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);
1737 list_for_each_entry_safe(req, tmp, &ctx->cq_overflow_list, compl.list) {
1738 if (!io_match_task(req, tsk, files))
1741 cqe = io_get_cqring(ctx);
1745 list_move(&req->compl.list, &list);
1747 WRITE_ONCE(cqe->user_data, req->user_data);
1748 WRITE_ONCE(cqe->res, req->result);
1749 WRITE_ONCE(cqe->flags, req->compl.cflags);
1751 ctx->cached_cq_overflow++;
1752 WRITE_ONCE(ctx->rings->cq_overflow,
1753 ctx->cached_cq_overflow);
1757 io_commit_cqring(ctx);
1758 io_cqring_mark_overflow(ctx);
1760 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1761 io_cqring_ev_posted(ctx);
1763 while (!list_empty(&list)) {
1764 req = list_first_entry(&list, struct io_kiocb, compl.list);
1765 list_del(&req->compl.list);
1772 static void __io_cqring_fill_event(struct io_kiocb *req, long res, long cflags)
1774 struct io_ring_ctx *ctx = req->ctx;
1775 struct io_uring_cqe *cqe;
1777 trace_io_uring_complete(ctx, req->user_data, res);
1780 * If we can't get a cq entry, userspace overflowed the
1781 * submission (by quite a lot). Increment the overflow count in
1784 cqe = io_get_cqring(ctx);
1786 WRITE_ONCE(cqe->user_data, req->user_data);
1787 WRITE_ONCE(cqe->res, res);
1788 WRITE_ONCE(cqe->flags, cflags);
1789 } else if (ctx->cq_overflow_flushed ||
1790 atomic_read(&req->task->io_uring->in_idle)) {
1792 * If we're in ring overflow flush mode, or in task cancel mode,
1793 * then we cannot store the request for later flushing, we need
1794 * to drop it on the floor.
1796 ctx->cached_cq_overflow++;
1797 WRITE_ONCE(ctx->rings->cq_overflow, ctx->cached_cq_overflow);
1799 if (list_empty(&ctx->cq_overflow_list)) {
1800 set_bit(0, &ctx->sq_check_overflow);
1801 set_bit(0, &ctx->cq_check_overflow);
1802 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
1806 req->compl.cflags = cflags;
1807 refcount_inc(&req->refs);
1808 list_add_tail(&req->compl.list, &ctx->cq_overflow_list);
1812 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1814 __io_cqring_fill_event(req, res, 0);
1817 static void io_cqring_add_event(struct io_kiocb *req, long res, long cflags)
1819 struct io_ring_ctx *ctx = req->ctx;
1820 unsigned long flags;
1822 spin_lock_irqsave(&ctx->completion_lock, flags);
1823 __io_cqring_fill_event(req, res, cflags);
1824 io_commit_cqring(ctx);
1825 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1827 io_cqring_ev_posted(ctx);
1830 static void io_submit_flush_completions(struct io_comp_state *cs)
1832 struct io_ring_ctx *ctx = cs->ctx;
1834 spin_lock_irq(&ctx->completion_lock);
1835 while (!list_empty(&cs->list)) {
1836 struct io_kiocb *req;
1838 req = list_first_entry(&cs->list, struct io_kiocb, compl.list);
1839 list_del(&req->compl.list);
1840 __io_cqring_fill_event(req, req->result, req->compl.cflags);
1843 * io_free_req() doesn't care about completion_lock unless one
1844 * of these flags is set. REQ_F_WORK_INITIALIZED is in the list
1845 * because of a potential deadlock with req->work.fs->lock
1847 if (req->flags & (REQ_F_FAIL_LINK|REQ_F_LINK_TIMEOUT
1848 |REQ_F_WORK_INITIALIZED)) {
1849 spin_unlock_irq(&ctx->completion_lock);
1851 spin_lock_irq(&ctx->completion_lock);
1856 io_commit_cqring(ctx);
1857 spin_unlock_irq(&ctx->completion_lock);
1859 io_cqring_ev_posted(ctx);
1863 static void __io_req_complete(struct io_kiocb *req, long res, unsigned cflags,
1864 struct io_comp_state *cs)
1867 io_cqring_add_event(req, res, cflags);
1872 req->compl.cflags = cflags;
1873 list_add_tail(&req->compl.list, &cs->list);
1875 io_submit_flush_completions(cs);
1879 static void io_req_complete(struct io_kiocb *req, long res)
1881 __io_req_complete(req, res, 0, NULL);
1884 static inline bool io_is_fallback_req(struct io_kiocb *req)
1886 return req == (struct io_kiocb *)
1887 ((unsigned long) req->ctx->fallback_req & ~1UL);
1890 static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx)
1892 struct io_kiocb *req;
1894 req = ctx->fallback_req;
1895 if (!test_and_set_bit_lock(0, (unsigned long *) &ctx->fallback_req))
1901 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx,
1902 struct io_submit_state *state)
1904 if (!state->free_reqs) {
1905 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1909 sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs));
1910 ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs);
1913 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1914 * retry single alloc to be on the safe side.
1916 if (unlikely(ret <= 0)) {
1917 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1918 if (!state->reqs[0])
1922 state->free_reqs = ret;
1926 return state->reqs[state->free_reqs];
1928 return io_get_fallback_req(ctx);
1931 static inline void io_put_file(struct io_kiocb *req, struct file *file,
1938 static void io_dismantle_req(struct io_kiocb *req)
1942 if (req->async_data)
1943 kfree(req->async_data);
1945 io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
1946 if (req->fixed_file_refs)
1947 percpu_ref_put(req->fixed_file_refs);
1948 io_req_clean_work(req);
1951 static void __io_free_req(struct io_kiocb *req)
1953 struct io_uring_task *tctx = req->task->io_uring;
1954 struct io_ring_ctx *ctx = req->ctx;
1956 io_dismantle_req(req);
1958 percpu_counter_dec(&tctx->inflight);
1959 if (atomic_read(&tctx->in_idle))
1960 wake_up(&tctx->wait);
1961 put_task_struct(req->task);
1963 if (likely(!io_is_fallback_req(req)))
1964 kmem_cache_free(req_cachep, req);
1966 clear_bit_unlock(0, (unsigned long *) &ctx->fallback_req);
1967 percpu_ref_put(&ctx->refs);
1970 static inline void io_remove_next_linked(struct io_kiocb *req)
1972 struct io_kiocb *nxt = req->link;
1974 req->link = nxt->link;
1978 static void io_kill_linked_timeout(struct io_kiocb *req)
1980 struct io_ring_ctx *ctx = req->ctx;
1981 struct io_kiocb *link;
1982 bool cancelled = false;
1983 unsigned long flags;
1985 spin_lock_irqsave(&ctx->completion_lock, flags);
1989 * Can happen if a linked timeout fired and link had been like
1990 * req -> link t-out -> link t-out [-> ...]
1992 if (link && (link->flags & REQ_F_LTIMEOUT_ACTIVE)) {
1993 struct io_timeout_data *io = link->async_data;
1996 io_remove_next_linked(req);
1997 link->timeout.head = NULL;
1998 ret = hrtimer_try_to_cancel(&io->timer);
2000 io_cqring_fill_event(link, -ECANCELED);
2001 io_commit_cqring(ctx);
2005 req->flags &= ~REQ_F_LINK_TIMEOUT;
2006 spin_unlock_irqrestore(&ctx->completion_lock, flags);
2009 io_cqring_ev_posted(ctx);
2015 static void io_fail_links(struct io_kiocb *req)
2017 struct io_kiocb *link, *nxt;
2018 struct io_ring_ctx *ctx = req->ctx;
2019 unsigned long flags;
2021 spin_lock_irqsave(&ctx->completion_lock, flags);
2029 trace_io_uring_fail_link(req, link);
2030 io_cqring_fill_event(link, -ECANCELED);
2033 * It's ok to free under spinlock as they're not linked anymore,
2034 * but avoid REQ_F_WORK_INITIALIZED because it may deadlock on
2037 if (link->flags & REQ_F_WORK_INITIALIZED)
2038 io_put_req_deferred(link, 2);
2040 io_double_put_req(link);
2043 io_commit_cqring(ctx);
2044 spin_unlock_irqrestore(&ctx->completion_lock, flags);
2046 io_cqring_ev_posted(ctx);
2049 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
2051 if (req->flags & REQ_F_LINK_TIMEOUT)
2052 io_kill_linked_timeout(req);
2055 * If LINK is set, we have dependent requests in this chain. If we
2056 * didn't fail this request, queue the first one up, moving any other
2057 * dependencies to the next request. In case of failure, fail the rest
2060 if (likely(!(req->flags & REQ_F_FAIL_LINK))) {
2061 struct io_kiocb *nxt = req->link;
2070 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
2072 if (likely(!(req->link) && !(req->flags & REQ_F_LINK_TIMEOUT)))
2074 return __io_req_find_next(req);
2077 static int io_req_task_work_add(struct io_kiocb *req)
2079 struct task_struct *tsk = req->task;
2080 struct io_ring_ctx *ctx = req->ctx;
2081 enum task_work_notify_mode notify;
2084 if (tsk->flags & PF_EXITING)
2088 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
2089 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
2090 * processing task_work. There's no reliable way to tell if TWA_RESUME
2094 if (!(ctx->flags & IORING_SETUP_SQPOLL))
2095 notify = TWA_SIGNAL;
2097 ret = task_work_add(tsk, &req->task_work, notify);
2099 wake_up_process(tsk);
2104 static void __io_req_task_cancel(struct io_kiocb *req, int error)
2106 struct io_ring_ctx *ctx = req->ctx;
2108 spin_lock_irq(&ctx->completion_lock);
2109 io_cqring_fill_event(req, error);
2110 io_commit_cqring(ctx);
2111 spin_unlock_irq(&ctx->completion_lock);
2113 io_cqring_ev_posted(ctx);
2114 req_set_fail_links(req);
2115 io_double_put_req(req);
2118 static void io_req_task_cancel(struct callback_head *cb)
2120 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2121 struct io_ring_ctx *ctx = req->ctx;
2123 __io_req_task_cancel(req, -ECANCELED);
2124 percpu_ref_put(&ctx->refs);
2127 static void __io_req_task_submit(struct io_kiocb *req)
2129 struct io_ring_ctx *ctx = req->ctx;
2131 if (!__io_sq_thread_acquire_mm(ctx) &&
2132 !__io_sq_thread_acquire_files(ctx)) {
2133 mutex_lock(&ctx->uring_lock);
2134 __io_queue_sqe(req, NULL);
2135 mutex_unlock(&ctx->uring_lock);
2137 __io_req_task_cancel(req, -EFAULT);
2141 static void io_req_task_submit(struct callback_head *cb)
2143 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2144 struct io_ring_ctx *ctx = req->ctx;
2146 __io_req_task_submit(req);
2147 percpu_ref_put(&ctx->refs);
2150 static void io_req_task_queue(struct io_kiocb *req)
2154 init_task_work(&req->task_work, io_req_task_submit);
2155 percpu_ref_get(&req->ctx->refs);
2157 ret = io_req_task_work_add(req);
2158 if (unlikely(ret)) {
2159 struct task_struct *tsk;
2161 init_task_work(&req->task_work, io_req_task_cancel);
2162 tsk = io_wq_get_task(req->ctx->io_wq);
2163 task_work_add(tsk, &req->task_work, TWA_NONE);
2164 wake_up_process(tsk);
2168 static inline void io_queue_next(struct io_kiocb *req)
2170 struct io_kiocb *nxt = io_req_find_next(req);
2173 io_req_task_queue(nxt);
2176 static void io_free_req(struct io_kiocb *req)
2183 void *reqs[IO_IOPOLL_BATCH];
2186 struct task_struct *task;
2190 static inline void io_init_req_batch(struct req_batch *rb)
2197 static void __io_req_free_batch_flush(struct io_ring_ctx *ctx,
2198 struct req_batch *rb)
2200 kmem_cache_free_bulk(req_cachep, rb->to_free, rb->reqs);
2201 percpu_ref_put_many(&ctx->refs, rb->to_free);
2205 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2206 struct req_batch *rb)
2209 __io_req_free_batch_flush(ctx, rb);
2211 struct io_uring_task *tctx = rb->task->io_uring;
2213 percpu_counter_sub(&tctx->inflight, rb->task_refs);
2214 put_task_struct_many(rb->task, rb->task_refs);
2219 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req)
2221 if (unlikely(io_is_fallback_req(req))) {
2227 if (req->task != rb->task) {
2229 struct io_uring_task *tctx = rb->task->io_uring;
2231 percpu_counter_sub(&tctx->inflight, rb->task_refs);
2232 put_task_struct_many(rb->task, rb->task_refs);
2234 rb->task = req->task;
2239 io_dismantle_req(req);
2240 rb->reqs[rb->to_free++] = req;
2241 if (unlikely(rb->to_free == ARRAY_SIZE(rb->reqs)))
2242 __io_req_free_batch_flush(req->ctx, rb);
2246 * Drop reference to request, return next in chain (if there is one) if this
2247 * was the last reference to this request.
2249 static struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2251 struct io_kiocb *nxt = NULL;
2253 if (refcount_dec_and_test(&req->refs)) {
2254 nxt = io_req_find_next(req);
2260 static void io_put_req(struct io_kiocb *req)
2262 if (refcount_dec_and_test(&req->refs))
2266 static void io_put_req_deferred_cb(struct callback_head *cb)
2268 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2273 static void io_free_req_deferred(struct io_kiocb *req)
2277 init_task_work(&req->task_work, io_put_req_deferred_cb);
2278 ret = io_req_task_work_add(req);
2279 if (unlikely(ret)) {
2280 struct task_struct *tsk;
2282 tsk = io_wq_get_task(req->ctx->io_wq);
2283 task_work_add(tsk, &req->task_work, TWA_NONE);
2284 wake_up_process(tsk);
2288 static inline void io_put_req_deferred(struct io_kiocb *req, int refs)
2290 if (refcount_sub_and_test(refs, &req->refs))
2291 io_free_req_deferred(req);
2294 static struct io_wq_work *io_steal_work(struct io_kiocb *req)
2296 struct io_kiocb *nxt;
2299 * A ref is owned by io-wq in which context we're. So, if that's the
2300 * last one, it's safe to steal next work. False negatives are Ok,
2301 * it just will be re-punted async in io_put_work()
2303 if (refcount_read(&req->refs) != 1)
2306 nxt = io_req_find_next(req);
2307 return nxt ? &nxt->work : NULL;
2310 static void io_double_put_req(struct io_kiocb *req)
2312 /* drop both submit and complete references */
2313 if (refcount_sub_and_test(2, &req->refs))
2317 static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush)
2319 struct io_rings *rings = ctx->rings;
2321 if (test_bit(0, &ctx->cq_check_overflow)) {
2323 * noflush == true is from the waitqueue handler, just ensure
2324 * we wake up the task, and the next invocation will flush the
2325 * entries. We cannot safely to it from here.
2330 io_cqring_overflow_flush(ctx, false, NULL, NULL);
2333 /* See comment at the top of this file */
2335 return ctx->cached_cq_tail - READ_ONCE(rings->cq.head);
2338 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2340 struct io_rings *rings = ctx->rings;
2342 /* make sure SQ entry isn't read before tail */
2343 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2346 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2348 unsigned int cflags;
2350 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2351 cflags |= IORING_CQE_F_BUFFER;
2352 req->flags &= ~REQ_F_BUFFER_SELECTED;
2357 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2359 struct io_buffer *kbuf;
2361 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2362 return io_put_kbuf(req, kbuf);
2365 static inline bool io_run_task_work(void)
2368 * Not safe to run on exiting task, and the task_work handling will
2369 * not add work to such a task.
2371 if (unlikely(current->flags & PF_EXITING))
2373 if (current->task_works) {
2374 __set_current_state(TASK_RUNNING);
2382 static void io_iopoll_queue(struct list_head *again)
2384 struct io_kiocb *req;
2387 req = list_first_entry(again, struct io_kiocb, inflight_entry);
2388 list_del(&req->inflight_entry);
2389 __io_complete_rw(req, -EAGAIN, 0, NULL);
2390 } while (!list_empty(again));
2394 * Find and free completed poll iocbs
2396 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2397 struct list_head *done)
2399 struct req_batch rb;
2400 struct io_kiocb *req;
2403 /* order with ->result store in io_complete_rw_iopoll() */
2406 io_init_req_batch(&rb);
2407 while (!list_empty(done)) {
2410 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2411 if (READ_ONCE(req->result) == -EAGAIN) {
2413 req->iopoll_completed = 0;
2414 list_move_tail(&req->inflight_entry, &again);
2417 list_del(&req->inflight_entry);
2419 if (req->flags & REQ_F_BUFFER_SELECTED)
2420 cflags = io_put_rw_kbuf(req);
2422 __io_cqring_fill_event(req, req->result, cflags);
2425 if (refcount_dec_and_test(&req->refs))
2426 io_req_free_batch(&rb, req);
2429 io_commit_cqring(ctx);
2430 if (ctx->flags & IORING_SETUP_SQPOLL)
2431 io_cqring_ev_posted(ctx);
2432 io_req_free_batch_finish(ctx, &rb);
2434 if (!list_empty(&again))
2435 io_iopoll_queue(&again);
2438 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2441 struct io_kiocb *req, *tmp;
2447 * Only spin for completions if we don't have multiple devices hanging
2448 * off our complete list, and we're under the requested amount.
2450 spin = !ctx->poll_multi_file && *nr_events < min;
2453 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2454 struct kiocb *kiocb = &req->rw.kiocb;
2457 * Move completed and retryable entries to our local lists.
2458 * If we find a request that requires polling, break out
2459 * and complete those lists first, if we have entries there.
2461 if (READ_ONCE(req->iopoll_completed)) {
2462 list_move_tail(&req->inflight_entry, &done);
2465 if (!list_empty(&done))
2468 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2472 /* iopoll may have completed current req */
2473 if (READ_ONCE(req->iopoll_completed))
2474 list_move_tail(&req->inflight_entry, &done);
2481 if (!list_empty(&done))
2482 io_iopoll_complete(ctx, nr_events, &done);
2488 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2489 * non-spinning poll check - we'll still enter the driver poll loop, but only
2490 * as a non-spinning completion check.
2492 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
2495 while (!list_empty(&ctx->iopoll_list) && !need_resched()) {
2498 ret = io_do_iopoll(ctx, nr_events, min);
2501 if (*nr_events >= min)
2509 * We can't just wait for polled events to come to us, we have to actively
2510 * find and complete them.
2512 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2514 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2517 mutex_lock(&ctx->uring_lock);
2518 while (!list_empty(&ctx->iopoll_list)) {
2519 unsigned int nr_events = 0;
2521 io_do_iopoll(ctx, &nr_events, 0);
2523 /* let it sleep and repeat later if can't complete a request */
2527 * Ensure we allow local-to-the-cpu processing to take place,
2528 * in this case we need to ensure that we reap all events.
2529 * Also let task_work, etc. to progress by releasing the mutex
2531 if (need_resched()) {
2532 mutex_unlock(&ctx->uring_lock);
2534 mutex_lock(&ctx->uring_lock);
2537 mutex_unlock(&ctx->uring_lock);
2540 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2542 unsigned int nr_events = 0;
2543 int iters = 0, ret = 0;
2546 * We disallow the app entering submit/complete with polling, but we
2547 * still need to lock the ring to prevent racing with polled issue
2548 * that got punted to a workqueue.
2550 mutex_lock(&ctx->uring_lock);
2553 * Don't enter poll loop if we already have events pending.
2554 * If we do, we can potentially be spinning for commands that
2555 * already triggered a CQE (eg in error).
2557 if (io_cqring_events(ctx, false))
2561 * If a submit got punted to a workqueue, we can have the
2562 * application entering polling for a command before it gets
2563 * issued. That app will hold the uring_lock for the duration
2564 * of the poll right here, so we need to take a breather every
2565 * now and then to ensure that the issue has a chance to add
2566 * the poll to the issued list. Otherwise we can spin here
2567 * forever, while the workqueue is stuck trying to acquire the
2570 if (!(++iters & 7)) {
2571 mutex_unlock(&ctx->uring_lock);
2573 mutex_lock(&ctx->uring_lock);
2576 ret = io_iopoll_getevents(ctx, &nr_events, min);
2580 } while (min && !nr_events && !need_resched());
2582 mutex_unlock(&ctx->uring_lock);
2586 static void kiocb_end_write(struct io_kiocb *req)
2589 * Tell lockdep we inherited freeze protection from submission
2592 if (req->flags & REQ_F_ISREG) {
2593 struct inode *inode = file_inode(req->file);
2595 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
2597 file_end_write(req->file);
2600 static void io_complete_rw_common(struct kiocb *kiocb, long res,
2601 struct io_comp_state *cs)
2603 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2606 if (kiocb->ki_flags & IOCB_WRITE)
2607 kiocb_end_write(req);
2609 if (res != req->result)
2610 req_set_fail_links(req);
2611 if (req->flags & REQ_F_BUFFER_SELECTED)
2612 cflags = io_put_rw_kbuf(req);
2613 __io_req_complete(req, res, cflags, cs);
2617 static bool io_resubmit_prep(struct io_kiocb *req, int error)
2619 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2620 ssize_t ret = -ECANCELED;
2621 struct iov_iter iter;
2629 switch (req->opcode) {
2630 case IORING_OP_READV:
2631 case IORING_OP_READ_FIXED:
2632 case IORING_OP_READ:
2635 case IORING_OP_WRITEV:
2636 case IORING_OP_WRITE_FIXED:
2637 case IORING_OP_WRITE:
2641 printk_once(KERN_WARNING "io_uring: bad opcode in resubmit %d\n",
2646 if (!req->async_data) {
2647 ret = io_import_iovec(rw, req, &iovec, &iter, false);
2650 ret = io_setup_async_rw(req, iovec, inline_vecs, &iter, false);
2658 req_set_fail_links(req);
2663 static bool io_rw_reissue(struct io_kiocb *req, long res)
2666 umode_t mode = file_inode(req->file)->i_mode;
2669 if (!S_ISBLK(mode) && !S_ISREG(mode))
2671 if ((res != -EAGAIN && res != -EOPNOTSUPP) || io_wq_current_is_worker())
2674 ret = io_sq_thread_acquire_mm_files(req->ctx, req);
2676 if (io_resubmit_prep(req, ret)) {
2677 refcount_inc(&req->refs);
2678 io_queue_async_work(req);
2686 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2687 struct io_comp_state *cs)
2689 if (!io_rw_reissue(req, res))
2690 io_complete_rw_common(&req->rw.kiocb, res, cs);
2693 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2695 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2697 __io_complete_rw(req, res, res2, NULL);
2700 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2702 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2704 if (kiocb->ki_flags & IOCB_WRITE)
2705 kiocb_end_write(req);
2707 if (res != -EAGAIN && res != req->result)
2708 req_set_fail_links(req);
2710 WRITE_ONCE(req->result, res);
2711 /* order with io_poll_complete() checking ->result */
2713 WRITE_ONCE(req->iopoll_completed, 1);
2717 * After the iocb has been issued, it's safe to be found on the poll list.
2718 * Adding the kiocb to the list AFTER submission ensures that we don't
2719 * find it from a io_iopoll_getevents() thread before the issuer is done
2720 * accessing the kiocb cookie.
2722 static void io_iopoll_req_issued(struct io_kiocb *req, bool in_async)
2724 struct io_ring_ctx *ctx = req->ctx;
2727 * Track whether we have multiple files in our lists. This will impact
2728 * how we do polling eventually, not spinning if we're on potentially
2729 * different devices.
2731 if (list_empty(&ctx->iopoll_list)) {
2732 ctx->poll_multi_file = false;
2733 } else if (!ctx->poll_multi_file) {
2734 struct io_kiocb *list_req;
2736 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2738 if (list_req->file != req->file)
2739 ctx->poll_multi_file = true;
2743 * For fast devices, IO may have already completed. If it has, add
2744 * it to the front so we find it first.
2746 if (READ_ONCE(req->iopoll_completed))
2747 list_add(&req->inflight_entry, &ctx->iopoll_list);
2749 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2752 * If IORING_SETUP_SQPOLL is enabled, sqes are either handled in sq thread
2753 * task context or in io worker task context. If current task context is
2754 * sq thread, we don't need to check whether should wake up sq thread.
2756 if (in_async && (ctx->flags & IORING_SETUP_SQPOLL) &&
2757 wq_has_sleeper(&ctx->sq_data->wait))
2758 wake_up(&ctx->sq_data->wait);
2761 static inline void __io_state_file_put(struct io_submit_state *state)
2763 fput_many(state->file, state->file_refs);
2764 state->file_refs = 0;
2767 static inline void io_state_file_put(struct io_submit_state *state)
2769 if (state->file_refs)
2770 __io_state_file_put(state);
2774 * Get as many references to a file as we have IOs left in this submission,
2775 * assuming most submissions are for one file, or at least that each file
2776 * has more than one submission.
2778 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2783 if (state->file_refs) {
2784 if (state->fd == fd) {
2788 __io_state_file_put(state);
2790 state->file = fget_many(fd, state->ios_left);
2791 if (unlikely(!state->file))
2795 state->file_refs = state->ios_left - 1;
2799 static bool io_bdev_nowait(struct block_device *bdev)
2801 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2805 * If we tracked the file through the SCM inflight mechanism, we could support
2806 * any file. For now, just ensure that anything potentially problematic is done
2809 static bool io_file_supports_async(struct file *file, int rw)
2811 umode_t mode = file_inode(file)->i_mode;
2813 if (S_ISBLK(mode)) {
2814 if (IS_ENABLED(CONFIG_BLOCK) &&
2815 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2819 if (S_ISCHR(mode) || S_ISSOCK(mode))
2821 if (S_ISREG(mode)) {
2822 if (IS_ENABLED(CONFIG_BLOCK) &&
2823 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2824 file->f_op != &io_uring_fops)
2829 /* any ->read/write should understand O_NONBLOCK */
2830 if (file->f_flags & O_NONBLOCK)
2833 if (!(file->f_mode & FMODE_NOWAIT))
2837 return file->f_op->read_iter != NULL;
2839 return file->f_op->write_iter != NULL;
2842 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2844 struct io_ring_ctx *ctx = req->ctx;
2845 struct kiocb *kiocb = &req->rw.kiocb;
2849 if (S_ISREG(file_inode(req->file)->i_mode))
2850 req->flags |= REQ_F_ISREG;
2852 kiocb->ki_pos = READ_ONCE(sqe->off);
2853 if (kiocb->ki_pos == -1 && !(req->file->f_mode & FMODE_STREAM)) {
2854 req->flags |= REQ_F_CUR_POS;
2855 kiocb->ki_pos = req->file->f_pos;
2857 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2858 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2859 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2863 ioprio = READ_ONCE(sqe->ioprio);
2865 ret = ioprio_check_cap(ioprio);
2869 kiocb->ki_ioprio = ioprio;
2871 kiocb->ki_ioprio = get_current_ioprio();
2873 /* don't allow async punt if RWF_NOWAIT was requested */
2874 if (kiocb->ki_flags & IOCB_NOWAIT)
2875 req->flags |= REQ_F_NOWAIT;
2877 if (ctx->flags & IORING_SETUP_IOPOLL) {
2878 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2879 !kiocb->ki_filp->f_op->iopoll)
2882 kiocb->ki_flags |= IOCB_HIPRI;
2883 kiocb->ki_complete = io_complete_rw_iopoll;
2884 req->iopoll_completed = 0;
2886 if (kiocb->ki_flags & IOCB_HIPRI)
2888 kiocb->ki_complete = io_complete_rw;
2891 req->rw.addr = READ_ONCE(sqe->addr);
2892 req->rw.len = READ_ONCE(sqe->len);
2893 req->buf_index = READ_ONCE(sqe->buf_index);
2897 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2903 case -ERESTARTNOINTR:
2904 case -ERESTARTNOHAND:
2905 case -ERESTART_RESTARTBLOCK:
2907 * We can't just restart the syscall, since previously
2908 * submitted sqes may already be in progress. Just fail this
2914 kiocb->ki_complete(kiocb, ret, 0);
2918 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2919 struct io_comp_state *cs)
2921 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2922 struct io_async_rw *io = req->async_data;
2924 /* add previously done IO, if any */
2925 if (io && io->bytes_done > 0) {
2927 ret = io->bytes_done;
2929 ret += io->bytes_done;
2932 if (req->flags & REQ_F_CUR_POS)
2933 req->file->f_pos = kiocb->ki_pos;
2934 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2935 __io_complete_rw(req, ret, 0, cs);
2937 io_rw_done(kiocb, ret);
2940 static ssize_t io_import_fixed(struct io_kiocb *req, int rw,
2941 struct iov_iter *iter)
2943 struct io_ring_ctx *ctx = req->ctx;
2944 size_t len = req->rw.len;
2945 struct io_mapped_ubuf *imu;
2946 u16 index, buf_index = req->buf_index;
2950 if (unlikely(buf_index >= ctx->nr_user_bufs))
2952 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2953 imu = &ctx->user_bufs[index];
2954 buf_addr = req->rw.addr;
2957 if (buf_addr + len < buf_addr)
2959 /* not inside the mapped region */
2960 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2964 * May not be a start of buffer, set size appropriately
2965 * and advance us to the beginning.
2967 offset = buf_addr - imu->ubuf;
2968 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2972 * Don't use iov_iter_advance() here, as it's really slow for
2973 * using the latter parts of a big fixed buffer - it iterates
2974 * over each segment manually. We can cheat a bit here, because
2977 * 1) it's a BVEC iter, we set it up
2978 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2979 * first and last bvec
2981 * So just find our index, and adjust the iterator afterwards.
2982 * If the offset is within the first bvec (or the whole first
2983 * bvec, just use iov_iter_advance(). This makes it easier
2984 * since we can just skip the first segment, which may not
2985 * be PAGE_SIZE aligned.
2987 const struct bio_vec *bvec = imu->bvec;
2989 if (offset <= bvec->bv_len) {
2990 iov_iter_advance(iter, offset);
2992 unsigned long seg_skip;
2994 /* skip first vec */
2995 offset -= bvec->bv_len;
2996 seg_skip = 1 + (offset >> PAGE_SHIFT);
2998 iter->bvec = bvec + seg_skip;
2999 iter->nr_segs -= seg_skip;
3000 iter->count -= bvec->bv_len + offset;
3001 iter->iov_offset = offset & ~PAGE_MASK;
3008 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
3011 mutex_unlock(&ctx->uring_lock);
3014 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
3017 * "Normal" inline submissions always hold the uring_lock, since we
3018 * grab it from the system call. Same is true for the SQPOLL offload.
3019 * The only exception is when we've detached the request and issue it
3020 * from an async worker thread, grab the lock for that case.
3023 mutex_lock(&ctx->uring_lock);
3026 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
3027 int bgid, struct io_buffer *kbuf,
3030 struct io_buffer *head;
3032 if (req->flags & REQ_F_BUFFER_SELECTED)
3035 io_ring_submit_lock(req->ctx, needs_lock);
3037 lockdep_assert_held(&req->ctx->uring_lock);
3039 head = idr_find(&req->ctx->io_buffer_idr, bgid);
3041 if (!list_empty(&head->list)) {
3042 kbuf = list_last_entry(&head->list, struct io_buffer,
3044 list_del(&kbuf->list);
3047 idr_remove(&req->ctx->io_buffer_idr, bgid);
3049 if (*len > kbuf->len)
3052 kbuf = ERR_PTR(-ENOBUFS);
3055 io_ring_submit_unlock(req->ctx, needs_lock);
3060 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
3063 struct io_buffer *kbuf;
3066 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3067 bgid = req->buf_index;
3068 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
3071 req->rw.addr = (u64) (unsigned long) kbuf;
3072 req->flags |= REQ_F_BUFFER_SELECTED;
3073 return u64_to_user_ptr(kbuf->addr);
3076 #ifdef CONFIG_COMPAT
3077 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
3080 struct compat_iovec __user *uiov;
3081 compat_ssize_t clen;
3085 uiov = u64_to_user_ptr(req->rw.addr);
3086 if (!access_ok(uiov, sizeof(*uiov)))
3088 if (__get_user(clen, &uiov->iov_len))
3094 buf = io_rw_buffer_select(req, &len, needs_lock);
3096 return PTR_ERR(buf);
3097 iov[0].iov_base = buf;
3098 iov[0].iov_len = (compat_size_t) len;
3103 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3106 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
3110 if (copy_from_user(iov, uiov, sizeof(*uiov)))
3113 len = iov[0].iov_len;
3116 buf = io_rw_buffer_select(req, &len, needs_lock);
3118 return PTR_ERR(buf);
3119 iov[0].iov_base = buf;
3120 iov[0].iov_len = len;
3124 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3127 if (req->flags & REQ_F_BUFFER_SELECTED) {
3128 struct io_buffer *kbuf;
3130 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3131 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
3132 iov[0].iov_len = kbuf->len;
3137 else if (req->rw.len > 1)
3140 #ifdef CONFIG_COMPAT
3141 if (req->ctx->compat)
3142 return io_compat_import(req, iov, needs_lock);
3145 return __io_iov_buffer_select(req, iov, needs_lock);
3148 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
3149 struct iovec **iovec, struct iov_iter *iter,
3152 void __user *buf = u64_to_user_ptr(req->rw.addr);
3153 size_t sqe_len = req->rw.len;
3157 opcode = req->opcode;
3158 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3160 return io_import_fixed(req, rw, iter);
3163 /* buffer index only valid with fixed read/write, or buffer select */
3164 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3167 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3168 if (req->flags & REQ_F_BUFFER_SELECT) {
3169 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3171 return PTR_ERR(buf);
3172 req->rw.len = sqe_len;
3175 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3180 if (req->flags & REQ_F_BUFFER_SELECT) {
3181 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3183 ret = (*iovec)->iov_len;
3184 iov_iter_init(iter, rw, *iovec, 1, ret);
3190 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3194 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3196 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3200 * For files that don't have ->read_iter() and ->write_iter(), handle them
3201 * by looping over ->read() or ->write() manually.
3203 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3205 struct kiocb *kiocb = &req->rw.kiocb;
3206 struct file *file = req->file;
3210 * Don't support polled IO through this interface, and we can't
3211 * support non-blocking either. For the latter, this just causes
3212 * the kiocb to be handled from an async context.
3214 if (kiocb->ki_flags & IOCB_HIPRI)
3216 if (kiocb->ki_flags & IOCB_NOWAIT)
3219 while (iov_iter_count(iter)) {
3223 if (!iov_iter_is_bvec(iter)) {
3224 iovec = iov_iter_iovec(iter);
3226 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3227 iovec.iov_len = req->rw.len;
3231 nr = file->f_op->read(file, iovec.iov_base,
3232 iovec.iov_len, io_kiocb_ppos(kiocb));
3234 nr = file->f_op->write(file, iovec.iov_base,
3235 iovec.iov_len, io_kiocb_ppos(kiocb));
3244 if (nr != iovec.iov_len)
3248 iov_iter_advance(iter, nr);
3254 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3255 const struct iovec *fast_iov, struct iov_iter *iter)
3257 struct io_async_rw *rw = req->async_data;
3259 memcpy(&rw->iter, iter, sizeof(*iter));
3260 rw->free_iovec = iovec;
3262 /* can only be fixed buffers, no need to do anything */
3263 if (iov_iter_is_bvec(iter))
3266 unsigned iov_off = 0;
3268 rw->iter.iov = rw->fast_iov;
3269 if (iter->iov != fast_iov) {
3270 iov_off = iter->iov - fast_iov;
3271 rw->iter.iov += iov_off;
3273 if (rw->fast_iov != fast_iov)
3274 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3275 sizeof(struct iovec) * iter->nr_segs);
3277 req->flags |= REQ_F_NEED_CLEANUP;
3281 static inline int __io_alloc_async_data(struct io_kiocb *req)
3283 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3284 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3285 return req->async_data == NULL;
3288 static int io_alloc_async_data(struct io_kiocb *req)
3290 if (!io_op_defs[req->opcode].needs_async_data)
3293 return __io_alloc_async_data(req);
3296 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3297 const struct iovec *fast_iov,
3298 struct iov_iter *iter, bool force)
3300 if (!force && !io_op_defs[req->opcode].needs_async_data)
3302 if (!req->async_data) {
3303 if (__io_alloc_async_data(req))
3306 io_req_map_rw(req, iovec, fast_iov, iter);
3311 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3313 struct io_async_rw *iorw = req->async_data;
3314 struct iovec *iov = iorw->fast_iov;
3317 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3318 if (unlikely(ret < 0))
3321 iorw->bytes_done = 0;
3322 iorw->free_iovec = iov;
3324 req->flags |= REQ_F_NEED_CLEANUP;
3328 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3332 ret = io_prep_rw(req, sqe);
3336 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3339 /* either don't need iovec imported or already have it */
3340 if (!req->async_data)
3342 return io_rw_prep_async(req, READ);
3346 * This is our waitqueue callback handler, registered through lock_page_async()
3347 * when we initially tried to do the IO with the iocb armed our waitqueue.
3348 * This gets called when the page is unlocked, and we generally expect that to
3349 * happen when the page IO is completed and the page is now uptodate. This will
3350 * queue a task_work based retry of the operation, attempting to copy the data
3351 * again. If the latter fails because the page was NOT uptodate, then we will
3352 * do a thread based blocking retry of the operation. That's the unexpected
3355 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3356 int sync, void *arg)
3358 struct wait_page_queue *wpq;
3359 struct io_kiocb *req = wait->private;
3360 struct wait_page_key *key = arg;
3363 wpq = container_of(wait, struct wait_page_queue, wait);
3365 if (!wake_page_match(wpq, key))
3368 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3369 list_del_init(&wait->entry);
3371 init_task_work(&req->task_work, io_req_task_submit);
3372 percpu_ref_get(&req->ctx->refs);
3374 /* submit ref gets dropped, acquire a new one */
3375 refcount_inc(&req->refs);
3376 ret = io_req_task_work_add(req);
3377 if (unlikely(ret)) {
3378 struct task_struct *tsk;
3380 /* queue just for cancelation */
3381 init_task_work(&req->task_work, io_req_task_cancel);
3382 tsk = io_wq_get_task(req->ctx->io_wq);
3383 task_work_add(tsk, &req->task_work, TWA_NONE);
3384 wake_up_process(tsk);
3390 * This controls whether a given IO request should be armed for async page
3391 * based retry. If we return false here, the request is handed to the async
3392 * worker threads for retry. If we're doing buffered reads on a regular file,
3393 * we prepare a private wait_page_queue entry and retry the operation. This
3394 * will either succeed because the page is now uptodate and unlocked, or it
3395 * will register a callback when the page is unlocked at IO completion. Through
3396 * that callback, io_uring uses task_work to setup a retry of the operation.
3397 * That retry will attempt the buffered read again. The retry will generally
3398 * succeed, or in rare cases where it fails, we then fall back to using the
3399 * async worker threads for a blocking retry.
3401 static bool io_rw_should_retry(struct io_kiocb *req)
3403 struct io_async_rw *rw = req->async_data;
3404 struct wait_page_queue *wait = &rw->wpq;
3405 struct kiocb *kiocb = &req->rw.kiocb;
3407 /* never retry for NOWAIT, we just complete with -EAGAIN */
3408 if (req->flags & REQ_F_NOWAIT)
3411 /* Only for buffered IO */
3412 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3416 * just use poll if we can, and don't attempt if the fs doesn't
3417 * support callback based unlocks
3419 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3422 wait->wait.func = io_async_buf_func;
3423 wait->wait.private = req;
3424 wait->wait.flags = 0;
3425 INIT_LIST_HEAD(&wait->wait.entry);
3426 kiocb->ki_flags |= IOCB_WAITQ;
3427 kiocb->ki_flags &= ~IOCB_NOWAIT;
3428 kiocb->ki_waitq = wait;
3432 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3434 if (req->file->f_op->read_iter)
3435 return call_read_iter(req->file, &req->rw.kiocb, iter);
3436 else if (req->file->f_op->read)
3437 return loop_rw_iter(READ, req, iter);
3442 static int io_read(struct io_kiocb *req, bool force_nonblock,
3443 struct io_comp_state *cs)
3445 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3446 struct kiocb *kiocb = &req->rw.kiocb;
3447 struct iov_iter __iter, *iter = &__iter;
3448 struct io_async_rw *rw = req->async_data;
3449 ssize_t io_size, ret, ret2;
3456 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3460 io_size = iov_iter_count(iter);
3461 req->result = io_size;
3464 /* Ensure we clear previously set non-block flag */
3465 if (!force_nonblock)
3466 kiocb->ki_flags &= ~IOCB_NOWAIT;
3468 kiocb->ki_flags |= IOCB_NOWAIT;
3471 /* If the file doesn't support async, just async punt */
3472 no_async = force_nonblock && !io_file_supports_async(req->file, READ);
3476 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size);
3480 ret = io_iter_do_read(req, iter);
3484 } else if (ret == -EIOCBQUEUED) {
3487 } else if (ret == -EAGAIN) {
3488 /* IOPOLL retry should happen for io-wq threads */
3489 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3491 /* no retry on NONBLOCK marked file */
3492 if (req->file->f_flags & O_NONBLOCK)
3494 /* some cases will consume bytes even on error returns */
3495 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3498 } else if (ret < 0) {
3499 /* make sure -ERESTARTSYS -> -EINTR is done */
3503 /* read it all, or we did blocking attempt. no retry. */
3504 if (!iov_iter_count(iter) || !force_nonblock ||
3505 (req->file->f_flags & O_NONBLOCK))
3510 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3517 rw = req->async_data;
3518 /* it's copied and will be cleaned with ->io */
3520 /* now use our persistent iterator, if we aren't already */
3523 rw->bytes_done += ret;
3524 /* if we can retry, do so with the callbacks armed */
3525 if (!io_rw_should_retry(req)) {
3526 kiocb->ki_flags &= ~IOCB_WAITQ;
3531 * Now retry read with the IOCB_WAITQ parts set in the iocb. If we
3532 * get -EIOCBQUEUED, then we'll get a notification when the desired
3533 * page gets unlocked. We can also get a partial read here, and if we
3534 * do, then just retry at the new offset.
3536 ret = io_iter_do_read(req, iter);
3537 if (ret == -EIOCBQUEUED) {
3540 } else if (ret > 0 && ret < io_size) {
3541 /* we got some bytes, but not all. retry. */
3545 kiocb_done(kiocb, ret, cs);
3548 /* it's reportedly faster than delegating the null check to kfree() */
3554 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3558 ret = io_prep_rw(req, sqe);
3562 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3565 /* either don't need iovec imported or already have it */
3566 if (!req->async_data)
3568 return io_rw_prep_async(req, WRITE);
3571 static int io_write(struct io_kiocb *req, bool force_nonblock,
3572 struct io_comp_state *cs)
3574 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3575 struct kiocb *kiocb = &req->rw.kiocb;
3576 struct iov_iter __iter, *iter = &__iter;
3577 struct io_async_rw *rw = req->async_data;
3578 ssize_t ret, ret2, io_size;
3584 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3588 io_size = iov_iter_count(iter);
3589 req->result = io_size;
3591 /* Ensure we clear previously set non-block flag */
3592 if (!force_nonblock)
3593 kiocb->ki_flags &= ~IOCB_NOWAIT;
3595 kiocb->ki_flags |= IOCB_NOWAIT;
3597 /* If the file doesn't support async, just async punt */
3598 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
3601 /* file path doesn't support NOWAIT for non-direct_IO */
3602 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3603 (req->flags & REQ_F_ISREG))
3606 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size);
3611 * Open-code file_start_write here to grab freeze protection,
3612 * which will be released by another thread in
3613 * io_complete_rw(). Fool lockdep by telling it the lock got
3614 * released so that it doesn't complain about the held lock when
3615 * we return to userspace.
3617 if (req->flags & REQ_F_ISREG) {
3618 sb_start_write(file_inode(req->file)->i_sb);
3619 __sb_writers_release(file_inode(req->file)->i_sb,
3622 kiocb->ki_flags |= IOCB_WRITE;
3624 if (req->file->f_op->write_iter)
3625 ret2 = call_write_iter(req->file, kiocb, iter);
3626 else if (req->file->f_op->write)
3627 ret2 = loop_rw_iter(WRITE, req, iter);
3632 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3633 * retry them without IOCB_NOWAIT.
3635 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3637 /* no retry on NONBLOCK marked file */
3638 if (ret2 == -EAGAIN && (req->file->f_flags & O_NONBLOCK))
3640 if (!force_nonblock || ret2 != -EAGAIN) {
3641 /* IOPOLL retry should happen for io-wq threads */
3642 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3645 kiocb_done(kiocb, ret2, cs);
3648 /* some cases will consume bytes even on error returns */
3649 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3650 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3655 /* it's reportedly faster than delegating the null check to kfree() */
3661 static int io_renameat_prep(struct io_kiocb *req,
3662 const struct io_uring_sqe *sqe)
3664 struct io_rename *ren = &req->rename;
3665 const char __user *oldf, *newf;
3667 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3670 ren->old_dfd = READ_ONCE(sqe->fd);
3671 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3672 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3673 ren->new_dfd = READ_ONCE(sqe->len);
3674 ren->flags = READ_ONCE(sqe->rename_flags);
3676 ren->oldpath = getname(oldf);
3677 if (IS_ERR(ren->oldpath))
3678 return PTR_ERR(ren->oldpath);
3680 ren->newpath = getname(newf);
3681 if (IS_ERR(ren->newpath)) {
3682 putname(ren->oldpath);
3683 return PTR_ERR(ren->newpath);
3686 req->flags |= REQ_F_NEED_CLEANUP;
3690 static int io_renameat(struct io_kiocb *req, bool force_nonblock)
3692 struct io_rename *ren = &req->rename;
3698 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3699 ren->newpath, ren->flags);
3701 req->flags &= ~REQ_F_NEED_CLEANUP;
3703 req_set_fail_links(req);
3704 io_req_complete(req, ret);
3708 static int io_unlinkat_prep(struct io_kiocb *req,
3709 const struct io_uring_sqe *sqe)
3711 struct io_unlink *un = &req->unlink;
3712 const char __user *fname;
3714 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3717 un->dfd = READ_ONCE(sqe->fd);
3719 un->flags = READ_ONCE(sqe->unlink_flags);
3720 if (un->flags & ~AT_REMOVEDIR)
3723 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3724 un->filename = getname(fname);
3725 if (IS_ERR(un->filename))
3726 return PTR_ERR(un->filename);
3728 req->flags |= REQ_F_NEED_CLEANUP;
3732 static int io_unlinkat(struct io_kiocb *req, bool force_nonblock)
3734 struct io_unlink *un = &req->unlink;
3740 if (un->flags & AT_REMOVEDIR)
3741 ret = do_rmdir(un->dfd, un->filename);
3743 ret = do_unlinkat(un->dfd, un->filename);
3745 req->flags &= ~REQ_F_NEED_CLEANUP;
3747 req_set_fail_links(req);
3748 io_req_complete(req, ret);
3752 static int io_shutdown_prep(struct io_kiocb *req,
3753 const struct io_uring_sqe *sqe)
3755 #if defined(CONFIG_NET)
3756 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3758 if (sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3762 req->shutdown.how = READ_ONCE(sqe->len);
3769 static int io_shutdown(struct io_kiocb *req, bool force_nonblock)
3771 #if defined(CONFIG_NET)
3772 struct socket *sock;
3778 sock = sock_from_file(req->file);
3779 if (unlikely(!sock))
3782 ret = __sys_shutdown_sock(sock, req->shutdown.how);
3784 req_set_fail_links(req);
3785 io_req_complete(req, ret);
3792 static int __io_splice_prep(struct io_kiocb *req,
3793 const struct io_uring_sqe *sqe)
3795 struct io_splice* sp = &req->splice;
3796 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3798 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3802 sp->len = READ_ONCE(sqe->len);
3803 sp->flags = READ_ONCE(sqe->splice_flags);
3805 if (unlikely(sp->flags & ~valid_flags))
3808 sp->file_in = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in),
3809 (sp->flags & SPLICE_F_FD_IN_FIXED));
3812 req->flags |= REQ_F_NEED_CLEANUP;
3814 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
3816 * Splice operation will be punted aync, and here need to
3817 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3819 io_req_init_async(req);
3820 req->work.flags |= IO_WQ_WORK_UNBOUND;
3826 static int io_tee_prep(struct io_kiocb *req,
3827 const struct io_uring_sqe *sqe)
3829 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3831 return __io_splice_prep(req, sqe);
3834 static int io_tee(struct io_kiocb *req, bool force_nonblock)
3836 struct io_splice *sp = &req->splice;
3837 struct file *in = sp->file_in;
3838 struct file *out = sp->file_out;
3839 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3845 ret = do_tee(in, out, sp->len, flags);
3847 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3848 req->flags &= ~REQ_F_NEED_CLEANUP;
3851 req_set_fail_links(req);
3852 io_req_complete(req, ret);
3856 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3858 struct io_splice* sp = &req->splice;
3860 sp->off_in = READ_ONCE(sqe->splice_off_in);
3861 sp->off_out = READ_ONCE(sqe->off);
3862 return __io_splice_prep(req, sqe);
3865 static int io_splice(struct io_kiocb *req, bool force_nonblock)
3867 struct io_splice *sp = &req->splice;
3868 struct file *in = sp->file_in;
3869 struct file *out = sp->file_out;
3870 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3871 loff_t *poff_in, *poff_out;
3877 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3878 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3881 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3883 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3884 req->flags &= ~REQ_F_NEED_CLEANUP;
3887 req_set_fail_links(req);
3888 io_req_complete(req, ret);
3893 * IORING_OP_NOP just posts a completion event, nothing else.
3895 static int io_nop(struct io_kiocb *req, struct io_comp_state *cs)
3897 struct io_ring_ctx *ctx = req->ctx;
3899 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3902 __io_req_complete(req, 0, 0, cs);
3906 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3908 struct io_ring_ctx *ctx = req->ctx;
3913 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3915 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3918 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3919 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3922 req->sync.off = READ_ONCE(sqe->off);
3923 req->sync.len = READ_ONCE(sqe->len);
3927 static int io_fsync(struct io_kiocb *req, bool force_nonblock)
3929 loff_t end = req->sync.off + req->sync.len;
3932 /* fsync always requires a blocking context */
3936 ret = vfs_fsync_range(req->file, req->sync.off,
3937 end > 0 ? end : LLONG_MAX,
3938 req->sync.flags & IORING_FSYNC_DATASYNC);
3940 req_set_fail_links(req);
3941 io_req_complete(req, ret);
3945 static int io_fallocate_prep(struct io_kiocb *req,
3946 const struct io_uring_sqe *sqe)
3948 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3950 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3953 req->sync.off = READ_ONCE(sqe->off);
3954 req->sync.len = READ_ONCE(sqe->addr);
3955 req->sync.mode = READ_ONCE(sqe->len);
3959 static int io_fallocate(struct io_kiocb *req, bool force_nonblock)
3963 /* fallocate always requiring blocking context */
3966 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3969 req_set_fail_links(req);
3970 io_req_complete(req, ret);
3974 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3976 const char __user *fname;
3979 if (unlikely(sqe->ioprio || sqe->buf_index))
3981 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3984 /* open.how should be already initialised */
3985 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3986 req->open.how.flags |= O_LARGEFILE;
3988 req->open.dfd = READ_ONCE(sqe->fd);
3989 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3990 req->open.filename = getname(fname);
3991 if (IS_ERR(req->open.filename)) {
3992 ret = PTR_ERR(req->open.filename);
3993 req->open.filename = NULL;
3996 req->open.nofile = rlimit(RLIMIT_NOFILE);
3997 req->open.ignore_nonblock = false;
3998 req->flags |= REQ_F_NEED_CLEANUP;
4002 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4006 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4008 mode = READ_ONCE(sqe->len);
4009 flags = READ_ONCE(sqe->open_flags);
4010 req->open.how = build_open_how(flags, mode);
4011 return __io_openat_prep(req, sqe);
4014 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4016 struct open_how __user *how;
4020 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4022 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4023 len = READ_ONCE(sqe->len);
4024 if (len < OPEN_HOW_SIZE_VER0)
4027 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
4032 return __io_openat_prep(req, sqe);
4035 static int io_openat2(struct io_kiocb *req, bool force_nonblock)
4037 struct open_flags op;
4041 if (force_nonblock && !req->open.ignore_nonblock)
4044 ret = build_open_flags(&req->open.how, &op);
4048 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
4052 file = do_filp_open(req->open.dfd, req->open.filename, &op);
4055 ret = PTR_ERR(file);
4057 * A work-around to ensure that /proc/self works that way
4058 * that it should - if we get -EOPNOTSUPP back, then assume
4059 * that proc_self_get_link() failed us because we're in async
4060 * context. We should be safe to retry this from the task
4061 * itself with force_nonblock == false set, as it should not
4062 * block on lookup. Would be nice to know this upfront and
4063 * avoid the async dance, but doesn't seem feasible.
4065 if (ret == -EOPNOTSUPP && io_wq_current_is_worker()) {
4066 req->open.ignore_nonblock = true;
4067 refcount_inc(&req->refs);
4068 io_req_task_queue(req);
4072 fsnotify_open(file);
4073 fd_install(ret, file);
4076 putname(req->open.filename);
4077 req->flags &= ~REQ_F_NEED_CLEANUP;
4079 req_set_fail_links(req);
4080 io_req_complete(req, ret);
4084 static int io_openat(struct io_kiocb *req, bool force_nonblock)
4086 return io_openat2(req, force_nonblock);
4089 static int io_remove_buffers_prep(struct io_kiocb *req,
4090 const struct io_uring_sqe *sqe)
4092 struct io_provide_buf *p = &req->pbuf;
4095 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
4098 tmp = READ_ONCE(sqe->fd);
4099 if (!tmp || tmp > USHRT_MAX)
4102 memset(p, 0, sizeof(*p));
4104 p->bgid = READ_ONCE(sqe->buf_group);
4108 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
4109 int bgid, unsigned nbufs)
4113 /* shouldn't happen */
4117 /* the head kbuf is the list itself */
4118 while (!list_empty(&buf->list)) {
4119 struct io_buffer *nxt;
4121 nxt = list_first_entry(&buf->list, struct io_buffer, list);
4122 list_del(&nxt->list);
4129 idr_remove(&ctx->io_buffer_idr, bgid);
4134 static int io_remove_buffers(struct io_kiocb *req, bool force_nonblock,
4135 struct io_comp_state *cs)
4137 struct io_provide_buf *p = &req->pbuf;
4138 struct io_ring_ctx *ctx = req->ctx;
4139 struct io_buffer *head;
4142 io_ring_submit_lock(ctx, !force_nonblock);
4144 lockdep_assert_held(&ctx->uring_lock);
4147 head = idr_find(&ctx->io_buffer_idr, p->bgid);
4149 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
4151 req_set_fail_links(req);
4153 /* need to hold the lock to complete IOPOLL requests */
4154 if (ctx->flags & IORING_SETUP_IOPOLL) {
4155 __io_req_complete(req, ret, 0, cs);
4156 io_ring_submit_unlock(ctx, !force_nonblock);
4158 io_ring_submit_unlock(ctx, !force_nonblock);
4159 __io_req_complete(req, ret, 0, cs);
4164 static int io_provide_buffers_prep(struct io_kiocb *req,
4165 const struct io_uring_sqe *sqe)
4167 struct io_provide_buf *p = &req->pbuf;
4170 if (sqe->ioprio || sqe->rw_flags)
4173 tmp = READ_ONCE(sqe->fd);
4174 if (!tmp || tmp > USHRT_MAX)
4177 p->addr = READ_ONCE(sqe->addr);
4178 p->len = READ_ONCE(sqe->len);
4180 if (!access_ok(u64_to_user_ptr(p->addr), (p->len * p->nbufs)))
4183 p->bgid = READ_ONCE(sqe->buf_group);
4184 tmp = READ_ONCE(sqe->off);
4185 if (tmp > USHRT_MAX)
4191 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
4193 struct io_buffer *buf;
4194 u64 addr = pbuf->addr;
4195 int i, bid = pbuf->bid;
4197 for (i = 0; i < pbuf->nbufs; i++) {
4198 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
4203 buf->len = pbuf->len;
4208 INIT_LIST_HEAD(&buf->list);
4211 list_add_tail(&buf->list, &(*head)->list);
4215 return i ? i : -ENOMEM;
4218 static int io_provide_buffers(struct io_kiocb *req, bool force_nonblock,
4219 struct io_comp_state *cs)
4221 struct io_provide_buf *p = &req->pbuf;
4222 struct io_ring_ctx *ctx = req->ctx;
4223 struct io_buffer *head, *list;
4226 io_ring_submit_lock(ctx, !force_nonblock);
4228 lockdep_assert_held(&ctx->uring_lock);
4230 list = head = idr_find(&ctx->io_buffer_idr, p->bgid);
4232 ret = io_add_buffers(p, &head);
4237 ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1,
4240 __io_remove_buffers(ctx, head, p->bgid, -1U);
4246 req_set_fail_links(req);
4248 /* need to hold the lock to complete IOPOLL requests */
4249 if (ctx->flags & IORING_SETUP_IOPOLL) {
4250 __io_req_complete(req, ret, 0, cs);
4251 io_ring_submit_unlock(ctx, !force_nonblock);
4253 io_ring_submit_unlock(ctx, !force_nonblock);
4254 __io_req_complete(req, ret, 0, cs);
4259 static int io_epoll_ctl_prep(struct io_kiocb *req,
4260 const struct io_uring_sqe *sqe)
4262 #if defined(CONFIG_EPOLL)
4263 if (sqe->ioprio || sqe->buf_index)
4265 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4268 req->epoll.epfd = READ_ONCE(sqe->fd);
4269 req->epoll.op = READ_ONCE(sqe->len);
4270 req->epoll.fd = READ_ONCE(sqe->off);
4272 if (ep_op_has_event(req->epoll.op)) {
4273 struct epoll_event __user *ev;
4275 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4276 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4286 static int io_epoll_ctl(struct io_kiocb *req, bool force_nonblock,
4287 struct io_comp_state *cs)
4289 #if defined(CONFIG_EPOLL)
4290 struct io_epoll *ie = &req->epoll;
4293 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4294 if (force_nonblock && ret == -EAGAIN)
4298 req_set_fail_links(req);
4299 __io_req_complete(req, ret, 0, cs);
4306 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4308 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4309 if (sqe->ioprio || sqe->buf_index || sqe->off)
4311 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4314 req->madvise.addr = READ_ONCE(sqe->addr);
4315 req->madvise.len = READ_ONCE(sqe->len);
4316 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4323 static int io_madvise(struct io_kiocb *req, bool force_nonblock)
4325 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4326 struct io_madvise *ma = &req->madvise;
4332 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4334 req_set_fail_links(req);
4335 io_req_complete(req, ret);
4342 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4344 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4346 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4349 req->fadvise.offset = READ_ONCE(sqe->off);
4350 req->fadvise.len = READ_ONCE(sqe->len);
4351 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4355 static int io_fadvise(struct io_kiocb *req, bool force_nonblock)
4357 struct io_fadvise *fa = &req->fadvise;
4360 if (force_nonblock) {
4361 switch (fa->advice) {
4362 case POSIX_FADV_NORMAL:
4363 case POSIX_FADV_RANDOM:
4364 case POSIX_FADV_SEQUENTIAL:
4371 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4373 req_set_fail_links(req);
4374 io_req_complete(req, ret);
4378 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4380 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4382 if (sqe->ioprio || sqe->buf_index)
4384 if (req->flags & REQ_F_FIXED_FILE)
4387 req->statx.dfd = READ_ONCE(sqe->fd);
4388 req->statx.mask = READ_ONCE(sqe->len);
4389 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4390 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4391 req->statx.flags = READ_ONCE(sqe->statx_flags);
4396 static int io_statx(struct io_kiocb *req, bool force_nonblock)
4398 struct io_statx *ctx = &req->statx;
4401 if (force_nonblock) {
4402 /* only need file table for an actual valid fd */
4403 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
4404 req->flags |= REQ_F_NO_FILE_TABLE;
4408 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4412 req_set_fail_links(req);
4413 io_req_complete(req, ret);
4417 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4420 * If we queue this for async, it must not be cancellable. That would
4421 * leave the 'file' in an undeterminate state, and here need to modify
4422 * io_wq_work.flags, so initialize io_wq_work firstly.
4424 io_req_init_async(req);
4425 req->work.flags |= IO_WQ_WORK_NO_CANCEL;
4427 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4429 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4430 sqe->rw_flags || sqe->buf_index)
4432 if (req->flags & REQ_F_FIXED_FILE)
4435 req->close.fd = READ_ONCE(sqe->fd);
4436 if ((req->file && req->file->f_op == &io_uring_fops))
4439 req->close.put_file = NULL;
4443 static int io_close(struct io_kiocb *req, bool force_nonblock,
4444 struct io_comp_state *cs)
4446 struct io_close *close = &req->close;
4449 /* might be already done during nonblock submission */
4450 if (!close->put_file) {
4451 ret = close_fd_get_file(close->fd, &close->put_file);
4453 return (ret == -ENOENT) ? -EBADF : ret;
4456 /* if the file has a flush method, be safe and punt to async */
4457 if (close->put_file->f_op->flush && force_nonblock) {
4458 /* was never set, but play safe */
4459 req->flags &= ~REQ_F_NOWAIT;
4460 /* avoid grabbing files - we don't need the files */
4461 req->flags |= REQ_F_NO_FILE_TABLE;
4465 /* No ->flush() or already async, safely close from here */
4466 ret = filp_close(close->put_file, req->work.identity->files);
4468 req_set_fail_links(req);
4469 fput(close->put_file);
4470 close->put_file = NULL;
4471 __io_req_complete(req, ret, 0, cs);
4475 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4477 struct io_ring_ctx *ctx = req->ctx;
4482 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4484 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4487 req->sync.off = READ_ONCE(sqe->off);
4488 req->sync.len = READ_ONCE(sqe->len);
4489 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4493 static int io_sync_file_range(struct io_kiocb *req, bool force_nonblock)
4497 /* sync_file_range always requires a blocking context */
4501 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4504 req_set_fail_links(req);
4505 io_req_complete(req, ret);
4509 #if defined(CONFIG_NET)
4510 static int io_setup_async_msg(struct io_kiocb *req,
4511 struct io_async_msghdr *kmsg)
4513 struct io_async_msghdr *async_msg = req->async_data;
4517 if (io_alloc_async_data(req)) {
4518 if (kmsg->iov != kmsg->fast_iov)
4522 async_msg = req->async_data;
4523 req->flags |= REQ_F_NEED_CLEANUP;
4524 memcpy(async_msg, kmsg, sizeof(*kmsg));
4528 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4529 struct io_async_msghdr *iomsg)
4531 iomsg->iov = iomsg->fast_iov;
4532 iomsg->msg.msg_name = &iomsg->addr;
4533 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4534 req->sr_msg.msg_flags, &iomsg->iov);
4537 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4539 struct io_async_msghdr *async_msg = req->async_data;
4540 struct io_sr_msg *sr = &req->sr_msg;
4543 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4546 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4547 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4548 sr->len = READ_ONCE(sqe->len);
4550 #ifdef CONFIG_COMPAT
4551 if (req->ctx->compat)
4552 sr->msg_flags |= MSG_CMSG_COMPAT;
4555 if (!async_msg || !io_op_defs[req->opcode].needs_async_data)
4557 ret = io_sendmsg_copy_hdr(req, async_msg);
4559 req->flags |= REQ_F_NEED_CLEANUP;
4563 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4564 struct io_comp_state *cs)
4566 struct io_async_msghdr iomsg, *kmsg;
4567 struct socket *sock;
4571 sock = sock_from_file(req->file);
4572 if (unlikely(!sock))
4575 if (req->async_data) {
4576 kmsg = req->async_data;
4577 kmsg->msg.msg_name = &kmsg->addr;
4578 /* if iov is set, it's allocated already */
4580 kmsg->iov = kmsg->fast_iov;
4581 kmsg->msg.msg_iter.iov = kmsg->iov;
4583 ret = io_sendmsg_copy_hdr(req, &iomsg);
4589 flags = req->sr_msg.msg_flags;
4590 if (flags & MSG_DONTWAIT)
4591 req->flags |= REQ_F_NOWAIT;
4592 else if (force_nonblock)
4593 flags |= MSG_DONTWAIT;
4595 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4596 if (force_nonblock && ret == -EAGAIN)
4597 return io_setup_async_msg(req, kmsg);
4598 if (ret == -ERESTARTSYS)
4601 if (kmsg->iov != kmsg->fast_iov)
4603 req->flags &= ~REQ_F_NEED_CLEANUP;
4605 req_set_fail_links(req);
4606 __io_req_complete(req, ret, 0, cs);
4610 static int io_send(struct io_kiocb *req, bool force_nonblock,
4611 struct io_comp_state *cs)
4613 struct io_sr_msg *sr = &req->sr_msg;
4616 struct socket *sock;
4620 sock = sock_from_file(req->file);
4621 if (unlikely(!sock))
4624 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4628 msg.msg_name = NULL;
4629 msg.msg_control = NULL;
4630 msg.msg_controllen = 0;
4631 msg.msg_namelen = 0;
4633 flags = req->sr_msg.msg_flags;
4634 if (flags & MSG_DONTWAIT)
4635 req->flags |= REQ_F_NOWAIT;
4636 else if (force_nonblock)
4637 flags |= MSG_DONTWAIT;
4639 msg.msg_flags = flags;
4640 ret = sock_sendmsg(sock, &msg);
4641 if (force_nonblock && ret == -EAGAIN)
4643 if (ret == -ERESTARTSYS)
4647 req_set_fail_links(req);
4648 __io_req_complete(req, ret, 0, cs);
4652 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4653 struct io_async_msghdr *iomsg)
4655 struct io_sr_msg *sr = &req->sr_msg;
4656 struct iovec __user *uiov;
4660 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4661 &iomsg->uaddr, &uiov, &iov_len);
4665 if (req->flags & REQ_F_BUFFER_SELECT) {
4668 if (copy_from_user(iomsg->iov, uiov, sizeof(*uiov)))
4670 sr->len = iomsg->iov[0].iov_len;
4671 iov_iter_init(&iomsg->msg.msg_iter, READ, iomsg->iov, 1,
4675 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4676 &iomsg->iov, &iomsg->msg.msg_iter,
4685 #ifdef CONFIG_COMPAT
4686 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4687 struct io_async_msghdr *iomsg)
4689 struct compat_msghdr __user *msg_compat;
4690 struct io_sr_msg *sr = &req->sr_msg;
4691 struct compat_iovec __user *uiov;
4696 msg_compat = (struct compat_msghdr __user *) sr->umsg;
4697 ret = __get_compat_msghdr(&iomsg->msg, msg_compat, &iomsg->uaddr,
4702 uiov = compat_ptr(ptr);
4703 if (req->flags & REQ_F_BUFFER_SELECT) {
4704 compat_ssize_t clen;
4708 if (!access_ok(uiov, sizeof(*uiov)))
4710 if (__get_user(clen, &uiov->iov_len))
4715 iomsg->iov[0].iov_len = clen;
4718 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4719 UIO_FASTIOV, &iomsg->iov,
4720 &iomsg->msg.msg_iter, true);
4729 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4730 struct io_async_msghdr *iomsg)
4732 iomsg->msg.msg_name = &iomsg->addr;
4733 iomsg->iov = iomsg->fast_iov;
4735 #ifdef CONFIG_COMPAT
4736 if (req->ctx->compat)
4737 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4740 return __io_recvmsg_copy_hdr(req, iomsg);
4743 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4746 struct io_sr_msg *sr = &req->sr_msg;
4747 struct io_buffer *kbuf;
4749 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4754 req->flags |= REQ_F_BUFFER_SELECTED;
4758 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4760 return io_put_kbuf(req, req->sr_msg.kbuf);
4763 static int io_recvmsg_prep(struct io_kiocb *req,
4764 const struct io_uring_sqe *sqe)
4766 struct io_async_msghdr *async_msg = req->async_data;
4767 struct io_sr_msg *sr = &req->sr_msg;
4770 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4773 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4774 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4775 sr->len = READ_ONCE(sqe->len);
4776 sr->bgid = READ_ONCE(sqe->buf_group);
4778 #ifdef CONFIG_COMPAT
4779 if (req->ctx->compat)
4780 sr->msg_flags |= MSG_CMSG_COMPAT;
4783 if (!async_msg || !io_op_defs[req->opcode].needs_async_data)
4785 ret = io_recvmsg_copy_hdr(req, async_msg);
4787 req->flags |= REQ_F_NEED_CLEANUP;
4791 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4792 struct io_comp_state *cs)
4794 struct io_async_msghdr iomsg, *kmsg;
4795 struct socket *sock;
4796 struct io_buffer *kbuf;
4798 int ret, cflags = 0;
4800 sock = sock_from_file(req->file);
4801 if (unlikely(!sock))
4804 if (req->async_data) {
4805 kmsg = req->async_data;
4806 kmsg->msg.msg_name = &kmsg->addr;
4807 /* if iov is set, it's allocated already */
4809 kmsg->iov = kmsg->fast_iov;
4810 kmsg->msg.msg_iter.iov = kmsg->iov;
4812 ret = io_recvmsg_copy_hdr(req, &iomsg);
4818 if (req->flags & REQ_F_BUFFER_SELECT) {
4819 kbuf = io_recv_buffer_select(req, !force_nonblock);
4821 return PTR_ERR(kbuf);
4822 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4823 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->iov,
4824 1, req->sr_msg.len);
4827 flags = req->sr_msg.msg_flags;
4828 if (flags & MSG_DONTWAIT)
4829 req->flags |= REQ_F_NOWAIT;
4830 else if (force_nonblock)
4831 flags |= MSG_DONTWAIT;
4833 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4834 kmsg->uaddr, flags);
4835 if (force_nonblock && ret == -EAGAIN)
4836 return io_setup_async_msg(req, kmsg);
4837 if (ret == -ERESTARTSYS)
4840 if (req->flags & REQ_F_BUFFER_SELECTED)
4841 cflags = io_put_recv_kbuf(req);
4842 if (kmsg->iov != kmsg->fast_iov)
4844 req->flags &= ~REQ_F_NEED_CLEANUP;
4846 req_set_fail_links(req);
4847 __io_req_complete(req, ret, cflags, cs);
4851 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4852 struct io_comp_state *cs)
4854 struct io_buffer *kbuf;
4855 struct io_sr_msg *sr = &req->sr_msg;
4857 void __user *buf = sr->buf;
4858 struct socket *sock;
4861 int ret, cflags = 0;
4863 sock = sock_from_file(req->file);
4864 if (unlikely(!sock))
4867 if (req->flags & REQ_F_BUFFER_SELECT) {
4868 kbuf = io_recv_buffer_select(req, !force_nonblock);
4870 return PTR_ERR(kbuf);
4871 buf = u64_to_user_ptr(kbuf->addr);
4874 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4878 msg.msg_name = NULL;
4879 msg.msg_control = NULL;
4880 msg.msg_controllen = 0;
4881 msg.msg_namelen = 0;
4882 msg.msg_iocb = NULL;
4885 flags = req->sr_msg.msg_flags;
4886 if (flags & MSG_DONTWAIT)
4887 req->flags |= REQ_F_NOWAIT;
4888 else if (force_nonblock)
4889 flags |= MSG_DONTWAIT;
4891 ret = sock_recvmsg(sock, &msg, flags);
4892 if (force_nonblock && ret == -EAGAIN)
4894 if (ret == -ERESTARTSYS)
4897 if (req->flags & REQ_F_BUFFER_SELECTED)
4898 cflags = io_put_recv_kbuf(req);
4900 req_set_fail_links(req);
4901 __io_req_complete(req, ret, cflags, cs);
4905 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4907 struct io_accept *accept = &req->accept;
4909 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4911 if (sqe->ioprio || sqe->len || sqe->buf_index)
4914 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4915 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4916 accept->flags = READ_ONCE(sqe->accept_flags);
4917 accept->nofile = rlimit(RLIMIT_NOFILE);
4921 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4922 struct io_comp_state *cs)
4924 struct io_accept *accept = &req->accept;
4925 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4928 if (req->file->f_flags & O_NONBLOCK)
4929 req->flags |= REQ_F_NOWAIT;
4931 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4932 accept->addr_len, accept->flags,
4934 if (ret == -EAGAIN && force_nonblock)
4937 if (ret == -ERESTARTSYS)
4939 req_set_fail_links(req);
4941 __io_req_complete(req, ret, 0, cs);
4945 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4947 struct io_connect *conn = &req->connect;
4948 struct io_async_connect *io = req->async_data;
4950 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4952 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4955 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4956 conn->addr_len = READ_ONCE(sqe->addr2);
4961 return move_addr_to_kernel(conn->addr, conn->addr_len,
4965 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4966 struct io_comp_state *cs)
4968 struct io_async_connect __io, *io;
4969 unsigned file_flags;
4972 if (req->async_data) {
4973 io = req->async_data;
4975 ret = move_addr_to_kernel(req->connect.addr,
4976 req->connect.addr_len,
4983 file_flags = force_nonblock ? O_NONBLOCK : 0;
4985 ret = __sys_connect_file(req->file, &io->address,
4986 req->connect.addr_len, file_flags);
4987 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4988 if (req->async_data)
4990 if (io_alloc_async_data(req)) {
4994 io = req->async_data;
4995 memcpy(req->async_data, &__io, sizeof(__io));
4998 if (ret == -ERESTARTSYS)
5002 req_set_fail_links(req);
5003 __io_req_complete(req, ret, 0, cs);
5006 #else /* !CONFIG_NET */
5007 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5012 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
5013 struct io_comp_state *cs)
5018 static int io_send(struct io_kiocb *req, bool force_nonblock,
5019 struct io_comp_state *cs)
5024 static int io_recvmsg_prep(struct io_kiocb *req,
5025 const struct io_uring_sqe *sqe)
5030 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
5031 struct io_comp_state *cs)
5036 static int io_recv(struct io_kiocb *req, bool force_nonblock,
5037 struct io_comp_state *cs)
5042 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5047 static int io_accept(struct io_kiocb *req, bool force_nonblock,
5048 struct io_comp_state *cs)
5053 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5058 static int io_connect(struct io_kiocb *req, bool force_nonblock,
5059 struct io_comp_state *cs)
5063 #endif /* CONFIG_NET */
5065 struct io_poll_table {
5066 struct poll_table_struct pt;
5067 struct io_kiocb *req;
5071 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
5072 __poll_t mask, task_work_func_t func)
5076 /* for instances that support it check for an event match first: */
5077 if (mask && !(mask & poll->events))
5080 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
5082 list_del_init(&poll->wait.entry);
5085 init_task_work(&req->task_work, func);
5086 percpu_ref_get(&req->ctx->refs);
5089 * If this fails, then the task is exiting. When a task exits, the
5090 * work gets canceled, so just cancel this request as well instead
5091 * of executing it. We can't safely execute it anyway, as we may not
5092 * have the needed state needed for it anyway.
5094 ret = io_req_task_work_add(req);
5095 if (unlikely(ret)) {
5096 struct task_struct *tsk;
5098 WRITE_ONCE(poll->canceled, true);
5099 tsk = io_wq_get_task(req->ctx->io_wq);
5100 task_work_add(tsk, &req->task_work, TWA_NONE);
5101 wake_up_process(tsk);
5106 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
5107 __acquires(&req->ctx->completion_lock)
5109 struct io_ring_ctx *ctx = req->ctx;
5111 if (!req->result && !READ_ONCE(poll->canceled)) {
5112 struct poll_table_struct pt = { ._key = poll->events };
5114 req->result = vfs_poll(req->file, &pt) & poll->events;
5117 spin_lock_irq(&ctx->completion_lock);
5118 if (!req->result && !READ_ONCE(poll->canceled)) {
5119 add_wait_queue(poll->head, &poll->wait);
5126 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
5128 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
5129 if (req->opcode == IORING_OP_POLL_ADD)
5130 return req->async_data;
5131 return req->apoll->double_poll;
5134 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
5136 if (req->opcode == IORING_OP_POLL_ADD)
5138 return &req->apoll->poll;
5141 static void io_poll_remove_double(struct io_kiocb *req)
5143 struct io_poll_iocb *poll = io_poll_get_double(req);
5145 lockdep_assert_held(&req->ctx->completion_lock);
5147 if (poll && poll->head) {
5148 struct wait_queue_head *head = poll->head;
5150 spin_lock(&head->lock);
5151 list_del_init(&poll->wait.entry);
5152 if (poll->wait.private)
5153 refcount_dec(&req->refs);
5155 spin_unlock(&head->lock);
5159 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
5161 struct io_ring_ctx *ctx = req->ctx;
5163 io_poll_remove_double(req);
5164 req->poll.done = true;
5165 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
5166 io_commit_cqring(ctx);
5169 static void io_poll_task_func(struct callback_head *cb)
5171 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
5172 struct io_ring_ctx *ctx = req->ctx;
5173 struct io_kiocb *nxt;
5175 if (io_poll_rewait(req, &req->poll)) {
5176 spin_unlock_irq(&ctx->completion_lock);
5178 hash_del(&req->hash_node);
5179 io_poll_complete(req, req->result, 0);
5180 spin_unlock_irq(&ctx->completion_lock);
5182 nxt = io_put_req_find_next(req);
5183 io_cqring_ev_posted(ctx);
5185 __io_req_task_submit(nxt);
5188 percpu_ref_put(&ctx->refs);
5191 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
5192 int sync, void *key)
5194 struct io_kiocb *req = wait->private;
5195 struct io_poll_iocb *poll = io_poll_get_single(req);
5196 __poll_t mask = key_to_poll(key);
5198 /* for instances that support it check for an event match first: */
5199 if (mask && !(mask & poll->events))
5202 list_del_init(&wait->entry);
5204 if (poll && poll->head) {
5207 spin_lock(&poll->head->lock);
5208 done = list_empty(&poll->wait.entry);
5210 list_del_init(&poll->wait.entry);
5211 /* make sure double remove sees this as being gone */
5212 wait->private = NULL;
5213 spin_unlock(&poll->head->lock);
5215 /* use wait func handler, so it matches the rq type */
5216 poll->wait.func(&poll->wait, mode, sync, key);
5219 refcount_dec(&req->refs);
5223 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
5224 wait_queue_func_t wake_func)
5228 poll->canceled = false;
5229 poll->events = events;
5230 INIT_LIST_HEAD(&poll->wait.entry);
5231 init_waitqueue_func_entry(&poll->wait, wake_func);
5234 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5235 struct wait_queue_head *head,
5236 struct io_poll_iocb **poll_ptr)
5238 struct io_kiocb *req = pt->req;
5241 * If poll->head is already set, it's because the file being polled
5242 * uses multiple waitqueues for poll handling (eg one for read, one
5243 * for write). Setup a separate io_poll_iocb if this happens.
5245 if (unlikely(poll->head)) {
5246 struct io_poll_iocb *poll_one = poll;
5248 /* already have a 2nd entry, fail a third attempt */
5250 pt->error = -EINVAL;
5253 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5255 pt->error = -ENOMEM;
5258 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5259 refcount_inc(&req->refs);
5260 poll->wait.private = req;
5267 if (poll->events & EPOLLEXCLUSIVE)
5268 add_wait_queue_exclusive(head, &poll->wait);
5270 add_wait_queue(head, &poll->wait);
5273 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5274 struct poll_table_struct *p)
5276 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5277 struct async_poll *apoll = pt->req->apoll;
5279 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5282 static void io_async_task_func(struct callback_head *cb)
5284 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
5285 struct async_poll *apoll = req->apoll;
5286 struct io_ring_ctx *ctx = req->ctx;
5288 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
5290 if (io_poll_rewait(req, &apoll->poll)) {
5291 spin_unlock_irq(&ctx->completion_lock);
5292 percpu_ref_put(&ctx->refs);
5296 /* If req is still hashed, it cannot have been canceled. Don't check. */
5297 if (hash_hashed(&req->hash_node))
5298 hash_del(&req->hash_node);
5300 io_poll_remove_double(req);
5301 spin_unlock_irq(&ctx->completion_lock);
5303 if (!READ_ONCE(apoll->poll.canceled))
5304 __io_req_task_submit(req);
5306 __io_req_task_cancel(req, -ECANCELED);
5308 percpu_ref_put(&ctx->refs);
5309 kfree(apoll->double_poll);
5313 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5316 struct io_kiocb *req = wait->private;
5317 struct io_poll_iocb *poll = &req->apoll->poll;
5319 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5322 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5325 static void io_poll_req_insert(struct io_kiocb *req)
5327 struct io_ring_ctx *ctx = req->ctx;
5328 struct hlist_head *list;
5330 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5331 hlist_add_head(&req->hash_node, list);
5334 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5335 struct io_poll_iocb *poll,
5336 struct io_poll_table *ipt, __poll_t mask,
5337 wait_queue_func_t wake_func)
5338 __acquires(&ctx->completion_lock)
5340 struct io_ring_ctx *ctx = req->ctx;
5341 bool cancel = false;
5343 INIT_HLIST_NODE(&req->hash_node);
5344 io_init_poll_iocb(poll, mask, wake_func);
5345 poll->file = req->file;
5346 poll->wait.private = req;
5348 ipt->pt._key = mask;
5350 ipt->error = -EINVAL;
5352 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5354 spin_lock_irq(&ctx->completion_lock);
5355 if (likely(poll->head)) {
5356 spin_lock(&poll->head->lock);
5357 if (unlikely(list_empty(&poll->wait.entry))) {
5363 if (mask || ipt->error)
5364 list_del_init(&poll->wait.entry);
5366 WRITE_ONCE(poll->canceled, true);
5367 else if (!poll->done) /* actually waiting for an event */
5368 io_poll_req_insert(req);
5369 spin_unlock(&poll->head->lock);
5375 static bool io_arm_poll_handler(struct io_kiocb *req)
5377 const struct io_op_def *def = &io_op_defs[req->opcode];
5378 struct io_ring_ctx *ctx = req->ctx;
5379 struct async_poll *apoll;
5380 struct io_poll_table ipt;
5384 if (!req->file || !file_can_poll(req->file))
5386 if (req->flags & REQ_F_POLLED)
5390 else if (def->pollout)
5394 /* if we can't nonblock try, then no point in arming a poll handler */
5395 if (!io_file_supports_async(req->file, rw))
5398 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5399 if (unlikely(!apoll))
5401 apoll->double_poll = NULL;
5403 req->flags |= REQ_F_POLLED;
5408 mask |= POLLIN | POLLRDNORM;
5410 mask |= POLLOUT | POLLWRNORM;
5412 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5413 if ((req->opcode == IORING_OP_RECVMSG) &&
5414 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5417 mask |= POLLERR | POLLPRI;
5419 ipt.pt._qproc = io_async_queue_proc;
5421 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5423 if (ret || ipt.error) {
5424 io_poll_remove_double(req);
5425 spin_unlock_irq(&ctx->completion_lock);
5426 kfree(apoll->double_poll);
5430 spin_unlock_irq(&ctx->completion_lock);
5431 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
5432 apoll->poll.events);
5436 static bool __io_poll_remove_one(struct io_kiocb *req,
5437 struct io_poll_iocb *poll)
5439 bool do_complete = false;
5441 spin_lock(&poll->head->lock);
5442 WRITE_ONCE(poll->canceled, true);
5443 if (!list_empty(&poll->wait.entry)) {
5444 list_del_init(&poll->wait.entry);
5447 spin_unlock(&poll->head->lock);
5448 hash_del(&req->hash_node);
5452 static bool io_poll_remove_one(struct io_kiocb *req)
5456 io_poll_remove_double(req);
5458 if (req->opcode == IORING_OP_POLL_ADD) {
5459 do_complete = __io_poll_remove_one(req, &req->poll);
5461 struct async_poll *apoll = req->apoll;
5463 /* non-poll requests have submit ref still */
5464 do_complete = __io_poll_remove_one(req, &apoll->poll);
5467 kfree(apoll->double_poll);
5473 io_cqring_fill_event(req, -ECANCELED);
5474 io_commit_cqring(req->ctx);
5475 req_set_fail_links(req);
5476 io_put_req_deferred(req, 1);
5483 * Returns true if we found and killed one or more poll requests
5485 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5486 struct files_struct *files)
5488 struct hlist_node *tmp;
5489 struct io_kiocb *req;
5492 spin_lock_irq(&ctx->completion_lock);
5493 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5494 struct hlist_head *list;
5496 list = &ctx->cancel_hash[i];
5497 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5498 if (io_match_task(req, tsk, files))
5499 posted += io_poll_remove_one(req);
5502 spin_unlock_irq(&ctx->completion_lock);
5505 io_cqring_ev_posted(ctx);
5510 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
5512 struct hlist_head *list;
5513 struct io_kiocb *req;
5515 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5516 hlist_for_each_entry(req, list, hash_node) {
5517 if (sqe_addr != req->user_data)
5519 if (io_poll_remove_one(req))
5527 static int io_poll_remove_prep(struct io_kiocb *req,
5528 const struct io_uring_sqe *sqe)
5530 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5532 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
5536 req->poll_remove.addr = READ_ONCE(sqe->addr);
5541 * Find a running poll command that matches one specified in sqe->addr,
5542 * and remove it if found.
5544 static int io_poll_remove(struct io_kiocb *req)
5546 struct io_ring_ctx *ctx = req->ctx;
5549 spin_lock_irq(&ctx->completion_lock);
5550 ret = io_poll_cancel(ctx, req->poll_remove.addr);
5551 spin_unlock_irq(&ctx->completion_lock);
5554 req_set_fail_links(req);
5555 io_req_complete(req, ret);
5559 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5562 struct io_kiocb *req = wait->private;
5563 struct io_poll_iocb *poll = &req->poll;
5565 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5568 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5569 struct poll_table_struct *p)
5571 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5573 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5576 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5578 struct io_poll_iocb *poll = &req->poll;
5581 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5583 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
5586 events = READ_ONCE(sqe->poll32_events);
5588 events = swahw32(events);
5590 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP |
5591 (events & EPOLLEXCLUSIVE);
5595 static int io_poll_add(struct io_kiocb *req)
5597 struct io_poll_iocb *poll = &req->poll;
5598 struct io_ring_ctx *ctx = req->ctx;
5599 struct io_poll_table ipt;
5602 ipt.pt._qproc = io_poll_queue_proc;
5604 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5607 if (mask) { /* no async, we'd stolen it */
5609 io_poll_complete(req, mask, 0);
5611 spin_unlock_irq(&ctx->completion_lock);
5614 io_cqring_ev_posted(ctx);
5620 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5622 struct io_timeout_data *data = container_of(timer,
5623 struct io_timeout_data, timer);
5624 struct io_kiocb *req = data->req;
5625 struct io_ring_ctx *ctx = req->ctx;
5626 unsigned long flags;
5628 spin_lock_irqsave(&ctx->completion_lock, flags);
5629 list_del_init(&req->timeout.list);
5630 atomic_set(&req->ctx->cq_timeouts,
5631 atomic_read(&req->ctx->cq_timeouts) + 1);
5633 io_cqring_fill_event(req, -ETIME);
5634 io_commit_cqring(ctx);
5635 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5637 io_cqring_ev_posted(ctx);
5638 req_set_fail_links(req);
5640 return HRTIMER_NORESTART;
5643 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5646 struct io_timeout_data *io;
5647 struct io_kiocb *req;
5650 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5651 if (user_data == req->user_data) {
5658 return ERR_PTR(ret);
5660 io = req->async_data;
5661 ret = hrtimer_try_to_cancel(&io->timer);
5663 return ERR_PTR(-EALREADY);
5664 list_del_init(&req->timeout.list);
5668 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5670 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5673 return PTR_ERR(req);
5675 req_set_fail_links(req);
5676 io_cqring_fill_event(req, -ECANCELED);
5677 io_put_req_deferred(req, 1);
5681 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5682 struct timespec64 *ts, enum hrtimer_mode mode)
5684 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5685 struct io_timeout_data *data;
5688 return PTR_ERR(req);
5690 req->timeout.off = 0; /* noseq */
5691 data = req->async_data;
5692 list_add_tail(&req->timeout.list, &ctx->timeout_list);
5693 hrtimer_init(&data->timer, CLOCK_MONOTONIC, mode);
5694 data->timer.function = io_timeout_fn;
5695 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
5699 static int io_timeout_remove_prep(struct io_kiocb *req,
5700 const struct io_uring_sqe *sqe)
5702 struct io_timeout_rem *tr = &req->timeout_rem;
5704 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5706 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5708 if (sqe->ioprio || sqe->buf_index || sqe->len)
5711 tr->addr = READ_ONCE(sqe->addr);
5712 tr->flags = READ_ONCE(sqe->timeout_flags);
5713 if (tr->flags & IORING_TIMEOUT_UPDATE) {
5714 if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS))
5716 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
5718 } else if (tr->flags) {
5719 /* timeout removal doesn't support flags */
5727 * Remove or update an existing timeout command
5729 static int io_timeout_remove(struct io_kiocb *req)
5731 struct io_timeout_rem *tr = &req->timeout_rem;
5732 struct io_ring_ctx *ctx = req->ctx;
5735 spin_lock_irq(&ctx->completion_lock);
5736 if (req->timeout_rem.flags & IORING_TIMEOUT_UPDATE) {
5737 enum hrtimer_mode mode = (tr->flags & IORING_TIMEOUT_ABS)
5738 ? HRTIMER_MODE_ABS : HRTIMER_MODE_REL;
5740 ret = io_timeout_update(ctx, tr->addr, &tr->ts, mode);
5742 ret = io_timeout_cancel(ctx, tr->addr);
5745 io_cqring_fill_event(req, ret);
5746 io_commit_cqring(ctx);
5747 spin_unlock_irq(&ctx->completion_lock);
5748 io_cqring_ev_posted(ctx);
5750 req_set_fail_links(req);
5755 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5756 bool is_timeout_link)
5758 struct io_timeout_data *data;
5760 u32 off = READ_ONCE(sqe->off);
5762 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5764 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5766 if (off && is_timeout_link)
5768 flags = READ_ONCE(sqe->timeout_flags);
5769 if (flags & ~IORING_TIMEOUT_ABS)
5772 req->timeout.off = off;
5774 if (!req->async_data && io_alloc_async_data(req))
5777 data = req->async_data;
5780 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5783 if (flags & IORING_TIMEOUT_ABS)
5784 data->mode = HRTIMER_MODE_ABS;
5786 data->mode = HRTIMER_MODE_REL;
5788 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5792 static int io_timeout(struct io_kiocb *req)
5794 struct io_ring_ctx *ctx = req->ctx;
5795 struct io_timeout_data *data = req->async_data;
5796 struct list_head *entry;
5797 u32 tail, off = req->timeout.off;
5799 spin_lock_irq(&ctx->completion_lock);
5802 * sqe->off holds how many events that need to occur for this
5803 * timeout event to be satisfied. If it isn't set, then this is
5804 * a pure timeout request, sequence isn't used.
5806 if (io_is_timeout_noseq(req)) {
5807 entry = ctx->timeout_list.prev;
5811 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5812 req->timeout.target_seq = tail + off;
5815 * Insertion sort, ensuring the first entry in the list is always
5816 * the one we need first.
5818 list_for_each_prev(entry, &ctx->timeout_list) {
5819 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5822 if (io_is_timeout_noseq(nxt))
5824 /* nxt.seq is behind @tail, otherwise would've been completed */
5825 if (off >= nxt->timeout.target_seq - tail)
5829 list_add(&req->timeout.list, entry);
5830 data->timer.function = io_timeout_fn;
5831 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5832 spin_unlock_irq(&ctx->completion_lock);
5836 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5838 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5840 return req->user_data == (unsigned long) data;
5843 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
5845 enum io_wq_cancel cancel_ret;
5848 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr, false);
5849 switch (cancel_ret) {
5850 case IO_WQ_CANCEL_OK:
5853 case IO_WQ_CANCEL_RUNNING:
5856 case IO_WQ_CANCEL_NOTFOUND:
5864 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5865 struct io_kiocb *req, __u64 sqe_addr,
5868 unsigned long flags;
5871 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
5872 if (ret != -ENOENT) {
5873 spin_lock_irqsave(&ctx->completion_lock, flags);
5877 spin_lock_irqsave(&ctx->completion_lock, flags);
5878 ret = io_timeout_cancel(ctx, sqe_addr);
5881 ret = io_poll_cancel(ctx, sqe_addr);
5885 io_cqring_fill_event(req, ret);
5886 io_commit_cqring(ctx);
5887 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5888 io_cqring_ev_posted(ctx);
5891 req_set_fail_links(req);
5895 static int io_async_cancel_prep(struct io_kiocb *req,
5896 const struct io_uring_sqe *sqe)
5898 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5900 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5902 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5905 req->cancel.addr = READ_ONCE(sqe->addr);
5909 static int io_async_cancel(struct io_kiocb *req)
5911 struct io_ring_ctx *ctx = req->ctx;
5913 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
5917 static int io_files_update_prep(struct io_kiocb *req,
5918 const struct io_uring_sqe *sqe)
5920 if (unlikely(req->ctx->flags & IORING_SETUP_SQPOLL))
5922 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5924 if (sqe->ioprio || sqe->rw_flags)
5927 req->files_update.offset = READ_ONCE(sqe->off);
5928 req->files_update.nr_args = READ_ONCE(sqe->len);
5929 if (!req->files_update.nr_args)
5931 req->files_update.arg = READ_ONCE(sqe->addr);
5935 static int io_files_update(struct io_kiocb *req, bool force_nonblock,
5936 struct io_comp_state *cs)
5938 struct io_ring_ctx *ctx = req->ctx;
5939 struct io_uring_files_update up;
5945 up.offset = req->files_update.offset;
5946 up.fds = req->files_update.arg;
5948 mutex_lock(&ctx->uring_lock);
5949 ret = __io_sqe_files_update(ctx, &up, req->files_update.nr_args);
5950 mutex_unlock(&ctx->uring_lock);
5953 req_set_fail_links(req);
5954 __io_req_complete(req, ret, 0, cs);
5958 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5960 switch (req->opcode) {
5963 case IORING_OP_READV:
5964 case IORING_OP_READ_FIXED:
5965 case IORING_OP_READ:
5966 return io_read_prep(req, sqe);
5967 case IORING_OP_WRITEV:
5968 case IORING_OP_WRITE_FIXED:
5969 case IORING_OP_WRITE:
5970 return io_write_prep(req, sqe);
5971 case IORING_OP_POLL_ADD:
5972 return io_poll_add_prep(req, sqe);
5973 case IORING_OP_POLL_REMOVE:
5974 return io_poll_remove_prep(req, sqe);
5975 case IORING_OP_FSYNC:
5976 return io_prep_fsync(req, sqe);
5977 case IORING_OP_SYNC_FILE_RANGE:
5978 return io_prep_sfr(req, sqe);
5979 case IORING_OP_SENDMSG:
5980 case IORING_OP_SEND:
5981 return io_sendmsg_prep(req, sqe);
5982 case IORING_OP_RECVMSG:
5983 case IORING_OP_RECV:
5984 return io_recvmsg_prep(req, sqe);
5985 case IORING_OP_CONNECT:
5986 return io_connect_prep(req, sqe);
5987 case IORING_OP_TIMEOUT:
5988 return io_timeout_prep(req, sqe, false);
5989 case IORING_OP_TIMEOUT_REMOVE:
5990 return io_timeout_remove_prep(req, sqe);
5991 case IORING_OP_ASYNC_CANCEL:
5992 return io_async_cancel_prep(req, sqe);
5993 case IORING_OP_LINK_TIMEOUT:
5994 return io_timeout_prep(req, sqe, true);
5995 case IORING_OP_ACCEPT:
5996 return io_accept_prep(req, sqe);
5997 case IORING_OP_FALLOCATE:
5998 return io_fallocate_prep(req, sqe);
5999 case IORING_OP_OPENAT:
6000 return io_openat_prep(req, sqe);
6001 case IORING_OP_CLOSE:
6002 return io_close_prep(req, sqe);
6003 case IORING_OP_FILES_UPDATE:
6004 return io_files_update_prep(req, sqe);
6005 case IORING_OP_STATX:
6006 return io_statx_prep(req, sqe);
6007 case IORING_OP_FADVISE:
6008 return io_fadvise_prep(req, sqe);
6009 case IORING_OP_MADVISE:
6010 return io_madvise_prep(req, sqe);
6011 case IORING_OP_OPENAT2:
6012 return io_openat2_prep(req, sqe);
6013 case IORING_OP_EPOLL_CTL:
6014 return io_epoll_ctl_prep(req, sqe);
6015 case IORING_OP_SPLICE:
6016 return io_splice_prep(req, sqe);
6017 case IORING_OP_PROVIDE_BUFFERS:
6018 return io_provide_buffers_prep(req, sqe);
6019 case IORING_OP_REMOVE_BUFFERS:
6020 return io_remove_buffers_prep(req, sqe);
6022 return io_tee_prep(req, sqe);
6023 case IORING_OP_SHUTDOWN:
6024 return io_shutdown_prep(req, sqe);
6025 case IORING_OP_RENAMEAT:
6026 return io_renameat_prep(req, sqe);
6027 case IORING_OP_UNLINKAT:
6028 return io_unlinkat_prep(req, sqe);
6031 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
6036 static int io_req_defer_prep(struct io_kiocb *req,
6037 const struct io_uring_sqe *sqe)
6041 if (io_alloc_async_data(req))
6043 return io_req_prep(req, sqe);
6046 static u32 io_get_sequence(struct io_kiocb *req)
6048 struct io_kiocb *pos;
6049 struct io_ring_ctx *ctx = req->ctx;
6050 u32 total_submitted, nr_reqs = 0;
6052 io_for_each_link(pos, req)
6055 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
6056 return total_submitted - nr_reqs;
6059 static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe)
6061 struct io_ring_ctx *ctx = req->ctx;
6062 struct io_defer_entry *de;
6066 /* Still need defer if there is pending req in defer list. */
6067 if (likely(list_empty_careful(&ctx->defer_list) &&
6068 !(req->flags & REQ_F_IO_DRAIN)))
6071 seq = io_get_sequence(req);
6072 /* Still a chance to pass the sequence check */
6073 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
6076 if (!req->async_data) {
6077 ret = io_req_defer_prep(req, sqe);
6081 io_prep_async_link(req);
6082 de = kmalloc(sizeof(*de), GFP_KERNEL);
6086 spin_lock_irq(&ctx->completion_lock);
6087 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
6088 spin_unlock_irq(&ctx->completion_lock);
6090 io_queue_async_work(req);
6091 return -EIOCBQUEUED;
6094 trace_io_uring_defer(ctx, req, req->user_data);
6097 list_add_tail(&de->list, &ctx->defer_list);
6098 spin_unlock_irq(&ctx->completion_lock);
6099 return -EIOCBQUEUED;
6102 static void io_req_drop_files(struct io_kiocb *req)
6104 struct io_ring_ctx *ctx = req->ctx;
6105 struct io_uring_task *tctx = req->task->io_uring;
6106 unsigned long flags;
6108 spin_lock_irqsave(&ctx->inflight_lock, flags);
6109 list_del(&req->inflight_entry);
6110 if (atomic_read(&tctx->in_idle))
6111 wake_up(&tctx->wait);
6112 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
6113 req->flags &= ~REQ_F_INFLIGHT;
6114 put_files_struct(req->work.identity->files);
6115 put_nsproxy(req->work.identity->nsproxy);
6116 req->work.flags &= ~IO_WQ_WORK_FILES;
6119 static void __io_clean_op(struct io_kiocb *req)
6121 if (req->flags & REQ_F_BUFFER_SELECTED) {
6122 switch (req->opcode) {
6123 case IORING_OP_READV:
6124 case IORING_OP_READ_FIXED:
6125 case IORING_OP_READ:
6126 kfree((void *)(unsigned long)req->rw.addr);
6128 case IORING_OP_RECVMSG:
6129 case IORING_OP_RECV:
6130 kfree(req->sr_msg.kbuf);
6133 req->flags &= ~REQ_F_BUFFER_SELECTED;
6136 if (req->flags & REQ_F_NEED_CLEANUP) {
6137 switch (req->opcode) {
6138 case IORING_OP_READV:
6139 case IORING_OP_READ_FIXED:
6140 case IORING_OP_READ:
6141 case IORING_OP_WRITEV:
6142 case IORING_OP_WRITE_FIXED:
6143 case IORING_OP_WRITE: {
6144 struct io_async_rw *io = req->async_data;
6146 kfree(io->free_iovec);
6149 case IORING_OP_RECVMSG:
6150 case IORING_OP_SENDMSG: {
6151 struct io_async_msghdr *io = req->async_data;
6152 if (io->iov != io->fast_iov)
6156 case IORING_OP_SPLICE:
6158 io_put_file(req, req->splice.file_in,
6159 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
6161 case IORING_OP_OPENAT:
6162 case IORING_OP_OPENAT2:
6163 if (req->open.filename)
6164 putname(req->open.filename);
6166 case IORING_OP_RENAMEAT:
6167 putname(req->rename.oldpath);
6168 putname(req->rename.newpath);
6170 case IORING_OP_UNLINKAT:
6171 putname(req->unlink.filename);
6174 req->flags &= ~REQ_F_NEED_CLEANUP;
6177 if (req->flags & REQ_F_INFLIGHT)
6178 io_req_drop_files(req);
6181 static int io_issue_sqe(struct io_kiocb *req, bool force_nonblock,
6182 struct io_comp_state *cs)
6184 struct io_ring_ctx *ctx = req->ctx;
6187 switch (req->opcode) {
6189 ret = io_nop(req, cs);
6191 case IORING_OP_READV:
6192 case IORING_OP_READ_FIXED:
6193 case IORING_OP_READ:
6194 ret = io_read(req, force_nonblock, cs);
6196 case IORING_OP_WRITEV:
6197 case IORING_OP_WRITE_FIXED:
6198 case IORING_OP_WRITE:
6199 ret = io_write(req, force_nonblock, cs);
6201 case IORING_OP_FSYNC:
6202 ret = io_fsync(req, force_nonblock);
6204 case IORING_OP_POLL_ADD:
6205 ret = io_poll_add(req);
6207 case IORING_OP_POLL_REMOVE:
6208 ret = io_poll_remove(req);
6210 case IORING_OP_SYNC_FILE_RANGE:
6211 ret = io_sync_file_range(req, force_nonblock);
6213 case IORING_OP_SENDMSG:
6214 ret = io_sendmsg(req, force_nonblock, cs);
6216 case IORING_OP_SEND:
6217 ret = io_send(req, force_nonblock, cs);
6219 case IORING_OP_RECVMSG:
6220 ret = io_recvmsg(req, force_nonblock, cs);
6222 case IORING_OP_RECV:
6223 ret = io_recv(req, force_nonblock, cs);
6225 case IORING_OP_TIMEOUT:
6226 ret = io_timeout(req);
6228 case IORING_OP_TIMEOUT_REMOVE:
6229 ret = io_timeout_remove(req);
6231 case IORING_OP_ACCEPT:
6232 ret = io_accept(req, force_nonblock, cs);
6234 case IORING_OP_CONNECT:
6235 ret = io_connect(req, force_nonblock, cs);
6237 case IORING_OP_ASYNC_CANCEL:
6238 ret = io_async_cancel(req);
6240 case IORING_OP_FALLOCATE:
6241 ret = io_fallocate(req, force_nonblock);
6243 case IORING_OP_OPENAT:
6244 ret = io_openat(req, force_nonblock);
6246 case IORING_OP_CLOSE:
6247 ret = io_close(req, force_nonblock, cs);
6249 case IORING_OP_FILES_UPDATE:
6250 ret = io_files_update(req, force_nonblock, cs);
6252 case IORING_OP_STATX:
6253 ret = io_statx(req, force_nonblock);
6255 case IORING_OP_FADVISE:
6256 ret = io_fadvise(req, force_nonblock);
6258 case IORING_OP_MADVISE:
6259 ret = io_madvise(req, force_nonblock);
6261 case IORING_OP_OPENAT2:
6262 ret = io_openat2(req, force_nonblock);
6264 case IORING_OP_EPOLL_CTL:
6265 ret = io_epoll_ctl(req, force_nonblock, cs);
6267 case IORING_OP_SPLICE:
6268 ret = io_splice(req, force_nonblock);
6270 case IORING_OP_PROVIDE_BUFFERS:
6271 ret = io_provide_buffers(req, force_nonblock, cs);
6273 case IORING_OP_REMOVE_BUFFERS:
6274 ret = io_remove_buffers(req, force_nonblock, cs);
6277 ret = io_tee(req, force_nonblock);
6279 case IORING_OP_SHUTDOWN:
6280 ret = io_shutdown(req, force_nonblock);
6282 case IORING_OP_RENAMEAT:
6283 ret = io_renameat(req, force_nonblock);
6285 case IORING_OP_UNLINKAT:
6286 ret = io_unlinkat(req, force_nonblock);
6296 /* If the op doesn't have a file, we're not polling for it */
6297 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
6298 const bool in_async = io_wq_current_is_worker();
6300 /* workqueue context doesn't hold uring_lock, grab it now */
6302 mutex_lock(&ctx->uring_lock);
6304 io_iopoll_req_issued(req, in_async);
6307 mutex_unlock(&ctx->uring_lock);
6313 static struct io_wq_work *io_wq_submit_work(struct io_wq_work *work)
6315 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6316 struct io_kiocb *timeout;
6319 timeout = io_prep_linked_timeout(req);
6321 io_queue_linked_timeout(timeout);
6323 /* if NO_CANCEL is set, we must still run the work */
6324 if ((work->flags & (IO_WQ_WORK_CANCEL|IO_WQ_WORK_NO_CANCEL)) ==
6325 IO_WQ_WORK_CANCEL) {
6331 ret = io_issue_sqe(req, false, NULL);
6333 * We can get EAGAIN for polled IO even though we're
6334 * forcing a sync submission from here, since we can't
6335 * wait for request slots on the block side.
6345 * io_iopoll_complete() does not hold completion_lock to complete
6346 * polled io, so here for polled io, just mark it done and still let
6347 * io_iopoll_complete() complete it.
6349 if (req->ctx->flags & IORING_SETUP_IOPOLL) {
6350 struct kiocb *kiocb = &req->rw.kiocb;
6352 kiocb_done(kiocb, ret, NULL);
6354 req_set_fail_links(req);
6355 io_req_complete(req, ret);
6359 return io_steal_work(req);
6362 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6365 struct fixed_file_table *table;
6367 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
6368 return table->files[index & IORING_FILE_TABLE_MASK];
6371 static struct file *io_file_get(struct io_submit_state *state,
6372 struct io_kiocb *req, int fd, bool fixed)
6374 struct io_ring_ctx *ctx = req->ctx;
6378 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6380 fd = array_index_nospec(fd, ctx->nr_user_files);
6381 file = io_file_from_index(ctx, fd);
6382 io_set_resource_node(req);
6384 trace_io_uring_file_get(ctx, fd);
6385 file = __io_file_get(state, fd);
6391 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6393 struct io_timeout_data *data = container_of(timer,
6394 struct io_timeout_data, timer);
6395 struct io_kiocb *prev, *req = data->req;
6396 struct io_ring_ctx *ctx = req->ctx;
6397 unsigned long flags;
6399 spin_lock_irqsave(&ctx->completion_lock, flags);
6400 prev = req->timeout.head;
6401 req->timeout.head = NULL;
6404 * We don't expect the list to be empty, that will only happen if we
6405 * race with the completion of the linked work.
6407 if (prev && refcount_inc_not_zero(&prev->refs))
6408 io_remove_next_linked(prev);
6411 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6414 req_set_fail_links(prev);
6415 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6418 io_req_complete(req, -ETIME);
6420 return HRTIMER_NORESTART;
6423 static void __io_queue_linked_timeout(struct io_kiocb *req)
6426 * If the back reference is NULL, then our linked request finished
6427 * before we got a chance to setup the timer
6429 if (req->timeout.head) {
6430 struct io_timeout_data *data = req->async_data;
6432 data->timer.function = io_link_timeout_fn;
6433 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6438 static void io_queue_linked_timeout(struct io_kiocb *req)
6440 struct io_ring_ctx *ctx = req->ctx;
6442 spin_lock_irq(&ctx->completion_lock);
6443 __io_queue_linked_timeout(req);
6444 spin_unlock_irq(&ctx->completion_lock);
6446 /* drop submission reference */
6450 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6452 struct io_kiocb *nxt = req->link;
6454 if (!nxt || (req->flags & REQ_F_LINK_TIMEOUT) ||
6455 nxt->opcode != IORING_OP_LINK_TIMEOUT)
6458 nxt->timeout.head = req;
6459 nxt->flags |= REQ_F_LTIMEOUT_ACTIVE;
6460 req->flags |= REQ_F_LINK_TIMEOUT;
6464 static void __io_queue_sqe(struct io_kiocb *req, struct io_comp_state *cs)
6466 struct io_kiocb *linked_timeout;
6467 const struct cred *old_creds = NULL;
6471 linked_timeout = io_prep_linked_timeout(req);
6473 if ((req->flags & REQ_F_WORK_INITIALIZED) &&
6474 (req->work.flags & IO_WQ_WORK_CREDS) &&
6475 req->work.identity->creds != current_cred()) {
6477 revert_creds(old_creds);
6478 if (old_creds == req->work.identity->creds)
6479 old_creds = NULL; /* restored original creds */
6481 old_creds = override_creds(req->work.identity->creds);
6484 ret = io_issue_sqe(req, true, cs);
6487 * We async punt it if the file wasn't marked NOWAIT, or if the file
6488 * doesn't support non-blocking read/write attempts
6490 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6491 if (!io_arm_poll_handler(req)) {
6493 * Queued up for async execution, worker will release
6494 * submit reference when the iocb is actually submitted.
6496 io_queue_async_work(req);
6500 io_queue_linked_timeout(linked_timeout);
6501 } else if (likely(!ret)) {
6502 /* drop submission reference */
6503 req = io_put_req_find_next(req);
6505 io_queue_linked_timeout(linked_timeout);
6508 if (!(req->flags & REQ_F_FORCE_ASYNC))
6510 io_queue_async_work(req);
6513 /* un-prep timeout, so it'll be killed as any other linked */
6514 req->flags &= ~REQ_F_LINK_TIMEOUT;
6515 req_set_fail_links(req);
6517 io_req_complete(req, ret);
6521 revert_creds(old_creds);
6524 static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6525 struct io_comp_state *cs)
6529 ret = io_req_defer(req, sqe);
6531 if (ret != -EIOCBQUEUED) {
6533 req_set_fail_links(req);
6535 io_req_complete(req, ret);
6537 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6538 if (!req->async_data) {
6539 ret = io_req_defer_prep(req, sqe);
6543 io_queue_async_work(req);
6546 ret = io_req_prep(req, sqe);
6550 __io_queue_sqe(req, cs);
6554 static inline void io_queue_link_head(struct io_kiocb *req,
6555 struct io_comp_state *cs)
6557 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
6559 io_req_complete(req, -ECANCELED);
6561 io_queue_sqe(req, NULL, cs);
6564 struct io_submit_link {
6565 struct io_kiocb *head;
6566 struct io_kiocb *last;
6569 static int io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6570 struct io_submit_link *link, struct io_comp_state *cs)
6572 struct io_ring_ctx *ctx = req->ctx;
6576 * If we already have a head request, queue this one for async
6577 * submittal once the head completes. If we don't have a head but
6578 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6579 * submitted sync once the chain is complete. If none of those
6580 * conditions are true (normal request), then just queue it.
6583 struct io_kiocb *head = link->head;
6586 * Taking sequential execution of a link, draining both sides
6587 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6588 * requests in the link. So, it drains the head and the
6589 * next after the link request. The last one is done via
6590 * drain_next flag to persist the effect across calls.
6592 if (req->flags & REQ_F_IO_DRAIN) {
6593 head->flags |= REQ_F_IO_DRAIN;
6594 ctx->drain_next = 1;
6596 ret = io_req_defer_prep(req, sqe);
6597 if (unlikely(ret)) {
6598 /* fail even hard links since we don't submit */
6599 head->flags |= REQ_F_FAIL_LINK;
6602 trace_io_uring_link(ctx, req, head);
6603 link->last->link = req;
6606 /* last request of a link, enqueue the link */
6607 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6608 io_queue_link_head(head, cs);
6612 if (unlikely(ctx->drain_next)) {
6613 req->flags |= REQ_F_IO_DRAIN;
6614 ctx->drain_next = 0;
6616 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6617 ret = io_req_defer_prep(req, sqe);
6619 req->flags |= REQ_F_FAIL_LINK;
6623 io_queue_sqe(req, sqe, cs);
6631 * Batched submission is done, ensure local IO is flushed out.
6633 static void io_submit_state_end(struct io_submit_state *state)
6635 if (!list_empty(&state->comp.list))
6636 io_submit_flush_completions(&state->comp);
6637 if (state->plug_started)
6638 blk_finish_plug(&state->plug);
6639 io_state_file_put(state);
6640 if (state->free_reqs)
6641 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
6645 * Start submission side cache.
6647 static void io_submit_state_start(struct io_submit_state *state,
6648 struct io_ring_ctx *ctx, unsigned int max_ios)
6650 state->plug_started = false;
6652 INIT_LIST_HEAD(&state->comp.list);
6653 state->comp.ctx = ctx;
6654 state->free_reqs = 0;
6655 state->file_refs = 0;
6656 state->ios_left = max_ios;
6659 static void io_commit_sqring(struct io_ring_ctx *ctx)
6661 struct io_rings *rings = ctx->rings;
6664 * Ensure any loads from the SQEs are done at this point,
6665 * since once we write the new head, the application could
6666 * write new data to them.
6668 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6672 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6673 * that is mapped by userspace. This means that care needs to be taken to
6674 * ensure that reads are stable, as we cannot rely on userspace always
6675 * being a good citizen. If members of the sqe are validated and then later
6676 * used, it's important that those reads are done through READ_ONCE() to
6677 * prevent a re-load down the line.
6679 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6681 u32 *sq_array = ctx->sq_array;
6685 * The cached sq head (or cq tail) serves two purposes:
6687 * 1) allows us to batch the cost of updating the user visible
6689 * 2) allows the kernel side to track the head on its own, even
6690 * though the application is the one updating it.
6692 head = READ_ONCE(sq_array[ctx->cached_sq_head & ctx->sq_mask]);
6693 if (likely(head < ctx->sq_entries))
6694 return &ctx->sq_sqes[head];
6696 /* drop invalid entries */
6697 ctx->cached_sq_dropped++;
6698 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6702 static inline void io_consume_sqe(struct io_ring_ctx *ctx)
6704 ctx->cached_sq_head++;
6708 * Check SQE restrictions (opcode and flags).
6710 * Returns 'true' if SQE is allowed, 'false' otherwise.
6712 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6713 struct io_kiocb *req,
6714 unsigned int sqe_flags)
6716 if (!ctx->restricted)
6719 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6722 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6723 ctx->restrictions.sqe_flags_required)
6726 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6727 ctx->restrictions.sqe_flags_required))
6733 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
6734 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
6735 IOSQE_BUFFER_SELECT)
6737 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6738 const struct io_uring_sqe *sqe,
6739 struct io_submit_state *state)
6741 unsigned int sqe_flags;
6744 req->opcode = READ_ONCE(sqe->opcode);
6745 req->user_data = READ_ONCE(sqe->user_data);
6746 req->async_data = NULL;
6751 req->fixed_file_refs = NULL;
6752 /* one is dropped after submission, the other at completion */
6753 refcount_set(&req->refs, 2);
6754 req->task = current;
6757 if (unlikely(req->opcode >= IORING_OP_LAST))
6760 if (unlikely(io_sq_thread_acquire_mm_files(ctx, req)))
6763 sqe_flags = READ_ONCE(sqe->flags);
6764 /* enforce forwards compatibility on users */
6765 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6768 if (unlikely(!io_check_restriction(ctx, req, sqe_flags)))
6771 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6772 !io_op_defs[req->opcode].buffer_select)
6775 id = READ_ONCE(sqe->personality);
6777 struct io_identity *iod;
6779 iod = idr_find(&ctx->personality_idr, id);
6782 refcount_inc(&iod->count);
6784 __io_req_init_async(req);
6785 get_cred(iod->creds);
6786 req->work.identity = iod;
6787 req->work.flags |= IO_WQ_WORK_CREDS;
6790 /* same numerical values with corresponding REQ_F_*, safe to copy */
6791 req->flags |= sqe_flags;
6794 * Plug now if we have more than 1 IO left after this, and the target
6795 * is potentially a read/write to block based storage.
6797 if (!state->plug_started && state->ios_left > 1 &&
6798 io_op_defs[req->opcode].plug) {
6799 blk_start_plug(&state->plug);
6800 state->plug_started = true;
6804 if (io_op_defs[req->opcode].needs_file) {
6805 bool fixed = req->flags & REQ_F_FIXED_FILE;
6807 req->file = io_file_get(state, req, READ_ONCE(sqe->fd), fixed);
6808 if (unlikely(!req->file &&
6809 !io_op_defs[req->opcode].needs_file_no_error))
6817 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6819 struct io_submit_state state;
6820 struct io_submit_link link;
6821 int i, submitted = 0;
6823 /* if we have a backlog and couldn't flush it all, return BUSY */
6824 if (test_bit(0, &ctx->sq_check_overflow)) {
6825 if (!list_empty(&ctx->cq_overflow_list) &&
6826 !io_cqring_overflow_flush(ctx, false, NULL, NULL))
6830 /* make sure SQ entry isn't read before tail */
6831 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6833 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6836 percpu_counter_add(¤t->io_uring->inflight, nr);
6837 refcount_add(nr, ¤t->usage);
6839 io_submit_state_start(&state, ctx, nr);
6842 for (i = 0; i < nr; i++) {
6843 const struct io_uring_sqe *sqe;
6844 struct io_kiocb *req;
6847 sqe = io_get_sqe(ctx);
6848 if (unlikely(!sqe)) {
6849 io_consume_sqe(ctx);
6852 req = io_alloc_req(ctx, &state);
6853 if (unlikely(!req)) {
6855 submitted = -EAGAIN;
6858 io_consume_sqe(ctx);
6859 /* will complete beyond this point, count as submitted */
6862 err = io_init_req(ctx, req, sqe, &state);
6863 if (unlikely(err)) {
6866 io_req_complete(req, err);
6870 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6871 true, io_async_submit(ctx));
6872 err = io_submit_sqe(req, sqe, &link, &state.comp);
6877 if (unlikely(submitted != nr)) {
6878 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6879 struct io_uring_task *tctx = current->io_uring;
6880 int unused = nr - ref_used;
6882 percpu_ref_put_many(&ctx->refs, unused);
6883 percpu_counter_sub(&tctx->inflight, unused);
6884 put_task_struct_many(current, unused);
6887 io_queue_link_head(link.head, &state.comp);
6888 io_submit_state_end(&state);
6890 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6891 io_commit_sqring(ctx);
6896 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6898 /* Tell userspace we may need a wakeup call */
6899 spin_lock_irq(&ctx->completion_lock);
6900 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6901 spin_unlock_irq(&ctx->completion_lock);
6904 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6906 spin_lock_irq(&ctx->completion_lock);
6907 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6908 spin_unlock_irq(&ctx->completion_lock);
6911 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
6913 unsigned int to_submit;
6916 to_submit = io_sqring_entries(ctx);
6917 /* if we're handling multiple rings, cap submit size for fairness */
6918 if (cap_entries && to_submit > 8)
6921 if (!list_empty(&ctx->iopoll_list) || to_submit) {
6922 unsigned nr_events = 0;
6924 mutex_lock(&ctx->uring_lock);
6925 if (!list_empty(&ctx->iopoll_list))
6926 io_do_iopoll(ctx, &nr_events, 0);
6928 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)))
6929 ret = io_submit_sqes(ctx, to_submit);
6930 mutex_unlock(&ctx->uring_lock);
6933 if (!io_sqring_full(ctx) && wq_has_sleeper(&ctx->sqo_sq_wait))
6934 wake_up(&ctx->sqo_sq_wait);
6939 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
6941 struct io_ring_ctx *ctx;
6942 unsigned sq_thread_idle = 0;
6944 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6945 if (sq_thread_idle < ctx->sq_thread_idle)
6946 sq_thread_idle = ctx->sq_thread_idle;
6949 sqd->sq_thread_idle = sq_thread_idle;
6952 static void io_sqd_init_new(struct io_sq_data *sqd)
6954 struct io_ring_ctx *ctx;
6956 while (!list_empty(&sqd->ctx_new_list)) {
6957 ctx = list_first_entry(&sqd->ctx_new_list, struct io_ring_ctx, sqd_list);
6958 list_move_tail(&ctx->sqd_list, &sqd->ctx_list);
6959 complete(&ctx->sq_thread_comp);
6962 io_sqd_update_thread_idle(sqd);
6965 static int io_sq_thread(void *data)
6967 struct cgroup_subsys_state *cur_css = NULL;
6968 struct files_struct *old_files = current->files;
6969 struct nsproxy *old_nsproxy = current->nsproxy;
6970 const struct cred *old_cred = NULL;
6971 struct io_sq_data *sqd = data;
6972 struct io_ring_ctx *ctx;
6973 unsigned long timeout = 0;
6977 current->files = NULL;
6978 current->nsproxy = NULL;
6979 task_unlock(current);
6981 while (!kthread_should_stop()) {
6983 bool cap_entries, sqt_spin, needs_sched;
6986 * Any changes to the sqd lists are synchronized through the
6987 * kthread parking. This synchronizes the thread vs users,
6988 * the users are synchronized on the sqd->ctx_lock.
6990 if (kthread_should_park()) {
6993 * When sq thread is unparked, in case the previous park operation
6994 * comes from io_put_sq_data(), which means that sq thread is going
6995 * to be stopped, so here needs to have a check.
6997 if (kthread_should_stop())
7001 if (unlikely(!list_empty(&sqd->ctx_new_list))) {
7002 io_sqd_init_new(sqd);
7003 timeout = jiffies + sqd->sq_thread_idle;
7007 cap_entries = !list_is_singular(&sqd->ctx_list);
7008 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
7009 if (current->cred != ctx->creds) {
7011 revert_creds(old_cred);
7012 old_cred = override_creds(ctx->creds);
7014 io_sq_thread_associate_blkcg(ctx, &cur_css);
7016 current->loginuid = ctx->loginuid;
7017 current->sessionid = ctx->sessionid;
7020 ret = __io_sq_thread(ctx, cap_entries);
7021 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
7024 io_sq_thread_drop_mm_files();
7027 if (sqt_spin || !time_after(jiffies, timeout)) {
7031 timeout = jiffies + sqd->sq_thread_idle;
7035 if (kthread_should_park())
7039 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
7040 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
7041 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
7042 !list_empty_careful(&ctx->iopoll_list)) {
7043 needs_sched = false;
7046 if (io_sqring_entries(ctx)) {
7047 needs_sched = false;
7053 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7054 io_ring_set_wakeup_flag(ctx);
7057 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7058 io_ring_clear_wakeup_flag(ctx);
7061 finish_wait(&sqd->wait, &wait);
7062 timeout = jiffies + sqd->sq_thread_idle;
7068 io_sq_thread_unassociate_blkcg();
7070 revert_creds(old_cred);
7073 current->files = old_files;
7074 current->nsproxy = old_nsproxy;
7075 task_unlock(current);
7082 struct io_wait_queue {
7083 struct wait_queue_entry wq;
7084 struct io_ring_ctx *ctx;
7086 unsigned nr_timeouts;
7089 static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush)
7091 struct io_ring_ctx *ctx = iowq->ctx;
7094 * Wake up if we have enough events, or if a timeout occurred since we
7095 * started waiting. For timeouts, we always want to return to userspace,
7096 * regardless of event count.
7098 return io_cqring_events(ctx, noflush) >= iowq->to_wait ||
7099 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
7102 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
7103 int wake_flags, void *key)
7105 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
7108 /* use noflush == true, as we can't safely rely on locking context */
7109 if (!io_should_wake(iowq, true))
7112 return autoremove_wake_function(curr, mode, wake_flags, key);
7115 static int io_run_task_work_sig(void)
7117 if (io_run_task_work())
7119 if (!signal_pending(current))
7121 if (test_tsk_thread_flag(current, TIF_NOTIFY_SIGNAL))
7122 return -ERESTARTSYS;
7127 * Wait until events become available, if we don't already have some. The
7128 * application must reap them itself, as they reside on the shared cq ring.
7130 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
7131 const sigset_t __user *sig, size_t sigsz,
7132 struct __kernel_timespec __user *uts)
7134 struct io_wait_queue iowq = {
7137 .func = io_wake_function,
7138 .entry = LIST_HEAD_INIT(iowq.wq.entry),
7141 .to_wait = min_events,
7143 struct io_rings *rings = ctx->rings;
7144 struct timespec64 ts;
7145 signed long timeout = 0;
7149 if (io_cqring_events(ctx, false) >= min_events)
7151 if (!io_run_task_work())
7156 #ifdef CONFIG_COMPAT
7157 if (in_compat_syscall())
7158 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
7162 ret = set_user_sigmask(sig, sigsz);
7169 if (get_timespec64(&ts, uts))
7171 timeout = timespec64_to_jiffies(&ts);
7174 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
7175 trace_io_uring_cqring_wait(ctx, min_events);
7177 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
7178 TASK_INTERRUPTIBLE);
7179 /* make sure we run task_work before checking for signals */
7180 ret = io_run_task_work_sig();
7185 if (io_should_wake(&iowq, false))
7188 timeout = schedule_timeout(timeout);
7197 finish_wait(&ctx->wait, &iowq.wq);
7199 restore_saved_sigmask_unless(ret == -EINTR);
7201 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
7204 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
7206 #if defined(CONFIG_UNIX)
7207 if (ctx->ring_sock) {
7208 struct sock *sock = ctx->ring_sock->sk;
7209 struct sk_buff *skb;
7211 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
7217 for (i = 0; i < ctx->nr_user_files; i++) {
7220 file = io_file_from_index(ctx, i);
7227 static void io_file_ref_kill(struct percpu_ref *ref)
7229 struct fixed_file_data *data;
7231 data = container_of(ref, struct fixed_file_data, refs);
7232 complete(&data->done);
7235 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7237 struct fixed_file_data *data = ctx->file_data;
7238 struct fixed_file_ref_node *ref_node = NULL;
7239 unsigned nr_tables, i;
7244 spin_lock_bh(&data->lock);
7245 ref_node = data->node;
7246 spin_unlock_bh(&data->lock);
7248 percpu_ref_kill(&ref_node->refs);
7250 percpu_ref_kill(&data->refs);
7252 /* wait for all refs nodes to complete */
7253 flush_delayed_work(&ctx->file_put_work);
7254 wait_for_completion(&data->done);
7256 __io_sqe_files_unregister(ctx);
7257 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
7258 for (i = 0; i < nr_tables; i++)
7259 kfree(data->table[i].files);
7261 percpu_ref_exit(&data->refs);
7263 ctx->file_data = NULL;
7264 ctx->nr_user_files = 0;
7268 static void io_put_sq_data(struct io_sq_data *sqd)
7270 if (refcount_dec_and_test(&sqd->refs)) {
7272 * The park is a bit of a work-around, without it we get
7273 * warning spews on shutdown with SQPOLL set and affinity
7274 * set to a single CPU.
7277 kthread_park(sqd->thread);
7278 kthread_stop(sqd->thread);
7285 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7287 struct io_ring_ctx *ctx_attach;
7288 struct io_sq_data *sqd;
7291 f = fdget(p->wq_fd);
7293 return ERR_PTR(-ENXIO);
7294 if (f.file->f_op != &io_uring_fops) {
7296 return ERR_PTR(-EINVAL);
7299 ctx_attach = f.file->private_data;
7300 sqd = ctx_attach->sq_data;
7303 return ERR_PTR(-EINVAL);
7306 refcount_inc(&sqd->refs);
7311 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p)
7313 struct io_sq_data *sqd;
7315 if (p->flags & IORING_SETUP_ATTACH_WQ)
7316 return io_attach_sq_data(p);
7318 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7320 return ERR_PTR(-ENOMEM);
7322 refcount_set(&sqd->refs, 1);
7323 INIT_LIST_HEAD(&sqd->ctx_list);
7324 INIT_LIST_HEAD(&sqd->ctx_new_list);
7325 mutex_init(&sqd->ctx_lock);
7326 mutex_init(&sqd->lock);
7327 init_waitqueue_head(&sqd->wait);
7331 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7332 __releases(&sqd->lock)
7336 kthread_unpark(sqd->thread);
7337 mutex_unlock(&sqd->lock);
7340 static void io_sq_thread_park(struct io_sq_data *sqd)
7341 __acquires(&sqd->lock)
7345 mutex_lock(&sqd->lock);
7346 kthread_park(sqd->thread);
7349 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
7351 struct io_sq_data *sqd = ctx->sq_data;
7356 * We may arrive here from the error branch in
7357 * io_sq_offload_create() where the kthread is created
7358 * without being waked up, thus wake it up now to make
7359 * sure the wait will complete.
7361 wake_up_process(sqd->thread);
7362 wait_for_completion(&ctx->sq_thread_comp);
7364 io_sq_thread_park(sqd);
7367 mutex_lock(&sqd->ctx_lock);
7368 list_del(&ctx->sqd_list);
7369 io_sqd_update_thread_idle(sqd);
7370 mutex_unlock(&sqd->ctx_lock);
7373 io_sq_thread_unpark(sqd);
7375 io_put_sq_data(sqd);
7376 ctx->sq_data = NULL;
7380 static void io_finish_async(struct io_ring_ctx *ctx)
7382 io_sq_thread_stop(ctx);
7385 io_wq_destroy(ctx->io_wq);
7390 #if defined(CONFIG_UNIX)
7392 * Ensure the UNIX gc is aware of our file set, so we are certain that
7393 * the io_uring can be safely unregistered on process exit, even if we have
7394 * loops in the file referencing.
7396 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7398 struct sock *sk = ctx->ring_sock->sk;
7399 struct scm_fp_list *fpl;
7400 struct sk_buff *skb;
7403 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7407 skb = alloc_skb(0, GFP_KERNEL);
7416 fpl->user = get_uid(ctx->user);
7417 for (i = 0; i < nr; i++) {
7418 struct file *file = io_file_from_index(ctx, i + offset);
7422 fpl->fp[nr_files] = get_file(file);
7423 unix_inflight(fpl->user, fpl->fp[nr_files]);
7428 fpl->max = SCM_MAX_FD;
7429 fpl->count = nr_files;
7430 UNIXCB(skb).fp = fpl;
7431 skb->destructor = unix_destruct_scm;
7432 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7433 skb_queue_head(&sk->sk_receive_queue, skb);
7435 for (i = 0; i < nr_files; i++)
7446 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7447 * causes regular reference counting to break down. We rely on the UNIX
7448 * garbage collection to take care of this problem for us.
7450 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7452 unsigned left, total;
7456 left = ctx->nr_user_files;
7458 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7460 ret = __io_sqe_files_scm(ctx, this_files, total);
7464 total += this_files;
7470 while (total < ctx->nr_user_files) {
7471 struct file *file = io_file_from_index(ctx, total);
7481 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7487 static int io_sqe_alloc_file_tables(struct fixed_file_data *file_data,
7488 unsigned nr_tables, unsigned nr_files)
7492 for (i = 0; i < nr_tables; i++) {
7493 struct fixed_file_table *table = &file_data->table[i];
7494 unsigned this_files;
7496 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
7497 table->files = kcalloc(this_files, sizeof(struct file *),
7501 nr_files -= this_files;
7507 for (i = 0; i < nr_tables; i++) {
7508 struct fixed_file_table *table = &file_data->table[i];
7509 kfree(table->files);
7514 static void io_ring_file_put(struct io_ring_ctx *ctx, struct file *file)
7516 #if defined(CONFIG_UNIX)
7517 struct sock *sock = ctx->ring_sock->sk;
7518 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7519 struct sk_buff *skb;
7522 __skb_queue_head_init(&list);
7525 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7526 * remove this entry and rearrange the file array.
7528 skb = skb_dequeue(head);
7530 struct scm_fp_list *fp;
7532 fp = UNIXCB(skb).fp;
7533 for (i = 0; i < fp->count; i++) {
7536 if (fp->fp[i] != file)
7539 unix_notinflight(fp->user, fp->fp[i]);
7540 left = fp->count - 1 - i;
7542 memmove(&fp->fp[i], &fp->fp[i + 1],
7543 left * sizeof(struct file *));
7550 __skb_queue_tail(&list, skb);
7560 __skb_queue_tail(&list, skb);
7562 skb = skb_dequeue(head);
7565 if (skb_peek(&list)) {
7566 spin_lock_irq(&head->lock);
7567 while ((skb = __skb_dequeue(&list)) != NULL)
7568 __skb_queue_tail(head, skb);
7569 spin_unlock_irq(&head->lock);
7576 struct io_file_put {
7577 struct list_head list;
7581 static void __io_file_put_work(struct fixed_file_ref_node *ref_node)
7583 struct fixed_file_data *file_data = ref_node->file_data;
7584 struct io_ring_ctx *ctx = file_data->ctx;
7585 struct io_file_put *pfile, *tmp;
7587 list_for_each_entry_safe(pfile, tmp, &ref_node->file_list, list) {
7588 list_del(&pfile->list);
7589 io_ring_file_put(ctx, pfile->file);
7593 percpu_ref_exit(&ref_node->refs);
7595 percpu_ref_put(&file_data->refs);
7598 static void io_file_put_work(struct work_struct *work)
7600 struct io_ring_ctx *ctx;
7601 struct llist_node *node;
7603 ctx = container_of(work, struct io_ring_ctx, file_put_work.work);
7604 node = llist_del_all(&ctx->file_put_llist);
7607 struct fixed_file_ref_node *ref_node;
7608 struct llist_node *next = node->next;
7610 ref_node = llist_entry(node, struct fixed_file_ref_node, llist);
7611 __io_file_put_work(ref_node);
7616 static void io_file_data_ref_zero(struct percpu_ref *ref)
7618 struct fixed_file_ref_node *ref_node;
7619 struct fixed_file_data *data;
7620 struct io_ring_ctx *ctx;
7621 bool first_add = false;
7624 ref_node = container_of(ref, struct fixed_file_ref_node, refs);
7625 data = ref_node->file_data;
7628 spin_lock_bh(&data->lock);
7629 ref_node->done = true;
7631 while (!list_empty(&data->ref_list)) {
7632 ref_node = list_first_entry(&data->ref_list,
7633 struct fixed_file_ref_node, node);
7634 /* recycle ref nodes in order */
7635 if (!ref_node->done)
7637 list_del(&ref_node->node);
7638 first_add |= llist_add(&ref_node->llist, &ctx->file_put_llist);
7640 spin_unlock_bh(&data->lock);
7642 if (percpu_ref_is_dying(&data->refs))
7646 mod_delayed_work(system_wq, &ctx->file_put_work, 0);
7648 queue_delayed_work(system_wq, &ctx->file_put_work, delay);
7651 static struct fixed_file_ref_node *alloc_fixed_file_ref_node(
7652 struct io_ring_ctx *ctx)
7654 struct fixed_file_ref_node *ref_node;
7656 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7658 return ERR_PTR(-ENOMEM);
7660 if (percpu_ref_init(&ref_node->refs, io_file_data_ref_zero,
7663 return ERR_PTR(-ENOMEM);
7665 INIT_LIST_HEAD(&ref_node->node);
7666 INIT_LIST_HEAD(&ref_node->file_list);
7667 ref_node->file_data = ctx->file_data;
7668 ref_node->done = false;
7672 static void destroy_fixed_file_ref_node(struct fixed_file_ref_node *ref_node)
7674 percpu_ref_exit(&ref_node->refs);
7678 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7681 __s32 __user *fds = (__s32 __user *) arg;
7682 unsigned nr_tables, i;
7684 int fd, ret = -ENOMEM;
7685 struct fixed_file_ref_node *ref_node;
7686 struct fixed_file_data *file_data;
7692 if (nr_args > IORING_MAX_FIXED_FILES)
7695 file_data = kzalloc(sizeof(*ctx->file_data), GFP_KERNEL);
7698 file_data->ctx = ctx;
7699 init_completion(&file_data->done);
7700 INIT_LIST_HEAD(&file_data->ref_list);
7701 spin_lock_init(&file_data->lock);
7703 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
7704 file_data->table = kcalloc(nr_tables, sizeof(*file_data->table),
7706 if (!file_data->table)
7709 if (percpu_ref_init(&file_data->refs, io_file_ref_kill,
7710 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
7713 if (io_sqe_alloc_file_tables(file_data, nr_tables, nr_args))
7715 ctx->file_data = file_data;
7717 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7718 struct fixed_file_table *table;
7721 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7725 /* allow sparse sets */
7735 * Don't allow io_uring instances to be registered. If UNIX
7736 * isn't enabled, then this causes a reference cycle and this
7737 * instance can never get freed. If UNIX is enabled we'll
7738 * handle it just fine, but there's still no point in allowing
7739 * a ring fd as it doesn't support regular read/write anyway.
7741 if (file->f_op == &io_uring_fops) {
7745 table = &file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7746 index = i & IORING_FILE_TABLE_MASK;
7747 table->files[index] = file;
7750 ret = io_sqe_files_scm(ctx);
7752 io_sqe_files_unregister(ctx);
7756 ref_node = alloc_fixed_file_ref_node(ctx);
7757 if (IS_ERR(ref_node)) {
7758 io_sqe_files_unregister(ctx);
7759 return PTR_ERR(ref_node);
7762 file_data->node = ref_node;
7763 spin_lock_bh(&file_data->lock);
7764 list_add_tail(&ref_node->node, &file_data->ref_list);
7765 spin_unlock_bh(&file_data->lock);
7766 percpu_ref_get(&file_data->refs);
7769 for (i = 0; i < ctx->nr_user_files; i++) {
7770 file = io_file_from_index(ctx, i);
7774 for (i = 0; i < nr_tables; i++)
7775 kfree(file_data->table[i].files);
7776 ctx->nr_user_files = 0;
7778 percpu_ref_exit(&file_data->refs);
7780 kfree(file_data->table);
7782 ctx->file_data = NULL;
7786 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7789 #if defined(CONFIG_UNIX)
7790 struct sock *sock = ctx->ring_sock->sk;
7791 struct sk_buff_head *head = &sock->sk_receive_queue;
7792 struct sk_buff *skb;
7795 * See if we can merge this file into an existing skb SCM_RIGHTS
7796 * file set. If there's no room, fall back to allocating a new skb
7797 * and filling it in.
7799 spin_lock_irq(&head->lock);
7800 skb = skb_peek(head);
7802 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7804 if (fpl->count < SCM_MAX_FD) {
7805 __skb_unlink(skb, head);
7806 spin_unlock_irq(&head->lock);
7807 fpl->fp[fpl->count] = get_file(file);
7808 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7810 spin_lock_irq(&head->lock);
7811 __skb_queue_head(head, skb);
7816 spin_unlock_irq(&head->lock);
7823 return __io_sqe_files_scm(ctx, 1, index);
7829 static int io_queue_file_removal(struct fixed_file_data *data,
7832 struct io_file_put *pfile;
7833 struct fixed_file_ref_node *ref_node = data->node;
7835 pfile = kzalloc(sizeof(*pfile), GFP_KERNEL);
7840 list_add(&pfile->list, &ref_node->file_list);
7845 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7846 struct io_uring_files_update *up,
7849 struct fixed_file_data *data = ctx->file_data;
7850 struct fixed_file_ref_node *ref_node;
7855 bool needs_switch = false;
7857 if (check_add_overflow(up->offset, nr_args, &done))
7859 if (done > ctx->nr_user_files)
7862 ref_node = alloc_fixed_file_ref_node(ctx);
7863 if (IS_ERR(ref_node))
7864 return PTR_ERR(ref_node);
7867 fds = u64_to_user_ptr(up->fds);
7869 struct fixed_file_table *table;
7873 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
7877 i = array_index_nospec(up->offset, ctx->nr_user_files);
7878 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7879 index = i & IORING_FILE_TABLE_MASK;
7880 if (table->files[index]) {
7881 file = table->files[index];
7882 err = io_queue_file_removal(data, file);
7885 table->files[index] = NULL;
7886 needs_switch = true;
7895 * Don't allow io_uring instances to be registered. If
7896 * UNIX isn't enabled, then this causes a reference
7897 * cycle and this instance can never get freed. If UNIX
7898 * is enabled we'll handle it just fine, but there's
7899 * still no point in allowing a ring fd as it doesn't
7900 * support regular read/write anyway.
7902 if (file->f_op == &io_uring_fops) {
7907 table->files[index] = file;
7908 err = io_sqe_file_register(ctx, file, i);
7910 table->files[index] = NULL;
7921 percpu_ref_kill(&data->node->refs);
7922 spin_lock_bh(&data->lock);
7923 list_add_tail(&ref_node->node, &data->ref_list);
7924 data->node = ref_node;
7925 spin_unlock_bh(&data->lock);
7926 percpu_ref_get(&ctx->file_data->refs);
7928 destroy_fixed_file_ref_node(ref_node);
7930 return done ? done : err;
7933 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
7936 struct io_uring_files_update up;
7938 if (!ctx->file_data)
7942 if (copy_from_user(&up, arg, sizeof(up)))
7947 return __io_sqe_files_update(ctx, &up, nr_args);
7950 static void io_free_work(struct io_wq_work *work)
7952 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7954 /* Consider that io_steal_work() relies on this ref */
7958 static int io_init_wq_offload(struct io_ring_ctx *ctx,
7959 struct io_uring_params *p)
7961 struct io_wq_data data;
7963 struct io_ring_ctx *ctx_attach;
7964 unsigned int concurrency;
7967 data.user = ctx->user;
7968 data.free_work = io_free_work;
7969 data.do_work = io_wq_submit_work;
7971 if (!(p->flags & IORING_SETUP_ATTACH_WQ)) {
7972 /* Do QD, or 4 * CPUS, whatever is smallest */
7973 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7975 ctx->io_wq = io_wq_create(concurrency, &data);
7976 if (IS_ERR(ctx->io_wq)) {
7977 ret = PTR_ERR(ctx->io_wq);
7983 f = fdget(p->wq_fd);
7987 if (f.file->f_op != &io_uring_fops) {
7992 ctx_attach = f.file->private_data;
7993 /* @io_wq is protected by holding the fd */
7994 if (!io_wq_get(ctx_attach->io_wq, &data)) {
7999 ctx->io_wq = ctx_attach->io_wq;
8005 static int io_uring_alloc_task_context(struct task_struct *task)
8007 struct io_uring_task *tctx;
8010 tctx = kmalloc(sizeof(*tctx), GFP_KERNEL);
8011 if (unlikely(!tctx))
8014 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
8015 if (unlikely(ret)) {
8021 init_waitqueue_head(&tctx->wait);
8023 atomic_set(&tctx->in_idle, 0);
8024 tctx->sqpoll = false;
8025 io_init_identity(&tctx->__identity);
8026 tctx->identity = &tctx->__identity;
8027 task->io_uring = tctx;
8031 void __io_uring_free(struct task_struct *tsk)
8033 struct io_uring_task *tctx = tsk->io_uring;
8035 WARN_ON_ONCE(!xa_empty(&tctx->xa));
8036 WARN_ON_ONCE(refcount_read(&tctx->identity->count) != 1);
8037 if (tctx->identity != &tctx->__identity)
8038 kfree(tctx->identity);
8039 percpu_counter_destroy(&tctx->inflight);
8041 tsk->io_uring = NULL;
8044 static int io_sq_offload_create(struct io_ring_ctx *ctx,
8045 struct io_uring_params *p)
8049 if (ctx->flags & IORING_SETUP_SQPOLL) {
8050 struct io_sq_data *sqd;
8053 if (!capable(CAP_SYS_ADMIN) && !capable(CAP_SYS_NICE))
8056 sqd = io_get_sq_data(p);
8063 io_sq_thread_park(sqd);
8064 mutex_lock(&sqd->ctx_lock);
8065 list_add(&ctx->sqd_list, &sqd->ctx_new_list);
8066 mutex_unlock(&sqd->ctx_lock);
8067 io_sq_thread_unpark(sqd);
8069 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
8070 if (!ctx->sq_thread_idle)
8071 ctx->sq_thread_idle = HZ;
8076 if (p->flags & IORING_SETUP_SQ_AFF) {
8077 int cpu = p->sq_thread_cpu;
8080 if (cpu >= nr_cpu_ids)
8082 if (!cpu_online(cpu))
8085 sqd->thread = kthread_create_on_cpu(io_sq_thread, sqd,
8086 cpu, "io_uring-sq");
8088 sqd->thread = kthread_create(io_sq_thread, sqd,
8091 if (IS_ERR(sqd->thread)) {
8092 ret = PTR_ERR(sqd->thread);
8096 ret = io_uring_alloc_task_context(sqd->thread);
8099 } else if (p->flags & IORING_SETUP_SQ_AFF) {
8100 /* Can't have SQ_AFF without SQPOLL */
8106 ret = io_init_wq_offload(ctx, p);
8112 io_finish_async(ctx);
8116 static void io_sq_offload_start(struct io_ring_ctx *ctx)
8118 struct io_sq_data *sqd = ctx->sq_data;
8120 if ((ctx->flags & IORING_SETUP_SQPOLL) && sqd->thread)
8121 wake_up_process(sqd->thread);
8124 static inline void __io_unaccount_mem(struct user_struct *user,
8125 unsigned long nr_pages)
8127 atomic_long_sub(nr_pages, &user->locked_vm);
8130 static inline int __io_account_mem(struct user_struct *user,
8131 unsigned long nr_pages)
8133 unsigned long page_limit, cur_pages, new_pages;
8135 /* Don't allow more pages than we can safely lock */
8136 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
8139 cur_pages = atomic_long_read(&user->locked_vm);
8140 new_pages = cur_pages + nr_pages;
8141 if (new_pages > page_limit)
8143 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
8144 new_pages) != cur_pages);
8149 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
8150 enum io_mem_account acct)
8153 __io_unaccount_mem(ctx->user, nr_pages);
8155 if (ctx->mm_account) {
8156 if (acct == ACCT_LOCKED) {
8157 mmap_write_lock(ctx->mm_account);
8158 ctx->mm_account->locked_vm -= nr_pages;
8159 mmap_write_unlock(ctx->mm_account);
8160 }else if (acct == ACCT_PINNED) {
8161 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
8166 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
8167 enum io_mem_account acct)
8171 if (ctx->limit_mem) {
8172 ret = __io_account_mem(ctx->user, nr_pages);
8177 if (ctx->mm_account) {
8178 if (acct == ACCT_LOCKED) {
8179 mmap_write_lock(ctx->mm_account);
8180 ctx->mm_account->locked_vm += nr_pages;
8181 mmap_write_unlock(ctx->mm_account);
8182 } else if (acct == ACCT_PINNED) {
8183 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
8190 static void io_mem_free(void *ptr)
8197 page = virt_to_head_page(ptr);
8198 if (put_page_testzero(page))
8199 free_compound_page(page);
8202 static void *io_mem_alloc(size_t size)
8204 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
8207 return (void *) __get_free_pages(gfp_flags, get_order(size));
8210 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8213 struct io_rings *rings;
8214 size_t off, sq_array_size;
8216 off = struct_size(rings, cqes, cq_entries);
8217 if (off == SIZE_MAX)
8221 off = ALIGN(off, SMP_CACHE_BYTES);
8229 sq_array_size = array_size(sizeof(u32), sq_entries);
8230 if (sq_array_size == SIZE_MAX)
8233 if (check_add_overflow(off, sq_array_size, &off))
8239 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
8243 pages = (size_t)1 << get_order(
8244 rings_size(sq_entries, cq_entries, NULL));
8245 pages += (size_t)1 << get_order(
8246 array_size(sizeof(struct io_uring_sqe), sq_entries));
8251 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
8255 if (!ctx->user_bufs)
8258 for (i = 0; i < ctx->nr_user_bufs; i++) {
8259 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8261 for (j = 0; j < imu->nr_bvecs; j++)
8262 unpin_user_page(imu->bvec[j].bv_page);
8264 if (imu->acct_pages)
8265 io_unaccount_mem(ctx, imu->acct_pages, ACCT_PINNED);
8270 kfree(ctx->user_bufs);
8271 ctx->user_bufs = NULL;
8272 ctx->nr_user_bufs = 0;
8276 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8277 void __user *arg, unsigned index)
8279 struct iovec __user *src;
8281 #ifdef CONFIG_COMPAT
8283 struct compat_iovec __user *ciovs;
8284 struct compat_iovec ciov;
8286 ciovs = (struct compat_iovec __user *) arg;
8287 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8290 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8291 dst->iov_len = ciov.iov_len;
8295 src = (struct iovec __user *) arg;
8296 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8302 * Not super efficient, but this is just a registration time. And we do cache
8303 * the last compound head, so generally we'll only do a full search if we don't
8306 * We check if the given compound head page has already been accounted, to
8307 * avoid double accounting it. This allows us to account the full size of the
8308 * page, not just the constituent pages of a huge page.
8310 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8311 int nr_pages, struct page *hpage)
8315 /* check current page array */
8316 for (i = 0; i < nr_pages; i++) {
8317 if (!PageCompound(pages[i]))
8319 if (compound_head(pages[i]) == hpage)
8323 /* check previously registered pages */
8324 for (i = 0; i < ctx->nr_user_bufs; i++) {
8325 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8327 for (j = 0; j < imu->nr_bvecs; j++) {
8328 if (!PageCompound(imu->bvec[j].bv_page))
8330 if (compound_head(imu->bvec[j].bv_page) == hpage)
8338 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8339 int nr_pages, struct io_mapped_ubuf *imu,
8340 struct page **last_hpage)
8344 for (i = 0; i < nr_pages; i++) {
8345 if (!PageCompound(pages[i])) {
8350 hpage = compound_head(pages[i]);
8351 if (hpage == *last_hpage)
8353 *last_hpage = hpage;
8354 if (headpage_already_acct(ctx, pages, i, hpage))
8356 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8360 if (!imu->acct_pages)
8363 ret = io_account_mem(ctx, imu->acct_pages, ACCT_PINNED);
8365 imu->acct_pages = 0;
8369 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
8372 struct vm_area_struct **vmas = NULL;
8373 struct page **pages = NULL;
8374 struct page *last_hpage = NULL;
8375 int i, j, got_pages = 0;
8380 if (!nr_args || nr_args > UIO_MAXIOV)
8383 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
8385 if (!ctx->user_bufs)
8388 for (i = 0; i < nr_args; i++) {
8389 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8390 unsigned long off, start, end, ubuf;
8395 ret = io_copy_iov(ctx, &iov, arg, i);
8400 * Don't impose further limits on the size and buffer
8401 * constraints here, we'll -EINVAL later when IO is
8402 * submitted if they are wrong.
8405 if (!iov.iov_base || !iov.iov_len)
8408 /* arbitrary limit, but we need something */
8409 if (iov.iov_len > SZ_1G)
8412 ubuf = (unsigned long) iov.iov_base;
8413 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8414 start = ubuf >> PAGE_SHIFT;
8415 nr_pages = end - start;
8418 if (!pages || nr_pages > got_pages) {
8421 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
8423 vmas = kvmalloc_array(nr_pages,
8424 sizeof(struct vm_area_struct *),
8426 if (!pages || !vmas) {
8430 got_pages = nr_pages;
8433 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
8440 mmap_read_lock(current->mm);
8441 pret = pin_user_pages(ubuf, nr_pages,
8442 FOLL_WRITE | FOLL_LONGTERM,
8444 if (pret == nr_pages) {
8445 /* don't support file backed memory */
8446 for (j = 0; j < nr_pages; j++) {
8447 struct vm_area_struct *vma = vmas[j];
8450 !is_file_hugepages(vma->vm_file)) {
8456 ret = pret < 0 ? pret : -EFAULT;
8458 mmap_read_unlock(current->mm);
8461 * if we did partial map, or found file backed vmas,
8462 * release any pages we did get
8465 unpin_user_pages(pages, pret);
8470 ret = io_buffer_account_pin(ctx, pages, pret, imu, &last_hpage);
8472 unpin_user_pages(pages, pret);
8477 off = ubuf & ~PAGE_MASK;
8479 for (j = 0; j < nr_pages; j++) {
8482 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8483 imu->bvec[j].bv_page = pages[j];
8484 imu->bvec[j].bv_len = vec_len;
8485 imu->bvec[j].bv_offset = off;
8489 /* store original address for later verification */
8491 imu->len = iov.iov_len;
8492 imu->nr_bvecs = nr_pages;
8494 ctx->nr_user_bufs++;
8502 io_sqe_buffer_unregister(ctx);
8506 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8508 __s32 __user *fds = arg;
8514 if (copy_from_user(&fd, fds, sizeof(*fds)))
8517 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8518 if (IS_ERR(ctx->cq_ev_fd)) {
8519 int ret = PTR_ERR(ctx->cq_ev_fd);
8520 ctx->cq_ev_fd = NULL;
8527 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8529 if (ctx->cq_ev_fd) {
8530 eventfd_ctx_put(ctx->cq_ev_fd);
8531 ctx->cq_ev_fd = NULL;
8538 static int __io_destroy_buffers(int id, void *p, void *data)
8540 struct io_ring_ctx *ctx = data;
8541 struct io_buffer *buf = p;
8543 __io_remove_buffers(ctx, buf, id, -1U);
8547 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8549 idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx);
8550 idr_destroy(&ctx->io_buffer_idr);
8553 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8555 io_finish_async(ctx);
8556 io_sqe_buffer_unregister(ctx);
8558 if (ctx->sqo_task) {
8559 put_task_struct(ctx->sqo_task);
8560 ctx->sqo_task = NULL;
8561 mmdrop(ctx->mm_account);
8562 ctx->mm_account = NULL;
8565 #ifdef CONFIG_BLK_CGROUP
8566 if (ctx->sqo_blkcg_css)
8567 css_put(ctx->sqo_blkcg_css);
8570 io_sqe_files_unregister(ctx);
8571 io_eventfd_unregister(ctx);
8572 io_destroy_buffers(ctx);
8573 idr_destroy(&ctx->personality_idr);
8575 #if defined(CONFIG_UNIX)
8576 if (ctx->ring_sock) {
8577 ctx->ring_sock->file = NULL; /* so that iput() is called */
8578 sock_release(ctx->ring_sock);
8582 io_mem_free(ctx->rings);
8583 io_mem_free(ctx->sq_sqes);
8585 percpu_ref_exit(&ctx->refs);
8586 free_uid(ctx->user);
8587 put_cred(ctx->creds);
8588 kfree(ctx->cancel_hash);
8589 kmem_cache_free(req_cachep, ctx->fallback_req);
8593 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8595 struct io_ring_ctx *ctx = file->private_data;
8598 poll_wait(file, &ctx->cq_wait, wait);
8600 * synchronizes with barrier from wq_has_sleeper call in
8604 if (!io_sqring_full(ctx))
8605 mask |= EPOLLOUT | EPOLLWRNORM;
8606 if (io_cqring_events(ctx, false))
8607 mask |= EPOLLIN | EPOLLRDNORM;
8612 static int io_uring_fasync(int fd, struct file *file, int on)
8614 struct io_ring_ctx *ctx = file->private_data;
8616 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8619 static int io_remove_personalities(int id, void *p, void *data)
8621 struct io_ring_ctx *ctx = data;
8622 struct io_identity *iod;
8624 iod = idr_remove(&ctx->personality_idr, id);
8626 put_cred(iod->creds);
8627 if (refcount_dec_and_test(&iod->count))
8633 static void io_ring_exit_work(struct work_struct *work)
8635 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
8639 * If we're doing polled IO and end up having requests being
8640 * submitted async (out-of-line), then completions can come in while
8641 * we're waiting for refs to drop. We need to reap these manually,
8642 * as nobody else will be looking for them.
8645 io_iopoll_try_reap_events(ctx);
8646 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
8647 io_ring_ctx_free(ctx);
8650 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8652 mutex_lock(&ctx->uring_lock);
8653 percpu_ref_kill(&ctx->refs);
8654 /* if force is set, the ring is going away. always drop after that */
8655 ctx->cq_overflow_flushed = 1;
8657 io_cqring_overflow_flush(ctx, true, NULL, NULL);
8658 mutex_unlock(&ctx->uring_lock);
8660 io_kill_timeouts(ctx, NULL, NULL);
8661 io_poll_remove_all(ctx, NULL, NULL);
8664 io_wq_cancel_all(ctx->io_wq);
8666 /* if we failed setting up the ctx, we might not have any rings */
8667 io_iopoll_try_reap_events(ctx);
8668 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
8671 * Do this upfront, so we won't have a grace period where the ring
8672 * is closed but resources aren't reaped yet. This can cause
8673 * spurious failure in setting up a new ring.
8675 io_unaccount_mem(ctx, ring_pages(ctx->sq_entries, ctx->cq_entries),
8678 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8680 * Use system_unbound_wq to avoid spawning tons of event kworkers
8681 * if we're exiting a ton of rings at the same time. It just adds
8682 * noise and overhead, there's no discernable change in runtime
8683 * over using system_wq.
8685 queue_work(system_unbound_wq, &ctx->exit_work);
8688 static int io_uring_release(struct inode *inode, struct file *file)
8690 struct io_ring_ctx *ctx = file->private_data;
8692 file->private_data = NULL;
8693 io_ring_ctx_wait_and_kill(ctx);
8697 struct io_task_cancel {
8698 struct task_struct *task;
8699 struct files_struct *files;
8702 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8704 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8705 struct io_task_cancel *cancel = data;
8708 if (cancel->files && (req->flags & REQ_F_LINK_TIMEOUT)) {
8709 unsigned long flags;
8710 struct io_ring_ctx *ctx = req->ctx;
8712 /* protect against races with linked timeouts */
8713 spin_lock_irqsave(&ctx->completion_lock, flags);
8714 ret = io_match_task(req, cancel->task, cancel->files);
8715 spin_unlock_irqrestore(&ctx->completion_lock, flags);
8717 ret = io_match_task(req, cancel->task, cancel->files);
8722 static void io_cancel_defer_files(struct io_ring_ctx *ctx,
8723 struct task_struct *task,
8724 struct files_struct *files)
8726 struct io_defer_entry *de = NULL;
8729 spin_lock_irq(&ctx->completion_lock);
8730 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8731 if (io_match_task(de->req, task, files)) {
8732 list_cut_position(&list, &ctx->defer_list, &de->list);
8736 spin_unlock_irq(&ctx->completion_lock);
8738 while (!list_empty(&list)) {
8739 de = list_first_entry(&list, struct io_defer_entry, list);
8740 list_del_init(&de->list);
8741 req_set_fail_links(de->req);
8742 io_put_req(de->req);
8743 io_req_complete(de->req, -ECANCELED);
8748 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
8749 struct task_struct *task,
8750 struct files_struct *files)
8752 while (!list_empty_careful(&ctx->inflight_list)) {
8753 struct io_task_cancel cancel = { .task = task, .files = files };
8754 struct io_kiocb *req;
8758 spin_lock_irq(&ctx->inflight_lock);
8759 list_for_each_entry(req, &ctx->inflight_list, inflight_entry) {
8760 if (req->task != task ||
8761 req->work.identity->files != files)
8767 prepare_to_wait(&task->io_uring->wait, &wait,
8768 TASK_UNINTERRUPTIBLE);
8769 spin_unlock_irq(&ctx->inflight_lock);
8771 /* We need to keep going until we don't find a matching req */
8775 io_wq_cancel_cb(ctx->io_wq, io_cancel_task_cb, &cancel, true);
8776 io_poll_remove_all(ctx, task, files);
8777 io_kill_timeouts(ctx, task, files);
8778 /* cancellations _may_ trigger task work */
8781 finish_wait(&task->io_uring->wait, &wait);
8785 static void __io_uring_cancel_task_requests(struct io_ring_ctx *ctx,
8786 struct task_struct *task)
8789 struct io_task_cancel cancel = { .task = task, .files = NULL, };
8790 enum io_wq_cancel cret;
8793 cret = io_wq_cancel_cb(ctx->io_wq, io_cancel_task_cb, &cancel, true);
8794 if (cret != IO_WQ_CANCEL_NOTFOUND)
8797 /* SQPOLL thread does its own polling */
8798 if (!(ctx->flags & IORING_SETUP_SQPOLL)) {
8799 while (!list_empty_careful(&ctx->iopoll_list)) {
8800 io_iopoll_try_reap_events(ctx);
8805 ret |= io_poll_remove_all(ctx, task, NULL);
8806 ret |= io_kill_timeouts(ctx, task, NULL);
8815 * We need to iteratively cancel requests, in case a request has dependent
8816 * hard links. These persist even for failure of cancelations, hence keep
8817 * looping until none are found.
8819 static void io_uring_cancel_task_requests(struct io_ring_ctx *ctx,
8820 struct files_struct *files)
8822 struct task_struct *task = current;
8824 if ((ctx->flags & IORING_SETUP_SQPOLL) && ctx->sq_data) {
8825 task = ctx->sq_data->thread;
8826 atomic_inc(&task->io_uring->in_idle);
8827 io_sq_thread_park(ctx->sq_data);
8830 io_cancel_defer_files(ctx, task, files);
8831 io_ring_submit_lock(ctx, (ctx->flags & IORING_SETUP_IOPOLL));
8832 io_cqring_overflow_flush(ctx, true, task, files);
8833 io_ring_submit_unlock(ctx, (ctx->flags & IORING_SETUP_IOPOLL));
8836 __io_uring_cancel_task_requests(ctx, task);
8838 io_uring_cancel_files(ctx, task, files);
8840 if ((ctx->flags & IORING_SETUP_SQPOLL) && ctx->sq_data) {
8841 atomic_dec(&task->io_uring->in_idle);
8843 * If the files that are going away are the ones in the thread
8844 * identity, clear them out.
8846 if (task->io_uring->identity->files == files)
8847 task->io_uring->identity->files = NULL;
8848 io_sq_thread_unpark(ctx->sq_data);
8853 * Note that this task has used io_uring. We use it for cancelation purposes.
8855 static int io_uring_add_task_file(struct io_ring_ctx *ctx, struct file *file)
8857 struct io_uring_task *tctx = current->io_uring;
8859 if (unlikely(!tctx)) {
8862 ret = io_uring_alloc_task_context(current);
8865 tctx = current->io_uring;
8867 if (tctx->last != file) {
8868 void *old = xa_load(&tctx->xa, (unsigned long)file);
8872 xa_store(&tctx->xa, (unsigned long)file, file, GFP_KERNEL);
8878 * This is race safe in that the task itself is doing this, hence it
8879 * cannot be going through the exit/cancel paths at the same time.
8880 * This cannot be modified while exit/cancel is running.
8882 if (!tctx->sqpoll && (ctx->flags & IORING_SETUP_SQPOLL))
8883 tctx->sqpoll = true;
8889 * Remove this io_uring_file -> task mapping.
8891 static void io_uring_del_task_file(struct file *file)
8893 struct io_uring_task *tctx = current->io_uring;
8895 if (tctx->last == file)
8897 file = xa_erase(&tctx->xa, (unsigned long)file);
8903 * Drop task note for this file if we're the only ones that hold it after
8906 static void io_uring_attempt_task_drop(struct file *file)
8908 if (!current->io_uring)
8911 * fput() is pending, will be 2 if the only other ref is our potential
8912 * task file note. If the task is exiting, drop regardless of count.
8914 if (fatal_signal_pending(current) || (current->flags & PF_EXITING) ||
8915 atomic_long_read(&file->f_count) == 2)
8916 io_uring_del_task_file(file);
8919 void __io_uring_files_cancel(struct files_struct *files)
8921 struct io_uring_task *tctx = current->io_uring;
8923 unsigned long index;
8925 /* make sure overflow events are dropped */
8926 atomic_inc(&tctx->in_idle);
8928 xa_for_each(&tctx->xa, index, file) {
8929 struct io_ring_ctx *ctx = file->private_data;
8931 io_uring_cancel_task_requests(ctx, files);
8933 io_uring_del_task_file(file);
8936 atomic_dec(&tctx->in_idle);
8939 static s64 tctx_inflight(struct io_uring_task *tctx)
8941 unsigned long index;
8945 inflight = percpu_counter_sum(&tctx->inflight);
8950 * If we have SQPOLL rings, then we need to iterate and find them, and
8951 * add the pending count for those.
8953 xa_for_each(&tctx->xa, index, file) {
8954 struct io_ring_ctx *ctx = file->private_data;
8956 if (ctx->flags & IORING_SETUP_SQPOLL) {
8957 struct io_uring_task *__tctx = ctx->sqo_task->io_uring;
8959 inflight += percpu_counter_sum(&__tctx->inflight);
8967 * Find any io_uring fd that this task has registered or done IO on, and cancel
8970 void __io_uring_task_cancel(void)
8972 struct io_uring_task *tctx = current->io_uring;
8976 /* make sure overflow events are dropped */
8977 atomic_inc(&tctx->in_idle);
8980 /* read completions before cancelations */
8981 inflight = tctx_inflight(tctx);
8984 __io_uring_files_cancel(NULL);
8986 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
8989 * If we've seen completions, retry. This avoids a race where
8990 * a completion comes in before we did prepare_to_wait().
8992 if (inflight != tctx_inflight(tctx))
8997 finish_wait(&tctx->wait, &wait);
8998 atomic_dec(&tctx->in_idle);
9001 static int io_uring_flush(struct file *file, void *data)
9003 io_uring_attempt_task_drop(file);
9007 static void *io_uring_validate_mmap_request(struct file *file,
9008 loff_t pgoff, size_t sz)
9010 struct io_ring_ctx *ctx = file->private_data;
9011 loff_t offset = pgoff << PAGE_SHIFT;
9016 case IORING_OFF_SQ_RING:
9017 case IORING_OFF_CQ_RING:
9020 case IORING_OFF_SQES:
9024 return ERR_PTR(-EINVAL);
9027 page = virt_to_head_page(ptr);
9028 if (sz > page_size(page))
9029 return ERR_PTR(-EINVAL);
9036 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9038 size_t sz = vma->vm_end - vma->vm_start;
9042 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
9044 return PTR_ERR(ptr);
9046 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
9047 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
9050 #else /* !CONFIG_MMU */
9052 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9054 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
9057 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
9059 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
9062 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
9063 unsigned long addr, unsigned long len,
9064 unsigned long pgoff, unsigned long flags)
9068 ptr = io_uring_validate_mmap_request(file, pgoff, len);
9070 return PTR_ERR(ptr);
9072 return (unsigned long) ptr;
9075 #endif /* !CONFIG_MMU */
9077 static void io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
9082 if (!io_sqring_full(ctx))
9085 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9087 if (!io_sqring_full(ctx))
9091 } while (!signal_pending(current));
9093 finish_wait(&ctx->sqo_sq_wait, &wait);
9096 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9097 struct __kernel_timespec __user **ts,
9098 const sigset_t __user **sig)
9100 struct io_uring_getevents_arg arg;
9103 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9104 * is just a pointer to the sigset_t.
9106 if (!(flags & IORING_ENTER_EXT_ARG)) {
9107 *sig = (const sigset_t __user *) argp;
9113 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9114 * timespec and sigset_t pointers if good.
9116 if (*argsz != sizeof(arg))
9118 if (copy_from_user(&arg, argp, sizeof(arg)))
9120 *sig = u64_to_user_ptr(arg.sigmask);
9121 *argsz = arg.sigmask_sz;
9122 *ts = u64_to_user_ptr(arg.ts);
9126 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9127 u32, min_complete, u32, flags, const void __user *, argp,
9130 struct io_ring_ctx *ctx;
9137 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9138 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG))
9146 if (f.file->f_op != &io_uring_fops)
9150 ctx = f.file->private_data;
9151 if (!percpu_ref_tryget(&ctx->refs))
9155 if (ctx->flags & IORING_SETUP_R_DISABLED)
9159 * For SQ polling, the thread will do all submissions and completions.
9160 * Just return the requested submit count, and wake the thread if
9164 if (ctx->flags & IORING_SETUP_SQPOLL) {
9165 io_ring_submit_lock(ctx, (ctx->flags & IORING_SETUP_IOPOLL));
9166 if (!list_empty_careful(&ctx->cq_overflow_list))
9167 io_cqring_overflow_flush(ctx, false, NULL, NULL);
9168 io_ring_submit_unlock(ctx, (ctx->flags & IORING_SETUP_IOPOLL));
9169 if (flags & IORING_ENTER_SQ_WAKEUP)
9170 wake_up(&ctx->sq_data->wait);
9171 if (flags & IORING_ENTER_SQ_WAIT)
9172 io_sqpoll_wait_sq(ctx);
9173 submitted = to_submit;
9174 } else if (to_submit) {
9175 ret = io_uring_add_task_file(ctx, f.file);
9178 mutex_lock(&ctx->uring_lock);
9179 submitted = io_submit_sqes(ctx, to_submit);
9180 mutex_unlock(&ctx->uring_lock);
9182 if (submitted != to_submit)
9185 if (flags & IORING_ENTER_GETEVENTS) {
9186 const sigset_t __user *sig;
9187 struct __kernel_timespec __user *ts;
9189 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
9193 min_complete = min(min_complete, ctx->cq_entries);
9196 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9197 * space applications don't need to do io completion events
9198 * polling again, they can rely on io_sq_thread to do polling
9199 * work, which can reduce cpu usage and uring_lock contention.
9201 if (ctx->flags & IORING_SETUP_IOPOLL &&
9202 !(ctx->flags & IORING_SETUP_SQPOLL)) {
9203 ret = io_iopoll_check(ctx, min_complete);
9205 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
9210 percpu_ref_put(&ctx->refs);
9213 return submitted ? submitted : ret;
9216 #ifdef CONFIG_PROC_FS
9217 static int io_uring_show_cred(int id, void *p, void *data)
9219 struct io_identity *iod = p;
9220 const struct cred *cred = iod->creds;
9221 struct seq_file *m = data;
9222 struct user_namespace *uns = seq_user_ns(m);
9223 struct group_info *gi;
9228 seq_printf(m, "%5d\n", id);
9229 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
9230 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
9231 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
9232 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
9233 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
9234 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
9235 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
9236 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
9237 seq_puts(m, "\n\tGroups:\t");
9238 gi = cred->group_info;
9239 for (g = 0; g < gi->ngroups; g++) {
9240 seq_put_decimal_ull(m, g ? " " : "",
9241 from_kgid_munged(uns, gi->gid[g]));
9243 seq_puts(m, "\n\tCapEff:\t");
9244 cap = cred->cap_effective;
9245 CAP_FOR_EACH_U32(__capi)
9246 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
9251 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
9253 struct io_sq_data *sq = NULL;
9258 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9259 * since fdinfo case grabs it in the opposite direction of normal use
9260 * cases. If we fail to get the lock, we just don't iterate any
9261 * structures that could be going away outside the io_uring mutex.
9263 has_lock = mutex_trylock(&ctx->uring_lock);
9265 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL))
9268 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9269 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9270 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9271 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9272 struct fixed_file_table *table;
9275 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
9276 f = table->files[i & IORING_FILE_TABLE_MASK];
9278 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9280 seq_printf(m, "%5u: <none>\n", i);
9282 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9283 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9284 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
9286 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
9287 (unsigned int) buf->len);
9289 if (has_lock && !idr_is_empty(&ctx->personality_idr)) {
9290 seq_printf(m, "Personalities:\n");
9291 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
9293 seq_printf(m, "PollList:\n");
9294 spin_lock_irq(&ctx->completion_lock);
9295 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9296 struct hlist_head *list = &ctx->cancel_hash[i];
9297 struct io_kiocb *req;
9299 hlist_for_each_entry(req, list, hash_node)
9300 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9301 req->task->task_works != NULL);
9303 spin_unlock_irq(&ctx->completion_lock);
9305 mutex_unlock(&ctx->uring_lock);
9308 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9310 struct io_ring_ctx *ctx = f->private_data;
9312 if (percpu_ref_tryget(&ctx->refs)) {
9313 __io_uring_show_fdinfo(ctx, m);
9314 percpu_ref_put(&ctx->refs);
9319 static const struct file_operations io_uring_fops = {
9320 .release = io_uring_release,
9321 .flush = io_uring_flush,
9322 .mmap = io_uring_mmap,
9324 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9325 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9327 .poll = io_uring_poll,
9328 .fasync = io_uring_fasync,
9329 #ifdef CONFIG_PROC_FS
9330 .show_fdinfo = io_uring_show_fdinfo,
9334 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9335 struct io_uring_params *p)
9337 struct io_rings *rings;
9338 size_t size, sq_array_offset;
9340 /* make sure these are sane, as we already accounted them */
9341 ctx->sq_entries = p->sq_entries;
9342 ctx->cq_entries = p->cq_entries;
9344 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9345 if (size == SIZE_MAX)
9348 rings = io_mem_alloc(size);
9353 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9354 rings->sq_ring_mask = p->sq_entries - 1;
9355 rings->cq_ring_mask = p->cq_entries - 1;
9356 rings->sq_ring_entries = p->sq_entries;
9357 rings->cq_ring_entries = p->cq_entries;
9358 ctx->sq_mask = rings->sq_ring_mask;
9359 ctx->cq_mask = rings->cq_ring_mask;
9361 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9362 if (size == SIZE_MAX) {
9363 io_mem_free(ctx->rings);
9368 ctx->sq_sqes = io_mem_alloc(size);
9369 if (!ctx->sq_sqes) {
9370 io_mem_free(ctx->rings);
9379 * Allocate an anonymous fd, this is what constitutes the application
9380 * visible backing of an io_uring instance. The application mmaps this
9381 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9382 * we have to tie this fd to a socket for file garbage collection purposes.
9384 static int io_uring_get_fd(struct io_ring_ctx *ctx)
9390 #if defined(CONFIG_UNIX)
9391 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9397 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9402 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9403 O_RDWR | O_CLOEXEC);
9406 ret = PTR_ERR(file);
9410 #if defined(CONFIG_UNIX)
9411 ctx->ring_sock->file = file;
9413 ret = io_uring_add_task_file(ctx, file);
9419 fd_install(fd, file);
9422 #if defined(CONFIG_UNIX)
9423 sock_release(ctx->ring_sock);
9424 ctx->ring_sock = NULL;
9429 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9430 struct io_uring_params __user *params)
9432 struct user_struct *user = NULL;
9433 struct io_ring_ctx *ctx;
9439 if (entries > IORING_MAX_ENTRIES) {
9440 if (!(p->flags & IORING_SETUP_CLAMP))
9442 entries = IORING_MAX_ENTRIES;
9446 * Use twice as many entries for the CQ ring. It's possible for the
9447 * application to drive a higher depth than the size of the SQ ring,
9448 * since the sqes are only used at submission time. This allows for
9449 * some flexibility in overcommitting a bit. If the application has
9450 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9451 * of CQ ring entries manually.
9453 p->sq_entries = roundup_pow_of_two(entries);
9454 if (p->flags & IORING_SETUP_CQSIZE) {
9456 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9457 * to a power-of-two, if it isn't already. We do NOT impose
9458 * any cq vs sq ring sizing.
9462 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9463 if (!(p->flags & IORING_SETUP_CLAMP))
9465 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9467 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9468 if (p->cq_entries < p->sq_entries)
9471 p->cq_entries = 2 * p->sq_entries;
9474 user = get_uid(current_user());
9475 limit_mem = !capable(CAP_IPC_LOCK);
9478 ret = __io_account_mem(user,
9479 ring_pages(p->sq_entries, p->cq_entries));
9486 ctx = io_ring_ctx_alloc(p);
9489 __io_unaccount_mem(user, ring_pages(p->sq_entries,
9494 ctx->compat = in_compat_syscall();
9496 ctx->creds = get_current_cred();
9498 ctx->loginuid = current->loginuid;
9499 ctx->sessionid = current->sessionid;
9501 ctx->sqo_task = get_task_struct(current);
9504 * This is just grabbed for accounting purposes. When a process exits,
9505 * the mm is exited and dropped before the files, hence we need to hang
9506 * on to this mm purely for the purposes of being able to unaccount
9507 * memory (locked/pinned vm). It's not used for anything else.
9509 mmgrab(current->mm);
9510 ctx->mm_account = current->mm;
9512 #ifdef CONFIG_BLK_CGROUP
9514 * The sq thread will belong to the original cgroup it was inited in.
9515 * If the cgroup goes offline (e.g. disabling the io controller), then
9516 * issued bios will be associated with the closest cgroup later in the
9520 ctx->sqo_blkcg_css = blkcg_css();
9521 ret = css_tryget_online(ctx->sqo_blkcg_css);
9524 /* don't init against a dying cgroup, have the user try again */
9525 ctx->sqo_blkcg_css = NULL;
9532 * Account memory _before_ installing the file descriptor. Once
9533 * the descriptor is installed, it can get closed at any time. Also
9534 * do this before hitting the general error path, as ring freeing
9535 * will un-account as well.
9537 io_account_mem(ctx, ring_pages(p->sq_entries, p->cq_entries),
9539 ctx->limit_mem = limit_mem;
9541 ret = io_allocate_scq_urings(ctx, p);
9545 ret = io_sq_offload_create(ctx, p);
9549 if (!(p->flags & IORING_SETUP_R_DISABLED))
9550 io_sq_offload_start(ctx);
9552 memset(&p->sq_off, 0, sizeof(p->sq_off));
9553 p->sq_off.head = offsetof(struct io_rings, sq.head);
9554 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9555 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9556 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9557 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9558 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9559 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9561 memset(&p->cq_off, 0, sizeof(p->cq_off));
9562 p->cq_off.head = offsetof(struct io_rings, cq.head);
9563 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9564 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9565 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9566 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9567 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9568 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9570 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9571 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9572 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9573 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
9574 IORING_FEAT_EXT_ARG;
9576 if (copy_to_user(params, p, sizeof(*p))) {
9582 * Install ring fd as the very last thing, so we don't risk someone
9583 * having closed it before we finish setup
9585 ret = io_uring_get_fd(ctx);
9589 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9592 io_ring_ctx_wait_and_kill(ctx);
9597 * Sets up an aio uring context, and returns the fd. Applications asks for a
9598 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9599 * params structure passed in.
9601 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9603 struct io_uring_params p;
9606 if (copy_from_user(&p, params, sizeof(p)))
9608 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9613 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9614 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9615 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9616 IORING_SETUP_R_DISABLED))
9619 return io_uring_create(entries, &p, params);
9622 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9623 struct io_uring_params __user *, params)
9625 return io_uring_setup(entries, params);
9628 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9630 struct io_uring_probe *p;
9634 size = struct_size(p, ops, nr_args);
9635 if (size == SIZE_MAX)
9637 p = kzalloc(size, GFP_KERNEL);
9642 if (copy_from_user(p, arg, size))
9645 if (memchr_inv(p, 0, size))
9648 p->last_op = IORING_OP_LAST - 1;
9649 if (nr_args > IORING_OP_LAST)
9650 nr_args = IORING_OP_LAST;
9652 for (i = 0; i < nr_args; i++) {
9654 if (!io_op_defs[i].not_supported)
9655 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9660 if (copy_to_user(arg, p, size))
9667 static int io_register_personality(struct io_ring_ctx *ctx)
9669 struct io_identity *id;
9672 id = kmalloc(sizeof(*id), GFP_KERNEL);
9676 io_init_identity(id);
9677 id->creds = get_current_cred();
9679 ret = idr_alloc_cyclic(&ctx->personality_idr, id, 1, USHRT_MAX, GFP_KERNEL);
9681 put_cred(id->creds);
9687 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
9689 struct io_identity *iod;
9691 iod = idr_remove(&ctx->personality_idr, id);
9693 put_cred(iod->creds);
9694 if (refcount_dec_and_test(&iod->count))
9702 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9703 unsigned int nr_args)
9705 struct io_uring_restriction *res;
9709 /* Restrictions allowed only if rings started disabled */
9710 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9713 /* We allow only a single restrictions registration */
9714 if (ctx->restrictions.registered)
9717 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9720 size = array_size(nr_args, sizeof(*res));
9721 if (size == SIZE_MAX)
9724 res = memdup_user(arg, size);
9726 return PTR_ERR(res);
9730 for (i = 0; i < nr_args; i++) {
9731 switch (res[i].opcode) {
9732 case IORING_RESTRICTION_REGISTER_OP:
9733 if (res[i].register_op >= IORING_REGISTER_LAST) {
9738 __set_bit(res[i].register_op,
9739 ctx->restrictions.register_op);
9741 case IORING_RESTRICTION_SQE_OP:
9742 if (res[i].sqe_op >= IORING_OP_LAST) {
9747 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9749 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9750 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9752 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9753 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9762 /* Reset all restrictions if an error happened */
9764 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9766 ctx->restrictions.registered = true;
9772 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9774 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9777 if (ctx->restrictions.registered)
9778 ctx->restricted = 1;
9780 ctx->flags &= ~IORING_SETUP_R_DISABLED;
9782 io_sq_offload_start(ctx);
9787 static bool io_register_op_must_quiesce(int op)
9790 case IORING_UNREGISTER_FILES:
9791 case IORING_REGISTER_FILES_UPDATE:
9792 case IORING_REGISTER_PROBE:
9793 case IORING_REGISTER_PERSONALITY:
9794 case IORING_UNREGISTER_PERSONALITY:
9801 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
9802 void __user *arg, unsigned nr_args)
9803 __releases(ctx->uring_lock)
9804 __acquires(ctx->uring_lock)
9809 * We're inside the ring mutex, if the ref is already dying, then
9810 * someone else killed the ctx or is already going through
9811 * io_uring_register().
9813 if (percpu_ref_is_dying(&ctx->refs))
9816 if (io_register_op_must_quiesce(opcode)) {
9817 percpu_ref_kill(&ctx->refs);
9820 * Drop uring mutex before waiting for references to exit. If
9821 * another thread is currently inside io_uring_enter() it might
9822 * need to grab the uring_lock to make progress. If we hold it
9823 * here across the drain wait, then we can deadlock. It's safe
9824 * to drop the mutex here, since no new references will come in
9825 * after we've killed the percpu ref.
9827 mutex_unlock(&ctx->uring_lock);
9829 ret = wait_for_completion_interruptible(&ctx->ref_comp);
9832 ret = io_run_task_work_sig();
9837 mutex_lock(&ctx->uring_lock);
9840 percpu_ref_resurrect(&ctx->refs);
9845 if (ctx->restricted) {
9846 if (opcode >= IORING_REGISTER_LAST) {
9851 if (!test_bit(opcode, ctx->restrictions.register_op)) {
9858 case IORING_REGISTER_BUFFERS:
9859 ret = io_sqe_buffer_register(ctx, arg, nr_args);
9861 case IORING_UNREGISTER_BUFFERS:
9865 ret = io_sqe_buffer_unregister(ctx);
9867 case IORING_REGISTER_FILES:
9868 ret = io_sqe_files_register(ctx, arg, nr_args);
9870 case IORING_UNREGISTER_FILES:
9874 ret = io_sqe_files_unregister(ctx);
9876 case IORING_REGISTER_FILES_UPDATE:
9877 ret = io_sqe_files_update(ctx, arg, nr_args);
9879 case IORING_REGISTER_EVENTFD:
9880 case IORING_REGISTER_EVENTFD_ASYNC:
9884 ret = io_eventfd_register(ctx, arg);
9887 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
9888 ctx->eventfd_async = 1;
9890 ctx->eventfd_async = 0;
9892 case IORING_UNREGISTER_EVENTFD:
9896 ret = io_eventfd_unregister(ctx);
9898 case IORING_REGISTER_PROBE:
9900 if (!arg || nr_args > 256)
9902 ret = io_probe(ctx, arg, nr_args);
9904 case IORING_REGISTER_PERSONALITY:
9908 ret = io_register_personality(ctx);
9910 case IORING_UNREGISTER_PERSONALITY:
9914 ret = io_unregister_personality(ctx, nr_args);
9916 case IORING_REGISTER_ENABLE_RINGS:
9920 ret = io_register_enable_rings(ctx);
9922 case IORING_REGISTER_RESTRICTIONS:
9923 ret = io_register_restrictions(ctx, arg, nr_args);
9931 if (io_register_op_must_quiesce(opcode)) {
9932 /* bring the ctx back to life */
9933 percpu_ref_reinit(&ctx->refs);
9935 reinit_completion(&ctx->ref_comp);
9940 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
9941 void __user *, arg, unsigned int, nr_args)
9943 struct io_ring_ctx *ctx;
9952 if (f.file->f_op != &io_uring_fops)
9955 ctx = f.file->private_data;
9957 mutex_lock(&ctx->uring_lock);
9958 ret = __io_uring_register(ctx, opcode, arg, nr_args);
9959 mutex_unlock(&ctx->uring_lock);
9960 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
9961 ctx->cq_ev_fd != NULL, ret);
9967 static int __init io_uring_init(void)
9969 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
9970 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
9971 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
9974 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
9975 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
9976 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
9977 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
9978 BUILD_BUG_SQE_ELEM(1, __u8, flags);
9979 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
9980 BUILD_BUG_SQE_ELEM(4, __s32, fd);
9981 BUILD_BUG_SQE_ELEM(8, __u64, off);
9982 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
9983 BUILD_BUG_SQE_ELEM(16, __u64, addr);
9984 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
9985 BUILD_BUG_SQE_ELEM(24, __u32, len);
9986 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
9987 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
9988 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
9989 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
9990 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
9991 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
9992 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
9993 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
9994 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
9995 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
9996 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
9997 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
9998 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
9999 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
10000 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
10001 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
10002 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
10003 BUILD_BUG_SQE_ELEM(42, __u16, personality);
10004 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
10006 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
10007 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
10008 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
10011 __initcall(io_uring_init);