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 __io_complete_rw(struct io_kiocb *req, long res, long res2,
996 struct io_comp_state *cs);
997 static void io_cqring_fill_event(struct io_kiocb *req, long res);
998 static void io_put_req(struct io_kiocb *req);
999 static void io_put_req_deferred(struct io_kiocb *req, int nr);
1000 static void io_double_put_req(struct io_kiocb *req);
1001 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
1002 static void __io_queue_linked_timeout(struct io_kiocb *req);
1003 static void io_queue_linked_timeout(struct io_kiocb *req);
1004 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
1005 struct io_uring_files_update *ip,
1007 static void __io_clean_op(struct io_kiocb *req);
1008 static struct file *io_file_get(struct io_submit_state *state,
1009 struct io_kiocb *req, int fd, bool fixed);
1010 static void __io_queue_sqe(struct io_kiocb *req, struct io_comp_state *cs);
1011 static void io_file_put_work(struct work_struct *work);
1013 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
1014 struct iovec **iovec, struct iov_iter *iter,
1016 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
1017 const struct iovec *fast_iov,
1018 struct iov_iter *iter, bool force);
1020 static struct kmem_cache *req_cachep;
1022 static const struct file_operations io_uring_fops;
1024 struct sock *io_uring_get_socket(struct file *file)
1026 #if defined(CONFIG_UNIX)
1027 if (file->f_op == &io_uring_fops) {
1028 struct io_ring_ctx *ctx = file->private_data;
1030 return ctx->ring_sock->sk;
1035 EXPORT_SYMBOL(io_uring_get_socket);
1037 #define io_for_each_link(pos, head) \
1038 for (pos = (head); pos; pos = pos->link)
1040 static inline void io_clean_op(struct io_kiocb *req)
1042 if (req->flags & (REQ_F_NEED_CLEANUP | REQ_F_BUFFER_SELECTED |
1047 static inline void io_set_resource_node(struct io_kiocb *req)
1049 struct io_ring_ctx *ctx = req->ctx;
1051 if (!req->fixed_file_refs) {
1052 req->fixed_file_refs = &ctx->file_data->node->refs;
1053 percpu_ref_get(req->fixed_file_refs);
1057 static bool io_match_task(struct io_kiocb *head,
1058 struct task_struct *task,
1059 struct files_struct *files)
1061 struct io_kiocb *req;
1063 if (task && head->task != task)
1068 io_for_each_link(req, head) {
1069 if ((req->flags & REQ_F_WORK_INITIALIZED) &&
1070 (req->work.flags & IO_WQ_WORK_FILES) &&
1071 req->work.identity->files == files)
1077 static void io_sq_thread_drop_mm_files(void)
1079 struct files_struct *files = current->files;
1080 struct mm_struct *mm = current->mm;
1083 kthread_unuse_mm(mm);
1088 struct nsproxy *nsproxy = current->nsproxy;
1091 current->files = NULL;
1092 current->nsproxy = NULL;
1093 task_unlock(current);
1094 put_files_struct(files);
1095 put_nsproxy(nsproxy);
1099 static int __io_sq_thread_acquire_files(struct io_ring_ctx *ctx)
1101 if (!current->files) {
1102 struct files_struct *files;
1103 struct nsproxy *nsproxy;
1105 task_lock(ctx->sqo_task);
1106 files = ctx->sqo_task->files;
1108 task_unlock(ctx->sqo_task);
1111 atomic_inc(&files->count);
1112 get_nsproxy(ctx->sqo_task->nsproxy);
1113 nsproxy = ctx->sqo_task->nsproxy;
1114 task_unlock(ctx->sqo_task);
1117 current->files = files;
1118 current->nsproxy = nsproxy;
1119 task_unlock(current);
1124 static int __io_sq_thread_acquire_mm(struct io_ring_ctx *ctx)
1126 struct mm_struct *mm;
1131 /* Should never happen */
1132 if (unlikely(!(ctx->flags & IORING_SETUP_SQPOLL)))
1135 task_lock(ctx->sqo_task);
1136 mm = ctx->sqo_task->mm;
1137 if (unlikely(!mm || !mmget_not_zero(mm)))
1139 task_unlock(ctx->sqo_task);
1149 static int io_sq_thread_acquire_mm_files(struct io_ring_ctx *ctx,
1150 struct io_kiocb *req)
1152 const struct io_op_def *def = &io_op_defs[req->opcode];
1155 if (def->work_flags & IO_WQ_WORK_MM) {
1156 ret = __io_sq_thread_acquire_mm(ctx);
1161 if (def->needs_file || (def->work_flags & IO_WQ_WORK_FILES)) {
1162 ret = __io_sq_thread_acquire_files(ctx);
1170 static void io_sq_thread_associate_blkcg(struct io_ring_ctx *ctx,
1171 struct cgroup_subsys_state **cur_css)
1174 #ifdef CONFIG_BLK_CGROUP
1175 /* puts the old one when swapping */
1176 if (*cur_css != ctx->sqo_blkcg_css) {
1177 kthread_associate_blkcg(ctx->sqo_blkcg_css);
1178 *cur_css = ctx->sqo_blkcg_css;
1183 static void io_sq_thread_unassociate_blkcg(void)
1185 #ifdef CONFIG_BLK_CGROUP
1186 kthread_associate_blkcg(NULL);
1190 static inline void req_set_fail_links(struct io_kiocb *req)
1192 if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
1193 req->flags |= REQ_F_FAIL_LINK;
1197 * None of these are dereferenced, they are simply used to check if any of
1198 * them have changed. If we're under current and check they are still the
1199 * same, we're fine to grab references to them for actual out-of-line use.
1201 static void io_init_identity(struct io_identity *id)
1203 id->files = current->files;
1204 id->mm = current->mm;
1205 #ifdef CONFIG_BLK_CGROUP
1207 id->blkcg_css = blkcg_css();
1210 id->creds = current_cred();
1211 id->nsproxy = current->nsproxy;
1212 id->fs = current->fs;
1213 id->fsize = rlimit(RLIMIT_FSIZE);
1215 id->loginuid = current->loginuid;
1216 id->sessionid = current->sessionid;
1218 refcount_set(&id->count, 1);
1221 static inline void __io_req_init_async(struct io_kiocb *req)
1223 memset(&req->work, 0, sizeof(req->work));
1224 req->flags |= REQ_F_WORK_INITIALIZED;
1228 * Note: must call io_req_init_async() for the first time you
1229 * touch any members of io_wq_work.
1231 static inline void io_req_init_async(struct io_kiocb *req)
1233 struct io_uring_task *tctx = current->io_uring;
1235 if (req->flags & REQ_F_WORK_INITIALIZED)
1238 __io_req_init_async(req);
1240 /* Grab a ref if this isn't our static identity */
1241 req->work.identity = tctx->identity;
1242 if (tctx->identity != &tctx->__identity)
1243 refcount_inc(&req->work.identity->count);
1246 static inline bool io_async_submit(struct io_ring_ctx *ctx)
1248 return ctx->flags & IORING_SETUP_SQPOLL;
1251 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1253 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1255 complete(&ctx->ref_comp);
1258 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1260 return !req->timeout.off;
1263 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1265 struct io_ring_ctx *ctx;
1268 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1272 ctx->fallback_req = kmem_cache_alloc(req_cachep, GFP_KERNEL);
1273 if (!ctx->fallback_req)
1277 * Use 5 bits less than the max cq entries, that should give us around
1278 * 32 entries per hash list if totally full and uniformly spread.
1280 hash_bits = ilog2(p->cq_entries);
1284 ctx->cancel_hash_bits = hash_bits;
1285 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1287 if (!ctx->cancel_hash)
1289 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1291 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1292 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1295 ctx->flags = p->flags;
1296 init_waitqueue_head(&ctx->sqo_sq_wait);
1297 INIT_LIST_HEAD(&ctx->sqd_list);
1298 init_waitqueue_head(&ctx->cq_wait);
1299 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1300 init_completion(&ctx->ref_comp);
1301 init_completion(&ctx->sq_thread_comp);
1302 idr_init(&ctx->io_buffer_idr);
1303 idr_init(&ctx->personality_idr);
1304 mutex_init(&ctx->uring_lock);
1305 init_waitqueue_head(&ctx->wait);
1306 spin_lock_init(&ctx->completion_lock);
1307 INIT_LIST_HEAD(&ctx->iopoll_list);
1308 INIT_LIST_HEAD(&ctx->defer_list);
1309 INIT_LIST_HEAD(&ctx->timeout_list);
1310 spin_lock_init(&ctx->inflight_lock);
1311 INIT_LIST_HEAD(&ctx->inflight_list);
1312 INIT_DELAYED_WORK(&ctx->file_put_work, io_file_put_work);
1313 init_llist_head(&ctx->file_put_llist);
1316 if (ctx->fallback_req)
1317 kmem_cache_free(req_cachep, ctx->fallback_req);
1318 kfree(ctx->cancel_hash);
1323 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1325 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1326 struct io_ring_ctx *ctx = req->ctx;
1328 return seq != ctx->cached_cq_tail
1329 + READ_ONCE(ctx->cached_cq_overflow);
1335 static void __io_commit_cqring(struct io_ring_ctx *ctx)
1337 struct io_rings *rings = ctx->rings;
1339 /* order cqe stores with ring update */
1340 smp_store_release(&rings->cq.tail, ctx->cached_cq_tail);
1342 if (wq_has_sleeper(&ctx->cq_wait)) {
1343 wake_up_interruptible(&ctx->cq_wait);
1344 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1348 static void io_put_identity(struct io_uring_task *tctx, struct io_kiocb *req)
1350 if (req->work.identity == &tctx->__identity)
1352 if (refcount_dec_and_test(&req->work.identity->count))
1353 kfree(req->work.identity);
1356 static void io_req_clean_work(struct io_kiocb *req)
1358 if (!(req->flags & REQ_F_WORK_INITIALIZED))
1361 req->flags &= ~REQ_F_WORK_INITIALIZED;
1363 if (req->work.flags & IO_WQ_WORK_MM) {
1364 mmdrop(req->work.identity->mm);
1365 req->work.flags &= ~IO_WQ_WORK_MM;
1367 #ifdef CONFIG_BLK_CGROUP
1368 if (req->work.flags & IO_WQ_WORK_BLKCG) {
1369 css_put(req->work.identity->blkcg_css);
1370 req->work.flags &= ~IO_WQ_WORK_BLKCG;
1373 if (req->work.flags & IO_WQ_WORK_CREDS) {
1374 put_cred(req->work.identity->creds);
1375 req->work.flags &= ~IO_WQ_WORK_CREDS;
1377 if (req->work.flags & IO_WQ_WORK_FS) {
1378 struct fs_struct *fs = req->work.identity->fs;
1380 spin_lock(&req->work.identity->fs->lock);
1383 spin_unlock(&req->work.identity->fs->lock);
1386 req->work.flags &= ~IO_WQ_WORK_FS;
1389 io_put_identity(req->task->io_uring, req);
1393 * Create a private copy of io_identity, since some fields don't match
1394 * the current context.
1396 static bool io_identity_cow(struct io_kiocb *req)
1398 struct io_uring_task *tctx = current->io_uring;
1399 const struct cred *creds = NULL;
1400 struct io_identity *id;
1402 if (req->work.flags & IO_WQ_WORK_CREDS)
1403 creds = req->work.identity->creds;
1405 id = kmemdup(req->work.identity, sizeof(*id), GFP_KERNEL);
1406 if (unlikely(!id)) {
1407 req->work.flags |= IO_WQ_WORK_CANCEL;
1412 * We can safely just re-init the creds we copied Either the field
1413 * matches the current one, or we haven't grabbed it yet. The only
1414 * exception is ->creds, through registered personalities, so handle
1415 * that one separately.
1417 io_init_identity(id);
1421 /* add one for this request */
1422 refcount_inc(&id->count);
1424 /* drop tctx and req identity references, if needed */
1425 if (tctx->identity != &tctx->__identity &&
1426 refcount_dec_and_test(&tctx->identity->count))
1427 kfree(tctx->identity);
1428 if (req->work.identity != &tctx->__identity &&
1429 refcount_dec_and_test(&req->work.identity->count))
1430 kfree(req->work.identity);
1432 req->work.identity = id;
1433 tctx->identity = id;
1437 static bool io_grab_identity(struct io_kiocb *req)
1439 const struct io_op_def *def = &io_op_defs[req->opcode];
1440 struct io_identity *id = req->work.identity;
1441 struct io_ring_ctx *ctx = req->ctx;
1443 if (def->work_flags & IO_WQ_WORK_FSIZE) {
1444 if (id->fsize != rlimit(RLIMIT_FSIZE))
1446 req->work.flags |= IO_WQ_WORK_FSIZE;
1448 #ifdef CONFIG_BLK_CGROUP
1449 if (!(req->work.flags & IO_WQ_WORK_BLKCG) &&
1450 (def->work_flags & IO_WQ_WORK_BLKCG)) {
1452 if (id->blkcg_css != blkcg_css()) {
1457 * This should be rare, either the cgroup is dying or the task
1458 * is moving cgroups. Just punt to root for the handful of ios.
1460 if (css_tryget_online(id->blkcg_css))
1461 req->work.flags |= IO_WQ_WORK_BLKCG;
1465 if (!(req->work.flags & IO_WQ_WORK_CREDS)) {
1466 if (id->creds != current_cred())
1468 get_cred(id->creds);
1469 req->work.flags |= IO_WQ_WORK_CREDS;
1472 if (!uid_eq(current->loginuid, id->loginuid) ||
1473 current->sessionid != id->sessionid)
1476 if (!(req->work.flags & IO_WQ_WORK_FS) &&
1477 (def->work_flags & IO_WQ_WORK_FS)) {
1478 if (current->fs != id->fs)
1480 spin_lock(&id->fs->lock);
1481 if (!id->fs->in_exec) {
1483 req->work.flags |= IO_WQ_WORK_FS;
1485 req->work.flags |= IO_WQ_WORK_CANCEL;
1487 spin_unlock(¤t->fs->lock);
1489 if (!(req->work.flags & IO_WQ_WORK_FILES) &&
1490 (def->work_flags & IO_WQ_WORK_FILES) &&
1491 !(req->flags & REQ_F_NO_FILE_TABLE)) {
1492 if (id->files != current->files ||
1493 id->nsproxy != current->nsproxy)
1495 atomic_inc(&id->files->count);
1496 get_nsproxy(id->nsproxy);
1497 req->flags |= REQ_F_INFLIGHT;
1499 spin_lock_irq(&ctx->inflight_lock);
1500 list_add(&req->inflight_entry, &ctx->inflight_list);
1501 spin_unlock_irq(&ctx->inflight_lock);
1502 req->work.flags |= IO_WQ_WORK_FILES;
1508 static void io_prep_async_work(struct io_kiocb *req)
1510 const struct io_op_def *def = &io_op_defs[req->opcode];
1511 struct io_ring_ctx *ctx = req->ctx;
1512 struct io_identity *id;
1514 io_req_init_async(req);
1515 id = req->work.identity;
1517 if (req->flags & REQ_F_FORCE_ASYNC)
1518 req->work.flags |= IO_WQ_WORK_CONCURRENT;
1520 if (req->flags & REQ_F_ISREG) {
1521 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1522 io_wq_hash_work(&req->work, file_inode(req->file));
1524 if (def->unbound_nonreg_file)
1525 req->work.flags |= IO_WQ_WORK_UNBOUND;
1528 /* ->mm can never change on us */
1529 if (!(req->work.flags & IO_WQ_WORK_MM) &&
1530 (def->work_flags & IO_WQ_WORK_MM)) {
1532 req->work.flags |= IO_WQ_WORK_MM;
1535 /* if we fail grabbing identity, we must COW, regrab, and retry */
1536 if (io_grab_identity(req))
1539 if (!io_identity_cow(req))
1542 /* can't fail at this point */
1543 if (!io_grab_identity(req))
1547 static void io_prep_async_link(struct io_kiocb *req)
1549 struct io_kiocb *cur;
1551 io_for_each_link(cur, req)
1552 io_prep_async_work(cur);
1555 static struct io_kiocb *__io_queue_async_work(struct io_kiocb *req)
1557 struct io_ring_ctx *ctx = req->ctx;
1558 struct io_kiocb *link = io_prep_linked_timeout(req);
1560 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1561 &req->work, req->flags);
1562 io_wq_enqueue(ctx->io_wq, &req->work);
1566 static void io_queue_async_work(struct io_kiocb *req)
1568 struct io_kiocb *link;
1570 /* init ->work of the whole link before punting */
1571 io_prep_async_link(req);
1572 link = __io_queue_async_work(req);
1575 io_queue_linked_timeout(link);
1578 static void io_kill_timeout(struct io_kiocb *req)
1580 struct io_timeout_data *io = req->async_data;
1583 ret = hrtimer_try_to_cancel(&io->timer);
1585 atomic_set(&req->ctx->cq_timeouts,
1586 atomic_read(&req->ctx->cq_timeouts) + 1);
1587 list_del_init(&req->timeout.list);
1588 io_cqring_fill_event(req, 0);
1589 io_put_req_deferred(req, 1);
1594 * Returns true if we found and killed one or more timeouts
1596 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
1597 struct files_struct *files)
1599 struct io_kiocb *req, *tmp;
1602 spin_lock_irq(&ctx->completion_lock);
1603 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
1604 if (io_match_task(req, tsk, files)) {
1605 io_kill_timeout(req);
1609 spin_unlock_irq(&ctx->completion_lock);
1610 return canceled != 0;
1613 static void __io_queue_deferred(struct io_ring_ctx *ctx)
1616 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1617 struct io_defer_entry, list);
1618 struct io_kiocb *link;
1620 if (req_need_defer(de->req, de->seq))
1622 list_del_init(&de->list);
1623 /* punt-init is done before queueing for defer */
1624 link = __io_queue_async_work(de->req);
1626 __io_queue_linked_timeout(link);
1627 /* drop submission reference */
1628 io_put_req_deferred(link, 1);
1631 } while (!list_empty(&ctx->defer_list));
1634 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1636 while (!list_empty(&ctx->timeout_list)) {
1637 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1638 struct io_kiocb, timeout.list);
1640 if (io_is_timeout_noseq(req))
1642 if (req->timeout.target_seq != ctx->cached_cq_tail
1643 - atomic_read(&ctx->cq_timeouts))
1646 list_del_init(&req->timeout.list);
1647 io_kill_timeout(req);
1651 static void io_commit_cqring(struct io_ring_ctx *ctx)
1653 io_flush_timeouts(ctx);
1654 __io_commit_cqring(ctx);
1656 if (unlikely(!list_empty(&ctx->defer_list)))
1657 __io_queue_deferred(ctx);
1660 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1662 struct io_rings *r = ctx->rings;
1664 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == r->sq_ring_entries;
1667 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1669 struct io_rings *rings = ctx->rings;
1672 tail = ctx->cached_cq_tail;
1674 * writes to the cq entry need to come after reading head; the
1675 * control dependency is enough as we're using WRITE_ONCE to
1678 if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries)
1681 ctx->cached_cq_tail++;
1682 return &rings->cqes[tail & ctx->cq_mask];
1685 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1689 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1691 if (!ctx->eventfd_async)
1693 return io_wq_current_is_worker();
1696 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1698 if (waitqueue_active(&ctx->wait))
1699 wake_up(&ctx->wait);
1700 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1701 wake_up(&ctx->sq_data->wait);
1702 if (io_should_trigger_evfd(ctx))
1703 eventfd_signal(ctx->cq_ev_fd, 1);
1706 static void io_cqring_mark_overflow(struct io_ring_ctx *ctx)
1708 if (list_empty(&ctx->cq_overflow_list)) {
1709 clear_bit(0, &ctx->sq_check_overflow);
1710 clear_bit(0, &ctx->cq_check_overflow);
1711 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
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 (list_empty_careful(&ctx->cq_overflow_list))
1729 if ((ctx->cached_cq_tail - READ_ONCE(rings->cq.head) ==
1730 rings->cq_ring_entries))
1734 spin_lock_irqsave(&ctx->completion_lock, flags);
1736 /* if force is set, the ring is going away. always drop after that */
1738 ctx->cq_overflow_flushed = 1;
1741 list_for_each_entry_safe(req, tmp, &ctx->cq_overflow_list, compl.list) {
1742 if (!io_match_task(req, tsk, files))
1745 cqe = io_get_cqring(ctx);
1749 list_move(&req->compl.list, &list);
1751 WRITE_ONCE(cqe->user_data, req->user_data);
1752 WRITE_ONCE(cqe->res, req->result);
1753 WRITE_ONCE(cqe->flags, req->compl.cflags);
1755 ctx->cached_cq_overflow++;
1756 WRITE_ONCE(ctx->rings->cq_overflow,
1757 ctx->cached_cq_overflow);
1761 io_commit_cqring(ctx);
1762 io_cqring_mark_overflow(ctx);
1764 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1765 io_cqring_ev_posted(ctx);
1767 while (!list_empty(&list)) {
1768 req = list_first_entry(&list, struct io_kiocb, compl.list);
1769 list_del(&req->compl.list);
1776 static void __io_cqring_fill_event(struct io_kiocb *req, long res, long cflags)
1778 struct io_ring_ctx *ctx = req->ctx;
1779 struct io_uring_cqe *cqe;
1781 trace_io_uring_complete(ctx, req->user_data, res);
1784 * If we can't get a cq entry, userspace overflowed the
1785 * submission (by quite a lot). Increment the overflow count in
1788 cqe = io_get_cqring(ctx);
1790 WRITE_ONCE(cqe->user_data, req->user_data);
1791 WRITE_ONCE(cqe->res, res);
1792 WRITE_ONCE(cqe->flags, cflags);
1793 } else if (ctx->cq_overflow_flushed ||
1794 atomic_read(&req->task->io_uring->in_idle)) {
1796 * If we're in ring overflow flush mode, or in task cancel mode,
1797 * then we cannot store the request for later flushing, we need
1798 * to drop it on the floor.
1800 ctx->cached_cq_overflow++;
1801 WRITE_ONCE(ctx->rings->cq_overflow, ctx->cached_cq_overflow);
1803 if (list_empty(&ctx->cq_overflow_list)) {
1804 set_bit(0, &ctx->sq_check_overflow);
1805 set_bit(0, &ctx->cq_check_overflow);
1806 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
1810 req->compl.cflags = cflags;
1811 refcount_inc(&req->refs);
1812 list_add_tail(&req->compl.list, &ctx->cq_overflow_list);
1816 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1818 __io_cqring_fill_event(req, res, 0);
1821 static void io_cqring_add_event(struct io_kiocb *req, long res, long cflags)
1823 struct io_ring_ctx *ctx = req->ctx;
1824 unsigned long flags;
1826 spin_lock_irqsave(&ctx->completion_lock, flags);
1827 __io_cqring_fill_event(req, res, cflags);
1828 io_commit_cqring(ctx);
1829 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1831 io_cqring_ev_posted(ctx);
1834 static void io_submit_flush_completions(struct io_comp_state *cs)
1836 struct io_ring_ctx *ctx = cs->ctx;
1838 spin_lock_irq(&ctx->completion_lock);
1839 while (!list_empty(&cs->list)) {
1840 struct io_kiocb *req;
1842 req = list_first_entry(&cs->list, struct io_kiocb, compl.list);
1843 list_del(&req->compl.list);
1844 __io_cqring_fill_event(req, req->result, req->compl.cflags);
1847 * io_free_req() doesn't care about completion_lock unless one
1848 * of these flags is set. REQ_F_WORK_INITIALIZED is in the list
1849 * because of a potential deadlock with req->work.fs->lock
1851 if (req->flags & (REQ_F_FAIL_LINK|REQ_F_LINK_TIMEOUT
1852 |REQ_F_WORK_INITIALIZED)) {
1853 spin_unlock_irq(&ctx->completion_lock);
1855 spin_lock_irq(&ctx->completion_lock);
1860 io_commit_cqring(ctx);
1861 spin_unlock_irq(&ctx->completion_lock);
1863 io_cqring_ev_posted(ctx);
1867 static void __io_req_complete(struct io_kiocb *req, long res, unsigned cflags,
1868 struct io_comp_state *cs)
1871 io_cqring_add_event(req, res, cflags);
1876 req->compl.cflags = cflags;
1877 list_add_tail(&req->compl.list, &cs->list);
1879 io_submit_flush_completions(cs);
1883 static void io_req_complete(struct io_kiocb *req, long res)
1885 __io_req_complete(req, res, 0, NULL);
1888 static inline bool io_is_fallback_req(struct io_kiocb *req)
1890 return req == (struct io_kiocb *)
1891 ((unsigned long) req->ctx->fallback_req & ~1UL);
1894 static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx)
1896 struct io_kiocb *req;
1898 req = ctx->fallback_req;
1899 if (!test_and_set_bit_lock(0, (unsigned long *) &ctx->fallback_req))
1905 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx,
1906 struct io_submit_state *state)
1908 if (!state->free_reqs) {
1909 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1913 sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs));
1914 ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs);
1917 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1918 * retry single alloc to be on the safe side.
1920 if (unlikely(ret <= 0)) {
1921 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1922 if (!state->reqs[0])
1926 state->free_reqs = ret;
1930 return state->reqs[state->free_reqs];
1932 return io_get_fallback_req(ctx);
1935 static inline void io_put_file(struct io_kiocb *req, struct file *file,
1942 static void io_dismantle_req(struct io_kiocb *req)
1946 if (req->async_data)
1947 kfree(req->async_data);
1949 io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
1950 if (req->fixed_file_refs)
1951 percpu_ref_put(req->fixed_file_refs);
1952 io_req_clean_work(req);
1955 static void __io_free_req(struct io_kiocb *req)
1957 struct io_uring_task *tctx = req->task->io_uring;
1958 struct io_ring_ctx *ctx = req->ctx;
1960 io_dismantle_req(req);
1962 percpu_counter_dec(&tctx->inflight);
1963 if (atomic_read(&tctx->in_idle))
1964 wake_up(&tctx->wait);
1965 put_task_struct(req->task);
1967 if (likely(!io_is_fallback_req(req)))
1968 kmem_cache_free(req_cachep, req);
1970 clear_bit_unlock(0, (unsigned long *) &ctx->fallback_req);
1971 percpu_ref_put(&ctx->refs);
1974 static inline void io_remove_next_linked(struct io_kiocb *req)
1976 struct io_kiocb *nxt = req->link;
1978 req->link = nxt->link;
1982 static void io_kill_linked_timeout(struct io_kiocb *req)
1984 struct io_ring_ctx *ctx = req->ctx;
1985 struct io_kiocb *link;
1986 bool cancelled = false;
1987 unsigned long flags;
1989 spin_lock_irqsave(&ctx->completion_lock, flags);
1993 * Can happen if a linked timeout fired and link had been like
1994 * req -> link t-out -> link t-out [-> ...]
1996 if (link && (link->flags & REQ_F_LTIMEOUT_ACTIVE)) {
1997 struct io_timeout_data *io = link->async_data;
2000 io_remove_next_linked(req);
2001 link->timeout.head = NULL;
2002 ret = hrtimer_try_to_cancel(&io->timer);
2004 io_cqring_fill_event(link, -ECANCELED);
2005 io_commit_cqring(ctx);
2009 req->flags &= ~REQ_F_LINK_TIMEOUT;
2010 spin_unlock_irqrestore(&ctx->completion_lock, flags);
2013 io_cqring_ev_posted(ctx);
2019 static void io_fail_links(struct io_kiocb *req)
2021 struct io_kiocb *link, *nxt;
2022 struct io_ring_ctx *ctx = req->ctx;
2023 unsigned long flags;
2025 spin_lock_irqsave(&ctx->completion_lock, flags);
2033 trace_io_uring_fail_link(req, link);
2034 io_cqring_fill_event(link, -ECANCELED);
2037 * It's ok to free under spinlock as they're not linked anymore,
2038 * but avoid REQ_F_WORK_INITIALIZED because it may deadlock on
2041 if (link->flags & REQ_F_WORK_INITIALIZED)
2042 io_put_req_deferred(link, 2);
2044 io_double_put_req(link);
2047 io_commit_cqring(ctx);
2048 spin_unlock_irqrestore(&ctx->completion_lock, flags);
2050 io_cqring_ev_posted(ctx);
2053 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
2055 if (req->flags & REQ_F_LINK_TIMEOUT)
2056 io_kill_linked_timeout(req);
2059 * If LINK is set, we have dependent requests in this chain. If we
2060 * didn't fail this request, queue the first one up, moving any other
2061 * dependencies to the next request. In case of failure, fail the rest
2064 if (likely(!(req->flags & REQ_F_FAIL_LINK))) {
2065 struct io_kiocb *nxt = req->link;
2074 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
2076 if (likely(!(req->link) && !(req->flags & REQ_F_LINK_TIMEOUT)))
2078 return __io_req_find_next(req);
2081 static int io_req_task_work_add(struct io_kiocb *req)
2083 struct task_struct *tsk = req->task;
2084 struct io_ring_ctx *ctx = req->ctx;
2085 enum task_work_notify_mode notify;
2088 if (tsk->flags & PF_EXITING)
2092 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
2093 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
2094 * processing task_work. There's no reliable way to tell if TWA_RESUME
2098 if (!(ctx->flags & IORING_SETUP_SQPOLL))
2099 notify = TWA_SIGNAL;
2101 ret = task_work_add(tsk, &req->task_work, notify);
2103 wake_up_process(tsk);
2108 static void __io_req_task_cancel(struct io_kiocb *req, int error)
2110 struct io_ring_ctx *ctx = req->ctx;
2112 spin_lock_irq(&ctx->completion_lock);
2113 io_cqring_fill_event(req, error);
2114 io_commit_cqring(ctx);
2115 spin_unlock_irq(&ctx->completion_lock);
2117 io_cqring_ev_posted(ctx);
2118 req_set_fail_links(req);
2119 io_double_put_req(req);
2122 static void io_req_task_cancel(struct callback_head *cb)
2124 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2125 struct io_ring_ctx *ctx = req->ctx;
2127 __io_req_task_cancel(req, -ECANCELED);
2128 percpu_ref_put(&ctx->refs);
2131 static void __io_req_task_submit(struct io_kiocb *req)
2133 struct io_ring_ctx *ctx = req->ctx;
2135 if (!__io_sq_thread_acquire_mm(ctx) &&
2136 !__io_sq_thread_acquire_files(ctx)) {
2137 mutex_lock(&ctx->uring_lock);
2138 __io_queue_sqe(req, NULL);
2139 mutex_unlock(&ctx->uring_lock);
2141 __io_req_task_cancel(req, -EFAULT);
2145 static void io_req_task_submit(struct callback_head *cb)
2147 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2148 struct io_ring_ctx *ctx = req->ctx;
2150 __io_req_task_submit(req);
2151 percpu_ref_put(&ctx->refs);
2154 static void io_req_task_queue(struct io_kiocb *req)
2158 init_task_work(&req->task_work, io_req_task_submit);
2159 percpu_ref_get(&req->ctx->refs);
2161 ret = io_req_task_work_add(req);
2162 if (unlikely(ret)) {
2163 struct task_struct *tsk;
2165 init_task_work(&req->task_work, io_req_task_cancel);
2166 tsk = io_wq_get_task(req->ctx->io_wq);
2167 task_work_add(tsk, &req->task_work, TWA_NONE);
2168 wake_up_process(tsk);
2172 static inline void io_queue_next(struct io_kiocb *req)
2174 struct io_kiocb *nxt = io_req_find_next(req);
2177 io_req_task_queue(nxt);
2180 static void io_free_req(struct io_kiocb *req)
2187 void *reqs[IO_IOPOLL_BATCH];
2190 struct task_struct *task;
2194 static inline void io_init_req_batch(struct req_batch *rb)
2201 static void __io_req_free_batch_flush(struct io_ring_ctx *ctx,
2202 struct req_batch *rb)
2204 kmem_cache_free_bulk(req_cachep, rb->to_free, rb->reqs);
2205 percpu_ref_put_many(&ctx->refs, rb->to_free);
2209 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2210 struct req_batch *rb)
2213 __io_req_free_batch_flush(ctx, rb);
2215 struct io_uring_task *tctx = rb->task->io_uring;
2217 percpu_counter_sub(&tctx->inflight, rb->task_refs);
2218 put_task_struct_many(rb->task, rb->task_refs);
2223 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req)
2225 if (unlikely(io_is_fallback_req(req))) {
2231 if (req->task != rb->task) {
2233 struct io_uring_task *tctx = rb->task->io_uring;
2235 percpu_counter_sub(&tctx->inflight, rb->task_refs);
2236 put_task_struct_many(rb->task, rb->task_refs);
2238 rb->task = req->task;
2243 io_dismantle_req(req);
2244 rb->reqs[rb->to_free++] = req;
2245 if (unlikely(rb->to_free == ARRAY_SIZE(rb->reqs)))
2246 __io_req_free_batch_flush(req->ctx, rb);
2250 * Drop reference to request, return next in chain (if there is one) if this
2251 * was the last reference to this request.
2253 static struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2255 struct io_kiocb *nxt = NULL;
2257 if (refcount_dec_and_test(&req->refs)) {
2258 nxt = io_req_find_next(req);
2264 static void io_put_req(struct io_kiocb *req)
2266 if (refcount_dec_and_test(&req->refs))
2270 static void io_put_req_deferred_cb(struct callback_head *cb)
2272 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2277 static void io_free_req_deferred(struct io_kiocb *req)
2281 init_task_work(&req->task_work, io_put_req_deferred_cb);
2282 ret = io_req_task_work_add(req);
2283 if (unlikely(ret)) {
2284 struct task_struct *tsk;
2286 tsk = io_wq_get_task(req->ctx->io_wq);
2287 task_work_add(tsk, &req->task_work, TWA_NONE);
2288 wake_up_process(tsk);
2292 static inline void io_put_req_deferred(struct io_kiocb *req, int refs)
2294 if (refcount_sub_and_test(refs, &req->refs))
2295 io_free_req_deferred(req);
2298 static struct io_wq_work *io_steal_work(struct io_kiocb *req)
2300 struct io_kiocb *nxt;
2303 * A ref is owned by io-wq in which context we're. So, if that's the
2304 * last one, it's safe to steal next work. False negatives are Ok,
2305 * it just will be re-punted async in io_put_work()
2307 if (refcount_read(&req->refs) != 1)
2310 nxt = io_req_find_next(req);
2311 return nxt ? &nxt->work : NULL;
2314 static void io_double_put_req(struct io_kiocb *req)
2316 /* drop both submit and complete references */
2317 if (refcount_sub_and_test(2, &req->refs))
2321 static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush)
2323 struct io_rings *rings = ctx->rings;
2325 if (test_bit(0, &ctx->cq_check_overflow)) {
2327 * noflush == true is from the waitqueue handler, just ensure
2328 * we wake up the task, and the next invocation will flush the
2329 * entries. We cannot safely to it from here.
2334 io_cqring_overflow_flush(ctx, false, NULL, NULL);
2337 /* See comment at the top of this file */
2339 return ctx->cached_cq_tail - READ_ONCE(rings->cq.head);
2342 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2344 struct io_rings *rings = ctx->rings;
2346 /* make sure SQ entry isn't read before tail */
2347 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2350 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2352 unsigned int cflags;
2354 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2355 cflags |= IORING_CQE_F_BUFFER;
2356 req->flags &= ~REQ_F_BUFFER_SELECTED;
2361 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2363 struct io_buffer *kbuf;
2365 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2366 return io_put_kbuf(req, kbuf);
2369 static inline bool io_run_task_work(void)
2372 * Not safe to run on exiting task, and the task_work handling will
2373 * not add work to such a task.
2375 if (unlikely(current->flags & PF_EXITING))
2377 if (current->task_works) {
2378 __set_current_state(TASK_RUNNING);
2386 static void io_iopoll_queue(struct list_head *again)
2388 struct io_kiocb *req;
2391 req = list_first_entry(again, struct io_kiocb, inflight_entry);
2392 list_del(&req->inflight_entry);
2393 __io_complete_rw(req, -EAGAIN, 0, NULL);
2394 } while (!list_empty(again));
2398 * Find and free completed poll iocbs
2400 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2401 struct list_head *done)
2403 struct req_batch rb;
2404 struct io_kiocb *req;
2407 /* order with ->result store in io_complete_rw_iopoll() */
2410 io_init_req_batch(&rb);
2411 while (!list_empty(done)) {
2414 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2415 if (READ_ONCE(req->result) == -EAGAIN) {
2417 req->iopoll_completed = 0;
2418 list_move_tail(&req->inflight_entry, &again);
2421 list_del(&req->inflight_entry);
2423 if (req->flags & REQ_F_BUFFER_SELECTED)
2424 cflags = io_put_rw_kbuf(req);
2426 __io_cqring_fill_event(req, req->result, cflags);
2429 if (refcount_dec_and_test(&req->refs))
2430 io_req_free_batch(&rb, req);
2433 io_commit_cqring(ctx);
2434 if (ctx->flags & IORING_SETUP_SQPOLL)
2435 io_cqring_ev_posted(ctx);
2436 io_req_free_batch_finish(ctx, &rb);
2438 if (!list_empty(&again))
2439 io_iopoll_queue(&again);
2442 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2445 struct io_kiocb *req, *tmp;
2451 * Only spin for completions if we don't have multiple devices hanging
2452 * off our complete list, and we're under the requested amount.
2454 spin = !ctx->poll_multi_file && *nr_events < min;
2457 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2458 struct kiocb *kiocb = &req->rw.kiocb;
2461 * Move completed and retryable entries to our local lists.
2462 * If we find a request that requires polling, break out
2463 * and complete those lists first, if we have entries there.
2465 if (READ_ONCE(req->iopoll_completed)) {
2466 list_move_tail(&req->inflight_entry, &done);
2469 if (!list_empty(&done))
2472 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2476 /* iopoll may have completed current req */
2477 if (READ_ONCE(req->iopoll_completed))
2478 list_move_tail(&req->inflight_entry, &done);
2485 if (!list_empty(&done))
2486 io_iopoll_complete(ctx, nr_events, &done);
2492 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2493 * non-spinning poll check - we'll still enter the driver poll loop, but only
2494 * as a non-spinning completion check.
2496 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
2499 while (!list_empty(&ctx->iopoll_list) && !need_resched()) {
2502 ret = io_do_iopoll(ctx, nr_events, min);
2505 if (*nr_events >= min)
2513 * We can't just wait for polled events to come to us, we have to actively
2514 * find and complete them.
2516 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2518 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2521 mutex_lock(&ctx->uring_lock);
2522 while (!list_empty(&ctx->iopoll_list)) {
2523 unsigned int nr_events = 0;
2525 io_do_iopoll(ctx, &nr_events, 0);
2527 /* let it sleep and repeat later if can't complete a request */
2531 * Ensure we allow local-to-the-cpu processing to take place,
2532 * in this case we need to ensure that we reap all events.
2533 * Also let task_work, etc. to progress by releasing the mutex
2535 if (need_resched()) {
2536 mutex_unlock(&ctx->uring_lock);
2538 mutex_lock(&ctx->uring_lock);
2541 mutex_unlock(&ctx->uring_lock);
2544 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2546 unsigned int nr_events = 0;
2547 int iters = 0, ret = 0;
2550 * We disallow the app entering submit/complete with polling, but we
2551 * still need to lock the ring to prevent racing with polled issue
2552 * that got punted to a workqueue.
2554 mutex_lock(&ctx->uring_lock);
2557 * Don't enter poll loop if we already have events pending.
2558 * If we do, we can potentially be spinning for commands that
2559 * already triggered a CQE (eg in error).
2561 if (io_cqring_events(ctx, false))
2565 * If a submit got punted to a workqueue, we can have the
2566 * application entering polling for a command before it gets
2567 * issued. That app will hold the uring_lock for the duration
2568 * of the poll right here, so we need to take a breather every
2569 * now and then to ensure that the issue has a chance to add
2570 * the poll to the issued list. Otherwise we can spin here
2571 * forever, while the workqueue is stuck trying to acquire the
2574 if (!(++iters & 7)) {
2575 mutex_unlock(&ctx->uring_lock);
2577 mutex_lock(&ctx->uring_lock);
2580 ret = io_iopoll_getevents(ctx, &nr_events, min);
2584 } while (min && !nr_events && !need_resched());
2586 mutex_unlock(&ctx->uring_lock);
2590 static void kiocb_end_write(struct io_kiocb *req)
2593 * Tell lockdep we inherited freeze protection from submission
2596 if (req->flags & REQ_F_ISREG) {
2597 struct inode *inode = file_inode(req->file);
2599 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
2601 file_end_write(req->file);
2604 static void io_complete_rw_common(struct kiocb *kiocb, long res,
2605 struct io_comp_state *cs)
2607 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2610 if (kiocb->ki_flags & IOCB_WRITE)
2611 kiocb_end_write(req);
2613 if (res != req->result)
2614 req_set_fail_links(req);
2615 if (req->flags & REQ_F_BUFFER_SELECTED)
2616 cflags = io_put_rw_kbuf(req);
2617 __io_req_complete(req, res, cflags, cs);
2621 static bool io_resubmit_prep(struct io_kiocb *req, int error)
2623 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2624 ssize_t ret = -ECANCELED;
2625 struct iov_iter iter;
2633 switch (req->opcode) {
2634 case IORING_OP_READV:
2635 case IORING_OP_READ_FIXED:
2636 case IORING_OP_READ:
2639 case IORING_OP_WRITEV:
2640 case IORING_OP_WRITE_FIXED:
2641 case IORING_OP_WRITE:
2645 printk_once(KERN_WARNING "io_uring: bad opcode in resubmit %d\n",
2650 if (!req->async_data) {
2651 ret = io_import_iovec(rw, req, &iovec, &iter, false);
2654 ret = io_setup_async_rw(req, iovec, inline_vecs, &iter, false);
2662 req_set_fail_links(req);
2667 static bool io_rw_reissue(struct io_kiocb *req, long res)
2670 umode_t mode = file_inode(req->file)->i_mode;
2673 if (!S_ISBLK(mode) && !S_ISREG(mode))
2675 if ((res != -EAGAIN && res != -EOPNOTSUPP) || io_wq_current_is_worker())
2678 ret = io_sq_thread_acquire_mm_files(req->ctx, req);
2680 if (io_resubmit_prep(req, ret)) {
2681 refcount_inc(&req->refs);
2682 io_queue_async_work(req);
2690 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2691 struct io_comp_state *cs)
2693 if (!io_rw_reissue(req, res))
2694 io_complete_rw_common(&req->rw.kiocb, res, cs);
2697 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2699 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2701 __io_complete_rw(req, res, res2, NULL);
2704 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2706 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2708 if (kiocb->ki_flags & IOCB_WRITE)
2709 kiocb_end_write(req);
2711 if (res != -EAGAIN && res != req->result)
2712 req_set_fail_links(req);
2714 WRITE_ONCE(req->result, res);
2715 /* order with io_poll_complete() checking ->result */
2717 WRITE_ONCE(req->iopoll_completed, 1);
2721 * After the iocb has been issued, it's safe to be found on the poll list.
2722 * Adding the kiocb to the list AFTER submission ensures that we don't
2723 * find it from a io_iopoll_getevents() thread before the issuer is done
2724 * accessing the kiocb cookie.
2726 static void io_iopoll_req_issued(struct io_kiocb *req, bool in_async)
2728 struct io_ring_ctx *ctx = req->ctx;
2731 * Track whether we have multiple files in our lists. This will impact
2732 * how we do polling eventually, not spinning if we're on potentially
2733 * different devices.
2735 if (list_empty(&ctx->iopoll_list)) {
2736 ctx->poll_multi_file = false;
2737 } else if (!ctx->poll_multi_file) {
2738 struct io_kiocb *list_req;
2740 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2742 if (list_req->file != req->file)
2743 ctx->poll_multi_file = true;
2747 * For fast devices, IO may have already completed. If it has, add
2748 * it to the front so we find it first.
2750 if (READ_ONCE(req->iopoll_completed))
2751 list_add(&req->inflight_entry, &ctx->iopoll_list);
2753 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2756 * If IORING_SETUP_SQPOLL is enabled, sqes are either handled in sq thread
2757 * task context or in io worker task context. If current task context is
2758 * sq thread, we don't need to check whether should wake up sq thread.
2760 if (in_async && (ctx->flags & IORING_SETUP_SQPOLL) &&
2761 wq_has_sleeper(&ctx->sq_data->wait))
2762 wake_up(&ctx->sq_data->wait);
2765 static inline void __io_state_file_put(struct io_submit_state *state)
2767 fput_many(state->file, state->file_refs);
2768 state->file_refs = 0;
2771 static inline void io_state_file_put(struct io_submit_state *state)
2773 if (state->file_refs)
2774 __io_state_file_put(state);
2778 * Get as many references to a file as we have IOs left in this submission,
2779 * assuming most submissions are for one file, or at least that each file
2780 * has more than one submission.
2782 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2787 if (state->file_refs) {
2788 if (state->fd == fd) {
2792 __io_state_file_put(state);
2794 state->file = fget_many(fd, state->ios_left);
2795 if (unlikely(!state->file))
2799 state->file_refs = state->ios_left - 1;
2803 static bool io_bdev_nowait(struct block_device *bdev)
2805 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2809 * If we tracked the file through the SCM inflight mechanism, we could support
2810 * any file. For now, just ensure that anything potentially problematic is done
2813 static bool io_file_supports_async(struct file *file, int rw)
2815 umode_t mode = file_inode(file)->i_mode;
2817 if (S_ISBLK(mode)) {
2818 if (IS_ENABLED(CONFIG_BLOCK) &&
2819 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2823 if (S_ISCHR(mode) || S_ISSOCK(mode))
2825 if (S_ISREG(mode)) {
2826 if (IS_ENABLED(CONFIG_BLOCK) &&
2827 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2828 file->f_op != &io_uring_fops)
2833 /* any ->read/write should understand O_NONBLOCK */
2834 if (file->f_flags & O_NONBLOCK)
2837 if (!(file->f_mode & FMODE_NOWAIT))
2841 return file->f_op->read_iter != NULL;
2843 return file->f_op->write_iter != NULL;
2846 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2848 struct io_ring_ctx *ctx = req->ctx;
2849 struct kiocb *kiocb = &req->rw.kiocb;
2853 if (S_ISREG(file_inode(req->file)->i_mode))
2854 req->flags |= REQ_F_ISREG;
2856 kiocb->ki_pos = READ_ONCE(sqe->off);
2857 if (kiocb->ki_pos == -1 && !(req->file->f_mode & FMODE_STREAM)) {
2858 req->flags |= REQ_F_CUR_POS;
2859 kiocb->ki_pos = req->file->f_pos;
2861 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2862 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2863 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2867 ioprio = READ_ONCE(sqe->ioprio);
2869 ret = ioprio_check_cap(ioprio);
2873 kiocb->ki_ioprio = ioprio;
2875 kiocb->ki_ioprio = get_current_ioprio();
2877 /* don't allow async punt if RWF_NOWAIT was requested */
2878 if (kiocb->ki_flags & IOCB_NOWAIT)
2879 req->flags |= REQ_F_NOWAIT;
2881 if (ctx->flags & IORING_SETUP_IOPOLL) {
2882 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2883 !kiocb->ki_filp->f_op->iopoll)
2886 kiocb->ki_flags |= IOCB_HIPRI;
2887 kiocb->ki_complete = io_complete_rw_iopoll;
2888 req->iopoll_completed = 0;
2890 if (kiocb->ki_flags & IOCB_HIPRI)
2892 kiocb->ki_complete = io_complete_rw;
2895 req->rw.addr = READ_ONCE(sqe->addr);
2896 req->rw.len = READ_ONCE(sqe->len);
2897 req->buf_index = READ_ONCE(sqe->buf_index);
2901 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2907 case -ERESTARTNOINTR:
2908 case -ERESTARTNOHAND:
2909 case -ERESTART_RESTARTBLOCK:
2911 * We can't just restart the syscall, since previously
2912 * submitted sqes may already be in progress. Just fail this
2918 kiocb->ki_complete(kiocb, ret, 0);
2922 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2923 struct io_comp_state *cs)
2925 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2926 struct io_async_rw *io = req->async_data;
2928 /* add previously done IO, if any */
2929 if (io && io->bytes_done > 0) {
2931 ret = io->bytes_done;
2933 ret += io->bytes_done;
2936 if (req->flags & REQ_F_CUR_POS)
2937 req->file->f_pos = kiocb->ki_pos;
2938 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2939 __io_complete_rw(req, ret, 0, cs);
2941 io_rw_done(kiocb, ret);
2944 static ssize_t io_import_fixed(struct io_kiocb *req, int rw,
2945 struct iov_iter *iter)
2947 struct io_ring_ctx *ctx = req->ctx;
2948 size_t len = req->rw.len;
2949 struct io_mapped_ubuf *imu;
2950 u16 index, buf_index = req->buf_index;
2954 if (unlikely(buf_index >= ctx->nr_user_bufs))
2956 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2957 imu = &ctx->user_bufs[index];
2958 buf_addr = req->rw.addr;
2961 if (buf_addr + len < buf_addr)
2963 /* not inside the mapped region */
2964 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2968 * May not be a start of buffer, set size appropriately
2969 * and advance us to the beginning.
2971 offset = buf_addr - imu->ubuf;
2972 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2976 * Don't use iov_iter_advance() here, as it's really slow for
2977 * using the latter parts of a big fixed buffer - it iterates
2978 * over each segment manually. We can cheat a bit here, because
2981 * 1) it's a BVEC iter, we set it up
2982 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2983 * first and last bvec
2985 * So just find our index, and adjust the iterator afterwards.
2986 * If the offset is within the first bvec (or the whole first
2987 * bvec, just use iov_iter_advance(). This makes it easier
2988 * since we can just skip the first segment, which may not
2989 * be PAGE_SIZE aligned.
2991 const struct bio_vec *bvec = imu->bvec;
2993 if (offset <= bvec->bv_len) {
2994 iov_iter_advance(iter, offset);
2996 unsigned long seg_skip;
2998 /* skip first vec */
2999 offset -= bvec->bv_len;
3000 seg_skip = 1 + (offset >> PAGE_SHIFT);
3002 iter->bvec = bvec + seg_skip;
3003 iter->nr_segs -= seg_skip;
3004 iter->count -= bvec->bv_len + offset;
3005 iter->iov_offset = offset & ~PAGE_MASK;
3012 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
3015 mutex_unlock(&ctx->uring_lock);
3018 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
3021 * "Normal" inline submissions always hold the uring_lock, since we
3022 * grab it from the system call. Same is true for the SQPOLL offload.
3023 * The only exception is when we've detached the request and issue it
3024 * from an async worker thread, grab the lock for that case.
3027 mutex_lock(&ctx->uring_lock);
3030 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
3031 int bgid, struct io_buffer *kbuf,
3034 struct io_buffer *head;
3036 if (req->flags & REQ_F_BUFFER_SELECTED)
3039 io_ring_submit_lock(req->ctx, needs_lock);
3041 lockdep_assert_held(&req->ctx->uring_lock);
3043 head = idr_find(&req->ctx->io_buffer_idr, bgid);
3045 if (!list_empty(&head->list)) {
3046 kbuf = list_last_entry(&head->list, struct io_buffer,
3048 list_del(&kbuf->list);
3051 idr_remove(&req->ctx->io_buffer_idr, bgid);
3053 if (*len > kbuf->len)
3056 kbuf = ERR_PTR(-ENOBUFS);
3059 io_ring_submit_unlock(req->ctx, needs_lock);
3064 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
3067 struct io_buffer *kbuf;
3070 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3071 bgid = req->buf_index;
3072 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
3075 req->rw.addr = (u64) (unsigned long) kbuf;
3076 req->flags |= REQ_F_BUFFER_SELECTED;
3077 return u64_to_user_ptr(kbuf->addr);
3080 #ifdef CONFIG_COMPAT
3081 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
3084 struct compat_iovec __user *uiov;
3085 compat_ssize_t clen;
3089 uiov = u64_to_user_ptr(req->rw.addr);
3090 if (!access_ok(uiov, sizeof(*uiov)))
3092 if (__get_user(clen, &uiov->iov_len))
3098 buf = io_rw_buffer_select(req, &len, needs_lock);
3100 return PTR_ERR(buf);
3101 iov[0].iov_base = buf;
3102 iov[0].iov_len = (compat_size_t) len;
3107 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3110 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
3114 if (copy_from_user(iov, uiov, sizeof(*uiov)))
3117 len = iov[0].iov_len;
3120 buf = io_rw_buffer_select(req, &len, needs_lock);
3122 return PTR_ERR(buf);
3123 iov[0].iov_base = buf;
3124 iov[0].iov_len = len;
3128 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3131 if (req->flags & REQ_F_BUFFER_SELECTED) {
3132 struct io_buffer *kbuf;
3134 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3135 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
3136 iov[0].iov_len = kbuf->len;
3141 else if (req->rw.len > 1)
3144 #ifdef CONFIG_COMPAT
3145 if (req->ctx->compat)
3146 return io_compat_import(req, iov, needs_lock);
3149 return __io_iov_buffer_select(req, iov, needs_lock);
3152 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
3153 struct iovec **iovec, struct iov_iter *iter,
3156 void __user *buf = u64_to_user_ptr(req->rw.addr);
3157 size_t sqe_len = req->rw.len;
3161 opcode = req->opcode;
3162 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3164 return io_import_fixed(req, rw, iter);
3167 /* buffer index only valid with fixed read/write, or buffer select */
3168 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3171 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3172 if (req->flags & REQ_F_BUFFER_SELECT) {
3173 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3175 return PTR_ERR(buf);
3176 req->rw.len = sqe_len;
3179 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3184 if (req->flags & REQ_F_BUFFER_SELECT) {
3185 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3187 ret = (*iovec)->iov_len;
3188 iov_iter_init(iter, rw, *iovec, 1, ret);
3194 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3198 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3200 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3204 * For files that don't have ->read_iter() and ->write_iter(), handle them
3205 * by looping over ->read() or ->write() manually.
3207 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3209 struct kiocb *kiocb = &req->rw.kiocb;
3210 struct file *file = req->file;
3214 * Don't support polled IO through this interface, and we can't
3215 * support non-blocking either. For the latter, this just causes
3216 * the kiocb to be handled from an async context.
3218 if (kiocb->ki_flags & IOCB_HIPRI)
3220 if (kiocb->ki_flags & IOCB_NOWAIT)
3223 while (iov_iter_count(iter)) {
3227 if (!iov_iter_is_bvec(iter)) {
3228 iovec = iov_iter_iovec(iter);
3230 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3231 iovec.iov_len = req->rw.len;
3235 nr = file->f_op->read(file, iovec.iov_base,
3236 iovec.iov_len, io_kiocb_ppos(kiocb));
3238 nr = file->f_op->write(file, iovec.iov_base,
3239 iovec.iov_len, io_kiocb_ppos(kiocb));
3248 if (nr != iovec.iov_len)
3252 iov_iter_advance(iter, nr);
3258 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3259 const struct iovec *fast_iov, struct iov_iter *iter)
3261 struct io_async_rw *rw = req->async_data;
3263 memcpy(&rw->iter, iter, sizeof(*iter));
3264 rw->free_iovec = iovec;
3266 /* can only be fixed buffers, no need to do anything */
3267 if (iov_iter_is_bvec(iter))
3270 unsigned iov_off = 0;
3272 rw->iter.iov = rw->fast_iov;
3273 if (iter->iov != fast_iov) {
3274 iov_off = iter->iov - fast_iov;
3275 rw->iter.iov += iov_off;
3277 if (rw->fast_iov != fast_iov)
3278 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3279 sizeof(struct iovec) * iter->nr_segs);
3281 req->flags |= REQ_F_NEED_CLEANUP;
3285 static inline int __io_alloc_async_data(struct io_kiocb *req)
3287 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3288 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3289 return req->async_data == NULL;
3292 static int io_alloc_async_data(struct io_kiocb *req)
3294 if (!io_op_defs[req->opcode].needs_async_data)
3297 return __io_alloc_async_data(req);
3300 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3301 const struct iovec *fast_iov,
3302 struct iov_iter *iter, bool force)
3304 if (!force && !io_op_defs[req->opcode].needs_async_data)
3306 if (!req->async_data) {
3307 if (__io_alloc_async_data(req))
3310 io_req_map_rw(req, iovec, fast_iov, iter);
3315 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3317 struct io_async_rw *iorw = req->async_data;
3318 struct iovec *iov = iorw->fast_iov;
3321 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3322 if (unlikely(ret < 0))
3325 iorw->bytes_done = 0;
3326 iorw->free_iovec = iov;
3328 req->flags |= REQ_F_NEED_CLEANUP;
3332 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3336 ret = io_prep_rw(req, sqe);
3340 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3343 /* either don't need iovec imported or already have it */
3344 if (!req->async_data)
3346 return io_rw_prep_async(req, READ);
3350 * This is our waitqueue callback handler, registered through lock_page_async()
3351 * when we initially tried to do the IO with the iocb armed our waitqueue.
3352 * This gets called when the page is unlocked, and we generally expect that to
3353 * happen when the page IO is completed and the page is now uptodate. This will
3354 * queue a task_work based retry of the operation, attempting to copy the data
3355 * again. If the latter fails because the page was NOT uptodate, then we will
3356 * do a thread based blocking retry of the operation. That's the unexpected
3359 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3360 int sync, void *arg)
3362 struct wait_page_queue *wpq;
3363 struct io_kiocb *req = wait->private;
3364 struct wait_page_key *key = arg;
3367 wpq = container_of(wait, struct wait_page_queue, wait);
3369 if (!wake_page_match(wpq, key))
3372 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3373 list_del_init(&wait->entry);
3375 init_task_work(&req->task_work, io_req_task_submit);
3376 percpu_ref_get(&req->ctx->refs);
3378 /* submit ref gets dropped, acquire a new one */
3379 refcount_inc(&req->refs);
3380 ret = io_req_task_work_add(req);
3381 if (unlikely(ret)) {
3382 struct task_struct *tsk;
3384 /* queue just for cancelation */
3385 init_task_work(&req->task_work, io_req_task_cancel);
3386 tsk = io_wq_get_task(req->ctx->io_wq);
3387 task_work_add(tsk, &req->task_work, TWA_NONE);
3388 wake_up_process(tsk);
3394 * This controls whether a given IO request should be armed for async page
3395 * based retry. If we return false here, the request is handed to the async
3396 * worker threads for retry. If we're doing buffered reads on a regular file,
3397 * we prepare a private wait_page_queue entry and retry the operation. This
3398 * will either succeed because the page is now uptodate and unlocked, or it
3399 * will register a callback when the page is unlocked at IO completion. Through
3400 * that callback, io_uring uses task_work to setup a retry of the operation.
3401 * That retry will attempt the buffered read again. The retry will generally
3402 * succeed, or in rare cases where it fails, we then fall back to using the
3403 * async worker threads for a blocking retry.
3405 static bool io_rw_should_retry(struct io_kiocb *req)
3407 struct io_async_rw *rw = req->async_data;
3408 struct wait_page_queue *wait = &rw->wpq;
3409 struct kiocb *kiocb = &req->rw.kiocb;
3411 /* never retry for NOWAIT, we just complete with -EAGAIN */
3412 if (req->flags & REQ_F_NOWAIT)
3415 /* Only for buffered IO */
3416 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3420 * just use poll if we can, and don't attempt if the fs doesn't
3421 * support callback based unlocks
3423 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3426 wait->wait.func = io_async_buf_func;
3427 wait->wait.private = req;
3428 wait->wait.flags = 0;
3429 INIT_LIST_HEAD(&wait->wait.entry);
3430 kiocb->ki_flags |= IOCB_WAITQ;
3431 kiocb->ki_flags &= ~IOCB_NOWAIT;
3432 kiocb->ki_waitq = wait;
3436 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3438 if (req->file->f_op->read_iter)
3439 return call_read_iter(req->file, &req->rw.kiocb, iter);
3440 else if (req->file->f_op->read)
3441 return loop_rw_iter(READ, req, iter);
3446 static int io_read(struct io_kiocb *req, bool force_nonblock,
3447 struct io_comp_state *cs)
3449 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3450 struct kiocb *kiocb = &req->rw.kiocb;
3451 struct iov_iter __iter, *iter = &__iter;
3452 struct io_async_rw *rw = req->async_data;
3453 ssize_t io_size, ret, ret2;
3460 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3464 io_size = iov_iter_count(iter);
3465 req->result = io_size;
3468 /* Ensure we clear previously set non-block flag */
3469 if (!force_nonblock)
3470 kiocb->ki_flags &= ~IOCB_NOWAIT;
3472 kiocb->ki_flags |= IOCB_NOWAIT;
3475 /* If the file doesn't support async, just async punt */
3476 no_async = force_nonblock && !io_file_supports_async(req->file, READ);
3480 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size);
3484 ret = io_iter_do_read(req, iter);
3488 } else if (ret == -EIOCBQUEUED) {
3491 } else if (ret == -EAGAIN) {
3492 /* IOPOLL retry should happen for io-wq threads */
3493 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3495 /* no retry on NONBLOCK marked file */
3496 if (req->file->f_flags & O_NONBLOCK)
3498 /* some cases will consume bytes even on error returns */
3499 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3502 } else if (ret < 0) {
3503 /* make sure -ERESTARTSYS -> -EINTR is done */
3507 /* read it all, or we did blocking attempt. no retry. */
3508 if (!iov_iter_count(iter) || !force_nonblock ||
3509 (req->file->f_flags & O_NONBLOCK))
3514 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3521 rw = req->async_data;
3522 /* it's copied and will be cleaned with ->io */
3524 /* now use our persistent iterator, if we aren't already */
3527 rw->bytes_done += ret;
3528 /* if we can retry, do so with the callbacks armed */
3529 if (!io_rw_should_retry(req)) {
3530 kiocb->ki_flags &= ~IOCB_WAITQ;
3535 * Now retry read with the IOCB_WAITQ parts set in the iocb. If we
3536 * get -EIOCBQUEUED, then we'll get a notification when the desired
3537 * page gets unlocked. We can also get a partial read here, and if we
3538 * do, then just retry at the new offset.
3540 ret = io_iter_do_read(req, iter);
3541 if (ret == -EIOCBQUEUED) {
3544 } else if (ret > 0 && ret < io_size) {
3545 /* we got some bytes, but not all. retry. */
3549 kiocb_done(kiocb, ret, cs);
3552 /* it's reportedly faster than delegating the null check to kfree() */
3558 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3562 ret = io_prep_rw(req, sqe);
3566 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3569 /* either don't need iovec imported or already have it */
3570 if (!req->async_data)
3572 return io_rw_prep_async(req, WRITE);
3575 static int io_write(struct io_kiocb *req, bool force_nonblock,
3576 struct io_comp_state *cs)
3578 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3579 struct kiocb *kiocb = &req->rw.kiocb;
3580 struct iov_iter __iter, *iter = &__iter;
3581 struct io_async_rw *rw = req->async_data;
3582 ssize_t ret, ret2, io_size;
3588 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3592 io_size = iov_iter_count(iter);
3593 req->result = io_size;
3595 /* Ensure we clear previously set non-block flag */
3596 if (!force_nonblock)
3597 kiocb->ki_flags &= ~IOCB_NOWAIT;
3599 kiocb->ki_flags |= IOCB_NOWAIT;
3601 /* If the file doesn't support async, just async punt */
3602 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
3605 /* file path doesn't support NOWAIT for non-direct_IO */
3606 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3607 (req->flags & REQ_F_ISREG))
3610 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size);
3615 * Open-code file_start_write here to grab freeze protection,
3616 * which will be released by another thread in
3617 * io_complete_rw(). Fool lockdep by telling it the lock got
3618 * released so that it doesn't complain about the held lock when
3619 * we return to userspace.
3621 if (req->flags & REQ_F_ISREG) {
3622 sb_start_write(file_inode(req->file)->i_sb);
3623 __sb_writers_release(file_inode(req->file)->i_sb,
3626 kiocb->ki_flags |= IOCB_WRITE;
3628 if (req->file->f_op->write_iter)
3629 ret2 = call_write_iter(req->file, kiocb, iter);
3630 else if (req->file->f_op->write)
3631 ret2 = loop_rw_iter(WRITE, req, iter);
3636 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3637 * retry them without IOCB_NOWAIT.
3639 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3641 /* no retry on NONBLOCK marked file */
3642 if (ret2 == -EAGAIN && (req->file->f_flags & O_NONBLOCK))
3644 if (!force_nonblock || ret2 != -EAGAIN) {
3645 /* IOPOLL retry should happen for io-wq threads */
3646 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3649 kiocb_done(kiocb, ret2, cs);
3652 /* some cases will consume bytes even on error returns */
3653 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3654 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3659 /* it's reportedly faster than delegating the null check to kfree() */
3665 static int io_renameat_prep(struct io_kiocb *req,
3666 const struct io_uring_sqe *sqe)
3668 struct io_rename *ren = &req->rename;
3669 const char __user *oldf, *newf;
3671 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3674 ren->old_dfd = READ_ONCE(sqe->fd);
3675 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3676 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3677 ren->new_dfd = READ_ONCE(sqe->len);
3678 ren->flags = READ_ONCE(sqe->rename_flags);
3680 ren->oldpath = getname(oldf);
3681 if (IS_ERR(ren->oldpath))
3682 return PTR_ERR(ren->oldpath);
3684 ren->newpath = getname(newf);
3685 if (IS_ERR(ren->newpath)) {
3686 putname(ren->oldpath);
3687 return PTR_ERR(ren->newpath);
3690 req->flags |= REQ_F_NEED_CLEANUP;
3694 static int io_renameat(struct io_kiocb *req, bool force_nonblock)
3696 struct io_rename *ren = &req->rename;
3702 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3703 ren->newpath, ren->flags);
3705 req->flags &= ~REQ_F_NEED_CLEANUP;
3707 req_set_fail_links(req);
3708 io_req_complete(req, ret);
3712 static int io_unlinkat_prep(struct io_kiocb *req,
3713 const struct io_uring_sqe *sqe)
3715 struct io_unlink *un = &req->unlink;
3716 const char __user *fname;
3718 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3721 un->dfd = READ_ONCE(sqe->fd);
3723 un->flags = READ_ONCE(sqe->unlink_flags);
3724 if (un->flags & ~AT_REMOVEDIR)
3727 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3728 un->filename = getname(fname);
3729 if (IS_ERR(un->filename))
3730 return PTR_ERR(un->filename);
3732 req->flags |= REQ_F_NEED_CLEANUP;
3736 static int io_unlinkat(struct io_kiocb *req, bool force_nonblock)
3738 struct io_unlink *un = &req->unlink;
3744 if (un->flags & AT_REMOVEDIR)
3745 ret = do_rmdir(un->dfd, un->filename);
3747 ret = do_unlinkat(un->dfd, un->filename);
3749 req->flags &= ~REQ_F_NEED_CLEANUP;
3751 req_set_fail_links(req);
3752 io_req_complete(req, ret);
3756 static int io_shutdown_prep(struct io_kiocb *req,
3757 const struct io_uring_sqe *sqe)
3759 #if defined(CONFIG_NET)
3760 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3762 if (sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3766 req->shutdown.how = READ_ONCE(sqe->len);
3773 static int io_shutdown(struct io_kiocb *req, bool force_nonblock)
3775 #if defined(CONFIG_NET)
3776 struct socket *sock;
3782 sock = sock_from_file(req->file);
3783 if (unlikely(!sock))
3786 ret = __sys_shutdown_sock(sock, req->shutdown.how);
3787 io_req_complete(req, ret);
3794 static int __io_splice_prep(struct io_kiocb *req,
3795 const struct io_uring_sqe *sqe)
3797 struct io_splice* sp = &req->splice;
3798 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3800 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3804 sp->len = READ_ONCE(sqe->len);
3805 sp->flags = READ_ONCE(sqe->splice_flags);
3807 if (unlikely(sp->flags & ~valid_flags))
3810 sp->file_in = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in),
3811 (sp->flags & SPLICE_F_FD_IN_FIXED));
3814 req->flags |= REQ_F_NEED_CLEANUP;
3816 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
3818 * Splice operation will be punted aync, and here need to
3819 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3821 io_req_init_async(req);
3822 req->work.flags |= IO_WQ_WORK_UNBOUND;
3828 static int io_tee_prep(struct io_kiocb *req,
3829 const struct io_uring_sqe *sqe)
3831 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3833 return __io_splice_prep(req, sqe);
3836 static int io_tee(struct io_kiocb *req, bool force_nonblock)
3838 struct io_splice *sp = &req->splice;
3839 struct file *in = sp->file_in;
3840 struct file *out = sp->file_out;
3841 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3847 ret = do_tee(in, out, sp->len, flags);
3849 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3850 req->flags &= ~REQ_F_NEED_CLEANUP;
3853 req_set_fail_links(req);
3854 io_req_complete(req, ret);
3858 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3860 struct io_splice* sp = &req->splice;
3862 sp->off_in = READ_ONCE(sqe->splice_off_in);
3863 sp->off_out = READ_ONCE(sqe->off);
3864 return __io_splice_prep(req, sqe);
3867 static int io_splice(struct io_kiocb *req, bool force_nonblock)
3869 struct io_splice *sp = &req->splice;
3870 struct file *in = sp->file_in;
3871 struct file *out = sp->file_out;
3872 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3873 loff_t *poff_in, *poff_out;
3879 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3880 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3883 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3885 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3886 req->flags &= ~REQ_F_NEED_CLEANUP;
3889 req_set_fail_links(req);
3890 io_req_complete(req, ret);
3895 * IORING_OP_NOP just posts a completion event, nothing else.
3897 static int io_nop(struct io_kiocb *req, struct io_comp_state *cs)
3899 struct io_ring_ctx *ctx = req->ctx;
3901 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3904 __io_req_complete(req, 0, 0, cs);
3908 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3910 struct io_ring_ctx *ctx = req->ctx;
3915 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3917 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3920 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3921 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3924 req->sync.off = READ_ONCE(sqe->off);
3925 req->sync.len = READ_ONCE(sqe->len);
3929 static int io_fsync(struct io_kiocb *req, bool force_nonblock)
3931 loff_t end = req->sync.off + req->sync.len;
3934 /* fsync always requires a blocking context */
3938 ret = vfs_fsync_range(req->file, req->sync.off,
3939 end > 0 ? end : LLONG_MAX,
3940 req->sync.flags & IORING_FSYNC_DATASYNC);
3942 req_set_fail_links(req);
3943 io_req_complete(req, ret);
3947 static int io_fallocate_prep(struct io_kiocb *req,
3948 const struct io_uring_sqe *sqe)
3950 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3952 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3955 req->sync.off = READ_ONCE(sqe->off);
3956 req->sync.len = READ_ONCE(sqe->addr);
3957 req->sync.mode = READ_ONCE(sqe->len);
3961 static int io_fallocate(struct io_kiocb *req, bool force_nonblock)
3965 /* fallocate always requiring blocking context */
3968 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3971 req_set_fail_links(req);
3972 io_req_complete(req, ret);
3976 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3978 const char __user *fname;
3981 if (unlikely(sqe->ioprio || sqe->buf_index))
3983 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3986 /* open.how should be already initialised */
3987 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3988 req->open.how.flags |= O_LARGEFILE;
3990 req->open.dfd = READ_ONCE(sqe->fd);
3991 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3992 req->open.filename = getname(fname);
3993 if (IS_ERR(req->open.filename)) {
3994 ret = PTR_ERR(req->open.filename);
3995 req->open.filename = NULL;
3998 req->open.nofile = rlimit(RLIMIT_NOFILE);
3999 req->open.ignore_nonblock = false;
4000 req->flags |= REQ_F_NEED_CLEANUP;
4004 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4008 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4010 mode = READ_ONCE(sqe->len);
4011 flags = READ_ONCE(sqe->open_flags);
4012 req->open.how = build_open_how(flags, mode);
4013 return __io_openat_prep(req, sqe);
4016 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4018 struct open_how __user *how;
4022 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4024 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4025 len = READ_ONCE(sqe->len);
4026 if (len < OPEN_HOW_SIZE_VER0)
4029 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
4034 return __io_openat_prep(req, sqe);
4037 static int io_openat2(struct io_kiocb *req, bool force_nonblock)
4039 struct open_flags op;
4043 if (force_nonblock && !req->open.ignore_nonblock)
4046 ret = build_open_flags(&req->open.how, &op);
4050 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
4054 file = do_filp_open(req->open.dfd, req->open.filename, &op);
4057 ret = PTR_ERR(file);
4059 * A work-around to ensure that /proc/self works that way
4060 * that it should - if we get -EOPNOTSUPP back, then assume
4061 * that proc_self_get_link() failed us because we're in async
4062 * context. We should be safe to retry this from the task
4063 * itself with force_nonblock == false set, as it should not
4064 * block on lookup. Would be nice to know this upfront and
4065 * avoid the async dance, but doesn't seem feasible.
4067 if (ret == -EOPNOTSUPP && io_wq_current_is_worker()) {
4068 req->open.ignore_nonblock = true;
4069 refcount_inc(&req->refs);
4070 io_req_task_queue(req);
4074 fsnotify_open(file);
4075 fd_install(ret, file);
4078 putname(req->open.filename);
4079 req->flags &= ~REQ_F_NEED_CLEANUP;
4081 req_set_fail_links(req);
4082 io_req_complete(req, ret);
4086 static int io_openat(struct io_kiocb *req, bool force_nonblock)
4088 return io_openat2(req, force_nonblock);
4091 static int io_remove_buffers_prep(struct io_kiocb *req,
4092 const struct io_uring_sqe *sqe)
4094 struct io_provide_buf *p = &req->pbuf;
4097 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
4100 tmp = READ_ONCE(sqe->fd);
4101 if (!tmp || tmp > USHRT_MAX)
4104 memset(p, 0, sizeof(*p));
4106 p->bgid = READ_ONCE(sqe->buf_group);
4110 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
4111 int bgid, unsigned nbufs)
4115 /* shouldn't happen */
4119 /* the head kbuf is the list itself */
4120 while (!list_empty(&buf->list)) {
4121 struct io_buffer *nxt;
4123 nxt = list_first_entry(&buf->list, struct io_buffer, list);
4124 list_del(&nxt->list);
4131 idr_remove(&ctx->io_buffer_idr, bgid);
4136 static int io_remove_buffers(struct io_kiocb *req, bool force_nonblock,
4137 struct io_comp_state *cs)
4139 struct io_provide_buf *p = &req->pbuf;
4140 struct io_ring_ctx *ctx = req->ctx;
4141 struct io_buffer *head;
4144 io_ring_submit_lock(ctx, !force_nonblock);
4146 lockdep_assert_held(&ctx->uring_lock);
4149 head = idr_find(&ctx->io_buffer_idr, p->bgid);
4151 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
4153 req_set_fail_links(req);
4155 /* need to hold the lock to complete IOPOLL requests */
4156 if (ctx->flags & IORING_SETUP_IOPOLL) {
4157 __io_req_complete(req, ret, 0, cs);
4158 io_ring_submit_unlock(ctx, !force_nonblock);
4160 io_ring_submit_unlock(ctx, !force_nonblock);
4161 __io_req_complete(req, ret, 0, cs);
4166 static int io_provide_buffers_prep(struct io_kiocb *req,
4167 const struct io_uring_sqe *sqe)
4169 struct io_provide_buf *p = &req->pbuf;
4172 if (sqe->ioprio || sqe->rw_flags)
4175 tmp = READ_ONCE(sqe->fd);
4176 if (!tmp || tmp > USHRT_MAX)
4179 p->addr = READ_ONCE(sqe->addr);
4180 p->len = READ_ONCE(sqe->len);
4182 if (!access_ok(u64_to_user_ptr(p->addr), (p->len * p->nbufs)))
4185 p->bgid = READ_ONCE(sqe->buf_group);
4186 tmp = READ_ONCE(sqe->off);
4187 if (tmp > USHRT_MAX)
4193 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
4195 struct io_buffer *buf;
4196 u64 addr = pbuf->addr;
4197 int i, bid = pbuf->bid;
4199 for (i = 0; i < pbuf->nbufs; i++) {
4200 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
4205 buf->len = pbuf->len;
4210 INIT_LIST_HEAD(&buf->list);
4213 list_add_tail(&buf->list, &(*head)->list);
4217 return i ? i : -ENOMEM;
4220 static int io_provide_buffers(struct io_kiocb *req, bool force_nonblock,
4221 struct io_comp_state *cs)
4223 struct io_provide_buf *p = &req->pbuf;
4224 struct io_ring_ctx *ctx = req->ctx;
4225 struct io_buffer *head, *list;
4228 io_ring_submit_lock(ctx, !force_nonblock);
4230 lockdep_assert_held(&ctx->uring_lock);
4232 list = head = idr_find(&ctx->io_buffer_idr, p->bgid);
4234 ret = io_add_buffers(p, &head);
4239 ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1,
4242 __io_remove_buffers(ctx, head, p->bgid, -1U);
4248 req_set_fail_links(req);
4250 /* need to hold the lock to complete IOPOLL requests */
4251 if (ctx->flags & IORING_SETUP_IOPOLL) {
4252 __io_req_complete(req, ret, 0, cs);
4253 io_ring_submit_unlock(ctx, !force_nonblock);
4255 io_ring_submit_unlock(ctx, !force_nonblock);
4256 __io_req_complete(req, ret, 0, cs);
4261 static int io_epoll_ctl_prep(struct io_kiocb *req,
4262 const struct io_uring_sqe *sqe)
4264 #if defined(CONFIG_EPOLL)
4265 if (sqe->ioprio || sqe->buf_index)
4267 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4270 req->epoll.epfd = READ_ONCE(sqe->fd);
4271 req->epoll.op = READ_ONCE(sqe->len);
4272 req->epoll.fd = READ_ONCE(sqe->off);
4274 if (ep_op_has_event(req->epoll.op)) {
4275 struct epoll_event __user *ev;
4277 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4278 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4288 static int io_epoll_ctl(struct io_kiocb *req, bool force_nonblock,
4289 struct io_comp_state *cs)
4291 #if defined(CONFIG_EPOLL)
4292 struct io_epoll *ie = &req->epoll;
4295 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4296 if (force_nonblock && ret == -EAGAIN)
4300 req_set_fail_links(req);
4301 __io_req_complete(req, ret, 0, cs);
4308 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4310 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4311 if (sqe->ioprio || sqe->buf_index || sqe->off)
4313 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4316 req->madvise.addr = READ_ONCE(sqe->addr);
4317 req->madvise.len = READ_ONCE(sqe->len);
4318 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4325 static int io_madvise(struct io_kiocb *req, bool force_nonblock)
4327 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4328 struct io_madvise *ma = &req->madvise;
4334 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4336 req_set_fail_links(req);
4337 io_req_complete(req, ret);
4344 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4346 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4348 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4351 req->fadvise.offset = READ_ONCE(sqe->off);
4352 req->fadvise.len = READ_ONCE(sqe->len);
4353 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4357 static int io_fadvise(struct io_kiocb *req, bool force_nonblock)
4359 struct io_fadvise *fa = &req->fadvise;
4362 if (force_nonblock) {
4363 switch (fa->advice) {
4364 case POSIX_FADV_NORMAL:
4365 case POSIX_FADV_RANDOM:
4366 case POSIX_FADV_SEQUENTIAL:
4373 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4375 req_set_fail_links(req);
4376 io_req_complete(req, ret);
4380 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4382 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4384 if (sqe->ioprio || sqe->buf_index)
4386 if (req->flags & REQ_F_FIXED_FILE)
4389 req->statx.dfd = READ_ONCE(sqe->fd);
4390 req->statx.mask = READ_ONCE(sqe->len);
4391 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4392 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4393 req->statx.flags = READ_ONCE(sqe->statx_flags);
4398 static int io_statx(struct io_kiocb *req, bool force_nonblock)
4400 struct io_statx *ctx = &req->statx;
4403 if (force_nonblock) {
4404 /* only need file table for an actual valid fd */
4405 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
4406 req->flags |= REQ_F_NO_FILE_TABLE;
4410 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4414 req_set_fail_links(req);
4415 io_req_complete(req, ret);
4419 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4422 * If we queue this for async, it must not be cancellable. That would
4423 * leave the 'file' in an undeterminate state, and here need to modify
4424 * io_wq_work.flags, so initialize io_wq_work firstly.
4426 io_req_init_async(req);
4427 req->work.flags |= IO_WQ_WORK_NO_CANCEL;
4429 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4431 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4432 sqe->rw_flags || sqe->buf_index)
4434 if (req->flags & REQ_F_FIXED_FILE)
4437 req->close.fd = READ_ONCE(sqe->fd);
4438 if ((req->file && req->file->f_op == &io_uring_fops))
4441 req->close.put_file = NULL;
4445 static int io_close(struct io_kiocb *req, bool force_nonblock,
4446 struct io_comp_state *cs)
4448 struct io_close *close = &req->close;
4451 /* might be already done during nonblock submission */
4452 if (!close->put_file) {
4453 ret = close_fd_get_file(close->fd, &close->put_file);
4455 return (ret == -ENOENT) ? -EBADF : ret;
4458 /* if the file has a flush method, be safe and punt to async */
4459 if (close->put_file->f_op->flush && force_nonblock) {
4460 /* was never set, but play safe */
4461 req->flags &= ~REQ_F_NOWAIT;
4462 /* avoid grabbing files - we don't need the files */
4463 req->flags |= REQ_F_NO_FILE_TABLE;
4467 /* No ->flush() or already async, safely close from here */
4468 ret = filp_close(close->put_file, req->work.identity->files);
4470 req_set_fail_links(req);
4471 fput(close->put_file);
4472 close->put_file = NULL;
4473 __io_req_complete(req, ret, 0, cs);
4477 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4479 struct io_ring_ctx *ctx = req->ctx;
4484 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4486 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4489 req->sync.off = READ_ONCE(sqe->off);
4490 req->sync.len = READ_ONCE(sqe->len);
4491 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4495 static int io_sync_file_range(struct io_kiocb *req, bool force_nonblock)
4499 /* sync_file_range always requires a blocking context */
4503 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4506 req_set_fail_links(req);
4507 io_req_complete(req, ret);
4511 #if defined(CONFIG_NET)
4512 static int io_setup_async_msg(struct io_kiocb *req,
4513 struct io_async_msghdr *kmsg)
4515 struct io_async_msghdr *async_msg = req->async_data;
4519 if (io_alloc_async_data(req)) {
4520 if (kmsg->iov != kmsg->fast_iov)
4524 async_msg = req->async_data;
4525 req->flags |= REQ_F_NEED_CLEANUP;
4526 memcpy(async_msg, kmsg, sizeof(*kmsg));
4530 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4531 struct io_async_msghdr *iomsg)
4533 iomsg->iov = iomsg->fast_iov;
4534 iomsg->msg.msg_name = &iomsg->addr;
4535 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4536 req->sr_msg.msg_flags, &iomsg->iov);
4539 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4541 struct io_async_msghdr *async_msg = req->async_data;
4542 struct io_sr_msg *sr = &req->sr_msg;
4545 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4548 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4549 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4550 sr->len = READ_ONCE(sqe->len);
4552 #ifdef CONFIG_COMPAT
4553 if (req->ctx->compat)
4554 sr->msg_flags |= MSG_CMSG_COMPAT;
4557 if (!async_msg || !io_op_defs[req->opcode].needs_async_data)
4559 ret = io_sendmsg_copy_hdr(req, async_msg);
4561 req->flags |= REQ_F_NEED_CLEANUP;
4565 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4566 struct io_comp_state *cs)
4568 struct io_async_msghdr iomsg, *kmsg;
4569 struct socket *sock;
4573 sock = sock_from_file(req->file);
4574 if (unlikely(!sock))
4577 if (req->async_data) {
4578 kmsg = req->async_data;
4579 kmsg->msg.msg_name = &kmsg->addr;
4580 /* if iov is set, it's allocated already */
4582 kmsg->iov = kmsg->fast_iov;
4583 kmsg->msg.msg_iter.iov = kmsg->iov;
4585 ret = io_sendmsg_copy_hdr(req, &iomsg);
4591 flags = req->sr_msg.msg_flags;
4592 if (flags & MSG_DONTWAIT)
4593 req->flags |= REQ_F_NOWAIT;
4594 else if (force_nonblock)
4595 flags |= MSG_DONTWAIT;
4597 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4598 if (force_nonblock && ret == -EAGAIN)
4599 return io_setup_async_msg(req, kmsg);
4600 if (ret == -ERESTARTSYS)
4603 if (kmsg->iov != kmsg->fast_iov)
4605 req->flags &= ~REQ_F_NEED_CLEANUP;
4607 req_set_fail_links(req);
4608 __io_req_complete(req, ret, 0, cs);
4612 static int io_send(struct io_kiocb *req, bool force_nonblock,
4613 struct io_comp_state *cs)
4615 struct io_sr_msg *sr = &req->sr_msg;
4618 struct socket *sock;
4622 sock = sock_from_file(req->file);
4623 if (unlikely(!sock))
4626 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4630 msg.msg_name = NULL;
4631 msg.msg_control = NULL;
4632 msg.msg_controllen = 0;
4633 msg.msg_namelen = 0;
4635 flags = req->sr_msg.msg_flags;
4636 if (flags & MSG_DONTWAIT)
4637 req->flags |= REQ_F_NOWAIT;
4638 else if (force_nonblock)
4639 flags |= MSG_DONTWAIT;
4641 msg.msg_flags = flags;
4642 ret = sock_sendmsg(sock, &msg);
4643 if (force_nonblock && ret == -EAGAIN)
4645 if (ret == -ERESTARTSYS)
4649 req_set_fail_links(req);
4650 __io_req_complete(req, ret, 0, cs);
4654 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4655 struct io_async_msghdr *iomsg)
4657 struct io_sr_msg *sr = &req->sr_msg;
4658 struct iovec __user *uiov;
4662 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4663 &iomsg->uaddr, &uiov, &iov_len);
4667 if (req->flags & REQ_F_BUFFER_SELECT) {
4670 if (copy_from_user(iomsg->iov, uiov, sizeof(*uiov)))
4672 sr->len = iomsg->iov[0].iov_len;
4673 iov_iter_init(&iomsg->msg.msg_iter, READ, iomsg->iov, 1,
4677 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4678 &iomsg->iov, &iomsg->msg.msg_iter,
4687 #ifdef CONFIG_COMPAT
4688 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4689 struct io_async_msghdr *iomsg)
4691 struct compat_msghdr __user *msg_compat;
4692 struct io_sr_msg *sr = &req->sr_msg;
4693 struct compat_iovec __user *uiov;
4698 msg_compat = (struct compat_msghdr __user *) sr->umsg;
4699 ret = __get_compat_msghdr(&iomsg->msg, msg_compat, &iomsg->uaddr,
4704 uiov = compat_ptr(ptr);
4705 if (req->flags & REQ_F_BUFFER_SELECT) {
4706 compat_ssize_t clen;
4710 if (!access_ok(uiov, sizeof(*uiov)))
4712 if (__get_user(clen, &uiov->iov_len))
4717 iomsg->iov[0].iov_len = clen;
4720 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4721 UIO_FASTIOV, &iomsg->iov,
4722 &iomsg->msg.msg_iter, true);
4731 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4732 struct io_async_msghdr *iomsg)
4734 iomsg->msg.msg_name = &iomsg->addr;
4735 iomsg->iov = iomsg->fast_iov;
4737 #ifdef CONFIG_COMPAT
4738 if (req->ctx->compat)
4739 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4742 return __io_recvmsg_copy_hdr(req, iomsg);
4745 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4748 struct io_sr_msg *sr = &req->sr_msg;
4749 struct io_buffer *kbuf;
4751 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4756 req->flags |= REQ_F_BUFFER_SELECTED;
4760 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4762 return io_put_kbuf(req, req->sr_msg.kbuf);
4765 static int io_recvmsg_prep(struct io_kiocb *req,
4766 const struct io_uring_sqe *sqe)
4768 struct io_async_msghdr *async_msg = req->async_data;
4769 struct io_sr_msg *sr = &req->sr_msg;
4772 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4775 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4776 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4777 sr->len = READ_ONCE(sqe->len);
4778 sr->bgid = READ_ONCE(sqe->buf_group);
4780 #ifdef CONFIG_COMPAT
4781 if (req->ctx->compat)
4782 sr->msg_flags |= MSG_CMSG_COMPAT;
4785 if (!async_msg || !io_op_defs[req->opcode].needs_async_data)
4787 ret = io_recvmsg_copy_hdr(req, async_msg);
4789 req->flags |= REQ_F_NEED_CLEANUP;
4793 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4794 struct io_comp_state *cs)
4796 struct io_async_msghdr iomsg, *kmsg;
4797 struct socket *sock;
4798 struct io_buffer *kbuf;
4800 int ret, cflags = 0;
4802 sock = sock_from_file(req->file);
4803 if (unlikely(!sock))
4806 if (req->async_data) {
4807 kmsg = req->async_data;
4808 kmsg->msg.msg_name = &kmsg->addr;
4809 /* if iov is set, it's allocated already */
4811 kmsg->iov = kmsg->fast_iov;
4812 kmsg->msg.msg_iter.iov = kmsg->iov;
4814 ret = io_recvmsg_copy_hdr(req, &iomsg);
4820 if (req->flags & REQ_F_BUFFER_SELECT) {
4821 kbuf = io_recv_buffer_select(req, !force_nonblock);
4823 return PTR_ERR(kbuf);
4824 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4825 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->iov,
4826 1, req->sr_msg.len);
4829 flags = req->sr_msg.msg_flags;
4830 if (flags & MSG_DONTWAIT)
4831 req->flags |= REQ_F_NOWAIT;
4832 else if (force_nonblock)
4833 flags |= MSG_DONTWAIT;
4835 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4836 kmsg->uaddr, flags);
4837 if (force_nonblock && ret == -EAGAIN)
4838 return io_setup_async_msg(req, kmsg);
4839 if (ret == -ERESTARTSYS)
4842 if (req->flags & REQ_F_BUFFER_SELECTED)
4843 cflags = io_put_recv_kbuf(req);
4844 if (kmsg->iov != kmsg->fast_iov)
4846 req->flags &= ~REQ_F_NEED_CLEANUP;
4848 req_set_fail_links(req);
4849 __io_req_complete(req, ret, cflags, cs);
4853 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4854 struct io_comp_state *cs)
4856 struct io_buffer *kbuf;
4857 struct io_sr_msg *sr = &req->sr_msg;
4859 void __user *buf = sr->buf;
4860 struct socket *sock;
4863 int ret, cflags = 0;
4865 sock = sock_from_file(req->file);
4866 if (unlikely(!sock))
4869 if (req->flags & REQ_F_BUFFER_SELECT) {
4870 kbuf = io_recv_buffer_select(req, !force_nonblock);
4872 return PTR_ERR(kbuf);
4873 buf = u64_to_user_ptr(kbuf->addr);
4876 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4880 msg.msg_name = NULL;
4881 msg.msg_control = NULL;
4882 msg.msg_controllen = 0;
4883 msg.msg_namelen = 0;
4884 msg.msg_iocb = NULL;
4887 flags = req->sr_msg.msg_flags;
4888 if (flags & MSG_DONTWAIT)
4889 req->flags |= REQ_F_NOWAIT;
4890 else if (force_nonblock)
4891 flags |= MSG_DONTWAIT;
4893 ret = sock_recvmsg(sock, &msg, flags);
4894 if (force_nonblock && ret == -EAGAIN)
4896 if (ret == -ERESTARTSYS)
4899 if (req->flags & REQ_F_BUFFER_SELECTED)
4900 cflags = io_put_recv_kbuf(req);
4902 req_set_fail_links(req);
4903 __io_req_complete(req, ret, cflags, cs);
4907 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4909 struct io_accept *accept = &req->accept;
4911 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4913 if (sqe->ioprio || sqe->len || sqe->buf_index)
4916 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4917 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4918 accept->flags = READ_ONCE(sqe->accept_flags);
4919 accept->nofile = rlimit(RLIMIT_NOFILE);
4923 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4924 struct io_comp_state *cs)
4926 struct io_accept *accept = &req->accept;
4927 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4930 if (req->file->f_flags & O_NONBLOCK)
4931 req->flags |= REQ_F_NOWAIT;
4933 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4934 accept->addr_len, accept->flags,
4936 if (ret == -EAGAIN && force_nonblock)
4939 if (ret == -ERESTARTSYS)
4941 req_set_fail_links(req);
4943 __io_req_complete(req, ret, 0, cs);
4947 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4949 struct io_connect *conn = &req->connect;
4950 struct io_async_connect *io = req->async_data;
4952 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4954 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4957 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4958 conn->addr_len = READ_ONCE(sqe->addr2);
4963 return move_addr_to_kernel(conn->addr, conn->addr_len,
4967 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4968 struct io_comp_state *cs)
4970 struct io_async_connect __io, *io;
4971 unsigned file_flags;
4974 if (req->async_data) {
4975 io = req->async_data;
4977 ret = move_addr_to_kernel(req->connect.addr,
4978 req->connect.addr_len,
4985 file_flags = force_nonblock ? O_NONBLOCK : 0;
4987 ret = __sys_connect_file(req->file, &io->address,
4988 req->connect.addr_len, file_flags);
4989 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4990 if (req->async_data)
4992 if (io_alloc_async_data(req)) {
4996 io = req->async_data;
4997 memcpy(req->async_data, &__io, sizeof(__io));
5000 if (ret == -ERESTARTSYS)
5004 req_set_fail_links(req);
5005 __io_req_complete(req, ret, 0, cs);
5008 #else /* !CONFIG_NET */
5009 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5014 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
5015 struct io_comp_state *cs)
5020 static int io_send(struct io_kiocb *req, bool force_nonblock,
5021 struct io_comp_state *cs)
5026 static int io_recvmsg_prep(struct io_kiocb *req,
5027 const struct io_uring_sqe *sqe)
5032 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
5033 struct io_comp_state *cs)
5038 static int io_recv(struct io_kiocb *req, bool force_nonblock,
5039 struct io_comp_state *cs)
5044 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5049 static int io_accept(struct io_kiocb *req, bool force_nonblock,
5050 struct io_comp_state *cs)
5055 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5060 static int io_connect(struct io_kiocb *req, bool force_nonblock,
5061 struct io_comp_state *cs)
5065 #endif /* CONFIG_NET */
5067 struct io_poll_table {
5068 struct poll_table_struct pt;
5069 struct io_kiocb *req;
5073 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
5074 __poll_t mask, task_work_func_t func)
5078 /* for instances that support it check for an event match first: */
5079 if (mask && !(mask & poll->events))
5082 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
5084 list_del_init(&poll->wait.entry);
5087 init_task_work(&req->task_work, func);
5088 percpu_ref_get(&req->ctx->refs);
5091 * If this fails, then the task is exiting. When a task exits, the
5092 * work gets canceled, so just cancel this request as well instead
5093 * of executing it. We can't safely execute it anyway, as we may not
5094 * have the needed state needed for it anyway.
5096 ret = io_req_task_work_add(req);
5097 if (unlikely(ret)) {
5098 struct task_struct *tsk;
5100 WRITE_ONCE(poll->canceled, true);
5101 tsk = io_wq_get_task(req->ctx->io_wq);
5102 task_work_add(tsk, &req->task_work, TWA_NONE);
5103 wake_up_process(tsk);
5108 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
5109 __acquires(&req->ctx->completion_lock)
5111 struct io_ring_ctx *ctx = req->ctx;
5113 if (!req->result && !READ_ONCE(poll->canceled)) {
5114 struct poll_table_struct pt = { ._key = poll->events };
5116 req->result = vfs_poll(req->file, &pt) & poll->events;
5119 spin_lock_irq(&ctx->completion_lock);
5120 if (!req->result && !READ_ONCE(poll->canceled)) {
5121 add_wait_queue(poll->head, &poll->wait);
5128 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
5130 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
5131 if (req->opcode == IORING_OP_POLL_ADD)
5132 return req->async_data;
5133 return req->apoll->double_poll;
5136 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
5138 if (req->opcode == IORING_OP_POLL_ADD)
5140 return &req->apoll->poll;
5143 static void io_poll_remove_double(struct io_kiocb *req)
5145 struct io_poll_iocb *poll = io_poll_get_double(req);
5147 lockdep_assert_held(&req->ctx->completion_lock);
5149 if (poll && poll->head) {
5150 struct wait_queue_head *head = poll->head;
5152 spin_lock(&head->lock);
5153 list_del_init(&poll->wait.entry);
5154 if (poll->wait.private)
5155 refcount_dec(&req->refs);
5157 spin_unlock(&head->lock);
5161 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
5163 struct io_ring_ctx *ctx = req->ctx;
5165 io_poll_remove_double(req);
5166 req->poll.done = true;
5167 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
5168 io_commit_cqring(ctx);
5171 static void io_poll_task_func(struct callback_head *cb)
5173 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
5174 struct io_ring_ctx *ctx = req->ctx;
5175 struct io_kiocb *nxt;
5177 if (io_poll_rewait(req, &req->poll)) {
5178 spin_unlock_irq(&ctx->completion_lock);
5180 hash_del(&req->hash_node);
5181 io_poll_complete(req, req->result, 0);
5182 spin_unlock_irq(&ctx->completion_lock);
5184 nxt = io_put_req_find_next(req);
5185 io_cqring_ev_posted(ctx);
5187 __io_req_task_submit(nxt);
5190 percpu_ref_put(&ctx->refs);
5193 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
5194 int sync, void *key)
5196 struct io_kiocb *req = wait->private;
5197 struct io_poll_iocb *poll = io_poll_get_single(req);
5198 __poll_t mask = key_to_poll(key);
5200 /* for instances that support it check for an event match first: */
5201 if (mask && !(mask & poll->events))
5204 list_del_init(&wait->entry);
5206 if (poll && poll->head) {
5209 spin_lock(&poll->head->lock);
5210 done = list_empty(&poll->wait.entry);
5212 list_del_init(&poll->wait.entry);
5213 /* make sure double remove sees this as being gone */
5214 wait->private = NULL;
5215 spin_unlock(&poll->head->lock);
5217 /* use wait func handler, so it matches the rq type */
5218 poll->wait.func(&poll->wait, mode, sync, key);
5221 refcount_dec(&req->refs);
5225 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
5226 wait_queue_func_t wake_func)
5230 poll->canceled = false;
5231 poll->events = events;
5232 INIT_LIST_HEAD(&poll->wait.entry);
5233 init_waitqueue_func_entry(&poll->wait, wake_func);
5236 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5237 struct wait_queue_head *head,
5238 struct io_poll_iocb **poll_ptr)
5240 struct io_kiocb *req = pt->req;
5243 * If poll->head is already set, it's because the file being polled
5244 * uses multiple waitqueues for poll handling (eg one for read, one
5245 * for write). Setup a separate io_poll_iocb if this happens.
5247 if (unlikely(poll->head)) {
5248 struct io_poll_iocb *poll_one = poll;
5250 /* already have a 2nd entry, fail a third attempt */
5252 pt->error = -EINVAL;
5255 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5257 pt->error = -ENOMEM;
5260 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5261 refcount_inc(&req->refs);
5262 poll->wait.private = req;
5269 if (poll->events & EPOLLEXCLUSIVE)
5270 add_wait_queue_exclusive(head, &poll->wait);
5272 add_wait_queue(head, &poll->wait);
5275 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5276 struct poll_table_struct *p)
5278 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5279 struct async_poll *apoll = pt->req->apoll;
5281 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5284 static void io_async_task_func(struct callback_head *cb)
5286 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
5287 struct async_poll *apoll = req->apoll;
5288 struct io_ring_ctx *ctx = req->ctx;
5290 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
5292 if (io_poll_rewait(req, &apoll->poll)) {
5293 spin_unlock_irq(&ctx->completion_lock);
5294 percpu_ref_put(&ctx->refs);
5298 /* If req is still hashed, it cannot have been canceled. Don't check. */
5299 if (hash_hashed(&req->hash_node))
5300 hash_del(&req->hash_node);
5302 io_poll_remove_double(req);
5303 spin_unlock_irq(&ctx->completion_lock);
5305 if (!READ_ONCE(apoll->poll.canceled))
5306 __io_req_task_submit(req);
5308 __io_req_task_cancel(req, -ECANCELED);
5310 percpu_ref_put(&ctx->refs);
5311 kfree(apoll->double_poll);
5315 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5318 struct io_kiocb *req = wait->private;
5319 struct io_poll_iocb *poll = &req->apoll->poll;
5321 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5324 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5327 static void io_poll_req_insert(struct io_kiocb *req)
5329 struct io_ring_ctx *ctx = req->ctx;
5330 struct hlist_head *list;
5332 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5333 hlist_add_head(&req->hash_node, list);
5336 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5337 struct io_poll_iocb *poll,
5338 struct io_poll_table *ipt, __poll_t mask,
5339 wait_queue_func_t wake_func)
5340 __acquires(&ctx->completion_lock)
5342 struct io_ring_ctx *ctx = req->ctx;
5343 bool cancel = false;
5345 INIT_HLIST_NODE(&req->hash_node);
5346 io_init_poll_iocb(poll, mask, wake_func);
5347 poll->file = req->file;
5348 poll->wait.private = req;
5350 ipt->pt._key = mask;
5352 ipt->error = -EINVAL;
5354 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5356 spin_lock_irq(&ctx->completion_lock);
5357 if (likely(poll->head)) {
5358 spin_lock(&poll->head->lock);
5359 if (unlikely(list_empty(&poll->wait.entry))) {
5365 if (mask || ipt->error)
5366 list_del_init(&poll->wait.entry);
5368 WRITE_ONCE(poll->canceled, true);
5369 else if (!poll->done) /* actually waiting for an event */
5370 io_poll_req_insert(req);
5371 spin_unlock(&poll->head->lock);
5377 static bool io_arm_poll_handler(struct io_kiocb *req)
5379 const struct io_op_def *def = &io_op_defs[req->opcode];
5380 struct io_ring_ctx *ctx = req->ctx;
5381 struct async_poll *apoll;
5382 struct io_poll_table ipt;
5386 if (!req->file || !file_can_poll(req->file))
5388 if (req->flags & REQ_F_POLLED)
5392 else if (def->pollout)
5396 /* if we can't nonblock try, then no point in arming a poll handler */
5397 if (!io_file_supports_async(req->file, rw))
5400 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5401 if (unlikely(!apoll))
5403 apoll->double_poll = NULL;
5405 req->flags |= REQ_F_POLLED;
5410 mask |= POLLIN | POLLRDNORM;
5412 mask |= POLLOUT | POLLWRNORM;
5414 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5415 if ((req->opcode == IORING_OP_RECVMSG) &&
5416 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5419 mask |= POLLERR | POLLPRI;
5421 ipt.pt._qproc = io_async_queue_proc;
5423 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5425 if (ret || ipt.error) {
5426 io_poll_remove_double(req);
5427 spin_unlock_irq(&ctx->completion_lock);
5428 kfree(apoll->double_poll);
5432 spin_unlock_irq(&ctx->completion_lock);
5433 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
5434 apoll->poll.events);
5438 static bool __io_poll_remove_one(struct io_kiocb *req,
5439 struct io_poll_iocb *poll)
5441 bool do_complete = false;
5443 spin_lock(&poll->head->lock);
5444 WRITE_ONCE(poll->canceled, true);
5445 if (!list_empty(&poll->wait.entry)) {
5446 list_del_init(&poll->wait.entry);
5449 spin_unlock(&poll->head->lock);
5450 hash_del(&req->hash_node);
5454 static bool io_poll_remove_one(struct io_kiocb *req)
5458 io_poll_remove_double(req);
5460 if (req->opcode == IORING_OP_POLL_ADD) {
5461 do_complete = __io_poll_remove_one(req, &req->poll);
5463 struct async_poll *apoll = req->apoll;
5465 /* non-poll requests have submit ref still */
5466 do_complete = __io_poll_remove_one(req, &apoll->poll);
5469 kfree(apoll->double_poll);
5475 io_cqring_fill_event(req, -ECANCELED);
5476 io_commit_cqring(req->ctx);
5477 req_set_fail_links(req);
5478 io_put_req_deferred(req, 1);
5485 * Returns true if we found and killed one or more poll requests
5487 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5488 struct files_struct *files)
5490 struct hlist_node *tmp;
5491 struct io_kiocb *req;
5494 spin_lock_irq(&ctx->completion_lock);
5495 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5496 struct hlist_head *list;
5498 list = &ctx->cancel_hash[i];
5499 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5500 if (io_match_task(req, tsk, files))
5501 posted += io_poll_remove_one(req);
5504 spin_unlock_irq(&ctx->completion_lock);
5507 io_cqring_ev_posted(ctx);
5512 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
5514 struct hlist_head *list;
5515 struct io_kiocb *req;
5517 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5518 hlist_for_each_entry(req, list, hash_node) {
5519 if (sqe_addr != req->user_data)
5521 if (io_poll_remove_one(req))
5529 static int io_poll_remove_prep(struct io_kiocb *req,
5530 const struct io_uring_sqe *sqe)
5532 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5534 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
5538 req->poll_remove.addr = READ_ONCE(sqe->addr);
5543 * Find a running poll command that matches one specified in sqe->addr,
5544 * and remove it if found.
5546 static int io_poll_remove(struct io_kiocb *req)
5548 struct io_ring_ctx *ctx = req->ctx;
5551 spin_lock_irq(&ctx->completion_lock);
5552 ret = io_poll_cancel(ctx, req->poll_remove.addr);
5553 spin_unlock_irq(&ctx->completion_lock);
5556 req_set_fail_links(req);
5557 io_req_complete(req, ret);
5561 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5564 struct io_kiocb *req = wait->private;
5565 struct io_poll_iocb *poll = &req->poll;
5567 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5570 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5571 struct poll_table_struct *p)
5573 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5575 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5578 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5580 struct io_poll_iocb *poll = &req->poll;
5583 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5585 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
5588 events = READ_ONCE(sqe->poll32_events);
5590 events = swahw32(events);
5592 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP |
5593 (events & EPOLLEXCLUSIVE);
5597 static int io_poll_add(struct io_kiocb *req)
5599 struct io_poll_iocb *poll = &req->poll;
5600 struct io_ring_ctx *ctx = req->ctx;
5601 struct io_poll_table ipt;
5604 ipt.pt._qproc = io_poll_queue_proc;
5606 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5609 if (mask) { /* no async, we'd stolen it */
5611 io_poll_complete(req, mask, 0);
5613 spin_unlock_irq(&ctx->completion_lock);
5616 io_cqring_ev_posted(ctx);
5622 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5624 struct io_timeout_data *data = container_of(timer,
5625 struct io_timeout_data, timer);
5626 struct io_kiocb *req = data->req;
5627 struct io_ring_ctx *ctx = req->ctx;
5628 unsigned long flags;
5630 spin_lock_irqsave(&ctx->completion_lock, flags);
5631 list_del_init(&req->timeout.list);
5632 atomic_set(&req->ctx->cq_timeouts,
5633 atomic_read(&req->ctx->cq_timeouts) + 1);
5635 io_cqring_fill_event(req, -ETIME);
5636 io_commit_cqring(ctx);
5637 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5639 io_cqring_ev_posted(ctx);
5640 req_set_fail_links(req);
5642 return HRTIMER_NORESTART;
5645 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5648 struct io_timeout_data *io;
5649 struct io_kiocb *req;
5652 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5653 if (user_data == req->user_data) {
5660 return ERR_PTR(ret);
5662 io = req->async_data;
5663 ret = hrtimer_try_to_cancel(&io->timer);
5665 return ERR_PTR(-EALREADY);
5666 list_del_init(&req->timeout.list);
5670 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5672 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5675 return PTR_ERR(req);
5677 req_set_fail_links(req);
5678 io_cqring_fill_event(req, -ECANCELED);
5679 io_put_req_deferred(req, 1);
5683 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5684 struct timespec64 *ts, enum hrtimer_mode mode)
5686 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5687 struct io_timeout_data *data;
5690 return PTR_ERR(req);
5692 req->timeout.off = 0; /* noseq */
5693 data = req->async_data;
5694 list_add_tail(&req->timeout.list, &ctx->timeout_list);
5695 hrtimer_init(&data->timer, CLOCK_MONOTONIC, mode);
5696 data->timer.function = io_timeout_fn;
5697 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
5701 static int io_timeout_remove_prep(struct io_kiocb *req,
5702 const struct io_uring_sqe *sqe)
5704 struct io_timeout_rem *tr = &req->timeout_rem;
5706 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5708 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5710 if (sqe->ioprio || sqe->buf_index || sqe->len)
5713 tr->addr = READ_ONCE(sqe->addr);
5714 tr->flags = READ_ONCE(sqe->timeout_flags);
5715 if (tr->flags & IORING_TIMEOUT_UPDATE) {
5716 if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS))
5718 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
5720 } else if (tr->flags) {
5721 /* timeout removal doesn't support flags */
5729 * Remove or update an existing timeout command
5731 static int io_timeout_remove(struct io_kiocb *req)
5733 struct io_timeout_rem *tr = &req->timeout_rem;
5734 struct io_ring_ctx *ctx = req->ctx;
5737 spin_lock_irq(&ctx->completion_lock);
5738 if (req->timeout_rem.flags & IORING_TIMEOUT_UPDATE) {
5739 enum hrtimer_mode mode = (tr->flags & IORING_TIMEOUT_ABS)
5740 ? HRTIMER_MODE_ABS : HRTIMER_MODE_REL;
5742 ret = io_timeout_update(ctx, tr->addr, &tr->ts, mode);
5744 ret = io_timeout_cancel(ctx, tr->addr);
5747 io_cqring_fill_event(req, ret);
5748 io_commit_cqring(ctx);
5749 spin_unlock_irq(&ctx->completion_lock);
5750 io_cqring_ev_posted(ctx);
5752 req_set_fail_links(req);
5757 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5758 bool is_timeout_link)
5760 struct io_timeout_data *data;
5762 u32 off = READ_ONCE(sqe->off);
5764 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5766 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5768 if (off && is_timeout_link)
5770 flags = READ_ONCE(sqe->timeout_flags);
5771 if (flags & ~IORING_TIMEOUT_ABS)
5774 req->timeout.off = off;
5776 if (!req->async_data && io_alloc_async_data(req))
5779 data = req->async_data;
5782 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5785 if (flags & IORING_TIMEOUT_ABS)
5786 data->mode = HRTIMER_MODE_ABS;
5788 data->mode = HRTIMER_MODE_REL;
5790 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5794 static int io_timeout(struct io_kiocb *req)
5796 struct io_ring_ctx *ctx = req->ctx;
5797 struct io_timeout_data *data = req->async_data;
5798 struct list_head *entry;
5799 u32 tail, off = req->timeout.off;
5801 spin_lock_irq(&ctx->completion_lock);
5804 * sqe->off holds how many events that need to occur for this
5805 * timeout event to be satisfied. If it isn't set, then this is
5806 * a pure timeout request, sequence isn't used.
5808 if (io_is_timeout_noseq(req)) {
5809 entry = ctx->timeout_list.prev;
5813 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5814 req->timeout.target_seq = tail + off;
5817 * Insertion sort, ensuring the first entry in the list is always
5818 * the one we need first.
5820 list_for_each_prev(entry, &ctx->timeout_list) {
5821 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5824 if (io_is_timeout_noseq(nxt))
5826 /* nxt.seq is behind @tail, otherwise would've been completed */
5827 if (off >= nxt->timeout.target_seq - tail)
5831 list_add(&req->timeout.list, entry);
5832 data->timer.function = io_timeout_fn;
5833 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5834 spin_unlock_irq(&ctx->completion_lock);
5838 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5840 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5842 return req->user_data == (unsigned long) data;
5845 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
5847 enum io_wq_cancel cancel_ret;
5850 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr, false);
5851 switch (cancel_ret) {
5852 case IO_WQ_CANCEL_OK:
5855 case IO_WQ_CANCEL_RUNNING:
5858 case IO_WQ_CANCEL_NOTFOUND:
5866 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5867 struct io_kiocb *req, __u64 sqe_addr,
5870 unsigned long flags;
5873 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
5874 if (ret != -ENOENT) {
5875 spin_lock_irqsave(&ctx->completion_lock, flags);
5879 spin_lock_irqsave(&ctx->completion_lock, flags);
5880 ret = io_timeout_cancel(ctx, sqe_addr);
5883 ret = io_poll_cancel(ctx, sqe_addr);
5887 io_cqring_fill_event(req, ret);
5888 io_commit_cqring(ctx);
5889 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5890 io_cqring_ev_posted(ctx);
5893 req_set_fail_links(req);
5897 static int io_async_cancel_prep(struct io_kiocb *req,
5898 const struct io_uring_sqe *sqe)
5900 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5902 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5904 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5907 req->cancel.addr = READ_ONCE(sqe->addr);
5911 static int io_async_cancel(struct io_kiocb *req)
5913 struct io_ring_ctx *ctx = req->ctx;
5915 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
5919 static int io_files_update_prep(struct io_kiocb *req,
5920 const struct io_uring_sqe *sqe)
5922 if (unlikely(req->ctx->flags & IORING_SETUP_SQPOLL))
5924 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5926 if (sqe->ioprio || sqe->rw_flags)
5929 req->files_update.offset = READ_ONCE(sqe->off);
5930 req->files_update.nr_args = READ_ONCE(sqe->len);
5931 if (!req->files_update.nr_args)
5933 req->files_update.arg = READ_ONCE(sqe->addr);
5937 static int io_files_update(struct io_kiocb *req, bool force_nonblock,
5938 struct io_comp_state *cs)
5940 struct io_ring_ctx *ctx = req->ctx;
5941 struct io_uring_files_update up;
5947 up.offset = req->files_update.offset;
5948 up.fds = req->files_update.arg;
5950 mutex_lock(&ctx->uring_lock);
5951 ret = __io_sqe_files_update(ctx, &up, req->files_update.nr_args);
5952 mutex_unlock(&ctx->uring_lock);
5955 req_set_fail_links(req);
5956 __io_req_complete(req, ret, 0, cs);
5960 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5962 switch (req->opcode) {
5965 case IORING_OP_READV:
5966 case IORING_OP_READ_FIXED:
5967 case IORING_OP_READ:
5968 return io_read_prep(req, sqe);
5969 case IORING_OP_WRITEV:
5970 case IORING_OP_WRITE_FIXED:
5971 case IORING_OP_WRITE:
5972 return io_write_prep(req, sqe);
5973 case IORING_OP_POLL_ADD:
5974 return io_poll_add_prep(req, sqe);
5975 case IORING_OP_POLL_REMOVE:
5976 return io_poll_remove_prep(req, sqe);
5977 case IORING_OP_FSYNC:
5978 return io_prep_fsync(req, sqe);
5979 case IORING_OP_SYNC_FILE_RANGE:
5980 return io_prep_sfr(req, sqe);
5981 case IORING_OP_SENDMSG:
5982 case IORING_OP_SEND:
5983 return io_sendmsg_prep(req, sqe);
5984 case IORING_OP_RECVMSG:
5985 case IORING_OP_RECV:
5986 return io_recvmsg_prep(req, sqe);
5987 case IORING_OP_CONNECT:
5988 return io_connect_prep(req, sqe);
5989 case IORING_OP_TIMEOUT:
5990 return io_timeout_prep(req, sqe, false);
5991 case IORING_OP_TIMEOUT_REMOVE:
5992 return io_timeout_remove_prep(req, sqe);
5993 case IORING_OP_ASYNC_CANCEL:
5994 return io_async_cancel_prep(req, sqe);
5995 case IORING_OP_LINK_TIMEOUT:
5996 return io_timeout_prep(req, sqe, true);
5997 case IORING_OP_ACCEPT:
5998 return io_accept_prep(req, sqe);
5999 case IORING_OP_FALLOCATE:
6000 return io_fallocate_prep(req, sqe);
6001 case IORING_OP_OPENAT:
6002 return io_openat_prep(req, sqe);
6003 case IORING_OP_CLOSE:
6004 return io_close_prep(req, sqe);
6005 case IORING_OP_FILES_UPDATE:
6006 return io_files_update_prep(req, sqe);
6007 case IORING_OP_STATX:
6008 return io_statx_prep(req, sqe);
6009 case IORING_OP_FADVISE:
6010 return io_fadvise_prep(req, sqe);
6011 case IORING_OP_MADVISE:
6012 return io_madvise_prep(req, sqe);
6013 case IORING_OP_OPENAT2:
6014 return io_openat2_prep(req, sqe);
6015 case IORING_OP_EPOLL_CTL:
6016 return io_epoll_ctl_prep(req, sqe);
6017 case IORING_OP_SPLICE:
6018 return io_splice_prep(req, sqe);
6019 case IORING_OP_PROVIDE_BUFFERS:
6020 return io_provide_buffers_prep(req, sqe);
6021 case IORING_OP_REMOVE_BUFFERS:
6022 return io_remove_buffers_prep(req, sqe);
6024 return io_tee_prep(req, sqe);
6025 case IORING_OP_SHUTDOWN:
6026 return io_shutdown_prep(req, sqe);
6027 case IORING_OP_RENAMEAT:
6028 return io_renameat_prep(req, sqe);
6029 case IORING_OP_UNLINKAT:
6030 return io_unlinkat_prep(req, sqe);
6033 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
6038 static int io_req_defer_prep(struct io_kiocb *req,
6039 const struct io_uring_sqe *sqe)
6043 if (io_alloc_async_data(req))
6045 return io_req_prep(req, sqe);
6048 static u32 io_get_sequence(struct io_kiocb *req)
6050 struct io_kiocb *pos;
6051 struct io_ring_ctx *ctx = req->ctx;
6052 u32 total_submitted, nr_reqs = 0;
6054 io_for_each_link(pos, req)
6057 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
6058 return total_submitted - nr_reqs;
6061 static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe)
6063 struct io_ring_ctx *ctx = req->ctx;
6064 struct io_defer_entry *de;
6068 /* Still need defer if there is pending req in defer list. */
6069 if (likely(list_empty_careful(&ctx->defer_list) &&
6070 !(req->flags & REQ_F_IO_DRAIN)))
6073 seq = io_get_sequence(req);
6074 /* Still a chance to pass the sequence check */
6075 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
6078 if (!req->async_data) {
6079 ret = io_req_defer_prep(req, sqe);
6083 io_prep_async_link(req);
6084 de = kmalloc(sizeof(*de), GFP_KERNEL);
6088 spin_lock_irq(&ctx->completion_lock);
6089 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
6090 spin_unlock_irq(&ctx->completion_lock);
6092 io_queue_async_work(req);
6093 return -EIOCBQUEUED;
6096 trace_io_uring_defer(ctx, req, req->user_data);
6099 list_add_tail(&de->list, &ctx->defer_list);
6100 spin_unlock_irq(&ctx->completion_lock);
6101 return -EIOCBQUEUED;
6104 static void io_req_drop_files(struct io_kiocb *req)
6106 struct io_ring_ctx *ctx = req->ctx;
6107 struct io_uring_task *tctx = req->task->io_uring;
6108 unsigned long flags;
6110 spin_lock_irqsave(&ctx->inflight_lock, flags);
6111 list_del(&req->inflight_entry);
6112 if (atomic_read(&tctx->in_idle))
6113 wake_up(&tctx->wait);
6114 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
6115 req->flags &= ~REQ_F_INFLIGHT;
6116 put_files_struct(req->work.identity->files);
6117 put_nsproxy(req->work.identity->nsproxy);
6118 req->work.flags &= ~IO_WQ_WORK_FILES;
6121 static void __io_clean_op(struct io_kiocb *req)
6123 if (req->flags & REQ_F_BUFFER_SELECTED) {
6124 switch (req->opcode) {
6125 case IORING_OP_READV:
6126 case IORING_OP_READ_FIXED:
6127 case IORING_OP_READ:
6128 kfree((void *)(unsigned long)req->rw.addr);
6130 case IORING_OP_RECVMSG:
6131 case IORING_OP_RECV:
6132 kfree(req->sr_msg.kbuf);
6135 req->flags &= ~REQ_F_BUFFER_SELECTED;
6138 if (req->flags & REQ_F_NEED_CLEANUP) {
6139 switch (req->opcode) {
6140 case IORING_OP_READV:
6141 case IORING_OP_READ_FIXED:
6142 case IORING_OP_READ:
6143 case IORING_OP_WRITEV:
6144 case IORING_OP_WRITE_FIXED:
6145 case IORING_OP_WRITE: {
6146 struct io_async_rw *io = req->async_data;
6148 kfree(io->free_iovec);
6151 case IORING_OP_RECVMSG:
6152 case IORING_OP_SENDMSG: {
6153 struct io_async_msghdr *io = req->async_data;
6154 if (io->iov != io->fast_iov)
6158 case IORING_OP_SPLICE:
6160 io_put_file(req, req->splice.file_in,
6161 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
6163 case IORING_OP_OPENAT:
6164 case IORING_OP_OPENAT2:
6165 if (req->open.filename)
6166 putname(req->open.filename);
6168 case IORING_OP_RENAMEAT:
6169 putname(req->rename.oldpath);
6170 putname(req->rename.newpath);
6172 case IORING_OP_UNLINKAT:
6173 putname(req->unlink.filename);
6176 req->flags &= ~REQ_F_NEED_CLEANUP;
6179 if (req->flags & REQ_F_INFLIGHT)
6180 io_req_drop_files(req);
6183 static int io_issue_sqe(struct io_kiocb *req, bool force_nonblock,
6184 struct io_comp_state *cs)
6186 struct io_ring_ctx *ctx = req->ctx;
6189 switch (req->opcode) {
6191 ret = io_nop(req, cs);
6193 case IORING_OP_READV:
6194 case IORING_OP_READ_FIXED:
6195 case IORING_OP_READ:
6196 ret = io_read(req, force_nonblock, cs);
6198 case IORING_OP_WRITEV:
6199 case IORING_OP_WRITE_FIXED:
6200 case IORING_OP_WRITE:
6201 ret = io_write(req, force_nonblock, cs);
6203 case IORING_OP_FSYNC:
6204 ret = io_fsync(req, force_nonblock);
6206 case IORING_OP_POLL_ADD:
6207 ret = io_poll_add(req);
6209 case IORING_OP_POLL_REMOVE:
6210 ret = io_poll_remove(req);
6212 case IORING_OP_SYNC_FILE_RANGE:
6213 ret = io_sync_file_range(req, force_nonblock);
6215 case IORING_OP_SENDMSG:
6216 ret = io_sendmsg(req, force_nonblock, cs);
6218 case IORING_OP_SEND:
6219 ret = io_send(req, force_nonblock, cs);
6221 case IORING_OP_RECVMSG:
6222 ret = io_recvmsg(req, force_nonblock, cs);
6224 case IORING_OP_RECV:
6225 ret = io_recv(req, force_nonblock, cs);
6227 case IORING_OP_TIMEOUT:
6228 ret = io_timeout(req);
6230 case IORING_OP_TIMEOUT_REMOVE:
6231 ret = io_timeout_remove(req);
6233 case IORING_OP_ACCEPT:
6234 ret = io_accept(req, force_nonblock, cs);
6236 case IORING_OP_CONNECT:
6237 ret = io_connect(req, force_nonblock, cs);
6239 case IORING_OP_ASYNC_CANCEL:
6240 ret = io_async_cancel(req);
6242 case IORING_OP_FALLOCATE:
6243 ret = io_fallocate(req, force_nonblock);
6245 case IORING_OP_OPENAT:
6246 ret = io_openat(req, force_nonblock);
6248 case IORING_OP_CLOSE:
6249 ret = io_close(req, force_nonblock, cs);
6251 case IORING_OP_FILES_UPDATE:
6252 ret = io_files_update(req, force_nonblock, cs);
6254 case IORING_OP_STATX:
6255 ret = io_statx(req, force_nonblock);
6257 case IORING_OP_FADVISE:
6258 ret = io_fadvise(req, force_nonblock);
6260 case IORING_OP_MADVISE:
6261 ret = io_madvise(req, force_nonblock);
6263 case IORING_OP_OPENAT2:
6264 ret = io_openat2(req, force_nonblock);
6266 case IORING_OP_EPOLL_CTL:
6267 ret = io_epoll_ctl(req, force_nonblock, cs);
6269 case IORING_OP_SPLICE:
6270 ret = io_splice(req, force_nonblock);
6272 case IORING_OP_PROVIDE_BUFFERS:
6273 ret = io_provide_buffers(req, force_nonblock, cs);
6275 case IORING_OP_REMOVE_BUFFERS:
6276 ret = io_remove_buffers(req, force_nonblock, cs);
6279 ret = io_tee(req, force_nonblock);
6281 case IORING_OP_SHUTDOWN:
6282 ret = io_shutdown(req, force_nonblock);
6284 case IORING_OP_RENAMEAT:
6285 ret = io_renameat(req, force_nonblock);
6287 case IORING_OP_UNLINKAT:
6288 ret = io_unlinkat(req, force_nonblock);
6298 /* If the op doesn't have a file, we're not polling for it */
6299 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
6300 const bool in_async = io_wq_current_is_worker();
6302 /* workqueue context doesn't hold uring_lock, grab it now */
6304 mutex_lock(&ctx->uring_lock);
6306 io_iopoll_req_issued(req, in_async);
6309 mutex_unlock(&ctx->uring_lock);
6315 static struct io_wq_work *io_wq_submit_work(struct io_wq_work *work)
6317 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6318 struct io_kiocb *timeout;
6321 timeout = io_prep_linked_timeout(req);
6323 io_queue_linked_timeout(timeout);
6325 /* if NO_CANCEL is set, we must still run the work */
6326 if ((work->flags & (IO_WQ_WORK_CANCEL|IO_WQ_WORK_NO_CANCEL)) ==
6327 IO_WQ_WORK_CANCEL) {
6333 ret = io_issue_sqe(req, false, NULL);
6335 * We can get EAGAIN for polled IO even though we're
6336 * forcing a sync submission from here, since we can't
6337 * wait for request slots on the block side.
6347 * io_iopoll_complete() does not hold completion_lock to complete
6348 * polled io, so here for polled io, just mark it done and still let
6349 * io_iopoll_complete() complete it.
6351 if (req->ctx->flags & IORING_SETUP_IOPOLL) {
6352 struct kiocb *kiocb = &req->rw.kiocb;
6354 kiocb_done(kiocb, ret, NULL);
6356 req_set_fail_links(req);
6357 io_req_complete(req, ret);
6361 return io_steal_work(req);
6364 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6367 struct fixed_file_table *table;
6369 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
6370 return table->files[index & IORING_FILE_TABLE_MASK];
6373 static struct file *io_file_get(struct io_submit_state *state,
6374 struct io_kiocb *req, int fd, bool fixed)
6376 struct io_ring_ctx *ctx = req->ctx;
6380 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6382 fd = array_index_nospec(fd, ctx->nr_user_files);
6383 file = io_file_from_index(ctx, fd);
6384 io_set_resource_node(req);
6386 trace_io_uring_file_get(ctx, fd);
6387 file = __io_file_get(state, fd);
6393 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6395 struct io_timeout_data *data = container_of(timer,
6396 struct io_timeout_data, timer);
6397 struct io_kiocb *prev, *req = data->req;
6398 struct io_ring_ctx *ctx = req->ctx;
6399 unsigned long flags;
6401 spin_lock_irqsave(&ctx->completion_lock, flags);
6402 prev = req->timeout.head;
6403 req->timeout.head = NULL;
6406 * We don't expect the list to be empty, that will only happen if we
6407 * race with the completion of the linked work.
6409 if (prev && refcount_inc_not_zero(&prev->refs))
6410 io_remove_next_linked(prev);
6413 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6416 req_set_fail_links(prev);
6417 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6420 io_req_complete(req, -ETIME);
6422 return HRTIMER_NORESTART;
6425 static void __io_queue_linked_timeout(struct io_kiocb *req)
6428 * If the back reference is NULL, then our linked request finished
6429 * before we got a chance to setup the timer
6431 if (req->timeout.head) {
6432 struct io_timeout_data *data = req->async_data;
6434 data->timer.function = io_link_timeout_fn;
6435 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6440 static void io_queue_linked_timeout(struct io_kiocb *req)
6442 struct io_ring_ctx *ctx = req->ctx;
6444 spin_lock_irq(&ctx->completion_lock);
6445 __io_queue_linked_timeout(req);
6446 spin_unlock_irq(&ctx->completion_lock);
6448 /* drop submission reference */
6452 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6454 struct io_kiocb *nxt = req->link;
6456 if (!nxt || (req->flags & REQ_F_LINK_TIMEOUT) ||
6457 nxt->opcode != IORING_OP_LINK_TIMEOUT)
6460 nxt->timeout.head = req;
6461 nxt->flags |= REQ_F_LTIMEOUT_ACTIVE;
6462 req->flags |= REQ_F_LINK_TIMEOUT;
6466 static void __io_queue_sqe(struct io_kiocb *req, struct io_comp_state *cs)
6468 struct io_kiocb *linked_timeout;
6469 const struct cred *old_creds = NULL;
6473 linked_timeout = io_prep_linked_timeout(req);
6475 if ((req->flags & REQ_F_WORK_INITIALIZED) &&
6476 (req->work.flags & IO_WQ_WORK_CREDS) &&
6477 req->work.identity->creds != current_cred()) {
6479 revert_creds(old_creds);
6480 if (old_creds == req->work.identity->creds)
6481 old_creds = NULL; /* restored original creds */
6483 old_creds = override_creds(req->work.identity->creds);
6486 ret = io_issue_sqe(req, true, cs);
6489 * We async punt it if the file wasn't marked NOWAIT, or if the file
6490 * doesn't support non-blocking read/write attempts
6492 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6493 if (!io_arm_poll_handler(req)) {
6495 * Queued up for async execution, worker will release
6496 * submit reference when the iocb is actually submitted.
6498 io_queue_async_work(req);
6502 io_queue_linked_timeout(linked_timeout);
6503 } else if (likely(!ret)) {
6504 /* drop submission reference */
6505 req = io_put_req_find_next(req);
6507 io_queue_linked_timeout(linked_timeout);
6510 if (!(req->flags & REQ_F_FORCE_ASYNC))
6512 io_queue_async_work(req);
6515 /* un-prep timeout, so it'll be killed as any other linked */
6516 req->flags &= ~REQ_F_LINK_TIMEOUT;
6517 req_set_fail_links(req);
6519 io_req_complete(req, ret);
6523 revert_creds(old_creds);
6526 static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6527 struct io_comp_state *cs)
6531 ret = io_req_defer(req, sqe);
6533 if (ret != -EIOCBQUEUED) {
6535 req_set_fail_links(req);
6537 io_req_complete(req, ret);
6539 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6540 if (!req->async_data) {
6541 ret = io_req_defer_prep(req, sqe);
6545 io_queue_async_work(req);
6548 ret = io_req_prep(req, sqe);
6552 __io_queue_sqe(req, cs);
6556 static inline void io_queue_link_head(struct io_kiocb *req,
6557 struct io_comp_state *cs)
6559 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
6561 io_req_complete(req, -ECANCELED);
6563 io_queue_sqe(req, NULL, cs);
6566 struct io_submit_link {
6567 struct io_kiocb *head;
6568 struct io_kiocb *last;
6571 static int io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6572 struct io_submit_link *link, struct io_comp_state *cs)
6574 struct io_ring_ctx *ctx = req->ctx;
6578 * If we already have a head request, queue this one for async
6579 * submittal once the head completes. If we don't have a head but
6580 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6581 * submitted sync once the chain is complete. If none of those
6582 * conditions are true (normal request), then just queue it.
6585 struct io_kiocb *head = link->head;
6588 * Taking sequential execution of a link, draining both sides
6589 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6590 * requests in the link. So, it drains the head and the
6591 * next after the link request. The last one is done via
6592 * drain_next flag to persist the effect across calls.
6594 if (req->flags & REQ_F_IO_DRAIN) {
6595 head->flags |= REQ_F_IO_DRAIN;
6596 ctx->drain_next = 1;
6598 ret = io_req_defer_prep(req, sqe);
6599 if (unlikely(ret)) {
6600 /* fail even hard links since we don't submit */
6601 head->flags |= REQ_F_FAIL_LINK;
6604 trace_io_uring_link(ctx, req, head);
6605 link->last->link = req;
6608 /* last request of a link, enqueue the link */
6609 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6610 io_queue_link_head(head, cs);
6614 if (unlikely(ctx->drain_next)) {
6615 req->flags |= REQ_F_IO_DRAIN;
6616 ctx->drain_next = 0;
6618 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6619 ret = io_req_defer_prep(req, sqe);
6621 req->flags |= REQ_F_FAIL_LINK;
6625 io_queue_sqe(req, sqe, cs);
6633 * Batched submission is done, ensure local IO is flushed out.
6635 static void io_submit_state_end(struct io_submit_state *state)
6637 if (!list_empty(&state->comp.list))
6638 io_submit_flush_completions(&state->comp);
6639 if (state->plug_started)
6640 blk_finish_plug(&state->plug);
6641 io_state_file_put(state);
6642 if (state->free_reqs)
6643 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
6647 * Start submission side cache.
6649 static void io_submit_state_start(struct io_submit_state *state,
6650 struct io_ring_ctx *ctx, unsigned int max_ios)
6652 state->plug_started = false;
6654 INIT_LIST_HEAD(&state->comp.list);
6655 state->comp.ctx = ctx;
6656 state->free_reqs = 0;
6657 state->file_refs = 0;
6658 state->ios_left = max_ios;
6661 static void io_commit_sqring(struct io_ring_ctx *ctx)
6663 struct io_rings *rings = ctx->rings;
6666 * Ensure any loads from the SQEs are done at this point,
6667 * since once we write the new head, the application could
6668 * write new data to them.
6670 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6674 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6675 * that is mapped by userspace. This means that care needs to be taken to
6676 * ensure that reads are stable, as we cannot rely on userspace always
6677 * being a good citizen. If members of the sqe are validated and then later
6678 * used, it's important that those reads are done through READ_ONCE() to
6679 * prevent a re-load down the line.
6681 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6683 u32 *sq_array = ctx->sq_array;
6687 * The cached sq head (or cq tail) serves two purposes:
6689 * 1) allows us to batch the cost of updating the user visible
6691 * 2) allows the kernel side to track the head on its own, even
6692 * though the application is the one updating it.
6694 head = READ_ONCE(sq_array[ctx->cached_sq_head & ctx->sq_mask]);
6695 if (likely(head < ctx->sq_entries))
6696 return &ctx->sq_sqes[head];
6698 /* drop invalid entries */
6699 ctx->cached_sq_dropped++;
6700 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6704 static inline void io_consume_sqe(struct io_ring_ctx *ctx)
6706 ctx->cached_sq_head++;
6710 * Check SQE restrictions (opcode and flags).
6712 * Returns 'true' if SQE is allowed, 'false' otherwise.
6714 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6715 struct io_kiocb *req,
6716 unsigned int sqe_flags)
6718 if (!ctx->restricted)
6721 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6724 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6725 ctx->restrictions.sqe_flags_required)
6728 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6729 ctx->restrictions.sqe_flags_required))
6735 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
6736 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
6737 IOSQE_BUFFER_SELECT)
6739 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6740 const struct io_uring_sqe *sqe,
6741 struct io_submit_state *state)
6743 unsigned int sqe_flags;
6746 req->opcode = READ_ONCE(sqe->opcode);
6747 req->user_data = READ_ONCE(sqe->user_data);
6748 req->async_data = NULL;
6753 req->fixed_file_refs = NULL;
6754 /* one is dropped after submission, the other at completion */
6755 refcount_set(&req->refs, 2);
6756 req->task = current;
6759 if (unlikely(req->opcode >= IORING_OP_LAST))
6762 if (unlikely(io_sq_thread_acquire_mm_files(ctx, req)))
6765 sqe_flags = READ_ONCE(sqe->flags);
6766 /* enforce forwards compatibility on users */
6767 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6770 if (unlikely(!io_check_restriction(ctx, req, sqe_flags)))
6773 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6774 !io_op_defs[req->opcode].buffer_select)
6777 id = READ_ONCE(sqe->personality);
6779 struct io_identity *iod;
6781 iod = idr_find(&ctx->personality_idr, id);
6784 refcount_inc(&iod->count);
6786 __io_req_init_async(req);
6787 get_cred(iod->creds);
6788 req->work.identity = iod;
6789 req->work.flags |= IO_WQ_WORK_CREDS;
6792 /* same numerical values with corresponding REQ_F_*, safe to copy */
6793 req->flags |= sqe_flags;
6796 * Plug now if we have more than 1 IO left after this, and the target
6797 * is potentially a read/write to block based storage.
6799 if (!state->plug_started && state->ios_left > 1 &&
6800 io_op_defs[req->opcode].plug) {
6801 blk_start_plug(&state->plug);
6802 state->plug_started = true;
6806 if (io_op_defs[req->opcode].needs_file) {
6807 bool fixed = req->flags & REQ_F_FIXED_FILE;
6809 req->file = io_file_get(state, req, READ_ONCE(sqe->fd), fixed);
6810 if (unlikely(!req->file &&
6811 !io_op_defs[req->opcode].needs_file_no_error))
6819 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6821 struct io_submit_state state;
6822 struct io_submit_link link;
6823 int i, submitted = 0;
6825 /* if we have a backlog and couldn't flush it all, return BUSY */
6826 if (test_bit(0, &ctx->sq_check_overflow)) {
6827 if (!list_empty(&ctx->cq_overflow_list) &&
6828 !io_cqring_overflow_flush(ctx, false, NULL, NULL))
6832 /* make sure SQ entry isn't read before tail */
6833 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6835 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6838 percpu_counter_add(¤t->io_uring->inflight, nr);
6839 refcount_add(nr, ¤t->usage);
6841 io_submit_state_start(&state, ctx, nr);
6844 for (i = 0; i < nr; i++) {
6845 const struct io_uring_sqe *sqe;
6846 struct io_kiocb *req;
6849 sqe = io_get_sqe(ctx);
6850 if (unlikely(!sqe)) {
6851 io_consume_sqe(ctx);
6854 req = io_alloc_req(ctx, &state);
6855 if (unlikely(!req)) {
6857 submitted = -EAGAIN;
6860 io_consume_sqe(ctx);
6861 /* will complete beyond this point, count as submitted */
6864 err = io_init_req(ctx, req, sqe, &state);
6865 if (unlikely(err)) {
6868 io_req_complete(req, err);
6872 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6873 true, io_async_submit(ctx));
6874 err = io_submit_sqe(req, sqe, &link, &state.comp);
6879 if (unlikely(submitted != nr)) {
6880 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6881 struct io_uring_task *tctx = current->io_uring;
6882 int unused = nr - ref_used;
6884 percpu_ref_put_many(&ctx->refs, unused);
6885 percpu_counter_sub(&tctx->inflight, unused);
6886 put_task_struct_many(current, unused);
6889 io_queue_link_head(link.head, &state.comp);
6890 io_submit_state_end(&state);
6892 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6893 io_commit_sqring(ctx);
6898 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6900 /* Tell userspace we may need a wakeup call */
6901 spin_lock_irq(&ctx->completion_lock);
6902 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6903 spin_unlock_irq(&ctx->completion_lock);
6906 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6908 spin_lock_irq(&ctx->completion_lock);
6909 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6910 spin_unlock_irq(&ctx->completion_lock);
6913 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
6915 unsigned int to_submit;
6918 to_submit = io_sqring_entries(ctx);
6919 /* if we're handling multiple rings, cap submit size for fairness */
6920 if (cap_entries && to_submit > 8)
6923 if (!list_empty(&ctx->iopoll_list) || to_submit) {
6924 unsigned nr_events = 0;
6926 mutex_lock(&ctx->uring_lock);
6927 if (!list_empty(&ctx->iopoll_list))
6928 io_do_iopoll(ctx, &nr_events, 0);
6930 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)))
6931 ret = io_submit_sqes(ctx, to_submit);
6932 mutex_unlock(&ctx->uring_lock);
6935 if (!io_sqring_full(ctx) && wq_has_sleeper(&ctx->sqo_sq_wait))
6936 wake_up(&ctx->sqo_sq_wait);
6941 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
6943 struct io_ring_ctx *ctx;
6944 unsigned sq_thread_idle = 0;
6946 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6947 if (sq_thread_idle < ctx->sq_thread_idle)
6948 sq_thread_idle = ctx->sq_thread_idle;
6951 sqd->sq_thread_idle = sq_thread_idle;
6954 static void io_sqd_init_new(struct io_sq_data *sqd)
6956 struct io_ring_ctx *ctx;
6958 while (!list_empty(&sqd->ctx_new_list)) {
6959 ctx = list_first_entry(&sqd->ctx_new_list, struct io_ring_ctx, sqd_list);
6960 list_move_tail(&ctx->sqd_list, &sqd->ctx_list);
6961 complete(&ctx->sq_thread_comp);
6964 io_sqd_update_thread_idle(sqd);
6967 static int io_sq_thread(void *data)
6969 struct cgroup_subsys_state *cur_css = NULL;
6970 struct files_struct *old_files = current->files;
6971 struct nsproxy *old_nsproxy = current->nsproxy;
6972 const struct cred *old_cred = NULL;
6973 struct io_sq_data *sqd = data;
6974 struct io_ring_ctx *ctx;
6975 unsigned long timeout = 0;
6979 current->files = NULL;
6980 current->nsproxy = NULL;
6981 task_unlock(current);
6983 while (!kthread_should_stop()) {
6985 bool cap_entries, sqt_spin, needs_sched;
6988 * Any changes to the sqd lists are synchronized through the
6989 * kthread parking. This synchronizes the thread vs users,
6990 * the users are synchronized on the sqd->ctx_lock.
6992 if (kthread_should_park()) {
6995 * When sq thread is unparked, in case the previous park operation
6996 * comes from io_put_sq_data(), which means that sq thread is going
6997 * to be stopped, so here needs to have a check.
6999 if (kthread_should_stop())
7003 if (unlikely(!list_empty(&sqd->ctx_new_list))) {
7004 io_sqd_init_new(sqd);
7005 timeout = jiffies + sqd->sq_thread_idle;
7009 cap_entries = !list_is_singular(&sqd->ctx_list);
7010 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
7011 if (current->cred != ctx->creds) {
7013 revert_creds(old_cred);
7014 old_cred = override_creds(ctx->creds);
7016 io_sq_thread_associate_blkcg(ctx, &cur_css);
7018 current->loginuid = ctx->loginuid;
7019 current->sessionid = ctx->sessionid;
7022 ret = __io_sq_thread(ctx, cap_entries);
7023 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
7026 io_sq_thread_drop_mm_files();
7029 if (sqt_spin || !time_after(jiffies, timeout)) {
7033 timeout = jiffies + sqd->sq_thread_idle;
7037 if (kthread_should_park())
7041 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
7042 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
7043 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
7044 !list_empty_careful(&ctx->iopoll_list)) {
7045 needs_sched = false;
7048 if (io_sqring_entries(ctx)) {
7049 needs_sched = false;
7055 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7056 io_ring_set_wakeup_flag(ctx);
7059 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7060 io_ring_clear_wakeup_flag(ctx);
7063 finish_wait(&sqd->wait, &wait);
7064 timeout = jiffies + sqd->sq_thread_idle;
7070 io_sq_thread_unassociate_blkcg();
7072 revert_creds(old_cred);
7075 current->files = old_files;
7076 current->nsproxy = old_nsproxy;
7077 task_unlock(current);
7084 struct io_wait_queue {
7085 struct wait_queue_entry wq;
7086 struct io_ring_ctx *ctx;
7088 unsigned nr_timeouts;
7091 static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush)
7093 struct io_ring_ctx *ctx = iowq->ctx;
7096 * Wake up if we have enough events, or if a timeout occurred since we
7097 * started waiting. For timeouts, we always want to return to userspace,
7098 * regardless of event count.
7100 return io_cqring_events(ctx, noflush) >= iowq->to_wait ||
7101 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
7104 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
7105 int wake_flags, void *key)
7107 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
7110 /* use noflush == true, as we can't safely rely on locking context */
7111 if (!io_should_wake(iowq, true))
7114 return autoremove_wake_function(curr, mode, wake_flags, key);
7117 static int io_run_task_work_sig(void)
7119 if (io_run_task_work())
7121 if (!signal_pending(current))
7123 if (test_tsk_thread_flag(current, TIF_NOTIFY_SIGNAL))
7124 return -ERESTARTSYS;
7129 * Wait until events become available, if we don't already have some. The
7130 * application must reap them itself, as they reside on the shared cq ring.
7132 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
7133 const sigset_t __user *sig, size_t sigsz,
7134 struct __kernel_timespec __user *uts)
7136 struct io_wait_queue iowq = {
7139 .func = io_wake_function,
7140 .entry = LIST_HEAD_INIT(iowq.wq.entry),
7143 .to_wait = min_events,
7145 struct io_rings *rings = ctx->rings;
7146 struct timespec64 ts;
7147 signed long timeout = 0;
7151 if (io_cqring_events(ctx, false) >= min_events)
7153 if (!io_run_task_work())
7158 #ifdef CONFIG_COMPAT
7159 if (in_compat_syscall())
7160 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
7164 ret = set_user_sigmask(sig, sigsz);
7171 if (get_timespec64(&ts, uts))
7173 timeout = timespec64_to_jiffies(&ts);
7176 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
7177 trace_io_uring_cqring_wait(ctx, min_events);
7179 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
7180 TASK_INTERRUPTIBLE);
7181 /* make sure we run task_work before checking for signals */
7182 ret = io_run_task_work_sig();
7187 if (io_should_wake(&iowq, false))
7190 timeout = schedule_timeout(timeout);
7199 finish_wait(&ctx->wait, &iowq.wq);
7201 restore_saved_sigmask_unless(ret == -EINTR);
7203 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
7206 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
7208 #if defined(CONFIG_UNIX)
7209 if (ctx->ring_sock) {
7210 struct sock *sock = ctx->ring_sock->sk;
7211 struct sk_buff *skb;
7213 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
7219 for (i = 0; i < ctx->nr_user_files; i++) {
7222 file = io_file_from_index(ctx, i);
7229 static void io_file_ref_kill(struct percpu_ref *ref)
7231 struct fixed_file_data *data;
7233 data = container_of(ref, struct fixed_file_data, refs);
7234 complete(&data->done);
7237 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7239 struct fixed_file_data *data = ctx->file_data;
7240 struct fixed_file_ref_node *ref_node = NULL;
7241 unsigned nr_tables, i;
7246 spin_lock_bh(&data->lock);
7247 ref_node = data->node;
7248 spin_unlock_bh(&data->lock);
7250 percpu_ref_kill(&ref_node->refs);
7252 percpu_ref_kill(&data->refs);
7254 /* wait for all refs nodes to complete */
7255 flush_delayed_work(&ctx->file_put_work);
7256 wait_for_completion(&data->done);
7258 __io_sqe_files_unregister(ctx);
7259 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
7260 for (i = 0; i < nr_tables; i++)
7261 kfree(data->table[i].files);
7263 percpu_ref_exit(&data->refs);
7265 ctx->file_data = NULL;
7266 ctx->nr_user_files = 0;
7270 static void io_put_sq_data(struct io_sq_data *sqd)
7272 if (refcount_dec_and_test(&sqd->refs)) {
7274 * The park is a bit of a work-around, without it we get
7275 * warning spews on shutdown with SQPOLL set and affinity
7276 * set to a single CPU.
7279 kthread_park(sqd->thread);
7280 kthread_stop(sqd->thread);
7287 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7289 struct io_ring_ctx *ctx_attach;
7290 struct io_sq_data *sqd;
7293 f = fdget(p->wq_fd);
7295 return ERR_PTR(-ENXIO);
7296 if (f.file->f_op != &io_uring_fops) {
7298 return ERR_PTR(-EINVAL);
7301 ctx_attach = f.file->private_data;
7302 sqd = ctx_attach->sq_data;
7305 return ERR_PTR(-EINVAL);
7308 refcount_inc(&sqd->refs);
7313 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p)
7315 struct io_sq_data *sqd;
7317 if (p->flags & IORING_SETUP_ATTACH_WQ)
7318 return io_attach_sq_data(p);
7320 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7322 return ERR_PTR(-ENOMEM);
7324 refcount_set(&sqd->refs, 1);
7325 INIT_LIST_HEAD(&sqd->ctx_list);
7326 INIT_LIST_HEAD(&sqd->ctx_new_list);
7327 mutex_init(&sqd->ctx_lock);
7328 mutex_init(&sqd->lock);
7329 init_waitqueue_head(&sqd->wait);
7333 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7334 __releases(&sqd->lock)
7338 kthread_unpark(sqd->thread);
7339 mutex_unlock(&sqd->lock);
7342 static void io_sq_thread_park(struct io_sq_data *sqd)
7343 __acquires(&sqd->lock)
7347 mutex_lock(&sqd->lock);
7348 kthread_park(sqd->thread);
7351 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
7353 struct io_sq_data *sqd = ctx->sq_data;
7358 * We may arrive here from the error branch in
7359 * io_sq_offload_create() where the kthread is created
7360 * without being waked up, thus wake it up now to make
7361 * sure the wait will complete.
7363 wake_up_process(sqd->thread);
7364 wait_for_completion(&ctx->sq_thread_comp);
7366 io_sq_thread_park(sqd);
7369 mutex_lock(&sqd->ctx_lock);
7370 list_del(&ctx->sqd_list);
7371 io_sqd_update_thread_idle(sqd);
7372 mutex_unlock(&sqd->ctx_lock);
7375 io_sq_thread_unpark(sqd);
7377 io_put_sq_data(sqd);
7378 ctx->sq_data = NULL;
7382 static void io_finish_async(struct io_ring_ctx *ctx)
7384 io_sq_thread_stop(ctx);
7387 io_wq_destroy(ctx->io_wq);
7392 #if defined(CONFIG_UNIX)
7394 * Ensure the UNIX gc is aware of our file set, so we are certain that
7395 * the io_uring can be safely unregistered on process exit, even if we have
7396 * loops in the file referencing.
7398 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7400 struct sock *sk = ctx->ring_sock->sk;
7401 struct scm_fp_list *fpl;
7402 struct sk_buff *skb;
7405 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7409 skb = alloc_skb(0, GFP_KERNEL);
7418 fpl->user = get_uid(ctx->user);
7419 for (i = 0; i < nr; i++) {
7420 struct file *file = io_file_from_index(ctx, i + offset);
7424 fpl->fp[nr_files] = get_file(file);
7425 unix_inflight(fpl->user, fpl->fp[nr_files]);
7430 fpl->max = SCM_MAX_FD;
7431 fpl->count = nr_files;
7432 UNIXCB(skb).fp = fpl;
7433 skb->destructor = unix_destruct_scm;
7434 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7435 skb_queue_head(&sk->sk_receive_queue, skb);
7437 for (i = 0; i < nr_files; i++)
7448 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7449 * causes regular reference counting to break down. We rely on the UNIX
7450 * garbage collection to take care of this problem for us.
7452 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7454 unsigned left, total;
7458 left = ctx->nr_user_files;
7460 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7462 ret = __io_sqe_files_scm(ctx, this_files, total);
7466 total += this_files;
7472 while (total < ctx->nr_user_files) {
7473 struct file *file = io_file_from_index(ctx, total);
7483 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7489 static int io_sqe_alloc_file_tables(struct fixed_file_data *file_data,
7490 unsigned nr_tables, unsigned nr_files)
7494 for (i = 0; i < nr_tables; i++) {
7495 struct fixed_file_table *table = &file_data->table[i];
7496 unsigned this_files;
7498 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
7499 table->files = kcalloc(this_files, sizeof(struct file *),
7503 nr_files -= this_files;
7509 for (i = 0; i < nr_tables; i++) {
7510 struct fixed_file_table *table = &file_data->table[i];
7511 kfree(table->files);
7516 static void io_ring_file_put(struct io_ring_ctx *ctx, struct file *file)
7518 #if defined(CONFIG_UNIX)
7519 struct sock *sock = ctx->ring_sock->sk;
7520 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7521 struct sk_buff *skb;
7524 __skb_queue_head_init(&list);
7527 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7528 * remove this entry and rearrange the file array.
7530 skb = skb_dequeue(head);
7532 struct scm_fp_list *fp;
7534 fp = UNIXCB(skb).fp;
7535 for (i = 0; i < fp->count; i++) {
7538 if (fp->fp[i] != file)
7541 unix_notinflight(fp->user, fp->fp[i]);
7542 left = fp->count - 1 - i;
7544 memmove(&fp->fp[i], &fp->fp[i + 1],
7545 left * sizeof(struct file *));
7552 __skb_queue_tail(&list, skb);
7562 __skb_queue_tail(&list, skb);
7564 skb = skb_dequeue(head);
7567 if (skb_peek(&list)) {
7568 spin_lock_irq(&head->lock);
7569 while ((skb = __skb_dequeue(&list)) != NULL)
7570 __skb_queue_tail(head, skb);
7571 spin_unlock_irq(&head->lock);
7578 struct io_file_put {
7579 struct list_head list;
7583 static void __io_file_put_work(struct fixed_file_ref_node *ref_node)
7585 struct fixed_file_data *file_data = ref_node->file_data;
7586 struct io_ring_ctx *ctx = file_data->ctx;
7587 struct io_file_put *pfile, *tmp;
7589 list_for_each_entry_safe(pfile, tmp, &ref_node->file_list, list) {
7590 list_del(&pfile->list);
7591 io_ring_file_put(ctx, pfile->file);
7595 percpu_ref_exit(&ref_node->refs);
7597 percpu_ref_put(&file_data->refs);
7600 static void io_file_put_work(struct work_struct *work)
7602 struct io_ring_ctx *ctx;
7603 struct llist_node *node;
7605 ctx = container_of(work, struct io_ring_ctx, file_put_work.work);
7606 node = llist_del_all(&ctx->file_put_llist);
7609 struct fixed_file_ref_node *ref_node;
7610 struct llist_node *next = node->next;
7612 ref_node = llist_entry(node, struct fixed_file_ref_node, llist);
7613 __io_file_put_work(ref_node);
7618 static void io_file_data_ref_zero(struct percpu_ref *ref)
7620 struct fixed_file_ref_node *ref_node;
7621 struct fixed_file_data *data;
7622 struct io_ring_ctx *ctx;
7623 bool first_add = false;
7626 ref_node = container_of(ref, struct fixed_file_ref_node, refs);
7627 data = ref_node->file_data;
7630 spin_lock_bh(&data->lock);
7631 ref_node->done = true;
7633 while (!list_empty(&data->ref_list)) {
7634 ref_node = list_first_entry(&data->ref_list,
7635 struct fixed_file_ref_node, node);
7636 /* recycle ref nodes in order */
7637 if (!ref_node->done)
7639 list_del(&ref_node->node);
7640 first_add |= llist_add(&ref_node->llist, &ctx->file_put_llist);
7642 spin_unlock_bh(&data->lock);
7644 if (percpu_ref_is_dying(&data->refs))
7648 mod_delayed_work(system_wq, &ctx->file_put_work, 0);
7650 queue_delayed_work(system_wq, &ctx->file_put_work, delay);
7653 static struct fixed_file_ref_node *alloc_fixed_file_ref_node(
7654 struct io_ring_ctx *ctx)
7656 struct fixed_file_ref_node *ref_node;
7658 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7660 return ERR_PTR(-ENOMEM);
7662 if (percpu_ref_init(&ref_node->refs, io_file_data_ref_zero,
7665 return ERR_PTR(-ENOMEM);
7667 INIT_LIST_HEAD(&ref_node->node);
7668 INIT_LIST_HEAD(&ref_node->file_list);
7669 ref_node->file_data = ctx->file_data;
7670 ref_node->done = false;
7674 static void destroy_fixed_file_ref_node(struct fixed_file_ref_node *ref_node)
7676 percpu_ref_exit(&ref_node->refs);
7680 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7683 __s32 __user *fds = (__s32 __user *) arg;
7684 unsigned nr_tables, i;
7686 int fd, ret = -ENOMEM;
7687 struct fixed_file_ref_node *ref_node;
7688 struct fixed_file_data *file_data;
7694 if (nr_args > IORING_MAX_FIXED_FILES)
7697 file_data = kzalloc(sizeof(*ctx->file_data), GFP_KERNEL);
7700 file_data->ctx = ctx;
7701 init_completion(&file_data->done);
7702 INIT_LIST_HEAD(&file_data->ref_list);
7703 spin_lock_init(&file_data->lock);
7705 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
7706 file_data->table = kcalloc(nr_tables, sizeof(*file_data->table),
7708 if (!file_data->table)
7711 if (percpu_ref_init(&file_data->refs, io_file_ref_kill,
7712 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
7715 if (io_sqe_alloc_file_tables(file_data, nr_tables, nr_args))
7717 ctx->file_data = file_data;
7719 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7720 struct fixed_file_table *table;
7723 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7727 /* allow sparse sets */
7737 * Don't allow io_uring instances to be registered. If UNIX
7738 * isn't enabled, then this causes a reference cycle and this
7739 * instance can never get freed. If UNIX is enabled we'll
7740 * handle it just fine, but there's still no point in allowing
7741 * a ring fd as it doesn't support regular read/write anyway.
7743 if (file->f_op == &io_uring_fops) {
7747 table = &file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7748 index = i & IORING_FILE_TABLE_MASK;
7749 table->files[index] = file;
7752 ret = io_sqe_files_scm(ctx);
7754 io_sqe_files_unregister(ctx);
7758 ref_node = alloc_fixed_file_ref_node(ctx);
7759 if (IS_ERR(ref_node)) {
7760 io_sqe_files_unregister(ctx);
7761 return PTR_ERR(ref_node);
7764 file_data->node = ref_node;
7765 spin_lock_bh(&file_data->lock);
7766 list_add_tail(&ref_node->node, &file_data->ref_list);
7767 spin_unlock_bh(&file_data->lock);
7768 percpu_ref_get(&file_data->refs);
7771 for (i = 0; i < ctx->nr_user_files; i++) {
7772 file = io_file_from_index(ctx, i);
7776 for (i = 0; i < nr_tables; i++)
7777 kfree(file_data->table[i].files);
7778 ctx->nr_user_files = 0;
7780 percpu_ref_exit(&file_data->refs);
7782 kfree(file_data->table);
7784 ctx->file_data = NULL;
7788 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7791 #if defined(CONFIG_UNIX)
7792 struct sock *sock = ctx->ring_sock->sk;
7793 struct sk_buff_head *head = &sock->sk_receive_queue;
7794 struct sk_buff *skb;
7797 * See if we can merge this file into an existing skb SCM_RIGHTS
7798 * file set. If there's no room, fall back to allocating a new skb
7799 * and filling it in.
7801 spin_lock_irq(&head->lock);
7802 skb = skb_peek(head);
7804 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7806 if (fpl->count < SCM_MAX_FD) {
7807 __skb_unlink(skb, head);
7808 spin_unlock_irq(&head->lock);
7809 fpl->fp[fpl->count] = get_file(file);
7810 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7812 spin_lock_irq(&head->lock);
7813 __skb_queue_head(head, skb);
7818 spin_unlock_irq(&head->lock);
7825 return __io_sqe_files_scm(ctx, 1, index);
7831 static int io_queue_file_removal(struct fixed_file_data *data,
7834 struct io_file_put *pfile;
7835 struct fixed_file_ref_node *ref_node = data->node;
7837 pfile = kzalloc(sizeof(*pfile), GFP_KERNEL);
7842 list_add(&pfile->list, &ref_node->file_list);
7847 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7848 struct io_uring_files_update *up,
7851 struct fixed_file_data *data = ctx->file_data;
7852 struct fixed_file_ref_node *ref_node;
7857 bool needs_switch = false;
7859 if (check_add_overflow(up->offset, nr_args, &done))
7861 if (done > ctx->nr_user_files)
7864 ref_node = alloc_fixed_file_ref_node(ctx);
7865 if (IS_ERR(ref_node))
7866 return PTR_ERR(ref_node);
7869 fds = u64_to_user_ptr(up->fds);
7871 struct fixed_file_table *table;
7875 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
7879 i = array_index_nospec(up->offset, ctx->nr_user_files);
7880 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7881 index = i & IORING_FILE_TABLE_MASK;
7882 if (table->files[index]) {
7883 file = table->files[index];
7884 err = io_queue_file_removal(data, file);
7887 table->files[index] = NULL;
7888 needs_switch = true;
7897 * Don't allow io_uring instances to be registered. If
7898 * UNIX isn't enabled, then this causes a reference
7899 * cycle and this instance can never get freed. If UNIX
7900 * is enabled we'll handle it just fine, but there's
7901 * still no point in allowing a ring fd as it doesn't
7902 * support regular read/write anyway.
7904 if (file->f_op == &io_uring_fops) {
7909 table->files[index] = file;
7910 err = io_sqe_file_register(ctx, file, i);
7912 table->files[index] = NULL;
7923 percpu_ref_kill(&data->node->refs);
7924 spin_lock_bh(&data->lock);
7925 list_add_tail(&ref_node->node, &data->ref_list);
7926 data->node = ref_node;
7927 spin_unlock_bh(&data->lock);
7928 percpu_ref_get(&ctx->file_data->refs);
7930 destroy_fixed_file_ref_node(ref_node);
7932 return done ? done : err;
7935 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
7938 struct io_uring_files_update up;
7940 if (!ctx->file_data)
7944 if (copy_from_user(&up, arg, sizeof(up)))
7949 return __io_sqe_files_update(ctx, &up, nr_args);
7952 static void io_free_work(struct io_wq_work *work)
7954 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7956 /* Consider that io_steal_work() relies on this ref */
7960 static int io_init_wq_offload(struct io_ring_ctx *ctx,
7961 struct io_uring_params *p)
7963 struct io_wq_data data;
7965 struct io_ring_ctx *ctx_attach;
7966 unsigned int concurrency;
7969 data.user = ctx->user;
7970 data.free_work = io_free_work;
7971 data.do_work = io_wq_submit_work;
7973 if (!(p->flags & IORING_SETUP_ATTACH_WQ)) {
7974 /* Do QD, or 4 * CPUS, whatever is smallest */
7975 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7977 ctx->io_wq = io_wq_create(concurrency, &data);
7978 if (IS_ERR(ctx->io_wq)) {
7979 ret = PTR_ERR(ctx->io_wq);
7985 f = fdget(p->wq_fd);
7989 if (f.file->f_op != &io_uring_fops) {
7994 ctx_attach = f.file->private_data;
7995 /* @io_wq is protected by holding the fd */
7996 if (!io_wq_get(ctx_attach->io_wq, &data)) {
8001 ctx->io_wq = ctx_attach->io_wq;
8007 static int io_uring_alloc_task_context(struct task_struct *task)
8009 struct io_uring_task *tctx;
8012 tctx = kmalloc(sizeof(*tctx), GFP_KERNEL);
8013 if (unlikely(!tctx))
8016 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
8017 if (unlikely(ret)) {
8023 init_waitqueue_head(&tctx->wait);
8025 atomic_set(&tctx->in_idle, 0);
8026 tctx->sqpoll = false;
8027 io_init_identity(&tctx->__identity);
8028 tctx->identity = &tctx->__identity;
8029 task->io_uring = tctx;
8033 void __io_uring_free(struct task_struct *tsk)
8035 struct io_uring_task *tctx = tsk->io_uring;
8037 WARN_ON_ONCE(!xa_empty(&tctx->xa));
8038 WARN_ON_ONCE(refcount_read(&tctx->identity->count) != 1);
8039 if (tctx->identity != &tctx->__identity)
8040 kfree(tctx->identity);
8041 percpu_counter_destroy(&tctx->inflight);
8043 tsk->io_uring = NULL;
8046 static int io_sq_offload_create(struct io_ring_ctx *ctx,
8047 struct io_uring_params *p)
8051 if (ctx->flags & IORING_SETUP_SQPOLL) {
8052 struct io_sq_data *sqd;
8055 if (!capable(CAP_SYS_ADMIN) && !capable(CAP_SYS_NICE))
8058 sqd = io_get_sq_data(p);
8065 io_sq_thread_park(sqd);
8066 mutex_lock(&sqd->ctx_lock);
8067 list_add(&ctx->sqd_list, &sqd->ctx_new_list);
8068 mutex_unlock(&sqd->ctx_lock);
8069 io_sq_thread_unpark(sqd);
8071 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
8072 if (!ctx->sq_thread_idle)
8073 ctx->sq_thread_idle = HZ;
8078 if (p->flags & IORING_SETUP_SQ_AFF) {
8079 int cpu = p->sq_thread_cpu;
8082 if (cpu >= nr_cpu_ids)
8084 if (!cpu_online(cpu))
8087 sqd->thread = kthread_create_on_cpu(io_sq_thread, sqd,
8088 cpu, "io_uring-sq");
8090 sqd->thread = kthread_create(io_sq_thread, sqd,
8093 if (IS_ERR(sqd->thread)) {
8094 ret = PTR_ERR(sqd->thread);
8098 ret = io_uring_alloc_task_context(sqd->thread);
8101 } else if (p->flags & IORING_SETUP_SQ_AFF) {
8102 /* Can't have SQ_AFF without SQPOLL */
8108 ret = io_init_wq_offload(ctx, p);
8114 io_finish_async(ctx);
8118 static void io_sq_offload_start(struct io_ring_ctx *ctx)
8120 struct io_sq_data *sqd = ctx->sq_data;
8122 if ((ctx->flags & IORING_SETUP_SQPOLL) && sqd->thread)
8123 wake_up_process(sqd->thread);
8126 static inline void __io_unaccount_mem(struct user_struct *user,
8127 unsigned long nr_pages)
8129 atomic_long_sub(nr_pages, &user->locked_vm);
8132 static inline int __io_account_mem(struct user_struct *user,
8133 unsigned long nr_pages)
8135 unsigned long page_limit, cur_pages, new_pages;
8137 /* Don't allow more pages than we can safely lock */
8138 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
8141 cur_pages = atomic_long_read(&user->locked_vm);
8142 new_pages = cur_pages + nr_pages;
8143 if (new_pages > page_limit)
8145 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
8146 new_pages) != cur_pages);
8151 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
8152 enum io_mem_account acct)
8155 __io_unaccount_mem(ctx->user, nr_pages);
8157 if (ctx->mm_account) {
8158 if (acct == ACCT_LOCKED)
8159 ctx->mm_account->locked_vm -= nr_pages;
8160 else if (acct == ACCT_PINNED)
8161 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
8165 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
8166 enum io_mem_account acct)
8170 if (ctx->limit_mem) {
8171 ret = __io_account_mem(ctx->user, nr_pages);
8176 if (ctx->mm_account) {
8177 if (acct == ACCT_LOCKED)
8178 ctx->mm_account->locked_vm += nr_pages;
8179 else if (acct == ACCT_PINNED)
8180 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
8186 static void io_mem_free(void *ptr)
8193 page = virt_to_head_page(ptr);
8194 if (put_page_testzero(page))
8195 free_compound_page(page);
8198 static void *io_mem_alloc(size_t size)
8200 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
8203 return (void *) __get_free_pages(gfp_flags, get_order(size));
8206 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8209 struct io_rings *rings;
8210 size_t off, sq_array_size;
8212 off = struct_size(rings, cqes, cq_entries);
8213 if (off == SIZE_MAX)
8217 off = ALIGN(off, SMP_CACHE_BYTES);
8225 sq_array_size = array_size(sizeof(u32), sq_entries);
8226 if (sq_array_size == SIZE_MAX)
8229 if (check_add_overflow(off, sq_array_size, &off))
8235 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
8239 pages = (size_t)1 << get_order(
8240 rings_size(sq_entries, cq_entries, NULL));
8241 pages += (size_t)1 << get_order(
8242 array_size(sizeof(struct io_uring_sqe), sq_entries));
8247 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
8251 if (!ctx->user_bufs)
8254 for (i = 0; i < ctx->nr_user_bufs; i++) {
8255 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8257 for (j = 0; j < imu->nr_bvecs; j++)
8258 unpin_user_page(imu->bvec[j].bv_page);
8260 if (imu->acct_pages)
8261 io_unaccount_mem(ctx, imu->acct_pages, ACCT_PINNED);
8266 kfree(ctx->user_bufs);
8267 ctx->user_bufs = NULL;
8268 ctx->nr_user_bufs = 0;
8272 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8273 void __user *arg, unsigned index)
8275 struct iovec __user *src;
8277 #ifdef CONFIG_COMPAT
8279 struct compat_iovec __user *ciovs;
8280 struct compat_iovec ciov;
8282 ciovs = (struct compat_iovec __user *) arg;
8283 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8286 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8287 dst->iov_len = ciov.iov_len;
8291 src = (struct iovec __user *) arg;
8292 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8298 * Not super efficient, but this is just a registration time. And we do cache
8299 * the last compound head, so generally we'll only do a full search if we don't
8302 * We check if the given compound head page has already been accounted, to
8303 * avoid double accounting it. This allows us to account the full size of the
8304 * page, not just the constituent pages of a huge page.
8306 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8307 int nr_pages, struct page *hpage)
8311 /* check current page array */
8312 for (i = 0; i < nr_pages; i++) {
8313 if (!PageCompound(pages[i]))
8315 if (compound_head(pages[i]) == hpage)
8319 /* check previously registered pages */
8320 for (i = 0; i < ctx->nr_user_bufs; i++) {
8321 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8323 for (j = 0; j < imu->nr_bvecs; j++) {
8324 if (!PageCompound(imu->bvec[j].bv_page))
8326 if (compound_head(imu->bvec[j].bv_page) == hpage)
8334 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8335 int nr_pages, struct io_mapped_ubuf *imu,
8336 struct page **last_hpage)
8340 for (i = 0; i < nr_pages; i++) {
8341 if (!PageCompound(pages[i])) {
8346 hpage = compound_head(pages[i]);
8347 if (hpage == *last_hpage)
8349 *last_hpage = hpage;
8350 if (headpage_already_acct(ctx, pages, i, hpage))
8352 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8356 if (!imu->acct_pages)
8359 ret = io_account_mem(ctx, imu->acct_pages, ACCT_PINNED);
8361 imu->acct_pages = 0;
8365 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
8368 struct vm_area_struct **vmas = NULL;
8369 struct page **pages = NULL;
8370 struct page *last_hpage = NULL;
8371 int i, j, got_pages = 0;
8376 if (!nr_args || nr_args > UIO_MAXIOV)
8379 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
8381 if (!ctx->user_bufs)
8384 for (i = 0; i < nr_args; i++) {
8385 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8386 unsigned long off, start, end, ubuf;
8391 ret = io_copy_iov(ctx, &iov, arg, i);
8396 * Don't impose further limits on the size and buffer
8397 * constraints here, we'll -EINVAL later when IO is
8398 * submitted if they are wrong.
8401 if (!iov.iov_base || !iov.iov_len)
8404 /* arbitrary limit, but we need something */
8405 if (iov.iov_len > SZ_1G)
8408 ubuf = (unsigned long) iov.iov_base;
8409 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8410 start = ubuf >> PAGE_SHIFT;
8411 nr_pages = end - start;
8414 if (!pages || nr_pages > got_pages) {
8417 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
8419 vmas = kvmalloc_array(nr_pages,
8420 sizeof(struct vm_area_struct *),
8422 if (!pages || !vmas) {
8426 got_pages = nr_pages;
8429 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
8436 mmap_read_lock(current->mm);
8437 pret = pin_user_pages(ubuf, nr_pages,
8438 FOLL_WRITE | FOLL_LONGTERM,
8440 if (pret == nr_pages) {
8441 /* don't support file backed memory */
8442 for (j = 0; j < nr_pages; j++) {
8443 struct vm_area_struct *vma = vmas[j];
8446 !is_file_hugepages(vma->vm_file)) {
8452 ret = pret < 0 ? pret : -EFAULT;
8454 mmap_read_unlock(current->mm);
8457 * if we did partial map, or found file backed vmas,
8458 * release any pages we did get
8461 unpin_user_pages(pages, pret);
8466 ret = io_buffer_account_pin(ctx, pages, pret, imu, &last_hpage);
8468 unpin_user_pages(pages, pret);
8473 off = ubuf & ~PAGE_MASK;
8475 for (j = 0; j < nr_pages; j++) {
8478 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8479 imu->bvec[j].bv_page = pages[j];
8480 imu->bvec[j].bv_len = vec_len;
8481 imu->bvec[j].bv_offset = off;
8485 /* store original address for later verification */
8487 imu->len = iov.iov_len;
8488 imu->nr_bvecs = nr_pages;
8490 ctx->nr_user_bufs++;
8498 io_sqe_buffer_unregister(ctx);
8502 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8504 __s32 __user *fds = arg;
8510 if (copy_from_user(&fd, fds, sizeof(*fds)))
8513 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8514 if (IS_ERR(ctx->cq_ev_fd)) {
8515 int ret = PTR_ERR(ctx->cq_ev_fd);
8516 ctx->cq_ev_fd = NULL;
8523 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8525 if (ctx->cq_ev_fd) {
8526 eventfd_ctx_put(ctx->cq_ev_fd);
8527 ctx->cq_ev_fd = NULL;
8534 static int __io_destroy_buffers(int id, void *p, void *data)
8536 struct io_ring_ctx *ctx = data;
8537 struct io_buffer *buf = p;
8539 __io_remove_buffers(ctx, buf, id, -1U);
8543 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8545 idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx);
8546 idr_destroy(&ctx->io_buffer_idr);
8549 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8551 io_finish_async(ctx);
8552 io_sqe_buffer_unregister(ctx);
8554 if (ctx->sqo_task) {
8555 put_task_struct(ctx->sqo_task);
8556 ctx->sqo_task = NULL;
8557 mmdrop(ctx->mm_account);
8558 ctx->mm_account = NULL;
8561 #ifdef CONFIG_BLK_CGROUP
8562 if (ctx->sqo_blkcg_css)
8563 css_put(ctx->sqo_blkcg_css);
8566 io_sqe_files_unregister(ctx);
8567 io_eventfd_unregister(ctx);
8568 io_destroy_buffers(ctx);
8569 idr_destroy(&ctx->personality_idr);
8571 #if defined(CONFIG_UNIX)
8572 if (ctx->ring_sock) {
8573 ctx->ring_sock->file = NULL; /* so that iput() is called */
8574 sock_release(ctx->ring_sock);
8578 io_mem_free(ctx->rings);
8579 io_mem_free(ctx->sq_sqes);
8581 percpu_ref_exit(&ctx->refs);
8582 free_uid(ctx->user);
8583 put_cred(ctx->creds);
8584 kfree(ctx->cancel_hash);
8585 kmem_cache_free(req_cachep, ctx->fallback_req);
8589 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8591 struct io_ring_ctx *ctx = file->private_data;
8594 poll_wait(file, &ctx->cq_wait, wait);
8596 * synchronizes with barrier from wq_has_sleeper call in
8600 if (!io_sqring_full(ctx))
8601 mask |= EPOLLOUT | EPOLLWRNORM;
8602 if (io_cqring_events(ctx, false))
8603 mask |= EPOLLIN | EPOLLRDNORM;
8608 static int io_uring_fasync(int fd, struct file *file, int on)
8610 struct io_ring_ctx *ctx = file->private_data;
8612 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8615 static int io_remove_personalities(int id, void *p, void *data)
8617 struct io_ring_ctx *ctx = data;
8618 struct io_identity *iod;
8620 iod = idr_remove(&ctx->personality_idr, id);
8622 put_cred(iod->creds);
8623 if (refcount_dec_and_test(&iod->count))
8629 static void io_ring_exit_work(struct work_struct *work)
8631 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
8635 * If we're doing polled IO and end up having requests being
8636 * submitted async (out-of-line), then completions can come in while
8637 * we're waiting for refs to drop. We need to reap these manually,
8638 * as nobody else will be looking for them.
8641 io_iopoll_try_reap_events(ctx);
8642 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
8643 io_ring_ctx_free(ctx);
8646 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8648 mutex_lock(&ctx->uring_lock);
8649 percpu_ref_kill(&ctx->refs);
8651 io_cqring_overflow_flush(ctx, true, NULL, NULL);
8652 mutex_unlock(&ctx->uring_lock);
8654 io_kill_timeouts(ctx, NULL, NULL);
8655 io_poll_remove_all(ctx, NULL, NULL);
8658 io_wq_cancel_all(ctx->io_wq);
8660 /* if we failed setting up the ctx, we might not have any rings */
8661 io_iopoll_try_reap_events(ctx);
8662 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
8665 * Do this upfront, so we won't have a grace period where the ring
8666 * is closed but resources aren't reaped yet. This can cause
8667 * spurious failure in setting up a new ring.
8669 io_unaccount_mem(ctx, ring_pages(ctx->sq_entries, ctx->cq_entries),
8672 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8674 * Use system_unbound_wq to avoid spawning tons of event kworkers
8675 * if we're exiting a ton of rings at the same time. It just adds
8676 * noise and overhead, there's no discernable change in runtime
8677 * over using system_wq.
8679 queue_work(system_unbound_wq, &ctx->exit_work);
8682 static int io_uring_release(struct inode *inode, struct file *file)
8684 struct io_ring_ctx *ctx = file->private_data;
8686 file->private_data = NULL;
8687 io_ring_ctx_wait_and_kill(ctx);
8691 struct io_task_cancel {
8692 struct task_struct *task;
8693 struct files_struct *files;
8696 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8698 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8699 struct io_task_cancel *cancel = data;
8702 if (cancel->files && (req->flags & REQ_F_LINK_TIMEOUT)) {
8703 unsigned long flags;
8704 struct io_ring_ctx *ctx = req->ctx;
8706 /* protect against races with linked timeouts */
8707 spin_lock_irqsave(&ctx->completion_lock, flags);
8708 ret = io_match_task(req, cancel->task, cancel->files);
8709 spin_unlock_irqrestore(&ctx->completion_lock, flags);
8711 ret = io_match_task(req, cancel->task, cancel->files);
8716 static void io_cancel_defer_files(struct io_ring_ctx *ctx,
8717 struct task_struct *task,
8718 struct files_struct *files)
8720 struct io_defer_entry *de = NULL;
8723 spin_lock_irq(&ctx->completion_lock);
8724 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8725 if (io_match_task(de->req, task, files)) {
8726 list_cut_position(&list, &ctx->defer_list, &de->list);
8730 spin_unlock_irq(&ctx->completion_lock);
8732 while (!list_empty(&list)) {
8733 de = list_first_entry(&list, struct io_defer_entry, list);
8734 list_del_init(&de->list);
8735 req_set_fail_links(de->req);
8736 io_put_req(de->req);
8737 io_req_complete(de->req, -ECANCELED);
8742 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
8743 struct task_struct *task,
8744 struct files_struct *files)
8746 while (!list_empty_careful(&ctx->inflight_list)) {
8747 struct io_task_cancel cancel = { .task = task, .files = files };
8748 struct io_kiocb *req;
8752 spin_lock_irq(&ctx->inflight_lock);
8753 list_for_each_entry(req, &ctx->inflight_list, inflight_entry) {
8754 if (req->task != task ||
8755 req->work.identity->files != files)
8761 prepare_to_wait(&task->io_uring->wait, &wait,
8762 TASK_UNINTERRUPTIBLE);
8763 spin_unlock_irq(&ctx->inflight_lock);
8765 /* We need to keep going until we don't find a matching req */
8769 io_wq_cancel_cb(ctx->io_wq, io_cancel_task_cb, &cancel, true);
8770 io_poll_remove_all(ctx, task, files);
8771 io_kill_timeouts(ctx, task, files);
8772 /* cancellations _may_ trigger task work */
8775 finish_wait(&task->io_uring->wait, &wait);
8779 static void __io_uring_cancel_task_requests(struct io_ring_ctx *ctx,
8780 struct task_struct *task)
8783 struct io_task_cancel cancel = { .task = task, .files = NULL, };
8784 enum io_wq_cancel cret;
8787 cret = io_wq_cancel_cb(ctx->io_wq, io_cancel_task_cb, &cancel, true);
8788 if (cret != IO_WQ_CANCEL_NOTFOUND)
8791 /* SQPOLL thread does its own polling */
8792 if (!(ctx->flags & IORING_SETUP_SQPOLL)) {
8793 while (!list_empty_careful(&ctx->iopoll_list)) {
8794 io_iopoll_try_reap_events(ctx);
8799 ret |= io_poll_remove_all(ctx, task, NULL);
8800 ret |= io_kill_timeouts(ctx, task, NULL);
8809 * We need to iteratively cancel requests, in case a request has dependent
8810 * hard links. These persist even for failure of cancelations, hence keep
8811 * looping until none are found.
8813 static void io_uring_cancel_task_requests(struct io_ring_ctx *ctx,
8814 struct files_struct *files)
8816 struct task_struct *task = current;
8818 if ((ctx->flags & IORING_SETUP_SQPOLL) && ctx->sq_data) {
8819 task = ctx->sq_data->thread;
8820 atomic_inc(&task->io_uring->in_idle);
8821 io_sq_thread_park(ctx->sq_data);
8824 io_cancel_defer_files(ctx, task, files);
8825 io_ring_submit_lock(ctx, (ctx->flags & IORING_SETUP_IOPOLL));
8826 io_cqring_overflow_flush(ctx, true, task, files);
8827 io_ring_submit_unlock(ctx, (ctx->flags & IORING_SETUP_IOPOLL));
8830 __io_uring_cancel_task_requests(ctx, task);
8832 io_uring_cancel_files(ctx, task, files);
8834 if ((ctx->flags & IORING_SETUP_SQPOLL) && ctx->sq_data) {
8835 atomic_dec(&task->io_uring->in_idle);
8837 * If the files that are going away are the ones in the thread
8838 * identity, clear them out.
8840 if (task->io_uring->identity->files == files)
8841 task->io_uring->identity->files = NULL;
8842 io_sq_thread_unpark(ctx->sq_data);
8847 * Note that this task has used io_uring. We use it for cancelation purposes.
8849 static int io_uring_add_task_file(struct io_ring_ctx *ctx, struct file *file)
8851 struct io_uring_task *tctx = current->io_uring;
8853 if (unlikely(!tctx)) {
8856 ret = io_uring_alloc_task_context(current);
8859 tctx = current->io_uring;
8861 if (tctx->last != file) {
8862 void *old = xa_load(&tctx->xa, (unsigned long)file);
8866 xa_store(&tctx->xa, (unsigned long)file, file, GFP_KERNEL);
8872 * This is race safe in that the task itself is doing this, hence it
8873 * cannot be going through the exit/cancel paths at the same time.
8874 * This cannot be modified while exit/cancel is running.
8876 if (!tctx->sqpoll && (ctx->flags & IORING_SETUP_SQPOLL))
8877 tctx->sqpoll = true;
8883 * Remove this io_uring_file -> task mapping.
8885 static void io_uring_del_task_file(struct file *file)
8887 struct io_uring_task *tctx = current->io_uring;
8889 if (tctx->last == file)
8891 file = xa_erase(&tctx->xa, (unsigned long)file);
8897 * Drop task note for this file if we're the only ones that hold it after
8900 static void io_uring_attempt_task_drop(struct file *file)
8902 if (!current->io_uring)
8905 * fput() is pending, will be 2 if the only other ref is our potential
8906 * task file note. If the task is exiting, drop regardless of count.
8908 if (fatal_signal_pending(current) || (current->flags & PF_EXITING) ||
8909 atomic_long_read(&file->f_count) == 2)
8910 io_uring_del_task_file(file);
8913 void __io_uring_files_cancel(struct files_struct *files)
8915 struct io_uring_task *tctx = current->io_uring;
8917 unsigned long index;
8919 /* make sure overflow events are dropped */
8920 atomic_inc(&tctx->in_idle);
8922 xa_for_each(&tctx->xa, index, file) {
8923 struct io_ring_ctx *ctx = file->private_data;
8925 io_uring_cancel_task_requests(ctx, files);
8927 io_uring_del_task_file(file);
8930 atomic_dec(&tctx->in_idle);
8933 static s64 tctx_inflight(struct io_uring_task *tctx)
8935 unsigned long index;
8939 inflight = percpu_counter_sum(&tctx->inflight);
8944 * If we have SQPOLL rings, then we need to iterate and find them, and
8945 * add the pending count for those.
8947 xa_for_each(&tctx->xa, index, file) {
8948 struct io_ring_ctx *ctx = file->private_data;
8950 if (ctx->flags & IORING_SETUP_SQPOLL) {
8951 struct io_uring_task *__tctx = ctx->sqo_task->io_uring;
8953 inflight += percpu_counter_sum(&__tctx->inflight);
8961 * Find any io_uring fd that this task has registered or done IO on, and cancel
8964 void __io_uring_task_cancel(void)
8966 struct io_uring_task *tctx = current->io_uring;
8970 /* make sure overflow events are dropped */
8971 atomic_inc(&tctx->in_idle);
8974 /* read completions before cancelations */
8975 inflight = tctx_inflight(tctx);
8978 __io_uring_files_cancel(NULL);
8980 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
8983 * If we've seen completions, retry. This avoids a race where
8984 * a completion comes in before we did prepare_to_wait().
8986 if (inflight != tctx_inflight(tctx))
8991 finish_wait(&tctx->wait, &wait);
8992 atomic_dec(&tctx->in_idle);
8995 static int io_uring_flush(struct file *file, void *data)
8997 io_uring_attempt_task_drop(file);
9001 static void *io_uring_validate_mmap_request(struct file *file,
9002 loff_t pgoff, size_t sz)
9004 struct io_ring_ctx *ctx = file->private_data;
9005 loff_t offset = pgoff << PAGE_SHIFT;
9010 case IORING_OFF_SQ_RING:
9011 case IORING_OFF_CQ_RING:
9014 case IORING_OFF_SQES:
9018 return ERR_PTR(-EINVAL);
9021 page = virt_to_head_page(ptr);
9022 if (sz > page_size(page))
9023 return ERR_PTR(-EINVAL);
9030 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9032 size_t sz = vma->vm_end - vma->vm_start;
9036 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
9038 return PTR_ERR(ptr);
9040 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
9041 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
9044 #else /* !CONFIG_MMU */
9046 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9048 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
9051 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
9053 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
9056 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
9057 unsigned long addr, unsigned long len,
9058 unsigned long pgoff, unsigned long flags)
9062 ptr = io_uring_validate_mmap_request(file, pgoff, len);
9064 return PTR_ERR(ptr);
9066 return (unsigned long) ptr;
9069 #endif /* !CONFIG_MMU */
9071 static void io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
9076 if (!io_sqring_full(ctx))
9079 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9081 if (!io_sqring_full(ctx))
9085 } while (!signal_pending(current));
9087 finish_wait(&ctx->sqo_sq_wait, &wait);
9090 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9091 struct __kernel_timespec __user **ts,
9092 const sigset_t __user **sig)
9094 struct io_uring_getevents_arg arg;
9097 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9098 * is just a pointer to the sigset_t.
9100 if (!(flags & IORING_ENTER_EXT_ARG)) {
9101 *sig = (const sigset_t __user *) argp;
9107 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9108 * timespec and sigset_t pointers if good.
9110 if (*argsz != sizeof(arg))
9112 if (copy_from_user(&arg, argp, sizeof(arg)))
9114 *sig = u64_to_user_ptr(arg.sigmask);
9115 *argsz = arg.sigmask_sz;
9116 *ts = u64_to_user_ptr(arg.ts);
9120 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9121 u32, min_complete, u32, flags, const void __user *, argp,
9124 struct io_ring_ctx *ctx;
9131 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9132 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG))
9140 if (f.file->f_op != &io_uring_fops)
9144 ctx = f.file->private_data;
9145 if (!percpu_ref_tryget(&ctx->refs))
9149 if (ctx->flags & IORING_SETUP_R_DISABLED)
9153 * For SQ polling, the thread will do all submissions and completions.
9154 * Just return the requested submit count, and wake the thread if
9158 if (ctx->flags & IORING_SETUP_SQPOLL) {
9159 io_ring_submit_lock(ctx, (ctx->flags & IORING_SETUP_IOPOLL));
9160 if (!list_empty_careful(&ctx->cq_overflow_list))
9161 io_cqring_overflow_flush(ctx, false, NULL, NULL);
9162 io_ring_submit_unlock(ctx, (ctx->flags & IORING_SETUP_IOPOLL));
9163 if (flags & IORING_ENTER_SQ_WAKEUP)
9164 wake_up(&ctx->sq_data->wait);
9165 if (flags & IORING_ENTER_SQ_WAIT)
9166 io_sqpoll_wait_sq(ctx);
9167 submitted = to_submit;
9168 } else if (to_submit) {
9169 ret = io_uring_add_task_file(ctx, f.file);
9172 mutex_lock(&ctx->uring_lock);
9173 submitted = io_submit_sqes(ctx, to_submit);
9174 mutex_unlock(&ctx->uring_lock);
9176 if (submitted != to_submit)
9179 if (flags & IORING_ENTER_GETEVENTS) {
9180 const sigset_t __user *sig;
9181 struct __kernel_timespec __user *ts;
9183 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
9187 min_complete = min(min_complete, ctx->cq_entries);
9190 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9191 * space applications don't need to do io completion events
9192 * polling again, they can rely on io_sq_thread to do polling
9193 * work, which can reduce cpu usage and uring_lock contention.
9195 if (ctx->flags & IORING_SETUP_IOPOLL &&
9196 !(ctx->flags & IORING_SETUP_SQPOLL)) {
9197 ret = io_iopoll_check(ctx, min_complete);
9199 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
9204 percpu_ref_put(&ctx->refs);
9207 return submitted ? submitted : ret;
9210 #ifdef CONFIG_PROC_FS
9211 static int io_uring_show_cred(int id, void *p, void *data)
9213 struct io_identity *iod = p;
9214 const struct cred *cred = iod->creds;
9215 struct seq_file *m = data;
9216 struct user_namespace *uns = seq_user_ns(m);
9217 struct group_info *gi;
9222 seq_printf(m, "%5d\n", id);
9223 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
9224 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
9225 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
9226 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
9227 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
9228 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
9229 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
9230 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
9231 seq_puts(m, "\n\tGroups:\t");
9232 gi = cred->group_info;
9233 for (g = 0; g < gi->ngroups; g++) {
9234 seq_put_decimal_ull(m, g ? " " : "",
9235 from_kgid_munged(uns, gi->gid[g]));
9237 seq_puts(m, "\n\tCapEff:\t");
9238 cap = cred->cap_effective;
9239 CAP_FOR_EACH_U32(__capi)
9240 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
9245 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
9247 struct io_sq_data *sq = NULL;
9252 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9253 * since fdinfo case grabs it in the opposite direction of normal use
9254 * cases. If we fail to get the lock, we just don't iterate any
9255 * structures that could be going away outside the io_uring mutex.
9257 has_lock = mutex_trylock(&ctx->uring_lock);
9259 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL))
9262 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9263 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9264 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9265 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9266 struct fixed_file_table *table;
9269 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
9270 f = table->files[i & IORING_FILE_TABLE_MASK];
9272 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9274 seq_printf(m, "%5u: <none>\n", i);
9276 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9277 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9278 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
9280 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
9281 (unsigned int) buf->len);
9283 if (has_lock && !idr_is_empty(&ctx->personality_idr)) {
9284 seq_printf(m, "Personalities:\n");
9285 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
9287 seq_printf(m, "PollList:\n");
9288 spin_lock_irq(&ctx->completion_lock);
9289 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9290 struct hlist_head *list = &ctx->cancel_hash[i];
9291 struct io_kiocb *req;
9293 hlist_for_each_entry(req, list, hash_node)
9294 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9295 req->task->task_works != NULL);
9297 spin_unlock_irq(&ctx->completion_lock);
9299 mutex_unlock(&ctx->uring_lock);
9302 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9304 struct io_ring_ctx *ctx = f->private_data;
9306 if (percpu_ref_tryget(&ctx->refs)) {
9307 __io_uring_show_fdinfo(ctx, m);
9308 percpu_ref_put(&ctx->refs);
9313 static const struct file_operations io_uring_fops = {
9314 .release = io_uring_release,
9315 .flush = io_uring_flush,
9316 .mmap = io_uring_mmap,
9318 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9319 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9321 .poll = io_uring_poll,
9322 .fasync = io_uring_fasync,
9323 #ifdef CONFIG_PROC_FS
9324 .show_fdinfo = io_uring_show_fdinfo,
9328 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9329 struct io_uring_params *p)
9331 struct io_rings *rings;
9332 size_t size, sq_array_offset;
9334 /* make sure these are sane, as we already accounted them */
9335 ctx->sq_entries = p->sq_entries;
9336 ctx->cq_entries = p->cq_entries;
9338 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9339 if (size == SIZE_MAX)
9342 rings = io_mem_alloc(size);
9347 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9348 rings->sq_ring_mask = p->sq_entries - 1;
9349 rings->cq_ring_mask = p->cq_entries - 1;
9350 rings->sq_ring_entries = p->sq_entries;
9351 rings->cq_ring_entries = p->cq_entries;
9352 ctx->sq_mask = rings->sq_ring_mask;
9353 ctx->cq_mask = rings->cq_ring_mask;
9355 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9356 if (size == SIZE_MAX) {
9357 io_mem_free(ctx->rings);
9362 ctx->sq_sqes = io_mem_alloc(size);
9363 if (!ctx->sq_sqes) {
9364 io_mem_free(ctx->rings);
9373 * Allocate an anonymous fd, this is what constitutes the application
9374 * visible backing of an io_uring instance. The application mmaps this
9375 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9376 * we have to tie this fd to a socket for file garbage collection purposes.
9378 static int io_uring_get_fd(struct io_ring_ctx *ctx)
9384 #if defined(CONFIG_UNIX)
9385 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9391 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9396 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9397 O_RDWR | O_CLOEXEC);
9400 ret = PTR_ERR(file);
9404 #if defined(CONFIG_UNIX)
9405 ctx->ring_sock->file = file;
9407 ret = io_uring_add_task_file(ctx, file);
9413 fd_install(fd, file);
9416 #if defined(CONFIG_UNIX)
9417 sock_release(ctx->ring_sock);
9418 ctx->ring_sock = NULL;
9423 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9424 struct io_uring_params __user *params)
9426 struct user_struct *user = NULL;
9427 struct io_ring_ctx *ctx;
9433 if (entries > IORING_MAX_ENTRIES) {
9434 if (!(p->flags & IORING_SETUP_CLAMP))
9436 entries = IORING_MAX_ENTRIES;
9440 * Use twice as many entries for the CQ ring. It's possible for the
9441 * application to drive a higher depth than the size of the SQ ring,
9442 * since the sqes are only used at submission time. This allows for
9443 * some flexibility in overcommitting a bit. If the application has
9444 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9445 * of CQ ring entries manually.
9447 p->sq_entries = roundup_pow_of_two(entries);
9448 if (p->flags & IORING_SETUP_CQSIZE) {
9450 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9451 * to a power-of-two, if it isn't already. We do NOT impose
9452 * any cq vs sq ring sizing.
9456 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9457 if (!(p->flags & IORING_SETUP_CLAMP))
9459 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9461 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9462 if (p->cq_entries < p->sq_entries)
9465 p->cq_entries = 2 * p->sq_entries;
9468 user = get_uid(current_user());
9469 limit_mem = !capable(CAP_IPC_LOCK);
9472 ret = __io_account_mem(user,
9473 ring_pages(p->sq_entries, p->cq_entries));
9480 ctx = io_ring_ctx_alloc(p);
9483 __io_unaccount_mem(user, ring_pages(p->sq_entries,
9488 ctx->compat = in_compat_syscall();
9490 ctx->creds = get_current_cred();
9492 ctx->loginuid = current->loginuid;
9493 ctx->sessionid = current->sessionid;
9495 ctx->sqo_task = get_task_struct(current);
9498 * This is just grabbed for accounting purposes. When a process exits,
9499 * the mm is exited and dropped before the files, hence we need to hang
9500 * on to this mm purely for the purposes of being able to unaccount
9501 * memory (locked/pinned vm). It's not used for anything else.
9503 mmgrab(current->mm);
9504 ctx->mm_account = current->mm;
9506 #ifdef CONFIG_BLK_CGROUP
9508 * The sq thread will belong to the original cgroup it was inited in.
9509 * If the cgroup goes offline (e.g. disabling the io controller), then
9510 * issued bios will be associated with the closest cgroup later in the
9514 ctx->sqo_blkcg_css = blkcg_css();
9515 ret = css_tryget_online(ctx->sqo_blkcg_css);
9518 /* don't init against a dying cgroup, have the user try again */
9519 ctx->sqo_blkcg_css = NULL;
9526 * Account memory _before_ installing the file descriptor. Once
9527 * the descriptor is installed, it can get closed at any time. Also
9528 * do this before hitting the general error path, as ring freeing
9529 * will un-account as well.
9531 io_account_mem(ctx, ring_pages(p->sq_entries, p->cq_entries),
9533 ctx->limit_mem = limit_mem;
9535 ret = io_allocate_scq_urings(ctx, p);
9539 ret = io_sq_offload_create(ctx, p);
9543 if (!(p->flags & IORING_SETUP_R_DISABLED))
9544 io_sq_offload_start(ctx);
9546 memset(&p->sq_off, 0, sizeof(p->sq_off));
9547 p->sq_off.head = offsetof(struct io_rings, sq.head);
9548 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9549 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9550 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9551 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9552 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9553 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9555 memset(&p->cq_off, 0, sizeof(p->cq_off));
9556 p->cq_off.head = offsetof(struct io_rings, cq.head);
9557 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9558 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9559 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9560 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9561 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9562 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9564 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9565 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9566 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9567 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
9568 IORING_FEAT_EXT_ARG;
9570 if (copy_to_user(params, p, sizeof(*p))) {
9576 * Install ring fd as the very last thing, so we don't risk someone
9577 * having closed it before we finish setup
9579 ret = io_uring_get_fd(ctx);
9583 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9586 io_ring_ctx_wait_and_kill(ctx);
9591 * Sets up an aio uring context, and returns the fd. Applications asks for a
9592 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9593 * params structure passed in.
9595 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9597 struct io_uring_params p;
9600 if (copy_from_user(&p, params, sizeof(p)))
9602 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9607 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9608 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9609 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9610 IORING_SETUP_R_DISABLED))
9613 return io_uring_create(entries, &p, params);
9616 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9617 struct io_uring_params __user *, params)
9619 return io_uring_setup(entries, params);
9622 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9624 struct io_uring_probe *p;
9628 size = struct_size(p, ops, nr_args);
9629 if (size == SIZE_MAX)
9631 p = kzalloc(size, GFP_KERNEL);
9636 if (copy_from_user(p, arg, size))
9639 if (memchr_inv(p, 0, size))
9642 p->last_op = IORING_OP_LAST - 1;
9643 if (nr_args > IORING_OP_LAST)
9644 nr_args = IORING_OP_LAST;
9646 for (i = 0; i < nr_args; i++) {
9648 if (!io_op_defs[i].not_supported)
9649 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9654 if (copy_to_user(arg, p, size))
9661 static int io_register_personality(struct io_ring_ctx *ctx)
9663 struct io_identity *id;
9666 id = kmalloc(sizeof(*id), GFP_KERNEL);
9670 io_init_identity(id);
9671 id->creds = get_current_cred();
9673 ret = idr_alloc_cyclic(&ctx->personality_idr, id, 1, USHRT_MAX, GFP_KERNEL);
9675 put_cred(id->creds);
9681 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
9683 struct io_identity *iod;
9685 iod = idr_remove(&ctx->personality_idr, id);
9687 put_cred(iod->creds);
9688 if (refcount_dec_and_test(&iod->count))
9696 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9697 unsigned int nr_args)
9699 struct io_uring_restriction *res;
9703 /* Restrictions allowed only if rings started disabled */
9704 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9707 /* We allow only a single restrictions registration */
9708 if (ctx->restrictions.registered)
9711 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9714 size = array_size(nr_args, sizeof(*res));
9715 if (size == SIZE_MAX)
9718 res = memdup_user(arg, size);
9720 return PTR_ERR(res);
9724 for (i = 0; i < nr_args; i++) {
9725 switch (res[i].opcode) {
9726 case IORING_RESTRICTION_REGISTER_OP:
9727 if (res[i].register_op >= IORING_REGISTER_LAST) {
9732 __set_bit(res[i].register_op,
9733 ctx->restrictions.register_op);
9735 case IORING_RESTRICTION_SQE_OP:
9736 if (res[i].sqe_op >= IORING_OP_LAST) {
9741 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9743 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9744 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9746 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9747 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9756 /* Reset all restrictions if an error happened */
9758 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9760 ctx->restrictions.registered = true;
9766 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9768 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9771 if (ctx->restrictions.registered)
9772 ctx->restricted = 1;
9774 ctx->flags &= ~IORING_SETUP_R_DISABLED;
9776 io_sq_offload_start(ctx);
9781 static bool io_register_op_must_quiesce(int op)
9784 case IORING_UNREGISTER_FILES:
9785 case IORING_REGISTER_FILES_UPDATE:
9786 case IORING_REGISTER_PROBE:
9787 case IORING_REGISTER_PERSONALITY:
9788 case IORING_UNREGISTER_PERSONALITY:
9795 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
9796 void __user *arg, unsigned nr_args)
9797 __releases(ctx->uring_lock)
9798 __acquires(ctx->uring_lock)
9803 * We're inside the ring mutex, if the ref is already dying, then
9804 * someone else killed the ctx or is already going through
9805 * io_uring_register().
9807 if (percpu_ref_is_dying(&ctx->refs))
9810 if (io_register_op_must_quiesce(opcode)) {
9811 percpu_ref_kill(&ctx->refs);
9814 * Drop uring mutex before waiting for references to exit. If
9815 * another thread is currently inside io_uring_enter() it might
9816 * need to grab the uring_lock to make progress. If we hold it
9817 * here across the drain wait, then we can deadlock. It's safe
9818 * to drop the mutex here, since no new references will come in
9819 * after we've killed the percpu ref.
9821 mutex_unlock(&ctx->uring_lock);
9823 ret = wait_for_completion_interruptible(&ctx->ref_comp);
9826 ret = io_run_task_work_sig();
9831 mutex_lock(&ctx->uring_lock);
9834 percpu_ref_resurrect(&ctx->refs);
9839 if (ctx->restricted) {
9840 if (opcode >= IORING_REGISTER_LAST) {
9845 if (!test_bit(opcode, ctx->restrictions.register_op)) {
9852 case IORING_REGISTER_BUFFERS:
9853 ret = io_sqe_buffer_register(ctx, arg, nr_args);
9855 case IORING_UNREGISTER_BUFFERS:
9859 ret = io_sqe_buffer_unregister(ctx);
9861 case IORING_REGISTER_FILES:
9862 ret = io_sqe_files_register(ctx, arg, nr_args);
9864 case IORING_UNREGISTER_FILES:
9868 ret = io_sqe_files_unregister(ctx);
9870 case IORING_REGISTER_FILES_UPDATE:
9871 ret = io_sqe_files_update(ctx, arg, nr_args);
9873 case IORING_REGISTER_EVENTFD:
9874 case IORING_REGISTER_EVENTFD_ASYNC:
9878 ret = io_eventfd_register(ctx, arg);
9881 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
9882 ctx->eventfd_async = 1;
9884 ctx->eventfd_async = 0;
9886 case IORING_UNREGISTER_EVENTFD:
9890 ret = io_eventfd_unregister(ctx);
9892 case IORING_REGISTER_PROBE:
9894 if (!arg || nr_args > 256)
9896 ret = io_probe(ctx, arg, nr_args);
9898 case IORING_REGISTER_PERSONALITY:
9902 ret = io_register_personality(ctx);
9904 case IORING_UNREGISTER_PERSONALITY:
9908 ret = io_unregister_personality(ctx, nr_args);
9910 case IORING_REGISTER_ENABLE_RINGS:
9914 ret = io_register_enable_rings(ctx);
9916 case IORING_REGISTER_RESTRICTIONS:
9917 ret = io_register_restrictions(ctx, arg, nr_args);
9925 if (io_register_op_must_quiesce(opcode)) {
9926 /* bring the ctx back to life */
9927 percpu_ref_reinit(&ctx->refs);
9929 reinit_completion(&ctx->ref_comp);
9934 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
9935 void __user *, arg, unsigned int, nr_args)
9937 struct io_ring_ctx *ctx;
9946 if (f.file->f_op != &io_uring_fops)
9949 ctx = f.file->private_data;
9951 mutex_lock(&ctx->uring_lock);
9952 ret = __io_uring_register(ctx, opcode, arg, nr_args);
9953 mutex_unlock(&ctx->uring_lock);
9954 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
9955 ctx->cq_ev_fd != NULL, ret);
9961 static int __init io_uring_init(void)
9963 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
9964 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
9965 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
9968 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
9969 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
9970 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
9971 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
9972 BUILD_BUG_SQE_ELEM(1, __u8, flags);
9973 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
9974 BUILD_BUG_SQE_ELEM(4, __s32, fd);
9975 BUILD_BUG_SQE_ELEM(8, __u64, off);
9976 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
9977 BUILD_BUG_SQE_ELEM(16, __u64, addr);
9978 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
9979 BUILD_BUG_SQE_ELEM(24, __u32, len);
9980 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
9981 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
9982 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
9983 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
9984 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
9985 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
9986 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
9987 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
9988 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
9989 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
9990 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
9991 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
9992 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
9993 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
9994 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
9995 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
9996 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
9997 BUILD_BUG_SQE_ELEM(42, __u16, personality);
9998 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
10000 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
10001 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
10002 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
10005 __initcall(io_uring_init);