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 struct io_uring_task *tctx = req->task->io_uring;
1653 unsigned long flags;
1655 spin_lock_irqsave(&ctx->inflight_lock, flags);
1656 list_del(&req->inflight_entry);
1657 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
1658 req->flags &= ~REQ_F_INFLIGHT;
1659 if (atomic_read(&tctx->in_idle))
1660 wake_up(&tctx->wait);
1664 static inline void io_put_task(struct task_struct *task, int nr)
1666 struct io_uring_task *tctx = task->io_uring;
1668 percpu_counter_sub(&tctx->inflight, nr);
1669 if (unlikely(atomic_read(&tctx->in_idle)))
1670 wake_up(&tctx->wait);
1671 put_task_struct_many(task, nr);
1674 static void __io_free_req(struct io_kiocb *req)
1676 struct io_ring_ctx *ctx = req->ctx;
1678 io_dismantle_req(req);
1679 io_put_task(req->task, 1);
1681 kmem_cache_free(req_cachep, req);
1682 percpu_ref_put(&ctx->refs);
1685 static inline void io_remove_next_linked(struct io_kiocb *req)
1687 struct io_kiocb *nxt = req->link;
1689 req->link = nxt->link;
1693 static void io_kill_linked_timeout(struct io_kiocb *req)
1695 struct io_ring_ctx *ctx = req->ctx;
1696 struct io_kiocb *link;
1697 bool cancelled = false;
1698 unsigned long flags;
1700 spin_lock_irqsave(&ctx->completion_lock, flags);
1704 * Can happen if a linked timeout fired and link had been like
1705 * req -> link t-out -> link t-out [-> ...]
1707 if (link && (link->flags & REQ_F_LTIMEOUT_ACTIVE)) {
1708 struct io_timeout_data *io = link->async_data;
1711 io_remove_next_linked(req);
1712 link->timeout.head = NULL;
1713 ret = hrtimer_try_to_cancel(&io->timer);
1715 io_cqring_fill_event(link, -ECANCELED);
1716 io_commit_cqring(ctx);
1720 req->flags &= ~REQ_F_LINK_TIMEOUT;
1721 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1724 io_cqring_ev_posted(ctx);
1730 static void io_fail_links(struct io_kiocb *req)
1732 struct io_kiocb *link, *nxt;
1733 struct io_ring_ctx *ctx = req->ctx;
1734 unsigned long flags;
1736 spin_lock_irqsave(&ctx->completion_lock, flags);
1744 trace_io_uring_fail_link(req, link);
1745 io_cqring_fill_event(link, -ECANCELED);
1747 io_put_req_deferred(link, 2);
1750 io_commit_cqring(ctx);
1751 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1753 io_cqring_ev_posted(ctx);
1756 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1758 if (req->flags & REQ_F_LINK_TIMEOUT)
1759 io_kill_linked_timeout(req);
1762 * If LINK is set, we have dependent requests in this chain. If we
1763 * didn't fail this request, queue the first one up, moving any other
1764 * dependencies to the next request. In case of failure, fail the rest
1767 if (likely(!(req->flags & REQ_F_FAIL_LINK))) {
1768 struct io_kiocb *nxt = req->link;
1777 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1779 if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK))))
1781 return __io_req_find_next(req);
1784 static void ctx_flush_and_put(struct io_ring_ctx *ctx)
1788 if (ctx->submit_state.comp.nr) {
1789 mutex_lock(&ctx->uring_lock);
1790 io_submit_flush_completions(&ctx->submit_state.comp, ctx);
1791 mutex_unlock(&ctx->uring_lock);
1793 percpu_ref_put(&ctx->refs);
1796 static bool __tctx_task_work(struct io_uring_task *tctx)
1798 struct io_ring_ctx *ctx = NULL;
1799 struct io_wq_work_list list;
1800 struct io_wq_work_node *node;
1802 if (wq_list_empty(&tctx->task_list))
1805 spin_lock_irq(&tctx->task_lock);
1806 list = tctx->task_list;
1807 INIT_WQ_LIST(&tctx->task_list);
1808 spin_unlock_irq(&tctx->task_lock);
1812 struct io_wq_work_node *next = node->next;
1813 struct io_kiocb *req;
1815 req = container_of(node, struct io_kiocb, io_task_work.node);
1816 if (req->ctx != ctx) {
1817 ctx_flush_and_put(ctx);
1819 percpu_ref_get(&ctx->refs);
1822 req->task_work.func(&req->task_work);
1826 ctx_flush_and_put(ctx);
1827 return list.first != NULL;
1830 static void tctx_task_work(struct callback_head *cb)
1832 struct io_uring_task *tctx = container_of(cb, struct io_uring_task, task_work);
1834 clear_bit(0, &tctx->task_state);
1836 while (__tctx_task_work(tctx))
1840 static int io_task_work_add(struct task_struct *tsk, struct io_kiocb *req,
1841 enum task_work_notify_mode notify)
1843 struct io_uring_task *tctx = tsk->io_uring;
1844 struct io_wq_work_node *node, *prev;
1845 unsigned long flags;
1848 WARN_ON_ONCE(!tctx);
1850 spin_lock_irqsave(&tctx->task_lock, flags);
1851 wq_list_add_tail(&req->io_task_work.node, &tctx->task_list);
1852 spin_unlock_irqrestore(&tctx->task_lock, flags);
1854 /* task_work already pending, we're done */
1855 if (test_bit(0, &tctx->task_state) ||
1856 test_and_set_bit(0, &tctx->task_state))
1859 if (!task_work_add(tsk, &tctx->task_work, notify))
1863 * Slow path - we failed, find and delete work. if the work is not
1864 * in the list, it got run and we're fine.
1867 spin_lock_irqsave(&tctx->task_lock, flags);
1868 wq_list_for_each(node, prev, &tctx->task_list) {
1869 if (&req->io_task_work.node == node) {
1870 wq_list_del(&tctx->task_list, node, prev);
1875 spin_unlock_irqrestore(&tctx->task_lock, flags);
1876 clear_bit(0, &tctx->task_state);
1880 static int io_req_task_work_add(struct io_kiocb *req)
1882 struct task_struct *tsk = req->task;
1883 struct io_ring_ctx *ctx = req->ctx;
1884 enum task_work_notify_mode notify;
1887 if (tsk->flags & PF_EXITING)
1891 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
1892 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
1893 * processing task_work. There's no reliable way to tell if TWA_RESUME
1897 if (!(ctx->flags & IORING_SETUP_SQPOLL))
1898 notify = TWA_SIGNAL;
1900 ret = io_task_work_add(tsk, req, notify);
1902 wake_up_process(tsk);
1907 static void io_req_task_work_add_fallback(struct io_kiocb *req,
1908 task_work_func_t cb)
1910 struct io_ring_ctx *ctx = req->ctx;
1911 struct callback_head *head;
1913 init_task_work(&req->task_work, cb);
1915 head = READ_ONCE(ctx->exit_task_work);
1916 req->task_work.next = head;
1917 } while (cmpxchg(&ctx->exit_task_work, head, &req->task_work) != head);
1920 static void __io_req_task_cancel(struct io_kiocb *req, int error)
1922 struct io_ring_ctx *ctx = req->ctx;
1924 spin_lock_irq(&ctx->completion_lock);
1925 io_cqring_fill_event(req, error);
1926 io_commit_cqring(ctx);
1927 spin_unlock_irq(&ctx->completion_lock);
1929 io_cqring_ev_posted(ctx);
1930 req_set_fail_links(req);
1931 io_double_put_req(req);
1934 static void io_req_task_cancel(struct callback_head *cb)
1936 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1937 struct io_ring_ctx *ctx = req->ctx;
1939 mutex_lock(&ctx->uring_lock);
1940 __io_req_task_cancel(req, req->result);
1941 mutex_unlock(&ctx->uring_lock);
1942 percpu_ref_put(&ctx->refs);
1945 static void __io_req_task_submit(struct io_kiocb *req)
1947 struct io_ring_ctx *ctx = req->ctx;
1949 /* ctx stays valid until unlock, even if we drop all ours ctx->refs */
1950 mutex_lock(&ctx->uring_lock);
1951 if (!(current->flags & PF_EXITING) && !current->in_execve)
1952 __io_queue_sqe(req);
1954 __io_req_task_cancel(req, -EFAULT);
1955 mutex_unlock(&ctx->uring_lock);
1958 static void io_req_task_submit(struct callback_head *cb)
1960 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1962 __io_req_task_submit(req);
1965 static void io_req_task_queue(struct io_kiocb *req)
1969 req->task_work.func = io_req_task_submit;
1970 ret = io_req_task_work_add(req);
1971 if (unlikely(ret)) {
1972 req->result = -ECANCELED;
1973 percpu_ref_get(&req->ctx->refs);
1974 io_req_task_work_add_fallback(req, io_req_task_cancel);
1978 static void io_req_task_queue_fail(struct io_kiocb *req, int ret)
1980 percpu_ref_get(&req->ctx->refs);
1982 req->task_work.func = io_req_task_cancel;
1984 if (unlikely(io_req_task_work_add(req)))
1985 io_req_task_work_add_fallback(req, io_req_task_cancel);
1988 static inline void io_queue_next(struct io_kiocb *req)
1990 struct io_kiocb *nxt = io_req_find_next(req);
1993 io_req_task_queue(nxt);
1996 static void io_free_req(struct io_kiocb *req)
2003 struct task_struct *task;
2008 static inline void io_init_req_batch(struct req_batch *rb)
2015 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2016 struct req_batch *rb)
2019 io_put_task(rb->task, rb->task_refs);
2021 percpu_ref_put_many(&ctx->refs, rb->ctx_refs);
2024 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req,
2025 struct io_submit_state *state)
2029 if (req->task != rb->task) {
2031 io_put_task(rb->task, rb->task_refs);
2032 rb->task = req->task;
2038 io_dismantle_req(req);
2039 if (state->free_reqs != ARRAY_SIZE(state->reqs))
2040 state->reqs[state->free_reqs++] = req;
2042 list_add(&req->compl.list, &state->comp.free_list);
2045 static void io_submit_flush_completions(struct io_comp_state *cs,
2046 struct io_ring_ctx *ctx)
2049 struct io_kiocb *req;
2050 struct req_batch rb;
2052 io_init_req_batch(&rb);
2053 spin_lock_irq(&ctx->completion_lock);
2054 for (i = 0; i < nr; i++) {
2056 __io_cqring_fill_event(req, req->result, req->compl.cflags);
2058 io_commit_cqring(ctx);
2059 spin_unlock_irq(&ctx->completion_lock);
2061 io_cqring_ev_posted(ctx);
2062 for (i = 0; i < nr; i++) {
2065 /* submission and completion refs */
2066 if (refcount_sub_and_test(2, &req->refs))
2067 io_req_free_batch(&rb, req, &ctx->submit_state);
2070 io_req_free_batch_finish(ctx, &rb);
2075 * Drop reference to request, return next in chain (if there is one) if this
2076 * was the last reference to this request.
2078 static struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2080 struct io_kiocb *nxt = NULL;
2082 if (refcount_dec_and_test(&req->refs)) {
2083 nxt = io_req_find_next(req);
2089 static void io_put_req(struct io_kiocb *req)
2091 if (refcount_dec_and_test(&req->refs))
2095 static void io_put_req_deferred_cb(struct callback_head *cb)
2097 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2102 static void io_free_req_deferred(struct io_kiocb *req)
2106 req->task_work.func = io_put_req_deferred_cb;
2107 ret = io_req_task_work_add(req);
2109 io_req_task_work_add_fallback(req, io_put_req_deferred_cb);
2112 static inline void io_put_req_deferred(struct io_kiocb *req, int refs)
2114 if (refcount_sub_and_test(refs, &req->refs))
2115 io_free_req_deferred(req);
2118 static void io_double_put_req(struct io_kiocb *req)
2120 /* drop both submit and complete references */
2121 if (refcount_sub_and_test(2, &req->refs))
2125 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2127 /* See comment at the top of this file */
2129 return __io_cqring_events(ctx);
2132 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2134 struct io_rings *rings = ctx->rings;
2136 /* make sure SQ entry isn't read before tail */
2137 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2140 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2142 unsigned int cflags;
2144 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2145 cflags |= IORING_CQE_F_BUFFER;
2146 req->flags &= ~REQ_F_BUFFER_SELECTED;
2151 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2153 struct io_buffer *kbuf;
2155 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2156 return io_put_kbuf(req, kbuf);
2159 static inline bool io_run_task_work(void)
2162 * Not safe to run on exiting task, and the task_work handling will
2163 * not add work to such a task.
2165 if (unlikely(current->flags & PF_EXITING))
2167 if (current->task_works) {
2168 __set_current_state(TASK_RUNNING);
2177 * Find and free completed poll iocbs
2179 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2180 struct list_head *done)
2182 struct req_batch rb;
2183 struct io_kiocb *req;
2185 /* order with ->result store in io_complete_rw_iopoll() */
2188 io_init_req_batch(&rb);
2189 while (!list_empty(done)) {
2192 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2193 list_del(&req->inflight_entry);
2195 if (READ_ONCE(req->result) == -EAGAIN) {
2196 req->iopoll_completed = 0;
2197 if (io_rw_reissue(req))
2201 if (req->flags & REQ_F_BUFFER_SELECTED)
2202 cflags = io_put_rw_kbuf(req);
2204 __io_cqring_fill_event(req, req->result, cflags);
2207 if (refcount_dec_and_test(&req->refs))
2208 io_req_free_batch(&rb, req, &ctx->submit_state);
2211 io_commit_cqring(ctx);
2212 io_cqring_ev_posted_iopoll(ctx);
2213 io_req_free_batch_finish(ctx, &rb);
2216 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2219 struct io_kiocb *req, *tmp;
2225 * Only spin for completions if we don't have multiple devices hanging
2226 * off our complete list, and we're under the requested amount.
2228 spin = !ctx->poll_multi_file && *nr_events < min;
2231 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2232 struct kiocb *kiocb = &req->rw.kiocb;
2235 * Move completed and retryable entries to our local lists.
2236 * If we find a request that requires polling, break out
2237 * and complete those lists first, if we have entries there.
2239 if (READ_ONCE(req->iopoll_completed)) {
2240 list_move_tail(&req->inflight_entry, &done);
2243 if (!list_empty(&done))
2246 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2250 /* iopoll may have completed current req */
2251 if (READ_ONCE(req->iopoll_completed))
2252 list_move_tail(&req->inflight_entry, &done);
2259 if (!list_empty(&done))
2260 io_iopoll_complete(ctx, nr_events, &done);
2266 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2267 * non-spinning poll check - we'll still enter the driver poll loop, but only
2268 * as a non-spinning completion check.
2270 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
2273 while (!list_empty(&ctx->iopoll_list) && !need_resched()) {
2276 ret = io_do_iopoll(ctx, nr_events, min);
2279 if (*nr_events >= min)
2287 * We can't just wait for polled events to come to us, we have to actively
2288 * find and complete them.
2290 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2292 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2295 mutex_lock(&ctx->uring_lock);
2296 while (!list_empty(&ctx->iopoll_list)) {
2297 unsigned int nr_events = 0;
2299 io_do_iopoll(ctx, &nr_events, 0);
2301 /* let it sleep and repeat later if can't complete a request */
2305 * Ensure we allow local-to-the-cpu processing to take place,
2306 * in this case we need to ensure that we reap all events.
2307 * Also let task_work, etc. to progress by releasing the mutex
2309 if (need_resched()) {
2310 mutex_unlock(&ctx->uring_lock);
2312 mutex_lock(&ctx->uring_lock);
2315 mutex_unlock(&ctx->uring_lock);
2318 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2320 unsigned int nr_events = 0;
2321 int iters = 0, ret = 0;
2324 * We disallow the app entering submit/complete with polling, but we
2325 * still need to lock the ring to prevent racing with polled issue
2326 * that got punted to a workqueue.
2328 mutex_lock(&ctx->uring_lock);
2331 * Don't enter poll loop if we already have events pending.
2332 * If we do, we can potentially be spinning for commands that
2333 * already triggered a CQE (eg in error).
2335 if (test_bit(0, &ctx->cq_check_overflow))
2336 __io_cqring_overflow_flush(ctx, false, NULL, NULL);
2337 if (io_cqring_events(ctx))
2341 * If a submit got punted to a workqueue, we can have the
2342 * application entering polling for a command before it gets
2343 * issued. That app will hold the uring_lock for the duration
2344 * of the poll right here, so we need to take a breather every
2345 * now and then to ensure that the issue has a chance to add
2346 * the poll to the issued list. Otherwise we can spin here
2347 * forever, while the workqueue is stuck trying to acquire the
2350 if (!(++iters & 7)) {
2351 mutex_unlock(&ctx->uring_lock);
2353 mutex_lock(&ctx->uring_lock);
2356 ret = io_iopoll_getevents(ctx, &nr_events, min);
2360 } while (min && !nr_events && !need_resched());
2362 mutex_unlock(&ctx->uring_lock);
2366 static void kiocb_end_write(struct io_kiocb *req)
2369 * Tell lockdep we inherited freeze protection from submission
2372 if (req->flags & REQ_F_ISREG) {
2373 struct inode *inode = file_inode(req->file);
2375 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
2377 file_end_write(req->file);
2381 static bool io_resubmit_prep(struct io_kiocb *req)
2383 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2385 struct iov_iter iter;
2387 /* already prepared */
2388 if (req->async_data)
2391 switch (req->opcode) {
2392 case IORING_OP_READV:
2393 case IORING_OP_READ_FIXED:
2394 case IORING_OP_READ:
2397 case IORING_OP_WRITEV:
2398 case IORING_OP_WRITE_FIXED:
2399 case IORING_OP_WRITE:
2403 printk_once(KERN_WARNING "io_uring: bad opcode in resubmit %d\n",
2408 ret = io_import_iovec(rw, req, &iovec, &iter, false);
2411 return !io_setup_async_rw(req, iovec, inline_vecs, &iter, false);
2414 static bool io_rw_should_reissue(struct io_kiocb *req)
2416 umode_t mode = file_inode(req->file)->i_mode;
2417 struct io_ring_ctx *ctx = req->ctx;
2419 if (!S_ISBLK(mode) && !S_ISREG(mode))
2421 if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
2422 !(ctx->flags & IORING_SETUP_IOPOLL)))
2425 * If ref is dying, we might be running poll reap from the exit work.
2426 * Don't attempt to reissue from that path, just let it fail with
2429 if (percpu_ref_is_dying(&ctx->refs))
2435 static bool io_rw_reissue(struct io_kiocb *req)
2438 if (!io_rw_should_reissue(req))
2441 lockdep_assert_held(&req->ctx->uring_lock);
2443 if (io_resubmit_prep(req)) {
2444 refcount_inc(&req->refs);
2445 io_queue_async_work(req);
2448 req_set_fail_links(req);
2453 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2454 unsigned int issue_flags)
2458 if ((res == -EAGAIN || res == -EOPNOTSUPP) && io_rw_reissue(req))
2460 if (res != req->result)
2461 req_set_fail_links(req);
2463 if (req->rw.kiocb.ki_flags & IOCB_WRITE)
2464 kiocb_end_write(req);
2465 if (req->flags & REQ_F_BUFFER_SELECTED)
2466 cflags = io_put_rw_kbuf(req);
2467 __io_req_complete(req, issue_flags, res, cflags);
2470 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2472 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2474 __io_complete_rw(req, res, res2, 0);
2477 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2479 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2482 /* Rewind iter, if we have one. iopoll path resubmits as usual */
2483 if (res == -EAGAIN && io_rw_should_reissue(req)) {
2484 struct io_async_rw *rw = req->async_data;
2487 iov_iter_revert(&rw->iter,
2488 req->result - iov_iter_count(&rw->iter));
2489 else if (!io_resubmit_prep(req))
2494 if (kiocb->ki_flags & IOCB_WRITE)
2495 kiocb_end_write(req);
2497 if (res != -EAGAIN && res != req->result)
2498 req_set_fail_links(req);
2500 WRITE_ONCE(req->result, res);
2501 /* order with io_poll_complete() checking ->result */
2503 WRITE_ONCE(req->iopoll_completed, 1);
2507 * After the iocb has been issued, it's safe to be found on the poll list.
2508 * Adding the kiocb to the list AFTER submission ensures that we don't
2509 * find it from a io_iopoll_getevents() thread before the issuer is done
2510 * accessing the kiocb cookie.
2512 static void io_iopoll_req_issued(struct io_kiocb *req, bool in_async)
2514 struct io_ring_ctx *ctx = req->ctx;
2517 * Track whether we have multiple files in our lists. This will impact
2518 * how we do polling eventually, not spinning if we're on potentially
2519 * different devices.
2521 if (list_empty(&ctx->iopoll_list)) {
2522 ctx->poll_multi_file = false;
2523 } else if (!ctx->poll_multi_file) {
2524 struct io_kiocb *list_req;
2526 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2528 if (list_req->file != req->file)
2529 ctx->poll_multi_file = true;
2533 * For fast devices, IO may have already completed. If it has, add
2534 * it to the front so we find it first.
2536 if (READ_ONCE(req->iopoll_completed))
2537 list_add(&req->inflight_entry, &ctx->iopoll_list);
2539 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2542 * If IORING_SETUP_SQPOLL is enabled, sqes are either handled in sq thread
2543 * task context or in io worker task context. If current task context is
2544 * sq thread, we don't need to check whether should wake up sq thread.
2546 if (in_async && (ctx->flags & IORING_SETUP_SQPOLL) &&
2547 wq_has_sleeper(&ctx->sq_data->wait))
2548 wake_up(&ctx->sq_data->wait);
2551 static inline void io_state_file_put(struct io_submit_state *state)
2553 if (state->file_refs) {
2554 fput_many(state->file, state->file_refs);
2555 state->file_refs = 0;
2560 * Get as many references to a file as we have IOs left in this submission,
2561 * assuming most submissions are for one file, or at least that each file
2562 * has more than one submission.
2564 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2569 if (state->file_refs) {
2570 if (state->fd == fd) {
2574 io_state_file_put(state);
2576 state->file = fget_many(fd, state->ios_left);
2577 if (unlikely(!state->file))
2581 state->file_refs = state->ios_left - 1;
2585 static bool io_bdev_nowait(struct block_device *bdev)
2587 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2591 * If we tracked the file through the SCM inflight mechanism, we could support
2592 * any file. For now, just ensure that anything potentially problematic is done
2595 static bool io_file_supports_async(struct file *file, int rw)
2597 umode_t mode = file_inode(file)->i_mode;
2599 if (S_ISBLK(mode)) {
2600 if (IS_ENABLED(CONFIG_BLOCK) &&
2601 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2605 if (S_ISCHR(mode) || S_ISSOCK(mode))
2607 if (S_ISREG(mode)) {
2608 if (IS_ENABLED(CONFIG_BLOCK) &&
2609 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2610 file->f_op != &io_uring_fops)
2615 /* any ->read/write should understand O_NONBLOCK */
2616 if (file->f_flags & O_NONBLOCK)
2619 if (!(file->f_mode & FMODE_NOWAIT))
2623 return file->f_op->read_iter != NULL;
2625 return file->f_op->write_iter != NULL;
2628 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2630 struct io_ring_ctx *ctx = req->ctx;
2631 struct kiocb *kiocb = &req->rw.kiocb;
2632 struct file *file = req->file;
2636 if (S_ISREG(file_inode(file)->i_mode))
2637 req->flags |= REQ_F_ISREG;
2639 kiocb->ki_pos = READ_ONCE(sqe->off);
2640 if (kiocb->ki_pos == -1 && !(file->f_mode & FMODE_STREAM)) {
2641 req->flags |= REQ_F_CUR_POS;
2642 kiocb->ki_pos = file->f_pos;
2644 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2645 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2646 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2650 /* don't allow async punt for O_NONBLOCK or RWF_NOWAIT */
2651 if ((kiocb->ki_flags & IOCB_NOWAIT) || (file->f_flags & O_NONBLOCK))
2652 req->flags |= REQ_F_NOWAIT;
2654 ioprio = READ_ONCE(sqe->ioprio);
2656 ret = ioprio_check_cap(ioprio);
2660 kiocb->ki_ioprio = ioprio;
2662 kiocb->ki_ioprio = get_current_ioprio();
2664 if (ctx->flags & IORING_SETUP_IOPOLL) {
2665 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2666 !kiocb->ki_filp->f_op->iopoll)
2669 kiocb->ki_flags |= IOCB_HIPRI;
2670 kiocb->ki_complete = io_complete_rw_iopoll;
2671 req->iopoll_completed = 0;
2673 if (kiocb->ki_flags & IOCB_HIPRI)
2675 kiocb->ki_complete = io_complete_rw;
2678 req->rw.addr = READ_ONCE(sqe->addr);
2679 req->rw.len = READ_ONCE(sqe->len);
2680 req->buf_index = READ_ONCE(sqe->buf_index);
2684 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2690 case -ERESTARTNOINTR:
2691 case -ERESTARTNOHAND:
2692 case -ERESTART_RESTARTBLOCK:
2694 * We can't just restart the syscall, since previously
2695 * submitted sqes may already be in progress. Just fail this
2701 kiocb->ki_complete(kiocb, ret, 0);
2705 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2706 unsigned int issue_flags)
2708 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2709 struct io_async_rw *io = req->async_data;
2711 /* add previously done IO, if any */
2712 if (io && io->bytes_done > 0) {
2714 ret = io->bytes_done;
2716 ret += io->bytes_done;
2719 if (req->flags & REQ_F_CUR_POS)
2720 req->file->f_pos = kiocb->ki_pos;
2721 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2722 __io_complete_rw(req, ret, 0, issue_flags);
2724 io_rw_done(kiocb, ret);
2727 static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
2729 struct io_ring_ctx *ctx = req->ctx;
2730 size_t len = req->rw.len;
2731 struct io_mapped_ubuf *imu;
2732 u16 index, buf_index = req->buf_index;
2736 if (unlikely(buf_index >= ctx->nr_user_bufs))
2738 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2739 imu = &ctx->user_bufs[index];
2740 buf_addr = req->rw.addr;
2743 if (buf_addr + len < buf_addr)
2745 /* not inside the mapped region */
2746 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2750 * May not be a start of buffer, set size appropriately
2751 * and advance us to the beginning.
2753 offset = buf_addr - imu->ubuf;
2754 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2758 * Don't use iov_iter_advance() here, as it's really slow for
2759 * using the latter parts of a big fixed buffer - it iterates
2760 * over each segment manually. We can cheat a bit here, because
2763 * 1) it's a BVEC iter, we set it up
2764 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2765 * first and last bvec
2767 * So just find our index, and adjust the iterator afterwards.
2768 * If the offset is within the first bvec (or the whole first
2769 * bvec, just use iov_iter_advance(). This makes it easier
2770 * since we can just skip the first segment, which may not
2771 * be PAGE_SIZE aligned.
2773 const struct bio_vec *bvec = imu->bvec;
2775 if (offset <= bvec->bv_len) {
2776 iov_iter_advance(iter, offset);
2778 unsigned long seg_skip;
2780 /* skip first vec */
2781 offset -= bvec->bv_len;
2782 seg_skip = 1 + (offset >> PAGE_SHIFT);
2784 iter->bvec = bvec + seg_skip;
2785 iter->nr_segs -= seg_skip;
2786 iter->count -= bvec->bv_len + offset;
2787 iter->iov_offset = offset & ~PAGE_MASK;
2794 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2797 mutex_unlock(&ctx->uring_lock);
2800 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2803 * "Normal" inline submissions always hold the uring_lock, since we
2804 * grab it from the system call. Same is true for the SQPOLL offload.
2805 * The only exception is when we've detached the request and issue it
2806 * from an async worker thread, grab the lock for that case.
2809 mutex_lock(&ctx->uring_lock);
2812 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2813 int bgid, struct io_buffer *kbuf,
2816 struct io_buffer *head;
2818 if (req->flags & REQ_F_BUFFER_SELECTED)
2821 io_ring_submit_lock(req->ctx, needs_lock);
2823 lockdep_assert_held(&req->ctx->uring_lock);
2825 head = idr_find(&req->ctx->io_buffer_idr, bgid);
2827 if (!list_empty(&head->list)) {
2828 kbuf = list_last_entry(&head->list, struct io_buffer,
2830 list_del(&kbuf->list);
2833 idr_remove(&req->ctx->io_buffer_idr, bgid);
2835 if (*len > kbuf->len)
2838 kbuf = ERR_PTR(-ENOBUFS);
2841 io_ring_submit_unlock(req->ctx, needs_lock);
2846 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2849 struct io_buffer *kbuf;
2852 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2853 bgid = req->buf_index;
2854 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2857 req->rw.addr = (u64) (unsigned long) kbuf;
2858 req->flags |= REQ_F_BUFFER_SELECTED;
2859 return u64_to_user_ptr(kbuf->addr);
2862 #ifdef CONFIG_COMPAT
2863 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2866 struct compat_iovec __user *uiov;
2867 compat_ssize_t clen;
2871 uiov = u64_to_user_ptr(req->rw.addr);
2872 if (!access_ok(uiov, sizeof(*uiov)))
2874 if (__get_user(clen, &uiov->iov_len))
2880 buf = io_rw_buffer_select(req, &len, needs_lock);
2882 return PTR_ERR(buf);
2883 iov[0].iov_base = buf;
2884 iov[0].iov_len = (compat_size_t) len;
2889 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2892 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2896 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2899 len = iov[0].iov_len;
2902 buf = io_rw_buffer_select(req, &len, needs_lock);
2904 return PTR_ERR(buf);
2905 iov[0].iov_base = buf;
2906 iov[0].iov_len = len;
2910 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2913 if (req->flags & REQ_F_BUFFER_SELECTED) {
2914 struct io_buffer *kbuf;
2916 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2917 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
2918 iov[0].iov_len = kbuf->len;
2921 if (req->rw.len != 1)
2924 #ifdef CONFIG_COMPAT
2925 if (req->ctx->compat)
2926 return io_compat_import(req, iov, needs_lock);
2929 return __io_iov_buffer_select(req, iov, needs_lock);
2932 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
2933 struct iov_iter *iter, bool needs_lock)
2935 void __user *buf = u64_to_user_ptr(req->rw.addr);
2936 size_t sqe_len = req->rw.len;
2937 u8 opcode = req->opcode;
2940 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
2942 return io_import_fixed(req, rw, iter);
2945 /* buffer index only valid with fixed read/write, or buffer select */
2946 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
2949 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
2950 if (req->flags & REQ_F_BUFFER_SELECT) {
2951 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
2953 return PTR_ERR(buf);
2954 req->rw.len = sqe_len;
2957 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
2962 if (req->flags & REQ_F_BUFFER_SELECT) {
2963 ret = io_iov_buffer_select(req, *iovec, needs_lock);
2965 iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
2970 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
2974 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
2976 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
2980 * For files that don't have ->read_iter() and ->write_iter(), handle them
2981 * by looping over ->read() or ->write() manually.
2983 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
2985 struct kiocb *kiocb = &req->rw.kiocb;
2986 struct file *file = req->file;
2990 * Don't support polled IO through this interface, and we can't
2991 * support non-blocking either. For the latter, this just causes
2992 * the kiocb to be handled from an async context.
2994 if (kiocb->ki_flags & IOCB_HIPRI)
2996 if (kiocb->ki_flags & IOCB_NOWAIT)
2999 while (iov_iter_count(iter)) {
3003 if (!iov_iter_is_bvec(iter)) {
3004 iovec = iov_iter_iovec(iter);
3006 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3007 iovec.iov_len = req->rw.len;
3011 nr = file->f_op->read(file, iovec.iov_base,
3012 iovec.iov_len, io_kiocb_ppos(kiocb));
3014 nr = file->f_op->write(file, iovec.iov_base,
3015 iovec.iov_len, io_kiocb_ppos(kiocb));
3024 if (nr != iovec.iov_len)
3028 iov_iter_advance(iter, nr);
3034 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3035 const struct iovec *fast_iov, struct iov_iter *iter)
3037 struct io_async_rw *rw = req->async_data;
3039 memcpy(&rw->iter, iter, sizeof(*iter));
3040 rw->free_iovec = iovec;
3042 /* can only be fixed buffers, no need to do anything */
3043 if (iov_iter_is_bvec(iter))
3046 unsigned iov_off = 0;
3048 rw->iter.iov = rw->fast_iov;
3049 if (iter->iov != fast_iov) {
3050 iov_off = iter->iov - fast_iov;
3051 rw->iter.iov += iov_off;
3053 if (rw->fast_iov != fast_iov)
3054 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3055 sizeof(struct iovec) * iter->nr_segs);
3057 req->flags |= REQ_F_NEED_CLEANUP;
3061 static inline int __io_alloc_async_data(struct io_kiocb *req)
3063 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3064 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3065 return req->async_data == NULL;
3068 static int io_alloc_async_data(struct io_kiocb *req)
3070 if (!io_op_defs[req->opcode].needs_async_data)
3073 return __io_alloc_async_data(req);
3076 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3077 const struct iovec *fast_iov,
3078 struct iov_iter *iter, bool force)
3080 if (!force && !io_op_defs[req->opcode].needs_async_data)
3082 if (!req->async_data) {
3083 if (__io_alloc_async_data(req)) {
3088 io_req_map_rw(req, iovec, fast_iov, iter);
3093 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3095 struct io_async_rw *iorw = req->async_data;
3096 struct iovec *iov = iorw->fast_iov;
3099 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3100 if (unlikely(ret < 0))
3103 iorw->bytes_done = 0;
3104 iorw->free_iovec = iov;
3106 req->flags |= REQ_F_NEED_CLEANUP;
3110 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3112 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3114 return io_prep_rw(req, sqe);
3118 * This is our waitqueue callback handler, registered through lock_page_async()
3119 * when we initially tried to do the IO with the iocb armed our waitqueue.
3120 * This gets called when the page is unlocked, and we generally expect that to
3121 * happen when the page IO is completed and the page is now uptodate. This will
3122 * queue a task_work based retry of the operation, attempting to copy the data
3123 * again. If the latter fails because the page was NOT uptodate, then we will
3124 * do a thread based blocking retry of the operation. That's the unexpected
3127 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3128 int sync, void *arg)
3130 struct wait_page_queue *wpq;
3131 struct io_kiocb *req = wait->private;
3132 struct wait_page_key *key = arg;
3134 wpq = container_of(wait, struct wait_page_queue, wait);
3136 if (!wake_page_match(wpq, key))
3139 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3140 list_del_init(&wait->entry);
3142 /* submit ref gets dropped, acquire a new one */
3143 refcount_inc(&req->refs);
3144 io_req_task_queue(req);
3149 * This controls whether a given IO request should be armed for async page
3150 * based retry. If we return false here, the request is handed to the async
3151 * worker threads for retry. If we're doing buffered reads on a regular file,
3152 * we prepare a private wait_page_queue entry and retry the operation. This
3153 * will either succeed because the page is now uptodate and unlocked, or it
3154 * will register a callback when the page is unlocked at IO completion. Through
3155 * that callback, io_uring uses task_work to setup a retry of the operation.
3156 * That retry will attempt the buffered read again. The retry will generally
3157 * succeed, or in rare cases where it fails, we then fall back to using the
3158 * async worker threads for a blocking retry.
3160 static bool io_rw_should_retry(struct io_kiocb *req)
3162 struct io_async_rw *rw = req->async_data;
3163 struct wait_page_queue *wait = &rw->wpq;
3164 struct kiocb *kiocb = &req->rw.kiocb;
3166 /* never retry for NOWAIT, we just complete with -EAGAIN */
3167 if (req->flags & REQ_F_NOWAIT)
3170 /* Only for buffered IO */
3171 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3175 * just use poll if we can, and don't attempt if the fs doesn't
3176 * support callback based unlocks
3178 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3181 wait->wait.func = io_async_buf_func;
3182 wait->wait.private = req;
3183 wait->wait.flags = 0;
3184 INIT_LIST_HEAD(&wait->wait.entry);
3185 kiocb->ki_flags |= IOCB_WAITQ;
3186 kiocb->ki_flags &= ~IOCB_NOWAIT;
3187 kiocb->ki_waitq = wait;
3191 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3193 if (req->file->f_op->read_iter)
3194 return call_read_iter(req->file, &req->rw.kiocb, iter);
3195 else if (req->file->f_op->read)
3196 return loop_rw_iter(READ, req, iter);
3201 static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3203 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3204 struct kiocb *kiocb = &req->rw.kiocb;
3205 struct iov_iter __iter, *iter = &__iter;
3206 struct io_async_rw *rw = req->async_data;
3207 ssize_t io_size, ret, ret2;
3208 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3214 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3218 io_size = iov_iter_count(iter);
3219 req->result = io_size;
3221 /* Ensure we clear previously set non-block flag */
3222 if (!force_nonblock)
3223 kiocb->ki_flags &= ~IOCB_NOWAIT;
3225 kiocb->ki_flags |= IOCB_NOWAIT;
3227 /* If the file doesn't support async, just async punt */
3228 if (force_nonblock && !io_file_supports_async(req->file, READ)) {
3229 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3230 return ret ?: -EAGAIN;
3233 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size);
3234 if (unlikely(ret)) {
3239 ret = io_iter_do_read(req, iter);
3241 if (ret == -EIOCBQUEUED) {
3242 if (req->async_data)
3243 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3245 } else if (ret == -EAGAIN) {
3246 /* IOPOLL retry should happen for io-wq threads */
3247 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3249 /* no retry on NONBLOCK nor RWF_NOWAIT */
3250 if (req->flags & REQ_F_NOWAIT)
3252 /* some cases will consume bytes even on error returns */
3253 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3255 } else if (ret <= 0 || ret == io_size || !force_nonblock ||
3256 (req->flags & REQ_F_NOWAIT) || !(req->flags & REQ_F_ISREG)) {
3257 /* read all, failed, already did sync or don't want to retry */
3261 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3266 rw = req->async_data;
3267 /* now use our persistent iterator, if we aren't already */
3272 rw->bytes_done += ret;
3273 /* if we can retry, do so with the callbacks armed */
3274 if (!io_rw_should_retry(req)) {
3275 kiocb->ki_flags &= ~IOCB_WAITQ;
3280 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3281 * we get -EIOCBQUEUED, then we'll get a notification when the
3282 * desired page gets unlocked. We can also get a partial read
3283 * here, and if we do, then just retry at the new offset.
3285 ret = io_iter_do_read(req, iter);
3286 if (ret == -EIOCBQUEUED)
3288 /* we got some bytes, but not all. retry. */
3289 } while (ret > 0 && ret < io_size);
3291 kiocb_done(kiocb, ret, issue_flags);
3293 /* it's faster to check here then delegate to kfree */
3299 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3301 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3303 return io_prep_rw(req, sqe);
3306 static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3308 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3309 struct kiocb *kiocb = &req->rw.kiocb;
3310 struct iov_iter __iter, *iter = &__iter;
3311 struct io_async_rw *rw = req->async_data;
3312 ssize_t ret, ret2, io_size;
3313 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3319 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3323 io_size = iov_iter_count(iter);
3324 req->result = io_size;
3326 /* Ensure we clear previously set non-block flag */
3327 if (!force_nonblock)
3328 kiocb->ki_flags &= ~IOCB_NOWAIT;
3330 kiocb->ki_flags |= IOCB_NOWAIT;
3332 /* If the file doesn't support async, just async punt */
3333 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
3336 /* file path doesn't support NOWAIT for non-direct_IO */
3337 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3338 (req->flags & REQ_F_ISREG))
3341 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size);
3346 * Open-code file_start_write here to grab freeze protection,
3347 * which will be released by another thread in
3348 * io_complete_rw(). Fool lockdep by telling it the lock got
3349 * released so that it doesn't complain about the held lock when
3350 * we return to userspace.
3352 if (req->flags & REQ_F_ISREG) {
3353 sb_start_write(file_inode(req->file)->i_sb);
3354 __sb_writers_release(file_inode(req->file)->i_sb,
3357 kiocb->ki_flags |= IOCB_WRITE;
3359 if (req->file->f_op->write_iter)
3360 ret2 = call_write_iter(req->file, kiocb, iter);
3361 else if (req->file->f_op->write)
3362 ret2 = loop_rw_iter(WRITE, req, iter);
3367 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3368 * retry them without IOCB_NOWAIT.
3370 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3372 /* no retry on NONBLOCK nor RWF_NOWAIT */
3373 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3375 if (ret2 == -EIOCBQUEUED && req->async_data)
3376 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3377 if (!force_nonblock || ret2 != -EAGAIN) {
3378 /* IOPOLL retry should happen for io-wq threads */
3379 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3382 kiocb_done(kiocb, ret2, issue_flags);
3385 /* some cases will consume bytes even on error returns */
3386 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3387 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3388 return ret ?: -EAGAIN;
3391 /* it's reportedly faster than delegating the null check to kfree() */
3397 static int io_renameat_prep(struct io_kiocb *req,
3398 const struct io_uring_sqe *sqe)
3400 struct io_rename *ren = &req->rename;
3401 const char __user *oldf, *newf;
3403 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3406 ren->old_dfd = READ_ONCE(sqe->fd);
3407 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3408 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3409 ren->new_dfd = READ_ONCE(sqe->len);
3410 ren->flags = READ_ONCE(sqe->rename_flags);
3412 ren->oldpath = getname(oldf);
3413 if (IS_ERR(ren->oldpath))
3414 return PTR_ERR(ren->oldpath);
3416 ren->newpath = getname(newf);
3417 if (IS_ERR(ren->newpath)) {
3418 putname(ren->oldpath);
3419 return PTR_ERR(ren->newpath);
3422 req->flags |= REQ_F_NEED_CLEANUP;
3426 static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3428 struct io_rename *ren = &req->rename;
3431 if (issue_flags & IO_URING_F_NONBLOCK)
3434 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3435 ren->newpath, ren->flags);
3437 req->flags &= ~REQ_F_NEED_CLEANUP;
3439 req_set_fail_links(req);
3440 io_req_complete(req, ret);
3444 static int io_unlinkat_prep(struct io_kiocb *req,
3445 const struct io_uring_sqe *sqe)
3447 struct io_unlink *un = &req->unlink;
3448 const char __user *fname;
3450 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3453 un->dfd = READ_ONCE(sqe->fd);
3455 un->flags = READ_ONCE(sqe->unlink_flags);
3456 if (un->flags & ~AT_REMOVEDIR)
3459 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3460 un->filename = getname(fname);
3461 if (IS_ERR(un->filename))
3462 return PTR_ERR(un->filename);
3464 req->flags |= REQ_F_NEED_CLEANUP;
3468 static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3470 struct io_unlink *un = &req->unlink;
3473 if (issue_flags & IO_URING_F_NONBLOCK)
3476 if (un->flags & AT_REMOVEDIR)
3477 ret = do_rmdir(un->dfd, un->filename);
3479 ret = do_unlinkat(un->dfd, un->filename);
3481 req->flags &= ~REQ_F_NEED_CLEANUP;
3483 req_set_fail_links(req);
3484 io_req_complete(req, ret);
3488 static int io_shutdown_prep(struct io_kiocb *req,
3489 const struct io_uring_sqe *sqe)
3491 #if defined(CONFIG_NET)
3492 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3494 if (sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3498 req->shutdown.how = READ_ONCE(sqe->len);
3505 static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3507 #if defined(CONFIG_NET)
3508 struct socket *sock;
3511 if (issue_flags & IO_URING_F_NONBLOCK)
3514 sock = sock_from_file(req->file);
3515 if (unlikely(!sock))
3518 ret = __sys_shutdown_sock(sock, req->shutdown.how);
3520 req_set_fail_links(req);
3521 io_req_complete(req, ret);
3528 static int __io_splice_prep(struct io_kiocb *req,
3529 const struct io_uring_sqe *sqe)
3531 struct io_splice* sp = &req->splice;
3532 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3534 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3538 sp->len = READ_ONCE(sqe->len);
3539 sp->flags = READ_ONCE(sqe->splice_flags);
3541 if (unlikely(sp->flags & ~valid_flags))
3544 sp->file_in = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in),
3545 (sp->flags & SPLICE_F_FD_IN_FIXED));
3548 req->flags |= REQ_F_NEED_CLEANUP;
3550 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
3552 * Splice operation will be punted aync, and here need to
3553 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3555 req->work.flags |= IO_WQ_WORK_UNBOUND;
3561 static int io_tee_prep(struct io_kiocb *req,
3562 const struct io_uring_sqe *sqe)
3564 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3566 return __io_splice_prep(req, sqe);
3569 static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
3571 struct io_splice *sp = &req->splice;
3572 struct file *in = sp->file_in;
3573 struct file *out = sp->file_out;
3574 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3577 if (issue_flags & IO_URING_F_NONBLOCK)
3580 ret = do_tee(in, out, sp->len, flags);
3582 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3583 req->flags &= ~REQ_F_NEED_CLEANUP;
3586 req_set_fail_links(req);
3587 io_req_complete(req, ret);
3591 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3593 struct io_splice* sp = &req->splice;
3595 sp->off_in = READ_ONCE(sqe->splice_off_in);
3596 sp->off_out = READ_ONCE(sqe->off);
3597 return __io_splice_prep(req, sqe);
3600 static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
3602 struct io_splice *sp = &req->splice;
3603 struct file *in = sp->file_in;
3604 struct file *out = sp->file_out;
3605 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3606 loff_t *poff_in, *poff_out;
3609 if (issue_flags & IO_URING_F_NONBLOCK)
3612 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3613 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3616 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3618 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3619 req->flags &= ~REQ_F_NEED_CLEANUP;
3622 req_set_fail_links(req);
3623 io_req_complete(req, ret);
3628 * IORING_OP_NOP just posts a completion event, nothing else.
3630 static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
3632 struct io_ring_ctx *ctx = req->ctx;
3634 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3637 __io_req_complete(req, issue_flags, 0, 0);
3641 static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3643 struct io_ring_ctx *ctx = req->ctx;
3648 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3650 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3653 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3654 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3657 req->sync.off = READ_ONCE(sqe->off);
3658 req->sync.len = READ_ONCE(sqe->len);
3662 static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
3664 loff_t end = req->sync.off + req->sync.len;
3667 /* fsync always requires a blocking context */
3668 if (issue_flags & IO_URING_F_NONBLOCK)
3671 ret = vfs_fsync_range(req->file, req->sync.off,
3672 end > 0 ? end : LLONG_MAX,
3673 req->sync.flags & IORING_FSYNC_DATASYNC);
3675 req_set_fail_links(req);
3676 io_req_complete(req, ret);
3680 static int io_fallocate_prep(struct io_kiocb *req,
3681 const struct io_uring_sqe *sqe)
3683 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3685 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3688 req->sync.off = READ_ONCE(sqe->off);
3689 req->sync.len = READ_ONCE(sqe->addr);
3690 req->sync.mode = READ_ONCE(sqe->len);
3694 static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
3698 /* fallocate always requiring blocking context */
3699 if (issue_flags & IO_URING_F_NONBLOCK)
3701 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3704 req_set_fail_links(req);
3705 io_req_complete(req, ret);
3709 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3711 const char __user *fname;
3714 if (unlikely(sqe->ioprio || sqe->buf_index))
3716 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3719 /* open.how should be already initialised */
3720 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3721 req->open.how.flags |= O_LARGEFILE;
3723 req->open.dfd = READ_ONCE(sqe->fd);
3724 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3725 req->open.filename = getname(fname);
3726 if (IS_ERR(req->open.filename)) {
3727 ret = PTR_ERR(req->open.filename);
3728 req->open.filename = NULL;
3731 req->open.nofile = rlimit(RLIMIT_NOFILE);
3732 req->flags |= REQ_F_NEED_CLEANUP;
3736 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3740 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3742 mode = READ_ONCE(sqe->len);
3743 flags = READ_ONCE(sqe->open_flags);
3744 req->open.how = build_open_how(flags, mode);
3745 return __io_openat_prep(req, sqe);
3748 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3750 struct open_how __user *how;
3754 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3756 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3757 len = READ_ONCE(sqe->len);
3758 if (len < OPEN_HOW_SIZE_VER0)
3761 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3766 return __io_openat_prep(req, sqe);
3769 static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
3771 struct open_flags op;
3774 bool resolve_nonblock;
3777 ret = build_open_flags(&req->open.how, &op);
3780 nonblock_set = op.open_flag & O_NONBLOCK;
3781 resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
3782 if (issue_flags & IO_URING_F_NONBLOCK) {
3784 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
3785 * it'll always -EAGAIN
3787 if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
3789 op.lookup_flags |= LOOKUP_CACHED;
3790 op.open_flag |= O_NONBLOCK;
3793 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3797 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3798 /* only retry if RESOLVE_CACHED wasn't already set by application */
3799 if ((!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)) &&
3800 file == ERR_PTR(-EAGAIN)) {
3802 * We could hang on to this 'fd', but seems like marginal
3803 * gain for something that is now known to be a slower path.
3804 * So just put it, and we'll get a new one when we retry.
3812 ret = PTR_ERR(file);
3814 if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
3815 file->f_flags &= ~O_NONBLOCK;
3816 fsnotify_open(file);
3817 fd_install(ret, file);
3820 putname(req->open.filename);
3821 req->flags &= ~REQ_F_NEED_CLEANUP;
3823 req_set_fail_links(req);
3824 io_req_complete(req, ret);
3828 static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
3830 return io_openat2(req, issue_flags & IO_URING_F_NONBLOCK);
3833 static int io_remove_buffers_prep(struct io_kiocb *req,
3834 const struct io_uring_sqe *sqe)
3836 struct io_provide_buf *p = &req->pbuf;
3839 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3842 tmp = READ_ONCE(sqe->fd);
3843 if (!tmp || tmp > USHRT_MAX)
3846 memset(p, 0, sizeof(*p));
3848 p->bgid = READ_ONCE(sqe->buf_group);
3852 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3853 int bgid, unsigned nbufs)
3857 /* shouldn't happen */
3861 /* the head kbuf is the list itself */
3862 while (!list_empty(&buf->list)) {
3863 struct io_buffer *nxt;
3865 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3866 list_del(&nxt->list);
3873 idr_remove(&ctx->io_buffer_idr, bgid);
3878 static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
3880 struct io_provide_buf *p = &req->pbuf;
3881 struct io_ring_ctx *ctx = req->ctx;
3882 struct io_buffer *head;
3884 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3886 io_ring_submit_lock(ctx, !force_nonblock);
3888 lockdep_assert_held(&ctx->uring_lock);
3891 head = idr_find(&ctx->io_buffer_idr, p->bgid);
3893 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3895 req_set_fail_links(req);
3897 /* need to hold the lock to complete IOPOLL requests */
3898 if (ctx->flags & IORING_SETUP_IOPOLL) {
3899 __io_req_complete(req, issue_flags, ret, 0);
3900 io_ring_submit_unlock(ctx, !force_nonblock);
3902 io_ring_submit_unlock(ctx, !force_nonblock);
3903 __io_req_complete(req, issue_flags, ret, 0);
3908 static int io_provide_buffers_prep(struct io_kiocb *req,
3909 const struct io_uring_sqe *sqe)
3911 struct io_provide_buf *p = &req->pbuf;
3914 if (sqe->ioprio || sqe->rw_flags)
3917 tmp = READ_ONCE(sqe->fd);
3918 if (!tmp || tmp > USHRT_MAX)
3921 p->addr = READ_ONCE(sqe->addr);
3922 p->len = READ_ONCE(sqe->len);
3924 if (!access_ok(u64_to_user_ptr(p->addr), (p->len * p->nbufs)))
3927 p->bgid = READ_ONCE(sqe->buf_group);
3928 tmp = READ_ONCE(sqe->off);
3929 if (tmp > USHRT_MAX)
3935 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3937 struct io_buffer *buf;
3938 u64 addr = pbuf->addr;
3939 int i, bid = pbuf->bid;
3941 for (i = 0; i < pbuf->nbufs; i++) {
3942 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
3947 buf->len = pbuf->len;
3952 INIT_LIST_HEAD(&buf->list);
3955 list_add_tail(&buf->list, &(*head)->list);
3959 return i ? i : -ENOMEM;
3962 static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
3964 struct io_provide_buf *p = &req->pbuf;
3965 struct io_ring_ctx *ctx = req->ctx;
3966 struct io_buffer *head, *list;
3968 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3970 io_ring_submit_lock(ctx, !force_nonblock);
3972 lockdep_assert_held(&ctx->uring_lock);
3974 list = head = idr_find(&ctx->io_buffer_idr, p->bgid);
3976 ret = io_add_buffers(p, &head);
3981 ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1,
3984 __io_remove_buffers(ctx, head, p->bgid, -1U);
3990 req_set_fail_links(req);
3992 /* need to hold the lock to complete IOPOLL requests */
3993 if (ctx->flags & IORING_SETUP_IOPOLL) {
3994 __io_req_complete(req, issue_flags, ret, 0);
3995 io_ring_submit_unlock(ctx, !force_nonblock);
3997 io_ring_submit_unlock(ctx, !force_nonblock);
3998 __io_req_complete(req, issue_flags, ret, 0);
4003 static int io_epoll_ctl_prep(struct io_kiocb *req,
4004 const struct io_uring_sqe *sqe)
4006 #if defined(CONFIG_EPOLL)
4007 if (sqe->ioprio || sqe->buf_index)
4009 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4012 req->epoll.epfd = READ_ONCE(sqe->fd);
4013 req->epoll.op = READ_ONCE(sqe->len);
4014 req->epoll.fd = READ_ONCE(sqe->off);
4016 if (ep_op_has_event(req->epoll.op)) {
4017 struct epoll_event __user *ev;
4019 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4020 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4030 static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4032 #if defined(CONFIG_EPOLL)
4033 struct io_epoll *ie = &req->epoll;
4035 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4037 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4038 if (force_nonblock && ret == -EAGAIN)
4042 req_set_fail_links(req);
4043 __io_req_complete(req, issue_flags, ret, 0);
4050 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4052 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4053 if (sqe->ioprio || sqe->buf_index || sqe->off)
4055 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4058 req->madvise.addr = READ_ONCE(sqe->addr);
4059 req->madvise.len = READ_ONCE(sqe->len);
4060 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4067 static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4069 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4070 struct io_madvise *ma = &req->madvise;
4073 if (issue_flags & IO_URING_F_NONBLOCK)
4076 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4078 req_set_fail_links(req);
4079 io_req_complete(req, ret);
4086 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4088 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4090 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4093 req->fadvise.offset = READ_ONCE(sqe->off);
4094 req->fadvise.len = READ_ONCE(sqe->len);
4095 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4099 static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4101 struct io_fadvise *fa = &req->fadvise;
4104 if (issue_flags & IO_URING_F_NONBLOCK) {
4105 switch (fa->advice) {
4106 case POSIX_FADV_NORMAL:
4107 case POSIX_FADV_RANDOM:
4108 case POSIX_FADV_SEQUENTIAL:
4115 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4117 req_set_fail_links(req);
4118 io_req_complete(req, ret);
4122 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4124 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4126 if (sqe->ioprio || sqe->buf_index)
4128 if (req->flags & REQ_F_FIXED_FILE)
4131 req->statx.dfd = READ_ONCE(sqe->fd);
4132 req->statx.mask = READ_ONCE(sqe->len);
4133 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4134 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4135 req->statx.flags = READ_ONCE(sqe->statx_flags);
4140 static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4142 struct io_statx *ctx = &req->statx;
4145 if (issue_flags & IO_URING_F_NONBLOCK) {
4146 /* only need file table for an actual valid fd */
4147 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
4148 req->flags |= REQ_F_NO_FILE_TABLE;
4152 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4156 req_set_fail_links(req);
4157 io_req_complete(req, ret);
4161 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4163 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4165 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4166 sqe->rw_flags || sqe->buf_index)
4168 if (req->flags & REQ_F_FIXED_FILE)
4171 req->close.fd = READ_ONCE(sqe->fd);
4175 static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4177 struct files_struct *files = current->files;
4178 struct io_close *close = &req->close;
4179 struct fdtable *fdt;
4185 spin_lock(&files->file_lock);
4186 fdt = files_fdtable(files);
4187 if (close->fd >= fdt->max_fds) {
4188 spin_unlock(&files->file_lock);
4191 file = fdt->fd[close->fd];
4193 spin_unlock(&files->file_lock);
4197 if (file->f_op == &io_uring_fops) {
4198 spin_unlock(&files->file_lock);
4203 /* if the file has a flush method, be safe and punt to async */
4204 if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4205 spin_unlock(&files->file_lock);
4209 ret = __close_fd_get_file(close->fd, &file);
4210 spin_unlock(&files->file_lock);
4217 /* No ->flush() or already async, safely close from here */
4218 ret = filp_close(file, current->files);
4221 req_set_fail_links(req);
4224 __io_req_complete(req, issue_flags, ret, 0);
4228 static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4230 struct io_ring_ctx *ctx = req->ctx;
4232 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4234 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4237 req->sync.off = READ_ONCE(sqe->off);
4238 req->sync.len = READ_ONCE(sqe->len);
4239 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4243 static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4247 /* sync_file_range always requires a blocking context */
4248 if (issue_flags & IO_URING_F_NONBLOCK)
4251 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4254 req_set_fail_links(req);
4255 io_req_complete(req, ret);
4259 #if defined(CONFIG_NET)
4260 static int io_setup_async_msg(struct io_kiocb *req,
4261 struct io_async_msghdr *kmsg)
4263 struct io_async_msghdr *async_msg = req->async_data;
4267 if (io_alloc_async_data(req)) {
4268 kfree(kmsg->free_iov);
4271 async_msg = req->async_data;
4272 req->flags |= REQ_F_NEED_CLEANUP;
4273 memcpy(async_msg, kmsg, sizeof(*kmsg));
4274 async_msg->msg.msg_name = &async_msg->addr;
4275 /* if were using fast_iov, set it to the new one */
4276 if (!async_msg->free_iov)
4277 async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4282 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4283 struct io_async_msghdr *iomsg)
4285 iomsg->msg.msg_name = &iomsg->addr;
4286 iomsg->free_iov = iomsg->fast_iov;
4287 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4288 req->sr_msg.msg_flags, &iomsg->free_iov);
4291 static int io_sendmsg_prep_async(struct io_kiocb *req)
4295 if (!io_op_defs[req->opcode].needs_async_data)
4297 ret = io_sendmsg_copy_hdr(req, req->async_data);
4299 req->flags |= REQ_F_NEED_CLEANUP;
4303 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4305 struct io_sr_msg *sr = &req->sr_msg;
4307 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4310 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4311 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4312 sr->len = READ_ONCE(sqe->len);
4314 #ifdef CONFIG_COMPAT
4315 if (req->ctx->compat)
4316 sr->msg_flags |= MSG_CMSG_COMPAT;
4321 static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4323 struct io_async_msghdr iomsg, *kmsg;
4324 struct socket *sock;
4328 sock = sock_from_file(req->file);
4329 if (unlikely(!sock))
4332 kmsg = req->async_data;
4334 ret = io_sendmsg_copy_hdr(req, &iomsg);
4340 flags = req->sr_msg.msg_flags;
4341 if (flags & MSG_DONTWAIT)
4342 req->flags |= REQ_F_NOWAIT;
4343 else if (issue_flags & IO_URING_F_NONBLOCK)
4344 flags |= MSG_DONTWAIT;
4346 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4347 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4348 return io_setup_async_msg(req, kmsg);
4349 if (ret == -ERESTARTSYS)
4352 /* fast path, check for non-NULL to avoid function call */
4354 kfree(kmsg->free_iov);
4355 req->flags &= ~REQ_F_NEED_CLEANUP;
4357 req_set_fail_links(req);
4358 __io_req_complete(req, issue_flags, ret, 0);
4362 static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4364 struct io_sr_msg *sr = &req->sr_msg;
4367 struct socket *sock;
4371 sock = sock_from_file(req->file);
4372 if (unlikely(!sock))
4375 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4379 msg.msg_name = NULL;
4380 msg.msg_control = NULL;
4381 msg.msg_controllen = 0;
4382 msg.msg_namelen = 0;
4384 flags = req->sr_msg.msg_flags;
4385 if (flags & MSG_DONTWAIT)
4386 req->flags |= REQ_F_NOWAIT;
4387 else if (issue_flags & IO_URING_F_NONBLOCK)
4388 flags |= MSG_DONTWAIT;
4390 msg.msg_flags = flags;
4391 ret = sock_sendmsg(sock, &msg);
4392 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4394 if (ret == -ERESTARTSYS)
4398 req_set_fail_links(req);
4399 __io_req_complete(req, issue_flags, ret, 0);
4403 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4404 struct io_async_msghdr *iomsg)
4406 struct io_sr_msg *sr = &req->sr_msg;
4407 struct iovec __user *uiov;
4411 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4412 &iomsg->uaddr, &uiov, &iov_len);
4416 if (req->flags & REQ_F_BUFFER_SELECT) {
4419 if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4421 sr->len = iomsg->fast_iov[0].iov_len;
4422 iomsg->free_iov = NULL;
4424 iomsg->free_iov = iomsg->fast_iov;
4425 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4426 &iomsg->free_iov, &iomsg->msg.msg_iter,
4435 #ifdef CONFIG_COMPAT
4436 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4437 struct io_async_msghdr *iomsg)
4439 struct compat_msghdr __user *msg_compat;
4440 struct io_sr_msg *sr = &req->sr_msg;
4441 struct compat_iovec __user *uiov;
4446 msg_compat = (struct compat_msghdr __user *) sr->umsg;
4447 ret = __get_compat_msghdr(&iomsg->msg, msg_compat, &iomsg->uaddr,
4452 uiov = compat_ptr(ptr);
4453 if (req->flags & REQ_F_BUFFER_SELECT) {
4454 compat_ssize_t clen;
4458 if (!access_ok(uiov, sizeof(*uiov)))
4460 if (__get_user(clen, &uiov->iov_len))
4465 iomsg->free_iov = NULL;
4467 iomsg->free_iov = iomsg->fast_iov;
4468 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4469 UIO_FASTIOV, &iomsg->free_iov,
4470 &iomsg->msg.msg_iter, true);
4479 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4480 struct io_async_msghdr *iomsg)
4482 iomsg->msg.msg_name = &iomsg->addr;
4484 #ifdef CONFIG_COMPAT
4485 if (req->ctx->compat)
4486 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4489 return __io_recvmsg_copy_hdr(req, iomsg);
4492 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4495 struct io_sr_msg *sr = &req->sr_msg;
4496 struct io_buffer *kbuf;
4498 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4503 req->flags |= REQ_F_BUFFER_SELECTED;
4507 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4509 return io_put_kbuf(req, req->sr_msg.kbuf);
4512 static int io_recvmsg_prep_async(struct io_kiocb *req)
4516 if (!io_op_defs[req->opcode].needs_async_data)
4518 ret = io_recvmsg_copy_hdr(req, req->async_data);
4520 req->flags |= REQ_F_NEED_CLEANUP;
4524 static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4526 struct io_sr_msg *sr = &req->sr_msg;
4528 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4531 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4532 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4533 sr->len = READ_ONCE(sqe->len);
4534 sr->bgid = READ_ONCE(sqe->buf_group);
4536 #ifdef CONFIG_COMPAT
4537 if (req->ctx->compat)
4538 sr->msg_flags |= MSG_CMSG_COMPAT;
4543 static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
4545 struct io_async_msghdr iomsg, *kmsg;
4546 struct socket *sock;
4547 struct io_buffer *kbuf;
4549 int ret, cflags = 0;
4550 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4552 sock = sock_from_file(req->file);
4553 if (unlikely(!sock))
4556 kmsg = req->async_data;
4558 ret = io_recvmsg_copy_hdr(req, &iomsg);
4564 if (req->flags & REQ_F_BUFFER_SELECT) {
4565 kbuf = io_recv_buffer_select(req, !force_nonblock);
4567 return PTR_ERR(kbuf);
4568 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4569 kmsg->fast_iov[0].iov_len = req->sr_msg.len;
4570 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
4571 1, req->sr_msg.len);
4574 flags = req->sr_msg.msg_flags;
4575 if (flags & MSG_DONTWAIT)
4576 req->flags |= REQ_F_NOWAIT;
4577 else if (force_nonblock)
4578 flags |= MSG_DONTWAIT;
4580 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4581 kmsg->uaddr, flags);
4582 if (force_nonblock && ret == -EAGAIN)
4583 return io_setup_async_msg(req, kmsg);
4584 if (ret == -ERESTARTSYS)
4587 if (req->flags & REQ_F_BUFFER_SELECTED)
4588 cflags = io_put_recv_kbuf(req);
4589 /* fast path, check for non-NULL to avoid function call */
4591 kfree(kmsg->free_iov);
4592 req->flags &= ~REQ_F_NEED_CLEANUP;
4594 req_set_fail_links(req);
4595 __io_req_complete(req, issue_flags, ret, cflags);
4599 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
4601 struct io_buffer *kbuf;
4602 struct io_sr_msg *sr = &req->sr_msg;
4604 void __user *buf = sr->buf;
4605 struct socket *sock;
4608 int ret, cflags = 0;
4609 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4611 sock = sock_from_file(req->file);
4612 if (unlikely(!sock))
4615 if (req->flags & REQ_F_BUFFER_SELECT) {
4616 kbuf = io_recv_buffer_select(req, !force_nonblock);
4618 return PTR_ERR(kbuf);
4619 buf = u64_to_user_ptr(kbuf->addr);
4622 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4626 msg.msg_name = NULL;
4627 msg.msg_control = NULL;
4628 msg.msg_controllen = 0;
4629 msg.msg_namelen = 0;
4630 msg.msg_iocb = NULL;
4633 flags = req->sr_msg.msg_flags;
4634 if (flags & MSG_DONTWAIT)
4635 req->flags |= REQ_F_NOWAIT;
4636 else if (force_nonblock)
4637 flags |= MSG_DONTWAIT;
4639 ret = sock_recvmsg(sock, &msg, flags);
4640 if (force_nonblock && ret == -EAGAIN)
4642 if (ret == -ERESTARTSYS)
4645 if (req->flags & REQ_F_BUFFER_SELECTED)
4646 cflags = io_put_recv_kbuf(req);
4648 req_set_fail_links(req);
4649 __io_req_complete(req, issue_flags, ret, cflags);
4653 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4655 struct io_accept *accept = &req->accept;
4657 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4659 if (sqe->ioprio || sqe->len || sqe->buf_index)
4662 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4663 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4664 accept->flags = READ_ONCE(sqe->accept_flags);
4665 accept->nofile = rlimit(RLIMIT_NOFILE);
4669 static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
4671 struct io_accept *accept = &req->accept;
4672 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4673 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4676 if (req->file->f_flags & O_NONBLOCK)
4677 req->flags |= REQ_F_NOWAIT;
4679 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4680 accept->addr_len, accept->flags,
4682 if (ret == -EAGAIN && force_nonblock)
4685 if (ret == -ERESTARTSYS)
4687 req_set_fail_links(req);
4689 __io_req_complete(req, issue_flags, ret, 0);
4693 static int io_connect_prep_async(struct io_kiocb *req)
4695 struct io_async_connect *io = req->async_data;
4696 struct io_connect *conn = &req->connect;
4698 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
4701 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4703 struct io_connect *conn = &req->connect;
4705 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4707 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4710 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4711 conn->addr_len = READ_ONCE(sqe->addr2);
4715 static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
4717 struct io_async_connect __io, *io;
4718 unsigned file_flags;
4720 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4722 if (req->async_data) {
4723 io = req->async_data;
4725 ret = move_addr_to_kernel(req->connect.addr,
4726 req->connect.addr_len,
4733 file_flags = force_nonblock ? O_NONBLOCK : 0;
4735 ret = __sys_connect_file(req->file, &io->address,
4736 req->connect.addr_len, file_flags);
4737 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4738 if (req->async_data)
4740 if (io_alloc_async_data(req)) {
4744 io = req->async_data;
4745 memcpy(req->async_data, &__io, sizeof(__io));
4748 if (ret == -ERESTARTSYS)
4752 req_set_fail_links(req);
4753 __io_req_complete(req, issue_flags, ret, 0);
4756 #else /* !CONFIG_NET */
4757 #define IO_NETOP_FN(op) \
4758 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
4760 return -EOPNOTSUPP; \
4763 #define IO_NETOP_PREP(op) \
4765 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
4767 return -EOPNOTSUPP; \
4770 #define IO_NETOP_PREP_ASYNC(op) \
4772 static int io_##op##_prep_async(struct io_kiocb *req) \
4774 return -EOPNOTSUPP; \
4777 IO_NETOP_PREP_ASYNC(sendmsg);
4778 IO_NETOP_PREP_ASYNC(recvmsg);
4779 IO_NETOP_PREP_ASYNC(connect);
4780 IO_NETOP_PREP(accept);
4783 #endif /* CONFIG_NET */
4785 struct io_poll_table {
4786 struct poll_table_struct pt;
4787 struct io_kiocb *req;
4791 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4792 __poll_t mask, task_work_func_t func)
4796 /* for instances that support it check for an event match first: */
4797 if (mask && !(mask & poll->events))
4800 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4802 list_del_init(&poll->wait.entry);
4805 req->task_work.func = func;
4806 percpu_ref_get(&req->ctx->refs);
4809 * If this fails, then the task is exiting. When a task exits, the
4810 * work gets canceled, so just cancel this request as well instead
4811 * of executing it. We can't safely execute it anyway, as we may not
4812 * have the needed state needed for it anyway.
4814 ret = io_req_task_work_add(req);
4815 if (unlikely(ret)) {
4816 WRITE_ONCE(poll->canceled, true);
4817 io_req_task_work_add_fallback(req, func);
4822 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4823 __acquires(&req->ctx->completion_lock)
4825 struct io_ring_ctx *ctx = req->ctx;
4827 if (!req->result && !READ_ONCE(poll->canceled)) {
4828 struct poll_table_struct pt = { ._key = poll->events };
4830 req->result = vfs_poll(req->file, &pt) & poll->events;
4833 spin_lock_irq(&ctx->completion_lock);
4834 if (!req->result && !READ_ONCE(poll->canceled)) {
4835 add_wait_queue(poll->head, &poll->wait);
4842 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4844 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4845 if (req->opcode == IORING_OP_POLL_ADD)
4846 return req->async_data;
4847 return req->apoll->double_poll;
4850 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4852 if (req->opcode == IORING_OP_POLL_ADD)
4854 return &req->apoll->poll;
4857 static void io_poll_remove_double(struct io_kiocb *req)
4859 struct io_poll_iocb *poll = io_poll_get_double(req);
4861 lockdep_assert_held(&req->ctx->completion_lock);
4863 if (poll && poll->head) {
4864 struct wait_queue_head *head = poll->head;
4866 spin_lock(&head->lock);
4867 list_del_init(&poll->wait.entry);
4868 if (poll->wait.private)
4869 refcount_dec(&req->refs);
4871 spin_unlock(&head->lock);
4875 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
4877 struct io_ring_ctx *ctx = req->ctx;
4879 io_poll_remove_double(req);
4880 req->poll.done = true;
4881 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
4882 io_commit_cqring(ctx);
4885 static void io_poll_task_func(struct callback_head *cb)
4887 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4888 struct io_ring_ctx *ctx = req->ctx;
4889 struct io_kiocb *nxt;
4891 if (io_poll_rewait(req, &req->poll)) {
4892 spin_unlock_irq(&ctx->completion_lock);
4894 hash_del(&req->hash_node);
4895 io_poll_complete(req, req->result, 0);
4896 spin_unlock_irq(&ctx->completion_lock);
4898 nxt = io_put_req_find_next(req);
4899 io_cqring_ev_posted(ctx);
4901 __io_req_task_submit(nxt);
4904 percpu_ref_put(&ctx->refs);
4907 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4908 int sync, void *key)
4910 struct io_kiocb *req = wait->private;
4911 struct io_poll_iocb *poll = io_poll_get_single(req);
4912 __poll_t mask = key_to_poll(key);
4914 /* for instances that support it check for an event match first: */
4915 if (mask && !(mask & poll->events))
4918 list_del_init(&wait->entry);
4920 if (poll && poll->head) {
4923 spin_lock(&poll->head->lock);
4924 done = list_empty(&poll->wait.entry);
4926 list_del_init(&poll->wait.entry);
4927 /* make sure double remove sees this as being gone */
4928 wait->private = NULL;
4929 spin_unlock(&poll->head->lock);
4931 /* use wait func handler, so it matches the rq type */
4932 poll->wait.func(&poll->wait, mode, sync, key);
4935 refcount_dec(&req->refs);
4939 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
4940 wait_queue_func_t wake_func)
4944 poll->canceled = false;
4945 poll->events = events;
4946 INIT_LIST_HEAD(&poll->wait.entry);
4947 init_waitqueue_func_entry(&poll->wait, wake_func);
4950 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
4951 struct wait_queue_head *head,
4952 struct io_poll_iocb **poll_ptr)
4954 struct io_kiocb *req = pt->req;
4957 * If poll->head is already set, it's because the file being polled
4958 * uses multiple waitqueues for poll handling (eg one for read, one
4959 * for write). Setup a separate io_poll_iocb if this happens.
4961 if (unlikely(poll->head)) {
4962 struct io_poll_iocb *poll_one = poll;
4964 /* already have a 2nd entry, fail a third attempt */
4966 pt->error = -EINVAL;
4969 /* double add on the same waitqueue head, ignore */
4970 if (poll->head == head)
4972 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
4974 pt->error = -ENOMEM;
4977 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
4978 refcount_inc(&req->refs);
4979 poll->wait.private = req;
4986 if (poll->events & EPOLLEXCLUSIVE)
4987 add_wait_queue_exclusive(head, &poll->wait);
4989 add_wait_queue(head, &poll->wait);
4992 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
4993 struct poll_table_struct *p)
4995 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
4996 struct async_poll *apoll = pt->req->apoll;
4998 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5001 static void io_async_task_func(struct callback_head *cb)
5003 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
5004 struct async_poll *apoll = req->apoll;
5005 struct io_ring_ctx *ctx = req->ctx;
5007 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
5009 if (io_poll_rewait(req, &apoll->poll)) {
5010 spin_unlock_irq(&ctx->completion_lock);
5011 percpu_ref_put(&ctx->refs);
5015 /* If req is still hashed, it cannot have been canceled. Don't check. */
5016 if (hash_hashed(&req->hash_node))
5017 hash_del(&req->hash_node);
5019 io_poll_remove_double(req);
5020 spin_unlock_irq(&ctx->completion_lock);
5022 if (!READ_ONCE(apoll->poll.canceled))
5023 __io_req_task_submit(req);
5025 __io_req_task_cancel(req, -ECANCELED);
5027 percpu_ref_put(&ctx->refs);
5028 kfree(apoll->double_poll);
5032 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5035 struct io_kiocb *req = wait->private;
5036 struct io_poll_iocb *poll = &req->apoll->poll;
5038 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5041 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5044 static void io_poll_req_insert(struct io_kiocb *req)
5046 struct io_ring_ctx *ctx = req->ctx;
5047 struct hlist_head *list;
5049 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5050 hlist_add_head(&req->hash_node, list);
5053 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5054 struct io_poll_iocb *poll,
5055 struct io_poll_table *ipt, __poll_t mask,
5056 wait_queue_func_t wake_func)
5057 __acquires(&ctx->completion_lock)
5059 struct io_ring_ctx *ctx = req->ctx;
5060 bool cancel = false;
5062 INIT_HLIST_NODE(&req->hash_node);
5063 io_init_poll_iocb(poll, mask, wake_func);
5064 poll->file = req->file;
5065 poll->wait.private = req;
5067 ipt->pt._key = mask;
5069 ipt->error = -EINVAL;
5071 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5073 spin_lock_irq(&ctx->completion_lock);
5074 if (likely(poll->head)) {
5075 spin_lock(&poll->head->lock);
5076 if (unlikely(list_empty(&poll->wait.entry))) {
5082 if (mask || ipt->error)
5083 list_del_init(&poll->wait.entry);
5085 WRITE_ONCE(poll->canceled, true);
5086 else if (!poll->done) /* actually waiting for an event */
5087 io_poll_req_insert(req);
5088 spin_unlock(&poll->head->lock);
5094 static bool io_arm_poll_handler(struct io_kiocb *req)
5096 const struct io_op_def *def = &io_op_defs[req->opcode];
5097 struct io_ring_ctx *ctx = req->ctx;
5098 struct async_poll *apoll;
5099 struct io_poll_table ipt;
5103 if (!req->file || !file_can_poll(req->file))
5105 if (req->flags & REQ_F_POLLED)
5109 else if (def->pollout)
5113 /* if we can't nonblock try, then no point in arming a poll handler */
5114 if (!io_file_supports_async(req->file, rw))
5117 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5118 if (unlikely(!apoll))
5120 apoll->double_poll = NULL;
5122 req->flags |= REQ_F_POLLED;
5127 mask |= POLLIN | POLLRDNORM;
5129 mask |= POLLOUT | POLLWRNORM;
5131 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5132 if ((req->opcode == IORING_OP_RECVMSG) &&
5133 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5136 mask |= POLLERR | POLLPRI;
5138 ipt.pt._qproc = io_async_queue_proc;
5140 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5142 if (ret || ipt.error) {
5143 io_poll_remove_double(req);
5144 spin_unlock_irq(&ctx->completion_lock);
5145 kfree(apoll->double_poll);
5149 spin_unlock_irq(&ctx->completion_lock);
5150 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
5151 apoll->poll.events);
5155 static bool __io_poll_remove_one(struct io_kiocb *req,
5156 struct io_poll_iocb *poll)
5158 bool do_complete = false;
5160 spin_lock(&poll->head->lock);
5161 WRITE_ONCE(poll->canceled, true);
5162 if (!list_empty(&poll->wait.entry)) {
5163 list_del_init(&poll->wait.entry);
5166 spin_unlock(&poll->head->lock);
5167 hash_del(&req->hash_node);
5171 static bool io_poll_remove_one(struct io_kiocb *req)
5175 io_poll_remove_double(req);
5177 if (req->opcode == IORING_OP_POLL_ADD) {
5178 do_complete = __io_poll_remove_one(req, &req->poll);
5180 struct async_poll *apoll = req->apoll;
5182 /* non-poll requests have submit ref still */
5183 do_complete = __io_poll_remove_one(req, &apoll->poll);
5186 kfree(apoll->double_poll);
5192 io_cqring_fill_event(req, -ECANCELED);
5193 io_commit_cqring(req->ctx);
5194 req_set_fail_links(req);
5195 io_put_req_deferred(req, 1);
5202 * Returns true if we found and killed one or more poll requests
5204 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5205 struct files_struct *files)
5207 struct hlist_node *tmp;
5208 struct io_kiocb *req;
5211 spin_lock_irq(&ctx->completion_lock);
5212 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5213 struct hlist_head *list;
5215 list = &ctx->cancel_hash[i];
5216 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5217 if (io_match_task(req, tsk, files))
5218 posted += io_poll_remove_one(req);
5221 spin_unlock_irq(&ctx->completion_lock);
5224 io_cqring_ev_posted(ctx);
5229 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
5231 struct hlist_head *list;
5232 struct io_kiocb *req;
5234 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5235 hlist_for_each_entry(req, list, hash_node) {
5236 if (sqe_addr != req->user_data)
5238 if (io_poll_remove_one(req))
5246 static int io_poll_remove_prep(struct io_kiocb *req,
5247 const struct io_uring_sqe *sqe)
5249 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5251 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
5255 req->poll_remove.addr = READ_ONCE(sqe->addr);
5260 * Find a running poll command that matches one specified in sqe->addr,
5261 * and remove it if found.
5263 static int io_poll_remove(struct io_kiocb *req, unsigned int issue_flags)
5265 struct io_ring_ctx *ctx = req->ctx;
5268 spin_lock_irq(&ctx->completion_lock);
5269 ret = io_poll_cancel(ctx, req->poll_remove.addr);
5270 spin_unlock_irq(&ctx->completion_lock);
5273 req_set_fail_links(req);
5274 io_req_complete(req, ret);
5278 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5281 struct io_kiocb *req = wait->private;
5282 struct io_poll_iocb *poll = &req->poll;
5284 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5287 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5288 struct poll_table_struct *p)
5290 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5292 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5295 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5297 struct io_poll_iocb *poll = &req->poll;
5300 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5302 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
5305 events = READ_ONCE(sqe->poll32_events);
5307 events = swahw32(events);
5309 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP |
5310 (events & EPOLLEXCLUSIVE);
5314 static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5316 struct io_poll_iocb *poll = &req->poll;
5317 struct io_ring_ctx *ctx = req->ctx;
5318 struct io_poll_table ipt;
5321 ipt.pt._qproc = io_poll_queue_proc;
5323 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5326 if (mask) { /* no async, we'd stolen it */
5328 io_poll_complete(req, mask, 0);
5330 spin_unlock_irq(&ctx->completion_lock);
5333 io_cqring_ev_posted(ctx);
5339 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5341 struct io_timeout_data *data = container_of(timer,
5342 struct io_timeout_data, timer);
5343 struct io_kiocb *req = data->req;
5344 struct io_ring_ctx *ctx = req->ctx;
5345 unsigned long flags;
5347 spin_lock_irqsave(&ctx->completion_lock, flags);
5348 list_del_init(&req->timeout.list);
5349 atomic_set(&req->ctx->cq_timeouts,
5350 atomic_read(&req->ctx->cq_timeouts) + 1);
5352 io_cqring_fill_event(req, -ETIME);
5353 io_commit_cqring(ctx);
5354 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5356 io_cqring_ev_posted(ctx);
5357 req_set_fail_links(req);
5359 return HRTIMER_NORESTART;
5362 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5365 struct io_timeout_data *io;
5366 struct io_kiocb *req;
5369 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5370 if (user_data == req->user_data) {
5377 return ERR_PTR(ret);
5379 io = req->async_data;
5380 ret = hrtimer_try_to_cancel(&io->timer);
5382 return ERR_PTR(-EALREADY);
5383 list_del_init(&req->timeout.list);
5387 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5389 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5392 return PTR_ERR(req);
5394 req_set_fail_links(req);
5395 io_cqring_fill_event(req, -ECANCELED);
5396 io_put_req_deferred(req, 1);
5400 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5401 struct timespec64 *ts, enum hrtimer_mode mode)
5403 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5404 struct io_timeout_data *data;
5407 return PTR_ERR(req);
5409 req->timeout.off = 0; /* noseq */
5410 data = req->async_data;
5411 list_add_tail(&req->timeout.list, &ctx->timeout_list);
5412 hrtimer_init(&data->timer, CLOCK_MONOTONIC, mode);
5413 data->timer.function = io_timeout_fn;
5414 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
5418 static int io_timeout_remove_prep(struct io_kiocb *req,
5419 const struct io_uring_sqe *sqe)
5421 struct io_timeout_rem *tr = &req->timeout_rem;
5423 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5425 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5427 if (sqe->ioprio || sqe->buf_index || sqe->len)
5430 tr->addr = READ_ONCE(sqe->addr);
5431 tr->flags = READ_ONCE(sqe->timeout_flags);
5432 if (tr->flags & IORING_TIMEOUT_UPDATE) {
5433 if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS))
5435 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
5437 } else if (tr->flags) {
5438 /* timeout removal doesn't support flags */
5445 static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
5447 return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
5452 * Remove or update an existing timeout command
5454 static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
5456 struct io_timeout_rem *tr = &req->timeout_rem;
5457 struct io_ring_ctx *ctx = req->ctx;
5460 spin_lock_irq(&ctx->completion_lock);
5461 if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE))
5462 ret = io_timeout_cancel(ctx, tr->addr);
5464 ret = io_timeout_update(ctx, tr->addr, &tr->ts,
5465 io_translate_timeout_mode(tr->flags));
5467 io_cqring_fill_event(req, ret);
5468 io_commit_cqring(ctx);
5469 spin_unlock_irq(&ctx->completion_lock);
5470 io_cqring_ev_posted(ctx);
5472 req_set_fail_links(req);
5477 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5478 bool is_timeout_link)
5480 struct io_timeout_data *data;
5482 u32 off = READ_ONCE(sqe->off);
5484 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5486 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5488 if (off && is_timeout_link)
5490 flags = READ_ONCE(sqe->timeout_flags);
5491 if (flags & ~IORING_TIMEOUT_ABS)
5494 req->timeout.off = off;
5496 if (!req->async_data && io_alloc_async_data(req))
5499 data = req->async_data;
5502 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5505 data->mode = io_translate_timeout_mode(flags);
5506 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5510 static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
5512 struct io_ring_ctx *ctx = req->ctx;
5513 struct io_timeout_data *data = req->async_data;
5514 struct list_head *entry;
5515 u32 tail, off = req->timeout.off;
5517 spin_lock_irq(&ctx->completion_lock);
5520 * sqe->off holds how many events that need to occur for this
5521 * timeout event to be satisfied. If it isn't set, then this is
5522 * a pure timeout request, sequence isn't used.
5524 if (io_is_timeout_noseq(req)) {
5525 entry = ctx->timeout_list.prev;
5529 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5530 req->timeout.target_seq = tail + off;
5532 /* Update the last seq here in case io_flush_timeouts() hasn't.
5533 * This is safe because ->completion_lock is held, and submissions
5534 * and completions are never mixed in the same ->completion_lock section.
5536 ctx->cq_last_tm_flush = tail;
5539 * Insertion sort, ensuring the first entry in the list is always
5540 * the one we need first.
5542 list_for_each_prev(entry, &ctx->timeout_list) {
5543 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5546 if (io_is_timeout_noseq(nxt))
5548 /* nxt.seq is behind @tail, otherwise would've been completed */
5549 if (off >= nxt->timeout.target_seq - tail)
5553 list_add(&req->timeout.list, entry);
5554 data->timer.function = io_timeout_fn;
5555 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5556 spin_unlock_irq(&ctx->completion_lock);
5560 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5562 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5564 return req->user_data == (unsigned long) data;
5567 static int io_async_cancel_one(struct io_uring_task *tctx, void *sqe_addr)
5569 enum io_wq_cancel cancel_ret;
5575 cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, sqe_addr, false);
5576 switch (cancel_ret) {
5577 case IO_WQ_CANCEL_OK:
5580 case IO_WQ_CANCEL_RUNNING:
5583 case IO_WQ_CANCEL_NOTFOUND:
5591 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5592 struct io_kiocb *req, __u64 sqe_addr,
5595 unsigned long flags;
5598 ret = io_async_cancel_one(req->task->io_uring,
5599 (void *) (unsigned long) sqe_addr);
5600 if (ret != -ENOENT) {
5601 spin_lock_irqsave(&ctx->completion_lock, flags);
5605 spin_lock_irqsave(&ctx->completion_lock, flags);
5606 ret = io_timeout_cancel(ctx, sqe_addr);
5609 ret = io_poll_cancel(ctx, sqe_addr);
5613 io_cqring_fill_event(req, ret);
5614 io_commit_cqring(ctx);
5615 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5616 io_cqring_ev_posted(ctx);
5619 req_set_fail_links(req);
5623 static int io_async_cancel_prep(struct io_kiocb *req,
5624 const struct io_uring_sqe *sqe)
5626 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5628 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5630 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5633 req->cancel.addr = READ_ONCE(sqe->addr);
5637 static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
5639 struct io_ring_ctx *ctx = req->ctx;
5641 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
5645 static int io_rsrc_update_prep(struct io_kiocb *req,
5646 const struct io_uring_sqe *sqe)
5648 if (unlikely(req->ctx->flags & IORING_SETUP_SQPOLL))
5650 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5652 if (sqe->ioprio || sqe->rw_flags)
5655 req->rsrc_update.offset = READ_ONCE(sqe->off);
5656 req->rsrc_update.nr_args = READ_ONCE(sqe->len);
5657 if (!req->rsrc_update.nr_args)
5659 req->rsrc_update.arg = READ_ONCE(sqe->addr);
5663 static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
5665 struct io_ring_ctx *ctx = req->ctx;
5666 struct io_uring_rsrc_update up;
5669 if (issue_flags & IO_URING_F_NONBLOCK)
5672 up.offset = req->rsrc_update.offset;
5673 up.data = req->rsrc_update.arg;
5675 mutex_lock(&ctx->uring_lock);
5676 ret = __io_sqe_files_update(ctx, &up, req->rsrc_update.nr_args);
5677 mutex_unlock(&ctx->uring_lock);
5680 req_set_fail_links(req);
5681 __io_req_complete(req, issue_flags, ret, 0);
5685 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5687 switch (req->opcode) {
5690 case IORING_OP_READV:
5691 case IORING_OP_READ_FIXED:
5692 case IORING_OP_READ:
5693 return io_read_prep(req, sqe);
5694 case IORING_OP_WRITEV:
5695 case IORING_OP_WRITE_FIXED:
5696 case IORING_OP_WRITE:
5697 return io_write_prep(req, sqe);
5698 case IORING_OP_POLL_ADD:
5699 return io_poll_add_prep(req, sqe);
5700 case IORING_OP_POLL_REMOVE:
5701 return io_poll_remove_prep(req, sqe);
5702 case IORING_OP_FSYNC:
5703 return io_fsync_prep(req, sqe);
5704 case IORING_OP_SYNC_FILE_RANGE:
5705 return io_sfr_prep(req, sqe);
5706 case IORING_OP_SENDMSG:
5707 case IORING_OP_SEND:
5708 return io_sendmsg_prep(req, sqe);
5709 case IORING_OP_RECVMSG:
5710 case IORING_OP_RECV:
5711 return io_recvmsg_prep(req, sqe);
5712 case IORING_OP_CONNECT:
5713 return io_connect_prep(req, sqe);
5714 case IORING_OP_TIMEOUT:
5715 return io_timeout_prep(req, sqe, false);
5716 case IORING_OP_TIMEOUT_REMOVE:
5717 return io_timeout_remove_prep(req, sqe);
5718 case IORING_OP_ASYNC_CANCEL:
5719 return io_async_cancel_prep(req, sqe);
5720 case IORING_OP_LINK_TIMEOUT:
5721 return io_timeout_prep(req, sqe, true);
5722 case IORING_OP_ACCEPT:
5723 return io_accept_prep(req, sqe);
5724 case IORING_OP_FALLOCATE:
5725 return io_fallocate_prep(req, sqe);
5726 case IORING_OP_OPENAT:
5727 return io_openat_prep(req, sqe);
5728 case IORING_OP_CLOSE:
5729 return io_close_prep(req, sqe);
5730 case IORING_OP_FILES_UPDATE:
5731 return io_rsrc_update_prep(req, sqe);
5732 case IORING_OP_STATX:
5733 return io_statx_prep(req, sqe);
5734 case IORING_OP_FADVISE:
5735 return io_fadvise_prep(req, sqe);
5736 case IORING_OP_MADVISE:
5737 return io_madvise_prep(req, sqe);
5738 case IORING_OP_OPENAT2:
5739 return io_openat2_prep(req, sqe);
5740 case IORING_OP_EPOLL_CTL:
5741 return io_epoll_ctl_prep(req, sqe);
5742 case IORING_OP_SPLICE:
5743 return io_splice_prep(req, sqe);
5744 case IORING_OP_PROVIDE_BUFFERS:
5745 return io_provide_buffers_prep(req, sqe);
5746 case IORING_OP_REMOVE_BUFFERS:
5747 return io_remove_buffers_prep(req, sqe);
5749 return io_tee_prep(req, sqe);
5750 case IORING_OP_SHUTDOWN:
5751 return io_shutdown_prep(req, sqe);
5752 case IORING_OP_RENAMEAT:
5753 return io_renameat_prep(req, sqe);
5754 case IORING_OP_UNLINKAT:
5755 return io_unlinkat_prep(req, sqe);
5758 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5763 static int io_req_prep_async(struct io_kiocb *req)
5765 switch (req->opcode) {
5766 case IORING_OP_READV:
5767 case IORING_OP_READ_FIXED:
5768 case IORING_OP_READ:
5769 return io_rw_prep_async(req, READ);
5770 case IORING_OP_WRITEV:
5771 case IORING_OP_WRITE_FIXED:
5772 case IORING_OP_WRITE:
5773 return io_rw_prep_async(req, WRITE);
5774 case IORING_OP_SENDMSG:
5775 case IORING_OP_SEND:
5776 return io_sendmsg_prep_async(req);
5777 case IORING_OP_RECVMSG:
5778 case IORING_OP_RECV:
5779 return io_recvmsg_prep_async(req);
5780 case IORING_OP_CONNECT:
5781 return io_connect_prep_async(req);
5786 static int io_req_defer_prep(struct io_kiocb *req)
5788 if (!io_op_defs[req->opcode].needs_async_data)
5790 /* some opcodes init it during the inital prep */
5791 if (req->async_data)
5793 if (__io_alloc_async_data(req))
5795 return io_req_prep_async(req);
5798 static u32 io_get_sequence(struct io_kiocb *req)
5800 struct io_kiocb *pos;
5801 struct io_ring_ctx *ctx = req->ctx;
5802 u32 total_submitted, nr_reqs = 0;
5804 io_for_each_link(pos, req)
5807 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
5808 return total_submitted - nr_reqs;
5811 static int io_req_defer(struct io_kiocb *req)
5813 struct io_ring_ctx *ctx = req->ctx;
5814 struct io_defer_entry *de;
5818 /* Still need defer if there is pending req in defer list. */
5819 if (likely(list_empty_careful(&ctx->defer_list) &&
5820 !(req->flags & REQ_F_IO_DRAIN)))
5823 seq = io_get_sequence(req);
5824 /* Still a chance to pass the sequence check */
5825 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
5828 ret = io_req_defer_prep(req);
5831 io_prep_async_link(req);
5832 de = kmalloc(sizeof(*de), GFP_KERNEL);
5836 spin_lock_irq(&ctx->completion_lock);
5837 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
5838 spin_unlock_irq(&ctx->completion_lock);
5840 io_queue_async_work(req);
5841 return -EIOCBQUEUED;
5844 trace_io_uring_defer(ctx, req, req->user_data);
5847 list_add_tail(&de->list, &ctx->defer_list);
5848 spin_unlock_irq(&ctx->completion_lock);
5849 return -EIOCBQUEUED;
5852 static void __io_clean_op(struct io_kiocb *req)
5854 if (req->flags & REQ_F_BUFFER_SELECTED) {
5855 switch (req->opcode) {
5856 case IORING_OP_READV:
5857 case IORING_OP_READ_FIXED:
5858 case IORING_OP_READ:
5859 kfree((void *)(unsigned long)req->rw.addr);
5861 case IORING_OP_RECVMSG:
5862 case IORING_OP_RECV:
5863 kfree(req->sr_msg.kbuf);
5866 req->flags &= ~REQ_F_BUFFER_SELECTED;
5869 if (req->flags & REQ_F_NEED_CLEANUP) {
5870 switch (req->opcode) {
5871 case IORING_OP_READV:
5872 case IORING_OP_READ_FIXED:
5873 case IORING_OP_READ:
5874 case IORING_OP_WRITEV:
5875 case IORING_OP_WRITE_FIXED:
5876 case IORING_OP_WRITE: {
5877 struct io_async_rw *io = req->async_data;
5879 kfree(io->free_iovec);
5882 case IORING_OP_RECVMSG:
5883 case IORING_OP_SENDMSG: {
5884 struct io_async_msghdr *io = req->async_data;
5886 kfree(io->free_iov);
5889 case IORING_OP_SPLICE:
5891 io_put_file(req, req->splice.file_in,
5892 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
5894 case IORING_OP_OPENAT:
5895 case IORING_OP_OPENAT2:
5896 if (req->open.filename)
5897 putname(req->open.filename);
5899 case IORING_OP_RENAMEAT:
5900 putname(req->rename.oldpath);
5901 putname(req->rename.newpath);
5903 case IORING_OP_UNLINKAT:
5904 putname(req->unlink.filename);
5907 req->flags &= ~REQ_F_NEED_CLEANUP;
5911 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
5913 struct io_ring_ctx *ctx = req->ctx;
5914 const struct cred *creds = NULL;
5917 if (req->work.personality) {
5918 const struct cred *new_creds;
5920 if (!(issue_flags & IO_URING_F_NONBLOCK))
5921 mutex_lock(&ctx->uring_lock);
5922 new_creds = idr_find(&ctx->personality_idr, req->work.personality);
5923 if (!(issue_flags & IO_URING_F_NONBLOCK))
5924 mutex_unlock(&ctx->uring_lock);
5927 creds = override_creds(new_creds);
5930 switch (req->opcode) {
5932 ret = io_nop(req, issue_flags);
5934 case IORING_OP_READV:
5935 case IORING_OP_READ_FIXED:
5936 case IORING_OP_READ:
5937 ret = io_read(req, issue_flags);
5939 case IORING_OP_WRITEV:
5940 case IORING_OP_WRITE_FIXED:
5941 case IORING_OP_WRITE:
5942 ret = io_write(req, issue_flags);
5944 case IORING_OP_FSYNC:
5945 ret = io_fsync(req, issue_flags);
5947 case IORING_OP_POLL_ADD:
5948 ret = io_poll_add(req, issue_flags);
5950 case IORING_OP_POLL_REMOVE:
5951 ret = io_poll_remove(req, issue_flags);
5953 case IORING_OP_SYNC_FILE_RANGE:
5954 ret = io_sync_file_range(req, issue_flags);
5956 case IORING_OP_SENDMSG:
5957 ret = io_sendmsg(req, issue_flags);
5959 case IORING_OP_SEND:
5960 ret = io_send(req, issue_flags);
5962 case IORING_OP_RECVMSG:
5963 ret = io_recvmsg(req, issue_flags);
5965 case IORING_OP_RECV:
5966 ret = io_recv(req, issue_flags);
5968 case IORING_OP_TIMEOUT:
5969 ret = io_timeout(req, issue_flags);
5971 case IORING_OP_TIMEOUT_REMOVE:
5972 ret = io_timeout_remove(req, issue_flags);
5974 case IORING_OP_ACCEPT:
5975 ret = io_accept(req, issue_flags);
5977 case IORING_OP_CONNECT:
5978 ret = io_connect(req, issue_flags);
5980 case IORING_OP_ASYNC_CANCEL:
5981 ret = io_async_cancel(req, issue_flags);
5983 case IORING_OP_FALLOCATE:
5984 ret = io_fallocate(req, issue_flags);
5986 case IORING_OP_OPENAT:
5987 ret = io_openat(req, issue_flags);
5989 case IORING_OP_CLOSE:
5990 ret = io_close(req, issue_flags);
5992 case IORING_OP_FILES_UPDATE:
5993 ret = io_files_update(req, issue_flags);
5995 case IORING_OP_STATX:
5996 ret = io_statx(req, issue_flags);
5998 case IORING_OP_FADVISE:
5999 ret = io_fadvise(req, issue_flags);
6001 case IORING_OP_MADVISE:
6002 ret = io_madvise(req, issue_flags);
6004 case IORING_OP_OPENAT2:
6005 ret = io_openat2(req, issue_flags);
6007 case IORING_OP_EPOLL_CTL:
6008 ret = io_epoll_ctl(req, issue_flags);
6010 case IORING_OP_SPLICE:
6011 ret = io_splice(req, issue_flags);
6013 case IORING_OP_PROVIDE_BUFFERS:
6014 ret = io_provide_buffers(req, issue_flags);
6016 case IORING_OP_REMOVE_BUFFERS:
6017 ret = io_remove_buffers(req, issue_flags);
6020 ret = io_tee(req, issue_flags);
6022 case IORING_OP_SHUTDOWN:
6023 ret = io_shutdown(req, issue_flags);
6025 case IORING_OP_RENAMEAT:
6026 ret = io_renameat(req, issue_flags);
6028 case IORING_OP_UNLINKAT:
6029 ret = io_unlinkat(req, issue_flags);
6037 revert_creds(creds);
6042 /* If the op doesn't have a file, we're not polling for it */
6043 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
6044 const bool in_async = io_wq_current_is_worker();
6046 /* workqueue context doesn't hold uring_lock, grab it now */
6048 mutex_lock(&ctx->uring_lock);
6050 io_iopoll_req_issued(req, in_async);
6053 mutex_unlock(&ctx->uring_lock);
6059 static void io_wq_submit_work(struct io_wq_work *work)
6061 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6062 struct io_kiocb *timeout;
6065 timeout = io_prep_linked_timeout(req);
6067 io_queue_linked_timeout(timeout);
6069 if (work->flags & IO_WQ_WORK_CANCEL)
6074 ret = io_issue_sqe(req, 0);
6076 * We can get EAGAIN for polled IO even though we're
6077 * forcing a sync submission from here, since we can't
6078 * wait for request slots on the block side.
6086 /* avoid locking problems by failing it from a clean context */
6088 /* io-wq is going to take one down */
6089 refcount_inc(&req->refs);
6090 io_req_task_queue_fail(req, ret);
6094 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6097 struct fixed_rsrc_table *table;
6099 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
6100 return table->files[index & IORING_FILE_TABLE_MASK];
6103 static struct file *io_file_get(struct io_submit_state *state,
6104 struct io_kiocb *req, int fd, bool fixed)
6106 struct io_ring_ctx *ctx = req->ctx;
6110 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6112 fd = array_index_nospec(fd, ctx->nr_user_files);
6113 file = io_file_from_index(ctx, fd);
6114 io_set_resource_node(req);
6116 trace_io_uring_file_get(ctx, fd);
6117 file = __io_file_get(state, fd);
6120 if (file && unlikely(file->f_op == &io_uring_fops))
6121 io_req_track_inflight(req);
6125 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6127 struct io_timeout_data *data = container_of(timer,
6128 struct io_timeout_data, timer);
6129 struct io_kiocb *prev, *req = data->req;
6130 struct io_ring_ctx *ctx = req->ctx;
6131 unsigned long flags;
6133 spin_lock_irqsave(&ctx->completion_lock, flags);
6134 prev = req->timeout.head;
6135 req->timeout.head = NULL;
6138 * We don't expect the list to be empty, that will only happen if we
6139 * race with the completion of the linked work.
6141 if (prev && refcount_inc_not_zero(&prev->refs))
6142 io_remove_next_linked(prev);
6145 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6148 req_set_fail_links(prev);
6149 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6150 io_put_req_deferred(prev, 1);
6152 io_req_complete_post(req, -ETIME, 0);
6153 io_put_req_deferred(req, 1);
6155 return HRTIMER_NORESTART;
6158 static void __io_queue_linked_timeout(struct io_kiocb *req)
6161 * If the back reference is NULL, then our linked request finished
6162 * before we got a chance to setup the timer
6164 if (req->timeout.head) {
6165 struct io_timeout_data *data = req->async_data;
6167 data->timer.function = io_link_timeout_fn;
6168 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6173 static void io_queue_linked_timeout(struct io_kiocb *req)
6175 struct io_ring_ctx *ctx = req->ctx;
6177 spin_lock_irq(&ctx->completion_lock);
6178 __io_queue_linked_timeout(req);
6179 spin_unlock_irq(&ctx->completion_lock);
6181 /* drop submission reference */
6185 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6187 struct io_kiocb *nxt = req->link;
6189 if (!nxt || (req->flags & REQ_F_LINK_TIMEOUT) ||
6190 nxt->opcode != IORING_OP_LINK_TIMEOUT)
6193 nxt->timeout.head = req;
6194 nxt->flags |= REQ_F_LTIMEOUT_ACTIVE;
6195 req->flags |= REQ_F_LINK_TIMEOUT;
6199 static void __io_queue_sqe(struct io_kiocb *req)
6201 struct io_kiocb *linked_timeout = io_prep_linked_timeout(req);
6204 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
6207 * We async punt it if the file wasn't marked NOWAIT, or if the file
6208 * doesn't support non-blocking read/write attempts
6210 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6211 if (!io_arm_poll_handler(req)) {
6213 * Queued up for async execution, worker will release
6214 * submit reference when the iocb is actually submitted.
6216 io_queue_async_work(req);
6218 } else if (likely(!ret)) {
6219 /* drop submission reference */
6220 if (req->flags & REQ_F_COMPLETE_INLINE) {
6221 struct io_ring_ctx *ctx = req->ctx;
6222 struct io_comp_state *cs = &ctx->submit_state.comp;
6224 cs->reqs[cs->nr++] = req;
6225 if (cs->nr == ARRAY_SIZE(cs->reqs))
6226 io_submit_flush_completions(cs, ctx);
6231 req_set_fail_links(req);
6233 io_req_complete(req, ret);
6236 io_queue_linked_timeout(linked_timeout);
6239 static void io_queue_sqe(struct io_kiocb *req)
6243 ret = io_req_defer(req);
6245 if (ret != -EIOCBQUEUED) {
6247 req_set_fail_links(req);
6249 io_req_complete(req, ret);
6251 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6252 ret = io_req_defer_prep(req);
6255 io_queue_async_work(req);
6257 __io_queue_sqe(req);
6262 * Check SQE restrictions (opcode and flags).
6264 * Returns 'true' if SQE is allowed, 'false' otherwise.
6266 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6267 struct io_kiocb *req,
6268 unsigned int sqe_flags)
6270 if (!ctx->restricted)
6273 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6276 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6277 ctx->restrictions.sqe_flags_required)
6280 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6281 ctx->restrictions.sqe_flags_required))
6287 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6288 const struct io_uring_sqe *sqe)
6290 struct io_submit_state *state;
6291 unsigned int sqe_flags;
6294 req->opcode = READ_ONCE(sqe->opcode);
6295 /* same numerical values with corresponding REQ_F_*, safe to copy */
6296 req->flags = sqe_flags = READ_ONCE(sqe->flags);
6297 req->user_data = READ_ONCE(sqe->user_data);
6298 req->async_data = NULL;
6302 req->fixed_rsrc_refs = NULL;
6303 /* one is dropped after submission, the other at completion */
6304 refcount_set(&req->refs, 2);
6305 req->task = current;
6308 /* enforce forwards compatibility on users */
6309 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS)) {
6314 if (unlikely(req->opcode >= IORING_OP_LAST))
6317 if (unlikely(!io_check_restriction(ctx, req, sqe_flags)))
6320 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6321 !io_op_defs[req->opcode].buffer_select)
6324 req->work.list.next = NULL;
6325 req->work.flags = 0;
6326 req->work.personality = READ_ONCE(sqe->personality);
6327 state = &ctx->submit_state;
6330 * Plug now if we have more than 1 IO left after this, and the target
6331 * is potentially a read/write to block based storage.
6333 if (!state->plug_started && state->ios_left > 1 &&
6334 io_op_defs[req->opcode].plug) {
6335 blk_start_plug(&state->plug);
6336 state->plug_started = true;
6339 if (io_op_defs[req->opcode].needs_file) {
6340 bool fixed = req->flags & REQ_F_FIXED_FILE;
6342 req->file = io_file_get(state, req, READ_ONCE(sqe->fd), fixed);
6343 if (unlikely(!req->file))
6351 static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
6352 const struct io_uring_sqe *sqe)
6354 struct io_submit_link *link = &ctx->submit_state.link;
6357 ret = io_init_req(ctx, req, sqe);
6358 if (unlikely(ret)) {
6361 io_req_complete(req, ret);
6363 /* fail even hard links since we don't submit */
6364 link->head->flags |= REQ_F_FAIL_LINK;
6365 io_put_req(link->head);
6366 io_req_complete(link->head, -ECANCELED);
6371 ret = io_req_prep(req, sqe);
6375 /* don't need @sqe from now on */
6376 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6377 true, ctx->flags & IORING_SETUP_SQPOLL);
6380 * If we already have a head request, queue this one for async
6381 * submittal once the head completes. If we don't have a head but
6382 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6383 * submitted sync once the chain is complete. If none of those
6384 * conditions are true (normal request), then just queue it.
6387 struct io_kiocb *head = link->head;
6390 * Taking sequential execution of a link, draining both sides
6391 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6392 * requests in the link. So, it drains the head and the
6393 * next after the link request. The last one is done via
6394 * drain_next flag to persist the effect across calls.
6396 if (req->flags & REQ_F_IO_DRAIN) {
6397 head->flags |= REQ_F_IO_DRAIN;
6398 ctx->drain_next = 1;
6400 ret = io_req_defer_prep(req);
6403 trace_io_uring_link(ctx, req, head);
6404 link->last->link = req;
6407 /* last request of a link, enqueue the link */
6408 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6413 if (unlikely(ctx->drain_next)) {
6414 req->flags |= REQ_F_IO_DRAIN;
6415 ctx->drain_next = 0;
6417 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6429 * Batched submission is done, ensure local IO is flushed out.
6431 static void io_submit_state_end(struct io_submit_state *state,
6432 struct io_ring_ctx *ctx)
6434 if (state->link.head)
6435 io_queue_sqe(state->link.head);
6437 io_submit_flush_completions(&state->comp, ctx);
6438 if (state->plug_started)
6439 blk_finish_plug(&state->plug);
6440 io_state_file_put(state);
6444 * Start submission side cache.
6446 static void io_submit_state_start(struct io_submit_state *state,
6447 unsigned int max_ios)
6449 state->plug_started = false;
6450 state->ios_left = max_ios;
6451 /* set only head, no need to init link_last in advance */
6452 state->link.head = NULL;
6455 static void io_commit_sqring(struct io_ring_ctx *ctx)
6457 struct io_rings *rings = ctx->rings;
6460 * Ensure any loads from the SQEs are done at this point,
6461 * since once we write the new head, the application could
6462 * write new data to them.
6464 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6468 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6469 * that is mapped by userspace. This means that care needs to be taken to
6470 * ensure that reads are stable, as we cannot rely on userspace always
6471 * being a good citizen. If members of the sqe are validated and then later
6472 * used, it's important that those reads are done through READ_ONCE() to
6473 * prevent a re-load down the line.
6475 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6477 u32 *sq_array = ctx->sq_array;
6481 * The cached sq head (or cq tail) serves two purposes:
6483 * 1) allows us to batch the cost of updating the user visible
6485 * 2) allows the kernel side to track the head on its own, even
6486 * though the application is the one updating it.
6488 head = READ_ONCE(sq_array[ctx->cached_sq_head++ & ctx->sq_mask]);
6489 if (likely(head < ctx->sq_entries))
6490 return &ctx->sq_sqes[head];
6492 /* drop invalid entries */
6493 ctx->cached_sq_dropped++;
6494 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6498 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6502 /* if we have a backlog and couldn't flush it all, return BUSY */
6503 if (test_bit(0, &ctx->sq_check_overflow)) {
6504 if (!__io_cqring_overflow_flush(ctx, false, NULL, NULL))
6508 /* make sure SQ entry isn't read before tail */
6509 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6511 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6514 percpu_counter_add(¤t->io_uring->inflight, nr);
6515 refcount_add(nr, ¤t->usage);
6516 io_submit_state_start(&ctx->submit_state, nr);
6518 while (submitted < nr) {
6519 const struct io_uring_sqe *sqe;
6520 struct io_kiocb *req;
6522 req = io_alloc_req(ctx);
6523 if (unlikely(!req)) {
6525 submitted = -EAGAIN;
6528 sqe = io_get_sqe(ctx);
6529 if (unlikely(!sqe)) {
6530 kmem_cache_free(req_cachep, req);
6533 /* will complete beyond this point, count as submitted */
6535 if (io_submit_sqe(ctx, req, sqe))
6539 if (unlikely(submitted != nr)) {
6540 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6541 struct io_uring_task *tctx = current->io_uring;
6542 int unused = nr - ref_used;
6544 percpu_ref_put_many(&ctx->refs, unused);
6545 percpu_counter_sub(&tctx->inflight, unused);
6546 put_task_struct_many(current, unused);
6549 io_submit_state_end(&ctx->submit_state, ctx);
6550 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6551 io_commit_sqring(ctx);
6556 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6558 /* Tell userspace we may need a wakeup call */
6559 spin_lock_irq(&ctx->completion_lock);
6560 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6561 spin_unlock_irq(&ctx->completion_lock);
6564 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6566 spin_lock_irq(&ctx->completion_lock);
6567 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6568 spin_unlock_irq(&ctx->completion_lock);
6571 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
6573 unsigned int to_submit;
6576 to_submit = io_sqring_entries(ctx);
6577 /* if we're handling multiple rings, cap submit size for fairness */
6578 if (cap_entries && to_submit > 8)
6581 if (!list_empty(&ctx->iopoll_list) || to_submit) {
6582 unsigned nr_events = 0;
6584 mutex_lock(&ctx->uring_lock);
6585 if (!list_empty(&ctx->iopoll_list))
6586 io_do_iopoll(ctx, &nr_events, 0);
6588 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)))
6589 ret = io_submit_sqes(ctx, to_submit);
6590 mutex_unlock(&ctx->uring_lock);
6593 if (!io_sqring_full(ctx) && wq_has_sleeper(&ctx->sqo_sq_wait))
6594 wake_up(&ctx->sqo_sq_wait);
6599 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
6601 struct io_ring_ctx *ctx;
6602 unsigned sq_thread_idle = 0;
6604 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6605 if (sq_thread_idle < ctx->sq_thread_idle)
6606 sq_thread_idle = ctx->sq_thread_idle;
6609 sqd->sq_thread_idle = sq_thread_idle;
6612 static void io_sqd_init_new(struct io_sq_data *sqd)
6614 struct io_ring_ctx *ctx;
6616 while (!list_empty(&sqd->ctx_new_list)) {
6617 ctx = list_first_entry(&sqd->ctx_new_list, struct io_ring_ctx, sqd_list);
6618 list_move_tail(&ctx->sqd_list, &sqd->ctx_list);
6619 complete(&ctx->sq_thread_comp);
6622 io_sqd_update_thread_idle(sqd);
6625 static bool io_sq_thread_should_stop(struct io_sq_data *sqd)
6627 return test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
6630 static bool io_sq_thread_should_park(struct io_sq_data *sqd)
6632 return test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
6635 static void io_sq_thread_parkme(struct io_sq_data *sqd)
6639 * TASK_PARKED is a special state; we must serialize against
6640 * possible pending wakeups to avoid store-store collisions on
6643 * Such a collision might possibly result in the task state
6644 * changin from TASK_PARKED and us failing the
6645 * wait_task_inactive() in kthread_park().
6647 set_special_state(TASK_PARKED);
6648 if (!test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state))
6652 * Thread is going to call schedule(), do not preempt it,
6653 * or the caller of kthread_park() may spend more time in
6654 * wait_task_inactive().
6657 complete(&sqd->completion);
6658 schedule_preempt_disabled();
6661 __set_current_state(TASK_RUNNING);
6664 static int io_sq_thread(void *data)
6666 struct io_sq_data *sqd = data;
6667 struct io_ring_ctx *ctx;
6668 unsigned long timeout = 0;
6669 char buf[TASK_COMM_LEN];
6672 sprintf(buf, "iou-sqp-%d", sqd->task_pid);
6673 set_task_comm(current, buf);
6674 sqd->thread = current;
6675 current->pf_io_worker = NULL;
6677 if (sqd->sq_cpu != -1)
6678 set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
6680 set_cpus_allowed_ptr(current, cpu_online_mask);
6681 current->flags |= PF_NO_SETAFFINITY;
6683 complete(&sqd->completion);
6685 wait_for_completion(&sqd->startup);
6687 while (!io_sq_thread_should_stop(sqd)) {
6689 bool cap_entries, sqt_spin, needs_sched;
6692 * Any changes to the sqd lists are synchronized through the
6693 * thread parking. This synchronizes the thread vs users,
6694 * the users are synchronized on the sqd->ctx_lock.
6696 if (io_sq_thread_should_park(sqd)) {
6697 io_sq_thread_parkme(sqd);
6700 if (unlikely(!list_empty(&sqd->ctx_new_list))) {
6701 io_sqd_init_new(sqd);
6702 timeout = jiffies + sqd->sq_thread_idle;
6704 if (fatal_signal_pending(current))
6707 cap_entries = !list_is_singular(&sqd->ctx_list);
6708 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6709 ret = __io_sq_thread(ctx, cap_entries);
6710 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
6714 if (sqt_spin || !time_after(jiffies, timeout)) {
6718 timeout = jiffies + sqd->sq_thread_idle;
6723 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
6724 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6725 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6726 !list_empty_careful(&ctx->iopoll_list)) {
6727 needs_sched = false;
6730 if (io_sqring_entries(ctx)) {
6731 needs_sched = false;
6736 if (needs_sched && !io_sq_thread_should_park(sqd)) {
6737 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6738 io_ring_set_wakeup_flag(ctx);
6741 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6742 io_ring_clear_wakeup_flag(ctx);
6745 finish_wait(&sqd->wait, &wait);
6746 timeout = jiffies + sqd->sq_thread_idle;
6749 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6750 io_uring_cancel_sqpoll(ctx);
6754 if (io_sq_thread_should_park(sqd))
6755 io_sq_thread_parkme(sqd);
6758 * Clear thread under lock so that concurrent parks work correctly
6760 complete(&sqd->completion);
6761 mutex_lock(&sqd->lock);
6763 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6765 io_ring_set_wakeup_flag(ctx);
6768 complete(&sqd->exited);
6769 mutex_unlock(&sqd->lock);
6773 struct io_wait_queue {
6774 struct wait_queue_entry wq;
6775 struct io_ring_ctx *ctx;
6777 unsigned nr_timeouts;
6780 static inline bool io_should_wake(struct io_wait_queue *iowq)
6782 struct io_ring_ctx *ctx = iowq->ctx;
6785 * Wake up if we have enough events, or if a timeout occurred since we
6786 * started waiting. For timeouts, we always want to return to userspace,
6787 * regardless of event count.
6789 return io_cqring_events(ctx) >= iowq->to_wait ||
6790 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6793 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6794 int wake_flags, void *key)
6796 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6800 * Cannot safely flush overflowed CQEs from here, ensure we wake up
6801 * the task, and the next invocation will do it.
6803 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->cq_check_overflow))
6804 return autoremove_wake_function(curr, mode, wake_flags, key);
6808 static int io_run_task_work_sig(void)
6810 if (io_run_task_work())
6812 if (!signal_pending(current))
6814 if (test_tsk_thread_flag(current, TIF_NOTIFY_SIGNAL))
6815 return -ERESTARTSYS;
6819 /* when returns >0, the caller should retry */
6820 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
6821 struct io_wait_queue *iowq,
6822 signed long *timeout)
6826 /* make sure we run task_work before checking for signals */
6827 ret = io_run_task_work_sig();
6828 if (ret || io_should_wake(iowq))
6830 /* let the caller flush overflows, retry */
6831 if (test_bit(0, &ctx->cq_check_overflow))
6834 *timeout = schedule_timeout(*timeout);
6835 return !*timeout ? -ETIME : 1;
6839 * Wait until events become available, if we don't already have some. The
6840 * application must reap them itself, as they reside on the shared cq ring.
6842 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6843 const sigset_t __user *sig, size_t sigsz,
6844 struct __kernel_timespec __user *uts)
6846 struct io_wait_queue iowq = {
6849 .func = io_wake_function,
6850 .entry = LIST_HEAD_INIT(iowq.wq.entry),
6853 .to_wait = min_events,
6855 struct io_rings *rings = ctx->rings;
6856 signed long timeout = MAX_SCHEDULE_TIMEOUT;
6860 io_cqring_overflow_flush(ctx, false, NULL, NULL);
6861 if (io_cqring_events(ctx) >= min_events)
6863 if (!io_run_task_work())
6868 #ifdef CONFIG_COMPAT
6869 if (in_compat_syscall())
6870 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
6874 ret = set_user_sigmask(sig, sigsz);
6881 struct timespec64 ts;
6883 if (get_timespec64(&ts, uts))
6885 timeout = timespec64_to_jiffies(&ts);
6888 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
6889 trace_io_uring_cqring_wait(ctx, min_events);
6891 io_cqring_overflow_flush(ctx, false, NULL, NULL);
6892 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
6893 TASK_INTERRUPTIBLE);
6894 ret = io_cqring_wait_schedule(ctx, &iowq, &timeout);
6895 finish_wait(&ctx->wait, &iowq.wq);
6898 restore_saved_sigmask_unless(ret == -EINTR);
6900 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
6903 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
6905 #if defined(CONFIG_UNIX)
6906 if (ctx->ring_sock) {
6907 struct sock *sock = ctx->ring_sock->sk;
6908 struct sk_buff *skb;
6910 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
6916 for (i = 0; i < ctx->nr_user_files; i++) {
6919 file = io_file_from_index(ctx, i);
6926 static void io_rsrc_data_ref_zero(struct percpu_ref *ref)
6928 struct fixed_rsrc_data *data;
6930 data = container_of(ref, struct fixed_rsrc_data, refs);
6931 complete(&data->done);
6934 static inline void io_rsrc_ref_lock(struct io_ring_ctx *ctx)
6936 spin_lock_bh(&ctx->rsrc_ref_lock);
6939 static inline void io_rsrc_ref_unlock(struct io_ring_ctx *ctx)
6941 spin_unlock_bh(&ctx->rsrc_ref_lock);
6944 static void io_sqe_rsrc_set_node(struct io_ring_ctx *ctx,
6945 struct fixed_rsrc_data *rsrc_data,
6946 struct fixed_rsrc_ref_node *ref_node)
6948 io_rsrc_ref_lock(ctx);
6949 rsrc_data->node = ref_node;
6950 list_add_tail(&ref_node->node, &ctx->rsrc_ref_list);
6951 io_rsrc_ref_unlock(ctx);
6952 percpu_ref_get(&rsrc_data->refs);
6955 static void io_sqe_rsrc_kill_node(struct io_ring_ctx *ctx, struct fixed_rsrc_data *data)
6957 struct fixed_rsrc_ref_node *ref_node = NULL;
6959 io_rsrc_ref_lock(ctx);
6960 ref_node = data->node;
6962 io_rsrc_ref_unlock(ctx);
6964 percpu_ref_kill(&ref_node->refs);
6967 static int io_rsrc_ref_quiesce(struct fixed_rsrc_data *data,
6968 struct io_ring_ctx *ctx,
6969 void (*rsrc_put)(struct io_ring_ctx *ctx,
6970 struct io_rsrc_put *prsrc))
6972 struct fixed_rsrc_ref_node *backup_node;
6978 data->quiesce = true;
6981 backup_node = alloc_fixed_rsrc_ref_node(ctx);
6984 backup_node->rsrc_data = data;
6985 backup_node->rsrc_put = rsrc_put;
6987 io_sqe_rsrc_kill_node(ctx, data);
6988 percpu_ref_kill(&data->refs);
6989 flush_delayed_work(&ctx->rsrc_put_work);
6991 ret = wait_for_completion_interruptible(&data->done);
6995 percpu_ref_resurrect(&data->refs);
6996 io_sqe_rsrc_set_node(ctx, data, backup_node);
6998 reinit_completion(&data->done);
6999 mutex_unlock(&ctx->uring_lock);
7000 ret = io_run_task_work_sig();
7001 mutex_lock(&ctx->uring_lock);
7003 data->quiesce = false;
7006 destroy_fixed_rsrc_ref_node(backup_node);
7010 static struct fixed_rsrc_data *alloc_fixed_rsrc_data(struct io_ring_ctx *ctx)
7012 struct fixed_rsrc_data *data;
7014 data = kzalloc(sizeof(*data), GFP_KERNEL);
7018 if (percpu_ref_init(&data->refs, io_rsrc_data_ref_zero,
7019 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) {
7024 init_completion(&data->done);
7028 static void free_fixed_rsrc_data(struct fixed_rsrc_data *data)
7030 percpu_ref_exit(&data->refs);
7035 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7037 struct fixed_rsrc_data *data = ctx->file_data;
7038 unsigned nr_tables, i;
7042 * percpu_ref_is_dying() is to stop parallel files unregister
7043 * Since we possibly drop uring lock later in this function to
7046 if (!data || percpu_ref_is_dying(&data->refs))
7048 ret = io_rsrc_ref_quiesce(data, ctx, io_ring_file_put);
7052 __io_sqe_files_unregister(ctx);
7053 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
7054 for (i = 0; i < nr_tables; i++)
7055 kfree(data->table[i].files);
7056 free_fixed_rsrc_data(data);
7057 ctx->file_data = NULL;
7058 ctx->nr_user_files = 0;
7062 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7063 __releases(&sqd->lock)
7067 if (sqd->thread == current)
7069 clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7070 wake_up_state(sqd->thread, TASK_PARKED);
7071 mutex_unlock(&sqd->lock);
7074 static bool io_sq_thread_park(struct io_sq_data *sqd)
7075 __acquires(&sqd->lock)
7077 if (sqd->thread == current)
7079 mutex_lock(&sqd->lock);
7081 mutex_unlock(&sqd->lock);
7084 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7085 wake_up_process(sqd->thread);
7086 wait_for_completion(&sqd->completion);
7090 static void io_sq_thread_stop(struct io_sq_data *sqd)
7092 if (test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state))
7094 mutex_lock(&sqd->lock);
7096 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7097 WARN_ON_ONCE(test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state));
7098 wake_up_process(sqd->thread);
7099 mutex_unlock(&sqd->lock);
7100 wait_for_completion(&sqd->exited);
7101 WARN_ON_ONCE(sqd->thread);
7103 mutex_unlock(&sqd->lock);
7107 static void io_put_sq_data(struct io_sq_data *sqd)
7109 if (refcount_dec_and_test(&sqd->refs)) {
7110 io_sq_thread_stop(sqd);
7115 static void io_sq_thread_finish(struct io_ring_ctx *ctx)
7117 struct io_sq_data *sqd = ctx->sq_data;
7120 complete(&sqd->startup);
7122 wait_for_completion(&ctx->sq_thread_comp);
7123 io_sq_thread_park(sqd);
7126 mutex_lock(&sqd->ctx_lock);
7127 list_del(&ctx->sqd_list);
7128 io_sqd_update_thread_idle(sqd);
7129 mutex_unlock(&sqd->ctx_lock);
7132 io_sq_thread_unpark(sqd);
7134 io_put_sq_data(sqd);
7135 ctx->sq_data = NULL;
7139 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7141 struct io_ring_ctx *ctx_attach;
7142 struct io_sq_data *sqd;
7145 f = fdget(p->wq_fd);
7147 return ERR_PTR(-ENXIO);
7148 if (f.file->f_op != &io_uring_fops) {
7150 return ERR_PTR(-EINVAL);
7153 ctx_attach = f.file->private_data;
7154 sqd = ctx_attach->sq_data;
7157 return ERR_PTR(-EINVAL);
7160 refcount_inc(&sqd->refs);
7165 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p)
7167 struct io_sq_data *sqd;
7169 if (p->flags & IORING_SETUP_ATTACH_WQ)
7170 return io_attach_sq_data(p);
7172 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7174 return ERR_PTR(-ENOMEM);
7176 refcount_set(&sqd->refs, 1);
7177 INIT_LIST_HEAD(&sqd->ctx_list);
7178 INIT_LIST_HEAD(&sqd->ctx_new_list);
7179 mutex_init(&sqd->ctx_lock);
7180 mutex_init(&sqd->lock);
7181 init_waitqueue_head(&sqd->wait);
7182 init_completion(&sqd->startup);
7183 init_completion(&sqd->completion);
7184 init_completion(&sqd->exited);
7188 #if defined(CONFIG_UNIX)
7190 * Ensure the UNIX gc is aware of our file set, so we are certain that
7191 * the io_uring can be safely unregistered on process exit, even if we have
7192 * loops in the file referencing.
7194 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7196 struct sock *sk = ctx->ring_sock->sk;
7197 struct scm_fp_list *fpl;
7198 struct sk_buff *skb;
7201 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7205 skb = alloc_skb(0, GFP_KERNEL);
7214 fpl->user = get_uid(current_user());
7215 for (i = 0; i < nr; i++) {
7216 struct file *file = io_file_from_index(ctx, i + offset);
7220 fpl->fp[nr_files] = get_file(file);
7221 unix_inflight(fpl->user, fpl->fp[nr_files]);
7226 fpl->max = SCM_MAX_FD;
7227 fpl->count = nr_files;
7228 UNIXCB(skb).fp = fpl;
7229 skb->destructor = unix_destruct_scm;
7230 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7231 skb_queue_head(&sk->sk_receive_queue, skb);
7233 for (i = 0; i < nr_files; i++)
7244 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7245 * causes regular reference counting to break down. We rely on the UNIX
7246 * garbage collection to take care of this problem for us.
7248 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7250 unsigned left, total;
7254 left = ctx->nr_user_files;
7256 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7258 ret = __io_sqe_files_scm(ctx, this_files, total);
7262 total += this_files;
7268 while (total < ctx->nr_user_files) {
7269 struct file *file = io_file_from_index(ctx, total);
7279 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7285 static int io_sqe_alloc_file_tables(struct fixed_rsrc_data *file_data,
7286 unsigned nr_tables, unsigned nr_files)
7290 for (i = 0; i < nr_tables; i++) {
7291 struct fixed_rsrc_table *table = &file_data->table[i];
7292 unsigned this_files;
7294 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
7295 table->files = kcalloc(this_files, sizeof(struct file *),
7299 nr_files -= this_files;
7305 for (i = 0; i < nr_tables; i++) {
7306 struct fixed_rsrc_table *table = &file_data->table[i];
7307 kfree(table->files);
7312 static void io_ring_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
7314 struct file *file = prsrc->file;
7315 #if defined(CONFIG_UNIX)
7316 struct sock *sock = ctx->ring_sock->sk;
7317 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7318 struct sk_buff *skb;
7321 __skb_queue_head_init(&list);
7324 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7325 * remove this entry and rearrange the file array.
7327 skb = skb_dequeue(head);
7329 struct scm_fp_list *fp;
7331 fp = UNIXCB(skb).fp;
7332 for (i = 0; i < fp->count; i++) {
7335 if (fp->fp[i] != file)
7338 unix_notinflight(fp->user, fp->fp[i]);
7339 left = fp->count - 1 - i;
7341 memmove(&fp->fp[i], &fp->fp[i + 1],
7342 left * sizeof(struct file *));
7349 __skb_queue_tail(&list, skb);
7359 __skb_queue_tail(&list, skb);
7361 skb = skb_dequeue(head);
7364 if (skb_peek(&list)) {
7365 spin_lock_irq(&head->lock);
7366 while ((skb = __skb_dequeue(&list)) != NULL)
7367 __skb_queue_tail(head, skb);
7368 spin_unlock_irq(&head->lock);
7375 static void __io_rsrc_put_work(struct fixed_rsrc_ref_node *ref_node)
7377 struct fixed_rsrc_data *rsrc_data = ref_node->rsrc_data;
7378 struct io_ring_ctx *ctx = rsrc_data->ctx;
7379 struct io_rsrc_put *prsrc, *tmp;
7381 list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
7382 list_del(&prsrc->list);
7383 ref_node->rsrc_put(ctx, prsrc);
7387 percpu_ref_exit(&ref_node->refs);
7389 percpu_ref_put(&rsrc_data->refs);
7392 static void io_rsrc_put_work(struct work_struct *work)
7394 struct io_ring_ctx *ctx;
7395 struct llist_node *node;
7397 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
7398 node = llist_del_all(&ctx->rsrc_put_llist);
7401 struct fixed_rsrc_ref_node *ref_node;
7402 struct llist_node *next = node->next;
7404 ref_node = llist_entry(node, struct fixed_rsrc_ref_node, llist);
7405 __io_rsrc_put_work(ref_node);
7410 static struct file **io_fixed_file_slot(struct fixed_rsrc_data *file_data,
7413 struct fixed_rsrc_table *table;
7415 table = &file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7416 return &table->files[i & IORING_FILE_TABLE_MASK];
7419 static void io_rsrc_node_ref_zero(struct percpu_ref *ref)
7421 struct fixed_rsrc_ref_node *ref_node;
7422 struct fixed_rsrc_data *data;
7423 struct io_ring_ctx *ctx;
7424 bool first_add = false;
7427 ref_node = container_of(ref, struct fixed_rsrc_ref_node, refs);
7428 data = ref_node->rsrc_data;
7431 io_rsrc_ref_lock(ctx);
7432 ref_node->done = true;
7434 while (!list_empty(&ctx->rsrc_ref_list)) {
7435 ref_node = list_first_entry(&ctx->rsrc_ref_list,
7436 struct fixed_rsrc_ref_node, node);
7437 /* recycle ref nodes in order */
7438 if (!ref_node->done)
7440 list_del(&ref_node->node);
7441 first_add |= llist_add(&ref_node->llist, &ctx->rsrc_put_llist);
7443 io_rsrc_ref_unlock(ctx);
7445 if (percpu_ref_is_dying(&data->refs))
7449 mod_delayed_work(system_wq, &ctx->rsrc_put_work, 0);
7451 queue_delayed_work(system_wq, &ctx->rsrc_put_work, delay);
7454 static struct fixed_rsrc_ref_node *alloc_fixed_rsrc_ref_node(
7455 struct io_ring_ctx *ctx)
7457 struct fixed_rsrc_ref_node *ref_node;
7459 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7463 if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
7468 INIT_LIST_HEAD(&ref_node->node);
7469 INIT_LIST_HEAD(&ref_node->rsrc_list);
7470 ref_node->done = false;
7474 static void init_fixed_file_ref_node(struct io_ring_ctx *ctx,
7475 struct fixed_rsrc_ref_node *ref_node)
7477 ref_node->rsrc_data = ctx->file_data;
7478 ref_node->rsrc_put = io_ring_file_put;
7481 static void destroy_fixed_rsrc_ref_node(struct fixed_rsrc_ref_node *ref_node)
7483 percpu_ref_exit(&ref_node->refs);
7488 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7491 __s32 __user *fds = (__s32 __user *) arg;
7492 unsigned nr_tables, i;
7494 int fd, ret = -ENOMEM;
7495 struct fixed_rsrc_ref_node *ref_node;
7496 struct fixed_rsrc_data *file_data;
7502 if (nr_args > IORING_MAX_FIXED_FILES)
7505 file_data = alloc_fixed_rsrc_data(ctx);
7508 ctx->file_data = file_data;
7510 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
7511 file_data->table = kcalloc(nr_tables, sizeof(*file_data->table),
7513 if (!file_data->table)
7516 if (io_sqe_alloc_file_tables(file_data, nr_tables, nr_args))
7519 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7520 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7524 /* allow sparse sets */
7534 * Don't allow io_uring instances to be registered. If UNIX
7535 * isn't enabled, then this causes a reference cycle and this
7536 * instance can never get freed. If UNIX is enabled we'll
7537 * handle it just fine, but there's still no point in allowing
7538 * a ring fd as it doesn't support regular read/write anyway.
7540 if (file->f_op == &io_uring_fops) {
7544 *io_fixed_file_slot(file_data, i) = file;
7547 ret = io_sqe_files_scm(ctx);
7549 io_sqe_files_unregister(ctx);
7553 ref_node = alloc_fixed_rsrc_ref_node(ctx);
7555 io_sqe_files_unregister(ctx);
7558 init_fixed_file_ref_node(ctx, ref_node);
7560 io_sqe_rsrc_set_node(ctx, file_data, ref_node);
7563 for (i = 0; i < ctx->nr_user_files; i++) {
7564 file = io_file_from_index(ctx, i);
7568 for (i = 0; i < nr_tables; i++)
7569 kfree(file_data->table[i].files);
7570 ctx->nr_user_files = 0;
7572 free_fixed_rsrc_data(ctx->file_data);
7573 ctx->file_data = NULL;
7577 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7580 #if defined(CONFIG_UNIX)
7581 struct sock *sock = ctx->ring_sock->sk;
7582 struct sk_buff_head *head = &sock->sk_receive_queue;
7583 struct sk_buff *skb;
7586 * See if we can merge this file into an existing skb SCM_RIGHTS
7587 * file set. If there's no room, fall back to allocating a new skb
7588 * and filling it in.
7590 spin_lock_irq(&head->lock);
7591 skb = skb_peek(head);
7593 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7595 if (fpl->count < SCM_MAX_FD) {
7596 __skb_unlink(skb, head);
7597 spin_unlock_irq(&head->lock);
7598 fpl->fp[fpl->count] = get_file(file);
7599 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7601 spin_lock_irq(&head->lock);
7602 __skb_queue_head(head, skb);
7607 spin_unlock_irq(&head->lock);
7614 return __io_sqe_files_scm(ctx, 1, index);
7620 static int io_queue_rsrc_removal(struct fixed_rsrc_data *data, void *rsrc)
7622 struct io_rsrc_put *prsrc;
7623 struct fixed_rsrc_ref_node *ref_node = data->node;
7625 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
7630 list_add(&prsrc->list, &ref_node->rsrc_list);
7635 static inline int io_queue_file_removal(struct fixed_rsrc_data *data,
7638 return io_queue_rsrc_removal(data, (void *)file);
7641 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7642 struct io_uring_rsrc_update *up,
7645 struct fixed_rsrc_data *data = ctx->file_data;
7646 struct fixed_rsrc_ref_node *ref_node;
7647 struct file *file, **file_slot;
7651 bool needs_switch = false;
7653 if (check_add_overflow(up->offset, nr_args, &done))
7655 if (done > ctx->nr_user_files)
7658 ref_node = alloc_fixed_rsrc_ref_node(ctx);
7661 init_fixed_file_ref_node(ctx, ref_node);
7663 fds = u64_to_user_ptr(up->data);
7664 for (done = 0; done < nr_args; done++) {
7666 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
7670 if (fd == IORING_REGISTER_FILES_SKIP)
7673 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
7674 file_slot = io_fixed_file_slot(ctx->file_data, i);
7677 err = io_queue_file_removal(data, *file_slot);
7681 needs_switch = true;
7690 * Don't allow io_uring instances to be registered. If
7691 * UNIX isn't enabled, then this causes a reference
7692 * cycle and this instance can never get freed. If UNIX
7693 * is enabled we'll handle it just fine, but there's
7694 * still no point in allowing a ring fd as it doesn't
7695 * support regular read/write anyway.
7697 if (file->f_op == &io_uring_fops) {
7703 err = io_sqe_file_register(ctx, file, i);
7713 percpu_ref_kill(&data->node->refs);
7714 io_sqe_rsrc_set_node(ctx, data, ref_node);
7716 destroy_fixed_rsrc_ref_node(ref_node);
7718 return done ? done : err;
7721 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
7724 struct io_uring_rsrc_update up;
7726 if (!ctx->file_data)
7730 if (copy_from_user(&up, arg, sizeof(up)))
7735 return __io_sqe_files_update(ctx, &up, nr_args);
7738 static struct io_wq_work *io_free_work(struct io_wq_work *work)
7740 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7742 req = io_put_req_find_next(req);
7743 return req ? &req->work : NULL;
7746 static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx)
7748 struct io_wq_hash *hash;
7749 struct io_wq_data data;
7750 unsigned int concurrency;
7752 hash = ctx->hash_map;
7754 hash = kzalloc(sizeof(*hash), GFP_KERNEL);
7756 return ERR_PTR(-ENOMEM);
7757 refcount_set(&hash->refs, 1);
7758 init_waitqueue_head(&hash->wait);
7759 ctx->hash_map = hash;
7763 data.free_work = io_free_work;
7764 data.do_work = io_wq_submit_work;
7766 /* Do QD, or 4 * CPUS, whatever is smallest */
7767 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7769 return io_wq_create(concurrency, &data);
7772 static int io_uring_alloc_task_context(struct task_struct *task,
7773 struct io_ring_ctx *ctx)
7775 struct io_uring_task *tctx;
7778 tctx = kmalloc(sizeof(*tctx), GFP_KERNEL);
7779 if (unlikely(!tctx))
7782 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
7783 if (unlikely(ret)) {
7788 tctx->io_wq = io_init_wq_offload(ctx);
7789 if (IS_ERR(tctx->io_wq)) {
7790 ret = PTR_ERR(tctx->io_wq);
7791 percpu_counter_destroy(&tctx->inflight);
7797 init_waitqueue_head(&tctx->wait);
7799 atomic_set(&tctx->in_idle, 0);
7800 tctx->sqpoll = false;
7801 task->io_uring = tctx;
7802 spin_lock_init(&tctx->task_lock);
7803 INIT_WQ_LIST(&tctx->task_list);
7804 tctx->task_state = 0;
7805 init_task_work(&tctx->task_work, tctx_task_work);
7809 void __io_uring_free(struct task_struct *tsk)
7811 struct io_uring_task *tctx = tsk->io_uring;
7813 WARN_ON_ONCE(!xa_empty(&tctx->xa));
7814 WARN_ON_ONCE(tctx->io_wq);
7816 percpu_counter_destroy(&tctx->inflight);
7818 tsk->io_uring = NULL;
7821 static int io_sq_thread_fork(struct io_sq_data *sqd, struct io_ring_ctx *ctx)
7825 clear_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7826 reinit_completion(&sqd->completion);
7828 sqd->task_pid = current->pid;
7829 current->flags |= PF_IO_WORKER;
7830 ret = io_wq_fork_thread(io_sq_thread, sqd);
7831 current->flags &= ~PF_IO_WORKER;
7836 wait_for_completion(&sqd->completion);
7837 return io_uring_alloc_task_context(sqd->thread, ctx);
7840 static int io_sq_offload_create(struct io_ring_ctx *ctx,
7841 struct io_uring_params *p)
7845 /* Retain compatibility with failing for an invalid attach attempt */
7846 if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
7847 IORING_SETUP_ATTACH_WQ) {
7850 f = fdget(p->wq_fd);
7853 if (f.file->f_op != &io_uring_fops) {
7859 if (ctx->flags & IORING_SETUP_SQPOLL) {
7860 struct io_sq_data *sqd;
7863 if (!capable(CAP_SYS_ADMIN) && !capable(CAP_SYS_NICE))
7866 sqd = io_get_sq_data(p);
7873 io_sq_thread_park(sqd);
7874 mutex_lock(&sqd->ctx_lock);
7875 list_add(&ctx->sqd_list, &sqd->ctx_new_list);
7876 mutex_unlock(&sqd->ctx_lock);
7877 io_sq_thread_unpark(sqd);
7879 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
7880 if (!ctx->sq_thread_idle)
7881 ctx->sq_thread_idle = HZ;
7886 if (p->flags & IORING_SETUP_SQ_AFF) {
7887 int cpu = p->sq_thread_cpu;
7890 if (cpu >= nr_cpu_ids)
7892 if (!cpu_online(cpu))
7900 sqd->task_pid = current->pid;
7901 current->flags |= PF_IO_WORKER;
7902 ret = io_wq_fork_thread(io_sq_thread, sqd);
7903 current->flags &= ~PF_IO_WORKER;
7908 wait_for_completion(&sqd->completion);
7909 ret = io_uring_alloc_task_context(sqd->thread, ctx);
7912 } else if (p->flags & IORING_SETUP_SQ_AFF) {
7913 /* Can't have SQ_AFF without SQPOLL */
7920 io_sq_thread_finish(ctx);
7924 static void io_sq_offload_start(struct io_ring_ctx *ctx)
7926 struct io_sq_data *sqd = ctx->sq_data;
7928 ctx->flags &= ~IORING_SETUP_R_DISABLED;
7929 if (ctx->flags & IORING_SETUP_SQPOLL)
7930 complete(&sqd->startup);
7933 static inline void __io_unaccount_mem(struct user_struct *user,
7934 unsigned long nr_pages)
7936 atomic_long_sub(nr_pages, &user->locked_vm);
7939 static inline int __io_account_mem(struct user_struct *user,
7940 unsigned long nr_pages)
7942 unsigned long page_limit, cur_pages, new_pages;
7944 /* Don't allow more pages than we can safely lock */
7945 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
7948 cur_pages = atomic_long_read(&user->locked_vm);
7949 new_pages = cur_pages + nr_pages;
7950 if (new_pages > page_limit)
7952 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
7953 new_pages) != cur_pages);
7958 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
7961 __io_unaccount_mem(ctx->user, nr_pages);
7963 if (ctx->mm_account)
7964 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
7967 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
7972 ret = __io_account_mem(ctx->user, nr_pages);
7977 if (ctx->mm_account)
7978 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
7983 static void io_mem_free(void *ptr)
7990 page = virt_to_head_page(ptr);
7991 if (put_page_testzero(page))
7992 free_compound_page(page);
7995 static void *io_mem_alloc(size_t size)
7997 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
7998 __GFP_NORETRY | __GFP_ACCOUNT;
8000 return (void *) __get_free_pages(gfp_flags, get_order(size));
8003 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8006 struct io_rings *rings;
8007 size_t off, sq_array_size;
8009 off = struct_size(rings, cqes, cq_entries);
8010 if (off == SIZE_MAX)
8014 off = ALIGN(off, SMP_CACHE_BYTES);
8022 sq_array_size = array_size(sizeof(u32), sq_entries);
8023 if (sq_array_size == SIZE_MAX)
8026 if (check_add_overflow(off, sq_array_size, &off))
8032 static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8036 if (!ctx->user_bufs)
8039 for (i = 0; i < ctx->nr_user_bufs; i++) {
8040 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8042 for (j = 0; j < imu->nr_bvecs; j++)
8043 unpin_user_page(imu->bvec[j].bv_page);
8045 if (imu->acct_pages)
8046 io_unaccount_mem(ctx, imu->acct_pages);
8051 kfree(ctx->user_bufs);
8052 ctx->user_bufs = NULL;
8053 ctx->nr_user_bufs = 0;
8057 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8058 void __user *arg, unsigned index)
8060 struct iovec __user *src;
8062 #ifdef CONFIG_COMPAT
8064 struct compat_iovec __user *ciovs;
8065 struct compat_iovec ciov;
8067 ciovs = (struct compat_iovec __user *) arg;
8068 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8071 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8072 dst->iov_len = ciov.iov_len;
8076 src = (struct iovec __user *) arg;
8077 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8083 * Not super efficient, but this is just a registration time. And we do cache
8084 * the last compound head, so generally we'll only do a full search if we don't
8087 * We check if the given compound head page has already been accounted, to
8088 * avoid double accounting it. This allows us to account the full size of the
8089 * page, not just the constituent pages of a huge page.
8091 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8092 int nr_pages, struct page *hpage)
8096 /* check current page array */
8097 for (i = 0; i < nr_pages; i++) {
8098 if (!PageCompound(pages[i]))
8100 if (compound_head(pages[i]) == hpage)
8104 /* check previously registered pages */
8105 for (i = 0; i < ctx->nr_user_bufs; i++) {
8106 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8108 for (j = 0; j < imu->nr_bvecs; j++) {
8109 if (!PageCompound(imu->bvec[j].bv_page))
8111 if (compound_head(imu->bvec[j].bv_page) == hpage)
8119 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8120 int nr_pages, struct io_mapped_ubuf *imu,
8121 struct page **last_hpage)
8125 for (i = 0; i < nr_pages; i++) {
8126 if (!PageCompound(pages[i])) {
8131 hpage = compound_head(pages[i]);
8132 if (hpage == *last_hpage)
8134 *last_hpage = hpage;
8135 if (headpage_already_acct(ctx, pages, i, hpage))
8137 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8141 if (!imu->acct_pages)
8144 ret = io_account_mem(ctx, imu->acct_pages);
8146 imu->acct_pages = 0;
8150 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
8151 struct io_mapped_ubuf *imu,
8152 struct page **last_hpage)
8154 struct vm_area_struct **vmas = NULL;
8155 struct page **pages = NULL;
8156 unsigned long off, start, end, ubuf;
8158 int ret, pret, nr_pages, i;
8160 ubuf = (unsigned long) iov->iov_base;
8161 end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8162 start = ubuf >> PAGE_SHIFT;
8163 nr_pages = end - start;
8167 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
8171 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
8176 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
8182 mmap_read_lock(current->mm);
8183 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
8185 if (pret == nr_pages) {
8186 /* don't support file backed memory */
8187 for (i = 0; i < nr_pages; i++) {
8188 struct vm_area_struct *vma = vmas[i];
8191 !is_file_hugepages(vma->vm_file)) {
8197 ret = pret < 0 ? pret : -EFAULT;
8199 mmap_read_unlock(current->mm);
8202 * if we did partial map, or found file backed vmas,
8203 * release any pages we did get
8206 unpin_user_pages(pages, pret);
8211 ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
8213 unpin_user_pages(pages, pret);
8218 off = ubuf & ~PAGE_MASK;
8219 size = iov->iov_len;
8220 for (i = 0; i < nr_pages; i++) {
8223 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8224 imu->bvec[i].bv_page = pages[i];
8225 imu->bvec[i].bv_len = vec_len;
8226 imu->bvec[i].bv_offset = off;
8230 /* store original address for later verification */
8232 imu->len = iov->iov_len;
8233 imu->nr_bvecs = nr_pages;
8241 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
8245 if (!nr_args || nr_args > UIO_MAXIOV)
8248 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
8250 if (!ctx->user_bufs)
8256 static int io_buffer_validate(struct iovec *iov)
8259 * Don't impose further limits on the size and buffer
8260 * constraints here, we'll -EINVAL later when IO is
8261 * submitted if they are wrong.
8263 if (!iov->iov_base || !iov->iov_len)
8266 /* arbitrary limit, but we need something */
8267 if (iov->iov_len > SZ_1G)
8273 static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
8274 unsigned int nr_args)
8278 struct page *last_hpage = NULL;
8280 ret = io_buffers_map_alloc(ctx, nr_args);
8284 for (i = 0; i < nr_args; i++) {
8285 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8287 ret = io_copy_iov(ctx, &iov, arg, i);
8291 ret = io_buffer_validate(&iov);
8295 ret = io_sqe_buffer_register(ctx, &iov, imu, &last_hpage);
8299 ctx->nr_user_bufs++;
8303 io_sqe_buffers_unregister(ctx);
8308 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8310 __s32 __user *fds = arg;
8316 if (copy_from_user(&fd, fds, sizeof(*fds)))
8319 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8320 if (IS_ERR(ctx->cq_ev_fd)) {
8321 int ret = PTR_ERR(ctx->cq_ev_fd);
8322 ctx->cq_ev_fd = NULL;
8329 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8331 if (ctx->cq_ev_fd) {
8332 eventfd_ctx_put(ctx->cq_ev_fd);
8333 ctx->cq_ev_fd = NULL;
8340 static int __io_destroy_buffers(int id, void *p, void *data)
8342 struct io_ring_ctx *ctx = data;
8343 struct io_buffer *buf = p;
8345 __io_remove_buffers(ctx, buf, id, -1U);
8349 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8351 idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx);
8352 idr_destroy(&ctx->io_buffer_idr);
8355 static void io_req_cache_free(struct list_head *list, struct task_struct *tsk)
8357 struct io_kiocb *req, *nxt;
8359 list_for_each_entry_safe(req, nxt, list, compl.list) {
8360 if (tsk && req->task != tsk)
8362 list_del(&req->compl.list);
8363 kmem_cache_free(req_cachep, req);
8367 static void io_req_caches_free(struct io_ring_ctx *ctx)
8369 struct io_submit_state *submit_state = &ctx->submit_state;
8370 struct io_comp_state *cs = &ctx->submit_state.comp;
8372 mutex_lock(&ctx->uring_lock);
8374 if (submit_state->free_reqs) {
8375 kmem_cache_free_bulk(req_cachep, submit_state->free_reqs,
8376 submit_state->reqs);
8377 submit_state->free_reqs = 0;
8380 spin_lock_irq(&ctx->completion_lock);
8381 list_splice_init(&cs->locked_free_list, &cs->free_list);
8382 cs->locked_free_nr = 0;
8383 spin_unlock_irq(&ctx->completion_lock);
8385 io_req_cache_free(&cs->free_list, NULL);
8387 mutex_unlock(&ctx->uring_lock);
8390 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8393 * Some may use context even when all refs and requests have been put,
8394 * and they are free to do so while still holding uring_lock, see
8395 * __io_req_task_submit(). Wait for them to finish.
8397 mutex_lock(&ctx->uring_lock);
8398 mutex_unlock(&ctx->uring_lock);
8400 io_sq_thread_finish(ctx);
8401 io_sqe_buffers_unregister(ctx);
8403 if (ctx->mm_account) {
8404 mmdrop(ctx->mm_account);
8405 ctx->mm_account = NULL;
8408 mutex_lock(&ctx->uring_lock);
8409 io_sqe_files_unregister(ctx);
8410 mutex_unlock(&ctx->uring_lock);
8411 io_eventfd_unregister(ctx);
8412 io_destroy_buffers(ctx);
8413 idr_destroy(&ctx->personality_idr);
8415 #if defined(CONFIG_UNIX)
8416 if (ctx->ring_sock) {
8417 ctx->ring_sock->file = NULL; /* so that iput() is called */
8418 sock_release(ctx->ring_sock);
8422 io_mem_free(ctx->rings);
8423 io_mem_free(ctx->sq_sqes);
8425 percpu_ref_exit(&ctx->refs);
8426 free_uid(ctx->user);
8427 io_req_caches_free(ctx);
8429 io_wq_put_hash(ctx->hash_map);
8430 kfree(ctx->cancel_hash);
8434 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8436 struct io_ring_ctx *ctx = file->private_data;
8439 poll_wait(file, &ctx->cq_wait, wait);
8441 * synchronizes with barrier from wq_has_sleeper call in
8445 if (!io_sqring_full(ctx))
8446 mask |= EPOLLOUT | EPOLLWRNORM;
8449 * Don't flush cqring overflow list here, just do a simple check.
8450 * Otherwise there could possible be ABBA deadlock:
8453 * lock(&ctx->uring_lock);
8455 * lock(&ctx->uring_lock);
8458 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
8459 * pushs them to do the flush.
8461 if (io_cqring_events(ctx) || test_bit(0, &ctx->cq_check_overflow))
8462 mask |= EPOLLIN | EPOLLRDNORM;
8467 static int io_uring_fasync(int fd, struct file *file, int on)
8469 struct io_ring_ctx *ctx = file->private_data;
8471 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8474 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8476 const struct cred *creds;
8478 creds = idr_remove(&ctx->personality_idr, id);
8487 static int io_remove_personalities(int id, void *p, void *data)
8489 struct io_ring_ctx *ctx = data;
8491 io_unregister_personality(ctx, id);
8495 static bool io_run_ctx_fallback(struct io_ring_ctx *ctx)
8497 struct callback_head *work, *next;
8498 bool executed = false;
8501 work = xchg(&ctx->exit_task_work, NULL);
8517 static void io_ring_exit_work(struct work_struct *work)
8519 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
8523 * If we're doing polled IO and end up having requests being
8524 * submitted async (out-of-line), then completions can come in while
8525 * we're waiting for refs to drop. We need to reap these manually,
8526 * as nobody else will be looking for them.
8529 io_uring_try_cancel_requests(ctx, NULL, NULL);
8530 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
8531 io_ring_ctx_free(ctx);
8534 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8536 mutex_lock(&ctx->uring_lock);
8537 percpu_ref_kill(&ctx->refs);
8538 /* if force is set, the ring is going away. always drop after that */
8539 ctx->cq_overflow_flushed = 1;
8541 __io_cqring_overflow_flush(ctx, true, NULL, NULL);
8542 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
8543 mutex_unlock(&ctx->uring_lock);
8545 io_kill_timeouts(ctx, NULL, NULL);
8546 io_poll_remove_all(ctx, NULL, NULL);
8548 /* if we failed setting up the ctx, we might not have any rings */
8549 io_iopoll_try_reap_events(ctx);
8551 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8553 * Use system_unbound_wq to avoid spawning tons of event kworkers
8554 * if we're exiting a ton of rings at the same time. It just adds
8555 * noise and overhead, there's no discernable change in runtime
8556 * over using system_wq.
8558 queue_work(system_unbound_wq, &ctx->exit_work);
8561 static int io_uring_release(struct inode *inode, struct file *file)
8563 struct io_ring_ctx *ctx = file->private_data;
8565 file->private_data = NULL;
8566 io_ring_ctx_wait_and_kill(ctx);
8570 struct io_task_cancel {
8571 struct task_struct *task;
8572 struct files_struct *files;
8575 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8577 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8578 struct io_task_cancel *cancel = data;
8581 if (cancel->files && (req->flags & REQ_F_LINK_TIMEOUT)) {
8582 unsigned long flags;
8583 struct io_ring_ctx *ctx = req->ctx;
8585 /* protect against races with linked timeouts */
8586 spin_lock_irqsave(&ctx->completion_lock, flags);
8587 ret = io_match_task(req, cancel->task, cancel->files);
8588 spin_unlock_irqrestore(&ctx->completion_lock, flags);
8590 ret = io_match_task(req, cancel->task, cancel->files);
8595 static void io_cancel_defer_files(struct io_ring_ctx *ctx,
8596 struct task_struct *task,
8597 struct files_struct *files)
8599 struct io_defer_entry *de = NULL;
8602 spin_lock_irq(&ctx->completion_lock);
8603 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8604 if (io_match_task(de->req, task, files)) {
8605 list_cut_position(&list, &ctx->defer_list, &de->list);
8609 spin_unlock_irq(&ctx->completion_lock);
8611 while (!list_empty(&list)) {
8612 de = list_first_entry(&list, struct io_defer_entry, list);
8613 list_del_init(&de->list);
8614 req_set_fail_links(de->req);
8615 io_put_req(de->req);
8616 io_req_complete(de->req, -ECANCELED);
8621 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
8622 struct task_struct *task,
8623 struct files_struct *files)
8625 struct io_task_cancel cancel = { .task = task, .files = files, };
8626 struct task_struct *tctx_task = task ?: current;
8627 struct io_uring_task *tctx = tctx_task->io_uring;
8630 enum io_wq_cancel cret;
8633 if (tctx && tctx->io_wq) {
8634 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
8636 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8639 /* SQPOLL thread does its own polling */
8640 if (!(ctx->flags & IORING_SETUP_SQPOLL) && !files) {
8641 while (!list_empty_careful(&ctx->iopoll_list)) {
8642 io_iopoll_try_reap_events(ctx);
8647 ret |= io_poll_remove_all(ctx, task, files);
8648 ret |= io_kill_timeouts(ctx, task, files);
8649 ret |= io_run_task_work();
8650 ret |= io_run_ctx_fallback(ctx);
8651 io_cqring_overflow_flush(ctx, true, task, files);
8658 static int io_uring_count_inflight(struct io_ring_ctx *ctx,
8659 struct task_struct *task,
8660 struct files_struct *files)
8662 struct io_kiocb *req;
8665 spin_lock_irq(&ctx->inflight_lock);
8666 list_for_each_entry(req, &ctx->inflight_list, inflight_entry)
8667 cnt += io_match_task(req, task, files);
8668 spin_unlock_irq(&ctx->inflight_lock);
8672 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
8673 struct task_struct *task,
8674 struct files_struct *files)
8676 while (!list_empty_careful(&ctx->inflight_list)) {
8680 inflight = io_uring_count_inflight(ctx, task, files);
8684 io_uring_try_cancel_requests(ctx, task, files);
8687 io_sq_thread_unpark(ctx->sq_data);
8688 prepare_to_wait(&task->io_uring->wait, &wait,
8689 TASK_UNINTERRUPTIBLE);
8690 if (inflight == io_uring_count_inflight(ctx, task, files))
8692 finish_wait(&task->io_uring->wait, &wait);
8694 io_sq_thread_park(ctx->sq_data);
8699 * We need to iteratively cancel requests, in case a request has dependent
8700 * hard links. These persist even for failure of cancelations, hence keep
8701 * looping until none are found.
8703 static void io_uring_cancel_task_requests(struct io_ring_ctx *ctx,
8704 struct files_struct *files)
8706 struct task_struct *task = current;
8707 bool did_park = false;
8709 if ((ctx->flags & IORING_SETUP_SQPOLL) && ctx->sq_data) {
8710 /* never started, nothing to cancel */
8711 if (ctx->flags & IORING_SETUP_R_DISABLED) {
8712 io_sq_offload_start(ctx);
8715 did_park = io_sq_thread_park(ctx->sq_data);
8717 task = ctx->sq_data->thread;
8718 atomic_inc(&task->io_uring->in_idle);
8722 io_cancel_defer_files(ctx, task, files);
8724 io_uring_cancel_files(ctx, task, files);
8726 io_uring_try_cancel_requests(ctx, task, NULL);
8729 atomic_dec(&task->io_uring->in_idle);
8730 io_sq_thread_unpark(ctx->sq_data);
8735 * Note that this task has used io_uring. We use it for cancelation purposes.
8737 static int io_uring_add_task_file(struct io_ring_ctx *ctx, struct file *file)
8739 struct io_uring_task *tctx = current->io_uring;
8742 if (unlikely(!tctx)) {
8743 ret = io_uring_alloc_task_context(current, ctx);
8746 tctx = current->io_uring;
8748 if (tctx->last != file) {
8749 void *old = xa_load(&tctx->xa, (unsigned long)file);
8753 ret = xa_err(xa_store(&tctx->xa, (unsigned long)file,
8764 * This is race safe in that the task itself is doing this, hence it
8765 * cannot be going through the exit/cancel paths at the same time.
8766 * This cannot be modified while exit/cancel is running.
8768 if (!tctx->sqpoll && (ctx->flags & IORING_SETUP_SQPOLL))
8769 tctx->sqpoll = true;
8775 * Remove this io_uring_file -> task mapping.
8777 static void io_uring_del_task_file(struct file *file)
8779 struct io_uring_task *tctx = current->io_uring;
8781 if (tctx->last == file)
8783 file = xa_erase(&tctx->xa, (unsigned long)file);
8788 static void io_uring_clean_tctx(struct io_uring_task *tctx)
8791 unsigned long index;
8793 xa_for_each(&tctx->xa, index, file)
8794 io_uring_del_task_file(file);
8796 io_wq_put_and_exit(tctx->io_wq);
8801 void __io_uring_files_cancel(struct files_struct *files)
8803 struct io_uring_task *tctx = current->io_uring;
8805 unsigned long index;
8807 /* make sure overflow events are dropped */
8808 atomic_inc(&tctx->in_idle);
8809 xa_for_each(&tctx->xa, index, file)
8810 io_uring_cancel_task_requests(file->private_data, files);
8811 atomic_dec(&tctx->in_idle);
8814 io_uring_clean_tctx(tctx);
8817 static s64 tctx_inflight(struct io_uring_task *tctx)
8819 return percpu_counter_sum(&tctx->inflight);
8822 static void io_uring_cancel_sqpoll(struct io_ring_ctx *ctx)
8824 struct io_sq_data *sqd = ctx->sq_data;
8825 struct io_uring_task *tctx;
8831 if (!io_sq_thread_park(sqd))
8833 tctx = ctx->sq_data->thread->io_uring;
8834 /* can happen on fork/alloc failure, just ignore that state */
8836 io_sq_thread_unpark(sqd);
8840 atomic_inc(&tctx->in_idle);
8842 /* read completions before cancelations */
8843 inflight = tctx_inflight(tctx);
8846 io_uring_cancel_task_requests(ctx, NULL);
8848 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
8850 * If we've seen completions, retry without waiting. This
8851 * avoids a race where a completion comes in before we did
8852 * prepare_to_wait().
8854 if (inflight == tctx_inflight(tctx))
8856 finish_wait(&tctx->wait, &wait);
8858 atomic_dec(&tctx->in_idle);
8859 io_sq_thread_unpark(sqd);
8863 * Find any io_uring fd that this task has registered or done IO on, and cancel
8866 void __io_uring_task_cancel(void)
8868 struct io_uring_task *tctx = current->io_uring;
8872 /* make sure overflow events are dropped */
8873 atomic_inc(&tctx->in_idle);
8877 unsigned long index;
8879 xa_for_each(&tctx->xa, index, file)
8880 io_uring_cancel_sqpoll(file->private_data);
8884 /* read completions before cancelations */
8885 inflight = tctx_inflight(tctx);
8888 __io_uring_files_cancel(NULL);
8890 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
8893 * If we've seen completions, retry without waiting. This
8894 * avoids a race where a completion comes in before we did
8895 * prepare_to_wait().
8897 if (inflight == tctx_inflight(tctx))
8899 finish_wait(&tctx->wait, &wait);
8902 atomic_dec(&tctx->in_idle);
8904 io_uring_clean_tctx(tctx);
8905 /* all current's requests should be gone, we can kill tctx */
8906 __io_uring_free(current);
8909 static void *io_uring_validate_mmap_request(struct file *file,
8910 loff_t pgoff, size_t sz)
8912 struct io_ring_ctx *ctx = file->private_data;
8913 loff_t offset = pgoff << PAGE_SHIFT;
8918 case IORING_OFF_SQ_RING:
8919 case IORING_OFF_CQ_RING:
8922 case IORING_OFF_SQES:
8926 return ERR_PTR(-EINVAL);
8929 page = virt_to_head_page(ptr);
8930 if (sz > page_size(page))
8931 return ERR_PTR(-EINVAL);
8938 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
8940 size_t sz = vma->vm_end - vma->vm_start;
8944 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
8946 return PTR_ERR(ptr);
8948 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
8949 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
8952 #else /* !CONFIG_MMU */
8954 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
8956 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
8959 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
8961 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
8964 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
8965 unsigned long addr, unsigned long len,
8966 unsigned long pgoff, unsigned long flags)
8970 ptr = io_uring_validate_mmap_request(file, pgoff, len);
8972 return PTR_ERR(ptr);
8974 return (unsigned long) ptr;
8977 #endif /* !CONFIG_MMU */
8979 static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
8985 if (!io_sqring_full(ctx))
8987 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
8989 if (!io_sqring_full(ctx))
8992 } while (!signal_pending(current));
8994 finish_wait(&ctx->sqo_sq_wait, &wait);
8998 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
8999 struct __kernel_timespec __user **ts,
9000 const sigset_t __user **sig)
9002 struct io_uring_getevents_arg arg;
9005 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9006 * is just a pointer to the sigset_t.
9008 if (!(flags & IORING_ENTER_EXT_ARG)) {
9009 *sig = (const sigset_t __user *) argp;
9015 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9016 * timespec and sigset_t pointers if good.
9018 if (*argsz != sizeof(arg))
9020 if (copy_from_user(&arg, argp, sizeof(arg)))
9022 *sig = u64_to_user_ptr(arg.sigmask);
9023 *argsz = arg.sigmask_sz;
9024 *ts = u64_to_user_ptr(arg.ts);
9028 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9029 u32, min_complete, u32, flags, const void __user *, argp,
9032 struct io_ring_ctx *ctx;
9039 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9040 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG))
9048 if (f.file->f_op != &io_uring_fops)
9052 ctx = f.file->private_data;
9053 if (!percpu_ref_tryget(&ctx->refs))
9057 if (ctx->flags & IORING_SETUP_R_DISABLED)
9061 * For SQ polling, the thread will do all submissions and completions.
9062 * Just return the requested submit count, and wake the thread if
9066 if (ctx->flags & IORING_SETUP_SQPOLL) {
9067 io_cqring_overflow_flush(ctx, false, NULL, NULL);
9069 if (unlikely(ctx->sqo_exec)) {
9070 ret = io_sq_thread_fork(ctx->sq_data, ctx);
9076 if (flags & IORING_ENTER_SQ_WAKEUP)
9077 wake_up(&ctx->sq_data->wait);
9078 if (flags & IORING_ENTER_SQ_WAIT) {
9079 ret = io_sqpoll_wait_sq(ctx);
9083 submitted = to_submit;
9084 } else if (to_submit) {
9085 ret = io_uring_add_task_file(ctx, f.file);
9088 mutex_lock(&ctx->uring_lock);
9089 submitted = io_submit_sqes(ctx, to_submit);
9090 mutex_unlock(&ctx->uring_lock);
9092 if (submitted != to_submit)
9095 if (flags & IORING_ENTER_GETEVENTS) {
9096 const sigset_t __user *sig;
9097 struct __kernel_timespec __user *ts;
9099 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
9103 min_complete = min(min_complete, ctx->cq_entries);
9106 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9107 * space applications don't need to do io completion events
9108 * polling again, they can rely on io_sq_thread to do polling
9109 * work, which can reduce cpu usage and uring_lock contention.
9111 if (ctx->flags & IORING_SETUP_IOPOLL &&
9112 !(ctx->flags & IORING_SETUP_SQPOLL)) {
9113 ret = io_iopoll_check(ctx, min_complete);
9115 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
9120 percpu_ref_put(&ctx->refs);
9123 return submitted ? submitted : ret;
9126 #ifdef CONFIG_PROC_FS
9127 static int io_uring_show_cred(int id, void *p, void *data)
9129 const struct cred *cred = p;
9130 struct seq_file *m = data;
9131 struct user_namespace *uns = seq_user_ns(m);
9132 struct group_info *gi;
9137 seq_printf(m, "%5d\n", id);
9138 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
9139 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
9140 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
9141 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
9142 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
9143 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
9144 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
9145 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
9146 seq_puts(m, "\n\tGroups:\t");
9147 gi = cred->group_info;
9148 for (g = 0; g < gi->ngroups; g++) {
9149 seq_put_decimal_ull(m, g ? " " : "",
9150 from_kgid_munged(uns, gi->gid[g]));
9152 seq_puts(m, "\n\tCapEff:\t");
9153 cap = cred->cap_effective;
9154 CAP_FOR_EACH_U32(__capi)
9155 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
9160 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
9162 struct io_sq_data *sq = NULL;
9167 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9168 * since fdinfo case grabs it in the opposite direction of normal use
9169 * cases. If we fail to get the lock, we just don't iterate any
9170 * structures that could be going away outside the io_uring mutex.
9172 has_lock = mutex_trylock(&ctx->uring_lock);
9174 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) {
9180 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9181 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9182 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9183 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9184 struct file *f = *io_fixed_file_slot(ctx->file_data, i);
9187 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9189 seq_printf(m, "%5u: <none>\n", i);
9191 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9192 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9193 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
9195 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
9196 (unsigned int) buf->len);
9198 if (has_lock && !idr_is_empty(&ctx->personality_idr)) {
9199 seq_printf(m, "Personalities:\n");
9200 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
9202 seq_printf(m, "PollList:\n");
9203 spin_lock_irq(&ctx->completion_lock);
9204 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9205 struct hlist_head *list = &ctx->cancel_hash[i];
9206 struct io_kiocb *req;
9208 hlist_for_each_entry(req, list, hash_node)
9209 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9210 req->task->task_works != NULL);
9212 spin_unlock_irq(&ctx->completion_lock);
9214 mutex_unlock(&ctx->uring_lock);
9217 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9219 struct io_ring_ctx *ctx = f->private_data;
9221 if (percpu_ref_tryget(&ctx->refs)) {
9222 __io_uring_show_fdinfo(ctx, m);
9223 percpu_ref_put(&ctx->refs);
9228 static const struct file_operations io_uring_fops = {
9229 .release = io_uring_release,
9230 .mmap = io_uring_mmap,
9232 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9233 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9235 .poll = io_uring_poll,
9236 .fasync = io_uring_fasync,
9237 #ifdef CONFIG_PROC_FS
9238 .show_fdinfo = io_uring_show_fdinfo,
9242 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9243 struct io_uring_params *p)
9245 struct io_rings *rings;
9246 size_t size, sq_array_offset;
9248 /* make sure these are sane, as we already accounted them */
9249 ctx->sq_entries = p->sq_entries;
9250 ctx->cq_entries = p->cq_entries;
9252 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9253 if (size == SIZE_MAX)
9256 rings = io_mem_alloc(size);
9261 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9262 rings->sq_ring_mask = p->sq_entries - 1;
9263 rings->cq_ring_mask = p->cq_entries - 1;
9264 rings->sq_ring_entries = p->sq_entries;
9265 rings->cq_ring_entries = p->cq_entries;
9266 ctx->sq_mask = rings->sq_ring_mask;
9267 ctx->cq_mask = rings->cq_ring_mask;
9269 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9270 if (size == SIZE_MAX) {
9271 io_mem_free(ctx->rings);
9276 ctx->sq_sqes = io_mem_alloc(size);
9277 if (!ctx->sq_sqes) {
9278 io_mem_free(ctx->rings);
9286 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
9290 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9294 ret = io_uring_add_task_file(ctx, file);
9299 fd_install(fd, file);
9304 * Allocate an anonymous fd, this is what constitutes the application
9305 * visible backing of an io_uring instance. The application mmaps this
9306 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9307 * we have to tie this fd to a socket for file garbage collection purposes.
9309 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
9312 #if defined(CONFIG_UNIX)
9315 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9318 return ERR_PTR(ret);
9321 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9322 O_RDWR | O_CLOEXEC);
9323 #if defined(CONFIG_UNIX)
9325 sock_release(ctx->ring_sock);
9326 ctx->ring_sock = NULL;
9328 ctx->ring_sock->file = file;
9334 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9335 struct io_uring_params __user *params)
9337 struct io_ring_ctx *ctx;
9343 if (entries > IORING_MAX_ENTRIES) {
9344 if (!(p->flags & IORING_SETUP_CLAMP))
9346 entries = IORING_MAX_ENTRIES;
9350 * Use twice as many entries for the CQ ring. It's possible for the
9351 * application to drive a higher depth than the size of the SQ ring,
9352 * since the sqes are only used at submission time. This allows for
9353 * some flexibility in overcommitting a bit. If the application has
9354 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9355 * of CQ ring entries manually.
9357 p->sq_entries = roundup_pow_of_two(entries);
9358 if (p->flags & IORING_SETUP_CQSIZE) {
9360 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9361 * to a power-of-two, if it isn't already. We do NOT impose
9362 * any cq vs sq ring sizing.
9366 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9367 if (!(p->flags & IORING_SETUP_CLAMP))
9369 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9371 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9372 if (p->cq_entries < p->sq_entries)
9375 p->cq_entries = 2 * p->sq_entries;
9378 ctx = io_ring_ctx_alloc(p);
9381 ctx->compat = in_compat_syscall();
9382 if (!capable(CAP_IPC_LOCK))
9383 ctx->user = get_uid(current_user());
9386 * This is just grabbed for accounting purposes. When a process exits,
9387 * the mm is exited and dropped before the files, hence we need to hang
9388 * on to this mm purely for the purposes of being able to unaccount
9389 * memory (locked/pinned vm). It's not used for anything else.
9391 mmgrab(current->mm);
9392 ctx->mm_account = current->mm;
9394 ret = io_allocate_scq_urings(ctx, p);
9398 ret = io_sq_offload_create(ctx, p);
9402 if (!(p->flags & IORING_SETUP_R_DISABLED))
9403 io_sq_offload_start(ctx);
9405 memset(&p->sq_off, 0, sizeof(p->sq_off));
9406 p->sq_off.head = offsetof(struct io_rings, sq.head);
9407 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9408 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9409 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9410 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9411 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9412 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9414 memset(&p->cq_off, 0, sizeof(p->cq_off));
9415 p->cq_off.head = offsetof(struct io_rings, cq.head);
9416 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9417 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9418 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9419 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9420 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9421 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9423 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9424 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9425 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9426 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
9427 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS;
9429 if (copy_to_user(params, p, sizeof(*p))) {
9434 file = io_uring_get_file(ctx);
9436 ret = PTR_ERR(file);
9441 * Install ring fd as the very last thing, so we don't risk someone
9442 * having closed it before we finish setup
9444 ret = io_uring_install_fd(ctx, file);
9446 /* fput will clean it up */
9451 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9454 io_ring_ctx_wait_and_kill(ctx);
9459 * Sets up an aio uring context, and returns the fd. Applications asks for a
9460 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9461 * params structure passed in.
9463 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9465 struct io_uring_params p;
9468 if (copy_from_user(&p, params, sizeof(p)))
9470 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9475 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9476 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9477 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9478 IORING_SETUP_R_DISABLED))
9481 return io_uring_create(entries, &p, params);
9484 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9485 struct io_uring_params __user *, params)
9487 return io_uring_setup(entries, params);
9490 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9492 struct io_uring_probe *p;
9496 size = struct_size(p, ops, nr_args);
9497 if (size == SIZE_MAX)
9499 p = kzalloc(size, GFP_KERNEL);
9504 if (copy_from_user(p, arg, size))
9507 if (memchr_inv(p, 0, size))
9510 p->last_op = IORING_OP_LAST - 1;
9511 if (nr_args > IORING_OP_LAST)
9512 nr_args = IORING_OP_LAST;
9514 for (i = 0; i < nr_args; i++) {
9516 if (!io_op_defs[i].not_supported)
9517 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9522 if (copy_to_user(arg, p, size))
9529 static int io_register_personality(struct io_ring_ctx *ctx)
9531 const struct cred *creds;
9534 creds = get_current_cred();
9536 ret = idr_alloc_cyclic(&ctx->personality_idr, (void *) creds, 1,
9537 USHRT_MAX, GFP_KERNEL);
9543 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9544 unsigned int nr_args)
9546 struct io_uring_restriction *res;
9550 /* Restrictions allowed only if rings started disabled */
9551 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9554 /* We allow only a single restrictions registration */
9555 if (ctx->restrictions.registered)
9558 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9561 size = array_size(nr_args, sizeof(*res));
9562 if (size == SIZE_MAX)
9565 res = memdup_user(arg, size);
9567 return PTR_ERR(res);
9571 for (i = 0; i < nr_args; i++) {
9572 switch (res[i].opcode) {
9573 case IORING_RESTRICTION_REGISTER_OP:
9574 if (res[i].register_op >= IORING_REGISTER_LAST) {
9579 __set_bit(res[i].register_op,
9580 ctx->restrictions.register_op);
9582 case IORING_RESTRICTION_SQE_OP:
9583 if (res[i].sqe_op >= IORING_OP_LAST) {
9588 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9590 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9591 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9593 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9594 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9603 /* Reset all restrictions if an error happened */
9605 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9607 ctx->restrictions.registered = true;
9613 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9615 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9618 if (ctx->restrictions.registered)
9619 ctx->restricted = 1;
9621 io_sq_offload_start(ctx);
9625 static bool io_register_op_must_quiesce(int op)
9628 case IORING_UNREGISTER_FILES:
9629 case IORING_REGISTER_FILES_UPDATE:
9630 case IORING_REGISTER_PROBE:
9631 case IORING_REGISTER_PERSONALITY:
9632 case IORING_UNREGISTER_PERSONALITY:
9639 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
9640 void __user *arg, unsigned nr_args)
9641 __releases(ctx->uring_lock)
9642 __acquires(ctx->uring_lock)
9647 * We're inside the ring mutex, if the ref is already dying, then
9648 * someone else killed the ctx or is already going through
9649 * io_uring_register().
9651 if (percpu_ref_is_dying(&ctx->refs))
9654 if (io_register_op_must_quiesce(opcode)) {
9655 percpu_ref_kill(&ctx->refs);
9658 * Drop uring mutex before waiting for references to exit. If
9659 * another thread is currently inside io_uring_enter() it might
9660 * need to grab the uring_lock to make progress. If we hold it
9661 * here across the drain wait, then we can deadlock. It's safe
9662 * to drop the mutex here, since no new references will come in
9663 * after we've killed the percpu ref.
9665 mutex_unlock(&ctx->uring_lock);
9667 ret = wait_for_completion_interruptible(&ctx->ref_comp);
9670 ret = io_run_task_work_sig();
9675 mutex_lock(&ctx->uring_lock);
9678 percpu_ref_resurrect(&ctx->refs);
9683 if (ctx->restricted) {
9684 if (opcode >= IORING_REGISTER_LAST) {
9689 if (!test_bit(opcode, ctx->restrictions.register_op)) {
9696 case IORING_REGISTER_BUFFERS:
9697 ret = io_sqe_buffers_register(ctx, arg, nr_args);
9699 case IORING_UNREGISTER_BUFFERS:
9703 ret = io_sqe_buffers_unregister(ctx);
9705 case IORING_REGISTER_FILES:
9706 ret = io_sqe_files_register(ctx, arg, nr_args);
9708 case IORING_UNREGISTER_FILES:
9712 ret = io_sqe_files_unregister(ctx);
9714 case IORING_REGISTER_FILES_UPDATE:
9715 ret = io_sqe_files_update(ctx, arg, nr_args);
9717 case IORING_REGISTER_EVENTFD:
9718 case IORING_REGISTER_EVENTFD_ASYNC:
9722 ret = io_eventfd_register(ctx, arg);
9725 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
9726 ctx->eventfd_async = 1;
9728 ctx->eventfd_async = 0;
9730 case IORING_UNREGISTER_EVENTFD:
9734 ret = io_eventfd_unregister(ctx);
9736 case IORING_REGISTER_PROBE:
9738 if (!arg || nr_args > 256)
9740 ret = io_probe(ctx, arg, nr_args);
9742 case IORING_REGISTER_PERSONALITY:
9746 ret = io_register_personality(ctx);
9748 case IORING_UNREGISTER_PERSONALITY:
9752 ret = io_unregister_personality(ctx, nr_args);
9754 case IORING_REGISTER_ENABLE_RINGS:
9758 ret = io_register_enable_rings(ctx);
9760 case IORING_REGISTER_RESTRICTIONS:
9761 ret = io_register_restrictions(ctx, arg, nr_args);
9769 if (io_register_op_must_quiesce(opcode)) {
9770 /* bring the ctx back to life */
9771 percpu_ref_reinit(&ctx->refs);
9773 reinit_completion(&ctx->ref_comp);
9778 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
9779 void __user *, arg, unsigned int, nr_args)
9781 struct io_ring_ctx *ctx;
9790 if (f.file->f_op != &io_uring_fops)
9793 ctx = f.file->private_data;
9797 mutex_lock(&ctx->uring_lock);
9798 ret = __io_uring_register(ctx, opcode, arg, nr_args);
9799 mutex_unlock(&ctx->uring_lock);
9800 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
9801 ctx->cq_ev_fd != NULL, ret);
9807 static int __init io_uring_init(void)
9809 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
9810 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
9811 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
9814 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
9815 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
9816 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
9817 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
9818 BUILD_BUG_SQE_ELEM(1, __u8, flags);
9819 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
9820 BUILD_BUG_SQE_ELEM(4, __s32, fd);
9821 BUILD_BUG_SQE_ELEM(8, __u64, off);
9822 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
9823 BUILD_BUG_SQE_ELEM(16, __u64, addr);
9824 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
9825 BUILD_BUG_SQE_ELEM(24, __u32, len);
9826 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
9827 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
9828 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
9829 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
9830 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
9831 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
9832 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
9833 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
9834 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
9835 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
9836 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
9837 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
9838 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
9839 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
9840 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
9841 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
9842 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
9843 BUILD_BUG_SQE_ELEM(42, __u16, personality);
9844 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
9846 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
9847 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
9848 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
9852 __initcall(io_uring_init);