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_dead: 1;
342 unsigned int sqo_exec: 1;
345 * Ring buffer of indices into array of io_uring_sqe, which is
346 * mmapped by the application using the IORING_OFF_SQES offset.
348 * This indirection could e.g. be used to assign fixed
349 * io_uring_sqe entries to operations and only submit them to
350 * the queue when needed.
352 * The kernel modifies neither the indices array nor the entries
356 unsigned cached_sq_head;
359 unsigned sq_thread_idle;
360 unsigned cached_sq_dropped;
361 unsigned cached_cq_overflow;
362 unsigned long sq_check_overflow;
364 /* hashed buffered write serialization */
365 struct io_wq_hash *hash_map;
367 struct list_head defer_list;
368 struct list_head timeout_list;
369 struct list_head cq_overflow_list;
371 struct io_uring_sqe *sq_sqes;
372 } ____cacheline_aligned_in_smp;
375 struct mutex uring_lock;
376 wait_queue_head_t wait;
377 } ____cacheline_aligned_in_smp;
379 struct io_submit_state submit_state;
381 struct io_rings *rings;
386 struct task_struct *sqo_task;
388 /* Only used for accounting purposes */
389 struct mm_struct *mm_account;
391 struct io_sq_data *sq_data; /* if using sq thread polling */
393 struct wait_queue_head sqo_sq_wait;
394 struct list_head sqd_list;
397 * If used, fixed file set. Writers must ensure that ->refs is dead,
398 * readers must ensure that ->refs is alive as long as the file* is
399 * used. Only updated through io_uring_register(2).
401 struct fixed_rsrc_data *file_data;
402 unsigned nr_user_files;
404 /* if used, fixed mapped user buffers */
405 unsigned nr_user_bufs;
406 struct io_mapped_ubuf *user_bufs;
408 struct user_struct *user;
410 struct completion ref_comp;
411 struct completion sq_thread_comp;
413 #if defined(CONFIG_UNIX)
414 struct socket *ring_sock;
417 struct idr io_buffer_idr;
419 struct idr personality_idr;
422 unsigned cached_cq_tail;
425 atomic_t cq_timeouts;
426 unsigned cq_last_tm_flush;
427 unsigned long cq_check_overflow;
428 struct wait_queue_head cq_wait;
429 struct fasync_struct *cq_fasync;
430 struct eventfd_ctx *cq_ev_fd;
431 } ____cacheline_aligned_in_smp;
434 spinlock_t completion_lock;
437 * ->iopoll_list is protected by the ctx->uring_lock for
438 * io_uring instances that don't use IORING_SETUP_SQPOLL.
439 * For SQPOLL, only the single threaded io_sq_thread() will
440 * manipulate the list, hence no extra locking is needed there.
442 struct list_head iopoll_list;
443 struct hlist_head *cancel_hash;
444 unsigned cancel_hash_bits;
445 bool poll_multi_file;
447 spinlock_t inflight_lock;
448 struct list_head inflight_list;
449 } ____cacheline_aligned_in_smp;
451 struct delayed_work rsrc_put_work;
452 struct llist_head rsrc_put_llist;
453 struct list_head rsrc_ref_list;
454 spinlock_t rsrc_ref_lock;
456 struct io_restriction restrictions;
459 struct callback_head *exit_task_work;
461 struct wait_queue_head hash_wait;
463 /* Keep this last, we don't need it for the fast path */
464 struct work_struct exit_work;
468 * First field must be the file pointer in all the
469 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
471 struct io_poll_iocb {
473 struct wait_queue_head *head;
477 struct wait_queue_entry wait;
480 struct io_poll_remove {
490 struct io_timeout_data {
491 struct io_kiocb *req;
492 struct hrtimer timer;
493 struct timespec64 ts;
494 enum hrtimer_mode mode;
499 struct sockaddr __user *addr;
500 int __user *addr_len;
502 unsigned long nofile;
522 struct list_head list;
523 /* head of the link, used by linked timeouts only */
524 struct io_kiocb *head;
527 struct io_timeout_rem {
532 struct timespec64 ts;
537 /* NOTE: kiocb has the file as the first member, so don't do it here */
545 struct sockaddr __user *addr;
552 struct user_msghdr __user *umsg;
558 struct io_buffer *kbuf;
564 struct filename *filename;
566 unsigned long nofile;
569 struct io_rsrc_update {
595 struct epoll_event event;
599 struct file *file_out;
600 struct file *file_in;
607 struct io_provide_buf {
621 const char __user *filename;
622 struct statx __user *buffer;
634 struct filename *oldpath;
635 struct filename *newpath;
643 struct filename *filename;
646 struct io_completion {
648 struct list_head list;
652 struct io_async_connect {
653 struct sockaddr_storage address;
656 struct io_async_msghdr {
657 struct iovec fast_iov[UIO_FASTIOV];
658 /* points to an allocated iov, if NULL we use fast_iov instead */
659 struct iovec *free_iov;
660 struct sockaddr __user *uaddr;
662 struct sockaddr_storage addr;
666 struct iovec fast_iov[UIO_FASTIOV];
667 const struct iovec *free_iovec;
668 struct iov_iter iter;
670 struct wait_page_queue wpq;
674 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
675 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
676 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
677 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
678 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
679 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
685 REQ_F_LINK_TIMEOUT_BIT,
687 REQ_F_NEED_CLEANUP_BIT,
689 REQ_F_BUFFER_SELECTED_BIT,
690 REQ_F_NO_FILE_TABLE_BIT,
691 REQ_F_LTIMEOUT_ACTIVE_BIT,
692 REQ_F_COMPLETE_INLINE_BIT,
694 /* not a real bit, just to check we're not overflowing the space */
700 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
701 /* drain existing IO first */
702 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
704 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
705 /* doesn't sever on completion < 0 */
706 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
708 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
709 /* IOSQE_BUFFER_SELECT */
710 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
712 /* fail rest of links */
713 REQ_F_FAIL_LINK = BIT(REQ_F_FAIL_LINK_BIT),
714 /* on inflight list */
715 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
716 /* read/write uses file position */
717 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
718 /* must not punt to workers */
719 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
720 /* has or had linked timeout */
721 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
723 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
725 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
726 /* already went through poll handler */
727 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
728 /* buffer already selected */
729 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
730 /* doesn't need file table for this request */
731 REQ_F_NO_FILE_TABLE = BIT(REQ_F_NO_FILE_TABLE_BIT),
732 /* linked timeout is active, i.e. prepared by link's head */
733 REQ_F_LTIMEOUT_ACTIVE = BIT(REQ_F_LTIMEOUT_ACTIVE_BIT),
734 /* completion is deferred through io_comp_state */
735 REQ_F_COMPLETE_INLINE = BIT(REQ_F_COMPLETE_INLINE_BIT),
739 struct io_poll_iocb poll;
740 struct io_poll_iocb *double_poll;
743 struct io_task_work {
744 struct io_wq_work_node node;
745 task_work_func_t func;
749 * NOTE! Each of the iocb union members has the file pointer
750 * as the first entry in their struct definition. So you can
751 * access the file pointer through any of the sub-structs,
752 * or directly as just 'ki_filp' in this struct.
758 struct io_poll_iocb poll;
759 struct io_poll_remove poll_remove;
760 struct io_accept accept;
762 struct io_cancel cancel;
763 struct io_timeout timeout;
764 struct io_timeout_rem timeout_rem;
765 struct io_connect connect;
766 struct io_sr_msg sr_msg;
768 struct io_close close;
769 struct io_rsrc_update rsrc_update;
770 struct io_fadvise fadvise;
771 struct io_madvise madvise;
772 struct io_epoll epoll;
773 struct io_splice splice;
774 struct io_provide_buf pbuf;
775 struct io_statx statx;
776 struct io_shutdown shutdown;
777 struct io_rename rename;
778 struct io_unlink unlink;
779 /* use only after cleaning per-op data, see io_clean_op() */
780 struct io_completion compl;
783 /* opcode allocated if it needs to store data for async defer */
786 /* polled IO has completed */
792 struct io_ring_ctx *ctx;
795 struct task_struct *task;
798 struct io_kiocb *link;
799 struct percpu_ref *fixed_rsrc_refs;
802 * 1. used with ctx->iopoll_list with reads/writes
803 * 2. to track reqs with ->files (see io_op_def::file_table)
805 struct list_head inflight_entry;
807 struct io_task_work io_task_work;
808 struct callback_head task_work;
810 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
811 struct hlist_node hash_node;
812 struct async_poll *apoll;
813 struct io_wq_work work;
816 struct io_defer_entry {
817 struct list_head list;
818 struct io_kiocb *req;
823 /* needs req->file assigned */
824 unsigned needs_file : 1;
825 /* hash wq insertion if file is a regular file */
826 unsigned hash_reg_file : 1;
827 /* unbound wq insertion if file is a non-regular file */
828 unsigned unbound_nonreg_file : 1;
829 /* opcode is not supported by this kernel */
830 unsigned not_supported : 1;
831 /* set if opcode supports polled "wait" */
833 unsigned pollout : 1;
834 /* op supports buffer selection */
835 unsigned buffer_select : 1;
836 /* must always have async data allocated */
837 unsigned needs_async_data : 1;
838 /* should block plug */
840 /* size of async data needed, if any */
841 unsigned short async_size;
844 static const struct io_op_def io_op_defs[] = {
845 [IORING_OP_NOP] = {},
846 [IORING_OP_READV] = {
848 .unbound_nonreg_file = 1,
851 .needs_async_data = 1,
853 .async_size = sizeof(struct io_async_rw),
855 [IORING_OP_WRITEV] = {
858 .unbound_nonreg_file = 1,
860 .needs_async_data = 1,
862 .async_size = sizeof(struct io_async_rw),
864 [IORING_OP_FSYNC] = {
867 [IORING_OP_READ_FIXED] = {
869 .unbound_nonreg_file = 1,
872 .async_size = sizeof(struct io_async_rw),
874 [IORING_OP_WRITE_FIXED] = {
877 .unbound_nonreg_file = 1,
880 .async_size = sizeof(struct io_async_rw),
882 [IORING_OP_POLL_ADD] = {
884 .unbound_nonreg_file = 1,
886 [IORING_OP_POLL_REMOVE] = {},
887 [IORING_OP_SYNC_FILE_RANGE] = {
890 [IORING_OP_SENDMSG] = {
892 .unbound_nonreg_file = 1,
894 .needs_async_data = 1,
895 .async_size = sizeof(struct io_async_msghdr),
897 [IORING_OP_RECVMSG] = {
899 .unbound_nonreg_file = 1,
902 .needs_async_data = 1,
903 .async_size = sizeof(struct io_async_msghdr),
905 [IORING_OP_TIMEOUT] = {
906 .needs_async_data = 1,
907 .async_size = sizeof(struct io_timeout_data),
909 [IORING_OP_TIMEOUT_REMOVE] = {
910 /* used by timeout updates' prep() */
912 [IORING_OP_ACCEPT] = {
914 .unbound_nonreg_file = 1,
917 [IORING_OP_ASYNC_CANCEL] = {},
918 [IORING_OP_LINK_TIMEOUT] = {
919 .needs_async_data = 1,
920 .async_size = sizeof(struct io_timeout_data),
922 [IORING_OP_CONNECT] = {
924 .unbound_nonreg_file = 1,
926 .needs_async_data = 1,
927 .async_size = sizeof(struct io_async_connect),
929 [IORING_OP_FALLOCATE] = {
932 [IORING_OP_OPENAT] = {},
933 [IORING_OP_CLOSE] = {},
934 [IORING_OP_FILES_UPDATE] = {},
935 [IORING_OP_STATX] = {},
938 .unbound_nonreg_file = 1,
942 .async_size = sizeof(struct io_async_rw),
944 [IORING_OP_WRITE] = {
946 .unbound_nonreg_file = 1,
949 .async_size = sizeof(struct io_async_rw),
951 [IORING_OP_FADVISE] = {
954 [IORING_OP_MADVISE] = {},
957 .unbound_nonreg_file = 1,
962 .unbound_nonreg_file = 1,
966 [IORING_OP_OPENAT2] = {
968 [IORING_OP_EPOLL_CTL] = {
969 .unbound_nonreg_file = 1,
971 [IORING_OP_SPLICE] = {
974 .unbound_nonreg_file = 1,
976 [IORING_OP_PROVIDE_BUFFERS] = {},
977 [IORING_OP_REMOVE_BUFFERS] = {},
981 .unbound_nonreg_file = 1,
983 [IORING_OP_SHUTDOWN] = {
986 [IORING_OP_RENAMEAT] = {},
987 [IORING_OP_UNLINKAT] = {},
990 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
991 struct task_struct *task,
992 struct files_struct *files);
993 static void io_uring_cancel_sqpoll(struct io_ring_ctx *ctx);
994 static void destroy_fixed_rsrc_ref_node(struct fixed_rsrc_ref_node *ref_node);
995 static struct fixed_rsrc_ref_node *alloc_fixed_rsrc_ref_node(
996 struct io_ring_ctx *ctx);
997 static void io_ring_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc);
999 static bool io_rw_reissue(struct io_kiocb *req);
1000 static void io_cqring_fill_event(struct io_kiocb *req, long res);
1001 static void io_put_req(struct io_kiocb *req);
1002 static void io_put_req_deferred(struct io_kiocb *req, int nr);
1003 static void io_double_put_req(struct io_kiocb *req);
1004 static void io_dismantle_req(struct io_kiocb *req);
1005 static void io_put_task(struct task_struct *task, int nr);
1006 static void io_queue_next(struct io_kiocb *req);
1007 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
1008 static void __io_queue_linked_timeout(struct io_kiocb *req);
1009 static void io_queue_linked_timeout(struct io_kiocb *req);
1010 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
1011 struct io_uring_rsrc_update *ip,
1013 static void __io_clean_op(struct io_kiocb *req);
1014 static struct file *io_file_get(struct io_submit_state *state,
1015 struct io_kiocb *req, int fd, bool fixed);
1016 static void __io_queue_sqe(struct io_kiocb *req);
1017 static void io_rsrc_put_work(struct work_struct *work);
1019 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
1020 struct iov_iter *iter, bool needs_lock);
1021 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
1022 const struct iovec *fast_iov,
1023 struct iov_iter *iter, bool force);
1024 static void io_req_task_queue(struct io_kiocb *req);
1025 static void io_submit_flush_completions(struct io_comp_state *cs,
1026 struct io_ring_ctx *ctx);
1028 static struct kmem_cache *req_cachep;
1030 static const struct file_operations io_uring_fops;
1032 struct sock *io_uring_get_socket(struct file *file)
1034 #if defined(CONFIG_UNIX)
1035 if (file->f_op == &io_uring_fops) {
1036 struct io_ring_ctx *ctx = file->private_data;
1038 return ctx->ring_sock->sk;
1043 EXPORT_SYMBOL(io_uring_get_socket);
1045 #define io_for_each_link(pos, head) \
1046 for (pos = (head); pos; pos = pos->link)
1048 static inline void io_clean_op(struct io_kiocb *req)
1050 if (req->flags & (REQ_F_NEED_CLEANUP | REQ_F_BUFFER_SELECTED))
1054 static inline void io_set_resource_node(struct io_kiocb *req)
1056 struct io_ring_ctx *ctx = req->ctx;
1058 if (!req->fixed_rsrc_refs) {
1059 req->fixed_rsrc_refs = &ctx->file_data->node->refs;
1060 percpu_ref_get(req->fixed_rsrc_refs);
1064 static bool io_match_task(struct io_kiocb *head,
1065 struct task_struct *task,
1066 struct files_struct *files)
1068 struct io_kiocb *req;
1070 if (task && head->task != task) {
1071 /* in terms of cancelation, always match if req task is dead */
1072 if (head->task->flags & PF_EXITING)
1079 io_for_each_link(req, head) {
1080 if (req->file && req->file->f_op == &io_uring_fops)
1082 if (req->task->files == files)
1088 static inline void req_set_fail_links(struct io_kiocb *req)
1090 if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
1091 req->flags |= REQ_F_FAIL_LINK;
1094 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1096 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1098 complete(&ctx->ref_comp);
1101 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1103 return !req->timeout.off;
1106 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1108 struct io_ring_ctx *ctx;
1111 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1116 * Use 5 bits less than the max cq entries, that should give us around
1117 * 32 entries per hash list if totally full and uniformly spread.
1119 hash_bits = ilog2(p->cq_entries);
1123 ctx->cancel_hash_bits = hash_bits;
1124 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1126 if (!ctx->cancel_hash)
1128 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1130 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1131 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1134 ctx->flags = p->flags;
1135 init_waitqueue_head(&ctx->sqo_sq_wait);
1136 INIT_LIST_HEAD(&ctx->sqd_list);
1137 init_waitqueue_head(&ctx->cq_wait);
1138 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1139 init_completion(&ctx->ref_comp);
1140 init_completion(&ctx->sq_thread_comp);
1141 idr_init(&ctx->io_buffer_idr);
1142 idr_init(&ctx->personality_idr);
1143 mutex_init(&ctx->uring_lock);
1144 init_waitqueue_head(&ctx->wait);
1145 spin_lock_init(&ctx->completion_lock);
1146 INIT_LIST_HEAD(&ctx->iopoll_list);
1147 INIT_LIST_HEAD(&ctx->defer_list);
1148 INIT_LIST_HEAD(&ctx->timeout_list);
1149 spin_lock_init(&ctx->inflight_lock);
1150 INIT_LIST_HEAD(&ctx->inflight_list);
1151 spin_lock_init(&ctx->rsrc_ref_lock);
1152 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
1153 INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
1154 init_llist_head(&ctx->rsrc_put_llist);
1155 INIT_LIST_HEAD(&ctx->submit_state.comp.free_list);
1156 INIT_LIST_HEAD(&ctx->submit_state.comp.locked_free_list);
1159 kfree(ctx->cancel_hash);
1164 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1166 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1167 struct io_ring_ctx *ctx = req->ctx;
1169 return seq != ctx->cached_cq_tail
1170 + READ_ONCE(ctx->cached_cq_overflow);
1176 static void io_req_clean_work(struct io_kiocb *req)
1178 if (req->flags & REQ_F_INFLIGHT) {
1179 struct io_ring_ctx *ctx = req->ctx;
1180 struct io_uring_task *tctx = req->task->io_uring;
1181 unsigned long flags;
1183 spin_lock_irqsave(&ctx->inflight_lock, flags);
1184 list_del(&req->inflight_entry);
1185 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
1186 req->flags &= ~REQ_F_INFLIGHT;
1187 if (atomic_read(&tctx->in_idle))
1188 wake_up(&tctx->wait);
1192 static void io_req_track_inflight(struct io_kiocb *req)
1194 struct io_ring_ctx *ctx = req->ctx;
1196 if (!(req->flags & REQ_F_INFLIGHT)) {
1197 req->flags |= REQ_F_INFLIGHT;
1199 spin_lock_irq(&ctx->inflight_lock);
1200 list_add(&req->inflight_entry, &ctx->inflight_list);
1201 spin_unlock_irq(&ctx->inflight_lock);
1205 static void io_prep_async_work(struct io_kiocb *req)
1207 const struct io_op_def *def = &io_op_defs[req->opcode];
1208 struct io_ring_ctx *ctx = req->ctx;
1210 if (req->flags & REQ_F_FORCE_ASYNC)
1211 req->work.flags |= IO_WQ_WORK_CONCURRENT;
1213 if (req->flags & REQ_F_ISREG) {
1214 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1215 io_wq_hash_work(&req->work, file_inode(req->file));
1217 if (def->unbound_nonreg_file)
1218 req->work.flags |= IO_WQ_WORK_UNBOUND;
1222 static void io_prep_async_link(struct io_kiocb *req)
1224 struct io_kiocb *cur;
1226 io_for_each_link(cur, req)
1227 io_prep_async_work(cur);
1230 static struct io_kiocb *__io_queue_async_work(struct io_kiocb *req)
1232 struct io_ring_ctx *ctx = req->ctx;
1233 struct io_kiocb *link = io_prep_linked_timeout(req);
1234 struct io_uring_task *tctx = req->task->io_uring;
1237 BUG_ON(!tctx->io_wq);
1239 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1240 &req->work, req->flags);
1241 io_wq_enqueue(tctx->io_wq, &req->work);
1245 static void io_queue_async_work(struct io_kiocb *req)
1247 struct io_kiocb *link;
1249 /* init ->work of the whole link before punting */
1250 io_prep_async_link(req);
1251 link = __io_queue_async_work(req);
1254 io_queue_linked_timeout(link);
1257 static void io_kill_timeout(struct io_kiocb *req)
1259 struct io_timeout_data *io = req->async_data;
1262 ret = hrtimer_try_to_cancel(&io->timer);
1264 atomic_set(&req->ctx->cq_timeouts,
1265 atomic_read(&req->ctx->cq_timeouts) + 1);
1266 list_del_init(&req->timeout.list);
1267 io_cqring_fill_event(req, 0);
1268 io_put_req_deferred(req, 1);
1273 * Returns true if we found and killed one or more timeouts
1275 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
1276 struct files_struct *files)
1278 struct io_kiocb *req, *tmp;
1281 spin_lock_irq(&ctx->completion_lock);
1282 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
1283 if (io_match_task(req, tsk, files)) {
1284 io_kill_timeout(req);
1288 spin_unlock_irq(&ctx->completion_lock);
1289 return canceled != 0;
1292 static void __io_queue_deferred(struct io_ring_ctx *ctx)
1295 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1296 struct io_defer_entry, list);
1298 if (req_need_defer(de->req, de->seq))
1300 list_del_init(&de->list);
1301 io_req_task_queue(de->req);
1303 } while (!list_empty(&ctx->defer_list));
1306 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1310 if (list_empty(&ctx->timeout_list))
1313 seq = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
1316 u32 events_needed, events_got;
1317 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1318 struct io_kiocb, timeout.list);
1320 if (io_is_timeout_noseq(req))
1324 * Since seq can easily wrap around over time, subtract
1325 * the last seq at which timeouts were flushed before comparing.
1326 * Assuming not more than 2^31-1 events have happened since,
1327 * these subtractions won't have wrapped, so we can check if
1328 * target is in [last_seq, current_seq] by comparing the two.
1330 events_needed = req->timeout.target_seq - ctx->cq_last_tm_flush;
1331 events_got = seq - ctx->cq_last_tm_flush;
1332 if (events_got < events_needed)
1335 list_del_init(&req->timeout.list);
1336 io_kill_timeout(req);
1337 } while (!list_empty(&ctx->timeout_list));
1339 ctx->cq_last_tm_flush = seq;
1342 static void io_commit_cqring(struct io_ring_ctx *ctx)
1344 io_flush_timeouts(ctx);
1346 /* order cqe stores with ring update */
1347 smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
1349 if (unlikely(!list_empty(&ctx->defer_list)))
1350 __io_queue_deferred(ctx);
1353 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1355 struct io_rings *r = ctx->rings;
1357 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == r->sq_ring_entries;
1360 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
1362 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
1365 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1367 struct io_rings *rings = ctx->rings;
1371 * writes to the cq entry need to come after reading head; the
1372 * control dependency is enough as we're using WRITE_ONCE to
1375 if (__io_cqring_events(ctx) == rings->cq_ring_entries)
1378 tail = ctx->cached_cq_tail++;
1379 return &rings->cqes[tail & ctx->cq_mask];
1382 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1386 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1388 if (!ctx->eventfd_async)
1390 return io_wq_current_is_worker();
1393 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1395 /* see waitqueue_active() comment */
1398 if (waitqueue_active(&ctx->wait))
1399 wake_up(&ctx->wait);
1400 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1401 wake_up(&ctx->sq_data->wait);
1402 if (io_should_trigger_evfd(ctx))
1403 eventfd_signal(ctx->cq_ev_fd, 1);
1404 if (waitqueue_active(&ctx->cq_wait)) {
1405 wake_up_interruptible(&ctx->cq_wait);
1406 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1410 static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx)
1412 /* see waitqueue_active() comment */
1415 if (ctx->flags & IORING_SETUP_SQPOLL) {
1416 if (waitqueue_active(&ctx->wait))
1417 wake_up(&ctx->wait);
1419 if (io_should_trigger_evfd(ctx))
1420 eventfd_signal(ctx->cq_ev_fd, 1);
1421 if (waitqueue_active(&ctx->cq_wait)) {
1422 wake_up_interruptible(&ctx->cq_wait);
1423 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1427 /* Returns true if there are no backlogged entries after the flush */
1428 static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force,
1429 struct task_struct *tsk,
1430 struct files_struct *files)
1432 struct io_rings *rings = ctx->rings;
1433 struct io_kiocb *req, *tmp;
1434 struct io_uring_cqe *cqe;
1435 unsigned long flags;
1436 bool all_flushed, posted;
1439 if (!force && __io_cqring_events(ctx) == rings->cq_ring_entries)
1443 spin_lock_irqsave(&ctx->completion_lock, flags);
1444 list_for_each_entry_safe(req, tmp, &ctx->cq_overflow_list, compl.list) {
1445 if (!io_match_task(req, tsk, files))
1448 cqe = io_get_cqring(ctx);
1452 list_move(&req->compl.list, &list);
1454 WRITE_ONCE(cqe->user_data, req->user_data);
1455 WRITE_ONCE(cqe->res, req->result);
1456 WRITE_ONCE(cqe->flags, req->compl.cflags);
1458 ctx->cached_cq_overflow++;
1459 WRITE_ONCE(ctx->rings->cq_overflow,
1460 ctx->cached_cq_overflow);
1465 all_flushed = list_empty(&ctx->cq_overflow_list);
1467 clear_bit(0, &ctx->sq_check_overflow);
1468 clear_bit(0, &ctx->cq_check_overflow);
1469 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
1473 io_commit_cqring(ctx);
1474 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1476 io_cqring_ev_posted(ctx);
1478 while (!list_empty(&list)) {
1479 req = list_first_entry(&list, struct io_kiocb, compl.list);
1480 list_del(&req->compl.list);
1487 static void io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force,
1488 struct task_struct *tsk,
1489 struct files_struct *files)
1491 if (test_bit(0, &ctx->cq_check_overflow)) {
1492 /* iopoll syncs against uring_lock, not completion_lock */
1493 if (ctx->flags & IORING_SETUP_IOPOLL)
1494 mutex_lock(&ctx->uring_lock);
1495 __io_cqring_overflow_flush(ctx, force, tsk, files);
1496 if (ctx->flags & IORING_SETUP_IOPOLL)
1497 mutex_unlock(&ctx->uring_lock);
1501 static void __io_cqring_fill_event(struct io_kiocb *req, long res, long cflags)
1503 struct io_ring_ctx *ctx = req->ctx;
1504 struct io_uring_cqe *cqe;
1506 trace_io_uring_complete(ctx, req->user_data, res);
1509 * If we can't get a cq entry, userspace overflowed the
1510 * submission (by quite a lot). Increment the overflow count in
1513 cqe = io_get_cqring(ctx);
1515 WRITE_ONCE(cqe->user_data, req->user_data);
1516 WRITE_ONCE(cqe->res, res);
1517 WRITE_ONCE(cqe->flags, cflags);
1518 } else if (ctx->cq_overflow_flushed ||
1519 atomic_read(&req->task->io_uring->in_idle)) {
1521 * If we're in ring overflow flush mode, or in task cancel mode,
1522 * then we cannot store the request for later flushing, we need
1523 * to drop it on the floor.
1525 ctx->cached_cq_overflow++;
1526 WRITE_ONCE(ctx->rings->cq_overflow, ctx->cached_cq_overflow);
1528 if (list_empty(&ctx->cq_overflow_list)) {
1529 set_bit(0, &ctx->sq_check_overflow);
1530 set_bit(0, &ctx->cq_check_overflow);
1531 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
1535 req->compl.cflags = cflags;
1536 refcount_inc(&req->refs);
1537 list_add_tail(&req->compl.list, &ctx->cq_overflow_list);
1541 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1543 __io_cqring_fill_event(req, res, 0);
1546 static inline void io_req_complete_post(struct io_kiocb *req, long res,
1547 unsigned int cflags)
1549 struct io_ring_ctx *ctx = req->ctx;
1550 unsigned long flags;
1552 spin_lock_irqsave(&ctx->completion_lock, flags);
1553 __io_cqring_fill_event(req, res, cflags);
1554 io_commit_cqring(ctx);
1556 * If we're the last reference to this request, add to our locked
1559 if (refcount_dec_and_test(&req->refs)) {
1560 struct io_comp_state *cs = &ctx->submit_state.comp;
1562 io_dismantle_req(req);
1563 io_put_task(req->task, 1);
1564 list_add(&req->compl.list, &cs->locked_free_list);
1565 cs->locked_free_nr++;
1568 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1570 io_cqring_ev_posted(ctx);
1573 percpu_ref_put(&ctx->refs);
1577 static void io_req_complete_state(struct io_kiocb *req, long res,
1578 unsigned int cflags)
1582 req->compl.cflags = cflags;
1583 req->flags |= REQ_F_COMPLETE_INLINE;
1586 static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags,
1587 long res, unsigned cflags)
1589 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1590 io_req_complete_state(req, res, cflags);
1592 io_req_complete_post(req, res, cflags);
1595 static inline void io_req_complete(struct io_kiocb *req, long res)
1597 __io_req_complete(req, 0, res, 0);
1600 static bool io_flush_cached_reqs(struct io_ring_ctx *ctx)
1602 struct io_submit_state *state = &ctx->submit_state;
1603 struct io_comp_state *cs = &state->comp;
1604 struct io_kiocb *req = NULL;
1607 * If we have more than a batch's worth of requests in our IRQ side
1608 * locked cache, grab the lock and move them over to our submission
1611 if (READ_ONCE(cs->locked_free_nr) > IO_COMPL_BATCH) {
1612 spin_lock_irq(&ctx->completion_lock);
1613 list_splice_init(&cs->locked_free_list, &cs->free_list);
1614 cs->locked_free_nr = 0;
1615 spin_unlock_irq(&ctx->completion_lock);
1618 while (!list_empty(&cs->free_list)) {
1619 req = list_first_entry(&cs->free_list, struct io_kiocb,
1621 list_del(&req->compl.list);
1622 state->reqs[state->free_reqs++] = req;
1623 if (state->free_reqs == ARRAY_SIZE(state->reqs))
1630 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
1632 struct io_submit_state *state = &ctx->submit_state;
1634 BUILD_BUG_ON(IO_REQ_ALLOC_BATCH > ARRAY_SIZE(state->reqs));
1636 if (!state->free_reqs) {
1637 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1640 if (io_flush_cached_reqs(ctx))
1643 ret = kmem_cache_alloc_bulk(req_cachep, gfp, IO_REQ_ALLOC_BATCH,
1647 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1648 * retry single alloc to be on the safe side.
1650 if (unlikely(ret <= 0)) {
1651 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1652 if (!state->reqs[0])
1656 state->free_reqs = ret;
1660 return state->reqs[state->free_reqs];
1663 static inline void io_put_file(struct io_kiocb *req, struct file *file,
1670 static void io_dismantle_req(struct io_kiocb *req)
1674 if (req->async_data)
1675 kfree(req->async_data);
1677 io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
1678 if (req->fixed_rsrc_refs)
1679 percpu_ref_put(req->fixed_rsrc_refs);
1680 io_req_clean_work(req);
1683 static inline void io_put_task(struct task_struct *task, int nr)
1685 struct io_uring_task *tctx = task->io_uring;
1687 percpu_counter_sub(&tctx->inflight, nr);
1688 if (unlikely(atomic_read(&tctx->in_idle)))
1689 wake_up(&tctx->wait);
1690 put_task_struct_many(task, nr);
1693 static void __io_free_req(struct io_kiocb *req)
1695 struct io_ring_ctx *ctx = req->ctx;
1697 io_dismantle_req(req);
1698 io_put_task(req->task, 1);
1700 kmem_cache_free(req_cachep, req);
1701 percpu_ref_put(&ctx->refs);
1704 static inline void io_remove_next_linked(struct io_kiocb *req)
1706 struct io_kiocb *nxt = req->link;
1708 req->link = nxt->link;
1712 static void io_kill_linked_timeout(struct io_kiocb *req)
1714 struct io_ring_ctx *ctx = req->ctx;
1715 struct io_kiocb *link;
1716 bool cancelled = false;
1717 unsigned long flags;
1719 spin_lock_irqsave(&ctx->completion_lock, flags);
1723 * Can happen if a linked timeout fired and link had been like
1724 * req -> link t-out -> link t-out [-> ...]
1726 if (link && (link->flags & REQ_F_LTIMEOUT_ACTIVE)) {
1727 struct io_timeout_data *io = link->async_data;
1730 io_remove_next_linked(req);
1731 link->timeout.head = NULL;
1732 ret = hrtimer_try_to_cancel(&io->timer);
1734 io_cqring_fill_event(link, -ECANCELED);
1735 io_commit_cqring(ctx);
1739 req->flags &= ~REQ_F_LINK_TIMEOUT;
1740 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1743 io_cqring_ev_posted(ctx);
1749 static void io_fail_links(struct io_kiocb *req)
1751 struct io_kiocb *link, *nxt;
1752 struct io_ring_ctx *ctx = req->ctx;
1753 unsigned long flags;
1755 spin_lock_irqsave(&ctx->completion_lock, flags);
1763 trace_io_uring_fail_link(req, link);
1764 io_cqring_fill_event(link, -ECANCELED);
1766 io_put_req_deferred(link, 2);
1769 io_commit_cqring(ctx);
1770 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1772 io_cqring_ev_posted(ctx);
1775 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1777 if (req->flags & REQ_F_LINK_TIMEOUT)
1778 io_kill_linked_timeout(req);
1781 * If LINK is set, we have dependent requests in this chain. If we
1782 * didn't fail this request, queue the first one up, moving any other
1783 * dependencies to the next request. In case of failure, fail the rest
1786 if (likely(!(req->flags & REQ_F_FAIL_LINK))) {
1787 struct io_kiocb *nxt = req->link;
1796 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1798 if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK))))
1800 return __io_req_find_next(req);
1803 static void ctx_flush_and_put(struct io_ring_ctx *ctx)
1807 if (ctx->submit_state.comp.nr) {
1808 mutex_lock(&ctx->uring_lock);
1809 io_submit_flush_completions(&ctx->submit_state.comp, ctx);
1810 mutex_unlock(&ctx->uring_lock);
1812 percpu_ref_put(&ctx->refs);
1815 static bool __tctx_task_work(struct io_uring_task *tctx)
1817 struct io_ring_ctx *ctx = NULL;
1818 struct io_wq_work_list list;
1819 struct io_wq_work_node *node;
1821 if (wq_list_empty(&tctx->task_list))
1824 spin_lock_irq(&tctx->task_lock);
1825 list = tctx->task_list;
1826 INIT_WQ_LIST(&tctx->task_list);
1827 spin_unlock_irq(&tctx->task_lock);
1831 struct io_wq_work_node *next = node->next;
1832 struct io_kiocb *req;
1834 req = container_of(node, struct io_kiocb, io_task_work.node);
1835 if (req->ctx != ctx) {
1836 ctx_flush_and_put(ctx);
1838 percpu_ref_get(&ctx->refs);
1841 req->task_work.func(&req->task_work);
1845 ctx_flush_and_put(ctx);
1846 return list.first != NULL;
1849 static void tctx_task_work(struct callback_head *cb)
1851 struct io_uring_task *tctx = container_of(cb, struct io_uring_task, task_work);
1853 clear_bit(0, &tctx->task_state);
1855 while (__tctx_task_work(tctx))
1859 static int io_task_work_add(struct task_struct *tsk, struct io_kiocb *req,
1860 enum task_work_notify_mode notify)
1862 struct io_uring_task *tctx = tsk->io_uring;
1863 struct io_wq_work_node *node, *prev;
1864 unsigned long flags;
1867 WARN_ON_ONCE(!tctx);
1869 spin_lock_irqsave(&tctx->task_lock, flags);
1870 wq_list_add_tail(&req->io_task_work.node, &tctx->task_list);
1871 spin_unlock_irqrestore(&tctx->task_lock, flags);
1873 /* task_work already pending, we're done */
1874 if (test_bit(0, &tctx->task_state) ||
1875 test_and_set_bit(0, &tctx->task_state))
1878 if (!task_work_add(tsk, &tctx->task_work, notify))
1882 * Slow path - we failed, find and delete work. if the work is not
1883 * in the list, it got run and we're fine.
1886 spin_lock_irqsave(&tctx->task_lock, flags);
1887 wq_list_for_each(node, prev, &tctx->task_list) {
1888 if (&req->io_task_work.node == node) {
1889 wq_list_del(&tctx->task_list, node, prev);
1894 spin_unlock_irqrestore(&tctx->task_lock, flags);
1895 clear_bit(0, &tctx->task_state);
1899 static int io_req_task_work_add(struct io_kiocb *req)
1901 struct task_struct *tsk = req->task;
1902 struct io_ring_ctx *ctx = req->ctx;
1903 enum task_work_notify_mode notify;
1906 if (tsk->flags & PF_EXITING)
1910 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
1911 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
1912 * processing task_work. There's no reliable way to tell if TWA_RESUME
1916 if (!(ctx->flags & IORING_SETUP_SQPOLL))
1917 notify = TWA_SIGNAL;
1919 ret = io_task_work_add(tsk, req, notify);
1921 wake_up_process(tsk);
1926 static void io_req_task_work_add_fallback(struct io_kiocb *req,
1927 task_work_func_t cb)
1929 struct io_ring_ctx *ctx = req->ctx;
1930 struct callback_head *head;
1932 init_task_work(&req->task_work, cb);
1934 head = READ_ONCE(ctx->exit_task_work);
1935 req->task_work.next = head;
1936 } while (cmpxchg(&ctx->exit_task_work, head, &req->task_work) != head);
1939 static void __io_req_task_cancel(struct io_kiocb *req, int error)
1941 struct io_ring_ctx *ctx = req->ctx;
1943 spin_lock_irq(&ctx->completion_lock);
1944 io_cqring_fill_event(req, error);
1945 io_commit_cqring(ctx);
1946 spin_unlock_irq(&ctx->completion_lock);
1948 io_cqring_ev_posted(ctx);
1949 req_set_fail_links(req);
1950 io_double_put_req(req);
1953 static void io_req_task_cancel(struct callback_head *cb)
1955 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1956 struct io_ring_ctx *ctx = req->ctx;
1958 mutex_lock(&ctx->uring_lock);
1959 __io_req_task_cancel(req, req->result);
1960 mutex_unlock(&ctx->uring_lock);
1961 percpu_ref_put(&ctx->refs);
1964 static void __io_req_task_submit(struct io_kiocb *req)
1966 struct io_ring_ctx *ctx = req->ctx;
1968 /* ctx stays valid until unlock, even if we drop all ours ctx->refs */
1969 mutex_lock(&ctx->uring_lock);
1970 if (!ctx->sqo_dead && !(current->flags & PF_EXITING) && !current->in_execve)
1971 __io_queue_sqe(req);
1973 __io_req_task_cancel(req, -EFAULT);
1974 mutex_unlock(&ctx->uring_lock);
1977 static void io_req_task_submit(struct callback_head *cb)
1979 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1981 __io_req_task_submit(req);
1984 static void io_req_task_queue(struct io_kiocb *req)
1988 req->task_work.func = io_req_task_submit;
1989 ret = io_req_task_work_add(req);
1990 if (unlikely(ret)) {
1991 req->result = -ECANCELED;
1992 percpu_ref_get(&req->ctx->refs);
1993 io_req_task_work_add_fallback(req, io_req_task_cancel);
1997 static void io_req_task_queue_fail(struct io_kiocb *req, int ret)
1999 percpu_ref_get(&req->ctx->refs);
2001 req->task_work.func = io_req_task_cancel;
2003 if (unlikely(io_req_task_work_add(req)))
2004 io_req_task_work_add_fallback(req, io_req_task_cancel);
2007 static inline void io_queue_next(struct io_kiocb *req)
2009 struct io_kiocb *nxt = io_req_find_next(req);
2012 io_req_task_queue(nxt);
2015 static void io_free_req(struct io_kiocb *req)
2022 struct task_struct *task;
2027 static inline void io_init_req_batch(struct req_batch *rb)
2034 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2035 struct req_batch *rb)
2038 io_put_task(rb->task, rb->task_refs);
2040 percpu_ref_put_many(&ctx->refs, rb->ctx_refs);
2043 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req,
2044 struct io_submit_state *state)
2048 if (req->task != rb->task) {
2050 io_put_task(rb->task, rb->task_refs);
2051 rb->task = req->task;
2057 io_dismantle_req(req);
2058 if (state->free_reqs != ARRAY_SIZE(state->reqs))
2059 state->reqs[state->free_reqs++] = req;
2061 list_add(&req->compl.list, &state->comp.free_list);
2064 static void io_submit_flush_completions(struct io_comp_state *cs,
2065 struct io_ring_ctx *ctx)
2068 struct io_kiocb *req;
2069 struct req_batch rb;
2071 io_init_req_batch(&rb);
2072 spin_lock_irq(&ctx->completion_lock);
2073 for (i = 0; i < nr; i++) {
2075 __io_cqring_fill_event(req, req->result, req->compl.cflags);
2077 io_commit_cqring(ctx);
2078 spin_unlock_irq(&ctx->completion_lock);
2080 io_cqring_ev_posted(ctx);
2081 for (i = 0; i < nr; i++) {
2084 /* submission and completion refs */
2085 if (refcount_sub_and_test(2, &req->refs))
2086 io_req_free_batch(&rb, req, &ctx->submit_state);
2089 io_req_free_batch_finish(ctx, &rb);
2094 * Drop reference to request, return next in chain (if there is one) if this
2095 * was the last reference to this request.
2097 static struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2099 struct io_kiocb *nxt = NULL;
2101 if (refcount_dec_and_test(&req->refs)) {
2102 nxt = io_req_find_next(req);
2108 static void io_put_req(struct io_kiocb *req)
2110 if (refcount_dec_and_test(&req->refs))
2114 static void io_put_req_deferred_cb(struct callback_head *cb)
2116 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2121 static void io_free_req_deferred(struct io_kiocb *req)
2125 req->task_work.func = io_put_req_deferred_cb;
2126 ret = io_req_task_work_add(req);
2128 io_req_task_work_add_fallback(req, io_put_req_deferred_cb);
2131 static inline void io_put_req_deferred(struct io_kiocb *req, int refs)
2133 if (refcount_sub_and_test(refs, &req->refs))
2134 io_free_req_deferred(req);
2137 static void io_double_put_req(struct io_kiocb *req)
2139 /* drop both submit and complete references */
2140 if (refcount_sub_and_test(2, &req->refs))
2144 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2146 /* See comment at the top of this file */
2148 return __io_cqring_events(ctx);
2151 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2153 struct io_rings *rings = ctx->rings;
2155 /* make sure SQ entry isn't read before tail */
2156 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2159 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2161 unsigned int cflags;
2163 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2164 cflags |= IORING_CQE_F_BUFFER;
2165 req->flags &= ~REQ_F_BUFFER_SELECTED;
2170 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2172 struct io_buffer *kbuf;
2174 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2175 return io_put_kbuf(req, kbuf);
2178 static inline bool io_run_task_work(void)
2181 * Not safe to run on exiting task, and the task_work handling will
2182 * not add work to such a task.
2184 if (unlikely(current->flags & PF_EXITING))
2186 if (current->task_works) {
2187 __set_current_state(TASK_RUNNING);
2196 * Find and free completed poll iocbs
2198 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2199 struct list_head *done)
2201 struct req_batch rb;
2202 struct io_kiocb *req;
2204 /* order with ->result store in io_complete_rw_iopoll() */
2207 io_init_req_batch(&rb);
2208 while (!list_empty(done)) {
2211 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2212 list_del(&req->inflight_entry);
2214 if (READ_ONCE(req->result) == -EAGAIN) {
2215 req->iopoll_completed = 0;
2216 if (io_rw_reissue(req))
2220 if (req->flags & REQ_F_BUFFER_SELECTED)
2221 cflags = io_put_rw_kbuf(req);
2223 __io_cqring_fill_event(req, req->result, cflags);
2226 if (refcount_dec_and_test(&req->refs))
2227 io_req_free_batch(&rb, req, &ctx->submit_state);
2230 io_commit_cqring(ctx);
2231 io_cqring_ev_posted_iopoll(ctx);
2232 io_req_free_batch_finish(ctx, &rb);
2235 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2238 struct io_kiocb *req, *tmp;
2244 * Only spin for completions if we don't have multiple devices hanging
2245 * off our complete list, and we're under the requested amount.
2247 spin = !ctx->poll_multi_file && *nr_events < min;
2250 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2251 struct kiocb *kiocb = &req->rw.kiocb;
2254 * Move completed and retryable entries to our local lists.
2255 * If we find a request that requires polling, break out
2256 * and complete those lists first, if we have entries there.
2258 if (READ_ONCE(req->iopoll_completed)) {
2259 list_move_tail(&req->inflight_entry, &done);
2262 if (!list_empty(&done))
2265 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2269 /* iopoll may have completed current req */
2270 if (READ_ONCE(req->iopoll_completed))
2271 list_move_tail(&req->inflight_entry, &done);
2278 if (!list_empty(&done))
2279 io_iopoll_complete(ctx, nr_events, &done);
2285 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2286 * non-spinning poll check - we'll still enter the driver poll loop, but only
2287 * as a non-spinning completion check.
2289 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
2292 while (!list_empty(&ctx->iopoll_list) && !need_resched()) {
2295 ret = io_do_iopoll(ctx, nr_events, min);
2298 if (*nr_events >= min)
2306 * We can't just wait for polled events to come to us, we have to actively
2307 * find and complete them.
2309 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2311 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2314 mutex_lock(&ctx->uring_lock);
2315 while (!list_empty(&ctx->iopoll_list)) {
2316 unsigned int nr_events = 0;
2318 io_do_iopoll(ctx, &nr_events, 0);
2320 /* let it sleep and repeat later if can't complete a request */
2324 * Ensure we allow local-to-the-cpu processing to take place,
2325 * in this case we need to ensure that we reap all events.
2326 * Also let task_work, etc. to progress by releasing the mutex
2328 if (need_resched()) {
2329 mutex_unlock(&ctx->uring_lock);
2331 mutex_lock(&ctx->uring_lock);
2334 mutex_unlock(&ctx->uring_lock);
2337 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2339 unsigned int nr_events = 0;
2340 int iters = 0, ret = 0;
2343 * We disallow the app entering submit/complete with polling, but we
2344 * still need to lock the ring to prevent racing with polled issue
2345 * that got punted to a workqueue.
2347 mutex_lock(&ctx->uring_lock);
2350 * Don't enter poll loop if we already have events pending.
2351 * If we do, we can potentially be spinning for commands that
2352 * already triggered a CQE (eg in error).
2354 if (test_bit(0, &ctx->cq_check_overflow))
2355 __io_cqring_overflow_flush(ctx, false, NULL, NULL);
2356 if (io_cqring_events(ctx))
2360 * If a submit got punted to a workqueue, we can have the
2361 * application entering polling for a command before it gets
2362 * issued. That app will hold the uring_lock for the duration
2363 * of the poll right here, so we need to take a breather every
2364 * now and then to ensure that the issue has a chance to add
2365 * the poll to the issued list. Otherwise we can spin here
2366 * forever, while the workqueue is stuck trying to acquire the
2369 if (!(++iters & 7)) {
2370 mutex_unlock(&ctx->uring_lock);
2372 mutex_lock(&ctx->uring_lock);
2375 ret = io_iopoll_getevents(ctx, &nr_events, min);
2379 } while (min && !nr_events && !need_resched());
2381 mutex_unlock(&ctx->uring_lock);
2385 static void kiocb_end_write(struct io_kiocb *req)
2388 * Tell lockdep we inherited freeze protection from submission
2391 if (req->flags & REQ_F_ISREG) {
2392 struct inode *inode = file_inode(req->file);
2394 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
2396 file_end_write(req->file);
2400 static bool io_resubmit_prep(struct io_kiocb *req)
2402 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2404 struct iov_iter iter;
2406 /* already prepared */
2407 if (req->async_data)
2410 switch (req->opcode) {
2411 case IORING_OP_READV:
2412 case IORING_OP_READ_FIXED:
2413 case IORING_OP_READ:
2416 case IORING_OP_WRITEV:
2417 case IORING_OP_WRITE_FIXED:
2418 case IORING_OP_WRITE:
2422 printk_once(KERN_WARNING "io_uring: bad opcode in resubmit %d\n",
2427 ret = io_import_iovec(rw, req, &iovec, &iter, false);
2430 return !io_setup_async_rw(req, iovec, inline_vecs, &iter, false);
2434 static bool io_rw_reissue(struct io_kiocb *req)
2437 umode_t mode = file_inode(req->file)->i_mode;
2439 if (!S_ISBLK(mode) && !S_ISREG(mode))
2441 if ((req->flags & REQ_F_NOWAIT) || io_wq_current_is_worker())
2444 * If ref is dying, we might be running poll reap from the exit work.
2445 * Don't attempt to reissue from that path, just let it fail with
2448 if (percpu_ref_is_dying(&req->ctx->refs))
2451 lockdep_assert_held(&req->ctx->uring_lock);
2453 if (io_resubmit_prep(req)) {
2454 refcount_inc(&req->refs);
2455 io_queue_async_work(req);
2458 req_set_fail_links(req);
2463 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2464 unsigned int issue_flags)
2468 if ((res == -EAGAIN || res == -EOPNOTSUPP) && io_rw_reissue(req))
2470 if (res != req->result)
2471 req_set_fail_links(req);
2473 if (req->rw.kiocb.ki_flags & IOCB_WRITE)
2474 kiocb_end_write(req);
2475 if (req->flags & REQ_F_BUFFER_SELECTED)
2476 cflags = io_put_rw_kbuf(req);
2477 __io_req_complete(req, issue_flags, res, cflags);
2480 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2482 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2484 __io_complete_rw(req, res, res2, 0);
2487 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2489 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2491 if (kiocb->ki_flags & IOCB_WRITE)
2492 kiocb_end_write(req);
2494 if (res != -EAGAIN && res != req->result)
2495 req_set_fail_links(req);
2497 WRITE_ONCE(req->result, res);
2498 /* order with io_poll_complete() checking ->result */
2500 WRITE_ONCE(req->iopoll_completed, 1);
2504 * After the iocb has been issued, it's safe to be found on the poll list.
2505 * Adding the kiocb to the list AFTER submission ensures that we don't
2506 * find it from a io_iopoll_getevents() thread before the issuer is done
2507 * accessing the kiocb cookie.
2509 static void io_iopoll_req_issued(struct io_kiocb *req, bool in_async)
2511 struct io_ring_ctx *ctx = req->ctx;
2514 * Track whether we have multiple files in our lists. This will impact
2515 * how we do polling eventually, not spinning if we're on potentially
2516 * different devices.
2518 if (list_empty(&ctx->iopoll_list)) {
2519 ctx->poll_multi_file = false;
2520 } else if (!ctx->poll_multi_file) {
2521 struct io_kiocb *list_req;
2523 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2525 if (list_req->file != req->file)
2526 ctx->poll_multi_file = true;
2530 * For fast devices, IO may have already completed. If it has, add
2531 * it to the front so we find it first.
2533 if (READ_ONCE(req->iopoll_completed))
2534 list_add(&req->inflight_entry, &ctx->iopoll_list);
2536 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2539 * If IORING_SETUP_SQPOLL is enabled, sqes are either handled in sq thread
2540 * task context or in io worker task context. If current task context is
2541 * sq thread, we don't need to check whether should wake up sq thread.
2543 if (in_async && (ctx->flags & IORING_SETUP_SQPOLL) &&
2544 wq_has_sleeper(&ctx->sq_data->wait))
2545 wake_up(&ctx->sq_data->wait);
2548 static inline void io_state_file_put(struct io_submit_state *state)
2550 if (state->file_refs) {
2551 fput_many(state->file, state->file_refs);
2552 state->file_refs = 0;
2557 * Get as many references to a file as we have IOs left in this submission,
2558 * assuming most submissions are for one file, or at least that each file
2559 * has more than one submission.
2561 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2566 if (state->file_refs) {
2567 if (state->fd == fd) {
2571 io_state_file_put(state);
2573 state->file = fget_many(fd, state->ios_left);
2574 if (unlikely(!state->file))
2578 state->file_refs = state->ios_left - 1;
2582 static bool io_bdev_nowait(struct block_device *bdev)
2584 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2588 * If we tracked the file through the SCM inflight mechanism, we could support
2589 * any file. For now, just ensure that anything potentially problematic is done
2592 static bool io_file_supports_async(struct file *file, int rw)
2594 umode_t mode = file_inode(file)->i_mode;
2596 if (S_ISBLK(mode)) {
2597 if (IS_ENABLED(CONFIG_BLOCK) &&
2598 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2602 if (S_ISCHR(mode) || S_ISSOCK(mode))
2604 if (S_ISREG(mode)) {
2605 if (IS_ENABLED(CONFIG_BLOCK) &&
2606 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2607 file->f_op != &io_uring_fops)
2612 /* any ->read/write should understand O_NONBLOCK */
2613 if (file->f_flags & O_NONBLOCK)
2616 if (!(file->f_mode & FMODE_NOWAIT))
2620 return file->f_op->read_iter != NULL;
2622 return file->f_op->write_iter != NULL;
2625 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2627 struct io_ring_ctx *ctx = req->ctx;
2628 struct kiocb *kiocb = &req->rw.kiocb;
2629 struct file *file = req->file;
2633 if (S_ISREG(file_inode(file)->i_mode))
2634 req->flags |= REQ_F_ISREG;
2636 kiocb->ki_pos = READ_ONCE(sqe->off);
2637 if (kiocb->ki_pos == -1 && !(file->f_mode & FMODE_STREAM)) {
2638 req->flags |= REQ_F_CUR_POS;
2639 kiocb->ki_pos = file->f_pos;
2641 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2642 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2643 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2647 /* don't allow async punt for O_NONBLOCK or RWF_NOWAIT */
2648 if ((kiocb->ki_flags & IOCB_NOWAIT) || (file->f_flags & O_NONBLOCK))
2649 req->flags |= REQ_F_NOWAIT;
2651 ioprio = READ_ONCE(sqe->ioprio);
2653 ret = ioprio_check_cap(ioprio);
2657 kiocb->ki_ioprio = ioprio;
2659 kiocb->ki_ioprio = get_current_ioprio();
2661 if (ctx->flags & IORING_SETUP_IOPOLL) {
2662 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2663 !kiocb->ki_filp->f_op->iopoll)
2666 kiocb->ki_flags |= IOCB_HIPRI;
2667 kiocb->ki_complete = io_complete_rw_iopoll;
2668 req->iopoll_completed = 0;
2670 if (kiocb->ki_flags & IOCB_HIPRI)
2672 kiocb->ki_complete = io_complete_rw;
2675 req->rw.addr = READ_ONCE(sqe->addr);
2676 req->rw.len = READ_ONCE(sqe->len);
2677 req->buf_index = READ_ONCE(sqe->buf_index);
2681 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2687 case -ERESTARTNOINTR:
2688 case -ERESTARTNOHAND:
2689 case -ERESTART_RESTARTBLOCK:
2691 * We can't just restart the syscall, since previously
2692 * submitted sqes may already be in progress. Just fail this
2698 kiocb->ki_complete(kiocb, ret, 0);
2702 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2703 unsigned int issue_flags)
2705 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2706 struct io_async_rw *io = req->async_data;
2708 /* add previously done IO, if any */
2709 if (io && io->bytes_done > 0) {
2711 ret = io->bytes_done;
2713 ret += io->bytes_done;
2716 if (req->flags & REQ_F_CUR_POS)
2717 req->file->f_pos = kiocb->ki_pos;
2718 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2719 __io_complete_rw(req, ret, 0, issue_flags);
2721 io_rw_done(kiocb, ret);
2724 static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
2726 struct io_ring_ctx *ctx = req->ctx;
2727 size_t len = req->rw.len;
2728 struct io_mapped_ubuf *imu;
2729 u16 index, buf_index = req->buf_index;
2733 if (unlikely(buf_index >= ctx->nr_user_bufs))
2735 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2736 imu = &ctx->user_bufs[index];
2737 buf_addr = req->rw.addr;
2740 if (buf_addr + len < buf_addr)
2742 /* not inside the mapped region */
2743 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2747 * May not be a start of buffer, set size appropriately
2748 * and advance us to the beginning.
2750 offset = buf_addr - imu->ubuf;
2751 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2755 * Don't use iov_iter_advance() here, as it's really slow for
2756 * using the latter parts of a big fixed buffer - it iterates
2757 * over each segment manually. We can cheat a bit here, because
2760 * 1) it's a BVEC iter, we set it up
2761 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2762 * first and last bvec
2764 * So just find our index, and adjust the iterator afterwards.
2765 * If the offset is within the first bvec (or the whole first
2766 * bvec, just use iov_iter_advance(). This makes it easier
2767 * since we can just skip the first segment, which may not
2768 * be PAGE_SIZE aligned.
2770 const struct bio_vec *bvec = imu->bvec;
2772 if (offset <= bvec->bv_len) {
2773 iov_iter_advance(iter, offset);
2775 unsigned long seg_skip;
2777 /* skip first vec */
2778 offset -= bvec->bv_len;
2779 seg_skip = 1 + (offset >> PAGE_SHIFT);
2781 iter->bvec = bvec + seg_skip;
2782 iter->nr_segs -= seg_skip;
2783 iter->count -= bvec->bv_len + offset;
2784 iter->iov_offset = offset & ~PAGE_MASK;
2791 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2794 mutex_unlock(&ctx->uring_lock);
2797 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2800 * "Normal" inline submissions always hold the uring_lock, since we
2801 * grab it from the system call. Same is true for the SQPOLL offload.
2802 * The only exception is when we've detached the request and issue it
2803 * from an async worker thread, grab the lock for that case.
2806 mutex_lock(&ctx->uring_lock);
2809 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2810 int bgid, struct io_buffer *kbuf,
2813 struct io_buffer *head;
2815 if (req->flags & REQ_F_BUFFER_SELECTED)
2818 io_ring_submit_lock(req->ctx, needs_lock);
2820 lockdep_assert_held(&req->ctx->uring_lock);
2822 head = idr_find(&req->ctx->io_buffer_idr, bgid);
2824 if (!list_empty(&head->list)) {
2825 kbuf = list_last_entry(&head->list, struct io_buffer,
2827 list_del(&kbuf->list);
2830 idr_remove(&req->ctx->io_buffer_idr, bgid);
2832 if (*len > kbuf->len)
2835 kbuf = ERR_PTR(-ENOBUFS);
2838 io_ring_submit_unlock(req->ctx, needs_lock);
2843 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2846 struct io_buffer *kbuf;
2849 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2850 bgid = req->buf_index;
2851 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2854 req->rw.addr = (u64) (unsigned long) kbuf;
2855 req->flags |= REQ_F_BUFFER_SELECTED;
2856 return u64_to_user_ptr(kbuf->addr);
2859 #ifdef CONFIG_COMPAT
2860 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2863 struct compat_iovec __user *uiov;
2864 compat_ssize_t clen;
2868 uiov = u64_to_user_ptr(req->rw.addr);
2869 if (!access_ok(uiov, sizeof(*uiov)))
2871 if (__get_user(clen, &uiov->iov_len))
2877 buf = io_rw_buffer_select(req, &len, needs_lock);
2879 return PTR_ERR(buf);
2880 iov[0].iov_base = buf;
2881 iov[0].iov_len = (compat_size_t) len;
2886 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2889 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2893 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2896 len = iov[0].iov_len;
2899 buf = io_rw_buffer_select(req, &len, needs_lock);
2901 return PTR_ERR(buf);
2902 iov[0].iov_base = buf;
2903 iov[0].iov_len = len;
2907 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2910 if (req->flags & REQ_F_BUFFER_SELECTED) {
2911 struct io_buffer *kbuf;
2913 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2914 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
2915 iov[0].iov_len = kbuf->len;
2918 if (req->rw.len != 1)
2921 #ifdef CONFIG_COMPAT
2922 if (req->ctx->compat)
2923 return io_compat_import(req, iov, needs_lock);
2926 return __io_iov_buffer_select(req, iov, needs_lock);
2929 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
2930 struct iov_iter *iter, bool needs_lock)
2932 void __user *buf = u64_to_user_ptr(req->rw.addr);
2933 size_t sqe_len = req->rw.len;
2934 u8 opcode = req->opcode;
2937 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
2939 return io_import_fixed(req, rw, iter);
2942 /* buffer index only valid with fixed read/write, or buffer select */
2943 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
2946 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
2947 if (req->flags & REQ_F_BUFFER_SELECT) {
2948 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
2950 return PTR_ERR(buf);
2951 req->rw.len = sqe_len;
2954 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
2959 if (req->flags & REQ_F_BUFFER_SELECT) {
2960 ret = io_iov_buffer_select(req, *iovec, needs_lock);
2962 iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
2967 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
2971 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
2973 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
2977 * For files that don't have ->read_iter() and ->write_iter(), handle them
2978 * by looping over ->read() or ->write() manually.
2980 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
2982 struct kiocb *kiocb = &req->rw.kiocb;
2983 struct file *file = req->file;
2987 * Don't support polled IO through this interface, and we can't
2988 * support non-blocking either. For the latter, this just causes
2989 * the kiocb to be handled from an async context.
2991 if (kiocb->ki_flags & IOCB_HIPRI)
2993 if (kiocb->ki_flags & IOCB_NOWAIT)
2996 while (iov_iter_count(iter)) {
3000 if (!iov_iter_is_bvec(iter)) {
3001 iovec = iov_iter_iovec(iter);
3003 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3004 iovec.iov_len = req->rw.len;
3008 nr = file->f_op->read(file, iovec.iov_base,
3009 iovec.iov_len, io_kiocb_ppos(kiocb));
3011 nr = file->f_op->write(file, iovec.iov_base,
3012 iovec.iov_len, io_kiocb_ppos(kiocb));
3021 if (nr != iovec.iov_len)
3025 iov_iter_advance(iter, nr);
3031 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3032 const struct iovec *fast_iov, struct iov_iter *iter)
3034 struct io_async_rw *rw = req->async_data;
3036 memcpy(&rw->iter, iter, sizeof(*iter));
3037 rw->free_iovec = iovec;
3039 /* can only be fixed buffers, no need to do anything */
3040 if (iov_iter_is_bvec(iter))
3043 unsigned iov_off = 0;
3045 rw->iter.iov = rw->fast_iov;
3046 if (iter->iov != fast_iov) {
3047 iov_off = iter->iov - fast_iov;
3048 rw->iter.iov += iov_off;
3050 if (rw->fast_iov != fast_iov)
3051 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3052 sizeof(struct iovec) * iter->nr_segs);
3054 req->flags |= REQ_F_NEED_CLEANUP;
3058 static inline int __io_alloc_async_data(struct io_kiocb *req)
3060 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3061 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3062 return req->async_data == NULL;
3065 static int io_alloc_async_data(struct io_kiocb *req)
3067 if (!io_op_defs[req->opcode].needs_async_data)
3070 return __io_alloc_async_data(req);
3073 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3074 const struct iovec *fast_iov,
3075 struct iov_iter *iter, bool force)
3077 if (!force && !io_op_defs[req->opcode].needs_async_data)
3079 if (!req->async_data) {
3080 if (__io_alloc_async_data(req)) {
3085 io_req_map_rw(req, iovec, fast_iov, iter);
3090 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3092 struct io_async_rw *iorw = req->async_data;
3093 struct iovec *iov = iorw->fast_iov;
3096 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3097 if (unlikely(ret < 0))
3100 iorw->bytes_done = 0;
3101 iorw->free_iovec = iov;
3103 req->flags |= REQ_F_NEED_CLEANUP;
3107 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3109 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3111 return io_prep_rw(req, sqe);
3115 * This is our waitqueue callback handler, registered through lock_page_async()
3116 * when we initially tried to do the IO with the iocb armed our waitqueue.
3117 * This gets called when the page is unlocked, and we generally expect that to
3118 * happen when the page IO is completed and the page is now uptodate. This will
3119 * queue a task_work based retry of the operation, attempting to copy the data
3120 * again. If the latter fails because the page was NOT uptodate, then we will
3121 * do a thread based blocking retry of the operation. That's the unexpected
3124 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3125 int sync, void *arg)
3127 struct wait_page_queue *wpq;
3128 struct io_kiocb *req = wait->private;
3129 struct wait_page_key *key = arg;
3131 wpq = container_of(wait, struct wait_page_queue, wait);
3133 if (!wake_page_match(wpq, key))
3136 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3137 list_del_init(&wait->entry);
3139 /* submit ref gets dropped, acquire a new one */
3140 refcount_inc(&req->refs);
3141 io_req_task_queue(req);
3146 * This controls whether a given IO request should be armed for async page
3147 * based retry. If we return false here, the request is handed to the async
3148 * worker threads for retry. If we're doing buffered reads on a regular file,
3149 * we prepare a private wait_page_queue entry and retry the operation. This
3150 * will either succeed because the page is now uptodate and unlocked, or it
3151 * will register a callback when the page is unlocked at IO completion. Through
3152 * that callback, io_uring uses task_work to setup a retry of the operation.
3153 * That retry will attempt the buffered read again. The retry will generally
3154 * succeed, or in rare cases where it fails, we then fall back to using the
3155 * async worker threads for a blocking retry.
3157 static bool io_rw_should_retry(struct io_kiocb *req)
3159 struct io_async_rw *rw = req->async_data;
3160 struct wait_page_queue *wait = &rw->wpq;
3161 struct kiocb *kiocb = &req->rw.kiocb;
3163 /* never retry for NOWAIT, we just complete with -EAGAIN */
3164 if (req->flags & REQ_F_NOWAIT)
3167 /* Only for buffered IO */
3168 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3172 * just use poll if we can, and don't attempt if the fs doesn't
3173 * support callback based unlocks
3175 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3178 wait->wait.func = io_async_buf_func;
3179 wait->wait.private = req;
3180 wait->wait.flags = 0;
3181 INIT_LIST_HEAD(&wait->wait.entry);
3182 kiocb->ki_flags |= IOCB_WAITQ;
3183 kiocb->ki_flags &= ~IOCB_NOWAIT;
3184 kiocb->ki_waitq = wait;
3188 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3190 if (req->file->f_op->read_iter)
3191 return call_read_iter(req->file, &req->rw.kiocb, iter);
3192 else if (req->file->f_op->read)
3193 return loop_rw_iter(READ, req, iter);
3198 static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3200 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3201 struct kiocb *kiocb = &req->rw.kiocb;
3202 struct iov_iter __iter, *iter = &__iter;
3203 struct io_async_rw *rw = req->async_data;
3204 ssize_t io_size, ret, ret2;
3205 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3211 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3215 io_size = iov_iter_count(iter);
3216 req->result = io_size;
3218 /* Ensure we clear previously set non-block flag */
3219 if (!force_nonblock)
3220 kiocb->ki_flags &= ~IOCB_NOWAIT;
3222 kiocb->ki_flags |= IOCB_NOWAIT;
3224 /* If the file doesn't support async, just async punt */
3225 if (force_nonblock && !io_file_supports_async(req->file, READ)) {
3226 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3227 return ret ?: -EAGAIN;
3230 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size);
3231 if (unlikely(ret)) {
3236 ret = io_iter_do_read(req, iter);
3238 if (ret == -EIOCBQUEUED) {
3240 } else if (ret == -EAGAIN) {
3241 /* IOPOLL retry should happen for io-wq threads */
3242 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3244 /* no retry on NONBLOCK nor RWF_NOWAIT */
3245 if (req->flags & REQ_F_NOWAIT)
3247 /* some cases will consume bytes even on error returns */
3248 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3250 } else if (ret <= 0 || ret == io_size || !force_nonblock ||
3251 (req->flags & REQ_F_NOWAIT) || !(req->flags & REQ_F_ISREG)) {
3252 /* read all, failed, already did sync or don't want to retry */
3256 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3261 rw = req->async_data;
3262 /* now use our persistent iterator, if we aren't already */
3267 rw->bytes_done += ret;
3268 /* if we can retry, do so with the callbacks armed */
3269 if (!io_rw_should_retry(req)) {
3270 kiocb->ki_flags &= ~IOCB_WAITQ;
3275 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3276 * we get -EIOCBQUEUED, then we'll get a notification when the
3277 * desired page gets unlocked. We can also get a partial read
3278 * here, and if we do, then just retry at the new offset.
3280 ret = io_iter_do_read(req, iter);
3281 if (ret == -EIOCBQUEUED)
3283 /* we got some bytes, but not all. retry. */
3284 } while (ret > 0 && ret < io_size);
3286 kiocb_done(kiocb, ret, issue_flags);
3288 /* it's faster to check here then delegate to kfree */
3294 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3296 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3298 return io_prep_rw(req, sqe);
3301 static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3303 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3304 struct kiocb *kiocb = &req->rw.kiocb;
3305 struct iov_iter __iter, *iter = &__iter;
3306 struct io_async_rw *rw = req->async_data;
3307 ssize_t ret, ret2, io_size;
3308 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3314 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3318 io_size = iov_iter_count(iter);
3319 req->result = io_size;
3321 /* Ensure we clear previously set non-block flag */
3322 if (!force_nonblock)
3323 kiocb->ki_flags &= ~IOCB_NOWAIT;
3325 kiocb->ki_flags |= IOCB_NOWAIT;
3327 /* If the file doesn't support async, just async punt */
3328 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
3331 /* file path doesn't support NOWAIT for non-direct_IO */
3332 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3333 (req->flags & REQ_F_ISREG))
3336 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size);
3341 * Open-code file_start_write here to grab freeze protection,
3342 * which will be released by another thread in
3343 * io_complete_rw(). Fool lockdep by telling it the lock got
3344 * released so that it doesn't complain about the held lock when
3345 * we return to userspace.
3347 if (req->flags & REQ_F_ISREG) {
3348 sb_start_write(file_inode(req->file)->i_sb);
3349 __sb_writers_release(file_inode(req->file)->i_sb,
3352 kiocb->ki_flags |= IOCB_WRITE;
3354 if (req->file->f_op->write_iter)
3355 ret2 = call_write_iter(req->file, kiocb, iter);
3356 else if (req->file->f_op->write)
3357 ret2 = loop_rw_iter(WRITE, req, iter);
3362 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3363 * retry them without IOCB_NOWAIT.
3365 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3367 /* no retry on NONBLOCK nor RWF_NOWAIT */
3368 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3370 if (!force_nonblock || ret2 != -EAGAIN) {
3371 /* IOPOLL retry should happen for io-wq threads */
3372 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3375 kiocb_done(kiocb, ret2, issue_flags);
3378 /* some cases will consume bytes even on error returns */
3379 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3380 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3381 return ret ?: -EAGAIN;
3384 /* it's reportedly faster than delegating the null check to kfree() */
3390 static int io_renameat_prep(struct io_kiocb *req,
3391 const struct io_uring_sqe *sqe)
3393 struct io_rename *ren = &req->rename;
3394 const char __user *oldf, *newf;
3396 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3399 ren->old_dfd = READ_ONCE(sqe->fd);
3400 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3401 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3402 ren->new_dfd = READ_ONCE(sqe->len);
3403 ren->flags = READ_ONCE(sqe->rename_flags);
3405 ren->oldpath = getname(oldf);
3406 if (IS_ERR(ren->oldpath))
3407 return PTR_ERR(ren->oldpath);
3409 ren->newpath = getname(newf);
3410 if (IS_ERR(ren->newpath)) {
3411 putname(ren->oldpath);
3412 return PTR_ERR(ren->newpath);
3415 req->flags |= REQ_F_NEED_CLEANUP;
3419 static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3421 struct io_rename *ren = &req->rename;
3424 if (issue_flags & IO_URING_F_NONBLOCK)
3427 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3428 ren->newpath, ren->flags);
3430 req->flags &= ~REQ_F_NEED_CLEANUP;
3432 req_set_fail_links(req);
3433 io_req_complete(req, ret);
3437 static int io_unlinkat_prep(struct io_kiocb *req,
3438 const struct io_uring_sqe *sqe)
3440 struct io_unlink *un = &req->unlink;
3441 const char __user *fname;
3443 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3446 un->dfd = READ_ONCE(sqe->fd);
3448 un->flags = READ_ONCE(sqe->unlink_flags);
3449 if (un->flags & ~AT_REMOVEDIR)
3452 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3453 un->filename = getname(fname);
3454 if (IS_ERR(un->filename))
3455 return PTR_ERR(un->filename);
3457 req->flags |= REQ_F_NEED_CLEANUP;
3461 static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3463 struct io_unlink *un = &req->unlink;
3466 if (issue_flags & IO_URING_F_NONBLOCK)
3469 if (un->flags & AT_REMOVEDIR)
3470 ret = do_rmdir(un->dfd, un->filename);
3472 ret = do_unlinkat(un->dfd, un->filename);
3474 req->flags &= ~REQ_F_NEED_CLEANUP;
3476 req_set_fail_links(req);
3477 io_req_complete(req, ret);
3481 static int io_shutdown_prep(struct io_kiocb *req,
3482 const struct io_uring_sqe *sqe)
3484 #if defined(CONFIG_NET)
3485 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3487 if (sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3491 req->shutdown.how = READ_ONCE(sqe->len);
3498 static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3500 #if defined(CONFIG_NET)
3501 struct socket *sock;
3504 if (issue_flags & IO_URING_F_NONBLOCK)
3507 sock = sock_from_file(req->file);
3508 if (unlikely(!sock))
3511 ret = __sys_shutdown_sock(sock, req->shutdown.how);
3513 req_set_fail_links(req);
3514 io_req_complete(req, ret);
3521 static int __io_splice_prep(struct io_kiocb *req,
3522 const struct io_uring_sqe *sqe)
3524 struct io_splice* sp = &req->splice;
3525 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3527 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3531 sp->len = READ_ONCE(sqe->len);
3532 sp->flags = READ_ONCE(sqe->splice_flags);
3534 if (unlikely(sp->flags & ~valid_flags))
3537 sp->file_in = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in),
3538 (sp->flags & SPLICE_F_FD_IN_FIXED));
3541 req->flags |= REQ_F_NEED_CLEANUP;
3543 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
3545 * Splice operation will be punted aync, and here need to
3546 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3548 req->work.flags |= IO_WQ_WORK_UNBOUND;
3554 static int io_tee_prep(struct io_kiocb *req,
3555 const struct io_uring_sqe *sqe)
3557 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3559 return __io_splice_prep(req, sqe);
3562 static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
3564 struct io_splice *sp = &req->splice;
3565 struct file *in = sp->file_in;
3566 struct file *out = sp->file_out;
3567 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3570 if (issue_flags & IO_URING_F_NONBLOCK)
3573 ret = do_tee(in, out, sp->len, flags);
3575 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3576 req->flags &= ~REQ_F_NEED_CLEANUP;
3579 req_set_fail_links(req);
3580 io_req_complete(req, ret);
3584 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3586 struct io_splice* sp = &req->splice;
3588 sp->off_in = READ_ONCE(sqe->splice_off_in);
3589 sp->off_out = READ_ONCE(sqe->off);
3590 return __io_splice_prep(req, sqe);
3593 static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
3595 struct io_splice *sp = &req->splice;
3596 struct file *in = sp->file_in;
3597 struct file *out = sp->file_out;
3598 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3599 loff_t *poff_in, *poff_out;
3602 if (issue_flags & IO_URING_F_NONBLOCK)
3605 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3606 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3609 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3611 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3612 req->flags &= ~REQ_F_NEED_CLEANUP;
3615 req_set_fail_links(req);
3616 io_req_complete(req, ret);
3621 * IORING_OP_NOP just posts a completion event, nothing else.
3623 static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
3625 struct io_ring_ctx *ctx = req->ctx;
3627 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3630 __io_req_complete(req, issue_flags, 0, 0);
3634 static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3636 struct io_ring_ctx *ctx = req->ctx;
3641 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3643 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3646 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3647 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3650 req->sync.off = READ_ONCE(sqe->off);
3651 req->sync.len = READ_ONCE(sqe->len);
3655 static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
3657 loff_t end = req->sync.off + req->sync.len;
3660 /* fsync always requires a blocking context */
3661 if (issue_flags & IO_URING_F_NONBLOCK)
3664 ret = vfs_fsync_range(req->file, req->sync.off,
3665 end > 0 ? end : LLONG_MAX,
3666 req->sync.flags & IORING_FSYNC_DATASYNC);
3668 req_set_fail_links(req);
3669 io_req_complete(req, ret);
3673 static int io_fallocate_prep(struct io_kiocb *req,
3674 const struct io_uring_sqe *sqe)
3676 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3678 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3681 req->sync.off = READ_ONCE(sqe->off);
3682 req->sync.len = READ_ONCE(sqe->addr);
3683 req->sync.mode = READ_ONCE(sqe->len);
3687 static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
3691 /* fallocate always requiring blocking context */
3692 if (issue_flags & IO_URING_F_NONBLOCK)
3694 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3697 req_set_fail_links(req);
3698 io_req_complete(req, ret);
3702 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3704 const char __user *fname;
3707 if (unlikely(sqe->ioprio || sqe->buf_index))
3709 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3712 /* open.how should be already initialised */
3713 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3714 req->open.how.flags |= O_LARGEFILE;
3716 req->open.dfd = READ_ONCE(sqe->fd);
3717 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3718 req->open.filename = getname(fname);
3719 if (IS_ERR(req->open.filename)) {
3720 ret = PTR_ERR(req->open.filename);
3721 req->open.filename = NULL;
3724 req->open.nofile = rlimit(RLIMIT_NOFILE);
3725 req->flags |= REQ_F_NEED_CLEANUP;
3729 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3733 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3735 mode = READ_ONCE(sqe->len);
3736 flags = READ_ONCE(sqe->open_flags);
3737 req->open.how = build_open_how(flags, mode);
3738 return __io_openat_prep(req, sqe);
3741 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3743 struct open_how __user *how;
3747 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3749 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3750 len = READ_ONCE(sqe->len);
3751 if (len < OPEN_HOW_SIZE_VER0)
3754 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3759 return __io_openat_prep(req, sqe);
3762 static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
3764 struct open_flags op;
3767 bool resolve_nonblock;
3770 ret = build_open_flags(&req->open.how, &op);
3773 nonblock_set = op.open_flag & O_NONBLOCK;
3774 resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
3775 if (issue_flags & IO_URING_F_NONBLOCK) {
3777 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
3778 * it'll always -EAGAIN
3780 if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
3782 op.lookup_flags |= LOOKUP_CACHED;
3783 op.open_flag |= O_NONBLOCK;
3786 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3790 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3791 /* only retry if RESOLVE_CACHED wasn't already set by application */
3792 if ((!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)) &&
3793 file == ERR_PTR(-EAGAIN)) {
3795 * We could hang on to this 'fd', but seems like marginal
3796 * gain for something that is now known to be a slower path.
3797 * So just put it, and we'll get a new one when we retry.
3805 ret = PTR_ERR(file);
3807 if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
3808 file->f_flags &= ~O_NONBLOCK;
3809 fsnotify_open(file);
3810 fd_install(ret, file);
3813 putname(req->open.filename);
3814 req->flags &= ~REQ_F_NEED_CLEANUP;
3816 req_set_fail_links(req);
3817 io_req_complete(req, ret);
3821 static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
3823 return io_openat2(req, issue_flags & IO_URING_F_NONBLOCK);
3826 static int io_remove_buffers_prep(struct io_kiocb *req,
3827 const struct io_uring_sqe *sqe)
3829 struct io_provide_buf *p = &req->pbuf;
3832 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3835 tmp = READ_ONCE(sqe->fd);
3836 if (!tmp || tmp > USHRT_MAX)
3839 memset(p, 0, sizeof(*p));
3841 p->bgid = READ_ONCE(sqe->buf_group);
3845 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3846 int bgid, unsigned nbufs)
3850 /* shouldn't happen */
3854 /* the head kbuf is the list itself */
3855 while (!list_empty(&buf->list)) {
3856 struct io_buffer *nxt;
3858 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3859 list_del(&nxt->list);
3866 idr_remove(&ctx->io_buffer_idr, bgid);
3871 static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
3873 struct io_provide_buf *p = &req->pbuf;
3874 struct io_ring_ctx *ctx = req->ctx;
3875 struct io_buffer *head;
3877 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3879 io_ring_submit_lock(ctx, !force_nonblock);
3881 lockdep_assert_held(&ctx->uring_lock);
3884 head = idr_find(&ctx->io_buffer_idr, p->bgid);
3886 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3888 req_set_fail_links(req);
3890 /* need to hold the lock to complete IOPOLL requests */
3891 if (ctx->flags & IORING_SETUP_IOPOLL) {
3892 __io_req_complete(req, issue_flags, ret, 0);
3893 io_ring_submit_unlock(ctx, !force_nonblock);
3895 io_ring_submit_unlock(ctx, !force_nonblock);
3896 __io_req_complete(req, issue_flags, ret, 0);
3901 static int io_provide_buffers_prep(struct io_kiocb *req,
3902 const struct io_uring_sqe *sqe)
3904 struct io_provide_buf *p = &req->pbuf;
3907 if (sqe->ioprio || sqe->rw_flags)
3910 tmp = READ_ONCE(sqe->fd);
3911 if (!tmp || tmp > USHRT_MAX)
3914 p->addr = READ_ONCE(sqe->addr);
3915 p->len = READ_ONCE(sqe->len);
3917 if (!access_ok(u64_to_user_ptr(p->addr), (p->len * p->nbufs)))
3920 p->bgid = READ_ONCE(sqe->buf_group);
3921 tmp = READ_ONCE(sqe->off);
3922 if (tmp > USHRT_MAX)
3928 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3930 struct io_buffer *buf;
3931 u64 addr = pbuf->addr;
3932 int i, bid = pbuf->bid;
3934 for (i = 0; i < pbuf->nbufs; i++) {
3935 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
3940 buf->len = pbuf->len;
3945 INIT_LIST_HEAD(&buf->list);
3948 list_add_tail(&buf->list, &(*head)->list);
3952 return i ? i : -ENOMEM;
3955 static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
3957 struct io_provide_buf *p = &req->pbuf;
3958 struct io_ring_ctx *ctx = req->ctx;
3959 struct io_buffer *head, *list;
3961 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3963 io_ring_submit_lock(ctx, !force_nonblock);
3965 lockdep_assert_held(&ctx->uring_lock);
3967 list = head = idr_find(&ctx->io_buffer_idr, p->bgid);
3969 ret = io_add_buffers(p, &head);
3974 ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1,
3977 __io_remove_buffers(ctx, head, p->bgid, -1U);
3983 req_set_fail_links(req);
3985 /* need to hold the lock to complete IOPOLL requests */
3986 if (ctx->flags & IORING_SETUP_IOPOLL) {
3987 __io_req_complete(req, issue_flags, ret, 0);
3988 io_ring_submit_unlock(ctx, !force_nonblock);
3990 io_ring_submit_unlock(ctx, !force_nonblock);
3991 __io_req_complete(req, issue_flags, ret, 0);
3996 static int io_epoll_ctl_prep(struct io_kiocb *req,
3997 const struct io_uring_sqe *sqe)
3999 #if defined(CONFIG_EPOLL)
4000 if (sqe->ioprio || sqe->buf_index)
4002 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4005 req->epoll.epfd = READ_ONCE(sqe->fd);
4006 req->epoll.op = READ_ONCE(sqe->len);
4007 req->epoll.fd = READ_ONCE(sqe->off);
4009 if (ep_op_has_event(req->epoll.op)) {
4010 struct epoll_event __user *ev;
4012 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4013 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4023 static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4025 #if defined(CONFIG_EPOLL)
4026 struct io_epoll *ie = &req->epoll;
4028 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4030 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4031 if (force_nonblock && ret == -EAGAIN)
4035 req_set_fail_links(req);
4036 __io_req_complete(req, issue_flags, ret, 0);
4043 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4045 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4046 if (sqe->ioprio || sqe->buf_index || sqe->off)
4048 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4051 req->madvise.addr = READ_ONCE(sqe->addr);
4052 req->madvise.len = READ_ONCE(sqe->len);
4053 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4060 static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4062 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4063 struct io_madvise *ma = &req->madvise;
4066 if (issue_flags & IO_URING_F_NONBLOCK)
4069 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4071 req_set_fail_links(req);
4072 io_req_complete(req, ret);
4079 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4081 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4083 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4086 req->fadvise.offset = READ_ONCE(sqe->off);
4087 req->fadvise.len = READ_ONCE(sqe->len);
4088 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4092 static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4094 struct io_fadvise *fa = &req->fadvise;
4097 if (issue_flags & IO_URING_F_NONBLOCK) {
4098 switch (fa->advice) {
4099 case POSIX_FADV_NORMAL:
4100 case POSIX_FADV_RANDOM:
4101 case POSIX_FADV_SEQUENTIAL:
4108 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4110 req_set_fail_links(req);
4111 io_req_complete(req, ret);
4115 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4117 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4119 if (sqe->ioprio || sqe->buf_index)
4121 if (req->flags & REQ_F_FIXED_FILE)
4124 req->statx.dfd = READ_ONCE(sqe->fd);
4125 req->statx.mask = READ_ONCE(sqe->len);
4126 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4127 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4128 req->statx.flags = READ_ONCE(sqe->statx_flags);
4133 static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4135 struct io_statx *ctx = &req->statx;
4138 if (issue_flags & IO_URING_F_NONBLOCK) {
4139 /* only need file table for an actual valid fd */
4140 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
4141 req->flags |= REQ_F_NO_FILE_TABLE;
4145 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4149 req_set_fail_links(req);
4150 io_req_complete(req, ret);
4154 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4156 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4158 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4159 sqe->rw_flags || sqe->buf_index)
4161 if (req->flags & REQ_F_FIXED_FILE)
4164 req->close.fd = READ_ONCE(sqe->fd);
4168 static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4170 struct files_struct *files = current->files;
4171 struct io_close *close = &req->close;
4172 struct fdtable *fdt;
4178 spin_lock(&files->file_lock);
4179 fdt = files_fdtable(files);
4180 if (close->fd >= fdt->max_fds) {
4181 spin_unlock(&files->file_lock);
4184 file = fdt->fd[close->fd];
4186 spin_unlock(&files->file_lock);
4190 if (file->f_op == &io_uring_fops) {
4191 spin_unlock(&files->file_lock);
4196 /* if the file has a flush method, be safe and punt to async */
4197 if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4198 spin_unlock(&files->file_lock);
4202 ret = __close_fd_get_file(close->fd, &file);
4203 spin_unlock(&files->file_lock);
4210 /* No ->flush() or already async, safely close from here */
4211 ret = filp_close(file, current->files);
4214 req_set_fail_links(req);
4217 __io_req_complete(req, issue_flags, ret, 0);
4221 static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4223 struct io_ring_ctx *ctx = req->ctx;
4225 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4227 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4230 req->sync.off = READ_ONCE(sqe->off);
4231 req->sync.len = READ_ONCE(sqe->len);
4232 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4236 static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4240 /* sync_file_range always requires a blocking context */
4241 if (issue_flags & IO_URING_F_NONBLOCK)
4244 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4247 req_set_fail_links(req);
4248 io_req_complete(req, ret);
4252 #if defined(CONFIG_NET)
4253 static int io_setup_async_msg(struct io_kiocb *req,
4254 struct io_async_msghdr *kmsg)
4256 struct io_async_msghdr *async_msg = req->async_data;
4260 if (io_alloc_async_data(req)) {
4261 kfree(kmsg->free_iov);
4264 async_msg = req->async_data;
4265 req->flags |= REQ_F_NEED_CLEANUP;
4266 memcpy(async_msg, kmsg, sizeof(*kmsg));
4267 async_msg->msg.msg_name = &async_msg->addr;
4268 /* if were using fast_iov, set it to the new one */
4269 if (!async_msg->free_iov)
4270 async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4275 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4276 struct io_async_msghdr *iomsg)
4278 iomsg->msg.msg_name = &iomsg->addr;
4279 iomsg->free_iov = iomsg->fast_iov;
4280 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4281 req->sr_msg.msg_flags, &iomsg->free_iov);
4284 static int io_sendmsg_prep_async(struct io_kiocb *req)
4288 if (!io_op_defs[req->opcode].needs_async_data)
4290 ret = io_sendmsg_copy_hdr(req, req->async_data);
4292 req->flags |= REQ_F_NEED_CLEANUP;
4296 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4298 struct io_sr_msg *sr = &req->sr_msg;
4300 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4303 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4304 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4305 sr->len = READ_ONCE(sqe->len);
4307 #ifdef CONFIG_COMPAT
4308 if (req->ctx->compat)
4309 sr->msg_flags |= MSG_CMSG_COMPAT;
4314 static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4316 struct io_async_msghdr iomsg, *kmsg;
4317 struct socket *sock;
4321 sock = sock_from_file(req->file);
4322 if (unlikely(!sock))
4325 kmsg = req->async_data;
4327 ret = io_sendmsg_copy_hdr(req, &iomsg);
4333 flags = req->sr_msg.msg_flags;
4334 if (flags & MSG_DONTWAIT)
4335 req->flags |= REQ_F_NOWAIT;
4336 else if (issue_flags & IO_URING_F_NONBLOCK)
4337 flags |= MSG_DONTWAIT;
4339 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4340 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4341 return io_setup_async_msg(req, kmsg);
4342 if (ret == -ERESTARTSYS)
4345 /* fast path, check for non-NULL to avoid function call */
4347 kfree(kmsg->free_iov);
4348 req->flags &= ~REQ_F_NEED_CLEANUP;
4350 req_set_fail_links(req);
4351 __io_req_complete(req, issue_flags, ret, 0);
4355 static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4357 struct io_sr_msg *sr = &req->sr_msg;
4360 struct socket *sock;
4364 sock = sock_from_file(req->file);
4365 if (unlikely(!sock))
4368 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4372 msg.msg_name = NULL;
4373 msg.msg_control = NULL;
4374 msg.msg_controllen = 0;
4375 msg.msg_namelen = 0;
4377 flags = req->sr_msg.msg_flags;
4378 if (flags & MSG_DONTWAIT)
4379 req->flags |= REQ_F_NOWAIT;
4380 else if (issue_flags & IO_URING_F_NONBLOCK)
4381 flags |= MSG_DONTWAIT;
4383 msg.msg_flags = flags;
4384 ret = sock_sendmsg(sock, &msg);
4385 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4387 if (ret == -ERESTARTSYS)
4391 req_set_fail_links(req);
4392 __io_req_complete(req, issue_flags, ret, 0);
4396 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4397 struct io_async_msghdr *iomsg)
4399 struct io_sr_msg *sr = &req->sr_msg;
4400 struct iovec __user *uiov;
4404 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4405 &iomsg->uaddr, &uiov, &iov_len);
4409 if (req->flags & REQ_F_BUFFER_SELECT) {
4412 if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4414 sr->len = iomsg->fast_iov[0].iov_len;
4415 iomsg->free_iov = NULL;
4417 iomsg->free_iov = iomsg->fast_iov;
4418 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4419 &iomsg->free_iov, &iomsg->msg.msg_iter,
4428 #ifdef CONFIG_COMPAT
4429 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4430 struct io_async_msghdr *iomsg)
4432 struct compat_msghdr __user *msg_compat;
4433 struct io_sr_msg *sr = &req->sr_msg;
4434 struct compat_iovec __user *uiov;
4439 msg_compat = (struct compat_msghdr __user *) sr->umsg;
4440 ret = __get_compat_msghdr(&iomsg->msg, msg_compat, &iomsg->uaddr,
4445 uiov = compat_ptr(ptr);
4446 if (req->flags & REQ_F_BUFFER_SELECT) {
4447 compat_ssize_t clen;
4451 if (!access_ok(uiov, sizeof(*uiov)))
4453 if (__get_user(clen, &uiov->iov_len))
4458 iomsg->free_iov = NULL;
4460 iomsg->free_iov = iomsg->fast_iov;
4461 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4462 UIO_FASTIOV, &iomsg->free_iov,
4463 &iomsg->msg.msg_iter, true);
4472 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4473 struct io_async_msghdr *iomsg)
4475 iomsg->msg.msg_name = &iomsg->addr;
4477 #ifdef CONFIG_COMPAT
4478 if (req->ctx->compat)
4479 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4482 return __io_recvmsg_copy_hdr(req, iomsg);
4485 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4488 struct io_sr_msg *sr = &req->sr_msg;
4489 struct io_buffer *kbuf;
4491 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4496 req->flags |= REQ_F_BUFFER_SELECTED;
4500 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4502 return io_put_kbuf(req, req->sr_msg.kbuf);
4505 static int io_recvmsg_prep_async(struct io_kiocb *req)
4509 if (!io_op_defs[req->opcode].needs_async_data)
4511 ret = io_recvmsg_copy_hdr(req, req->async_data);
4513 req->flags |= REQ_F_NEED_CLEANUP;
4517 static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4519 struct io_sr_msg *sr = &req->sr_msg;
4521 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4524 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4525 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4526 sr->len = READ_ONCE(sqe->len);
4527 sr->bgid = READ_ONCE(sqe->buf_group);
4529 #ifdef CONFIG_COMPAT
4530 if (req->ctx->compat)
4531 sr->msg_flags |= MSG_CMSG_COMPAT;
4536 static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
4538 struct io_async_msghdr iomsg, *kmsg;
4539 struct socket *sock;
4540 struct io_buffer *kbuf;
4542 int ret, cflags = 0;
4543 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4545 sock = sock_from_file(req->file);
4546 if (unlikely(!sock))
4549 kmsg = req->async_data;
4551 ret = io_recvmsg_copy_hdr(req, &iomsg);
4557 if (req->flags & REQ_F_BUFFER_SELECT) {
4558 kbuf = io_recv_buffer_select(req, !force_nonblock);
4560 return PTR_ERR(kbuf);
4561 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4562 kmsg->fast_iov[0].iov_len = req->sr_msg.len;
4563 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
4564 1, req->sr_msg.len);
4567 flags = req->sr_msg.msg_flags;
4568 if (flags & MSG_DONTWAIT)
4569 req->flags |= REQ_F_NOWAIT;
4570 else if (force_nonblock)
4571 flags |= MSG_DONTWAIT;
4573 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4574 kmsg->uaddr, flags);
4575 if (force_nonblock && ret == -EAGAIN)
4576 return io_setup_async_msg(req, kmsg);
4577 if (ret == -ERESTARTSYS)
4580 if (req->flags & REQ_F_BUFFER_SELECTED)
4581 cflags = io_put_recv_kbuf(req);
4582 /* fast path, check for non-NULL to avoid function call */
4584 kfree(kmsg->free_iov);
4585 req->flags &= ~REQ_F_NEED_CLEANUP;
4587 req_set_fail_links(req);
4588 __io_req_complete(req, issue_flags, ret, cflags);
4592 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
4594 struct io_buffer *kbuf;
4595 struct io_sr_msg *sr = &req->sr_msg;
4597 void __user *buf = sr->buf;
4598 struct socket *sock;
4601 int ret, cflags = 0;
4602 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4604 sock = sock_from_file(req->file);
4605 if (unlikely(!sock))
4608 if (req->flags & REQ_F_BUFFER_SELECT) {
4609 kbuf = io_recv_buffer_select(req, !force_nonblock);
4611 return PTR_ERR(kbuf);
4612 buf = u64_to_user_ptr(kbuf->addr);
4615 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4619 msg.msg_name = NULL;
4620 msg.msg_control = NULL;
4621 msg.msg_controllen = 0;
4622 msg.msg_namelen = 0;
4623 msg.msg_iocb = NULL;
4626 flags = req->sr_msg.msg_flags;
4627 if (flags & MSG_DONTWAIT)
4628 req->flags |= REQ_F_NOWAIT;
4629 else if (force_nonblock)
4630 flags |= MSG_DONTWAIT;
4632 ret = sock_recvmsg(sock, &msg, flags);
4633 if (force_nonblock && ret == -EAGAIN)
4635 if (ret == -ERESTARTSYS)
4638 if (req->flags & REQ_F_BUFFER_SELECTED)
4639 cflags = io_put_recv_kbuf(req);
4641 req_set_fail_links(req);
4642 __io_req_complete(req, issue_flags, ret, cflags);
4646 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4648 struct io_accept *accept = &req->accept;
4650 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4652 if (sqe->ioprio || sqe->len || sqe->buf_index)
4655 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4656 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4657 accept->flags = READ_ONCE(sqe->accept_flags);
4658 accept->nofile = rlimit(RLIMIT_NOFILE);
4662 static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
4664 struct io_accept *accept = &req->accept;
4665 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4666 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4669 if (req->file->f_flags & O_NONBLOCK)
4670 req->flags |= REQ_F_NOWAIT;
4672 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4673 accept->addr_len, accept->flags,
4675 if (ret == -EAGAIN && force_nonblock)
4678 if (ret == -ERESTARTSYS)
4680 req_set_fail_links(req);
4682 __io_req_complete(req, issue_flags, ret, 0);
4686 static int io_connect_prep_async(struct io_kiocb *req)
4688 struct io_async_connect *io = req->async_data;
4689 struct io_connect *conn = &req->connect;
4691 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
4694 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4696 struct io_connect *conn = &req->connect;
4698 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4700 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4703 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4704 conn->addr_len = READ_ONCE(sqe->addr2);
4708 static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
4710 struct io_async_connect __io, *io;
4711 unsigned file_flags;
4713 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4715 if (req->async_data) {
4716 io = req->async_data;
4718 ret = move_addr_to_kernel(req->connect.addr,
4719 req->connect.addr_len,
4726 file_flags = force_nonblock ? O_NONBLOCK : 0;
4728 ret = __sys_connect_file(req->file, &io->address,
4729 req->connect.addr_len, file_flags);
4730 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4731 if (req->async_data)
4733 if (io_alloc_async_data(req)) {
4737 io = req->async_data;
4738 memcpy(req->async_data, &__io, sizeof(__io));
4741 if (ret == -ERESTARTSYS)
4745 req_set_fail_links(req);
4746 __io_req_complete(req, issue_flags, ret, 0);
4749 #else /* !CONFIG_NET */
4750 #define IO_NETOP_FN(op) \
4751 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
4753 return -EOPNOTSUPP; \
4756 #define IO_NETOP_PREP(op) \
4758 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
4760 return -EOPNOTSUPP; \
4763 #define IO_NETOP_PREP_ASYNC(op) \
4765 static int io_##op##_prep_async(struct io_kiocb *req) \
4767 return -EOPNOTSUPP; \
4770 IO_NETOP_PREP_ASYNC(sendmsg);
4771 IO_NETOP_PREP_ASYNC(recvmsg);
4772 IO_NETOP_PREP_ASYNC(connect);
4773 IO_NETOP_PREP(accept);
4776 #endif /* CONFIG_NET */
4778 struct io_poll_table {
4779 struct poll_table_struct pt;
4780 struct io_kiocb *req;
4784 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4785 __poll_t mask, task_work_func_t func)
4789 /* for instances that support it check for an event match first: */
4790 if (mask && !(mask & poll->events))
4793 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4795 list_del_init(&poll->wait.entry);
4798 req->task_work.func = func;
4799 percpu_ref_get(&req->ctx->refs);
4802 * If this fails, then the task is exiting. When a task exits, the
4803 * work gets canceled, so just cancel this request as well instead
4804 * of executing it. We can't safely execute it anyway, as we may not
4805 * have the needed state needed for it anyway.
4807 ret = io_req_task_work_add(req);
4808 if (unlikely(ret)) {
4809 WRITE_ONCE(poll->canceled, true);
4810 io_req_task_work_add_fallback(req, func);
4815 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4816 __acquires(&req->ctx->completion_lock)
4818 struct io_ring_ctx *ctx = req->ctx;
4820 if (!req->result && !READ_ONCE(poll->canceled)) {
4821 struct poll_table_struct pt = { ._key = poll->events };
4823 req->result = vfs_poll(req->file, &pt) & poll->events;
4826 spin_lock_irq(&ctx->completion_lock);
4827 if (!req->result && !READ_ONCE(poll->canceled)) {
4828 add_wait_queue(poll->head, &poll->wait);
4835 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4837 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4838 if (req->opcode == IORING_OP_POLL_ADD)
4839 return req->async_data;
4840 return req->apoll->double_poll;
4843 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4845 if (req->opcode == IORING_OP_POLL_ADD)
4847 return &req->apoll->poll;
4850 static void io_poll_remove_double(struct io_kiocb *req)
4852 struct io_poll_iocb *poll = io_poll_get_double(req);
4854 lockdep_assert_held(&req->ctx->completion_lock);
4856 if (poll && poll->head) {
4857 struct wait_queue_head *head = poll->head;
4859 spin_lock(&head->lock);
4860 list_del_init(&poll->wait.entry);
4861 if (poll->wait.private)
4862 refcount_dec(&req->refs);
4864 spin_unlock(&head->lock);
4868 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
4870 struct io_ring_ctx *ctx = req->ctx;
4872 io_poll_remove_double(req);
4873 req->poll.done = true;
4874 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
4875 io_commit_cqring(ctx);
4878 static void io_poll_task_func(struct callback_head *cb)
4880 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4881 struct io_ring_ctx *ctx = req->ctx;
4882 struct io_kiocb *nxt;
4884 if (io_poll_rewait(req, &req->poll)) {
4885 spin_unlock_irq(&ctx->completion_lock);
4887 hash_del(&req->hash_node);
4888 io_poll_complete(req, req->result, 0);
4889 spin_unlock_irq(&ctx->completion_lock);
4891 nxt = io_put_req_find_next(req);
4892 io_cqring_ev_posted(ctx);
4894 __io_req_task_submit(nxt);
4897 percpu_ref_put(&ctx->refs);
4900 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4901 int sync, void *key)
4903 struct io_kiocb *req = wait->private;
4904 struct io_poll_iocb *poll = io_poll_get_single(req);
4905 __poll_t mask = key_to_poll(key);
4907 /* for instances that support it check for an event match first: */
4908 if (mask && !(mask & poll->events))
4911 list_del_init(&wait->entry);
4913 if (poll && poll->head) {
4916 spin_lock(&poll->head->lock);
4917 done = list_empty(&poll->wait.entry);
4919 list_del_init(&poll->wait.entry);
4920 /* make sure double remove sees this as being gone */
4921 wait->private = NULL;
4922 spin_unlock(&poll->head->lock);
4924 /* use wait func handler, so it matches the rq type */
4925 poll->wait.func(&poll->wait, mode, sync, key);
4928 refcount_dec(&req->refs);
4932 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
4933 wait_queue_func_t wake_func)
4937 poll->canceled = false;
4938 poll->events = events;
4939 INIT_LIST_HEAD(&poll->wait.entry);
4940 init_waitqueue_func_entry(&poll->wait, wake_func);
4943 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
4944 struct wait_queue_head *head,
4945 struct io_poll_iocb **poll_ptr)
4947 struct io_kiocb *req = pt->req;
4950 * If poll->head is already set, it's because the file being polled
4951 * uses multiple waitqueues for poll handling (eg one for read, one
4952 * for write). Setup a separate io_poll_iocb if this happens.
4954 if (unlikely(poll->head)) {
4955 struct io_poll_iocb *poll_one = poll;
4957 /* already have a 2nd entry, fail a third attempt */
4959 pt->error = -EINVAL;
4962 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
4964 pt->error = -ENOMEM;
4967 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
4968 refcount_inc(&req->refs);
4969 poll->wait.private = req;
4976 if (poll->events & EPOLLEXCLUSIVE)
4977 add_wait_queue_exclusive(head, &poll->wait);
4979 add_wait_queue(head, &poll->wait);
4982 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
4983 struct poll_table_struct *p)
4985 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
4986 struct async_poll *apoll = pt->req->apoll;
4988 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
4991 static void io_async_task_func(struct callback_head *cb)
4993 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4994 struct async_poll *apoll = req->apoll;
4995 struct io_ring_ctx *ctx = req->ctx;
4997 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
4999 if (io_poll_rewait(req, &apoll->poll)) {
5000 spin_unlock_irq(&ctx->completion_lock);
5001 percpu_ref_put(&ctx->refs);
5005 /* If req is still hashed, it cannot have been canceled. Don't check. */
5006 if (hash_hashed(&req->hash_node))
5007 hash_del(&req->hash_node);
5009 io_poll_remove_double(req);
5010 spin_unlock_irq(&ctx->completion_lock);
5012 if (!READ_ONCE(apoll->poll.canceled))
5013 __io_req_task_submit(req);
5015 __io_req_task_cancel(req, -ECANCELED);
5017 percpu_ref_put(&ctx->refs);
5018 kfree(apoll->double_poll);
5022 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5025 struct io_kiocb *req = wait->private;
5026 struct io_poll_iocb *poll = &req->apoll->poll;
5028 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5031 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5034 static void io_poll_req_insert(struct io_kiocb *req)
5036 struct io_ring_ctx *ctx = req->ctx;
5037 struct hlist_head *list;
5039 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5040 hlist_add_head(&req->hash_node, list);
5043 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5044 struct io_poll_iocb *poll,
5045 struct io_poll_table *ipt, __poll_t mask,
5046 wait_queue_func_t wake_func)
5047 __acquires(&ctx->completion_lock)
5049 struct io_ring_ctx *ctx = req->ctx;
5050 bool cancel = false;
5052 INIT_HLIST_NODE(&req->hash_node);
5053 io_init_poll_iocb(poll, mask, wake_func);
5054 poll->file = req->file;
5055 poll->wait.private = req;
5057 ipt->pt._key = mask;
5059 ipt->error = -EINVAL;
5061 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5063 spin_lock_irq(&ctx->completion_lock);
5064 if (likely(poll->head)) {
5065 spin_lock(&poll->head->lock);
5066 if (unlikely(list_empty(&poll->wait.entry))) {
5072 if (mask || ipt->error)
5073 list_del_init(&poll->wait.entry);
5075 WRITE_ONCE(poll->canceled, true);
5076 else if (!poll->done) /* actually waiting for an event */
5077 io_poll_req_insert(req);
5078 spin_unlock(&poll->head->lock);
5084 static bool io_arm_poll_handler(struct io_kiocb *req)
5086 const struct io_op_def *def = &io_op_defs[req->opcode];
5087 struct io_ring_ctx *ctx = req->ctx;
5088 struct async_poll *apoll;
5089 struct io_poll_table ipt;
5093 if (!req->file || !file_can_poll(req->file))
5095 if (req->flags & REQ_F_POLLED)
5099 else if (def->pollout)
5103 /* if we can't nonblock try, then no point in arming a poll handler */
5104 if (!io_file_supports_async(req->file, rw))
5107 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5108 if (unlikely(!apoll))
5110 apoll->double_poll = NULL;
5112 req->flags |= REQ_F_POLLED;
5117 mask |= POLLIN | POLLRDNORM;
5119 mask |= POLLOUT | POLLWRNORM;
5121 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5122 if ((req->opcode == IORING_OP_RECVMSG) &&
5123 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5126 mask |= POLLERR | POLLPRI;
5128 ipt.pt._qproc = io_async_queue_proc;
5130 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5132 if (ret || ipt.error) {
5133 io_poll_remove_double(req);
5134 spin_unlock_irq(&ctx->completion_lock);
5135 kfree(apoll->double_poll);
5139 spin_unlock_irq(&ctx->completion_lock);
5140 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
5141 apoll->poll.events);
5145 static bool __io_poll_remove_one(struct io_kiocb *req,
5146 struct io_poll_iocb *poll)
5148 bool do_complete = false;
5150 spin_lock(&poll->head->lock);
5151 WRITE_ONCE(poll->canceled, true);
5152 if (!list_empty(&poll->wait.entry)) {
5153 list_del_init(&poll->wait.entry);
5156 spin_unlock(&poll->head->lock);
5157 hash_del(&req->hash_node);
5161 static bool io_poll_remove_one(struct io_kiocb *req)
5165 io_poll_remove_double(req);
5167 if (req->opcode == IORING_OP_POLL_ADD) {
5168 do_complete = __io_poll_remove_one(req, &req->poll);
5170 struct async_poll *apoll = req->apoll;
5172 /* non-poll requests have submit ref still */
5173 do_complete = __io_poll_remove_one(req, &apoll->poll);
5176 kfree(apoll->double_poll);
5182 io_cqring_fill_event(req, -ECANCELED);
5183 io_commit_cqring(req->ctx);
5184 req_set_fail_links(req);
5185 io_put_req_deferred(req, 1);
5192 * Returns true if we found and killed one or more poll requests
5194 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5195 struct files_struct *files)
5197 struct hlist_node *tmp;
5198 struct io_kiocb *req;
5201 spin_lock_irq(&ctx->completion_lock);
5202 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5203 struct hlist_head *list;
5205 list = &ctx->cancel_hash[i];
5206 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5207 if (io_match_task(req, tsk, files))
5208 posted += io_poll_remove_one(req);
5211 spin_unlock_irq(&ctx->completion_lock);
5214 io_cqring_ev_posted(ctx);
5219 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
5221 struct hlist_head *list;
5222 struct io_kiocb *req;
5224 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5225 hlist_for_each_entry(req, list, hash_node) {
5226 if (sqe_addr != req->user_data)
5228 if (io_poll_remove_one(req))
5236 static int io_poll_remove_prep(struct io_kiocb *req,
5237 const struct io_uring_sqe *sqe)
5239 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5241 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
5245 req->poll_remove.addr = READ_ONCE(sqe->addr);
5250 * Find a running poll command that matches one specified in sqe->addr,
5251 * and remove it if found.
5253 static int io_poll_remove(struct io_kiocb *req, unsigned int issue_flags)
5255 struct io_ring_ctx *ctx = req->ctx;
5258 spin_lock_irq(&ctx->completion_lock);
5259 ret = io_poll_cancel(ctx, req->poll_remove.addr);
5260 spin_unlock_irq(&ctx->completion_lock);
5263 req_set_fail_links(req);
5264 io_req_complete(req, ret);
5268 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5271 struct io_kiocb *req = wait->private;
5272 struct io_poll_iocb *poll = &req->poll;
5274 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5277 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5278 struct poll_table_struct *p)
5280 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5282 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5285 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5287 struct io_poll_iocb *poll = &req->poll;
5290 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5292 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
5295 events = READ_ONCE(sqe->poll32_events);
5297 events = swahw32(events);
5299 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP |
5300 (events & EPOLLEXCLUSIVE);
5304 static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5306 struct io_poll_iocb *poll = &req->poll;
5307 struct io_ring_ctx *ctx = req->ctx;
5308 struct io_poll_table ipt;
5311 ipt.pt._qproc = io_poll_queue_proc;
5313 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5316 if (mask) { /* no async, we'd stolen it */
5318 io_poll_complete(req, mask, 0);
5320 spin_unlock_irq(&ctx->completion_lock);
5323 io_cqring_ev_posted(ctx);
5329 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5331 struct io_timeout_data *data = container_of(timer,
5332 struct io_timeout_data, timer);
5333 struct io_kiocb *req = data->req;
5334 struct io_ring_ctx *ctx = req->ctx;
5335 unsigned long flags;
5337 spin_lock_irqsave(&ctx->completion_lock, flags);
5338 list_del_init(&req->timeout.list);
5339 atomic_set(&req->ctx->cq_timeouts,
5340 atomic_read(&req->ctx->cq_timeouts) + 1);
5342 io_cqring_fill_event(req, -ETIME);
5343 io_commit_cqring(ctx);
5344 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5346 io_cqring_ev_posted(ctx);
5347 req_set_fail_links(req);
5349 return HRTIMER_NORESTART;
5352 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5355 struct io_timeout_data *io;
5356 struct io_kiocb *req;
5359 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5360 if (user_data == req->user_data) {
5367 return ERR_PTR(ret);
5369 io = req->async_data;
5370 ret = hrtimer_try_to_cancel(&io->timer);
5372 return ERR_PTR(-EALREADY);
5373 list_del_init(&req->timeout.list);
5377 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5379 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5382 return PTR_ERR(req);
5384 req_set_fail_links(req);
5385 io_cqring_fill_event(req, -ECANCELED);
5386 io_put_req_deferred(req, 1);
5390 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5391 struct timespec64 *ts, enum hrtimer_mode mode)
5393 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5394 struct io_timeout_data *data;
5397 return PTR_ERR(req);
5399 req->timeout.off = 0; /* noseq */
5400 data = req->async_data;
5401 list_add_tail(&req->timeout.list, &ctx->timeout_list);
5402 hrtimer_init(&data->timer, CLOCK_MONOTONIC, mode);
5403 data->timer.function = io_timeout_fn;
5404 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
5408 static int io_timeout_remove_prep(struct io_kiocb *req,
5409 const struct io_uring_sqe *sqe)
5411 struct io_timeout_rem *tr = &req->timeout_rem;
5413 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5415 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5417 if (sqe->ioprio || sqe->buf_index || sqe->len)
5420 tr->addr = READ_ONCE(sqe->addr);
5421 tr->flags = READ_ONCE(sqe->timeout_flags);
5422 if (tr->flags & IORING_TIMEOUT_UPDATE) {
5423 if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS))
5425 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
5427 } else if (tr->flags) {
5428 /* timeout removal doesn't support flags */
5435 static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
5437 return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
5442 * Remove or update an existing timeout command
5444 static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
5446 struct io_timeout_rem *tr = &req->timeout_rem;
5447 struct io_ring_ctx *ctx = req->ctx;
5450 spin_lock_irq(&ctx->completion_lock);
5451 if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE))
5452 ret = io_timeout_cancel(ctx, tr->addr);
5454 ret = io_timeout_update(ctx, tr->addr, &tr->ts,
5455 io_translate_timeout_mode(tr->flags));
5457 io_cqring_fill_event(req, ret);
5458 io_commit_cqring(ctx);
5459 spin_unlock_irq(&ctx->completion_lock);
5460 io_cqring_ev_posted(ctx);
5462 req_set_fail_links(req);
5467 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5468 bool is_timeout_link)
5470 struct io_timeout_data *data;
5472 u32 off = READ_ONCE(sqe->off);
5474 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5476 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5478 if (off && is_timeout_link)
5480 flags = READ_ONCE(sqe->timeout_flags);
5481 if (flags & ~IORING_TIMEOUT_ABS)
5484 req->timeout.off = off;
5486 if (!req->async_data && io_alloc_async_data(req))
5489 data = req->async_data;
5492 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5495 data->mode = io_translate_timeout_mode(flags);
5496 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5500 static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
5502 struct io_ring_ctx *ctx = req->ctx;
5503 struct io_timeout_data *data = req->async_data;
5504 struct list_head *entry;
5505 u32 tail, off = req->timeout.off;
5507 spin_lock_irq(&ctx->completion_lock);
5510 * sqe->off holds how many events that need to occur for this
5511 * timeout event to be satisfied. If it isn't set, then this is
5512 * a pure timeout request, sequence isn't used.
5514 if (io_is_timeout_noseq(req)) {
5515 entry = ctx->timeout_list.prev;
5519 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5520 req->timeout.target_seq = tail + off;
5522 /* Update the last seq here in case io_flush_timeouts() hasn't.
5523 * This is safe because ->completion_lock is held, and submissions
5524 * and completions are never mixed in the same ->completion_lock section.
5526 ctx->cq_last_tm_flush = tail;
5529 * Insertion sort, ensuring the first entry in the list is always
5530 * the one we need first.
5532 list_for_each_prev(entry, &ctx->timeout_list) {
5533 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5536 if (io_is_timeout_noseq(nxt))
5538 /* nxt.seq is behind @tail, otherwise would've been completed */
5539 if (off >= nxt->timeout.target_seq - tail)
5543 list_add(&req->timeout.list, entry);
5544 data->timer.function = io_timeout_fn;
5545 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5546 spin_unlock_irq(&ctx->completion_lock);
5550 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5552 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5554 return req->user_data == (unsigned long) data;
5557 static int io_async_cancel_one(struct io_uring_task *tctx, void *sqe_addr)
5559 enum io_wq_cancel cancel_ret;
5565 cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, sqe_addr, false);
5566 switch (cancel_ret) {
5567 case IO_WQ_CANCEL_OK:
5570 case IO_WQ_CANCEL_RUNNING:
5573 case IO_WQ_CANCEL_NOTFOUND:
5581 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5582 struct io_kiocb *req, __u64 sqe_addr,
5585 unsigned long flags;
5588 ret = io_async_cancel_one(req->task->io_uring,
5589 (void *) (unsigned long) sqe_addr);
5590 if (ret != -ENOENT) {
5591 spin_lock_irqsave(&ctx->completion_lock, flags);
5595 spin_lock_irqsave(&ctx->completion_lock, flags);
5596 ret = io_timeout_cancel(ctx, sqe_addr);
5599 ret = io_poll_cancel(ctx, sqe_addr);
5603 io_cqring_fill_event(req, ret);
5604 io_commit_cqring(ctx);
5605 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5606 io_cqring_ev_posted(ctx);
5609 req_set_fail_links(req);
5613 static int io_async_cancel_prep(struct io_kiocb *req,
5614 const struct io_uring_sqe *sqe)
5616 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5618 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5620 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5623 req->cancel.addr = READ_ONCE(sqe->addr);
5627 static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
5629 struct io_ring_ctx *ctx = req->ctx;
5631 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
5635 static int io_rsrc_update_prep(struct io_kiocb *req,
5636 const struct io_uring_sqe *sqe)
5638 if (unlikely(req->ctx->flags & IORING_SETUP_SQPOLL))
5640 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5642 if (sqe->ioprio || sqe->rw_flags)
5645 req->rsrc_update.offset = READ_ONCE(sqe->off);
5646 req->rsrc_update.nr_args = READ_ONCE(sqe->len);
5647 if (!req->rsrc_update.nr_args)
5649 req->rsrc_update.arg = READ_ONCE(sqe->addr);
5653 static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
5655 struct io_ring_ctx *ctx = req->ctx;
5656 struct io_uring_rsrc_update up;
5659 if (issue_flags & IO_URING_F_NONBLOCK)
5662 up.offset = req->rsrc_update.offset;
5663 up.data = req->rsrc_update.arg;
5665 mutex_lock(&ctx->uring_lock);
5666 ret = __io_sqe_files_update(ctx, &up, req->rsrc_update.nr_args);
5667 mutex_unlock(&ctx->uring_lock);
5670 req_set_fail_links(req);
5671 __io_req_complete(req, issue_flags, ret, 0);
5675 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5677 switch (req->opcode) {
5680 case IORING_OP_READV:
5681 case IORING_OP_READ_FIXED:
5682 case IORING_OP_READ:
5683 return io_read_prep(req, sqe);
5684 case IORING_OP_WRITEV:
5685 case IORING_OP_WRITE_FIXED:
5686 case IORING_OP_WRITE:
5687 return io_write_prep(req, sqe);
5688 case IORING_OP_POLL_ADD:
5689 return io_poll_add_prep(req, sqe);
5690 case IORING_OP_POLL_REMOVE:
5691 return io_poll_remove_prep(req, sqe);
5692 case IORING_OP_FSYNC:
5693 return io_fsync_prep(req, sqe);
5694 case IORING_OP_SYNC_FILE_RANGE:
5695 return io_sfr_prep(req, sqe);
5696 case IORING_OP_SENDMSG:
5697 case IORING_OP_SEND:
5698 return io_sendmsg_prep(req, sqe);
5699 case IORING_OP_RECVMSG:
5700 case IORING_OP_RECV:
5701 return io_recvmsg_prep(req, sqe);
5702 case IORING_OP_CONNECT:
5703 return io_connect_prep(req, sqe);
5704 case IORING_OP_TIMEOUT:
5705 return io_timeout_prep(req, sqe, false);
5706 case IORING_OP_TIMEOUT_REMOVE:
5707 return io_timeout_remove_prep(req, sqe);
5708 case IORING_OP_ASYNC_CANCEL:
5709 return io_async_cancel_prep(req, sqe);
5710 case IORING_OP_LINK_TIMEOUT:
5711 return io_timeout_prep(req, sqe, true);
5712 case IORING_OP_ACCEPT:
5713 return io_accept_prep(req, sqe);
5714 case IORING_OP_FALLOCATE:
5715 return io_fallocate_prep(req, sqe);
5716 case IORING_OP_OPENAT:
5717 return io_openat_prep(req, sqe);
5718 case IORING_OP_CLOSE:
5719 return io_close_prep(req, sqe);
5720 case IORING_OP_FILES_UPDATE:
5721 return io_rsrc_update_prep(req, sqe);
5722 case IORING_OP_STATX:
5723 return io_statx_prep(req, sqe);
5724 case IORING_OP_FADVISE:
5725 return io_fadvise_prep(req, sqe);
5726 case IORING_OP_MADVISE:
5727 return io_madvise_prep(req, sqe);
5728 case IORING_OP_OPENAT2:
5729 return io_openat2_prep(req, sqe);
5730 case IORING_OP_EPOLL_CTL:
5731 return io_epoll_ctl_prep(req, sqe);
5732 case IORING_OP_SPLICE:
5733 return io_splice_prep(req, sqe);
5734 case IORING_OP_PROVIDE_BUFFERS:
5735 return io_provide_buffers_prep(req, sqe);
5736 case IORING_OP_REMOVE_BUFFERS:
5737 return io_remove_buffers_prep(req, sqe);
5739 return io_tee_prep(req, sqe);
5740 case IORING_OP_SHUTDOWN:
5741 return io_shutdown_prep(req, sqe);
5742 case IORING_OP_RENAMEAT:
5743 return io_renameat_prep(req, sqe);
5744 case IORING_OP_UNLINKAT:
5745 return io_unlinkat_prep(req, sqe);
5748 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5753 static int io_req_prep_async(struct io_kiocb *req)
5755 switch (req->opcode) {
5756 case IORING_OP_READV:
5757 case IORING_OP_READ_FIXED:
5758 case IORING_OP_READ:
5759 return io_rw_prep_async(req, READ);
5760 case IORING_OP_WRITEV:
5761 case IORING_OP_WRITE_FIXED:
5762 case IORING_OP_WRITE:
5763 return io_rw_prep_async(req, WRITE);
5764 case IORING_OP_SENDMSG:
5765 case IORING_OP_SEND:
5766 return io_sendmsg_prep_async(req);
5767 case IORING_OP_RECVMSG:
5768 case IORING_OP_RECV:
5769 return io_recvmsg_prep_async(req);
5770 case IORING_OP_CONNECT:
5771 return io_connect_prep_async(req);
5776 static int io_req_defer_prep(struct io_kiocb *req)
5778 if (!io_op_defs[req->opcode].needs_async_data)
5780 /* some opcodes init it during the inital prep */
5781 if (req->async_data)
5783 if (__io_alloc_async_data(req))
5785 return io_req_prep_async(req);
5788 static u32 io_get_sequence(struct io_kiocb *req)
5790 struct io_kiocb *pos;
5791 struct io_ring_ctx *ctx = req->ctx;
5792 u32 total_submitted, nr_reqs = 0;
5794 io_for_each_link(pos, req)
5797 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
5798 return total_submitted - nr_reqs;
5801 static int io_req_defer(struct io_kiocb *req)
5803 struct io_ring_ctx *ctx = req->ctx;
5804 struct io_defer_entry *de;
5808 /* Still need defer if there is pending req in defer list. */
5809 if (likely(list_empty_careful(&ctx->defer_list) &&
5810 !(req->flags & REQ_F_IO_DRAIN)))
5813 seq = io_get_sequence(req);
5814 /* Still a chance to pass the sequence check */
5815 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
5818 ret = io_req_defer_prep(req);
5821 io_prep_async_link(req);
5822 de = kmalloc(sizeof(*de), GFP_KERNEL);
5826 spin_lock_irq(&ctx->completion_lock);
5827 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
5828 spin_unlock_irq(&ctx->completion_lock);
5830 io_queue_async_work(req);
5831 return -EIOCBQUEUED;
5834 trace_io_uring_defer(ctx, req, req->user_data);
5837 list_add_tail(&de->list, &ctx->defer_list);
5838 spin_unlock_irq(&ctx->completion_lock);
5839 return -EIOCBQUEUED;
5842 static void __io_clean_op(struct io_kiocb *req)
5844 if (req->flags & REQ_F_BUFFER_SELECTED) {
5845 switch (req->opcode) {
5846 case IORING_OP_READV:
5847 case IORING_OP_READ_FIXED:
5848 case IORING_OP_READ:
5849 kfree((void *)(unsigned long)req->rw.addr);
5851 case IORING_OP_RECVMSG:
5852 case IORING_OP_RECV:
5853 kfree(req->sr_msg.kbuf);
5856 req->flags &= ~REQ_F_BUFFER_SELECTED;
5859 if (req->flags & REQ_F_NEED_CLEANUP) {
5860 switch (req->opcode) {
5861 case IORING_OP_READV:
5862 case IORING_OP_READ_FIXED:
5863 case IORING_OP_READ:
5864 case IORING_OP_WRITEV:
5865 case IORING_OP_WRITE_FIXED:
5866 case IORING_OP_WRITE: {
5867 struct io_async_rw *io = req->async_data;
5869 kfree(io->free_iovec);
5872 case IORING_OP_RECVMSG:
5873 case IORING_OP_SENDMSG: {
5874 struct io_async_msghdr *io = req->async_data;
5876 kfree(io->free_iov);
5879 case IORING_OP_SPLICE:
5881 io_put_file(req, req->splice.file_in,
5882 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
5884 case IORING_OP_OPENAT:
5885 case IORING_OP_OPENAT2:
5886 if (req->open.filename)
5887 putname(req->open.filename);
5889 case IORING_OP_RENAMEAT:
5890 putname(req->rename.oldpath);
5891 putname(req->rename.newpath);
5893 case IORING_OP_UNLINKAT:
5894 putname(req->unlink.filename);
5897 req->flags &= ~REQ_F_NEED_CLEANUP;
5901 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
5903 struct io_ring_ctx *ctx = req->ctx;
5904 const struct cred *creds = NULL;
5907 if (req->work.personality) {
5908 const struct cred *new_creds;
5910 if (!(issue_flags & IO_URING_F_NONBLOCK))
5911 mutex_lock(&ctx->uring_lock);
5912 new_creds = idr_find(&ctx->personality_idr, req->work.personality);
5913 if (!(issue_flags & IO_URING_F_NONBLOCK))
5914 mutex_unlock(&ctx->uring_lock);
5917 creds = override_creds(new_creds);
5920 switch (req->opcode) {
5922 ret = io_nop(req, issue_flags);
5924 case IORING_OP_READV:
5925 case IORING_OP_READ_FIXED:
5926 case IORING_OP_READ:
5927 ret = io_read(req, issue_flags);
5929 case IORING_OP_WRITEV:
5930 case IORING_OP_WRITE_FIXED:
5931 case IORING_OP_WRITE:
5932 ret = io_write(req, issue_flags);
5934 case IORING_OP_FSYNC:
5935 ret = io_fsync(req, issue_flags);
5937 case IORING_OP_POLL_ADD:
5938 ret = io_poll_add(req, issue_flags);
5940 case IORING_OP_POLL_REMOVE:
5941 ret = io_poll_remove(req, issue_flags);
5943 case IORING_OP_SYNC_FILE_RANGE:
5944 ret = io_sync_file_range(req, issue_flags);
5946 case IORING_OP_SENDMSG:
5947 ret = io_sendmsg(req, issue_flags);
5949 case IORING_OP_SEND:
5950 ret = io_send(req, issue_flags);
5952 case IORING_OP_RECVMSG:
5953 ret = io_recvmsg(req, issue_flags);
5955 case IORING_OP_RECV:
5956 ret = io_recv(req, issue_flags);
5958 case IORING_OP_TIMEOUT:
5959 ret = io_timeout(req, issue_flags);
5961 case IORING_OP_TIMEOUT_REMOVE:
5962 ret = io_timeout_remove(req, issue_flags);
5964 case IORING_OP_ACCEPT:
5965 ret = io_accept(req, issue_flags);
5967 case IORING_OP_CONNECT:
5968 ret = io_connect(req, issue_flags);
5970 case IORING_OP_ASYNC_CANCEL:
5971 ret = io_async_cancel(req, issue_flags);
5973 case IORING_OP_FALLOCATE:
5974 ret = io_fallocate(req, issue_flags);
5976 case IORING_OP_OPENAT:
5977 ret = io_openat(req, issue_flags);
5979 case IORING_OP_CLOSE:
5980 ret = io_close(req, issue_flags);
5982 case IORING_OP_FILES_UPDATE:
5983 ret = io_files_update(req, issue_flags);
5985 case IORING_OP_STATX:
5986 ret = io_statx(req, issue_flags);
5988 case IORING_OP_FADVISE:
5989 ret = io_fadvise(req, issue_flags);
5991 case IORING_OP_MADVISE:
5992 ret = io_madvise(req, issue_flags);
5994 case IORING_OP_OPENAT2:
5995 ret = io_openat2(req, issue_flags);
5997 case IORING_OP_EPOLL_CTL:
5998 ret = io_epoll_ctl(req, issue_flags);
6000 case IORING_OP_SPLICE:
6001 ret = io_splice(req, issue_flags);
6003 case IORING_OP_PROVIDE_BUFFERS:
6004 ret = io_provide_buffers(req, issue_flags);
6006 case IORING_OP_REMOVE_BUFFERS:
6007 ret = io_remove_buffers(req, issue_flags);
6010 ret = io_tee(req, issue_flags);
6012 case IORING_OP_SHUTDOWN:
6013 ret = io_shutdown(req, issue_flags);
6015 case IORING_OP_RENAMEAT:
6016 ret = io_renameat(req, issue_flags);
6018 case IORING_OP_UNLINKAT:
6019 ret = io_unlinkat(req, issue_flags);
6027 revert_creds(creds);
6032 /* If the op doesn't have a file, we're not polling for it */
6033 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
6034 const bool in_async = io_wq_current_is_worker();
6036 /* workqueue context doesn't hold uring_lock, grab it now */
6038 mutex_lock(&ctx->uring_lock);
6040 io_iopoll_req_issued(req, in_async);
6043 mutex_unlock(&ctx->uring_lock);
6049 static void io_wq_submit_work(struct io_wq_work *work)
6051 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6052 struct io_kiocb *timeout;
6055 timeout = io_prep_linked_timeout(req);
6057 io_queue_linked_timeout(timeout);
6059 if (work->flags & IO_WQ_WORK_CANCEL)
6064 ret = io_issue_sqe(req, 0);
6066 * We can get EAGAIN for polled IO even though we're
6067 * forcing a sync submission from here, since we can't
6068 * wait for request slots on the block side.
6076 /* avoid locking problems by failing it from a clean context */
6078 /* io-wq is going to take one down */
6079 refcount_inc(&req->refs);
6080 io_req_task_queue_fail(req, ret);
6084 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6087 struct fixed_rsrc_table *table;
6089 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
6090 return table->files[index & IORING_FILE_TABLE_MASK];
6093 static struct file *io_file_get(struct io_submit_state *state,
6094 struct io_kiocb *req, int fd, bool fixed)
6096 struct io_ring_ctx *ctx = req->ctx;
6100 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6102 fd = array_index_nospec(fd, ctx->nr_user_files);
6103 file = io_file_from_index(ctx, fd);
6104 io_set_resource_node(req);
6106 trace_io_uring_file_get(ctx, fd);
6107 file = __io_file_get(state, fd);
6110 if (file && unlikely(file->f_op == &io_uring_fops))
6111 io_req_track_inflight(req);
6115 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6117 struct io_timeout_data *data = container_of(timer,
6118 struct io_timeout_data, timer);
6119 struct io_kiocb *prev, *req = data->req;
6120 struct io_ring_ctx *ctx = req->ctx;
6121 unsigned long flags;
6123 spin_lock_irqsave(&ctx->completion_lock, flags);
6124 prev = req->timeout.head;
6125 req->timeout.head = NULL;
6128 * We don't expect the list to be empty, that will only happen if we
6129 * race with the completion of the linked work.
6131 if (prev && refcount_inc_not_zero(&prev->refs))
6132 io_remove_next_linked(prev);
6135 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6138 req_set_fail_links(prev);
6139 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6140 io_put_req_deferred(prev, 1);
6142 io_req_complete_post(req, -ETIME, 0);
6143 io_put_req_deferred(req, 1);
6145 return HRTIMER_NORESTART;
6148 static void __io_queue_linked_timeout(struct io_kiocb *req)
6151 * If the back reference is NULL, then our linked request finished
6152 * before we got a chance to setup the timer
6154 if (req->timeout.head) {
6155 struct io_timeout_data *data = req->async_data;
6157 data->timer.function = io_link_timeout_fn;
6158 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6163 static void io_queue_linked_timeout(struct io_kiocb *req)
6165 struct io_ring_ctx *ctx = req->ctx;
6167 spin_lock_irq(&ctx->completion_lock);
6168 __io_queue_linked_timeout(req);
6169 spin_unlock_irq(&ctx->completion_lock);
6171 /* drop submission reference */
6175 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6177 struct io_kiocb *nxt = req->link;
6179 if (!nxt || (req->flags & REQ_F_LINK_TIMEOUT) ||
6180 nxt->opcode != IORING_OP_LINK_TIMEOUT)
6183 nxt->timeout.head = req;
6184 nxt->flags |= REQ_F_LTIMEOUT_ACTIVE;
6185 req->flags |= REQ_F_LINK_TIMEOUT;
6189 static void __io_queue_sqe(struct io_kiocb *req)
6191 struct io_kiocb *linked_timeout = io_prep_linked_timeout(req);
6194 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
6197 * We async punt it if the file wasn't marked NOWAIT, or if the file
6198 * doesn't support non-blocking read/write attempts
6200 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6201 if (!io_arm_poll_handler(req)) {
6203 * Queued up for async execution, worker will release
6204 * submit reference when the iocb is actually submitted.
6206 io_queue_async_work(req);
6208 } else if (likely(!ret)) {
6209 /* drop submission reference */
6210 if (req->flags & REQ_F_COMPLETE_INLINE) {
6211 struct io_ring_ctx *ctx = req->ctx;
6212 struct io_comp_state *cs = &ctx->submit_state.comp;
6214 cs->reqs[cs->nr++] = req;
6215 if (cs->nr == ARRAY_SIZE(cs->reqs))
6216 io_submit_flush_completions(cs, ctx);
6221 req_set_fail_links(req);
6223 io_req_complete(req, ret);
6226 io_queue_linked_timeout(linked_timeout);
6229 static void io_queue_sqe(struct io_kiocb *req)
6233 ret = io_req_defer(req);
6235 if (ret != -EIOCBQUEUED) {
6237 req_set_fail_links(req);
6239 io_req_complete(req, ret);
6241 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6242 ret = io_req_defer_prep(req);
6245 io_queue_async_work(req);
6247 __io_queue_sqe(req);
6252 * Check SQE restrictions (opcode and flags).
6254 * Returns 'true' if SQE is allowed, 'false' otherwise.
6256 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6257 struct io_kiocb *req,
6258 unsigned int sqe_flags)
6260 if (!ctx->restricted)
6263 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6266 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6267 ctx->restrictions.sqe_flags_required)
6270 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6271 ctx->restrictions.sqe_flags_required))
6277 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6278 const struct io_uring_sqe *sqe)
6280 struct io_submit_state *state;
6281 unsigned int sqe_flags;
6284 req->opcode = READ_ONCE(sqe->opcode);
6285 /* same numerical values with corresponding REQ_F_*, safe to copy */
6286 req->flags = sqe_flags = READ_ONCE(sqe->flags);
6287 req->user_data = READ_ONCE(sqe->user_data);
6288 req->async_data = NULL;
6292 req->fixed_rsrc_refs = NULL;
6293 /* one is dropped after submission, the other at completion */
6294 refcount_set(&req->refs, 2);
6295 req->task = current;
6298 /* enforce forwards compatibility on users */
6299 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS)) {
6304 if (unlikely(req->opcode >= IORING_OP_LAST))
6307 if (unlikely(!io_check_restriction(ctx, req, sqe_flags)))
6310 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6311 !io_op_defs[req->opcode].buffer_select)
6314 req->work.list.next = NULL;
6315 req->work.flags = 0;
6316 req->work.personality = READ_ONCE(sqe->personality);
6317 state = &ctx->submit_state;
6320 * Plug now if we have more than 1 IO left after this, and the target
6321 * is potentially a read/write to block based storage.
6323 if (!state->plug_started && state->ios_left > 1 &&
6324 io_op_defs[req->opcode].plug) {
6325 blk_start_plug(&state->plug);
6326 state->plug_started = true;
6329 if (io_op_defs[req->opcode].needs_file) {
6330 bool fixed = req->flags & REQ_F_FIXED_FILE;
6332 req->file = io_file_get(state, req, READ_ONCE(sqe->fd), fixed);
6333 if (unlikely(!req->file))
6341 static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
6342 const struct io_uring_sqe *sqe)
6344 struct io_submit_link *link = &ctx->submit_state.link;
6347 ret = io_init_req(ctx, req, sqe);
6348 if (unlikely(ret)) {
6351 io_req_complete(req, ret);
6353 /* fail even hard links since we don't submit */
6354 link->head->flags |= REQ_F_FAIL_LINK;
6355 io_put_req(link->head);
6356 io_req_complete(link->head, -ECANCELED);
6361 ret = io_req_prep(req, sqe);
6365 /* don't need @sqe from now on */
6366 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6367 true, ctx->flags & IORING_SETUP_SQPOLL);
6370 * If we already have a head request, queue this one for async
6371 * submittal once the head completes. If we don't have a head but
6372 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6373 * submitted sync once the chain is complete. If none of those
6374 * conditions are true (normal request), then just queue it.
6377 struct io_kiocb *head = link->head;
6380 * Taking sequential execution of a link, draining both sides
6381 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6382 * requests in the link. So, it drains the head and the
6383 * next after the link request. The last one is done via
6384 * drain_next flag to persist the effect across calls.
6386 if (req->flags & REQ_F_IO_DRAIN) {
6387 head->flags |= REQ_F_IO_DRAIN;
6388 ctx->drain_next = 1;
6390 ret = io_req_defer_prep(req);
6393 trace_io_uring_link(ctx, req, head);
6394 link->last->link = req;
6397 /* last request of a link, enqueue the link */
6398 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6403 if (unlikely(ctx->drain_next)) {
6404 req->flags |= REQ_F_IO_DRAIN;
6405 ctx->drain_next = 0;
6407 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6419 * Batched submission is done, ensure local IO is flushed out.
6421 static void io_submit_state_end(struct io_submit_state *state,
6422 struct io_ring_ctx *ctx)
6424 if (state->link.head)
6425 io_queue_sqe(state->link.head);
6427 io_submit_flush_completions(&state->comp, ctx);
6428 if (state->plug_started)
6429 blk_finish_plug(&state->plug);
6430 io_state_file_put(state);
6434 * Start submission side cache.
6436 static void io_submit_state_start(struct io_submit_state *state,
6437 unsigned int max_ios)
6439 state->plug_started = false;
6440 state->ios_left = max_ios;
6441 /* set only head, no need to init link_last in advance */
6442 state->link.head = NULL;
6445 static void io_commit_sqring(struct io_ring_ctx *ctx)
6447 struct io_rings *rings = ctx->rings;
6450 * Ensure any loads from the SQEs are done at this point,
6451 * since once we write the new head, the application could
6452 * write new data to them.
6454 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6458 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6459 * that is mapped by userspace. This means that care needs to be taken to
6460 * ensure that reads are stable, as we cannot rely on userspace always
6461 * being a good citizen. If members of the sqe are validated and then later
6462 * used, it's important that those reads are done through READ_ONCE() to
6463 * prevent a re-load down the line.
6465 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6467 u32 *sq_array = ctx->sq_array;
6471 * The cached sq head (or cq tail) serves two purposes:
6473 * 1) allows us to batch the cost of updating the user visible
6475 * 2) allows the kernel side to track the head on its own, even
6476 * though the application is the one updating it.
6478 head = READ_ONCE(sq_array[ctx->cached_sq_head++ & ctx->sq_mask]);
6479 if (likely(head < ctx->sq_entries))
6480 return &ctx->sq_sqes[head];
6482 /* drop invalid entries */
6483 ctx->cached_sq_dropped++;
6484 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6488 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6492 /* if we have a backlog and couldn't flush it all, return BUSY */
6493 if (test_bit(0, &ctx->sq_check_overflow)) {
6494 if (!__io_cqring_overflow_flush(ctx, false, NULL, NULL))
6498 /* make sure SQ entry isn't read before tail */
6499 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6501 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6504 percpu_counter_add(¤t->io_uring->inflight, nr);
6505 refcount_add(nr, ¤t->usage);
6506 io_submit_state_start(&ctx->submit_state, nr);
6508 while (submitted < nr) {
6509 const struct io_uring_sqe *sqe;
6510 struct io_kiocb *req;
6512 req = io_alloc_req(ctx);
6513 if (unlikely(!req)) {
6515 submitted = -EAGAIN;
6518 sqe = io_get_sqe(ctx);
6519 if (unlikely(!sqe)) {
6520 kmem_cache_free(req_cachep, req);
6523 /* will complete beyond this point, count as submitted */
6525 if (io_submit_sqe(ctx, req, sqe))
6529 if (unlikely(submitted != nr)) {
6530 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6531 struct io_uring_task *tctx = current->io_uring;
6532 int unused = nr - ref_used;
6534 percpu_ref_put_many(&ctx->refs, unused);
6535 percpu_counter_sub(&tctx->inflight, unused);
6536 put_task_struct_many(current, unused);
6539 io_submit_state_end(&ctx->submit_state, ctx);
6540 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6541 io_commit_sqring(ctx);
6546 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6548 /* Tell userspace we may need a wakeup call */
6549 spin_lock_irq(&ctx->completion_lock);
6550 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6551 spin_unlock_irq(&ctx->completion_lock);
6554 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6556 spin_lock_irq(&ctx->completion_lock);
6557 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6558 spin_unlock_irq(&ctx->completion_lock);
6561 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
6563 unsigned int to_submit;
6566 to_submit = io_sqring_entries(ctx);
6567 /* if we're handling multiple rings, cap submit size for fairness */
6568 if (cap_entries && to_submit > 8)
6571 if (!list_empty(&ctx->iopoll_list) || to_submit) {
6572 unsigned nr_events = 0;
6574 mutex_lock(&ctx->uring_lock);
6575 if (!list_empty(&ctx->iopoll_list))
6576 io_do_iopoll(ctx, &nr_events, 0);
6578 if (to_submit && !ctx->sqo_dead &&
6579 likely(!percpu_ref_is_dying(&ctx->refs)))
6580 ret = io_submit_sqes(ctx, to_submit);
6581 mutex_unlock(&ctx->uring_lock);
6584 if (!io_sqring_full(ctx) && wq_has_sleeper(&ctx->sqo_sq_wait))
6585 wake_up(&ctx->sqo_sq_wait);
6590 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
6592 struct io_ring_ctx *ctx;
6593 unsigned sq_thread_idle = 0;
6595 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6596 if (sq_thread_idle < ctx->sq_thread_idle)
6597 sq_thread_idle = ctx->sq_thread_idle;
6600 sqd->sq_thread_idle = sq_thread_idle;
6603 static void io_sqd_init_new(struct io_sq_data *sqd)
6605 struct io_ring_ctx *ctx;
6607 while (!list_empty(&sqd->ctx_new_list)) {
6608 ctx = list_first_entry(&sqd->ctx_new_list, struct io_ring_ctx, sqd_list);
6609 list_move_tail(&ctx->sqd_list, &sqd->ctx_list);
6610 complete(&ctx->sq_thread_comp);
6613 io_sqd_update_thread_idle(sqd);
6616 static bool io_sq_thread_should_stop(struct io_sq_data *sqd)
6618 return test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
6621 static bool io_sq_thread_should_park(struct io_sq_data *sqd)
6623 return test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
6626 static void io_sq_thread_parkme(struct io_sq_data *sqd)
6630 * TASK_PARKED is a special state; we must serialize against
6631 * possible pending wakeups to avoid store-store collisions on
6634 * Such a collision might possibly result in the task state
6635 * changin from TASK_PARKED and us failing the
6636 * wait_task_inactive() in kthread_park().
6638 set_special_state(TASK_PARKED);
6639 if (!test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state))
6643 * Thread is going to call schedule(), do not preempt it,
6644 * or the caller of kthread_park() may spend more time in
6645 * wait_task_inactive().
6648 complete(&sqd->completion);
6649 schedule_preempt_disabled();
6652 __set_current_state(TASK_RUNNING);
6655 static int io_sq_thread(void *data)
6657 struct io_sq_data *sqd = data;
6658 struct io_ring_ctx *ctx;
6659 unsigned long timeout = 0;
6660 char buf[TASK_COMM_LEN];
6663 sprintf(buf, "iou-sqp-%d", sqd->task_pid);
6664 set_task_comm(current, buf);
6665 sqd->thread = current;
6666 current->pf_io_worker = NULL;
6668 if (sqd->sq_cpu != -1)
6669 set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
6671 set_cpus_allowed_ptr(current, cpu_online_mask);
6672 current->flags |= PF_NO_SETAFFINITY;
6674 complete(&sqd->completion);
6676 wait_for_completion(&sqd->startup);
6678 while (!io_sq_thread_should_stop(sqd)) {
6680 bool cap_entries, sqt_spin, needs_sched;
6683 * Any changes to the sqd lists are synchronized through the
6684 * thread parking. This synchronizes the thread vs users,
6685 * the users are synchronized on the sqd->ctx_lock.
6687 if (io_sq_thread_should_park(sqd)) {
6688 io_sq_thread_parkme(sqd);
6691 if (unlikely(!list_empty(&sqd->ctx_new_list))) {
6692 io_sqd_init_new(sqd);
6693 timeout = jiffies + sqd->sq_thread_idle;
6695 if (fatal_signal_pending(current))
6698 cap_entries = !list_is_singular(&sqd->ctx_list);
6699 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6700 ret = __io_sq_thread(ctx, cap_entries);
6701 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
6705 if (sqt_spin || !time_after(jiffies, timeout)) {
6709 timeout = jiffies + sqd->sq_thread_idle;
6714 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
6715 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6716 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6717 !list_empty_careful(&ctx->iopoll_list)) {
6718 needs_sched = false;
6721 if (io_sqring_entries(ctx)) {
6722 needs_sched = false;
6727 if (needs_sched && !io_sq_thread_should_park(sqd)) {
6728 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6729 io_ring_set_wakeup_flag(ctx);
6732 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6733 io_ring_clear_wakeup_flag(ctx);
6736 finish_wait(&sqd->wait, &wait);
6737 timeout = jiffies + sqd->sq_thread_idle;
6740 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6741 io_uring_cancel_sqpoll(ctx);
6745 if (io_sq_thread_should_park(sqd))
6746 io_sq_thread_parkme(sqd);
6749 * Clear thread under lock so that concurrent parks work correctly
6751 complete(&sqd->completion);
6752 mutex_lock(&sqd->lock);
6754 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6756 io_ring_set_wakeup_flag(ctx);
6759 complete(&sqd->exited);
6760 mutex_unlock(&sqd->lock);
6764 struct io_wait_queue {
6765 struct wait_queue_entry wq;
6766 struct io_ring_ctx *ctx;
6768 unsigned nr_timeouts;
6771 static inline bool io_should_wake(struct io_wait_queue *iowq)
6773 struct io_ring_ctx *ctx = iowq->ctx;
6776 * Wake up if we have enough events, or if a timeout occurred since we
6777 * started waiting. For timeouts, we always want to return to userspace,
6778 * regardless of event count.
6780 return io_cqring_events(ctx) >= iowq->to_wait ||
6781 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6784 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6785 int wake_flags, void *key)
6787 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6791 * Cannot safely flush overflowed CQEs from here, ensure we wake up
6792 * the task, and the next invocation will do it.
6794 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->cq_check_overflow))
6795 return autoremove_wake_function(curr, mode, wake_flags, key);
6799 static int io_run_task_work_sig(void)
6801 if (io_run_task_work())
6803 if (!signal_pending(current))
6805 if (test_tsk_thread_flag(current, TIF_NOTIFY_SIGNAL))
6806 return -ERESTARTSYS;
6810 /* when returns >0, the caller should retry */
6811 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
6812 struct io_wait_queue *iowq,
6813 signed long *timeout)
6817 /* make sure we run task_work before checking for signals */
6818 ret = io_run_task_work_sig();
6819 if (ret || io_should_wake(iowq))
6821 /* let the caller flush overflows, retry */
6822 if (test_bit(0, &ctx->cq_check_overflow))
6825 *timeout = schedule_timeout(*timeout);
6826 return !*timeout ? -ETIME : 1;
6830 * Wait until events become available, if we don't already have some. The
6831 * application must reap them itself, as they reside on the shared cq ring.
6833 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6834 const sigset_t __user *sig, size_t sigsz,
6835 struct __kernel_timespec __user *uts)
6837 struct io_wait_queue iowq = {
6840 .func = io_wake_function,
6841 .entry = LIST_HEAD_INIT(iowq.wq.entry),
6844 .to_wait = min_events,
6846 struct io_rings *rings = ctx->rings;
6847 signed long timeout = MAX_SCHEDULE_TIMEOUT;
6851 io_cqring_overflow_flush(ctx, false, NULL, NULL);
6852 if (io_cqring_events(ctx) >= min_events)
6854 if (!io_run_task_work())
6859 #ifdef CONFIG_COMPAT
6860 if (in_compat_syscall())
6861 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
6865 ret = set_user_sigmask(sig, sigsz);
6872 struct timespec64 ts;
6874 if (get_timespec64(&ts, uts))
6876 timeout = timespec64_to_jiffies(&ts);
6879 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
6880 trace_io_uring_cqring_wait(ctx, min_events);
6882 io_cqring_overflow_flush(ctx, false, NULL, NULL);
6883 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
6884 TASK_INTERRUPTIBLE);
6885 ret = io_cqring_wait_schedule(ctx, &iowq, &timeout);
6886 finish_wait(&ctx->wait, &iowq.wq);
6889 restore_saved_sigmask_unless(ret == -EINTR);
6891 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
6894 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
6896 #if defined(CONFIG_UNIX)
6897 if (ctx->ring_sock) {
6898 struct sock *sock = ctx->ring_sock->sk;
6899 struct sk_buff *skb;
6901 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
6907 for (i = 0; i < ctx->nr_user_files; i++) {
6910 file = io_file_from_index(ctx, i);
6917 static void io_rsrc_data_ref_zero(struct percpu_ref *ref)
6919 struct fixed_rsrc_data *data;
6921 data = container_of(ref, struct fixed_rsrc_data, refs);
6922 complete(&data->done);
6925 static inline void io_rsrc_ref_lock(struct io_ring_ctx *ctx)
6927 spin_lock_bh(&ctx->rsrc_ref_lock);
6930 static inline void io_rsrc_ref_unlock(struct io_ring_ctx *ctx)
6932 spin_unlock_bh(&ctx->rsrc_ref_lock);
6935 static void io_sqe_rsrc_set_node(struct io_ring_ctx *ctx,
6936 struct fixed_rsrc_data *rsrc_data,
6937 struct fixed_rsrc_ref_node *ref_node)
6939 io_rsrc_ref_lock(ctx);
6940 rsrc_data->node = ref_node;
6941 list_add_tail(&ref_node->node, &ctx->rsrc_ref_list);
6942 io_rsrc_ref_unlock(ctx);
6943 percpu_ref_get(&rsrc_data->refs);
6946 static void io_sqe_rsrc_kill_node(struct io_ring_ctx *ctx, struct fixed_rsrc_data *data)
6948 struct fixed_rsrc_ref_node *ref_node = NULL;
6950 io_rsrc_ref_lock(ctx);
6951 ref_node = data->node;
6953 io_rsrc_ref_unlock(ctx);
6955 percpu_ref_kill(&ref_node->refs);
6958 static int io_rsrc_ref_quiesce(struct fixed_rsrc_data *data,
6959 struct io_ring_ctx *ctx,
6960 void (*rsrc_put)(struct io_ring_ctx *ctx,
6961 struct io_rsrc_put *prsrc))
6963 struct fixed_rsrc_ref_node *backup_node;
6969 data->quiesce = true;
6972 backup_node = alloc_fixed_rsrc_ref_node(ctx);
6975 backup_node->rsrc_data = data;
6976 backup_node->rsrc_put = rsrc_put;
6978 io_sqe_rsrc_kill_node(ctx, data);
6979 percpu_ref_kill(&data->refs);
6980 flush_delayed_work(&ctx->rsrc_put_work);
6982 ret = wait_for_completion_interruptible(&data->done);
6986 percpu_ref_resurrect(&data->refs);
6987 io_sqe_rsrc_set_node(ctx, data, backup_node);
6989 reinit_completion(&data->done);
6990 mutex_unlock(&ctx->uring_lock);
6991 ret = io_run_task_work_sig();
6992 mutex_lock(&ctx->uring_lock);
6994 data->quiesce = false;
6997 destroy_fixed_rsrc_ref_node(backup_node);
7001 static struct fixed_rsrc_data *alloc_fixed_rsrc_data(struct io_ring_ctx *ctx)
7003 struct fixed_rsrc_data *data;
7005 data = kzalloc(sizeof(*data), GFP_KERNEL);
7009 if (percpu_ref_init(&data->refs, io_rsrc_data_ref_zero,
7010 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) {
7015 init_completion(&data->done);
7019 static void free_fixed_rsrc_data(struct fixed_rsrc_data *data)
7021 percpu_ref_exit(&data->refs);
7026 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7028 struct fixed_rsrc_data *data = ctx->file_data;
7029 unsigned nr_tables, i;
7033 * percpu_ref_is_dying() is to stop parallel files unregister
7034 * Since we possibly drop uring lock later in this function to
7037 if (!data || percpu_ref_is_dying(&data->refs))
7039 ret = io_rsrc_ref_quiesce(data, ctx, io_ring_file_put);
7043 __io_sqe_files_unregister(ctx);
7044 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
7045 for (i = 0; i < nr_tables; i++)
7046 kfree(data->table[i].files);
7047 free_fixed_rsrc_data(data);
7048 ctx->file_data = NULL;
7049 ctx->nr_user_files = 0;
7053 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7054 __releases(&sqd->lock)
7058 if (sqd->thread == current)
7060 clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7061 wake_up_state(sqd->thread, TASK_PARKED);
7062 mutex_unlock(&sqd->lock);
7065 static bool io_sq_thread_park(struct io_sq_data *sqd)
7066 __acquires(&sqd->lock)
7068 if (sqd->thread == current)
7070 mutex_lock(&sqd->lock);
7072 mutex_unlock(&sqd->lock);
7075 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7076 wake_up_process(sqd->thread);
7077 wait_for_completion(&sqd->completion);
7081 static void io_sq_thread_stop(struct io_sq_data *sqd)
7083 if (test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state))
7085 mutex_lock(&sqd->lock);
7087 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7088 WARN_ON_ONCE(test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state));
7089 wake_up_process(sqd->thread);
7090 mutex_unlock(&sqd->lock);
7091 wait_for_completion(&sqd->exited);
7092 WARN_ON_ONCE(sqd->thread);
7094 mutex_unlock(&sqd->lock);
7098 static void io_put_sq_data(struct io_sq_data *sqd)
7100 if (refcount_dec_and_test(&sqd->refs)) {
7101 io_sq_thread_stop(sqd);
7106 static void io_sq_thread_finish(struct io_ring_ctx *ctx)
7108 struct io_sq_data *sqd = ctx->sq_data;
7111 complete(&sqd->startup);
7113 wait_for_completion(&ctx->sq_thread_comp);
7114 io_sq_thread_park(sqd);
7117 mutex_lock(&sqd->ctx_lock);
7118 list_del(&ctx->sqd_list);
7119 io_sqd_update_thread_idle(sqd);
7120 mutex_unlock(&sqd->ctx_lock);
7123 io_sq_thread_unpark(sqd);
7125 io_put_sq_data(sqd);
7126 ctx->sq_data = NULL;
7130 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7132 struct io_ring_ctx *ctx_attach;
7133 struct io_sq_data *sqd;
7136 f = fdget(p->wq_fd);
7138 return ERR_PTR(-ENXIO);
7139 if (f.file->f_op != &io_uring_fops) {
7141 return ERR_PTR(-EINVAL);
7144 ctx_attach = f.file->private_data;
7145 sqd = ctx_attach->sq_data;
7148 return ERR_PTR(-EINVAL);
7151 refcount_inc(&sqd->refs);
7156 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p)
7158 struct io_sq_data *sqd;
7160 if (p->flags & IORING_SETUP_ATTACH_WQ)
7161 return io_attach_sq_data(p);
7163 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7165 return ERR_PTR(-ENOMEM);
7167 refcount_set(&sqd->refs, 1);
7168 INIT_LIST_HEAD(&sqd->ctx_list);
7169 INIT_LIST_HEAD(&sqd->ctx_new_list);
7170 mutex_init(&sqd->ctx_lock);
7171 mutex_init(&sqd->lock);
7172 init_waitqueue_head(&sqd->wait);
7173 init_completion(&sqd->startup);
7174 init_completion(&sqd->completion);
7175 init_completion(&sqd->exited);
7179 #if defined(CONFIG_UNIX)
7181 * Ensure the UNIX gc is aware of our file set, so we are certain that
7182 * the io_uring can be safely unregistered on process exit, even if we have
7183 * loops in the file referencing.
7185 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7187 struct sock *sk = ctx->ring_sock->sk;
7188 struct scm_fp_list *fpl;
7189 struct sk_buff *skb;
7192 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7196 skb = alloc_skb(0, GFP_KERNEL);
7205 fpl->user = get_uid(current_user());
7206 for (i = 0; i < nr; i++) {
7207 struct file *file = io_file_from_index(ctx, i + offset);
7211 fpl->fp[nr_files] = get_file(file);
7212 unix_inflight(fpl->user, fpl->fp[nr_files]);
7217 fpl->max = SCM_MAX_FD;
7218 fpl->count = nr_files;
7219 UNIXCB(skb).fp = fpl;
7220 skb->destructor = unix_destruct_scm;
7221 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7222 skb_queue_head(&sk->sk_receive_queue, skb);
7224 for (i = 0; i < nr_files; i++)
7235 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7236 * causes regular reference counting to break down. We rely on the UNIX
7237 * garbage collection to take care of this problem for us.
7239 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7241 unsigned left, total;
7245 left = ctx->nr_user_files;
7247 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7249 ret = __io_sqe_files_scm(ctx, this_files, total);
7253 total += this_files;
7259 while (total < ctx->nr_user_files) {
7260 struct file *file = io_file_from_index(ctx, total);
7270 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7276 static int io_sqe_alloc_file_tables(struct fixed_rsrc_data *file_data,
7277 unsigned nr_tables, unsigned nr_files)
7281 for (i = 0; i < nr_tables; i++) {
7282 struct fixed_rsrc_table *table = &file_data->table[i];
7283 unsigned this_files;
7285 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
7286 table->files = kcalloc(this_files, sizeof(struct file *),
7290 nr_files -= this_files;
7296 for (i = 0; i < nr_tables; i++) {
7297 struct fixed_rsrc_table *table = &file_data->table[i];
7298 kfree(table->files);
7303 static void io_ring_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
7305 struct file *file = prsrc->file;
7306 #if defined(CONFIG_UNIX)
7307 struct sock *sock = ctx->ring_sock->sk;
7308 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7309 struct sk_buff *skb;
7312 __skb_queue_head_init(&list);
7315 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7316 * remove this entry and rearrange the file array.
7318 skb = skb_dequeue(head);
7320 struct scm_fp_list *fp;
7322 fp = UNIXCB(skb).fp;
7323 for (i = 0; i < fp->count; i++) {
7326 if (fp->fp[i] != file)
7329 unix_notinflight(fp->user, fp->fp[i]);
7330 left = fp->count - 1 - i;
7332 memmove(&fp->fp[i], &fp->fp[i + 1],
7333 left * sizeof(struct file *));
7340 __skb_queue_tail(&list, skb);
7350 __skb_queue_tail(&list, skb);
7352 skb = skb_dequeue(head);
7355 if (skb_peek(&list)) {
7356 spin_lock_irq(&head->lock);
7357 while ((skb = __skb_dequeue(&list)) != NULL)
7358 __skb_queue_tail(head, skb);
7359 spin_unlock_irq(&head->lock);
7366 static void __io_rsrc_put_work(struct fixed_rsrc_ref_node *ref_node)
7368 struct fixed_rsrc_data *rsrc_data = ref_node->rsrc_data;
7369 struct io_ring_ctx *ctx = rsrc_data->ctx;
7370 struct io_rsrc_put *prsrc, *tmp;
7372 list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
7373 list_del(&prsrc->list);
7374 ref_node->rsrc_put(ctx, prsrc);
7378 percpu_ref_exit(&ref_node->refs);
7380 percpu_ref_put(&rsrc_data->refs);
7383 static void io_rsrc_put_work(struct work_struct *work)
7385 struct io_ring_ctx *ctx;
7386 struct llist_node *node;
7388 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
7389 node = llist_del_all(&ctx->rsrc_put_llist);
7392 struct fixed_rsrc_ref_node *ref_node;
7393 struct llist_node *next = node->next;
7395 ref_node = llist_entry(node, struct fixed_rsrc_ref_node, llist);
7396 __io_rsrc_put_work(ref_node);
7401 static struct file **io_fixed_file_slot(struct fixed_rsrc_data *file_data,
7404 struct fixed_rsrc_table *table;
7406 table = &file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7407 return &table->files[i & IORING_FILE_TABLE_MASK];
7410 static void io_rsrc_node_ref_zero(struct percpu_ref *ref)
7412 struct fixed_rsrc_ref_node *ref_node;
7413 struct fixed_rsrc_data *data;
7414 struct io_ring_ctx *ctx;
7415 bool first_add = false;
7418 ref_node = container_of(ref, struct fixed_rsrc_ref_node, refs);
7419 data = ref_node->rsrc_data;
7422 io_rsrc_ref_lock(ctx);
7423 ref_node->done = true;
7425 while (!list_empty(&ctx->rsrc_ref_list)) {
7426 ref_node = list_first_entry(&ctx->rsrc_ref_list,
7427 struct fixed_rsrc_ref_node, node);
7428 /* recycle ref nodes in order */
7429 if (!ref_node->done)
7431 list_del(&ref_node->node);
7432 first_add |= llist_add(&ref_node->llist, &ctx->rsrc_put_llist);
7434 io_rsrc_ref_unlock(ctx);
7436 if (percpu_ref_is_dying(&data->refs))
7440 mod_delayed_work(system_wq, &ctx->rsrc_put_work, 0);
7442 queue_delayed_work(system_wq, &ctx->rsrc_put_work, delay);
7445 static struct fixed_rsrc_ref_node *alloc_fixed_rsrc_ref_node(
7446 struct io_ring_ctx *ctx)
7448 struct fixed_rsrc_ref_node *ref_node;
7450 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7454 if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
7459 INIT_LIST_HEAD(&ref_node->node);
7460 INIT_LIST_HEAD(&ref_node->rsrc_list);
7461 ref_node->done = false;
7465 static void init_fixed_file_ref_node(struct io_ring_ctx *ctx,
7466 struct fixed_rsrc_ref_node *ref_node)
7468 ref_node->rsrc_data = ctx->file_data;
7469 ref_node->rsrc_put = io_ring_file_put;
7472 static void destroy_fixed_rsrc_ref_node(struct fixed_rsrc_ref_node *ref_node)
7474 percpu_ref_exit(&ref_node->refs);
7479 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7482 __s32 __user *fds = (__s32 __user *) arg;
7483 unsigned nr_tables, i;
7485 int fd, ret = -ENOMEM;
7486 struct fixed_rsrc_ref_node *ref_node;
7487 struct fixed_rsrc_data *file_data;
7493 if (nr_args > IORING_MAX_FIXED_FILES)
7496 file_data = alloc_fixed_rsrc_data(ctx);
7499 ctx->file_data = file_data;
7501 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
7502 file_data->table = kcalloc(nr_tables, sizeof(*file_data->table),
7504 if (!file_data->table)
7507 if (io_sqe_alloc_file_tables(file_data, nr_tables, nr_args))
7510 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7511 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7515 /* allow sparse sets */
7525 * Don't allow io_uring instances to be registered. If UNIX
7526 * isn't enabled, then this causes a reference cycle and this
7527 * instance can never get freed. If UNIX is enabled we'll
7528 * handle it just fine, but there's still no point in allowing
7529 * a ring fd as it doesn't support regular read/write anyway.
7531 if (file->f_op == &io_uring_fops) {
7535 *io_fixed_file_slot(file_data, i) = file;
7538 ret = io_sqe_files_scm(ctx);
7540 io_sqe_files_unregister(ctx);
7544 ref_node = alloc_fixed_rsrc_ref_node(ctx);
7546 io_sqe_files_unregister(ctx);
7549 init_fixed_file_ref_node(ctx, ref_node);
7551 io_sqe_rsrc_set_node(ctx, file_data, ref_node);
7554 for (i = 0; i < ctx->nr_user_files; i++) {
7555 file = io_file_from_index(ctx, i);
7559 for (i = 0; i < nr_tables; i++)
7560 kfree(file_data->table[i].files);
7561 ctx->nr_user_files = 0;
7563 free_fixed_rsrc_data(ctx->file_data);
7564 ctx->file_data = NULL;
7568 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7571 #if defined(CONFIG_UNIX)
7572 struct sock *sock = ctx->ring_sock->sk;
7573 struct sk_buff_head *head = &sock->sk_receive_queue;
7574 struct sk_buff *skb;
7577 * See if we can merge this file into an existing skb SCM_RIGHTS
7578 * file set. If there's no room, fall back to allocating a new skb
7579 * and filling it in.
7581 spin_lock_irq(&head->lock);
7582 skb = skb_peek(head);
7584 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7586 if (fpl->count < SCM_MAX_FD) {
7587 __skb_unlink(skb, head);
7588 spin_unlock_irq(&head->lock);
7589 fpl->fp[fpl->count] = get_file(file);
7590 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7592 spin_lock_irq(&head->lock);
7593 __skb_queue_head(head, skb);
7598 spin_unlock_irq(&head->lock);
7605 return __io_sqe_files_scm(ctx, 1, index);
7611 static int io_queue_rsrc_removal(struct fixed_rsrc_data *data, void *rsrc)
7613 struct io_rsrc_put *prsrc;
7614 struct fixed_rsrc_ref_node *ref_node = data->node;
7616 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
7621 list_add(&prsrc->list, &ref_node->rsrc_list);
7626 static inline int io_queue_file_removal(struct fixed_rsrc_data *data,
7629 return io_queue_rsrc_removal(data, (void *)file);
7632 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7633 struct io_uring_rsrc_update *up,
7636 struct fixed_rsrc_data *data = ctx->file_data;
7637 struct fixed_rsrc_ref_node *ref_node;
7638 struct file *file, **file_slot;
7642 bool needs_switch = false;
7644 if (check_add_overflow(up->offset, nr_args, &done))
7646 if (done > ctx->nr_user_files)
7649 ref_node = alloc_fixed_rsrc_ref_node(ctx);
7652 init_fixed_file_ref_node(ctx, ref_node);
7654 fds = u64_to_user_ptr(up->data);
7655 for (done = 0; done < nr_args; done++) {
7657 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
7661 if (fd == IORING_REGISTER_FILES_SKIP)
7664 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
7665 file_slot = io_fixed_file_slot(ctx->file_data, i);
7668 err = io_queue_file_removal(data, *file_slot);
7672 needs_switch = true;
7681 * Don't allow io_uring instances to be registered. If
7682 * UNIX isn't enabled, then this causes a reference
7683 * cycle and this instance can never get freed. If UNIX
7684 * is enabled we'll handle it just fine, but there's
7685 * still no point in allowing a ring fd as it doesn't
7686 * support regular read/write anyway.
7688 if (file->f_op == &io_uring_fops) {
7694 err = io_sqe_file_register(ctx, file, i);
7704 percpu_ref_kill(&data->node->refs);
7705 io_sqe_rsrc_set_node(ctx, data, ref_node);
7707 destroy_fixed_rsrc_ref_node(ref_node);
7709 return done ? done : err;
7712 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
7715 struct io_uring_rsrc_update up;
7717 if (!ctx->file_data)
7721 if (copy_from_user(&up, arg, sizeof(up)))
7726 return __io_sqe_files_update(ctx, &up, nr_args);
7729 static struct io_wq_work *io_free_work(struct io_wq_work *work)
7731 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7733 req = io_put_req_find_next(req);
7734 return req ? &req->work : NULL;
7737 static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx)
7739 struct io_wq_hash *hash;
7740 struct io_wq_data data;
7741 unsigned int concurrency;
7743 hash = ctx->hash_map;
7745 hash = kzalloc(sizeof(*hash), GFP_KERNEL);
7747 return ERR_PTR(-ENOMEM);
7748 refcount_set(&hash->refs, 1);
7749 init_waitqueue_head(&hash->wait);
7750 ctx->hash_map = hash;
7754 data.free_work = io_free_work;
7755 data.do_work = io_wq_submit_work;
7757 /* Do QD, or 4 * CPUS, whatever is smallest */
7758 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7760 return io_wq_create(concurrency, &data);
7763 static int io_uring_alloc_task_context(struct task_struct *task,
7764 struct io_ring_ctx *ctx)
7766 struct io_uring_task *tctx;
7769 tctx = kmalloc(sizeof(*tctx), GFP_KERNEL);
7770 if (unlikely(!tctx))
7773 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
7774 if (unlikely(ret)) {
7779 tctx->io_wq = io_init_wq_offload(ctx);
7780 if (IS_ERR(tctx->io_wq)) {
7781 ret = PTR_ERR(tctx->io_wq);
7782 percpu_counter_destroy(&tctx->inflight);
7788 init_waitqueue_head(&tctx->wait);
7790 atomic_set(&tctx->in_idle, 0);
7791 tctx->sqpoll = false;
7792 task->io_uring = tctx;
7793 spin_lock_init(&tctx->task_lock);
7794 INIT_WQ_LIST(&tctx->task_list);
7795 tctx->task_state = 0;
7796 init_task_work(&tctx->task_work, tctx_task_work);
7800 void __io_uring_free(struct task_struct *tsk)
7802 struct io_uring_task *tctx = tsk->io_uring;
7804 WARN_ON_ONCE(!xa_empty(&tctx->xa));
7805 WARN_ON_ONCE(tctx->io_wq);
7807 percpu_counter_destroy(&tctx->inflight);
7809 tsk->io_uring = NULL;
7812 static int io_sq_thread_fork(struct io_sq_data *sqd, struct io_ring_ctx *ctx)
7816 clear_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7817 reinit_completion(&sqd->completion);
7818 ctx->sqo_dead = ctx->sqo_exec = 0;
7819 sqd->task_pid = current->pid;
7820 current->flags |= PF_IO_WORKER;
7821 ret = io_wq_fork_thread(io_sq_thread, sqd);
7822 current->flags &= ~PF_IO_WORKER;
7827 wait_for_completion(&sqd->completion);
7828 return io_uring_alloc_task_context(sqd->thread, ctx);
7831 static int io_sq_offload_create(struct io_ring_ctx *ctx,
7832 struct io_uring_params *p)
7836 /* Retain compatibility with failing for an invalid attach attempt */
7837 if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
7838 IORING_SETUP_ATTACH_WQ) {
7841 f = fdget(p->wq_fd);
7844 if (f.file->f_op != &io_uring_fops) {
7850 if (ctx->flags & IORING_SETUP_SQPOLL) {
7851 struct io_sq_data *sqd;
7854 if (!capable(CAP_SYS_ADMIN) && !capable(CAP_SYS_NICE))
7857 sqd = io_get_sq_data(p);
7864 io_sq_thread_park(sqd);
7865 mutex_lock(&sqd->ctx_lock);
7866 list_add(&ctx->sqd_list, &sqd->ctx_new_list);
7867 mutex_unlock(&sqd->ctx_lock);
7868 io_sq_thread_unpark(sqd);
7870 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
7871 if (!ctx->sq_thread_idle)
7872 ctx->sq_thread_idle = HZ;
7877 if (p->flags & IORING_SETUP_SQ_AFF) {
7878 int cpu = p->sq_thread_cpu;
7881 if (cpu >= nr_cpu_ids)
7883 if (!cpu_online(cpu))
7891 sqd->task_pid = current->pid;
7892 current->flags |= PF_IO_WORKER;
7893 ret = io_wq_fork_thread(io_sq_thread, sqd);
7894 current->flags &= ~PF_IO_WORKER;
7899 wait_for_completion(&sqd->completion);
7900 ret = io_uring_alloc_task_context(sqd->thread, ctx);
7903 } else if (p->flags & IORING_SETUP_SQ_AFF) {
7904 /* Can't have SQ_AFF without SQPOLL */
7911 io_sq_thread_finish(ctx);
7915 static void io_sq_offload_start(struct io_ring_ctx *ctx)
7917 struct io_sq_data *sqd = ctx->sq_data;
7919 if (ctx->flags & IORING_SETUP_SQPOLL)
7920 complete(&sqd->startup);
7923 static inline void __io_unaccount_mem(struct user_struct *user,
7924 unsigned long nr_pages)
7926 atomic_long_sub(nr_pages, &user->locked_vm);
7929 static inline int __io_account_mem(struct user_struct *user,
7930 unsigned long nr_pages)
7932 unsigned long page_limit, cur_pages, new_pages;
7934 /* Don't allow more pages than we can safely lock */
7935 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
7938 cur_pages = atomic_long_read(&user->locked_vm);
7939 new_pages = cur_pages + nr_pages;
7940 if (new_pages > page_limit)
7942 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
7943 new_pages) != cur_pages);
7948 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
7951 __io_unaccount_mem(ctx->user, nr_pages);
7953 if (ctx->mm_account)
7954 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
7957 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
7962 ret = __io_account_mem(ctx->user, nr_pages);
7967 if (ctx->mm_account)
7968 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
7973 static void io_mem_free(void *ptr)
7980 page = virt_to_head_page(ptr);
7981 if (put_page_testzero(page))
7982 free_compound_page(page);
7985 static void *io_mem_alloc(size_t size)
7987 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
7988 __GFP_NORETRY | __GFP_ACCOUNT;
7990 return (void *) __get_free_pages(gfp_flags, get_order(size));
7993 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
7996 struct io_rings *rings;
7997 size_t off, sq_array_size;
7999 off = struct_size(rings, cqes, cq_entries);
8000 if (off == SIZE_MAX)
8004 off = ALIGN(off, SMP_CACHE_BYTES);
8012 sq_array_size = array_size(sizeof(u32), sq_entries);
8013 if (sq_array_size == SIZE_MAX)
8016 if (check_add_overflow(off, sq_array_size, &off))
8022 static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8026 if (!ctx->user_bufs)
8029 for (i = 0; i < ctx->nr_user_bufs; i++) {
8030 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8032 for (j = 0; j < imu->nr_bvecs; j++)
8033 unpin_user_page(imu->bvec[j].bv_page);
8035 if (imu->acct_pages)
8036 io_unaccount_mem(ctx, imu->acct_pages);
8041 kfree(ctx->user_bufs);
8042 ctx->user_bufs = NULL;
8043 ctx->nr_user_bufs = 0;
8047 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8048 void __user *arg, unsigned index)
8050 struct iovec __user *src;
8052 #ifdef CONFIG_COMPAT
8054 struct compat_iovec __user *ciovs;
8055 struct compat_iovec ciov;
8057 ciovs = (struct compat_iovec __user *) arg;
8058 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8061 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8062 dst->iov_len = ciov.iov_len;
8066 src = (struct iovec __user *) arg;
8067 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8073 * Not super efficient, but this is just a registration time. And we do cache
8074 * the last compound head, so generally we'll only do a full search if we don't
8077 * We check if the given compound head page has already been accounted, to
8078 * avoid double accounting it. This allows us to account the full size of the
8079 * page, not just the constituent pages of a huge page.
8081 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8082 int nr_pages, struct page *hpage)
8086 /* check current page array */
8087 for (i = 0; i < nr_pages; i++) {
8088 if (!PageCompound(pages[i]))
8090 if (compound_head(pages[i]) == hpage)
8094 /* check previously registered pages */
8095 for (i = 0; i < ctx->nr_user_bufs; i++) {
8096 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8098 for (j = 0; j < imu->nr_bvecs; j++) {
8099 if (!PageCompound(imu->bvec[j].bv_page))
8101 if (compound_head(imu->bvec[j].bv_page) == hpage)
8109 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8110 int nr_pages, struct io_mapped_ubuf *imu,
8111 struct page **last_hpage)
8115 for (i = 0; i < nr_pages; i++) {
8116 if (!PageCompound(pages[i])) {
8121 hpage = compound_head(pages[i]);
8122 if (hpage == *last_hpage)
8124 *last_hpage = hpage;
8125 if (headpage_already_acct(ctx, pages, i, hpage))
8127 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8131 if (!imu->acct_pages)
8134 ret = io_account_mem(ctx, imu->acct_pages);
8136 imu->acct_pages = 0;
8140 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
8141 struct io_mapped_ubuf *imu,
8142 struct page **last_hpage)
8144 struct vm_area_struct **vmas = NULL;
8145 struct page **pages = NULL;
8146 unsigned long off, start, end, ubuf;
8148 int ret, pret, nr_pages, i;
8150 ubuf = (unsigned long) iov->iov_base;
8151 end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8152 start = ubuf >> PAGE_SHIFT;
8153 nr_pages = end - start;
8157 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
8161 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
8166 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
8172 mmap_read_lock(current->mm);
8173 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
8175 if (pret == nr_pages) {
8176 /* don't support file backed memory */
8177 for (i = 0; i < nr_pages; i++) {
8178 struct vm_area_struct *vma = vmas[i];
8181 !is_file_hugepages(vma->vm_file)) {
8187 ret = pret < 0 ? pret : -EFAULT;
8189 mmap_read_unlock(current->mm);
8192 * if we did partial map, or found file backed vmas,
8193 * release any pages we did get
8196 unpin_user_pages(pages, pret);
8201 ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
8203 unpin_user_pages(pages, pret);
8208 off = ubuf & ~PAGE_MASK;
8209 size = iov->iov_len;
8210 for (i = 0; i < nr_pages; i++) {
8213 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8214 imu->bvec[i].bv_page = pages[i];
8215 imu->bvec[i].bv_len = vec_len;
8216 imu->bvec[i].bv_offset = off;
8220 /* store original address for later verification */
8222 imu->len = iov->iov_len;
8223 imu->nr_bvecs = nr_pages;
8231 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
8235 if (!nr_args || nr_args > UIO_MAXIOV)
8238 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
8240 if (!ctx->user_bufs)
8246 static int io_buffer_validate(struct iovec *iov)
8249 * Don't impose further limits on the size and buffer
8250 * constraints here, we'll -EINVAL later when IO is
8251 * submitted if they are wrong.
8253 if (!iov->iov_base || !iov->iov_len)
8256 /* arbitrary limit, but we need something */
8257 if (iov->iov_len > SZ_1G)
8263 static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
8264 unsigned int nr_args)
8268 struct page *last_hpage = NULL;
8270 ret = io_buffers_map_alloc(ctx, nr_args);
8274 for (i = 0; i < nr_args; i++) {
8275 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8277 ret = io_copy_iov(ctx, &iov, arg, i);
8281 ret = io_buffer_validate(&iov);
8285 ret = io_sqe_buffer_register(ctx, &iov, imu, &last_hpage);
8289 ctx->nr_user_bufs++;
8293 io_sqe_buffers_unregister(ctx);
8298 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8300 __s32 __user *fds = arg;
8306 if (copy_from_user(&fd, fds, sizeof(*fds)))
8309 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8310 if (IS_ERR(ctx->cq_ev_fd)) {
8311 int ret = PTR_ERR(ctx->cq_ev_fd);
8312 ctx->cq_ev_fd = NULL;
8319 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8321 if (ctx->cq_ev_fd) {
8322 eventfd_ctx_put(ctx->cq_ev_fd);
8323 ctx->cq_ev_fd = NULL;
8330 static int __io_destroy_buffers(int id, void *p, void *data)
8332 struct io_ring_ctx *ctx = data;
8333 struct io_buffer *buf = p;
8335 __io_remove_buffers(ctx, buf, id, -1U);
8339 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8341 idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx);
8342 idr_destroy(&ctx->io_buffer_idr);
8345 static void io_req_cache_free(struct list_head *list, struct task_struct *tsk)
8347 struct io_kiocb *req, *nxt;
8349 list_for_each_entry_safe(req, nxt, list, compl.list) {
8350 if (tsk && req->task != tsk)
8352 list_del(&req->compl.list);
8353 kmem_cache_free(req_cachep, req);
8357 static void io_req_caches_free(struct io_ring_ctx *ctx)
8359 struct io_submit_state *submit_state = &ctx->submit_state;
8360 struct io_comp_state *cs = &ctx->submit_state.comp;
8362 mutex_lock(&ctx->uring_lock);
8364 if (submit_state->free_reqs) {
8365 kmem_cache_free_bulk(req_cachep, submit_state->free_reqs,
8366 submit_state->reqs);
8367 submit_state->free_reqs = 0;
8370 spin_lock_irq(&ctx->completion_lock);
8371 list_splice_init(&cs->locked_free_list, &cs->free_list);
8372 cs->locked_free_nr = 0;
8373 spin_unlock_irq(&ctx->completion_lock);
8375 io_req_cache_free(&cs->free_list, NULL);
8377 mutex_unlock(&ctx->uring_lock);
8380 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8383 * Some may use context even when all refs and requests have been put,
8384 * and they are free to do so while still holding uring_lock, see
8385 * __io_req_task_submit(). Wait for them to finish.
8387 mutex_lock(&ctx->uring_lock);
8388 mutex_unlock(&ctx->uring_lock);
8390 io_sq_thread_finish(ctx);
8391 io_sqe_buffers_unregister(ctx);
8393 if (ctx->mm_account) {
8394 mmdrop(ctx->mm_account);
8395 ctx->mm_account = NULL;
8398 mutex_lock(&ctx->uring_lock);
8399 io_sqe_files_unregister(ctx);
8400 mutex_unlock(&ctx->uring_lock);
8401 io_eventfd_unregister(ctx);
8402 io_destroy_buffers(ctx);
8403 idr_destroy(&ctx->personality_idr);
8405 #if defined(CONFIG_UNIX)
8406 if (ctx->ring_sock) {
8407 ctx->ring_sock->file = NULL; /* so that iput() is called */
8408 sock_release(ctx->ring_sock);
8412 io_mem_free(ctx->rings);
8413 io_mem_free(ctx->sq_sqes);
8415 percpu_ref_exit(&ctx->refs);
8416 free_uid(ctx->user);
8417 io_req_caches_free(ctx);
8419 io_wq_put_hash(ctx->hash_map);
8420 kfree(ctx->cancel_hash);
8424 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8426 struct io_ring_ctx *ctx = file->private_data;
8429 poll_wait(file, &ctx->cq_wait, wait);
8431 * synchronizes with barrier from wq_has_sleeper call in
8435 if (!io_sqring_full(ctx))
8436 mask |= EPOLLOUT | EPOLLWRNORM;
8439 * Don't flush cqring overflow list here, just do a simple check.
8440 * Otherwise there could possible be ABBA deadlock:
8443 * lock(&ctx->uring_lock);
8445 * lock(&ctx->uring_lock);
8448 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
8449 * pushs them to do the flush.
8451 if (io_cqring_events(ctx) || test_bit(0, &ctx->cq_check_overflow))
8452 mask |= EPOLLIN | EPOLLRDNORM;
8457 static int io_uring_fasync(int fd, struct file *file, int on)
8459 struct io_ring_ctx *ctx = file->private_data;
8461 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8464 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8466 const struct cred *creds;
8468 creds = idr_remove(&ctx->personality_idr, id);
8477 static int io_remove_personalities(int id, void *p, void *data)
8479 struct io_ring_ctx *ctx = data;
8481 io_unregister_personality(ctx, id);
8485 static bool io_run_ctx_fallback(struct io_ring_ctx *ctx)
8487 struct callback_head *work, *head, *next;
8488 bool executed = false;
8493 work = READ_ONCE(ctx->exit_task_work);
8494 } while (cmpxchg(&ctx->exit_task_work, work, head) != work);
8511 static void io_ring_exit_work(struct work_struct *work)
8513 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
8517 * If we're doing polled IO and end up having requests being
8518 * submitted async (out-of-line), then completions can come in while
8519 * we're waiting for refs to drop. We need to reap these manually,
8520 * as nobody else will be looking for them.
8523 io_uring_try_cancel_requests(ctx, NULL, NULL);
8524 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
8525 io_ring_ctx_free(ctx);
8528 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8530 mutex_lock(&ctx->uring_lock);
8531 percpu_ref_kill(&ctx->refs);
8533 if (WARN_ON_ONCE((ctx->flags & IORING_SETUP_SQPOLL) && !ctx->sqo_dead))
8536 /* if force is set, the ring is going away. always drop after that */
8537 ctx->cq_overflow_flushed = 1;
8539 __io_cqring_overflow_flush(ctx, true, NULL, NULL);
8540 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
8541 mutex_unlock(&ctx->uring_lock);
8543 io_kill_timeouts(ctx, NULL, NULL);
8544 io_poll_remove_all(ctx, NULL, NULL);
8546 /* if we failed setting up the ctx, we might not have any rings */
8547 io_iopoll_try_reap_events(ctx);
8549 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8551 * Use system_unbound_wq to avoid spawning tons of event kworkers
8552 * if we're exiting a ton of rings at the same time. It just adds
8553 * noise and overhead, there's no discernable change in runtime
8554 * over using system_wq.
8556 queue_work(system_unbound_wq, &ctx->exit_work);
8559 static int io_uring_release(struct inode *inode, struct file *file)
8561 struct io_ring_ctx *ctx = file->private_data;
8563 file->private_data = NULL;
8564 io_ring_ctx_wait_and_kill(ctx);
8568 struct io_task_cancel {
8569 struct task_struct *task;
8570 struct files_struct *files;
8573 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8575 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8576 struct io_task_cancel *cancel = data;
8579 if (cancel->files && (req->flags & REQ_F_LINK_TIMEOUT)) {
8580 unsigned long flags;
8581 struct io_ring_ctx *ctx = req->ctx;
8583 /* protect against races with linked timeouts */
8584 spin_lock_irqsave(&ctx->completion_lock, flags);
8585 ret = io_match_task(req, cancel->task, cancel->files);
8586 spin_unlock_irqrestore(&ctx->completion_lock, flags);
8588 ret = io_match_task(req, cancel->task, cancel->files);
8593 static void io_cancel_defer_files(struct io_ring_ctx *ctx,
8594 struct task_struct *task,
8595 struct files_struct *files)
8597 struct io_defer_entry *de = NULL;
8600 spin_lock_irq(&ctx->completion_lock);
8601 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8602 if (io_match_task(de->req, task, files)) {
8603 list_cut_position(&list, &ctx->defer_list, &de->list);
8607 spin_unlock_irq(&ctx->completion_lock);
8609 while (!list_empty(&list)) {
8610 de = list_first_entry(&list, struct io_defer_entry, list);
8611 list_del_init(&de->list);
8612 req_set_fail_links(de->req);
8613 io_put_req(de->req);
8614 io_req_complete(de->req, -ECANCELED);
8619 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
8620 struct task_struct *task,
8621 struct files_struct *files)
8623 struct io_task_cancel cancel = { .task = task, .files = files, };
8624 struct io_uring_task *tctx = current->io_uring;
8627 enum io_wq_cancel cret;
8630 if (tctx && tctx->io_wq) {
8631 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
8633 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8636 /* SQPOLL thread does its own polling */
8637 if (!(ctx->flags & IORING_SETUP_SQPOLL) && !files) {
8638 while (!list_empty_careful(&ctx->iopoll_list)) {
8639 io_iopoll_try_reap_events(ctx);
8644 ret |= io_poll_remove_all(ctx, task, files);
8645 ret |= io_kill_timeouts(ctx, task, files);
8646 ret |= io_run_task_work();
8647 ret |= io_run_ctx_fallback(ctx);
8648 io_cqring_overflow_flush(ctx, true, task, files);
8655 static int io_uring_count_inflight(struct io_ring_ctx *ctx,
8656 struct task_struct *task,
8657 struct files_struct *files)
8659 struct io_kiocb *req;
8662 spin_lock_irq(&ctx->inflight_lock);
8663 list_for_each_entry(req, &ctx->inflight_list, inflight_entry)
8664 cnt += io_match_task(req, task, files);
8665 spin_unlock_irq(&ctx->inflight_lock);
8669 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
8670 struct task_struct *task,
8671 struct files_struct *files)
8673 while (!list_empty_careful(&ctx->inflight_list)) {
8677 inflight = io_uring_count_inflight(ctx, task, files);
8681 io_uring_try_cancel_requests(ctx, task, files);
8684 io_sq_thread_unpark(ctx->sq_data);
8685 prepare_to_wait(&task->io_uring->wait, &wait,
8686 TASK_UNINTERRUPTIBLE);
8687 if (inflight == io_uring_count_inflight(ctx, task, files))
8689 finish_wait(&task->io_uring->wait, &wait);
8691 io_sq_thread_park(ctx->sq_data);
8695 static void io_disable_sqo_submit(struct io_ring_ctx *ctx)
8697 mutex_lock(&ctx->uring_lock);
8699 mutex_unlock(&ctx->uring_lock);
8701 /* make sure callers enter the ring to get error */
8703 io_ring_set_wakeup_flag(ctx);
8707 * We need to iteratively cancel requests, in case a request has dependent
8708 * hard links. These persist even for failure of cancelations, hence keep
8709 * looping until none are found.
8711 static void io_uring_cancel_task_requests(struct io_ring_ctx *ctx,
8712 struct files_struct *files)
8714 struct task_struct *task = current;
8715 bool did_park = false;
8717 if ((ctx->flags & IORING_SETUP_SQPOLL) && ctx->sq_data) {
8718 io_disable_sqo_submit(ctx);
8719 did_park = io_sq_thread_park(ctx->sq_data);
8721 task = ctx->sq_data->thread;
8722 atomic_inc(&task->io_uring->in_idle);
8726 io_cancel_defer_files(ctx, task, files);
8728 io_uring_cancel_files(ctx, task, files);
8730 io_uring_try_cancel_requests(ctx, task, NULL);
8733 atomic_dec(&task->io_uring->in_idle);
8734 io_sq_thread_unpark(ctx->sq_data);
8739 * Note that this task has used io_uring. We use it for cancelation purposes.
8741 static int io_uring_add_task_file(struct io_ring_ctx *ctx, struct file *file)
8743 struct io_uring_task *tctx = current->io_uring;
8746 if (unlikely(!tctx)) {
8747 ret = io_uring_alloc_task_context(current, ctx);
8750 tctx = current->io_uring;
8752 if (tctx->last != file) {
8753 void *old = xa_load(&tctx->xa, (unsigned long)file);
8757 ret = xa_err(xa_store(&tctx->xa, (unsigned long)file,
8764 /* one and only SQPOLL file note, held by sqo_task */
8765 WARN_ON_ONCE((ctx->flags & IORING_SETUP_SQPOLL) &&
8766 current != ctx->sqo_task);
8772 * This is race safe in that the task itself is doing this, hence it
8773 * cannot be going through the exit/cancel paths at the same time.
8774 * This cannot be modified while exit/cancel is running.
8776 if (!tctx->sqpoll && (ctx->flags & IORING_SETUP_SQPOLL))
8777 tctx->sqpoll = true;
8783 * Remove this io_uring_file -> task mapping.
8785 static void io_uring_del_task_file(struct file *file)
8787 struct io_uring_task *tctx = current->io_uring;
8789 if (tctx->last == file)
8791 file = xa_erase(&tctx->xa, (unsigned long)file);
8796 static void io_uring_clean_tctx(struct io_uring_task *tctx)
8799 unsigned long index;
8801 xa_for_each(&tctx->xa, index, file)
8802 io_uring_del_task_file(file);
8804 io_wq_put_and_exit(tctx->io_wq);
8809 void __io_uring_files_cancel(struct files_struct *files)
8811 struct io_uring_task *tctx = current->io_uring;
8813 unsigned long index;
8815 /* make sure overflow events are dropped */
8816 atomic_inc(&tctx->in_idle);
8817 xa_for_each(&tctx->xa, index, file)
8818 io_uring_cancel_task_requests(file->private_data, files);
8819 atomic_dec(&tctx->in_idle);
8822 io_uring_clean_tctx(tctx);
8825 static s64 tctx_inflight(struct io_uring_task *tctx)
8827 return percpu_counter_sum(&tctx->inflight);
8830 static void io_uring_cancel_sqpoll(struct io_ring_ctx *ctx)
8832 struct io_sq_data *sqd = ctx->sq_data;
8833 struct io_uring_task *tctx;
8839 io_disable_sqo_submit(ctx);
8840 if (!io_sq_thread_park(sqd))
8842 tctx = ctx->sq_data->thread->io_uring;
8843 /* can happen on fork/alloc failure, just ignore that state */
8845 io_sq_thread_unpark(sqd);
8849 atomic_inc(&tctx->in_idle);
8851 /* read completions before cancelations */
8852 inflight = tctx_inflight(tctx);
8855 io_uring_cancel_task_requests(ctx, NULL);
8857 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
8859 * If we've seen completions, retry without waiting. This
8860 * avoids a race where a completion comes in before we did
8861 * prepare_to_wait().
8863 if (inflight == tctx_inflight(tctx))
8865 finish_wait(&tctx->wait, &wait);
8867 atomic_dec(&tctx->in_idle);
8868 io_sq_thread_unpark(sqd);
8872 * Find any io_uring fd that this task has registered or done IO on, and cancel
8875 void __io_uring_task_cancel(void)
8877 struct io_uring_task *tctx = current->io_uring;
8881 /* make sure overflow events are dropped */
8882 atomic_inc(&tctx->in_idle);
8884 /* trigger io_disable_sqo_submit() */
8887 unsigned long index;
8889 xa_for_each(&tctx->xa, index, file)
8890 io_uring_cancel_sqpoll(file->private_data);
8894 /* read completions before cancelations */
8895 inflight = tctx_inflight(tctx);
8898 __io_uring_files_cancel(NULL);
8900 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
8903 * If we've seen completions, retry without waiting. This
8904 * avoids a race where a completion comes in before we did
8905 * prepare_to_wait().
8907 if (inflight == tctx_inflight(tctx))
8909 finish_wait(&tctx->wait, &wait);
8912 atomic_dec(&tctx->in_idle);
8914 io_uring_clean_tctx(tctx);
8915 /* all current's requests should be gone, we can kill tctx */
8916 __io_uring_free(current);
8919 static void *io_uring_validate_mmap_request(struct file *file,
8920 loff_t pgoff, size_t sz)
8922 struct io_ring_ctx *ctx = file->private_data;
8923 loff_t offset = pgoff << PAGE_SHIFT;
8928 case IORING_OFF_SQ_RING:
8929 case IORING_OFF_CQ_RING:
8932 case IORING_OFF_SQES:
8936 return ERR_PTR(-EINVAL);
8939 page = virt_to_head_page(ptr);
8940 if (sz > page_size(page))
8941 return ERR_PTR(-EINVAL);
8948 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
8950 size_t sz = vma->vm_end - vma->vm_start;
8954 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
8956 return PTR_ERR(ptr);
8958 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
8959 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
8962 #else /* !CONFIG_MMU */
8964 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
8966 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
8969 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
8971 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
8974 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
8975 unsigned long addr, unsigned long len,
8976 unsigned long pgoff, unsigned long flags)
8980 ptr = io_uring_validate_mmap_request(file, pgoff, len);
8982 return PTR_ERR(ptr);
8984 return (unsigned long) ptr;
8987 #endif /* !CONFIG_MMU */
8989 static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
8995 if (!io_sqring_full(ctx))
8998 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9000 if (unlikely(ctx->sqo_dead)) {
9005 if (!io_sqring_full(ctx))
9009 } while (!signal_pending(current));
9011 finish_wait(&ctx->sqo_sq_wait, &wait);
9016 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9017 struct __kernel_timespec __user **ts,
9018 const sigset_t __user **sig)
9020 struct io_uring_getevents_arg arg;
9023 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9024 * is just a pointer to the sigset_t.
9026 if (!(flags & IORING_ENTER_EXT_ARG)) {
9027 *sig = (const sigset_t __user *) argp;
9033 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9034 * timespec and sigset_t pointers if good.
9036 if (*argsz != sizeof(arg))
9038 if (copy_from_user(&arg, argp, sizeof(arg)))
9040 *sig = u64_to_user_ptr(arg.sigmask);
9041 *argsz = arg.sigmask_sz;
9042 *ts = u64_to_user_ptr(arg.ts);
9046 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9047 u32, min_complete, u32, flags, const void __user *, argp,
9050 struct io_ring_ctx *ctx;
9057 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9058 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG))
9066 if (f.file->f_op != &io_uring_fops)
9070 ctx = f.file->private_data;
9071 if (!percpu_ref_tryget(&ctx->refs))
9075 if (ctx->flags & IORING_SETUP_R_DISABLED)
9079 * For SQ polling, the thread will do all submissions and completions.
9080 * Just return the requested submit count, and wake the thread if
9084 if (ctx->flags & IORING_SETUP_SQPOLL) {
9085 io_cqring_overflow_flush(ctx, false, NULL, NULL);
9087 if (unlikely(ctx->sqo_exec)) {
9088 ret = io_sq_thread_fork(ctx->sq_data, ctx);
9094 if (unlikely(ctx->sqo_dead))
9096 if (flags & IORING_ENTER_SQ_WAKEUP)
9097 wake_up(&ctx->sq_data->wait);
9098 if (flags & IORING_ENTER_SQ_WAIT) {
9099 ret = io_sqpoll_wait_sq(ctx);
9103 submitted = to_submit;
9104 } else if (to_submit) {
9105 ret = io_uring_add_task_file(ctx, f.file);
9108 mutex_lock(&ctx->uring_lock);
9109 submitted = io_submit_sqes(ctx, to_submit);
9110 mutex_unlock(&ctx->uring_lock);
9112 if (submitted != to_submit)
9115 if (flags & IORING_ENTER_GETEVENTS) {
9116 const sigset_t __user *sig;
9117 struct __kernel_timespec __user *ts;
9119 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
9123 min_complete = min(min_complete, ctx->cq_entries);
9126 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9127 * space applications don't need to do io completion events
9128 * polling again, they can rely on io_sq_thread to do polling
9129 * work, which can reduce cpu usage and uring_lock contention.
9131 if (ctx->flags & IORING_SETUP_IOPOLL &&
9132 !(ctx->flags & IORING_SETUP_SQPOLL)) {
9133 ret = io_iopoll_check(ctx, min_complete);
9135 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
9140 percpu_ref_put(&ctx->refs);
9143 return submitted ? submitted : ret;
9146 #ifdef CONFIG_PROC_FS
9147 static int io_uring_show_cred(int id, void *p, void *data)
9149 const struct cred *cred = p;
9150 struct seq_file *m = data;
9151 struct user_namespace *uns = seq_user_ns(m);
9152 struct group_info *gi;
9157 seq_printf(m, "%5d\n", id);
9158 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
9159 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
9160 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
9161 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
9162 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
9163 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
9164 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
9165 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
9166 seq_puts(m, "\n\tGroups:\t");
9167 gi = cred->group_info;
9168 for (g = 0; g < gi->ngroups; g++) {
9169 seq_put_decimal_ull(m, g ? " " : "",
9170 from_kgid_munged(uns, gi->gid[g]));
9172 seq_puts(m, "\n\tCapEff:\t");
9173 cap = cred->cap_effective;
9174 CAP_FOR_EACH_U32(__capi)
9175 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
9180 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
9182 struct io_sq_data *sq = NULL;
9187 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9188 * since fdinfo case grabs it in the opposite direction of normal use
9189 * cases. If we fail to get the lock, we just don't iterate any
9190 * structures that could be going away outside the io_uring mutex.
9192 has_lock = mutex_trylock(&ctx->uring_lock);
9194 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) {
9200 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9201 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9202 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9203 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9204 struct file *f = *io_fixed_file_slot(ctx->file_data, i);
9207 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9209 seq_printf(m, "%5u: <none>\n", i);
9211 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9212 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9213 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
9215 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
9216 (unsigned int) buf->len);
9218 if (has_lock && !idr_is_empty(&ctx->personality_idr)) {
9219 seq_printf(m, "Personalities:\n");
9220 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
9222 seq_printf(m, "PollList:\n");
9223 spin_lock_irq(&ctx->completion_lock);
9224 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9225 struct hlist_head *list = &ctx->cancel_hash[i];
9226 struct io_kiocb *req;
9228 hlist_for_each_entry(req, list, hash_node)
9229 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9230 req->task->task_works != NULL);
9232 spin_unlock_irq(&ctx->completion_lock);
9234 mutex_unlock(&ctx->uring_lock);
9237 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9239 struct io_ring_ctx *ctx = f->private_data;
9241 if (percpu_ref_tryget(&ctx->refs)) {
9242 __io_uring_show_fdinfo(ctx, m);
9243 percpu_ref_put(&ctx->refs);
9248 static const struct file_operations io_uring_fops = {
9249 .release = io_uring_release,
9250 .mmap = io_uring_mmap,
9252 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9253 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9255 .poll = io_uring_poll,
9256 .fasync = io_uring_fasync,
9257 #ifdef CONFIG_PROC_FS
9258 .show_fdinfo = io_uring_show_fdinfo,
9262 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9263 struct io_uring_params *p)
9265 struct io_rings *rings;
9266 size_t size, sq_array_offset;
9268 /* make sure these are sane, as we already accounted them */
9269 ctx->sq_entries = p->sq_entries;
9270 ctx->cq_entries = p->cq_entries;
9272 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9273 if (size == SIZE_MAX)
9276 rings = io_mem_alloc(size);
9281 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9282 rings->sq_ring_mask = p->sq_entries - 1;
9283 rings->cq_ring_mask = p->cq_entries - 1;
9284 rings->sq_ring_entries = p->sq_entries;
9285 rings->cq_ring_entries = p->cq_entries;
9286 ctx->sq_mask = rings->sq_ring_mask;
9287 ctx->cq_mask = rings->cq_ring_mask;
9289 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9290 if (size == SIZE_MAX) {
9291 io_mem_free(ctx->rings);
9296 ctx->sq_sqes = io_mem_alloc(size);
9297 if (!ctx->sq_sqes) {
9298 io_mem_free(ctx->rings);
9306 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
9310 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9314 ret = io_uring_add_task_file(ctx, file);
9319 fd_install(fd, file);
9324 * Allocate an anonymous fd, this is what constitutes the application
9325 * visible backing of an io_uring instance. The application mmaps this
9326 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9327 * we have to tie this fd to a socket for file garbage collection purposes.
9329 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
9332 #if defined(CONFIG_UNIX)
9335 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9338 return ERR_PTR(ret);
9341 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9342 O_RDWR | O_CLOEXEC);
9343 #if defined(CONFIG_UNIX)
9345 sock_release(ctx->ring_sock);
9346 ctx->ring_sock = NULL;
9348 ctx->ring_sock->file = file;
9354 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9355 struct io_uring_params __user *params)
9357 struct io_ring_ctx *ctx;
9363 if (entries > IORING_MAX_ENTRIES) {
9364 if (!(p->flags & IORING_SETUP_CLAMP))
9366 entries = IORING_MAX_ENTRIES;
9370 * Use twice as many entries for the CQ ring. It's possible for the
9371 * application to drive a higher depth than the size of the SQ ring,
9372 * since the sqes are only used at submission time. This allows for
9373 * some flexibility in overcommitting a bit. If the application has
9374 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9375 * of CQ ring entries manually.
9377 p->sq_entries = roundup_pow_of_two(entries);
9378 if (p->flags & IORING_SETUP_CQSIZE) {
9380 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9381 * to a power-of-two, if it isn't already. We do NOT impose
9382 * any cq vs sq ring sizing.
9386 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9387 if (!(p->flags & IORING_SETUP_CLAMP))
9389 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9391 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9392 if (p->cq_entries < p->sq_entries)
9395 p->cq_entries = 2 * p->sq_entries;
9398 ctx = io_ring_ctx_alloc(p);
9401 ctx->compat = in_compat_syscall();
9402 if (!capable(CAP_IPC_LOCK))
9403 ctx->user = get_uid(current_user());
9404 ctx->sqo_task = current;
9407 * This is just grabbed for accounting purposes. When a process exits,
9408 * the mm is exited and dropped before the files, hence we need to hang
9409 * on to this mm purely for the purposes of being able to unaccount
9410 * memory (locked/pinned vm). It's not used for anything else.
9412 mmgrab(current->mm);
9413 ctx->mm_account = current->mm;
9415 ret = io_allocate_scq_urings(ctx, p);
9419 ret = io_sq_offload_create(ctx, p);
9423 if (!(p->flags & IORING_SETUP_R_DISABLED))
9424 io_sq_offload_start(ctx);
9426 memset(&p->sq_off, 0, sizeof(p->sq_off));
9427 p->sq_off.head = offsetof(struct io_rings, sq.head);
9428 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9429 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9430 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9431 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9432 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9433 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9435 memset(&p->cq_off, 0, sizeof(p->cq_off));
9436 p->cq_off.head = offsetof(struct io_rings, cq.head);
9437 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9438 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9439 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9440 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9441 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9442 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9444 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9445 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9446 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9447 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
9448 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS;
9450 if (copy_to_user(params, p, sizeof(*p))) {
9455 file = io_uring_get_file(ctx);
9457 ret = PTR_ERR(file);
9462 * Install ring fd as the very last thing, so we don't risk someone
9463 * having closed it before we finish setup
9465 ret = io_uring_install_fd(ctx, file);
9467 io_disable_sqo_submit(ctx);
9468 /* fput will clean it up */
9473 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9476 io_disable_sqo_submit(ctx);
9477 io_ring_ctx_wait_and_kill(ctx);
9482 * Sets up an aio uring context, and returns the fd. Applications asks for a
9483 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9484 * params structure passed in.
9486 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9488 struct io_uring_params p;
9491 if (copy_from_user(&p, params, sizeof(p)))
9493 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9498 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9499 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9500 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9501 IORING_SETUP_R_DISABLED))
9504 return io_uring_create(entries, &p, params);
9507 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9508 struct io_uring_params __user *, params)
9510 return io_uring_setup(entries, params);
9513 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9515 struct io_uring_probe *p;
9519 size = struct_size(p, ops, nr_args);
9520 if (size == SIZE_MAX)
9522 p = kzalloc(size, GFP_KERNEL);
9527 if (copy_from_user(p, arg, size))
9530 if (memchr_inv(p, 0, size))
9533 p->last_op = IORING_OP_LAST - 1;
9534 if (nr_args > IORING_OP_LAST)
9535 nr_args = IORING_OP_LAST;
9537 for (i = 0; i < nr_args; i++) {
9539 if (!io_op_defs[i].not_supported)
9540 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9545 if (copy_to_user(arg, p, size))
9552 static int io_register_personality(struct io_ring_ctx *ctx)
9554 const struct cred *creds;
9557 creds = get_current_cred();
9559 ret = idr_alloc_cyclic(&ctx->personality_idr, (void *) creds, 1,
9560 USHRT_MAX, GFP_KERNEL);
9566 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9567 unsigned int nr_args)
9569 struct io_uring_restriction *res;
9573 /* Restrictions allowed only if rings started disabled */
9574 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9577 /* We allow only a single restrictions registration */
9578 if (ctx->restrictions.registered)
9581 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9584 size = array_size(nr_args, sizeof(*res));
9585 if (size == SIZE_MAX)
9588 res = memdup_user(arg, size);
9590 return PTR_ERR(res);
9594 for (i = 0; i < nr_args; i++) {
9595 switch (res[i].opcode) {
9596 case IORING_RESTRICTION_REGISTER_OP:
9597 if (res[i].register_op >= IORING_REGISTER_LAST) {
9602 __set_bit(res[i].register_op,
9603 ctx->restrictions.register_op);
9605 case IORING_RESTRICTION_SQE_OP:
9606 if (res[i].sqe_op >= IORING_OP_LAST) {
9611 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9613 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9614 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9616 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9617 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9626 /* Reset all restrictions if an error happened */
9628 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9630 ctx->restrictions.registered = true;
9636 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9638 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9641 if (ctx->restrictions.registered)
9642 ctx->restricted = 1;
9644 ctx->flags &= ~IORING_SETUP_R_DISABLED;
9646 io_sq_offload_start(ctx);
9651 static bool io_register_op_must_quiesce(int op)
9654 case IORING_UNREGISTER_FILES:
9655 case IORING_REGISTER_FILES_UPDATE:
9656 case IORING_REGISTER_PROBE:
9657 case IORING_REGISTER_PERSONALITY:
9658 case IORING_UNREGISTER_PERSONALITY:
9665 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
9666 void __user *arg, unsigned nr_args)
9667 __releases(ctx->uring_lock)
9668 __acquires(ctx->uring_lock)
9673 * We're inside the ring mutex, if the ref is already dying, then
9674 * someone else killed the ctx or is already going through
9675 * io_uring_register().
9677 if (percpu_ref_is_dying(&ctx->refs))
9680 if (io_register_op_must_quiesce(opcode)) {
9681 percpu_ref_kill(&ctx->refs);
9684 * Drop uring mutex before waiting for references to exit. If
9685 * another thread is currently inside io_uring_enter() it might
9686 * need to grab the uring_lock to make progress. If we hold it
9687 * here across the drain wait, then we can deadlock. It's safe
9688 * to drop the mutex here, since no new references will come in
9689 * after we've killed the percpu ref.
9691 mutex_unlock(&ctx->uring_lock);
9693 ret = wait_for_completion_interruptible(&ctx->ref_comp);
9696 ret = io_run_task_work_sig();
9701 mutex_lock(&ctx->uring_lock);
9704 percpu_ref_resurrect(&ctx->refs);
9709 if (ctx->restricted) {
9710 if (opcode >= IORING_REGISTER_LAST) {
9715 if (!test_bit(opcode, ctx->restrictions.register_op)) {
9722 case IORING_REGISTER_BUFFERS:
9723 ret = io_sqe_buffers_register(ctx, arg, nr_args);
9725 case IORING_UNREGISTER_BUFFERS:
9729 ret = io_sqe_buffers_unregister(ctx);
9731 case IORING_REGISTER_FILES:
9732 ret = io_sqe_files_register(ctx, arg, nr_args);
9734 case IORING_UNREGISTER_FILES:
9738 ret = io_sqe_files_unregister(ctx);
9740 case IORING_REGISTER_FILES_UPDATE:
9741 ret = io_sqe_files_update(ctx, arg, nr_args);
9743 case IORING_REGISTER_EVENTFD:
9744 case IORING_REGISTER_EVENTFD_ASYNC:
9748 ret = io_eventfd_register(ctx, arg);
9751 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
9752 ctx->eventfd_async = 1;
9754 ctx->eventfd_async = 0;
9756 case IORING_UNREGISTER_EVENTFD:
9760 ret = io_eventfd_unregister(ctx);
9762 case IORING_REGISTER_PROBE:
9764 if (!arg || nr_args > 256)
9766 ret = io_probe(ctx, arg, nr_args);
9768 case IORING_REGISTER_PERSONALITY:
9772 ret = io_register_personality(ctx);
9774 case IORING_UNREGISTER_PERSONALITY:
9778 ret = io_unregister_personality(ctx, nr_args);
9780 case IORING_REGISTER_ENABLE_RINGS:
9784 ret = io_register_enable_rings(ctx);
9786 case IORING_REGISTER_RESTRICTIONS:
9787 ret = io_register_restrictions(ctx, arg, nr_args);
9795 if (io_register_op_must_quiesce(opcode)) {
9796 /* bring the ctx back to life */
9797 percpu_ref_reinit(&ctx->refs);
9799 reinit_completion(&ctx->ref_comp);
9804 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
9805 void __user *, arg, unsigned int, nr_args)
9807 struct io_ring_ctx *ctx;
9816 if (f.file->f_op != &io_uring_fops)
9819 ctx = f.file->private_data;
9823 mutex_lock(&ctx->uring_lock);
9824 ret = __io_uring_register(ctx, opcode, arg, nr_args);
9825 mutex_unlock(&ctx->uring_lock);
9826 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
9827 ctx->cq_ev_fd != NULL, ret);
9833 static int __init io_uring_init(void)
9835 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
9836 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
9837 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
9840 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
9841 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
9842 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
9843 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
9844 BUILD_BUG_SQE_ELEM(1, __u8, flags);
9845 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
9846 BUILD_BUG_SQE_ELEM(4, __s32, fd);
9847 BUILD_BUG_SQE_ELEM(8, __u64, off);
9848 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
9849 BUILD_BUG_SQE_ELEM(16, __u64, addr);
9850 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
9851 BUILD_BUG_SQE_ELEM(24, __u32, len);
9852 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
9853 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
9854 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
9855 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
9856 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
9857 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
9858 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
9859 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
9860 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
9861 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
9862 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
9863 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
9864 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
9865 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
9866 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
9867 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
9868 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
9869 BUILD_BUG_SQE_ELEM(42, __u16, personality);
9870 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
9872 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
9873 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
9874 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
9878 __initcall(io_uring_init);