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/splice.h>
78 #include <linux/task_work.h>
79 #include <linux/pagemap.h>
80 #include <linux/io_uring.h>
81 #include <linux/freezer.h>
83 #define CREATE_TRACE_POINTS
84 #include <trace/events/io_uring.h>
86 #include <uapi/linux/io_uring.h>
91 #define IORING_MAX_ENTRIES 32768
92 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
95 * Shift of 9 is 512 entries, or exactly one page on 64-bit archs
97 #define IORING_FILE_TABLE_SHIFT 9
98 #define IORING_MAX_FILES_TABLE (1U << IORING_FILE_TABLE_SHIFT)
99 #define IORING_FILE_TABLE_MASK (IORING_MAX_FILES_TABLE - 1)
100 #define IORING_MAX_FIXED_FILES (64 * IORING_MAX_FILES_TABLE)
101 #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
102 IORING_REGISTER_LAST + IORING_OP_LAST)
104 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
105 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
109 u32 head ____cacheline_aligned_in_smp;
110 u32 tail ____cacheline_aligned_in_smp;
114 * This data is shared with the application through the mmap at offsets
115 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
117 * The offsets to the member fields are published through struct
118 * io_sqring_offsets when calling io_uring_setup.
122 * Head and tail offsets into the ring; the offsets need to be
123 * masked to get valid indices.
125 * The kernel controls head of the sq ring and the tail of the cq ring,
126 * and the application controls tail of the sq ring and the head of the
129 struct io_uring sq, cq;
131 * Bitmasks to apply to head and tail offsets (constant, equals
134 u32 sq_ring_mask, cq_ring_mask;
135 /* Ring sizes (constant, power of 2) */
136 u32 sq_ring_entries, cq_ring_entries;
138 * Number of invalid entries dropped by the kernel due to
139 * invalid index stored in array
141 * Written by the kernel, shouldn't be modified by the
142 * application (i.e. get number of "new events" by comparing to
145 * After a new SQ head value was read by the application this
146 * counter includes all submissions that were dropped reaching
147 * the new SQ head (and possibly more).
153 * Written by the kernel, shouldn't be modified by the
156 * The application needs a full memory barrier before checking
157 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
163 * Written by the application, shouldn't be modified by the
168 * Number of completion events lost because the queue was full;
169 * this should be avoided by the application by making sure
170 * there are not more requests pending than there is space in
171 * the completion queue.
173 * Written by the kernel, shouldn't be modified by the
174 * application (i.e. get number of "new events" by comparing to
177 * As completion events come in out of order this counter is not
178 * ordered with any other data.
182 * Ring buffer of completion events.
184 * The kernel writes completion events fresh every time they are
185 * produced, so the application is allowed to modify pending
188 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
191 enum io_uring_cmd_flags {
192 IO_URING_F_NONBLOCK = 1,
193 IO_URING_F_COMPLETE_DEFER = 2,
196 struct io_mapped_ubuf {
199 struct bio_vec *bvec;
200 unsigned int nr_bvecs;
201 unsigned long acct_pages;
207 struct list_head list;
214 struct fixed_rsrc_table {
218 struct fixed_rsrc_ref_node {
219 struct percpu_ref refs;
220 struct list_head node;
221 struct list_head rsrc_list;
222 struct fixed_rsrc_data *rsrc_data;
223 void (*rsrc_put)(struct io_ring_ctx *ctx,
224 struct io_rsrc_put *prsrc);
225 struct llist_node llist;
229 struct fixed_rsrc_data {
230 struct fixed_rsrc_table *table;
231 struct io_ring_ctx *ctx;
233 struct fixed_rsrc_ref_node *node;
234 struct percpu_ref refs;
235 struct completion done;
240 struct list_head list;
246 struct io_restriction {
247 DECLARE_BITMAP(register_op, IORING_REGISTER_LAST);
248 DECLARE_BITMAP(sqe_op, IORING_OP_LAST);
249 u8 sqe_flags_allowed;
250 u8 sqe_flags_required;
255 IO_SQ_THREAD_SHOULD_STOP = 0,
256 IO_SQ_THREAD_SHOULD_PARK,
263 /* ctx's that are using this sqd */
264 struct list_head ctx_list;
265 struct list_head ctx_new_list;
266 struct mutex ctx_lock;
268 struct task_struct *thread;
269 struct wait_queue_head wait;
271 unsigned sq_thread_idle;
276 struct completion startup;
277 struct completion parked;
278 struct completion exited;
281 #define IO_IOPOLL_BATCH 8
282 #define IO_COMPL_BATCH 32
283 #define IO_REQ_CACHE_SIZE 32
284 #define IO_REQ_ALLOC_BATCH 8
286 struct io_comp_state {
287 struct io_kiocb *reqs[IO_COMPL_BATCH];
289 unsigned int locked_free_nr;
290 /* inline/task_work completion list, under ->uring_lock */
291 struct list_head free_list;
292 /* IRQ completion list, under ->completion_lock */
293 struct list_head locked_free_list;
296 struct io_submit_link {
297 struct io_kiocb *head;
298 struct io_kiocb *last;
301 struct io_submit_state {
302 struct blk_plug plug;
303 struct io_submit_link link;
306 * io_kiocb alloc cache
308 void *reqs[IO_REQ_CACHE_SIZE];
309 unsigned int free_reqs;
314 * Batch completion logic
316 struct io_comp_state comp;
319 * File reference cache
323 unsigned int file_refs;
324 unsigned int ios_left;
329 struct percpu_ref refs;
330 } ____cacheline_aligned_in_smp;
334 unsigned int compat: 1;
335 unsigned int cq_overflow_flushed: 1;
336 unsigned int drain_next: 1;
337 unsigned int eventfd_async: 1;
338 unsigned int restricted: 1;
339 unsigned int sqo_exec: 1;
342 * Ring buffer of indices into array of io_uring_sqe, which is
343 * mmapped by the application using the IORING_OFF_SQES offset.
345 * This indirection could e.g. be used to assign fixed
346 * io_uring_sqe entries to operations and only submit them to
347 * the queue when needed.
349 * The kernel modifies neither the indices array nor the entries
353 unsigned cached_sq_head;
356 unsigned sq_thread_idle;
357 unsigned cached_sq_dropped;
358 unsigned cached_cq_overflow;
359 unsigned long sq_check_overflow;
361 /* hashed buffered write serialization */
362 struct io_wq_hash *hash_map;
364 struct list_head defer_list;
365 struct list_head timeout_list;
366 struct list_head cq_overflow_list;
368 struct io_uring_sqe *sq_sqes;
369 } ____cacheline_aligned_in_smp;
372 struct mutex uring_lock;
373 wait_queue_head_t wait;
374 } ____cacheline_aligned_in_smp;
376 struct io_submit_state submit_state;
378 struct io_rings *rings;
380 /* Only used for accounting purposes */
381 struct mm_struct *mm_account;
383 struct io_sq_data *sq_data; /* if using sq thread polling */
385 struct wait_queue_head sqo_sq_wait;
386 struct list_head sqd_list;
389 * If used, fixed file set. Writers must ensure that ->refs is dead,
390 * readers must ensure that ->refs is alive as long as the file* is
391 * used. Only updated through io_uring_register(2).
393 struct fixed_rsrc_data *file_data;
394 unsigned nr_user_files;
396 /* if used, fixed mapped user buffers */
397 unsigned nr_user_bufs;
398 struct io_mapped_ubuf *user_bufs;
400 struct user_struct *user;
402 struct completion ref_comp;
403 struct completion sq_thread_comp;
405 #if defined(CONFIG_UNIX)
406 struct socket *ring_sock;
409 struct idr io_buffer_idr;
411 struct idr personality_idr;
414 unsigned cached_cq_tail;
417 atomic_t cq_timeouts;
418 unsigned cq_last_tm_flush;
419 unsigned long cq_check_overflow;
420 struct wait_queue_head cq_wait;
421 struct fasync_struct *cq_fasync;
422 struct eventfd_ctx *cq_ev_fd;
423 } ____cacheline_aligned_in_smp;
426 spinlock_t completion_lock;
429 * ->iopoll_list is protected by the ctx->uring_lock for
430 * io_uring instances that don't use IORING_SETUP_SQPOLL.
431 * For SQPOLL, only the single threaded io_sq_thread() will
432 * manipulate the list, hence no extra locking is needed there.
434 struct list_head iopoll_list;
435 struct hlist_head *cancel_hash;
436 unsigned cancel_hash_bits;
437 bool poll_multi_file;
439 spinlock_t inflight_lock;
440 struct list_head inflight_list;
441 } ____cacheline_aligned_in_smp;
443 struct delayed_work rsrc_put_work;
444 struct llist_head rsrc_put_llist;
445 struct list_head rsrc_ref_list;
446 spinlock_t rsrc_ref_lock;
448 struct io_restriction restrictions;
451 struct callback_head *exit_task_work;
453 struct wait_queue_head hash_wait;
455 /* Keep this last, we don't need it for the fast path */
456 struct work_struct exit_work;
460 * First field must be the file pointer in all the
461 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
463 struct io_poll_iocb {
465 struct wait_queue_head *head;
469 struct wait_queue_entry wait;
472 struct io_poll_remove {
482 struct io_timeout_data {
483 struct io_kiocb *req;
484 struct hrtimer timer;
485 struct timespec64 ts;
486 enum hrtimer_mode mode;
491 struct sockaddr __user *addr;
492 int __user *addr_len;
494 unsigned long nofile;
514 struct list_head list;
515 /* head of the link, used by linked timeouts only */
516 struct io_kiocb *head;
519 struct io_timeout_rem {
524 struct timespec64 ts;
529 /* NOTE: kiocb has the file as the first member, so don't do it here */
537 struct sockaddr __user *addr;
544 struct user_msghdr __user *umsg;
550 struct io_buffer *kbuf;
556 struct filename *filename;
558 unsigned long nofile;
561 struct io_rsrc_update {
587 struct epoll_event event;
591 struct file *file_out;
592 struct file *file_in;
599 struct io_provide_buf {
613 const char __user *filename;
614 struct statx __user *buffer;
626 struct filename *oldpath;
627 struct filename *newpath;
635 struct filename *filename;
638 struct io_completion {
640 struct list_head list;
644 struct io_async_connect {
645 struct sockaddr_storage address;
648 struct io_async_msghdr {
649 struct iovec fast_iov[UIO_FASTIOV];
650 /* points to an allocated iov, if NULL we use fast_iov instead */
651 struct iovec *free_iov;
652 struct sockaddr __user *uaddr;
654 struct sockaddr_storage addr;
658 struct iovec fast_iov[UIO_FASTIOV];
659 const struct iovec *free_iovec;
660 struct iov_iter iter;
662 struct wait_page_queue wpq;
666 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
667 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
668 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
669 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
670 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
671 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
677 REQ_F_LINK_TIMEOUT_BIT,
679 REQ_F_NEED_CLEANUP_BIT,
681 REQ_F_BUFFER_SELECTED_BIT,
682 REQ_F_NO_FILE_TABLE_BIT,
683 REQ_F_LTIMEOUT_ACTIVE_BIT,
684 REQ_F_COMPLETE_INLINE_BIT,
686 /* not a real bit, just to check we're not overflowing the space */
692 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
693 /* drain existing IO first */
694 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
696 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
697 /* doesn't sever on completion < 0 */
698 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
700 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
701 /* IOSQE_BUFFER_SELECT */
702 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
704 /* fail rest of links */
705 REQ_F_FAIL_LINK = BIT(REQ_F_FAIL_LINK_BIT),
706 /* on inflight list, should be cancelled and waited on exit reliably */
707 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
708 /* read/write uses file position */
709 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
710 /* must not punt to workers */
711 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
712 /* has or had linked timeout */
713 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
715 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
717 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
718 /* already went through poll handler */
719 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
720 /* buffer already selected */
721 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
722 /* doesn't need file table for this request */
723 REQ_F_NO_FILE_TABLE = BIT(REQ_F_NO_FILE_TABLE_BIT),
724 /* linked timeout is active, i.e. prepared by link's head */
725 REQ_F_LTIMEOUT_ACTIVE = BIT(REQ_F_LTIMEOUT_ACTIVE_BIT),
726 /* completion is deferred through io_comp_state */
727 REQ_F_COMPLETE_INLINE = BIT(REQ_F_COMPLETE_INLINE_BIT),
731 struct io_poll_iocb poll;
732 struct io_poll_iocb *double_poll;
735 struct io_task_work {
736 struct io_wq_work_node node;
737 task_work_func_t func;
741 * NOTE! Each of the iocb union members has the file pointer
742 * as the first entry in their struct definition. So you can
743 * access the file pointer through any of the sub-structs,
744 * or directly as just 'ki_filp' in this struct.
750 struct io_poll_iocb poll;
751 struct io_poll_remove poll_remove;
752 struct io_accept accept;
754 struct io_cancel cancel;
755 struct io_timeout timeout;
756 struct io_timeout_rem timeout_rem;
757 struct io_connect connect;
758 struct io_sr_msg sr_msg;
760 struct io_close close;
761 struct io_rsrc_update rsrc_update;
762 struct io_fadvise fadvise;
763 struct io_madvise madvise;
764 struct io_epoll epoll;
765 struct io_splice splice;
766 struct io_provide_buf pbuf;
767 struct io_statx statx;
768 struct io_shutdown shutdown;
769 struct io_rename rename;
770 struct io_unlink unlink;
771 /* use only after cleaning per-op data, see io_clean_op() */
772 struct io_completion compl;
775 /* opcode allocated if it needs to store data for async defer */
778 /* polled IO has completed */
784 struct io_ring_ctx *ctx;
787 struct task_struct *task;
790 struct io_kiocb *link;
791 struct percpu_ref *fixed_rsrc_refs;
794 * 1. used with ctx->iopoll_list with reads/writes
795 * 2. to track reqs with ->files (see io_op_def::file_table)
797 struct list_head inflight_entry;
799 struct io_task_work io_task_work;
800 struct callback_head task_work;
802 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
803 struct hlist_node hash_node;
804 struct async_poll *apoll;
805 struct io_wq_work work;
808 struct io_defer_entry {
809 struct list_head list;
810 struct io_kiocb *req;
815 /* needs req->file assigned */
816 unsigned needs_file : 1;
817 /* hash wq insertion if file is a regular file */
818 unsigned hash_reg_file : 1;
819 /* unbound wq insertion if file is a non-regular file */
820 unsigned unbound_nonreg_file : 1;
821 /* opcode is not supported by this kernel */
822 unsigned not_supported : 1;
823 /* set if opcode supports polled "wait" */
825 unsigned pollout : 1;
826 /* op supports buffer selection */
827 unsigned buffer_select : 1;
828 /* must always have async data allocated */
829 unsigned needs_async_data : 1;
830 /* should block plug */
832 /* size of async data needed, if any */
833 unsigned short async_size;
836 static const struct io_op_def io_op_defs[] = {
837 [IORING_OP_NOP] = {},
838 [IORING_OP_READV] = {
840 .unbound_nonreg_file = 1,
843 .needs_async_data = 1,
845 .async_size = sizeof(struct io_async_rw),
847 [IORING_OP_WRITEV] = {
850 .unbound_nonreg_file = 1,
852 .needs_async_data = 1,
854 .async_size = sizeof(struct io_async_rw),
856 [IORING_OP_FSYNC] = {
859 [IORING_OP_READ_FIXED] = {
861 .unbound_nonreg_file = 1,
864 .async_size = sizeof(struct io_async_rw),
866 [IORING_OP_WRITE_FIXED] = {
869 .unbound_nonreg_file = 1,
872 .async_size = sizeof(struct io_async_rw),
874 [IORING_OP_POLL_ADD] = {
876 .unbound_nonreg_file = 1,
878 [IORING_OP_POLL_REMOVE] = {},
879 [IORING_OP_SYNC_FILE_RANGE] = {
882 [IORING_OP_SENDMSG] = {
884 .unbound_nonreg_file = 1,
886 .needs_async_data = 1,
887 .async_size = sizeof(struct io_async_msghdr),
889 [IORING_OP_RECVMSG] = {
891 .unbound_nonreg_file = 1,
894 .needs_async_data = 1,
895 .async_size = sizeof(struct io_async_msghdr),
897 [IORING_OP_TIMEOUT] = {
898 .needs_async_data = 1,
899 .async_size = sizeof(struct io_timeout_data),
901 [IORING_OP_TIMEOUT_REMOVE] = {
902 /* used by timeout updates' prep() */
904 [IORING_OP_ACCEPT] = {
906 .unbound_nonreg_file = 1,
909 [IORING_OP_ASYNC_CANCEL] = {},
910 [IORING_OP_LINK_TIMEOUT] = {
911 .needs_async_data = 1,
912 .async_size = sizeof(struct io_timeout_data),
914 [IORING_OP_CONNECT] = {
916 .unbound_nonreg_file = 1,
918 .needs_async_data = 1,
919 .async_size = sizeof(struct io_async_connect),
921 [IORING_OP_FALLOCATE] = {
924 [IORING_OP_OPENAT] = {},
925 [IORING_OP_CLOSE] = {},
926 [IORING_OP_FILES_UPDATE] = {},
927 [IORING_OP_STATX] = {},
930 .unbound_nonreg_file = 1,
934 .async_size = sizeof(struct io_async_rw),
936 [IORING_OP_WRITE] = {
938 .unbound_nonreg_file = 1,
941 .async_size = sizeof(struct io_async_rw),
943 [IORING_OP_FADVISE] = {
946 [IORING_OP_MADVISE] = {},
949 .unbound_nonreg_file = 1,
954 .unbound_nonreg_file = 1,
958 [IORING_OP_OPENAT2] = {
960 [IORING_OP_EPOLL_CTL] = {
961 .unbound_nonreg_file = 1,
963 [IORING_OP_SPLICE] = {
966 .unbound_nonreg_file = 1,
968 [IORING_OP_PROVIDE_BUFFERS] = {},
969 [IORING_OP_REMOVE_BUFFERS] = {},
973 .unbound_nonreg_file = 1,
975 [IORING_OP_SHUTDOWN] = {
978 [IORING_OP_RENAMEAT] = {},
979 [IORING_OP_UNLINKAT] = {},
982 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
983 struct task_struct *task,
984 struct files_struct *files);
985 static void io_uring_cancel_sqpoll(struct io_ring_ctx *ctx);
986 static void destroy_fixed_rsrc_ref_node(struct fixed_rsrc_ref_node *ref_node);
987 static struct fixed_rsrc_ref_node *alloc_fixed_rsrc_ref_node(
988 struct io_ring_ctx *ctx);
989 static void io_ring_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc);
991 static bool io_rw_reissue(struct io_kiocb *req);
992 static void io_cqring_fill_event(struct io_kiocb *req, long res);
993 static void io_put_req(struct io_kiocb *req);
994 static void io_put_req_deferred(struct io_kiocb *req, int nr);
995 static void io_double_put_req(struct io_kiocb *req);
996 static void io_dismantle_req(struct io_kiocb *req);
997 static void io_put_task(struct task_struct *task, int nr);
998 static void io_queue_next(struct io_kiocb *req);
999 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
1000 static void __io_queue_linked_timeout(struct io_kiocb *req);
1001 static void io_queue_linked_timeout(struct io_kiocb *req);
1002 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
1003 struct io_uring_rsrc_update *ip,
1005 static void __io_clean_op(struct io_kiocb *req);
1006 static struct file *io_file_get(struct io_submit_state *state,
1007 struct io_kiocb *req, int fd, bool fixed);
1008 static void __io_queue_sqe(struct io_kiocb *req);
1009 static void io_rsrc_put_work(struct work_struct *work);
1011 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
1012 struct iov_iter *iter, bool needs_lock);
1013 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
1014 const struct iovec *fast_iov,
1015 struct iov_iter *iter, bool force);
1016 static void io_req_task_queue(struct io_kiocb *req);
1017 static void io_submit_flush_completions(struct io_comp_state *cs,
1018 struct io_ring_ctx *ctx);
1020 static struct kmem_cache *req_cachep;
1022 static const struct file_operations io_uring_fops;
1024 struct sock *io_uring_get_socket(struct file *file)
1026 #if defined(CONFIG_UNIX)
1027 if (file->f_op == &io_uring_fops) {
1028 struct io_ring_ctx *ctx = file->private_data;
1030 return ctx->ring_sock->sk;
1035 EXPORT_SYMBOL(io_uring_get_socket);
1037 #define io_for_each_link(pos, head) \
1038 for (pos = (head); pos; pos = pos->link)
1040 static inline void io_clean_op(struct io_kiocb *req)
1042 if (req->flags & (REQ_F_NEED_CLEANUP | REQ_F_BUFFER_SELECTED))
1046 static inline void io_set_resource_node(struct io_kiocb *req)
1048 struct io_ring_ctx *ctx = req->ctx;
1050 if (!req->fixed_rsrc_refs) {
1051 req->fixed_rsrc_refs = &ctx->file_data->node->refs;
1052 percpu_ref_get(req->fixed_rsrc_refs);
1056 static bool io_match_task(struct io_kiocb *head,
1057 struct task_struct *task,
1058 struct files_struct *files)
1060 struct io_kiocb *req;
1062 if (task && head->task != task) {
1063 /* in terms of cancelation, always match if req task is dead */
1064 if (head->task->flags & PF_EXITING)
1071 io_for_each_link(req, head) {
1072 if (req->flags & REQ_F_INFLIGHT)
1074 if (req->task->files == files)
1080 static inline void req_set_fail_links(struct io_kiocb *req)
1082 if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
1083 req->flags |= REQ_F_FAIL_LINK;
1086 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1088 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1090 complete(&ctx->ref_comp);
1093 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1095 return !req->timeout.off;
1098 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1100 struct io_ring_ctx *ctx;
1103 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1108 * Use 5 bits less than the max cq entries, that should give us around
1109 * 32 entries per hash list if totally full and uniformly spread.
1111 hash_bits = ilog2(p->cq_entries);
1115 ctx->cancel_hash_bits = hash_bits;
1116 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1118 if (!ctx->cancel_hash)
1120 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1122 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1123 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1126 ctx->flags = p->flags;
1127 init_waitqueue_head(&ctx->sqo_sq_wait);
1128 INIT_LIST_HEAD(&ctx->sqd_list);
1129 init_waitqueue_head(&ctx->cq_wait);
1130 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1131 init_completion(&ctx->ref_comp);
1132 init_completion(&ctx->sq_thread_comp);
1133 idr_init(&ctx->io_buffer_idr);
1134 idr_init(&ctx->personality_idr);
1135 mutex_init(&ctx->uring_lock);
1136 init_waitqueue_head(&ctx->wait);
1137 spin_lock_init(&ctx->completion_lock);
1138 INIT_LIST_HEAD(&ctx->iopoll_list);
1139 INIT_LIST_HEAD(&ctx->defer_list);
1140 INIT_LIST_HEAD(&ctx->timeout_list);
1141 spin_lock_init(&ctx->inflight_lock);
1142 INIT_LIST_HEAD(&ctx->inflight_list);
1143 spin_lock_init(&ctx->rsrc_ref_lock);
1144 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
1145 INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
1146 init_llist_head(&ctx->rsrc_put_llist);
1147 INIT_LIST_HEAD(&ctx->submit_state.comp.free_list);
1148 INIT_LIST_HEAD(&ctx->submit_state.comp.locked_free_list);
1151 kfree(ctx->cancel_hash);
1156 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1158 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1159 struct io_ring_ctx *ctx = req->ctx;
1161 return seq != ctx->cached_cq_tail
1162 + READ_ONCE(ctx->cached_cq_overflow);
1168 static void io_req_track_inflight(struct io_kiocb *req)
1170 struct io_ring_ctx *ctx = req->ctx;
1172 if (!(req->flags & REQ_F_INFLIGHT)) {
1173 req->flags |= REQ_F_INFLIGHT;
1175 spin_lock_irq(&ctx->inflight_lock);
1176 list_add(&req->inflight_entry, &ctx->inflight_list);
1177 spin_unlock_irq(&ctx->inflight_lock);
1181 static void io_prep_async_work(struct io_kiocb *req)
1183 const struct io_op_def *def = &io_op_defs[req->opcode];
1184 struct io_ring_ctx *ctx = req->ctx;
1186 if (req->flags & REQ_F_FORCE_ASYNC)
1187 req->work.flags |= IO_WQ_WORK_CONCURRENT;
1189 if (req->flags & REQ_F_ISREG) {
1190 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1191 io_wq_hash_work(&req->work, file_inode(req->file));
1193 if (def->unbound_nonreg_file)
1194 req->work.flags |= IO_WQ_WORK_UNBOUND;
1198 static void io_prep_async_link(struct io_kiocb *req)
1200 struct io_kiocb *cur;
1202 io_for_each_link(cur, req)
1203 io_prep_async_work(cur);
1206 static void io_queue_async_work(struct io_kiocb *req)
1208 struct io_ring_ctx *ctx = req->ctx;
1209 struct io_kiocb *link = io_prep_linked_timeout(req);
1210 struct io_uring_task *tctx = req->task->io_uring;
1213 BUG_ON(!tctx->io_wq);
1215 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1216 &req->work, req->flags);
1217 /* init ->work of the whole link before punting */
1218 io_prep_async_link(req);
1219 io_wq_enqueue(tctx->io_wq, &req->work);
1221 io_queue_linked_timeout(link);
1224 static void io_kill_timeout(struct io_kiocb *req)
1226 struct io_timeout_data *io = req->async_data;
1229 ret = hrtimer_try_to_cancel(&io->timer);
1231 atomic_set(&req->ctx->cq_timeouts,
1232 atomic_read(&req->ctx->cq_timeouts) + 1);
1233 list_del_init(&req->timeout.list);
1234 io_cqring_fill_event(req, 0);
1235 io_put_req_deferred(req, 1);
1240 * Returns true if we found and killed one or more timeouts
1242 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
1243 struct files_struct *files)
1245 struct io_kiocb *req, *tmp;
1248 spin_lock_irq(&ctx->completion_lock);
1249 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
1250 if (io_match_task(req, tsk, files)) {
1251 io_kill_timeout(req);
1255 spin_unlock_irq(&ctx->completion_lock);
1256 return canceled != 0;
1259 static void __io_queue_deferred(struct io_ring_ctx *ctx)
1262 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1263 struct io_defer_entry, list);
1265 if (req_need_defer(de->req, de->seq))
1267 list_del_init(&de->list);
1268 io_req_task_queue(de->req);
1270 } while (!list_empty(&ctx->defer_list));
1273 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1277 if (list_empty(&ctx->timeout_list))
1280 seq = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
1283 u32 events_needed, events_got;
1284 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1285 struct io_kiocb, timeout.list);
1287 if (io_is_timeout_noseq(req))
1291 * Since seq can easily wrap around over time, subtract
1292 * the last seq at which timeouts were flushed before comparing.
1293 * Assuming not more than 2^31-1 events have happened since,
1294 * these subtractions won't have wrapped, so we can check if
1295 * target is in [last_seq, current_seq] by comparing the two.
1297 events_needed = req->timeout.target_seq - ctx->cq_last_tm_flush;
1298 events_got = seq - ctx->cq_last_tm_flush;
1299 if (events_got < events_needed)
1302 list_del_init(&req->timeout.list);
1303 io_kill_timeout(req);
1304 } while (!list_empty(&ctx->timeout_list));
1306 ctx->cq_last_tm_flush = seq;
1309 static void io_commit_cqring(struct io_ring_ctx *ctx)
1311 io_flush_timeouts(ctx);
1313 /* order cqe stores with ring update */
1314 smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
1316 if (unlikely(!list_empty(&ctx->defer_list)))
1317 __io_queue_deferred(ctx);
1320 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1322 struct io_rings *r = ctx->rings;
1324 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == r->sq_ring_entries;
1327 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
1329 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
1332 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1334 struct io_rings *rings = ctx->rings;
1338 * writes to the cq entry need to come after reading head; the
1339 * control dependency is enough as we're using WRITE_ONCE to
1342 if (__io_cqring_events(ctx) == rings->cq_ring_entries)
1345 tail = ctx->cached_cq_tail++;
1346 return &rings->cqes[tail & ctx->cq_mask];
1349 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1353 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1355 if (!ctx->eventfd_async)
1357 return io_wq_current_is_worker();
1360 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1362 /* see waitqueue_active() comment */
1365 if (waitqueue_active(&ctx->wait))
1366 wake_up(&ctx->wait);
1367 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1368 wake_up(&ctx->sq_data->wait);
1369 if (io_should_trigger_evfd(ctx))
1370 eventfd_signal(ctx->cq_ev_fd, 1);
1371 if (waitqueue_active(&ctx->cq_wait)) {
1372 wake_up_interruptible(&ctx->cq_wait);
1373 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1377 static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx)
1379 /* see waitqueue_active() comment */
1382 if (ctx->flags & IORING_SETUP_SQPOLL) {
1383 if (waitqueue_active(&ctx->wait))
1384 wake_up(&ctx->wait);
1386 if (io_should_trigger_evfd(ctx))
1387 eventfd_signal(ctx->cq_ev_fd, 1);
1388 if (waitqueue_active(&ctx->cq_wait)) {
1389 wake_up_interruptible(&ctx->cq_wait);
1390 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1394 /* Returns true if there are no backlogged entries after the flush */
1395 static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force,
1396 struct task_struct *tsk,
1397 struct files_struct *files)
1399 struct io_rings *rings = ctx->rings;
1400 struct io_kiocb *req, *tmp;
1401 struct io_uring_cqe *cqe;
1402 unsigned long flags;
1403 bool all_flushed, posted;
1406 if (!force && __io_cqring_events(ctx) == rings->cq_ring_entries)
1410 spin_lock_irqsave(&ctx->completion_lock, flags);
1411 list_for_each_entry_safe(req, tmp, &ctx->cq_overflow_list, compl.list) {
1412 if (!io_match_task(req, tsk, files))
1415 cqe = io_get_cqring(ctx);
1419 list_move(&req->compl.list, &list);
1421 WRITE_ONCE(cqe->user_data, req->user_data);
1422 WRITE_ONCE(cqe->res, req->result);
1423 WRITE_ONCE(cqe->flags, req->compl.cflags);
1425 ctx->cached_cq_overflow++;
1426 WRITE_ONCE(ctx->rings->cq_overflow,
1427 ctx->cached_cq_overflow);
1432 all_flushed = list_empty(&ctx->cq_overflow_list);
1434 clear_bit(0, &ctx->sq_check_overflow);
1435 clear_bit(0, &ctx->cq_check_overflow);
1436 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
1440 io_commit_cqring(ctx);
1441 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1443 io_cqring_ev_posted(ctx);
1445 while (!list_empty(&list)) {
1446 req = list_first_entry(&list, struct io_kiocb, compl.list);
1447 list_del(&req->compl.list);
1454 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force,
1455 struct task_struct *tsk,
1456 struct files_struct *files)
1460 if (test_bit(0, &ctx->cq_check_overflow)) {
1461 /* iopoll syncs against uring_lock, not completion_lock */
1462 if (ctx->flags & IORING_SETUP_IOPOLL)
1463 mutex_lock(&ctx->uring_lock);
1464 ret = __io_cqring_overflow_flush(ctx, force, tsk, files);
1465 if (ctx->flags & IORING_SETUP_IOPOLL)
1466 mutex_unlock(&ctx->uring_lock);
1472 static void __io_cqring_fill_event(struct io_kiocb *req, long res, long cflags)
1474 struct io_ring_ctx *ctx = req->ctx;
1475 struct io_uring_cqe *cqe;
1477 trace_io_uring_complete(ctx, req->user_data, res);
1480 * If we can't get a cq entry, userspace overflowed the
1481 * submission (by quite a lot). Increment the overflow count in
1484 cqe = io_get_cqring(ctx);
1486 WRITE_ONCE(cqe->user_data, req->user_data);
1487 WRITE_ONCE(cqe->res, res);
1488 WRITE_ONCE(cqe->flags, cflags);
1489 } else if (ctx->cq_overflow_flushed ||
1490 atomic_read(&req->task->io_uring->in_idle)) {
1492 * If we're in ring overflow flush mode, or in task cancel mode,
1493 * then we cannot store the request for later flushing, we need
1494 * to drop it on the floor.
1496 ctx->cached_cq_overflow++;
1497 WRITE_ONCE(ctx->rings->cq_overflow, ctx->cached_cq_overflow);
1499 if (list_empty(&ctx->cq_overflow_list)) {
1500 set_bit(0, &ctx->sq_check_overflow);
1501 set_bit(0, &ctx->cq_check_overflow);
1502 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
1506 req->compl.cflags = cflags;
1507 refcount_inc(&req->refs);
1508 list_add_tail(&req->compl.list, &ctx->cq_overflow_list);
1512 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1514 __io_cqring_fill_event(req, res, 0);
1517 static inline void io_req_complete_post(struct io_kiocb *req, long res,
1518 unsigned int cflags)
1520 struct io_ring_ctx *ctx = req->ctx;
1521 unsigned long flags;
1523 spin_lock_irqsave(&ctx->completion_lock, flags);
1524 __io_cqring_fill_event(req, res, cflags);
1525 io_commit_cqring(ctx);
1527 * If we're the last reference to this request, add to our locked
1530 if (refcount_dec_and_test(&req->refs)) {
1531 struct io_comp_state *cs = &ctx->submit_state.comp;
1533 io_dismantle_req(req);
1534 io_put_task(req->task, 1);
1535 list_add(&req->compl.list, &cs->locked_free_list);
1536 cs->locked_free_nr++;
1539 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1541 io_cqring_ev_posted(ctx);
1544 percpu_ref_put(&ctx->refs);
1548 static void io_req_complete_state(struct io_kiocb *req, long res,
1549 unsigned int cflags)
1553 req->compl.cflags = cflags;
1554 req->flags |= REQ_F_COMPLETE_INLINE;
1557 static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags,
1558 long res, unsigned cflags)
1560 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1561 io_req_complete_state(req, res, cflags);
1563 io_req_complete_post(req, res, cflags);
1566 static inline void io_req_complete(struct io_kiocb *req, long res)
1568 __io_req_complete(req, 0, res, 0);
1571 static bool io_flush_cached_reqs(struct io_ring_ctx *ctx)
1573 struct io_submit_state *state = &ctx->submit_state;
1574 struct io_comp_state *cs = &state->comp;
1575 struct io_kiocb *req = NULL;
1578 * If we have more than a batch's worth of requests in our IRQ side
1579 * locked cache, grab the lock and move them over to our submission
1582 if (READ_ONCE(cs->locked_free_nr) > IO_COMPL_BATCH) {
1583 spin_lock_irq(&ctx->completion_lock);
1584 list_splice_init(&cs->locked_free_list, &cs->free_list);
1585 cs->locked_free_nr = 0;
1586 spin_unlock_irq(&ctx->completion_lock);
1589 while (!list_empty(&cs->free_list)) {
1590 req = list_first_entry(&cs->free_list, struct io_kiocb,
1592 list_del(&req->compl.list);
1593 state->reqs[state->free_reqs++] = req;
1594 if (state->free_reqs == ARRAY_SIZE(state->reqs))
1601 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
1603 struct io_submit_state *state = &ctx->submit_state;
1605 BUILD_BUG_ON(IO_REQ_ALLOC_BATCH > ARRAY_SIZE(state->reqs));
1607 if (!state->free_reqs) {
1608 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1611 if (io_flush_cached_reqs(ctx))
1614 ret = kmem_cache_alloc_bulk(req_cachep, gfp, IO_REQ_ALLOC_BATCH,
1618 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1619 * retry single alloc to be on the safe side.
1621 if (unlikely(ret <= 0)) {
1622 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1623 if (!state->reqs[0])
1627 state->free_reqs = ret;
1631 return state->reqs[state->free_reqs];
1634 static inline void io_put_file(struct io_kiocb *req, struct file *file,
1641 static void io_dismantle_req(struct io_kiocb *req)
1645 if (req->async_data)
1646 kfree(req->async_data);
1648 io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
1649 if (req->fixed_rsrc_refs)
1650 percpu_ref_put(req->fixed_rsrc_refs);
1652 if (req->flags & REQ_F_INFLIGHT) {
1653 struct io_ring_ctx *ctx = req->ctx;
1654 unsigned long flags;
1656 spin_lock_irqsave(&ctx->inflight_lock, flags);
1657 list_del(&req->inflight_entry);
1658 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
1659 req->flags &= ~REQ_F_INFLIGHT;
1663 /* must to be called somewhat shortly after putting a request */
1664 static inline void io_put_task(struct task_struct *task, int nr)
1666 struct io_uring_task *tctx = task->io_uring;
1668 percpu_counter_sub(&tctx->inflight, nr);
1669 if (unlikely(atomic_read(&tctx->in_idle)))
1670 wake_up(&tctx->wait);
1671 put_task_struct_many(task, nr);
1674 static void __io_free_req(struct io_kiocb *req)
1676 struct io_ring_ctx *ctx = req->ctx;
1678 io_dismantle_req(req);
1679 io_put_task(req->task, 1);
1681 kmem_cache_free(req_cachep, req);
1682 percpu_ref_put(&ctx->refs);
1685 static inline void io_remove_next_linked(struct io_kiocb *req)
1687 struct io_kiocb *nxt = req->link;
1689 req->link = nxt->link;
1693 static void io_kill_linked_timeout(struct io_kiocb *req)
1695 struct io_ring_ctx *ctx = req->ctx;
1696 struct io_kiocb *link;
1697 bool cancelled = false;
1698 unsigned long flags;
1700 spin_lock_irqsave(&ctx->completion_lock, flags);
1704 * Can happen if a linked timeout fired and link had been like
1705 * req -> link t-out -> link t-out [-> ...]
1707 if (link && (link->flags & REQ_F_LTIMEOUT_ACTIVE)) {
1708 struct io_timeout_data *io = link->async_data;
1711 io_remove_next_linked(req);
1712 link->timeout.head = NULL;
1713 ret = hrtimer_try_to_cancel(&io->timer);
1715 io_cqring_fill_event(link, -ECANCELED);
1716 io_commit_cqring(ctx);
1720 req->flags &= ~REQ_F_LINK_TIMEOUT;
1721 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1724 io_cqring_ev_posted(ctx);
1730 static void io_fail_links(struct io_kiocb *req)
1732 struct io_kiocb *link, *nxt;
1733 struct io_ring_ctx *ctx = req->ctx;
1734 unsigned long flags;
1736 spin_lock_irqsave(&ctx->completion_lock, flags);
1744 trace_io_uring_fail_link(req, link);
1745 io_cqring_fill_event(link, -ECANCELED);
1747 io_put_req_deferred(link, 2);
1750 io_commit_cqring(ctx);
1751 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1753 io_cqring_ev_posted(ctx);
1756 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1758 if (req->flags & REQ_F_LINK_TIMEOUT)
1759 io_kill_linked_timeout(req);
1762 * If LINK is set, we have dependent requests in this chain. If we
1763 * didn't fail this request, queue the first one up, moving any other
1764 * dependencies to the next request. In case of failure, fail the rest
1767 if (likely(!(req->flags & REQ_F_FAIL_LINK))) {
1768 struct io_kiocb *nxt = req->link;
1777 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1779 if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK))))
1781 return __io_req_find_next(req);
1784 static void ctx_flush_and_put(struct io_ring_ctx *ctx)
1788 if (ctx->submit_state.comp.nr) {
1789 mutex_lock(&ctx->uring_lock);
1790 io_submit_flush_completions(&ctx->submit_state.comp, ctx);
1791 mutex_unlock(&ctx->uring_lock);
1793 percpu_ref_put(&ctx->refs);
1796 static bool __tctx_task_work(struct io_uring_task *tctx)
1798 struct io_ring_ctx *ctx = NULL;
1799 struct io_wq_work_list list;
1800 struct io_wq_work_node *node;
1802 if (wq_list_empty(&tctx->task_list))
1805 spin_lock_irq(&tctx->task_lock);
1806 list = tctx->task_list;
1807 INIT_WQ_LIST(&tctx->task_list);
1808 spin_unlock_irq(&tctx->task_lock);
1812 struct io_wq_work_node *next = node->next;
1813 struct io_kiocb *req;
1815 req = container_of(node, struct io_kiocb, io_task_work.node);
1816 if (req->ctx != ctx) {
1817 ctx_flush_and_put(ctx);
1819 percpu_ref_get(&ctx->refs);
1822 req->task_work.func(&req->task_work);
1826 ctx_flush_and_put(ctx);
1827 return list.first != NULL;
1830 static void tctx_task_work(struct callback_head *cb)
1832 struct io_uring_task *tctx = container_of(cb, struct io_uring_task, task_work);
1834 clear_bit(0, &tctx->task_state);
1836 while (__tctx_task_work(tctx))
1840 static int io_task_work_add(struct task_struct *tsk, struct io_kiocb *req,
1841 enum task_work_notify_mode notify)
1843 struct io_uring_task *tctx = tsk->io_uring;
1844 struct io_wq_work_node *node, *prev;
1845 unsigned long flags;
1848 WARN_ON_ONCE(!tctx);
1850 spin_lock_irqsave(&tctx->task_lock, flags);
1851 wq_list_add_tail(&req->io_task_work.node, &tctx->task_list);
1852 spin_unlock_irqrestore(&tctx->task_lock, flags);
1854 /* task_work already pending, we're done */
1855 if (test_bit(0, &tctx->task_state) ||
1856 test_and_set_bit(0, &tctx->task_state))
1859 if (!task_work_add(tsk, &tctx->task_work, notify))
1863 * Slow path - we failed, find and delete work. if the work is not
1864 * in the list, it got run and we're fine.
1867 spin_lock_irqsave(&tctx->task_lock, flags);
1868 wq_list_for_each(node, prev, &tctx->task_list) {
1869 if (&req->io_task_work.node == node) {
1870 wq_list_del(&tctx->task_list, node, prev);
1875 spin_unlock_irqrestore(&tctx->task_lock, flags);
1876 clear_bit(0, &tctx->task_state);
1880 static int io_req_task_work_add(struct io_kiocb *req)
1882 struct task_struct *tsk = req->task;
1883 struct io_ring_ctx *ctx = req->ctx;
1884 enum task_work_notify_mode notify;
1887 if (tsk->flags & PF_EXITING)
1891 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
1892 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
1893 * processing task_work. There's no reliable way to tell if TWA_RESUME
1897 if (!(ctx->flags & IORING_SETUP_SQPOLL))
1898 notify = TWA_SIGNAL;
1900 ret = io_task_work_add(tsk, req, notify);
1902 wake_up_process(tsk);
1907 static void io_req_task_work_add_fallback(struct io_kiocb *req,
1908 task_work_func_t cb)
1910 struct io_ring_ctx *ctx = req->ctx;
1911 struct callback_head *head;
1913 init_task_work(&req->task_work, cb);
1915 head = READ_ONCE(ctx->exit_task_work);
1916 req->task_work.next = head;
1917 } while (cmpxchg(&ctx->exit_task_work, head, &req->task_work) != head);
1920 static void __io_req_task_cancel(struct io_kiocb *req, int error)
1922 struct io_ring_ctx *ctx = req->ctx;
1924 spin_lock_irq(&ctx->completion_lock);
1925 io_cqring_fill_event(req, error);
1926 io_commit_cqring(ctx);
1927 spin_unlock_irq(&ctx->completion_lock);
1929 io_cqring_ev_posted(ctx);
1930 req_set_fail_links(req);
1931 io_double_put_req(req);
1934 static void io_req_task_cancel(struct callback_head *cb)
1936 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1937 struct io_ring_ctx *ctx = req->ctx;
1939 mutex_lock(&ctx->uring_lock);
1940 __io_req_task_cancel(req, req->result);
1941 mutex_unlock(&ctx->uring_lock);
1942 percpu_ref_put(&ctx->refs);
1945 static void __io_req_task_submit(struct io_kiocb *req)
1947 struct io_ring_ctx *ctx = req->ctx;
1949 /* ctx stays valid until unlock, even if we drop all ours ctx->refs */
1950 mutex_lock(&ctx->uring_lock);
1951 if (!(current->flags & PF_EXITING) && !current->in_execve)
1952 __io_queue_sqe(req);
1954 __io_req_task_cancel(req, -EFAULT);
1955 mutex_unlock(&ctx->uring_lock);
1958 static void io_req_task_submit(struct callback_head *cb)
1960 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1962 __io_req_task_submit(req);
1965 static void io_req_task_queue(struct io_kiocb *req)
1969 req->task_work.func = io_req_task_submit;
1970 ret = io_req_task_work_add(req);
1971 if (unlikely(ret)) {
1972 req->result = -ECANCELED;
1973 percpu_ref_get(&req->ctx->refs);
1974 io_req_task_work_add_fallback(req, io_req_task_cancel);
1978 static void io_req_task_queue_fail(struct io_kiocb *req, int ret)
1980 percpu_ref_get(&req->ctx->refs);
1982 req->task_work.func = io_req_task_cancel;
1984 if (unlikely(io_req_task_work_add(req)))
1985 io_req_task_work_add_fallback(req, io_req_task_cancel);
1988 static inline void io_queue_next(struct io_kiocb *req)
1990 struct io_kiocb *nxt = io_req_find_next(req);
1993 io_req_task_queue(nxt);
1996 static void io_free_req(struct io_kiocb *req)
2003 struct task_struct *task;
2008 static inline void io_init_req_batch(struct req_batch *rb)
2015 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2016 struct req_batch *rb)
2019 io_put_task(rb->task, rb->task_refs);
2021 percpu_ref_put_many(&ctx->refs, rb->ctx_refs);
2024 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req,
2025 struct io_submit_state *state)
2029 if (req->task != rb->task) {
2031 io_put_task(rb->task, rb->task_refs);
2032 rb->task = req->task;
2038 io_dismantle_req(req);
2039 if (state->free_reqs != ARRAY_SIZE(state->reqs))
2040 state->reqs[state->free_reqs++] = req;
2042 list_add(&req->compl.list, &state->comp.free_list);
2045 static void io_submit_flush_completions(struct io_comp_state *cs,
2046 struct io_ring_ctx *ctx)
2049 struct io_kiocb *req;
2050 struct req_batch rb;
2052 io_init_req_batch(&rb);
2053 spin_lock_irq(&ctx->completion_lock);
2054 for (i = 0; i < nr; i++) {
2056 __io_cqring_fill_event(req, req->result, req->compl.cflags);
2058 io_commit_cqring(ctx);
2059 spin_unlock_irq(&ctx->completion_lock);
2061 io_cqring_ev_posted(ctx);
2062 for (i = 0; i < nr; i++) {
2065 /* submission and completion refs */
2066 if (refcount_sub_and_test(2, &req->refs))
2067 io_req_free_batch(&rb, req, &ctx->submit_state);
2070 io_req_free_batch_finish(ctx, &rb);
2075 * Drop reference to request, return next in chain (if there is one) if this
2076 * was the last reference to this request.
2078 static struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2080 struct io_kiocb *nxt = NULL;
2082 if (refcount_dec_and_test(&req->refs)) {
2083 nxt = io_req_find_next(req);
2089 static void io_put_req(struct io_kiocb *req)
2091 if (refcount_dec_and_test(&req->refs))
2095 static void io_put_req_deferred_cb(struct callback_head *cb)
2097 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2102 static void io_free_req_deferred(struct io_kiocb *req)
2106 req->task_work.func = io_put_req_deferred_cb;
2107 ret = io_req_task_work_add(req);
2109 io_req_task_work_add_fallback(req, io_put_req_deferred_cb);
2112 static inline void io_put_req_deferred(struct io_kiocb *req, int refs)
2114 if (refcount_sub_and_test(refs, &req->refs))
2115 io_free_req_deferred(req);
2118 static void io_double_put_req(struct io_kiocb *req)
2120 /* drop both submit and complete references */
2121 if (refcount_sub_and_test(2, &req->refs))
2125 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2127 /* See comment at the top of this file */
2129 return __io_cqring_events(ctx);
2132 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2134 struct io_rings *rings = ctx->rings;
2136 /* make sure SQ entry isn't read before tail */
2137 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2140 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2142 unsigned int cflags;
2144 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2145 cflags |= IORING_CQE_F_BUFFER;
2146 req->flags &= ~REQ_F_BUFFER_SELECTED;
2151 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2153 struct io_buffer *kbuf;
2155 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2156 return io_put_kbuf(req, kbuf);
2159 static inline bool io_run_task_work(void)
2162 * Not safe to run on exiting task, and the task_work handling will
2163 * not add work to such a task.
2165 if (unlikely(current->flags & PF_EXITING))
2167 if (current->task_works) {
2168 __set_current_state(TASK_RUNNING);
2177 * Find and free completed poll iocbs
2179 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2180 struct list_head *done)
2182 struct req_batch rb;
2183 struct io_kiocb *req;
2185 /* order with ->result store in io_complete_rw_iopoll() */
2188 io_init_req_batch(&rb);
2189 while (!list_empty(done)) {
2192 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2193 list_del(&req->inflight_entry);
2195 if (READ_ONCE(req->result) == -EAGAIN) {
2196 req->iopoll_completed = 0;
2197 if (io_rw_reissue(req))
2201 if (req->flags & REQ_F_BUFFER_SELECTED)
2202 cflags = io_put_rw_kbuf(req);
2204 __io_cqring_fill_event(req, req->result, cflags);
2207 if (refcount_dec_and_test(&req->refs))
2208 io_req_free_batch(&rb, req, &ctx->submit_state);
2211 io_commit_cqring(ctx);
2212 io_cqring_ev_posted_iopoll(ctx);
2213 io_req_free_batch_finish(ctx, &rb);
2216 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2219 struct io_kiocb *req, *tmp;
2225 * Only spin for completions if we don't have multiple devices hanging
2226 * off our complete list, and we're under the requested amount.
2228 spin = !ctx->poll_multi_file && *nr_events < min;
2231 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2232 struct kiocb *kiocb = &req->rw.kiocb;
2235 * Move completed and retryable entries to our local lists.
2236 * If we find a request that requires polling, break out
2237 * and complete those lists first, if we have entries there.
2239 if (READ_ONCE(req->iopoll_completed)) {
2240 list_move_tail(&req->inflight_entry, &done);
2243 if (!list_empty(&done))
2246 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2250 /* iopoll may have completed current req */
2251 if (READ_ONCE(req->iopoll_completed))
2252 list_move_tail(&req->inflight_entry, &done);
2259 if (!list_empty(&done))
2260 io_iopoll_complete(ctx, nr_events, &done);
2266 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2267 * non-spinning poll check - we'll still enter the driver poll loop, but only
2268 * as a non-spinning completion check.
2270 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
2273 while (!list_empty(&ctx->iopoll_list) && !need_resched()) {
2276 ret = io_do_iopoll(ctx, nr_events, min);
2279 if (*nr_events >= min)
2287 * We can't just wait for polled events to come to us, we have to actively
2288 * find and complete them.
2290 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2292 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2295 mutex_lock(&ctx->uring_lock);
2296 while (!list_empty(&ctx->iopoll_list)) {
2297 unsigned int nr_events = 0;
2299 io_do_iopoll(ctx, &nr_events, 0);
2301 /* let it sleep and repeat later if can't complete a request */
2305 * Ensure we allow local-to-the-cpu processing to take place,
2306 * in this case we need to ensure that we reap all events.
2307 * Also let task_work, etc. to progress by releasing the mutex
2309 if (need_resched()) {
2310 mutex_unlock(&ctx->uring_lock);
2312 mutex_lock(&ctx->uring_lock);
2315 mutex_unlock(&ctx->uring_lock);
2318 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2320 unsigned int nr_events = 0;
2321 int iters = 0, ret = 0;
2324 * We disallow the app entering submit/complete with polling, but we
2325 * still need to lock the ring to prevent racing with polled issue
2326 * that got punted to a workqueue.
2328 mutex_lock(&ctx->uring_lock);
2331 * Don't enter poll loop if we already have events pending.
2332 * If we do, we can potentially be spinning for commands that
2333 * already triggered a CQE (eg in error).
2335 if (test_bit(0, &ctx->cq_check_overflow))
2336 __io_cqring_overflow_flush(ctx, false, NULL, NULL);
2337 if (io_cqring_events(ctx))
2341 * If a submit got punted to a workqueue, we can have the
2342 * application entering polling for a command before it gets
2343 * issued. That app will hold the uring_lock for the duration
2344 * of the poll right here, so we need to take a breather every
2345 * now and then to ensure that the issue has a chance to add
2346 * the poll to the issued list. Otherwise we can spin here
2347 * forever, while the workqueue is stuck trying to acquire the
2350 if (!(++iters & 7)) {
2351 mutex_unlock(&ctx->uring_lock);
2353 mutex_lock(&ctx->uring_lock);
2356 ret = io_iopoll_getevents(ctx, &nr_events, min);
2360 } while (min && !nr_events && !need_resched());
2362 mutex_unlock(&ctx->uring_lock);
2366 static void kiocb_end_write(struct io_kiocb *req)
2369 * Tell lockdep we inherited freeze protection from submission
2372 if (req->flags & REQ_F_ISREG) {
2373 struct inode *inode = file_inode(req->file);
2375 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
2377 file_end_write(req->file);
2381 static bool io_resubmit_prep(struct io_kiocb *req)
2383 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2385 struct iov_iter iter;
2387 /* already prepared */
2388 if (req->async_data)
2391 switch (req->opcode) {
2392 case IORING_OP_READV:
2393 case IORING_OP_READ_FIXED:
2394 case IORING_OP_READ:
2397 case IORING_OP_WRITEV:
2398 case IORING_OP_WRITE_FIXED:
2399 case IORING_OP_WRITE:
2403 printk_once(KERN_WARNING "io_uring: bad opcode in resubmit %d\n",
2408 ret = io_import_iovec(rw, req, &iovec, &iter, false);
2411 return !io_setup_async_rw(req, iovec, inline_vecs, &iter, false);
2414 static bool io_rw_should_reissue(struct io_kiocb *req)
2416 umode_t mode = file_inode(req->file)->i_mode;
2417 struct io_ring_ctx *ctx = req->ctx;
2419 if (!S_ISBLK(mode) && !S_ISREG(mode))
2421 if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
2422 !(ctx->flags & IORING_SETUP_IOPOLL)))
2425 * If ref is dying, we might be running poll reap from the exit work.
2426 * Don't attempt to reissue from that path, just let it fail with
2429 if (percpu_ref_is_dying(&ctx->refs))
2435 static bool io_rw_reissue(struct io_kiocb *req)
2438 if (!io_rw_should_reissue(req))
2441 lockdep_assert_held(&req->ctx->uring_lock);
2443 if (io_resubmit_prep(req)) {
2444 refcount_inc(&req->refs);
2445 io_queue_async_work(req);
2448 req_set_fail_links(req);
2453 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2454 unsigned int issue_flags)
2458 if ((res == -EAGAIN || res == -EOPNOTSUPP) && io_rw_reissue(req))
2460 if (res != req->result)
2461 req_set_fail_links(req);
2463 if (req->rw.kiocb.ki_flags & IOCB_WRITE)
2464 kiocb_end_write(req);
2465 if (req->flags & REQ_F_BUFFER_SELECTED)
2466 cflags = io_put_rw_kbuf(req);
2467 __io_req_complete(req, issue_flags, res, cflags);
2470 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2472 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2474 __io_complete_rw(req, res, res2, 0);
2477 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2479 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2482 /* Rewind iter, if we have one. iopoll path resubmits as usual */
2483 if (res == -EAGAIN && io_rw_should_reissue(req)) {
2484 struct io_async_rw *rw = req->async_data;
2487 iov_iter_revert(&rw->iter,
2488 req->result - iov_iter_count(&rw->iter));
2489 else if (!io_resubmit_prep(req))
2494 if (kiocb->ki_flags & IOCB_WRITE)
2495 kiocb_end_write(req);
2497 if (res != -EAGAIN && res != req->result)
2498 req_set_fail_links(req);
2500 WRITE_ONCE(req->result, res);
2501 /* order with io_poll_complete() checking ->result */
2503 WRITE_ONCE(req->iopoll_completed, 1);
2507 * After the iocb has been issued, it's safe to be found on the poll list.
2508 * Adding the kiocb to the list AFTER submission ensures that we don't
2509 * find it from a io_iopoll_getevents() thread before the issuer is done
2510 * accessing the kiocb cookie.
2512 static void io_iopoll_req_issued(struct io_kiocb *req, bool in_async)
2514 struct io_ring_ctx *ctx = req->ctx;
2517 * Track whether we have multiple files in our lists. This will impact
2518 * how we do polling eventually, not spinning if we're on potentially
2519 * different devices.
2521 if (list_empty(&ctx->iopoll_list)) {
2522 ctx->poll_multi_file = false;
2523 } else if (!ctx->poll_multi_file) {
2524 struct io_kiocb *list_req;
2526 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2528 if (list_req->file != req->file)
2529 ctx->poll_multi_file = true;
2533 * For fast devices, IO may have already completed. If it has, add
2534 * it to the front so we find it first.
2536 if (READ_ONCE(req->iopoll_completed))
2537 list_add(&req->inflight_entry, &ctx->iopoll_list);
2539 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2542 * If IORING_SETUP_SQPOLL is enabled, sqes are either handled in sq thread
2543 * task context or in io worker task context. If current task context is
2544 * sq thread, we don't need to check whether should wake up sq thread.
2546 if (in_async && (ctx->flags & IORING_SETUP_SQPOLL) &&
2547 wq_has_sleeper(&ctx->sq_data->wait))
2548 wake_up(&ctx->sq_data->wait);
2551 static inline void io_state_file_put(struct io_submit_state *state)
2553 if (state->file_refs) {
2554 fput_many(state->file, state->file_refs);
2555 state->file_refs = 0;
2560 * Get as many references to a file as we have IOs left in this submission,
2561 * assuming most submissions are for one file, or at least that each file
2562 * has more than one submission.
2564 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2569 if (state->file_refs) {
2570 if (state->fd == fd) {
2574 io_state_file_put(state);
2576 state->file = fget_many(fd, state->ios_left);
2577 if (unlikely(!state->file))
2581 state->file_refs = state->ios_left - 1;
2585 static bool io_bdev_nowait(struct block_device *bdev)
2587 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2591 * If we tracked the file through the SCM inflight mechanism, we could support
2592 * any file. For now, just ensure that anything potentially problematic is done
2595 static bool io_file_supports_async(struct file *file, int rw)
2597 umode_t mode = file_inode(file)->i_mode;
2599 if (S_ISBLK(mode)) {
2600 if (IS_ENABLED(CONFIG_BLOCK) &&
2601 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2605 if (S_ISCHR(mode) || S_ISSOCK(mode))
2607 if (S_ISREG(mode)) {
2608 if (IS_ENABLED(CONFIG_BLOCK) &&
2609 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2610 file->f_op != &io_uring_fops)
2615 /* any ->read/write should understand O_NONBLOCK */
2616 if (file->f_flags & O_NONBLOCK)
2619 if (!(file->f_mode & FMODE_NOWAIT))
2623 return file->f_op->read_iter != NULL;
2625 return file->f_op->write_iter != NULL;
2628 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2630 struct io_ring_ctx *ctx = req->ctx;
2631 struct kiocb *kiocb = &req->rw.kiocb;
2632 struct file *file = req->file;
2636 if (S_ISREG(file_inode(file)->i_mode))
2637 req->flags |= REQ_F_ISREG;
2639 kiocb->ki_pos = READ_ONCE(sqe->off);
2640 if (kiocb->ki_pos == -1 && !(file->f_mode & FMODE_STREAM)) {
2641 req->flags |= REQ_F_CUR_POS;
2642 kiocb->ki_pos = file->f_pos;
2644 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2645 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2646 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2650 /* don't allow async punt for O_NONBLOCK or RWF_NOWAIT */
2651 if ((kiocb->ki_flags & IOCB_NOWAIT) || (file->f_flags & O_NONBLOCK))
2652 req->flags |= REQ_F_NOWAIT;
2654 ioprio = READ_ONCE(sqe->ioprio);
2656 ret = ioprio_check_cap(ioprio);
2660 kiocb->ki_ioprio = ioprio;
2662 kiocb->ki_ioprio = get_current_ioprio();
2664 if (ctx->flags & IORING_SETUP_IOPOLL) {
2665 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2666 !kiocb->ki_filp->f_op->iopoll)
2669 kiocb->ki_flags |= IOCB_HIPRI;
2670 kiocb->ki_complete = io_complete_rw_iopoll;
2671 req->iopoll_completed = 0;
2673 if (kiocb->ki_flags & IOCB_HIPRI)
2675 kiocb->ki_complete = io_complete_rw;
2678 req->rw.addr = READ_ONCE(sqe->addr);
2679 req->rw.len = READ_ONCE(sqe->len);
2680 req->buf_index = READ_ONCE(sqe->buf_index);
2684 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2690 case -ERESTARTNOINTR:
2691 case -ERESTARTNOHAND:
2692 case -ERESTART_RESTARTBLOCK:
2694 * We can't just restart the syscall, since previously
2695 * submitted sqes may already be in progress. Just fail this
2701 kiocb->ki_complete(kiocb, ret, 0);
2705 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2706 unsigned int issue_flags)
2708 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2709 struct io_async_rw *io = req->async_data;
2711 /* add previously done IO, if any */
2712 if (io && io->bytes_done > 0) {
2714 ret = io->bytes_done;
2716 ret += io->bytes_done;
2719 if (req->flags & REQ_F_CUR_POS)
2720 req->file->f_pos = kiocb->ki_pos;
2721 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2722 __io_complete_rw(req, ret, 0, issue_flags);
2724 io_rw_done(kiocb, ret);
2727 static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
2729 struct io_ring_ctx *ctx = req->ctx;
2730 size_t len = req->rw.len;
2731 struct io_mapped_ubuf *imu;
2732 u16 index, buf_index = req->buf_index;
2736 if (unlikely(buf_index >= ctx->nr_user_bufs))
2738 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2739 imu = &ctx->user_bufs[index];
2740 buf_addr = req->rw.addr;
2743 if (buf_addr + len < buf_addr)
2745 /* not inside the mapped region */
2746 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2750 * May not be a start of buffer, set size appropriately
2751 * and advance us to the beginning.
2753 offset = buf_addr - imu->ubuf;
2754 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2758 * Don't use iov_iter_advance() here, as it's really slow for
2759 * using the latter parts of a big fixed buffer - it iterates
2760 * over each segment manually. We can cheat a bit here, because
2763 * 1) it's a BVEC iter, we set it up
2764 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2765 * first and last bvec
2767 * So just find our index, and adjust the iterator afterwards.
2768 * If the offset is within the first bvec (or the whole first
2769 * bvec, just use iov_iter_advance(). This makes it easier
2770 * since we can just skip the first segment, which may not
2771 * be PAGE_SIZE aligned.
2773 const struct bio_vec *bvec = imu->bvec;
2775 if (offset <= bvec->bv_len) {
2776 iov_iter_advance(iter, offset);
2778 unsigned long seg_skip;
2780 /* skip first vec */
2781 offset -= bvec->bv_len;
2782 seg_skip = 1 + (offset >> PAGE_SHIFT);
2784 iter->bvec = bvec + seg_skip;
2785 iter->nr_segs -= seg_skip;
2786 iter->count -= bvec->bv_len + offset;
2787 iter->iov_offset = offset & ~PAGE_MASK;
2794 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2797 mutex_unlock(&ctx->uring_lock);
2800 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2803 * "Normal" inline submissions always hold the uring_lock, since we
2804 * grab it from the system call. Same is true for the SQPOLL offload.
2805 * The only exception is when we've detached the request and issue it
2806 * from an async worker thread, grab the lock for that case.
2809 mutex_lock(&ctx->uring_lock);
2812 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2813 int bgid, struct io_buffer *kbuf,
2816 struct io_buffer *head;
2818 if (req->flags & REQ_F_BUFFER_SELECTED)
2821 io_ring_submit_lock(req->ctx, needs_lock);
2823 lockdep_assert_held(&req->ctx->uring_lock);
2825 head = idr_find(&req->ctx->io_buffer_idr, bgid);
2827 if (!list_empty(&head->list)) {
2828 kbuf = list_last_entry(&head->list, struct io_buffer,
2830 list_del(&kbuf->list);
2833 idr_remove(&req->ctx->io_buffer_idr, bgid);
2835 if (*len > kbuf->len)
2838 kbuf = ERR_PTR(-ENOBUFS);
2841 io_ring_submit_unlock(req->ctx, needs_lock);
2846 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2849 struct io_buffer *kbuf;
2852 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2853 bgid = req->buf_index;
2854 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2857 req->rw.addr = (u64) (unsigned long) kbuf;
2858 req->flags |= REQ_F_BUFFER_SELECTED;
2859 return u64_to_user_ptr(kbuf->addr);
2862 #ifdef CONFIG_COMPAT
2863 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2866 struct compat_iovec __user *uiov;
2867 compat_ssize_t clen;
2871 uiov = u64_to_user_ptr(req->rw.addr);
2872 if (!access_ok(uiov, sizeof(*uiov)))
2874 if (__get_user(clen, &uiov->iov_len))
2880 buf = io_rw_buffer_select(req, &len, needs_lock);
2882 return PTR_ERR(buf);
2883 iov[0].iov_base = buf;
2884 iov[0].iov_len = (compat_size_t) len;
2889 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2892 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2896 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2899 len = iov[0].iov_len;
2902 buf = io_rw_buffer_select(req, &len, needs_lock);
2904 return PTR_ERR(buf);
2905 iov[0].iov_base = buf;
2906 iov[0].iov_len = len;
2910 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2913 if (req->flags & REQ_F_BUFFER_SELECTED) {
2914 struct io_buffer *kbuf;
2916 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2917 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
2918 iov[0].iov_len = kbuf->len;
2921 if (req->rw.len != 1)
2924 #ifdef CONFIG_COMPAT
2925 if (req->ctx->compat)
2926 return io_compat_import(req, iov, needs_lock);
2929 return __io_iov_buffer_select(req, iov, needs_lock);
2932 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
2933 struct iov_iter *iter, bool needs_lock)
2935 void __user *buf = u64_to_user_ptr(req->rw.addr);
2936 size_t sqe_len = req->rw.len;
2937 u8 opcode = req->opcode;
2940 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
2942 return io_import_fixed(req, rw, iter);
2945 /* buffer index only valid with fixed read/write, or buffer select */
2946 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
2949 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
2950 if (req->flags & REQ_F_BUFFER_SELECT) {
2951 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
2953 return PTR_ERR(buf);
2954 req->rw.len = sqe_len;
2957 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
2962 if (req->flags & REQ_F_BUFFER_SELECT) {
2963 ret = io_iov_buffer_select(req, *iovec, needs_lock);
2965 iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
2970 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
2974 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
2976 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
2980 * For files that don't have ->read_iter() and ->write_iter(), handle them
2981 * by looping over ->read() or ->write() manually.
2983 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
2985 struct kiocb *kiocb = &req->rw.kiocb;
2986 struct file *file = req->file;
2990 * Don't support polled IO through this interface, and we can't
2991 * support non-blocking either. For the latter, this just causes
2992 * the kiocb to be handled from an async context.
2994 if (kiocb->ki_flags & IOCB_HIPRI)
2996 if (kiocb->ki_flags & IOCB_NOWAIT)
2999 while (iov_iter_count(iter)) {
3003 if (!iov_iter_is_bvec(iter)) {
3004 iovec = iov_iter_iovec(iter);
3006 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3007 iovec.iov_len = req->rw.len;
3011 nr = file->f_op->read(file, iovec.iov_base,
3012 iovec.iov_len, io_kiocb_ppos(kiocb));
3014 nr = file->f_op->write(file, iovec.iov_base,
3015 iovec.iov_len, io_kiocb_ppos(kiocb));
3024 if (nr != iovec.iov_len)
3028 iov_iter_advance(iter, nr);
3034 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3035 const struct iovec *fast_iov, struct iov_iter *iter)
3037 struct io_async_rw *rw = req->async_data;
3039 memcpy(&rw->iter, iter, sizeof(*iter));
3040 rw->free_iovec = iovec;
3042 /* can only be fixed buffers, no need to do anything */
3043 if (iov_iter_is_bvec(iter))
3046 unsigned iov_off = 0;
3048 rw->iter.iov = rw->fast_iov;
3049 if (iter->iov != fast_iov) {
3050 iov_off = iter->iov - fast_iov;
3051 rw->iter.iov += iov_off;
3053 if (rw->fast_iov != fast_iov)
3054 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3055 sizeof(struct iovec) * iter->nr_segs);
3057 req->flags |= REQ_F_NEED_CLEANUP;
3061 static inline int __io_alloc_async_data(struct io_kiocb *req)
3063 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3064 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3065 return req->async_data == NULL;
3068 static int io_alloc_async_data(struct io_kiocb *req)
3070 if (!io_op_defs[req->opcode].needs_async_data)
3073 return __io_alloc_async_data(req);
3076 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3077 const struct iovec *fast_iov,
3078 struct iov_iter *iter, bool force)
3080 if (!force && !io_op_defs[req->opcode].needs_async_data)
3082 if (!req->async_data) {
3083 if (__io_alloc_async_data(req)) {
3088 io_req_map_rw(req, iovec, fast_iov, iter);
3093 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3095 struct io_async_rw *iorw = req->async_data;
3096 struct iovec *iov = iorw->fast_iov;
3099 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3100 if (unlikely(ret < 0))
3103 iorw->bytes_done = 0;
3104 iorw->free_iovec = iov;
3106 req->flags |= REQ_F_NEED_CLEANUP;
3110 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3112 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3114 return io_prep_rw(req, sqe);
3118 * This is our waitqueue callback handler, registered through lock_page_async()
3119 * when we initially tried to do the IO with the iocb armed our waitqueue.
3120 * This gets called when the page is unlocked, and we generally expect that to
3121 * happen when the page IO is completed and the page is now uptodate. This will
3122 * queue a task_work based retry of the operation, attempting to copy the data
3123 * again. If the latter fails because the page was NOT uptodate, then we will
3124 * do a thread based blocking retry of the operation. That's the unexpected
3127 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3128 int sync, void *arg)
3130 struct wait_page_queue *wpq;
3131 struct io_kiocb *req = wait->private;
3132 struct wait_page_key *key = arg;
3134 wpq = container_of(wait, struct wait_page_queue, wait);
3136 if (!wake_page_match(wpq, key))
3139 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3140 list_del_init(&wait->entry);
3142 /* submit ref gets dropped, acquire a new one */
3143 refcount_inc(&req->refs);
3144 io_req_task_queue(req);
3149 * This controls whether a given IO request should be armed for async page
3150 * based retry. If we return false here, the request is handed to the async
3151 * worker threads for retry. If we're doing buffered reads on a regular file,
3152 * we prepare a private wait_page_queue entry and retry the operation. This
3153 * will either succeed because the page is now uptodate and unlocked, or it
3154 * will register a callback when the page is unlocked at IO completion. Through
3155 * that callback, io_uring uses task_work to setup a retry of the operation.
3156 * That retry will attempt the buffered read again. The retry will generally
3157 * succeed, or in rare cases where it fails, we then fall back to using the
3158 * async worker threads for a blocking retry.
3160 static bool io_rw_should_retry(struct io_kiocb *req)
3162 struct io_async_rw *rw = req->async_data;
3163 struct wait_page_queue *wait = &rw->wpq;
3164 struct kiocb *kiocb = &req->rw.kiocb;
3166 /* never retry for NOWAIT, we just complete with -EAGAIN */
3167 if (req->flags & REQ_F_NOWAIT)
3170 /* Only for buffered IO */
3171 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3175 * just use poll if we can, and don't attempt if the fs doesn't
3176 * support callback based unlocks
3178 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3181 wait->wait.func = io_async_buf_func;
3182 wait->wait.private = req;
3183 wait->wait.flags = 0;
3184 INIT_LIST_HEAD(&wait->wait.entry);
3185 kiocb->ki_flags |= IOCB_WAITQ;
3186 kiocb->ki_flags &= ~IOCB_NOWAIT;
3187 kiocb->ki_waitq = wait;
3191 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3193 if (req->file->f_op->read_iter)
3194 return call_read_iter(req->file, &req->rw.kiocb, iter);
3195 else if (req->file->f_op->read)
3196 return loop_rw_iter(READ, req, iter);
3201 static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3203 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3204 struct kiocb *kiocb = &req->rw.kiocb;
3205 struct iov_iter __iter, *iter = &__iter;
3206 struct io_async_rw *rw = req->async_data;
3207 ssize_t io_size, ret, ret2;
3208 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3214 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3218 io_size = iov_iter_count(iter);
3219 req->result = io_size;
3221 /* Ensure we clear previously set non-block flag */
3222 if (!force_nonblock)
3223 kiocb->ki_flags &= ~IOCB_NOWAIT;
3225 kiocb->ki_flags |= IOCB_NOWAIT;
3227 /* If the file doesn't support async, just async punt */
3228 if (force_nonblock && !io_file_supports_async(req->file, READ)) {
3229 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3230 return ret ?: -EAGAIN;
3233 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size);
3234 if (unlikely(ret)) {
3239 ret = io_iter_do_read(req, iter);
3241 if (ret == -EIOCBQUEUED) {
3242 if (req->async_data)
3243 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3245 } else if (ret == -EAGAIN) {
3246 /* IOPOLL retry should happen for io-wq threads */
3247 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3249 /* no retry on NONBLOCK nor RWF_NOWAIT */
3250 if (req->flags & REQ_F_NOWAIT)
3252 /* some cases will consume bytes even on error returns */
3253 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3255 } else if (ret <= 0 || ret == io_size || !force_nonblock ||
3256 (req->flags & REQ_F_NOWAIT) || !(req->flags & REQ_F_ISREG)) {
3257 /* read all, failed, already did sync or don't want to retry */
3261 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3266 rw = req->async_data;
3267 /* now use our persistent iterator, if we aren't already */
3272 rw->bytes_done += ret;
3273 /* if we can retry, do so with the callbacks armed */
3274 if (!io_rw_should_retry(req)) {
3275 kiocb->ki_flags &= ~IOCB_WAITQ;
3280 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3281 * we get -EIOCBQUEUED, then we'll get a notification when the
3282 * desired page gets unlocked. We can also get a partial read
3283 * here, and if we do, then just retry at the new offset.
3285 ret = io_iter_do_read(req, iter);
3286 if (ret == -EIOCBQUEUED)
3288 /* we got some bytes, but not all. retry. */
3289 kiocb->ki_flags &= ~IOCB_WAITQ;
3290 } while (ret > 0 && ret < io_size);
3292 kiocb_done(kiocb, ret, issue_flags);
3294 /* it's faster to check here then delegate to kfree */
3300 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3302 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3304 return io_prep_rw(req, sqe);
3307 static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3309 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3310 struct kiocb *kiocb = &req->rw.kiocb;
3311 struct iov_iter __iter, *iter = &__iter;
3312 struct io_async_rw *rw = req->async_data;
3313 ssize_t ret, ret2, io_size;
3314 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3320 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3324 io_size = iov_iter_count(iter);
3325 req->result = io_size;
3327 /* Ensure we clear previously set non-block flag */
3328 if (!force_nonblock)
3329 kiocb->ki_flags &= ~IOCB_NOWAIT;
3331 kiocb->ki_flags |= IOCB_NOWAIT;
3333 /* If the file doesn't support async, just async punt */
3334 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
3337 /* file path doesn't support NOWAIT for non-direct_IO */
3338 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3339 (req->flags & REQ_F_ISREG))
3342 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size);
3347 * Open-code file_start_write here to grab freeze protection,
3348 * which will be released by another thread in
3349 * io_complete_rw(). Fool lockdep by telling it the lock got
3350 * released so that it doesn't complain about the held lock when
3351 * we return to userspace.
3353 if (req->flags & REQ_F_ISREG) {
3354 sb_start_write(file_inode(req->file)->i_sb);
3355 __sb_writers_release(file_inode(req->file)->i_sb,
3358 kiocb->ki_flags |= IOCB_WRITE;
3360 if (req->file->f_op->write_iter)
3361 ret2 = call_write_iter(req->file, kiocb, iter);
3362 else if (req->file->f_op->write)
3363 ret2 = loop_rw_iter(WRITE, req, iter);
3368 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3369 * retry them without IOCB_NOWAIT.
3371 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3373 /* no retry on NONBLOCK nor RWF_NOWAIT */
3374 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3376 if (ret2 == -EIOCBQUEUED && req->async_data)
3377 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3378 if (!force_nonblock || ret2 != -EAGAIN) {
3379 /* IOPOLL retry should happen for io-wq threads */
3380 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3383 kiocb_done(kiocb, ret2, issue_flags);
3386 /* some cases will consume bytes even on error returns */
3387 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3388 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3389 return ret ?: -EAGAIN;
3392 /* it's reportedly faster than delegating the null check to kfree() */
3398 static int io_renameat_prep(struct io_kiocb *req,
3399 const struct io_uring_sqe *sqe)
3401 struct io_rename *ren = &req->rename;
3402 const char __user *oldf, *newf;
3404 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3407 ren->old_dfd = READ_ONCE(sqe->fd);
3408 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3409 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3410 ren->new_dfd = READ_ONCE(sqe->len);
3411 ren->flags = READ_ONCE(sqe->rename_flags);
3413 ren->oldpath = getname(oldf);
3414 if (IS_ERR(ren->oldpath))
3415 return PTR_ERR(ren->oldpath);
3417 ren->newpath = getname(newf);
3418 if (IS_ERR(ren->newpath)) {
3419 putname(ren->oldpath);
3420 return PTR_ERR(ren->newpath);
3423 req->flags |= REQ_F_NEED_CLEANUP;
3427 static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3429 struct io_rename *ren = &req->rename;
3432 if (issue_flags & IO_URING_F_NONBLOCK)
3435 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3436 ren->newpath, ren->flags);
3438 req->flags &= ~REQ_F_NEED_CLEANUP;
3440 req_set_fail_links(req);
3441 io_req_complete(req, ret);
3445 static int io_unlinkat_prep(struct io_kiocb *req,
3446 const struct io_uring_sqe *sqe)
3448 struct io_unlink *un = &req->unlink;
3449 const char __user *fname;
3451 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3454 un->dfd = READ_ONCE(sqe->fd);
3456 un->flags = READ_ONCE(sqe->unlink_flags);
3457 if (un->flags & ~AT_REMOVEDIR)
3460 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3461 un->filename = getname(fname);
3462 if (IS_ERR(un->filename))
3463 return PTR_ERR(un->filename);
3465 req->flags |= REQ_F_NEED_CLEANUP;
3469 static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3471 struct io_unlink *un = &req->unlink;
3474 if (issue_flags & IO_URING_F_NONBLOCK)
3477 if (un->flags & AT_REMOVEDIR)
3478 ret = do_rmdir(un->dfd, un->filename);
3480 ret = do_unlinkat(un->dfd, un->filename);
3482 req->flags &= ~REQ_F_NEED_CLEANUP;
3484 req_set_fail_links(req);
3485 io_req_complete(req, ret);
3489 static int io_shutdown_prep(struct io_kiocb *req,
3490 const struct io_uring_sqe *sqe)
3492 #if defined(CONFIG_NET)
3493 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3495 if (sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3499 req->shutdown.how = READ_ONCE(sqe->len);
3506 static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3508 #if defined(CONFIG_NET)
3509 struct socket *sock;
3512 if (issue_flags & IO_URING_F_NONBLOCK)
3515 sock = sock_from_file(req->file);
3516 if (unlikely(!sock))
3519 ret = __sys_shutdown_sock(sock, req->shutdown.how);
3521 req_set_fail_links(req);
3522 io_req_complete(req, ret);
3529 static int __io_splice_prep(struct io_kiocb *req,
3530 const struct io_uring_sqe *sqe)
3532 struct io_splice* sp = &req->splice;
3533 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3535 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3539 sp->len = READ_ONCE(sqe->len);
3540 sp->flags = READ_ONCE(sqe->splice_flags);
3542 if (unlikely(sp->flags & ~valid_flags))
3545 sp->file_in = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in),
3546 (sp->flags & SPLICE_F_FD_IN_FIXED));
3549 req->flags |= REQ_F_NEED_CLEANUP;
3551 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
3553 * Splice operation will be punted aync, and here need to
3554 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3556 req->work.flags |= IO_WQ_WORK_UNBOUND;
3562 static int io_tee_prep(struct io_kiocb *req,
3563 const struct io_uring_sqe *sqe)
3565 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3567 return __io_splice_prep(req, sqe);
3570 static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
3572 struct io_splice *sp = &req->splice;
3573 struct file *in = sp->file_in;
3574 struct file *out = sp->file_out;
3575 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3578 if (issue_flags & IO_URING_F_NONBLOCK)
3581 ret = do_tee(in, out, sp->len, flags);
3583 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3584 req->flags &= ~REQ_F_NEED_CLEANUP;
3587 req_set_fail_links(req);
3588 io_req_complete(req, ret);
3592 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3594 struct io_splice* sp = &req->splice;
3596 sp->off_in = READ_ONCE(sqe->splice_off_in);
3597 sp->off_out = READ_ONCE(sqe->off);
3598 return __io_splice_prep(req, sqe);
3601 static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
3603 struct io_splice *sp = &req->splice;
3604 struct file *in = sp->file_in;
3605 struct file *out = sp->file_out;
3606 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3607 loff_t *poff_in, *poff_out;
3610 if (issue_flags & IO_URING_F_NONBLOCK)
3613 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3614 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3617 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3619 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3620 req->flags &= ~REQ_F_NEED_CLEANUP;
3623 req_set_fail_links(req);
3624 io_req_complete(req, ret);
3629 * IORING_OP_NOP just posts a completion event, nothing else.
3631 static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
3633 struct io_ring_ctx *ctx = req->ctx;
3635 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3638 __io_req_complete(req, issue_flags, 0, 0);
3642 static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3644 struct io_ring_ctx *ctx = req->ctx;
3649 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3651 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3654 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3655 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3658 req->sync.off = READ_ONCE(sqe->off);
3659 req->sync.len = READ_ONCE(sqe->len);
3663 static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
3665 loff_t end = req->sync.off + req->sync.len;
3668 /* fsync always requires a blocking context */
3669 if (issue_flags & IO_URING_F_NONBLOCK)
3672 ret = vfs_fsync_range(req->file, req->sync.off,
3673 end > 0 ? end : LLONG_MAX,
3674 req->sync.flags & IORING_FSYNC_DATASYNC);
3676 req_set_fail_links(req);
3677 io_req_complete(req, ret);
3681 static int io_fallocate_prep(struct io_kiocb *req,
3682 const struct io_uring_sqe *sqe)
3684 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3686 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3689 req->sync.off = READ_ONCE(sqe->off);
3690 req->sync.len = READ_ONCE(sqe->addr);
3691 req->sync.mode = READ_ONCE(sqe->len);
3695 static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
3699 /* fallocate always requiring blocking context */
3700 if (issue_flags & IO_URING_F_NONBLOCK)
3702 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3705 req_set_fail_links(req);
3706 io_req_complete(req, ret);
3710 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3712 const char __user *fname;
3715 if (unlikely(sqe->ioprio || sqe->buf_index))
3717 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3720 /* open.how should be already initialised */
3721 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3722 req->open.how.flags |= O_LARGEFILE;
3724 req->open.dfd = READ_ONCE(sqe->fd);
3725 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3726 req->open.filename = getname(fname);
3727 if (IS_ERR(req->open.filename)) {
3728 ret = PTR_ERR(req->open.filename);
3729 req->open.filename = NULL;
3732 req->open.nofile = rlimit(RLIMIT_NOFILE);
3733 req->flags |= REQ_F_NEED_CLEANUP;
3737 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3741 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3743 mode = READ_ONCE(sqe->len);
3744 flags = READ_ONCE(sqe->open_flags);
3745 req->open.how = build_open_how(flags, mode);
3746 return __io_openat_prep(req, sqe);
3749 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3751 struct open_how __user *how;
3755 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3757 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3758 len = READ_ONCE(sqe->len);
3759 if (len < OPEN_HOW_SIZE_VER0)
3762 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3767 return __io_openat_prep(req, sqe);
3770 static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
3772 struct open_flags op;
3775 bool resolve_nonblock;
3778 ret = build_open_flags(&req->open.how, &op);
3781 nonblock_set = op.open_flag & O_NONBLOCK;
3782 resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
3783 if (issue_flags & IO_URING_F_NONBLOCK) {
3785 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
3786 * it'll always -EAGAIN
3788 if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
3790 op.lookup_flags |= LOOKUP_CACHED;
3791 op.open_flag |= O_NONBLOCK;
3794 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3798 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3799 /* only retry if RESOLVE_CACHED wasn't already set by application */
3800 if ((!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)) &&
3801 file == ERR_PTR(-EAGAIN)) {
3803 * We could hang on to this 'fd', but seems like marginal
3804 * gain for something that is now known to be a slower path.
3805 * So just put it, and we'll get a new one when we retry.
3813 ret = PTR_ERR(file);
3815 if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
3816 file->f_flags &= ~O_NONBLOCK;
3817 fsnotify_open(file);
3818 fd_install(ret, file);
3821 putname(req->open.filename);
3822 req->flags &= ~REQ_F_NEED_CLEANUP;
3824 req_set_fail_links(req);
3825 io_req_complete(req, ret);
3829 static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
3831 return io_openat2(req, issue_flags);
3834 static int io_remove_buffers_prep(struct io_kiocb *req,
3835 const struct io_uring_sqe *sqe)
3837 struct io_provide_buf *p = &req->pbuf;
3840 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3843 tmp = READ_ONCE(sqe->fd);
3844 if (!tmp || tmp > USHRT_MAX)
3847 memset(p, 0, sizeof(*p));
3849 p->bgid = READ_ONCE(sqe->buf_group);
3853 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3854 int bgid, unsigned nbufs)
3858 /* shouldn't happen */
3862 /* the head kbuf is the list itself */
3863 while (!list_empty(&buf->list)) {
3864 struct io_buffer *nxt;
3866 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3867 list_del(&nxt->list);
3874 idr_remove(&ctx->io_buffer_idr, bgid);
3879 static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
3881 struct io_provide_buf *p = &req->pbuf;
3882 struct io_ring_ctx *ctx = req->ctx;
3883 struct io_buffer *head;
3885 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3887 io_ring_submit_lock(ctx, !force_nonblock);
3889 lockdep_assert_held(&ctx->uring_lock);
3892 head = idr_find(&ctx->io_buffer_idr, p->bgid);
3894 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3896 req_set_fail_links(req);
3898 /* need to hold the lock to complete IOPOLL requests */
3899 if (ctx->flags & IORING_SETUP_IOPOLL) {
3900 __io_req_complete(req, issue_flags, ret, 0);
3901 io_ring_submit_unlock(ctx, !force_nonblock);
3903 io_ring_submit_unlock(ctx, !force_nonblock);
3904 __io_req_complete(req, issue_flags, ret, 0);
3909 static int io_provide_buffers_prep(struct io_kiocb *req,
3910 const struct io_uring_sqe *sqe)
3912 struct io_provide_buf *p = &req->pbuf;
3915 if (sqe->ioprio || sqe->rw_flags)
3918 tmp = READ_ONCE(sqe->fd);
3919 if (!tmp || tmp > USHRT_MAX)
3922 p->addr = READ_ONCE(sqe->addr);
3923 p->len = READ_ONCE(sqe->len);
3925 if (!access_ok(u64_to_user_ptr(p->addr), (p->len * p->nbufs)))
3928 p->bgid = READ_ONCE(sqe->buf_group);
3929 tmp = READ_ONCE(sqe->off);
3930 if (tmp > USHRT_MAX)
3936 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3938 struct io_buffer *buf;
3939 u64 addr = pbuf->addr;
3940 int i, bid = pbuf->bid;
3942 for (i = 0; i < pbuf->nbufs; i++) {
3943 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
3948 buf->len = pbuf->len;
3953 INIT_LIST_HEAD(&buf->list);
3956 list_add_tail(&buf->list, &(*head)->list);
3960 return i ? i : -ENOMEM;
3963 static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
3965 struct io_provide_buf *p = &req->pbuf;
3966 struct io_ring_ctx *ctx = req->ctx;
3967 struct io_buffer *head, *list;
3969 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3971 io_ring_submit_lock(ctx, !force_nonblock);
3973 lockdep_assert_held(&ctx->uring_lock);
3975 list = head = idr_find(&ctx->io_buffer_idr, p->bgid);
3977 ret = io_add_buffers(p, &head);
3982 ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1,
3985 __io_remove_buffers(ctx, head, p->bgid, -1U);
3991 req_set_fail_links(req);
3993 /* need to hold the lock to complete IOPOLL requests */
3994 if (ctx->flags & IORING_SETUP_IOPOLL) {
3995 __io_req_complete(req, issue_flags, ret, 0);
3996 io_ring_submit_unlock(ctx, !force_nonblock);
3998 io_ring_submit_unlock(ctx, !force_nonblock);
3999 __io_req_complete(req, issue_flags, ret, 0);
4004 static int io_epoll_ctl_prep(struct io_kiocb *req,
4005 const struct io_uring_sqe *sqe)
4007 #if defined(CONFIG_EPOLL)
4008 if (sqe->ioprio || sqe->buf_index)
4010 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4013 req->epoll.epfd = READ_ONCE(sqe->fd);
4014 req->epoll.op = READ_ONCE(sqe->len);
4015 req->epoll.fd = READ_ONCE(sqe->off);
4017 if (ep_op_has_event(req->epoll.op)) {
4018 struct epoll_event __user *ev;
4020 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4021 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4031 static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4033 #if defined(CONFIG_EPOLL)
4034 struct io_epoll *ie = &req->epoll;
4036 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4038 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4039 if (force_nonblock && ret == -EAGAIN)
4043 req_set_fail_links(req);
4044 __io_req_complete(req, issue_flags, ret, 0);
4051 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4053 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4054 if (sqe->ioprio || sqe->buf_index || sqe->off)
4056 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4059 req->madvise.addr = READ_ONCE(sqe->addr);
4060 req->madvise.len = READ_ONCE(sqe->len);
4061 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4068 static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4070 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4071 struct io_madvise *ma = &req->madvise;
4074 if (issue_flags & IO_URING_F_NONBLOCK)
4077 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4079 req_set_fail_links(req);
4080 io_req_complete(req, ret);
4087 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4089 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4091 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4094 req->fadvise.offset = READ_ONCE(sqe->off);
4095 req->fadvise.len = READ_ONCE(sqe->len);
4096 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4100 static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4102 struct io_fadvise *fa = &req->fadvise;
4105 if (issue_flags & IO_URING_F_NONBLOCK) {
4106 switch (fa->advice) {
4107 case POSIX_FADV_NORMAL:
4108 case POSIX_FADV_RANDOM:
4109 case POSIX_FADV_SEQUENTIAL:
4116 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4118 req_set_fail_links(req);
4119 io_req_complete(req, ret);
4123 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4125 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4127 if (sqe->ioprio || sqe->buf_index)
4129 if (req->flags & REQ_F_FIXED_FILE)
4132 req->statx.dfd = READ_ONCE(sqe->fd);
4133 req->statx.mask = READ_ONCE(sqe->len);
4134 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4135 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4136 req->statx.flags = READ_ONCE(sqe->statx_flags);
4141 static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4143 struct io_statx *ctx = &req->statx;
4146 if (issue_flags & IO_URING_F_NONBLOCK) {
4147 /* only need file table for an actual valid fd */
4148 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
4149 req->flags |= REQ_F_NO_FILE_TABLE;
4153 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4157 req_set_fail_links(req);
4158 io_req_complete(req, ret);
4162 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4164 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4166 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4167 sqe->rw_flags || sqe->buf_index)
4169 if (req->flags & REQ_F_FIXED_FILE)
4172 req->close.fd = READ_ONCE(sqe->fd);
4176 static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4178 struct files_struct *files = current->files;
4179 struct io_close *close = &req->close;
4180 struct fdtable *fdt;
4186 spin_lock(&files->file_lock);
4187 fdt = files_fdtable(files);
4188 if (close->fd >= fdt->max_fds) {
4189 spin_unlock(&files->file_lock);
4192 file = fdt->fd[close->fd];
4194 spin_unlock(&files->file_lock);
4198 if (file->f_op == &io_uring_fops) {
4199 spin_unlock(&files->file_lock);
4204 /* if the file has a flush method, be safe and punt to async */
4205 if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4206 spin_unlock(&files->file_lock);
4210 ret = __close_fd_get_file(close->fd, &file);
4211 spin_unlock(&files->file_lock);
4218 /* No ->flush() or already async, safely close from here */
4219 ret = filp_close(file, current->files);
4222 req_set_fail_links(req);
4225 __io_req_complete(req, issue_flags, ret, 0);
4229 static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4231 struct io_ring_ctx *ctx = req->ctx;
4233 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4235 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4238 req->sync.off = READ_ONCE(sqe->off);
4239 req->sync.len = READ_ONCE(sqe->len);
4240 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4244 static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4248 /* sync_file_range always requires a blocking context */
4249 if (issue_flags & IO_URING_F_NONBLOCK)
4252 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4255 req_set_fail_links(req);
4256 io_req_complete(req, ret);
4260 #if defined(CONFIG_NET)
4261 static int io_setup_async_msg(struct io_kiocb *req,
4262 struct io_async_msghdr *kmsg)
4264 struct io_async_msghdr *async_msg = req->async_data;
4268 if (io_alloc_async_data(req)) {
4269 kfree(kmsg->free_iov);
4272 async_msg = req->async_data;
4273 req->flags |= REQ_F_NEED_CLEANUP;
4274 memcpy(async_msg, kmsg, sizeof(*kmsg));
4275 async_msg->msg.msg_name = &async_msg->addr;
4276 /* if were using fast_iov, set it to the new one */
4277 if (!async_msg->free_iov)
4278 async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4283 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4284 struct io_async_msghdr *iomsg)
4286 iomsg->msg.msg_name = &iomsg->addr;
4287 iomsg->free_iov = iomsg->fast_iov;
4288 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4289 req->sr_msg.msg_flags, &iomsg->free_iov);
4292 static int io_sendmsg_prep_async(struct io_kiocb *req)
4296 if (!io_op_defs[req->opcode].needs_async_data)
4298 ret = io_sendmsg_copy_hdr(req, req->async_data);
4300 req->flags |= REQ_F_NEED_CLEANUP;
4304 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4306 struct io_sr_msg *sr = &req->sr_msg;
4308 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4311 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4312 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4313 sr->len = READ_ONCE(sqe->len);
4315 #ifdef CONFIG_COMPAT
4316 if (req->ctx->compat)
4317 sr->msg_flags |= MSG_CMSG_COMPAT;
4322 static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4324 struct io_async_msghdr iomsg, *kmsg;
4325 struct socket *sock;
4329 sock = sock_from_file(req->file);
4330 if (unlikely(!sock))
4333 kmsg = req->async_data;
4335 ret = io_sendmsg_copy_hdr(req, &iomsg);
4341 flags = req->sr_msg.msg_flags;
4342 if (flags & MSG_DONTWAIT)
4343 req->flags |= REQ_F_NOWAIT;
4344 else if (issue_flags & IO_URING_F_NONBLOCK)
4345 flags |= MSG_DONTWAIT;
4347 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4348 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4349 return io_setup_async_msg(req, kmsg);
4350 if (ret == -ERESTARTSYS)
4353 /* fast path, check for non-NULL to avoid function call */
4355 kfree(kmsg->free_iov);
4356 req->flags &= ~REQ_F_NEED_CLEANUP;
4358 req_set_fail_links(req);
4359 __io_req_complete(req, issue_flags, ret, 0);
4363 static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4365 struct io_sr_msg *sr = &req->sr_msg;
4368 struct socket *sock;
4372 sock = sock_from_file(req->file);
4373 if (unlikely(!sock))
4376 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4380 msg.msg_name = NULL;
4381 msg.msg_control = NULL;
4382 msg.msg_controllen = 0;
4383 msg.msg_namelen = 0;
4385 flags = req->sr_msg.msg_flags;
4386 if (flags & MSG_DONTWAIT)
4387 req->flags |= REQ_F_NOWAIT;
4388 else if (issue_flags & IO_URING_F_NONBLOCK)
4389 flags |= MSG_DONTWAIT;
4391 msg.msg_flags = flags;
4392 ret = sock_sendmsg(sock, &msg);
4393 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4395 if (ret == -ERESTARTSYS)
4399 req_set_fail_links(req);
4400 __io_req_complete(req, issue_flags, ret, 0);
4404 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4405 struct io_async_msghdr *iomsg)
4407 struct io_sr_msg *sr = &req->sr_msg;
4408 struct iovec __user *uiov;
4412 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4413 &iomsg->uaddr, &uiov, &iov_len);
4417 if (req->flags & REQ_F_BUFFER_SELECT) {
4420 if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4422 sr->len = iomsg->fast_iov[0].iov_len;
4423 iomsg->free_iov = NULL;
4425 iomsg->free_iov = iomsg->fast_iov;
4426 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4427 &iomsg->free_iov, &iomsg->msg.msg_iter,
4436 #ifdef CONFIG_COMPAT
4437 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4438 struct io_async_msghdr *iomsg)
4440 struct compat_msghdr __user *msg_compat;
4441 struct io_sr_msg *sr = &req->sr_msg;
4442 struct compat_iovec __user *uiov;
4447 msg_compat = (struct compat_msghdr __user *) sr->umsg;
4448 ret = __get_compat_msghdr(&iomsg->msg, msg_compat, &iomsg->uaddr,
4453 uiov = compat_ptr(ptr);
4454 if (req->flags & REQ_F_BUFFER_SELECT) {
4455 compat_ssize_t clen;
4459 if (!access_ok(uiov, sizeof(*uiov)))
4461 if (__get_user(clen, &uiov->iov_len))
4466 iomsg->free_iov = NULL;
4468 iomsg->free_iov = iomsg->fast_iov;
4469 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4470 UIO_FASTIOV, &iomsg->free_iov,
4471 &iomsg->msg.msg_iter, true);
4480 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4481 struct io_async_msghdr *iomsg)
4483 iomsg->msg.msg_name = &iomsg->addr;
4485 #ifdef CONFIG_COMPAT
4486 if (req->ctx->compat)
4487 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4490 return __io_recvmsg_copy_hdr(req, iomsg);
4493 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4496 struct io_sr_msg *sr = &req->sr_msg;
4497 struct io_buffer *kbuf;
4499 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4504 req->flags |= REQ_F_BUFFER_SELECTED;
4508 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4510 return io_put_kbuf(req, req->sr_msg.kbuf);
4513 static int io_recvmsg_prep_async(struct io_kiocb *req)
4517 if (!io_op_defs[req->opcode].needs_async_data)
4519 ret = io_recvmsg_copy_hdr(req, req->async_data);
4521 req->flags |= REQ_F_NEED_CLEANUP;
4525 static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4527 struct io_sr_msg *sr = &req->sr_msg;
4529 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4532 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4533 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4534 sr->len = READ_ONCE(sqe->len);
4535 sr->bgid = READ_ONCE(sqe->buf_group);
4537 #ifdef CONFIG_COMPAT
4538 if (req->ctx->compat)
4539 sr->msg_flags |= MSG_CMSG_COMPAT;
4544 static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
4546 struct io_async_msghdr iomsg, *kmsg;
4547 struct socket *sock;
4548 struct io_buffer *kbuf;
4550 int ret, cflags = 0;
4551 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4553 sock = sock_from_file(req->file);
4554 if (unlikely(!sock))
4557 kmsg = req->async_data;
4559 ret = io_recvmsg_copy_hdr(req, &iomsg);
4565 if (req->flags & REQ_F_BUFFER_SELECT) {
4566 kbuf = io_recv_buffer_select(req, !force_nonblock);
4568 return PTR_ERR(kbuf);
4569 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4570 kmsg->fast_iov[0].iov_len = req->sr_msg.len;
4571 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
4572 1, req->sr_msg.len);
4575 flags = req->sr_msg.msg_flags;
4576 if (flags & MSG_DONTWAIT)
4577 req->flags |= REQ_F_NOWAIT;
4578 else if (force_nonblock)
4579 flags |= MSG_DONTWAIT;
4581 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4582 kmsg->uaddr, flags);
4583 if (force_nonblock && ret == -EAGAIN)
4584 return io_setup_async_msg(req, kmsg);
4585 if (ret == -ERESTARTSYS)
4588 if (req->flags & REQ_F_BUFFER_SELECTED)
4589 cflags = io_put_recv_kbuf(req);
4590 /* fast path, check for non-NULL to avoid function call */
4592 kfree(kmsg->free_iov);
4593 req->flags &= ~REQ_F_NEED_CLEANUP;
4595 req_set_fail_links(req);
4596 __io_req_complete(req, issue_flags, ret, cflags);
4600 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
4602 struct io_buffer *kbuf;
4603 struct io_sr_msg *sr = &req->sr_msg;
4605 void __user *buf = sr->buf;
4606 struct socket *sock;
4609 int ret, cflags = 0;
4610 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4612 sock = sock_from_file(req->file);
4613 if (unlikely(!sock))
4616 if (req->flags & REQ_F_BUFFER_SELECT) {
4617 kbuf = io_recv_buffer_select(req, !force_nonblock);
4619 return PTR_ERR(kbuf);
4620 buf = u64_to_user_ptr(kbuf->addr);
4623 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4627 msg.msg_name = NULL;
4628 msg.msg_control = NULL;
4629 msg.msg_controllen = 0;
4630 msg.msg_namelen = 0;
4631 msg.msg_iocb = NULL;
4634 flags = req->sr_msg.msg_flags;
4635 if (flags & MSG_DONTWAIT)
4636 req->flags |= REQ_F_NOWAIT;
4637 else if (force_nonblock)
4638 flags |= MSG_DONTWAIT;
4640 ret = sock_recvmsg(sock, &msg, flags);
4641 if (force_nonblock && ret == -EAGAIN)
4643 if (ret == -ERESTARTSYS)
4646 if (req->flags & REQ_F_BUFFER_SELECTED)
4647 cflags = io_put_recv_kbuf(req);
4649 req_set_fail_links(req);
4650 __io_req_complete(req, issue_flags, ret, cflags);
4654 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4656 struct io_accept *accept = &req->accept;
4658 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4660 if (sqe->ioprio || sqe->len || sqe->buf_index)
4663 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4664 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4665 accept->flags = READ_ONCE(sqe->accept_flags);
4666 accept->nofile = rlimit(RLIMIT_NOFILE);
4670 static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
4672 struct io_accept *accept = &req->accept;
4673 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4674 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4677 if (req->file->f_flags & O_NONBLOCK)
4678 req->flags |= REQ_F_NOWAIT;
4680 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4681 accept->addr_len, accept->flags,
4683 if (ret == -EAGAIN && force_nonblock)
4686 if (ret == -ERESTARTSYS)
4688 req_set_fail_links(req);
4690 __io_req_complete(req, issue_flags, ret, 0);
4694 static int io_connect_prep_async(struct io_kiocb *req)
4696 struct io_async_connect *io = req->async_data;
4697 struct io_connect *conn = &req->connect;
4699 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
4702 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4704 struct io_connect *conn = &req->connect;
4706 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4708 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4711 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4712 conn->addr_len = READ_ONCE(sqe->addr2);
4716 static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
4718 struct io_async_connect __io, *io;
4719 unsigned file_flags;
4721 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4723 if (req->async_data) {
4724 io = req->async_data;
4726 ret = move_addr_to_kernel(req->connect.addr,
4727 req->connect.addr_len,
4734 file_flags = force_nonblock ? O_NONBLOCK : 0;
4736 ret = __sys_connect_file(req->file, &io->address,
4737 req->connect.addr_len, file_flags);
4738 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4739 if (req->async_data)
4741 if (io_alloc_async_data(req)) {
4745 io = req->async_data;
4746 memcpy(req->async_data, &__io, sizeof(__io));
4749 if (ret == -ERESTARTSYS)
4753 req_set_fail_links(req);
4754 __io_req_complete(req, issue_flags, ret, 0);
4757 #else /* !CONFIG_NET */
4758 #define IO_NETOP_FN(op) \
4759 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
4761 return -EOPNOTSUPP; \
4764 #define IO_NETOP_PREP(op) \
4766 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
4768 return -EOPNOTSUPP; \
4771 #define IO_NETOP_PREP_ASYNC(op) \
4773 static int io_##op##_prep_async(struct io_kiocb *req) \
4775 return -EOPNOTSUPP; \
4778 IO_NETOP_PREP_ASYNC(sendmsg);
4779 IO_NETOP_PREP_ASYNC(recvmsg);
4780 IO_NETOP_PREP_ASYNC(connect);
4781 IO_NETOP_PREP(accept);
4784 #endif /* CONFIG_NET */
4786 struct io_poll_table {
4787 struct poll_table_struct pt;
4788 struct io_kiocb *req;
4792 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4793 __poll_t mask, task_work_func_t func)
4797 /* for instances that support it check for an event match first: */
4798 if (mask && !(mask & poll->events))
4801 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4803 list_del_init(&poll->wait.entry);
4806 req->task_work.func = func;
4807 percpu_ref_get(&req->ctx->refs);
4810 * If this fails, then the task is exiting. When a task exits, the
4811 * work gets canceled, so just cancel this request as well instead
4812 * of executing it. We can't safely execute it anyway, as we may not
4813 * have the needed state needed for it anyway.
4815 ret = io_req_task_work_add(req);
4816 if (unlikely(ret)) {
4817 WRITE_ONCE(poll->canceled, true);
4818 io_req_task_work_add_fallback(req, func);
4823 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4824 __acquires(&req->ctx->completion_lock)
4826 struct io_ring_ctx *ctx = req->ctx;
4828 if (!req->result && !READ_ONCE(poll->canceled)) {
4829 struct poll_table_struct pt = { ._key = poll->events };
4831 req->result = vfs_poll(req->file, &pt) & poll->events;
4834 spin_lock_irq(&ctx->completion_lock);
4835 if (!req->result && !READ_ONCE(poll->canceled)) {
4836 add_wait_queue(poll->head, &poll->wait);
4843 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4845 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4846 if (req->opcode == IORING_OP_POLL_ADD)
4847 return req->async_data;
4848 return req->apoll->double_poll;
4851 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4853 if (req->opcode == IORING_OP_POLL_ADD)
4855 return &req->apoll->poll;
4858 static void io_poll_remove_double(struct io_kiocb *req)
4860 struct io_poll_iocb *poll = io_poll_get_double(req);
4862 lockdep_assert_held(&req->ctx->completion_lock);
4864 if (poll && poll->head) {
4865 struct wait_queue_head *head = poll->head;
4867 spin_lock(&head->lock);
4868 list_del_init(&poll->wait.entry);
4869 if (poll->wait.private)
4870 refcount_dec(&req->refs);
4872 spin_unlock(&head->lock);
4876 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
4878 struct io_ring_ctx *ctx = req->ctx;
4880 io_poll_remove_double(req);
4881 req->poll.done = true;
4882 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
4883 io_commit_cqring(ctx);
4886 static void io_poll_task_func(struct callback_head *cb)
4888 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4889 struct io_ring_ctx *ctx = req->ctx;
4890 struct io_kiocb *nxt;
4892 if (io_poll_rewait(req, &req->poll)) {
4893 spin_unlock_irq(&ctx->completion_lock);
4895 hash_del(&req->hash_node);
4896 io_poll_complete(req, req->result, 0);
4897 spin_unlock_irq(&ctx->completion_lock);
4899 nxt = io_put_req_find_next(req);
4900 io_cqring_ev_posted(ctx);
4902 __io_req_task_submit(nxt);
4905 percpu_ref_put(&ctx->refs);
4908 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4909 int sync, void *key)
4911 struct io_kiocb *req = wait->private;
4912 struct io_poll_iocb *poll = io_poll_get_single(req);
4913 __poll_t mask = key_to_poll(key);
4915 /* for instances that support it check for an event match first: */
4916 if (mask && !(mask & poll->events))
4919 list_del_init(&wait->entry);
4921 if (poll && poll->head) {
4924 spin_lock(&poll->head->lock);
4925 done = list_empty(&poll->wait.entry);
4927 list_del_init(&poll->wait.entry);
4928 /* make sure double remove sees this as being gone */
4929 wait->private = NULL;
4930 spin_unlock(&poll->head->lock);
4932 /* use wait func handler, so it matches the rq type */
4933 poll->wait.func(&poll->wait, mode, sync, key);
4936 refcount_dec(&req->refs);
4940 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
4941 wait_queue_func_t wake_func)
4945 poll->canceled = false;
4946 poll->events = events;
4947 INIT_LIST_HEAD(&poll->wait.entry);
4948 init_waitqueue_func_entry(&poll->wait, wake_func);
4951 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
4952 struct wait_queue_head *head,
4953 struct io_poll_iocb **poll_ptr)
4955 struct io_kiocb *req = pt->req;
4958 * If poll->head is already set, it's because the file being polled
4959 * uses multiple waitqueues for poll handling (eg one for read, one
4960 * for write). Setup a separate io_poll_iocb if this happens.
4962 if (unlikely(poll->head)) {
4963 struct io_poll_iocb *poll_one = poll;
4965 /* already have a 2nd entry, fail a third attempt */
4967 pt->error = -EINVAL;
4970 /* double add on the same waitqueue head, ignore */
4971 if (poll->head == head)
4973 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
4975 pt->error = -ENOMEM;
4978 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
4979 refcount_inc(&req->refs);
4980 poll->wait.private = req;
4987 if (poll->events & EPOLLEXCLUSIVE)
4988 add_wait_queue_exclusive(head, &poll->wait);
4990 add_wait_queue(head, &poll->wait);
4993 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
4994 struct poll_table_struct *p)
4996 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
4997 struct async_poll *apoll = pt->req->apoll;
4999 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5002 static void io_async_task_func(struct callback_head *cb)
5004 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
5005 struct async_poll *apoll = req->apoll;
5006 struct io_ring_ctx *ctx = req->ctx;
5008 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
5010 if (io_poll_rewait(req, &apoll->poll)) {
5011 spin_unlock_irq(&ctx->completion_lock);
5012 percpu_ref_put(&ctx->refs);
5016 /* If req is still hashed, it cannot have been canceled. Don't check. */
5017 if (hash_hashed(&req->hash_node))
5018 hash_del(&req->hash_node);
5020 io_poll_remove_double(req);
5021 spin_unlock_irq(&ctx->completion_lock);
5023 if (!READ_ONCE(apoll->poll.canceled))
5024 __io_req_task_submit(req);
5026 __io_req_task_cancel(req, -ECANCELED);
5028 percpu_ref_put(&ctx->refs);
5029 kfree(apoll->double_poll);
5033 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5036 struct io_kiocb *req = wait->private;
5037 struct io_poll_iocb *poll = &req->apoll->poll;
5039 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5042 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5045 static void io_poll_req_insert(struct io_kiocb *req)
5047 struct io_ring_ctx *ctx = req->ctx;
5048 struct hlist_head *list;
5050 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5051 hlist_add_head(&req->hash_node, list);
5054 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5055 struct io_poll_iocb *poll,
5056 struct io_poll_table *ipt, __poll_t mask,
5057 wait_queue_func_t wake_func)
5058 __acquires(&ctx->completion_lock)
5060 struct io_ring_ctx *ctx = req->ctx;
5061 bool cancel = false;
5063 INIT_HLIST_NODE(&req->hash_node);
5064 io_init_poll_iocb(poll, mask, wake_func);
5065 poll->file = req->file;
5066 poll->wait.private = req;
5068 ipt->pt._key = mask;
5070 ipt->error = -EINVAL;
5072 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5074 spin_lock_irq(&ctx->completion_lock);
5075 if (likely(poll->head)) {
5076 spin_lock(&poll->head->lock);
5077 if (unlikely(list_empty(&poll->wait.entry))) {
5083 if (mask || ipt->error)
5084 list_del_init(&poll->wait.entry);
5086 WRITE_ONCE(poll->canceled, true);
5087 else if (!poll->done) /* actually waiting for an event */
5088 io_poll_req_insert(req);
5089 spin_unlock(&poll->head->lock);
5095 static bool io_arm_poll_handler(struct io_kiocb *req)
5097 const struct io_op_def *def = &io_op_defs[req->opcode];
5098 struct io_ring_ctx *ctx = req->ctx;
5099 struct async_poll *apoll;
5100 struct io_poll_table ipt;
5104 if (!req->file || !file_can_poll(req->file))
5106 if (req->flags & REQ_F_POLLED)
5110 else if (def->pollout)
5114 /* if we can't nonblock try, then no point in arming a poll handler */
5115 if (!io_file_supports_async(req->file, rw))
5118 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5119 if (unlikely(!apoll))
5121 apoll->double_poll = NULL;
5123 req->flags |= REQ_F_POLLED;
5128 mask |= POLLIN | POLLRDNORM;
5130 mask |= POLLOUT | POLLWRNORM;
5132 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5133 if ((req->opcode == IORING_OP_RECVMSG) &&
5134 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5137 mask |= POLLERR | POLLPRI;
5139 ipt.pt._qproc = io_async_queue_proc;
5141 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5143 if (ret || ipt.error) {
5144 io_poll_remove_double(req);
5145 spin_unlock_irq(&ctx->completion_lock);
5146 kfree(apoll->double_poll);
5150 spin_unlock_irq(&ctx->completion_lock);
5151 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
5152 apoll->poll.events);
5156 static bool __io_poll_remove_one(struct io_kiocb *req,
5157 struct io_poll_iocb *poll)
5159 bool do_complete = false;
5161 spin_lock(&poll->head->lock);
5162 WRITE_ONCE(poll->canceled, true);
5163 if (!list_empty(&poll->wait.entry)) {
5164 list_del_init(&poll->wait.entry);
5167 spin_unlock(&poll->head->lock);
5168 hash_del(&req->hash_node);
5172 static bool io_poll_remove_one(struct io_kiocb *req)
5176 io_poll_remove_double(req);
5178 if (req->opcode == IORING_OP_POLL_ADD) {
5179 do_complete = __io_poll_remove_one(req, &req->poll);
5181 struct async_poll *apoll = req->apoll;
5183 /* non-poll requests have submit ref still */
5184 do_complete = __io_poll_remove_one(req, &apoll->poll);
5187 kfree(apoll->double_poll);
5193 io_cqring_fill_event(req, -ECANCELED);
5194 io_commit_cqring(req->ctx);
5195 req_set_fail_links(req);
5196 io_put_req_deferred(req, 1);
5203 * Returns true if we found and killed one or more poll requests
5205 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5206 struct files_struct *files)
5208 struct hlist_node *tmp;
5209 struct io_kiocb *req;
5212 spin_lock_irq(&ctx->completion_lock);
5213 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5214 struct hlist_head *list;
5216 list = &ctx->cancel_hash[i];
5217 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5218 if (io_match_task(req, tsk, files))
5219 posted += io_poll_remove_one(req);
5222 spin_unlock_irq(&ctx->completion_lock);
5225 io_cqring_ev_posted(ctx);
5230 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
5232 struct hlist_head *list;
5233 struct io_kiocb *req;
5235 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5236 hlist_for_each_entry(req, list, hash_node) {
5237 if (sqe_addr != req->user_data)
5239 if (io_poll_remove_one(req))
5247 static int io_poll_remove_prep(struct io_kiocb *req,
5248 const struct io_uring_sqe *sqe)
5250 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5252 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
5256 req->poll_remove.addr = READ_ONCE(sqe->addr);
5261 * Find a running poll command that matches one specified in sqe->addr,
5262 * and remove it if found.
5264 static int io_poll_remove(struct io_kiocb *req, unsigned int issue_flags)
5266 struct io_ring_ctx *ctx = req->ctx;
5269 spin_lock_irq(&ctx->completion_lock);
5270 ret = io_poll_cancel(ctx, req->poll_remove.addr);
5271 spin_unlock_irq(&ctx->completion_lock);
5274 req_set_fail_links(req);
5275 io_req_complete(req, ret);
5279 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5282 struct io_kiocb *req = wait->private;
5283 struct io_poll_iocb *poll = &req->poll;
5285 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5288 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5289 struct poll_table_struct *p)
5291 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5293 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5296 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5298 struct io_poll_iocb *poll = &req->poll;
5301 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5303 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
5306 events = READ_ONCE(sqe->poll32_events);
5308 events = swahw32(events);
5310 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP |
5311 (events & EPOLLEXCLUSIVE);
5315 static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5317 struct io_poll_iocb *poll = &req->poll;
5318 struct io_ring_ctx *ctx = req->ctx;
5319 struct io_poll_table ipt;
5322 ipt.pt._qproc = io_poll_queue_proc;
5324 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5327 if (mask) { /* no async, we'd stolen it */
5329 io_poll_complete(req, mask, 0);
5331 spin_unlock_irq(&ctx->completion_lock);
5334 io_cqring_ev_posted(ctx);
5340 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5342 struct io_timeout_data *data = container_of(timer,
5343 struct io_timeout_data, timer);
5344 struct io_kiocb *req = data->req;
5345 struct io_ring_ctx *ctx = req->ctx;
5346 unsigned long flags;
5348 spin_lock_irqsave(&ctx->completion_lock, flags);
5349 list_del_init(&req->timeout.list);
5350 atomic_set(&req->ctx->cq_timeouts,
5351 atomic_read(&req->ctx->cq_timeouts) + 1);
5353 io_cqring_fill_event(req, -ETIME);
5354 io_commit_cqring(ctx);
5355 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5357 io_cqring_ev_posted(ctx);
5358 req_set_fail_links(req);
5360 return HRTIMER_NORESTART;
5363 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5366 struct io_timeout_data *io;
5367 struct io_kiocb *req;
5370 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5371 if (user_data == req->user_data) {
5378 return ERR_PTR(ret);
5380 io = req->async_data;
5381 ret = hrtimer_try_to_cancel(&io->timer);
5383 return ERR_PTR(-EALREADY);
5384 list_del_init(&req->timeout.list);
5388 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5390 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5393 return PTR_ERR(req);
5395 req_set_fail_links(req);
5396 io_cqring_fill_event(req, -ECANCELED);
5397 io_put_req_deferred(req, 1);
5401 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5402 struct timespec64 *ts, enum hrtimer_mode mode)
5404 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5405 struct io_timeout_data *data;
5408 return PTR_ERR(req);
5410 req->timeout.off = 0; /* noseq */
5411 data = req->async_data;
5412 list_add_tail(&req->timeout.list, &ctx->timeout_list);
5413 hrtimer_init(&data->timer, CLOCK_MONOTONIC, mode);
5414 data->timer.function = io_timeout_fn;
5415 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
5419 static int io_timeout_remove_prep(struct io_kiocb *req,
5420 const struct io_uring_sqe *sqe)
5422 struct io_timeout_rem *tr = &req->timeout_rem;
5424 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5426 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5428 if (sqe->ioprio || sqe->buf_index || sqe->len)
5431 tr->addr = READ_ONCE(sqe->addr);
5432 tr->flags = READ_ONCE(sqe->timeout_flags);
5433 if (tr->flags & IORING_TIMEOUT_UPDATE) {
5434 if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS))
5436 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
5438 } else if (tr->flags) {
5439 /* timeout removal doesn't support flags */
5446 static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
5448 return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
5453 * Remove or update an existing timeout command
5455 static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
5457 struct io_timeout_rem *tr = &req->timeout_rem;
5458 struct io_ring_ctx *ctx = req->ctx;
5461 spin_lock_irq(&ctx->completion_lock);
5462 if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE))
5463 ret = io_timeout_cancel(ctx, tr->addr);
5465 ret = io_timeout_update(ctx, tr->addr, &tr->ts,
5466 io_translate_timeout_mode(tr->flags));
5468 io_cqring_fill_event(req, ret);
5469 io_commit_cqring(ctx);
5470 spin_unlock_irq(&ctx->completion_lock);
5471 io_cqring_ev_posted(ctx);
5473 req_set_fail_links(req);
5478 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5479 bool is_timeout_link)
5481 struct io_timeout_data *data;
5483 u32 off = READ_ONCE(sqe->off);
5485 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5487 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5489 if (off && is_timeout_link)
5491 flags = READ_ONCE(sqe->timeout_flags);
5492 if (flags & ~IORING_TIMEOUT_ABS)
5495 req->timeout.off = off;
5497 if (!req->async_data && io_alloc_async_data(req))
5500 data = req->async_data;
5503 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5506 data->mode = io_translate_timeout_mode(flags);
5507 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5508 io_req_track_inflight(req);
5512 static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
5514 struct io_ring_ctx *ctx = req->ctx;
5515 struct io_timeout_data *data = req->async_data;
5516 struct list_head *entry;
5517 u32 tail, off = req->timeout.off;
5519 spin_lock_irq(&ctx->completion_lock);
5522 * sqe->off holds how many events that need to occur for this
5523 * timeout event to be satisfied. If it isn't set, then this is
5524 * a pure timeout request, sequence isn't used.
5526 if (io_is_timeout_noseq(req)) {
5527 entry = ctx->timeout_list.prev;
5531 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5532 req->timeout.target_seq = tail + off;
5534 /* Update the last seq here in case io_flush_timeouts() hasn't.
5535 * This is safe because ->completion_lock is held, and submissions
5536 * and completions are never mixed in the same ->completion_lock section.
5538 ctx->cq_last_tm_flush = tail;
5541 * Insertion sort, ensuring the first entry in the list is always
5542 * the one we need first.
5544 list_for_each_prev(entry, &ctx->timeout_list) {
5545 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5548 if (io_is_timeout_noseq(nxt))
5550 /* nxt.seq is behind @tail, otherwise would've been completed */
5551 if (off >= nxt->timeout.target_seq - tail)
5555 list_add(&req->timeout.list, entry);
5556 data->timer.function = io_timeout_fn;
5557 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5558 spin_unlock_irq(&ctx->completion_lock);
5562 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5564 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5566 return req->user_data == (unsigned long) data;
5569 static int io_async_cancel_one(struct io_uring_task *tctx, void *sqe_addr)
5571 enum io_wq_cancel cancel_ret;
5577 cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, sqe_addr, false);
5578 switch (cancel_ret) {
5579 case IO_WQ_CANCEL_OK:
5582 case IO_WQ_CANCEL_RUNNING:
5585 case IO_WQ_CANCEL_NOTFOUND:
5593 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5594 struct io_kiocb *req, __u64 sqe_addr,
5597 unsigned long flags;
5600 ret = io_async_cancel_one(req->task->io_uring,
5601 (void *) (unsigned long) sqe_addr);
5602 if (ret != -ENOENT) {
5603 spin_lock_irqsave(&ctx->completion_lock, flags);
5607 spin_lock_irqsave(&ctx->completion_lock, flags);
5608 ret = io_timeout_cancel(ctx, sqe_addr);
5611 ret = io_poll_cancel(ctx, sqe_addr);
5615 io_cqring_fill_event(req, ret);
5616 io_commit_cqring(ctx);
5617 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5618 io_cqring_ev_posted(ctx);
5621 req_set_fail_links(req);
5625 static int io_async_cancel_prep(struct io_kiocb *req,
5626 const struct io_uring_sqe *sqe)
5628 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5630 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5632 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5635 req->cancel.addr = READ_ONCE(sqe->addr);
5639 static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
5641 struct io_ring_ctx *ctx = req->ctx;
5643 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
5647 static int io_rsrc_update_prep(struct io_kiocb *req,
5648 const struct io_uring_sqe *sqe)
5650 if (unlikely(req->ctx->flags & IORING_SETUP_SQPOLL))
5652 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5654 if (sqe->ioprio || sqe->rw_flags)
5657 req->rsrc_update.offset = READ_ONCE(sqe->off);
5658 req->rsrc_update.nr_args = READ_ONCE(sqe->len);
5659 if (!req->rsrc_update.nr_args)
5661 req->rsrc_update.arg = READ_ONCE(sqe->addr);
5665 static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
5667 struct io_ring_ctx *ctx = req->ctx;
5668 struct io_uring_rsrc_update up;
5671 if (issue_flags & IO_URING_F_NONBLOCK)
5674 up.offset = req->rsrc_update.offset;
5675 up.data = req->rsrc_update.arg;
5677 mutex_lock(&ctx->uring_lock);
5678 ret = __io_sqe_files_update(ctx, &up, req->rsrc_update.nr_args);
5679 mutex_unlock(&ctx->uring_lock);
5682 req_set_fail_links(req);
5683 __io_req_complete(req, issue_flags, ret, 0);
5687 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5689 switch (req->opcode) {
5692 case IORING_OP_READV:
5693 case IORING_OP_READ_FIXED:
5694 case IORING_OP_READ:
5695 return io_read_prep(req, sqe);
5696 case IORING_OP_WRITEV:
5697 case IORING_OP_WRITE_FIXED:
5698 case IORING_OP_WRITE:
5699 return io_write_prep(req, sqe);
5700 case IORING_OP_POLL_ADD:
5701 return io_poll_add_prep(req, sqe);
5702 case IORING_OP_POLL_REMOVE:
5703 return io_poll_remove_prep(req, sqe);
5704 case IORING_OP_FSYNC:
5705 return io_fsync_prep(req, sqe);
5706 case IORING_OP_SYNC_FILE_RANGE:
5707 return io_sfr_prep(req, sqe);
5708 case IORING_OP_SENDMSG:
5709 case IORING_OP_SEND:
5710 return io_sendmsg_prep(req, sqe);
5711 case IORING_OP_RECVMSG:
5712 case IORING_OP_RECV:
5713 return io_recvmsg_prep(req, sqe);
5714 case IORING_OP_CONNECT:
5715 return io_connect_prep(req, sqe);
5716 case IORING_OP_TIMEOUT:
5717 return io_timeout_prep(req, sqe, false);
5718 case IORING_OP_TIMEOUT_REMOVE:
5719 return io_timeout_remove_prep(req, sqe);
5720 case IORING_OP_ASYNC_CANCEL:
5721 return io_async_cancel_prep(req, sqe);
5722 case IORING_OP_LINK_TIMEOUT:
5723 return io_timeout_prep(req, sqe, true);
5724 case IORING_OP_ACCEPT:
5725 return io_accept_prep(req, sqe);
5726 case IORING_OP_FALLOCATE:
5727 return io_fallocate_prep(req, sqe);
5728 case IORING_OP_OPENAT:
5729 return io_openat_prep(req, sqe);
5730 case IORING_OP_CLOSE:
5731 return io_close_prep(req, sqe);
5732 case IORING_OP_FILES_UPDATE:
5733 return io_rsrc_update_prep(req, sqe);
5734 case IORING_OP_STATX:
5735 return io_statx_prep(req, sqe);
5736 case IORING_OP_FADVISE:
5737 return io_fadvise_prep(req, sqe);
5738 case IORING_OP_MADVISE:
5739 return io_madvise_prep(req, sqe);
5740 case IORING_OP_OPENAT2:
5741 return io_openat2_prep(req, sqe);
5742 case IORING_OP_EPOLL_CTL:
5743 return io_epoll_ctl_prep(req, sqe);
5744 case IORING_OP_SPLICE:
5745 return io_splice_prep(req, sqe);
5746 case IORING_OP_PROVIDE_BUFFERS:
5747 return io_provide_buffers_prep(req, sqe);
5748 case IORING_OP_REMOVE_BUFFERS:
5749 return io_remove_buffers_prep(req, sqe);
5751 return io_tee_prep(req, sqe);
5752 case IORING_OP_SHUTDOWN:
5753 return io_shutdown_prep(req, sqe);
5754 case IORING_OP_RENAMEAT:
5755 return io_renameat_prep(req, sqe);
5756 case IORING_OP_UNLINKAT:
5757 return io_unlinkat_prep(req, sqe);
5760 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5765 static int io_req_prep_async(struct io_kiocb *req)
5767 switch (req->opcode) {
5768 case IORING_OP_READV:
5769 case IORING_OP_READ_FIXED:
5770 case IORING_OP_READ:
5771 return io_rw_prep_async(req, READ);
5772 case IORING_OP_WRITEV:
5773 case IORING_OP_WRITE_FIXED:
5774 case IORING_OP_WRITE:
5775 return io_rw_prep_async(req, WRITE);
5776 case IORING_OP_SENDMSG:
5777 case IORING_OP_SEND:
5778 return io_sendmsg_prep_async(req);
5779 case IORING_OP_RECVMSG:
5780 case IORING_OP_RECV:
5781 return io_recvmsg_prep_async(req);
5782 case IORING_OP_CONNECT:
5783 return io_connect_prep_async(req);
5788 static int io_req_defer_prep(struct io_kiocb *req)
5790 if (!io_op_defs[req->opcode].needs_async_data)
5792 /* some opcodes init it during the inital prep */
5793 if (req->async_data)
5795 if (__io_alloc_async_data(req))
5797 return io_req_prep_async(req);
5800 static u32 io_get_sequence(struct io_kiocb *req)
5802 struct io_kiocb *pos;
5803 struct io_ring_ctx *ctx = req->ctx;
5804 u32 total_submitted, nr_reqs = 0;
5806 io_for_each_link(pos, req)
5809 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
5810 return total_submitted - nr_reqs;
5813 static int io_req_defer(struct io_kiocb *req)
5815 struct io_ring_ctx *ctx = req->ctx;
5816 struct io_defer_entry *de;
5820 /* Still need defer if there is pending req in defer list. */
5821 if (likely(list_empty_careful(&ctx->defer_list) &&
5822 !(req->flags & REQ_F_IO_DRAIN)))
5825 seq = io_get_sequence(req);
5826 /* Still a chance to pass the sequence check */
5827 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
5830 ret = io_req_defer_prep(req);
5833 io_prep_async_link(req);
5834 de = kmalloc(sizeof(*de), GFP_KERNEL);
5838 spin_lock_irq(&ctx->completion_lock);
5839 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
5840 spin_unlock_irq(&ctx->completion_lock);
5842 io_queue_async_work(req);
5843 return -EIOCBQUEUED;
5846 trace_io_uring_defer(ctx, req, req->user_data);
5849 list_add_tail(&de->list, &ctx->defer_list);
5850 spin_unlock_irq(&ctx->completion_lock);
5851 return -EIOCBQUEUED;
5854 static void __io_clean_op(struct io_kiocb *req)
5856 if (req->flags & REQ_F_BUFFER_SELECTED) {
5857 switch (req->opcode) {
5858 case IORING_OP_READV:
5859 case IORING_OP_READ_FIXED:
5860 case IORING_OP_READ:
5861 kfree((void *)(unsigned long)req->rw.addr);
5863 case IORING_OP_RECVMSG:
5864 case IORING_OP_RECV:
5865 kfree(req->sr_msg.kbuf);
5868 req->flags &= ~REQ_F_BUFFER_SELECTED;
5871 if (req->flags & REQ_F_NEED_CLEANUP) {
5872 switch (req->opcode) {
5873 case IORING_OP_READV:
5874 case IORING_OP_READ_FIXED:
5875 case IORING_OP_READ:
5876 case IORING_OP_WRITEV:
5877 case IORING_OP_WRITE_FIXED:
5878 case IORING_OP_WRITE: {
5879 struct io_async_rw *io = req->async_data;
5881 kfree(io->free_iovec);
5884 case IORING_OP_RECVMSG:
5885 case IORING_OP_SENDMSG: {
5886 struct io_async_msghdr *io = req->async_data;
5888 kfree(io->free_iov);
5891 case IORING_OP_SPLICE:
5893 io_put_file(req, req->splice.file_in,
5894 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
5896 case IORING_OP_OPENAT:
5897 case IORING_OP_OPENAT2:
5898 if (req->open.filename)
5899 putname(req->open.filename);
5901 case IORING_OP_RENAMEAT:
5902 putname(req->rename.oldpath);
5903 putname(req->rename.newpath);
5905 case IORING_OP_UNLINKAT:
5906 putname(req->unlink.filename);
5909 req->flags &= ~REQ_F_NEED_CLEANUP;
5913 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
5915 struct io_ring_ctx *ctx = req->ctx;
5916 const struct cred *creds = NULL;
5919 if (req->work.personality) {
5920 const struct cred *new_creds;
5922 if (!(issue_flags & IO_URING_F_NONBLOCK))
5923 mutex_lock(&ctx->uring_lock);
5924 new_creds = idr_find(&ctx->personality_idr, req->work.personality);
5925 if (!(issue_flags & IO_URING_F_NONBLOCK))
5926 mutex_unlock(&ctx->uring_lock);
5929 creds = override_creds(new_creds);
5932 switch (req->opcode) {
5934 ret = io_nop(req, issue_flags);
5936 case IORING_OP_READV:
5937 case IORING_OP_READ_FIXED:
5938 case IORING_OP_READ:
5939 ret = io_read(req, issue_flags);
5941 case IORING_OP_WRITEV:
5942 case IORING_OP_WRITE_FIXED:
5943 case IORING_OP_WRITE:
5944 ret = io_write(req, issue_flags);
5946 case IORING_OP_FSYNC:
5947 ret = io_fsync(req, issue_flags);
5949 case IORING_OP_POLL_ADD:
5950 ret = io_poll_add(req, issue_flags);
5952 case IORING_OP_POLL_REMOVE:
5953 ret = io_poll_remove(req, issue_flags);
5955 case IORING_OP_SYNC_FILE_RANGE:
5956 ret = io_sync_file_range(req, issue_flags);
5958 case IORING_OP_SENDMSG:
5959 ret = io_sendmsg(req, issue_flags);
5961 case IORING_OP_SEND:
5962 ret = io_send(req, issue_flags);
5964 case IORING_OP_RECVMSG:
5965 ret = io_recvmsg(req, issue_flags);
5967 case IORING_OP_RECV:
5968 ret = io_recv(req, issue_flags);
5970 case IORING_OP_TIMEOUT:
5971 ret = io_timeout(req, issue_flags);
5973 case IORING_OP_TIMEOUT_REMOVE:
5974 ret = io_timeout_remove(req, issue_flags);
5976 case IORING_OP_ACCEPT:
5977 ret = io_accept(req, issue_flags);
5979 case IORING_OP_CONNECT:
5980 ret = io_connect(req, issue_flags);
5982 case IORING_OP_ASYNC_CANCEL:
5983 ret = io_async_cancel(req, issue_flags);
5985 case IORING_OP_FALLOCATE:
5986 ret = io_fallocate(req, issue_flags);
5988 case IORING_OP_OPENAT:
5989 ret = io_openat(req, issue_flags);
5991 case IORING_OP_CLOSE:
5992 ret = io_close(req, issue_flags);
5994 case IORING_OP_FILES_UPDATE:
5995 ret = io_files_update(req, issue_flags);
5997 case IORING_OP_STATX:
5998 ret = io_statx(req, issue_flags);
6000 case IORING_OP_FADVISE:
6001 ret = io_fadvise(req, issue_flags);
6003 case IORING_OP_MADVISE:
6004 ret = io_madvise(req, issue_flags);
6006 case IORING_OP_OPENAT2:
6007 ret = io_openat2(req, issue_flags);
6009 case IORING_OP_EPOLL_CTL:
6010 ret = io_epoll_ctl(req, issue_flags);
6012 case IORING_OP_SPLICE:
6013 ret = io_splice(req, issue_flags);
6015 case IORING_OP_PROVIDE_BUFFERS:
6016 ret = io_provide_buffers(req, issue_flags);
6018 case IORING_OP_REMOVE_BUFFERS:
6019 ret = io_remove_buffers(req, issue_flags);
6022 ret = io_tee(req, issue_flags);
6024 case IORING_OP_SHUTDOWN:
6025 ret = io_shutdown(req, issue_flags);
6027 case IORING_OP_RENAMEAT:
6028 ret = io_renameat(req, issue_flags);
6030 case IORING_OP_UNLINKAT:
6031 ret = io_unlinkat(req, issue_flags);
6039 revert_creds(creds);
6044 /* If the op doesn't have a file, we're not polling for it */
6045 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
6046 const bool in_async = io_wq_current_is_worker();
6048 /* workqueue context doesn't hold uring_lock, grab it now */
6050 mutex_lock(&ctx->uring_lock);
6052 io_iopoll_req_issued(req, in_async);
6055 mutex_unlock(&ctx->uring_lock);
6061 static void io_wq_submit_work(struct io_wq_work *work)
6063 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6064 struct io_kiocb *timeout;
6067 timeout = io_prep_linked_timeout(req);
6069 io_queue_linked_timeout(timeout);
6071 if (work->flags & IO_WQ_WORK_CANCEL)
6076 ret = io_issue_sqe(req, 0);
6078 * We can get EAGAIN for polled IO even though we're
6079 * forcing a sync submission from here, since we can't
6080 * wait for request slots on the block side.
6088 /* avoid locking problems by failing it from a clean context */
6090 /* io-wq is going to take one down */
6091 refcount_inc(&req->refs);
6092 io_req_task_queue_fail(req, ret);
6096 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6099 struct fixed_rsrc_table *table;
6101 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
6102 return table->files[index & IORING_FILE_TABLE_MASK];
6105 static struct file *io_file_get(struct io_submit_state *state,
6106 struct io_kiocb *req, int fd, bool fixed)
6108 struct io_ring_ctx *ctx = req->ctx;
6112 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6114 fd = array_index_nospec(fd, ctx->nr_user_files);
6115 file = io_file_from_index(ctx, fd);
6116 io_set_resource_node(req);
6118 trace_io_uring_file_get(ctx, fd);
6119 file = __io_file_get(state, fd);
6122 if (file && unlikely(file->f_op == &io_uring_fops))
6123 io_req_track_inflight(req);
6127 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6129 struct io_timeout_data *data = container_of(timer,
6130 struct io_timeout_data, timer);
6131 struct io_kiocb *prev, *req = data->req;
6132 struct io_ring_ctx *ctx = req->ctx;
6133 unsigned long flags;
6135 spin_lock_irqsave(&ctx->completion_lock, flags);
6136 prev = req->timeout.head;
6137 req->timeout.head = NULL;
6140 * We don't expect the list to be empty, that will only happen if we
6141 * race with the completion of the linked work.
6143 if (prev && refcount_inc_not_zero(&prev->refs))
6144 io_remove_next_linked(prev);
6147 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6150 req_set_fail_links(prev);
6151 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6152 io_put_req_deferred(prev, 1);
6154 io_req_complete_post(req, -ETIME, 0);
6155 io_put_req_deferred(req, 1);
6157 return HRTIMER_NORESTART;
6160 static void __io_queue_linked_timeout(struct io_kiocb *req)
6163 * If the back reference is NULL, then our linked request finished
6164 * before we got a chance to setup the timer
6166 if (req->timeout.head) {
6167 struct io_timeout_data *data = req->async_data;
6169 data->timer.function = io_link_timeout_fn;
6170 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6175 static void io_queue_linked_timeout(struct io_kiocb *req)
6177 struct io_ring_ctx *ctx = req->ctx;
6179 spin_lock_irq(&ctx->completion_lock);
6180 __io_queue_linked_timeout(req);
6181 spin_unlock_irq(&ctx->completion_lock);
6183 /* drop submission reference */
6187 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6189 struct io_kiocb *nxt = req->link;
6191 if (!nxt || (req->flags & REQ_F_LINK_TIMEOUT) ||
6192 nxt->opcode != IORING_OP_LINK_TIMEOUT)
6195 nxt->timeout.head = req;
6196 nxt->flags |= REQ_F_LTIMEOUT_ACTIVE;
6197 req->flags |= REQ_F_LINK_TIMEOUT;
6201 static void __io_queue_sqe(struct io_kiocb *req)
6203 struct io_kiocb *linked_timeout = io_prep_linked_timeout(req);
6206 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
6209 * We async punt it if the file wasn't marked NOWAIT, or if the file
6210 * doesn't support non-blocking read/write attempts
6212 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6213 if (!io_arm_poll_handler(req)) {
6215 * Queued up for async execution, worker will release
6216 * submit reference when the iocb is actually submitted.
6218 io_queue_async_work(req);
6220 } else if (likely(!ret)) {
6221 /* drop submission reference */
6222 if (req->flags & REQ_F_COMPLETE_INLINE) {
6223 struct io_ring_ctx *ctx = req->ctx;
6224 struct io_comp_state *cs = &ctx->submit_state.comp;
6226 cs->reqs[cs->nr++] = req;
6227 if (cs->nr == ARRAY_SIZE(cs->reqs))
6228 io_submit_flush_completions(cs, ctx);
6233 req_set_fail_links(req);
6235 io_req_complete(req, ret);
6238 io_queue_linked_timeout(linked_timeout);
6241 static void io_queue_sqe(struct io_kiocb *req)
6245 ret = io_req_defer(req);
6247 if (ret != -EIOCBQUEUED) {
6249 req_set_fail_links(req);
6251 io_req_complete(req, ret);
6253 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6254 ret = io_req_defer_prep(req);
6257 io_queue_async_work(req);
6259 __io_queue_sqe(req);
6264 * Check SQE restrictions (opcode and flags).
6266 * Returns 'true' if SQE is allowed, 'false' otherwise.
6268 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6269 struct io_kiocb *req,
6270 unsigned int sqe_flags)
6272 if (!ctx->restricted)
6275 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6278 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6279 ctx->restrictions.sqe_flags_required)
6282 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6283 ctx->restrictions.sqe_flags_required))
6289 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6290 const struct io_uring_sqe *sqe)
6292 struct io_submit_state *state;
6293 unsigned int sqe_flags;
6296 req->opcode = READ_ONCE(sqe->opcode);
6297 /* same numerical values with corresponding REQ_F_*, safe to copy */
6298 req->flags = sqe_flags = READ_ONCE(sqe->flags);
6299 req->user_data = READ_ONCE(sqe->user_data);
6300 req->async_data = NULL;
6304 req->fixed_rsrc_refs = NULL;
6305 /* one is dropped after submission, the other at completion */
6306 refcount_set(&req->refs, 2);
6307 req->task = current;
6310 /* enforce forwards compatibility on users */
6311 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS)) {
6316 if (unlikely(req->opcode >= IORING_OP_LAST))
6319 if (unlikely(!io_check_restriction(ctx, req, sqe_flags)))
6322 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6323 !io_op_defs[req->opcode].buffer_select)
6326 req->work.list.next = NULL;
6327 req->work.flags = 0;
6328 req->work.personality = READ_ONCE(sqe->personality);
6329 state = &ctx->submit_state;
6332 * Plug now if we have more than 1 IO left after this, and the target
6333 * is potentially a read/write to block based storage.
6335 if (!state->plug_started && state->ios_left > 1 &&
6336 io_op_defs[req->opcode].plug) {
6337 blk_start_plug(&state->plug);
6338 state->plug_started = true;
6341 if (io_op_defs[req->opcode].needs_file) {
6342 bool fixed = req->flags & REQ_F_FIXED_FILE;
6344 req->file = io_file_get(state, req, READ_ONCE(sqe->fd), fixed);
6345 if (unlikely(!req->file))
6353 static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
6354 const struct io_uring_sqe *sqe)
6356 struct io_submit_link *link = &ctx->submit_state.link;
6359 ret = io_init_req(ctx, req, sqe);
6360 if (unlikely(ret)) {
6363 io_req_complete(req, ret);
6365 /* fail even hard links since we don't submit */
6366 link->head->flags |= REQ_F_FAIL_LINK;
6367 io_put_req(link->head);
6368 io_req_complete(link->head, -ECANCELED);
6373 ret = io_req_prep(req, sqe);
6377 /* don't need @sqe from now on */
6378 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6379 true, ctx->flags & IORING_SETUP_SQPOLL);
6382 * If we already have a head request, queue this one for async
6383 * submittal once the head completes. If we don't have a head but
6384 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6385 * submitted sync once the chain is complete. If none of those
6386 * conditions are true (normal request), then just queue it.
6389 struct io_kiocb *head = link->head;
6392 * Taking sequential execution of a link, draining both sides
6393 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6394 * requests in the link. So, it drains the head and the
6395 * next after the link request. The last one is done via
6396 * drain_next flag to persist the effect across calls.
6398 if (req->flags & REQ_F_IO_DRAIN) {
6399 head->flags |= REQ_F_IO_DRAIN;
6400 ctx->drain_next = 1;
6402 ret = io_req_defer_prep(req);
6405 trace_io_uring_link(ctx, req, head);
6406 link->last->link = req;
6409 /* last request of a link, enqueue the link */
6410 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6415 if (unlikely(ctx->drain_next)) {
6416 req->flags |= REQ_F_IO_DRAIN;
6417 ctx->drain_next = 0;
6419 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6431 * Batched submission is done, ensure local IO is flushed out.
6433 static void io_submit_state_end(struct io_submit_state *state,
6434 struct io_ring_ctx *ctx)
6436 if (state->link.head)
6437 io_queue_sqe(state->link.head);
6439 io_submit_flush_completions(&state->comp, ctx);
6440 if (state->plug_started)
6441 blk_finish_plug(&state->plug);
6442 io_state_file_put(state);
6446 * Start submission side cache.
6448 static void io_submit_state_start(struct io_submit_state *state,
6449 unsigned int max_ios)
6451 state->plug_started = false;
6452 state->ios_left = max_ios;
6453 /* set only head, no need to init link_last in advance */
6454 state->link.head = NULL;
6457 static void io_commit_sqring(struct io_ring_ctx *ctx)
6459 struct io_rings *rings = ctx->rings;
6462 * Ensure any loads from the SQEs are done at this point,
6463 * since once we write the new head, the application could
6464 * write new data to them.
6466 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6470 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6471 * that is mapped by userspace. This means that care needs to be taken to
6472 * ensure that reads are stable, as we cannot rely on userspace always
6473 * being a good citizen. If members of the sqe are validated and then later
6474 * used, it's important that those reads are done through READ_ONCE() to
6475 * prevent a re-load down the line.
6477 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6479 u32 *sq_array = ctx->sq_array;
6483 * The cached sq head (or cq tail) serves two purposes:
6485 * 1) allows us to batch the cost of updating the user visible
6487 * 2) allows the kernel side to track the head on its own, even
6488 * though the application is the one updating it.
6490 head = READ_ONCE(sq_array[ctx->cached_sq_head++ & ctx->sq_mask]);
6491 if (likely(head < ctx->sq_entries))
6492 return &ctx->sq_sqes[head];
6494 /* drop invalid entries */
6495 ctx->cached_sq_dropped++;
6496 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6500 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6504 /* if we have a backlog and couldn't flush it all, return BUSY */
6505 if (test_bit(0, &ctx->sq_check_overflow)) {
6506 if (!__io_cqring_overflow_flush(ctx, false, NULL, NULL))
6510 /* make sure SQ entry isn't read before tail */
6511 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6513 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6516 percpu_counter_add(¤t->io_uring->inflight, nr);
6517 refcount_add(nr, ¤t->usage);
6518 io_submit_state_start(&ctx->submit_state, nr);
6520 while (submitted < nr) {
6521 const struct io_uring_sqe *sqe;
6522 struct io_kiocb *req;
6524 req = io_alloc_req(ctx);
6525 if (unlikely(!req)) {
6527 submitted = -EAGAIN;
6530 sqe = io_get_sqe(ctx);
6531 if (unlikely(!sqe)) {
6532 kmem_cache_free(req_cachep, req);
6535 /* will complete beyond this point, count as submitted */
6537 if (io_submit_sqe(ctx, req, sqe))
6541 if (unlikely(submitted != nr)) {
6542 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6543 struct io_uring_task *tctx = current->io_uring;
6544 int unused = nr - ref_used;
6546 percpu_ref_put_many(&ctx->refs, unused);
6547 percpu_counter_sub(&tctx->inflight, unused);
6548 put_task_struct_many(current, unused);
6551 io_submit_state_end(&ctx->submit_state, ctx);
6552 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6553 io_commit_sqring(ctx);
6558 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6560 /* Tell userspace we may need a wakeup call */
6561 spin_lock_irq(&ctx->completion_lock);
6562 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6563 spin_unlock_irq(&ctx->completion_lock);
6566 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6568 spin_lock_irq(&ctx->completion_lock);
6569 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6570 spin_unlock_irq(&ctx->completion_lock);
6573 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
6575 unsigned int to_submit;
6578 to_submit = io_sqring_entries(ctx);
6579 /* if we're handling multiple rings, cap submit size for fairness */
6580 if (cap_entries && to_submit > 8)
6583 if (!list_empty(&ctx->iopoll_list) || to_submit) {
6584 unsigned nr_events = 0;
6586 mutex_lock(&ctx->uring_lock);
6587 if (!list_empty(&ctx->iopoll_list))
6588 io_do_iopoll(ctx, &nr_events, 0);
6590 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)))
6591 ret = io_submit_sqes(ctx, to_submit);
6592 mutex_unlock(&ctx->uring_lock);
6595 if (!io_sqring_full(ctx) && wq_has_sleeper(&ctx->sqo_sq_wait))
6596 wake_up(&ctx->sqo_sq_wait);
6601 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
6603 struct io_ring_ctx *ctx;
6604 unsigned sq_thread_idle = 0;
6606 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6607 if (sq_thread_idle < ctx->sq_thread_idle)
6608 sq_thread_idle = ctx->sq_thread_idle;
6611 sqd->sq_thread_idle = sq_thread_idle;
6614 static void io_sqd_init_new(struct io_sq_data *sqd)
6616 struct io_ring_ctx *ctx;
6618 while (!list_empty(&sqd->ctx_new_list)) {
6619 ctx = list_first_entry(&sqd->ctx_new_list, struct io_ring_ctx, sqd_list);
6620 list_move_tail(&ctx->sqd_list, &sqd->ctx_list);
6621 complete(&ctx->sq_thread_comp);
6624 io_sqd_update_thread_idle(sqd);
6627 static bool io_sq_thread_should_stop(struct io_sq_data *sqd)
6629 return test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
6632 static bool io_sq_thread_should_park(struct io_sq_data *sqd)
6634 return test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
6637 static void io_sq_thread_parkme(struct io_sq_data *sqd)
6641 * TASK_PARKED is a special state; we must serialize against
6642 * possible pending wakeups to avoid store-store collisions on
6645 * Such a collision might possibly result in the task state
6646 * changin from TASK_PARKED and us failing the
6647 * wait_task_inactive() in kthread_park().
6649 set_special_state(TASK_PARKED);
6650 if (!test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state))
6654 * Thread is going to call schedule(), do not preempt it,
6655 * or the caller of kthread_park() may spend more time in
6656 * wait_task_inactive().
6659 complete(&sqd->parked);
6660 schedule_preempt_disabled();
6663 __set_current_state(TASK_RUNNING);
6666 static int io_sq_thread(void *data)
6668 struct io_sq_data *sqd = data;
6669 struct io_ring_ctx *ctx;
6670 unsigned long timeout = 0;
6671 char buf[TASK_COMM_LEN];
6674 sprintf(buf, "iou-sqp-%d", sqd->task_pid);
6675 set_task_comm(current, buf);
6676 current->pf_io_worker = NULL;
6678 if (sqd->sq_cpu != -1)
6679 set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
6681 set_cpus_allowed_ptr(current, cpu_online_mask);
6682 current->flags |= PF_NO_SETAFFINITY;
6684 wait_for_completion(&sqd->startup);
6686 while (!io_sq_thread_should_stop(sqd)) {
6688 bool cap_entries, sqt_spin, needs_sched;
6691 * Any changes to the sqd lists are synchronized through the
6692 * thread parking. This synchronizes the thread vs users,
6693 * the users are synchronized on the sqd->ctx_lock.
6695 if (io_sq_thread_should_park(sqd)) {
6696 io_sq_thread_parkme(sqd);
6699 if (unlikely(!list_empty(&sqd->ctx_new_list))) {
6700 io_sqd_init_new(sqd);
6701 timeout = jiffies + sqd->sq_thread_idle;
6703 if (fatal_signal_pending(current))
6706 cap_entries = !list_is_singular(&sqd->ctx_list);
6707 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6708 ret = __io_sq_thread(ctx, cap_entries);
6709 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
6713 if (sqt_spin || !time_after(jiffies, timeout)) {
6717 timeout = jiffies + sqd->sq_thread_idle;
6722 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
6723 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6724 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6725 !list_empty_careful(&ctx->iopoll_list)) {
6726 needs_sched = false;
6729 if (io_sqring_entries(ctx)) {
6730 needs_sched = false;
6735 if (needs_sched && !io_sq_thread_should_park(sqd)) {
6736 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6737 io_ring_set_wakeup_flag(ctx);
6741 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6742 io_ring_clear_wakeup_flag(ctx);
6745 finish_wait(&sqd->wait, &wait);
6746 timeout = jiffies + sqd->sq_thread_idle;
6749 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6750 io_uring_cancel_sqpoll(ctx);
6755 * Ensure that we park properly if racing with someone trying to park
6756 * while we're exiting. If we fail to grab the lock, check park and
6757 * park if necessary. The ordering with the park bit and the lock
6758 * ensures that we catch this reliably.
6760 if (!mutex_trylock(&sqd->lock)) {
6761 if (io_sq_thread_should_park(sqd))
6762 io_sq_thread_parkme(sqd);
6763 mutex_lock(&sqd->lock);
6767 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6769 io_ring_set_wakeup_flag(ctx);
6772 complete(&sqd->exited);
6773 mutex_unlock(&sqd->lock);
6777 struct io_wait_queue {
6778 struct wait_queue_entry wq;
6779 struct io_ring_ctx *ctx;
6781 unsigned nr_timeouts;
6784 static inline bool io_should_wake(struct io_wait_queue *iowq)
6786 struct io_ring_ctx *ctx = iowq->ctx;
6789 * Wake up if we have enough events, or if a timeout occurred since we
6790 * started waiting. For timeouts, we always want to return to userspace,
6791 * regardless of event count.
6793 return io_cqring_events(ctx) >= iowq->to_wait ||
6794 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6797 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6798 int wake_flags, void *key)
6800 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6804 * Cannot safely flush overflowed CQEs from here, ensure we wake up
6805 * the task, and the next invocation will do it.
6807 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->cq_check_overflow))
6808 return autoremove_wake_function(curr, mode, wake_flags, key);
6812 static int io_run_task_work_sig(void)
6814 if (io_run_task_work())
6816 if (!signal_pending(current))
6818 if (test_tsk_thread_flag(current, TIF_NOTIFY_SIGNAL))
6819 return -ERESTARTSYS;
6823 /* when returns >0, the caller should retry */
6824 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
6825 struct io_wait_queue *iowq,
6826 signed long *timeout)
6830 /* make sure we run task_work before checking for signals */
6831 ret = io_run_task_work_sig();
6832 if (ret || io_should_wake(iowq))
6834 /* let the caller flush overflows, retry */
6835 if (test_bit(0, &ctx->cq_check_overflow))
6838 *timeout = schedule_timeout(*timeout);
6839 return !*timeout ? -ETIME : 1;
6843 * Wait until events become available, if we don't already have some. The
6844 * application must reap them itself, as they reside on the shared cq ring.
6846 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6847 const sigset_t __user *sig, size_t sigsz,
6848 struct __kernel_timespec __user *uts)
6850 struct io_wait_queue iowq = {
6853 .func = io_wake_function,
6854 .entry = LIST_HEAD_INIT(iowq.wq.entry),
6857 .to_wait = min_events,
6859 struct io_rings *rings = ctx->rings;
6860 signed long timeout = MAX_SCHEDULE_TIMEOUT;
6864 io_cqring_overflow_flush(ctx, false, NULL, NULL);
6865 if (io_cqring_events(ctx) >= min_events)
6867 if (!io_run_task_work())
6872 #ifdef CONFIG_COMPAT
6873 if (in_compat_syscall())
6874 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
6878 ret = set_user_sigmask(sig, sigsz);
6885 struct timespec64 ts;
6887 if (get_timespec64(&ts, uts))
6889 timeout = timespec64_to_jiffies(&ts);
6892 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
6893 trace_io_uring_cqring_wait(ctx, min_events);
6895 /* if we can't even flush overflow, don't wait for more */
6896 if (!io_cqring_overflow_flush(ctx, false, NULL, NULL)) {
6900 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
6901 TASK_INTERRUPTIBLE);
6902 ret = io_cqring_wait_schedule(ctx, &iowq, &timeout);
6903 finish_wait(&ctx->wait, &iowq.wq);
6907 restore_saved_sigmask_unless(ret == -EINTR);
6909 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
6912 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
6914 #if defined(CONFIG_UNIX)
6915 if (ctx->ring_sock) {
6916 struct sock *sock = ctx->ring_sock->sk;
6917 struct sk_buff *skb;
6919 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
6925 for (i = 0; i < ctx->nr_user_files; i++) {
6928 file = io_file_from_index(ctx, i);
6935 static void io_rsrc_data_ref_zero(struct percpu_ref *ref)
6937 struct fixed_rsrc_data *data;
6939 data = container_of(ref, struct fixed_rsrc_data, refs);
6940 complete(&data->done);
6943 static inline void io_rsrc_ref_lock(struct io_ring_ctx *ctx)
6945 spin_lock_bh(&ctx->rsrc_ref_lock);
6948 static inline void io_rsrc_ref_unlock(struct io_ring_ctx *ctx)
6950 spin_unlock_bh(&ctx->rsrc_ref_lock);
6953 static void io_sqe_rsrc_set_node(struct io_ring_ctx *ctx,
6954 struct fixed_rsrc_data *rsrc_data,
6955 struct fixed_rsrc_ref_node *ref_node)
6957 io_rsrc_ref_lock(ctx);
6958 rsrc_data->node = ref_node;
6959 list_add_tail(&ref_node->node, &ctx->rsrc_ref_list);
6960 io_rsrc_ref_unlock(ctx);
6961 percpu_ref_get(&rsrc_data->refs);
6964 static void io_sqe_rsrc_kill_node(struct io_ring_ctx *ctx, struct fixed_rsrc_data *data)
6966 struct fixed_rsrc_ref_node *ref_node = NULL;
6968 io_rsrc_ref_lock(ctx);
6969 ref_node = data->node;
6971 io_rsrc_ref_unlock(ctx);
6973 percpu_ref_kill(&ref_node->refs);
6976 static int io_rsrc_ref_quiesce(struct fixed_rsrc_data *data,
6977 struct io_ring_ctx *ctx,
6978 void (*rsrc_put)(struct io_ring_ctx *ctx,
6979 struct io_rsrc_put *prsrc))
6981 struct fixed_rsrc_ref_node *backup_node;
6987 data->quiesce = true;
6990 backup_node = alloc_fixed_rsrc_ref_node(ctx);
6993 backup_node->rsrc_data = data;
6994 backup_node->rsrc_put = rsrc_put;
6996 io_sqe_rsrc_kill_node(ctx, data);
6997 percpu_ref_kill(&data->refs);
6998 flush_delayed_work(&ctx->rsrc_put_work);
7000 ret = wait_for_completion_interruptible(&data->done);
7004 percpu_ref_resurrect(&data->refs);
7005 io_sqe_rsrc_set_node(ctx, data, backup_node);
7007 reinit_completion(&data->done);
7008 mutex_unlock(&ctx->uring_lock);
7009 ret = io_run_task_work_sig();
7010 mutex_lock(&ctx->uring_lock);
7012 data->quiesce = false;
7015 destroy_fixed_rsrc_ref_node(backup_node);
7019 static struct fixed_rsrc_data *alloc_fixed_rsrc_data(struct io_ring_ctx *ctx)
7021 struct fixed_rsrc_data *data;
7023 data = kzalloc(sizeof(*data), GFP_KERNEL);
7027 if (percpu_ref_init(&data->refs, io_rsrc_data_ref_zero,
7028 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) {
7033 init_completion(&data->done);
7037 static void free_fixed_rsrc_data(struct fixed_rsrc_data *data)
7039 percpu_ref_exit(&data->refs);
7044 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7046 struct fixed_rsrc_data *data = ctx->file_data;
7047 unsigned nr_tables, i;
7051 * percpu_ref_is_dying() is to stop parallel files unregister
7052 * Since we possibly drop uring lock later in this function to
7055 if (!data || percpu_ref_is_dying(&data->refs))
7057 ret = io_rsrc_ref_quiesce(data, ctx, io_ring_file_put);
7061 __io_sqe_files_unregister(ctx);
7062 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
7063 for (i = 0; i < nr_tables; i++)
7064 kfree(data->table[i].files);
7065 free_fixed_rsrc_data(data);
7066 ctx->file_data = NULL;
7067 ctx->nr_user_files = 0;
7071 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7072 __releases(&sqd->lock)
7074 if (sqd->thread == current)
7076 clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7078 wake_up_state(sqd->thread, TASK_PARKED);
7079 mutex_unlock(&sqd->lock);
7082 static void io_sq_thread_park(struct io_sq_data *sqd)
7083 __acquires(&sqd->lock)
7085 if (sqd->thread == current)
7087 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7088 mutex_lock(&sqd->lock);
7090 wake_up_process(sqd->thread);
7091 wait_for_completion(&sqd->parked);
7095 static void io_sq_thread_stop(struct io_sq_data *sqd)
7097 if (test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state))
7099 mutex_lock(&sqd->lock);
7101 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7102 WARN_ON_ONCE(test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state));
7103 wake_up_process(sqd->thread);
7104 mutex_unlock(&sqd->lock);
7105 wait_for_completion(&sqd->exited);
7106 WARN_ON_ONCE(sqd->thread);
7108 mutex_unlock(&sqd->lock);
7112 static void io_put_sq_data(struct io_sq_data *sqd)
7114 if (refcount_dec_and_test(&sqd->refs)) {
7115 io_sq_thread_stop(sqd);
7120 static void io_sq_thread_finish(struct io_ring_ctx *ctx)
7122 struct io_sq_data *sqd = ctx->sq_data;
7125 complete(&sqd->startup);
7127 wait_for_completion(&ctx->sq_thread_comp);
7128 io_sq_thread_park(sqd);
7131 mutex_lock(&sqd->ctx_lock);
7132 list_del(&ctx->sqd_list);
7133 io_sqd_update_thread_idle(sqd);
7134 mutex_unlock(&sqd->ctx_lock);
7137 io_sq_thread_unpark(sqd);
7139 io_put_sq_data(sqd);
7140 ctx->sq_data = NULL;
7144 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7146 struct io_ring_ctx *ctx_attach;
7147 struct io_sq_data *sqd;
7150 f = fdget(p->wq_fd);
7152 return ERR_PTR(-ENXIO);
7153 if (f.file->f_op != &io_uring_fops) {
7155 return ERR_PTR(-EINVAL);
7158 ctx_attach = f.file->private_data;
7159 sqd = ctx_attach->sq_data;
7162 return ERR_PTR(-EINVAL);
7165 refcount_inc(&sqd->refs);
7170 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p)
7172 struct io_sq_data *sqd;
7174 if (p->flags & IORING_SETUP_ATTACH_WQ)
7175 return io_attach_sq_data(p);
7177 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7179 return ERR_PTR(-ENOMEM);
7181 refcount_set(&sqd->refs, 1);
7182 INIT_LIST_HEAD(&sqd->ctx_list);
7183 INIT_LIST_HEAD(&sqd->ctx_new_list);
7184 mutex_init(&sqd->ctx_lock);
7185 mutex_init(&sqd->lock);
7186 init_waitqueue_head(&sqd->wait);
7187 init_completion(&sqd->startup);
7188 init_completion(&sqd->parked);
7189 init_completion(&sqd->exited);
7193 #if defined(CONFIG_UNIX)
7195 * Ensure the UNIX gc is aware of our file set, so we are certain that
7196 * the io_uring can be safely unregistered on process exit, even if we have
7197 * loops in the file referencing.
7199 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7201 struct sock *sk = ctx->ring_sock->sk;
7202 struct scm_fp_list *fpl;
7203 struct sk_buff *skb;
7206 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7210 skb = alloc_skb(0, GFP_KERNEL);
7219 fpl->user = get_uid(current_user());
7220 for (i = 0; i < nr; i++) {
7221 struct file *file = io_file_from_index(ctx, i + offset);
7225 fpl->fp[nr_files] = get_file(file);
7226 unix_inflight(fpl->user, fpl->fp[nr_files]);
7231 fpl->max = SCM_MAX_FD;
7232 fpl->count = nr_files;
7233 UNIXCB(skb).fp = fpl;
7234 skb->destructor = unix_destruct_scm;
7235 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7236 skb_queue_head(&sk->sk_receive_queue, skb);
7238 for (i = 0; i < nr_files; i++)
7249 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7250 * causes regular reference counting to break down. We rely on the UNIX
7251 * garbage collection to take care of this problem for us.
7253 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7255 unsigned left, total;
7259 left = ctx->nr_user_files;
7261 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7263 ret = __io_sqe_files_scm(ctx, this_files, total);
7267 total += this_files;
7273 while (total < ctx->nr_user_files) {
7274 struct file *file = io_file_from_index(ctx, total);
7284 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7290 static int io_sqe_alloc_file_tables(struct fixed_rsrc_data *file_data,
7291 unsigned nr_tables, unsigned nr_files)
7295 for (i = 0; i < nr_tables; i++) {
7296 struct fixed_rsrc_table *table = &file_data->table[i];
7297 unsigned this_files;
7299 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
7300 table->files = kcalloc(this_files, sizeof(struct file *),
7304 nr_files -= this_files;
7310 for (i = 0; i < nr_tables; i++) {
7311 struct fixed_rsrc_table *table = &file_data->table[i];
7312 kfree(table->files);
7317 static void io_ring_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
7319 struct file *file = prsrc->file;
7320 #if defined(CONFIG_UNIX)
7321 struct sock *sock = ctx->ring_sock->sk;
7322 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7323 struct sk_buff *skb;
7326 __skb_queue_head_init(&list);
7329 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7330 * remove this entry and rearrange the file array.
7332 skb = skb_dequeue(head);
7334 struct scm_fp_list *fp;
7336 fp = UNIXCB(skb).fp;
7337 for (i = 0; i < fp->count; i++) {
7340 if (fp->fp[i] != file)
7343 unix_notinflight(fp->user, fp->fp[i]);
7344 left = fp->count - 1 - i;
7346 memmove(&fp->fp[i], &fp->fp[i + 1],
7347 left * sizeof(struct file *));
7354 __skb_queue_tail(&list, skb);
7364 __skb_queue_tail(&list, skb);
7366 skb = skb_dequeue(head);
7369 if (skb_peek(&list)) {
7370 spin_lock_irq(&head->lock);
7371 while ((skb = __skb_dequeue(&list)) != NULL)
7372 __skb_queue_tail(head, skb);
7373 spin_unlock_irq(&head->lock);
7380 static void __io_rsrc_put_work(struct fixed_rsrc_ref_node *ref_node)
7382 struct fixed_rsrc_data *rsrc_data = ref_node->rsrc_data;
7383 struct io_ring_ctx *ctx = rsrc_data->ctx;
7384 struct io_rsrc_put *prsrc, *tmp;
7386 list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
7387 list_del(&prsrc->list);
7388 ref_node->rsrc_put(ctx, prsrc);
7392 percpu_ref_exit(&ref_node->refs);
7394 percpu_ref_put(&rsrc_data->refs);
7397 static void io_rsrc_put_work(struct work_struct *work)
7399 struct io_ring_ctx *ctx;
7400 struct llist_node *node;
7402 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
7403 node = llist_del_all(&ctx->rsrc_put_llist);
7406 struct fixed_rsrc_ref_node *ref_node;
7407 struct llist_node *next = node->next;
7409 ref_node = llist_entry(node, struct fixed_rsrc_ref_node, llist);
7410 __io_rsrc_put_work(ref_node);
7415 static struct file **io_fixed_file_slot(struct fixed_rsrc_data *file_data,
7418 struct fixed_rsrc_table *table;
7420 table = &file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7421 return &table->files[i & IORING_FILE_TABLE_MASK];
7424 static void io_rsrc_node_ref_zero(struct percpu_ref *ref)
7426 struct fixed_rsrc_ref_node *ref_node;
7427 struct fixed_rsrc_data *data;
7428 struct io_ring_ctx *ctx;
7429 bool first_add = false;
7432 ref_node = container_of(ref, struct fixed_rsrc_ref_node, refs);
7433 data = ref_node->rsrc_data;
7436 io_rsrc_ref_lock(ctx);
7437 ref_node->done = true;
7439 while (!list_empty(&ctx->rsrc_ref_list)) {
7440 ref_node = list_first_entry(&ctx->rsrc_ref_list,
7441 struct fixed_rsrc_ref_node, node);
7442 /* recycle ref nodes in order */
7443 if (!ref_node->done)
7445 list_del(&ref_node->node);
7446 first_add |= llist_add(&ref_node->llist, &ctx->rsrc_put_llist);
7448 io_rsrc_ref_unlock(ctx);
7450 if (percpu_ref_is_dying(&data->refs))
7454 mod_delayed_work(system_wq, &ctx->rsrc_put_work, 0);
7456 queue_delayed_work(system_wq, &ctx->rsrc_put_work, delay);
7459 static struct fixed_rsrc_ref_node *alloc_fixed_rsrc_ref_node(
7460 struct io_ring_ctx *ctx)
7462 struct fixed_rsrc_ref_node *ref_node;
7464 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7468 if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
7473 INIT_LIST_HEAD(&ref_node->node);
7474 INIT_LIST_HEAD(&ref_node->rsrc_list);
7475 ref_node->done = false;
7479 static void init_fixed_file_ref_node(struct io_ring_ctx *ctx,
7480 struct fixed_rsrc_ref_node *ref_node)
7482 ref_node->rsrc_data = ctx->file_data;
7483 ref_node->rsrc_put = io_ring_file_put;
7486 static void destroy_fixed_rsrc_ref_node(struct fixed_rsrc_ref_node *ref_node)
7488 percpu_ref_exit(&ref_node->refs);
7493 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7496 __s32 __user *fds = (__s32 __user *) arg;
7497 unsigned nr_tables, i;
7499 int fd, ret = -ENOMEM;
7500 struct fixed_rsrc_ref_node *ref_node;
7501 struct fixed_rsrc_data *file_data;
7507 if (nr_args > IORING_MAX_FIXED_FILES)
7510 file_data = alloc_fixed_rsrc_data(ctx);
7513 ctx->file_data = file_data;
7515 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
7516 file_data->table = kcalloc(nr_tables, sizeof(*file_data->table),
7518 if (!file_data->table)
7521 if (io_sqe_alloc_file_tables(file_data, nr_tables, nr_args))
7524 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7525 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7529 /* allow sparse sets */
7539 * Don't allow io_uring instances to be registered. If UNIX
7540 * isn't enabled, then this causes a reference cycle and this
7541 * instance can never get freed. If UNIX is enabled we'll
7542 * handle it just fine, but there's still no point in allowing
7543 * a ring fd as it doesn't support regular read/write anyway.
7545 if (file->f_op == &io_uring_fops) {
7549 *io_fixed_file_slot(file_data, i) = file;
7552 ret = io_sqe_files_scm(ctx);
7554 io_sqe_files_unregister(ctx);
7558 ref_node = alloc_fixed_rsrc_ref_node(ctx);
7560 io_sqe_files_unregister(ctx);
7563 init_fixed_file_ref_node(ctx, ref_node);
7565 io_sqe_rsrc_set_node(ctx, file_data, ref_node);
7568 for (i = 0; i < ctx->nr_user_files; i++) {
7569 file = io_file_from_index(ctx, i);
7573 for (i = 0; i < nr_tables; i++)
7574 kfree(file_data->table[i].files);
7575 ctx->nr_user_files = 0;
7577 free_fixed_rsrc_data(ctx->file_data);
7578 ctx->file_data = NULL;
7582 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7585 #if defined(CONFIG_UNIX)
7586 struct sock *sock = ctx->ring_sock->sk;
7587 struct sk_buff_head *head = &sock->sk_receive_queue;
7588 struct sk_buff *skb;
7591 * See if we can merge this file into an existing skb SCM_RIGHTS
7592 * file set. If there's no room, fall back to allocating a new skb
7593 * and filling it in.
7595 spin_lock_irq(&head->lock);
7596 skb = skb_peek(head);
7598 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7600 if (fpl->count < SCM_MAX_FD) {
7601 __skb_unlink(skb, head);
7602 spin_unlock_irq(&head->lock);
7603 fpl->fp[fpl->count] = get_file(file);
7604 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7606 spin_lock_irq(&head->lock);
7607 __skb_queue_head(head, skb);
7612 spin_unlock_irq(&head->lock);
7619 return __io_sqe_files_scm(ctx, 1, index);
7625 static int io_queue_rsrc_removal(struct fixed_rsrc_data *data, void *rsrc)
7627 struct io_rsrc_put *prsrc;
7628 struct fixed_rsrc_ref_node *ref_node = data->node;
7630 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
7635 list_add(&prsrc->list, &ref_node->rsrc_list);
7640 static inline int io_queue_file_removal(struct fixed_rsrc_data *data,
7643 return io_queue_rsrc_removal(data, (void *)file);
7646 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7647 struct io_uring_rsrc_update *up,
7650 struct fixed_rsrc_data *data = ctx->file_data;
7651 struct fixed_rsrc_ref_node *ref_node;
7652 struct file *file, **file_slot;
7656 bool needs_switch = false;
7658 if (check_add_overflow(up->offset, nr_args, &done))
7660 if (done > ctx->nr_user_files)
7663 ref_node = alloc_fixed_rsrc_ref_node(ctx);
7666 init_fixed_file_ref_node(ctx, ref_node);
7668 fds = u64_to_user_ptr(up->data);
7669 for (done = 0; done < nr_args; done++) {
7671 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
7675 if (fd == IORING_REGISTER_FILES_SKIP)
7678 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
7679 file_slot = io_fixed_file_slot(ctx->file_data, i);
7682 err = io_queue_file_removal(data, *file_slot);
7686 needs_switch = true;
7695 * Don't allow io_uring instances to be registered. If
7696 * UNIX isn't enabled, then this causes a reference
7697 * cycle and this instance can never get freed. If UNIX
7698 * is enabled we'll handle it just fine, but there's
7699 * still no point in allowing a ring fd as it doesn't
7700 * support regular read/write anyway.
7702 if (file->f_op == &io_uring_fops) {
7708 err = io_sqe_file_register(ctx, file, i);
7718 percpu_ref_kill(&data->node->refs);
7719 io_sqe_rsrc_set_node(ctx, data, ref_node);
7721 destroy_fixed_rsrc_ref_node(ref_node);
7723 return done ? done : err;
7726 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
7729 struct io_uring_rsrc_update up;
7731 if (!ctx->file_data)
7735 if (copy_from_user(&up, arg, sizeof(up)))
7740 return __io_sqe_files_update(ctx, &up, nr_args);
7743 static struct io_wq_work *io_free_work(struct io_wq_work *work)
7745 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7747 req = io_put_req_find_next(req);
7748 return req ? &req->work : NULL;
7751 static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx)
7753 struct io_wq_hash *hash;
7754 struct io_wq_data data;
7755 unsigned int concurrency;
7757 hash = ctx->hash_map;
7759 hash = kzalloc(sizeof(*hash), GFP_KERNEL);
7761 return ERR_PTR(-ENOMEM);
7762 refcount_set(&hash->refs, 1);
7763 init_waitqueue_head(&hash->wait);
7764 ctx->hash_map = hash;
7768 data.free_work = io_free_work;
7769 data.do_work = io_wq_submit_work;
7771 /* Do QD, or 4 * CPUS, whatever is smallest */
7772 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7774 return io_wq_create(concurrency, &data);
7777 static int io_uring_alloc_task_context(struct task_struct *task,
7778 struct io_ring_ctx *ctx)
7780 struct io_uring_task *tctx;
7783 tctx = kmalloc(sizeof(*tctx), GFP_KERNEL);
7784 if (unlikely(!tctx))
7787 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
7788 if (unlikely(ret)) {
7793 tctx->io_wq = io_init_wq_offload(ctx);
7794 if (IS_ERR(tctx->io_wq)) {
7795 ret = PTR_ERR(tctx->io_wq);
7796 percpu_counter_destroy(&tctx->inflight);
7802 init_waitqueue_head(&tctx->wait);
7804 atomic_set(&tctx->in_idle, 0);
7805 tctx->sqpoll = false;
7806 task->io_uring = tctx;
7807 spin_lock_init(&tctx->task_lock);
7808 INIT_WQ_LIST(&tctx->task_list);
7809 tctx->task_state = 0;
7810 init_task_work(&tctx->task_work, tctx_task_work);
7814 void __io_uring_free(struct task_struct *tsk)
7816 struct io_uring_task *tctx = tsk->io_uring;
7818 WARN_ON_ONCE(!xa_empty(&tctx->xa));
7819 WARN_ON_ONCE(tctx->io_wq);
7821 percpu_counter_destroy(&tctx->inflight);
7823 tsk->io_uring = NULL;
7826 static int io_sq_thread_fork(struct io_sq_data *sqd, struct io_ring_ctx *ctx)
7828 struct task_struct *tsk;
7831 clear_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7832 reinit_completion(&sqd->parked);
7834 sqd->task_pid = current->pid;
7835 tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE);
7837 return PTR_ERR(tsk);
7838 ret = io_uring_alloc_task_context(tsk, ctx);
7840 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7842 wake_up_new_task(tsk);
7846 static int io_sq_offload_create(struct io_ring_ctx *ctx,
7847 struct io_uring_params *p)
7851 /* Retain compatibility with failing for an invalid attach attempt */
7852 if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
7853 IORING_SETUP_ATTACH_WQ) {
7856 f = fdget(p->wq_fd);
7859 if (f.file->f_op != &io_uring_fops) {
7865 if (ctx->flags & IORING_SETUP_SQPOLL) {
7866 struct task_struct *tsk;
7867 struct io_sq_data *sqd;
7870 if (!capable(CAP_SYS_ADMIN) && !capable(CAP_SYS_NICE))
7873 sqd = io_get_sq_data(p);
7880 io_sq_thread_park(sqd);
7881 mutex_lock(&sqd->ctx_lock);
7882 list_add(&ctx->sqd_list, &sqd->ctx_new_list);
7883 mutex_unlock(&sqd->ctx_lock);
7884 io_sq_thread_unpark(sqd);
7886 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
7887 if (!ctx->sq_thread_idle)
7888 ctx->sq_thread_idle = HZ;
7893 if (p->flags & IORING_SETUP_SQ_AFF) {
7894 int cpu = p->sq_thread_cpu;
7897 if (cpu >= nr_cpu_ids)
7899 if (!cpu_online(cpu))
7907 sqd->task_pid = current->pid;
7908 tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE);
7913 ret = io_uring_alloc_task_context(tsk, ctx);
7915 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7917 wake_up_new_task(tsk);
7920 } else if (p->flags & IORING_SETUP_SQ_AFF) {
7921 /* Can't have SQ_AFF without SQPOLL */
7928 io_sq_thread_finish(ctx);
7932 static void io_sq_offload_start(struct io_ring_ctx *ctx)
7934 struct io_sq_data *sqd = ctx->sq_data;
7936 ctx->flags &= ~IORING_SETUP_R_DISABLED;
7937 if (ctx->flags & IORING_SETUP_SQPOLL)
7938 complete(&sqd->startup);
7941 static inline void __io_unaccount_mem(struct user_struct *user,
7942 unsigned long nr_pages)
7944 atomic_long_sub(nr_pages, &user->locked_vm);
7947 static inline int __io_account_mem(struct user_struct *user,
7948 unsigned long nr_pages)
7950 unsigned long page_limit, cur_pages, new_pages;
7952 /* Don't allow more pages than we can safely lock */
7953 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
7956 cur_pages = atomic_long_read(&user->locked_vm);
7957 new_pages = cur_pages + nr_pages;
7958 if (new_pages > page_limit)
7960 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
7961 new_pages) != cur_pages);
7966 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
7969 __io_unaccount_mem(ctx->user, nr_pages);
7971 if (ctx->mm_account)
7972 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
7975 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
7980 ret = __io_account_mem(ctx->user, nr_pages);
7985 if (ctx->mm_account)
7986 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
7991 static void io_mem_free(void *ptr)
7998 page = virt_to_head_page(ptr);
7999 if (put_page_testzero(page))
8000 free_compound_page(page);
8003 static void *io_mem_alloc(size_t size)
8005 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
8006 __GFP_NORETRY | __GFP_ACCOUNT;
8008 return (void *) __get_free_pages(gfp_flags, get_order(size));
8011 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8014 struct io_rings *rings;
8015 size_t off, sq_array_size;
8017 off = struct_size(rings, cqes, cq_entries);
8018 if (off == SIZE_MAX)
8022 off = ALIGN(off, SMP_CACHE_BYTES);
8030 sq_array_size = array_size(sizeof(u32), sq_entries);
8031 if (sq_array_size == SIZE_MAX)
8034 if (check_add_overflow(off, sq_array_size, &off))
8040 static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8044 if (!ctx->user_bufs)
8047 for (i = 0; i < ctx->nr_user_bufs; i++) {
8048 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8050 for (j = 0; j < imu->nr_bvecs; j++)
8051 unpin_user_page(imu->bvec[j].bv_page);
8053 if (imu->acct_pages)
8054 io_unaccount_mem(ctx, imu->acct_pages);
8059 kfree(ctx->user_bufs);
8060 ctx->user_bufs = NULL;
8061 ctx->nr_user_bufs = 0;
8065 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8066 void __user *arg, unsigned index)
8068 struct iovec __user *src;
8070 #ifdef CONFIG_COMPAT
8072 struct compat_iovec __user *ciovs;
8073 struct compat_iovec ciov;
8075 ciovs = (struct compat_iovec __user *) arg;
8076 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8079 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8080 dst->iov_len = ciov.iov_len;
8084 src = (struct iovec __user *) arg;
8085 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8091 * Not super efficient, but this is just a registration time. And we do cache
8092 * the last compound head, so generally we'll only do a full search if we don't
8095 * We check if the given compound head page has already been accounted, to
8096 * avoid double accounting it. This allows us to account the full size of the
8097 * page, not just the constituent pages of a huge page.
8099 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8100 int nr_pages, struct page *hpage)
8104 /* check current page array */
8105 for (i = 0; i < nr_pages; i++) {
8106 if (!PageCompound(pages[i]))
8108 if (compound_head(pages[i]) == hpage)
8112 /* check previously registered pages */
8113 for (i = 0; i < ctx->nr_user_bufs; i++) {
8114 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8116 for (j = 0; j < imu->nr_bvecs; j++) {
8117 if (!PageCompound(imu->bvec[j].bv_page))
8119 if (compound_head(imu->bvec[j].bv_page) == hpage)
8127 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8128 int nr_pages, struct io_mapped_ubuf *imu,
8129 struct page **last_hpage)
8133 for (i = 0; i < nr_pages; i++) {
8134 if (!PageCompound(pages[i])) {
8139 hpage = compound_head(pages[i]);
8140 if (hpage == *last_hpage)
8142 *last_hpage = hpage;
8143 if (headpage_already_acct(ctx, pages, i, hpage))
8145 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8149 if (!imu->acct_pages)
8152 ret = io_account_mem(ctx, imu->acct_pages);
8154 imu->acct_pages = 0;
8158 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
8159 struct io_mapped_ubuf *imu,
8160 struct page **last_hpage)
8162 struct vm_area_struct **vmas = NULL;
8163 struct page **pages = NULL;
8164 unsigned long off, start, end, ubuf;
8166 int ret, pret, nr_pages, i;
8168 ubuf = (unsigned long) iov->iov_base;
8169 end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8170 start = ubuf >> PAGE_SHIFT;
8171 nr_pages = end - start;
8175 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
8179 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
8184 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
8190 mmap_read_lock(current->mm);
8191 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
8193 if (pret == nr_pages) {
8194 /* don't support file backed memory */
8195 for (i = 0; i < nr_pages; i++) {
8196 struct vm_area_struct *vma = vmas[i];
8199 !is_file_hugepages(vma->vm_file)) {
8205 ret = pret < 0 ? pret : -EFAULT;
8207 mmap_read_unlock(current->mm);
8210 * if we did partial map, or found file backed vmas,
8211 * release any pages we did get
8214 unpin_user_pages(pages, pret);
8219 ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
8221 unpin_user_pages(pages, pret);
8226 off = ubuf & ~PAGE_MASK;
8227 size = iov->iov_len;
8228 for (i = 0; i < nr_pages; i++) {
8231 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8232 imu->bvec[i].bv_page = pages[i];
8233 imu->bvec[i].bv_len = vec_len;
8234 imu->bvec[i].bv_offset = off;
8238 /* store original address for later verification */
8240 imu->len = iov->iov_len;
8241 imu->nr_bvecs = nr_pages;
8249 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
8253 if (!nr_args || nr_args > UIO_MAXIOV)
8256 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
8258 if (!ctx->user_bufs)
8264 static int io_buffer_validate(struct iovec *iov)
8267 * Don't impose further limits on the size and buffer
8268 * constraints here, we'll -EINVAL later when IO is
8269 * submitted if they are wrong.
8271 if (!iov->iov_base || !iov->iov_len)
8274 /* arbitrary limit, but we need something */
8275 if (iov->iov_len > SZ_1G)
8281 static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
8282 unsigned int nr_args)
8286 struct page *last_hpage = NULL;
8288 ret = io_buffers_map_alloc(ctx, nr_args);
8292 for (i = 0; i < nr_args; i++) {
8293 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8295 ret = io_copy_iov(ctx, &iov, arg, i);
8299 ret = io_buffer_validate(&iov);
8303 ret = io_sqe_buffer_register(ctx, &iov, imu, &last_hpage);
8307 ctx->nr_user_bufs++;
8311 io_sqe_buffers_unregister(ctx);
8316 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8318 __s32 __user *fds = arg;
8324 if (copy_from_user(&fd, fds, sizeof(*fds)))
8327 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8328 if (IS_ERR(ctx->cq_ev_fd)) {
8329 int ret = PTR_ERR(ctx->cq_ev_fd);
8330 ctx->cq_ev_fd = NULL;
8337 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8339 if (ctx->cq_ev_fd) {
8340 eventfd_ctx_put(ctx->cq_ev_fd);
8341 ctx->cq_ev_fd = NULL;
8348 static int __io_destroy_buffers(int id, void *p, void *data)
8350 struct io_ring_ctx *ctx = data;
8351 struct io_buffer *buf = p;
8353 __io_remove_buffers(ctx, buf, id, -1U);
8357 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8359 idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx);
8360 idr_destroy(&ctx->io_buffer_idr);
8363 static void io_req_cache_free(struct list_head *list, struct task_struct *tsk)
8365 struct io_kiocb *req, *nxt;
8367 list_for_each_entry_safe(req, nxt, list, compl.list) {
8368 if (tsk && req->task != tsk)
8370 list_del(&req->compl.list);
8371 kmem_cache_free(req_cachep, req);
8375 static void io_req_caches_free(struct io_ring_ctx *ctx)
8377 struct io_submit_state *submit_state = &ctx->submit_state;
8378 struct io_comp_state *cs = &ctx->submit_state.comp;
8380 mutex_lock(&ctx->uring_lock);
8382 if (submit_state->free_reqs) {
8383 kmem_cache_free_bulk(req_cachep, submit_state->free_reqs,
8384 submit_state->reqs);
8385 submit_state->free_reqs = 0;
8388 spin_lock_irq(&ctx->completion_lock);
8389 list_splice_init(&cs->locked_free_list, &cs->free_list);
8390 cs->locked_free_nr = 0;
8391 spin_unlock_irq(&ctx->completion_lock);
8393 io_req_cache_free(&cs->free_list, NULL);
8395 mutex_unlock(&ctx->uring_lock);
8398 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8401 * Some may use context even when all refs and requests have been put,
8402 * and they are free to do so while still holding uring_lock, see
8403 * __io_req_task_submit(). Wait for them to finish.
8405 mutex_lock(&ctx->uring_lock);
8406 mutex_unlock(&ctx->uring_lock);
8408 io_sq_thread_finish(ctx);
8409 io_sqe_buffers_unregister(ctx);
8411 if (ctx->mm_account) {
8412 mmdrop(ctx->mm_account);
8413 ctx->mm_account = NULL;
8416 mutex_lock(&ctx->uring_lock);
8417 io_sqe_files_unregister(ctx);
8418 mutex_unlock(&ctx->uring_lock);
8419 io_eventfd_unregister(ctx);
8420 io_destroy_buffers(ctx);
8421 idr_destroy(&ctx->personality_idr);
8423 #if defined(CONFIG_UNIX)
8424 if (ctx->ring_sock) {
8425 ctx->ring_sock->file = NULL; /* so that iput() is called */
8426 sock_release(ctx->ring_sock);
8430 io_mem_free(ctx->rings);
8431 io_mem_free(ctx->sq_sqes);
8433 percpu_ref_exit(&ctx->refs);
8434 free_uid(ctx->user);
8435 io_req_caches_free(ctx);
8437 io_wq_put_hash(ctx->hash_map);
8438 kfree(ctx->cancel_hash);
8442 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8444 struct io_ring_ctx *ctx = file->private_data;
8447 poll_wait(file, &ctx->cq_wait, wait);
8449 * synchronizes with barrier from wq_has_sleeper call in
8453 if (!io_sqring_full(ctx))
8454 mask |= EPOLLOUT | EPOLLWRNORM;
8457 * Don't flush cqring overflow list here, just do a simple check.
8458 * Otherwise there could possible be ABBA deadlock:
8461 * lock(&ctx->uring_lock);
8463 * lock(&ctx->uring_lock);
8466 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
8467 * pushs them to do the flush.
8469 if (io_cqring_events(ctx) || test_bit(0, &ctx->cq_check_overflow))
8470 mask |= EPOLLIN | EPOLLRDNORM;
8475 static int io_uring_fasync(int fd, struct file *file, int on)
8477 struct io_ring_ctx *ctx = file->private_data;
8479 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8482 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8484 const struct cred *creds;
8486 creds = idr_remove(&ctx->personality_idr, id);
8495 static int io_remove_personalities(int id, void *p, void *data)
8497 struct io_ring_ctx *ctx = data;
8499 io_unregister_personality(ctx, id);
8503 static bool io_run_ctx_fallback(struct io_ring_ctx *ctx)
8505 struct callback_head *work, *next;
8506 bool executed = false;
8509 work = xchg(&ctx->exit_task_work, NULL);
8525 static void io_ring_exit_work(struct work_struct *work)
8527 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
8531 * If we're doing polled IO and end up having requests being
8532 * submitted async (out-of-line), then completions can come in while
8533 * we're waiting for refs to drop. We need to reap these manually,
8534 * as nobody else will be looking for them.
8537 io_uring_try_cancel_requests(ctx, NULL, NULL);
8538 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
8539 io_ring_ctx_free(ctx);
8542 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8544 mutex_lock(&ctx->uring_lock);
8545 percpu_ref_kill(&ctx->refs);
8546 /* if force is set, the ring is going away. always drop after that */
8547 ctx->cq_overflow_flushed = 1;
8549 __io_cqring_overflow_flush(ctx, true, NULL, NULL);
8550 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
8551 mutex_unlock(&ctx->uring_lock);
8553 io_kill_timeouts(ctx, NULL, NULL);
8554 io_poll_remove_all(ctx, NULL, NULL);
8556 /* if we failed setting up the ctx, we might not have any rings */
8557 io_iopoll_try_reap_events(ctx);
8559 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8561 * Use system_unbound_wq to avoid spawning tons of event kworkers
8562 * if we're exiting a ton of rings at the same time. It just adds
8563 * noise and overhead, there's no discernable change in runtime
8564 * over using system_wq.
8566 queue_work(system_unbound_wq, &ctx->exit_work);
8569 static int io_uring_release(struct inode *inode, struct file *file)
8571 struct io_ring_ctx *ctx = file->private_data;
8573 file->private_data = NULL;
8574 io_ring_ctx_wait_and_kill(ctx);
8578 struct io_task_cancel {
8579 struct task_struct *task;
8580 struct files_struct *files;
8583 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8585 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8586 struct io_task_cancel *cancel = data;
8589 if (cancel->files && (req->flags & REQ_F_LINK_TIMEOUT)) {
8590 unsigned long flags;
8591 struct io_ring_ctx *ctx = req->ctx;
8593 /* protect against races with linked timeouts */
8594 spin_lock_irqsave(&ctx->completion_lock, flags);
8595 ret = io_match_task(req, cancel->task, cancel->files);
8596 spin_unlock_irqrestore(&ctx->completion_lock, flags);
8598 ret = io_match_task(req, cancel->task, cancel->files);
8603 static void io_cancel_defer_files(struct io_ring_ctx *ctx,
8604 struct task_struct *task,
8605 struct files_struct *files)
8607 struct io_defer_entry *de = NULL;
8610 spin_lock_irq(&ctx->completion_lock);
8611 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8612 if (io_match_task(de->req, task, files)) {
8613 list_cut_position(&list, &ctx->defer_list, &de->list);
8617 spin_unlock_irq(&ctx->completion_lock);
8619 while (!list_empty(&list)) {
8620 de = list_first_entry(&list, struct io_defer_entry, list);
8621 list_del_init(&de->list);
8622 req_set_fail_links(de->req);
8623 io_put_req(de->req);
8624 io_req_complete(de->req, -ECANCELED);
8629 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
8630 struct task_struct *task,
8631 struct files_struct *files)
8633 struct io_task_cancel cancel = { .task = task, .files = files, };
8634 struct task_struct *tctx_task = task ?: current;
8635 struct io_uring_task *tctx = tctx_task->io_uring;
8638 enum io_wq_cancel cret;
8641 if (tctx && tctx->io_wq) {
8642 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
8644 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8647 /* SQPOLL thread does its own polling */
8648 if (!(ctx->flags & IORING_SETUP_SQPOLL) && !files) {
8649 while (!list_empty_careful(&ctx->iopoll_list)) {
8650 io_iopoll_try_reap_events(ctx);
8655 ret |= io_poll_remove_all(ctx, task, files);
8656 ret |= io_kill_timeouts(ctx, task, files);
8657 ret |= io_run_task_work();
8658 ret |= io_run_ctx_fallback(ctx);
8659 io_cqring_overflow_flush(ctx, true, task, files);
8666 static int io_uring_count_inflight(struct io_ring_ctx *ctx,
8667 struct task_struct *task,
8668 struct files_struct *files)
8670 struct io_kiocb *req;
8673 spin_lock_irq(&ctx->inflight_lock);
8674 list_for_each_entry(req, &ctx->inflight_list, inflight_entry)
8675 cnt += io_match_task(req, task, files);
8676 spin_unlock_irq(&ctx->inflight_lock);
8680 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
8681 struct task_struct *task,
8682 struct files_struct *files)
8684 while (!list_empty_careful(&ctx->inflight_list)) {
8688 inflight = io_uring_count_inflight(ctx, task, files);
8692 io_uring_try_cancel_requests(ctx, task, files);
8695 io_sq_thread_unpark(ctx->sq_data);
8696 prepare_to_wait(&task->io_uring->wait, &wait,
8697 TASK_UNINTERRUPTIBLE);
8698 if (inflight == io_uring_count_inflight(ctx, task, files))
8700 finish_wait(&task->io_uring->wait, &wait);
8702 io_sq_thread_park(ctx->sq_data);
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;
8716 if ((ctx->flags & IORING_SETUP_SQPOLL) && ctx->sq_data) {
8717 /* never started, nothing to cancel */
8718 if (ctx->flags & IORING_SETUP_R_DISABLED) {
8719 io_sq_offload_start(ctx);
8722 io_sq_thread_park(ctx->sq_data);
8723 task = ctx->sq_data->thread;
8725 atomic_inc(&task->io_uring->in_idle);
8728 io_cancel_defer_files(ctx, task, files);
8730 io_uring_cancel_files(ctx, task, files);
8732 io_uring_try_cancel_requests(ctx, task, NULL);
8735 atomic_dec(&task->io_uring->in_idle);
8737 io_sq_thread_unpark(ctx->sq_data);
8741 * Note that this task has used io_uring. We use it for cancelation purposes.
8743 static int io_uring_add_task_file(struct io_ring_ctx *ctx, struct file *file)
8745 struct io_uring_task *tctx = current->io_uring;
8748 if (unlikely(!tctx)) {
8749 ret = io_uring_alloc_task_context(current, ctx);
8752 tctx = current->io_uring;
8754 if (tctx->last != file) {
8755 void *old = xa_load(&tctx->xa, (unsigned long)file);
8759 ret = xa_err(xa_store(&tctx->xa, (unsigned long)file,
8770 * This is race safe in that the task itself is doing this, hence it
8771 * cannot be going through the exit/cancel paths at the same time.
8772 * This cannot be modified while exit/cancel is running.
8774 if (!tctx->sqpoll && (ctx->flags & IORING_SETUP_SQPOLL))
8775 tctx->sqpoll = true;
8781 * Remove this io_uring_file -> task mapping.
8783 static void io_uring_del_task_file(struct file *file)
8785 struct io_uring_task *tctx = current->io_uring;
8787 if (tctx->last == file)
8789 file = xa_erase(&tctx->xa, (unsigned long)file);
8794 static void io_uring_clean_tctx(struct io_uring_task *tctx)
8797 unsigned long index;
8799 xa_for_each(&tctx->xa, index, file)
8800 io_uring_del_task_file(file);
8802 io_wq_put_and_exit(tctx->io_wq);
8807 void __io_uring_files_cancel(struct files_struct *files)
8809 struct io_uring_task *tctx = current->io_uring;
8811 unsigned long index;
8813 /* make sure overflow events are dropped */
8814 atomic_inc(&tctx->in_idle);
8815 xa_for_each(&tctx->xa, index, file)
8816 io_uring_cancel_task_requests(file->private_data, files);
8817 atomic_dec(&tctx->in_idle);
8820 io_uring_clean_tctx(tctx);
8823 static s64 tctx_inflight(struct io_uring_task *tctx)
8825 return percpu_counter_sum(&tctx->inflight);
8828 static void io_uring_cancel_sqpoll(struct io_ring_ctx *ctx)
8830 struct io_sq_data *sqd = ctx->sq_data;
8831 struct io_uring_task *tctx;
8837 io_sq_thread_park(sqd);
8838 if (!sqd->thread || !sqd->thread->io_uring) {
8839 io_sq_thread_unpark(sqd);
8842 tctx = ctx->sq_data->thread->io_uring;
8843 atomic_inc(&tctx->in_idle);
8845 /* read completions before cancelations */
8846 inflight = tctx_inflight(tctx);
8849 io_uring_cancel_task_requests(ctx, NULL);
8851 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
8853 * If we've seen completions, retry without waiting. This
8854 * avoids a race where a completion comes in before we did
8855 * prepare_to_wait().
8857 if (inflight == tctx_inflight(tctx))
8859 finish_wait(&tctx->wait, &wait);
8861 atomic_dec(&tctx->in_idle);
8862 io_sq_thread_unpark(sqd);
8866 * Find any io_uring fd that this task has registered or done IO on, and cancel
8869 void __io_uring_task_cancel(void)
8871 struct io_uring_task *tctx = current->io_uring;
8875 /* make sure overflow events are dropped */
8876 atomic_inc(&tctx->in_idle);
8880 unsigned long index;
8882 xa_for_each(&tctx->xa, index, file)
8883 io_uring_cancel_sqpoll(file->private_data);
8887 /* read completions before cancelations */
8888 inflight = tctx_inflight(tctx);
8891 __io_uring_files_cancel(NULL);
8893 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
8896 * If we've seen completions, retry without waiting. This
8897 * avoids a race where a completion comes in before we did
8898 * prepare_to_wait().
8900 if (inflight == tctx_inflight(tctx))
8902 finish_wait(&tctx->wait, &wait);
8905 atomic_dec(&tctx->in_idle);
8907 io_uring_clean_tctx(tctx);
8908 /* all current's requests should be gone, we can kill tctx */
8909 __io_uring_free(current);
8912 static void *io_uring_validate_mmap_request(struct file *file,
8913 loff_t pgoff, size_t sz)
8915 struct io_ring_ctx *ctx = file->private_data;
8916 loff_t offset = pgoff << PAGE_SHIFT;
8921 case IORING_OFF_SQ_RING:
8922 case IORING_OFF_CQ_RING:
8925 case IORING_OFF_SQES:
8929 return ERR_PTR(-EINVAL);
8932 page = virt_to_head_page(ptr);
8933 if (sz > page_size(page))
8934 return ERR_PTR(-EINVAL);
8941 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
8943 size_t sz = vma->vm_end - vma->vm_start;
8947 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
8949 return PTR_ERR(ptr);
8951 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
8952 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
8955 #else /* !CONFIG_MMU */
8957 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
8959 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
8962 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
8964 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
8967 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
8968 unsigned long addr, unsigned long len,
8969 unsigned long pgoff, unsigned long flags)
8973 ptr = io_uring_validate_mmap_request(file, pgoff, len);
8975 return PTR_ERR(ptr);
8977 return (unsigned long) ptr;
8980 #endif /* !CONFIG_MMU */
8982 static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
8988 if (!io_sqring_full(ctx))
8990 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
8992 if (!io_sqring_full(ctx))
8995 } while (!signal_pending(current));
8997 finish_wait(&ctx->sqo_sq_wait, &wait);
9001 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9002 struct __kernel_timespec __user **ts,
9003 const sigset_t __user **sig)
9005 struct io_uring_getevents_arg arg;
9008 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9009 * is just a pointer to the sigset_t.
9011 if (!(flags & IORING_ENTER_EXT_ARG)) {
9012 *sig = (const sigset_t __user *) argp;
9018 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9019 * timespec and sigset_t pointers if good.
9021 if (*argsz != sizeof(arg))
9023 if (copy_from_user(&arg, argp, sizeof(arg)))
9025 *sig = u64_to_user_ptr(arg.sigmask);
9026 *argsz = arg.sigmask_sz;
9027 *ts = u64_to_user_ptr(arg.ts);
9031 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9032 u32, min_complete, u32, flags, const void __user *, argp,
9035 struct io_ring_ctx *ctx;
9042 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9043 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG))
9051 if (f.file->f_op != &io_uring_fops)
9055 ctx = f.file->private_data;
9056 if (!percpu_ref_tryget(&ctx->refs))
9060 if (ctx->flags & IORING_SETUP_R_DISABLED)
9064 * For SQ polling, the thread will do all submissions and completions.
9065 * Just return the requested submit count, and wake the thread if
9069 if (ctx->flags & IORING_SETUP_SQPOLL) {
9070 io_cqring_overflow_flush(ctx, false, NULL, NULL);
9072 if (unlikely(ctx->sqo_exec)) {
9073 ret = io_sq_thread_fork(ctx->sq_data, ctx);
9079 if (flags & IORING_ENTER_SQ_WAKEUP)
9080 wake_up(&ctx->sq_data->wait);
9081 if (flags & IORING_ENTER_SQ_WAIT) {
9082 ret = io_sqpoll_wait_sq(ctx);
9086 submitted = to_submit;
9087 } else if (to_submit) {
9088 ret = io_uring_add_task_file(ctx, f.file);
9091 mutex_lock(&ctx->uring_lock);
9092 submitted = io_submit_sqes(ctx, to_submit);
9093 mutex_unlock(&ctx->uring_lock);
9095 if (submitted != to_submit)
9098 if (flags & IORING_ENTER_GETEVENTS) {
9099 const sigset_t __user *sig;
9100 struct __kernel_timespec __user *ts;
9102 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
9106 min_complete = min(min_complete, ctx->cq_entries);
9109 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9110 * space applications don't need to do io completion events
9111 * polling again, they can rely on io_sq_thread to do polling
9112 * work, which can reduce cpu usage and uring_lock contention.
9114 if (ctx->flags & IORING_SETUP_IOPOLL &&
9115 !(ctx->flags & IORING_SETUP_SQPOLL)) {
9116 ret = io_iopoll_check(ctx, min_complete);
9118 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
9123 percpu_ref_put(&ctx->refs);
9126 return submitted ? submitted : ret;
9129 #ifdef CONFIG_PROC_FS
9130 static int io_uring_show_cred(int id, void *p, void *data)
9132 const struct cred *cred = p;
9133 struct seq_file *m = data;
9134 struct user_namespace *uns = seq_user_ns(m);
9135 struct group_info *gi;
9140 seq_printf(m, "%5d\n", id);
9141 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
9142 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
9143 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
9144 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
9145 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
9146 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
9147 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
9148 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
9149 seq_puts(m, "\n\tGroups:\t");
9150 gi = cred->group_info;
9151 for (g = 0; g < gi->ngroups; g++) {
9152 seq_put_decimal_ull(m, g ? " " : "",
9153 from_kgid_munged(uns, gi->gid[g]));
9155 seq_puts(m, "\n\tCapEff:\t");
9156 cap = cred->cap_effective;
9157 CAP_FOR_EACH_U32(__capi)
9158 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
9163 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
9165 struct io_sq_data *sq = NULL;
9170 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9171 * since fdinfo case grabs it in the opposite direction of normal use
9172 * cases. If we fail to get the lock, we just don't iterate any
9173 * structures that could be going away outside the io_uring mutex.
9175 has_lock = mutex_trylock(&ctx->uring_lock);
9177 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) {
9183 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9184 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9185 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9186 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9187 struct file *f = *io_fixed_file_slot(ctx->file_data, i);
9190 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9192 seq_printf(m, "%5u: <none>\n", i);
9194 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9195 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9196 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
9198 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
9199 (unsigned int) buf->len);
9201 if (has_lock && !idr_is_empty(&ctx->personality_idr)) {
9202 seq_printf(m, "Personalities:\n");
9203 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
9205 seq_printf(m, "PollList:\n");
9206 spin_lock_irq(&ctx->completion_lock);
9207 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9208 struct hlist_head *list = &ctx->cancel_hash[i];
9209 struct io_kiocb *req;
9211 hlist_for_each_entry(req, list, hash_node)
9212 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9213 req->task->task_works != NULL);
9215 spin_unlock_irq(&ctx->completion_lock);
9217 mutex_unlock(&ctx->uring_lock);
9220 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9222 struct io_ring_ctx *ctx = f->private_data;
9224 if (percpu_ref_tryget(&ctx->refs)) {
9225 __io_uring_show_fdinfo(ctx, m);
9226 percpu_ref_put(&ctx->refs);
9231 static const struct file_operations io_uring_fops = {
9232 .release = io_uring_release,
9233 .mmap = io_uring_mmap,
9235 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9236 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9238 .poll = io_uring_poll,
9239 .fasync = io_uring_fasync,
9240 #ifdef CONFIG_PROC_FS
9241 .show_fdinfo = io_uring_show_fdinfo,
9245 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9246 struct io_uring_params *p)
9248 struct io_rings *rings;
9249 size_t size, sq_array_offset;
9251 /* make sure these are sane, as we already accounted them */
9252 ctx->sq_entries = p->sq_entries;
9253 ctx->cq_entries = p->cq_entries;
9255 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9256 if (size == SIZE_MAX)
9259 rings = io_mem_alloc(size);
9264 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9265 rings->sq_ring_mask = p->sq_entries - 1;
9266 rings->cq_ring_mask = p->cq_entries - 1;
9267 rings->sq_ring_entries = p->sq_entries;
9268 rings->cq_ring_entries = p->cq_entries;
9269 ctx->sq_mask = rings->sq_ring_mask;
9270 ctx->cq_mask = rings->cq_ring_mask;
9272 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9273 if (size == SIZE_MAX) {
9274 io_mem_free(ctx->rings);
9279 ctx->sq_sqes = io_mem_alloc(size);
9280 if (!ctx->sq_sqes) {
9281 io_mem_free(ctx->rings);
9289 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
9293 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9297 ret = io_uring_add_task_file(ctx, file);
9302 fd_install(fd, file);
9307 * Allocate an anonymous fd, this is what constitutes the application
9308 * visible backing of an io_uring instance. The application mmaps this
9309 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9310 * we have to tie this fd to a socket for file garbage collection purposes.
9312 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
9315 #if defined(CONFIG_UNIX)
9318 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9321 return ERR_PTR(ret);
9324 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9325 O_RDWR | O_CLOEXEC);
9326 #if defined(CONFIG_UNIX)
9328 sock_release(ctx->ring_sock);
9329 ctx->ring_sock = NULL;
9331 ctx->ring_sock->file = file;
9337 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9338 struct io_uring_params __user *params)
9340 struct io_ring_ctx *ctx;
9346 if (entries > IORING_MAX_ENTRIES) {
9347 if (!(p->flags & IORING_SETUP_CLAMP))
9349 entries = IORING_MAX_ENTRIES;
9353 * Use twice as many entries for the CQ ring. It's possible for the
9354 * application to drive a higher depth than the size of the SQ ring,
9355 * since the sqes are only used at submission time. This allows for
9356 * some flexibility in overcommitting a bit. If the application has
9357 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9358 * of CQ ring entries manually.
9360 p->sq_entries = roundup_pow_of_two(entries);
9361 if (p->flags & IORING_SETUP_CQSIZE) {
9363 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9364 * to a power-of-two, if it isn't already. We do NOT impose
9365 * any cq vs sq ring sizing.
9369 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9370 if (!(p->flags & IORING_SETUP_CLAMP))
9372 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9374 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9375 if (p->cq_entries < p->sq_entries)
9378 p->cq_entries = 2 * p->sq_entries;
9381 ctx = io_ring_ctx_alloc(p);
9384 ctx->compat = in_compat_syscall();
9385 if (!capable(CAP_IPC_LOCK))
9386 ctx->user = get_uid(current_user());
9389 * This is just grabbed for accounting purposes. When a process exits,
9390 * the mm is exited and dropped before the files, hence we need to hang
9391 * on to this mm purely for the purposes of being able to unaccount
9392 * memory (locked/pinned vm). It's not used for anything else.
9394 mmgrab(current->mm);
9395 ctx->mm_account = current->mm;
9397 ret = io_allocate_scq_urings(ctx, p);
9401 ret = io_sq_offload_create(ctx, p);
9405 if (!(p->flags & IORING_SETUP_R_DISABLED))
9406 io_sq_offload_start(ctx);
9408 memset(&p->sq_off, 0, sizeof(p->sq_off));
9409 p->sq_off.head = offsetof(struct io_rings, sq.head);
9410 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9411 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9412 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9413 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9414 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9415 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9417 memset(&p->cq_off, 0, sizeof(p->cq_off));
9418 p->cq_off.head = offsetof(struct io_rings, cq.head);
9419 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9420 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9421 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9422 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9423 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9424 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9426 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9427 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9428 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9429 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
9430 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS;
9432 if (copy_to_user(params, p, sizeof(*p))) {
9437 file = io_uring_get_file(ctx);
9439 ret = PTR_ERR(file);
9444 * Install ring fd as the very last thing, so we don't risk someone
9445 * having closed it before we finish setup
9447 ret = io_uring_install_fd(ctx, file);
9449 /* fput will clean it up */
9454 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9457 io_ring_ctx_wait_and_kill(ctx);
9462 * Sets up an aio uring context, and returns the fd. Applications asks for a
9463 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9464 * params structure passed in.
9466 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9468 struct io_uring_params p;
9471 if (copy_from_user(&p, params, sizeof(p)))
9473 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9478 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9479 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9480 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9481 IORING_SETUP_R_DISABLED))
9484 return io_uring_create(entries, &p, params);
9487 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9488 struct io_uring_params __user *, params)
9490 return io_uring_setup(entries, params);
9493 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9495 struct io_uring_probe *p;
9499 size = struct_size(p, ops, nr_args);
9500 if (size == SIZE_MAX)
9502 p = kzalloc(size, GFP_KERNEL);
9507 if (copy_from_user(p, arg, size))
9510 if (memchr_inv(p, 0, size))
9513 p->last_op = IORING_OP_LAST - 1;
9514 if (nr_args > IORING_OP_LAST)
9515 nr_args = IORING_OP_LAST;
9517 for (i = 0; i < nr_args; i++) {
9519 if (!io_op_defs[i].not_supported)
9520 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9525 if (copy_to_user(arg, p, size))
9532 static int io_register_personality(struct io_ring_ctx *ctx)
9534 const struct cred *creds;
9537 creds = get_current_cred();
9539 ret = idr_alloc_cyclic(&ctx->personality_idr, (void *) creds, 1,
9540 USHRT_MAX, GFP_KERNEL);
9546 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9547 unsigned int nr_args)
9549 struct io_uring_restriction *res;
9553 /* Restrictions allowed only if rings started disabled */
9554 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9557 /* We allow only a single restrictions registration */
9558 if (ctx->restrictions.registered)
9561 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9564 size = array_size(nr_args, sizeof(*res));
9565 if (size == SIZE_MAX)
9568 res = memdup_user(arg, size);
9570 return PTR_ERR(res);
9574 for (i = 0; i < nr_args; i++) {
9575 switch (res[i].opcode) {
9576 case IORING_RESTRICTION_REGISTER_OP:
9577 if (res[i].register_op >= IORING_REGISTER_LAST) {
9582 __set_bit(res[i].register_op,
9583 ctx->restrictions.register_op);
9585 case IORING_RESTRICTION_SQE_OP:
9586 if (res[i].sqe_op >= IORING_OP_LAST) {
9591 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9593 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9594 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9596 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9597 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9606 /* Reset all restrictions if an error happened */
9608 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9610 ctx->restrictions.registered = true;
9616 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9618 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9621 if (ctx->restrictions.registered)
9622 ctx->restricted = 1;
9624 io_sq_offload_start(ctx);
9628 static bool io_register_op_must_quiesce(int op)
9631 case IORING_UNREGISTER_FILES:
9632 case IORING_REGISTER_FILES_UPDATE:
9633 case IORING_REGISTER_PROBE:
9634 case IORING_REGISTER_PERSONALITY:
9635 case IORING_UNREGISTER_PERSONALITY:
9642 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
9643 void __user *arg, unsigned nr_args)
9644 __releases(ctx->uring_lock)
9645 __acquires(ctx->uring_lock)
9650 * We're inside the ring mutex, if the ref is already dying, then
9651 * someone else killed the ctx or is already going through
9652 * io_uring_register().
9654 if (percpu_ref_is_dying(&ctx->refs))
9657 if (io_register_op_must_quiesce(opcode)) {
9658 percpu_ref_kill(&ctx->refs);
9661 * Drop uring mutex before waiting for references to exit. If
9662 * another thread is currently inside io_uring_enter() it might
9663 * need to grab the uring_lock to make progress. If we hold it
9664 * here across the drain wait, then we can deadlock. It's safe
9665 * to drop the mutex here, since no new references will come in
9666 * after we've killed the percpu ref.
9668 mutex_unlock(&ctx->uring_lock);
9670 ret = wait_for_completion_interruptible(&ctx->ref_comp);
9673 ret = io_run_task_work_sig();
9678 mutex_lock(&ctx->uring_lock);
9681 percpu_ref_resurrect(&ctx->refs);
9686 if (ctx->restricted) {
9687 if (opcode >= IORING_REGISTER_LAST) {
9692 if (!test_bit(opcode, ctx->restrictions.register_op)) {
9699 case IORING_REGISTER_BUFFERS:
9700 ret = io_sqe_buffers_register(ctx, arg, nr_args);
9702 case IORING_UNREGISTER_BUFFERS:
9706 ret = io_sqe_buffers_unregister(ctx);
9708 case IORING_REGISTER_FILES:
9709 ret = io_sqe_files_register(ctx, arg, nr_args);
9711 case IORING_UNREGISTER_FILES:
9715 ret = io_sqe_files_unregister(ctx);
9717 case IORING_REGISTER_FILES_UPDATE:
9718 ret = io_sqe_files_update(ctx, arg, nr_args);
9720 case IORING_REGISTER_EVENTFD:
9721 case IORING_REGISTER_EVENTFD_ASYNC:
9725 ret = io_eventfd_register(ctx, arg);
9728 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
9729 ctx->eventfd_async = 1;
9731 ctx->eventfd_async = 0;
9733 case IORING_UNREGISTER_EVENTFD:
9737 ret = io_eventfd_unregister(ctx);
9739 case IORING_REGISTER_PROBE:
9741 if (!arg || nr_args > 256)
9743 ret = io_probe(ctx, arg, nr_args);
9745 case IORING_REGISTER_PERSONALITY:
9749 ret = io_register_personality(ctx);
9751 case IORING_UNREGISTER_PERSONALITY:
9755 ret = io_unregister_personality(ctx, nr_args);
9757 case IORING_REGISTER_ENABLE_RINGS:
9761 ret = io_register_enable_rings(ctx);
9763 case IORING_REGISTER_RESTRICTIONS:
9764 ret = io_register_restrictions(ctx, arg, nr_args);
9772 if (io_register_op_must_quiesce(opcode)) {
9773 /* bring the ctx back to life */
9774 percpu_ref_reinit(&ctx->refs);
9776 reinit_completion(&ctx->ref_comp);
9781 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
9782 void __user *, arg, unsigned int, nr_args)
9784 struct io_ring_ctx *ctx;
9793 if (f.file->f_op != &io_uring_fops)
9796 ctx = f.file->private_data;
9800 mutex_lock(&ctx->uring_lock);
9801 ret = __io_uring_register(ctx, opcode, arg, nr_args);
9802 mutex_unlock(&ctx->uring_lock);
9803 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
9804 ctx->cq_ev_fd != NULL, ret);
9810 static int __init io_uring_init(void)
9812 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
9813 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
9814 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
9817 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
9818 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
9819 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
9820 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
9821 BUILD_BUG_SQE_ELEM(1, __u8, flags);
9822 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
9823 BUILD_BUG_SQE_ELEM(4, __s32, fd);
9824 BUILD_BUG_SQE_ELEM(8, __u64, off);
9825 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
9826 BUILD_BUG_SQE_ELEM(16, __u64, addr);
9827 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
9828 BUILD_BUG_SQE_ELEM(24, __u32, len);
9829 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
9830 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
9831 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
9832 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
9833 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
9834 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
9835 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
9836 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
9837 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
9838 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
9839 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
9840 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
9841 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
9842 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
9843 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
9844 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
9845 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
9846 BUILD_BUG_SQE_ELEM(42, __u16, personality);
9847 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
9849 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
9850 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
9851 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
9855 __initcall(io_uring_init);