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)
107 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
108 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
112 u32 head ____cacheline_aligned_in_smp;
113 u32 tail ____cacheline_aligned_in_smp;
117 * This data is shared with the application through the mmap at offsets
118 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
120 * The offsets to the member fields are published through struct
121 * io_sqring_offsets when calling io_uring_setup.
125 * Head and tail offsets into the ring; the offsets need to be
126 * masked to get valid indices.
128 * The kernel controls head of the sq ring and the tail of the cq ring,
129 * and the application controls tail of the sq ring and the head of the
132 struct io_uring sq, cq;
134 * Bitmasks to apply to head and tail offsets (constant, equals
137 u32 sq_ring_mask, cq_ring_mask;
138 /* Ring sizes (constant, power of 2) */
139 u32 sq_ring_entries, cq_ring_entries;
141 * Number of invalid entries dropped by the kernel due to
142 * invalid index stored in array
144 * Written by the kernel, shouldn't be modified by the
145 * application (i.e. get number of "new events" by comparing to
148 * After a new SQ head value was read by the application this
149 * counter includes all submissions that were dropped reaching
150 * the new SQ head (and possibly more).
156 * Written by the kernel, shouldn't be modified by the
159 * The application needs a full memory barrier before checking
160 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
166 * Written by the application, shouldn't be modified by the
171 * Number of completion events lost because the queue was full;
172 * this should be avoided by the application by making sure
173 * there are not more requests pending than there is space in
174 * the completion queue.
176 * Written by the kernel, shouldn't be modified by the
177 * application (i.e. get number of "new events" by comparing to
180 * As completion events come in out of order this counter is not
181 * ordered with any other data.
185 * Ring buffer of completion events.
187 * The kernel writes completion events fresh every time they are
188 * produced, so the application is allowed to modify pending
191 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
194 enum io_uring_cmd_flags {
195 IO_URING_F_NONBLOCK = 1,
196 IO_URING_F_COMPLETE_DEFER = 2,
199 struct io_mapped_ubuf {
202 struct bio_vec *bvec;
203 unsigned int nr_bvecs;
204 unsigned long acct_pages;
210 struct list_head list;
217 struct fixed_rsrc_table {
221 struct fixed_rsrc_ref_node {
222 struct percpu_ref refs;
223 struct list_head node;
224 struct list_head rsrc_list;
225 struct fixed_rsrc_data *rsrc_data;
226 void (*rsrc_put)(struct io_ring_ctx *ctx,
227 struct io_rsrc_put *prsrc);
228 struct llist_node llist;
232 struct fixed_rsrc_data {
233 struct fixed_rsrc_table *table;
234 struct io_ring_ctx *ctx;
236 struct fixed_rsrc_ref_node *node;
237 struct percpu_ref refs;
238 struct completion done;
242 struct list_head list;
248 struct io_restriction {
249 DECLARE_BITMAP(register_op, IORING_REGISTER_LAST);
250 DECLARE_BITMAP(sqe_op, IORING_OP_LAST);
251 u8 sqe_flags_allowed;
252 u8 sqe_flags_required;
260 /* ctx's that are using this sqd */
261 struct list_head ctx_list;
262 struct list_head ctx_new_list;
263 struct mutex ctx_lock;
265 struct task_struct *thread;
266 struct wait_queue_head wait;
268 unsigned sq_thread_idle;
271 #define IO_IOPOLL_BATCH 8
272 #define IO_COMPL_BATCH 32
273 #define IO_REQ_CACHE_SIZE 32
274 #define IO_REQ_ALLOC_BATCH 8
276 struct io_comp_state {
277 struct io_kiocb *reqs[IO_COMPL_BATCH];
279 unsigned int locked_free_nr;
280 /* inline/task_work completion list, under ->uring_lock */
281 struct list_head free_list;
282 /* IRQ completion list, under ->completion_lock */
283 struct list_head locked_free_list;
286 struct io_submit_link {
287 struct io_kiocb *head;
288 struct io_kiocb *last;
291 struct io_submit_state {
292 struct blk_plug plug;
293 struct io_submit_link link;
296 * io_kiocb alloc cache
298 void *reqs[IO_REQ_CACHE_SIZE];
299 unsigned int free_reqs;
304 * Batch completion logic
306 struct io_comp_state comp;
309 * File reference cache
313 unsigned int file_refs;
314 unsigned int ios_left;
319 struct percpu_ref refs;
320 } ____cacheline_aligned_in_smp;
324 unsigned int compat: 1;
325 unsigned int limit_mem: 1;
326 unsigned int cq_overflow_flushed: 1;
327 unsigned int drain_next: 1;
328 unsigned int eventfd_async: 1;
329 unsigned int restricted: 1;
330 unsigned int sqo_dead: 1;
333 * Ring buffer of indices into array of io_uring_sqe, which is
334 * mmapped by the application using the IORING_OFF_SQES offset.
336 * This indirection could e.g. be used to assign fixed
337 * io_uring_sqe entries to operations and only submit them to
338 * the queue when needed.
340 * The kernel modifies neither the indices array nor the entries
344 unsigned cached_sq_head;
347 unsigned sq_thread_idle;
348 unsigned cached_sq_dropped;
349 unsigned cached_cq_overflow;
350 unsigned long sq_check_overflow;
352 struct list_head defer_list;
353 struct list_head timeout_list;
354 struct list_head cq_overflow_list;
356 struct io_uring_sqe *sq_sqes;
357 } ____cacheline_aligned_in_smp;
360 struct mutex uring_lock;
361 wait_queue_head_t wait;
362 } ____cacheline_aligned_in_smp;
364 struct io_submit_state submit_state;
366 struct io_rings *rings;
372 * For SQPOLL usage - we hold a reference to the parent task, so we
373 * have access to the ->files
375 struct task_struct *sqo_task;
377 /* Only used for accounting purposes */
378 struct mm_struct *mm_account;
380 #ifdef CONFIG_BLK_CGROUP
381 struct cgroup_subsys_state *sqo_blkcg_css;
384 struct io_sq_data *sq_data; /* if using sq thread polling */
386 struct wait_queue_head sqo_sq_wait;
387 struct list_head sqd_list;
390 * If used, fixed file set. Writers must ensure that ->refs is dead,
391 * readers must ensure that ->refs is alive as long as the file* is
392 * used. Only updated through io_uring_register(2).
394 struct fixed_rsrc_data *file_data;
395 unsigned nr_user_files;
397 /* if used, fixed mapped user buffers */
398 unsigned nr_user_bufs;
399 struct io_mapped_ubuf *user_bufs;
401 struct user_struct *user;
403 const struct cred *creds;
407 unsigned int sessionid;
410 struct completion ref_comp;
411 struct completion sq_thread_comp;
413 #if defined(CONFIG_UNIX)
414 struct socket *ring_sock;
417 struct idr io_buffer_idr;
419 struct idr personality_idr;
422 unsigned cached_cq_tail;
425 atomic_t cq_timeouts;
426 unsigned cq_last_tm_flush;
427 unsigned long cq_check_overflow;
428 struct wait_queue_head cq_wait;
429 struct fasync_struct *cq_fasync;
430 struct eventfd_ctx *cq_ev_fd;
431 } ____cacheline_aligned_in_smp;
434 spinlock_t completion_lock;
437 * ->iopoll_list is protected by the ctx->uring_lock for
438 * io_uring instances that don't use IORING_SETUP_SQPOLL.
439 * For SQPOLL, only the single threaded io_sq_thread() will
440 * manipulate the list, hence no extra locking is needed there.
442 struct list_head iopoll_list;
443 struct hlist_head *cancel_hash;
444 unsigned cancel_hash_bits;
445 bool poll_multi_file;
447 spinlock_t inflight_lock;
448 struct list_head inflight_list;
449 } ____cacheline_aligned_in_smp;
451 struct delayed_work rsrc_put_work;
452 struct llist_head rsrc_put_llist;
453 struct list_head rsrc_ref_list;
454 spinlock_t rsrc_ref_lock;
456 struct io_restriction restrictions;
458 /* Keep this last, we don't need it for the fast path */
459 struct work_struct exit_work;
463 * First field must be the file pointer in all the
464 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
466 struct io_poll_iocb {
468 struct wait_queue_head *head;
472 struct wait_queue_entry wait;
475 struct io_poll_remove {
485 struct io_timeout_data {
486 struct io_kiocb *req;
487 struct hrtimer timer;
488 struct timespec64 ts;
489 enum hrtimer_mode mode;
494 struct sockaddr __user *addr;
495 int __user *addr_len;
497 unsigned long nofile;
517 struct list_head list;
518 /* head of the link, used by linked timeouts only */
519 struct io_kiocb *head;
522 struct io_timeout_rem {
527 struct timespec64 ts;
532 /* NOTE: kiocb has the file as the first member, so don't do it here */
540 struct sockaddr __user *addr;
547 struct user_msghdr __user *umsg;
553 struct io_buffer *kbuf;
559 struct filename *filename;
561 unsigned long nofile;
564 struct io_rsrc_update {
590 struct epoll_event event;
594 struct file *file_out;
595 struct file *file_in;
602 struct io_provide_buf {
616 const char __user *filename;
617 struct statx __user *buffer;
629 struct filename *oldpath;
630 struct filename *newpath;
638 struct filename *filename;
641 struct io_completion {
643 struct list_head list;
647 struct io_async_connect {
648 struct sockaddr_storage address;
651 struct io_async_msghdr {
652 struct iovec fast_iov[UIO_FASTIOV];
653 /* points to an allocated iov, if NULL we use fast_iov instead */
654 struct iovec *free_iov;
655 struct sockaddr __user *uaddr;
657 struct sockaddr_storage addr;
661 struct iovec fast_iov[UIO_FASTIOV];
662 const struct iovec *free_iovec;
663 struct iov_iter iter;
665 struct wait_page_queue wpq;
669 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
670 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
671 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
672 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
673 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
674 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
680 REQ_F_LINK_TIMEOUT_BIT,
682 REQ_F_NEED_CLEANUP_BIT,
684 REQ_F_BUFFER_SELECTED_BIT,
685 REQ_F_NO_FILE_TABLE_BIT,
686 REQ_F_WORK_INITIALIZED_BIT,
687 REQ_F_LTIMEOUT_ACTIVE_BIT,
688 REQ_F_COMPLETE_INLINE_BIT,
690 /* not a real bit, just to check we're not overflowing the space */
696 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
697 /* drain existing IO first */
698 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
700 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
701 /* doesn't sever on completion < 0 */
702 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
704 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
705 /* IOSQE_BUFFER_SELECT */
706 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
708 /* fail rest of links */
709 REQ_F_FAIL_LINK = BIT(REQ_F_FAIL_LINK_BIT),
710 /* on inflight list */
711 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
712 /* read/write uses file position */
713 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
714 /* must not punt to workers */
715 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
716 /* has or had linked timeout */
717 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
719 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
721 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
722 /* already went through poll handler */
723 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
724 /* buffer already selected */
725 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
726 /* doesn't need file table for this request */
727 REQ_F_NO_FILE_TABLE = BIT(REQ_F_NO_FILE_TABLE_BIT),
728 /* io_wq_work is initialized */
729 REQ_F_WORK_INITIALIZED = BIT(REQ_F_WORK_INITIALIZED_BIT),
730 /* linked timeout is active, i.e. prepared by link's head */
731 REQ_F_LTIMEOUT_ACTIVE = BIT(REQ_F_LTIMEOUT_ACTIVE_BIT),
732 /* completion is deferred through io_comp_state */
733 REQ_F_COMPLETE_INLINE = BIT(REQ_F_COMPLETE_INLINE_BIT),
737 struct io_poll_iocb poll;
738 struct io_poll_iocb *double_poll;
741 struct io_task_work {
742 struct io_wq_work_node node;
743 task_work_func_t func;
747 * NOTE! Each of the iocb union members has the file pointer
748 * as the first entry in their struct definition. So you can
749 * access the file pointer through any of the sub-structs,
750 * or directly as just 'ki_filp' in this struct.
756 struct io_poll_iocb poll;
757 struct io_poll_remove poll_remove;
758 struct io_accept accept;
760 struct io_cancel cancel;
761 struct io_timeout timeout;
762 struct io_timeout_rem timeout_rem;
763 struct io_connect connect;
764 struct io_sr_msg sr_msg;
766 struct io_close close;
767 struct io_rsrc_update rsrc_update;
768 struct io_fadvise fadvise;
769 struct io_madvise madvise;
770 struct io_epoll epoll;
771 struct io_splice splice;
772 struct io_provide_buf pbuf;
773 struct io_statx statx;
774 struct io_shutdown shutdown;
775 struct io_rename rename;
776 struct io_unlink unlink;
777 /* use only after cleaning per-op data, see io_clean_op() */
778 struct io_completion compl;
781 /* opcode allocated if it needs to store data for async defer */
784 /* polled IO has completed */
790 struct io_ring_ctx *ctx;
793 struct task_struct *task;
796 struct io_kiocb *link;
797 struct percpu_ref *fixed_rsrc_refs;
800 * 1. used with ctx->iopoll_list with reads/writes
801 * 2. to track reqs with ->files (see io_op_def::file_table)
803 struct list_head inflight_entry;
805 struct io_task_work io_task_work;
806 struct callback_head task_work;
808 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
809 struct hlist_node hash_node;
810 struct async_poll *apoll;
811 struct io_wq_work work;
814 struct io_defer_entry {
815 struct list_head list;
816 struct io_kiocb *req;
821 /* needs req->file assigned */
822 unsigned needs_file : 1;
823 /* hash wq insertion if file is a regular file */
824 unsigned hash_reg_file : 1;
825 /* unbound wq insertion if file is a non-regular file */
826 unsigned unbound_nonreg_file : 1;
827 /* opcode is not supported by this kernel */
828 unsigned not_supported : 1;
829 /* set if opcode supports polled "wait" */
831 unsigned pollout : 1;
832 /* op supports buffer selection */
833 unsigned buffer_select : 1;
834 /* must always have async data allocated */
835 unsigned needs_async_data : 1;
836 /* should block plug */
838 /* size of async data needed, if any */
839 unsigned short async_size;
843 static const struct io_op_def io_op_defs[] = {
844 [IORING_OP_NOP] = {},
845 [IORING_OP_READV] = {
847 .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,
855 [IORING_OP_WRITEV] = {
858 .unbound_nonreg_file = 1,
860 .needs_async_data = 1,
862 .async_size = sizeof(struct io_async_rw),
863 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG |
866 [IORING_OP_FSYNC] = {
868 .work_flags = IO_WQ_WORK_BLKCG,
870 [IORING_OP_READ_FIXED] = {
872 .unbound_nonreg_file = 1,
875 .async_size = sizeof(struct io_async_rw),
876 .work_flags = IO_WQ_WORK_BLKCG | IO_WQ_WORK_MM,
878 [IORING_OP_WRITE_FIXED] = {
881 .unbound_nonreg_file = 1,
884 .async_size = sizeof(struct io_async_rw),
885 .work_flags = IO_WQ_WORK_BLKCG | IO_WQ_WORK_FSIZE |
888 [IORING_OP_POLL_ADD] = {
890 .unbound_nonreg_file = 1,
892 [IORING_OP_POLL_REMOVE] = {},
893 [IORING_OP_SYNC_FILE_RANGE] = {
895 .work_flags = IO_WQ_WORK_BLKCG,
897 [IORING_OP_SENDMSG] = {
899 .unbound_nonreg_file = 1,
901 .needs_async_data = 1,
902 .async_size = sizeof(struct io_async_msghdr),
903 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
905 [IORING_OP_RECVMSG] = {
907 .unbound_nonreg_file = 1,
910 .needs_async_data = 1,
911 .async_size = sizeof(struct io_async_msghdr),
912 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
914 [IORING_OP_TIMEOUT] = {
915 .needs_async_data = 1,
916 .async_size = sizeof(struct io_timeout_data),
917 .work_flags = IO_WQ_WORK_MM,
919 [IORING_OP_TIMEOUT_REMOVE] = {
920 /* used by timeout updates' prep() */
921 .work_flags = IO_WQ_WORK_MM,
923 [IORING_OP_ACCEPT] = {
925 .unbound_nonreg_file = 1,
927 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_FILES,
929 [IORING_OP_ASYNC_CANCEL] = {},
930 [IORING_OP_LINK_TIMEOUT] = {
931 .needs_async_data = 1,
932 .async_size = sizeof(struct io_timeout_data),
933 .work_flags = IO_WQ_WORK_MM,
935 [IORING_OP_CONNECT] = {
937 .unbound_nonreg_file = 1,
939 .needs_async_data = 1,
940 .async_size = sizeof(struct io_async_connect),
941 .work_flags = IO_WQ_WORK_MM,
943 [IORING_OP_FALLOCATE] = {
945 .work_flags = IO_WQ_WORK_BLKCG | IO_WQ_WORK_FSIZE,
947 [IORING_OP_OPENAT] = {
948 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_BLKCG |
949 IO_WQ_WORK_FS | IO_WQ_WORK_MM,
951 [IORING_OP_CLOSE] = {
952 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_BLKCG,
954 [IORING_OP_FILES_UPDATE] = {
955 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_MM,
957 [IORING_OP_STATX] = {
958 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_MM |
959 IO_WQ_WORK_FS | IO_WQ_WORK_BLKCG,
963 .unbound_nonreg_file = 1,
967 .async_size = sizeof(struct io_async_rw),
968 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
970 [IORING_OP_WRITE] = {
972 .unbound_nonreg_file = 1,
975 .async_size = sizeof(struct io_async_rw),
976 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG |
979 [IORING_OP_FADVISE] = {
981 .work_flags = IO_WQ_WORK_BLKCG,
983 [IORING_OP_MADVISE] = {
984 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
988 .unbound_nonreg_file = 1,
990 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
994 .unbound_nonreg_file = 1,
997 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
999 [IORING_OP_OPENAT2] = {
1000 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_FS |
1001 IO_WQ_WORK_BLKCG | IO_WQ_WORK_MM,
1003 [IORING_OP_EPOLL_CTL] = {
1004 .unbound_nonreg_file = 1,
1005 .work_flags = IO_WQ_WORK_FILES,
1007 [IORING_OP_SPLICE] = {
1010 .unbound_nonreg_file = 1,
1011 .work_flags = IO_WQ_WORK_BLKCG,
1013 [IORING_OP_PROVIDE_BUFFERS] = {},
1014 [IORING_OP_REMOVE_BUFFERS] = {},
1018 .unbound_nonreg_file = 1,
1020 [IORING_OP_SHUTDOWN] = {
1023 [IORING_OP_RENAMEAT] = {
1024 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_FILES |
1025 IO_WQ_WORK_FS | IO_WQ_WORK_BLKCG,
1027 [IORING_OP_UNLINKAT] = {
1028 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_FILES |
1029 IO_WQ_WORK_FS | IO_WQ_WORK_BLKCG,
1033 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
1034 struct task_struct *task,
1035 struct files_struct *files);
1036 static void destroy_fixed_rsrc_ref_node(struct fixed_rsrc_ref_node *ref_node);
1037 static struct fixed_rsrc_ref_node *alloc_fixed_rsrc_ref_node(
1038 struct io_ring_ctx *ctx);
1039 static void init_fixed_file_ref_node(struct io_ring_ctx *ctx,
1040 struct fixed_rsrc_ref_node *ref_node);
1042 static bool io_rw_reissue(struct io_kiocb *req);
1043 static void io_cqring_fill_event(struct io_kiocb *req, long res);
1044 static void io_put_req(struct io_kiocb *req);
1045 static void io_put_req_deferred(struct io_kiocb *req, int nr);
1046 static void io_double_put_req(struct io_kiocb *req);
1047 static void io_dismantle_req(struct io_kiocb *req);
1048 static void io_put_task(struct task_struct *task, int nr);
1049 static void io_queue_next(struct io_kiocb *req);
1050 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
1051 static void __io_queue_linked_timeout(struct io_kiocb *req);
1052 static void io_queue_linked_timeout(struct io_kiocb *req);
1053 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
1054 struct io_uring_rsrc_update *ip,
1056 static void __io_clean_op(struct io_kiocb *req);
1057 static struct file *io_file_get(struct io_submit_state *state,
1058 struct io_kiocb *req, int fd, bool fixed);
1059 static void __io_queue_sqe(struct io_kiocb *req);
1060 static void io_rsrc_put_work(struct work_struct *work);
1062 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
1063 struct iov_iter *iter, bool needs_lock);
1064 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
1065 const struct iovec *fast_iov,
1066 struct iov_iter *iter, bool force);
1067 static void io_req_task_queue(struct io_kiocb *req);
1068 static void io_submit_flush_completions(struct io_comp_state *cs,
1069 struct io_ring_ctx *ctx);
1071 static struct kmem_cache *req_cachep;
1073 static const struct file_operations io_uring_fops;
1075 struct sock *io_uring_get_socket(struct file *file)
1077 #if defined(CONFIG_UNIX)
1078 if (file->f_op == &io_uring_fops) {
1079 struct io_ring_ctx *ctx = file->private_data;
1081 return ctx->ring_sock->sk;
1086 EXPORT_SYMBOL(io_uring_get_socket);
1088 #define io_for_each_link(pos, head) \
1089 for (pos = (head); pos; pos = pos->link)
1091 static inline void io_clean_op(struct io_kiocb *req)
1093 if (req->flags & (REQ_F_NEED_CLEANUP | REQ_F_BUFFER_SELECTED))
1097 static inline void io_set_resource_node(struct io_kiocb *req)
1099 struct io_ring_ctx *ctx = req->ctx;
1101 if (!req->fixed_rsrc_refs) {
1102 req->fixed_rsrc_refs = &ctx->file_data->node->refs;
1103 percpu_ref_get(req->fixed_rsrc_refs);
1107 static bool io_match_task(struct io_kiocb *head,
1108 struct task_struct *task,
1109 struct files_struct *files)
1111 struct io_kiocb *req;
1113 if (task && head->task != task) {
1114 /* in terms of cancelation, always match if req task is dead */
1115 if (head->task->flags & PF_EXITING)
1122 io_for_each_link(req, head) {
1123 if (!(req->flags & REQ_F_WORK_INITIALIZED))
1125 if (req->file && req->file->f_op == &io_uring_fops)
1127 if ((req->work.flags & IO_WQ_WORK_FILES) &&
1128 req->work.identity->files == files)
1134 static void io_sq_thread_drop_mm_files(void)
1136 struct files_struct *files = current->files;
1137 struct mm_struct *mm = current->mm;
1140 kthread_unuse_mm(mm);
1145 struct nsproxy *nsproxy = current->nsproxy;
1148 current->files = NULL;
1149 current->nsproxy = NULL;
1150 task_unlock(current);
1151 put_files_struct(files);
1152 put_nsproxy(nsproxy);
1156 static int __io_sq_thread_acquire_files(struct io_ring_ctx *ctx)
1158 if (!current->files) {
1159 struct files_struct *files;
1160 struct nsproxy *nsproxy;
1162 task_lock(ctx->sqo_task);
1163 files = ctx->sqo_task->files;
1165 task_unlock(ctx->sqo_task);
1168 atomic_inc(&files->count);
1169 get_nsproxy(ctx->sqo_task->nsproxy);
1170 nsproxy = ctx->sqo_task->nsproxy;
1171 task_unlock(ctx->sqo_task);
1174 current->files = files;
1175 current->nsproxy = nsproxy;
1176 task_unlock(current);
1181 static int __io_sq_thread_acquire_mm(struct io_ring_ctx *ctx)
1183 struct mm_struct *mm;
1188 task_lock(ctx->sqo_task);
1189 mm = ctx->sqo_task->mm;
1190 if (unlikely(!mm || !mmget_not_zero(mm)))
1192 task_unlock(ctx->sqo_task);
1202 static int __io_sq_thread_acquire_mm_files(struct io_ring_ctx *ctx,
1203 struct io_kiocb *req)
1205 const struct io_op_def *def = &io_op_defs[req->opcode];
1208 if (def->work_flags & IO_WQ_WORK_MM) {
1209 ret = __io_sq_thread_acquire_mm(ctx);
1214 if (def->needs_file || (def->work_flags & IO_WQ_WORK_FILES)) {
1215 ret = __io_sq_thread_acquire_files(ctx);
1223 static inline int io_sq_thread_acquire_mm_files(struct io_ring_ctx *ctx,
1224 struct io_kiocb *req)
1226 if (!(ctx->flags & IORING_SETUP_SQPOLL))
1228 return __io_sq_thread_acquire_mm_files(ctx, req);
1231 static void io_sq_thread_associate_blkcg(struct io_ring_ctx *ctx,
1232 struct cgroup_subsys_state **cur_css)
1235 #ifdef CONFIG_BLK_CGROUP
1236 /* puts the old one when swapping */
1237 if (*cur_css != ctx->sqo_blkcg_css) {
1238 kthread_associate_blkcg(ctx->sqo_blkcg_css);
1239 *cur_css = ctx->sqo_blkcg_css;
1244 static void io_sq_thread_unassociate_blkcg(void)
1246 #ifdef CONFIG_BLK_CGROUP
1247 kthread_associate_blkcg(NULL);
1251 static inline void req_set_fail_links(struct io_kiocb *req)
1253 if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
1254 req->flags |= REQ_F_FAIL_LINK;
1258 * None of these are dereferenced, they are simply used to check if any of
1259 * them have changed. If we're under current and check they are still the
1260 * same, we're fine to grab references to them for actual out-of-line use.
1262 static void io_init_identity(struct io_identity *id)
1264 id->files = current->files;
1265 id->mm = current->mm;
1266 #ifdef CONFIG_BLK_CGROUP
1268 id->blkcg_css = blkcg_css();
1271 id->creds = current_cred();
1272 id->nsproxy = current->nsproxy;
1273 id->fs = current->fs;
1274 id->fsize = rlimit(RLIMIT_FSIZE);
1276 id->loginuid = current->loginuid;
1277 id->sessionid = current->sessionid;
1279 refcount_set(&id->count, 1);
1282 static inline void __io_req_init_async(struct io_kiocb *req)
1284 memset(&req->work, 0, sizeof(req->work));
1285 req->flags |= REQ_F_WORK_INITIALIZED;
1289 * Note: must call io_req_init_async() for the first time you
1290 * touch any members of io_wq_work.
1292 static inline void io_req_init_async(struct io_kiocb *req)
1294 struct io_uring_task *tctx = current->io_uring;
1296 if (req->flags & REQ_F_WORK_INITIALIZED)
1299 __io_req_init_async(req);
1301 /* Grab a ref if this isn't our static identity */
1302 req->work.identity = tctx->identity;
1303 if (tctx->identity != &tctx->__identity)
1304 refcount_inc(&req->work.identity->count);
1307 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1309 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1311 complete(&ctx->ref_comp);
1314 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1316 return !req->timeout.off;
1319 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1321 struct io_ring_ctx *ctx;
1324 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1329 * Use 5 bits less than the max cq entries, that should give us around
1330 * 32 entries per hash list if totally full and uniformly spread.
1332 hash_bits = ilog2(p->cq_entries);
1336 ctx->cancel_hash_bits = hash_bits;
1337 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1339 if (!ctx->cancel_hash)
1341 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1343 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1344 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1347 ctx->flags = p->flags;
1348 init_waitqueue_head(&ctx->sqo_sq_wait);
1349 INIT_LIST_HEAD(&ctx->sqd_list);
1350 init_waitqueue_head(&ctx->cq_wait);
1351 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1352 init_completion(&ctx->ref_comp);
1353 init_completion(&ctx->sq_thread_comp);
1354 idr_init(&ctx->io_buffer_idr);
1355 idr_init(&ctx->personality_idr);
1356 mutex_init(&ctx->uring_lock);
1357 init_waitqueue_head(&ctx->wait);
1358 spin_lock_init(&ctx->completion_lock);
1359 INIT_LIST_HEAD(&ctx->iopoll_list);
1360 INIT_LIST_HEAD(&ctx->defer_list);
1361 INIT_LIST_HEAD(&ctx->timeout_list);
1362 spin_lock_init(&ctx->inflight_lock);
1363 INIT_LIST_HEAD(&ctx->inflight_list);
1364 spin_lock_init(&ctx->rsrc_ref_lock);
1365 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
1366 INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
1367 init_llist_head(&ctx->rsrc_put_llist);
1368 INIT_LIST_HEAD(&ctx->submit_state.comp.free_list);
1369 INIT_LIST_HEAD(&ctx->submit_state.comp.locked_free_list);
1372 kfree(ctx->cancel_hash);
1377 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1379 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1380 struct io_ring_ctx *ctx = req->ctx;
1382 return seq != ctx->cached_cq_tail
1383 + READ_ONCE(ctx->cached_cq_overflow);
1389 static void io_put_identity(struct io_uring_task *tctx, struct io_kiocb *req)
1391 if (req->work.identity == &tctx->__identity)
1393 if (refcount_dec_and_test(&req->work.identity->count))
1394 kfree(req->work.identity);
1397 static void io_req_clean_work(struct io_kiocb *req)
1399 if (!(req->flags & REQ_F_WORK_INITIALIZED))
1402 if (req->work.flags & IO_WQ_WORK_MM)
1403 mmdrop(req->work.identity->mm);
1404 #ifdef CONFIG_BLK_CGROUP
1405 if (req->work.flags & IO_WQ_WORK_BLKCG)
1406 css_put(req->work.identity->blkcg_css);
1408 if (req->work.flags & IO_WQ_WORK_CREDS)
1409 put_cred(req->work.identity->creds);
1410 if (req->work.flags & IO_WQ_WORK_FS) {
1411 struct fs_struct *fs = req->work.identity->fs;
1413 spin_lock(&req->work.identity->fs->lock);
1416 spin_unlock(&req->work.identity->fs->lock);
1420 if (req->work.flags & IO_WQ_WORK_FILES) {
1421 put_files_struct(req->work.identity->files);
1422 put_nsproxy(req->work.identity->nsproxy);
1424 if (req->flags & REQ_F_INFLIGHT) {
1425 struct io_ring_ctx *ctx = req->ctx;
1426 struct io_uring_task *tctx = req->task->io_uring;
1427 unsigned long flags;
1429 spin_lock_irqsave(&ctx->inflight_lock, flags);
1430 list_del(&req->inflight_entry);
1431 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
1432 req->flags &= ~REQ_F_INFLIGHT;
1433 if (atomic_read(&tctx->in_idle))
1434 wake_up(&tctx->wait);
1437 req->flags &= ~REQ_F_WORK_INITIALIZED;
1438 req->work.flags &= ~(IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG | IO_WQ_WORK_FS |
1439 IO_WQ_WORK_CREDS | IO_WQ_WORK_FILES);
1440 io_put_identity(req->task->io_uring, req);
1444 * Create a private copy of io_identity, since some fields don't match
1445 * the current context.
1447 static bool io_identity_cow(struct io_kiocb *req)
1449 struct io_uring_task *tctx = current->io_uring;
1450 const struct cred *creds = NULL;
1451 struct io_identity *id;
1453 if (req->work.flags & IO_WQ_WORK_CREDS)
1454 creds = req->work.identity->creds;
1456 id = kmemdup(req->work.identity, sizeof(*id), GFP_KERNEL);
1457 if (unlikely(!id)) {
1458 req->work.flags |= IO_WQ_WORK_CANCEL;
1463 * We can safely just re-init the creds we copied Either the field
1464 * matches the current one, or we haven't grabbed it yet. The only
1465 * exception is ->creds, through registered personalities, so handle
1466 * that one separately.
1468 io_init_identity(id);
1472 /* add one for this request */
1473 refcount_inc(&id->count);
1475 /* drop tctx and req identity references, if needed */
1476 if (tctx->identity != &tctx->__identity &&
1477 refcount_dec_and_test(&tctx->identity->count))
1478 kfree(tctx->identity);
1479 if (req->work.identity != &tctx->__identity &&
1480 refcount_dec_and_test(&req->work.identity->count))
1481 kfree(req->work.identity);
1483 req->work.identity = id;
1484 tctx->identity = id;
1488 static void io_req_track_inflight(struct io_kiocb *req)
1490 struct io_ring_ctx *ctx = req->ctx;
1492 if (!(req->flags & REQ_F_INFLIGHT)) {
1493 io_req_init_async(req);
1494 req->flags |= REQ_F_INFLIGHT;
1496 spin_lock_irq(&ctx->inflight_lock);
1497 list_add(&req->inflight_entry, &ctx->inflight_list);
1498 spin_unlock_irq(&ctx->inflight_lock);
1502 static bool io_grab_identity(struct io_kiocb *req)
1504 const struct io_op_def *def = &io_op_defs[req->opcode];
1505 struct io_identity *id = req->work.identity;
1507 if (def->work_flags & IO_WQ_WORK_FSIZE) {
1508 if (id->fsize != rlimit(RLIMIT_FSIZE))
1510 req->work.flags |= IO_WQ_WORK_FSIZE;
1512 #ifdef CONFIG_BLK_CGROUP
1513 if (!(req->work.flags & IO_WQ_WORK_BLKCG) &&
1514 (def->work_flags & IO_WQ_WORK_BLKCG)) {
1516 if (id->blkcg_css != blkcg_css()) {
1521 * This should be rare, either the cgroup is dying or the task
1522 * is moving cgroups. Just punt to root for the handful of ios.
1524 if (css_tryget_online(id->blkcg_css))
1525 req->work.flags |= IO_WQ_WORK_BLKCG;
1529 if (!(req->work.flags & IO_WQ_WORK_CREDS)) {
1530 if (id->creds != current_cred())
1532 get_cred(id->creds);
1533 req->work.flags |= IO_WQ_WORK_CREDS;
1536 if (!uid_eq(current->loginuid, id->loginuid) ||
1537 current->sessionid != id->sessionid)
1540 if (!(req->work.flags & IO_WQ_WORK_FS) &&
1541 (def->work_flags & IO_WQ_WORK_FS)) {
1542 if (current->fs != id->fs)
1544 spin_lock(&id->fs->lock);
1545 if (!id->fs->in_exec) {
1547 req->work.flags |= IO_WQ_WORK_FS;
1549 req->work.flags |= IO_WQ_WORK_CANCEL;
1551 spin_unlock(¤t->fs->lock);
1553 if (!(req->work.flags & IO_WQ_WORK_FILES) &&
1554 (def->work_flags & IO_WQ_WORK_FILES) &&
1555 !(req->flags & REQ_F_NO_FILE_TABLE)) {
1556 if (id->files != current->files ||
1557 id->nsproxy != current->nsproxy)
1559 atomic_inc(&id->files->count);
1560 get_nsproxy(id->nsproxy);
1561 req->work.flags |= IO_WQ_WORK_FILES;
1562 io_req_track_inflight(req);
1564 if (!(req->work.flags & IO_WQ_WORK_MM) &&
1565 (def->work_flags & IO_WQ_WORK_MM)) {
1566 if (id->mm != current->mm)
1569 req->work.flags |= IO_WQ_WORK_MM;
1575 static void io_prep_async_work(struct io_kiocb *req)
1577 const struct io_op_def *def = &io_op_defs[req->opcode];
1578 struct io_ring_ctx *ctx = req->ctx;
1580 io_req_init_async(req);
1582 if (req->flags & REQ_F_FORCE_ASYNC)
1583 req->work.flags |= IO_WQ_WORK_CONCURRENT;
1585 if (req->flags & REQ_F_ISREG) {
1586 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1587 io_wq_hash_work(&req->work, file_inode(req->file));
1589 if (def->unbound_nonreg_file)
1590 req->work.flags |= IO_WQ_WORK_UNBOUND;
1593 /* if we fail grabbing identity, we must COW, regrab, and retry */
1594 if (io_grab_identity(req))
1597 if (!io_identity_cow(req))
1600 /* can't fail at this point */
1601 if (!io_grab_identity(req))
1605 static void io_prep_async_link(struct io_kiocb *req)
1607 struct io_kiocb *cur;
1609 io_for_each_link(cur, req)
1610 io_prep_async_work(cur);
1613 static struct io_kiocb *__io_queue_async_work(struct io_kiocb *req)
1615 struct io_ring_ctx *ctx = req->ctx;
1616 struct io_kiocb *link = io_prep_linked_timeout(req);
1618 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1619 &req->work, req->flags);
1620 io_wq_enqueue(ctx->io_wq, &req->work);
1624 static void io_queue_async_work(struct io_kiocb *req)
1626 struct io_kiocb *link;
1628 /* init ->work of the whole link before punting */
1629 io_prep_async_link(req);
1630 link = __io_queue_async_work(req);
1633 io_queue_linked_timeout(link);
1636 static void io_kill_timeout(struct io_kiocb *req)
1638 struct io_timeout_data *io = req->async_data;
1641 ret = hrtimer_try_to_cancel(&io->timer);
1643 atomic_set(&req->ctx->cq_timeouts,
1644 atomic_read(&req->ctx->cq_timeouts) + 1);
1645 list_del_init(&req->timeout.list);
1646 io_cqring_fill_event(req, 0);
1647 io_put_req_deferred(req, 1);
1652 * Returns true if we found and killed one or more timeouts
1654 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
1655 struct files_struct *files)
1657 struct io_kiocb *req, *tmp;
1660 spin_lock_irq(&ctx->completion_lock);
1661 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
1662 if (io_match_task(req, tsk, files)) {
1663 io_kill_timeout(req);
1667 spin_unlock_irq(&ctx->completion_lock);
1668 return canceled != 0;
1671 static void __io_queue_deferred(struct io_ring_ctx *ctx)
1674 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1675 struct io_defer_entry, list);
1677 if (req_need_defer(de->req, de->seq))
1679 list_del_init(&de->list);
1680 io_req_task_queue(de->req);
1682 } while (!list_empty(&ctx->defer_list));
1685 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1689 if (list_empty(&ctx->timeout_list))
1692 seq = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
1695 u32 events_needed, events_got;
1696 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1697 struct io_kiocb, timeout.list);
1699 if (io_is_timeout_noseq(req))
1703 * Since seq can easily wrap around over time, subtract
1704 * the last seq at which timeouts were flushed before comparing.
1705 * Assuming not more than 2^31-1 events have happened since,
1706 * these subtractions won't have wrapped, so we can check if
1707 * target is in [last_seq, current_seq] by comparing the two.
1709 events_needed = req->timeout.target_seq - ctx->cq_last_tm_flush;
1710 events_got = seq - ctx->cq_last_tm_flush;
1711 if (events_got < events_needed)
1714 list_del_init(&req->timeout.list);
1715 io_kill_timeout(req);
1716 } while (!list_empty(&ctx->timeout_list));
1718 ctx->cq_last_tm_flush = seq;
1721 static void io_commit_cqring(struct io_ring_ctx *ctx)
1723 io_flush_timeouts(ctx);
1725 /* order cqe stores with ring update */
1726 smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
1728 if (unlikely(!list_empty(&ctx->defer_list)))
1729 __io_queue_deferred(ctx);
1732 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1734 struct io_rings *r = ctx->rings;
1736 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == r->sq_ring_entries;
1739 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
1741 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
1744 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1746 struct io_rings *rings = ctx->rings;
1750 * writes to the cq entry need to come after reading head; the
1751 * control dependency is enough as we're using WRITE_ONCE to
1754 if (__io_cqring_events(ctx) == rings->cq_ring_entries)
1757 tail = ctx->cached_cq_tail++;
1758 return &rings->cqes[tail & ctx->cq_mask];
1761 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1765 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1767 if (!ctx->eventfd_async)
1769 return io_wq_current_is_worker();
1772 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1774 /* see waitqueue_active() comment */
1777 if (waitqueue_active(&ctx->wait))
1778 wake_up(&ctx->wait);
1779 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1780 wake_up(&ctx->sq_data->wait);
1781 if (io_should_trigger_evfd(ctx))
1782 eventfd_signal(ctx->cq_ev_fd, 1);
1783 if (waitqueue_active(&ctx->cq_wait)) {
1784 wake_up_interruptible(&ctx->cq_wait);
1785 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1789 static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx)
1791 /* see waitqueue_active() comment */
1794 if (ctx->flags & IORING_SETUP_SQPOLL) {
1795 if (waitqueue_active(&ctx->wait))
1796 wake_up(&ctx->wait);
1798 if (io_should_trigger_evfd(ctx))
1799 eventfd_signal(ctx->cq_ev_fd, 1);
1800 if (waitqueue_active(&ctx->cq_wait)) {
1801 wake_up_interruptible(&ctx->cq_wait);
1802 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1806 /* Returns true if there are no backlogged entries after the flush */
1807 static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force,
1808 struct task_struct *tsk,
1809 struct files_struct *files)
1811 struct io_rings *rings = ctx->rings;
1812 struct io_kiocb *req, *tmp;
1813 struct io_uring_cqe *cqe;
1814 unsigned long flags;
1815 bool all_flushed, posted;
1818 if (!force && __io_cqring_events(ctx) == rings->cq_ring_entries)
1822 spin_lock_irqsave(&ctx->completion_lock, flags);
1823 list_for_each_entry_safe(req, tmp, &ctx->cq_overflow_list, compl.list) {
1824 if (!io_match_task(req, tsk, files))
1827 cqe = io_get_cqring(ctx);
1831 list_move(&req->compl.list, &list);
1833 WRITE_ONCE(cqe->user_data, req->user_data);
1834 WRITE_ONCE(cqe->res, req->result);
1835 WRITE_ONCE(cqe->flags, req->compl.cflags);
1837 ctx->cached_cq_overflow++;
1838 WRITE_ONCE(ctx->rings->cq_overflow,
1839 ctx->cached_cq_overflow);
1844 all_flushed = list_empty(&ctx->cq_overflow_list);
1846 clear_bit(0, &ctx->sq_check_overflow);
1847 clear_bit(0, &ctx->cq_check_overflow);
1848 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
1852 io_commit_cqring(ctx);
1853 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1855 io_cqring_ev_posted(ctx);
1857 while (!list_empty(&list)) {
1858 req = list_first_entry(&list, struct io_kiocb, compl.list);
1859 list_del(&req->compl.list);
1866 static void io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force,
1867 struct task_struct *tsk,
1868 struct files_struct *files)
1870 if (test_bit(0, &ctx->cq_check_overflow)) {
1871 /* iopoll syncs against uring_lock, not completion_lock */
1872 if (ctx->flags & IORING_SETUP_IOPOLL)
1873 mutex_lock(&ctx->uring_lock);
1874 __io_cqring_overflow_flush(ctx, force, tsk, files);
1875 if (ctx->flags & IORING_SETUP_IOPOLL)
1876 mutex_unlock(&ctx->uring_lock);
1880 static void __io_cqring_fill_event(struct io_kiocb *req, long res, long cflags)
1882 struct io_ring_ctx *ctx = req->ctx;
1883 struct io_uring_cqe *cqe;
1885 trace_io_uring_complete(ctx, req->user_data, res);
1888 * If we can't get a cq entry, userspace overflowed the
1889 * submission (by quite a lot). Increment the overflow count in
1892 cqe = io_get_cqring(ctx);
1894 WRITE_ONCE(cqe->user_data, req->user_data);
1895 WRITE_ONCE(cqe->res, res);
1896 WRITE_ONCE(cqe->flags, cflags);
1897 } else if (ctx->cq_overflow_flushed ||
1898 atomic_read(&req->task->io_uring->in_idle)) {
1900 * If we're in ring overflow flush mode, or in task cancel mode,
1901 * then we cannot store the request for later flushing, we need
1902 * to drop it on the floor.
1904 ctx->cached_cq_overflow++;
1905 WRITE_ONCE(ctx->rings->cq_overflow, ctx->cached_cq_overflow);
1907 if (list_empty(&ctx->cq_overflow_list)) {
1908 set_bit(0, &ctx->sq_check_overflow);
1909 set_bit(0, &ctx->cq_check_overflow);
1910 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
1914 req->compl.cflags = cflags;
1915 refcount_inc(&req->refs);
1916 list_add_tail(&req->compl.list, &ctx->cq_overflow_list);
1920 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1922 __io_cqring_fill_event(req, res, 0);
1925 static inline void io_req_complete_post(struct io_kiocb *req, long res,
1926 unsigned int cflags)
1928 struct io_ring_ctx *ctx = req->ctx;
1929 unsigned long flags;
1931 spin_lock_irqsave(&ctx->completion_lock, flags);
1932 __io_cqring_fill_event(req, res, cflags);
1933 io_commit_cqring(ctx);
1935 * If we're the last reference to this request, add to our locked
1938 if (refcount_dec_and_test(&req->refs)) {
1939 struct io_comp_state *cs = &ctx->submit_state.comp;
1941 io_dismantle_req(req);
1942 io_put_task(req->task, 1);
1943 list_add(&req->compl.list, &cs->locked_free_list);
1944 cs->locked_free_nr++;
1947 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1949 io_cqring_ev_posted(ctx);
1952 percpu_ref_put(&ctx->refs);
1956 static void io_req_complete_state(struct io_kiocb *req, long res,
1957 unsigned int cflags)
1961 req->compl.cflags = cflags;
1962 req->flags |= REQ_F_COMPLETE_INLINE;
1965 static inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags,
1966 long res, unsigned cflags)
1968 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1969 io_req_complete_state(req, res, cflags);
1971 io_req_complete_post(req, res, cflags);
1974 static inline void io_req_complete(struct io_kiocb *req, long res)
1976 __io_req_complete(req, 0, res, 0);
1979 static bool io_flush_cached_reqs(struct io_ring_ctx *ctx)
1981 struct io_submit_state *state = &ctx->submit_state;
1982 struct io_comp_state *cs = &state->comp;
1983 struct io_kiocb *req = NULL;
1986 * If we have more than a batch's worth of requests in our IRQ side
1987 * locked cache, grab the lock and move them over to our submission
1990 if (READ_ONCE(cs->locked_free_nr) > IO_COMPL_BATCH) {
1991 spin_lock_irq(&ctx->completion_lock);
1992 list_splice_init(&cs->locked_free_list, &cs->free_list);
1993 cs->locked_free_nr = 0;
1994 spin_unlock_irq(&ctx->completion_lock);
1997 while (!list_empty(&cs->free_list)) {
1998 req = list_first_entry(&cs->free_list, struct io_kiocb,
2000 list_del(&req->compl.list);
2001 state->reqs[state->free_reqs++] = req;
2002 if (state->free_reqs == ARRAY_SIZE(state->reqs))
2009 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
2011 struct io_submit_state *state = &ctx->submit_state;
2013 BUILD_BUG_ON(IO_REQ_ALLOC_BATCH > ARRAY_SIZE(state->reqs));
2015 if (!state->free_reqs) {
2016 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
2019 if (io_flush_cached_reqs(ctx))
2022 ret = kmem_cache_alloc_bulk(req_cachep, gfp, IO_REQ_ALLOC_BATCH,
2026 * Bulk alloc is all-or-nothing. If we fail to get a batch,
2027 * retry single alloc to be on the safe side.
2029 if (unlikely(ret <= 0)) {
2030 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
2031 if (!state->reqs[0])
2035 state->free_reqs = ret;
2039 return state->reqs[state->free_reqs];
2042 static inline void io_put_file(struct io_kiocb *req, struct file *file,
2049 static void io_dismantle_req(struct io_kiocb *req)
2053 if (req->async_data)
2054 kfree(req->async_data);
2056 io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
2057 if (req->fixed_rsrc_refs)
2058 percpu_ref_put(req->fixed_rsrc_refs);
2059 io_req_clean_work(req);
2062 static inline void io_put_task(struct task_struct *task, int nr)
2064 struct io_uring_task *tctx = task->io_uring;
2066 percpu_counter_sub(&tctx->inflight, nr);
2067 if (unlikely(atomic_read(&tctx->in_idle)))
2068 wake_up(&tctx->wait);
2069 put_task_struct_many(task, nr);
2072 static void __io_free_req(struct io_kiocb *req)
2074 struct io_ring_ctx *ctx = req->ctx;
2076 io_dismantle_req(req);
2077 io_put_task(req->task, 1);
2079 kmem_cache_free(req_cachep, req);
2080 percpu_ref_put(&ctx->refs);
2083 static inline void io_remove_next_linked(struct io_kiocb *req)
2085 struct io_kiocb *nxt = req->link;
2087 req->link = nxt->link;
2091 static void io_kill_linked_timeout(struct io_kiocb *req)
2093 struct io_ring_ctx *ctx = req->ctx;
2094 struct io_kiocb *link;
2095 bool cancelled = false;
2096 unsigned long flags;
2098 spin_lock_irqsave(&ctx->completion_lock, flags);
2102 * Can happen if a linked timeout fired and link had been like
2103 * req -> link t-out -> link t-out [-> ...]
2105 if (link && (link->flags & REQ_F_LTIMEOUT_ACTIVE)) {
2106 struct io_timeout_data *io = link->async_data;
2109 io_remove_next_linked(req);
2110 link->timeout.head = NULL;
2111 ret = hrtimer_try_to_cancel(&io->timer);
2113 io_cqring_fill_event(link, -ECANCELED);
2114 io_commit_cqring(ctx);
2118 req->flags &= ~REQ_F_LINK_TIMEOUT;
2119 spin_unlock_irqrestore(&ctx->completion_lock, flags);
2122 io_cqring_ev_posted(ctx);
2128 static void io_fail_links(struct io_kiocb *req)
2130 struct io_kiocb *link, *nxt;
2131 struct io_ring_ctx *ctx = req->ctx;
2132 unsigned long flags;
2134 spin_lock_irqsave(&ctx->completion_lock, flags);
2142 trace_io_uring_fail_link(req, link);
2143 io_cqring_fill_event(link, -ECANCELED);
2146 * It's ok to free under spinlock as they're not linked anymore,
2147 * but avoid REQ_F_WORK_INITIALIZED because it may deadlock on
2150 if (link->flags & REQ_F_WORK_INITIALIZED)
2151 io_put_req_deferred(link, 2);
2153 io_double_put_req(link);
2156 io_commit_cqring(ctx);
2157 spin_unlock_irqrestore(&ctx->completion_lock, flags);
2159 io_cqring_ev_posted(ctx);
2162 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
2164 if (req->flags & REQ_F_LINK_TIMEOUT)
2165 io_kill_linked_timeout(req);
2168 * If LINK is set, we have dependent requests in this chain. If we
2169 * didn't fail this request, queue the first one up, moving any other
2170 * dependencies to the next request. In case of failure, fail the rest
2173 if (likely(!(req->flags & REQ_F_FAIL_LINK))) {
2174 struct io_kiocb *nxt = req->link;
2183 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
2185 if (likely(!(req->flags & (REQ_F_LINK|REQ_F_HARDLINK))))
2187 return __io_req_find_next(req);
2190 static bool __tctx_task_work(struct io_uring_task *tctx)
2192 struct io_ring_ctx *ctx = NULL;
2193 struct io_wq_work_list list;
2194 struct io_wq_work_node *node;
2196 if (wq_list_empty(&tctx->task_list))
2199 spin_lock_irq(&tctx->task_lock);
2200 list = tctx->task_list;
2201 INIT_WQ_LIST(&tctx->task_list);
2202 spin_unlock_irq(&tctx->task_lock);
2206 struct io_wq_work_node *next = node->next;
2207 struct io_ring_ctx *this_ctx;
2208 struct io_kiocb *req;
2210 req = container_of(node, struct io_kiocb, io_task_work.node);
2211 this_ctx = req->ctx;
2212 req->task_work.func(&req->task_work);
2217 } else if (ctx != this_ctx) {
2218 mutex_lock(&ctx->uring_lock);
2219 io_submit_flush_completions(&ctx->submit_state.comp, ctx);
2220 mutex_unlock(&ctx->uring_lock);
2225 if (ctx && ctx->submit_state.comp.nr) {
2226 mutex_lock(&ctx->uring_lock);
2227 io_submit_flush_completions(&ctx->submit_state.comp, ctx);
2228 mutex_unlock(&ctx->uring_lock);
2231 return list.first != NULL;
2234 static void tctx_task_work(struct callback_head *cb)
2236 struct io_uring_task *tctx = container_of(cb, struct io_uring_task, task_work);
2238 while (__tctx_task_work(tctx))
2241 clear_bit(0, &tctx->task_state);
2244 static int io_task_work_add(struct task_struct *tsk, struct io_kiocb *req,
2245 enum task_work_notify_mode notify)
2247 struct io_uring_task *tctx = tsk->io_uring;
2248 struct io_wq_work_node *node, *prev;
2249 unsigned long flags;
2252 WARN_ON_ONCE(!tctx);
2254 spin_lock_irqsave(&tctx->task_lock, flags);
2255 wq_list_add_tail(&req->io_task_work.node, &tctx->task_list);
2256 spin_unlock_irqrestore(&tctx->task_lock, flags);
2258 /* task_work already pending, we're done */
2259 if (test_bit(0, &tctx->task_state) ||
2260 test_and_set_bit(0, &tctx->task_state))
2263 if (!task_work_add(tsk, &tctx->task_work, notify))
2267 * Slow path - we failed, find and delete work. if the work is not
2268 * in the list, it got run and we're fine.
2271 spin_lock_irqsave(&tctx->task_lock, flags);
2272 wq_list_for_each(node, prev, &tctx->task_list) {
2273 if (&req->io_task_work.node == node) {
2274 wq_list_del(&tctx->task_list, node, prev);
2279 spin_unlock_irqrestore(&tctx->task_lock, flags);
2280 clear_bit(0, &tctx->task_state);
2284 static int io_req_task_work_add(struct io_kiocb *req)
2286 struct task_struct *tsk = req->task;
2287 struct io_ring_ctx *ctx = req->ctx;
2288 enum task_work_notify_mode notify;
2291 if (tsk->flags & PF_EXITING)
2295 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
2296 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
2297 * processing task_work. There's no reliable way to tell if TWA_RESUME
2301 if (!(ctx->flags & IORING_SETUP_SQPOLL))
2302 notify = TWA_SIGNAL;
2304 ret = io_task_work_add(tsk, req, notify);
2306 wake_up_process(tsk);
2311 static void io_req_task_work_add_fallback(struct io_kiocb *req,
2312 task_work_func_t cb)
2314 struct task_struct *tsk = io_wq_get_task(req->ctx->io_wq);
2316 init_task_work(&req->task_work, cb);
2317 task_work_add(tsk, &req->task_work, TWA_NONE);
2318 wake_up_process(tsk);
2321 static void __io_req_task_cancel(struct io_kiocb *req, int error)
2323 struct io_ring_ctx *ctx = req->ctx;
2325 spin_lock_irq(&ctx->completion_lock);
2326 io_cqring_fill_event(req, error);
2327 io_commit_cqring(ctx);
2328 spin_unlock_irq(&ctx->completion_lock);
2330 io_cqring_ev_posted(ctx);
2331 req_set_fail_links(req);
2332 io_double_put_req(req);
2335 static void io_req_task_cancel(struct callback_head *cb)
2337 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2338 struct io_ring_ctx *ctx = req->ctx;
2340 __io_req_task_cancel(req, -ECANCELED);
2341 percpu_ref_put(&ctx->refs);
2344 static void __io_req_task_submit(struct io_kiocb *req)
2346 struct io_ring_ctx *ctx = req->ctx;
2348 /* ctx stays valid until unlock, even if we drop all ours ctx->refs */
2349 mutex_lock(&ctx->uring_lock);
2350 if (!ctx->sqo_dead && !(current->flags & PF_EXITING) &&
2351 !io_sq_thread_acquire_mm_files(ctx, req))
2352 __io_queue_sqe(req);
2354 __io_req_task_cancel(req, -EFAULT);
2355 mutex_unlock(&ctx->uring_lock);
2358 static void io_req_task_submit(struct callback_head *cb)
2360 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2362 __io_req_task_submit(req);
2365 static void io_req_task_queue(struct io_kiocb *req)
2369 req->task_work.func = io_req_task_submit;
2370 ret = io_req_task_work_add(req);
2371 if (unlikely(ret)) {
2372 percpu_ref_get(&req->ctx->refs);
2373 io_req_task_work_add_fallback(req, io_req_task_cancel);
2377 static inline void io_queue_next(struct io_kiocb *req)
2379 struct io_kiocb *nxt = io_req_find_next(req);
2382 io_req_task_queue(nxt);
2385 static void io_free_req(struct io_kiocb *req)
2392 struct task_struct *task;
2397 static inline void io_init_req_batch(struct req_batch *rb)
2404 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2405 struct req_batch *rb)
2408 io_put_task(rb->task, rb->task_refs);
2410 percpu_ref_put_many(&ctx->refs, rb->ctx_refs);
2413 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req,
2414 struct io_submit_state *state)
2418 if (req->task != rb->task) {
2420 io_put_task(rb->task, rb->task_refs);
2421 rb->task = req->task;
2427 io_dismantle_req(req);
2428 if (state->free_reqs != ARRAY_SIZE(state->reqs))
2429 state->reqs[state->free_reqs++] = req;
2431 list_add(&req->compl.list, &state->comp.free_list);
2434 static void io_submit_flush_completions(struct io_comp_state *cs,
2435 struct io_ring_ctx *ctx)
2438 struct io_kiocb *req;
2439 struct req_batch rb;
2441 io_init_req_batch(&rb);
2442 spin_lock_irq(&ctx->completion_lock);
2443 for (i = 0; i < nr; i++) {
2445 __io_cqring_fill_event(req, req->result, req->compl.cflags);
2447 io_commit_cqring(ctx);
2448 spin_unlock_irq(&ctx->completion_lock);
2450 io_cqring_ev_posted(ctx);
2451 for (i = 0; i < nr; i++) {
2454 /* submission and completion refs */
2455 if (refcount_sub_and_test(2, &req->refs))
2456 io_req_free_batch(&rb, req, &ctx->submit_state);
2459 io_req_free_batch_finish(ctx, &rb);
2464 * Drop reference to request, return next in chain (if there is one) if this
2465 * was the last reference to this request.
2467 static struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2469 struct io_kiocb *nxt = NULL;
2471 if (refcount_dec_and_test(&req->refs)) {
2472 nxt = io_req_find_next(req);
2478 static void io_put_req(struct io_kiocb *req)
2480 if (refcount_dec_and_test(&req->refs))
2484 static void io_put_req_deferred_cb(struct callback_head *cb)
2486 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2491 static void io_free_req_deferred(struct io_kiocb *req)
2495 req->task_work.func = io_put_req_deferred_cb;
2496 ret = io_req_task_work_add(req);
2498 io_req_task_work_add_fallback(req, io_put_req_deferred_cb);
2501 static inline void io_put_req_deferred(struct io_kiocb *req, int refs)
2503 if (refcount_sub_and_test(refs, &req->refs))
2504 io_free_req_deferred(req);
2507 static void io_double_put_req(struct io_kiocb *req)
2509 /* drop both submit and complete references */
2510 if (refcount_sub_and_test(2, &req->refs))
2514 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
2516 /* See comment at the top of this file */
2518 return __io_cqring_events(ctx);
2521 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2523 struct io_rings *rings = ctx->rings;
2525 /* make sure SQ entry isn't read before tail */
2526 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2529 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2531 unsigned int cflags;
2533 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2534 cflags |= IORING_CQE_F_BUFFER;
2535 req->flags &= ~REQ_F_BUFFER_SELECTED;
2540 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2542 struct io_buffer *kbuf;
2544 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2545 return io_put_kbuf(req, kbuf);
2548 static inline bool io_run_task_work(void)
2551 * Not safe to run on exiting task, and the task_work handling will
2552 * not add work to such a task.
2554 if (unlikely(current->flags & PF_EXITING))
2556 if (current->task_works) {
2557 __set_current_state(TASK_RUNNING);
2566 * Find and free completed poll iocbs
2568 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2569 struct list_head *done)
2571 struct req_batch rb;
2572 struct io_kiocb *req;
2574 /* order with ->result store in io_complete_rw_iopoll() */
2577 io_init_req_batch(&rb);
2578 while (!list_empty(done)) {
2581 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2582 list_del(&req->inflight_entry);
2584 if (READ_ONCE(req->result) == -EAGAIN) {
2585 req->iopoll_completed = 0;
2586 if (io_rw_reissue(req))
2590 if (req->flags & REQ_F_BUFFER_SELECTED)
2591 cflags = io_put_rw_kbuf(req);
2593 __io_cqring_fill_event(req, req->result, cflags);
2596 if (refcount_dec_and_test(&req->refs))
2597 io_req_free_batch(&rb, req, &ctx->submit_state);
2600 io_commit_cqring(ctx);
2601 io_cqring_ev_posted_iopoll(ctx);
2602 io_req_free_batch_finish(ctx, &rb);
2605 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2608 struct io_kiocb *req, *tmp;
2614 * Only spin for completions if we don't have multiple devices hanging
2615 * off our complete list, and we're under the requested amount.
2617 spin = !ctx->poll_multi_file && *nr_events < min;
2620 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2621 struct kiocb *kiocb = &req->rw.kiocb;
2624 * Move completed and retryable entries to our local lists.
2625 * If we find a request that requires polling, break out
2626 * and complete those lists first, if we have entries there.
2628 if (READ_ONCE(req->iopoll_completed)) {
2629 list_move_tail(&req->inflight_entry, &done);
2632 if (!list_empty(&done))
2635 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2639 /* iopoll may have completed current req */
2640 if (READ_ONCE(req->iopoll_completed))
2641 list_move_tail(&req->inflight_entry, &done);
2648 if (!list_empty(&done))
2649 io_iopoll_complete(ctx, nr_events, &done);
2655 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2656 * non-spinning poll check - we'll still enter the driver poll loop, but only
2657 * as a non-spinning completion check.
2659 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
2662 while (!list_empty(&ctx->iopoll_list) && !need_resched()) {
2665 ret = io_do_iopoll(ctx, nr_events, min);
2668 if (*nr_events >= min)
2676 * We can't just wait for polled events to come to us, we have to actively
2677 * find and complete them.
2679 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2681 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2684 mutex_lock(&ctx->uring_lock);
2685 while (!list_empty(&ctx->iopoll_list)) {
2686 unsigned int nr_events = 0;
2688 io_do_iopoll(ctx, &nr_events, 0);
2690 /* let it sleep and repeat later if can't complete a request */
2694 * Ensure we allow local-to-the-cpu processing to take place,
2695 * in this case we need to ensure that we reap all events.
2696 * Also let task_work, etc. to progress by releasing the mutex
2698 if (need_resched()) {
2699 mutex_unlock(&ctx->uring_lock);
2701 mutex_lock(&ctx->uring_lock);
2704 mutex_unlock(&ctx->uring_lock);
2707 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2709 unsigned int nr_events = 0;
2710 int iters = 0, ret = 0;
2713 * We disallow the app entering submit/complete with polling, but we
2714 * still need to lock the ring to prevent racing with polled issue
2715 * that got punted to a workqueue.
2717 mutex_lock(&ctx->uring_lock);
2720 * Don't enter poll loop if we already have events pending.
2721 * If we do, we can potentially be spinning for commands that
2722 * already triggered a CQE (eg in error).
2724 if (test_bit(0, &ctx->cq_check_overflow))
2725 __io_cqring_overflow_flush(ctx, false, NULL, NULL);
2726 if (io_cqring_events(ctx))
2730 * If a submit got punted to a workqueue, we can have the
2731 * application entering polling for a command before it gets
2732 * issued. That app will hold the uring_lock for the duration
2733 * of the poll right here, so we need to take a breather every
2734 * now and then to ensure that the issue has a chance to add
2735 * the poll to the issued list. Otherwise we can spin here
2736 * forever, while the workqueue is stuck trying to acquire the
2739 if (!(++iters & 7)) {
2740 mutex_unlock(&ctx->uring_lock);
2742 mutex_lock(&ctx->uring_lock);
2745 ret = io_iopoll_getevents(ctx, &nr_events, min);
2749 } while (min && !nr_events && !need_resched());
2751 mutex_unlock(&ctx->uring_lock);
2755 static void kiocb_end_write(struct io_kiocb *req)
2758 * Tell lockdep we inherited freeze protection from submission
2761 if (req->flags & REQ_F_ISREG) {
2762 struct inode *inode = file_inode(req->file);
2764 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
2766 file_end_write(req->file);
2770 static bool io_resubmit_prep(struct io_kiocb *req)
2772 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2774 struct iov_iter iter;
2776 /* already prepared */
2777 if (req->async_data)
2780 switch (req->opcode) {
2781 case IORING_OP_READV:
2782 case IORING_OP_READ_FIXED:
2783 case IORING_OP_READ:
2786 case IORING_OP_WRITEV:
2787 case IORING_OP_WRITE_FIXED:
2788 case IORING_OP_WRITE:
2792 printk_once(KERN_WARNING "io_uring: bad opcode in resubmit %d\n",
2797 ret = io_import_iovec(rw, req, &iovec, &iter, false);
2800 return !io_setup_async_rw(req, iovec, inline_vecs, &iter, false);
2804 static bool io_rw_reissue(struct io_kiocb *req)
2807 umode_t mode = file_inode(req->file)->i_mode;
2810 if (!S_ISBLK(mode) && !S_ISREG(mode))
2812 if ((req->flags & REQ_F_NOWAIT) || io_wq_current_is_worker())
2815 lockdep_assert_held(&req->ctx->uring_lock);
2817 ret = io_sq_thread_acquire_mm_files(req->ctx, req);
2819 if (!ret && io_resubmit_prep(req)) {
2820 refcount_inc(&req->refs);
2821 io_queue_async_work(req);
2824 req_set_fail_links(req);
2829 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2830 unsigned int issue_flags)
2834 if ((res == -EAGAIN || res == -EOPNOTSUPP) && io_rw_reissue(req))
2836 if (res != req->result)
2837 req_set_fail_links(req);
2839 if (req->rw.kiocb.ki_flags & IOCB_WRITE)
2840 kiocb_end_write(req);
2841 if (req->flags & REQ_F_BUFFER_SELECTED)
2842 cflags = io_put_rw_kbuf(req);
2843 __io_req_complete(req, issue_flags, res, cflags);
2846 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2848 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2850 __io_complete_rw(req, res, res2, 0);
2853 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2855 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2857 if (kiocb->ki_flags & IOCB_WRITE)
2858 kiocb_end_write(req);
2860 if (res != -EAGAIN && res != req->result)
2861 req_set_fail_links(req);
2863 WRITE_ONCE(req->result, res);
2864 /* order with io_poll_complete() checking ->result */
2866 WRITE_ONCE(req->iopoll_completed, 1);
2870 * After the iocb has been issued, it's safe to be found on the poll list.
2871 * Adding the kiocb to the list AFTER submission ensures that we don't
2872 * find it from a io_iopoll_getevents() thread before the issuer is done
2873 * accessing the kiocb cookie.
2875 static void io_iopoll_req_issued(struct io_kiocb *req, bool in_async)
2877 struct io_ring_ctx *ctx = req->ctx;
2880 * Track whether we have multiple files in our lists. This will impact
2881 * how we do polling eventually, not spinning if we're on potentially
2882 * different devices.
2884 if (list_empty(&ctx->iopoll_list)) {
2885 ctx->poll_multi_file = false;
2886 } else if (!ctx->poll_multi_file) {
2887 struct io_kiocb *list_req;
2889 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2891 if (list_req->file != req->file)
2892 ctx->poll_multi_file = true;
2896 * For fast devices, IO may have already completed. If it has, add
2897 * it to the front so we find it first.
2899 if (READ_ONCE(req->iopoll_completed))
2900 list_add(&req->inflight_entry, &ctx->iopoll_list);
2902 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2905 * If IORING_SETUP_SQPOLL is enabled, sqes are either handled in sq thread
2906 * task context or in io worker task context. If current task context is
2907 * sq thread, we don't need to check whether should wake up sq thread.
2909 if (in_async && (ctx->flags & IORING_SETUP_SQPOLL) &&
2910 wq_has_sleeper(&ctx->sq_data->wait))
2911 wake_up(&ctx->sq_data->wait);
2914 static inline void io_state_file_put(struct io_submit_state *state)
2916 if (state->file_refs) {
2917 fput_many(state->file, state->file_refs);
2918 state->file_refs = 0;
2923 * Get as many references to a file as we have IOs left in this submission,
2924 * assuming most submissions are for one file, or at least that each file
2925 * has more than one submission.
2927 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2932 if (state->file_refs) {
2933 if (state->fd == fd) {
2937 io_state_file_put(state);
2939 state->file = fget_many(fd, state->ios_left);
2940 if (unlikely(!state->file))
2944 state->file_refs = state->ios_left - 1;
2948 static bool io_bdev_nowait(struct block_device *bdev)
2950 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2954 * If we tracked the file through the SCM inflight mechanism, we could support
2955 * any file. For now, just ensure that anything potentially problematic is done
2958 static bool io_file_supports_async(struct file *file, int rw)
2960 umode_t mode = file_inode(file)->i_mode;
2962 if (S_ISBLK(mode)) {
2963 if (IS_ENABLED(CONFIG_BLOCK) &&
2964 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2968 if (S_ISCHR(mode) || S_ISSOCK(mode))
2970 if (S_ISREG(mode)) {
2971 if (IS_ENABLED(CONFIG_BLOCK) &&
2972 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2973 file->f_op != &io_uring_fops)
2978 /* any ->read/write should understand O_NONBLOCK */
2979 if (file->f_flags & O_NONBLOCK)
2982 if (!(file->f_mode & FMODE_NOWAIT))
2986 return file->f_op->read_iter != NULL;
2988 return file->f_op->write_iter != NULL;
2991 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2993 struct io_ring_ctx *ctx = req->ctx;
2994 struct kiocb *kiocb = &req->rw.kiocb;
2995 struct file *file = req->file;
2999 if (S_ISREG(file_inode(file)->i_mode))
3000 req->flags |= REQ_F_ISREG;
3002 kiocb->ki_pos = READ_ONCE(sqe->off);
3003 if (kiocb->ki_pos == -1 && !(file->f_mode & FMODE_STREAM)) {
3004 req->flags |= REQ_F_CUR_POS;
3005 kiocb->ki_pos = file->f_pos;
3007 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
3008 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
3009 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
3013 /* don't allow async punt for O_NONBLOCK or RWF_NOWAIT */
3014 if ((kiocb->ki_flags & IOCB_NOWAIT) || (file->f_flags & O_NONBLOCK))
3015 req->flags |= REQ_F_NOWAIT;
3017 ioprio = READ_ONCE(sqe->ioprio);
3019 ret = ioprio_check_cap(ioprio);
3023 kiocb->ki_ioprio = ioprio;
3025 kiocb->ki_ioprio = get_current_ioprio();
3027 if (ctx->flags & IORING_SETUP_IOPOLL) {
3028 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
3029 !kiocb->ki_filp->f_op->iopoll)
3032 kiocb->ki_flags |= IOCB_HIPRI;
3033 kiocb->ki_complete = io_complete_rw_iopoll;
3034 req->iopoll_completed = 0;
3036 if (kiocb->ki_flags & IOCB_HIPRI)
3038 kiocb->ki_complete = io_complete_rw;
3041 req->rw.addr = READ_ONCE(sqe->addr);
3042 req->rw.len = READ_ONCE(sqe->len);
3043 req->buf_index = READ_ONCE(sqe->buf_index);
3047 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
3053 case -ERESTARTNOINTR:
3054 case -ERESTARTNOHAND:
3055 case -ERESTART_RESTARTBLOCK:
3057 * We can't just restart the syscall, since previously
3058 * submitted sqes may already be in progress. Just fail this
3064 kiocb->ki_complete(kiocb, ret, 0);
3068 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
3069 unsigned int issue_flags)
3071 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
3072 struct io_async_rw *io = req->async_data;
3074 /* add previously done IO, if any */
3075 if (io && io->bytes_done > 0) {
3077 ret = io->bytes_done;
3079 ret += io->bytes_done;
3082 if (req->flags & REQ_F_CUR_POS)
3083 req->file->f_pos = kiocb->ki_pos;
3084 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
3085 __io_complete_rw(req, ret, 0, issue_flags);
3087 io_rw_done(kiocb, ret);
3090 static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter)
3092 struct io_ring_ctx *ctx = req->ctx;
3093 size_t len = req->rw.len;
3094 struct io_mapped_ubuf *imu;
3095 u16 index, buf_index = req->buf_index;
3099 if (unlikely(buf_index >= ctx->nr_user_bufs))
3101 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
3102 imu = &ctx->user_bufs[index];
3103 buf_addr = req->rw.addr;
3106 if (buf_addr + len < buf_addr)
3108 /* not inside the mapped region */
3109 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
3113 * May not be a start of buffer, set size appropriately
3114 * and advance us to the beginning.
3116 offset = buf_addr - imu->ubuf;
3117 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
3121 * Don't use iov_iter_advance() here, as it's really slow for
3122 * using the latter parts of a big fixed buffer - it iterates
3123 * over each segment manually. We can cheat a bit here, because
3126 * 1) it's a BVEC iter, we set it up
3127 * 2) all bvecs are PAGE_SIZE in size, except potentially the
3128 * first and last bvec
3130 * So just find our index, and adjust the iterator afterwards.
3131 * If the offset is within the first bvec (or the whole first
3132 * bvec, just use iov_iter_advance(). This makes it easier
3133 * since we can just skip the first segment, which may not
3134 * be PAGE_SIZE aligned.
3136 const struct bio_vec *bvec = imu->bvec;
3138 if (offset <= bvec->bv_len) {
3139 iov_iter_advance(iter, offset);
3141 unsigned long seg_skip;
3143 /* skip first vec */
3144 offset -= bvec->bv_len;
3145 seg_skip = 1 + (offset >> PAGE_SHIFT);
3147 iter->bvec = bvec + seg_skip;
3148 iter->nr_segs -= seg_skip;
3149 iter->count -= bvec->bv_len + offset;
3150 iter->iov_offset = offset & ~PAGE_MASK;
3157 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
3160 mutex_unlock(&ctx->uring_lock);
3163 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
3166 * "Normal" inline submissions always hold the uring_lock, since we
3167 * grab it from the system call. Same is true for the SQPOLL offload.
3168 * The only exception is when we've detached the request and issue it
3169 * from an async worker thread, grab the lock for that case.
3172 mutex_lock(&ctx->uring_lock);
3175 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
3176 int bgid, struct io_buffer *kbuf,
3179 struct io_buffer *head;
3181 if (req->flags & REQ_F_BUFFER_SELECTED)
3184 io_ring_submit_lock(req->ctx, needs_lock);
3186 lockdep_assert_held(&req->ctx->uring_lock);
3188 head = idr_find(&req->ctx->io_buffer_idr, bgid);
3190 if (!list_empty(&head->list)) {
3191 kbuf = list_last_entry(&head->list, struct io_buffer,
3193 list_del(&kbuf->list);
3196 idr_remove(&req->ctx->io_buffer_idr, bgid);
3198 if (*len > kbuf->len)
3201 kbuf = ERR_PTR(-ENOBUFS);
3204 io_ring_submit_unlock(req->ctx, needs_lock);
3209 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
3212 struct io_buffer *kbuf;
3215 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3216 bgid = req->buf_index;
3217 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
3220 req->rw.addr = (u64) (unsigned long) kbuf;
3221 req->flags |= REQ_F_BUFFER_SELECTED;
3222 return u64_to_user_ptr(kbuf->addr);
3225 #ifdef CONFIG_COMPAT
3226 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
3229 struct compat_iovec __user *uiov;
3230 compat_ssize_t clen;
3234 uiov = u64_to_user_ptr(req->rw.addr);
3235 if (!access_ok(uiov, sizeof(*uiov)))
3237 if (__get_user(clen, &uiov->iov_len))
3243 buf = io_rw_buffer_select(req, &len, needs_lock);
3245 return PTR_ERR(buf);
3246 iov[0].iov_base = buf;
3247 iov[0].iov_len = (compat_size_t) len;
3252 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3255 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
3259 if (copy_from_user(iov, uiov, sizeof(*uiov)))
3262 len = iov[0].iov_len;
3265 buf = io_rw_buffer_select(req, &len, needs_lock);
3267 return PTR_ERR(buf);
3268 iov[0].iov_base = buf;
3269 iov[0].iov_len = len;
3273 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3276 if (req->flags & REQ_F_BUFFER_SELECTED) {
3277 struct io_buffer *kbuf;
3279 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3280 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
3281 iov[0].iov_len = kbuf->len;
3284 if (req->rw.len != 1)
3287 #ifdef CONFIG_COMPAT
3288 if (req->ctx->compat)
3289 return io_compat_import(req, iov, needs_lock);
3292 return __io_iov_buffer_select(req, iov, needs_lock);
3295 static int io_import_iovec(int rw, struct io_kiocb *req, struct iovec **iovec,
3296 struct iov_iter *iter, bool needs_lock)
3298 void __user *buf = u64_to_user_ptr(req->rw.addr);
3299 size_t sqe_len = req->rw.len;
3300 u8 opcode = req->opcode;
3303 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3305 return io_import_fixed(req, rw, iter);
3308 /* buffer index only valid with fixed read/write, or buffer select */
3309 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3312 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3313 if (req->flags & REQ_F_BUFFER_SELECT) {
3314 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3316 return PTR_ERR(buf);
3317 req->rw.len = sqe_len;
3320 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3325 if (req->flags & REQ_F_BUFFER_SELECT) {
3326 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3328 iov_iter_init(iter, rw, *iovec, 1, (*iovec)->iov_len);
3333 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3337 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3339 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3343 * For files that don't have ->read_iter() and ->write_iter(), handle them
3344 * by looping over ->read() or ->write() manually.
3346 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3348 struct kiocb *kiocb = &req->rw.kiocb;
3349 struct file *file = req->file;
3353 * Don't support polled IO through this interface, and we can't
3354 * support non-blocking either. For the latter, this just causes
3355 * the kiocb to be handled from an async context.
3357 if (kiocb->ki_flags & IOCB_HIPRI)
3359 if (kiocb->ki_flags & IOCB_NOWAIT)
3362 while (iov_iter_count(iter)) {
3366 if (!iov_iter_is_bvec(iter)) {
3367 iovec = iov_iter_iovec(iter);
3369 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3370 iovec.iov_len = req->rw.len;
3374 nr = file->f_op->read(file, iovec.iov_base,
3375 iovec.iov_len, io_kiocb_ppos(kiocb));
3377 nr = file->f_op->write(file, iovec.iov_base,
3378 iovec.iov_len, io_kiocb_ppos(kiocb));
3387 if (nr != iovec.iov_len)
3391 iov_iter_advance(iter, nr);
3397 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3398 const struct iovec *fast_iov, struct iov_iter *iter)
3400 struct io_async_rw *rw = req->async_data;
3402 memcpy(&rw->iter, iter, sizeof(*iter));
3403 rw->free_iovec = iovec;
3405 /* can only be fixed buffers, no need to do anything */
3406 if (iov_iter_is_bvec(iter))
3409 unsigned iov_off = 0;
3411 rw->iter.iov = rw->fast_iov;
3412 if (iter->iov != fast_iov) {
3413 iov_off = iter->iov - fast_iov;
3414 rw->iter.iov += iov_off;
3416 if (rw->fast_iov != fast_iov)
3417 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3418 sizeof(struct iovec) * iter->nr_segs);
3420 req->flags |= REQ_F_NEED_CLEANUP;
3424 static inline int __io_alloc_async_data(struct io_kiocb *req)
3426 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3427 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3428 return req->async_data == NULL;
3431 static int io_alloc_async_data(struct io_kiocb *req)
3433 if (!io_op_defs[req->opcode].needs_async_data)
3436 return __io_alloc_async_data(req);
3439 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3440 const struct iovec *fast_iov,
3441 struct iov_iter *iter, bool force)
3443 if (!force && !io_op_defs[req->opcode].needs_async_data)
3445 if (!req->async_data) {
3446 if (__io_alloc_async_data(req)) {
3451 io_req_map_rw(req, iovec, fast_iov, iter);
3456 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3458 struct io_async_rw *iorw = req->async_data;
3459 struct iovec *iov = iorw->fast_iov;
3462 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3463 if (unlikely(ret < 0))
3466 iorw->bytes_done = 0;
3467 iorw->free_iovec = iov;
3469 req->flags |= REQ_F_NEED_CLEANUP;
3473 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3477 ret = io_prep_rw(req, sqe);
3481 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3484 /* either don't need iovec imported or already have it */
3485 if (!req->async_data)
3487 return io_rw_prep_async(req, READ);
3491 * This is our waitqueue callback handler, registered through lock_page_async()
3492 * when we initially tried to do the IO with the iocb armed our waitqueue.
3493 * This gets called when the page is unlocked, and we generally expect that to
3494 * happen when the page IO is completed and the page is now uptodate. This will
3495 * queue a task_work based retry of the operation, attempting to copy the data
3496 * again. If the latter fails because the page was NOT uptodate, then we will
3497 * do a thread based blocking retry of the operation. That's the unexpected
3500 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3501 int sync, void *arg)
3503 struct wait_page_queue *wpq;
3504 struct io_kiocb *req = wait->private;
3505 struct wait_page_key *key = arg;
3507 wpq = container_of(wait, struct wait_page_queue, wait);
3509 if (!wake_page_match(wpq, key))
3512 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3513 list_del_init(&wait->entry);
3515 /* submit ref gets dropped, acquire a new one */
3516 refcount_inc(&req->refs);
3517 io_req_task_queue(req);
3522 * This controls whether a given IO request should be armed for async page
3523 * based retry. If we return false here, the request is handed to the async
3524 * worker threads for retry. If we're doing buffered reads on a regular file,
3525 * we prepare a private wait_page_queue entry and retry the operation. This
3526 * will either succeed because the page is now uptodate and unlocked, or it
3527 * will register a callback when the page is unlocked at IO completion. Through
3528 * that callback, io_uring uses task_work to setup a retry of the operation.
3529 * That retry will attempt the buffered read again. The retry will generally
3530 * succeed, or in rare cases where it fails, we then fall back to using the
3531 * async worker threads for a blocking retry.
3533 static bool io_rw_should_retry(struct io_kiocb *req)
3535 struct io_async_rw *rw = req->async_data;
3536 struct wait_page_queue *wait = &rw->wpq;
3537 struct kiocb *kiocb = &req->rw.kiocb;
3539 /* never retry for NOWAIT, we just complete with -EAGAIN */
3540 if (req->flags & REQ_F_NOWAIT)
3543 /* Only for buffered IO */
3544 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3548 * just use poll if we can, and don't attempt if the fs doesn't
3549 * support callback based unlocks
3551 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3554 wait->wait.func = io_async_buf_func;
3555 wait->wait.private = req;
3556 wait->wait.flags = 0;
3557 INIT_LIST_HEAD(&wait->wait.entry);
3558 kiocb->ki_flags |= IOCB_WAITQ;
3559 kiocb->ki_flags &= ~IOCB_NOWAIT;
3560 kiocb->ki_waitq = wait;
3564 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3566 if (req->file->f_op->read_iter)
3567 return call_read_iter(req->file, &req->rw.kiocb, iter);
3568 else if (req->file->f_op->read)
3569 return loop_rw_iter(READ, req, iter);
3574 static int io_read(struct io_kiocb *req, unsigned int issue_flags)
3576 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3577 struct kiocb *kiocb = &req->rw.kiocb;
3578 struct iov_iter __iter, *iter = &__iter;
3579 struct io_async_rw *rw = req->async_data;
3580 ssize_t io_size, ret, ret2;
3581 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3587 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3591 io_size = iov_iter_count(iter);
3592 req->result = io_size;
3594 /* Ensure we clear previously set non-block flag */
3595 if (!force_nonblock)
3596 kiocb->ki_flags &= ~IOCB_NOWAIT;
3598 kiocb->ki_flags |= IOCB_NOWAIT;
3600 /* If the file doesn't support async, just async punt */
3601 if (force_nonblock && !io_file_supports_async(req->file, READ)) {
3602 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3603 return ret ?: -EAGAIN;
3606 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size);
3607 if (unlikely(ret)) {
3612 ret = io_iter_do_read(req, iter);
3614 if (ret == -EIOCBQUEUED) {
3616 } else if (ret == -EAGAIN) {
3617 /* IOPOLL retry should happen for io-wq threads */
3618 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3620 /* no retry on NONBLOCK nor RWF_NOWAIT */
3621 if (req->flags & REQ_F_NOWAIT)
3623 /* some cases will consume bytes even on error returns */
3624 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3626 } else if (ret <= 0 || ret == io_size || !force_nonblock ||
3627 (req->flags & REQ_F_NOWAIT) || !(req->flags & REQ_F_ISREG)) {
3628 /* read all, failed, already did sync or don't want to retry */
3632 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3637 rw = req->async_data;
3638 /* now use our persistent iterator, if we aren't already */
3643 rw->bytes_done += ret;
3644 /* if we can retry, do so with the callbacks armed */
3645 if (!io_rw_should_retry(req)) {
3646 kiocb->ki_flags &= ~IOCB_WAITQ;
3651 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
3652 * we get -EIOCBQUEUED, then we'll get a notification when the
3653 * desired page gets unlocked. We can also get a partial read
3654 * here, and if we do, then just retry at the new offset.
3656 ret = io_iter_do_read(req, iter);
3657 if (ret == -EIOCBQUEUED)
3659 /* we got some bytes, but not all. retry. */
3660 } while (ret > 0 && ret < io_size);
3662 kiocb_done(kiocb, ret, issue_flags);
3664 /* it's faster to check here then delegate to kfree */
3670 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3674 ret = io_prep_rw(req, sqe);
3678 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3681 /* either don't need iovec imported or already have it */
3682 if (!req->async_data)
3684 return io_rw_prep_async(req, WRITE);
3687 static int io_write(struct io_kiocb *req, unsigned int issue_flags)
3689 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3690 struct kiocb *kiocb = &req->rw.kiocb;
3691 struct iov_iter __iter, *iter = &__iter;
3692 struct io_async_rw *rw = req->async_data;
3693 ssize_t ret, ret2, io_size;
3694 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
3700 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3704 io_size = iov_iter_count(iter);
3705 req->result = io_size;
3707 /* Ensure we clear previously set non-block flag */
3708 if (!force_nonblock)
3709 kiocb->ki_flags &= ~IOCB_NOWAIT;
3711 kiocb->ki_flags |= IOCB_NOWAIT;
3713 /* If the file doesn't support async, just async punt */
3714 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
3717 /* file path doesn't support NOWAIT for non-direct_IO */
3718 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3719 (req->flags & REQ_F_ISREG))
3722 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size);
3727 * Open-code file_start_write here to grab freeze protection,
3728 * which will be released by another thread in
3729 * io_complete_rw(). Fool lockdep by telling it the lock got
3730 * released so that it doesn't complain about the held lock when
3731 * we return to userspace.
3733 if (req->flags & REQ_F_ISREG) {
3734 sb_start_write(file_inode(req->file)->i_sb);
3735 __sb_writers_release(file_inode(req->file)->i_sb,
3738 kiocb->ki_flags |= IOCB_WRITE;
3740 if (req->file->f_op->write_iter)
3741 ret2 = call_write_iter(req->file, kiocb, iter);
3742 else if (req->file->f_op->write)
3743 ret2 = loop_rw_iter(WRITE, req, iter);
3748 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3749 * retry them without IOCB_NOWAIT.
3751 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3753 /* no retry on NONBLOCK nor RWF_NOWAIT */
3754 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
3756 if (!force_nonblock || ret2 != -EAGAIN) {
3757 /* IOPOLL retry should happen for io-wq threads */
3758 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3761 kiocb_done(kiocb, ret2, issue_flags);
3764 /* some cases will consume bytes even on error returns */
3765 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3766 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3767 return ret ?: -EAGAIN;
3770 /* it's reportedly faster than delegating the null check to kfree() */
3776 static int io_renameat_prep(struct io_kiocb *req,
3777 const struct io_uring_sqe *sqe)
3779 struct io_rename *ren = &req->rename;
3780 const char __user *oldf, *newf;
3782 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3785 ren->old_dfd = READ_ONCE(sqe->fd);
3786 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3787 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3788 ren->new_dfd = READ_ONCE(sqe->len);
3789 ren->flags = READ_ONCE(sqe->rename_flags);
3791 ren->oldpath = getname(oldf);
3792 if (IS_ERR(ren->oldpath))
3793 return PTR_ERR(ren->oldpath);
3795 ren->newpath = getname(newf);
3796 if (IS_ERR(ren->newpath)) {
3797 putname(ren->oldpath);
3798 return PTR_ERR(ren->newpath);
3801 req->flags |= REQ_F_NEED_CLEANUP;
3805 static int io_renameat(struct io_kiocb *req, unsigned int issue_flags)
3807 struct io_rename *ren = &req->rename;
3810 if (issue_flags & IO_URING_F_NONBLOCK)
3813 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3814 ren->newpath, ren->flags);
3816 req->flags &= ~REQ_F_NEED_CLEANUP;
3818 req_set_fail_links(req);
3819 io_req_complete(req, ret);
3823 static int io_unlinkat_prep(struct io_kiocb *req,
3824 const struct io_uring_sqe *sqe)
3826 struct io_unlink *un = &req->unlink;
3827 const char __user *fname;
3829 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3832 un->dfd = READ_ONCE(sqe->fd);
3834 un->flags = READ_ONCE(sqe->unlink_flags);
3835 if (un->flags & ~AT_REMOVEDIR)
3838 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3839 un->filename = getname(fname);
3840 if (IS_ERR(un->filename))
3841 return PTR_ERR(un->filename);
3843 req->flags |= REQ_F_NEED_CLEANUP;
3847 static int io_unlinkat(struct io_kiocb *req, unsigned int issue_flags)
3849 struct io_unlink *un = &req->unlink;
3852 if (issue_flags & IO_URING_F_NONBLOCK)
3855 if (un->flags & AT_REMOVEDIR)
3856 ret = do_rmdir(un->dfd, un->filename);
3858 ret = do_unlinkat(un->dfd, un->filename);
3860 req->flags &= ~REQ_F_NEED_CLEANUP;
3862 req_set_fail_links(req);
3863 io_req_complete(req, ret);
3867 static int io_shutdown_prep(struct io_kiocb *req,
3868 const struct io_uring_sqe *sqe)
3870 #if defined(CONFIG_NET)
3871 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3873 if (sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3877 req->shutdown.how = READ_ONCE(sqe->len);
3884 static int io_shutdown(struct io_kiocb *req, unsigned int issue_flags)
3886 #if defined(CONFIG_NET)
3887 struct socket *sock;
3890 if (issue_flags & IO_URING_F_NONBLOCK)
3893 sock = sock_from_file(req->file);
3894 if (unlikely(!sock))
3897 ret = __sys_shutdown_sock(sock, req->shutdown.how);
3899 req_set_fail_links(req);
3900 io_req_complete(req, ret);
3907 static int __io_splice_prep(struct io_kiocb *req,
3908 const struct io_uring_sqe *sqe)
3910 struct io_splice* sp = &req->splice;
3911 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3913 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3917 sp->len = READ_ONCE(sqe->len);
3918 sp->flags = READ_ONCE(sqe->splice_flags);
3920 if (unlikely(sp->flags & ~valid_flags))
3923 sp->file_in = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in),
3924 (sp->flags & SPLICE_F_FD_IN_FIXED));
3927 req->flags |= REQ_F_NEED_CLEANUP;
3929 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
3931 * Splice operation will be punted aync, and here need to
3932 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3934 io_req_init_async(req);
3935 req->work.flags |= IO_WQ_WORK_UNBOUND;
3941 static int io_tee_prep(struct io_kiocb *req,
3942 const struct io_uring_sqe *sqe)
3944 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3946 return __io_splice_prep(req, sqe);
3949 static int io_tee(struct io_kiocb *req, unsigned int issue_flags)
3951 struct io_splice *sp = &req->splice;
3952 struct file *in = sp->file_in;
3953 struct file *out = sp->file_out;
3954 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3957 if (issue_flags & IO_URING_F_NONBLOCK)
3960 ret = do_tee(in, out, sp->len, flags);
3962 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3963 req->flags &= ~REQ_F_NEED_CLEANUP;
3966 req_set_fail_links(req);
3967 io_req_complete(req, ret);
3971 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3973 struct io_splice* sp = &req->splice;
3975 sp->off_in = READ_ONCE(sqe->splice_off_in);
3976 sp->off_out = READ_ONCE(sqe->off);
3977 return __io_splice_prep(req, sqe);
3980 static int io_splice(struct io_kiocb *req, unsigned int issue_flags)
3982 struct io_splice *sp = &req->splice;
3983 struct file *in = sp->file_in;
3984 struct file *out = sp->file_out;
3985 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3986 loff_t *poff_in, *poff_out;
3989 if (issue_flags & IO_URING_F_NONBLOCK)
3992 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3993 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3996 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3998 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3999 req->flags &= ~REQ_F_NEED_CLEANUP;
4002 req_set_fail_links(req);
4003 io_req_complete(req, ret);
4008 * IORING_OP_NOP just posts a completion event, nothing else.
4010 static int io_nop(struct io_kiocb *req, unsigned int issue_flags)
4012 struct io_ring_ctx *ctx = req->ctx;
4014 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4017 __io_req_complete(req, issue_flags, 0, 0);
4021 static int io_fsync_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4023 struct io_ring_ctx *ctx = req->ctx;
4028 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4030 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4033 req->sync.flags = READ_ONCE(sqe->fsync_flags);
4034 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
4037 req->sync.off = READ_ONCE(sqe->off);
4038 req->sync.len = READ_ONCE(sqe->len);
4042 static int io_fsync(struct io_kiocb *req, unsigned int issue_flags)
4044 loff_t end = req->sync.off + req->sync.len;
4047 /* fsync always requires a blocking context */
4048 if (issue_flags & IO_URING_F_NONBLOCK)
4051 ret = vfs_fsync_range(req->file, req->sync.off,
4052 end > 0 ? end : LLONG_MAX,
4053 req->sync.flags & IORING_FSYNC_DATASYNC);
4055 req_set_fail_links(req);
4056 io_req_complete(req, ret);
4060 static int io_fallocate_prep(struct io_kiocb *req,
4061 const struct io_uring_sqe *sqe)
4063 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
4065 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4068 req->sync.off = READ_ONCE(sqe->off);
4069 req->sync.len = READ_ONCE(sqe->addr);
4070 req->sync.mode = READ_ONCE(sqe->len);
4074 static int io_fallocate(struct io_kiocb *req, unsigned int issue_flags)
4078 /* fallocate always requiring blocking context */
4079 if (issue_flags & IO_URING_F_NONBLOCK)
4081 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
4084 req_set_fail_links(req);
4085 io_req_complete(req, ret);
4089 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4091 const char __user *fname;
4094 if (unlikely(sqe->ioprio || sqe->buf_index))
4096 if (unlikely(req->flags & REQ_F_FIXED_FILE))
4099 /* open.how should be already initialised */
4100 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
4101 req->open.how.flags |= O_LARGEFILE;
4103 req->open.dfd = READ_ONCE(sqe->fd);
4104 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
4105 req->open.filename = getname(fname);
4106 if (IS_ERR(req->open.filename)) {
4107 ret = PTR_ERR(req->open.filename);
4108 req->open.filename = NULL;
4111 req->open.nofile = rlimit(RLIMIT_NOFILE);
4112 req->flags |= REQ_F_NEED_CLEANUP;
4116 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4120 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4122 mode = READ_ONCE(sqe->len);
4123 flags = READ_ONCE(sqe->open_flags);
4124 req->open.how = build_open_how(flags, mode);
4125 return __io_openat_prep(req, sqe);
4128 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4130 struct open_how __user *how;
4134 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4136 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4137 len = READ_ONCE(sqe->len);
4138 if (len < OPEN_HOW_SIZE_VER0)
4141 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
4146 return __io_openat_prep(req, sqe);
4149 static int io_openat2(struct io_kiocb *req, unsigned int issue_flags)
4151 struct open_flags op;
4154 bool resolve_nonblock;
4157 ret = build_open_flags(&req->open.how, &op);
4160 nonblock_set = op.open_flag & O_NONBLOCK;
4161 resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
4162 if (issue_flags & IO_URING_F_NONBLOCK) {
4164 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
4165 * it'll always -EAGAIN
4167 if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
4169 op.lookup_flags |= LOOKUP_CACHED;
4170 op.open_flag |= O_NONBLOCK;
4173 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
4177 file = do_filp_open(req->open.dfd, req->open.filename, &op);
4178 /* only retry if RESOLVE_CACHED wasn't already set by application */
4179 if ((!resolve_nonblock && (issue_flags & IO_URING_F_NONBLOCK)) &&
4180 file == ERR_PTR(-EAGAIN)) {
4182 * We could hang on to this 'fd', but seems like marginal
4183 * gain for something that is now known to be a slower path.
4184 * So just put it, and we'll get a new one when we retry.
4192 ret = PTR_ERR(file);
4194 if ((issue_flags & IO_URING_F_NONBLOCK) && !nonblock_set)
4195 file->f_flags &= ~O_NONBLOCK;
4196 fsnotify_open(file);
4197 fd_install(ret, file);
4200 putname(req->open.filename);
4201 req->flags &= ~REQ_F_NEED_CLEANUP;
4203 req_set_fail_links(req);
4204 io_req_complete(req, ret);
4208 static int io_openat(struct io_kiocb *req, unsigned int issue_flags)
4210 return io_openat2(req, issue_flags & IO_URING_F_NONBLOCK);
4213 static int io_remove_buffers_prep(struct io_kiocb *req,
4214 const struct io_uring_sqe *sqe)
4216 struct io_provide_buf *p = &req->pbuf;
4219 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
4222 tmp = READ_ONCE(sqe->fd);
4223 if (!tmp || tmp > USHRT_MAX)
4226 memset(p, 0, sizeof(*p));
4228 p->bgid = READ_ONCE(sqe->buf_group);
4232 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
4233 int bgid, unsigned nbufs)
4237 /* shouldn't happen */
4241 /* the head kbuf is the list itself */
4242 while (!list_empty(&buf->list)) {
4243 struct io_buffer *nxt;
4245 nxt = list_first_entry(&buf->list, struct io_buffer, list);
4246 list_del(&nxt->list);
4253 idr_remove(&ctx->io_buffer_idr, bgid);
4258 static int io_remove_buffers(struct io_kiocb *req, unsigned int issue_flags)
4260 struct io_provide_buf *p = &req->pbuf;
4261 struct io_ring_ctx *ctx = req->ctx;
4262 struct io_buffer *head;
4264 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4266 io_ring_submit_lock(ctx, !force_nonblock);
4268 lockdep_assert_held(&ctx->uring_lock);
4271 head = idr_find(&ctx->io_buffer_idr, p->bgid);
4273 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
4275 req_set_fail_links(req);
4277 /* need to hold the lock to complete IOPOLL requests */
4278 if (ctx->flags & IORING_SETUP_IOPOLL) {
4279 __io_req_complete(req, issue_flags, ret, 0);
4280 io_ring_submit_unlock(ctx, !force_nonblock);
4282 io_ring_submit_unlock(ctx, !force_nonblock);
4283 __io_req_complete(req, issue_flags, ret, 0);
4288 static int io_provide_buffers_prep(struct io_kiocb *req,
4289 const struct io_uring_sqe *sqe)
4291 struct io_provide_buf *p = &req->pbuf;
4294 if (sqe->ioprio || sqe->rw_flags)
4297 tmp = READ_ONCE(sqe->fd);
4298 if (!tmp || tmp > USHRT_MAX)
4301 p->addr = READ_ONCE(sqe->addr);
4302 p->len = READ_ONCE(sqe->len);
4304 if (!access_ok(u64_to_user_ptr(p->addr), (p->len * p->nbufs)))
4307 p->bgid = READ_ONCE(sqe->buf_group);
4308 tmp = READ_ONCE(sqe->off);
4309 if (tmp > USHRT_MAX)
4315 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
4317 struct io_buffer *buf;
4318 u64 addr = pbuf->addr;
4319 int i, bid = pbuf->bid;
4321 for (i = 0; i < pbuf->nbufs; i++) {
4322 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
4327 buf->len = pbuf->len;
4332 INIT_LIST_HEAD(&buf->list);
4335 list_add_tail(&buf->list, &(*head)->list);
4339 return i ? i : -ENOMEM;
4342 static int io_provide_buffers(struct io_kiocb *req, unsigned int issue_flags)
4344 struct io_provide_buf *p = &req->pbuf;
4345 struct io_ring_ctx *ctx = req->ctx;
4346 struct io_buffer *head, *list;
4348 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4350 io_ring_submit_lock(ctx, !force_nonblock);
4352 lockdep_assert_held(&ctx->uring_lock);
4354 list = head = idr_find(&ctx->io_buffer_idr, p->bgid);
4356 ret = io_add_buffers(p, &head);
4361 ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1,
4364 __io_remove_buffers(ctx, head, p->bgid, -1U);
4370 req_set_fail_links(req);
4372 /* need to hold the lock to complete IOPOLL requests */
4373 if (ctx->flags & IORING_SETUP_IOPOLL) {
4374 __io_req_complete(req, issue_flags, ret, 0);
4375 io_ring_submit_unlock(ctx, !force_nonblock);
4377 io_ring_submit_unlock(ctx, !force_nonblock);
4378 __io_req_complete(req, issue_flags, ret, 0);
4383 static int io_epoll_ctl_prep(struct io_kiocb *req,
4384 const struct io_uring_sqe *sqe)
4386 #if defined(CONFIG_EPOLL)
4387 if (sqe->ioprio || sqe->buf_index)
4389 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4392 req->epoll.epfd = READ_ONCE(sqe->fd);
4393 req->epoll.op = READ_ONCE(sqe->len);
4394 req->epoll.fd = READ_ONCE(sqe->off);
4396 if (ep_op_has_event(req->epoll.op)) {
4397 struct epoll_event __user *ev;
4399 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4400 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4410 static int io_epoll_ctl(struct io_kiocb *req, unsigned int issue_flags)
4412 #if defined(CONFIG_EPOLL)
4413 struct io_epoll *ie = &req->epoll;
4415 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4417 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4418 if (force_nonblock && ret == -EAGAIN)
4422 req_set_fail_links(req);
4423 __io_req_complete(req, issue_flags, ret, 0);
4430 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4432 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4433 if (sqe->ioprio || sqe->buf_index || sqe->off)
4435 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4438 req->madvise.addr = READ_ONCE(sqe->addr);
4439 req->madvise.len = READ_ONCE(sqe->len);
4440 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4447 static int io_madvise(struct io_kiocb *req, unsigned int issue_flags)
4449 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4450 struct io_madvise *ma = &req->madvise;
4453 if (issue_flags & IO_URING_F_NONBLOCK)
4456 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4458 req_set_fail_links(req);
4459 io_req_complete(req, ret);
4466 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4468 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4470 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4473 req->fadvise.offset = READ_ONCE(sqe->off);
4474 req->fadvise.len = READ_ONCE(sqe->len);
4475 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4479 static int io_fadvise(struct io_kiocb *req, unsigned int issue_flags)
4481 struct io_fadvise *fa = &req->fadvise;
4484 if (issue_flags & IO_URING_F_NONBLOCK) {
4485 switch (fa->advice) {
4486 case POSIX_FADV_NORMAL:
4487 case POSIX_FADV_RANDOM:
4488 case POSIX_FADV_SEQUENTIAL:
4495 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4497 req_set_fail_links(req);
4498 io_req_complete(req, ret);
4502 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4504 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4506 if (sqe->ioprio || sqe->buf_index)
4508 if (req->flags & REQ_F_FIXED_FILE)
4511 req->statx.dfd = READ_ONCE(sqe->fd);
4512 req->statx.mask = READ_ONCE(sqe->len);
4513 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4514 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4515 req->statx.flags = READ_ONCE(sqe->statx_flags);
4520 static int io_statx(struct io_kiocb *req, unsigned int issue_flags)
4522 struct io_statx *ctx = &req->statx;
4525 if (issue_flags & IO_URING_F_NONBLOCK) {
4526 /* only need file table for an actual valid fd */
4527 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
4528 req->flags |= REQ_F_NO_FILE_TABLE;
4532 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4536 req_set_fail_links(req);
4537 io_req_complete(req, ret);
4541 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4543 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4545 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4546 sqe->rw_flags || sqe->buf_index)
4548 if (req->flags & REQ_F_FIXED_FILE)
4551 req->close.fd = READ_ONCE(sqe->fd);
4555 static int io_close(struct io_kiocb *req, unsigned int issue_flags)
4557 struct files_struct *files = current->files;
4558 struct io_close *close = &req->close;
4559 struct fdtable *fdt;
4565 spin_lock(&files->file_lock);
4566 fdt = files_fdtable(files);
4567 if (close->fd >= fdt->max_fds) {
4568 spin_unlock(&files->file_lock);
4571 file = fdt->fd[close->fd];
4573 spin_unlock(&files->file_lock);
4577 if (file->f_op == &io_uring_fops) {
4578 spin_unlock(&files->file_lock);
4583 /* if the file has a flush method, be safe and punt to async */
4584 if (file->f_op->flush && (issue_flags & IO_URING_F_NONBLOCK)) {
4585 spin_unlock(&files->file_lock);
4589 ret = __close_fd_get_file(close->fd, &file);
4590 spin_unlock(&files->file_lock);
4597 /* No ->flush() or already async, safely close from here */
4598 ret = filp_close(file, current->files);
4601 req_set_fail_links(req);
4604 __io_req_complete(req, issue_flags, ret, 0);
4608 static int io_sfr_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4610 struct io_ring_ctx *ctx = req->ctx;
4612 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4614 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4617 req->sync.off = READ_ONCE(sqe->off);
4618 req->sync.len = READ_ONCE(sqe->len);
4619 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4623 static int io_sync_file_range(struct io_kiocb *req, unsigned int issue_flags)
4627 /* sync_file_range always requires a blocking context */
4628 if (issue_flags & IO_URING_F_NONBLOCK)
4631 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4634 req_set_fail_links(req);
4635 io_req_complete(req, ret);
4639 #if defined(CONFIG_NET)
4640 static int io_setup_async_msg(struct io_kiocb *req,
4641 struct io_async_msghdr *kmsg)
4643 struct io_async_msghdr *async_msg = req->async_data;
4647 if (io_alloc_async_data(req)) {
4648 kfree(kmsg->free_iov);
4651 async_msg = req->async_data;
4652 req->flags |= REQ_F_NEED_CLEANUP;
4653 memcpy(async_msg, kmsg, sizeof(*kmsg));
4654 async_msg->msg.msg_name = &async_msg->addr;
4655 /* if were using fast_iov, set it to the new one */
4656 if (!async_msg->free_iov)
4657 async_msg->msg.msg_iter.iov = async_msg->fast_iov;
4662 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4663 struct io_async_msghdr *iomsg)
4665 iomsg->msg.msg_name = &iomsg->addr;
4666 iomsg->free_iov = iomsg->fast_iov;
4667 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4668 req->sr_msg.msg_flags, &iomsg->free_iov);
4671 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4673 struct io_async_msghdr *async_msg = req->async_data;
4674 struct io_sr_msg *sr = &req->sr_msg;
4677 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4680 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4681 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4682 sr->len = READ_ONCE(sqe->len);
4684 #ifdef CONFIG_COMPAT
4685 if (req->ctx->compat)
4686 sr->msg_flags |= MSG_CMSG_COMPAT;
4689 if (!async_msg || !io_op_defs[req->opcode].needs_async_data)
4691 ret = io_sendmsg_copy_hdr(req, async_msg);
4693 req->flags |= REQ_F_NEED_CLEANUP;
4697 static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
4699 struct io_async_msghdr iomsg, *kmsg;
4700 struct socket *sock;
4704 sock = sock_from_file(req->file);
4705 if (unlikely(!sock))
4708 kmsg = req->async_data;
4710 ret = io_sendmsg_copy_hdr(req, &iomsg);
4716 flags = req->sr_msg.msg_flags;
4717 if (flags & MSG_DONTWAIT)
4718 req->flags |= REQ_F_NOWAIT;
4719 else if (issue_flags & IO_URING_F_NONBLOCK)
4720 flags |= MSG_DONTWAIT;
4722 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4723 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4724 return io_setup_async_msg(req, kmsg);
4725 if (ret == -ERESTARTSYS)
4728 /* fast path, check for non-NULL to avoid function call */
4730 kfree(kmsg->free_iov);
4731 req->flags &= ~REQ_F_NEED_CLEANUP;
4733 req_set_fail_links(req);
4734 __io_req_complete(req, issue_flags, ret, 0);
4738 static int io_send(struct io_kiocb *req, unsigned int issue_flags)
4740 struct io_sr_msg *sr = &req->sr_msg;
4743 struct socket *sock;
4747 sock = sock_from_file(req->file);
4748 if (unlikely(!sock))
4751 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4755 msg.msg_name = NULL;
4756 msg.msg_control = NULL;
4757 msg.msg_controllen = 0;
4758 msg.msg_namelen = 0;
4760 flags = req->sr_msg.msg_flags;
4761 if (flags & MSG_DONTWAIT)
4762 req->flags |= REQ_F_NOWAIT;
4763 else if (issue_flags & IO_URING_F_NONBLOCK)
4764 flags |= MSG_DONTWAIT;
4766 msg.msg_flags = flags;
4767 ret = sock_sendmsg(sock, &msg);
4768 if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
4770 if (ret == -ERESTARTSYS)
4774 req_set_fail_links(req);
4775 __io_req_complete(req, issue_flags, ret, 0);
4779 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4780 struct io_async_msghdr *iomsg)
4782 struct io_sr_msg *sr = &req->sr_msg;
4783 struct iovec __user *uiov;
4787 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4788 &iomsg->uaddr, &uiov, &iov_len);
4792 if (req->flags & REQ_F_BUFFER_SELECT) {
4795 if (copy_from_user(iomsg->fast_iov, uiov, sizeof(*uiov)))
4797 sr->len = iomsg->fast_iov[0].iov_len;
4798 iomsg->free_iov = NULL;
4800 iomsg->free_iov = iomsg->fast_iov;
4801 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4802 &iomsg->free_iov, &iomsg->msg.msg_iter,
4811 #ifdef CONFIG_COMPAT
4812 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4813 struct io_async_msghdr *iomsg)
4815 struct compat_msghdr __user *msg_compat;
4816 struct io_sr_msg *sr = &req->sr_msg;
4817 struct compat_iovec __user *uiov;
4822 msg_compat = (struct compat_msghdr __user *) sr->umsg;
4823 ret = __get_compat_msghdr(&iomsg->msg, msg_compat, &iomsg->uaddr,
4828 uiov = compat_ptr(ptr);
4829 if (req->flags & REQ_F_BUFFER_SELECT) {
4830 compat_ssize_t clen;
4834 if (!access_ok(uiov, sizeof(*uiov)))
4836 if (__get_user(clen, &uiov->iov_len))
4841 iomsg->free_iov = NULL;
4843 iomsg->free_iov = iomsg->fast_iov;
4844 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4845 UIO_FASTIOV, &iomsg->free_iov,
4846 &iomsg->msg.msg_iter, true);
4855 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4856 struct io_async_msghdr *iomsg)
4858 iomsg->msg.msg_name = &iomsg->addr;
4860 #ifdef CONFIG_COMPAT
4861 if (req->ctx->compat)
4862 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4865 return __io_recvmsg_copy_hdr(req, iomsg);
4868 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4871 struct io_sr_msg *sr = &req->sr_msg;
4872 struct io_buffer *kbuf;
4874 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4879 req->flags |= REQ_F_BUFFER_SELECTED;
4883 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4885 return io_put_kbuf(req, req->sr_msg.kbuf);
4888 static int io_recvmsg_prep(struct io_kiocb *req,
4889 const struct io_uring_sqe *sqe)
4891 struct io_async_msghdr *async_msg = req->async_data;
4892 struct io_sr_msg *sr = &req->sr_msg;
4895 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4898 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4899 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4900 sr->len = READ_ONCE(sqe->len);
4901 sr->bgid = READ_ONCE(sqe->buf_group);
4903 #ifdef CONFIG_COMPAT
4904 if (req->ctx->compat)
4905 sr->msg_flags |= MSG_CMSG_COMPAT;
4908 if (!async_msg || !io_op_defs[req->opcode].needs_async_data)
4910 ret = io_recvmsg_copy_hdr(req, async_msg);
4912 req->flags |= REQ_F_NEED_CLEANUP;
4916 static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
4918 struct io_async_msghdr iomsg, *kmsg;
4919 struct socket *sock;
4920 struct io_buffer *kbuf;
4922 int ret, cflags = 0;
4923 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4925 sock = sock_from_file(req->file);
4926 if (unlikely(!sock))
4929 kmsg = req->async_data;
4931 ret = io_recvmsg_copy_hdr(req, &iomsg);
4937 if (req->flags & REQ_F_BUFFER_SELECT) {
4938 kbuf = io_recv_buffer_select(req, !force_nonblock);
4940 return PTR_ERR(kbuf);
4941 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4942 kmsg->fast_iov[0].iov_len = req->sr_msg.len;
4943 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->fast_iov,
4944 1, req->sr_msg.len);
4947 flags = req->sr_msg.msg_flags;
4948 if (flags & MSG_DONTWAIT)
4949 req->flags |= REQ_F_NOWAIT;
4950 else if (force_nonblock)
4951 flags |= MSG_DONTWAIT;
4953 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4954 kmsg->uaddr, flags);
4955 if (force_nonblock && ret == -EAGAIN)
4956 return io_setup_async_msg(req, kmsg);
4957 if (ret == -ERESTARTSYS)
4960 if (req->flags & REQ_F_BUFFER_SELECTED)
4961 cflags = io_put_recv_kbuf(req);
4962 /* fast path, check for non-NULL to avoid function call */
4964 kfree(kmsg->free_iov);
4965 req->flags &= ~REQ_F_NEED_CLEANUP;
4967 req_set_fail_links(req);
4968 __io_req_complete(req, issue_flags, ret, cflags);
4972 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
4974 struct io_buffer *kbuf;
4975 struct io_sr_msg *sr = &req->sr_msg;
4977 void __user *buf = sr->buf;
4978 struct socket *sock;
4981 int ret, cflags = 0;
4982 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
4984 sock = sock_from_file(req->file);
4985 if (unlikely(!sock))
4988 if (req->flags & REQ_F_BUFFER_SELECT) {
4989 kbuf = io_recv_buffer_select(req, !force_nonblock);
4991 return PTR_ERR(kbuf);
4992 buf = u64_to_user_ptr(kbuf->addr);
4995 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4999 msg.msg_name = NULL;
5000 msg.msg_control = NULL;
5001 msg.msg_controllen = 0;
5002 msg.msg_namelen = 0;
5003 msg.msg_iocb = NULL;
5006 flags = req->sr_msg.msg_flags;
5007 if (flags & MSG_DONTWAIT)
5008 req->flags |= REQ_F_NOWAIT;
5009 else if (force_nonblock)
5010 flags |= MSG_DONTWAIT;
5012 ret = sock_recvmsg(sock, &msg, flags);
5013 if (force_nonblock && ret == -EAGAIN)
5015 if (ret == -ERESTARTSYS)
5018 if (req->flags & REQ_F_BUFFER_SELECTED)
5019 cflags = io_put_recv_kbuf(req);
5021 req_set_fail_links(req);
5022 __io_req_complete(req, issue_flags, ret, cflags);
5026 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5028 struct io_accept *accept = &req->accept;
5030 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5032 if (sqe->ioprio || sqe->len || sqe->buf_index)
5035 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
5036 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
5037 accept->flags = READ_ONCE(sqe->accept_flags);
5038 accept->nofile = rlimit(RLIMIT_NOFILE);
5042 static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
5044 struct io_accept *accept = &req->accept;
5045 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
5046 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
5049 if (req->file->f_flags & O_NONBLOCK)
5050 req->flags |= REQ_F_NOWAIT;
5052 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
5053 accept->addr_len, accept->flags,
5055 if (ret == -EAGAIN && force_nonblock)
5058 if (ret == -ERESTARTSYS)
5060 req_set_fail_links(req);
5062 __io_req_complete(req, issue_flags, ret, 0);
5066 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5068 struct io_connect *conn = &req->connect;
5069 struct io_async_connect *io = req->async_data;
5071 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5073 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
5076 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
5077 conn->addr_len = READ_ONCE(sqe->addr2);
5082 return move_addr_to_kernel(conn->addr, conn->addr_len,
5086 static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
5088 struct io_async_connect __io, *io;
5089 unsigned file_flags;
5091 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
5093 if (req->async_data) {
5094 io = req->async_data;
5096 ret = move_addr_to_kernel(req->connect.addr,
5097 req->connect.addr_len,
5104 file_flags = force_nonblock ? O_NONBLOCK : 0;
5106 ret = __sys_connect_file(req->file, &io->address,
5107 req->connect.addr_len, file_flags);
5108 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
5109 if (req->async_data)
5111 if (io_alloc_async_data(req)) {
5115 io = req->async_data;
5116 memcpy(req->async_data, &__io, sizeof(__io));
5119 if (ret == -ERESTARTSYS)
5123 req_set_fail_links(req);
5124 __io_req_complete(req, issue_flags, ret, 0);
5127 #else /* !CONFIG_NET */
5128 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5133 static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
5138 static int io_send(struct io_kiocb *req, unsigned int issue_flags)
5143 static int io_recvmsg_prep(struct io_kiocb *req,
5144 const struct io_uring_sqe *sqe)
5149 static int io_recvmsg(struct io_kiocb *req, unsigned int issue_flags)
5154 static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
5159 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5164 static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
5169 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5174 static int io_connect(struct io_kiocb *req, unsigned int issue_flags)
5178 #endif /* CONFIG_NET */
5180 struct io_poll_table {
5181 struct poll_table_struct pt;
5182 struct io_kiocb *req;
5186 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
5187 __poll_t mask, task_work_func_t func)
5191 /* for instances that support it check for an event match first: */
5192 if (mask && !(mask & poll->events))
5195 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
5197 list_del_init(&poll->wait.entry);
5200 req->task_work.func = func;
5201 percpu_ref_get(&req->ctx->refs);
5204 * If this fails, then the task is exiting. When a task exits, the
5205 * work gets canceled, so just cancel this request as well instead
5206 * of executing it. We can't safely execute it anyway, as we may not
5207 * have the needed state needed for it anyway.
5209 ret = io_req_task_work_add(req);
5210 if (unlikely(ret)) {
5211 WRITE_ONCE(poll->canceled, true);
5212 io_req_task_work_add_fallback(req, func);
5217 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
5218 __acquires(&req->ctx->completion_lock)
5220 struct io_ring_ctx *ctx = req->ctx;
5222 if (!req->result && !READ_ONCE(poll->canceled)) {
5223 struct poll_table_struct pt = { ._key = poll->events };
5225 req->result = vfs_poll(req->file, &pt) & poll->events;
5228 spin_lock_irq(&ctx->completion_lock);
5229 if (!req->result && !READ_ONCE(poll->canceled)) {
5230 add_wait_queue(poll->head, &poll->wait);
5237 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
5239 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
5240 if (req->opcode == IORING_OP_POLL_ADD)
5241 return req->async_data;
5242 return req->apoll->double_poll;
5245 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
5247 if (req->opcode == IORING_OP_POLL_ADD)
5249 return &req->apoll->poll;
5252 static void io_poll_remove_double(struct io_kiocb *req)
5254 struct io_poll_iocb *poll = io_poll_get_double(req);
5256 lockdep_assert_held(&req->ctx->completion_lock);
5258 if (poll && poll->head) {
5259 struct wait_queue_head *head = poll->head;
5261 spin_lock(&head->lock);
5262 list_del_init(&poll->wait.entry);
5263 if (poll->wait.private)
5264 refcount_dec(&req->refs);
5266 spin_unlock(&head->lock);
5270 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
5272 struct io_ring_ctx *ctx = req->ctx;
5274 io_poll_remove_double(req);
5275 req->poll.done = true;
5276 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
5277 io_commit_cqring(ctx);
5280 static void io_poll_task_func(struct callback_head *cb)
5282 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
5283 struct io_ring_ctx *ctx = req->ctx;
5284 struct io_kiocb *nxt;
5286 if (io_poll_rewait(req, &req->poll)) {
5287 spin_unlock_irq(&ctx->completion_lock);
5289 hash_del(&req->hash_node);
5290 io_poll_complete(req, req->result, 0);
5291 spin_unlock_irq(&ctx->completion_lock);
5293 nxt = io_put_req_find_next(req);
5294 io_cqring_ev_posted(ctx);
5296 __io_req_task_submit(nxt);
5299 percpu_ref_put(&ctx->refs);
5302 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
5303 int sync, void *key)
5305 struct io_kiocb *req = wait->private;
5306 struct io_poll_iocb *poll = io_poll_get_single(req);
5307 __poll_t mask = key_to_poll(key);
5309 /* for instances that support it check for an event match first: */
5310 if (mask && !(mask & poll->events))
5313 list_del_init(&wait->entry);
5315 if (poll && poll->head) {
5318 spin_lock(&poll->head->lock);
5319 done = list_empty(&poll->wait.entry);
5321 list_del_init(&poll->wait.entry);
5322 /* make sure double remove sees this as being gone */
5323 wait->private = NULL;
5324 spin_unlock(&poll->head->lock);
5326 /* use wait func handler, so it matches the rq type */
5327 poll->wait.func(&poll->wait, mode, sync, key);
5330 refcount_dec(&req->refs);
5334 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
5335 wait_queue_func_t wake_func)
5339 poll->canceled = false;
5340 poll->events = events;
5341 INIT_LIST_HEAD(&poll->wait.entry);
5342 init_waitqueue_func_entry(&poll->wait, wake_func);
5345 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5346 struct wait_queue_head *head,
5347 struct io_poll_iocb **poll_ptr)
5349 struct io_kiocb *req = pt->req;
5352 * If poll->head is already set, it's because the file being polled
5353 * uses multiple waitqueues for poll handling (eg one for read, one
5354 * for write). Setup a separate io_poll_iocb if this happens.
5356 if (unlikely(poll->head)) {
5357 struct io_poll_iocb *poll_one = poll;
5359 /* already have a 2nd entry, fail a third attempt */
5361 pt->error = -EINVAL;
5364 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5366 pt->error = -ENOMEM;
5369 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5370 refcount_inc(&req->refs);
5371 poll->wait.private = req;
5378 if (poll->events & EPOLLEXCLUSIVE)
5379 add_wait_queue_exclusive(head, &poll->wait);
5381 add_wait_queue(head, &poll->wait);
5384 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5385 struct poll_table_struct *p)
5387 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5388 struct async_poll *apoll = pt->req->apoll;
5390 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5393 static void io_async_task_func(struct callback_head *cb)
5395 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
5396 struct async_poll *apoll = req->apoll;
5397 struct io_ring_ctx *ctx = req->ctx;
5399 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
5401 if (io_poll_rewait(req, &apoll->poll)) {
5402 spin_unlock_irq(&ctx->completion_lock);
5403 percpu_ref_put(&ctx->refs);
5407 /* If req is still hashed, it cannot have been canceled. Don't check. */
5408 if (hash_hashed(&req->hash_node))
5409 hash_del(&req->hash_node);
5411 io_poll_remove_double(req);
5412 spin_unlock_irq(&ctx->completion_lock);
5414 if (!READ_ONCE(apoll->poll.canceled))
5415 __io_req_task_submit(req);
5417 __io_req_task_cancel(req, -ECANCELED);
5419 percpu_ref_put(&ctx->refs);
5420 kfree(apoll->double_poll);
5424 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5427 struct io_kiocb *req = wait->private;
5428 struct io_poll_iocb *poll = &req->apoll->poll;
5430 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5433 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5436 static void io_poll_req_insert(struct io_kiocb *req)
5438 struct io_ring_ctx *ctx = req->ctx;
5439 struct hlist_head *list;
5441 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5442 hlist_add_head(&req->hash_node, list);
5445 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5446 struct io_poll_iocb *poll,
5447 struct io_poll_table *ipt, __poll_t mask,
5448 wait_queue_func_t wake_func)
5449 __acquires(&ctx->completion_lock)
5451 struct io_ring_ctx *ctx = req->ctx;
5452 bool cancel = false;
5454 INIT_HLIST_NODE(&req->hash_node);
5455 io_init_poll_iocb(poll, mask, wake_func);
5456 poll->file = req->file;
5457 poll->wait.private = req;
5459 ipt->pt._key = mask;
5461 ipt->error = -EINVAL;
5463 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5465 spin_lock_irq(&ctx->completion_lock);
5466 if (likely(poll->head)) {
5467 spin_lock(&poll->head->lock);
5468 if (unlikely(list_empty(&poll->wait.entry))) {
5474 if (mask || ipt->error)
5475 list_del_init(&poll->wait.entry);
5477 WRITE_ONCE(poll->canceled, true);
5478 else if (!poll->done) /* actually waiting for an event */
5479 io_poll_req_insert(req);
5480 spin_unlock(&poll->head->lock);
5486 static bool io_arm_poll_handler(struct io_kiocb *req)
5488 const struct io_op_def *def = &io_op_defs[req->opcode];
5489 struct io_ring_ctx *ctx = req->ctx;
5490 struct async_poll *apoll;
5491 struct io_poll_table ipt;
5495 if (!req->file || !file_can_poll(req->file))
5497 if (req->flags & REQ_F_POLLED)
5501 else if (def->pollout)
5505 /* if we can't nonblock try, then no point in arming a poll handler */
5506 if (!io_file_supports_async(req->file, rw))
5509 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5510 if (unlikely(!apoll))
5512 apoll->double_poll = NULL;
5514 req->flags |= REQ_F_POLLED;
5519 mask |= POLLIN | POLLRDNORM;
5521 mask |= POLLOUT | POLLWRNORM;
5523 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5524 if ((req->opcode == IORING_OP_RECVMSG) &&
5525 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5528 mask |= POLLERR | POLLPRI;
5530 ipt.pt._qproc = io_async_queue_proc;
5532 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5534 if (ret || ipt.error) {
5535 io_poll_remove_double(req);
5536 spin_unlock_irq(&ctx->completion_lock);
5537 kfree(apoll->double_poll);
5541 spin_unlock_irq(&ctx->completion_lock);
5542 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
5543 apoll->poll.events);
5547 static bool __io_poll_remove_one(struct io_kiocb *req,
5548 struct io_poll_iocb *poll)
5550 bool do_complete = false;
5552 spin_lock(&poll->head->lock);
5553 WRITE_ONCE(poll->canceled, true);
5554 if (!list_empty(&poll->wait.entry)) {
5555 list_del_init(&poll->wait.entry);
5558 spin_unlock(&poll->head->lock);
5559 hash_del(&req->hash_node);
5563 static bool io_poll_remove_one(struct io_kiocb *req)
5567 io_poll_remove_double(req);
5569 if (req->opcode == IORING_OP_POLL_ADD) {
5570 do_complete = __io_poll_remove_one(req, &req->poll);
5572 struct async_poll *apoll = req->apoll;
5574 /* non-poll requests have submit ref still */
5575 do_complete = __io_poll_remove_one(req, &apoll->poll);
5578 kfree(apoll->double_poll);
5584 io_cqring_fill_event(req, -ECANCELED);
5585 io_commit_cqring(req->ctx);
5586 req_set_fail_links(req);
5587 io_put_req_deferred(req, 1);
5594 * Returns true if we found and killed one or more poll requests
5596 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5597 struct files_struct *files)
5599 struct hlist_node *tmp;
5600 struct io_kiocb *req;
5603 spin_lock_irq(&ctx->completion_lock);
5604 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5605 struct hlist_head *list;
5607 list = &ctx->cancel_hash[i];
5608 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5609 if (io_match_task(req, tsk, files))
5610 posted += io_poll_remove_one(req);
5613 spin_unlock_irq(&ctx->completion_lock);
5616 io_cqring_ev_posted(ctx);
5621 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
5623 struct hlist_head *list;
5624 struct io_kiocb *req;
5626 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5627 hlist_for_each_entry(req, list, hash_node) {
5628 if (sqe_addr != req->user_data)
5630 if (io_poll_remove_one(req))
5638 static int io_poll_remove_prep(struct io_kiocb *req,
5639 const struct io_uring_sqe *sqe)
5641 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5643 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
5647 req->poll_remove.addr = READ_ONCE(sqe->addr);
5652 * Find a running poll command that matches one specified in sqe->addr,
5653 * and remove it if found.
5655 static int io_poll_remove(struct io_kiocb *req, unsigned int issue_flags)
5657 struct io_ring_ctx *ctx = req->ctx;
5660 spin_lock_irq(&ctx->completion_lock);
5661 ret = io_poll_cancel(ctx, req->poll_remove.addr);
5662 spin_unlock_irq(&ctx->completion_lock);
5665 req_set_fail_links(req);
5666 io_req_complete(req, ret);
5670 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5673 struct io_kiocb *req = wait->private;
5674 struct io_poll_iocb *poll = &req->poll;
5676 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5679 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5680 struct poll_table_struct *p)
5682 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5684 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5687 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5689 struct io_poll_iocb *poll = &req->poll;
5692 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5694 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
5697 events = READ_ONCE(sqe->poll32_events);
5699 events = swahw32(events);
5701 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP |
5702 (events & EPOLLEXCLUSIVE);
5706 static int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
5708 struct io_poll_iocb *poll = &req->poll;
5709 struct io_ring_ctx *ctx = req->ctx;
5710 struct io_poll_table ipt;
5713 ipt.pt._qproc = io_poll_queue_proc;
5715 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5718 if (mask) { /* no async, we'd stolen it */
5720 io_poll_complete(req, mask, 0);
5722 spin_unlock_irq(&ctx->completion_lock);
5725 io_cqring_ev_posted(ctx);
5731 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5733 struct io_timeout_data *data = container_of(timer,
5734 struct io_timeout_data, timer);
5735 struct io_kiocb *req = data->req;
5736 struct io_ring_ctx *ctx = req->ctx;
5737 unsigned long flags;
5739 spin_lock_irqsave(&ctx->completion_lock, flags);
5740 list_del_init(&req->timeout.list);
5741 atomic_set(&req->ctx->cq_timeouts,
5742 atomic_read(&req->ctx->cq_timeouts) + 1);
5744 io_cqring_fill_event(req, -ETIME);
5745 io_commit_cqring(ctx);
5746 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5748 io_cqring_ev_posted(ctx);
5749 req_set_fail_links(req);
5751 return HRTIMER_NORESTART;
5754 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5757 struct io_timeout_data *io;
5758 struct io_kiocb *req;
5761 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5762 if (user_data == req->user_data) {
5769 return ERR_PTR(ret);
5771 io = req->async_data;
5772 ret = hrtimer_try_to_cancel(&io->timer);
5774 return ERR_PTR(-EALREADY);
5775 list_del_init(&req->timeout.list);
5779 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5781 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5784 return PTR_ERR(req);
5786 req_set_fail_links(req);
5787 io_cqring_fill_event(req, -ECANCELED);
5788 io_put_req_deferred(req, 1);
5792 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5793 struct timespec64 *ts, enum hrtimer_mode mode)
5795 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5796 struct io_timeout_data *data;
5799 return PTR_ERR(req);
5801 req->timeout.off = 0; /* noseq */
5802 data = req->async_data;
5803 list_add_tail(&req->timeout.list, &ctx->timeout_list);
5804 hrtimer_init(&data->timer, CLOCK_MONOTONIC, mode);
5805 data->timer.function = io_timeout_fn;
5806 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
5810 static int io_timeout_remove_prep(struct io_kiocb *req,
5811 const struct io_uring_sqe *sqe)
5813 struct io_timeout_rem *tr = &req->timeout_rem;
5815 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5817 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5819 if (sqe->ioprio || sqe->buf_index || sqe->len)
5822 tr->addr = READ_ONCE(sqe->addr);
5823 tr->flags = READ_ONCE(sqe->timeout_flags);
5824 if (tr->flags & IORING_TIMEOUT_UPDATE) {
5825 if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS))
5827 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
5829 } else if (tr->flags) {
5830 /* timeout removal doesn't support flags */
5837 static inline enum hrtimer_mode io_translate_timeout_mode(unsigned int flags)
5839 return (flags & IORING_TIMEOUT_ABS) ? HRTIMER_MODE_ABS
5844 * Remove or update an existing timeout command
5846 static int io_timeout_remove(struct io_kiocb *req, unsigned int issue_flags)
5848 struct io_timeout_rem *tr = &req->timeout_rem;
5849 struct io_ring_ctx *ctx = req->ctx;
5852 spin_lock_irq(&ctx->completion_lock);
5853 if (!(req->timeout_rem.flags & IORING_TIMEOUT_UPDATE))
5854 ret = io_timeout_cancel(ctx, tr->addr);
5856 ret = io_timeout_update(ctx, tr->addr, &tr->ts,
5857 io_translate_timeout_mode(tr->flags));
5859 io_cqring_fill_event(req, ret);
5860 io_commit_cqring(ctx);
5861 spin_unlock_irq(&ctx->completion_lock);
5862 io_cqring_ev_posted(ctx);
5864 req_set_fail_links(req);
5869 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5870 bool is_timeout_link)
5872 struct io_timeout_data *data;
5874 u32 off = READ_ONCE(sqe->off);
5876 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5878 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5880 if (off && is_timeout_link)
5882 flags = READ_ONCE(sqe->timeout_flags);
5883 if (flags & ~IORING_TIMEOUT_ABS)
5886 req->timeout.off = off;
5888 if (!req->async_data && io_alloc_async_data(req))
5891 data = req->async_data;
5894 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5897 data->mode = io_translate_timeout_mode(flags);
5898 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5902 static int io_timeout(struct io_kiocb *req, unsigned int issue_flags)
5904 struct io_ring_ctx *ctx = req->ctx;
5905 struct io_timeout_data *data = req->async_data;
5906 struct list_head *entry;
5907 u32 tail, off = req->timeout.off;
5909 spin_lock_irq(&ctx->completion_lock);
5912 * sqe->off holds how many events that need to occur for this
5913 * timeout event to be satisfied. If it isn't set, then this is
5914 * a pure timeout request, sequence isn't used.
5916 if (io_is_timeout_noseq(req)) {
5917 entry = ctx->timeout_list.prev;
5921 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5922 req->timeout.target_seq = tail + off;
5924 /* Update the last seq here in case io_flush_timeouts() hasn't.
5925 * This is safe because ->completion_lock is held, and submissions
5926 * and completions are never mixed in the same ->completion_lock section.
5928 ctx->cq_last_tm_flush = tail;
5931 * Insertion sort, ensuring the first entry in the list is always
5932 * the one we need first.
5934 list_for_each_prev(entry, &ctx->timeout_list) {
5935 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5938 if (io_is_timeout_noseq(nxt))
5940 /* nxt.seq is behind @tail, otherwise would've been completed */
5941 if (off >= nxt->timeout.target_seq - tail)
5945 list_add(&req->timeout.list, entry);
5946 data->timer.function = io_timeout_fn;
5947 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5948 spin_unlock_irq(&ctx->completion_lock);
5952 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5954 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5956 return req->user_data == (unsigned long) data;
5959 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
5961 enum io_wq_cancel cancel_ret;
5964 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr, false);
5965 switch (cancel_ret) {
5966 case IO_WQ_CANCEL_OK:
5969 case IO_WQ_CANCEL_RUNNING:
5972 case IO_WQ_CANCEL_NOTFOUND:
5980 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5981 struct io_kiocb *req, __u64 sqe_addr,
5984 unsigned long flags;
5987 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
5988 if (ret != -ENOENT) {
5989 spin_lock_irqsave(&ctx->completion_lock, flags);
5993 spin_lock_irqsave(&ctx->completion_lock, flags);
5994 ret = io_timeout_cancel(ctx, sqe_addr);
5997 ret = io_poll_cancel(ctx, sqe_addr);
6001 io_cqring_fill_event(req, ret);
6002 io_commit_cqring(ctx);
6003 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6004 io_cqring_ev_posted(ctx);
6007 req_set_fail_links(req);
6011 static int io_async_cancel_prep(struct io_kiocb *req,
6012 const struct io_uring_sqe *sqe)
6014 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
6016 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
6018 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
6021 req->cancel.addr = READ_ONCE(sqe->addr);
6025 static int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
6027 struct io_ring_ctx *ctx = req->ctx;
6029 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
6033 static int io_rsrc_update_prep(struct io_kiocb *req,
6034 const struct io_uring_sqe *sqe)
6036 if (unlikely(req->ctx->flags & IORING_SETUP_SQPOLL))
6038 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
6040 if (sqe->ioprio || sqe->rw_flags)
6043 req->rsrc_update.offset = READ_ONCE(sqe->off);
6044 req->rsrc_update.nr_args = READ_ONCE(sqe->len);
6045 if (!req->rsrc_update.nr_args)
6047 req->rsrc_update.arg = READ_ONCE(sqe->addr);
6051 static int io_files_update(struct io_kiocb *req, unsigned int issue_flags)
6053 struct io_ring_ctx *ctx = req->ctx;
6054 struct io_uring_rsrc_update up;
6057 if (issue_flags & IO_URING_F_NONBLOCK)
6060 up.offset = req->rsrc_update.offset;
6061 up.data = req->rsrc_update.arg;
6063 mutex_lock(&ctx->uring_lock);
6064 ret = __io_sqe_files_update(ctx, &up, req->rsrc_update.nr_args);
6065 mutex_unlock(&ctx->uring_lock);
6068 req_set_fail_links(req);
6069 __io_req_complete(req, issue_flags, ret, 0);
6073 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
6075 switch (req->opcode) {
6078 case IORING_OP_READV:
6079 case IORING_OP_READ_FIXED:
6080 case IORING_OP_READ:
6081 return io_read_prep(req, sqe);
6082 case IORING_OP_WRITEV:
6083 case IORING_OP_WRITE_FIXED:
6084 case IORING_OP_WRITE:
6085 return io_write_prep(req, sqe);
6086 case IORING_OP_POLL_ADD:
6087 return io_poll_add_prep(req, sqe);
6088 case IORING_OP_POLL_REMOVE:
6089 return io_poll_remove_prep(req, sqe);
6090 case IORING_OP_FSYNC:
6091 return io_fsync_prep(req, sqe);
6092 case IORING_OP_SYNC_FILE_RANGE:
6093 return io_sfr_prep(req, sqe);
6094 case IORING_OP_SENDMSG:
6095 case IORING_OP_SEND:
6096 return io_sendmsg_prep(req, sqe);
6097 case IORING_OP_RECVMSG:
6098 case IORING_OP_RECV:
6099 return io_recvmsg_prep(req, sqe);
6100 case IORING_OP_CONNECT:
6101 return io_connect_prep(req, sqe);
6102 case IORING_OP_TIMEOUT:
6103 return io_timeout_prep(req, sqe, false);
6104 case IORING_OP_TIMEOUT_REMOVE:
6105 return io_timeout_remove_prep(req, sqe);
6106 case IORING_OP_ASYNC_CANCEL:
6107 return io_async_cancel_prep(req, sqe);
6108 case IORING_OP_LINK_TIMEOUT:
6109 return io_timeout_prep(req, sqe, true);
6110 case IORING_OP_ACCEPT:
6111 return io_accept_prep(req, sqe);
6112 case IORING_OP_FALLOCATE:
6113 return io_fallocate_prep(req, sqe);
6114 case IORING_OP_OPENAT:
6115 return io_openat_prep(req, sqe);
6116 case IORING_OP_CLOSE:
6117 return io_close_prep(req, sqe);
6118 case IORING_OP_FILES_UPDATE:
6119 return io_rsrc_update_prep(req, sqe);
6120 case IORING_OP_STATX:
6121 return io_statx_prep(req, sqe);
6122 case IORING_OP_FADVISE:
6123 return io_fadvise_prep(req, sqe);
6124 case IORING_OP_MADVISE:
6125 return io_madvise_prep(req, sqe);
6126 case IORING_OP_OPENAT2:
6127 return io_openat2_prep(req, sqe);
6128 case IORING_OP_EPOLL_CTL:
6129 return io_epoll_ctl_prep(req, sqe);
6130 case IORING_OP_SPLICE:
6131 return io_splice_prep(req, sqe);
6132 case IORING_OP_PROVIDE_BUFFERS:
6133 return io_provide_buffers_prep(req, sqe);
6134 case IORING_OP_REMOVE_BUFFERS:
6135 return io_remove_buffers_prep(req, sqe);
6137 return io_tee_prep(req, sqe);
6138 case IORING_OP_SHUTDOWN:
6139 return io_shutdown_prep(req, sqe);
6140 case IORING_OP_RENAMEAT:
6141 return io_renameat_prep(req, sqe);
6142 case IORING_OP_UNLINKAT:
6143 return io_unlinkat_prep(req, sqe);
6146 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
6151 static int io_req_defer_prep(struct io_kiocb *req,
6152 const struct io_uring_sqe *sqe)
6156 if (io_alloc_async_data(req))
6158 return io_req_prep(req, sqe);
6161 static u32 io_get_sequence(struct io_kiocb *req)
6163 struct io_kiocb *pos;
6164 struct io_ring_ctx *ctx = req->ctx;
6165 u32 total_submitted, nr_reqs = 0;
6167 io_for_each_link(pos, req)
6170 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
6171 return total_submitted - nr_reqs;
6174 static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe)
6176 struct io_ring_ctx *ctx = req->ctx;
6177 struct io_defer_entry *de;
6181 /* Still need defer if there is pending req in defer list. */
6182 if (likely(list_empty_careful(&ctx->defer_list) &&
6183 !(req->flags & REQ_F_IO_DRAIN)))
6186 seq = io_get_sequence(req);
6187 /* Still a chance to pass the sequence check */
6188 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
6191 if (!req->async_data) {
6192 ret = io_req_defer_prep(req, sqe);
6196 io_prep_async_link(req);
6197 de = kmalloc(sizeof(*de), GFP_KERNEL);
6201 spin_lock_irq(&ctx->completion_lock);
6202 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
6203 spin_unlock_irq(&ctx->completion_lock);
6205 io_queue_async_work(req);
6206 return -EIOCBQUEUED;
6209 trace_io_uring_defer(ctx, req, req->user_data);
6212 list_add_tail(&de->list, &ctx->defer_list);
6213 spin_unlock_irq(&ctx->completion_lock);
6214 return -EIOCBQUEUED;
6217 static void __io_clean_op(struct io_kiocb *req)
6219 if (req->flags & REQ_F_BUFFER_SELECTED) {
6220 switch (req->opcode) {
6221 case IORING_OP_READV:
6222 case IORING_OP_READ_FIXED:
6223 case IORING_OP_READ:
6224 kfree((void *)(unsigned long)req->rw.addr);
6226 case IORING_OP_RECVMSG:
6227 case IORING_OP_RECV:
6228 kfree(req->sr_msg.kbuf);
6231 req->flags &= ~REQ_F_BUFFER_SELECTED;
6234 if (req->flags & REQ_F_NEED_CLEANUP) {
6235 switch (req->opcode) {
6236 case IORING_OP_READV:
6237 case IORING_OP_READ_FIXED:
6238 case IORING_OP_READ:
6239 case IORING_OP_WRITEV:
6240 case IORING_OP_WRITE_FIXED:
6241 case IORING_OP_WRITE: {
6242 struct io_async_rw *io = req->async_data;
6244 kfree(io->free_iovec);
6247 case IORING_OP_RECVMSG:
6248 case IORING_OP_SENDMSG: {
6249 struct io_async_msghdr *io = req->async_data;
6251 kfree(io->free_iov);
6254 case IORING_OP_SPLICE:
6256 io_put_file(req, req->splice.file_in,
6257 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
6259 case IORING_OP_OPENAT:
6260 case IORING_OP_OPENAT2:
6261 if (req->open.filename)
6262 putname(req->open.filename);
6264 case IORING_OP_RENAMEAT:
6265 putname(req->rename.oldpath);
6266 putname(req->rename.newpath);
6268 case IORING_OP_UNLINKAT:
6269 putname(req->unlink.filename);
6272 req->flags &= ~REQ_F_NEED_CLEANUP;
6276 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
6278 struct io_ring_ctx *ctx = req->ctx;
6281 switch (req->opcode) {
6283 ret = io_nop(req, issue_flags);
6285 case IORING_OP_READV:
6286 case IORING_OP_READ_FIXED:
6287 case IORING_OP_READ:
6288 ret = io_read(req, issue_flags);
6290 case IORING_OP_WRITEV:
6291 case IORING_OP_WRITE_FIXED:
6292 case IORING_OP_WRITE:
6293 ret = io_write(req, issue_flags);
6295 case IORING_OP_FSYNC:
6296 ret = io_fsync(req, issue_flags);
6298 case IORING_OP_POLL_ADD:
6299 ret = io_poll_add(req, issue_flags);
6301 case IORING_OP_POLL_REMOVE:
6302 ret = io_poll_remove(req, issue_flags);
6304 case IORING_OP_SYNC_FILE_RANGE:
6305 ret = io_sync_file_range(req, issue_flags);
6307 case IORING_OP_SENDMSG:
6308 ret = io_sendmsg(req, issue_flags);
6310 case IORING_OP_SEND:
6311 ret = io_send(req, issue_flags);
6313 case IORING_OP_RECVMSG:
6314 ret = io_recvmsg(req, issue_flags);
6316 case IORING_OP_RECV:
6317 ret = io_recv(req, issue_flags);
6319 case IORING_OP_TIMEOUT:
6320 ret = io_timeout(req, issue_flags);
6322 case IORING_OP_TIMEOUT_REMOVE:
6323 ret = io_timeout_remove(req, issue_flags);
6325 case IORING_OP_ACCEPT:
6326 ret = io_accept(req, issue_flags);
6328 case IORING_OP_CONNECT:
6329 ret = io_connect(req, issue_flags);
6331 case IORING_OP_ASYNC_CANCEL:
6332 ret = io_async_cancel(req, issue_flags);
6334 case IORING_OP_FALLOCATE:
6335 ret = io_fallocate(req, issue_flags);
6337 case IORING_OP_OPENAT:
6338 ret = io_openat(req, issue_flags);
6340 case IORING_OP_CLOSE:
6341 ret = io_close(req, issue_flags);
6343 case IORING_OP_FILES_UPDATE:
6344 ret = io_files_update(req, issue_flags);
6346 case IORING_OP_STATX:
6347 ret = io_statx(req, issue_flags);
6349 case IORING_OP_FADVISE:
6350 ret = io_fadvise(req, issue_flags);
6352 case IORING_OP_MADVISE:
6353 ret = io_madvise(req, issue_flags);
6355 case IORING_OP_OPENAT2:
6356 ret = io_openat2(req, issue_flags);
6358 case IORING_OP_EPOLL_CTL:
6359 ret = io_epoll_ctl(req, issue_flags);
6361 case IORING_OP_SPLICE:
6362 ret = io_splice(req, issue_flags);
6364 case IORING_OP_PROVIDE_BUFFERS:
6365 ret = io_provide_buffers(req, issue_flags);
6367 case IORING_OP_REMOVE_BUFFERS:
6368 ret = io_remove_buffers(req, issue_flags);
6371 ret = io_tee(req, issue_flags);
6373 case IORING_OP_SHUTDOWN:
6374 ret = io_shutdown(req, issue_flags);
6376 case IORING_OP_RENAMEAT:
6377 ret = io_renameat(req, issue_flags);
6379 case IORING_OP_UNLINKAT:
6380 ret = io_unlinkat(req, issue_flags);
6390 /* If the op doesn't have a file, we're not polling for it */
6391 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
6392 const bool in_async = io_wq_current_is_worker();
6394 /* workqueue context doesn't hold uring_lock, grab it now */
6396 mutex_lock(&ctx->uring_lock);
6398 io_iopoll_req_issued(req, in_async);
6401 mutex_unlock(&ctx->uring_lock);
6407 static void io_wq_submit_work(struct io_wq_work *work)
6409 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6410 struct io_kiocb *timeout;
6413 timeout = io_prep_linked_timeout(req);
6415 io_queue_linked_timeout(timeout);
6417 if (work->flags & IO_WQ_WORK_CANCEL)
6422 ret = io_issue_sqe(req, 0);
6424 * We can get EAGAIN for polled IO even though we're
6425 * forcing a sync submission from here, since we can't
6426 * wait for request slots on the block side.
6435 struct io_ring_ctx *lock_ctx = NULL;
6437 if (req->ctx->flags & IORING_SETUP_IOPOLL)
6438 lock_ctx = req->ctx;
6441 * io_iopoll_complete() does not hold completion_lock to
6442 * complete polled io, so here for polled io, we can not call
6443 * io_req_complete() directly, otherwise there maybe concurrent
6444 * access to cqring, defer_list, etc, which is not safe. Given
6445 * that io_iopoll_complete() is always called under uring_lock,
6446 * so here for polled io, we also get uring_lock to complete
6450 mutex_lock(&lock_ctx->uring_lock);
6452 req_set_fail_links(req);
6453 io_req_complete(req, ret);
6456 mutex_unlock(&lock_ctx->uring_lock);
6460 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6463 struct fixed_rsrc_table *table;
6465 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
6466 return table->files[index & IORING_FILE_TABLE_MASK];
6469 static struct file *io_file_get(struct io_submit_state *state,
6470 struct io_kiocb *req, int fd, bool fixed)
6472 struct io_ring_ctx *ctx = req->ctx;
6476 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6478 fd = array_index_nospec(fd, ctx->nr_user_files);
6479 file = io_file_from_index(ctx, fd);
6480 io_set_resource_node(req);
6482 trace_io_uring_file_get(ctx, fd);
6483 file = __io_file_get(state, fd);
6486 if (file && unlikely(file->f_op == &io_uring_fops))
6487 io_req_track_inflight(req);
6491 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6493 struct io_timeout_data *data = container_of(timer,
6494 struct io_timeout_data, timer);
6495 struct io_kiocb *prev, *req = data->req;
6496 struct io_ring_ctx *ctx = req->ctx;
6497 unsigned long flags;
6499 spin_lock_irqsave(&ctx->completion_lock, flags);
6500 prev = req->timeout.head;
6501 req->timeout.head = NULL;
6504 * We don't expect the list to be empty, that will only happen if we
6505 * race with the completion of the linked work.
6507 if (prev && refcount_inc_not_zero(&prev->refs))
6508 io_remove_next_linked(prev);
6511 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6514 req_set_fail_links(prev);
6515 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6516 io_put_req_deferred(prev, 1);
6518 io_req_complete_post(req, -ETIME, 0);
6519 io_put_req_deferred(req, 1);
6521 return HRTIMER_NORESTART;
6524 static void __io_queue_linked_timeout(struct io_kiocb *req)
6527 * If the back reference is NULL, then our linked request finished
6528 * before we got a chance to setup the timer
6530 if (req->timeout.head) {
6531 struct io_timeout_data *data = req->async_data;
6533 data->timer.function = io_link_timeout_fn;
6534 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6539 static void io_queue_linked_timeout(struct io_kiocb *req)
6541 struct io_ring_ctx *ctx = req->ctx;
6543 spin_lock_irq(&ctx->completion_lock);
6544 __io_queue_linked_timeout(req);
6545 spin_unlock_irq(&ctx->completion_lock);
6547 /* drop submission reference */
6551 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6553 struct io_kiocb *nxt = req->link;
6555 if (!nxt || (req->flags & REQ_F_LINK_TIMEOUT) ||
6556 nxt->opcode != IORING_OP_LINK_TIMEOUT)
6559 nxt->timeout.head = req;
6560 nxt->flags |= REQ_F_LTIMEOUT_ACTIVE;
6561 req->flags |= REQ_F_LINK_TIMEOUT;
6565 static void __io_queue_sqe(struct io_kiocb *req)
6567 struct io_kiocb *linked_timeout = io_prep_linked_timeout(req);
6568 const struct cred *old_creds = NULL;
6571 if ((req->flags & REQ_F_WORK_INITIALIZED) &&
6572 (req->work.flags & IO_WQ_WORK_CREDS) &&
6573 req->work.identity->creds != current_cred())
6574 old_creds = override_creds(req->work.identity->creds);
6576 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
6579 revert_creds(old_creds);
6582 * We async punt it if the file wasn't marked NOWAIT, or if the file
6583 * doesn't support non-blocking read/write attempts
6585 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6586 if (!io_arm_poll_handler(req)) {
6588 * Queued up for async execution, worker will release
6589 * submit reference when the iocb is actually submitted.
6591 io_queue_async_work(req);
6593 } else if (likely(!ret)) {
6594 /* drop submission reference */
6595 if (req->flags & REQ_F_COMPLETE_INLINE) {
6596 struct io_ring_ctx *ctx = req->ctx;
6597 struct io_comp_state *cs = &ctx->submit_state.comp;
6599 cs->reqs[cs->nr++] = req;
6600 if (cs->nr == ARRAY_SIZE(cs->reqs))
6601 io_submit_flush_completions(cs, ctx);
6606 req_set_fail_links(req);
6608 io_req_complete(req, ret);
6611 io_queue_linked_timeout(linked_timeout);
6614 static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe)
6618 ret = io_req_defer(req, sqe);
6620 if (ret != -EIOCBQUEUED) {
6622 req_set_fail_links(req);
6624 io_req_complete(req, ret);
6626 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6627 if (!req->async_data) {
6628 ret = io_req_defer_prep(req, sqe);
6632 io_queue_async_work(req);
6635 ret = io_req_prep(req, sqe);
6639 __io_queue_sqe(req);
6643 static inline void io_queue_link_head(struct io_kiocb *req)
6645 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
6647 io_req_complete(req, -ECANCELED);
6649 io_queue_sqe(req, NULL);
6653 * Check SQE restrictions (opcode and flags).
6655 * Returns 'true' if SQE is allowed, 'false' otherwise.
6657 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6658 struct io_kiocb *req,
6659 unsigned int sqe_flags)
6661 if (!ctx->restricted)
6664 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6667 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6668 ctx->restrictions.sqe_flags_required)
6671 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6672 ctx->restrictions.sqe_flags_required))
6678 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6679 const struct io_uring_sqe *sqe)
6681 struct io_submit_state *state;
6682 unsigned int sqe_flags;
6685 req->opcode = READ_ONCE(sqe->opcode);
6686 /* same numerical values with corresponding REQ_F_*, safe to copy */
6687 req->flags = sqe_flags = READ_ONCE(sqe->flags);
6688 req->user_data = READ_ONCE(sqe->user_data);
6689 req->async_data = NULL;
6693 req->fixed_rsrc_refs = NULL;
6694 /* one is dropped after submission, the other at completion */
6695 refcount_set(&req->refs, 2);
6696 req->task = current;
6699 /* enforce forwards compatibility on users */
6700 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6703 if (unlikely(req->opcode >= IORING_OP_LAST))
6706 if (unlikely(io_sq_thread_acquire_mm_files(ctx, req)))
6709 if (unlikely(!io_check_restriction(ctx, req, sqe_flags)))
6712 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6713 !io_op_defs[req->opcode].buffer_select)
6716 id = READ_ONCE(sqe->personality);
6718 struct io_identity *iod;
6720 iod = idr_find(&ctx->personality_idr, id);
6723 refcount_inc(&iod->count);
6725 __io_req_init_async(req);
6726 get_cred(iod->creds);
6727 req->work.identity = iod;
6728 req->work.flags |= IO_WQ_WORK_CREDS;
6731 state = &ctx->submit_state;
6734 * Plug now if we have more than 1 IO left after this, and the target
6735 * is potentially a read/write to block based storage.
6737 if (!state->plug_started && state->ios_left > 1 &&
6738 io_op_defs[req->opcode].plug) {
6739 blk_start_plug(&state->plug);
6740 state->plug_started = true;
6743 if (io_op_defs[req->opcode].needs_file) {
6744 bool fixed = req->flags & REQ_F_FIXED_FILE;
6746 req->file = io_file_get(state, req, READ_ONCE(sqe->fd), fixed);
6747 if (unlikely(!req->file))
6755 static int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
6756 const struct io_uring_sqe *sqe)
6758 struct io_submit_link *link = &ctx->submit_state.link;
6761 ret = io_init_req(ctx, req, sqe);
6762 if (unlikely(ret)) {
6765 io_req_complete(req, ret);
6769 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6770 true, ctx->flags & IORING_SETUP_SQPOLL);
6773 * If we already have a head request, queue this one for async
6774 * submittal once the head completes. If we don't have a head but
6775 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6776 * submitted sync once the chain is complete. If none of those
6777 * conditions are true (normal request), then just queue it.
6780 struct io_kiocb *head = link->head;
6783 * Taking sequential execution of a link, draining both sides
6784 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6785 * requests in the link. So, it drains the head and the
6786 * next after the link request. The last one is done via
6787 * drain_next flag to persist the effect across calls.
6789 if (req->flags & REQ_F_IO_DRAIN) {
6790 head->flags |= REQ_F_IO_DRAIN;
6791 ctx->drain_next = 1;
6793 ret = io_req_defer_prep(req, sqe);
6794 if (unlikely(ret)) {
6795 /* fail even hard links since we don't submit */
6796 head->flags |= REQ_F_FAIL_LINK;
6799 trace_io_uring_link(ctx, req, head);
6800 link->last->link = req;
6803 /* last request of a link, enqueue the link */
6804 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6805 io_queue_link_head(head);
6809 if (unlikely(ctx->drain_next)) {
6810 req->flags |= REQ_F_IO_DRAIN;
6811 ctx->drain_next = 0;
6813 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6814 ret = io_req_defer_prep(req, sqe);
6816 req->flags |= REQ_F_FAIL_LINK;
6820 io_queue_sqe(req, sqe);
6828 * Batched submission is done, ensure local IO is flushed out.
6830 static void io_submit_state_end(struct io_submit_state *state,
6831 struct io_ring_ctx *ctx)
6833 if (state->link.head)
6834 io_queue_link_head(state->link.head);
6836 io_submit_flush_completions(&state->comp, ctx);
6837 if (state->plug_started)
6838 blk_finish_plug(&state->plug);
6839 io_state_file_put(state);
6843 * Start submission side cache.
6845 static void io_submit_state_start(struct io_submit_state *state,
6846 unsigned int max_ios)
6848 state->plug_started = false;
6849 state->ios_left = max_ios;
6850 /* set only head, no need to init link_last in advance */
6851 state->link.head = NULL;
6854 static void io_commit_sqring(struct io_ring_ctx *ctx)
6856 struct io_rings *rings = ctx->rings;
6859 * Ensure any loads from the SQEs are done at this point,
6860 * since once we write the new head, the application could
6861 * write new data to them.
6863 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6867 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6868 * that is mapped by userspace. This means that care needs to be taken to
6869 * ensure that reads are stable, as we cannot rely on userspace always
6870 * being a good citizen. If members of the sqe are validated and then later
6871 * used, it's important that those reads are done through READ_ONCE() to
6872 * prevent a re-load down the line.
6874 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6876 u32 *sq_array = ctx->sq_array;
6880 * The cached sq head (or cq tail) serves two purposes:
6882 * 1) allows us to batch the cost of updating the user visible
6884 * 2) allows the kernel side to track the head on its own, even
6885 * though the application is the one updating it.
6887 head = READ_ONCE(sq_array[ctx->cached_sq_head++ & ctx->sq_mask]);
6888 if (likely(head < ctx->sq_entries))
6889 return &ctx->sq_sqes[head];
6891 /* drop invalid entries */
6892 ctx->cached_sq_dropped++;
6893 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6897 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6901 /* if we have a backlog and couldn't flush it all, return BUSY */
6902 if (test_bit(0, &ctx->sq_check_overflow)) {
6903 if (!__io_cqring_overflow_flush(ctx, false, NULL, NULL))
6907 /* make sure SQ entry isn't read before tail */
6908 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6910 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6913 percpu_counter_add(¤t->io_uring->inflight, nr);
6914 refcount_add(nr, ¤t->usage);
6915 io_submit_state_start(&ctx->submit_state, nr);
6917 while (submitted < nr) {
6918 const struct io_uring_sqe *sqe;
6919 struct io_kiocb *req;
6921 req = io_alloc_req(ctx);
6922 if (unlikely(!req)) {
6924 submitted = -EAGAIN;
6927 sqe = io_get_sqe(ctx);
6928 if (unlikely(!sqe)) {
6929 kmem_cache_free(req_cachep, req);
6932 /* will complete beyond this point, count as submitted */
6934 if (io_submit_sqe(ctx, req, sqe))
6938 if (unlikely(submitted != nr)) {
6939 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6940 struct io_uring_task *tctx = current->io_uring;
6941 int unused = nr - ref_used;
6943 percpu_ref_put_many(&ctx->refs, unused);
6944 percpu_counter_sub(&tctx->inflight, unused);
6945 put_task_struct_many(current, unused);
6948 io_submit_state_end(&ctx->submit_state, ctx);
6949 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6950 io_commit_sqring(ctx);
6955 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6957 /* Tell userspace we may need a wakeup call */
6958 spin_lock_irq(&ctx->completion_lock);
6959 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6960 spin_unlock_irq(&ctx->completion_lock);
6963 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6965 spin_lock_irq(&ctx->completion_lock);
6966 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6967 spin_unlock_irq(&ctx->completion_lock);
6970 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
6972 unsigned int to_submit;
6975 to_submit = io_sqring_entries(ctx);
6976 /* if we're handling multiple rings, cap submit size for fairness */
6977 if (cap_entries && to_submit > 8)
6980 if (!list_empty(&ctx->iopoll_list) || to_submit) {
6981 unsigned nr_events = 0;
6983 mutex_lock(&ctx->uring_lock);
6984 if (!list_empty(&ctx->iopoll_list))
6985 io_do_iopoll(ctx, &nr_events, 0);
6987 if (to_submit && !ctx->sqo_dead &&
6988 likely(!percpu_ref_is_dying(&ctx->refs)))
6989 ret = io_submit_sqes(ctx, to_submit);
6990 mutex_unlock(&ctx->uring_lock);
6993 if (!io_sqring_full(ctx) && wq_has_sleeper(&ctx->sqo_sq_wait))
6994 wake_up(&ctx->sqo_sq_wait);
6999 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
7001 struct io_ring_ctx *ctx;
7002 unsigned sq_thread_idle = 0;
7004 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
7005 if (sq_thread_idle < ctx->sq_thread_idle)
7006 sq_thread_idle = ctx->sq_thread_idle;
7009 sqd->sq_thread_idle = sq_thread_idle;
7012 static void io_sqd_init_new(struct io_sq_data *sqd)
7014 struct io_ring_ctx *ctx;
7016 while (!list_empty(&sqd->ctx_new_list)) {
7017 ctx = list_first_entry(&sqd->ctx_new_list, struct io_ring_ctx, sqd_list);
7018 list_move_tail(&ctx->sqd_list, &sqd->ctx_list);
7019 complete(&ctx->sq_thread_comp);
7022 io_sqd_update_thread_idle(sqd);
7025 static int io_sq_thread(void *data)
7027 struct cgroup_subsys_state *cur_css = NULL;
7028 struct files_struct *old_files = current->files;
7029 struct nsproxy *old_nsproxy = current->nsproxy;
7030 const struct cred *old_cred = NULL;
7031 struct io_sq_data *sqd = data;
7032 struct io_ring_ctx *ctx;
7033 unsigned long timeout = 0;
7037 current->files = NULL;
7038 current->nsproxy = NULL;
7039 task_unlock(current);
7041 while (!kthread_should_stop()) {
7043 bool cap_entries, sqt_spin, needs_sched;
7046 * Any changes to the sqd lists are synchronized through the
7047 * kthread parking. This synchronizes the thread vs users,
7048 * the users are synchronized on the sqd->ctx_lock.
7050 if (kthread_should_park()) {
7053 * When sq thread is unparked, in case the previous park operation
7054 * comes from io_put_sq_data(), which means that sq thread is going
7055 * to be stopped, so here needs to have a check.
7057 if (kthread_should_stop())
7061 if (unlikely(!list_empty(&sqd->ctx_new_list))) {
7062 io_sqd_init_new(sqd);
7063 timeout = jiffies + sqd->sq_thread_idle;
7067 cap_entries = !list_is_singular(&sqd->ctx_list);
7068 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
7069 if (current->cred != ctx->creds) {
7071 revert_creds(old_cred);
7072 old_cred = override_creds(ctx->creds);
7074 io_sq_thread_associate_blkcg(ctx, &cur_css);
7076 current->loginuid = ctx->loginuid;
7077 current->sessionid = ctx->sessionid;
7080 ret = __io_sq_thread(ctx, cap_entries);
7081 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
7084 io_sq_thread_drop_mm_files();
7087 if (sqt_spin || !time_after(jiffies, timeout)) {
7089 io_sq_thread_drop_mm_files();
7092 timeout = jiffies + sqd->sq_thread_idle;
7097 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
7098 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
7099 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
7100 !list_empty_careful(&ctx->iopoll_list)) {
7101 needs_sched = false;
7104 if (io_sqring_entries(ctx)) {
7105 needs_sched = false;
7110 if (needs_sched && !kthread_should_park()) {
7111 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7112 io_ring_set_wakeup_flag(ctx);
7115 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7116 io_ring_clear_wakeup_flag(ctx);
7119 finish_wait(&sqd->wait, &wait);
7120 timeout = jiffies + sqd->sq_thread_idle;
7124 io_sq_thread_drop_mm_files();
7127 io_sq_thread_unassociate_blkcg();
7129 revert_creds(old_cred);
7132 current->files = old_files;
7133 current->nsproxy = old_nsproxy;
7134 task_unlock(current);
7141 struct io_wait_queue {
7142 struct wait_queue_entry wq;
7143 struct io_ring_ctx *ctx;
7145 unsigned nr_timeouts;
7148 static inline bool io_should_wake(struct io_wait_queue *iowq)
7150 struct io_ring_ctx *ctx = iowq->ctx;
7153 * Wake up if we have enough events, or if a timeout occurred since we
7154 * started waiting. For timeouts, we always want to return to userspace,
7155 * regardless of event count.
7157 return io_cqring_events(ctx) >= iowq->to_wait ||
7158 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
7161 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
7162 int wake_flags, void *key)
7164 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
7168 * Cannot safely flush overflowed CQEs from here, ensure we wake up
7169 * the task, and the next invocation will do it.
7171 if (io_should_wake(iowq) || test_bit(0, &iowq->ctx->cq_check_overflow))
7172 return autoremove_wake_function(curr, mode, wake_flags, key);
7176 static int io_run_task_work_sig(void)
7178 if (io_run_task_work())
7180 if (!signal_pending(current))
7182 if (test_tsk_thread_flag(current, TIF_NOTIFY_SIGNAL))
7183 return -ERESTARTSYS;
7187 /* when returns >0, the caller should retry */
7188 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
7189 struct io_wait_queue *iowq,
7190 signed long *timeout)
7194 /* make sure we run task_work before checking for signals */
7195 ret = io_run_task_work_sig();
7196 if (ret || io_should_wake(iowq))
7198 /* let the caller flush overflows, retry */
7199 if (test_bit(0, &ctx->cq_check_overflow))
7202 *timeout = schedule_timeout(*timeout);
7203 return !*timeout ? -ETIME : 1;
7207 * Wait until events become available, if we don't already have some. The
7208 * application must reap them itself, as they reside on the shared cq ring.
7210 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
7211 const sigset_t __user *sig, size_t sigsz,
7212 struct __kernel_timespec __user *uts)
7214 struct io_wait_queue iowq = {
7217 .func = io_wake_function,
7218 .entry = LIST_HEAD_INIT(iowq.wq.entry),
7221 .to_wait = min_events,
7223 struct io_rings *rings = ctx->rings;
7224 signed long timeout = MAX_SCHEDULE_TIMEOUT;
7228 io_cqring_overflow_flush(ctx, false, NULL, NULL);
7229 if (io_cqring_events(ctx) >= min_events)
7231 if (!io_run_task_work())
7236 #ifdef CONFIG_COMPAT
7237 if (in_compat_syscall())
7238 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
7242 ret = set_user_sigmask(sig, sigsz);
7249 struct timespec64 ts;
7251 if (get_timespec64(&ts, uts))
7253 timeout = timespec64_to_jiffies(&ts);
7256 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
7257 trace_io_uring_cqring_wait(ctx, min_events);
7259 io_cqring_overflow_flush(ctx, false, NULL, NULL);
7260 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
7261 TASK_INTERRUPTIBLE);
7262 ret = io_cqring_wait_schedule(ctx, &iowq, &timeout);
7263 finish_wait(&ctx->wait, &iowq.wq);
7266 restore_saved_sigmask_unless(ret == -EINTR);
7268 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
7271 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
7273 #if defined(CONFIG_UNIX)
7274 if (ctx->ring_sock) {
7275 struct sock *sock = ctx->ring_sock->sk;
7276 struct sk_buff *skb;
7278 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
7284 for (i = 0; i < ctx->nr_user_files; i++) {
7287 file = io_file_from_index(ctx, i);
7294 static void io_rsrc_data_ref_zero(struct percpu_ref *ref)
7296 struct fixed_rsrc_data *data;
7298 data = container_of(ref, struct fixed_rsrc_data, refs);
7299 complete(&data->done);
7302 static inline void io_rsrc_ref_lock(struct io_ring_ctx *ctx)
7304 spin_lock_bh(&ctx->rsrc_ref_lock);
7307 static inline void io_rsrc_ref_unlock(struct io_ring_ctx *ctx)
7309 spin_unlock_bh(&ctx->rsrc_ref_lock);
7312 static void io_sqe_rsrc_set_node(struct io_ring_ctx *ctx,
7313 struct fixed_rsrc_data *rsrc_data,
7314 struct fixed_rsrc_ref_node *ref_node)
7316 io_rsrc_ref_lock(ctx);
7317 rsrc_data->node = ref_node;
7318 list_add_tail(&ref_node->node, &ctx->rsrc_ref_list);
7319 io_rsrc_ref_unlock(ctx);
7320 percpu_ref_get(&rsrc_data->refs);
7323 static int io_rsrc_ref_quiesce(struct fixed_rsrc_data *data,
7324 struct io_ring_ctx *ctx,
7325 struct fixed_rsrc_ref_node *backup_node)
7327 struct fixed_rsrc_ref_node *ref_node;
7330 io_rsrc_ref_lock(ctx);
7331 ref_node = data->node;
7332 io_rsrc_ref_unlock(ctx);
7334 percpu_ref_kill(&ref_node->refs);
7336 percpu_ref_kill(&data->refs);
7338 /* wait for all refs nodes to complete */
7339 flush_delayed_work(&ctx->rsrc_put_work);
7341 ret = wait_for_completion_interruptible(&data->done);
7344 ret = io_run_task_work_sig();
7346 percpu_ref_resurrect(&data->refs);
7347 reinit_completion(&data->done);
7348 io_sqe_rsrc_set_node(ctx, data, backup_node);
7353 destroy_fixed_rsrc_ref_node(backup_node);
7357 static struct fixed_rsrc_data *alloc_fixed_rsrc_data(struct io_ring_ctx *ctx)
7359 struct fixed_rsrc_data *data;
7361 data = kzalloc(sizeof(*data), GFP_KERNEL);
7365 if (percpu_ref_init(&data->refs, io_rsrc_data_ref_zero,
7366 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) {
7371 init_completion(&data->done);
7375 static void free_fixed_rsrc_data(struct fixed_rsrc_data *data)
7377 percpu_ref_exit(&data->refs);
7382 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7384 struct fixed_rsrc_data *data = ctx->file_data;
7385 struct fixed_rsrc_ref_node *backup_node;
7386 unsigned nr_tables, i;
7391 backup_node = alloc_fixed_rsrc_ref_node(ctx);
7394 init_fixed_file_ref_node(ctx, backup_node);
7396 ret = io_rsrc_ref_quiesce(data, ctx, backup_node);
7400 __io_sqe_files_unregister(ctx);
7401 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
7402 for (i = 0; i < nr_tables; i++)
7403 kfree(data->table[i].files);
7404 free_fixed_rsrc_data(data);
7405 ctx->file_data = NULL;
7406 ctx->nr_user_files = 0;
7410 static void io_put_sq_data(struct io_sq_data *sqd)
7412 if (refcount_dec_and_test(&sqd->refs)) {
7414 * The park is a bit of a work-around, without it we get
7415 * warning spews on shutdown with SQPOLL set and affinity
7416 * set to a single CPU.
7419 kthread_park(sqd->thread);
7420 kthread_stop(sqd->thread);
7427 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7429 struct io_ring_ctx *ctx_attach;
7430 struct io_sq_data *sqd;
7433 f = fdget(p->wq_fd);
7435 return ERR_PTR(-ENXIO);
7436 if (f.file->f_op != &io_uring_fops) {
7438 return ERR_PTR(-EINVAL);
7441 ctx_attach = f.file->private_data;
7442 sqd = ctx_attach->sq_data;
7445 return ERR_PTR(-EINVAL);
7448 refcount_inc(&sqd->refs);
7453 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p)
7455 struct io_sq_data *sqd;
7457 if (p->flags & IORING_SETUP_ATTACH_WQ)
7458 return io_attach_sq_data(p);
7460 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7462 return ERR_PTR(-ENOMEM);
7464 refcount_set(&sqd->refs, 1);
7465 INIT_LIST_HEAD(&sqd->ctx_list);
7466 INIT_LIST_HEAD(&sqd->ctx_new_list);
7467 mutex_init(&sqd->ctx_lock);
7468 mutex_init(&sqd->lock);
7469 init_waitqueue_head(&sqd->wait);
7473 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7474 __releases(&sqd->lock)
7478 kthread_unpark(sqd->thread);
7479 mutex_unlock(&sqd->lock);
7482 static void io_sq_thread_park(struct io_sq_data *sqd)
7483 __acquires(&sqd->lock)
7487 mutex_lock(&sqd->lock);
7488 kthread_park(sqd->thread);
7491 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
7493 struct io_sq_data *sqd = ctx->sq_data;
7498 * We may arrive here from the error branch in
7499 * io_sq_offload_create() where the kthread is created
7500 * without being waked up, thus wake it up now to make
7501 * sure the wait will complete.
7503 wake_up_process(sqd->thread);
7504 wait_for_completion(&ctx->sq_thread_comp);
7506 io_sq_thread_park(sqd);
7509 mutex_lock(&sqd->ctx_lock);
7510 list_del(&ctx->sqd_list);
7511 io_sqd_update_thread_idle(sqd);
7512 mutex_unlock(&sqd->ctx_lock);
7515 io_sq_thread_unpark(sqd);
7517 io_put_sq_data(sqd);
7518 ctx->sq_data = NULL;
7522 static void io_finish_async(struct io_ring_ctx *ctx)
7524 io_sq_thread_stop(ctx);
7527 io_wq_destroy(ctx->io_wq);
7532 #if defined(CONFIG_UNIX)
7534 * Ensure the UNIX gc is aware of our file set, so we are certain that
7535 * the io_uring can be safely unregistered on process exit, even if we have
7536 * loops in the file referencing.
7538 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7540 struct sock *sk = ctx->ring_sock->sk;
7541 struct scm_fp_list *fpl;
7542 struct sk_buff *skb;
7545 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7549 skb = alloc_skb(0, GFP_KERNEL);
7558 fpl->user = get_uid(ctx->user);
7559 for (i = 0; i < nr; i++) {
7560 struct file *file = io_file_from_index(ctx, i + offset);
7564 fpl->fp[nr_files] = get_file(file);
7565 unix_inflight(fpl->user, fpl->fp[nr_files]);
7570 fpl->max = SCM_MAX_FD;
7571 fpl->count = nr_files;
7572 UNIXCB(skb).fp = fpl;
7573 skb->destructor = unix_destruct_scm;
7574 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7575 skb_queue_head(&sk->sk_receive_queue, skb);
7577 for (i = 0; i < nr_files; i++)
7588 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7589 * causes regular reference counting to break down. We rely on the UNIX
7590 * garbage collection to take care of this problem for us.
7592 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7594 unsigned left, total;
7598 left = ctx->nr_user_files;
7600 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7602 ret = __io_sqe_files_scm(ctx, this_files, total);
7606 total += this_files;
7612 while (total < ctx->nr_user_files) {
7613 struct file *file = io_file_from_index(ctx, total);
7623 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7629 static int io_sqe_alloc_file_tables(struct fixed_rsrc_data *file_data,
7630 unsigned nr_tables, unsigned nr_files)
7634 for (i = 0; i < nr_tables; i++) {
7635 struct fixed_rsrc_table *table = &file_data->table[i];
7636 unsigned this_files;
7638 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
7639 table->files = kcalloc(this_files, sizeof(struct file *),
7643 nr_files -= this_files;
7649 for (i = 0; i < nr_tables; i++) {
7650 struct fixed_rsrc_table *table = &file_data->table[i];
7651 kfree(table->files);
7656 static void io_ring_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc)
7658 struct file *file = prsrc->file;
7659 #if defined(CONFIG_UNIX)
7660 struct sock *sock = ctx->ring_sock->sk;
7661 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7662 struct sk_buff *skb;
7665 __skb_queue_head_init(&list);
7668 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7669 * remove this entry and rearrange the file array.
7671 skb = skb_dequeue(head);
7673 struct scm_fp_list *fp;
7675 fp = UNIXCB(skb).fp;
7676 for (i = 0; i < fp->count; i++) {
7679 if (fp->fp[i] != file)
7682 unix_notinflight(fp->user, fp->fp[i]);
7683 left = fp->count - 1 - i;
7685 memmove(&fp->fp[i], &fp->fp[i + 1],
7686 left * sizeof(struct file *));
7693 __skb_queue_tail(&list, skb);
7703 __skb_queue_tail(&list, skb);
7705 skb = skb_dequeue(head);
7708 if (skb_peek(&list)) {
7709 spin_lock_irq(&head->lock);
7710 while ((skb = __skb_dequeue(&list)) != NULL)
7711 __skb_queue_tail(head, skb);
7712 spin_unlock_irq(&head->lock);
7719 static void __io_rsrc_put_work(struct fixed_rsrc_ref_node *ref_node)
7721 struct fixed_rsrc_data *rsrc_data = ref_node->rsrc_data;
7722 struct io_ring_ctx *ctx = rsrc_data->ctx;
7723 struct io_rsrc_put *prsrc, *tmp;
7725 list_for_each_entry_safe(prsrc, tmp, &ref_node->rsrc_list, list) {
7726 list_del(&prsrc->list);
7727 ref_node->rsrc_put(ctx, prsrc);
7731 percpu_ref_exit(&ref_node->refs);
7733 percpu_ref_put(&rsrc_data->refs);
7736 static void io_rsrc_put_work(struct work_struct *work)
7738 struct io_ring_ctx *ctx;
7739 struct llist_node *node;
7741 ctx = container_of(work, struct io_ring_ctx, rsrc_put_work.work);
7742 node = llist_del_all(&ctx->rsrc_put_llist);
7745 struct fixed_rsrc_ref_node *ref_node;
7746 struct llist_node *next = node->next;
7748 ref_node = llist_entry(node, struct fixed_rsrc_ref_node, llist);
7749 __io_rsrc_put_work(ref_node);
7754 static struct file **io_fixed_file_slot(struct fixed_rsrc_data *file_data,
7757 struct fixed_rsrc_table *table;
7759 table = &file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7760 return &table->files[i & IORING_FILE_TABLE_MASK];
7763 static void io_rsrc_node_ref_zero(struct percpu_ref *ref)
7765 struct fixed_rsrc_ref_node *ref_node;
7766 struct fixed_rsrc_data *data;
7767 struct io_ring_ctx *ctx;
7768 bool first_add = false;
7771 ref_node = container_of(ref, struct fixed_rsrc_ref_node, refs);
7772 data = ref_node->rsrc_data;
7775 io_rsrc_ref_lock(ctx);
7776 ref_node->done = true;
7778 while (!list_empty(&ctx->rsrc_ref_list)) {
7779 ref_node = list_first_entry(&ctx->rsrc_ref_list,
7780 struct fixed_rsrc_ref_node, node);
7781 /* recycle ref nodes in order */
7782 if (!ref_node->done)
7784 list_del(&ref_node->node);
7785 first_add |= llist_add(&ref_node->llist, &ctx->rsrc_put_llist);
7787 io_rsrc_ref_unlock(ctx);
7789 if (percpu_ref_is_dying(&data->refs))
7793 mod_delayed_work(system_wq, &ctx->rsrc_put_work, 0);
7795 queue_delayed_work(system_wq, &ctx->rsrc_put_work, delay);
7798 static struct fixed_rsrc_ref_node *alloc_fixed_rsrc_ref_node(
7799 struct io_ring_ctx *ctx)
7801 struct fixed_rsrc_ref_node *ref_node;
7803 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7807 if (percpu_ref_init(&ref_node->refs, io_rsrc_node_ref_zero,
7812 INIT_LIST_HEAD(&ref_node->node);
7813 INIT_LIST_HEAD(&ref_node->rsrc_list);
7814 ref_node->done = false;
7818 static void init_fixed_file_ref_node(struct io_ring_ctx *ctx,
7819 struct fixed_rsrc_ref_node *ref_node)
7821 ref_node->rsrc_data = ctx->file_data;
7822 ref_node->rsrc_put = io_ring_file_put;
7825 static void destroy_fixed_rsrc_ref_node(struct fixed_rsrc_ref_node *ref_node)
7827 percpu_ref_exit(&ref_node->refs);
7832 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7835 __s32 __user *fds = (__s32 __user *) arg;
7836 unsigned nr_tables, i;
7838 int fd, ret = -ENOMEM;
7839 struct fixed_rsrc_ref_node *ref_node;
7840 struct fixed_rsrc_data *file_data;
7846 if (nr_args > IORING_MAX_FIXED_FILES)
7849 file_data = alloc_fixed_rsrc_data(ctx);
7852 ctx->file_data = file_data;
7854 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
7855 file_data->table = kcalloc(nr_tables, sizeof(*file_data->table),
7857 if (!file_data->table)
7860 if (io_sqe_alloc_file_tables(file_data, nr_tables, nr_args))
7863 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7864 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7868 /* allow sparse sets */
7878 * Don't allow io_uring instances to be registered. If UNIX
7879 * isn't enabled, then this causes a reference cycle and this
7880 * instance can never get freed. If UNIX is enabled we'll
7881 * handle it just fine, but there's still no point in allowing
7882 * a ring fd as it doesn't support regular read/write anyway.
7884 if (file->f_op == &io_uring_fops) {
7888 *io_fixed_file_slot(file_data, i) = file;
7891 ret = io_sqe_files_scm(ctx);
7893 io_sqe_files_unregister(ctx);
7897 ref_node = alloc_fixed_rsrc_ref_node(ctx);
7899 io_sqe_files_unregister(ctx);
7902 init_fixed_file_ref_node(ctx, ref_node);
7904 io_sqe_rsrc_set_node(ctx, file_data, ref_node);
7907 for (i = 0; i < ctx->nr_user_files; i++) {
7908 file = io_file_from_index(ctx, i);
7912 for (i = 0; i < nr_tables; i++)
7913 kfree(file_data->table[i].files);
7914 ctx->nr_user_files = 0;
7916 free_fixed_rsrc_data(ctx->file_data);
7917 ctx->file_data = NULL;
7921 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7924 #if defined(CONFIG_UNIX)
7925 struct sock *sock = ctx->ring_sock->sk;
7926 struct sk_buff_head *head = &sock->sk_receive_queue;
7927 struct sk_buff *skb;
7930 * See if we can merge this file into an existing skb SCM_RIGHTS
7931 * file set. If there's no room, fall back to allocating a new skb
7932 * and filling it in.
7934 spin_lock_irq(&head->lock);
7935 skb = skb_peek(head);
7937 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7939 if (fpl->count < SCM_MAX_FD) {
7940 __skb_unlink(skb, head);
7941 spin_unlock_irq(&head->lock);
7942 fpl->fp[fpl->count] = get_file(file);
7943 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7945 spin_lock_irq(&head->lock);
7946 __skb_queue_head(head, skb);
7951 spin_unlock_irq(&head->lock);
7958 return __io_sqe_files_scm(ctx, 1, index);
7964 static int io_queue_rsrc_removal(struct fixed_rsrc_data *data, void *rsrc)
7966 struct io_rsrc_put *prsrc;
7967 struct fixed_rsrc_ref_node *ref_node = data->node;
7969 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL);
7974 list_add(&prsrc->list, &ref_node->rsrc_list);
7979 static inline int io_queue_file_removal(struct fixed_rsrc_data *data,
7982 return io_queue_rsrc_removal(data, (void *)file);
7985 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7986 struct io_uring_rsrc_update *up,
7989 struct fixed_rsrc_data *data = ctx->file_data;
7990 struct fixed_rsrc_ref_node *ref_node;
7991 struct file *file, **file_slot;
7995 bool needs_switch = false;
7997 if (check_add_overflow(up->offset, nr_args, &done))
7999 if (done > ctx->nr_user_files)
8002 ref_node = alloc_fixed_rsrc_ref_node(ctx);
8005 init_fixed_file_ref_node(ctx, ref_node);
8007 fds = u64_to_user_ptr(up->data);
8008 for (done = 0; done < nr_args; done++) {
8010 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
8014 if (fd == IORING_REGISTER_FILES_SKIP)
8017 i = array_index_nospec(up->offset + done, ctx->nr_user_files);
8018 file_slot = io_fixed_file_slot(ctx->file_data, i);
8021 err = io_queue_file_removal(data, *file_slot);
8025 needs_switch = true;
8034 * Don't allow io_uring instances to be registered. If
8035 * UNIX isn't enabled, then this causes a reference
8036 * cycle and this instance can never get freed. If UNIX
8037 * is enabled we'll handle it just fine, but there's
8038 * still no point in allowing a ring fd as it doesn't
8039 * support regular read/write anyway.
8041 if (file->f_op == &io_uring_fops) {
8047 err = io_sqe_file_register(ctx, file, i);
8057 percpu_ref_kill(&data->node->refs);
8058 io_sqe_rsrc_set_node(ctx, data, ref_node);
8060 destroy_fixed_rsrc_ref_node(ref_node);
8062 return done ? done : err;
8065 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
8068 struct io_uring_rsrc_update up;
8070 if (!ctx->file_data)
8074 if (copy_from_user(&up, arg, sizeof(up)))
8079 return __io_sqe_files_update(ctx, &up, nr_args);
8082 static struct io_wq_work *io_free_work(struct io_wq_work *work)
8084 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8086 req = io_put_req_find_next(req);
8087 return req ? &req->work : NULL;
8090 static int io_init_wq_offload(struct io_ring_ctx *ctx,
8091 struct io_uring_params *p)
8093 struct io_wq_data data;
8095 struct io_ring_ctx *ctx_attach;
8096 unsigned int concurrency;
8099 data.user = ctx->user;
8100 data.free_work = io_free_work;
8101 data.do_work = io_wq_submit_work;
8103 if (!(p->flags & IORING_SETUP_ATTACH_WQ)) {
8104 /* Do QD, or 4 * CPUS, whatever is smallest */
8105 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
8107 ctx->io_wq = io_wq_create(concurrency, &data);
8108 if (IS_ERR(ctx->io_wq)) {
8109 ret = PTR_ERR(ctx->io_wq);
8115 f = fdget(p->wq_fd);
8119 if (f.file->f_op != &io_uring_fops) {
8124 ctx_attach = f.file->private_data;
8125 /* @io_wq is protected by holding the fd */
8126 if (!io_wq_get(ctx_attach->io_wq, &data)) {
8131 ctx->io_wq = ctx_attach->io_wq;
8137 static int io_uring_alloc_task_context(struct task_struct *task)
8139 struct io_uring_task *tctx;
8142 tctx = kmalloc(sizeof(*tctx), GFP_KERNEL);
8143 if (unlikely(!tctx))
8146 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
8147 if (unlikely(ret)) {
8153 init_waitqueue_head(&tctx->wait);
8155 atomic_set(&tctx->in_idle, 0);
8156 tctx->sqpoll = false;
8157 io_init_identity(&tctx->__identity);
8158 tctx->identity = &tctx->__identity;
8159 task->io_uring = tctx;
8160 spin_lock_init(&tctx->task_lock);
8161 INIT_WQ_LIST(&tctx->task_list);
8162 tctx->task_state = 0;
8163 init_task_work(&tctx->task_work, tctx_task_work);
8167 void __io_uring_free(struct task_struct *tsk)
8169 struct io_uring_task *tctx = tsk->io_uring;
8171 WARN_ON_ONCE(!xa_empty(&tctx->xa));
8172 WARN_ON_ONCE(refcount_read(&tctx->identity->count) != 1);
8173 if (tctx->identity != &tctx->__identity)
8174 kfree(tctx->identity);
8175 percpu_counter_destroy(&tctx->inflight);
8177 tsk->io_uring = NULL;
8180 static int io_sq_offload_create(struct io_ring_ctx *ctx,
8181 struct io_uring_params *p)
8185 if (ctx->flags & IORING_SETUP_SQPOLL) {
8186 struct io_sq_data *sqd;
8189 if (!capable(CAP_SYS_ADMIN) && !capable(CAP_SYS_NICE))
8192 sqd = io_get_sq_data(p);
8199 io_sq_thread_park(sqd);
8200 mutex_lock(&sqd->ctx_lock);
8201 list_add(&ctx->sqd_list, &sqd->ctx_new_list);
8202 mutex_unlock(&sqd->ctx_lock);
8203 io_sq_thread_unpark(sqd);
8205 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
8206 if (!ctx->sq_thread_idle)
8207 ctx->sq_thread_idle = HZ;
8212 if (p->flags & IORING_SETUP_SQ_AFF) {
8213 int cpu = p->sq_thread_cpu;
8216 if (cpu >= nr_cpu_ids)
8218 if (!cpu_online(cpu))
8221 sqd->thread = kthread_create_on_cpu(io_sq_thread, sqd,
8222 cpu, "io_uring-sq");
8224 sqd->thread = kthread_create(io_sq_thread, sqd,
8227 if (IS_ERR(sqd->thread)) {
8228 ret = PTR_ERR(sqd->thread);
8232 ret = io_uring_alloc_task_context(sqd->thread);
8235 } else if (p->flags & IORING_SETUP_SQ_AFF) {
8236 /* Can't have SQ_AFF without SQPOLL */
8242 ret = io_init_wq_offload(ctx, p);
8248 io_finish_async(ctx);
8252 static void io_sq_offload_start(struct io_ring_ctx *ctx)
8254 struct io_sq_data *sqd = ctx->sq_data;
8256 if ((ctx->flags & IORING_SETUP_SQPOLL) && sqd->thread)
8257 wake_up_process(sqd->thread);
8260 static inline void __io_unaccount_mem(struct user_struct *user,
8261 unsigned long nr_pages)
8263 atomic_long_sub(nr_pages, &user->locked_vm);
8266 static inline int __io_account_mem(struct user_struct *user,
8267 unsigned long nr_pages)
8269 unsigned long page_limit, cur_pages, new_pages;
8271 /* Don't allow more pages than we can safely lock */
8272 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
8275 cur_pages = atomic_long_read(&user->locked_vm);
8276 new_pages = cur_pages + nr_pages;
8277 if (new_pages > page_limit)
8279 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
8280 new_pages) != cur_pages);
8285 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8288 __io_unaccount_mem(ctx->user, nr_pages);
8290 if (ctx->mm_account)
8291 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
8294 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages)
8298 if (ctx->limit_mem) {
8299 ret = __io_account_mem(ctx->user, nr_pages);
8304 if (ctx->mm_account)
8305 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
8310 static void io_mem_free(void *ptr)
8317 page = virt_to_head_page(ptr);
8318 if (put_page_testzero(page))
8319 free_compound_page(page);
8322 static void *io_mem_alloc(size_t size)
8324 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
8325 __GFP_NORETRY | __GFP_ACCOUNT;
8327 return (void *) __get_free_pages(gfp_flags, get_order(size));
8330 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8333 struct io_rings *rings;
8334 size_t off, sq_array_size;
8336 off = struct_size(rings, cqes, cq_entries);
8337 if (off == SIZE_MAX)
8341 off = ALIGN(off, SMP_CACHE_BYTES);
8349 sq_array_size = array_size(sizeof(u32), sq_entries);
8350 if (sq_array_size == SIZE_MAX)
8353 if (check_add_overflow(off, sq_array_size, &off))
8359 static int io_sqe_buffers_unregister(struct io_ring_ctx *ctx)
8363 if (!ctx->user_bufs)
8366 for (i = 0; i < ctx->nr_user_bufs; i++) {
8367 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8369 for (j = 0; j < imu->nr_bvecs; j++)
8370 unpin_user_page(imu->bvec[j].bv_page);
8372 if (imu->acct_pages)
8373 io_unaccount_mem(ctx, imu->acct_pages);
8378 kfree(ctx->user_bufs);
8379 ctx->user_bufs = NULL;
8380 ctx->nr_user_bufs = 0;
8384 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8385 void __user *arg, unsigned index)
8387 struct iovec __user *src;
8389 #ifdef CONFIG_COMPAT
8391 struct compat_iovec __user *ciovs;
8392 struct compat_iovec ciov;
8394 ciovs = (struct compat_iovec __user *) arg;
8395 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8398 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8399 dst->iov_len = ciov.iov_len;
8403 src = (struct iovec __user *) arg;
8404 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8410 * Not super efficient, but this is just a registration time. And we do cache
8411 * the last compound head, so generally we'll only do a full search if we don't
8414 * We check if the given compound head page has already been accounted, to
8415 * avoid double accounting it. This allows us to account the full size of the
8416 * page, not just the constituent pages of a huge page.
8418 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8419 int nr_pages, struct page *hpage)
8423 /* check current page array */
8424 for (i = 0; i < nr_pages; i++) {
8425 if (!PageCompound(pages[i]))
8427 if (compound_head(pages[i]) == hpage)
8431 /* check previously registered pages */
8432 for (i = 0; i < ctx->nr_user_bufs; i++) {
8433 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8435 for (j = 0; j < imu->nr_bvecs; j++) {
8436 if (!PageCompound(imu->bvec[j].bv_page))
8438 if (compound_head(imu->bvec[j].bv_page) == hpage)
8446 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8447 int nr_pages, struct io_mapped_ubuf *imu,
8448 struct page **last_hpage)
8452 for (i = 0; i < nr_pages; i++) {
8453 if (!PageCompound(pages[i])) {
8458 hpage = compound_head(pages[i]);
8459 if (hpage == *last_hpage)
8461 *last_hpage = hpage;
8462 if (headpage_already_acct(ctx, pages, i, hpage))
8464 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8468 if (!imu->acct_pages)
8471 ret = io_account_mem(ctx, imu->acct_pages);
8473 imu->acct_pages = 0;
8477 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov,
8478 struct io_mapped_ubuf *imu,
8479 struct page **last_hpage)
8481 struct vm_area_struct **vmas = NULL;
8482 struct page **pages = NULL;
8483 unsigned long off, start, end, ubuf;
8485 int ret, pret, nr_pages, i;
8487 ubuf = (unsigned long) iov->iov_base;
8488 end = (ubuf + iov->iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8489 start = ubuf >> PAGE_SHIFT;
8490 nr_pages = end - start;
8494 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
8498 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *),
8503 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
8509 mmap_read_lock(current->mm);
8510 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
8512 if (pret == nr_pages) {
8513 /* don't support file backed memory */
8514 for (i = 0; i < nr_pages; i++) {
8515 struct vm_area_struct *vma = vmas[i];
8518 !is_file_hugepages(vma->vm_file)) {
8524 ret = pret < 0 ? pret : -EFAULT;
8526 mmap_read_unlock(current->mm);
8529 * if we did partial map, or found file backed vmas,
8530 * release any pages we did get
8533 unpin_user_pages(pages, pret);
8538 ret = io_buffer_account_pin(ctx, pages, pret, imu, last_hpage);
8540 unpin_user_pages(pages, pret);
8545 off = ubuf & ~PAGE_MASK;
8546 size = iov->iov_len;
8547 for (i = 0; i < nr_pages; i++) {
8550 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8551 imu->bvec[i].bv_page = pages[i];
8552 imu->bvec[i].bv_len = vec_len;
8553 imu->bvec[i].bv_offset = off;
8557 /* store original address for later verification */
8559 imu->len = iov->iov_len;
8560 imu->nr_bvecs = nr_pages;
8568 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args)
8572 if (!nr_args || nr_args > UIO_MAXIOV)
8575 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
8577 if (!ctx->user_bufs)
8583 static int io_buffer_validate(struct iovec *iov)
8586 * Don't impose further limits on the size and buffer
8587 * constraints here, we'll -EINVAL later when IO is
8588 * submitted if they are wrong.
8590 if (!iov->iov_base || !iov->iov_len)
8593 /* arbitrary limit, but we need something */
8594 if (iov->iov_len > SZ_1G)
8600 static int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg,
8601 unsigned int nr_args)
8605 struct page *last_hpage = NULL;
8607 ret = io_buffers_map_alloc(ctx, nr_args);
8611 for (i = 0; i < nr_args; i++) {
8612 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8614 ret = io_copy_iov(ctx, &iov, arg, i);
8618 ret = io_buffer_validate(&iov);
8622 ret = io_sqe_buffer_register(ctx, &iov, imu, &last_hpage);
8626 ctx->nr_user_bufs++;
8630 io_sqe_buffers_unregister(ctx);
8635 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8637 __s32 __user *fds = arg;
8643 if (copy_from_user(&fd, fds, sizeof(*fds)))
8646 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8647 if (IS_ERR(ctx->cq_ev_fd)) {
8648 int ret = PTR_ERR(ctx->cq_ev_fd);
8649 ctx->cq_ev_fd = NULL;
8656 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8658 if (ctx->cq_ev_fd) {
8659 eventfd_ctx_put(ctx->cq_ev_fd);
8660 ctx->cq_ev_fd = NULL;
8667 static int __io_destroy_buffers(int id, void *p, void *data)
8669 struct io_ring_ctx *ctx = data;
8670 struct io_buffer *buf = p;
8672 __io_remove_buffers(ctx, buf, id, -1U);
8676 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8678 idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx);
8679 idr_destroy(&ctx->io_buffer_idr);
8682 static void io_req_cache_free(struct list_head *list, struct task_struct *tsk)
8684 struct io_kiocb *req, *nxt;
8686 list_for_each_entry_safe(req, nxt, list, compl.list) {
8687 if (tsk && req->task != tsk)
8689 list_del(&req->compl.list);
8690 kmem_cache_free(req_cachep, req);
8694 static void io_req_caches_free(struct io_ring_ctx *ctx, struct task_struct *tsk)
8696 struct io_submit_state *submit_state = &ctx->submit_state;
8698 mutex_lock(&ctx->uring_lock);
8700 if (submit_state->free_reqs)
8701 kmem_cache_free_bulk(req_cachep, submit_state->free_reqs,
8702 submit_state->reqs);
8704 io_req_cache_free(&submit_state->comp.free_list, NULL);
8706 spin_lock_irq(&ctx->completion_lock);
8707 io_req_cache_free(&submit_state->comp.locked_free_list, NULL);
8708 spin_unlock_irq(&ctx->completion_lock);
8710 mutex_unlock(&ctx->uring_lock);
8713 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8716 * Some may use context even when all refs and requests have been put,
8717 * and they are free to do so while still holding uring_lock, see
8718 * __io_req_task_submit(). Wait for them to finish.
8720 mutex_lock(&ctx->uring_lock);
8721 mutex_unlock(&ctx->uring_lock);
8723 io_finish_async(ctx);
8724 io_sqe_buffers_unregister(ctx);
8726 if (ctx->sqo_task) {
8727 put_task_struct(ctx->sqo_task);
8728 ctx->sqo_task = NULL;
8729 mmdrop(ctx->mm_account);
8730 ctx->mm_account = NULL;
8733 #ifdef CONFIG_BLK_CGROUP
8734 if (ctx->sqo_blkcg_css)
8735 css_put(ctx->sqo_blkcg_css);
8738 io_sqe_files_unregister(ctx);
8739 io_eventfd_unregister(ctx);
8740 io_destroy_buffers(ctx);
8741 idr_destroy(&ctx->personality_idr);
8743 #if defined(CONFIG_UNIX)
8744 if (ctx->ring_sock) {
8745 ctx->ring_sock->file = NULL; /* so that iput() is called */
8746 sock_release(ctx->ring_sock);
8750 io_mem_free(ctx->rings);
8751 io_mem_free(ctx->sq_sqes);
8753 percpu_ref_exit(&ctx->refs);
8754 free_uid(ctx->user);
8755 put_cred(ctx->creds);
8756 io_req_caches_free(ctx, NULL);
8757 kfree(ctx->cancel_hash);
8761 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8763 struct io_ring_ctx *ctx = file->private_data;
8766 poll_wait(file, &ctx->cq_wait, wait);
8768 * synchronizes with barrier from wq_has_sleeper call in
8772 if (!io_sqring_full(ctx))
8773 mask |= EPOLLOUT | EPOLLWRNORM;
8776 * Don't flush cqring overflow list here, just do a simple check.
8777 * Otherwise there could possible be ABBA deadlock:
8780 * lock(&ctx->uring_lock);
8782 * lock(&ctx->uring_lock);
8785 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
8786 * pushs them to do the flush.
8788 if (io_cqring_events(ctx) || test_bit(0, &ctx->cq_check_overflow))
8789 mask |= EPOLLIN | EPOLLRDNORM;
8794 static int io_uring_fasync(int fd, struct file *file, int on)
8796 struct io_ring_ctx *ctx = file->private_data;
8798 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8801 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
8803 struct io_identity *iod;
8805 iod = idr_remove(&ctx->personality_idr, id);
8807 put_cred(iod->creds);
8808 if (refcount_dec_and_test(&iod->count))
8816 static int io_remove_personalities(int id, void *p, void *data)
8818 struct io_ring_ctx *ctx = data;
8820 io_unregister_personality(ctx, id);
8824 static void io_ring_exit_work(struct work_struct *work)
8826 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
8830 * If we're doing polled IO and end up having requests being
8831 * submitted async (out-of-line), then completions can come in while
8832 * we're waiting for refs to drop. We need to reap these manually,
8833 * as nobody else will be looking for them.
8836 io_uring_try_cancel_requests(ctx, NULL, NULL);
8837 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
8838 io_ring_ctx_free(ctx);
8841 static bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
8843 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8845 return req->ctx == data;
8848 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8850 mutex_lock(&ctx->uring_lock);
8851 percpu_ref_kill(&ctx->refs);
8853 if (WARN_ON_ONCE((ctx->flags & IORING_SETUP_SQPOLL) && !ctx->sqo_dead))
8856 /* if force is set, the ring is going away. always drop after that */
8857 ctx->cq_overflow_flushed = 1;
8859 __io_cqring_overflow_flush(ctx, true, NULL, NULL);
8860 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
8861 mutex_unlock(&ctx->uring_lock);
8863 io_kill_timeouts(ctx, NULL, NULL);
8864 io_poll_remove_all(ctx, NULL, NULL);
8867 io_wq_cancel_cb(ctx->io_wq, io_cancel_ctx_cb, ctx, true);
8869 /* if we failed setting up the ctx, we might not have any rings */
8870 io_iopoll_try_reap_events(ctx);
8872 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8874 * Use system_unbound_wq to avoid spawning tons of event kworkers
8875 * if we're exiting a ton of rings at the same time. It just adds
8876 * noise and overhead, there's no discernable change in runtime
8877 * over using system_wq.
8879 queue_work(system_unbound_wq, &ctx->exit_work);
8882 static int io_uring_release(struct inode *inode, struct file *file)
8884 struct io_ring_ctx *ctx = file->private_data;
8886 file->private_data = NULL;
8887 io_ring_ctx_wait_and_kill(ctx);
8891 struct io_task_cancel {
8892 struct task_struct *task;
8893 struct files_struct *files;
8896 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8898 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8899 struct io_task_cancel *cancel = data;
8902 if (cancel->files && (req->flags & REQ_F_LINK_TIMEOUT)) {
8903 unsigned long flags;
8904 struct io_ring_ctx *ctx = req->ctx;
8906 /* protect against races with linked timeouts */
8907 spin_lock_irqsave(&ctx->completion_lock, flags);
8908 ret = io_match_task(req, cancel->task, cancel->files);
8909 spin_unlock_irqrestore(&ctx->completion_lock, flags);
8911 ret = io_match_task(req, cancel->task, cancel->files);
8916 static void io_cancel_defer_files(struct io_ring_ctx *ctx,
8917 struct task_struct *task,
8918 struct files_struct *files)
8920 struct io_defer_entry *de = NULL;
8923 spin_lock_irq(&ctx->completion_lock);
8924 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8925 if (io_match_task(de->req, task, files)) {
8926 list_cut_position(&list, &ctx->defer_list, &de->list);
8930 spin_unlock_irq(&ctx->completion_lock);
8932 while (!list_empty(&list)) {
8933 de = list_first_entry(&list, struct io_defer_entry, list);
8934 list_del_init(&de->list);
8935 req_set_fail_links(de->req);
8936 io_put_req(de->req);
8937 io_req_complete(de->req, -ECANCELED);
8942 static void io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
8943 struct task_struct *task,
8944 struct files_struct *files)
8946 struct io_task_cancel cancel = { .task = task, .files = files, };
8949 enum io_wq_cancel cret;
8953 cret = io_wq_cancel_cb(ctx->io_wq, io_cancel_task_cb,
8955 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
8958 /* SQPOLL thread does its own polling */
8959 if (!(ctx->flags & IORING_SETUP_SQPOLL) && !files) {
8960 while (!list_empty_careful(&ctx->iopoll_list)) {
8961 io_iopoll_try_reap_events(ctx);
8966 ret |= io_poll_remove_all(ctx, task, files);
8967 ret |= io_kill_timeouts(ctx, task, files);
8968 ret |= io_run_task_work();
8969 io_cqring_overflow_flush(ctx, true, task, files);
8976 static int io_uring_count_inflight(struct io_ring_ctx *ctx,
8977 struct task_struct *task,
8978 struct files_struct *files)
8980 struct io_kiocb *req;
8983 spin_lock_irq(&ctx->inflight_lock);
8984 list_for_each_entry(req, &ctx->inflight_list, inflight_entry)
8985 cnt += io_match_task(req, task, files);
8986 spin_unlock_irq(&ctx->inflight_lock);
8990 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
8991 struct task_struct *task,
8992 struct files_struct *files)
8994 while (!list_empty_careful(&ctx->inflight_list)) {
8998 inflight = io_uring_count_inflight(ctx, task, files);
9002 io_uring_try_cancel_requests(ctx, task, files);
9005 io_sq_thread_unpark(ctx->sq_data);
9006 prepare_to_wait(&task->io_uring->wait, &wait,
9007 TASK_UNINTERRUPTIBLE);
9008 if (inflight == io_uring_count_inflight(ctx, task, files))
9010 finish_wait(&task->io_uring->wait, &wait);
9012 io_sq_thread_park(ctx->sq_data);
9016 static void io_disable_sqo_submit(struct io_ring_ctx *ctx)
9018 mutex_lock(&ctx->uring_lock);
9020 mutex_unlock(&ctx->uring_lock);
9022 /* make sure callers enter the ring to get error */
9024 io_ring_set_wakeup_flag(ctx);
9028 * We need to iteratively cancel requests, in case a request has dependent
9029 * hard links. These persist even for failure of cancelations, hence keep
9030 * looping until none are found.
9032 static void io_uring_cancel_task_requests(struct io_ring_ctx *ctx,
9033 struct files_struct *files)
9035 struct task_struct *task = current;
9037 if ((ctx->flags & IORING_SETUP_SQPOLL) && ctx->sq_data) {
9038 io_disable_sqo_submit(ctx);
9039 task = ctx->sq_data->thread;
9040 atomic_inc(&task->io_uring->in_idle);
9041 io_sq_thread_park(ctx->sq_data);
9044 io_cancel_defer_files(ctx, task, files);
9046 io_uring_cancel_files(ctx, task, files);
9048 io_uring_try_cancel_requests(ctx, task, NULL);
9050 if ((ctx->flags & IORING_SETUP_SQPOLL) && ctx->sq_data) {
9051 atomic_dec(&task->io_uring->in_idle);
9053 * If the files that are going away are the ones in the thread
9054 * identity, clear them out.
9056 if (task->io_uring->identity->files == files)
9057 task->io_uring->identity->files = NULL;
9058 io_sq_thread_unpark(ctx->sq_data);
9063 * Note that this task has used io_uring. We use it for cancelation purposes.
9065 static int io_uring_add_task_file(struct io_ring_ctx *ctx, struct file *file)
9067 struct io_uring_task *tctx = current->io_uring;
9070 if (unlikely(!tctx)) {
9071 ret = io_uring_alloc_task_context(current);
9074 tctx = current->io_uring;
9076 if (tctx->last != file) {
9077 void *old = xa_load(&tctx->xa, (unsigned long)file);
9081 ret = xa_err(xa_store(&tctx->xa, (unsigned long)file,
9088 /* one and only SQPOLL file note, held by sqo_task */
9089 WARN_ON_ONCE((ctx->flags & IORING_SETUP_SQPOLL) &&
9090 current != ctx->sqo_task);
9096 * This is race safe in that the task itself is doing this, hence it
9097 * cannot be going through the exit/cancel paths at the same time.
9098 * This cannot be modified while exit/cancel is running.
9100 if (!tctx->sqpoll && (ctx->flags & IORING_SETUP_SQPOLL))
9101 tctx->sqpoll = true;
9107 * Remove this io_uring_file -> task mapping.
9109 static void io_uring_del_task_file(struct file *file)
9111 struct io_uring_task *tctx = current->io_uring;
9113 if (tctx->last == file)
9115 file = xa_erase(&tctx->xa, (unsigned long)file);
9120 static void io_uring_remove_task_files(struct io_uring_task *tctx)
9123 unsigned long index;
9125 xa_for_each(&tctx->xa, index, file)
9126 io_uring_del_task_file(file);
9129 void __io_uring_files_cancel(struct files_struct *files)
9131 struct io_uring_task *tctx = current->io_uring;
9133 unsigned long index;
9135 /* make sure overflow events are dropped */
9136 atomic_inc(&tctx->in_idle);
9137 xa_for_each(&tctx->xa, index, file)
9138 io_uring_cancel_task_requests(file->private_data, files);
9139 atomic_dec(&tctx->in_idle);
9142 io_uring_remove_task_files(tctx);
9145 static s64 tctx_inflight(struct io_uring_task *tctx)
9147 return percpu_counter_sum(&tctx->inflight);
9150 static void io_uring_cancel_sqpoll(struct io_ring_ctx *ctx)
9152 struct io_uring_task *tctx;
9158 tctx = ctx->sq_data->thread->io_uring;
9159 io_disable_sqo_submit(ctx);
9161 atomic_inc(&tctx->in_idle);
9163 /* read completions before cancelations */
9164 inflight = tctx_inflight(tctx);
9167 io_uring_cancel_task_requests(ctx, NULL);
9169 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
9171 * If we've seen completions, retry without waiting. This
9172 * avoids a race where a completion comes in before we did
9173 * prepare_to_wait().
9175 if (inflight == tctx_inflight(tctx))
9177 finish_wait(&tctx->wait, &wait);
9179 atomic_dec(&tctx->in_idle);
9183 * Find any io_uring fd that this task has registered or done IO on, and cancel
9186 void __io_uring_task_cancel(void)
9188 struct io_uring_task *tctx = current->io_uring;
9192 /* make sure overflow events are dropped */
9193 atomic_inc(&tctx->in_idle);
9195 /* trigger io_disable_sqo_submit() */
9198 unsigned long index;
9200 xa_for_each(&tctx->xa, index, file)
9201 io_uring_cancel_sqpoll(file->private_data);
9205 /* read completions before cancelations */
9206 inflight = tctx_inflight(tctx);
9209 __io_uring_files_cancel(NULL);
9211 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
9214 * If we've seen completions, retry without waiting. This
9215 * avoids a race where a completion comes in before we did
9216 * prepare_to_wait().
9218 if (inflight == tctx_inflight(tctx))
9220 finish_wait(&tctx->wait, &wait);
9223 atomic_dec(&tctx->in_idle);
9225 io_uring_remove_task_files(tctx);
9228 static int io_uring_flush(struct file *file, void *data)
9230 struct io_uring_task *tctx = current->io_uring;
9231 struct io_ring_ctx *ctx = file->private_data;
9233 if (fatal_signal_pending(current) || (current->flags & PF_EXITING)) {
9234 io_uring_cancel_task_requests(ctx, NULL);
9235 io_req_caches_free(ctx, current);
9241 /* we should have cancelled and erased it before PF_EXITING */
9242 WARN_ON_ONCE((current->flags & PF_EXITING) &&
9243 xa_load(&tctx->xa, (unsigned long)file));
9246 * fput() is pending, will be 2 if the only other ref is our potential
9247 * task file note. If the task is exiting, drop regardless of count.
9249 if (atomic_long_read(&file->f_count) != 2)
9252 if (ctx->flags & IORING_SETUP_SQPOLL) {
9253 /* there is only one file note, which is owned by sqo_task */
9254 WARN_ON_ONCE(ctx->sqo_task != current &&
9255 xa_load(&tctx->xa, (unsigned long)file));
9256 /* sqo_dead check is for when this happens after cancellation */
9257 WARN_ON_ONCE(ctx->sqo_task == current && !ctx->sqo_dead &&
9258 !xa_load(&tctx->xa, (unsigned long)file));
9260 io_disable_sqo_submit(ctx);
9263 if (!(ctx->flags & IORING_SETUP_SQPOLL) || ctx->sqo_task == current)
9264 io_uring_del_task_file(file);
9268 static void *io_uring_validate_mmap_request(struct file *file,
9269 loff_t pgoff, size_t sz)
9271 struct io_ring_ctx *ctx = file->private_data;
9272 loff_t offset = pgoff << PAGE_SHIFT;
9277 case IORING_OFF_SQ_RING:
9278 case IORING_OFF_CQ_RING:
9281 case IORING_OFF_SQES:
9285 return ERR_PTR(-EINVAL);
9288 page = virt_to_head_page(ptr);
9289 if (sz > page_size(page))
9290 return ERR_PTR(-EINVAL);
9297 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9299 size_t sz = vma->vm_end - vma->vm_start;
9303 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
9305 return PTR_ERR(ptr);
9307 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
9308 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
9311 #else /* !CONFIG_MMU */
9313 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9315 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
9318 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
9320 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
9323 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
9324 unsigned long addr, unsigned long len,
9325 unsigned long pgoff, unsigned long flags)
9329 ptr = io_uring_validate_mmap_request(file, pgoff, len);
9331 return PTR_ERR(ptr);
9333 return (unsigned long) ptr;
9336 #endif /* !CONFIG_MMU */
9338 static int io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
9344 if (!io_sqring_full(ctx))
9347 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9349 if (unlikely(ctx->sqo_dead)) {
9354 if (!io_sqring_full(ctx))
9358 } while (!signal_pending(current));
9360 finish_wait(&ctx->sqo_sq_wait, &wait);
9365 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9366 struct __kernel_timespec __user **ts,
9367 const sigset_t __user **sig)
9369 struct io_uring_getevents_arg arg;
9372 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9373 * is just a pointer to the sigset_t.
9375 if (!(flags & IORING_ENTER_EXT_ARG)) {
9376 *sig = (const sigset_t __user *) argp;
9382 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9383 * timespec and sigset_t pointers if good.
9385 if (*argsz != sizeof(arg))
9387 if (copy_from_user(&arg, argp, sizeof(arg)))
9389 *sig = u64_to_user_ptr(arg.sigmask);
9390 *argsz = arg.sigmask_sz;
9391 *ts = u64_to_user_ptr(arg.ts);
9395 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9396 u32, min_complete, u32, flags, const void __user *, argp,
9399 struct io_ring_ctx *ctx;
9406 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9407 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG))
9415 if (f.file->f_op != &io_uring_fops)
9419 ctx = f.file->private_data;
9420 if (!percpu_ref_tryget(&ctx->refs))
9424 if (ctx->flags & IORING_SETUP_R_DISABLED)
9428 * For SQ polling, the thread will do all submissions and completions.
9429 * Just return the requested submit count, and wake the thread if
9433 if (ctx->flags & IORING_SETUP_SQPOLL) {
9434 io_cqring_overflow_flush(ctx, false, NULL, NULL);
9437 if (unlikely(ctx->sqo_dead))
9439 if (flags & IORING_ENTER_SQ_WAKEUP)
9440 wake_up(&ctx->sq_data->wait);
9441 if (flags & IORING_ENTER_SQ_WAIT) {
9442 ret = io_sqpoll_wait_sq(ctx);
9446 submitted = to_submit;
9447 } else if (to_submit) {
9448 ret = io_uring_add_task_file(ctx, f.file);
9451 mutex_lock(&ctx->uring_lock);
9452 submitted = io_submit_sqes(ctx, to_submit);
9453 mutex_unlock(&ctx->uring_lock);
9455 if (submitted != to_submit)
9458 if (flags & IORING_ENTER_GETEVENTS) {
9459 const sigset_t __user *sig;
9460 struct __kernel_timespec __user *ts;
9462 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
9466 min_complete = min(min_complete, ctx->cq_entries);
9469 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9470 * space applications don't need to do io completion events
9471 * polling again, they can rely on io_sq_thread to do polling
9472 * work, which can reduce cpu usage and uring_lock contention.
9474 if (ctx->flags & IORING_SETUP_IOPOLL &&
9475 !(ctx->flags & IORING_SETUP_SQPOLL)) {
9476 ret = io_iopoll_check(ctx, min_complete);
9478 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
9483 percpu_ref_put(&ctx->refs);
9486 return submitted ? submitted : ret;
9489 #ifdef CONFIG_PROC_FS
9490 static int io_uring_show_cred(int id, void *p, void *data)
9492 struct io_identity *iod = p;
9493 const struct cred *cred = iod->creds;
9494 struct seq_file *m = data;
9495 struct user_namespace *uns = seq_user_ns(m);
9496 struct group_info *gi;
9501 seq_printf(m, "%5d\n", id);
9502 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
9503 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
9504 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
9505 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
9506 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
9507 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
9508 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
9509 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
9510 seq_puts(m, "\n\tGroups:\t");
9511 gi = cred->group_info;
9512 for (g = 0; g < gi->ngroups; g++) {
9513 seq_put_decimal_ull(m, g ? " " : "",
9514 from_kgid_munged(uns, gi->gid[g]));
9516 seq_puts(m, "\n\tCapEff:\t");
9517 cap = cred->cap_effective;
9518 CAP_FOR_EACH_U32(__capi)
9519 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
9524 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
9526 struct io_sq_data *sq = NULL;
9531 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9532 * since fdinfo case grabs it in the opposite direction of normal use
9533 * cases. If we fail to get the lock, we just don't iterate any
9534 * structures that could be going away outside the io_uring mutex.
9536 has_lock = mutex_trylock(&ctx->uring_lock);
9538 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL))
9541 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9542 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9543 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9544 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9545 struct file *f = *io_fixed_file_slot(ctx->file_data, i);
9548 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9550 seq_printf(m, "%5u: <none>\n", i);
9552 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9553 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9554 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
9556 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
9557 (unsigned int) buf->len);
9559 if (has_lock && !idr_is_empty(&ctx->personality_idr)) {
9560 seq_printf(m, "Personalities:\n");
9561 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
9563 seq_printf(m, "PollList:\n");
9564 spin_lock_irq(&ctx->completion_lock);
9565 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9566 struct hlist_head *list = &ctx->cancel_hash[i];
9567 struct io_kiocb *req;
9569 hlist_for_each_entry(req, list, hash_node)
9570 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9571 req->task->task_works != NULL);
9573 spin_unlock_irq(&ctx->completion_lock);
9575 mutex_unlock(&ctx->uring_lock);
9578 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9580 struct io_ring_ctx *ctx = f->private_data;
9582 if (percpu_ref_tryget(&ctx->refs)) {
9583 __io_uring_show_fdinfo(ctx, m);
9584 percpu_ref_put(&ctx->refs);
9589 static const struct file_operations io_uring_fops = {
9590 .release = io_uring_release,
9591 .flush = io_uring_flush,
9592 .mmap = io_uring_mmap,
9594 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9595 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9597 .poll = io_uring_poll,
9598 .fasync = io_uring_fasync,
9599 #ifdef CONFIG_PROC_FS
9600 .show_fdinfo = io_uring_show_fdinfo,
9604 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9605 struct io_uring_params *p)
9607 struct io_rings *rings;
9608 size_t size, sq_array_offset;
9610 /* make sure these are sane, as we already accounted them */
9611 ctx->sq_entries = p->sq_entries;
9612 ctx->cq_entries = p->cq_entries;
9614 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9615 if (size == SIZE_MAX)
9618 rings = io_mem_alloc(size);
9623 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9624 rings->sq_ring_mask = p->sq_entries - 1;
9625 rings->cq_ring_mask = p->cq_entries - 1;
9626 rings->sq_ring_entries = p->sq_entries;
9627 rings->cq_ring_entries = p->cq_entries;
9628 ctx->sq_mask = rings->sq_ring_mask;
9629 ctx->cq_mask = rings->cq_ring_mask;
9631 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9632 if (size == SIZE_MAX) {
9633 io_mem_free(ctx->rings);
9638 ctx->sq_sqes = io_mem_alloc(size);
9639 if (!ctx->sq_sqes) {
9640 io_mem_free(ctx->rings);
9648 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
9652 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9656 ret = io_uring_add_task_file(ctx, file);
9661 fd_install(fd, file);
9666 * Allocate an anonymous fd, this is what constitutes the application
9667 * visible backing of an io_uring instance. The application mmaps this
9668 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9669 * we have to tie this fd to a socket for file garbage collection purposes.
9671 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
9674 #if defined(CONFIG_UNIX)
9677 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9680 return ERR_PTR(ret);
9683 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9684 O_RDWR | O_CLOEXEC);
9685 #if defined(CONFIG_UNIX)
9687 sock_release(ctx->ring_sock);
9688 ctx->ring_sock = NULL;
9690 ctx->ring_sock->file = file;
9696 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9697 struct io_uring_params __user *params)
9699 struct user_struct *user = NULL;
9700 struct io_ring_ctx *ctx;
9706 if (entries > IORING_MAX_ENTRIES) {
9707 if (!(p->flags & IORING_SETUP_CLAMP))
9709 entries = IORING_MAX_ENTRIES;
9713 * Use twice as many entries for the CQ ring. It's possible for the
9714 * application to drive a higher depth than the size of the SQ ring,
9715 * since the sqes are only used at submission time. This allows for
9716 * some flexibility in overcommitting a bit. If the application has
9717 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9718 * of CQ ring entries manually.
9720 p->sq_entries = roundup_pow_of_two(entries);
9721 if (p->flags & IORING_SETUP_CQSIZE) {
9723 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9724 * to a power-of-two, if it isn't already. We do NOT impose
9725 * any cq vs sq ring sizing.
9729 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9730 if (!(p->flags & IORING_SETUP_CLAMP))
9732 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9734 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9735 if (p->cq_entries < p->sq_entries)
9738 p->cq_entries = 2 * p->sq_entries;
9741 user = get_uid(current_user());
9743 ctx = io_ring_ctx_alloc(p);
9748 ctx->compat = in_compat_syscall();
9749 ctx->limit_mem = !capable(CAP_IPC_LOCK);
9751 ctx->creds = get_current_cred();
9753 ctx->loginuid = current->loginuid;
9754 ctx->sessionid = current->sessionid;
9756 ctx->sqo_task = get_task_struct(current);
9759 * This is just grabbed for accounting purposes. When a process exits,
9760 * the mm is exited and dropped before the files, hence we need to hang
9761 * on to this mm purely for the purposes of being able to unaccount
9762 * memory (locked/pinned vm). It's not used for anything else.
9764 mmgrab(current->mm);
9765 ctx->mm_account = current->mm;
9767 #ifdef CONFIG_BLK_CGROUP
9769 * The sq thread will belong to the original cgroup it was inited in.
9770 * If the cgroup goes offline (e.g. disabling the io controller), then
9771 * issued bios will be associated with the closest cgroup later in the
9775 ctx->sqo_blkcg_css = blkcg_css();
9776 ret = css_tryget_online(ctx->sqo_blkcg_css);
9779 /* don't init against a dying cgroup, have the user try again */
9780 ctx->sqo_blkcg_css = NULL;
9785 ret = io_allocate_scq_urings(ctx, p);
9789 ret = io_sq_offload_create(ctx, p);
9793 if (!(p->flags & IORING_SETUP_R_DISABLED))
9794 io_sq_offload_start(ctx);
9796 memset(&p->sq_off, 0, sizeof(p->sq_off));
9797 p->sq_off.head = offsetof(struct io_rings, sq.head);
9798 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9799 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9800 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9801 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9802 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9803 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9805 memset(&p->cq_off, 0, sizeof(p->cq_off));
9806 p->cq_off.head = offsetof(struct io_rings, cq.head);
9807 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9808 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9809 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9810 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9811 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9812 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9814 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9815 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9816 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9817 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
9818 IORING_FEAT_EXT_ARG;
9820 if (copy_to_user(params, p, sizeof(*p))) {
9825 file = io_uring_get_file(ctx);
9827 ret = PTR_ERR(file);
9832 * Install ring fd as the very last thing, so we don't risk someone
9833 * having closed it before we finish setup
9835 ret = io_uring_install_fd(ctx, file);
9837 io_disable_sqo_submit(ctx);
9838 /* fput will clean it up */
9843 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9846 io_disable_sqo_submit(ctx);
9847 io_ring_ctx_wait_and_kill(ctx);
9852 * Sets up an aio uring context, and returns the fd. Applications asks for a
9853 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9854 * params structure passed in.
9856 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9858 struct io_uring_params p;
9861 if (copy_from_user(&p, params, sizeof(p)))
9863 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9868 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9869 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9870 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9871 IORING_SETUP_R_DISABLED))
9874 return io_uring_create(entries, &p, params);
9877 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9878 struct io_uring_params __user *, params)
9880 return io_uring_setup(entries, params);
9883 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9885 struct io_uring_probe *p;
9889 size = struct_size(p, ops, nr_args);
9890 if (size == SIZE_MAX)
9892 p = kzalloc(size, GFP_KERNEL);
9897 if (copy_from_user(p, arg, size))
9900 if (memchr_inv(p, 0, size))
9903 p->last_op = IORING_OP_LAST - 1;
9904 if (nr_args > IORING_OP_LAST)
9905 nr_args = IORING_OP_LAST;
9907 for (i = 0; i < nr_args; i++) {
9909 if (!io_op_defs[i].not_supported)
9910 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9915 if (copy_to_user(arg, p, size))
9922 static int io_register_personality(struct io_ring_ctx *ctx)
9924 struct io_identity *id;
9927 id = kmalloc(sizeof(*id), GFP_KERNEL);
9931 io_init_identity(id);
9932 id->creds = get_current_cred();
9934 ret = idr_alloc_cyclic(&ctx->personality_idr, id, 1, USHRT_MAX, GFP_KERNEL);
9936 put_cred(id->creds);
9942 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9943 unsigned int nr_args)
9945 struct io_uring_restriction *res;
9949 /* Restrictions allowed only if rings started disabled */
9950 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9953 /* We allow only a single restrictions registration */
9954 if (ctx->restrictions.registered)
9957 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9960 size = array_size(nr_args, sizeof(*res));
9961 if (size == SIZE_MAX)
9964 res = memdup_user(arg, size);
9966 return PTR_ERR(res);
9970 for (i = 0; i < nr_args; i++) {
9971 switch (res[i].opcode) {
9972 case IORING_RESTRICTION_REGISTER_OP:
9973 if (res[i].register_op >= IORING_REGISTER_LAST) {
9978 __set_bit(res[i].register_op,
9979 ctx->restrictions.register_op);
9981 case IORING_RESTRICTION_SQE_OP:
9982 if (res[i].sqe_op >= IORING_OP_LAST) {
9987 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9989 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9990 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9992 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9993 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
10002 /* Reset all restrictions if an error happened */
10004 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
10006 ctx->restrictions.registered = true;
10012 static int io_register_enable_rings(struct io_ring_ctx *ctx)
10014 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
10017 if (ctx->restrictions.registered)
10018 ctx->restricted = 1;
10020 ctx->flags &= ~IORING_SETUP_R_DISABLED;
10022 io_sq_offload_start(ctx);
10027 static bool io_register_op_must_quiesce(int op)
10030 case IORING_UNREGISTER_FILES:
10031 case IORING_REGISTER_FILES_UPDATE:
10032 case IORING_REGISTER_PROBE:
10033 case IORING_REGISTER_PERSONALITY:
10034 case IORING_UNREGISTER_PERSONALITY:
10041 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
10042 void __user *arg, unsigned nr_args)
10043 __releases(ctx->uring_lock)
10044 __acquires(ctx->uring_lock)
10049 * We're inside the ring mutex, if the ref is already dying, then
10050 * someone else killed the ctx or is already going through
10051 * io_uring_register().
10053 if (percpu_ref_is_dying(&ctx->refs))
10056 if (io_register_op_must_quiesce(opcode)) {
10057 percpu_ref_kill(&ctx->refs);
10060 * Drop uring mutex before waiting for references to exit. If
10061 * another thread is currently inside io_uring_enter() it might
10062 * need to grab the uring_lock to make progress. If we hold it
10063 * here across the drain wait, then we can deadlock. It's safe
10064 * to drop the mutex here, since no new references will come in
10065 * after we've killed the percpu ref.
10067 mutex_unlock(&ctx->uring_lock);
10069 ret = wait_for_completion_interruptible(&ctx->ref_comp);
10072 ret = io_run_task_work_sig();
10077 mutex_lock(&ctx->uring_lock);
10080 percpu_ref_resurrect(&ctx->refs);
10085 if (ctx->restricted) {
10086 if (opcode >= IORING_REGISTER_LAST) {
10091 if (!test_bit(opcode, ctx->restrictions.register_op)) {
10098 case IORING_REGISTER_BUFFERS:
10099 ret = io_sqe_buffers_register(ctx, arg, nr_args);
10101 case IORING_UNREGISTER_BUFFERS:
10103 if (arg || nr_args)
10105 ret = io_sqe_buffers_unregister(ctx);
10107 case IORING_REGISTER_FILES:
10108 ret = io_sqe_files_register(ctx, arg, nr_args);
10110 case IORING_UNREGISTER_FILES:
10112 if (arg || nr_args)
10114 ret = io_sqe_files_unregister(ctx);
10116 case IORING_REGISTER_FILES_UPDATE:
10117 ret = io_sqe_files_update(ctx, arg, nr_args);
10119 case IORING_REGISTER_EVENTFD:
10120 case IORING_REGISTER_EVENTFD_ASYNC:
10124 ret = io_eventfd_register(ctx, arg);
10127 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
10128 ctx->eventfd_async = 1;
10130 ctx->eventfd_async = 0;
10132 case IORING_UNREGISTER_EVENTFD:
10134 if (arg || nr_args)
10136 ret = io_eventfd_unregister(ctx);
10138 case IORING_REGISTER_PROBE:
10140 if (!arg || nr_args > 256)
10142 ret = io_probe(ctx, arg, nr_args);
10144 case IORING_REGISTER_PERSONALITY:
10146 if (arg || nr_args)
10148 ret = io_register_personality(ctx);
10150 case IORING_UNREGISTER_PERSONALITY:
10154 ret = io_unregister_personality(ctx, nr_args);
10156 case IORING_REGISTER_ENABLE_RINGS:
10158 if (arg || nr_args)
10160 ret = io_register_enable_rings(ctx);
10162 case IORING_REGISTER_RESTRICTIONS:
10163 ret = io_register_restrictions(ctx, arg, nr_args);
10171 if (io_register_op_must_quiesce(opcode)) {
10172 /* bring the ctx back to life */
10173 percpu_ref_reinit(&ctx->refs);
10175 reinit_completion(&ctx->ref_comp);
10180 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
10181 void __user *, arg, unsigned int, nr_args)
10183 struct io_ring_ctx *ctx;
10192 if (f.file->f_op != &io_uring_fops)
10195 ctx = f.file->private_data;
10197 mutex_lock(&ctx->uring_lock);
10198 ret = __io_uring_register(ctx, opcode, arg, nr_args);
10199 mutex_unlock(&ctx->uring_lock);
10200 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
10201 ctx->cq_ev_fd != NULL, ret);
10207 static int __init io_uring_init(void)
10209 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
10210 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
10211 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
10214 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
10215 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
10216 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
10217 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
10218 BUILD_BUG_SQE_ELEM(1, __u8, flags);
10219 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
10220 BUILD_BUG_SQE_ELEM(4, __s32, fd);
10221 BUILD_BUG_SQE_ELEM(8, __u64, off);
10222 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
10223 BUILD_BUG_SQE_ELEM(16, __u64, addr);
10224 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
10225 BUILD_BUG_SQE_ELEM(24, __u32, len);
10226 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
10227 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
10228 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
10229 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
10230 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
10231 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
10232 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
10233 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
10234 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
10235 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
10236 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
10237 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
10238 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
10239 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
10240 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
10241 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
10242 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
10243 BUILD_BUG_SQE_ELEM(42, __u16, personality);
10244 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
10246 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
10247 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
10248 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
10252 __initcall(io_uring_init);