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/kthread.h>
61 #include <linux/blkdev.h>
62 #include <linux/bvec.h>
63 #include <linux/net.h>
65 #include <net/af_unix.h>
67 #include <linux/anon_inodes.h>
68 #include <linux/sched/mm.h>
69 #include <linux/uaccess.h>
70 #include <linux/nospec.h>
71 #include <linux/sizes.h>
72 #include <linux/hugetlb.h>
73 #include <linux/highmem.h>
74 #include <linux/namei.h>
75 #include <linux/fsnotify.h>
76 #include <linux/fadvise.h>
77 #include <linux/eventpoll.h>
78 #include <linux/fs_struct.h>
79 #include <linux/splice.h>
80 #include <linux/task_work.h>
81 #include <linux/pagemap.h>
82 #include <linux/io_uring.h>
83 #include <linux/blk-cgroup.h>
84 #include <linux/audit.h>
86 #define CREATE_TRACE_POINTS
87 #include <trace/events/io_uring.h>
89 #include <uapi/linux/io_uring.h>
94 #define IORING_MAX_ENTRIES 32768
95 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
98 * Shift of 9 is 512 entries, or exactly one page on 64-bit archs
100 #define IORING_FILE_TABLE_SHIFT 9
101 #define IORING_MAX_FILES_TABLE (1U << IORING_FILE_TABLE_SHIFT)
102 #define IORING_FILE_TABLE_MASK (IORING_MAX_FILES_TABLE - 1)
103 #define IORING_MAX_FIXED_FILES (64 * IORING_MAX_FILES_TABLE)
104 #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
105 IORING_REGISTER_LAST + IORING_OP_LAST)
108 u32 head ____cacheline_aligned_in_smp;
109 u32 tail ____cacheline_aligned_in_smp;
113 * This data is shared with the application through the mmap at offsets
114 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
116 * The offsets to the member fields are published through struct
117 * io_sqring_offsets when calling io_uring_setup.
121 * Head and tail offsets into the ring; the offsets need to be
122 * masked to get valid indices.
124 * The kernel controls head of the sq ring and the tail of the cq ring,
125 * and the application controls tail of the sq ring and the head of the
128 struct io_uring sq, cq;
130 * Bitmasks to apply to head and tail offsets (constant, equals
133 u32 sq_ring_mask, cq_ring_mask;
134 /* Ring sizes (constant, power of 2) */
135 u32 sq_ring_entries, cq_ring_entries;
137 * Number of invalid entries dropped by the kernel due to
138 * invalid index stored in array
140 * Written by the kernel, shouldn't be modified by the
141 * application (i.e. get number of "new events" by comparing to
144 * After a new SQ head value was read by the application this
145 * counter includes all submissions that were dropped reaching
146 * the new SQ head (and possibly more).
152 * Written by the kernel, shouldn't be modified by the
155 * The application needs a full memory barrier before checking
156 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
162 * Written by the application, shouldn't be modified by the
167 * Number of completion events lost because the queue was full;
168 * this should be avoided by the application by making sure
169 * there are not more requests pending than there is space in
170 * the completion queue.
172 * Written by the kernel, shouldn't be modified by the
173 * application (i.e. get number of "new events" by comparing to
176 * As completion events come in out of order this counter is not
177 * ordered with any other data.
181 * Ring buffer of completion events.
183 * The kernel writes completion events fresh every time they are
184 * produced, so the application is allowed to modify pending
187 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
190 enum io_uring_cmd_flags {
191 IO_URING_F_NONBLOCK = 1,
192 IO_URING_F_COMPLETE_DEFER = 2,
195 struct io_mapped_ubuf {
198 struct bio_vec *bvec;
199 unsigned int nr_bvecs;
200 unsigned long acct_pages;
206 struct list_head list;
213 struct fixed_rsrc_table {
217 struct fixed_rsrc_ref_node {
218 struct percpu_ref refs;
219 struct list_head node;
220 struct list_head rsrc_list;
221 struct fixed_rsrc_data *rsrc_data;
222 void (*rsrc_put)(struct io_ring_ctx *ctx,
223 struct io_rsrc_put *prsrc);
224 struct llist_node llist;
228 struct fixed_rsrc_data {
229 struct fixed_rsrc_table *table;
230 struct io_ring_ctx *ctx;
232 struct fixed_rsrc_ref_node *node;
233 struct percpu_ref refs;
234 struct completion done;
238 struct list_head list;
244 struct io_restriction {
245 DECLARE_BITMAP(register_op, IORING_REGISTER_LAST);
246 DECLARE_BITMAP(sqe_op, IORING_OP_LAST);
247 u8 sqe_flags_allowed;
248 u8 sqe_flags_required;
256 /* ctx's that are using this sqd */
257 struct list_head ctx_list;
258 struct list_head ctx_new_list;
259 struct mutex ctx_lock;
261 struct task_struct *thread;
262 struct wait_queue_head wait;
264 unsigned sq_thread_idle;
267 #define IO_IOPOLL_BATCH 8
268 #define IO_COMPL_BATCH 32
269 #define IO_REQ_CACHE_SIZE 32
270 #define IO_REQ_ALLOC_BATCH 8
272 struct io_comp_state {
273 struct io_kiocb *reqs[IO_COMPL_BATCH];
275 unsigned int locked_free_nr;
276 /* inline/task_work completion list, under ->uring_lock */
277 struct list_head free_list;
278 /* IRQ completion list, under ->completion_lock */
279 struct list_head locked_free_list;
282 struct io_submit_state {
283 struct blk_plug plug;
286 * io_kiocb alloc cache
288 void *reqs[IO_REQ_CACHE_SIZE];
289 unsigned int free_reqs;
294 * Batch completion logic
296 struct io_comp_state comp;
299 * File reference cache
303 unsigned int file_refs;
304 unsigned int ios_left;
309 struct percpu_ref refs;
310 } ____cacheline_aligned_in_smp;
314 unsigned int compat: 1;
315 unsigned int limit_mem: 1;
316 unsigned int cq_overflow_flushed: 1;
317 unsigned int drain_next: 1;
318 unsigned int eventfd_async: 1;
319 unsigned int restricted: 1;
320 unsigned int sqo_dead: 1;
323 * Ring buffer of indices into array of io_uring_sqe, which is
324 * mmapped by the application using the IORING_OFF_SQES offset.
326 * This indirection could e.g. be used to assign fixed
327 * io_uring_sqe entries to operations and only submit them to
328 * the queue when needed.
330 * The kernel modifies neither the indices array nor the entries
334 unsigned cached_sq_head;
337 unsigned sq_thread_idle;
338 unsigned cached_sq_dropped;
339 unsigned cached_cq_overflow;
340 unsigned long sq_check_overflow;
342 struct list_head defer_list;
343 struct list_head timeout_list;
344 struct list_head cq_overflow_list;
346 struct io_uring_sqe *sq_sqes;
347 } ____cacheline_aligned_in_smp;
350 struct mutex uring_lock;
351 wait_queue_head_t wait;
352 } ____cacheline_aligned_in_smp;
354 struct io_submit_state submit_state;
356 struct io_rings *rings;
362 * For SQPOLL usage - we hold a reference to the parent task, so we
363 * have access to the ->files
365 struct task_struct *sqo_task;
367 /* Only used for accounting purposes */
368 struct mm_struct *mm_account;
370 #ifdef CONFIG_BLK_CGROUP
371 struct cgroup_subsys_state *sqo_blkcg_css;
374 struct io_sq_data *sq_data; /* if using sq thread polling */
376 struct wait_queue_head sqo_sq_wait;
377 struct list_head sqd_list;
380 * If used, fixed file set. Writers must ensure that ->refs is dead,
381 * readers must ensure that ->refs is alive as long as the file* is
382 * used. Only updated through io_uring_register(2).
384 struct fixed_rsrc_data *file_data;
385 unsigned nr_user_files;
387 /* if used, fixed mapped user buffers */
388 unsigned nr_user_bufs;
389 struct io_mapped_ubuf *user_bufs;
391 struct user_struct *user;
393 const struct cred *creds;
397 unsigned int sessionid;
400 struct completion ref_comp;
401 struct completion sq_thread_comp;
403 #if defined(CONFIG_UNIX)
404 struct socket *ring_sock;
407 struct idr io_buffer_idr;
409 struct idr personality_idr;
412 unsigned cached_cq_tail;
415 atomic_t cq_timeouts;
416 unsigned cq_last_tm_flush;
417 unsigned long cq_check_overflow;
418 struct wait_queue_head cq_wait;
419 struct fasync_struct *cq_fasync;
420 struct eventfd_ctx *cq_ev_fd;
421 } ____cacheline_aligned_in_smp;
424 spinlock_t completion_lock;
427 * ->iopoll_list is protected by the ctx->uring_lock for
428 * io_uring instances that don't use IORING_SETUP_SQPOLL.
429 * For SQPOLL, only the single threaded io_sq_thread() will
430 * manipulate the list, hence no extra locking is needed there.
432 struct list_head iopoll_list;
433 struct hlist_head *cancel_hash;
434 unsigned cancel_hash_bits;
435 bool poll_multi_file;
437 spinlock_t inflight_lock;
438 struct list_head inflight_list;
439 } ____cacheline_aligned_in_smp;
441 struct delayed_work rsrc_put_work;
442 struct llist_head rsrc_put_llist;
443 struct list_head rsrc_ref_list;
444 spinlock_t rsrc_ref_lock;
446 struct io_restriction restrictions;
448 /* Keep this last, we don't need it for the fast path */
449 struct work_struct exit_work;
453 * First field must be the file pointer in all the
454 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
456 struct io_poll_iocb {
458 struct wait_queue_head *head;
462 struct wait_queue_entry wait;
465 struct io_poll_remove {
475 struct io_timeout_data {
476 struct io_kiocb *req;
477 struct hrtimer timer;
478 struct timespec64 ts;
479 enum hrtimer_mode mode;
484 struct sockaddr __user *addr;
485 int __user *addr_len;
487 unsigned long nofile;
507 struct list_head list;
508 /* head of the link, used by linked timeouts only */
509 struct io_kiocb *head;
512 struct io_timeout_rem {
517 struct timespec64 ts;
522 /* NOTE: kiocb has the file as the first member, so don't do it here */
530 struct sockaddr __user *addr;
537 struct user_msghdr __user *umsg;
543 struct io_buffer *kbuf;
549 struct filename *filename;
551 unsigned long nofile;
554 struct io_rsrc_update {
580 struct epoll_event event;
584 struct file *file_out;
585 struct file *file_in;
592 struct io_provide_buf {
606 const char __user *filename;
607 struct statx __user *buffer;
619 struct filename *oldpath;
620 struct filename *newpath;
628 struct filename *filename;
631 struct io_completion {
633 struct list_head list;
637 struct io_async_connect {
638 struct sockaddr_storage address;
641 struct io_async_msghdr {
642 struct iovec fast_iov[UIO_FASTIOV];
643 /* points to an allocated iov, if NULL we use fast_iov instead */
644 struct iovec *free_iov;
645 struct sockaddr __user *uaddr;
647 struct sockaddr_storage addr;
651 struct iovec fast_iov[UIO_FASTIOV];
652 const struct iovec *free_iovec;
653 struct iov_iter iter;
655 struct wait_page_queue wpq;
659 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
660 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
661 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
662 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
663 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
664 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
670 REQ_F_LINK_TIMEOUT_BIT,
672 REQ_F_NEED_CLEANUP_BIT,
674 REQ_F_BUFFER_SELECTED_BIT,
675 REQ_F_NO_FILE_TABLE_BIT,
676 REQ_F_WORK_INITIALIZED_BIT,
677 REQ_F_LTIMEOUT_ACTIVE_BIT,
678 REQ_F_COMPLETE_INLINE_BIT,
680 /* not a real bit, just to check we're not overflowing the space */
686 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
687 /* drain existing IO first */
688 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
690 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
691 /* doesn't sever on completion < 0 */
692 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
694 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
695 /* IOSQE_BUFFER_SELECT */
696 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
698 /* fail rest of links */
699 REQ_F_FAIL_LINK = BIT(REQ_F_FAIL_LINK_BIT),
700 /* on inflight list */
701 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
702 /* read/write uses file position */
703 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
704 /* must not punt to workers */
705 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
706 /* has or had linked timeout */
707 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
709 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
711 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
712 /* already went through poll handler */
713 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
714 /* buffer already selected */
715 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
716 /* doesn't need file table for this request */
717 REQ_F_NO_FILE_TABLE = BIT(REQ_F_NO_FILE_TABLE_BIT),
718 /* io_wq_work is initialized */
719 REQ_F_WORK_INITIALIZED = BIT(REQ_F_WORK_INITIALIZED_BIT),
720 /* linked timeout is active, i.e. prepared by link's head */
721 REQ_F_LTIMEOUT_ACTIVE = BIT(REQ_F_LTIMEOUT_ACTIVE_BIT),
722 /* completion is deferred through io_comp_state */
723 REQ_F_COMPLETE_INLINE = BIT(REQ_F_COMPLETE_INLINE_BIT),
727 struct io_poll_iocb poll;
728 struct io_poll_iocb *double_poll;
731 struct io_task_work {
732 struct io_wq_work_node node;
733 task_work_func_t func;
737 * NOTE! Each of the iocb union members has the file pointer
738 * as the first entry in their struct definition. So you can
739 * access the file pointer through any of the sub-structs,
740 * or directly as just 'ki_filp' in this struct.
746 struct io_poll_iocb poll;
747 struct io_poll_remove poll_remove;
748 struct io_accept accept;
750 struct io_cancel cancel;
751 struct io_timeout timeout;
752 struct io_timeout_rem timeout_rem;
753 struct io_connect connect;
754 struct io_sr_msg sr_msg;
756 struct io_close close;
757 struct io_rsrc_update rsrc_update;
758 struct io_fadvise fadvise;
759 struct io_madvise madvise;
760 struct io_epoll epoll;
761 struct io_splice splice;
762 struct io_provide_buf pbuf;
763 struct io_statx statx;
764 struct io_shutdown shutdown;
765 struct io_rename rename;
766 struct io_unlink unlink;
767 /* use only after cleaning per-op data, see io_clean_op() */
768 struct io_completion compl;
771 /* opcode allocated if it needs to store data for async defer */
774 /* polled IO has completed */
780 struct io_ring_ctx *ctx;
783 struct task_struct *task;
786 struct io_kiocb *link;
787 struct percpu_ref *fixed_rsrc_refs;
790 * 1. used with ctx->iopoll_list with reads/writes
791 * 2. to track reqs with ->files (see io_op_def::file_table)
793 struct list_head inflight_entry;
795 struct io_task_work io_task_work;
796 struct callback_head task_work;
798 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
799 struct hlist_node hash_node;
800 struct async_poll *apoll;
801 struct io_wq_work work;
804 struct io_defer_entry {
805 struct list_head list;
806 struct io_kiocb *req;
811 /* needs req->file assigned */
812 unsigned needs_file : 1;
813 /* hash wq insertion if file is a regular file */
814 unsigned hash_reg_file : 1;
815 /* unbound wq insertion if file is a non-regular file */
816 unsigned unbound_nonreg_file : 1;
817 /* opcode is not supported by this kernel */
818 unsigned not_supported : 1;
819 /* set if opcode supports polled "wait" */
821 unsigned pollout : 1;
822 /* op supports buffer selection */
823 unsigned buffer_select : 1;
824 /* must always have async data allocated */
825 unsigned needs_async_data : 1;
826 /* should block plug */
828 /* size of async data needed, if any */
829 unsigned short async_size;
833 static const struct io_op_def io_op_defs[] = {
834 [IORING_OP_NOP] = {},
835 [IORING_OP_READV] = {
837 .unbound_nonreg_file = 1,
840 .needs_async_data = 1,
842 .async_size = sizeof(struct io_async_rw),
843 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
845 [IORING_OP_WRITEV] = {
848 .unbound_nonreg_file = 1,
850 .needs_async_data = 1,
852 .async_size = sizeof(struct io_async_rw),
853 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG |
856 [IORING_OP_FSYNC] = {
858 .work_flags = IO_WQ_WORK_BLKCG,
860 [IORING_OP_READ_FIXED] = {
862 .unbound_nonreg_file = 1,
865 .async_size = sizeof(struct io_async_rw),
866 .work_flags = IO_WQ_WORK_BLKCG | IO_WQ_WORK_MM,
868 [IORING_OP_WRITE_FIXED] = {
871 .unbound_nonreg_file = 1,
874 .async_size = sizeof(struct io_async_rw),
875 .work_flags = IO_WQ_WORK_BLKCG | IO_WQ_WORK_FSIZE |
878 [IORING_OP_POLL_ADD] = {
880 .unbound_nonreg_file = 1,
882 [IORING_OP_POLL_REMOVE] = {},
883 [IORING_OP_SYNC_FILE_RANGE] = {
885 .work_flags = IO_WQ_WORK_BLKCG,
887 [IORING_OP_SENDMSG] = {
889 .unbound_nonreg_file = 1,
891 .needs_async_data = 1,
892 .async_size = sizeof(struct io_async_msghdr),
893 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
895 [IORING_OP_RECVMSG] = {
897 .unbound_nonreg_file = 1,
900 .needs_async_data = 1,
901 .async_size = sizeof(struct io_async_msghdr),
902 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
904 [IORING_OP_TIMEOUT] = {
905 .needs_async_data = 1,
906 .async_size = sizeof(struct io_timeout_data),
907 .work_flags = IO_WQ_WORK_MM,
909 [IORING_OP_TIMEOUT_REMOVE] = {
910 /* used by timeout updates' prep() */
911 .work_flags = IO_WQ_WORK_MM,
913 [IORING_OP_ACCEPT] = {
915 .unbound_nonreg_file = 1,
917 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_FILES,
919 [IORING_OP_ASYNC_CANCEL] = {},
920 [IORING_OP_LINK_TIMEOUT] = {
921 .needs_async_data = 1,
922 .async_size = sizeof(struct io_timeout_data),
923 .work_flags = IO_WQ_WORK_MM,
925 [IORING_OP_CONNECT] = {
927 .unbound_nonreg_file = 1,
929 .needs_async_data = 1,
930 .async_size = sizeof(struct io_async_connect),
931 .work_flags = IO_WQ_WORK_MM,
933 [IORING_OP_FALLOCATE] = {
935 .work_flags = IO_WQ_WORK_BLKCG | IO_WQ_WORK_FSIZE,
937 [IORING_OP_OPENAT] = {
938 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_BLKCG |
939 IO_WQ_WORK_FS | IO_WQ_WORK_MM,
941 [IORING_OP_CLOSE] = {
942 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_BLKCG,
944 [IORING_OP_FILES_UPDATE] = {
945 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_MM,
947 [IORING_OP_STATX] = {
948 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_MM |
949 IO_WQ_WORK_FS | IO_WQ_WORK_BLKCG,
953 .unbound_nonreg_file = 1,
957 .async_size = sizeof(struct io_async_rw),
958 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
960 [IORING_OP_WRITE] = {
962 .unbound_nonreg_file = 1,
965 .async_size = sizeof(struct io_async_rw),
966 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG |
969 [IORING_OP_FADVISE] = {
971 .work_flags = IO_WQ_WORK_BLKCG,
973 [IORING_OP_MADVISE] = {
974 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
978 .unbound_nonreg_file = 1,
980 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
984 .unbound_nonreg_file = 1,
987 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
989 [IORING_OP_OPENAT2] = {
990 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_FS |
991 IO_WQ_WORK_BLKCG | IO_WQ_WORK_MM,
993 [IORING_OP_EPOLL_CTL] = {
994 .unbound_nonreg_file = 1,
995 .work_flags = IO_WQ_WORK_FILES,
997 [IORING_OP_SPLICE] = {
1000 .unbound_nonreg_file = 1,
1001 .work_flags = IO_WQ_WORK_BLKCG,
1003 [IORING_OP_PROVIDE_BUFFERS] = {},
1004 [IORING_OP_REMOVE_BUFFERS] = {},
1008 .unbound_nonreg_file = 1,
1010 [IORING_OP_SHUTDOWN] = {
1013 [IORING_OP_RENAMEAT] = {
1014 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_FILES |
1015 IO_WQ_WORK_FS | IO_WQ_WORK_BLKCG,
1017 [IORING_OP_UNLINKAT] = {
1018 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_FILES |
1019 IO_WQ_WORK_FS | IO_WQ_WORK_BLKCG,
1023 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
1024 struct task_struct *task,
1025 struct files_struct *files);
1026 static void destroy_fixed_rsrc_ref_node(struct fixed_rsrc_ref_node *ref_node);
1027 static struct fixed_rsrc_ref_node *alloc_fixed_rsrc_ref_node(
1028 struct io_ring_ctx *ctx);
1029 static void init_fixed_file_ref_node(struct io_ring_ctx *ctx,
1030 struct fixed_rsrc_ref_node *ref_node);
1032 static bool io_rw_reissue(struct io_kiocb *req);
1033 static void io_cqring_fill_event(struct io_kiocb *req, long res);
1034 static void io_put_req(struct io_kiocb *req);
1035 static void io_put_req_deferred(struct io_kiocb *req, int nr);
1036 static void io_double_put_req(struct io_kiocb *req);
1037 static void io_dismantle_req(struct io_kiocb *req);
1038 static void io_put_task(struct task_struct *task, int nr);
1039 static void io_queue_next(struct io_kiocb *req);
1040 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
1041 static void __io_queue_linked_timeout(struct io_kiocb *req);
1042 static void io_queue_linked_timeout(struct io_kiocb *req);
1043 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
1044 struct io_uring_rsrc_update *ip,
1046 static void __io_clean_op(struct io_kiocb *req);
1047 static struct file *io_file_get(struct io_submit_state *state,
1048 struct io_kiocb *req, int fd, bool fixed);
1049 static void __io_queue_sqe(struct io_kiocb *req);
1050 static void io_rsrc_put_work(struct work_struct *work);
1052 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
1053 struct iov_iter *iter, bool needs_lock);
1054 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
1055 const struct iovec *fast_iov,
1056 struct iov_iter *iter, bool force);
1057 static void io_req_task_queue(struct io_kiocb *req);
1058 static void io_submit_flush_completions(struct io_comp_state *cs,
1059 struct io_ring_ctx *ctx);
1061 static struct kmem_cache *req_cachep;
1063 static const struct file_operations io_uring_fops;
1065 struct sock *io_uring_get_socket(struct file *file)
1067 #if defined(CONFIG_UNIX)
1068 if (file->f_op == &io_uring_fops) {
1069 struct io_ring_ctx *ctx = file->private_data;
1071 return ctx->ring_sock->sk;
1076 EXPORT_SYMBOL(io_uring_get_socket);
1078 #define io_for_each_link(pos, head) \
1079 for (pos = (head); pos; pos = pos->link)
1081 static inline void io_clean_op(struct io_kiocb *req)
1083 if (req->flags & (REQ_F_NEED_CLEANUP | REQ_F_BUFFER_SELECTED))
1087 static inline void io_set_resource_node(struct io_kiocb *req)
1089 struct io_ring_ctx *ctx = req->ctx;
1091 if (!req->fixed_rsrc_refs) {
1092 req->fixed_rsrc_refs = &ctx->file_data->node->refs;
1093 percpu_ref_get(req->fixed_rsrc_refs);
1097 static bool io_match_task(struct io_kiocb *head,
1098 struct task_struct *task,
1099 struct files_struct *files)
1101 struct io_kiocb *req;
1103 if (task && head->task != task) {
1104 /* in terms of cancelation, always match if req task is dead */
1105 if (head->task->flags & PF_EXITING)
1112 io_for_each_link(req, head) {
1113 if (!(req->flags & REQ_F_WORK_INITIALIZED))
1115 if (req->file && req->file->f_op == &io_uring_fops)
1117 if ((req->work.flags & IO_WQ_WORK_FILES) &&
1118 req->work.identity->files == files)
1124 static void io_sq_thread_drop_mm_files(void)
1126 struct files_struct *files = current->files;
1127 struct mm_struct *mm = current->mm;
1130 kthread_unuse_mm(mm);
1135 struct nsproxy *nsproxy = current->nsproxy;
1138 current->files = NULL;
1139 current->nsproxy = NULL;
1140 task_unlock(current);
1141 put_files_struct(files);
1142 put_nsproxy(nsproxy);
1146 static int __io_sq_thread_acquire_files(struct io_ring_ctx *ctx)
1148 if (current->flags & PF_EXITING)
1151 if (!current->files) {
1152 struct files_struct *files;
1153 struct nsproxy *nsproxy;
1155 task_lock(ctx->sqo_task);
1156 files = ctx->sqo_task->files;
1158 task_unlock(ctx->sqo_task);
1161 atomic_inc(&files->count);
1162 get_nsproxy(ctx->sqo_task->nsproxy);
1163 nsproxy = ctx->sqo_task->nsproxy;
1164 task_unlock(ctx->sqo_task);
1167 current->files = files;
1168 current->nsproxy = nsproxy;
1169 task_unlock(current);
1174 static int __io_sq_thread_acquire_mm(struct io_ring_ctx *ctx)
1176 struct mm_struct *mm;
1178 if (current->flags & PF_EXITING)
1183 /* Should never happen */
1184 if (unlikely(!(ctx->flags & IORING_SETUP_SQPOLL)))
1187 task_lock(ctx->sqo_task);
1188 mm = ctx->sqo_task->mm;
1189 if (unlikely(!mm || !mmget_not_zero(mm)))
1191 task_unlock(ctx->sqo_task);
1201 static int io_sq_thread_acquire_mm_files(struct io_ring_ctx *ctx,
1202 struct io_kiocb *req)
1204 const struct io_op_def *def = &io_op_defs[req->opcode];
1207 if (def->work_flags & IO_WQ_WORK_MM) {
1208 ret = __io_sq_thread_acquire_mm(ctx);
1213 if (def->needs_file || (def->work_flags & IO_WQ_WORK_FILES)) {
1214 ret = __io_sq_thread_acquire_files(ctx);
1222 static void io_sq_thread_associate_blkcg(struct io_ring_ctx *ctx,
1223 struct cgroup_subsys_state **cur_css)
1226 #ifdef CONFIG_BLK_CGROUP
1227 /* puts the old one when swapping */
1228 if (*cur_css != ctx->sqo_blkcg_css) {
1229 kthread_associate_blkcg(ctx->sqo_blkcg_css);
1230 *cur_css = ctx->sqo_blkcg_css;
1235 static void io_sq_thread_unassociate_blkcg(void)
1237 #ifdef CONFIG_BLK_CGROUP
1238 kthread_associate_blkcg(NULL);
1242 static inline void req_set_fail_links(struct io_kiocb *req)
1244 if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
1245 req->flags |= REQ_F_FAIL_LINK;
1249 * None of these are dereferenced, they are simply used to check if any of
1250 * them have changed. If we're under current and check they are still the
1251 * same, we're fine to grab references to them for actual out-of-line use.
1253 static void io_init_identity(struct io_identity *id)
1255 id->files = current->files;
1256 id->mm = current->mm;
1257 #ifdef CONFIG_BLK_CGROUP
1259 id->blkcg_css = blkcg_css();
1262 id->creds = current_cred();
1263 id->nsproxy = current->nsproxy;
1264 id->fs = current->fs;
1265 id->fsize = rlimit(RLIMIT_FSIZE);
1267 id->loginuid = current->loginuid;
1268 id->sessionid = current->sessionid;
1270 refcount_set(&id->count, 1);
1273 static inline void __io_req_init_async(struct io_kiocb *req)
1275 memset(&req->work, 0, sizeof(req->work));
1276 req->flags |= REQ_F_WORK_INITIALIZED;
1280 * Note: must call io_req_init_async() for the first time you
1281 * touch any members of io_wq_work.
1283 static inline void io_req_init_async(struct io_kiocb *req)
1285 struct io_uring_task *tctx = current->io_uring;
1287 if (req->flags & REQ_F_WORK_INITIALIZED)
1290 __io_req_init_async(req);
1292 /* Grab a ref if this isn't our static identity */
1293 req->work.identity = tctx->identity;
1294 if (tctx->identity != &tctx->__identity)
1295 refcount_inc(&req->work.identity->count);
1298 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1300 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1302 complete(&ctx->ref_comp);
1305 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1307 return !req->timeout.off;
1310 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1312 struct io_ring_ctx *ctx;
1315 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1320 * Use 5 bits less than the max cq entries, that should give us around
1321 * 32 entries per hash list if totally full and uniformly spread.
1323 hash_bits = ilog2(p->cq_entries);
1327 ctx->cancel_hash_bits = hash_bits;
1328 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1330 if (!ctx->cancel_hash)
1332 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1334 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1335 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1338 ctx->flags = p->flags;
1339 init_waitqueue_head(&ctx->sqo_sq_wait);
1340 INIT_LIST_HEAD(&ctx->sqd_list);
1341 init_waitqueue_head(&ctx->cq_wait);
1342 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1343 init_completion(&ctx->ref_comp);
1344 init_completion(&ctx->sq_thread_comp);
1345 idr_init(&ctx->io_buffer_idr);
1346 idr_init(&ctx->personality_idr);
1347 mutex_init(&ctx->uring_lock);
1348 init_waitqueue_head(&ctx->wait);
1349 spin_lock_init(&ctx->completion_lock);
1350 INIT_LIST_HEAD(&ctx->iopoll_list);
1351 INIT_LIST_HEAD(&ctx->defer_list);
1352 INIT_LIST_HEAD(&ctx->timeout_list);
1353 spin_lock_init(&ctx->inflight_lock);
1354 INIT_LIST_HEAD(&ctx->inflight_list);
1355 spin_lock_init(&ctx->rsrc_ref_lock);
1356 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
1357 INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
1358 init_llist_head(&ctx->rsrc_put_llist);
1359 INIT_LIST_HEAD(&ctx->submit_state.comp.free_list);
1360 INIT_LIST_HEAD(&ctx->submit_state.comp.locked_free_list);
1363 kfree(ctx->cancel_hash);
1368 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1370 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1371 struct io_ring_ctx *ctx = req->ctx;
1373 return seq != ctx->cached_cq_tail
1374 + READ_ONCE(ctx->cached_cq_overflow);
1380 static void io_put_identity(struct io_uring_task *tctx, struct io_kiocb *req)
1382 if (req->work.identity == &tctx->__identity)
1384 if (refcount_dec_and_test(&req->work.identity->count))
1385 kfree(req->work.identity);
1388 static void io_req_clean_work(struct io_kiocb *req)
1390 if (!(req->flags & REQ_F_WORK_INITIALIZED))
1393 if (req->work.flags & IO_WQ_WORK_MM)
1394 mmdrop(req->work.identity->mm);
1395 #ifdef CONFIG_BLK_CGROUP
1396 if (req->work.flags & IO_WQ_WORK_BLKCG)
1397 css_put(req->work.identity->blkcg_css);
1399 if (req->work.flags & IO_WQ_WORK_CREDS)
1400 put_cred(req->work.identity->creds);
1401 if (req->work.flags & IO_WQ_WORK_FS) {
1402 struct fs_struct *fs = req->work.identity->fs;
1404 spin_lock(&req->work.identity->fs->lock);
1407 spin_unlock(&req->work.identity->fs->lock);
1411 if (req->work.flags & IO_WQ_WORK_FILES) {
1412 put_files_struct(req->work.identity->files);
1413 put_nsproxy(req->work.identity->nsproxy);
1415 if (req->flags & REQ_F_INFLIGHT) {
1416 struct io_ring_ctx *ctx = req->ctx;
1417 struct io_uring_task *tctx = req->task->io_uring;
1418 unsigned long flags;
1420 spin_lock_irqsave(&ctx->inflight_lock, flags);
1421 list_del(&req->inflight_entry);
1422 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
1423 req->flags &= ~REQ_F_INFLIGHT;
1424 if (atomic_read(&tctx->in_idle))
1425 wake_up(&tctx->wait);
1428 req->flags &= ~REQ_F_WORK_INITIALIZED;
1429 req->work.flags &= ~(IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG | IO_WQ_WORK_FS |
1430 IO_WQ_WORK_CREDS | IO_WQ_WORK_FILES);
1431 io_put_identity(req->task->io_uring, req);
1435 * Create a private copy of io_identity, since some fields don't match
1436 * the current context.
1438 static bool io_identity_cow(struct io_kiocb *req)
1440 struct io_uring_task *tctx = current->io_uring;
1441 const struct cred *creds = NULL;
1442 struct io_identity *id;
1444 if (req->work.flags & IO_WQ_WORK_CREDS)
1445 creds = req->work.identity->creds;
1447 id = kmemdup(req->work.identity, sizeof(*id), GFP_KERNEL);
1448 if (unlikely(!id)) {
1449 req->work.flags |= IO_WQ_WORK_CANCEL;
1454 * We can safely just re-init the creds we copied Either the field
1455 * matches the current one, or we haven't grabbed it yet. The only
1456 * exception is ->creds, through registered personalities, so handle
1457 * that one separately.
1459 io_init_identity(id);
1463 /* add one for this request */
1464 refcount_inc(&id->count);
1466 /* drop tctx and req identity references, if needed */
1467 if (tctx->identity != &tctx->__identity &&
1468 refcount_dec_and_test(&tctx->identity->count))
1469 kfree(tctx->identity);
1470 if (req->work.identity != &tctx->__identity &&
1471 refcount_dec_and_test(&req->work.identity->count))
1472 kfree(req->work.identity);
1474 req->work.identity = id;
1475 tctx->identity = id;
1479 static void io_req_track_inflight(struct io_kiocb *req)
1481 struct io_ring_ctx *ctx = req->ctx;
1483 if (!(req->flags & REQ_F_INFLIGHT)) {
1484 io_req_init_async(req);
1485 req->flags |= REQ_F_INFLIGHT;
1487 spin_lock_irq(&ctx->inflight_lock);
1488 list_add(&req->inflight_entry, &ctx->inflight_list);
1489 spin_unlock_irq(&ctx->inflight_lock);
1493 static bool io_grab_identity(struct io_kiocb *req)
1495 const struct io_op_def *def = &io_op_defs[req->opcode];
1496 struct io_identity *id = req->work.identity;
1498 if (def->work_flags & IO_WQ_WORK_FSIZE) {
1499 if (id->fsize != rlimit(RLIMIT_FSIZE))
1501 req->work.flags |= IO_WQ_WORK_FSIZE;
1503 #ifdef CONFIG_BLK_CGROUP
1504 if (!(req->work.flags & IO_WQ_WORK_BLKCG) &&
1505 (def->work_flags & IO_WQ_WORK_BLKCG)) {
1507 if (id->blkcg_css != blkcg_css()) {
1512 * This should be rare, either the cgroup is dying or the task
1513 * is moving cgroups. Just punt to root for the handful of ios.
1515 if (css_tryget_online(id->blkcg_css))
1516 req->work.flags |= IO_WQ_WORK_BLKCG;
1520 if (!(req->work.flags & IO_WQ_WORK_CREDS)) {
1521 if (id->creds != current_cred())
1523 get_cred(id->creds);
1524 req->work.flags |= IO_WQ_WORK_CREDS;
1527 if (!uid_eq(current->loginuid, id->loginuid) ||
1528 current->sessionid != id->sessionid)
1531 if (!(req->work.flags & IO_WQ_WORK_FS) &&
1532 (def->work_flags & IO_WQ_WORK_FS)) {
1533 if (current->fs != id->fs)
1535 spin_lock(&id->fs->lock);
1536 if (!id->fs->in_exec) {
1538 req->work.flags |= IO_WQ_WORK_FS;
1540 req->work.flags |= IO_WQ_WORK_CANCEL;
1542 spin_unlock(¤t->fs->lock);
1544 if (!(req->work.flags & IO_WQ_WORK_FILES) &&
1545 (def->work_flags & IO_WQ_WORK_FILES) &&
1546 !(req->flags & REQ_F_NO_FILE_TABLE)) {
1547 if (id->files != current->files ||
1548 id->nsproxy != current->nsproxy)
1550 atomic_inc(&id->files->count);
1551 get_nsproxy(id->nsproxy);
1552 req->work.flags |= IO_WQ_WORK_FILES;
1553 io_req_track_inflight(req);
1555 if (!(req->work.flags & IO_WQ_WORK_MM) &&
1556 (def->work_flags & IO_WQ_WORK_MM)) {
1557 if (id->mm != current->mm)
1560 req->work.flags |= IO_WQ_WORK_MM;
1566 static void io_prep_async_work(struct io_kiocb *req)
1568 const struct io_op_def *def = &io_op_defs[req->opcode];
1569 struct io_ring_ctx *ctx = req->ctx;
1571 io_req_init_async(req);
1573 if (req->flags & REQ_F_FORCE_ASYNC)
1574 req->work.flags |= IO_WQ_WORK_CONCURRENT;
1576 if (req->flags & REQ_F_ISREG) {
1577 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1578 io_wq_hash_work(&req->work, file_inode(req->file));
1580 if (def->unbound_nonreg_file)
1581 req->work.flags |= IO_WQ_WORK_UNBOUND;
1584 /* if we fail grabbing identity, we must COW, regrab, and retry */
1585 if (io_grab_identity(req))
1588 if (!io_identity_cow(req))
1591 /* can't fail at this point */
1592 if (!io_grab_identity(req))
1596 static void io_prep_async_link(struct io_kiocb *req)
1598 struct io_kiocb *cur;
1600 io_for_each_link(cur, req)
1601 io_prep_async_work(cur);
1604 static struct io_kiocb *__io_queue_async_work(struct io_kiocb *req)
1606 struct io_ring_ctx *ctx = req->ctx;
1607 struct io_kiocb *link = io_prep_linked_timeout(req);
1609 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1610 &req->work, req->flags);
1611 io_wq_enqueue(ctx->io_wq, &req->work);
1615 static void io_queue_async_work(struct io_kiocb *req)
1617 struct io_kiocb *link;
1619 /* init ->work of the whole link before punting */
1620 io_prep_async_link(req);
1621 link = __io_queue_async_work(req);
1624 io_queue_linked_timeout(link);
1627 static void io_kill_timeout(struct io_kiocb *req)
1629 struct io_timeout_data *io = req->async_data;
1632 ret = hrtimer_try_to_cancel(&io->timer);
1634 atomic_set(&req->ctx->cq_timeouts,
1635 atomic_read(&req->ctx->cq_timeouts) + 1);
1636 list_del_init(&req->timeout.list);
1637 io_cqring_fill_event(req, 0);
1638 io_put_req_deferred(req, 1);
1643 * Returns true if we found and killed one or more timeouts
1645 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
1646 struct files_struct *files)
1648 struct io_kiocb *req, *tmp;
1651 spin_lock_irq(&ctx->completion_lock);
1652 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
1653 if (io_match_task(req, tsk, files)) {
1654 io_kill_timeout(req);
1658 spin_unlock_irq(&ctx->completion_lock);
1659 return canceled != 0;
1662 static void __io_queue_deferred(struct io_ring_ctx *ctx)
1665 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1666 struct io_defer_entry, list);
1668 if (req_need_defer(de->req, de->seq))
1670 list_del_init(&de->list);
1671 io_req_task_queue(de->req);
1673 } while (!list_empty(&ctx->defer_list));
1676 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1680 if (list_empty(&ctx->timeout_list))
1683 seq = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
1686 u32 events_needed, events_got;
1687 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1688 struct io_kiocb, timeout.list);
1690 if (io_is_timeout_noseq(req))
1694 * Since seq can easily wrap around over time, subtract
1695 * the last seq at which timeouts were flushed before comparing.
1696 * Assuming not more than 2^31-1 events have happened since,
1697 * these subtractions won't have wrapped, so we can check if
1698 * target is in [last_seq, current_seq] by comparing the two.
1700 events_needed = req->timeout.target_seq - ctx->cq_last_tm_flush;
1701 events_got = seq - ctx->cq_last_tm_flush;
1702 if (events_got < events_needed)
1705 list_del_init(&req->timeout.list);
1706 io_kill_timeout(req);
1707 } while (!list_empty(&ctx->timeout_list));
1709 ctx->cq_last_tm_flush = seq;
1712 static void io_commit_cqring(struct io_ring_ctx *ctx)
1714 io_flush_timeouts(ctx);
1716 /* order cqe stores with ring update */
1717 smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
1719 if (unlikely(!list_empty(&ctx->defer_list)))
1720 __io_queue_deferred(ctx);
1723 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1725 struct io_rings *r = ctx->rings;
1727 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == r->sq_ring_entries;
1730 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
1732 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
1735 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1737 struct io_rings *rings = ctx->rings;
1741 * writes to the cq entry need to come after reading head; the
1742 * control dependency is enough as we're using WRITE_ONCE to
1745 if (__io_cqring_events(ctx) == rings->cq_ring_entries)
1748 tail = ctx->cached_cq_tail++;
1749 return &rings->cqes[tail & ctx->cq_mask];
1752 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1756 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1758 if (!ctx->eventfd_async)
1760 return io_wq_current_is_worker();
1763 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1765 /* see waitqueue_active() comment */
1768 if (waitqueue_active(&ctx->wait))
1769 wake_up(&ctx->wait);
1770 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1771 wake_up(&ctx->sq_data->wait);
1772 if (io_should_trigger_evfd(ctx))
1773 eventfd_signal(ctx->cq_ev_fd, 1);
1774 if (waitqueue_active(&ctx->cq_wait)) {
1775 wake_up_interruptible(&ctx->cq_wait);
1776 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1780 static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx)
1782 /* see waitqueue_active() comment */
1785 if (ctx->flags & IORING_SETUP_SQPOLL) {
1786 if (waitqueue_active(&ctx->wait))
1787 wake_up(&ctx->wait);
1789 if (io_should_trigger_evfd(ctx))
1790 eventfd_signal(ctx->cq_ev_fd, 1);
1791 if (waitqueue_active(&ctx->cq_wait)) {
1792 wake_up_interruptible(&ctx->cq_wait);
1793 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1797 /* Returns true if there are no backlogged entries after the flush */
1798 static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force,
1799 struct task_struct *tsk,
1800 struct files_struct *files)
1802 struct io_rings *rings = ctx->rings;
1803 struct io_kiocb *req, *tmp;
1804 struct io_uring_cqe *cqe;
1805 unsigned long flags;
1806 bool all_flushed, posted;
1809 if (!force && __io_cqring_events(ctx) == rings->cq_ring_entries)
1813 spin_lock_irqsave(&ctx->completion_lock, flags);
1814 list_for_each_entry_safe(req, tmp, &ctx->cq_overflow_list, compl.list) {
1815 if (!io_match_task(req, tsk, files))
1818 cqe = io_get_cqring(ctx);
1822 list_move(&req->compl.list, &list);
1824 WRITE_ONCE(cqe->user_data, req->user_data);
1825 WRITE_ONCE(cqe->res, req->result);
1826 WRITE_ONCE(cqe->flags, req->compl.cflags);
1828 ctx->cached_cq_overflow++;
1829 WRITE_ONCE(ctx->rings->cq_overflow,
1830 ctx->cached_cq_overflow);
1835 all_flushed = list_empty(&ctx->cq_overflow_list);
1837 clear_bit(0, &ctx->sq_check_overflow);
1838 clear_bit(0, &ctx->cq_check_overflow);
1839 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
1843 io_commit_cqring(ctx);
1844 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1846 io_cqring_ev_posted(ctx);
1848 while (!list_empty(&list)) {
1849 req = list_first_entry(&list, struct io_kiocb, compl.list);
1850 list_del(&req->compl.list);
1857 static void io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force,
1858 struct task_struct *tsk,
1859 struct files_struct *files)
1861 if (test_bit(0, &ctx->cq_check_overflow)) {
1862 /* iopoll syncs against uring_lock, not completion_lock */
1863 if (ctx->flags & IORING_SETUP_IOPOLL)
1864 mutex_lock(&ctx->uring_lock);
1865 __io_cqring_overflow_flush(ctx, force, tsk, files);
1866 if (ctx->flags & IORING_SETUP_IOPOLL)
1867 mutex_unlock(&ctx->uring_lock);
1871 static void __io_cqring_fill_event(struct io_kiocb *req, long res, long cflags)
1873 struct io_ring_ctx *ctx = req->ctx;
1874 struct io_uring_cqe *cqe;
1876 trace_io_uring_complete(ctx, req->user_data, res);
1879 * If we can't get a cq entry, userspace overflowed the
1880 * submission (by quite a lot). Increment the overflow count in
1883 cqe = io_get_cqring(ctx);
1885 WRITE_ONCE(cqe->user_data, req->user_data);
1886 WRITE_ONCE(cqe->res, res);
1887 WRITE_ONCE(cqe->flags, cflags);
1888 } else if (ctx->cq_overflow_flushed ||
1889 atomic_read(&req->task->io_uring->in_idle)) {
1891 * If we're in ring overflow flush mode, or in task cancel mode,
1892 * then we cannot store the request for later flushing, we need
1893 * to drop it on the floor.
1895 ctx->cached_cq_overflow++;
1896 WRITE_ONCE(ctx->rings->cq_overflow, ctx->cached_cq_overflow);
1898 if (list_empty(&ctx->cq_overflow_list)) {
1899 set_bit(0, &ctx->sq_check_overflow);
1900 set_bit(0, &ctx->cq_check_overflow);
1901 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
1905 req->compl.cflags = cflags;
1906 refcount_inc(&req->refs);
1907 list_add_tail(&req->compl.list, &ctx->cq_overflow_list);
1911 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1913 __io_cqring_fill_event(req, res, 0);
1916 static inline void io_req_complete_post(struct io_kiocb *req, long res,
1917 unsigned int cflags)
1919 struct io_ring_ctx *ctx = req->ctx;
1920 unsigned long flags;
1922 spin_lock_irqsave(&ctx->completion_lock, flags);
1923 __io_cqring_fill_event(req, res, cflags);
1924 io_commit_cqring(ctx);
1926 * If we're the last reference to this request, add to our locked
1929 if (refcount_dec_and_test(&req->refs)) {
1930 struct io_comp_state *cs = &ctx->submit_state.comp;
1932 io_dismantle_req(req);
1933 io_put_task(req->task, 1);
1934 list_add(&req->compl.list, &cs->locked_free_list);
1935 cs->locked_free_nr++;
1938 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1940 io_cqring_ev_posted(ctx);
1943 percpu_ref_put(&ctx->refs);
1947 static void io_req_complete_state(struct io_kiocb *req, long res,
1948 unsigned int cflags)
1952 req->compl.cflags = cflags;
1953 req->flags |= REQ_F_COMPLETE_INLINE;
1956 static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags,
1957 long res, unsigned cflags)
1959 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1960 io_req_complete_state(req, res, cflags);
1962 io_req_complete_post(req, res, cflags);
1965 static inline void io_req_complete(struct io_kiocb *req, long res)
1967 __io_req_complete(req, 0, res, 0);
1970 static bool io_flush_cached_reqs(struct io_ring_ctx *ctx)
1972 struct io_submit_state *state = &ctx->submit_state;
1973 struct io_comp_state *cs = &state->comp;
1974 struct io_kiocb *req = NULL;
1977 * If we have more than a batch's worth of requests in our IRQ side
1978 * locked cache, grab the lock and move them over to our submission
1981 if (READ_ONCE(cs->locked_free_nr) > IO_COMPL_BATCH) {
1982 spin_lock_irq(&ctx->completion_lock);
1983 list_splice_init(&cs->locked_free_list, &cs->free_list);
1984 cs->locked_free_nr = 0;
1985 spin_unlock_irq(&ctx->completion_lock);
1988 while (!list_empty(&cs->free_list)) {
1989 req = list_first_entry(&cs->free_list, struct io_kiocb,
1991 list_del(&req->compl.list);
1992 state->reqs[state->free_reqs++] = req;
1993 if (state->free_reqs == ARRAY_SIZE(state->reqs))
2000 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
2002 struct io_submit_state *state = &ctx->submit_state;
2004 BUILD_BUG_ON(IO_REQ_ALLOC_BATCH > ARRAY_SIZE(state->reqs));
2006 if (!state->free_reqs) {
2007 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
2010 if (io_flush_cached_reqs(ctx))
2013 ret = kmem_cache_alloc_bulk(req_cachep, gfp, IO_REQ_ALLOC_BATCH,
2017 * Bulk alloc is all-or-nothing. If we fail to get a batch,
2018 * retry single alloc to be on the safe side.
2020 if (unlikely(ret <= 0)) {
2021 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
2022 if (!state->reqs[0])
2026 state->free_reqs = ret;
2030 return state->reqs[state->free_reqs];
2033 static inline void io_put_file(struct io_kiocb *req, struct file *file,
2040 static void io_dismantle_req(struct io_kiocb *req)
2044 if (req->async_data)
2045 kfree(req->async_data);
2047 io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
2048 if (req->fixed_rsrc_refs)
2049 percpu_ref_put(req->fixed_rsrc_refs);
2050 io_req_clean_work(req);
2053 static inline void io_put_task(struct task_struct *task, int nr)
2055 struct io_uring_task *tctx = task->io_uring;
2057 percpu_counter_sub(&tctx->inflight, nr);
2058 if (unlikely(atomic_read(&tctx->in_idle)))
2059 wake_up(&tctx->wait);
2060 put_task_struct_many(task, nr);
2063 static void __io_free_req(struct io_kiocb *req)
2065 struct io_ring_ctx *ctx = req->ctx;
2067 io_dismantle_req(req);
2068 io_put_task(req->task, 1);
2070 kmem_cache_free(req_cachep, req);
2071 percpu_ref_put(&ctx->refs);
2074 static inline void io_remove_next_linked(struct io_kiocb *req)
2076 struct io_kiocb *nxt = req->link;
2078 req->link = nxt->link;
2082 static void io_kill_linked_timeout(struct io_kiocb *req)
2084 struct io_ring_ctx *ctx = req->ctx;
2085 struct io_kiocb *link;
2086 bool cancelled = false;
2087 unsigned long flags;
2089 spin_lock_irqsave(&ctx->completion_lock, flags);
2093 * Can happen if a linked timeout fired and link had been like
2094 * req -> link t-out -> link t-out [-> ...]
2096 if (link && (link->flags & REQ_F_LTIMEOUT_ACTIVE)) {
2097 struct io_timeout_data *io = link->async_data;
2100 io_remove_next_linked(req);
2101 link->timeout.head = NULL;
2102 ret = hrtimer_try_to_cancel(&io->timer);
2104 io_cqring_fill_event(link, -ECANCELED);
2105 io_commit_cqring(ctx);
2109 req->flags &= ~REQ_F_LINK_TIMEOUT;
2110 spin_unlock_irqrestore(&ctx->completion_lock, flags);
2113 io_cqring_ev_posted(ctx);
2119 static void io_fail_links(struct io_kiocb *req)
2121 struct io_kiocb *link, *nxt;
2122 struct io_ring_ctx *ctx = req->ctx;
2123 unsigned long flags;
2125 spin_lock_irqsave(&ctx->completion_lock, flags);
2133 trace_io_uring_fail_link(req, link);
2134 io_cqring_fill_event(link, -ECANCELED);
2137 * It's ok to free under spinlock as they're not linked anymore,
2138 * but avoid REQ_F_WORK_INITIALIZED because it may deadlock on
2141 if (link->flags & REQ_F_WORK_INITIALIZED)
2142 io_put_req_deferred(link, 2);
2144 io_double_put_req(link);
2147 io_commit_cqring(ctx);
2148 spin_unlock_irqrestore(&ctx->completion_lock, flags);
2150 io_cqring_ev_posted(ctx);
2153 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
2155 if (req->flags & REQ_F_LINK_TIMEOUT)
2156 io_kill_linked_timeout(req);
2159 * If LINK is set, we have dependent requests in this chain. If we
2160 * didn't fail this request, queue the first one up, moving any other
2161 * dependencies to the next request. In case of failure, fail the rest
2164 if (likely(!(req->flags & REQ_F_FAIL_LINK))) {
2165 struct io_kiocb *nxt = req->link;
2174 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
2176 if (likely(!(req->link) && !(req->flags & REQ_F_LINK_TIMEOUT)))
2178 return __io_req_find_next(req);
2181 static bool __tctx_task_work(struct io_uring_task *tctx)
2183 struct io_ring_ctx *ctx = NULL;
2184 struct io_wq_work_list list;
2185 struct io_wq_work_node *node;
2187 if (wq_list_empty(&tctx->task_list))
2190 spin_lock(&tctx->task_lock);
2191 list = tctx->task_list;
2192 INIT_WQ_LIST(&tctx->task_list);
2193 spin_unlock(&tctx->task_lock);
2197 struct io_wq_work_node *next = node->next;
2198 struct io_ring_ctx *this_ctx;
2199 struct io_kiocb *req;
2201 req = container_of(node, struct io_kiocb, io_task_work.node);
2202 this_ctx = req->ctx;
2203 req->task_work.func(&req->task_work);
2208 } else if (ctx != this_ctx) {
2209 mutex_lock(&ctx->uring_lock);
2210 io_submit_flush_completions(&ctx->submit_state.comp, ctx);
2211 mutex_unlock(&ctx->uring_lock);
2216 if (ctx && ctx->submit_state.comp.nr) {
2217 mutex_lock(&ctx->uring_lock);
2218 io_submit_flush_completions(&ctx->submit_state.comp, ctx);
2219 mutex_unlock(&ctx->uring_lock);
2222 return list.first != NULL;
2225 static void tctx_task_work(struct callback_head *cb)
2227 struct io_uring_task *tctx = container_of(cb, struct io_uring_task, task_work);
2229 while (__tctx_task_work(tctx))
2232 clear_bit(0, &tctx->task_state);
2235 static int io_task_work_add(struct task_struct *tsk, struct io_kiocb *req,
2236 enum task_work_notify_mode notify)
2238 struct io_uring_task *tctx = tsk->io_uring;
2239 struct io_wq_work_node *node, *prev;
2242 WARN_ON_ONCE(!tctx);
2244 spin_lock(&tctx->task_lock);
2245 wq_list_add_tail(&req->io_task_work.node, &tctx->task_list);
2246 spin_unlock(&tctx->task_lock);
2248 /* task_work already pending, we're done */
2249 if (test_bit(0, &tctx->task_state) ||
2250 test_and_set_bit(0, &tctx->task_state))
2253 if (!task_work_add(tsk, &tctx->task_work, notify))
2257 * Slow path - we failed, find and delete work. if the work is not
2258 * in the list, it got run and we're fine.
2261 spin_lock(&tctx->task_lock);
2262 wq_list_for_each(node, prev, &tctx->task_list) {
2263 if (&req->io_task_work.node == node) {
2264 wq_list_del(&tctx->task_list, node, prev);
2269 spin_unlock(&tctx->task_lock);
2270 clear_bit(0, &tctx->task_state);
2274 static int io_req_task_work_add(struct io_kiocb *req)
2276 struct task_struct *tsk = req->task;
2277 struct io_ring_ctx *ctx = req->ctx;
2278 enum task_work_notify_mode notify;
2281 if (tsk->flags & PF_EXITING)
2285 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
2286 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
2287 * processing task_work. There's no reliable way to tell if TWA_RESUME
2291 if (!(ctx->flags & IORING_SETUP_SQPOLL))
2292 notify = TWA_SIGNAL;
2294 ret = io_task_work_add(tsk, req, notify);
2296 wake_up_process(tsk);
2301 static void io_req_task_work_add_fallback(struct io_kiocb *req,
2302 task_work_func_t cb)
2304 struct task_struct *tsk = io_wq_get_task(req->ctx->io_wq);
2306 init_task_work(&req->task_work, cb);
2307 task_work_add(tsk, &req->task_work, TWA_NONE);
2308 wake_up_process(tsk);
2311 static void __io_req_task_cancel(struct io_kiocb *req, int error)
2313 struct io_ring_ctx *ctx = req->ctx;
2315 spin_lock_irq(&ctx->completion_lock);
2316 io_cqring_fill_event(req, error);
2317 io_commit_cqring(ctx);
2318 spin_unlock_irq(&ctx->completion_lock);
2320 io_cqring_ev_posted(ctx);
2321 req_set_fail_links(req);
2322 io_double_put_req(req);
2325 static void io_req_task_cancel(struct callback_head *cb)
2327 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2328 struct io_ring_ctx *ctx = req->ctx;
2330 __io_req_task_cancel(req, -ECANCELED);
2331 percpu_ref_put(&ctx->refs);
2334 static void __io_req_task_submit(struct io_kiocb *req)
2336 struct io_ring_ctx *ctx = req->ctx;
2338 mutex_lock(&ctx->uring_lock);
2339 if (!ctx->sqo_dead &&
2340 !__io_sq_thread_acquire_mm(ctx) &&
2341 !__io_sq_thread_acquire_files(ctx))
2342 __io_queue_sqe(req);
2344 __io_req_task_cancel(req, -EFAULT);
2345 mutex_unlock(&ctx->uring_lock);
2348 static void io_req_task_submit(struct callback_head *cb)
2350 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2351 struct io_ring_ctx *ctx = req->ctx;
2353 __io_req_task_submit(req);
2354 percpu_ref_put(&ctx->refs);
2357 static void io_req_task_queue(struct io_kiocb *req)
2361 req->task_work.func = io_req_task_submit;
2362 percpu_ref_get(&req->ctx->refs);
2364 ret = io_req_task_work_add(req);
2366 io_req_task_work_add_fallback(req, io_req_task_cancel);
2369 static inline void io_queue_next(struct io_kiocb *req)
2371 struct io_kiocb *nxt = io_req_find_next(req);
2374 io_req_task_queue(nxt);
2377 static void io_free_req(struct io_kiocb *req)
2384 struct task_struct *task;
2389 static inline void io_init_req_batch(struct req_batch *rb)
2396 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2397 struct req_batch *rb)
2400 io_put_task(rb->task, rb->task_refs);
2402 percpu_ref_put_many(&ctx->refs, rb->ctx_refs);
2405 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req,
2406 struct io_submit_state *state)
2410 if (req->task != rb->task) {
2412 io_put_task(rb->task, rb->task_refs);
2413 rb->task = req->task;
2419 io_dismantle_req(req);
2420 if (state->free_reqs != ARRAY_SIZE(state->reqs)) {
2421 state->reqs[state->free_reqs++] = req;
2423 struct io_comp_state *cs = &req->ctx->submit_state.comp;
2425 list_add(&req->compl.list, &cs->free_list);
2429 static void io_submit_flush_completions(struct io_comp_state *cs,
2430 struct io_ring_ctx *ctx)
2433 struct io_kiocb *req;
2434 struct req_batch rb;
2436 io_init_req_batch(&rb);
2437 spin_lock_irq(&ctx->completion_lock);
2438 for (i = 0; i < nr; i++) {
2440 __io_cqring_fill_event(req, req->result, req->compl.cflags);
2442 io_commit_cqring(ctx);
2443 spin_unlock_irq(&ctx->completion_lock);
2445 io_cqring_ev_posted(ctx);
2446 for (i = 0; i < nr; i++) {
2449 /* submission and completion refs */
2450 if (refcount_sub_and_test(2, &req->refs))
2451 io_req_free_batch(&rb, req, &ctx->submit_state);
2454 io_req_free_batch_finish(ctx, &rb);
2459 * Drop reference to request, return next in chain (if there is one) if this
2460 * was the last reference to this request.
2462 static struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2464 struct io_kiocb *nxt = NULL;
2466 if (refcount_dec_and_test(&req->refs)) {
2467 nxt = io_req_find_next(req);
2473 static void io_put_req(struct io_kiocb *req)
2475 if (refcount_dec_and_test(&req->refs))
2479 static void io_put_req_deferred_cb(struct callback_head *cb)
2481 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2486 static void io_free_req_deferred(struct io_kiocb *req)
2490 req->task_work.func = io_put_req_deferred_cb;
2491 ret = io_req_task_work_add(req);
2493 io_req_task_work_add_fallback(req, io_put_req_deferred_cb);
2496 static inline void io_put_req_deferred(struct io_kiocb *req, int refs)
2498 if (refcount_sub_and_test(refs, &req->refs))
2499 io_free_req_deferred(req);
2502 static void io_double_put_req(struct io_kiocb *req)
2504 /* drop both submit and complete references */
2505 if (refcount_sub_and_test(2, &req->refs))
2509 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2511 /* See comment at the top of this file */
2513 return __io_cqring_events(ctx);
2516 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2518 struct io_rings *rings = ctx->rings;
2520 /* make sure SQ entry isn't read before tail */
2521 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2524 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2526 unsigned int cflags;
2528 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2529 cflags |= IORING_CQE_F_BUFFER;
2530 req->flags &= ~REQ_F_BUFFER_SELECTED;
2535 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2537 struct io_buffer *kbuf;
2539 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2540 return io_put_kbuf(req, kbuf);
2543 static inline bool io_run_task_work(void)
2546 * Not safe to run on exiting task, and the task_work handling will
2547 * not add work to such a task.
2549 if (unlikely(current->flags & PF_EXITING))
2551 if (current->task_works) {
2552 __set_current_state(TASK_RUNNING);
2561 * Find and free completed poll iocbs
2563 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2564 struct list_head *done)
2566 struct req_batch rb;
2567 struct io_kiocb *req;
2569 /* order with ->result store in io_complete_rw_iopoll() */
2572 io_init_req_batch(&rb);
2573 while (!list_empty(done)) {
2576 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2577 list_del(&req->inflight_entry);
2579 if (READ_ONCE(req->result) == -EAGAIN) {
2580 req->iopoll_completed = 0;
2581 if (io_rw_reissue(req))
2585 if (req->flags & REQ_F_BUFFER_SELECTED)
2586 cflags = io_put_rw_kbuf(req);
2588 __io_cqring_fill_event(req, req->result, cflags);
2591 if (refcount_dec_and_test(&req->refs))
2592 io_req_free_batch(&rb, req, &ctx->submit_state);
2595 io_commit_cqring(ctx);
2596 io_cqring_ev_posted_iopoll(ctx);
2597 io_req_free_batch_finish(ctx, &rb);
2600 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2603 struct io_kiocb *req, *tmp;
2609 * Only spin for completions if we don't have multiple devices hanging
2610 * off our complete list, and we're under the requested amount.
2612 spin = !ctx->poll_multi_file && *nr_events < min;
2615 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2616 struct kiocb *kiocb = &req->rw.kiocb;
2619 * Move completed and retryable entries to our local lists.
2620 * If we find a request that requires polling, break out
2621 * and complete those lists first, if we have entries there.
2623 if (READ_ONCE(req->iopoll_completed)) {
2624 list_move_tail(&req->inflight_entry, &done);
2627 if (!list_empty(&done))
2630 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2634 /* iopoll may have completed current req */
2635 if (READ_ONCE(req->iopoll_completed))
2636 list_move_tail(&req->inflight_entry, &done);
2643 if (!list_empty(&done))
2644 io_iopoll_complete(ctx, nr_events, &done);
2650 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2651 * non-spinning poll check - we'll still enter the driver poll loop, but only
2652 * as a non-spinning completion check.
2654 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
2657 while (!list_empty(&ctx->iopoll_list) && !need_resched()) {
2660 ret = io_do_iopoll(ctx, nr_events, min);
2663 if (*nr_events >= min)
2671 * We can't just wait for polled events to come to us, we have to actively
2672 * find and complete them.
2674 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2676 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2679 mutex_lock(&ctx->uring_lock);
2680 while (!list_empty(&ctx->iopoll_list)) {
2681 unsigned int nr_events = 0;
2683 io_do_iopoll(ctx, &nr_events, 0);
2685 /* let it sleep and repeat later if can't complete a request */
2689 * Ensure we allow local-to-the-cpu processing to take place,
2690 * in this case we need to ensure that we reap all events.
2691 * Also let task_work, etc. to progress by releasing the mutex
2693 if (need_resched()) {
2694 mutex_unlock(&ctx->uring_lock);
2696 mutex_lock(&ctx->uring_lock);
2699 mutex_unlock(&ctx->uring_lock);
2702 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2704 unsigned int nr_events = 0;
2705 int iters = 0, ret = 0;
2708 * We disallow the app entering submit/complete with polling, but we
2709 * still need to lock the ring to prevent racing with polled issue
2710 * that got punted to a workqueue.
2712 mutex_lock(&ctx->uring_lock);
2715 * Don't enter poll loop if we already have events pending.
2716 * If we do, we can potentially be spinning for commands that
2717 * already triggered a CQE (eg in error).
2719 if (test_bit(0, &ctx->cq_check_overflow))
2720 __io_cqring_overflow_flush(ctx, false, NULL, NULL);
2721 if (io_cqring_events(ctx))
2725 * If a submit got punted to a workqueue, we can have the
2726 * application entering polling for a command before it gets
2727 * issued. That app will hold the uring_lock for the duration
2728 * of the poll right here, so we need to take a breather every
2729 * now and then to ensure that the issue has a chance to add
2730 * the poll to the issued list. Otherwise we can spin here
2731 * forever, while the workqueue is stuck trying to acquire the
2734 if (!(++iters & 7)) {
2735 mutex_unlock(&ctx->uring_lock);
2737 mutex_lock(&ctx->uring_lock);
2740 ret = io_iopoll_getevents(ctx, &nr_events, min);
2744 } while (min && !nr_events && !need_resched());
2746 mutex_unlock(&ctx->uring_lock);
2750 static void kiocb_end_write(struct io_kiocb *req)
2753 * Tell lockdep we inherited freeze protection from submission
2756 if (req->flags & REQ_F_ISREG) {
2757 struct inode *inode = file_inode(req->file);
2759 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
2761 file_end_write(req->file);
2765 static bool io_resubmit_prep(struct io_kiocb *req)
2767 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2769 struct iov_iter iter;
2771 /* already prepared */
2772 if (req->async_data)
2775 switch (req->opcode) {
2776 case IORING_OP_READV:
2777 case IORING_OP_READ_FIXED:
2778 case IORING_OP_READ:
2781 case IORING_OP_WRITEV:
2782 case IORING_OP_WRITE_FIXED:
2783 case IORING_OP_WRITE:
2787 printk_once(KERN_WARNING "io_uring: bad opcode in resubmit %d\n",
2792 ret = io_import_iovec(rw, req, &iovec, &iter, false);
2795 return !io_setup_async_rw(req, iovec, inline_vecs, &iter, false);
2799 static bool io_rw_reissue(struct io_kiocb *req)
2802 umode_t mode = file_inode(req->file)->i_mode;
2805 if (!S_ISBLK(mode) && !S_ISREG(mode))
2807 if ((req->flags & REQ_F_NOWAIT) || io_wq_current_is_worker())
2810 lockdep_assert_held(&req->ctx->uring_lock);
2812 ret = io_sq_thread_acquire_mm_files(req->ctx, req);
2814 if (!ret && io_resubmit_prep(req)) {
2815 refcount_inc(&req->refs);
2816 io_queue_async_work(req);
2819 req_set_fail_links(req);
2824 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2825 unsigned int issue_flags)
2829 if ((res == -EAGAIN || res == -EOPNOTSUPP) && io_rw_reissue(req))
2831 if (res != req->result)
2832 req_set_fail_links(req);
2834 if (req->rw.kiocb.ki_flags & IOCB_WRITE)
2835 kiocb_end_write(req);
2836 if (req->flags & REQ_F_BUFFER_SELECTED)
2837 cflags = io_put_rw_kbuf(req);
2838 __io_req_complete(req, issue_flags, res, cflags);
2841 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2843 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2845 __io_complete_rw(req, res, res2, 0);
2848 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2850 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2852 if (kiocb->ki_flags & IOCB_WRITE)
2853 kiocb_end_write(req);
2855 if (res != -EAGAIN && res != req->result)
2856 req_set_fail_links(req);
2858 WRITE_ONCE(req->result, res);
2859 /* order with io_poll_complete() checking ->result */
2861 WRITE_ONCE(req->iopoll_completed, 1);
2865 * After the iocb has been issued, it's safe to be found on the poll list.
2866 * Adding the kiocb to the list AFTER submission ensures that we don't
2867 * find it from a io_iopoll_getevents() thread before the issuer is done
2868 * accessing the kiocb cookie.
2870 static void io_iopoll_req_issued(struct io_kiocb *req, bool in_async)
2872 struct io_ring_ctx *ctx = req->ctx;
2875 * Track whether we have multiple files in our lists. This will impact
2876 * how we do polling eventually, not spinning if we're on potentially
2877 * different devices.
2879 if (list_empty(&ctx->iopoll_list)) {
2880 ctx->poll_multi_file = false;
2881 } else if (!ctx->poll_multi_file) {
2882 struct io_kiocb *list_req;
2884 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2886 if (list_req->file != req->file)
2887 ctx->poll_multi_file = true;
2891 * For fast devices, IO may have already completed. If it has, add
2892 * it to the front so we find it first.
2894 if (READ_ONCE(req->iopoll_completed))
2895 list_add(&req->inflight_entry, &ctx->iopoll_list);
2897 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2900 * If IORING_SETUP_SQPOLL is enabled, sqes are either handled in sq thread
2901 * task context or in io worker task context. If current task context is
2902 * sq thread, we don't need to check whether should wake up sq thread.
2904 if (in_async && (ctx->flags & IORING_SETUP_SQPOLL) &&
2905 wq_has_sleeper(&ctx->sq_data->wait))
2906 wake_up(&ctx->sq_data->wait);
2909 static inline void io_state_file_put(struct io_submit_state *state)
2911 if (state->file_refs) {
2912 fput_many(state->file, state->file_refs);
2913 state->file_refs = 0;
2918 * Get as many references to a file as we have IOs left in this submission,
2919 * assuming most submissions are for one file, or at least that each file
2920 * has more than one submission.
2922 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2927 if (state->file_refs) {
2928 if (state->fd == fd) {
2932 io_state_file_put(state);
2934 state->file = fget_many(fd, state->ios_left);
2935 if (unlikely(!state->file))
2939 state->file_refs = state->ios_left - 1;
2943 static bool io_bdev_nowait(struct block_device *bdev)
2945 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2949 * If we tracked the file through the SCM inflight mechanism, we could support
2950 * any file. For now, just ensure that anything potentially problematic is done
2953 static bool io_file_supports_async(struct file *file, int rw)
2955 umode_t mode = file_inode(file)->i_mode;
2957 if (S_ISBLK(mode)) {
2958 if (IS_ENABLED(CONFIG_BLOCK) &&
2959 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2963 if (S_ISCHR(mode) || S_ISSOCK(mode))
2965 if (S_ISREG(mode)) {
2966 if (IS_ENABLED(CONFIG_BLOCK) &&
2967 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2968 file->f_op != &io_uring_fops)
2973 /* any ->read/write should understand O_NONBLOCK */
2974 if (file->f_flags & O_NONBLOCK)
2977 if (!(file->f_mode & FMODE_NOWAIT))
2981 return file->f_op->read_iter != NULL;
2983 return file->f_op->write_iter != NULL;
2986 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2988 struct io_ring_ctx *ctx = req->ctx;
2989 struct kiocb *kiocb = &req->rw.kiocb;
2990 struct file *file = req->file;
2994 if (S_ISREG(file_inode(file)->i_mode))
2995 req->flags |= REQ_F_ISREG;
2997 kiocb->ki_pos = READ_ONCE(sqe->off);
2998 if (kiocb->ki_pos == -1 && !(file->f_mode & FMODE_STREAM)) {
2999 req->flags |= REQ_F_CUR_POS;
3000 kiocb->ki_pos = file->f_pos;
3002 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
3003 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
3004 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
3008 /* don't allow async punt for O_NONBLOCK or RWF_NOWAIT */
3009 if ((kiocb->ki_flags & IOCB_NOWAIT) || (file->f_flags & O_NONBLOCK))
3010 req->flags |= REQ_F_NOWAIT;
3012 ioprio = READ_ONCE(sqe->ioprio);
3014 ret = ioprio_check_cap(ioprio);
3018 kiocb->ki_ioprio = ioprio;
3020 kiocb->ki_ioprio = get_current_ioprio();
3022 if (ctx->flags & IORING_SETUP_IOPOLL) {
3023 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
3024 !kiocb->ki_filp->f_op->iopoll)
3027 kiocb->ki_flags |= IOCB_HIPRI;
3028 kiocb->ki_complete = io_complete_rw_iopoll;
3029 req->iopoll_completed = 0;
3031 if (kiocb->ki_flags & IOCB_HIPRI)
3033 kiocb->ki_complete = io_complete_rw;
3036 req->rw.addr = READ_ONCE(sqe->addr);
3037 req->rw.len = READ_ONCE(sqe->len);
3038 req->buf_index = READ_ONCE(sqe->buf_index);
3042 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
3048 case -ERESTARTNOINTR:
3049 case -ERESTARTNOHAND:
3050 case -ERESTART_RESTARTBLOCK:
3052 * We can't just restart the syscall, since previously
3053 * submitted sqes may already be in progress. Just fail this
3059 kiocb->ki_complete(kiocb, ret, 0);
3063 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
3064 unsigned int issue_flags)
3066 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
3067 struct io_async_rw *io = req->async_data;
3069 /* add previously done IO, if any */
3070 if (io && io->bytes_done > 0) {
3072 ret = io->bytes_done;
3074 ret += io->bytes_done;
3077 if (req->flags & REQ_F_CUR_POS)
3078 req->file->f_pos = kiocb->ki_pos;
3079 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
3080 __io_complete_rw(req, ret, 0, issue_flags);
3082 io_rw_done(kiocb, ret);
3085 static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
3087 struct io_ring_ctx *ctx = req->ctx;
3088 size_t len = req->rw.len;
3089 struct io_mapped_ubuf *imu;
3090 u16 index, buf_index = req->buf_index;
3094 if (unlikely(buf_index >= ctx->nr_user_bufs))
3096 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
3097 imu = &ctx->user_bufs[index];
3098 buf_addr = req->rw.addr;
3101 if (buf_addr + len < buf_addr)
3103 /* not inside the mapped region */
3104 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
3108 * May not be a start of buffer, set size appropriately
3109 * and advance us to the beginning.
3111 offset = buf_addr - imu->ubuf;
3112 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
3116 * Don't use iov_iter_advance() here, as it's really slow for
3117 * using the latter parts of a big fixed buffer - it iterates
3118 * over each segment manually. We can cheat a bit here, because
3121 * 1) it's a BVEC iter, we set it up
3122 * 2) all bvecs are PAGE_SIZE in size, except potentially the
3123 * first and last bvec
3125 * So just find our index, and adjust the iterator afterwards.
3126 * If the offset is within the first bvec (or the whole first
3127 * bvec, just use iov_iter_advance(). This makes it easier
3128 * since we can just skip the first segment, which may not
3129 * be PAGE_SIZE aligned.
3131 const struct bio_vec *bvec = imu->bvec;
3133 if (offset <= bvec->bv_len) {
3134 iov_iter_advance(iter, offset);
3136 unsigned long seg_skip;
3138 /* skip first vec */
3139 offset -= bvec->bv_len;
3140 seg_skip = 1 + (offset >> PAGE_SHIFT);
3142 iter->bvec = bvec + seg_skip;
3143 iter->nr_segs -= seg_skip;
3144 iter->count -= bvec->bv_len + offset;
3145 iter->iov_offset = offset & ~PAGE_MASK;
3152 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
3155 mutex_unlock(&ctx->uring_lock);
3158 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
3161 * "Normal" inline submissions always hold the uring_lock, since we
3162 * grab it from the system call. Same is true for the SQPOLL offload.
3163 * The only exception is when we've detached the request and issue it
3164 * from an async worker thread, grab the lock for that case.
3167 mutex_lock(&ctx->uring_lock);
3170 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
3171 int bgid, struct io_buffer *kbuf,
3174 struct io_buffer *head;
3176 if (req->flags & REQ_F_BUFFER_SELECTED)
3179 io_ring_submit_lock(req->ctx, needs_lock);
3181 lockdep_assert_held(&req->ctx->uring_lock);
3183 head = idr_find(&req->ctx->io_buffer_idr, bgid);
3185 if (!list_empty(&head->list)) {
3186 kbuf = list_last_entry(&head->list, struct io_buffer,
3188 list_del(&kbuf->list);
3191 idr_remove(&req->ctx->io_buffer_idr, bgid);
3193 if (*len > kbuf->len)
3196 kbuf = ERR_PTR(-ENOBUFS);
3199 io_ring_submit_unlock(req->ctx, needs_lock);
3204 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
3207 struct io_buffer *kbuf;
3210 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3211 bgid = req->buf_index;
3212 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
3215 req->rw.addr = (u64) (unsigned long) kbuf;
3216 req->flags |= REQ_F_BUFFER_SELECTED;
3217 return u64_to_user_ptr(kbuf->addr);
3220 #ifdef CONFIG_COMPAT
3221 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
3224 struct compat_iovec __user *uiov;
3225 compat_ssize_t clen;
3229 uiov = u64_to_user_ptr(req->rw.addr);
3230 if (!access_ok(uiov, sizeof(*uiov)))
3232 if (__get_user(clen, &uiov->iov_len))
3238 buf = io_rw_buffer_select(req, &len, needs_lock);
3240 return PTR_ERR(buf);
3241 iov[0].iov_base = buf;
3242 iov[0].iov_len = (compat_size_t) len;
3247 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3250 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
3254 if (copy_from_user(iov, uiov, sizeof(*uiov)))
3257 len = iov[0].iov_len;
3260 buf = io_rw_buffer_select(req, &len, needs_lock);
3262 return PTR_ERR(buf);
3263 iov[0].iov_base = buf;
3264 iov[0].iov_len = len;
3268 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3271 if (req->flags & REQ_F_BUFFER_SELECTED) {
3272 struct io_buffer *kbuf;
3274 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3275 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
3276 iov[0].iov_len = kbuf->len;
3279 if (req->rw.len != 1)
3282 #ifdef CONFIG_COMPAT
3283 if (req->ctx->compat)
3284 return io_compat_import(req, iov, needs_lock);
3287 return __io_iov_buffer_select(req, iov, needs_lock);
3290 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
3291 struct iov_iter *iter, bool needs_lock)
3293 void __user *buf = u64_to_user_ptr(req->rw.addr);
3294 size_t sqe_len = req->rw.len;
3295 u8 opcode = req->opcode;
3298 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3300 return io_import_fixed(req, rw, iter);
3303 /* buffer index only valid with fixed read/write, or buffer select */
3304 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3307 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3308 if (req->flags & REQ_F_BUFFER_SELECT) {
3309 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3311 return PTR_ERR(buf);
3312 req->rw.len = sqe_len;
3315 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3320 if (req->flags & REQ_F_BUFFER_SELECT) {
3321 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3323 iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
3328 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3332 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3334 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3338 * For files that don't have ->read_iter() and ->write_iter(), handle them
3339 * by looping over ->read() or ->write() manually.
3341 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3343 struct kiocb *kiocb = &req->rw.kiocb;
3344 struct file *file = req->file;
3348 * Don't support polled IO through this interface, and we can't
3349 * support non-blocking either. For the latter, this just causes
3350 * the kiocb to be handled from an async context.
3352 if (kiocb->ki_flags & IOCB_HIPRI)
3354 if (kiocb->ki_flags & IOCB_NOWAIT)
3357 while (iov_iter_count(iter)) {
3361 if (!iov_iter_is_bvec(iter)) {
3362 iovec = iov_iter_iovec(iter);
3364 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3365 iovec.iov_len = req->rw.len;
3369 nr = file->f_op->read(file, iovec.iov_base,
3370 iovec.iov_len, io_kiocb_ppos(kiocb));
3372 nr = file->f_op->write(file, iovec.iov_base,
3373 iovec.iov_len, io_kiocb_ppos(kiocb));
3382 if (nr != iovec.iov_len)
3386 iov_iter_advance(iter, nr);
3392 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3393 const struct iovec *fast_iov, struct iov_iter *iter)
3395 struct io_async_rw *rw = req->async_data;
3397 memcpy(&rw->iter, iter, sizeof(*iter));
3398 rw->free_iovec = iovec;
3400 /* can only be fixed buffers, no need to do anything */
3401 if (iov_iter_is_bvec(iter))
3404 unsigned iov_off = 0;
3406 rw->iter.iov = rw->fast_iov;
3407 if (iter->iov != fast_iov) {
3408 iov_off = iter->iov - fast_iov;
3409 rw->iter.iov += iov_off;
3411 if (rw->fast_iov != fast_iov)
3412 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3413 sizeof(struct iovec) * iter->nr_segs);
3415 req->flags |= REQ_F_NEED_CLEANUP;
3419 static inline int __io_alloc_async_data(struct io_kiocb *req)
3421 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3422 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3423 return req->async_data == NULL;
3426 static int io_alloc_async_data(struct io_kiocb *req)
3428 if (!io_op_defs[req->opcode].needs_async_data)
3431 return __io_alloc_async_data(req);
3434 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3435 const struct iovec *fast_iov,
3436 struct iov_iter *iter, bool force)
3438 if (!force && !io_op_defs[req->opcode].needs_async_data)
3440 if (!req->async_data) {
3441 if (__io_alloc_async_data(req)) {
3446 io_req_map_rw(req, iovec, fast_iov, iter);
3451 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3453 struct io_async_rw *iorw = req->async_data;
3454 struct iovec *iov = iorw->fast_iov;
3457 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3458 if (unlikely(ret < 0))
3461 iorw->bytes_done = 0;
3462 iorw->free_iovec = iov;
3464 req->flags |= REQ_F_NEED_CLEANUP;
3468 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3472 ret = io_prep_rw(req, sqe);
3476 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3479 /* either don't need iovec imported or already have it */
3480 if (!req->async_data)
3482 return io_rw_prep_async(req, READ);
3486 * This is our waitqueue callback handler, registered through lock_page_async()
3487 * when we initially tried to do the IO with the iocb armed our waitqueue.
3488 * This gets called when the page is unlocked, and we generally expect that to
3489 * happen when the page IO is completed and the page is now uptodate. This will
3490 * queue a task_work based retry of the operation, attempting to copy the data
3491 * again. If the latter fails because the page was NOT uptodate, then we will
3492 * do a thread based blocking retry of the operation. That's the unexpected
3495 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3496 int sync, void *arg)
3498 struct wait_page_queue *wpq;
3499 struct io_kiocb *req = wait->private;
3500 struct wait_page_key *key = arg;
3503 wpq = container_of(wait, struct wait_page_queue, wait);
3505 if (!wake_page_match(wpq, key))
3508 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3509 list_del_init(&wait->entry);
3511 req->task_work.func = io_req_task_submit;
3512 percpu_ref_get(&req->ctx->refs);
3514 /* submit ref gets dropped, acquire a new one */
3515 refcount_inc(&req->refs);
3516 ret = io_req_task_work_add(req);
3518 io_req_task_work_add_fallback(req, io_req_task_cancel);
3523 * This controls whether a given IO request should be armed for async page
3524 * based retry. If we return false here, the request is handed to the async
3525 * worker threads for retry. If we're doing buffered reads on a regular file,
3526 * we prepare a private wait_page_queue entry and retry the operation. This
3527 * will either succeed because the page is now uptodate and unlocked, or it
3528 * will register a callback when the page is unlocked at IO completion. Through
3529 * that callback, io_uring uses task_work to setup a retry of the operation.
3530 * That retry will attempt the buffered read again. The retry will generally
3531 * succeed, or in rare cases where it fails, we then fall back to using the
3532 * async worker threads for a blocking retry.
3534 static bool io_rw_should_retry(struct io_kiocb *req)
3536 struct io_async_rw *rw = req->async_data;
3537 struct wait_page_queue *wait = &rw->wpq;
3538 struct kiocb *kiocb = &req->rw.kiocb;
3540 /* never retry for NOWAIT, we just complete with -EAGAIN */
3541 if (req->flags & REQ_F_NOWAIT)
3544 /* Only for buffered IO */
3545 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3549 * just use poll if we can, and don't attempt if the fs doesn't
3550 * support callback based unlocks
3552 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3555 wait->wait.func = io_async_buf_func;
3556 wait->wait.private = req;
3557 wait->wait.flags = 0;
3558 INIT_LIST_HEAD(&wait->wait.entry);
3559 kiocb->ki_flags |= IOCB_WAITQ;
3560 kiocb->ki_flags &= ~IOCB_NOWAIT;
3561 kiocb->ki_waitq = wait;
3565 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3567 if (req->file->f_op->read_iter)
3568 return call_read_iter(req->file, &req->rw.kiocb, iter);
3569 else if (req->file->f_op->read)
3570 return loop_rw_iter(READ, req, iter);
3575 static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3577 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3578 struct kiocb *kiocb = &req->rw.kiocb;
3579 struct iov_iter __iter, *iter = &__iter;
3580 struct io_async_rw *rw = req->async_data;
3581 ssize_t io_size, ret, ret2;
3582 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3588 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3592 io_size = iov_iter_count(iter);
3593 req->result = io_size;
3595 /* Ensure we clear previously set non-block flag */
3596 if (!force_nonblock)
3597 kiocb->ki_flags &= ~IOCB_NOWAIT;
3599 kiocb->ki_flags |= IOCB_NOWAIT;
3601 /* If the file doesn't support async, just async punt */
3602 if (force_nonblock && !io_file_supports_async(req->file, READ)) {
3603 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3604 return ret ?: -EAGAIN;
3607 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size);
3608 if (unlikely(ret)) {
3613 ret = io_iter_do_read(req, iter);
3615 if (ret == -EIOCBQUEUED) {
3616 /* it's faster to check here then delegate to kfree */
3620 } else if (ret == -EAGAIN) {
3621 /* IOPOLL retry should happen for io-wq threads */
3622 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3624 /* no retry on NONBLOCK nor RWF_NOWAIT */
3625 if (req->flags & REQ_F_NOWAIT)
3627 /* some cases will consume bytes even on error returns */
3628 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3630 } else if (ret <= 0 || ret == io_size || !force_nonblock ||
3631 (req->flags & REQ_F_NOWAIT) || !(req->flags & REQ_F_ISREG)) {
3632 /* read all, failed, already did sync or don't want to retry */
3636 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3640 rw = req->async_data;
3641 /* now use our persistent iterator, if we aren't already */
3646 rw->bytes_done += ret;
3647 /* if we can retry, do so with the callbacks armed */
3648 if (!io_rw_should_retry(req)) {
3649 kiocb->ki_flags &= ~IOCB_WAITQ;
3654 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3655 * we get -EIOCBQUEUED, then we'll get a notification when the
3656 * desired page gets unlocked. We can also get a partial read
3657 * here, and if we do, then just retry at the new offset.
3659 ret = io_iter_do_read(req, iter);
3660 if (ret == -EIOCBQUEUED)
3662 /* we got some bytes, but not all. retry. */
3663 } while (ret > 0 && ret < io_size);
3665 kiocb_done(kiocb, ret, issue_flags);
3669 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3673 ret = io_prep_rw(req, sqe);
3677 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3680 /* either don't need iovec imported or already have it */
3681 if (!req->async_data)
3683 return io_rw_prep_async(req, WRITE);
3686 static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3688 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3689 struct kiocb *kiocb = &req->rw.kiocb;
3690 struct iov_iter __iter, *iter = &__iter;
3691 struct io_async_rw *rw = req->async_data;
3692 ssize_t ret, ret2, io_size;
3693 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3699 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3703 io_size = iov_iter_count(iter);
3704 req->result = io_size;
3706 /* Ensure we clear previously set non-block flag */
3707 if (!force_nonblock)
3708 kiocb->ki_flags &= ~IOCB_NOWAIT;
3710 kiocb->ki_flags |= IOCB_NOWAIT;
3712 /* If the file doesn't support async, just async punt */
3713 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
3716 /* file path doesn't support NOWAIT for non-direct_IO */
3717 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3718 (req->flags & REQ_F_ISREG))
3721 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size);
3726 * Open-code file_start_write here to grab freeze protection,
3727 * which will be released by another thread in
3728 * io_complete_rw(). Fool lockdep by telling it the lock got
3729 * released so that it doesn't complain about the held lock when
3730 * we return to userspace.
3732 if (req->flags & REQ_F_ISREG) {
3733 sb_start_write(file_inode(req->file)->i_sb);
3734 __sb_writers_release(file_inode(req->file)->i_sb,
3737 kiocb->ki_flags |= IOCB_WRITE;
3739 if (req->file->f_op->write_iter)
3740 ret2 = call_write_iter(req->file, kiocb, iter);
3741 else if (req->file->f_op->write)
3742 ret2 = loop_rw_iter(WRITE, req, iter);
3747 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3748 * retry them without IOCB_NOWAIT.
3750 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3752 /* no retry on NONBLOCK nor RWF_NOWAIT */
3753 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3755 if (!force_nonblock || ret2 != -EAGAIN) {
3756 /* IOPOLL retry should happen for io-wq threads */
3757 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3760 kiocb_done(kiocb, ret2, issue_flags);
3763 /* some cases will consume bytes even on error returns */
3764 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3765 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3766 return ret ?: -EAGAIN;
3769 /* it's reportedly faster than delegating the null check to kfree() */
3775 static int io_renameat_prep(struct io_kiocb *req,
3776 const struct io_uring_sqe *sqe)
3778 struct io_rename *ren = &req->rename;
3779 const char __user *oldf, *newf;
3781 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3784 ren->old_dfd = READ_ONCE(sqe->fd);
3785 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3786 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3787 ren->new_dfd = READ_ONCE(sqe->len);
3788 ren->flags = READ_ONCE(sqe->rename_flags);
3790 ren->oldpath = getname(oldf);
3791 if (IS_ERR(ren->oldpath))
3792 return PTR_ERR(ren->oldpath);
3794 ren->newpath = getname(newf);
3795 if (IS_ERR(ren->newpath)) {
3796 putname(ren->oldpath);
3797 return PTR_ERR(ren->newpath);
3800 req->flags |= REQ_F_NEED_CLEANUP;
3804 static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3806 struct io_rename *ren = &req->rename;
3809 if (issue_flags & IO_URING_F_NONBLOCK)
3812 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3813 ren->newpath, ren->flags);
3815 req->flags &= ~REQ_F_NEED_CLEANUP;
3817 req_set_fail_links(req);
3818 io_req_complete(req, ret);
3822 static int io_unlinkat_prep(struct io_kiocb *req,
3823 const struct io_uring_sqe *sqe)
3825 struct io_unlink *un = &req->unlink;
3826 const char __user *fname;
3828 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3831 un->dfd = READ_ONCE(sqe->fd);
3833 un->flags = READ_ONCE(sqe->unlink_flags);
3834 if (un->flags & ~AT_REMOVEDIR)
3837 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3838 un->filename = getname(fname);
3839 if (IS_ERR(un->filename))
3840 return PTR_ERR(un->filename);
3842 req->flags |= REQ_F_NEED_CLEANUP;
3846 static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3848 struct io_unlink *un = &req->unlink;
3851 if (issue_flags & IO_URING_F_NONBLOCK)
3854 if (un->flags & AT_REMOVEDIR)
3855 ret = do_rmdir(un->dfd, un->filename);
3857 ret = do_unlinkat(un->dfd, un->filename);
3859 req->flags &= ~REQ_F_NEED_CLEANUP;
3861 req_set_fail_links(req);
3862 io_req_complete(req, ret);
3866 static int io_shutdown_prep(struct io_kiocb *req,
3867 const struct io_uring_sqe *sqe)
3869 #if defined(CONFIG_NET)
3870 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3872 if (sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3876 req->shutdown.how = READ_ONCE(sqe->len);
3883 static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3885 #if defined(CONFIG_NET)
3886 struct socket *sock;
3889 if (issue_flags & IO_URING_F_NONBLOCK)
3892 sock = sock_from_file(req->file);
3893 if (unlikely(!sock))
3896 ret = __sys_shutdown_sock(sock, req->shutdown.how);
3898 req_set_fail_links(req);
3899 io_req_complete(req, ret);
3906 static int __io_splice_prep(struct io_kiocb *req,
3907 const struct io_uring_sqe *sqe)
3909 struct io_splice* sp = &req->splice;
3910 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3912 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3916 sp->len = READ_ONCE(sqe->len);
3917 sp->flags = READ_ONCE(sqe->splice_flags);
3919 if (unlikely(sp->flags & ~valid_flags))
3922 sp->file_in = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in),
3923 (sp->flags & SPLICE_F_FD_IN_FIXED));
3926 req->flags |= REQ_F_NEED_CLEANUP;
3928 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
3930 * Splice operation will be punted aync, and here need to
3931 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3933 io_req_init_async(req);
3934 req->work.flags |= IO_WQ_WORK_UNBOUND;
3940 static int io_tee_prep(struct io_kiocb *req,
3941 const struct io_uring_sqe *sqe)
3943 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3945 return __io_splice_prep(req, sqe);
3948 static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
3950 struct io_splice *sp = &req->splice;
3951 struct file *in = sp->file_in;
3952 struct file *out = sp->file_out;
3953 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3956 if (issue_flags & IO_URING_F_NONBLOCK)
3959 ret = do_tee(in, out, sp->len, flags);
3961 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3962 req->flags &= ~REQ_F_NEED_CLEANUP;
3965 req_set_fail_links(req);
3966 io_req_complete(req, ret);
3970 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3972 struct io_splice* sp = &req->splice;
3974 sp->off_in = READ_ONCE(sqe->splice_off_in);
3975 sp->off_out = READ_ONCE(sqe->off);
3976 return __io_splice_prep(req, sqe);
3979 static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
3981 struct io_splice *sp = &req->splice;
3982 struct file *in = sp->file_in;
3983 struct file *out = sp->file_out;
3984 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3985 loff_t *poff_in, *poff_out;
3988 if (issue_flags & IO_URING_F_NONBLOCK)
3991 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3992 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3995 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3997 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3998 req->flags &= ~REQ_F_NEED_CLEANUP;
4001 req_set_fail_links(req);
4002 io_req_complete(req, ret);
4007 * IORING_OP_NOP just posts a completion event, nothing else.
4009 static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
4011 struct io_ring_ctx *ctx = req->ctx;
4013 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4016 __io_req_complete(req, issue_flags, 0, 0);
4020 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4022 struct io_ring_ctx *ctx = req->ctx;
4027 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4029 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4032 req->sync.flags = READ_ONCE(sqe->fsync_flags);
4033 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
4036 req->sync.off = READ_ONCE(sqe->off);
4037 req->sync.len = READ_ONCE(sqe->len);
4041 static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
4043 loff_t end = req->sync.off + req->sync.len;
4046 /* fsync always requires a blocking context */
4047 if (issue_flags & IO_URING_F_NONBLOCK)
4050 ret = vfs_fsync_range(req->file, req->sync.off,
4051 end > 0 ? end : LLONG_MAX,
4052 req->sync.flags & IORING_FSYNC_DATASYNC);
4054 req_set_fail_links(req);
4055 io_req_complete(req, ret);
4059 static int io_fallocate_prep(struct io_kiocb *req,
4060 const struct io_uring_sqe *sqe)
4062 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
4064 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4067 req->sync.off = READ_ONCE(sqe->off);
4068 req->sync.len = READ_ONCE(sqe->addr);
4069 req->sync.mode = READ_ONCE(sqe->len);
4073 static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
4077 /* fallocate always requiring blocking context */
4078 if (issue_flags & IO_URING_F_NONBLOCK)
4080 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
4083 req_set_fail_links(req);
4084 io_req_complete(req, ret);
4088 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4090 const char __user *fname;
4093 if (unlikely(sqe->ioprio || sqe->buf_index))
4095 if (unlikely(req->flags & REQ_F_FIXED_FILE))
4098 /* open.how should be already initialised */
4099 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
4100 req->open.how.flags |= O_LARGEFILE;
4102 req->open.dfd = READ_ONCE(sqe->fd);
4103 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
4104 req->open.filename = getname(fname);
4105 if (IS_ERR(req->open.filename)) {
4106 ret = PTR_ERR(req->open.filename);
4107 req->open.filename = NULL;
4110 req->open.nofile = rlimit(RLIMIT_NOFILE);
4111 req->flags |= REQ_F_NEED_CLEANUP;
4115 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4119 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4121 mode = READ_ONCE(sqe->len);
4122 flags = READ_ONCE(sqe->open_flags);
4123 req->open.how = build_open_how(flags, mode);
4124 return __io_openat_prep(req, sqe);
4127 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4129 struct open_how __user *how;
4133 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4135 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4136 len = READ_ONCE(sqe->len);
4137 if (len < OPEN_HOW_SIZE_VER0)
4140 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
4145 return __io_openat_prep(req, sqe);
4148 static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
4150 struct open_flags op;
4153 bool resolve_nonblock;
4156 ret = build_open_flags(&req->open.how, &op);
4159 nonblock_set = op.open_flag & O_NONBLOCK;
4160 resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
4161 if (issue_flags & IO_URING_F_NONBLOCK) {
4163 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
4164 * it'll always -EAGAIN
4166 if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
4168 op.lookup_flags |= LOOKUP_CACHED;
4169 op.open_flag |= O_NONBLOCK;
4172 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
4176 file = do_filp_open(req->open.dfd, req->open.filename, &op);
4177 /* only retry if RESOLVE_CACHED wasn't already set by application */
4178 if ((!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)) &&
4179 file == ERR_PTR(-EAGAIN)) {
4181 * We could hang on to this 'fd', but seems like marginal
4182 * gain for something that is now known to be a slower path.
4183 * So just put it, and we'll get a new one when we retry.
4191 ret = PTR_ERR(file);
4193 if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
4194 file->f_flags &= ~O_NONBLOCK;
4195 fsnotify_open(file);
4196 fd_install(ret, file);
4199 putname(req->open.filename);
4200 req->flags &= ~REQ_F_NEED_CLEANUP;
4202 req_set_fail_links(req);
4203 io_req_complete(req, ret);
4207 static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
4209 return io_openat2(req, issue_flags & IO_URING_F_NONBLOCK);
4212 static int io_remove_buffers_prep(struct io_kiocb *req,
4213 const struct io_uring_sqe *sqe)
4215 struct io_provide_buf *p = &req->pbuf;
4218 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
4221 tmp = READ_ONCE(sqe->fd);
4222 if (!tmp || tmp > USHRT_MAX)
4225 memset(p, 0, sizeof(*p));
4227 p->bgid = READ_ONCE(sqe->buf_group);
4231 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
4232 int bgid, unsigned nbufs)
4236 /* shouldn't happen */
4240 /* the head kbuf is the list itself */
4241 while (!list_empty(&buf->list)) {
4242 struct io_buffer *nxt;
4244 nxt = list_first_entry(&buf->list, struct io_buffer, list);
4245 list_del(&nxt->list);
4252 idr_remove(&ctx->io_buffer_idr, bgid);
4257 static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
4259 struct io_provide_buf *p = &req->pbuf;
4260 struct io_ring_ctx *ctx = req->ctx;
4261 struct io_buffer *head;
4263 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4265 io_ring_submit_lock(ctx, !force_nonblock);
4267 lockdep_assert_held(&ctx->uring_lock);
4270 head = idr_find(&ctx->io_buffer_idr, p->bgid);
4272 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
4274 req_set_fail_links(req);
4276 /* need to hold the lock to complete IOPOLL requests */
4277 if (ctx->flags & IORING_SETUP_IOPOLL) {
4278 __io_req_complete(req, issue_flags, ret, 0);
4279 io_ring_submit_unlock(ctx, !force_nonblock);
4281 io_ring_submit_unlock(ctx, !force_nonblock);
4282 __io_req_complete(req, issue_flags, ret, 0);
4287 static int io_provide_buffers_prep(struct io_kiocb *req,
4288 const struct io_uring_sqe *sqe)
4290 struct io_provide_buf *p = &req->pbuf;
4293 if (sqe->ioprio || sqe->rw_flags)
4296 tmp = READ_ONCE(sqe->fd);
4297 if (!tmp || tmp > USHRT_MAX)
4300 p->addr = READ_ONCE(sqe->addr);
4301 p->len = READ_ONCE(sqe->len);
4303 if (!access_ok(u64_to_user_ptr(p->addr), (p->len * p->nbufs)))
4306 p->bgid = READ_ONCE(sqe->buf_group);
4307 tmp = READ_ONCE(sqe->off);
4308 if (tmp > USHRT_MAX)
4314 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
4316 struct io_buffer *buf;
4317 u64 addr = pbuf->addr;
4318 int i, bid = pbuf->bid;
4320 for (i = 0; i < pbuf->nbufs; i++) {
4321 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
4326 buf->len = pbuf->len;
4331 INIT_LIST_HEAD(&buf->list);
4334 list_add_tail(&buf->list, &(*head)->list);
4338 return i ? i : -ENOMEM;
4341 static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
4343 struct io_provide_buf *p = &req->pbuf;
4344 struct io_ring_ctx *ctx = req->ctx;
4345 struct io_buffer *head, *list;
4347 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4349 io_ring_submit_lock(ctx, !force_nonblock);
4351 lockdep_assert_held(&ctx->uring_lock);
4353 list = head = idr_find(&ctx->io_buffer_idr, p->bgid);
4355 ret = io_add_buffers(p, &head);
4360 ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1,
4363 __io_remove_buffers(ctx, head, p->bgid, -1U);
4369 req_set_fail_links(req);
4371 /* need to hold the lock to complete IOPOLL requests */
4372 if (ctx->flags & IORING_SETUP_IOPOLL) {
4373 __io_req_complete(req, issue_flags, ret, 0);
4374 io_ring_submit_unlock(ctx, !force_nonblock);
4376 io_ring_submit_unlock(ctx, !force_nonblock);
4377 __io_req_complete(req, issue_flags, ret, 0);
4382 static int io_epoll_ctl_prep(struct io_kiocb *req,
4383 const struct io_uring_sqe *sqe)
4385 #if defined(CONFIG_EPOLL)
4386 if (sqe->ioprio || sqe->buf_index)
4388 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4391 req->epoll.epfd = READ_ONCE(sqe->fd);
4392 req->epoll.op = READ_ONCE(sqe->len);
4393 req->epoll.fd = READ_ONCE(sqe->off);
4395 if (ep_op_has_event(req->epoll.op)) {
4396 struct epoll_event __user *ev;
4398 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4399 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4409 static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4411 #if defined(CONFIG_EPOLL)
4412 struct io_epoll *ie = &req->epoll;
4414 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4416 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4417 if (force_nonblock && ret == -EAGAIN)
4421 req_set_fail_links(req);
4422 __io_req_complete(req, issue_flags, ret, 0);
4429 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4431 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4432 if (sqe->ioprio || sqe->buf_index || sqe->off)
4434 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4437 req->madvise.addr = READ_ONCE(sqe->addr);
4438 req->madvise.len = READ_ONCE(sqe->len);
4439 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4446 static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4448 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4449 struct io_madvise *ma = &req->madvise;
4452 if (issue_flags & IO_URING_F_NONBLOCK)
4455 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4457 req_set_fail_links(req);
4458 io_req_complete(req, ret);
4465 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4467 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4469 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4472 req->fadvise.offset = READ_ONCE(sqe->off);
4473 req->fadvise.len = READ_ONCE(sqe->len);
4474 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4478 static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4480 struct io_fadvise *fa = &req->fadvise;
4483 if (issue_flags & IO_URING_F_NONBLOCK) {
4484 switch (fa->advice) {
4485 case POSIX_FADV_NORMAL:
4486 case POSIX_FADV_RANDOM:
4487 case POSIX_FADV_SEQUENTIAL:
4494 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4496 req_set_fail_links(req);
4497 io_req_complete(req, ret);
4501 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4503 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4505 if (sqe->ioprio || sqe->buf_index)
4507 if (req->flags & REQ_F_FIXED_FILE)
4510 req->statx.dfd = READ_ONCE(sqe->fd);
4511 req->statx.mask = READ_ONCE(sqe->len);
4512 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4513 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4514 req->statx.flags = READ_ONCE(sqe->statx_flags);
4519 static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4521 struct io_statx *ctx = &req->statx;
4524 if (issue_flags & IO_URING_F_NONBLOCK) {
4525 /* only need file table for an actual valid fd */
4526 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
4527 req->flags |= REQ_F_NO_FILE_TABLE;
4531 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4535 req_set_fail_links(req);
4536 io_req_complete(req, ret);
4540 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4542 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4544 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4545 sqe->rw_flags || sqe->buf_index)
4547 if (req->flags & REQ_F_FIXED_FILE)
4550 req->close.fd = READ_ONCE(sqe->fd);
4554 static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4556 struct files_struct *files = current->files;
4557 struct io_close *close = &req->close;
4558 struct fdtable *fdt;
4564 spin_lock(&files->file_lock);
4565 fdt = files_fdtable(files);
4566 if (close->fd >= fdt->max_fds) {
4567 spin_unlock(&files->file_lock);
4570 file = fdt->fd[close->fd];
4572 spin_unlock(&files->file_lock);
4576 if (file->f_op == &io_uring_fops) {
4577 spin_unlock(&files->file_lock);
4582 /* if the file has a flush method, be safe and punt to async */
4583 if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4584 spin_unlock(&files->file_lock);
4588 ret = __close_fd_get_file(close->fd, &file);
4589 spin_unlock(&files->file_lock);
4596 /* No ->flush() or already async, safely close from here */
4597 ret = filp_close(file, current->files);
4600 req_set_fail_links(req);
4603 __io_req_complete(req, issue_flags, ret, 0);
4607 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4609 struct io_ring_ctx *ctx = req->ctx;
4614 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4616 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4619 req->sync.off = READ_ONCE(sqe->off);
4620 req->sync.len = READ_ONCE(sqe->len);
4621 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4625 static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4629 /* sync_file_range always requires a blocking context */
4630 if (issue_flags & IO_URING_F_NONBLOCK)
4633 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4636 req_set_fail_links(req);
4637 io_req_complete(req, ret);
4641 #if defined(CONFIG_NET)
4642 static int io_setup_async_msg(struct io_kiocb *req,
4643 struct io_async_msghdr *kmsg)
4645 struct io_async_msghdr *async_msg = req->async_data;
4649 if (io_alloc_async_data(req)) {
4650 kfree(kmsg->free_iov);
4653 async_msg = req->async_data;
4654 req->flags |= REQ_F_NEED_CLEANUP;
4655 memcpy(async_msg, kmsg, sizeof(*kmsg));
4656 async_msg->msg.msg_name = &async_msg->addr;
4657 /* if were using fast_iov, set it to the new one */
4658 if (!async_msg->free_iov)
4659 async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4664 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4665 struct io_async_msghdr *iomsg)
4667 iomsg->msg.msg_name = &iomsg->addr;
4668 iomsg->free_iov = iomsg->fast_iov;
4669 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4670 req->sr_msg.msg_flags, &iomsg->free_iov);
4673 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4675 struct io_async_msghdr *async_msg = req->async_data;
4676 struct io_sr_msg *sr = &req->sr_msg;
4679 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4682 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4683 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4684 sr->len = READ_ONCE(sqe->len);
4686 #ifdef CONFIG_COMPAT
4687 if (req->ctx->compat)
4688 sr->msg_flags |= MSG_CMSG_COMPAT;
4691 if (!async_msg || !io_op_defs[req->opcode].needs_async_data)
4693 ret = io_sendmsg_copy_hdr(req, async_msg);
4695 req->flags |= REQ_F_NEED_CLEANUP;
4699 static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4701 struct io_async_msghdr iomsg, *kmsg;
4702 struct socket *sock;
4706 sock = sock_from_file(req->file);
4707 if (unlikely(!sock))
4710 kmsg = req->async_data;
4712 ret = io_sendmsg_copy_hdr(req, &iomsg);
4718 flags = req->sr_msg.msg_flags;
4719 if (flags & MSG_DONTWAIT)
4720 req->flags |= REQ_F_NOWAIT;
4721 else if (issue_flags & IO_URING_F_NONBLOCK)
4722 flags |= MSG_DONTWAIT;
4724 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4725 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4726 return io_setup_async_msg(req, kmsg);
4727 if (ret == -ERESTARTSYS)
4730 /* fast path, check for non-NULL to avoid function call */
4732 kfree(kmsg->free_iov);
4733 req->flags &= ~REQ_F_NEED_CLEANUP;
4735 req_set_fail_links(req);
4736 __io_req_complete(req, issue_flags, ret, 0);
4740 static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4742 struct io_sr_msg *sr = &req->sr_msg;
4745 struct socket *sock;
4749 sock = sock_from_file(req->file);
4750 if (unlikely(!sock))
4753 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4757 msg.msg_name = NULL;
4758 msg.msg_control = NULL;
4759 msg.msg_controllen = 0;
4760 msg.msg_namelen = 0;
4762 flags = req->sr_msg.msg_flags;
4763 if (flags & MSG_DONTWAIT)
4764 req->flags |= REQ_F_NOWAIT;
4765 else if (issue_flags & IO_URING_F_NONBLOCK)
4766 flags |= MSG_DONTWAIT;
4768 msg.msg_flags = flags;
4769 ret = sock_sendmsg(sock, &msg);
4770 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4772 if (ret == -ERESTARTSYS)
4776 req_set_fail_links(req);
4777 __io_req_complete(req, issue_flags, ret, 0);
4781 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4782 struct io_async_msghdr *iomsg)
4784 struct io_sr_msg *sr = &req->sr_msg;
4785 struct iovec __user *uiov;
4789 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4790 &iomsg->uaddr, &uiov, &iov_len);
4794 if (req->flags & REQ_F_BUFFER_SELECT) {
4797 if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4799 sr->len = iomsg->fast_iov[0].iov_len;
4800 iomsg->free_iov = NULL;
4802 iomsg->free_iov = iomsg->fast_iov;
4803 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4804 &iomsg->free_iov, &iomsg->msg.msg_iter,
4813 #ifdef CONFIG_COMPAT
4814 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4815 struct io_async_msghdr *iomsg)
4817 struct compat_msghdr __user *msg_compat;
4818 struct io_sr_msg *sr = &req->sr_msg;
4819 struct compat_iovec __user *uiov;
4824 msg_compat = (struct compat_msghdr __user *) sr->umsg;
4825 ret = __get_compat_msghdr(&iomsg->msg, msg_compat, &iomsg->uaddr,
4830 uiov = compat_ptr(ptr);
4831 if (req->flags & REQ_F_BUFFER_SELECT) {
4832 compat_ssize_t clen;
4836 if (!access_ok(uiov, sizeof(*uiov)))
4838 if (__get_user(clen, &uiov->iov_len))
4843 iomsg->free_iov = NULL;
4845 iomsg->free_iov = iomsg->fast_iov;
4846 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4847 UIO_FASTIOV, &iomsg->free_iov,
4848 &iomsg->msg.msg_iter, true);
4857 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4858 struct io_async_msghdr *iomsg)
4860 iomsg->msg.msg_name = &iomsg->addr;
4862 #ifdef CONFIG_COMPAT
4863 if (req->ctx->compat)
4864 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4867 return __io_recvmsg_copy_hdr(req, iomsg);
4870 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4873 struct io_sr_msg *sr = &req->sr_msg;
4874 struct io_buffer *kbuf;
4876 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4881 req->flags |= REQ_F_BUFFER_SELECTED;
4885 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4887 return io_put_kbuf(req, req->sr_msg.kbuf);
4890 static int io_recvmsg_prep(struct io_kiocb *req,
4891 const struct io_uring_sqe *sqe)
4893 struct io_async_msghdr *async_msg = req->async_data;
4894 struct io_sr_msg *sr = &req->sr_msg;
4897 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4900 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4901 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4902 sr->len = READ_ONCE(sqe->len);
4903 sr->bgid = READ_ONCE(sqe->buf_group);
4905 #ifdef CONFIG_COMPAT
4906 if (req->ctx->compat)
4907 sr->msg_flags |= MSG_CMSG_COMPAT;
4910 if (!async_msg || !io_op_defs[req->opcode].needs_async_data)
4912 ret = io_recvmsg_copy_hdr(req, async_msg);
4914 req->flags |= REQ_F_NEED_CLEANUP;
4918 static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
4920 struct io_async_msghdr iomsg, *kmsg;
4921 struct socket *sock;
4922 struct io_buffer *kbuf;
4924 int ret, cflags = 0;
4925 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4927 sock = sock_from_file(req->file);
4928 if (unlikely(!sock))
4931 kmsg = req->async_data;
4933 ret = io_recvmsg_copy_hdr(req, &iomsg);
4939 if (req->flags & REQ_F_BUFFER_SELECT) {
4940 kbuf = io_recv_buffer_select(req, !force_nonblock);
4942 return PTR_ERR(kbuf);
4943 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4944 kmsg->fast_iov[0].iov_len = req->sr_msg.len;
4945 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
4946 1, req->sr_msg.len);
4949 flags = req->sr_msg.msg_flags;
4950 if (flags & MSG_DONTWAIT)
4951 req->flags |= REQ_F_NOWAIT;
4952 else if (force_nonblock)
4953 flags |= MSG_DONTWAIT;
4955 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4956 kmsg->uaddr, flags);
4957 if (force_nonblock && ret == -EAGAIN)
4958 return io_setup_async_msg(req, kmsg);
4959 if (ret == -ERESTARTSYS)
4962 if (req->flags & REQ_F_BUFFER_SELECTED)
4963 cflags = io_put_recv_kbuf(req);
4964 /* fast path, check for non-NULL to avoid function call */
4966 kfree(kmsg->free_iov);
4967 req->flags &= ~REQ_F_NEED_CLEANUP;
4969 req_set_fail_links(req);
4970 __io_req_complete(req, issue_flags, ret, cflags);
4974 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
4976 struct io_buffer *kbuf;
4977 struct io_sr_msg *sr = &req->sr_msg;
4979 void __user *buf = sr->buf;
4980 struct socket *sock;
4983 int ret, cflags = 0;
4984 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4986 sock = sock_from_file(req->file);
4987 if (unlikely(!sock))
4990 if (req->flags & REQ_F_BUFFER_SELECT) {
4991 kbuf = io_recv_buffer_select(req, !force_nonblock);
4993 return PTR_ERR(kbuf);
4994 buf = u64_to_user_ptr(kbuf->addr);
4997 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
5001 msg.msg_name = NULL;
5002 msg.msg_control = NULL;
5003 msg.msg_controllen = 0;
5004 msg.msg_namelen = 0;
5005 msg.msg_iocb = NULL;
5008 flags = req->sr_msg.msg_flags;
5009 if (flags & MSG_DONTWAIT)
5010 req->flags |= REQ_F_NOWAIT;
5011 else if (force_nonblock)
5012 flags |= MSG_DONTWAIT;
5014 ret = sock_recvmsg(sock, &msg, flags);
5015 if (force_nonblock && ret == -EAGAIN)
5017 if (ret == -ERESTARTSYS)
5020 if (req->flags & REQ_F_BUFFER_SELECTED)
5021 cflags = io_put_recv_kbuf(req);
5023 req_set_fail_links(req);
5024 __io_req_complete(req, issue_flags, ret, cflags);
5028 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5030 struct io_accept *accept = &req->accept;
5032 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5034 if (sqe->ioprio || sqe->len || sqe->buf_index)
5037 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
5038 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
5039 accept->flags = READ_ONCE(sqe->accept_flags);
5040 accept->nofile = rlimit(RLIMIT_NOFILE);
5044 static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
5046 struct io_accept *accept = &req->accept;
5047 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
5048 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
5051 if (req->file->f_flags & O_NONBLOCK)
5052 req->flags |= REQ_F_NOWAIT;
5054 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
5055 accept->addr_len, accept->flags,
5057 if (ret == -EAGAIN && force_nonblock)
5060 if (ret == -ERESTARTSYS)
5062 req_set_fail_links(req);
5064 __io_req_complete(req, issue_flags, ret, 0);
5068 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5070 struct io_connect *conn = &req->connect;
5071 struct io_async_connect *io = req->async_data;
5073 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5075 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
5078 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
5079 conn->addr_len = READ_ONCE(sqe->addr2);
5084 return move_addr_to_kernel(conn->addr, conn->addr_len,
5088 static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
5090 struct io_async_connect __io, *io;
5091 unsigned file_flags;
5093 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
5095 if (req->async_data) {
5096 io = req->async_data;
5098 ret = move_addr_to_kernel(req->connect.addr,
5099 req->connect.addr_len,
5106 file_flags = force_nonblock ? O_NONBLOCK : 0;
5108 ret = __sys_connect_file(req->file, &io->address,
5109 req->connect.addr_len, file_flags);
5110 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
5111 if (req->async_data)
5113 if (io_alloc_async_data(req)) {
5117 io = req->async_data;
5118 memcpy(req->async_data, &__io, sizeof(__io));
5121 if (ret == -ERESTARTSYS)
5125 req_set_fail_links(req);
5126 __io_req_complete(req, issue_flags, ret, 0);
5129 #else /* !CONFIG_NET */
5130 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5135 static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
5140 static int io_send(struct io_kiocb *req, unsigned int issue_flags)
5145 static int io_recvmsg_prep(struct io_kiocb *req,
5146 const struct io_uring_sqe *sqe)
5151 static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
5156 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
5161 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5166 static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
5171 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5176 static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
5180 #endif /* CONFIG_NET */
5182 struct io_poll_table {
5183 struct poll_table_struct pt;
5184 struct io_kiocb *req;
5188 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
5189 __poll_t mask, task_work_func_t func)
5193 /* for instances that support it check for an event match first: */
5194 if (mask && !(mask & poll->events))
5197 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
5199 list_del_init(&poll->wait.entry);
5202 req->task_work.func = func;
5203 percpu_ref_get(&req->ctx->refs);
5206 * If this fails, then the task is exiting. When a task exits, the
5207 * work gets canceled, so just cancel this request as well instead
5208 * of executing it. We can't safely execute it anyway, as we may not
5209 * have the needed state needed for it anyway.
5211 ret = io_req_task_work_add(req);
5212 if (unlikely(ret)) {
5213 WRITE_ONCE(poll->canceled, true);
5214 io_req_task_work_add_fallback(req, func);
5219 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
5220 __acquires(&req->ctx->completion_lock)
5222 struct io_ring_ctx *ctx = req->ctx;
5224 if (!req->result && !READ_ONCE(poll->canceled)) {
5225 struct poll_table_struct pt = { ._key = poll->events };
5227 req->result = vfs_poll(req->file, &pt) & poll->events;
5230 spin_lock_irq(&ctx->completion_lock);
5231 if (!req->result && !READ_ONCE(poll->canceled)) {
5232 add_wait_queue(poll->head, &poll->wait);
5239 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
5241 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
5242 if (req->opcode == IORING_OP_POLL_ADD)
5243 return req->async_data;
5244 return req->apoll->double_poll;
5247 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
5249 if (req->opcode == IORING_OP_POLL_ADD)
5251 return &req->apoll->poll;
5254 static void io_poll_remove_double(struct io_kiocb *req)
5256 struct io_poll_iocb *poll = io_poll_get_double(req);
5258 lockdep_assert_held(&req->ctx->completion_lock);
5260 if (poll && poll->head) {
5261 struct wait_queue_head *head = poll->head;
5263 spin_lock(&head->lock);
5264 list_del_init(&poll->wait.entry);
5265 if (poll->wait.private)
5266 refcount_dec(&req->refs);
5268 spin_unlock(&head->lock);
5272 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
5274 struct io_ring_ctx *ctx = req->ctx;
5276 io_poll_remove_double(req);
5277 req->poll.done = true;
5278 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
5279 io_commit_cqring(ctx);
5282 static void io_poll_task_func(struct callback_head *cb)
5284 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
5285 struct io_ring_ctx *ctx = req->ctx;
5286 struct io_kiocb *nxt;
5288 if (io_poll_rewait(req, &req->poll)) {
5289 spin_unlock_irq(&ctx->completion_lock);
5291 hash_del(&req->hash_node);
5292 io_poll_complete(req, req->result, 0);
5293 spin_unlock_irq(&ctx->completion_lock);
5295 nxt = io_put_req_find_next(req);
5296 io_cqring_ev_posted(ctx);
5298 __io_req_task_submit(nxt);
5301 percpu_ref_put(&ctx->refs);
5304 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
5305 int sync, void *key)
5307 struct io_kiocb *req = wait->private;
5308 struct io_poll_iocb *poll = io_poll_get_single(req);
5309 __poll_t mask = key_to_poll(key);
5311 /* for instances that support it check for an event match first: */
5312 if (mask && !(mask & poll->events))
5315 list_del_init(&wait->entry);
5317 if (poll && poll->head) {
5320 spin_lock(&poll->head->lock);
5321 done = list_empty(&poll->wait.entry);
5323 list_del_init(&poll->wait.entry);
5324 /* make sure double remove sees this as being gone */
5325 wait->private = NULL;
5326 spin_unlock(&poll->head->lock);
5328 /* use wait func handler, so it matches the rq type */
5329 poll->wait.func(&poll->wait, mode, sync, key);
5332 refcount_dec(&req->refs);
5336 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
5337 wait_queue_func_t wake_func)
5341 poll->canceled = false;
5342 poll->events = events;
5343 INIT_LIST_HEAD(&poll->wait.entry);
5344 init_waitqueue_func_entry(&poll->wait, wake_func);
5347 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5348 struct wait_queue_head *head,
5349 struct io_poll_iocb **poll_ptr)
5351 struct io_kiocb *req = pt->req;
5354 * If poll->head is already set, it's because the file being polled
5355 * uses multiple waitqueues for poll handling (eg one for read, one
5356 * for write). Setup a separate io_poll_iocb if this happens.
5358 if (unlikely(poll->head)) {
5359 struct io_poll_iocb *poll_one = poll;
5361 /* already have a 2nd entry, fail a third attempt */
5363 pt->error = -EINVAL;
5366 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5368 pt->error = -ENOMEM;
5371 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5372 refcount_inc(&req->refs);
5373 poll->wait.private = req;
5380 if (poll->events & EPOLLEXCLUSIVE)
5381 add_wait_queue_exclusive(head, &poll->wait);
5383 add_wait_queue(head, &poll->wait);
5386 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5387 struct poll_table_struct *p)
5389 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5390 struct async_poll *apoll = pt->req->apoll;
5392 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5395 static void io_async_task_func(struct callback_head *cb)
5397 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
5398 struct async_poll *apoll = req->apoll;
5399 struct io_ring_ctx *ctx = req->ctx;
5401 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
5403 if (io_poll_rewait(req, &apoll->poll)) {
5404 spin_unlock_irq(&ctx->completion_lock);
5405 percpu_ref_put(&ctx->refs);
5409 /* If req is still hashed, it cannot have been canceled. Don't check. */
5410 if (hash_hashed(&req->hash_node))
5411 hash_del(&req->hash_node);
5413 io_poll_remove_double(req);
5414 spin_unlock_irq(&ctx->completion_lock);
5416 if (!READ_ONCE(apoll->poll.canceled))
5417 __io_req_task_submit(req);
5419 __io_req_task_cancel(req, -ECANCELED);
5421 percpu_ref_put(&ctx->refs);
5422 kfree(apoll->double_poll);
5426 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5429 struct io_kiocb *req = wait->private;
5430 struct io_poll_iocb *poll = &req->apoll->poll;
5432 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5435 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5438 static void io_poll_req_insert(struct io_kiocb *req)
5440 struct io_ring_ctx *ctx = req->ctx;
5441 struct hlist_head *list;
5443 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5444 hlist_add_head(&req->hash_node, list);
5447 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5448 struct io_poll_iocb *poll,
5449 struct io_poll_table *ipt, __poll_t mask,
5450 wait_queue_func_t wake_func)
5451 __acquires(&ctx->completion_lock)
5453 struct io_ring_ctx *ctx = req->ctx;
5454 bool cancel = false;
5456 INIT_HLIST_NODE(&req->hash_node);
5457 io_init_poll_iocb(poll, mask, wake_func);
5458 poll->file = req->file;
5459 poll->wait.private = req;
5461 ipt->pt._key = mask;
5463 ipt->error = -EINVAL;
5465 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5467 spin_lock_irq(&ctx->completion_lock);
5468 if (likely(poll->head)) {
5469 spin_lock(&poll->head->lock);
5470 if (unlikely(list_empty(&poll->wait.entry))) {
5476 if (mask || ipt->error)
5477 list_del_init(&poll->wait.entry);
5479 WRITE_ONCE(poll->canceled, true);
5480 else if (!poll->done) /* actually waiting for an event */
5481 io_poll_req_insert(req);
5482 spin_unlock(&poll->head->lock);
5488 static bool io_arm_poll_handler(struct io_kiocb *req)
5490 const struct io_op_def *def = &io_op_defs[req->opcode];
5491 struct io_ring_ctx *ctx = req->ctx;
5492 struct async_poll *apoll;
5493 struct io_poll_table ipt;
5497 if (!req->file || !file_can_poll(req->file))
5499 if (req->flags & REQ_F_POLLED)
5503 else if (def->pollout)
5507 /* if we can't nonblock try, then no point in arming a poll handler */
5508 if (!io_file_supports_async(req->file, rw))
5511 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5512 if (unlikely(!apoll))
5514 apoll->double_poll = NULL;
5516 req->flags |= REQ_F_POLLED;
5521 mask |= POLLIN | POLLRDNORM;
5523 mask |= POLLOUT | POLLWRNORM;
5525 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5526 if ((req->opcode == IORING_OP_RECVMSG) &&
5527 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5530 mask |= POLLERR | POLLPRI;
5532 ipt.pt._qproc = io_async_queue_proc;
5534 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5536 if (ret || ipt.error) {
5537 io_poll_remove_double(req);
5538 spin_unlock_irq(&ctx->completion_lock);
5539 kfree(apoll->double_poll);
5543 spin_unlock_irq(&ctx->completion_lock);
5544 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
5545 apoll->poll.events);
5549 static bool __io_poll_remove_one(struct io_kiocb *req,
5550 struct io_poll_iocb *poll)
5552 bool do_complete = false;
5554 spin_lock(&poll->head->lock);
5555 WRITE_ONCE(poll->canceled, true);
5556 if (!list_empty(&poll->wait.entry)) {
5557 list_del_init(&poll->wait.entry);
5560 spin_unlock(&poll->head->lock);
5561 hash_del(&req->hash_node);
5565 static bool io_poll_remove_one(struct io_kiocb *req)
5569 io_poll_remove_double(req);
5571 if (req->opcode == IORING_OP_POLL_ADD) {
5572 do_complete = __io_poll_remove_one(req, &req->poll);
5574 struct async_poll *apoll = req->apoll;
5576 /* non-poll requests have submit ref still */
5577 do_complete = __io_poll_remove_one(req, &apoll->poll);
5580 kfree(apoll->double_poll);
5586 io_cqring_fill_event(req, -ECANCELED);
5587 io_commit_cqring(req->ctx);
5588 req_set_fail_links(req);
5589 io_put_req_deferred(req, 1);
5596 * Returns true if we found and killed one or more poll requests
5598 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5599 struct files_struct *files)
5601 struct hlist_node *tmp;
5602 struct io_kiocb *req;
5605 spin_lock_irq(&ctx->completion_lock);
5606 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5607 struct hlist_head *list;
5609 list = &ctx->cancel_hash[i];
5610 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5611 if (io_match_task(req, tsk, files))
5612 posted += io_poll_remove_one(req);
5615 spin_unlock_irq(&ctx->completion_lock);
5618 io_cqring_ev_posted(ctx);
5623 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
5625 struct hlist_head *list;
5626 struct io_kiocb *req;
5628 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5629 hlist_for_each_entry(req, list, hash_node) {
5630 if (sqe_addr != req->user_data)
5632 if (io_poll_remove_one(req))
5640 static int io_poll_remove_prep(struct io_kiocb *req,
5641 const struct io_uring_sqe *sqe)
5643 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5645 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
5649 req->poll_remove.addr = READ_ONCE(sqe->addr);
5654 * Find a running poll command that matches one specified in sqe->addr,
5655 * and remove it if found.
5657 static int io_poll_remove(struct io_kiocb *req, unsigned int issue_flags)
5659 struct io_ring_ctx *ctx = req->ctx;
5662 spin_lock_irq(&ctx->completion_lock);
5663 ret = io_poll_cancel(ctx, req->poll_remove.addr);
5664 spin_unlock_irq(&ctx->completion_lock);
5667 req_set_fail_links(req);
5668 io_req_complete(req, ret);
5672 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5675 struct io_kiocb *req = wait->private;
5676 struct io_poll_iocb *poll = &req->poll;
5678 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5681 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5682 struct poll_table_struct *p)
5684 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5686 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5689 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5691 struct io_poll_iocb *poll = &req->poll;
5694 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5696 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
5699 events = READ_ONCE(sqe->poll32_events);
5701 events = swahw32(events);
5703 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP |
5704 (events & EPOLLEXCLUSIVE);
5708 static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5710 struct io_poll_iocb *poll = &req->poll;
5711 struct io_ring_ctx *ctx = req->ctx;
5712 struct io_poll_table ipt;
5715 ipt.pt._qproc = io_poll_queue_proc;
5717 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5720 if (mask) { /* no async, we'd stolen it */
5722 io_poll_complete(req, mask, 0);
5724 spin_unlock_irq(&ctx->completion_lock);
5727 io_cqring_ev_posted(ctx);
5733 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5735 struct io_timeout_data *data = container_of(timer,
5736 struct io_timeout_data, timer);
5737 struct io_kiocb *req = data->req;
5738 struct io_ring_ctx *ctx = req->ctx;
5739 unsigned long flags;
5741 spin_lock_irqsave(&ctx->completion_lock, flags);
5742 list_del_init(&req->timeout.list);
5743 atomic_set(&req->ctx->cq_timeouts,
5744 atomic_read(&req->ctx->cq_timeouts) + 1);
5746 io_cqring_fill_event(req, -ETIME);
5747 io_commit_cqring(ctx);
5748 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5750 io_cqring_ev_posted(ctx);
5751 req_set_fail_links(req);
5753 return HRTIMER_NORESTART;
5756 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5759 struct io_timeout_data *io;
5760 struct io_kiocb *req;
5763 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5764 if (user_data == req->user_data) {
5771 return ERR_PTR(ret);
5773 io = req->async_data;
5774 ret = hrtimer_try_to_cancel(&io->timer);
5776 return ERR_PTR(-EALREADY);
5777 list_del_init(&req->timeout.list);
5781 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5783 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5786 return PTR_ERR(req);
5788 req_set_fail_links(req);
5789 io_cqring_fill_event(req, -ECANCELED);
5790 io_put_req_deferred(req, 1);
5794 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5795 struct timespec64 *ts, enum hrtimer_mode mode)
5797 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5798 struct io_timeout_data *data;
5801 return PTR_ERR(req);
5803 req->timeout.off = 0; /* noseq */
5804 data = req->async_data;
5805 list_add_tail(&req->timeout.list, &ctx->timeout_list);
5806 hrtimer_init(&data->timer, CLOCK_MONOTONIC, mode);
5807 data->timer.function = io_timeout_fn;
5808 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
5812 static int io_timeout_remove_prep(struct io_kiocb *req,
5813 const struct io_uring_sqe *sqe)
5815 struct io_timeout_rem *tr = &req->timeout_rem;
5817 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5819 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5821 if (sqe->ioprio || sqe->buf_index || sqe->len)
5824 tr->addr = READ_ONCE(sqe->addr);
5825 tr->flags = READ_ONCE(sqe->timeout_flags);
5826 if (tr->flags & IORING_TIMEOUT_UPDATE) {
5827 if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS))
5829 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
5831 } else if (tr->flags) {
5832 /* timeout removal doesn't support flags */
5839 static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
5841 return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
5846 * Remove or update an existing timeout command
5848 static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
5850 struct io_timeout_rem *tr = &req->timeout_rem;
5851 struct io_ring_ctx *ctx = req->ctx;
5854 spin_lock_irq(&ctx->completion_lock);
5855 if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE))
5856 ret = io_timeout_cancel(ctx, tr->addr);
5858 ret = io_timeout_update(ctx, tr->addr, &tr->ts,
5859 io_translate_timeout_mode(tr->flags));
5861 io_cqring_fill_event(req, ret);
5862 io_commit_cqring(ctx);
5863 spin_unlock_irq(&ctx->completion_lock);
5864 io_cqring_ev_posted(ctx);
5866 req_set_fail_links(req);
5871 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5872 bool is_timeout_link)
5874 struct io_timeout_data *data;
5876 u32 off = READ_ONCE(sqe->off);
5878 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5880 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5882 if (off && is_timeout_link)
5884 flags = READ_ONCE(sqe->timeout_flags);
5885 if (flags & ~IORING_TIMEOUT_ABS)
5888 req->timeout.off = off;
5890 if (!req->async_data && io_alloc_async_data(req))
5893 data = req->async_data;
5896 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5899 data->mode = io_translate_timeout_mode(flags);
5900 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5904 static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
5906 struct io_ring_ctx *ctx = req->ctx;
5907 struct io_timeout_data *data = req->async_data;
5908 struct list_head *entry;
5909 u32 tail, off = req->timeout.off;
5911 spin_lock_irq(&ctx->completion_lock);
5914 * sqe->off holds how many events that need to occur for this
5915 * timeout event to be satisfied. If it isn't set, then this is
5916 * a pure timeout request, sequence isn't used.
5918 if (io_is_timeout_noseq(req)) {
5919 entry = ctx->timeout_list.prev;
5923 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5924 req->timeout.target_seq = tail + off;
5926 /* Update the last seq here in case io_flush_timeouts() hasn't.
5927 * This is safe because ->completion_lock is held, and submissions
5928 * and completions are never mixed in the same ->completion_lock section.
5930 ctx->cq_last_tm_flush = tail;
5933 * Insertion sort, ensuring the first entry in the list is always
5934 * the one we need first.
5936 list_for_each_prev(entry, &ctx->timeout_list) {
5937 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5940 if (io_is_timeout_noseq(nxt))
5942 /* nxt.seq is behind @tail, otherwise would've been completed */
5943 if (off >= nxt->timeout.target_seq - tail)
5947 list_add(&req->timeout.list, entry);
5948 data->timer.function = io_timeout_fn;
5949 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5950 spin_unlock_irq(&ctx->completion_lock);
5954 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5956 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5958 return req->user_data == (unsigned long) data;
5961 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
5963 enum io_wq_cancel cancel_ret;
5966 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr, false);
5967 switch (cancel_ret) {
5968 case IO_WQ_CANCEL_OK:
5971 case IO_WQ_CANCEL_RUNNING:
5974 case IO_WQ_CANCEL_NOTFOUND:
5982 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5983 struct io_kiocb *req, __u64 sqe_addr,
5986 unsigned long flags;
5989 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
5990 if (ret != -ENOENT) {
5991 spin_lock_irqsave(&ctx->completion_lock, flags);
5995 spin_lock_irqsave(&ctx->completion_lock, flags);
5996 ret = io_timeout_cancel(ctx, sqe_addr);
5999 ret = io_poll_cancel(ctx, sqe_addr);
6003 io_cqring_fill_event(req, ret);
6004 io_commit_cqring(ctx);
6005 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6006 io_cqring_ev_posted(ctx);
6009 req_set_fail_links(req);
6013 static int io_async_cancel_prep(struct io_kiocb *req,
6014 const struct io_uring_sqe *sqe)
6016 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
6018 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
6020 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
6023 req->cancel.addr = READ_ONCE(sqe->addr);
6027 static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
6029 struct io_ring_ctx *ctx = req->ctx;
6031 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
6035 static int io_rsrc_update_prep(struct io_kiocb *req,
6036 const struct io_uring_sqe *sqe)
6038 if (unlikely(req->ctx->flags & IORING_SETUP_SQPOLL))
6040 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
6042 if (sqe->ioprio || sqe->rw_flags)
6045 req->rsrc_update.offset = READ_ONCE(sqe->off);
6046 req->rsrc_update.nr_args = READ_ONCE(sqe->len);
6047 if (!req->rsrc_update.nr_args)
6049 req->rsrc_update.arg = READ_ONCE(sqe->addr);
6053 static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
6055 struct io_ring_ctx *ctx = req->ctx;
6056 struct io_uring_rsrc_update up;
6059 if (issue_flags & IO_URING_F_NONBLOCK)
6062 up.offset = req->rsrc_update.offset;
6063 up.data = req->rsrc_update.arg;
6065 mutex_lock(&ctx->uring_lock);
6066 ret = __io_sqe_files_update(ctx, &up, req->rsrc_update.nr_args);
6067 mutex_unlock(&ctx->uring_lock);
6070 req_set_fail_links(req);
6071 __io_req_complete(req, issue_flags, ret, 0);
6075 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
6077 switch (req->opcode) {
6080 case IORING_OP_READV:
6081 case IORING_OP_READ_FIXED:
6082 case IORING_OP_READ:
6083 return io_read_prep(req, sqe);
6084 case IORING_OP_WRITEV:
6085 case IORING_OP_WRITE_FIXED:
6086 case IORING_OP_WRITE:
6087 return io_write_prep(req, sqe);
6088 case IORING_OP_POLL_ADD:
6089 return io_poll_add_prep(req, sqe);
6090 case IORING_OP_POLL_REMOVE:
6091 return io_poll_remove_prep(req, sqe);
6092 case IORING_OP_FSYNC:
6093 return io_prep_fsync(req, sqe);
6094 case IORING_OP_SYNC_FILE_RANGE:
6095 return io_prep_sfr(req, sqe);
6096 case IORING_OP_SENDMSG:
6097 case IORING_OP_SEND:
6098 return io_sendmsg_prep(req, sqe);
6099 case IORING_OP_RECVMSG:
6100 case IORING_OP_RECV:
6101 return io_recvmsg_prep(req, sqe);
6102 case IORING_OP_CONNECT:
6103 return io_connect_prep(req, sqe);
6104 case IORING_OP_TIMEOUT:
6105 return io_timeout_prep(req, sqe, false);
6106 case IORING_OP_TIMEOUT_REMOVE:
6107 return io_timeout_remove_prep(req, sqe);
6108 case IORING_OP_ASYNC_CANCEL:
6109 return io_async_cancel_prep(req, sqe);
6110 case IORING_OP_LINK_TIMEOUT:
6111 return io_timeout_prep(req, sqe, true);
6112 case IORING_OP_ACCEPT:
6113 return io_accept_prep(req, sqe);
6114 case IORING_OP_FALLOCATE:
6115 return io_fallocate_prep(req, sqe);
6116 case IORING_OP_OPENAT:
6117 return io_openat_prep(req, sqe);
6118 case IORING_OP_CLOSE:
6119 return io_close_prep(req, sqe);
6120 case IORING_OP_FILES_UPDATE:
6121 return io_rsrc_update_prep(req, sqe);
6122 case IORING_OP_STATX:
6123 return io_statx_prep(req, sqe);
6124 case IORING_OP_FADVISE:
6125 return io_fadvise_prep(req, sqe);
6126 case IORING_OP_MADVISE:
6127 return io_madvise_prep(req, sqe);
6128 case IORING_OP_OPENAT2:
6129 return io_openat2_prep(req, sqe);
6130 case IORING_OP_EPOLL_CTL:
6131 return io_epoll_ctl_prep(req, sqe);
6132 case IORING_OP_SPLICE:
6133 return io_splice_prep(req, sqe);
6134 case IORING_OP_PROVIDE_BUFFERS:
6135 return io_provide_buffers_prep(req, sqe);
6136 case IORING_OP_REMOVE_BUFFERS:
6137 return io_remove_buffers_prep(req, sqe);
6139 return io_tee_prep(req, sqe);
6140 case IORING_OP_SHUTDOWN:
6141 return io_shutdown_prep(req, sqe);
6142 case IORING_OP_RENAMEAT:
6143 return io_renameat_prep(req, sqe);
6144 case IORING_OP_UNLINKAT:
6145 return io_unlinkat_prep(req, sqe);
6148 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
6153 static int io_req_defer_prep(struct io_kiocb *req,
6154 const struct io_uring_sqe *sqe)
6158 if (io_alloc_async_data(req))
6160 return io_req_prep(req, sqe);
6163 static u32 io_get_sequence(struct io_kiocb *req)
6165 struct io_kiocb *pos;
6166 struct io_ring_ctx *ctx = req->ctx;
6167 u32 total_submitted, nr_reqs = 0;
6169 io_for_each_link(pos, req)
6172 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
6173 return total_submitted - nr_reqs;
6176 static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe)
6178 struct io_ring_ctx *ctx = req->ctx;
6179 struct io_defer_entry *de;
6183 /* Still need defer if there is pending req in defer list. */
6184 if (likely(list_empty_careful(&ctx->defer_list) &&
6185 !(req->flags & REQ_F_IO_DRAIN)))
6188 seq = io_get_sequence(req);
6189 /* Still a chance to pass the sequence check */
6190 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
6193 if (!req->async_data) {
6194 ret = io_req_defer_prep(req, sqe);
6198 io_prep_async_link(req);
6199 de = kmalloc(sizeof(*de), GFP_KERNEL);
6203 spin_lock_irq(&ctx->completion_lock);
6204 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
6205 spin_unlock_irq(&ctx->completion_lock);
6207 io_queue_async_work(req);
6208 return -EIOCBQUEUED;
6211 trace_io_uring_defer(ctx, req, req->user_data);
6214 list_add_tail(&de->list, &ctx->defer_list);
6215 spin_unlock_irq(&ctx->completion_lock);
6216 return -EIOCBQUEUED;
6219 static void __io_clean_op(struct io_kiocb *req)
6221 if (req->flags & REQ_F_BUFFER_SELECTED) {
6222 switch (req->opcode) {
6223 case IORING_OP_READV:
6224 case IORING_OP_READ_FIXED:
6225 case IORING_OP_READ:
6226 kfree((void *)(unsigned long)req->rw.addr);
6228 case IORING_OP_RECVMSG:
6229 case IORING_OP_RECV:
6230 kfree(req->sr_msg.kbuf);
6233 req->flags &= ~REQ_F_BUFFER_SELECTED;
6236 if (req->flags & REQ_F_NEED_CLEANUP) {
6237 switch (req->opcode) {
6238 case IORING_OP_READV:
6239 case IORING_OP_READ_FIXED:
6240 case IORING_OP_READ:
6241 case IORING_OP_WRITEV:
6242 case IORING_OP_WRITE_FIXED:
6243 case IORING_OP_WRITE: {
6244 struct io_async_rw *io = req->async_data;
6246 kfree(io->free_iovec);
6249 case IORING_OP_RECVMSG:
6250 case IORING_OP_SENDMSG: {
6251 struct io_async_msghdr *io = req->async_data;
6253 kfree(io->free_iov);
6256 case IORING_OP_SPLICE:
6258 io_put_file(req, req->splice.file_in,
6259 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
6261 case IORING_OP_OPENAT:
6262 case IORING_OP_OPENAT2:
6263 if (req->open.filename)
6264 putname(req->open.filename);
6266 case IORING_OP_RENAMEAT:
6267 putname(req->rename.oldpath);
6268 putname(req->rename.newpath);
6270 case IORING_OP_UNLINKAT:
6271 putname(req->unlink.filename);
6274 req->flags &= ~REQ_F_NEED_CLEANUP;
6278 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
6280 struct io_ring_ctx *ctx = req->ctx;
6283 switch (req->opcode) {
6285 ret = io_nop(req, issue_flags);
6287 case IORING_OP_READV:
6288 case IORING_OP_READ_FIXED:
6289 case IORING_OP_READ:
6290 ret = io_read(req, issue_flags);
6292 case IORING_OP_WRITEV:
6293 case IORING_OP_WRITE_FIXED:
6294 case IORING_OP_WRITE:
6295 ret = io_write(req, issue_flags);
6297 case IORING_OP_FSYNC:
6298 ret = io_fsync(req, issue_flags);
6300 case IORING_OP_POLL_ADD:
6301 ret = io_poll_add(req, issue_flags);
6303 case IORING_OP_POLL_REMOVE:
6304 ret = io_poll_remove(req, issue_flags);
6306 case IORING_OP_SYNC_FILE_RANGE:
6307 ret = io_sync_file_range(req, issue_flags);
6309 case IORING_OP_SENDMSG:
6310 ret = io_sendmsg(req, issue_flags);
6312 case IORING_OP_SEND:
6313 ret = io_send(req, issue_flags);
6315 case IORING_OP_RECVMSG:
6316 ret = io_recvmsg(req, issue_flags);
6318 case IORING_OP_RECV:
6319 ret = io_recv(req, issue_flags);
6321 case IORING_OP_TIMEOUT:
6322 ret = io_timeout(req, issue_flags);
6324 case IORING_OP_TIMEOUT_REMOVE:
6325 ret = io_timeout_remove(req, issue_flags);
6327 case IORING_OP_ACCEPT:
6328 ret = io_accept(req, issue_flags);
6330 case IORING_OP_CONNECT:
6331 ret = io_connect(req, issue_flags);
6333 case IORING_OP_ASYNC_CANCEL:
6334 ret = io_async_cancel(req, issue_flags);
6336 case IORING_OP_FALLOCATE:
6337 ret = io_fallocate(req, issue_flags);
6339 case IORING_OP_OPENAT:
6340 ret = io_openat(req, issue_flags);
6342 case IORING_OP_CLOSE:
6343 ret = io_close(req, issue_flags);
6345 case IORING_OP_FILES_UPDATE:
6346 ret = io_files_update(req, issue_flags);
6348 case IORING_OP_STATX:
6349 ret = io_statx(req, issue_flags);
6351 case IORING_OP_FADVISE:
6352 ret = io_fadvise(req, issue_flags);
6354 case IORING_OP_MADVISE:
6355 ret = io_madvise(req, issue_flags);
6357 case IORING_OP_OPENAT2:
6358 ret = io_openat2(req, issue_flags);
6360 case IORING_OP_EPOLL_CTL:
6361 ret = io_epoll_ctl(req, issue_flags);
6363 case IORING_OP_SPLICE:
6364 ret = io_splice(req, issue_flags);
6366 case IORING_OP_PROVIDE_BUFFERS:
6367 ret = io_provide_buffers(req, issue_flags);
6369 case IORING_OP_REMOVE_BUFFERS:
6370 ret = io_remove_buffers(req, issue_flags);
6373 ret = io_tee(req, issue_flags);
6375 case IORING_OP_SHUTDOWN:
6376 ret = io_shutdown(req, issue_flags);
6378 case IORING_OP_RENAMEAT:
6379 ret = io_renameat(req, issue_flags);
6381 case IORING_OP_UNLINKAT:
6382 ret = io_unlinkat(req, issue_flags);
6392 /* If the op doesn't have a file, we're not polling for it */
6393 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
6394 const bool in_async = io_wq_current_is_worker();
6396 /* workqueue context doesn't hold uring_lock, grab it now */
6398 mutex_lock(&ctx->uring_lock);
6400 io_iopoll_req_issued(req, in_async);
6403 mutex_unlock(&ctx->uring_lock);
6409 static void io_wq_submit_work(struct io_wq_work *work)
6411 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6412 struct io_kiocb *timeout;
6415 timeout = io_prep_linked_timeout(req);
6417 io_queue_linked_timeout(timeout);
6419 if (work->flags & IO_WQ_WORK_CANCEL)
6424 ret = io_issue_sqe(req, 0);
6426 * We can get EAGAIN for polled IO even though we're
6427 * forcing a sync submission from here, since we can't
6428 * wait for request slots on the block side.
6437 struct io_ring_ctx *lock_ctx = NULL;
6439 if (req->ctx->flags & IORING_SETUP_IOPOLL)
6440 lock_ctx = req->ctx;
6443 * io_iopoll_complete() does not hold completion_lock to
6444 * complete polled io, so here for polled io, we can not call
6445 * io_req_complete() directly, otherwise there maybe concurrent
6446 * access to cqring, defer_list, etc, which is not safe. Given
6447 * that io_iopoll_complete() is always called under uring_lock,
6448 * so here for polled io, we also get uring_lock to complete
6452 mutex_lock(&lock_ctx->uring_lock);
6454 req_set_fail_links(req);
6455 io_req_complete(req, ret);
6458 mutex_unlock(&lock_ctx->uring_lock);
6462 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6465 struct fixed_rsrc_table *table;
6467 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
6468 return table->files[index & IORING_FILE_TABLE_MASK];
6471 static struct file *io_file_get(struct io_submit_state *state,
6472 struct io_kiocb *req, int fd, bool fixed)
6474 struct io_ring_ctx *ctx = req->ctx;
6478 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6480 fd = array_index_nospec(fd, ctx->nr_user_files);
6481 file = io_file_from_index(ctx, fd);
6482 io_set_resource_node(req);
6484 trace_io_uring_file_get(ctx, fd);
6485 file = __io_file_get(state, fd);
6488 if (file && unlikely(file->f_op == &io_uring_fops))
6489 io_req_track_inflight(req);
6493 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6495 struct io_timeout_data *data = container_of(timer,
6496 struct io_timeout_data, timer);
6497 struct io_kiocb *prev, *req = data->req;
6498 struct io_ring_ctx *ctx = req->ctx;
6499 unsigned long flags;
6501 spin_lock_irqsave(&ctx->completion_lock, flags);
6502 prev = req->timeout.head;
6503 req->timeout.head = NULL;
6506 * We don't expect the list to be empty, that will only happen if we
6507 * race with the completion of the linked work.
6509 if (prev && refcount_inc_not_zero(&prev->refs))
6510 io_remove_next_linked(prev);
6513 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6516 req_set_fail_links(prev);
6517 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6518 io_put_req_deferred(prev, 1);
6520 io_req_complete_post(req, -ETIME, 0);
6521 io_put_req_deferred(req, 1);
6523 return HRTIMER_NORESTART;
6526 static void __io_queue_linked_timeout(struct io_kiocb *req)
6529 * If the back reference is NULL, then our linked request finished
6530 * before we got a chance to setup the timer
6532 if (req->timeout.head) {
6533 struct io_timeout_data *data = req->async_data;
6535 data->timer.function = io_link_timeout_fn;
6536 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6541 static void io_queue_linked_timeout(struct io_kiocb *req)
6543 struct io_ring_ctx *ctx = req->ctx;
6545 spin_lock_irq(&ctx->completion_lock);
6546 __io_queue_linked_timeout(req);
6547 spin_unlock_irq(&ctx->completion_lock);
6549 /* drop submission reference */
6553 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6555 struct io_kiocb *nxt = req->link;
6557 if (!nxt || (req->flags & REQ_F_LINK_TIMEOUT) ||
6558 nxt->opcode != IORING_OP_LINK_TIMEOUT)
6561 nxt->timeout.head = req;
6562 nxt->flags |= REQ_F_LTIMEOUT_ACTIVE;
6563 req->flags |= REQ_F_LINK_TIMEOUT;
6567 static void __io_queue_sqe(struct io_kiocb *req)
6569 struct io_kiocb *linked_timeout;
6570 const struct cred *old_creds = NULL;
6574 linked_timeout = io_prep_linked_timeout(req);
6576 if ((req->flags & REQ_F_WORK_INITIALIZED) &&
6577 (req->work.flags & IO_WQ_WORK_CREDS) &&
6578 req->work.identity->creds != current_cred()) {
6580 revert_creds(old_creds);
6581 if (old_creds == req->work.identity->creds)
6582 old_creds = NULL; /* restored original creds */
6584 old_creds = override_creds(req->work.identity->creds);
6587 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
6590 * We async punt it if the file wasn't marked NOWAIT, or if the file
6591 * doesn't support non-blocking read/write attempts
6593 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6594 if (!io_arm_poll_handler(req)) {
6596 * Queued up for async execution, worker will release
6597 * submit reference when the iocb is actually submitted.
6599 io_queue_async_work(req);
6603 io_queue_linked_timeout(linked_timeout);
6604 } else if (likely(!ret)) {
6605 /* drop submission reference */
6606 if (req->flags & REQ_F_COMPLETE_INLINE) {
6607 struct io_ring_ctx *ctx = req->ctx;
6608 struct io_comp_state *cs = &ctx->submit_state.comp;
6610 cs->reqs[cs->nr++] = req;
6611 if (cs->nr == IO_COMPL_BATCH)
6612 io_submit_flush_completions(cs, ctx);
6615 req = io_put_req_find_next(req);
6619 io_queue_linked_timeout(linked_timeout);
6622 if (!(req->flags & REQ_F_FORCE_ASYNC))
6624 io_queue_async_work(req);
6627 /* un-prep timeout, so it'll be killed as any other linked */
6628 req->flags &= ~REQ_F_LINK_TIMEOUT;
6629 req_set_fail_links(req);
6631 io_req_complete(req, ret);
6635 revert_creds(old_creds);
6638 static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe)
6642 ret = io_req_defer(req, sqe);
6644 if (ret != -EIOCBQUEUED) {
6646 req_set_fail_links(req);
6648 io_req_complete(req, ret);
6650 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6651 if (!req->async_data) {
6652 ret = io_req_defer_prep(req, sqe);
6656 io_queue_async_work(req);
6659 ret = io_req_prep(req, sqe);
6663 __io_queue_sqe(req);
6667 static inline void io_queue_link_head(struct io_kiocb *req)
6669 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
6671 io_req_complete(req, -ECANCELED);
6673 io_queue_sqe(req, NULL);
6676 struct io_submit_link {
6677 struct io_kiocb *head;
6678 struct io_kiocb *last;
6681 static int io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6682 struct io_submit_link *link)
6684 struct io_ring_ctx *ctx = req->ctx;
6688 * If we already have a head request, queue this one for async
6689 * submittal once the head completes. If we don't have a head but
6690 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6691 * submitted sync once the chain is complete. If none of those
6692 * conditions are true (normal request), then just queue it.
6695 struct io_kiocb *head = link->head;
6698 * Taking sequential execution of a link, draining both sides
6699 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6700 * requests in the link. So, it drains the head and the
6701 * next after the link request. The last one is done via
6702 * drain_next flag to persist the effect across calls.
6704 if (req->flags & REQ_F_IO_DRAIN) {
6705 head->flags |= REQ_F_IO_DRAIN;
6706 ctx->drain_next = 1;
6708 ret = io_req_defer_prep(req, sqe);
6709 if (unlikely(ret)) {
6710 /* fail even hard links since we don't submit */
6711 head->flags |= REQ_F_FAIL_LINK;
6714 trace_io_uring_link(ctx, req, head);
6715 link->last->link = req;
6718 /* last request of a link, enqueue the link */
6719 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6720 io_queue_link_head(head);
6724 if (unlikely(ctx->drain_next)) {
6725 req->flags |= REQ_F_IO_DRAIN;
6726 ctx->drain_next = 0;
6728 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6729 ret = io_req_defer_prep(req, sqe);
6731 req->flags |= REQ_F_FAIL_LINK;
6735 io_queue_sqe(req, sqe);
6743 * Batched submission is done, ensure local IO is flushed out.
6745 static void io_submit_state_end(struct io_submit_state *state,
6746 struct io_ring_ctx *ctx)
6749 io_submit_flush_completions(&state->comp, ctx);
6750 if (state->plug_started)
6751 blk_finish_plug(&state->plug);
6752 io_state_file_put(state);
6756 * Start submission side cache.
6758 static void io_submit_state_start(struct io_submit_state *state,
6759 unsigned int max_ios)
6761 state->plug_started = false;
6762 state->ios_left = max_ios;
6765 static void io_commit_sqring(struct io_ring_ctx *ctx)
6767 struct io_rings *rings = ctx->rings;
6770 * Ensure any loads from the SQEs are done at this point,
6771 * since once we write the new head, the application could
6772 * write new data to them.
6774 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6778 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6779 * that is mapped by userspace. This means that care needs to be taken to
6780 * ensure that reads are stable, as we cannot rely on userspace always
6781 * being a good citizen. If members of the sqe are validated and then later
6782 * used, it's important that those reads are done through READ_ONCE() to
6783 * prevent a re-load down the line.
6785 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6787 u32 *sq_array = ctx->sq_array;
6791 * The cached sq head (or cq tail) serves two purposes:
6793 * 1) allows us to batch the cost of updating the user visible
6795 * 2) allows the kernel side to track the head on its own, even
6796 * though the application is the one updating it.
6798 head = READ_ONCE(sq_array[ctx->cached_sq_head & ctx->sq_mask]);
6799 if (likely(head < ctx->sq_entries))
6800 return &ctx->sq_sqes[head];
6802 /* drop invalid entries */
6803 ctx->cached_sq_dropped++;
6804 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6808 static inline void io_consume_sqe(struct io_ring_ctx *ctx)
6810 ctx->cached_sq_head++;
6814 * Check SQE restrictions (opcode and flags).
6816 * Returns 'true' if SQE is allowed, 'false' otherwise.
6818 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6819 struct io_kiocb *req,
6820 unsigned int sqe_flags)
6822 if (!ctx->restricted)
6825 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6828 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6829 ctx->restrictions.sqe_flags_required)
6832 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6833 ctx->restrictions.sqe_flags_required))
6839 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
6840 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
6841 IOSQE_BUFFER_SELECT)
6843 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6844 const struct io_uring_sqe *sqe)
6846 struct io_submit_state *state;
6847 unsigned int sqe_flags;
6850 req->opcode = READ_ONCE(sqe->opcode);
6851 req->user_data = READ_ONCE(sqe->user_data);
6852 req->async_data = NULL;
6857 req->fixed_rsrc_refs = NULL;
6858 /* one is dropped after submission, the other at completion */
6859 refcount_set(&req->refs, 2);
6860 req->task = current;
6863 if (unlikely(req->opcode >= IORING_OP_LAST))
6866 if (unlikely(io_sq_thread_acquire_mm_files(ctx, req)))
6869 sqe_flags = READ_ONCE(sqe->flags);
6870 /* enforce forwards compatibility on users */
6871 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6874 if (unlikely(!io_check_restriction(ctx, req, sqe_flags)))
6877 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6878 !io_op_defs[req->opcode].buffer_select)
6881 id = READ_ONCE(sqe->personality);
6883 struct io_identity *iod;
6885 iod = idr_find(&ctx->personality_idr, id);
6888 refcount_inc(&iod->count);
6890 __io_req_init_async(req);
6891 get_cred(iod->creds);
6892 req->work.identity = iod;
6893 req->work.flags |= IO_WQ_WORK_CREDS;
6896 /* same numerical values with corresponding REQ_F_*, safe to copy */
6897 req->flags |= sqe_flags;
6898 state = &ctx->submit_state;
6901 * Plug now if we have more than 1 IO left after this, and the target
6902 * is potentially a read/write to block based storage.
6904 if (!state->plug_started && state->ios_left > 1 &&
6905 io_op_defs[req->opcode].plug) {
6906 blk_start_plug(&state->plug);
6907 state->plug_started = true;
6911 if (io_op_defs[req->opcode].needs_file) {
6912 bool fixed = req->flags & REQ_F_FIXED_FILE;
6914 req->file = io_file_get(state, req, READ_ONCE(sqe->fd), fixed);
6915 if (unlikely(!req->file))
6923 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6925 struct io_submit_link link;
6926 int i, submitted = 0;
6928 /* if we have a backlog and couldn't flush it all, return BUSY */
6929 if (test_bit(0, &ctx->sq_check_overflow)) {
6930 if (!__io_cqring_overflow_flush(ctx, false, NULL, NULL))
6934 /* make sure SQ entry isn't read before tail */
6935 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6937 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6940 percpu_counter_add(¤t->io_uring->inflight, nr);
6941 refcount_add(nr, ¤t->usage);
6943 io_submit_state_start(&ctx->submit_state, nr);
6946 for (i = 0; i < nr; i++) {
6947 const struct io_uring_sqe *sqe;
6948 struct io_kiocb *req;
6951 sqe = io_get_sqe(ctx);
6952 if (unlikely(!sqe)) {
6953 io_consume_sqe(ctx);
6956 req = io_alloc_req(ctx);
6957 if (unlikely(!req)) {
6959 submitted = -EAGAIN;
6962 io_consume_sqe(ctx);
6963 /* will complete beyond this point, count as submitted */
6966 err = io_init_req(ctx, req, sqe);
6967 if (unlikely(err)) {
6970 io_req_complete(req, err);
6974 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6975 true, ctx->flags & IORING_SETUP_SQPOLL);
6976 err = io_submit_sqe(req, sqe, &link);
6981 if (unlikely(submitted != nr)) {
6982 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6983 struct io_uring_task *tctx = current->io_uring;
6984 int unused = nr - ref_used;
6986 percpu_ref_put_many(&ctx->refs, unused);
6987 percpu_counter_sub(&tctx->inflight, unused);
6988 put_task_struct_many(current, unused);
6991 io_queue_link_head(link.head);
6992 io_submit_state_end(&ctx->submit_state, ctx);
6994 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6995 io_commit_sqring(ctx);
7000 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
7002 /* Tell userspace we may need a wakeup call */
7003 spin_lock_irq(&ctx->completion_lock);
7004 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
7005 spin_unlock_irq(&ctx->completion_lock);
7008 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
7010 spin_lock_irq(&ctx->completion_lock);
7011 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
7012 spin_unlock_irq(&ctx->completion_lock);
7015 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
7017 unsigned int to_submit;
7020 to_submit = io_sqring_entries(ctx);
7021 /* if we're handling multiple rings, cap submit size for fairness */
7022 if (cap_entries && to_submit > 8)
7025 if (!list_empty(&ctx->iopoll_list) || to_submit) {
7026 unsigned nr_events = 0;
7028 mutex_lock(&ctx->uring_lock);
7029 if (!list_empty(&ctx->iopoll_list))
7030 io_do_iopoll(ctx, &nr_events, 0);
7032 if (to_submit && !ctx->sqo_dead &&
7033 likely(!percpu_ref_is_dying(&ctx->refs)))
7034 ret = io_submit_sqes(ctx, to_submit);
7035 mutex_unlock(&ctx->uring_lock);
7038 if (!io_sqring_full(ctx) && wq_has_sleeper(&ctx->sqo_sq_wait))
7039 wake_up(&ctx->sqo_sq_wait);
7044 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
7046 struct io_ring_ctx *ctx;
7047 unsigned sq_thread_idle = 0;
7049 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
7050 if (sq_thread_idle < ctx->sq_thread_idle)
7051 sq_thread_idle = ctx->sq_thread_idle;
7054 sqd->sq_thread_idle = sq_thread_idle;
7057 static void io_sqd_init_new(struct io_sq_data *sqd)
7059 struct io_ring_ctx *ctx;
7061 while (!list_empty(&sqd->ctx_new_list)) {
7062 ctx = list_first_entry(&sqd->ctx_new_list, struct io_ring_ctx, sqd_list);
7063 list_move_tail(&ctx->sqd_list, &sqd->ctx_list);
7064 complete(&ctx->sq_thread_comp);
7067 io_sqd_update_thread_idle(sqd);
7070 static int io_sq_thread(void *data)
7072 struct cgroup_subsys_state *cur_css = NULL;
7073 struct files_struct *old_files = current->files;
7074 struct nsproxy *old_nsproxy = current->nsproxy;
7075 const struct cred *old_cred = NULL;
7076 struct io_sq_data *sqd = data;
7077 struct io_ring_ctx *ctx;
7078 unsigned long timeout = 0;
7082 current->files = NULL;
7083 current->nsproxy = NULL;
7084 task_unlock(current);
7086 while (!kthread_should_stop()) {
7088 bool cap_entries, sqt_spin, needs_sched;
7091 * Any changes to the sqd lists are synchronized through the
7092 * kthread parking. This synchronizes the thread vs users,
7093 * the users are synchronized on the sqd->ctx_lock.
7095 if (kthread_should_park()) {
7098 * When sq thread is unparked, in case the previous park operation
7099 * comes from io_put_sq_data(), which means that sq thread is going
7100 * to be stopped, so here needs to have a check.
7102 if (kthread_should_stop())
7106 if (unlikely(!list_empty(&sqd->ctx_new_list))) {
7107 io_sqd_init_new(sqd);
7108 timeout = jiffies + sqd->sq_thread_idle;
7112 cap_entries = !list_is_singular(&sqd->ctx_list);
7113 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
7114 if (current->cred != ctx->creds) {
7116 revert_creds(old_cred);
7117 old_cred = override_creds(ctx->creds);
7119 io_sq_thread_associate_blkcg(ctx, &cur_css);
7121 current->loginuid = ctx->loginuid;
7122 current->sessionid = ctx->sessionid;
7125 ret = __io_sq_thread(ctx, cap_entries);
7126 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
7129 io_sq_thread_drop_mm_files();
7132 if (sqt_spin || !time_after(jiffies, timeout)) {
7134 io_sq_thread_drop_mm_files();
7137 timeout = jiffies + sqd->sq_thread_idle;
7142 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
7143 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
7144 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
7145 !list_empty_careful(&ctx->iopoll_list)) {
7146 needs_sched = false;
7149 if (io_sqring_entries(ctx)) {
7150 needs_sched = false;
7155 if (needs_sched && !kthread_should_park()) {
7156 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7157 io_ring_set_wakeup_flag(ctx);
7160 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7161 io_ring_clear_wakeup_flag(ctx);
7164 finish_wait(&sqd->wait, &wait);
7165 timeout = jiffies + sqd->sq_thread_idle;
7169 io_sq_thread_drop_mm_files();
7172 io_sq_thread_unassociate_blkcg();
7174 revert_creds(old_cred);
7177 current->files = old_files;
7178 current->nsproxy = old_nsproxy;
7179 task_unlock(current);
7186 struct io_wait_queue {
7187 struct wait_queue_entry wq;
7188 struct io_ring_ctx *ctx;
7190 unsigned nr_timeouts;
7193 static inline bool io_should_wake(struct io_wait_queue *iowq)
7195 struct io_ring_ctx *ctx = iowq->ctx;
7198 * Wake up if we have enough events, or if a timeout occurred since we
7199 * started waiting. For timeouts, we always want to return to userspace,
7200 * regardless of event count.
7202 return io_cqring_events(ctx) >= iowq->to_wait ||
7203 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
7206 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
7207 int wake_flags, void *key)
7209 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
7213 * Cannot safely flush overflowed CQEs from here, ensure we wake up
7214 * the task, and the next invocation will do it.
7216 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->cq_check_overflow))
7217 return autoremove_wake_function(curr, mode, wake_flags, key);
7221 static int io_run_task_work_sig(void)
7223 if (io_run_task_work())
7225 if (!signal_pending(current))
7227 if (test_tsk_thread_flag(current, TIF_NOTIFY_SIGNAL))
7228 return -ERESTARTSYS;
7232 /* when returns >0, the caller should retry */
7233 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
7234 struct io_wait_queue *iowq,
7235 signed long *timeout)
7239 /* make sure we run task_work before checking for signals */
7240 ret = io_run_task_work_sig();
7241 if (ret || io_should_wake(iowq))
7243 /* let the caller flush overflows, retry */
7244 if (test_bit(0, &ctx->cq_check_overflow))
7247 *timeout = schedule_timeout(*timeout);
7248 return !*timeout ? -ETIME : 1;
7252 * Wait until events become available, if we don't already have some. The
7253 * application must reap them itself, as they reside on the shared cq ring.
7255 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
7256 const sigset_t __user *sig, size_t sigsz,
7257 struct __kernel_timespec __user *uts)
7259 struct io_wait_queue iowq = {
7262 .func = io_wake_function,
7263 .entry = LIST_HEAD_INIT(iowq.wq.entry),
7266 .to_wait = min_events,
7268 struct io_rings *rings = ctx->rings;
7269 signed long timeout = MAX_SCHEDULE_TIMEOUT;
7273 io_cqring_overflow_flush(ctx, false, NULL, NULL);
7274 if (io_cqring_events(ctx) >= min_events)
7276 if (!io_run_task_work())
7281 #ifdef CONFIG_COMPAT
7282 if (in_compat_syscall())
7283 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
7287 ret = set_user_sigmask(sig, sigsz);
7294 struct timespec64 ts;
7296 if (get_timespec64(&ts, uts))
7298 timeout = timespec64_to_jiffies(&ts);
7301 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
7302 trace_io_uring_cqring_wait(ctx, min_events);
7304 io_cqring_overflow_flush(ctx, false, NULL, NULL);
7305 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
7306 TASK_INTERRUPTIBLE);
7307 ret = io_cqring_wait_schedule(ctx, &iowq, &timeout);
7308 finish_wait(&ctx->wait, &iowq.wq);
7311 restore_saved_sigmask_unless(ret == -EINTR);
7313 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
7316 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
7318 #if defined(CONFIG_UNIX)
7319 if (ctx->ring_sock) {
7320 struct sock *sock = ctx->ring_sock->sk;
7321 struct sk_buff *skb;
7323 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
7329 for (i = 0; i < ctx->nr_user_files; i++) {
7332 file = io_file_from_index(ctx, i);
7339 static void io_rsrc_data_ref_zero(struct percpu_ref *ref)
7341 struct fixed_rsrc_data *data;
7343 data = container_of(ref, struct fixed_rsrc_data, refs);
7344 complete(&data->done);
7347 static inline void io_rsrc_ref_lock(struct io_ring_ctx *ctx)
7349 spin_lock_bh(&ctx->rsrc_ref_lock);
7352 static inline void io_rsrc_ref_unlock(struct io_ring_ctx *ctx)
7354 spin_unlock_bh(&ctx->rsrc_ref_lock);
7357 static void io_sqe_rsrc_set_node(struct io_ring_ctx *ctx,
7358 struct fixed_rsrc_data *rsrc_data,
7359 struct fixed_rsrc_ref_node *ref_node)
7361 io_rsrc_ref_lock(ctx);
7362 rsrc_data->node = ref_node;
7363 list_add_tail(&ref_node->node, &ctx->rsrc_ref_list);
7364 io_rsrc_ref_unlock(ctx);
7365 percpu_ref_get(&rsrc_data->refs);
7368 static int io_rsrc_ref_quiesce(struct fixed_rsrc_data *data,
7369 struct io_ring_ctx *ctx,
7370 struct fixed_rsrc_ref_node *backup_node)
7372 struct fixed_rsrc_ref_node *ref_node;
7375 io_rsrc_ref_lock(ctx);
7376 ref_node = data->node;
7377 io_rsrc_ref_unlock(ctx);
7379 percpu_ref_kill(&ref_node->refs);
7381 percpu_ref_kill(&data->refs);
7383 /* wait for all refs nodes to complete */
7384 flush_delayed_work(&ctx->rsrc_put_work);
7386 ret = wait_for_completion_interruptible(&data->done);
7389 ret = io_run_task_work_sig();
7391 percpu_ref_resurrect(&data->refs);
7392 reinit_completion(&data->done);
7393 io_sqe_rsrc_set_node(ctx, data, backup_node);
7398 destroy_fixed_rsrc_ref_node(backup_node);
7402 static struct fixed_rsrc_data *alloc_fixed_rsrc_data(struct io_ring_ctx *ctx)
7404 struct fixed_rsrc_data *data;
7406 data = kzalloc(sizeof(*data), GFP_KERNEL);
7410 if (percpu_ref_init(&data->refs, io_rsrc_data_ref_zero,
7411 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) {
7416 init_completion(&data->done);
7420 static void free_fixed_rsrc_data(struct fixed_rsrc_data *data)
7422 percpu_ref_exit(&data->refs);
7427 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7429 struct fixed_rsrc_data *data = ctx->file_data;
7430 struct fixed_rsrc_ref_node *backup_node;
7431 unsigned nr_tables, i;
7436 backup_node = alloc_fixed_rsrc_ref_node(ctx);
7439 init_fixed_file_ref_node(ctx, backup_node);
7441 ret = io_rsrc_ref_quiesce(data, ctx, backup_node);
7445 __io_sqe_files_unregister(ctx);
7446 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
7447 for (i = 0; i < nr_tables; i++)
7448 kfree(data->table[i].files);
7449 free_fixed_rsrc_data(data);
7450 ctx->file_data = NULL;
7451 ctx->nr_user_files = 0;
7455 static void io_put_sq_data(struct io_sq_data *sqd)
7457 if (refcount_dec_and_test(&sqd->refs)) {
7459 * The park is a bit of a work-around, without it we get
7460 * warning spews on shutdown with SQPOLL set and affinity
7461 * set to a single CPU.
7464 kthread_park(sqd->thread);
7465 kthread_stop(sqd->thread);
7472 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7474 struct io_ring_ctx *ctx_attach;
7475 struct io_sq_data *sqd;
7478 f = fdget(p->wq_fd);
7480 return ERR_PTR(-ENXIO);
7481 if (f.file->f_op != &io_uring_fops) {
7483 return ERR_PTR(-EINVAL);
7486 ctx_attach = f.file->private_data;
7487 sqd = ctx_attach->sq_data;
7490 return ERR_PTR(-EINVAL);
7493 refcount_inc(&sqd->refs);
7498 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p)
7500 struct io_sq_data *sqd;
7502 if (p->flags & IORING_SETUP_ATTACH_WQ)
7503 return io_attach_sq_data(p);
7505 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7507 return ERR_PTR(-ENOMEM);
7509 refcount_set(&sqd->refs, 1);
7510 INIT_LIST_HEAD(&sqd->ctx_list);
7511 INIT_LIST_HEAD(&sqd->ctx_new_list);
7512 mutex_init(&sqd->ctx_lock);
7513 mutex_init(&sqd->lock);
7514 init_waitqueue_head(&sqd->wait);
7518 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7519 __releases(&sqd->lock)
7523 kthread_unpark(sqd->thread);
7524 mutex_unlock(&sqd->lock);
7527 static void io_sq_thread_park(struct io_sq_data *sqd)
7528 __acquires(&sqd->lock)
7532 mutex_lock(&sqd->lock);
7533 kthread_park(sqd->thread);
7536 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
7538 struct io_sq_data *sqd = ctx->sq_data;
7543 * We may arrive here from the error branch in
7544 * io_sq_offload_create() where the kthread is created
7545 * without being waked up, thus wake it up now to make
7546 * sure the wait will complete.
7548 wake_up_process(sqd->thread);
7549 wait_for_completion(&ctx->sq_thread_comp);
7551 io_sq_thread_park(sqd);
7554 mutex_lock(&sqd->ctx_lock);
7555 list_del(&ctx->sqd_list);
7556 io_sqd_update_thread_idle(sqd);
7557 mutex_unlock(&sqd->ctx_lock);
7560 io_sq_thread_unpark(sqd);
7562 io_put_sq_data(sqd);
7563 ctx->sq_data = NULL;
7567 static void io_finish_async(struct io_ring_ctx *ctx)
7569 io_sq_thread_stop(ctx);
7572 io_wq_destroy(ctx->io_wq);
7577 #if defined(CONFIG_UNIX)
7579 * Ensure the UNIX gc is aware of our file set, so we are certain that
7580 * the io_uring can be safely unregistered on process exit, even if we have
7581 * loops in the file referencing.
7583 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7585 struct sock *sk = ctx->ring_sock->sk;
7586 struct scm_fp_list *fpl;
7587 struct sk_buff *skb;
7590 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7594 skb = alloc_skb(0, GFP_KERNEL);
7603 fpl->user = get_uid(ctx->user);
7604 for (i = 0; i < nr; i++) {
7605 struct file *file = io_file_from_index(ctx, i + offset);
7609 fpl->fp[nr_files] = get_file(file);
7610 unix_inflight(fpl->user, fpl->fp[nr_files]);
7615 fpl->max = SCM_MAX_FD;
7616 fpl->count = nr_files;
7617 UNIXCB(skb).fp = fpl;
7618 skb->destructor = unix_destruct_scm;
7619 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7620 skb_queue_head(&sk->sk_receive_queue, skb);
7622 for (i = 0; i < nr_files; i++)
7633 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7634 * causes regular reference counting to break down. We rely on the UNIX
7635 * garbage collection to take care of this problem for us.
7637 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7639 unsigned left, total;
7643 left = ctx->nr_user_files;
7645 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7647 ret = __io_sqe_files_scm(ctx, this_files, total);
7651 total += this_files;
7657 while (total < ctx->nr_user_files) {
7658 struct file *file = io_file_from_index(ctx, total);
7668 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7674 static int io_sqe_alloc_file_tables(struct fixed_rsrc_data *file_data,
7675 unsigned nr_tables, unsigned nr_files)
7679 for (i = 0; i < nr_tables; i++) {
7680 struct fixed_rsrc_table *table = &file_data->table[i];
7681 unsigned this_files;
7683 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
7684 table->files = kcalloc(this_files, sizeof(struct file *),
7688 nr_files -= this_files;
7694 for (i = 0; i < nr_tables; i++) {
7695 struct fixed_rsrc_table *table = &file_data->table[i];
7696 kfree(table->files);
7701 static void io_ring_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
7703 struct file *file = prsrc->file;
7704 #if defined(CONFIG_UNIX)
7705 struct sock *sock = ctx->ring_sock->sk;
7706 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7707 struct sk_buff *skb;
7710 __skb_queue_head_init(&list);
7713 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7714 * remove this entry and rearrange the file array.
7716 skb = skb_dequeue(head);
7718 struct scm_fp_list *fp;
7720 fp = UNIXCB(skb).fp;
7721 for (i = 0; i < fp->count; i++) {
7724 if (fp->fp[i] != file)
7727 unix_notinflight(fp->user, fp->fp[i]);
7728 left = fp->count - 1 - i;
7730 memmove(&fp->fp[i], &fp->fp[i + 1],
7731 left * sizeof(struct file *));
7738 __skb_queue_tail(&list, skb);
7748 __skb_queue_tail(&list, skb);
7750 skb = skb_dequeue(head);
7753 if (skb_peek(&list)) {
7754 spin_lock_irq(&head->lock);
7755 while ((skb = __skb_dequeue(&list)) != NULL)
7756 __skb_queue_tail(head, skb);
7757 spin_unlock_irq(&head->lock);
7764 static void __io_rsrc_put_work(struct fixed_rsrc_ref_node *ref_node)
7766 struct fixed_rsrc_data *rsrc_data = ref_node->rsrc_data;
7767 struct io_ring_ctx *ctx = rsrc_data->ctx;
7768 struct io_rsrc_put *prsrc, *tmp;
7770 list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
7771 list_del(&prsrc->list);
7772 ref_node->rsrc_put(ctx, prsrc);
7776 percpu_ref_exit(&ref_node->refs);
7778 percpu_ref_put(&rsrc_data->refs);
7781 static void io_rsrc_put_work(struct work_struct *work)
7783 struct io_ring_ctx *ctx;
7784 struct llist_node *node;
7786 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
7787 node = llist_del_all(&ctx->rsrc_put_llist);
7790 struct fixed_rsrc_ref_node *ref_node;
7791 struct llist_node *next = node->next;
7793 ref_node = llist_entry(node, struct fixed_rsrc_ref_node, llist);
7794 __io_rsrc_put_work(ref_node);
7799 static struct file **io_fixed_file_slot(struct fixed_rsrc_data *file_data,
7802 struct fixed_rsrc_table *table;
7804 table = &file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7805 return &table->files[i & IORING_FILE_TABLE_MASK];
7808 static void io_rsrc_node_ref_zero(struct percpu_ref *ref)
7810 struct fixed_rsrc_ref_node *ref_node;
7811 struct fixed_rsrc_data *data;
7812 struct io_ring_ctx *ctx;
7813 bool first_add = false;
7816 ref_node = container_of(ref, struct fixed_rsrc_ref_node, refs);
7817 data = ref_node->rsrc_data;
7820 io_rsrc_ref_lock(ctx);
7821 ref_node->done = true;
7823 while (!list_empty(&ctx->rsrc_ref_list)) {
7824 ref_node = list_first_entry(&ctx->rsrc_ref_list,
7825 struct fixed_rsrc_ref_node, node);
7826 /* recycle ref nodes in order */
7827 if (!ref_node->done)
7829 list_del(&ref_node->node);
7830 first_add |= llist_add(&ref_node->llist, &ctx->rsrc_put_llist);
7832 io_rsrc_ref_unlock(ctx);
7834 if (percpu_ref_is_dying(&data->refs))
7838 mod_delayed_work(system_wq, &ctx->rsrc_put_work, 0);
7840 queue_delayed_work(system_wq, &ctx->rsrc_put_work, delay);
7843 static struct fixed_rsrc_ref_node *alloc_fixed_rsrc_ref_node(
7844 struct io_ring_ctx *ctx)
7846 struct fixed_rsrc_ref_node *ref_node;
7848 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7852 if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
7857 INIT_LIST_HEAD(&ref_node->node);
7858 INIT_LIST_HEAD(&ref_node->rsrc_list);
7859 ref_node->done = false;
7863 static void init_fixed_file_ref_node(struct io_ring_ctx *ctx,
7864 struct fixed_rsrc_ref_node *ref_node)
7866 ref_node->rsrc_data = ctx->file_data;
7867 ref_node->rsrc_put = io_ring_file_put;
7870 static void destroy_fixed_rsrc_ref_node(struct fixed_rsrc_ref_node *ref_node)
7872 percpu_ref_exit(&ref_node->refs);
7877 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7880 __s32 __user *fds = (__s32 __user *) arg;
7881 unsigned nr_tables, i;
7883 int fd, ret = -ENOMEM;
7884 struct fixed_rsrc_ref_node *ref_node;
7885 struct fixed_rsrc_data *file_data;
7891 if (nr_args > IORING_MAX_FIXED_FILES)
7894 file_data = alloc_fixed_rsrc_data(ctx);
7897 ctx->file_data = file_data;
7899 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
7900 file_data->table = kcalloc(nr_tables, sizeof(*file_data->table),
7902 if (!file_data->table)
7905 if (io_sqe_alloc_file_tables(file_data, nr_tables, nr_args))
7908 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7909 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7913 /* allow sparse sets */
7923 * Don't allow io_uring instances to be registered. If UNIX
7924 * isn't enabled, then this causes a reference cycle and this
7925 * instance can never get freed. If UNIX is enabled we'll
7926 * handle it just fine, but there's still no point in allowing
7927 * a ring fd as it doesn't support regular read/write anyway.
7929 if (file->f_op == &io_uring_fops) {
7933 *io_fixed_file_slot(file_data, i) = file;
7936 ret = io_sqe_files_scm(ctx);
7938 io_sqe_files_unregister(ctx);
7942 ref_node = alloc_fixed_rsrc_ref_node(ctx);
7944 io_sqe_files_unregister(ctx);
7947 init_fixed_file_ref_node(ctx, ref_node);
7949 io_sqe_rsrc_set_node(ctx, file_data, ref_node);
7952 for (i = 0; i < ctx->nr_user_files; i++) {
7953 file = io_file_from_index(ctx, i);
7957 for (i = 0; i < nr_tables; i++)
7958 kfree(file_data->table[i].files);
7959 ctx->nr_user_files = 0;
7961 free_fixed_rsrc_data(ctx->file_data);
7962 ctx->file_data = NULL;
7966 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7969 #if defined(CONFIG_UNIX)
7970 struct sock *sock = ctx->ring_sock->sk;
7971 struct sk_buff_head *head = &sock->sk_receive_queue;
7972 struct sk_buff *skb;
7975 * See if we can merge this file into an existing skb SCM_RIGHTS
7976 * file set. If there's no room, fall back to allocating a new skb
7977 * and filling it in.
7979 spin_lock_irq(&head->lock);
7980 skb = skb_peek(head);
7982 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7984 if (fpl->count < SCM_MAX_FD) {
7985 __skb_unlink(skb, head);
7986 spin_unlock_irq(&head->lock);
7987 fpl->fp[fpl->count] = get_file(file);
7988 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7990 spin_lock_irq(&head->lock);
7991 __skb_queue_head(head, skb);
7996 spin_unlock_irq(&head->lock);
8003 return __io_sqe_files_scm(ctx, 1, index);
8009 static int io_queue_rsrc_removal(struct fixed_rsrc_data *data, void *rsrc)
8011 struct io_rsrc_put *prsrc;
8012 struct fixed_rsrc_ref_node *ref_node = data->node;
8014 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
8019 list_add(&prsrc->list, &ref_node->rsrc_list);
8024 static inline int io_queue_file_removal(struct fixed_rsrc_data *data,
8027 return io_queue_rsrc_removal(data, (void *)file);
8030 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
8031 struct io_uring_rsrc_update *up,
8034 struct fixed_rsrc_data *data = ctx->file_data;
8035 struct fixed_rsrc_ref_node *ref_node;
8036 struct file *file, **file_slot;
8040 bool needs_switch = false;
8042 if (check_add_overflow(up->offset, nr_args, &done))
8044 if (done > ctx->nr_user_files)
8047 ref_node = alloc_fixed_rsrc_ref_node(ctx);
8050 init_fixed_file_ref_node(ctx, ref_node);
8052 fds = u64_to_user_ptr(up->data);
8053 for (done = 0; done < nr_args; done++) {
8055 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
8059 if (fd == IORING_REGISTER_FILES_SKIP)
8062 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
8063 file_slot = io_fixed_file_slot(ctx->file_data, i);
8066 err = io_queue_file_removal(data, *file_slot);
8070 needs_switch = true;
8079 * Don't allow io_uring instances to be registered. If
8080 * UNIX isn't enabled, then this causes a reference
8081 * cycle and this instance can never get freed. If UNIX
8082 * is enabled we'll handle it just fine, but there's
8083 * still no point in allowing a ring fd as it doesn't
8084 * support regular read/write anyway.
8086 if (file->f_op == &io_uring_fops) {
8092 err = io_sqe_file_register(ctx, file, i);
8102 percpu_ref_kill(&data->node->refs);
8103 io_sqe_rsrc_set_node(ctx, data, ref_node);
8105 destroy_fixed_rsrc_ref_node(ref_node);
8107 return done ? done : err;
8110 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
8113 struct io_uring_rsrc_update up;
8115 if (!ctx->file_data)
8119 if (copy_from_user(&up, arg, sizeof(up)))
8124 return __io_sqe_files_update(ctx, &up, nr_args);
8127 static struct io_wq_work *io_free_work(struct io_wq_work *work)
8129 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8131 req = io_put_req_find_next(req);
8132 return req ? &req->work : NULL;
8135 static int io_init_wq_offload(struct io_ring_ctx *ctx,
8136 struct io_uring_params *p)
8138 struct io_wq_data data;
8140 struct io_ring_ctx *ctx_attach;
8141 unsigned int concurrency;
8144 data.user = ctx->user;
8145 data.free_work = io_free_work;
8146 data.do_work = io_wq_submit_work;
8148 if (!(p->flags & IORING_SETUP_ATTACH_WQ)) {
8149 /* Do QD, or 4 * CPUS, whatever is smallest */
8150 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
8152 ctx->io_wq = io_wq_create(concurrency, &data);
8153 if (IS_ERR(ctx->io_wq)) {
8154 ret = PTR_ERR(ctx->io_wq);
8160 f = fdget(p->wq_fd);
8164 if (f.file->f_op != &io_uring_fops) {
8169 ctx_attach = f.file->private_data;
8170 /* @io_wq is protected by holding the fd */
8171 if (!io_wq_get(ctx_attach->io_wq, &data)) {
8176 ctx->io_wq = ctx_attach->io_wq;
8182 static int io_uring_alloc_task_context(struct task_struct *task)
8184 struct io_uring_task *tctx;
8187 tctx = kmalloc(sizeof(*tctx), GFP_KERNEL);
8188 if (unlikely(!tctx))
8191 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
8192 if (unlikely(ret)) {
8198 init_waitqueue_head(&tctx->wait);
8200 atomic_set(&tctx->in_idle, 0);
8201 tctx->sqpoll = false;
8202 io_init_identity(&tctx->__identity);
8203 tctx->identity = &tctx->__identity;
8204 task->io_uring = tctx;
8205 spin_lock_init(&tctx->task_lock);
8206 INIT_WQ_LIST(&tctx->task_list);
8207 tctx->task_state = 0;
8208 init_task_work(&tctx->task_work, tctx_task_work);
8212 void __io_uring_free(struct task_struct *tsk)
8214 struct io_uring_task *tctx = tsk->io_uring;
8216 WARN_ON_ONCE(!xa_empty(&tctx->xa));
8217 WARN_ON_ONCE(refcount_read(&tctx->identity->count) != 1);
8218 if (tctx->identity != &tctx->__identity)
8219 kfree(tctx->identity);
8220 percpu_counter_destroy(&tctx->inflight);
8222 tsk->io_uring = NULL;
8225 static int io_sq_offload_create(struct io_ring_ctx *ctx,
8226 struct io_uring_params *p)
8230 if (ctx->flags & IORING_SETUP_SQPOLL) {
8231 struct io_sq_data *sqd;
8234 if (!capable(CAP_SYS_ADMIN) && !capable(CAP_SYS_NICE))
8237 sqd = io_get_sq_data(p);
8244 io_sq_thread_park(sqd);
8245 mutex_lock(&sqd->ctx_lock);
8246 list_add(&ctx->sqd_list, &sqd->ctx_new_list);
8247 mutex_unlock(&sqd->ctx_lock);
8248 io_sq_thread_unpark(sqd);
8250 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
8251 if (!ctx->sq_thread_idle)
8252 ctx->sq_thread_idle = HZ;
8257 if (p->flags & IORING_SETUP_SQ_AFF) {
8258 int cpu = p->sq_thread_cpu;
8261 if (cpu >= nr_cpu_ids)
8263 if (!cpu_online(cpu))
8266 sqd->thread = kthread_create_on_cpu(io_sq_thread, sqd,
8267 cpu, "io_uring-sq");
8269 sqd->thread = kthread_create(io_sq_thread, sqd,
8272 if (IS_ERR(sqd->thread)) {
8273 ret = PTR_ERR(sqd->thread);
8277 ret = io_uring_alloc_task_context(sqd->thread);
8280 } else if (p->flags & IORING_SETUP_SQ_AFF) {
8281 /* Can't have SQ_AFF without SQPOLL */
8287 ret = io_init_wq_offload(ctx, p);
8293 io_finish_async(ctx);
8297 static void io_sq_offload_start(struct io_ring_ctx *ctx)
8299 struct io_sq_data *sqd = ctx->sq_data;
8301 if ((ctx->flags & IORING_SETUP_SQPOLL) && sqd->thread)
8302 wake_up_process(sqd->thread);
8305 static inline void __io_unaccount_mem(struct user_struct *user,
8306 unsigned long nr_pages)
8308 atomic_long_sub(nr_pages, &user->locked_vm);
8311 static inline int __io_account_mem(struct user_struct *user,
8312 unsigned long nr_pages)
8314 unsigned long page_limit, cur_pages, new_pages;
8316 /* Don't allow more pages than we can safely lock */
8317 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
8320 cur_pages = atomic_long_read(&user->locked_vm);
8321 new_pages = cur_pages + nr_pages;
8322 if (new_pages > page_limit)
8324 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
8325 new_pages) != cur_pages);
8330 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8333 __io_unaccount_mem(ctx->user, nr_pages);
8335 if (ctx->mm_account)
8336 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
8339 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8343 if (ctx->limit_mem) {
8344 ret = __io_account_mem(ctx->user, nr_pages);
8349 if (ctx->mm_account)
8350 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
8355 static void io_mem_free(void *ptr)
8362 page = virt_to_head_page(ptr);
8363 if (put_page_testzero(page))
8364 free_compound_page(page);
8367 static void *io_mem_alloc(size_t size)
8369 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
8370 __GFP_NORETRY | __GFP_ACCOUNT;
8372 return (void *) __get_free_pages(gfp_flags, get_order(size));
8375 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8378 struct io_rings *rings;
8379 size_t off, sq_array_size;
8381 off = struct_size(rings, cqes, cq_entries);
8382 if (off == SIZE_MAX)
8386 off = ALIGN(off, SMP_CACHE_BYTES);
8394 sq_array_size = array_size(sizeof(u32), sq_entries);
8395 if (sq_array_size == SIZE_MAX)
8398 if (check_add_overflow(off, sq_array_size, &off))
8404 static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8408 if (!ctx->user_bufs)
8411 for (i = 0; i < ctx->nr_user_bufs; i++) {
8412 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8414 for (j = 0; j < imu->nr_bvecs; j++)
8415 unpin_user_page(imu->bvec[j].bv_page);
8417 if (imu->acct_pages)
8418 io_unaccount_mem(ctx, imu->acct_pages);
8423 kfree(ctx->user_bufs);
8424 ctx->user_bufs = NULL;
8425 ctx->nr_user_bufs = 0;
8429 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8430 void __user *arg, unsigned index)
8432 struct iovec __user *src;
8434 #ifdef CONFIG_COMPAT
8436 struct compat_iovec __user *ciovs;
8437 struct compat_iovec ciov;
8439 ciovs = (struct compat_iovec __user *) arg;
8440 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8443 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8444 dst->iov_len = ciov.iov_len;
8448 src = (struct iovec __user *) arg;
8449 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8455 * Not super efficient, but this is just a registration time. And we do cache
8456 * the last compound head, so generally we'll only do a full search if we don't
8459 * We check if the given compound head page has already been accounted, to
8460 * avoid double accounting it. This allows us to account the full size of the
8461 * page, not just the constituent pages of a huge page.
8463 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8464 int nr_pages, struct page *hpage)
8468 /* check current page array */
8469 for (i = 0; i < nr_pages; i++) {
8470 if (!PageCompound(pages[i]))
8472 if (compound_head(pages[i]) == hpage)
8476 /* check previously registered pages */
8477 for (i = 0; i < ctx->nr_user_bufs; i++) {
8478 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8480 for (j = 0; j < imu->nr_bvecs; j++) {
8481 if (!PageCompound(imu->bvec[j].bv_page))
8483 if (compound_head(imu->bvec[j].bv_page) == hpage)
8491 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8492 int nr_pages, struct io_mapped_ubuf *imu,
8493 struct page **last_hpage)
8497 for (i = 0; i < nr_pages; i++) {
8498 if (!PageCompound(pages[i])) {
8503 hpage = compound_head(pages[i]);
8504 if (hpage == *last_hpage)
8506 *last_hpage = hpage;
8507 if (headpage_already_acct(ctx, pages, i, hpage))
8509 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8513 if (!imu->acct_pages)
8516 ret = io_account_mem(ctx, imu->acct_pages);
8518 imu->acct_pages = 0;
8522 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
8523 struct io_mapped_ubuf *imu,
8524 struct page **last_hpage)
8526 struct vm_area_struct **vmas = NULL;
8527 struct page **pages = NULL;
8528 unsigned long off, start, end, ubuf;
8530 int ret, pret, nr_pages, i;
8532 ubuf = (unsigned long) iov->iov_base;
8533 end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8534 start = ubuf >> PAGE_SHIFT;
8535 nr_pages = end - start;
8539 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
8543 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
8548 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
8554 mmap_read_lock(current->mm);
8555 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
8557 if (pret == nr_pages) {
8558 /* don't support file backed memory */
8559 for (i = 0; i < nr_pages; i++) {
8560 struct vm_area_struct *vma = vmas[i];
8563 !is_file_hugepages(vma->vm_file)) {
8569 ret = pret < 0 ? pret : -EFAULT;
8571 mmap_read_unlock(current->mm);
8574 * if we did partial map, or found file backed vmas,
8575 * release any pages we did get
8578 unpin_user_pages(pages, pret);
8583 ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
8585 unpin_user_pages(pages, pret);
8590 off = ubuf & ~PAGE_MASK;
8591 size = iov->iov_len;
8592 for (i = 0; i < nr_pages; i++) {
8595 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8596 imu->bvec[i].bv_page = pages[i];
8597 imu->bvec[i].bv_len = vec_len;
8598 imu->bvec[i].bv_offset = off;
8602 /* store original address for later verification */
8604 imu->len = iov->iov_len;
8605 imu->nr_bvecs = nr_pages;
8613 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
8617 if (!nr_args || nr_args > UIO_MAXIOV)
8620 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
8622 if (!ctx->user_bufs)
8628 static int io_buffer_validate(struct iovec *iov)
8631 * Don't impose further limits on the size and buffer
8632 * constraints here, we'll -EINVAL later when IO is
8633 * submitted if they are wrong.
8635 if (!iov->iov_base || !iov->iov_len)
8638 /* arbitrary limit, but we need something */
8639 if (iov->iov_len > SZ_1G)
8645 static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
8646 unsigned int nr_args)
8650 struct page *last_hpage = NULL;
8652 ret = io_buffers_map_alloc(ctx, nr_args);
8656 for (i = 0; i < nr_args; i++) {
8657 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8659 ret = io_copy_iov(ctx, &iov, arg, i);
8663 ret = io_buffer_validate(&iov);
8667 ret = io_sqe_buffer_register(ctx, &iov, imu, &last_hpage);
8671 ctx->nr_user_bufs++;
8675 io_sqe_buffers_unregister(ctx);
8680 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8682 __s32 __user *fds = arg;
8688 if (copy_from_user(&fd, fds, sizeof(*fds)))
8691 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8692 if (IS_ERR(ctx->cq_ev_fd)) {
8693 int ret = PTR_ERR(ctx->cq_ev_fd);
8694 ctx->cq_ev_fd = NULL;
8701 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8703 if (ctx->cq_ev_fd) {
8704 eventfd_ctx_put(ctx->cq_ev_fd);
8705 ctx->cq_ev_fd = NULL;
8712 static int __io_destroy_buffers(int id, void *p, void *data)
8714 struct io_ring_ctx *ctx = data;
8715 struct io_buffer *buf = p;
8717 __io_remove_buffers(ctx, buf, id, -1U);
8721 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8723 idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx);
8724 idr_destroy(&ctx->io_buffer_idr);
8727 static void io_req_cache_free(struct list_head *list)
8729 while (!list_empty(list)) {
8730 struct io_kiocb *req;
8732 req = list_first_entry(list, struct io_kiocb, compl.list);
8733 list_del(&req->compl.list);
8734 kmem_cache_free(req_cachep, req);
8738 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8740 struct io_submit_state *submit_state = &ctx->submit_state;
8742 io_finish_async(ctx);
8743 io_sqe_buffers_unregister(ctx);
8745 if (ctx->sqo_task) {
8746 put_task_struct(ctx->sqo_task);
8747 ctx->sqo_task = NULL;
8748 mmdrop(ctx->mm_account);
8749 ctx->mm_account = NULL;
8752 if (submit_state->free_reqs)
8753 kmem_cache_free_bulk(req_cachep, submit_state->free_reqs,
8754 submit_state->reqs);
8756 #ifdef CONFIG_BLK_CGROUP
8757 if (ctx->sqo_blkcg_css)
8758 css_put(ctx->sqo_blkcg_css);
8761 io_sqe_files_unregister(ctx);
8762 io_eventfd_unregister(ctx);
8763 io_destroy_buffers(ctx);
8764 idr_destroy(&ctx->personality_idr);
8766 #if defined(CONFIG_UNIX)
8767 if (ctx->ring_sock) {
8768 ctx->ring_sock->file = NULL; /* so that iput() is called */
8769 sock_release(ctx->ring_sock);
8773 io_mem_free(ctx->rings);
8774 io_mem_free(ctx->sq_sqes);
8776 percpu_ref_exit(&ctx->refs);
8777 free_uid(ctx->user);
8778 put_cred(ctx->creds);
8779 kfree(ctx->cancel_hash);
8780 io_req_cache_free(&ctx->submit_state.comp.free_list);
8781 io_req_cache_free(&ctx->submit_state.comp.locked_free_list);
8785 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8787 struct io_ring_ctx *ctx = file->private_data;
8790 poll_wait(file, &ctx->cq_wait, wait);
8792 * synchronizes with barrier from wq_has_sleeper call in
8796 if (!io_sqring_full(ctx))
8797 mask |= EPOLLOUT | EPOLLWRNORM;
8800 * Don't flush cqring overflow list here, just do a simple check.
8801 * Otherwise there could possible be ABBA deadlock:
8804 * lock(&ctx->uring_lock);
8806 * lock(&ctx->uring_lock);
8809 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
8810 * pushs them to do the flush.
8812 if (io_cqring_events(ctx) || test_bit(0, &ctx->cq_check_overflow))
8813 mask |= EPOLLIN | EPOLLRDNORM;
8818 static int io_uring_fasync(int fd, struct file *file, int on)
8820 struct io_ring_ctx *ctx = file->private_data;
8822 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8825 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8827 struct io_identity *iod;
8829 iod = idr_remove(&ctx->personality_idr, id);
8831 put_cred(iod->creds);
8832 if (refcount_dec_and_test(&iod->count))
8840 static int io_remove_personalities(int id, void *p, void *data)
8842 struct io_ring_ctx *ctx = data;
8844 io_unregister_personality(ctx, id);
8848 static void io_ring_exit_work(struct work_struct *work)
8850 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
8854 * If we're doing polled IO and end up having requests being
8855 * submitted async (out-of-line), then completions can come in while
8856 * we're waiting for refs to drop. We need to reap these manually,
8857 * as nobody else will be looking for them.
8860 io_uring_try_cancel_requests(ctx, NULL, NULL);
8861 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
8862 io_ring_ctx_free(ctx);
8865 static bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
8867 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8869 return req->ctx == data;
8872 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8874 mutex_lock(&ctx->uring_lock);
8875 percpu_ref_kill(&ctx->refs);
8877 if (WARN_ON_ONCE((ctx->flags & IORING_SETUP_SQPOLL) && !ctx->sqo_dead))
8880 /* if force is set, the ring is going away. always drop after that */
8881 ctx->cq_overflow_flushed = 1;
8883 __io_cqring_overflow_flush(ctx, true, NULL, NULL);
8884 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
8885 mutex_unlock(&ctx->uring_lock);
8887 io_kill_timeouts(ctx, NULL, NULL);
8888 io_poll_remove_all(ctx, NULL, NULL);
8891 io_wq_cancel_cb(ctx->io_wq, io_cancel_ctx_cb, ctx, true);
8893 /* if we failed setting up the ctx, we might not have any rings */
8894 io_iopoll_try_reap_events(ctx);
8896 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8898 * Use system_unbound_wq to avoid spawning tons of event kworkers
8899 * if we're exiting a ton of rings at the same time. It just adds
8900 * noise and overhead, there's no discernable change in runtime
8901 * over using system_wq.
8903 queue_work(system_unbound_wq, &ctx->exit_work);
8906 static int io_uring_release(struct inode *inode, struct file *file)
8908 struct io_ring_ctx *ctx = file->private_data;
8910 file->private_data = NULL;
8911 io_ring_ctx_wait_and_kill(ctx);
8915 struct io_task_cancel {
8916 struct task_struct *task;
8917 struct files_struct *files;
8920 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8922 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8923 struct io_task_cancel *cancel = data;
8926 if (cancel->files && (req->flags & REQ_F_LINK_TIMEOUT)) {
8927 unsigned long flags;
8928 struct io_ring_ctx *ctx = req->ctx;
8930 /* protect against races with linked timeouts */
8931 spin_lock_irqsave(&ctx->completion_lock, flags);
8932 ret = io_match_task(req, cancel->task, cancel->files);
8933 spin_unlock_irqrestore(&ctx->completion_lock, flags);
8935 ret = io_match_task(req, cancel->task, cancel->files);
8940 static void io_cancel_defer_files(struct io_ring_ctx *ctx,
8941 struct task_struct *task,
8942 struct files_struct *files)
8944 struct io_defer_entry *de = NULL;
8947 spin_lock_irq(&ctx->completion_lock);
8948 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8949 if (io_match_task(de->req, task, files)) {
8950 list_cut_position(&list, &ctx->defer_list, &de->list);
8954 spin_unlock_irq(&ctx->completion_lock);
8956 while (!list_empty(&list)) {
8957 de = list_first_entry(&list, struct io_defer_entry, list);
8958 list_del_init(&de->list);
8959 req_set_fail_links(de->req);
8960 io_put_req(de->req);
8961 io_req_complete(de->req, -ECANCELED);
8966 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
8967 struct task_struct *task,
8968 struct files_struct *files)
8970 struct io_task_cancel cancel = { .task = task, .files = files, };
8973 enum io_wq_cancel cret;
8977 cret = io_wq_cancel_cb(ctx->io_wq, io_cancel_task_cb,
8979 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8982 /* SQPOLL thread does its own polling */
8983 if (!(ctx->flags & IORING_SETUP_SQPOLL) && !files) {
8984 while (!list_empty_careful(&ctx->iopoll_list)) {
8985 io_iopoll_try_reap_events(ctx);
8990 ret |= io_poll_remove_all(ctx, task, files);
8991 ret |= io_kill_timeouts(ctx, task, files);
8992 ret |= io_run_task_work();
8993 io_cqring_overflow_flush(ctx, true, task, files);
9000 static int io_uring_count_inflight(struct io_ring_ctx *ctx,
9001 struct task_struct *task,
9002 struct files_struct *files)
9004 struct io_kiocb *req;
9007 spin_lock_irq(&ctx->inflight_lock);
9008 list_for_each_entry(req, &ctx->inflight_list, inflight_entry)
9009 cnt += io_match_task(req, task, files);
9010 spin_unlock_irq(&ctx->inflight_lock);
9014 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
9015 struct task_struct *task,
9016 struct files_struct *files)
9018 while (!list_empty_careful(&ctx->inflight_list)) {
9022 inflight = io_uring_count_inflight(ctx, task, files);
9026 io_uring_try_cancel_requests(ctx, task, files);
9029 io_sq_thread_unpark(ctx->sq_data);
9030 prepare_to_wait(&task->io_uring->wait, &wait,
9031 TASK_UNINTERRUPTIBLE);
9032 if (inflight == io_uring_count_inflight(ctx, task, files))
9034 finish_wait(&task->io_uring->wait, &wait);
9036 io_sq_thread_park(ctx->sq_data);
9040 static void io_disable_sqo_submit(struct io_ring_ctx *ctx)
9042 mutex_lock(&ctx->uring_lock);
9044 mutex_unlock(&ctx->uring_lock);
9046 /* make sure callers enter the ring to get error */
9048 io_ring_set_wakeup_flag(ctx);
9052 * We need to iteratively cancel requests, in case a request has dependent
9053 * hard links. These persist even for failure of cancelations, hence keep
9054 * looping until none are found.
9056 static void io_uring_cancel_task_requests(struct io_ring_ctx *ctx,
9057 struct files_struct *files)
9059 struct task_struct *task = current;
9061 if ((ctx->flags & IORING_SETUP_SQPOLL) && ctx->sq_data) {
9062 io_disable_sqo_submit(ctx);
9063 task = ctx->sq_data->thread;
9064 atomic_inc(&task->io_uring->in_idle);
9065 io_sq_thread_park(ctx->sq_data);
9068 io_cancel_defer_files(ctx, task, files);
9070 io_uring_cancel_files(ctx, task, files);
9072 io_uring_try_cancel_requests(ctx, task, NULL);
9074 if ((ctx->flags & IORING_SETUP_SQPOLL) && ctx->sq_data) {
9075 atomic_dec(&task->io_uring->in_idle);
9077 * If the files that are going away are the ones in the thread
9078 * identity, clear them out.
9080 if (task->io_uring->identity->files == files)
9081 task->io_uring->identity->files = NULL;
9082 io_sq_thread_unpark(ctx->sq_data);
9087 * Note that this task has used io_uring. We use it for cancelation purposes.
9089 static int io_uring_add_task_file(struct io_ring_ctx *ctx, struct file *file)
9091 struct io_uring_task *tctx = current->io_uring;
9094 if (unlikely(!tctx)) {
9095 ret = io_uring_alloc_task_context(current);
9098 tctx = current->io_uring;
9100 if (tctx->last != file) {
9101 void *old = xa_load(&tctx->xa, (unsigned long)file);
9105 ret = xa_err(xa_store(&tctx->xa, (unsigned long)file,
9112 /* one and only SQPOLL file note, held by sqo_task */
9113 WARN_ON_ONCE((ctx->flags & IORING_SETUP_SQPOLL) &&
9114 current != ctx->sqo_task);
9120 * This is race safe in that the task itself is doing this, hence it
9121 * cannot be going through the exit/cancel paths at the same time.
9122 * This cannot be modified while exit/cancel is running.
9124 if (!tctx->sqpoll && (ctx->flags & IORING_SETUP_SQPOLL))
9125 tctx->sqpoll = true;
9131 * Remove this io_uring_file -> task mapping.
9133 static void io_uring_del_task_file(struct file *file)
9135 struct io_uring_task *tctx = current->io_uring;
9137 if (tctx->last == file)
9139 file = xa_erase(&tctx->xa, (unsigned long)file);
9144 static void io_uring_remove_task_files(struct io_uring_task *tctx)
9147 unsigned long index;
9149 xa_for_each(&tctx->xa, index, file)
9150 io_uring_del_task_file(file);
9153 void __io_uring_files_cancel(struct files_struct *files)
9155 struct io_uring_task *tctx = current->io_uring;
9157 unsigned long index;
9159 /* make sure overflow events are dropped */
9160 atomic_inc(&tctx->in_idle);
9161 xa_for_each(&tctx->xa, index, file)
9162 io_uring_cancel_task_requests(file->private_data, files);
9163 atomic_dec(&tctx->in_idle);
9166 io_uring_remove_task_files(tctx);
9169 static s64 tctx_inflight(struct io_uring_task *tctx)
9171 return percpu_counter_sum(&tctx->inflight);
9174 static void io_uring_cancel_sqpoll(struct io_ring_ctx *ctx)
9176 struct io_uring_task *tctx;
9182 tctx = ctx->sq_data->thread->io_uring;
9183 io_disable_sqo_submit(ctx);
9185 atomic_inc(&tctx->in_idle);
9187 /* read completions before cancelations */
9188 inflight = tctx_inflight(tctx);
9191 io_uring_cancel_task_requests(ctx, NULL);
9193 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
9195 * If we've seen completions, retry without waiting. This
9196 * avoids a race where a completion comes in before we did
9197 * prepare_to_wait().
9199 if (inflight == tctx_inflight(tctx))
9201 finish_wait(&tctx->wait, &wait);
9203 atomic_dec(&tctx->in_idle);
9207 * Find any io_uring fd that this task has registered or done IO on, and cancel
9210 void __io_uring_task_cancel(void)
9212 struct io_uring_task *tctx = current->io_uring;
9216 /* make sure overflow events are dropped */
9217 atomic_inc(&tctx->in_idle);
9219 /* trigger io_disable_sqo_submit() */
9222 unsigned long index;
9224 xa_for_each(&tctx->xa, index, file)
9225 io_uring_cancel_sqpoll(file->private_data);
9229 /* read completions before cancelations */
9230 inflight = tctx_inflight(tctx);
9233 __io_uring_files_cancel(NULL);
9235 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
9238 * If we've seen completions, retry without waiting. This
9239 * avoids a race where a completion comes in before we did
9240 * prepare_to_wait().
9242 if (inflight == tctx_inflight(tctx))
9244 finish_wait(&tctx->wait, &wait);
9247 atomic_dec(&tctx->in_idle);
9249 io_uring_remove_task_files(tctx);
9252 static int io_uring_flush(struct file *file, void *data)
9254 struct io_uring_task *tctx = current->io_uring;
9255 struct io_ring_ctx *ctx = file->private_data;
9257 if (fatal_signal_pending(current) || (current->flags & PF_EXITING))
9258 io_uring_cancel_task_requests(ctx, NULL);
9263 /* we should have cancelled and erased it before PF_EXITING */
9264 WARN_ON_ONCE((current->flags & PF_EXITING) &&
9265 xa_load(&tctx->xa, (unsigned long)file));
9268 * fput() is pending, will be 2 if the only other ref is our potential
9269 * task file note. If the task is exiting, drop regardless of count.
9271 if (atomic_long_read(&file->f_count) != 2)
9274 if (ctx->flags & IORING_SETUP_SQPOLL) {
9275 /* there is only one file note, which is owned by sqo_task */
9276 WARN_ON_ONCE(ctx->sqo_task != current &&
9277 xa_load(&tctx->xa, (unsigned long)file));
9278 /* sqo_dead check is for when this happens after cancellation */
9279 WARN_ON_ONCE(ctx->sqo_task == current && !ctx->sqo_dead &&
9280 !xa_load(&tctx->xa, (unsigned long)file));
9282 io_disable_sqo_submit(ctx);
9285 if (!(ctx->flags & IORING_SETUP_SQPOLL) || ctx->sqo_task == current)
9286 io_uring_del_task_file(file);
9290 static void *io_uring_validate_mmap_request(struct file *file,
9291 loff_t pgoff, size_t sz)
9293 struct io_ring_ctx *ctx = file->private_data;
9294 loff_t offset = pgoff << PAGE_SHIFT;
9299 case IORING_OFF_SQ_RING:
9300 case IORING_OFF_CQ_RING:
9303 case IORING_OFF_SQES:
9307 return ERR_PTR(-EINVAL);
9310 page = virt_to_head_page(ptr);
9311 if (sz > page_size(page))
9312 return ERR_PTR(-EINVAL);
9319 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9321 size_t sz = vma->vm_end - vma->vm_start;
9325 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
9327 return PTR_ERR(ptr);
9329 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
9330 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
9333 #else /* !CONFIG_MMU */
9335 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9337 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
9340 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
9342 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
9345 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
9346 unsigned long addr, unsigned long len,
9347 unsigned long pgoff, unsigned long flags)
9351 ptr = io_uring_validate_mmap_request(file, pgoff, len);
9353 return PTR_ERR(ptr);
9355 return (unsigned long) ptr;
9358 #endif /* !CONFIG_MMU */
9360 static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
9366 if (!io_sqring_full(ctx))
9369 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9371 if (unlikely(ctx->sqo_dead)) {
9376 if (!io_sqring_full(ctx))
9380 } while (!signal_pending(current));
9382 finish_wait(&ctx->sqo_sq_wait, &wait);
9387 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9388 struct __kernel_timespec __user **ts,
9389 const sigset_t __user **sig)
9391 struct io_uring_getevents_arg arg;
9394 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9395 * is just a pointer to the sigset_t.
9397 if (!(flags & IORING_ENTER_EXT_ARG)) {
9398 *sig = (const sigset_t __user *) argp;
9404 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9405 * timespec and sigset_t pointers if good.
9407 if (*argsz != sizeof(arg))
9409 if (copy_from_user(&arg, argp, sizeof(arg)))
9411 *sig = u64_to_user_ptr(arg.sigmask);
9412 *argsz = arg.sigmask_sz;
9413 *ts = u64_to_user_ptr(arg.ts);
9417 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9418 u32, min_complete, u32, flags, const void __user *, argp,
9421 struct io_ring_ctx *ctx;
9428 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9429 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG))
9437 if (f.file->f_op != &io_uring_fops)
9441 ctx = f.file->private_data;
9442 if (!percpu_ref_tryget(&ctx->refs))
9446 if (ctx->flags & IORING_SETUP_R_DISABLED)
9450 * For SQ polling, the thread will do all submissions and completions.
9451 * Just return the requested submit count, and wake the thread if
9455 if (ctx->flags & IORING_SETUP_SQPOLL) {
9456 io_cqring_overflow_flush(ctx, false, NULL, NULL);
9459 if (unlikely(ctx->sqo_dead))
9461 if (flags & IORING_ENTER_SQ_WAKEUP)
9462 wake_up(&ctx->sq_data->wait);
9463 if (flags & IORING_ENTER_SQ_WAIT) {
9464 ret = io_sqpoll_wait_sq(ctx);
9468 submitted = to_submit;
9469 } else if (to_submit) {
9470 ret = io_uring_add_task_file(ctx, f.file);
9473 mutex_lock(&ctx->uring_lock);
9474 submitted = io_submit_sqes(ctx, to_submit);
9475 mutex_unlock(&ctx->uring_lock);
9477 if (submitted != to_submit)
9480 if (flags & IORING_ENTER_GETEVENTS) {
9481 const sigset_t __user *sig;
9482 struct __kernel_timespec __user *ts;
9484 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
9488 min_complete = min(min_complete, ctx->cq_entries);
9491 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9492 * space applications don't need to do io completion events
9493 * polling again, they can rely on io_sq_thread to do polling
9494 * work, which can reduce cpu usage and uring_lock contention.
9496 if (ctx->flags & IORING_SETUP_IOPOLL &&
9497 !(ctx->flags & IORING_SETUP_SQPOLL)) {
9498 ret = io_iopoll_check(ctx, min_complete);
9500 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
9505 percpu_ref_put(&ctx->refs);
9508 return submitted ? submitted : ret;
9511 #ifdef CONFIG_PROC_FS
9512 static int io_uring_show_cred(int id, void *p, void *data)
9514 struct io_identity *iod = p;
9515 const struct cred *cred = iod->creds;
9516 struct seq_file *m = data;
9517 struct user_namespace *uns = seq_user_ns(m);
9518 struct group_info *gi;
9523 seq_printf(m, "%5d\n", id);
9524 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
9525 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
9526 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
9527 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
9528 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
9529 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
9530 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
9531 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
9532 seq_puts(m, "\n\tGroups:\t");
9533 gi = cred->group_info;
9534 for (g = 0; g < gi->ngroups; g++) {
9535 seq_put_decimal_ull(m, g ? " " : "",
9536 from_kgid_munged(uns, gi->gid[g]));
9538 seq_puts(m, "\n\tCapEff:\t");
9539 cap = cred->cap_effective;
9540 CAP_FOR_EACH_U32(__capi)
9541 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
9546 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
9548 struct io_sq_data *sq = NULL;
9553 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9554 * since fdinfo case grabs it in the opposite direction of normal use
9555 * cases. If we fail to get the lock, we just don't iterate any
9556 * structures that could be going away outside the io_uring mutex.
9558 has_lock = mutex_trylock(&ctx->uring_lock);
9560 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL))
9563 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9564 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9565 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9566 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9567 struct file *f = *io_fixed_file_slot(ctx->file_data, i);
9570 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9572 seq_printf(m, "%5u: <none>\n", i);
9574 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9575 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9576 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
9578 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
9579 (unsigned int) buf->len);
9581 if (has_lock && !idr_is_empty(&ctx->personality_idr)) {
9582 seq_printf(m, "Personalities:\n");
9583 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
9585 seq_printf(m, "PollList:\n");
9586 spin_lock_irq(&ctx->completion_lock);
9587 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9588 struct hlist_head *list = &ctx->cancel_hash[i];
9589 struct io_kiocb *req;
9591 hlist_for_each_entry(req, list, hash_node)
9592 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9593 req->task->task_works != NULL);
9595 spin_unlock_irq(&ctx->completion_lock);
9597 mutex_unlock(&ctx->uring_lock);
9600 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9602 struct io_ring_ctx *ctx = f->private_data;
9604 if (percpu_ref_tryget(&ctx->refs)) {
9605 __io_uring_show_fdinfo(ctx, m);
9606 percpu_ref_put(&ctx->refs);
9611 static const struct file_operations io_uring_fops = {
9612 .release = io_uring_release,
9613 .flush = io_uring_flush,
9614 .mmap = io_uring_mmap,
9616 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9617 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9619 .poll = io_uring_poll,
9620 .fasync = io_uring_fasync,
9621 #ifdef CONFIG_PROC_FS
9622 .show_fdinfo = io_uring_show_fdinfo,
9626 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9627 struct io_uring_params *p)
9629 struct io_rings *rings;
9630 size_t size, sq_array_offset;
9632 /* make sure these are sane, as we already accounted them */
9633 ctx->sq_entries = p->sq_entries;
9634 ctx->cq_entries = p->cq_entries;
9636 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9637 if (size == SIZE_MAX)
9640 rings = io_mem_alloc(size);
9645 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9646 rings->sq_ring_mask = p->sq_entries - 1;
9647 rings->cq_ring_mask = p->cq_entries - 1;
9648 rings->sq_ring_entries = p->sq_entries;
9649 rings->cq_ring_entries = p->cq_entries;
9650 ctx->sq_mask = rings->sq_ring_mask;
9651 ctx->cq_mask = rings->cq_ring_mask;
9653 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9654 if (size == SIZE_MAX) {
9655 io_mem_free(ctx->rings);
9660 ctx->sq_sqes = io_mem_alloc(size);
9661 if (!ctx->sq_sqes) {
9662 io_mem_free(ctx->rings);
9670 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
9674 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9678 ret = io_uring_add_task_file(ctx, file);
9683 fd_install(fd, file);
9688 * Allocate an anonymous fd, this is what constitutes the application
9689 * visible backing of an io_uring instance. The application mmaps this
9690 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9691 * we have to tie this fd to a socket for file garbage collection purposes.
9693 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
9696 #if defined(CONFIG_UNIX)
9699 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9702 return ERR_PTR(ret);
9705 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9706 O_RDWR | O_CLOEXEC);
9707 #if defined(CONFIG_UNIX)
9709 sock_release(ctx->ring_sock);
9710 ctx->ring_sock = NULL;
9712 ctx->ring_sock->file = file;
9718 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9719 struct io_uring_params __user *params)
9721 struct user_struct *user = NULL;
9722 struct io_ring_ctx *ctx;
9728 if (entries > IORING_MAX_ENTRIES) {
9729 if (!(p->flags & IORING_SETUP_CLAMP))
9731 entries = IORING_MAX_ENTRIES;
9735 * Use twice as many entries for the CQ ring. It's possible for the
9736 * application to drive a higher depth than the size of the SQ ring,
9737 * since the sqes are only used at submission time. This allows for
9738 * some flexibility in overcommitting a bit. If the application has
9739 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9740 * of CQ ring entries manually.
9742 p->sq_entries = roundup_pow_of_two(entries);
9743 if (p->flags & IORING_SETUP_CQSIZE) {
9745 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9746 * to a power-of-two, if it isn't already. We do NOT impose
9747 * any cq vs sq ring sizing.
9751 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9752 if (!(p->flags & IORING_SETUP_CLAMP))
9754 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9756 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9757 if (p->cq_entries < p->sq_entries)
9760 p->cq_entries = 2 * p->sq_entries;
9763 user = get_uid(current_user());
9765 ctx = io_ring_ctx_alloc(p);
9770 ctx->compat = in_compat_syscall();
9771 ctx->limit_mem = !capable(CAP_IPC_LOCK);
9773 ctx->creds = get_current_cred();
9775 ctx->loginuid = current->loginuid;
9776 ctx->sessionid = current->sessionid;
9778 ctx->sqo_task = get_task_struct(current);
9781 * This is just grabbed for accounting purposes. When a process exits,
9782 * the mm is exited and dropped before the files, hence we need to hang
9783 * on to this mm purely for the purposes of being able to unaccount
9784 * memory (locked/pinned vm). It's not used for anything else.
9786 mmgrab(current->mm);
9787 ctx->mm_account = current->mm;
9789 #ifdef CONFIG_BLK_CGROUP
9791 * The sq thread will belong to the original cgroup it was inited in.
9792 * If the cgroup goes offline (e.g. disabling the io controller), then
9793 * issued bios will be associated with the closest cgroup later in the
9797 ctx->sqo_blkcg_css = blkcg_css();
9798 ret = css_tryget_online(ctx->sqo_blkcg_css);
9801 /* don't init against a dying cgroup, have the user try again */
9802 ctx->sqo_blkcg_css = NULL;
9807 ret = io_allocate_scq_urings(ctx, p);
9811 ret = io_sq_offload_create(ctx, p);
9815 if (!(p->flags & IORING_SETUP_R_DISABLED))
9816 io_sq_offload_start(ctx);
9818 memset(&p->sq_off, 0, sizeof(p->sq_off));
9819 p->sq_off.head = offsetof(struct io_rings, sq.head);
9820 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9821 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9822 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9823 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9824 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9825 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9827 memset(&p->cq_off, 0, sizeof(p->cq_off));
9828 p->cq_off.head = offsetof(struct io_rings, cq.head);
9829 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9830 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9831 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9832 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9833 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9834 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9836 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9837 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9838 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9839 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
9840 IORING_FEAT_EXT_ARG;
9842 if (copy_to_user(params, p, sizeof(*p))) {
9847 file = io_uring_get_file(ctx);
9849 ret = PTR_ERR(file);
9854 * Install ring fd as the very last thing, so we don't risk someone
9855 * having closed it before we finish setup
9857 ret = io_uring_install_fd(ctx, file);
9859 io_disable_sqo_submit(ctx);
9860 /* fput will clean it up */
9865 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9868 io_disable_sqo_submit(ctx);
9869 io_ring_ctx_wait_and_kill(ctx);
9874 * Sets up an aio uring context, and returns the fd. Applications asks for a
9875 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9876 * params structure passed in.
9878 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9880 struct io_uring_params p;
9883 if (copy_from_user(&p, params, sizeof(p)))
9885 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9890 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9891 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9892 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9893 IORING_SETUP_R_DISABLED))
9896 return io_uring_create(entries, &p, params);
9899 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9900 struct io_uring_params __user *, params)
9902 return io_uring_setup(entries, params);
9905 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9907 struct io_uring_probe *p;
9911 size = struct_size(p, ops, nr_args);
9912 if (size == SIZE_MAX)
9914 p = kzalloc(size, GFP_KERNEL);
9919 if (copy_from_user(p, arg, size))
9922 if (memchr_inv(p, 0, size))
9925 p->last_op = IORING_OP_LAST - 1;
9926 if (nr_args > IORING_OP_LAST)
9927 nr_args = IORING_OP_LAST;
9929 for (i = 0; i < nr_args; i++) {
9931 if (!io_op_defs[i].not_supported)
9932 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9937 if (copy_to_user(arg, p, size))
9944 static int io_register_personality(struct io_ring_ctx *ctx)
9946 struct io_identity *id;
9949 id = kmalloc(sizeof(*id), GFP_KERNEL);
9953 io_init_identity(id);
9954 id->creds = get_current_cred();
9956 ret = idr_alloc_cyclic(&ctx->personality_idr, id, 1, USHRT_MAX, GFP_KERNEL);
9958 put_cred(id->creds);
9964 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9965 unsigned int nr_args)
9967 struct io_uring_restriction *res;
9971 /* Restrictions allowed only if rings started disabled */
9972 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9975 /* We allow only a single restrictions registration */
9976 if (ctx->restrictions.registered)
9979 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9982 size = array_size(nr_args, sizeof(*res));
9983 if (size == SIZE_MAX)
9986 res = memdup_user(arg, size);
9988 return PTR_ERR(res);
9992 for (i = 0; i < nr_args; i++) {
9993 switch (res[i].opcode) {
9994 case IORING_RESTRICTION_REGISTER_OP:
9995 if (res[i].register_op >= IORING_REGISTER_LAST) {
10000 __set_bit(res[i].register_op,
10001 ctx->restrictions.register_op);
10003 case IORING_RESTRICTION_SQE_OP:
10004 if (res[i].sqe_op >= IORING_OP_LAST) {
10009 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
10011 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
10012 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
10014 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
10015 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
10024 /* Reset all restrictions if an error happened */
10026 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
10028 ctx->restrictions.registered = true;
10034 static int io_register_enable_rings(struct io_ring_ctx *ctx)
10036 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
10039 if (ctx->restrictions.registered)
10040 ctx->restricted = 1;
10042 ctx->flags &= ~IORING_SETUP_R_DISABLED;
10044 io_sq_offload_start(ctx);
10049 static bool io_register_op_must_quiesce(int op)
10052 case IORING_UNREGISTER_FILES:
10053 case IORING_REGISTER_FILES_UPDATE:
10054 case IORING_REGISTER_PROBE:
10055 case IORING_REGISTER_PERSONALITY:
10056 case IORING_UNREGISTER_PERSONALITY:
10063 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
10064 void __user *arg, unsigned nr_args)
10065 __releases(ctx->uring_lock)
10066 __acquires(ctx->uring_lock)
10071 * We're inside the ring mutex, if the ref is already dying, then
10072 * someone else killed the ctx or is already going through
10073 * io_uring_register().
10075 if (percpu_ref_is_dying(&ctx->refs))
10078 if (io_register_op_must_quiesce(opcode)) {
10079 percpu_ref_kill(&ctx->refs);
10082 * Drop uring mutex before waiting for references to exit. If
10083 * another thread is currently inside io_uring_enter() it might
10084 * need to grab the uring_lock to make progress. If we hold it
10085 * here across the drain wait, then we can deadlock. It's safe
10086 * to drop the mutex here, since no new references will come in
10087 * after we've killed the percpu ref.
10089 mutex_unlock(&ctx->uring_lock);
10091 ret = wait_for_completion_interruptible(&ctx->ref_comp);
10094 ret = io_run_task_work_sig();
10099 mutex_lock(&ctx->uring_lock);
10102 percpu_ref_resurrect(&ctx->refs);
10107 if (ctx->restricted) {
10108 if (opcode >= IORING_REGISTER_LAST) {
10113 if (!test_bit(opcode, ctx->restrictions.register_op)) {
10120 case IORING_REGISTER_BUFFERS:
10121 ret = io_sqe_buffers_register(ctx, arg, nr_args);
10123 case IORING_UNREGISTER_BUFFERS:
10125 if (arg || nr_args)
10127 ret = io_sqe_buffers_unregister(ctx);
10129 case IORING_REGISTER_FILES:
10130 ret = io_sqe_files_register(ctx, arg, nr_args);
10132 case IORING_UNREGISTER_FILES:
10134 if (arg || nr_args)
10136 ret = io_sqe_files_unregister(ctx);
10138 case IORING_REGISTER_FILES_UPDATE:
10139 ret = io_sqe_files_update(ctx, arg, nr_args);
10141 case IORING_REGISTER_EVENTFD:
10142 case IORING_REGISTER_EVENTFD_ASYNC:
10146 ret = io_eventfd_register(ctx, arg);
10149 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
10150 ctx->eventfd_async = 1;
10152 ctx->eventfd_async = 0;
10154 case IORING_UNREGISTER_EVENTFD:
10156 if (arg || nr_args)
10158 ret = io_eventfd_unregister(ctx);
10160 case IORING_REGISTER_PROBE:
10162 if (!arg || nr_args > 256)
10164 ret = io_probe(ctx, arg, nr_args);
10166 case IORING_REGISTER_PERSONALITY:
10168 if (arg || nr_args)
10170 ret = io_register_personality(ctx);
10172 case IORING_UNREGISTER_PERSONALITY:
10176 ret = io_unregister_personality(ctx, nr_args);
10178 case IORING_REGISTER_ENABLE_RINGS:
10180 if (arg || nr_args)
10182 ret = io_register_enable_rings(ctx);
10184 case IORING_REGISTER_RESTRICTIONS:
10185 ret = io_register_restrictions(ctx, arg, nr_args);
10193 if (io_register_op_must_quiesce(opcode)) {
10194 /* bring the ctx back to life */
10195 percpu_ref_reinit(&ctx->refs);
10197 reinit_completion(&ctx->ref_comp);
10202 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
10203 void __user *, arg, unsigned int, nr_args)
10205 struct io_ring_ctx *ctx;
10214 if (f.file->f_op != &io_uring_fops)
10217 ctx = f.file->private_data;
10219 mutex_lock(&ctx->uring_lock);
10220 ret = __io_uring_register(ctx, opcode, arg, nr_args);
10221 mutex_unlock(&ctx->uring_lock);
10222 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
10223 ctx->cq_ev_fd != NULL, ret);
10229 static int __init io_uring_init(void)
10231 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
10232 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
10233 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
10236 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
10237 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
10238 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
10239 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
10240 BUILD_BUG_SQE_ELEM(1, __u8, flags);
10241 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
10242 BUILD_BUG_SQE_ELEM(4, __s32, fd);
10243 BUILD_BUG_SQE_ELEM(8, __u64, off);
10244 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
10245 BUILD_BUG_SQE_ELEM(16, __u64, addr);
10246 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
10247 BUILD_BUG_SQE_ELEM(24, __u32, len);
10248 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
10249 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
10250 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
10251 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
10252 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
10253 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
10254 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
10255 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
10256 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
10257 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
10258 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
10259 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
10260 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
10261 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
10262 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
10263 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
10264 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
10265 BUILD_BUG_SQE_ELEM(42, __u16, personality);
10266 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
10268 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
10269 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
10270 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
10274 __initcall(io_uring_init);