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 ((ctx->cached_cq_tail - READ_ONCE(rings->cq.head) ==
1728 rings->cq_ring_entries))
1732 spin_lock_irqsave(&ctx->completion_lock, flags);
1735 list_for_each_entry_safe(req, tmp, &ctx->cq_overflow_list, compl.list) {
1736 if (!io_match_task(req, tsk, files))
1739 cqe = io_get_cqring(ctx);
1743 list_move(&req->compl.list, &list);
1745 WRITE_ONCE(cqe->user_data, req->user_data);
1746 WRITE_ONCE(cqe->res, req->result);
1747 WRITE_ONCE(cqe->flags, req->compl.cflags);
1749 ctx->cached_cq_overflow++;
1750 WRITE_ONCE(ctx->rings->cq_overflow,
1751 ctx->cached_cq_overflow);
1755 io_commit_cqring(ctx);
1756 io_cqring_mark_overflow(ctx);
1758 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1759 io_cqring_ev_posted(ctx);
1761 while (!list_empty(&list)) {
1762 req = list_first_entry(&list, struct io_kiocb, compl.list);
1763 list_del(&req->compl.list);
1770 static void __io_cqring_fill_event(struct io_kiocb *req, long res, long cflags)
1772 struct io_ring_ctx *ctx = req->ctx;
1773 struct io_uring_cqe *cqe;
1775 trace_io_uring_complete(ctx, req->user_data, res);
1778 * If we can't get a cq entry, userspace overflowed the
1779 * submission (by quite a lot). Increment the overflow count in
1782 cqe = io_get_cqring(ctx);
1784 WRITE_ONCE(cqe->user_data, req->user_data);
1785 WRITE_ONCE(cqe->res, res);
1786 WRITE_ONCE(cqe->flags, cflags);
1787 } else if (ctx->cq_overflow_flushed ||
1788 atomic_read(&req->task->io_uring->in_idle)) {
1790 * If we're in ring overflow flush mode, or in task cancel mode,
1791 * then we cannot store the request for later flushing, we need
1792 * to drop it on the floor.
1794 ctx->cached_cq_overflow++;
1795 WRITE_ONCE(ctx->rings->cq_overflow, ctx->cached_cq_overflow);
1797 if (list_empty(&ctx->cq_overflow_list)) {
1798 set_bit(0, &ctx->sq_check_overflow);
1799 set_bit(0, &ctx->cq_check_overflow);
1800 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
1804 req->compl.cflags = cflags;
1805 refcount_inc(&req->refs);
1806 list_add_tail(&req->compl.list, &ctx->cq_overflow_list);
1810 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1812 __io_cqring_fill_event(req, res, 0);
1815 static void io_cqring_add_event(struct io_kiocb *req, long res, long cflags)
1817 struct io_ring_ctx *ctx = req->ctx;
1818 unsigned long flags;
1820 spin_lock_irqsave(&ctx->completion_lock, flags);
1821 __io_cqring_fill_event(req, res, cflags);
1822 io_commit_cqring(ctx);
1823 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1825 io_cqring_ev_posted(ctx);
1828 static void io_submit_flush_completions(struct io_comp_state *cs)
1830 struct io_ring_ctx *ctx = cs->ctx;
1832 spin_lock_irq(&ctx->completion_lock);
1833 while (!list_empty(&cs->list)) {
1834 struct io_kiocb *req;
1836 req = list_first_entry(&cs->list, struct io_kiocb, compl.list);
1837 list_del(&req->compl.list);
1838 __io_cqring_fill_event(req, req->result, req->compl.cflags);
1841 * io_free_req() doesn't care about completion_lock unless one
1842 * of these flags is set. REQ_F_WORK_INITIALIZED is in the list
1843 * because of a potential deadlock with req->work.fs->lock
1845 if (req->flags & (REQ_F_FAIL_LINK|REQ_F_LINK_TIMEOUT
1846 |REQ_F_WORK_INITIALIZED)) {
1847 spin_unlock_irq(&ctx->completion_lock);
1849 spin_lock_irq(&ctx->completion_lock);
1854 io_commit_cqring(ctx);
1855 spin_unlock_irq(&ctx->completion_lock);
1857 io_cqring_ev_posted(ctx);
1861 static void __io_req_complete(struct io_kiocb *req, long res, unsigned cflags,
1862 struct io_comp_state *cs)
1865 io_cqring_add_event(req, res, cflags);
1870 req->compl.cflags = cflags;
1871 list_add_tail(&req->compl.list, &cs->list);
1873 io_submit_flush_completions(cs);
1877 static void io_req_complete(struct io_kiocb *req, long res)
1879 __io_req_complete(req, res, 0, NULL);
1882 static inline bool io_is_fallback_req(struct io_kiocb *req)
1884 return req == (struct io_kiocb *)
1885 ((unsigned long) req->ctx->fallback_req & ~1UL);
1888 static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx)
1890 struct io_kiocb *req;
1892 req = ctx->fallback_req;
1893 if (!test_and_set_bit_lock(0, (unsigned long *) &ctx->fallback_req))
1899 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx,
1900 struct io_submit_state *state)
1902 if (!state->free_reqs) {
1903 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1907 sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs));
1908 ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs);
1911 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1912 * retry single alloc to be on the safe side.
1914 if (unlikely(ret <= 0)) {
1915 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1916 if (!state->reqs[0])
1920 state->free_reqs = ret;
1924 return state->reqs[state->free_reqs];
1926 return io_get_fallback_req(ctx);
1929 static inline void io_put_file(struct io_kiocb *req, struct file *file,
1936 static void io_dismantle_req(struct io_kiocb *req)
1940 if (req->async_data)
1941 kfree(req->async_data);
1943 io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
1944 if (req->fixed_file_refs)
1945 percpu_ref_put(req->fixed_file_refs);
1946 io_req_clean_work(req);
1949 static void __io_free_req(struct io_kiocb *req)
1951 struct io_uring_task *tctx = req->task->io_uring;
1952 struct io_ring_ctx *ctx = req->ctx;
1954 io_dismantle_req(req);
1956 percpu_counter_dec(&tctx->inflight);
1957 if (atomic_read(&tctx->in_idle))
1958 wake_up(&tctx->wait);
1959 put_task_struct(req->task);
1961 if (likely(!io_is_fallback_req(req)))
1962 kmem_cache_free(req_cachep, req);
1964 clear_bit_unlock(0, (unsigned long *) &ctx->fallback_req);
1965 percpu_ref_put(&ctx->refs);
1968 static inline void io_remove_next_linked(struct io_kiocb *req)
1970 struct io_kiocb *nxt = req->link;
1972 req->link = nxt->link;
1976 static void io_kill_linked_timeout(struct io_kiocb *req)
1978 struct io_ring_ctx *ctx = req->ctx;
1979 struct io_kiocb *link;
1980 bool cancelled = false;
1981 unsigned long flags;
1983 spin_lock_irqsave(&ctx->completion_lock, flags);
1987 * Can happen if a linked timeout fired and link had been like
1988 * req -> link t-out -> link t-out [-> ...]
1990 if (link && (link->flags & REQ_F_LTIMEOUT_ACTIVE)) {
1991 struct io_timeout_data *io = link->async_data;
1994 io_remove_next_linked(req);
1995 link->timeout.head = NULL;
1996 ret = hrtimer_try_to_cancel(&io->timer);
1998 io_cqring_fill_event(link, -ECANCELED);
1999 io_commit_cqring(ctx);
2003 req->flags &= ~REQ_F_LINK_TIMEOUT;
2004 spin_unlock_irqrestore(&ctx->completion_lock, flags);
2007 io_cqring_ev_posted(ctx);
2013 static void io_fail_links(struct io_kiocb *req)
2015 struct io_kiocb *link, *nxt;
2016 struct io_ring_ctx *ctx = req->ctx;
2017 unsigned long flags;
2019 spin_lock_irqsave(&ctx->completion_lock, flags);
2027 trace_io_uring_fail_link(req, link);
2028 io_cqring_fill_event(link, -ECANCELED);
2031 * It's ok to free under spinlock as they're not linked anymore,
2032 * but avoid REQ_F_WORK_INITIALIZED because it may deadlock on
2035 if (link->flags & REQ_F_WORK_INITIALIZED)
2036 io_put_req_deferred(link, 2);
2038 io_double_put_req(link);
2041 io_commit_cqring(ctx);
2042 spin_unlock_irqrestore(&ctx->completion_lock, flags);
2044 io_cqring_ev_posted(ctx);
2047 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
2049 if (req->flags & REQ_F_LINK_TIMEOUT)
2050 io_kill_linked_timeout(req);
2053 * If LINK is set, we have dependent requests in this chain. If we
2054 * didn't fail this request, queue the first one up, moving any other
2055 * dependencies to the next request. In case of failure, fail the rest
2058 if (likely(!(req->flags & REQ_F_FAIL_LINK))) {
2059 struct io_kiocb *nxt = req->link;
2068 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
2070 if (likely(!(req->link) && !(req->flags & REQ_F_LINK_TIMEOUT)))
2072 return __io_req_find_next(req);
2075 static int io_req_task_work_add(struct io_kiocb *req)
2077 struct task_struct *tsk = req->task;
2078 struct io_ring_ctx *ctx = req->ctx;
2079 enum task_work_notify_mode notify;
2082 if (tsk->flags & PF_EXITING)
2086 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
2087 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
2088 * processing task_work. There's no reliable way to tell if TWA_RESUME
2092 if (!(ctx->flags & IORING_SETUP_SQPOLL))
2093 notify = TWA_SIGNAL;
2095 ret = task_work_add(tsk, &req->task_work, notify);
2097 wake_up_process(tsk);
2102 static void __io_req_task_cancel(struct io_kiocb *req, int error)
2104 struct io_ring_ctx *ctx = req->ctx;
2106 spin_lock_irq(&ctx->completion_lock);
2107 io_cqring_fill_event(req, error);
2108 io_commit_cqring(ctx);
2109 spin_unlock_irq(&ctx->completion_lock);
2111 io_cqring_ev_posted(ctx);
2112 req_set_fail_links(req);
2113 io_double_put_req(req);
2116 static void io_req_task_cancel(struct callback_head *cb)
2118 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2119 struct io_ring_ctx *ctx = req->ctx;
2121 __io_req_task_cancel(req, -ECANCELED);
2122 percpu_ref_put(&ctx->refs);
2125 static void __io_req_task_submit(struct io_kiocb *req)
2127 struct io_ring_ctx *ctx = req->ctx;
2129 if (!__io_sq_thread_acquire_mm(ctx) &&
2130 !__io_sq_thread_acquire_files(ctx)) {
2131 mutex_lock(&ctx->uring_lock);
2132 __io_queue_sqe(req, NULL);
2133 mutex_unlock(&ctx->uring_lock);
2135 __io_req_task_cancel(req, -EFAULT);
2139 static void io_req_task_submit(struct callback_head *cb)
2141 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2142 struct io_ring_ctx *ctx = req->ctx;
2144 __io_req_task_submit(req);
2145 percpu_ref_put(&ctx->refs);
2148 static void io_req_task_queue(struct io_kiocb *req)
2152 init_task_work(&req->task_work, io_req_task_submit);
2153 percpu_ref_get(&req->ctx->refs);
2155 ret = io_req_task_work_add(req);
2156 if (unlikely(ret)) {
2157 struct task_struct *tsk;
2159 init_task_work(&req->task_work, io_req_task_cancel);
2160 tsk = io_wq_get_task(req->ctx->io_wq);
2161 task_work_add(tsk, &req->task_work, TWA_NONE);
2162 wake_up_process(tsk);
2166 static inline void io_queue_next(struct io_kiocb *req)
2168 struct io_kiocb *nxt = io_req_find_next(req);
2171 io_req_task_queue(nxt);
2174 static void io_free_req(struct io_kiocb *req)
2181 void *reqs[IO_IOPOLL_BATCH];
2184 struct task_struct *task;
2188 static inline void io_init_req_batch(struct req_batch *rb)
2195 static void __io_req_free_batch_flush(struct io_ring_ctx *ctx,
2196 struct req_batch *rb)
2198 kmem_cache_free_bulk(req_cachep, rb->to_free, rb->reqs);
2199 percpu_ref_put_many(&ctx->refs, rb->to_free);
2203 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2204 struct req_batch *rb)
2207 __io_req_free_batch_flush(ctx, rb);
2209 struct io_uring_task *tctx = rb->task->io_uring;
2211 percpu_counter_sub(&tctx->inflight, rb->task_refs);
2212 put_task_struct_many(rb->task, rb->task_refs);
2217 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req)
2219 if (unlikely(io_is_fallback_req(req))) {
2225 if (req->task != rb->task) {
2227 struct io_uring_task *tctx = rb->task->io_uring;
2229 percpu_counter_sub(&tctx->inflight, rb->task_refs);
2230 put_task_struct_many(rb->task, rb->task_refs);
2232 rb->task = req->task;
2237 io_dismantle_req(req);
2238 rb->reqs[rb->to_free++] = req;
2239 if (unlikely(rb->to_free == ARRAY_SIZE(rb->reqs)))
2240 __io_req_free_batch_flush(req->ctx, rb);
2244 * Drop reference to request, return next in chain (if there is one) if this
2245 * was the last reference to this request.
2247 static struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2249 struct io_kiocb *nxt = NULL;
2251 if (refcount_dec_and_test(&req->refs)) {
2252 nxt = io_req_find_next(req);
2258 static void io_put_req(struct io_kiocb *req)
2260 if (refcount_dec_and_test(&req->refs))
2264 static void io_put_req_deferred_cb(struct callback_head *cb)
2266 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2271 static void io_free_req_deferred(struct io_kiocb *req)
2275 init_task_work(&req->task_work, io_put_req_deferred_cb);
2276 ret = io_req_task_work_add(req);
2277 if (unlikely(ret)) {
2278 struct task_struct *tsk;
2280 tsk = io_wq_get_task(req->ctx->io_wq);
2281 task_work_add(tsk, &req->task_work, TWA_NONE);
2282 wake_up_process(tsk);
2286 static inline void io_put_req_deferred(struct io_kiocb *req, int refs)
2288 if (refcount_sub_and_test(refs, &req->refs))
2289 io_free_req_deferred(req);
2292 static struct io_wq_work *io_steal_work(struct io_kiocb *req)
2294 struct io_kiocb *nxt;
2297 * A ref is owned by io-wq in which context we're. So, if that's the
2298 * last one, it's safe to steal next work. False negatives are Ok,
2299 * it just will be re-punted async in io_put_work()
2301 if (refcount_read(&req->refs) != 1)
2304 nxt = io_req_find_next(req);
2305 return nxt ? &nxt->work : NULL;
2308 static void io_double_put_req(struct io_kiocb *req)
2310 /* drop both submit and complete references */
2311 if (refcount_sub_and_test(2, &req->refs))
2315 static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush)
2317 struct io_rings *rings = ctx->rings;
2319 if (test_bit(0, &ctx->cq_check_overflow)) {
2321 * noflush == true is from the waitqueue handler, just ensure
2322 * we wake up the task, and the next invocation will flush the
2323 * entries. We cannot safely to it from here.
2328 io_cqring_overflow_flush(ctx, false, NULL, NULL);
2331 /* See comment at the top of this file */
2333 return ctx->cached_cq_tail - READ_ONCE(rings->cq.head);
2336 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2338 struct io_rings *rings = ctx->rings;
2340 /* make sure SQ entry isn't read before tail */
2341 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2344 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2346 unsigned int cflags;
2348 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2349 cflags |= IORING_CQE_F_BUFFER;
2350 req->flags &= ~REQ_F_BUFFER_SELECTED;
2355 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2357 struct io_buffer *kbuf;
2359 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2360 return io_put_kbuf(req, kbuf);
2363 static inline bool io_run_task_work(void)
2366 * Not safe to run on exiting task, and the task_work handling will
2367 * not add work to such a task.
2369 if (unlikely(current->flags & PF_EXITING))
2371 if (current->task_works) {
2372 __set_current_state(TASK_RUNNING);
2380 static void io_iopoll_queue(struct list_head *again)
2382 struct io_kiocb *req;
2385 req = list_first_entry(again, struct io_kiocb, inflight_entry);
2386 list_del(&req->inflight_entry);
2387 __io_complete_rw(req, -EAGAIN, 0, NULL);
2388 } while (!list_empty(again));
2392 * Find and free completed poll iocbs
2394 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2395 struct list_head *done)
2397 struct req_batch rb;
2398 struct io_kiocb *req;
2401 /* order with ->result store in io_complete_rw_iopoll() */
2404 io_init_req_batch(&rb);
2405 while (!list_empty(done)) {
2408 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2409 if (READ_ONCE(req->result) == -EAGAIN) {
2411 req->iopoll_completed = 0;
2412 list_move_tail(&req->inflight_entry, &again);
2415 list_del(&req->inflight_entry);
2417 if (req->flags & REQ_F_BUFFER_SELECTED)
2418 cflags = io_put_rw_kbuf(req);
2420 __io_cqring_fill_event(req, req->result, cflags);
2423 if (refcount_dec_and_test(&req->refs))
2424 io_req_free_batch(&rb, req);
2427 io_commit_cqring(ctx);
2428 if (ctx->flags & IORING_SETUP_SQPOLL)
2429 io_cqring_ev_posted(ctx);
2430 io_req_free_batch_finish(ctx, &rb);
2432 if (!list_empty(&again))
2433 io_iopoll_queue(&again);
2436 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2439 struct io_kiocb *req, *tmp;
2445 * Only spin for completions if we don't have multiple devices hanging
2446 * off our complete list, and we're under the requested amount.
2448 spin = !ctx->poll_multi_file && *nr_events < min;
2451 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2452 struct kiocb *kiocb = &req->rw.kiocb;
2455 * Move completed and retryable entries to our local lists.
2456 * If we find a request that requires polling, break out
2457 * and complete those lists first, if we have entries there.
2459 if (READ_ONCE(req->iopoll_completed)) {
2460 list_move_tail(&req->inflight_entry, &done);
2463 if (!list_empty(&done))
2466 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2470 /* iopoll may have completed current req */
2471 if (READ_ONCE(req->iopoll_completed))
2472 list_move_tail(&req->inflight_entry, &done);
2479 if (!list_empty(&done))
2480 io_iopoll_complete(ctx, nr_events, &done);
2486 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2487 * non-spinning poll check - we'll still enter the driver poll loop, but only
2488 * as a non-spinning completion check.
2490 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
2493 while (!list_empty(&ctx->iopoll_list) && !need_resched()) {
2496 ret = io_do_iopoll(ctx, nr_events, min);
2499 if (*nr_events >= min)
2507 * We can't just wait for polled events to come to us, we have to actively
2508 * find and complete them.
2510 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2512 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2515 mutex_lock(&ctx->uring_lock);
2516 while (!list_empty(&ctx->iopoll_list)) {
2517 unsigned int nr_events = 0;
2519 io_do_iopoll(ctx, &nr_events, 0);
2521 /* let it sleep and repeat later if can't complete a request */
2525 * Ensure we allow local-to-the-cpu processing to take place,
2526 * in this case we need to ensure that we reap all events.
2527 * Also let task_work, etc. to progress by releasing the mutex
2529 if (need_resched()) {
2530 mutex_unlock(&ctx->uring_lock);
2532 mutex_lock(&ctx->uring_lock);
2535 mutex_unlock(&ctx->uring_lock);
2538 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2540 unsigned int nr_events = 0;
2541 int iters = 0, ret = 0;
2544 * We disallow the app entering submit/complete with polling, but we
2545 * still need to lock the ring to prevent racing with polled issue
2546 * that got punted to a workqueue.
2548 mutex_lock(&ctx->uring_lock);
2551 * Don't enter poll loop if we already have events pending.
2552 * If we do, we can potentially be spinning for commands that
2553 * already triggered a CQE (eg in error).
2555 if (io_cqring_events(ctx, false))
2559 * If a submit got punted to a workqueue, we can have the
2560 * application entering polling for a command before it gets
2561 * issued. That app will hold the uring_lock for the duration
2562 * of the poll right here, so we need to take a breather every
2563 * now and then to ensure that the issue has a chance to add
2564 * the poll to the issued list. Otherwise we can spin here
2565 * forever, while the workqueue is stuck trying to acquire the
2568 if (!(++iters & 7)) {
2569 mutex_unlock(&ctx->uring_lock);
2571 mutex_lock(&ctx->uring_lock);
2574 ret = io_iopoll_getevents(ctx, &nr_events, min);
2578 } while (min && !nr_events && !need_resched());
2580 mutex_unlock(&ctx->uring_lock);
2584 static void kiocb_end_write(struct io_kiocb *req)
2587 * Tell lockdep we inherited freeze protection from submission
2590 if (req->flags & REQ_F_ISREG) {
2591 struct inode *inode = file_inode(req->file);
2593 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
2595 file_end_write(req->file);
2598 static void io_complete_rw_common(struct kiocb *kiocb, long res,
2599 struct io_comp_state *cs)
2601 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2604 if (kiocb->ki_flags & IOCB_WRITE)
2605 kiocb_end_write(req);
2607 if (res != req->result)
2608 req_set_fail_links(req);
2609 if (req->flags & REQ_F_BUFFER_SELECTED)
2610 cflags = io_put_rw_kbuf(req);
2611 __io_req_complete(req, res, cflags, cs);
2615 static bool io_resubmit_prep(struct io_kiocb *req, int error)
2617 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2618 ssize_t ret = -ECANCELED;
2619 struct iov_iter iter;
2627 switch (req->opcode) {
2628 case IORING_OP_READV:
2629 case IORING_OP_READ_FIXED:
2630 case IORING_OP_READ:
2633 case IORING_OP_WRITEV:
2634 case IORING_OP_WRITE_FIXED:
2635 case IORING_OP_WRITE:
2639 printk_once(KERN_WARNING "io_uring: bad opcode in resubmit %d\n",
2644 if (!req->async_data) {
2645 ret = io_import_iovec(rw, req, &iovec, &iter, false);
2648 ret = io_setup_async_rw(req, iovec, inline_vecs, &iter, false);
2656 req_set_fail_links(req);
2661 static bool io_rw_reissue(struct io_kiocb *req, long res)
2664 umode_t mode = file_inode(req->file)->i_mode;
2667 if (!S_ISBLK(mode) && !S_ISREG(mode))
2669 if ((res != -EAGAIN && res != -EOPNOTSUPP) || io_wq_current_is_worker())
2672 ret = io_sq_thread_acquire_mm_files(req->ctx, req);
2674 if (io_resubmit_prep(req, ret)) {
2675 refcount_inc(&req->refs);
2676 io_queue_async_work(req);
2684 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2685 struct io_comp_state *cs)
2687 if (!io_rw_reissue(req, res))
2688 io_complete_rw_common(&req->rw.kiocb, res, cs);
2691 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2693 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2695 __io_complete_rw(req, res, res2, NULL);
2698 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2700 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2702 if (kiocb->ki_flags & IOCB_WRITE)
2703 kiocb_end_write(req);
2705 if (res != -EAGAIN && res != req->result)
2706 req_set_fail_links(req);
2708 WRITE_ONCE(req->result, res);
2709 /* order with io_poll_complete() checking ->result */
2711 WRITE_ONCE(req->iopoll_completed, 1);
2715 * After the iocb has been issued, it's safe to be found on the poll list.
2716 * Adding the kiocb to the list AFTER submission ensures that we don't
2717 * find it from a io_iopoll_getevents() thread before the issuer is done
2718 * accessing the kiocb cookie.
2720 static void io_iopoll_req_issued(struct io_kiocb *req, bool in_async)
2722 struct io_ring_ctx *ctx = req->ctx;
2725 * Track whether we have multiple files in our lists. This will impact
2726 * how we do polling eventually, not spinning if we're on potentially
2727 * different devices.
2729 if (list_empty(&ctx->iopoll_list)) {
2730 ctx->poll_multi_file = false;
2731 } else if (!ctx->poll_multi_file) {
2732 struct io_kiocb *list_req;
2734 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2736 if (list_req->file != req->file)
2737 ctx->poll_multi_file = true;
2741 * For fast devices, IO may have already completed. If it has, add
2742 * it to the front so we find it first.
2744 if (READ_ONCE(req->iopoll_completed))
2745 list_add(&req->inflight_entry, &ctx->iopoll_list);
2747 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2750 * If IORING_SETUP_SQPOLL is enabled, sqes are either handled in sq thread
2751 * task context or in io worker task context. If current task context is
2752 * sq thread, we don't need to check whether should wake up sq thread.
2754 if (in_async && (ctx->flags & IORING_SETUP_SQPOLL) &&
2755 wq_has_sleeper(&ctx->sq_data->wait))
2756 wake_up(&ctx->sq_data->wait);
2759 static inline void __io_state_file_put(struct io_submit_state *state)
2761 fput_many(state->file, state->file_refs);
2762 state->file_refs = 0;
2765 static inline void io_state_file_put(struct io_submit_state *state)
2767 if (state->file_refs)
2768 __io_state_file_put(state);
2772 * Get as many references to a file as we have IOs left in this submission,
2773 * assuming most submissions are for one file, or at least that each file
2774 * has more than one submission.
2776 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2781 if (state->file_refs) {
2782 if (state->fd == fd) {
2786 __io_state_file_put(state);
2788 state->file = fget_many(fd, state->ios_left);
2789 if (unlikely(!state->file))
2793 state->file_refs = state->ios_left - 1;
2797 static bool io_bdev_nowait(struct block_device *bdev)
2799 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2803 * If we tracked the file through the SCM inflight mechanism, we could support
2804 * any file. For now, just ensure that anything potentially problematic is done
2807 static bool io_file_supports_async(struct file *file, int rw)
2809 umode_t mode = file_inode(file)->i_mode;
2811 if (S_ISBLK(mode)) {
2812 if (IS_ENABLED(CONFIG_BLOCK) &&
2813 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2817 if (S_ISCHR(mode) || S_ISSOCK(mode))
2819 if (S_ISREG(mode)) {
2820 if (IS_ENABLED(CONFIG_BLOCK) &&
2821 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2822 file->f_op != &io_uring_fops)
2827 /* any ->read/write should understand O_NONBLOCK */
2828 if (file->f_flags & O_NONBLOCK)
2831 if (!(file->f_mode & FMODE_NOWAIT))
2835 return file->f_op->read_iter != NULL;
2837 return file->f_op->write_iter != NULL;
2840 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2842 struct io_ring_ctx *ctx = req->ctx;
2843 struct kiocb *kiocb = &req->rw.kiocb;
2847 if (S_ISREG(file_inode(req->file)->i_mode))
2848 req->flags |= REQ_F_ISREG;
2850 kiocb->ki_pos = READ_ONCE(sqe->off);
2851 if (kiocb->ki_pos == -1 && !(req->file->f_mode & FMODE_STREAM)) {
2852 req->flags |= REQ_F_CUR_POS;
2853 kiocb->ki_pos = req->file->f_pos;
2855 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2856 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2857 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2861 ioprio = READ_ONCE(sqe->ioprio);
2863 ret = ioprio_check_cap(ioprio);
2867 kiocb->ki_ioprio = ioprio;
2869 kiocb->ki_ioprio = get_current_ioprio();
2871 /* don't allow async punt if RWF_NOWAIT was requested */
2872 if (kiocb->ki_flags & IOCB_NOWAIT)
2873 req->flags |= REQ_F_NOWAIT;
2875 if (ctx->flags & IORING_SETUP_IOPOLL) {
2876 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2877 !kiocb->ki_filp->f_op->iopoll)
2880 kiocb->ki_flags |= IOCB_HIPRI;
2881 kiocb->ki_complete = io_complete_rw_iopoll;
2882 req->iopoll_completed = 0;
2884 if (kiocb->ki_flags & IOCB_HIPRI)
2886 kiocb->ki_complete = io_complete_rw;
2889 req->rw.addr = READ_ONCE(sqe->addr);
2890 req->rw.len = READ_ONCE(sqe->len);
2891 req->buf_index = READ_ONCE(sqe->buf_index);
2895 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2901 case -ERESTARTNOINTR:
2902 case -ERESTARTNOHAND:
2903 case -ERESTART_RESTARTBLOCK:
2905 * We can't just restart the syscall, since previously
2906 * submitted sqes may already be in progress. Just fail this
2912 kiocb->ki_complete(kiocb, ret, 0);
2916 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2917 struct io_comp_state *cs)
2919 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2920 struct io_async_rw *io = req->async_data;
2922 /* add previously done IO, if any */
2923 if (io && io->bytes_done > 0) {
2925 ret = io->bytes_done;
2927 ret += io->bytes_done;
2930 if (req->flags & REQ_F_CUR_POS)
2931 req->file->f_pos = kiocb->ki_pos;
2932 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2933 __io_complete_rw(req, ret, 0, cs);
2935 io_rw_done(kiocb, ret);
2938 static ssize_t io_import_fixed(struct io_kiocb *req, int rw,
2939 struct iov_iter *iter)
2941 struct io_ring_ctx *ctx = req->ctx;
2942 size_t len = req->rw.len;
2943 struct io_mapped_ubuf *imu;
2944 u16 index, buf_index = req->buf_index;
2948 if (unlikely(buf_index >= ctx->nr_user_bufs))
2950 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2951 imu = &ctx->user_bufs[index];
2952 buf_addr = req->rw.addr;
2955 if (buf_addr + len < buf_addr)
2957 /* not inside the mapped region */
2958 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2962 * May not be a start of buffer, set size appropriately
2963 * and advance us to the beginning.
2965 offset = buf_addr - imu->ubuf;
2966 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2970 * Don't use iov_iter_advance() here, as it's really slow for
2971 * using the latter parts of a big fixed buffer - it iterates
2972 * over each segment manually. We can cheat a bit here, because
2975 * 1) it's a BVEC iter, we set it up
2976 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2977 * first and last bvec
2979 * So just find our index, and adjust the iterator afterwards.
2980 * If the offset is within the first bvec (or the whole first
2981 * bvec, just use iov_iter_advance(). This makes it easier
2982 * since we can just skip the first segment, which may not
2983 * be PAGE_SIZE aligned.
2985 const struct bio_vec *bvec = imu->bvec;
2987 if (offset <= bvec->bv_len) {
2988 iov_iter_advance(iter, offset);
2990 unsigned long seg_skip;
2992 /* skip first vec */
2993 offset -= bvec->bv_len;
2994 seg_skip = 1 + (offset >> PAGE_SHIFT);
2996 iter->bvec = bvec + seg_skip;
2997 iter->nr_segs -= seg_skip;
2998 iter->count -= bvec->bv_len + offset;
2999 iter->iov_offset = offset & ~PAGE_MASK;
3006 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
3009 mutex_unlock(&ctx->uring_lock);
3012 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
3015 * "Normal" inline submissions always hold the uring_lock, since we
3016 * grab it from the system call. Same is true for the SQPOLL offload.
3017 * The only exception is when we've detached the request and issue it
3018 * from an async worker thread, grab the lock for that case.
3021 mutex_lock(&ctx->uring_lock);
3024 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
3025 int bgid, struct io_buffer *kbuf,
3028 struct io_buffer *head;
3030 if (req->flags & REQ_F_BUFFER_SELECTED)
3033 io_ring_submit_lock(req->ctx, needs_lock);
3035 lockdep_assert_held(&req->ctx->uring_lock);
3037 head = idr_find(&req->ctx->io_buffer_idr, bgid);
3039 if (!list_empty(&head->list)) {
3040 kbuf = list_last_entry(&head->list, struct io_buffer,
3042 list_del(&kbuf->list);
3045 idr_remove(&req->ctx->io_buffer_idr, bgid);
3047 if (*len > kbuf->len)
3050 kbuf = ERR_PTR(-ENOBUFS);
3053 io_ring_submit_unlock(req->ctx, needs_lock);
3058 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
3061 struct io_buffer *kbuf;
3064 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3065 bgid = req->buf_index;
3066 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
3069 req->rw.addr = (u64) (unsigned long) kbuf;
3070 req->flags |= REQ_F_BUFFER_SELECTED;
3071 return u64_to_user_ptr(kbuf->addr);
3074 #ifdef CONFIG_COMPAT
3075 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
3078 struct compat_iovec __user *uiov;
3079 compat_ssize_t clen;
3083 uiov = u64_to_user_ptr(req->rw.addr);
3084 if (!access_ok(uiov, sizeof(*uiov)))
3086 if (__get_user(clen, &uiov->iov_len))
3092 buf = io_rw_buffer_select(req, &len, needs_lock);
3094 return PTR_ERR(buf);
3095 iov[0].iov_base = buf;
3096 iov[0].iov_len = (compat_size_t) len;
3101 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3104 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
3108 if (copy_from_user(iov, uiov, sizeof(*uiov)))
3111 len = iov[0].iov_len;
3114 buf = io_rw_buffer_select(req, &len, needs_lock);
3116 return PTR_ERR(buf);
3117 iov[0].iov_base = buf;
3118 iov[0].iov_len = len;
3122 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3125 if (req->flags & REQ_F_BUFFER_SELECTED) {
3126 struct io_buffer *kbuf;
3128 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3129 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
3130 iov[0].iov_len = kbuf->len;
3135 else if (req->rw.len > 1)
3138 #ifdef CONFIG_COMPAT
3139 if (req->ctx->compat)
3140 return io_compat_import(req, iov, needs_lock);
3143 return __io_iov_buffer_select(req, iov, needs_lock);
3146 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
3147 struct iovec **iovec, struct iov_iter *iter,
3150 void __user *buf = u64_to_user_ptr(req->rw.addr);
3151 size_t sqe_len = req->rw.len;
3155 opcode = req->opcode;
3156 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3158 return io_import_fixed(req, rw, iter);
3161 /* buffer index only valid with fixed read/write, or buffer select */
3162 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3165 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3166 if (req->flags & REQ_F_BUFFER_SELECT) {
3167 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3169 return PTR_ERR(buf);
3170 req->rw.len = sqe_len;
3173 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3178 if (req->flags & REQ_F_BUFFER_SELECT) {
3179 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3181 ret = (*iovec)->iov_len;
3182 iov_iter_init(iter, rw, *iovec, 1, ret);
3188 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3192 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3194 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3198 * For files that don't have ->read_iter() and ->write_iter(), handle them
3199 * by looping over ->read() or ->write() manually.
3201 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3203 struct kiocb *kiocb = &req->rw.kiocb;
3204 struct file *file = req->file;
3208 * Don't support polled IO through this interface, and we can't
3209 * support non-blocking either. For the latter, this just causes
3210 * the kiocb to be handled from an async context.
3212 if (kiocb->ki_flags & IOCB_HIPRI)
3214 if (kiocb->ki_flags & IOCB_NOWAIT)
3217 while (iov_iter_count(iter)) {
3221 if (!iov_iter_is_bvec(iter)) {
3222 iovec = iov_iter_iovec(iter);
3224 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3225 iovec.iov_len = req->rw.len;
3229 nr = file->f_op->read(file, iovec.iov_base,
3230 iovec.iov_len, io_kiocb_ppos(kiocb));
3232 nr = file->f_op->write(file, iovec.iov_base,
3233 iovec.iov_len, io_kiocb_ppos(kiocb));
3242 if (nr != iovec.iov_len)
3246 iov_iter_advance(iter, nr);
3252 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3253 const struct iovec *fast_iov, struct iov_iter *iter)
3255 struct io_async_rw *rw = req->async_data;
3257 memcpy(&rw->iter, iter, sizeof(*iter));
3258 rw->free_iovec = iovec;
3260 /* can only be fixed buffers, no need to do anything */
3261 if (iov_iter_is_bvec(iter))
3264 unsigned iov_off = 0;
3266 rw->iter.iov = rw->fast_iov;
3267 if (iter->iov != fast_iov) {
3268 iov_off = iter->iov - fast_iov;
3269 rw->iter.iov += iov_off;
3271 if (rw->fast_iov != fast_iov)
3272 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3273 sizeof(struct iovec) * iter->nr_segs);
3275 req->flags |= REQ_F_NEED_CLEANUP;
3279 static inline int __io_alloc_async_data(struct io_kiocb *req)
3281 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3282 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3283 return req->async_data == NULL;
3286 static int io_alloc_async_data(struct io_kiocb *req)
3288 if (!io_op_defs[req->opcode].needs_async_data)
3291 return __io_alloc_async_data(req);
3294 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3295 const struct iovec *fast_iov,
3296 struct iov_iter *iter, bool force)
3298 if (!force && !io_op_defs[req->opcode].needs_async_data)
3300 if (!req->async_data) {
3301 if (__io_alloc_async_data(req))
3304 io_req_map_rw(req, iovec, fast_iov, iter);
3309 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3311 struct io_async_rw *iorw = req->async_data;
3312 struct iovec *iov = iorw->fast_iov;
3315 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3316 if (unlikely(ret < 0))
3319 iorw->bytes_done = 0;
3320 iorw->free_iovec = iov;
3322 req->flags |= REQ_F_NEED_CLEANUP;
3326 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3330 ret = io_prep_rw(req, sqe);
3334 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3337 /* either don't need iovec imported or already have it */
3338 if (!req->async_data)
3340 return io_rw_prep_async(req, READ);
3344 * This is our waitqueue callback handler, registered through lock_page_async()
3345 * when we initially tried to do the IO with the iocb armed our waitqueue.
3346 * This gets called when the page is unlocked, and we generally expect that to
3347 * happen when the page IO is completed and the page is now uptodate. This will
3348 * queue a task_work based retry of the operation, attempting to copy the data
3349 * again. If the latter fails because the page was NOT uptodate, then we will
3350 * do a thread based blocking retry of the operation. That's the unexpected
3353 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3354 int sync, void *arg)
3356 struct wait_page_queue *wpq;
3357 struct io_kiocb *req = wait->private;
3358 struct wait_page_key *key = arg;
3361 wpq = container_of(wait, struct wait_page_queue, wait);
3363 if (!wake_page_match(wpq, key))
3366 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3367 list_del_init(&wait->entry);
3369 init_task_work(&req->task_work, io_req_task_submit);
3370 percpu_ref_get(&req->ctx->refs);
3372 /* submit ref gets dropped, acquire a new one */
3373 refcount_inc(&req->refs);
3374 ret = io_req_task_work_add(req);
3375 if (unlikely(ret)) {
3376 struct task_struct *tsk;
3378 /* queue just for cancelation */
3379 init_task_work(&req->task_work, io_req_task_cancel);
3380 tsk = io_wq_get_task(req->ctx->io_wq);
3381 task_work_add(tsk, &req->task_work, TWA_NONE);
3382 wake_up_process(tsk);
3388 * This controls whether a given IO request should be armed for async page
3389 * based retry. If we return false here, the request is handed to the async
3390 * worker threads for retry. If we're doing buffered reads on a regular file,
3391 * we prepare a private wait_page_queue entry and retry the operation. This
3392 * will either succeed because the page is now uptodate and unlocked, or it
3393 * will register a callback when the page is unlocked at IO completion. Through
3394 * that callback, io_uring uses task_work to setup a retry of the operation.
3395 * That retry will attempt the buffered read again. The retry will generally
3396 * succeed, or in rare cases where it fails, we then fall back to using the
3397 * async worker threads for a blocking retry.
3399 static bool io_rw_should_retry(struct io_kiocb *req)
3401 struct io_async_rw *rw = req->async_data;
3402 struct wait_page_queue *wait = &rw->wpq;
3403 struct kiocb *kiocb = &req->rw.kiocb;
3405 /* never retry for NOWAIT, we just complete with -EAGAIN */
3406 if (req->flags & REQ_F_NOWAIT)
3409 /* Only for buffered IO */
3410 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3414 * just use poll if we can, and don't attempt if the fs doesn't
3415 * support callback based unlocks
3417 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3420 wait->wait.func = io_async_buf_func;
3421 wait->wait.private = req;
3422 wait->wait.flags = 0;
3423 INIT_LIST_HEAD(&wait->wait.entry);
3424 kiocb->ki_flags |= IOCB_WAITQ;
3425 kiocb->ki_flags &= ~IOCB_NOWAIT;
3426 kiocb->ki_waitq = wait;
3430 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3432 if (req->file->f_op->read_iter)
3433 return call_read_iter(req->file, &req->rw.kiocb, iter);
3434 else if (req->file->f_op->read)
3435 return loop_rw_iter(READ, req, iter);
3440 static int io_read(struct io_kiocb *req, bool force_nonblock,
3441 struct io_comp_state *cs)
3443 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3444 struct kiocb *kiocb = &req->rw.kiocb;
3445 struct iov_iter __iter, *iter = &__iter;
3446 struct io_async_rw *rw = req->async_data;
3447 ssize_t io_size, ret, ret2;
3454 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3458 io_size = iov_iter_count(iter);
3459 req->result = io_size;
3462 /* Ensure we clear previously set non-block flag */
3463 if (!force_nonblock)
3464 kiocb->ki_flags &= ~IOCB_NOWAIT;
3466 kiocb->ki_flags |= IOCB_NOWAIT;
3469 /* If the file doesn't support async, just async punt */
3470 no_async = force_nonblock && !io_file_supports_async(req->file, READ);
3474 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size);
3478 ret = io_iter_do_read(req, iter);
3482 } else if (ret == -EIOCBQUEUED) {
3485 } else if (ret == -EAGAIN) {
3486 /* IOPOLL retry should happen for io-wq threads */
3487 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3489 /* no retry on NONBLOCK marked file */
3490 if (req->file->f_flags & O_NONBLOCK)
3492 /* some cases will consume bytes even on error returns */
3493 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3496 } else if (ret < 0) {
3497 /* make sure -ERESTARTSYS -> -EINTR is done */
3501 /* read it all, or we did blocking attempt. no retry. */
3502 if (!iov_iter_count(iter) || !force_nonblock ||
3503 (req->file->f_flags & O_NONBLOCK))
3508 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3515 rw = req->async_data;
3516 /* it's copied and will be cleaned with ->io */
3518 /* now use our persistent iterator, if we aren't already */
3521 rw->bytes_done += ret;
3522 /* if we can retry, do so with the callbacks armed */
3523 if (!io_rw_should_retry(req)) {
3524 kiocb->ki_flags &= ~IOCB_WAITQ;
3529 * Now retry read with the IOCB_WAITQ parts set in the iocb. If we
3530 * get -EIOCBQUEUED, then we'll get a notification when the desired
3531 * page gets unlocked. We can also get a partial read here, and if we
3532 * do, then just retry at the new offset.
3534 ret = io_iter_do_read(req, iter);
3535 if (ret == -EIOCBQUEUED) {
3538 } else if (ret > 0 && ret < io_size) {
3539 /* we got some bytes, but not all. retry. */
3543 kiocb_done(kiocb, ret, cs);
3546 /* it's reportedly faster than delegating the null check to kfree() */
3552 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3556 ret = io_prep_rw(req, sqe);
3560 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3563 /* either don't need iovec imported or already have it */
3564 if (!req->async_data)
3566 return io_rw_prep_async(req, WRITE);
3569 static int io_write(struct io_kiocb *req, bool force_nonblock,
3570 struct io_comp_state *cs)
3572 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3573 struct kiocb *kiocb = &req->rw.kiocb;
3574 struct iov_iter __iter, *iter = &__iter;
3575 struct io_async_rw *rw = req->async_data;
3576 ssize_t ret, ret2, io_size;
3582 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3586 io_size = iov_iter_count(iter);
3587 req->result = io_size;
3589 /* Ensure we clear previously set non-block flag */
3590 if (!force_nonblock)
3591 kiocb->ki_flags &= ~IOCB_NOWAIT;
3593 kiocb->ki_flags |= IOCB_NOWAIT;
3595 /* If the file doesn't support async, just async punt */
3596 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
3599 /* file path doesn't support NOWAIT for non-direct_IO */
3600 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3601 (req->flags & REQ_F_ISREG))
3604 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size);
3609 * Open-code file_start_write here to grab freeze protection,
3610 * which will be released by another thread in
3611 * io_complete_rw(). Fool lockdep by telling it the lock got
3612 * released so that it doesn't complain about the held lock when
3613 * we return to userspace.
3615 if (req->flags & REQ_F_ISREG) {
3616 sb_start_write(file_inode(req->file)->i_sb);
3617 __sb_writers_release(file_inode(req->file)->i_sb,
3620 kiocb->ki_flags |= IOCB_WRITE;
3622 if (req->file->f_op->write_iter)
3623 ret2 = call_write_iter(req->file, kiocb, iter);
3624 else if (req->file->f_op->write)
3625 ret2 = loop_rw_iter(WRITE, req, iter);
3630 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3631 * retry them without IOCB_NOWAIT.
3633 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3635 /* no retry on NONBLOCK marked file */
3636 if (ret2 == -EAGAIN && (req->file->f_flags & O_NONBLOCK))
3638 if (!force_nonblock || ret2 != -EAGAIN) {
3639 /* IOPOLL retry should happen for io-wq threads */
3640 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3643 kiocb_done(kiocb, ret2, cs);
3646 /* some cases will consume bytes even on error returns */
3647 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3648 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3653 /* it's reportedly faster than delegating the null check to kfree() */
3659 static int io_renameat_prep(struct io_kiocb *req,
3660 const struct io_uring_sqe *sqe)
3662 struct io_rename *ren = &req->rename;
3663 const char __user *oldf, *newf;
3665 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3668 ren->old_dfd = READ_ONCE(sqe->fd);
3669 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3670 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3671 ren->new_dfd = READ_ONCE(sqe->len);
3672 ren->flags = READ_ONCE(sqe->rename_flags);
3674 ren->oldpath = getname(oldf);
3675 if (IS_ERR(ren->oldpath))
3676 return PTR_ERR(ren->oldpath);
3678 ren->newpath = getname(newf);
3679 if (IS_ERR(ren->newpath)) {
3680 putname(ren->oldpath);
3681 return PTR_ERR(ren->newpath);
3684 req->flags |= REQ_F_NEED_CLEANUP;
3688 static int io_renameat(struct io_kiocb *req, bool force_nonblock)
3690 struct io_rename *ren = &req->rename;
3696 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3697 ren->newpath, ren->flags);
3699 req->flags &= ~REQ_F_NEED_CLEANUP;
3701 req_set_fail_links(req);
3702 io_req_complete(req, ret);
3706 static int io_unlinkat_prep(struct io_kiocb *req,
3707 const struct io_uring_sqe *sqe)
3709 struct io_unlink *un = &req->unlink;
3710 const char __user *fname;
3712 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3715 un->dfd = READ_ONCE(sqe->fd);
3717 un->flags = READ_ONCE(sqe->unlink_flags);
3718 if (un->flags & ~AT_REMOVEDIR)
3721 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3722 un->filename = getname(fname);
3723 if (IS_ERR(un->filename))
3724 return PTR_ERR(un->filename);
3726 req->flags |= REQ_F_NEED_CLEANUP;
3730 static int io_unlinkat(struct io_kiocb *req, bool force_nonblock)
3732 struct io_unlink *un = &req->unlink;
3738 if (un->flags & AT_REMOVEDIR)
3739 ret = do_rmdir(un->dfd, un->filename);
3741 ret = do_unlinkat(un->dfd, un->filename);
3743 req->flags &= ~REQ_F_NEED_CLEANUP;
3745 req_set_fail_links(req);
3746 io_req_complete(req, ret);
3750 static int io_shutdown_prep(struct io_kiocb *req,
3751 const struct io_uring_sqe *sqe)
3753 #if defined(CONFIG_NET)
3754 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3756 if (sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3760 req->shutdown.how = READ_ONCE(sqe->len);
3767 static int io_shutdown(struct io_kiocb *req, bool force_nonblock)
3769 #if defined(CONFIG_NET)
3770 struct socket *sock;
3776 sock = sock_from_file(req->file);
3777 if (unlikely(!sock))
3780 ret = __sys_shutdown_sock(sock, req->shutdown.how);
3782 req_set_fail_links(req);
3783 io_req_complete(req, ret);
3790 static int __io_splice_prep(struct io_kiocb *req,
3791 const struct io_uring_sqe *sqe)
3793 struct io_splice* sp = &req->splice;
3794 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3796 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3800 sp->len = READ_ONCE(sqe->len);
3801 sp->flags = READ_ONCE(sqe->splice_flags);
3803 if (unlikely(sp->flags & ~valid_flags))
3806 sp->file_in = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in),
3807 (sp->flags & SPLICE_F_FD_IN_FIXED));
3810 req->flags |= REQ_F_NEED_CLEANUP;
3812 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
3814 * Splice operation will be punted aync, and here need to
3815 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3817 io_req_init_async(req);
3818 req->work.flags |= IO_WQ_WORK_UNBOUND;
3824 static int io_tee_prep(struct io_kiocb *req,
3825 const struct io_uring_sqe *sqe)
3827 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3829 return __io_splice_prep(req, sqe);
3832 static int io_tee(struct io_kiocb *req, bool force_nonblock)
3834 struct io_splice *sp = &req->splice;
3835 struct file *in = sp->file_in;
3836 struct file *out = sp->file_out;
3837 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3843 ret = do_tee(in, out, sp->len, flags);
3845 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3846 req->flags &= ~REQ_F_NEED_CLEANUP;
3849 req_set_fail_links(req);
3850 io_req_complete(req, ret);
3854 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3856 struct io_splice* sp = &req->splice;
3858 sp->off_in = READ_ONCE(sqe->splice_off_in);
3859 sp->off_out = READ_ONCE(sqe->off);
3860 return __io_splice_prep(req, sqe);
3863 static int io_splice(struct io_kiocb *req, bool force_nonblock)
3865 struct io_splice *sp = &req->splice;
3866 struct file *in = sp->file_in;
3867 struct file *out = sp->file_out;
3868 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3869 loff_t *poff_in, *poff_out;
3875 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3876 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3879 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3881 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3882 req->flags &= ~REQ_F_NEED_CLEANUP;
3885 req_set_fail_links(req);
3886 io_req_complete(req, ret);
3891 * IORING_OP_NOP just posts a completion event, nothing else.
3893 static int io_nop(struct io_kiocb *req, struct io_comp_state *cs)
3895 struct io_ring_ctx *ctx = req->ctx;
3897 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3900 __io_req_complete(req, 0, 0, cs);
3904 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3906 struct io_ring_ctx *ctx = req->ctx;
3911 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3913 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3916 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3917 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3920 req->sync.off = READ_ONCE(sqe->off);
3921 req->sync.len = READ_ONCE(sqe->len);
3925 static int io_fsync(struct io_kiocb *req, bool force_nonblock)
3927 loff_t end = req->sync.off + req->sync.len;
3930 /* fsync always requires a blocking context */
3934 ret = vfs_fsync_range(req->file, req->sync.off,
3935 end > 0 ? end : LLONG_MAX,
3936 req->sync.flags & IORING_FSYNC_DATASYNC);
3938 req_set_fail_links(req);
3939 io_req_complete(req, ret);
3943 static int io_fallocate_prep(struct io_kiocb *req,
3944 const struct io_uring_sqe *sqe)
3946 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3948 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3951 req->sync.off = READ_ONCE(sqe->off);
3952 req->sync.len = READ_ONCE(sqe->addr);
3953 req->sync.mode = READ_ONCE(sqe->len);
3957 static int io_fallocate(struct io_kiocb *req, bool force_nonblock)
3961 /* fallocate always requiring blocking context */
3964 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3967 req_set_fail_links(req);
3968 io_req_complete(req, ret);
3972 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3974 const char __user *fname;
3977 if (unlikely(sqe->ioprio || sqe->buf_index))
3979 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3982 /* open.how should be already initialised */
3983 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3984 req->open.how.flags |= O_LARGEFILE;
3986 req->open.dfd = READ_ONCE(sqe->fd);
3987 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3988 req->open.filename = getname(fname);
3989 if (IS_ERR(req->open.filename)) {
3990 ret = PTR_ERR(req->open.filename);
3991 req->open.filename = NULL;
3994 req->open.nofile = rlimit(RLIMIT_NOFILE);
3995 req->open.ignore_nonblock = false;
3996 req->flags |= REQ_F_NEED_CLEANUP;
4000 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4004 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4006 mode = READ_ONCE(sqe->len);
4007 flags = READ_ONCE(sqe->open_flags);
4008 req->open.how = build_open_how(flags, mode);
4009 return __io_openat_prep(req, sqe);
4012 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4014 struct open_how __user *how;
4018 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4020 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4021 len = READ_ONCE(sqe->len);
4022 if (len < OPEN_HOW_SIZE_VER0)
4025 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
4030 return __io_openat_prep(req, sqe);
4033 static int io_openat2(struct io_kiocb *req, bool force_nonblock)
4035 struct open_flags op;
4039 if (force_nonblock && !req->open.ignore_nonblock)
4042 ret = build_open_flags(&req->open.how, &op);
4046 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
4050 file = do_filp_open(req->open.dfd, req->open.filename, &op);
4053 ret = PTR_ERR(file);
4055 * A work-around to ensure that /proc/self works that way
4056 * that it should - if we get -EOPNOTSUPP back, then assume
4057 * that proc_self_get_link() failed us because we're in async
4058 * context. We should be safe to retry this from the task
4059 * itself with force_nonblock == false set, as it should not
4060 * block on lookup. Would be nice to know this upfront and
4061 * avoid the async dance, but doesn't seem feasible.
4063 if (ret == -EOPNOTSUPP && io_wq_current_is_worker()) {
4064 req->open.ignore_nonblock = true;
4065 refcount_inc(&req->refs);
4066 io_req_task_queue(req);
4070 fsnotify_open(file);
4071 fd_install(ret, file);
4074 putname(req->open.filename);
4075 req->flags &= ~REQ_F_NEED_CLEANUP;
4077 req_set_fail_links(req);
4078 io_req_complete(req, ret);
4082 static int io_openat(struct io_kiocb *req, bool force_nonblock)
4084 return io_openat2(req, force_nonblock);
4087 static int io_remove_buffers_prep(struct io_kiocb *req,
4088 const struct io_uring_sqe *sqe)
4090 struct io_provide_buf *p = &req->pbuf;
4093 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
4096 tmp = READ_ONCE(sqe->fd);
4097 if (!tmp || tmp > USHRT_MAX)
4100 memset(p, 0, sizeof(*p));
4102 p->bgid = READ_ONCE(sqe->buf_group);
4106 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
4107 int bgid, unsigned nbufs)
4111 /* shouldn't happen */
4115 /* the head kbuf is the list itself */
4116 while (!list_empty(&buf->list)) {
4117 struct io_buffer *nxt;
4119 nxt = list_first_entry(&buf->list, struct io_buffer, list);
4120 list_del(&nxt->list);
4127 idr_remove(&ctx->io_buffer_idr, bgid);
4132 static int io_remove_buffers(struct io_kiocb *req, bool force_nonblock,
4133 struct io_comp_state *cs)
4135 struct io_provide_buf *p = &req->pbuf;
4136 struct io_ring_ctx *ctx = req->ctx;
4137 struct io_buffer *head;
4140 io_ring_submit_lock(ctx, !force_nonblock);
4142 lockdep_assert_held(&ctx->uring_lock);
4145 head = idr_find(&ctx->io_buffer_idr, p->bgid);
4147 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
4149 req_set_fail_links(req);
4151 /* need to hold the lock to complete IOPOLL requests */
4152 if (ctx->flags & IORING_SETUP_IOPOLL) {
4153 __io_req_complete(req, ret, 0, cs);
4154 io_ring_submit_unlock(ctx, !force_nonblock);
4156 io_ring_submit_unlock(ctx, !force_nonblock);
4157 __io_req_complete(req, ret, 0, cs);
4162 static int io_provide_buffers_prep(struct io_kiocb *req,
4163 const struct io_uring_sqe *sqe)
4165 struct io_provide_buf *p = &req->pbuf;
4168 if (sqe->ioprio || sqe->rw_flags)
4171 tmp = READ_ONCE(sqe->fd);
4172 if (!tmp || tmp > USHRT_MAX)
4175 p->addr = READ_ONCE(sqe->addr);
4176 p->len = READ_ONCE(sqe->len);
4178 if (!access_ok(u64_to_user_ptr(p->addr), (p->len * p->nbufs)))
4181 p->bgid = READ_ONCE(sqe->buf_group);
4182 tmp = READ_ONCE(sqe->off);
4183 if (tmp > USHRT_MAX)
4189 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
4191 struct io_buffer *buf;
4192 u64 addr = pbuf->addr;
4193 int i, bid = pbuf->bid;
4195 for (i = 0; i < pbuf->nbufs; i++) {
4196 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
4201 buf->len = pbuf->len;
4206 INIT_LIST_HEAD(&buf->list);
4209 list_add_tail(&buf->list, &(*head)->list);
4213 return i ? i : -ENOMEM;
4216 static int io_provide_buffers(struct io_kiocb *req, bool force_nonblock,
4217 struct io_comp_state *cs)
4219 struct io_provide_buf *p = &req->pbuf;
4220 struct io_ring_ctx *ctx = req->ctx;
4221 struct io_buffer *head, *list;
4224 io_ring_submit_lock(ctx, !force_nonblock);
4226 lockdep_assert_held(&ctx->uring_lock);
4228 list = head = idr_find(&ctx->io_buffer_idr, p->bgid);
4230 ret = io_add_buffers(p, &head);
4235 ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1,
4238 __io_remove_buffers(ctx, head, p->bgid, -1U);
4244 req_set_fail_links(req);
4246 /* need to hold the lock to complete IOPOLL requests */
4247 if (ctx->flags & IORING_SETUP_IOPOLL) {
4248 __io_req_complete(req, ret, 0, cs);
4249 io_ring_submit_unlock(ctx, !force_nonblock);
4251 io_ring_submit_unlock(ctx, !force_nonblock);
4252 __io_req_complete(req, ret, 0, cs);
4257 static int io_epoll_ctl_prep(struct io_kiocb *req,
4258 const struct io_uring_sqe *sqe)
4260 #if defined(CONFIG_EPOLL)
4261 if (sqe->ioprio || sqe->buf_index)
4263 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4266 req->epoll.epfd = READ_ONCE(sqe->fd);
4267 req->epoll.op = READ_ONCE(sqe->len);
4268 req->epoll.fd = READ_ONCE(sqe->off);
4270 if (ep_op_has_event(req->epoll.op)) {
4271 struct epoll_event __user *ev;
4273 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4274 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4284 static int io_epoll_ctl(struct io_kiocb *req, bool force_nonblock,
4285 struct io_comp_state *cs)
4287 #if defined(CONFIG_EPOLL)
4288 struct io_epoll *ie = &req->epoll;
4291 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4292 if (force_nonblock && ret == -EAGAIN)
4296 req_set_fail_links(req);
4297 __io_req_complete(req, ret, 0, cs);
4304 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4306 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4307 if (sqe->ioprio || sqe->buf_index || sqe->off)
4309 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4312 req->madvise.addr = READ_ONCE(sqe->addr);
4313 req->madvise.len = READ_ONCE(sqe->len);
4314 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4321 static int io_madvise(struct io_kiocb *req, bool force_nonblock)
4323 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4324 struct io_madvise *ma = &req->madvise;
4330 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4332 req_set_fail_links(req);
4333 io_req_complete(req, ret);
4340 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4342 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4344 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4347 req->fadvise.offset = READ_ONCE(sqe->off);
4348 req->fadvise.len = READ_ONCE(sqe->len);
4349 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4353 static int io_fadvise(struct io_kiocb *req, bool force_nonblock)
4355 struct io_fadvise *fa = &req->fadvise;
4358 if (force_nonblock) {
4359 switch (fa->advice) {
4360 case POSIX_FADV_NORMAL:
4361 case POSIX_FADV_RANDOM:
4362 case POSIX_FADV_SEQUENTIAL:
4369 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4371 req_set_fail_links(req);
4372 io_req_complete(req, ret);
4376 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4378 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4380 if (sqe->ioprio || sqe->buf_index)
4382 if (req->flags & REQ_F_FIXED_FILE)
4385 req->statx.dfd = READ_ONCE(sqe->fd);
4386 req->statx.mask = READ_ONCE(sqe->len);
4387 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4388 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4389 req->statx.flags = READ_ONCE(sqe->statx_flags);
4394 static int io_statx(struct io_kiocb *req, bool force_nonblock)
4396 struct io_statx *ctx = &req->statx;
4399 if (force_nonblock) {
4400 /* only need file table for an actual valid fd */
4401 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
4402 req->flags |= REQ_F_NO_FILE_TABLE;
4406 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4410 req_set_fail_links(req);
4411 io_req_complete(req, ret);
4415 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4418 * If we queue this for async, it must not be cancellable. That would
4419 * leave the 'file' in an undeterminate state, and here need to modify
4420 * io_wq_work.flags, so initialize io_wq_work firstly.
4422 io_req_init_async(req);
4423 req->work.flags |= IO_WQ_WORK_NO_CANCEL;
4425 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4427 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4428 sqe->rw_flags || sqe->buf_index)
4430 if (req->flags & REQ_F_FIXED_FILE)
4433 req->close.fd = READ_ONCE(sqe->fd);
4434 if ((req->file && req->file->f_op == &io_uring_fops))
4437 req->close.put_file = NULL;
4441 static int io_close(struct io_kiocb *req, bool force_nonblock,
4442 struct io_comp_state *cs)
4444 struct io_close *close = &req->close;
4447 /* might be already done during nonblock submission */
4448 if (!close->put_file) {
4449 ret = close_fd_get_file(close->fd, &close->put_file);
4451 return (ret == -ENOENT) ? -EBADF : ret;
4454 /* if the file has a flush method, be safe and punt to async */
4455 if (close->put_file->f_op->flush && force_nonblock) {
4456 /* was never set, but play safe */
4457 req->flags &= ~REQ_F_NOWAIT;
4458 /* avoid grabbing files - we don't need the files */
4459 req->flags |= REQ_F_NO_FILE_TABLE;
4463 /* No ->flush() or already async, safely close from here */
4464 ret = filp_close(close->put_file, req->work.identity->files);
4466 req_set_fail_links(req);
4467 fput(close->put_file);
4468 close->put_file = NULL;
4469 __io_req_complete(req, ret, 0, cs);
4473 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4475 struct io_ring_ctx *ctx = req->ctx;
4480 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4482 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4485 req->sync.off = READ_ONCE(sqe->off);
4486 req->sync.len = READ_ONCE(sqe->len);
4487 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4491 static int io_sync_file_range(struct io_kiocb *req, bool force_nonblock)
4495 /* sync_file_range always requires a blocking context */
4499 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4502 req_set_fail_links(req);
4503 io_req_complete(req, ret);
4507 #if defined(CONFIG_NET)
4508 static int io_setup_async_msg(struct io_kiocb *req,
4509 struct io_async_msghdr *kmsg)
4511 struct io_async_msghdr *async_msg = req->async_data;
4515 if (io_alloc_async_data(req)) {
4516 if (kmsg->iov != kmsg->fast_iov)
4520 async_msg = req->async_data;
4521 req->flags |= REQ_F_NEED_CLEANUP;
4522 memcpy(async_msg, kmsg, sizeof(*kmsg));
4526 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4527 struct io_async_msghdr *iomsg)
4529 iomsg->iov = iomsg->fast_iov;
4530 iomsg->msg.msg_name = &iomsg->addr;
4531 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4532 req->sr_msg.msg_flags, &iomsg->iov);
4535 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4537 struct io_async_msghdr *async_msg = req->async_data;
4538 struct io_sr_msg *sr = &req->sr_msg;
4541 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4544 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4545 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4546 sr->len = READ_ONCE(sqe->len);
4548 #ifdef CONFIG_COMPAT
4549 if (req->ctx->compat)
4550 sr->msg_flags |= MSG_CMSG_COMPAT;
4553 if (!async_msg || !io_op_defs[req->opcode].needs_async_data)
4555 ret = io_sendmsg_copy_hdr(req, async_msg);
4557 req->flags |= REQ_F_NEED_CLEANUP;
4561 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4562 struct io_comp_state *cs)
4564 struct io_async_msghdr iomsg, *kmsg;
4565 struct socket *sock;
4569 sock = sock_from_file(req->file);
4570 if (unlikely(!sock))
4573 if (req->async_data) {
4574 kmsg = req->async_data;
4575 kmsg->msg.msg_name = &kmsg->addr;
4576 /* if iov is set, it's allocated already */
4578 kmsg->iov = kmsg->fast_iov;
4579 kmsg->msg.msg_iter.iov = kmsg->iov;
4581 ret = io_sendmsg_copy_hdr(req, &iomsg);
4587 flags = req->sr_msg.msg_flags;
4588 if (flags & MSG_DONTWAIT)
4589 req->flags |= REQ_F_NOWAIT;
4590 else if (force_nonblock)
4591 flags |= MSG_DONTWAIT;
4593 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4594 if (force_nonblock && ret == -EAGAIN)
4595 return io_setup_async_msg(req, kmsg);
4596 if (ret == -ERESTARTSYS)
4599 if (kmsg->iov != kmsg->fast_iov)
4601 req->flags &= ~REQ_F_NEED_CLEANUP;
4603 req_set_fail_links(req);
4604 __io_req_complete(req, ret, 0, cs);
4608 static int io_send(struct io_kiocb *req, bool force_nonblock,
4609 struct io_comp_state *cs)
4611 struct io_sr_msg *sr = &req->sr_msg;
4614 struct socket *sock;
4618 sock = sock_from_file(req->file);
4619 if (unlikely(!sock))
4622 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4626 msg.msg_name = NULL;
4627 msg.msg_control = NULL;
4628 msg.msg_controllen = 0;
4629 msg.msg_namelen = 0;
4631 flags = req->sr_msg.msg_flags;
4632 if (flags & MSG_DONTWAIT)
4633 req->flags |= REQ_F_NOWAIT;
4634 else if (force_nonblock)
4635 flags |= MSG_DONTWAIT;
4637 msg.msg_flags = flags;
4638 ret = sock_sendmsg(sock, &msg);
4639 if (force_nonblock && ret == -EAGAIN)
4641 if (ret == -ERESTARTSYS)
4645 req_set_fail_links(req);
4646 __io_req_complete(req, ret, 0, cs);
4650 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4651 struct io_async_msghdr *iomsg)
4653 struct io_sr_msg *sr = &req->sr_msg;
4654 struct iovec __user *uiov;
4658 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4659 &iomsg->uaddr, &uiov, &iov_len);
4663 if (req->flags & REQ_F_BUFFER_SELECT) {
4666 if (copy_from_user(iomsg->iov, uiov, sizeof(*uiov)))
4668 sr->len = iomsg->iov[0].iov_len;
4669 iov_iter_init(&iomsg->msg.msg_iter, READ, iomsg->iov, 1,
4673 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4674 &iomsg->iov, &iomsg->msg.msg_iter,
4683 #ifdef CONFIG_COMPAT
4684 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4685 struct io_async_msghdr *iomsg)
4687 struct compat_msghdr __user *msg_compat;
4688 struct io_sr_msg *sr = &req->sr_msg;
4689 struct compat_iovec __user *uiov;
4694 msg_compat = (struct compat_msghdr __user *) sr->umsg;
4695 ret = __get_compat_msghdr(&iomsg->msg, msg_compat, &iomsg->uaddr,
4700 uiov = compat_ptr(ptr);
4701 if (req->flags & REQ_F_BUFFER_SELECT) {
4702 compat_ssize_t clen;
4706 if (!access_ok(uiov, sizeof(*uiov)))
4708 if (__get_user(clen, &uiov->iov_len))
4713 iomsg->iov[0].iov_len = clen;
4716 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4717 UIO_FASTIOV, &iomsg->iov,
4718 &iomsg->msg.msg_iter, true);
4727 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4728 struct io_async_msghdr *iomsg)
4730 iomsg->msg.msg_name = &iomsg->addr;
4731 iomsg->iov = iomsg->fast_iov;
4733 #ifdef CONFIG_COMPAT
4734 if (req->ctx->compat)
4735 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4738 return __io_recvmsg_copy_hdr(req, iomsg);
4741 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4744 struct io_sr_msg *sr = &req->sr_msg;
4745 struct io_buffer *kbuf;
4747 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4752 req->flags |= REQ_F_BUFFER_SELECTED;
4756 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4758 return io_put_kbuf(req, req->sr_msg.kbuf);
4761 static int io_recvmsg_prep(struct io_kiocb *req,
4762 const struct io_uring_sqe *sqe)
4764 struct io_async_msghdr *async_msg = req->async_data;
4765 struct io_sr_msg *sr = &req->sr_msg;
4768 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4771 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4772 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4773 sr->len = READ_ONCE(sqe->len);
4774 sr->bgid = READ_ONCE(sqe->buf_group);
4776 #ifdef CONFIG_COMPAT
4777 if (req->ctx->compat)
4778 sr->msg_flags |= MSG_CMSG_COMPAT;
4781 if (!async_msg || !io_op_defs[req->opcode].needs_async_data)
4783 ret = io_recvmsg_copy_hdr(req, async_msg);
4785 req->flags |= REQ_F_NEED_CLEANUP;
4789 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4790 struct io_comp_state *cs)
4792 struct io_async_msghdr iomsg, *kmsg;
4793 struct socket *sock;
4794 struct io_buffer *kbuf;
4796 int ret, cflags = 0;
4798 sock = sock_from_file(req->file);
4799 if (unlikely(!sock))
4802 if (req->async_data) {
4803 kmsg = req->async_data;
4804 kmsg->msg.msg_name = &kmsg->addr;
4805 /* if iov is set, it's allocated already */
4807 kmsg->iov = kmsg->fast_iov;
4808 kmsg->msg.msg_iter.iov = kmsg->iov;
4810 ret = io_recvmsg_copy_hdr(req, &iomsg);
4816 if (req->flags & REQ_F_BUFFER_SELECT) {
4817 kbuf = io_recv_buffer_select(req, !force_nonblock);
4819 return PTR_ERR(kbuf);
4820 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4821 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->iov,
4822 1, req->sr_msg.len);
4825 flags = req->sr_msg.msg_flags;
4826 if (flags & MSG_DONTWAIT)
4827 req->flags |= REQ_F_NOWAIT;
4828 else if (force_nonblock)
4829 flags |= MSG_DONTWAIT;
4831 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4832 kmsg->uaddr, flags);
4833 if (force_nonblock && ret == -EAGAIN)
4834 return io_setup_async_msg(req, kmsg);
4835 if (ret == -ERESTARTSYS)
4838 if (req->flags & REQ_F_BUFFER_SELECTED)
4839 cflags = io_put_recv_kbuf(req);
4840 if (kmsg->iov != kmsg->fast_iov)
4842 req->flags &= ~REQ_F_NEED_CLEANUP;
4844 req_set_fail_links(req);
4845 __io_req_complete(req, ret, cflags, cs);
4849 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4850 struct io_comp_state *cs)
4852 struct io_buffer *kbuf;
4853 struct io_sr_msg *sr = &req->sr_msg;
4855 void __user *buf = sr->buf;
4856 struct socket *sock;
4859 int ret, cflags = 0;
4861 sock = sock_from_file(req->file);
4862 if (unlikely(!sock))
4865 if (req->flags & REQ_F_BUFFER_SELECT) {
4866 kbuf = io_recv_buffer_select(req, !force_nonblock);
4868 return PTR_ERR(kbuf);
4869 buf = u64_to_user_ptr(kbuf->addr);
4872 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4876 msg.msg_name = NULL;
4877 msg.msg_control = NULL;
4878 msg.msg_controllen = 0;
4879 msg.msg_namelen = 0;
4880 msg.msg_iocb = NULL;
4883 flags = req->sr_msg.msg_flags;
4884 if (flags & MSG_DONTWAIT)
4885 req->flags |= REQ_F_NOWAIT;
4886 else if (force_nonblock)
4887 flags |= MSG_DONTWAIT;
4889 ret = sock_recvmsg(sock, &msg, flags);
4890 if (force_nonblock && ret == -EAGAIN)
4892 if (ret == -ERESTARTSYS)
4895 if (req->flags & REQ_F_BUFFER_SELECTED)
4896 cflags = io_put_recv_kbuf(req);
4898 req_set_fail_links(req);
4899 __io_req_complete(req, ret, cflags, cs);
4903 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4905 struct io_accept *accept = &req->accept;
4907 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4909 if (sqe->ioprio || sqe->len || sqe->buf_index)
4912 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4913 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4914 accept->flags = READ_ONCE(sqe->accept_flags);
4915 accept->nofile = rlimit(RLIMIT_NOFILE);
4919 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4920 struct io_comp_state *cs)
4922 struct io_accept *accept = &req->accept;
4923 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4926 if (req->file->f_flags & O_NONBLOCK)
4927 req->flags |= REQ_F_NOWAIT;
4929 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4930 accept->addr_len, accept->flags,
4932 if (ret == -EAGAIN && force_nonblock)
4935 if (ret == -ERESTARTSYS)
4937 req_set_fail_links(req);
4939 __io_req_complete(req, ret, 0, cs);
4943 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4945 struct io_connect *conn = &req->connect;
4946 struct io_async_connect *io = req->async_data;
4948 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4950 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4953 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4954 conn->addr_len = READ_ONCE(sqe->addr2);
4959 return move_addr_to_kernel(conn->addr, conn->addr_len,
4963 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4964 struct io_comp_state *cs)
4966 struct io_async_connect __io, *io;
4967 unsigned file_flags;
4970 if (req->async_data) {
4971 io = req->async_data;
4973 ret = move_addr_to_kernel(req->connect.addr,
4974 req->connect.addr_len,
4981 file_flags = force_nonblock ? O_NONBLOCK : 0;
4983 ret = __sys_connect_file(req->file, &io->address,
4984 req->connect.addr_len, file_flags);
4985 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4986 if (req->async_data)
4988 if (io_alloc_async_data(req)) {
4992 io = req->async_data;
4993 memcpy(req->async_data, &__io, sizeof(__io));
4996 if (ret == -ERESTARTSYS)
5000 req_set_fail_links(req);
5001 __io_req_complete(req, ret, 0, cs);
5004 #else /* !CONFIG_NET */
5005 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5010 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
5011 struct io_comp_state *cs)
5016 static int io_send(struct io_kiocb *req, bool force_nonblock,
5017 struct io_comp_state *cs)
5022 static int io_recvmsg_prep(struct io_kiocb *req,
5023 const struct io_uring_sqe *sqe)
5028 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
5029 struct io_comp_state *cs)
5034 static int io_recv(struct io_kiocb *req, bool force_nonblock,
5035 struct io_comp_state *cs)
5040 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5045 static int io_accept(struct io_kiocb *req, bool force_nonblock,
5046 struct io_comp_state *cs)
5051 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5056 static int io_connect(struct io_kiocb *req, bool force_nonblock,
5057 struct io_comp_state *cs)
5061 #endif /* CONFIG_NET */
5063 struct io_poll_table {
5064 struct poll_table_struct pt;
5065 struct io_kiocb *req;
5069 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
5070 __poll_t mask, task_work_func_t func)
5074 /* for instances that support it check for an event match first: */
5075 if (mask && !(mask & poll->events))
5078 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
5080 list_del_init(&poll->wait.entry);
5083 init_task_work(&req->task_work, func);
5084 percpu_ref_get(&req->ctx->refs);
5087 * If this fails, then the task is exiting. When a task exits, the
5088 * work gets canceled, so just cancel this request as well instead
5089 * of executing it. We can't safely execute it anyway, as we may not
5090 * have the needed state needed for it anyway.
5092 ret = io_req_task_work_add(req);
5093 if (unlikely(ret)) {
5094 struct task_struct *tsk;
5096 WRITE_ONCE(poll->canceled, true);
5097 tsk = io_wq_get_task(req->ctx->io_wq);
5098 task_work_add(tsk, &req->task_work, TWA_NONE);
5099 wake_up_process(tsk);
5104 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
5105 __acquires(&req->ctx->completion_lock)
5107 struct io_ring_ctx *ctx = req->ctx;
5109 if (!req->result && !READ_ONCE(poll->canceled)) {
5110 struct poll_table_struct pt = { ._key = poll->events };
5112 req->result = vfs_poll(req->file, &pt) & poll->events;
5115 spin_lock_irq(&ctx->completion_lock);
5116 if (!req->result && !READ_ONCE(poll->canceled)) {
5117 add_wait_queue(poll->head, &poll->wait);
5124 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
5126 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
5127 if (req->opcode == IORING_OP_POLL_ADD)
5128 return req->async_data;
5129 return req->apoll->double_poll;
5132 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
5134 if (req->opcode == IORING_OP_POLL_ADD)
5136 return &req->apoll->poll;
5139 static void io_poll_remove_double(struct io_kiocb *req)
5141 struct io_poll_iocb *poll = io_poll_get_double(req);
5143 lockdep_assert_held(&req->ctx->completion_lock);
5145 if (poll && poll->head) {
5146 struct wait_queue_head *head = poll->head;
5148 spin_lock(&head->lock);
5149 list_del_init(&poll->wait.entry);
5150 if (poll->wait.private)
5151 refcount_dec(&req->refs);
5153 spin_unlock(&head->lock);
5157 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
5159 struct io_ring_ctx *ctx = req->ctx;
5161 io_poll_remove_double(req);
5162 req->poll.done = true;
5163 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
5164 io_commit_cqring(ctx);
5167 static void io_poll_task_func(struct callback_head *cb)
5169 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
5170 struct io_ring_ctx *ctx = req->ctx;
5171 struct io_kiocb *nxt;
5173 if (io_poll_rewait(req, &req->poll)) {
5174 spin_unlock_irq(&ctx->completion_lock);
5176 hash_del(&req->hash_node);
5177 io_poll_complete(req, req->result, 0);
5178 spin_unlock_irq(&ctx->completion_lock);
5180 nxt = io_put_req_find_next(req);
5181 io_cqring_ev_posted(ctx);
5183 __io_req_task_submit(nxt);
5186 percpu_ref_put(&ctx->refs);
5189 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
5190 int sync, void *key)
5192 struct io_kiocb *req = wait->private;
5193 struct io_poll_iocb *poll = io_poll_get_single(req);
5194 __poll_t mask = key_to_poll(key);
5196 /* for instances that support it check for an event match first: */
5197 if (mask && !(mask & poll->events))
5200 list_del_init(&wait->entry);
5202 if (poll && poll->head) {
5205 spin_lock(&poll->head->lock);
5206 done = list_empty(&poll->wait.entry);
5208 list_del_init(&poll->wait.entry);
5209 /* make sure double remove sees this as being gone */
5210 wait->private = NULL;
5211 spin_unlock(&poll->head->lock);
5213 /* use wait func handler, so it matches the rq type */
5214 poll->wait.func(&poll->wait, mode, sync, key);
5217 refcount_dec(&req->refs);
5221 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
5222 wait_queue_func_t wake_func)
5226 poll->canceled = false;
5227 poll->events = events;
5228 INIT_LIST_HEAD(&poll->wait.entry);
5229 init_waitqueue_func_entry(&poll->wait, wake_func);
5232 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5233 struct wait_queue_head *head,
5234 struct io_poll_iocb **poll_ptr)
5236 struct io_kiocb *req = pt->req;
5239 * If poll->head is already set, it's because the file being polled
5240 * uses multiple waitqueues for poll handling (eg one for read, one
5241 * for write). Setup a separate io_poll_iocb if this happens.
5243 if (unlikely(poll->head)) {
5244 struct io_poll_iocb *poll_one = poll;
5246 /* already have a 2nd entry, fail a third attempt */
5248 pt->error = -EINVAL;
5251 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5253 pt->error = -ENOMEM;
5256 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5257 refcount_inc(&req->refs);
5258 poll->wait.private = req;
5265 if (poll->events & EPOLLEXCLUSIVE)
5266 add_wait_queue_exclusive(head, &poll->wait);
5268 add_wait_queue(head, &poll->wait);
5271 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5272 struct poll_table_struct *p)
5274 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5275 struct async_poll *apoll = pt->req->apoll;
5277 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5280 static void io_async_task_func(struct callback_head *cb)
5282 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
5283 struct async_poll *apoll = req->apoll;
5284 struct io_ring_ctx *ctx = req->ctx;
5286 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
5288 if (io_poll_rewait(req, &apoll->poll)) {
5289 spin_unlock_irq(&ctx->completion_lock);
5290 percpu_ref_put(&ctx->refs);
5294 /* If req is still hashed, it cannot have been canceled. Don't check. */
5295 if (hash_hashed(&req->hash_node))
5296 hash_del(&req->hash_node);
5298 io_poll_remove_double(req);
5299 spin_unlock_irq(&ctx->completion_lock);
5301 if (!READ_ONCE(apoll->poll.canceled))
5302 __io_req_task_submit(req);
5304 __io_req_task_cancel(req, -ECANCELED);
5306 percpu_ref_put(&ctx->refs);
5307 kfree(apoll->double_poll);
5311 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5314 struct io_kiocb *req = wait->private;
5315 struct io_poll_iocb *poll = &req->apoll->poll;
5317 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5320 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5323 static void io_poll_req_insert(struct io_kiocb *req)
5325 struct io_ring_ctx *ctx = req->ctx;
5326 struct hlist_head *list;
5328 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5329 hlist_add_head(&req->hash_node, list);
5332 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5333 struct io_poll_iocb *poll,
5334 struct io_poll_table *ipt, __poll_t mask,
5335 wait_queue_func_t wake_func)
5336 __acquires(&ctx->completion_lock)
5338 struct io_ring_ctx *ctx = req->ctx;
5339 bool cancel = false;
5341 INIT_HLIST_NODE(&req->hash_node);
5342 io_init_poll_iocb(poll, mask, wake_func);
5343 poll->file = req->file;
5344 poll->wait.private = req;
5346 ipt->pt._key = mask;
5348 ipt->error = -EINVAL;
5350 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5352 spin_lock_irq(&ctx->completion_lock);
5353 if (likely(poll->head)) {
5354 spin_lock(&poll->head->lock);
5355 if (unlikely(list_empty(&poll->wait.entry))) {
5361 if (mask || ipt->error)
5362 list_del_init(&poll->wait.entry);
5364 WRITE_ONCE(poll->canceled, true);
5365 else if (!poll->done) /* actually waiting for an event */
5366 io_poll_req_insert(req);
5367 spin_unlock(&poll->head->lock);
5373 static bool io_arm_poll_handler(struct io_kiocb *req)
5375 const struct io_op_def *def = &io_op_defs[req->opcode];
5376 struct io_ring_ctx *ctx = req->ctx;
5377 struct async_poll *apoll;
5378 struct io_poll_table ipt;
5382 if (!req->file || !file_can_poll(req->file))
5384 if (req->flags & REQ_F_POLLED)
5388 else if (def->pollout)
5392 /* if we can't nonblock try, then no point in arming a poll handler */
5393 if (!io_file_supports_async(req->file, rw))
5396 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5397 if (unlikely(!apoll))
5399 apoll->double_poll = NULL;
5401 req->flags |= REQ_F_POLLED;
5406 mask |= POLLIN | POLLRDNORM;
5408 mask |= POLLOUT | POLLWRNORM;
5410 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5411 if ((req->opcode == IORING_OP_RECVMSG) &&
5412 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5415 mask |= POLLERR | POLLPRI;
5417 ipt.pt._qproc = io_async_queue_proc;
5419 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5421 if (ret || ipt.error) {
5422 io_poll_remove_double(req);
5423 spin_unlock_irq(&ctx->completion_lock);
5424 kfree(apoll->double_poll);
5428 spin_unlock_irq(&ctx->completion_lock);
5429 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
5430 apoll->poll.events);
5434 static bool __io_poll_remove_one(struct io_kiocb *req,
5435 struct io_poll_iocb *poll)
5437 bool do_complete = false;
5439 spin_lock(&poll->head->lock);
5440 WRITE_ONCE(poll->canceled, true);
5441 if (!list_empty(&poll->wait.entry)) {
5442 list_del_init(&poll->wait.entry);
5445 spin_unlock(&poll->head->lock);
5446 hash_del(&req->hash_node);
5450 static bool io_poll_remove_one(struct io_kiocb *req)
5454 io_poll_remove_double(req);
5456 if (req->opcode == IORING_OP_POLL_ADD) {
5457 do_complete = __io_poll_remove_one(req, &req->poll);
5459 struct async_poll *apoll = req->apoll;
5461 /* non-poll requests have submit ref still */
5462 do_complete = __io_poll_remove_one(req, &apoll->poll);
5465 kfree(apoll->double_poll);
5471 io_cqring_fill_event(req, -ECANCELED);
5472 io_commit_cqring(req->ctx);
5473 req_set_fail_links(req);
5474 io_put_req_deferred(req, 1);
5481 * Returns true if we found and killed one or more poll requests
5483 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5484 struct files_struct *files)
5486 struct hlist_node *tmp;
5487 struct io_kiocb *req;
5490 spin_lock_irq(&ctx->completion_lock);
5491 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5492 struct hlist_head *list;
5494 list = &ctx->cancel_hash[i];
5495 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5496 if (io_match_task(req, tsk, files))
5497 posted += io_poll_remove_one(req);
5500 spin_unlock_irq(&ctx->completion_lock);
5503 io_cqring_ev_posted(ctx);
5508 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
5510 struct hlist_head *list;
5511 struct io_kiocb *req;
5513 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5514 hlist_for_each_entry(req, list, hash_node) {
5515 if (sqe_addr != req->user_data)
5517 if (io_poll_remove_one(req))
5525 static int io_poll_remove_prep(struct io_kiocb *req,
5526 const struct io_uring_sqe *sqe)
5528 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5530 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
5534 req->poll_remove.addr = READ_ONCE(sqe->addr);
5539 * Find a running poll command that matches one specified in sqe->addr,
5540 * and remove it if found.
5542 static int io_poll_remove(struct io_kiocb *req)
5544 struct io_ring_ctx *ctx = req->ctx;
5547 spin_lock_irq(&ctx->completion_lock);
5548 ret = io_poll_cancel(ctx, req->poll_remove.addr);
5549 spin_unlock_irq(&ctx->completion_lock);
5552 req_set_fail_links(req);
5553 io_req_complete(req, ret);
5557 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5560 struct io_kiocb *req = wait->private;
5561 struct io_poll_iocb *poll = &req->poll;
5563 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5566 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5567 struct poll_table_struct *p)
5569 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5571 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5574 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5576 struct io_poll_iocb *poll = &req->poll;
5579 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5581 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
5584 events = READ_ONCE(sqe->poll32_events);
5586 events = swahw32(events);
5588 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP |
5589 (events & EPOLLEXCLUSIVE);
5593 static int io_poll_add(struct io_kiocb *req)
5595 struct io_poll_iocb *poll = &req->poll;
5596 struct io_ring_ctx *ctx = req->ctx;
5597 struct io_poll_table ipt;
5600 ipt.pt._qproc = io_poll_queue_proc;
5602 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5605 if (mask) { /* no async, we'd stolen it */
5607 io_poll_complete(req, mask, 0);
5609 spin_unlock_irq(&ctx->completion_lock);
5612 io_cqring_ev_posted(ctx);
5618 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5620 struct io_timeout_data *data = container_of(timer,
5621 struct io_timeout_data, timer);
5622 struct io_kiocb *req = data->req;
5623 struct io_ring_ctx *ctx = req->ctx;
5624 unsigned long flags;
5626 spin_lock_irqsave(&ctx->completion_lock, flags);
5627 list_del_init(&req->timeout.list);
5628 atomic_set(&req->ctx->cq_timeouts,
5629 atomic_read(&req->ctx->cq_timeouts) + 1);
5631 io_cqring_fill_event(req, -ETIME);
5632 io_commit_cqring(ctx);
5633 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5635 io_cqring_ev_posted(ctx);
5636 req_set_fail_links(req);
5638 return HRTIMER_NORESTART;
5641 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5644 struct io_timeout_data *io;
5645 struct io_kiocb *req;
5648 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5649 if (user_data == req->user_data) {
5656 return ERR_PTR(ret);
5658 io = req->async_data;
5659 ret = hrtimer_try_to_cancel(&io->timer);
5661 return ERR_PTR(-EALREADY);
5662 list_del_init(&req->timeout.list);
5666 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5668 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5671 return PTR_ERR(req);
5673 req_set_fail_links(req);
5674 io_cqring_fill_event(req, -ECANCELED);
5675 io_put_req_deferred(req, 1);
5679 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5680 struct timespec64 *ts, enum hrtimer_mode mode)
5682 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5683 struct io_timeout_data *data;
5686 return PTR_ERR(req);
5688 req->timeout.off = 0; /* noseq */
5689 data = req->async_data;
5690 list_add_tail(&req->timeout.list, &ctx->timeout_list);
5691 hrtimer_init(&data->timer, CLOCK_MONOTONIC, mode);
5692 data->timer.function = io_timeout_fn;
5693 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
5697 static int io_timeout_remove_prep(struct io_kiocb *req,
5698 const struct io_uring_sqe *sqe)
5700 struct io_timeout_rem *tr = &req->timeout_rem;
5702 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5704 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5706 if (sqe->ioprio || sqe->buf_index || sqe->len)
5709 tr->addr = READ_ONCE(sqe->addr);
5710 tr->flags = READ_ONCE(sqe->timeout_flags);
5711 if (tr->flags & IORING_TIMEOUT_UPDATE) {
5712 if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS))
5714 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
5716 } else if (tr->flags) {
5717 /* timeout removal doesn't support flags */
5725 * Remove or update an existing timeout command
5727 static int io_timeout_remove(struct io_kiocb *req)
5729 struct io_timeout_rem *tr = &req->timeout_rem;
5730 struct io_ring_ctx *ctx = req->ctx;
5733 spin_lock_irq(&ctx->completion_lock);
5734 if (req->timeout_rem.flags & IORING_TIMEOUT_UPDATE) {
5735 enum hrtimer_mode mode = (tr->flags & IORING_TIMEOUT_ABS)
5736 ? HRTIMER_MODE_ABS : HRTIMER_MODE_REL;
5738 ret = io_timeout_update(ctx, tr->addr, &tr->ts, mode);
5740 ret = io_timeout_cancel(ctx, tr->addr);
5743 io_cqring_fill_event(req, ret);
5744 io_commit_cqring(ctx);
5745 spin_unlock_irq(&ctx->completion_lock);
5746 io_cqring_ev_posted(ctx);
5748 req_set_fail_links(req);
5753 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5754 bool is_timeout_link)
5756 struct io_timeout_data *data;
5758 u32 off = READ_ONCE(sqe->off);
5760 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5762 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5764 if (off && is_timeout_link)
5766 flags = READ_ONCE(sqe->timeout_flags);
5767 if (flags & ~IORING_TIMEOUT_ABS)
5770 req->timeout.off = off;
5772 if (!req->async_data && io_alloc_async_data(req))
5775 data = req->async_data;
5778 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5781 if (flags & IORING_TIMEOUT_ABS)
5782 data->mode = HRTIMER_MODE_ABS;
5784 data->mode = HRTIMER_MODE_REL;
5786 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5790 static int io_timeout(struct io_kiocb *req)
5792 struct io_ring_ctx *ctx = req->ctx;
5793 struct io_timeout_data *data = req->async_data;
5794 struct list_head *entry;
5795 u32 tail, off = req->timeout.off;
5797 spin_lock_irq(&ctx->completion_lock);
5800 * sqe->off holds how many events that need to occur for this
5801 * timeout event to be satisfied. If it isn't set, then this is
5802 * a pure timeout request, sequence isn't used.
5804 if (io_is_timeout_noseq(req)) {
5805 entry = ctx->timeout_list.prev;
5809 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5810 req->timeout.target_seq = tail + off;
5813 * Insertion sort, ensuring the first entry in the list is always
5814 * the one we need first.
5816 list_for_each_prev(entry, &ctx->timeout_list) {
5817 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5820 if (io_is_timeout_noseq(nxt))
5822 /* nxt.seq is behind @tail, otherwise would've been completed */
5823 if (off >= nxt->timeout.target_seq - tail)
5827 list_add(&req->timeout.list, entry);
5828 data->timer.function = io_timeout_fn;
5829 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5830 spin_unlock_irq(&ctx->completion_lock);
5834 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5836 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5838 return req->user_data == (unsigned long) data;
5841 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
5843 enum io_wq_cancel cancel_ret;
5846 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr, false);
5847 switch (cancel_ret) {
5848 case IO_WQ_CANCEL_OK:
5851 case IO_WQ_CANCEL_RUNNING:
5854 case IO_WQ_CANCEL_NOTFOUND:
5862 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5863 struct io_kiocb *req, __u64 sqe_addr,
5866 unsigned long flags;
5869 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
5870 if (ret != -ENOENT) {
5871 spin_lock_irqsave(&ctx->completion_lock, flags);
5875 spin_lock_irqsave(&ctx->completion_lock, flags);
5876 ret = io_timeout_cancel(ctx, sqe_addr);
5879 ret = io_poll_cancel(ctx, sqe_addr);
5883 io_cqring_fill_event(req, ret);
5884 io_commit_cqring(ctx);
5885 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5886 io_cqring_ev_posted(ctx);
5889 req_set_fail_links(req);
5893 static int io_async_cancel_prep(struct io_kiocb *req,
5894 const struct io_uring_sqe *sqe)
5896 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5898 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5900 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5903 req->cancel.addr = READ_ONCE(sqe->addr);
5907 static int io_async_cancel(struct io_kiocb *req)
5909 struct io_ring_ctx *ctx = req->ctx;
5911 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
5915 static int io_files_update_prep(struct io_kiocb *req,
5916 const struct io_uring_sqe *sqe)
5918 if (unlikely(req->ctx->flags & IORING_SETUP_SQPOLL))
5920 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5922 if (sqe->ioprio || sqe->rw_flags)
5925 req->files_update.offset = READ_ONCE(sqe->off);
5926 req->files_update.nr_args = READ_ONCE(sqe->len);
5927 if (!req->files_update.nr_args)
5929 req->files_update.arg = READ_ONCE(sqe->addr);
5933 static int io_files_update(struct io_kiocb *req, bool force_nonblock,
5934 struct io_comp_state *cs)
5936 struct io_ring_ctx *ctx = req->ctx;
5937 struct io_uring_files_update up;
5943 up.offset = req->files_update.offset;
5944 up.fds = req->files_update.arg;
5946 mutex_lock(&ctx->uring_lock);
5947 ret = __io_sqe_files_update(ctx, &up, req->files_update.nr_args);
5948 mutex_unlock(&ctx->uring_lock);
5951 req_set_fail_links(req);
5952 __io_req_complete(req, ret, 0, cs);
5956 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5958 switch (req->opcode) {
5961 case IORING_OP_READV:
5962 case IORING_OP_READ_FIXED:
5963 case IORING_OP_READ:
5964 return io_read_prep(req, sqe);
5965 case IORING_OP_WRITEV:
5966 case IORING_OP_WRITE_FIXED:
5967 case IORING_OP_WRITE:
5968 return io_write_prep(req, sqe);
5969 case IORING_OP_POLL_ADD:
5970 return io_poll_add_prep(req, sqe);
5971 case IORING_OP_POLL_REMOVE:
5972 return io_poll_remove_prep(req, sqe);
5973 case IORING_OP_FSYNC:
5974 return io_prep_fsync(req, sqe);
5975 case IORING_OP_SYNC_FILE_RANGE:
5976 return io_prep_sfr(req, sqe);
5977 case IORING_OP_SENDMSG:
5978 case IORING_OP_SEND:
5979 return io_sendmsg_prep(req, sqe);
5980 case IORING_OP_RECVMSG:
5981 case IORING_OP_RECV:
5982 return io_recvmsg_prep(req, sqe);
5983 case IORING_OP_CONNECT:
5984 return io_connect_prep(req, sqe);
5985 case IORING_OP_TIMEOUT:
5986 return io_timeout_prep(req, sqe, false);
5987 case IORING_OP_TIMEOUT_REMOVE:
5988 return io_timeout_remove_prep(req, sqe);
5989 case IORING_OP_ASYNC_CANCEL:
5990 return io_async_cancel_prep(req, sqe);
5991 case IORING_OP_LINK_TIMEOUT:
5992 return io_timeout_prep(req, sqe, true);
5993 case IORING_OP_ACCEPT:
5994 return io_accept_prep(req, sqe);
5995 case IORING_OP_FALLOCATE:
5996 return io_fallocate_prep(req, sqe);
5997 case IORING_OP_OPENAT:
5998 return io_openat_prep(req, sqe);
5999 case IORING_OP_CLOSE:
6000 return io_close_prep(req, sqe);
6001 case IORING_OP_FILES_UPDATE:
6002 return io_files_update_prep(req, sqe);
6003 case IORING_OP_STATX:
6004 return io_statx_prep(req, sqe);
6005 case IORING_OP_FADVISE:
6006 return io_fadvise_prep(req, sqe);
6007 case IORING_OP_MADVISE:
6008 return io_madvise_prep(req, sqe);
6009 case IORING_OP_OPENAT2:
6010 return io_openat2_prep(req, sqe);
6011 case IORING_OP_EPOLL_CTL:
6012 return io_epoll_ctl_prep(req, sqe);
6013 case IORING_OP_SPLICE:
6014 return io_splice_prep(req, sqe);
6015 case IORING_OP_PROVIDE_BUFFERS:
6016 return io_provide_buffers_prep(req, sqe);
6017 case IORING_OP_REMOVE_BUFFERS:
6018 return io_remove_buffers_prep(req, sqe);
6020 return io_tee_prep(req, sqe);
6021 case IORING_OP_SHUTDOWN:
6022 return io_shutdown_prep(req, sqe);
6023 case IORING_OP_RENAMEAT:
6024 return io_renameat_prep(req, sqe);
6025 case IORING_OP_UNLINKAT:
6026 return io_unlinkat_prep(req, sqe);
6029 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
6034 static int io_req_defer_prep(struct io_kiocb *req,
6035 const struct io_uring_sqe *sqe)
6039 if (io_alloc_async_data(req))
6041 return io_req_prep(req, sqe);
6044 static u32 io_get_sequence(struct io_kiocb *req)
6046 struct io_kiocb *pos;
6047 struct io_ring_ctx *ctx = req->ctx;
6048 u32 total_submitted, nr_reqs = 0;
6050 io_for_each_link(pos, req)
6053 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
6054 return total_submitted - nr_reqs;
6057 static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe)
6059 struct io_ring_ctx *ctx = req->ctx;
6060 struct io_defer_entry *de;
6064 /* Still need defer if there is pending req in defer list. */
6065 if (likely(list_empty_careful(&ctx->defer_list) &&
6066 !(req->flags & REQ_F_IO_DRAIN)))
6069 seq = io_get_sequence(req);
6070 /* Still a chance to pass the sequence check */
6071 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
6074 if (!req->async_data) {
6075 ret = io_req_defer_prep(req, sqe);
6079 io_prep_async_link(req);
6080 de = kmalloc(sizeof(*de), GFP_KERNEL);
6084 spin_lock_irq(&ctx->completion_lock);
6085 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
6086 spin_unlock_irq(&ctx->completion_lock);
6088 io_queue_async_work(req);
6089 return -EIOCBQUEUED;
6092 trace_io_uring_defer(ctx, req, req->user_data);
6095 list_add_tail(&de->list, &ctx->defer_list);
6096 spin_unlock_irq(&ctx->completion_lock);
6097 return -EIOCBQUEUED;
6100 static void io_req_drop_files(struct io_kiocb *req)
6102 struct io_ring_ctx *ctx = req->ctx;
6103 struct io_uring_task *tctx = req->task->io_uring;
6104 unsigned long flags;
6106 spin_lock_irqsave(&ctx->inflight_lock, flags);
6107 list_del(&req->inflight_entry);
6108 if (atomic_read(&tctx->in_idle))
6109 wake_up(&tctx->wait);
6110 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
6111 req->flags &= ~REQ_F_INFLIGHT;
6112 put_files_struct(req->work.identity->files);
6113 put_nsproxy(req->work.identity->nsproxy);
6114 req->work.flags &= ~IO_WQ_WORK_FILES;
6117 static void __io_clean_op(struct io_kiocb *req)
6119 if (req->flags & REQ_F_BUFFER_SELECTED) {
6120 switch (req->opcode) {
6121 case IORING_OP_READV:
6122 case IORING_OP_READ_FIXED:
6123 case IORING_OP_READ:
6124 kfree((void *)(unsigned long)req->rw.addr);
6126 case IORING_OP_RECVMSG:
6127 case IORING_OP_RECV:
6128 kfree(req->sr_msg.kbuf);
6131 req->flags &= ~REQ_F_BUFFER_SELECTED;
6134 if (req->flags & REQ_F_NEED_CLEANUP) {
6135 switch (req->opcode) {
6136 case IORING_OP_READV:
6137 case IORING_OP_READ_FIXED:
6138 case IORING_OP_READ:
6139 case IORING_OP_WRITEV:
6140 case IORING_OP_WRITE_FIXED:
6141 case IORING_OP_WRITE: {
6142 struct io_async_rw *io = req->async_data;
6144 kfree(io->free_iovec);
6147 case IORING_OP_RECVMSG:
6148 case IORING_OP_SENDMSG: {
6149 struct io_async_msghdr *io = req->async_data;
6150 if (io->iov != io->fast_iov)
6154 case IORING_OP_SPLICE:
6156 io_put_file(req, req->splice.file_in,
6157 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
6159 case IORING_OP_OPENAT:
6160 case IORING_OP_OPENAT2:
6161 if (req->open.filename)
6162 putname(req->open.filename);
6164 case IORING_OP_RENAMEAT:
6165 putname(req->rename.oldpath);
6166 putname(req->rename.newpath);
6168 case IORING_OP_UNLINKAT:
6169 putname(req->unlink.filename);
6172 req->flags &= ~REQ_F_NEED_CLEANUP;
6175 if (req->flags & REQ_F_INFLIGHT)
6176 io_req_drop_files(req);
6179 static int io_issue_sqe(struct io_kiocb *req, bool force_nonblock,
6180 struct io_comp_state *cs)
6182 struct io_ring_ctx *ctx = req->ctx;
6185 switch (req->opcode) {
6187 ret = io_nop(req, cs);
6189 case IORING_OP_READV:
6190 case IORING_OP_READ_FIXED:
6191 case IORING_OP_READ:
6192 ret = io_read(req, force_nonblock, cs);
6194 case IORING_OP_WRITEV:
6195 case IORING_OP_WRITE_FIXED:
6196 case IORING_OP_WRITE:
6197 ret = io_write(req, force_nonblock, cs);
6199 case IORING_OP_FSYNC:
6200 ret = io_fsync(req, force_nonblock);
6202 case IORING_OP_POLL_ADD:
6203 ret = io_poll_add(req);
6205 case IORING_OP_POLL_REMOVE:
6206 ret = io_poll_remove(req);
6208 case IORING_OP_SYNC_FILE_RANGE:
6209 ret = io_sync_file_range(req, force_nonblock);
6211 case IORING_OP_SENDMSG:
6212 ret = io_sendmsg(req, force_nonblock, cs);
6214 case IORING_OP_SEND:
6215 ret = io_send(req, force_nonblock, cs);
6217 case IORING_OP_RECVMSG:
6218 ret = io_recvmsg(req, force_nonblock, cs);
6220 case IORING_OP_RECV:
6221 ret = io_recv(req, force_nonblock, cs);
6223 case IORING_OP_TIMEOUT:
6224 ret = io_timeout(req);
6226 case IORING_OP_TIMEOUT_REMOVE:
6227 ret = io_timeout_remove(req);
6229 case IORING_OP_ACCEPT:
6230 ret = io_accept(req, force_nonblock, cs);
6232 case IORING_OP_CONNECT:
6233 ret = io_connect(req, force_nonblock, cs);
6235 case IORING_OP_ASYNC_CANCEL:
6236 ret = io_async_cancel(req);
6238 case IORING_OP_FALLOCATE:
6239 ret = io_fallocate(req, force_nonblock);
6241 case IORING_OP_OPENAT:
6242 ret = io_openat(req, force_nonblock);
6244 case IORING_OP_CLOSE:
6245 ret = io_close(req, force_nonblock, cs);
6247 case IORING_OP_FILES_UPDATE:
6248 ret = io_files_update(req, force_nonblock, cs);
6250 case IORING_OP_STATX:
6251 ret = io_statx(req, force_nonblock);
6253 case IORING_OP_FADVISE:
6254 ret = io_fadvise(req, force_nonblock);
6256 case IORING_OP_MADVISE:
6257 ret = io_madvise(req, force_nonblock);
6259 case IORING_OP_OPENAT2:
6260 ret = io_openat2(req, force_nonblock);
6262 case IORING_OP_EPOLL_CTL:
6263 ret = io_epoll_ctl(req, force_nonblock, cs);
6265 case IORING_OP_SPLICE:
6266 ret = io_splice(req, force_nonblock);
6268 case IORING_OP_PROVIDE_BUFFERS:
6269 ret = io_provide_buffers(req, force_nonblock, cs);
6271 case IORING_OP_REMOVE_BUFFERS:
6272 ret = io_remove_buffers(req, force_nonblock, cs);
6275 ret = io_tee(req, force_nonblock);
6277 case IORING_OP_SHUTDOWN:
6278 ret = io_shutdown(req, force_nonblock);
6280 case IORING_OP_RENAMEAT:
6281 ret = io_renameat(req, force_nonblock);
6283 case IORING_OP_UNLINKAT:
6284 ret = io_unlinkat(req, force_nonblock);
6294 /* If the op doesn't have a file, we're not polling for it */
6295 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
6296 const bool in_async = io_wq_current_is_worker();
6298 /* workqueue context doesn't hold uring_lock, grab it now */
6300 mutex_lock(&ctx->uring_lock);
6302 io_iopoll_req_issued(req, in_async);
6305 mutex_unlock(&ctx->uring_lock);
6311 static struct io_wq_work *io_wq_submit_work(struct io_wq_work *work)
6313 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6314 struct io_kiocb *timeout;
6317 timeout = io_prep_linked_timeout(req);
6319 io_queue_linked_timeout(timeout);
6321 /* if NO_CANCEL is set, we must still run the work */
6322 if ((work->flags & (IO_WQ_WORK_CANCEL|IO_WQ_WORK_NO_CANCEL)) ==
6323 IO_WQ_WORK_CANCEL) {
6329 ret = io_issue_sqe(req, false, NULL);
6331 * We can get EAGAIN for polled IO even though we're
6332 * forcing a sync submission from here, since we can't
6333 * wait for request slots on the block side.
6343 * io_iopoll_complete() does not hold completion_lock to complete
6344 * polled io, so here for polled io, just mark it done and still let
6345 * io_iopoll_complete() complete it.
6347 if (req->ctx->flags & IORING_SETUP_IOPOLL) {
6348 struct kiocb *kiocb = &req->rw.kiocb;
6350 kiocb_done(kiocb, ret, NULL);
6352 req_set_fail_links(req);
6353 io_req_complete(req, ret);
6357 return io_steal_work(req);
6360 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6363 struct fixed_file_table *table;
6365 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
6366 return table->files[index & IORING_FILE_TABLE_MASK];
6369 static struct file *io_file_get(struct io_submit_state *state,
6370 struct io_kiocb *req, int fd, bool fixed)
6372 struct io_ring_ctx *ctx = req->ctx;
6376 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6378 fd = array_index_nospec(fd, ctx->nr_user_files);
6379 file = io_file_from_index(ctx, fd);
6380 io_set_resource_node(req);
6382 trace_io_uring_file_get(ctx, fd);
6383 file = __io_file_get(state, fd);
6389 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6391 struct io_timeout_data *data = container_of(timer,
6392 struct io_timeout_data, timer);
6393 struct io_kiocb *prev, *req = data->req;
6394 struct io_ring_ctx *ctx = req->ctx;
6395 unsigned long flags;
6397 spin_lock_irqsave(&ctx->completion_lock, flags);
6398 prev = req->timeout.head;
6399 req->timeout.head = NULL;
6402 * We don't expect the list to be empty, that will only happen if we
6403 * race with the completion of the linked work.
6405 if (prev && refcount_inc_not_zero(&prev->refs))
6406 io_remove_next_linked(prev);
6409 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6412 req_set_fail_links(prev);
6413 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6416 io_req_complete(req, -ETIME);
6418 return HRTIMER_NORESTART;
6421 static void __io_queue_linked_timeout(struct io_kiocb *req)
6424 * If the back reference is NULL, then our linked request finished
6425 * before we got a chance to setup the timer
6427 if (req->timeout.head) {
6428 struct io_timeout_data *data = req->async_data;
6430 data->timer.function = io_link_timeout_fn;
6431 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6436 static void io_queue_linked_timeout(struct io_kiocb *req)
6438 struct io_ring_ctx *ctx = req->ctx;
6440 spin_lock_irq(&ctx->completion_lock);
6441 __io_queue_linked_timeout(req);
6442 spin_unlock_irq(&ctx->completion_lock);
6444 /* drop submission reference */
6448 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6450 struct io_kiocb *nxt = req->link;
6452 if (!nxt || (req->flags & REQ_F_LINK_TIMEOUT) ||
6453 nxt->opcode != IORING_OP_LINK_TIMEOUT)
6456 nxt->timeout.head = req;
6457 nxt->flags |= REQ_F_LTIMEOUT_ACTIVE;
6458 req->flags |= REQ_F_LINK_TIMEOUT;
6462 static void __io_queue_sqe(struct io_kiocb *req, struct io_comp_state *cs)
6464 struct io_kiocb *linked_timeout;
6465 const struct cred *old_creds = NULL;
6469 linked_timeout = io_prep_linked_timeout(req);
6471 if ((req->flags & REQ_F_WORK_INITIALIZED) &&
6472 (req->work.flags & IO_WQ_WORK_CREDS) &&
6473 req->work.identity->creds != current_cred()) {
6475 revert_creds(old_creds);
6476 if (old_creds == req->work.identity->creds)
6477 old_creds = NULL; /* restored original creds */
6479 old_creds = override_creds(req->work.identity->creds);
6482 ret = io_issue_sqe(req, true, cs);
6485 * We async punt it if the file wasn't marked NOWAIT, or if the file
6486 * doesn't support non-blocking read/write attempts
6488 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6489 if (!io_arm_poll_handler(req)) {
6491 * Queued up for async execution, worker will release
6492 * submit reference when the iocb is actually submitted.
6494 io_queue_async_work(req);
6498 io_queue_linked_timeout(linked_timeout);
6499 } else if (likely(!ret)) {
6500 /* drop submission reference */
6501 req = io_put_req_find_next(req);
6503 io_queue_linked_timeout(linked_timeout);
6506 if (!(req->flags & REQ_F_FORCE_ASYNC))
6508 io_queue_async_work(req);
6511 /* un-prep timeout, so it'll be killed as any other linked */
6512 req->flags &= ~REQ_F_LINK_TIMEOUT;
6513 req_set_fail_links(req);
6515 io_req_complete(req, ret);
6519 revert_creds(old_creds);
6522 static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6523 struct io_comp_state *cs)
6527 ret = io_req_defer(req, sqe);
6529 if (ret != -EIOCBQUEUED) {
6531 req_set_fail_links(req);
6533 io_req_complete(req, ret);
6535 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6536 if (!req->async_data) {
6537 ret = io_req_defer_prep(req, sqe);
6541 io_queue_async_work(req);
6544 ret = io_req_prep(req, sqe);
6548 __io_queue_sqe(req, cs);
6552 static inline void io_queue_link_head(struct io_kiocb *req,
6553 struct io_comp_state *cs)
6555 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
6557 io_req_complete(req, -ECANCELED);
6559 io_queue_sqe(req, NULL, cs);
6562 struct io_submit_link {
6563 struct io_kiocb *head;
6564 struct io_kiocb *last;
6567 static int io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6568 struct io_submit_link *link, struct io_comp_state *cs)
6570 struct io_ring_ctx *ctx = req->ctx;
6574 * If we already have a head request, queue this one for async
6575 * submittal once the head completes. If we don't have a head but
6576 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6577 * submitted sync once the chain is complete. If none of those
6578 * conditions are true (normal request), then just queue it.
6581 struct io_kiocb *head = link->head;
6584 * Taking sequential execution of a link, draining both sides
6585 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6586 * requests in the link. So, it drains the head and the
6587 * next after the link request. The last one is done via
6588 * drain_next flag to persist the effect across calls.
6590 if (req->flags & REQ_F_IO_DRAIN) {
6591 head->flags |= REQ_F_IO_DRAIN;
6592 ctx->drain_next = 1;
6594 ret = io_req_defer_prep(req, sqe);
6595 if (unlikely(ret)) {
6596 /* fail even hard links since we don't submit */
6597 head->flags |= REQ_F_FAIL_LINK;
6600 trace_io_uring_link(ctx, req, head);
6601 link->last->link = req;
6604 /* last request of a link, enqueue the link */
6605 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6606 io_queue_link_head(head, cs);
6610 if (unlikely(ctx->drain_next)) {
6611 req->flags |= REQ_F_IO_DRAIN;
6612 ctx->drain_next = 0;
6614 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6615 ret = io_req_defer_prep(req, sqe);
6617 req->flags |= REQ_F_FAIL_LINK;
6621 io_queue_sqe(req, sqe, cs);
6629 * Batched submission is done, ensure local IO is flushed out.
6631 static void io_submit_state_end(struct io_submit_state *state)
6633 if (!list_empty(&state->comp.list))
6634 io_submit_flush_completions(&state->comp);
6635 if (state->plug_started)
6636 blk_finish_plug(&state->plug);
6637 io_state_file_put(state);
6638 if (state->free_reqs)
6639 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
6643 * Start submission side cache.
6645 static void io_submit_state_start(struct io_submit_state *state,
6646 struct io_ring_ctx *ctx, unsigned int max_ios)
6648 state->plug_started = false;
6650 INIT_LIST_HEAD(&state->comp.list);
6651 state->comp.ctx = ctx;
6652 state->free_reqs = 0;
6653 state->file_refs = 0;
6654 state->ios_left = max_ios;
6657 static void io_commit_sqring(struct io_ring_ctx *ctx)
6659 struct io_rings *rings = ctx->rings;
6662 * Ensure any loads from the SQEs are done at this point,
6663 * since once we write the new head, the application could
6664 * write new data to them.
6666 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6670 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6671 * that is mapped by userspace. This means that care needs to be taken to
6672 * ensure that reads are stable, as we cannot rely on userspace always
6673 * being a good citizen. If members of the sqe are validated and then later
6674 * used, it's important that those reads are done through READ_ONCE() to
6675 * prevent a re-load down the line.
6677 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6679 u32 *sq_array = ctx->sq_array;
6683 * The cached sq head (or cq tail) serves two purposes:
6685 * 1) allows us to batch the cost of updating the user visible
6687 * 2) allows the kernel side to track the head on its own, even
6688 * though the application is the one updating it.
6690 head = READ_ONCE(sq_array[ctx->cached_sq_head & ctx->sq_mask]);
6691 if (likely(head < ctx->sq_entries))
6692 return &ctx->sq_sqes[head];
6694 /* drop invalid entries */
6695 ctx->cached_sq_dropped++;
6696 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6700 static inline void io_consume_sqe(struct io_ring_ctx *ctx)
6702 ctx->cached_sq_head++;
6706 * Check SQE restrictions (opcode and flags).
6708 * Returns 'true' if SQE is allowed, 'false' otherwise.
6710 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6711 struct io_kiocb *req,
6712 unsigned int sqe_flags)
6714 if (!ctx->restricted)
6717 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6720 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6721 ctx->restrictions.sqe_flags_required)
6724 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6725 ctx->restrictions.sqe_flags_required))
6731 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
6732 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
6733 IOSQE_BUFFER_SELECT)
6735 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6736 const struct io_uring_sqe *sqe,
6737 struct io_submit_state *state)
6739 unsigned int sqe_flags;
6742 req->opcode = READ_ONCE(sqe->opcode);
6743 req->user_data = READ_ONCE(sqe->user_data);
6744 req->async_data = NULL;
6749 req->fixed_file_refs = NULL;
6750 /* one is dropped after submission, the other at completion */
6751 refcount_set(&req->refs, 2);
6752 req->task = current;
6755 if (unlikely(req->opcode >= IORING_OP_LAST))
6758 if (unlikely(io_sq_thread_acquire_mm_files(ctx, req)))
6761 sqe_flags = READ_ONCE(sqe->flags);
6762 /* enforce forwards compatibility on users */
6763 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6766 if (unlikely(!io_check_restriction(ctx, req, sqe_flags)))
6769 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6770 !io_op_defs[req->opcode].buffer_select)
6773 id = READ_ONCE(sqe->personality);
6775 struct io_identity *iod;
6777 iod = idr_find(&ctx->personality_idr, id);
6780 refcount_inc(&iod->count);
6782 __io_req_init_async(req);
6783 get_cred(iod->creds);
6784 req->work.identity = iod;
6785 req->work.flags |= IO_WQ_WORK_CREDS;
6788 /* same numerical values with corresponding REQ_F_*, safe to copy */
6789 req->flags |= sqe_flags;
6792 * Plug now if we have more than 1 IO left after this, and the target
6793 * is potentially a read/write to block based storage.
6795 if (!state->plug_started && state->ios_left > 1 &&
6796 io_op_defs[req->opcode].plug) {
6797 blk_start_plug(&state->plug);
6798 state->plug_started = true;
6802 if (io_op_defs[req->opcode].needs_file) {
6803 bool fixed = req->flags & REQ_F_FIXED_FILE;
6805 req->file = io_file_get(state, req, READ_ONCE(sqe->fd), fixed);
6806 if (unlikely(!req->file &&
6807 !io_op_defs[req->opcode].needs_file_no_error))
6815 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6817 struct io_submit_state state;
6818 struct io_submit_link link;
6819 int i, submitted = 0;
6821 /* if we have a backlog and couldn't flush it all, return BUSY */
6822 if (test_bit(0, &ctx->sq_check_overflow)) {
6823 if (!io_cqring_overflow_flush(ctx, false, NULL, NULL))
6827 /* make sure SQ entry isn't read before tail */
6828 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6830 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6833 percpu_counter_add(¤t->io_uring->inflight, nr);
6834 refcount_add(nr, ¤t->usage);
6836 io_submit_state_start(&state, ctx, nr);
6839 for (i = 0; i < nr; i++) {
6840 const struct io_uring_sqe *sqe;
6841 struct io_kiocb *req;
6844 sqe = io_get_sqe(ctx);
6845 if (unlikely(!sqe)) {
6846 io_consume_sqe(ctx);
6849 req = io_alloc_req(ctx, &state);
6850 if (unlikely(!req)) {
6852 submitted = -EAGAIN;
6855 io_consume_sqe(ctx);
6856 /* will complete beyond this point, count as submitted */
6859 err = io_init_req(ctx, req, sqe, &state);
6860 if (unlikely(err)) {
6863 io_req_complete(req, err);
6867 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6868 true, io_async_submit(ctx));
6869 err = io_submit_sqe(req, sqe, &link, &state.comp);
6874 if (unlikely(submitted != nr)) {
6875 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6876 struct io_uring_task *tctx = current->io_uring;
6877 int unused = nr - ref_used;
6879 percpu_ref_put_many(&ctx->refs, unused);
6880 percpu_counter_sub(&tctx->inflight, unused);
6881 put_task_struct_many(current, unused);
6884 io_queue_link_head(link.head, &state.comp);
6885 io_submit_state_end(&state);
6887 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6888 io_commit_sqring(ctx);
6893 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6895 /* Tell userspace we may need a wakeup call */
6896 spin_lock_irq(&ctx->completion_lock);
6897 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6898 spin_unlock_irq(&ctx->completion_lock);
6901 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6903 spin_lock_irq(&ctx->completion_lock);
6904 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6905 spin_unlock_irq(&ctx->completion_lock);
6908 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
6910 unsigned int to_submit;
6913 to_submit = io_sqring_entries(ctx);
6914 /* if we're handling multiple rings, cap submit size for fairness */
6915 if (cap_entries && to_submit > 8)
6918 if (!list_empty(&ctx->iopoll_list) || to_submit) {
6919 unsigned nr_events = 0;
6921 mutex_lock(&ctx->uring_lock);
6922 if (!list_empty(&ctx->iopoll_list))
6923 io_do_iopoll(ctx, &nr_events, 0);
6925 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)))
6926 ret = io_submit_sqes(ctx, to_submit);
6927 mutex_unlock(&ctx->uring_lock);
6930 if (!io_sqring_full(ctx) && wq_has_sleeper(&ctx->sqo_sq_wait))
6931 wake_up(&ctx->sqo_sq_wait);
6936 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
6938 struct io_ring_ctx *ctx;
6939 unsigned sq_thread_idle = 0;
6941 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6942 if (sq_thread_idle < ctx->sq_thread_idle)
6943 sq_thread_idle = ctx->sq_thread_idle;
6946 sqd->sq_thread_idle = sq_thread_idle;
6949 static void io_sqd_init_new(struct io_sq_data *sqd)
6951 struct io_ring_ctx *ctx;
6953 while (!list_empty(&sqd->ctx_new_list)) {
6954 ctx = list_first_entry(&sqd->ctx_new_list, struct io_ring_ctx, sqd_list);
6955 list_move_tail(&ctx->sqd_list, &sqd->ctx_list);
6956 complete(&ctx->sq_thread_comp);
6959 io_sqd_update_thread_idle(sqd);
6962 static int io_sq_thread(void *data)
6964 struct cgroup_subsys_state *cur_css = NULL;
6965 struct files_struct *old_files = current->files;
6966 struct nsproxy *old_nsproxy = current->nsproxy;
6967 const struct cred *old_cred = NULL;
6968 struct io_sq_data *sqd = data;
6969 struct io_ring_ctx *ctx;
6970 unsigned long timeout = 0;
6974 current->files = NULL;
6975 current->nsproxy = NULL;
6976 task_unlock(current);
6978 while (!kthread_should_stop()) {
6980 bool cap_entries, sqt_spin, needs_sched;
6983 * Any changes to the sqd lists are synchronized through the
6984 * kthread parking. This synchronizes the thread vs users,
6985 * the users are synchronized on the sqd->ctx_lock.
6987 if (kthread_should_park()) {
6990 * When sq thread is unparked, in case the previous park operation
6991 * comes from io_put_sq_data(), which means that sq thread is going
6992 * to be stopped, so here needs to have a check.
6994 if (kthread_should_stop())
6998 if (unlikely(!list_empty(&sqd->ctx_new_list))) {
6999 io_sqd_init_new(sqd);
7000 timeout = jiffies + sqd->sq_thread_idle;
7004 cap_entries = !list_is_singular(&sqd->ctx_list);
7005 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
7006 if (current->cred != ctx->creds) {
7008 revert_creds(old_cred);
7009 old_cred = override_creds(ctx->creds);
7011 io_sq_thread_associate_blkcg(ctx, &cur_css);
7013 current->loginuid = ctx->loginuid;
7014 current->sessionid = ctx->sessionid;
7017 ret = __io_sq_thread(ctx, cap_entries);
7018 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
7021 io_sq_thread_drop_mm_files();
7024 if (sqt_spin || !time_after(jiffies, timeout)) {
7028 timeout = jiffies + sqd->sq_thread_idle;
7032 if (kthread_should_park())
7036 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
7037 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
7038 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
7039 !list_empty_careful(&ctx->iopoll_list)) {
7040 needs_sched = false;
7043 if (io_sqring_entries(ctx)) {
7044 needs_sched = false;
7050 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7051 io_ring_set_wakeup_flag(ctx);
7054 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7055 io_ring_clear_wakeup_flag(ctx);
7058 finish_wait(&sqd->wait, &wait);
7059 timeout = jiffies + sqd->sq_thread_idle;
7065 io_sq_thread_unassociate_blkcg();
7067 revert_creds(old_cred);
7070 current->files = old_files;
7071 current->nsproxy = old_nsproxy;
7072 task_unlock(current);
7079 struct io_wait_queue {
7080 struct wait_queue_entry wq;
7081 struct io_ring_ctx *ctx;
7083 unsigned nr_timeouts;
7086 static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush)
7088 struct io_ring_ctx *ctx = iowq->ctx;
7091 * Wake up if we have enough events, or if a timeout occurred since we
7092 * started waiting. For timeouts, we always want to return to userspace,
7093 * regardless of event count.
7095 return io_cqring_events(ctx, noflush) >= iowq->to_wait ||
7096 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
7099 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
7100 int wake_flags, void *key)
7102 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
7105 /* use noflush == true, as we can't safely rely on locking context */
7106 if (!io_should_wake(iowq, true))
7109 return autoremove_wake_function(curr, mode, wake_flags, key);
7112 static int io_run_task_work_sig(void)
7114 if (io_run_task_work())
7116 if (!signal_pending(current))
7118 if (test_tsk_thread_flag(current, TIF_NOTIFY_SIGNAL))
7119 return -ERESTARTSYS;
7124 * Wait until events become available, if we don't already have some. The
7125 * application must reap them itself, as they reside on the shared cq ring.
7127 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
7128 const sigset_t __user *sig, size_t sigsz,
7129 struct __kernel_timespec __user *uts)
7131 struct io_wait_queue iowq = {
7134 .func = io_wake_function,
7135 .entry = LIST_HEAD_INIT(iowq.wq.entry),
7138 .to_wait = min_events,
7140 struct io_rings *rings = ctx->rings;
7141 struct timespec64 ts;
7142 signed long timeout = 0;
7146 if (io_cqring_events(ctx, false) >= min_events)
7148 if (!io_run_task_work())
7153 #ifdef CONFIG_COMPAT
7154 if (in_compat_syscall())
7155 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
7159 ret = set_user_sigmask(sig, sigsz);
7166 if (get_timespec64(&ts, uts))
7168 timeout = timespec64_to_jiffies(&ts);
7171 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
7172 trace_io_uring_cqring_wait(ctx, min_events);
7174 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
7175 TASK_INTERRUPTIBLE);
7176 /* make sure we run task_work before checking for signals */
7177 ret = io_run_task_work_sig();
7182 if (io_should_wake(&iowq, false))
7185 timeout = schedule_timeout(timeout);
7194 finish_wait(&ctx->wait, &iowq.wq);
7196 restore_saved_sigmask_unless(ret == -EINTR);
7198 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
7201 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
7203 #if defined(CONFIG_UNIX)
7204 if (ctx->ring_sock) {
7205 struct sock *sock = ctx->ring_sock->sk;
7206 struct sk_buff *skb;
7208 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
7214 for (i = 0; i < ctx->nr_user_files; i++) {
7217 file = io_file_from_index(ctx, i);
7224 static void io_file_ref_kill(struct percpu_ref *ref)
7226 struct fixed_file_data *data;
7228 data = container_of(ref, struct fixed_file_data, refs);
7229 complete(&data->done);
7232 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7234 struct fixed_file_data *data = ctx->file_data;
7235 struct fixed_file_ref_node *ref_node = NULL;
7236 unsigned nr_tables, i;
7241 spin_lock_bh(&data->lock);
7242 ref_node = data->node;
7243 spin_unlock_bh(&data->lock);
7245 percpu_ref_kill(&ref_node->refs);
7247 percpu_ref_kill(&data->refs);
7249 /* wait for all refs nodes to complete */
7250 flush_delayed_work(&ctx->file_put_work);
7251 wait_for_completion(&data->done);
7253 __io_sqe_files_unregister(ctx);
7254 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
7255 for (i = 0; i < nr_tables; i++)
7256 kfree(data->table[i].files);
7258 percpu_ref_exit(&data->refs);
7260 ctx->file_data = NULL;
7261 ctx->nr_user_files = 0;
7265 static void io_put_sq_data(struct io_sq_data *sqd)
7267 if (refcount_dec_and_test(&sqd->refs)) {
7269 * The park is a bit of a work-around, without it we get
7270 * warning spews on shutdown with SQPOLL set and affinity
7271 * set to a single CPU.
7274 kthread_park(sqd->thread);
7275 kthread_stop(sqd->thread);
7282 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7284 struct io_ring_ctx *ctx_attach;
7285 struct io_sq_data *sqd;
7288 f = fdget(p->wq_fd);
7290 return ERR_PTR(-ENXIO);
7291 if (f.file->f_op != &io_uring_fops) {
7293 return ERR_PTR(-EINVAL);
7296 ctx_attach = f.file->private_data;
7297 sqd = ctx_attach->sq_data;
7300 return ERR_PTR(-EINVAL);
7303 refcount_inc(&sqd->refs);
7308 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p)
7310 struct io_sq_data *sqd;
7312 if (p->flags & IORING_SETUP_ATTACH_WQ)
7313 return io_attach_sq_data(p);
7315 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7317 return ERR_PTR(-ENOMEM);
7319 refcount_set(&sqd->refs, 1);
7320 INIT_LIST_HEAD(&sqd->ctx_list);
7321 INIT_LIST_HEAD(&sqd->ctx_new_list);
7322 mutex_init(&sqd->ctx_lock);
7323 mutex_init(&sqd->lock);
7324 init_waitqueue_head(&sqd->wait);
7328 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7329 __releases(&sqd->lock)
7333 kthread_unpark(sqd->thread);
7334 mutex_unlock(&sqd->lock);
7337 static void io_sq_thread_park(struct io_sq_data *sqd)
7338 __acquires(&sqd->lock)
7342 mutex_lock(&sqd->lock);
7343 kthread_park(sqd->thread);
7346 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
7348 struct io_sq_data *sqd = ctx->sq_data;
7353 * We may arrive here from the error branch in
7354 * io_sq_offload_create() where the kthread is created
7355 * without being waked up, thus wake it up now to make
7356 * sure the wait will complete.
7358 wake_up_process(sqd->thread);
7359 wait_for_completion(&ctx->sq_thread_comp);
7361 io_sq_thread_park(sqd);
7364 mutex_lock(&sqd->ctx_lock);
7365 list_del(&ctx->sqd_list);
7366 io_sqd_update_thread_idle(sqd);
7367 mutex_unlock(&sqd->ctx_lock);
7370 io_sq_thread_unpark(sqd);
7372 io_put_sq_data(sqd);
7373 ctx->sq_data = NULL;
7377 static void io_finish_async(struct io_ring_ctx *ctx)
7379 io_sq_thread_stop(ctx);
7382 io_wq_destroy(ctx->io_wq);
7387 #if defined(CONFIG_UNIX)
7389 * Ensure the UNIX gc is aware of our file set, so we are certain that
7390 * the io_uring can be safely unregistered on process exit, even if we have
7391 * loops in the file referencing.
7393 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7395 struct sock *sk = ctx->ring_sock->sk;
7396 struct scm_fp_list *fpl;
7397 struct sk_buff *skb;
7400 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7404 skb = alloc_skb(0, GFP_KERNEL);
7413 fpl->user = get_uid(ctx->user);
7414 for (i = 0; i < nr; i++) {
7415 struct file *file = io_file_from_index(ctx, i + offset);
7419 fpl->fp[nr_files] = get_file(file);
7420 unix_inflight(fpl->user, fpl->fp[nr_files]);
7425 fpl->max = SCM_MAX_FD;
7426 fpl->count = nr_files;
7427 UNIXCB(skb).fp = fpl;
7428 skb->destructor = unix_destruct_scm;
7429 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7430 skb_queue_head(&sk->sk_receive_queue, skb);
7432 for (i = 0; i < nr_files; i++)
7443 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7444 * causes regular reference counting to break down. We rely on the UNIX
7445 * garbage collection to take care of this problem for us.
7447 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7449 unsigned left, total;
7453 left = ctx->nr_user_files;
7455 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7457 ret = __io_sqe_files_scm(ctx, this_files, total);
7461 total += this_files;
7467 while (total < ctx->nr_user_files) {
7468 struct file *file = io_file_from_index(ctx, total);
7478 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7484 static int io_sqe_alloc_file_tables(struct fixed_file_data *file_data,
7485 unsigned nr_tables, unsigned nr_files)
7489 for (i = 0; i < nr_tables; i++) {
7490 struct fixed_file_table *table = &file_data->table[i];
7491 unsigned this_files;
7493 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
7494 table->files = kcalloc(this_files, sizeof(struct file *),
7498 nr_files -= this_files;
7504 for (i = 0; i < nr_tables; i++) {
7505 struct fixed_file_table *table = &file_data->table[i];
7506 kfree(table->files);
7511 static void io_ring_file_put(struct io_ring_ctx *ctx, struct file *file)
7513 #if defined(CONFIG_UNIX)
7514 struct sock *sock = ctx->ring_sock->sk;
7515 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7516 struct sk_buff *skb;
7519 __skb_queue_head_init(&list);
7522 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7523 * remove this entry and rearrange the file array.
7525 skb = skb_dequeue(head);
7527 struct scm_fp_list *fp;
7529 fp = UNIXCB(skb).fp;
7530 for (i = 0; i < fp->count; i++) {
7533 if (fp->fp[i] != file)
7536 unix_notinflight(fp->user, fp->fp[i]);
7537 left = fp->count - 1 - i;
7539 memmove(&fp->fp[i], &fp->fp[i + 1],
7540 left * sizeof(struct file *));
7547 __skb_queue_tail(&list, skb);
7557 __skb_queue_tail(&list, skb);
7559 skb = skb_dequeue(head);
7562 if (skb_peek(&list)) {
7563 spin_lock_irq(&head->lock);
7564 while ((skb = __skb_dequeue(&list)) != NULL)
7565 __skb_queue_tail(head, skb);
7566 spin_unlock_irq(&head->lock);
7573 struct io_file_put {
7574 struct list_head list;
7578 static void __io_file_put_work(struct fixed_file_ref_node *ref_node)
7580 struct fixed_file_data *file_data = ref_node->file_data;
7581 struct io_ring_ctx *ctx = file_data->ctx;
7582 struct io_file_put *pfile, *tmp;
7584 list_for_each_entry_safe(pfile, tmp, &ref_node->file_list, list) {
7585 list_del(&pfile->list);
7586 io_ring_file_put(ctx, pfile->file);
7590 percpu_ref_exit(&ref_node->refs);
7592 percpu_ref_put(&file_data->refs);
7595 static void io_file_put_work(struct work_struct *work)
7597 struct io_ring_ctx *ctx;
7598 struct llist_node *node;
7600 ctx = container_of(work, struct io_ring_ctx, file_put_work.work);
7601 node = llist_del_all(&ctx->file_put_llist);
7604 struct fixed_file_ref_node *ref_node;
7605 struct llist_node *next = node->next;
7607 ref_node = llist_entry(node, struct fixed_file_ref_node, llist);
7608 __io_file_put_work(ref_node);
7613 static void io_file_data_ref_zero(struct percpu_ref *ref)
7615 struct fixed_file_ref_node *ref_node;
7616 struct fixed_file_data *data;
7617 struct io_ring_ctx *ctx;
7618 bool first_add = false;
7621 ref_node = container_of(ref, struct fixed_file_ref_node, refs);
7622 data = ref_node->file_data;
7625 spin_lock_bh(&data->lock);
7626 ref_node->done = true;
7628 while (!list_empty(&data->ref_list)) {
7629 ref_node = list_first_entry(&data->ref_list,
7630 struct fixed_file_ref_node, node);
7631 /* recycle ref nodes in order */
7632 if (!ref_node->done)
7634 list_del(&ref_node->node);
7635 first_add |= llist_add(&ref_node->llist, &ctx->file_put_llist);
7637 spin_unlock_bh(&data->lock);
7639 if (percpu_ref_is_dying(&data->refs))
7643 mod_delayed_work(system_wq, &ctx->file_put_work, 0);
7645 queue_delayed_work(system_wq, &ctx->file_put_work, delay);
7648 static struct fixed_file_ref_node *alloc_fixed_file_ref_node(
7649 struct io_ring_ctx *ctx)
7651 struct fixed_file_ref_node *ref_node;
7653 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7655 return ERR_PTR(-ENOMEM);
7657 if (percpu_ref_init(&ref_node->refs, io_file_data_ref_zero,
7660 return ERR_PTR(-ENOMEM);
7662 INIT_LIST_HEAD(&ref_node->node);
7663 INIT_LIST_HEAD(&ref_node->file_list);
7664 ref_node->file_data = ctx->file_data;
7665 ref_node->done = false;
7669 static void destroy_fixed_file_ref_node(struct fixed_file_ref_node *ref_node)
7671 percpu_ref_exit(&ref_node->refs);
7675 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7678 __s32 __user *fds = (__s32 __user *) arg;
7679 unsigned nr_tables, i;
7681 int fd, ret = -ENOMEM;
7682 struct fixed_file_ref_node *ref_node;
7683 struct fixed_file_data *file_data;
7689 if (nr_args > IORING_MAX_FIXED_FILES)
7692 file_data = kzalloc(sizeof(*ctx->file_data), GFP_KERNEL);
7695 file_data->ctx = ctx;
7696 init_completion(&file_data->done);
7697 INIT_LIST_HEAD(&file_data->ref_list);
7698 spin_lock_init(&file_data->lock);
7700 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
7701 file_data->table = kcalloc(nr_tables, sizeof(*file_data->table),
7703 if (!file_data->table)
7706 if (percpu_ref_init(&file_data->refs, io_file_ref_kill,
7707 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
7710 if (io_sqe_alloc_file_tables(file_data, nr_tables, nr_args))
7712 ctx->file_data = file_data;
7714 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7715 struct fixed_file_table *table;
7718 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7722 /* allow sparse sets */
7732 * Don't allow io_uring instances to be registered. If UNIX
7733 * isn't enabled, then this causes a reference cycle and this
7734 * instance can never get freed. If UNIX is enabled we'll
7735 * handle it just fine, but there's still no point in allowing
7736 * a ring fd as it doesn't support regular read/write anyway.
7738 if (file->f_op == &io_uring_fops) {
7742 table = &file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7743 index = i & IORING_FILE_TABLE_MASK;
7744 table->files[index] = file;
7747 ret = io_sqe_files_scm(ctx);
7749 io_sqe_files_unregister(ctx);
7753 ref_node = alloc_fixed_file_ref_node(ctx);
7754 if (IS_ERR(ref_node)) {
7755 io_sqe_files_unregister(ctx);
7756 return PTR_ERR(ref_node);
7759 file_data->node = ref_node;
7760 spin_lock_bh(&file_data->lock);
7761 list_add_tail(&ref_node->node, &file_data->ref_list);
7762 spin_unlock_bh(&file_data->lock);
7763 percpu_ref_get(&file_data->refs);
7766 for (i = 0; i < ctx->nr_user_files; i++) {
7767 file = io_file_from_index(ctx, i);
7771 for (i = 0; i < nr_tables; i++)
7772 kfree(file_data->table[i].files);
7773 ctx->nr_user_files = 0;
7775 percpu_ref_exit(&file_data->refs);
7777 kfree(file_data->table);
7779 ctx->file_data = NULL;
7783 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7786 #if defined(CONFIG_UNIX)
7787 struct sock *sock = ctx->ring_sock->sk;
7788 struct sk_buff_head *head = &sock->sk_receive_queue;
7789 struct sk_buff *skb;
7792 * See if we can merge this file into an existing skb SCM_RIGHTS
7793 * file set. If there's no room, fall back to allocating a new skb
7794 * and filling it in.
7796 spin_lock_irq(&head->lock);
7797 skb = skb_peek(head);
7799 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7801 if (fpl->count < SCM_MAX_FD) {
7802 __skb_unlink(skb, head);
7803 spin_unlock_irq(&head->lock);
7804 fpl->fp[fpl->count] = get_file(file);
7805 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7807 spin_lock_irq(&head->lock);
7808 __skb_queue_head(head, skb);
7813 spin_unlock_irq(&head->lock);
7820 return __io_sqe_files_scm(ctx, 1, index);
7826 static int io_queue_file_removal(struct fixed_file_data *data,
7829 struct io_file_put *pfile;
7830 struct fixed_file_ref_node *ref_node = data->node;
7832 pfile = kzalloc(sizeof(*pfile), GFP_KERNEL);
7837 list_add(&pfile->list, &ref_node->file_list);
7842 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7843 struct io_uring_files_update *up,
7846 struct fixed_file_data *data = ctx->file_data;
7847 struct fixed_file_ref_node *ref_node;
7852 bool needs_switch = false;
7854 if (check_add_overflow(up->offset, nr_args, &done))
7856 if (done > ctx->nr_user_files)
7859 ref_node = alloc_fixed_file_ref_node(ctx);
7860 if (IS_ERR(ref_node))
7861 return PTR_ERR(ref_node);
7864 fds = u64_to_user_ptr(up->fds);
7866 struct fixed_file_table *table;
7870 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
7874 i = array_index_nospec(up->offset, ctx->nr_user_files);
7875 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7876 index = i & IORING_FILE_TABLE_MASK;
7877 if (table->files[index]) {
7878 file = table->files[index];
7879 err = io_queue_file_removal(data, file);
7882 table->files[index] = NULL;
7883 needs_switch = true;
7892 * Don't allow io_uring instances to be registered. If
7893 * UNIX isn't enabled, then this causes a reference
7894 * cycle and this instance can never get freed. If UNIX
7895 * is enabled we'll handle it just fine, but there's
7896 * still no point in allowing a ring fd as it doesn't
7897 * support regular read/write anyway.
7899 if (file->f_op == &io_uring_fops) {
7904 table->files[index] = file;
7905 err = io_sqe_file_register(ctx, file, i);
7907 table->files[index] = NULL;
7918 percpu_ref_kill(&data->node->refs);
7919 spin_lock_bh(&data->lock);
7920 list_add_tail(&ref_node->node, &data->ref_list);
7921 data->node = ref_node;
7922 spin_unlock_bh(&data->lock);
7923 percpu_ref_get(&ctx->file_data->refs);
7925 destroy_fixed_file_ref_node(ref_node);
7927 return done ? done : err;
7930 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
7933 struct io_uring_files_update up;
7935 if (!ctx->file_data)
7939 if (copy_from_user(&up, arg, sizeof(up)))
7944 return __io_sqe_files_update(ctx, &up, nr_args);
7947 static void io_free_work(struct io_wq_work *work)
7949 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7951 /* Consider that io_steal_work() relies on this ref */
7955 static int io_init_wq_offload(struct io_ring_ctx *ctx,
7956 struct io_uring_params *p)
7958 struct io_wq_data data;
7960 struct io_ring_ctx *ctx_attach;
7961 unsigned int concurrency;
7964 data.user = ctx->user;
7965 data.free_work = io_free_work;
7966 data.do_work = io_wq_submit_work;
7968 if (!(p->flags & IORING_SETUP_ATTACH_WQ)) {
7969 /* Do QD, or 4 * CPUS, whatever is smallest */
7970 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7972 ctx->io_wq = io_wq_create(concurrency, &data);
7973 if (IS_ERR(ctx->io_wq)) {
7974 ret = PTR_ERR(ctx->io_wq);
7980 f = fdget(p->wq_fd);
7984 if (f.file->f_op != &io_uring_fops) {
7989 ctx_attach = f.file->private_data;
7990 /* @io_wq is protected by holding the fd */
7991 if (!io_wq_get(ctx_attach->io_wq, &data)) {
7996 ctx->io_wq = ctx_attach->io_wq;
8002 static int io_uring_alloc_task_context(struct task_struct *task)
8004 struct io_uring_task *tctx;
8007 tctx = kmalloc(sizeof(*tctx), GFP_KERNEL);
8008 if (unlikely(!tctx))
8011 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
8012 if (unlikely(ret)) {
8018 init_waitqueue_head(&tctx->wait);
8020 atomic_set(&tctx->in_idle, 0);
8021 tctx->sqpoll = false;
8022 io_init_identity(&tctx->__identity);
8023 tctx->identity = &tctx->__identity;
8024 task->io_uring = tctx;
8028 void __io_uring_free(struct task_struct *tsk)
8030 struct io_uring_task *tctx = tsk->io_uring;
8032 WARN_ON_ONCE(!xa_empty(&tctx->xa));
8033 WARN_ON_ONCE(refcount_read(&tctx->identity->count) != 1);
8034 if (tctx->identity != &tctx->__identity)
8035 kfree(tctx->identity);
8036 percpu_counter_destroy(&tctx->inflight);
8038 tsk->io_uring = NULL;
8041 static int io_sq_offload_create(struct io_ring_ctx *ctx,
8042 struct io_uring_params *p)
8046 if (ctx->flags & IORING_SETUP_SQPOLL) {
8047 struct io_sq_data *sqd;
8050 if (!capable(CAP_SYS_ADMIN) && !capable(CAP_SYS_NICE))
8053 sqd = io_get_sq_data(p);
8060 io_sq_thread_park(sqd);
8061 mutex_lock(&sqd->ctx_lock);
8062 list_add(&ctx->sqd_list, &sqd->ctx_new_list);
8063 mutex_unlock(&sqd->ctx_lock);
8064 io_sq_thread_unpark(sqd);
8066 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
8067 if (!ctx->sq_thread_idle)
8068 ctx->sq_thread_idle = HZ;
8073 if (p->flags & IORING_SETUP_SQ_AFF) {
8074 int cpu = p->sq_thread_cpu;
8077 if (cpu >= nr_cpu_ids)
8079 if (!cpu_online(cpu))
8082 sqd->thread = kthread_create_on_cpu(io_sq_thread, sqd,
8083 cpu, "io_uring-sq");
8085 sqd->thread = kthread_create(io_sq_thread, sqd,
8088 if (IS_ERR(sqd->thread)) {
8089 ret = PTR_ERR(sqd->thread);
8093 ret = io_uring_alloc_task_context(sqd->thread);
8096 } else if (p->flags & IORING_SETUP_SQ_AFF) {
8097 /* Can't have SQ_AFF without SQPOLL */
8103 ret = io_init_wq_offload(ctx, p);
8109 io_finish_async(ctx);
8113 static void io_sq_offload_start(struct io_ring_ctx *ctx)
8115 struct io_sq_data *sqd = ctx->sq_data;
8117 if ((ctx->flags & IORING_SETUP_SQPOLL) && sqd->thread)
8118 wake_up_process(sqd->thread);
8121 static inline void __io_unaccount_mem(struct user_struct *user,
8122 unsigned long nr_pages)
8124 atomic_long_sub(nr_pages, &user->locked_vm);
8127 static inline int __io_account_mem(struct user_struct *user,
8128 unsigned long nr_pages)
8130 unsigned long page_limit, cur_pages, new_pages;
8132 /* Don't allow more pages than we can safely lock */
8133 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
8136 cur_pages = atomic_long_read(&user->locked_vm);
8137 new_pages = cur_pages + nr_pages;
8138 if (new_pages > page_limit)
8140 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
8141 new_pages) != cur_pages);
8146 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
8147 enum io_mem_account acct)
8150 __io_unaccount_mem(ctx->user, nr_pages);
8152 if (ctx->mm_account) {
8153 if (acct == ACCT_LOCKED) {
8154 mmap_write_lock(ctx->mm_account);
8155 ctx->mm_account->locked_vm -= nr_pages;
8156 mmap_write_unlock(ctx->mm_account);
8157 }else if (acct == ACCT_PINNED) {
8158 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
8163 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
8164 enum io_mem_account acct)
8168 if (ctx->limit_mem) {
8169 ret = __io_account_mem(ctx->user, nr_pages);
8174 if (ctx->mm_account) {
8175 if (acct == ACCT_LOCKED) {
8176 mmap_write_lock(ctx->mm_account);
8177 ctx->mm_account->locked_vm += nr_pages;
8178 mmap_write_unlock(ctx->mm_account);
8179 } else if (acct == ACCT_PINNED) {
8180 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
8187 static void io_mem_free(void *ptr)
8194 page = virt_to_head_page(ptr);
8195 if (put_page_testzero(page))
8196 free_compound_page(page);
8199 static void *io_mem_alloc(size_t size)
8201 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
8204 return (void *) __get_free_pages(gfp_flags, get_order(size));
8207 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8210 struct io_rings *rings;
8211 size_t off, sq_array_size;
8213 off = struct_size(rings, cqes, cq_entries);
8214 if (off == SIZE_MAX)
8218 off = ALIGN(off, SMP_CACHE_BYTES);
8226 sq_array_size = array_size(sizeof(u32), sq_entries);
8227 if (sq_array_size == SIZE_MAX)
8230 if (check_add_overflow(off, sq_array_size, &off))
8236 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
8240 pages = (size_t)1 << get_order(
8241 rings_size(sq_entries, cq_entries, NULL));
8242 pages += (size_t)1 << get_order(
8243 array_size(sizeof(struct io_uring_sqe), sq_entries));
8248 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
8252 if (!ctx->user_bufs)
8255 for (i = 0; i < ctx->nr_user_bufs; i++) {
8256 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8258 for (j = 0; j < imu->nr_bvecs; j++)
8259 unpin_user_page(imu->bvec[j].bv_page);
8261 if (imu->acct_pages)
8262 io_unaccount_mem(ctx, imu->acct_pages, ACCT_PINNED);
8267 kfree(ctx->user_bufs);
8268 ctx->user_bufs = NULL;
8269 ctx->nr_user_bufs = 0;
8273 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8274 void __user *arg, unsigned index)
8276 struct iovec __user *src;
8278 #ifdef CONFIG_COMPAT
8280 struct compat_iovec __user *ciovs;
8281 struct compat_iovec ciov;
8283 ciovs = (struct compat_iovec __user *) arg;
8284 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8287 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8288 dst->iov_len = ciov.iov_len;
8292 src = (struct iovec __user *) arg;
8293 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8299 * Not super efficient, but this is just a registration time. And we do cache
8300 * the last compound head, so generally we'll only do a full search if we don't
8303 * We check if the given compound head page has already been accounted, to
8304 * avoid double accounting it. This allows us to account the full size of the
8305 * page, not just the constituent pages of a huge page.
8307 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8308 int nr_pages, struct page *hpage)
8312 /* check current page array */
8313 for (i = 0; i < nr_pages; i++) {
8314 if (!PageCompound(pages[i]))
8316 if (compound_head(pages[i]) == hpage)
8320 /* check previously registered pages */
8321 for (i = 0; i < ctx->nr_user_bufs; i++) {
8322 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8324 for (j = 0; j < imu->nr_bvecs; j++) {
8325 if (!PageCompound(imu->bvec[j].bv_page))
8327 if (compound_head(imu->bvec[j].bv_page) == hpage)
8335 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8336 int nr_pages, struct io_mapped_ubuf *imu,
8337 struct page **last_hpage)
8341 for (i = 0; i < nr_pages; i++) {
8342 if (!PageCompound(pages[i])) {
8347 hpage = compound_head(pages[i]);
8348 if (hpage == *last_hpage)
8350 *last_hpage = hpage;
8351 if (headpage_already_acct(ctx, pages, i, hpage))
8353 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8357 if (!imu->acct_pages)
8360 ret = io_account_mem(ctx, imu->acct_pages, ACCT_PINNED);
8362 imu->acct_pages = 0;
8366 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
8369 struct vm_area_struct **vmas = NULL;
8370 struct page **pages = NULL;
8371 struct page *last_hpage = NULL;
8372 int i, j, got_pages = 0;
8377 if (!nr_args || nr_args > UIO_MAXIOV)
8380 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
8382 if (!ctx->user_bufs)
8385 for (i = 0; i < nr_args; i++) {
8386 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8387 unsigned long off, start, end, ubuf;
8392 ret = io_copy_iov(ctx, &iov, arg, i);
8397 * Don't impose further limits on the size and buffer
8398 * constraints here, we'll -EINVAL later when IO is
8399 * submitted if they are wrong.
8402 if (!iov.iov_base || !iov.iov_len)
8405 /* arbitrary limit, but we need something */
8406 if (iov.iov_len > SZ_1G)
8409 ubuf = (unsigned long) iov.iov_base;
8410 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8411 start = ubuf >> PAGE_SHIFT;
8412 nr_pages = end - start;
8415 if (!pages || nr_pages > got_pages) {
8418 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
8420 vmas = kvmalloc_array(nr_pages,
8421 sizeof(struct vm_area_struct *),
8423 if (!pages || !vmas) {
8427 got_pages = nr_pages;
8430 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
8437 mmap_read_lock(current->mm);
8438 pret = pin_user_pages(ubuf, nr_pages,
8439 FOLL_WRITE | FOLL_LONGTERM,
8441 if (pret == nr_pages) {
8442 /* don't support file backed memory */
8443 for (j = 0; j < nr_pages; j++) {
8444 struct vm_area_struct *vma = vmas[j];
8447 !is_file_hugepages(vma->vm_file)) {
8453 ret = pret < 0 ? pret : -EFAULT;
8455 mmap_read_unlock(current->mm);
8458 * if we did partial map, or found file backed vmas,
8459 * release any pages we did get
8462 unpin_user_pages(pages, pret);
8467 ret = io_buffer_account_pin(ctx, pages, pret, imu, &last_hpage);
8469 unpin_user_pages(pages, pret);
8474 off = ubuf & ~PAGE_MASK;
8476 for (j = 0; j < nr_pages; j++) {
8479 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8480 imu->bvec[j].bv_page = pages[j];
8481 imu->bvec[j].bv_len = vec_len;
8482 imu->bvec[j].bv_offset = off;
8486 /* store original address for later verification */
8488 imu->len = iov.iov_len;
8489 imu->nr_bvecs = nr_pages;
8491 ctx->nr_user_bufs++;
8499 io_sqe_buffer_unregister(ctx);
8503 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8505 __s32 __user *fds = arg;
8511 if (copy_from_user(&fd, fds, sizeof(*fds)))
8514 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8515 if (IS_ERR(ctx->cq_ev_fd)) {
8516 int ret = PTR_ERR(ctx->cq_ev_fd);
8517 ctx->cq_ev_fd = NULL;
8524 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8526 if (ctx->cq_ev_fd) {
8527 eventfd_ctx_put(ctx->cq_ev_fd);
8528 ctx->cq_ev_fd = NULL;
8535 static int __io_destroy_buffers(int id, void *p, void *data)
8537 struct io_ring_ctx *ctx = data;
8538 struct io_buffer *buf = p;
8540 __io_remove_buffers(ctx, buf, id, -1U);
8544 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8546 idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx);
8547 idr_destroy(&ctx->io_buffer_idr);
8550 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8552 io_finish_async(ctx);
8553 io_sqe_buffer_unregister(ctx);
8555 if (ctx->sqo_task) {
8556 put_task_struct(ctx->sqo_task);
8557 ctx->sqo_task = NULL;
8558 mmdrop(ctx->mm_account);
8559 ctx->mm_account = NULL;
8562 #ifdef CONFIG_BLK_CGROUP
8563 if (ctx->sqo_blkcg_css)
8564 css_put(ctx->sqo_blkcg_css);
8567 io_sqe_files_unregister(ctx);
8568 io_eventfd_unregister(ctx);
8569 io_destroy_buffers(ctx);
8570 idr_destroy(&ctx->personality_idr);
8572 #if defined(CONFIG_UNIX)
8573 if (ctx->ring_sock) {
8574 ctx->ring_sock->file = NULL; /* so that iput() is called */
8575 sock_release(ctx->ring_sock);
8579 io_mem_free(ctx->rings);
8580 io_mem_free(ctx->sq_sqes);
8582 percpu_ref_exit(&ctx->refs);
8583 free_uid(ctx->user);
8584 put_cred(ctx->creds);
8585 kfree(ctx->cancel_hash);
8586 kmem_cache_free(req_cachep, ctx->fallback_req);
8590 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8592 struct io_ring_ctx *ctx = file->private_data;
8595 poll_wait(file, &ctx->cq_wait, wait);
8597 * synchronizes with barrier from wq_has_sleeper call in
8601 if (!io_sqring_full(ctx))
8602 mask |= EPOLLOUT | EPOLLWRNORM;
8603 if (io_cqring_events(ctx, false))
8604 mask |= EPOLLIN | EPOLLRDNORM;
8609 static int io_uring_fasync(int fd, struct file *file, int on)
8611 struct io_ring_ctx *ctx = file->private_data;
8613 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8616 static int io_remove_personalities(int id, void *p, void *data)
8618 struct io_ring_ctx *ctx = data;
8619 struct io_identity *iod;
8621 iod = idr_remove(&ctx->personality_idr, id);
8623 put_cred(iod->creds);
8624 if (refcount_dec_and_test(&iod->count))
8630 static void io_ring_exit_work(struct work_struct *work)
8632 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
8636 * If we're doing polled IO and end up having requests being
8637 * submitted async (out-of-line), then completions can come in while
8638 * we're waiting for refs to drop. We need to reap these manually,
8639 * as nobody else will be looking for them.
8642 io_iopoll_try_reap_events(ctx);
8643 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
8644 io_ring_ctx_free(ctx);
8647 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8649 mutex_lock(&ctx->uring_lock);
8650 percpu_ref_kill(&ctx->refs);
8651 /* if force is set, the ring is going away. always drop after that */
8652 ctx->cq_overflow_flushed = 1;
8654 io_cqring_overflow_flush(ctx, true, NULL, NULL);
8655 mutex_unlock(&ctx->uring_lock);
8657 io_kill_timeouts(ctx, NULL, NULL);
8658 io_poll_remove_all(ctx, NULL, NULL);
8661 io_wq_cancel_all(ctx->io_wq);
8663 /* if we failed setting up the ctx, we might not have any rings */
8664 io_iopoll_try_reap_events(ctx);
8665 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
8668 * Do this upfront, so we won't have a grace period where the ring
8669 * is closed but resources aren't reaped yet. This can cause
8670 * spurious failure in setting up a new ring.
8672 io_unaccount_mem(ctx, ring_pages(ctx->sq_entries, ctx->cq_entries),
8675 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8677 * Use system_unbound_wq to avoid spawning tons of event kworkers
8678 * if we're exiting a ton of rings at the same time. It just adds
8679 * noise and overhead, there's no discernable change in runtime
8680 * over using system_wq.
8682 queue_work(system_unbound_wq, &ctx->exit_work);
8685 static int io_uring_release(struct inode *inode, struct file *file)
8687 struct io_ring_ctx *ctx = file->private_data;
8689 file->private_data = NULL;
8690 io_ring_ctx_wait_and_kill(ctx);
8694 struct io_task_cancel {
8695 struct task_struct *task;
8696 struct files_struct *files;
8699 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8701 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8702 struct io_task_cancel *cancel = data;
8705 if (cancel->files && (req->flags & REQ_F_LINK_TIMEOUT)) {
8706 unsigned long flags;
8707 struct io_ring_ctx *ctx = req->ctx;
8709 /* protect against races with linked timeouts */
8710 spin_lock_irqsave(&ctx->completion_lock, flags);
8711 ret = io_match_task(req, cancel->task, cancel->files);
8712 spin_unlock_irqrestore(&ctx->completion_lock, flags);
8714 ret = io_match_task(req, cancel->task, cancel->files);
8719 static void io_cancel_defer_files(struct io_ring_ctx *ctx,
8720 struct task_struct *task,
8721 struct files_struct *files)
8723 struct io_defer_entry *de = NULL;
8726 spin_lock_irq(&ctx->completion_lock);
8727 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8728 if (io_match_task(de->req, task, files)) {
8729 list_cut_position(&list, &ctx->defer_list, &de->list);
8733 spin_unlock_irq(&ctx->completion_lock);
8735 while (!list_empty(&list)) {
8736 de = list_first_entry(&list, struct io_defer_entry, list);
8737 list_del_init(&de->list);
8738 req_set_fail_links(de->req);
8739 io_put_req(de->req);
8740 io_req_complete(de->req, -ECANCELED);
8745 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
8746 struct task_struct *task,
8747 struct files_struct *files)
8749 while (!list_empty_careful(&ctx->inflight_list)) {
8750 struct io_task_cancel cancel = { .task = task, .files = files };
8751 struct io_kiocb *req;
8755 spin_lock_irq(&ctx->inflight_lock);
8756 list_for_each_entry(req, &ctx->inflight_list, inflight_entry) {
8757 if (req->task != task ||
8758 req->work.identity->files != files)
8764 prepare_to_wait(&task->io_uring->wait, &wait,
8765 TASK_UNINTERRUPTIBLE);
8766 spin_unlock_irq(&ctx->inflight_lock);
8768 /* We need to keep going until we don't find a matching req */
8772 io_wq_cancel_cb(ctx->io_wq, io_cancel_task_cb, &cancel, true);
8773 io_poll_remove_all(ctx, task, files);
8774 io_kill_timeouts(ctx, task, files);
8775 /* cancellations _may_ trigger task work */
8778 finish_wait(&task->io_uring->wait, &wait);
8782 static void __io_uring_cancel_task_requests(struct io_ring_ctx *ctx,
8783 struct task_struct *task)
8786 struct io_task_cancel cancel = { .task = task, .files = NULL, };
8787 enum io_wq_cancel cret;
8790 cret = io_wq_cancel_cb(ctx->io_wq, io_cancel_task_cb, &cancel, true);
8791 if (cret != IO_WQ_CANCEL_NOTFOUND)
8794 /* SQPOLL thread does its own polling */
8795 if (!(ctx->flags & IORING_SETUP_SQPOLL)) {
8796 while (!list_empty_careful(&ctx->iopoll_list)) {
8797 io_iopoll_try_reap_events(ctx);
8802 ret |= io_poll_remove_all(ctx, task, NULL);
8803 ret |= io_kill_timeouts(ctx, task, NULL);
8812 * We need to iteratively cancel requests, in case a request has dependent
8813 * hard links. These persist even for failure of cancelations, hence keep
8814 * looping until none are found.
8816 static void io_uring_cancel_task_requests(struct io_ring_ctx *ctx,
8817 struct files_struct *files)
8819 struct task_struct *task = current;
8821 if ((ctx->flags & IORING_SETUP_SQPOLL) && ctx->sq_data) {
8822 task = ctx->sq_data->thread;
8823 atomic_inc(&task->io_uring->in_idle);
8824 io_sq_thread_park(ctx->sq_data);
8827 io_cancel_defer_files(ctx, task, files);
8828 io_ring_submit_lock(ctx, (ctx->flags & IORING_SETUP_IOPOLL));
8829 io_cqring_overflow_flush(ctx, true, task, files);
8830 io_ring_submit_unlock(ctx, (ctx->flags & IORING_SETUP_IOPOLL));
8833 __io_uring_cancel_task_requests(ctx, task);
8835 io_uring_cancel_files(ctx, task, files);
8837 if ((ctx->flags & IORING_SETUP_SQPOLL) && ctx->sq_data) {
8838 atomic_dec(&task->io_uring->in_idle);
8840 * If the files that are going away are the ones in the thread
8841 * identity, clear them out.
8843 if (task->io_uring->identity->files == files)
8844 task->io_uring->identity->files = NULL;
8845 io_sq_thread_unpark(ctx->sq_data);
8850 * Note that this task has used io_uring. We use it for cancelation purposes.
8852 static int io_uring_add_task_file(struct io_ring_ctx *ctx, struct file *file)
8854 struct io_uring_task *tctx = current->io_uring;
8856 if (unlikely(!tctx)) {
8859 ret = io_uring_alloc_task_context(current);
8862 tctx = current->io_uring;
8864 if (tctx->last != file) {
8865 void *old = xa_load(&tctx->xa, (unsigned long)file);
8869 xa_store(&tctx->xa, (unsigned long)file, file, GFP_KERNEL);
8875 * This is race safe in that the task itself is doing this, hence it
8876 * cannot be going through the exit/cancel paths at the same time.
8877 * This cannot be modified while exit/cancel is running.
8879 if (!tctx->sqpoll && (ctx->flags & IORING_SETUP_SQPOLL))
8880 tctx->sqpoll = true;
8886 * Remove this io_uring_file -> task mapping.
8888 static void io_uring_del_task_file(struct file *file)
8890 struct io_uring_task *tctx = current->io_uring;
8892 if (tctx->last == file)
8894 file = xa_erase(&tctx->xa, (unsigned long)file);
8900 * Drop task note for this file if we're the only ones that hold it after
8903 static void io_uring_attempt_task_drop(struct file *file)
8905 if (!current->io_uring)
8908 * fput() is pending, will be 2 if the only other ref is our potential
8909 * task file note. If the task is exiting, drop regardless of count.
8911 if (fatal_signal_pending(current) || (current->flags & PF_EXITING) ||
8912 atomic_long_read(&file->f_count) == 2)
8913 io_uring_del_task_file(file);
8916 void __io_uring_files_cancel(struct files_struct *files)
8918 struct io_uring_task *tctx = current->io_uring;
8920 unsigned long index;
8922 /* make sure overflow events are dropped */
8923 atomic_inc(&tctx->in_idle);
8925 xa_for_each(&tctx->xa, index, file) {
8926 struct io_ring_ctx *ctx = file->private_data;
8928 io_uring_cancel_task_requests(ctx, files);
8930 io_uring_del_task_file(file);
8933 atomic_dec(&tctx->in_idle);
8936 static s64 tctx_inflight(struct io_uring_task *tctx)
8938 unsigned long index;
8942 inflight = percpu_counter_sum(&tctx->inflight);
8947 * If we have SQPOLL rings, then we need to iterate and find them, and
8948 * add the pending count for those.
8950 xa_for_each(&tctx->xa, index, file) {
8951 struct io_ring_ctx *ctx = file->private_data;
8953 if (ctx->flags & IORING_SETUP_SQPOLL) {
8954 struct io_uring_task *__tctx = ctx->sqo_task->io_uring;
8956 inflight += percpu_counter_sum(&__tctx->inflight);
8964 * Find any io_uring fd that this task has registered or done IO on, and cancel
8967 void __io_uring_task_cancel(void)
8969 struct io_uring_task *tctx = current->io_uring;
8973 /* make sure overflow events are dropped */
8974 atomic_inc(&tctx->in_idle);
8977 /* read completions before cancelations */
8978 inflight = tctx_inflight(tctx);
8981 __io_uring_files_cancel(NULL);
8983 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
8986 * If we've seen completions, retry. This avoids a race where
8987 * a completion comes in before we did prepare_to_wait().
8989 if (inflight != tctx_inflight(tctx))
8994 finish_wait(&tctx->wait, &wait);
8995 atomic_dec(&tctx->in_idle);
8998 static int io_uring_flush(struct file *file, void *data)
9000 io_uring_attempt_task_drop(file);
9004 static void *io_uring_validate_mmap_request(struct file *file,
9005 loff_t pgoff, size_t sz)
9007 struct io_ring_ctx *ctx = file->private_data;
9008 loff_t offset = pgoff << PAGE_SHIFT;
9013 case IORING_OFF_SQ_RING:
9014 case IORING_OFF_CQ_RING:
9017 case IORING_OFF_SQES:
9021 return ERR_PTR(-EINVAL);
9024 page = virt_to_head_page(ptr);
9025 if (sz > page_size(page))
9026 return ERR_PTR(-EINVAL);
9033 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9035 size_t sz = vma->vm_end - vma->vm_start;
9039 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
9041 return PTR_ERR(ptr);
9043 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
9044 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
9047 #else /* !CONFIG_MMU */
9049 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9051 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
9054 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
9056 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
9059 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
9060 unsigned long addr, unsigned long len,
9061 unsigned long pgoff, unsigned long flags)
9065 ptr = io_uring_validate_mmap_request(file, pgoff, len);
9067 return PTR_ERR(ptr);
9069 return (unsigned long) ptr;
9072 #endif /* !CONFIG_MMU */
9074 static void io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
9079 if (!io_sqring_full(ctx))
9082 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9084 if (!io_sqring_full(ctx))
9088 } while (!signal_pending(current));
9090 finish_wait(&ctx->sqo_sq_wait, &wait);
9093 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9094 struct __kernel_timespec __user **ts,
9095 const sigset_t __user **sig)
9097 struct io_uring_getevents_arg arg;
9100 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9101 * is just a pointer to the sigset_t.
9103 if (!(flags & IORING_ENTER_EXT_ARG)) {
9104 *sig = (const sigset_t __user *) argp;
9110 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9111 * timespec and sigset_t pointers if good.
9113 if (*argsz != sizeof(arg))
9115 if (copy_from_user(&arg, argp, sizeof(arg)))
9117 *sig = u64_to_user_ptr(arg.sigmask);
9118 *argsz = arg.sigmask_sz;
9119 *ts = u64_to_user_ptr(arg.ts);
9123 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9124 u32, min_complete, u32, flags, const void __user *, argp,
9127 struct io_ring_ctx *ctx;
9134 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9135 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG))
9143 if (f.file->f_op != &io_uring_fops)
9147 ctx = f.file->private_data;
9148 if (!percpu_ref_tryget(&ctx->refs))
9152 if (ctx->flags & IORING_SETUP_R_DISABLED)
9156 * For SQ polling, the thread will do all submissions and completions.
9157 * Just return the requested submit count, and wake the thread if
9161 if (ctx->flags & IORING_SETUP_SQPOLL) {
9162 io_ring_submit_lock(ctx, (ctx->flags & IORING_SETUP_IOPOLL));
9163 if (!list_empty_careful(&ctx->cq_overflow_list))
9164 io_cqring_overflow_flush(ctx, false, NULL, NULL);
9165 io_ring_submit_unlock(ctx, (ctx->flags & IORING_SETUP_IOPOLL));
9166 if (flags & IORING_ENTER_SQ_WAKEUP)
9167 wake_up(&ctx->sq_data->wait);
9168 if (flags & IORING_ENTER_SQ_WAIT)
9169 io_sqpoll_wait_sq(ctx);
9170 submitted = to_submit;
9171 } else if (to_submit) {
9172 ret = io_uring_add_task_file(ctx, f.file);
9175 mutex_lock(&ctx->uring_lock);
9176 submitted = io_submit_sqes(ctx, to_submit);
9177 mutex_unlock(&ctx->uring_lock);
9179 if (submitted != to_submit)
9182 if (flags & IORING_ENTER_GETEVENTS) {
9183 const sigset_t __user *sig;
9184 struct __kernel_timespec __user *ts;
9186 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
9190 min_complete = min(min_complete, ctx->cq_entries);
9193 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9194 * space applications don't need to do io completion events
9195 * polling again, they can rely on io_sq_thread to do polling
9196 * work, which can reduce cpu usage and uring_lock contention.
9198 if (ctx->flags & IORING_SETUP_IOPOLL &&
9199 !(ctx->flags & IORING_SETUP_SQPOLL)) {
9200 ret = io_iopoll_check(ctx, min_complete);
9202 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
9207 percpu_ref_put(&ctx->refs);
9210 return submitted ? submitted : ret;
9213 #ifdef CONFIG_PROC_FS
9214 static int io_uring_show_cred(int id, void *p, void *data)
9216 struct io_identity *iod = p;
9217 const struct cred *cred = iod->creds;
9218 struct seq_file *m = data;
9219 struct user_namespace *uns = seq_user_ns(m);
9220 struct group_info *gi;
9225 seq_printf(m, "%5d\n", id);
9226 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
9227 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
9228 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
9229 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
9230 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
9231 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
9232 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
9233 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
9234 seq_puts(m, "\n\tGroups:\t");
9235 gi = cred->group_info;
9236 for (g = 0; g < gi->ngroups; g++) {
9237 seq_put_decimal_ull(m, g ? " " : "",
9238 from_kgid_munged(uns, gi->gid[g]));
9240 seq_puts(m, "\n\tCapEff:\t");
9241 cap = cred->cap_effective;
9242 CAP_FOR_EACH_U32(__capi)
9243 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
9248 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
9250 struct io_sq_data *sq = NULL;
9255 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9256 * since fdinfo case grabs it in the opposite direction of normal use
9257 * cases. If we fail to get the lock, we just don't iterate any
9258 * structures that could be going away outside the io_uring mutex.
9260 has_lock = mutex_trylock(&ctx->uring_lock);
9262 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL))
9265 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9266 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9267 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9268 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9269 struct fixed_file_table *table;
9272 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
9273 f = table->files[i & IORING_FILE_TABLE_MASK];
9275 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9277 seq_printf(m, "%5u: <none>\n", i);
9279 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9280 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9281 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
9283 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
9284 (unsigned int) buf->len);
9286 if (has_lock && !idr_is_empty(&ctx->personality_idr)) {
9287 seq_printf(m, "Personalities:\n");
9288 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
9290 seq_printf(m, "PollList:\n");
9291 spin_lock_irq(&ctx->completion_lock);
9292 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9293 struct hlist_head *list = &ctx->cancel_hash[i];
9294 struct io_kiocb *req;
9296 hlist_for_each_entry(req, list, hash_node)
9297 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9298 req->task->task_works != NULL);
9300 spin_unlock_irq(&ctx->completion_lock);
9302 mutex_unlock(&ctx->uring_lock);
9305 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9307 struct io_ring_ctx *ctx = f->private_data;
9309 if (percpu_ref_tryget(&ctx->refs)) {
9310 __io_uring_show_fdinfo(ctx, m);
9311 percpu_ref_put(&ctx->refs);
9316 static const struct file_operations io_uring_fops = {
9317 .release = io_uring_release,
9318 .flush = io_uring_flush,
9319 .mmap = io_uring_mmap,
9321 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9322 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9324 .poll = io_uring_poll,
9325 .fasync = io_uring_fasync,
9326 #ifdef CONFIG_PROC_FS
9327 .show_fdinfo = io_uring_show_fdinfo,
9331 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9332 struct io_uring_params *p)
9334 struct io_rings *rings;
9335 size_t size, sq_array_offset;
9337 /* make sure these are sane, as we already accounted them */
9338 ctx->sq_entries = p->sq_entries;
9339 ctx->cq_entries = p->cq_entries;
9341 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9342 if (size == SIZE_MAX)
9345 rings = io_mem_alloc(size);
9350 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9351 rings->sq_ring_mask = p->sq_entries - 1;
9352 rings->cq_ring_mask = p->cq_entries - 1;
9353 rings->sq_ring_entries = p->sq_entries;
9354 rings->cq_ring_entries = p->cq_entries;
9355 ctx->sq_mask = rings->sq_ring_mask;
9356 ctx->cq_mask = rings->cq_ring_mask;
9358 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9359 if (size == SIZE_MAX) {
9360 io_mem_free(ctx->rings);
9365 ctx->sq_sqes = io_mem_alloc(size);
9366 if (!ctx->sq_sqes) {
9367 io_mem_free(ctx->rings);
9376 * Allocate an anonymous fd, this is what constitutes the application
9377 * visible backing of an io_uring instance. The application mmaps this
9378 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9379 * we have to tie this fd to a socket for file garbage collection purposes.
9381 static int io_uring_get_fd(struct io_ring_ctx *ctx)
9387 #if defined(CONFIG_UNIX)
9388 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9394 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9399 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9400 O_RDWR | O_CLOEXEC);
9403 ret = PTR_ERR(file);
9407 #if defined(CONFIG_UNIX)
9408 ctx->ring_sock->file = file;
9410 ret = io_uring_add_task_file(ctx, file);
9416 fd_install(fd, file);
9419 #if defined(CONFIG_UNIX)
9420 sock_release(ctx->ring_sock);
9421 ctx->ring_sock = NULL;
9426 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9427 struct io_uring_params __user *params)
9429 struct user_struct *user = NULL;
9430 struct io_ring_ctx *ctx;
9436 if (entries > IORING_MAX_ENTRIES) {
9437 if (!(p->flags & IORING_SETUP_CLAMP))
9439 entries = IORING_MAX_ENTRIES;
9443 * Use twice as many entries for the CQ ring. It's possible for the
9444 * application to drive a higher depth than the size of the SQ ring,
9445 * since the sqes are only used at submission time. This allows for
9446 * some flexibility in overcommitting a bit. If the application has
9447 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9448 * of CQ ring entries manually.
9450 p->sq_entries = roundup_pow_of_two(entries);
9451 if (p->flags & IORING_SETUP_CQSIZE) {
9453 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9454 * to a power-of-two, if it isn't already. We do NOT impose
9455 * any cq vs sq ring sizing.
9459 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9460 if (!(p->flags & IORING_SETUP_CLAMP))
9462 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9464 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9465 if (p->cq_entries < p->sq_entries)
9468 p->cq_entries = 2 * p->sq_entries;
9471 user = get_uid(current_user());
9472 limit_mem = !capable(CAP_IPC_LOCK);
9475 ret = __io_account_mem(user,
9476 ring_pages(p->sq_entries, p->cq_entries));
9483 ctx = io_ring_ctx_alloc(p);
9486 __io_unaccount_mem(user, ring_pages(p->sq_entries,
9491 ctx->compat = in_compat_syscall();
9493 ctx->creds = get_current_cred();
9495 ctx->loginuid = current->loginuid;
9496 ctx->sessionid = current->sessionid;
9498 ctx->sqo_task = get_task_struct(current);
9501 * This is just grabbed for accounting purposes. When a process exits,
9502 * the mm is exited and dropped before the files, hence we need to hang
9503 * on to this mm purely for the purposes of being able to unaccount
9504 * memory (locked/pinned vm). It's not used for anything else.
9506 mmgrab(current->mm);
9507 ctx->mm_account = current->mm;
9509 #ifdef CONFIG_BLK_CGROUP
9511 * The sq thread will belong to the original cgroup it was inited in.
9512 * If the cgroup goes offline (e.g. disabling the io controller), then
9513 * issued bios will be associated with the closest cgroup later in the
9517 ctx->sqo_blkcg_css = blkcg_css();
9518 ret = css_tryget_online(ctx->sqo_blkcg_css);
9521 /* don't init against a dying cgroup, have the user try again */
9522 ctx->sqo_blkcg_css = NULL;
9529 * Account memory _before_ installing the file descriptor. Once
9530 * the descriptor is installed, it can get closed at any time. Also
9531 * do this before hitting the general error path, as ring freeing
9532 * will un-account as well.
9534 io_account_mem(ctx, ring_pages(p->sq_entries, p->cq_entries),
9536 ctx->limit_mem = limit_mem;
9538 ret = io_allocate_scq_urings(ctx, p);
9542 ret = io_sq_offload_create(ctx, p);
9546 if (!(p->flags & IORING_SETUP_R_DISABLED))
9547 io_sq_offload_start(ctx);
9549 memset(&p->sq_off, 0, sizeof(p->sq_off));
9550 p->sq_off.head = offsetof(struct io_rings, sq.head);
9551 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9552 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9553 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9554 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9555 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9556 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9558 memset(&p->cq_off, 0, sizeof(p->cq_off));
9559 p->cq_off.head = offsetof(struct io_rings, cq.head);
9560 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9561 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9562 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9563 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9564 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9565 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9567 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9568 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9569 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9570 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
9571 IORING_FEAT_EXT_ARG;
9573 if (copy_to_user(params, p, sizeof(*p))) {
9579 * Install ring fd as the very last thing, so we don't risk someone
9580 * having closed it before we finish setup
9582 ret = io_uring_get_fd(ctx);
9586 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9589 io_ring_ctx_wait_and_kill(ctx);
9594 * Sets up an aio uring context, and returns the fd. Applications asks for a
9595 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9596 * params structure passed in.
9598 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9600 struct io_uring_params p;
9603 if (copy_from_user(&p, params, sizeof(p)))
9605 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9610 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9611 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9612 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9613 IORING_SETUP_R_DISABLED))
9616 return io_uring_create(entries, &p, params);
9619 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9620 struct io_uring_params __user *, params)
9622 return io_uring_setup(entries, params);
9625 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9627 struct io_uring_probe *p;
9631 size = struct_size(p, ops, nr_args);
9632 if (size == SIZE_MAX)
9634 p = kzalloc(size, GFP_KERNEL);
9639 if (copy_from_user(p, arg, size))
9642 if (memchr_inv(p, 0, size))
9645 p->last_op = IORING_OP_LAST - 1;
9646 if (nr_args > IORING_OP_LAST)
9647 nr_args = IORING_OP_LAST;
9649 for (i = 0; i < nr_args; i++) {
9651 if (!io_op_defs[i].not_supported)
9652 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9657 if (copy_to_user(arg, p, size))
9664 static int io_register_personality(struct io_ring_ctx *ctx)
9666 struct io_identity *id;
9669 id = kmalloc(sizeof(*id), GFP_KERNEL);
9673 io_init_identity(id);
9674 id->creds = get_current_cred();
9676 ret = idr_alloc_cyclic(&ctx->personality_idr, id, 1, USHRT_MAX, GFP_KERNEL);
9678 put_cred(id->creds);
9684 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
9686 struct io_identity *iod;
9688 iod = idr_remove(&ctx->personality_idr, id);
9690 put_cred(iod->creds);
9691 if (refcount_dec_and_test(&iod->count))
9699 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9700 unsigned int nr_args)
9702 struct io_uring_restriction *res;
9706 /* Restrictions allowed only if rings started disabled */
9707 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9710 /* We allow only a single restrictions registration */
9711 if (ctx->restrictions.registered)
9714 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9717 size = array_size(nr_args, sizeof(*res));
9718 if (size == SIZE_MAX)
9721 res = memdup_user(arg, size);
9723 return PTR_ERR(res);
9727 for (i = 0; i < nr_args; i++) {
9728 switch (res[i].opcode) {
9729 case IORING_RESTRICTION_REGISTER_OP:
9730 if (res[i].register_op >= IORING_REGISTER_LAST) {
9735 __set_bit(res[i].register_op,
9736 ctx->restrictions.register_op);
9738 case IORING_RESTRICTION_SQE_OP:
9739 if (res[i].sqe_op >= IORING_OP_LAST) {
9744 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9746 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9747 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9749 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9750 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9759 /* Reset all restrictions if an error happened */
9761 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9763 ctx->restrictions.registered = true;
9769 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9771 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9774 if (ctx->restrictions.registered)
9775 ctx->restricted = 1;
9777 ctx->flags &= ~IORING_SETUP_R_DISABLED;
9779 io_sq_offload_start(ctx);
9784 static bool io_register_op_must_quiesce(int op)
9787 case IORING_UNREGISTER_FILES:
9788 case IORING_REGISTER_FILES_UPDATE:
9789 case IORING_REGISTER_PROBE:
9790 case IORING_REGISTER_PERSONALITY:
9791 case IORING_UNREGISTER_PERSONALITY:
9798 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
9799 void __user *arg, unsigned nr_args)
9800 __releases(ctx->uring_lock)
9801 __acquires(ctx->uring_lock)
9806 * We're inside the ring mutex, if the ref is already dying, then
9807 * someone else killed the ctx or is already going through
9808 * io_uring_register().
9810 if (percpu_ref_is_dying(&ctx->refs))
9813 if (io_register_op_must_quiesce(opcode)) {
9814 percpu_ref_kill(&ctx->refs);
9817 * Drop uring mutex before waiting for references to exit. If
9818 * another thread is currently inside io_uring_enter() it might
9819 * need to grab the uring_lock to make progress. If we hold it
9820 * here across the drain wait, then we can deadlock. It's safe
9821 * to drop the mutex here, since no new references will come in
9822 * after we've killed the percpu ref.
9824 mutex_unlock(&ctx->uring_lock);
9826 ret = wait_for_completion_interruptible(&ctx->ref_comp);
9829 ret = io_run_task_work_sig();
9834 mutex_lock(&ctx->uring_lock);
9837 percpu_ref_resurrect(&ctx->refs);
9842 if (ctx->restricted) {
9843 if (opcode >= IORING_REGISTER_LAST) {
9848 if (!test_bit(opcode, ctx->restrictions.register_op)) {
9855 case IORING_REGISTER_BUFFERS:
9856 ret = io_sqe_buffer_register(ctx, arg, nr_args);
9858 case IORING_UNREGISTER_BUFFERS:
9862 ret = io_sqe_buffer_unregister(ctx);
9864 case IORING_REGISTER_FILES:
9865 ret = io_sqe_files_register(ctx, arg, nr_args);
9867 case IORING_UNREGISTER_FILES:
9871 ret = io_sqe_files_unregister(ctx);
9873 case IORING_REGISTER_FILES_UPDATE:
9874 ret = io_sqe_files_update(ctx, arg, nr_args);
9876 case IORING_REGISTER_EVENTFD:
9877 case IORING_REGISTER_EVENTFD_ASYNC:
9881 ret = io_eventfd_register(ctx, arg);
9884 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
9885 ctx->eventfd_async = 1;
9887 ctx->eventfd_async = 0;
9889 case IORING_UNREGISTER_EVENTFD:
9893 ret = io_eventfd_unregister(ctx);
9895 case IORING_REGISTER_PROBE:
9897 if (!arg || nr_args > 256)
9899 ret = io_probe(ctx, arg, nr_args);
9901 case IORING_REGISTER_PERSONALITY:
9905 ret = io_register_personality(ctx);
9907 case IORING_UNREGISTER_PERSONALITY:
9911 ret = io_unregister_personality(ctx, nr_args);
9913 case IORING_REGISTER_ENABLE_RINGS:
9917 ret = io_register_enable_rings(ctx);
9919 case IORING_REGISTER_RESTRICTIONS:
9920 ret = io_register_restrictions(ctx, arg, nr_args);
9928 if (io_register_op_must_quiesce(opcode)) {
9929 /* bring the ctx back to life */
9930 percpu_ref_reinit(&ctx->refs);
9932 reinit_completion(&ctx->ref_comp);
9937 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
9938 void __user *, arg, unsigned int, nr_args)
9940 struct io_ring_ctx *ctx;
9949 if (f.file->f_op != &io_uring_fops)
9952 ctx = f.file->private_data;
9954 mutex_lock(&ctx->uring_lock);
9955 ret = __io_uring_register(ctx, opcode, arg, nr_args);
9956 mutex_unlock(&ctx->uring_lock);
9957 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
9958 ctx->cq_ev_fd != NULL, ret);
9964 static int __init io_uring_init(void)
9966 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
9967 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
9968 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
9971 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
9972 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
9973 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
9974 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
9975 BUILD_BUG_SQE_ELEM(1, __u8, flags);
9976 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
9977 BUILD_BUG_SQE_ELEM(4, __s32, fd);
9978 BUILD_BUG_SQE_ELEM(8, __u64, off);
9979 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
9980 BUILD_BUG_SQE_ELEM(16, __u64, addr);
9981 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
9982 BUILD_BUG_SQE_ELEM(24, __u32, len);
9983 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
9984 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
9985 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
9986 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
9987 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
9988 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
9989 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
9990 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
9991 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
9992 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
9993 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
9994 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
9995 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
9996 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
9997 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
9998 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
9999 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
10000 BUILD_BUG_SQE_ELEM(42, __u16, personality);
10001 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
10003 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
10004 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
10005 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
10008 __initcall(io_uring_init);