_wait);
1136 INIT_LIST_HEAD(&ctx->sqd_list);
1137 init_waitqueue_head(&ctx->cq_wait);
1138 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1139 init_completion(&ctx->ref_comp);
1140 init_completion(&ctx->sq_thread_comp);
1141 idr_init(&ctx->io_buffer_idr);
1142 xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
1143 mutex_init(&ctx->uring_lock);
1144 init_waitqueue_head(&ctx->wait);
1145 spin_lock_init(&ctx->completion_lock);
1146 INIT_LIST_HEAD(&ctx->iopoll_list);
1147 INIT_LIST_HEAD(&ctx->defer_list);
1148 INIT_LIST_HEAD(&ctx->timeout_list);
1149 spin_lock_init(&ctx->inflight_lock);
1150 INIT_LIST_HEAD(&ctx->inflight_list);
1151 spin_lock_init(&ctx->rsrc_ref_lock);
1152 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
1153 INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
1154 init_llist_head(&ctx->rsrc_put_llist);
1155 INIT_LIST_HEAD(&ctx->tctx_list);
1156 INIT_LIST_HEAD(&ctx->submit_state.comp.free_list);
1157 INIT_LIST_HEAD(&ctx->submit_state.comp.locked_free_list);
1160 kfree(ctx->cancel_hash);
1165 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1167 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1168 struct io_ring_ctx *ctx = req->ctx;
1170 return seq != ctx->cached_cq_tail
1171 + READ_ONCE(ctx->cached_cq_overflow);
1177 static void io_req_track_inflight(struct io_kiocb *req)
1179 struct io_ring_ctx *ctx = req->ctx;
1181 if (!(req->flags & REQ_F_INFLIGHT)) {
1182 req->flags |= REQ_F_INFLIGHT;
1184 spin_lock_irq(&ctx->inflight_lock);
1185 list_add(&req->inflight_entry, &ctx->inflight_list);
1186 spin_unlock_irq(&ctx->inflight_lock);
1190 static void io_prep_async_work(struct io_kiocb *req)
1192 const struct io_op_def *def = &io_op_defs[req->opcode];
1193 struct io_ring_ctx *ctx = req->ctx;
1195 if (!req->work.creds)
1196 req->work.creds = get_current_cred();
1198 if (req->flags & REQ_F_FORCE_ASYNC)
1199 req->work.flags |= IO_WQ_WORK_CONCURRENT;
1201 if (req->flags & REQ_F_ISREG) {
1202 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1203 io_wq_hash_work(&req->work, file_inode(req->file));
1205 if (def->unbound_nonreg_file)
1206 req->work.flags |= IO_WQ_WORK_UNBOUND;
1210 static void io_prep_async_link(struct io_kiocb *req)
1212 struct io_kiocb *cur;
1214 io_for_each_link(cur, req)
1215 io_prep_async_work(cur);
1218 static void io_queue_async_work(struct io_kiocb *req)
1220 struct io_ring_ctx *ctx = req->ctx;
1221 struct io_kiocb *link = io_prep_linked_timeout(req);
1222 struct io_uring_task *tctx = req->task->io_uring;
1225 BUG_ON(!tctx->io_wq);
1227 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1228 &req->work, req->flags);
1229 /* init ->work of the whole link before punting */
1230 io_prep_async_link(req);
1231 io_wq_enqueue(tctx->io_wq, &req->work);
1233 io_queue_linked_timeout(link);
1236 static void io_kill_timeout(struct io_kiocb *req)
1238 struct io_timeout_data *io = req->async_data;
1241 ret = hrtimer_try_to_cancel(&io->timer);
1243 atomic_set(&req->ctx->cq_timeouts,
1244 atomic_read(&req->ctx->cq_timeouts) + 1);
1245 list_del_init(&req->timeout.list);
1246 io_cqring_fill_event(req, 0);
1247 io_put_req_deferred(req, 1);
1252 * Returns true if we found and killed one or more timeouts
1254 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
1255 struct files_struct *files)
1257 struct io_kiocb *req, *tmp;
1260 spin_lock_irq(&ctx->completion_lock);
1261 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
1262 if (io_match_task(req, tsk, files)) {
1263 io_kill_timeout(req);
1267 spin_unlock_irq(&ctx->completion_lock);
1268 return canceled != 0;
1271 static void __io_queue_deferred(struct io_ring_ctx *ctx)
1274 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1275 struct io_defer_entry, list);
1277 if (req_need_defer(de->req, de->seq))
1279 list_del_init(&de->list);
1280 io_req_task_queue(de->req);
1282 } while (!list_empty(&ctx->defer_list));
1285 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1289 if (list_empty(&ctx->timeout_list))
1292 seq = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
1295 u32 events_needed, events_got;
1296 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1297 struct io_kiocb, timeout.list);
1299 if (io_is_timeout_noseq(req))
1303 * Since seq can easily wrap around over time, subtract
1304 * the last seq at which timeouts were flushed before comparing.
1305 * Assuming not more than 2^31-1 events have happened since,
1306 * these subtractions won't have wrapped, so we can check if
1307 * target is in [last_seq, current_seq] by comparing the two.
1309 events_needed = req->timeout.target_seq - ctx->cq_last_tm_flush;
1310 events_got = seq - ctx->cq_last_tm_flush;
1311 if (events_got < events_needed)
1314 list_del_init(&req->timeout.list);
1315 io_kill_timeout(req);
1316 } while (!list_empty(&ctx->timeout_list));
1318 ctx->cq_last_tm_flush = seq;
1321 static void io_commit_cqring(struct io_ring_ctx *ctx)
1323 io_flush_timeouts(ctx);
1325 /* order cqe stores with ring update */
1326 smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
1328 if (unlikely(!list_empty(&ctx->defer_list)))
1329 __io_queue_deferred(ctx);
1332 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1334 struct io_rings *r = ctx->rings;
1336 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == r->sq_ring_entries;
1339 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
1341 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
1344 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1346 struct io_rings *rings = ctx->rings;
1350 * writes to the cq entry need to come after reading head; the
1351 * control dependency is enough as we're using WRITE_ONCE to
1354 if (__io_cqring_events(ctx) == rings->cq_ring_entries)
1357 tail = ctx->cached_cq_tail++;
1358 return &rings->cqes[tail & ctx->cq_mask];
1361 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1365 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1367 if (!ctx->eventfd_async)
1369 return io_wq_current_is_worker();
1372 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1374 /* see waitqueue_active() comment */
1377 if (waitqueue_active(&ctx->wait))
1378 wake_up(&ctx->wait);
1379 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1380 wake_up(&ctx->sq_data->wait);
1381 if (io_should_trigger_evfd(ctx))
1382 eventfd_signal(ctx->cq_ev_fd, 1);
1383 if (waitqueue_active(&ctx->cq_wait)) {
1384 wake_up_interruptible(&ctx->cq_wait);
1385 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1389 static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx)
1391 /* see waitqueue_active() comment */
1394 if (ctx->flags & IORING_SETUP_SQPOLL) {
1395 if (waitqueue_active(&ctx->wait))
1396 wake_up(&ctx->wait);
1398 if (io_should_trigger_evfd(ctx))
1399 eventfd_signal(ctx->cq_ev_fd, 1);
1400 if (waitqueue_active(&ctx->cq_wait)) {
1401 wake_up_interruptible(&ctx->cq_wait);
1402 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1406 /* Returns true if there are no backlogged entries after the flush */
1407 static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force,
1408 struct task_struct *tsk,
1409 struct files_struct *files)
1411 struct io_rings *rings = ctx->rings;
1412 struct io_kiocb *req, *tmp;
1413 struct io_uring_cqe *cqe;
1414 unsigned long flags;
1415 bool all_flushed, posted;
1418 if (!force && __io_cqring_events(ctx) == rings->cq_ring_entries)
1422 spin_lock_irqsave(&ctx->completion_lock, flags);
1423 list_for_each_entry_safe(req, tmp, &ctx->cq_overflow_list, compl.list) {
1424 if (!io_match_task(req, tsk, files))
1427 cqe = io_get_cqring(ctx);
1431 list_move(&req->compl.list, &list);
1433 WRITE_ONCE(cqe->user_data, req->user_data);
1434 WRITE_ONCE(cqe->res, req->result);
1435 WRITE_ONCE(cqe->flags, req->compl.cflags);
1437 ctx->cached_cq_overflow++;
1438 WRITE_ONCE(ctx->rings->cq_overflow,
1439 ctx->cached_cq_overflow);
1444 all_flushed = list_empty(&ctx->cq_overflow_list);
1446 clear_bit(0, &ctx->sq_check_overflow);
1447 clear_bit(0, &ctx->cq_check_overflow);
1448 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
1452 io_commit_cqring(ctx);
1453 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1455 io_cqring_ev_posted(ctx);
1457 while (!list_empty(&list)) {
1458 req = list_first_entry(&list, struct io_kiocb, compl.list);
1459 list_del(&req->compl.list);
1466 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force,
1467 struct task_struct *tsk,
1468 struct files_struct *files)
1472 if (test_bit(0, &ctx->cq_check_overflow)) {
1473 /* iopoll syncs against uring_lock, not completion_lock */
1474 if (ctx->flags & IORING_SETUP_IOPOLL)
1475 mutex_lock(&ctx->uring_lock);
1476 ret = __io_cqring_overflow_flush(ctx, force, tsk, files);
1477 if (ctx->flags & IORING_SETUP_IOPOLL)
1478 mutex_unlock(&ctx->uring_lock);
1484 static void __io_cqring_fill_event(struct io_kiocb *req, long res, long cflags)
1486 struct io_ring_ctx *ctx = req->ctx;
1487 struct io_uring_cqe *cqe;
1489 trace_io_uring_complete(ctx, req->user_data, res);
1492 * If we can't get a cq entry, userspace overflowed the
1493 * submission (by quite a lot). Increment the overflow count in
1496 cqe = io_get_cqring(ctx);
1498 WRITE_ONCE(cqe->user_data, req->user_data);
1499 WRITE_ONCE(cqe->res, res);
1500 WRITE_ONCE(cqe->flags, cflags);
1501 } else if (ctx->cq_overflow_flushed ||
1502 atomic_read(&req->task->io_uring->in_idle)) {
1504 * If we're in ring overflow flush mode, or in task cancel mode,
1505 * then we cannot store the request for later flushing, we need
1506 * to drop it on the floor.
1508 ctx->cached_cq_overflow++;
1509 WRITE_ONCE(ctx->rings->cq_overflow, ctx->cached_cq_overflow);
1511 if (list_empty(&ctx->cq_overflow_list)) {
1512 set_bit(0, &ctx->sq_check_overflow);
1513 set_bit(0, &ctx->cq_check_overflow);
1514 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
1518 req->compl.cflags = cflags;
1519 refcount_inc(&req->refs);
1520 list_add_tail(&req->compl.list, &ctx->cq_overflow_list);
1524 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1526 __io_cqring_fill_event(req, res, 0);
1529 static void io_req_complete_post(struct io_kiocb *req, long res,
1530 unsigned int cflags)
1532 struct io_ring_ctx *ctx = req->ctx;
1533 unsigned long flags;
1535 spin_lock_irqsave(&ctx->completion_lock, flags);
1536 __io_cqring_fill_event(req, res, cflags);
1538 * If we're the last reference to this request, add to our locked
1541 if (refcount_dec_and_test(&req->refs)) {
1542 struct io_comp_state *cs = &ctx->submit_state.comp;
1544 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
1545 if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL_LINK))
1546 io_disarm_next(req);
1548 io_req_task_queue(req->link);
1552 io_dismantle_req(req);
1553 io_put_task(req->task, 1);
1554 list_add(&req->compl.list, &cs->locked_free_list);
1555 cs->locked_free_nr++;
1558 io_commit_cqring(ctx);
1559 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1560 io_cqring_ev_posted(ctx);
1563 percpu_ref_put(&ctx->refs);
1566 static void io_req_complete_state(struct io_kiocb *req, long res,
1567 unsigned int cflags)
1571 req->compl.cflags = cflags;
1572 req->flags |= REQ_F_COMPLETE_INLINE;
1575 static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags,
1576 long res, unsigned cflags)
1578 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1579 io_req_complete_state(req, res, cflags);
1581 io_req_complete_post(req, res, cflags);
1584 static inline void io_req_complete(struct io_kiocb *req, long res)
1586 __io_req_complete(req, 0, res, 0);
1589 static bool io_flush_cached_reqs(struct io_ring_ctx *ctx)
1591 struct io_submit_state *state = &ctx->submit_state;
1592 struct io_comp_state *cs = &state->comp;
1593 struct io_kiocb *req = NULL;
1596 * If we have more than a batch's worth of requests in our IRQ side
1597 * locked cache, grab the lock and move them over to our submission
1600 if (READ_ONCE(cs->locked_free_nr) > IO_COMPL_BATCH) {
1601 spin_lock_irq(&ctx->completion_lock);
1602 list_splice_init(&cs->locked_free_list, &cs->free_list);
1603 cs->locked_free_nr = 0;
1604 spin_unlock_irq(&ctx->completion_lock);
1607 while (!list_empty(&cs->free_list)) {
1608 req = list_first_entry(&cs->free_list, struct io_kiocb,
1610 list_del(&req->compl.list);
1611 state->reqs[state->free_reqs++] = req;
1612 if (state->free_reqs == ARRAY_SIZE(state->reqs))
1619 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
1621 struct io_submit_state *state = &ctx->submit_state;
1623 BUILD_BUG_ON(IO_REQ_ALLOC_BATCH > ARRAY_SIZE(state->reqs));
1625 if (!state->free_reqs) {
1626 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1629 if (io_flush_cached_reqs(ctx))
1632 ret = kmem_cache_alloc_bulk(req_cachep, gfp, IO_REQ_ALLOC_BATCH,
1636 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1637 * retry single alloc to be on the safe side.
1639 if (unlikely(ret <= 0)) {
1640 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1641 if (!state->reqs[0])
1645 state->free_reqs = ret;
1649 return state->reqs[state->free_reqs];
1652 static inline void io_put_file(struct io_kiocb *req, struct file *file,
1659 static void io_dismantle_req(struct io_kiocb *req)
1663 if (req->async_data)
1664 kfree(req->async_data);
1666 io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
1667 if (req->fixed_rsrc_refs)
1668 percpu_ref_put(req->fixed_rsrc_refs);
1669 if (req->work.creds) {
1670 put_cred(req->work.creds);
1671 req->work.creds = NULL;
1674 if (req->flags & REQ_F_INFLIGHT) {
1675 struct io_ring_ctx *ctx = req->ctx;
1676 unsigned long flags;
1678 spin_lock_irqsave(&ctx->inflight_lock, flags);
1679 list_del(&req->inflight_entry);
1680 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
1681 req->flags &= ~REQ_F_INFLIGHT;
1685 /* must to be called somewhat shortly after putting a request */
1686 static inline void io_put_task(struct task_struct *task, int nr)
1688 struct io_uring_task *tctx = task->io_uring;
1690 percpu_counter_sub(&tctx->inflight, nr);
1691 if (unlikely(atomic_read(&tctx->in_idle)))
1692 wake_up(&tctx->wait);
1693 put_task_struct_many(task, nr);
1696 static void __io_free_req(struct io_kiocb *req)
1698 struct io_ring_ctx *ctx = req->ctx;
1700 io_dismantle_req(req);
1701 io_put_task(req->task, 1);
1703 kmem_cache_free(req_cachep, req);
1704 percpu_ref_put(&ctx->refs);
1707 static inline void io_remove_next_linked(struct io_kiocb *req)
1709 struct io_kiocb *nxt = req->link;
1711 req->link = nxt->link;
1715 static bool io_kill_linked_timeout(struct io_kiocb *req)
1716 __must_hold(&req->ctx->completion_lock)
1718 struct io_kiocb *link = req->link;
1719 bool cancelled = false;
1722 * Can happen if a linked timeout fired and link had been like
1723 * req -> link t-out -> link t-out [-> ...]
1725 if (link && (link->flags & REQ_F_LTIMEOUT_ACTIVE)) {
1726 struct io_timeout_data *io = link->async_data;
1729 io_remove_next_linked(req);
1730 link->timeout.head = NULL;
1731 ret = hrtimer_try_to_cancel(&io->timer);
1733 io_cqring_fill_event(link, -ECANCELED);
1734 io_put_req_deferred(link, 1);
1738 req->flags &= ~REQ_F_LINK_TIMEOUT;
1742 static void io_fail_links(struct io_kiocb *req)
1743 __must_hold(&req->ctx->completion_lock)
1745 struct io_kiocb *nxt, *link = req->link;
1752 trace_io_uring_fail_link(req, link);
1753 io_cqring_fill_event(link, -ECANCELED);
1754 io_put_req_deferred(link, 2);
1759 static bool io_disarm_next(struct io_kiocb *req)
1760 __must_hold(&req->ctx->completion_lock)
1762 bool posted = false;
1764 if (likely(req->flags & REQ_F_LINK_TIMEOUT))
1765 posted = io_kill_linked_timeout(req);
1766 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
1767 posted |= (req->link != NULL);
1773 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1775 struct io_kiocb *nxt;
1778 * If LINK is set, we have dependent requests in this chain. If we
1779 * didn't fail this request, queue the first one up, moving any other
1780 * dependencies to the next request. In case of failure, fail the rest
1783 if (req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_FAIL_LINK)) {
1784 struct io_ring_ctx *ctx = req->ctx;
1785 unsigned long flags;
1788 spin_lock_irqsave(&ctx->completion_lock, flags);
1789 posted = io_disarm_next(req);
1791 io_commit_cqring(req->ctx);
1792 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1794 io_cqring_ev_posted(ctx);
1801 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1803 if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK))))
1805 return __io_req_find_next(req);
1808 static void ctx_flush_and_put(struct io_ring_ctx *ctx)
1812 if (ctx->submit_state.comp.nr) {
1813 mutex_lock(&ctx->uring_lock);
1814 io_submit_flush_completions(&ctx->submit_state.comp, ctx);
1815 mutex_unlock(&ctx->uring_lock);
1817 percpu_ref_put(&ctx->refs);
1820 static bool __tctx_task_work(struct io_uring_task *tctx)
1822 struct io_ring_ctx *ctx = NULL;
1823 struct io_wq_work_list list;
1824 struct io_wq_work_node *node;
1826 if (wq_list_empty(&tctx->task_list))
1829 spin_lock_irq(&tctx->task_lock);
1830 list = tctx->task_list;
1831 INIT_WQ_LIST(&tctx->task_list);
1832 spin_unlock_irq(&tctx->task_lock);
1836 struct io_wq_work_node *next = node->next;
1837 struct io_kiocb *req;
1839 req = container_of(node, struct io_kiocb, io_task_work.node);
1840 if (req->ctx != ctx) {
1841 ctx_flush_and_put(ctx);
1843 percpu_ref_get(&ctx->refs);
1846 req->task_work.func(&req->task_work);
1850 ctx_flush_and_put(ctx);
1851 return list.first != NULL;
1854 static void tctx_task_work(struct callback_head *cb)
1856 struct io_uring_task *tctx = container_of(cb, struct io_uring_task, task_work);
1858 clear_bit(0, &tctx->task_state);
1860 while (__tctx_task_work(tctx))
1864 static int io_task_work_add(struct task_struct *tsk, struct io_kiocb *req,
1865 enum task_work_notify_mode notify)
1867 struct io_uring_task *tctx = tsk->io_uring;
1868 struct io_wq_work_node *node, *prev;
1869 unsigned long flags;
1872 WARN_ON_ONCE(!tctx);
1874 spin_lock_irqsave(&tctx->task_lock, flags);
1875 wq_list_add_tail(&req->io_task_work.node, &tctx->task_list);
1876 spin_unlock_irqrestore(&tctx->task_lock, flags);
1878 /* task_work already pending, we're done */
1879 if (test_bit(0, &tctx->task_state) ||
1880 test_and_set_bit(0, &tctx->task_state))
1883 if (!task_work_add(tsk, &tctx->task_work, notify))
1887 * Slow path - we failed, find and delete work. if the work is not
1888 * in the list, it got run and we're fine.
1891 spin_lock_irqsave(&tctx->task_lock, flags);
1892 wq_list_for_each(node, prev, &tctx->task_list) {
1893 if (&req->io_task_work.node == node) {
1894 wq_list_del(&tctx->task_list, node, prev);
1899 spin_unlock_irqrestore(&tctx->task_lock, flags);
1900 clear_bit(0, &tctx->task_state);
1904 static int io_req_task_work_add(struct io_kiocb *req)
1906 struct task_struct *tsk = req->task;
1907 struct io_ring_ctx *ctx = req->ctx;
1908 enum task_work_notify_mode notify;
1911 if (tsk->flags & PF_EXITING)
1915 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
1916 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
1917 * processing task_work. There's no reliable way to tell if TWA_RESUME
1921 if (!(ctx->flags & IORING_SETUP_SQPOLL))
1922 notify = TWA_SIGNAL;
1924 ret = io_task_work_add(tsk, req, notify);
1926 wake_up_process(tsk);
1931 static void io_req_task_work_add_fallback(struct io_kiocb *req,
1932 task_work_func_t cb)
1934 struct io_ring_ctx *ctx = req->ctx;
1935 struct callback_head *head;
1937 init_task_work(&req->task_work, cb);
1939 head = READ_ONCE(ctx->exit_task_work);
1940 req->task_work.next = head;
1941 } while (cmpxchg(&ctx->exit_task_work, head, &req->task_work) != head);
1944 static void __io_req_task_cancel(struct io_kiocb *req, int error)
1946 struct io_ring_ctx *ctx = req->ctx;
1948 spin_lock_irq(&ctx->completion_lock);
1949 io_cqring_fill_event(req, error);
1950 io_commit_cqring(ctx);
1951 spin_unlock_irq(&ctx->completion_lock);
1953 io_cqring_ev_posted(ctx);
1954 req_set_fail_links(req);
1955 io_double_put_req(req);
1958 static void io_req_task_cancel(struct callback_head *cb)
1960 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1961 struct io_ring_ctx *ctx = req->ctx;
1963 mutex_lock(&ctx->uring_lock);
1964 __io_req_task_cancel(req, req->result);
1965 mutex_unlock(&ctx->uring_lock);
1966 percpu_ref_put(&ctx->refs);
1969 static void __io_req_task_submit(struct io_kiocb *req)
1971 struct io_ring_ctx *ctx = req->ctx;
1973 /* ctx stays valid until unlock, even if we drop all ours ctx->refs */
1974 mutex_lock(&ctx->uring_lock);
1975 if (!(current->flags & PF_EXITING) && !current->in_execve)
1976 __io_queue_sqe(req);
1978 __io_req_task_cancel(req, -EFAULT);
1979 mutex_unlock(&ctx->uring_lock);
1982 static void io_req_task_submit(struct callback_head *cb)
1984 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1986 __io_req_task_submit(req);
1989 static void io_req_task_queue(struct io_kiocb *req)
1993 req->task_work.func = io_req_task_submit;
1994 ret = io_req_task_work_add(req);
1995 if (unlikely(ret)) {
1996 req->result = -ECANCELED;
1997 percpu_ref_get(&req->ctx->refs);
1998 io_req_task_work_add_fallback(req, io_req_task_cancel);
2002 static void io_req_task_queue_fail(struct io_kiocb *req, int ret)
2004 percpu_ref_get(&req->ctx->refs);
2006 req->task_work.func = io_req_task_cancel;
2008 if (unlikely(io_req_task_work_add(req)))
2009 io_req_task_work_add_fallback(req, io_req_task_cancel);
2012 static inline void io_queue_next(struct io_kiocb *req)
2014 struct io_kiocb *nxt = io_req_find_next(req);
2017 io_req_task_queue(nxt);
2020 static void io_free_req(struct io_kiocb *req)
2027 struct task_struct *task;
2032 static inline void io_init_req_batch(struct req_batch *rb)
2039 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2040 struct req_batch *rb)
2043 io_put_task(rb->task, rb->task_refs);
2045 percpu_ref_put_many(&ctx->refs, rb->ctx_refs);
2048 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req,
2049 struct io_submit_state *state)
2053 if (req->task != rb->task) {
2055 io_put_task(rb->task, rb->task_refs);
2056 rb->task = req->task;
2062 io_dismantle_req(req);
2063 if (state->free_reqs != ARRAY_SIZE(state->reqs))
2064 state->reqs[state->free_reqs++] = req;
2066 list_add(&req->compl.list, &state->comp.free_list);
2069 static void io_submit_flush_completions(struct io_comp_state *cs,
2070 struct io_ring_ctx *ctx)
2073 struct io_kiocb *req;
2074 struct req_batch rb;
2076 io_init_req_batch(&rb);
2077 spin_lock_irq(&ctx->completion_lock);
2078 for (i = 0; i < nr; i++) {
2080 __io_cqring_fill_event(req, req->result, req->compl.cflags);
2082 io_commit_cqring(ctx);
2083 spin_unlock_irq(&ctx->completion_lock);
2085 io_cqring_ev_posted(ctx);
2086 for (i = 0; i < nr; i++) {
2089 /* submission and completion refs */
2090 if (refcount_sub_and_test(2, &req->refs))
2091 io_req_free_batch(&rb, req, &ctx->submit_state);
2094 io_req_free_batch_finish(ctx, &rb);
2099 * Drop reference to request, return next in chain (if there is one) if this
2100 * was the last reference to this request.
2102 static struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2104 struct io_kiocb *nxt = NULL;
2106 if (refcount_dec_and_test(&req->refs)) {
2107 nxt = io_req_find_next(req);
2113 static void io_put_req(struct io_kiocb *req)
2115 if (refcount_dec_and_test(&req->refs))
2119 static void io_put_req_deferred_cb(struct callback_head *cb)
2121 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2126 static void io_free_req_deferred(struct io_kiocb *req)
2130 req->task_work.func = io_put_req_deferred_cb;
2131 ret = io_req_task_work_add(req);
2133 io_req_task_work_add_fallback(req, io_put_req_deferred_cb);
2136 static inline void io_put_req_deferred(struct io_kiocb *req, int refs)
2138 if (refcount_sub_and_test(refs, &req->refs))
2139 io_free_req_deferred(req);
2142 static void io_double_put_req(struct io_kiocb *req)
2144 /* drop both submit and complete references */
2145 if (refcount_sub_and_test(2, &req->refs))
2149 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2151 /* See comment at the top of this file */
2153 return __io_cqring_events(ctx);
2156 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2158 struct io_rings *rings = ctx->rings;
2160 /* make sure SQ entry isn't read before tail */
2161 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2164 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2166 unsigned int cflags;
2168 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2169 cflags |= IORING_CQE_F_BUFFER;
2170 req->flags &= ~REQ_F_BUFFER_SELECTED;
2175 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2177 struct io_buffer *kbuf;
2179 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2180 return io_put_kbuf(req, kbuf);
2183 static inline bool io_run_task_work(void)
2186 * Not safe to run on exiting task, and the task_work handling will
2187 * not add work to such a task.
2189 if (unlikely(current->flags & PF_EXITING))
2191 if (current->task_works) {
2192 __set_current_state(TASK_RUNNING);
2201 * Find and free completed poll iocbs
2203 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2204 struct list_head *done)
2206 struct req_batch rb;
2207 struct io_kiocb *req;
2209 /* order with ->result store in io_complete_rw_iopoll() */
2212 io_init_req_batch(&rb);
2213 while (!list_empty(done)) {
2216 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2217 list_del(&req->inflight_entry);
2219 if (READ_ONCE(req->result) == -EAGAIN) {
2220 req->iopoll_completed = 0;
2221 if (io_rw_reissue(req))
2225 if (req->flags & REQ_F_BUFFER_SELECTED)
2226 cflags = io_put_rw_kbuf(req);
2228 __io_cqring_fill_event(req, req->result, cflags);
2231 if (refcount_dec_and_test(&req->refs))
2232 io_req_free_batch(&rb, req, &ctx->submit_state);
2235 io_commit_cqring(ctx);
2236 io_cqring_ev_posted_iopoll(ctx);
2237 io_req_free_batch_finish(ctx, &rb);
2240 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2243 struct io_kiocb *req, *tmp;
2249 * Only spin for completions if we don't have multiple devices hanging
2250 * off our complete list, and we're under the requested amount.
2252 spin = !ctx->poll_multi_file && *nr_events < min;
2255 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2256 struct kiocb *kiocb = &req->rw.kiocb;
2259 * Move completed and retryable entries to our local lists.
2260 * If we find a request that requires polling, break out
2261 * and complete those lists first, if we have entries there.
2263 if (READ_ONCE(req->iopoll_completed)) {
2264 list_move_tail(&req->inflight_entry, &done);
2267 if (!list_empty(&done))
2270 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2274 /* iopoll may have completed current req */
2275 if (READ_ONCE(req->iopoll_completed))
2276 list_move_tail(&req->inflight_entry, &done);
2283 if (!list_empty(&done))
2284 io_iopoll_complete(ctx, nr_events, &done);
2290 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2291 * non-spinning poll check - we'll still enter the driver poll loop, but only
2292 * as a non-spinning completion check.
2294 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
2297 while (!list_empty(&ctx->iopoll_list) && !need_resched()) {
2300 ret = io_do_iopoll(ctx, nr_events, min);
2303 if (*nr_events >= min)
2311 * We can't just wait for polled events to come to us, we have to actively
2312 * find and complete them.
2314 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2316 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2319 mutex_lock(&ctx->uring_lock);
2320 while (!list_empty(&ctx->iopoll_list)) {
2321 unsigned int nr_events = 0;
2323 io_do_iopoll(ctx, &nr_events, 0);
2325 /* let it sleep and repeat later if can't complete a request */
2329 * Ensure we allow local-to-the-cpu processing to take place,
2330 * in this case we need to ensure that we reap all events.
2331 * Also let task_work, etc. to progress by releasing the mutex
2333 if (need_resched()) {
2334 mutex_unlock(&ctx->uring_lock);
2336 mutex_lock(&ctx->uring_lock);
2339 mutex_unlock(&ctx->uring_lock);
2342 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2344 unsigned int nr_events = 0;
2345 int iters = 0, ret = 0;
2348 * We disallow the app entering submit/complete with polling, but we
2349 * still need to lock the ring to prevent racing with polled issue
2350 * that got punted to a workqueue.
2352 mutex_lock(&ctx->uring_lock);
2355 * Don't enter poll loop if we already have events pending.
2356 * If we do, we can potentially be spinning for commands that
2357 * already triggered a CQE (eg in error).
2359 if (test_bit(0, &ctx->cq_check_overflow))
2360 __io_cqring_overflow_flush(ctx, false, NULL, NULL);
2361 if (io_cqring_events(ctx))
2365 * If a submit got punted to a workqueue, we can have the
2366 * application entering polling for a command before it gets
2367 * issued. That app will hold the uring_lock for the duration
2368 * of the poll right here, so we need to take a breather every
2369 * now and then to ensure that the issue has a chance to add
2370 * the poll to the issued list. Otherwise we can spin here
2371 * forever, while the workqueue is stuck trying to acquire the
2374 if (!(++iters & 7)) {
2375 mutex_unlock(&ctx->uring_lock);
2377 mutex_lock(&ctx->uring_lock);
2380 ret = io_iopoll_getevents(ctx, &nr_events, min);
2384 } while (min && !nr_events && !need_resched());
2386 mutex_unlock(&ctx->uring_lock);
2390 static void kiocb_end_write(struct io_kiocb *req)
2393 * Tell lockdep we inherited freeze protection from submission
2396 if (req->flags & REQ_F_ISREG) {
2397 struct inode *inode = file_inode(req->file);
2399 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
2401 file_end_write(req->file);
2405 static bool io_resubmit_prep(struct io_kiocb *req)
2407 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2409 struct iov_iter iter;
2411 /* already prepared */
2412 if (req->async_data)
2415 switch (req->opcode) {
2416 case IORING_OP_READV:
2417 case IORING_OP_READ_FIXED:
2418 case IORING_OP_READ:
2421 case IORING_OP_WRITEV:
2422 case IORING_OP_WRITE_FIXED:
2423 case IORING_OP_WRITE:
2427 printk_once(KERN_WARNING "io_uring: bad opcode in resubmit %d\n",
2432 ret = io_import_iovec(rw, req, &iovec, &iter, false);
2435 return !io_setup_async_rw(req, iovec, inline_vecs, &iter, false);
2438 static bool io_rw_should_reissue(struct io_kiocb *req)
2440 umode_t mode = file_inode(req->file)->i_mode;
2441 struct io_ring_ctx *ctx = req->ctx;
2443 if (!S_ISBLK(mode) && !S_ISREG(mode))
2445 if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
2446 !(ctx->flags & IORING_SETUP_IOPOLL)))
2449 * If ref is dying, we might be running poll reap from the exit work.
2450 * Don't attempt to reissue from that path, just let it fail with
2453 if (percpu_ref_is_dying(&ctx->refs))
2459 static bool io_rw_reissue(struct io_kiocb *req)
2462 if (!io_rw_should_reissue(req))
2465 lockdep_assert_held(&req->ctx->uring_lock);
2467 if (io_resubmit_prep(req)) {
2468 refcount_inc(&req->refs);
2469 io_queue_async_work(req);
2472 req_set_fail_links(req);
2477 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2478 unsigned int issue_flags)
2482 if ((res == -EAGAIN || res == -EOPNOTSUPP) && io_rw_reissue(req))
2484 if (res != req->result)
2485 req_set_fail_links(req);
2487 if (req->rw.kiocb.ki_flags & IOCB_WRITE)
2488 kiocb_end_write(req);
2489 if (req->flags & REQ_F_BUFFER_SELECTED)
2490 cflags = io_put_rw_kbuf(req);
2491 __io_req_complete(req, issue_flags, res, cflags);
2494 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2496 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2498 __io_complete_rw(req, res, res2, 0);
2501 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2503 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2506 /* Rewind iter, if we have one. iopoll path resubmits as usual */
2507 if (res == -EAGAIN && io_rw_should_reissue(req)) {
2508 struct io_async_rw *rw = req->async_data;
2511 iov_iter_revert(&rw->iter,
2512 req->result - iov_iter_count(&rw->iter));
2513 else if (!io_resubmit_prep(req))
2518 if (kiocb->ki_flags & IOCB_WRITE)
2519 kiocb_end_write(req);
2521 if (res != -EAGAIN && res != req->result)
2522 req_set_fail_links(req);
2524 WRITE_ONCE(req->result, res);
2525 /* order with io_poll_complete() checking ->result */
2527 WRITE_ONCE(req->iopoll_completed, 1);
2531 * After the iocb has been issued, it's safe to be found on the poll list.
2532 * Adding the kiocb to the list AFTER submission ensures that we don't
2533 * find it from a io_iopoll_getevents() thread before the issuer is done
2534 * accessing the kiocb cookie.
2536 static void io_iopoll_req_issued(struct io_kiocb *req, bool in_async)
2538 struct io_ring_ctx *ctx = req->ctx;
2541 * Track whether we have multiple files in our lists. This will impact
2542 * how we do polling eventually, not spinning if we're on potentially
2543 * different devices.
2545 if (list_empty(&ctx->iopoll_list)) {
2546 ctx->poll_multi_file = false;
2547 } else if (!ctx->poll_multi_file) {
2548 struct io_kiocb *list_req;
2550 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2552 if (list_req->file != req->file)
2553 ctx->poll_multi_file = true;
2557 * For fast devices, IO may have already completed. If it has, add
2558 * it to the front so we find it first.
2560 if (READ_ONCE(req->iopoll_completed))
2561 list_add(&req->inflight_entry, &ctx->iopoll_list);
2563 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2566 * If IORING_SETUP_SQPOLL is enabled, sqes are either handled in sq thread
2567 * task context or in io worker task context. If current task context is
2568 * sq thread, we don't need to check whether should wake up sq thread.
2570 if (in_async && (ctx->flags & IORING_SETUP_SQPOLL) &&
2571 wq_has_sleeper(&ctx->sq_data->wait))
2572 wake_up(&ctx->sq_data->wait);
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)))
2629 if (S_ISCHR(mode) || S_ISSOCK(mode))
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 int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2654 struct io_ring_ctx *ctx = req->ctx;
2655 struct kiocb *kiocb = &req->rw.kiocb;
2656 struct file *file = req->file;
2660 if (S_ISREG(file_inode(file)->i_mode))
2661 req->flags |= REQ_F_ISREG;
2663 kiocb->ki_pos = READ_ONCE(sqe->off);
2664 if (kiocb->ki_pos == -1 && !(file->f_mode & FMODE_STREAM)) {
2665 req->flags |= REQ_F_CUR_POS;
2666 kiocb->ki_pos = file->f_pos;
2668 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2669 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2670 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2674 /* don't allow async punt for O_NONBLOCK or RWF_NOWAIT */
2675 if ((kiocb->ki_flags & IOCB_NOWAIT) || (file->f_flags & O_NONBLOCK))
2676 req->flags |= REQ_F_NOWAIT;
2678 ioprio = READ_ONCE(sqe->ioprio);
2680 ret = ioprio_check_cap(ioprio);
2684 kiocb->ki_ioprio = ioprio;
2686 kiocb->ki_ioprio = get_current_ioprio();
2688 if (ctx->flags & IORING_SETUP_IOPOLL) {
2689 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2690 !kiocb->ki_filp->f_op->iopoll)
2693 kiocb->ki_flags |= IOCB_HIPRI;
2694 kiocb->ki_complete = io_complete_rw_iopoll;
2695 req->iopoll_completed = 0;
2697 if (kiocb->ki_flags & IOCB_HIPRI)
2699 kiocb->ki_complete = io_complete_rw;
2702 req->rw.addr = READ_ONCE(sqe->addr);
2703 req->rw.len = READ_ONCE(sqe->len);
2704 req->buf_index = READ_ONCE(sqe->buf_index);
2708 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2714 case -ERESTARTNOINTR:
2715 case -ERESTARTNOHAND:
2716 case -ERESTART_RESTARTBLOCK:
2718 * We can't just restart the syscall, since previously
2719 * submitted sqes may already be in progress. Just fail this
2725 kiocb->ki_complete(kiocb, ret, 0);
2729 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2730 unsigned int issue_flags)
2732 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2733 struct io_async_rw *io = req->async_data;
2735 /* add previously done IO, if any */
2736 if (io && io->bytes_done > 0) {
2738 ret = io->bytes_done;
2740 ret += io->bytes_done;
2743 if (req->flags & REQ_F_CUR_POS)
2744 req->file->f_pos = kiocb->ki_pos;
2745 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2746 __io_complete_rw(req, ret, 0, issue_flags);
2748 io_rw_done(kiocb, ret);
2751 static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
2753 struct io_ring_ctx *ctx = req->ctx;
2754 size_t len = req->rw.len;
2755 struct io_mapped_ubuf *imu;
2756 u16 index, buf_index = req->buf_index;
2760 if (unlikely(buf_index >= ctx->nr_user_bufs))
2762 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2763 imu = &ctx->user_bufs[index];
2764 buf_addr = req->rw.addr;
2767 if (buf_addr + len < buf_addr)
2769 /* not inside the mapped region */
2770 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2774 * May not be a start of buffer, set size appropriately
2775 * and advance us to the beginning.
2777 offset = buf_addr - imu->ubuf;
2778 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2782 * Don't use iov_iter_advance() here, as it's really slow for
2783 * using the latter parts of a big fixed buffer - it iterates
2784 * over each segment manually. We can cheat a bit here, because
2787 * 1) it's a BVEC iter, we set it up
2788 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2789 * first and last bvec
2791 * So just find our index, and adjust the iterator afterwards.
2792 * If the offset is within the first bvec (or the whole first
2793 * bvec, just use iov_iter_advance(). This makes it easier
2794 * since we can just skip the first segment, which may not
2795 * be PAGE_SIZE aligned.
2797 const struct bio_vec *bvec = imu->bvec;
2799 if (offset <= bvec->bv_len) {
2800 iov_iter_advance(iter, offset);
2802 unsigned long seg_skip;
2804 /* skip first vec */
2805 offset -= bvec->bv_len;
2806 seg_skip = 1 + (offset >> PAGE_SHIFT);
2808 iter->bvec = bvec + seg_skip;
2809 iter->nr_segs -= seg_skip;
2810 iter->count -= bvec->bv_len + offset;
2811 iter->iov_offset = offset & ~PAGE_MASK;
2818 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2821 mutex_unlock(&ctx->uring_lock);
2824 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2827 * "Normal" inline submissions always hold the uring_lock, since we
2828 * grab it from the system call. Same is true for the SQPOLL offload.
2829 * The only exception is when we've detached the request and issue it
2830 * from an async worker thread, grab the lock for that case.
2833 mutex_lock(&ctx->uring_lock);
2836 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2837 int bgid, struct io_buffer *kbuf,
2840 struct io_buffer *head;
2842 if (req->flags & REQ_F_BUFFER_SELECTED)
2845 io_ring_submit_lock(req->ctx, needs_lock);
2847 lockdep_assert_held(&req->ctx->uring_lock);
2849 head = idr_find(&req->ctx->io_buffer_idr, bgid);
2851 if (!list_empty(&head->list)) {
2852 kbuf = list_last_entry(&head->list, struct io_buffer,
2854 list_del(&kbuf->list);
2857 idr_remove(&req->ctx->io_buffer_idr, bgid);
2859 if (*len > kbuf->len)
2862 kbuf = ERR_PTR(-ENOBUFS);
2865 io_ring_submit_unlock(req->ctx, needs_lock);
2870 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2873 struct io_buffer *kbuf;
2876 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2877 bgid = req->buf_index;
2878 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2881 req->rw.addr = (u64) (unsigned long) kbuf;
2882 req->flags |= REQ_F_BUFFER_SELECTED;
2883 return u64_to_user_ptr(kbuf->addr);
2886 #ifdef CONFIG_COMPAT
2887 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2890 struct compat_iovec __user *uiov;
2891 compat_ssize_t clen;
2895 uiov = u64_to_user_ptr(req->rw.addr);
2896 if (!access_ok(uiov, sizeof(*uiov)))
2898 if (__get_user(clen, &uiov->iov_len))
2904 buf = io_rw_buffer_select(req, &len, needs_lock);
2906 return PTR_ERR(buf);
2907 iov[0].iov_base = buf;
2908 iov[0].iov_len = (compat_size_t) len;
2913 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2916 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2920 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2923 len = iov[0].iov_len;
2926 buf = io_rw_buffer_select(req, &len, needs_lock);
2928 return PTR_ERR(buf);
2929 iov[0].iov_base = buf;
2930 iov[0].iov_len = len;
2934 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2937 if (req->flags & REQ_F_BUFFER_SELECTED) {
2938 struct io_buffer *kbuf;
2940 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2941 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
2942 iov[0].iov_len = kbuf->len;
2945 if (req->rw.len != 1)
2948 #ifdef CONFIG_COMPAT
2949 if (req->ctx->compat)
2950 return io_compat_import(req, iov, needs_lock);
2953 return __io_iov_buffer_select(req, iov, needs_lock);
2956 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
2957 struct iov_iter *iter, bool needs_lock)
2959 void __user *buf = u64_to_user_ptr(req->rw.addr);
2960 size_t sqe_len = req->rw.len;
2961 u8 opcode = req->opcode;
2964 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
2966 return io_import_fixed(req, rw, iter);
2969 /* buffer index only valid with fixed read/write, or buffer select */
2970 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
2973 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
2974 if (req->flags & REQ_F_BUFFER_SELECT) {
2975 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
2977 return PTR_ERR(buf);
2978 req->rw.len = sqe_len;
2981 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
2986 if (req->flags & REQ_F_BUFFER_SELECT) {
2987 ret = io_iov_buffer_select(req, *iovec, needs_lock);
2989 iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
2994 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
2998 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3000 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3004 * For files that don't have ->read_iter() and ->write_iter(), handle them
3005 * by looping over ->read() or ->write() manually.
3007 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3009 struct kiocb *kiocb = &req->rw.kiocb;
3010 struct file *file = req->file;
3014 * Don't support polled IO through this interface, and we can't
3015 * support non-blocking either. For the latter, this just causes
3016 * the kiocb to be handled from an async context.
3018 if (kiocb->ki_flags & IOCB_HIPRI)
3020 if (kiocb->ki_flags & IOCB_NOWAIT)
3023 while (iov_iter_count(iter)) {
3027 if (!iov_iter_is_bvec(iter)) {
3028 iovec = iov_iter_iovec(iter);
3030 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3031 iovec.iov_len = req->rw.len;
3035 nr = file->f_op->read(file, iovec.iov_base,
3036 iovec.iov_len, io_kiocb_ppos(kiocb));
3038 nr = file->f_op->write(file, iovec.iov_base,
3039 iovec.iov_len, io_kiocb_ppos(kiocb));
3048 if (nr != iovec.iov_len)
3052 iov_iter_advance(iter, nr);
3058 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3059 const struct iovec *fast_iov, struct iov_iter *iter)
3061 struct io_async_rw *rw = req->async_data;
3063 memcpy(&rw->iter, iter, sizeof(*iter));
3064 rw->free_iovec = iovec;
3066 /* can only be fixed buffers, no need to do anything */
3067 if (iov_iter_is_bvec(iter))
3070 unsigned iov_off = 0;
3072 rw->iter.iov = rw->fast_iov;
3073 if (iter->iov != fast_iov) {
3074 iov_off = iter->iov - fast_iov;
3075 rw->iter.iov += iov_off;
3077 if (rw->fast_iov != fast_iov)
3078 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3079 sizeof(struct iovec) * iter->nr_segs);
3081 req->flags |= REQ_F_NEED_CLEANUP;
3085 static inline int __io_alloc_async_data(struct io_kiocb *req)
3087 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3088 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3089 return req->async_data == NULL;
3092 static int io_alloc_async_data(struct io_kiocb *req)
3094 if (!io_op_defs[req->opcode].needs_async_data)
3097 return __io_alloc_async_data(req);
3100 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3101 const struct iovec *fast_iov,
3102 struct iov_iter *iter, bool force)
3104 if (!force && !io_op_defs[req->opcode].needs_async_data)
3106 if (!req->async_data) {
3107 if (__io_alloc_async_data(req)) {
3112 io_req_map_rw(req, iovec, fast_iov, iter);
3117 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3119 struct io_async_rw *iorw = req->async_data;
3120 struct iovec *iov = iorw->fast_iov;
3123 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3124 if (unlikely(ret < 0))
3127 iorw->bytes_done = 0;
3128 iorw->free_iovec = iov;
3130 req->flags |= REQ_F_NEED_CLEANUP;
3134 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3136 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3138 return io_prep_rw(req, sqe);
3142 * This is our waitqueue callback handler, registered through lock_page_async()
3143 * when we initially tried to do the IO with the iocb armed our waitqueue.
3144 * This gets called when the page is unlocked, and we generally expect that to
3145 * happen when the page IO is completed and the page is now uptodate. This will
3146 * queue a task_work based retry of the operation, attempting to copy the data
3147 * again. If the latter fails because the page was NOT uptodate, then we will
3148 * do a thread based blocking retry of the operation. That's the unexpected
3151 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3152 int sync, void *arg)
3154 struct wait_page_queue *wpq;
3155 struct io_kiocb *req = wait->private;
3156 struct wait_page_key *key = arg;
3158 wpq = container_of(wait, struct wait_page_queue, wait);
3160 if (!wake_page_match(wpq, key))
3163 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3164 list_del_init(&wait->entry);
3166 /* submit ref gets dropped, acquire a new one */
3167 refcount_inc(&req->refs);
3168 io_req_task_queue(req);
3173 * This controls whether a given IO request should be armed for async page
3174 * based retry. If we return false here, the request is handed to the async
3175 * worker threads for retry. If we're doing buffered reads on a regular file,
3176 * we prepare a private wait_page_queue entry and retry the operation. This
3177 * will either succeed because the page is now uptodate and unlocked, or it
3178 * will register a callback when the page is unlocked at IO completion. Through
3179 * that callback, io_uring uses task_work to setup a retry of the operation.
3180 * That retry will attempt the buffered read again. The retry will generally
3181 * succeed, or in rare cases where it fails, we then fall back to using the
3182 * async worker threads for a blocking retry.
3184 static bool io_rw_should_retry(struct io_kiocb *req)
3186 struct io_async_rw *rw = req->async_data;
3187 struct wait_page_queue *wait = &rw->wpq;
3188 struct kiocb *kiocb = &req->rw.kiocb;
3190 /* never retry for NOWAIT, we just complete with -EAGAIN */
3191 if (req->flags & REQ_F_NOWAIT)
3194 /* Only for buffered IO */
3195 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3199 * just use poll if we can, and don't attempt if the fs doesn't
3200 * support callback based unlocks
3202 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3205 wait->wait.func = io_async_buf_func;
3206 wait->wait.private = req;
3207 wait->wait.flags = 0;
3208 INIT_LIST_HEAD(&wait->wait.entry);
3209 kiocb->ki_flags |= IOCB_WAITQ;
3210 kiocb->ki_flags &= ~IOCB_NOWAIT;
3211 kiocb->ki_waitq = wait;
3215 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3217 if (req->file->f_op->read_iter)
3218 return call_read_iter(req->file, &req->rw.kiocb, iter);
3219 else if (req->file->f_op->read)
3220 return loop_rw_iter(READ, req, iter);
3225 static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3227 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3228 struct kiocb *kiocb = &req->rw.kiocb;
3229 struct iov_iter __iter, *iter = &__iter;
3230 struct io_async_rw *rw = req->async_data;
3231 ssize_t io_size, ret, ret2;
3232 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3238 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3242 io_size = iov_iter_count(iter);
3243 req->result = io_size;
3245 /* Ensure we clear previously set non-block flag */
3246 if (!force_nonblock)
3247 kiocb->ki_flags &= ~IOCB_NOWAIT;
3249 kiocb->ki_flags |= IOCB_NOWAIT;
3251 /* If the file doesn't support async, just async punt */
3252 if (force_nonblock && !io_file_supports_async(req->file, READ)) {
3253 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3254 return ret ?: -EAGAIN;
3257 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size);
3258 if (unlikely(ret)) {
3263 ret = io_iter_do_read(req, iter);
3265 if (ret == -EIOCBQUEUED) {
3266 if (req->async_data)
3267 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3269 } else if (ret == -EAGAIN) {
3270 /* IOPOLL retry should happen for io-wq threads */
3271 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3273 /* no retry on NONBLOCK nor RWF_NOWAIT */
3274 if (req->flags & REQ_F_NOWAIT)
3276 /* some cases will consume bytes even on error returns */
3277 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3279 } else if (ret <= 0 || ret == io_size || !force_nonblock ||
3280 (req->flags & REQ_F_NOWAIT) || !(req->flags & REQ_F_ISREG)) {
3281 /* read all, failed, already did sync or don't want to retry */
3285 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3290 rw = req->async_data;
3291 /* now use our persistent iterator, if we aren't already */
3296 rw->bytes_done += ret;
3297 /* if we can retry, do so with the callbacks armed */
3298 if (!io_rw_should_retry(req)) {
3299 kiocb->ki_flags &= ~IOCB_WAITQ;
3304 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3305 * we get -EIOCBQUEUED, then we'll get a notification when the
3306 * desired page gets unlocked. We can also get a partial read
3307 * here, and if we do, then just retry at the new offset.
3309 ret = io_iter_do_read(req, iter);
3310 if (ret == -EIOCBQUEUED)
3312 /* we got some bytes, but not all. retry. */
3313 kiocb->ki_flags &= ~IOCB_WAITQ;
3314 } while (ret > 0 && ret < io_size);
3316 kiocb_done(kiocb, ret, issue_flags);
3318 /* it's faster to check here then delegate to kfree */
3324 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3326 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3328 return io_prep_rw(req, sqe);
3331 static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3333 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3334 struct kiocb *kiocb = &req->rw.kiocb;
3335 struct iov_iter __iter, *iter = &__iter;
3336 struct io_async_rw *rw = req->async_data;
3337 ssize_t ret, ret2, io_size;
3338 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3344 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3348 io_size = iov_iter_count(iter);
3349 req->result = io_size;
3351 /* Ensure we clear previously set non-block flag */
3352 if (!force_nonblock)
3353 kiocb->ki_flags &= ~IOCB_NOWAIT;
3355 kiocb->ki_flags |= IOCB_NOWAIT;
3357 /* If the file doesn't support async, just async punt */
3358 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
3361 /* file path doesn't support NOWAIT for non-direct_IO */
3362 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3363 (req->flags & REQ_F_ISREG))
3366 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size);
3371 * Open-code file_start_write here to grab freeze protection,
3372 * which will be released by another thread in
3373 * io_complete_rw(). Fool lockdep by telling it the lock got
3374 * released so that it doesn't complain about the held lock when
3375 * we return to userspace.
3377 if (req->flags & REQ_F_ISREG) {
3378 sb_start_write(file_inode(req->file)->i_sb);
3379 __sb_writers_release(file_inode(req->file)->i_sb,
3382 kiocb->ki_flags |= IOCB_WRITE;
3384 if (req->file->f_op->write_iter)
3385 ret2 = call_write_iter(req->file, kiocb, iter);
3386 else if (req->file->f_op->write)
3387 ret2 = loop_rw_iter(WRITE, req, iter);
3392 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3393 * retry them without IOCB_NOWAIT.
3395 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3397 /* no retry on NONBLOCK nor RWF_NOWAIT */
3398 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3400 if (ret2 == -EIOCBQUEUED && req->async_data)
3401 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3402 if (!force_nonblock || ret2 != -EAGAIN) {
3403 /* IOPOLL retry should happen for io-wq threads */
3404 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3407 kiocb_done(kiocb, ret2, issue_flags);
3410 /* some cases will consume bytes even on error returns */
3411 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3412 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3413 return ret ?: -EAGAIN;
3416 /* it's reportedly faster than delegating the null check to kfree() */
3422 static int io_renameat_prep(struct io_kiocb *req,
3423 const struct io_uring_sqe *sqe)
3425 struct io_rename *ren = &req->rename;
3426 const char __user *oldf, *newf;
3428 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3431 ren->old_dfd = READ_ONCE(sqe->fd);
3432 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3433 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3434 ren->new_dfd = READ_ONCE(sqe->len);
3435 ren->flags = READ_ONCE(sqe->rename_flags);
3437 ren->oldpath = getname(oldf);
3438 if (IS_ERR(ren->oldpath))
3439 return PTR_ERR(ren->oldpath);
3441 ren->newpath = getname(newf);
3442 if (IS_ERR(ren->newpath)) {
3443 putname(ren->oldpath);
3444 return PTR_ERR(ren->newpath);
3447 req->flags |= REQ_F_NEED_CLEANUP;
3451 static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3453 struct io_rename *ren = &req->rename;
3456 if (issue_flags & IO_URING_F_NONBLOCK)
3459 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3460 ren->newpath, ren->flags);
3462 req->flags &= ~REQ_F_NEED_CLEANUP;
3464 req_set_fail_links(req);
3465 io_req_complete(req, ret);
3469 static int io_unlinkat_prep(struct io_kiocb *req,
3470 const struct io_uring_sqe *sqe)
3472 struct io_unlink *un = &req->unlink;
3473 const char __user *fname;
3475 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3478 un->dfd = READ_ONCE(sqe->fd);
3480 un->flags = READ_ONCE(sqe->unlink_flags);
3481 if (un->flags & ~AT_REMOVEDIR)
3484 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3485 un->filename = getname(fname);
3486 if (IS_ERR(un->filename))
3487 return PTR_ERR(un->filename);
3489 req->flags |= REQ_F_NEED_CLEANUP;
3493 static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3495 struct io_unlink *un = &req->unlink;
3498 if (issue_flags & IO_URING_F_NONBLOCK)
3501 if (un->flags & AT_REMOVEDIR)
3502 ret = do_rmdir(un->dfd, un->filename);
3504 ret = do_unlinkat(un->dfd, un->filename);
3506 req->flags &= ~REQ_F_NEED_CLEANUP;
3508 req_set_fail_links(req);
3509 io_req_complete(req, ret);
3513 static int io_shutdown_prep(struct io_kiocb *req,
3514 const struct io_uring_sqe *sqe)
3516 #if defined(CONFIG_NET)
3517 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3519 if (sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3523 req->shutdown.how = READ_ONCE(sqe->len);
3530 static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3532 #if defined(CONFIG_NET)
3533 struct socket *sock;
3536 if (issue_flags & IO_URING_F_NONBLOCK)
3539 sock = sock_from_file(req->file);
3540 if (unlikely(!sock))
3543 ret = __sys_shutdown_sock(sock, req->shutdown.how);
3545 req_set_fail_links(req);
3546 io_req_complete(req, ret);
3553 static int __io_splice_prep(struct io_kiocb *req,
3554 const struct io_uring_sqe *sqe)
3556 struct io_splice* sp = &req->splice;
3557 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3559 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3563 sp->len = READ_ONCE(sqe->len);
3564 sp->flags = READ_ONCE(sqe->splice_flags);
3566 if (unlikely(sp->flags & ~valid_flags))
3569 sp->file_in = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in),
3570 (sp->flags & SPLICE_F_FD_IN_FIXED));
3573 req->flags |= REQ_F_NEED_CLEANUP;
3575 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
3577 * Splice operation will be punted aync, and here need to
3578 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3580 req->work.flags |= IO_WQ_WORK_UNBOUND;
3586 static int io_tee_prep(struct io_kiocb *req,
3587 const struct io_uring_sqe *sqe)
3589 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3591 return __io_splice_prep(req, sqe);
3594 static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
3596 struct io_splice *sp = &req->splice;
3597 struct file *in = sp->file_in;
3598 struct file *out = sp->file_out;
3599 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3602 if (issue_flags & IO_URING_F_NONBLOCK)
3605 ret = do_tee(in, out, sp->len, flags);
3607 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3608 req->flags &= ~REQ_F_NEED_CLEANUP;
3611 req_set_fail_links(req);
3612 io_req_complete(req, ret);
3616 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3618 struct io_splice* sp = &req->splice;
3620 sp->off_in = READ_ONCE(sqe->splice_off_in);
3621 sp->off_out = READ_ONCE(sqe->off);
3622 return __io_splice_prep(req, sqe);
3625 static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
3627 struct io_splice *sp = &req->splice;
3628 struct file *in = sp->file_in;
3629 struct file *out = sp->file_out;
3630 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3631 loff_t *poff_in, *poff_out;
3634 if (issue_flags & IO_URING_F_NONBLOCK)
3637 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3638 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3641 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3643 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3644 req->flags &= ~REQ_F_NEED_CLEANUP;
3647 req_set_fail_links(req);
3648 io_req_complete(req, ret);
3653 * IORING_OP_NOP just posts a completion event, nothing else.
3655 static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
3657 struct io_ring_ctx *ctx = req->ctx;
3659 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3662 __io_req_complete(req, issue_flags, 0, 0);
3666 static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3668 struct io_ring_ctx *ctx = req->ctx;
3673 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3675 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3678 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3679 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3682 req->sync.off = READ_ONCE(sqe->off);
3683 req->sync.len = READ_ONCE(sqe->len);
3687 static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
3689 loff_t end = req->sync.off + req->sync.len;
3692 /* fsync always requires a blocking context */
3693 if (issue_flags & IO_URING_F_NONBLOCK)
3696 ret = vfs_fsync_range(req->file, req->sync.off,
3697 end > 0 ? end : LLONG_MAX,
3698 req->sync.flags & IORING_FSYNC_DATASYNC);
3700 req_set_fail_links(req);
3701 io_req_complete(req, ret);
3705 static int io_fallocate_prep(struct io_kiocb *req,
3706 const struct io_uring_sqe *sqe)
3708 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3710 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3713 req->sync.off = READ_ONCE(sqe->off);
3714 req->sync.len = READ_ONCE(sqe->addr);
3715 req->sync.mode = READ_ONCE(sqe->len);
3719 static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
3723 /* fallocate always requiring blocking context */
3724 if (issue_flags & IO_URING_F_NONBLOCK)
3726 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3729 req_set_fail_links(req);
3730 io_req_complete(req, ret);
3734 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3736 const char __user *fname;
3739 if (unlikely(sqe->ioprio || sqe->buf_index))
3741 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3744 /* open.how should be already initialised */
3745 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3746 req->open.how.flags |= O_LARGEFILE;
3748 req->open.dfd = READ_ONCE(sqe->fd);
3749 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3750 req->open.filename = getname(fname);
3751 if (IS_ERR(req->open.filename)) {
3752 ret = PTR_ERR(req->open.filename);
3753 req->open.filename = NULL;
3756 req->open.nofile = rlimit(RLIMIT_NOFILE);
3757 req->flags |= REQ_F_NEED_CLEANUP;
3761 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3765 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3767 mode = READ_ONCE(sqe->len);
3768 flags = READ_ONCE(sqe->open_flags);
3769 req->open.how = build_open_how(flags, mode);
3770 return __io_openat_prep(req, sqe);
3773 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3775 struct open_how __user *how;
3779 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3781 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3782 len = READ_ONCE(sqe->len);
3783 if (len < OPEN_HOW_SIZE_VER0)
3786 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3791 return __io_openat_prep(req, sqe);
3794 static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
3796 struct open_flags op;
3799 bool resolve_nonblock;
3802 ret = build_open_flags(&req->open.how, &op);
3805 nonblock_set = op.open_flag & O_NONBLOCK;
3806 resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
3807 if (issue_flags & IO_URING_F_NONBLOCK) {
3809 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
3810 * it'll always -EAGAIN
3812 if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
3814 op.lookup_flags |= LOOKUP_CACHED;
3815 op.open_flag |= O_NONBLOCK;
3818 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3822 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3823 /* only retry if RESOLVE_CACHED wasn't already set by application */
3824 if ((!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)) &&
3825 file == ERR_PTR(-EAGAIN)) {
3827 * We could hang on to this 'fd', but seems like marginal
3828 * gain for something that is now known to be a slower path.
3829 * So just put it, and we'll get a new one when we retry.
3837 ret = PTR_ERR(file);
3839 if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
3840 file->f_flags &= ~O_NONBLOCK;
3841 fsnotify_open(file);
3842 fd_install(ret, file);
3845 putname(req->open.filename);
3846 req->flags &= ~REQ_F_NEED_CLEANUP;
3848 req_set_fail_links(req);
3849 io_req_complete(req, ret);
3853 static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
3855 return io_openat2(req, issue_flags);
3858 static int io_remove_buffers_prep(struct io_kiocb *req,
3859 const struct io_uring_sqe *sqe)
3861 struct io_provide_buf *p = &req->pbuf;
3864 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3867 tmp = READ_ONCE(sqe->fd);
3868 if (!tmp || tmp > USHRT_MAX)
3871 memset(p, 0, sizeof(*p));
3873 p->bgid = READ_ONCE(sqe->buf_group);
3877 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3878 int bgid, unsigned nbufs)
3882 /* shouldn't happen */
3886 /* the head kbuf is the list itself */
3887 while (!list_empty(&buf->list)) {
3888 struct io_buffer *nxt;
3890 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3891 list_del(&nxt->list);
3898 idr_remove(&ctx->io_buffer_idr, bgid);
3903 static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
3905 struct io_provide_buf *p = &req->pbuf;
3906 struct io_ring_ctx *ctx = req->ctx;
3907 struct io_buffer *head;
3909 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3911 io_ring_submit_lock(ctx, !force_nonblock);
3913 lockdep_assert_held(&ctx->uring_lock);
3916 head = idr_find(&ctx->io_buffer_idr, p->bgid);
3918 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3920 req_set_fail_links(req);
3922 /* need to hold the lock to complete IOPOLL requests */
3923 if (ctx->flags & IORING_SETUP_IOPOLL) {
3924 __io_req_complete(req, issue_flags, ret, 0);
3925 io_ring_submit_unlock(ctx, !force_nonblock);
3927 io_ring_submit_unlock(ctx, !force_nonblock);
3928 __io_req_complete(req, issue_flags, ret, 0);
3933 static int io_provide_buffers_prep(struct io_kiocb *req,
3934 const struct io_uring_sqe *sqe)
3936 struct io_provide_buf *p = &req->pbuf;
3939 if (sqe->ioprio || sqe->rw_flags)
3942 tmp = READ_ONCE(sqe->fd);
3943 if (!tmp || tmp > USHRT_MAX)
3946 p->addr = READ_ONCE(sqe->addr);
3947 p->len = READ_ONCE(sqe->len);
3949 if (!access_ok(u64_to_user_ptr(p->addr), (p->len * p->nbufs)))
3952 p->bgid = READ_ONCE(sqe->buf_group);
3953 tmp = READ_ONCE(sqe->off);
3954 if (tmp > USHRT_MAX)
3960 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3962 struct io_buffer *buf;
3963 u64 addr = pbuf->addr;
3964 int i, bid = pbuf->bid;
3966 for (i = 0; i < pbuf->nbufs; i++) {
3967 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
3972 buf->len = pbuf->len;
3977 INIT_LIST_HEAD(&buf->list);
3980 list_add_tail(&buf->list, &(*head)->list);
3984 return i ? i : -ENOMEM;
3987 static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
3989 struct io_provide_buf *p = &req->pbuf;
3990 struct io_ring_ctx *ctx = req->ctx;
3991 struct io_buffer *head, *list;
3993 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3995 io_ring_submit_lock(ctx, !force_nonblock);
3997 lockdep_assert_held(&ctx->uring_lock);
3999 list = head = idr_find(&ctx->io_buffer_idr, p->bgid);
4001 ret = io_add_buffers(p, &head);
4006 ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1,
4009 __io_remove_buffers(ctx, head, p->bgid, -1U);
4015 req_set_fail_links(req);
4017 /* need to hold the lock to complete IOPOLL requests */
4018 if (ctx->flags & IORING_SETUP_IOPOLL) {
4019 __io_req_complete(req, issue_flags, ret, 0);
4020 io_ring_submit_unlock(ctx, !force_nonblock);
4022 io_ring_submit_unlock(ctx, !force_nonblock);
4023 __io_req_complete(req, issue_flags, ret, 0);
4028 static int io_epoll_ctl_prep(struct io_kiocb *req,
4029 const struct io_uring_sqe *sqe)
4031 #if defined(CONFIG_EPOLL)
4032 if (sqe->ioprio || sqe->buf_index)
4034 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4037 req->epoll.epfd = READ_ONCE(sqe->fd);
4038 req->epoll.op = READ_ONCE(sqe->len);
4039 req->epoll.fd = READ_ONCE(sqe->off);
4041 if (ep_op_has_event(req->epoll.op)) {
4042 struct epoll_event __user *ev;
4044 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4045 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4055 static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4057 #if defined(CONFIG_EPOLL)
4058 struct io_epoll *ie = &req->epoll;
4060 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4062 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4063 if (force_nonblock && ret == -EAGAIN)
4067 req_set_fail_links(req);
4068 __io_req_complete(req, issue_flags, ret, 0);
4075 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4077 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4078 if (sqe->ioprio || sqe->buf_index || sqe->off)
4080 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4083 req->madvise.addr = READ_ONCE(sqe->addr);
4084 req->madvise.len = READ_ONCE(sqe->len);
4085 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4092 static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4094 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4095 struct io_madvise *ma = &req->madvise;
4098 if (issue_flags & IO_URING_F_NONBLOCK)
4101 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4103 req_set_fail_links(req);
4104 io_req_complete(req, ret);
4111 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4113 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4115 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4118 req->fadvise.offset = READ_ONCE(sqe->off);
4119 req->fadvise.len = READ_ONCE(sqe->len);
4120 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4124 static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4126 struct io_fadvise *fa = &req->fadvise;
4129 if (issue_flags & IO_URING_F_NONBLOCK) {
4130 switch (fa->advice) {
4131 case POSIX_FADV_NORMAL:
4132 case POSIX_FADV_RANDOM:
4133 case POSIX_FADV_SEQUENTIAL:
4140 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4142 req_set_fail_links(req);
4143 io_req_complete(req, ret);
4147 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4149 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4151 if (sqe->ioprio || sqe->buf_index)
4153 if (req->flags & REQ_F_FIXED_FILE)
4156 req->statx.dfd = READ_ONCE(sqe->fd);
4157 req->statx.mask = READ_ONCE(sqe->len);
4158 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4159 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4160 req->statx.flags = READ_ONCE(sqe->statx_flags);
4165 static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4167 struct io_statx *ctx = &req->statx;
4170 if (issue_flags & IO_URING_F_NONBLOCK) {
4171 /* only need file table for an actual valid fd */
4172 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
4173 req->flags |= REQ_F_NO_FILE_TABLE;
4177 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4181 req_set_fail_links(req);
4182 io_req_complete(req, ret);
4186 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4188 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4190 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4191 sqe->rw_flags || sqe->buf_index)
4193 if (req->flags & REQ_F_FIXED_FILE)
4196 req->close.fd = READ_ONCE(sqe->fd);
4200 static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4202 struct files_struct *files = current->files;
4203 struct io_close *close = &req->close;
4204 struct fdtable *fdt;
4210 spin_lock(&files->file_lock);
4211 fdt = files_fdtable(files);
4212 if (close->fd >= fdt->max_fds) {
4213 spin_unlock(&files->file_lock);
4216 file = fdt->fd[close->fd];
4218 spin_unlock(&files->file_lock);
4222 if (file->f_op == &io_uring_fops) {
4223 spin_unlock(&files->file_lock);
4228 /* if the file has a flush method, be safe and punt to async */
4229 if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4230 spin_unlock(&files->file_lock);
4234 ret = __close_fd_get_file(close->fd, &file);
4235 spin_unlock(&files->file_lock);
4242 /* No ->flush() or already async, safely close from here */
4243 ret = filp_close(file, current->files);
4246 req_set_fail_links(req);
4249 __io_req_complete(req, issue_flags, ret, 0);
4253 static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4255 struct io_ring_ctx *ctx = req->ctx;
4257 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4259 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4262 req->sync.off = READ_ONCE(sqe->off);
4263 req->sync.len = READ_ONCE(sqe->len);
4264 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4268 static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4272 /* sync_file_range always requires a blocking context */
4273 if (issue_flags & IO_URING_F_NONBLOCK)
4276 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4279 req_set_fail_links(req);
4280 io_req_complete(req, ret);
4284 #if defined(CONFIG_NET)
4285 static int io_setup_async_msg(struct io_kiocb *req,
4286 struct io_async_msghdr *kmsg)
4288 struct io_async_msghdr *async_msg = req->async_data;
4292 if (io_alloc_async_data(req)) {
4293 kfree(kmsg->free_iov);
4296 async_msg = req->async_data;
4297 req->flags |= REQ_F_NEED_CLEANUP;
4298 memcpy(async_msg, kmsg, sizeof(*kmsg));
4299 async_msg->msg.msg_name = &async_msg->addr;
4300 /* if were using fast_iov, set it to the new one */
4301 if (!async_msg->free_iov)
4302 async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4307 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4308 struct io_async_msghdr *iomsg)
4310 iomsg->msg.msg_name = &iomsg->addr;
4311 iomsg->free_iov = iomsg->fast_iov;
4312 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4313 req->sr_msg.msg_flags, &iomsg->free_iov);
4316 static int io_sendmsg_prep_async(struct io_kiocb *req)
4320 if (!io_op_defs[req->opcode].needs_async_data)
4322 ret = io_sendmsg_copy_hdr(req, req->async_data);
4324 req->flags |= REQ_F_NEED_CLEANUP;
4328 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4330 struct io_sr_msg *sr = &req->sr_msg;
4332 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4335 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4336 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4337 sr->len = READ_ONCE(sqe->len);
4339 #ifdef CONFIG_COMPAT
4340 if (req->ctx->compat)
4341 sr->msg_flags |= MSG_CMSG_COMPAT;
4346 static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4348 struct io_async_msghdr iomsg, *kmsg;
4349 struct socket *sock;
4353 sock = sock_from_file(req->file);
4354 if (unlikely(!sock))
4357 kmsg = req->async_data;
4359 ret = io_sendmsg_copy_hdr(req, &iomsg);
4365 flags = req->sr_msg.msg_flags;
4366 if (flags & MSG_DONTWAIT)
4367 req->flags |= REQ_F_NOWAIT;
4368 else if (issue_flags & IO_URING_F_NONBLOCK)
4369 flags |= MSG_DONTWAIT;
4371 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4372 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4373 return io_setup_async_msg(req, kmsg);
4374 if (ret == -ERESTARTSYS)
4377 /* fast path, check for non-NULL to avoid function call */
4379 kfree(kmsg->free_iov);
4380 req->flags &= ~REQ_F_NEED_CLEANUP;
4382 req_set_fail_links(req);
4383 __io_req_complete(req, issue_flags, ret, 0);
4387 static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4389 struct io_sr_msg *sr = &req->sr_msg;
4392 struct socket *sock;
4396 sock = sock_from_file(req->file);
4397 if (unlikely(!sock))
4400 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4404 msg.msg_name = NULL;
4405 msg.msg_control = NULL;
4406 msg.msg_controllen = 0;
4407 msg.msg_namelen = 0;
4409 flags = req->sr_msg.msg_flags;
4410 if (flags & MSG_DONTWAIT)
4411 req->flags |= REQ_F_NOWAIT;
4412 else if (issue_flags & IO_URING_F_NONBLOCK)
4413 flags |= MSG_DONTWAIT;
4415 msg.msg_flags = flags;
4416 ret = sock_sendmsg(sock, &msg);
4417 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4419 if (ret == -ERESTARTSYS)
4423 req_set_fail_links(req);
4424 __io_req_complete(req, issue_flags, ret, 0);
4428 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4429 struct io_async_msghdr *iomsg)
4431 struct io_sr_msg *sr = &req->sr_msg;
4432 struct iovec __user *uiov;
4436 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4437 &iomsg->uaddr, &uiov, &iov_len);
4441 if (req->flags & REQ_F_BUFFER_SELECT) {
4444 if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4446 sr->len = iomsg->fast_iov[0].iov_len;
4447 iomsg->free_iov = NULL;
4449 iomsg->free_iov = iomsg->fast_iov;
4450 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4451 &iomsg->free_iov, &iomsg->msg.msg_iter,
4460 #ifdef CONFIG_COMPAT
4461 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4462 struct io_async_msghdr *iomsg)
4464 struct compat_msghdr __user *msg_compat;
4465 struct io_sr_msg *sr = &req->sr_msg;
4466 struct compat_iovec __user *uiov;
4471 msg_compat = (struct compat_msghdr __user *) sr->umsg;
4472 ret = __get_compat_msghdr(&iomsg->msg, msg_compat, &iomsg->uaddr,
4477 uiov = compat_ptr(ptr);
4478 if (req->flags & REQ_F_BUFFER_SELECT) {
4479 compat_ssize_t clen;
4483 if (!access_ok(uiov, sizeof(*uiov)))
4485 if (__get_user(clen, &uiov->iov_len))
4490 iomsg->free_iov = NULL;
4492 iomsg->free_iov = iomsg->fast_iov;
4493 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4494 UIO_FASTIOV, &iomsg->free_iov,
4495 &iomsg->msg.msg_iter, true);
4504 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4505 struct io_async_msghdr *iomsg)
4507 iomsg->msg.msg_name = &iomsg->addr;
4509 #ifdef CONFIG_COMPAT
4510 if (req->ctx->compat)
4511 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4514 return __io_recvmsg_copy_hdr(req, iomsg);
4517 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4520 struct io_sr_msg *sr = &req->sr_msg;
4521 struct io_buffer *kbuf;
4523 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4528 req->flags |= REQ_F_BUFFER_SELECTED;
4532 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4534 return io_put_kbuf(req, req->sr_msg.kbuf);
4537 static int io_recvmsg_prep_async(struct io_kiocb *req)
4541 if (!io_op_defs[req->opcode].needs_async_data)
4543 ret = io_recvmsg_copy_hdr(req, req->async_data);
4545 req->flags |= REQ_F_NEED_CLEANUP;
4549 static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4551 struct io_sr_msg *sr = &req->sr_msg;
4553 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4556 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4557 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4558 sr->len = READ_ONCE(sqe->len);
4559 sr->bgid = READ_ONCE(sqe->buf_group);
4561 #ifdef CONFIG_COMPAT
4562 if (req->ctx->compat)
4563 sr->msg_flags |= MSG_CMSG_COMPAT;
4568 static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
4570 struct io_async_msghdr iomsg, *kmsg;
4571 struct socket *sock;
4572 struct io_buffer *kbuf;
4574 int ret, cflags = 0;
4575 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4577 sock = sock_from_file(req->file);
4578 if (unlikely(!sock))
4581 kmsg = req->async_data;
4583 ret = io_recvmsg_copy_hdr(req, &iomsg);
4589 if (req->flags & REQ_F_BUFFER_SELECT) {
4590 kbuf = io_recv_buffer_select(req, !force_nonblock);
4592 return PTR_ERR(kbuf);
4593 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4594 kmsg->fast_iov[0].iov_len = req->sr_msg.len;
4595 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
4596 1, req->sr_msg.len);
4599 flags = req->sr_msg.msg_flags;
4600 if (flags & MSG_DONTWAIT)
4601 req->flags |= REQ_F_NOWAIT;
4602 else if (force_nonblock)
4603 flags |= MSG_DONTWAIT;
4605 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4606 kmsg->uaddr, flags);
4607 if (force_nonblock && ret == -EAGAIN)
4608 return io_setup_async_msg(req, kmsg);
4609 if (ret == -ERESTARTSYS)
4612 if (req->flags & REQ_F_BUFFER_SELECTED)
4613 cflags = io_put_recv_kbuf(req);
4614 /* fast path, check for non-NULL to avoid function call */
4616 kfree(kmsg->free_iov);
4617 req->flags &= ~REQ_F_NEED_CLEANUP;
4619 req_set_fail_links(req);
4620 __io_req_complete(req, issue_flags, ret, cflags);
4624 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
4626 struct io_buffer *kbuf;
4627 struct io_sr_msg *sr = &req->sr_msg;
4629 void __user *buf = sr->buf;
4630 struct socket *sock;
4633 int ret, cflags = 0;
4634 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4636 sock = sock_from_file(req->file);
4637 if (unlikely(!sock))
4640 if (req->flags & REQ_F_BUFFER_SELECT) {
4641 kbuf = io_recv_buffer_select(req, !force_nonblock);
4643 return PTR_ERR(kbuf);
4644 buf = u64_to_user_ptr(kbuf->addr);
4647 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4651 msg.msg_name = NULL;
4652 msg.msg_control = NULL;
4653 msg.msg_controllen = 0;
4654 msg.msg_namelen = 0;
4655 msg.msg_iocb = NULL;
4658 flags = req->sr_msg.msg_flags;
4659 if (flags & MSG_DONTWAIT)
4660 req->flags |= REQ_F_NOWAIT;
4661 else if (force_nonblock)
4662 flags |= MSG_DONTWAIT;
4664 ret = sock_recvmsg(sock, &msg, flags);
4665 if (force_nonblock && ret == -EAGAIN)
4667 if (ret == -ERESTARTSYS)
4670 if (req->flags & REQ_F_BUFFER_SELECTED)
4671 cflags = io_put_recv_kbuf(req);
4673 req_set_fail_links(req);
4674 __io_req_complete(req, issue_flags, ret, cflags);
4678 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4680 struct io_accept *accept = &req->accept;
4682 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4684 if (sqe->ioprio || sqe->len || sqe->buf_index)
4687 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4688 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4689 accept->flags = READ_ONCE(sqe->accept_flags);
4690 accept->nofile = rlimit(RLIMIT_NOFILE);
4694 static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
4696 struct io_accept *accept = &req->accept;
4697 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4698 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4701 if (req->file->f_flags & O_NONBLOCK)
4702 req->flags |= REQ_F_NOWAIT;
4704 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4705 accept->addr_len, accept->flags,
4707 if (ret == -EAGAIN && force_nonblock)
4710 if (ret == -ERESTARTSYS)
4712 req_set_fail_links(req);
4714 __io_req_complete(req, issue_flags, ret, 0);
4718 static int io_connect_prep_async(struct io_kiocb *req)
4720 struct io_async_connect *io = req->async_data;
4721 struct io_connect *conn = &req->connect;
4723 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
4726 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4728 struct io_connect *conn = &req->connect;
4730 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4732 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4735 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4736 conn->addr_len = READ_ONCE(sqe->addr2);
4740 static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
4742 struct io_async_connect __io, *io;
4743 unsigned file_flags;
4745 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4747 if (req->async_data) {
4748 io = req->async_data;
4750 ret = move_addr_to_kernel(req->connect.addr,
4751 req->connect.addr_len,
4758 file_flags = force_nonblock ? O_NONBLOCK : 0;
4760 ret = __sys_connect_file(req->file, &io->address,
4761 req->connect.addr_len, file_flags);
4762 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4763 if (req->async_data)
4765 if (io_alloc_async_data(req)) {
4769 io = req->async_data;
4770 memcpy(req->async_data, &__io, sizeof(__io));
4773 if (ret == -ERESTARTSYS)
4777 req_set_fail_links(req);
4778 __io_req_complete(req, issue_flags, ret, 0);
4781 #else /* !CONFIG_NET */
4782 #define IO_NETOP_FN(op) \
4783 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
4785 return -EOPNOTSUPP; \
4788 #define IO_NETOP_PREP(op) \
4790 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
4792 return -EOPNOTSUPP; \
4795 #define IO_NETOP_PREP_ASYNC(op) \
4797 static int io_##op##_prep_async(struct io_kiocb *req) \
4799 return -EOPNOTSUPP; \
4802 IO_NETOP_PREP_ASYNC(sendmsg);
4803 IO_NETOP_PREP_ASYNC(recvmsg);
4804 IO_NETOP_PREP_ASYNC(connect);
4805 IO_NETOP_PREP(accept);
4808 #endif /* CONFIG_NET */
4810 struct io_poll_table {
4811 struct poll_table_struct pt;
4812 struct io_kiocb *req;
4816 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4817 __poll_t mask, task_work_func_t func)
4821 /* for instances that support it check for an event match first: */
4822 if (mask && !(mask & poll->events))
4825 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4827 list_del_init(&poll->wait.entry);
4830 req->task_work.func = func;
4831 percpu_ref_get(&req->ctx->refs);
4834 * If this fails, then the task is exiting. When a task exits, the
4835 * work gets canceled, so just cancel this request as well instead
4836 * of executing it. We can't safely execute it anyway, as we may not
4837 * have the needed state needed for it anyway.
4839 ret = io_req_task_work_add(req);
4840 if (unlikely(ret)) {
4841 WRITE_ONCE(poll->canceled, true);
4842 io_req_task_work_add_fallback(req, func);
4847 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4848 __acquires(&req->ctx->completion_lock)
4850 struct io_ring_ctx *ctx = req->ctx;
4852 if (!req->result && !READ_ONCE(poll->canceled)) {
4853 struct poll_table_struct pt = { ._key = poll->events };
4855 req->result = vfs_poll(req->file, &pt) & poll->events;
4858 spin_lock_irq(&ctx->completion_lock);
4859 if (!req->result && !READ_ONCE(poll->canceled)) {
4860 add_wait_queue(poll->head, &poll->wait);
4867 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4869 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4870 if (req->opcode == IORING_OP_POLL_ADD)
4871 return req->async_data;
4872 return req->apoll->double_poll;
4875 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4877 if (req->opcode == IORING_OP_POLL_ADD)
4879 return &req->apoll->poll;
4882 static void io_poll_remove_double(struct io_kiocb *req)
4884 struct io_poll_iocb *poll = io_poll_get_double(req);
4886 lockdep_assert_held(&req->ctx->completion_lock);
4888 if (poll && poll->head) {
4889 struct wait_queue_head *head = poll->head;
4891 spin_lock(&head->lock);
4892 list_del_init(&poll->wait.entry);
4893 if (poll->wait.private)
4894 refcount_dec(&req->refs);
4896 spin_unlock(&head->lock);
4900 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
4902 struct io_ring_ctx *ctx = req->ctx;
4904 io_poll_remove_double(req);
4905 req->poll.done = true;
4906 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
4907 io_commit_cqring(ctx);
4910 static void io_poll_task_func(struct callback_head *cb)
4912 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4913 struct io_ring_ctx *ctx = req->ctx;
4914 struct io_kiocb *nxt;
4916 if (io_poll_rewait(req, &req->poll)) {
4917 spin_unlock_irq(&ctx->completion_lock);
4919 hash_del(&req->hash_node);
4920 io_poll_complete(req, req->result, 0);
4921 spin_unlock_irq(&ctx->completion_lock);
4923 nxt = io_put_req_find_next(req);
4924 io_cqring_ev_posted(ctx);
4926 __io_req_task_submit(nxt);
4929 percpu_ref_put(&ctx->refs);
4932 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4933 int sync, void *key)
4935 struct io_kiocb *req = wait->private;
4936 struct io_poll_iocb *poll = io_poll_get_single(req);
4937 __poll_t mask = key_to_poll(key);
4939 /* for instances that support it check for an event match first: */
4940 if (mask && !(mask & poll->events))
4943 list_del_init(&wait->entry);
4945 if (poll && poll->head) {
4948 spin_lock(&poll->head->lock);
4949 done = list_empty(&poll->wait.entry);
4951 list_del_init(&poll->wait.entry);
4952 /* make sure double remove sees this as being gone */
4953 wait->private = NULL;
4954 spin_unlock(&poll->head->lock);
4956 /* use wait func handler, so it matches the rq type */
4957 poll->wait.func(&poll->wait, mode, sync, key);
4960 refcount_dec(&req->refs);
4964 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
4965 wait_queue_func_t wake_func)
4969 poll->canceled = false;
4970 poll->events = events;
4971 INIT_LIST_HEAD(&poll->wait.entry);
4972 init_waitqueue_func_entry(&poll->wait, wake_func);
4975 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
4976 struct wait_queue_head *head,
4977 struct io_poll_iocb **poll_ptr)
4979 struct io_kiocb *req = pt->req;
4982 * If poll->head is already set, it's because the file being polled
4983 * uses multiple waitqueues for poll handling (eg one for read, one
4984 * for write). Setup a separate io_poll_iocb if this happens.
4986 if (unlikely(poll->head)) {
4987 struct io_poll_iocb *poll_one = poll;
4989 /* already have a 2nd entry, fail a third attempt */
4991 pt->error = -EINVAL;
4994 /* double add on the same waitqueue head, ignore */
4995 if (poll->head == head)
4997 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
4999 pt->error = -ENOMEM;
5002 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5003 refcount_inc(&req->refs);
5004 poll->wait.private = req;
5011 if (poll->events & EPOLLEXCLUSIVE)
5012 add_wait_queue_exclusive(head, &poll->wait);
5014 add_wait_queue(head, &poll->wait);
5017 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5018 struct poll_table_struct *p)
5020 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5021 struct async_poll *apoll = pt->req->apoll;
5023 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5026 static void io_async_task_func(struct callback_head *cb)
5028 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
5029 struct async_poll *apoll = req->apoll;
5030 struct io_ring_ctx *ctx = req->ctx;
5032 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
5034 if (io_poll_rewait(req, &apoll->poll)) {
5035 spin_unlock_irq(&ctx->completion_lock);
5036 percpu_ref_put(&ctx->refs);
5040 /* If req is still hashed, it cannot have been canceled. Don't check. */
5041 if (hash_hashed(&req->hash_node))
5042 hash_del(&req->hash_node);
5044 io_poll_remove_double(req);
5045 spin_unlock_irq(&ctx->completion_lock);
5047 if (!READ_ONCE(apoll->poll.canceled))
5048 __io_req_task_submit(req);
5050 __io_req_task_cancel(req, -ECANCELED);
5052 percpu_ref_put(&ctx->refs);
5053 kfree(apoll->double_poll);
5057 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5060 struct io_kiocb *req = wait->private;
5061 struct io_poll_iocb *poll = &req->apoll->poll;
5063 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5066 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5069 static void io_poll_req_insert(struct io_kiocb *req)
5071 struct io_ring_ctx *ctx = req->ctx;
5072 struct hlist_head *list;
5074 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5075 hlist_add_head(&req->hash_node, list);
5078 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5079 struct io_poll_iocb *poll,
5080 struct io_poll_table *ipt, __poll_t mask,
5081 wait_queue_func_t wake_func)
5082 __acquires(&ctx->completion_lock)
5084 struct io_ring_ctx *ctx = req->ctx;
5085 bool cancel = false;
5087 INIT_HLIST_NODE(&req->hash_node);
5088 io_init_poll_iocb(poll, mask, wake_func);
5089 poll->file = req->file;
5090 poll->wait.private = req;
5092 ipt->pt._key = mask;
5094 ipt->error = -EINVAL;
5096 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5098 spin_lock_irq(&ctx->completion_lock);
5099 if (likely(poll->head)) {
5100 spin_lock(&poll->head->lock);
5101 if (unlikely(list_empty(&poll->wait.entry))) {
5107 if (mask || ipt->error)
5108 list_del_init(&poll->wait.entry);
5110 WRITE_ONCE(poll->canceled, true);
5111 else if (!poll->done) /* actually waiting for an event */
5112 io_poll_req_insert(req);
5113 spin_unlock(&poll->head->lock);
5119 static bool io_arm_poll_handler(struct io_kiocb *req)
5121 const struct io_op_def *def = &io_op_defs[req->opcode];
5122 struct io_ring_ctx *ctx = req->ctx;
5123 struct async_poll *apoll;
5124 struct io_poll_table ipt;
5128 if (!req->file || !file_can_poll(req->file))
5130 if (req->flags & REQ_F_POLLED)
5134 else if (def->pollout)
5138 /* if we can't nonblock try, then no point in arming a poll handler */
5139 if (!io_file_supports_async(req->file, rw))
5142 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5143 if (unlikely(!apoll))
5145 apoll->double_poll = NULL;
5147 req->flags |= REQ_F_POLLED;
5152 mask |= POLLIN | POLLRDNORM;
5154 mask |= POLLOUT | POLLWRNORM;
5156 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5157 if ((req->opcode == IORING_OP_RECVMSG) &&
5158 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5161 mask |= POLLERR | POLLPRI;
5163 ipt.pt._qproc = io_async_queue_proc;
5165 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5167 if (ret || ipt.error) {
5168 io_poll_remove_double(req);
5169 spin_unlock_irq(&ctx->completion_lock);
5170 kfree(apoll->double_poll);
5174 spin_unlock_irq(&ctx->completion_lock);
5175 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
5176 apoll->poll.events);
5180 static bool __io_poll_remove_one(struct io_kiocb *req,
5181 struct io_poll_iocb *poll)
5183 bool do_complete = false;
5185 spin_lock(&poll->head->lock);
5186 WRITE_ONCE(poll->canceled, true);
5187 if (!list_empty(&poll->wait.entry)) {
5188 list_del_init(&poll->wait.entry);
5191 spin_unlock(&poll->head->lock);
5192 hash_del(&req->hash_node);
5196 static bool io_poll_remove_one(struct io_kiocb *req)
5200 io_poll_remove_double(req);
5202 if (req->opcode == IORING_OP_POLL_ADD) {
5203 do_complete = __io_poll_remove_one(req, &req->poll);
5205 struct async_poll *apoll = req->apoll;
5207 /* non-poll requests have submit ref still */
5208 do_complete = __io_poll_remove_one(req, &apoll->poll);
5211 kfree(apoll->double_poll);
5217 io_cqring_fill_event(req, -ECANCELED);
5218 io_commit_cqring(req->ctx);
5219 req_set_fail_links(req);
5220 io_put_req_deferred(req, 1);
5227 * Returns true if we found and killed one or more poll requests
5229 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5230 struct files_struct *files)
5232 struct hlist_node *tmp;
5233 struct io_kiocb *req;
5236 spin_lock_irq(&ctx->completion_lock);
5237 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5238 struct hlist_head *list;
5240 list = &ctx->cancel_hash[i];
5241 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5242 if (io_match_task(req, tsk, files))
5243 posted += io_poll_remove_one(req);
5246 spin_unlock_irq(&ctx->completion_lock);
5249 io_cqring_ev_posted(ctx);
5254 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
5256 struct hlist_head *list;
5257 struct io_kiocb *req;
5259 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5260 hlist_for_each_entry(req, list, hash_node) {
5261 if (sqe_addr != req->user_data)
5263 if (io_poll_remove_one(req))
5271 static int io_poll_remove_prep(struct io_kiocb *req,
5272 const struct io_uring_sqe *sqe)
5274 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5276 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
5280 req->poll_remove.addr = READ_ONCE(sqe->addr);
5285 * Find a running poll command that matches one specified in sqe->addr,
5286 * and remove it if found.
5288 static int io_poll_remove(struct io_kiocb *req, unsigned int issue_flags)
5290 struct io_ring_ctx *ctx = req->ctx;
5293 spin_lock_irq(&ctx->completion_lock);
5294 ret = io_poll_cancel(ctx, req->poll_remove.addr);
5295 spin_unlock_irq(&ctx->completion_lock);
5298 req_set_fail_links(req);
5299 io_req_complete(req, ret);
5303 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5306 struct io_kiocb *req = wait->private;
5307 struct io_poll_iocb *poll = &req->poll;
5309 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5312 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5313 struct poll_table_struct *p)
5315 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5317 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5320 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5322 struct io_poll_iocb *poll = &req->poll;
5325 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5327 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
5330 events = READ_ONCE(sqe->poll32_events);
5332 events = swahw32(events);
5334 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP |
5335 (events & EPOLLEXCLUSIVE);
5339 static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5341 struct io_poll_iocb *poll = &req->poll;
5342 struct io_ring_ctx *ctx = req->ctx;
5343 struct io_poll_table ipt;
5346 ipt.pt._qproc = io_poll_queue_proc;
5348 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5351 if (mask) { /* no async, we'd stolen it */
5353 io_poll_complete(req, mask, 0);
5355 spin_unlock_irq(&ctx->completion_lock);
5358 io_cqring_ev_posted(ctx);
5364 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5366 struct io_timeout_data *data = container_of(timer,
5367 struct io_timeout_data, timer);
5368 struct io_kiocb *req = data->req;
5369 struct io_ring_ctx *ctx = req->ctx;
5370 unsigned long flags;
5372 spin_lock_irqsave(&ctx->completion_lock, flags);
5373 list_del_init(&req->timeout.list);
5374 atomic_set(&req->ctx->cq_timeouts,
5375 atomic_read(&req->ctx->cq_timeouts) + 1);
5377 io_cqring_fill_event(req, -ETIME);
5378 io_commit_cqring(ctx);
5379 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5381 io_cqring_ev_posted(ctx);
5382 req_set_fail_links(req);
5384 return HRTIMER_NORESTART;
5387 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5390 struct io_timeout_data *io;
5391 struct io_kiocb *req;
5394 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5395 if (user_data == req->user_data) {
5402 return ERR_PTR(ret);
5404 io = req->async_data;
5405 ret = hrtimer_try_to_cancel(&io->timer);
5407 return ERR_PTR(-EALREADY);
5408 list_del_init(&req->timeout.list);
5412 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5414 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5417 return PTR_ERR(req);
5419 req_set_fail_links(req);
5420 io_cqring_fill_event(req, -ECANCELED);
5421 io_put_req_deferred(req, 1);
5425 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5426 struct timespec64 *ts, enum hrtimer_mode mode)
5428 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5429 struct io_timeout_data *data;
5432 return PTR_ERR(req);
5434 req->timeout.off = 0; /* noseq */
5435 data = req->async_data;
5436 list_add_tail(&req->timeout.list, &ctx->timeout_list);
5437 hrtimer_init(&data->timer, CLOCK_MONOTONIC, mode);
5438 data->timer.function = io_timeout_fn;
5439 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
5443 static int io_timeout_remove_prep(struct io_kiocb *req,
5444 const struct io_uring_sqe *sqe)
5446 struct io_timeout_rem *tr = &req->timeout_rem;
5448 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5450 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5452 if (sqe->ioprio || sqe->buf_index || sqe->len)
5455 tr->addr = READ_ONCE(sqe->addr);
5456 tr->flags = READ_ONCE(sqe->timeout_flags);
5457 if (tr->flags & IORING_TIMEOUT_UPDATE) {
5458 if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS))
5460 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
5462 } else if (tr->flags) {
5463 /* timeout removal doesn't support flags */
5470 static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
5472 return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
5477 * Remove or update an existing timeout command
5479 static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
5481 struct io_timeout_rem *tr = &req->timeout_rem;
5482 struct io_ring_ctx *ctx = req->ctx;
5485 spin_lock_irq(&ctx->completion_lock);
5486 if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE))
5487 ret = io_timeout_cancel(ctx, tr->addr);
5489 ret = io_timeout_update(ctx, tr->addr, &tr->ts,
5490 io_translate_timeout_mode(tr->flags));
5492 io_cqring_fill_event(req, ret);
5493 io_commit_cqring(ctx);
5494 spin_unlock_irq(&ctx->completion_lock);
5495 io_cqring_ev_posted(ctx);
5497 req_set_fail_links(req);
5502 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5503 bool is_timeout_link)
5505 struct io_timeout_data *data;
5507 u32 off = READ_ONCE(sqe->off);
5509 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5511 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5513 if (off && is_timeout_link)
5515 flags = READ_ONCE(sqe->timeout_flags);
5516 if (flags & ~IORING_TIMEOUT_ABS)
5519 req->timeout.off = off;
5521 if (!req->async_data && io_alloc_async_data(req))
5524 data = req->async_data;
5527 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5530 data->mode = io_translate_timeout_mode(flags);
5531 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5532 io_req_track_inflight(req);
5536 static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
5538 struct io_ring_ctx *ctx = req->ctx;
5539 struct io_timeout_data *data = req->async_data;
5540 struct list_head *entry;
5541 u32 tail, off = req->timeout.off;
5543 spin_lock_irq(&ctx->completion_lock);
5546 * sqe->off holds how many events that need to occur for this
5547 * timeout event to be satisfied. If it isn't set, then this is
5548 * a pure timeout request, sequence isn't used.
5550 if (io_is_timeout_noseq(req)) {
5551 entry = ctx->timeout_list.prev;
5555 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5556 req->timeout.target_seq = tail + off;
5558 /* Update the last seq here in case io_flush_timeouts() hasn't.
5559 * This is safe because ->completion_lock is held, and submissions
5560 * and completions are never mixed in the same ->completion_lock section.
5562 ctx->cq_last_tm_flush = tail;
5565 * Insertion sort, ensuring the first entry in the list is always
5566 * the one we need first.
5568 list_for_each_prev(entry, &ctx->timeout_list) {
5569 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5572 if (io_is_timeout_noseq(nxt))
5574 /* nxt.seq is behind @tail, otherwise would've been completed */
5575 if (off >= nxt->timeout.target_seq - tail)
5579 list_add(&req->timeout.list, entry);
5580 data->timer.function = io_timeout_fn;
5581 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5582 spin_unlock_irq(&ctx->completion_lock);
5586 struct io_cancel_data {
5587 struct io_ring_ctx *ctx;
5591 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5593 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5594 struct io_cancel_data *cd = data;
5596 return req->ctx == cd->ctx && req->user_data == cd->user_data;
5599 static int io_async_cancel_one(struct io_uring_task *tctx, u64 user_data,
5600 struct io_ring_ctx *ctx)
5602 struct io_cancel_data data = { .ctx = ctx, .user_data = user_data, };
5603 enum io_wq_cancel cancel_ret;
5606 if (!tctx || !tctx->io_wq)
5609 cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, &data, false);
5610 switch (cancel_ret) {
5611 case IO_WQ_CANCEL_OK:
5614 case IO_WQ_CANCEL_RUNNING:
5617 case IO_WQ_CANCEL_NOTFOUND:
5625 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5626 struct io_kiocb *req, __u64 sqe_addr,
5629 unsigned long flags;
5632 ret = io_async_cancel_one(req->task->io_uring, sqe_addr, ctx);
5633 if (ret != -ENOENT) {
5634 spin_lock_irqsave(&ctx->completion_lock, flags);
5638 spin_lock_irqsave(&ctx->completion_lock, flags);
5639 ret = io_timeout_cancel(ctx, sqe_addr);
5642 ret = io_poll_cancel(ctx, sqe_addr);
5646 io_cqring_fill_event(req, ret);
5647 io_commit_cqring(ctx);
5648 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5649 io_cqring_ev_posted(ctx);
5652 req_set_fail_links(req);
5656 static int io_async_cancel_prep(struct io_kiocb *req,
5657 const struct io_uring_sqe *sqe)
5659 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5661 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5663 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5666 req->cancel.addr = READ_ONCE(sqe->addr);
5670 static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
5672 struct io_ring_ctx *ctx = req->ctx;
5674 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
5678 static int io_rsrc_update_prep(struct io_kiocb *req,
5679 const struct io_uring_sqe *sqe)
5681 if (unlikely(req->ctx->flags & IORING_SETUP_SQPOLL))
5683 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5685 if (sqe->ioprio || sqe->rw_flags)
5688 req->rsrc_update.offset = READ_ONCE(sqe->off);
5689 req->rsrc_update.nr_args = READ_ONCE(sqe->len);
5690 if (!req->rsrc_update.nr_args)
5692 req->rsrc_update.arg = READ_ONCE(sqe->addr);
5696 static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
5698 struct io_ring_ctx *ctx = req->ctx;
5699 struct io_uring_rsrc_update up;
5702 if (issue_flags & IO_URING_F_NONBLOCK)
5705 up.offset = req->rsrc_update.offset;
5706 up.data = req->rsrc_update.arg;
5708 mutex_lock(&ctx->uring_lock);
5709 ret = __io_sqe_files_update(ctx, &up, req->rsrc_update.nr_args);
5710 mutex_unlock(&ctx->uring_lock);
5713 req_set_fail_links(req);
5714 __io_req_complete(req, issue_flags, ret, 0);
5718 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5720 switch (req->opcode) {
5723 case IORING_OP_READV:
5724 case IORING_OP_READ_FIXED:
5725 case IORING_OP_READ:
5726 return io_read_prep(req, sqe);
5727 case IORING_OP_WRITEV:
5728 case IORING_OP_WRITE_FIXED:
5729 case IORING_OP_WRITE:
5730 return io_write_prep(req, sqe);
5731 case IORING_OP_POLL_ADD:
5732 return io_poll_add_prep(req, sqe);
5733 case IORING_OP_POLL_REMOVE:
5734 return io_poll_remove_prep(req, sqe);
5735 case IORING_OP_FSYNC:
5736 return io_fsync_prep(req, sqe);
5737 case IORING_OP_SYNC_FILE_RANGE:
5738 return io_sfr_prep(req, sqe);
5739 case IORING_OP_SENDMSG:
5740 case IORING_OP_SEND:
5741 return io_sendmsg_prep(req, sqe);
5742 case IORING_OP_RECVMSG:
5743 case IORING_OP_RECV:
5744 return io_recvmsg_prep(req, sqe);
5745 case IORING_OP_CONNECT:
5746 return io_connect_prep(req, sqe);
5747 case IORING_OP_TIMEOUT:
5748 return io_timeout_prep(req, sqe, false);
5749 case IORING_OP_TIMEOUT_REMOVE:
5750 return io_timeout_remove_prep(req, sqe);
5751 case IORING_OP_ASYNC_CANCEL:
5752 return io_async_cancel_prep(req, sqe);
5753 case IORING_OP_LINK_TIMEOUT:
5754 return io_timeout_prep(req, sqe, true);
5755 case IORING_OP_ACCEPT:
5756 return io_accept_prep(req, sqe);
5757 case IORING_OP_FALLOCATE:
5758 return io_fallocate_prep(req, sqe);
5759 case IORING_OP_OPENAT:
5760 return io_openat_prep(req, sqe);
5761 case IORING_OP_CLOSE:
5762 return io_close_prep(req, sqe);
5763 case IORING_OP_FILES_UPDATE:
5764 return io_rsrc_update_prep(req, sqe);
5765 case IORING_OP_STATX:
5766 return io_statx_prep(req, sqe);
5767 case IORING_OP_FADVISE:
5768 return io_fadvise_prep(req, sqe);
5769 case IORING_OP_MADVISE:
5770 return io_madvise_prep(req, sqe);
5771 case IORING_OP_OPENAT2:
5772 return io_openat2_prep(req, sqe);
5773 case IORING_OP_EPOLL_CTL:
5774 return io_epoll_ctl_prep(req, sqe);
5775 case IORING_OP_SPLICE:
5776 return io_splice_prep(req, sqe);
5777 case IORING_OP_PROVIDE_BUFFERS:
5778 return io_provide_buffers_prep(req, sqe);
5779 case IORING_OP_REMOVE_BUFFERS:
5780 return io_remove_buffers_prep(req, sqe);
5782 return io_tee_prep(req, sqe);
5783 case IORING_OP_SHUTDOWN:
5784 return io_shutdown_prep(req, sqe);
5785 case IORING_OP_RENAMEAT:
5786 return io_renameat_prep(req, sqe);
5787 case IORING_OP_UNLINKAT:
5788 return io_unlinkat_prep(req, sqe);
5791 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5796 static int io_req_prep_async(struct io_kiocb *req)
5798 switch (req->opcode) {
5799 case IORING_OP_READV:
5800 case IORING_OP_READ_FIXED:
5801 case IORING_OP_READ:
5802 return io_rw_prep_async(req, READ);
5803 case IORING_OP_WRITEV:
5804 case IORING_OP_WRITE_FIXED:
5805 case IORING_OP_WRITE:
5806 return io_rw_prep_async(req, WRITE);
5807 case IORING_OP_SENDMSG:
5808 case IORING_OP_SEND:
5809 return io_sendmsg_prep_async(req);
5810 case IORING_OP_RECVMSG:
5811 case IORING_OP_RECV:
5812 return io_recvmsg_prep_async(req);
5813 case IORING_OP_CONNECT:
5814 return io_connect_prep_async(req);
5819 static int io_req_defer_prep(struct io_kiocb *req)
5821 if (!io_op_defs[req->opcode].needs_async_data)
5823 /* some opcodes init it during the inital prep */
5824 if (req->async_data)
5826 if (__io_alloc_async_data(req))
5828 return io_req_prep_async(req);
5831 static u32 io_get_sequence(struct io_kiocb *req)
5833 struct io_kiocb *pos;
5834 struct io_ring_ctx *ctx = req->ctx;
5835 u32 total_submitted, nr_reqs = 0;
5837 io_for_each_link(pos, req)
5840 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
5841 return total_submitted - nr_reqs;
5844 static int io_req_defer(struct io_kiocb *req)
5846 struct io_ring_ctx *ctx = req->ctx;
5847 struct io_defer_entry *de;
5851 /* Still need defer if there is pending req in defer list. */
5852 if (likely(list_empty_careful(&ctx->defer_list) &&
5853 !(req->flags & REQ_F_IO_DRAIN)))
5856 seq = io_get_sequence(req);
5857 /* Still a chance to pass the sequence check */
5858 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
5861 ret = io_req_defer_prep(req);
5864 io_prep_async_link(req);
5865 de = kmalloc(sizeof(*de), GFP_KERNEL);
5869 spin_lock_irq(&ctx->completion_lock);
5870 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
5871 spin_unlock_irq(&ctx->completion_lock);
5873 io_queue_async_work(req);
5874 return -EIOCBQUEUED;
5877 trace_io_uring_defer(ctx, req, req->user_data);
5880 list_add_tail(&de->list, &ctx->defer_list);
5881 spin_unlock_irq(&ctx->completion_lock);
5882 return -EIOCBQUEUED;
5885 static void __io_clean_op(struct io_kiocb *req)
5887 if (req->flags & REQ_F_BUFFER_SELECTED) {
5888 switch (req->opcode) {
5889 case IORING_OP_READV:
5890 case IORING_OP_READ_FIXED:
5891 case IORING_OP_READ:
5892 kfree((void *)(unsigned long)req->rw.addr);
5894 case IORING_OP_RECVMSG:
5895 case IORING_OP_RECV:
5896 kfree(req->sr_msg.kbuf);
5899 req->flags &= ~REQ_F_BUFFER_SELECTED;
5902 if (req->flags & REQ_F_NEED_CLEANUP) {
5903 switch (req->opcode) {
5904 case IORING_OP_READV:
5905 case IORING_OP_READ_FIXED:
5906 case IORING_OP_READ:
5907 case IORING_OP_WRITEV:
5908 case IORING_OP_WRITE_FIXED:
5909 case IORING_OP_WRITE: {
5910 struct io_async_rw *io = req->async_data;
5912 kfree(io->free_iovec);
5915 case IORING_OP_RECVMSG:
5916 case IORING_OP_SENDMSG: {
5917 struct io_async_msghdr *io = req->async_data;
5919 kfree(io->free_iov);
5922 case IORING_OP_SPLICE:
5924 io_put_file(req, req->splice.file_in,
5925 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
5927 case IORING_OP_OPENAT:
5928 case IORING_OP_OPENAT2:
5929 if (req->open.filename)
5930 putname(req->open.filename);
5932 case IORING_OP_RENAMEAT:
5933 putname(req->rename.oldpath);
5934 putname(req->rename.newpath);
5936 case IORING_OP_UNLINKAT:
5937 putname(req->unlink.filename);
5940 req->flags &= ~REQ_F_NEED_CLEANUP;
5944 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
5946 struct io_ring_ctx *ctx = req->ctx;
5947 const struct cred *creds = NULL;
5950 if (req->work.creds && req->work.creds != current_cred())
5951 creds = override_creds(req->work.creds);
5953 switch (req->opcode) {
5955 ret = io_nop(req, issue_flags);
5957 case IORING_OP_READV:
5958 case IORING_OP_READ_FIXED:
5959 case IORING_OP_READ:
5960 ret = io_read(req, issue_flags);
5962 case IORING_OP_WRITEV:
5963 case IORING_OP_WRITE_FIXED:
5964 case IORING_OP_WRITE:
5965 ret = io_write(req, issue_flags);
5967 case IORING_OP_FSYNC:
5968 ret = io_fsync(req, issue_flags);
5970 case IORING_OP_POLL_ADD:
5971 ret = io_poll_add(req, issue_flags);
5973 case IORING_OP_POLL_REMOVE:
5974 ret = io_poll_remove(req, issue_flags);
5976 case IORING_OP_SYNC_FILE_RANGE:
5977 ret = io_sync_file_range(req, issue_flags);
5979 case IORING_OP_SENDMSG:
5980 ret = io_sendmsg(req, issue_flags);
5982 case IORING_OP_SEND:
5983 ret = io_send(req, issue_flags);
5985 case IORING_OP_RECVMSG:
5986 ret = io_recvmsg(req, issue_flags);
5988 case IORING_OP_RECV:
5989 ret = io_recv(req, issue_flags);
5991 case IORING_OP_TIMEOUT:
5992 ret = io_timeout(req, issue_flags);
5994 case IORING_OP_TIMEOUT_REMOVE:
5995 ret = io_timeout_remove(req, issue_flags);
5997 case IORING_OP_ACCEPT:
5998 ret = io_accept(req, issue_flags);
6000 case IORING_OP_CONNECT:
6001 ret = io_connect(req, issue_flags);
6003 case IORING_OP_ASYNC_CANCEL:
6004 ret = io_async_cancel(req, issue_flags);
6006 case IORING_OP_FALLOCATE:
6007 ret = io_fallocate(req, issue_flags);
6009 case IORING_OP_OPENAT:
6010 ret = io_openat(req, issue_flags);
6012 case IORING_OP_CLOSE:
6013 ret = io_close(req, issue_flags);
6015 case IORING_OP_FILES_UPDATE:
6016 ret = io_files_update(req, issue_flags);
6018 case IORING_OP_STATX:
6019 ret = io_statx(req, issue_flags);
6021 case IORING_OP_FADVISE:
6022 ret = io_fadvise(req, issue_flags);
6024 case IORING_OP_MADVISE:
6025 ret = io_madvise(req, issue_flags);
6027 case IORING_OP_OPENAT2:
6028 ret = io_openat2(req, issue_flags);
6030 case IORING_OP_EPOLL_CTL:
6031 ret = io_epoll_ctl(req, issue_flags);
6033 case IORING_OP_SPLICE:
6034 ret = io_splice(req, issue_flags);
6036 case IORING_OP_PROVIDE_BUFFERS:
6037 ret = io_provide_buffers(req, issue_flags);
6039 case IORING_OP_REMOVE_BUFFERS:
6040 ret = io_remove_buffers(req, issue_flags);
6043 ret = io_tee(req, issue_flags);
6045 case IORING_OP_SHUTDOWN:
6046 ret = io_shutdown(req, issue_flags);
6048 case IORING_OP_RENAMEAT:
6049 ret = io_renameat(req, issue_flags);
6051 case IORING_OP_UNLINKAT:
6052 ret = io_unlinkat(req, issue_flags);
6060 revert_creds(creds);
6065 /* If the op doesn't have a file, we're not polling for it */
6066 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
6067 const bool in_async = io_wq_current_is_worker();
6069 /* workqueue context doesn't hold uring_lock, grab it now */
6071 mutex_lock(&ctx->uring_lock);
6073 io_iopoll_req_issued(req, in_async);
6076 mutex_unlock(&ctx->uring_lock);
6082 static void io_wq_submit_work(struct io_wq_work *work)
6084 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6085 struct io_kiocb *timeout;
6088 timeout = io_prep_linked_timeout(req);
6090 io_queue_linked_timeout(timeout);
6092 if (work->flags & IO_WQ_WORK_CANCEL)
6097 ret = io_issue_sqe(req, 0);
6099 * We can get EAGAIN for polled IO even though we're
6100 * forcing a sync submission from here, since we can't
6101 * wait for request slots on the block side.
6109 /* avoid locking problems by failing it from a clean context */
6111 /* io-wq is going to take one down */
6112 refcount_inc(&req->refs);
6113 io_req_task_queue_fail(req, ret);
6117 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6120 struct fixed_rsrc_table *table;
6122 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
6123 return table->files[index & IORING_FILE_TABLE_MASK];
6126 static struct file *io_file_get(struct io_submit_state *state,
6127 struct io_kiocb *req, int fd, bool fixed)
6129 struct io_ring_ctx *ctx = req->ctx;
6133 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6135 fd = array_index_nospec(fd, ctx->nr_user_files);
6136 file = io_file_from_index(ctx, fd);
6137 io_set_resource_node(req);
6139 trace_io_uring_file_get(ctx, fd);
6140 file = __io_file_get(state, fd);
6143 if (file && unlikely(file->f_op == &io_uring_fops))
6144 io_req_track_inflight(req);
6148 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6150 struct io_timeout_data *data = container_of(timer,
6151 struct io_timeout_data, timer);
6152 struct io_kiocb *prev, *req = data->req;
6153 struct io_ring_ctx *ctx = req->ctx;
6154 unsigned long flags;
6156 spin_lock_irqsave(&ctx->completion_lock, flags);
6157 prev = req->timeout.head;
6158 req->timeout.head = NULL;
6161 * We don't expect the list to be empty, that will only happen if we
6162 * race with the completion of the linked work.
6164 if (prev && refcount_inc_not_zero(&prev->refs))
6165 io_remove_next_linked(prev);
6168 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6171 req_set_fail_links(prev);
6172 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6173 io_put_req_deferred(prev, 1);
6175 io_req_complete_post(req, -ETIME, 0);
6176 io_put_req_deferred(req, 1);
6178 return HRTIMER_NORESTART;
6181 static void __io_queue_linked_timeout(struct io_kiocb *req)
6184 * If the back reference is NULL, then our linked request finished
6185 * before we got a chance to setup the timer
6187 if (req->timeout.head) {
6188 struct io_timeout_data *data = req->async_data;
6190 data->timer.function = io_link_timeout_fn;
6191 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6196 static void io_queue_linked_timeout(struct io_kiocb *req)
6198 struct io_ring_ctx *ctx = req->ctx;
6200 spin_lock_irq(&ctx->completion_lock);
6201 __io_queue_linked_timeout(req);
6202 spin_unlock_irq(&ctx->completion_lock);
6204 /* drop submission reference */
6208 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6210 struct io_kiocb *nxt = req->link;
6212 if (!nxt || (req->flags & REQ_F_LINK_TIMEOUT) ||
6213 nxt->opcode != IORING_OP_LINK_TIMEOUT)
6216 nxt->timeout.head = req;
6217 nxt->flags |= REQ_F_LTIMEOUT_ACTIVE;
6218 req->flags |= REQ_F_LINK_TIMEOUT;
6222 static void __io_queue_sqe(struct io_kiocb *req)
6224 struct io_kiocb *linked_timeout = io_prep_linked_timeout(req);
6227 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
6230 * We async punt it if the file wasn't marked NOWAIT, or if the file
6231 * doesn't support non-blocking read/write attempts
6233 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6234 if (!io_arm_poll_handler(req)) {
6236 * Queued up for async execution, worker will release
6237 * submit reference when the iocb is actually submitted.
6239 io_queue_async_work(req);
6241 } else if (likely(!ret)) {
6242 /* drop submission reference */
6243 if (req->flags & REQ_F_COMPLETE_INLINE) {
6244 struct io_ring_ctx *ctx = req->ctx;
6245 struct io_comp_state *cs = &ctx->submit_state.comp;
6247 cs->reqs[cs->nr++] = req;
6248 if (cs->nr == ARRAY_SIZE(cs->reqs))
6249 io_submit_flush_completions(cs, ctx);
6254 req_set_fail_links(req);
6256 io_req_complete(req, ret);
6259 io_queue_linked_timeout(linked_timeout);
6262 static void io_queue_sqe(struct io_kiocb *req)
6266 ret = io_req_defer(req);
6268 if (ret != -EIOCBQUEUED) {
6270 req_set_fail_links(req);
6272 io_req_complete(req, ret);
6274 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6275 ret = io_req_defer_prep(req);
6278 io_queue_async_work(req);
6280 __io_queue_sqe(req);
6285 * Check SQE restrictions (opcode and flags).
6287 * Returns 'true' if SQE is allowed, 'false' otherwise.
6289 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6290 struct io_kiocb *req,
6291 unsigned int sqe_flags)
6293 if (!ctx->restricted)
6296 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6299 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6300 ctx->restrictions.sqe_flags_required)
6303 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6304 ctx->restrictions.sqe_flags_required))
6310 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6311 const struct io_uring_sqe *sqe)
6313 struct io_submit_state *state;
6314 unsigned int sqe_flags;
6315 int personality, ret = 0;
6317 req->opcode = READ_ONCE(sqe->opcode);
6318 /* same numerical values with corresponding REQ_F_*, safe to copy */
6319 req->flags = sqe_flags = READ_ONCE(sqe->flags);
6320 req->user_data = READ_ONCE(sqe->user_data);
6321 req->async_data = NULL;
6325 req->fixed_rsrc_refs = NULL;
6326 /* one is dropped after submission, the other at completion */
6327 refcount_set(&req->refs, 2);
6328 req->task = current;
6330 req->work.list.next = NULL;
6331 req->work.creds = NULL;
6332 req->work.flags = 0;
6334 /* enforce forwards compatibility on users */
6335 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS)) {
6340 if (unlikely(req->opcode >= IORING_OP_LAST))
6343 if (unlikely(!io_check_restriction(ctx, req, sqe_flags)))
6346 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6347 !io_op_defs[req->opcode].buffer_select)
6350 personality = READ_ONCE(sqe->personality);
6352 req->work.creds = xa_load(&ctx->personalities, personality);
6353 if (!req->work.creds)
6355 get_cred(req->work.creds);
6357 state = &ctx->submit_state;
6360 * Plug now if we have more than 1 IO left after this, and the target
6361 * is potentially a read/write to block based storage.
6363 if (!state->plug_started && state->ios_left > 1 &&
6364 io_op_defs[req->opcode].plug) {
6365 blk_start_plug(&state->plug);
6366 state->plug_started = true;
6369 if (io_op_defs[req->opcode].needs_file) {
6370 bool fixed = req->flags & REQ_F_FIXED_FILE;
6372 req->file = io_file_get(state, req, READ_ONCE(sqe->fd), fixed);
6373 if (unlikely(!req->file))
6381 static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
6382 const struct io_uring_sqe *sqe)
6384 struct io_submit_link *link = &ctx->submit_state.link;
6387 ret = io_init_req(ctx, req, sqe);
6388 if (unlikely(ret)) {
6391 io_req_complete(req, ret);
6393 /* fail even hard links since we don't submit */
6394 link->head->flags |= REQ_F_FAIL_LINK;
6395 io_put_req(link->head);
6396 io_req_complete(link->head, -ECANCELED);
6401 ret = io_req_prep(req, sqe);
6405 /* don't need @sqe from now on */
6406 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6407 true, ctx->flags & IORING_SETUP_SQPOLL);
6410 * If we already have a head request, queue this one for async
6411 * submittal once the head completes. If we don't have a head but
6412 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6413 * submitted sync once the chain is complete. If none of those
6414 * conditions are true (normal request), then just queue it.
6417 struct io_kiocb *head = link->head;
6420 * Taking sequential execution of a link, draining both sides
6421 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6422 * requests in the link. So, it drains the head and the
6423 * next after the link request. The last one is done via
6424 * drain_next flag to persist the effect across calls.
6426 if (req->flags & REQ_F_IO_DRAIN) {
6427 head->flags |= REQ_F_IO_DRAIN;
6428 ctx->drain_next = 1;
6430 ret = io_req_defer_prep(req);
6433 trace_io_uring_link(ctx, req, head);
6434 link->last->link = req;
6437 /* last request of a link, enqueue the link */
6438 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6443 if (unlikely(ctx->drain_next)) {
6444 req->flags |= REQ_F_IO_DRAIN;
6445 ctx->drain_next = 0;
6447 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6459 * Batched submission is done, ensure local IO is flushed out.
6461 static void io_submit_state_end(struct io_submit_state *state,
6462 struct io_ring_ctx *ctx)
6464 if (state->link.head)
6465 io_queue_sqe(state->link.head);
6467 io_submit_flush_completions(&state->comp, ctx);
6468 if (state->plug_started)
6469 blk_finish_plug(&state->plug);
6470 io_state_file_put(state);
6474 * Start submission side cache.
6476 static void io_submit_state_start(struct io_submit_state *state,
6477 unsigned int max_ios)
6479 state->plug_started = false;
6480 state->ios_left = max_ios;
6481 /* set only head, no need to init link_last in advance */
6482 state->link.head = NULL;
6485 static void io_commit_sqring(struct io_ring_ctx *ctx)
6487 struct io_rings *rings = ctx->rings;
6490 * Ensure any loads from the SQEs are done at this point,
6491 * since once we write the new head, the application could
6492 * write new data to them.
6494 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6498 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6499 * that is mapped by userspace. This means that care needs to be taken to
6500 * ensure that reads are stable, as we cannot rely on userspace always
6501 * being a good citizen. If members of the sqe are validated and then later
6502 * used, it's important that those reads are done through READ_ONCE() to
6503 * prevent a re-load down the line.
6505 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6507 u32 *sq_array = ctx->sq_array;
6511 * The cached sq head (or cq tail) serves two purposes:
6513 * 1) allows us to batch the cost of updating the user visible
6515 * 2) allows the kernel side to track the head on its own, even
6516 * though the application is the one updating it.
6518 head = READ_ONCE(sq_array[ctx->cached_sq_head++ & ctx->sq_mask]);
6519 if (likely(head < ctx->sq_entries))
6520 return &ctx->sq_sqes[head];
6522 /* drop invalid entries */
6523 ctx->cached_sq_dropped++;
6524 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6528 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6532 /* if we have a backlog and couldn't flush it all, return BUSY */
6533 if (test_bit(0, &ctx->sq_check_overflow)) {
6534 if (!__io_cqring_overflow_flush(ctx, false, NULL, NULL))
6538 /* make sure SQ entry isn't read before tail */
6539 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6541 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6544 percpu_counter_add(¤t->io_uring->inflight, nr);
6545 refcount_add(nr, ¤t->usage);
6546 io_submit_state_start(&ctx->submit_state, nr);
6548 while (submitted < nr) {
6549 const struct io_uring_sqe *sqe;
6550 struct io_kiocb *req;
6552 req = io_alloc_req(ctx);
6553 if (unlikely(!req)) {
6555 submitted = -EAGAIN;
6558 sqe = io_get_sqe(ctx);
6559 if (unlikely(!sqe)) {
6560 kmem_cache_free(req_cachep, req);
6563 /* will complete beyond this point, count as submitted */
6565 if (io_submit_sqe(ctx, req, sqe))
6569 if (unlikely(submitted != nr)) {
6570 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6571 struct io_uring_task *tctx = current->io_uring;
6572 int unused = nr - ref_used;
6574 percpu_ref_put_many(&ctx->refs, unused);
6575 percpu_counter_sub(&tctx->inflight, unused);
6576 put_task_struct_many(current, unused);
6579 io_submit_state_end(&ctx->submit_state, ctx);
6580 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6581 io_commit_sqring(ctx);
6586 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6588 /* Tell userspace we may need a wakeup call */
6589 spin_lock_irq(&ctx->completion_lock);
6590 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6591 spin_unlock_irq(&ctx->completion_lock);
6594 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6596 spin_lock_irq(&ctx->completion_lock);
6597 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6598 spin_unlock_irq(&ctx->completion_lock);
6601 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
6603 unsigned int to_submit;
6606 to_submit = io_sqring_entries(ctx);
6607 /* if we're handling multiple rings, cap submit size for fairness */
6608 if (cap_entries && to_submit > 8)
6611 if (!list_empty(&ctx->iopoll_list) || to_submit) {
6612 unsigned nr_events = 0;
6614 mutex_lock(&ctx->uring_lock);
6615 if (!list_empty(&ctx->iopoll_list))
6616 io_do_iopoll(ctx, &nr_events, 0);
6618 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)) &&
6619 !(ctx->flags & IORING_SETUP_R_DISABLED))
6620 ret = io_submit_sqes(ctx, to_submit);
6621 mutex_unlock(&ctx->uring_lock);
6624 if (!io_sqring_full(ctx) && wq_has_sleeper(&ctx->sqo_sq_wait))
6625 wake_up(&ctx->sqo_sq_wait);
6630 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
6632 struct io_ring_ctx *ctx;
6633 unsigned sq_thread_idle = 0;
6635 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6636 if (sq_thread_idle < ctx->sq_thread_idle)
6637 sq_thread_idle = ctx->sq_thread_idle;
6640 sqd->sq_thread_idle = sq_thread_idle;
6643 static void io_sqd_init_new(struct io_sq_data *sqd)
6645 struct io_ring_ctx *ctx;
6647 while (!list_empty(&sqd->ctx_new_list)) {
6648 ctx = list_first_entry(&sqd->ctx_new_list, struct io_ring_ctx, sqd_list);
6649 list_move_tail(&ctx->sqd_list, &sqd->ctx_list);
6650 complete(&ctx->sq_thread_comp);
6653 io_sqd_update_thread_idle(sqd);
6656 static int io_sq_thread(void *data)
6658 struct io_sq_data *sqd = data;
6659 struct io_ring_ctx *ctx;
6660 unsigned long timeout = 0;
6661 char buf[TASK_COMM_LEN];
6664 sprintf(buf, "iou-sqp-%d", sqd->task_pid);
6665 set_task_comm(current, buf);
6666 current->pf_io_worker = NULL;
6668 if (sqd->sq_cpu != -1)
6669 set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
6671 set_cpus_allowed_ptr(current, cpu_online_mask);
6672 current->flags |= PF_NO_SETAFFINITY;
6674 wait_for_completion(&sqd->startup);
6676 down_read(&sqd->rw_lock);
6678 while (!test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state)) {
6680 bool cap_entries, sqt_spin, needs_sched;
6682 if (test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state)) {
6683 up_read(&sqd->rw_lock);
6685 down_read(&sqd->rw_lock);
6688 if (unlikely(!list_empty(&sqd->ctx_new_list))) {
6689 io_sqd_init_new(sqd);
6690 timeout = jiffies + sqd->sq_thread_idle;
6692 if (fatal_signal_pending(current))
6695 cap_entries = !list_is_singular(&sqd->ctx_list);
6696 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6697 const struct cred *creds = NULL;
6699 if (ctx->sq_creds != current_cred())
6700 creds = override_creds(ctx->sq_creds);
6701 ret = __io_sq_thread(ctx, cap_entries);
6703 revert_creds(creds);
6704 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
6708 if (sqt_spin || !time_after(jiffies, timeout)) {
6712 timeout = jiffies + sqd->sq_thread_idle;
6717 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
6718 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6719 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6720 !list_empty_careful(&ctx->iopoll_list)) {
6721 needs_sched = false;
6724 if (io_sqring_entries(ctx)) {
6725 needs_sched = false;
6730 if (needs_sched && !test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state)) {
6731 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6732 io_ring_set_wakeup_flag(ctx);
6734 up_read(&sqd->rw_lock);
6736 down_read(&sqd->rw_lock);
6737 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6738 io_ring_clear_wakeup_flag(ctx);
6741 finish_wait(&sqd->wait, &wait);
6742 timeout = jiffies + sqd->sq_thread_idle;
6745 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6746 io_uring_cancel_sqpoll(ctx);
6747 up_read(&sqd->rw_lock);
6751 down_write(&sqd->rw_lock);
6753 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6754 io_ring_set_wakeup_flag(ctx);
6755 up_write(&sqd->rw_lock);
6756 complete(&sqd->exited);
6760 struct io_wait_queue {
6761 struct wait_queue_entry wq;
6762 struct io_ring_ctx *ctx;
6764 unsigned nr_timeouts;
6767 static inline bool io_should_wake(struct io_wait_queue *iowq)
6769 struct io_ring_ctx *ctx = iowq->ctx;
6772 * Wake up if we have enough events, or if a timeout occurred since we
6773 * started waiting. For timeouts, we always want to return to userspace,
6774 * regardless of event count.
6776 return io_cqring_events(ctx) >= iowq->to_wait ||
6777 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6780 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6781 int wake_flags, void *key)
6783 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6787 * Cannot safely flush overflowed CQEs from here, ensure we wake up
6788 * the task, and the next invocation will do it.
6790 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->cq_check_overflow))
6791 return autoremove_wake_function(curr, mode, wake_flags, key);
6795 static int io_run_task_work_sig(void)
6797 if (io_run_task_work())
6799 if (!signal_pending(current))
6801 if (test_tsk_thread_flag(current, TIF_NOTIFY_SIGNAL))
6802 return -ERESTARTSYS;
6806 /* when returns >0, the caller should retry */
6807 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
6808 struct io_wait_queue *iowq,
6809 signed long *timeout)
6813 /* make sure we run task_work before checking for signals */
6814 ret = io_run_task_work_sig();
6815 if (ret || io_should_wake(iowq))
6817 /* let the caller flush overflows, retry */
6818 if (test_bit(0, &ctx->cq_check_overflow))
6821 *timeout = schedule_timeout(*timeout);
6822 return !*timeout ? -ETIME : 1;
6826 * Wait until events become available, if we don't already have some. The
6827 * application must reap them itself, as they reside on the shared cq ring.
6829 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6830 const sigset_t __user *sig, size_t sigsz,
6831 struct __kernel_timespec __user *uts)
6833 struct io_wait_queue iowq = {
6836 .func = io_wake_function,
6837 .entry = LIST_HEAD_INIT(iowq.wq.entry),
6840 .to_wait = min_events,
6842 struct io_rings *rings = ctx->rings;
6843 signed long timeout = MAX_SCHEDULE_TIMEOUT;
6847 io_cqring_overflow_flush(ctx, false, NULL, NULL);
6848 if (io_cqring_events(ctx) >= min_events)
6850 if (!io_run_task_work())
6855 #ifdef CONFIG_COMPAT
6856 if (in_compat_syscall())
6857 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
6861 ret = set_user_sigmask(sig, sigsz);
6868 struct timespec64 ts;
6870 if (get_timespec64(&ts, uts))
6872 timeout = timespec64_to_jiffies(&ts);
6875 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
6876 trace_io_uring_cqring_wait(ctx, min_events);
6878 /* if we can't even flush overflow, don't wait for more */
6879 if (!io_cqring_overflow_flush(ctx, false, NULL, NULL)) {
6883 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
6884 TASK_INTERRUPTIBLE);
6885 ret = io_cqring_wait_schedule(ctx, &iowq, &timeout);
6886 finish_wait(&ctx->wait, &iowq.wq);
6890 restore_saved_sigmask_unless(ret == -EINTR);
6892 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
6895 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
6897 #if defined(CONFIG_UNIX)
6898 if (ctx->ring_sock) {
6899 struct sock *sock = ctx->ring_sock->sk;
6900 struct sk_buff *skb;
6902 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
6908 for (i = 0; i < ctx->nr_user_files; i++) {
6911 file = io_file_from_index(ctx, i);
6918 static void io_rsrc_data_ref_zero(struct percpu_ref *ref)
6920 struct fixed_rsrc_data *data;
6922 data = container_of(ref, struct fixed_rsrc_data, refs);
6923 complete(&data->done);
6926 static inline void io_rsrc_ref_lock(struct io_ring_ctx *ctx)
6928 spin_lock_bh(&ctx->rsrc_ref_lock);
6931 static inline void io_rsrc_ref_unlock(struct io_ring_ctx *ctx)
6933 spin_unlock_bh(&ctx->rsrc_ref_lock);
6936 static void io_sqe_rsrc_set_node(struct io_ring_ctx *ctx,
6937 struct fixed_rsrc_data *rsrc_data,
6938 struct fixed_rsrc_ref_node *ref_node)
6940 io_rsrc_ref_lock(ctx);
6941 rsrc_data->node = ref_node;
6942 list_add_tail(&ref_node->node, &ctx->rsrc_ref_list);
6943 io_rsrc_ref_unlock(ctx);
6944 percpu_ref_get(&rsrc_data->refs);
6947 static void io_sqe_rsrc_kill_node(struct io_ring_ctx *ctx, struct fixed_rsrc_data *data)
6949 struct fixed_rsrc_ref_node *ref_node = NULL;
6951 io_rsrc_ref_lock(ctx);
6952 ref_node = data->node;
6954 io_rsrc_ref_unlock(ctx);
6956 percpu_ref_kill(&ref_node->refs);
6959 static int io_rsrc_ref_quiesce(struct fixed_rsrc_data *data,
6960 struct io_ring_ctx *ctx,
6961 void (*rsrc_put)(struct io_ring_ctx *ctx,
6962 struct io_rsrc_put *prsrc))
6964 struct fixed_rsrc_ref_node *backup_node;
6970 data->quiesce = true;
6973 backup_node = alloc_fixed_rsrc_ref_node(ctx);
6976 backup_node->rsrc_data = data;
6977 backup_node->rsrc_put = rsrc_put;
6979 io_sqe_rsrc_kill_node(ctx, data);
6980 percpu_ref_kill(&data->refs);
6981 flush_delayed_work(&ctx->rsrc_put_work);
6983 ret = wait_for_completion_interruptible(&data->done);
6987 percpu_ref_resurrect(&data->refs);
6988 io_sqe_rsrc_set_node(ctx, data, backup_node);
6990 reinit_completion(&data->done);
6991 mutex_unlock(&ctx->uring_lock);
6992 ret = io_run_task_work_sig();
6993 mutex_lock(&ctx->uring_lock);
6995 data->quiesce = false;
6998 destroy_fixed_rsrc_ref_node(backup_node);
7002 static struct fixed_rsrc_data *alloc_fixed_rsrc_data(struct io_ring_ctx *ctx)
7004 struct fixed_rsrc_data *data;
7006 data = kzalloc(sizeof(*data), GFP_KERNEL);
7010 if (percpu_ref_init(&data->refs, io_rsrc_data_ref_zero,
7011 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) {
7016 init_completion(&data->done);
7020 static void free_fixed_rsrc_data(struct fixed_rsrc_data *data)
7022 percpu_ref_exit(&data->refs);
7027 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7029 struct fixed_rsrc_data *data = ctx->file_data;
7030 unsigned nr_tables, i;
7034 * percpu_ref_is_dying() is to stop parallel files unregister
7035 * Since we possibly drop uring lock later in this function to
7038 if (!data || percpu_ref_is_dying(&data->refs))
7040 ret = io_rsrc_ref_quiesce(data, ctx, io_ring_file_put);
7044 __io_sqe_files_unregister(ctx);
7045 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
7046 for (i = 0; i < nr_tables; i++)
7047 kfree(data->table[i].files);
7048 free_fixed_rsrc_data(data);
7049 ctx->file_data = NULL;
7050 ctx->nr_user_files = 0;
7054 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7055 __releases(&sqd->rw_lock)
7057 if (sqd->thread == current)
7059 clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7060 up_write(&sqd->rw_lock);
7063 static void io_sq_thread_park(struct io_sq_data *sqd)
7064 __acquires(&sqd->rw_lock)
7066 if (sqd->thread == current)
7068 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7069 down_write(&sqd->rw_lock);
7070 /* set again for consistency, in case concurrent parks are happening */
7071 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7073 wake_up_process(sqd->thread);
7076 static void io_sq_thread_stop(struct io_sq_data *sqd)
7078 if (test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state))
7080 down_write(&sqd->rw_lock);
7081 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7083 wake_up_process(sqd->thread);
7084 up_write(&sqd->rw_lock);
7085 wait_for_completion(&sqd->exited);
7088 static void io_put_sq_data(struct io_sq_data *sqd)
7090 if (refcount_dec_and_test(&sqd->refs)) {
7091 io_sq_thread_stop(sqd);
7096 static void io_sq_thread_finish(struct io_ring_ctx *ctx)
7098 struct io_sq_data *sqd = ctx->sq_data;
7101 complete(&sqd->startup);
7103 wait_for_completion(&ctx->sq_thread_comp);
7105 io_sq_thread_park(sqd);
7106 list_del(&ctx->sqd_list);
7107 io_sqd_update_thread_idle(sqd);
7108 io_sq_thread_unpark(sqd);
7110 io_put_sq_data(sqd);
7111 ctx->sq_data = NULL;
7113 put_cred(ctx->sq_creds);
7117 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7119 struct io_ring_ctx *ctx_attach;
7120 struct io_sq_data *sqd;
7123 f = fdget(p->wq_fd);
7125 return ERR_PTR(-ENXIO);
7126 if (f.file->f_op != &io_uring_fops) {
7128 return ERR_PTR(-EINVAL);
7131 ctx_attach = f.file->private_data;
7132 sqd = ctx_attach->sq_data;
7135 return ERR_PTR(-EINVAL);
7138 refcount_inc(&sqd->refs);
7143 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p)
7145 struct io_sq_data *sqd;
7147 if (p->flags & IORING_SETUP_ATTACH_WQ)
7148 return io_attach_sq_data(p);
7150 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7152 return ERR_PTR(-ENOMEM);
7154 refcount_set(&sqd->refs, 1);
7155 INIT_LIST_HEAD(&sqd->ctx_list);
7156 INIT_LIST_HEAD(&sqd->ctx_new_list);
7157 init_rwsem(&sqd->rw_lock);
7158 init_waitqueue_head(&sqd->wait);
7159 init_completion(&sqd->startup);
7160 init_completion(&sqd->exited);
7164 #if defined(CONFIG_UNIX)
7166 * Ensure the UNIX gc is aware of our file set, so we are certain that
7167 * the io_uring can be safely unregistered on process exit, even if we have
7168 * loops in the file referencing.
7170 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7172 struct sock *sk = ctx->ring_sock->sk;
7173 struct scm_fp_list *fpl;
7174 struct sk_buff *skb;
7177 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7181 skb = alloc_skb(0, GFP_KERNEL);
7190 fpl->user = get_uid(current_user());
7191 for (i = 0; i < nr; i++) {
7192 struct file *file = io_file_from_index(ctx, i + offset);
7196 fpl->fp[nr_files] = get_file(file);
7197 unix_inflight(fpl->user, fpl->fp[nr_files]);
7202 fpl->max = SCM_MAX_FD;
7203 fpl->count = nr_files;
7204 UNIXCB(skb).fp = fpl;
7205 skb->destructor = unix_destruct_scm;
7206 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7207 skb_queue_head(&sk->sk_receive_queue, skb);
7209 for (i = 0; i < nr_files; i++)
7220 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7221 * causes regular reference counting to break down. We rely on the UNIX
7222 * garbage collection to take care of this problem for us.
7224 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7226 unsigned left, total;
7230 left = ctx->nr_user_files;
7232 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7234 ret = __io_sqe_files_scm(ctx, this_files, total);
7238 total += this_files;
7244 while (total < ctx->nr_user_files) {
7245 struct file *file = io_file_from_index(ctx, total);
7255 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7261 static int io_sqe_alloc_file_tables(struct fixed_rsrc_data *file_data,
7262 unsigned nr_tables, unsigned nr_files)
7266 for (i = 0; i < nr_tables; i++) {
7267 struct fixed_rsrc_table *table = &file_data->table[i];
7268 unsigned this_files;
7270 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
7271 table->files = kcalloc(this_files, sizeof(struct file *),
7275 nr_files -= this_files;
7281 for (i = 0; i < nr_tables; i++) {
7282 struct fixed_rsrc_table *table = &file_data->table[i];
7283 kfree(table->files);
7288 static void io_ring_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
7290 struct file *file = prsrc->file;
7291 #if defined(CONFIG_UNIX)
7292 struct sock *sock = ctx->ring_sock->sk;
7293 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7294 struct sk_buff *skb;
7297 __skb_queue_head_init(&list);
7300 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7301 * remove this entry and rearrange the file array.
7303 skb = skb_dequeue(head);
7305 struct scm_fp_list *fp;
7307 fp = UNIXCB(skb).fp;
7308 for (i = 0; i < fp->count; i++) {
7311 if (fp->fp[i] != file)
7314 unix_notinflight(fp->user, fp->fp[i]);
7315 left = fp->count - 1 - i;
7317 memmove(&fp->fp[i], &fp->fp[i + 1],
7318 left * sizeof(struct file *));
7325 __skb_queue_tail(&list, skb);
7335 __skb_queue_tail(&list, skb);
7337 skb = skb_dequeue(head);
7340 if (skb_peek(&list)) {
7341 spin_lock_irq(&head->lock);
7342 while ((skb = __skb_dequeue(&list)) != NULL)
7343 __skb_queue_tail(head, skb);
7344 spin_unlock_irq(&head->lock);
7351 static void __io_rsrc_put_work(struct fixed_rsrc_ref_node *ref_node)
7353 struct fixed_rsrc_data *rsrc_data = ref_node->rsrc_data;
7354 struct io_ring_ctx *ctx = rsrc_data->ctx;
7355 struct io_rsrc_put *prsrc, *tmp;
7357 list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
7358 list_del(&prsrc->list);
7359 ref_node->rsrc_put(ctx, prsrc);
7363 percpu_ref_exit(&ref_node->refs);
7365 percpu_ref_put(&rsrc_data->refs);
7368 static void io_rsrc_put_work(struct work_struct *work)
7370 struct io_ring_ctx *ctx;
7371 struct llist_node *node;
7373 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
7374 node = llist_del_all(&ctx->rsrc_put_llist);
7377 struct fixed_rsrc_ref_node *ref_node;
7378 struct llist_node *next = node->next;
7380 ref_node = llist_entry(node, struct fixed_rsrc_ref_node, llist);
7381 __io_rsrc_put_work(ref_node);
7386 static struct file **io_fixed_file_slot(struct fixed_rsrc_data *file_data,
7389 struct fixed_rsrc_table *table;
7391 table = &file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7392 return &table->files[i & IORING_FILE_TABLE_MASK];
7395 static void io_rsrc_node_ref_zero(struct percpu_ref *ref)
7397 struct fixed_rsrc_ref_node *ref_node;
7398 struct fixed_rsrc_data *data;
7399 struct io_ring_ctx *ctx;
7400 bool first_add = false;
7403 ref_node = container_of(ref, struct fixed_rsrc_ref_node, refs);
7404 data = ref_node->rsrc_data;
7407 io_rsrc_ref_lock(ctx);
7408 ref_node->done = true;
7410 while (!list_empty(&ctx->rsrc_ref_list)) {
7411 ref_node = list_first_entry(&ctx->rsrc_ref_list,
7412 struct fixed_rsrc_ref_node, node);
7413 /* recycle ref nodes in order */
7414 if (!ref_node->done)
7416 list_del(&ref_node->node);
7417 first_add |= llist_add(&ref_node->llist, &ctx->rsrc_put_llist);
7419 io_rsrc_ref_unlock(ctx);
7421 if (percpu_ref_is_dying(&data->refs))
7425 mod_delayed_work(system_wq, &ctx->rsrc_put_work, 0);
7427 queue_delayed_work(system_wq, &ctx->rsrc_put_work, delay);
7430 static struct fixed_rsrc_ref_node *alloc_fixed_rsrc_ref_node(
7431 struct io_ring_ctx *ctx)
7433 struct fixed_rsrc_ref_node *ref_node;
7435 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7439 if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
7444 INIT_LIST_HEAD(&ref_node->node);
7445 INIT_LIST_HEAD(&ref_node->rsrc_list);
7446 ref_node->done = false;
7450 static void init_fixed_file_ref_node(struct io_ring_ctx *ctx,
7451 struct fixed_rsrc_ref_node *ref_node)
7453 ref_node->rsrc_data = ctx->file_data;
7454 ref_node->rsrc_put = io_ring_file_put;
7457 static void destroy_fixed_rsrc_ref_node(struct fixed_rsrc_ref_node *ref_node)
7459 percpu_ref_exit(&ref_node->refs);
7464 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7467 __s32 __user *fds = (__s32 __user *) arg;
7468 unsigned nr_tables, i;
7470 int fd, ret = -ENOMEM;
7471 struct fixed_rsrc_ref_node *ref_node;
7472 struct fixed_rsrc_data *file_data;
7478 if (nr_args > IORING_MAX_FIXED_FILES)
7481 file_data = alloc_fixed_rsrc_data(ctx);
7484 ctx->file_data = file_data;
7486 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
7487 file_data->table = kcalloc(nr_tables, sizeof(*file_data->table),
7489 if (!file_data->table)
7492 if (io_sqe_alloc_file_tables(file_data, nr_tables, nr_args))
7495 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7496 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7500 /* allow sparse sets */
7510 * Don't allow io_uring instances to be registered. If UNIX
7511 * isn't enabled, then this causes a reference cycle and this
7512 * instance can never get freed. If UNIX is enabled we'll
7513 * handle it just fine, but there's still no point in allowing
7514 * a ring fd as it doesn't support regular read/write anyway.
7516 if (file->f_op == &io_uring_fops) {
7520 *io_fixed_file_slot(file_data, i) = file;
7523 ret = io_sqe_files_scm(ctx);
7525 io_sqe_files_unregister(ctx);
7529 ref_node = alloc_fixed_rsrc_ref_node(ctx);
7531 io_sqe_files_unregister(ctx);
7534 init_fixed_file_ref_node(ctx, ref_node);
7536 io_sqe_rsrc_set_node(ctx, file_data, ref_node);
7539 for (i = 0; i < ctx->nr_user_files; i++) {
7540 file = io_file_from_index(ctx, i);
7544 for (i = 0; i < nr_tables; i++)
7545 kfree(file_data->table[i].files);
7546 ctx->nr_user_files = 0;
7548 free_fixed_rsrc_data(ctx->file_data);
7549 ctx->file_data = NULL;
7553 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7556 #if defined(CONFIG_UNIX)
7557 struct sock *sock = ctx->ring_sock->sk;
7558 struct sk_buff_head *head = &sock->sk_receive_queue;
7559 struct sk_buff *skb;
7562 * See if we can merge this file into an existing skb SCM_RIGHTS
7563 * file set. If there's no room, fall back to allocating a new skb
7564 * and filling it in.
7566 spin_lock_irq(&head->lock);
7567 skb = skb_peek(head);
7569 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7571 if (fpl->count < SCM_MAX_FD) {
7572 __skb_unlink(skb, head);
7573 spin_unlock_irq(&head->lock);
7574 fpl->fp[fpl->count] = get_file(file);
7575 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7577 spin_lock_irq(&head->lock);
7578 __skb_queue_head(head, skb);
7583 spin_unlock_irq(&head->lock);
7590 return __io_sqe_files_scm(ctx, 1, index);
7596 static int io_queue_rsrc_removal(struct fixed_rsrc_data *data, void *rsrc)
7598 struct io_rsrc_put *prsrc;
7599 struct fixed_rsrc_ref_node *ref_node = data->node;
7601 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
7606 list_add(&prsrc->list, &ref_node->rsrc_list);
7611 static inline int io_queue_file_removal(struct fixed_rsrc_data *data,
7614 return io_queue_rsrc_removal(data, (void *)file);
7617 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7618 struct io_uring_rsrc_update *up,
7621 struct fixed_rsrc_data *data = ctx->file_data;
7622 struct fixed_rsrc_ref_node *ref_node;
7623 struct file *file, **file_slot;
7627 bool needs_switch = false;
7629 if (check_add_overflow(up->offset, nr_args, &done))
7631 if (done > ctx->nr_user_files)
7634 ref_node = alloc_fixed_rsrc_ref_node(ctx);
7637 init_fixed_file_ref_node(ctx, ref_node);
7639 fds = u64_to_user_ptr(up->data);
7640 for (done = 0; done < nr_args; done++) {
7642 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
7646 if (fd == IORING_REGISTER_FILES_SKIP)
7649 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
7650 file_slot = io_fixed_file_slot(ctx->file_data, i);
7653 err = io_queue_file_removal(data, *file_slot);
7657 needs_switch = true;
7666 * Don't allow io_uring instances to be registered. If
7667 * UNIX isn't enabled, then this causes a reference
7668 * cycle and this instance can never get freed. If UNIX
7669 * is enabled we'll handle it just fine, but there's
7670 * still no point in allowing a ring fd as it doesn't
7671 * support regular read/write anyway.
7673 if (file->f_op == &io_uring_fops) {
7679 err = io_sqe_file_register(ctx, file, i);
7689 percpu_ref_kill(&data->node->refs);
7690 io_sqe_rsrc_set_node(ctx, data, ref_node);
7692 destroy_fixed_rsrc_ref_node(ref_node);
7694 return done ? done : err;
7697 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
7700 struct io_uring_rsrc_update up;
7702 if (!ctx->file_data)
7706 if (copy_from_user(&up, arg, sizeof(up)))
7711 return __io_sqe_files_update(ctx, &up, nr_args);
7714 static struct io_wq_work *io_free_work(struct io_wq_work *work)
7716 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7718 req = io_put_req_find_next(req);
7719 return req ? &req->work : NULL;
7722 static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx)
7724 struct io_wq_hash *hash;
7725 struct io_wq_data data;
7726 unsigned int concurrency;
7728 hash = ctx->hash_map;
7730 hash = kzalloc(sizeof(*hash), GFP_KERNEL);
7732 return ERR_PTR(-ENOMEM);
7733 refcount_set(&hash->refs, 1);
7734 init_waitqueue_head(&hash->wait);
7735 ctx->hash_map = hash;
7739 data.free_work = io_free_work;
7740 data.do_work = io_wq_submit_work;
7742 /* Do QD, or 4 * CPUS, whatever is smallest */
7743 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7745 return io_wq_create(concurrency, &data);
7748 static int io_uring_alloc_task_context(struct task_struct *task,
7749 struct io_ring_ctx *ctx)
7751 struct io_uring_task *tctx;
7754 tctx = kmalloc(sizeof(*tctx), GFP_KERNEL);
7755 if (unlikely(!tctx))
7758 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
7759 if (unlikely(ret)) {
7764 tctx->io_wq = io_init_wq_offload(ctx);
7765 if (IS_ERR(tctx->io_wq)) {
7766 ret = PTR_ERR(tctx->io_wq);
7767 percpu_counter_destroy(&tctx->inflight);
7773 init_waitqueue_head(&tctx->wait);
7775 atomic_set(&tctx->in_idle, 0);
7776 tctx->sqpoll = false;
7777 task->io_uring = tctx;
7778 spin_lock_init(&tctx->task_lock);
7779 INIT_WQ_LIST(&tctx->task_list);
7780 tctx->task_state = 0;
7781 init_task_work(&tctx->task_work, tctx_task_work);
7785 void __io_uring_free(struct task_struct *tsk)
7787 struct io_uring_task *tctx = tsk->io_uring;
7789 WARN_ON_ONCE(!xa_empty(&tctx->xa));
7790 WARN_ON_ONCE(tctx->io_wq);
7792 percpu_counter_destroy(&tctx->inflight);
7794 tsk->io_uring = NULL;
7797 static int io_sq_offload_create(struct io_ring_ctx *ctx,
7798 struct io_uring_params *p)
7802 /* Retain compatibility with failing for an invalid attach attempt */
7803 if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
7804 IORING_SETUP_ATTACH_WQ) {
7807 f = fdget(p->wq_fd);
7810 if (f.file->f_op != &io_uring_fops) {
7816 if (ctx->flags & IORING_SETUP_SQPOLL) {
7817 struct task_struct *tsk;
7818 struct io_sq_data *sqd;
7821 if (!capable(CAP_SYS_ADMIN) && !capable(CAP_SYS_NICE))
7824 sqd = io_get_sq_data(p);
7830 ctx->sq_creds = get_current_cred();
7832 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
7833 if (!ctx->sq_thread_idle)
7834 ctx->sq_thread_idle = HZ;
7836 io_sq_thread_park(sqd);
7837 list_add(&ctx->sqd_list, &sqd->ctx_new_list);
7838 io_sq_thread_unpark(sqd);
7843 if (p->flags & IORING_SETUP_SQ_AFF) {
7844 int cpu = p->sq_thread_cpu;
7847 if (cpu >= nr_cpu_ids)
7849 if (!cpu_online(cpu))
7857 sqd->task_pid = current->pid;
7858 tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE);
7865 ret = io_uring_alloc_task_context(tsk, ctx);
7866 wake_up_new_task(tsk);
7869 complete(&sqd->startup);
7870 } else if (p->flags & IORING_SETUP_SQ_AFF) {
7871 /* Can't have SQ_AFF without SQPOLL */
7878 io_sq_thread_finish(ctx);
7881 complete(&ctx->sq_data->exited);
7885 static inline void __io_unaccount_mem(struct user_struct *user,
7886 unsigned long nr_pages)
7888 atomic_long_sub(nr_pages, &user->locked_vm);
7891 static inline int __io_account_mem(struct user_struct *user,
7892 unsigned long nr_pages)
7894 unsigned long page_limit, cur_pages, new_pages;
7896 /* Don't allow more pages than we can safely lock */
7897 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
7900 cur_pages = atomic_long_read(&user->locked_vm);
7901 new_pages = cur_pages + nr_pages;
7902 if (new_pages > page_limit)
7904 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
7905 new_pages) != cur_pages);
7910 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
7913 __io_unaccount_mem(ctx->user, nr_pages);
7915 if (ctx->mm_account)
7916 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
7919 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
7924 ret = __io_account_mem(ctx->user, nr_pages);
7929 if (ctx->mm_account)
7930 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
7935 static void io_mem_free(void *ptr)
7942 page = virt_to_head_page(ptr);
7943 if (put_page_testzero(page))
7944 free_compound_page(page);
7947 static void *io_mem_alloc(size_t size)
7949 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
7950 __GFP_NORETRY | __GFP_ACCOUNT;
7952 return (void *) __get_free_pages(gfp_flags, get_order(size));
7955 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
7958 struct io_rings *rings;
7959 size_t off, sq_array_size;
7961 off = struct_size(rings, cqes, cq_entries);
7962 if (off == SIZE_MAX)
7966 off = ALIGN(off, SMP_CACHE_BYTES);
7974 sq_array_size = array_size(sizeof(u32), sq_entries);
7975 if (sq_array_size == SIZE_MAX)
7978 if (check_add_overflow(off, sq_array_size, &off))
7984 static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
7988 if (!ctx->user_bufs)
7991 for (i = 0; i < ctx->nr_user_bufs; i++) {
7992 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7994 for (j = 0; j < imu->nr_bvecs; j++)
7995 unpin_user_page(imu->bvec[j].bv_page);
7997 if (imu->acct_pages)
7998 io_unaccount_mem(ctx, imu->acct_pages);
8003 kfree(ctx->user_bufs);
8004 ctx->user_bufs = NULL;
8005 ctx->nr_user_bufs = 0;
8009 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8010 void __user *arg, unsigned index)
8012 struct iovec __user *src;
8014 #ifdef CONFIG_COMPAT
8016 struct compat_iovec __user *ciovs;
8017 struct compat_iovec ciov;
8019 ciovs = (struct compat_iovec __user *) arg;
8020 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8023 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8024 dst->iov_len = ciov.iov_len;
8028 src = (struct iovec __user *) arg;
8029 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8035 * Not super efficient, but this is just a registration time. And we do cache
8036 * the last compound head, so generally we'll only do a full search if we don't
8039 * We check if the given compound head page has already been accounted, to
8040 * avoid double accounting it. This allows us to account the full size of the
8041 * page, not just the constituent pages of a huge page.
8043 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8044 int nr_pages, struct page *hpage)
8048 /* check current page array */
8049 for (i = 0; i < nr_pages; i++) {
8050 if (!PageCompound(pages[i]))
8052 if (compound_head(pages[i]) == hpage)
8056 /* check previously registered pages */
8057 for (i = 0; i < ctx->nr_user_bufs; i++) {
8058 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8060 for (j = 0; j < imu->nr_bvecs; j++) {
8061 if (!PageCompound(imu->bvec[j].bv_page))
8063 if (compound_head(imu->bvec[j].bv_page) == hpage)
8071 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8072 int nr_pages, struct io_mapped_ubuf *imu,
8073 struct page **last_hpage)
8077 for (i = 0; i < nr_pages; i++) {
8078 if (!PageCompound(pages[i])) {
8083 hpage = compound_head(pages[i]);
8084 if (hpage == *last_hpage)
8086 *last_hpage = hpage;
8087 if (headpage_already_acct(ctx, pages, i, hpage))
8089 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8093 if (!imu->acct_pages)
8096 ret = io_account_mem(ctx, imu->acct_pages);
8098 imu->acct_pages = 0;
8102 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
8103 struct io_mapped_ubuf *imu,
8104 struct page **last_hpage)
8106 struct vm_area_struct **vmas = NULL;
8107 struct page **pages = NULL;
8108 unsigned long off, start, end, ubuf;
8110 int ret, pret, nr_pages, i;
8112 ubuf = (unsigned long) iov->iov_base;
8113 end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8114 start = ubuf >> PAGE_SHIFT;
8115 nr_pages = end - start;
8119 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
8123 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
8128 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
8134 mmap_read_lock(current->mm);
8135 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
8137 if (pret == nr_pages) {
8138 /* don't support file backed memory */
8139 for (i = 0; i < nr_pages; i++) {
8140 struct vm_area_struct *vma = vmas[i];
8143 !is_file_hugepages(vma->vm_file)) {
8149 ret = pret < 0 ? pret : -EFAULT;
8151 mmap_read_unlock(current->mm);
8154 * if we did partial map, or found file backed vmas,
8155 * release any pages we did get
8158 unpin_user_pages(pages, pret);
8163 ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
8165 unpin_user_pages(pages, pret);
8170 off = ubuf & ~PAGE_MASK;
8171 size = iov->iov_len;
8172 for (i = 0; i < nr_pages; i++) {
8175 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8176 imu->bvec[i].bv_page = pages[i];
8177 imu->bvec[i].bv_len = vec_len;
8178 imu->bvec[i].bv_offset = off;
8182 /* store original address for later verification */
8184 imu->len = iov->iov_len;
8185 imu->nr_bvecs = nr_pages;
8193 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
8197 if (!nr_args || nr_args > UIO_MAXIOV)
8200 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
8202 if (!ctx->user_bufs)
8208 static int io_buffer_validate(struct iovec *iov)
8211 * Don't impose further limits on the size and buffer
8212 * constraints here, we'll -EINVAL later when IO is
8213 * submitted if they are wrong.
8215 if (!iov->iov_base || !iov->iov_len)
8218 /* arbitrary limit, but we need something */
8219 if (iov->iov_len > SZ_1G)
8225 static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
8226 unsigned int nr_args)
8230 struct page *last_hpage = NULL;
8232 ret = io_buffers_map_alloc(ctx, nr_args);
8236 for (i = 0; i < nr_args; i++) {
8237 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8239 ret = io_copy_iov(ctx, &iov, arg, i);
8243 ret = io_buffer_validate(&iov);
8247 ret = io_sqe_buffer_register(ctx, &iov, imu, &last_hpage);
8251 ctx->nr_user_bufs++;
8255 io_sqe_buffers_unregister(ctx);
8260 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8262 __s32 __user *fds = arg;
8268 if (copy_from_user(&fd, fds, sizeof(*fds)))
8271 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8272 if (IS_ERR(ctx->cq_ev_fd)) {
8273 int ret = PTR_ERR(ctx->cq_ev_fd);
8274 ctx->cq_ev_fd = NULL;
8281 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8283 if (ctx->cq_ev_fd) {
8284 eventfd_ctx_put(ctx->cq_ev_fd);
8285 ctx->cq_ev_fd = NULL;
8292 static int __io_destroy_buffers(int id, void *p, void *data)
8294 struct io_ring_ctx *ctx = data;
8295 struct io_buffer *buf = p;
8297 __io_remove_buffers(ctx, buf, id, -1U);
8301 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8303 idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx);
8304 idr_destroy(&ctx->io_buffer_idr);
8307 static void io_req_cache_free(struct list_head *list, struct task_struct *tsk)
8309 struct io_kiocb *req, *nxt;
8311 list_for_each_entry_safe(req, nxt, list, compl.list) {
8312 if (tsk && req->task != tsk)
8314 list_del(&req->compl.list);
8315 kmem_cache_free(req_cachep, req);
8319 static void io_req_caches_free(struct io_ring_ctx *ctx)
8321 struct io_submit_state *submit_state = &ctx->submit_state;
8322 struct io_comp_state *cs = &ctx->submit_state.comp;
8324 mutex_lock(&ctx->uring_lock);
8326 if (submit_state->free_reqs) {
8327 kmem_cache_free_bulk(req_cachep, submit_state->free_reqs,
8328 submit_state->reqs);
8329 submit_state->free_reqs = 0;
8332 spin_lock_irq(&ctx->completion_lock);
8333 list_splice_init(&cs->locked_free_list, &cs->free_list);
8334 cs->locked_free_nr = 0;
8335 spin_unlock_irq(&ctx->completion_lock);
8337 io_req_cache_free(&cs->free_list, NULL);
8339 mutex_unlock(&ctx->uring_lock);
8342 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8345 * Some may use context even when all refs and requests have been put,
8346 * and they are free to do so while still holding uring_lock, see
8347 * __io_req_task_submit(). Wait for them to finish.
8349 mutex_lock(&ctx->uring_lock);
8350 mutex_unlock(&ctx->uring_lock);
8352 io_sq_thread_finish(ctx);
8353 io_sqe_buffers_unregister(ctx);
8355 if (ctx->mm_account) {
8356 mmdrop(ctx->mm_account);
8357 ctx->mm_account = NULL;
8360 mutex_lock(&ctx->uring_lock);
8361 io_sqe_files_unregister(ctx);
8362 mutex_unlock(&ctx->uring_lock);
8363 io_eventfd_unregister(ctx);
8364 io_destroy_buffers(ctx);
8366 #if defined(CONFIG_UNIX)
8367 if (ctx->ring_sock) {
8368 ctx->ring_sock->file = NULL; /* so that iput() is called */
8369 sock_release(ctx->ring_sock);
8373 io_mem_free(ctx->rings);
8374 io_mem_free(ctx->sq_sqes);
8376 percpu_ref_exit(&ctx->refs);
8377 free_uid(ctx->user);
8378 io_req_caches_free(ctx);
8380 io_wq_put_hash(ctx->hash_map);
8381 kfree(ctx->cancel_hash);
8385 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8387 struct io_ring_ctx *ctx = file->private_data;
8390 poll_wait(file, &ctx->cq_wait, wait);
8392 * synchronizes with barrier from wq_has_sleeper call in
8396 if (!io_sqring_full(ctx))
8397 mask |= EPOLLOUT | EPOLLWRNORM;
8400 * Don't flush cqring overflow list here, just do a simple check.
8401 * Otherwise there could possible be ABBA deadlock:
8404 * lock(&ctx->uring_lock);
8406 * lock(&ctx->uring_lock);
8409 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
8410 * pushs them to do the flush.
8412 if (io_cqring_events(ctx) || test_bit(0, &ctx->cq_check_overflow))
8413 mask |= EPOLLIN | EPOLLRDNORM;
8418 static int io_uring_fasync(int fd, struct file *file, int on)
8420 struct io_ring_ctx *ctx = file->private_data;
8422 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8425 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8427 const struct cred *creds;
8429 creds = xa_erase(&ctx->personalities, id);
8438 static bool io_run_ctx_fallback(struct io_ring_ctx *ctx)
8440 struct callback_head *work, *next;
8441 bool executed = false;
8444 work = xchg(&ctx->exit_task_work, NULL);
8460 struct io_tctx_exit {
8461 struct callback_head task_work;
8462 struct completion completion;
8463 struct io_ring_ctx *ctx;
8466 static void io_tctx_exit_cb(struct callback_head *cb)
8468 struct io_uring_task *tctx = current->io_uring;
8469 struct io_tctx_exit *work;
8471 work = container_of(cb, struct io_tctx_exit, task_work);
8473 * When @in_idle, we're in cancellation and it's racy to remove the
8474 * node. It'll be removed by the end of cancellation, just ignore it.
8476 if (!atomic_read(&tctx->in_idle))
8477 io_uring_del_task_file((unsigned long)work->ctx);
8478 complete(&work->completion);
8481 static void io_ring_exit_work(struct work_struct *work)
8483 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
8484 unsigned long timeout = jiffies + HZ * 60 * 5;
8485 struct io_tctx_exit exit;
8486 struct io_tctx_node *node;
8490 * If we're doing polled IO and end up having requests being
8491 * submitted async (out-of-line), then completions can come in while
8492 * we're waiting for refs to drop. We need to reap these manually,
8493 * as nobody else will be looking for them.
8496 io_uring_try_cancel_requests(ctx, NULL, NULL);
8498 WARN_ON_ONCE(time_after(jiffies, timeout));
8499 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
8501 mutex_lock(&ctx->uring_lock);
8502 while (!list_empty(&ctx->tctx_list)) {
8503 WARN_ON_ONCE(time_after(jiffies, timeout));
8505 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
8508 init_completion(&exit.completion);
8509 init_task_work(&exit.task_work, io_tctx_exit_cb);
8510 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
8511 if (WARN_ON_ONCE(ret))
8513 wake_up_process(node->task);
8515 mutex_unlock(&ctx->uring_lock);
8516 wait_for_completion(&exit.completion);
8518 mutex_lock(&ctx->uring_lock);
8520 mutex_unlock(&ctx->uring_lock);
8522 io_ring_ctx_free(ctx);
8525 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8527 unsigned long index;
8528 struct creds *creds;
8530 mutex_lock(&ctx->uring_lock);
8531 percpu_ref_kill(&ctx->refs);
8532 /* if force is set, the ring is going away. always drop after that */
8533 ctx->cq_overflow_flushed = 1;
8535 __io_cqring_overflow_flush(ctx, true, NULL, NULL);
8536 xa_for_each(&ctx->personalities, index, creds)
8537 io_unregister_personality(ctx, index);
8538 mutex_unlock(&ctx->uring_lock);
8540 io_kill_timeouts(ctx, NULL, NULL);
8541 io_poll_remove_all(ctx, NULL, NULL);
8543 /* if we failed setting up the ctx, we might not have any rings */
8544 io_iopoll_try_reap_events(ctx);
8546 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8548 * Use system_unbound_wq to avoid spawning tons of event kworkers
8549 * if we're exiting a ton of rings at the same time. It just adds
8550 * noise and overhead, there's no discernable change in runtime
8551 * over using system_wq.
8553 queue_work(system_unbound_wq, &ctx->exit_work);
8556 static int io_uring_release(struct inode *inode, struct file *file)
8558 struct io_ring_ctx *ctx = file->private_data;
8560 file->private_data = NULL;
8561 io_ring_ctx_wait_and_kill(ctx);
8565 struct io_task_cancel {
8566 struct task_struct *task;
8567 struct files_struct *files;
8570 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8572 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8573 struct io_task_cancel *cancel = data;
8576 if (cancel->files && (req->flags & REQ_F_LINK_TIMEOUT)) {
8577 unsigned long flags;
8578 struct io_ring_ctx *ctx = req->ctx;
8580 /* protect against races with linked timeouts */
8581 spin_lock_irqsave(&ctx->completion_lock, flags);
8582 ret = io_match_task(req, cancel->task, cancel->files);
8583 spin_unlock_irqrestore(&ctx->completion_lock, flags);
8585 ret = io_match_task(req, cancel->task, cancel->files);
8590 static void io_cancel_defer_files(struct io_ring_ctx *ctx,
8591 struct task_struct *task,
8592 struct files_struct *files)
8594 struct io_defer_entry *de = NULL;
8597 spin_lock_irq(&ctx->completion_lock);
8598 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8599 if (io_match_task(de->req, task, files)) {
8600 list_cut_position(&list, &ctx->defer_list, &de->list);
8604 spin_unlock_irq(&ctx->completion_lock);
8606 while (!list_empty(&list)) {
8607 de = list_first_entry(&list, struct io_defer_entry, list);
8608 list_del_init(&de->list);
8609 req_set_fail_links(de->req);
8610 io_put_req(de->req);
8611 io_req_complete(de->req, -ECANCELED);
8616 static bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
8618 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8620 return req->ctx == data;
8623 static bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
8625 struct io_tctx_node *node;
8626 enum io_wq_cancel cret;
8629 mutex_lock(&ctx->uring_lock);
8630 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
8631 struct io_uring_task *tctx = node->task->io_uring;
8634 * io_wq will stay alive while we hold uring_lock, because it's
8635 * killed after ctx nodes, which requires to take the lock.
8637 if (!tctx || !tctx->io_wq)
8639 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
8640 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8642 mutex_unlock(&ctx->uring_lock);
8647 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
8648 struct task_struct *task,
8649 struct files_struct *files)
8651 struct io_task_cancel cancel = { .task = task, .files = files, };
8652 struct io_uring_task *tctx = task ? task->io_uring : NULL;
8655 enum io_wq_cancel cret;
8659 ret |= io_uring_try_cancel_iowq(ctx);
8660 } else if (tctx && tctx->io_wq) {
8662 * Cancels requests of all rings, not only @ctx, but
8663 * it's fine as the task is in exit/exec.
8665 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
8667 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8670 /* SQPOLL thread does its own polling */
8671 if (!(ctx->flags & IORING_SETUP_SQPOLL) && !files) {
8672 while (!list_empty_careful(&ctx->iopoll_list)) {
8673 io_iopoll_try_reap_events(ctx);
8678 ret |= io_poll_remove_all(ctx, task, files);
8679 ret |= io_kill_timeouts(ctx, task, files);
8680 ret |= io_run_task_work();
8681 ret |= io_run_ctx_fallback(ctx);
8682 io_cqring_overflow_flush(ctx, true, task, files);
8689 static int io_uring_count_inflight(struct io_ring_ctx *ctx,
8690 struct task_struct *task,
8691 struct files_struct *files)
8693 struct io_kiocb *req;
8696 spin_lock_irq(&ctx->inflight_lock);
8697 list_for_each_entry(req, &ctx->inflight_list, inflight_entry)
8698 cnt += io_match_task(req, task, files);
8699 spin_unlock_irq(&ctx->inflight_lock);
8703 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
8704 struct task_struct *task,
8705 struct files_struct *files)
8707 while (!list_empty_careful(&ctx->inflight_list)) {
8711 inflight = io_uring_count_inflight(ctx, task, files);
8715 io_uring_try_cancel_requests(ctx, task, files);
8718 io_sq_thread_unpark(ctx->sq_data);
8719 prepare_to_wait(&task->io_uring->wait, &wait,
8720 TASK_UNINTERRUPTIBLE);
8721 if (inflight == io_uring_count_inflight(ctx, task, files))
8723 finish_wait(&task->io_uring->wait, &wait);
8725 io_sq_thread_park(ctx->sq_data);
8730 * We need to iteratively cancel requests, in case a request has dependent
8731 * hard links. These persist even for failure of cancelations, hence keep
8732 * looping until none are found.
8734 static void io_uring_cancel_task_requests(struct io_ring_ctx *ctx,
8735 struct files_struct *files)
8737 struct task_struct *task = current;
8739 if ((ctx->flags & IORING_SETUP_SQPOLL) && ctx->sq_data) {
8740 io_sq_thread_park(ctx->sq_data);
8741 task = ctx->sq_data->thread;
8743 atomic_inc(&task->io_uring->in_idle);
8746 io_cancel_defer_files(ctx, task, files);
8748 io_uring_cancel_files(ctx, task, files);
8750 io_uring_try_cancel_requests(ctx, task, NULL);
8753 atomic_dec(&task->io_uring->in_idle);
8755 io_sq_thread_unpark(ctx->sq_data);
8759 * Note that this task has used io_uring. We use it for cancelation purposes.
8761 static int io_uring_add_task_file(struct io_ring_ctx *ctx)
8763 struct io_uring_task *tctx = current->io_uring;
8764 struct io_tctx_node *node;
8767 if (unlikely(!tctx)) {
8768 ret = io_uring_alloc_task_context(current, ctx);
8771 tctx = current->io_uring;
8773 if (tctx->last != ctx) {
8774 void *old = xa_load(&tctx->xa, (unsigned long)ctx);
8777 node = kmalloc(sizeof(*node), GFP_KERNEL);
8781 node->task = current;
8783 ret = xa_err(xa_store(&tctx->xa, (unsigned long)ctx,
8790 mutex_lock(&ctx->uring_lock);
8791 list_add(&node->ctx_node, &ctx->tctx_list);
8792 mutex_unlock(&ctx->uring_lock);
8798 * This is race safe in that the task itself is doing this, hence it
8799 * cannot be going through the exit/cancel paths at the same time.
8800 * This cannot be modified while exit/cancel is running.
8802 if (!tctx->sqpoll && (ctx->flags & IORING_SETUP_SQPOLL))
8803 tctx->sqpoll = true;
8809 * Remove this io_uring_file -> task mapping.
8811 static void io_uring_del_task_file(unsigned long index)
8813 struct io_uring_task *tctx = current->io_uring;
8814 struct io_tctx_node *node;
8818 node = xa_erase(&tctx->xa, index);
8822 WARN_ON_ONCE(current != node->task);
8823 WARN_ON_ONCE(list_empty(&node->ctx_node));
8825 mutex_lock(&node->ctx->uring_lock);
8826 list_del(&node->ctx_node);
8827 mutex_unlock(&node->ctx->uring_lock);
8829 if (tctx->last == node->ctx)
8834 static void io_uring_clean_tctx(struct io_uring_task *tctx)
8836 struct io_tctx_node *node;
8837 unsigned long index;
8839 xa_for_each(&tctx->xa, index, node)
8840 io_uring_del_task_file(index);
8842 io_wq_put_and_exit(tctx->io_wq);
8847 void __io_uring_files_cancel(struct files_struct *files)
8849 struct io_uring_task *tctx = current->io_uring;
8850 struct io_tctx_node *node;
8851 unsigned long index;
8853 /* make sure overflow events are dropped */
8854 atomic_inc(&tctx->in_idle);
8855 xa_for_each(&tctx->xa, index, node)
8856 io_uring_cancel_task_requests(node->ctx, files);
8857 atomic_dec(&tctx->in_idle);
8860 io_uring_clean_tctx(tctx);
8863 static s64 tctx_inflight(struct io_uring_task *tctx)
8865 return percpu_counter_sum(&tctx->inflight);
8868 static void io_uring_cancel_sqpoll(struct io_ring_ctx *ctx)
8870 struct io_sq_data *sqd = ctx->sq_data;
8871 struct io_uring_task *tctx;
8877 io_sq_thread_park(sqd);
8878 if (!sqd->thread || !sqd->thread->io_uring) {
8879 io_sq_thread_unpark(sqd);
8882 tctx = ctx->sq_data->thread->io_uring;
8883 atomic_inc(&tctx->in_idle);
8885 /* read completions before cancelations */
8886 inflight = tctx_inflight(tctx);
8889 io_uring_cancel_task_requests(ctx, NULL);
8891 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
8893 * If we've seen completions, retry without waiting. This
8894 * avoids a race where a completion comes in before we did
8895 * prepare_to_wait().
8897 if (inflight == tctx_inflight(tctx))
8899 finish_wait(&tctx->wait, &wait);
8901 atomic_dec(&tctx->in_idle);
8902 io_sq_thread_unpark(sqd);
8906 * Find any io_uring fd that this task has registered or done IO on, and cancel
8909 void __io_uring_task_cancel(void)
8911 struct io_uring_task *tctx = current->io_uring;
8915 /* make sure overflow events are dropped */
8916 atomic_inc(&tctx->in_idle);
8919 struct io_tctx_node *node;
8920 unsigned long index;
8922 xa_for_each(&tctx->xa, index, node)
8923 io_uring_cancel_sqpoll(node->ctx);
8927 /* read completions before cancelations */
8928 inflight = tctx_inflight(tctx);
8931 __io_uring_files_cancel(NULL);
8933 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
8936 * If we've seen completions, retry without waiting. This
8937 * avoids a race where a completion comes in before we did
8938 * prepare_to_wait().
8940 if (inflight == tctx_inflight(tctx))
8942 finish_wait(&tctx->wait, &wait);
8945 atomic_dec(&tctx->in_idle);
8947 io_uring_clean_tctx(tctx);
8948 /* all current's requests should be gone, we can kill tctx */
8949 __io_uring_free(current);
8952 static void *io_uring_validate_mmap_request(struct file *file,
8953 loff_t pgoff, size_t sz)
8955 struct io_ring_ctx *ctx = file->private_data;
8956 loff_t offset = pgoff << PAGE_SHIFT;
8961 case IORING_OFF_SQ_RING:
8962 case IORING_OFF_CQ_RING:
8965 case IORING_OFF_SQES:
8969 return ERR_PTR(-EINVAL);
8972 page = virt_to_head_page(ptr);
8973 if (sz > page_size(page))
8974 return ERR_PTR(-EINVAL);
8981 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
8983 size_t sz = vma->vm_end - vma->vm_start;
8987 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
8989 return PTR_ERR(ptr);
8991 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
8992 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
8995 #else /* !CONFIG_MMU */
8997 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
8999 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
9002 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
9004 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
9007 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
9008 unsigned long addr, unsigned long len,
9009 unsigned long pgoff, unsigned long flags)
9013 ptr = io_uring_validate_mmap_request(file, pgoff, len);
9015 return PTR_ERR(ptr);
9017 return (unsigned long) ptr;
9020 #endif /* !CONFIG_MMU */
9022 static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
9027 if (!io_sqring_full(ctx))
9029 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9031 if (!io_sqring_full(ctx))
9034 } while (!signal_pending(current));
9036 finish_wait(&ctx->sqo_sq_wait, &wait);
9040 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9041 struct __kernel_timespec __user **ts,
9042 const sigset_t __user **sig)
9044 struct io_uring_getevents_arg arg;
9047 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9048 * is just a pointer to the sigset_t.
9050 if (!(flags & IORING_ENTER_EXT_ARG)) {
9051 *sig = (const sigset_t __user *) argp;
9057 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9058 * timespec and sigset_t pointers if good.
9060 if (*argsz != sizeof(arg))
9062 if (copy_from_user(&arg, argp, sizeof(arg)))
9064 *sig = u64_to_user_ptr(arg.sigmask);
9065 *argsz = arg.sigmask_sz;
9066 *ts = u64_to_user_ptr(arg.ts);
9070 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9071 u32, min_complete, u32, flags, const void __user *, argp,
9074 struct io_ring_ctx *ctx;
9081 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9082 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG))
9090 if (f.file->f_op != &io_uring_fops)
9094 ctx = f.file->private_data;
9095 if (!percpu_ref_tryget(&ctx->refs))
9099 if (ctx->flags & IORING_SETUP_R_DISABLED)
9103 * For SQ polling, the thread will do all submissions and completions.
9104 * Just return the requested submit count, and wake the thread if
9108 if (ctx->flags & IORING_SETUP_SQPOLL) {
9109 io_cqring_overflow_flush(ctx, false, NULL, NULL);
9112 if (unlikely(ctx->sq_data->thread == NULL)) {
9115 if (flags & IORING_ENTER_SQ_WAKEUP)
9116 wake_up(&ctx->sq_data->wait);
9117 if (flags & IORING_ENTER_SQ_WAIT) {
9118 ret = io_sqpoll_wait_sq(ctx);
9122 submitted = to_submit;
9123 } else if (to_submit) {
9124 ret = io_uring_add_task_file(ctx);
9127 mutex_lock(&ctx->uring_lock);
9128 submitted = io_submit_sqes(ctx, to_submit);
9129 mutex_unlock(&ctx->uring_lock);
9131 if (submitted != to_submit)
9134 if (flags & IORING_ENTER_GETEVENTS) {
9135 const sigset_t __user *sig;
9136 struct __kernel_timespec __user *ts;
9138 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
9142 min_complete = min(min_complete, ctx->cq_entries);
9145 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9146 * space applications don't need to do io completion events
9147 * polling again, they can rely on io_sq_thread to do polling
9148 * work, which can reduce cpu usage and uring_lock contention.
9150 if (ctx->flags & IORING_SETUP_IOPOLL &&
9151 !(ctx->flags & IORING_SETUP_SQPOLL)) {
9152 ret = io_iopoll_check(ctx, min_complete);
9154 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
9159 percpu_ref_put(&ctx->refs);
9162 return submitted ? submitted : ret;
9165 #ifdef CONFIG_PROC_FS
9166 static int io_uring_show_cred(struct seq_file *m, unsigned int id,
9167 const struct cred *cred)
9169 struct user_namespace *uns = seq_user_ns(m);
9170 struct group_info *gi;
9175 seq_printf(m, "%5d\n", id);
9176 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
9177 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
9178 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
9179 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
9180 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
9181 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
9182 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
9183 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
9184 seq_puts(m, "\n\tGroups:\t");
9185 gi = cred->group_info;
9186 for (g = 0; g < gi->ngroups; g++) {
9187 seq_put_decimal_ull(m, g ? " " : "",
9188 from_kgid_munged(uns, gi->gid[g]));
9190 seq_puts(m, "\n\tCapEff:\t");
9191 cap = cred->cap_effective;
9192 CAP_FOR_EACH_U32(__capi)
9193 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
9198 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
9200 struct io_sq_data *sq = NULL;
9205 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9206 * since fdinfo case grabs it in the opposite direction of normal use
9207 * cases. If we fail to get the lock, we just don't iterate any
9208 * structures that could be going away outside the io_uring mutex.
9210 has_lock = mutex_trylock(&ctx->uring_lock);
9212 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) {
9218 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9219 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9220 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9221 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9222 struct file *f = *io_fixed_file_slot(ctx->file_data, i);
9225 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9227 seq_printf(m, "%5u: <none>\n", i);
9229 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9230 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9231 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
9233 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
9234 (unsigned int) buf->len);
9236 if (has_lock && !xa_empty(&ctx->personalities)) {
9237 unsigned long index;
9238 const struct cred *cred;
9240 seq_printf(m, "Personalities:\n");
9241 xa_for_each(&ctx->personalities, index, cred)
9242 io_uring_show_cred(m, index, cred);
9244 seq_printf(m, "PollList:\n");
9245 spin_lock_irq(&ctx->completion_lock);
9246 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9247 struct hlist_head *list = &ctx->cancel_hash[i];
9248 struct io_kiocb *req;
9250 hlist_for_each_entry(req, list, hash_node)
9251 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9252 req->task->task_works != NULL);
9254 spin_unlock_irq(&ctx->completion_lock);
9256 mutex_unlock(&ctx->uring_lock);
9259 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9261 struct io_ring_ctx *ctx = f->private_data;
9263 if (percpu_ref_tryget(&ctx->refs)) {
9264 __io_uring_show_fdinfo(ctx, m);
9265 percpu_ref_put(&ctx->refs);
9270 static const struct file_operations io_uring_fops = {
9271 .release = io_uring_release,
9272 .mmap = io_uring_mmap,
9274 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9275 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9277 .poll = io_uring_poll,
9278 .fasync = io_uring_fasync,
9279 #ifdef CONFIG_PROC_FS
9280 .show_fdinfo = io_uring_show_fdinfo,
9284 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9285 struct io_uring_params *p)
9287 struct io_rings *rings;
9288 size_t size, sq_array_offset;
9290 /* make sure these are sane, as we already accounted them */
9291 ctx->sq_entries = p->sq_entries;
9292 ctx->cq_entries = p->cq_entries;
9294 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9295 if (size == SIZE_MAX)
9298 rings = io_mem_alloc(size);
9303 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9304 rings->sq_ring_mask = p->sq_entries - 1;
9305 rings->cq_ring_mask = p->cq_entries - 1;
9306 rings->sq_ring_entries = p->sq_entries;
9307 rings->cq_ring_entries = p->cq_entries;
9308 ctx->sq_mask = rings->sq_ring_mask;
9309 ctx->cq_mask = rings->cq_ring_mask;
9311 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9312 if (size == SIZE_MAX) {
9313 io_mem_free(ctx->rings);
9318 ctx->sq_sqes = io_mem_alloc(size);
9319 if (!ctx->sq_sqes) {
9320 io_mem_free(ctx->rings);
9328 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
9332 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9336 ret = io_uring_add_task_file(ctx);
9341 fd_install(fd, file);
9346 * Allocate an anonymous fd, this is what constitutes the application
9347 * visible backing of an io_uring instance. The application mmaps this
9348 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9349 * we have to tie this fd to a socket for file garbage collection purposes.
9351 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
9354 #if defined(CONFIG_UNIX)
9357 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9360 return ERR_PTR(ret);
9363 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9364 O_RDWR | O_CLOEXEC);
9365 #if defined(CONFIG_UNIX)
9367 sock_release(ctx->ring_sock);
9368 ctx->ring_sock = NULL;
9370 ctx->ring_sock->file = file;
9376 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9377 struct io_uring_params __user *params)
9379 struct io_ring_ctx *ctx;
9385 if (entries > IORING_MAX_ENTRIES) {
9386 if (!(p->flags & IORING_SETUP_CLAMP))
9388 entries = IORING_MAX_ENTRIES;
9392 * Use twice as many entries for the CQ ring. It's possible for the
9393 * application to drive a higher depth than the size of the SQ ring,
9394 * since the sqes are only used at submission time. This allows for
9395 * some flexibility in overcommitting a bit. If the application has
9396 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9397 * of CQ ring entries manually.
9399 p->sq_entries = roundup_pow_of_two(entries);
9400 if (p->flags & IORING_SETUP_CQSIZE) {
9402 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9403 * to a power-of-two, if it isn't already. We do NOT impose
9404 * any cq vs sq ring sizing.
9408 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9409 if (!(p->flags & IORING_SETUP_CLAMP))
9411 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9413 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9414 if (p->cq_entries < p->sq_entries)
9417 p->cq_entries = 2 * p->sq_entries;
9420 ctx = io_ring_ctx_alloc(p);
9423 ctx->compat = in_compat_syscall();
9424 if (!capable(CAP_IPC_LOCK))
9425 ctx->user = get_uid(current_user());
9428 * This is just grabbed for accounting purposes. When a process exits,
9429 * the mm is exited and dropped before the files, hence we need to hang
9430 * on to this mm purely for the purposes of being able to unaccount
9431 * memory (locked/pinned vm). It's not used for anything else.
9433 mmgrab(current->mm);
9434 ctx->mm_account = current->mm;
9436 ret = io_allocate_scq_urings(ctx, p);
9440 ret = io_sq_offload_create(ctx, p);
9444 memset(&p->sq_off, 0, sizeof(p->sq_off));
9445 p->sq_off.head = offsetof(struct io_rings, sq.head);
9446 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9447 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9448 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9449 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9450 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9451 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9453 memset(&p->cq_off, 0, sizeof(p->cq_off));
9454 p->cq_off.head = offsetof(struct io_rings, cq.head);
9455 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9456 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9457 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9458 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9459 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9460 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9462 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9463 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9464 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9465 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
9466 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS;
9468 if (copy_to_user(params, p, sizeof(*p))) {
9473 file = io_uring_get_file(ctx);
9475 ret = PTR_ERR(file);
9480 * Install ring fd as the very last thing, so we don't risk someone
9481 * having closed it before we finish setup
9483 ret = io_uring_install_fd(ctx, file);
9485 /* fput will clean it up */
9490 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9493 io_ring_ctx_wait_and_kill(ctx);
9498 * Sets up an aio uring context, and returns the fd. Applications asks for a
9499 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9500 * params structure passed in.
9502 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9504 struct io_uring_params p;
9507 if (copy_from_user(&p, params, sizeof(p)))
9509 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9514 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9515 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9516 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9517 IORING_SETUP_R_DISABLED))
9520 return io_uring_create(entries, &p, params);
9523 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9524 struct io_uring_params __user *, params)
9526 return io_uring_setup(entries, params);
9529 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9531 struct io_uring_probe *p;
9535 size = struct_size(p, ops, nr_args);
9536 if (size == SIZE_MAX)
9538 p = kzalloc(size, GFP_KERNEL);
9543 if (copy_from_user(p, arg, size))
9546 if (memchr_inv(p, 0, size))
9549 p->last_op = IORING_OP_LAST - 1;
9550 if (nr_args > IORING_OP_LAST)
9551 nr_args = IORING_OP_LAST;
9553 for (i = 0; i < nr_args; i++) {
9555 if (!io_op_defs[i].not_supported)
9556 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9561 if (copy_to_user(arg, p, size))
9568 static int io_register_personality(struct io_ring_ctx *ctx)
9570 const struct cred *creds;
9574 creds = get_current_cred();
9576 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
9577 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
9584 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9585 unsigned int nr_args)
9587 struct io_uring_restriction *res;
9591 /* Restrictions allowed only if rings started disabled */
9592 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9595 /* We allow only a single restrictions registration */
9596 if (ctx->restrictions.registered)
9599 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9602 size = array_size(nr_args, sizeof(*res));
9603 if (size == SIZE_MAX)
9606 res = memdup_user(arg, size);
9608 return PTR_ERR(res);
9612 for (i = 0; i < nr_args; i++) {
9613 switch (res[i].opcode) {
9614 case IORING_RESTRICTION_REGISTER_OP:
9615 if (res[i].register_op >= IORING_REGISTER_LAST) {
9620 __set_bit(res[i].register_op,
9621 ctx->restrictions.register_op);
9623 case IORING_RESTRICTION_SQE_OP:
9624 if (res[i].sqe_op >= IORING_OP_LAST) {
9629 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9631 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9632 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9634 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9635 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9644 /* Reset all restrictions if an error happened */
9646 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9648 ctx->restrictions.registered = true;
9654 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9656 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9659 if (ctx->restrictions.registered)
9660 ctx->restricted = 1;
9662 ctx->flags &= ~IORING_SETUP_R_DISABLED;
9663 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
9664 wake_up(&ctx->sq_data->wait);
9668 static bool io_register_op_must_quiesce(int op)
9671 case IORING_UNREGISTER_FILES:
9672 case IORING_REGISTER_FILES_UPDATE:
9673 case IORING_REGISTER_PROBE:
9674 case IORING_REGISTER_PERSONALITY:
9675 case IORING_UNREGISTER_PERSONALITY:
9682 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
9683 void __user *arg, unsigned nr_args)
9684 __releases(ctx->uring_lock)
9685 __acquires(ctx->uring_lock)
9690 * We're inside the ring mutex, if the ref is already dying, then
9691 * someone else killed the ctx or is already going through
9692 * io_uring_register().
9694 if (percpu_ref_is_dying(&ctx->refs))
9697 if (io_register_op_must_quiesce(opcode)) {
9698 percpu_ref_kill(&ctx->refs);
9701 * Drop uring mutex before waiting for references to exit. If
9702 * another thread is currently inside io_uring_enter() it might
9703 * need to grab the uring_lock to make progress. If we hold it
9704 * here across the drain wait, then we can deadlock. It's safe
9705 * to drop the mutex here, since no new references will come in
9706 * after we've killed the percpu ref.
9708 mutex_unlock(&ctx->uring_lock);
9710 ret = wait_for_completion_interruptible(&ctx->ref_comp);
9713 ret = io_run_task_work_sig();
9718 mutex_lock(&ctx->uring_lock);
9721 percpu_ref_resurrect(&ctx->refs);
9726 if (ctx->restricted) {
9727 if (opcode >= IORING_REGISTER_LAST) {
9732 if (!test_bit(opcode, ctx->restrictions.register_op)) {
9739 case IORING_REGISTER_BUFFERS:
9740 ret = io_sqe_buffers_register(ctx, arg, nr_args);
9742 case IORING_UNREGISTER_BUFFERS:
9746 ret = io_sqe_buffers_unregister(ctx);
9748 case IORING_REGISTER_FILES:
9749 ret = io_sqe_files_register(ctx, arg, nr_args);
9751 case IORING_UNREGISTER_FILES:
9755 ret = io_sqe_files_unregister(ctx);
9757 case IORING_REGISTER_FILES_UPDATE:
9758 ret = io_sqe_files_update(ctx, arg, nr_args);
9760 case IORING_REGISTER_EVENTFD:
9761 case IORING_REGISTER_EVENTFD_ASYNC:
9765 ret = io_eventfd_register(ctx, arg);
9768 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
9769 ctx->eventfd_async = 1;
9771 ctx->eventfd_async = 0;
9773 case IORING_UNREGISTER_EVENTFD:
9777 ret = io_eventfd_unregister(ctx);
9779 case IORING_REGISTER_PROBE:
9781 if (!arg || nr_args > 256)
9783 ret = io_probe(ctx, arg, nr_args);
9785 case IORING_REGISTER_PERSONALITY:
9789 ret = io_register_personality(ctx);
9791 case IORING_UNREGISTER_PERSONALITY:
9795 ret = io_unregister_personality(ctx, nr_args);
9797 case IORING_REGISTER_ENABLE_RINGS:
9801 ret = io_register_enable_rings(ctx);
9803 case IORING_REGISTER_RESTRICTIONS:
9804 ret = io_register_restrictions(ctx, arg, nr_args);
9812 if (io_register_op_must_quiesce(opcode)) {
9813 /* bring the ctx back to life */
9814 percpu_ref_reinit(&ctx->refs);
9816 reinit_completion(&ctx->ref_comp);
9821 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
9822 void __user *, arg, unsigned int, nr_args)
9824 struct io_ring_ctx *ctx;
9833 if (f.file->f_op != &io_uring_fops)
9836 ctx = f.file->private_data;
9840 mutex_lock(&ctx->uring_lock);
9841 ret = __io_uring_register(ctx, opcode, arg, nr_args);
9842 mutex_unlock(&ctx->uring_lock);
9843 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
9844 ctx->cq_ev_fd != NULL, ret);
9850 static int __init io_uring_init(void)
9852 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
9853 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
9854 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
9857 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
9858 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
9859 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
9860 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
9861 BUILD_BUG_SQE_ELEM(1, __u8, flags);
9862 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
9863 BUILD_BUG_SQE_ELEM(4, __s32, fd);
9864 BUILD_BUG_SQE_ELEM(8, __u64, off);
9865 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
9866 BUILD_BUG_SQE_ELEM(16, __u64, addr);
9867 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
9868 BUILD_BUG_SQE_ELEM(24, __u32, len);
9869 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
9870 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
9871 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
9872 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
9873 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
9874 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
9875 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
9876 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
9877 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
9878 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
9879 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
9880 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
9881 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
9882 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
9883 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
9884 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
9885 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
9886 BUILD_BUG_SQE_ELEM(42, __u16, personality);
9887 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
9889 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
9890 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
9891 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
9895 __initcall(io_uring_init);