1 // SPDX-License-Identifier: GPL-2.0
3 * Shared application/kernel submission and completion ring pairs, for
4 * supporting fast/efficient IO.
6 * A note on the read/write ordering memory barriers that are matched between
7 * the application and kernel side.
9 * After the application reads the CQ ring tail, it must use an
10 * appropriate smp_rmb() to pair with the smp_wmb() the kernel uses
11 * before writing the tail (using smp_load_acquire to read the tail will
12 * do). It also needs a smp_mb() before updating CQ head (ordering the
13 * entry load(s) with the head store), pairing with an implicit barrier
14 * through a control-dependency in io_get_cqring (smp_store_release to
15 * store head will do). Failure to do so could lead to reading invalid
18 * Likewise, the application must use an appropriate smp_wmb() before
19 * writing the SQ tail (ordering SQ entry stores with the tail store),
20 * which pairs with smp_load_acquire in io_get_sqring (smp_store_release
21 * to store the tail will do). And it needs a barrier ordering the SQ
22 * head load before writing new SQ entries (smp_load_acquire to read
25 * When using the SQ poll thread (IORING_SETUP_SQPOLL), the application
26 * needs to check the SQ flags for IORING_SQ_NEED_WAKEUP *after*
27 * updating the SQ tail; a full memory barrier smp_mb() is needed
30 * Also see the examples in the liburing library:
32 * git://git.kernel.dk/liburing
34 * io_uring also uses READ/WRITE_ONCE() for _any_ store or load that happens
35 * from data shared between the kernel and application. This is done both
36 * for ordering purposes, but also to ensure that once a value is loaded from
37 * data that the application could potentially modify, it remains stable.
39 * Copyright (C) 2018-2019 Jens Axboe
40 * Copyright (c) 2018-2019 Christoph Hellwig
42 #include <linux/kernel.h>
43 #include <linux/init.h>
44 #include <linux/errno.h>
45 #include <linux/syscalls.h>
46 #include <linux/compat.h>
47 #include <net/compat.h>
48 #include <linux/refcount.h>
49 #include <linux/uio.h>
50 #include <linux/bits.h>
52 #include <linux/sched/signal.h>
54 #include <linux/file.h>
55 #include <linux/fdtable.h>
57 #include <linux/mman.h>
58 #include <linux/percpu.h>
59 #include <linux/slab.h>
60 #include <linux/kthread.h>
61 #include <linux/blkdev.h>
62 #include <linux/bvec.h>
63 #include <linux/net.h>
65 #include <net/af_unix.h>
67 #include <linux/anon_inodes.h>
68 #include <linux/sched/mm.h>
69 #include <linux/uaccess.h>
70 #include <linux/nospec.h>
71 #include <linux/sizes.h>
72 #include <linux/hugetlb.h>
73 #include <linux/highmem.h>
74 #include <linux/namei.h>
75 #include <linux/fsnotify.h>
76 #include <linux/fadvise.h>
77 #include <linux/eventpoll.h>
78 #include <linux/fs_struct.h>
79 #include <linux/splice.h>
80 #include <linux/task_work.h>
81 #include <linux/pagemap.h>
82 #include <linux/io_uring.h>
83 #include <linux/blk-cgroup.h>
84 #include <linux/audit.h>
86 #define CREATE_TRACE_POINTS
87 #include <trace/events/io_uring.h>
89 #include <uapi/linux/io_uring.h>
94 #define IORING_MAX_ENTRIES 32768
95 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
98 * Shift of 9 is 512 entries, or exactly one page on 64-bit archs
100 #define IORING_FILE_TABLE_SHIFT 9
101 #define IORING_MAX_FILES_TABLE (1U << IORING_FILE_TABLE_SHIFT)
102 #define IORING_FILE_TABLE_MASK (IORING_MAX_FILES_TABLE - 1)
103 #define IORING_MAX_FIXED_FILES (64 * IORING_MAX_FILES_TABLE)
104 #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
105 IORING_REGISTER_LAST + IORING_OP_LAST)
108 u32 head ____cacheline_aligned_in_smp;
109 u32 tail ____cacheline_aligned_in_smp;
113 * This data is shared with the application through the mmap at offsets
114 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
116 * The offsets to the member fields are published through struct
117 * io_sqring_offsets when calling io_uring_setup.
121 * Head and tail offsets into the ring; the offsets need to be
122 * masked to get valid indices.
124 * The kernel controls head of the sq ring and the tail of the cq ring,
125 * and the application controls tail of the sq ring and the head of the
128 struct io_uring sq, cq;
130 * Bitmasks to apply to head and tail offsets (constant, equals
133 u32 sq_ring_mask, cq_ring_mask;
134 /* Ring sizes (constant, power of 2) */
135 u32 sq_ring_entries, cq_ring_entries;
137 * Number of invalid entries dropped by the kernel due to
138 * invalid index stored in array
140 * Written by the kernel, shouldn't be modified by the
141 * application (i.e. get number of "new events" by comparing to
144 * After a new SQ head value was read by the application this
145 * counter includes all submissions that were dropped reaching
146 * the new SQ head (and possibly more).
152 * Written by the kernel, shouldn't be modified by the
155 * The application needs a full memory barrier before checking
156 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
162 * Written by the application, shouldn't be modified by the
167 * Number of completion events lost because the queue was full;
168 * this should be avoided by the application by making sure
169 * there are not more requests pending than there is space in
170 * the completion queue.
172 * Written by the kernel, shouldn't be modified by the
173 * application (i.e. get number of "new events" by comparing to
176 * As completion events come in out of order this counter is not
177 * ordered with any other data.
181 * Ring buffer of completion events.
183 * The kernel writes completion events fresh every time they are
184 * produced, so the application is allowed to modify pending
187 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
190 struct io_mapped_ubuf {
193 struct bio_vec *bvec;
194 unsigned int nr_bvecs;
195 unsigned long acct_pages;
198 struct fixed_file_table {
202 struct fixed_file_ref_node {
203 struct percpu_ref refs;
204 struct list_head node;
205 struct list_head file_list;
206 struct fixed_file_data *file_data;
207 struct llist_node llist;
211 struct fixed_file_data {
212 struct fixed_file_table *table;
213 struct io_ring_ctx *ctx;
215 struct fixed_file_ref_node *node;
216 struct percpu_ref refs;
217 struct completion done;
218 struct list_head ref_list;
223 struct list_head list;
229 struct io_restriction {
230 DECLARE_BITMAP(register_op, IORING_REGISTER_LAST);
231 DECLARE_BITMAP(sqe_op, IORING_OP_LAST);
232 u8 sqe_flags_allowed;
233 u8 sqe_flags_required;
241 /* ctx's that are using this sqd */
242 struct list_head ctx_list;
243 struct list_head ctx_new_list;
244 struct mutex ctx_lock;
246 struct task_struct *thread;
247 struct wait_queue_head wait;
249 unsigned sq_thread_idle;
254 struct percpu_ref refs;
255 } ____cacheline_aligned_in_smp;
259 unsigned int compat: 1;
260 unsigned int limit_mem: 1;
261 unsigned int cq_overflow_flushed: 1;
262 unsigned int drain_next: 1;
263 unsigned int eventfd_async: 1;
264 unsigned int restricted: 1;
267 * Ring buffer of indices into array of io_uring_sqe, which is
268 * mmapped by the application using the IORING_OFF_SQES offset.
270 * This indirection could e.g. be used to assign fixed
271 * io_uring_sqe entries to operations and only submit them to
272 * the queue when needed.
274 * The kernel modifies neither the indices array nor the entries
278 unsigned cached_sq_head;
281 unsigned sq_thread_idle;
282 unsigned cached_sq_dropped;
283 unsigned cached_cq_overflow;
284 unsigned long sq_check_overflow;
286 struct list_head defer_list;
287 struct list_head timeout_list;
288 struct list_head cq_overflow_list;
290 struct io_uring_sqe *sq_sqes;
291 } ____cacheline_aligned_in_smp;
293 struct io_rings *rings;
299 * For SQPOLL usage - we hold a reference to the parent task, so we
300 * have access to the ->files
302 struct task_struct *sqo_task;
304 /* Only used for accounting purposes */
305 struct mm_struct *mm_account;
307 #ifdef CONFIG_BLK_CGROUP
308 struct cgroup_subsys_state *sqo_blkcg_css;
311 struct io_sq_data *sq_data; /* if using sq thread polling */
313 struct wait_queue_head sqo_sq_wait;
314 struct list_head sqd_list;
317 * If used, fixed file set. Writers must ensure that ->refs is dead,
318 * readers must ensure that ->refs is alive as long as the file* is
319 * used. Only updated through io_uring_register(2).
321 struct fixed_file_data *file_data;
322 unsigned nr_user_files;
324 /* if used, fixed mapped user buffers */
325 unsigned nr_user_bufs;
326 struct io_mapped_ubuf *user_bufs;
328 struct user_struct *user;
330 const struct cred *creds;
334 unsigned int sessionid;
337 struct completion ref_comp;
338 struct completion sq_thread_comp;
340 /* if all else fails... */
341 struct io_kiocb *fallback_req;
343 #if defined(CONFIG_UNIX)
344 struct socket *ring_sock;
347 struct idr io_buffer_idr;
349 struct idr personality_idr;
352 unsigned cached_cq_tail;
355 atomic_t cq_timeouts;
356 unsigned long cq_check_overflow;
357 struct wait_queue_head cq_wait;
358 struct fasync_struct *cq_fasync;
359 struct eventfd_ctx *cq_ev_fd;
360 } ____cacheline_aligned_in_smp;
363 struct mutex uring_lock;
364 wait_queue_head_t wait;
365 } ____cacheline_aligned_in_smp;
368 spinlock_t completion_lock;
371 * ->iopoll_list is protected by the ctx->uring_lock for
372 * io_uring instances that don't use IORING_SETUP_SQPOLL.
373 * For SQPOLL, only the single threaded io_sq_thread() will
374 * manipulate the list, hence no extra locking is needed there.
376 struct list_head iopoll_list;
377 struct hlist_head *cancel_hash;
378 unsigned cancel_hash_bits;
379 bool poll_multi_file;
381 spinlock_t inflight_lock;
382 struct list_head inflight_list;
383 } ____cacheline_aligned_in_smp;
385 struct delayed_work file_put_work;
386 struct llist_head file_put_llist;
388 struct work_struct exit_work;
389 struct io_restriction restrictions;
393 * First field must be the file pointer in all the
394 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
396 struct io_poll_iocb {
398 struct wait_queue_head *head;
402 struct wait_queue_entry wait;
405 struct io_poll_remove {
412 struct file *put_file;
416 struct io_timeout_data {
417 struct io_kiocb *req;
418 struct hrtimer timer;
419 struct timespec64 ts;
420 enum hrtimer_mode mode;
425 struct sockaddr __user *addr;
426 int __user *addr_len;
428 unsigned long nofile;
448 struct list_head list;
449 /* head of the link, used by linked timeouts only */
450 struct io_kiocb *head;
453 struct io_timeout_rem {
458 struct timespec64 ts;
463 /* NOTE: kiocb has the file as the first member, so don't do it here */
471 struct sockaddr __user *addr;
478 struct user_msghdr __user *umsg;
484 struct io_buffer *kbuf;
490 bool ignore_nonblock;
491 struct filename *filename;
493 unsigned long nofile;
496 struct io_files_update {
522 struct epoll_event event;
526 struct file *file_out;
527 struct file *file_in;
534 struct io_provide_buf {
548 const char __user *filename;
549 struct statx __user *buffer;
561 struct filename *oldpath;
562 struct filename *newpath;
570 struct filename *filename;
573 struct io_completion {
575 struct list_head list;
579 struct io_async_connect {
580 struct sockaddr_storage address;
583 struct io_async_msghdr {
584 struct iovec fast_iov[UIO_FASTIOV];
586 struct sockaddr __user *uaddr;
588 struct sockaddr_storage addr;
592 struct iovec fast_iov[UIO_FASTIOV];
593 const struct iovec *free_iovec;
594 struct iov_iter iter;
596 struct wait_page_queue wpq;
600 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
601 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
602 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
603 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
604 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
605 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
611 REQ_F_LINK_TIMEOUT_BIT,
613 REQ_F_NEED_CLEANUP_BIT,
615 REQ_F_BUFFER_SELECTED_BIT,
616 REQ_F_NO_FILE_TABLE_BIT,
617 REQ_F_WORK_INITIALIZED_BIT,
618 REQ_F_LTIMEOUT_ACTIVE_BIT,
620 /* not a real bit, just to check we're not overflowing the space */
626 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
627 /* drain existing IO first */
628 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
630 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
631 /* doesn't sever on completion < 0 */
632 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
634 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
635 /* IOSQE_BUFFER_SELECT */
636 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
638 /* fail rest of links */
639 REQ_F_FAIL_LINK = BIT(REQ_F_FAIL_LINK_BIT),
640 /* on inflight list */
641 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
642 /* read/write uses file position */
643 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
644 /* must not punt to workers */
645 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
646 /* has or had linked timeout */
647 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
649 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
651 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
652 /* already went through poll handler */
653 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
654 /* buffer already selected */
655 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
656 /* doesn't need file table for this request */
657 REQ_F_NO_FILE_TABLE = BIT(REQ_F_NO_FILE_TABLE_BIT),
658 /* io_wq_work is initialized */
659 REQ_F_WORK_INITIALIZED = BIT(REQ_F_WORK_INITIALIZED_BIT),
660 /* linked timeout is active, i.e. prepared by link's head */
661 REQ_F_LTIMEOUT_ACTIVE = BIT(REQ_F_LTIMEOUT_ACTIVE_BIT),
665 struct io_poll_iocb poll;
666 struct io_poll_iocb *double_poll;
670 * NOTE! Each of the iocb union members has the file pointer
671 * as the first entry in their struct definition. So you can
672 * access the file pointer through any of the sub-structs,
673 * or directly as just 'ki_filp' in this struct.
679 struct io_poll_iocb poll;
680 struct io_poll_remove poll_remove;
681 struct io_accept accept;
683 struct io_cancel cancel;
684 struct io_timeout timeout;
685 struct io_timeout_rem timeout_rem;
686 struct io_connect connect;
687 struct io_sr_msg sr_msg;
689 struct io_close close;
690 struct io_files_update files_update;
691 struct io_fadvise fadvise;
692 struct io_madvise madvise;
693 struct io_epoll epoll;
694 struct io_splice splice;
695 struct io_provide_buf pbuf;
696 struct io_statx statx;
697 struct io_shutdown shutdown;
698 struct io_rename rename;
699 struct io_unlink unlink;
700 /* use only after cleaning per-op data, see io_clean_op() */
701 struct io_completion compl;
704 /* opcode allocated if it needs to store data for async defer */
707 /* polled IO has completed */
713 struct io_ring_ctx *ctx;
716 struct task_struct *task;
719 struct io_kiocb *link;
720 struct percpu_ref *fixed_file_refs;
723 * 1. used with ctx->iopoll_list with reads/writes
724 * 2. to track reqs with ->files (see io_op_def::file_table)
726 struct list_head inflight_entry;
727 struct callback_head task_work;
728 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
729 struct hlist_node hash_node;
730 struct async_poll *apoll;
731 struct io_wq_work work;
734 struct io_defer_entry {
735 struct list_head list;
736 struct io_kiocb *req;
740 #define IO_IOPOLL_BATCH 8
742 struct io_comp_state {
744 struct list_head list;
745 struct io_ring_ctx *ctx;
748 struct io_submit_state {
749 struct blk_plug plug;
752 * io_kiocb alloc cache
754 void *reqs[IO_IOPOLL_BATCH];
755 unsigned int free_reqs;
760 * Batch completion logic
762 struct io_comp_state comp;
765 * File reference cache
769 unsigned int file_refs;
770 unsigned int ios_left;
774 /* needs req->file assigned */
775 unsigned needs_file : 1;
776 /* don't fail if file grab fails */
777 unsigned needs_file_no_error : 1;
778 /* hash wq insertion if file is a regular file */
779 unsigned hash_reg_file : 1;
780 /* unbound wq insertion if file is a non-regular file */
781 unsigned unbound_nonreg_file : 1;
782 /* opcode is not supported by this kernel */
783 unsigned not_supported : 1;
784 /* set if opcode supports polled "wait" */
786 unsigned pollout : 1;
787 /* op supports buffer selection */
788 unsigned buffer_select : 1;
789 /* must always have async data allocated */
790 unsigned needs_async_data : 1;
791 /* should block plug */
793 /* size of async data needed, if any */
794 unsigned short async_size;
798 static const struct io_op_def io_op_defs[] = {
799 [IORING_OP_NOP] = {},
800 [IORING_OP_READV] = {
802 .unbound_nonreg_file = 1,
805 .needs_async_data = 1,
807 .async_size = sizeof(struct io_async_rw),
808 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
810 [IORING_OP_WRITEV] = {
813 .unbound_nonreg_file = 1,
815 .needs_async_data = 1,
817 .async_size = sizeof(struct io_async_rw),
818 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG |
821 [IORING_OP_FSYNC] = {
823 .work_flags = IO_WQ_WORK_BLKCG,
825 [IORING_OP_READ_FIXED] = {
827 .unbound_nonreg_file = 1,
830 .async_size = sizeof(struct io_async_rw),
831 .work_flags = IO_WQ_WORK_BLKCG | IO_WQ_WORK_MM,
833 [IORING_OP_WRITE_FIXED] = {
836 .unbound_nonreg_file = 1,
839 .async_size = sizeof(struct io_async_rw),
840 .work_flags = IO_WQ_WORK_BLKCG | IO_WQ_WORK_FSIZE |
843 [IORING_OP_POLL_ADD] = {
845 .unbound_nonreg_file = 1,
847 [IORING_OP_POLL_REMOVE] = {},
848 [IORING_OP_SYNC_FILE_RANGE] = {
850 .work_flags = IO_WQ_WORK_BLKCG,
852 [IORING_OP_SENDMSG] = {
854 .unbound_nonreg_file = 1,
856 .needs_async_data = 1,
857 .async_size = sizeof(struct io_async_msghdr),
858 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
860 [IORING_OP_RECVMSG] = {
862 .unbound_nonreg_file = 1,
865 .needs_async_data = 1,
866 .async_size = sizeof(struct io_async_msghdr),
867 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
869 [IORING_OP_TIMEOUT] = {
870 .needs_async_data = 1,
871 .async_size = sizeof(struct io_timeout_data),
872 .work_flags = IO_WQ_WORK_MM,
874 [IORING_OP_TIMEOUT_REMOVE] = {
875 /* used by timeout updates' prep() */
876 .work_flags = IO_WQ_WORK_MM,
878 [IORING_OP_ACCEPT] = {
880 .unbound_nonreg_file = 1,
882 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_FILES,
884 [IORING_OP_ASYNC_CANCEL] = {},
885 [IORING_OP_LINK_TIMEOUT] = {
886 .needs_async_data = 1,
887 .async_size = sizeof(struct io_timeout_data),
888 .work_flags = IO_WQ_WORK_MM,
890 [IORING_OP_CONNECT] = {
892 .unbound_nonreg_file = 1,
894 .needs_async_data = 1,
895 .async_size = sizeof(struct io_async_connect),
896 .work_flags = IO_WQ_WORK_MM,
898 [IORING_OP_FALLOCATE] = {
900 .work_flags = IO_WQ_WORK_BLKCG | IO_WQ_WORK_FSIZE,
902 [IORING_OP_OPENAT] = {
903 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_BLKCG |
904 IO_WQ_WORK_FS | IO_WQ_WORK_MM,
906 [IORING_OP_CLOSE] = {
908 .needs_file_no_error = 1,
909 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_BLKCG,
911 [IORING_OP_FILES_UPDATE] = {
912 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_MM,
914 [IORING_OP_STATX] = {
915 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_MM |
916 IO_WQ_WORK_FS | IO_WQ_WORK_BLKCG,
920 .unbound_nonreg_file = 1,
924 .async_size = sizeof(struct io_async_rw),
925 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
927 [IORING_OP_WRITE] = {
929 .unbound_nonreg_file = 1,
932 .async_size = sizeof(struct io_async_rw),
933 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG |
936 [IORING_OP_FADVISE] = {
938 .work_flags = IO_WQ_WORK_BLKCG,
940 [IORING_OP_MADVISE] = {
941 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
945 .unbound_nonreg_file = 1,
947 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
951 .unbound_nonreg_file = 1,
954 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_BLKCG,
956 [IORING_OP_OPENAT2] = {
957 .work_flags = IO_WQ_WORK_FILES | IO_WQ_WORK_FS |
958 IO_WQ_WORK_BLKCG | IO_WQ_WORK_MM,
960 [IORING_OP_EPOLL_CTL] = {
961 .unbound_nonreg_file = 1,
962 .work_flags = IO_WQ_WORK_FILES,
964 [IORING_OP_SPLICE] = {
967 .unbound_nonreg_file = 1,
968 .work_flags = IO_WQ_WORK_BLKCG,
970 [IORING_OP_PROVIDE_BUFFERS] = {},
971 [IORING_OP_REMOVE_BUFFERS] = {},
975 .unbound_nonreg_file = 1,
977 [IORING_OP_SHUTDOWN] = {
980 [IORING_OP_RENAMEAT] = {
981 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_FILES |
982 IO_WQ_WORK_FS | IO_WQ_WORK_BLKCG,
984 [IORING_OP_UNLINKAT] = {
985 .work_flags = IO_WQ_WORK_MM | IO_WQ_WORK_FILES |
986 IO_WQ_WORK_FS | IO_WQ_WORK_BLKCG,
990 enum io_mem_account {
995 static void destroy_fixed_file_ref_node(struct fixed_file_ref_node *ref_node);
996 static struct fixed_file_ref_node *alloc_fixed_file_ref_node(
997 struct io_ring_ctx *ctx);
999 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
1000 struct io_comp_state *cs);
1001 static void io_cqring_fill_event(struct io_kiocb *req, long res);
1002 static void io_put_req(struct io_kiocb *req);
1003 static void io_put_req_deferred(struct io_kiocb *req, int nr);
1004 static void io_double_put_req(struct io_kiocb *req);
1005 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
1006 static void __io_queue_linked_timeout(struct io_kiocb *req);
1007 static void io_queue_linked_timeout(struct io_kiocb *req);
1008 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
1009 struct io_uring_files_update *ip,
1011 static void __io_clean_op(struct io_kiocb *req);
1012 static struct file *io_file_get(struct io_submit_state *state,
1013 struct io_kiocb *req, int fd, bool fixed);
1014 static void __io_queue_sqe(struct io_kiocb *req, struct io_comp_state *cs);
1015 static void io_file_put_work(struct work_struct *work);
1017 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
1018 struct iovec **iovec, struct iov_iter *iter,
1020 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
1021 const struct iovec *fast_iov,
1022 struct iov_iter *iter, bool force);
1024 static struct kmem_cache *req_cachep;
1026 static const struct file_operations io_uring_fops;
1028 struct sock *io_uring_get_socket(struct file *file)
1030 #if defined(CONFIG_UNIX)
1031 if (file->f_op == &io_uring_fops) {
1032 struct io_ring_ctx *ctx = file->private_data;
1034 return ctx->ring_sock->sk;
1039 EXPORT_SYMBOL(io_uring_get_socket);
1041 #define io_for_each_link(pos, head) \
1042 for (pos = (head); pos; pos = pos->link)
1044 static inline void io_clean_op(struct io_kiocb *req)
1046 if (req->flags & (REQ_F_NEED_CLEANUP | REQ_F_BUFFER_SELECTED |
1051 static inline void io_set_resource_node(struct io_kiocb *req)
1053 struct io_ring_ctx *ctx = req->ctx;
1055 if (!req->fixed_file_refs) {
1056 req->fixed_file_refs = &ctx->file_data->node->refs;
1057 percpu_ref_get(req->fixed_file_refs);
1061 static bool io_match_task(struct io_kiocb *head,
1062 struct task_struct *task,
1063 struct files_struct *files)
1065 struct io_kiocb *req;
1067 if (task && head->task != task)
1072 io_for_each_link(req, head) {
1073 if ((req->flags & REQ_F_WORK_INITIALIZED) &&
1074 (req->work.flags & IO_WQ_WORK_FILES) &&
1075 req->work.identity->files == files)
1081 static void io_sq_thread_drop_mm_files(void)
1083 struct files_struct *files = current->files;
1084 struct mm_struct *mm = current->mm;
1087 kthread_unuse_mm(mm);
1092 struct nsproxy *nsproxy = current->nsproxy;
1095 current->files = NULL;
1096 current->nsproxy = NULL;
1097 task_unlock(current);
1098 put_files_struct(files);
1099 put_nsproxy(nsproxy);
1103 static int __io_sq_thread_acquire_files(struct io_ring_ctx *ctx)
1105 if (!current->files) {
1106 struct files_struct *files;
1107 struct nsproxy *nsproxy;
1109 task_lock(ctx->sqo_task);
1110 files = ctx->sqo_task->files;
1112 task_unlock(ctx->sqo_task);
1115 atomic_inc(&files->count);
1116 get_nsproxy(ctx->sqo_task->nsproxy);
1117 nsproxy = ctx->sqo_task->nsproxy;
1118 task_unlock(ctx->sqo_task);
1121 current->files = files;
1122 current->nsproxy = nsproxy;
1123 task_unlock(current);
1128 static int __io_sq_thread_acquire_mm(struct io_ring_ctx *ctx)
1130 struct mm_struct *mm;
1135 /* Should never happen */
1136 if (unlikely(!(ctx->flags & IORING_SETUP_SQPOLL)))
1139 task_lock(ctx->sqo_task);
1140 mm = ctx->sqo_task->mm;
1141 if (unlikely(!mm || !mmget_not_zero(mm)))
1143 task_unlock(ctx->sqo_task);
1153 static int io_sq_thread_acquire_mm_files(struct io_ring_ctx *ctx,
1154 struct io_kiocb *req)
1156 const struct io_op_def *def = &io_op_defs[req->opcode];
1159 if (def->work_flags & IO_WQ_WORK_MM) {
1160 ret = __io_sq_thread_acquire_mm(ctx);
1165 if (def->needs_file || (def->work_flags & IO_WQ_WORK_FILES)) {
1166 ret = __io_sq_thread_acquire_files(ctx);
1174 static void io_sq_thread_associate_blkcg(struct io_ring_ctx *ctx,
1175 struct cgroup_subsys_state **cur_css)
1178 #ifdef CONFIG_BLK_CGROUP
1179 /* puts the old one when swapping */
1180 if (*cur_css != ctx->sqo_blkcg_css) {
1181 kthread_associate_blkcg(ctx->sqo_blkcg_css);
1182 *cur_css = ctx->sqo_blkcg_css;
1187 static void io_sq_thread_unassociate_blkcg(void)
1189 #ifdef CONFIG_BLK_CGROUP
1190 kthread_associate_blkcg(NULL);
1194 static inline void req_set_fail_links(struct io_kiocb *req)
1196 if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
1197 req->flags |= REQ_F_FAIL_LINK;
1201 * None of these are dereferenced, they are simply used to check if any of
1202 * them have changed. If we're under current and check they are still the
1203 * same, we're fine to grab references to them for actual out-of-line use.
1205 static void io_init_identity(struct io_identity *id)
1207 id->files = current->files;
1208 id->mm = current->mm;
1209 #ifdef CONFIG_BLK_CGROUP
1211 id->blkcg_css = blkcg_css();
1214 id->creds = current_cred();
1215 id->nsproxy = current->nsproxy;
1216 id->fs = current->fs;
1217 id->fsize = rlimit(RLIMIT_FSIZE);
1219 id->loginuid = current->loginuid;
1220 id->sessionid = current->sessionid;
1222 refcount_set(&id->count, 1);
1225 static inline void __io_req_init_async(struct io_kiocb *req)
1227 memset(&req->work, 0, sizeof(req->work));
1228 req->flags |= REQ_F_WORK_INITIALIZED;
1232 * Note: must call io_req_init_async() for the first time you
1233 * touch any members of io_wq_work.
1235 static inline void io_req_init_async(struct io_kiocb *req)
1237 struct io_uring_task *tctx = current->io_uring;
1239 if (req->flags & REQ_F_WORK_INITIALIZED)
1242 __io_req_init_async(req);
1244 /* Grab a ref if this isn't our static identity */
1245 req->work.identity = tctx->identity;
1246 if (tctx->identity != &tctx->__identity)
1247 refcount_inc(&req->work.identity->count);
1250 static inline bool io_async_submit(struct io_ring_ctx *ctx)
1252 return ctx->flags & IORING_SETUP_SQPOLL;
1255 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1257 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1259 complete(&ctx->ref_comp);
1262 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1264 return !req->timeout.off;
1267 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1269 struct io_ring_ctx *ctx;
1272 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1276 ctx->fallback_req = kmem_cache_alloc(req_cachep, GFP_KERNEL);
1277 if (!ctx->fallback_req)
1281 * Use 5 bits less than the max cq entries, that should give us around
1282 * 32 entries per hash list if totally full and uniformly spread.
1284 hash_bits = ilog2(p->cq_entries);
1288 ctx->cancel_hash_bits = hash_bits;
1289 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1291 if (!ctx->cancel_hash)
1293 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1295 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1296 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1299 ctx->flags = p->flags;
1300 init_waitqueue_head(&ctx->sqo_sq_wait);
1301 INIT_LIST_HEAD(&ctx->sqd_list);
1302 init_waitqueue_head(&ctx->cq_wait);
1303 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1304 init_completion(&ctx->ref_comp);
1305 init_completion(&ctx->sq_thread_comp);
1306 idr_init(&ctx->io_buffer_idr);
1307 idr_init(&ctx->personality_idr);
1308 mutex_init(&ctx->uring_lock);
1309 init_waitqueue_head(&ctx->wait);
1310 spin_lock_init(&ctx->completion_lock);
1311 INIT_LIST_HEAD(&ctx->iopoll_list);
1312 INIT_LIST_HEAD(&ctx->defer_list);
1313 INIT_LIST_HEAD(&ctx->timeout_list);
1314 spin_lock_init(&ctx->inflight_lock);
1315 INIT_LIST_HEAD(&ctx->inflight_list);
1316 INIT_DELAYED_WORK(&ctx->file_put_work, io_file_put_work);
1317 init_llist_head(&ctx->file_put_llist);
1320 if (ctx->fallback_req)
1321 kmem_cache_free(req_cachep, ctx->fallback_req);
1322 kfree(ctx->cancel_hash);
1327 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1329 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1330 struct io_ring_ctx *ctx = req->ctx;
1332 return seq != ctx->cached_cq_tail
1333 + READ_ONCE(ctx->cached_cq_overflow);
1339 static void __io_commit_cqring(struct io_ring_ctx *ctx)
1341 struct io_rings *rings = ctx->rings;
1343 /* order cqe stores with ring update */
1344 smp_store_release(&rings->cq.tail, ctx->cached_cq_tail);
1346 if (wq_has_sleeper(&ctx->cq_wait)) {
1347 wake_up_interruptible(&ctx->cq_wait);
1348 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1352 static void io_put_identity(struct io_uring_task *tctx, struct io_kiocb *req)
1354 if (req->work.identity == &tctx->__identity)
1356 if (refcount_dec_and_test(&req->work.identity->count))
1357 kfree(req->work.identity);
1360 static void io_req_clean_work(struct io_kiocb *req)
1362 if (!(req->flags & REQ_F_WORK_INITIALIZED))
1365 req->flags &= ~REQ_F_WORK_INITIALIZED;
1367 if (req->work.flags & IO_WQ_WORK_MM) {
1368 mmdrop(req->work.identity->mm);
1369 req->work.flags &= ~IO_WQ_WORK_MM;
1371 #ifdef CONFIG_BLK_CGROUP
1372 if (req->work.flags & IO_WQ_WORK_BLKCG) {
1373 css_put(req->work.identity->blkcg_css);
1374 req->work.flags &= ~IO_WQ_WORK_BLKCG;
1377 if (req->work.flags & IO_WQ_WORK_CREDS) {
1378 put_cred(req->work.identity->creds);
1379 req->work.flags &= ~IO_WQ_WORK_CREDS;
1381 if (req->work.flags & IO_WQ_WORK_FS) {
1382 struct fs_struct *fs = req->work.identity->fs;
1384 spin_lock(&req->work.identity->fs->lock);
1387 spin_unlock(&req->work.identity->fs->lock);
1390 req->work.flags &= ~IO_WQ_WORK_FS;
1393 io_put_identity(req->task->io_uring, req);
1397 * Create a private copy of io_identity, since some fields don't match
1398 * the current context.
1400 static bool io_identity_cow(struct io_kiocb *req)
1402 struct io_uring_task *tctx = current->io_uring;
1403 const struct cred *creds = NULL;
1404 struct io_identity *id;
1406 if (req->work.flags & IO_WQ_WORK_CREDS)
1407 creds = req->work.identity->creds;
1409 id = kmemdup(req->work.identity, sizeof(*id), GFP_KERNEL);
1410 if (unlikely(!id)) {
1411 req->work.flags |= IO_WQ_WORK_CANCEL;
1416 * We can safely just re-init the creds we copied Either the field
1417 * matches the current one, or we haven't grabbed it yet. The only
1418 * exception is ->creds, through registered personalities, so handle
1419 * that one separately.
1421 io_init_identity(id);
1425 /* add one for this request */
1426 refcount_inc(&id->count);
1428 /* drop tctx and req identity references, if needed */
1429 if (tctx->identity != &tctx->__identity &&
1430 refcount_dec_and_test(&tctx->identity->count))
1431 kfree(tctx->identity);
1432 if (req->work.identity != &tctx->__identity &&
1433 refcount_dec_and_test(&req->work.identity->count))
1434 kfree(req->work.identity);
1436 req->work.identity = id;
1437 tctx->identity = id;
1441 static bool io_grab_identity(struct io_kiocb *req)
1443 const struct io_op_def *def = &io_op_defs[req->opcode];
1444 struct io_identity *id = req->work.identity;
1445 struct io_ring_ctx *ctx = req->ctx;
1447 if (def->work_flags & IO_WQ_WORK_FSIZE) {
1448 if (id->fsize != rlimit(RLIMIT_FSIZE))
1450 req->work.flags |= IO_WQ_WORK_FSIZE;
1452 #ifdef CONFIG_BLK_CGROUP
1453 if (!(req->work.flags & IO_WQ_WORK_BLKCG) &&
1454 (def->work_flags & IO_WQ_WORK_BLKCG)) {
1456 if (id->blkcg_css != blkcg_css()) {
1461 * This should be rare, either the cgroup is dying or the task
1462 * is moving cgroups. Just punt to root for the handful of ios.
1464 if (css_tryget_online(id->blkcg_css))
1465 req->work.flags |= IO_WQ_WORK_BLKCG;
1469 if (!(req->work.flags & IO_WQ_WORK_CREDS)) {
1470 if (id->creds != current_cred())
1472 get_cred(id->creds);
1473 req->work.flags |= IO_WQ_WORK_CREDS;
1476 if (!uid_eq(current->loginuid, id->loginuid) ||
1477 current->sessionid != id->sessionid)
1480 if (!(req->work.flags & IO_WQ_WORK_FS) &&
1481 (def->work_flags & IO_WQ_WORK_FS)) {
1482 if (current->fs != id->fs)
1484 spin_lock(&id->fs->lock);
1485 if (!id->fs->in_exec) {
1487 req->work.flags |= IO_WQ_WORK_FS;
1489 req->work.flags |= IO_WQ_WORK_CANCEL;
1491 spin_unlock(¤t->fs->lock);
1493 if (!(req->work.flags & IO_WQ_WORK_FILES) &&
1494 (def->work_flags & IO_WQ_WORK_FILES) &&
1495 !(req->flags & REQ_F_NO_FILE_TABLE)) {
1496 if (id->files != current->files ||
1497 id->nsproxy != current->nsproxy)
1499 atomic_inc(&id->files->count);
1500 get_nsproxy(id->nsproxy);
1501 req->flags |= REQ_F_INFLIGHT;
1503 spin_lock_irq(&ctx->inflight_lock);
1504 list_add(&req->inflight_entry, &ctx->inflight_list);
1505 spin_unlock_irq(&ctx->inflight_lock);
1506 req->work.flags |= IO_WQ_WORK_FILES;
1508 if (!(req->work.flags & IO_WQ_WORK_MM) &&
1509 (def->work_flags & IO_WQ_WORK_MM)) {
1510 if (id->mm != current->mm)
1513 req->work.flags |= IO_WQ_WORK_MM;
1519 static void io_prep_async_work(struct io_kiocb *req)
1521 const struct io_op_def *def = &io_op_defs[req->opcode];
1522 struct io_ring_ctx *ctx = req->ctx;
1523 struct io_identity *id;
1525 io_req_init_async(req);
1526 id = req->work.identity;
1528 if (req->flags & REQ_F_FORCE_ASYNC)
1529 req->work.flags |= IO_WQ_WORK_CONCURRENT;
1531 if (req->flags & REQ_F_ISREG) {
1532 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1533 io_wq_hash_work(&req->work, file_inode(req->file));
1535 if (def->unbound_nonreg_file)
1536 req->work.flags |= IO_WQ_WORK_UNBOUND;
1539 /* if we fail grabbing identity, we must COW, regrab, and retry */
1540 if (io_grab_identity(req))
1543 if (!io_identity_cow(req))
1546 /* can't fail at this point */
1547 if (!io_grab_identity(req))
1551 static void io_prep_async_link(struct io_kiocb *req)
1553 struct io_kiocb *cur;
1555 io_for_each_link(cur, req)
1556 io_prep_async_work(cur);
1559 static struct io_kiocb *__io_queue_async_work(struct io_kiocb *req)
1561 struct io_ring_ctx *ctx = req->ctx;
1562 struct io_kiocb *link = io_prep_linked_timeout(req);
1564 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1565 &req->work, req->flags);
1566 io_wq_enqueue(ctx->io_wq, &req->work);
1570 static void io_queue_async_work(struct io_kiocb *req)
1572 struct io_kiocb *link;
1574 /* init ->work of the whole link before punting */
1575 io_prep_async_link(req);
1576 link = __io_queue_async_work(req);
1579 io_queue_linked_timeout(link);
1582 static void io_kill_timeout(struct io_kiocb *req)
1584 struct io_timeout_data *io = req->async_data;
1587 ret = hrtimer_try_to_cancel(&io->timer);
1589 atomic_set(&req->ctx->cq_timeouts,
1590 atomic_read(&req->ctx->cq_timeouts) + 1);
1591 list_del_init(&req->timeout.list);
1592 io_cqring_fill_event(req, 0);
1593 io_put_req_deferred(req, 1);
1598 * Returns true if we found and killed one or more timeouts
1600 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
1601 struct files_struct *files)
1603 struct io_kiocb *req, *tmp;
1606 spin_lock_irq(&ctx->completion_lock);
1607 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
1608 if (io_match_task(req, tsk, files)) {
1609 io_kill_timeout(req);
1613 spin_unlock_irq(&ctx->completion_lock);
1614 return canceled != 0;
1617 static void __io_queue_deferred(struct io_ring_ctx *ctx)
1620 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1621 struct io_defer_entry, list);
1622 struct io_kiocb *link;
1624 if (req_need_defer(de->req, de->seq))
1626 list_del_init(&de->list);
1627 /* punt-init is done before queueing for defer */
1628 link = __io_queue_async_work(de->req);
1630 __io_queue_linked_timeout(link);
1631 /* drop submission reference */
1632 io_put_req_deferred(link, 1);
1635 } while (!list_empty(&ctx->defer_list));
1638 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1640 while (!list_empty(&ctx->timeout_list)) {
1641 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1642 struct io_kiocb, timeout.list);
1644 if (io_is_timeout_noseq(req))
1646 if (req->timeout.target_seq != ctx->cached_cq_tail
1647 - atomic_read(&ctx->cq_timeouts))
1650 list_del_init(&req->timeout.list);
1651 io_kill_timeout(req);
1655 static void io_commit_cqring(struct io_ring_ctx *ctx)
1657 io_flush_timeouts(ctx);
1658 __io_commit_cqring(ctx);
1660 if (unlikely(!list_empty(&ctx->defer_list)))
1661 __io_queue_deferred(ctx);
1664 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1666 struct io_rings *r = ctx->rings;
1668 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == r->sq_ring_entries;
1671 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1673 struct io_rings *rings = ctx->rings;
1676 tail = ctx->cached_cq_tail;
1678 * writes to the cq entry need to come after reading head; the
1679 * control dependency is enough as we're using WRITE_ONCE to
1682 if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries)
1685 ctx->cached_cq_tail++;
1686 return &rings->cqes[tail & ctx->cq_mask];
1689 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1693 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1695 if (!ctx->eventfd_async)
1697 return io_wq_current_is_worker();
1700 static inline unsigned __io_cqring_events(struct io_ring_ctx *ctx)
1702 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
1705 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1707 if (waitqueue_active(&ctx->wait))
1708 wake_up(&ctx->wait);
1709 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1710 wake_up(&ctx->sq_data->wait);
1711 if (io_should_trigger_evfd(ctx))
1712 eventfd_signal(ctx->cq_ev_fd, 1);
1715 /* Returns true if there are no backlogged entries after the flush */
1716 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force,
1717 struct task_struct *tsk,
1718 struct files_struct *files)
1720 struct io_rings *rings = ctx->rings;
1721 struct io_kiocb *req, *tmp;
1722 struct io_uring_cqe *cqe;
1723 unsigned long flags;
1727 if (!force && __io_cqring_events(ctx) == rings->cq_ring_entries)
1730 spin_lock_irqsave(&ctx->completion_lock, flags);
1731 list_for_each_entry_safe(req, tmp, &ctx->cq_overflow_list, compl.list) {
1732 if (!io_match_task(req, tsk, files))
1735 cqe = io_get_cqring(ctx);
1739 list_move(&req->compl.list, &list);
1741 WRITE_ONCE(cqe->user_data, req->user_data);
1742 WRITE_ONCE(cqe->res, req->result);
1743 WRITE_ONCE(cqe->flags, req->compl.cflags);
1745 ctx->cached_cq_overflow++;
1746 WRITE_ONCE(ctx->rings->cq_overflow,
1747 ctx->cached_cq_overflow);
1751 all_flushed = list_empty(&ctx->cq_overflow_list);
1753 clear_bit(0, &ctx->sq_check_overflow);
1754 clear_bit(0, &ctx->cq_check_overflow);
1755 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
1758 io_commit_cqring(ctx);
1759 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1760 io_cqring_ev_posted(ctx);
1762 while (!list_empty(&list)) {
1763 req = list_first_entry(&list, struct io_kiocb, compl.list);
1764 list_del(&req->compl.list);
1771 static void __io_cqring_fill_event(struct io_kiocb *req, long res, long cflags)
1773 struct io_ring_ctx *ctx = req->ctx;
1774 struct io_uring_cqe *cqe;
1776 trace_io_uring_complete(ctx, req->user_data, res);
1779 * If we can't get a cq entry, userspace overflowed the
1780 * submission (by quite a lot). Increment the overflow count in
1783 cqe = io_get_cqring(ctx);
1785 WRITE_ONCE(cqe->user_data, req->user_data);
1786 WRITE_ONCE(cqe->res, res);
1787 WRITE_ONCE(cqe->flags, cflags);
1788 } else if (ctx->cq_overflow_flushed ||
1789 atomic_read(&req->task->io_uring->in_idle)) {
1791 * If we're in ring overflow flush mode, or in task cancel mode,
1792 * then we cannot store the request for later flushing, we need
1793 * to drop it on the floor.
1795 ctx->cached_cq_overflow++;
1796 WRITE_ONCE(ctx->rings->cq_overflow, ctx->cached_cq_overflow);
1798 if (list_empty(&ctx->cq_overflow_list)) {
1799 set_bit(0, &ctx->sq_check_overflow);
1800 set_bit(0, &ctx->cq_check_overflow);
1801 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
1805 req->compl.cflags = cflags;
1806 refcount_inc(&req->refs);
1807 list_add_tail(&req->compl.list, &ctx->cq_overflow_list);
1811 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1813 __io_cqring_fill_event(req, res, 0);
1816 static void io_cqring_add_event(struct io_kiocb *req, long res, long cflags)
1818 struct io_ring_ctx *ctx = req->ctx;
1819 unsigned long flags;
1821 spin_lock_irqsave(&ctx->completion_lock, flags);
1822 __io_cqring_fill_event(req, res, cflags);
1823 io_commit_cqring(ctx);
1824 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1826 io_cqring_ev_posted(ctx);
1829 static void io_submit_flush_completions(struct io_comp_state *cs)
1831 struct io_ring_ctx *ctx = cs->ctx;
1833 spin_lock_irq(&ctx->completion_lock);
1834 while (!list_empty(&cs->list)) {
1835 struct io_kiocb *req;
1837 req = list_first_entry(&cs->list, struct io_kiocb, compl.list);
1838 list_del(&req->compl.list);
1839 __io_cqring_fill_event(req, req->result, req->compl.cflags);
1842 * io_free_req() doesn't care about completion_lock unless one
1843 * of these flags is set. REQ_F_WORK_INITIALIZED is in the list
1844 * because of a potential deadlock with req->work.fs->lock
1846 if (req->flags & (REQ_F_FAIL_LINK|REQ_F_LINK_TIMEOUT
1847 |REQ_F_WORK_INITIALIZED)) {
1848 spin_unlock_irq(&ctx->completion_lock);
1850 spin_lock_irq(&ctx->completion_lock);
1855 io_commit_cqring(ctx);
1856 spin_unlock_irq(&ctx->completion_lock);
1858 io_cqring_ev_posted(ctx);
1862 static void __io_req_complete(struct io_kiocb *req, long res, unsigned cflags,
1863 struct io_comp_state *cs)
1866 io_cqring_add_event(req, res, cflags);
1871 req->compl.cflags = cflags;
1872 list_add_tail(&req->compl.list, &cs->list);
1874 io_submit_flush_completions(cs);
1878 static void io_req_complete(struct io_kiocb *req, long res)
1880 __io_req_complete(req, res, 0, NULL);
1883 static inline bool io_is_fallback_req(struct io_kiocb *req)
1885 return req == (struct io_kiocb *)
1886 ((unsigned long) req->ctx->fallback_req & ~1UL);
1889 static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx)
1891 struct io_kiocb *req;
1893 req = ctx->fallback_req;
1894 if (!test_and_set_bit_lock(0, (unsigned long *) &ctx->fallback_req))
1900 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx,
1901 struct io_submit_state *state)
1903 if (!state->free_reqs) {
1904 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1908 sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs));
1909 ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs);
1912 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1913 * retry single alloc to be on the safe side.
1915 if (unlikely(ret <= 0)) {
1916 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1917 if (!state->reqs[0])
1921 state->free_reqs = ret;
1925 return state->reqs[state->free_reqs];
1927 return io_get_fallback_req(ctx);
1930 static inline void io_put_file(struct io_kiocb *req, struct file *file,
1937 static void io_dismantle_req(struct io_kiocb *req)
1941 if (req->async_data)
1942 kfree(req->async_data);
1944 io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
1945 if (req->fixed_file_refs)
1946 percpu_ref_put(req->fixed_file_refs);
1947 io_req_clean_work(req);
1950 static void __io_free_req(struct io_kiocb *req)
1952 struct io_uring_task *tctx = req->task->io_uring;
1953 struct io_ring_ctx *ctx = req->ctx;
1955 io_dismantle_req(req);
1957 percpu_counter_dec(&tctx->inflight);
1958 if (atomic_read(&tctx->in_idle))
1959 wake_up(&tctx->wait);
1960 put_task_struct(req->task);
1962 if (likely(!io_is_fallback_req(req)))
1963 kmem_cache_free(req_cachep, req);
1965 clear_bit_unlock(0, (unsigned long *) &ctx->fallback_req);
1966 percpu_ref_put(&ctx->refs);
1969 static inline void io_remove_next_linked(struct io_kiocb *req)
1971 struct io_kiocb *nxt = req->link;
1973 req->link = nxt->link;
1977 static void io_kill_linked_timeout(struct io_kiocb *req)
1979 struct io_ring_ctx *ctx = req->ctx;
1980 struct io_kiocb *link;
1981 bool cancelled = false;
1982 unsigned long flags;
1984 spin_lock_irqsave(&ctx->completion_lock, flags);
1988 * Can happen if a linked timeout fired and link had been like
1989 * req -> link t-out -> link t-out [-> ...]
1991 if (link && (link->flags & REQ_F_LTIMEOUT_ACTIVE)) {
1992 struct io_timeout_data *io = link->async_data;
1995 io_remove_next_linked(req);
1996 link->timeout.head = NULL;
1997 ret = hrtimer_try_to_cancel(&io->timer);
1999 io_cqring_fill_event(link, -ECANCELED);
2000 io_commit_cqring(ctx);
2004 req->flags &= ~REQ_F_LINK_TIMEOUT;
2005 spin_unlock_irqrestore(&ctx->completion_lock, flags);
2008 io_cqring_ev_posted(ctx);
2014 static void io_fail_links(struct io_kiocb *req)
2016 struct io_kiocb *link, *nxt;
2017 struct io_ring_ctx *ctx = req->ctx;
2018 unsigned long flags;
2020 spin_lock_irqsave(&ctx->completion_lock, flags);
2028 trace_io_uring_fail_link(req, link);
2029 io_cqring_fill_event(link, -ECANCELED);
2032 * It's ok to free under spinlock as they're not linked anymore,
2033 * but avoid REQ_F_WORK_INITIALIZED because it may deadlock on
2036 if (link->flags & REQ_F_WORK_INITIALIZED)
2037 io_put_req_deferred(link, 2);
2039 io_double_put_req(link);
2042 io_commit_cqring(ctx);
2043 spin_unlock_irqrestore(&ctx->completion_lock, flags);
2045 io_cqring_ev_posted(ctx);
2048 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
2050 if (req->flags & REQ_F_LINK_TIMEOUT)
2051 io_kill_linked_timeout(req);
2054 * If LINK is set, we have dependent requests in this chain. If we
2055 * didn't fail this request, queue the first one up, moving any other
2056 * dependencies to the next request. In case of failure, fail the rest
2059 if (likely(!(req->flags & REQ_F_FAIL_LINK))) {
2060 struct io_kiocb *nxt = req->link;
2069 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
2071 if (likely(!(req->link) && !(req->flags & REQ_F_LINK_TIMEOUT)))
2073 return __io_req_find_next(req);
2076 static int io_req_task_work_add(struct io_kiocb *req)
2078 struct task_struct *tsk = req->task;
2079 struct io_ring_ctx *ctx = req->ctx;
2080 enum task_work_notify_mode notify;
2083 if (tsk->flags & PF_EXITING)
2087 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
2088 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
2089 * processing task_work. There's no reliable way to tell if TWA_RESUME
2093 if (!(ctx->flags & IORING_SETUP_SQPOLL))
2094 notify = TWA_SIGNAL;
2096 ret = task_work_add(tsk, &req->task_work, notify);
2098 wake_up_process(tsk);
2103 static void __io_req_task_cancel(struct io_kiocb *req, int error)
2105 struct io_ring_ctx *ctx = req->ctx;
2107 spin_lock_irq(&ctx->completion_lock);
2108 io_cqring_fill_event(req, error);
2109 io_commit_cqring(ctx);
2110 spin_unlock_irq(&ctx->completion_lock);
2112 io_cqring_ev_posted(ctx);
2113 req_set_fail_links(req);
2114 io_double_put_req(req);
2117 static void io_req_task_cancel(struct callback_head *cb)
2119 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2120 struct io_ring_ctx *ctx = req->ctx;
2122 __io_req_task_cancel(req, -ECANCELED);
2123 percpu_ref_put(&ctx->refs);
2126 static void __io_req_task_submit(struct io_kiocb *req)
2128 struct io_ring_ctx *ctx = req->ctx;
2131 fail = __io_sq_thread_acquire_mm(ctx) ||
2132 __io_sq_thread_acquire_files(ctx);
2133 mutex_lock(&ctx->uring_lock);
2135 __io_queue_sqe(req, NULL);
2137 __io_req_task_cancel(req, -EFAULT);
2138 mutex_unlock(&ctx->uring_lock);
2141 static void io_req_task_submit(struct callback_head *cb)
2143 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2144 struct io_ring_ctx *ctx = req->ctx;
2146 __io_req_task_submit(req);
2147 percpu_ref_put(&ctx->refs);
2150 static void io_req_task_queue(struct io_kiocb *req)
2154 init_task_work(&req->task_work, io_req_task_submit);
2155 percpu_ref_get(&req->ctx->refs);
2157 ret = io_req_task_work_add(req);
2158 if (unlikely(ret)) {
2159 struct task_struct *tsk;
2161 init_task_work(&req->task_work, io_req_task_cancel);
2162 tsk = io_wq_get_task(req->ctx->io_wq);
2163 task_work_add(tsk, &req->task_work, TWA_NONE);
2164 wake_up_process(tsk);
2168 static inline void io_queue_next(struct io_kiocb *req)
2170 struct io_kiocb *nxt = io_req_find_next(req);
2173 io_req_task_queue(nxt);
2176 static void io_free_req(struct io_kiocb *req)
2183 void *reqs[IO_IOPOLL_BATCH];
2186 struct task_struct *task;
2190 static inline void io_init_req_batch(struct req_batch *rb)
2197 static void __io_req_free_batch_flush(struct io_ring_ctx *ctx,
2198 struct req_batch *rb)
2200 kmem_cache_free_bulk(req_cachep, rb->to_free, rb->reqs);
2201 percpu_ref_put_many(&ctx->refs, rb->to_free);
2205 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2206 struct req_batch *rb)
2209 __io_req_free_batch_flush(ctx, rb);
2211 struct io_uring_task *tctx = rb->task->io_uring;
2213 percpu_counter_sub(&tctx->inflight, rb->task_refs);
2214 put_task_struct_many(rb->task, rb->task_refs);
2219 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req)
2221 if (unlikely(io_is_fallback_req(req))) {
2227 if (req->task != rb->task) {
2229 struct io_uring_task *tctx = rb->task->io_uring;
2231 percpu_counter_sub(&tctx->inflight, rb->task_refs);
2232 put_task_struct_many(rb->task, rb->task_refs);
2234 rb->task = req->task;
2239 io_dismantle_req(req);
2240 rb->reqs[rb->to_free++] = req;
2241 if (unlikely(rb->to_free == ARRAY_SIZE(rb->reqs)))
2242 __io_req_free_batch_flush(req->ctx, rb);
2246 * Drop reference to request, return next in chain (if there is one) if this
2247 * was the last reference to this request.
2249 static struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2251 struct io_kiocb *nxt = NULL;
2253 if (refcount_dec_and_test(&req->refs)) {
2254 nxt = io_req_find_next(req);
2260 static void io_put_req(struct io_kiocb *req)
2262 if (refcount_dec_and_test(&req->refs))
2266 static void io_put_req_deferred_cb(struct callback_head *cb)
2268 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2273 static void io_free_req_deferred(struct io_kiocb *req)
2277 init_task_work(&req->task_work, io_put_req_deferred_cb);
2278 ret = io_req_task_work_add(req);
2279 if (unlikely(ret)) {
2280 struct task_struct *tsk;
2282 tsk = io_wq_get_task(req->ctx->io_wq);
2283 task_work_add(tsk, &req->task_work, TWA_NONE);
2284 wake_up_process(tsk);
2288 static inline void io_put_req_deferred(struct io_kiocb *req, int refs)
2290 if (refcount_sub_and_test(refs, &req->refs))
2291 io_free_req_deferred(req);
2294 static struct io_wq_work *io_steal_work(struct io_kiocb *req)
2296 struct io_kiocb *nxt;
2299 * A ref is owned by io-wq in which context we're. So, if that's the
2300 * last one, it's safe to steal next work. False negatives are Ok,
2301 * it just will be re-punted async in io_put_work()
2303 if (refcount_read(&req->refs) != 1)
2306 nxt = io_req_find_next(req);
2307 return nxt ? &nxt->work : NULL;
2310 static void io_double_put_req(struct io_kiocb *req)
2312 /* drop both submit and complete references */
2313 if (refcount_sub_and_test(2, &req->refs))
2317 static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush)
2319 if (test_bit(0, &ctx->cq_check_overflow)) {
2321 * noflush == true is from the waitqueue handler, just ensure
2322 * we wake up the task, and the next invocation will flush the
2323 * entries. We cannot safely to it from here.
2328 io_cqring_overflow_flush(ctx, false, NULL, NULL);
2331 /* See comment at the top of this file */
2333 return __io_cqring_events(ctx);
2336 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2338 struct io_rings *rings = ctx->rings;
2340 /* make sure SQ entry isn't read before tail */
2341 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2344 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2346 unsigned int cflags;
2348 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2349 cflags |= IORING_CQE_F_BUFFER;
2350 req->flags &= ~REQ_F_BUFFER_SELECTED;
2355 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2357 struct io_buffer *kbuf;
2359 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2360 return io_put_kbuf(req, kbuf);
2363 static inline bool io_run_task_work(void)
2366 * Not safe to run on exiting task, and the task_work handling will
2367 * not add work to such a task.
2369 if (unlikely(current->flags & PF_EXITING))
2371 if (current->task_works) {
2372 __set_current_state(TASK_RUNNING);
2380 static void io_iopoll_queue(struct list_head *again)
2382 struct io_kiocb *req;
2385 req = list_first_entry(again, struct io_kiocb, inflight_entry);
2386 list_del(&req->inflight_entry);
2387 __io_complete_rw(req, -EAGAIN, 0, NULL);
2388 } while (!list_empty(again));
2392 * Find and free completed poll iocbs
2394 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2395 struct list_head *done)
2397 struct req_batch rb;
2398 struct io_kiocb *req;
2401 /* order with ->result store in io_complete_rw_iopoll() */
2404 io_init_req_batch(&rb);
2405 while (!list_empty(done)) {
2408 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2409 if (READ_ONCE(req->result) == -EAGAIN) {
2411 req->iopoll_completed = 0;
2412 list_move_tail(&req->inflight_entry, &again);
2415 list_del(&req->inflight_entry);
2417 if (req->flags & REQ_F_BUFFER_SELECTED)
2418 cflags = io_put_rw_kbuf(req);
2420 __io_cqring_fill_event(req, req->result, cflags);
2423 if (refcount_dec_and_test(&req->refs))
2424 io_req_free_batch(&rb, req);
2427 io_commit_cqring(ctx);
2428 if (ctx->flags & IORING_SETUP_SQPOLL)
2429 io_cqring_ev_posted(ctx);
2430 io_req_free_batch_finish(ctx, &rb);
2432 if (!list_empty(&again))
2433 io_iopoll_queue(&again);
2436 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2439 struct io_kiocb *req, *tmp;
2445 * Only spin for completions if we don't have multiple devices hanging
2446 * off our complete list, and we're under the requested amount.
2448 spin = !ctx->poll_multi_file && *nr_events < min;
2451 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2452 struct kiocb *kiocb = &req->rw.kiocb;
2455 * Move completed and retryable entries to our local lists.
2456 * If we find a request that requires polling, break out
2457 * and complete those lists first, if we have entries there.
2459 if (READ_ONCE(req->iopoll_completed)) {
2460 list_move_tail(&req->inflight_entry, &done);
2463 if (!list_empty(&done))
2466 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2470 /* iopoll may have completed current req */
2471 if (READ_ONCE(req->iopoll_completed))
2472 list_move_tail(&req->inflight_entry, &done);
2479 if (!list_empty(&done))
2480 io_iopoll_complete(ctx, nr_events, &done);
2486 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2487 * non-spinning poll check - we'll still enter the driver poll loop, but only
2488 * as a non-spinning completion check.
2490 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
2493 while (!list_empty(&ctx->iopoll_list) && !need_resched()) {
2496 ret = io_do_iopoll(ctx, nr_events, min);
2499 if (*nr_events >= min)
2507 * We can't just wait for polled events to come to us, we have to actively
2508 * find and complete them.
2510 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2512 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2515 mutex_lock(&ctx->uring_lock);
2516 while (!list_empty(&ctx->iopoll_list)) {
2517 unsigned int nr_events = 0;
2519 io_do_iopoll(ctx, &nr_events, 0);
2521 /* let it sleep and repeat later if can't complete a request */
2525 * Ensure we allow local-to-the-cpu processing to take place,
2526 * in this case we need to ensure that we reap all events.
2527 * Also let task_work, etc. to progress by releasing the mutex
2529 if (need_resched()) {
2530 mutex_unlock(&ctx->uring_lock);
2532 mutex_lock(&ctx->uring_lock);
2535 mutex_unlock(&ctx->uring_lock);
2538 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2540 unsigned int nr_events = 0;
2541 int iters = 0, ret = 0;
2544 * We disallow the app entering submit/complete with polling, but we
2545 * still need to lock the ring to prevent racing with polled issue
2546 * that got punted to a workqueue.
2548 mutex_lock(&ctx->uring_lock);
2551 * Don't enter poll loop if we already have events pending.
2552 * If we do, we can potentially be spinning for commands that
2553 * already triggered a CQE (eg in error).
2555 if (io_cqring_events(ctx, false))
2559 * If a submit got punted to a workqueue, we can have the
2560 * application entering polling for a command before it gets
2561 * issued. That app will hold the uring_lock for the duration
2562 * of the poll right here, so we need to take a breather every
2563 * now and then to ensure that the issue has a chance to add
2564 * the poll to the issued list. Otherwise we can spin here
2565 * forever, while the workqueue is stuck trying to acquire the
2568 if (!(++iters & 7)) {
2569 mutex_unlock(&ctx->uring_lock);
2571 mutex_lock(&ctx->uring_lock);
2574 ret = io_iopoll_getevents(ctx, &nr_events, min);
2578 } while (min && !nr_events && !need_resched());
2580 mutex_unlock(&ctx->uring_lock);
2584 static void kiocb_end_write(struct io_kiocb *req)
2587 * Tell lockdep we inherited freeze protection from submission
2590 if (req->flags & REQ_F_ISREG) {
2591 struct inode *inode = file_inode(req->file);
2593 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
2595 file_end_write(req->file);
2598 static void io_complete_rw_common(struct kiocb *kiocb, long res,
2599 struct io_comp_state *cs)
2601 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2604 if (kiocb->ki_flags & IOCB_WRITE)
2605 kiocb_end_write(req);
2607 if (res != req->result)
2608 req_set_fail_links(req);
2609 if (req->flags & REQ_F_BUFFER_SELECTED)
2610 cflags = io_put_rw_kbuf(req);
2611 __io_req_complete(req, res, cflags, cs);
2615 static bool io_resubmit_prep(struct io_kiocb *req, int error)
2617 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2618 ssize_t ret = -ECANCELED;
2619 struct iov_iter iter;
2627 switch (req->opcode) {
2628 case IORING_OP_READV:
2629 case IORING_OP_READ_FIXED:
2630 case IORING_OP_READ:
2633 case IORING_OP_WRITEV:
2634 case IORING_OP_WRITE_FIXED:
2635 case IORING_OP_WRITE:
2639 printk_once(KERN_WARNING "io_uring: bad opcode in resubmit %d\n",
2644 if (!req->async_data) {
2645 ret = io_import_iovec(rw, req, &iovec, &iter, false);
2648 ret = io_setup_async_rw(req, iovec, inline_vecs, &iter, false);
2656 req_set_fail_links(req);
2661 static bool io_rw_reissue(struct io_kiocb *req, long res)
2664 umode_t mode = file_inode(req->file)->i_mode;
2667 if (!S_ISBLK(mode) && !S_ISREG(mode))
2669 if ((res != -EAGAIN && res != -EOPNOTSUPP) || io_wq_current_is_worker())
2672 ret = io_sq_thread_acquire_mm_files(req->ctx, req);
2674 if (io_resubmit_prep(req, ret)) {
2675 refcount_inc(&req->refs);
2676 io_queue_async_work(req);
2684 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2685 struct io_comp_state *cs)
2687 if (!io_rw_reissue(req, res))
2688 io_complete_rw_common(&req->rw.kiocb, res, cs);
2691 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2693 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2695 __io_complete_rw(req, res, res2, NULL);
2698 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2700 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2702 if (kiocb->ki_flags & IOCB_WRITE)
2703 kiocb_end_write(req);
2705 if (res != -EAGAIN && res != req->result)
2706 req_set_fail_links(req);
2708 WRITE_ONCE(req->result, res);
2709 /* order with io_poll_complete() checking ->result */
2711 WRITE_ONCE(req->iopoll_completed, 1);
2715 * After the iocb has been issued, it's safe to be found on the poll list.
2716 * Adding the kiocb to the list AFTER submission ensures that we don't
2717 * find it from a io_iopoll_getevents() thread before the issuer is done
2718 * accessing the kiocb cookie.
2720 static void io_iopoll_req_issued(struct io_kiocb *req, bool in_async)
2722 struct io_ring_ctx *ctx = req->ctx;
2725 * Track whether we have multiple files in our lists. This will impact
2726 * how we do polling eventually, not spinning if we're on potentially
2727 * different devices.
2729 if (list_empty(&ctx->iopoll_list)) {
2730 ctx->poll_multi_file = false;
2731 } else if (!ctx->poll_multi_file) {
2732 struct io_kiocb *list_req;
2734 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2736 if (list_req->file != req->file)
2737 ctx->poll_multi_file = true;
2741 * For fast devices, IO may have already completed. If it has, add
2742 * it to the front so we find it first.
2744 if (READ_ONCE(req->iopoll_completed))
2745 list_add(&req->inflight_entry, &ctx->iopoll_list);
2747 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2750 * If IORING_SETUP_SQPOLL is enabled, sqes are either handled in sq thread
2751 * task context or in io worker task context. If current task context is
2752 * sq thread, we don't need to check whether should wake up sq thread.
2754 if (in_async && (ctx->flags & IORING_SETUP_SQPOLL) &&
2755 wq_has_sleeper(&ctx->sq_data->wait))
2756 wake_up(&ctx->sq_data->wait);
2759 static inline void __io_state_file_put(struct io_submit_state *state)
2761 fput_many(state->file, state->file_refs);
2762 state->file_refs = 0;
2765 static inline void io_state_file_put(struct io_submit_state *state)
2767 if (state->file_refs)
2768 __io_state_file_put(state);
2772 * Get as many references to a file as we have IOs left in this submission,
2773 * assuming most submissions are for one file, or at least that each file
2774 * has more than one submission.
2776 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2781 if (state->file_refs) {
2782 if (state->fd == fd) {
2786 __io_state_file_put(state);
2788 state->file = fget_many(fd, state->ios_left);
2789 if (unlikely(!state->file))
2793 state->file_refs = state->ios_left - 1;
2797 static bool io_bdev_nowait(struct block_device *bdev)
2799 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2803 * If we tracked the file through the SCM inflight mechanism, we could support
2804 * any file. For now, just ensure that anything potentially problematic is done
2807 static bool io_file_supports_async(struct file *file, int rw)
2809 umode_t mode = file_inode(file)->i_mode;
2811 if (S_ISBLK(mode)) {
2812 if (IS_ENABLED(CONFIG_BLOCK) &&
2813 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2817 if (S_ISCHR(mode) || S_ISSOCK(mode))
2819 if (S_ISREG(mode)) {
2820 if (IS_ENABLED(CONFIG_BLOCK) &&
2821 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2822 file->f_op != &io_uring_fops)
2827 /* any ->read/write should understand O_NONBLOCK */
2828 if (file->f_flags & O_NONBLOCK)
2831 if (!(file->f_mode & FMODE_NOWAIT))
2835 return file->f_op->read_iter != NULL;
2837 return file->f_op->write_iter != NULL;
2840 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2842 struct io_ring_ctx *ctx = req->ctx;
2843 struct kiocb *kiocb = &req->rw.kiocb;
2847 if (S_ISREG(file_inode(req->file)->i_mode))
2848 req->flags |= REQ_F_ISREG;
2850 kiocb->ki_pos = READ_ONCE(sqe->off);
2851 if (kiocb->ki_pos == -1 && !(req->file->f_mode & FMODE_STREAM)) {
2852 req->flags |= REQ_F_CUR_POS;
2853 kiocb->ki_pos = req->file->f_pos;
2855 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2856 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2857 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2861 ioprio = READ_ONCE(sqe->ioprio);
2863 ret = ioprio_check_cap(ioprio);
2867 kiocb->ki_ioprio = ioprio;
2869 kiocb->ki_ioprio = get_current_ioprio();
2871 /* don't allow async punt if RWF_NOWAIT was requested */
2872 if (kiocb->ki_flags & IOCB_NOWAIT)
2873 req->flags |= REQ_F_NOWAIT;
2875 if (ctx->flags & IORING_SETUP_IOPOLL) {
2876 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2877 !kiocb->ki_filp->f_op->iopoll)
2880 kiocb->ki_flags |= IOCB_HIPRI;
2881 kiocb->ki_complete = io_complete_rw_iopoll;
2882 req->iopoll_completed = 0;
2884 if (kiocb->ki_flags & IOCB_HIPRI)
2886 kiocb->ki_complete = io_complete_rw;
2889 req->rw.addr = READ_ONCE(sqe->addr);
2890 req->rw.len = READ_ONCE(sqe->len);
2891 req->buf_index = READ_ONCE(sqe->buf_index);
2895 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2901 case -ERESTARTNOINTR:
2902 case -ERESTARTNOHAND:
2903 case -ERESTART_RESTARTBLOCK:
2905 * We can't just restart the syscall, since previously
2906 * submitted sqes may already be in progress. Just fail this
2912 kiocb->ki_complete(kiocb, ret, 0);
2916 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2917 struct io_comp_state *cs)
2919 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2920 struct io_async_rw *io = req->async_data;
2922 /* add previously done IO, if any */
2923 if (io && io->bytes_done > 0) {
2925 ret = io->bytes_done;
2927 ret += io->bytes_done;
2930 if (req->flags & REQ_F_CUR_POS)
2931 req->file->f_pos = kiocb->ki_pos;
2932 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2933 __io_complete_rw(req, ret, 0, cs);
2935 io_rw_done(kiocb, ret);
2938 static ssize_t io_import_fixed(struct io_kiocb *req, int rw,
2939 struct iov_iter *iter)
2941 struct io_ring_ctx *ctx = req->ctx;
2942 size_t len = req->rw.len;
2943 struct io_mapped_ubuf *imu;
2944 u16 index, buf_index = req->buf_index;
2948 if (unlikely(buf_index >= ctx->nr_user_bufs))
2950 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2951 imu = &ctx->user_bufs[index];
2952 buf_addr = req->rw.addr;
2955 if (buf_addr + len < buf_addr)
2957 /* not inside the mapped region */
2958 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2962 * May not be a start of buffer, set size appropriately
2963 * and advance us to the beginning.
2965 offset = buf_addr - imu->ubuf;
2966 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2970 * Don't use iov_iter_advance() here, as it's really slow for
2971 * using the latter parts of a big fixed buffer - it iterates
2972 * over each segment manually. We can cheat a bit here, because
2975 * 1) it's a BVEC iter, we set it up
2976 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2977 * first and last bvec
2979 * So just find our index, and adjust the iterator afterwards.
2980 * If the offset is within the first bvec (or the whole first
2981 * bvec, just use iov_iter_advance(). This makes it easier
2982 * since we can just skip the first segment, which may not
2983 * be PAGE_SIZE aligned.
2985 const struct bio_vec *bvec = imu->bvec;
2987 if (offset <= bvec->bv_len) {
2988 iov_iter_advance(iter, offset);
2990 unsigned long seg_skip;
2992 /* skip first vec */
2993 offset -= bvec->bv_len;
2994 seg_skip = 1 + (offset >> PAGE_SHIFT);
2996 iter->bvec = bvec + seg_skip;
2997 iter->nr_segs -= seg_skip;
2998 iter->count -= bvec->bv_len + offset;
2999 iter->iov_offset = offset & ~PAGE_MASK;
3006 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
3009 mutex_unlock(&ctx->uring_lock);
3012 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
3015 * "Normal" inline submissions always hold the uring_lock, since we
3016 * grab it from the system call. Same is true for the SQPOLL offload.
3017 * The only exception is when we've detached the request and issue it
3018 * from an async worker thread, grab the lock for that case.
3021 mutex_lock(&ctx->uring_lock);
3024 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
3025 int bgid, struct io_buffer *kbuf,
3028 struct io_buffer *head;
3030 if (req->flags & REQ_F_BUFFER_SELECTED)
3033 io_ring_submit_lock(req->ctx, needs_lock);
3035 lockdep_assert_held(&req->ctx->uring_lock);
3037 head = idr_find(&req->ctx->io_buffer_idr, bgid);
3039 if (!list_empty(&head->list)) {
3040 kbuf = list_last_entry(&head->list, struct io_buffer,
3042 list_del(&kbuf->list);
3045 idr_remove(&req->ctx->io_buffer_idr, bgid);
3047 if (*len > kbuf->len)
3050 kbuf = ERR_PTR(-ENOBUFS);
3053 io_ring_submit_unlock(req->ctx, needs_lock);
3058 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
3061 struct io_buffer *kbuf;
3064 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3065 bgid = req->buf_index;
3066 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
3069 req->rw.addr = (u64) (unsigned long) kbuf;
3070 req->flags |= REQ_F_BUFFER_SELECTED;
3071 return u64_to_user_ptr(kbuf->addr);
3074 #ifdef CONFIG_COMPAT
3075 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
3078 struct compat_iovec __user *uiov;
3079 compat_ssize_t clen;
3083 uiov = u64_to_user_ptr(req->rw.addr);
3084 if (!access_ok(uiov, sizeof(*uiov)))
3086 if (__get_user(clen, &uiov->iov_len))
3092 buf = io_rw_buffer_select(req, &len, needs_lock);
3094 return PTR_ERR(buf);
3095 iov[0].iov_base = buf;
3096 iov[0].iov_len = (compat_size_t) len;
3101 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3104 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
3108 if (copy_from_user(iov, uiov, sizeof(*uiov)))
3111 len = iov[0].iov_len;
3114 buf = io_rw_buffer_select(req, &len, needs_lock);
3116 return PTR_ERR(buf);
3117 iov[0].iov_base = buf;
3118 iov[0].iov_len = len;
3122 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3125 if (req->flags & REQ_F_BUFFER_SELECTED) {
3126 struct io_buffer *kbuf;
3128 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3129 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
3130 iov[0].iov_len = kbuf->len;
3133 if (req->rw.len != 1)
3136 #ifdef CONFIG_COMPAT
3137 if (req->ctx->compat)
3138 return io_compat_import(req, iov, needs_lock);
3141 return __io_iov_buffer_select(req, iov, needs_lock);
3144 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
3145 struct iovec **iovec, struct iov_iter *iter,
3148 void __user *buf = u64_to_user_ptr(req->rw.addr);
3149 size_t sqe_len = req->rw.len;
3153 opcode = req->opcode;
3154 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3156 return io_import_fixed(req, rw, iter);
3159 /* buffer index only valid with fixed read/write, or buffer select */
3160 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3163 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3164 if (req->flags & REQ_F_BUFFER_SELECT) {
3165 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3167 return PTR_ERR(buf);
3168 req->rw.len = sqe_len;
3171 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3176 if (req->flags & REQ_F_BUFFER_SELECT) {
3177 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3179 ret = (*iovec)->iov_len;
3180 iov_iter_init(iter, rw, *iovec, 1, ret);
3186 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3190 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3192 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3196 * For files that don't have ->read_iter() and ->write_iter(), handle them
3197 * by looping over ->read() or ->write() manually.
3199 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3201 struct kiocb *kiocb = &req->rw.kiocb;
3202 struct file *file = req->file;
3206 * Don't support polled IO through this interface, and we can't
3207 * support non-blocking either. For the latter, this just causes
3208 * the kiocb to be handled from an async context.
3210 if (kiocb->ki_flags & IOCB_HIPRI)
3212 if (kiocb->ki_flags & IOCB_NOWAIT)
3215 while (iov_iter_count(iter)) {
3219 if (!iov_iter_is_bvec(iter)) {
3220 iovec = iov_iter_iovec(iter);
3222 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3223 iovec.iov_len = req->rw.len;
3227 nr = file->f_op->read(file, iovec.iov_base,
3228 iovec.iov_len, io_kiocb_ppos(kiocb));
3230 nr = file->f_op->write(file, iovec.iov_base,
3231 iovec.iov_len, io_kiocb_ppos(kiocb));
3240 if (nr != iovec.iov_len)
3244 iov_iter_advance(iter, nr);
3250 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3251 const struct iovec *fast_iov, struct iov_iter *iter)
3253 struct io_async_rw *rw = req->async_data;
3255 memcpy(&rw->iter, iter, sizeof(*iter));
3256 rw->free_iovec = iovec;
3258 /* can only be fixed buffers, no need to do anything */
3259 if (iov_iter_is_bvec(iter))
3262 unsigned iov_off = 0;
3264 rw->iter.iov = rw->fast_iov;
3265 if (iter->iov != fast_iov) {
3266 iov_off = iter->iov - fast_iov;
3267 rw->iter.iov += iov_off;
3269 if (rw->fast_iov != fast_iov)
3270 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3271 sizeof(struct iovec) * iter->nr_segs);
3273 req->flags |= REQ_F_NEED_CLEANUP;
3277 static inline int __io_alloc_async_data(struct io_kiocb *req)
3279 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3280 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3281 return req->async_data == NULL;
3284 static int io_alloc_async_data(struct io_kiocb *req)
3286 if (!io_op_defs[req->opcode].needs_async_data)
3289 return __io_alloc_async_data(req);
3292 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3293 const struct iovec *fast_iov,
3294 struct iov_iter *iter, bool force)
3296 if (!force && !io_op_defs[req->opcode].needs_async_data)
3298 if (!req->async_data) {
3299 if (__io_alloc_async_data(req))
3302 io_req_map_rw(req, iovec, fast_iov, iter);
3307 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3309 struct io_async_rw *iorw = req->async_data;
3310 struct iovec *iov = iorw->fast_iov;
3313 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3314 if (unlikely(ret < 0))
3317 iorw->bytes_done = 0;
3318 iorw->free_iovec = iov;
3320 req->flags |= REQ_F_NEED_CLEANUP;
3324 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3328 ret = io_prep_rw(req, sqe);
3332 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3335 /* either don't need iovec imported or already have it */
3336 if (!req->async_data)
3338 return io_rw_prep_async(req, READ);
3342 * This is our waitqueue callback handler, registered through lock_page_async()
3343 * when we initially tried to do the IO with the iocb armed our waitqueue.
3344 * This gets called when the page is unlocked, and we generally expect that to
3345 * happen when the page IO is completed and the page is now uptodate. This will
3346 * queue a task_work based retry of the operation, attempting to copy the data
3347 * again. If the latter fails because the page was NOT uptodate, then we will
3348 * do a thread based blocking retry of the operation. That's the unexpected
3351 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3352 int sync, void *arg)
3354 struct wait_page_queue *wpq;
3355 struct io_kiocb *req = wait->private;
3356 struct wait_page_key *key = arg;
3359 wpq = container_of(wait, struct wait_page_queue, wait);
3361 if (!wake_page_match(wpq, key))
3364 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3365 list_del_init(&wait->entry);
3367 init_task_work(&req->task_work, io_req_task_submit);
3368 percpu_ref_get(&req->ctx->refs);
3370 /* submit ref gets dropped, acquire a new one */
3371 refcount_inc(&req->refs);
3372 ret = io_req_task_work_add(req);
3373 if (unlikely(ret)) {
3374 struct task_struct *tsk;
3376 /* queue just for cancelation */
3377 init_task_work(&req->task_work, io_req_task_cancel);
3378 tsk = io_wq_get_task(req->ctx->io_wq);
3379 task_work_add(tsk, &req->task_work, TWA_NONE);
3380 wake_up_process(tsk);
3386 * This controls whether a given IO request should be armed for async page
3387 * based retry. If we return false here, the request is handed to the async
3388 * worker threads for retry. If we're doing buffered reads on a regular file,
3389 * we prepare a private wait_page_queue entry and retry the operation. This
3390 * will either succeed because the page is now uptodate and unlocked, or it
3391 * will register a callback when the page is unlocked at IO completion. Through
3392 * that callback, io_uring uses task_work to setup a retry of the operation.
3393 * That retry will attempt the buffered read again. The retry will generally
3394 * succeed, or in rare cases where it fails, we then fall back to using the
3395 * async worker threads for a blocking retry.
3397 static bool io_rw_should_retry(struct io_kiocb *req)
3399 struct io_async_rw *rw = req->async_data;
3400 struct wait_page_queue *wait = &rw->wpq;
3401 struct kiocb *kiocb = &req->rw.kiocb;
3403 /* never retry for NOWAIT, we just complete with -EAGAIN */
3404 if (req->flags & REQ_F_NOWAIT)
3407 /* Only for buffered IO */
3408 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3412 * just use poll if we can, and don't attempt if the fs doesn't
3413 * support callback based unlocks
3415 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3418 wait->wait.func = io_async_buf_func;
3419 wait->wait.private = req;
3420 wait->wait.flags = 0;
3421 INIT_LIST_HEAD(&wait->wait.entry);
3422 kiocb->ki_flags |= IOCB_WAITQ;
3423 kiocb->ki_flags &= ~IOCB_NOWAIT;
3424 kiocb->ki_waitq = wait;
3428 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3430 if (req->file->f_op->read_iter)
3431 return call_read_iter(req->file, &req->rw.kiocb, iter);
3432 else if (req->file->f_op->read)
3433 return loop_rw_iter(READ, req, iter);
3438 static int io_read(struct io_kiocb *req, bool force_nonblock,
3439 struct io_comp_state *cs)
3441 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3442 struct kiocb *kiocb = &req->rw.kiocb;
3443 struct iov_iter __iter, *iter = &__iter;
3444 struct io_async_rw *rw = req->async_data;
3445 ssize_t io_size, ret, ret2;
3452 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3456 io_size = iov_iter_count(iter);
3457 req->result = io_size;
3460 /* Ensure we clear previously set non-block flag */
3461 if (!force_nonblock)
3462 kiocb->ki_flags &= ~IOCB_NOWAIT;
3464 kiocb->ki_flags |= IOCB_NOWAIT;
3467 /* If the file doesn't support async, just async punt */
3468 no_async = force_nonblock && !io_file_supports_async(req->file, READ);
3472 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size);
3476 ret = io_iter_do_read(req, iter);
3480 } else if (ret == -EIOCBQUEUED) {
3483 } else if (ret == -EAGAIN) {
3484 /* IOPOLL retry should happen for io-wq threads */
3485 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3487 /* no retry on NONBLOCK marked file */
3488 if (req->file->f_flags & O_NONBLOCK)
3490 /* some cases will consume bytes even on error returns */
3491 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3494 } else if (ret < 0) {
3495 /* make sure -ERESTARTSYS -> -EINTR is done */
3499 /* read it all, or we did blocking attempt. no retry. */
3500 if (!iov_iter_count(iter) || !force_nonblock ||
3501 (req->file->f_flags & O_NONBLOCK))
3506 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3513 rw = req->async_data;
3514 /* it's copied and will be cleaned with ->io */
3516 /* now use our persistent iterator, if we aren't already */
3519 rw->bytes_done += ret;
3520 /* if we can retry, do so with the callbacks armed */
3521 if (!io_rw_should_retry(req)) {
3522 kiocb->ki_flags &= ~IOCB_WAITQ;
3527 * Now retry read with the IOCB_WAITQ parts set in the iocb. If we
3528 * get -EIOCBQUEUED, then we'll get a notification when the desired
3529 * page gets unlocked. We can also get a partial read here, and if we
3530 * do, then just retry at the new offset.
3532 ret = io_iter_do_read(req, iter);
3533 if (ret == -EIOCBQUEUED) {
3536 } else if (ret > 0 && ret < io_size) {
3537 /* we got some bytes, but not all. retry. */
3541 kiocb_done(kiocb, ret, cs);
3544 /* it's reportedly faster than delegating the null check to kfree() */
3550 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3554 ret = io_prep_rw(req, sqe);
3558 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3561 /* either don't need iovec imported or already have it */
3562 if (!req->async_data)
3564 return io_rw_prep_async(req, WRITE);
3567 static int io_write(struct io_kiocb *req, bool force_nonblock,
3568 struct io_comp_state *cs)
3570 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3571 struct kiocb *kiocb = &req->rw.kiocb;
3572 struct iov_iter __iter, *iter = &__iter;
3573 struct io_async_rw *rw = req->async_data;
3574 ssize_t ret, ret2, io_size;
3580 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3584 io_size = iov_iter_count(iter);
3585 req->result = io_size;
3587 /* Ensure we clear previously set non-block flag */
3588 if (!force_nonblock)
3589 kiocb->ki_flags &= ~IOCB_NOWAIT;
3591 kiocb->ki_flags |= IOCB_NOWAIT;
3593 /* If the file doesn't support async, just async punt */
3594 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
3597 /* file path doesn't support NOWAIT for non-direct_IO */
3598 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3599 (req->flags & REQ_F_ISREG))
3602 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size);
3607 * Open-code file_start_write here to grab freeze protection,
3608 * which will be released by another thread in
3609 * io_complete_rw(). Fool lockdep by telling it the lock got
3610 * released so that it doesn't complain about the held lock when
3611 * we return to userspace.
3613 if (req->flags & REQ_F_ISREG) {
3614 sb_start_write(file_inode(req->file)->i_sb);
3615 __sb_writers_release(file_inode(req->file)->i_sb,
3618 kiocb->ki_flags |= IOCB_WRITE;
3620 if (req->file->f_op->write_iter)
3621 ret2 = call_write_iter(req->file, kiocb, iter);
3622 else if (req->file->f_op->write)
3623 ret2 = loop_rw_iter(WRITE, req, iter);
3628 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3629 * retry them without IOCB_NOWAIT.
3631 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3633 /* no retry on NONBLOCK marked file */
3634 if (ret2 == -EAGAIN && (req->file->f_flags & O_NONBLOCK))
3636 if (!force_nonblock || ret2 != -EAGAIN) {
3637 /* IOPOLL retry should happen for io-wq threads */
3638 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3641 kiocb_done(kiocb, ret2, cs);
3644 /* some cases will consume bytes even on error returns */
3645 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3646 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3651 /* it's reportedly faster than delegating the null check to kfree() */
3657 static int io_renameat_prep(struct io_kiocb *req,
3658 const struct io_uring_sqe *sqe)
3660 struct io_rename *ren = &req->rename;
3661 const char __user *oldf, *newf;
3663 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3666 ren->old_dfd = READ_ONCE(sqe->fd);
3667 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3668 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3669 ren->new_dfd = READ_ONCE(sqe->len);
3670 ren->flags = READ_ONCE(sqe->rename_flags);
3672 ren->oldpath = getname(oldf);
3673 if (IS_ERR(ren->oldpath))
3674 return PTR_ERR(ren->oldpath);
3676 ren->newpath = getname(newf);
3677 if (IS_ERR(ren->newpath)) {
3678 putname(ren->oldpath);
3679 return PTR_ERR(ren->newpath);
3682 req->flags |= REQ_F_NEED_CLEANUP;
3686 static int io_renameat(struct io_kiocb *req, bool force_nonblock)
3688 struct io_rename *ren = &req->rename;
3694 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3695 ren->newpath, ren->flags);
3697 req->flags &= ~REQ_F_NEED_CLEANUP;
3699 req_set_fail_links(req);
3700 io_req_complete(req, ret);
3704 static int io_unlinkat_prep(struct io_kiocb *req,
3705 const struct io_uring_sqe *sqe)
3707 struct io_unlink *un = &req->unlink;
3708 const char __user *fname;
3710 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3713 un->dfd = READ_ONCE(sqe->fd);
3715 un->flags = READ_ONCE(sqe->unlink_flags);
3716 if (un->flags & ~AT_REMOVEDIR)
3719 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3720 un->filename = getname(fname);
3721 if (IS_ERR(un->filename))
3722 return PTR_ERR(un->filename);
3724 req->flags |= REQ_F_NEED_CLEANUP;
3728 static int io_unlinkat(struct io_kiocb *req, bool force_nonblock)
3730 struct io_unlink *un = &req->unlink;
3736 if (un->flags & AT_REMOVEDIR)
3737 ret = do_rmdir(un->dfd, un->filename);
3739 ret = do_unlinkat(un->dfd, un->filename);
3741 req->flags &= ~REQ_F_NEED_CLEANUP;
3743 req_set_fail_links(req);
3744 io_req_complete(req, ret);
3748 static int io_shutdown_prep(struct io_kiocb *req,
3749 const struct io_uring_sqe *sqe)
3751 #if defined(CONFIG_NET)
3752 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3754 if (sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3758 req->shutdown.how = READ_ONCE(sqe->len);
3765 static int io_shutdown(struct io_kiocb *req, bool force_nonblock)
3767 #if defined(CONFIG_NET)
3768 struct socket *sock;
3774 sock = sock_from_file(req->file);
3775 if (unlikely(!sock))
3778 ret = __sys_shutdown_sock(sock, req->shutdown.how);
3780 req_set_fail_links(req);
3781 io_req_complete(req, ret);
3788 static int __io_splice_prep(struct io_kiocb *req,
3789 const struct io_uring_sqe *sqe)
3791 struct io_splice* sp = &req->splice;
3792 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3794 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3798 sp->len = READ_ONCE(sqe->len);
3799 sp->flags = READ_ONCE(sqe->splice_flags);
3801 if (unlikely(sp->flags & ~valid_flags))
3804 sp->file_in = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in),
3805 (sp->flags & SPLICE_F_FD_IN_FIXED));
3808 req->flags |= REQ_F_NEED_CLEANUP;
3810 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
3812 * Splice operation will be punted aync, and here need to
3813 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3815 io_req_init_async(req);
3816 req->work.flags |= IO_WQ_WORK_UNBOUND;
3822 static int io_tee_prep(struct io_kiocb *req,
3823 const struct io_uring_sqe *sqe)
3825 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3827 return __io_splice_prep(req, sqe);
3830 static int io_tee(struct io_kiocb *req, bool force_nonblock)
3832 struct io_splice *sp = &req->splice;
3833 struct file *in = sp->file_in;
3834 struct file *out = sp->file_out;
3835 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3841 ret = do_tee(in, out, sp->len, flags);
3843 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3844 req->flags &= ~REQ_F_NEED_CLEANUP;
3847 req_set_fail_links(req);
3848 io_req_complete(req, ret);
3852 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3854 struct io_splice* sp = &req->splice;
3856 sp->off_in = READ_ONCE(sqe->splice_off_in);
3857 sp->off_out = READ_ONCE(sqe->off);
3858 return __io_splice_prep(req, sqe);
3861 static int io_splice(struct io_kiocb *req, bool force_nonblock)
3863 struct io_splice *sp = &req->splice;
3864 struct file *in = sp->file_in;
3865 struct file *out = sp->file_out;
3866 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3867 loff_t *poff_in, *poff_out;
3873 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3874 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3877 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3879 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3880 req->flags &= ~REQ_F_NEED_CLEANUP;
3883 req_set_fail_links(req);
3884 io_req_complete(req, ret);
3889 * IORING_OP_NOP just posts a completion event, nothing else.
3891 static int io_nop(struct io_kiocb *req, struct io_comp_state *cs)
3893 struct io_ring_ctx *ctx = req->ctx;
3895 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3898 __io_req_complete(req, 0, 0, cs);
3902 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3904 struct io_ring_ctx *ctx = req->ctx;
3909 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3911 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3914 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3915 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3918 req->sync.off = READ_ONCE(sqe->off);
3919 req->sync.len = READ_ONCE(sqe->len);
3923 static int io_fsync(struct io_kiocb *req, bool force_nonblock)
3925 loff_t end = req->sync.off + req->sync.len;
3928 /* fsync always requires a blocking context */
3932 ret = vfs_fsync_range(req->file, req->sync.off,
3933 end > 0 ? end : LLONG_MAX,
3934 req->sync.flags & IORING_FSYNC_DATASYNC);
3936 req_set_fail_links(req);
3937 io_req_complete(req, ret);
3941 static int io_fallocate_prep(struct io_kiocb *req,
3942 const struct io_uring_sqe *sqe)
3944 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3946 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3949 req->sync.off = READ_ONCE(sqe->off);
3950 req->sync.len = READ_ONCE(sqe->addr);
3951 req->sync.mode = READ_ONCE(sqe->len);
3955 static int io_fallocate(struct io_kiocb *req, bool force_nonblock)
3959 /* fallocate always requiring blocking context */
3962 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3965 req_set_fail_links(req);
3966 io_req_complete(req, ret);
3970 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3972 const char __user *fname;
3975 if (unlikely(sqe->ioprio || sqe->buf_index))
3977 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3980 /* open.how should be already initialised */
3981 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3982 req->open.how.flags |= O_LARGEFILE;
3984 req->open.dfd = READ_ONCE(sqe->fd);
3985 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3986 req->open.filename = getname(fname);
3987 if (IS_ERR(req->open.filename)) {
3988 ret = PTR_ERR(req->open.filename);
3989 req->open.filename = NULL;
3992 req->open.nofile = rlimit(RLIMIT_NOFILE);
3993 req->open.ignore_nonblock = false;
3994 req->flags |= REQ_F_NEED_CLEANUP;
3998 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4002 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4004 mode = READ_ONCE(sqe->len);
4005 flags = READ_ONCE(sqe->open_flags);
4006 req->open.how = build_open_how(flags, mode);
4007 return __io_openat_prep(req, sqe);
4010 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4012 struct open_how __user *how;
4016 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4018 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4019 len = READ_ONCE(sqe->len);
4020 if (len < OPEN_HOW_SIZE_VER0)
4023 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
4028 return __io_openat_prep(req, sqe);
4031 static int io_openat2(struct io_kiocb *req, bool force_nonblock)
4033 struct open_flags op;
4037 if (force_nonblock && !req->open.ignore_nonblock)
4040 ret = build_open_flags(&req->open.how, &op);
4044 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
4048 file = do_filp_open(req->open.dfd, req->open.filename, &op);
4051 ret = PTR_ERR(file);
4053 * A work-around to ensure that /proc/self works that way
4054 * that it should - if we get -EOPNOTSUPP back, then assume
4055 * that proc_self_get_link() failed us because we're in async
4056 * context. We should be safe to retry this from the task
4057 * itself with force_nonblock == false set, as it should not
4058 * block on lookup. Would be nice to know this upfront and
4059 * avoid the async dance, but doesn't seem feasible.
4061 if (ret == -EOPNOTSUPP && io_wq_current_is_worker()) {
4062 req->open.ignore_nonblock = true;
4063 refcount_inc(&req->refs);
4064 io_req_task_queue(req);
4068 fsnotify_open(file);
4069 fd_install(ret, file);
4072 putname(req->open.filename);
4073 req->flags &= ~REQ_F_NEED_CLEANUP;
4075 req_set_fail_links(req);
4076 io_req_complete(req, ret);
4080 static int io_openat(struct io_kiocb *req, bool force_nonblock)
4082 return io_openat2(req, force_nonblock);
4085 static int io_remove_buffers_prep(struct io_kiocb *req,
4086 const struct io_uring_sqe *sqe)
4088 struct io_provide_buf *p = &req->pbuf;
4091 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
4094 tmp = READ_ONCE(sqe->fd);
4095 if (!tmp || tmp > USHRT_MAX)
4098 memset(p, 0, sizeof(*p));
4100 p->bgid = READ_ONCE(sqe->buf_group);
4104 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
4105 int bgid, unsigned nbufs)
4109 /* shouldn't happen */
4113 /* the head kbuf is the list itself */
4114 while (!list_empty(&buf->list)) {
4115 struct io_buffer *nxt;
4117 nxt = list_first_entry(&buf->list, struct io_buffer, list);
4118 list_del(&nxt->list);
4125 idr_remove(&ctx->io_buffer_idr, bgid);
4130 static int io_remove_buffers(struct io_kiocb *req, bool force_nonblock,
4131 struct io_comp_state *cs)
4133 struct io_provide_buf *p = &req->pbuf;
4134 struct io_ring_ctx *ctx = req->ctx;
4135 struct io_buffer *head;
4138 io_ring_submit_lock(ctx, !force_nonblock);
4140 lockdep_assert_held(&ctx->uring_lock);
4143 head = idr_find(&ctx->io_buffer_idr, p->bgid);
4145 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
4147 req_set_fail_links(req);
4149 /* need to hold the lock to complete IOPOLL requests */
4150 if (ctx->flags & IORING_SETUP_IOPOLL) {
4151 __io_req_complete(req, ret, 0, cs);
4152 io_ring_submit_unlock(ctx, !force_nonblock);
4154 io_ring_submit_unlock(ctx, !force_nonblock);
4155 __io_req_complete(req, ret, 0, cs);
4160 static int io_provide_buffers_prep(struct io_kiocb *req,
4161 const struct io_uring_sqe *sqe)
4163 struct io_provide_buf *p = &req->pbuf;
4166 if (sqe->ioprio || sqe->rw_flags)
4169 tmp = READ_ONCE(sqe->fd);
4170 if (!tmp || tmp > USHRT_MAX)
4173 p->addr = READ_ONCE(sqe->addr);
4174 p->len = READ_ONCE(sqe->len);
4176 if (!access_ok(u64_to_user_ptr(p->addr), (p->len * p->nbufs)))
4179 p->bgid = READ_ONCE(sqe->buf_group);
4180 tmp = READ_ONCE(sqe->off);
4181 if (tmp > USHRT_MAX)
4187 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
4189 struct io_buffer *buf;
4190 u64 addr = pbuf->addr;
4191 int i, bid = pbuf->bid;
4193 for (i = 0; i < pbuf->nbufs; i++) {
4194 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
4199 buf->len = pbuf->len;
4204 INIT_LIST_HEAD(&buf->list);
4207 list_add_tail(&buf->list, &(*head)->list);
4211 return i ? i : -ENOMEM;
4214 static int io_provide_buffers(struct io_kiocb *req, bool force_nonblock,
4215 struct io_comp_state *cs)
4217 struct io_provide_buf *p = &req->pbuf;
4218 struct io_ring_ctx *ctx = req->ctx;
4219 struct io_buffer *head, *list;
4222 io_ring_submit_lock(ctx, !force_nonblock);
4224 lockdep_assert_held(&ctx->uring_lock);
4226 list = head = idr_find(&ctx->io_buffer_idr, p->bgid);
4228 ret = io_add_buffers(p, &head);
4233 ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1,
4236 __io_remove_buffers(ctx, head, p->bgid, -1U);
4242 req_set_fail_links(req);
4244 /* need to hold the lock to complete IOPOLL requests */
4245 if (ctx->flags & IORING_SETUP_IOPOLL) {
4246 __io_req_complete(req, ret, 0, cs);
4247 io_ring_submit_unlock(ctx, !force_nonblock);
4249 io_ring_submit_unlock(ctx, !force_nonblock);
4250 __io_req_complete(req, ret, 0, cs);
4255 static int io_epoll_ctl_prep(struct io_kiocb *req,
4256 const struct io_uring_sqe *sqe)
4258 #if defined(CONFIG_EPOLL)
4259 if (sqe->ioprio || sqe->buf_index)
4261 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4264 req->epoll.epfd = READ_ONCE(sqe->fd);
4265 req->epoll.op = READ_ONCE(sqe->len);
4266 req->epoll.fd = READ_ONCE(sqe->off);
4268 if (ep_op_has_event(req->epoll.op)) {
4269 struct epoll_event __user *ev;
4271 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4272 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4282 static int io_epoll_ctl(struct io_kiocb *req, bool force_nonblock,
4283 struct io_comp_state *cs)
4285 #if defined(CONFIG_EPOLL)
4286 struct io_epoll *ie = &req->epoll;
4289 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4290 if (force_nonblock && ret == -EAGAIN)
4294 req_set_fail_links(req);
4295 __io_req_complete(req, ret, 0, cs);
4302 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4304 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4305 if (sqe->ioprio || sqe->buf_index || sqe->off)
4307 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4310 req->madvise.addr = READ_ONCE(sqe->addr);
4311 req->madvise.len = READ_ONCE(sqe->len);
4312 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4319 static int io_madvise(struct io_kiocb *req, bool force_nonblock)
4321 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4322 struct io_madvise *ma = &req->madvise;
4328 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4330 req_set_fail_links(req);
4331 io_req_complete(req, ret);
4338 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4340 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4342 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4345 req->fadvise.offset = READ_ONCE(sqe->off);
4346 req->fadvise.len = READ_ONCE(sqe->len);
4347 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4351 static int io_fadvise(struct io_kiocb *req, bool force_nonblock)
4353 struct io_fadvise *fa = &req->fadvise;
4356 if (force_nonblock) {
4357 switch (fa->advice) {
4358 case POSIX_FADV_NORMAL:
4359 case POSIX_FADV_RANDOM:
4360 case POSIX_FADV_SEQUENTIAL:
4367 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4369 req_set_fail_links(req);
4370 io_req_complete(req, ret);
4374 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4376 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4378 if (sqe->ioprio || sqe->buf_index)
4380 if (req->flags & REQ_F_FIXED_FILE)
4383 req->statx.dfd = READ_ONCE(sqe->fd);
4384 req->statx.mask = READ_ONCE(sqe->len);
4385 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4386 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4387 req->statx.flags = READ_ONCE(sqe->statx_flags);
4392 static int io_statx(struct io_kiocb *req, bool force_nonblock)
4394 struct io_statx *ctx = &req->statx;
4397 if (force_nonblock) {
4398 /* only need file table for an actual valid fd */
4399 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
4400 req->flags |= REQ_F_NO_FILE_TABLE;
4404 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4408 req_set_fail_links(req);
4409 io_req_complete(req, ret);
4413 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4416 * If we queue this for async, it must not be cancellable. That would
4417 * leave the 'file' in an undeterminate state, and here need to modify
4418 * io_wq_work.flags, so initialize io_wq_work firstly.
4420 io_req_init_async(req);
4421 req->work.flags |= IO_WQ_WORK_NO_CANCEL;
4423 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4425 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4426 sqe->rw_flags || sqe->buf_index)
4428 if (req->flags & REQ_F_FIXED_FILE)
4431 req->close.fd = READ_ONCE(sqe->fd);
4432 if ((req->file && req->file->f_op == &io_uring_fops))
4435 req->close.put_file = NULL;
4439 static int io_close(struct io_kiocb *req, bool force_nonblock,
4440 struct io_comp_state *cs)
4442 struct io_close *close = &req->close;
4445 /* might be already done during nonblock submission */
4446 if (!close->put_file) {
4447 ret = close_fd_get_file(close->fd, &close->put_file);
4449 return (ret == -ENOENT) ? -EBADF : ret;
4452 /* if the file has a flush method, be safe and punt to async */
4453 if (close->put_file->f_op->flush && force_nonblock) {
4454 /* was never set, but play safe */
4455 req->flags &= ~REQ_F_NOWAIT;
4456 /* avoid grabbing files - we don't need the files */
4457 req->flags |= REQ_F_NO_FILE_TABLE;
4461 /* No ->flush() or already async, safely close from here */
4462 ret = filp_close(close->put_file, req->work.identity->files);
4464 req_set_fail_links(req);
4465 fput(close->put_file);
4466 close->put_file = NULL;
4467 __io_req_complete(req, ret, 0, cs);
4471 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4473 struct io_ring_ctx *ctx = req->ctx;
4478 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4480 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4483 req->sync.off = READ_ONCE(sqe->off);
4484 req->sync.len = READ_ONCE(sqe->len);
4485 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4489 static int io_sync_file_range(struct io_kiocb *req, bool force_nonblock)
4493 /* sync_file_range always requires a blocking context */
4497 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4500 req_set_fail_links(req);
4501 io_req_complete(req, ret);
4505 #if defined(CONFIG_NET)
4506 static int io_setup_async_msg(struct io_kiocb *req,
4507 struct io_async_msghdr *kmsg)
4509 struct io_async_msghdr *async_msg = req->async_data;
4513 if (io_alloc_async_data(req)) {
4514 if (kmsg->iov != kmsg->fast_iov)
4518 async_msg = req->async_data;
4519 req->flags |= REQ_F_NEED_CLEANUP;
4520 memcpy(async_msg, kmsg, sizeof(*kmsg));
4524 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4525 struct io_async_msghdr *iomsg)
4527 iomsg->iov = iomsg->fast_iov;
4528 iomsg->msg.msg_name = &iomsg->addr;
4529 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4530 req->sr_msg.msg_flags, &iomsg->iov);
4533 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4535 struct io_async_msghdr *async_msg = req->async_data;
4536 struct io_sr_msg *sr = &req->sr_msg;
4539 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4542 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4543 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4544 sr->len = READ_ONCE(sqe->len);
4546 #ifdef CONFIG_COMPAT
4547 if (req->ctx->compat)
4548 sr->msg_flags |= MSG_CMSG_COMPAT;
4551 if (!async_msg || !io_op_defs[req->opcode].needs_async_data)
4553 ret = io_sendmsg_copy_hdr(req, async_msg);
4555 req->flags |= REQ_F_NEED_CLEANUP;
4559 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4560 struct io_comp_state *cs)
4562 struct io_async_msghdr iomsg, *kmsg;
4563 struct socket *sock;
4567 sock = sock_from_file(req->file);
4568 if (unlikely(!sock))
4571 if (req->async_data) {
4572 kmsg = req->async_data;
4573 kmsg->msg.msg_name = &kmsg->addr;
4574 /* if iov is set, it's allocated already */
4576 kmsg->iov = kmsg->fast_iov;
4577 kmsg->msg.msg_iter.iov = kmsg->iov;
4579 ret = io_sendmsg_copy_hdr(req, &iomsg);
4585 flags = req->sr_msg.msg_flags;
4586 if (flags & MSG_DONTWAIT)
4587 req->flags |= REQ_F_NOWAIT;
4588 else if (force_nonblock)
4589 flags |= MSG_DONTWAIT;
4591 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4592 if (force_nonblock && ret == -EAGAIN)
4593 return io_setup_async_msg(req, kmsg);
4594 if (ret == -ERESTARTSYS)
4597 if (kmsg->iov != kmsg->fast_iov)
4599 req->flags &= ~REQ_F_NEED_CLEANUP;
4601 req_set_fail_links(req);
4602 __io_req_complete(req, ret, 0, cs);
4606 static int io_send(struct io_kiocb *req, bool force_nonblock,
4607 struct io_comp_state *cs)
4609 struct io_sr_msg *sr = &req->sr_msg;
4612 struct socket *sock;
4616 sock = sock_from_file(req->file);
4617 if (unlikely(!sock))
4620 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4624 msg.msg_name = NULL;
4625 msg.msg_control = NULL;
4626 msg.msg_controllen = 0;
4627 msg.msg_namelen = 0;
4629 flags = req->sr_msg.msg_flags;
4630 if (flags & MSG_DONTWAIT)
4631 req->flags |= REQ_F_NOWAIT;
4632 else if (force_nonblock)
4633 flags |= MSG_DONTWAIT;
4635 msg.msg_flags = flags;
4636 ret = sock_sendmsg(sock, &msg);
4637 if (force_nonblock && ret == -EAGAIN)
4639 if (ret == -ERESTARTSYS)
4643 req_set_fail_links(req);
4644 __io_req_complete(req, ret, 0, cs);
4648 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4649 struct io_async_msghdr *iomsg)
4651 struct io_sr_msg *sr = &req->sr_msg;
4652 struct iovec __user *uiov;
4656 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4657 &iomsg->uaddr, &uiov, &iov_len);
4661 if (req->flags & REQ_F_BUFFER_SELECT) {
4664 if (copy_from_user(iomsg->iov, uiov, sizeof(*uiov)))
4666 sr->len = iomsg->iov[0].iov_len;
4667 iov_iter_init(&iomsg->msg.msg_iter, READ, iomsg->iov, 1,
4671 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4672 &iomsg->iov, &iomsg->msg.msg_iter,
4681 #ifdef CONFIG_COMPAT
4682 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4683 struct io_async_msghdr *iomsg)
4685 struct compat_msghdr __user *msg_compat;
4686 struct io_sr_msg *sr = &req->sr_msg;
4687 struct compat_iovec __user *uiov;
4692 msg_compat = (struct compat_msghdr __user *) sr->umsg;
4693 ret = __get_compat_msghdr(&iomsg->msg, msg_compat, &iomsg->uaddr,
4698 uiov = compat_ptr(ptr);
4699 if (req->flags & REQ_F_BUFFER_SELECT) {
4700 compat_ssize_t clen;
4704 if (!access_ok(uiov, sizeof(*uiov)))
4706 if (__get_user(clen, &uiov->iov_len))
4711 iomsg->iov[0].iov_len = clen;
4714 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4715 UIO_FASTIOV, &iomsg->iov,
4716 &iomsg->msg.msg_iter, true);
4725 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4726 struct io_async_msghdr *iomsg)
4728 iomsg->msg.msg_name = &iomsg->addr;
4729 iomsg->iov = iomsg->fast_iov;
4731 #ifdef CONFIG_COMPAT
4732 if (req->ctx->compat)
4733 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4736 return __io_recvmsg_copy_hdr(req, iomsg);
4739 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4742 struct io_sr_msg *sr = &req->sr_msg;
4743 struct io_buffer *kbuf;
4745 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4750 req->flags |= REQ_F_BUFFER_SELECTED;
4754 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4756 return io_put_kbuf(req, req->sr_msg.kbuf);
4759 static int io_recvmsg_prep(struct io_kiocb *req,
4760 const struct io_uring_sqe *sqe)
4762 struct io_async_msghdr *async_msg = req->async_data;
4763 struct io_sr_msg *sr = &req->sr_msg;
4766 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4769 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4770 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4771 sr->len = READ_ONCE(sqe->len);
4772 sr->bgid = READ_ONCE(sqe->buf_group);
4774 #ifdef CONFIG_COMPAT
4775 if (req->ctx->compat)
4776 sr->msg_flags |= MSG_CMSG_COMPAT;
4779 if (!async_msg || !io_op_defs[req->opcode].needs_async_data)
4781 ret = io_recvmsg_copy_hdr(req, async_msg);
4783 req->flags |= REQ_F_NEED_CLEANUP;
4787 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4788 struct io_comp_state *cs)
4790 struct io_async_msghdr iomsg, *kmsg;
4791 struct socket *sock;
4792 struct io_buffer *kbuf;
4794 int ret, cflags = 0;
4796 sock = sock_from_file(req->file);
4797 if (unlikely(!sock))
4800 if (req->async_data) {
4801 kmsg = req->async_data;
4802 kmsg->msg.msg_name = &kmsg->addr;
4803 /* if iov is set, it's allocated already */
4805 kmsg->iov = kmsg->fast_iov;
4806 kmsg->msg.msg_iter.iov = kmsg->iov;
4808 ret = io_recvmsg_copy_hdr(req, &iomsg);
4814 if (req->flags & REQ_F_BUFFER_SELECT) {
4815 kbuf = io_recv_buffer_select(req, !force_nonblock);
4817 return PTR_ERR(kbuf);
4818 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4819 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->iov,
4820 1, req->sr_msg.len);
4823 flags = req->sr_msg.msg_flags;
4824 if (flags & MSG_DONTWAIT)
4825 req->flags |= REQ_F_NOWAIT;
4826 else if (force_nonblock)
4827 flags |= MSG_DONTWAIT;
4829 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4830 kmsg->uaddr, flags);
4831 if (force_nonblock && ret == -EAGAIN)
4832 return io_setup_async_msg(req, kmsg);
4833 if (ret == -ERESTARTSYS)
4836 if (req->flags & REQ_F_BUFFER_SELECTED)
4837 cflags = io_put_recv_kbuf(req);
4838 if (kmsg->iov != kmsg->fast_iov)
4840 req->flags &= ~REQ_F_NEED_CLEANUP;
4842 req_set_fail_links(req);
4843 __io_req_complete(req, ret, cflags, cs);
4847 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4848 struct io_comp_state *cs)
4850 struct io_buffer *kbuf;
4851 struct io_sr_msg *sr = &req->sr_msg;
4853 void __user *buf = sr->buf;
4854 struct socket *sock;
4857 int ret, cflags = 0;
4859 sock = sock_from_file(req->file);
4860 if (unlikely(!sock))
4863 if (req->flags & REQ_F_BUFFER_SELECT) {
4864 kbuf = io_recv_buffer_select(req, !force_nonblock);
4866 return PTR_ERR(kbuf);
4867 buf = u64_to_user_ptr(kbuf->addr);
4870 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4874 msg.msg_name = NULL;
4875 msg.msg_control = NULL;
4876 msg.msg_controllen = 0;
4877 msg.msg_namelen = 0;
4878 msg.msg_iocb = NULL;
4881 flags = req->sr_msg.msg_flags;
4882 if (flags & MSG_DONTWAIT)
4883 req->flags |= REQ_F_NOWAIT;
4884 else if (force_nonblock)
4885 flags |= MSG_DONTWAIT;
4887 ret = sock_recvmsg(sock, &msg, flags);
4888 if (force_nonblock && ret == -EAGAIN)
4890 if (ret == -ERESTARTSYS)
4893 if (req->flags & REQ_F_BUFFER_SELECTED)
4894 cflags = io_put_recv_kbuf(req);
4896 req_set_fail_links(req);
4897 __io_req_complete(req, ret, cflags, cs);
4901 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4903 struct io_accept *accept = &req->accept;
4905 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4907 if (sqe->ioprio || sqe->len || sqe->buf_index)
4910 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4911 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4912 accept->flags = READ_ONCE(sqe->accept_flags);
4913 accept->nofile = rlimit(RLIMIT_NOFILE);
4917 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4918 struct io_comp_state *cs)
4920 struct io_accept *accept = &req->accept;
4921 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4924 if (req->file->f_flags & O_NONBLOCK)
4925 req->flags |= REQ_F_NOWAIT;
4927 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4928 accept->addr_len, accept->flags,
4930 if (ret == -EAGAIN && force_nonblock)
4933 if (ret == -ERESTARTSYS)
4935 req_set_fail_links(req);
4937 __io_req_complete(req, ret, 0, cs);
4941 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4943 struct io_connect *conn = &req->connect;
4944 struct io_async_connect *io = req->async_data;
4946 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4948 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4951 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4952 conn->addr_len = READ_ONCE(sqe->addr2);
4957 return move_addr_to_kernel(conn->addr, conn->addr_len,
4961 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4962 struct io_comp_state *cs)
4964 struct io_async_connect __io, *io;
4965 unsigned file_flags;
4968 if (req->async_data) {
4969 io = req->async_data;
4971 ret = move_addr_to_kernel(req->connect.addr,
4972 req->connect.addr_len,
4979 file_flags = force_nonblock ? O_NONBLOCK : 0;
4981 ret = __sys_connect_file(req->file, &io->address,
4982 req->connect.addr_len, file_flags);
4983 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4984 if (req->async_data)
4986 if (io_alloc_async_data(req)) {
4990 io = req->async_data;
4991 memcpy(req->async_data, &__io, sizeof(__io));
4994 if (ret == -ERESTARTSYS)
4998 req_set_fail_links(req);
4999 __io_req_complete(req, ret, 0, cs);
5002 #else /* !CONFIG_NET */
5003 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5008 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
5009 struct io_comp_state *cs)
5014 static int io_send(struct io_kiocb *req, bool force_nonblock,
5015 struct io_comp_state *cs)
5020 static int io_recvmsg_prep(struct io_kiocb *req,
5021 const struct io_uring_sqe *sqe)
5026 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
5027 struct io_comp_state *cs)
5032 static int io_recv(struct io_kiocb *req, bool force_nonblock,
5033 struct io_comp_state *cs)
5038 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5043 static int io_accept(struct io_kiocb *req, bool force_nonblock,
5044 struct io_comp_state *cs)
5049 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5054 static int io_connect(struct io_kiocb *req, bool force_nonblock,
5055 struct io_comp_state *cs)
5059 #endif /* CONFIG_NET */
5061 struct io_poll_table {
5062 struct poll_table_struct pt;
5063 struct io_kiocb *req;
5067 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
5068 __poll_t mask, task_work_func_t func)
5072 /* for instances that support it check for an event match first: */
5073 if (mask && !(mask & poll->events))
5076 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
5078 list_del_init(&poll->wait.entry);
5081 init_task_work(&req->task_work, func);
5082 percpu_ref_get(&req->ctx->refs);
5085 * If this fails, then the task is exiting. When a task exits, the
5086 * work gets canceled, so just cancel this request as well instead
5087 * of executing it. We can't safely execute it anyway, as we may not
5088 * have the needed state needed for it anyway.
5090 ret = io_req_task_work_add(req);
5091 if (unlikely(ret)) {
5092 struct task_struct *tsk;
5094 WRITE_ONCE(poll->canceled, true);
5095 tsk = io_wq_get_task(req->ctx->io_wq);
5096 task_work_add(tsk, &req->task_work, TWA_NONE);
5097 wake_up_process(tsk);
5102 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
5103 __acquires(&req->ctx->completion_lock)
5105 struct io_ring_ctx *ctx = req->ctx;
5107 if (!req->result && !READ_ONCE(poll->canceled)) {
5108 struct poll_table_struct pt = { ._key = poll->events };
5110 req->result = vfs_poll(req->file, &pt) & poll->events;
5113 spin_lock_irq(&ctx->completion_lock);
5114 if (!req->result && !READ_ONCE(poll->canceled)) {
5115 add_wait_queue(poll->head, &poll->wait);
5122 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
5124 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
5125 if (req->opcode == IORING_OP_POLL_ADD)
5126 return req->async_data;
5127 return req->apoll->double_poll;
5130 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
5132 if (req->opcode == IORING_OP_POLL_ADD)
5134 return &req->apoll->poll;
5137 static void io_poll_remove_double(struct io_kiocb *req)
5139 struct io_poll_iocb *poll = io_poll_get_double(req);
5141 lockdep_assert_held(&req->ctx->completion_lock);
5143 if (poll && poll->head) {
5144 struct wait_queue_head *head = poll->head;
5146 spin_lock(&head->lock);
5147 list_del_init(&poll->wait.entry);
5148 if (poll->wait.private)
5149 refcount_dec(&req->refs);
5151 spin_unlock(&head->lock);
5155 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
5157 struct io_ring_ctx *ctx = req->ctx;
5159 io_poll_remove_double(req);
5160 req->poll.done = true;
5161 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
5162 io_commit_cqring(ctx);
5165 static void io_poll_task_func(struct callback_head *cb)
5167 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
5168 struct io_ring_ctx *ctx = req->ctx;
5169 struct io_kiocb *nxt;
5171 if (io_poll_rewait(req, &req->poll)) {
5172 spin_unlock_irq(&ctx->completion_lock);
5174 hash_del(&req->hash_node);
5175 io_poll_complete(req, req->result, 0);
5176 spin_unlock_irq(&ctx->completion_lock);
5178 nxt = io_put_req_find_next(req);
5179 io_cqring_ev_posted(ctx);
5181 __io_req_task_submit(nxt);
5184 percpu_ref_put(&ctx->refs);
5187 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
5188 int sync, void *key)
5190 struct io_kiocb *req = wait->private;
5191 struct io_poll_iocb *poll = io_poll_get_single(req);
5192 __poll_t mask = key_to_poll(key);
5194 /* for instances that support it check for an event match first: */
5195 if (mask && !(mask & poll->events))
5198 list_del_init(&wait->entry);
5200 if (poll && poll->head) {
5203 spin_lock(&poll->head->lock);
5204 done = list_empty(&poll->wait.entry);
5206 list_del_init(&poll->wait.entry);
5207 /* make sure double remove sees this as being gone */
5208 wait->private = NULL;
5209 spin_unlock(&poll->head->lock);
5211 /* use wait func handler, so it matches the rq type */
5212 poll->wait.func(&poll->wait, mode, sync, key);
5215 refcount_dec(&req->refs);
5219 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
5220 wait_queue_func_t wake_func)
5224 poll->canceled = false;
5225 poll->events = events;
5226 INIT_LIST_HEAD(&poll->wait.entry);
5227 init_waitqueue_func_entry(&poll->wait, wake_func);
5230 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5231 struct wait_queue_head *head,
5232 struct io_poll_iocb **poll_ptr)
5234 struct io_kiocb *req = pt->req;
5237 * If poll->head is already set, it's because the file being polled
5238 * uses multiple waitqueues for poll handling (eg one for read, one
5239 * for write). Setup a separate io_poll_iocb if this happens.
5241 if (unlikely(poll->head)) {
5242 struct io_poll_iocb *poll_one = poll;
5244 /* already have a 2nd entry, fail a third attempt */
5246 pt->error = -EINVAL;
5249 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5251 pt->error = -ENOMEM;
5254 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5255 refcount_inc(&req->refs);
5256 poll->wait.private = req;
5263 if (poll->events & EPOLLEXCLUSIVE)
5264 add_wait_queue_exclusive(head, &poll->wait);
5266 add_wait_queue(head, &poll->wait);
5269 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5270 struct poll_table_struct *p)
5272 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5273 struct async_poll *apoll = pt->req->apoll;
5275 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5278 static void io_async_task_func(struct callback_head *cb)
5280 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
5281 struct async_poll *apoll = req->apoll;
5282 struct io_ring_ctx *ctx = req->ctx;
5284 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
5286 if (io_poll_rewait(req, &apoll->poll)) {
5287 spin_unlock_irq(&ctx->completion_lock);
5288 percpu_ref_put(&ctx->refs);
5292 /* If req is still hashed, it cannot have been canceled. Don't check. */
5293 if (hash_hashed(&req->hash_node))
5294 hash_del(&req->hash_node);
5296 io_poll_remove_double(req);
5297 spin_unlock_irq(&ctx->completion_lock);
5299 if (!READ_ONCE(apoll->poll.canceled))
5300 __io_req_task_submit(req);
5302 __io_req_task_cancel(req, -ECANCELED);
5304 percpu_ref_put(&ctx->refs);
5305 kfree(apoll->double_poll);
5309 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5312 struct io_kiocb *req = wait->private;
5313 struct io_poll_iocb *poll = &req->apoll->poll;
5315 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5318 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5321 static void io_poll_req_insert(struct io_kiocb *req)
5323 struct io_ring_ctx *ctx = req->ctx;
5324 struct hlist_head *list;
5326 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5327 hlist_add_head(&req->hash_node, list);
5330 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5331 struct io_poll_iocb *poll,
5332 struct io_poll_table *ipt, __poll_t mask,
5333 wait_queue_func_t wake_func)
5334 __acquires(&ctx->completion_lock)
5336 struct io_ring_ctx *ctx = req->ctx;
5337 bool cancel = false;
5339 INIT_HLIST_NODE(&req->hash_node);
5340 io_init_poll_iocb(poll, mask, wake_func);
5341 poll->file = req->file;
5342 poll->wait.private = req;
5344 ipt->pt._key = mask;
5346 ipt->error = -EINVAL;
5348 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5350 spin_lock_irq(&ctx->completion_lock);
5351 if (likely(poll->head)) {
5352 spin_lock(&poll->head->lock);
5353 if (unlikely(list_empty(&poll->wait.entry))) {
5359 if (mask || ipt->error)
5360 list_del_init(&poll->wait.entry);
5362 WRITE_ONCE(poll->canceled, true);
5363 else if (!poll->done) /* actually waiting for an event */
5364 io_poll_req_insert(req);
5365 spin_unlock(&poll->head->lock);
5371 static bool io_arm_poll_handler(struct io_kiocb *req)
5373 const struct io_op_def *def = &io_op_defs[req->opcode];
5374 struct io_ring_ctx *ctx = req->ctx;
5375 struct async_poll *apoll;
5376 struct io_poll_table ipt;
5380 if (!req->file || !file_can_poll(req->file))
5382 if (req->flags & REQ_F_POLLED)
5386 else if (def->pollout)
5390 /* if we can't nonblock try, then no point in arming a poll handler */
5391 if (!io_file_supports_async(req->file, rw))
5394 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5395 if (unlikely(!apoll))
5397 apoll->double_poll = NULL;
5399 req->flags |= REQ_F_POLLED;
5404 mask |= POLLIN | POLLRDNORM;
5406 mask |= POLLOUT | POLLWRNORM;
5408 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5409 if ((req->opcode == IORING_OP_RECVMSG) &&
5410 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5413 mask |= POLLERR | POLLPRI;
5415 ipt.pt._qproc = io_async_queue_proc;
5417 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5419 if (ret || ipt.error) {
5420 io_poll_remove_double(req);
5421 spin_unlock_irq(&ctx->completion_lock);
5422 kfree(apoll->double_poll);
5426 spin_unlock_irq(&ctx->completion_lock);
5427 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
5428 apoll->poll.events);
5432 static bool __io_poll_remove_one(struct io_kiocb *req,
5433 struct io_poll_iocb *poll)
5435 bool do_complete = false;
5437 spin_lock(&poll->head->lock);
5438 WRITE_ONCE(poll->canceled, true);
5439 if (!list_empty(&poll->wait.entry)) {
5440 list_del_init(&poll->wait.entry);
5443 spin_unlock(&poll->head->lock);
5444 hash_del(&req->hash_node);
5448 static bool io_poll_remove_one(struct io_kiocb *req)
5452 io_poll_remove_double(req);
5454 if (req->opcode == IORING_OP_POLL_ADD) {
5455 do_complete = __io_poll_remove_one(req, &req->poll);
5457 struct async_poll *apoll = req->apoll;
5459 /* non-poll requests have submit ref still */
5460 do_complete = __io_poll_remove_one(req, &apoll->poll);
5463 kfree(apoll->double_poll);
5469 io_cqring_fill_event(req, -ECANCELED);
5470 io_commit_cqring(req->ctx);
5471 req_set_fail_links(req);
5472 io_put_req_deferred(req, 1);
5479 * Returns true if we found and killed one or more poll requests
5481 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5482 struct files_struct *files)
5484 struct hlist_node *tmp;
5485 struct io_kiocb *req;
5488 spin_lock_irq(&ctx->completion_lock);
5489 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5490 struct hlist_head *list;
5492 list = &ctx->cancel_hash[i];
5493 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5494 if (io_match_task(req, tsk, files))
5495 posted += io_poll_remove_one(req);
5498 spin_unlock_irq(&ctx->completion_lock);
5501 io_cqring_ev_posted(ctx);
5506 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
5508 struct hlist_head *list;
5509 struct io_kiocb *req;
5511 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5512 hlist_for_each_entry(req, list, hash_node) {
5513 if (sqe_addr != req->user_data)
5515 if (io_poll_remove_one(req))
5523 static int io_poll_remove_prep(struct io_kiocb *req,
5524 const struct io_uring_sqe *sqe)
5526 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5528 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
5532 req->poll_remove.addr = READ_ONCE(sqe->addr);
5537 * Find a running poll command that matches one specified in sqe->addr,
5538 * and remove it if found.
5540 static int io_poll_remove(struct io_kiocb *req)
5542 struct io_ring_ctx *ctx = req->ctx;
5545 spin_lock_irq(&ctx->completion_lock);
5546 ret = io_poll_cancel(ctx, req->poll_remove.addr);
5547 spin_unlock_irq(&ctx->completion_lock);
5550 req_set_fail_links(req);
5551 io_req_complete(req, ret);
5555 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5558 struct io_kiocb *req = wait->private;
5559 struct io_poll_iocb *poll = &req->poll;
5561 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5564 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5565 struct poll_table_struct *p)
5567 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5569 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5572 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5574 struct io_poll_iocb *poll = &req->poll;
5577 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5579 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
5582 events = READ_ONCE(sqe->poll32_events);
5584 events = swahw32(events);
5586 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP |
5587 (events & EPOLLEXCLUSIVE);
5591 static int io_poll_add(struct io_kiocb *req)
5593 struct io_poll_iocb *poll = &req->poll;
5594 struct io_ring_ctx *ctx = req->ctx;
5595 struct io_poll_table ipt;
5598 ipt.pt._qproc = io_poll_queue_proc;
5600 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5603 if (mask) { /* no async, we'd stolen it */
5605 io_poll_complete(req, mask, 0);
5607 spin_unlock_irq(&ctx->completion_lock);
5610 io_cqring_ev_posted(ctx);
5616 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5618 struct io_timeout_data *data = container_of(timer,
5619 struct io_timeout_data, timer);
5620 struct io_kiocb *req = data->req;
5621 struct io_ring_ctx *ctx = req->ctx;
5622 unsigned long flags;
5624 spin_lock_irqsave(&ctx->completion_lock, flags);
5625 list_del_init(&req->timeout.list);
5626 atomic_set(&req->ctx->cq_timeouts,
5627 atomic_read(&req->ctx->cq_timeouts) + 1);
5629 io_cqring_fill_event(req, -ETIME);
5630 io_commit_cqring(ctx);
5631 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5633 io_cqring_ev_posted(ctx);
5634 req_set_fail_links(req);
5636 return HRTIMER_NORESTART;
5639 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5642 struct io_timeout_data *io;
5643 struct io_kiocb *req;
5646 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5647 if (user_data == req->user_data) {
5654 return ERR_PTR(ret);
5656 io = req->async_data;
5657 ret = hrtimer_try_to_cancel(&io->timer);
5659 return ERR_PTR(-EALREADY);
5660 list_del_init(&req->timeout.list);
5664 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5666 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5669 return PTR_ERR(req);
5671 req_set_fail_links(req);
5672 io_cqring_fill_event(req, -ECANCELED);
5673 io_put_req_deferred(req, 1);
5677 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5678 struct timespec64 *ts, enum hrtimer_mode mode)
5680 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5681 struct io_timeout_data *data;
5684 return PTR_ERR(req);
5686 req->timeout.off = 0; /* noseq */
5687 data = req->async_data;
5688 list_add_tail(&req->timeout.list, &ctx->timeout_list);
5689 hrtimer_init(&data->timer, CLOCK_MONOTONIC, mode);
5690 data->timer.function = io_timeout_fn;
5691 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
5695 static int io_timeout_remove_prep(struct io_kiocb *req,
5696 const struct io_uring_sqe *sqe)
5698 struct io_timeout_rem *tr = &req->timeout_rem;
5700 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5702 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5704 if (sqe->ioprio || sqe->buf_index || sqe->len)
5707 tr->addr = READ_ONCE(sqe->addr);
5708 tr->flags = READ_ONCE(sqe->timeout_flags);
5709 if (tr->flags & IORING_TIMEOUT_UPDATE) {
5710 if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS))
5712 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
5714 } else if (tr->flags) {
5715 /* timeout removal doesn't support flags */
5723 * Remove or update an existing timeout command
5725 static int io_timeout_remove(struct io_kiocb *req)
5727 struct io_timeout_rem *tr = &req->timeout_rem;
5728 struct io_ring_ctx *ctx = req->ctx;
5731 spin_lock_irq(&ctx->completion_lock);
5732 if (req->timeout_rem.flags & IORING_TIMEOUT_UPDATE) {
5733 enum hrtimer_mode mode = (tr->flags & IORING_TIMEOUT_ABS)
5734 ? HRTIMER_MODE_ABS : HRTIMER_MODE_REL;
5736 ret = io_timeout_update(ctx, tr->addr, &tr->ts, mode);
5738 ret = io_timeout_cancel(ctx, tr->addr);
5741 io_cqring_fill_event(req, ret);
5742 io_commit_cqring(ctx);
5743 spin_unlock_irq(&ctx->completion_lock);
5744 io_cqring_ev_posted(ctx);
5746 req_set_fail_links(req);
5751 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5752 bool is_timeout_link)
5754 struct io_timeout_data *data;
5756 u32 off = READ_ONCE(sqe->off);
5758 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5760 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5762 if (off && is_timeout_link)
5764 flags = READ_ONCE(sqe->timeout_flags);
5765 if (flags & ~IORING_TIMEOUT_ABS)
5768 req->timeout.off = off;
5770 if (!req->async_data && io_alloc_async_data(req))
5773 data = req->async_data;
5776 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5779 if (flags & IORING_TIMEOUT_ABS)
5780 data->mode = HRTIMER_MODE_ABS;
5782 data->mode = HRTIMER_MODE_REL;
5784 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5788 static int io_timeout(struct io_kiocb *req)
5790 struct io_ring_ctx *ctx = req->ctx;
5791 struct io_timeout_data *data = req->async_data;
5792 struct list_head *entry;
5793 u32 tail, off = req->timeout.off;
5795 spin_lock_irq(&ctx->completion_lock);
5798 * sqe->off holds how many events that need to occur for this
5799 * timeout event to be satisfied. If it isn't set, then this is
5800 * a pure timeout request, sequence isn't used.
5802 if (io_is_timeout_noseq(req)) {
5803 entry = ctx->timeout_list.prev;
5807 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5808 req->timeout.target_seq = tail + off;
5811 * Insertion sort, ensuring the first entry in the list is always
5812 * the one we need first.
5814 list_for_each_prev(entry, &ctx->timeout_list) {
5815 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5818 if (io_is_timeout_noseq(nxt))
5820 /* nxt.seq is behind @tail, otherwise would've been completed */
5821 if (off >= nxt->timeout.target_seq - tail)
5825 list_add(&req->timeout.list, entry);
5826 data->timer.function = io_timeout_fn;
5827 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5828 spin_unlock_irq(&ctx->completion_lock);
5832 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5834 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5836 return req->user_data == (unsigned long) data;
5839 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
5841 enum io_wq_cancel cancel_ret;
5844 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr, false);
5845 switch (cancel_ret) {
5846 case IO_WQ_CANCEL_OK:
5849 case IO_WQ_CANCEL_RUNNING:
5852 case IO_WQ_CANCEL_NOTFOUND:
5860 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5861 struct io_kiocb *req, __u64 sqe_addr,
5864 unsigned long flags;
5867 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
5868 if (ret != -ENOENT) {
5869 spin_lock_irqsave(&ctx->completion_lock, flags);
5873 spin_lock_irqsave(&ctx->completion_lock, flags);
5874 ret = io_timeout_cancel(ctx, sqe_addr);
5877 ret = io_poll_cancel(ctx, sqe_addr);
5881 io_cqring_fill_event(req, ret);
5882 io_commit_cqring(ctx);
5883 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5884 io_cqring_ev_posted(ctx);
5887 req_set_fail_links(req);
5891 static int io_async_cancel_prep(struct io_kiocb *req,
5892 const struct io_uring_sqe *sqe)
5894 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5896 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5898 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5901 req->cancel.addr = READ_ONCE(sqe->addr);
5905 static int io_async_cancel(struct io_kiocb *req)
5907 struct io_ring_ctx *ctx = req->ctx;
5909 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
5913 static int io_files_update_prep(struct io_kiocb *req,
5914 const struct io_uring_sqe *sqe)
5916 if (unlikely(req->ctx->flags & IORING_SETUP_SQPOLL))
5918 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5920 if (sqe->ioprio || sqe->rw_flags)
5923 req->files_update.offset = READ_ONCE(sqe->off);
5924 req->files_update.nr_args = READ_ONCE(sqe->len);
5925 if (!req->files_update.nr_args)
5927 req->files_update.arg = READ_ONCE(sqe->addr);
5931 static int io_files_update(struct io_kiocb *req, bool force_nonblock,
5932 struct io_comp_state *cs)
5934 struct io_ring_ctx *ctx = req->ctx;
5935 struct io_uring_files_update up;
5941 up.offset = req->files_update.offset;
5942 up.fds = req->files_update.arg;
5944 mutex_lock(&ctx->uring_lock);
5945 ret = __io_sqe_files_update(ctx, &up, req->files_update.nr_args);
5946 mutex_unlock(&ctx->uring_lock);
5949 req_set_fail_links(req);
5950 __io_req_complete(req, ret, 0, cs);
5954 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5956 switch (req->opcode) {
5959 case IORING_OP_READV:
5960 case IORING_OP_READ_FIXED:
5961 case IORING_OP_READ:
5962 return io_read_prep(req, sqe);
5963 case IORING_OP_WRITEV:
5964 case IORING_OP_WRITE_FIXED:
5965 case IORING_OP_WRITE:
5966 return io_write_prep(req, sqe);
5967 case IORING_OP_POLL_ADD:
5968 return io_poll_add_prep(req, sqe);
5969 case IORING_OP_POLL_REMOVE:
5970 return io_poll_remove_prep(req, sqe);
5971 case IORING_OP_FSYNC:
5972 return io_prep_fsync(req, sqe);
5973 case IORING_OP_SYNC_FILE_RANGE:
5974 return io_prep_sfr(req, sqe);
5975 case IORING_OP_SENDMSG:
5976 case IORING_OP_SEND:
5977 return io_sendmsg_prep(req, sqe);
5978 case IORING_OP_RECVMSG:
5979 case IORING_OP_RECV:
5980 return io_recvmsg_prep(req, sqe);
5981 case IORING_OP_CONNECT:
5982 return io_connect_prep(req, sqe);
5983 case IORING_OP_TIMEOUT:
5984 return io_timeout_prep(req, sqe, false);
5985 case IORING_OP_TIMEOUT_REMOVE:
5986 return io_timeout_remove_prep(req, sqe);
5987 case IORING_OP_ASYNC_CANCEL:
5988 return io_async_cancel_prep(req, sqe);
5989 case IORING_OP_LINK_TIMEOUT:
5990 return io_timeout_prep(req, sqe, true);
5991 case IORING_OP_ACCEPT:
5992 return io_accept_prep(req, sqe);
5993 case IORING_OP_FALLOCATE:
5994 return io_fallocate_prep(req, sqe);
5995 case IORING_OP_OPENAT:
5996 return io_openat_prep(req, sqe);
5997 case IORING_OP_CLOSE:
5998 return io_close_prep(req, sqe);
5999 case IORING_OP_FILES_UPDATE:
6000 return io_files_update_prep(req, sqe);
6001 case IORING_OP_STATX:
6002 return io_statx_prep(req, sqe);
6003 case IORING_OP_FADVISE:
6004 return io_fadvise_prep(req, sqe);
6005 case IORING_OP_MADVISE:
6006 return io_madvise_prep(req, sqe);
6007 case IORING_OP_OPENAT2:
6008 return io_openat2_prep(req, sqe);
6009 case IORING_OP_EPOLL_CTL:
6010 return io_epoll_ctl_prep(req, sqe);
6011 case IORING_OP_SPLICE:
6012 return io_splice_prep(req, sqe);
6013 case IORING_OP_PROVIDE_BUFFERS:
6014 return io_provide_buffers_prep(req, sqe);
6015 case IORING_OP_REMOVE_BUFFERS:
6016 return io_remove_buffers_prep(req, sqe);
6018 return io_tee_prep(req, sqe);
6019 case IORING_OP_SHUTDOWN:
6020 return io_shutdown_prep(req, sqe);
6021 case IORING_OP_RENAMEAT:
6022 return io_renameat_prep(req, sqe);
6023 case IORING_OP_UNLINKAT:
6024 return io_unlinkat_prep(req, sqe);
6027 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
6032 static int io_req_defer_prep(struct io_kiocb *req,
6033 const struct io_uring_sqe *sqe)
6037 if (io_alloc_async_data(req))
6039 return io_req_prep(req, sqe);
6042 static u32 io_get_sequence(struct io_kiocb *req)
6044 struct io_kiocb *pos;
6045 struct io_ring_ctx *ctx = req->ctx;
6046 u32 total_submitted, nr_reqs = 0;
6048 io_for_each_link(pos, req)
6051 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
6052 return total_submitted - nr_reqs;
6055 static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe)
6057 struct io_ring_ctx *ctx = req->ctx;
6058 struct io_defer_entry *de;
6062 /* Still need defer if there is pending req in defer list. */
6063 if (likely(list_empty_careful(&ctx->defer_list) &&
6064 !(req->flags & REQ_F_IO_DRAIN)))
6067 seq = io_get_sequence(req);
6068 /* Still a chance to pass the sequence check */
6069 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
6072 if (!req->async_data) {
6073 ret = io_req_defer_prep(req, sqe);
6077 io_prep_async_link(req);
6078 de = kmalloc(sizeof(*de), GFP_KERNEL);
6082 spin_lock_irq(&ctx->completion_lock);
6083 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
6084 spin_unlock_irq(&ctx->completion_lock);
6086 io_queue_async_work(req);
6087 return -EIOCBQUEUED;
6090 trace_io_uring_defer(ctx, req, req->user_data);
6093 list_add_tail(&de->list, &ctx->defer_list);
6094 spin_unlock_irq(&ctx->completion_lock);
6095 return -EIOCBQUEUED;
6098 static void io_req_drop_files(struct io_kiocb *req)
6100 struct io_ring_ctx *ctx = req->ctx;
6101 struct io_uring_task *tctx = req->task->io_uring;
6102 unsigned long flags;
6104 put_files_struct(req->work.identity->files);
6105 put_nsproxy(req->work.identity->nsproxy);
6106 spin_lock_irqsave(&ctx->inflight_lock, flags);
6107 list_del(&req->inflight_entry);
6108 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
6109 req->flags &= ~REQ_F_INFLIGHT;
6110 req->work.flags &= ~IO_WQ_WORK_FILES;
6111 if (atomic_read(&tctx->in_idle))
6112 wake_up(&tctx->wait);
6115 static void __io_clean_op(struct io_kiocb *req)
6117 if (req->flags & REQ_F_BUFFER_SELECTED) {
6118 switch (req->opcode) {
6119 case IORING_OP_READV:
6120 case IORING_OP_READ_FIXED:
6121 case IORING_OP_READ:
6122 kfree((void *)(unsigned long)req->rw.addr);
6124 case IORING_OP_RECVMSG:
6125 case IORING_OP_RECV:
6126 kfree(req->sr_msg.kbuf);
6129 req->flags &= ~REQ_F_BUFFER_SELECTED;
6132 if (req->flags & REQ_F_NEED_CLEANUP) {
6133 switch (req->opcode) {
6134 case IORING_OP_READV:
6135 case IORING_OP_READ_FIXED:
6136 case IORING_OP_READ:
6137 case IORING_OP_WRITEV:
6138 case IORING_OP_WRITE_FIXED:
6139 case IORING_OP_WRITE: {
6140 struct io_async_rw *io = req->async_data;
6142 kfree(io->free_iovec);
6145 case IORING_OP_RECVMSG:
6146 case IORING_OP_SENDMSG: {
6147 struct io_async_msghdr *io = req->async_data;
6148 if (io->iov != io->fast_iov)
6152 case IORING_OP_SPLICE:
6154 io_put_file(req, req->splice.file_in,
6155 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
6157 case IORING_OP_OPENAT:
6158 case IORING_OP_OPENAT2:
6159 if (req->open.filename)
6160 putname(req->open.filename);
6162 case IORING_OP_RENAMEAT:
6163 putname(req->rename.oldpath);
6164 putname(req->rename.newpath);
6166 case IORING_OP_UNLINKAT:
6167 putname(req->unlink.filename);
6170 req->flags &= ~REQ_F_NEED_CLEANUP;
6173 if (req->flags & REQ_F_INFLIGHT)
6174 io_req_drop_files(req);
6177 static int io_issue_sqe(struct io_kiocb *req, bool force_nonblock,
6178 struct io_comp_state *cs)
6180 struct io_ring_ctx *ctx = req->ctx;
6183 switch (req->opcode) {
6185 ret = io_nop(req, cs);
6187 case IORING_OP_READV:
6188 case IORING_OP_READ_FIXED:
6189 case IORING_OP_READ:
6190 ret = io_read(req, force_nonblock, cs);
6192 case IORING_OP_WRITEV:
6193 case IORING_OP_WRITE_FIXED:
6194 case IORING_OP_WRITE:
6195 ret = io_write(req, force_nonblock, cs);
6197 case IORING_OP_FSYNC:
6198 ret = io_fsync(req, force_nonblock);
6200 case IORING_OP_POLL_ADD:
6201 ret = io_poll_add(req);
6203 case IORING_OP_POLL_REMOVE:
6204 ret = io_poll_remove(req);
6206 case IORING_OP_SYNC_FILE_RANGE:
6207 ret = io_sync_file_range(req, force_nonblock);
6209 case IORING_OP_SENDMSG:
6210 ret = io_sendmsg(req, force_nonblock, cs);
6212 case IORING_OP_SEND:
6213 ret = io_send(req, force_nonblock, cs);
6215 case IORING_OP_RECVMSG:
6216 ret = io_recvmsg(req, force_nonblock, cs);
6218 case IORING_OP_RECV:
6219 ret = io_recv(req, force_nonblock, cs);
6221 case IORING_OP_TIMEOUT:
6222 ret = io_timeout(req);
6224 case IORING_OP_TIMEOUT_REMOVE:
6225 ret = io_timeout_remove(req);
6227 case IORING_OP_ACCEPT:
6228 ret = io_accept(req, force_nonblock, cs);
6230 case IORING_OP_CONNECT:
6231 ret = io_connect(req, force_nonblock, cs);
6233 case IORING_OP_ASYNC_CANCEL:
6234 ret = io_async_cancel(req);
6236 case IORING_OP_FALLOCATE:
6237 ret = io_fallocate(req, force_nonblock);
6239 case IORING_OP_OPENAT:
6240 ret = io_openat(req, force_nonblock);
6242 case IORING_OP_CLOSE:
6243 ret = io_close(req, force_nonblock, cs);
6245 case IORING_OP_FILES_UPDATE:
6246 ret = io_files_update(req, force_nonblock, cs);
6248 case IORING_OP_STATX:
6249 ret = io_statx(req, force_nonblock);
6251 case IORING_OP_FADVISE:
6252 ret = io_fadvise(req, force_nonblock);
6254 case IORING_OP_MADVISE:
6255 ret = io_madvise(req, force_nonblock);
6257 case IORING_OP_OPENAT2:
6258 ret = io_openat2(req, force_nonblock);
6260 case IORING_OP_EPOLL_CTL:
6261 ret = io_epoll_ctl(req, force_nonblock, cs);
6263 case IORING_OP_SPLICE:
6264 ret = io_splice(req, force_nonblock);
6266 case IORING_OP_PROVIDE_BUFFERS:
6267 ret = io_provide_buffers(req, force_nonblock, cs);
6269 case IORING_OP_REMOVE_BUFFERS:
6270 ret = io_remove_buffers(req, force_nonblock, cs);
6273 ret = io_tee(req, force_nonblock);
6275 case IORING_OP_SHUTDOWN:
6276 ret = io_shutdown(req, force_nonblock);
6278 case IORING_OP_RENAMEAT:
6279 ret = io_renameat(req, force_nonblock);
6281 case IORING_OP_UNLINKAT:
6282 ret = io_unlinkat(req, force_nonblock);
6292 /* If the op doesn't have a file, we're not polling for it */
6293 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
6294 const bool in_async = io_wq_current_is_worker();
6296 /* workqueue context doesn't hold uring_lock, grab it now */
6298 mutex_lock(&ctx->uring_lock);
6300 io_iopoll_req_issued(req, in_async);
6303 mutex_unlock(&ctx->uring_lock);
6309 static struct io_wq_work *io_wq_submit_work(struct io_wq_work *work)
6311 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6312 struct io_kiocb *timeout;
6315 timeout = io_prep_linked_timeout(req);
6317 io_queue_linked_timeout(timeout);
6319 /* if NO_CANCEL is set, we must still run the work */
6320 if ((work->flags & (IO_WQ_WORK_CANCEL|IO_WQ_WORK_NO_CANCEL)) ==
6321 IO_WQ_WORK_CANCEL) {
6327 ret = io_issue_sqe(req, false, NULL);
6329 * We can get EAGAIN for polled IO even though we're
6330 * forcing a sync submission from here, since we can't
6331 * wait for request slots on the block side.
6340 struct io_ring_ctx *lock_ctx = NULL;
6342 if (req->ctx->flags & IORING_SETUP_IOPOLL)
6343 lock_ctx = req->ctx;
6346 * io_iopoll_complete() does not hold completion_lock to
6347 * complete polled io, so here for polled io, we can not call
6348 * io_req_complete() directly, otherwise there maybe concurrent
6349 * access to cqring, defer_list, etc, which is not safe. Given
6350 * that io_iopoll_complete() is always called under uring_lock,
6351 * so here for polled io, we also get uring_lock to complete
6355 mutex_lock(&lock_ctx->uring_lock);
6357 req_set_fail_links(req);
6358 io_req_complete(req, ret);
6361 mutex_unlock(&lock_ctx->uring_lock);
6364 return io_steal_work(req);
6367 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6370 struct fixed_file_table *table;
6372 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
6373 return table->files[index & IORING_FILE_TABLE_MASK];
6376 static struct file *io_file_get(struct io_submit_state *state,
6377 struct io_kiocb *req, int fd, bool fixed)
6379 struct io_ring_ctx *ctx = req->ctx;
6383 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6385 fd = array_index_nospec(fd, ctx->nr_user_files);
6386 file = io_file_from_index(ctx, fd);
6387 io_set_resource_node(req);
6389 trace_io_uring_file_get(ctx, fd);
6390 file = __io_file_get(state, fd);
6396 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6398 struct io_timeout_data *data = container_of(timer,
6399 struct io_timeout_data, timer);
6400 struct io_kiocb *prev, *req = data->req;
6401 struct io_ring_ctx *ctx = req->ctx;
6402 unsigned long flags;
6404 spin_lock_irqsave(&ctx->completion_lock, flags);
6405 prev = req->timeout.head;
6406 req->timeout.head = NULL;
6409 * We don't expect the list to be empty, that will only happen if we
6410 * race with the completion of the linked work.
6412 if (prev && refcount_inc_not_zero(&prev->refs))
6413 io_remove_next_linked(prev);
6416 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6419 req_set_fail_links(prev);
6420 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6423 io_req_complete(req, -ETIME);
6425 return HRTIMER_NORESTART;
6428 static void __io_queue_linked_timeout(struct io_kiocb *req)
6431 * If the back reference is NULL, then our linked request finished
6432 * before we got a chance to setup the timer
6434 if (req->timeout.head) {
6435 struct io_timeout_data *data = req->async_data;
6437 data->timer.function = io_link_timeout_fn;
6438 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6443 static void io_queue_linked_timeout(struct io_kiocb *req)
6445 struct io_ring_ctx *ctx = req->ctx;
6447 spin_lock_irq(&ctx->completion_lock);
6448 __io_queue_linked_timeout(req);
6449 spin_unlock_irq(&ctx->completion_lock);
6451 /* drop submission reference */
6455 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6457 struct io_kiocb *nxt = req->link;
6459 if (!nxt || (req->flags & REQ_F_LINK_TIMEOUT) ||
6460 nxt->opcode != IORING_OP_LINK_TIMEOUT)
6463 nxt->timeout.head = req;
6464 nxt->flags |= REQ_F_LTIMEOUT_ACTIVE;
6465 req->flags |= REQ_F_LINK_TIMEOUT;
6469 static void __io_queue_sqe(struct io_kiocb *req, struct io_comp_state *cs)
6471 struct io_kiocb *linked_timeout;
6472 const struct cred *old_creds = NULL;
6476 linked_timeout = io_prep_linked_timeout(req);
6478 if ((req->flags & REQ_F_WORK_INITIALIZED) &&
6479 (req->work.flags & IO_WQ_WORK_CREDS) &&
6480 req->work.identity->creds != current_cred()) {
6482 revert_creds(old_creds);
6483 if (old_creds == req->work.identity->creds)
6484 old_creds = NULL; /* restored original creds */
6486 old_creds = override_creds(req->work.identity->creds);
6489 ret = io_issue_sqe(req, true, cs);
6492 * We async punt it if the file wasn't marked NOWAIT, or if the file
6493 * doesn't support non-blocking read/write attempts
6495 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6496 if (!io_arm_poll_handler(req)) {
6498 * Queued up for async execution, worker will release
6499 * submit reference when the iocb is actually submitted.
6501 io_queue_async_work(req);
6505 io_queue_linked_timeout(linked_timeout);
6506 } else if (likely(!ret)) {
6507 /* drop submission reference */
6508 req = io_put_req_find_next(req);
6510 io_queue_linked_timeout(linked_timeout);
6513 if (!(req->flags & REQ_F_FORCE_ASYNC))
6515 io_queue_async_work(req);
6518 /* un-prep timeout, so it'll be killed as any other linked */
6519 req->flags &= ~REQ_F_LINK_TIMEOUT;
6520 req_set_fail_links(req);
6522 io_req_complete(req, ret);
6526 revert_creds(old_creds);
6529 static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6530 struct io_comp_state *cs)
6534 ret = io_req_defer(req, sqe);
6536 if (ret != -EIOCBQUEUED) {
6538 req_set_fail_links(req);
6540 io_req_complete(req, ret);
6542 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6543 if (!req->async_data) {
6544 ret = io_req_defer_prep(req, sqe);
6548 io_queue_async_work(req);
6551 ret = io_req_prep(req, sqe);
6555 __io_queue_sqe(req, cs);
6559 static inline void io_queue_link_head(struct io_kiocb *req,
6560 struct io_comp_state *cs)
6562 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
6564 io_req_complete(req, -ECANCELED);
6566 io_queue_sqe(req, NULL, cs);
6569 struct io_submit_link {
6570 struct io_kiocb *head;
6571 struct io_kiocb *last;
6574 static int io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6575 struct io_submit_link *link, struct io_comp_state *cs)
6577 struct io_ring_ctx *ctx = req->ctx;
6581 * If we already have a head request, queue this one for async
6582 * submittal once the head completes. If we don't have a head but
6583 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6584 * submitted sync once the chain is complete. If none of those
6585 * conditions are true (normal request), then just queue it.
6588 struct io_kiocb *head = link->head;
6591 * Taking sequential execution of a link, draining both sides
6592 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6593 * requests in the link. So, it drains the head and the
6594 * next after the link request. The last one is done via
6595 * drain_next flag to persist the effect across calls.
6597 if (req->flags & REQ_F_IO_DRAIN) {
6598 head->flags |= REQ_F_IO_DRAIN;
6599 ctx->drain_next = 1;
6601 ret = io_req_defer_prep(req, sqe);
6602 if (unlikely(ret)) {
6603 /* fail even hard links since we don't submit */
6604 head->flags |= REQ_F_FAIL_LINK;
6607 trace_io_uring_link(ctx, req, head);
6608 link->last->link = req;
6611 /* last request of a link, enqueue the link */
6612 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6613 io_queue_link_head(head, cs);
6617 if (unlikely(ctx->drain_next)) {
6618 req->flags |= REQ_F_IO_DRAIN;
6619 ctx->drain_next = 0;
6621 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6622 ret = io_req_defer_prep(req, sqe);
6624 req->flags |= REQ_F_FAIL_LINK;
6628 io_queue_sqe(req, sqe, cs);
6636 * Batched submission is done, ensure local IO is flushed out.
6638 static void io_submit_state_end(struct io_submit_state *state)
6640 if (!list_empty(&state->comp.list))
6641 io_submit_flush_completions(&state->comp);
6642 if (state->plug_started)
6643 blk_finish_plug(&state->plug);
6644 io_state_file_put(state);
6645 if (state->free_reqs)
6646 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
6650 * Start submission side cache.
6652 static void io_submit_state_start(struct io_submit_state *state,
6653 struct io_ring_ctx *ctx, unsigned int max_ios)
6655 state->plug_started = false;
6657 INIT_LIST_HEAD(&state->comp.list);
6658 state->comp.ctx = ctx;
6659 state->free_reqs = 0;
6660 state->file_refs = 0;
6661 state->ios_left = max_ios;
6664 static void io_commit_sqring(struct io_ring_ctx *ctx)
6666 struct io_rings *rings = ctx->rings;
6669 * Ensure any loads from the SQEs are done at this point,
6670 * since once we write the new head, the application could
6671 * write new data to them.
6673 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6677 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6678 * that is mapped by userspace. This means that care needs to be taken to
6679 * ensure that reads are stable, as we cannot rely on userspace always
6680 * being a good citizen. If members of the sqe are validated and then later
6681 * used, it's important that those reads are done through READ_ONCE() to
6682 * prevent a re-load down the line.
6684 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6686 u32 *sq_array = ctx->sq_array;
6690 * The cached sq head (or cq tail) serves two purposes:
6692 * 1) allows us to batch the cost of updating the user visible
6694 * 2) allows the kernel side to track the head on its own, even
6695 * though the application is the one updating it.
6697 head = READ_ONCE(sq_array[ctx->cached_sq_head & ctx->sq_mask]);
6698 if (likely(head < ctx->sq_entries))
6699 return &ctx->sq_sqes[head];
6701 /* drop invalid entries */
6702 ctx->cached_sq_dropped++;
6703 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6707 static inline void io_consume_sqe(struct io_ring_ctx *ctx)
6709 ctx->cached_sq_head++;
6713 * Check SQE restrictions (opcode and flags).
6715 * Returns 'true' if SQE is allowed, 'false' otherwise.
6717 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6718 struct io_kiocb *req,
6719 unsigned int sqe_flags)
6721 if (!ctx->restricted)
6724 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6727 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6728 ctx->restrictions.sqe_flags_required)
6731 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6732 ctx->restrictions.sqe_flags_required))
6738 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
6739 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
6740 IOSQE_BUFFER_SELECT)
6742 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6743 const struct io_uring_sqe *sqe,
6744 struct io_submit_state *state)
6746 unsigned int sqe_flags;
6749 req->opcode = READ_ONCE(sqe->opcode);
6750 req->user_data = READ_ONCE(sqe->user_data);
6751 req->async_data = NULL;
6756 req->fixed_file_refs = NULL;
6757 /* one is dropped after submission, the other at completion */
6758 refcount_set(&req->refs, 2);
6759 req->task = current;
6762 if (unlikely(req->opcode >= IORING_OP_LAST))
6765 if (unlikely(io_sq_thread_acquire_mm_files(ctx, req)))
6768 sqe_flags = READ_ONCE(sqe->flags);
6769 /* enforce forwards compatibility on users */
6770 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6773 if (unlikely(!io_check_restriction(ctx, req, sqe_flags)))
6776 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6777 !io_op_defs[req->opcode].buffer_select)
6780 id = READ_ONCE(sqe->personality);
6782 struct io_identity *iod;
6784 iod = idr_find(&ctx->personality_idr, id);
6787 refcount_inc(&iod->count);
6789 __io_req_init_async(req);
6790 get_cred(iod->creds);
6791 req->work.identity = iod;
6792 req->work.flags |= IO_WQ_WORK_CREDS;
6795 /* same numerical values with corresponding REQ_F_*, safe to copy */
6796 req->flags |= sqe_flags;
6799 * Plug now if we have more than 1 IO left after this, and the target
6800 * is potentially a read/write to block based storage.
6802 if (!state->plug_started && state->ios_left > 1 &&
6803 io_op_defs[req->opcode].plug) {
6804 blk_start_plug(&state->plug);
6805 state->plug_started = true;
6809 if (io_op_defs[req->opcode].needs_file) {
6810 bool fixed = req->flags & REQ_F_FIXED_FILE;
6812 req->file = io_file_get(state, req, READ_ONCE(sqe->fd), fixed);
6813 if (unlikely(!req->file &&
6814 !io_op_defs[req->opcode].needs_file_no_error))
6822 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6824 struct io_submit_state state;
6825 struct io_submit_link link;
6826 int i, submitted = 0;
6828 /* if we have a backlog and couldn't flush it all, return BUSY */
6829 if (test_bit(0, &ctx->sq_check_overflow)) {
6830 if (!io_cqring_overflow_flush(ctx, false, NULL, NULL))
6834 /* make sure SQ entry isn't read before tail */
6835 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6837 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6840 percpu_counter_add(¤t->io_uring->inflight, nr);
6841 refcount_add(nr, ¤t->usage);
6843 io_submit_state_start(&state, ctx, nr);
6846 for (i = 0; i < nr; i++) {
6847 const struct io_uring_sqe *sqe;
6848 struct io_kiocb *req;
6851 sqe = io_get_sqe(ctx);
6852 if (unlikely(!sqe)) {
6853 io_consume_sqe(ctx);
6856 req = io_alloc_req(ctx, &state);
6857 if (unlikely(!req)) {
6859 submitted = -EAGAIN;
6862 io_consume_sqe(ctx);
6863 /* will complete beyond this point, count as submitted */
6866 err = io_init_req(ctx, req, sqe, &state);
6867 if (unlikely(err)) {
6870 io_req_complete(req, err);
6874 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6875 true, io_async_submit(ctx));
6876 err = io_submit_sqe(req, sqe, &link, &state.comp);
6881 if (unlikely(submitted != nr)) {
6882 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6883 struct io_uring_task *tctx = current->io_uring;
6884 int unused = nr - ref_used;
6886 percpu_ref_put_many(&ctx->refs, unused);
6887 percpu_counter_sub(&tctx->inflight, unused);
6888 put_task_struct_many(current, unused);
6891 io_queue_link_head(link.head, &state.comp);
6892 io_submit_state_end(&state);
6894 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6895 io_commit_sqring(ctx);
6900 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6902 /* Tell userspace we may need a wakeup call */
6903 spin_lock_irq(&ctx->completion_lock);
6904 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6905 spin_unlock_irq(&ctx->completion_lock);
6908 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6910 spin_lock_irq(&ctx->completion_lock);
6911 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6912 spin_unlock_irq(&ctx->completion_lock);
6915 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
6917 unsigned int to_submit;
6920 to_submit = io_sqring_entries(ctx);
6921 /* if we're handling multiple rings, cap submit size for fairness */
6922 if (cap_entries && to_submit > 8)
6925 if (!list_empty(&ctx->iopoll_list) || to_submit) {
6926 unsigned nr_events = 0;
6928 mutex_lock(&ctx->uring_lock);
6929 if (!list_empty(&ctx->iopoll_list))
6930 io_do_iopoll(ctx, &nr_events, 0);
6932 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)))
6933 ret = io_submit_sqes(ctx, to_submit);
6934 mutex_unlock(&ctx->uring_lock);
6937 if (!io_sqring_full(ctx) && wq_has_sleeper(&ctx->sqo_sq_wait))
6938 wake_up(&ctx->sqo_sq_wait);
6943 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
6945 struct io_ring_ctx *ctx;
6946 unsigned sq_thread_idle = 0;
6948 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6949 if (sq_thread_idle < ctx->sq_thread_idle)
6950 sq_thread_idle = ctx->sq_thread_idle;
6953 sqd->sq_thread_idle = sq_thread_idle;
6956 static void io_sqd_init_new(struct io_sq_data *sqd)
6958 struct io_ring_ctx *ctx;
6960 while (!list_empty(&sqd->ctx_new_list)) {
6961 ctx = list_first_entry(&sqd->ctx_new_list, struct io_ring_ctx, sqd_list);
6962 list_move_tail(&ctx->sqd_list, &sqd->ctx_list);
6963 complete(&ctx->sq_thread_comp);
6966 io_sqd_update_thread_idle(sqd);
6969 static int io_sq_thread(void *data)
6971 struct cgroup_subsys_state *cur_css = NULL;
6972 struct files_struct *old_files = current->files;
6973 struct nsproxy *old_nsproxy = current->nsproxy;
6974 const struct cred *old_cred = NULL;
6975 struct io_sq_data *sqd = data;
6976 struct io_ring_ctx *ctx;
6977 unsigned long timeout = 0;
6981 current->files = NULL;
6982 current->nsproxy = NULL;
6983 task_unlock(current);
6985 while (!kthread_should_stop()) {
6987 bool cap_entries, sqt_spin, needs_sched;
6990 * Any changes to the sqd lists are synchronized through the
6991 * kthread parking. This synchronizes the thread vs users,
6992 * the users are synchronized on the sqd->ctx_lock.
6994 if (kthread_should_park()) {
6997 * When sq thread is unparked, in case the previous park operation
6998 * comes from io_put_sq_data(), which means that sq thread is going
6999 * to be stopped, so here needs to have a check.
7001 if (kthread_should_stop())
7005 if (unlikely(!list_empty(&sqd->ctx_new_list))) {
7006 io_sqd_init_new(sqd);
7007 timeout = jiffies + sqd->sq_thread_idle;
7011 cap_entries = !list_is_singular(&sqd->ctx_list);
7012 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
7013 if (current->cred != ctx->creds) {
7015 revert_creds(old_cred);
7016 old_cred = override_creds(ctx->creds);
7018 io_sq_thread_associate_blkcg(ctx, &cur_css);
7020 current->loginuid = ctx->loginuid;
7021 current->sessionid = ctx->sessionid;
7024 ret = __io_sq_thread(ctx, cap_entries);
7025 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
7028 io_sq_thread_drop_mm_files();
7031 if (sqt_spin || !time_after(jiffies, timeout)) {
7035 timeout = jiffies + sqd->sq_thread_idle;
7039 if (kthread_should_park())
7043 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
7044 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
7045 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
7046 !list_empty_careful(&ctx->iopoll_list)) {
7047 needs_sched = false;
7050 if (io_sqring_entries(ctx)) {
7051 needs_sched = false;
7057 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7058 io_ring_set_wakeup_flag(ctx);
7061 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7062 io_ring_clear_wakeup_flag(ctx);
7065 finish_wait(&sqd->wait, &wait);
7066 timeout = jiffies + sqd->sq_thread_idle;
7072 io_sq_thread_unassociate_blkcg();
7074 revert_creds(old_cred);
7077 current->files = old_files;
7078 current->nsproxy = old_nsproxy;
7079 task_unlock(current);
7086 struct io_wait_queue {
7087 struct wait_queue_entry wq;
7088 struct io_ring_ctx *ctx;
7090 unsigned nr_timeouts;
7093 static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush)
7095 struct io_ring_ctx *ctx = iowq->ctx;
7098 * Wake up if we have enough events, or if a timeout occurred since we
7099 * started waiting. For timeouts, we always want to return to userspace,
7100 * regardless of event count.
7102 return io_cqring_events(ctx, noflush) >= iowq->to_wait ||
7103 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
7106 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
7107 int wake_flags, void *key)
7109 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
7112 /* use noflush == true, as we can't safely rely on locking context */
7113 if (!io_should_wake(iowq, true))
7116 return autoremove_wake_function(curr, mode, wake_flags, key);
7119 static int io_run_task_work_sig(void)
7121 if (io_run_task_work())
7123 if (!signal_pending(current))
7125 if (test_tsk_thread_flag(current, TIF_NOTIFY_SIGNAL))
7126 return -ERESTARTSYS;
7131 * Wait until events become available, if we don't already have some. The
7132 * application must reap them itself, as they reside on the shared cq ring.
7134 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
7135 const sigset_t __user *sig, size_t sigsz,
7136 struct __kernel_timespec __user *uts)
7138 struct io_wait_queue iowq = {
7141 .func = io_wake_function,
7142 .entry = LIST_HEAD_INIT(iowq.wq.entry),
7145 .to_wait = min_events,
7147 struct io_rings *rings = ctx->rings;
7148 struct timespec64 ts;
7149 signed long timeout = 0;
7153 if (io_cqring_events(ctx, false) >= min_events)
7155 if (!io_run_task_work())
7160 #ifdef CONFIG_COMPAT
7161 if (in_compat_syscall())
7162 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
7166 ret = set_user_sigmask(sig, sigsz);
7173 if (get_timespec64(&ts, uts))
7175 timeout = timespec64_to_jiffies(&ts);
7178 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
7179 trace_io_uring_cqring_wait(ctx, min_events);
7181 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
7182 TASK_INTERRUPTIBLE);
7183 /* make sure we run task_work before checking for signals */
7184 ret = io_run_task_work_sig();
7189 if (io_should_wake(&iowq, false))
7192 timeout = schedule_timeout(timeout);
7201 finish_wait(&ctx->wait, &iowq.wq);
7203 restore_saved_sigmask_unless(ret == -EINTR);
7205 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
7208 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
7210 #if defined(CONFIG_UNIX)
7211 if (ctx->ring_sock) {
7212 struct sock *sock = ctx->ring_sock->sk;
7213 struct sk_buff *skb;
7215 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
7221 for (i = 0; i < ctx->nr_user_files; i++) {
7224 file = io_file_from_index(ctx, i);
7231 static void io_file_ref_kill(struct percpu_ref *ref)
7233 struct fixed_file_data *data;
7235 data = container_of(ref, struct fixed_file_data, refs);
7236 complete(&data->done);
7239 static void io_sqe_files_set_node(struct fixed_file_data *file_data,
7240 struct fixed_file_ref_node *ref_node)
7242 spin_lock_bh(&file_data->lock);
7243 file_data->node = ref_node;
7244 list_add_tail(&ref_node->node, &file_data->ref_list);
7245 spin_unlock_bh(&file_data->lock);
7246 percpu_ref_get(&file_data->refs);
7249 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7251 struct fixed_file_data *data = ctx->file_data;
7252 struct fixed_file_ref_node *backup_node, *ref_node = NULL;
7253 unsigned nr_tables, i;
7258 backup_node = alloc_fixed_file_ref_node(ctx);
7262 spin_lock_bh(&data->lock);
7263 ref_node = data->node;
7264 spin_unlock_bh(&data->lock);
7266 percpu_ref_kill(&ref_node->refs);
7268 percpu_ref_kill(&data->refs);
7270 /* wait for all refs nodes to complete */
7271 flush_delayed_work(&ctx->file_put_work);
7273 ret = wait_for_completion_interruptible(&data->done);
7276 ret = io_run_task_work_sig();
7278 percpu_ref_resurrect(&data->refs);
7279 reinit_completion(&data->done);
7280 io_sqe_files_set_node(data, backup_node);
7285 __io_sqe_files_unregister(ctx);
7286 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
7287 for (i = 0; i < nr_tables; i++)
7288 kfree(data->table[i].files);
7290 percpu_ref_exit(&data->refs);
7292 ctx->file_data = NULL;
7293 ctx->nr_user_files = 0;
7294 destroy_fixed_file_ref_node(backup_node);
7298 static void io_put_sq_data(struct io_sq_data *sqd)
7300 if (refcount_dec_and_test(&sqd->refs)) {
7302 * The park is a bit of a work-around, without it we get
7303 * warning spews on shutdown with SQPOLL set and affinity
7304 * set to a single CPU.
7307 kthread_park(sqd->thread);
7308 kthread_stop(sqd->thread);
7315 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7317 struct io_ring_ctx *ctx_attach;
7318 struct io_sq_data *sqd;
7321 f = fdget(p->wq_fd);
7323 return ERR_PTR(-ENXIO);
7324 if (f.file->f_op != &io_uring_fops) {
7326 return ERR_PTR(-EINVAL);
7329 ctx_attach = f.file->private_data;
7330 sqd = ctx_attach->sq_data;
7333 return ERR_PTR(-EINVAL);
7336 refcount_inc(&sqd->refs);
7341 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p)
7343 struct io_sq_data *sqd;
7345 if (p->flags & IORING_SETUP_ATTACH_WQ)
7346 return io_attach_sq_data(p);
7348 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7350 return ERR_PTR(-ENOMEM);
7352 refcount_set(&sqd->refs, 1);
7353 INIT_LIST_HEAD(&sqd->ctx_list);
7354 INIT_LIST_HEAD(&sqd->ctx_new_list);
7355 mutex_init(&sqd->ctx_lock);
7356 mutex_init(&sqd->lock);
7357 init_waitqueue_head(&sqd->wait);
7361 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7362 __releases(&sqd->lock)
7366 kthread_unpark(sqd->thread);
7367 mutex_unlock(&sqd->lock);
7370 static void io_sq_thread_park(struct io_sq_data *sqd)
7371 __acquires(&sqd->lock)
7375 mutex_lock(&sqd->lock);
7376 kthread_park(sqd->thread);
7379 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
7381 struct io_sq_data *sqd = ctx->sq_data;
7386 * We may arrive here from the error branch in
7387 * io_sq_offload_create() where the kthread is created
7388 * without being waked up, thus wake it up now to make
7389 * sure the wait will complete.
7391 wake_up_process(sqd->thread);
7392 wait_for_completion(&ctx->sq_thread_comp);
7394 io_sq_thread_park(sqd);
7397 mutex_lock(&sqd->ctx_lock);
7398 list_del(&ctx->sqd_list);
7399 io_sqd_update_thread_idle(sqd);
7400 mutex_unlock(&sqd->ctx_lock);
7403 io_sq_thread_unpark(sqd);
7405 io_put_sq_data(sqd);
7406 ctx->sq_data = NULL;
7410 static void io_finish_async(struct io_ring_ctx *ctx)
7412 io_sq_thread_stop(ctx);
7415 io_wq_destroy(ctx->io_wq);
7420 #if defined(CONFIG_UNIX)
7422 * Ensure the UNIX gc is aware of our file set, so we are certain that
7423 * the io_uring can be safely unregistered on process exit, even if we have
7424 * loops in the file referencing.
7426 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7428 struct sock *sk = ctx->ring_sock->sk;
7429 struct scm_fp_list *fpl;
7430 struct sk_buff *skb;
7433 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7437 skb = alloc_skb(0, GFP_KERNEL);
7446 fpl->user = get_uid(ctx->user);
7447 for (i = 0; i < nr; i++) {
7448 struct file *file = io_file_from_index(ctx, i + offset);
7452 fpl->fp[nr_files] = get_file(file);
7453 unix_inflight(fpl->user, fpl->fp[nr_files]);
7458 fpl->max = SCM_MAX_FD;
7459 fpl->count = nr_files;
7460 UNIXCB(skb).fp = fpl;
7461 skb->destructor = unix_destruct_scm;
7462 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7463 skb_queue_head(&sk->sk_receive_queue, skb);
7465 for (i = 0; i < nr_files; i++)
7476 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7477 * causes regular reference counting to break down. We rely on the UNIX
7478 * garbage collection to take care of this problem for us.
7480 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7482 unsigned left, total;
7486 left = ctx->nr_user_files;
7488 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7490 ret = __io_sqe_files_scm(ctx, this_files, total);
7494 total += this_files;
7500 while (total < ctx->nr_user_files) {
7501 struct file *file = io_file_from_index(ctx, total);
7511 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7517 static int io_sqe_alloc_file_tables(struct fixed_file_data *file_data,
7518 unsigned nr_tables, unsigned nr_files)
7522 for (i = 0; i < nr_tables; i++) {
7523 struct fixed_file_table *table = &file_data->table[i];
7524 unsigned this_files;
7526 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
7527 table->files = kcalloc(this_files, sizeof(struct file *),
7531 nr_files -= this_files;
7537 for (i = 0; i < nr_tables; i++) {
7538 struct fixed_file_table *table = &file_data->table[i];
7539 kfree(table->files);
7544 static void io_ring_file_put(struct io_ring_ctx *ctx, struct file *file)
7546 #if defined(CONFIG_UNIX)
7547 struct sock *sock = ctx->ring_sock->sk;
7548 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7549 struct sk_buff *skb;
7552 __skb_queue_head_init(&list);
7555 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7556 * remove this entry and rearrange the file array.
7558 skb = skb_dequeue(head);
7560 struct scm_fp_list *fp;
7562 fp = UNIXCB(skb).fp;
7563 for (i = 0; i < fp->count; i++) {
7566 if (fp->fp[i] != file)
7569 unix_notinflight(fp->user, fp->fp[i]);
7570 left = fp->count - 1 - i;
7572 memmove(&fp->fp[i], &fp->fp[i + 1],
7573 left * sizeof(struct file *));
7580 __skb_queue_tail(&list, skb);
7590 __skb_queue_tail(&list, skb);
7592 skb = skb_dequeue(head);
7595 if (skb_peek(&list)) {
7596 spin_lock_irq(&head->lock);
7597 while ((skb = __skb_dequeue(&list)) != NULL)
7598 __skb_queue_tail(head, skb);
7599 spin_unlock_irq(&head->lock);
7606 struct io_file_put {
7607 struct list_head list;
7611 static void __io_file_put_work(struct fixed_file_ref_node *ref_node)
7613 struct fixed_file_data *file_data = ref_node->file_data;
7614 struct io_ring_ctx *ctx = file_data->ctx;
7615 struct io_file_put *pfile, *tmp;
7617 list_for_each_entry_safe(pfile, tmp, &ref_node->file_list, list) {
7618 list_del(&pfile->list);
7619 io_ring_file_put(ctx, pfile->file);
7623 percpu_ref_exit(&ref_node->refs);
7625 percpu_ref_put(&file_data->refs);
7628 static void io_file_put_work(struct work_struct *work)
7630 struct io_ring_ctx *ctx;
7631 struct llist_node *node;
7633 ctx = container_of(work, struct io_ring_ctx, file_put_work.work);
7634 node = llist_del_all(&ctx->file_put_llist);
7637 struct fixed_file_ref_node *ref_node;
7638 struct llist_node *next = node->next;
7640 ref_node = llist_entry(node, struct fixed_file_ref_node, llist);
7641 __io_file_put_work(ref_node);
7646 static void io_file_data_ref_zero(struct percpu_ref *ref)
7648 struct fixed_file_ref_node *ref_node;
7649 struct fixed_file_data *data;
7650 struct io_ring_ctx *ctx;
7651 bool first_add = false;
7654 ref_node = container_of(ref, struct fixed_file_ref_node, refs);
7655 data = ref_node->file_data;
7658 spin_lock_bh(&data->lock);
7659 ref_node->done = true;
7661 while (!list_empty(&data->ref_list)) {
7662 ref_node = list_first_entry(&data->ref_list,
7663 struct fixed_file_ref_node, node);
7664 /* recycle ref nodes in order */
7665 if (!ref_node->done)
7667 list_del(&ref_node->node);
7668 first_add |= llist_add(&ref_node->llist, &ctx->file_put_llist);
7670 spin_unlock_bh(&data->lock);
7672 if (percpu_ref_is_dying(&data->refs))
7676 mod_delayed_work(system_wq, &ctx->file_put_work, 0);
7678 queue_delayed_work(system_wq, &ctx->file_put_work, delay);
7681 static struct fixed_file_ref_node *alloc_fixed_file_ref_node(
7682 struct io_ring_ctx *ctx)
7684 struct fixed_file_ref_node *ref_node;
7686 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7688 return ERR_PTR(-ENOMEM);
7690 if (percpu_ref_init(&ref_node->refs, io_file_data_ref_zero,
7693 return ERR_PTR(-ENOMEM);
7695 INIT_LIST_HEAD(&ref_node->node);
7696 INIT_LIST_HEAD(&ref_node->file_list);
7697 ref_node->file_data = ctx->file_data;
7698 ref_node->done = false;
7702 static void destroy_fixed_file_ref_node(struct fixed_file_ref_node *ref_node)
7704 percpu_ref_exit(&ref_node->refs);
7708 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7711 __s32 __user *fds = (__s32 __user *) arg;
7712 unsigned nr_tables, i;
7714 int fd, ret = -ENOMEM;
7715 struct fixed_file_ref_node *ref_node;
7716 struct fixed_file_data *file_data;
7722 if (nr_args > IORING_MAX_FIXED_FILES)
7725 file_data = kzalloc(sizeof(*ctx->file_data), GFP_KERNEL);
7728 file_data->ctx = ctx;
7729 init_completion(&file_data->done);
7730 INIT_LIST_HEAD(&file_data->ref_list);
7731 spin_lock_init(&file_data->lock);
7733 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
7734 file_data->table = kcalloc(nr_tables, sizeof(*file_data->table),
7736 if (!file_data->table)
7739 if (percpu_ref_init(&file_data->refs, io_file_ref_kill,
7740 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
7743 if (io_sqe_alloc_file_tables(file_data, nr_tables, nr_args))
7745 ctx->file_data = file_data;
7747 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7748 struct fixed_file_table *table;
7751 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7755 /* allow sparse sets */
7765 * Don't allow io_uring instances to be registered. If UNIX
7766 * isn't enabled, then this causes a reference cycle and this
7767 * instance can never get freed. If UNIX is enabled we'll
7768 * handle it just fine, but there's still no point in allowing
7769 * a ring fd as it doesn't support regular read/write anyway.
7771 if (file->f_op == &io_uring_fops) {
7775 table = &file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7776 index = i & IORING_FILE_TABLE_MASK;
7777 table->files[index] = file;
7780 ret = io_sqe_files_scm(ctx);
7782 io_sqe_files_unregister(ctx);
7786 ref_node = alloc_fixed_file_ref_node(ctx);
7787 if (IS_ERR(ref_node)) {
7788 io_sqe_files_unregister(ctx);
7789 return PTR_ERR(ref_node);
7792 io_sqe_files_set_node(file_data, ref_node);
7795 for (i = 0; i < ctx->nr_user_files; i++) {
7796 file = io_file_from_index(ctx, i);
7800 for (i = 0; i < nr_tables; i++)
7801 kfree(file_data->table[i].files);
7802 ctx->nr_user_files = 0;
7804 percpu_ref_exit(&file_data->refs);
7806 kfree(file_data->table);
7808 ctx->file_data = NULL;
7812 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7815 #if defined(CONFIG_UNIX)
7816 struct sock *sock = ctx->ring_sock->sk;
7817 struct sk_buff_head *head = &sock->sk_receive_queue;
7818 struct sk_buff *skb;
7821 * See if we can merge this file into an existing skb SCM_RIGHTS
7822 * file set. If there's no room, fall back to allocating a new skb
7823 * and filling it in.
7825 spin_lock_irq(&head->lock);
7826 skb = skb_peek(head);
7828 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7830 if (fpl->count < SCM_MAX_FD) {
7831 __skb_unlink(skb, head);
7832 spin_unlock_irq(&head->lock);
7833 fpl->fp[fpl->count] = get_file(file);
7834 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7836 spin_lock_irq(&head->lock);
7837 __skb_queue_head(head, skb);
7842 spin_unlock_irq(&head->lock);
7849 return __io_sqe_files_scm(ctx, 1, index);
7855 static int io_queue_file_removal(struct fixed_file_data *data,
7858 struct io_file_put *pfile;
7859 struct fixed_file_ref_node *ref_node = data->node;
7861 pfile = kzalloc(sizeof(*pfile), GFP_KERNEL);
7866 list_add(&pfile->list, &ref_node->file_list);
7871 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7872 struct io_uring_files_update *up,
7875 struct fixed_file_data *data = ctx->file_data;
7876 struct fixed_file_ref_node *ref_node;
7881 bool needs_switch = false;
7883 if (check_add_overflow(up->offset, nr_args, &done))
7885 if (done > ctx->nr_user_files)
7888 ref_node = alloc_fixed_file_ref_node(ctx);
7889 if (IS_ERR(ref_node))
7890 return PTR_ERR(ref_node);
7893 fds = u64_to_user_ptr(up->fds);
7895 struct fixed_file_table *table;
7899 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
7903 i = array_index_nospec(up->offset, ctx->nr_user_files);
7904 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7905 index = i & IORING_FILE_TABLE_MASK;
7906 if (table->files[index]) {
7907 file = table->files[index];
7908 err = io_queue_file_removal(data, file);
7911 table->files[index] = NULL;
7912 needs_switch = true;
7921 * Don't allow io_uring instances to be registered. If
7922 * UNIX isn't enabled, then this causes a reference
7923 * cycle and this instance can never get freed. If UNIX
7924 * is enabled we'll handle it just fine, but there's
7925 * still no point in allowing a ring fd as it doesn't
7926 * support regular read/write anyway.
7928 if (file->f_op == &io_uring_fops) {
7933 table->files[index] = file;
7934 err = io_sqe_file_register(ctx, file, i);
7936 table->files[index] = NULL;
7947 percpu_ref_kill(&data->node->refs);
7948 io_sqe_files_set_node(data, ref_node);
7950 destroy_fixed_file_ref_node(ref_node);
7952 return done ? done : err;
7955 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
7958 struct io_uring_files_update up;
7960 if (!ctx->file_data)
7964 if (copy_from_user(&up, arg, sizeof(up)))
7969 return __io_sqe_files_update(ctx, &up, nr_args);
7972 static void io_free_work(struct io_wq_work *work)
7974 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7976 /* Consider that io_steal_work() relies on this ref */
7980 static int io_init_wq_offload(struct io_ring_ctx *ctx,
7981 struct io_uring_params *p)
7983 struct io_wq_data data;
7985 struct io_ring_ctx *ctx_attach;
7986 unsigned int concurrency;
7989 data.user = ctx->user;
7990 data.free_work = io_free_work;
7991 data.do_work = io_wq_submit_work;
7993 if (!(p->flags & IORING_SETUP_ATTACH_WQ)) {
7994 /* Do QD, or 4 * CPUS, whatever is smallest */
7995 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7997 ctx->io_wq = io_wq_create(concurrency, &data);
7998 if (IS_ERR(ctx->io_wq)) {
7999 ret = PTR_ERR(ctx->io_wq);
8005 f = fdget(p->wq_fd);
8009 if (f.file->f_op != &io_uring_fops) {
8014 ctx_attach = f.file->private_data;
8015 /* @io_wq is protected by holding the fd */
8016 if (!io_wq_get(ctx_attach->io_wq, &data)) {
8021 ctx->io_wq = ctx_attach->io_wq;
8027 static int io_uring_alloc_task_context(struct task_struct *task)
8029 struct io_uring_task *tctx;
8032 tctx = kmalloc(sizeof(*tctx), GFP_KERNEL);
8033 if (unlikely(!tctx))
8036 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
8037 if (unlikely(ret)) {
8043 init_waitqueue_head(&tctx->wait);
8045 atomic_set(&tctx->in_idle, 0);
8046 tctx->sqpoll = false;
8047 io_init_identity(&tctx->__identity);
8048 tctx->identity = &tctx->__identity;
8049 task->io_uring = tctx;
8053 void __io_uring_free(struct task_struct *tsk)
8055 struct io_uring_task *tctx = tsk->io_uring;
8057 WARN_ON_ONCE(!xa_empty(&tctx->xa));
8058 WARN_ON_ONCE(refcount_read(&tctx->identity->count) != 1);
8059 if (tctx->identity != &tctx->__identity)
8060 kfree(tctx->identity);
8061 percpu_counter_destroy(&tctx->inflight);
8063 tsk->io_uring = NULL;
8066 static int io_sq_offload_create(struct io_ring_ctx *ctx,
8067 struct io_uring_params *p)
8071 if (ctx->flags & IORING_SETUP_SQPOLL) {
8072 struct io_sq_data *sqd;
8075 if (!capable(CAP_SYS_ADMIN) && !capable(CAP_SYS_NICE))
8078 sqd = io_get_sq_data(p);
8085 io_sq_thread_park(sqd);
8086 mutex_lock(&sqd->ctx_lock);
8087 list_add(&ctx->sqd_list, &sqd->ctx_new_list);
8088 mutex_unlock(&sqd->ctx_lock);
8089 io_sq_thread_unpark(sqd);
8091 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
8092 if (!ctx->sq_thread_idle)
8093 ctx->sq_thread_idle = HZ;
8098 if (p->flags & IORING_SETUP_SQ_AFF) {
8099 int cpu = p->sq_thread_cpu;
8102 if (cpu >= nr_cpu_ids)
8104 if (!cpu_online(cpu))
8107 sqd->thread = kthread_create_on_cpu(io_sq_thread, sqd,
8108 cpu, "io_uring-sq");
8110 sqd->thread = kthread_create(io_sq_thread, sqd,
8113 if (IS_ERR(sqd->thread)) {
8114 ret = PTR_ERR(sqd->thread);
8118 ret = io_uring_alloc_task_context(sqd->thread);
8121 } else if (p->flags & IORING_SETUP_SQ_AFF) {
8122 /* Can't have SQ_AFF without SQPOLL */
8128 ret = io_init_wq_offload(ctx, p);
8134 io_finish_async(ctx);
8138 static void io_sq_offload_start(struct io_ring_ctx *ctx)
8140 struct io_sq_data *sqd = ctx->sq_data;
8142 if ((ctx->flags & IORING_SETUP_SQPOLL) && sqd->thread)
8143 wake_up_process(sqd->thread);
8146 static inline void __io_unaccount_mem(struct user_struct *user,
8147 unsigned long nr_pages)
8149 atomic_long_sub(nr_pages, &user->locked_vm);
8152 static inline int __io_account_mem(struct user_struct *user,
8153 unsigned long nr_pages)
8155 unsigned long page_limit, cur_pages, new_pages;
8157 /* Don't allow more pages than we can safely lock */
8158 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
8161 cur_pages = atomic_long_read(&user->locked_vm);
8162 new_pages = cur_pages + nr_pages;
8163 if (new_pages > page_limit)
8165 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
8166 new_pages) != cur_pages);
8171 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
8172 enum io_mem_account acct)
8175 __io_unaccount_mem(ctx->user, nr_pages);
8177 if (ctx->mm_account) {
8178 if (acct == ACCT_LOCKED) {
8179 mmap_write_lock(ctx->mm_account);
8180 ctx->mm_account->locked_vm -= nr_pages;
8181 mmap_write_unlock(ctx->mm_account);
8182 }else if (acct == ACCT_PINNED) {
8183 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
8188 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
8189 enum io_mem_account acct)
8193 if (ctx->limit_mem) {
8194 ret = __io_account_mem(ctx->user, nr_pages);
8199 if (ctx->mm_account) {
8200 if (acct == ACCT_LOCKED) {
8201 mmap_write_lock(ctx->mm_account);
8202 ctx->mm_account->locked_vm += nr_pages;
8203 mmap_write_unlock(ctx->mm_account);
8204 } else if (acct == ACCT_PINNED) {
8205 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
8212 static void io_mem_free(void *ptr)
8219 page = virt_to_head_page(ptr);
8220 if (put_page_testzero(page))
8221 free_compound_page(page);
8224 static void *io_mem_alloc(size_t size)
8226 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
8229 return (void *) __get_free_pages(gfp_flags, get_order(size));
8232 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8235 struct io_rings *rings;
8236 size_t off, sq_array_size;
8238 off = struct_size(rings, cqes, cq_entries);
8239 if (off == SIZE_MAX)
8243 off = ALIGN(off, SMP_CACHE_BYTES);
8251 sq_array_size = array_size(sizeof(u32), sq_entries);
8252 if (sq_array_size == SIZE_MAX)
8255 if (check_add_overflow(off, sq_array_size, &off))
8261 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
8265 pages = (size_t)1 << get_order(
8266 rings_size(sq_entries, cq_entries, NULL));
8267 pages += (size_t)1 << get_order(
8268 array_size(sizeof(struct io_uring_sqe), sq_entries));
8273 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
8277 if (!ctx->user_bufs)
8280 for (i = 0; i < ctx->nr_user_bufs; i++) {
8281 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8283 for (j = 0; j < imu->nr_bvecs; j++)
8284 unpin_user_page(imu->bvec[j].bv_page);
8286 if (imu->acct_pages)
8287 io_unaccount_mem(ctx, imu->acct_pages, ACCT_PINNED);
8292 kfree(ctx->user_bufs);
8293 ctx->user_bufs = NULL;
8294 ctx->nr_user_bufs = 0;
8298 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8299 void __user *arg, unsigned index)
8301 struct iovec __user *src;
8303 #ifdef CONFIG_COMPAT
8305 struct compat_iovec __user *ciovs;
8306 struct compat_iovec ciov;
8308 ciovs = (struct compat_iovec __user *) arg;
8309 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8312 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8313 dst->iov_len = ciov.iov_len;
8317 src = (struct iovec __user *) arg;
8318 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8324 * Not super efficient, but this is just a registration time. And we do cache
8325 * the last compound head, so generally we'll only do a full search if we don't
8328 * We check if the given compound head page has already been accounted, to
8329 * avoid double accounting it. This allows us to account the full size of the
8330 * page, not just the constituent pages of a huge page.
8332 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8333 int nr_pages, struct page *hpage)
8337 /* check current page array */
8338 for (i = 0; i < nr_pages; i++) {
8339 if (!PageCompound(pages[i]))
8341 if (compound_head(pages[i]) == hpage)
8345 /* check previously registered pages */
8346 for (i = 0; i < ctx->nr_user_bufs; i++) {
8347 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8349 for (j = 0; j < imu->nr_bvecs; j++) {
8350 if (!PageCompound(imu->bvec[j].bv_page))
8352 if (compound_head(imu->bvec[j].bv_page) == hpage)
8360 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8361 int nr_pages, struct io_mapped_ubuf *imu,
8362 struct page **last_hpage)
8366 for (i = 0; i < nr_pages; i++) {
8367 if (!PageCompound(pages[i])) {
8372 hpage = compound_head(pages[i]);
8373 if (hpage == *last_hpage)
8375 *last_hpage = hpage;
8376 if (headpage_already_acct(ctx, pages, i, hpage))
8378 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8382 if (!imu->acct_pages)
8385 ret = io_account_mem(ctx, imu->acct_pages, ACCT_PINNED);
8387 imu->acct_pages = 0;
8391 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
8394 struct vm_area_struct **vmas = NULL;
8395 struct page **pages = NULL;
8396 struct page *last_hpage = NULL;
8397 int i, j, got_pages = 0;
8402 if (!nr_args || nr_args > UIO_MAXIOV)
8405 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
8407 if (!ctx->user_bufs)
8410 for (i = 0; i < nr_args; i++) {
8411 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8412 unsigned long off, start, end, ubuf;
8417 ret = io_copy_iov(ctx, &iov, arg, i);
8422 * Don't impose further limits on the size and buffer
8423 * constraints here, we'll -EINVAL later when IO is
8424 * submitted if they are wrong.
8427 if (!iov.iov_base || !iov.iov_len)
8430 /* arbitrary limit, but we need something */
8431 if (iov.iov_len > SZ_1G)
8434 ubuf = (unsigned long) iov.iov_base;
8435 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8436 start = ubuf >> PAGE_SHIFT;
8437 nr_pages = end - start;
8440 if (!pages || nr_pages > got_pages) {
8443 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
8445 vmas = kvmalloc_array(nr_pages,
8446 sizeof(struct vm_area_struct *),
8448 if (!pages || !vmas) {
8452 got_pages = nr_pages;
8455 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
8462 mmap_read_lock(current->mm);
8463 pret = pin_user_pages(ubuf, nr_pages,
8464 FOLL_WRITE | FOLL_LONGTERM,
8466 if (pret == nr_pages) {
8467 /* don't support file backed memory */
8468 for (j = 0; j < nr_pages; j++) {
8469 struct vm_area_struct *vma = vmas[j];
8472 !is_file_hugepages(vma->vm_file)) {
8478 ret = pret < 0 ? pret : -EFAULT;
8480 mmap_read_unlock(current->mm);
8483 * if we did partial map, or found file backed vmas,
8484 * release any pages we did get
8487 unpin_user_pages(pages, pret);
8492 ret = io_buffer_account_pin(ctx, pages, pret, imu, &last_hpage);
8494 unpin_user_pages(pages, pret);
8499 off = ubuf & ~PAGE_MASK;
8501 for (j = 0; j < nr_pages; j++) {
8504 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8505 imu->bvec[j].bv_page = pages[j];
8506 imu->bvec[j].bv_len = vec_len;
8507 imu->bvec[j].bv_offset = off;
8511 /* store original address for later verification */
8513 imu->len = iov.iov_len;
8514 imu->nr_bvecs = nr_pages;
8516 ctx->nr_user_bufs++;
8524 io_sqe_buffer_unregister(ctx);
8528 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8530 __s32 __user *fds = arg;
8536 if (copy_from_user(&fd, fds, sizeof(*fds)))
8539 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8540 if (IS_ERR(ctx->cq_ev_fd)) {
8541 int ret = PTR_ERR(ctx->cq_ev_fd);
8542 ctx->cq_ev_fd = NULL;
8549 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8551 if (ctx->cq_ev_fd) {
8552 eventfd_ctx_put(ctx->cq_ev_fd);
8553 ctx->cq_ev_fd = NULL;
8560 static int __io_destroy_buffers(int id, void *p, void *data)
8562 struct io_ring_ctx *ctx = data;
8563 struct io_buffer *buf = p;
8565 __io_remove_buffers(ctx, buf, id, -1U);
8569 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8571 idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx);
8572 idr_destroy(&ctx->io_buffer_idr);
8575 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8577 io_finish_async(ctx);
8578 io_sqe_buffer_unregister(ctx);
8580 if (ctx->sqo_task) {
8581 put_task_struct(ctx->sqo_task);
8582 ctx->sqo_task = NULL;
8583 mmdrop(ctx->mm_account);
8584 ctx->mm_account = NULL;
8587 #ifdef CONFIG_BLK_CGROUP
8588 if (ctx->sqo_blkcg_css)
8589 css_put(ctx->sqo_blkcg_css);
8592 io_sqe_files_unregister(ctx);
8593 io_eventfd_unregister(ctx);
8594 io_destroy_buffers(ctx);
8595 idr_destroy(&ctx->personality_idr);
8597 #if defined(CONFIG_UNIX)
8598 if (ctx->ring_sock) {
8599 ctx->ring_sock->file = NULL; /* so that iput() is called */
8600 sock_release(ctx->ring_sock);
8604 io_mem_free(ctx->rings);
8605 io_mem_free(ctx->sq_sqes);
8607 percpu_ref_exit(&ctx->refs);
8608 free_uid(ctx->user);
8609 put_cred(ctx->creds);
8610 kfree(ctx->cancel_hash);
8611 kmem_cache_free(req_cachep, ctx->fallback_req);
8615 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8617 struct io_ring_ctx *ctx = file->private_data;
8620 poll_wait(file, &ctx->cq_wait, wait);
8622 * synchronizes with barrier from wq_has_sleeper call in
8626 if (!io_sqring_full(ctx))
8627 mask |= EPOLLOUT | EPOLLWRNORM;
8628 if (io_cqring_events(ctx, false))
8629 mask |= EPOLLIN | EPOLLRDNORM;
8634 static int io_uring_fasync(int fd, struct file *file, int on)
8636 struct io_ring_ctx *ctx = file->private_data;
8638 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8641 static int io_remove_personalities(int id, void *p, void *data)
8643 struct io_ring_ctx *ctx = data;
8644 struct io_identity *iod;
8646 iod = idr_remove(&ctx->personality_idr, id);
8648 put_cred(iod->creds);
8649 if (refcount_dec_and_test(&iod->count))
8655 static void io_ring_exit_work(struct work_struct *work)
8657 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
8661 * If we're doing polled IO and end up having requests being
8662 * submitted async (out-of-line), then completions can come in while
8663 * we're waiting for refs to drop. We need to reap these manually,
8664 * as nobody else will be looking for them.
8667 io_iopoll_try_reap_events(ctx);
8668 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
8669 io_ring_ctx_free(ctx);
8672 static bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
8674 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8676 return req->ctx == data;
8679 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8681 mutex_lock(&ctx->uring_lock);
8682 percpu_ref_kill(&ctx->refs);
8683 /* if force is set, the ring is going away. always drop after that */
8684 ctx->cq_overflow_flushed = 1;
8686 io_cqring_overflow_flush(ctx, true, NULL, NULL);
8687 mutex_unlock(&ctx->uring_lock);
8689 io_kill_timeouts(ctx, NULL, NULL);
8690 io_poll_remove_all(ctx, NULL, NULL);
8693 io_wq_cancel_cb(ctx->io_wq, io_cancel_ctx_cb, ctx, true);
8695 /* if we failed setting up the ctx, we might not have any rings */
8696 io_iopoll_try_reap_events(ctx);
8697 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
8700 * Do this upfront, so we won't have a grace period where the ring
8701 * is closed but resources aren't reaped yet. This can cause
8702 * spurious failure in setting up a new ring.
8704 io_unaccount_mem(ctx, ring_pages(ctx->sq_entries, ctx->cq_entries),
8707 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8709 * Use system_unbound_wq to avoid spawning tons of event kworkers
8710 * if we're exiting a ton of rings at the same time. It just adds
8711 * noise and overhead, there's no discernable change in runtime
8712 * over using system_wq.
8714 queue_work(system_unbound_wq, &ctx->exit_work);
8717 static int io_uring_release(struct inode *inode, struct file *file)
8719 struct io_ring_ctx *ctx = file->private_data;
8721 file->private_data = NULL;
8722 io_ring_ctx_wait_and_kill(ctx);
8726 struct io_task_cancel {
8727 struct task_struct *task;
8728 struct files_struct *files;
8731 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8733 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8734 struct io_task_cancel *cancel = data;
8737 if (cancel->files && (req->flags & REQ_F_LINK_TIMEOUT)) {
8738 unsigned long flags;
8739 struct io_ring_ctx *ctx = req->ctx;
8741 /* protect against races with linked timeouts */
8742 spin_lock_irqsave(&ctx->completion_lock, flags);
8743 ret = io_match_task(req, cancel->task, cancel->files);
8744 spin_unlock_irqrestore(&ctx->completion_lock, flags);
8746 ret = io_match_task(req, cancel->task, cancel->files);
8751 static void io_cancel_defer_files(struct io_ring_ctx *ctx,
8752 struct task_struct *task,
8753 struct files_struct *files)
8755 struct io_defer_entry *de = NULL;
8758 spin_lock_irq(&ctx->completion_lock);
8759 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8760 if (io_match_task(de->req, task, files)) {
8761 list_cut_position(&list, &ctx->defer_list, &de->list);
8765 spin_unlock_irq(&ctx->completion_lock);
8767 while (!list_empty(&list)) {
8768 de = list_first_entry(&list, struct io_defer_entry, list);
8769 list_del_init(&de->list);
8770 req_set_fail_links(de->req);
8771 io_put_req(de->req);
8772 io_req_complete(de->req, -ECANCELED);
8777 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
8778 struct task_struct *task,
8779 struct files_struct *files)
8781 while (!list_empty_careful(&ctx->inflight_list)) {
8782 struct io_task_cancel cancel = { .task = task, .files = files };
8783 struct io_kiocb *req;
8787 spin_lock_irq(&ctx->inflight_lock);
8788 list_for_each_entry(req, &ctx->inflight_list, inflight_entry) {
8789 if (req->task != task ||
8790 req->work.identity->files != files)
8796 prepare_to_wait(&task->io_uring->wait, &wait,
8797 TASK_UNINTERRUPTIBLE);
8798 spin_unlock_irq(&ctx->inflight_lock);
8800 /* We need to keep going until we don't find a matching req */
8804 io_wq_cancel_cb(ctx->io_wq, io_cancel_task_cb, &cancel, true);
8805 io_poll_remove_all(ctx, task, files);
8806 io_kill_timeouts(ctx, task, files);
8807 /* cancellations _may_ trigger task work */
8810 finish_wait(&task->io_uring->wait, &wait);
8814 static void __io_uring_cancel_task_requests(struct io_ring_ctx *ctx,
8815 struct task_struct *task)
8818 struct io_task_cancel cancel = { .task = task, .files = NULL, };
8819 enum io_wq_cancel cret;
8822 cret = io_wq_cancel_cb(ctx->io_wq, io_cancel_task_cb, &cancel, true);
8823 if (cret != IO_WQ_CANCEL_NOTFOUND)
8826 /* SQPOLL thread does its own polling */
8827 if (!(ctx->flags & IORING_SETUP_SQPOLL)) {
8828 while (!list_empty_careful(&ctx->iopoll_list)) {
8829 io_iopoll_try_reap_events(ctx);
8834 ret |= io_poll_remove_all(ctx, task, NULL);
8835 ret |= io_kill_timeouts(ctx, task, NULL);
8836 ret |= io_run_task_work();
8844 * We need to iteratively cancel requests, in case a request has dependent
8845 * hard links. These persist even for failure of cancelations, hence keep
8846 * looping until none are found.
8848 static void io_uring_cancel_task_requests(struct io_ring_ctx *ctx,
8849 struct files_struct *files)
8851 struct task_struct *task = current;
8853 if ((ctx->flags & IORING_SETUP_SQPOLL) && ctx->sq_data) {
8854 task = ctx->sq_data->thread;
8855 atomic_inc(&task->io_uring->in_idle);
8856 io_sq_thread_park(ctx->sq_data);
8859 io_cancel_defer_files(ctx, task, files);
8860 io_ring_submit_lock(ctx, (ctx->flags & IORING_SETUP_IOPOLL));
8861 io_cqring_overflow_flush(ctx, true, task, files);
8862 io_ring_submit_unlock(ctx, (ctx->flags & IORING_SETUP_IOPOLL));
8865 __io_uring_cancel_task_requests(ctx, task);
8867 io_uring_cancel_files(ctx, task, files);
8869 if ((ctx->flags & IORING_SETUP_SQPOLL) && ctx->sq_data) {
8870 atomic_dec(&task->io_uring->in_idle);
8872 * If the files that are going away are the ones in the thread
8873 * identity, clear them out.
8875 if (task->io_uring->identity->files == files)
8876 task->io_uring->identity->files = NULL;
8877 io_sq_thread_unpark(ctx->sq_data);
8882 * Note that this task has used io_uring. We use it for cancelation purposes.
8884 static int io_uring_add_task_file(struct io_ring_ctx *ctx, struct file *file)
8886 struct io_uring_task *tctx = current->io_uring;
8889 if (unlikely(!tctx)) {
8890 ret = io_uring_alloc_task_context(current);
8893 tctx = current->io_uring;
8895 if (tctx->last != file) {
8896 void *old = xa_load(&tctx->xa, (unsigned long)file);
8900 ret = xa_err(xa_store(&tctx->xa, (unsigned long)file,
8911 * This is race safe in that the task itself is doing this, hence it
8912 * cannot be going through the exit/cancel paths at the same time.
8913 * This cannot be modified while exit/cancel is running.
8915 if (!tctx->sqpoll && (ctx->flags & IORING_SETUP_SQPOLL))
8916 tctx->sqpoll = true;
8922 * Remove this io_uring_file -> task mapping.
8924 static void io_uring_del_task_file(struct file *file)
8926 struct io_uring_task *tctx = current->io_uring;
8928 if (tctx->last == file)
8930 file = xa_erase(&tctx->xa, (unsigned long)file);
8936 * Drop task note for this file if we're the only ones that hold it after
8939 static void io_uring_attempt_task_drop(struct file *file)
8941 if (!current->io_uring)
8944 * fput() is pending, will be 2 if the only other ref is our potential
8945 * task file note. If the task is exiting, drop regardless of count.
8947 if (fatal_signal_pending(current) || (current->flags & PF_EXITING) ||
8948 atomic_long_read(&file->f_count) == 2)
8949 io_uring_del_task_file(file);
8952 void __io_uring_files_cancel(struct files_struct *files)
8954 struct io_uring_task *tctx = current->io_uring;
8956 unsigned long index;
8958 /* make sure overflow events are dropped */
8959 atomic_inc(&tctx->in_idle);
8961 xa_for_each(&tctx->xa, index, file) {
8962 struct io_ring_ctx *ctx = file->private_data;
8964 io_uring_cancel_task_requests(ctx, files);
8966 io_uring_del_task_file(file);
8969 atomic_dec(&tctx->in_idle);
8972 static s64 tctx_inflight(struct io_uring_task *tctx)
8974 unsigned long index;
8978 inflight = percpu_counter_sum(&tctx->inflight);
8983 * If we have SQPOLL rings, then we need to iterate and find them, and
8984 * add the pending count for those.
8986 xa_for_each(&tctx->xa, index, file) {
8987 struct io_ring_ctx *ctx = file->private_data;
8989 if (ctx->flags & IORING_SETUP_SQPOLL) {
8990 struct io_uring_task *__tctx = ctx->sqo_task->io_uring;
8992 inflight += percpu_counter_sum(&__tctx->inflight);
9000 * Find any io_uring fd that this task has registered or done IO on, and cancel
9003 void __io_uring_task_cancel(void)
9005 struct io_uring_task *tctx = current->io_uring;
9009 /* make sure overflow events are dropped */
9010 atomic_inc(&tctx->in_idle);
9013 /* read completions before cancelations */
9014 inflight = tctx_inflight(tctx);
9017 __io_uring_files_cancel(NULL);
9019 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
9022 * If we've seen completions, retry. This avoids a race where
9023 * a completion comes in before we did prepare_to_wait().
9025 if (inflight != tctx_inflight(tctx))
9028 finish_wait(&tctx->wait, &wait);
9031 atomic_dec(&tctx->in_idle);
9034 static int io_uring_flush(struct file *file, void *data)
9036 io_uring_attempt_task_drop(file);
9040 static void *io_uring_validate_mmap_request(struct file *file,
9041 loff_t pgoff, size_t sz)
9043 struct io_ring_ctx *ctx = file->private_data;
9044 loff_t offset = pgoff << PAGE_SHIFT;
9049 case IORING_OFF_SQ_RING:
9050 case IORING_OFF_CQ_RING:
9053 case IORING_OFF_SQES:
9057 return ERR_PTR(-EINVAL);
9060 page = virt_to_head_page(ptr);
9061 if (sz > page_size(page))
9062 return ERR_PTR(-EINVAL);
9069 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9071 size_t sz = vma->vm_end - vma->vm_start;
9075 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
9077 return PTR_ERR(ptr);
9079 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
9080 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
9083 #else /* !CONFIG_MMU */
9085 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9087 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
9090 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
9092 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
9095 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
9096 unsigned long addr, unsigned long len,
9097 unsigned long pgoff, unsigned long flags)
9101 ptr = io_uring_validate_mmap_request(file, pgoff, len);
9103 return PTR_ERR(ptr);
9105 return (unsigned long) ptr;
9108 #endif /* !CONFIG_MMU */
9110 static void io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
9115 if (!io_sqring_full(ctx))
9118 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9120 if (!io_sqring_full(ctx))
9124 } while (!signal_pending(current));
9126 finish_wait(&ctx->sqo_sq_wait, &wait);
9129 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9130 struct __kernel_timespec __user **ts,
9131 const sigset_t __user **sig)
9133 struct io_uring_getevents_arg arg;
9136 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9137 * is just a pointer to the sigset_t.
9139 if (!(flags & IORING_ENTER_EXT_ARG)) {
9140 *sig = (const sigset_t __user *) argp;
9146 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9147 * timespec and sigset_t pointers if good.
9149 if (*argsz != sizeof(arg))
9151 if (copy_from_user(&arg, argp, sizeof(arg)))
9153 *sig = u64_to_user_ptr(arg.sigmask);
9154 *argsz = arg.sigmask_sz;
9155 *ts = u64_to_user_ptr(arg.ts);
9159 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9160 u32, min_complete, u32, flags, const void __user *, argp,
9163 struct io_ring_ctx *ctx;
9170 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9171 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG))
9179 if (f.file->f_op != &io_uring_fops)
9183 ctx = f.file->private_data;
9184 if (!percpu_ref_tryget(&ctx->refs))
9188 if (ctx->flags & IORING_SETUP_R_DISABLED)
9192 * For SQ polling, the thread will do all submissions and completions.
9193 * Just return the requested submit count, and wake the thread if
9197 if (ctx->flags & IORING_SETUP_SQPOLL) {
9198 if (!list_empty_careful(&ctx->cq_overflow_list)) {
9199 bool needs_lock = ctx->flags & IORING_SETUP_IOPOLL;
9201 io_ring_submit_lock(ctx, needs_lock);
9202 io_cqring_overflow_flush(ctx, false, NULL, NULL);
9203 io_ring_submit_unlock(ctx, needs_lock);
9205 if (flags & IORING_ENTER_SQ_WAKEUP)
9206 wake_up(&ctx->sq_data->wait);
9207 if (flags & IORING_ENTER_SQ_WAIT)
9208 io_sqpoll_wait_sq(ctx);
9209 submitted = to_submit;
9210 } else if (to_submit) {
9211 ret = io_uring_add_task_file(ctx, f.file);
9214 mutex_lock(&ctx->uring_lock);
9215 submitted = io_submit_sqes(ctx, to_submit);
9216 mutex_unlock(&ctx->uring_lock);
9218 if (submitted != to_submit)
9221 if (flags & IORING_ENTER_GETEVENTS) {
9222 const sigset_t __user *sig;
9223 struct __kernel_timespec __user *ts;
9225 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
9229 min_complete = min(min_complete, ctx->cq_entries);
9232 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9233 * space applications don't need to do io completion events
9234 * polling again, they can rely on io_sq_thread to do polling
9235 * work, which can reduce cpu usage and uring_lock contention.
9237 if (ctx->flags & IORING_SETUP_IOPOLL &&
9238 !(ctx->flags & IORING_SETUP_SQPOLL)) {
9239 ret = io_iopoll_check(ctx, min_complete);
9241 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
9246 percpu_ref_put(&ctx->refs);
9249 return submitted ? submitted : ret;
9252 #ifdef CONFIG_PROC_FS
9253 static int io_uring_show_cred(int id, void *p, void *data)
9255 struct io_identity *iod = p;
9256 const struct cred *cred = iod->creds;
9257 struct seq_file *m = data;
9258 struct user_namespace *uns = seq_user_ns(m);
9259 struct group_info *gi;
9264 seq_printf(m, "%5d\n", id);
9265 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
9266 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
9267 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
9268 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
9269 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
9270 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
9271 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
9272 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
9273 seq_puts(m, "\n\tGroups:\t");
9274 gi = cred->group_info;
9275 for (g = 0; g < gi->ngroups; g++) {
9276 seq_put_decimal_ull(m, g ? " " : "",
9277 from_kgid_munged(uns, gi->gid[g]));
9279 seq_puts(m, "\n\tCapEff:\t");
9280 cap = cred->cap_effective;
9281 CAP_FOR_EACH_U32(__capi)
9282 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
9287 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
9289 struct io_sq_data *sq = NULL;
9294 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9295 * since fdinfo case grabs it in the opposite direction of normal use
9296 * cases. If we fail to get the lock, we just don't iterate any
9297 * structures that could be going away outside the io_uring mutex.
9299 has_lock = mutex_trylock(&ctx->uring_lock);
9301 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL))
9304 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9305 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9306 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9307 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9308 struct fixed_file_table *table;
9311 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
9312 f = table->files[i & IORING_FILE_TABLE_MASK];
9314 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9316 seq_printf(m, "%5u: <none>\n", i);
9318 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9319 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9320 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
9322 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
9323 (unsigned int) buf->len);
9325 if (has_lock && !idr_is_empty(&ctx->personality_idr)) {
9326 seq_printf(m, "Personalities:\n");
9327 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
9329 seq_printf(m, "PollList:\n");
9330 spin_lock_irq(&ctx->completion_lock);
9331 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9332 struct hlist_head *list = &ctx->cancel_hash[i];
9333 struct io_kiocb *req;
9335 hlist_for_each_entry(req, list, hash_node)
9336 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9337 req->task->task_works != NULL);
9339 spin_unlock_irq(&ctx->completion_lock);
9341 mutex_unlock(&ctx->uring_lock);
9344 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9346 struct io_ring_ctx *ctx = f->private_data;
9348 if (percpu_ref_tryget(&ctx->refs)) {
9349 __io_uring_show_fdinfo(ctx, m);
9350 percpu_ref_put(&ctx->refs);
9355 static const struct file_operations io_uring_fops = {
9356 .release = io_uring_release,
9357 .flush = io_uring_flush,
9358 .mmap = io_uring_mmap,
9360 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9361 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9363 .poll = io_uring_poll,
9364 .fasync = io_uring_fasync,
9365 #ifdef CONFIG_PROC_FS
9366 .show_fdinfo = io_uring_show_fdinfo,
9370 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9371 struct io_uring_params *p)
9373 struct io_rings *rings;
9374 size_t size, sq_array_offset;
9376 /* make sure these are sane, as we already accounted them */
9377 ctx->sq_entries = p->sq_entries;
9378 ctx->cq_entries = p->cq_entries;
9380 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9381 if (size == SIZE_MAX)
9384 rings = io_mem_alloc(size);
9389 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9390 rings->sq_ring_mask = p->sq_entries - 1;
9391 rings->cq_ring_mask = p->cq_entries - 1;
9392 rings->sq_ring_entries = p->sq_entries;
9393 rings->cq_ring_entries = p->cq_entries;
9394 ctx->sq_mask = rings->sq_ring_mask;
9395 ctx->cq_mask = rings->cq_ring_mask;
9397 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9398 if (size == SIZE_MAX) {
9399 io_mem_free(ctx->rings);
9404 ctx->sq_sqes = io_mem_alloc(size);
9405 if (!ctx->sq_sqes) {
9406 io_mem_free(ctx->rings);
9414 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
9418 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9422 ret = io_uring_add_task_file(ctx, file);
9427 fd_install(fd, file);
9432 * Allocate an anonymous fd, this is what constitutes the application
9433 * visible backing of an io_uring instance. The application mmaps this
9434 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9435 * we have to tie this fd to a socket for file garbage collection purposes.
9437 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
9440 #if defined(CONFIG_UNIX)
9443 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9446 return ERR_PTR(ret);
9449 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9450 O_RDWR | O_CLOEXEC);
9451 #if defined(CONFIG_UNIX)
9453 sock_release(ctx->ring_sock);
9454 ctx->ring_sock = NULL;
9456 ctx->ring_sock->file = file;
9462 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9463 struct io_uring_params __user *params)
9465 struct user_struct *user = NULL;
9466 struct io_ring_ctx *ctx;
9473 if (entries > IORING_MAX_ENTRIES) {
9474 if (!(p->flags & IORING_SETUP_CLAMP))
9476 entries = IORING_MAX_ENTRIES;
9480 * Use twice as many entries for the CQ ring. It's possible for the
9481 * application to drive a higher depth than the size of the SQ ring,
9482 * since the sqes are only used at submission time. This allows for
9483 * some flexibility in overcommitting a bit. If the application has
9484 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9485 * of CQ ring entries manually.
9487 p->sq_entries = roundup_pow_of_two(entries);
9488 if (p->flags & IORING_SETUP_CQSIZE) {
9490 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9491 * to a power-of-two, if it isn't already. We do NOT impose
9492 * any cq vs sq ring sizing.
9496 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9497 if (!(p->flags & IORING_SETUP_CLAMP))
9499 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9501 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9502 if (p->cq_entries < p->sq_entries)
9505 p->cq_entries = 2 * p->sq_entries;
9508 user = get_uid(current_user());
9509 limit_mem = !capable(CAP_IPC_LOCK);
9512 ret = __io_account_mem(user,
9513 ring_pages(p->sq_entries, p->cq_entries));
9520 ctx = io_ring_ctx_alloc(p);
9523 __io_unaccount_mem(user, ring_pages(p->sq_entries,
9528 ctx->compat = in_compat_syscall();
9530 ctx->creds = get_current_cred();
9532 ctx->loginuid = current->loginuid;
9533 ctx->sessionid = current->sessionid;
9535 ctx->sqo_task = get_task_struct(current);
9538 * This is just grabbed for accounting purposes. When a process exits,
9539 * the mm is exited and dropped before the files, hence we need to hang
9540 * on to this mm purely for the purposes of being able to unaccount
9541 * memory (locked/pinned vm). It's not used for anything else.
9543 mmgrab(current->mm);
9544 ctx->mm_account = current->mm;
9546 #ifdef CONFIG_BLK_CGROUP
9548 * The sq thread will belong to the original cgroup it was inited in.
9549 * If the cgroup goes offline (e.g. disabling the io controller), then
9550 * issued bios will be associated with the closest cgroup later in the
9554 ctx->sqo_blkcg_css = blkcg_css();
9555 ret = css_tryget_online(ctx->sqo_blkcg_css);
9558 /* don't init against a dying cgroup, have the user try again */
9559 ctx->sqo_blkcg_css = NULL;
9566 * Account memory _before_ installing the file descriptor. Once
9567 * the descriptor is installed, it can get closed at any time. Also
9568 * do this before hitting the general error path, as ring freeing
9569 * will un-account as well.
9571 io_account_mem(ctx, ring_pages(p->sq_entries, p->cq_entries),
9573 ctx->limit_mem = limit_mem;
9575 ret = io_allocate_scq_urings(ctx, p);
9579 ret = io_sq_offload_create(ctx, p);
9583 if (!(p->flags & IORING_SETUP_R_DISABLED))
9584 io_sq_offload_start(ctx);
9586 memset(&p->sq_off, 0, sizeof(p->sq_off));
9587 p->sq_off.head = offsetof(struct io_rings, sq.head);
9588 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9589 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9590 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9591 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9592 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9593 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9595 memset(&p->cq_off, 0, sizeof(p->cq_off));
9596 p->cq_off.head = offsetof(struct io_rings, cq.head);
9597 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9598 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9599 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9600 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9601 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9602 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9604 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9605 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9606 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9607 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
9608 IORING_FEAT_EXT_ARG;
9610 if (copy_to_user(params, p, sizeof(*p))) {
9615 file = io_uring_get_file(ctx);
9617 ret = PTR_ERR(file);
9622 * Install ring fd as the very last thing, so we don't risk someone
9623 * having closed it before we finish setup
9625 ret = io_uring_install_fd(ctx, file);
9627 /* fput will clean it up */
9632 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9635 io_ring_ctx_wait_and_kill(ctx);
9640 * Sets up an aio uring context, and returns the fd. Applications asks for a
9641 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9642 * params structure passed in.
9644 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9646 struct io_uring_params p;
9649 if (copy_from_user(&p, params, sizeof(p)))
9651 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9656 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9657 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9658 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9659 IORING_SETUP_R_DISABLED))
9662 return io_uring_create(entries, &p, params);
9665 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9666 struct io_uring_params __user *, params)
9668 return io_uring_setup(entries, params);
9671 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9673 struct io_uring_probe *p;
9677 size = struct_size(p, ops, nr_args);
9678 if (size == SIZE_MAX)
9680 p = kzalloc(size, GFP_KERNEL);
9685 if (copy_from_user(p, arg, size))
9688 if (memchr_inv(p, 0, size))
9691 p->last_op = IORING_OP_LAST - 1;
9692 if (nr_args > IORING_OP_LAST)
9693 nr_args = IORING_OP_LAST;
9695 for (i = 0; i < nr_args; i++) {
9697 if (!io_op_defs[i].not_supported)
9698 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9703 if (copy_to_user(arg, p, size))
9710 static int io_register_personality(struct io_ring_ctx *ctx)
9712 struct io_identity *id;
9715 id = kmalloc(sizeof(*id), GFP_KERNEL);
9719 io_init_identity(id);
9720 id->creds = get_current_cred();
9722 ret = idr_alloc_cyclic(&ctx->personality_idr, id, 1, USHRT_MAX, GFP_KERNEL);
9724 put_cred(id->creds);
9730 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
9732 struct io_identity *iod;
9734 iod = idr_remove(&ctx->personality_idr, id);
9736 put_cred(iod->creds);
9737 if (refcount_dec_and_test(&iod->count))
9745 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9746 unsigned int nr_args)
9748 struct io_uring_restriction *res;
9752 /* Restrictions allowed only if rings started disabled */
9753 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9756 /* We allow only a single restrictions registration */
9757 if (ctx->restrictions.registered)
9760 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9763 size = array_size(nr_args, sizeof(*res));
9764 if (size == SIZE_MAX)
9767 res = memdup_user(arg, size);
9769 return PTR_ERR(res);
9773 for (i = 0; i < nr_args; i++) {
9774 switch (res[i].opcode) {
9775 case IORING_RESTRICTION_REGISTER_OP:
9776 if (res[i].register_op >= IORING_REGISTER_LAST) {
9781 __set_bit(res[i].register_op,
9782 ctx->restrictions.register_op);
9784 case IORING_RESTRICTION_SQE_OP:
9785 if (res[i].sqe_op >= IORING_OP_LAST) {
9790 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9792 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9793 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9795 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9796 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9805 /* Reset all restrictions if an error happened */
9807 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9809 ctx->restrictions.registered = true;
9815 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9817 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9820 if (ctx->restrictions.registered)
9821 ctx->restricted = 1;
9823 ctx->flags &= ~IORING_SETUP_R_DISABLED;
9825 io_sq_offload_start(ctx);
9830 static bool io_register_op_must_quiesce(int op)
9833 case IORING_UNREGISTER_FILES:
9834 case IORING_REGISTER_FILES_UPDATE:
9835 case IORING_REGISTER_PROBE:
9836 case IORING_REGISTER_PERSONALITY:
9837 case IORING_UNREGISTER_PERSONALITY:
9844 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
9845 void __user *arg, unsigned nr_args)
9846 __releases(ctx->uring_lock)
9847 __acquires(ctx->uring_lock)
9852 * We're inside the ring mutex, if the ref is already dying, then
9853 * someone else killed the ctx or is already going through
9854 * io_uring_register().
9856 if (percpu_ref_is_dying(&ctx->refs))
9859 if (io_register_op_must_quiesce(opcode)) {
9860 percpu_ref_kill(&ctx->refs);
9863 * Drop uring mutex before waiting for references to exit. If
9864 * another thread is currently inside io_uring_enter() it might
9865 * need to grab the uring_lock to make progress. If we hold it
9866 * here across the drain wait, then we can deadlock. It's safe
9867 * to drop the mutex here, since no new references will come in
9868 * after we've killed the percpu ref.
9870 mutex_unlock(&ctx->uring_lock);
9872 ret = wait_for_completion_interruptible(&ctx->ref_comp);
9875 ret = io_run_task_work_sig();
9880 mutex_lock(&ctx->uring_lock);
9883 percpu_ref_resurrect(&ctx->refs);
9888 if (ctx->restricted) {
9889 if (opcode >= IORING_REGISTER_LAST) {
9894 if (!test_bit(opcode, ctx->restrictions.register_op)) {
9901 case IORING_REGISTER_BUFFERS:
9902 ret = io_sqe_buffer_register(ctx, arg, nr_args);
9904 case IORING_UNREGISTER_BUFFERS:
9908 ret = io_sqe_buffer_unregister(ctx);
9910 case IORING_REGISTER_FILES:
9911 ret = io_sqe_files_register(ctx, arg, nr_args);
9913 case IORING_UNREGISTER_FILES:
9917 ret = io_sqe_files_unregister(ctx);
9919 case IORING_REGISTER_FILES_UPDATE:
9920 ret = io_sqe_files_update(ctx, arg, nr_args);
9922 case IORING_REGISTER_EVENTFD:
9923 case IORING_REGISTER_EVENTFD_ASYNC:
9927 ret = io_eventfd_register(ctx, arg);
9930 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
9931 ctx->eventfd_async = 1;
9933 ctx->eventfd_async = 0;
9935 case IORING_UNREGISTER_EVENTFD:
9939 ret = io_eventfd_unregister(ctx);
9941 case IORING_REGISTER_PROBE:
9943 if (!arg || nr_args > 256)
9945 ret = io_probe(ctx, arg, nr_args);
9947 case IORING_REGISTER_PERSONALITY:
9951 ret = io_register_personality(ctx);
9953 case IORING_UNREGISTER_PERSONALITY:
9957 ret = io_unregister_personality(ctx, nr_args);
9959 case IORING_REGISTER_ENABLE_RINGS:
9963 ret = io_register_enable_rings(ctx);
9965 case IORING_REGISTER_RESTRICTIONS:
9966 ret = io_register_restrictions(ctx, arg, nr_args);
9974 if (io_register_op_must_quiesce(opcode)) {
9975 /* bring the ctx back to life */
9976 percpu_ref_reinit(&ctx->refs);
9978 reinit_completion(&ctx->ref_comp);
9983 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
9984 void __user *, arg, unsigned int, nr_args)
9986 struct io_ring_ctx *ctx;
9995 if (f.file->f_op != &io_uring_fops)
9998 ctx = f.file->private_data;
10000 mutex_lock(&ctx->uring_lock);
10001 ret = __io_uring_register(ctx, opcode, arg, nr_args);
10002 mutex_unlock(&ctx->uring_lock);
10003 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
10004 ctx->cq_ev_fd != NULL, ret);
10010 static int __init io_uring_init(void)
10012 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
10013 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
10014 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
10017 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
10018 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
10019 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
10020 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
10021 BUILD_BUG_SQE_ELEM(1, __u8, flags);
10022 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
10023 BUILD_BUG_SQE_ELEM(4, __s32, fd);
10024 BUILD_BUG_SQE_ELEM(8, __u64, off);
10025 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
10026 BUILD_BUG_SQE_ELEM(16, __u64, addr);
10027 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
10028 BUILD_BUG_SQE_ELEM(24, __u32, len);
10029 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
10030 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
10031 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
10032 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
10033 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
10034 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
10035 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
10036 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
10037 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
10038 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
10039 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
10040 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
10041 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
10042 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
10043 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
10044 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
10045 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
10046 BUILD_BUG_SQE_ELEM(42, __u16, personality);
10047 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
10049 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
10050 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
10051 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
10054 __initcall(io_uring_init);