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);
2428 static bool io_rw_reissue(struct io_kiocb *req)
2431 umode_t mode = file_inode(req->file)->i_mode;
2433 if (!S_ISBLK(mode) && !S_ISREG(mode))
2435 if ((req->flags & REQ_F_NOWAIT) || io_wq_current_is_worker())
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(&req->ctx->refs))
2445 lockdep_assert_held(&req->ctx->uring_lock);
2447 if (io_resubmit_prep(req)) {
2448 refcount_inc(&req->refs);
2449 io_queue_async_work(req);
2452 req_set_fail_links(req);
2457 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2458 unsigned int issue_flags)
2462 if ((res == -EAGAIN || res == -EOPNOTSUPP) && io_rw_reissue(req))
2464 if (res != req->result)
2465 req_set_fail_links(req);
2467 if (req->rw.kiocb.ki_flags & IOCB_WRITE)
2468 kiocb_end_write(req);
2469 if (req->flags & REQ_F_BUFFER_SELECTED)
2470 cflags = io_put_rw_kbuf(req);
2471 __io_req_complete(req, issue_flags, res, cflags);
2474 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2476 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2478 __io_complete_rw(req, res, res2, 0);
2481 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2483 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2485 if (kiocb->ki_flags & IOCB_WRITE)
2486 kiocb_end_write(req);
2488 if (res != -EAGAIN && res != req->result)
2489 req_set_fail_links(req);
2491 WRITE_ONCE(req->result, res);
2492 /* order with io_poll_complete() checking ->result */
2494 WRITE_ONCE(req->iopoll_completed, 1);
2498 * After the iocb has been issued, it's safe to be found on the poll list.
2499 * Adding the kiocb to the list AFTER submission ensures that we don't
2500 * find it from a io_iopoll_getevents() thread before the issuer is done
2501 * accessing the kiocb cookie.
2503 static void io_iopoll_req_issued(struct io_kiocb *req, bool in_async)
2505 struct io_ring_ctx *ctx = req->ctx;
2508 * Track whether we have multiple files in our lists. This will impact
2509 * how we do polling eventually, not spinning if we're on potentially
2510 * different devices.
2512 if (list_empty(&ctx->iopoll_list)) {
2513 ctx->poll_multi_file = false;
2514 } else if (!ctx->poll_multi_file) {
2515 struct io_kiocb *list_req;
2517 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2519 if (list_req->file != req->file)
2520 ctx->poll_multi_file = true;
2524 * For fast devices, IO may have already completed. If it has, add
2525 * it to the front so we find it first.
2527 if (READ_ONCE(req->iopoll_completed))
2528 list_add(&req->inflight_entry, &ctx->iopoll_list);
2530 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2533 * If IORING_SETUP_SQPOLL is enabled, sqes are either handled in sq thread
2534 * task context or in io worker task context. If current task context is
2535 * sq thread, we don't need to check whether should wake up sq thread.
2537 if (in_async && (ctx->flags & IORING_SETUP_SQPOLL) &&
2538 wq_has_sleeper(&ctx->sq_data->wait))
2539 wake_up(&ctx->sq_data->wait);
2542 static inline void io_state_file_put(struct io_submit_state *state)
2544 if (state->file_refs) {
2545 fput_many(state->file, state->file_refs);
2546 state->file_refs = 0;
2551 * Get as many references to a file as we have IOs left in this submission,
2552 * assuming most submissions are for one file, or at least that each file
2553 * has more than one submission.
2555 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2560 if (state->file_refs) {
2561 if (state->fd == fd) {
2565 io_state_file_put(state);
2567 state->file = fget_many(fd, state->ios_left);
2568 if (unlikely(!state->file))
2572 state->file_refs = state->ios_left - 1;
2576 static bool io_bdev_nowait(struct block_device *bdev)
2578 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2582 * If we tracked the file through the SCM inflight mechanism, we could support
2583 * any file. For now, just ensure that anything potentially problematic is done
2586 static bool io_file_supports_async(struct file *file, int rw)
2588 umode_t mode = file_inode(file)->i_mode;
2590 if (S_ISBLK(mode)) {
2591 if (IS_ENABLED(CONFIG_BLOCK) &&
2592 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2596 if (S_ISCHR(mode) || S_ISSOCK(mode))
2598 if (S_ISREG(mode)) {
2599 if (IS_ENABLED(CONFIG_BLOCK) &&
2600 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2601 file->f_op != &io_uring_fops)
2606 /* any ->read/write should understand O_NONBLOCK */
2607 if (file->f_flags & O_NONBLOCK)
2610 if (!(file->f_mode & FMODE_NOWAIT))
2614 return file->f_op->read_iter != NULL;
2616 return file->f_op->write_iter != NULL;
2619 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2621 struct io_ring_ctx *ctx = req->ctx;
2622 struct kiocb *kiocb = &req->rw.kiocb;
2623 struct file *file = req->file;
2627 if (S_ISREG(file_inode(file)->i_mode))
2628 req->flags |= REQ_F_ISREG;
2630 kiocb->ki_pos = READ_ONCE(sqe->off);
2631 if (kiocb->ki_pos == -1 && !(file->f_mode & FMODE_STREAM)) {
2632 req->flags |= REQ_F_CUR_POS;
2633 kiocb->ki_pos = file->f_pos;
2635 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2636 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2637 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2641 /* don't allow async punt for O_NONBLOCK or RWF_NOWAIT */
2642 if ((kiocb->ki_flags & IOCB_NOWAIT) || (file->f_flags & O_NONBLOCK))
2643 req->flags |= REQ_F_NOWAIT;
2645 ioprio = READ_ONCE(sqe->ioprio);
2647 ret = ioprio_check_cap(ioprio);
2651 kiocb->ki_ioprio = ioprio;
2653 kiocb->ki_ioprio = get_current_ioprio();
2655 if (ctx->flags & IORING_SETUP_IOPOLL) {
2656 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2657 !kiocb->ki_filp->f_op->iopoll)
2660 kiocb->ki_flags |= IOCB_HIPRI;
2661 kiocb->ki_complete = io_complete_rw_iopoll;
2662 req->iopoll_completed = 0;
2664 if (kiocb->ki_flags & IOCB_HIPRI)
2666 kiocb->ki_complete = io_complete_rw;
2669 req->rw.addr = READ_ONCE(sqe->addr);
2670 req->rw.len = READ_ONCE(sqe->len);
2671 req->buf_index = READ_ONCE(sqe->buf_index);
2675 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2681 case -ERESTARTNOINTR:
2682 case -ERESTARTNOHAND:
2683 case -ERESTART_RESTARTBLOCK:
2685 * We can't just restart the syscall, since previously
2686 * submitted sqes may already be in progress. Just fail this
2692 kiocb->ki_complete(kiocb, ret, 0);
2696 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2697 unsigned int issue_flags)
2699 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2700 struct io_async_rw *io = req->async_data;
2702 /* add previously done IO, if any */
2703 if (io && io->bytes_done > 0) {
2705 ret = io->bytes_done;
2707 ret += io->bytes_done;
2710 if (req->flags & REQ_F_CUR_POS)
2711 req->file->f_pos = kiocb->ki_pos;
2712 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2713 __io_complete_rw(req, ret, 0, issue_flags);
2715 io_rw_done(kiocb, ret);
2718 static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
2720 struct io_ring_ctx *ctx = req->ctx;
2721 size_t len = req->rw.len;
2722 struct io_mapped_ubuf *imu;
2723 u16 index, buf_index = req->buf_index;
2727 if (unlikely(buf_index >= ctx->nr_user_bufs))
2729 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2730 imu = &ctx->user_bufs[index];
2731 buf_addr = req->rw.addr;
2734 if (buf_addr + len < buf_addr)
2736 /* not inside the mapped region */
2737 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2741 * May not be a start of buffer, set size appropriately
2742 * and advance us to the beginning.
2744 offset = buf_addr - imu->ubuf;
2745 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2749 * Don't use iov_iter_advance() here, as it's really slow for
2750 * using the latter parts of a big fixed buffer - it iterates
2751 * over each segment manually. We can cheat a bit here, because
2754 * 1) it's a BVEC iter, we set it up
2755 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2756 * first and last bvec
2758 * So just find our index, and adjust the iterator afterwards.
2759 * If the offset is within the first bvec (or the whole first
2760 * bvec, just use iov_iter_advance(). This makes it easier
2761 * since we can just skip the first segment, which may not
2762 * be PAGE_SIZE aligned.
2764 const struct bio_vec *bvec = imu->bvec;
2766 if (offset <= bvec->bv_len) {
2767 iov_iter_advance(iter, offset);
2769 unsigned long seg_skip;
2771 /* skip first vec */
2772 offset -= bvec->bv_len;
2773 seg_skip = 1 + (offset >> PAGE_SHIFT);
2775 iter->bvec = bvec + seg_skip;
2776 iter->nr_segs -= seg_skip;
2777 iter->count -= bvec->bv_len + offset;
2778 iter->iov_offset = offset & ~PAGE_MASK;
2785 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2788 mutex_unlock(&ctx->uring_lock);
2791 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2794 * "Normal" inline submissions always hold the uring_lock, since we
2795 * grab it from the system call. Same is true for the SQPOLL offload.
2796 * The only exception is when we've detached the request and issue it
2797 * from an async worker thread, grab the lock for that case.
2800 mutex_lock(&ctx->uring_lock);
2803 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2804 int bgid, struct io_buffer *kbuf,
2807 struct io_buffer *head;
2809 if (req->flags & REQ_F_BUFFER_SELECTED)
2812 io_ring_submit_lock(req->ctx, needs_lock);
2814 lockdep_assert_held(&req->ctx->uring_lock);
2816 head = idr_find(&req->ctx->io_buffer_idr, bgid);
2818 if (!list_empty(&head->list)) {
2819 kbuf = list_last_entry(&head->list, struct io_buffer,
2821 list_del(&kbuf->list);
2824 idr_remove(&req->ctx->io_buffer_idr, bgid);
2826 if (*len > kbuf->len)
2829 kbuf = ERR_PTR(-ENOBUFS);
2832 io_ring_submit_unlock(req->ctx, needs_lock);
2837 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2840 struct io_buffer *kbuf;
2843 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2844 bgid = req->buf_index;
2845 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2848 req->rw.addr = (u64) (unsigned long) kbuf;
2849 req->flags |= REQ_F_BUFFER_SELECTED;
2850 return u64_to_user_ptr(kbuf->addr);
2853 #ifdef CONFIG_COMPAT
2854 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2857 struct compat_iovec __user *uiov;
2858 compat_ssize_t clen;
2862 uiov = u64_to_user_ptr(req->rw.addr);
2863 if (!access_ok(uiov, sizeof(*uiov)))
2865 if (__get_user(clen, &uiov->iov_len))
2871 buf = io_rw_buffer_select(req, &len, needs_lock);
2873 return PTR_ERR(buf);
2874 iov[0].iov_base = buf;
2875 iov[0].iov_len = (compat_size_t) len;
2880 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2883 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2887 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2890 len = iov[0].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 = len;
2901 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2904 if (req->flags & REQ_F_BUFFER_SELECTED) {
2905 struct io_buffer *kbuf;
2907 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2908 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
2909 iov[0].iov_len = kbuf->len;
2912 if (req->rw.len != 1)
2915 #ifdef CONFIG_COMPAT
2916 if (req->ctx->compat)
2917 return io_compat_import(req, iov, needs_lock);
2920 return __io_iov_buffer_select(req, iov, needs_lock);
2923 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
2924 struct iov_iter *iter, bool needs_lock)
2926 void __user *buf = u64_to_user_ptr(req->rw.addr);
2927 size_t sqe_len = req->rw.len;
2928 u8 opcode = req->opcode;
2931 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
2933 return io_import_fixed(req, rw, iter);
2936 /* buffer index only valid with fixed read/write, or buffer select */
2937 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
2940 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
2941 if (req->flags & REQ_F_BUFFER_SELECT) {
2942 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
2944 return PTR_ERR(buf);
2945 req->rw.len = sqe_len;
2948 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
2953 if (req->flags & REQ_F_BUFFER_SELECT) {
2954 ret = io_iov_buffer_select(req, *iovec, needs_lock);
2956 iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
2961 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
2965 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
2967 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
2971 * For files that don't have ->read_iter() and ->write_iter(), handle them
2972 * by looping over ->read() or ->write() manually.
2974 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
2976 struct kiocb *kiocb = &req->rw.kiocb;
2977 struct file *file = req->file;
2981 * Don't support polled IO through this interface, and we can't
2982 * support non-blocking either. For the latter, this just causes
2983 * the kiocb to be handled from an async context.
2985 if (kiocb->ki_flags & IOCB_HIPRI)
2987 if (kiocb->ki_flags & IOCB_NOWAIT)
2990 while (iov_iter_count(iter)) {
2994 if (!iov_iter_is_bvec(iter)) {
2995 iovec = iov_iter_iovec(iter);
2997 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
2998 iovec.iov_len = req->rw.len;
3002 nr = file->f_op->read(file, iovec.iov_base,
3003 iovec.iov_len, io_kiocb_ppos(kiocb));
3005 nr = file->f_op->write(file, iovec.iov_base,
3006 iovec.iov_len, io_kiocb_ppos(kiocb));
3015 if (nr != iovec.iov_len)
3019 iov_iter_advance(iter, nr);
3025 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3026 const struct iovec *fast_iov, struct iov_iter *iter)
3028 struct io_async_rw *rw = req->async_data;
3030 memcpy(&rw->iter, iter, sizeof(*iter));
3031 rw->free_iovec = iovec;
3033 /* can only be fixed buffers, no need to do anything */
3034 if (iov_iter_is_bvec(iter))
3037 unsigned iov_off = 0;
3039 rw->iter.iov = rw->fast_iov;
3040 if (iter->iov != fast_iov) {
3041 iov_off = iter->iov - fast_iov;
3042 rw->iter.iov += iov_off;
3044 if (rw->fast_iov != fast_iov)
3045 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3046 sizeof(struct iovec) * iter->nr_segs);
3048 req->flags |= REQ_F_NEED_CLEANUP;
3052 static inline int __io_alloc_async_data(struct io_kiocb *req)
3054 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3055 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3056 return req->async_data == NULL;
3059 static int io_alloc_async_data(struct io_kiocb *req)
3061 if (!io_op_defs[req->opcode].needs_async_data)
3064 return __io_alloc_async_data(req);
3067 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3068 const struct iovec *fast_iov,
3069 struct iov_iter *iter, bool force)
3071 if (!force && !io_op_defs[req->opcode].needs_async_data)
3073 if (!req->async_data) {
3074 if (__io_alloc_async_data(req)) {
3079 io_req_map_rw(req, iovec, fast_iov, iter);
3084 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3086 struct io_async_rw *iorw = req->async_data;
3087 struct iovec *iov = iorw->fast_iov;
3090 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3091 if (unlikely(ret < 0))
3094 iorw->bytes_done = 0;
3095 iorw->free_iovec = iov;
3097 req->flags |= REQ_F_NEED_CLEANUP;
3101 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3103 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3105 return io_prep_rw(req, sqe);
3109 * This is our waitqueue callback handler, registered through lock_page_async()
3110 * when we initially tried to do the IO with the iocb armed our waitqueue.
3111 * This gets called when the page is unlocked, and we generally expect that to
3112 * happen when the page IO is completed and the page is now uptodate. This will
3113 * queue a task_work based retry of the operation, attempting to copy the data
3114 * again. If the latter fails because the page was NOT uptodate, then we will
3115 * do a thread based blocking retry of the operation. That's the unexpected
3118 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3119 int sync, void *arg)
3121 struct wait_page_queue *wpq;
3122 struct io_kiocb *req = wait->private;
3123 struct wait_page_key *key = arg;
3125 wpq = container_of(wait, struct wait_page_queue, wait);
3127 if (!wake_page_match(wpq, key))
3130 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3131 list_del_init(&wait->entry);
3133 /* submit ref gets dropped, acquire a new one */
3134 refcount_inc(&req->refs);
3135 io_req_task_queue(req);
3140 * This controls whether a given IO request should be armed for async page
3141 * based retry. If we return false here, the request is handed to the async
3142 * worker threads for retry. If we're doing buffered reads on a regular file,
3143 * we prepare a private wait_page_queue entry and retry the operation. This
3144 * will either succeed because the page is now uptodate and unlocked, or it
3145 * will register a callback when the page is unlocked at IO completion. Through
3146 * that callback, io_uring uses task_work to setup a retry of the operation.
3147 * That retry will attempt the buffered read again. The retry will generally
3148 * succeed, or in rare cases where it fails, we then fall back to using the
3149 * async worker threads for a blocking retry.
3151 static bool io_rw_should_retry(struct io_kiocb *req)
3153 struct io_async_rw *rw = req->async_data;
3154 struct wait_page_queue *wait = &rw->wpq;
3155 struct kiocb *kiocb = &req->rw.kiocb;
3157 /* never retry for NOWAIT, we just complete with -EAGAIN */
3158 if (req->flags & REQ_F_NOWAIT)
3161 /* Only for buffered IO */
3162 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3166 * just use poll if we can, and don't attempt if the fs doesn't
3167 * support callback based unlocks
3169 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3172 wait->wait.func = io_async_buf_func;
3173 wait->wait.private = req;
3174 wait->wait.flags = 0;
3175 INIT_LIST_HEAD(&wait->wait.entry);
3176 kiocb->ki_flags |= IOCB_WAITQ;
3177 kiocb->ki_flags &= ~IOCB_NOWAIT;
3178 kiocb->ki_waitq = wait;
3182 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3184 if (req->file->f_op->read_iter)
3185 return call_read_iter(req->file, &req->rw.kiocb, iter);
3186 else if (req->file->f_op->read)
3187 return loop_rw_iter(READ, req, iter);
3192 static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3194 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3195 struct kiocb *kiocb = &req->rw.kiocb;
3196 struct iov_iter __iter, *iter = &__iter;
3197 struct io_async_rw *rw = req->async_data;
3198 ssize_t io_size, ret, ret2;
3199 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3205 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3209 io_size = iov_iter_count(iter);
3210 req->result = io_size;
3212 /* Ensure we clear previously set non-block flag */
3213 if (!force_nonblock)
3214 kiocb->ki_flags &= ~IOCB_NOWAIT;
3216 kiocb->ki_flags |= IOCB_NOWAIT;
3218 /* If the file doesn't support async, just async punt */
3219 if (force_nonblock && !io_file_supports_async(req->file, READ)) {
3220 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3221 return ret ?: -EAGAIN;
3224 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size);
3225 if (unlikely(ret)) {
3230 ret = io_iter_do_read(req, iter);
3232 if (ret == -EIOCBQUEUED) {
3234 } else if (ret == -EAGAIN) {
3235 /* IOPOLL retry should happen for io-wq threads */
3236 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3238 /* no retry on NONBLOCK nor RWF_NOWAIT */
3239 if (req->flags & REQ_F_NOWAIT)
3241 /* some cases will consume bytes even on error returns */
3242 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3244 } else if (ret <= 0 || ret == io_size || !force_nonblock ||
3245 (req->flags & REQ_F_NOWAIT) || !(req->flags & REQ_F_ISREG)) {
3246 /* read all, failed, already did sync or don't want to retry */
3250 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3255 rw = req->async_data;
3256 /* now use our persistent iterator, if we aren't already */
3261 rw->bytes_done += ret;
3262 /* if we can retry, do so with the callbacks armed */
3263 if (!io_rw_should_retry(req)) {
3264 kiocb->ki_flags &= ~IOCB_WAITQ;
3269 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3270 * we get -EIOCBQUEUED, then we'll get a notification when the
3271 * desired page gets unlocked. We can also get a partial read
3272 * here, and if we do, then just retry at the new offset.
3274 ret = io_iter_do_read(req, iter);
3275 if (ret == -EIOCBQUEUED)
3277 /* we got some bytes, but not all. retry. */
3278 } while (ret > 0 && ret < io_size);
3280 kiocb_done(kiocb, ret, issue_flags);
3282 /* it's faster to check here then delegate to kfree */
3288 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3290 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3292 return io_prep_rw(req, sqe);
3295 static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3297 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3298 struct kiocb *kiocb = &req->rw.kiocb;
3299 struct iov_iter __iter, *iter = &__iter;
3300 struct io_async_rw *rw = req->async_data;
3301 ssize_t ret, ret2, io_size;
3302 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3308 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3312 io_size = iov_iter_count(iter);
3313 req->result = io_size;
3315 /* Ensure we clear previously set non-block flag */
3316 if (!force_nonblock)
3317 kiocb->ki_flags &= ~IOCB_NOWAIT;
3319 kiocb->ki_flags |= IOCB_NOWAIT;
3321 /* If the file doesn't support async, just async punt */
3322 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
3325 /* file path doesn't support NOWAIT for non-direct_IO */
3326 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3327 (req->flags & REQ_F_ISREG))
3330 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size);
3335 * Open-code file_start_write here to grab freeze protection,
3336 * which will be released by another thread in
3337 * io_complete_rw(). Fool lockdep by telling it the lock got
3338 * released so that it doesn't complain about the held lock when
3339 * we return to userspace.
3341 if (req->flags & REQ_F_ISREG) {
3342 sb_start_write(file_inode(req->file)->i_sb);
3343 __sb_writers_release(file_inode(req->file)->i_sb,
3346 kiocb->ki_flags |= IOCB_WRITE;
3348 if (req->file->f_op->write_iter)
3349 ret2 = call_write_iter(req->file, kiocb, iter);
3350 else if (req->file->f_op->write)
3351 ret2 = loop_rw_iter(WRITE, req, iter);
3356 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3357 * retry them without IOCB_NOWAIT.
3359 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3361 /* no retry on NONBLOCK nor RWF_NOWAIT */
3362 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3364 if (!force_nonblock || ret2 != -EAGAIN) {
3365 /* IOPOLL retry should happen for io-wq threads */
3366 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3369 kiocb_done(kiocb, ret2, issue_flags);
3372 /* some cases will consume bytes even on error returns */
3373 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3374 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3375 return ret ?: -EAGAIN;
3378 /* it's reportedly faster than delegating the null check to kfree() */
3384 static int io_renameat_prep(struct io_kiocb *req,
3385 const struct io_uring_sqe *sqe)
3387 struct io_rename *ren = &req->rename;
3388 const char __user *oldf, *newf;
3390 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3393 ren->old_dfd = READ_ONCE(sqe->fd);
3394 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3395 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3396 ren->new_dfd = READ_ONCE(sqe->len);
3397 ren->flags = READ_ONCE(sqe->rename_flags);
3399 ren->oldpath = getname(oldf);
3400 if (IS_ERR(ren->oldpath))
3401 return PTR_ERR(ren->oldpath);
3403 ren->newpath = getname(newf);
3404 if (IS_ERR(ren->newpath)) {
3405 putname(ren->oldpath);
3406 return PTR_ERR(ren->newpath);
3409 req->flags |= REQ_F_NEED_CLEANUP;
3413 static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3415 struct io_rename *ren = &req->rename;
3418 if (issue_flags & IO_URING_F_NONBLOCK)
3421 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3422 ren->newpath, ren->flags);
3424 req->flags &= ~REQ_F_NEED_CLEANUP;
3426 req_set_fail_links(req);
3427 io_req_complete(req, ret);
3431 static int io_unlinkat_prep(struct io_kiocb *req,
3432 const struct io_uring_sqe *sqe)
3434 struct io_unlink *un = &req->unlink;
3435 const char __user *fname;
3437 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3440 un->dfd = READ_ONCE(sqe->fd);
3442 un->flags = READ_ONCE(sqe->unlink_flags);
3443 if (un->flags & ~AT_REMOVEDIR)
3446 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3447 un->filename = getname(fname);
3448 if (IS_ERR(un->filename))
3449 return PTR_ERR(un->filename);
3451 req->flags |= REQ_F_NEED_CLEANUP;
3455 static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3457 struct io_unlink *un = &req->unlink;
3460 if (issue_flags & IO_URING_F_NONBLOCK)
3463 if (un->flags & AT_REMOVEDIR)
3464 ret = do_rmdir(un->dfd, un->filename);
3466 ret = do_unlinkat(un->dfd, un->filename);
3468 req->flags &= ~REQ_F_NEED_CLEANUP;
3470 req_set_fail_links(req);
3471 io_req_complete(req, ret);
3475 static int io_shutdown_prep(struct io_kiocb *req,
3476 const struct io_uring_sqe *sqe)
3478 #if defined(CONFIG_NET)
3479 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3481 if (sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3485 req->shutdown.how = READ_ONCE(sqe->len);
3492 static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3494 #if defined(CONFIG_NET)
3495 struct socket *sock;
3498 if (issue_flags & IO_URING_F_NONBLOCK)
3501 sock = sock_from_file(req->file);
3502 if (unlikely(!sock))
3505 ret = __sys_shutdown_sock(sock, req->shutdown.how);
3507 req_set_fail_links(req);
3508 io_req_complete(req, ret);
3515 static int __io_splice_prep(struct io_kiocb *req,
3516 const struct io_uring_sqe *sqe)
3518 struct io_splice* sp = &req->splice;
3519 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3521 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3525 sp->len = READ_ONCE(sqe->len);
3526 sp->flags = READ_ONCE(sqe->splice_flags);
3528 if (unlikely(sp->flags & ~valid_flags))
3531 sp->file_in = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in),
3532 (sp->flags & SPLICE_F_FD_IN_FIXED));
3535 req->flags |= REQ_F_NEED_CLEANUP;
3537 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
3539 * Splice operation will be punted aync, and here need to
3540 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3542 req->work.flags |= IO_WQ_WORK_UNBOUND;
3548 static int io_tee_prep(struct io_kiocb *req,
3549 const struct io_uring_sqe *sqe)
3551 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3553 return __io_splice_prep(req, sqe);
3556 static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
3558 struct io_splice *sp = &req->splice;
3559 struct file *in = sp->file_in;
3560 struct file *out = sp->file_out;
3561 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3564 if (issue_flags & IO_URING_F_NONBLOCK)
3567 ret = do_tee(in, out, sp->len, flags);
3569 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3570 req->flags &= ~REQ_F_NEED_CLEANUP;
3573 req_set_fail_links(req);
3574 io_req_complete(req, ret);
3578 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3580 struct io_splice* sp = &req->splice;
3582 sp->off_in = READ_ONCE(sqe->splice_off_in);
3583 sp->off_out = READ_ONCE(sqe->off);
3584 return __io_splice_prep(req, sqe);
3587 static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
3589 struct io_splice *sp = &req->splice;
3590 struct file *in = sp->file_in;
3591 struct file *out = sp->file_out;
3592 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3593 loff_t *poff_in, *poff_out;
3596 if (issue_flags & IO_URING_F_NONBLOCK)
3599 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3600 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3603 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3605 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3606 req->flags &= ~REQ_F_NEED_CLEANUP;
3609 req_set_fail_links(req);
3610 io_req_complete(req, ret);
3615 * IORING_OP_NOP just posts a completion event, nothing else.
3617 static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
3619 struct io_ring_ctx *ctx = req->ctx;
3621 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3624 __io_req_complete(req, issue_flags, 0, 0);
3628 static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3630 struct io_ring_ctx *ctx = req->ctx;
3635 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3637 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3640 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3641 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3644 req->sync.off = READ_ONCE(sqe->off);
3645 req->sync.len = READ_ONCE(sqe->len);
3649 static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
3651 loff_t end = req->sync.off + req->sync.len;
3654 /* fsync always requires a blocking context */
3655 if (issue_flags & IO_URING_F_NONBLOCK)
3658 ret = vfs_fsync_range(req->file, req->sync.off,
3659 end > 0 ? end : LLONG_MAX,
3660 req->sync.flags & IORING_FSYNC_DATASYNC);
3662 req_set_fail_links(req);
3663 io_req_complete(req, ret);
3667 static int io_fallocate_prep(struct io_kiocb *req,
3668 const struct io_uring_sqe *sqe)
3670 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3672 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3675 req->sync.off = READ_ONCE(sqe->off);
3676 req->sync.len = READ_ONCE(sqe->addr);
3677 req->sync.mode = READ_ONCE(sqe->len);
3681 static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
3685 /* fallocate always requiring blocking context */
3686 if (issue_flags & IO_URING_F_NONBLOCK)
3688 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3691 req_set_fail_links(req);
3692 io_req_complete(req, ret);
3696 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3698 const char __user *fname;
3701 if (unlikely(sqe->ioprio || sqe->buf_index))
3703 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3706 /* open.how should be already initialised */
3707 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3708 req->open.how.flags |= O_LARGEFILE;
3710 req->open.dfd = READ_ONCE(sqe->fd);
3711 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3712 req->open.filename = getname(fname);
3713 if (IS_ERR(req->open.filename)) {
3714 ret = PTR_ERR(req->open.filename);
3715 req->open.filename = NULL;
3718 req->open.nofile = rlimit(RLIMIT_NOFILE);
3719 req->flags |= REQ_F_NEED_CLEANUP;
3723 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3727 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3729 mode = READ_ONCE(sqe->len);
3730 flags = READ_ONCE(sqe->open_flags);
3731 req->open.how = build_open_how(flags, mode);
3732 return __io_openat_prep(req, sqe);
3735 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3737 struct open_how __user *how;
3741 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3743 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3744 len = READ_ONCE(sqe->len);
3745 if (len < OPEN_HOW_SIZE_VER0)
3748 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3753 return __io_openat_prep(req, sqe);
3756 static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
3758 struct open_flags op;
3761 bool resolve_nonblock;
3764 ret = build_open_flags(&req->open.how, &op);
3767 nonblock_set = op.open_flag & O_NONBLOCK;
3768 resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
3769 if (issue_flags & IO_URING_F_NONBLOCK) {
3771 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
3772 * it'll always -EAGAIN
3774 if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
3776 op.lookup_flags |= LOOKUP_CACHED;
3777 op.open_flag |= O_NONBLOCK;
3780 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3784 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3785 /* only retry if RESOLVE_CACHED wasn't already set by application */
3786 if ((!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)) &&
3787 file == ERR_PTR(-EAGAIN)) {
3789 * We could hang on to this 'fd', but seems like marginal
3790 * gain for something that is now known to be a slower path.
3791 * So just put it, and we'll get a new one when we retry.
3799 ret = PTR_ERR(file);
3801 if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
3802 file->f_flags &= ~O_NONBLOCK;
3803 fsnotify_open(file);
3804 fd_install(ret, file);
3807 putname(req->open.filename);
3808 req->flags &= ~REQ_F_NEED_CLEANUP;
3810 req_set_fail_links(req);
3811 io_req_complete(req, ret);
3815 static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
3817 return io_openat2(req, issue_flags & IO_URING_F_NONBLOCK);
3820 static int io_remove_buffers_prep(struct io_kiocb *req,
3821 const struct io_uring_sqe *sqe)
3823 struct io_provide_buf *p = &req->pbuf;
3826 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3829 tmp = READ_ONCE(sqe->fd);
3830 if (!tmp || tmp > USHRT_MAX)
3833 memset(p, 0, sizeof(*p));
3835 p->bgid = READ_ONCE(sqe->buf_group);
3839 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3840 int bgid, unsigned nbufs)
3844 /* shouldn't happen */
3848 /* the head kbuf is the list itself */
3849 while (!list_empty(&buf->list)) {
3850 struct io_buffer *nxt;
3852 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3853 list_del(&nxt->list);
3860 idr_remove(&ctx->io_buffer_idr, bgid);
3865 static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
3867 struct io_provide_buf *p = &req->pbuf;
3868 struct io_ring_ctx *ctx = req->ctx;
3869 struct io_buffer *head;
3871 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3873 io_ring_submit_lock(ctx, !force_nonblock);
3875 lockdep_assert_held(&ctx->uring_lock);
3878 head = idr_find(&ctx->io_buffer_idr, p->bgid);
3880 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3882 req_set_fail_links(req);
3884 /* need to hold the lock to complete IOPOLL requests */
3885 if (ctx->flags & IORING_SETUP_IOPOLL) {
3886 __io_req_complete(req, issue_flags, ret, 0);
3887 io_ring_submit_unlock(ctx, !force_nonblock);
3889 io_ring_submit_unlock(ctx, !force_nonblock);
3890 __io_req_complete(req, issue_flags, ret, 0);
3895 static int io_provide_buffers_prep(struct io_kiocb *req,
3896 const struct io_uring_sqe *sqe)
3898 struct io_provide_buf *p = &req->pbuf;
3901 if (sqe->ioprio || sqe->rw_flags)
3904 tmp = READ_ONCE(sqe->fd);
3905 if (!tmp || tmp > USHRT_MAX)
3908 p->addr = READ_ONCE(sqe->addr);
3909 p->len = READ_ONCE(sqe->len);
3911 if (!access_ok(u64_to_user_ptr(p->addr), (p->len * p->nbufs)))
3914 p->bgid = READ_ONCE(sqe->buf_group);
3915 tmp = READ_ONCE(sqe->off);
3916 if (tmp > USHRT_MAX)
3922 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3924 struct io_buffer *buf;
3925 u64 addr = pbuf->addr;
3926 int i, bid = pbuf->bid;
3928 for (i = 0; i < pbuf->nbufs; i++) {
3929 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
3934 buf->len = pbuf->len;
3939 INIT_LIST_HEAD(&buf->list);
3942 list_add_tail(&buf->list, &(*head)->list);
3946 return i ? i : -ENOMEM;
3949 static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
3951 struct io_provide_buf *p = &req->pbuf;
3952 struct io_ring_ctx *ctx = req->ctx;
3953 struct io_buffer *head, *list;
3955 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3957 io_ring_submit_lock(ctx, !force_nonblock);
3959 lockdep_assert_held(&ctx->uring_lock);
3961 list = head = idr_find(&ctx->io_buffer_idr, p->bgid);
3963 ret = io_add_buffers(p, &head);
3968 ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1,
3971 __io_remove_buffers(ctx, head, p->bgid, -1U);
3977 req_set_fail_links(req);
3979 /* need to hold the lock to complete IOPOLL requests */
3980 if (ctx->flags & IORING_SETUP_IOPOLL) {
3981 __io_req_complete(req, issue_flags, ret, 0);
3982 io_ring_submit_unlock(ctx, !force_nonblock);
3984 io_ring_submit_unlock(ctx, !force_nonblock);
3985 __io_req_complete(req, issue_flags, ret, 0);
3990 static int io_epoll_ctl_prep(struct io_kiocb *req,
3991 const struct io_uring_sqe *sqe)
3993 #if defined(CONFIG_EPOLL)
3994 if (sqe->ioprio || sqe->buf_index)
3996 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
3999 req->epoll.epfd = READ_ONCE(sqe->fd);
4000 req->epoll.op = READ_ONCE(sqe->len);
4001 req->epoll.fd = READ_ONCE(sqe->off);
4003 if (ep_op_has_event(req->epoll.op)) {
4004 struct epoll_event __user *ev;
4006 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4007 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4017 static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4019 #if defined(CONFIG_EPOLL)
4020 struct io_epoll *ie = &req->epoll;
4022 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4024 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4025 if (force_nonblock && ret == -EAGAIN)
4029 req_set_fail_links(req);
4030 __io_req_complete(req, issue_flags, ret, 0);
4037 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4039 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4040 if (sqe->ioprio || sqe->buf_index || sqe->off)
4042 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4045 req->madvise.addr = READ_ONCE(sqe->addr);
4046 req->madvise.len = READ_ONCE(sqe->len);
4047 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4054 static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4056 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4057 struct io_madvise *ma = &req->madvise;
4060 if (issue_flags & IO_URING_F_NONBLOCK)
4063 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4065 req_set_fail_links(req);
4066 io_req_complete(req, ret);
4073 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4075 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4077 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4080 req->fadvise.offset = READ_ONCE(sqe->off);
4081 req->fadvise.len = READ_ONCE(sqe->len);
4082 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4086 static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4088 struct io_fadvise *fa = &req->fadvise;
4091 if (issue_flags & IO_URING_F_NONBLOCK) {
4092 switch (fa->advice) {
4093 case POSIX_FADV_NORMAL:
4094 case POSIX_FADV_RANDOM:
4095 case POSIX_FADV_SEQUENTIAL:
4102 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4104 req_set_fail_links(req);
4105 io_req_complete(req, ret);
4109 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4111 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4113 if (sqe->ioprio || sqe->buf_index)
4115 if (req->flags & REQ_F_FIXED_FILE)
4118 req->statx.dfd = READ_ONCE(sqe->fd);
4119 req->statx.mask = READ_ONCE(sqe->len);
4120 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4121 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4122 req->statx.flags = READ_ONCE(sqe->statx_flags);
4127 static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4129 struct io_statx *ctx = &req->statx;
4132 if (issue_flags & IO_URING_F_NONBLOCK) {
4133 /* only need file table for an actual valid fd */
4134 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
4135 req->flags |= REQ_F_NO_FILE_TABLE;
4139 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4143 req_set_fail_links(req);
4144 io_req_complete(req, ret);
4148 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4150 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4152 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4153 sqe->rw_flags || sqe->buf_index)
4155 if (req->flags & REQ_F_FIXED_FILE)
4158 req->close.fd = READ_ONCE(sqe->fd);
4162 static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4164 struct files_struct *files = current->files;
4165 struct io_close *close = &req->close;
4166 struct fdtable *fdt;
4172 spin_lock(&files->file_lock);
4173 fdt = files_fdtable(files);
4174 if (close->fd >= fdt->max_fds) {
4175 spin_unlock(&files->file_lock);
4178 file = fdt->fd[close->fd];
4180 spin_unlock(&files->file_lock);
4184 if (file->f_op == &io_uring_fops) {
4185 spin_unlock(&files->file_lock);
4190 /* if the file has a flush method, be safe and punt to async */
4191 if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4192 spin_unlock(&files->file_lock);
4196 ret = __close_fd_get_file(close->fd, &file);
4197 spin_unlock(&files->file_lock);
4204 /* No ->flush() or already async, safely close from here */
4205 ret = filp_close(file, current->files);
4208 req_set_fail_links(req);
4211 __io_req_complete(req, issue_flags, ret, 0);
4215 static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4217 struct io_ring_ctx *ctx = req->ctx;
4219 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4221 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4224 req->sync.off = READ_ONCE(sqe->off);
4225 req->sync.len = READ_ONCE(sqe->len);
4226 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4230 static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4234 /* sync_file_range always requires a blocking context */
4235 if (issue_flags & IO_URING_F_NONBLOCK)
4238 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4241 req_set_fail_links(req);
4242 io_req_complete(req, ret);
4246 #if defined(CONFIG_NET)
4247 static int io_setup_async_msg(struct io_kiocb *req,
4248 struct io_async_msghdr *kmsg)
4250 struct io_async_msghdr *async_msg = req->async_data;
4254 if (io_alloc_async_data(req)) {
4255 kfree(kmsg->free_iov);
4258 async_msg = req->async_data;
4259 req->flags |= REQ_F_NEED_CLEANUP;
4260 memcpy(async_msg, kmsg, sizeof(*kmsg));
4261 async_msg->msg.msg_name = &async_msg->addr;
4262 /* if were using fast_iov, set it to the new one */
4263 if (!async_msg->free_iov)
4264 async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4269 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4270 struct io_async_msghdr *iomsg)
4272 iomsg->msg.msg_name = &iomsg->addr;
4273 iomsg->free_iov = iomsg->fast_iov;
4274 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4275 req->sr_msg.msg_flags, &iomsg->free_iov);
4278 static int io_sendmsg_prep_async(struct io_kiocb *req)
4282 if (!io_op_defs[req->opcode].needs_async_data)
4284 ret = io_sendmsg_copy_hdr(req, req->async_data);
4286 req->flags |= REQ_F_NEED_CLEANUP;
4290 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4292 struct io_sr_msg *sr = &req->sr_msg;
4294 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4297 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4298 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4299 sr->len = READ_ONCE(sqe->len);
4301 #ifdef CONFIG_COMPAT
4302 if (req->ctx->compat)
4303 sr->msg_flags |= MSG_CMSG_COMPAT;
4308 static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4310 struct io_async_msghdr iomsg, *kmsg;
4311 struct socket *sock;
4315 sock = sock_from_file(req->file);
4316 if (unlikely(!sock))
4319 kmsg = req->async_data;
4321 ret = io_sendmsg_copy_hdr(req, &iomsg);
4327 flags = req->sr_msg.msg_flags;
4328 if (flags & MSG_DONTWAIT)
4329 req->flags |= REQ_F_NOWAIT;
4330 else if (issue_flags & IO_URING_F_NONBLOCK)
4331 flags |= MSG_DONTWAIT;
4333 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4334 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4335 return io_setup_async_msg(req, kmsg);
4336 if (ret == -ERESTARTSYS)
4339 /* fast path, check for non-NULL to avoid function call */
4341 kfree(kmsg->free_iov);
4342 req->flags &= ~REQ_F_NEED_CLEANUP;
4344 req_set_fail_links(req);
4345 __io_req_complete(req, issue_flags, ret, 0);
4349 static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4351 struct io_sr_msg *sr = &req->sr_msg;
4354 struct socket *sock;
4358 sock = sock_from_file(req->file);
4359 if (unlikely(!sock))
4362 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4366 msg.msg_name = NULL;
4367 msg.msg_control = NULL;
4368 msg.msg_controllen = 0;
4369 msg.msg_namelen = 0;
4371 flags = req->sr_msg.msg_flags;
4372 if (flags & MSG_DONTWAIT)
4373 req->flags |= REQ_F_NOWAIT;
4374 else if (issue_flags & IO_URING_F_NONBLOCK)
4375 flags |= MSG_DONTWAIT;
4377 msg.msg_flags = flags;
4378 ret = sock_sendmsg(sock, &msg);
4379 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4381 if (ret == -ERESTARTSYS)
4385 req_set_fail_links(req);
4386 __io_req_complete(req, issue_flags, ret, 0);
4390 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4391 struct io_async_msghdr *iomsg)
4393 struct io_sr_msg *sr = &req->sr_msg;
4394 struct iovec __user *uiov;
4398 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4399 &iomsg->uaddr, &uiov, &iov_len);
4403 if (req->flags & REQ_F_BUFFER_SELECT) {
4406 if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4408 sr->len = iomsg->fast_iov[0].iov_len;
4409 iomsg->free_iov = NULL;
4411 iomsg->free_iov = iomsg->fast_iov;
4412 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4413 &iomsg->free_iov, &iomsg->msg.msg_iter,
4422 #ifdef CONFIG_COMPAT
4423 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4424 struct io_async_msghdr *iomsg)
4426 struct compat_msghdr __user *msg_compat;
4427 struct io_sr_msg *sr = &req->sr_msg;
4428 struct compat_iovec __user *uiov;
4433 msg_compat = (struct compat_msghdr __user *) sr->umsg;
4434 ret = __get_compat_msghdr(&iomsg->msg, msg_compat, &iomsg->uaddr,
4439 uiov = compat_ptr(ptr);
4440 if (req->flags & REQ_F_BUFFER_SELECT) {
4441 compat_ssize_t clen;
4445 if (!access_ok(uiov, sizeof(*uiov)))
4447 if (__get_user(clen, &uiov->iov_len))
4452 iomsg->free_iov = NULL;
4454 iomsg->free_iov = iomsg->fast_iov;
4455 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4456 UIO_FASTIOV, &iomsg->free_iov,
4457 &iomsg->msg.msg_iter, true);
4466 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4467 struct io_async_msghdr *iomsg)
4469 iomsg->msg.msg_name = &iomsg->addr;
4471 #ifdef CONFIG_COMPAT
4472 if (req->ctx->compat)
4473 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4476 return __io_recvmsg_copy_hdr(req, iomsg);
4479 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4482 struct io_sr_msg *sr = &req->sr_msg;
4483 struct io_buffer *kbuf;
4485 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4490 req->flags |= REQ_F_BUFFER_SELECTED;
4494 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4496 return io_put_kbuf(req, req->sr_msg.kbuf);
4499 static int io_recvmsg_prep_async(struct io_kiocb *req)
4503 if (!io_op_defs[req->opcode].needs_async_data)
4505 ret = io_recvmsg_copy_hdr(req, req->async_data);
4507 req->flags |= REQ_F_NEED_CLEANUP;
4511 static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4513 struct io_sr_msg *sr = &req->sr_msg;
4515 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4518 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4519 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4520 sr->len = READ_ONCE(sqe->len);
4521 sr->bgid = READ_ONCE(sqe->buf_group);
4523 #ifdef CONFIG_COMPAT
4524 if (req->ctx->compat)
4525 sr->msg_flags |= MSG_CMSG_COMPAT;
4530 static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
4532 struct io_async_msghdr iomsg, *kmsg;
4533 struct socket *sock;
4534 struct io_buffer *kbuf;
4536 int ret, cflags = 0;
4537 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4539 sock = sock_from_file(req->file);
4540 if (unlikely(!sock))
4543 kmsg = req->async_data;
4545 ret = io_recvmsg_copy_hdr(req, &iomsg);
4551 if (req->flags & REQ_F_BUFFER_SELECT) {
4552 kbuf = io_recv_buffer_select(req, !force_nonblock);
4554 return PTR_ERR(kbuf);
4555 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4556 kmsg->fast_iov[0].iov_len = req->sr_msg.len;
4557 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
4558 1, req->sr_msg.len);
4561 flags = req->sr_msg.msg_flags;
4562 if (flags & MSG_DONTWAIT)
4563 req->flags |= REQ_F_NOWAIT;
4564 else if (force_nonblock)
4565 flags |= MSG_DONTWAIT;
4567 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4568 kmsg->uaddr, flags);
4569 if (force_nonblock && ret == -EAGAIN)
4570 return io_setup_async_msg(req, kmsg);
4571 if (ret == -ERESTARTSYS)
4574 if (req->flags & REQ_F_BUFFER_SELECTED)
4575 cflags = io_put_recv_kbuf(req);
4576 /* fast path, check for non-NULL to avoid function call */
4578 kfree(kmsg->free_iov);
4579 req->flags &= ~REQ_F_NEED_CLEANUP;
4581 req_set_fail_links(req);
4582 __io_req_complete(req, issue_flags, ret, cflags);
4586 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
4588 struct io_buffer *kbuf;
4589 struct io_sr_msg *sr = &req->sr_msg;
4591 void __user *buf = sr->buf;
4592 struct socket *sock;
4595 int ret, cflags = 0;
4596 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4598 sock = sock_from_file(req->file);
4599 if (unlikely(!sock))
4602 if (req->flags & REQ_F_BUFFER_SELECT) {
4603 kbuf = io_recv_buffer_select(req, !force_nonblock);
4605 return PTR_ERR(kbuf);
4606 buf = u64_to_user_ptr(kbuf->addr);
4609 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4613 msg.msg_name = NULL;
4614 msg.msg_control = NULL;
4615 msg.msg_controllen = 0;
4616 msg.msg_namelen = 0;
4617 msg.msg_iocb = NULL;
4620 flags = req->sr_msg.msg_flags;
4621 if (flags & MSG_DONTWAIT)
4622 req->flags |= REQ_F_NOWAIT;
4623 else if (force_nonblock)
4624 flags |= MSG_DONTWAIT;
4626 ret = sock_recvmsg(sock, &msg, flags);
4627 if (force_nonblock && ret == -EAGAIN)
4629 if (ret == -ERESTARTSYS)
4632 if (req->flags & REQ_F_BUFFER_SELECTED)
4633 cflags = io_put_recv_kbuf(req);
4635 req_set_fail_links(req);
4636 __io_req_complete(req, issue_flags, ret, cflags);
4640 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4642 struct io_accept *accept = &req->accept;
4644 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4646 if (sqe->ioprio || sqe->len || sqe->buf_index)
4649 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4650 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4651 accept->flags = READ_ONCE(sqe->accept_flags);
4652 accept->nofile = rlimit(RLIMIT_NOFILE);
4656 static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
4658 struct io_accept *accept = &req->accept;
4659 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4660 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4663 if (req->file->f_flags & O_NONBLOCK)
4664 req->flags |= REQ_F_NOWAIT;
4666 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4667 accept->addr_len, accept->flags,
4669 if (ret == -EAGAIN && force_nonblock)
4672 if (ret == -ERESTARTSYS)
4674 req_set_fail_links(req);
4676 __io_req_complete(req, issue_flags, ret, 0);
4680 static int io_connect_prep_async(struct io_kiocb *req)
4682 struct io_async_connect *io = req->async_data;
4683 struct io_connect *conn = &req->connect;
4685 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
4688 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4690 struct io_connect *conn = &req->connect;
4692 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4694 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4697 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4698 conn->addr_len = READ_ONCE(sqe->addr2);
4702 static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
4704 struct io_async_connect __io, *io;
4705 unsigned file_flags;
4707 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4709 if (req->async_data) {
4710 io = req->async_data;
4712 ret = move_addr_to_kernel(req->connect.addr,
4713 req->connect.addr_len,
4720 file_flags = force_nonblock ? O_NONBLOCK : 0;
4722 ret = __sys_connect_file(req->file, &io->address,
4723 req->connect.addr_len, file_flags);
4724 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4725 if (req->async_data)
4727 if (io_alloc_async_data(req)) {
4731 io = req->async_data;
4732 memcpy(req->async_data, &__io, sizeof(__io));
4735 if (ret == -ERESTARTSYS)
4739 req_set_fail_links(req);
4740 __io_req_complete(req, issue_flags, ret, 0);
4743 #else /* !CONFIG_NET */
4744 #define IO_NETOP_FN(op) \
4745 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
4747 return -EOPNOTSUPP; \
4750 #define IO_NETOP_PREP(op) \
4752 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
4754 return -EOPNOTSUPP; \
4757 #define IO_NETOP_PREP_ASYNC(op) \
4759 static int io_##op##_prep_async(struct io_kiocb *req) \
4761 return -EOPNOTSUPP; \
4764 IO_NETOP_PREP_ASYNC(sendmsg);
4765 IO_NETOP_PREP_ASYNC(recvmsg);
4766 IO_NETOP_PREP_ASYNC(connect);
4767 IO_NETOP_PREP(accept);
4770 #endif /* CONFIG_NET */
4772 struct io_poll_table {
4773 struct poll_table_struct pt;
4774 struct io_kiocb *req;
4778 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4779 __poll_t mask, task_work_func_t func)
4783 /* for instances that support it check for an event match first: */
4784 if (mask && !(mask & poll->events))
4787 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4789 list_del_init(&poll->wait.entry);
4792 req->task_work.func = func;
4793 percpu_ref_get(&req->ctx->refs);
4796 * If this fails, then the task is exiting. When a task exits, the
4797 * work gets canceled, so just cancel this request as well instead
4798 * of executing it. We can't safely execute it anyway, as we may not
4799 * have the needed state needed for it anyway.
4801 ret = io_req_task_work_add(req);
4802 if (unlikely(ret)) {
4803 WRITE_ONCE(poll->canceled, true);
4804 io_req_task_work_add_fallback(req, func);
4809 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4810 __acquires(&req->ctx->completion_lock)
4812 struct io_ring_ctx *ctx = req->ctx;
4814 if (!req->result && !READ_ONCE(poll->canceled)) {
4815 struct poll_table_struct pt = { ._key = poll->events };
4817 req->result = vfs_poll(req->file, &pt) & poll->events;
4820 spin_lock_irq(&ctx->completion_lock);
4821 if (!req->result && !READ_ONCE(poll->canceled)) {
4822 add_wait_queue(poll->head, &poll->wait);
4829 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4831 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4832 if (req->opcode == IORING_OP_POLL_ADD)
4833 return req->async_data;
4834 return req->apoll->double_poll;
4837 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4839 if (req->opcode == IORING_OP_POLL_ADD)
4841 return &req->apoll->poll;
4844 static void io_poll_remove_double(struct io_kiocb *req)
4846 struct io_poll_iocb *poll = io_poll_get_double(req);
4848 lockdep_assert_held(&req->ctx->completion_lock);
4850 if (poll && poll->head) {
4851 struct wait_queue_head *head = poll->head;
4853 spin_lock(&head->lock);
4854 list_del_init(&poll->wait.entry);
4855 if (poll->wait.private)
4856 refcount_dec(&req->refs);
4858 spin_unlock(&head->lock);
4862 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
4864 struct io_ring_ctx *ctx = req->ctx;
4866 io_poll_remove_double(req);
4867 req->poll.done = true;
4868 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
4869 io_commit_cqring(ctx);
4872 static void io_poll_task_func(struct callback_head *cb)
4874 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4875 struct io_ring_ctx *ctx = req->ctx;
4876 struct io_kiocb *nxt;
4878 if (io_poll_rewait(req, &req->poll)) {
4879 spin_unlock_irq(&ctx->completion_lock);
4881 hash_del(&req->hash_node);
4882 io_poll_complete(req, req->result, 0);
4883 spin_unlock_irq(&ctx->completion_lock);
4885 nxt = io_put_req_find_next(req);
4886 io_cqring_ev_posted(ctx);
4888 __io_req_task_submit(nxt);
4891 percpu_ref_put(&ctx->refs);
4894 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4895 int sync, void *key)
4897 struct io_kiocb *req = wait->private;
4898 struct io_poll_iocb *poll = io_poll_get_single(req);
4899 __poll_t mask = key_to_poll(key);
4901 /* for instances that support it check for an event match first: */
4902 if (mask && !(mask & poll->events))
4905 list_del_init(&wait->entry);
4907 if (poll && poll->head) {
4910 spin_lock(&poll->head->lock);
4911 done = list_empty(&poll->wait.entry);
4913 list_del_init(&poll->wait.entry);
4914 /* make sure double remove sees this as being gone */
4915 wait->private = NULL;
4916 spin_unlock(&poll->head->lock);
4918 /* use wait func handler, so it matches the rq type */
4919 poll->wait.func(&poll->wait, mode, sync, key);
4922 refcount_dec(&req->refs);
4926 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
4927 wait_queue_func_t wake_func)
4931 poll->canceled = false;
4932 poll->events = events;
4933 INIT_LIST_HEAD(&poll->wait.entry);
4934 init_waitqueue_func_entry(&poll->wait, wake_func);
4937 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
4938 struct wait_queue_head *head,
4939 struct io_poll_iocb **poll_ptr)
4941 struct io_kiocb *req = pt->req;
4944 * If poll->head is already set, it's because the file being polled
4945 * uses multiple waitqueues for poll handling (eg one for read, one
4946 * for write). Setup a separate io_poll_iocb if this happens.
4948 if (unlikely(poll->head)) {
4949 struct io_poll_iocb *poll_one = poll;
4951 /* already have a 2nd entry, fail a third attempt */
4953 pt->error = -EINVAL;
4956 /* double add on the same waitqueue head, ignore */
4957 if (poll->head == head)
4959 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
4961 pt->error = -ENOMEM;
4964 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
4965 refcount_inc(&req->refs);
4966 poll->wait.private = req;
4973 if (poll->events & EPOLLEXCLUSIVE)
4974 add_wait_queue_exclusive(head, &poll->wait);
4976 add_wait_queue(head, &poll->wait);
4979 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
4980 struct poll_table_struct *p)
4982 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
4983 struct async_poll *apoll = pt->req->apoll;
4985 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
4988 static void io_async_task_func(struct callback_head *cb)
4990 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4991 struct async_poll *apoll = req->apoll;
4992 struct io_ring_ctx *ctx = req->ctx;
4994 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
4996 if (io_poll_rewait(req, &apoll->poll)) {
4997 spin_unlock_irq(&ctx->completion_lock);
4998 percpu_ref_put(&ctx->refs);
5002 /* If req is still hashed, it cannot have been canceled. Don't check. */
5003 if (hash_hashed(&req->hash_node))
5004 hash_del(&req->hash_node);
5006 io_poll_remove_double(req);
5007 spin_unlock_irq(&ctx->completion_lock);
5009 if (!READ_ONCE(apoll->poll.canceled))
5010 __io_req_task_submit(req);
5012 __io_req_task_cancel(req, -ECANCELED);
5014 percpu_ref_put(&ctx->refs);
5015 kfree(apoll->double_poll);
5019 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5022 struct io_kiocb *req = wait->private;
5023 struct io_poll_iocb *poll = &req->apoll->poll;
5025 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5028 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5031 static void io_poll_req_insert(struct io_kiocb *req)
5033 struct io_ring_ctx *ctx = req->ctx;
5034 struct hlist_head *list;
5036 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5037 hlist_add_head(&req->hash_node, list);
5040 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5041 struct io_poll_iocb *poll,
5042 struct io_poll_table *ipt, __poll_t mask,
5043 wait_queue_func_t wake_func)
5044 __acquires(&ctx->completion_lock)
5046 struct io_ring_ctx *ctx = req->ctx;
5047 bool cancel = false;
5049 INIT_HLIST_NODE(&req->hash_node);
5050 io_init_poll_iocb(poll, mask, wake_func);
5051 poll->file = req->file;
5052 poll->wait.private = req;
5054 ipt->pt._key = mask;
5056 ipt->error = -EINVAL;
5058 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5060 spin_lock_irq(&ctx->completion_lock);
5061 if (likely(poll->head)) {
5062 spin_lock(&poll->head->lock);
5063 if (unlikely(list_empty(&poll->wait.entry))) {
5069 if (mask || ipt->error)
5070 list_del_init(&poll->wait.entry);
5072 WRITE_ONCE(poll->canceled, true);
5073 else if (!poll->done) /* actually waiting for an event */
5074 io_poll_req_insert(req);
5075 spin_unlock(&poll->head->lock);
5081 static bool io_arm_poll_handler(struct io_kiocb *req)
5083 const struct io_op_def *def = &io_op_defs[req->opcode];
5084 struct io_ring_ctx *ctx = req->ctx;
5085 struct async_poll *apoll;
5086 struct io_poll_table ipt;
5090 if (!req->file || !file_can_poll(req->file))
5092 if (req->flags & REQ_F_POLLED)
5096 else if (def->pollout)
5100 /* if we can't nonblock try, then no point in arming a poll handler */
5101 if (!io_file_supports_async(req->file, rw))
5104 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5105 if (unlikely(!apoll))
5107 apoll->double_poll = NULL;
5109 req->flags |= REQ_F_POLLED;
5114 mask |= POLLIN | POLLRDNORM;
5116 mask |= POLLOUT | POLLWRNORM;
5118 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5119 if ((req->opcode == IORING_OP_RECVMSG) &&
5120 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5123 mask |= POLLERR | POLLPRI;
5125 ipt.pt._qproc = io_async_queue_proc;
5127 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5129 if (ret || ipt.error) {
5130 io_poll_remove_double(req);
5131 spin_unlock_irq(&ctx->completion_lock);
5132 kfree(apoll->double_poll);
5136 spin_unlock_irq(&ctx->completion_lock);
5137 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
5138 apoll->poll.events);
5142 static bool __io_poll_remove_one(struct io_kiocb *req,
5143 struct io_poll_iocb *poll)
5145 bool do_complete = false;
5147 spin_lock(&poll->head->lock);
5148 WRITE_ONCE(poll->canceled, true);
5149 if (!list_empty(&poll->wait.entry)) {
5150 list_del_init(&poll->wait.entry);
5153 spin_unlock(&poll->head->lock);
5154 hash_del(&req->hash_node);
5158 static bool io_poll_remove_one(struct io_kiocb *req)
5162 io_poll_remove_double(req);
5164 if (req->opcode == IORING_OP_POLL_ADD) {
5165 do_complete = __io_poll_remove_one(req, &req->poll);
5167 struct async_poll *apoll = req->apoll;
5169 /* non-poll requests have submit ref still */
5170 do_complete = __io_poll_remove_one(req, &apoll->poll);
5173 kfree(apoll->double_poll);
5179 io_cqring_fill_event(req, -ECANCELED);
5180 io_commit_cqring(req->ctx);
5181 req_set_fail_links(req);
5182 io_put_req_deferred(req, 1);
5189 * Returns true if we found and killed one or more poll requests
5191 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5192 struct files_struct *files)
5194 struct hlist_node *tmp;
5195 struct io_kiocb *req;
5198 spin_lock_irq(&ctx->completion_lock);
5199 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5200 struct hlist_head *list;
5202 list = &ctx->cancel_hash[i];
5203 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5204 if (io_match_task(req, tsk, files))
5205 posted += io_poll_remove_one(req);
5208 spin_unlock_irq(&ctx->completion_lock);
5211 io_cqring_ev_posted(ctx);
5216 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
5218 struct hlist_head *list;
5219 struct io_kiocb *req;
5221 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5222 hlist_for_each_entry(req, list, hash_node) {
5223 if (sqe_addr != req->user_data)
5225 if (io_poll_remove_one(req))
5233 static int io_poll_remove_prep(struct io_kiocb *req,
5234 const struct io_uring_sqe *sqe)
5236 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5238 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
5242 req->poll_remove.addr = READ_ONCE(sqe->addr);
5247 * Find a running poll command that matches one specified in sqe->addr,
5248 * and remove it if found.
5250 static int io_poll_remove(struct io_kiocb *req, unsigned int issue_flags)
5252 struct io_ring_ctx *ctx = req->ctx;
5255 spin_lock_irq(&ctx->completion_lock);
5256 ret = io_poll_cancel(ctx, req->poll_remove.addr);
5257 spin_unlock_irq(&ctx->completion_lock);
5260 req_set_fail_links(req);
5261 io_req_complete(req, ret);
5265 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5268 struct io_kiocb *req = wait->private;
5269 struct io_poll_iocb *poll = &req->poll;
5271 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5274 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5275 struct poll_table_struct *p)
5277 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5279 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5282 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5284 struct io_poll_iocb *poll = &req->poll;
5287 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5289 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
5292 events = READ_ONCE(sqe->poll32_events);
5294 events = swahw32(events);
5296 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP |
5297 (events & EPOLLEXCLUSIVE);
5301 static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5303 struct io_poll_iocb *poll = &req->poll;
5304 struct io_ring_ctx *ctx = req->ctx;
5305 struct io_poll_table ipt;
5308 ipt.pt._qproc = io_poll_queue_proc;
5310 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5313 if (mask) { /* no async, we'd stolen it */
5315 io_poll_complete(req, mask, 0);
5317 spin_unlock_irq(&ctx->completion_lock);
5320 io_cqring_ev_posted(ctx);
5326 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5328 struct io_timeout_data *data = container_of(timer,
5329 struct io_timeout_data, timer);
5330 struct io_kiocb *req = data->req;
5331 struct io_ring_ctx *ctx = req->ctx;
5332 unsigned long flags;
5334 spin_lock_irqsave(&ctx->completion_lock, flags);
5335 list_del_init(&req->timeout.list);
5336 atomic_set(&req->ctx->cq_timeouts,
5337 atomic_read(&req->ctx->cq_timeouts) + 1);
5339 io_cqring_fill_event(req, -ETIME);
5340 io_commit_cqring(ctx);
5341 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5343 io_cqring_ev_posted(ctx);
5344 req_set_fail_links(req);
5346 return HRTIMER_NORESTART;
5349 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5352 struct io_timeout_data *io;
5353 struct io_kiocb *req;
5356 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5357 if (user_data == req->user_data) {
5364 return ERR_PTR(ret);
5366 io = req->async_data;
5367 ret = hrtimer_try_to_cancel(&io->timer);
5369 return ERR_PTR(-EALREADY);
5370 list_del_init(&req->timeout.list);
5374 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5376 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5379 return PTR_ERR(req);
5381 req_set_fail_links(req);
5382 io_cqring_fill_event(req, -ECANCELED);
5383 io_put_req_deferred(req, 1);
5387 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5388 struct timespec64 *ts, enum hrtimer_mode mode)
5390 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5391 struct io_timeout_data *data;
5394 return PTR_ERR(req);
5396 req->timeout.off = 0; /* noseq */
5397 data = req->async_data;
5398 list_add_tail(&req->timeout.list, &ctx->timeout_list);
5399 hrtimer_init(&data->timer, CLOCK_MONOTONIC, mode);
5400 data->timer.function = io_timeout_fn;
5401 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
5405 static int io_timeout_remove_prep(struct io_kiocb *req,
5406 const struct io_uring_sqe *sqe)
5408 struct io_timeout_rem *tr = &req->timeout_rem;
5410 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5412 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5414 if (sqe->ioprio || sqe->buf_index || sqe->len)
5417 tr->addr = READ_ONCE(sqe->addr);
5418 tr->flags = READ_ONCE(sqe->timeout_flags);
5419 if (tr->flags & IORING_TIMEOUT_UPDATE) {
5420 if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS))
5422 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
5424 } else if (tr->flags) {
5425 /* timeout removal doesn't support flags */
5432 static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
5434 return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
5439 * Remove or update an existing timeout command
5441 static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
5443 struct io_timeout_rem *tr = &req->timeout_rem;
5444 struct io_ring_ctx *ctx = req->ctx;
5447 spin_lock_irq(&ctx->completion_lock);
5448 if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE))
5449 ret = io_timeout_cancel(ctx, tr->addr);
5451 ret = io_timeout_update(ctx, tr->addr, &tr->ts,
5452 io_translate_timeout_mode(tr->flags));
5454 io_cqring_fill_event(req, ret);
5455 io_commit_cqring(ctx);
5456 spin_unlock_irq(&ctx->completion_lock);
5457 io_cqring_ev_posted(ctx);
5459 req_set_fail_links(req);
5464 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5465 bool is_timeout_link)
5467 struct io_timeout_data *data;
5469 u32 off = READ_ONCE(sqe->off);
5471 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5473 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5475 if (off && is_timeout_link)
5477 flags = READ_ONCE(sqe->timeout_flags);
5478 if (flags & ~IORING_TIMEOUT_ABS)
5481 req->timeout.off = off;
5483 if (!req->async_data && io_alloc_async_data(req))
5486 data = req->async_data;
5489 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5492 data->mode = io_translate_timeout_mode(flags);
5493 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5497 static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
5499 struct io_ring_ctx *ctx = req->ctx;
5500 struct io_timeout_data *data = req->async_data;
5501 struct list_head *entry;
5502 u32 tail, off = req->timeout.off;
5504 spin_lock_irq(&ctx->completion_lock);
5507 * sqe->off holds how many events that need to occur for this
5508 * timeout event to be satisfied. If it isn't set, then this is
5509 * a pure timeout request, sequence isn't used.
5511 if (io_is_timeout_noseq(req)) {
5512 entry = ctx->timeout_list.prev;
5516 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5517 req->timeout.target_seq = tail + off;
5519 /* Update the last seq here in case io_flush_timeouts() hasn't.
5520 * This is safe because ->completion_lock is held, and submissions
5521 * and completions are never mixed in the same ->completion_lock section.
5523 ctx->cq_last_tm_flush = tail;
5526 * Insertion sort, ensuring the first entry in the list is always
5527 * the one we need first.
5529 list_for_each_prev(entry, &ctx->timeout_list) {
5530 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5533 if (io_is_timeout_noseq(nxt))
5535 /* nxt.seq is behind @tail, otherwise would've been completed */
5536 if (off >= nxt->timeout.target_seq - tail)
5540 list_add(&req->timeout.list, entry);
5541 data->timer.function = io_timeout_fn;
5542 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5543 spin_unlock_irq(&ctx->completion_lock);
5547 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5549 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5551 return req->user_data == (unsigned long) data;
5554 static int io_async_cancel_one(struct io_uring_task *tctx, void *sqe_addr)
5556 enum io_wq_cancel cancel_ret;
5562 cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, sqe_addr, false);
5563 switch (cancel_ret) {
5564 case IO_WQ_CANCEL_OK:
5567 case IO_WQ_CANCEL_RUNNING:
5570 case IO_WQ_CANCEL_NOTFOUND:
5578 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5579 struct io_kiocb *req, __u64 sqe_addr,
5582 unsigned long flags;
5585 ret = io_async_cancel_one(req->task->io_uring,
5586 (void *) (unsigned long) sqe_addr);
5587 if (ret != -ENOENT) {
5588 spin_lock_irqsave(&ctx->completion_lock, flags);
5592 spin_lock_irqsave(&ctx->completion_lock, flags);
5593 ret = io_timeout_cancel(ctx, sqe_addr);
5596 ret = io_poll_cancel(ctx, sqe_addr);
5600 io_cqring_fill_event(req, ret);
5601 io_commit_cqring(ctx);
5602 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5603 io_cqring_ev_posted(ctx);
5606 req_set_fail_links(req);
5610 static int io_async_cancel_prep(struct io_kiocb *req,
5611 const struct io_uring_sqe *sqe)
5613 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5615 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5617 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5620 req->cancel.addr = READ_ONCE(sqe->addr);
5624 static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
5626 struct io_ring_ctx *ctx = req->ctx;
5628 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
5632 static int io_rsrc_update_prep(struct io_kiocb *req,
5633 const struct io_uring_sqe *sqe)
5635 if (unlikely(req->ctx->flags & IORING_SETUP_SQPOLL))
5637 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5639 if (sqe->ioprio || sqe->rw_flags)
5642 req->rsrc_update.offset = READ_ONCE(sqe->off);
5643 req->rsrc_update.nr_args = READ_ONCE(sqe->len);
5644 if (!req->rsrc_update.nr_args)
5646 req->rsrc_update.arg = READ_ONCE(sqe->addr);
5650 static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
5652 struct io_ring_ctx *ctx = req->ctx;
5653 struct io_uring_rsrc_update up;
5656 if (issue_flags & IO_URING_F_NONBLOCK)
5659 up.offset = req->rsrc_update.offset;
5660 up.data = req->rsrc_update.arg;
5662 mutex_lock(&ctx->uring_lock);
5663 ret = __io_sqe_files_update(ctx, &up, req->rsrc_update.nr_args);
5664 mutex_unlock(&ctx->uring_lock);
5667 req_set_fail_links(req);
5668 __io_req_complete(req, issue_flags, ret, 0);
5672 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5674 switch (req->opcode) {
5677 case IORING_OP_READV:
5678 case IORING_OP_READ_FIXED:
5679 case IORING_OP_READ:
5680 return io_read_prep(req, sqe);
5681 case IORING_OP_WRITEV:
5682 case IORING_OP_WRITE_FIXED:
5683 case IORING_OP_WRITE:
5684 return io_write_prep(req, sqe);
5685 case IORING_OP_POLL_ADD:
5686 return io_poll_add_prep(req, sqe);
5687 case IORING_OP_POLL_REMOVE:
5688 return io_poll_remove_prep(req, sqe);
5689 case IORING_OP_FSYNC:
5690 return io_fsync_prep(req, sqe);
5691 case IORING_OP_SYNC_FILE_RANGE:
5692 return io_sfr_prep(req, sqe);
5693 case IORING_OP_SENDMSG:
5694 case IORING_OP_SEND:
5695 return io_sendmsg_prep(req, sqe);
5696 case IORING_OP_RECVMSG:
5697 case IORING_OP_RECV:
5698 return io_recvmsg_prep(req, sqe);
5699 case IORING_OP_CONNECT:
5700 return io_connect_prep(req, sqe);
5701 case IORING_OP_TIMEOUT:
5702 return io_timeout_prep(req, sqe, false);
5703 case IORING_OP_TIMEOUT_REMOVE:
5704 return io_timeout_remove_prep(req, sqe);
5705 case IORING_OP_ASYNC_CANCEL:
5706 return io_async_cancel_prep(req, sqe);
5707 case IORING_OP_LINK_TIMEOUT:
5708 return io_timeout_prep(req, sqe, true);
5709 case IORING_OP_ACCEPT:
5710 return io_accept_prep(req, sqe);
5711 case IORING_OP_FALLOCATE:
5712 return io_fallocate_prep(req, sqe);
5713 case IORING_OP_OPENAT:
5714 return io_openat_prep(req, sqe);
5715 case IORING_OP_CLOSE:
5716 return io_close_prep(req, sqe);
5717 case IORING_OP_FILES_UPDATE:
5718 return io_rsrc_update_prep(req, sqe);
5719 case IORING_OP_STATX:
5720 return io_statx_prep(req, sqe);
5721 case IORING_OP_FADVISE:
5722 return io_fadvise_prep(req, sqe);
5723 case IORING_OP_MADVISE:
5724 return io_madvise_prep(req, sqe);
5725 case IORING_OP_OPENAT2:
5726 return io_openat2_prep(req, sqe);
5727 case IORING_OP_EPOLL_CTL:
5728 return io_epoll_ctl_prep(req, sqe);
5729 case IORING_OP_SPLICE:
5730 return io_splice_prep(req, sqe);
5731 case IORING_OP_PROVIDE_BUFFERS:
5732 return io_provide_buffers_prep(req, sqe);
5733 case IORING_OP_REMOVE_BUFFERS:
5734 return io_remove_buffers_prep(req, sqe);
5736 return io_tee_prep(req, sqe);
5737 case IORING_OP_SHUTDOWN:
5738 return io_shutdown_prep(req, sqe);
5739 case IORING_OP_RENAMEAT:
5740 return io_renameat_prep(req, sqe);
5741 case IORING_OP_UNLINKAT:
5742 return io_unlinkat_prep(req, sqe);
5745 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5750 static int io_req_prep_async(struct io_kiocb *req)
5752 switch (req->opcode) {
5753 case IORING_OP_READV:
5754 case IORING_OP_READ_FIXED:
5755 case IORING_OP_READ:
5756 return io_rw_prep_async(req, READ);
5757 case IORING_OP_WRITEV:
5758 case IORING_OP_WRITE_FIXED:
5759 case IORING_OP_WRITE:
5760 return io_rw_prep_async(req, WRITE);
5761 case IORING_OP_SENDMSG:
5762 case IORING_OP_SEND:
5763 return io_sendmsg_prep_async(req);
5764 case IORING_OP_RECVMSG:
5765 case IORING_OP_RECV:
5766 return io_recvmsg_prep_async(req);
5767 case IORING_OP_CONNECT:
5768 return io_connect_prep_async(req);
5773 static int io_req_defer_prep(struct io_kiocb *req)
5775 if (!io_op_defs[req->opcode].needs_async_data)
5777 /* some opcodes init it during the inital prep */
5778 if (req->async_data)
5780 if (__io_alloc_async_data(req))
5782 return io_req_prep_async(req);
5785 static u32 io_get_sequence(struct io_kiocb *req)
5787 struct io_kiocb *pos;
5788 struct io_ring_ctx *ctx = req->ctx;
5789 u32 total_submitted, nr_reqs = 0;
5791 io_for_each_link(pos, req)
5794 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
5795 return total_submitted - nr_reqs;
5798 static int io_req_defer(struct io_kiocb *req)
5800 struct io_ring_ctx *ctx = req->ctx;
5801 struct io_defer_entry *de;
5805 /* Still need defer if there is pending req in defer list. */
5806 if (likely(list_empty_careful(&ctx->defer_list) &&
5807 !(req->flags & REQ_F_IO_DRAIN)))
5810 seq = io_get_sequence(req);
5811 /* Still a chance to pass the sequence check */
5812 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
5815 ret = io_req_defer_prep(req);
5818 io_prep_async_link(req);
5819 de = kmalloc(sizeof(*de), GFP_KERNEL);
5823 spin_lock_irq(&ctx->completion_lock);
5824 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
5825 spin_unlock_irq(&ctx->completion_lock);
5827 io_queue_async_work(req);
5828 return -EIOCBQUEUED;
5831 trace_io_uring_defer(ctx, req, req->user_data);
5834 list_add_tail(&de->list, &ctx->defer_list);
5835 spin_unlock_irq(&ctx->completion_lock);
5836 return -EIOCBQUEUED;
5839 static void __io_clean_op(struct io_kiocb *req)
5841 if (req->flags & REQ_F_BUFFER_SELECTED) {
5842 switch (req->opcode) {
5843 case IORING_OP_READV:
5844 case IORING_OP_READ_FIXED:
5845 case IORING_OP_READ:
5846 kfree((void *)(unsigned long)req->rw.addr);
5848 case IORING_OP_RECVMSG:
5849 case IORING_OP_RECV:
5850 kfree(req->sr_msg.kbuf);
5853 req->flags &= ~REQ_F_BUFFER_SELECTED;
5856 if (req->flags & REQ_F_NEED_CLEANUP) {
5857 switch (req->opcode) {
5858 case IORING_OP_READV:
5859 case IORING_OP_READ_FIXED:
5860 case IORING_OP_READ:
5861 case IORING_OP_WRITEV:
5862 case IORING_OP_WRITE_FIXED:
5863 case IORING_OP_WRITE: {
5864 struct io_async_rw *io = req->async_data;
5866 kfree(io->free_iovec);
5869 case IORING_OP_RECVMSG:
5870 case IORING_OP_SENDMSG: {
5871 struct io_async_msghdr *io = req->async_data;
5873 kfree(io->free_iov);
5876 case IORING_OP_SPLICE:
5878 io_put_file(req, req->splice.file_in,
5879 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
5881 case IORING_OP_OPENAT:
5882 case IORING_OP_OPENAT2:
5883 if (req->open.filename)
5884 putname(req->open.filename);
5886 case IORING_OP_RENAMEAT:
5887 putname(req->rename.oldpath);
5888 putname(req->rename.newpath);
5890 case IORING_OP_UNLINKAT:
5891 putname(req->unlink.filename);
5894 req->flags &= ~REQ_F_NEED_CLEANUP;
5898 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
5900 struct io_ring_ctx *ctx = req->ctx;
5901 const struct cred *creds = NULL;
5904 if (req->work.personality) {
5905 const struct cred *new_creds;
5907 if (!(issue_flags & IO_URING_F_NONBLOCK))
5908 mutex_lock(&ctx->uring_lock);
5909 new_creds = idr_find(&ctx->personality_idr, req->work.personality);
5910 if (!(issue_flags & IO_URING_F_NONBLOCK))
5911 mutex_unlock(&ctx->uring_lock);
5914 creds = override_creds(new_creds);
5917 switch (req->opcode) {
5919 ret = io_nop(req, issue_flags);
5921 case IORING_OP_READV:
5922 case IORING_OP_READ_FIXED:
5923 case IORING_OP_READ:
5924 ret = io_read(req, issue_flags);
5926 case IORING_OP_WRITEV:
5927 case IORING_OP_WRITE_FIXED:
5928 case IORING_OP_WRITE:
5929 ret = io_write(req, issue_flags);
5931 case IORING_OP_FSYNC:
5932 ret = io_fsync(req, issue_flags);
5934 case IORING_OP_POLL_ADD:
5935 ret = io_poll_add(req, issue_flags);
5937 case IORING_OP_POLL_REMOVE:
5938 ret = io_poll_remove(req, issue_flags);
5940 case IORING_OP_SYNC_FILE_RANGE:
5941 ret = io_sync_file_range(req, issue_flags);
5943 case IORING_OP_SENDMSG:
5944 ret = io_sendmsg(req, issue_flags);
5946 case IORING_OP_SEND:
5947 ret = io_send(req, issue_flags);
5949 case IORING_OP_RECVMSG:
5950 ret = io_recvmsg(req, issue_flags);
5952 case IORING_OP_RECV:
5953 ret = io_recv(req, issue_flags);
5955 case IORING_OP_TIMEOUT:
5956 ret = io_timeout(req, issue_flags);
5958 case IORING_OP_TIMEOUT_REMOVE:
5959 ret = io_timeout_remove(req, issue_flags);
5961 case IORING_OP_ACCEPT:
5962 ret = io_accept(req, issue_flags);
5964 case IORING_OP_CONNECT:
5965 ret = io_connect(req, issue_flags);
5967 case IORING_OP_ASYNC_CANCEL:
5968 ret = io_async_cancel(req, issue_flags);
5970 case IORING_OP_FALLOCATE:
5971 ret = io_fallocate(req, issue_flags);
5973 case IORING_OP_OPENAT:
5974 ret = io_openat(req, issue_flags);
5976 case IORING_OP_CLOSE:
5977 ret = io_close(req, issue_flags);
5979 case IORING_OP_FILES_UPDATE:
5980 ret = io_files_update(req, issue_flags);
5982 case IORING_OP_STATX:
5983 ret = io_statx(req, issue_flags);
5985 case IORING_OP_FADVISE:
5986 ret = io_fadvise(req, issue_flags);
5988 case IORING_OP_MADVISE:
5989 ret = io_madvise(req, issue_flags);
5991 case IORING_OP_OPENAT2:
5992 ret = io_openat2(req, issue_flags);
5994 case IORING_OP_EPOLL_CTL:
5995 ret = io_epoll_ctl(req, issue_flags);
5997 case IORING_OP_SPLICE:
5998 ret = io_splice(req, issue_flags);
6000 case IORING_OP_PROVIDE_BUFFERS:
6001 ret = io_provide_buffers(req, issue_flags);
6003 case IORING_OP_REMOVE_BUFFERS:
6004 ret = io_remove_buffers(req, issue_flags);
6007 ret = io_tee(req, issue_flags);
6009 case IORING_OP_SHUTDOWN:
6010 ret = io_shutdown(req, issue_flags);
6012 case IORING_OP_RENAMEAT:
6013 ret = io_renameat(req, issue_flags);
6015 case IORING_OP_UNLINKAT:
6016 ret = io_unlinkat(req, issue_flags);
6024 revert_creds(creds);
6029 /* If the op doesn't have a file, we're not polling for it */
6030 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
6031 const bool in_async = io_wq_current_is_worker();
6033 /* workqueue context doesn't hold uring_lock, grab it now */
6035 mutex_lock(&ctx->uring_lock);
6037 io_iopoll_req_issued(req, in_async);
6040 mutex_unlock(&ctx->uring_lock);
6046 static void io_wq_submit_work(struct io_wq_work *work)
6048 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6049 struct io_kiocb *timeout;
6052 timeout = io_prep_linked_timeout(req);
6054 io_queue_linked_timeout(timeout);
6056 if (work->flags & IO_WQ_WORK_CANCEL)
6061 ret = io_issue_sqe(req, 0);
6063 * We can get EAGAIN for polled IO even though we're
6064 * forcing a sync submission from here, since we can't
6065 * wait for request slots on the block side.
6073 /* avoid locking problems by failing it from a clean context */
6075 /* io-wq is going to take one down */
6076 refcount_inc(&req->refs);
6077 io_req_task_queue_fail(req, ret);
6081 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6084 struct fixed_rsrc_table *table;
6086 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
6087 return table->files[index & IORING_FILE_TABLE_MASK];
6090 static struct file *io_file_get(struct io_submit_state *state,
6091 struct io_kiocb *req, int fd, bool fixed)
6093 struct io_ring_ctx *ctx = req->ctx;
6097 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6099 fd = array_index_nospec(fd, ctx->nr_user_files);
6100 file = io_file_from_index(ctx, fd);
6101 io_set_resource_node(req);
6103 trace_io_uring_file_get(ctx, fd);
6104 file = __io_file_get(state, fd);
6107 if (file && unlikely(file->f_op == &io_uring_fops))
6108 io_req_track_inflight(req);
6112 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6114 struct io_timeout_data *data = container_of(timer,
6115 struct io_timeout_data, timer);
6116 struct io_kiocb *prev, *req = data->req;
6117 struct io_ring_ctx *ctx = req->ctx;
6118 unsigned long flags;
6120 spin_lock_irqsave(&ctx->completion_lock, flags);
6121 prev = req->timeout.head;
6122 req->timeout.head = NULL;
6125 * We don't expect the list to be empty, that will only happen if we
6126 * race with the completion of the linked work.
6128 if (prev && refcount_inc_not_zero(&prev->refs))
6129 io_remove_next_linked(prev);
6132 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6135 req_set_fail_links(prev);
6136 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6137 io_put_req_deferred(prev, 1);
6139 io_req_complete_post(req, -ETIME, 0);
6140 io_put_req_deferred(req, 1);
6142 return HRTIMER_NORESTART;
6145 static void __io_queue_linked_timeout(struct io_kiocb *req)
6148 * If the back reference is NULL, then our linked request finished
6149 * before we got a chance to setup the timer
6151 if (req->timeout.head) {
6152 struct io_timeout_data *data = req->async_data;
6154 data->timer.function = io_link_timeout_fn;
6155 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6160 static void io_queue_linked_timeout(struct io_kiocb *req)
6162 struct io_ring_ctx *ctx = req->ctx;
6164 spin_lock_irq(&ctx->completion_lock);
6165 __io_queue_linked_timeout(req);
6166 spin_unlock_irq(&ctx->completion_lock);
6168 /* drop submission reference */
6172 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6174 struct io_kiocb *nxt = req->link;
6176 if (!nxt || (req->flags & REQ_F_LINK_TIMEOUT) ||
6177 nxt->opcode != IORING_OP_LINK_TIMEOUT)
6180 nxt->timeout.head = req;
6181 nxt->flags |= REQ_F_LTIMEOUT_ACTIVE;
6182 req->flags |= REQ_F_LINK_TIMEOUT;
6186 static void __io_queue_sqe(struct io_kiocb *req)
6188 struct io_kiocb *linked_timeout = io_prep_linked_timeout(req);
6191 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
6194 * We async punt it if the file wasn't marked NOWAIT, or if the file
6195 * doesn't support non-blocking read/write attempts
6197 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6198 if (!io_arm_poll_handler(req)) {
6200 * Queued up for async execution, worker will release
6201 * submit reference when the iocb is actually submitted.
6203 io_queue_async_work(req);
6205 } else if (likely(!ret)) {
6206 /* drop submission reference */
6207 if (req->flags & REQ_F_COMPLETE_INLINE) {
6208 struct io_ring_ctx *ctx = req->ctx;
6209 struct io_comp_state *cs = &ctx->submit_state.comp;
6211 cs->reqs[cs->nr++] = req;
6212 if (cs->nr == ARRAY_SIZE(cs->reqs))
6213 io_submit_flush_completions(cs, ctx);
6218 req_set_fail_links(req);
6220 io_req_complete(req, ret);
6223 io_queue_linked_timeout(linked_timeout);
6226 static void io_queue_sqe(struct io_kiocb *req)
6230 ret = io_req_defer(req);
6232 if (ret != -EIOCBQUEUED) {
6234 req_set_fail_links(req);
6236 io_req_complete(req, ret);
6238 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6239 ret = io_req_defer_prep(req);
6242 io_queue_async_work(req);
6244 __io_queue_sqe(req);
6249 * Check SQE restrictions (opcode and flags).
6251 * Returns 'true' if SQE is allowed, 'false' otherwise.
6253 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6254 struct io_kiocb *req,
6255 unsigned int sqe_flags)
6257 if (!ctx->restricted)
6260 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6263 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6264 ctx->restrictions.sqe_flags_required)
6267 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6268 ctx->restrictions.sqe_flags_required))
6274 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6275 const struct io_uring_sqe *sqe)
6277 struct io_submit_state *state;
6278 unsigned int sqe_flags;
6281 req->opcode = READ_ONCE(sqe->opcode);
6282 /* same numerical values with corresponding REQ_F_*, safe to copy */
6283 req->flags = sqe_flags = READ_ONCE(sqe->flags);
6284 req->user_data = READ_ONCE(sqe->user_data);
6285 req->async_data = NULL;
6289 req->fixed_rsrc_refs = NULL;
6290 /* one is dropped after submission, the other at completion */
6291 refcount_set(&req->refs, 2);
6292 req->task = current;
6295 /* enforce forwards compatibility on users */
6296 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS)) {
6301 if (unlikely(req->opcode >= IORING_OP_LAST))
6304 if (unlikely(!io_check_restriction(ctx, req, sqe_flags)))
6307 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6308 !io_op_defs[req->opcode].buffer_select)
6311 req->work.list.next = NULL;
6312 req->work.flags = 0;
6313 req->work.personality = READ_ONCE(sqe->personality);
6314 state = &ctx->submit_state;
6317 * Plug now if we have more than 1 IO left after this, and the target
6318 * is potentially a read/write to block based storage.
6320 if (!state->plug_started && state->ios_left > 1 &&
6321 io_op_defs[req->opcode].plug) {
6322 blk_start_plug(&state->plug);
6323 state->plug_started = true;
6326 if (io_op_defs[req->opcode].needs_file) {
6327 bool fixed = req->flags & REQ_F_FIXED_FILE;
6329 req->file = io_file_get(state, req, READ_ONCE(sqe->fd), fixed);
6330 if (unlikely(!req->file))
6338 static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
6339 const struct io_uring_sqe *sqe)
6341 struct io_submit_link *link = &ctx->submit_state.link;
6344 ret = io_init_req(ctx, req, sqe);
6345 if (unlikely(ret)) {
6348 io_req_complete(req, ret);
6350 /* fail even hard links since we don't submit */
6351 link->head->flags |= REQ_F_FAIL_LINK;
6352 io_put_req(link->head);
6353 io_req_complete(link->head, -ECANCELED);
6358 ret = io_req_prep(req, sqe);
6362 /* don't need @sqe from now on */
6363 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6364 true, ctx->flags & IORING_SETUP_SQPOLL);
6367 * If we already have a head request, queue this one for async
6368 * submittal once the head completes. If we don't have a head but
6369 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6370 * submitted sync once the chain is complete. If none of those
6371 * conditions are true (normal request), then just queue it.
6374 struct io_kiocb *head = link->head;
6377 * Taking sequential execution of a link, draining both sides
6378 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6379 * requests in the link. So, it drains the head and the
6380 * next after the link request. The last one is done via
6381 * drain_next flag to persist the effect across calls.
6383 if (req->flags & REQ_F_IO_DRAIN) {
6384 head->flags |= REQ_F_IO_DRAIN;
6385 ctx->drain_next = 1;
6387 ret = io_req_defer_prep(req);
6390 trace_io_uring_link(ctx, req, head);
6391 link->last->link = req;
6394 /* last request of a link, enqueue the link */
6395 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6400 if (unlikely(ctx->drain_next)) {
6401 req->flags |= REQ_F_IO_DRAIN;
6402 ctx->drain_next = 0;
6404 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6416 * Batched submission is done, ensure local IO is flushed out.
6418 static void io_submit_state_end(struct io_submit_state *state,
6419 struct io_ring_ctx *ctx)
6421 if (state->link.head)
6422 io_queue_sqe(state->link.head);
6424 io_submit_flush_completions(&state->comp, ctx);
6425 if (state->plug_started)
6426 blk_finish_plug(&state->plug);
6427 io_state_file_put(state);
6431 * Start submission side cache.
6433 static void io_submit_state_start(struct io_submit_state *state,
6434 unsigned int max_ios)
6436 state->plug_started = false;
6437 state->ios_left = max_ios;
6438 /* set only head, no need to init link_last in advance */
6439 state->link.head = NULL;
6442 static void io_commit_sqring(struct io_ring_ctx *ctx)
6444 struct io_rings *rings = ctx->rings;
6447 * Ensure any loads from the SQEs are done at this point,
6448 * since once we write the new head, the application could
6449 * write new data to them.
6451 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6455 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6456 * that is mapped by userspace. This means that care needs to be taken to
6457 * ensure that reads are stable, as we cannot rely on userspace always
6458 * being a good citizen. If members of the sqe are validated and then later
6459 * used, it's important that those reads are done through READ_ONCE() to
6460 * prevent a re-load down the line.
6462 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6464 u32 *sq_array = ctx->sq_array;
6468 * The cached sq head (or cq tail) serves two purposes:
6470 * 1) allows us to batch the cost of updating the user visible
6472 * 2) allows the kernel side to track the head on its own, even
6473 * though the application is the one updating it.
6475 head = READ_ONCE(sq_array[ctx->cached_sq_head++ & ctx->sq_mask]);
6476 if (likely(head < ctx->sq_entries))
6477 return &ctx->sq_sqes[head];
6479 /* drop invalid entries */
6480 ctx->cached_sq_dropped++;
6481 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6485 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6489 /* if we have a backlog and couldn't flush it all, return BUSY */
6490 if (test_bit(0, &ctx->sq_check_overflow)) {
6491 if (!__io_cqring_overflow_flush(ctx, false, NULL, NULL))
6495 /* make sure SQ entry isn't read before tail */
6496 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6498 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6501 percpu_counter_add(¤t->io_uring->inflight, nr);
6502 refcount_add(nr, ¤t->usage);
6503 io_submit_state_start(&ctx->submit_state, nr);
6505 while (submitted < nr) {
6506 const struct io_uring_sqe *sqe;
6507 struct io_kiocb *req;
6509 req = io_alloc_req(ctx);
6510 if (unlikely(!req)) {
6512 submitted = -EAGAIN;
6515 sqe = io_get_sqe(ctx);
6516 if (unlikely(!sqe)) {
6517 kmem_cache_free(req_cachep, req);
6520 /* will complete beyond this point, count as submitted */
6522 if (io_submit_sqe(ctx, req, sqe))
6526 if (unlikely(submitted != nr)) {
6527 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6528 struct io_uring_task *tctx = current->io_uring;
6529 int unused = nr - ref_used;
6531 percpu_ref_put_many(&ctx->refs, unused);
6532 percpu_counter_sub(&tctx->inflight, unused);
6533 put_task_struct_many(current, unused);
6536 io_submit_state_end(&ctx->submit_state, ctx);
6537 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6538 io_commit_sqring(ctx);
6543 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6545 /* Tell userspace we may need a wakeup call */
6546 spin_lock_irq(&ctx->completion_lock);
6547 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6548 spin_unlock_irq(&ctx->completion_lock);
6551 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6553 spin_lock_irq(&ctx->completion_lock);
6554 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6555 spin_unlock_irq(&ctx->completion_lock);
6558 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
6560 unsigned int to_submit;
6563 to_submit = io_sqring_entries(ctx);
6564 /* if we're handling multiple rings, cap submit size for fairness */
6565 if (cap_entries && to_submit > 8)
6568 if (!list_empty(&ctx->iopoll_list) || to_submit) {
6569 unsigned nr_events = 0;
6571 mutex_lock(&ctx->uring_lock);
6572 if (!list_empty(&ctx->iopoll_list))
6573 io_do_iopoll(ctx, &nr_events, 0);
6575 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)))
6576 ret = io_submit_sqes(ctx, to_submit);
6577 mutex_unlock(&ctx->uring_lock);
6580 if (!io_sqring_full(ctx) && wq_has_sleeper(&ctx->sqo_sq_wait))
6581 wake_up(&ctx->sqo_sq_wait);
6586 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
6588 struct io_ring_ctx *ctx;
6589 unsigned sq_thread_idle = 0;
6591 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6592 if (sq_thread_idle < ctx->sq_thread_idle)
6593 sq_thread_idle = ctx->sq_thread_idle;
6596 sqd->sq_thread_idle = sq_thread_idle;
6599 static void io_sqd_init_new(struct io_sq_data *sqd)
6601 struct io_ring_ctx *ctx;
6603 while (!list_empty(&sqd->ctx_new_list)) {
6604 ctx = list_first_entry(&sqd->ctx_new_list, struct io_ring_ctx, sqd_list);
6605 list_move_tail(&ctx->sqd_list, &sqd->ctx_list);
6606 complete(&ctx->sq_thread_comp);
6609 io_sqd_update_thread_idle(sqd);
6612 static bool io_sq_thread_should_stop(struct io_sq_data *sqd)
6614 return test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
6617 static bool io_sq_thread_should_park(struct io_sq_data *sqd)
6619 return test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
6622 static void io_sq_thread_parkme(struct io_sq_data *sqd)
6626 * TASK_PARKED is a special state; we must serialize against
6627 * possible pending wakeups to avoid store-store collisions on
6630 * Such a collision might possibly result in the task state
6631 * changin from TASK_PARKED and us failing the
6632 * wait_task_inactive() in kthread_park().
6634 set_special_state(TASK_PARKED);
6635 if (!test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state))
6639 * Thread is going to call schedule(), do not preempt it,
6640 * or the caller of kthread_park() may spend more time in
6641 * wait_task_inactive().
6644 complete(&sqd->completion);
6645 schedule_preempt_disabled();
6648 __set_current_state(TASK_RUNNING);
6651 static int io_sq_thread(void *data)
6653 struct io_sq_data *sqd = data;
6654 struct io_ring_ctx *ctx;
6655 unsigned long timeout = 0;
6656 char buf[TASK_COMM_LEN];
6659 sprintf(buf, "iou-sqp-%d", sqd->task_pid);
6660 set_task_comm(current, buf);
6661 sqd->thread = current;
6662 current->pf_io_worker = NULL;
6664 if (sqd->sq_cpu != -1)
6665 set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
6667 set_cpus_allowed_ptr(current, cpu_online_mask);
6668 current->flags |= PF_NO_SETAFFINITY;
6670 complete(&sqd->completion);
6672 wait_for_completion(&sqd->startup);
6674 while (!io_sq_thread_should_stop(sqd)) {
6676 bool cap_entries, sqt_spin, needs_sched;
6679 * Any changes to the sqd lists are synchronized through the
6680 * thread parking. This synchronizes the thread vs users,
6681 * the users are synchronized on the sqd->ctx_lock.
6683 if (io_sq_thread_should_park(sqd)) {
6684 io_sq_thread_parkme(sqd);
6687 if (unlikely(!list_empty(&sqd->ctx_new_list))) {
6688 io_sqd_init_new(sqd);
6689 timeout = jiffies + sqd->sq_thread_idle;
6691 if (fatal_signal_pending(current))
6694 cap_entries = !list_is_singular(&sqd->ctx_list);
6695 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6696 ret = __io_sq_thread(ctx, cap_entries);
6697 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
6701 if (sqt_spin || !time_after(jiffies, timeout)) {
6705 timeout = jiffies + sqd->sq_thread_idle;
6710 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
6711 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6712 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6713 !list_empty_careful(&ctx->iopoll_list)) {
6714 needs_sched = false;
6717 if (io_sqring_entries(ctx)) {
6718 needs_sched = false;
6723 if (needs_sched && !io_sq_thread_should_park(sqd)) {
6724 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6725 io_ring_set_wakeup_flag(ctx);
6728 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6729 io_ring_clear_wakeup_flag(ctx);
6732 finish_wait(&sqd->wait, &wait);
6733 timeout = jiffies + sqd->sq_thread_idle;
6736 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6737 io_uring_cancel_sqpoll(ctx);
6741 if (io_sq_thread_should_park(sqd))
6742 io_sq_thread_parkme(sqd);
6745 * Clear thread under lock so that concurrent parks work correctly
6747 complete(&sqd->completion);
6748 mutex_lock(&sqd->lock);
6750 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6752 io_ring_set_wakeup_flag(ctx);
6755 complete(&sqd->exited);
6756 mutex_unlock(&sqd->lock);
6760 struct io_wait_queue {
6761 struct wait_queue_entry wq;
6762 struct io_ring_ctx *ctx;
6764 unsigned nr_timeouts;
6767 static inline bool io_should_wake(struct io_wait_queue *iowq)
6769 struct io_ring_ctx *ctx = iowq->ctx;
6772 * Wake up if we have enough events, or if a timeout occurred since we
6773 * started waiting. For timeouts, we always want to return to userspace,
6774 * regardless of event count.
6776 return io_cqring_events(ctx) >= iowq->to_wait ||
6777 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6780 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6781 int wake_flags, void *key)
6783 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6787 * Cannot safely flush overflowed CQEs from here, ensure we wake up
6788 * the task, and the next invocation will do it.
6790 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->cq_check_overflow))
6791 return autoremove_wake_function(curr, mode, wake_flags, key);
6795 static int io_run_task_work_sig(void)
6797 if (io_run_task_work())
6799 if (!signal_pending(current))
6801 if (test_tsk_thread_flag(current, TIF_NOTIFY_SIGNAL))
6802 return -ERESTARTSYS;
6806 /* when returns >0, the caller should retry */
6807 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
6808 struct io_wait_queue *iowq,
6809 signed long *timeout)
6813 /* make sure we run task_work before checking for signals */
6814 ret = io_run_task_work_sig();
6815 if (ret || io_should_wake(iowq))
6817 /* let the caller flush overflows, retry */
6818 if (test_bit(0, &ctx->cq_check_overflow))
6821 *timeout = schedule_timeout(*timeout);
6822 return !*timeout ? -ETIME : 1;
6826 * Wait until events become available, if we don't already have some. The
6827 * application must reap them itself, as they reside on the shared cq ring.
6829 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6830 const sigset_t __user *sig, size_t sigsz,
6831 struct __kernel_timespec __user *uts)
6833 struct io_wait_queue iowq = {
6836 .func = io_wake_function,
6837 .entry = LIST_HEAD_INIT(iowq.wq.entry),
6840 .to_wait = min_events,
6842 struct io_rings *rings = ctx->rings;
6843 signed long timeout = MAX_SCHEDULE_TIMEOUT;
6847 io_cqring_overflow_flush(ctx, false, NULL, NULL);
6848 if (io_cqring_events(ctx) >= min_events)
6850 if (!io_run_task_work())
6855 #ifdef CONFIG_COMPAT
6856 if (in_compat_syscall())
6857 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
6861 ret = set_user_sigmask(sig, sigsz);
6868 struct timespec64 ts;
6870 if (get_timespec64(&ts, uts))
6872 timeout = timespec64_to_jiffies(&ts);
6875 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
6876 trace_io_uring_cqring_wait(ctx, min_events);
6878 io_cqring_overflow_flush(ctx, false, NULL, NULL);
6879 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
6880 TASK_INTERRUPTIBLE);
6881 ret = io_cqring_wait_schedule(ctx, &iowq, &timeout);
6882 finish_wait(&ctx->wait, &iowq.wq);
6885 restore_saved_sigmask_unless(ret == -EINTR);
6887 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
6890 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
6892 #if defined(CONFIG_UNIX)
6893 if (ctx->ring_sock) {
6894 struct sock *sock = ctx->ring_sock->sk;
6895 struct sk_buff *skb;
6897 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
6903 for (i = 0; i < ctx->nr_user_files; i++) {
6906 file = io_file_from_index(ctx, i);
6913 static void io_rsrc_data_ref_zero(struct percpu_ref *ref)
6915 struct fixed_rsrc_data *data;
6917 data = container_of(ref, struct fixed_rsrc_data, refs);
6918 complete(&data->done);
6921 static inline void io_rsrc_ref_lock(struct io_ring_ctx *ctx)
6923 spin_lock_bh(&ctx->rsrc_ref_lock);
6926 static inline void io_rsrc_ref_unlock(struct io_ring_ctx *ctx)
6928 spin_unlock_bh(&ctx->rsrc_ref_lock);
6931 static void io_sqe_rsrc_set_node(struct io_ring_ctx *ctx,
6932 struct fixed_rsrc_data *rsrc_data,
6933 struct fixed_rsrc_ref_node *ref_node)
6935 io_rsrc_ref_lock(ctx);
6936 rsrc_data->node = ref_node;
6937 list_add_tail(&ref_node->node, &ctx->rsrc_ref_list);
6938 io_rsrc_ref_unlock(ctx);
6939 percpu_ref_get(&rsrc_data->refs);
6942 static void io_sqe_rsrc_kill_node(struct io_ring_ctx *ctx, struct fixed_rsrc_data *data)
6944 struct fixed_rsrc_ref_node *ref_node = NULL;
6946 io_rsrc_ref_lock(ctx);
6947 ref_node = data->node;
6949 io_rsrc_ref_unlock(ctx);
6951 percpu_ref_kill(&ref_node->refs);
6954 static int io_rsrc_ref_quiesce(struct fixed_rsrc_data *data,
6955 struct io_ring_ctx *ctx,
6956 void (*rsrc_put)(struct io_ring_ctx *ctx,
6957 struct io_rsrc_put *prsrc))
6959 struct fixed_rsrc_ref_node *backup_node;
6965 data->quiesce = true;
6968 backup_node = alloc_fixed_rsrc_ref_node(ctx);
6971 backup_node->rsrc_data = data;
6972 backup_node->rsrc_put = rsrc_put;
6974 io_sqe_rsrc_kill_node(ctx, data);
6975 percpu_ref_kill(&data->refs);
6976 flush_delayed_work(&ctx->rsrc_put_work);
6978 ret = wait_for_completion_interruptible(&data->done);
6982 percpu_ref_resurrect(&data->refs);
6983 io_sqe_rsrc_set_node(ctx, data, backup_node);
6985 reinit_completion(&data->done);
6986 mutex_unlock(&ctx->uring_lock);
6987 ret = io_run_task_work_sig();
6988 mutex_lock(&ctx->uring_lock);
6990 data->quiesce = false;
6993 destroy_fixed_rsrc_ref_node(backup_node);
6997 static struct fixed_rsrc_data *alloc_fixed_rsrc_data(struct io_ring_ctx *ctx)
6999 struct fixed_rsrc_data *data;
7001 data = kzalloc(sizeof(*data), GFP_KERNEL);
7005 if (percpu_ref_init(&data->refs, io_rsrc_data_ref_zero,
7006 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) {
7011 init_completion(&data->done);
7015 static void free_fixed_rsrc_data(struct fixed_rsrc_data *data)
7017 percpu_ref_exit(&data->refs);
7022 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7024 struct fixed_rsrc_data *data = ctx->file_data;
7025 unsigned nr_tables, i;
7029 * percpu_ref_is_dying() is to stop parallel files unregister
7030 * Since we possibly drop uring lock later in this function to
7033 if (!data || percpu_ref_is_dying(&data->refs))
7035 ret = io_rsrc_ref_quiesce(data, ctx, io_ring_file_put);
7039 __io_sqe_files_unregister(ctx);
7040 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
7041 for (i = 0; i < nr_tables; i++)
7042 kfree(data->table[i].files);
7043 free_fixed_rsrc_data(data);
7044 ctx->file_data = NULL;
7045 ctx->nr_user_files = 0;
7049 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7050 __releases(&sqd->lock)
7054 if (sqd->thread == current)
7056 clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7057 wake_up_state(sqd->thread, TASK_PARKED);
7058 mutex_unlock(&sqd->lock);
7061 static bool io_sq_thread_park(struct io_sq_data *sqd)
7062 __acquires(&sqd->lock)
7064 if (sqd->thread == current)
7066 mutex_lock(&sqd->lock);
7068 mutex_unlock(&sqd->lock);
7071 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7072 wake_up_process(sqd->thread);
7073 wait_for_completion(&sqd->completion);
7077 static void io_sq_thread_stop(struct io_sq_data *sqd)
7079 if (test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state))
7081 mutex_lock(&sqd->lock);
7083 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7084 WARN_ON_ONCE(test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state));
7085 wake_up_process(sqd->thread);
7086 mutex_unlock(&sqd->lock);
7087 wait_for_completion(&sqd->exited);
7088 WARN_ON_ONCE(sqd->thread);
7090 mutex_unlock(&sqd->lock);
7094 static void io_put_sq_data(struct io_sq_data *sqd)
7096 if (refcount_dec_and_test(&sqd->refs)) {
7097 io_sq_thread_stop(sqd);
7102 static void io_sq_thread_finish(struct io_ring_ctx *ctx)
7104 struct io_sq_data *sqd = ctx->sq_data;
7107 complete(&sqd->startup);
7109 wait_for_completion(&ctx->sq_thread_comp);
7110 io_sq_thread_park(sqd);
7113 mutex_lock(&sqd->ctx_lock);
7114 list_del(&ctx->sqd_list);
7115 io_sqd_update_thread_idle(sqd);
7116 mutex_unlock(&sqd->ctx_lock);
7119 io_sq_thread_unpark(sqd);
7121 io_put_sq_data(sqd);
7122 ctx->sq_data = NULL;
7126 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7128 struct io_ring_ctx *ctx_attach;
7129 struct io_sq_data *sqd;
7132 f = fdget(p->wq_fd);
7134 return ERR_PTR(-ENXIO);
7135 if (f.file->f_op != &io_uring_fops) {
7137 return ERR_PTR(-EINVAL);
7140 ctx_attach = f.file->private_data;
7141 sqd = ctx_attach->sq_data;
7144 return ERR_PTR(-EINVAL);
7147 refcount_inc(&sqd->refs);
7152 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p)
7154 struct io_sq_data *sqd;
7156 if (p->flags & IORING_SETUP_ATTACH_WQ)
7157 return io_attach_sq_data(p);
7159 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7161 return ERR_PTR(-ENOMEM);
7163 refcount_set(&sqd->refs, 1);
7164 INIT_LIST_HEAD(&sqd->ctx_list);
7165 INIT_LIST_HEAD(&sqd->ctx_new_list);
7166 mutex_init(&sqd->ctx_lock);
7167 mutex_init(&sqd->lock);
7168 init_waitqueue_head(&sqd->wait);
7169 init_completion(&sqd->startup);
7170 init_completion(&sqd->completion);
7171 init_completion(&sqd->exited);
7175 #if defined(CONFIG_UNIX)
7177 * Ensure the UNIX gc is aware of our file set, so we are certain that
7178 * the io_uring can be safely unregistered on process exit, even if we have
7179 * loops in the file referencing.
7181 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7183 struct sock *sk = ctx->ring_sock->sk;
7184 struct scm_fp_list *fpl;
7185 struct sk_buff *skb;
7188 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7192 skb = alloc_skb(0, GFP_KERNEL);
7201 fpl->user = get_uid(current_user());
7202 for (i = 0; i < nr; i++) {
7203 struct file *file = io_file_from_index(ctx, i + offset);
7207 fpl->fp[nr_files] = get_file(file);
7208 unix_inflight(fpl->user, fpl->fp[nr_files]);
7213 fpl->max = SCM_MAX_FD;
7214 fpl->count = nr_files;
7215 UNIXCB(skb).fp = fpl;
7216 skb->destructor = unix_destruct_scm;
7217 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7218 skb_queue_head(&sk->sk_receive_queue, skb);
7220 for (i = 0; i < nr_files; i++)
7231 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7232 * causes regular reference counting to break down. We rely on the UNIX
7233 * garbage collection to take care of this problem for us.
7235 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7237 unsigned left, total;
7241 left = ctx->nr_user_files;
7243 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7245 ret = __io_sqe_files_scm(ctx, this_files, total);
7249 total += this_files;
7255 while (total < ctx->nr_user_files) {
7256 struct file *file = io_file_from_index(ctx, total);
7266 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7272 static int io_sqe_alloc_file_tables(struct fixed_rsrc_data *file_data,
7273 unsigned nr_tables, unsigned nr_files)
7277 for (i = 0; i < nr_tables; i++) {
7278 struct fixed_rsrc_table *table = &file_data->table[i];
7279 unsigned this_files;
7281 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
7282 table->files = kcalloc(this_files, sizeof(struct file *),
7286 nr_files -= this_files;
7292 for (i = 0; i < nr_tables; i++) {
7293 struct fixed_rsrc_table *table = &file_data->table[i];
7294 kfree(table->files);
7299 static void io_ring_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
7301 struct file *file = prsrc->file;
7302 #if defined(CONFIG_UNIX)
7303 struct sock *sock = ctx->ring_sock->sk;
7304 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7305 struct sk_buff *skb;
7308 __skb_queue_head_init(&list);
7311 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7312 * remove this entry and rearrange the file array.
7314 skb = skb_dequeue(head);
7316 struct scm_fp_list *fp;
7318 fp = UNIXCB(skb).fp;
7319 for (i = 0; i < fp->count; i++) {
7322 if (fp->fp[i] != file)
7325 unix_notinflight(fp->user, fp->fp[i]);
7326 left = fp->count - 1 - i;
7328 memmove(&fp->fp[i], &fp->fp[i + 1],
7329 left * sizeof(struct file *));
7336 __skb_queue_tail(&list, skb);
7346 __skb_queue_tail(&list, skb);
7348 skb = skb_dequeue(head);
7351 if (skb_peek(&list)) {
7352 spin_lock_irq(&head->lock);
7353 while ((skb = __skb_dequeue(&list)) != NULL)
7354 __skb_queue_tail(head, skb);
7355 spin_unlock_irq(&head->lock);
7362 static void __io_rsrc_put_work(struct fixed_rsrc_ref_node *ref_node)
7364 struct fixed_rsrc_data *rsrc_data = ref_node->rsrc_data;
7365 struct io_ring_ctx *ctx = rsrc_data->ctx;
7366 struct io_rsrc_put *prsrc, *tmp;
7368 list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
7369 list_del(&prsrc->list);
7370 ref_node->rsrc_put(ctx, prsrc);
7374 percpu_ref_exit(&ref_node->refs);
7376 percpu_ref_put(&rsrc_data->refs);
7379 static void io_rsrc_put_work(struct work_struct *work)
7381 struct io_ring_ctx *ctx;
7382 struct llist_node *node;
7384 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
7385 node = llist_del_all(&ctx->rsrc_put_llist);
7388 struct fixed_rsrc_ref_node *ref_node;
7389 struct llist_node *next = node->next;
7391 ref_node = llist_entry(node, struct fixed_rsrc_ref_node, llist);
7392 __io_rsrc_put_work(ref_node);
7397 static struct file **io_fixed_file_slot(struct fixed_rsrc_data *file_data,
7400 struct fixed_rsrc_table *table;
7402 table = &file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7403 return &table->files[i & IORING_FILE_TABLE_MASK];
7406 static void io_rsrc_node_ref_zero(struct percpu_ref *ref)
7408 struct fixed_rsrc_ref_node *ref_node;
7409 struct fixed_rsrc_data *data;
7410 struct io_ring_ctx *ctx;
7411 bool first_add = false;
7414 ref_node = container_of(ref, struct fixed_rsrc_ref_node, refs);
7415 data = ref_node->rsrc_data;
7418 io_rsrc_ref_lock(ctx);
7419 ref_node->done = true;
7421 while (!list_empty(&ctx->rsrc_ref_list)) {
7422 ref_node = list_first_entry(&ctx->rsrc_ref_list,
7423 struct fixed_rsrc_ref_node, node);
7424 /* recycle ref nodes in order */
7425 if (!ref_node->done)
7427 list_del(&ref_node->node);
7428 first_add |= llist_add(&ref_node->llist, &ctx->rsrc_put_llist);
7430 io_rsrc_ref_unlock(ctx);
7432 if (percpu_ref_is_dying(&data->refs))
7436 mod_delayed_work(system_wq, &ctx->rsrc_put_work, 0);
7438 queue_delayed_work(system_wq, &ctx->rsrc_put_work, delay);
7441 static struct fixed_rsrc_ref_node *alloc_fixed_rsrc_ref_node(
7442 struct io_ring_ctx *ctx)
7444 struct fixed_rsrc_ref_node *ref_node;
7446 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7450 if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
7455 INIT_LIST_HEAD(&ref_node->node);
7456 INIT_LIST_HEAD(&ref_node->rsrc_list);
7457 ref_node->done = false;
7461 static void init_fixed_file_ref_node(struct io_ring_ctx *ctx,
7462 struct fixed_rsrc_ref_node *ref_node)
7464 ref_node->rsrc_data = ctx->file_data;
7465 ref_node->rsrc_put = io_ring_file_put;
7468 static void destroy_fixed_rsrc_ref_node(struct fixed_rsrc_ref_node *ref_node)
7470 percpu_ref_exit(&ref_node->refs);
7475 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7478 __s32 __user *fds = (__s32 __user *) arg;
7479 unsigned nr_tables, i;
7481 int fd, ret = -ENOMEM;
7482 struct fixed_rsrc_ref_node *ref_node;
7483 struct fixed_rsrc_data *file_data;
7489 if (nr_args > IORING_MAX_FIXED_FILES)
7492 file_data = alloc_fixed_rsrc_data(ctx);
7495 ctx->file_data = file_data;
7497 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
7498 file_data->table = kcalloc(nr_tables, sizeof(*file_data->table),
7500 if (!file_data->table)
7503 if (io_sqe_alloc_file_tables(file_data, nr_tables, nr_args))
7506 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7507 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7511 /* allow sparse sets */
7521 * Don't allow io_uring instances to be registered. If UNIX
7522 * isn't enabled, then this causes a reference cycle and this
7523 * instance can never get freed. If UNIX is enabled we'll
7524 * handle it just fine, but there's still no point in allowing
7525 * a ring fd as it doesn't support regular read/write anyway.
7527 if (file->f_op == &io_uring_fops) {
7531 *io_fixed_file_slot(file_data, i) = file;
7534 ret = io_sqe_files_scm(ctx);
7536 io_sqe_files_unregister(ctx);
7540 ref_node = alloc_fixed_rsrc_ref_node(ctx);
7542 io_sqe_files_unregister(ctx);
7545 init_fixed_file_ref_node(ctx, ref_node);
7547 io_sqe_rsrc_set_node(ctx, file_data, ref_node);
7550 for (i = 0; i < ctx->nr_user_files; i++) {
7551 file = io_file_from_index(ctx, i);
7555 for (i = 0; i < nr_tables; i++)
7556 kfree(file_data->table[i].files);
7557 ctx->nr_user_files = 0;
7559 free_fixed_rsrc_data(ctx->file_data);
7560 ctx->file_data = NULL;
7564 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7567 #if defined(CONFIG_UNIX)
7568 struct sock *sock = ctx->ring_sock->sk;
7569 struct sk_buff_head *head = &sock->sk_receive_queue;
7570 struct sk_buff *skb;
7573 * See if we can merge this file into an existing skb SCM_RIGHTS
7574 * file set. If there's no room, fall back to allocating a new skb
7575 * and filling it in.
7577 spin_lock_irq(&head->lock);
7578 skb = skb_peek(head);
7580 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7582 if (fpl->count < SCM_MAX_FD) {
7583 __skb_unlink(skb, head);
7584 spin_unlock_irq(&head->lock);
7585 fpl->fp[fpl->count] = get_file(file);
7586 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7588 spin_lock_irq(&head->lock);
7589 __skb_queue_head(head, skb);
7594 spin_unlock_irq(&head->lock);
7601 return __io_sqe_files_scm(ctx, 1, index);
7607 static int io_queue_rsrc_removal(struct fixed_rsrc_data *data, void *rsrc)
7609 struct io_rsrc_put *prsrc;
7610 struct fixed_rsrc_ref_node *ref_node = data->node;
7612 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
7617 list_add(&prsrc->list, &ref_node->rsrc_list);
7622 static inline int io_queue_file_removal(struct fixed_rsrc_data *data,
7625 return io_queue_rsrc_removal(data, (void *)file);
7628 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7629 struct io_uring_rsrc_update *up,
7632 struct fixed_rsrc_data *data = ctx->file_data;
7633 struct fixed_rsrc_ref_node *ref_node;
7634 struct file *file, **file_slot;
7638 bool needs_switch = false;
7640 if (check_add_overflow(up->offset, nr_args, &done))
7642 if (done > ctx->nr_user_files)
7645 ref_node = alloc_fixed_rsrc_ref_node(ctx);
7648 init_fixed_file_ref_node(ctx, ref_node);
7650 fds = u64_to_user_ptr(up->data);
7651 for (done = 0; done < nr_args; done++) {
7653 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
7657 if (fd == IORING_REGISTER_FILES_SKIP)
7660 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
7661 file_slot = io_fixed_file_slot(ctx->file_data, i);
7664 err = io_queue_file_removal(data, *file_slot);
7668 needs_switch = true;
7677 * Don't allow io_uring instances to be registered. If
7678 * UNIX isn't enabled, then this causes a reference
7679 * cycle and this instance can never get freed. If UNIX
7680 * is enabled we'll handle it just fine, but there's
7681 * still no point in allowing a ring fd as it doesn't
7682 * support regular read/write anyway.
7684 if (file->f_op == &io_uring_fops) {
7690 err = io_sqe_file_register(ctx, file, i);
7700 percpu_ref_kill(&data->node->refs);
7701 io_sqe_rsrc_set_node(ctx, data, ref_node);
7703 destroy_fixed_rsrc_ref_node(ref_node);
7705 return done ? done : err;
7708 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
7711 struct io_uring_rsrc_update up;
7713 if (!ctx->file_data)
7717 if (copy_from_user(&up, arg, sizeof(up)))
7722 return __io_sqe_files_update(ctx, &up, nr_args);
7725 static struct io_wq_work *io_free_work(struct io_wq_work *work)
7727 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7729 req = io_put_req_find_next(req);
7730 return req ? &req->work : NULL;
7733 static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx)
7735 struct io_wq_hash *hash;
7736 struct io_wq_data data;
7737 unsigned int concurrency;
7739 hash = ctx->hash_map;
7741 hash = kzalloc(sizeof(*hash), GFP_KERNEL);
7743 return ERR_PTR(-ENOMEM);
7744 refcount_set(&hash->refs, 1);
7745 init_waitqueue_head(&hash->wait);
7746 ctx->hash_map = hash;
7750 data.free_work = io_free_work;
7751 data.do_work = io_wq_submit_work;
7753 /* Do QD, or 4 * CPUS, whatever is smallest */
7754 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7756 return io_wq_create(concurrency, &data);
7759 static int io_uring_alloc_task_context(struct task_struct *task,
7760 struct io_ring_ctx *ctx)
7762 struct io_uring_task *tctx;
7765 tctx = kmalloc(sizeof(*tctx), GFP_KERNEL);
7766 if (unlikely(!tctx))
7769 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
7770 if (unlikely(ret)) {
7775 tctx->io_wq = io_init_wq_offload(ctx);
7776 if (IS_ERR(tctx->io_wq)) {
7777 ret = PTR_ERR(tctx->io_wq);
7778 percpu_counter_destroy(&tctx->inflight);
7784 init_waitqueue_head(&tctx->wait);
7786 atomic_set(&tctx->in_idle, 0);
7787 tctx->sqpoll = false;
7788 task->io_uring = tctx;
7789 spin_lock_init(&tctx->task_lock);
7790 INIT_WQ_LIST(&tctx->task_list);
7791 tctx->task_state = 0;
7792 init_task_work(&tctx->task_work, tctx_task_work);
7796 void __io_uring_free(struct task_struct *tsk)
7798 struct io_uring_task *tctx = tsk->io_uring;
7800 WARN_ON_ONCE(!xa_empty(&tctx->xa));
7801 WARN_ON_ONCE(tctx->io_wq);
7803 percpu_counter_destroy(&tctx->inflight);
7805 tsk->io_uring = NULL;
7808 static int io_sq_thread_fork(struct io_sq_data *sqd, struct io_ring_ctx *ctx)
7812 clear_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7813 reinit_completion(&sqd->completion);
7815 sqd->task_pid = current->pid;
7816 current->flags |= PF_IO_WORKER;
7817 ret = io_wq_fork_thread(io_sq_thread, sqd);
7818 current->flags &= ~PF_IO_WORKER;
7823 wait_for_completion(&sqd->completion);
7824 return io_uring_alloc_task_context(sqd->thread, ctx);
7827 static int io_sq_offload_create(struct io_ring_ctx *ctx,
7828 struct io_uring_params *p)
7832 /* Retain compatibility with failing for an invalid attach attempt */
7833 if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
7834 IORING_SETUP_ATTACH_WQ) {
7837 f = fdget(p->wq_fd);
7840 if (f.file->f_op != &io_uring_fops) {
7846 if (ctx->flags & IORING_SETUP_SQPOLL) {
7847 struct io_sq_data *sqd;
7850 if (!capable(CAP_SYS_ADMIN) && !capable(CAP_SYS_NICE))
7853 sqd = io_get_sq_data(p);
7860 io_sq_thread_park(sqd);
7861 mutex_lock(&sqd->ctx_lock);
7862 list_add(&ctx->sqd_list, &sqd->ctx_new_list);
7863 mutex_unlock(&sqd->ctx_lock);
7864 io_sq_thread_unpark(sqd);
7866 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
7867 if (!ctx->sq_thread_idle)
7868 ctx->sq_thread_idle = HZ;
7873 if (p->flags & IORING_SETUP_SQ_AFF) {
7874 int cpu = p->sq_thread_cpu;
7877 if (cpu >= nr_cpu_ids)
7879 if (!cpu_online(cpu))
7887 sqd->task_pid = current->pid;
7888 current->flags |= PF_IO_WORKER;
7889 ret = io_wq_fork_thread(io_sq_thread, sqd);
7890 current->flags &= ~PF_IO_WORKER;
7895 wait_for_completion(&sqd->completion);
7896 ret = io_uring_alloc_task_context(sqd->thread, ctx);
7899 } else if (p->flags & IORING_SETUP_SQ_AFF) {
7900 /* Can't have SQ_AFF without SQPOLL */
7907 io_sq_thread_finish(ctx);
7911 static void io_sq_offload_start(struct io_ring_ctx *ctx)
7913 struct io_sq_data *sqd = ctx->sq_data;
7915 ctx->flags &= ~IORING_SETUP_R_DISABLED;
7916 if (ctx->flags & IORING_SETUP_SQPOLL)
7917 complete(&sqd->startup);
7920 static inline void __io_unaccount_mem(struct user_struct *user,
7921 unsigned long nr_pages)
7923 atomic_long_sub(nr_pages, &user->locked_vm);
7926 static inline int __io_account_mem(struct user_struct *user,
7927 unsigned long nr_pages)
7929 unsigned long page_limit, cur_pages, new_pages;
7931 /* Don't allow more pages than we can safely lock */
7932 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
7935 cur_pages = atomic_long_read(&user->locked_vm);
7936 new_pages = cur_pages + nr_pages;
7937 if (new_pages > page_limit)
7939 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
7940 new_pages) != cur_pages);
7945 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
7948 __io_unaccount_mem(ctx->user, nr_pages);
7950 if (ctx->mm_account)
7951 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
7954 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
7959 ret = __io_account_mem(ctx->user, nr_pages);
7964 if (ctx->mm_account)
7965 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
7970 static void io_mem_free(void *ptr)
7977 page = virt_to_head_page(ptr);
7978 if (put_page_testzero(page))
7979 free_compound_page(page);
7982 static void *io_mem_alloc(size_t size)
7984 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
7985 __GFP_NORETRY | __GFP_ACCOUNT;
7987 return (void *) __get_free_pages(gfp_flags, get_order(size));
7990 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
7993 struct io_rings *rings;
7994 size_t off, sq_array_size;
7996 off = struct_size(rings, cqes, cq_entries);
7997 if (off == SIZE_MAX)
8001 off = ALIGN(off, SMP_CACHE_BYTES);
8009 sq_array_size = array_size(sizeof(u32), sq_entries);
8010 if (sq_array_size == SIZE_MAX)
8013 if (check_add_overflow(off, sq_array_size, &off))
8019 static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8023 if (!ctx->user_bufs)
8026 for (i = 0; i < ctx->nr_user_bufs; i++) {
8027 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8029 for (j = 0; j < imu->nr_bvecs; j++)
8030 unpin_user_page(imu->bvec[j].bv_page);
8032 if (imu->acct_pages)
8033 io_unaccount_mem(ctx, imu->acct_pages);
8038 kfree(ctx->user_bufs);
8039 ctx->user_bufs = NULL;
8040 ctx->nr_user_bufs = 0;
8044 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8045 void __user *arg, unsigned index)
8047 struct iovec __user *src;
8049 #ifdef CONFIG_COMPAT
8051 struct compat_iovec __user *ciovs;
8052 struct compat_iovec ciov;
8054 ciovs = (struct compat_iovec __user *) arg;
8055 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8058 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8059 dst->iov_len = ciov.iov_len;
8063 src = (struct iovec __user *) arg;
8064 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8070 * Not super efficient, but this is just a registration time. And we do cache
8071 * the last compound head, so generally we'll only do a full search if we don't
8074 * We check if the given compound head page has already been accounted, to
8075 * avoid double accounting it. This allows us to account the full size of the
8076 * page, not just the constituent pages of a huge page.
8078 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8079 int nr_pages, struct page *hpage)
8083 /* check current page array */
8084 for (i = 0; i < nr_pages; i++) {
8085 if (!PageCompound(pages[i]))
8087 if (compound_head(pages[i]) == hpage)
8091 /* check previously registered pages */
8092 for (i = 0; i < ctx->nr_user_bufs; i++) {
8093 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8095 for (j = 0; j < imu->nr_bvecs; j++) {
8096 if (!PageCompound(imu->bvec[j].bv_page))
8098 if (compound_head(imu->bvec[j].bv_page) == hpage)
8106 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8107 int nr_pages, struct io_mapped_ubuf *imu,
8108 struct page **last_hpage)
8112 for (i = 0; i < nr_pages; i++) {
8113 if (!PageCompound(pages[i])) {
8118 hpage = compound_head(pages[i]);
8119 if (hpage == *last_hpage)
8121 *last_hpage = hpage;
8122 if (headpage_already_acct(ctx, pages, i, hpage))
8124 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8128 if (!imu->acct_pages)
8131 ret = io_account_mem(ctx, imu->acct_pages);
8133 imu->acct_pages = 0;
8137 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
8138 struct io_mapped_ubuf *imu,
8139 struct page **last_hpage)
8141 struct vm_area_struct **vmas = NULL;
8142 struct page **pages = NULL;
8143 unsigned long off, start, end, ubuf;
8145 int ret, pret, nr_pages, i;
8147 ubuf = (unsigned long) iov->iov_base;
8148 end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8149 start = ubuf >> PAGE_SHIFT;
8150 nr_pages = end - start;
8154 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
8158 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
8163 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
8169 mmap_read_lock(current->mm);
8170 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
8172 if (pret == nr_pages) {
8173 /* don't support file backed memory */
8174 for (i = 0; i < nr_pages; i++) {
8175 struct vm_area_struct *vma = vmas[i];
8178 !is_file_hugepages(vma->vm_file)) {
8184 ret = pret < 0 ? pret : -EFAULT;
8186 mmap_read_unlock(current->mm);
8189 * if we did partial map, or found file backed vmas,
8190 * release any pages we did get
8193 unpin_user_pages(pages, pret);
8198 ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
8200 unpin_user_pages(pages, pret);
8205 off = ubuf & ~PAGE_MASK;
8206 size = iov->iov_len;
8207 for (i = 0; i < nr_pages; i++) {
8210 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8211 imu->bvec[i].bv_page = pages[i];
8212 imu->bvec[i].bv_len = vec_len;
8213 imu->bvec[i].bv_offset = off;
8217 /* store original address for later verification */
8219 imu->len = iov->iov_len;
8220 imu->nr_bvecs = nr_pages;
8228 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
8232 if (!nr_args || nr_args > UIO_MAXIOV)
8235 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
8237 if (!ctx->user_bufs)
8243 static int io_buffer_validate(struct iovec *iov)
8246 * Don't impose further limits on the size and buffer
8247 * constraints here, we'll -EINVAL later when IO is
8248 * submitted if they are wrong.
8250 if (!iov->iov_base || !iov->iov_len)
8253 /* arbitrary limit, but we need something */
8254 if (iov->iov_len > SZ_1G)
8260 static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
8261 unsigned int nr_args)
8265 struct page *last_hpage = NULL;
8267 ret = io_buffers_map_alloc(ctx, nr_args);
8271 for (i = 0; i < nr_args; i++) {
8272 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8274 ret = io_copy_iov(ctx, &iov, arg, i);
8278 ret = io_buffer_validate(&iov);
8282 ret = io_sqe_buffer_register(ctx, &iov, imu, &last_hpage);
8286 ctx->nr_user_bufs++;
8290 io_sqe_buffers_unregister(ctx);
8295 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8297 __s32 __user *fds = arg;
8303 if (copy_from_user(&fd, fds, sizeof(*fds)))
8306 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8307 if (IS_ERR(ctx->cq_ev_fd)) {
8308 int ret = PTR_ERR(ctx->cq_ev_fd);
8309 ctx->cq_ev_fd = NULL;
8316 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8318 if (ctx->cq_ev_fd) {
8319 eventfd_ctx_put(ctx->cq_ev_fd);
8320 ctx->cq_ev_fd = NULL;
8327 static int __io_destroy_buffers(int id, void *p, void *data)
8329 struct io_ring_ctx *ctx = data;
8330 struct io_buffer *buf = p;
8332 __io_remove_buffers(ctx, buf, id, -1U);
8336 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8338 idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx);
8339 idr_destroy(&ctx->io_buffer_idr);
8342 static void io_req_cache_free(struct list_head *list, struct task_struct *tsk)
8344 struct io_kiocb *req, *nxt;
8346 list_for_each_entry_safe(req, nxt, list, compl.list) {
8347 if (tsk && req->task != tsk)
8349 list_del(&req->compl.list);
8350 kmem_cache_free(req_cachep, req);
8354 static void io_req_caches_free(struct io_ring_ctx *ctx)
8356 struct io_submit_state *submit_state = &ctx->submit_state;
8357 struct io_comp_state *cs = &ctx->submit_state.comp;
8359 mutex_lock(&ctx->uring_lock);
8361 if (submit_state->free_reqs) {
8362 kmem_cache_free_bulk(req_cachep, submit_state->free_reqs,
8363 submit_state->reqs);
8364 submit_state->free_reqs = 0;
8367 spin_lock_irq(&ctx->completion_lock);
8368 list_splice_init(&cs->locked_free_list, &cs->free_list);
8369 cs->locked_free_nr = 0;
8370 spin_unlock_irq(&ctx->completion_lock);
8372 io_req_cache_free(&cs->free_list, NULL);
8374 mutex_unlock(&ctx->uring_lock);
8377 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8380 * Some may use context even when all refs and requests have been put,
8381 * and they are free to do so while still holding uring_lock, see
8382 * __io_req_task_submit(). Wait for them to finish.
8384 mutex_lock(&ctx->uring_lock);
8385 mutex_unlock(&ctx->uring_lock);
8387 io_sq_thread_finish(ctx);
8388 io_sqe_buffers_unregister(ctx);
8390 if (ctx->mm_account) {
8391 mmdrop(ctx->mm_account);
8392 ctx->mm_account = NULL;
8395 mutex_lock(&ctx->uring_lock);
8396 io_sqe_files_unregister(ctx);
8397 mutex_unlock(&ctx->uring_lock);
8398 io_eventfd_unregister(ctx);
8399 io_destroy_buffers(ctx);
8400 idr_destroy(&ctx->personality_idr);
8402 #if defined(CONFIG_UNIX)
8403 if (ctx->ring_sock) {
8404 ctx->ring_sock->file = NULL; /* so that iput() is called */
8405 sock_release(ctx->ring_sock);
8409 io_mem_free(ctx->rings);
8410 io_mem_free(ctx->sq_sqes);
8412 percpu_ref_exit(&ctx->refs);
8413 free_uid(ctx->user);
8414 io_req_caches_free(ctx);
8416 io_wq_put_hash(ctx->hash_map);
8417 kfree(ctx->cancel_hash);
8421 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8423 struct io_ring_ctx *ctx = file->private_data;
8426 poll_wait(file, &ctx->cq_wait, wait);
8428 * synchronizes with barrier from wq_has_sleeper call in
8432 if (!io_sqring_full(ctx))
8433 mask |= EPOLLOUT | EPOLLWRNORM;
8436 * Don't flush cqring overflow list here, just do a simple check.
8437 * Otherwise there could possible be ABBA deadlock:
8440 * lock(&ctx->uring_lock);
8442 * lock(&ctx->uring_lock);
8445 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
8446 * pushs them to do the flush.
8448 if (io_cqring_events(ctx) || test_bit(0, &ctx->cq_check_overflow))
8449 mask |= EPOLLIN | EPOLLRDNORM;
8454 static int io_uring_fasync(int fd, struct file *file, int on)
8456 struct io_ring_ctx *ctx = file->private_data;
8458 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8461 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8463 const struct cred *creds;
8465 creds = idr_remove(&ctx->personality_idr, id);
8474 static int io_remove_personalities(int id, void *p, void *data)
8476 struct io_ring_ctx *ctx = data;
8478 io_unregister_personality(ctx, id);
8482 static bool io_run_ctx_fallback(struct io_ring_ctx *ctx)
8484 struct callback_head *work, *next;
8485 bool executed = false;
8488 work = xchg(&ctx->exit_task_work, NULL);
8504 static void io_ring_exit_work(struct work_struct *work)
8506 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
8510 * If we're doing polled IO and end up having requests being
8511 * submitted async (out-of-line), then completions can come in while
8512 * we're waiting for refs to drop. We need to reap these manually,
8513 * as nobody else will be looking for them.
8516 io_uring_try_cancel_requests(ctx, NULL, NULL);
8517 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
8518 io_ring_ctx_free(ctx);
8521 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8523 mutex_lock(&ctx->uring_lock);
8524 percpu_ref_kill(&ctx->refs);
8525 /* if force is set, the ring is going away. always drop after that */
8526 ctx->cq_overflow_flushed = 1;
8528 __io_cqring_overflow_flush(ctx, true, NULL, NULL);
8529 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
8530 mutex_unlock(&ctx->uring_lock);
8532 io_kill_timeouts(ctx, NULL, NULL);
8533 io_poll_remove_all(ctx, NULL, NULL);
8535 /* if we failed setting up the ctx, we might not have any rings */
8536 io_iopoll_try_reap_events(ctx);
8538 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8540 * Use system_unbound_wq to avoid spawning tons of event kworkers
8541 * if we're exiting a ton of rings at the same time. It just adds
8542 * noise and overhead, there's no discernable change in runtime
8543 * over using system_wq.
8545 queue_work(system_unbound_wq, &ctx->exit_work);
8548 static int io_uring_release(struct inode *inode, struct file *file)
8550 struct io_ring_ctx *ctx = file->private_data;
8552 file->private_data = NULL;
8553 io_ring_ctx_wait_and_kill(ctx);
8557 struct io_task_cancel {
8558 struct task_struct *task;
8559 struct files_struct *files;
8562 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8564 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8565 struct io_task_cancel *cancel = data;
8568 if (cancel->files && (req->flags & REQ_F_LINK_TIMEOUT)) {
8569 unsigned long flags;
8570 struct io_ring_ctx *ctx = req->ctx;
8572 /* protect against races with linked timeouts */
8573 spin_lock_irqsave(&ctx->completion_lock, flags);
8574 ret = io_match_task(req, cancel->task, cancel->files);
8575 spin_unlock_irqrestore(&ctx->completion_lock, flags);
8577 ret = io_match_task(req, cancel->task, cancel->files);
8582 static void io_cancel_defer_files(struct io_ring_ctx *ctx,
8583 struct task_struct *task,
8584 struct files_struct *files)
8586 struct io_defer_entry *de = NULL;
8589 spin_lock_irq(&ctx->completion_lock);
8590 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8591 if (io_match_task(de->req, task, files)) {
8592 list_cut_position(&list, &ctx->defer_list, &de->list);
8596 spin_unlock_irq(&ctx->completion_lock);
8598 while (!list_empty(&list)) {
8599 de = list_first_entry(&list, struct io_defer_entry, list);
8600 list_del_init(&de->list);
8601 req_set_fail_links(de->req);
8602 io_put_req(de->req);
8603 io_req_complete(de->req, -ECANCELED);
8608 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
8609 struct task_struct *task,
8610 struct files_struct *files)
8612 struct io_task_cancel cancel = { .task = task, .files = files, };
8613 struct io_uring_task *tctx = current->io_uring;
8616 enum io_wq_cancel cret;
8619 if (tctx && tctx->io_wq) {
8620 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
8622 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8625 /* SQPOLL thread does its own polling */
8626 if (!(ctx->flags & IORING_SETUP_SQPOLL) && !files) {
8627 while (!list_empty_careful(&ctx->iopoll_list)) {
8628 io_iopoll_try_reap_events(ctx);
8633 ret |= io_poll_remove_all(ctx, task, files);
8634 ret |= io_kill_timeouts(ctx, task, files);
8635 ret |= io_run_task_work();
8636 ret |= io_run_ctx_fallback(ctx);
8637 io_cqring_overflow_flush(ctx, true, task, files);
8644 static int io_uring_count_inflight(struct io_ring_ctx *ctx,
8645 struct task_struct *task,
8646 struct files_struct *files)
8648 struct io_kiocb *req;
8651 spin_lock_irq(&ctx->inflight_lock);
8652 list_for_each_entry(req, &ctx->inflight_list, inflight_entry)
8653 cnt += io_match_task(req, task, files);
8654 spin_unlock_irq(&ctx->inflight_lock);
8658 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
8659 struct task_struct *task,
8660 struct files_struct *files)
8662 while (!list_empty_careful(&ctx->inflight_list)) {
8666 inflight = io_uring_count_inflight(ctx, task, files);
8670 io_uring_try_cancel_requests(ctx, task, files);
8673 io_sq_thread_unpark(ctx->sq_data);
8674 prepare_to_wait(&task->io_uring->wait, &wait,
8675 TASK_UNINTERRUPTIBLE);
8676 if (inflight == io_uring_count_inflight(ctx, task, files))
8678 finish_wait(&task->io_uring->wait, &wait);
8680 io_sq_thread_park(ctx->sq_data);
8685 * We need to iteratively cancel requests, in case a request has dependent
8686 * hard links. These persist even for failure of cancelations, hence keep
8687 * looping until none are found.
8689 static void io_uring_cancel_task_requests(struct io_ring_ctx *ctx,
8690 struct files_struct *files)
8692 struct task_struct *task = current;
8693 bool did_park = false;
8695 if ((ctx->flags & IORING_SETUP_SQPOLL) && ctx->sq_data) {
8696 /* never started, nothing to cancel */
8697 if (ctx->flags & IORING_SETUP_R_DISABLED) {
8698 io_sq_offload_start(ctx);
8701 did_park = io_sq_thread_park(ctx->sq_data);
8703 task = ctx->sq_data->thread;
8704 atomic_inc(&task->io_uring->in_idle);
8708 io_cancel_defer_files(ctx, task, files);
8710 io_uring_cancel_files(ctx, task, files);
8712 io_uring_try_cancel_requests(ctx, task, NULL);
8715 atomic_dec(&task->io_uring->in_idle);
8716 io_sq_thread_unpark(ctx->sq_data);
8721 * Note that this task has used io_uring. We use it for cancelation purposes.
8723 static int io_uring_add_task_file(struct io_ring_ctx *ctx, struct file *file)
8725 struct io_uring_task *tctx = current->io_uring;
8728 if (unlikely(!tctx)) {
8729 ret = io_uring_alloc_task_context(current, ctx);
8732 tctx = current->io_uring;
8734 if (tctx->last != file) {
8735 void *old = xa_load(&tctx->xa, (unsigned long)file);
8739 ret = xa_err(xa_store(&tctx->xa, (unsigned long)file,
8750 * This is race safe in that the task itself is doing this, hence it
8751 * cannot be going through the exit/cancel paths at the same time.
8752 * This cannot be modified while exit/cancel is running.
8754 if (!tctx->sqpoll && (ctx->flags & IORING_SETUP_SQPOLL))
8755 tctx->sqpoll = true;
8761 * Remove this io_uring_file -> task mapping.
8763 static void io_uring_del_task_file(struct file *file)
8765 struct io_uring_task *tctx = current->io_uring;
8767 if (tctx->last == file)
8769 file = xa_erase(&tctx->xa, (unsigned long)file);
8774 static void io_uring_clean_tctx(struct io_uring_task *tctx)
8777 unsigned long index;
8779 xa_for_each(&tctx->xa, index, file)
8780 io_uring_del_task_file(file);
8782 io_wq_put_and_exit(tctx->io_wq);
8787 void __io_uring_files_cancel(struct files_struct *files)
8789 struct io_uring_task *tctx = current->io_uring;
8791 unsigned long index;
8793 /* make sure overflow events are dropped */
8794 atomic_inc(&tctx->in_idle);
8795 xa_for_each(&tctx->xa, index, file)
8796 io_uring_cancel_task_requests(file->private_data, files);
8797 atomic_dec(&tctx->in_idle);
8800 io_uring_clean_tctx(tctx);
8803 static s64 tctx_inflight(struct io_uring_task *tctx)
8805 return percpu_counter_sum(&tctx->inflight);
8808 static void io_uring_cancel_sqpoll(struct io_ring_ctx *ctx)
8810 struct io_sq_data *sqd = ctx->sq_data;
8811 struct io_uring_task *tctx;
8817 if (!io_sq_thread_park(sqd))
8819 tctx = ctx->sq_data->thread->io_uring;
8820 /* can happen on fork/alloc failure, just ignore that state */
8822 io_sq_thread_unpark(sqd);
8826 atomic_inc(&tctx->in_idle);
8828 /* read completions before cancelations */
8829 inflight = tctx_inflight(tctx);
8832 io_uring_cancel_task_requests(ctx, NULL);
8834 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
8836 * If we've seen completions, retry without waiting. This
8837 * avoids a race where a completion comes in before we did
8838 * prepare_to_wait().
8840 if (inflight == tctx_inflight(tctx))
8842 finish_wait(&tctx->wait, &wait);
8844 atomic_dec(&tctx->in_idle);
8845 io_sq_thread_unpark(sqd);
8849 * Find any io_uring fd that this task has registered or done IO on, and cancel
8852 void __io_uring_task_cancel(void)
8854 struct io_uring_task *tctx = current->io_uring;
8858 /* make sure overflow events are dropped */
8859 atomic_inc(&tctx->in_idle);
8863 unsigned long index;
8865 xa_for_each(&tctx->xa, index, file)
8866 io_uring_cancel_sqpoll(file->private_data);
8870 /* read completions before cancelations */
8871 inflight = tctx_inflight(tctx);
8874 __io_uring_files_cancel(NULL);
8876 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
8879 * If we've seen completions, retry without waiting. This
8880 * avoids a race where a completion comes in before we did
8881 * prepare_to_wait().
8883 if (inflight == tctx_inflight(tctx))
8885 finish_wait(&tctx->wait, &wait);
8888 atomic_dec(&tctx->in_idle);
8890 io_uring_clean_tctx(tctx);
8891 /* all current's requests should be gone, we can kill tctx */
8892 __io_uring_free(current);
8895 static void *io_uring_validate_mmap_request(struct file *file,
8896 loff_t pgoff, size_t sz)
8898 struct io_ring_ctx *ctx = file->private_data;
8899 loff_t offset = pgoff << PAGE_SHIFT;
8904 case IORING_OFF_SQ_RING:
8905 case IORING_OFF_CQ_RING:
8908 case IORING_OFF_SQES:
8912 return ERR_PTR(-EINVAL);
8915 page = virt_to_head_page(ptr);
8916 if (sz > page_size(page))
8917 return ERR_PTR(-EINVAL);
8924 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
8926 size_t sz = vma->vm_end - vma->vm_start;
8930 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
8932 return PTR_ERR(ptr);
8934 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
8935 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
8938 #else /* !CONFIG_MMU */
8940 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
8942 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
8945 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
8947 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
8950 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
8951 unsigned long addr, unsigned long len,
8952 unsigned long pgoff, unsigned long flags)
8956 ptr = io_uring_validate_mmap_request(file, pgoff, len);
8958 return PTR_ERR(ptr);
8960 return (unsigned long) ptr;
8963 #endif /* !CONFIG_MMU */
8965 static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
8971 if (!io_sqring_full(ctx))
8973 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
8975 if (!io_sqring_full(ctx))
8978 } while (!signal_pending(current));
8980 finish_wait(&ctx->sqo_sq_wait, &wait);
8984 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
8985 struct __kernel_timespec __user **ts,
8986 const sigset_t __user **sig)
8988 struct io_uring_getevents_arg arg;
8991 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
8992 * is just a pointer to the sigset_t.
8994 if (!(flags & IORING_ENTER_EXT_ARG)) {
8995 *sig = (const sigset_t __user *) argp;
9001 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9002 * timespec and sigset_t pointers if good.
9004 if (*argsz != sizeof(arg))
9006 if (copy_from_user(&arg, argp, sizeof(arg)))
9008 *sig = u64_to_user_ptr(arg.sigmask);
9009 *argsz = arg.sigmask_sz;
9010 *ts = u64_to_user_ptr(arg.ts);
9014 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9015 u32, min_complete, u32, flags, const void __user *, argp,
9018 struct io_ring_ctx *ctx;
9025 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9026 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG))
9034 if (f.file->f_op != &io_uring_fops)
9038 ctx = f.file->private_data;
9039 if (!percpu_ref_tryget(&ctx->refs))
9043 if (ctx->flags & IORING_SETUP_R_DISABLED)
9047 * For SQ polling, the thread will do all submissions and completions.
9048 * Just return the requested submit count, and wake the thread if
9052 if (ctx->flags & IORING_SETUP_SQPOLL) {
9053 io_cqring_overflow_flush(ctx, false, NULL, NULL);
9055 if (unlikely(ctx->sqo_exec)) {
9056 ret = io_sq_thread_fork(ctx->sq_data, ctx);
9062 if (flags & IORING_ENTER_SQ_WAKEUP)
9063 wake_up(&ctx->sq_data->wait);
9064 if (flags & IORING_ENTER_SQ_WAIT) {
9065 ret = io_sqpoll_wait_sq(ctx);
9069 submitted = to_submit;
9070 } else if (to_submit) {
9071 ret = io_uring_add_task_file(ctx, f.file);
9074 mutex_lock(&ctx->uring_lock);
9075 submitted = io_submit_sqes(ctx, to_submit);
9076 mutex_unlock(&ctx->uring_lock);
9078 if (submitted != to_submit)
9081 if (flags & IORING_ENTER_GETEVENTS) {
9082 const sigset_t __user *sig;
9083 struct __kernel_timespec __user *ts;
9085 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
9089 min_complete = min(min_complete, ctx->cq_entries);
9092 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9093 * space applications don't need to do io completion events
9094 * polling again, they can rely on io_sq_thread to do polling
9095 * work, which can reduce cpu usage and uring_lock contention.
9097 if (ctx->flags & IORING_SETUP_IOPOLL &&
9098 !(ctx->flags & IORING_SETUP_SQPOLL)) {
9099 ret = io_iopoll_check(ctx, min_complete);
9101 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
9106 percpu_ref_put(&ctx->refs);
9109 return submitted ? submitted : ret;
9112 #ifdef CONFIG_PROC_FS
9113 static int io_uring_show_cred(int id, void *p, void *data)
9115 const struct cred *cred = p;
9116 struct seq_file *m = data;
9117 struct user_namespace *uns = seq_user_ns(m);
9118 struct group_info *gi;
9123 seq_printf(m, "%5d\n", id);
9124 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
9125 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
9126 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
9127 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
9128 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
9129 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
9130 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
9131 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
9132 seq_puts(m, "\n\tGroups:\t");
9133 gi = cred->group_info;
9134 for (g = 0; g < gi->ngroups; g++) {
9135 seq_put_decimal_ull(m, g ? " " : "",
9136 from_kgid_munged(uns, gi->gid[g]));
9138 seq_puts(m, "\n\tCapEff:\t");
9139 cap = cred->cap_effective;
9140 CAP_FOR_EACH_U32(__capi)
9141 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
9146 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
9148 struct io_sq_data *sq = NULL;
9153 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9154 * since fdinfo case grabs it in the opposite direction of normal use
9155 * cases. If we fail to get the lock, we just don't iterate any
9156 * structures that could be going away outside the io_uring mutex.
9158 has_lock = mutex_trylock(&ctx->uring_lock);
9160 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) {
9166 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9167 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9168 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9169 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9170 struct file *f = *io_fixed_file_slot(ctx->file_data, i);
9173 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9175 seq_printf(m, "%5u: <none>\n", i);
9177 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9178 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9179 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
9181 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
9182 (unsigned int) buf->len);
9184 if (has_lock && !idr_is_empty(&ctx->personality_idr)) {
9185 seq_printf(m, "Personalities:\n");
9186 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
9188 seq_printf(m, "PollList:\n");
9189 spin_lock_irq(&ctx->completion_lock);
9190 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9191 struct hlist_head *list = &ctx->cancel_hash[i];
9192 struct io_kiocb *req;
9194 hlist_for_each_entry(req, list, hash_node)
9195 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9196 req->task->task_works != NULL);
9198 spin_unlock_irq(&ctx->completion_lock);
9200 mutex_unlock(&ctx->uring_lock);
9203 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9205 struct io_ring_ctx *ctx = f->private_data;
9207 if (percpu_ref_tryget(&ctx->refs)) {
9208 __io_uring_show_fdinfo(ctx, m);
9209 percpu_ref_put(&ctx->refs);
9214 static const struct file_operations io_uring_fops = {
9215 .release = io_uring_release,
9216 .mmap = io_uring_mmap,
9218 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9219 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9221 .poll = io_uring_poll,
9222 .fasync = io_uring_fasync,
9223 #ifdef CONFIG_PROC_FS
9224 .show_fdinfo = io_uring_show_fdinfo,
9228 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9229 struct io_uring_params *p)
9231 struct io_rings *rings;
9232 size_t size, sq_array_offset;
9234 /* make sure these are sane, as we already accounted them */
9235 ctx->sq_entries = p->sq_entries;
9236 ctx->cq_entries = p->cq_entries;
9238 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9239 if (size == SIZE_MAX)
9242 rings = io_mem_alloc(size);
9247 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9248 rings->sq_ring_mask = p->sq_entries - 1;
9249 rings->cq_ring_mask = p->cq_entries - 1;
9250 rings->sq_ring_entries = p->sq_entries;
9251 rings->cq_ring_entries = p->cq_entries;
9252 ctx->sq_mask = rings->sq_ring_mask;
9253 ctx->cq_mask = rings->cq_ring_mask;
9255 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9256 if (size == SIZE_MAX) {
9257 io_mem_free(ctx->rings);
9262 ctx->sq_sqes = io_mem_alloc(size);
9263 if (!ctx->sq_sqes) {
9264 io_mem_free(ctx->rings);
9272 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
9276 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9280 ret = io_uring_add_task_file(ctx, file);
9285 fd_install(fd, file);
9290 * Allocate an anonymous fd, this is what constitutes the application
9291 * visible backing of an io_uring instance. The application mmaps this
9292 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9293 * we have to tie this fd to a socket for file garbage collection purposes.
9295 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
9298 #if defined(CONFIG_UNIX)
9301 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9304 return ERR_PTR(ret);
9307 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9308 O_RDWR | O_CLOEXEC);
9309 #if defined(CONFIG_UNIX)
9311 sock_release(ctx->ring_sock);
9312 ctx->ring_sock = NULL;
9314 ctx->ring_sock->file = file;
9320 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9321 struct io_uring_params __user *params)
9323 struct io_ring_ctx *ctx;
9329 if (entries > IORING_MAX_ENTRIES) {
9330 if (!(p->flags & IORING_SETUP_CLAMP))
9332 entries = IORING_MAX_ENTRIES;
9336 * Use twice as many entries for the CQ ring. It's possible for the
9337 * application to drive a higher depth than the size of the SQ ring,
9338 * since the sqes are only used at submission time. This allows for
9339 * some flexibility in overcommitting a bit. If the application has
9340 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9341 * of CQ ring entries manually.
9343 p->sq_entries = roundup_pow_of_two(entries);
9344 if (p->flags & IORING_SETUP_CQSIZE) {
9346 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9347 * to a power-of-two, if it isn't already. We do NOT impose
9348 * any cq vs sq ring sizing.
9352 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9353 if (!(p->flags & IORING_SETUP_CLAMP))
9355 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9357 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9358 if (p->cq_entries < p->sq_entries)
9361 p->cq_entries = 2 * p->sq_entries;
9364 ctx = io_ring_ctx_alloc(p);
9367 ctx->compat = in_compat_syscall();
9368 if (!capable(CAP_IPC_LOCK))
9369 ctx->user = get_uid(current_user());
9372 * This is just grabbed for accounting purposes. When a process exits,
9373 * the mm is exited and dropped before the files, hence we need to hang
9374 * on to this mm purely for the purposes of being able to unaccount
9375 * memory (locked/pinned vm). It's not used for anything else.
9377 mmgrab(current->mm);
9378 ctx->mm_account = current->mm;
9380 ret = io_allocate_scq_urings(ctx, p);
9384 ret = io_sq_offload_create(ctx, p);
9388 if (!(p->flags & IORING_SETUP_R_DISABLED))
9389 io_sq_offload_start(ctx);
9391 memset(&p->sq_off, 0, sizeof(p->sq_off));
9392 p->sq_off.head = offsetof(struct io_rings, sq.head);
9393 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9394 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9395 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9396 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9397 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9398 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9400 memset(&p->cq_off, 0, sizeof(p->cq_off));
9401 p->cq_off.head = offsetof(struct io_rings, cq.head);
9402 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9403 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9404 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9405 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9406 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9407 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9409 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9410 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9411 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9412 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
9413 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS;
9415 if (copy_to_user(params, p, sizeof(*p))) {
9420 file = io_uring_get_file(ctx);
9422 ret = PTR_ERR(file);
9427 * Install ring fd as the very last thing, so we don't risk someone
9428 * having closed it before we finish setup
9430 ret = io_uring_install_fd(ctx, file);
9432 /* fput will clean it up */
9437 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9440 io_ring_ctx_wait_and_kill(ctx);
9445 * Sets up an aio uring context, and returns the fd. Applications asks for a
9446 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9447 * params structure passed in.
9449 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9451 struct io_uring_params p;
9454 if (copy_from_user(&p, params, sizeof(p)))
9456 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9461 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9462 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9463 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9464 IORING_SETUP_R_DISABLED))
9467 return io_uring_create(entries, &p, params);
9470 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9471 struct io_uring_params __user *, params)
9473 return io_uring_setup(entries, params);
9476 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9478 struct io_uring_probe *p;
9482 size = struct_size(p, ops, nr_args);
9483 if (size == SIZE_MAX)
9485 p = kzalloc(size, GFP_KERNEL);
9490 if (copy_from_user(p, arg, size))
9493 if (memchr_inv(p, 0, size))
9496 p->last_op = IORING_OP_LAST - 1;
9497 if (nr_args > IORING_OP_LAST)
9498 nr_args = IORING_OP_LAST;
9500 for (i = 0; i < nr_args; i++) {
9502 if (!io_op_defs[i].not_supported)
9503 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9508 if (copy_to_user(arg, p, size))
9515 static int io_register_personality(struct io_ring_ctx *ctx)
9517 const struct cred *creds;
9520 creds = get_current_cred();
9522 ret = idr_alloc_cyclic(&ctx->personality_idr, (void *) creds, 1,
9523 USHRT_MAX, GFP_KERNEL);
9529 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9530 unsigned int nr_args)
9532 struct io_uring_restriction *res;
9536 /* Restrictions allowed only if rings started disabled */
9537 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9540 /* We allow only a single restrictions registration */
9541 if (ctx->restrictions.registered)
9544 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9547 size = array_size(nr_args, sizeof(*res));
9548 if (size == SIZE_MAX)
9551 res = memdup_user(arg, size);
9553 return PTR_ERR(res);
9557 for (i = 0; i < nr_args; i++) {
9558 switch (res[i].opcode) {
9559 case IORING_RESTRICTION_REGISTER_OP:
9560 if (res[i].register_op >= IORING_REGISTER_LAST) {
9565 __set_bit(res[i].register_op,
9566 ctx->restrictions.register_op);
9568 case IORING_RESTRICTION_SQE_OP:
9569 if (res[i].sqe_op >= IORING_OP_LAST) {
9574 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9576 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9577 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9579 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9580 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9589 /* Reset all restrictions if an error happened */
9591 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9593 ctx->restrictions.registered = true;
9599 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9601 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9604 if (ctx->restrictions.registered)
9605 ctx->restricted = 1;
9607 io_sq_offload_start(ctx);
9611 static bool io_register_op_must_quiesce(int op)
9614 case IORING_UNREGISTER_FILES:
9615 case IORING_REGISTER_FILES_UPDATE:
9616 case IORING_REGISTER_PROBE:
9617 case IORING_REGISTER_PERSONALITY:
9618 case IORING_UNREGISTER_PERSONALITY:
9625 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
9626 void __user *arg, unsigned nr_args)
9627 __releases(ctx->uring_lock)
9628 __acquires(ctx->uring_lock)
9633 * We're inside the ring mutex, if the ref is already dying, then
9634 * someone else killed the ctx or is already going through
9635 * io_uring_register().
9637 if (percpu_ref_is_dying(&ctx->refs))
9640 if (io_register_op_must_quiesce(opcode)) {
9641 percpu_ref_kill(&ctx->refs);
9644 * Drop uring mutex before waiting for references to exit. If
9645 * another thread is currently inside io_uring_enter() it might
9646 * need to grab the uring_lock to make progress. If we hold it
9647 * here across the drain wait, then we can deadlock. It's safe
9648 * to drop the mutex here, since no new references will come in
9649 * after we've killed the percpu ref.
9651 mutex_unlock(&ctx->uring_lock);
9653 ret = wait_for_completion_interruptible(&ctx->ref_comp);
9656 ret = io_run_task_work_sig();
9661 mutex_lock(&ctx->uring_lock);
9664 percpu_ref_resurrect(&ctx->refs);
9669 if (ctx->restricted) {
9670 if (opcode >= IORING_REGISTER_LAST) {
9675 if (!test_bit(opcode, ctx->restrictions.register_op)) {
9682 case IORING_REGISTER_BUFFERS:
9683 ret = io_sqe_buffers_register(ctx, arg, nr_args);
9685 case IORING_UNREGISTER_BUFFERS:
9689 ret = io_sqe_buffers_unregister(ctx);
9691 case IORING_REGISTER_FILES:
9692 ret = io_sqe_files_register(ctx, arg, nr_args);
9694 case IORING_UNREGISTER_FILES:
9698 ret = io_sqe_files_unregister(ctx);
9700 case IORING_REGISTER_FILES_UPDATE:
9701 ret = io_sqe_files_update(ctx, arg, nr_args);
9703 case IORING_REGISTER_EVENTFD:
9704 case IORING_REGISTER_EVENTFD_ASYNC:
9708 ret = io_eventfd_register(ctx, arg);
9711 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
9712 ctx->eventfd_async = 1;
9714 ctx->eventfd_async = 0;
9716 case IORING_UNREGISTER_EVENTFD:
9720 ret = io_eventfd_unregister(ctx);
9722 case IORING_REGISTER_PROBE:
9724 if (!arg || nr_args > 256)
9726 ret = io_probe(ctx, arg, nr_args);
9728 case IORING_REGISTER_PERSONALITY:
9732 ret = io_register_personality(ctx);
9734 case IORING_UNREGISTER_PERSONALITY:
9738 ret = io_unregister_personality(ctx, nr_args);
9740 case IORING_REGISTER_ENABLE_RINGS:
9744 ret = io_register_enable_rings(ctx);
9746 case IORING_REGISTER_RESTRICTIONS:
9747 ret = io_register_restrictions(ctx, arg, nr_args);
9755 if (io_register_op_must_quiesce(opcode)) {
9756 /* bring the ctx back to life */
9757 percpu_ref_reinit(&ctx->refs);
9759 reinit_completion(&ctx->ref_comp);
9764 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
9765 void __user *, arg, unsigned int, nr_args)
9767 struct io_ring_ctx *ctx;
9776 if (f.file->f_op != &io_uring_fops)
9779 ctx = f.file->private_data;
9783 mutex_lock(&ctx->uring_lock);
9784 ret = __io_uring_register(ctx, opcode, arg, nr_args);
9785 mutex_unlock(&ctx->uring_lock);
9786 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
9787 ctx->cq_ev_fd != NULL, ret);
9793 static int __init io_uring_init(void)
9795 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
9796 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
9797 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
9800 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
9801 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
9802 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
9803 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
9804 BUILD_BUG_SQE_ELEM(1, __u8, flags);
9805 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
9806 BUILD_BUG_SQE_ELEM(4, __s32, fd);
9807 BUILD_BUG_SQE_ELEM(8, __u64, off);
9808 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
9809 BUILD_BUG_SQE_ELEM(16, __u64, addr);
9810 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
9811 BUILD_BUG_SQE_ELEM(24, __u32, len);
9812 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
9813 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
9814 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
9815 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
9816 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
9817 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
9818 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
9819 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
9820 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
9821 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
9822 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
9823 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
9824 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
9825 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
9826 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
9827 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
9828 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
9829 BUILD_BUG_SQE_ELEM(42, __u16, personality);
9830 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
9832 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
9833 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
9834 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
9838 __initcall(io_uring_init);