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 completion;
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 void io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force,
1455 struct task_struct *tsk,
1456 struct files_struct *files)
1458 if (test_bit(0, &ctx->cq_check_overflow)) {
1459 /* iopoll syncs against uring_lock, not completion_lock */
1460 if (ctx->flags & IORING_SETUP_IOPOLL)
1461 mutex_lock(&ctx->uring_lock);
1462 __io_cqring_overflow_flush(ctx, force, tsk, files);
1463 if (ctx->flags & IORING_SETUP_IOPOLL)
1464 mutex_unlock(&ctx->uring_lock);
1468 static void __io_cqring_fill_event(struct io_kiocb *req, long res, long cflags)
1470 struct io_ring_ctx *ctx = req->ctx;
1471 struct io_uring_cqe *cqe;
1473 trace_io_uring_complete(ctx, req->user_data, res);
1476 * If we can't get a cq entry, userspace overflowed the
1477 * submission (by quite a lot). Increment the overflow count in
1480 cqe = io_get_cqring(ctx);
1482 WRITE_ONCE(cqe->user_data, req->user_data);
1483 WRITE_ONCE(cqe->res, res);
1484 WRITE_ONCE(cqe->flags, cflags);
1485 } else if (ctx->cq_overflow_flushed ||
1486 atomic_read(&req->task->io_uring->in_idle)) {
1488 * If we're in ring overflow flush mode, or in task cancel mode,
1489 * then we cannot store the request for later flushing, we need
1490 * to drop it on the floor.
1492 ctx->cached_cq_overflow++;
1493 WRITE_ONCE(ctx->rings->cq_overflow, ctx->cached_cq_overflow);
1495 if (list_empty(&ctx->cq_overflow_list)) {
1496 set_bit(0, &ctx->sq_check_overflow);
1497 set_bit(0, &ctx->cq_check_overflow);
1498 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
1502 req->compl.cflags = cflags;
1503 refcount_inc(&req->refs);
1504 list_add_tail(&req->compl.list, &ctx->cq_overflow_list);
1508 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1510 __io_cqring_fill_event(req, res, 0);
1513 static inline void io_req_complete_post(struct io_kiocb *req, long res,
1514 unsigned int cflags)
1516 struct io_ring_ctx *ctx = req->ctx;
1517 unsigned long flags;
1519 spin_lock_irqsave(&ctx->completion_lock, flags);
1520 __io_cqring_fill_event(req, res, cflags);
1521 io_commit_cqring(ctx);
1523 * If we're the last reference to this request, add to our locked
1526 if (refcount_dec_and_test(&req->refs)) {
1527 struct io_comp_state *cs = &ctx->submit_state.comp;
1529 io_dismantle_req(req);
1530 io_put_task(req->task, 1);
1531 list_add(&req->compl.list, &cs->locked_free_list);
1532 cs->locked_free_nr++;
1535 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1537 io_cqring_ev_posted(ctx);
1540 percpu_ref_put(&ctx->refs);
1544 static void io_req_complete_state(struct io_kiocb *req, long res,
1545 unsigned int cflags)
1549 req->compl.cflags = cflags;
1550 req->flags |= REQ_F_COMPLETE_INLINE;
1553 static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags,
1554 long res, unsigned cflags)
1556 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1557 io_req_complete_state(req, res, cflags);
1559 io_req_complete_post(req, res, cflags);
1562 static inline void io_req_complete(struct io_kiocb *req, long res)
1564 __io_req_complete(req, 0, res, 0);
1567 static bool io_flush_cached_reqs(struct io_ring_ctx *ctx)
1569 struct io_submit_state *state = &ctx->submit_state;
1570 struct io_comp_state *cs = &state->comp;
1571 struct io_kiocb *req = NULL;
1574 * If we have more than a batch's worth of requests in our IRQ side
1575 * locked cache, grab the lock and move them over to our submission
1578 if (READ_ONCE(cs->locked_free_nr) > IO_COMPL_BATCH) {
1579 spin_lock_irq(&ctx->completion_lock);
1580 list_splice_init(&cs->locked_free_list, &cs->free_list);
1581 cs->locked_free_nr = 0;
1582 spin_unlock_irq(&ctx->completion_lock);
1585 while (!list_empty(&cs->free_list)) {
1586 req = list_first_entry(&cs->free_list, struct io_kiocb,
1588 list_del(&req->compl.list);
1589 state->reqs[state->free_reqs++] = req;
1590 if (state->free_reqs == ARRAY_SIZE(state->reqs))
1597 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
1599 struct io_submit_state *state = &ctx->submit_state;
1601 BUILD_BUG_ON(IO_REQ_ALLOC_BATCH > ARRAY_SIZE(state->reqs));
1603 if (!state->free_reqs) {
1604 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1607 if (io_flush_cached_reqs(ctx))
1610 ret = kmem_cache_alloc_bulk(req_cachep, gfp, IO_REQ_ALLOC_BATCH,
1614 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1615 * retry single alloc to be on the safe side.
1617 if (unlikely(ret <= 0)) {
1618 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1619 if (!state->reqs[0])
1623 state->free_reqs = ret;
1627 return state->reqs[state->free_reqs];
1630 static inline void io_put_file(struct io_kiocb *req, struct file *file,
1637 static void io_dismantle_req(struct io_kiocb *req)
1641 if (req->async_data)
1642 kfree(req->async_data);
1644 io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
1645 if (req->fixed_rsrc_refs)
1646 percpu_ref_put(req->fixed_rsrc_refs);
1648 if (req->flags & REQ_F_INFLIGHT) {
1649 struct io_ring_ctx *ctx = req->ctx;
1650 unsigned long flags;
1652 spin_lock_irqsave(&ctx->inflight_lock, flags);
1653 list_del(&req->inflight_entry);
1654 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
1655 req->flags &= ~REQ_F_INFLIGHT;
1659 /* must to be called somewhat shortly after putting a request */
1660 static inline void io_put_task(struct task_struct *task, int nr)
1662 struct io_uring_task *tctx = task->io_uring;
1664 percpu_counter_sub(&tctx->inflight, nr);
1665 if (unlikely(atomic_read(&tctx->in_idle)))
1666 wake_up(&tctx->wait);
1667 put_task_struct_many(task, nr);
1670 static void __io_free_req(struct io_kiocb *req)
1672 struct io_ring_ctx *ctx = req->ctx;
1674 io_dismantle_req(req);
1675 io_put_task(req->task, 1);
1677 kmem_cache_free(req_cachep, req);
1678 percpu_ref_put(&ctx->refs);
1681 static inline void io_remove_next_linked(struct io_kiocb *req)
1683 struct io_kiocb *nxt = req->link;
1685 req->link = nxt->link;
1689 static void io_kill_linked_timeout(struct io_kiocb *req)
1691 struct io_ring_ctx *ctx = req->ctx;
1692 struct io_kiocb *link;
1693 bool cancelled = false;
1694 unsigned long flags;
1696 spin_lock_irqsave(&ctx->completion_lock, flags);
1700 * Can happen if a linked timeout fired and link had been like
1701 * req -> link t-out -> link t-out [-> ...]
1703 if (link && (link->flags & REQ_F_LTIMEOUT_ACTIVE)) {
1704 struct io_timeout_data *io = link->async_data;
1707 io_remove_next_linked(req);
1708 link->timeout.head = NULL;
1709 ret = hrtimer_try_to_cancel(&io->timer);
1711 io_cqring_fill_event(link, -ECANCELED);
1712 io_commit_cqring(ctx);
1716 req->flags &= ~REQ_F_LINK_TIMEOUT;
1717 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1720 io_cqring_ev_posted(ctx);
1726 static void io_fail_links(struct io_kiocb *req)
1728 struct io_kiocb *link, *nxt;
1729 struct io_ring_ctx *ctx = req->ctx;
1730 unsigned long flags;
1732 spin_lock_irqsave(&ctx->completion_lock, flags);
1740 trace_io_uring_fail_link(req, link);
1741 io_cqring_fill_event(link, -ECANCELED);
1743 io_put_req_deferred(link, 2);
1746 io_commit_cqring(ctx);
1747 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1749 io_cqring_ev_posted(ctx);
1752 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1754 if (req->flags & REQ_F_LINK_TIMEOUT)
1755 io_kill_linked_timeout(req);
1758 * If LINK is set, we have dependent requests in this chain. If we
1759 * didn't fail this request, queue the first one up, moving any other
1760 * dependencies to the next request. In case of failure, fail the rest
1763 if (likely(!(req->flags & REQ_F_FAIL_LINK))) {
1764 struct io_kiocb *nxt = req->link;
1773 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1775 if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK))))
1777 return __io_req_find_next(req);
1780 static void ctx_flush_and_put(struct io_ring_ctx *ctx)
1784 if (ctx->submit_state.comp.nr) {
1785 mutex_lock(&ctx->uring_lock);
1786 io_submit_flush_completions(&ctx->submit_state.comp, ctx);
1787 mutex_unlock(&ctx->uring_lock);
1789 percpu_ref_put(&ctx->refs);
1792 static bool __tctx_task_work(struct io_uring_task *tctx)
1794 struct io_ring_ctx *ctx = NULL;
1795 struct io_wq_work_list list;
1796 struct io_wq_work_node *node;
1798 if (wq_list_empty(&tctx->task_list))
1801 spin_lock_irq(&tctx->task_lock);
1802 list = tctx->task_list;
1803 INIT_WQ_LIST(&tctx->task_list);
1804 spin_unlock_irq(&tctx->task_lock);
1808 struct io_wq_work_node *next = node->next;
1809 struct io_kiocb *req;
1811 req = container_of(node, struct io_kiocb, io_task_work.node);
1812 if (req->ctx != ctx) {
1813 ctx_flush_and_put(ctx);
1815 percpu_ref_get(&ctx->refs);
1818 req->task_work.func(&req->task_work);
1822 ctx_flush_and_put(ctx);
1823 return list.first != NULL;
1826 static void tctx_task_work(struct callback_head *cb)
1828 struct io_uring_task *tctx = container_of(cb, struct io_uring_task, task_work);
1830 clear_bit(0, &tctx->task_state);
1832 while (__tctx_task_work(tctx))
1836 static int io_task_work_add(struct task_struct *tsk, struct io_kiocb *req,
1837 enum task_work_notify_mode notify)
1839 struct io_uring_task *tctx = tsk->io_uring;
1840 struct io_wq_work_node *node, *prev;
1841 unsigned long flags;
1844 WARN_ON_ONCE(!tctx);
1846 spin_lock_irqsave(&tctx->task_lock, flags);
1847 wq_list_add_tail(&req->io_task_work.node, &tctx->task_list);
1848 spin_unlock_irqrestore(&tctx->task_lock, flags);
1850 /* task_work already pending, we're done */
1851 if (test_bit(0, &tctx->task_state) ||
1852 test_and_set_bit(0, &tctx->task_state))
1855 if (!task_work_add(tsk, &tctx->task_work, notify))
1859 * Slow path - we failed, find and delete work. if the work is not
1860 * in the list, it got run and we're fine.
1863 spin_lock_irqsave(&tctx->task_lock, flags);
1864 wq_list_for_each(node, prev, &tctx->task_list) {
1865 if (&req->io_task_work.node == node) {
1866 wq_list_del(&tctx->task_list, node, prev);
1871 spin_unlock_irqrestore(&tctx->task_lock, flags);
1872 clear_bit(0, &tctx->task_state);
1876 static int io_req_task_work_add(struct io_kiocb *req)
1878 struct task_struct *tsk = req->task;
1879 struct io_ring_ctx *ctx = req->ctx;
1880 enum task_work_notify_mode notify;
1883 if (tsk->flags & PF_EXITING)
1887 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
1888 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
1889 * processing task_work. There's no reliable way to tell if TWA_RESUME
1893 if (!(ctx->flags & IORING_SETUP_SQPOLL))
1894 notify = TWA_SIGNAL;
1896 ret = io_task_work_add(tsk, req, notify);
1898 wake_up_process(tsk);
1903 static void io_req_task_work_add_fallback(struct io_kiocb *req,
1904 task_work_func_t cb)
1906 struct io_ring_ctx *ctx = req->ctx;
1907 struct callback_head *head;
1909 init_task_work(&req->task_work, cb);
1911 head = READ_ONCE(ctx->exit_task_work);
1912 req->task_work.next = head;
1913 } while (cmpxchg(&ctx->exit_task_work, head, &req->task_work) != head);
1916 static void __io_req_task_cancel(struct io_kiocb *req, int error)
1918 struct io_ring_ctx *ctx = req->ctx;
1920 spin_lock_irq(&ctx->completion_lock);
1921 io_cqring_fill_event(req, error);
1922 io_commit_cqring(ctx);
1923 spin_unlock_irq(&ctx->completion_lock);
1925 io_cqring_ev_posted(ctx);
1926 req_set_fail_links(req);
1927 io_double_put_req(req);
1930 static void io_req_task_cancel(struct callback_head *cb)
1932 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1933 struct io_ring_ctx *ctx = req->ctx;
1935 mutex_lock(&ctx->uring_lock);
1936 __io_req_task_cancel(req, req->result);
1937 mutex_unlock(&ctx->uring_lock);
1938 percpu_ref_put(&ctx->refs);
1941 static void __io_req_task_submit(struct io_kiocb *req)
1943 struct io_ring_ctx *ctx = req->ctx;
1945 /* ctx stays valid until unlock, even if we drop all ours ctx->refs */
1946 mutex_lock(&ctx->uring_lock);
1947 if (!(current->flags & PF_EXITING) && !current->in_execve)
1948 __io_queue_sqe(req);
1950 __io_req_task_cancel(req, -EFAULT);
1951 mutex_unlock(&ctx->uring_lock);
1954 static void io_req_task_submit(struct callback_head *cb)
1956 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1958 __io_req_task_submit(req);
1961 static void io_req_task_queue(struct io_kiocb *req)
1965 req->task_work.func = io_req_task_submit;
1966 ret = io_req_task_work_add(req);
1967 if (unlikely(ret)) {
1968 req->result = -ECANCELED;
1969 percpu_ref_get(&req->ctx->refs);
1970 io_req_task_work_add_fallback(req, io_req_task_cancel);
1974 static void io_req_task_queue_fail(struct io_kiocb *req, int ret)
1976 percpu_ref_get(&req->ctx->refs);
1978 req->task_work.func = io_req_task_cancel;
1980 if (unlikely(io_req_task_work_add(req)))
1981 io_req_task_work_add_fallback(req, io_req_task_cancel);
1984 static inline void io_queue_next(struct io_kiocb *req)
1986 struct io_kiocb *nxt = io_req_find_next(req);
1989 io_req_task_queue(nxt);
1992 static void io_free_req(struct io_kiocb *req)
1999 struct task_struct *task;
2004 static inline void io_init_req_batch(struct req_batch *rb)
2011 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2012 struct req_batch *rb)
2015 io_put_task(rb->task, rb->task_refs);
2017 percpu_ref_put_many(&ctx->refs, rb->ctx_refs);
2020 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req,
2021 struct io_submit_state *state)
2025 if (req->task != rb->task) {
2027 io_put_task(rb->task, rb->task_refs);
2028 rb->task = req->task;
2034 io_dismantle_req(req);
2035 if (state->free_reqs != ARRAY_SIZE(state->reqs))
2036 state->reqs[state->free_reqs++] = req;
2038 list_add(&req->compl.list, &state->comp.free_list);
2041 static void io_submit_flush_completions(struct io_comp_state *cs,
2042 struct io_ring_ctx *ctx)
2045 struct io_kiocb *req;
2046 struct req_batch rb;
2048 io_init_req_batch(&rb);
2049 spin_lock_irq(&ctx->completion_lock);
2050 for (i = 0; i < nr; i++) {
2052 __io_cqring_fill_event(req, req->result, req->compl.cflags);
2054 io_commit_cqring(ctx);
2055 spin_unlock_irq(&ctx->completion_lock);
2057 io_cqring_ev_posted(ctx);
2058 for (i = 0; i < nr; i++) {
2061 /* submission and completion refs */
2062 if (refcount_sub_and_test(2, &req->refs))
2063 io_req_free_batch(&rb, req, &ctx->submit_state);
2066 io_req_free_batch_finish(ctx, &rb);
2071 * Drop reference to request, return next in chain (if there is one) if this
2072 * was the last reference to this request.
2074 static struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2076 struct io_kiocb *nxt = NULL;
2078 if (refcount_dec_and_test(&req->refs)) {
2079 nxt = io_req_find_next(req);
2085 static void io_put_req(struct io_kiocb *req)
2087 if (refcount_dec_and_test(&req->refs))
2091 static void io_put_req_deferred_cb(struct callback_head *cb)
2093 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2098 static void io_free_req_deferred(struct io_kiocb *req)
2102 req->task_work.func = io_put_req_deferred_cb;
2103 ret = io_req_task_work_add(req);
2105 io_req_task_work_add_fallback(req, io_put_req_deferred_cb);
2108 static inline void io_put_req_deferred(struct io_kiocb *req, int refs)
2110 if (refcount_sub_and_test(refs, &req->refs))
2111 io_free_req_deferred(req);
2114 static void io_double_put_req(struct io_kiocb *req)
2116 /* drop both submit and complete references */
2117 if (refcount_sub_and_test(2, &req->refs))
2121 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2123 /* See comment at the top of this file */
2125 return __io_cqring_events(ctx);
2128 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2130 struct io_rings *rings = ctx->rings;
2132 /* make sure SQ entry isn't read before tail */
2133 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2136 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2138 unsigned int cflags;
2140 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2141 cflags |= IORING_CQE_F_BUFFER;
2142 req->flags &= ~REQ_F_BUFFER_SELECTED;
2147 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2149 struct io_buffer *kbuf;
2151 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2152 return io_put_kbuf(req, kbuf);
2155 static inline bool io_run_task_work(void)
2158 * Not safe to run on exiting task, and the task_work handling will
2159 * not add work to such a task.
2161 if (unlikely(current->flags & PF_EXITING))
2163 if (current->task_works) {
2164 __set_current_state(TASK_RUNNING);
2173 * Find and free completed poll iocbs
2175 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2176 struct list_head *done)
2178 struct req_batch rb;
2179 struct io_kiocb *req;
2181 /* order with ->result store in io_complete_rw_iopoll() */
2184 io_init_req_batch(&rb);
2185 while (!list_empty(done)) {
2188 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2189 list_del(&req->inflight_entry);
2191 if (READ_ONCE(req->result) == -EAGAIN) {
2192 req->iopoll_completed = 0;
2193 if (io_rw_reissue(req))
2197 if (req->flags & REQ_F_BUFFER_SELECTED)
2198 cflags = io_put_rw_kbuf(req);
2200 __io_cqring_fill_event(req, req->result, cflags);
2203 if (refcount_dec_and_test(&req->refs))
2204 io_req_free_batch(&rb, req, &ctx->submit_state);
2207 io_commit_cqring(ctx);
2208 io_cqring_ev_posted_iopoll(ctx);
2209 io_req_free_batch_finish(ctx, &rb);
2212 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2215 struct io_kiocb *req, *tmp;
2221 * Only spin for completions if we don't have multiple devices hanging
2222 * off our complete list, and we're under the requested amount.
2224 spin = !ctx->poll_multi_file && *nr_events < min;
2227 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2228 struct kiocb *kiocb = &req->rw.kiocb;
2231 * Move completed and retryable entries to our local lists.
2232 * If we find a request that requires polling, break out
2233 * and complete those lists first, if we have entries there.
2235 if (READ_ONCE(req->iopoll_completed)) {
2236 list_move_tail(&req->inflight_entry, &done);
2239 if (!list_empty(&done))
2242 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2246 /* iopoll may have completed current req */
2247 if (READ_ONCE(req->iopoll_completed))
2248 list_move_tail(&req->inflight_entry, &done);
2255 if (!list_empty(&done))
2256 io_iopoll_complete(ctx, nr_events, &done);
2262 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2263 * non-spinning poll check - we'll still enter the driver poll loop, but only
2264 * as a non-spinning completion check.
2266 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
2269 while (!list_empty(&ctx->iopoll_list) && !need_resched()) {
2272 ret = io_do_iopoll(ctx, nr_events, min);
2275 if (*nr_events >= min)
2283 * We can't just wait for polled events to come to us, we have to actively
2284 * find and complete them.
2286 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2288 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2291 mutex_lock(&ctx->uring_lock);
2292 while (!list_empty(&ctx->iopoll_list)) {
2293 unsigned int nr_events = 0;
2295 io_do_iopoll(ctx, &nr_events, 0);
2297 /* let it sleep and repeat later if can't complete a request */
2301 * Ensure we allow local-to-the-cpu processing to take place,
2302 * in this case we need to ensure that we reap all events.
2303 * Also let task_work, etc. to progress by releasing the mutex
2305 if (need_resched()) {
2306 mutex_unlock(&ctx->uring_lock);
2308 mutex_lock(&ctx->uring_lock);
2311 mutex_unlock(&ctx->uring_lock);
2314 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2316 unsigned int nr_events = 0;
2317 int iters = 0, ret = 0;
2320 * We disallow the app entering submit/complete with polling, but we
2321 * still need to lock the ring to prevent racing with polled issue
2322 * that got punted to a workqueue.
2324 mutex_lock(&ctx->uring_lock);
2327 * Don't enter poll loop if we already have events pending.
2328 * If we do, we can potentially be spinning for commands that
2329 * already triggered a CQE (eg in error).
2331 if (test_bit(0, &ctx->cq_check_overflow))
2332 __io_cqring_overflow_flush(ctx, false, NULL, NULL);
2333 if (io_cqring_events(ctx))
2337 * If a submit got punted to a workqueue, we can have the
2338 * application entering polling for a command before it gets
2339 * issued. That app will hold the uring_lock for the duration
2340 * of the poll right here, so we need to take a breather every
2341 * now and then to ensure that the issue has a chance to add
2342 * the poll to the issued list. Otherwise we can spin here
2343 * forever, while the workqueue is stuck trying to acquire the
2346 if (!(++iters & 7)) {
2347 mutex_unlock(&ctx->uring_lock);
2349 mutex_lock(&ctx->uring_lock);
2352 ret = io_iopoll_getevents(ctx, &nr_events, min);
2356 } while (min && !nr_events && !need_resched());
2358 mutex_unlock(&ctx->uring_lock);
2362 static void kiocb_end_write(struct io_kiocb *req)
2365 * Tell lockdep we inherited freeze protection from submission
2368 if (req->flags & REQ_F_ISREG) {
2369 struct inode *inode = file_inode(req->file);
2371 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
2373 file_end_write(req->file);
2377 static bool io_resubmit_prep(struct io_kiocb *req)
2379 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2381 struct iov_iter iter;
2383 /* already prepared */
2384 if (req->async_data)
2387 switch (req->opcode) {
2388 case IORING_OP_READV:
2389 case IORING_OP_READ_FIXED:
2390 case IORING_OP_READ:
2393 case IORING_OP_WRITEV:
2394 case IORING_OP_WRITE_FIXED:
2395 case IORING_OP_WRITE:
2399 printk_once(KERN_WARNING "io_uring: bad opcode in resubmit %d\n",
2404 ret = io_import_iovec(rw, req, &iovec, &iter, false);
2407 return !io_setup_async_rw(req, iovec, inline_vecs, &iter, false);
2410 static bool io_rw_should_reissue(struct io_kiocb *req)
2412 umode_t mode = file_inode(req->file)->i_mode;
2413 struct io_ring_ctx *ctx = req->ctx;
2415 if (!S_ISBLK(mode) && !S_ISREG(mode))
2417 if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
2418 !(ctx->flags & IORING_SETUP_IOPOLL)))
2421 * If ref is dying, we might be running poll reap from the exit work.
2422 * Don't attempt to reissue from that path, just let it fail with
2425 if (percpu_ref_is_dying(&ctx->refs))
2431 static bool io_rw_reissue(struct io_kiocb *req)
2434 if (!io_rw_should_reissue(req))
2437 lockdep_assert_held(&req->ctx->uring_lock);
2439 if (io_resubmit_prep(req)) {
2440 refcount_inc(&req->refs);
2441 io_queue_async_work(req);
2444 req_set_fail_links(req);
2449 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2450 unsigned int issue_flags)
2454 if ((res == -EAGAIN || res == -EOPNOTSUPP) && io_rw_reissue(req))
2456 if (res != req->result)
2457 req_set_fail_links(req);
2459 if (req->rw.kiocb.ki_flags & IOCB_WRITE)
2460 kiocb_end_write(req);
2461 if (req->flags & REQ_F_BUFFER_SELECTED)
2462 cflags = io_put_rw_kbuf(req);
2463 __io_req_complete(req, issue_flags, res, cflags);
2466 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2468 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2470 __io_complete_rw(req, res, res2, 0);
2473 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2475 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2478 /* Rewind iter, if we have one. iopoll path resubmits as usual */
2479 if (res == -EAGAIN && io_rw_should_reissue(req)) {
2480 struct io_async_rw *rw = req->async_data;
2483 iov_iter_revert(&rw->iter,
2484 req->result - iov_iter_count(&rw->iter));
2485 else if (!io_resubmit_prep(req))
2490 if (kiocb->ki_flags & IOCB_WRITE)
2491 kiocb_end_write(req);
2493 if (res != -EAGAIN && res != req->result)
2494 req_set_fail_links(req);
2496 WRITE_ONCE(req->result, res);
2497 /* order with io_poll_complete() checking ->result */
2499 WRITE_ONCE(req->iopoll_completed, 1);
2503 * After the iocb has been issued, it's safe to be found on the poll list.
2504 * Adding the kiocb to the list AFTER submission ensures that we don't
2505 * find it from a io_iopoll_getevents() thread before the issuer is done
2506 * accessing the kiocb cookie.
2508 static void io_iopoll_req_issued(struct io_kiocb *req, bool in_async)
2510 struct io_ring_ctx *ctx = req->ctx;
2513 * Track whether we have multiple files in our lists. This will impact
2514 * how we do polling eventually, not spinning if we're on potentially
2515 * different devices.
2517 if (list_empty(&ctx->iopoll_list)) {
2518 ctx->poll_multi_file = false;
2519 } else if (!ctx->poll_multi_file) {
2520 struct io_kiocb *list_req;
2522 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2524 if (list_req->file != req->file)
2525 ctx->poll_multi_file = true;
2529 * For fast devices, IO may have already completed. If it has, add
2530 * it to the front so we find it first.
2532 if (READ_ONCE(req->iopoll_completed))
2533 list_add(&req->inflight_entry, &ctx->iopoll_list);
2535 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2538 * If IORING_SETUP_SQPOLL is enabled, sqes are either handled in sq thread
2539 * task context or in io worker task context. If current task context is
2540 * sq thread, we don't need to check whether should wake up sq thread.
2542 if (in_async && (ctx->flags & IORING_SETUP_SQPOLL) &&
2543 wq_has_sleeper(&ctx->sq_data->wait))
2544 wake_up(&ctx->sq_data->wait);
2547 static inline void io_state_file_put(struct io_submit_state *state)
2549 if (state->file_refs) {
2550 fput_many(state->file, state->file_refs);
2551 state->file_refs = 0;
2556 * Get as many references to a file as we have IOs left in this submission,
2557 * assuming most submissions are for one file, or at least that each file
2558 * has more than one submission.
2560 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2565 if (state->file_refs) {
2566 if (state->fd == fd) {
2570 io_state_file_put(state);
2572 state->file = fget_many(fd, state->ios_left);
2573 if (unlikely(!state->file))
2577 state->file_refs = state->ios_left - 1;
2581 static bool io_bdev_nowait(struct block_device *bdev)
2583 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2587 * If we tracked the file through the SCM inflight mechanism, we could support
2588 * any file. For now, just ensure that anything potentially problematic is done
2591 static bool io_file_supports_async(struct file *file, int rw)
2593 umode_t mode = file_inode(file)->i_mode;
2595 if (S_ISBLK(mode)) {
2596 if (IS_ENABLED(CONFIG_BLOCK) &&
2597 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2601 if (S_ISCHR(mode) || S_ISSOCK(mode))
2603 if (S_ISREG(mode)) {
2604 if (IS_ENABLED(CONFIG_BLOCK) &&
2605 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2606 file->f_op != &io_uring_fops)
2611 /* any ->read/write should understand O_NONBLOCK */
2612 if (file->f_flags & O_NONBLOCK)
2615 if (!(file->f_mode & FMODE_NOWAIT))
2619 return file->f_op->read_iter != NULL;
2621 return file->f_op->write_iter != NULL;
2624 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2626 struct io_ring_ctx *ctx = req->ctx;
2627 struct kiocb *kiocb = &req->rw.kiocb;
2628 struct file *file = req->file;
2632 if (S_ISREG(file_inode(file)->i_mode))
2633 req->flags |= REQ_F_ISREG;
2635 kiocb->ki_pos = READ_ONCE(sqe->off);
2636 if (kiocb->ki_pos == -1 && !(file->f_mode & FMODE_STREAM)) {
2637 req->flags |= REQ_F_CUR_POS;
2638 kiocb->ki_pos = file->f_pos;
2640 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2641 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2642 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2646 /* don't allow async punt for O_NONBLOCK or RWF_NOWAIT */
2647 if ((kiocb->ki_flags & IOCB_NOWAIT) || (file->f_flags & O_NONBLOCK))
2648 req->flags |= REQ_F_NOWAIT;
2650 ioprio = READ_ONCE(sqe->ioprio);
2652 ret = ioprio_check_cap(ioprio);
2656 kiocb->ki_ioprio = ioprio;
2658 kiocb->ki_ioprio = get_current_ioprio();
2660 if (ctx->flags & IORING_SETUP_IOPOLL) {
2661 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2662 !kiocb->ki_filp->f_op->iopoll)
2665 kiocb->ki_flags |= IOCB_HIPRI;
2666 kiocb->ki_complete = io_complete_rw_iopoll;
2667 req->iopoll_completed = 0;
2669 if (kiocb->ki_flags & IOCB_HIPRI)
2671 kiocb->ki_complete = io_complete_rw;
2674 req->rw.addr = READ_ONCE(sqe->addr);
2675 req->rw.len = READ_ONCE(sqe->len);
2676 req->buf_index = READ_ONCE(sqe->buf_index);
2680 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2686 case -ERESTARTNOINTR:
2687 case -ERESTARTNOHAND:
2688 case -ERESTART_RESTARTBLOCK:
2690 * We can't just restart the syscall, since previously
2691 * submitted sqes may already be in progress. Just fail this
2697 kiocb->ki_complete(kiocb, ret, 0);
2701 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2702 unsigned int issue_flags)
2704 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2705 struct io_async_rw *io = req->async_data;
2707 /* add previously done IO, if any */
2708 if (io && io->bytes_done > 0) {
2710 ret = io->bytes_done;
2712 ret += io->bytes_done;
2715 if (req->flags & REQ_F_CUR_POS)
2716 req->file->f_pos = kiocb->ki_pos;
2717 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2718 __io_complete_rw(req, ret, 0, issue_flags);
2720 io_rw_done(kiocb, ret);
2723 static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
2725 struct io_ring_ctx *ctx = req->ctx;
2726 size_t len = req->rw.len;
2727 struct io_mapped_ubuf *imu;
2728 u16 index, buf_index = req->buf_index;
2732 if (unlikely(buf_index >= ctx->nr_user_bufs))
2734 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2735 imu = &ctx->user_bufs[index];
2736 buf_addr = req->rw.addr;
2739 if (buf_addr + len < buf_addr)
2741 /* not inside the mapped region */
2742 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2746 * May not be a start of buffer, set size appropriately
2747 * and advance us to the beginning.
2749 offset = buf_addr - imu->ubuf;
2750 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2754 * Don't use iov_iter_advance() here, as it's really slow for
2755 * using the latter parts of a big fixed buffer - it iterates
2756 * over each segment manually. We can cheat a bit here, because
2759 * 1) it's a BVEC iter, we set it up
2760 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2761 * first and last bvec
2763 * So just find our index, and adjust the iterator afterwards.
2764 * If the offset is within the first bvec (or the whole first
2765 * bvec, just use iov_iter_advance(). This makes it easier
2766 * since we can just skip the first segment, which may not
2767 * be PAGE_SIZE aligned.
2769 const struct bio_vec *bvec = imu->bvec;
2771 if (offset <= bvec->bv_len) {
2772 iov_iter_advance(iter, offset);
2774 unsigned long seg_skip;
2776 /* skip first vec */
2777 offset -= bvec->bv_len;
2778 seg_skip = 1 + (offset >> PAGE_SHIFT);
2780 iter->bvec = bvec + seg_skip;
2781 iter->nr_segs -= seg_skip;
2782 iter->count -= bvec->bv_len + offset;
2783 iter->iov_offset = offset & ~PAGE_MASK;
2790 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2793 mutex_unlock(&ctx->uring_lock);
2796 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2799 * "Normal" inline submissions always hold the uring_lock, since we
2800 * grab it from the system call. Same is true for the SQPOLL offload.
2801 * The only exception is when we've detached the request and issue it
2802 * from an async worker thread, grab the lock for that case.
2805 mutex_lock(&ctx->uring_lock);
2808 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2809 int bgid, struct io_buffer *kbuf,
2812 struct io_buffer *head;
2814 if (req->flags & REQ_F_BUFFER_SELECTED)
2817 io_ring_submit_lock(req->ctx, needs_lock);
2819 lockdep_assert_held(&req->ctx->uring_lock);
2821 head = idr_find(&req->ctx->io_buffer_idr, bgid);
2823 if (!list_empty(&head->list)) {
2824 kbuf = list_last_entry(&head->list, struct io_buffer,
2826 list_del(&kbuf->list);
2829 idr_remove(&req->ctx->io_buffer_idr, bgid);
2831 if (*len > kbuf->len)
2834 kbuf = ERR_PTR(-ENOBUFS);
2837 io_ring_submit_unlock(req->ctx, needs_lock);
2842 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2845 struct io_buffer *kbuf;
2848 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2849 bgid = req->buf_index;
2850 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2853 req->rw.addr = (u64) (unsigned long) kbuf;
2854 req->flags |= REQ_F_BUFFER_SELECTED;
2855 return u64_to_user_ptr(kbuf->addr);
2858 #ifdef CONFIG_COMPAT
2859 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2862 struct compat_iovec __user *uiov;
2863 compat_ssize_t clen;
2867 uiov = u64_to_user_ptr(req->rw.addr);
2868 if (!access_ok(uiov, sizeof(*uiov)))
2870 if (__get_user(clen, &uiov->iov_len))
2876 buf = io_rw_buffer_select(req, &len, needs_lock);
2878 return PTR_ERR(buf);
2879 iov[0].iov_base = buf;
2880 iov[0].iov_len = (compat_size_t) len;
2885 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2888 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2892 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2895 len = iov[0].iov_len;
2898 buf = io_rw_buffer_select(req, &len, needs_lock);
2900 return PTR_ERR(buf);
2901 iov[0].iov_base = buf;
2902 iov[0].iov_len = len;
2906 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2909 if (req->flags & REQ_F_BUFFER_SELECTED) {
2910 struct io_buffer *kbuf;
2912 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2913 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
2914 iov[0].iov_len = kbuf->len;
2917 if (req->rw.len != 1)
2920 #ifdef CONFIG_COMPAT
2921 if (req->ctx->compat)
2922 return io_compat_import(req, iov, needs_lock);
2925 return __io_iov_buffer_select(req, iov, needs_lock);
2928 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
2929 struct iov_iter *iter, bool needs_lock)
2931 void __user *buf = u64_to_user_ptr(req->rw.addr);
2932 size_t sqe_len = req->rw.len;
2933 u8 opcode = req->opcode;
2936 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
2938 return io_import_fixed(req, rw, iter);
2941 /* buffer index only valid with fixed read/write, or buffer select */
2942 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
2945 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
2946 if (req->flags & REQ_F_BUFFER_SELECT) {
2947 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
2949 return PTR_ERR(buf);
2950 req->rw.len = sqe_len;
2953 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
2958 if (req->flags & REQ_F_BUFFER_SELECT) {
2959 ret = io_iov_buffer_select(req, *iovec, needs_lock);
2961 iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
2966 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
2970 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
2972 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
2976 * For files that don't have ->read_iter() and ->write_iter(), handle them
2977 * by looping over ->read() or ->write() manually.
2979 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
2981 struct kiocb *kiocb = &req->rw.kiocb;
2982 struct file *file = req->file;
2986 * Don't support polled IO through this interface, and we can't
2987 * support non-blocking either. For the latter, this just causes
2988 * the kiocb to be handled from an async context.
2990 if (kiocb->ki_flags & IOCB_HIPRI)
2992 if (kiocb->ki_flags & IOCB_NOWAIT)
2995 while (iov_iter_count(iter)) {
2999 if (!iov_iter_is_bvec(iter)) {
3000 iovec = iov_iter_iovec(iter);
3002 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3003 iovec.iov_len = req->rw.len;
3007 nr = file->f_op->read(file, iovec.iov_base,
3008 iovec.iov_len, io_kiocb_ppos(kiocb));
3010 nr = file->f_op->write(file, iovec.iov_base,
3011 iovec.iov_len, io_kiocb_ppos(kiocb));
3020 if (nr != iovec.iov_len)
3024 iov_iter_advance(iter, nr);
3030 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3031 const struct iovec *fast_iov, struct iov_iter *iter)
3033 struct io_async_rw *rw = req->async_data;
3035 memcpy(&rw->iter, iter, sizeof(*iter));
3036 rw->free_iovec = iovec;
3038 /* can only be fixed buffers, no need to do anything */
3039 if (iov_iter_is_bvec(iter))
3042 unsigned iov_off = 0;
3044 rw->iter.iov = rw->fast_iov;
3045 if (iter->iov != fast_iov) {
3046 iov_off = iter->iov - fast_iov;
3047 rw->iter.iov += iov_off;
3049 if (rw->fast_iov != fast_iov)
3050 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3051 sizeof(struct iovec) * iter->nr_segs);
3053 req->flags |= REQ_F_NEED_CLEANUP;
3057 static inline int __io_alloc_async_data(struct io_kiocb *req)
3059 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3060 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3061 return req->async_data == NULL;
3064 static int io_alloc_async_data(struct io_kiocb *req)
3066 if (!io_op_defs[req->opcode].needs_async_data)
3069 return __io_alloc_async_data(req);
3072 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3073 const struct iovec *fast_iov,
3074 struct iov_iter *iter, bool force)
3076 if (!force && !io_op_defs[req->opcode].needs_async_data)
3078 if (!req->async_data) {
3079 if (__io_alloc_async_data(req)) {
3084 io_req_map_rw(req, iovec, fast_iov, iter);
3089 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3091 struct io_async_rw *iorw = req->async_data;
3092 struct iovec *iov = iorw->fast_iov;
3095 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3096 if (unlikely(ret < 0))
3099 iorw->bytes_done = 0;
3100 iorw->free_iovec = iov;
3102 req->flags |= REQ_F_NEED_CLEANUP;
3106 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3108 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3110 return io_prep_rw(req, sqe);
3114 * This is our waitqueue callback handler, registered through lock_page_async()
3115 * when we initially tried to do the IO with the iocb armed our waitqueue.
3116 * This gets called when the page is unlocked, and we generally expect that to
3117 * happen when the page IO is completed and the page is now uptodate. This will
3118 * queue a task_work based retry of the operation, attempting to copy the data
3119 * again. If the latter fails because the page was NOT uptodate, then we will
3120 * do a thread based blocking retry of the operation. That's the unexpected
3123 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3124 int sync, void *arg)
3126 struct wait_page_queue *wpq;
3127 struct io_kiocb *req = wait->private;
3128 struct wait_page_key *key = arg;
3130 wpq = container_of(wait, struct wait_page_queue, wait);
3132 if (!wake_page_match(wpq, key))
3135 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3136 list_del_init(&wait->entry);
3138 /* submit ref gets dropped, acquire a new one */
3139 refcount_inc(&req->refs);
3140 io_req_task_queue(req);
3145 * This controls whether a given IO request should be armed for async page
3146 * based retry. If we return false here, the request is handed to the async
3147 * worker threads for retry. If we're doing buffered reads on a regular file,
3148 * we prepare a private wait_page_queue entry and retry the operation. This
3149 * will either succeed because the page is now uptodate and unlocked, or it
3150 * will register a callback when the page is unlocked at IO completion. Through
3151 * that callback, io_uring uses task_work to setup a retry of the operation.
3152 * That retry will attempt the buffered read again. The retry will generally
3153 * succeed, or in rare cases where it fails, we then fall back to using the
3154 * async worker threads for a blocking retry.
3156 static bool io_rw_should_retry(struct io_kiocb *req)
3158 struct io_async_rw *rw = req->async_data;
3159 struct wait_page_queue *wait = &rw->wpq;
3160 struct kiocb *kiocb = &req->rw.kiocb;
3162 /* never retry for NOWAIT, we just complete with -EAGAIN */
3163 if (req->flags & REQ_F_NOWAIT)
3166 /* Only for buffered IO */
3167 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3171 * just use poll if we can, and don't attempt if the fs doesn't
3172 * support callback based unlocks
3174 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3177 wait->wait.func = io_async_buf_func;
3178 wait->wait.private = req;
3179 wait->wait.flags = 0;
3180 INIT_LIST_HEAD(&wait->wait.entry);
3181 kiocb->ki_flags |= IOCB_WAITQ;
3182 kiocb->ki_flags &= ~IOCB_NOWAIT;
3183 kiocb->ki_waitq = wait;
3187 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3189 if (req->file->f_op->read_iter)
3190 return call_read_iter(req->file, &req->rw.kiocb, iter);
3191 else if (req->file->f_op->read)
3192 return loop_rw_iter(READ, req, iter);
3197 static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3199 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3200 struct kiocb *kiocb = &req->rw.kiocb;
3201 struct iov_iter __iter, *iter = &__iter;
3202 struct io_async_rw *rw = req->async_data;
3203 ssize_t io_size, ret, ret2;
3204 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3210 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3214 io_size = iov_iter_count(iter);
3215 req->result = io_size;
3217 /* Ensure we clear previously set non-block flag */
3218 if (!force_nonblock)
3219 kiocb->ki_flags &= ~IOCB_NOWAIT;
3221 kiocb->ki_flags |= IOCB_NOWAIT;
3223 /* If the file doesn't support async, just async punt */
3224 if (force_nonblock && !io_file_supports_async(req->file, READ)) {
3225 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3226 return ret ?: -EAGAIN;
3229 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size);
3230 if (unlikely(ret)) {
3235 ret = io_iter_do_read(req, iter);
3237 if (ret == -EIOCBQUEUED) {
3238 if (req->async_data)
3239 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3241 } else if (ret == -EAGAIN) {
3242 /* IOPOLL retry should happen for io-wq threads */
3243 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3245 /* no retry on NONBLOCK nor RWF_NOWAIT */
3246 if (req->flags & REQ_F_NOWAIT)
3248 /* some cases will consume bytes even on error returns */
3249 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3251 } else if (ret <= 0 || ret == io_size || !force_nonblock ||
3252 (req->flags & REQ_F_NOWAIT) || !(req->flags & REQ_F_ISREG)) {
3253 /* read all, failed, already did sync or don't want to retry */
3257 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3262 rw = req->async_data;
3263 /* now use our persistent iterator, if we aren't already */
3268 rw->bytes_done += ret;
3269 /* if we can retry, do so with the callbacks armed */
3270 if (!io_rw_should_retry(req)) {
3271 kiocb->ki_flags &= ~IOCB_WAITQ;
3276 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3277 * we get -EIOCBQUEUED, then we'll get a notification when the
3278 * desired page gets unlocked. We can also get a partial read
3279 * here, and if we do, then just retry at the new offset.
3281 ret = io_iter_do_read(req, iter);
3282 if (ret == -EIOCBQUEUED)
3284 /* we got some bytes, but not all. retry. */
3285 } while (ret > 0 && ret < io_size);
3287 kiocb_done(kiocb, ret, issue_flags);
3289 /* it's faster to check here then delegate to kfree */
3295 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3297 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3299 return io_prep_rw(req, sqe);
3302 static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3304 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3305 struct kiocb *kiocb = &req->rw.kiocb;
3306 struct iov_iter __iter, *iter = &__iter;
3307 struct io_async_rw *rw = req->async_data;
3308 ssize_t ret, ret2, io_size;
3309 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3315 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3319 io_size = iov_iter_count(iter);
3320 req->result = io_size;
3322 /* Ensure we clear previously set non-block flag */
3323 if (!force_nonblock)
3324 kiocb->ki_flags &= ~IOCB_NOWAIT;
3326 kiocb->ki_flags |= IOCB_NOWAIT;
3328 /* If the file doesn't support async, just async punt */
3329 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
3332 /* file path doesn't support NOWAIT for non-direct_IO */
3333 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3334 (req->flags & REQ_F_ISREG))
3337 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size);
3342 * Open-code file_start_write here to grab freeze protection,
3343 * which will be released by another thread in
3344 * io_complete_rw(). Fool lockdep by telling it the lock got
3345 * released so that it doesn't complain about the held lock when
3346 * we return to userspace.
3348 if (req->flags & REQ_F_ISREG) {
3349 sb_start_write(file_inode(req->file)->i_sb);
3350 __sb_writers_release(file_inode(req->file)->i_sb,
3353 kiocb->ki_flags |= IOCB_WRITE;
3355 if (req->file->f_op->write_iter)
3356 ret2 = call_write_iter(req->file, kiocb, iter);
3357 else if (req->file->f_op->write)
3358 ret2 = loop_rw_iter(WRITE, req, iter);
3363 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3364 * retry them without IOCB_NOWAIT.
3366 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3368 /* no retry on NONBLOCK nor RWF_NOWAIT */
3369 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3371 if (ret2 == -EIOCBQUEUED && req->async_data)
3372 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3373 if (!force_nonblock || ret2 != -EAGAIN) {
3374 /* IOPOLL retry should happen for io-wq threads */
3375 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3378 kiocb_done(kiocb, ret2, issue_flags);
3381 /* some cases will consume bytes even on error returns */
3382 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3383 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3384 return ret ?: -EAGAIN;
3387 /* it's reportedly faster than delegating the null check to kfree() */
3393 static int io_renameat_prep(struct io_kiocb *req,
3394 const struct io_uring_sqe *sqe)
3396 struct io_rename *ren = &req->rename;
3397 const char __user *oldf, *newf;
3399 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3402 ren->old_dfd = READ_ONCE(sqe->fd);
3403 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3404 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3405 ren->new_dfd = READ_ONCE(sqe->len);
3406 ren->flags = READ_ONCE(sqe->rename_flags);
3408 ren->oldpath = getname(oldf);
3409 if (IS_ERR(ren->oldpath))
3410 return PTR_ERR(ren->oldpath);
3412 ren->newpath = getname(newf);
3413 if (IS_ERR(ren->newpath)) {
3414 putname(ren->oldpath);
3415 return PTR_ERR(ren->newpath);
3418 req->flags |= REQ_F_NEED_CLEANUP;
3422 static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3424 struct io_rename *ren = &req->rename;
3427 if (issue_flags & IO_URING_F_NONBLOCK)
3430 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3431 ren->newpath, ren->flags);
3433 req->flags &= ~REQ_F_NEED_CLEANUP;
3435 req_set_fail_links(req);
3436 io_req_complete(req, ret);
3440 static int io_unlinkat_prep(struct io_kiocb *req,
3441 const struct io_uring_sqe *sqe)
3443 struct io_unlink *un = &req->unlink;
3444 const char __user *fname;
3446 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3449 un->dfd = READ_ONCE(sqe->fd);
3451 un->flags = READ_ONCE(sqe->unlink_flags);
3452 if (un->flags & ~AT_REMOVEDIR)
3455 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3456 un->filename = getname(fname);
3457 if (IS_ERR(un->filename))
3458 return PTR_ERR(un->filename);
3460 req->flags |= REQ_F_NEED_CLEANUP;
3464 static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3466 struct io_unlink *un = &req->unlink;
3469 if (issue_flags & IO_URING_F_NONBLOCK)
3472 if (un->flags & AT_REMOVEDIR)
3473 ret = do_rmdir(un->dfd, un->filename);
3475 ret = do_unlinkat(un->dfd, un->filename);
3477 req->flags &= ~REQ_F_NEED_CLEANUP;
3479 req_set_fail_links(req);
3480 io_req_complete(req, ret);
3484 static int io_shutdown_prep(struct io_kiocb *req,
3485 const struct io_uring_sqe *sqe)
3487 #if defined(CONFIG_NET)
3488 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3490 if (sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3494 req->shutdown.how = READ_ONCE(sqe->len);
3501 static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3503 #if defined(CONFIG_NET)
3504 struct socket *sock;
3507 if (issue_flags & IO_URING_F_NONBLOCK)
3510 sock = sock_from_file(req->file);
3511 if (unlikely(!sock))
3514 ret = __sys_shutdown_sock(sock, req->shutdown.how);
3516 req_set_fail_links(req);
3517 io_req_complete(req, ret);
3524 static int __io_splice_prep(struct io_kiocb *req,
3525 const struct io_uring_sqe *sqe)
3527 struct io_splice* sp = &req->splice;
3528 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3530 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3534 sp->len = READ_ONCE(sqe->len);
3535 sp->flags = READ_ONCE(sqe->splice_flags);
3537 if (unlikely(sp->flags & ~valid_flags))
3540 sp->file_in = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in),
3541 (sp->flags & SPLICE_F_FD_IN_FIXED));
3544 req->flags |= REQ_F_NEED_CLEANUP;
3546 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
3548 * Splice operation will be punted aync, and here need to
3549 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3551 req->work.flags |= IO_WQ_WORK_UNBOUND;
3557 static int io_tee_prep(struct io_kiocb *req,
3558 const struct io_uring_sqe *sqe)
3560 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3562 return __io_splice_prep(req, sqe);
3565 static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
3567 struct io_splice *sp = &req->splice;
3568 struct file *in = sp->file_in;
3569 struct file *out = sp->file_out;
3570 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3573 if (issue_flags & IO_URING_F_NONBLOCK)
3576 ret = do_tee(in, out, sp->len, flags);
3578 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3579 req->flags &= ~REQ_F_NEED_CLEANUP;
3582 req_set_fail_links(req);
3583 io_req_complete(req, ret);
3587 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3589 struct io_splice* sp = &req->splice;
3591 sp->off_in = READ_ONCE(sqe->splice_off_in);
3592 sp->off_out = READ_ONCE(sqe->off);
3593 return __io_splice_prep(req, sqe);
3596 static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
3598 struct io_splice *sp = &req->splice;
3599 struct file *in = sp->file_in;
3600 struct file *out = sp->file_out;
3601 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3602 loff_t *poff_in, *poff_out;
3605 if (issue_flags & IO_URING_F_NONBLOCK)
3608 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3609 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3612 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3614 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3615 req->flags &= ~REQ_F_NEED_CLEANUP;
3618 req_set_fail_links(req);
3619 io_req_complete(req, ret);
3624 * IORING_OP_NOP just posts a completion event, nothing else.
3626 static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
3628 struct io_ring_ctx *ctx = req->ctx;
3630 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3633 __io_req_complete(req, issue_flags, 0, 0);
3637 static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3639 struct io_ring_ctx *ctx = req->ctx;
3644 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3646 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3649 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3650 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3653 req->sync.off = READ_ONCE(sqe->off);
3654 req->sync.len = READ_ONCE(sqe->len);
3658 static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
3660 loff_t end = req->sync.off + req->sync.len;
3663 /* fsync always requires a blocking context */
3664 if (issue_flags & IO_URING_F_NONBLOCK)
3667 ret = vfs_fsync_range(req->file, req->sync.off,
3668 end > 0 ? end : LLONG_MAX,
3669 req->sync.flags & IORING_FSYNC_DATASYNC);
3671 req_set_fail_links(req);
3672 io_req_complete(req, ret);
3676 static int io_fallocate_prep(struct io_kiocb *req,
3677 const struct io_uring_sqe *sqe)
3679 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3681 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3684 req->sync.off = READ_ONCE(sqe->off);
3685 req->sync.len = READ_ONCE(sqe->addr);
3686 req->sync.mode = READ_ONCE(sqe->len);
3690 static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
3694 /* fallocate always requiring blocking context */
3695 if (issue_flags & IO_URING_F_NONBLOCK)
3697 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3700 req_set_fail_links(req);
3701 io_req_complete(req, ret);
3705 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3707 const char __user *fname;
3710 if (unlikely(sqe->ioprio || sqe->buf_index))
3712 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3715 /* open.how should be already initialised */
3716 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3717 req->open.how.flags |= O_LARGEFILE;
3719 req->open.dfd = READ_ONCE(sqe->fd);
3720 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3721 req->open.filename = getname(fname);
3722 if (IS_ERR(req->open.filename)) {
3723 ret = PTR_ERR(req->open.filename);
3724 req->open.filename = NULL;
3727 req->open.nofile = rlimit(RLIMIT_NOFILE);
3728 req->flags |= REQ_F_NEED_CLEANUP;
3732 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3736 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3738 mode = READ_ONCE(sqe->len);
3739 flags = READ_ONCE(sqe->open_flags);
3740 req->open.how = build_open_how(flags, mode);
3741 return __io_openat_prep(req, sqe);
3744 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3746 struct open_how __user *how;
3750 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3752 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3753 len = READ_ONCE(sqe->len);
3754 if (len < OPEN_HOW_SIZE_VER0)
3757 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3762 return __io_openat_prep(req, sqe);
3765 static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
3767 struct open_flags op;
3770 bool resolve_nonblock;
3773 ret = build_open_flags(&req->open.how, &op);
3776 nonblock_set = op.open_flag & O_NONBLOCK;
3777 resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
3778 if (issue_flags & IO_URING_F_NONBLOCK) {
3780 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
3781 * it'll always -EAGAIN
3783 if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
3785 op.lookup_flags |= LOOKUP_CACHED;
3786 op.open_flag |= O_NONBLOCK;
3789 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3793 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3794 /* only retry if RESOLVE_CACHED wasn't already set by application */
3795 if ((!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)) &&
3796 file == ERR_PTR(-EAGAIN)) {
3798 * We could hang on to this 'fd', but seems like marginal
3799 * gain for something that is now known to be a slower path.
3800 * So just put it, and we'll get a new one when we retry.
3808 ret = PTR_ERR(file);
3810 if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
3811 file->f_flags &= ~O_NONBLOCK;
3812 fsnotify_open(file);
3813 fd_install(ret, file);
3816 putname(req->open.filename);
3817 req->flags &= ~REQ_F_NEED_CLEANUP;
3819 req_set_fail_links(req);
3820 io_req_complete(req, ret);
3824 static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
3826 return io_openat2(req, issue_flags & IO_URING_F_NONBLOCK);
3829 static int io_remove_buffers_prep(struct io_kiocb *req,
3830 const struct io_uring_sqe *sqe)
3832 struct io_provide_buf *p = &req->pbuf;
3835 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3838 tmp = READ_ONCE(sqe->fd);
3839 if (!tmp || tmp > USHRT_MAX)
3842 memset(p, 0, sizeof(*p));
3844 p->bgid = READ_ONCE(sqe->buf_group);
3848 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3849 int bgid, unsigned nbufs)
3853 /* shouldn't happen */
3857 /* the head kbuf is the list itself */
3858 while (!list_empty(&buf->list)) {
3859 struct io_buffer *nxt;
3861 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3862 list_del(&nxt->list);
3869 idr_remove(&ctx->io_buffer_idr, bgid);
3874 static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
3876 struct io_provide_buf *p = &req->pbuf;
3877 struct io_ring_ctx *ctx = req->ctx;
3878 struct io_buffer *head;
3880 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3882 io_ring_submit_lock(ctx, !force_nonblock);
3884 lockdep_assert_held(&ctx->uring_lock);
3887 head = idr_find(&ctx->io_buffer_idr, p->bgid);
3889 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3891 req_set_fail_links(req);
3893 /* need to hold the lock to complete IOPOLL requests */
3894 if (ctx->flags & IORING_SETUP_IOPOLL) {
3895 __io_req_complete(req, issue_flags, ret, 0);
3896 io_ring_submit_unlock(ctx, !force_nonblock);
3898 io_ring_submit_unlock(ctx, !force_nonblock);
3899 __io_req_complete(req, issue_flags, ret, 0);
3904 static int io_provide_buffers_prep(struct io_kiocb *req,
3905 const struct io_uring_sqe *sqe)
3907 struct io_provide_buf *p = &req->pbuf;
3910 if (sqe->ioprio || sqe->rw_flags)
3913 tmp = READ_ONCE(sqe->fd);
3914 if (!tmp || tmp > USHRT_MAX)
3917 p->addr = READ_ONCE(sqe->addr);
3918 p->len = READ_ONCE(sqe->len);
3920 if (!access_ok(u64_to_user_ptr(p->addr), (p->len * p->nbufs)))
3923 p->bgid = READ_ONCE(sqe->buf_group);
3924 tmp = READ_ONCE(sqe->off);
3925 if (tmp > USHRT_MAX)
3931 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3933 struct io_buffer *buf;
3934 u64 addr = pbuf->addr;
3935 int i, bid = pbuf->bid;
3937 for (i = 0; i < pbuf->nbufs; i++) {
3938 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
3943 buf->len = pbuf->len;
3948 INIT_LIST_HEAD(&buf->list);
3951 list_add_tail(&buf->list, &(*head)->list);
3955 return i ? i : -ENOMEM;
3958 static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
3960 struct io_provide_buf *p = &req->pbuf;
3961 struct io_ring_ctx *ctx = req->ctx;
3962 struct io_buffer *head, *list;
3964 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3966 io_ring_submit_lock(ctx, !force_nonblock);
3968 lockdep_assert_held(&ctx->uring_lock);
3970 list = head = idr_find(&ctx->io_buffer_idr, p->bgid);
3972 ret = io_add_buffers(p, &head);
3977 ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1,
3980 __io_remove_buffers(ctx, head, p->bgid, -1U);
3986 req_set_fail_links(req);
3988 /* need to hold the lock to complete IOPOLL requests */
3989 if (ctx->flags & IORING_SETUP_IOPOLL) {
3990 __io_req_complete(req, issue_flags, ret, 0);
3991 io_ring_submit_unlock(ctx, !force_nonblock);
3993 io_ring_submit_unlock(ctx, !force_nonblock);
3994 __io_req_complete(req, issue_flags, ret, 0);
3999 static int io_epoll_ctl_prep(struct io_kiocb *req,
4000 const struct io_uring_sqe *sqe)
4002 #if defined(CONFIG_EPOLL)
4003 if (sqe->ioprio || sqe->buf_index)
4005 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4008 req->epoll.epfd = READ_ONCE(sqe->fd);
4009 req->epoll.op = READ_ONCE(sqe->len);
4010 req->epoll.fd = READ_ONCE(sqe->off);
4012 if (ep_op_has_event(req->epoll.op)) {
4013 struct epoll_event __user *ev;
4015 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4016 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4026 static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4028 #if defined(CONFIG_EPOLL)
4029 struct io_epoll *ie = &req->epoll;
4031 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4033 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4034 if (force_nonblock && ret == -EAGAIN)
4038 req_set_fail_links(req);
4039 __io_req_complete(req, issue_flags, ret, 0);
4046 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4048 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4049 if (sqe->ioprio || sqe->buf_index || sqe->off)
4051 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4054 req->madvise.addr = READ_ONCE(sqe->addr);
4055 req->madvise.len = READ_ONCE(sqe->len);
4056 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4063 static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4065 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4066 struct io_madvise *ma = &req->madvise;
4069 if (issue_flags & IO_URING_F_NONBLOCK)
4072 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4074 req_set_fail_links(req);
4075 io_req_complete(req, ret);
4082 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4084 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4086 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4089 req->fadvise.offset = READ_ONCE(sqe->off);
4090 req->fadvise.len = READ_ONCE(sqe->len);
4091 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4095 static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4097 struct io_fadvise *fa = &req->fadvise;
4100 if (issue_flags & IO_URING_F_NONBLOCK) {
4101 switch (fa->advice) {
4102 case POSIX_FADV_NORMAL:
4103 case POSIX_FADV_RANDOM:
4104 case POSIX_FADV_SEQUENTIAL:
4111 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4113 req_set_fail_links(req);
4114 io_req_complete(req, ret);
4118 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4120 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4122 if (sqe->ioprio || sqe->buf_index)
4124 if (req->flags & REQ_F_FIXED_FILE)
4127 req->statx.dfd = READ_ONCE(sqe->fd);
4128 req->statx.mask = READ_ONCE(sqe->len);
4129 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4130 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4131 req->statx.flags = READ_ONCE(sqe->statx_flags);
4136 static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4138 struct io_statx *ctx = &req->statx;
4141 if (issue_flags & IO_URING_F_NONBLOCK) {
4142 /* only need file table for an actual valid fd */
4143 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
4144 req->flags |= REQ_F_NO_FILE_TABLE;
4148 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4152 req_set_fail_links(req);
4153 io_req_complete(req, ret);
4157 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4159 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4161 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4162 sqe->rw_flags || sqe->buf_index)
4164 if (req->flags & REQ_F_FIXED_FILE)
4167 req->close.fd = READ_ONCE(sqe->fd);
4171 static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4173 struct files_struct *files = current->files;
4174 struct io_close *close = &req->close;
4175 struct fdtable *fdt;
4181 spin_lock(&files->file_lock);
4182 fdt = files_fdtable(files);
4183 if (close->fd >= fdt->max_fds) {
4184 spin_unlock(&files->file_lock);
4187 file = fdt->fd[close->fd];
4189 spin_unlock(&files->file_lock);
4193 if (file->f_op == &io_uring_fops) {
4194 spin_unlock(&files->file_lock);
4199 /* if the file has a flush method, be safe and punt to async */
4200 if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4201 spin_unlock(&files->file_lock);
4205 ret = __close_fd_get_file(close->fd, &file);
4206 spin_unlock(&files->file_lock);
4213 /* No ->flush() or already async, safely close from here */
4214 ret = filp_close(file, current->files);
4217 req_set_fail_links(req);
4220 __io_req_complete(req, issue_flags, ret, 0);
4224 static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4226 struct io_ring_ctx *ctx = req->ctx;
4228 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4230 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4233 req->sync.off = READ_ONCE(sqe->off);
4234 req->sync.len = READ_ONCE(sqe->len);
4235 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4239 static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4243 /* sync_file_range always requires a blocking context */
4244 if (issue_flags & IO_URING_F_NONBLOCK)
4247 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4250 req_set_fail_links(req);
4251 io_req_complete(req, ret);
4255 #if defined(CONFIG_NET)
4256 static int io_setup_async_msg(struct io_kiocb *req,
4257 struct io_async_msghdr *kmsg)
4259 struct io_async_msghdr *async_msg = req->async_data;
4263 if (io_alloc_async_data(req)) {
4264 kfree(kmsg->free_iov);
4267 async_msg = req->async_data;
4268 req->flags |= REQ_F_NEED_CLEANUP;
4269 memcpy(async_msg, kmsg, sizeof(*kmsg));
4270 async_msg->msg.msg_name = &async_msg->addr;
4271 /* if were using fast_iov, set it to the new one */
4272 if (!async_msg->free_iov)
4273 async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4278 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4279 struct io_async_msghdr *iomsg)
4281 iomsg->msg.msg_name = &iomsg->addr;
4282 iomsg->free_iov = iomsg->fast_iov;
4283 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4284 req->sr_msg.msg_flags, &iomsg->free_iov);
4287 static int io_sendmsg_prep_async(struct io_kiocb *req)
4291 if (!io_op_defs[req->opcode].needs_async_data)
4293 ret = io_sendmsg_copy_hdr(req, req->async_data);
4295 req->flags |= REQ_F_NEED_CLEANUP;
4299 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4301 struct io_sr_msg *sr = &req->sr_msg;
4303 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4306 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4307 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4308 sr->len = READ_ONCE(sqe->len);
4310 #ifdef CONFIG_COMPAT
4311 if (req->ctx->compat)
4312 sr->msg_flags |= MSG_CMSG_COMPAT;
4317 static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4319 struct io_async_msghdr iomsg, *kmsg;
4320 struct socket *sock;
4324 sock = sock_from_file(req->file);
4325 if (unlikely(!sock))
4328 kmsg = req->async_data;
4330 ret = io_sendmsg_copy_hdr(req, &iomsg);
4336 flags = req->sr_msg.msg_flags;
4337 if (flags & MSG_DONTWAIT)
4338 req->flags |= REQ_F_NOWAIT;
4339 else if (issue_flags & IO_URING_F_NONBLOCK)
4340 flags |= MSG_DONTWAIT;
4342 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4343 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4344 return io_setup_async_msg(req, kmsg);
4345 if (ret == -ERESTARTSYS)
4348 /* fast path, check for non-NULL to avoid function call */
4350 kfree(kmsg->free_iov);
4351 req->flags &= ~REQ_F_NEED_CLEANUP;
4353 req_set_fail_links(req);
4354 __io_req_complete(req, issue_flags, ret, 0);
4358 static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4360 struct io_sr_msg *sr = &req->sr_msg;
4363 struct socket *sock;
4367 sock = sock_from_file(req->file);
4368 if (unlikely(!sock))
4371 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4375 msg.msg_name = NULL;
4376 msg.msg_control = NULL;
4377 msg.msg_controllen = 0;
4378 msg.msg_namelen = 0;
4380 flags = req->sr_msg.msg_flags;
4381 if (flags & MSG_DONTWAIT)
4382 req->flags |= REQ_F_NOWAIT;
4383 else if (issue_flags & IO_URING_F_NONBLOCK)
4384 flags |= MSG_DONTWAIT;
4386 msg.msg_flags = flags;
4387 ret = sock_sendmsg(sock, &msg);
4388 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4390 if (ret == -ERESTARTSYS)
4394 req_set_fail_links(req);
4395 __io_req_complete(req, issue_flags, ret, 0);
4399 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4400 struct io_async_msghdr *iomsg)
4402 struct io_sr_msg *sr = &req->sr_msg;
4403 struct iovec __user *uiov;
4407 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4408 &iomsg->uaddr, &uiov, &iov_len);
4412 if (req->flags & REQ_F_BUFFER_SELECT) {
4415 if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4417 sr->len = iomsg->fast_iov[0].iov_len;
4418 iomsg->free_iov = NULL;
4420 iomsg->free_iov = iomsg->fast_iov;
4421 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4422 &iomsg->free_iov, &iomsg->msg.msg_iter,
4431 #ifdef CONFIG_COMPAT
4432 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4433 struct io_async_msghdr *iomsg)
4435 struct compat_msghdr __user *msg_compat;
4436 struct io_sr_msg *sr = &req->sr_msg;
4437 struct compat_iovec __user *uiov;
4442 msg_compat = (struct compat_msghdr __user *) sr->umsg;
4443 ret = __get_compat_msghdr(&iomsg->msg, msg_compat, &iomsg->uaddr,
4448 uiov = compat_ptr(ptr);
4449 if (req->flags & REQ_F_BUFFER_SELECT) {
4450 compat_ssize_t clen;
4454 if (!access_ok(uiov, sizeof(*uiov)))
4456 if (__get_user(clen, &uiov->iov_len))
4461 iomsg->free_iov = NULL;
4463 iomsg->free_iov = iomsg->fast_iov;
4464 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4465 UIO_FASTIOV, &iomsg->free_iov,
4466 &iomsg->msg.msg_iter, true);
4475 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4476 struct io_async_msghdr *iomsg)
4478 iomsg->msg.msg_name = &iomsg->addr;
4480 #ifdef CONFIG_COMPAT
4481 if (req->ctx->compat)
4482 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4485 return __io_recvmsg_copy_hdr(req, iomsg);
4488 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4491 struct io_sr_msg *sr = &req->sr_msg;
4492 struct io_buffer *kbuf;
4494 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4499 req->flags |= REQ_F_BUFFER_SELECTED;
4503 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4505 return io_put_kbuf(req, req->sr_msg.kbuf);
4508 static int io_recvmsg_prep_async(struct io_kiocb *req)
4512 if (!io_op_defs[req->opcode].needs_async_data)
4514 ret = io_recvmsg_copy_hdr(req, req->async_data);
4516 req->flags |= REQ_F_NEED_CLEANUP;
4520 static int io_recvmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4522 struct io_sr_msg *sr = &req->sr_msg;
4524 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4527 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4528 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4529 sr->len = READ_ONCE(sqe->len);
4530 sr->bgid = READ_ONCE(sqe->buf_group);
4532 #ifdef CONFIG_COMPAT
4533 if (req->ctx->compat)
4534 sr->msg_flags |= MSG_CMSG_COMPAT;
4539 static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
4541 struct io_async_msghdr iomsg, *kmsg;
4542 struct socket *sock;
4543 struct io_buffer *kbuf;
4545 int ret, cflags = 0;
4546 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4548 sock = sock_from_file(req->file);
4549 if (unlikely(!sock))
4552 kmsg = req->async_data;
4554 ret = io_recvmsg_copy_hdr(req, &iomsg);
4560 if (req->flags & REQ_F_BUFFER_SELECT) {
4561 kbuf = io_recv_buffer_select(req, !force_nonblock);
4563 return PTR_ERR(kbuf);
4564 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4565 kmsg->fast_iov[0].iov_len = req->sr_msg.len;
4566 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
4567 1, req->sr_msg.len);
4570 flags = req->sr_msg.msg_flags;
4571 if (flags & MSG_DONTWAIT)
4572 req->flags |= REQ_F_NOWAIT;
4573 else if (force_nonblock)
4574 flags |= MSG_DONTWAIT;
4576 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4577 kmsg->uaddr, flags);
4578 if (force_nonblock && ret == -EAGAIN)
4579 return io_setup_async_msg(req, kmsg);
4580 if (ret == -ERESTARTSYS)
4583 if (req->flags & REQ_F_BUFFER_SELECTED)
4584 cflags = io_put_recv_kbuf(req);
4585 /* fast path, check for non-NULL to avoid function call */
4587 kfree(kmsg->free_iov);
4588 req->flags &= ~REQ_F_NEED_CLEANUP;
4590 req_set_fail_links(req);
4591 __io_req_complete(req, issue_flags, ret, cflags);
4595 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
4597 struct io_buffer *kbuf;
4598 struct io_sr_msg *sr = &req->sr_msg;
4600 void __user *buf = sr->buf;
4601 struct socket *sock;
4604 int ret, cflags = 0;
4605 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4607 sock = sock_from_file(req->file);
4608 if (unlikely(!sock))
4611 if (req->flags & REQ_F_BUFFER_SELECT) {
4612 kbuf = io_recv_buffer_select(req, !force_nonblock);
4614 return PTR_ERR(kbuf);
4615 buf = u64_to_user_ptr(kbuf->addr);
4618 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4622 msg.msg_name = NULL;
4623 msg.msg_control = NULL;
4624 msg.msg_controllen = 0;
4625 msg.msg_namelen = 0;
4626 msg.msg_iocb = NULL;
4629 flags = req->sr_msg.msg_flags;
4630 if (flags & MSG_DONTWAIT)
4631 req->flags |= REQ_F_NOWAIT;
4632 else if (force_nonblock)
4633 flags |= MSG_DONTWAIT;
4635 ret = sock_recvmsg(sock, &msg, flags);
4636 if (force_nonblock && ret == -EAGAIN)
4638 if (ret == -ERESTARTSYS)
4641 if (req->flags & REQ_F_BUFFER_SELECTED)
4642 cflags = io_put_recv_kbuf(req);
4644 req_set_fail_links(req);
4645 __io_req_complete(req, issue_flags, ret, cflags);
4649 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4651 struct io_accept *accept = &req->accept;
4653 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4655 if (sqe->ioprio || sqe->len || sqe->buf_index)
4658 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4659 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4660 accept->flags = READ_ONCE(sqe->accept_flags);
4661 accept->nofile = rlimit(RLIMIT_NOFILE);
4665 static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
4667 struct io_accept *accept = &req->accept;
4668 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4669 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4672 if (req->file->f_flags & O_NONBLOCK)
4673 req->flags |= REQ_F_NOWAIT;
4675 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4676 accept->addr_len, accept->flags,
4678 if (ret == -EAGAIN && force_nonblock)
4681 if (ret == -ERESTARTSYS)
4683 req_set_fail_links(req);
4685 __io_req_complete(req, issue_flags, ret, 0);
4689 static int io_connect_prep_async(struct io_kiocb *req)
4691 struct io_async_connect *io = req->async_data;
4692 struct io_connect *conn = &req->connect;
4694 return move_addr_to_kernel(conn->addr, conn->addr_len, &io->address);
4697 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4699 struct io_connect *conn = &req->connect;
4701 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4703 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4706 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4707 conn->addr_len = READ_ONCE(sqe->addr2);
4711 static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
4713 struct io_async_connect __io, *io;
4714 unsigned file_flags;
4716 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4718 if (req->async_data) {
4719 io = req->async_data;
4721 ret = move_addr_to_kernel(req->connect.addr,
4722 req->connect.addr_len,
4729 file_flags = force_nonblock ? O_NONBLOCK : 0;
4731 ret = __sys_connect_file(req->file, &io->address,
4732 req->connect.addr_len, file_flags);
4733 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4734 if (req->async_data)
4736 if (io_alloc_async_data(req)) {
4740 io = req->async_data;
4741 memcpy(req->async_data, &__io, sizeof(__io));
4744 if (ret == -ERESTARTSYS)
4748 req_set_fail_links(req);
4749 __io_req_complete(req, issue_flags, ret, 0);
4752 #else /* !CONFIG_NET */
4753 #define IO_NETOP_FN(op) \
4754 static int io_##op(struct io_kiocb *req, unsigned int issue_flags) \
4756 return -EOPNOTSUPP; \
4759 #define IO_NETOP_PREP(op) \
4761 static int io_##op##_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) \
4763 return -EOPNOTSUPP; \
4766 #define IO_NETOP_PREP_ASYNC(op) \
4768 static int io_##op##_prep_async(struct io_kiocb *req) \
4770 return -EOPNOTSUPP; \
4773 IO_NETOP_PREP_ASYNC(sendmsg);
4774 IO_NETOP_PREP_ASYNC(recvmsg);
4775 IO_NETOP_PREP_ASYNC(connect);
4776 IO_NETOP_PREP(accept);
4779 #endif /* CONFIG_NET */
4781 struct io_poll_table {
4782 struct poll_table_struct pt;
4783 struct io_kiocb *req;
4787 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4788 __poll_t mask, task_work_func_t func)
4792 /* for instances that support it check for an event match first: */
4793 if (mask && !(mask & poll->events))
4796 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4798 list_del_init(&poll->wait.entry);
4801 req->task_work.func = func;
4802 percpu_ref_get(&req->ctx->refs);
4805 * If this fails, then the task is exiting. When a task exits, the
4806 * work gets canceled, so just cancel this request as well instead
4807 * of executing it. We can't safely execute it anyway, as we may not
4808 * have the needed state needed for it anyway.
4810 ret = io_req_task_work_add(req);
4811 if (unlikely(ret)) {
4812 WRITE_ONCE(poll->canceled, true);
4813 io_req_task_work_add_fallback(req, func);
4818 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4819 __acquires(&req->ctx->completion_lock)
4821 struct io_ring_ctx *ctx = req->ctx;
4823 if (!req->result && !READ_ONCE(poll->canceled)) {
4824 struct poll_table_struct pt = { ._key = poll->events };
4826 req->result = vfs_poll(req->file, &pt) & poll->events;
4829 spin_lock_irq(&ctx->completion_lock);
4830 if (!req->result && !READ_ONCE(poll->canceled)) {
4831 add_wait_queue(poll->head, &poll->wait);
4838 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4840 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4841 if (req->opcode == IORING_OP_POLL_ADD)
4842 return req->async_data;
4843 return req->apoll->double_poll;
4846 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4848 if (req->opcode == IORING_OP_POLL_ADD)
4850 return &req->apoll->poll;
4853 static void io_poll_remove_double(struct io_kiocb *req)
4855 struct io_poll_iocb *poll = io_poll_get_double(req);
4857 lockdep_assert_held(&req->ctx->completion_lock);
4859 if (poll && poll->head) {
4860 struct wait_queue_head *head = poll->head;
4862 spin_lock(&head->lock);
4863 list_del_init(&poll->wait.entry);
4864 if (poll->wait.private)
4865 refcount_dec(&req->refs);
4867 spin_unlock(&head->lock);
4871 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
4873 struct io_ring_ctx *ctx = req->ctx;
4875 io_poll_remove_double(req);
4876 req->poll.done = true;
4877 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
4878 io_commit_cqring(ctx);
4881 static void io_poll_task_func(struct callback_head *cb)
4883 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4884 struct io_ring_ctx *ctx = req->ctx;
4885 struct io_kiocb *nxt;
4887 if (io_poll_rewait(req, &req->poll)) {
4888 spin_unlock_irq(&ctx->completion_lock);
4890 hash_del(&req->hash_node);
4891 io_poll_complete(req, req->result, 0);
4892 spin_unlock_irq(&ctx->completion_lock);
4894 nxt = io_put_req_find_next(req);
4895 io_cqring_ev_posted(ctx);
4897 __io_req_task_submit(nxt);
4900 percpu_ref_put(&ctx->refs);
4903 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4904 int sync, void *key)
4906 struct io_kiocb *req = wait->private;
4907 struct io_poll_iocb *poll = io_poll_get_single(req);
4908 __poll_t mask = key_to_poll(key);
4910 /* for instances that support it check for an event match first: */
4911 if (mask && !(mask & poll->events))
4914 list_del_init(&wait->entry);
4916 if (poll && poll->head) {
4919 spin_lock(&poll->head->lock);
4920 done = list_empty(&poll->wait.entry);
4922 list_del_init(&poll->wait.entry);
4923 /* make sure double remove sees this as being gone */
4924 wait->private = NULL;
4925 spin_unlock(&poll->head->lock);
4927 /* use wait func handler, so it matches the rq type */
4928 poll->wait.func(&poll->wait, mode, sync, key);
4931 refcount_dec(&req->refs);
4935 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
4936 wait_queue_func_t wake_func)
4940 poll->canceled = false;
4941 poll->events = events;
4942 INIT_LIST_HEAD(&poll->wait.entry);
4943 init_waitqueue_func_entry(&poll->wait, wake_func);
4946 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
4947 struct wait_queue_head *head,
4948 struct io_poll_iocb **poll_ptr)
4950 struct io_kiocb *req = pt->req;
4953 * If poll->head is already set, it's because the file being polled
4954 * uses multiple waitqueues for poll handling (eg one for read, one
4955 * for write). Setup a separate io_poll_iocb if this happens.
4957 if (unlikely(poll->head)) {
4958 struct io_poll_iocb *poll_one = poll;
4960 /* already have a 2nd entry, fail a third attempt */
4962 pt->error = -EINVAL;
4965 /* double add on the same waitqueue head, ignore */
4966 if (poll->head == head)
4968 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
4970 pt->error = -ENOMEM;
4973 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
4974 refcount_inc(&req->refs);
4975 poll->wait.private = req;
4982 if (poll->events & EPOLLEXCLUSIVE)
4983 add_wait_queue_exclusive(head, &poll->wait);
4985 add_wait_queue(head, &poll->wait);
4988 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
4989 struct poll_table_struct *p)
4991 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
4992 struct async_poll *apoll = pt->req->apoll;
4994 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
4997 static void io_async_task_func(struct callback_head *cb)
4999 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
5000 struct async_poll *apoll = req->apoll;
5001 struct io_ring_ctx *ctx = req->ctx;
5003 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
5005 if (io_poll_rewait(req, &apoll->poll)) {
5006 spin_unlock_irq(&ctx->completion_lock);
5007 percpu_ref_put(&ctx->refs);
5011 /* If req is still hashed, it cannot have been canceled. Don't check. */
5012 if (hash_hashed(&req->hash_node))
5013 hash_del(&req->hash_node);
5015 io_poll_remove_double(req);
5016 spin_unlock_irq(&ctx->completion_lock);
5018 if (!READ_ONCE(apoll->poll.canceled))
5019 __io_req_task_submit(req);
5021 __io_req_task_cancel(req, -ECANCELED);
5023 percpu_ref_put(&ctx->refs);
5024 kfree(apoll->double_poll);
5028 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5031 struct io_kiocb *req = wait->private;
5032 struct io_poll_iocb *poll = &req->apoll->poll;
5034 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5037 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5040 static void io_poll_req_insert(struct io_kiocb *req)
5042 struct io_ring_ctx *ctx = req->ctx;
5043 struct hlist_head *list;
5045 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5046 hlist_add_head(&req->hash_node, list);
5049 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5050 struct io_poll_iocb *poll,
5051 struct io_poll_table *ipt, __poll_t mask,
5052 wait_queue_func_t wake_func)
5053 __acquires(&ctx->completion_lock)
5055 struct io_ring_ctx *ctx = req->ctx;
5056 bool cancel = false;
5058 INIT_HLIST_NODE(&req->hash_node);
5059 io_init_poll_iocb(poll, mask, wake_func);
5060 poll->file = req->file;
5061 poll->wait.private = req;
5063 ipt->pt._key = mask;
5065 ipt->error = -EINVAL;
5067 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5069 spin_lock_irq(&ctx->completion_lock);
5070 if (likely(poll->head)) {
5071 spin_lock(&poll->head->lock);
5072 if (unlikely(list_empty(&poll->wait.entry))) {
5078 if (mask || ipt->error)
5079 list_del_init(&poll->wait.entry);
5081 WRITE_ONCE(poll->canceled, true);
5082 else if (!poll->done) /* actually waiting for an event */
5083 io_poll_req_insert(req);
5084 spin_unlock(&poll->head->lock);
5090 static bool io_arm_poll_handler(struct io_kiocb *req)
5092 const struct io_op_def *def = &io_op_defs[req->opcode];
5093 struct io_ring_ctx *ctx = req->ctx;
5094 struct async_poll *apoll;
5095 struct io_poll_table ipt;
5099 if (!req->file || !file_can_poll(req->file))
5101 if (req->flags & REQ_F_POLLED)
5105 else if (def->pollout)
5109 /* if we can't nonblock try, then no point in arming a poll handler */
5110 if (!io_file_supports_async(req->file, rw))
5113 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5114 if (unlikely(!apoll))
5116 apoll->double_poll = NULL;
5118 req->flags |= REQ_F_POLLED;
5123 mask |= POLLIN | POLLRDNORM;
5125 mask |= POLLOUT | POLLWRNORM;
5127 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5128 if ((req->opcode == IORING_OP_RECVMSG) &&
5129 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5132 mask |= POLLERR | POLLPRI;
5134 ipt.pt._qproc = io_async_queue_proc;
5136 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5138 if (ret || ipt.error) {
5139 io_poll_remove_double(req);
5140 spin_unlock_irq(&ctx->completion_lock);
5141 kfree(apoll->double_poll);
5145 spin_unlock_irq(&ctx->completion_lock);
5146 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
5147 apoll->poll.events);
5151 static bool __io_poll_remove_one(struct io_kiocb *req,
5152 struct io_poll_iocb *poll)
5154 bool do_complete = false;
5156 spin_lock(&poll->head->lock);
5157 WRITE_ONCE(poll->canceled, true);
5158 if (!list_empty(&poll->wait.entry)) {
5159 list_del_init(&poll->wait.entry);
5162 spin_unlock(&poll->head->lock);
5163 hash_del(&req->hash_node);
5167 static bool io_poll_remove_one(struct io_kiocb *req)
5171 io_poll_remove_double(req);
5173 if (req->opcode == IORING_OP_POLL_ADD) {
5174 do_complete = __io_poll_remove_one(req, &req->poll);
5176 struct async_poll *apoll = req->apoll;
5178 /* non-poll requests have submit ref still */
5179 do_complete = __io_poll_remove_one(req, &apoll->poll);
5182 kfree(apoll->double_poll);
5188 io_cqring_fill_event(req, -ECANCELED);
5189 io_commit_cqring(req->ctx);
5190 req_set_fail_links(req);
5191 io_put_req_deferred(req, 1);
5198 * Returns true if we found and killed one or more poll requests
5200 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5201 struct files_struct *files)
5203 struct hlist_node *tmp;
5204 struct io_kiocb *req;
5207 spin_lock_irq(&ctx->completion_lock);
5208 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5209 struct hlist_head *list;
5211 list = &ctx->cancel_hash[i];
5212 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5213 if (io_match_task(req, tsk, files))
5214 posted += io_poll_remove_one(req);
5217 spin_unlock_irq(&ctx->completion_lock);
5220 io_cqring_ev_posted(ctx);
5225 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
5227 struct hlist_head *list;
5228 struct io_kiocb *req;
5230 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5231 hlist_for_each_entry(req, list, hash_node) {
5232 if (sqe_addr != req->user_data)
5234 if (io_poll_remove_one(req))
5242 static int io_poll_remove_prep(struct io_kiocb *req,
5243 const struct io_uring_sqe *sqe)
5245 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5247 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
5251 req->poll_remove.addr = READ_ONCE(sqe->addr);
5256 * Find a running poll command that matches one specified in sqe->addr,
5257 * and remove it if found.
5259 static int io_poll_remove(struct io_kiocb *req, unsigned int issue_flags)
5261 struct io_ring_ctx *ctx = req->ctx;
5264 spin_lock_irq(&ctx->completion_lock);
5265 ret = io_poll_cancel(ctx, req->poll_remove.addr);
5266 spin_unlock_irq(&ctx->completion_lock);
5269 req_set_fail_links(req);
5270 io_req_complete(req, ret);
5274 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5277 struct io_kiocb *req = wait->private;
5278 struct io_poll_iocb *poll = &req->poll;
5280 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5283 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5284 struct poll_table_struct *p)
5286 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5288 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5291 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5293 struct io_poll_iocb *poll = &req->poll;
5296 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5298 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
5301 events = READ_ONCE(sqe->poll32_events);
5303 events = swahw32(events);
5305 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP |
5306 (events & EPOLLEXCLUSIVE);
5310 static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5312 struct io_poll_iocb *poll = &req->poll;
5313 struct io_ring_ctx *ctx = req->ctx;
5314 struct io_poll_table ipt;
5317 ipt.pt._qproc = io_poll_queue_proc;
5319 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5322 if (mask) { /* no async, we'd stolen it */
5324 io_poll_complete(req, mask, 0);
5326 spin_unlock_irq(&ctx->completion_lock);
5329 io_cqring_ev_posted(ctx);
5335 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5337 struct io_timeout_data *data = container_of(timer,
5338 struct io_timeout_data, timer);
5339 struct io_kiocb *req = data->req;
5340 struct io_ring_ctx *ctx = req->ctx;
5341 unsigned long flags;
5343 spin_lock_irqsave(&ctx->completion_lock, flags);
5344 list_del_init(&req->timeout.list);
5345 atomic_set(&req->ctx->cq_timeouts,
5346 atomic_read(&req->ctx->cq_timeouts) + 1);
5348 io_cqring_fill_event(req, -ETIME);
5349 io_commit_cqring(ctx);
5350 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5352 io_cqring_ev_posted(ctx);
5353 req_set_fail_links(req);
5355 return HRTIMER_NORESTART;
5358 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5361 struct io_timeout_data *io;
5362 struct io_kiocb *req;
5365 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5366 if (user_data == req->user_data) {
5373 return ERR_PTR(ret);
5375 io = req->async_data;
5376 ret = hrtimer_try_to_cancel(&io->timer);
5378 return ERR_PTR(-EALREADY);
5379 list_del_init(&req->timeout.list);
5383 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5385 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5388 return PTR_ERR(req);
5390 req_set_fail_links(req);
5391 io_cqring_fill_event(req, -ECANCELED);
5392 io_put_req_deferred(req, 1);
5396 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5397 struct timespec64 *ts, enum hrtimer_mode mode)
5399 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5400 struct io_timeout_data *data;
5403 return PTR_ERR(req);
5405 req->timeout.off = 0; /* noseq */
5406 data = req->async_data;
5407 list_add_tail(&req->timeout.list, &ctx->timeout_list);
5408 hrtimer_init(&data->timer, CLOCK_MONOTONIC, mode);
5409 data->timer.function = io_timeout_fn;
5410 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
5414 static int io_timeout_remove_prep(struct io_kiocb *req,
5415 const struct io_uring_sqe *sqe)
5417 struct io_timeout_rem *tr = &req->timeout_rem;
5419 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5421 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5423 if (sqe->ioprio || sqe->buf_index || sqe->len)
5426 tr->addr = READ_ONCE(sqe->addr);
5427 tr->flags = READ_ONCE(sqe->timeout_flags);
5428 if (tr->flags & IORING_TIMEOUT_UPDATE) {
5429 if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS))
5431 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
5433 } else if (tr->flags) {
5434 /* timeout removal doesn't support flags */
5441 static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
5443 return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
5448 * Remove or update an existing timeout command
5450 static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
5452 struct io_timeout_rem *tr = &req->timeout_rem;
5453 struct io_ring_ctx *ctx = req->ctx;
5456 spin_lock_irq(&ctx->completion_lock);
5457 if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE))
5458 ret = io_timeout_cancel(ctx, tr->addr);
5460 ret = io_timeout_update(ctx, tr->addr, &tr->ts,
5461 io_translate_timeout_mode(tr->flags));
5463 io_cqring_fill_event(req, ret);
5464 io_commit_cqring(ctx);
5465 spin_unlock_irq(&ctx->completion_lock);
5466 io_cqring_ev_posted(ctx);
5468 req_set_fail_links(req);
5473 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5474 bool is_timeout_link)
5476 struct io_timeout_data *data;
5478 u32 off = READ_ONCE(sqe->off);
5480 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5482 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5484 if (off && is_timeout_link)
5486 flags = READ_ONCE(sqe->timeout_flags);
5487 if (flags & ~IORING_TIMEOUT_ABS)
5490 req->timeout.off = off;
5492 if (!req->async_data && io_alloc_async_data(req))
5495 data = req->async_data;
5498 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5501 data->mode = io_translate_timeout_mode(flags);
5502 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5506 static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
5508 struct io_ring_ctx *ctx = req->ctx;
5509 struct io_timeout_data *data = req->async_data;
5510 struct list_head *entry;
5511 u32 tail, off = req->timeout.off;
5513 spin_lock_irq(&ctx->completion_lock);
5516 * sqe->off holds how many events that need to occur for this
5517 * timeout event to be satisfied. If it isn't set, then this is
5518 * a pure timeout request, sequence isn't used.
5520 if (io_is_timeout_noseq(req)) {
5521 entry = ctx->timeout_list.prev;
5525 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5526 req->timeout.target_seq = tail + off;
5528 /* Update the last seq here in case io_flush_timeouts() hasn't.
5529 * This is safe because ->completion_lock is held, and submissions
5530 * and completions are never mixed in the same ->completion_lock section.
5532 ctx->cq_last_tm_flush = tail;
5535 * Insertion sort, ensuring the first entry in the list is always
5536 * the one we need first.
5538 list_for_each_prev(entry, &ctx->timeout_list) {
5539 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5542 if (io_is_timeout_noseq(nxt))
5544 /* nxt.seq is behind @tail, otherwise would've been completed */
5545 if (off >= nxt->timeout.target_seq - tail)
5549 list_add(&req->timeout.list, entry);
5550 data->timer.function = io_timeout_fn;
5551 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5552 spin_unlock_irq(&ctx->completion_lock);
5556 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5558 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5560 return req->user_data == (unsigned long) data;
5563 static int io_async_cancel_one(struct io_uring_task *tctx, void *sqe_addr)
5565 enum io_wq_cancel cancel_ret;
5571 cancel_ret = io_wq_cancel_cb(tctx->io_wq, io_cancel_cb, sqe_addr, false);
5572 switch (cancel_ret) {
5573 case IO_WQ_CANCEL_OK:
5576 case IO_WQ_CANCEL_RUNNING:
5579 case IO_WQ_CANCEL_NOTFOUND:
5587 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5588 struct io_kiocb *req, __u64 sqe_addr,
5591 unsigned long flags;
5594 ret = io_async_cancel_one(req->task->io_uring,
5595 (void *) (unsigned long) sqe_addr);
5596 if (ret != -ENOENT) {
5597 spin_lock_irqsave(&ctx->completion_lock, flags);
5601 spin_lock_irqsave(&ctx->completion_lock, flags);
5602 ret = io_timeout_cancel(ctx, sqe_addr);
5605 ret = io_poll_cancel(ctx, sqe_addr);
5609 io_cqring_fill_event(req, ret);
5610 io_commit_cqring(ctx);
5611 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5612 io_cqring_ev_posted(ctx);
5615 req_set_fail_links(req);
5619 static int io_async_cancel_prep(struct io_kiocb *req,
5620 const struct io_uring_sqe *sqe)
5622 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5624 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5626 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5629 req->cancel.addr = READ_ONCE(sqe->addr);
5633 static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
5635 struct io_ring_ctx *ctx = req->ctx;
5637 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
5641 static int io_rsrc_update_prep(struct io_kiocb *req,
5642 const struct io_uring_sqe *sqe)
5644 if (unlikely(req->ctx->flags & IORING_SETUP_SQPOLL))
5646 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5648 if (sqe->ioprio || sqe->rw_flags)
5651 req->rsrc_update.offset = READ_ONCE(sqe->off);
5652 req->rsrc_update.nr_args = READ_ONCE(sqe->len);
5653 if (!req->rsrc_update.nr_args)
5655 req->rsrc_update.arg = READ_ONCE(sqe->addr);
5659 static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
5661 struct io_ring_ctx *ctx = req->ctx;
5662 struct io_uring_rsrc_update up;
5665 if (issue_flags & IO_URING_F_NONBLOCK)
5668 up.offset = req->rsrc_update.offset;
5669 up.data = req->rsrc_update.arg;
5671 mutex_lock(&ctx->uring_lock);
5672 ret = __io_sqe_files_update(ctx, &up, req->rsrc_update.nr_args);
5673 mutex_unlock(&ctx->uring_lock);
5676 req_set_fail_links(req);
5677 __io_req_complete(req, issue_flags, ret, 0);
5681 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5683 switch (req->opcode) {
5686 case IORING_OP_READV:
5687 case IORING_OP_READ_FIXED:
5688 case IORING_OP_READ:
5689 return io_read_prep(req, sqe);
5690 case IORING_OP_WRITEV:
5691 case IORING_OP_WRITE_FIXED:
5692 case IORING_OP_WRITE:
5693 return io_write_prep(req, sqe);
5694 case IORING_OP_POLL_ADD:
5695 return io_poll_add_prep(req, sqe);
5696 case IORING_OP_POLL_REMOVE:
5697 return io_poll_remove_prep(req, sqe);
5698 case IORING_OP_FSYNC:
5699 return io_fsync_prep(req, sqe);
5700 case IORING_OP_SYNC_FILE_RANGE:
5701 return io_sfr_prep(req, sqe);
5702 case IORING_OP_SENDMSG:
5703 case IORING_OP_SEND:
5704 return io_sendmsg_prep(req, sqe);
5705 case IORING_OP_RECVMSG:
5706 case IORING_OP_RECV:
5707 return io_recvmsg_prep(req, sqe);
5708 case IORING_OP_CONNECT:
5709 return io_connect_prep(req, sqe);
5710 case IORING_OP_TIMEOUT:
5711 return io_timeout_prep(req, sqe, false);
5712 case IORING_OP_TIMEOUT_REMOVE:
5713 return io_timeout_remove_prep(req, sqe);
5714 case IORING_OP_ASYNC_CANCEL:
5715 return io_async_cancel_prep(req, sqe);
5716 case IORING_OP_LINK_TIMEOUT:
5717 return io_timeout_prep(req, sqe, true);
5718 case IORING_OP_ACCEPT:
5719 return io_accept_prep(req, sqe);
5720 case IORING_OP_FALLOCATE:
5721 return io_fallocate_prep(req, sqe);
5722 case IORING_OP_OPENAT:
5723 return io_openat_prep(req, sqe);
5724 case IORING_OP_CLOSE:
5725 return io_close_prep(req, sqe);
5726 case IORING_OP_FILES_UPDATE:
5727 return io_rsrc_update_prep(req, sqe);
5728 case IORING_OP_STATX:
5729 return io_statx_prep(req, sqe);
5730 case IORING_OP_FADVISE:
5731 return io_fadvise_prep(req, sqe);
5732 case IORING_OP_MADVISE:
5733 return io_madvise_prep(req, sqe);
5734 case IORING_OP_OPENAT2:
5735 return io_openat2_prep(req, sqe);
5736 case IORING_OP_EPOLL_CTL:
5737 return io_epoll_ctl_prep(req, sqe);
5738 case IORING_OP_SPLICE:
5739 return io_splice_prep(req, sqe);
5740 case IORING_OP_PROVIDE_BUFFERS:
5741 return io_provide_buffers_prep(req, sqe);
5742 case IORING_OP_REMOVE_BUFFERS:
5743 return io_remove_buffers_prep(req, sqe);
5745 return io_tee_prep(req, sqe);
5746 case IORING_OP_SHUTDOWN:
5747 return io_shutdown_prep(req, sqe);
5748 case IORING_OP_RENAMEAT:
5749 return io_renameat_prep(req, sqe);
5750 case IORING_OP_UNLINKAT:
5751 return io_unlinkat_prep(req, sqe);
5754 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5759 static int io_req_prep_async(struct io_kiocb *req)
5761 switch (req->opcode) {
5762 case IORING_OP_READV:
5763 case IORING_OP_READ_FIXED:
5764 case IORING_OP_READ:
5765 return io_rw_prep_async(req, READ);
5766 case IORING_OP_WRITEV:
5767 case IORING_OP_WRITE_FIXED:
5768 case IORING_OP_WRITE:
5769 return io_rw_prep_async(req, WRITE);
5770 case IORING_OP_SENDMSG:
5771 case IORING_OP_SEND:
5772 return io_sendmsg_prep_async(req);
5773 case IORING_OP_RECVMSG:
5774 case IORING_OP_RECV:
5775 return io_recvmsg_prep_async(req);
5776 case IORING_OP_CONNECT:
5777 return io_connect_prep_async(req);
5782 static int io_req_defer_prep(struct io_kiocb *req)
5784 if (!io_op_defs[req->opcode].needs_async_data)
5786 /* some opcodes init it during the inital prep */
5787 if (req->async_data)
5789 if (__io_alloc_async_data(req))
5791 return io_req_prep_async(req);
5794 static u32 io_get_sequence(struct io_kiocb *req)
5796 struct io_kiocb *pos;
5797 struct io_ring_ctx *ctx = req->ctx;
5798 u32 total_submitted, nr_reqs = 0;
5800 io_for_each_link(pos, req)
5803 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
5804 return total_submitted - nr_reqs;
5807 static int io_req_defer(struct io_kiocb *req)
5809 struct io_ring_ctx *ctx = req->ctx;
5810 struct io_defer_entry *de;
5814 /* Still need defer if there is pending req in defer list. */
5815 if (likely(list_empty_careful(&ctx->defer_list) &&
5816 !(req->flags & REQ_F_IO_DRAIN)))
5819 seq = io_get_sequence(req);
5820 /* Still a chance to pass the sequence check */
5821 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
5824 ret = io_req_defer_prep(req);
5827 io_prep_async_link(req);
5828 de = kmalloc(sizeof(*de), GFP_KERNEL);
5832 spin_lock_irq(&ctx->completion_lock);
5833 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
5834 spin_unlock_irq(&ctx->completion_lock);
5836 io_queue_async_work(req);
5837 return -EIOCBQUEUED;
5840 trace_io_uring_defer(ctx, req, req->user_data);
5843 list_add_tail(&de->list, &ctx->defer_list);
5844 spin_unlock_irq(&ctx->completion_lock);
5845 return -EIOCBQUEUED;
5848 static void __io_clean_op(struct io_kiocb *req)
5850 if (req->flags & REQ_F_BUFFER_SELECTED) {
5851 switch (req->opcode) {
5852 case IORING_OP_READV:
5853 case IORING_OP_READ_FIXED:
5854 case IORING_OP_READ:
5855 kfree((void *)(unsigned long)req->rw.addr);
5857 case IORING_OP_RECVMSG:
5858 case IORING_OP_RECV:
5859 kfree(req->sr_msg.kbuf);
5862 req->flags &= ~REQ_F_BUFFER_SELECTED;
5865 if (req->flags & REQ_F_NEED_CLEANUP) {
5866 switch (req->opcode) {
5867 case IORING_OP_READV:
5868 case IORING_OP_READ_FIXED:
5869 case IORING_OP_READ:
5870 case IORING_OP_WRITEV:
5871 case IORING_OP_WRITE_FIXED:
5872 case IORING_OP_WRITE: {
5873 struct io_async_rw *io = req->async_data;
5875 kfree(io->free_iovec);
5878 case IORING_OP_RECVMSG:
5879 case IORING_OP_SENDMSG: {
5880 struct io_async_msghdr *io = req->async_data;
5882 kfree(io->free_iov);
5885 case IORING_OP_SPLICE:
5887 io_put_file(req, req->splice.file_in,
5888 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
5890 case IORING_OP_OPENAT:
5891 case IORING_OP_OPENAT2:
5892 if (req->open.filename)
5893 putname(req->open.filename);
5895 case IORING_OP_RENAMEAT:
5896 putname(req->rename.oldpath);
5897 putname(req->rename.newpath);
5899 case IORING_OP_UNLINKAT:
5900 putname(req->unlink.filename);
5903 req->flags &= ~REQ_F_NEED_CLEANUP;
5907 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
5909 struct io_ring_ctx *ctx = req->ctx;
5910 const struct cred *creds = NULL;
5913 if (req->work.personality) {
5914 const struct cred *new_creds;
5916 if (!(issue_flags & IO_URING_F_NONBLOCK))
5917 mutex_lock(&ctx->uring_lock);
5918 new_creds = idr_find(&ctx->personality_idr, req->work.personality);
5919 if (!(issue_flags & IO_URING_F_NONBLOCK))
5920 mutex_unlock(&ctx->uring_lock);
5923 creds = override_creds(new_creds);
5926 switch (req->opcode) {
5928 ret = io_nop(req, issue_flags);
5930 case IORING_OP_READV:
5931 case IORING_OP_READ_FIXED:
5932 case IORING_OP_READ:
5933 ret = io_read(req, issue_flags);
5935 case IORING_OP_WRITEV:
5936 case IORING_OP_WRITE_FIXED:
5937 case IORING_OP_WRITE:
5938 ret = io_write(req, issue_flags);
5940 case IORING_OP_FSYNC:
5941 ret = io_fsync(req, issue_flags);
5943 case IORING_OP_POLL_ADD:
5944 ret = io_poll_add(req, issue_flags);
5946 case IORING_OP_POLL_REMOVE:
5947 ret = io_poll_remove(req, issue_flags);
5949 case IORING_OP_SYNC_FILE_RANGE:
5950 ret = io_sync_file_range(req, issue_flags);
5952 case IORING_OP_SENDMSG:
5953 ret = io_sendmsg(req, issue_flags);
5955 case IORING_OP_SEND:
5956 ret = io_send(req, issue_flags);
5958 case IORING_OP_RECVMSG:
5959 ret = io_recvmsg(req, issue_flags);
5961 case IORING_OP_RECV:
5962 ret = io_recv(req, issue_flags);
5964 case IORING_OP_TIMEOUT:
5965 ret = io_timeout(req, issue_flags);
5967 case IORING_OP_TIMEOUT_REMOVE:
5968 ret = io_timeout_remove(req, issue_flags);
5970 case IORING_OP_ACCEPT:
5971 ret = io_accept(req, issue_flags);
5973 case IORING_OP_CONNECT:
5974 ret = io_connect(req, issue_flags);
5976 case IORING_OP_ASYNC_CANCEL:
5977 ret = io_async_cancel(req, issue_flags);
5979 case IORING_OP_FALLOCATE:
5980 ret = io_fallocate(req, issue_flags);
5982 case IORING_OP_OPENAT:
5983 ret = io_openat(req, issue_flags);
5985 case IORING_OP_CLOSE:
5986 ret = io_close(req, issue_flags);
5988 case IORING_OP_FILES_UPDATE:
5989 ret = io_files_update(req, issue_flags);
5991 case IORING_OP_STATX:
5992 ret = io_statx(req, issue_flags);
5994 case IORING_OP_FADVISE:
5995 ret = io_fadvise(req, issue_flags);
5997 case IORING_OP_MADVISE:
5998 ret = io_madvise(req, issue_flags);
6000 case IORING_OP_OPENAT2:
6001 ret = io_openat2(req, issue_flags);
6003 case IORING_OP_EPOLL_CTL:
6004 ret = io_epoll_ctl(req, issue_flags);
6006 case IORING_OP_SPLICE:
6007 ret = io_splice(req, issue_flags);
6009 case IORING_OP_PROVIDE_BUFFERS:
6010 ret = io_provide_buffers(req, issue_flags);
6012 case IORING_OP_REMOVE_BUFFERS:
6013 ret = io_remove_buffers(req, issue_flags);
6016 ret = io_tee(req, issue_flags);
6018 case IORING_OP_SHUTDOWN:
6019 ret = io_shutdown(req, issue_flags);
6021 case IORING_OP_RENAMEAT:
6022 ret = io_renameat(req, issue_flags);
6024 case IORING_OP_UNLINKAT:
6025 ret = io_unlinkat(req, issue_flags);
6033 revert_creds(creds);
6038 /* If the op doesn't have a file, we're not polling for it */
6039 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
6040 const bool in_async = io_wq_current_is_worker();
6042 /* workqueue context doesn't hold uring_lock, grab it now */
6044 mutex_lock(&ctx->uring_lock);
6046 io_iopoll_req_issued(req, in_async);
6049 mutex_unlock(&ctx->uring_lock);
6055 static void io_wq_submit_work(struct io_wq_work *work)
6057 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6058 struct io_kiocb *timeout;
6061 timeout = io_prep_linked_timeout(req);
6063 io_queue_linked_timeout(timeout);
6065 if (work->flags & IO_WQ_WORK_CANCEL)
6070 ret = io_issue_sqe(req, 0);
6072 * We can get EAGAIN for polled IO even though we're
6073 * forcing a sync submission from here, since we can't
6074 * wait for request slots on the block side.
6082 /* avoid locking problems by failing it from a clean context */
6084 /* io-wq is going to take one down */
6085 refcount_inc(&req->refs);
6086 io_req_task_queue_fail(req, ret);
6090 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6093 struct fixed_rsrc_table *table;
6095 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
6096 return table->files[index & IORING_FILE_TABLE_MASK];
6099 static struct file *io_file_get(struct io_submit_state *state,
6100 struct io_kiocb *req, int fd, bool fixed)
6102 struct io_ring_ctx *ctx = req->ctx;
6106 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6108 fd = array_index_nospec(fd, ctx->nr_user_files);
6109 file = io_file_from_index(ctx, fd);
6110 io_set_resource_node(req);
6112 trace_io_uring_file_get(ctx, fd);
6113 file = __io_file_get(state, fd);
6116 if (file && unlikely(file->f_op == &io_uring_fops))
6117 io_req_track_inflight(req);
6121 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6123 struct io_timeout_data *data = container_of(timer,
6124 struct io_timeout_data, timer);
6125 struct io_kiocb *prev, *req = data->req;
6126 struct io_ring_ctx *ctx = req->ctx;
6127 unsigned long flags;
6129 spin_lock_irqsave(&ctx->completion_lock, flags);
6130 prev = req->timeout.head;
6131 req->timeout.head = NULL;
6134 * We don't expect the list to be empty, that will only happen if we
6135 * race with the completion of the linked work.
6137 if (prev && refcount_inc_not_zero(&prev->refs))
6138 io_remove_next_linked(prev);
6141 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6144 req_set_fail_links(prev);
6145 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6146 io_put_req_deferred(prev, 1);
6148 io_req_complete_post(req, -ETIME, 0);
6149 io_put_req_deferred(req, 1);
6151 return HRTIMER_NORESTART;
6154 static void __io_queue_linked_timeout(struct io_kiocb *req)
6157 * If the back reference is NULL, then our linked request finished
6158 * before we got a chance to setup the timer
6160 if (req->timeout.head) {
6161 struct io_timeout_data *data = req->async_data;
6163 data->timer.function = io_link_timeout_fn;
6164 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6169 static void io_queue_linked_timeout(struct io_kiocb *req)
6171 struct io_ring_ctx *ctx = req->ctx;
6173 spin_lock_irq(&ctx->completion_lock);
6174 __io_queue_linked_timeout(req);
6175 spin_unlock_irq(&ctx->completion_lock);
6177 /* drop submission reference */
6181 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6183 struct io_kiocb *nxt = req->link;
6185 if (!nxt || (req->flags & REQ_F_LINK_TIMEOUT) ||
6186 nxt->opcode != IORING_OP_LINK_TIMEOUT)
6189 nxt->timeout.head = req;
6190 nxt->flags |= REQ_F_LTIMEOUT_ACTIVE;
6191 req->flags |= REQ_F_LINK_TIMEOUT;
6195 static void __io_queue_sqe(struct io_kiocb *req)
6197 struct io_kiocb *linked_timeout = io_prep_linked_timeout(req);
6200 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
6203 * We async punt it if the file wasn't marked NOWAIT, or if the file
6204 * doesn't support non-blocking read/write attempts
6206 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6207 if (!io_arm_poll_handler(req)) {
6209 * Queued up for async execution, worker will release
6210 * submit reference when the iocb is actually submitted.
6212 io_queue_async_work(req);
6214 } else if (likely(!ret)) {
6215 /* drop submission reference */
6216 if (req->flags & REQ_F_COMPLETE_INLINE) {
6217 struct io_ring_ctx *ctx = req->ctx;
6218 struct io_comp_state *cs = &ctx->submit_state.comp;
6220 cs->reqs[cs->nr++] = req;
6221 if (cs->nr == ARRAY_SIZE(cs->reqs))
6222 io_submit_flush_completions(cs, ctx);
6227 req_set_fail_links(req);
6229 io_req_complete(req, ret);
6232 io_queue_linked_timeout(linked_timeout);
6235 static void io_queue_sqe(struct io_kiocb *req)
6239 ret = io_req_defer(req);
6241 if (ret != -EIOCBQUEUED) {
6243 req_set_fail_links(req);
6245 io_req_complete(req, ret);
6247 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6248 ret = io_req_defer_prep(req);
6251 io_queue_async_work(req);
6253 __io_queue_sqe(req);
6258 * Check SQE restrictions (opcode and flags).
6260 * Returns 'true' if SQE is allowed, 'false' otherwise.
6262 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6263 struct io_kiocb *req,
6264 unsigned int sqe_flags)
6266 if (!ctx->restricted)
6269 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6272 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6273 ctx->restrictions.sqe_flags_required)
6276 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6277 ctx->restrictions.sqe_flags_required))
6283 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6284 const struct io_uring_sqe *sqe)
6286 struct io_submit_state *state;
6287 unsigned int sqe_flags;
6290 req->opcode = READ_ONCE(sqe->opcode);
6291 /* same numerical values with corresponding REQ_F_*, safe to copy */
6292 req->flags = sqe_flags = READ_ONCE(sqe->flags);
6293 req->user_data = READ_ONCE(sqe->user_data);
6294 req->async_data = NULL;
6298 req->fixed_rsrc_refs = NULL;
6299 /* one is dropped after submission, the other at completion */
6300 refcount_set(&req->refs, 2);
6301 req->task = current;
6304 /* enforce forwards compatibility on users */
6305 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS)) {
6310 if (unlikely(req->opcode >= IORING_OP_LAST))
6313 if (unlikely(!io_check_restriction(ctx, req, sqe_flags)))
6316 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6317 !io_op_defs[req->opcode].buffer_select)
6320 req->work.list.next = NULL;
6321 req->work.flags = 0;
6322 req->work.personality = READ_ONCE(sqe->personality);
6323 state = &ctx->submit_state;
6326 * Plug now if we have more than 1 IO left after this, and the target
6327 * is potentially a read/write to block based storage.
6329 if (!state->plug_started && state->ios_left > 1 &&
6330 io_op_defs[req->opcode].plug) {
6331 blk_start_plug(&state->plug);
6332 state->plug_started = true;
6335 if (io_op_defs[req->opcode].needs_file) {
6336 bool fixed = req->flags & REQ_F_FIXED_FILE;
6338 req->file = io_file_get(state, req, READ_ONCE(sqe->fd), fixed);
6339 if (unlikely(!req->file))
6347 static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
6348 const struct io_uring_sqe *sqe)
6350 struct io_submit_link *link = &ctx->submit_state.link;
6353 ret = io_init_req(ctx, req, sqe);
6354 if (unlikely(ret)) {
6357 io_req_complete(req, ret);
6359 /* fail even hard links since we don't submit */
6360 link->head->flags |= REQ_F_FAIL_LINK;
6361 io_put_req(link->head);
6362 io_req_complete(link->head, -ECANCELED);
6367 ret = io_req_prep(req, sqe);
6371 /* don't need @sqe from now on */
6372 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6373 true, ctx->flags & IORING_SETUP_SQPOLL);
6376 * If we already have a head request, queue this one for async
6377 * submittal once the head completes. If we don't have a head but
6378 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6379 * submitted sync once the chain is complete. If none of those
6380 * conditions are true (normal request), then just queue it.
6383 struct io_kiocb *head = link->head;
6386 * Taking sequential execution of a link, draining both sides
6387 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6388 * requests in the link. So, it drains the head and the
6389 * next after the link request. The last one is done via
6390 * drain_next flag to persist the effect across calls.
6392 if (req->flags & REQ_F_IO_DRAIN) {
6393 head->flags |= REQ_F_IO_DRAIN;
6394 ctx->drain_next = 1;
6396 ret = io_req_defer_prep(req);
6399 trace_io_uring_link(ctx, req, head);
6400 link->last->link = req;
6403 /* last request of a link, enqueue the link */
6404 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6409 if (unlikely(ctx->drain_next)) {
6410 req->flags |= REQ_F_IO_DRAIN;
6411 ctx->drain_next = 0;
6413 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6425 * Batched submission is done, ensure local IO is flushed out.
6427 static void io_submit_state_end(struct io_submit_state *state,
6428 struct io_ring_ctx *ctx)
6430 if (state->link.head)
6431 io_queue_sqe(state->link.head);
6433 io_submit_flush_completions(&state->comp, ctx);
6434 if (state->plug_started)
6435 blk_finish_plug(&state->plug);
6436 io_state_file_put(state);
6440 * Start submission side cache.
6442 static void io_submit_state_start(struct io_submit_state *state,
6443 unsigned int max_ios)
6445 state->plug_started = false;
6446 state->ios_left = max_ios;
6447 /* set only head, no need to init link_last in advance */
6448 state->link.head = NULL;
6451 static void io_commit_sqring(struct io_ring_ctx *ctx)
6453 struct io_rings *rings = ctx->rings;
6456 * Ensure any loads from the SQEs are done at this point,
6457 * since once we write the new head, the application could
6458 * write new data to them.
6460 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6464 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6465 * that is mapped by userspace. This means that care needs to be taken to
6466 * ensure that reads are stable, as we cannot rely on userspace always
6467 * being a good citizen. If members of the sqe are validated and then later
6468 * used, it's important that those reads are done through READ_ONCE() to
6469 * prevent a re-load down the line.
6471 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6473 u32 *sq_array = ctx->sq_array;
6477 * The cached sq head (or cq tail) serves two purposes:
6479 * 1) allows us to batch the cost of updating the user visible
6481 * 2) allows the kernel side to track the head on its own, even
6482 * though the application is the one updating it.
6484 head = READ_ONCE(sq_array[ctx->cached_sq_head++ & ctx->sq_mask]);
6485 if (likely(head < ctx->sq_entries))
6486 return &ctx->sq_sqes[head];
6488 /* drop invalid entries */
6489 ctx->cached_sq_dropped++;
6490 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6494 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6498 /* if we have a backlog and couldn't flush it all, return BUSY */
6499 if (test_bit(0, &ctx->sq_check_overflow)) {
6500 if (!__io_cqring_overflow_flush(ctx, false, NULL, NULL))
6504 /* make sure SQ entry isn't read before tail */
6505 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6507 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6510 percpu_counter_add(¤t->io_uring->inflight, nr);
6511 refcount_add(nr, ¤t->usage);
6512 io_submit_state_start(&ctx->submit_state, nr);
6514 while (submitted < nr) {
6515 const struct io_uring_sqe *sqe;
6516 struct io_kiocb *req;
6518 req = io_alloc_req(ctx);
6519 if (unlikely(!req)) {
6521 submitted = -EAGAIN;
6524 sqe = io_get_sqe(ctx);
6525 if (unlikely(!sqe)) {
6526 kmem_cache_free(req_cachep, req);
6529 /* will complete beyond this point, count as submitted */
6531 if (io_submit_sqe(ctx, req, sqe))
6535 if (unlikely(submitted != nr)) {
6536 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6537 struct io_uring_task *tctx = current->io_uring;
6538 int unused = nr - ref_used;
6540 percpu_ref_put_many(&ctx->refs, unused);
6541 percpu_counter_sub(&tctx->inflight, unused);
6542 put_task_struct_many(current, unused);
6545 io_submit_state_end(&ctx->submit_state, ctx);
6546 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6547 io_commit_sqring(ctx);
6552 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6554 /* Tell userspace we may need a wakeup call */
6555 spin_lock_irq(&ctx->completion_lock);
6556 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6557 spin_unlock_irq(&ctx->completion_lock);
6560 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6562 spin_lock_irq(&ctx->completion_lock);
6563 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6564 spin_unlock_irq(&ctx->completion_lock);
6567 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
6569 unsigned int to_submit;
6572 to_submit = io_sqring_entries(ctx);
6573 /* if we're handling multiple rings, cap submit size for fairness */
6574 if (cap_entries && to_submit > 8)
6577 if (!list_empty(&ctx->iopoll_list) || to_submit) {
6578 unsigned nr_events = 0;
6580 mutex_lock(&ctx->uring_lock);
6581 if (!list_empty(&ctx->iopoll_list))
6582 io_do_iopoll(ctx, &nr_events, 0);
6584 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)))
6585 ret = io_submit_sqes(ctx, to_submit);
6586 mutex_unlock(&ctx->uring_lock);
6589 if (!io_sqring_full(ctx) && wq_has_sleeper(&ctx->sqo_sq_wait))
6590 wake_up(&ctx->sqo_sq_wait);
6595 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
6597 struct io_ring_ctx *ctx;
6598 unsigned sq_thread_idle = 0;
6600 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6601 if (sq_thread_idle < ctx->sq_thread_idle)
6602 sq_thread_idle = ctx->sq_thread_idle;
6605 sqd->sq_thread_idle = sq_thread_idle;
6608 static void io_sqd_init_new(struct io_sq_data *sqd)
6610 struct io_ring_ctx *ctx;
6612 while (!list_empty(&sqd->ctx_new_list)) {
6613 ctx = list_first_entry(&sqd->ctx_new_list, struct io_ring_ctx, sqd_list);
6614 list_move_tail(&ctx->sqd_list, &sqd->ctx_list);
6615 complete(&ctx->sq_thread_comp);
6618 io_sqd_update_thread_idle(sqd);
6621 static bool io_sq_thread_should_stop(struct io_sq_data *sqd)
6623 return test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
6626 static bool io_sq_thread_should_park(struct io_sq_data *sqd)
6628 return test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
6631 static void io_sq_thread_parkme(struct io_sq_data *sqd)
6635 * TASK_PARKED is a special state; we must serialize against
6636 * possible pending wakeups to avoid store-store collisions on
6639 * Such a collision might possibly result in the task state
6640 * changin from TASK_PARKED and us failing the
6641 * wait_task_inactive() in kthread_park().
6643 set_special_state(TASK_PARKED);
6644 if (!test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state))
6648 * Thread is going to call schedule(), do not preempt it,
6649 * or the caller of kthread_park() may spend more time in
6650 * wait_task_inactive().
6653 complete(&sqd->completion);
6654 schedule_preempt_disabled();
6657 __set_current_state(TASK_RUNNING);
6660 static int io_sq_thread(void *data)
6662 struct io_sq_data *sqd = data;
6663 struct io_ring_ctx *ctx;
6664 unsigned long timeout = 0;
6665 char buf[TASK_COMM_LEN];
6668 sprintf(buf, "iou-sqp-%d", sqd->task_pid);
6669 set_task_comm(current, buf);
6670 sqd->thread = current;
6671 current->pf_io_worker = NULL;
6673 if (sqd->sq_cpu != -1)
6674 set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
6676 set_cpus_allowed_ptr(current, cpu_online_mask);
6677 current->flags |= PF_NO_SETAFFINITY;
6679 complete(&sqd->completion);
6681 wait_for_completion(&sqd->startup);
6683 while (!io_sq_thread_should_stop(sqd)) {
6685 bool cap_entries, sqt_spin, needs_sched;
6688 * Any changes to the sqd lists are synchronized through the
6689 * thread parking. This synchronizes the thread vs users,
6690 * the users are synchronized on the sqd->ctx_lock.
6692 if (io_sq_thread_should_park(sqd)) {
6693 io_sq_thread_parkme(sqd);
6696 if (unlikely(!list_empty(&sqd->ctx_new_list))) {
6697 io_sqd_init_new(sqd);
6698 timeout = jiffies + sqd->sq_thread_idle;
6700 if (fatal_signal_pending(current))
6703 cap_entries = !list_is_singular(&sqd->ctx_list);
6704 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6705 ret = __io_sq_thread(ctx, cap_entries);
6706 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
6710 if (sqt_spin || !time_after(jiffies, timeout)) {
6714 timeout = jiffies + sqd->sq_thread_idle;
6719 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
6720 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6721 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6722 !list_empty_careful(&ctx->iopoll_list)) {
6723 needs_sched = false;
6726 if (io_sqring_entries(ctx)) {
6727 needs_sched = false;
6732 if (needs_sched && !io_sq_thread_should_park(sqd)) {
6733 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6734 io_ring_set_wakeup_flag(ctx);
6738 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6739 io_ring_clear_wakeup_flag(ctx);
6742 finish_wait(&sqd->wait, &wait);
6743 timeout = jiffies + sqd->sq_thread_idle;
6746 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6747 io_uring_cancel_sqpoll(ctx);
6751 if (io_sq_thread_should_park(sqd))
6752 io_sq_thread_parkme(sqd);
6755 * Clear thread under lock so that concurrent parks work correctly
6757 complete(&sqd->completion);
6758 mutex_lock(&sqd->lock);
6760 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6762 io_ring_set_wakeup_flag(ctx);
6765 complete(&sqd->exited);
6766 mutex_unlock(&sqd->lock);
6770 struct io_wait_queue {
6771 struct wait_queue_entry wq;
6772 struct io_ring_ctx *ctx;
6774 unsigned nr_timeouts;
6777 static inline bool io_should_wake(struct io_wait_queue *iowq)
6779 struct io_ring_ctx *ctx = iowq->ctx;
6782 * Wake up if we have enough events, or if a timeout occurred since we
6783 * started waiting. For timeouts, we always want to return to userspace,
6784 * regardless of event count.
6786 return io_cqring_events(ctx) >= iowq->to_wait ||
6787 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6790 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6791 int wake_flags, void *key)
6793 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6797 * Cannot safely flush overflowed CQEs from here, ensure we wake up
6798 * the task, and the next invocation will do it.
6800 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->cq_check_overflow))
6801 return autoremove_wake_function(curr, mode, wake_flags, key);
6805 static int io_run_task_work_sig(void)
6807 if (io_run_task_work())
6809 if (!signal_pending(current))
6811 if (test_tsk_thread_flag(current, TIF_NOTIFY_SIGNAL))
6812 return -ERESTARTSYS;
6816 /* when returns >0, the caller should retry */
6817 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
6818 struct io_wait_queue *iowq,
6819 signed long *timeout)
6823 /* make sure we run task_work before checking for signals */
6824 ret = io_run_task_work_sig();
6825 if (ret || io_should_wake(iowq))
6827 /* let the caller flush overflows, retry */
6828 if (test_bit(0, &ctx->cq_check_overflow))
6831 *timeout = schedule_timeout(*timeout);
6832 return !*timeout ? -ETIME : 1;
6836 * Wait until events become available, if we don't already have some. The
6837 * application must reap them itself, as they reside on the shared cq ring.
6839 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6840 const sigset_t __user *sig, size_t sigsz,
6841 struct __kernel_timespec __user *uts)
6843 struct io_wait_queue iowq = {
6846 .func = io_wake_function,
6847 .entry = LIST_HEAD_INIT(iowq.wq.entry),
6850 .to_wait = min_events,
6852 struct io_rings *rings = ctx->rings;
6853 signed long timeout = MAX_SCHEDULE_TIMEOUT;
6857 io_cqring_overflow_flush(ctx, false, NULL, NULL);
6858 if (io_cqring_events(ctx) >= min_events)
6860 if (!io_run_task_work())
6865 #ifdef CONFIG_COMPAT
6866 if (in_compat_syscall())
6867 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
6871 ret = set_user_sigmask(sig, sigsz);
6878 struct timespec64 ts;
6880 if (get_timespec64(&ts, uts))
6882 timeout = timespec64_to_jiffies(&ts);
6885 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
6886 trace_io_uring_cqring_wait(ctx, min_events);
6888 io_cqring_overflow_flush(ctx, false, NULL, NULL);
6889 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
6890 TASK_INTERRUPTIBLE);
6891 ret = io_cqring_wait_schedule(ctx, &iowq, &timeout);
6892 finish_wait(&ctx->wait, &iowq.wq);
6895 restore_saved_sigmask_unless(ret == -EINTR);
6897 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
6900 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
6902 #if defined(CONFIG_UNIX)
6903 if (ctx->ring_sock) {
6904 struct sock *sock = ctx->ring_sock->sk;
6905 struct sk_buff *skb;
6907 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
6913 for (i = 0; i < ctx->nr_user_files; i++) {
6916 file = io_file_from_index(ctx, i);
6923 static void io_rsrc_data_ref_zero(struct percpu_ref *ref)
6925 struct fixed_rsrc_data *data;
6927 data = container_of(ref, struct fixed_rsrc_data, refs);
6928 complete(&data->done);
6931 static inline void io_rsrc_ref_lock(struct io_ring_ctx *ctx)
6933 spin_lock_bh(&ctx->rsrc_ref_lock);
6936 static inline void io_rsrc_ref_unlock(struct io_ring_ctx *ctx)
6938 spin_unlock_bh(&ctx->rsrc_ref_lock);
6941 static void io_sqe_rsrc_set_node(struct io_ring_ctx *ctx,
6942 struct fixed_rsrc_data *rsrc_data,
6943 struct fixed_rsrc_ref_node *ref_node)
6945 io_rsrc_ref_lock(ctx);
6946 rsrc_data->node = ref_node;
6947 list_add_tail(&ref_node->node, &ctx->rsrc_ref_list);
6948 io_rsrc_ref_unlock(ctx);
6949 percpu_ref_get(&rsrc_data->refs);
6952 static void io_sqe_rsrc_kill_node(struct io_ring_ctx *ctx, struct fixed_rsrc_data *data)
6954 struct fixed_rsrc_ref_node *ref_node = NULL;
6956 io_rsrc_ref_lock(ctx);
6957 ref_node = data->node;
6959 io_rsrc_ref_unlock(ctx);
6961 percpu_ref_kill(&ref_node->refs);
6964 static int io_rsrc_ref_quiesce(struct fixed_rsrc_data *data,
6965 struct io_ring_ctx *ctx,
6966 void (*rsrc_put)(struct io_ring_ctx *ctx,
6967 struct io_rsrc_put *prsrc))
6969 struct fixed_rsrc_ref_node *backup_node;
6975 data->quiesce = true;
6978 backup_node = alloc_fixed_rsrc_ref_node(ctx);
6981 backup_node->rsrc_data = data;
6982 backup_node->rsrc_put = rsrc_put;
6984 io_sqe_rsrc_kill_node(ctx, data);
6985 percpu_ref_kill(&data->refs);
6986 flush_delayed_work(&ctx->rsrc_put_work);
6988 ret = wait_for_completion_interruptible(&data->done);
6992 percpu_ref_resurrect(&data->refs);
6993 io_sqe_rsrc_set_node(ctx, data, backup_node);
6995 reinit_completion(&data->done);
6996 mutex_unlock(&ctx->uring_lock);
6997 ret = io_run_task_work_sig();
6998 mutex_lock(&ctx->uring_lock);
7000 data->quiesce = false;
7003 destroy_fixed_rsrc_ref_node(backup_node);
7007 static struct fixed_rsrc_data *alloc_fixed_rsrc_data(struct io_ring_ctx *ctx)
7009 struct fixed_rsrc_data *data;
7011 data = kzalloc(sizeof(*data), GFP_KERNEL);
7015 if (percpu_ref_init(&data->refs, io_rsrc_data_ref_zero,
7016 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) {
7021 init_completion(&data->done);
7025 static void free_fixed_rsrc_data(struct fixed_rsrc_data *data)
7027 percpu_ref_exit(&data->refs);
7032 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7034 struct fixed_rsrc_data *data = ctx->file_data;
7035 unsigned nr_tables, i;
7039 * percpu_ref_is_dying() is to stop parallel files unregister
7040 * Since we possibly drop uring lock later in this function to
7043 if (!data || percpu_ref_is_dying(&data->refs))
7045 ret = io_rsrc_ref_quiesce(data, ctx, io_ring_file_put);
7049 __io_sqe_files_unregister(ctx);
7050 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
7051 for (i = 0; i < nr_tables; i++)
7052 kfree(data->table[i].files);
7053 free_fixed_rsrc_data(data);
7054 ctx->file_data = NULL;
7055 ctx->nr_user_files = 0;
7059 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7060 __releases(&sqd->lock)
7064 if (sqd->thread == current)
7066 clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7067 wake_up_state(sqd->thread, TASK_PARKED);
7068 mutex_unlock(&sqd->lock);
7071 static bool io_sq_thread_park(struct io_sq_data *sqd)
7072 __acquires(&sqd->lock)
7074 if (sqd->thread == current)
7076 mutex_lock(&sqd->lock);
7078 mutex_unlock(&sqd->lock);
7081 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
7082 wake_up_process(sqd->thread);
7083 wait_for_completion(&sqd->completion);
7087 static void io_sq_thread_stop(struct io_sq_data *sqd)
7089 if (test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state))
7091 mutex_lock(&sqd->lock);
7093 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7094 WARN_ON_ONCE(test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state));
7095 wake_up_process(sqd->thread);
7096 mutex_unlock(&sqd->lock);
7097 wait_for_completion(&sqd->exited);
7098 WARN_ON_ONCE(sqd->thread);
7100 mutex_unlock(&sqd->lock);
7104 static void io_put_sq_data(struct io_sq_data *sqd)
7106 if (refcount_dec_and_test(&sqd->refs)) {
7107 io_sq_thread_stop(sqd);
7112 static void io_sq_thread_finish(struct io_ring_ctx *ctx)
7114 struct io_sq_data *sqd = ctx->sq_data;
7117 complete(&sqd->startup);
7119 wait_for_completion(&ctx->sq_thread_comp);
7120 io_sq_thread_park(sqd);
7123 mutex_lock(&sqd->ctx_lock);
7124 list_del(&ctx->sqd_list);
7125 io_sqd_update_thread_idle(sqd);
7126 mutex_unlock(&sqd->ctx_lock);
7129 io_sq_thread_unpark(sqd);
7131 io_put_sq_data(sqd);
7132 ctx->sq_data = NULL;
7136 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7138 struct io_ring_ctx *ctx_attach;
7139 struct io_sq_data *sqd;
7142 f = fdget(p->wq_fd);
7144 return ERR_PTR(-ENXIO);
7145 if (f.file->f_op != &io_uring_fops) {
7147 return ERR_PTR(-EINVAL);
7150 ctx_attach = f.file->private_data;
7151 sqd = ctx_attach->sq_data;
7154 return ERR_PTR(-EINVAL);
7157 refcount_inc(&sqd->refs);
7162 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p)
7164 struct io_sq_data *sqd;
7166 if (p->flags & IORING_SETUP_ATTACH_WQ)
7167 return io_attach_sq_data(p);
7169 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7171 return ERR_PTR(-ENOMEM);
7173 refcount_set(&sqd->refs, 1);
7174 INIT_LIST_HEAD(&sqd->ctx_list);
7175 INIT_LIST_HEAD(&sqd->ctx_new_list);
7176 mutex_init(&sqd->ctx_lock);
7177 mutex_init(&sqd->lock);
7178 init_waitqueue_head(&sqd->wait);
7179 init_completion(&sqd->startup);
7180 init_completion(&sqd->completion);
7181 init_completion(&sqd->exited);
7185 #if defined(CONFIG_UNIX)
7187 * Ensure the UNIX gc is aware of our file set, so we are certain that
7188 * the io_uring can be safely unregistered on process exit, even if we have
7189 * loops in the file referencing.
7191 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7193 struct sock *sk = ctx->ring_sock->sk;
7194 struct scm_fp_list *fpl;
7195 struct sk_buff *skb;
7198 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7202 skb = alloc_skb(0, GFP_KERNEL);
7211 fpl->user = get_uid(current_user());
7212 for (i = 0; i < nr; i++) {
7213 struct file *file = io_file_from_index(ctx, i + offset);
7217 fpl->fp[nr_files] = get_file(file);
7218 unix_inflight(fpl->user, fpl->fp[nr_files]);
7223 fpl->max = SCM_MAX_FD;
7224 fpl->count = nr_files;
7225 UNIXCB(skb).fp = fpl;
7226 skb->destructor = unix_destruct_scm;
7227 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7228 skb_queue_head(&sk->sk_receive_queue, skb);
7230 for (i = 0; i < nr_files; i++)
7241 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7242 * causes regular reference counting to break down. We rely on the UNIX
7243 * garbage collection to take care of this problem for us.
7245 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7247 unsigned left, total;
7251 left = ctx->nr_user_files;
7253 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7255 ret = __io_sqe_files_scm(ctx, this_files, total);
7259 total += this_files;
7265 while (total < ctx->nr_user_files) {
7266 struct file *file = io_file_from_index(ctx, total);
7276 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7282 static int io_sqe_alloc_file_tables(struct fixed_rsrc_data *file_data,
7283 unsigned nr_tables, unsigned nr_files)
7287 for (i = 0; i < nr_tables; i++) {
7288 struct fixed_rsrc_table *table = &file_data->table[i];
7289 unsigned this_files;
7291 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
7292 table->files = kcalloc(this_files, sizeof(struct file *),
7296 nr_files -= this_files;
7302 for (i = 0; i < nr_tables; i++) {
7303 struct fixed_rsrc_table *table = &file_data->table[i];
7304 kfree(table->files);
7309 static void io_ring_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
7311 struct file *file = prsrc->file;
7312 #if defined(CONFIG_UNIX)
7313 struct sock *sock = ctx->ring_sock->sk;
7314 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7315 struct sk_buff *skb;
7318 __skb_queue_head_init(&list);
7321 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7322 * remove this entry and rearrange the file array.
7324 skb = skb_dequeue(head);
7326 struct scm_fp_list *fp;
7328 fp = UNIXCB(skb).fp;
7329 for (i = 0; i < fp->count; i++) {
7332 if (fp->fp[i] != file)
7335 unix_notinflight(fp->user, fp->fp[i]);
7336 left = fp->count - 1 - i;
7338 memmove(&fp->fp[i], &fp->fp[i + 1],
7339 left * sizeof(struct file *));
7346 __skb_queue_tail(&list, skb);
7356 __skb_queue_tail(&list, skb);
7358 skb = skb_dequeue(head);
7361 if (skb_peek(&list)) {
7362 spin_lock_irq(&head->lock);
7363 while ((skb = __skb_dequeue(&list)) != NULL)
7364 __skb_queue_tail(head, skb);
7365 spin_unlock_irq(&head->lock);
7372 static void __io_rsrc_put_work(struct fixed_rsrc_ref_node *ref_node)
7374 struct fixed_rsrc_data *rsrc_data = ref_node->rsrc_data;
7375 struct io_ring_ctx *ctx = rsrc_data->ctx;
7376 struct io_rsrc_put *prsrc, *tmp;
7378 list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
7379 list_del(&prsrc->list);
7380 ref_node->rsrc_put(ctx, prsrc);
7384 percpu_ref_exit(&ref_node->refs);
7386 percpu_ref_put(&rsrc_data->refs);
7389 static void io_rsrc_put_work(struct work_struct *work)
7391 struct io_ring_ctx *ctx;
7392 struct llist_node *node;
7394 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
7395 node = llist_del_all(&ctx->rsrc_put_llist);
7398 struct fixed_rsrc_ref_node *ref_node;
7399 struct llist_node *next = node->next;
7401 ref_node = llist_entry(node, struct fixed_rsrc_ref_node, llist);
7402 __io_rsrc_put_work(ref_node);
7407 static struct file **io_fixed_file_slot(struct fixed_rsrc_data *file_data,
7410 struct fixed_rsrc_table *table;
7412 table = &file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7413 return &table->files[i & IORING_FILE_TABLE_MASK];
7416 static void io_rsrc_node_ref_zero(struct percpu_ref *ref)
7418 struct fixed_rsrc_ref_node *ref_node;
7419 struct fixed_rsrc_data *data;
7420 struct io_ring_ctx *ctx;
7421 bool first_add = false;
7424 ref_node = container_of(ref, struct fixed_rsrc_ref_node, refs);
7425 data = ref_node->rsrc_data;
7428 io_rsrc_ref_lock(ctx);
7429 ref_node->done = true;
7431 while (!list_empty(&ctx->rsrc_ref_list)) {
7432 ref_node = list_first_entry(&ctx->rsrc_ref_list,
7433 struct fixed_rsrc_ref_node, node);
7434 /* recycle ref nodes in order */
7435 if (!ref_node->done)
7437 list_del(&ref_node->node);
7438 first_add |= llist_add(&ref_node->llist, &ctx->rsrc_put_llist);
7440 io_rsrc_ref_unlock(ctx);
7442 if (percpu_ref_is_dying(&data->refs))
7446 mod_delayed_work(system_wq, &ctx->rsrc_put_work, 0);
7448 queue_delayed_work(system_wq, &ctx->rsrc_put_work, delay);
7451 static struct fixed_rsrc_ref_node *alloc_fixed_rsrc_ref_node(
7452 struct io_ring_ctx *ctx)
7454 struct fixed_rsrc_ref_node *ref_node;
7456 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7460 if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
7465 INIT_LIST_HEAD(&ref_node->node);
7466 INIT_LIST_HEAD(&ref_node->rsrc_list);
7467 ref_node->done = false;
7471 static void init_fixed_file_ref_node(struct io_ring_ctx *ctx,
7472 struct fixed_rsrc_ref_node *ref_node)
7474 ref_node->rsrc_data = ctx->file_data;
7475 ref_node->rsrc_put = io_ring_file_put;
7478 static void destroy_fixed_rsrc_ref_node(struct fixed_rsrc_ref_node *ref_node)
7480 percpu_ref_exit(&ref_node->refs);
7485 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7488 __s32 __user *fds = (__s32 __user *) arg;
7489 unsigned nr_tables, i;
7491 int fd, ret = -ENOMEM;
7492 struct fixed_rsrc_ref_node *ref_node;
7493 struct fixed_rsrc_data *file_data;
7499 if (nr_args > IORING_MAX_FIXED_FILES)
7502 file_data = alloc_fixed_rsrc_data(ctx);
7505 ctx->file_data = file_data;
7507 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
7508 file_data->table = kcalloc(nr_tables, sizeof(*file_data->table),
7510 if (!file_data->table)
7513 if (io_sqe_alloc_file_tables(file_data, nr_tables, nr_args))
7516 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7517 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7521 /* allow sparse sets */
7531 * Don't allow io_uring instances to be registered. If UNIX
7532 * isn't enabled, then this causes a reference cycle and this
7533 * instance can never get freed. If UNIX is enabled we'll
7534 * handle it just fine, but there's still no point in allowing
7535 * a ring fd as it doesn't support regular read/write anyway.
7537 if (file->f_op == &io_uring_fops) {
7541 *io_fixed_file_slot(file_data, i) = file;
7544 ret = io_sqe_files_scm(ctx);
7546 io_sqe_files_unregister(ctx);
7550 ref_node = alloc_fixed_rsrc_ref_node(ctx);
7552 io_sqe_files_unregister(ctx);
7555 init_fixed_file_ref_node(ctx, ref_node);
7557 io_sqe_rsrc_set_node(ctx, file_data, ref_node);
7560 for (i = 0; i < ctx->nr_user_files; i++) {
7561 file = io_file_from_index(ctx, i);
7565 for (i = 0; i < nr_tables; i++)
7566 kfree(file_data->table[i].files);
7567 ctx->nr_user_files = 0;
7569 free_fixed_rsrc_data(ctx->file_data);
7570 ctx->file_data = NULL;
7574 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7577 #if defined(CONFIG_UNIX)
7578 struct sock *sock = ctx->ring_sock->sk;
7579 struct sk_buff_head *head = &sock->sk_receive_queue;
7580 struct sk_buff *skb;
7583 * See if we can merge this file into an existing skb SCM_RIGHTS
7584 * file set. If there's no room, fall back to allocating a new skb
7585 * and filling it in.
7587 spin_lock_irq(&head->lock);
7588 skb = skb_peek(head);
7590 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7592 if (fpl->count < SCM_MAX_FD) {
7593 __skb_unlink(skb, head);
7594 spin_unlock_irq(&head->lock);
7595 fpl->fp[fpl->count] = get_file(file);
7596 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7598 spin_lock_irq(&head->lock);
7599 __skb_queue_head(head, skb);
7604 spin_unlock_irq(&head->lock);
7611 return __io_sqe_files_scm(ctx, 1, index);
7617 static int io_queue_rsrc_removal(struct fixed_rsrc_data *data, void *rsrc)
7619 struct io_rsrc_put *prsrc;
7620 struct fixed_rsrc_ref_node *ref_node = data->node;
7622 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
7627 list_add(&prsrc->list, &ref_node->rsrc_list);
7632 static inline int io_queue_file_removal(struct fixed_rsrc_data *data,
7635 return io_queue_rsrc_removal(data, (void *)file);
7638 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7639 struct io_uring_rsrc_update *up,
7642 struct fixed_rsrc_data *data = ctx->file_data;
7643 struct fixed_rsrc_ref_node *ref_node;
7644 struct file *file, **file_slot;
7648 bool needs_switch = false;
7650 if (check_add_overflow(up->offset, nr_args, &done))
7652 if (done > ctx->nr_user_files)
7655 ref_node = alloc_fixed_rsrc_ref_node(ctx);
7658 init_fixed_file_ref_node(ctx, ref_node);
7660 fds = u64_to_user_ptr(up->data);
7661 for (done = 0; done < nr_args; done++) {
7663 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
7667 if (fd == IORING_REGISTER_FILES_SKIP)
7670 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
7671 file_slot = io_fixed_file_slot(ctx->file_data, i);
7674 err = io_queue_file_removal(data, *file_slot);
7678 needs_switch = true;
7687 * Don't allow io_uring instances to be registered. If
7688 * UNIX isn't enabled, then this causes a reference
7689 * cycle and this instance can never get freed. If UNIX
7690 * is enabled we'll handle it just fine, but there's
7691 * still no point in allowing a ring fd as it doesn't
7692 * support regular read/write anyway.
7694 if (file->f_op == &io_uring_fops) {
7700 err = io_sqe_file_register(ctx, file, i);
7710 percpu_ref_kill(&data->node->refs);
7711 io_sqe_rsrc_set_node(ctx, data, ref_node);
7713 destroy_fixed_rsrc_ref_node(ref_node);
7715 return done ? done : err;
7718 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
7721 struct io_uring_rsrc_update up;
7723 if (!ctx->file_data)
7727 if (copy_from_user(&up, arg, sizeof(up)))
7732 return __io_sqe_files_update(ctx, &up, nr_args);
7735 static struct io_wq_work *io_free_work(struct io_wq_work *work)
7737 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7739 req = io_put_req_find_next(req);
7740 return req ? &req->work : NULL;
7743 static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx)
7745 struct io_wq_hash *hash;
7746 struct io_wq_data data;
7747 unsigned int concurrency;
7749 hash = ctx->hash_map;
7751 hash = kzalloc(sizeof(*hash), GFP_KERNEL);
7753 return ERR_PTR(-ENOMEM);
7754 refcount_set(&hash->refs, 1);
7755 init_waitqueue_head(&hash->wait);
7756 ctx->hash_map = hash;
7760 data.free_work = io_free_work;
7761 data.do_work = io_wq_submit_work;
7763 /* Do QD, or 4 * CPUS, whatever is smallest */
7764 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7766 return io_wq_create(concurrency, &data);
7769 static int io_uring_alloc_task_context(struct task_struct *task,
7770 struct io_ring_ctx *ctx)
7772 struct io_uring_task *tctx;
7775 tctx = kmalloc(sizeof(*tctx), GFP_KERNEL);
7776 if (unlikely(!tctx))
7779 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
7780 if (unlikely(ret)) {
7785 tctx->io_wq = io_init_wq_offload(ctx);
7786 if (IS_ERR(tctx->io_wq)) {
7787 ret = PTR_ERR(tctx->io_wq);
7788 percpu_counter_destroy(&tctx->inflight);
7794 init_waitqueue_head(&tctx->wait);
7796 atomic_set(&tctx->in_idle, 0);
7797 tctx->sqpoll = false;
7798 task->io_uring = tctx;
7799 spin_lock_init(&tctx->task_lock);
7800 INIT_WQ_LIST(&tctx->task_list);
7801 tctx->task_state = 0;
7802 init_task_work(&tctx->task_work, tctx_task_work);
7806 void __io_uring_free(struct task_struct *tsk)
7808 struct io_uring_task *tctx = tsk->io_uring;
7810 WARN_ON_ONCE(!xa_empty(&tctx->xa));
7811 WARN_ON_ONCE(tctx->io_wq);
7813 percpu_counter_destroy(&tctx->inflight);
7815 tsk->io_uring = NULL;
7818 static int io_sq_thread_fork(struct io_sq_data *sqd, struct io_ring_ctx *ctx)
7822 clear_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
7823 reinit_completion(&sqd->completion);
7825 sqd->task_pid = current->pid;
7826 current->flags |= PF_IO_WORKER;
7827 ret = io_wq_fork_thread(io_sq_thread, sqd);
7828 current->flags &= ~PF_IO_WORKER;
7833 wait_for_completion(&sqd->completion);
7834 return io_uring_alloc_task_context(sqd->thread, ctx);
7837 static int io_sq_offload_create(struct io_ring_ctx *ctx,
7838 struct io_uring_params *p)
7842 /* Retain compatibility with failing for an invalid attach attempt */
7843 if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
7844 IORING_SETUP_ATTACH_WQ) {
7847 f = fdget(p->wq_fd);
7850 if (f.file->f_op != &io_uring_fops) {
7856 if (ctx->flags & IORING_SETUP_SQPOLL) {
7857 struct io_sq_data *sqd;
7860 if (!capable(CAP_SYS_ADMIN) && !capable(CAP_SYS_NICE))
7863 sqd = io_get_sq_data(p);
7870 io_sq_thread_park(sqd);
7871 mutex_lock(&sqd->ctx_lock);
7872 list_add(&ctx->sqd_list, &sqd->ctx_new_list);
7873 mutex_unlock(&sqd->ctx_lock);
7874 io_sq_thread_unpark(sqd);
7876 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
7877 if (!ctx->sq_thread_idle)
7878 ctx->sq_thread_idle = HZ;
7883 if (p->flags & IORING_SETUP_SQ_AFF) {
7884 int cpu = p->sq_thread_cpu;
7887 if (cpu >= nr_cpu_ids)
7889 if (!cpu_online(cpu))
7897 sqd->task_pid = current->pid;
7898 current->flags |= PF_IO_WORKER;
7899 ret = io_wq_fork_thread(io_sq_thread, sqd);
7900 current->flags &= ~PF_IO_WORKER;
7905 wait_for_completion(&sqd->completion);
7906 ret = io_uring_alloc_task_context(sqd->thread, ctx);
7909 } else if (p->flags & IORING_SETUP_SQ_AFF) {
7910 /* Can't have SQ_AFF without SQPOLL */
7917 io_sq_thread_finish(ctx);
7921 static void io_sq_offload_start(struct io_ring_ctx *ctx)
7923 struct io_sq_data *sqd = ctx->sq_data;
7925 ctx->flags &= ~IORING_SETUP_R_DISABLED;
7926 if (ctx->flags & IORING_SETUP_SQPOLL)
7927 complete(&sqd->startup);
7930 static inline void __io_unaccount_mem(struct user_struct *user,
7931 unsigned long nr_pages)
7933 atomic_long_sub(nr_pages, &user->locked_vm);
7936 static inline int __io_account_mem(struct user_struct *user,
7937 unsigned long nr_pages)
7939 unsigned long page_limit, cur_pages, new_pages;
7941 /* Don't allow more pages than we can safely lock */
7942 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
7945 cur_pages = atomic_long_read(&user->locked_vm);
7946 new_pages = cur_pages + nr_pages;
7947 if (new_pages > page_limit)
7949 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
7950 new_pages) != cur_pages);
7955 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
7958 __io_unaccount_mem(ctx->user, nr_pages);
7960 if (ctx->mm_account)
7961 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
7964 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
7969 ret = __io_account_mem(ctx->user, nr_pages);
7974 if (ctx->mm_account)
7975 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
7980 static void io_mem_free(void *ptr)
7987 page = virt_to_head_page(ptr);
7988 if (put_page_testzero(page))
7989 free_compound_page(page);
7992 static void *io_mem_alloc(size_t size)
7994 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
7995 __GFP_NORETRY | __GFP_ACCOUNT;
7997 return (void *) __get_free_pages(gfp_flags, get_order(size));
8000 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8003 struct io_rings *rings;
8004 size_t off, sq_array_size;
8006 off = struct_size(rings, cqes, cq_entries);
8007 if (off == SIZE_MAX)
8011 off = ALIGN(off, SMP_CACHE_BYTES);
8019 sq_array_size = array_size(sizeof(u32), sq_entries);
8020 if (sq_array_size == SIZE_MAX)
8023 if (check_add_overflow(off, sq_array_size, &off))
8029 static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8033 if (!ctx->user_bufs)
8036 for (i = 0; i < ctx->nr_user_bufs; i++) {
8037 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8039 for (j = 0; j < imu->nr_bvecs; j++)
8040 unpin_user_page(imu->bvec[j].bv_page);
8042 if (imu->acct_pages)
8043 io_unaccount_mem(ctx, imu->acct_pages);
8048 kfree(ctx->user_bufs);
8049 ctx->user_bufs = NULL;
8050 ctx->nr_user_bufs = 0;
8054 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8055 void __user *arg, unsigned index)
8057 struct iovec __user *src;
8059 #ifdef CONFIG_COMPAT
8061 struct compat_iovec __user *ciovs;
8062 struct compat_iovec ciov;
8064 ciovs = (struct compat_iovec __user *) arg;
8065 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8068 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8069 dst->iov_len = ciov.iov_len;
8073 src = (struct iovec __user *) arg;
8074 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8080 * Not super efficient, but this is just a registration time. And we do cache
8081 * the last compound head, so generally we'll only do a full search if we don't
8084 * We check if the given compound head page has already been accounted, to
8085 * avoid double accounting it. This allows us to account the full size of the
8086 * page, not just the constituent pages of a huge page.
8088 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8089 int nr_pages, struct page *hpage)
8093 /* check current page array */
8094 for (i = 0; i < nr_pages; i++) {
8095 if (!PageCompound(pages[i]))
8097 if (compound_head(pages[i]) == hpage)
8101 /* check previously registered pages */
8102 for (i = 0; i < ctx->nr_user_bufs; i++) {
8103 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8105 for (j = 0; j < imu->nr_bvecs; j++) {
8106 if (!PageCompound(imu->bvec[j].bv_page))
8108 if (compound_head(imu->bvec[j].bv_page) == hpage)
8116 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8117 int nr_pages, struct io_mapped_ubuf *imu,
8118 struct page **last_hpage)
8122 for (i = 0; i < nr_pages; i++) {
8123 if (!PageCompound(pages[i])) {
8128 hpage = compound_head(pages[i]);
8129 if (hpage == *last_hpage)
8131 *last_hpage = hpage;
8132 if (headpage_already_acct(ctx, pages, i, hpage))
8134 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8138 if (!imu->acct_pages)
8141 ret = io_account_mem(ctx, imu->acct_pages);
8143 imu->acct_pages = 0;
8147 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
8148 struct io_mapped_ubuf *imu,
8149 struct page **last_hpage)
8151 struct vm_area_struct **vmas = NULL;
8152 struct page **pages = NULL;
8153 unsigned long off, start, end, ubuf;
8155 int ret, pret, nr_pages, i;
8157 ubuf = (unsigned long) iov->iov_base;
8158 end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8159 start = ubuf >> PAGE_SHIFT;
8160 nr_pages = end - start;
8164 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
8168 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
8173 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
8179 mmap_read_lock(current->mm);
8180 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
8182 if (pret == nr_pages) {
8183 /* don't support file backed memory */
8184 for (i = 0; i < nr_pages; i++) {
8185 struct vm_area_struct *vma = vmas[i];
8188 !is_file_hugepages(vma->vm_file)) {
8194 ret = pret < 0 ? pret : -EFAULT;
8196 mmap_read_unlock(current->mm);
8199 * if we did partial map, or found file backed vmas,
8200 * release any pages we did get
8203 unpin_user_pages(pages, pret);
8208 ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
8210 unpin_user_pages(pages, pret);
8215 off = ubuf & ~PAGE_MASK;
8216 size = iov->iov_len;
8217 for (i = 0; i < nr_pages; i++) {
8220 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8221 imu->bvec[i].bv_page = pages[i];
8222 imu->bvec[i].bv_len = vec_len;
8223 imu->bvec[i].bv_offset = off;
8227 /* store original address for later verification */
8229 imu->len = iov->iov_len;
8230 imu->nr_bvecs = nr_pages;
8238 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
8242 if (!nr_args || nr_args > UIO_MAXIOV)
8245 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
8247 if (!ctx->user_bufs)
8253 static int io_buffer_validate(struct iovec *iov)
8256 * Don't impose further limits on the size and buffer
8257 * constraints here, we'll -EINVAL later when IO is
8258 * submitted if they are wrong.
8260 if (!iov->iov_base || !iov->iov_len)
8263 /* arbitrary limit, but we need something */
8264 if (iov->iov_len > SZ_1G)
8270 static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
8271 unsigned int nr_args)
8275 struct page *last_hpage = NULL;
8277 ret = io_buffers_map_alloc(ctx, nr_args);
8281 for (i = 0; i < nr_args; i++) {
8282 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8284 ret = io_copy_iov(ctx, &iov, arg, i);
8288 ret = io_buffer_validate(&iov);
8292 ret = io_sqe_buffer_register(ctx, &iov, imu, &last_hpage);
8296 ctx->nr_user_bufs++;
8300 io_sqe_buffers_unregister(ctx);
8305 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8307 __s32 __user *fds = arg;
8313 if (copy_from_user(&fd, fds, sizeof(*fds)))
8316 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8317 if (IS_ERR(ctx->cq_ev_fd)) {
8318 int ret = PTR_ERR(ctx->cq_ev_fd);
8319 ctx->cq_ev_fd = NULL;
8326 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8328 if (ctx->cq_ev_fd) {
8329 eventfd_ctx_put(ctx->cq_ev_fd);
8330 ctx->cq_ev_fd = NULL;
8337 static int __io_destroy_buffers(int id, void *p, void *data)
8339 struct io_ring_ctx *ctx = data;
8340 struct io_buffer *buf = p;
8342 __io_remove_buffers(ctx, buf, id, -1U);
8346 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8348 idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx);
8349 idr_destroy(&ctx->io_buffer_idr);
8352 static void io_req_cache_free(struct list_head *list, struct task_struct *tsk)
8354 struct io_kiocb *req, *nxt;
8356 list_for_each_entry_safe(req, nxt, list, compl.list) {
8357 if (tsk && req->task != tsk)
8359 list_del(&req->compl.list);
8360 kmem_cache_free(req_cachep, req);
8364 static void io_req_caches_free(struct io_ring_ctx *ctx)
8366 struct io_submit_state *submit_state = &ctx->submit_state;
8367 struct io_comp_state *cs = &ctx->submit_state.comp;
8369 mutex_lock(&ctx->uring_lock);
8371 if (submit_state->free_reqs) {
8372 kmem_cache_free_bulk(req_cachep, submit_state->free_reqs,
8373 submit_state->reqs);
8374 submit_state->free_reqs = 0;
8377 spin_lock_irq(&ctx->completion_lock);
8378 list_splice_init(&cs->locked_free_list, &cs->free_list);
8379 cs->locked_free_nr = 0;
8380 spin_unlock_irq(&ctx->completion_lock);
8382 io_req_cache_free(&cs->free_list, NULL);
8384 mutex_unlock(&ctx->uring_lock);
8387 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8390 * Some may use context even when all refs and requests have been put,
8391 * and they are free to do so while still holding uring_lock, see
8392 * __io_req_task_submit(). Wait for them to finish.
8394 mutex_lock(&ctx->uring_lock);
8395 mutex_unlock(&ctx->uring_lock);
8397 io_sq_thread_finish(ctx);
8398 io_sqe_buffers_unregister(ctx);
8400 if (ctx->mm_account) {
8401 mmdrop(ctx->mm_account);
8402 ctx->mm_account = NULL;
8405 mutex_lock(&ctx->uring_lock);
8406 io_sqe_files_unregister(ctx);
8407 mutex_unlock(&ctx->uring_lock);
8408 io_eventfd_unregister(ctx);
8409 io_destroy_buffers(ctx);
8410 idr_destroy(&ctx->personality_idr);
8412 #if defined(CONFIG_UNIX)
8413 if (ctx->ring_sock) {
8414 ctx->ring_sock->file = NULL; /* so that iput() is called */
8415 sock_release(ctx->ring_sock);
8419 io_mem_free(ctx->rings);
8420 io_mem_free(ctx->sq_sqes);
8422 percpu_ref_exit(&ctx->refs);
8423 free_uid(ctx->user);
8424 io_req_caches_free(ctx);
8426 io_wq_put_hash(ctx->hash_map);
8427 kfree(ctx->cancel_hash);
8431 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8433 struct io_ring_ctx *ctx = file->private_data;
8436 poll_wait(file, &ctx->cq_wait, wait);
8438 * synchronizes with barrier from wq_has_sleeper call in
8442 if (!io_sqring_full(ctx))
8443 mask |= EPOLLOUT | EPOLLWRNORM;
8446 * Don't flush cqring overflow list here, just do a simple check.
8447 * Otherwise there could possible be ABBA deadlock:
8450 * lock(&ctx->uring_lock);
8452 * lock(&ctx->uring_lock);
8455 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
8456 * pushs them to do the flush.
8458 if (io_cqring_events(ctx) || test_bit(0, &ctx->cq_check_overflow))
8459 mask |= EPOLLIN | EPOLLRDNORM;
8464 static int io_uring_fasync(int fd, struct file *file, int on)
8466 struct io_ring_ctx *ctx = file->private_data;
8468 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8471 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8473 const struct cred *creds;
8475 creds = idr_remove(&ctx->personality_idr, id);
8484 static int io_remove_personalities(int id, void *p, void *data)
8486 struct io_ring_ctx *ctx = data;
8488 io_unregister_personality(ctx, id);
8492 static bool io_run_ctx_fallback(struct io_ring_ctx *ctx)
8494 struct callback_head *work, *next;
8495 bool executed = false;
8498 work = xchg(&ctx->exit_task_work, NULL);
8514 static void io_ring_exit_work(struct work_struct *work)
8516 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
8520 * If we're doing polled IO and end up having requests being
8521 * submitted async (out-of-line), then completions can come in while
8522 * we're waiting for refs to drop. We need to reap these manually,
8523 * as nobody else will be looking for them.
8526 io_uring_try_cancel_requests(ctx, NULL, NULL);
8527 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
8528 io_ring_ctx_free(ctx);
8531 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8533 mutex_lock(&ctx->uring_lock);
8534 percpu_ref_kill(&ctx->refs);
8535 /* if force is set, the ring is going away. always drop after that */
8536 ctx->cq_overflow_flushed = 1;
8538 __io_cqring_overflow_flush(ctx, true, NULL, NULL);
8539 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
8540 mutex_unlock(&ctx->uring_lock);
8542 io_kill_timeouts(ctx, NULL, NULL);
8543 io_poll_remove_all(ctx, NULL, NULL);
8545 /* if we failed setting up the ctx, we might not have any rings */
8546 io_iopoll_try_reap_events(ctx);
8548 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8550 * Use system_unbound_wq to avoid spawning tons of event kworkers
8551 * if we're exiting a ton of rings at the same time. It just adds
8552 * noise and overhead, there's no discernable change in runtime
8553 * over using system_wq.
8555 queue_work(system_unbound_wq, &ctx->exit_work);
8558 static int io_uring_release(struct inode *inode, struct file *file)
8560 struct io_ring_ctx *ctx = file->private_data;
8562 file->private_data = NULL;
8563 io_ring_ctx_wait_and_kill(ctx);
8567 struct io_task_cancel {
8568 struct task_struct *task;
8569 struct files_struct *files;
8572 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8574 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8575 struct io_task_cancel *cancel = data;
8578 if (cancel->files && (req->flags & REQ_F_LINK_TIMEOUT)) {
8579 unsigned long flags;
8580 struct io_ring_ctx *ctx = req->ctx;
8582 /* protect against races with linked timeouts */
8583 spin_lock_irqsave(&ctx->completion_lock, flags);
8584 ret = io_match_task(req, cancel->task, cancel->files);
8585 spin_unlock_irqrestore(&ctx->completion_lock, flags);
8587 ret = io_match_task(req, cancel->task, cancel->files);
8592 static void io_cancel_defer_files(struct io_ring_ctx *ctx,
8593 struct task_struct *task,
8594 struct files_struct *files)
8596 struct io_defer_entry *de = NULL;
8599 spin_lock_irq(&ctx->completion_lock);
8600 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8601 if (io_match_task(de->req, task, files)) {
8602 list_cut_position(&list, &ctx->defer_list, &de->list);
8606 spin_unlock_irq(&ctx->completion_lock);
8608 while (!list_empty(&list)) {
8609 de = list_first_entry(&list, struct io_defer_entry, list);
8610 list_del_init(&de->list);
8611 req_set_fail_links(de->req);
8612 io_put_req(de->req);
8613 io_req_complete(de->req, -ECANCELED);
8618 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
8619 struct task_struct *task,
8620 struct files_struct *files)
8622 struct io_task_cancel cancel = { .task = task, .files = files, };
8623 struct task_struct *tctx_task = task ?: current;
8624 struct io_uring_task *tctx = tctx_task->io_uring;
8627 enum io_wq_cancel cret;
8630 if (tctx && tctx->io_wq) {
8631 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
8633 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8636 /* SQPOLL thread does its own polling */
8637 if (!(ctx->flags & IORING_SETUP_SQPOLL) && !files) {
8638 while (!list_empty_careful(&ctx->iopoll_list)) {
8639 io_iopoll_try_reap_events(ctx);
8644 ret |= io_poll_remove_all(ctx, task, files);
8645 ret |= io_kill_timeouts(ctx, task, files);
8646 ret |= io_run_task_work();
8647 ret |= io_run_ctx_fallback(ctx);
8648 io_cqring_overflow_flush(ctx, true, task, files);
8655 static int io_uring_count_inflight(struct io_ring_ctx *ctx,
8656 struct task_struct *task,
8657 struct files_struct *files)
8659 struct io_kiocb *req;
8662 spin_lock_irq(&ctx->inflight_lock);
8663 list_for_each_entry(req, &ctx->inflight_list, inflight_entry)
8664 cnt += io_match_task(req, task, files);
8665 spin_unlock_irq(&ctx->inflight_lock);
8669 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
8670 struct task_struct *task,
8671 struct files_struct *files)
8673 while (!list_empty_careful(&ctx->inflight_list)) {
8677 inflight = io_uring_count_inflight(ctx, task, files);
8681 io_uring_try_cancel_requests(ctx, task, files);
8684 io_sq_thread_unpark(ctx->sq_data);
8685 prepare_to_wait(&task->io_uring->wait, &wait,
8686 TASK_UNINTERRUPTIBLE);
8687 if (inflight == io_uring_count_inflight(ctx, task, files))
8689 finish_wait(&task->io_uring->wait, &wait);
8691 io_sq_thread_park(ctx->sq_data);
8696 * We need to iteratively cancel requests, in case a request has dependent
8697 * hard links. These persist even for failure of cancelations, hence keep
8698 * looping until none are found.
8700 static void io_uring_cancel_task_requests(struct io_ring_ctx *ctx,
8701 struct files_struct *files)
8703 struct task_struct *task = current;
8704 bool did_park = false;
8706 if ((ctx->flags & IORING_SETUP_SQPOLL) && ctx->sq_data) {
8707 /* never started, nothing to cancel */
8708 if (ctx->flags & IORING_SETUP_R_DISABLED) {
8709 io_sq_offload_start(ctx);
8712 did_park = io_sq_thread_park(ctx->sq_data);
8714 task = ctx->sq_data->thread;
8715 atomic_inc(&task->io_uring->in_idle);
8719 io_cancel_defer_files(ctx, task, files);
8721 io_uring_cancel_files(ctx, task, files);
8723 io_uring_try_cancel_requests(ctx, task, NULL);
8726 atomic_dec(&task->io_uring->in_idle);
8727 io_sq_thread_unpark(ctx->sq_data);
8732 * Note that this task has used io_uring. We use it for cancelation purposes.
8734 static int io_uring_add_task_file(struct io_ring_ctx *ctx, struct file *file)
8736 struct io_uring_task *tctx = current->io_uring;
8739 if (unlikely(!tctx)) {
8740 ret = io_uring_alloc_task_context(current, ctx);
8743 tctx = current->io_uring;
8745 if (tctx->last != file) {
8746 void *old = xa_load(&tctx->xa, (unsigned long)file);
8750 ret = xa_err(xa_store(&tctx->xa, (unsigned long)file,
8761 * This is race safe in that the task itself is doing this, hence it
8762 * cannot be going through the exit/cancel paths at the same time.
8763 * This cannot be modified while exit/cancel is running.
8765 if (!tctx->sqpoll && (ctx->flags & IORING_SETUP_SQPOLL))
8766 tctx->sqpoll = true;
8772 * Remove this io_uring_file -> task mapping.
8774 static void io_uring_del_task_file(struct file *file)
8776 struct io_uring_task *tctx = current->io_uring;
8778 if (tctx->last == file)
8780 file = xa_erase(&tctx->xa, (unsigned long)file);
8785 static void io_uring_clean_tctx(struct io_uring_task *tctx)
8788 unsigned long index;
8790 xa_for_each(&tctx->xa, index, file)
8791 io_uring_del_task_file(file);
8793 io_wq_put_and_exit(tctx->io_wq);
8798 void __io_uring_files_cancel(struct files_struct *files)
8800 struct io_uring_task *tctx = current->io_uring;
8802 unsigned long index;
8804 /* make sure overflow events are dropped */
8805 atomic_inc(&tctx->in_idle);
8806 xa_for_each(&tctx->xa, index, file)
8807 io_uring_cancel_task_requests(file->private_data, files);
8808 atomic_dec(&tctx->in_idle);
8811 io_uring_clean_tctx(tctx);
8814 static s64 tctx_inflight(struct io_uring_task *tctx)
8816 return percpu_counter_sum(&tctx->inflight);
8819 static void io_uring_cancel_sqpoll(struct io_ring_ctx *ctx)
8821 struct io_sq_data *sqd = ctx->sq_data;
8822 struct io_uring_task *tctx;
8828 if (!io_sq_thread_park(sqd))
8830 tctx = ctx->sq_data->thread->io_uring;
8831 /* can happen on fork/alloc failure, just ignore that state */
8833 io_sq_thread_unpark(sqd);
8837 atomic_inc(&tctx->in_idle);
8839 /* read completions before cancelations */
8840 inflight = tctx_inflight(tctx);
8843 io_uring_cancel_task_requests(ctx, NULL);
8845 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
8847 * If we've seen completions, retry without waiting. This
8848 * avoids a race where a completion comes in before we did
8849 * prepare_to_wait().
8851 if (inflight == tctx_inflight(tctx))
8853 finish_wait(&tctx->wait, &wait);
8855 atomic_dec(&tctx->in_idle);
8856 io_sq_thread_unpark(sqd);
8860 * Find any io_uring fd that this task has registered or done IO on, and cancel
8863 void __io_uring_task_cancel(void)
8865 struct io_uring_task *tctx = current->io_uring;
8869 /* make sure overflow events are dropped */
8870 atomic_inc(&tctx->in_idle);
8874 unsigned long index;
8876 xa_for_each(&tctx->xa, index, file)
8877 io_uring_cancel_sqpoll(file->private_data);
8881 /* read completions before cancelations */
8882 inflight = tctx_inflight(tctx);
8885 __io_uring_files_cancel(NULL);
8887 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
8890 * If we've seen completions, retry without waiting. This
8891 * avoids a race where a completion comes in before we did
8892 * prepare_to_wait().
8894 if (inflight == tctx_inflight(tctx))
8896 finish_wait(&tctx->wait, &wait);
8899 atomic_dec(&tctx->in_idle);
8901 io_uring_clean_tctx(tctx);
8902 /* all current's requests should be gone, we can kill tctx */
8903 __io_uring_free(current);
8906 static void *io_uring_validate_mmap_request(struct file *file,
8907 loff_t pgoff, size_t sz)
8909 struct io_ring_ctx *ctx = file->private_data;
8910 loff_t offset = pgoff << PAGE_SHIFT;
8915 case IORING_OFF_SQ_RING:
8916 case IORING_OFF_CQ_RING:
8919 case IORING_OFF_SQES:
8923 return ERR_PTR(-EINVAL);
8926 page = virt_to_head_page(ptr);
8927 if (sz > page_size(page))
8928 return ERR_PTR(-EINVAL);
8935 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
8937 size_t sz = vma->vm_end - vma->vm_start;
8941 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
8943 return PTR_ERR(ptr);
8945 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
8946 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
8949 #else /* !CONFIG_MMU */
8951 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
8953 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
8956 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
8958 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
8961 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
8962 unsigned long addr, unsigned long len,
8963 unsigned long pgoff, unsigned long flags)
8967 ptr = io_uring_validate_mmap_request(file, pgoff, len);
8969 return PTR_ERR(ptr);
8971 return (unsigned long) ptr;
8974 #endif /* !CONFIG_MMU */
8976 static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
8982 if (!io_sqring_full(ctx))
8984 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
8986 if (!io_sqring_full(ctx))
8989 } while (!signal_pending(current));
8991 finish_wait(&ctx->sqo_sq_wait, &wait);
8995 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
8996 struct __kernel_timespec __user **ts,
8997 const sigset_t __user **sig)
8999 struct io_uring_getevents_arg arg;
9002 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9003 * is just a pointer to the sigset_t.
9005 if (!(flags & IORING_ENTER_EXT_ARG)) {
9006 *sig = (const sigset_t __user *) argp;
9012 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9013 * timespec and sigset_t pointers if good.
9015 if (*argsz != sizeof(arg))
9017 if (copy_from_user(&arg, argp, sizeof(arg)))
9019 *sig = u64_to_user_ptr(arg.sigmask);
9020 *argsz = arg.sigmask_sz;
9021 *ts = u64_to_user_ptr(arg.ts);
9025 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9026 u32, min_complete, u32, flags, const void __user *, argp,
9029 struct io_ring_ctx *ctx;
9036 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9037 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG))
9045 if (f.file->f_op != &io_uring_fops)
9049 ctx = f.file->private_data;
9050 if (!percpu_ref_tryget(&ctx->refs))
9054 if (ctx->flags & IORING_SETUP_R_DISABLED)
9058 * For SQ polling, the thread will do all submissions and completions.
9059 * Just return the requested submit count, and wake the thread if
9063 if (ctx->flags & IORING_SETUP_SQPOLL) {
9064 io_cqring_overflow_flush(ctx, false, NULL, NULL);
9066 if (unlikely(ctx->sqo_exec)) {
9067 ret = io_sq_thread_fork(ctx->sq_data, ctx);
9073 if (flags & IORING_ENTER_SQ_WAKEUP)
9074 wake_up(&ctx->sq_data->wait);
9075 if (flags & IORING_ENTER_SQ_WAIT) {
9076 ret = io_sqpoll_wait_sq(ctx);
9080 submitted = to_submit;
9081 } else if (to_submit) {
9082 ret = io_uring_add_task_file(ctx, f.file);
9085 mutex_lock(&ctx->uring_lock);
9086 submitted = io_submit_sqes(ctx, to_submit);
9087 mutex_unlock(&ctx->uring_lock);
9089 if (submitted != to_submit)
9092 if (flags & IORING_ENTER_GETEVENTS) {
9093 const sigset_t __user *sig;
9094 struct __kernel_timespec __user *ts;
9096 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
9100 min_complete = min(min_complete, ctx->cq_entries);
9103 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9104 * space applications don't need to do io completion events
9105 * polling again, they can rely on io_sq_thread to do polling
9106 * work, which can reduce cpu usage and uring_lock contention.
9108 if (ctx->flags & IORING_SETUP_IOPOLL &&
9109 !(ctx->flags & IORING_SETUP_SQPOLL)) {
9110 ret = io_iopoll_check(ctx, min_complete);
9112 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
9117 percpu_ref_put(&ctx->refs);
9120 return submitted ? submitted : ret;
9123 #ifdef CONFIG_PROC_FS
9124 static int io_uring_show_cred(int id, void *p, void *data)
9126 const struct cred *cred = p;
9127 struct seq_file *m = data;
9128 struct user_namespace *uns = seq_user_ns(m);
9129 struct group_info *gi;
9134 seq_printf(m, "%5d\n", id);
9135 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
9136 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
9137 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
9138 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
9139 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
9140 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
9141 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
9142 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
9143 seq_puts(m, "\n\tGroups:\t");
9144 gi = cred->group_info;
9145 for (g = 0; g < gi->ngroups; g++) {
9146 seq_put_decimal_ull(m, g ? " " : "",
9147 from_kgid_munged(uns, gi->gid[g]));
9149 seq_puts(m, "\n\tCapEff:\t");
9150 cap = cred->cap_effective;
9151 CAP_FOR_EACH_U32(__capi)
9152 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
9157 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
9159 struct io_sq_data *sq = NULL;
9164 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9165 * since fdinfo case grabs it in the opposite direction of normal use
9166 * cases. If we fail to get the lock, we just don't iterate any
9167 * structures that could be going away outside the io_uring mutex.
9169 has_lock = mutex_trylock(&ctx->uring_lock);
9171 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL)) {
9177 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9178 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9179 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9180 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9181 struct file *f = *io_fixed_file_slot(ctx->file_data, i);
9184 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9186 seq_printf(m, "%5u: <none>\n", i);
9188 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9189 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9190 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
9192 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
9193 (unsigned int) buf->len);
9195 if (has_lock && !idr_is_empty(&ctx->personality_idr)) {
9196 seq_printf(m, "Personalities:\n");
9197 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
9199 seq_printf(m, "PollList:\n");
9200 spin_lock_irq(&ctx->completion_lock);
9201 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9202 struct hlist_head *list = &ctx->cancel_hash[i];
9203 struct io_kiocb *req;
9205 hlist_for_each_entry(req, list, hash_node)
9206 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9207 req->task->task_works != NULL);
9209 spin_unlock_irq(&ctx->completion_lock);
9211 mutex_unlock(&ctx->uring_lock);
9214 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9216 struct io_ring_ctx *ctx = f->private_data;
9218 if (percpu_ref_tryget(&ctx->refs)) {
9219 __io_uring_show_fdinfo(ctx, m);
9220 percpu_ref_put(&ctx->refs);
9225 static const struct file_operations io_uring_fops = {
9226 .release = io_uring_release,
9227 .mmap = io_uring_mmap,
9229 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9230 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9232 .poll = io_uring_poll,
9233 .fasync = io_uring_fasync,
9234 #ifdef CONFIG_PROC_FS
9235 .show_fdinfo = io_uring_show_fdinfo,
9239 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9240 struct io_uring_params *p)
9242 struct io_rings *rings;
9243 size_t size, sq_array_offset;
9245 /* make sure these are sane, as we already accounted them */
9246 ctx->sq_entries = p->sq_entries;
9247 ctx->cq_entries = p->cq_entries;
9249 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9250 if (size == SIZE_MAX)
9253 rings = io_mem_alloc(size);
9258 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9259 rings->sq_ring_mask = p->sq_entries - 1;
9260 rings->cq_ring_mask = p->cq_entries - 1;
9261 rings->sq_ring_entries = p->sq_entries;
9262 rings->cq_ring_entries = p->cq_entries;
9263 ctx->sq_mask = rings->sq_ring_mask;
9264 ctx->cq_mask = rings->cq_ring_mask;
9266 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9267 if (size == SIZE_MAX) {
9268 io_mem_free(ctx->rings);
9273 ctx->sq_sqes = io_mem_alloc(size);
9274 if (!ctx->sq_sqes) {
9275 io_mem_free(ctx->rings);
9283 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
9287 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9291 ret = io_uring_add_task_file(ctx, file);
9296 fd_install(fd, file);
9301 * Allocate an anonymous fd, this is what constitutes the application
9302 * visible backing of an io_uring instance. The application mmaps this
9303 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9304 * we have to tie this fd to a socket for file garbage collection purposes.
9306 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
9309 #if defined(CONFIG_UNIX)
9312 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9315 return ERR_PTR(ret);
9318 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9319 O_RDWR | O_CLOEXEC);
9320 #if defined(CONFIG_UNIX)
9322 sock_release(ctx->ring_sock);
9323 ctx->ring_sock = NULL;
9325 ctx->ring_sock->file = file;
9331 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9332 struct io_uring_params __user *params)
9334 struct io_ring_ctx *ctx;
9340 if (entries > IORING_MAX_ENTRIES) {
9341 if (!(p->flags & IORING_SETUP_CLAMP))
9343 entries = IORING_MAX_ENTRIES;
9347 * Use twice as many entries for the CQ ring. It's possible for the
9348 * application to drive a higher depth than the size of the SQ ring,
9349 * since the sqes are only used at submission time. This allows for
9350 * some flexibility in overcommitting a bit. If the application has
9351 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9352 * of CQ ring entries manually.
9354 p->sq_entries = roundup_pow_of_two(entries);
9355 if (p->flags & IORING_SETUP_CQSIZE) {
9357 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9358 * to a power-of-two, if it isn't already. We do NOT impose
9359 * any cq vs sq ring sizing.
9363 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9364 if (!(p->flags & IORING_SETUP_CLAMP))
9366 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9368 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9369 if (p->cq_entries < p->sq_entries)
9372 p->cq_entries = 2 * p->sq_entries;
9375 ctx = io_ring_ctx_alloc(p);
9378 ctx->compat = in_compat_syscall();
9379 if (!capable(CAP_IPC_LOCK))
9380 ctx->user = get_uid(current_user());
9383 * This is just grabbed for accounting purposes. When a process exits,
9384 * the mm is exited and dropped before the files, hence we need to hang
9385 * on to this mm purely for the purposes of being able to unaccount
9386 * memory (locked/pinned vm). It's not used for anything else.
9388 mmgrab(current->mm);
9389 ctx->mm_account = current->mm;
9391 ret = io_allocate_scq_urings(ctx, p);
9395 ret = io_sq_offload_create(ctx, p);
9399 if (!(p->flags & IORING_SETUP_R_DISABLED))
9400 io_sq_offload_start(ctx);
9402 memset(&p->sq_off, 0, sizeof(p->sq_off));
9403 p->sq_off.head = offsetof(struct io_rings, sq.head);
9404 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9405 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9406 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9407 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9408 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9409 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9411 memset(&p->cq_off, 0, sizeof(p->cq_off));
9412 p->cq_off.head = offsetof(struct io_rings, cq.head);
9413 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9414 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9415 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9416 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9417 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9418 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9420 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9421 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9422 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9423 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
9424 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS;
9426 if (copy_to_user(params, p, sizeof(*p))) {
9431 file = io_uring_get_file(ctx);
9433 ret = PTR_ERR(file);
9438 * Install ring fd as the very last thing, so we don't risk someone
9439 * having closed it before we finish setup
9441 ret = io_uring_install_fd(ctx, file);
9443 /* fput will clean it up */
9448 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9451 io_ring_ctx_wait_and_kill(ctx);
9456 * Sets up an aio uring context, and returns the fd. Applications asks for a
9457 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9458 * params structure passed in.
9460 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9462 struct io_uring_params p;
9465 if (copy_from_user(&p, params, sizeof(p)))
9467 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9472 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9473 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9474 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9475 IORING_SETUP_R_DISABLED))
9478 return io_uring_create(entries, &p, params);
9481 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9482 struct io_uring_params __user *, params)
9484 return io_uring_setup(entries, params);
9487 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9489 struct io_uring_probe *p;
9493 size = struct_size(p, ops, nr_args);
9494 if (size == SIZE_MAX)
9496 p = kzalloc(size, GFP_KERNEL);
9501 if (copy_from_user(p, arg, size))
9504 if (memchr_inv(p, 0, size))
9507 p->last_op = IORING_OP_LAST - 1;
9508 if (nr_args > IORING_OP_LAST)
9509 nr_args = IORING_OP_LAST;
9511 for (i = 0; i < nr_args; i++) {
9513 if (!io_op_defs[i].not_supported)
9514 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9519 if (copy_to_user(arg, p, size))
9526 static int io_register_personality(struct io_ring_ctx *ctx)
9528 const struct cred *creds;
9531 creds = get_current_cred();
9533 ret = idr_alloc_cyclic(&ctx->personality_idr, (void *) creds, 1,
9534 USHRT_MAX, GFP_KERNEL);
9540 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9541 unsigned int nr_args)
9543 struct io_uring_restriction *res;
9547 /* Restrictions allowed only if rings started disabled */
9548 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9551 /* We allow only a single restrictions registration */
9552 if (ctx->restrictions.registered)
9555 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9558 size = array_size(nr_args, sizeof(*res));
9559 if (size == SIZE_MAX)
9562 res = memdup_user(arg, size);
9564 return PTR_ERR(res);
9568 for (i = 0; i < nr_args; i++) {
9569 switch (res[i].opcode) {
9570 case IORING_RESTRICTION_REGISTER_OP:
9571 if (res[i].register_op >= IORING_REGISTER_LAST) {
9576 __set_bit(res[i].register_op,
9577 ctx->restrictions.register_op);
9579 case IORING_RESTRICTION_SQE_OP:
9580 if (res[i].sqe_op >= IORING_OP_LAST) {
9585 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9587 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9588 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9590 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9591 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9600 /* Reset all restrictions if an error happened */
9602 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9604 ctx->restrictions.registered = true;
9610 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9612 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9615 if (ctx->restrictions.registered)
9616 ctx->restricted = 1;
9618 io_sq_offload_start(ctx);
9622 static bool io_register_op_must_quiesce(int op)
9625 case IORING_UNREGISTER_FILES:
9626 case IORING_REGISTER_FILES_UPDATE:
9627 case IORING_REGISTER_PROBE:
9628 case IORING_REGISTER_PERSONALITY:
9629 case IORING_UNREGISTER_PERSONALITY:
9636 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
9637 void __user *arg, unsigned nr_args)
9638 __releases(ctx->uring_lock)
9639 __acquires(ctx->uring_lock)
9644 * We're inside the ring mutex, if the ref is already dying, then
9645 * someone else killed the ctx or is already going through
9646 * io_uring_register().
9648 if (percpu_ref_is_dying(&ctx->refs))
9651 if (io_register_op_must_quiesce(opcode)) {
9652 percpu_ref_kill(&ctx->refs);
9655 * Drop uring mutex before waiting for references to exit. If
9656 * another thread is currently inside io_uring_enter() it might
9657 * need to grab the uring_lock to make progress. If we hold it
9658 * here across the drain wait, then we can deadlock. It's safe
9659 * to drop the mutex here, since no new references will come in
9660 * after we've killed the percpu ref.
9662 mutex_unlock(&ctx->uring_lock);
9664 ret = wait_for_completion_interruptible(&ctx->ref_comp);
9667 ret = io_run_task_work_sig();
9672 mutex_lock(&ctx->uring_lock);
9675 percpu_ref_resurrect(&ctx->refs);
9680 if (ctx->restricted) {
9681 if (opcode >= IORING_REGISTER_LAST) {
9686 if (!test_bit(opcode, ctx->restrictions.register_op)) {
9693 case IORING_REGISTER_BUFFERS:
9694 ret = io_sqe_buffers_register(ctx, arg, nr_args);
9696 case IORING_UNREGISTER_BUFFERS:
9700 ret = io_sqe_buffers_unregister(ctx);
9702 case IORING_REGISTER_FILES:
9703 ret = io_sqe_files_register(ctx, arg, nr_args);
9705 case IORING_UNREGISTER_FILES:
9709 ret = io_sqe_files_unregister(ctx);
9711 case IORING_REGISTER_FILES_UPDATE:
9712 ret = io_sqe_files_update(ctx, arg, nr_args);
9714 case IORING_REGISTER_EVENTFD:
9715 case IORING_REGISTER_EVENTFD_ASYNC:
9719 ret = io_eventfd_register(ctx, arg);
9722 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
9723 ctx->eventfd_async = 1;
9725 ctx->eventfd_async = 0;
9727 case IORING_UNREGISTER_EVENTFD:
9731 ret = io_eventfd_unregister(ctx);
9733 case IORING_REGISTER_PROBE:
9735 if (!arg || nr_args > 256)
9737 ret = io_probe(ctx, arg, nr_args);
9739 case IORING_REGISTER_PERSONALITY:
9743 ret = io_register_personality(ctx);
9745 case IORING_UNREGISTER_PERSONALITY:
9749 ret = io_unregister_personality(ctx, nr_args);
9751 case IORING_REGISTER_ENABLE_RINGS:
9755 ret = io_register_enable_rings(ctx);
9757 case IORING_REGISTER_RESTRICTIONS:
9758 ret = io_register_restrictions(ctx, arg, nr_args);
9766 if (io_register_op_must_quiesce(opcode)) {
9767 /* bring the ctx back to life */
9768 percpu_ref_reinit(&ctx->refs);
9770 reinit_completion(&ctx->ref_comp);
9775 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
9776 void __user *, arg, unsigned int, nr_args)
9778 struct io_ring_ctx *ctx;
9787 if (f.file->f_op != &io_uring_fops)
9790 ctx = f.file->private_data;
9794 mutex_lock(&ctx->uring_lock);
9795 ret = __io_uring_register(ctx, opcode, arg, nr_args);
9796 mutex_unlock(&ctx->uring_lock);
9797 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
9798 ctx->cq_ev_fd != NULL, ret);
9804 static int __init io_uring_init(void)
9806 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
9807 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
9808 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
9811 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
9812 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
9813 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
9814 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
9815 BUILD_BUG_SQE_ELEM(1, __u8, flags);
9816 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
9817 BUILD_BUG_SQE_ELEM(4, __s32, fd);
9818 BUILD_BUG_SQE_ELEM(8, __u64, off);
9819 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
9820 BUILD_BUG_SQE_ELEM(16, __u64, addr);
9821 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
9822 BUILD_BUG_SQE_ELEM(24, __u32, len);
9823 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
9824 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
9825 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
9826 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
9827 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
9828 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
9829 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
9830 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
9831 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
9832 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
9833 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
9834 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
9835 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
9836 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
9837 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
9838 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
9839 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
9840 BUILD_BUG_SQE_ELEM(42, __u16, personality);
9841 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
9843 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
9844 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
9845 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
9849 __initcall(io_uring_init);