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_track_inflight(struct io_kiocb *req)
1172 struct io_ring_ctx *ctx = req->ctx;
1174 if (!(req->flags & REQ_F_INFLIGHT)) {
1175 req->flags |= REQ_F_INFLIGHT;
1177 spin_lock_irq(&ctx->inflight_lock);
1178 list_add(&req->inflight_entry, &ctx->inflight_list);
1179 spin_unlock_irq(&ctx->inflight_lock);
1183 static void io_prep_async_work(struct io_kiocb *req)
1185 const struct io_op_def *def = &io_op_defs[req->opcode];
1186 struct io_ring_ctx *ctx = req->ctx;
1188 if (req->flags & REQ_F_FORCE_ASYNC)
1189 req->work.flags |= IO_WQ_WORK_CONCURRENT;
1191 if (req->flags & REQ_F_ISREG) {
1192 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1193 io_wq_hash_work(&req->work, file_inode(req->file));
1195 if (def->unbound_nonreg_file)
1196 req->work.flags |= IO_WQ_WORK_UNBOUND;
1200 static void io_prep_async_link(struct io_kiocb *req)
1202 struct io_kiocb *cur;
1204 io_for_each_link(cur, req)
1205 io_prep_async_work(cur);
1208 static void io_queue_async_work(struct io_kiocb *req)
1210 struct io_ring_ctx *ctx = req->ctx;
1211 struct io_kiocb *link = io_prep_linked_timeout(req);
1212 struct io_uring_task *tctx = req->task->io_uring;
1215 BUG_ON(!tctx->io_wq);
1217 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1218 &req->work, req->flags);
1219 /* init ->work of the whole link before punting */
1220 io_prep_async_link(req);
1221 io_wq_enqueue(tctx->io_wq, &req->work);
1223 io_queue_linked_timeout(link);
1226 static void io_kill_timeout(struct io_kiocb *req)
1228 struct io_timeout_data *io = req->async_data;
1231 ret = hrtimer_try_to_cancel(&io->timer);
1233 atomic_set(&req->ctx->cq_timeouts,
1234 atomic_read(&req->ctx->cq_timeouts) + 1);
1235 list_del_init(&req->timeout.list);
1236 io_cqring_fill_event(req, 0);
1237 io_put_req_deferred(req, 1);
1242 * Returns true if we found and killed one or more timeouts
1244 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
1245 struct files_struct *files)
1247 struct io_kiocb *req, *tmp;
1250 spin_lock_irq(&ctx->completion_lock);
1251 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
1252 if (io_match_task(req, tsk, files)) {
1253 io_kill_timeout(req);
1257 spin_unlock_irq(&ctx->completion_lock);
1258 return canceled != 0;
1261 static void __io_queue_deferred(struct io_ring_ctx *ctx)
1264 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1265 struct io_defer_entry, list);
1267 if (req_need_defer(de->req, de->seq))
1269 list_del_init(&de->list);
1270 io_req_task_queue(de->req);
1272 } while (!list_empty(&ctx->defer_list));
1275 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1279 if (list_empty(&ctx->timeout_list))
1282 seq = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
1285 u32 events_needed, events_got;
1286 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1287 struct io_kiocb, timeout.list);
1289 if (io_is_timeout_noseq(req))
1293 * Since seq can easily wrap around over time, subtract
1294 * the last seq at which timeouts were flushed before comparing.
1295 * Assuming not more than 2^31-1 events have happened since,
1296 * these subtractions won't have wrapped, so we can check if
1297 * target is in [last_seq, current_seq] by comparing the two.
1299 events_needed = req->timeout.target_seq - ctx->cq_last_tm_flush;
1300 events_got = seq - ctx->cq_last_tm_flush;
1301 if (events_got < events_needed)
1304 list_del_init(&req->timeout.list);
1305 io_kill_timeout(req);
1306 } while (!list_empty(&ctx->timeout_list));
1308 ctx->cq_last_tm_flush = seq;
1311 static void io_commit_cqring(struct io_ring_ctx *ctx)
1313 io_flush_timeouts(ctx);
1315 /* order cqe stores with ring update */
1316 smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
1318 if (unlikely(!list_empty(&ctx->defer_list)))
1319 __io_queue_deferred(ctx);
1322 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1324 struct io_rings *r = ctx->rings;
1326 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == r->sq_ring_entries;
1329 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
1331 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
1334 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1336 struct io_rings *rings = ctx->rings;
1340 * writes to the cq entry need to come after reading head; the
1341 * control dependency is enough as we're using WRITE_ONCE to
1344 if (__io_cqring_events(ctx) == rings->cq_ring_entries)
1347 tail = ctx->cached_cq_tail++;
1348 return &rings->cqes[tail & ctx->cq_mask];
1351 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1355 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1357 if (!ctx->eventfd_async)
1359 return io_wq_current_is_worker();
1362 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1364 /* see waitqueue_active() comment */
1367 if (waitqueue_active(&ctx->wait))
1368 wake_up(&ctx->wait);
1369 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1370 wake_up(&ctx->sq_data->wait);
1371 if (io_should_trigger_evfd(ctx))
1372 eventfd_signal(ctx->cq_ev_fd, 1);
1373 if (waitqueue_active(&ctx->cq_wait)) {
1374 wake_up_interruptible(&ctx->cq_wait);
1375 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1379 static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx)
1381 /* see waitqueue_active() comment */
1384 if (ctx->flags & IORING_SETUP_SQPOLL) {
1385 if (waitqueue_active(&ctx->wait))
1386 wake_up(&ctx->wait);
1388 if (io_should_trigger_evfd(ctx))
1389 eventfd_signal(ctx->cq_ev_fd, 1);
1390 if (waitqueue_active(&ctx->cq_wait)) {
1391 wake_up_interruptible(&ctx->cq_wait);
1392 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1396 /* Returns true if there are no backlogged entries after the flush */
1397 static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force,
1398 struct task_struct *tsk,
1399 struct files_struct *files)
1401 struct io_rings *rings = ctx->rings;
1402 struct io_kiocb *req, *tmp;
1403 struct io_uring_cqe *cqe;
1404 unsigned long flags;
1405 bool all_flushed, posted;
1408 if (!force && __io_cqring_events(ctx) == rings->cq_ring_entries)
1412 spin_lock_irqsave(&ctx->completion_lock, flags);
1413 list_for_each_entry_safe(req, tmp, &ctx->cq_overflow_list, compl.list) {
1414 if (!io_match_task(req, tsk, files))
1417 cqe = io_get_cqring(ctx);
1421 list_move(&req->compl.list, &list);
1423 WRITE_ONCE(cqe->user_data, req->user_data);
1424 WRITE_ONCE(cqe->res, req->result);
1425 WRITE_ONCE(cqe->flags, req->compl.cflags);
1427 ctx->cached_cq_overflow++;
1428 WRITE_ONCE(ctx->rings->cq_overflow,
1429 ctx->cached_cq_overflow);
1434 all_flushed = list_empty(&ctx->cq_overflow_list);
1436 clear_bit(0, &ctx->sq_check_overflow);
1437 clear_bit(0, &ctx->cq_check_overflow);
1438 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
1442 io_commit_cqring(ctx);
1443 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1445 io_cqring_ev_posted(ctx);
1447 while (!list_empty(&list)) {
1448 req = list_first_entry(&list, struct io_kiocb, compl.list);
1449 list_del(&req->compl.list);
1456 static void io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force,
1457 struct task_struct *tsk,
1458 struct files_struct *files)
1460 if (test_bit(0, &ctx->cq_check_overflow)) {
1461 /* iopoll syncs against uring_lock, not completion_lock */
1462 if (ctx->flags & IORING_SETUP_IOPOLL)
1463 mutex_lock(&ctx->uring_lock);
1464 __io_cqring_overflow_flush(ctx, force, tsk, files);
1465 if (ctx->flags & IORING_SETUP_IOPOLL)
1466 mutex_unlock(&ctx->uring_lock);
1470 static void __io_cqring_fill_event(struct io_kiocb *req, long res, long cflags)
1472 struct io_ring_ctx *ctx = req->ctx;
1473 struct io_uring_cqe *cqe;
1475 trace_io_uring_complete(ctx, req->user_data, res);
1478 * If we can't get a cq entry, userspace overflowed the
1479 * submission (by quite a lot). Increment the overflow count in
1482 cqe = io_get_cqring(ctx);
1484 WRITE_ONCE(cqe->user_data, req->user_data);
1485 WRITE_ONCE(cqe->res, res);
1486 WRITE_ONCE(cqe->flags, cflags);
1487 } else if (ctx->cq_overflow_flushed ||
1488 atomic_read(&req->task->io_uring->in_idle)) {
1490 * If we're in ring overflow flush mode, or in task cancel mode,
1491 * then we cannot store the request for later flushing, we need
1492 * to drop it on the floor.
1494 ctx->cached_cq_overflow++;
1495 WRITE_ONCE(ctx->rings->cq_overflow, ctx->cached_cq_overflow);
1497 if (list_empty(&ctx->cq_overflow_list)) {
1498 set_bit(0, &ctx->sq_check_overflow);
1499 set_bit(0, &ctx->cq_check_overflow);
1500 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
1504 req->compl.cflags = cflags;
1505 refcount_inc(&req->refs);
1506 list_add_tail(&req->compl.list, &ctx->cq_overflow_list);
1510 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1512 __io_cqring_fill_event(req, res, 0);
1515 static inline void io_req_complete_post(struct io_kiocb *req, long res,
1516 unsigned int cflags)
1518 struct io_ring_ctx *ctx = req->ctx;
1519 unsigned long flags;
1521 spin_lock_irqsave(&ctx->completion_lock, flags);
1522 __io_cqring_fill_event(req, res, cflags);
1523 io_commit_cqring(ctx);
1525 * If we're the last reference to this request, add to our locked
1528 if (refcount_dec_and_test(&req->refs)) {
1529 struct io_comp_state *cs = &ctx->submit_state.comp;
1531 io_dismantle_req(req);
1532 io_put_task(req->task, 1);
1533 list_add(&req->compl.list, &cs->locked_free_list);
1534 cs->locked_free_nr++;
1537 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1539 io_cqring_ev_posted(ctx);
1542 percpu_ref_put(&ctx->refs);
1546 static void io_req_complete_state(struct io_kiocb *req, long res,
1547 unsigned int cflags)
1551 req->compl.cflags = cflags;
1552 req->flags |= REQ_F_COMPLETE_INLINE;
1555 static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags,
1556 long res, unsigned cflags)
1558 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1559 io_req_complete_state(req, res, cflags);
1561 io_req_complete_post(req, res, cflags);
1564 static inline void io_req_complete(struct io_kiocb *req, long res)
1566 __io_req_complete(req, 0, res, 0);
1569 static bool io_flush_cached_reqs(struct io_ring_ctx *ctx)
1571 struct io_submit_state *state = &ctx->submit_state;
1572 struct io_comp_state *cs = &state->comp;
1573 struct io_kiocb *req = NULL;
1576 * If we have more than a batch's worth of requests in our IRQ side
1577 * locked cache, grab the lock and move them over to our submission
1580 if (READ_ONCE(cs->locked_free_nr) > IO_COMPL_BATCH) {
1581 spin_lock_irq(&ctx->completion_lock);
1582 list_splice_init(&cs->locked_free_list, &cs->free_list);
1583 cs->locked_free_nr = 0;
1584 spin_unlock_irq(&ctx->completion_lock);
1587 while (!list_empty(&cs->free_list)) {
1588 req = list_first_entry(&cs->free_list, struct io_kiocb,
1590 list_del(&req->compl.list);
1591 state->reqs[state->free_reqs++] = req;
1592 if (state->free_reqs == ARRAY_SIZE(state->reqs))
1599 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
1601 struct io_submit_state *state = &ctx->submit_state;
1603 BUILD_BUG_ON(IO_REQ_ALLOC_BATCH > ARRAY_SIZE(state->reqs));
1605 if (!state->free_reqs) {
1606 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1609 if (io_flush_cached_reqs(ctx))
1612 ret = kmem_cache_alloc_bulk(req_cachep, gfp, IO_REQ_ALLOC_BATCH,
1616 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1617 * retry single alloc to be on the safe side.
1619 if (unlikely(ret <= 0)) {
1620 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1621 if (!state->reqs[0])
1625 state->free_reqs = ret;
1629 return state->reqs[state->free_reqs];
1632 static inline void io_put_file(struct io_kiocb *req, struct file *file,
1639 static void io_dismantle_req(struct io_kiocb *req)
1643 if (req->async_data)
1644 kfree(req->async_data);
1646 io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
1647 if (req->fixed_rsrc_refs)
1648 percpu_ref_put(req->fixed_rsrc_refs);
1650 if (req->flags & REQ_F_INFLIGHT) {
1651 struct io_ring_ctx *ctx = req->ctx;
1652 unsigned long flags;
1654 spin_lock_irqsave(&ctx->inflight_lock, flags);
1655 list_del(&req->inflight_entry);
1656 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
1657 req->flags &= ~REQ_F_INFLIGHT;
1661 /* must to be called somewhat shortly after putting a request */
1662 static inline void io_put_task(struct task_struct *task, int nr)
1664 struct io_uring_task *tctx = task->io_uring;
1666 percpu_counter_sub(&tctx->inflight, nr);
1667 if (unlikely(atomic_read(&tctx->in_idle)))
1668 wake_up(&tctx->wait);
1669 put_task_struct_many(task, nr);
1672 static void __io_free_req(struct io_kiocb *req)
1674 struct io_ring_ctx *ctx = req->ctx;
1676 io_dismantle_req(req);
1677 io_put_task(req->task, 1);
1679 kmem_cache_free(req_cachep, req);
1680 percpu_ref_put(&ctx->refs);
1683 static inline void io_remove_next_linked(struct io_kiocb *req)
1685 struct io_kiocb *nxt = req->link;
1687 req->link = nxt->link;
1691 static void io_kill_linked_timeout(struct io_kiocb *req)
1693 struct io_ring_ctx *ctx = req->ctx;
1694 struct io_kiocb *link;
1695 bool cancelled = false;
1696 unsigned long flags;
1698 spin_lock_irqsave(&ctx->completion_lock, flags);
1702 * Can happen if a linked timeout fired and link had been like
1703 * req -> link t-out -> link t-out [-> ...]
1705 if (link && (link->flags & REQ_F_LTIMEOUT_ACTIVE)) {
1706 struct io_timeout_data *io = link->async_data;
1709 io_remove_next_linked(req);
1710 link->timeout.head = NULL;
1711 ret = hrtimer_try_to_cancel(&io->timer);
1713 io_cqring_fill_event(link, -ECANCELED);
1714 io_commit_cqring(ctx);
1718 req->flags &= ~REQ_F_LINK_TIMEOUT;
1719 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1722 io_cqring_ev_posted(ctx);
1728 static void io_fail_links(struct io_kiocb *req)
1730 struct io_kiocb *link, *nxt;
1731 struct io_ring_ctx *ctx = req->ctx;
1732 unsigned long flags;
1734 spin_lock_irqsave(&ctx->completion_lock, flags);
1742 trace_io_uring_fail_link(req, link);
1743 io_cqring_fill_event(link, -ECANCELED);
1745 io_put_req_deferred(link, 2);
1748 io_commit_cqring(ctx);
1749 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1751 io_cqring_ev_posted(ctx);
1754 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1756 if (req->flags & REQ_F_LINK_TIMEOUT)
1757 io_kill_linked_timeout(req);
1760 * If LINK is set, we have dependent requests in this chain. If we
1761 * didn't fail this request, queue the first one up, moving any other
1762 * dependencies to the next request. In case of failure, fail the rest
1765 if (likely(!(req->flags & REQ_F_FAIL_LINK))) {
1766 struct io_kiocb *nxt = req->link;
1775 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1777 if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK))))
1779 return __io_req_find_next(req);
1782 static void ctx_flush_and_put(struct io_ring_ctx *ctx)
1786 if (ctx->submit_state.comp.nr) {
1787 mutex_lock(&ctx->uring_lock);
1788 io_submit_flush_completions(&ctx->submit_state.comp, ctx);
1789 mutex_unlock(&ctx->uring_lock);
1791 percpu_ref_put(&ctx->refs);
1794 static bool __tctx_task_work(struct io_uring_task *tctx)
1796 struct io_ring_ctx *ctx = NULL;
1797 struct io_wq_work_list list;
1798 struct io_wq_work_node *node;
1800 if (wq_list_empty(&tctx->task_list))
1803 spin_lock_irq(&tctx->task_lock);
1804 list = tctx->task_list;
1805 INIT_WQ_LIST(&tctx->task_list);
1806 spin_unlock_irq(&tctx->task_lock);
1810 struct io_wq_work_node *next = node->next;
1811 struct io_kiocb *req;
1813 req = container_of(node, struct io_kiocb, io_task_work.node);
1814 if (req->ctx != ctx) {
1815 ctx_flush_and_put(ctx);
1817 percpu_ref_get(&ctx->refs);
1820 req->task_work.func(&req->task_work);
1824 ctx_flush_and_put(ctx);
1825 return list.first != NULL;
1828 static void tctx_task_work(struct callback_head *cb)
1830 struct io_uring_task *tctx = container_of(cb, struct io_uring_task, task_work);
1832 clear_bit(0, &tctx->task_state);
1834 while (__tctx_task_work(tctx))
1838 static int io_task_work_add(struct task_struct *tsk, struct io_kiocb *req,
1839 enum task_work_notify_mode notify)
1841 struct io_uring_task *tctx = tsk->io_uring;
1842 struct io_wq_work_node *node, *prev;
1843 unsigned long flags;
1846 WARN_ON_ONCE(!tctx);
1848 spin_lock_irqsave(&tctx->task_lock, flags);
1849 wq_list_add_tail(&req->io_task_work.node, &tctx->task_list);
1850 spin_unlock_irqrestore(&tctx->task_lock, flags);
1852 /* task_work already pending, we're done */
1853 if (test_bit(0, &tctx->task_state) ||
1854 test_and_set_bit(0, &tctx->task_state))
1857 if (!task_work_add(tsk, &tctx->task_work, notify))
1861 * Slow path - we failed, find and delete work. if the work is not
1862 * in the list, it got run and we're fine.
1865 spin_lock_irqsave(&tctx->task_lock, flags);
1866 wq_list_for_each(node, prev, &tctx->task_list) {
1867 if (&req->io_task_work.node == node) {
1868 wq_list_del(&tctx->task_list, node, prev);
1873 spin_unlock_irqrestore(&tctx->task_lock, flags);
1874 clear_bit(0, &tctx->task_state);
1878 static int io_req_task_work_add(struct io_kiocb *req)
1880 struct task_struct *tsk = req->task;
1881 struct io_ring_ctx *ctx = req->ctx;
1882 enum task_work_notify_mode notify;
1885 if (tsk->flags & PF_EXITING)
1889 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
1890 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
1891 * processing task_work. There's no reliable way to tell if TWA_RESUME
1895 if (!(ctx->flags & IORING_SETUP_SQPOLL))
1896 notify = TWA_SIGNAL;
1898 ret = io_task_work_add(tsk, req, notify);
1900 wake_up_process(tsk);
1905 static void io_req_task_work_add_fallback(struct io_kiocb *req,
1906 task_work_func_t cb)
1908 struct io_ring_ctx *ctx = req->ctx;
1909 struct callback_head *head;
1911 init_task_work(&req->task_work, cb);
1913 head = READ_ONCE(ctx->exit_task_work);
1914 req->task_work.next = head;
1915 } while (cmpxchg(&ctx->exit_task_work, head, &req->task_work) != head);
1918 static void __io_req_task_cancel(struct io_kiocb *req, int error)
1920 struct io_ring_ctx *ctx = req->ctx;
1922 spin_lock_irq(&ctx->completion_lock);
1923 io_cqring_fill_event(req, error);
1924 io_commit_cqring(ctx);
1925 spin_unlock_irq(&ctx->completion_lock);
1927 io_cqring_ev_posted(ctx);
1928 req_set_fail_links(req);
1929 io_double_put_req(req);
1932 static void io_req_task_cancel(struct callback_head *cb)
1934 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1935 struct io_ring_ctx *ctx = req->ctx;
1937 mutex_lock(&ctx->uring_lock);
1938 __io_req_task_cancel(req, req->result);
1939 mutex_unlock(&ctx->uring_lock);
1940 percpu_ref_put(&ctx->refs);
1943 static void __io_req_task_submit(struct io_kiocb *req)
1945 struct io_ring_ctx *ctx = req->ctx;
1947 /* ctx stays valid until unlock, even if we drop all ours ctx->refs */
1948 mutex_lock(&ctx->uring_lock);
1949 if (!(current->flags & PF_EXITING) && !current->in_execve)
1950 __io_queue_sqe(req);
1952 __io_req_task_cancel(req, -EFAULT);
1953 mutex_unlock(&ctx->uring_lock);
1956 static void io_req_task_submit(struct callback_head *cb)
1958 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1960 __io_req_task_submit(req);
1963 static void io_req_task_queue(struct io_kiocb *req)
1967 req->task_work.func = io_req_task_submit;
1968 ret = io_req_task_work_add(req);
1969 if (unlikely(ret)) {
1970 req->result = -ECANCELED;
1971 percpu_ref_get(&req->ctx->refs);
1972 io_req_task_work_add_fallback(req, io_req_task_cancel);
1976 static void io_req_task_queue_fail(struct io_kiocb *req, int ret)
1978 percpu_ref_get(&req->ctx->refs);
1980 req->task_work.func = io_req_task_cancel;
1982 if (unlikely(io_req_task_work_add(req)))
1983 io_req_task_work_add_fallback(req, io_req_task_cancel);
1986 static inline void io_queue_next(struct io_kiocb *req)
1988 struct io_kiocb *nxt = io_req_find_next(req);
1991 io_req_task_queue(nxt);
1994 static void io_free_req(struct io_kiocb *req)
2001 struct task_struct *task;
2006 static inline void io_init_req_batch(struct req_batch *rb)
2013 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2014 struct req_batch *rb)
2017 io_put_task(rb->task, rb->task_refs);
2019 percpu_ref_put_many(&ctx->refs, rb->ctx_refs);
2022 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req,
2023 struct io_submit_state *state)
2027 if (req->task != rb->task) {
2029 io_put_task(rb->task, rb->task_refs);
2030 rb->task = req->task;
2036 io_dismantle_req(req);
2037 if (state->free_reqs != ARRAY_SIZE(state->reqs))
2038 state->reqs[state->free_reqs++] = req;
2040 list_add(&req->compl.list, &state->comp.free_list);
2043 static void io_submit_flush_completions(struct io_comp_state *cs,
2044 struct io_ring_ctx *ctx)
2047 struct io_kiocb *req;
2048 struct req_batch rb;
2050 io_init_req_batch(&rb);
2051 spin_lock_irq(&ctx->completion_lock);
2052 for (i = 0; i < nr; i++) {
2054 __io_cqring_fill_event(req, req->result, req->compl.cflags);
2056 io_commit_cqring(ctx);
2057 spin_unlock_irq(&ctx->completion_lock);
2059 io_cqring_ev_posted(ctx);
2060 for (i = 0; i < nr; i++) {
2063 /* submission and completion refs */
2064 if (refcount_sub_and_test(2, &req->refs))
2065 io_req_free_batch(&rb, req, &ctx->submit_state);
2068 io_req_free_batch_finish(ctx, &rb);
2073 * Drop reference to request, return next in chain (if there is one) if this
2074 * was the last reference to this request.
2076 static struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2078 struct io_kiocb *nxt = NULL;
2080 if (refcount_dec_and_test(&req->refs)) {
2081 nxt = io_req_find_next(req);
2087 static void io_put_req(struct io_kiocb *req)
2089 if (refcount_dec_and_test(&req->refs))
2093 static void io_put_req_deferred_cb(struct callback_head *cb)
2095 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2100 static void io_free_req_deferred(struct io_kiocb *req)
2104 req->task_work.func = io_put_req_deferred_cb;
2105 ret = io_req_task_work_add(req);
2107 io_req_task_work_add_fallback(req, io_put_req_deferred_cb);
2110 static inline void io_put_req_deferred(struct io_kiocb *req, int refs)
2112 if (refcount_sub_and_test(refs, &req->refs))
2113 io_free_req_deferred(req);
2116 static void io_double_put_req(struct io_kiocb *req)
2118 /* drop both submit and complete references */
2119 if (refcount_sub_and_test(2, &req->refs))
2123 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2125 /* See comment at the top of this file */
2127 return __io_cqring_events(ctx);
2130 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2132 struct io_rings *rings = ctx->rings;
2134 /* make sure SQ entry isn't read before tail */
2135 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2138 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2140 unsigned int cflags;
2142 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2143 cflags |= IORING_CQE_F_BUFFER;
2144 req->flags &= ~REQ_F_BUFFER_SELECTED;
2149 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2151 struct io_buffer *kbuf;
2153 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2154 return io_put_kbuf(req, kbuf);
2157 static inline bool io_run_task_work(void)
2160 * Not safe to run on exiting task, and the task_work handling will
2161 * not add work to such a task.
2163 if (unlikely(current->flags & PF_EXITING))
2165 if (current->task_works) {
2166 __set_current_state(TASK_RUNNING);
2175 * Find and free completed poll iocbs
2177 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2178 struct list_head *done)
2180 struct req_batch rb;
2181 struct io_kiocb *req;
2183 /* order with ->result store in io_complete_rw_iopoll() */
2186 io_init_req_batch(&rb);
2187 while (!list_empty(done)) {
2190 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2191 list_del(&req->inflight_entry);
2193 if (READ_ONCE(req->result) == -EAGAIN) {
2194 req->iopoll_completed = 0;
2195 if (io_rw_reissue(req))
2199 if (req->flags & REQ_F_BUFFER_SELECTED)
2200 cflags = io_put_rw_kbuf(req);
2202 __io_cqring_fill_event(req, req->result, cflags);
2205 if (refcount_dec_and_test(&req->refs))
2206 io_req_free_batch(&rb, req, &ctx->submit_state);
2209 io_commit_cqring(ctx);
2210 io_cqring_ev_posted_iopoll(ctx);
2211 io_req_free_batch_finish(ctx, &rb);
2214 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2217 struct io_kiocb *req, *tmp;
2223 * Only spin for completions if we don't have multiple devices hanging
2224 * off our complete list, and we're under the requested amount.
2226 spin = !ctx->poll_multi_file && *nr_events < min;
2229 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2230 struct kiocb *kiocb = &req->rw.kiocb;
2233 * Move completed and retryable entries to our local lists.
2234 * If we find a request that requires polling, break out
2235 * and complete those lists first, if we have entries there.
2237 if (READ_ONCE(req->iopoll_completed)) {
2238 list_move_tail(&req->inflight_entry, &done);
2241 if (!list_empty(&done))
2244 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2248 /* iopoll may have completed current req */
2249 if (READ_ONCE(req->iopoll_completed))
2250 list_move_tail(&req->inflight_entry, &done);
2257 if (!list_empty(&done))
2258 io_iopoll_complete(ctx, nr_events, &done);
2264 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2265 * non-spinning poll check - we'll still enter the driver poll loop, but only
2266 * as a non-spinning completion check.
2268 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
2271 while (!list_empty(&ctx->iopoll_list) && !need_resched()) {
2274 ret = io_do_iopoll(ctx, nr_events, min);
2277 if (*nr_events >= min)
2285 * We can't just wait for polled events to come to us, we have to actively
2286 * find and complete them.
2288 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2290 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2293 mutex_lock(&ctx->uring_lock);
2294 while (!list_empty(&ctx->iopoll_list)) {
2295 unsigned int nr_events = 0;
2297 io_do_iopoll(ctx, &nr_events, 0);
2299 /* let it sleep and repeat later if can't complete a request */
2303 * Ensure we allow local-to-the-cpu processing to take place,
2304 * in this case we need to ensure that we reap all events.
2305 * Also let task_work, etc. to progress by releasing the mutex
2307 if (need_resched()) {
2308 mutex_unlock(&ctx->uring_lock);
2310 mutex_lock(&ctx->uring_lock);
2313 mutex_unlock(&ctx->uring_lock);
2316 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2318 unsigned int nr_events = 0;
2319 int iters = 0, ret = 0;
2322 * We disallow the app entering submit/complete with polling, but we
2323 * still need to lock the ring to prevent racing with polled issue
2324 * that got punted to a workqueue.
2326 mutex_lock(&ctx->uring_lock);
2329 * Don't enter poll loop if we already have events pending.
2330 * If we do, we can potentially be spinning for commands that
2331 * already triggered a CQE (eg in error).
2333 if (test_bit(0, &ctx->cq_check_overflow))
2334 __io_cqring_overflow_flush(ctx, false, NULL, NULL);
2335 if (io_cqring_events(ctx))
2339 * If a submit got punted to a workqueue, we can have the
2340 * application entering polling for a command before it gets
2341 * issued. That app will hold the uring_lock for the duration
2342 * of the poll right here, so we need to take a breather every
2343 * now and then to ensure that the issue has a chance to add
2344 * the poll to the issued list. Otherwise we can spin here
2345 * forever, while the workqueue is stuck trying to acquire the
2348 if (!(++iters & 7)) {
2349 mutex_unlock(&ctx->uring_lock);
2351 mutex_lock(&ctx->uring_lock);
2354 ret = io_iopoll_getevents(ctx, &nr_events, min);
2358 } while (min && !nr_events && !need_resched());
2360 mutex_unlock(&ctx->uring_lock);
2364 static void kiocb_end_write(struct io_kiocb *req)
2367 * Tell lockdep we inherited freeze protection from submission
2370 if (req->flags & REQ_F_ISREG) {
2371 struct inode *inode = file_inode(req->file);
2373 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
2375 file_end_write(req->file);
2379 static bool io_resubmit_prep(struct io_kiocb *req)
2381 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2383 struct iov_iter iter;
2385 /* already prepared */
2386 if (req->async_data)
2389 switch (req->opcode) {
2390 case IORING_OP_READV:
2391 case IORING_OP_READ_FIXED:
2392 case IORING_OP_READ:
2395 case IORING_OP_WRITEV:
2396 case IORING_OP_WRITE_FIXED:
2397 case IORING_OP_WRITE:
2401 printk_once(KERN_WARNING "io_uring: bad opcode in resubmit %d\n",
2406 ret = io_import_iovec(rw, req, &iovec, &iter, false);
2409 return !io_setup_async_rw(req, iovec, inline_vecs, &iter, false);
2412 static bool io_rw_should_reissue(struct io_kiocb *req)
2414 umode_t mode = file_inode(req->file)->i_mode;
2415 struct io_ring_ctx *ctx = req->ctx;
2417 if (!S_ISBLK(mode) && !S_ISREG(mode))
2419 if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
2420 !(ctx->flags & IORING_SETUP_IOPOLL)))
2423 * If ref is dying, we might be running poll reap from the exit work.
2424 * Don't attempt to reissue from that path, just let it fail with
2427 if (percpu_ref_is_dying(&ctx->refs))
2433 static bool io_rw_reissue(struct io_kiocb *req)
2436 if (!io_rw_should_reissue(req))
2439 lockdep_assert_held(&req->ctx->uring_lock);
2441 if (io_resubmit_prep(req)) {
2442 refcount_inc(&req->refs);
2443 io_queue_async_work(req);
2446 req_set_fail_links(req);
2451 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2452 unsigned int issue_flags)
2456 if ((res == -EAGAIN || res == -EOPNOTSUPP) && io_rw_reissue(req))
2458 if (res != req->result)
2459 req_set_fail_links(req);
2461 if (req->rw.kiocb.ki_flags & IOCB_WRITE)
2462 kiocb_end_write(req);
2463 if (req->flags & REQ_F_BUFFER_SELECTED)
2464 cflags = io_put_rw_kbuf(req);
2465 __io_req_complete(req, issue_flags, res, cflags);
2468 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2470 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2472 __io_complete_rw(req, res, res2, 0);
2475 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2477 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2480 /* Rewind iter, if we have one. iopoll path resubmits as usual */
2481 if (res == -EAGAIN && io_rw_should_reissue(req)) {
2482 struct io_async_rw *rw = req->async_data;
2485 iov_iter_revert(&rw->iter,
2486 req->result - iov_iter_count(&rw->iter));
2487 else if (!io_resubmit_prep(req))
2492 if (kiocb->ki_flags & IOCB_WRITE)
2493 kiocb_end_write(req);
2495 if (res != -EAGAIN && res != req->result)
2496 req_set_fail_links(req);
2498 WRITE_ONCE(req->result, res);
2499 /* order with io_poll_complete() checking ->result */
2501 WRITE_ONCE(req->iopoll_completed, 1);
2505 * After the iocb has been issued, it's safe to be found on the poll list.
2506 * Adding the kiocb to the list AFTER submission ensures that we don't
2507 * find it from a io_iopoll_getevents() thread before the issuer is done
2508 * accessing the kiocb cookie.
2510 static void io_iopoll_req_issued(struct io_kiocb *req, bool in_async)
2512 struct io_ring_ctx *ctx = req->ctx;
2515 * Track whether we have multiple files in our lists. This will impact
2516 * how we do polling eventually, not spinning if we're on potentially
2517 * different devices.
2519 if (list_empty(&ctx->iopoll_list)) {
2520 ctx->poll_multi_file = false;
2521 } else if (!ctx->poll_multi_file) {
2522 struct io_kiocb *list_req;
2524 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2526 if (list_req->file != req->file)
2527 ctx->poll_multi_file = true;
2531 * For fast devices, IO may have already completed. If it has, add
2532 * it to the front so we find it first.
2534 if (READ_ONCE(req->iopoll_completed))
2535 list_add(&req->inflight_entry, &ctx->iopoll_list);
2537 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2540 * If IORING_SETUP_SQPOLL is enabled, sqes are either handled in sq thread
2541 * task context or in io worker task context. If current task context is
2542 * sq thread, we don't need to check whether should wake up sq thread.
2544 if (in_async && (ctx->flags & IORING_SETUP_SQPOLL) &&
2545 wq_has_sleeper(&ctx->sq_data->wait))
2546 wake_up(&ctx->sq_data->wait);
2549 static inline void io_state_file_put(struct io_submit_state *state)
2551 if (state->file_refs) {
2552 fput_many(state->file, state->file_refs);
2553 state->file_refs = 0;
2558 * Get as many references to a file as we have IOs left in this submission,
2559 * assuming most submissions are for one file, or at least that each file
2560 * has more than one submission.
2562 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2567 if (state->file_refs) {
2568 if (state->fd == fd) {
2572 io_state_file_put(state);
2574 state->file = fget_many(fd, state->ios_left);
2575 if (unlikely(!state->file))
2579 state->file_refs = state->ios_left - 1;
2583 static bool io_bdev_nowait(struct block_device *bdev)
2585 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2589 * If we tracked the file through the SCM inflight mechanism, we could support
2590 * any file. For now, just ensure that anything potentially problematic is done
2593 static bool io_file_supports_async(struct file *file, int rw)
2595 umode_t mode = file_inode(file)->i_mode;
2597 if (S_ISBLK(mode)) {
2598 if (IS_ENABLED(CONFIG_BLOCK) &&
2599 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2603 if (S_ISCHR(mode) || S_ISSOCK(mode))
2605 if (S_ISREG(mode)) {
2606 if (IS_ENABLED(CONFIG_BLOCK) &&
2607 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2608 file->f_op != &io_uring_fops)
2613 /* any ->read/write should understand O_NONBLOCK */
2614 if (file->f_flags & O_NONBLOCK)
2617 if (!(file->f_mode & FMODE_NOWAIT))
2621 return file->f_op->read_iter != NULL;
2623 return file->f_op->write_iter != NULL;
2626 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2628 struct io_ring_ctx *ctx = req->ctx;
2629 struct kiocb *kiocb = &req->rw.kiocb;
2630 struct file *file = req->file;
2634 if (S_ISREG(file_inode(file)->i_mode))
2635 req->flags |= REQ_F_ISREG;
2637 kiocb->ki_pos = READ_ONCE(sqe->off);
2638 if (kiocb->ki_pos == -1 && !(file->f_mode & FMODE_STREAM)) {
2639 req->flags |= REQ_F_CUR_POS;
2640 kiocb->ki_pos = file->f_pos;
2642 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2643 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2644 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2648 /* don't allow async punt for O_NONBLOCK or RWF_NOWAIT */
2649 if ((kiocb->ki_flags & IOCB_NOWAIT) || (file->f_flags & O_NONBLOCK))
2650 req->flags |= REQ_F_NOWAIT;
2652 ioprio = READ_ONCE(sqe->ioprio);
2654 ret = ioprio_check_cap(ioprio);
2658 kiocb->ki_ioprio = ioprio;
2660 kiocb->ki_ioprio = get_current_ioprio();
2662 if (ctx->flags & IORING_SETUP_IOPOLL) {
2663 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2664 !kiocb->ki_filp->f_op->iopoll)
2667 kiocb->ki_flags |= IOCB_HIPRI;
2668 kiocb->ki_complete = io_complete_rw_iopoll;
2669 req->iopoll_completed = 0;
2671 if (kiocb->ki_flags & IOCB_HIPRI)
2673 kiocb->ki_complete = io_complete_rw;
2676 req->rw.addr = READ_ONCE(sqe->addr);
2677 req->rw.len = READ_ONCE(sqe->len);
2678 req->buf_index = READ_ONCE(sqe->buf_index);
2682 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2688 case -ERESTARTNOINTR:
2689 case -ERESTARTNOHAND:
2690 case -ERESTART_RESTARTBLOCK:
2692 * We can't just restart the syscall, since previously
2693 * submitted sqes may already be in progress. Just fail this
2699 kiocb->ki_complete(kiocb, ret, 0);
2703 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2704 unsigned int issue_flags)
2706 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2707 struct io_async_rw *io = req->async_data;
2709 /* add previously done IO, if any */
2710 if (io && io->bytes_done > 0) {
2712 ret = io->bytes_done;
2714 ret += io->bytes_done;
2717 if (req->flags & REQ_F_CUR_POS)
2718 req->file->f_pos = kiocb->ki_pos;
2719 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2720 __io_complete_rw(req, ret, 0, issue_flags);
2722 io_rw_done(kiocb, ret);
2725 static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
2727 struct io_ring_ctx *ctx = req->ctx;
2728 size_t len = req->rw.len;
2729 struct io_mapped_ubuf *imu;
2730 u16 index, buf_index = req->buf_index;
2734 if (unlikely(buf_index >= ctx->nr_user_bufs))
2736 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2737 imu = &ctx->user_bufs[index];
2738 buf_addr = req->rw.addr;
2741 if (buf_addr + len < buf_addr)
2743 /* not inside the mapped region */
2744 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2748 * May not be a start of buffer, set size appropriately
2749 * and advance us to the beginning.
2751 offset = buf_addr - imu->ubuf;
2752 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2756 * Don't use iov_iter_advance() here, as it's really slow for
2757 * using the latter parts of a big fixed buffer - it iterates
2758 * over each segment manually. We can cheat a bit here, because
2761 * 1) it's a BVEC iter, we set it up
2762 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2763 * first and last bvec
2765 * So just find our index, and adjust the iterator afterwards.
2766 * If the offset is within the first bvec (or the whole first
2767 * bvec, just use iov_iter_advance(). This makes it easier
2768 * since we can just skip the first segment, which may not
2769 * be PAGE_SIZE aligned.
2771 const struct bio_vec *bvec = imu->bvec;
2773 if (offset <= bvec->bv_len) {
2774 iov_iter_advance(iter, offset);
2776 unsigned long seg_skip;
2778 /* skip first vec */
2779 offset -= bvec->bv_len;
2780 seg_skip = 1 + (offset >> PAGE_SHIFT);
2782 iter->bvec = bvec + seg_skip;
2783 iter->nr_segs -= seg_skip;
2784 iter->count -= bvec->bv_len + offset;
2785 iter->iov_offset = offset & ~PAGE_MASK;
2792 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2795 mutex_unlock(&ctx->uring_lock);
2798 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2801 * "Normal" inline submissions always hold the uring_lock, since we
2802 * grab it from the system call. Same is true for the SQPOLL offload.
2803 * The only exception is when we've detached the request and issue it
2804 * from an async worker thread, grab the lock for that case.
2807 mutex_lock(&ctx->uring_lock);
2810 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2811 int bgid, struct io_buffer *kbuf,
2814 struct io_buffer *head;
2816 if (req->flags & REQ_F_BUFFER_SELECTED)
2819 io_ring_submit_lock(req->ctx, needs_lock);
2821 lockdep_assert_held(&req->ctx->uring_lock);
2823 head = idr_find(&req->ctx->io_buffer_idr, bgid);
2825 if (!list_empty(&head->list)) {
2826 kbuf = list_last_entry(&head->list, struct io_buffer,
2828 list_del(&kbuf->list);
2831 idr_remove(&req->ctx->io_buffer_idr, bgid);
2833 if (*len > kbuf->len)
2836 kbuf = ERR_PTR(-ENOBUFS);
2839 io_ring_submit_unlock(req->ctx, needs_lock);
2844 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2847 struct io_buffer *kbuf;
2850 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2851 bgid = req->buf_index;
2852 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2855 req->rw.addr = (u64) (unsigned long) kbuf;
2856 req->flags |= REQ_F_BUFFER_SELECTED;
2857 return u64_to_user_ptr(kbuf->addr);
2860 #ifdef CONFIG_COMPAT
2861 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2864 struct compat_iovec __user *uiov;
2865 compat_ssize_t clen;
2869 uiov = u64_to_user_ptr(req->rw.addr);
2870 if (!access_ok(uiov, sizeof(*uiov)))
2872 if (__get_user(clen, &uiov->iov_len))
2878 buf = io_rw_buffer_select(req, &len, needs_lock);
2880 return PTR_ERR(buf);
2881 iov[0].iov_base = buf;
2882 iov[0].iov_len = (compat_size_t) len;
2887 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2890 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2894 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2897 len = iov[0].iov_len;
2900 buf = io_rw_buffer_select(req, &len, needs_lock);
2902 return PTR_ERR(buf);
2903 iov[0].iov_base = buf;
2904 iov[0].iov_len = len;
2908 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2911 if (req->flags & REQ_F_BUFFER_SELECTED) {
2912 struct io_buffer *kbuf;
2914 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2915 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
2916 iov[0].iov_len = kbuf->len;
2919 if (req->rw.len != 1)
2922 #ifdef CONFIG_COMPAT
2923 if (req->ctx->compat)
2924 return io_compat_import(req, iov, needs_lock);
2927 return __io_iov_buffer_select(req, iov, needs_lock);
2930 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
2931 struct iov_iter *iter, bool needs_lock)
2933 void __user *buf = u64_to_user_ptr(req->rw.addr);
2934 size_t sqe_len = req->rw.len;
2935 u8 opcode = req->opcode;
2938 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
2940 return io_import_fixed(req, rw, iter);
2943 /* buffer index only valid with fixed read/write, or buffer select */
2944 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
2947 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
2948 if (req->flags & REQ_F_BUFFER_SELECT) {
2949 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
2951 return PTR_ERR(buf);
2952 req->rw.len = sqe_len;
2955 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
2960 if (req->flags & REQ_F_BUFFER_SELECT) {
2961 ret = io_iov_buffer_select(req, *iovec, needs_lock);
2963 iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
2968 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
2972 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
2974 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
2978 * For files that don't have ->read_iter() and ->write_iter(), handle them
2979 * by looping over ->read() or ->write() manually.
2981 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
2983 struct kiocb *kiocb = &req->rw.kiocb;
2984 struct file *file = req->file;
2988 * Don't support polled IO through this interface, and we can't
2989 * support non-blocking either. For the latter, this just causes
2990 * the kiocb to be handled from an async context.
2992 if (kiocb->ki_flags & IOCB_HIPRI)
2994 if (kiocb->ki_flags & IOCB_NOWAIT)
2997 while (iov_iter_count(iter)) {
3001 if (!iov_iter_is_bvec(iter)) {
3002 iovec = iov_iter_iovec(iter);
3004 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3005 iovec.iov_len = req->rw.len;
3009 nr = file->f_op->read(file, iovec.iov_base,
3010 iovec.iov_len, io_kiocb_ppos(kiocb));
3012 nr = file->f_op->write(file, iovec.iov_base,
3013 iovec.iov_len, io_kiocb_ppos(kiocb));
3022 if (nr != iovec.iov_len)
3026 iov_iter_advance(iter, nr);
3032 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3033 const struct iovec *fast_iov, struct iov_iter *iter)
3035 struct io_async_rw *rw = req->async_data;
3037 memcpy(&rw->iter, iter, sizeof(*iter));
3038 rw->free_iovec = iovec;
3040 /* can only be fixed buffers, no need to do anything */
3041 if (iov_iter_is_bvec(iter))
3044 unsigned iov_off = 0;
3046 rw->iter.iov = rw->fast_iov;
3047 if (iter->iov != fast_iov) {
3048 iov_off = iter->iov - fast_iov;
3049 rw->iter.iov += iov_off;
3051 if (rw->fast_iov != fast_iov)
3052 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3053 sizeof(struct iovec) * iter->nr_segs);
3055 req->flags |= REQ_F_NEED_CLEANUP;
3059 static inline int __io_alloc_async_data(struct io_kiocb *req)
3061 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3062 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3063 return req->async_data == NULL;
3066 static int io_alloc_async_data(struct io_kiocb *req)
3068 if (!io_op_defs[req->opcode].needs_async_data)
3071 return __io_alloc_async_data(req);
3074 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3075 const struct iovec *fast_iov,
3076 struct iov_iter *iter, bool force)
3078 if (!force && !io_op_defs[req->opcode].needs_async_data)
3080 if (!req->async_data) {
3081 if (__io_alloc_async_data(req)) {
3086 io_req_map_rw(req, iovec, fast_iov, iter);
3091 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3093 struct io_async_rw *iorw = req->async_data;
3094 struct iovec *iov = iorw->fast_iov;
3097 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3098 if (unlikely(ret < 0))
3101 iorw->bytes_done = 0;
3102 iorw->free_iovec = iov;
3104 req->flags |= REQ_F_NEED_CLEANUP;
3108 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3110 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3112 return io_prep_rw(req, sqe);
3116 * This is our waitqueue callback handler, registered through lock_page_async()
3117 * when we initially tried to do the IO with the iocb armed our waitqueue.
3118 * This gets called when the page is unlocked, and we generally expect that to
3119 * happen when the page IO is completed and the page is now uptodate. This will
3120 * queue a task_work based retry of the operation, attempting to copy the data
3121 * again. If the latter fails because the page was NOT uptodate, then we will
3122 * do a thread based blocking retry of the operation. That's the unexpected
3125 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3126 int sync, void *arg)
3128 struct wait_page_queue *wpq;
3129 struct io_kiocb *req = wait->private;
3130 struct wait_page_key *key = arg;
3132 wpq = container_of(wait, struct wait_page_queue, wait);
3134 if (!wake_page_match(wpq, key))
3137 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3138 list_del_init(&wait->entry);
3140 /* submit ref gets dropped, acquire a new one */
3141 refcount_inc(&req->refs);
3142 io_req_task_queue(req);
3147 * This controls whether a given IO request should be armed for async page
3148 * based retry. If we return false here, the request is handed to the async
3149 * worker threads for retry. If we're doing buffered reads on a regular file,
3150 * we prepare a private wait_page_queue entry and retry the operation. This
3151 * will either succeed because the page is now uptodate and unlocked, or it
3152 * will register a callback when the page is unlocked at IO completion. Through
3153 * that callback, io_uring uses task_work to setup a retry of the operation.
3154 * That retry will attempt the buffered read again. The retry will generally
3155 * succeed, or in rare cases where it fails, we then fall back to using the
3156 * async worker threads for a blocking retry.
3158 static bool io_rw_should_retry(struct io_kiocb *req)
3160 struct io_async_rw *rw = req->async_data;
3161 struct wait_page_queue *wait = &rw->wpq;
3162 struct kiocb *kiocb = &req->rw.kiocb;
3164 /* never retry for NOWAIT, we just complete with -EAGAIN */
3165 if (req->flags & REQ_F_NOWAIT)
3168 /* Only for buffered IO */
3169 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3173 * just use poll if we can, and don't attempt if the fs doesn't
3174 * support callback based unlocks
3176 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3179 wait->wait.func = io_async_buf_func;
3180 wait->wait.private = req;
3181 wait->wait.flags = 0;
3182 INIT_LIST_HEAD(&wait->wait.entry);
3183 kiocb->ki_flags |= IOCB_WAITQ;
3184 kiocb->ki_flags &= ~IOCB_NOWAIT;
3185 kiocb->ki_waitq = wait;
3189 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3191 if (req->file->f_op->read_iter)
3192 return call_read_iter(req->file, &req->rw.kiocb, iter);
3193 else if (req->file->f_op->read)
3194 return loop_rw_iter(READ, req, iter);
3199 static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3201 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3202 struct kiocb *kiocb = &req->rw.kiocb;
3203 struct iov_iter __iter, *iter = &__iter;
3204 struct io_async_rw *rw = req->async_data;
3205 ssize_t io_size, ret, ret2;
3206 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3212 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3216 io_size = iov_iter_count(iter);
3217 req->result = io_size;
3219 /* Ensure we clear previously set non-block flag */
3220 if (!force_nonblock)
3221 kiocb->ki_flags &= ~IOCB_NOWAIT;
3223 kiocb->ki_flags |= IOCB_NOWAIT;
3225 /* If the file doesn't support async, just async punt */
3226 if (force_nonblock && !io_file_supports_async(req->file, READ)) {
3227 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3228 return ret ?: -EAGAIN;
3231 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size);
3232 if (unlikely(ret)) {
3237 ret = io_iter_do_read(req, iter);
3239 if (ret == -EIOCBQUEUED) {
3240 if (req->async_data)
3241 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3243 } else if (ret == -EAGAIN) {
3244 /* IOPOLL retry should happen for io-wq threads */
3245 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3247 /* no retry on NONBLOCK nor RWF_NOWAIT */
3248 if (req->flags & REQ_F_NOWAIT)
3250 /* some cases will consume bytes even on error returns */
3251 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3253 } else if (ret <= 0 || ret == io_size || !force_nonblock ||
3254 (req->flags & REQ_F_NOWAIT) || !(req->flags & REQ_F_ISREG)) {
3255 /* read all, failed, already did sync or don't want to retry */
3259 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3264 rw = req->async_data;
3265 /* now use our persistent iterator, if we aren't already */
3270 rw->bytes_done += ret;
3271 /* if we can retry, do so with the callbacks armed */
3272 if (!io_rw_should_retry(req)) {
3273 kiocb->ki_flags &= ~IOCB_WAITQ;
3278 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3279 * we get -EIOCBQUEUED, then we'll get a notification when the
3280 * desired page gets unlocked. We can also get a partial read
3281 * here, and if we do, then just retry at the new offset.
3283 ret = io_iter_do_read(req, iter);
3284 if (ret == -EIOCBQUEUED)
3286 /* we got some bytes, but not all. retry. */
3287 } while (ret > 0 && ret < io_size);
3289 kiocb_done(kiocb, ret, issue_flags);
3291 /* it's faster to check here then delegate to kfree */
3297 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3299 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3301 return io_prep_rw(req, sqe);
3304 static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3306 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3307 struct kiocb *kiocb = &req->rw.kiocb;
3308 struct iov_iter __iter, *iter = &__iter;
3309 struct io_async_rw *rw = req->async_data;
3310 ssize_t ret, ret2, io_size;
3311 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3317 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3321 io_size = iov_iter_count(iter);
3322 req->result = io_size;
3324 /* Ensure we clear previously set non-block flag */
3325 if (!force_nonblock)
3326 kiocb->ki_flags &= ~IOCB_NOWAIT;
3328 kiocb->ki_flags |= IOCB_NOWAIT;
3330 /* If the file doesn't support async, just async punt */
3331 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
3334 /* file path doesn't support NOWAIT for non-direct_IO */
3335 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3336 (req->flags & REQ_F_ISREG))
3339 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size);
3344 * Open-code file_start_write here to grab freeze protection,
3345 * which will be released by another thread in
3346 * io_complete_rw(). Fool lockdep by telling it the lock got
3347 * released so that it doesn't complain about the held lock when
3348 * we return to userspace.
3350 if (req->flags & REQ_F_ISREG) {
3351 sb_start_write(file_inode(req->file)->i_sb);
3352 __sb_writers_release(file_inode(req->file)->i_sb,
3355 kiocb->ki_flags |= IOCB_WRITE;
3357 if (req->file->f_op->write_iter)
3358 ret2 = call_write_iter(req->file, kiocb, iter);
3359 else if (req->file->f_op->write)
3360 ret2 = loop_rw_iter(WRITE, req, iter);
3365 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3366 * retry them without IOCB_NOWAIT.
3368 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3370 /* no retry on NONBLOCK nor RWF_NOWAIT */
3371 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3373 if (ret2 == -EIOCBQUEUED && req->async_data)
3374 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3375 if (!force_nonblock || ret2 != -EAGAIN) {
3376 /* IOPOLL retry should happen for io-wq threads */
3377 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3380 kiocb_done(kiocb, ret2, issue_flags);
3383 /* some cases will consume bytes even on error returns */
3384 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3385 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3386 return ret ?: -EAGAIN;
3389 /* it's reportedly faster than delegating the null check to kfree() */
3395 static int io_renameat_prep(struct io_kiocb *req,
3396 const struct io_uring_sqe *sqe)
3398 struct io_rename *ren = &req->rename;
3399 const char __user *oldf, *newf;
3401 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3404 ren->old_dfd = READ_ONCE(sqe->fd);
3405 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3406 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3407 ren->new_dfd = READ_ONCE(sqe->len);
3408 ren->flags = READ_ONCE(sqe->rename_flags);
3410 ren->oldpath = getname(oldf);
3411 if (IS_ERR(ren->oldpath))
3412 return PTR_ERR(ren->oldpath);
3414 ren->newpath = getname(newf);
3415 if (IS_ERR(ren->newpath)) {
3416 putname(ren->oldpath);
3417 return PTR_ERR(ren->newpath);
3420 req->flags |= REQ_F_NEED_CLEANUP;
3424 static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3426 struct io_rename *ren = &req->rename;
3429 if (issue_flags & IO_URING_F_NONBLOCK)
3432 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3433 ren->newpath, ren->flags);
3435 req->flags &= ~REQ_F_NEED_CLEANUP;
3437 req_set_fail_links(req);
3438 io_req_complete(req, ret);
3442 static int io_unlinkat_prep(struct io_kiocb *req,
3443 const struct io_uring_sqe *sqe)
3445 struct io_unlink *un = &req->unlink;
3446 const char __user *fname;
3448 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3451 un->dfd = READ_ONCE(sqe->fd);
3453 un->flags = READ_ONCE(sqe->unlink_flags);
3454 if (un->flags & ~AT_REMOVEDIR)
3457 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3458 un->filename = getname(fname);
3459 if (IS_ERR(un->filename))
3460 return PTR_ERR(un->filename);
3462 req->flags |= REQ_F_NEED_CLEANUP;
3466 static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3468 struct io_unlink *un = &req->unlink;
3471 if (issue_flags & IO_URING_F_NONBLOCK)
3474 if (un->flags & AT_REMOVEDIR)
3475 ret = do_rmdir(un->dfd, un->filename);
3477 ret = do_unlinkat(un->dfd, un->filename);
3479 req->flags &= ~REQ_F_NEED_CLEANUP;
3481 req_set_fail_links(req);
3482 io_req_complete(req, ret);
3486 static int io_shutdown_prep(struct io_kiocb *req,
3487 const struct io_uring_sqe *sqe)
3489 #if defined(CONFIG_NET)
3490 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3492 if (sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3496 req->shutdown.how = READ_ONCE(sqe->len);
3503 static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3505 #if defined(CONFIG_NET)
3506 struct socket *sock;
3509 if (issue_flags & IO_URING_F_NONBLOCK)
3512 sock = sock_from_file(req->file);
3513 if (unlikely(!sock))
3516 ret = __sys_shutdown_sock(sock, req->shutdown.how);
3518 req_set_fail_links(req);
3519 io_req_complete(req, ret);
3526 static int __io_splice_prep(struct io_kiocb *req,
3527 const struct io_uring_sqe *sqe)
3529 struct io_splice* sp = &req->splice;
3530 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3532 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3536 sp->len = READ_ONCE(sqe->len);
3537 sp->flags = READ_ONCE(sqe->splice_flags);
3539 if (unlikely(sp->flags & ~valid_flags))
3542 sp->file_in = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in),
3543 (sp->flags & SPLICE_F_FD_IN_FIXED));
3546 req->flags |= REQ_F_NEED_CLEANUP;
3548 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
3550 * Splice operation will be punted aync, and here need to
3551 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3553 req->work.flags |= IO_WQ_WORK_UNBOUND;
3559 static int io_tee_prep(struct io_kiocb *req,
3560 const struct io_uring_sqe *sqe)
3562 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3564 return __io_splice_prep(req, sqe);
3567 static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
3569 struct io_splice *sp = &req->splice;
3570 struct file *in = sp->file_in;
3571 struct file *out = sp->file_out;
3572 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3575 if (issue_flags & IO_URING_F_NONBLOCK)
3578 ret = do_tee(in, out, sp->len, flags);
3580 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3581 req->flags &= ~REQ_F_NEED_CLEANUP;
3584 req_set_fail_links(req);
3585 io_req_complete(req, ret);
3589 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3591 struct io_splice* sp = &req->splice;
3593 sp->off_in = READ_ONCE(sqe->splice_off_in);
3594 sp->off_out = READ_ONCE(sqe->off);
3595 return __io_splice_prep(req, sqe);
3598 static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
3600 struct io_splice *sp = &req->splice;
3601 struct file *in = sp->file_in;
3602 struct file *out = sp->file_out;
3603 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3604 loff_t *poff_in, *poff_out;
3607 if (issue_flags & IO_URING_F_NONBLOCK)
3610 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3611 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3614 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3616 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3617 req->flags &= ~REQ_F_NEED_CLEANUP;
3620 req_set_fail_links(req);
3621 io_req_complete(req, ret);
3626 * IORING_OP_NOP just posts a completion event, nothing else.
3628 static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
3630 struct io_ring_ctx *ctx = req->ctx;
3632 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3635 __io_req_complete(req, issue_flags, 0, 0);
3639 static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3641 struct io_ring_ctx *ctx = req->ctx;
3646 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3648 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3651 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3652 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3655 req->sync.off = READ_ONCE(sqe->off);
3656 req->sync.len = READ_ONCE(sqe->len);
3660 static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
3662 loff_t end = req->sync.off + req->sync.len;
3665 /* fsync always requires a blocking context */
3666 if (issue_flags & IO_URING_F_NONBLOCK)
3669 ret = vfs_fsync_range(req->file, req->sync.off,
3670 end > 0 ? end : LLONG_MAX,
3671 req->sync.flags & IORING_FSYNC_DATASYNC);
3673 req_set_fail_links(req);
3674 io_req_complete(req, ret);
3678 static int io_fallocate_prep(struct io_kiocb *req,
3679 const struct io_uring_sqe *sqe)
3681 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3683 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3686 req->sync.off = READ_ONCE(sqe->off);
3687 req->sync.len = READ_ONCE(sqe->addr);
3688 req->sync.mode = READ_ONCE(sqe->len);
3692 static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
3696 /* fallocate always requiring blocking context */
3697 if (issue_flags & IO_URING_F_NONBLOCK)
3699 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3702 req_set_fail_links(req);
3703 io_req_complete(req, ret);
3707 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3709 const char __user *fname;
3712 if (unlikely(sqe->ioprio || sqe->buf_index))
3714 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3717 /* open.how should be already initialised */
3718 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3719 req->open.how.flags |= O_LARGEFILE;
3721 req->open.dfd = READ_ONCE(sqe->fd);
3722 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3723 req->open.filename = getname(fname);
3724 if (IS_ERR(req->open.filename)) {
3725 ret = PTR_ERR(req->open.filename);
3726 req->open.filename = NULL;
3729 req->open.nofile = rlimit(RLIMIT_NOFILE);
3730 req->flags |= REQ_F_NEED_CLEANUP;
3734 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3738 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3740 mode = READ_ONCE(sqe->len);
3741 flags = READ_ONCE(sqe->open_flags);
3742 req->open.how = build_open_how(flags, mode);
3743 return __io_openat_prep(req, sqe);
3746 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3748 struct open_how __user *how;
3752 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3754 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3755 len = READ_ONCE(sqe->len);
3756 if (len < OPEN_HOW_SIZE_VER0)
3759 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3764 return __io_openat_prep(req, sqe);
3767 static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
3769 struct open_flags op;
3772 bool resolve_nonblock;
3775 ret = build_open_flags(&req->open.how, &op);
3778 nonblock_set = op.open_flag & O_NONBLOCK;
3779 resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
3780 if (issue_flags & IO_URING_F_NONBLOCK) {
3782 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
3783 * it'll always -EAGAIN
3785 if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
3787 op.lookup_flags |= LOOKUP_CACHED;
3788 op.open_flag |= O_NONBLOCK;
3791 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3795 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3796 /* only retry if RESOLVE_CACHED wasn't already set by application */
3797 if ((!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)) &&
3798 file == ERR_PTR(-EAGAIN)) {
3800 * We could hang on to this 'fd', but seems like marginal
3801 * gain for something that is now known to be a slower path.
3802 * So just put it, and we'll get a new one when we retry.
3810 ret = PTR_ERR(file);
3812 if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
3813 file->f_flags &= ~O_NONBLOCK;
3814 fsnotify_open(file);
3815 fd_install(ret, file);
3818 putname(req->open.filename);
3819 req->flags &= ~REQ_F_NEED_CLEANUP;
3821 req_set_fail_links(req);
3822 io_req_complete(req, ret);
3826 static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
3828 return io_openat2(req, issue_flags & IO_URING_F_NONBLOCK);
3831 static int io_remove_buffers_prep(struct io_kiocb *req,
3832 const struct io_uring_sqe *sqe)
3834 struct io_provide_buf *p = &req->pbuf;
3837 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3840 tmp = READ_ONCE(sqe->fd);
3841 if (!tmp || tmp > USHRT_MAX)
3844 memset(p, 0, sizeof(*p));
3846 p->bgid = READ_ONCE(sqe->buf_group);
3850 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3851 int bgid, unsigned nbufs)
3855 /* shouldn't happen */
3859 /* the head kbuf is the list itself */
3860 while (!list_empty(&buf->list)) {
3861 struct io_buffer *nxt;
3863 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3864 list_del(&nxt->list);
3871 idr_remove(&ctx->io_buffer_idr, bgid);
3876 static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
3878 struct io_provide_buf *p = &req->pbuf;
3879 struct io_ring_ctx *ctx = req->ctx;
3880 struct io_buffer *head;
3882 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3884 io_ring_submit_lock(ctx, !force_nonblock);
3886 lockdep_assert_held(&ctx->uring_lock);
3889 head = idr_find(&ctx->io_buffer_idr, p->bgid);
3891 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3893 req_set_fail_links(req);
3895 /* need to hold the lock to complete IOPOLL requests */
3896 if (ctx->flags & IORING_SETUP_IOPOLL) {
3897 __io_req_complete(req, issue_flags, ret, 0);
3898 io_ring_submit_unlock(ctx, !force_nonblock);
3900 io_ring_submit_unlock(ctx, !force_nonblock);
3901 __io_req_complete(req, issue_flags, ret, 0);
3906 static int io_provide_buffers_prep(struct io_kiocb *req,
3907 const struct io_uring_sqe *sqe)
3909 struct io_provide_buf *p = &req->pbuf;
3912 if (sqe->ioprio || sqe->rw_flags)
3915 tmp = READ_ONCE(sqe->fd);
3916 if (!tmp || tmp > USHRT_MAX)
3919 p->addr = READ_ONCE(sqe->addr);
3920 p->len = READ_ONCE(sqe->len);
3922 if (!access_ok(u64_to_user_ptr(p->addr), (p->len * p->nbufs)))
3925 p->bgid = READ_ONCE(sqe->buf_group);
3926 tmp = READ_ONCE(sqe->off);
3927 if (tmp > USHRT_MAX)
3933 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3935 struct io_buffer *buf;
3936 u64 addr = pbuf->addr;
3937 int i, bid = pbuf->bid;
3939 for (i = 0; i < pbuf->nbufs; i++) {
3940 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
3945 buf->len = pbuf->len;
3950 INIT_LIST_HEAD(&buf->list);
3953 list_add_tail(&buf->list, &(*head)->list);
3957 return i ? i : -ENOMEM;
3960 static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
3962 struct io_provide_buf *p = &req->pbuf;
3963 struct io_ring_ctx *ctx = req->ctx;
3964 struct io_buffer *head, *list;
3966 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3968 io_ring_submit_lock(ctx, !force_nonblock);
3970 lockdep_assert_held(&ctx->uring_lock);
3972 list = head = idr_find(&ctx->io_buffer_idr, p->bgid);
3974 ret = io_add_buffers(p, &head);
3979 ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1,
3982 __io_remove_buffers(ctx, head, p->bgid, -1U);
3988 req_set_fail_links(req);
3990 /* need to hold the lock to complete IOPOLL requests */
3991 if (ctx->flags & IORING_SETUP_IOPOLL) {
3992 __io_req_complete(req, issue_flags, ret, 0);
3993 io_ring_submit_unlock(ctx, !force_nonblock);
3995 io_ring_submit_unlock(ctx, !force_nonblock);
3996 __io_req_complete(req, issue_flags, ret, 0);
4001 static int io_epoll_ctl_prep(struct io_kiocb *req,
4002 const struct io_uring_sqe *sqe)
4004 #if defined(CONFIG_EPOLL)
4005 if (sqe->ioprio || sqe->buf_index)
4007 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4010 req->epoll.epfd = READ_ONCE(sqe->fd);
4011 req->epoll.op = READ_ONCE(sqe->len);
4012 req->epoll.fd = READ_ONCE(sqe->off);
4014 if (ep_op_has_event(req->epoll.op)) {
4015 struct epoll_event __user *ev;
4017 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4018 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4028 static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4030 #if defined(CONFIG_EPOLL)
4031 struct io_epoll *ie = &req->epoll;
4033 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4035 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4036 if (force_nonblock && ret == -EAGAIN)
4040 req_set_fail_links(req);
4041 __io_req_complete(req, issue_flags, ret, 0);
4048 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4050 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4051 if (sqe->ioprio || sqe->buf_index || sqe->off)
4053 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4056 req->madvise.addr = READ_ONCE(sqe->addr);
4057 req->madvise.len = READ_ONCE(sqe->len);
4058 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4065 static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4067 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4068 struct io_madvise *ma = &req->madvise;
4071 if (issue_flags & IO_URING_F_NONBLOCK)
4074 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4076 req_set_fail_links(req);
4077 io_req_complete(req, ret);
4084 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4086 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4088 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4091 req->fadvise.offset = READ_ONCE(sqe->off);
4092 req->fadvise.len = READ_ONCE(sqe->len);
4093 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4097 static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4099 struct io_fadvise *fa = &req->fadvise;
4102 if (issue_flags & IO_URING_F_NONBLOCK) {
4103 switch (fa->advice) {
4104 case POSIX_FADV_NORMAL:
4105 case POSIX_FADV_RANDOM:
4106 case POSIX_FADV_SEQUENTIAL:
4113 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4115 req_set_fail_links(req);
4116 io_req_complete(req, ret);
4120 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4122 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4124 if (sqe->ioprio || sqe->buf_index)
4126 if (req->flags & REQ_F_FIXED_FILE)
4129 req->statx.dfd = READ_ONCE(sqe->fd);
4130 req->statx.mask = READ_ONCE(sqe->len);
4131 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4132 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4133 req->statx.flags = READ_ONCE(sqe->statx_flags);
4138 static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4140 struct io_statx *ctx = &req->statx;
4143 if (issue_flags & IO_URING_F_NONBLOCK) {
4144 /* only need file table for an actual valid fd */
4145 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
4146 req->flags |= REQ_F_NO_FILE_TABLE;
4150 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4154 req_set_fail_links(req);
4155 io_req_complete(req, ret);
4159 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4161 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4163 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4164 sqe->rw_flags || sqe->buf_index)
4166 if (req->flags & REQ_F_FIXED_FILE)
4169 req->close.fd = READ_ONCE(sqe->fd);
4173 static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4175 struct files_struct *files = current->files;
4176 struct io_close *close = &req->close;
4177 struct fdtable *fdt;
4183 spin_lock(&files->file_lock);
4184 fdt = files_fdtable(files);
4185 if (close->fd >= fdt->max_fds) {
4186 spin_unlock(&files->file_lock);
4189 file = fdt->fd[close->fd];
4191 spin_unlock(&files->file_lock);
4195 if (file->f_op == &io_uring_fops) {
4196 spin_unlock(&files->file_lock);
4201 /* if the file has a flush method, be safe and punt to async */
4202 if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4203 spin_unlock(&files->file_lock);
4207 ret = __close_fd_get_file(close->fd, &file);
4208 spin_unlock(&files->file_lock);
4215 /* No ->flush() or already async, safely close from here */
4216 ret = filp_close(file, current->files);
4219 req_set_fail_links(req);
4222 __io_req_complete(req, issue_flags, ret, 0);
4226 static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4228 struct io_ring_ctx *ctx = req->ctx;
4230 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4232 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4235 req->sync.off = READ_ONCE(sqe->off);
4236 req->sync.len = READ_ONCE(sqe->len);
4237 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4241 static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4245 /* sync_file_range always requires a blocking context */
4246 if (issue_flags & IO_URING_F_NONBLOCK)
4249 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4252 req_set_fail_links(req);
4253 io_req_complete(req, ret);
4257 #if defined(CONFIG_NET)
4258 static int io_setup_async_msg(struct io_kiocb *req,
4259 struct io_async_msghdr *kmsg)
4261 struct io_async_msghdr *async_msg = req->async_data;
4265 if (io_alloc_async_data(req)) {
4266 kfree(kmsg->free_iov);
4269 async_msg = req->async_data;
4270 req->flags |= REQ_F_NEED_CLEANUP;
4271 memcpy(async_msg, kmsg, sizeof(*kmsg));
4272 async_msg->msg.msg_name = &async_msg->addr;
4273 /* if were using fast_iov, set it to the new one */
4274 if (!async_msg->free_iov)
4275 async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4280 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4281 struct io_async_msghdr *iomsg)
4283 iomsg->msg.msg_name = &iomsg->addr;
4284 iomsg->free_iov = iomsg->fast_iov;
4285 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4286 req->sr_msg.msg_flags, &iomsg->free_iov);
4289 static int io_sendmsg_prep_async(struct io_kiocb *req)
4293 if (!io_op_defs[req->opcode].needs_async_data)
4295 ret = io_sendmsg_copy_hdr(req, req->async_data);
4297 req->flags |= REQ_F_NEED_CLEANUP;
4301 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4303 struct io_sr_msg *sr = &req->sr_msg;
4305 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4308 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4309 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4310 sr->len = READ_ONCE(sqe->len);
4312 #ifdef CONFIG_COMPAT
4313 if (req->ctx->compat)
4314 sr->msg_flags |= MSG_CMSG_COMPAT;
4319 static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4321 struct io_async_msghdr iomsg, *kmsg;
4322 struct socket *sock;
4326 sock = sock_from_file(req->file);
4327 if (unlikely(!sock))
4330 kmsg = req->async_data;
4332 ret = io_sendmsg_copy_hdr(req, &iomsg);
4338 flags = req->sr_msg.msg_flags;
4339 if (flags & MSG_DONTWAIT)
4340 req->flags |= REQ_F_NOWAIT;
4341 else if (issue_flags & IO_URING_F_NONBLOCK)
4342 flags |= MSG_DONTWAIT;
4344 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4345 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4346 return io_setup_async_msg(req, kmsg);
4347 if (ret == -ERESTARTSYS)
4350 /* fast path, check for non-NULL to avoid function call */
4352 kfree(kmsg->free_iov);
4353 req->flags &= ~REQ_F_NEED_CLEANUP;
4355 req_set_fail_links(req);
4356 __io_req_complete(req, issue_flags, ret, 0);
4360 static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4362 struct io_sr_msg *sr = &req->sr_msg;
4365 struct socket *sock;
4369 sock = sock_from_file(req->file);
4370 if (unlikely(!sock))
4373 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4377 msg.msg_name = NULL;
4378 msg.msg_control = NULL;
4379 msg.msg_controllen = 0;
4380 msg.msg_namelen = 0;
4382 flags = req->sr_msg.msg_flags;
4383 if (flags & MSG_DONTWAIT)
4384 req->flags |= REQ_F_NOWAIT;
4385 else if (issue_flags & IO_URING_F_NONBLOCK)
4386 flags |= MSG_DONTWAIT;
4388 msg.msg_flags = flags;
4389 ret = sock_sendmsg(sock, &msg);
4390 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4392 if (ret == -ERESTARTSYS)
4396 req_set_fail_links(req);
4397 __io_req_complete(req, issue_flags, ret, 0);
4401 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4402 struct io_async_msghdr *iomsg)
4404 struct io_sr_msg *sr = &req->sr_msg;
4405 struct iovec __user *uiov;
4409 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4410 &iomsg->uaddr, &uiov, &iov_len);
4414 if (req->flags & REQ_F_BUFFER_SELECT) {
4417 if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4419 sr->len = iomsg->fast_iov[0].iov_len;
4420 iomsg->free_iov = NULL;
4422 iomsg->free_iov = iomsg->fast_iov;
4423 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4424 &iomsg->free_iov, &iomsg->msg.msg_iter,
4433 #ifdef CONFIG_COMPAT
4434 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4435 struct io_async_msghdr *iomsg)
4437 struct compat_msghdr __user *msg_compat;
4438 struct io_sr_msg *sr = &req->sr_msg;
4439 struct compat_iovec __user *uiov;
4444 msg_compat = (struct compat_msghdr __user *) sr->umsg;
4445 ret = __get_compat_msghdr(&iomsg->msg, msg_compat, &iomsg->uaddr,
4450 uiov = compat_ptr(ptr);
4451 if (req->flags & REQ_F_BUFFER_SELECT) {
4452 compat_ssize_t clen;
4456 if (!access_ok(uiov, sizeof(*uiov)))
4458 if (__get_user(clen, &uiov->iov_len))
4463 iomsg->free_iov = NULL;
4465 iomsg->free_iov = iomsg->fast_iov;
4466 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4467 UIO_FASTIOV, &iomsg->free_iov,
4468 &iomsg->msg.msg_iter, true);
4477 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4478 struct io_async_msghdr *iomsg)
4480 iomsg->msg.msg_name = &iomsg->addr;
4482 #ifdef CONFIG_COMPAT
4483 if (req->ctx->compat)
4484 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4487 return __io_recvmsg_copy_hdr(req, iomsg);
4490 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4493 struct io_sr_msg *sr = &req->sr_msg;
4494 struct io_buffer *kbuf;
4496 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4501 req->flags |= REQ_F_BUFFER_SELECTED;
4505 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4507 return io_put_kbuf(req, req->sr_msg.kbuf);
4510 static int io_recvmsg_prep_async(struct io_kiocb *req)
4514 if (!io_op_defs[req->opcode].needs_async_data)
4516 ret = io_recvmsg_copy_hdr(req, req->async_data);
4518 req->flags |= REQ_F_NEED_CLEANUP;
4522 static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4524 struct io_sr_msg *sr = &req->sr_msg;
4526 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4529 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4530 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4531 sr->len = READ_ONCE(sqe->len);
4532 sr->bgid = READ_ONCE(sqe->buf_group);
4534 #ifdef CONFIG_COMPAT
4535 if (req->ctx->compat)
4536 sr->msg_flags |= MSG_CMSG_COMPAT;
4541 static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
4543 struct io_async_msghdr iomsg, *kmsg;
4544 struct socket *sock;
4545 struct io_buffer *kbuf;
4547 int ret, cflags = 0;
4548 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4550 sock = sock_from_file(req->file);
4551 if (unlikely(!sock))
4554 kmsg = req->async_data;
4556 ret = io_recvmsg_copy_hdr(req, &iomsg);
4562 if (req->flags & REQ_F_BUFFER_SELECT) {
4563 kbuf = io_recv_buffer_select(req, !force_nonblock);
4565 return PTR_ERR(kbuf);
4566 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4567 kmsg->fast_iov[0].iov_len = req->sr_msg.len;
4568 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
4569 1, req->sr_msg.len);
4572 flags = req->sr_msg.msg_flags;
4573 if (flags & MSG_DONTWAIT)
4574 req->flags |= REQ_F_NOWAIT;
4575 else if (force_nonblock)
4576 flags |= MSG_DONTWAIT;
4578 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4579 kmsg->uaddr, flags);
4580 if (force_nonblock && ret == -EAGAIN)
4581 return io_setup_async_msg(req, kmsg);
4582 if (ret == -ERESTARTSYS)
4585 if (req->flags & REQ_F_BUFFER_SELECTED)
4586 cflags = io_put_recv_kbuf(req);
4587 /* fast path, check for non-NULL to avoid function call */
4589 kfree(kmsg->free_iov);
4590 req->flags &= ~REQ_F_NEED_CLEANUP;
4592 req_set_fail_links(req);
4593 __io_req_complete(req, issue_flags, ret, cflags);
4597 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
4599 struct io_buffer *kbuf;
4600 struct io_sr_msg *sr = &req->sr_msg;
4602 void __user *buf = sr->buf;
4603 struct socket *sock;
4606 int ret, cflags = 0;
4607 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4609 sock = sock_from_file(req->file);
4610 if (unlikely(!sock))
4613 if (req->flags & REQ_F_BUFFER_SELECT) {
4614 kbuf = io_recv_buffer_select(req, !force_nonblock);
4616 return PTR_ERR(kbuf);
4617 buf = u64_to_user_ptr(kbuf->addr);
4620 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4624 msg.msg_name = NULL;
4625 msg.msg_control = NULL;
4626 msg.msg_controllen = 0;
4627 msg.msg_namelen = 0;
4628 msg.msg_iocb = NULL;
4631 flags = req->sr_msg.msg_flags;
4632 if (flags & MSG_DONTWAIT)
4633 req->flags |= REQ_F_NOWAIT;
4634 else if (force_nonblock)
4635 flags |= MSG_DONTWAIT;
4637 ret = sock_recvmsg(sock, &msg, flags);
4638 if (force_nonblock && ret == -EAGAIN)
4640 if (ret == -ERESTARTSYS)
4643 if (req->flags & REQ_F_BUFFER_SELECTED)
4644 cflags = io_put_recv_kbuf(req);
4646 req_set_fail_links(req);
4647 __io_req_complete(req, issue_flags, ret, cflags);
4651 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4653 struct io_accept *accept = &req->accept;
4655 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4657 if (sqe->ioprio || sqe->len || sqe->buf_index)
4660 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4661 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4662 accept->flags = READ_ONCE(sqe->accept_flags);
4663 accept->nofile = rlimit(RLIMIT_NOFILE);
4667 static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
4669 struct io_accept *accept = &req->accept;
4670 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4671 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4674 if (req->file->f_flags & O_NONBLOCK)
4675 req->flags |= REQ_F_NOWAIT;
4677 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4678 accept->addr_len, accept->flags,
4680 if (ret == -EAGAIN && force_nonblock)
4683 if (ret == -ERESTARTSYS)
4685 req_set_fail_links(req);
4687 __io_req_complete(req, issue_flags, ret, 0);
4691 static int io_connect_prep_async(struct io_kiocb *req)
4693 struct io_async_connect *io = req->async_data;
4694 struct io_connect *conn = &req->connect;
4696 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
4699 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4701 struct io_connect *conn = &req->connect;
4703 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4705 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4708 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4709 conn->addr_len = READ_ONCE(sqe->addr2);
4713 static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
4715 struct io_async_connect __io, *io;
4716 unsigned file_flags;
4718 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4720 if (req->async_data) {
4721 io = req->async_data;
4723 ret = move_addr_to_kernel(req->connect.addr,
4724 req->connect.addr_len,
4731 file_flags = force_nonblock ? O_NONBLOCK : 0;
4733 ret = __sys_connect_file(req->file, &io->address,
4734 req->connect.addr_len, file_flags);
4735 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4736 if (req->async_data)
4738 if (io_alloc_async_data(req)) {
4742 io = req->async_data;
4743 memcpy(req->async_data, &__io, sizeof(__io));
4746 if (ret == -ERESTARTSYS)
4750 req_set_fail_links(req);
4751 __io_req_complete(req, issue_flags, ret, 0);
4754 #else /* !CONFIG_NET */
4755 #define IO_NETOP_FN(op) \
4756 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
4758 return -EOPNOTSUPP; \
4761 #define IO_NETOP_PREP(op) \
4763 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
4765 return -EOPNOTSUPP; \
4768 #define IO_NETOP_PREP_ASYNC(op) \
4770 static int io_##op##_prep_async(struct io_kiocb *req) \
4772 return -EOPNOTSUPP; \
4775 IO_NETOP_PREP_ASYNC(sendmsg);
4776 IO_NETOP_PREP_ASYNC(recvmsg);
4777 IO_NETOP_PREP_ASYNC(connect);
4778 IO_NETOP_PREP(accept);
4781 #endif /* CONFIG_NET */
4783 struct io_poll_table {
4784 struct poll_table_struct pt;
4785 struct io_kiocb *req;
4789 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4790 __poll_t mask, task_work_func_t func)
4794 /* for instances that support it check for an event match first: */
4795 if (mask && !(mask & poll->events))
4798 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4800 list_del_init(&poll->wait.entry);
4803 req->task_work.func = func;
4804 percpu_ref_get(&req->ctx->refs);
4807 * If this fails, then the task is exiting. When a task exits, the
4808 * work gets canceled, so just cancel this request as well instead
4809 * of executing it. We can't safely execute it anyway, as we may not
4810 * have the needed state needed for it anyway.
4812 ret = io_req_task_work_add(req);
4813 if (unlikely(ret)) {
4814 WRITE_ONCE(poll->canceled, true);
4815 io_req_task_work_add_fallback(req, func);
4820 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4821 __acquires(&req->ctx->completion_lock)
4823 struct io_ring_ctx *ctx = req->ctx;
4825 if (!req->result && !READ_ONCE(poll->canceled)) {
4826 struct poll_table_struct pt = { ._key = poll->events };
4828 req->result = vfs_poll(req->file, &pt) & poll->events;
4831 spin_lock_irq(&ctx->completion_lock);
4832 if (!req->result && !READ_ONCE(poll->canceled)) {
4833 add_wait_queue(poll->head, &poll->wait);
4840 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4842 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4843 if (req->opcode == IORING_OP_POLL_ADD)
4844 return req->async_data;
4845 return req->apoll->double_poll;
4848 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4850 if (req->opcode == IORING_OP_POLL_ADD)
4852 return &req->apoll->poll;
4855 static void io_poll_remove_double(struct io_kiocb *req)
4857 struct io_poll_iocb *poll = io_poll_get_double(req);
4859 lockdep_assert_held(&req->ctx->completion_lock);
4861 if (poll && poll->head) {
4862 struct wait_queue_head *head = poll->head;
4864 spin_lock(&head->lock);
4865 list_del_init(&poll->wait.entry);
4866 if (poll->wait.private)
4867 refcount_dec(&req->refs);
4869 spin_unlock(&head->lock);
4873 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
4875 struct io_ring_ctx *ctx = req->ctx;
4877 io_poll_remove_double(req);
4878 req->poll.done = true;
4879 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
4880 io_commit_cqring(ctx);
4883 static void io_poll_task_func(struct callback_head *cb)
4885 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4886 struct io_ring_ctx *ctx = req->ctx;
4887 struct io_kiocb *nxt;
4889 if (io_poll_rewait(req, &req->poll)) {
4890 spin_unlock_irq(&ctx->completion_lock);
4892 hash_del(&req->hash_node);
4893 io_poll_complete(req, req->result, 0);
4894 spin_unlock_irq(&ctx->completion_lock);
4896 nxt = io_put_req_find_next(req);
4897 io_cqring_ev_posted(ctx);
4899 __io_req_task_submit(nxt);
4902 percpu_ref_put(&ctx->refs);
4905 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4906 int sync, void *key)
4908 struct io_kiocb *req = wait->private;
4909 struct io_poll_iocb *poll = io_poll_get_single(req);
4910 __poll_t mask = key_to_poll(key);
4912 /* for instances that support it check for an event match first: */
4913 if (mask && !(mask & poll->events))
4916 list_del_init(&wait->entry);
4918 if (poll && poll->head) {
4921 spin_lock(&poll->head->lock);
4922 done = list_empty(&poll->wait.entry);
4924 list_del_init(&poll->wait.entry);
4925 /* make sure double remove sees this as being gone */
4926 wait->private = NULL;
4927 spin_unlock(&poll->head->lock);
4929 /* use wait func handler, so it matches the rq type */
4930 poll->wait.func(&poll->wait, mode, sync, key);
4933 refcount_dec(&req->refs);
4937 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
4938 wait_queue_func_t wake_func)
4942 poll->canceled = false;
4943 poll->events = events;
4944 INIT_LIST_HEAD(&poll->wait.entry);
4945 init_waitqueue_func_entry(&poll->wait, wake_func);
4948 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
4949 struct wait_queue_head *head,
4950 struct io_poll_iocb **poll_ptr)
4952 struct io_kiocb *req = pt->req;
4955 * If poll->head is already set, it's because the file being polled
4956 * uses multiple waitqueues for poll handling (eg one for read, one
4957 * for write). Setup a separate io_poll_iocb if this happens.
4959 if (unlikely(poll->head)) {
4960 struct io_poll_iocb *poll_one = poll;
4962 /* already have a 2nd entry, fail a third attempt */
4964 pt->error = -EINVAL;
4967 /* double add on the same waitqueue head, ignore */
4968 if (poll->head == head)
4970 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
4972 pt->error = -ENOMEM;
4975 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
4976 refcount_inc(&req->refs);
4977 poll->wait.private = req;
4984 if (poll->events & EPOLLEXCLUSIVE)
4985 add_wait_queue_exclusive(head, &poll->wait);
4987 add_wait_queue(head, &poll->wait);
4990 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
4991 struct poll_table_struct *p)
4993 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
4994 struct async_poll *apoll = pt->req->apoll;
4996 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
4999 static void io_async_task_func(struct callback_head *cb)
5001 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
5002 struct async_poll *apoll = req->apoll;
5003 struct io_ring_ctx *ctx = req->ctx;
5005 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
5007 if (io_poll_rewait(req, &apoll->poll)) {
5008 spin_unlock_irq(&ctx->completion_lock);
5009 percpu_ref_put(&ctx->refs);
5013 /* If req is still hashed, it cannot have been canceled. Don't check. */
5014 if (hash_hashed(&req->hash_node))
5015 hash_del(&req->hash_node);
5017 io_poll_remove_double(req);
5018 spin_unlock_irq(&ctx->completion_lock);
5020 if (!READ_ONCE(apoll->poll.canceled))
5021 __io_req_task_submit(req);
5023 __io_req_task_cancel(req, -ECANCELED);
5025 percpu_ref_put(&ctx->refs);
5026 kfree(apoll->double_poll);
5030 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5033 struct io_kiocb *req = wait->private;
5034 struct io_poll_iocb *poll = &req->apoll->poll;
5036 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5039 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5042 static void io_poll_req_insert(struct io_kiocb *req)
5044 struct io_ring_ctx *ctx = req->ctx;
5045 struct hlist_head *list;
5047 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5048 hlist_add_head(&req->hash_node, list);
5051 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5052 struct io_poll_iocb *poll,
5053 struct io_poll_table *ipt, __poll_t mask,
5054 wait_queue_func_t wake_func)
5055 __acquires(&ctx->completion_lock)
5057 struct io_ring_ctx *ctx = req->ctx;
5058 bool cancel = false;
5060 INIT_HLIST_NODE(&req->hash_node);
5061 io_init_poll_iocb(poll, mask, wake_func);
5062 poll->file = req->file;
5063 poll->wait.private = req;
5065 ipt->pt._key = mask;
5067 ipt->error = -EINVAL;
5069 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5071 spin_lock_irq(&ctx->completion_lock);
5072 if (likely(poll->head)) {
5073 spin_lock(&poll->head->lock);
5074 if (unlikely(list_empty(&poll->wait.entry))) {
5080 if (mask || ipt->error)
5081 list_del_init(&poll->wait.entry);
5083 WRITE_ONCE(poll->canceled, true);
5084 else if (!poll->done) /* actually waiting for an event */
5085 io_poll_req_insert(req);
5086 spin_unlock(&poll->head->lock);
5092 static bool io_arm_poll_handler(struct io_kiocb *req)
5094 const struct io_op_def *def = &io_op_defs[req->opcode];
5095 struct io_ring_ctx *ctx = req->ctx;
5096 struct async_poll *apoll;
5097 struct io_poll_table ipt;
5101 if (!req->file || !file_can_poll(req->file))
5103 if (req->flags & REQ_F_POLLED)
5107 else if (def->pollout)
5111 /* if we can't nonblock try, then no point in arming a poll handler */
5112 if (!io_file_supports_async(req->file, rw))
5115 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5116 if (unlikely(!apoll))
5118 apoll->double_poll = NULL;
5120 req->flags |= REQ_F_POLLED;
5125 mask |= POLLIN | POLLRDNORM;
5127 mask |= POLLOUT | POLLWRNORM;
5129 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5130 if ((req->opcode == IORING_OP_RECVMSG) &&
5131 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5134 mask |= POLLERR | POLLPRI;
5136 ipt.pt._qproc = io_async_queue_proc;
5138 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5140 if (ret || ipt.error) {
5141 io_poll_remove_double(req);
5142 spin_unlock_irq(&ctx->completion_lock);
5143 kfree(apoll->double_poll);
5147 spin_unlock_irq(&ctx->completion_lock);
5148 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
5149 apoll->poll.events);
5153 static bool __io_poll_remove_one(struct io_kiocb *req,
5154 struct io_poll_iocb *poll)
5156 bool do_complete = false;
5158 spin_lock(&poll->head->lock);
5159 WRITE_ONCE(poll->canceled, true);
5160 if (!list_empty(&poll->wait.entry)) {
5161 list_del_init(&poll->wait.entry);
5164 spin_unlock(&poll->head->lock);
5165 hash_del(&req->hash_node);
5169 static bool io_poll_remove_one(struct io_kiocb *req)
5173 io_poll_remove_double(req);
5175 if (req->opcode == IORING_OP_POLL_ADD) {
5176 do_complete = __io_poll_remove_one(req, &req->poll);
5178 struct async_poll *apoll = req->apoll;
5180 /* non-poll requests have submit ref still */
5181 do_complete = __io_poll_remove_one(req, &apoll->poll);
5184 kfree(apoll->double_poll);
5190 io_cqring_fill_event(req, -ECANCELED);
5191 io_commit_cqring(req->ctx);
5192 req_set_fail_links(req);
5193 io_put_req_deferred(req, 1);
5200 * Returns true if we found and killed one or more poll requests
5202 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5203 struct files_struct *files)
5205 struct hlist_node *tmp;
5206 struct io_kiocb *req;
5209 spin_lock_irq(&ctx->completion_lock);
5210 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5211 struct hlist_head *list;
5213 list = &ctx->cancel_hash[i];
5214 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5215 if (io_match_task(req, tsk, files))
5216 posted += io_poll_remove_one(req);
5219 spin_unlock_irq(&ctx->completion_lock);
5222 io_cqring_ev_posted(ctx);
5227 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
5229 struct hlist_head *list;
5230 struct io_kiocb *req;
5232 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5233 hlist_for_each_entry(req, list, hash_node) {
5234 if (sqe_addr != req->user_data)
5236 if (io_poll_remove_one(req))
5244 static int io_poll_remove_prep(struct io_kiocb *req,
5245 const struct io_uring_sqe *sqe)
5247 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5249 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
5253 req->poll_remove.addr = READ_ONCE(sqe->addr);
5258 * Find a running poll command that matches one specified in sqe->addr,
5259 * and remove it if found.
5261 static int io_poll_remove(struct io_kiocb *req, unsigned int issue_flags)
5263 struct io_ring_ctx *ctx = req->ctx;
5266 spin_lock_irq(&ctx->completion_lock);
5267 ret = io_poll_cancel(ctx, req->poll_remove.addr);
5268 spin_unlock_irq(&ctx->completion_lock);
5271 req_set_fail_links(req);
5272 io_req_complete(req, ret);
5276 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5279 struct io_kiocb *req = wait->private;
5280 struct io_poll_iocb *poll = &req->poll;
5282 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5285 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5286 struct poll_table_struct *p)
5288 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5290 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5293 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5295 struct io_poll_iocb *poll = &req->poll;
5298 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5300 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
5303 events = READ_ONCE(sqe->poll32_events);
5305 events = swahw32(events);
5307 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP |
5308 (events & EPOLLEXCLUSIVE);
5312 static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5314 struct io_poll_iocb *poll = &req->poll;
5315 struct io_ring_ctx *ctx = req->ctx;
5316 struct io_poll_table ipt;
5319 ipt.pt._qproc = io_poll_queue_proc;
5321 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5324 if (mask) { /* no async, we'd stolen it */
5326 io_poll_complete(req, mask, 0);
5328 spin_unlock_irq(&ctx->completion_lock);
5331 io_cqring_ev_posted(ctx);
5337 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5339 struct io_timeout_data *data = container_of(timer,
5340 struct io_timeout_data, timer);
5341 struct io_kiocb *req = data->req;
5342 struct io_ring_ctx *ctx = req->ctx;
5343 unsigned long flags;
5345 spin_lock_irqsave(&ctx->completion_lock, flags);
5346 list_del_init(&req->timeout.list);
5347 atomic_set(&req->ctx->cq_timeouts,
5348 atomic_read(&req->ctx->cq_timeouts) + 1);
5350 io_cqring_fill_event(req, -ETIME);
5351 io_commit_cqring(ctx);
5352 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5354 io_cqring_ev_posted(ctx);
5355 req_set_fail_links(req);
5357 return HRTIMER_NORESTART;
5360 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5363 struct io_timeout_data *io;
5364 struct io_kiocb *req;
5367 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5368 if (user_data == req->user_data) {
5375 return ERR_PTR(ret);
5377 io = req->async_data;
5378 ret = hrtimer_try_to_cancel(&io->timer);
5380 return ERR_PTR(-EALREADY);
5381 list_del_init(&req->timeout.list);
5385 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5387 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5390 return PTR_ERR(req);
5392 req_set_fail_links(req);
5393 io_cqring_fill_event(req, -ECANCELED);
5394 io_put_req_deferred(req, 1);
5398 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5399 struct timespec64 *ts, enum hrtimer_mode mode)
5401 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5402 struct io_timeout_data *data;
5405 return PTR_ERR(req);
5407 req->timeout.off = 0; /* noseq */
5408 data = req->async_data;
5409 list_add_tail(&req->timeout.list, &ctx->timeout_list);
5410 hrtimer_init(&data->timer, CLOCK_MONOTONIC, mode);
5411 data->timer.function = io_timeout_fn;
5412 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
5416 static int io_timeout_remove_prep(struct io_kiocb *req,
5417 const struct io_uring_sqe *sqe)
5419 struct io_timeout_rem *tr = &req->timeout_rem;
5421 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5423 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5425 if (sqe->ioprio || sqe->buf_index || sqe->len)
5428 tr->addr = READ_ONCE(sqe->addr);
5429 tr->flags = READ_ONCE(sqe->timeout_flags);
5430 if (tr->flags & IORING_TIMEOUT_UPDATE) {
5431 if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS))
5433 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
5435 } else if (tr->flags) {
5436 /* timeout removal doesn't support flags */
5443 static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
5445 return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
5450 * Remove or update an existing timeout command
5452 static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
5454 struct io_timeout_rem *tr = &req->timeout_rem;
5455 struct io_ring_ctx *ctx = req->ctx;
5458 spin_lock_irq(&ctx->completion_lock);
5459 if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE))
5460 ret = io_timeout_cancel(ctx, tr->addr);
5462 ret = io_timeout_update(ctx, tr->addr, &tr->ts,
5463 io_translate_timeout_mode(tr->flags));
5465 io_cqring_fill_event(req, ret);
5466 io_commit_cqring(ctx);
5467 spin_unlock_irq(&ctx->completion_lock);
5468 io_cqring_ev_posted(ctx);
5470 req_set_fail_links(req);
5475 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5476 bool is_timeout_link)
5478 struct io_timeout_data *data;
5480 u32 off = READ_ONCE(sqe->off);
5482 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5484 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5486 if (off && is_timeout_link)
5488 flags = READ_ONCE(sqe->timeout_flags);
5489 if (flags & ~IORING_TIMEOUT_ABS)
5492 req->timeout.off = off;
5494 if (!req->async_data && io_alloc_async_data(req))
5497 data = req->async_data;
5500 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5503 data->mode = io_translate_timeout_mode(flags);
5504 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5508 static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
5510 struct io_ring_ctx *ctx = req->ctx;
5511 struct io_timeout_data *data = req->async_data;
5512 struct list_head *entry;
5513 u32 tail, off = req->timeout.off;
5515 spin_lock_irq(&ctx->completion_lock);
5518 * sqe->off holds how many events that need to occur for this
5519 * timeout event to be satisfied. If it isn't set, then this is
5520 * a pure timeout request, sequence isn't used.
5522 if (io_is_timeout_noseq(req)) {
5523 entry = ctx->timeout_list.prev;
5527 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5528 req->timeout.target_seq = tail + off;
5530 /* Update the last seq here in case io_flush_timeouts() hasn't.
5531 * This is safe because ->completion_lock is held, and submissions
5532 * and completions are never mixed in the same ->completion_lock section.
5534 ctx->cq_last_tm_flush = tail;
5537 * Insertion sort, ensuring the first entry in the list is always
5538 * the one we need first.
5540 list_for_each_prev(entry, &ctx->timeout_list) {
5541 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5544 if (io_is_timeout_noseq(nxt))
5546 /* nxt.seq is behind @tail, otherwise would've been completed */
5547 if (off >= nxt->timeout.target_seq - tail)
5551 list_add(&req->timeout.list, entry);
5552 data->timer.function = io_timeout_fn;
5553 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5554 spin_unlock_irq(&ctx->completion_lock);
5558 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5560 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5562 return req->user_data == (unsigned long) data;
5565 static int io_async_cancel_one(struct io_uring_task *tctx, void *sqe_addr)
5567 enum io_wq_cancel cancel_ret;
5573 cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, sqe_addr, false);
5574 switch (cancel_ret) {
5575 case IO_WQ_CANCEL_OK:
5578 case IO_WQ_CANCEL_RUNNING:
5581 case IO_WQ_CANCEL_NOTFOUND:
5589 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5590 struct io_kiocb *req, __u64 sqe_addr,
5593 unsigned long flags;
5596 ret = io_async_cancel_one(req->task->io_uring,
5597 (void *) (unsigned long) sqe_addr);
5598 if (ret != -ENOENT) {
5599 spin_lock_irqsave(&ctx->completion_lock, flags);
5603 spin_lock_irqsave(&ctx->completion_lock, flags);
5604 ret = io_timeout_cancel(ctx, sqe_addr);
5607 ret = io_poll_cancel(ctx, sqe_addr);
5611 io_cqring_fill_event(req, ret);
5612 io_commit_cqring(ctx);
5613 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5614 io_cqring_ev_posted(ctx);
5617 req_set_fail_links(req);
5621 static int io_async_cancel_prep(struct io_kiocb *req,
5622 const struct io_uring_sqe *sqe)
5624 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5626 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5628 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5631 req->cancel.addr = READ_ONCE(sqe->addr);
5635 static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
5637 struct io_ring_ctx *ctx = req->ctx;
5639 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
5643 static int io_rsrc_update_prep(struct io_kiocb *req,
5644 const struct io_uring_sqe *sqe)
5646 if (unlikely(req->ctx->flags & IORING_SETUP_SQPOLL))
5648 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5650 if (sqe->ioprio || sqe->rw_flags)
5653 req->rsrc_update.offset = READ_ONCE(sqe->off);
5654 req->rsrc_update.nr_args = READ_ONCE(sqe->len);
5655 if (!req->rsrc_update.nr_args)
5657 req->rsrc_update.arg = READ_ONCE(sqe->addr);
5661 static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
5663 struct io_ring_ctx *ctx = req->ctx;
5664 struct io_uring_rsrc_update up;
5667 if (issue_flags & IO_URING_F_NONBLOCK)
5670 up.offset = req->rsrc_update.offset;
5671 up.data = req->rsrc_update.arg;
5673 mutex_lock(&ctx->uring_lock);
5674 ret = __io_sqe_files_update(ctx, &up, req->rsrc_update.nr_args);
5675 mutex_unlock(&ctx->uring_lock);
5678 req_set_fail_links(req);
5679 __io_req_complete(req, issue_flags, ret, 0);
5683 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5685 switch (req->opcode) {
5688 case IORING_OP_READV:
5689 case IORING_OP_READ_FIXED:
5690 case IORING_OP_READ:
5691 return io_read_prep(req, sqe);
5692 case IORING_OP_WRITEV:
5693 case IORING_OP_WRITE_FIXED:
5694 case IORING_OP_WRITE:
5695 return io_write_prep(req, sqe);
5696 case IORING_OP_POLL_ADD:
5697 return io_poll_add_prep(req, sqe);
5698 case IORING_OP_POLL_REMOVE:
5699 return io_poll_remove_prep(req, sqe);
5700 case IORING_OP_FSYNC:
5701 return io_fsync_prep(req, sqe);
5702 case IORING_OP_SYNC_FILE_RANGE:
5703 return io_sfr_prep(req, sqe);
5704 case IORING_OP_SENDMSG:
5705 case IORING_OP_SEND:
5706 return io_sendmsg_prep(req, sqe);
5707 case IORING_OP_RECVMSG:
5708 case IORING_OP_RECV:
5709 return io_recvmsg_prep(req, sqe);
5710 case IORING_OP_CONNECT:
5711 return io_connect_prep(req, sqe);
5712 case IORING_OP_TIMEOUT:
5713 return io_timeout_prep(req, sqe, false);
5714 case IORING_OP_TIMEOUT_REMOVE:
5715 return io_timeout_remove_prep(req, sqe);
5716 case IORING_OP_ASYNC_CANCEL:
5717 return io_async_cancel_prep(req, sqe);
5718 case IORING_OP_LINK_TIMEOUT:
5719 return io_timeout_prep(req, sqe, true);
5720 case IORING_OP_ACCEPT:
5721 return io_accept_prep(req, sqe);
5722 case IORING_OP_FALLOCATE:
5723 return io_fallocate_prep(req, sqe);
5724 case IORING_OP_OPENAT:
5725 return io_openat_prep(req, sqe);
5726 case IORING_OP_CLOSE:
5727 return io_close_prep(req, sqe);
5728 case IORING_OP_FILES_UPDATE:
5729 return io_rsrc_update_prep(req, sqe);
5730 case IORING_OP_STATX:
5731 return io_statx_prep(req, sqe);
5732 case IORING_OP_FADVISE:
5733 return io_fadvise_prep(req, sqe);
5734 case IORING_OP_MADVISE:
5735 return io_madvise_prep(req, sqe);
5736 case IORING_OP_OPENAT2:
5737 return io_openat2_prep(req, sqe);
5738 case IORING_OP_EPOLL_CTL:
5739 return io_epoll_ctl_prep(req, sqe);
5740 case IORING_OP_SPLICE:
5741 return io_splice_prep(req, sqe);
5742 case IORING_OP_PROVIDE_BUFFERS:
5743 return io_provide_buffers_prep(req, sqe);
5744 case IORING_OP_REMOVE_BUFFERS:
5745 return io_remove_buffers_prep(req, sqe);
5747 return io_tee_prep(req, sqe);
5748 case IORING_OP_SHUTDOWN:
5749 return io_shutdown_prep(req, sqe);
5750 case IORING_OP_RENAMEAT:
5751 return io_renameat_prep(req, sqe);
5752 case IORING_OP_UNLINKAT:
5753 return io_unlinkat_prep(req, sqe);
5756 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5761 static int io_req_prep_async(struct io_kiocb *req)
5763 switch (req->opcode) {
5764 case IORING_OP_READV:
5765 case IORING_OP_READ_FIXED:
5766 case IORING_OP_READ:
5767 return io_rw_prep_async(req, READ);
5768 case IORING_OP_WRITEV:
5769 case IORING_OP_WRITE_FIXED:
5770 case IORING_OP_WRITE:
5771 return io_rw_prep_async(req, WRITE);
5772 case IORING_OP_SENDMSG:
5773 case IORING_OP_SEND:
5774 return io_sendmsg_prep_async(req);
5775 case IORING_OP_RECVMSG:
5776 case IORING_OP_RECV:
5777 return io_recvmsg_prep_async(req);
5778 case IORING_OP_CONNECT:
5779 return io_connect_prep_async(req);
5784 static int io_req_defer_prep(struct io_kiocb *req)
5786 if (!io_op_defs[req->opcode].needs_async_data)
5788 /* some opcodes init it during the inital prep */
5789 if (req->async_data)
5791 if (__io_alloc_async_data(req))
5793 return io_req_prep_async(req);
5796 static u32 io_get_sequence(struct io_kiocb *req)
5798 struct io_kiocb *pos;
5799 struct io_ring_ctx *ctx = req->ctx;
5800 u32 total_submitted, nr_reqs = 0;
5802 io_for_each_link(pos, req)
5805 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
5806 return total_submitted - nr_reqs;
5809 static int io_req_defer(struct io_kiocb *req)
5811 struct io_ring_ctx *ctx = req->ctx;
5812 struct io_defer_entry *de;
5816 /* Still need defer if there is pending req in defer list. */
5817 if (likely(list_empty_careful(&ctx->defer_list) &&
5818 !(req->flags & REQ_F_IO_DRAIN)))
5821 seq = io_get_sequence(req);
5822 /* Still a chance to pass the sequence check */
5823 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
5826 ret = io_req_defer_prep(req);
5829 io_prep_async_link(req);
5830 de = kmalloc(sizeof(*de), GFP_KERNEL);
5834 spin_lock_irq(&ctx->completion_lock);
5835 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
5836 spin_unlock_irq(&ctx->completion_lock);
5838 io_queue_async_work(req);
5839 return -EIOCBQUEUED;
5842 trace_io_uring_defer(ctx, req, req->user_data);
5845 list_add_tail(&de->list, &ctx->defer_list);
5846 spin_unlock_irq(&ctx->completion_lock);
5847 return -EIOCBQUEUED;
5850 static void __io_clean_op(struct io_kiocb *req)
5852 if (req->flags & REQ_F_BUFFER_SELECTED) {
5853 switch (req->opcode) {
5854 case IORING_OP_READV:
5855 case IORING_OP_READ_FIXED:
5856 case IORING_OP_READ:
5857 kfree((void *)(unsigned long)req->rw.addr);
5859 case IORING_OP_RECVMSG:
5860 case IORING_OP_RECV:
5861 kfree(req->sr_msg.kbuf);
5864 req->flags &= ~REQ_F_BUFFER_SELECTED;
5867 if (req->flags & REQ_F_NEED_CLEANUP) {
5868 switch (req->opcode) {
5869 case IORING_OP_READV:
5870 case IORING_OP_READ_FIXED:
5871 case IORING_OP_READ:
5872 case IORING_OP_WRITEV:
5873 case IORING_OP_WRITE_FIXED:
5874 case IORING_OP_WRITE: {
5875 struct io_async_rw *io = req->async_data;
5877 kfree(io->free_iovec);
5880 case IORING_OP_RECVMSG:
5881 case IORING_OP_SENDMSG: {
5882 struct io_async_msghdr *io = req->async_data;
5884 kfree(io->free_iov);
5887 case IORING_OP_SPLICE:
5889 io_put_file(req, req->splice.file_in,
5890 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
5892 case IORING_OP_OPENAT:
5893 case IORING_OP_OPENAT2:
5894 if (req->open.filename)
5895 putname(req->open.filename);
5897 case IORING_OP_RENAMEAT:
5898 putname(req->rename.oldpath);
5899 putname(req->rename.newpath);
5901 case IORING_OP_UNLINKAT:
5902 putname(req->unlink.filename);
5905 req->flags &= ~REQ_F_NEED_CLEANUP;
5909 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
5911 struct io_ring_ctx *ctx = req->ctx;
5912 const struct cred *creds = NULL;
5915 if (req->work.personality) {
5916 const struct cred *new_creds;
5918 if (!(issue_flags & IO_URING_F_NONBLOCK))
5919 mutex_lock(&ctx->uring_lock);
5920 new_creds = idr_find(&ctx->personality_idr, req->work.personality);
5921 if (!(issue_flags & IO_URING_F_NONBLOCK))
5922 mutex_unlock(&ctx->uring_lock);
5925 creds = override_creds(new_creds);
5928 switch (req->opcode) {
5930 ret = io_nop(req, issue_flags);
5932 case IORING_OP_READV:
5933 case IORING_OP_READ_FIXED:
5934 case IORING_OP_READ:
5935 ret = io_read(req, issue_flags);
5937 case IORING_OP_WRITEV:
5938 case IORING_OP_WRITE_FIXED:
5939 case IORING_OP_WRITE:
5940 ret = io_write(req, issue_flags);
5942 case IORING_OP_FSYNC:
5943 ret = io_fsync(req, issue_flags);
5945 case IORING_OP_POLL_ADD:
5946 ret = io_poll_add(req, issue_flags);
5948 case IORING_OP_POLL_REMOVE:
5949 ret = io_poll_remove(req, issue_flags);
5951 case IORING_OP_SYNC_FILE_RANGE:
5952 ret = io_sync_file_range(req, issue_flags);
5954 case IORING_OP_SENDMSG:
5955 ret = io_sendmsg(req, issue_flags);
5957 case IORING_OP_SEND:
5958 ret = io_send(req, issue_flags);
5960 case IORING_OP_RECVMSG:
5961 ret = io_recvmsg(req, issue_flags);
5963 case IORING_OP_RECV:
5964 ret = io_recv(req, issue_flags);
5966 case IORING_OP_TIMEOUT:
5967 ret = io_timeout(req, issue_flags);
5969 case IORING_OP_TIMEOUT_REMOVE:
5970 ret = io_timeout_remove(req, issue_flags);
5972 case IORING_OP_ACCEPT:
5973 ret = io_accept(req, issue_flags);
5975 case IORING_OP_CONNECT:
5976 ret = io_connect(req, issue_flags);
5978 case IORING_OP_ASYNC_CANCEL:
5979 ret = io_async_cancel(req, issue_flags);
5981 case IORING_OP_FALLOCATE:
5982 ret = io_fallocate(req, issue_flags);
5984 case IORING_OP_OPENAT:
5985 ret = io_openat(req, issue_flags);
5987 case IORING_OP_CLOSE:
5988 ret = io_close(req, issue_flags);
5990 case IORING_OP_FILES_UPDATE:
5991 ret = io_files_update(req, issue_flags);
5993 case IORING_OP_STATX:
5994 ret = io_statx(req, issue_flags);
5996 case IORING_OP_FADVISE:
5997 ret = io_fadvise(req, issue_flags);
5999 case IORING_OP_MADVISE:
6000 ret = io_madvise(req, issue_flags);
6002 case IORING_OP_OPENAT2:
6003 ret = io_openat2(req, issue_flags);
6005 case IORING_OP_EPOLL_CTL:
6006 ret = io_epoll_ctl(req, issue_flags);
6008 case IORING_OP_SPLICE:
6009 ret = io_splice(req, issue_flags);
6011 case IORING_OP_PROVIDE_BUFFERS:
6012 ret = io_provide_buffers(req, issue_flags);
6014 case IORING_OP_REMOVE_BUFFERS:
6015 ret = io_remove_buffers(req, issue_flags);
6018 ret = io_tee(req, issue_flags);
6020 case IORING_OP_SHUTDOWN:
6021 ret = io_shutdown(req, issue_flags);
6023 case IORING_OP_RENAMEAT:
6024 ret = io_renameat(req, issue_flags);
6026 case IORING_OP_UNLINKAT:
6027 ret = io_unlinkat(req, issue_flags);
6035 revert_creds(creds);
6040 /* If the op doesn't have a file, we're not polling for it */
6041 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
6042 const bool in_async = io_wq_current_is_worker();
6044 /* workqueue context doesn't hold uring_lock, grab it now */
6046 mutex_lock(&ctx->uring_lock);
6048 io_iopoll_req_issued(req, in_async);
6051 mutex_unlock(&ctx->uring_lock);
6057 static void io_wq_submit_work(struct io_wq_work *work)
6059 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6060 struct io_kiocb *timeout;
6063 timeout = io_prep_linked_timeout(req);
6065 io_queue_linked_timeout(timeout);
6067 if (work->flags & IO_WQ_WORK_CANCEL)
6072 ret = io_issue_sqe(req, 0);
6074 * We can get EAGAIN for polled IO even though we're
6075 * forcing a sync submission from here, since we can't
6076 * wait for request slots on the block side.
6084 /* avoid locking problems by failing it from a clean context */
6086 /* io-wq is going to take one down */
6087 refcount_inc(&req->refs);
6088 io_req_task_queue_fail(req, ret);
6092 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6095 struct fixed_rsrc_table *table;
6097 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
6098 return table->files[index & IORING_FILE_TABLE_MASK];
6101 static struct file *io_file_get(struct io_submit_state *state,
6102 struct io_kiocb *req, int fd, bool fixed)
6104 struct io_ring_ctx *ctx = req->ctx;
6108 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6110 fd = array_index_nospec(fd, ctx->nr_user_files);
6111 file = io_file_from_index(ctx, fd);
6112 io_set_resource_node(req);
6114 trace_io_uring_file_get(ctx, fd);
6115 file = __io_file_get(state, fd);
6118 if (file && unlikely(file->f_op == &io_uring_fops))
6119 io_req_track_inflight(req);
6123 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6125 struct io_timeout_data *data = container_of(timer,
6126 struct io_timeout_data, timer);
6127 struct io_kiocb *prev, *req = data->req;
6128 struct io_ring_ctx *ctx = req->ctx;
6129 unsigned long flags;
6131 spin_lock_irqsave(&ctx->completion_lock, flags);
6132 prev = req->timeout.head;
6133 req->timeout.head = NULL;
6136 * We don't expect the list to be empty, that will only happen if we
6137 * race with the completion of the linked work.
6139 if (prev && refcount_inc_not_zero(&prev->refs))
6140 io_remove_next_linked(prev);
6143 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6146 req_set_fail_links(prev);
6147 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6148 io_put_req_deferred(prev, 1);
6150 io_req_complete_post(req, -ETIME, 0);
6151 io_put_req_deferred(req, 1);
6153 return HRTIMER_NORESTART;
6156 static void __io_queue_linked_timeout(struct io_kiocb *req)
6159 * If the back reference is NULL, then our linked request finished
6160 * before we got a chance to setup the timer
6162 if (req->timeout.head) {
6163 struct io_timeout_data *data = req->async_data;
6165 data->timer.function = io_link_timeout_fn;
6166 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6171 static void io_queue_linked_timeout(struct io_kiocb *req)
6173 struct io_ring_ctx *ctx = req->ctx;
6175 spin_lock_irq(&ctx->completion_lock);
6176 __io_queue_linked_timeout(req);
6177 spin_unlock_irq(&ctx->completion_lock);
6179 /* drop submission reference */
6183 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6185 struct io_kiocb *nxt = req->link;
6187 if (!nxt || (req->flags & REQ_F_LINK_TIMEOUT) ||
6188 nxt->opcode != IORING_OP_LINK_TIMEOUT)
6191 nxt->timeout.head = req;
6192 nxt->flags |= REQ_F_LTIMEOUT_ACTIVE;
6193 req->flags |= REQ_F_LINK_TIMEOUT;
6197 static void __io_queue_sqe(struct io_kiocb *req)
6199 struct io_kiocb *linked_timeout = io_prep_linked_timeout(req);
6202 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
6205 * We async punt it if the file wasn't marked NOWAIT, or if the file
6206 * doesn't support non-blocking read/write attempts
6208 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6209 if (!io_arm_poll_handler(req)) {
6211 * Queued up for async execution, worker will release
6212 * submit reference when the iocb is actually submitted.
6214 io_queue_async_work(req);
6216 } else if (likely(!ret)) {
6217 /* drop submission reference */
6218 if (req->flags & REQ_F_COMPLETE_INLINE) {
6219 struct io_ring_ctx *ctx = req->ctx;
6220 struct io_comp_state *cs = &ctx->submit_state.comp;
6222 cs->reqs[cs->nr++] = req;
6223 if (cs->nr == ARRAY_SIZE(cs->reqs))
6224 io_submit_flush_completions(cs, ctx);
6229 req_set_fail_links(req);
6231 io_req_complete(req, ret);
6234 io_queue_linked_timeout(linked_timeout);
6237 static void io_queue_sqe(struct io_kiocb *req)
6241 ret = io_req_defer(req);
6243 if (ret != -EIOCBQUEUED) {
6245 req_set_fail_links(req);
6247 io_req_complete(req, ret);
6249 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6250 ret = io_req_defer_prep(req);
6253 io_queue_async_work(req);
6255 __io_queue_sqe(req);
6260 * Check SQE restrictions (opcode and flags).
6262 * Returns 'true' if SQE is allowed, 'false' otherwise.
6264 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6265 struct io_kiocb *req,
6266 unsigned int sqe_flags)
6268 if (!ctx->restricted)
6271 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6274 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6275 ctx->restrictions.sqe_flags_required)
6278 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6279 ctx->restrictions.sqe_flags_required))
6285 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6286 const struct io_uring_sqe *sqe)
6288 struct io_submit_state *state;
6289 unsigned int sqe_flags;
6292 req->opcode = READ_ONCE(sqe->opcode);
6293 /* same numerical values with corresponding REQ_F_*, safe to copy */
6294 req->flags = sqe_flags = READ_ONCE(sqe->flags);
6295 req->user_data = READ_ONCE(sqe->user_data);
6296 req->async_data = NULL;
6300 req->fixed_rsrc_refs = NULL;
6301 /* one is dropped after submission, the other at completion */
6302 refcount_set(&req->refs, 2);
6303 req->task = current;
6306 /* enforce forwards compatibility on users */
6307 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS)) {
6312 if (unlikely(req->opcode >= IORING_OP_LAST))
6315 if (unlikely(!io_check_restriction(ctx, req, sqe_flags)))
6318 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6319 !io_op_defs[req->opcode].buffer_select)
6322 req->work.list.next = NULL;
6323 req->work.flags = 0;
6324 req->work.personality = READ_ONCE(sqe->personality);
6325 state = &ctx->submit_state;
6328 * Plug now if we have more than 1 IO left after this, and the target
6329 * is potentially a read/write to block based storage.
6331 if (!state->plug_started && state->ios_left > 1 &&
6332 io_op_defs[req->opcode].plug) {
6333 blk_start_plug(&state->plug);
6334 state->plug_started = true;
6337 if (io_op_defs[req->opcode].needs_file) {
6338 bool fixed = req->flags & REQ_F_FIXED_FILE;
6340 req->file = io_file_get(state, req, READ_ONCE(sqe->fd), fixed);
6341 if (unlikely(!req->file))
6349 static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
6350 const struct io_uring_sqe *sqe)
6352 struct io_submit_link *link = &ctx->submit_state.link;
6355 ret = io_init_req(ctx, req, sqe);
6356 if (unlikely(ret)) {
6359 io_req_complete(req, ret);
6361 /* fail even hard links since we don't submit */
6362 link->head->flags |= REQ_F_FAIL_LINK;
6363 io_put_req(link->head);
6364 io_req_complete(link->head, -ECANCELED);
6369 ret = io_req_prep(req, sqe);
6373 /* don't need @sqe from now on */
6374 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6375 true, ctx->flags & IORING_SETUP_SQPOLL);
6378 * If we already have a head request, queue this one for async
6379 * submittal once the head completes. If we don't have a head but
6380 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6381 * submitted sync once the chain is complete. If none of those
6382 * conditions are true (normal request), then just queue it.
6385 struct io_kiocb *head = link->head;
6388 * Taking sequential execution of a link, draining both sides
6389 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6390 * requests in the link. So, it drains the head and the
6391 * next after the link request. The last one is done via
6392 * drain_next flag to persist the effect across calls.
6394 if (req->flags & REQ_F_IO_DRAIN) {
6395 head->flags |= REQ_F_IO_DRAIN;
6396 ctx->drain_next = 1;
6398 ret = io_req_defer_prep(req);
6401 trace_io_uring_link(ctx, req, head);
6402 link->last->link = req;
6405 /* last request of a link, enqueue the link */
6406 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6411 if (unlikely(ctx->drain_next)) {
6412 req->flags |= REQ_F_IO_DRAIN;
6413 ctx->drain_next = 0;
6415 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6427 * Batched submission is done, ensure local IO is flushed out.
6429 static void io_submit_state_end(struct io_submit_state *state,
6430 struct io_ring_ctx *ctx)
6432 if (state->link.head)
6433 io_queue_sqe(state->link.head);
6435 io_submit_flush_completions(&state->comp, ctx);
6436 if (state->plug_started)
6437 blk_finish_plug(&state->plug);
6438 io_state_file_put(state);
6442 * Start submission side cache.
6444 static void io_submit_state_start(struct io_submit_state *state,
6445 unsigned int max_ios)
6447 state->plug_started = false;
6448 state->ios_left = max_ios;
6449 /* set only head, no need to init link_last in advance */
6450 state->link.head = NULL;
6453 static void io_commit_sqring(struct io_ring_ctx *ctx)
6455 struct io_rings *rings = ctx->rings;
6458 * Ensure any loads from the SQEs are done at this point,
6459 * since once we write the new head, the application could
6460 * write new data to them.
6462 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6466 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6467 * that is mapped by userspace. This means that care needs to be taken to
6468 * ensure that reads are stable, as we cannot rely on userspace always
6469 * being a good citizen. If members of the sqe are validated and then later
6470 * used, it's important that those reads are done through READ_ONCE() to
6471 * prevent a re-load down the line.
6473 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6475 u32 *sq_array = ctx->sq_array;
6479 * The cached sq head (or cq tail) serves two purposes:
6481 * 1) allows us to batch the cost of updating the user visible
6483 * 2) allows the kernel side to track the head on its own, even
6484 * though the application is the one updating it.
6486 head = READ_ONCE(sq_array[ctx->cached_sq_head++ & ctx->sq_mask]);
6487 if (likely(head < ctx->sq_entries))
6488 return &ctx->sq_sqes[head];
6490 /* drop invalid entries */
6491 ctx->cached_sq_dropped++;
6492 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6496 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6500 /* if we have a backlog and couldn't flush it all, return BUSY */
6501 if (test_bit(0, &ctx->sq_check_overflow)) {
6502 if (!__io_cqring_overflow_flush(ctx, false, NULL, NULL))
6506 /* make sure SQ entry isn't read before tail */
6507 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6509 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6512 percpu_counter_add(¤t->io_uring->inflight, nr);
6513 refcount_add(nr, ¤t->usage);
6514 io_submit_state_start(&ctx->submit_state, nr);
6516 while (submitted < nr) {
6517 const struct io_uring_sqe *sqe;
6518 struct io_kiocb *req;
6520 req = io_alloc_req(ctx);
6521 if (unlikely(!req)) {
6523 submitted = -EAGAIN;
6526 sqe = io_get_sqe(ctx);
6527 if (unlikely(!sqe)) {
6528 kmem_cache_free(req_cachep, req);
6531 /* will complete beyond this point, count as submitted */
6533 if (io_submit_sqe(ctx, req, sqe))
6537 if (unlikely(submitted != nr)) {
6538 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6539 struct io_uring_task *tctx = current->io_uring;
6540 int unused = nr - ref_used;
6542 percpu_ref_put_many(&ctx->refs, unused);
6543 percpu_counter_sub(&tctx->inflight, unused);
6544 put_task_struct_many(current, unused);
6547 io_submit_state_end(&ctx->submit_state, ctx);
6548 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6549 io_commit_sqring(ctx);
6554 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6556 /* Tell userspace we may need a wakeup call */
6557 spin_lock_irq(&ctx->completion_lock);
6558 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6559 spin_unlock_irq(&ctx->completion_lock);
6562 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6564 spin_lock_irq(&ctx->completion_lock);
6565 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6566 spin_unlock_irq(&ctx->completion_lock);
6569 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
6571 unsigned int to_submit;
6574 to_submit = io_sqring_entries(ctx);
6575 /* if we're handling multiple rings, cap submit size for fairness */
6576 if (cap_entries && to_submit > 8)
6579 if (!list_empty(&ctx->iopoll_list) || to_submit) {
6580 unsigned nr_events = 0;
6582 mutex_lock(&ctx->uring_lock);
6583 if (!list_empty(&ctx->iopoll_list))
6584 io_do_iopoll(ctx, &nr_events, 0);
6586 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)))
6587 ret = io_submit_sqes(ctx, to_submit);
6588 mutex_unlock(&ctx->uring_lock);
6591 if (!io_sqring_full(ctx) && wq_has_sleeper(&ctx->sqo_sq_wait))
6592 wake_up(&ctx->sqo_sq_wait);
6597 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
6599 struct io_ring_ctx *ctx;
6600 unsigned sq_thread_idle = 0;
6602 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6603 if (sq_thread_idle < ctx->sq_thread_idle)
6604 sq_thread_idle = ctx->sq_thread_idle;
6607 sqd->sq_thread_idle = sq_thread_idle;
6610 static void io_sqd_init_new(struct io_sq_data *sqd)
6612 struct io_ring_ctx *ctx;
6614 while (!list_empty(&sqd->ctx_new_list)) {
6615 ctx = list_first_entry(&sqd->ctx_new_list, struct io_ring_ctx, sqd_list);
6616 list_move_tail(&ctx->sqd_list, &sqd->ctx_list);
6617 complete(&ctx->sq_thread_comp);
6620 io_sqd_update_thread_idle(sqd);
6623 static bool io_sq_thread_should_stop(struct io_sq_data *sqd)
6625 return test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
6628 static bool io_sq_thread_should_park(struct io_sq_data *sqd)
6630 return test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
6633 static void io_sq_thread_parkme(struct io_sq_data *sqd)
6637 * TASK_PARKED is a special state; we must serialize against
6638 * possible pending wakeups to avoid store-store collisions on
6641 * Such a collision might possibly result in the task state
6642 * changin from TASK_PARKED and us failing the
6643 * wait_task_inactive() in kthread_park().
6645 set_special_state(TASK_PARKED);
6646 if (!test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state))
6650 * Thread is going to call schedule(), do not preempt it,
6651 * or the caller of kthread_park() may spend more time in
6652 * wait_task_inactive().
6655 complete(&sqd->completion);
6656 schedule_preempt_disabled();
6659 __set_current_state(TASK_RUNNING);
6662 static int io_sq_thread(void *data)
6664 struct io_sq_data *sqd = data;
6665 struct io_ring_ctx *ctx;
6666 unsigned long timeout = 0;
6667 char buf[TASK_COMM_LEN];
6670 sprintf(buf, "iou-sqp-%d", sqd->task_pid);
6671 set_task_comm(current, buf);
6672 sqd->thread = current;
6673 current->pf_io_worker = NULL;
6675 if (sqd->sq_cpu != -1)
6676 set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
6678 set_cpus_allowed_ptr(current, cpu_online_mask);
6679 current->flags |= PF_NO_SETAFFINITY;
6681 complete(&sqd->completion);
6683 wait_for_completion(&sqd->startup);
6685 while (!io_sq_thread_should_stop(sqd)) {
6687 bool cap_entries, sqt_spin, needs_sched;
6690 * Any changes to the sqd lists are synchronized through the
6691 * thread parking. This synchronizes the thread vs users,
6692 * the users are synchronized on the sqd->ctx_lock.
6694 if (io_sq_thread_should_park(sqd)) {
6695 io_sq_thread_parkme(sqd);
6698 if (unlikely(!list_empty(&sqd->ctx_new_list))) {
6699 io_sqd_init_new(sqd);
6700 timeout = jiffies + sqd->sq_thread_idle;
6702 if (fatal_signal_pending(current))
6705 cap_entries = !list_is_singular(&sqd->ctx_list);
6706 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6707 ret = __io_sq_thread(ctx, cap_entries);
6708 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
6712 if (sqt_spin || !time_after(jiffies, timeout)) {
6716 timeout = jiffies + sqd->sq_thread_idle;
6721 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
6722 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6723 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6724 !list_empty_careful(&ctx->iopoll_list)) {
6725 needs_sched = false;
6728 if (io_sqring_entries(ctx)) {
6729 needs_sched = false;
6734 if (needs_sched && !io_sq_thread_should_park(sqd)) {
6735 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6736 io_ring_set_wakeup_flag(ctx);
6739 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6740 io_ring_clear_wakeup_flag(ctx);
6743 finish_wait(&sqd->wait, &wait);
6744 timeout = jiffies + sqd->sq_thread_idle;
6747 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6748 io_uring_cancel_sqpoll(ctx);
6752 if (io_sq_thread_should_park(sqd))
6753 io_sq_thread_parkme(sqd);
6756 * Clear thread under lock so that concurrent parks work correctly
6758 complete(&sqd->completion);
6759 mutex_lock(&sqd->lock);
6761 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6763 io_ring_set_wakeup_flag(ctx);
6766 complete(&sqd->exited);
6767 mutex_unlock(&sqd->lock);
6771 struct io_wait_queue {
6772 struct wait_queue_entry wq;
6773 struct io_ring_ctx *ctx;
6775 unsigned nr_timeouts;
6778 static inline bool io_should_wake(struct io_wait_queue *iowq)
6780 struct io_ring_ctx *ctx = iowq->ctx;
6783 * Wake up if we have enough events, or if a timeout occurred since we
6784 * started waiting. For timeouts, we always want to return to userspace,
6785 * regardless of event count.
6787 return io_cqring_events(ctx) >= iowq->to_wait ||
6788 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6791 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6792 int wake_flags, void *key)
6794 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6798 * Cannot safely flush overflowed CQEs from here, ensure we wake up
6799 * the task, and the next invocation will do it.
6801 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->cq_check_overflow))
6802 return autoremove_wake_function(curr, mode, wake_flags, key);
6806 static int io_run_task_work_sig(void)
6808 if (io_run_task_work())
6810 if (!signal_pending(current))
6812 if (test_tsk_thread_flag(current, TIF_NOTIFY_SIGNAL))
6813 return -ERESTARTSYS;
6817 /* when returns >0, the caller should retry */
6818 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
6819 struct io_wait_queue *iowq,
6820 signed long *timeout)
6824 /* make sure we run task_work before checking for signals */
6825 ret = io_run_task_work_sig();
6826 if (ret || io_should_wake(iowq))
6828 /* let the caller flush overflows, retry */
6829 if (test_bit(0, &ctx->cq_check_overflow))
6832 *timeout = schedule_timeout(*timeout);
6833 return !*timeout ? -ETIME : 1;
6837 * Wait until events become available, if we don't already have some. The
6838 * application must reap them itself, as they reside on the shared cq ring.
6840 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6841 const sigset_t __user *sig, size_t sigsz,
6842 struct __kernel_timespec __user *uts)
6844 struct io_wait_queue iowq = {
6847 .func = io_wake_function,
6848 .entry = LIST_HEAD_INIT(iowq.wq.entry),
6851 .to_wait = min_events,
6853 struct io_rings *rings = ctx->rings;
6854 signed long timeout = MAX_SCHEDULE_TIMEOUT;
6858 io_cqring_overflow_flush(ctx, false, NULL, NULL);
6859 if (io_cqring_events(ctx) >= min_events)
6861 if (!io_run_task_work())
6866 #ifdef CONFIG_COMPAT
6867 if (in_compat_syscall())
6868 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
6872 ret = set_user_sigmask(sig, sigsz);
6879 struct timespec64 ts;
6881 if (get_timespec64(&ts, uts))
6883 timeout = timespec64_to_jiffies(&ts);
6886 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
6887 trace_io_uring_cqring_wait(ctx, min_events);
6889 io_cqring_overflow_flush(ctx, false, NULL, NULL);
6890 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
6891 TASK_INTERRUPTIBLE);
6892 ret = io_cqring_wait_schedule(ctx, &iowq, &timeout);
6893 finish_wait(&ctx->wait, &iowq.wq);
6896 restore_saved_sigmask_unless(ret == -EINTR);
6898 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
6901 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
6903 #if defined(CONFIG_UNIX)
6904 if (ctx->ring_sock) {
6905 struct sock *sock = ctx->ring_sock->sk;
6906 struct sk_buff *skb;
6908 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
6914 for (i = 0; i < ctx->nr_user_files; i++) {
6917 file = io_file_from_index(ctx, i);
6924 static void io_rsrc_data_ref_zero(struct percpu_ref *ref)
6926 struct fixed_rsrc_data *data;
6928 data = container_of(ref, struct fixed_rsrc_data, refs);
6929 complete(&data->done);
6932 static inline void io_rsrc_ref_lock(struct io_ring_ctx *ctx)
6934 spin_lock_bh(&ctx->rsrc_ref_lock);
6937 static inline void io_rsrc_ref_unlock(struct io_ring_ctx *ctx)
6939 spin_unlock_bh(&ctx->rsrc_ref_lock);
6942 static void io_sqe_rsrc_set_node(struct io_ring_ctx *ctx,
6943 struct fixed_rsrc_data *rsrc_data,
6944 struct fixed_rsrc_ref_node *ref_node)
6946 io_rsrc_ref_lock(ctx);
6947 rsrc_data->node = ref_node;
6948 list_add_tail(&ref_node->node, &ctx->rsrc_ref_list);
6949 io_rsrc_ref_unlock(ctx);
6950 percpu_ref_get(&rsrc_data->refs);
6953 static void io_sqe_rsrc_kill_node(struct io_ring_ctx *ctx, struct fixed_rsrc_data *data)
6955 struct fixed_rsrc_ref_node *ref_node = NULL;
6957 io_rsrc_ref_lock(ctx);
6958 ref_node = data->node;
6960 io_rsrc_ref_unlock(ctx);
6962 percpu_ref_kill(&ref_node->refs);
6965 static int io_rsrc_ref_quiesce(struct fixed_rsrc_data *data,
6966 struct io_ring_ctx *ctx,
6967 void (*rsrc_put)(struct io_ring_ctx *ctx,
6968 struct io_rsrc_put *prsrc))
6970 struct fixed_rsrc_ref_node *backup_node;
6976 data->quiesce = true;
6979 backup_node = alloc_fixed_rsrc_ref_node(ctx);
6982 backup_node->rsrc_data = data;
6983 backup_node->rsrc_put = rsrc_put;
6985 io_sqe_rsrc_kill_node(ctx, data);
6986 percpu_ref_kill(&data->refs);
6987 flush_delayed_work(&ctx->rsrc_put_work);
6989 ret = wait_for_completion_interruptible(&data->done);
6993 percpu_ref_resurrect(&data->refs);
6994 io_sqe_rsrc_set_node(ctx, data, backup_node);
6996 reinit_completion(&data->done);
6997 mutex_unlock(&ctx->uring_lock);
6998 ret = io_run_task_work_sig();
6999 mutex_lock(&ctx->uring_lock);
7001 data->quiesce = false;
7004 destroy_fixed_rsrc_ref_node(backup_node);
7008 static struct fixed_rsrc_data *alloc_fixed_rsrc_data(struct io_ring_ctx *ctx)
7010 struct fixed_rsrc_data *data;
7012 data = kzalloc(sizeof(*data), GFP_KERNEL);
7016 if (percpu_ref_init(&data->refs, io_rsrc_data_ref_zero,
7017 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) {
7022 init_completion(&data->done);
7026 static void free_fixed_rsrc_data(struct fixed_rsrc_data *data)
7028 percpu_ref_exit(&data->refs);
7033 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7035 struct fixed_rsrc_data *data = ctx->file_data;
7036 unsigned nr_tables, i;
7040 * percpu_ref_is_dying() is to stop parallel files unregister
7041 * Since we possibly drop uring lock later in this function to
7044 if (!data || percpu_ref_is_dying(&data->refs))
7046 ret = io_rsrc_ref_quiesce(data, ctx, io_ring_file_put);
7050 __io_sqe_files_unregister(ctx);
7051 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
7052 for (i = 0; i < nr_tables; i++)
7053 kfree(data->table[i].files);
7054 free_fixed_rsrc_data(data);
7055 ctx->file_data = NULL;
7056 ctx->nr_user_files = 0;
7060 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7061 __releases(&sqd->lock)
7065 if (sqd->thread == current)
7067 clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7068 wake_up_state(sqd->thread, TASK_PARKED);
7069 mutex_unlock(&sqd->lock);
7072 static bool io_sq_thread_park(struct io_sq_data *sqd)
7073 __acquires(&sqd->lock)
7075 if (sqd->thread == current)
7077 mutex_lock(&sqd->lock);
7079 mutex_unlock(&sqd->lock);
7082 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7083 wake_up_process(sqd->thread);
7084 wait_for_completion(&sqd->completion);
7088 static void io_sq_thread_stop(struct io_sq_data *sqd)
7090 if (test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state))
7092 mutex_lock(&sqd->lock);
7094 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7095 WARN_ON_ONCE(test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state));
7096 wake_up_process(sqd->thread);
7097 mutex_unlock(&sqd->lock);
7098 wait_for_completion(&sqd->exited);
7099 WARN_ON_ONCE(sqd->thread);
7101 mutex_unlock(&sqd->lock);
7105 static void io_put_sq_data(struct io_sq_data *sqd)
7107 if (refcount_dec_and_test(&sqd->refs)) {
7108 io_sq_thread_stop(sqd);
7113 static void io_sq_thread_finish(struct io_ring_ctx *ctx)
7115 struct io_sq_data *sqd = ctx->sq_data;
7118 complete(&sqd->startup);
7120 wait_for_completion(&ctx->sq_thread_comp);
7121 io_sq_thread_park(sqd);
7124 mutex_lock(&sqd->ctx_lock);
7125 list_del(&ctx->sqd_list);
7126 io_sqd_update_thread_idle(sqd);
7127 mutex_unlock(&sqd->ctx_lock);
7130 io_sq_thread_unpark(sqd);
7132 io_put_sq_data(sqd);
7133 ctx->sq_data = NULL;
7137 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7139 struct io_ring_ctx *ctx_attach;
7140 struct io_sq_data *sqd;
7143 f = fdget(p->wq_fd);
7145 return ERR_PTR(-ENXIO);
7146 if (f.file->f_op != &io_uring_fops) {
7148 return ERR_PTR(-EINVAL);
7151 ctx_attach = f.file->private_data;
7152 sqd = ctx_attach->sq_data;
7155 return ERR_PTR(-EINVAL);
7158 refcount_inc(&sqd->refs);
7163 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p)
7165 struct io_sq_data *sqd;
7167 if (p->flags & IORING_SETUP_ATTACH_WQ)
7168 return io_attach_sq_data(p);
7170 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7172 return ERR_PTR(-ENOMEM);
7174 refcount_set(&sqd->refs, 1);
7175 INIT_LIST_HEAD(&sqd->ctx_list);
7176 INIT_LIST_HEAD(&sqd->ctx_new_list);
7177 mutex_init(&sqd->ctx_lock);
7178 mutex_init(&sqd->lock);
7179 init_waitqueue_head(&sqd->wait);
7180 init_completion(&sqd->startup);
7181 init_completion(&sqd->completion);
7182 init_completion(&sqd->exited);
7186 #if defined(CONFIG_UNIX)
7188 * Ensure the UNIX gc is aware of our file set, so we are certain that
7189 * the io_uring can be safely unregistered on process exit, even if we have
7190 * loops in the file referencing.
7192 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7194 struct sock *sk = ctx->ring_sock->sk;
7195 struct scm_fp_list *fpl;
7196 struct sk_buff *skb;
7199 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7203 skb = alloc_skb(0, GFP_KERNEL);
7212 fpl->user = get_uid(current_user());
7213 for (i = 0; i < nr; i++) {
7214 struct file *file = io_file_from_index(ctx, i + offset);
7218 fpl->fp[nr_files] = get_file(file);
7219 unix_inflight(fpl->user, fpl->fp[nr_files]);
7224 fpl->max = SCM_MAX_FD;
7225 fpl->count = nr_files;
7226 UNIXCB(skb).fp = fpl;
7227 skb->destructor = unix_destruct_scm;
7228 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7229 skb_queue_head(&sk->sk_receive_queue, skb);
7231 for (i = 0; i < nr_files; i++)
7242 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7243 * causes regular reference counting to break down. We rely on the UNIX
7244 * garbage collection to take care of this problem for us.
7246 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7248 unsigned left, total;
7252 left = ctx->nr_user_files;
7254 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7256 ret = __io_sqe_files_scm(ctx, this_files, total);
7260 total += this_files;
7266 while (total < ctx->nr_user_files) {
7267 struct file *file = io_file_from_index(ctx, total);
7277 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7283 static int io_sqe_alloc_file_tables(struct fixed_rsrc_data *file_data,
7284 unsigned nr_tables, unsigned nr_files)
7288 for (i = 0; i < nr_tables; i++) {
7289 struct fixed_rsrc_table *table = &file_data->table[i];
7290 unsigned this_files;
7292 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
7293 table->files = kcalloc(this_files, sizeof(struct file *),
7297 nr_files -= this_files;
7303 for (i = 0; i < nr_tables; i++) {
7304 struct fixed_rsrc_table *table = &file_data->table[i];
7305 kfree(table->files);
7310 static void io_ring_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
7312 struct file *file = prsrc->file;
7313 #if defined(CONFIG_UNIX)
7314 struct sock *sock = ctx->ring_sock->sk;
7315 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7316 struct sk_buff *skb;
7319 __skb_queue_head_init(&list);
7322 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7323 * remove this entry and rearrange the file array.
7325 skb = skb_dequeue(head);
7327 struct scm_fp_list *fp;
7329 fp = UNIXCB(skb).fp;
7330 for (i = 0; i < fp->count; i++) {
7333 if (fp->fp[i] != file)
7336 unix_notinflight(fp->user, fp->fp[i]);
7337 left = fp->count - 1 - i;
7339 memmove(&fp->fp[i], &fp->fp[i + 1],
7340 left * sizeof(struct file *));
7347 __skb_queue_tail(&list, skb);
7357 __skb_queue_tail(&list, skb);
7359 skb = skb_dequeue(head);
7362 if (skb_peek(&list)) {
7363 spin_lock_irq(&head->lock);
7364 while ((skb = __skb_dequeue(&list)) != NULL)
7365 __skb_queue_tail(head, skb);
7366 spin_unlock_irq(&head->lock);
7373 static void __io_rsrc_put_work(struct fixed_rsrc_ref_node *ref_node)
7375 struct fixed_rsrc_data *rsrc_data = ref_node->rsrc_data;
7376 struct io_ring_ctx *ctx = rsrc_data->ctx;
7377 struct io_rsrc_put *prsrc, *tmp;
7379 list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
7380 list_del(&prsrc->list);
7381 ref_node->rsrc_put(ctx, prsrc);
7385 percpu_ref_exit(&ref_node->refs);
7387 percpu_ref_put(&rsrc_data->refs);
7390 static void io_rsrc_put_work(struct work_struct *work)
7392 struct io_ring_ctx *ctx;
7393 struct llist_node *node;
7395 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
7396 node = llist_del_all(&ctx->rsrc_put_llist);
7399 struct fixed_rsrc_ref_node *ref_node;
7400 struct llist_node *next = node->next;
7402 ref_node = llist_entry(node, struct fixed_rsrc_ref_node, llist);
7403 __io_rsrc_put_work(ref_node);
7408 static struct file **io_fixed_file_slot(struct fixed_rsrc_data *file_data,
7411 struct fixed_rsrc_table *table;
7413 table = &file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7414 return &table->files[i & IORING_FILE_TABLE_MASK];
7417 static void io_rsrc_node_ref_zero(struct percpu_ref *ref)
7419 struct fixed_rsrc_ref_node *ref_node;
7420 struct fixed_rsrc_data *data;
7421 struct io_ring_ctx *ctx;
7422 bool first_add = false;
7425 ref_node = container_of(ref, struct fixed_rsrc_ref_node, refs);
7426 data = ref_node->rsrc_data;
7429 io_rsrc_ref_lock(ctx);
7430 ref_node->done = true;
7432 while (!list_empty(&ctx->rsrc_ref_list)) {
7433 ref_node = list_first_entry(&ctx->rsrc_ref_list,
7434 struct fixed_rsrc_ref_node, node);
7435 /* recycle ref nodes in order */
7436 if (!ref_node->done)
7438 list_del(&ref_node->node);
7439 first_add |= llist_add(&ref_node->llist, &ctx->rsrc_put_llist);
7441 io_rsrc_ref_unlock(ctx);
7443 if (percpu_ref_is_dying(&data->refs))
7447 mod_delayed_work(system_wq, &ctx->rsrc_put_work, 0);
7449 queue_delayed_work(system_wq, &ctx->rsrc_put_work, delay);
7452 static struct fixed_rsrc_ref_node *alloc_fixed_rsrc_ref_node(
7453 struct io_ring_ctx *ctx)
7455 struct fixed_rsrc_ref_node *ref_node;
7457 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7461 if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
7466 INIT_LIST_HEAD(&ref_node->node);
7467 INIT_LIST_HEAD(&ref_node->rsrc_list);
7468 ref_node->done = false;
7472 static void init_fixed_file_ref_node(struct io_ring_ctx *ctx,
7473 struct fixed_rsrc_ref_node *ref_node)
7475 ref_node->rsrc_data = ctx->file_data;
7476 ref_node->rsrc_put = io_ring_file_put;
7479 static void destroy_fixed_rsrc_ref_node(struct fixed_rsrc_ref_node *ref_node)
7481 percpu_ref_exit(&ref_node->refs);
7486 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7489 __s32 __user *fds = (__s32 __user *) arg;
7490 unsigned nr_tables, i;
7492 int fd, ret = -ENOMEM;
7493 struct fixed_rsrc_ref_node *ref_node;
7494 struct fixed_rsrc_data *file_data;
7500 if (nr_args > IORING_MAX_FIXED_FILES)
7503 file_data = alloc_fixed_rsrc_data(ctx);
7506 ctx->file_data = file_data;
7508 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
7509 file_data->table = kcalloc(nr_tables, sizeof(*file_data->table),
7511 if (!file_data->table)
7514 if (io_sqe_alloc_file_tables(file_data, nr_tables, nr_args))
7517 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7518 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7522 /* allow sparse sets */
7532 * Don't allow io_uring instances to be registered. If UNIX
7533 * isn't enabled, then this causes a reference cycle and this
7534 * instance can never get freed. If UNIX is enabled we'll
7535 * handle it just fine, but there's still no point in allowing
7536 * a ring fd as it doesn't support regular read/write anyway.
7538 if (file->f_op == &io_uring_fops) {
7542 *io_fixed_file_slot(file_data, i) = file;
7545 ret = io_sqe_files_scm(ctx);
7547 io_sqe_files_unregister(ctx);
7551 ref_node = alloc_fixed_rsrc_ref_node(ctx);
7553 io_sqe_files_unregister(ctx);
7556 init_fixed_file_ref_node(ctx, ref_node);
7558 io_sqe_rsrc_set_node(ctx, file_data, ref_node);
7561 for (i = 0; i < ctx->nr_user_files; i++) {
7562 file = io_file_from_index(ctx, i);
7566 for (i = 0; i < nr_tables; i++)
7567 kfree(file_data->table[i].files);
7568 ctx->nr_user_files = 0;
7570 free_fixed_rsrc_data(ctx->file_data);
7571 ctx->file_data = NULL;
7575 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7578 #if defined(CONFIG_UNIX)
7579 struct sock *sock = ctx->ring_sock->sk;
7580 struct sk_buff_head *head = &sock->sk_receive_queue;
7581 struct sk_buff *skb;
7584 * See if we can merge this file into an existing skb SCM_RIGHTS
7585 * file set. If there's no room, fall back to allocating a new skb
7586 * and filling it in.
7588 spin_lock_irq(&head->lock);
7589 skb = skb_peek(head);
7591 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7593 if (fpl->count < SCM_MAX_FD) {
7594 __skb_unlink(skb, head);
7595 spin_unlock_irq(&head->lock);
7596 fpl->fp[fpl->count] = get_file(file);
7597 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7599 spin_lock_irq(&head->lock);
7600 __skb_queue_head(head, skb);
7605 spin_unlock_irq(&head->lock);
7612 return __io_sqe_files_scm(ctx, 1, index);
7618 static int io_queue_rsrc_removal(struct fixed_rsrc_data *data, void *rsrc)
7620 struct io_rsrc_put *prsrc;
7621 struct fixed_rsrc_ref_node *ref_node = data->node;
7623 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
7628 list_add(&prsrc->list, &ref_node->rsrc_list);
7633 static inline int io_queue_file_removal(struct fixed_rsrc_data *data,
7636 return io_queue_rsrc_removal(data, (void *)file);
7639 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7640 struct io_uring_rsrc_update *up,
7643 struct fixed_rsrc_data *data = ctx->file_data;
7644 struct fixed_rsrc_ref_node *ref_node;
7645 struct file *file, **file_slot;
7649 bool needs_switch = false;
7651 if (check_add_overflow(up->offset, nr_args, &done))
7653 if (done > ctx->nr_user_files)
7656 ref_node = alloc_fixed_rsrc_ref_node(ctx);
7659 init_fixed_file_ref_node(ctx, ref_node);
7661 fds = u64_to_user_ptr(up->data);
7662 for (done = 0; done < nr_args; done++) {
7664 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
7668 if (fd == IORING_REGISTER_FILES_SKIP)
7671 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
7672 file_slot = io_fixed_file_slot(ctx->file_data, i);
7675 err = io_queue_file_removal(data, *file_slot);
7679 needs_switch = true;
7688 * Don't allow io_uring instances to be registered. If
7689 * UNIX isn't enabled, then this causes a reference
7690 * cycle and this instance can never get freed. If UNIX
7691 * is enabled we'll handle it just fine, but there's
7692 * still no point in allowing a ring fd as it doesn't
7693 * support regular read/write anyway.
7695 if (file->f_op == &io_uring_fops) {
7701 err = io_sqe_file_register(ctx, file, i);
7711 percpu_ref_kill(&data->node->refs);
7712 io_sqe_rsrc_set_node(ctx, data, ref_node);
7714 destroy_fixed_rsrc_ref_node(ref_node);
7716 return done ? done : err;
7719 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
7722 struct io_uring_rsrc_update up;
7724 if (!ctx->file_data)
7728 if (copy_from_user(&up, arg, sizeof(up)))
7733 return __io_sqe_files_update(ctx, &up, nr_args);
7736 static struct io_wq_work *io_free_work(struct io_wq_work *work)
7738 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7740 req = io_put_req_find_next(req);
7741 return req ? &req->work : NULL;
7744 static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx)
7746 struct io_wq_hash *hash;
7747 struct io_wq_data data;
7748 unsigned int concurrency;
7750 hash = ctx->hash_map;
7752 hash = kzalloc(sizeof(*hash), GFP_KERNEL);
7754 return ERR_PTR(-ENOMEM);
7755 refcount_set(&hash->refs, 1);
7756 init_waitqueue_head(&hash->wait);
7757 ctx->hash_map = hash;
7761 data.free_work = io_free_work;
7762 data.do_work = io_wq_submit_work;
7764 /* Do QD, or 4 * CPUS, whatever is smallest */
7765 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7767 return io_wq_create(concurrency, &data);
7770 static int io_uring_alloc_task_context(struct task_struct *task,
7771 struct io_ring_ctx *ctx)
7773 struct io_uring_task *tctx;
7776 tctx = kmalloc(sizeof(*tctx), GFP_KERNEL);
7777 if (unlikely(!tctx))
7780 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
7781 if (unlikely(ret)) {
7786 tctx->io_wq = io_init_wq_offload(ctx);
7787 if (IS_ERR(tctx->io_wq)) {
7788 ret = PTR_ERR(tctx->io_wq);
7789 percpu_counter_destroy(&tctx->inflight);
7795 init_waitqueue_head(&tctx->wait);
7797 atomic_set(&tctx->in_idle, 0);
7798 tctx->sqpoll = false;
7799 task->io_uring = tctx;
7800 spin_lock_init(&tctx->task_lock);
7801 INIT_WQ_LIST(&tctx->task_list);
7802 tctx->task_state = 0;
7803 init_task_work(&tctx->task_work, tctx_task_work);
7807 void __io_uring_free(struct task_struct *tsk)
7809 struct io_uring_task *tctx = tsk->io_uring;
7811 WARN_ON_ONCE(!xa_empty(&tctx->xa));
7812 WARN_ON_ONCE(tctx->io_wq);
7814 percpu_counter_destroy(&tctx->inflight);
7816 tsk->io_uring = NULL;
7819 static int io_sq_thread_fork(struct io_sq_data *sqd, struct io_ring_ctx *ctx)
7823 clear_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7824 reinit_completion(&sqd->completion);
7826 sqd->task_pid = current->pid;
7827 current->flags |= PF_IO_WORKER;
7828 ret = io_wq_fork_thread(io_sq_thread, sqd);
7829 current->flags &= ~PF_IO_WORKER;
7834 wait_for_completion(&sqd->completion);
7835 return io_uring_alloc_task_context(sqd->thread, ctx);
7838 static int io_sq_offload_create(struct io_ring_ctx *ctx,
7839 struct io_uring_params *p)
7843 /* Retain compatibility with failing for an invalid attach attempt */
7844 if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
7845 IORING_SETUP_ATTACH_WQ) {
7848 f = fdget(p->wq_fd);
7851 if (f.file->f_op != &io_uring_fops) {
7857 if (ctx->flags & IORING_SETUP_SQPOLL) {
7858 struct io_sq_data *sqd;
7861 if (!capable(CAP_SYS_ADMIN) && !capable(CAP_SYS_NICE))
7864 sqd = io_get_sq_data(p);
7871 io_sq_thread_park(sqd);
7872 mutex_lock(&sqd->ctx_lock);
7873 list_add(&ctx->sqd_list, &sqd->ctx_new_list);
7874 mutex_unlock(&sqd->ctx_lock);
7875 io_sq_thread_unpark(sqd);
7877 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
7878 if (!ctx->sq_thread_idle)
7879 ctx->sq_thread_idle = HZ;
7884 if (p->flags & IORING_SETUP_SQ_AFF) {
7885 int cpu = p->sq_thread_cpu;
7888 if (cpu >= nr_cpu_ids)
7890 if (!cpu_online(cpu))
7898 sqd->task_pid = current->pid;
7899 current->flags |= PF_IO_WORKER;
7900 ret = io_wq_fork_thread(io_sq_thread, sqd);
7901 current->flags &= ~PF_IO_WORKER;
7906 wait_for_completion(&sqd->completion);
7907 ret = io_uring_alloc_task_context(sqd->thread, ctx);
7910 } else if (p->flags & IORING_SETUP_SQ_AFF) {
7911 /* Can't have SQ_AFF without SQPOLL */
7918 io_sq_thread_finish(ctx);
7922 static void io_sq_offload_start(struct io_ring_ctx *ctx)
7924 struct io_sq_data *sqd = ctx->sq_data;
7926 ctx->flags &= ~IORING_SETUP_R_DISABLED;
7927 if (ctx->flags & IORING_SETUP_SQPOLL)
7928 complete(&sqd->startup);
7931 static inline void __io_unaccount_mem(struct user_struct *user,
7932 unsigned long nr_pages)
7934 atomic_long_sub(nr_pages, &user->locked_vm);
7937 static inline int __io_account_mem(struct user_struct *user,
7938 unsigned long nr_pages)
7940 unsigned long page_limit, cur_pages, new_pages;
7942 /* Don't allow more pages than we can safely lock */
7943 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
7946 cur_pages = atomic_long_read(&user->locked_vm);
7947 new_pages = cur_pages + nr_pages;
7948 if (new_pages > page_limit)
7950 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
7951 new_pages) != cur_pages);
7956 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
7959 __io_unaccount_mem(ctx->user, nr_pages);
7961 if (ctx->mm_account)
7962 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
7965 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
7970 ret = __io_account_mem(ctx->user, nr_pages);
7975 if (ctx->mm_account)
7976 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
7981 static void io_mem_free(void *ptr)
7988 page = virt_to_head_page(ptr);
7989 if (put_page_testzero(page))
7990 free_compound_page(page);
7993 static void *io_mem_alloc(size_t size)
7995 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
7996 __GFP_NORETRY | __GFP_ACCOUNT;
7998 return (void *) __get_free_pages(gfp_flags, get_order(size));
8001 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8004 struct io_rings *rings;
8005 size_t off, sq_array_size;
8007 off = struct_size(rings, cqes, cq_entries);
8008 if (off == SIZE_MAX)
8012 off = ALIGN(off, SMP_CACHE_BYTES);
8020 sq_array_size = array_size(sizeof(u32), sq_entries);
8021 if (sq_array_size == SIZE_MAX)
8024 if (check_add_overflow(off, sq_array_size, &off))
8030 static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8034 if (!ctx->user_bufs)
8037 for (i = 0; i < ctx->nr_user_bufs; i++) {
8038 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8040 for (j = 0; j < imu->nr_bvecs; j++)
8041 unpin_user_page(imu->bvec[j].bv_page);
8043 if (imu->acct_pages)
8044 io_unaccount_mem(ctx, imu->acct_pages);
8049 kfree(ctx->user_bufs);
8050 ctx->user_bufs = NULL;
8051 ctx->nr_user_bufs = 0;
8055 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8056 void __user *arg, unsigned index)
8058 struct iovec __user *src;
8060 #ifdef CONFIG_COMPAT
8062 struct compat_iovec __user *ciovs;
8063 struct compat_iovec ciov;
8065 ciovs = (struct compat_iovec __user *) arg;
8066 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8069 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8070 dst->iov_len = ciov.iov_len;
8074 src = (struct iovec __user *) arg;
8075 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8081 * Not super efficient, but this is just a registration time. And we do cache
8082 * the last compound head, so generally we'll only do a full search if we don't
8085 * We check if the given compound head page has already been accounted, to
8086 * avoid double accounting it. This allows us to account the full size of the
8087 * page, not just the constituent pages of a huge page.
8089 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8090 int nr_pages, struct page *hpage)
8094 /* check current page array */
8095 for (i = 0; i < nr_pages; i++) {
8096 if (!PageCompound(pages[i]))
8098 if (compound_head(pages[i]) == hpage)
8102 /* check previously registered pages */
8103 for (i = 0; i < ctx->nr_user_bufs; i++) {
8104 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8106 for (j = 0; j < imu->nr_bvecs; j++) {
8107 if (!PageCompound(imu->bvec[j].bv_page))
8109 if (compound_head(imu->bvec[j].bv_page) == hpage)
8117 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8118 int nr_pages, struct io_mapped_ubuf *imu,
8119 struct page **last_hpage)
8123 for (i = 0; i < nr_pages; i++) {
8124 if (!PageCompound(pages[i])) {
8129 hpage = compound_head(pages[i]);
8130 if (hpage == *last_hpage)
8132 *last_hpage = hpage;
8133 if (headpage_already_acct(ctx, pages, i, hpage))
8135 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8139 if (!imu->acct_pages)
8142 ret = io_account_mem(ctx, imu->acct_pages);
8144 imu->acct_pages = 0;
8148 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
8149 struct io_mapped_ubuf *imu,
8150 struct page **last_hpage)
8152 struct vm_area_struct **vmas = NULL;
8153 struct page **pages = NULL;
8154 unsigned long off, start, end, ubuf;
8156 int ret, pret, nr_pages, i;
8158 ubuf = (unsigned long) iov->iov_base;
8159 end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8160 start = ubuf >> PAGE_SHIFT;
8161 nr_pages = end - start;
8165 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
8169 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
8174 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
8180 mmap_read_lock(current->mm);
8181 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
8183 if (pret == nr_pages) {
8184 /* don't support file backed memory */
8185 for (i = 0; i < nr_pages; i++) {
8186 struct vm_area_struct *vma = vmas[i];
8189 !is_file_hugepages(vma->vm_file)) {
8195 ret = pret < 0 ? pret : -EFAULT;
8197 mmap_read_unlock(current->mm);
8200 * if we did partial map, or found file backed vmas,
8201 * release any pages we did get
8204 unpin_user_pages(pages, pret);
8209 ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
8211 unpin_user_pages(pages, pret);
8216 off = ubuf & ~PAGE_MASK;
8217 size = iov->iov_len;
8218 for (i = 0; i < nr_pages; i++) {
8221 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8222 imu->bvec[i].bv_page = pages[i];
8223 imu->bvec[i].bv_len = vec_len;
8224 imu->bvec[i].bv_offset = off;
8228 /* store original address for later verification */
8230 imu->len = iov->iov_len;
8231 imu->nr_bvecs = nr_pages;
8239 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
8243 if (!nr_args || nr_args > UIO_MAXIOV)
8246 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
8248 if (!ctx->user_bufs)
8254 static int io_buffer_validate(struct iovec *iov)
8257 * Don't impose further limits on the size and buffer
8258 * constraints here, we'll -EINVAL later when IO is
8259 * submitted if they are wrong.
8261 if (!iov->iov_base || !iov->iov_len)
8264 /* arbitrary limit, but we need something */
8265 if (iov->iov_len > SZ_1G)
8271 static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
8272 unsigned int nr_args)
8276 struct page *last_hpage = NULL;
8278 ret = io_buffers_map_alloc(ctx, nr_args);
8282 for (i = 0; i < nr_args; i++) {
8283 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8285 ret = io_copy_iov(ctx, &iov, arg, i);
8289 ret = io_buffer_validate(&iov);
8293 ret = io_sqe_buffer_register(ctx, &iov, imu, &last_hpage);
8297 ctx->nr_user_bufs++;
8301 io_sqe_buffers_unregister(ctx);
8306 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8308 __s32 __user *fds = arg;
8314 if (copy_from_user(&fd, fds, sizeof(*fds)))
8317 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8318 if (IS_ERR(ctx->cq_ev_fd)) {
8319 int ret = PTR_ERR(ctx->cq_ev_fd);
8320 ctx->cq_ev_fd = NULL;
8327 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8329 if (ctx->cq_ev_fd) {
8330 eventfd_ctx_put(ctx->cq_ev_fd);
8331 ctx->cq_ev_fd = NULL;
8338 static int __io_destroy_buffers(int id, void *p, void *data)
8340 struct io_ring_ctx *ctx = data;
8341 struct io_buffer *buf = p;
8343 __io_remove_buffers(ctx, buf, id, -1U);
8347 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8349 idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx);
8350 idr_destroy(&ctx->io_buffer_idr);
8353 static void io_req_cache_free(struct list_head *list, struct task_struct *tsk)
8355 struct io_kiocb *req, *nxt;
8357 list_for_each_entry_safe(req, nxt, list, compl.list) {
8358 if (tsk && req->task != tsk)
8360 list_del(&req->compl.list);
8361 kmem_cache_free(req_cachep, req);
8365 static void io_req_caches_free(struct io_ring_ctx *ctx)
8367 struct io_submit_state *submit_state = &ctx->submit_state;
8368 struct io_comp_state *cs = &ctx->submit_state.comp;
8370 mutex_lock(&ctx->uring_lock);
8372 if (submit_state->free_reqs) {
8373 kmem_cache_free_bulk(req_cachep, submit_state->free_reqs,
8374 submit_state->reqs);
8375 submit_state->free_reqs = 0;
8378 spin_lock_irq(&ctx->completion_lock);
8379 list_splice_init(&cs->locked_free_list, &cs->free_list);
8380 cs->locked_free_nr = 0;
8381 spin_unlock_irq(&ctx->completion_lock);
8383 io_req_cache_free(&cs->free_list, NULL);
8385 mutex_unlock(&ctx->uring_lock);
8388 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8391 * Some may use context even when all refs and requests have been put,
8392 * and they are free to do so while still holding uring_lock, see
8393 * __io_req_task_submit(). Wait for them to finish.
8395 mutex_lock(&ctx->uring_lock);
8396 mutex_unlock(&ctx->uring_lock);
8398 io_sq_thread_finish(ctx);
8399 io_sqe_buffers_unregister(ctx);
8401 if (ctx->mm_account) {
8402 mmdrop(ctx->mm_account);
8403 ctx->mm_account = NULL;
8406 mutex_lock(&ctx->uring_lock);
8407 io_sqe_files_unregister(ctx);
8408 mutex_unlock(&ctx->uring_lock);
8409 io_eventfd_unregister(ctx);
8410 io_destroy_buffers(ctx);
8411 idr_destroy(&ctx->personality_idr);
8413 #if defined(CONFIG_UNIX)
8414 if (ctx->ring_sock) {
8415 ctx->ring_sock->file = NULL; /* so that iput() is called */
8416 sock_release(ctx->ring_sock);
8420 io_mem_free(ctx->rings);
8421 io_mem_free(ctx->sq_sqes);
8423 percpu_ref_exit(&ctx->refs);
8424 free_uid(ctx->user);
8425 io_req_caches_free(ctx);
8427 io_wq_put_hash(ctx->hash_map);
8428 kfree(ctx->cancel_hash);
8432 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8434 struct io_ring_ctx *ctx = file->private_data;
8437 poll_wait(file, &ctx->cq_wait, wait);
8439 * synchronizes with barrier from wq_has_sleeper call in
8443 if (!io_sqring_full(ctx))
8444 mask |= EPOLLOUT | EPOLLWRNORM;
8447 * Don't flush cqring overflow list here, just do a simple check.
8448 * Otherwise there could possible be ABBA deadlock:
8451 * lock(&ctx->uring_lock);
8453 * lock(&ctx->uring_lock);
8456 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
8457 * pushs them to do the flush.
8459 if (io_cqring_events(ctx) || test_bit(0, &ctx->cq_check_overflow))
8460 mask |= EPOLLIN | EPOLLRDNORM;
8465 static int io_uring_fasync(int fd, struct file *file, int on)
8467 struct io_ring_ctx *ctx = file->private_data;
8469 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8472 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8474 const struct cred *creds;
8476 creds = idr_remove(&ctx->personality_idr, id);
8485 static int io_remove_personalities(int id, void *p, void *data)
8487 struct io_ring_ctx *ctx = data;
8489 io_unregister_personality(ctx, id);
8493 static bool io_run_ctx_fallback(struct io_ring_ctx *ctx)
8495 struct callback_head *work, *next;
8496 bool executed = false;
8499 work = xchg(&ctx->exit_task_work, NULL);
8515 static void io_ring_exit_work(struct work_struct *work)
8517 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
8521 * If we're doing polled IO and end up having requests being
8522 * submitted async (out-of-line), then completions can come in while
8523 * we're waiting for refs to drop. We need to reap these manually,
8524 * as nobody else will be looking for them.
8527 io_uring_try_cancel_requests(ctx, NULL, NULL);
8528 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
8529 io_ring_ctx_free(ctx);
8532 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8534 mutex_lock(&ctx->uring_lock);
8535 percpu_ref_kill(&ctx->refs);
8536 /* if force is set, the ring is going away. always drop after that */
8537 ctx->cq_overflow_flushed = 1;
8539 __io_cqring_overflow_flush(ctx, true, NULL, NULL);
8540 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
8541 mutex_unlock(&ctx->uring_lock);
8543 io_kill_timeouts(ctx, NULL, NULL);
8544 io_poll_remove_all(ctx, NULL, NULL);
8546 /* if we failed setting up the ctx, we might not have any rings */
8547 io_iopoll_try_reap_events(ctx);
8549 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8551 * Use system_unbound_wq to avoid spawning tons of event kworkers
8552 * if we're exiting a ton of rings at the same time. It just adds
8553 * noise and overhead, there's no discernable change in runtime
8554 * over using system_wq.
8556 queue_work(system_unbound_wq, &ctx->exit_work);
8559 static int io_uring_release(struct inode *inode, struct file *file)
8561 struct io_ring_ctx *ctx = file->private_data;
8563 file->private_data = NULL;
8564 io_ring_ctx_wait_and_kill(ctx);
8568 struct io_task_cancel {
8569 struct task_struct *task;
8570 struct files_struct *files;
8573 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8575 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8576 struct io_task_cancel *cancel = data;
8579 if (cancel->files && (req->flags & REQ_F_LINK_TIMEOUT)) {
8580 unsigned long flags;
8581 struct io_ring_ctx *ctx = req->ctx;
8583 /* protect against races with linked timeouts */
8584 spin_lock_irqsave(&ctx->completion_lock, flags);
8585 ret = io_match_task(req, cancel->task, cancel->files);
8586 spin_unlock_irqrestore(&ctx->completion_lock, flags);
8588 ret = io_match_task(req, cancel->task, cancel->files);
8593 static void io_cancel_defer_files(struct io_ring_ctx *ctx,
8594 struct task_struct *task,
8595 struct files_struct *files)
8597 struct io_defer_entry *de = NULL;
8600 spin_lock_irq(&ctx->completion_lock);
8601 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8602 if (io_match_task(de->req, task, files)) {
8603 list_cut_position(&list, &ctx->defer_list, &de->list);
8607 spin_unlock_irq(&ctx->completion_lock);
8609 while (!list_empty(&list)) {
8610 de = list_first_entry(&list, struct io_defer_entry, list);
8611 list_del_init(&de->list);
8612 req_set_fail_links(de->req);
8613 io_put_req(de->req);
8614 io_req_complete(de->req, -ECANCELED);
8619 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
8620 struct task_struct *task,
8621 struct files_struct *files)
8623 struct io_task_cancel cancel = { .task = task, .files = files, };
8624 struct task_struct *tctx_task = task ?: current;
8625 struct io_uring_task *tctx = tctx_task->io_uring;
8628 enum io_wq_cancel cret;
8631 if (tctx && tctx->io_wq) {
8632 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
8634 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8637 /* SQPOLL thread does its own polling */
8638 if (!(ctx->flags & IORING_SETUP_SQPOLL) && !files) {
8639 while (!list_empty_careful(&ctx->iopoll_list)) {
8640 io_iopoll_try_reap_events(ctx);
8645 ret |= io_poll_remove_all(ctx, task, files);
8646 ret |= io_kill_timeouts(ctx, task, files);
8647 ret |= io_run_task_work();
8648 ret |= io_run_ctx_fallback(ctx);
8649 io_cqring_overflow_flush(ctx, true, task, files);
8656 static int io_uring_count_inflight(struct io_ring_ctx *ctx,
8657 struct task_struct *task,
8658 struct files_struct *files)
8660 struct io_kiocb *req;
8663 spin_lock_irq(&ctx->inflight_lock);
8664 list_for_each_entry(req, &ctx->inflight_list, inflight_entry)
8665 cnt += io_match_task(req, task, files);
8666 spin_unlock_irq(&ctx->inflight_lock);
8670 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
8671 struct task_struct *task,
8672 struct files_struct *files)
8674 while (!list_empty_careful(&ctx->inflight_list)) {
8678 inflight = io_uring_count_inflight(ctx, task, files);
8682 io_uring_try_cancel_requests(ctx, task, files);
8685 io_sq_thread_unpark(ctx->sq_data);
8686 prepare_to_wait(&task->io_uring->wait, &wait,
8687 TASK_UNINTERRUPTIBLE);
8688 if (inflight == io_uring_count_inflight(ctx, task, files))
8690 finish_wait(&task->io_uring->wait, &wait);
8692 io_sq_thread_park(ctx->sq_data);
8697 * We need to iteratively cancel requests, in case a request has dependent
8698 * hard links. These persist even for failure of cancelations, hence keep
8699 * looping until none are found.
8701 static void io_uring_cancel_task_requests(struct io_ring_ctx *ctx,
8702 struct files_struct *files)
8704 struct task_struct *task = current;
8705 bool did_park = false;
8707 if ((ctx->flags & IORING_SETUP_SQPOLL) && ctx->sq_data) {
8708 /* never started, nothing to cancel */
8709 if (ctx->flags & IORING_SETUP_R_DISABLED) {
8710 io_sq_offload_start(ctx);
8713 did_park = io_sq_thread_park(ctx->sq_data);
8715 task = ctx->sq_data->thread;
8716 atomic_inc(&task->io_uring->in_idle);
8720 io_cancel_defer_files(ctx, task, files);
8722 io_uring_cancel_files(ctx, task, files);
8724 io_uring_try_cancel_requests(ctx, task, NULL);
8727 atomic_dec(&task->io_uring->in_idle);
8728 io_sq_thread_unpark(ctx->sq_data);
8733 * Note that this task has used io_uring. We use it for cancelation purposes.
8735 static int io_uring_add_task_file(struct io_ring_ctx *ctx, struct file *file)
8737 struct io_uring_task *tctx = current->io_uring;
8740 if (unlikely(!tctx)) {
8741 ret = io_uring_alloc_task_context(current, ctx);
8744 tctx = current->io_uring;
8746 if (tctx->last != file) {
8747 void *old = xa_load(&tctx->xa, (unsigned long)file);
8751 ret = xa_err(xa_store(&tctx->xa, (unsigned long)file,
8762 * This is race safe in that the task itself is doing this, hence it
8763 * cannot be going through the exit/cancel paths at the same time.
8764 * This cannot be modified while exit/cancel is running.
8766 if (!tctx->sqpoll && (ctx->flags & IORING_SETUP_SQPOLL))
8767 tctx->sqpoll = true;
8773 * Remove this io_uring_file -> task mapping.
8775 static void io_uring_del_task_file(struct file *file)
8777 struct io_uring_task *tctx = current->io_uring;
8779 if (tctx->last == file)
8781 file = xa_erase(&tctx->xa, (unsigned long)file);
8786 static void io_uring_clean_tctx(struct io_uring_task *tctx)
8789 unsigned long index;
8791 xa_for_each(&tctx->xa, index, file)
8792 io_uring_del_task_file(file);
8794 io_wq_put_and_exit(tctx->io_wq);
8799 void __io_uring_files_cancel(struct files_struct *files)
8801 struct io_uring_task *tctx = current->io_uring;
8803 unsigned long index;
8805 /* make sure overflow events are dropped */
8806 atomic_inc(&tctx->in_idle);
8807 xa_for_each(&tctx->xa, index, file)
8808 io_uring_cancel_task_requests(file->private_data, files);
8809 atomic_dec(&tctx->in_idle);
8812 io_uring_clean_tctx(tctx);
8815 static s64 tctx_inflight(struct io_uring_task *tctx)
8817 return percpu_counter_sum(&tctx->inflight);
8820 static void io_uring_cancel_sqpoll(struct io_ring_ctx *ctx)
8822 struct io_sq_data *sqd = ctx->sq_data;
8823 struct io_uring_task *tctx;
8829 if (!io_sq_thread_park(sqd))
8831 tctx = ctx->sq_data->thread->io_uring;
8832 /* can happen on fork/alloc failure, just ignore that state */
8834 io_sq_thread_unpark(sqd);
8838 atomic_inc(&tctx->in_idle);
8840 /* read completions before cancelations */
8841 inflight = tctx_inflight(tctx);
8844 io_uring_cancel_task_requests(ctx, NULL);
8846 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
8848 * If we've seen completions, retry without waiting. This
8849 * avoids a race where a completion comes in before we did
8850 * prepare_to_wait().
8852 if (inflight == tctx_inflight(tctx))
8854 finish_wait(&tctx->wait, &wait);
8856 atomic_dec(&tctx->in_idle);
8857 io_sq_thread_unpark(sqd);
8861 * Find any io_uring fd that this task has registered or done IO on, and cancel
8864 void __io_uring_task_cancel(void)
8866 struct io_uring_task *tctx = current->io_uring;
8870 /* make sure overflow events are dropped */
8871 atomic_inc(&tctx->in_idle);
8875 unsigned long index;
8877 xa_for_each(&tctx->xa, index, file)
8878 io_uring_cancel_sqpoll(file->private_data);
8882 /* read completions before cancelations */
8883 inflight = tctx_inflight(tctx);
8886 __io_uring_files_cancel(NULL);
8888 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
8891 * If we've seen completions, retry without waiting. This
8892 * avoids a race where a completion comes in before we did
8893 * prepare_to_wait().
8895 if (inflight == tctx_inflight(tctx))
8897 finish_wait(&tctx->wait, &wait);
8900 atomic_dec(&tctx->in_idle);
8902 io_uring_clean_tctx(tctx);
8903 /* all current's requests should be gone, we can kill tctx */
8904 __io_uring_free(current);
8907 static void *io_uring_validate_mmap_request(struct file *file,
8908 loff_t pgoff, size_t sz)
8910 struct io_ring_ctx *ctx = file->private_data;
8911 loff_t offset = pgoff << PAGE_SHIFT;
8916 case IORING_OFF_SQ_RING:
8917 case IORING_OFF_CQ_RING:
8920 case IORING_OFF_SQES:
8924 return ERR_PTR(-EINVAL);
8927 page = virt_to_head_page(ptr);
8928 if (sz > page_size(page))
8929 return ERR_PTR(-EINVAL);
8936 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
8938 size_t sz = vma->vm_end - vma->vm_start;
8942 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
8944 return PTR_ERR(ptr);
8946 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
8947 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
8950 #else /* !CONFIG_MMU */
8952 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
8954 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
8957 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
8959 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
8962 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
8963 unsigned long addr, unsigned long len,
8964 unsigned long pgoff, unsigned long flags)
8968 ptr = io_uring_validate_mmap_request(file, pgoff, len);
8970 return PTR_ERR(ptr);
8972 return (unsigned long) ptr;
8975 #endif /* !CONFIG_MMU */
8977 static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
8983 if (!io_sqring_full(ctx))
8985 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
8987 if (!io_sqring_full(ctx))
8990 } while (!signal_pending(current));
8992 finish_wait(&ctx->sqo_sq_wait, &wait);
8996 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
8997 struct __kernel_timespec __user **ts,
8998 const sigset_t __user **sig)
9000 struct io_uring_getevents_arg arg;
9003 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9004 * is just a pointer to the sigset_t.
9006 if (!(flags & IORING_ENTER_EXT_ARG)) {
9007 *sig = (const sigset_t __user *) argp;
9013 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9014 * timespec and sigset_t pointers if good.
9016 if (*argsz != sizeof(arg))
9018 if (copy_from_user(&arg, argp, sizeof(arg)))
9020 *sig = u64_to_user_ptr(arg.sigmask);
9021 *argsz = arg.sigmask_sz;
9022 *ts = u64_to_user_ptr(arg.ts);
9026 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9027 u32, min_complete, u32, flags, const void __user *, argp,
9030 struct io_ring_ctx *ctx;
9037 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9038 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG))
9046 if (f.file->f_op != &io_uring_fops)
9050 ctx = f.file->private_data;
9051 if (!percpu_ref_tryget(&ctx->refs))
9055 if (ctx->flags & IORING_SETUP_R_DISABLED)
9059 * For SQ polling, the thread will do all submissions and completions.
9060 * Just return the requested submit count, and wake the thread if
9064 if (ctx->flags & IORING_SETUP_SQPOLL) {
9065 io_cqring_overflow_flush(ctx, false, NULL, NULL);
9067 if (unlikely(ctx->sqo_exec)) {
9068 ret = io_sq_thread_fork(ctx->sq_data, ctx);
9074 if (flags & IORING_ENTER_SQ_WAKEUP)
9075 wake_up(&ctx->sq_data->wait);
9076 if (flags & IORING_ENTER_SQ_WAIT) {
9077 ret = io_sqpoll_wait_sq(ctx);
9081 submitted = to_submit;
9082 } else if (to_submit) {
9083 ret = io_uring_add_task_file(ctx, f.file);
9086 mutex_lock(&ctx->uring_lock);
9087 submitted = io_submit_sqes(ctx, to_submit);
9088 mutex_unlock(&ctx->uring_lock);
9090 if (submitted != to_submit)
9093 if (flags & IORING_ENTER_GETEVENTS) {
9094 const sigset_t __user *sig;
9095 struct __kernel_timespec __user *ts;
9097 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
9101 min_complete = min(min_complete, ctx->cq_entries);
9104 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9105 * space applications don't need to do io completion events
9106 * polling again, they can rely on io_sq_thread to do polling
9107 * work, which can reduce cpu usage and uring_lock contention.
9109 if (ctx->flags & IORING_SETUP_IOPOLL &&
9110 !(ctx->flags & IORING_SETUP_SQPOLL)) {
9111 ret = io_iopoll_check(ctx, min_complete);
9113 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
9118 percpu_ref_put(&ctx->refs);
9121 return submitted ? submitted : ret;
9124 #ifdef CONFIG_PROC_FS
9125 static int io_uring_show_cred(int id, void *p, void *data)
9127 const struct cred *cred = p;
9128 struct seq_file *m = data;
9129 struct user_namespace *uns = seq_user_ns(m);
9130 struct group_info *gi;
9135 seq_printf(m, "%5d\n", id);
9136 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
9137 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
9138 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
9139 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
9140 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
9141 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
9142 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
9143 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
9144 seq_puts(m, "\n\tGroups:\t");
9145 gi = cred->group_info;
9146 for (g = 0; g < gi->ngroups; g++) {
9147 seq_put_decimal_ull(m, g ? " " : "",
9148 from_kgid_munged(uns, gi->gid[g]));
9150 seq_puts(m, "\n\tCapEff:\t");
9151 cap = cred->cap_effective;
9152 CAP_FOR_EACH_U32(__capi)
9153 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
9158 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
9160 struct io_sq_data *sq = NULL;
9165 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9166 * since fdinfo case grabs it in the opposite direction of normal use
9167 * cases. If we fail to get the lock, we just don't iterate any
9168 * structures that could be going away outside the io_uring mutex.
9170 has_lock = mutex_trylock(&ctx->uring_lock);
9172 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) {
9178 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9179 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9180 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9181 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9182 struct file *f = *io_fixed_file_slot(ctx->file_data, i);
9185 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9187 seq_printf(m, "%5u: <none>\n", i);
9189 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9190 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9191 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
9193 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
9194 (unsigned int) buf->len);
9196 if (has_lock && !idr_is_empty(&ctx->personality_idr)) {
9197 seq_printf(m, "Personalities:\n");
9198 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
9200 seq_printf(m, "PollList:\n");
9201 spin_lock_irq(&ctx->completion_lock);
9202 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9203 struct hlist_head *list = &ctx->cancel_hash[i];
9204 struct io_kiocb *req;
9206 hlist_for_each_entry(req, list, hash_node)
9207 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9208 req->task->task_works != NULL);
9210 spin_unlock_irq(&ctx->completion_lock);
9212 mutex_unlock(&ctx->uring_lock);
9215 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9217 struct io_ring_ctx *ctx = f->private_data;
9219 if (percpu_ref_tryget(&ctx->refs)) {
9220 __io_uring_show_fdinfo(ctx, m);
9221 percpu_ref_put(&ctx->refs);
9226 static const struct file_operations io_uring_fops = {
9227 .release = io_uring_release,
9228 .mmap = io_uring_mmap,
9230 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9231 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9233 .poll = io_uring_poll,
9234 .fasync = io_uring_fasync,
9235 #ifdef CONFIG_PROC_FS
9236 .show_fdinfo = io_uring_show_fdinfo,
9240 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9241 struct io_uring_params *p)
9243 struct io_rings *rings;
9244 size_t size, sq_array_offset;
9246 /* make sure these are sane, as we already accounted them */
9247 ctx->sq_entries = p->sq_entries;
9248 ctx->cq_entries = p->cq_entries;
9250 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9251 if (size == SIZE_MAX)
9254 rings = io_mem_alloc(size);
9259 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9260 rings->sq_ring_mask = p->sq_entries - 1;
9261 rings->cq_ring_mask = p->cq_entries - 1;
9262 rings->sq_ring_entries = p->sq_entries;
9263 rings->cq_ring_entries = p->cq_entries;
9264 ctx->sq_mask = rings->sq_ring_mask;
9265 ctx->cq_mask = rings->cq_ring_mask;
9267 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9268 if (size == SIZE_MAX) {
9269 io_mem_free(ctx->rings);
9274 ctx->sq_sqes = io_mem_alloc(size);
9275 if (!ctx->sq_sqes) {
9276 io_mem_free(ctx->rings);
9284 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
9288 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9292 ret = io_uring_add_task_file(ctx, file);
9297 fd_install(fd, file);
9302 * Allocate an anonymous fd, this is what constitutes the application
9303 * visible backing of an io_uring instance. The application mmaps this
9304 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9305 * we have to tie this fd to a socket for file garbage collection purposes.
9307 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
9310 #if defined(CONFIG_UNIX)
9313 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9316 return ERR_PTR(ret);
9319 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9320 O_RDWR | O_CLOEXEC);
9321 #if defined(CONFIG_UNIX)
9323 sock_release(ctx->ring_sock);
9324 ctx->ring_sock = NULL;
9326 ctx->ring_sock->file = file;
9332 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9333 struct io_uring_params __user *params)
9335 struct io_ring_ctx *ctx;
9341 if (entries > IORING_MAX_ENTRIES) {
9342 if (!(p->flags & IORING_SETUP_CLAMP))
9344 entries = IORING_MAX_ENTRIES;
9348 * Use twice as many entries for the CQ ring. It's possible for the
9349 * application to drive a higher depth than the size of the SQ ring,
9350 * since the sqes are only used at submission time. This allows for
9351 * some flexibility in overcommitting a bit. If the application has
9352 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9353 * of CQ ring entries manually.
9355 p->sq_entries = roundup_pow_of_two(entries);
9356 if (p->flags & IORING_SETUP_CQSIZE) {
9358 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9359 * to a power-of-two, if it isn't already. We do NOT impose
9360 * any cq vs sq ring sizing.
9364 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9365 if (!(p->flags & IORING_SETUP_CLAMP))
9367 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9369 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9370 if (p->cq_entries < p->sq_entries)
9373 p->cq_entries = 2 * p->sq_entries;
9376 ctx = io_ring_ctx_alloc(p);
9379 ctx->compat = in_compat_syscall();
9380 if (!capable(CAP_IPC_LOCK))
9381 ctx->user = get_uid(current_user());
9384 * This is just grabbed for accounting purposes. When a process exits,
9385 * the mm is exited and dropped before the files, hence we need to hang
9386 * on to this mm purely for the purposes of being able to unaccount
9387 * memory (locked/pinned vm). It's not used for anything else.
9389 mmgrab(current->mm);
9390 ctx->mm_account = current->mm;
9392 ret = io_allocate_scq_urings(ctx, p);
9396 ret = io_sq_offload_create(ctx, p);
9400 if (!(p->flags & IORING_SETUP_R_DISABLED))
9401 io_sq_offload_start(ctx);
9403 memset(&p->sq_off, 0, sizeof(p->sq_off));
9404 p->sq_off.head = offsetof(struct io_rings, sq.head);
9405 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9406 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9407 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9408 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9409 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9410 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9412 memset(&p->cq_off, 0, sizeof(p->cq_off));
9413 p->cq_off.head = offsetof(struct io_rings, cq.head);
9414 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9415 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9416 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9417 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9418 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9419 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9421 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9422 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9423 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9424 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
9425 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS;
9427 if (copy_to_user(params, p, sizeof(*p))) {
9432 file = io_uring_get_file(ctx);
9434 ret = PTR_ERR(file);
9439 * Install ring fd as the very last thing, so we don't risk someone
9440 * having closed it before we finish setup
9442 ret = io_uring_install_fd(ctx, file);
9444 /* fput will clean it up */
9449 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9452 io_ring_ctx_wait_and_kill(ctx);
9457 * Sets up an aio uring context, and returns the fd. Applications asks for a
9458 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9459 * params structure passed in.
9461 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9463 struct io_uring_params p;
9466 if (copy_from_user(&p, params, sizeof(p)))
9468 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9473 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9474 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9475 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9476 IORING_SETUP_R_DISABLED))
9479 return io_uring_create(entries, &p, params);
9482 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9483 struct io_uring_params __user *, params)
9485 return io_uring_setup(entries, params);
9488 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9490 struct io_uring_probe *p;
9494 size = struct_size(p, ops, nr_args);
9495 if (size == SIZE_MAX)
9497 p = kzalloc(size, GFP_KERNEL);
9502 if (copy_from_user(p, arg, size))
9505 if (memchr_inv(p, 0, size))
9508 p->last_op = IORING_OP_LAST - 1;
9509 if (nr_args > IORING_OP_LAST)
9510 nr_args = IORING_OP_LAST;
9512 for (i = 0; i < nr_args; i++) {
9514 if (!io_op_defs[i].not_supported)
9515 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9520 if (copy_to_user(arg, p, size))
9527 static int io_register_personality(struct io_ring_ctx *ctx)
9529 const struct cred *creds;
9532 creds = get_current_cred();
9534 ret = idr_alloc_cyclic(&ctx->personality_idr, (void *) creds, 1,
9535 USHRT_MAX, GFP_KERNEL);
9541 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9542 unsigned int nr_args)
9544 struct io_uring_restriction *res;
9548 /* Restrictions allowed only if rings started disabled */
9549 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9552 /* We allow only a single restrictions registration */
9553 if (ctx->restrictions.registered)
9556 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9559 size = array_size(nr_args, sizeof(*res));
9560 if (size == SIZE_MAX)
9563 res = memdup_user(arg, size);
9565 return PTR_ERR(res);
9569 for (i = 0; i < nr_args; i++) {
9570 switch (res[i].opcode) {
9571 case IORING_RESTRICTION_REGISTER_OP:
9572 if (res[i].register_op >= IORING_REGISTER_LAST) {
9577 __set_bit(res[i].register_op,
9578 ctx->restrictions.register_op);
9580 case IORING_RESTRICTION_SQE_OP:
9581 if (res[i].sqe_op >= IORING_OP_LAST) {
9586 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9588 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9589 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9591 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9592 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9601 /* Reset all restrictions if an error happened */
9603 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9605 ctx->restrictions.registered = true;
9611 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9613 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9616 if (ctx->restrictions.registered)
9617 ctx->restricted = 1;
9619 io_sq_offload_start(ctx);
9623 static bool io_register_op_must_quiesce(int op)
9626 case IORING_UNREGISTER_FILES:
9627 case IORING_REGISTER_FILES_UPDATE:
9628 case IORING_REGISTER_PROBE:
9629 case IORING_REGISTER_PERSONALITY:
9630 case IORING_UNREGISTER_PERSONALITY:
9637 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
9638 void __user *arg, unsigned nr_args)
9639 __releases(ctx->uring_lock)
9640 __acquires(ctx->uring_lock)
9645 * We're inside the ring mutex, if the ref is already dying, then
9646 * someone else killed the ctx or is already going through
9647 * io_uring_register().
9649 if (percpu_ref_is_dying(&ctx->refs))
9652 if (io_register_op_must_quiesce(opcode)) {
9653 percpu_ref_kill(&ctx->refs);
9656 * Drop uring mutex before waiting for references to exit. If
9657 * another thread is currently inside io_uring_enter() it might
9658 * need to grab the uring_lock to make progress. If we hold it
9659 * here across the drain wait, then we can deadlock. It's safe
9660 * to drop the mutex here, since no new references will come in
9661 * after we've killed the percpu ref.
9663 mutex_unlock(&ctx->uring_lock);
9665 ret = wait_for_completion_interruptible(&ctx->ref_comp);
9668 ret = io_run_task_work_sig();
9673 mutex_lock(&ctx->uring_lock);
9676 percpu_ref_resurrect(&ctx->refs);
9681 if (ctx->restricted) {
9682 if (opcode >= IORING_REGISTER_LAST) {
9687 if (!test_bit(opcode, ctx->restrictions.register_op)) {
9694 case IORING_REGISTER_BUFFERS:
9695 ret = io_sqe_buffers_register(ctx, arg, nr_args);
9697 case IORING_UNREGISTER_BUFFERS:
9701 ret = io_sqe_buffers_unregister(ctx);
9703 case IORING_REGISTER_FILES:
9704 ret = io_sqe_files_register(ctx, arg, nr_args);
9706 case IORING_UNREGISTER_FILES:
9710 ret = io_sqe_files_unregister(ctx);
9712 case IORING_REGISTER_FILES_UPDATE:
9713 ret = io_sqe_files_update(ctx, arg, nr_args);
9715 case IORING_REGISTER_EVENTFD:
9716 case IORING_REGISTER_EVENTFD_ASYNC:
9720 ret = io_eventfd_register(ctx, arg);
9723 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
9724 ctx->eventfd_async = 1;
9726 ctx->eventfd_async = 0;
9728 case IORING_UNREGISTER_EVENTFD:
9732 ret = io_eventfd_unregister(ctx);
9734 case IORING_REGISTER_PROBE:
9736 if (!arg || nr_args > 256)
9738 ret = io_probe(ctx, arg, nr_args);
9740 case IORING_REGISTER_PERSONALITY:
9744 ret = io_register_personality(ctx);
9746 case IORING_UNREGISTER_PERSONALITY:
9750 ret = io_unregister_personality(ctx, nr_args);
9752 case IORING_REGISTER_ENABLE_RINGS:
9756 ret = io_register_enable_rings(ctx);
9758 case IORING_REGISTER_RESTRICTIONS:
9759 ret = io_register_restrictions(ctx, arg, nr_args);
9767 if (io_register_op_must_quiesce(opcode)) {
9768 /* bring the ctx back to life */
9769 percpu_ref_reinit(&ctx->refs);
9771 reinit_completion(&ctx->ref_comp);
9776 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
9777 void __user *, arg, unsigned int, nr_args)
9779 struct io_ring_ctx *ctx;
9788 if (f.file->f_op != &io_uring_fops)
9791 ctx = f.file->private_data;
9795 mutex_lock(&ctx->uring_lock);
9796 ret = __io_uring_register(ctx, opcode, arg, nr_args);
9797 mutex_unlock(&ctx->uring_lock);
9798 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
9799 ctx->cq_ev_fd != NULL, ret);
9805 static int __init io_uring_init(void)
9807 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
9808 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
9809 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
9812 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
9813 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
9814 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
9815 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
9816 BUILD_BUG_SQE_ELEM(1, __u8, flags);
9817 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
9818 BUILD_BUG_SQE_ELEM(4, __s32, fd);
9819 BUILD_BUG_SQE_ELEM(8, __u64, off);
9820 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
9821 BUILD_BUG_SQE_ELEM(16, __u64, addr);
9822 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
9823 BUILD_BUG_SQE_ELEM(24, __u32, len);
9824 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
9825 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
9826 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
9827 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
9828 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
9829 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
9830 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
9831 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
9832 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
9833 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
9834 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
9835 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
9836 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
9837 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
9838 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
9839 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
9840 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
9841 BUILD_BUG_SQE_ELEM(42, __u16, personality);
9842 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
9844 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
9845 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
9846 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
9850 __initcall(io_uring_init);