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/fs_struct.h>
78 #include <linux/splice.h>
79 #include <linux/task_work.h>
80 #include <linux/pagemap.h>
81 #include <linux/io_uring.h>
82 #include <linux/blk-cgroup.h>
83 #include <linux/audit.h>
85 #define CREATE_TRACE_POINTS
86 #include <trace/events/io_uring.h>
88 #include <uapi/linux/io_uring.h>
93 #define IORING_MAX_ENTRIES 32768
94 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
97 * Shift of 9 is 512 entries, or exactly one page on 64-bit archs
99 #define IORING_FILE_TABLE_SHIFT 9
100 #define IORING_MAX_FILES_TABLE (1U << IORING_FILE_TABLE_SHIFT)
101 #define IORING_FILE_TABLE_MASK (IORING_MAX_FILES_TABLE - 1)
102 #define IORING_MAX_FIXED_FILES (64 * IORING_MAX_FILES_TABLE)
103 #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
104 IORING_REGISTER_LAST + IORING_OP_LAST)
106 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
107 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
111 u32 head ____cacheline_aligned_in_smp;
112 u32 tail ____cacheline_aligned_in_smp;
116 * This data is shared with the application through the mmap at offsets
117 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
119 * The offsets to the member fields are published through struct
120 * io_sqring_offsets when calling io_uring_setup.
124 * Head and tail offsets into the ring; the offsets need to be
125 * masked to get valid indices.
127 * The kernel controls head of the sq ring and the tail of the cq ring,
128 * and the application controls tail of the sq ring and the head of the
131 struct io_uring sq, cq;
133 * Bitmasks to apply to head and tail offsets (constant, equals
136 u32 sq_ring_mask, cq_ring_mask;
137 /* Ring sizes (constant, power of 2) */
138 u32 sq_ring_entries, cq_ring_entries;
140 * Number of invalid entries dropped by the kernel due to
141 * invalid index stored in array
143 * Written by the kernel, shouldn't be modified by the
144 * application (i.e. get number of "new events" by comparing to
147 * After a new SQ head value was read by the application this
148 * counter includes all submissions that were dropped reaching
149 * the new SQ head (and possibly more).
155 * Written by the kernel, shouldn't be modified by the
158 * The application needs a full memory barrier before checking
159 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
165 * Written by the application, shouldn't be modified by the
170 * Number of completion events lost because the queue was full;
171 * this should be avoided by the application by making sure
172 * there are not more requests pending than there is space in
173 * the completion queue.
175 * Written by the kernel, shouldn't be modified by the
176 * application (i.e. get number of "new events" by comparing to
179 * As completion events come in out of order this counter is not
180 * ordered with any other data.
184 * Ring buffer of completion events.
186 * The kernel writes completion events fresh every time they are
187 * produced, so the application is allowed to modify pending
190 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
193 enum io_uring_cmd_flags {
194 IO_URING_F_NONBLOCK = 1,
195 IO_URING_F_COMPLETE_DEFER = 2,
198 struct io_mapped_ubuf {
201 struct bio_vec *bvec;
202 unsigned int nr_bvecs;
203 unsigned long acct_pages;
209 struct list_head list;
216 struct fixed_rsrc_table {
220 struct fixed_rsrc_ref_node {
221 struct percpu_ref refs;
222 struct list_head node;
223 struct list_head rsrc_list;
224 struct fixed_rsrc_data *rsrc_data;
225 void (*rsrc_put)(struct io_ring_ctx *ctx,
226 struct io_rsrc_put *prsrc);
227 struct llist_node llist;
231 struct fixed_rsrc_data {
232 struct fixed_rsrc_table *table;
233 struct io_ring_ctx *ctx;
235 struct fixed_rsrc_ref_node *node;
236 struct percpu_ref refs;
237 struct completion done;
242 struct list_head list;
248 struct io_restriction {
249 DECLARE_BITMAP(register_op, IORING_REGISTER_LAST);
250 DECLARE_BITMAP(sqe_op, IORING_OP_LAST);
251 u8 sqe_flags_allowed;
252 u8 sqe_flags_required;
257 IO_SQ_THREAD_SHOULD_STOP = 0,
258 IO_SQ_THREAD_SHOULD_PARK,
265 /* ctx's that are using this sqd */
266 struct list_head ctx_list;
267 struct list_head ctx_new_list;
268 struct mutex ctx_lock;
270 struct task_struct *thread;
271 struct wait_queue_head wait;
273 unsigned sq_thread_idle;
278 struct completion startup;
279 struct completion completion;
280 struct completion exited;
283 #define IO_IOPOLL_BATCH 8
284 #define IO_COMPL_BATCH 32
285 #define IO_REQ_CACHE_SIZE 32
286 #define IO_REQ_ALLOC_BATCH 8
288 struct io_comp_state {
289 struct io_kiocb *reqs[IO_COMPL_BATCH];
291 unsigned int locked_free_nr;
292 /* inline/task_work completion list, under ->uring_lock */
293 struct list_head free_list;
294 /* IRQ completion list, under ->completion_lock */
295 struct list_head locked_free_list;
298 struct io_submit_link {
299 struct io_kiocb *head;
300 struct io_kiocb *last;
303 struct io_submit_state {
304 struct blk_plug plug;
305 struct io_submit_link link;
308 * io_kiocb alloc cache
310 void *reqs[IO_REQ_CACHE_SIZE];
311 unsigned int free_reqs;
316 * Batch completion logic
318 struct io_comp_state comp;
321 * File reference cache
325 unsigned int file_refs;
326 unsigned int ios_left;
331 struct percpu_ref refs;
332 } ____cacheline_aligned_in_smp;
336 unsigned int compat: 1;
337 unsigned int cq_overflow_flushed: 1;
338 unsigned int drain_next: 1;
339 unsigned int eventfd_async: 1;
340 unsigned int restricted: 1;
341 unsigned int sqo_exec: 1;
344 * Ring buffer of indices into array of io_uring_sqe, which is
345 * mmapped by the application using the IORING_OFF_SQES offset.
347 * This indirection could e.g. be used to assign fixed
348 * io_uring_sqe entries to operations and only submit them to
349 * the queue when needed.
351 * The kernel modifies neither the indices array nor the entries
355 unsigned cached_sq_head;
358 unsigned sq_thread_idle;
359 unsigned cached_sq_dropped;
360 unsigned cached_cq_overflow;
361 unsigned long sq_check_overflow;
363 /* hashed buffered write serialization */
364 struct io_wq_hash *hash_map;
366 struct list_head defer_list;
367 struct list_head timeout_list;
368 struct list_head cq_overflow_list;
370 struct io_uring_sqe *sq_sqes;
371 } ____cacheline_aligned_in_smp;
374 struct mutex uring_lock;
375 wait_queue_head_t wait;
376 } ____cacheline_aligned_in_smp;
378 struct io_submit_state submit_state;
380 struct io_rings *rings;
382 /* Only used for accounting purposes */
383 struct mm_struct *mm_account;
385 struct io_sq_data *sq_data; /* if using sq thread polling */
387 struct wait_queue_head sqo_sq_wait;
388 struct list_head sqd_list;
391 * If used, fixed file set. Writers must ensure that ->refs is dead,
392 * readers must ensure that ->refs is alive as long as the file* is
393 * used. Only updated through io_uring_register(2).
395 struct fixed_rsrc_data *file_data;
396 unsigned nr_user_files;
398 /* if used, fixed mapped user buffers */
399 unsigned nr_user_bufs;
400 struct io_mapped_ubuf *user_bufs;
402 struct user_struct *user;
404 struct completion ref_comp;
405 struct completion sq_thread_comp;
407 #if defined(CONFIG_UNIX)
408 struct socket *ring_sock;
411 struct idr io_buffer_idr;
413 struct idr personality_idr;
416 unsigned cached_cq_tail;
419 atomic_t cq_timeouts;
420 unsigned cq_last_tm_flush;
421 unsigned long cq_check_overflow;
422 struct wait_queue_head cq_wait;
423 struct fasync_struct *cq_fasync;
424 struct eventfd_ctx *cq_ev_fd;
425 } ____cacheline_aligned_in_smp;
428 spinlock_t completion_lock;
431 * ->iopoll_list is protected by the ctx->uring_lock for
432 * io_uring instances that don't use IORING_SETUP_SQPOLL.
433 * For SQPOLL, only the single threaded io_sq_thread() will
434 * manipulate the list, hence no extra locking is needed there.
436 struct list_head iopoll_list;
437 struct hlist_head *cancel_hash;
438 unsigned cancel_hash_bits;
439 bool poll_multi_file;
441 spinlock_t inflight_lock;
442 struct list_head inflight_list;
443 } ____cacheline_aligned_in_smp;
445 struct delayed_work rsrc_put_work;
446 struct llist_head rsrc_put_llist;
447 struct list_head rsrc_ref_list;
448 spinlock_t rsrc_ref_lock;
450 struct io_restriction restrictions;
453 struct callback_head *exit_task_work;
455 struct wait_queue_head hash_wait;
457 /* Keep this last, we don't need it for the fast path */
458 struct work_struct exit_work;
462 * First field must be the file pointer in all the
463 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
465 struct io_poll_iocb {
467 struct wait_queue_head *head;
471 struct wait_queue_entry wait;
474 struct io_poll_remove {
484 struct io_timeout_data {
485 struct io_kiocb *req;
486 struct hrtimer timer;
487 struct timespec64 ts;
488 enum hrtimer_mode mode;
493 struct sockaddr __user *addr;
494 int __user *addr_len;
496 unsigned long nofile;
516 struct list_head list;
517 /* head of the link, used by linked timeouts only */
518 struct io_kiocb *head;
521 struct io_timeout_rem {
526 struct timespec64 ts;
531 /* NOTE: kiocb has the file as the first member, so don't do it here */
539 struct sockaddr __user *addr;
546 struct user_msghdr __user *umsg;
552 struct io_buffer *kbuf;
558 struct filename *filename;
560 unsigned long nofile;
563 struct io_rsrc_update {
589 struct epoll_event event;
593 struct file *file_out;
594 struct file *file_in;
601 struct io_provide_buf {
615 const char __user *filename;
616 struct statx __user *buffer;
628 struct filename *oldpath;
629 struct filename *newpath;
637 struct filename *filename;
640 struct io_completion {
642 struct list_head list;
646 struct io_async_connect {
647 struct sockaddr_storage address;
650 struct io_async_msghdr {
651 struct iovec fast_iov[UIO_FASTIOV];
652 /* points to an allocated iov, if NULL we use fast_iov instead */
653 struct iovec *free_iov;
654 struct sockaddr __user *uaddr;
656 struct sockaddr_storage addr;
660 struct iovec fast_iov[UIO_FASTIOV];
661 const struct iovec *free_iovec;
662 struct iov_iter iter;
664 struct wait_page_queue wpq;
668 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
669 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
670 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
671 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
672 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
673 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
679 REQ_F_LINK_TIMEOUT_BIT,
681 REQ_F_NEED_CLEANUP_BIT,
683 REQ_F_BUFFER_SELECTED_BIT,
684 REQ_F_NO_FILE_TABLE_BIT,
685 REQ_F_LTIMEOUT_ACTIVE_BIT,
686 REQ_F_COMPLETE_INLINE_BIT,
688 /* not a real bit, just to check we're not overflowing the space */
694 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
695 /* drain existing IO first */
696 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
698 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
699 /* doesn't sever on completion < 0 */
700 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
702 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
703 /* IOSQE_BUFFER_SELECT */
704 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
706 /* fail rest of links */
707 REQ_F_FAIL_LINK = BIT(REQ_F_FAIL_LINK_BIT),
708 /* on inflight list */
709 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
710 /* read/write uses file position */
711 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
712 /* must not punt to workers */
713 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
714 /* has or had linked timeout */
715 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
717 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
719 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
720 /* already went through poll handler */
721 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
722 /* buffer already selected */
723 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
724 /* doesn't need file table for this request */
725 REQ_F_NO_FILE_TABLE = BIT(REQ_F_NO_FILE_TABLE_BIT),
726 /* linked timeout is active, i.e. prepared by link's head */
727 REQ_F_LTIMEOUT_ACTIVE = BIT(REQ_F_LTIMEOUT_ACTIVE_BIT),
728 /* completion is deferred through io_comp_state */
729 REQ_F_COMPLETE_INLINE = BIT(REQ_F_COMPLETE_INLINE_BIT),
733 struct io_poll_iocb poll;
734 struct io_poll_iocb *double_poll;
737 struct io_task_work {
738 struct io_wq_work_node node;
739 task_work_func_t func;
743 * NOTE! Each of the iocb union members has the file pointer
744 * as the first entry in their struct definition. So you can
745 * access the file pointer through any of the sub-structs,
746 * or directly as just 'ki_filp' in this struct.
752 struct io_poll_iocb poll;
753 struct io_poll_remove poll_remove;
754 struct io_accept accept;
756 struct io_cancel cancel;
757 struct io_timeout timeout;
758 struct io_timeout_rem timeout_rem;
759 struct io_connect connect;
760 struct io_sr_msg sr_msg;
762 struct io_close close;
763 struct io_rsrc_update rsrc_update;
764 struct io_fadvise fadvise;
765 struct io_madvise madvise;
766 struct io_epoll epoll;
767 struct io_splice splice;
768 struct io_provide_buf pbuf;
769 struct io_statx statx;
770 struct io_shutdown shutdown;
771 struct io_rename rename;
772 struct io_unlink unlink;
773 /* use only after cleaning per-op data, see io_clean_op() */
774 struct io_completion compl;
777 /* opcode allocated if it needs to store data for async defer */
780 /* polled IO has completed */
786 struct io_ring_ctx *ctx;
789 struct task_struct *task;
792 struct io_kiocb *link;
793 struct percpu_ref *fixed_rsrc_refs;
796 * 1. used with ctx->iopoll_list with reads/writes
797 * 2. to track reqs with ->files (see io_op_def::file_table)
799 struct list_head inflight_entry;
801 struct io_task_work io_task_work;
802 struct callback_head task_work;
804 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
805 struct hlist_node hash_node;
806 struct async_poll *apoll;
807 struct io_wq_work work;
810 struct io_defer_entry {
811 struct list_head list;
812 struct io_kiocb *req;
817 /* needs req->file assigned */
818 unsigned needs_file : 1;
819 /* hash wq insertion if file is a regular file */
820 unsigned hash_reg_file : 1;
821 /* unbound wq insertion if file is a non-regular file */
822 unsigned unbound_nonreg_file : 1;
823 /* opcode is not supported by this kernel */
824 unsigned not_supported : 1;
825 /* set if opcode supports polled "wait" */
827 unsigned pollout : 1;
828 /* op supports buffer selection */
829 unsigned buffer_select : 1;
830 /* must always have async data allocated */
831 unsigned needs_async_data : 1;
832 /* should block plug */
834 /* size of async data needed, if any */
835 unsigned short async_size;
838 static const struct io_op_def io_op_defs[] = {
839 [IORING_OP_NOP] = {},
840 [IORING_OP_READV] = {
842 .unbound_nonreg_file = 1,
845 .needs_async_data = 1,
847 .async_size = sizeof(struct io_async_rw),
849 [IORING_OP_WRITEV] = {
852 .unbound_nonreg_file = 1,
854 .needs_async_data = 1,
856 .async_size = sizeof(struct io_async_rw),
858 [IORING_OP_FSYNC] = {
861 [IORING_OP_READ_FIXED] = {
863 .unbound_nonreg_file = 1,
866 .async_size = sizeof(struct io_async_rw),
868 [IORING_OP_WRITE_FIXED] = {
871 .unbound_nonreg_file = 1,
874 .async_size = sizeof(struct io_async_rw),
876 [IORING_OP_POLL_ADD] = {
878 .unbound_nonreg_file = 1,
880 [IORING_OP_POLL_REMOVE] = {},
881 [IORING_OP_SYNC_FILE_RANGE] = {
884 [IORING_OP_SENDMSG] = {
886 .unbound_nonreg_file = 1,
888 .needs_async_data = 1,
889 .async_size = sizeof(struct io_async_msghdr),
891 [IORING_OP_RECVMSG] = {
893 .unbound_nonreg_file = 1,
896 .needs_async_data = 1,
897 .async_size = sizeof(struct io_async_msghdr),
899 [IORING_OP_TIMEOUT] = {
900 .needs_async_data = 1,
901 .async_size = sizeof(struct io_timeout_data),
903 [IORING_OP_TIMEOUT_REMOVE] = {
904 /* used by timeout updates' prep() */
906 [IORING_OP_ACCEPT] = {
908 .unbound_nonreg_file = 1,
911 [IORING_OP_ASYNC_CANCEL] = {},
912 [IORING_OP_LINK_TIMEOUT] = {
913 .needs_async_data = 1,
914 .async_size = sizeof(struct io_timeout_data),
916 [IORING_OP_CONNECT] = {
918 .unbound_nonreg_file = 1,
920 .needs_async_data = 1,
921 .async_size = sizeof(struct io_async_connect),
923 [IORING_OP_FALLOCATE] = {
926 [IORING_OP_OPENAT] = {},
927 [IORING_OP_CLOSE] = {},
928 [IORING_OP_FILES_UPDATE] = {},
929 [IORING_OP_STATX] = {},
932 .unbound_nonreg_file = 1,
936 .async_size = sizeof(struct io_async_rw),
938 [IORING_OP_WRITE] = {
940 .unbound_nonreg_file = 1,
943 .async_size = sizeof(struct io_async_rw),
945 [IORING_OP_FADVISE] = {
948 [IORING_OP_MADVISE] = {},
951 .unbound_nonreg_file = 1,
956 .unbound_nonreg_file = 1,
960 [IORING_OP_OPENAT2] = {
962 [IORING_OP_EPOLL_CTL] = {
963 .unbound_nonreg_file = 1,
965 [IORING_OP_SPLICE] = {
968 .unbound_nonreg_file = 1,
970 [IORING_OP_PROVIDE_BUFFERS] = {},
971 [IORING_OP_REMOVE_BUFFERS] = {},
975 .unbound_nonreg_file = 1,
977 [IORING_OP_SHUTDOWN] = {
980 [IORING_OP_RENAMEAT] = {},
981 [IORING_OP_UNLINKAT] = {},
984 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
985 struct task_struct *task,
986 struct files_struct *files);
987 static void io_uring_cancel_sqpoll(struct io_ring_ctx *ctx);
988 static void destroy_fixed_rsrc_ref_node(struct fixed_rsrc_ref_node *ref_node);
989 static struct fixed_rsrc_ref_node *alloc_fixed_rsrc_ref_node(
990 struct io_ring_ctx *ctx);
991 static void io_ring_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc);
993 static bool io_rw_reissue(struct io_kiocb *req);
994 static void io_cqring_fill_event(struct io_kiocb *req, long res);
995 static void io_put_req(struct io_kiocb *req);
996 static void io_put_req_deferred(struct io_kiocb *req, int nr);
997 static void io_double_put_req(struct io_kiocb *req);
998 static void io_dismantle_req(struct io_kiocb *req);
999 static void io_put_task(struct task_struct *task, int nr);
1000 static void io_queue_next(struct io_kiocb *req);
1001 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
1002 static void __io_queue_linked_timeout(struct io_kiocb *req);
1003 static void io_queue_linked_timeout(struct io_kiocb *req);
1004 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
1005 struct io_uring_rsrc_update *ip,
1007 static void __io_clean_op(struct io_kiocb *req);
1008 static struct file *io_file_get(struct io_submit_state *state,
1009 struct io_kiocb *req, int fd, bool fixed);
1010 static void __io_queue_sqe(struct io_kiocb *req);
1011 static void io_rsrc_put_work(struct work_struct *work);
1013 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
1014 struct iov_iter *iter, bool needs_lock);
1015 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
1016 const struct iovec *fast_iov,
1017 struct iov_iter *iter, bool force);
1018 static void io_req_task_queue(struct io_kiocb *req);
1019 static void io_submit_flush_completions(struct io_comp_state *cs,
1020 struct io_ring_ctx *ctx);
1022 static struct kmem_cache *req_cachep;
1024 static const struct file_operations io_uring_fops;
1026 struct sock *io_uring_get_socket(struct file *file)
1028 #if defined(CONFIG_UNIX)
1029 if (file->f_op == &io_uring_fops) {
1030 struct io_ring_ctx *ctx = file->private_data;
1032 return ctx->ring_sock->sk;
1037 EXPORT_SYMBOL(io_uring_get_socket);
1039 #define io_for_each_link(pos, head) \
1040 for (pos = (head); pos; pos = pos->link)
1042 static inline void io_clean_op(struct io_kiocb *req)
1044 if (req->flags & (REQ_F_NEED_CLEANUP | REQ_F_BUFFER_SELECTED))
1048 static inline void io_set_resource_node(struct io_kiocb *req)
1050 struct io_ring_ctx *ctx = req->ctx;
1052 if (!req->fixed_rsrc_refs) {
1053 req->fixed_rsrc_refs = &ctx->file_data->node->refs;
1054 percpu_ref_get(req->fixed_rsrc_refs);
1058 static bool io_match_task(struct io_kiocb *head,
1059 struct task_struct *task,
1060 struct files_struct *files)
1062 struct io_kiocb *req;
1064 if (task && head->task != task) {
1065 /* in terms of cancelation, always match if req task is dead */
1066 if (head->task->flags & PF_EXITING)
1073 io_for_each_link(req, head) {
1074 if (req->file && req->file->f_op == &io_uring_fops)
1076 if (req->task->files == files)
1082 static inline void req_set_fail_links(struct io_kiocb *req)
1084 if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
1085 req->flags |= REQ_F_FAIL_LINK;
1088 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1090 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1092 complete(&ctx->ref_comp);
1095 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1097 return !req->timeout.off;
1100 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1102 struct io_ring_ctx *ctx;
1105 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1110 * Use 5 bits less than the max cq entries, that should give us around
1111 * 32 entries per hash list if totally full and uniformly spread.
1113 hash_bits = ilog2(p->cq_entries);
1117 ctx->cancel_hash_bits = hash_bits;
1118 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1120 if (!ctx->cancel_hash)
1122 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1124 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1125 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1128 ctx->flags = p->flags;
1129 init_waitqueue_head(&ctx->sqo_sq_wait);
1130 INIT_LIST_HEAD(&ctx->sqd_list);
1131 init_waitqueue_head(&ctx->cq_wait);
1132 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1133 init_completion(&ctx->ref_comp);
1134 init_completion(&ctx->sq_thread_comp);
1135 idr_init(&ctx->io_buffer_idr);
1136 idr_init(&ctx->personality_idr);
1137 mutex_init(&ctx->uring_lock);
1138 init_waitqueue_head(&ctx->wait);
1139 spin_lock_init(&ctx->completion_lock);
1140 INIT_LIST_HEAD(&ctx->iopoll_list);
1141 INIT_LIST_HEAD(&ctx->defer_list);
1142 INIT_LIST_HEAD(&ctx->timeout_list);
1143 spin_lock_init(&ctx->inflight_lock);
1144 INIT_LIST_HEAD(&ctx->inflight_list);
1145 spin_lock_init(&ctx->rsrc_ref_lock);
1146 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
1147 INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
1148 init_llist_head(&ctx->rsrc_put_llist);
1149 INIT_LIST_HEAD(&ctx->submit_state.comp.free_list);
1150 INIT_LIST_HEAD(&ctx->submit_state.comp.locked_free_list);
1153 kfree(ctx->cancel_hash);
1158 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1160 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1161 struct io_ring_ctx *ctx = req->ctx;
1163 return seq != ctx->cached_cq_tail
1164 + READ_ONCE(ctx->cached_cq_overflow);
1170 static void io_req_clean_work(struct io_kiocb *req)
1172 if (req->flags & REQ_F_INFLIGHT) {
1173 struct io_ring_ctx *ctx = req->ctx;
1174 struct io_uring_task *tctx = req->task->io_uring;
1175 unsigned long flags;
1177 spin_lock_irqsave(&ctx->inflight_lock, flags);
1178 list_del(&req->inflight_entry);
1179 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
1180 req->flags &= ~REQ_F_INFLIGHT;
1181 if (atomic_read(&tctx->in_idle))
1182 wake_up(&tctx->wait);
1186 static void io_req_track_inflight(struct io_kiocb *req)
1188 struct io_ring_ctx *ctx = req->ctx;
1190 if (!(req->flags & REQ_F_INFLIGHT)) {
1191 req->flags |= REQ_F_INFLIGHT;
1193 spin_lock_irq(&ctx->inflight_lock);
1194 list_add(&req->inflight_entry, &ctx->inflight_list);
1195 spin_unlock_irq(&ctx->inflight_lock);
1199 static void io_prep_async_work(struct io_kiocb *req)
1201 const struct io_op_def *def = &io_op_defs[req->opcode];
1202 struct io_ring_ctx *ctx = req->ctx;
1204 if (req->flags & REQ_F_FORCE_ASYNC)
1205 req->work.flags |= IO_WQ_WORK_CONCURRENT;
1207 if (req->flags & REQ_F_ISREG) {
1208 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1209 io_wq_hash_work(&req->work, file_inode(req->file));
1211 if (def->unbound_nonreg_file)
1212 req->work.flags |= IO_WQ_WORK_UNBOUND;
1216 static void io_prep_async_link(struct io_kiocb *req)
1218 struct io_kiocb *cur;
1220 io_for_each_link(cur, req)
1221 io_prep_async_work(cur);
1224 static struct io_kiocb *__io_queue_async_work(struct io_kiocb *req)
1226 struct io_ring_ctx *ctx = req->ctx;
1227 struct io_kiocb *link = io_prep_linked_timeout(req);
1228 struct io_uring_task *tctx = req->task->io_uring;
1231 BUG_ON(!tctx->io_wq);
1233 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1234 &req->work, req->flags);
1235 io_wq_enqueue(tctx->io_wq, &req->work);
1239 static void io_queue_async_work(struct io_kiocb *req)
1241 struct io_kiocb *link;
1243 /* init ->work of the whole link before punting */
1244 io_prep_async_link(req);
1245 link = __io_queue_async_work(req);
1248 io_queue_linked_timeout(link);
1251 static void io_kill_timeout(struct io_kiocb *req)
1253 struct io_timeout_data *io = req->async_data;
1256 ret = hrtimer_try_to_cancel(&io->timer);
1258 atomic_set(&req->ctx->cq_timeouts,
1259 atomic_read(&req->ctx->cq_timeouts) + 1);
1260 list_del_init(&req->timeout.list);
1261 io_cqring_fill_event(req, 0);
1262 io_put_req_deferred(req, 1);
1267 * Returns true if we found and killed one or more timeouts
1269 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
1270 struct files_struct *files)
1272 struct io_kiocb *req, *tmp;
1275 spin_lock_irq(&ctx->completion_lock);
1276 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
1277 if (io_match_task(req, tsk, files)) {
1278 io_kill_timeout(req);
1282 spin_unlock_irq(&ctx->completion_lock);
1283 return canceled != 0;
1286 static void __io_queue_deferred(struct io_ring_ctx *ctx)
1289 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1290 struct io_defer_entry, list);
1292 if (req_need_defer(de->req, de->seq))
1294 list_del_init(&de->list);
1295 io_req_task_queue(de->req);
1297 } while (!list_empty(&ctx->defer_list));
1300 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1304 if (list_empty(&ctx->timeout_list))
1307 seq = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
1310 u32 events_needed, events_got;
1311 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1312 struct io_kiocb, timeout.list);
1314 if (io_is_timeout_noseq(req))
1318 * Since seq can easily wrap around over time, subtract
1319 * the last seq at which timeouts were flushed before comparing.
1320 * Assuming not more than 2^31-1 events have happened since,
1321 * these subtractions won't have wrapped, so we can check if
1322 * target is in [last_seq, current_seq] by comparing the two.
1324 events_needed = req->timeout.target_seq - ctx->cq_last_tm_flush;
1325 events_got = seq - ctx->cq_last_tm_flush;
1326 if (events_got < events_needed)
1329 list_del_init(&req->timeout.list);
1330 io_kill_timeout(req);
1331 } while (!list_empty(&ctx->timeout_list));
1333 ctx->cq_last_tm_flush = seq;
1336 static void io_commit_cqring(struct io_ring_ctx *ctx)
1338 io_flush_timeouts(ctx);
1340 /* order cqe stores with ring update */
1341 smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
1343 if (unlikely(!list_empty(&ctx->defer_list)))
1344 __io_queue_deferred(ctx);
1347 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1349 struct io_rings *r = ctx->rings;
1351 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == r->sq_ring_entries;
1354 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
1356 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
1359 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1361 struct io_rings *rings = ctx->rings;
1365 * writes to the cq entry need to come after reading head; the
1366 * control dependency is enough as we're using WRITE_ONCE to
1369 if (__io_cqring_events(ctx) == rings->cq_ring_entries)
1372 tail = ctx->cached_cq_tail++;
1373 return &rings->cqes[tail & ctx->cq_mask];
1376 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1380 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1382 if (!ctx->eventfd_async)
1384 return io_wq_current_is_worker();
1387 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1389 /* see waitqueue_active() comment */
1392 if (waitqueue_active(&ctx->wait))
1393 wake_up(&ctx->wait);
1394 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1395 wake_up(&ctx->sq_data->wait);
1396 if (io_should_trigger_evfd(ctx))
1397 eventfd_signal(ctx->cq_ev_fd, 1);
1398 if (waitqueue_active(&ctx->cq_wait)) {
1399 wake_up_interruptible(&ctx->cq_wait);
1400 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1404 static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx)
1406 /* see waitqueue_active() comment */
1409 if (ctx->flags & IORING_SETUP_SQPOLL) {
1410 if (waitqueue_active(&ctx->wait))
1411 wake_up(&ctx->wait);
1413 if (io_should_trigger_evfd(ctx))
1414 eventfd_signal(ctx->cq_ev_fd, 1);
1415 if (waitqueue_active(&ctx->cq_wait)) {
1416 wake_up_interruptible(&ctx->cq_wait);
1417 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1421 /* Returns true if there are no backlogged entries after the flush */
1422 static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force,
1423 struct task_struct *tsk,
1424 struct files_struct *files)
1426 struct io_rings *rings = ctx->rings;
1427 struct io_kiocb *req, *tmp;
1428 struct io_uring_cqe *cqe;
1429 unsigned long flags;
1430 bool all_flushed, posted;
1433 if (!force && __io_cqring_events(ctx) == rings->cq_ring_entries)
1437 spin_lock_irqsave(&ctx->completion_lock, flags);
1438 list_for_each_entry_safe(req, tmp, &ctx->cq_overflow_list, compl.list) {
1439 if (!io_match_task(req, tsk, files))
1442 cqe = io_get_cqring(ctx);
1446 list_move(&req->compl.list, &list);
1448 WRITE_ONCE(cqe->user_data, req->user_data);
1449 WRITE_ONCE(cqe->res, req->result);
1450 WRITE_ONCE(cqe->flags, req->compl.cflags);
1452 ctx->cached_cq_overflow++;
1453 WRITE_ONCE(ctx->rings->cq_overflow,
1454 ctx->cached_cq_overflow);
1459 all_flushed = list_empty(&ctx->cq_overflow_list);
1461 clear_bit(0, &ctx->sq_check_overflow);
1462 clear_bit(0, &ctx->cq_check_overflow);
1463 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
1467 io_commit_cqring(ctx);
1468 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1470 io_cqring_ev_posted(ctx);
1472 while (!list_empty(&list)) {
1473 req = list_first_entry(&list, struct io_kiocb, compl.list);
1474 list_del(&req->compl.list);
1481 static void io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force,
1482 struct task_struct *tsk,
1483 struct files_struct *files)
1485 if (test_bit(0, &ctx->cq_check_overflow)) {
1486 /* iopoll syncs against uring_lock, not completion_lock */
1487 if (ctx->flags & IORING_SETUP_IOPOLL)
1488 mutex_lock(&ctx->uring_lock);
1489 __io_cqring_overflow_flush(ctx, force, tsk, files);
1490 if (ctx->flags & IORING_SETUP_IOPOLL)
1491 mutex_unlock(&ctx->uring_lock);
1495 static void __io_cqring_fill_event(struct io_kiocb *req, long res, long cflags)
1497 struct io_ring_ctx *ctx = req->ctx;
1498 struct io_uring_cqe *cqe;
1500 trace_io_uring_complete(ctx, req->user_data, res);
1503 * If we can't get a cq entry, userspace overflowed the
1504 * submission (by quite a lot). Increment the overflow count in
1507 cqe = io_get_cqring(ctx);
1509 WRITE_ONCE(cqe->user_data, req->user_data);
1510 WRITE_ONCE(cqe->res, res);
1511 WRITE_ONCE(cqe->flags, cflags);
1512 } else if (ctx->cq_overflow_flushed ||
1513 atomic_read(&req->task->io_uring->in_idle)) {
1515 * If we're in ring overflow flush mode, or in task cancel mode,
1516 * then we cannot store the request for later flushing, we need
1517 * to drop it on the floor.
1519 ctx->cached_cq_overflow++;
1520 WRITE_ONCE(ctx->rings->cq_overflow, ctx->cached_cq_overflow);
1522 if (list_empty(&ctx->cq_overflow_list)) {
1523 set_bit(0, &ctx->sq_check_overflow);
1524 set_bit(0, &ctx->cq_check_overflow);
1525 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
1529 req->compl.cflags = cflags;
1530 refcount_inc(&req->refs);
1531 list_add_tail(&req->compl.list, &ctx->cq_overflow_list);
1535 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1537 __io_cqring_fill_event(req, res, 0);
1540 static inline void io_req_complete_post(struct io_kiocb *req, long res,
1541 unsigned int cflags)
1543 struct io_ring_ctx *ctx = req->ctx;
1544 unsigned long flags;
1546 spin_lock_irqsave(&ctx->completion_lock, flags);
1547 __io_cqring_fill_event(req, res, cflags);
1548 io_commit_cqring(ctx);
1550 * If we're the last reference to this request, add to our locked
1553 if (refcount_dec_and_test(&req->refs)) {
1554 struct io_comp_state *cs = &ctx->submit_state.comp;
1556 io_dismantle_req(req);
1557 io_put_task(req->task, 1);
1558 list_add(&req->compl.list, &cs->locked_free_list);
1559 cs->locked_free_nr++;
1562 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1564 io_cqring_ev_posted(ctx);
1567 percpu_ref_put(&ctx->refs);
1571 static void io_req_complete_state(struct io_kiocb *req, long res,
1572 unsigned int cflags)
1576 req->compl.cflags = cflags;
1577 req->flags |= REQ_F_COMPLETE_INLINE;
1580 static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags,
1581 long res, unsigned cflags)
1583 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1584 io_req_complete_state(req, res, cflags);
1586 io_req_complete_post(req, res, cflags);
1589 static inline void io_req_complete(struct io_kiocb *req, long res)
1591 __io_req_complete(req, 0, res, 0);
1594 static bool io_flush_cached_reqs(struct io_ring_ctx *ctx)
1596 struct io_submit_state *state = &ctx->submit_state;
1597 struct io_comp_state *cs = &state->comp;
1598 struct io_kiocb *req = NULL;
1601 * If we have more than a batch's worth of requests in our IRQ side
1602 * locked cache, grab the lock and move them over to our submission
1605 if (READ_ONCE(cs->locked_free_nr) > IO_COMPL_BATCH) {
1606 spin_lock_irq(&ctx->completion_lock);
1607 list_splice_init(&cs->locked_free_list, &cs->free_list);
1608 cs->locked_free_nr = 0;
1609 spin_unlock_irq(&ctx->completion_lock);
1612 while (!list_empty(&cs->free_list)) {
1613 req = list_first_entry(&cs->free_list, struct io_kiocb,
1615 list_del(&req->compl.list);
1616 state->reqs[state->free_reqs++] = req;
1617 if (state->free_reqs == ARRAY_SIZE(state->reqs))
1624 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
1626 struct io_submit_state *state = &ctx->submit_state;
1628 BUILD_BUG_ON(IO_REQ_ALLOC_BATCH > ARRAY_SIZE(state->reqs));
1630 if (!state->free_reqs) {
1631 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1634 if (io_flush_cached_reqs(ctx))
1637 ret = kmem_cache_alloc_bulk(req_cachep, gfp, IO_REQ_ALLOC_BATCH,
1641 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1642 * retry single alloc to be on the safe side.
1644 if (unlikely(ret <= 0)) {
1645 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1646 if (!state->reqs[0])
1650 state->free_reqs = ret;
1654 return state->reqs[state->free_reqs];
1657 static inline void io_put_file(struct io_kiocb *req, struct file *file,
1664 static void io_dismantle_req(struct io_kiocb *req)
1668 if (req->async_data)
1669 kfree(req->async_data);
1671 io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
1672 if (req->fixed_rsrc_refs)
1673 percpu_ref_put(req->fixed_rsrc_refs);
1674 io_req_clean_work(req);
1677 static inline void io_put_task(struct task_struct *task, int nr)
1679 struct io_uring_task *tctx = task->io_uring;
1681 percpu_counter_sub(&tctx->inflight, nr);
1682 if (unlikely(atomic_read(&tctx->in_idle)))
1683 wake_up(&tctx->wait);
1684 put_task_struct_many(task, nr);
1687 static void __io_free_req(struct io_kiocb *req)
1689 struct io_ring_ctx *ctx = req->ctx;
1691 io_dismantle_req(req);
1692 io_put_task(req->task, 1);
1694 kmem_cache_free(req_cachep, req);
1695 percpu_ref_put(&ctx->refs);
1698 static inline void io_remove_next_linked(struct io_kiocb *req)
1700 struct io_kiocb *nxt = req->link;
1702 req->link = nxt->link;
1706 static void io_kill_linked_timeout(struct io_kiocb *req)
1708 struct io_ring_ctx *ctx = req->ctx;
1709 struct io_kiocb *link;
1710 bool cancelled = false;
1711 unsigned long flags;
1713 spin_lock_irqsave(&ctx->completion_lock, flags);
1717 * Can happen if a linked timeout fired and link had been like
1718 * req -> link t-out -> link t-out [-> ...]
1720 if (link && (link->flags & REQ_F_LTIMEOUT_ACTIVE)) {
1721 struct io_timeout_data *io = link->async_data;
1724 io_remove_next_linked(req);
1725 link->timeout.head = NULL;
1726 ret = hrtimer_try_to_cancel(&io->timer);
1728 io_cqring_fill_event(link, -ECANCELED);
1729 io_commit_cqring(ctx);
1733 req->flags &= ~REQ_F_LINK_TIMEOUT;
1734 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1737 io_cqring_ev_posted(ctx);
1743 static void io_fail_links(struct io_kiocb *req)
1745 struct io_kiocb *link, *nxt;
1746 struct io_ring_ctx *ctx = req->ctx;
1747 unsigned long flags;
1749 spin_lock_irqsave(&ctx->completion_lock, flags);
1757 trace_io_uring_fail_link(req, link);
1758 io_cqring_fill_event(link, -ECANCELED);
1760 io_put_req_deferred(link, 2);
1763 io_commit_cqring(ctx);
1764 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1766 io_cqring_ev_posted(ctx);
1769 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1771 if (req->flags & REQ_F_LINK_TIMEOUT)
1772 io_kill_linked_timeout(req);
1775 * If LINK is set, we have dependent requests in this chain. If we
1776 * didn't fail this request, queue the first one up, moving any other
1777 * dependencies to the next request. In case of failure, fail the rest
1780 if (likely(!(req->flags & REQ_F_FAIL_LINK))) {
1781 struct io_kiocb *nxt = req->link;
1790 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1792 if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK))))
1794 return __io_req_find_next(req);
1797 static void ctx_flush_and_put(struct io_ring_ctx *ctx)
1801 if (ctx->submit_state.comp.nr) {
1802 mutex_lock(&ctx->uring_lock);
1803 io_submit_flush_completions(&ctx->submit_state.comp, ctx);
1804 mutex_unlock(&ctx->uring_lock);
1806 percpu_ref_put(&ctx->refs);
1809 static bool __tctx_task_work(struct io_uring_task *tctx)
1811 struct io_ring_ctx *ctx = NULL;
1812 struct io_wq_work_list list;
1813 struct io_wq_work_node *node;
1815 if (wq_list_empty(&tctx->task_list))
1818 spin_lock_irq(&tctx->task_lock);
1819 list = tctx->task_list;
1820 INIT_WQ_LIST(&tctx->task_list);
1821 spin_unlock_irq(&tctx->task_lock);
1825 struct io_wq_work_node *next = node->next;
1826 struct io_kiocb *req;
1828 req = container_of(node, struct io_kiocb, io_task_work.node);
1829 if (req->ctx != ctx) {
1830 ctx_flush_and_put(ctx);
1832 percpu_ref_get(&ctx->refs);
1835 req->task_work.func(&req->task_work);
1839 ctx_flush_and_put(ctx);
1840 return list.first != NULL;
1843 static void tctx_task_work(struct callback_head *cb)
1845 struct io_uring_task *tctx = container_of(cb, struct io_uring_task, task_work);
1847 clear_bit(0, &tctx->task_state);
1849 while (__tctx_task_work(tctx))
1853 static int io_task_work_add(struct task_struct *tsk, struct io_kiocb *req,
1854 enum task_work_notify_mode notify)
1856 struct io_uring_task *tctx = tsk->io_uring;
1857 struct io_wq_work_node *node, *prev;
1858 unsigned long flags;
1861 WARN_ON_ONCE(!tctx);
1863 spin_lock_irqsave(&tctx->task_lock, flags);
1864 wq_list_add_tail(&req->io_task_work.node, &tctx->task_list);
1865 spin_unlock_irqrestore(&tctx->task_lock, flags);
1867 /* task_work already pending, we're done */
1868 if (test_bit(0, &tctx->task_state) ||
1869 test_and_set_bit(0, &tctx->task_state))
1872 if (!task_work_add(tsk, &tctx->task_work, notify))
1876 * Slow path - we failed, find and delete work. if the work is not
1877 * in the list, it got run and we're fine.
1880 spin_lock_irqsave(&tctx->task_lock, flags);
1881 wq_list_for_each(node, prev, &tctx->task_list) {
1882 if (&req->io_task_work.node == node) {
1883 wq_list_del(&tctx->task_list, node, prev);
1888 spin_unlock_irqrestore(&tctx->task_lock, flags);
1889 clear_bit(0, &tctx->task_state);
1893 static int io_req_task_work_add(struct io_kiocb *req)
1895 struct task_struct *tsk = req->task;
1896 struct io_ring_ctx *ctx = req->ctx;
1897 enum task_work_notify_mode notify;
1900 if (tsk->flags & PF_EXITING)
1904 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
1905 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
1906 * processing task_work. There's no reliable way to tell if TWA_RESUME
1910 if (!(ctx->flags & IORING_SETUP_SQPOLL))
1911 notify = TWA_SIGNAL;
1913 ret = io_task_work_add(tsk, req, notify);
1915 wake_up_process(tsk);
1920 static void io_req_task_work_add_fallback(struct io_kiocb *req,
1921 task_work_func_t cb)
1923 struct io_ring_ctx *ctx = req->ctx;
1924 struct callback_head *head;
1926 init_task_work(&req->task_work, cb);
1928 head = READ_ONCE(ctx->exit_task_work);
1929 req->task_work.next = head;
1930 } while (cmpxchg(&ctx->exit_task_work, head, &req->task_work) != head);
1933 static void __io_req_task_cancel(struct io_kiocb *req, int error)
1935 struct io_ring_ctx *ctx = req->ctx;
1937 spin_lock_irq(&ctx->completion_lock);
1938 io_cqring_fill_event(req, error);
1939 io_commit_cqring(ctx);
1940 spin_unlock_irq(&ctx->completion_lock);
1942 io_cqring_ev_posted(ctx);
1943 req_set_fail_links(req);
1944 io_double_put_req(req);
1947 static void io_req_task_cancel(struct callback_head *cb)
1949 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1950 struct io_ring_ctx *ctx = req->ctx;
1952 mutex_lock(&ctx->uring_lock);
1953 __io_req_task_cancel(req, req->result);
1954 mutex_unlock(&ctx->uring_lock);
1955 percpu_ref_put(&ctx->refs);
1958 static void __io_req_task_submit(struct io_kiocb *req)
1960 struct io_ring_ctx *ctx = req->ctx;
1962 /* ctx stays valid until unlock, even if we drop all ours ctx->refs */
1963 mutex_lock(&ctx->uring_lock);
1964 if (!(current->flags & PF_EXITING) && !current->in_execve)
1965 __io_queue_sqe(req);
1967 __io_req_task_cancel(req, -EFAULT);
1968 mutex_unlock(&ctx->uring_lock);
1971 static void io_req_task_submit(struct callback_head *cb)
1973 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1975 __io_req_task_submit(req);
1978 static void io_req_task_queue(struct io_kiocb *req)
1982 req->task_work.func = io_req_task_submit;
1983 ret = io_req_task_work_add(req);
1984 if (unlikely(ret)) {
1985 req->result = -ECANCELED;
1986 percpu_ref_get(&req->ctx->refs);
1987 io_req_task_work_add_fallback(req, io_req_task_cancel);
1991 static void io_req_task_queue_fail(struct io_kiocb *req, int ret)
1993 percpu_ref_get(&req->ctx->refs);
1995 req->task_work.func = io_req_task_cancel;
1997 if (unlikely(io_req_task_work_add(req)))
1998 io_req_task_work_add_fallback(req, io_req_task_cancel);
2001 static inline void io_queue_next(struct io_kiocb *req)
2003 struct io_kiocb *nxt = io_req_find_next(req);
2006 io_req_task_queue(nxt);
2009 static void io_free_req(struct io_kiocb *req)
2016 struct task_struct *task;
2021 static inline void io_init_req_batch(struct req_batch *rb)
2028 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2029 struct req_batch *rb)
2032 io_put_task(rb->task, rb->task_refs);
2034 percpu_ref_put_many(&ctx->refs, rb->ctx_refs);
2037 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req,
2038 struct io_submit_state *state)
2042 if (req->task != rb->task) {
2044 io_put_task(rb->task, rb->task_refs);
2045 rb->task = req->task;
2051 io_dismantle_req(req);
2052 if (state->free_reqs != ARRAY_SIZE(state->reqs))
2053 state->reqs[state->free_reqs++] = req;
2055 list_add(&req->compl.list, &state->comp.free_list);
2058 static void io_submit_flush_completions(struct io_comp_state *cs,
2059 struct io_ring_ctx *ctx)
2062 struct io_kiocb *req;
2063 struct req_batch rb;
2065 io_init_req_batch(&rb);
2066 spin_lock_irq(&ctx->completion_lock);
2067 for (i = 0; i < nr; i++) {
2069 __io_cqring_fill_event(req, req->result, req->compl.cflags);
2071 io_commit_cqring(ctx);
2072 spin_unlock_irq(&ctx->completion_lock);
2074 io_cqring_ev_posted(ctx);
2075 for (i = 0; i < nr; i++) {
2078 /* submission and completion refs */
2079 if (refcount_sub_and_test(2, &req->refs))
2080 io_req_free_batch(&rb, req, &ctx->submit_state);
2083 io_req_free_batch_finish(ctx, &rb);
2088 * Drop reference to request, return next in chain (if there is one) if this
2089 * was the last reference to this request.
2091 static struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2093 struct io_kiocb *nxt = NULL;
2095 if (refcount_dec_and_test(&req->refs)) {
2096 nxt = io_req_find_next(req);
2102 static void io_put_req(struct io_kiocb *req)
2104 if (refcount_dec_and_test(&req->refs))
2108 static void io_put_req_deferred_cb(struct callback_head *cb)
2110 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2115 static void io_free_req_deferred(struct io_kiocb *req)
2119 req->task_work.func = io_put_req_deferred_cb;
2120 ret = io_req_task_work_add(req);
2122 io_req_task_work_add_fallback(req, io_put_req_deferred_cb);
2125 static inline void io_put_req_deferred(struct io_kiocb *req, int refs)
2127 if (refcount_sub_and_test(refs, &req->refs))
2128 io_free_req_deferred(req);
2131 static void io_double_put_req(struct io_kiocb *req)
2133 /* drop both submit and complete references */
2134 if (refcount_sub_and_test(2, &req->refs))
2138 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2140 /* See comment at the top of this file */
2142 return __io_cqring_events(ctx);
2145 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2147 struct io_rings *rings = ctx->rings;
2149 /* make sure SQ entry isn't read before tail */
2150 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2153 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2155 unsigned int cflags;
2157 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2158 cflags |= IORING_CQE_F_BUFFER;
2159 req->flags &= ~REQ_F_BUFFER_SELECTED;
2164 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2166 struct io_buffer *kbuf;
2168 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2169 return io_put_kbuf(req, kbuf);
2172 static inline bool io_run_task_work(void)
2175 * Not safe to run on exiting task, and the task_work handling will
2176 * not add work to such a task.
2178 if (unlikely(current->flags & PF_EXITING))
2180 if (current->task_works) {
2181 __set_current_state(TASK_RUNNING);
2190 * Find and free completed poll iocbs
2192 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2193 struct list_head *done)
2195 struct req_batch rb;
2196 struct io_kiocb *req;
2198 /* order with ->result store in io_complete_rw_iopoll() */
2201 io_init_req_batch(&rb);
2202 while (!list_empty(done)) {
2205 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2206 list_del(&req->inflight_entry);
2208 if (READ_ONCE(req->result) == -EAGAIN) {
2209 req->iopoll_completed = 0;
2210 if (io_rw_reissue(req))
2214 if (req->flags & REQ_F_BUFFER_SELECTED)
2215 cflags = io_put_rw_kbuf(req);
2217 __io_cqring_fill_event(req, req->result, cflags);
2220 if (refcount_dec_and_test(&req->refs))
2221 io_req_free_batch(&rb, req, &ctx->submit_state);
2224 io_commit_cqring(ctx);
2225 io_cqring_ev_posted_iopoll(ctx);
2226 io_req_free_batch_finish(ctx, &rb);
2229 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2232 struct io_kiocb *req, *tmp;
2238 * Only spin for completions if we don't have multiple devices hanging
2239 * off our complete list, and we're under the requested amount.
2241 spin = !ctx->poll_multi_file && *nr_events < min;
2244 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2245 struct kiocb *kiocb = &req->rw.kiocb;
2248 * Move completed and retryable entries to our local lists.
2249 * If we find a request that requires polling, break out
2250 * and complete those lists first, if we have entries there.
2252 if (READ_ONCE(req->iopoll_completed)) {
2253 list_move_tail(&req->inflight_entry, &done);
2256 if (!list_empty(&done))
2259 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2263 /* iopoll may have completed current req */
2264 if (READ_ONCE(req->iopoll_completed))
2265 list_move_tail(&req->inflight_entry, &done);
2272 if (!list_empty(&done))
2273 io_iopoll_complete(ctx, nr_events, &done);
2279 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2280 * non-spinning poll check - we'll still enter the driver poll loop, but only
2281 * as a non-spinning completion check.
2283 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
2286 while (!list_empty(&ctx->iopoll_list) && !need_resched()) {
2289 ret = io_do_iopoll(ctx, nr_events, min);
2292 if (*nr_events >= min)
2300 * We can't just wait for polled events to come to us, we have to actively
2301 * find and complete them.
2303 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2305 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2308 mutex_lock(&ctx->uring_lock);
2309 while (!list_empty(&ctx->iopoll_list)) {
2310 unsigned int nr_events = 0;
2312 io_do_iopoll(ctx, &nr_events, 0);
2314 /* let it sleep and repeat later if can't complete a request */
2318 * Ensure we allow local-to-the-cpu processing to take place,
2319 * in this case we need to ensure that we reap all events.
2320 * Also let task_work, etc. to progress by releasing the mutex
2322 if (need_resched()) {
2323 mutex_unlock(&ctx->uring_lock);
2325 mutex_lock(&ctx->uring_lock);
2328 mutex_unlock(&ctx->uring_lock);
2331 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2333 unsigned int nr_events = 0;
2334 int iters = 0, ret = 0;
2337 * We disallow the app entering submit/complete with polling, but we
2338 * still need to lock the ring to prevent racing with polled issue
2339 * that got punted to a workqueue.
2341 mutex_lock(&ctx->uring_lock);
2344 * Don't enter poll loop if we already have events pending.
2345 * If we do, we can potentially be spinning for commands that
2346 * already triggered a CQE (eg in error).
2348 if (test_bit(0, &ctx->cq_check_overflow))
2349 __io_cqring_overflow_flush(ctx, false, NULL, NULL);
2350 if (io_cqring_events(ctx))
2354 * If a submit got punted to a workqueue, we can have the
2355 * application entering polling for a command before it gets
2356 * issued. That app will hold the uring_lock for the duration
2357 * of the poll right here, so we need to take a breather every
2358 * now and then to ensure that the issue has a chance to add
2359 * the poll to the issued list. Otherwise we can spin here
2360 * forever, while the workqueue is stuck trying to acquire the
2363 if (!(++iters & 7)) {
2364 mutex_unlock(&ctx->uring_lock);
2366 mutex_lock(&ctx->uring_lock);
2369 ret = io_iopoll_getevents(ctx, &nr_events, min);
2373 } while (min && !nr_events && !need_resched());
2375 mutex_unlock(&ctx->uring_lock);
2379 static void kiocb_end_write(struct io_kiocb *req)
2382 * Tell lockdep we inherited freeze protection from submission
2385 if (req->flags & REQ_F_ISREG) {
2386 struct inode *inode = file_inode(req->file);
2388 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
2390 file_end_write(req->file);
2394 static bool io_resubmit_prep(struct io_kiocb *req)
2396 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2398 struct iov_iter iter;
2400 /* already prepared */
2401 if (req->async_data)
2404 switch (req->opcode) {
2405 case IORING_OP_READV:
2406 case IORING_OP_READ_FIXED:
2407 case IORING_OP_READ:
2410 case IORING_OP_WRITEV:
2411 case IORING_OP_WRITE_FIXED:
2412 case IORING_OP_WRITE:
2416 printk_once(KERN_WARNING "io_uring: bad opcode in resubmit %d\n",
2421 ret = io_import_iovec(rw, req, &iovec, &iter, false);
2424 return !io_setup_async_rw(req, iovec, inline_vecs, &iter, false);
2427 static bool io_rw_should_reissue(struct io_kiocb *req)
2429 umode_t mode = file_inode(req->file)->i_mode;
2430 struct io_ring_ctx *ctx = req->ctx;
2432 if (!S_ISBLK(mode) && !S_ISREG(mode))
2434 if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
2435 !(ctx->flags & IORING_SETUP_IOPOLL)))
2438 * If ref is dying, we might be running poll reap from the exit work.
2439 * Don't attempt to reissue from that path, just let it fail with
2442 if (percpu_ref_is_dying(&ctx->refs))
2448 static bool io_rw_reissue(struct io_kiocb *req)
2451 if (!io_rw_should_reissue(req))
2454 lockdep_assert_held(&req->ctx->uring_lock);
2456 if (io_resubmit_prep(req)) {
2457 refcount_inc(&req->refs);
2458 io_queue_async_work(req);
2461 req_set_fail_links(req);
2466 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2467 unsigned int issue_flags)
2471 if ((res == -EAGAIN || res == -EOPNOTSUPP) && io_rw_reissue(req))
2473 if (res != req->result)
2474 req_set_fail_links(req);
2476 if (req->rw.kiocb.ki_flags & IOCB_WRITE)
2477 kiocb_end_write(req);
2478 if (req->flags & REQ_F_BUFFER_SELECTED)
2479 cflags = io_put_rw_kbuf(req);
2480 __io_req_complete(req, issue_flags, res, cflags);
2483 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2485 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2487 __io_complete_rw(req, res, res2, 0);
2490 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2492 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2495 /* Rewind iter, if we have one. iopoll path resubmits as usual */
2496 if (res == -EAGAIN && io_rw_should_reissue(req)) {
2497 struct io_async_rw *rw = req->async_data;
2500 iov_iter_revert(&rw->iter,
2501 req->result - iov_iter_count(&rw->iter));
2502 else if (!io_resubmit_prep(req))
2507 if (kiocb->ki_flags & IOCB_WRITE)
2508 kiocb_end_write(req);
2510 if (res != -EAGAIN && res != req->result)
2511 req_set_fail_links(req);
2513 WRITE_ONCE(req->result, res);
2514 /* order with io_poll_complete() checking ->result */
2516 WRITE_ONCE(req->iopoll_completed, 1);
2520 * After the iocb has been issued, it's safe to be found on the poll list.
2521 * Adding the kiocb to the list AFTER submission ensures that we don't
2522 * find it from a io_iopoll_getevents() thread before the issuer is done
2523 * accessing the kiocb cookie.
2525 static void io_iopoll_req_issued(struct io_kiocb *req, bool in_async)
2527 struct io_ring_ctx *ctx = req->ctx;
2530 * Track whether we have multiple files in our lists. This will impact
2531 * how we do polling eventually, not spinning if we're on potentially
2532 * different devices.
2534 if (list_empty(&ctx->iopoll_list)) {
2535 ctx->poll_multi_file = false;
2536 } else if (!ctx->poll_multi_file) {
2537 struct io_kiocb *list_req;
2539 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2541 if (list_req->file != req->file)
2542 ctx->poll_multi_file = true;
2546 * For fast devices, IO may have already completed. If it has, add
2547 * it to the front so we find it first.
2549 if (READ_ONCE(req->iopoll_completed))
2550 list_add(&req->inflight_entry, &ctx->iopoll_list);
2552 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2555 * If IORING_SETUP_SQPOLL is enabled, sqes are either handled in sq thread
2556 * task context or in io worker task context. If current task context is
2557 * sq thread, we don't need to check whether should wake up sq thread.
2559 if (in_async && (ctx->flags & IORING_SETUP_SQPOLL) &&
2560 wq_has_sleeper(&ctx->sq_data->wait))
2561 wake_up(&ctx->sq_data->wait);
2564 static inline void io_state_file_put(struct io_submit_state *state)
2566 if (state->file_refs) {
2567 fput_many(state->file, state->file_refs);
2568 state->file_refs = 0;
2573 * Get as many references to a file as we have IOs left in this submission,
2574 * assuming most submissions are for one file, or at least that each file
2575 * has more than one submission.
2577 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2582 if (state->file_refs) {
2583 if (state->fd == fd) {
2587 io_state_file_put(state);
2589 state->file = fget_many(fd, state->ios_left);
2590 if (unlikely(!state->file))
2594 state->file_refs = state->ios_left - 1;
2598 static bool io_bdev_nowait(struct block_device *bdev)
2600 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2604 * If we tracked the file through the SCM inflight mechanism, we could support
2605 * any file. For now, just ensure that anything potentially problematic is done
2608 static bool io_file_supports_async(struct file *file, int rw)
2610 umode_t mode = file_inode(file)->i_mode;
2612 if (S_ISBLK(mode)) {
2613 if (IS_ENABLED(CONFIG_BLOCK) &&
2614 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2618 if (S_ISCHR(mode) || S_ISSOCK(mode))
2620 if (S_ISREG(mode)) {
2621 if (IS_ENABLED(CONFIG_BLOCK) &&
2622 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2623 file->f_op != &io_uring_fops)
2628 /* any ->read/write should understand O_NONBLOCK */
2629 if (file->f_flags & O_NONBLOCK)
2632 if (!(file->f_mode & FMODE_NOWAIT))
2636 return file->f_op->read_iter != NULL;
2638 return file->f_op->write_iter != NULL;
2641 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2643 struct io_ring_ctx *ctx = req->ctx;
2644 struct kiocb *kiocb = &req->rw.kiocb;
2645 struct file *file = req->file;
2649 if (S_ISREG(file_inode(file)->i_mode))
2650 req->flags |= REQ_F_ISREG;
2652 kiocb->ki_pos = READ_ONCE(sqe->off);
2653 if (kiocb->ki_pos == -1 && !(file->f_mode & FMODE_STREAM)) {
2654 req->flags |= REQ_F_CUR_POS;
2655 kiocb->ki_pos = file->f_pos;
2657 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2658 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2659 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2663 /* don't allow async punt for O_NONBLOCK or RWF_NOWAIT */
2664 if ((kiocb->ki_flags & IOCB_NOWAIT) || (file->f_flags & O_NONBLOCK))
2665 req->flags |= REQ_F_NOWAIT;
2667 ioprio = READ_ONCE(sqe->ioprio);
2669 ret = ioprio_check_cap(ioprio);
2673 kiocb->ki_ioprio = ioprio;
2675 kiocb->ki_ioprio = get_current_ioprio();
2677 if (ctx->flags & IORING_SETUP_IOPOLL) {
2678 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2679 !kiocb->ki_filp->f_op->iopoll)
2682 kiocb->ki_flags |= IOCB_HIPRI;
2683 kiocb->ki_complete = io_complete_rw_iopoll;
2684 req->iopoll_completed = 0;
2686 if (kiocb->ki_flags & IOCB_HIPRI)
2688 kiocb->ki_complete = io_complete_rw;
2691 req->rw.addr = READ_ONCE(sqe->addr);
2692 req->rw.len = READ_ONCE(sqe->len);
2693 req->buf_index = READ_ONCE(sqe->buf_index);
2697 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2703 case -ERESTARTNOINTR:
2704 case -ERESTARTNOHAND:
2705 case -ERESTART_RESTARTBLOCK:
2707 * We can't just restart the syscall, since previously
2708 * submitted sqes may already be in progress. Just fail this
2714 kiocb->ki_complete(kiocb, ret, 0);
2718 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2719 unsigned int issue_flags)
2721 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2722 struct io_async_rw *io = req->async_data;
2724 /* add previously done IO, if any */
2725 if (io && io->bytes_done > 0) {
2727 ret = io->bytes_done;
2729 ret += io->bytes_done;
2732 if (req->flags & REQ_F_CUR_POS)
2733 req->file->f_pos = kiocb->ki_pos;
2734 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2735 __io_complete_rw(req, ret, 0, issue_flags);
2737 io_rw_done(kiocb, ret);
2740 static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
2742 struct io_ring_ctx *ctx = req->ctx;
2743 size_t len = req->rw.len;
2744 struct io_mapped_ubuf *imu;
2745 u16 index, buf_index = req->buf_index;
2749 if (unlikely(buf_index >= ctx->nr_user_bufs))
2751 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2752 imu = &ctx->user_bufs[index];
2753 buf_addr = req->rw.addr;
2756 if (buf_addr + len < buf_addr)
2758 /* not inside the mapped region */
2759 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2763 * May not be a start of buffer, set size appropriately
2764 * and advance us to the beginning.
2766 offset = buf_addr - imu->ubuf;
2767 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2771 * Don't use iov_iter_advance() here, as it's really slow for
2772 * using the latter parts of a big fixed buffer - it iterates
2773 * over each segment manually. We can cheat a bit here, because
2776 * 1) it's a BVEC iter, we set it up
2777 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2778 * first and last bvec
2780 * So just find our index, and adjust the iterator afterwards.
2781 * If the offset is within the first bvec (or the whole first
2782 * bvec, just use iov_iter_advance(). This makes it easier
2783 * since we can just skip the first segment, which may not
2784 * be PAGE_SIZE aligned.
2786 const struct bio_vec *bvec = imu->bvec;
2788 if (offset <= bvec->bv_len) {
2789 iov_iter_advance(iter, offset);
2791 unsigned long seg_skip;
2793 /* skip first vec */
2794 offset -= bvec->bv_len;
2795 seg_skip = 1 + (offset >> PAGE_SHIFT);
2797 iter->bvec = bvec + seg_skip;
2798 iter->nr_segs -= seg_skip;
2799 iter->count -= bvec->bv_len + offset;
2800 iter->iov_offset = offset & ~PAGE_MASK;
2807 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2810 mutex_unlock(&ctx->uring_lock);
2813 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2816 * "Normal" inline submissions always hold the uring_lock, since we
2817 * grab it from the system call. Same is true for the SQPOLL offload.
2818 * The only exception is when we've detached the request and issue it
2819 * from an async worker thread, grab the lock for that case.
2822 mutex_lock(&ctx->uring_lock);
2825 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2826 int bgid, struct io_buffer *kbuf,
2829 struct io_buffer *head;
2831 if (req->flags & REQ_F_BUFFER_SELECTED)
2834 io_ring_submit_lock(req->ctx, needs_lock);
2836 lockdep_assert_held(&req->ctx->uring_lock);
2838 head = idr_find(&req->ctx->io_buffer_idr, bgid);
2840 if (!list_empty(&head->list)) {
2841 kbuf = list_last_entry(&head->list, struct io_buffer,
2843 list_del(&kbuf->list);
2846 idr_remove(&req->ctx->io_buffer_idr, bgid);
2848 if (*len > kbuf->len)
2851 kbuf = ERR_PTR(-ENOBUFS);
2854 io_ring_submit_unlock(req->ctx, needs_lock);
2859 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2862 struct io_buffer *kbuf;
2865 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2866 bgid = req->buf_index;
2867 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2870 req->rw.addr = (u64) (unsigned long) kbuf;
2871 req->flags |= REQ_F_BUFFER_SELECTED;
2872 return u64_to_user_ptr(kbuf->addr);
2875 #ifdef CONFIG_COMPAT
2876 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2879 struct compat_iovec __user *uiov;
2880 compat_ssize_t clen;
2884 uiov = u64_to_user_ptr(req->rw.addr);
2885 if (!access_ok(uiov, sizeof(*uiov)))
2887 if (__get_user(clen, &uiov->iov_len))
2893 buf = io_rw_buffer_select(req, &len, needs_lock);
2895 return PTR_ERR(buf);
2896 iov[0].iov_base = buf;
2897 iov[0].iov_len = (compat_size_t) len;
2902 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2905 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2909 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2912 len = iov[0].iov_len;
2915 buf = io_rw_buffer_select(req, &len, needs_lock);
2917 return PTR_ERR(buf);
2918 iov[0].iov_base = buf;
2919 iov[0].iov_len = len;
2923 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2926 if (req->flags & REQ_F_BUFFER_SELECTED) {
2927 struct io_buffer *kbuf;
2929 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2930 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
2931 iov[0].iov_len = kbuf->len;
2934 if (req->rw.len != 1)
2937 #ifdef CONFIG_COMPAT
2938 if (req->ctx->compat)
2939 return io_compat_import(req, iov, needs_lock);
2942 return __io_iov_buffer_select(req, iov, needs_lock);
2945 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
2946 struct iov_iter *iter, bool needs_lock)
2948 void __user *buf = u64_to_user_ptr(req->rw.addr);
2949 size_t sqe_len = req->rw.len;
2950 u8 opcode = req->opcode;
2953 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
2955 return io_import_fixed(req, rw, iter);
2958 /* buffer index only valid with fixed read/write, or buffer select */
2959 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
2962 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
2963 if (req->flags & REQ_F_BUFFER_SELECT) {
2964 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
2966 return PTR_ERR(buf);
2967 req->rw.len = sqe_len;
2970 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
2975 if (req->flags & REQ_F_BUFFER_SELECT) {
2976 ret = io_iov_buffer_select(req, *iovec, needs_lock);
2978 iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
2983 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
2987 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
2989 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
2993 * For files that don't have ->read_iter() and ->write_iter(), handle them
2994 * by looping over ->read() or ->write() manually.
2996 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
2998 struct kiocb *kiocb = &req->rw.kiocb;
2999 struct file *file = req->file;
3003 * Don't support polled IO through this interface, and we can't
3004 * support non-blocking either. For the latter, this just causes
3005 * the kiocb to be handled from an async context.
3007 if (kiocb->ki_flags & IOCB_HIPRI)
3009 if (kiocb->ki_flags & IOCB_NOWAIT)
3012 while (iov_iter_count(iter)) {
3016 if (!iov_iter_is_bvec(iter)) {
3017 iovec = iov_iter_iovec(iter);
3019 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3020 iovec.iov_len = req->rw.len;
3024 nr = file->f_op->read(file, iovec.iov_base,
3025 iovec.iov_len, io_kiocb_ppos(kiocb));
3027 nr = file->f_op->write(file, iovec.iov_base,
3028 iovec.iov_len, io_kiocb_ppos(kiocb));
3037 if (nr != iovec.iov_len)
3041 iov_iter_advance(iter, nr);
3047 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3048 const struct iovec *fast_iov, struct iov_iter *iter)
3050 struct io_async_rw *rw = req->async_data;
3052 memcpy(&rw->iter, iter, sizeof(*iter));
3053 rw->free_iovec = iovec;
3055 /* can only be fixed buffers, no need to do anything */
3056 if (iov_iter_is_bvec(iter))
3059 unsigned iov_off = 0;
3061 rw->iter.iov = rw->fast_iov;
3062 if (iter->iov != fast_iov) {
3063 iov_off = iter->iov - fast_iov;
3064 rw->iter.iov += iov_off;
3066 if (rw->fast_iov != fast_iov)
3067 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3068 sizeof(struct iovec) * iter->nr_segs);
3070 req->flags |= REQ_F_NEED_CLEANUP;
3074 static inline int __io_alloc_async_data(struct io_kiocb *req)
3076 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3077 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3078 return req->async_data == NULL;
3081 static int io_alloc_async_data(struct io_kiocb *req)
3083 if (!io_op_defs[req->opcode].needs_async_data)
3086 return __io_alloc_async_data(req);
3089 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3090 const struct iovec *fast_iov,
3091 struct iov_iter *iter, bool force)
3093 if (!force && !io_op_defs[req->opcode].needs_async_data)
3095 if (!req->async_data) {
3096 if (__io_alloc_async_data(req)) {
3101 io_req_map_rw(req, iovec, fast_iov, iter);
3106 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3108 struct io_async_rw *iorw = req->async_data;
3109 struct iovec *iov = iorw->fast_iov;
3112 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3113 if (unlikely(ret < 0))
3116 iorw->bytes_done = 0;
3117 iorw->free_iovec = iov;
3119 req->flags |= REQ_F_NEED_CLEANUP;
3123 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3125 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3127 return io_prep_rw(req, sqe);
3131 * This is our waitqueue callback handler, registered through lock_page_async()
3132 * when we initially tried to do the IO with the iocb armed our waitqueue.
3133 * This gets called when the page is unlocked, and we generally expect that to
3134 * happen when the page IO is completed and the page is now uptodate. This will
3135 * queue a task_work based retry of the operation, attempting to copy the data
3136 * again. If the latter fails because the page was NOT uptodate, then we will
3137 * do a thread based blocking retry of the operation. That's the unexpected
3140 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3141 int sync, void *arg)
3143 struct wait_page_queue *wpq;
3144 struct io_kiocb *req = wait->private;
3145 struct wait_page_key *key = arg;
3147 wpq = container_of(wait, struct wait_page_queue, wait);
3149 if (!wake_page_match(wpq, key))
3152 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3153 list_del_init(&wait->entry);
3155 /* submit ref gets dropped, acquire a new one */
3156 refcount_inc(&req->refs);
3157 io_req_task_queue(req);
3162 * This controls whether a given IO request should be armed for async page
3163 * based retry. If we return false here, the request is handed to the async
3164 * worker threads for retry. If we're doing buffered reads on a regular file,
3165 * we prepare a private wait_page_queue entry and retry the operation. This
3166 * will either succeed because the page is now uptodate and unlocked, or it
3167 * will register a callback when the page is unlocked at IO completion. Through
3168 * that callback, io_uring uses task_work to setup a retry of the operation.
3169 * That retry will attempt the buffered read again. The retry will generally
3170 * succeed, or in rare cases where it fails, we then fall back to using the
3171 * async worker threads for a blocking retry.
3173 static bool io_rw_should_retry(struct io_kiocb *req)
3175 struct io_async_rw *rw = req->async_data;
3176 struct wait_page_queue *wait = &rw->wpq;
3177 struct kiocb *kiocb = &req->rw.kiocb;
3179 /* never retry for NOWAIT, we just complete with -EAGAIN */
3180 if (req->flags & REQ_F_NOWAIT)
3183 /* Only for buffered IO */
3184 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3188 * just use poll if we can, and don't attempt if the fs doesn't
3189 * support callback based unlocks
3191 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3194 wait->wait.func = io_async_buf_func;
3195 wait->wait.private = req;
3196 wait->wait.flags = 0;
3197 INIT_LIST_HEAD(&wait->wait.entry);
3198 kiocb->ki_flags |= IOCB_WAITQ;
3199 kiocb->ki_flags &= ~IOCB_NOWAIT;
3200 kiocb->ki_waitq = wait;
3204 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3206 if (req->file->f_op->read_iter)
3207 return call_read_iter(req->file, &req->rw.kiocb, iter);
3208 else if (req->file->f_op->read)
3209 return loop_rw_iter(READ, req, iter);
3214 static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3216 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3217 struct kiocb *kiocb = &req->rw.kiocb;
3218 struct iov_iter __iter, *iter = &__iter;
3219 struct io_async_rw *rw = req->async_data;
3220 ssize_t io_size, ret, ret2;
3221 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3227 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3231 io_size = iov_iter_count(iter);
3232 req->result = io_size;
3234 /* Ensure we clear previously set non-block flag */
3235 if (!force_nonblock)
3236 kiocb->ki_flags &= ~IOCB_NOWAIT;
3238 kiocb->ki_flags |= IOCB_NOWAIT;
3240 /* If the file doesn't support async, just async punt */
3241 if (force_nonblock && !io_file_supports_async(req->file, READ)) {
3242 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3243 return ret ?: -EAGAIN;
3246 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size);
3247 if (unlikely(ret)) {
3252 ret = io_iter_do_read(req, iter);
3254 if (ret == -EIOCBQUEUED) {
3255 if (req->async_data)
3256 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3258 } else if (ret == -EAGAIN) {
3259 /* IOPOLL retry should happen for io-wq threads */
3260 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3262 /* no retry on NONBLOCK nor RWF_NOWAIT */
3263 if (req->flags & REQ_F_NOWAIT)
3265 /* some cases will consume bytes even on error returns */
3266 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3268 } else if (ret <= 0 || ret == io_size || !force_nonblock ||
3269 (req->flags & REQ_F_NOWAIT) || !(req->flags & REQ_F_ISREG)) {
3270 /* read all, failed, already did sync or don't want to retry */
3274 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3279 rw = req->async_data;
3280 /* now use our persistent iterator, if we aren't already */
3285 rw->bytes_done += ret;
3286 /* if we can retry, do so with the callbacks armed */
3287 if (!io_rw_should_retry(req)) {
3288 kiocb->ki_flags &= ~IOCB_WAITQ;
3293 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3294 * we get -EIOCBQUEUED, then we'll get a notification when the
3295 * desired page gets unlocked. We can also get a partial read
3296 * here, and if we do, then just retry at the new offset.
3298 ret = io_iter_do_read(req, iter);
3299 if (ret == -EIOCBQUEUED)
3301 /* we got some bytes, but not all. retry. */
3302 } while (ret > 0 && ret < io_size);
3304 kiocb_done(kiocb, ret, issue_flags);
3306 /* it's faster to check here then delegate to kfree */
3312 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3314 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3316 return io_prep_rw(req, sqe);
3319 static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3321 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3322 struct kiocb *kiocb = &req->rw.kiocb;
3323 struct iov_iter __iter, *iter = &__iter;
3324 struct io_async_rw *rw = req->async_data;
3325 ssize_t ret, ret2, io_size;
3326 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3332 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3336 io_size = iov_iter_count(iter);
3337 req->result = io_size;
3339 /* Ensure we clear previously set non-block flag */
3340 if (!force_nonblock)
3341 kiocb->ki_flags &= ~IOCB_NOWAIT;
3343 kiocb->ki_flags |= IOCB_NOWAIT;
3345 /* If the file doesn't support async, just async punt */
3346 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
3349 /* file path doesn't support NOWAIT for non-direct_IO */
3350 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3351 (req->flags & REQ_F_ISREG))
3354 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size);
3359 * Open-code file_start_write here to grab freeze protection,
3360 * which will be released by another thread in
3361 * io_complete_rw(). Fool lockdep by telling it the lock got
3362 * released so that it doesn't complain about the held lock when
3363 * we return to userspace.
3365 if (req->flags & REQ_F_ISREG) {
3366 sb_start_write(file_inode(req->file)->i_sb);
3367 __sb_writers_release(file_inode(req->file)->i_sb,
3370 kiocb->ki_flags |= IOCB_WRITE;
3372 if (req->file->f_op->write_iter)
3373 ret2 = call_write_iter(req->file, kiocb, iter);
3374 else if (req->file->f_op->write)
3375 ret2 = loop_rw_iter(WRITE, req, iter);
3380 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3381 * retry them without IOCB_NOWAIT.
3383 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3385 /* no retry on NONBLOCK nor RWF_NOWAIT */
3386 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3388 if (ret2 == -EIOCBQUEUED && req->async_data)
3389 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3390 if (!force_nonblock || ret2 != -EAGAIN) {
3391 /* IOPOLL retry should happen for io-wq threads */
3392 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3395 kiocb_done(kiocb, ret2, issue_flags);
3398 /* some cases will consume bytes even on error returns */
3399 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3400 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3401 return ret ?: -EAGAIN;
3404 /* it's reportedly faster than delegating the null check to kfree() */
3410 static int io_renameat_prep(struct io_kiocb *req,
3411 const struct io_uring_sqe *sqe)
3413 struct io_rename *ren = &req->rename;
3414 const char __user *oldf, *newf;
3416 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3419 ren->old_dfd = READ_ONCE(sqe->fd);
3420 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3421 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3422 ren->new_dfd = READ_ONCE(sqe->len);
3423 ren->flags = READ_ONCE(sqe->rename_flags);
3425 ren->oldpath = getname(oldf);
3426 if (IS_ERR(ren->oldpath))
3427 return PTR_ERR(ren->oldpath);
3429 ren->newpath = getname(newf);
3430 if (IS_ERR(ren->newpath)) {
3431 putname(ren->oldpath);
3432 return PTR_ERR(ren->newpath);
3435 req->flags |= REQ_F_NEED_CLEANUP;
3439 static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3441 struct io_rename *ren = &req->rename;
3444 if (issue_flags & IO_URING_F_NONBLOCK)
3447 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3448 ren->newpath, ren->flags);
3450 req->flags &= ~REQ_F_NEED_CLEANUP;
3452 req_set_fail_links(req);
3453 io_req_complete(req, ret);
3457 static int io_unlinkat_prep(struct io_kiocb *req,
3458 const struct io_uring_sqe *sqe)
3460 struct io_unlink *un = &req->unlink;
3461 const char __user *fname;
3463 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3466 un->dfd = READ_ONCE(sqe->fd);
3468 un->flags = READ_ONCE(sqe->unlink_flags);
3469 if (un->flags & ~AT_REMOVEDIR)
3472 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3473 un->filename = getname(fname);
3474 if (IS_ERR(un->filename))
3475 return PTR_ERR(un->filename);
3477 req->flags |= REQ_F_NEED_CLEANUP;
3481 static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3483 struct io_unlink *un = &req->unlink;
3486 if (issue_flags & IO_URING_F_NONBLOCK)
3489 if (un->flags & AT_REMOVEDIR)
3490 ret = do_rmdir(un->dfd, un->filename);
3492 ret = do_unlinkat(un->dfd, un->filename);
3494 req->flags &= ~REQ_F_NEED_CLEANUP;
3496 req_set_fail_links(req);
3497 io_req_complete(req, ret);
3501 static int io_shutdown_prep(struct io_kiocb *req,
3502 const struct io_uring_sqe *sqe)
3504 #if defined(CONFIG_NET)
3505 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3507 if (sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3511 req->shutdown.how = READ_ONCE(sqe->len);
3518 static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3520 #if defined(CONFIG_NET)
3521 struct socket *sock;
3524 if (issue_flags & IO_URING_F_NONBLOCK)
3527 sock = sock_from_file(req->file);
3528 if (unlikely(!sock))
3531 ret = __sys_shutdown_sock(sock, req->shutdown.how);
3533 req_set_fail_links(req);
3534 io_req_complete(req, ret);
3541 static int __io_splice_prep(struct io_kiocb *req,
3542 const struct io_uring_sqe *sqe)
3544 struct io_splice* sp = &req->splice;
3545 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3547 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3551 sp->len = READ_ONCE(sqe->len);
3552 sp->flags = READ_ONCE(sqe->splice_flags);
3554 if (unlikely(sp->flags & ~valid_flags))
3557 sp->file_in = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in),
3558 (sp->flags & SPLICE_F_FD_IN_FIXED));
3561 req->flags |= REQ_F_NEED_CLEANUP;
3563 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
3565 * Splice operation will be punted aync, and here need to
3566 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3568 req->work.flags |= IO_WQ_WORK_UNBOUND;
3574 static int io_tee_prep(struct io_kiocb *req,
3575 const struct io_uring_sqe *sqe)
3577 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3579 return __io_splice_prep(req, sqe);
3582 static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
3584 struct io_splice *sp = &req->splice;
3585 struct file *in = sp->file_in;
3586 struct file *out = sp->file_out;
3587 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3590 if (issue_flags & IO_URING_F_NONBLOCK)
3593 ret = do_tee(in, out, sp->len, flags);
3595 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3596 req->flags &= ~REQ_F_NEED_CLEANUP;
3599 req_set_fail_links(req);
3600 io_req_complete(req, ret);
3604 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3606 struct io_splice* sp = &req->splice;
3608 sp->off_in = READ_ONCE(sqe->splice_off_in);
3609 sp->off_out = READ_ONCE(sqe->off);
3610 return __io_splice_prep(req, sqe);
3613 static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
3615 struct io_splice *sp = &req->splice;
3616 struct file *in = sp->file_in;
3617 struct file *out = sp->file_out;
3618 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3619 loff_t *poff_in, *poff_out;
3622 if (issue_flags & IO_URING_F_NONBLOCK)
3625 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3626 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3629 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3631 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3632 req->flags &= ~REQ_F_NEED_CLEANUP;
3635 req_set_fail_links(req);
3636 io_req_complete(req, ret);
3641 * IORING_OP_NOP just posts a completion event, nothing else.
3643 static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
3645 struct io_ring_ctx *ctx = req->ctx;
3647 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3650 __io_req_complete(req, issue_flags, 0, 0);
3654 static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3656 struct io_ring_ctx *ctx = req->ctx;
3661 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3663 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3666 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3667 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3670 req->sync.off = READ_ONCE(sqe->off);
3671 req->sync.len = READ_ONCE(sqe->len);
3675 static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
3677 loff_t end = req->sync.off + req->sync.len;
3680 /* fsync always requires a blocking context */
3681 if (issue_flags & IO_URING_F_NONBLOCK)
3684 ret = vfs_fsync_range(req->file, req->sync.off,
3685 end > 0 ? end : LLONG_MAX,
3686 req->sync.flags & IORING_FSYNC_DATASYNC);
3688 req_set_fail_links(req);
3689 io_req_complete(req, ret);
3693 static int io_fallocate_prep(struct io_kiocb *req,
3694 const struct io_uring_sqe *sqe)
3696 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3698 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3701 req->sync.off = READ_ONCE(sqe->off);
3702 req->sync.len = READ_ONCE(sqe->addr);
3703 req->sync.mode = READ_ONCE(sqe->len);
3707 static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
3711 /* fallocate always requiring blocking context */
3712 if (issue_flags & IO_URING_F_NONBLOCK)
3714 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3717 req_set_fail_links(req);
3718 io_req_complete(req, ret);
3722 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3724 const char __user *fname;
3727 if (unlikely(sqe->ioprio || sqe->buf_index))
3729 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3732 /* open.how should be already initialised */
3733 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3734 req->open.how.flags |= O_LARGEFILE;
3736 req->open.dfd = READ_ONCE(sqe->fd);
3737 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3738 req->open.filename = getname(fname);
3739 if (IS_ERR(req->open.filename)) {
3740 ret = PTR_ERR(req->open.filename);
3741 req->open.filename = NULL;
3744 req->open.nofile = rlimit(RLIMIT_NOFILE);
3745 req->flags |= REQ_F_NEED_CLEANUP;
3749 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3753 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3755 mode = READ_ONCE(sqe->len);
3756 flags = READ_ONCE(sqe->open_flags);
3757 req->open.how = build_open_how(flags, mode);
3758 return __io_openat_prep(req, sqe);
3761 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3763 struct open_how __user *how;
3767 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3769 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3770 len = READ_ONCE(sqe->len);
3771 if (len < OPEN_HOW_SIZE_VER0)
3774 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3779 return __io_openat_prep(req, sqe);
3782 static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
3784 struct open_flags op;
3787 bool resolve_nonblock;
3790 ret = build_open_flags(&req->open.how, &op);
3793 nonblock_set = op.open_flag & O_NONBLOCK;
3794 resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
3795 if (issue_flags & IO_URING_F_NONBLOCK) {
3797 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
3798 * it'll always -EAGAIN
3800 if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
3802 op.lookup_flags |= LOOKUP_CACHED;
3803 op.open_flag |= O_NONBLOCK;
3806 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3810 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3811 /* only retry if RESOLVE_CACHED wasn't already set by application */
3812 if ((!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)) &&
3813 file == ERR_PTR(-EAGAIN)) {
3815 * We could hang on to this 'fd', but seems like marginal
3816 * gain for something that is now known to be a slower path.
3817 * So just put it, and we'll get a new one when we retry.
3825 ret = PTR_ERR(file);
3827 if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
3828 file->f_flags &= ~O_NONBLOCK;
3829 fsnotify_open(file);
3830 fd_install(ret, file);
3833 putname(req->open.filename);
3834 req->flags &= ~REQ_F_NEED_CLEANUP;
3836 req_set_fail_links(req);
3837 io_req_complete(req, ret);
3841 static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
3843 return io_openat2(req, issue_flags & IO_URING_F_NONBLOCK);
3846 static int io_remove_buffers_prep(struct io_kiocb *req,
3847 const struct io_uring_sqe *sqe)
3849 struct io_provide_buf *p = &req->pbuf;
3852 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3855 tmp = READ_ONCE(sqe->fd);
3856 if (!tmp || tmp > USHRT_MAX)
3859 memset(p, 0, sizeof(*p));
3861 p->bgid = READ_ONCE(sqe->buf_group);
3865 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3866 int bgid, unsigned nbufs)
3870 /* shouldn't happen */
3874 /* the head kbuf is the list itself */
3875 while (!list_empty(&buf->list)) {
3876 struct io_buffer *nxt;
3878 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3879 list_del(&nxt->list);
3886 idr_remove(&ctx->io_buffer_idr, bgid);
3891 static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
3893 struct io_provide_buf *p = &req->pbuf;
3894 struct io_ring_ctx *ctx = req->ctx;
3895 struct io_buffer *head;
3897 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3899 io_ring_submit_lock(ctx, !force_nonblock);
3901 lockdep_assert_held(&ctx->uring_lock);
3904 head = idr_find(&ctx->io_buffer_idr, p->bgid);
3906 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3908 req_set_fail_links(req);
3910 /* need to hold the lock to complete IOPOLL requests */
3911 if (ctx->flags & IORING_SETUP_IOPOLL) {
3912 __io_req_complete(req, issue_flags, ret, 0);
3913 io_ring_submit_unlock(ctx, !force_nonblock);
3915 io_ring_submit_unlock(ctx, !force_nonblock);
3916 __io_req_complete(req, issue_flags, ret, 0);
3921 static int io_provide_buffers_prep(struct io_kiocb *req,
3922 const struct io_uring_sqe *sqe)
3924 struct io_provide_buf *p = &req->pbuf;
3927 if (sqe->ioprio || sqe->rw_flags)
3930 tmp = READ_ONCE(sqe->fd);
3931 if (!tmp || tmp > USHRT_MAX)
3934 p->addr = READ_ONCE(sqe->addr);
3935 p->len = READ_ONCE(sqe->len);
3937 if (!access_ok(u64_to_user_ptr(p->addr), (p->len * p->nbufs)))
3940 p->bgid = READ_ONCE(sqe->buf_group);
3941 tmp = READ_ONCE(sqe->off);
3942 if (tmp > USHRT_MAX)
3948 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3950 struct io_buffer *buf;
3951 u64 addr = pbuf->addr;
3952 int i, bid = pbuf->bid;
3954 for (i = 0; i < pbuf->nbufs; i++) {
3955 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
3960 buf->len = pbuf->len;
3965 INIT_LIST_HEAD(&buf->list);
3968 list_add_tail(&buf->list, &(*head)->list);
3972 return i ? i : -ENOMEM;
3975 static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
3977 struct io_provide_buf *p = &req->pbuf;
3978 struct io_ring_ctx *ctx = req->ctx;
3979 struct io_buffer *head, *list;
3981 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3983 io_ring_submit_lock(ctx, !force_nonblock);
3985 lockdep_assert_held(&ctx->uring_lock);
3987 list = head = idr_find(&ctx->io_buffer_idr, p->bgid);
3989 ret = io_add_buffers(p, &head);
3994 ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1,
3997 __io_remove_buffers(ctx, head, p->bgid, -1U);
4003 req_set_fail_links(req);
4005 /* need to hold the lock to complete IOPOLL requests */
4006 if (ctx->flags & IORING_SETUP_IOPOLL) {
4007 __io_req_complete(req, issue_flags, ret, 0);
4008 io_ring_submit_unlock(ctx, !force_nonblock);
4010 io_ring_submit_unlock(ctx, !force_nonblock);
4011 __io_req_complete(req, issue_flags, ret, 0);
4016 static int io_epoll_ctl_prep(struct io_kiocb *req,
4017 const struct io_uring_sqe *sqe)
4019 #if defined(CONFIG_EPOLL)
4020 if (sqe->ioprio || sqe->buf_index)
4022 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4025 req->epoll.epfd = READ_ONCE(sqe->fd);
4026 req->epoll.op = READ_ONCE(sqe->len);
4027 req->epoll.fd = READ_ONCE(sqe->off);
4029 if (ep_op_has_event(req->epoll.op)) {
4030 struct epoll_event __user *ev;
4032 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4033 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4043 static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4045 #if defined(CONFIG_EPOLL)
4046 struct io_epoll *ie = &req->epoll;
4048 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4050 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4051 if (force_nonblock && ret == -EAGAIN)
4055 req_set_fail_links(req);
4056 __io_req_complete(req, issue_flags, ret, 0);
4063 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4065 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4066 if (sqe->ioprio || sqe->buf_index || sqe->off)
4068 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4071 req->madvise.addr = READ_ONCE(sqe->addr);
4072 req->madvise.len = READ_ONCE(sqe->len);
4073 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4080 static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4082 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4083 struct io_madvise *ma = &req->madvise;
4086 if (issue_flags & IO_URING_F_NONBLOCK)
4089 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4091 req_set_fail_links(req);
4092 io_req_complete(req, ret);
4099 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4101 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4103 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4106 req->fadvise.offset = READ_ONCE(sqe->off);
4107 req->fadvise.len = READ_ONCE(sqe->len);
4108 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4112 static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4114 struct io_fadvise *fa = &req->fadvise;
4117 if (issue_flags & IO_URING_F_NONBLOCK) {
4118 switch (fa->advice) {
4119 case POSIX_FADV_NORMAL:
4120 case POSIX_FADV_RANDOM:
4121 case POSIX_FADV_SEQUENTIAL:
4128 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4130 req_set_fail_links(req);
4131 io_req_complete(req, ret);
4135 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4137 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4139 if (sqe->ioprio || sqe->buf_index)
4141 if (req->flags & REQ_F_FIXED_FILE)
4144 req->statx.dfd = READ_ONCE(sqe->fd);
4145 req->statx.mask = READ_ONCE(sqe->len);
4146 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4147 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4148 req->statx.flags = READ_ONCE(sqe->statx_flags);
4153 static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4155 struct io_statx *ctx = &req->statx;
4158 if (issue_flags & IO_URING_F_NONBLOCK) {
4159 /* only need file table for an actual valid fd */
4160 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
4161 req->flags |= REQ_F_NO_FILE_TABLE;
4165 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4169 req_set_fail_links(req);
4170 io_req_complete(req, ret);
4174 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4176 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4178 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4179 sqe->rw_flags || sqe->buf_index)
4181 if (req->flags & REQ_F_FIXED_FILE)
4184 req->close.fd = READ_ONCE(sqe->fd);
4188 static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4190 struct files_struct *files = current->files;
4191 struct io_close *close = &req->close;
4192 struct fdtable *fdt;
4198 spin_lock(&files->file_lock);
4199 fdt = files_fdtable(files);
4200 if (close->fd >= fdt->max_fds) {
4201 spin_unlock(&files->file_lock);
4204 file = fdt->fd[close->fd];
4206 spin_unlock(&files->file_lock);
4210 if (file->f_op == &io_uring_fops) {
4211 spin_unlock(&files->file_lock);
4216 /* if the file has a flush method, be safe and punt to async */
4217 if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4218 spin_unlock(&files->file_lock);
4222 ret = __close_fd_get_file(close->fd, &file);
4223 spin_unlock(&files->file_lock);
4230 /* No ->flush() or already async, safely close from here */
4231 ret = filp_close(file, current->files);
4234 req_set_fail_links(req);
4237 __io_req_complete(req, issue_flags, ret, 0);
4241 static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4243 struct io_ring_ctx *ctx = req->ctx;
4245 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4247 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4250 req->sync.off = READ_ONCE(sqe->off);
4251 req->sync.len = READ_ONCE(sqe->len);
4252 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4256 static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4260 /* sync_file_range always requires a blocking context */
4261 if (issue_flags & IO_URING_F_NONBLOCK)
4264 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4267 req_set_fail_links(req);
4268 io_req_complete(req, ret);
4272 #if defined(CONFIG_NET)
4273 static int io_setup_async_msg(struct io_kiocb *req,
4274 struct io_async_msghdr *kmsg)
4276 struct io_async_msghdr *async_msg = req->async_data;
4280 if (io_alloc_async_data(req)) {
4281 kfree(kmsg->free_iov);
4284 async_msg = req->async_data;
4285 req->flags |= REQ_F_NEED_CLEANUP;
4286 memcpy(async_msg, kmsg, sizeof(*kmsg));
4287 async_msg->msg.msg_name = &async_msg->addr;
4288 /* if were using fast_iov, set it to the new one */
4289 if (!async_msg->free_iov)
4290 async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4295 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4296 struct io_async_msghdr *iomsg)
4298 iomsg->msg.msg_name = &iomsg->addr;
4299 iomsg->free_iov = iomsg->fast_iov;
4300 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4301 req->sr_msg.msg_flags, &iomsg->free_iov);
4304 static int io_sendmsg_prep_async(struct io_kiocb *req)
4308 if (!io_op_defs[req->opcode].needs_async_data)
4310 ret = io_sendmsg_copy_hdr(req, req->async_data);
4312 req->flags |= REQ_F_NEED_CLEANUP;
4316 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4318 struct io_sr_msg *sr = &req->sr_msg;
4320 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4323 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4324 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4325 sr->len = READ_ONCE(sqe->len);
4327 #ifdef CONFIG_COMPAT
4328 if (req->ctx->compat)
4329 sr->msg_flags |= MSG_CMSG_COMPAT;
4334 static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4336 struct io_async_msghdr iomsg, *kmsg;
4337 struct socket *sock;
4341 sock = sock_from_file(req->file);
4342 if (unlikely(!sock))
4345 kmsg = req->async_data;
4347 ret = io_sendmsg_copy_hdr(req, &iomsg);
4353 flags = req->sr_msg.msg_flags;
4354 if (flags & MSG_DONTWAIT)
4355 req->flags |= REQ_F_NOWAIT;
4356 else if (issue_flags & IO_URING_F_NONBLOCK)
4357 flags |= MSG_DONTWAIT;
4359 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4360 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4361 return io_setup_async_msg(req, kmsg);
4362 if (ret == -ERESTARTSYS)
4365 /* fast path, check for non-NULL to avoid function call */
4367 kfree(kmsg->free_iov);
4368 req->flags &= ~REQ_F_NEED_CLEANUP;
4370 req_set_fail_links(req);
4371 __io_req_complete(req, issue_flags, ret, 0);
4375 static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4377 struct io_sr_msg *sr = &req->sr_msg;
4380 struct socket *sock;
4384 sock = sock_from_file(req->file);
4385 if (unlikely(!sock))
4388 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4392 msg.msg_name = NULL;
4393 msg.msg_control = NULL;
4394 msg.msg_controllen = 0;
4395 msg.msg_namelen = 0;
4397 flags = req->sr_msg.msg_flags;
4398 if (flags & MSG_DONTWAIT)
4399 req->flags |= REQ_F_NOWAIT;
4400 else if (issue_flags & IO_URING_F_NONBLOCK)
4401 flags |= MSG_DONTWAIT;
4403 msg.msg_flags = flags;
4404 ret = sock_sendmsg(sock, &msg);
4405 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4407 if (ret == -ERESTARTSYS)
4411 req_set_fail_links(req);
4412 __io_req_complete(req, issue_flags, ret, 0);
4416 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4417 struct io_async_msghdr *iomsg)
4419 struct io_sr_msg *sr = &req->sr_msg;
4420 struct iovec __user *uiov;
4424 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4425 &iomsg->uaddr, &uiov, &iov_len);
4429 if (req->flags & REQ_F_BUFFER_SELECT) {
4432 if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4434 sr->len = iomsg->fast_iov[0].iov_len;
4435 iomsg->free_iov = NULL;
4437 iomsg->free_iov = iomsg->fast_iov;
4438 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4439 &iomsg->free_iov, &iomsg->msg.msg_iter,
4448 #ifdef CONFIG_COMPAT
4449 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4450 struct io_async_msghdr *iomsg)
4452 struct compat_msghdr __user *msg_compat;
4453 struct io_sr_msg *sr = &req->sr_msg;
4454 struct compat_iovec __user *uiov;
4459 msg_compat = (struct compat_msghdr __user *) sr->umsg;
4460 ret = __get_compat_msghdr(&iomsg->msg, msg_compat, &iomsg->uaddr,
4465 uiov = compat_ptr(ptr);
4466 if (req->flags & REQ_F_BUFFER_SELECT) {
4467 compat_ssize_t clen;
4471 if (!access_ok(uiov, sizeof(*uiov)))
4473 if (__get_user(clen, &uiov->iov_len))
4478 iomsg->free_iov = NULL;
4480 iomsg->free_iov = iomsg->fast_iov;
4481 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4482 UIO_FASTIOV, &iomsg->free_iov,
4483 &iomsg->msg.msg_iter, true);
4492 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4493 struct io_async_msghdr *iomsg)
4495 iomsg->msg.msg_name = &iomsg->addr;
4497 #ifdef CONFIG_COMPAT
4498 if (req->ctx->compat)
4499 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4502 return __io_recvmsg_copy_hdr(req, iomsg);
4505 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4508 struct io_sr_msg *sr = &req->sr_msg;
4509 struct io_buffer *kbuf;
4511 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4516 req->flags |= REQ_F_BUFFER_SELECTED;
4520 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4522 return io_put_kbuf(req, req->sr_msg.kbuf);
4525 static int io_recvmsg_prep_async(struct io_kiocb *req)
4529 if (!io_op_defs[req->opcode].needs_async_data)
4531 ret = io_recvmsg_copy_hdr(req, req->async_data);
4533 req->flags |= REQ_F_NEED_CLEANUP;
4537 static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4539 struct io_sr_msg *sr = &req->sr_msg;
4541 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4544 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4545 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4546 sr->len = READ_ONCE(sqe->len);
4547 sr->bgid = READ_ONCE(sqe->buf_group);
4549 #ifdef CONFIG_COMPAT
4550 if (req->ctx->compat)
4551 sr->msg_flags |= MSG_CMSG_COMPAT;
4556 static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
4558 struct io_async_msghdr iomsg, *kmsg;
4559 struct socket *sock;
4560 struct io_buffer *kbuf;
4562 int ret, cflags = 0;
4563 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4565 sock = sock_from_file(req->file);
4566 if (unlikely(!sock))
4569 kmsg = req->async_data;
4571 ret = io_recvmsg_copy_hdr(req, &iomsg);
4577 if (req->flags & REQ_F_BUFFER_SELECT) {
4578 kbuf = io_recv_buffer_select(req, !force_nonblock);
4580 return PTR_ERR(kbuf);
4581 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4582 kmsg->fast_iov[0].iov_len = req->sr_msg.len;
4583 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
4584 1, req->sr_msg.len);
4587 flags = req->sr_msg.msg_flags;
4588 if (flags & MSG_DONTWAIT)
4589 req->flags |= REQ_F_NOWAIT;
4590 else if (force_nonblock)
4591 flags |= MSG_DONTWAIT;
4593 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4594 kmsg->uaddr, flags);
4595 if (force_nonblock && ret == -EAGAIN)
4596 return io_setup_async_msg(req, kmsg);
4597 if (ret == -ERESTARTSYS)
4600 if (req->flags & REQ_F_BUFFER_SELECTED)
4601 cflags = io_put_recv_kbuf(req);
4602 /* fast path, check for non-NULL to avoid function call */
4604 kfree(kmsg->free_iov);
4605 req->flags &= ~REQ_F_NEED_CLEANUP;
4607 req_set_fail_links(req);
4608 __io_req_complete(req, issue_flags, ret, cflags);
4612 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
4614 struct io_buffer *kbuf;
4615 struct io_sr_msg *sr = &req->sr_msg;
4617 void __user *buf = sr->buf;
4618 struct socket *sock;
4621 int ret, cflags = 0;
4622 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4624 sock = sock_from_file(req->file);
4625 if (unlikely(!sock))
4628 if (req->flags & REQ_F_BUFFER_SELECT) {
4629 kbuf = io_recv_buffer_select(req, !force_nonblock);
4631 return PTR_ERR(kbuf);
4632 buf = u64_to_user_ptr(kbuf->addr);
4635 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4639 msg.msg_name = NULL;
4640 msg.msg_control = NULL;
4641 msg.msg_controllen = 0;
4642 msg.msg_namelen = 0;
4643 msg.msg_iocb = NULL;
4646 flags = req->sr_msg.msg_flags;
4647 if (flags & MSG_DONTWAIT)
4648 req->flags |= REQ_F_NOWAIT;
4649 else if (force_nonblock)
4650 flags |= MSG_DONTWAIT;
4652 ret = sock_recvmsg(sock, &msg, flags);
4653 if (force_nonblock && ret == -EAGAIN)
4655 if (ret == -ERESTARTSYS)
4658 if (req->flags & REQ_F_BUFFER_SELECTED)
4659 cflags = io_put_recv_kbuf(req);
4661 req_set_fail_links(req);
4662 __io_req_complete(req, issue_flags, ret, cflags);
4666 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4668 struct io_accept *accept = &req->accept;
4670 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4672 if (sqe->ioprio || sqe->len || sqe->buf_index)
4675 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4676 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4677 accept->flags = READ_ONCE(sqe->accept_flags);
4678 accept->nofile = rlimit(RLIMIT_NOFILE);
4682 static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
4684 struct io_accept *accept = &req->accept;
4685 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4686 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4689 if (req->file->f_flags & O_NONBLOCK)
4690 req->flags |= REQ_F_NOWAIT;
4692 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4693 accept->addr_len, accept->flags,
4695 if (ret == -EAGAIN && force_nonblock)
4698 if (ret == -ERESTARTSYS)
4700 req_set_fail_links(req);
4702 __io_req_complete(req, issue_flags, ret, 0);
4706 static int io_connect_prep_async(struct io_kiocb *req)
4708 struct io_async_connect *io = req->async_data;
4709 struct io_connect *conn = &req->connect;
4711 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
4714 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4716 struct io_connect *conn = &req->connect;
4718 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4720 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4723 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4724 conn->addr_len = READ_ONCE(sqe->addr2);
4728 static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
4730 struct io_async_connect __io, *io;
4731 unsigned file_flags;
4733 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4735 if (req->async_data) {
4736 io = req->async_data;
4738 ret = move_addr_to_kernel(req->connect.addr,
4739 req->connect.addr_len,
4746 file_flags = force_nonblock ? O_NONBLOCK : 0;
4748 ret = __sys_connect_file(req->file, &io->address,
4749 req->connect.addr_len, file_flags);
4750 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4751 if (req->async_data)
4753 if (io_alloc_async_data(req)) {
4757 io = req->async_data;
4758 memcpy(req->async_data, &__io, sizeof(__io));
4761 if (ret == -ERESTARTSYS)
4765 req_set_fail_links(req);
4766 __io_req_complete(req, issue_flags, ret, 0);
4769 #else /* !CONFIG_NET */
4770 #define IO_NETOP_FN(op) \
4771 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
4773 return -EOPNOTSUPP; \
4776 #define IO_NETOP_PREP(op) \
4778 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
4780 return -EOPNOTSUPP; \
4783 #define IO_NETOP_PREP_ASYNC(op) \
4785 static int io_##op##_prep_async(struct io_kiocb *req) \
4787 return -EOPNOTSUPP; \
4790 IO_NETOP_PREP_ASYNC(sendmsg);
4791 IO_NETOP_PREP_ASYNC(recvmsg);
4792 IO_NETOP_PREP_ASYNC(connect);
4793 IO_NETOP_PREP(accept);
4796 #endif /* CONFIG_NET */
4798 struct io_poll_table {
4799 struct poll_table_struct pt;
4800 struct io_kiocb *req;
4804 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4805 __poll_t mask, task_work_func_t func)
4809 /* for instances that support it check for an event match first: */
4810 if (mask && !(mask & poll->events))
4813 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4815 list_del_init(&poll->wait.entry);
4818 req->task_work.func = func;
4819 percpu_ref_get(&req->ctx->refs);
4822 * If this fails, then the task is exiting. When a task exits, the
4823 * work gets canceled, so just cancel this request as well instead
4824 * of executing it. We can't safely execute it anyway, as we may not
4825 * have the needed state needed for it anyway.
4827 ret = io_req_task_work_add(req);
4828 if (unlikely(ret)) {
4829 WRITE_ONCE(poll->canceled, true);
4830 io_req_task_work_add_fallback(req, func);
4835 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4836 __acquires(&req->ctx->completion_lock)
4838 struct io_ring_ctx *ctx = req->ctx;
4840 if (!req->result && !READ_ONCE(poll->canceled)) {
4841 struct poll_table_struct pt = { ._key = poll->events };
4843 req->result = vfs_poll(req->file, &pt) & poll->events;
4846 spin_lock_irq(&ctx->completion_lock);
4847 if (!req->result && !READ_ONCE(poll->canceled)) {
4848 add_wait_queue(poll->head, &poll->wait);
4855 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4857 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4858 if (req->opcode == IORING_OP_POLL_ADD)
4859 return req->async_data;
4860 return req->apoll->double_poll;
4863 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4865 if (req->opcode == IORING_OP_POLL_ADD)
4867 return &req->apoll->poll;
4870 static void io_poll_remove_double(struct io_kiocb *req)
4872 struct io_poll_iocb *poll = io_poll_get_double(req);
4874 lockdep_assert_held(&req->ctx->completion_lock);
4876 if (poll && poll->head) {
4877 struct wait_queue_head *head = poll->head;
4879 spin_lock(&head->lock);
4880 list_del_init(&poll->wait.entry);
4881 if (poll->wait.private)
4882 refcount_dec(&req->refs);
4884 spin_unlock(&head->lock);
4888 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
4890 struct io_ring_ctx *ctx = req->ctx;
4892 io_poll_remove_double(req);
4893 req->poll.done = true;
4894 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
4895 io_commit_cqring(ctx);
4898 static void io_poll_task_func(struct callback_head *cb)
4900 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4901 struct io_ring_ctx *ctx = req->ctx;
4902 struct io_kiocb *nxt;
4904 if (io_poll_rewait(req, &req->poll)) {
4905 spin_unlock_irq(&ctx->completion_lock);
4907 hash_del(&req->hash_node);
4908 io_poll_complete(req, req->result, 0);
4909 spin_unlock_irq(&ctx->completion_lock);
4911 nxt = io_put_req_find_next(req);
4912 io_cqring_ev_posted(ctx);
4914 __io_req_task_submit(nxt);
4917 percpu_ref_put(&ctx->refs);
4920 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4921 int sync, void *key)
4923 struct io_kiocb *req = wait->private;
4924 struct io_poll_iocb *poll = io_poll_get_single(req);
4925 __poll_t mask = key_to_poll(key);
4927 /* for instances that support it check for an event match first: */
4928 if (mask && !(mask & poll->events))
4931 list_del_init(&wait->entry);
4933 if (poll && poll->head) {
4936 spin_lock(&poll->head->lock);
4937 done = list_empty(&poll->wait.entry);
4939 list_del_init(&poll->wait.entry);
4940 /* make sure double remove sees this as being gone */
4941 wait->private = NULL;
4942 spin_unlock(&poll->head->lock);
4944 /* use wait func handler, so it matches the rq type */
4945 poll->wait.func(&poll->wait, mode, sync, key);
4948 refcount_dec(&req->refs);
4952 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
4953 wait_queue_func_t wake_func)
4957 poll->canceled = false;
4958 poll->events = events;
4959 INIT_LIST_HEAD(&poll->wait.entry);
4960 init_waitqueue_func_entry(&poll->wait, wake_func);
4963 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
4964 struct wait_queue_head *head,
4965 struct io_poll_iocb **poll_ptr)
4967 struct io_kiocb *req = pt->req;
4970 * If poll->head is already set, it's because the file being polled
4971 * uses multiple waitqueues for poll handling (eg one for read, one
4972 * for write). Setup a separate io_poll_iocb if this happens.
4974 if (unlikely(poll->head)) {
4975 struct io_poll_iocb *poll_one = poll;
4977 /* already have a 2nd entry, fail a third attempt */
4979 pt->error = -EINVAL;
4982 /* double add on the same waitqueue head, ignore */
4983 if (poll->head == head)
4985 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
4987 pt->error = -ENOMEM;
4990 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
4991 refcount_inc(&req->refs);
4992 poll->wait.private = req;
4999 if (poll->events & EPOLLEXCLUSIVE)
5000 add_wait_queue_exclusive(head, &poll->wait);
5002 add_wait_queue(head, &poll->wait);
5005 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5006 struct poll_table_struct *p)
5008 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5009 struct async_poll *apoll = pt->req->apoll;
5011 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5014 static void io_async_task_func(struct callback_head *cb)
5016 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
5017 struct async_poll *apoll = req->apoll;
5018 struct io_ring_ctx *ctx = req->ctx;
5020 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
5022 if (io_poll_rewait(req, &apoll->poll)) {
5023 spin_unlock_irq(&ctx->completion_lock);
5024 percpu_ref_put(&ctx->refs);
5028 /* If req is still hashed, it cannot have been canceled. Don't check. */
5029 if (hash_hashed(&req->hash_node))
5030 hash_del(&req->hash_node);
5032 io_poll_remove_double(req);
5033 spin_unlock_irq(&ctx->completion_lock);
5035 if (!READ_ONCE(apoll->poll.canceled))
5036 __io_req_task_submit(req);
5038 __io_req_task_cancel(req, -ECANCELED);
5040 percpu_ref_put(&ctx->refs);
5041 kfree(apoll->double_poll);
5045 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5048 struct io_kiocb *req = wait->private;
5049 struct io_poll_iocb *poll = &req->apoll->poll;
5051 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5054 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5057 static void io_poll_req_insert(struct io_kiocb *req)
5059 struct io_ring_ctx *ctx = req->ctx;
5060 struct hlist_head *list;
5062 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5063 hlist_add_head(&req->hash_node, list);
5066 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5067 struct io_poll_iocb *poll,
5068 struct io_poll_table *ipt, __poll_t mask,
5069 wait_queue_func_t wake_func)
5070 __acquires(&ctx->completion_lock)
5072 struct io_ring_ctx *ctx = req->ctx;
5073 bool cancel = false;
5075 INIT_HLIST_NODE(&req->hash_node);
5076 io_init_poll_iocb(poll, mask, wake_func);
5077 poll->file = req->file;
5078 poll->wait.private = req;
5080 ipt->pt._key = mask;
5082 ipt->error = -EINVAL;
5084 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5086 spin_lock_irq(&ctx->completion_lock);
5087 if (likely(poll->head)) {
5088 spin_lock(&poll->head->lock);
5089 if (unlikely(list_empty(&poll->wait.entry))) {
5095 if (mask || ipt->error)
5096 list_del_init(&poll->wait.entry);
5098 WRITE_ONCE(poll->canceled, true);
5099 else if (!poll->done) /* actually waiting for an event */
5100 io_poll_req_insert(req);
5101 spin_unlock(&poll->head->lock);
5107 static bool io_arm_poll_handler(struct io_kiocb *req)
5109 const struct io_op_def *def = &io_op_defs[req->opcode];
5110 struct io_ring_ctx *ctx = req->ctx;
5111 struct async_poll *apoll;
5112 struct io_poll_table ipt;
5116 if (!req->file || !file_can_poll(req->file))
5118 if (req->flags & REQ_F_POLLED)
5122 else if (def->pollout)
5126 /* if we can't nonblock try, then no point in arming a poll handler */
5127 if (!io_file_supports_async(req->file, rw))
5130 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5131 if (unlikely(!apoll))
5133 apoll->double_poll = NULL;
5135 req->flags |= REQ_F_POLLED;
5140 mask |= POLLIN | POLLRDNORM;
5142 mask |= POLLOUT | POLLWRNORM;
5144 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5145 if ((req->opcode == IORING_OP_RECVMSG) &&
5146 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5149 mask |= POLLERR | POLLPRI;
5151 ipt.pt._qproc = io_async_queue_proc;
5153 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5155 if (ret || ipt.error) {
5156 io_poll_remove_double(req);
5157 spin_unlock_irq(&ctx->completion_lock);
5158 kfree(apoll->double_poll);
5162 spin_unlock_irq(&ctx->completion_lock);
5163 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
5164 apoll->poll.events);
5168 static bool __io_poll_remove_one(struct io_kiocb *req,
5169 struct io_poll_iocb *poll)
5171 bool do_complete = false;
5173 spin_lock(&poll->head->lock);
5174 WRITE_ONCE(poll->canceled, true);
5175 if (!list_empty(&poll->wait.entry)) {
5176 list_del_init(&poll->wait.entry);
5179 spin_unlock(&poll->head->lock);
5180 hash_del(&req->hash_node);
5184 static bool io_poll_remove_one(struct io_kiocb *req)
5188 io_poll_remove_double(req);
5190 if (req->opcode == IORING_OP_POLL_ADD) {
5191 do_complete = __io_poll_remove_one(req, &req->poll);
5193 struct async_poll *apoll = req->apoll;
5195 /* non-poll requests have submit ref still */
5196 do_complete = __io_poll_remove_one(req, &apoll->poll);
5199 kfree(apoll->double_poll);
5205 io_cqring_fill_event(req, -ECANCELED);
5206 io_commit_cqring(req->ctx);
5207 req_set_fail_links(req);
5208 io_put_req_deferred(req, 1);
5215 * Returns true if we found and killed one or more poll requests
5217 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5218 struct files_struct *files)
5220 struct hlist_node *tmp;
5221 struct io_kiocb *req;
5224 spin_lock_irq(&ctx->completion_lock);
5225 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5226 struct hlist_head *list;
5228 list = &ctx->cancel_hash[i];
5229 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5230 if (io_match_task(req, tsk, files))
5231 posted += io_poll_remove_one(req);
5234 spin_unlock_irq(&ctx->completion_lock);
5237 io_cqring_ev_posted(ctx);
5242 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
5244 struct hlist_head *list;
5245 struct io_kiocb *req;
5247 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5248 hlist_for_each_entry(req, list, hash_node) {
5249 if (sqe_addr != req->user_data)
5251 if (io_poll_remove_one(req))
5259 static int io_poll_remove_prep(struct io_kiocb *req,
5260 const struct io_uring_sqe *sqe)
5262 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5264 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
5268 req->poll_remove.addr = READ_ONCE(sqe->addr);
5273 * Find a running poll command that matches one specified in sqe->addr,
5274 * and remove it if found.
5276 static int io_poll_remove(struct io_kiocb *req, unsigned int issue_flags)
5278 struct io_ring_ctx *ctx = req->ctx;
5281 spin_lock_irq(&ctx->completion_lock);
5282 ret = io_poll_cancel(ctx, req->poll_remove.addr);
5283 spin_unlock_irq(&ctx->completion_lock);
5286 req_set_fail_links(req);
5287 io_req_complete(req, ret);
5291 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5294 struct io_kiocb *req = wait->private;
5295 struct io_poll_iocb *poll = &req->poll;
5297 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5300 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5301 struct poll_table_struct *p)
5303 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5305 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5308 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5310 struct io_poll_iocb *poll = &req->poll;
5313 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5315 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
5318 events = READ_ONCE(sqe->poll32_events);
5320 events = swahw32(events);
5322 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP |
5323 (events & EPOLLEXCLUSIVE);
5327 static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5329 struct io_poll_iocb *poll = &req->poll;
5330 struct io_ring_ctx *ctx = req->ctx;
5331 struct io_poll_table ipt;
5334 ipt.pt._qproc = io_poll_queue_proc;
5336 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5339 if (mask) { /* no async, we'd stolen it */
5341 io_poll_complete(req, mask, 0);
5343 spin_unlock_irq(&ctx->completion_lock);
5346 io_cqring_ev_posted(ctx);
5352 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5354 struct io_timeout_data *data = container_of(timer,
5355 struct io_timeout_data, timer);
5356 struct io_kiocb *req = data->req;
5357 struct io_ring_ctx *ctx = req->ctx;
5358 unsigned long flags;
5360 spin_lock_irqsave(&ctx->completion_lock, flags);
5361 list_del_init(&req->timeout.list);
5362 atomic_set(&req->ctx->cq_timeouts,
5363 atomic_read(&req->ctx->cq_timeouts) + 1);
5365 io_cqring_fill_event(req, -ETIME);
5366 io_commit_cqring(ctx);
5367 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5369 io_cqring_ev_posted(ctx);
5370 req_set_fail_links(req);
5372 return HRTIMER_NORESTART;
5375 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5378 struct io_timeout_data *io;
5379 struct io_kiocb *req;
5382 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5383 if (user_data == req->user_data) {
5390 return ERR_PTR(ret);
5392 io = req->async_data;
5393 ret = hrtimer_try_to_cancel(&io->timer);
5395 return ERR_PTR(-EALREADY);
5396 list_del_init(&req->timeout.list);
5400 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5402 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5405 return PTR_ERR(req);
5407 req_set_fail_links(req);
5408 io_cqring_fill_event(req, -ECANCELED);
5409 io_put_req_deferred(req, 1);
5413 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5414 struct timespec64 *ts, enum hrtimer_mode mode)
5416 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5417 struct io_timeout_data *data;
5420 return PTR_ERR(req);
5422 req->timeout.off = 0; /* noseq */
5423 data = req->async_data;
5424 list_add_tail(&req->timeout.list, &ctx->timeout_list);
5425 hrtimer_init(&data->timer, CLOCK_MONOTONIC, mode);
5426 data->timer.function = io_timeout_fn;
5427 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
5431 static int io_timeout_remove_prep(struct io_kiocb *req,
5432 const struct io_uring_sqe *sqe)
5434 struct io_timeout_rem *tr = &req->timeout_rem;
5436 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5438 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5440 if (sqe->ioprio || sqe->buf_index || sqe->len)
5443 tr->addr = READ_ONCE(sqe->addr);
5444 tr->flags = READ_ONCE(sqe->timeout_flags);
5445 if (tr->flags & IORING_TIMEOUT_UPDATE) {
5446 if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS))
5448 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
5450 } else if (tr->flags) {
5451 /* timeout removal doesn't support flags */
5458 static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
5460 return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
5465 * Remove or update an existing timeout command
5467 static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
5469 struct io_timeout_rem *tr = &req->timeout_rem;
5470 struct io_ring_ctx *ctx = req->ctx;
5473 spin_lock_irq(&ctx->completion_lock);
5474 if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE))
5475 ret = io_timeout_cancel(ctx, tr->addr);
5477 ret = io_timeout_update(ctx, tr->addr, &tr->ts,
5478 io_translate_timeout_mode(tr->flags));
5480 io_cqring_fill_event(req, ret);
5481 io_commit_cqring(ctx);
5482 spin_unlock_irq(&ctx->completion_lock);
5483 io_cqring_ev_posted(ctx);
5485 req_set_fail_links(req);
5490 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5491 bool is_timeout_link)
5493 struct io_timeout_data *data;
5495 u32 off = READ_ONCE(sqe->off);
5497 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5499 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5501 if (off && is_timeout_link)
5503 flags = READ_ONCE(sqe->timeout_flags);
5504 if (flags & ~IORING_TIMEOUT_ABS)
5507 req->timeout.off = off;
5509 if (!req->async_data && io_alloc_async_data(req))
5512 data = req->async_data;
5515 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5518 data->mode = io_translate_timeout_mode(flags);
5519 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5523 static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
5525 struct io_ring_ctx *ctx = req->ctx;
5526 struct io_timeout_data *data = req->async_data;
5527 struct list_head *entry;
5528 u32 tail, off = req->timeout.off;
5530 spin_lock_irq(&ctx->completion_lock);
5533 * sqe->off holds how many events that need to occur for this
5534 * timeout event to be satisfied. If it isn't set, then this is
5535 * a pure timeout request, sequence isn't used.
5537 if (io_is_timeout_noseq(req)) {
5538 entry = ctx->timeout_list.prev;
5542 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5543 req->timeout.target_seq = tail + off;
5545 /* Update the last seq here in case io_flush_timeouts() hasn't.
5546 * This is safe because ->completion_lock is held, and submissions
5547 * and completions are never mixed in the same ->completion_lock section.
5549 ctx->cq_last_tm_flush = tail;
5552 * Insertion sort, ensuring the first entry in the list is always
5553 * the one we need first.
5555 list_for_each_prev(entry, &ctx->timeout_list) {
5556 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5559 if (io_is_timeout_noseq(nxt))
5561 /* nxt.seq is behind @tail, otherwise would've been completed */
5562 if (off >= nxt->timeout.target_seq - tail)
5566 list_add(&req->timeout.list, entry);
5567 data->timer.function = io_timeout_fn;
5568 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5569 spin_unlock_irq(&ctx->completion_lock);
5573 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5575 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5577 return req->user_data == (unsigned long) data;
5580 static int io_async_cancel_one(struct io_uring_task *tctx, void *sqe_addr)
5582 enum io_wq_cancel cancel_ret;
5588 cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, sqe_addr, false);
5589 switch (cancel_ret) {
5590 case IO_WQ_CANCEL_OK:
5593 case IO_WQ_CANCEL_RUNNING:
5596 case IO_WQ_CANCEL_NOTFOUND:
5604 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5605 struct io_kiocb *req, __u64 sqe_addr,
5608 unsigned long flags;
5611 ret = io_async_cancel_one(req->task->io_uring,
5612 (void *) (unsigned long) sqe_addr);
5613 if (ret != -ENOENT) {
5614 spin_lock_irqsave(&ctx->completion_lock, flags);
5618 spin_lock_irqsave(&ctx->completion_lock, flags);
5619 ret = io_timeout_cancel(ctx, sqe_addr);
5622 ret = io_poll_cancel(ctx, sqe_addr);
5626 io_cqring_fill_event(req, ret);
5627 io_commit_cqring(ctx);
5628 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5629 io_cqring_ev_posted(ctx);
5632 req_set_fail_links(req);
5636 static int io_async_cancel_prep(struct io_kiocb *req,
5637 const struct io_uring_sqe *sqe)
5639 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5641 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5643 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5646 req->cancel.addr = READ_ONCE(sqe->addr);
5650 static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
5652 struct io_ring_ctx *ctx = req->ctx;
5654 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
5658 static int io_rsrc_update_prep(struct io_kiocb *req,
5659 const struct io_uring_sqe *sqe)
5661 if (unlikely(req->ctx->flags & IORING_SETUP_SQPOLL))
5663 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5665 if (sqe->ioprio || sqe->rw_flags)
5668 req->rsrc_update.offset = READ_ONCE(sqe->off);
5669 req->rsrc_update.nr_args = READ_ONCE(sqe->len);
5670 if (!req->rsrc_update.nr_args)
5672 req->rsrc_update.arg = READ_ONCE(sqe->addr);
5676 static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
5678 struct io_ring_ctx *ctx = req->ctx;
5679 struct io_uring_rsrc_update up;
5682 if (issue_flags & IO_URING_F_NONBLOCK)
5685 up.offset = req->rsrc_update.offset;
5686 up.data = req->rsrc_update.arg;
5688 mutex_lock(&ctx->uring_lock);
5689 ret = __io_sqe_files_update(ctx, &up, req->rsrc_update.nr_args);
5690 mutex_unlock(&ctx->uring_lock);
5693 req_set_fail_links(req);
5694 __io_req_complete(req, issue_flags, ret, 0);
5698 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5700 switch (req->opcode) {
5703 case IORING_OP_READV:
5704 case IORING_OP_READ_FIXED:
5705 case IORING_OP_READ:
5706 return io_read_prep(req, sqe);
5707 case IORING_OP_WRITEV:
5708 case IORING_OP_WRITE_FIXED:
5709 case IORING_OP_WRITE:
5710 return io_write_prep(req, sqe);
5711 case IORING_OP_POLL_ADD:
5712 return io_poll_add_prep(req, sqe);
5713 case IORING_OP_POLL_REMOVE:
5714 return io_poll_remove_prep(req, sqe);
5715 case IORING_OP_FSYNC:
5716 return io_fsync_prep(req, sqe);
5717 case IORING_OP_SYNC_FILE_RANGE:
5718 return io_sfr_prep(req, sqe);
5719 case IORING_OP_SENDMSG:
5720 case IORING_OP_SEND:
5721 return io_sendmsg_prep(req, sqe);
5722 case IORING_OP_RECVMSG:
5723 case IORING_OP_RECV:
5724 return io_recvmsg_prep(req, sqe);
5725 case IORING_OP_CONNECT:
5726 return io_connect_prep(req, sqe);
5727 case IORING_OP_TIMEOUT:
5728 return io_timeout_prep(req, sqe, false);
5729 case IORING_OP_TIMEOUT_REMOVE:
5730 return io_timeout_remove_prep(req, sqe);
5731 case IORING_OP_ASYNC_CANCEL:
5732 return io_async_cancel_prep(req, sqe);
5733 case IORING_OP_LINK_TIMEOUT:
5734 return io_timeout_prep(req, sqe, true);
5735 case IORING_OP_ACCEPT:
5736 return io_accept_prep(req, sqe);
5737 case IORING_OP_FALLOCATE:
5738 return io_fallocate_prep(req, sqe);
5739 case IORING_OP_OPENAT:
5740 return io_openat_prep(req, sqe);
5741 case IORING_OP_CLOSE:
5742 return io_close_prep(req, sqe);
5743 case IORING_OP_FILES_UPDATE:
5744 return io_rsrc_update_prep(req, sqe);
5745 case IORING_OP_STATX:
5746 return io_statx_prep(req, sqe);
5747 case IORING_OP_FADVISE:
5748 return io_fadvise_prep(req, sqe);
5749 case IORING_OP_MADVISE:
5750 return io_madvise_prep(req, sqe);
5751 case IORING_OP_OPENAT2:
5752 return io_openat2_prep(req, sqe);
5753 case IORING_OP_EPOLL_CTL:
5754 return io_epoll_ctl_prep(req, sqe);
5755 case IORING_OP_SPLICE:
5756 return io_splice_prep(req, sqe);
5757 case IORING_OP_PROVIDE_BUFFERS:
5758 return io_provide_buffers_prep(req, sqe);
5759 case IORING_OP_REMOVE_BUFFERS:
5760 return io_remove_buffers_prep(req, sqe);
5762 return io_tee_prep(req, sqe);
5763 case IORING_OP_SHUTDOWN:
5764 return io_shutdown_prep(req, sqe);
5765 case IORING_OP_RENAMEAT:
5766 return io_renameat_prep(req, sqe);
5767 case IORING_OP_UNLINKAT:
5768 return io_unlinkat_prep(req, sqe);
5771 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5776 static int io_req_prep_async(struct io_kiocb *req)
5778 switch (req->opcode) {
5779 case IORING_OP_READV:
5780 case IORING_OP_READ_FIXED:
5781 case IORING_OP_READ:
5782 return io_rw_prep_async(req, READ);
5783 case IORING_OP_WRITEV:
5784 case IORING_OP_WRITE_FIXED:
5785 case IORING_OP_WRITE:
5786 return io_rw_prep_async(req, WRITE);
5787 case IORING_OP_SENDMSG:
5788 case IORING_OP_SEND:
5789 return io_sendmsg_prep_async(req);
5790 case IORING_OP_RECVMSG:
5791 case IORING_OP_RECV:
5792 return io_recvmsg_prep_async(req);
5793 case IORING_OP_CONNECT:
5794 return io_connect_prep_async(req);
5799 static int io_req_defer_prep(struct io_kiocb *req)
5801 if (!io_op_defs[req->opcode].needs_async_data)
5803 /* some opcodes init it during the inital prep */
5804 if (req->async_data)
5806 if (__io_alloc_async_data(req))
5808 return io_req_prep_async(req);
5811 static u32 io_get_sequence(struct io_kiocb *req)
5813 struct io_kiocb *pos;
5814 struct io_ring_ctx *ctx = req->ctx;
5815 u32 total_submitted, nr_reqs = 0;
5817 io_for_each_link(pos, req)
5820 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
5821 return total_submitted - nr_reqs;
5824 static int io_req_defer(struct io_kiocb *req)
5826 struct io_ring_ctx *ctx = req->ctx;
5827 struct io_defer_entry *de;
5831 /* Still need defer if there is pending req in defer list. */
5832 if (likely(list_empty_careful(&ctx->defer_list) &&
5833 !(req->flags & REQ_F_IO_DRAIN)))
5836 seq = io_get_sequence(req);
5837 /* Still a chance to pass the sequence check */
5838 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
5841 ret = io_req_defer_prep(req);
5844 io_prep_async_link(req);
5845 de = kmalloc(sizeof(*de), GFP_KERNEL);
5849 spin_lock_irq(&ctx->completion_lock);
5850 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
5851 spin_unlock_irq(&ctx->completion_lock);
5853 io_queue_async_work(req);
5854 return -EIOCBQUEUED;
5857 trace_io_uring_defer(ctx, req, req->user_data);
5860 list_add_tail(&de->list, &ctx->defer_list);
5861 spin_unlock_irq(&ctx->completion_lock);
5862 return -EIOCBQUEUED;
5865 static void __io_clean_op(struct io_kiocb *req)
5867 if (req->flags & REQ_F_BUFFER_SELECTED) {
5868 switch (req->opcode) {
5869 case IORING_OP_READV:
5870 case IORING_OP_READ_FIXED:
5871 case IORING_OP_READ:
5872 kfree((void *)(unsigned long)req->rw.addr);
5874 case IORING_OP_RECVMSG:
5875 case IORING_OP_RECV:
5876 kfree(req->sr_msg.kbuf);
5879 req->flags &= ~REQ_F_BUFFER_SELECTED;
5882 if (req->flags & REQ_F_NEED_CLEANUP) {
5883 switch (req->opcode) {
5884 case IORING_OP_READV:
5885 case IORING_OP_READ_FIXED:
5886 case IORING_OP_READ:
5887 case IORING_OP_WRITEV:
5888 case IORING_OP_WRITE_FIXED:
5889 case IORING_OP_WRITE: {
5890 struct io_async_rw *io = req->async_data;
5892 kfree(io->free_iovec);
5895 case IORING_OP_RECVMSG:
5896 case IORING_OP_SENDMSG: {
5897 struct io_async_msghdr *io = req->async_data;
5899 kfree(io->free_iov);
5902 case IORING_OP_SPLICE:
5904 io_put_file(req, req->splice.file_in,
5905 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
5907 case IORING_OP_OPENAT:
5908 case IORING_OP_OPENAT2:
5909 if (req->open.filename)
5910 putname(req->open.filename);
5912 case IORING_OP_RENAMEAT:
5913 putname(req->rename.oldpath);
5914 putname(req->rename.newpath);
5916 case IORING_OP_UNLINKAT:
5917 putname(req->unlink.filename);
5920 req->flags &= ~REQ_F_NEED_CLEANUP;
5924 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
5926 struct io_ring_ctx *ctx = req->ctx;
5927 const struct cred *creds = NULL;
5930 if (req->work.personality) {
5931 const struct cred *new_creds;
5933 if (!(issue_flags & IO_URING_F_NONBLOCK))
5934 mutex_lock(&ctx->uring_lock);
5935 new_creds = idr_find(&ctx->personality_idr, req->work.personality);
5936 if (!(issue_flags & IO_URING_F_NONBLOCK))
5937 mutex_unlock(&ctx->uring_lock);
5940 creds = override_creds(new_creds);
5943 switch (req->opcode) {
5945 ret = io_nop(req, issue_flags);
5947 case IORING_OP_READV:
5948 case IORING_OP_READ_FIXED:
5949 case IORING_OP_READ:
5950 ret = io_read(req, issue_flags);
5952 case IORING_OP_WRITEV:
5953 case IORING_OP_WRITE_FIXED:
5954 case IORING_OP_WRITE:
5955 ret = io_write(req, issue_flags);
5957 case IORING_OP_FSYNC:
5958 ret = io_fsync(req, issue_flags);
5960 case IORING_OP_POLL_ADD:
5961 ret = io_poll_add(req, issue_flags);
5963 case IORING_OP_POLL_REMOVE:
5964 ret = io_poll_remove(req, issue_flags);
5966 case IORING_OP_SYNC_FILE_RANGE:
5967 ret = io_sync_file_range(req, issue_flags);
5969 case IORING_OP_SENDMSG:
5970 ret = io_sendmsg(req, issue_flags);
5972 case IORING_OP_SEND:
5973 ret = io_send(req, issue_flags);
5975 case IORING_OP_RECVMSG:
5976 ret = io_recvmsg(req, issue_flags);
5978 case IORING_OP_RECV:
5979 ret = io_recv(req, issue_flags);
5981 case IORING_OP_TIMEOUT:
5982 ret = io_timeout(req, issue_flags);
5984 case IORING_OP_TIMEOUT_REMOVE:
5985 ret = io_timeout_remove(req, issue_flags);
5987 case IORING_OP_ACCEPT:
5988 ret = io_accept(req, issue_flags);
5990 case IORING_OP_CONNECT:
5991 ret = io_connect(req, issue_flags);
5993 case IORING_OP_ASYNC_CANCEL:
5994 ret = io_async_cancel(req, issue_flags);
5996 case IORING_OP_FALLOCATE:
5997 ret = io_fallocate(req, issue_flags);
5999 case IORING_OP_OPENAT:
6000 ret = io_openat(req, issue_flags);
6002 case IORING_OP_CLOSE:
6003 ret = io_close(req, issue_flags);
6005 case IORING_OP_FILES_UPDATE:
6006 ret = io_files_update(req, issue_flags);
6008 case IORING_OP_STATX:
6009 ret = io_statx(req, issue_flags);
6011 case IORING_OP_FADVISE:
6012 ret = io_fadvise(req, issue_flags);
6014 case IORING_OP_MADVISE:
6015 ret = io_madvise(req, issue_flags);
6017 case IORING_OP_OPENAT2:
6018 ret = io_openat2(req, issue_flags);
6020 case IORING_OP_EPOLL_CTL:
6021 ret = io_epoll_ctl(req, issue_flags);
6023 case IORING_OP_SPLICE:
6024 ret = io_splice(req, issue_flags);
6026 case IORING_OP_PROVIDE_BUFFERS:
6027 ret = io_provide_buffers(req, issue_flags);
6029 case IORING_OP_REMOVE_BUFFERS:
6030 ret = io_remove_buffers(req, issue_flags);
6033 ret = io_tee(req, issue_flags);
6035 case IORING_OP_SHUTDOWN:
6036 ret = io_shutdown(req, issue_flags);
6038 case IORING_OP_RENAMEAT:
6039 ret = io_renameat(req, issue_flags);
6041 case IORING_OP_UNLINKAT:
6042 ret = io_unlinkat(req, issue_flags);
6050 revert_creds(creds);
6055 /* If the op doesn't have a file, we're not polling for it */
6056 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
6057 const bool in_async = io_wq_current_is_worker();
6059 /* workqueue context doesn't hold uring_lock, grab it now */
6061 mutex_lock(&ctx->uring_lock);
6063 io_iopoll_req_issued(req, in_async);
6066 mutex_unlock(&ctx->uring_lock);
6072 static void io_wq_submit_work(struct io_wq_work *work)
6074 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6075 struct io_kiocb *timeout;
6078 timeout = io_prep_linked_timeout(req);
6080 io_queue_linked_timeout(timeout);
6082 if (work->flags & IO_WQ_WORK_CANCEL)
6087 ret = io_issue_sqe(req, 0);
6089 * We can get EAGAIN for polled IO even though we're
6090 * forcing a sync submission from here, since we can't
6091 * wait for request slots on the block side.
6099 /* avoid locking problems by failing it from a clean context */
6101 /* io-wq is going to take one down */
6102 refcount_inc(&req->refs);
6103 io_req_task_queue_fail(req, ret);
6107 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6110 struct fixed_rsrc_table *table;
6112 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
6113 return table->files[index & IORING_FILE_TABLE_MASK];
6116 static struct file *io_file_get(struct io_submit_state *state,
6117 struct io_kiocb *req, int fd, bool fixed)
6119 struct io_ring_ctx *ctx = req->ctx;
6123 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6125 fd = array_index_nospec(fd, ctx->nr_user_files);
6126 file = io_file_from_index(ctx, fd);
6127 io_set_resource_node(req);
6129 trace_io_uring_file_get(ctx, fd);
6130 file = __io_file_get(state, fd);
6133 if (file && unlikely(file->f_op == &io_uring_fops))
6134 io_req_track_inflight(req);
6138 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6140 struct io_timeout_data *data = container_of(timer,
6141 struct io_timeout_data, timer);
6142 struct io_kiocb *prev, *req = data->req;
6143 struct io_ring_ctx *ctx = req->ctx;
6144 unsigned long flags;
6146 spin_lock_irqsave(&ctx->completion_lock, flags);
6147 prev = req->timeout.head;
6148 req->timeout.head = NULL;
6151 * We don't expect the list to be empty, that will only happen if we
6152 * race with the completion of the linked work.
6154 if (prev && refcount_inc_not_zero(&prev->refs))
6155 io_remove_next_linked(prev);
6158 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6161 req_set_fail_links(prev);
6162 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6163 io_put_req_deferred(prev, 1);
6165 io_req_complete_post(req, -ETIME, 0);
6166 io_put_req_deferred(req, 1);
6168 return HRTIMER_NORESTART;
6171 static void __io_queue_linked_timeout(struct io_kiocb *req)
6174 * If the back reference is NULL, then our linked request finished
6175 * before we got a chance to setup the timer
6177 if (req->timeout.head) {
6178 struct io_timeout_data *data = req->async_data;
6180 data->timer.function = io_link_timeout_fn;
6181 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6186 static void io_queue_linked_timeout(struct io_kiocb *req)
6188 struct io_ring_ctx *ctx = req->ctx;
6190 spin_lock_irq(&ctx->completion_lock);
6191 __io_queue_linked_timeout(req);
6192 spin_unlock_irq(&ctx->completion_lock);
6194 /* drop submission reference */
6198 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6200 struct io_kiocb *nxt = req->link;
6202 if (!nxt || (req->flags & REQ_F_LINK_TIMEOUT) ||
6203 nxt->opcode != IORING_OP_LINK_TIMEOUT)
6206 nxt->timeout.head = req;
6207 nxt->flags |= REQ_F_LTIMEOUT_ACTIVE;
6208 req->flags |= REQ_F_LINK_TIMEOUT;
6212 static void __io_queue_sqe(struct io_kiocb *req)
6214 struct io_kiocb *linked_timeout = io_prep_linked_timeout(req);
6217 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
6220 * We async punt it if the file wasn't marked NOWAIT, or if the file
6221 * doesn't support non-blocking read/write attempts
6223 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6224 if (!io_arm_poll_handler(req)) {
6226 * Queued up for async execution, worker will release
6227 * submit reference when the iocb is actually submitted.
6229 io_queue_async_work(req);
6231 } else if (likely(!ret)) {
6232 /* drop submission reference */
6233 if (req->flags & REQ_F_COMPLETE_INLINE) {
6234 struct io_ring_ctx *ctx = req->ctx;
6235 struct io_comp_state *cs = &ctx->submit_state.comp;
6237 cs->reqs[cs->nr++] = req;
6238 if (cs->nr == ARRAY_SIZE(cs->reqs))
6239 io_submit_flush_completions(cs, ctx);
6244 req_set_fail_links(req);
6246 io_req_complete(req, ret);
6249 io_queue_linked_timeout(linked_timeout);
6252 static void io_queue_sqe(struct io_kiocb *req)
6256 ret = io_req_defer(req);
6258 if (ret != -EIOCBQUEUED) {
6260 req_set_fail_links(req);
6262 io_req_complete(req, ret);
6264 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6265 ret = io_req_defer_prep(req);
6268 io_queue_async_work(req);
6270 __io_queue_sqe(req);
6275 * Check SQE restrictions (opcode and flags).
6277 * Returns 'true' if SQE is allowed, 'false' otherwise.
6279 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6280 struct io_kiocb *req,
6281 unsigned int sqe_flags)
6283 if (!ctx->restricted)
6286 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6289 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6290 ctx->restrictions.sqe_flags_required)
6293 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6294 ctx->restrictions.sqe_flags_required))
6300 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6301 const struct io_uring_sqe *sqe)
6303 struct io_submit_state *state;
6304 unsigned int sqe_flags;
6307 req->opcode = READ_ONCE(sqe->opcode);
6308 /* same numerical values with corresponding REQ_F_*, safe to copy */
6309 req->flags = sqe_flags = READ_ONCE(sqe->flags);
6310 req->user_data = READ_ONCE(sqe->user_data);
6311 req->async_data = NULL;
6315 req->fixed_rsrc_refs = NULL;
6316 /* one is dropped after submission, the other at completion */
6317 refcount_set(&req->refs, 2);
6318 req->task = current;
6321 /* enforce forwards compatibility on users */
6322 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS)) {
6327 if (unlikely(req->opcode >= IORING_OP_LAST))
6330 if (unlikely(!io_check_restriction(ctx, req, sqe_flags)))
6333 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6334 !io_op_defs[req->opcode].buffer_select)
6337 req->work.list.next = NULL;
6338 req->work.flags = 0;
6339 req->work.personality = READ_ONCE(sqe->personality);
6340 state = &ctx->submit_state;
6343 * Plug now if we have more than 1 IO left after this, and the target
6344 * is potentially a read/write to block based storage.
6346 if (!state->plug_started && state->ios_left > 1 &&
6347 io_op_defs[req->opcode].plug) {
6348 blk_start_plug(&state->plug);
6349 state->plug_started = true;
6352 if (io_op_defs[req->opcode].needs_file) {
6353 bool fixed = req->flags & REQ_F_FIXED_FILE;
6355 req->file = io_file_get(state, req, READ_ONCE(sqe->fd), fixed);
6356 if (unlikely(!req->file))
6364 static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
6365 const struct io_uring_sqe *sqe)
6367 struct io_submit_link *link = &ctx->submit_state.link;
6370 ret = io_init_req(ctx, req, sqe);
6371 if (unlikely(ret)) {
6374 io_req_complete(req, ret);
6376 /* fail even hard links since we don't submit */
6377 link->head->flags |= REQ_F_FAIL_LINK;
6378 io_put_req(link->head);
6379 io_req_complete(link->head, -ECANCELED);
6384 ret = io_req_prep(req, sqe);
6388 /* don't need @sqe from now on */
6389 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6390 true, ctx->flags & IORING_SETUP_SQPOLL);
6393 * If we already have a head request, queue this one for async
6394 * submittal once the head completes. If we don't have a head but
6395 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6396 * submitted sync once the chain is complete. If none of those
6397 * conditions are true (normal request), then just queue it.
6400 struct io_kiocb *head = link->head;
6403 * Taking sequential execution of a link, draining both sides
6404 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6405 * requests in the link. So, it drains the head and the
6406 * next after the link request. The last one is done via
6407 * drain_next flag to persist the effect across calls.
6409 if (req->flags & REQ_F_IO_DRAIN) {
6410 head->flags |= REQ_F_IO_DRAIN;
6411 ctx->drain_next = 1;
6413 ret = io_req_defer_prep(req);
6416 trace_io_uring_link(ctx, req, head);
6417 link->last->link = req;
6420 /* last request of a link, enqueue the link */
6421 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6426 if (unlikely(ctx->drain_next)) {
6427 req->flags |= REQ_F_IO_DRAIN;
6428 ctx->drain_next = 0;
6430 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6442 * Batched submission is done, ensure local IO is flushed out.
6444 static void io_submit_state_end(struct io_submit_state *state,
6445 struct io_ring_ctx *ctx)
6447 if (state->link.head)
6448 io_queue_sqe(state->link.head);
6450 io_submit_flush_completions(&state->comp, ctx);
6451 if (state->plug_started)
6452 blk_finish_plug(&state->plug);
6453 io_state_file_put(state);
6457 * Start submission side cache.
6459 static void io_submit_state_start(struct io_submit_state *state,
6460 unsigned int max_ios)
6462 state->plug_started = false;
6463 state->ios_left = max_ios;
6464 /* set only head, no need to init link_last in advance */
6465 state->link.head = NULL;
6468 static void io_commit_sqring(struct io_ring_ctx *ctx)
6470 struct io_rings *rings = ctx->rings;
6473 * Ensure any loads from the SQEs are done at this point,
6474 * since once we write the new head, the application could
6475 * write new data to them.
6477 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6481 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6482 * that is mapped by userspace. This means that care needs to be taken to
6483 * ensure that reads are stable, as we cannot rely on userspace always
6484 * being a good citizen. If members of the sqe are validated and then later
6485 * used, it's important that those reads are done through READ_ONCE() to
6486 * prevent a re-load down the line.
6488 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6490 u32 *sq_array = ctx->sq_array;
6494 * The cached sq head (or cq tail) serves two purposes:
6496 * 1) allows us to batch the cost of updating the user visible
6498 * 2) allows the kernel side to track the head on its own, even
6499 * though the application is the one updating it.
6501 head = READ_ONCE(sq_array[ctx->cached_sq_head++ & ctx->sq_mask]);
6502 if (likely(head < ctx->sq_entries))
6503 return &ctx->sq_sqes[head];
6505 /* drop invalid entries */
6506 ctx->cached_sq_dropped++;
6507 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6511 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6515 /* if we have a backlog and couldn't flush it all, return BUSY */
6516 if (test_bit(0, &ctx->sq_check_overflow)) {
6517 if (!__io_cqring_overflow_flush(ctx, false, NULL, NULL))
6521 /* make sure SQ entry isn't read before tail */
6522 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6524 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6527 percpu_counter_add(¤t->io_uring->inflight, nr);
6528 refcount_add(nr, ¤t->usage);
6529 io_submit_state_start(&ctx->submit_state, nr);
6531 while (submitted < nr) {
6532 const struct io_uring_sqe *sqe;
6533 struct io_kiocb *req;
6535 req = io_alloc_req(ctx);
6536 if (unlikely(!req)) {
6538 submitted = -EAGAIN;
6541 sqe = io_get_sqe(ctx);
6542 if (unlikely(!sqe)) {
6543 kmem_cache_free(req_cachep, req);
6546 /* will complete beyond this point, count as submitted */
6548 if (io_submit_sqe(ctx, req, sqe))
6552 if (unlikely(submitted != nr)) {
6553 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6554 struct io_uring_task *tctx = current->io_uring;
6555 int unused = nr - ref_used;
6557 percpu_ref_put_many(&ctx->refs, unused);
6558 percpu_counter_sub(&tctx->inflight, unused);
6559 put_task_struct_many(current, unused);
6562 io_submit_state_end(&ctx->submit_state, ctx);
6563 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6564 io_commit_sqring(ctx);
6569 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6571 /* Tell userspace we may need a wakeup call */
6572 spin_lock_irq(&ctx->completion_lock);
6573 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6574 spin_unlock_irq(&ctx->completion_lock);
6577 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6579 spin_lock_irq(&ctx->completion_lock);
6580 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6581 spin_unlock_irq(&ctx->completion_lock);
6584 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
6586 unsigned int to_submit;
6589 to_submit = io_sqring_entries(ctx);
6590 /* if we're handling multiple rings, cap submit size for fairness */
6591 if (cap_entries && to_submit > 8)
6594 if (!list_empty(&ctx->iopoll_list) || to_submit) {
6595 unsigned nr_events = 0;
6597 mutex_lock(&ctx->uring_lock);
6598 if (!list_empty(&ctx->iopoll_list))
6599 io_do_iopoll(ctx, &nr_events, 0);
6601 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)))
6602 ret = io_submit_sqes(ctx, to_submit);
6603 mutex_unlock(&ctx->uring_lock);
6606 if (!io_sqring_full(ctx) && wq_has_sleeper(&ctx->sqo_sq_wait))
6607 wake_up(&ctx->sqo_sq_wait);
6612 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
6614 struct io_ring_ctx *ctx;
6615 unsigned sq_thread_idle = 0;
6617 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6618 if (sq_thread_idle < ctx->sq_thread_idle)
6619 sq_thread_idle = ctx->sq_thread_idle;
6622 sqd->sq_thread_idle = sq_thread_idle;
6625 static void io_sqd_init_new(struct io_sq_data *sqd)
6627 struct io_ring_ctx *ctx;
6629 while (!list_empty(&sqd->ctx_new_list)) {
6630 ctx = list_first_entry(&sqd->ctx_new_list, struct io_ring_ctx, sqd_list);
6631 list_move_tail(&ctx->sqd_list, &sqd->ctx_list);
6632 complete(&ctx->sq_thread_comp);
6635 io_sqd_update_thread_idle(sqd);
6638 static bool io_sq_thread_should_stop(struct io_sq_data *sqd)
6640 return test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
6643 static bool io_sq_thread_should_park(struct io_sq_data *sqd)
6645 return test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
6648 static void io_sq_thread_parkme(struct io_sq_data *sqd)
6652 * TASK_PARKED is a special state; we must serialize against
6653 * possible pending wakeups to avoid store-store collisions on
6656 * Such a collision might possibly result in the task state
6657 * changin from TASK_PARKED and us failing the
6658 * wait_task_inactive() in kthread_park().
6660 set_special_state(TASK_PARKED);
6661 if (!test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state))
6665 * Thread is going to call schedule(), do not preempt it,
6666 * or the caller of kthread_park() may spend more time in
6667 * wait_task_inactive().
6670 complete(&sqd->completion);
6671 schedule_preempt_disabled();
6674 __set_current_state(TASK_RUNNING);
6677 static int io_sq_thread(void *data)
6679 struct io_sq_data *sqd = data;
6680 struct io_ring_ctx *ctx;
6681 unsigned long timeout = 0;
6682 char buf[TASK_COMM_LEN];
6685 sprintf(buf, "iou-sqp-%d", sqd->task_pid);
6686 set_task_comm(current, buf);
6687 sqd->thread = current;
6688 current->pf_io_worker = NULL;
6690 if (sqd->sq_cpu != -1)
6691 set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
6693 set_cpus_allowed_ptr(current, cpu_online_mask);
6694 current->flags |= PF_NO_SETAFFINITY;
6696 complete(&sqd->completion);
6698 wait_for_completion(&sqd->startup);
6700 while (!io_sq_thread_should_stop(sqd)) {
6702 bool cap_entries, sqt_spin, needs_sched;
6705 * Any changes to the sqd lists are synchronized through the
6706 * thread parking. This synchronizes the thread vs users,
6707 * the users are synchronized on the sqd->ctx_lock.
6709 if (io_sq_thread_should_park(sqd)) {
6710 io_sq_thread_parkme(sqd);
6713 if (unlikely(!list_empty(&sqd->ctx_new_list))) {
6714 io_sqd_init_new(sqd);
6715 timeout = jiffies + sqd->sq_thread_idle;
6717 if (fatal_signal_pending(current))
6720 cap_entries = !list_is_singular(&sqd->ctx_list);
6721 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6722 ret = __io_sq_thread(ctx, cap_entries);
6723 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
6727 if (sqt_spin || !time_after(jiffies, timeout)) {
6731 timeout = jiffies + sqd->sq_thread_idle;
6736 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
6737 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6738 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6739 !list_empty_careful(&ctx->iopoll_list)) {
6740 needs_sched = false;
6743 if (io_sqring_entries(ctx)) {
6744 needs_sched = false;
6749 if (needs_sched && !io_sq_thread_should_park(sqd)) {
6750 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6751 io_ring_set_wakeup_flag(ctx);
6754 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6755 io_ring_clear_wakeup_flag(ctx);
6758 finish_wait(&sqd->wait, &wait);
6759 timeout = jiffies + sqd->sq_thread_idle;
6762 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6763 io_uring_cancel_sqpoll(ctx);
6767 if (io_sq_thread_should_park(sqd))
6768 io_sq_thread_parkme(sqd);
6771 * Clear thread under lock so that concurrent parks work correctly
6773 complete(&sqd->completion);
6774 mutex_lock(&sqd->lock);
6776 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6778 io_ring_set_wakeup_flag(ctx);
6781 complete(&sqd->exited);
6782 mutex_unlock(&sqd->lock);
6786 struct io_wait_queue {
6787 struct wait_queue_entry wq;
6788 struct io_ring_ctx *ctx;
6790 unsigned nr_timeouts;
6793 static inline bool io_should_wake(struct io_wait_queue *iowq)
6795 struct io_ring_ctx *ctx = iowq->ctx;
6798 * Wake up if we have enough events, or if a timeout occurred since we
6799 * started waiting. For timeouts, we always want to return to userspace,
6800 * regardless of event count.
6802 return io_cqring_events(ctx) >= iowq->to_wait ||
6803 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6806 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6807 int wake_flags, void *key)
6809 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6813 * Cannot safely flush overflowed CQEs from here, ensure we wake up
6814 * the task, and the next invocation will do it.
6816 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->cq_check_overflow))
6817 return autoremove_wake_function(curr, mode, wake_flags, key);
6821 static int io_run_task_work_sig(void)
6823 if (io_run_task_work())
6825 if (!signal_pending(current))
6827 if (test_tsk_thread_flag(current, TIF_NOTIFY_SIGNAL))
6828 return -ERESTARTSYS;
6832 /* when returns >0, the caller should retry */
6833 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
6834 struct io_wait_queue *iowq,
6835 signed long *timeout)
6839 /* make sure we run task_work before checking for signals */
6840 ret = io_run_task_work_sig();
6841 if (ret || io_should_wake(iowq))
6843 /* let the caller flush overflows, retry */
6844 if (test_bit(0, &ctx->cq_check_overflow))
6847 *timeout = schedule_timeout(*timeout);
6848 return !*timeout ? -ETIME : 1;
6852 * Wait until events become available, if we don't already have some. The
6853 * application must reap them itself, as they reside on the shared cq ring.
6855 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6856 const sigset_t __user *sig, size_t sigsz,
6857 struct __kernel_timespec __user *uts)
6859 struct io_wait_queue iowq = {
6862 .func = io_wake_function,
6863 .entry = LIST_HEAD_INIT(iowq.wq.entry),
6866 .to_wait = min_events,
6868 struct io_rings *rings = ctx->rings;
6869 signed long timeout = MAX_SCHEDULE_TIMEOUT;
6873 io_cqring_overflow_flush(ctx, false, NULL, NULL);
6874 if (io_cqring_events(ctx) >= min_events)
6876 if (!io_run_task_work())
6881 #ifdef CONFIG_COMPAT
6882 if (in_compat_syscall())
6883 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
6887 ret = set_user_sigmask(sig, sigsz);
6894 struct timespec64 ts;
6896 if (get_timespec64(&ts, uts))
6898 timeout = timespec64_to_jiffies(&ts);
6901 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
6902 trace_io_uring_cqring_wait(ctx, min_events);
6904 io_cqring_overflow_flush(ctx, false, NULL, NULL);
6905 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
6906 TASK_INTERRUPTIBLE);
6907 ret = io_cqring_wait_schedule(ctx, &iowq, &timeout);
6908 finish_wait(&ctx->wait, &iowq.wq);
6911 restore_saved_sigmask_unless(ret == -EINTR);
6913 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
6916 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
6918 #if defined(CONFIG_UNIX)
6919 if (ctx->ring_sock) {
6920 struct sock *sock = ctx->ring_sock->sk;
6921 struct sk_buff *skb;
6923 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
6929 for (i = 0; i < ctx->nr_user_files; i++) {
6932 file = io_file_from_index(ctx, i);
6939 static void io_rsrc_data_ref_zero(struct percpu_ref *ref)
6941 struct fixed_rsrc_data *data;
6943 data = container_of(ref, struct fixed_rsrc_data, refs);
6944 complete(&data->done);
6947 static inline void io_rsrc_ref_lock(struct io_ring_ctx *ctx)
6949 spin_lock_bh(&ctx->rsrc_ref_lock);
6952 static inline void io_rsrc_ref_unlock(struct io_ring_ctx *ctx)
6954 spin_unlock_bh(&ctx->rsrc_ref_lock);
6957 static void io_sqe_rsrc_set_node(struct io_ring_ctx *ctx,
6958 struct fixed_rsrc_data *rsrc_data,
6959 struct fixed_rsrc_ref_node *ref_node)
6961 io_rsrc_ref_lock(ctx);
6962 rsrc_data->node = ref_node;
6963 list_add_tail(&ref_node->node, &ctx->rsrc_ref_list);
6964 io_rsrc_ref_unlock(ctx);
6965 percpu_ref_get(&rsrc_data->refs);
6968 static void io_sqe_rsrc_kill_node(struct io_ring_ctx *ctx, struct fixed_rsrc_data *data)
6970 struct fixed_rsrc_ref_node *ref_node = NULL;
6972 io_rsrc_ref_lock(ctx);
6973 ref_node = data->node;
6975 io_rsrc_ref_unlock(ctx);
6977 percpu_ref_kill(&ref_node->refs);
6980 static int io_rsrc_ref_quiesce(struct fixed_rsrc_data *data,
6981 struct io_ring_ctx *ctx,
6982 void (*rsrc_put)(struct io_ring_ctx *ctx,
6983 struct io_rsrc_put *prsrc))
6985 struct fixed_rsrc_ref_node *backup_node;
6991 data->quiesce = true;
6994 backup_node = alloc_fixed_rsrc_ref_node(ctx);
6997 backup_node->rsrc_data = data;
6998 backup_node->rsrc_put = rsrc_put;
7000 io_sqe_rsrc_kill_node(ctx, data);
7001 percpu_ref_kill(&data->refs);
7002 flush_delayed_work(&ctx->rsrc_put_work);
7004 ret = wait_for_completion_interruptible(&data->done);
7008 percpu_ref_resurrect(&data->refs);
7009 io_sqe_rsrc_set_node(ctx, data, backup_node);
7011 reinit_completion(&data->done);
7012 mutex_unlock(&ctx->uring_lock);
7013 ret = io_run_task_work_sig();
7014 mutex_lock(&ctx->uring_lock);
7016 data->quiesce = false;
7019 destroy_fixed_rsrc_ref_node(backup_node);
7023 static struct fixed_rsrc_data *alloc_fixed_rsrc_data(struct io_ring_ctx *ctx)
7025 struct fixed_rsrc_data *data;
7027 data = kzalloc(sizeof(*data), GFP_KERNEL);
7031 if (percpu_ref_init(&data->refs, io_rsrc_data_ref_zero,
7032 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) {
7037 init_completion(&data->done);
7041 static void free_fixed_rsrc_data(struct fixed_rsrc_data *data)
7043 percpu_ref_exit(&data->refs);
7048 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7050 struct fixed_rsrc_data *data = ctx->file_data;
7051 unsigned nr_tables, i;
7055 * percpu_ref_is_dying() is to stop parallel files unregister
7056 * Since we possibly drop uring lock later in this function to
7059 if (!data || percpu_ref_is_dying(&data->refs))
7061 ret = io_rsrc_ref_quiesce(data, ctx, io_ring_file_put);
7065 __io_sqe_files_unregister(ctx);
7066 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
7067 for (i = 0; i < nr_tables; i++)
7068 kfree(data->table[i].files);
7069 free_fixed_rsrc_data(data);
7070 ctx->file_data = NULL;
7071 ctx->nr_user_files = 0;
7075 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7076 __releases(&sqd->lock)
7080 if (sqd->thread == current)
7082 clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7083 wake_up_state(sqd->thread, TASK_PARKED);
7084 mutex_unlock(&sqd->lock);
7087 static bool io_sq_thread_park(struct io_sq_data *sqd)
7088 __acquires(&sqd->lock)
7090 if (sqd->thread == current)
7092 mutex_lock(&sqd->lock);
7094 mutex_unlock(&sqd->lock);
7097 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7098 wake_up_process(sqd->thread);
7099 wait_for_completion(&sqd->completion);
7103 static void io_sq_thread_stop(struct io_sq_data *sqd)
7105 if (test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state))
7107 mutex_lock(&sqd->lock);
7109 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7110 WARN_ON_ONCE(test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state));
7111 wake_up_process(sqd->thread);
7112 mutex_unlock(&sqd->lock);
7113 wait_for_completion(&sqd->exited);
7114 WARN_ON_ONCE(sqd->thread);
7116 mutex_unlock(&sqd->lock);
7120 static void io_put_sq_data(struct io_sq_data *sqd)
7122 if (refcount_dec_and_test(&sqd->refs)) {
7123 io_sq_thread_stop(sqd);
7128 static void io_sq_thread_finish(struct io_ring_ctx *ctx)
7130 struct io_sq_data *sqd = ctx->sq_data;
7133 complete(&sqd->startup);
7135 wait_for_completion(&ctx->sq_thread_comp);
7136 io_sq_thread_park(sqd);
7139 mutex_lock(&sqd->ctx_lock);
7140 list_del(&ctx->sqd_list);
7141 io_sqd_update_thread_idle(sqd);
7142 mutex_unlock(&sqd->ctx_lock);
7145 io_sq_thread_unpark(sqd);
7147 io_put_sq_data(sqd);
7148 ctx->sq_data = NULL;
7152 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7154 struct io_ring_ctx *ctx_attach;
7155 struct io_sq_data *sqd;
7158 f = fdget(p->wq_fd);
7160 return ERR_PTR(-ENXIO);
7161 if (f.file->f_op != &io_uring_fops) {
7163 return ERR_PTR(-EINVAL);
7166 ctx_attach = f.file->private_data;
7167 sqd = ctx_attach->sq_data;
7170 return ERR_PTR(-EINVAL);
7173 refcount_inc(&sqd->refs);
7178 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p)
7180 struct io_sq_data *sqd;
7182 if (p->flags & IORING_SETUP_ATTACH_WQ)
7183 return io_attach_sq_data(p);
7185 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7187 return ERR_PTR(-ENOMEM);
7189 refcount_set(&sqd->refs, 1);
7190 INIT_LIST_HEAD(&sqd->ctx_list);
7191 INIT_LIST_HEAD(&sqd->ctx_new_list);
7192 mutex_init(&sqd->ctx_lock);
7193 mutex_init(&sqd->lock);
7194 init_waitqueue_head(&sqd->wait);
7195 init_completion(&sqd->startup);
7196 init_completion(&sqd->completion);
7197 init_completion(&sqd->exited);
7201 #if defined(CONFIG_UNIX)
7203 * Ensure the UNIX gc is aware of our file set, so we are certain that
7204 * the io_uring can be safely unregistered on process exit, even if we have
7205 * loops in the file referencing.
7207 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7209 struct sock *sk = ctx->ring_sock->sk;
7210 struct scm_fp_list *fpl;
7211 struct sk_buff *skb;
7214 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7218 skb = alloc_skb(0, GFP_KERNEL);
7227 fpl->user = get_uid(current_user());
7228 for (i = 0; i < nr; i++) {
7229 struct file *file = io_file_from_index(ctx, i + offset);
7233 fpl->fp[nr_files] = get_file(file);
7234 unix_inflight(fpl->user, fpl->fp[nr_files]);
7239 fpl->max = SCM_MAX_FD;
7240 fpl->count = nr_files;
7241 UNIXCB(skb).fp = fpl;
7242 skb->destructor = unix_destruct_scm;
7243 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7244 skb_queue_head(&sk->sk_receive_queue, skb);
7246 for (i = 0; i < nr_files; i++)
7257 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7258 * causes regular reference counting to break down. We rely on the UNIX
7259 * garbage collection to take care of this problem for us.
7261 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7263 unsigned left, total;
7267 left = ctx->nr_user_files;
7269 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7271 ret = __io_sqe_files_scm(ctx, this_files, total);
7275 total += this_files;
7281 while (total < ctx->nr_user_files) {
7282 struct file *file = io_file_from_index(ctx, total);
7292 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7298 static int io_sqe_alloc_file_tables(struct fixed_rsrc_data *file_data,
7299 unsigned nr_tables, unsigned nr_files)
7303 for (i = 0; i < nr_tables; i++) {
7304 struct fixed_rsrc_table *table = &file_data->table[i];
7305 unsigned this_files;
7307 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
7308 table->files = kcalloc(this_files, sizeof(struct file *),
7312 nr_files -= this_files;
7318 for (i = 0; i < nr_tables; i++) {
7319 struct fixed_rsrc_table *table = &file_data->table[i];
7320 kfree(table->files);
7325 static void io_ring_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
7327 struct file *file = prsrc->file;
7328 #if defined(CONFIG_UNIX)
7329 struct sock *sock = ctx->ring_sock->sk;
7330 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7331 struct sk_buff *skb;
7334 __skb_queue_head_init(&list);
7337 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7338 * remove this entry and rearrange the file array.
7340 skb = skb_dequeue(head);
7342 struct scm_fp_list *fp;
7344 fp = UNIXCB(skb).fp;
7345 for (i = 0; i < fp->count; i++) {
7348 if (fp->fp[i] != file)
7351 unix_notinflight(fp->user, fp->fp[i]);
7352 left = fp->count - 1 - i;
7354 memmove(&fp->fp[i], &fp->fp[i + 1],
7355 left * sizeof(struct file *));
7362 __skb_queue_tail(&list, skb);
7372 __skb_queue_tail(&list, skb);
7374 skb = skb_dequeue(head);
7377 if (skb_peek(&list)) {
7378 spin_lock_irq(&head->lock);
7379 while ((skb = __skb_dequeue(&list)) != NULL)
7380 __skb_queue_tail(head, skb);
7381 spin_unlock_irq(&head->lock);
7388 static void __io_rsrc_put_work(struct fixed_rsrc_ref_node *ref_node)
7390 struct fixed_rsrc_data *rsrc_data = ref_node->rsrc_data;
7391 struct io_ring_ctx *ctx = rsrc_data->ctx;
7392 struct io_rsrc_put *prsrc, *tmp;
7394 list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
7395 list_del(&prsrc->list);
7396 ref_node->rsrc_put(ctx, prsrc);
7400 percpu_ref_exit(&ref_node->refs);
7402 percpu_ref_put(&rsrc_data->refs);
7405 static void io_rsrc_put_work(struct work_struct *work)
7407 struct io_ring_ctx *ctx;
7408 struct llist_node *node;
7410 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
7411 node = llist_del_all(&ctx->rsrc_put_llist);
7414 struct fixed_rsrc_ref_node *ref_node;
7415 struct llist_node *next = node->next;
7417 ref_node = llist_entry(node, struct fixed_rsrc_ref_node, llist);
7418 __io_rsrc_put_work(ref_node);
7423 static struct file **io_fixed_file_slot(struct fixed_rsrc_data *file_data,
7426 struct fixed_rsrc_table *table;
7428 table = &file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7429 return &table->files[i & IORING_FILE_TABLE_MASK];
7432 static void io_rsrc_node_ref_zero(struct percpu_ref *ref)
7434 struct fixed_rsrc_ref_node *ref_node;
7435 struct fixed_rsrc_data *data;
7436 struct io_ring_ctx *ctx;
7437 bool first_add = false;
7440 ref_node = container_of(ref, struct fixed_rsrc_ref_node, refs);
7441 data = ref_node->rsrc_data;
7444 io_rsrc_ref_lock(ctx);
7445 ref_node->done = true;
7447 while (!list_empty(&ctx->rsrc_ref_list)) {
7448 ref_node = list_first_entry(&ctx->rsrc_ref_list,
7449 struct fixed_rsrc_ref_node, node);
7450 /* recycle ref nodes in order */
7451 if (!ref_node->done)
7453 list_del(&ref_node->node);
7454 first_add |= llist_add(&ref_node->llist, &ctx->rsrc_put_llist);
7456 io_rsrc_ref_unlock(ctx);
7458 if (percpu_ref_is_dying(&data->refs))
7462 mod_delayed_work(system_wq, &ctx->rsrc_put_work, 0);
7464 queue_delayed_work(system_wq, &ctx->rsrc_put_work, delay);
7467 static struct fixed_rsrc_ref_node *alloc_fixed_rsrc_ref_node(
7468 struct io_ring_ctx *ctx)
7470 struct fixed_rsrc_ref_node *ref_node;
7472 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7476 if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
7481 INIT_LIST_HEAD(&ref_node->node);
7482 INIT_LIST_HEAD(&ref_node->rsrc_list);
7483 ref_node->done = false;
7487 static void init_fixed_file_ref_node(struct io_ring_ctx *ctx,
7488 struct fixed_rsrc_ref_node *ref_node)
7490 ref_node->rsrc_data = ctx->file_data;
7491 ref_node->rsrc_put = io_ring_file_put;
7494 static void destroy_fixed_rsrc_ref_node(struct fixed_rsrc_ref_node *ref_node)
7496 percpu_ref_exit(&ref_node->refs);
7501 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7504 __s32 __user *fds = (__s32 __user *) arg;
7505 unsigned nr_tables, i;
7507 int fd, ret = -ENOMEM;
7508 struct fixed_rsrc_ref_node *ref_node;
7509 struct fixed_rsrc_data *file_data;
7515 if (nr_args > IORING_MAX_FIXED_FILES)
7518 file_data = alloc_fixed_rsrc_data(ctx);
7521 ctx->file_data = file_data;
7523 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
7524 file_data->table = kcalloc(nr_tables, sizeof(*file_data->table),
7526 if (!file_data->table)
7529 if (io_sqe_alloc_file_tables(file_data, nr_tables, nr_args))
7532 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7533 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7537 /* allow sparse sets */
7547 * Don't allow io_uring instances to be registered. If UNIX
7548 * isn't enabled, then this causes a reference cycle and this
7549 * instance can never get freed. If UNIX is enabled we'll
7550 * handle it just fine, but there's still no point in allowing
7551 * a ring fd as it doesn't support regular read/write anyway.
7553 if (file->f_op == &io_uring_fops) {
7557 *io_fixed_file_slot(file_data, i) = file;
7560 ret = io_sqe_files_scm(ctx);
7562 io_sqe_files_unregister(ctx);
7566 ref_node = alloc_fixed_rsrc_ref_node(ctx);
7568 io_sqe_files_unregister(ctx);
7571 init_fixed_file_ref_node(ctx, ref_node);
7573 io_sqe_rsrc_set_node(ctx, file_data, ref_node);
7576 for (i = 0; i < ctx->nr_user_files; i++) {
7577 file = io_file_from_index(ctx, i);
7581 for (i = 0; i < nr_tables; i++)
7582 kfree(file_data->table[i].files);
7583 ctx->nr_user_files = 0;
7585 free_fixed_rsrc_data(ctx->file_data);
7586 ctx->file_data = NULL;
7590 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7593 #if defined(CONFIG_UNIX)
7594 struct sock *sock = ctx->ring_sock->sk;
7595 struct sk_buff_head *head = &sock->sk_receive_queue;
7596 struct sk_buff *skb;
7599 * See if we can merge this file into an existing skb SCM_RIGHTS
7600 * file set. If there's no room, fall back to allocating a new skb
7601 * and filling it in.
7603 spin_lock_irq(&head->lock);
7604 skb = skb_peek(head);
7606 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7608 if (fpl->count < SCM_MAX_FD) {
7609 __skb_unlink(skb, head);
7610 spin_unlock_irq(&head->lock);
7611 fpl->fp[fpl->count] = get_file(file);
7612 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7614 spin_lock_irq(&head->lock);
7615 __skb_queue_head(head, skb);
7620 spin_unlock_irq(&head->lock);
7627 return __io_sqe_files_scm(ctx, 1, index);
7633 static int io_queue_rsrc_removal(struct fixed_rsrc_data *data, void *rsrc)
7635 struct io_rsrc_put *prsrc;
7636 struct fixed_rsrc_ref_node *ref_node = data->node;
7638 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
7643 list_add(&prsrc->list, &ref_node->rsrc_list);
7648 static inline int io_queue_file_removal(struct fixed_rsrc_data *data,
7651 return io_queue_rsrc_removal(data, (void *)file);
7654 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7655 struct io_uring_rsrc_update *up,
7658 struct fixed_rsrc_data *data = ctx->file_data;
7659 struct fixed_rsrc_ref_node *ref_node;
7660 struct file *file, **file_slot;
7664 bool needs_switch = false;
7666 if (check_add_overflow(up->offset, nr_args, &done))
7668 if (done > ctx->nr_user_files)
7671 ref_node = alloc_fixed_rsrc_ref_node(ctx);
7674 init_fixed_file_ref_node(ctx, ref_node);
7676 fds = u64_to_user_ptr(up->data);
7677 for (done = 0; done < nr_args; done++) {
7679 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
7683 if (fd == IORING_REGISTER_FILES_SKIP)
7686 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
7687 file_slot = io_fixed_file_slot(ctx->file_data, i);
7690 err = io_queue_file_removal(data, *file_slot);
7694 needs_switch = true;
7703 * Don't allow io_uring instances to be registered. If
7704 * UNIX isn't enabled, then this causes a reference
7705 * cycle and this instance can never get freed. If UNIX
7706 * is enabled we'll handle it just fine, but there's
7707 * still no point in allowing a ring fd as it doesn't
7708 * support regular read/write anyway.
7710 if (file->f_op == &io_uring_fops) {
7716 err = io_sqe_file_register(ctx, file, i);
7726 percpu_ref_kill(&data->node->refs);
7727 io_sqe_rsrc_set_node(ctx, data, ref_node);
7729 destroy_fixed_rsrc_ref_node(ref_node);
7731 return done ? done : err;
7734 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
7737 struct io_uring_rsrc_update up;
7739 if (!ctx->file_data)
7743 if (copy_from_user(&up, arg, sizeof(up)))
7748 return __io_sqe_files_update(ctx, &up, nr_args);
7751 static struct io_wq_work *io_free_work(struct io_wq_work *work)
7753 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7755 req = io_put_req_find_next(req);
7756 return req ? &req->work : NULL;
7759 static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx)
7761 struct io_wq_hash *hash;
7762 struct io_wq_data data;
7763 unsigned int concurrency;
7765 hash = ctx->hash_map;
7767 hash = kzalloc(sizeof(*hash), GFP_KERNEL);
7769 return ERR_PTR(-ENOMEM);
7770 refcount_set(&hash->refs, 1);
7771 init_waitqueue_head(&hash->wait);
7772 ctx->hash_map = hash;
7776 data.free_work = io_free_work;
7777 data.do_work = io_wq_submit_work;
7779 /* Do QD, or 4 * CPUS, whatever is smallest */
7780 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7782 return io_wq_create(concurrency, &data);
7785 static int io_uring_alloc_task_context(struct task_struct *task,
7786 struct io_ring_ctx *ctx)
7788 struct io_uring_task *tctx;
7791 tctx = kmalloc(sizeof(*tctx), GFP_KERNEL);
7792 if (unlikely(!tctx))
7795 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
7796 if (unlikely(ret)) {
7801 tctx->io_wq = io_init_wq_offload(ctx);
7802 if (IS_ERR(tctx->io_wq)) {
7803 ret = PTR_ERR(tctx->io_wq);
7804 percpu_counter_destroy(&tctx->inflight);
7810 init_waitqueue_head(&tctx->wait);
7812 atomic_set(&tctx->in_idle, 0);
7813 tctx->sqpoll = false;
7814 task->io_uring = tctx;
7815 spin_lock_init(&tctx->task_lock);
7816 INIT_WQ_LIST(&tctx->task_list);
7817 tctx->task_state = 0;
7818 init_task_work(&tctx->task_work, tctx_task_work);
7822 void __io_uring_free(struct task_struct *tsk)
7824 struct io_uring_task *tctx = tsk->io_uring;
7826 WARN_ON_ONCE(!xa_empty(&tctx->xa));
7827 WARN_ON_ONCE(tctx->io_wq);
7829 percpu_counter_destroy(&tctx->inflight);
7831 tsk->io_uring = NULL;
7834 static int io_sq_thread_fork(struct io_sq_data *sqd, struct io_ring_ctx *ctx)
7838 clear_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7839 reinit_completion(&sqd->completion);
7841 sqd->task_pid = current->pid;
7842 current->flags |= PF_IO_WORKER;
7843 ret = io_wq_fork_thread(io_sq_thread, sqd);
7844 current->flags &= ~PF_IO_WORKER;
7849 wait_for_completion(&sqd->completion);
7850 return io_uring_alloc_task_context(sqd->thread, ctx);
7853 static int io_sq_offload_create(struct io_ring_ctx *ctx,
7854 struct io_uring_params *p)
7858 /* Retain compatibility with failing for an invalid attach attempt */
7859 if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
7860 IORING_SETUP_ATTACH_WQ) {
7863 f = fdget(p->wq_fd);
7866 if (f.file->f_op != &io_uring_fops) {
7872 if (ctx->flags & IORING_SETUP_SQPOLL) {
7873 struct io_sq_data *sqd;
7876 if (!capable(CAP_SYS_ADMIN) && !capable(CAP_SYS_NICE))
7879 sqd = io_get_sq_data(p);
7886 io_sq_thread_park(sqd);
7887 mutex_lock(&sqd->ctx_lock);
7888 list_add(&ctx->sqd_list, &sqd->ctx_new_list);
7889 mutex_unlock(&sqd->ctx_lock);
7890 io_sq_thread_unpark(sqd);
7892 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
7893 if (!ctx->sq_thread_idle)
7894 ctx->sq_thread_idle = HZ;
7899 if (p->flags & IORING_SETUP_SQ_AFF) {
7900 int cpu = p->sq_thread_cpu;
7903 if (cpu >= nr_cpu_ids)
7905 if (!cpu_online(cpu))
7913 sqd->task_pid = current->pid;
7914 current->flags |= PF_IO_WORKER;
7915 ret = io_wq_fork_thread(io_sq_thread, sqd);
7916 current->flags &= ~PF_IO_WORKER;
7921 wait_for_completion(&sqd->completion);
7922 ret = io_uring_alloc_task_context(sqd->thread, ctx);
7925 } else if (p->flags & IORING_SETUP_SQ_AFF) {
7926 /* Can't have SQ_AFF without SQPOLL */
7933 io_sq_thread_finish(ctx);
7937 static void io_sq_offload_start(struct io_ring_ctx *ctx)
7939 struct io_sq_data *sqd = ctx->sq_data;
7941 ctx->flags &= ~IORING_SETUP_R_DISABLED;
7942 if (ctx->flags & IORING_SETUP_SQPOLL)
7943 complete(&sqd->startup);
7946 static inline void __io_unaccount_mem(struct user_struct *user,
7947 unsigned long nr_pages)
7949 atomic_long_sub(nr_pages, &user->locked_vm);
7952 static inline int __io_account_mem(struct user_struct *user,
7953 unsigned long nr_pages)
7955 unsigned long page_limit, cur_pages, new_pages;
7957 /* Don't allow more pages than we can safely lock */
7958 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
7961 cur_pages = atomic_long_read(&user->locked_vm);
7962 new_pages = cur_pages + nr_pages;
7963 if (new_pages > page_limit)
7965 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
7966 new_pages) != cur_pages);
7971 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
7974 __io_unaccount_mem(ctx->user, nr_pages);
7976 if (ctx->mm_account)
7977 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
7980 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
7985 ret = __io_account_mem(ctx->user, nr_pages);
7990 if (ctx->mm_account)
7991 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
7996 static void io_mem_free(void *ptr)
8003 page = virt_to_head_page(ptr);
8004 if (put_page_testzero(page))
8005 free_compound_page(page);
8008 static void *io_mem_alloc(size_t size)
8010 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
8011 __GFP_NORETRY | __GFP_ACCOUNT;
8013 return (void *) __get_free_pages(gfp_flags, get_order(size));
8016 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8019 struct io_rings *rings;
8020 size_t off, sq_array_size;
8022 off = struct_size(rings, cqes, cq_entries);
8023 if (off == SIZE_MAX)
8027 off = ALIGN(off, SMP_CACHE_BYTES);
8035 sq_array_size = array_size(sizeof(u32), sq_entries);
8036 if (sq_array_size == SIZE_MAX)
8039 if (check_add_overflow(off, sq_array_size, &off))
8045 static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8049 if (!ctx->user_bufs)
8052 for (i = 0; i < ctx->nr_user_bufs; i++) {
8053 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8055 for (j = 0; j < imu->nr_bvecs; j++)
8056 unpin_user_page(imu->bvec[j].bv_page);
8058 if (imu->acct_pages)
8059 io_unaccount_mem(ctx, imu->acct_pages);
8064 kfree(ctx->user_bufs);
8065 ctx->user_bufs = NULL;
8066 ctx->nr_user_bufs = 0;
8070 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8071 void __user *arg, unsigned index)
8073 struct iovec __user *src;
8075 #ifdef CONFIG_COMPAT
8077 struct compat_iovec __user *ciovs;
8078 struct compat_iovec ciov;
8080 ciovs = (struct compat_iovec __user *) arg;
8081 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8084 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8085 dst->iov_len = ciov.iov_len;
8089 src = (struct iovec __user *) arg;
8090 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8096 * Not super efficient, but this is just a registration time. And we do cache
8097 * the last compound head, so generally we'll only do a full search if we don't
8100 * We check if the given compound head page has already been accounted, to
8101 * avoid double accounting it. This allows us to account the full size of the
8102 * page, not just the constituent pages of a huge page.
8104 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8105 int nr_pages, struct page *hpage)
8109 /* check current page array */
8110 for (i = 0; i < nr_pages; i++) {
8111 if (!PageCompound(pages[i]))
8113 if (compound_head(pages[i]) == hpage)
8117 /* check previously registered pages */
8118 for (i = 0; i < ctx->nr_user_bufs; i++) {
8119 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8121 for (j = 0; j < imu->nr_bvecs; j++) {
8122 if (!PageCompound(imu->bvec[j].bv_page))
8124 if (compound_head(imu->bvec[j].bv_page) == hpage)
8132 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8133 int nr_pages, struct io_mapped_ubuf *imu,
8134 struct page **last_hpage)
8138 for (i = 0; i < nr_pages; i++) {
8139 if (!PageCompound(pages[i])) {
8144 hpage = compound_head(pages[i]);
8145 if (hpage == *last_hpage)
8147 *last_hpage = hpage;
8148 if (headpage_already_acct(ctx, pages, i, hpage))
8150 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8154 if (!imu->acct_pages)
8157 ret = io_account_mem(ctx, imu->acct_pages);
8159 imu->acct_pages = 0;
8163 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
8164 struct io_mapped_ubuf *imu,
8165 struct page **last_hpage)
8167 struct vm_area_struct **vmas = NULL;
8168 struct page **pages = NULL;
8169 unsigned long off, start, end, ubuf;
8171 int ret, pret, nr_pages, i;
8173 ubuf = (unsigned long) iov->iov_base;
8174 end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8175 start = ubuf >> PAGE_SHIFT;
8176 nr_pages = end - start;
8180 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
8184 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
8189 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
8195 mmap_read_lock(current->mm);
8196 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
8198 if (pret == nr_pages) {
8199 /* don't support file backed memory */
8200 for (i = 0; i < nr_pages; i++) {
8201 struct vm_area_struct *vma = vmas[i];
8204 !is_file_hugepages(vma->vm_file)) {
8210 ret = pret < 0 ? pret : -EFAULT;
8212 mmap_read_unlock(current->mm);
8215 * if we did partial map, or found file backed vmas,
8216 * release any pages we did get
8219 unpin_user_pages(pages, pret);
8224 ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
8226 unpin_user_pages(pages, pret);
8231 off = ubuf & ~PAGE_MASK;
8232 size = iov->iov_len;
8233 for (i = 0; i < nr_pages; i++) {
8236 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8237 imu->bvec[i].bv_page = pages[i];
8238 imu->bvec[i].bv_len = vec_len;
8239 imu->bvec[i].bv_offset = off;
8243 /* store original address for later verification */
8245 imu->len = iov->iov_len;
8246 imu->nr_bvecs = nr_pages;
8254 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
8258 if (!nr_args || nr_args > UIO_MAXIOV)
8261 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
8263 if (!ctx->user_bufs)
8269 static int io_buffer_validate(struct iovec *iov)
8272 * Don't impose further limits on the size and buffer
8273 * constraints here, we'll -EINVAL later when IO is
8274 * submitted if they are wrong.
8276 if (!iov->iov_base || !iov->iov_len)
8279 /* arbitrary limit, but we need something */
8280 if (iov->iov_len > SZ_1G)
8286 static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
8287 unsigned int nr_args)
8291 struct page *last_hpage = NULL;
8293 ret = io_buffers_map_alloc(ctx, nr_args);
8297 for (i = 0; i < nr_args; i++) {
8298 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8300 ret = io_copy_iov(ctx, &iov, arg, i);
8304 ret = io_buffer_validate(&iov);
8308 ret = io_sqe_buffer_register(ctx, &iov, imu, &last_hpage);
8312 ctx->nr_user_bufs++;
8316 io_sqe_buffers_unregister(ctx);
8321 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8323 __s32 __user *fds = arg;
8329 if (copy_from_user(&fd, fds, sizeof(*fds)))
8332 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8333 if (IS_ERR(ctx->cq_ev_fd)) {
8334 int ret = PTR_ERR(ctx->cq_ev_fd);
8335 ctx->cq_ev_fd = NULL;
8342 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8344 if (ctx->cq_ev_fd) {
8345 eventfd_ctx_put(ctx->cq_ev_fd);
8346 ctx->cq_ev_fd = NULL;
8353 static int __io_destroy_buffers(int id, void *p, void *data)
8355 struct io_ring_ctx *ctx = data;
8356 struct io_buffer *buf = p;
8358 __io_remove_buffers(ctx, buf, id, -1U);
8362 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8364 idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx);
8365 idr_destroy(&ctx->io_buffer_idr);
8368 static void io_req_cache_free(struct list_head *list, struct task_struct *tsk)
8370 struct io_kiocb *req, *nxt;
8372 list_for_each_entry_safe(req, nxt, list, compl.list) {
8373 if (tsk && req->task != tsk)
8375 list_del(&req->compl.list);
8376 kmem_cache_free(req_cachep, req);
8380 static void io_req_caches_free(struct io_ring_ctx *ctx)
8382 struct io_submit_state *submit_state = &ctx->submit_state;
8383 struct io_comp_state *cs = &ctx->submit_state.comp;
8385 mutex_lock(&ctx->uring_lock);
8387 if (submit_state->free_reqs) {
8388 kmem_cache_free_bulk(req_cachep, submit_state->free_reqs,
8389 submit_state->reqs);
8390 submit_state->free_reqs = 0;
8393 spin_lock_irq(&ctx->completion_lock);
8394 list_splice_init(&cs->locked_free_list, &cs->free_list);
8395 cs->locked_free_nr = 0;
8396 spin_unlock_irq(&ctx->completion_lock);
8398 io_req_cache_free(&cs->free_list, NULL);
8400 mutex_unlock(&ctx->uring_lock);
8403 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8406 * Some may use context even when all refs and requests have been put,
8407 * and they are free to do so while still holding uring_lock, see
8408 * __io_req_task_submit(). Wait for them to finish.
8410 mutex_lock(&ctx->uring_lock);
8411 mutex_unlock(&ctx->uring_lock);
8413 io_sq_thread_finish(ctx);
8414 io_sqe_buffers_unregister(ctx);
8416 if (ctx->mm_account) {
8417 mmdrop(ctx->mm_account);
8418 ctx->mm_account = NULL;
8421 mutex_lock(&ctx->uring_lock);
8422 io_sqe_files_unregister(ctx);
8423 mutex_unlock(&ctx->uring_lock);
8424 io_eventfd_unregister(ctx);
8425 io_destroy_buffers(ctx);
8426 idr_destroy(&ctx->personality_idr);
8428 #if defined(CONFIG_UNIX)
8429 if (ctx->ring_sock) {
8430 ctx->ring_sock->file = NULL; /* so that iput() is called */
8431 sock_release(ctx->ring_sock);
8435 io_mem_free(ctx->rings);
8436 io_mem_free(ctx->sq_sqes);
8438 percpu_ref_exit(&ctx->refs);
8439 free_uid(ctx->user);
8440 io_req_caches_free(ctx);
8442 io_wq_put_hash(ctx->hash_map);
8443 kfree(ctx->cancel_hash);
8447 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8449 struct io_ring_ctx *ctx = file->private_data;
8452 poll_wait(file, &ctx->cq_wait, wait);
8454 * synchronizes with barrier from wq_has_sleeper call in
8458 if (!io_sqring_full(ctx))
8459 mask |= EPOLLOUT | EPOLLWRNORM;
8462 * Don't flush cqring overflow list here, just do a simple check.
8463 * Otherwise there could possible be ABBA deadlock:
8466 * lock(&ctx->uring_lock);
8468 * lock(&ctx->uring_lock);
8471 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
8472 * pushs them to do the flush.
8474 if (io_cqring_events(ctx) || test_bit(0, &ctx->cq_check_overflow))
8475 mask |= EPOLLIN | EPOLLRDNORM;
8480 static int io_uring_fasync(int fd, struct file *file, int on)
8482 struct io_ring_ctx *ctx = file->private_data;
8484 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8487 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8489 const struct cred *creds;
8491 creds = idr_remove(&ctx->personality_idr, id);
8500 static int io_remove_personalities(int id, void *p, void *data)
8502 struct io_ring_ctx *ctx = data;
8504 io_unregister_personality(ctx, id);
8508 static bool io_run_ctx_fallback(struct io_ring_ctx *ctx)
8510 struct callback_head *work, *next;
8511 bool executed = false;
8514 work = xchg(&ctx->exit_task_work, NULL);
8530 static void io_ring_exit_work(struct work_struct *work)
8532 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
8536 * If we're doing polled IO and end up having requests being
8537 * submitted async (out-of-line), then completions can come in while
8538 * we're waiting for refs to drop. We need to reap these manually,
8539 * as nobody else will be looking for them.
8542 io_uring_try_cancel_requests(ctx, NULL, NULL);
8543 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
8544 io_ring_ctx_free(ctx);
8547 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8549 mutex_lock(&ctx->uring_lock);
8550 percpu_ref_kill(&ctx->refs);
8551 /* if force is set, the ring is going away. always drop after that */
8552 ctx->cq_overflow_flushed = 1;
8554 __io_cqring_overflow_flush(ctx, true, NULL, NULL);
8555 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
8556 mutex_unlock(&ctx->uring_lock);
8558 io_kill_timeouts(ctx, NULL, NULL);
8559 io_poll_remove_all(ctx, NULL, NULL);
8561 /* if we failed setting up the ctx, we might not have any rings */
8562 io_iopoll_try_reap_events(ctx);
8564 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8566 * Use system_unbound_wq to avoid spawning tons of event kworkers
8567 * if we're exiting a ton of rings at the same time. It just adds
8568 * noise and overhead, there's no discernable change in runtime
8569 * over using system_wq.
8571 queue_work(system_unbound_wq, &ctx->exit_work);
8574 static int io_uring_release(struct inode *inode, struct file *file)
8576 struct io_ring_ctx *ctx = file->private_data;
8578 file->private_data = NULL;
8579 io_ring_ctx_wait_and_kill(ctx);
8583 struct io_task_cancel {
8584 struct task_struct *task;
8585 struct files_struct *files;
8588 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8590 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8591 struct io_task_cancel *cancel = data;
8594 if (cancel->files && (req->flags & REQ_F_LINK_TIMEOUT)) {
8595 unsigned long flags;
8596 struct io_ring_ctx *ctx = req->ctx;
8598 /* protect against races with linked timeouts */
8599 spin_lock_irqsave(&ctx->completion_lock, flags);
8600 ret = io_match_task(req, cancel->task, cancel->files);
8601 spin_unlock_irqrestore(&ctx->completion_lock, flags);
8603 ret = io_match_task(req, cancel->task, cancel->files);
8608 static void io_cancel_defer_files(struct io_ring_ctx *ctx,
8609 struct task_struct *task,
8610 struct files_struct *files)
8612 struct io_defer_entry *de = NULL;
8615 spin_lock_irq(&ctx->completion_lock);
8616 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8617 if (io_match_task(de->req, task, files)) {
8618 list_cut_position(&list, &ctx->defer_list, &de->list);
8622 spin_unlock_irq(&ctx->completion_lock);
8624 while (!list_empty(&list)) {
8625 de = list_first_entry(&list, struct io_defer_entry, list);
8626 list_del_init(&de->list);
8627 req_set_fail_links(de->req);
8628 io_put_req(de->req);
8629 io_req_complete(de->req, -ECANCELED);
8634 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
8635 struct task_struct *task,
8636 struct files_struct *files)
8638 struct io_task_cancel cancel = { .task = task, .files = files, };
8639 struct task_struct *tctx_task = task ?: current;
8640 struct io_uring_task *tctx = tctx_task->io_uring;
8643 enum io_wq_cancel cret;
8646 if (tctx && tctx->io_wq) {
8647 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
8649 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8652 /* SQPOLL thread does its own polling */
8653 if (!(ctx->flags & IORING_SETUP_SQPOLL) && !files) {
8654 while (!list_empty_careful(&ctx->iopoll_list)) {
8655 io_iopoll_try_reap_events(ctx);
8660 ret |= io_poll_remove_all(ctx, task, files);
8661 ret |= io_kill_timeouts(ctx, task, files);
8662 ret |= io_run_task_work();
8663 ret |= io_run_ctx_fallback(ctx);
8664 io_cqring_overflow_flush(ctx, true, task, files);
8671 static int io_uring_count_inflight(struct io_ring_ctx *ctx,
8672 struct task_struct *task,
8673 struct files_struct *files)
8675 struct io_kiocb *req;
8678 spin_lock_irq(&ctx->inflight_lock);
8679 list_for_each_entry(req, &ctx->inflight_list, inflight_entry)
8680 cnt += io_match_task(req, task, files);
8681 spin_unlock_irq(&ctx->inflight_lock);
8685 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
8686 struct task_struct *task,
8687 struct files_struct *files)
8689 while (!list_empty_careful(&ctx->inflight_list)) {
8693 inflight = io_uring_count_inflight(ctx, task, files);
8697 io_uring_try_cancel_requests(ctx, task, files);
8700 io_sq_thread_unpark(ctx->sq_data);
8701 prepare_to_wait(&task->io_uring->wait, &wait,
8702 TASK_UNINTERRUPTIBLE);
8703 if (inflight == io_uring_count_inflight(ctx, task, files))
8705 finish_wait(&task->io_uring->wait, &wait);
8707 io_sq_thread_park(ctx->sq_data);
8712 * We need to iteratively cancel requests, in case a request has dependent
8713 * hard links. These persist even for failure of cancelations, hence keep
8714 * looping until none are found.
8716 static void io_uring_cancel_task_requests(struct io_ring_ctx *ctx,
8717 struct files_struct *files)
8719 struct task_struct *task = current;
8720 bool did_park = false;
8722 if ((ctx->flags & IORING_SETUP_SQPOLL) && ctx->sq_data) {
8723 /* never started, nothing to cancel */
8724 if (ctx->flags & IORING_SETUP_R_DISABLED) {
8725 io_sq_offload_start(ctx);
8728 did_park = io_sq_thread_park(ctx->sq_data);
8730 task = ctx->sq_data->thread;
8731 atomic_inc(&task->io_uring->in_idle);
8735 io_cancel_defer_files(ctx, task, files);
8737 io_uring_cancel_files(ctx, task, files);
8739 io_uring_try_cancel_requests(ctx, task, NULL);
8742 atomic_dec(&task->io_uring->in_idle);
8743 io_sq_thread_unpark(ctx->sq_data);
8748 * Note that this task has used io_uring. We use it for cancelation purposes.
8750 static int io_uring_add_task_file(struct io_ring_ctx *ctx, struct file *file)
8752 struct io_uring_task *tctx = current->io_uring;
8755 if (unlikely(!tctx)) {
8756 ret = io_uring_alloc_task_context(current, ctx);
8759 tctx = current->io_uring;
8761 if (tctx->last != file) {
8762 void *old = xa_load(&tctx->xa, (unsigned long)file);
8766 ret = xa_err(xa_store(&tctx->xa, (unsigned long)file,
8777 * This is race safe in that the task itself is doing this, hence it
8778 * cannot be going through the exit/cancel paths at the same time.
8779 * This cannot be modified while exit/cancel is running.
8781 if (!tctx->sqpoll && (ctx->flags & IORING_SETUP_SQPOLL))
8782 tctx->sqpoll = true;
8788 * Remove this io_uring_file -> task mapping.
8790 static void io_uring_del_task_file(struct file *file)
8792 struct io_uring_task *tctx = current->io_uring;
8794 if (tctx->last == file)
8796 file = xa_erase(&tctx->xa, (unsigned long)file);
8801 static void io_uring_clean_tctx(struct io_uring_task *tctx)
8804 unsigned long index;
8806 xa_for_each(&tctx->xa, index, file)
8807 io_uring_del_task_file(file);
8809 io_wq_put_and_exit(tctx->io_wq);
8814 void __io_uring_files_cancel(struct files_struct *files)
8816 struct io_uring_task *tctx = current->io_uring;
8818 unsigned long index;
8820 /* make sure overflow events are dropped */
8821 atomic_inc(&tctx->in_idle);
8822 xa_for_each(&tctx->xa, index, file)
8823 io_uring_cancel_task_requests(file->private_data, files);
8824 atomic_dec(&tctx->in_idle);
8827 io_uring_clean_tctx(tctx);
8830 static s64 tctx_inflight(struct io_uring_task *tctx)
8832 return percpu_counter_sum(&tctx->inflight);
8835 static void io_uring_cancel_sqpoll(struct io_ring_ctx *ctx)
8837 struct io_sq_data *sqd = ctx->sq_data;
8838 struct io_uring_task *tctx;
8844 if (!io_sq_thread_park(sqd))
8846 tctx = ctx->sq_data->thread->io_uring;
8847 /* can happen on fork/alloc failure, just ignore that state */
8849 io_sq_thread_unpark(sqd);
8853 atomic_inc(&tctx->in_idle);
8855 /* read completions before cancelations */
8856 inflight = tctx_inflight(tctx);
8859 io_uring_cancel_task_requests(ctx, NULL);
8861 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
8863 * If we've seen completions, retry without waiting. This
8864 * avoids a race where a completion comes in before we did
8865 * prepare_to_wait().
8867 if (inflight == tctx_inflight(tctx))
8869 finish_wait(&tctx->wait, &wait);
8871 atomic_dec(&tctx->in_idle);
8872 io_sq_thread_unpark(sqd);
8876 * Find any io_uring fd that this task has registered or done IO on, and cancel
8879 void __io_uring_task_cancel(void)
8881 struct io_uring_task *tctx = current->io_uring;
8885 /* make sure overflow events are dropped */
8886 atomic_inc(&tctx->in_idle);
8890 unsigned long index;
8892 xa_for_each(&tctx->xa, index, file)
8893 io_uring_cancel_sqpoll(file->private_data);
8897 /* read completions before cancelations */
8898 inflight = tctx_inflight(tctx);
8901 __io_uring_files_cancel(NULL);
8903 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
8906 * If we've seen completions, retry without waiting. This
8907 * avoids a race where a completion comes in before we did
8908 * prepare_to_wait().
8910 if (inflight == tctx_inflight(tctx))
8912 finish_wait(&tctx->wait, &wait);
8915 atomic_dec(&tctx->in_idle);
8917 io_uring_clean_tctx(tctx);
8918 /* all current's requests should be gone, we can kill tctx */
8919 __io_uring_free(current);
8922 static void *io_uring_validate_mmap_request(struct file *file,
8923 loff_t pgoff, size_t sz)
8925 struct io_ring_ctx *ctx = file->private_data;
8926 loff_t offset = pgoff << PAGE_SHIFT;
8931 case IORING_OFF_SQ_RING:
8932 case IORING_OFF_CQ_RING:
8935 case IORING_OFF_SQES:
8939 return ERR_PTR(-EINVAL);
8942 page = virt_to_head_page(ptr);
8943 if (sz > page_size(page))
8944 return ERR_PTR(-EINVAL);
8951 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
8953 size_t sz = vma->vm_end - vma->vm_start;
8957 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
8959 return PTR_ERR(ptr);
8961 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
8962 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
8965 #else /* !CONFIG_MMU */
8967 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
8969 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
8972 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
8974 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
8977 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
8978 unsigned long addr, unsigned long len,
8979 unsigned long pgoff, unsigned long flags)
8983 ptr = io_uring_validate_mmap_request(file, pgoff, len);
8985 return PTR_ERR(ptr);
8987 return (unsigned long) ptr;
8990 #endif /* !CONFIG_MMU */
8992 static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
8998 if (!io_sqring_full(ctx))
9000 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9002 if (!io_sqring_full(ctx))
9005 } while (!signal_pending(current));
9007 finish_wait(&ctx->sqo_sq_wait, &wait);
9011 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9012 struct __kernel_timespec __user **ts,
9013 const sigset_t __user **sig)
9015 struct io_uring_getevents_arg arg;
9018 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9019 * is just a pointer to the sigset_t.
9021 if (!(flags & IORING_ENTER_EXT_ARG)) {
9022 *sig = (const sigset_t __user *) argp;
9028 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9029 * timespec and sigset_t pointers if good.
9031 if (*argsz != sizeof(arg))
9033 if (copy_from_user(&arg, argp, sizeof(arg)))
9035 *sig = u64_to_user_ptr(arg.sigmask);
9036 *argsz = arg.sigmask_sz;
9037 *ts = u64_to_user_ptr(arg.ts);
9041 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9042 u32, min_complete, u32, flags, const void __user *, argp,
9045 struct io_ring_ctx *ctx;
9052 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9053 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG))
9061 if (f.file->f_op != &io_uring_fops)
9065 ctx = f.file->private_data;
9066 if (!percpu_ref_tryget(&ctx->refs))
9070 if (ctx->flags & IORING_SETUP_R_DISABLED)
9074 * For SQ polling, the thread will do all submissions and completions.
9075 * Just return the requested submit count, and wake the thread if
9079 if (ctx->flags & IORING_SETUP_SQPOLL) {
9080 io_cqring_overflow_flush(ctx, false, NULL, NULL);
9082 if (unlikely(ctx->sqo_exec)) {
9083 ret = io_sq_thread_fork(ctx->sq_data, ctx);
9089 if (flags & IORING_ENTER_SQ_WAKEUP)
9090 wake_up(&ctx->sq_data->wait);
9091 if (flags & IORING_ENTER_SQ_WAIT) {
9092 ret = io_sqpoll_wait_sq(ctx);
9096 submitted = to_submit;
9097 } else if (to_submit) {
9098 ret = io_uring_add_task_file(ctx, f.file);
9101 mutex_lock(&ctx->uring_lock);
9102 submitted = io_submit_sqes(ctx, to_submit);
9103 mutex_unlock(&ctx->uring_lock);
9105 if (submitted != to_submit)
9108 if (flags & IORING_ENTER_GETEVENTS) {
9109 const sigset_t __user *sig;
9110 struct __kernel_timespec __user *ts;
9112 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
9116 min_complete = min(min_complete, ctx->cq_entries);
9119 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9120 * space applications don't need to do io completion events
9121 * polling again, they can rely on io_sq_thread to do polling
9122 * work, which can reduce cpu usage and uring_lock contention.
9124 if (ctx->flags & IORING_SETUP_IOPOLL &&
9125 !(ctx->flags & IORING_SETUP_SQPOLL)) {
9126 ret = io_iopoll_check(ctx, min_complete);
9128 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
9133 percpu_ref_put(&ctx->refs);
9136 return submitted ? submitted : ret;
9139 #ifdef CONFIG_PROC_FS
9140 static int io_uring_show_cred(int id, void *p, void *data)
9142 const struct cred *cred = p;
9143 struct seq_file *m = data;
9144 struct user_namespace *uns = seq_user_ns(m);
9145 struct group_info *gi;
9150 seq_printf(m, "%5d\n", id);
9151 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
9152 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
9153 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
9154 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
9155 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
9156 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
9157 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
9158 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
9159 seq_puts(m, "\n\tGroups:\t");
9160 gi = cred->group_info;
9161 for (g = 0; g < gi->ngroups; g++) {
9162 seq_put_decimal_ull(m, g ? " " : "",
9163 from_kgid_munged(uns, gi->gid[g]));
9165 seq_puts(m, "\n\tCapEff:\t");
9166 cap = cred->cap_effective;
9167 CAP_FOR_EACH_U32(__capi)
9168 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
9173 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
9175 struct io_sq_data *sq = NULL;
9180 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9181 * since fdinfo case grabs it in the opposite direction of normal use
9182 * cases. If we fail to get the lock, we just don't iterate any
9183 * structures that could be going away outside the io_uring mutex.
9185 has_lock = mutex_trylock(&ctx->uring_lock);
9187 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) {
9193 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9194 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9195 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9196 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9197 struct file *f = *io_fixed_file_slot(ctx->file_data, i);
9200 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9202 seq_printf(m, "%5u: <none>\n", i);
9204 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9205 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9206 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
9208 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
9209 (unsigned int) buf->len);
9211 if (has_lock && !idr_is_empty(&ctx->personality_idr)) {
9212 seq_printf(m, "Personalities:\n");
9213 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
9215 seq_printf(m, "PollList:\n");
9216 spin_lock_irq(&ctx->completion_lock);
9217 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9218 struct hlist_head *list = &ctx->cancel_hash[i];
9219 struct io_kiocb *req;
9221 hlist_for_each_entry(req, list, hash_node)
9222 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9223 req->task->task_works != NULL);
9225 spin_unlock_irq(&ctx->completion_lock);
9227 mutex_unlock(&ctx->uring_lock);
9230 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9232 struct io_ring_ctx *ctx = f->private_data;
9234 if (percpu_ref_tryget(&ctx->refs)) {
9235 __io_uring_show_fdinfo(ctx, m);
9236 percpu_ref_put(&ctx->refs);
9241 static const struct file_operations io_uring_fops = {
9242 .release = io_uring_release,
9243 .mmap = io_uring_mmap,
9245 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9246 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9248 .poll = io_uring_poll,
9249 .fasync = io_uring_fasync,
9250 #ifdef CONFIG_PROC_FS
9251 .show_fdinfo = io_uring_show_fdinfo,
9255 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9256 struct io_uring_params *p)
9258 struct io_rings *rings;
9259 size_t size, sq_array_offset;
9261 /* make sure these are sane, as we already accounted them */
9262 ctx->sq_entries = p->sq_entries;
9263 ctx->cq_entries = p->cq_entries;
9265 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9266 if (size == SIZE_MAX)
9269 rings = io_mem_alloc(size);
9274 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9275 rings->sq_ring_mask = p->sq_entries - 1;
9276 rings->cq_ring_mask = p->cq_entries - 1;
9277 rings->sq_ring_entries = p->sq_entries;
9278 rings->cq_ring_entries = p->cq_entries;
9279 ctx->sq_mask = rings->sq_ring_mask;
9280 ctx->cq_mask = rings->cq_ring_mask;
9282 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9283 if (size == SIZE_MAX) {
9284 io_mem_free(ctx->rings);
9289 ctx->sq_sqes = io_mem_alloc(size);
9290 if (!ctx->sq_sqes) {
9291 io_mem_free(ctx->rings);
9299 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
9303 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9307 ret = io_uring_add_task_file(ctx, file);
9312 fd_install(fd, file);
9317 * Allocate an anonymous fd, this is what constitutes the application
9318 * visible backing of an io_uring instance. The application mmaps this
9319 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9320 * we have to tie this fd to a socket for file garbage collection purposes.
9322 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
9325 #if defined(CONFIG_UNIX)
9328 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9331 return ERR_PTR(ret);
9334 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9335 O_RDWR | O_CLOEXEC);
9336 #if defined(CONFIG_UNIX)
9338 sock_release(ctx->ring_sock);
9339 ctx->ring_sock = NULL;
9341 ctx->ring_sock->file = file;
9347 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9348 struct io_uring_params __user *params)
9350 struct io_ring_ctx *ctx;
9356 if (entries > IORING_MAX_ENTRIES) {
9357 if (!(p->flags & IORING_SETUP_CLAMP))
9359 entries = IORING_MAX_ENTRIES;
9363 * Use twice as many entries for the CQ ring. It's possible for the
9364 * application to drive a higher depth than the size of the SQ ring,
9365 * since the sqes are only used at submission time. This allows for
9366 * some flexibility in overcommitting a bit. If the application has
9367 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9368 * of CQ ring entries manually.
9370 p->sq_entries = roundup_pow_of_two(entries);
9371 if (p->flags & IORING_SETUP_CQSIZE) {
9373 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9374 * to a power-of-two, if it isn't already. We do NOT impose
9375 * any cq vs sq ring sizing.
9379 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9380 if (!(p->flags & IORING_SETUP_CLAMP))
9382 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9384 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9385 if (p->cq_entries < p->sq_entries)
9388 p->cq_entries = 2 * p->sq_entries;
9391 ctx = io_ring_ctx_alloc(p);
9394 ctx->compat = in_compat_syscall();
9395 if (!capable(CAP_IPC_LOCK))
9396 ctx->user = get_uid(current_user());
9399 * This is just grabbed for accounting purposes. When a process exits,
9400 * the mm is exited and dropped before the files, hence we need to hang
9401 * on to this mm purely for the purposes of being able to unaccount
9402 * memory (locked/pinned vm). It's not used for anything else.
9404 mmgrab(current->mm);
9405 ctx->mm_account = current->mm;
9407 ret = io_allocate_scq_urings(ctx, p);
9411 ret = io_sq_offload_create(ctx, p);
9415 if (!(p->flags & IORING_SETUP_R_DISABLED))
9416 io_sq_offload_start(ctx);
9418 memset(&p->sq_off, 0, sizeof(p->sq_off));
9419 p->sq_off.head = offsetof(struct io_rings, sq.head);
9420 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9421 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9422 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9423 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9424 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9425 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9427 memset(&p->cq_off, 0, sizeof(p->cq_off));
9428 p->cq_off.head = offsetof(struct io_rings, cq.head);
9429 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9430 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9431 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9432 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9433 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9434 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9436 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9437 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9438 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9439 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
9440 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS;
9442 if (copy_to_user(params, p, sizeof(*p))) {
9447 file = io_uring_get_file(ctx);
9449 ret = PTR_ERR(file);
9454 * Install ring fd as the very last thing, so we don't risk someone
9455 * having closed it before we finish setup
9457 ret = io_uring_install_fd(ctx, file);
9459 /* fput will clean it up */
9464 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9467 io_ring_ctx_wait_and_kill(ctx);
9472 * Sets up an aio uring context, and returns the fd. Applications asks for a
9473 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9474 * params structure passed in.
9476 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9478 struct io_uring_params p;
9481 if (copy_from_user(&p, params, sizeof(p)))
9483 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9488 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9489 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9490 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9491 IORING_SETUP_R_DISABLED))
9494 return io_uring_create(entries, &p, params);
9497 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9498 struct io_uring_params __user *, params)
9500 return io_uring_setup(entries, params);
9503 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9505 struct io_uring_probe *p;
9509 size = struct_size(p, ops, nr_args);
9510 if (size == SIZE_MAX)
9512 p = kzalloc(size, GFP_KERNEL);
9517 if (copy_from_user(p, arg, size))
9520 if (memchr_inv(p, 0, size))
9523 p->last_op = IORING_OP_LAST - 1;
9524 if (nr_args > IORING_OP_LAST)
9525 nr_args = IORING_OP_LAST;
9527 for (i = 0; i < nr_args; i++) {
9529 if (!io_op_defs[i].not_supported)
9530 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9535 if (copy_to_user(arg, p, size))
9542 static int io_register_personality(struct io_ring_ctx *ctx)
9544 const struct cred *creds;
9547 creds = get_current_cred();
9549 ret = idr_alloc_cyclic(&ctx->personality_idr, (void *) creds, 1,
9550 USHRT_MAX, GFP_KERNEL);
9556 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9557 unsigned int nr_args)
9559 struct io_uring_restriction *res;
9563 /* Restrictions allowed only if rings started disabled */
9564 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9567 /* We allow only a single restrictions registration */
9568 if (ctx->restrictions.registered)
9571 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9574 size = array_size(nr_args, sizeof(*res));
9575 if (size == SIZE_MAX)
9578 res = memdup_user(arg, size);
9580 return PTR_ERR(res);
9584 for (i = 0; i < nr_args; i++) {
9585 switch (res[i].opcode) {
9586 case IORING_RESTRICTION_REGISTER_OP:
9587 if (res[i].register_op >= IORING_REGISTER_LAST) {
9592 __set_bit(res[i].register_op,
9593 ctx->restrictions.register_op);
9595 case IORING_RESTRICTION_SQE_OP:
9596 if (res[i].sqe_op >= IORING_OP_LAST) {
9601 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9603 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9604 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9606 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9607 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9616 /* Reset all restrictions if an error happened */
9618 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9620 ctx->restrictions.registered = true;
9626 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9628 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9631 if (ctx->restrictions.registered)
9632 ctx->restricted = 1;
9634 io_sq_offload_start(ctx);
9638 static bool io_register_op_must_quiesce(int op)
9641 case IORING_UNREGISTER_FILES:
9642 case IORING_REGISTER_FILES_UPDATE:
9643 case IORING_REGISTER_PROBE:
9644 case IORING_REGISTER_PERSONALITY:
9645 case IORING_UNREGISTER_PERSONALITY:
9652 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
9653 void __user *arg, unsigned nr_args)
9654 __releases(ctx->uring_lock)
9655 __acquires(ctx->uring_lock)
9660 * We're inside the ring mutex, if the ref is already dying, then
9661 * someone else killed the ctx or is already going through
9662 * io_uring_register().
9664 if (percpu_ref_is_dying(&ctx->refs))
9667 if (io_register_op_must_quiesce(opcode)) {
9668 percpu_ref_kill(&ctx->refs);
9671 * Drop uring mutex before waiting for references to exit. If
9672 * another thread is currently inside io_uring_enter() it might
9673 * need to grab the uring_lock to make progress. If we hold it
9674 * here across the drain wait, then we can deadlock. It's safe
9675 * to drop the mutex here, since no new references will come in
9676 * after we've killed the percpu ref.
9678 mutex_unlock(&ctx->uring_lock);
9680 ret = wait_for_completion_interruptible(&ctx->ref_comp);
9683 ret = io_run_task_work_sig();
9688 mutex_lock(&ctx->uring_lock);
9691 percpu_ref_resurrect(&ctx->refs);
9696 if (ctx->restricted) {
9697 if (opcode >= IORING_REGISTER_LAST) {
9702 if (!test_bit(opcode, ctx->restrictions.register_op)) {
9709 case IORING_REGISTER_BUFFERS:
9710 ret = io_sqe_buffers_register(ctx, arg, nr_args);
9712 case IORING_UNREGISTER_BUFFERS:
9716 ret = io_sqe_buffers_unregister(ctx);
9718 case IORING_REGISTER_FILES:
9719 ret = io_sqe_files_register(ctx, arg, nr_args);
9721 case IORING_UNREGISTER_FILES:
9725 ret = io_sqe_files_unregister(ctx);
9727 case IORING_REGISTER_FILES_UPDATE:
9728 ret = io_sqe_files_update(ctx, arg, nr_args);
9730 case IORING_REGISTER_EVENTFD:
9731 case IORING_REGISTER_EVENTFD_ASYNC:
9735 ret = io_eventfd_register(ctx, arg);
9738 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
9739 ctx->eventfd_async = 1;
9741 ctx->eventfd_async = 0;
9743 case IORING_UNREGISTER_EVENTFD:
9747 ret = io_eventfd_unregister(ctx);
9749 case IORING_REGISTER_PROBE:
9751 if (!arg || nr_args > 256)
9753 ret = io_probe(ctx, arg, nr_args);
9755 case IORING_REGISTER_PERSONALITY:
9759 ret = io_register_personality(ctx);
9761 case IORING_UNREGISTER_PERSONALITY:
9765 ret = io_unregister_personality(ctx, nr_args);
9767 case IORING_REGISTER_ENABLE_RINGS:
9771 ret = io_register_enable_rings(ctx);
9773 case IORING_REGISTER_RESTRICTIONS:
9774 ret = io_register_restrictions(ctx, arg, nr_args);
9782 if (io_register_op_must_quiesce(opcode)) {
9783 /* bring the ctx back to life */
9784 percpu_ref_reinit(&ctx->refs);
9786 reinit_completion(&ctx->ref_comp);
9791 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
9792 void __user *, arg, unsigned int, nr_args)
9794 struct io_ring_ctx *ctx;
9803 if (f.file->f_op != &io_uring_fops)
9806 ctx = f.file->private_data;
9810 mutex_lock(&ctx->uring_lock);
9811 ret = __io_uring_register(ctx, opcode, arg, nr_args);
9812 mutex_unlock(&ctx->uring_lock);
9813 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
9814 ctx->cq_ev_fd != NULL, ret);
9820 static int __init io_uring_init(void)
9822 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
9823 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
9824 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
9827 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
9828 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
9829 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
9830 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
9831 BUILD_BUG_SQE_ELEM(1, __u8, flags);
9832 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
9833 BUILD_BUG_SQE_ELEM(4, __s32, fd);
9834 BUILD_BUG_SQE_ELEM(8, __u64, off);
9835 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
9836 BUILD_BUG_SQE_ELEM(16, __u64, addr);
9837 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
9838 BUILD_BUG_SQE_ELEM(24, __u32, len);
9839 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
9840 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
9841 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
9842 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
9843 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
9844 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
9845 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
9846 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
9847 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
9848 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
9849 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
9850 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
9851 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
9852 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
9853 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
9854 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
9855 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
9856 BUILD_BUG_SQE_ELEM(42, __u16, personality);
9857 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
9859 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
9860 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
9861 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
9865 __initcall(io_uring_init);