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 inline unsigned __io_cqring_events(struct io_ring_ctx *ctx)
1698 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
1701 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1703 if (waitqueue_active(&ctx->wait))
1704 wake_up(&ctx->wait);
1705 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1706 wake_up(&ctx->sq_data->wait);
1707 if (io_should_trigger_evfd(ctx))
1708 eventfd_signal(ctx->cq_ev_fd, 1);
1711 static void io_cqring_mark_overflow(struct io_ring_ctx *ctx)
1713 if (list_empty(&ctx->cq_overflow_list)) {
1714 clear_bit(0, &ctx->sq_check_overflow);
1715 clear_bit(0, &ctx->cq_check_overflow);
1716 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
1720 /* Returns true if there are no backlogged entries after the flush */
1721 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force,
1722 struct task_struct *tsk,
1723 struct files_struct *files)
1725 struct io_rings *rings = ctx->rings;
1726 struct io_kiocb *req, *tmp;
1727 struct io_uring_cqe *cqe;
1728 unsigned long flags;
1731 if (!force && __io_cqring_events(ctx) == rings->cq_ring_entries)
1734 spin_lock_irqsave(&ctx->completion_lock, flags);
1736 list_for_each_entry_safe(req, tmp, &ctx->cq_overflow_list, compl.list) {
1737 if (!io_match_task(req, tsk, files))
1740 cqe = io_get_cqring(ctx);
1744 list_move(&req->compl.list, &list);
1746 WRITE_ONCE(cqe->user_data, req->user_data);
1747 WRITE_ONCE(cqe->res, req->result);
1748 WRITE_ONCE(cqe->flags, req->compl.cflags);
1750 ctx->cached_cq_overflow++;
1751 WRITE_ONCE(ctx->rings->cq_overflow,
1752 ctx->cached_cq_overflow);
1756 io_commit_cqring(ctx);
1757 io_cqring_mark_overflow(ctx);
1759 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1760 io_cqring_ev_posted(ctx);
1762 while (!list_empty(&list)) {
1763 req = list_first_entry(&list, struct io_kiocb, compl.list);
1764 list_del(&req->compl.list);
1771 static void __io_cqring_fill_event(struct io_kiocb *req, long res, long cflags)
1773 struct io_ring_ctx *ctx = req->ctx;
1774 struct io_uring_cqe *cqe;
1776 trace_io_uring_complete(ctx, req->user_data, res);
1779 * If we can't get a cq entry, userspace overflowed the
1780 * submission (by quite a lot). Increment the overflow count in
1783 cqe = io_get_cqring(ctx);
1785 WRITE_ONCE(cqe->user_data, req->user_data);
1786 WRITE_ONCE(cqe->res, res);
1787 WRITE_ONCE(cqe->flags, cflags);
1788 } else if (ctx->cq_overflow_flushed ||
1789 atomic_read(&req->task->io_uring->in_idle)) {
1791 * If we're in ring overflow flush mode, or in task cancel mode,
1792 * then we cannot store the request for later flushing, we need
1793 * to drop it on the floor.
1795 ctx->cached_cq_overflow++;
1796 WRITE_ONCE(ctx->rings->cq_overflow, ctx->cached_cq_overflow);
1798 if (list_empty(&ctx->cq_overflow_list)) {
1799 set_bit(0, &ctx->sq_check_overflow);
1800 set_bit(0, &ctx->cq_check_overflow);
1801 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
1805 req->compl.cflags = cflags;
1806 refcount_inc(&req->refs);
1807 list_add_tail(&req->compl.list, &ctx->cq_overflow_list);
1811 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1813 __io_cqring_fill_event(req, res, 0);
1816 static void io_cqring_add_event(struct io_kiocb *req, long res, long cflags)
1818 struct io_ring_ctx *ctx = req->ctx;
1819 unsigned long flags;
1821 spin_lock_irqsave(&ctx->completion_lock, flags);
1822 __io_cqring_fill_event(req, res, cflags);
1823 io_commit_cqring(ctx);
1824 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1826 io_cqring_ev_posted(ctx);
1829 static void io_submit_flush_completions(struct io_comp_state *cs)
1831 struct io_ring_ctx *ctx = cs->ctx;
1833 spin_lock_irq(&ctx->completion_lock);
1834 while (!list_empty(&cs->list)) {
1835 struct io_kiocb *req;
1837 req = list_first_entry(&cs->list, struct io_kiocb, compl.list);
1838 list_del(&req->compl.list);
1839 __io_cqring_fill_event(req, req->result, req->compl.cflags);
1842 * io_free_req() doesn't care about completion_lock unless one
1843 * of these flags is set. REQ_F_WORK_INITIALIZED is in the list
1844 * because of a potential deadlock with req->work.fs->lock
1846 if (req->flags & (REQ_F_FAIL_LINK|REQ_F_LINK_TIMEOUT
1847 |REQ_F_WORK_INITIALIZED)) {
1848 spin_unlock_irq(&ctx->completion_lock);
1850 spin_lock_irq(&ctx->completion_lock);
1855 io_commit_cqring(ctx);
1856 spin_unlock_irq(&ctx->completion_lock);
1858 io_cqring_ev_posted(ctx);
1862 static void __io_req_complete(struct io_kiocb *req, long res, unsigned cflags,
1863 struct io_comp_state *cs)
1866 io_cqring_add_event(req, res, cflags);
1871 req->compl.cflags = cflags;
1872 list_add_tail(&req->compl.list, &cs->list);
1874 io_submit_flush_completions(cs);
1878 static void io_req_complete(struct io_kiocb *req, long res)
1880 __io_req_complete(req, res, 0, NULL);
1883 static inline bool io_is_fallback_req(struct io_kiocb *req)
1885 return req == (struct io_kiocb *)
1886 ((unsigned long) req->ctx->fallback_req & ~1UL);
1889 static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx)
1891 struct io_kiocb *req;
1893 req = ctx->fallback_req;
1894 if (!test_and_set_bit_lock(0, (unsigned long *) &ctx->fallback_req))
1900 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx,
1901 struct io_submit_state *state)
1903 if (!state->free_reqs) {
1904 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1908 sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs));
1909 ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs);
1912 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1913 * retry single alloc to be on the safe side.
1915 if (unlikely(ret <= 0)) {
1916 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1917 if (!state->reqs[0])
1921 state->free_reqs = ret;
1925 return state->reqs[state->free_reqs];
1927 return io_get_fallback_req(ctx);
1930 static inline void io_put_file(struct io_kiocb *req, struct file *file,
1937 static void io_dismantle_req(struct io_kiocb *req)
1941 if (req->async_data)
1942 kfree(req->async_data);
1944 io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
1945 if (req->fixed_file_refs)
1946 percpu_ref_put(req->fixed_file_refs);
1947 io_req_clean_work(req);
1950 static void __io_free_req(struct io_kiocb *req)
1952 struct io_uring_task *tctx = req->task->io_uring;
1953 struct io_ring_ctx *ctx = req->ctx;
1955 io_dismantle_req(req);
1957 percpu_counter_dec(&tctx->inflight);
1958 if (atomic_read(&tctx->in_idle))
1959 wake_up(&tctx->wait);
1960 put_task_struct(req->task);
1962 if (likely(!io_is_fallback_req(req)))
1963 kmem_cache_free(req_cachep, req);
1965 clear_bit_unlock(0, (unsigned long *) &ctx->fallback_req);
1966 percpu_ref_put(&ctx->refs);
1969 static inline void io_remove_next_linked(struct io_kiocb *req)
1971 struct io_kiocb *nxt = req->link;
1973 req->link = nxt->link;
1977 static void io_kill_linked_timeout(struct io_kiocb *req)
1979 struct io_ring_ctx *ctx = req->ctx;
1980 struct io_kiocb *link;
1981 bool cancelled = false;
1982 unsigned long flags;
1984 spin_lock_irqsave(&ctx->completion_lock, flags);
1988 * Can happen if a linked timeout fired and link had been like
1989 * req -> link t-out -> link t-out [-> ...]
1991 if (link && (link->flags & REQ_F_LTIMEOUT_ACTIVE)) {
1992 struct io_timeout_data *io = link->async_data;
1995 io_remove_next_linked(req);
1996 link->timeout.head = NULL;
1997 ret = hrtimer_try_to_cancel(&io->timer);
1999 io_cqring_fill_event(link, -ECANCELED);
2000 io_commit_cqring(ctx);
2004 req->flags &= ~REQ_F_LINK_TIMEOUT;
2005 spin_unlock_irqrestore(&ctx->completion_lock, flags);
2008 io_cqring_ev_posted(ctx);
2014 static void io_fail_links(struct io_kiocb *req)
2016 struct io_kiocb *link, *nxt;
2017 struct io_ring_ctx *ctx = req->ctx;
2018 unsigned long flags;
2020 spin_lock_irqsave(&ctx->completion_lock, flags);
2028 trace_io_uring_fail_link(req, link);
2029 io_cqring_fill_event(link, -ECANCELED);
2032 * It's ok to free under spinlock as they're not linked anymore,
2033 * but avoid REQ_F_WORK_INITIALIZED because it may deadlock on
2036 if (link->flags & REQ_F_WORK_INITIALIZED)
2037 io_put_req_deferred(link, 2);
2039 io_double_put_req(link);
2042 io_commit_cqring(ctx);
2043 spin_unlock_irqrestore(&ctx->completion_lock, flags);
2045 io_cqring_ev_posted(ctx);
2048 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
2050 if (req->flags & REQ_F_LINK_TIMEOUT)
2051 io_kill_linked_timeout(req);
2054 * If LINK is set, we have dependent requests in this chain. If we
2055 * didn't fail this request, queue the first one up, moving any other
2056 * dependencies to the next request. In case of failure, fail the rest
2059 if (likely(!(req->flags & REQ_F_FAIL_LINK))) {
2060 struct io_kiocb *nxt = req->link;
2069 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
2071 if (likely(!(req->link) && !(req->flags & REQ_F_LINK_TIMEOUT)))
2073 return __io_req_find_next(req);
2076 static int io_req_task_work_add(struct io_kiocb *req)
2078 struct task_struct *tsk = req->task;
2079 struct io_ring_ctx *ctx = req->ctx;
2080 enum task_work_notify_mode notify;
2083 if (tsk->flags & PF_EXITING)
2087 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
2088 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
2089 * processing task_work. There's no reliable way to tell if TWA_RESUME
2093 if (!(ctx->flags & IORING_SETUP_SQPOLL))
2094 notify = TWA_SIGNAL;
2096 ret = task_work_add(tsk, &req->task_work, notify);
2098 wake_up_process(tsk);
2103 static void __io_req_task_cancel(struct io_kiocb *req, int error)
2105 struct io_ring_ctx *ctx = req->ctx;
2107 spin_lock_irq(&ctx->completion_lock);
2108 io_cqring_fill_event(req, error);
2109 io_commit_cqring(ctx);
2110 spin_unlock_irq(&ctx->completion_lock);
2112 io_cqring_ev_posted(ctx);
2113 req_set_fail_links(req);
2114 io_double_put_req(req);
2117 static void io_req_task_cancel(struct callback_head *cb)
2119 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2120 struct io_ring_ctx *ctx = req->ctx;
2122 __io_req_task_cancel(req, -ECANCELED);
2123 percpu_ref_put(&ctx->refs);
2126 static void __io_req_task_submit(struct io_kiocb *req)
2128 struct io_ring_ctx *ctx = req->ctx;
2130 if (!__io_sq_thread_acquire_mm(ctx) &&
2131 !__io_sq_thread_acquire_files(ctx)) {
2132 mutex_lock(&ctx->uring_lock);
2133 __io_queue_sqe(req, NULL);
2134 mutex_unlock(&ctx->uring_lock);
2136 __io_req_task_cancel(req, -EFAULT);
2140 static void io_req_task_submit(struct callback_head *cb)
2142 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2143 struct io_ring_ctx *ctx = req->ctx;
2145 __io_req_task_submit(req);
2146 percpu_ref_put(&ctx->refs);
2149 static void io_req_task_queue(struct io_kiocb *req)
2153 init_task_work(&req->task_work, io_req_task_submit);
2154 percpu_ref_get(&req->ctx->refs);
2156 ret = io_req_task_work_add(req);
2157 if (unlikely(ret)) {
2158 struct task_struct *tsk;
2160 init_task_work(&req->task_work, io_req_task_cancel);
2161 tsk = io_wq_get_task(req->ctx->io_wq);
2162 task_work_add(tsk, &req->task_work, TWA_NONE);
2163 wake_up_process(tsk);
2167 static inline void io_queue_next(struct io_kiocb *req)
2169 struct io_kiocb *nxt = io_req_find_next(req);
2172 io_req_task_queue(nxt);
2175 static void io_free_req(struct io_kiocb *req)
2182 void *reqs[IO_IOPOLL_BATCH];
2185 struct task_struct *task;
2189 static inline void io_init_req_batch(struct req_batch *rb)
2196 static void __io_req_free_batch_flush(struct io_ring_ctx *ctx,
2197 struct req_batch *rb)
2199 kmem_cache_free_bulk(req_cachep, rb->to_free, rb->reqs);
2200 percpu_ref_put_many(&ctx->refs, rb->to_free);
2204 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2205 struct req_batch *rb)
2208 __io_req_free_batch_flush(ctx, rb);
2210 struct io_uring_task *tctx = rb->task->io_uring;
2212 percpu_counter_sub(&tctx->inflight, rb->task_refs);
2213 put_task_struct_many(rb->task, rb->task_refs);
2218 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req)
2220 if (unlikely(io_is_fallback_req(req))) {
2226 if (req->task != rb->task) {
2228 struct io_uring_task *tctx = rb->task->io_uring;
2230 percpu_counter_sub(&tctx->inflight, rb->task_refs);
2231 put_task_struct_many(rb->task, rb->task_refs);
2233 rb->task = req->task;
2238 io_dismantle_req(req);
2239 rb->reqs[rb->to_free++] = req;
2240 if (unlikely(rb->to_free == ARRAY_SIZE(rb->reqs)))
2241 __io_req_free_batch_flush(req->ctx, rb);
2245 * Drop reference to request, return next in chain (if there is one) if this
2246 * was the last reference to this request.
2248 static struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2250 struct io_kiocb *nxt = NULL;
2252 if (refcount_dec_and_test(&req->refs)) {
2253 nxt = io_req_find_next(req);
2259 static void io_put_req(struct io_kiocb *req)
2261 if (refcount_dec_and_test(&req->refs))
2265 static void io_put_req_deferred_cb(struct callback_head *cb)
2267 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2272 static void io_free_req_deferred(struct io_kiocb *req)
2276 init_task_work(&req->task_work, io_put_req_deferred_cb);
2277 ret = io_req_task_work_add(req);
2278 if (unlikely(ret)) {
2279 struct task_struct *tsk;
2281 tsk = io_wq_get_task(req->ctx->io_wq);
2282 task_work_add(tsk, &req->task_work, TWA_NONE);
2283 wake_up_process(tsk);
2287 static inline void io_put_req_deferred(struct io_kiocb *req, int refs)
2289 if (refcount_sub_and_test(refs, &req->refs))
2290 io_free_req_deferred(req);
2293 static struct io_wq_work *io_steal_work(struct io_kiocb *req)
2295 struct io_kiocb *nxt;
2298 * A ref is owned by io-wq in which context we're. So, if that's the
2299 * last one, it's safe to steal next work. False negatives are Ok,
2300 * it just will be re-punted async in io_put_work()
2302 if (refcount_read(&req->refs) != 1)
2305 nxt = io_req_find_next(req);
2306 return nxt ? &nxt->work : NULL;
2309 static void io_double_put_req(struct io_kiocb *req)
2311 /* drop both submit and complete references */
2312 if (refcount_sub_and_test(2, &req->refs))
2316 static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush)
2318 if (test_bit(0, &ctx->cq_check_overflow)) {
2320 * noflush == true is from the waitqueue handler, just ensure
2321 * we wake up the task, and the next invocation will flush the
2322 * entries. We cannot safely to it from here.
2327 io_cqring_overflow_flush(ctx, false, NULL, NULL);
2330 /* See comment at the top of this file */
2332 return __io_cqring_events(ctx);
2335 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2337 struct io_rings *rings = ctx->rings;
2339 /* make sure SQ entry isn't read before tail */
2340 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2343 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2345 unsigned int cflags;
2347 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2348 cflags |= IORING_CQE_F_BUFFER;
2349 req->flags &= ~REQ_F_BUFFER_SELECTED;
2354 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2356 struct io_buffer *kbuf;
2358 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2359 return io_put_kbuf(req, kbuf);
2362 static inline bool io_run_task_work(void)
2365 * Not safe to run on exiting task, and the task_work handling will
2366 * not add work to such a task.
2368 if (unlikely(current->flags & PF_EXITING))
2370 if (current->task_works) {
2371 __set_current_state(TASK_RUNNING);
2379 static void io_iopoll_queue(struct list_head *again)
2381 struct io_kiocb *req;
2384 req = list_first_entry(again, struct io_kiocb, inflight_entry);
2385 list_del(&req->inflight_entry);
2386 __io_complete_rw(req, -EAGAIN, 0, NULL);
2387 } while (!list_empty(again));
2391 * Find and free completed poll iocbs
2393 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2394 struct list_head *done)
2396 struct req_batch rb;
2397 struct io_kiocb *req;
2400 /* order with ->result store in io_complete_rw_iopoll() */
2403 io_init_req_batch(&rb);
2404 while (!list_empty(done)) {
2407 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2408 if (READ_ONCE(req->result) == -EAGAIN) {
2410 req->iopoll_completed = 0;
2411 list_move_tail(&req->inflight_entry, &again);
2414 list_del(&req->inflight_entry);
2416 if (req->flags & REQ_F_BUFFER_SELECTED)
2417 cflags = io_put_rw_kbuf(req);
2419 __io_cqring_fill_event(req, req->result, cflags);
2422 if (refcount_dec_and_test(&req->refs))
2423 io_req_free_batch(&rb, req);
2426 io_commit_cqring(ctx);
2427 if (ctx->flags & IORING_SETUP_SQPOLL)
2428 io_cqring_ev_posted(ctx);
2429 io_req_free_batch_finish(ctx, &rb);
2431 if (!list_empty(&again))
2432 io_iopoll_queue(&again);
2435 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2438 struct io_kiocb *req, *tmp;
2444 * Only spin for completions if we don't have multiple devices hanging
2445 * off our complete list, and we're under the requested amount.
2447 spin = !ctx->poll_multi_file && *nr_events < min;
2450 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2451 struct kiocb *kiocb = &req->rw.kiocb;
2454 * Move completed and retryable entries to our local lists.
2455 * If we find a request that requires polling, break out
2456 * and complete those lists first, if we have entries there.
2458 if (READ_ONCE(req->iopoll_completed)) {
2459 list_move_tail(&req->inflight_entry, &done);
2462 if (!list_empty(&done))
2465 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2469 /* iopoll may have completed current req */
2470 if (READ_ONCE(req->iopoll_completed))
2471 list_move_tail(&req->inflight_entry, &done);
2478 if (!list_empty(&done))
2479 io_iopoll_complete(ctx, nr_events, &done);
2485 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2486 * non-spinning poll check - we'll still enter the driver poll loop, but only
2487 * as a non-spinning completion check.
2489 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
2492 while (!list_empty(&ctx->iopoll_list) && !need_resched()) {
2495 ret = io_do_iopoll(ctx, nr_events, min);
2498 if (*nr_events >= min)
2506 * We can't just wait for polled events to come to us, we have to actively
2507 * find and complete them.
2509 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2511 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2514 mutex_lock(&ctx->uring_lock);
2515 while (!list_empty(&ctx->iopoll_list)) {
2516 unsigned int nr_events = 0;
2518 io_do_iopoll(ctx, &nr_events, 0);
2520 /* let it sleep and repeat later if can't complete a request */
2524 * Ensure we allow local-to-the-cpu processing to take place,
2525 * in this case we need to ensure that we reap all events.
2526 * Also let task_work, etc. to progress by releasing the mutex
2528 if (need_resched()) {
2529 mutex_unlock(&ctx->uring_lock);
2531 mutex_lock(&ctx->uring_lock);
2534 mutex_unlock(&ctx->uring_lock);
2537 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2539 unsigned int nr_events = 0;
2540 int iters = 0, ret = 0;
2543 * We disallow the app entering submit/complete with polling, but we
2544 * still need to lock the ring to prevent racing with polled issue
2545 * that got punted to a workqueue.
2547 mutex_lock(&ctx->uring_lock);
2550 * Don't enter poll loop if we already have events pending.
2551 * If we do, we can potentially be spinning for commands that
2552 * already triggered a CQE (eg in error).
2554 if (io_cqring_events(ctx, false))
2558 * If a submit got punted to a workqueue, we can have the
2559 * application entering polling for a command before it gets
2560 * issued. That app will hold the uring_lock for the duration
2561 * of the poll right here, so we need to take a breather every
2562 * now and then to ensure that the issue has a chance to add
2563 * the poll to the issued list. Otherwise we can spin here
2564 * forever, while the workqueue is stuck trying to acquire the
2567 if (!(++iters & 7)) {
2568 mutex_unlock(&ctx->uring_lock);
2570 mutex_lock(&ctx->uring_lock);
2573 ret = io_iopoll_getevents(ctx, &nr_events, min);
2577 } while (min && !nr_events && !need_resched());
2579 mutex_unlock(&ctx->uring_lock);
2583 static void kiocb_end_write(struct io_kiocb *req)
2586 * Tell lockdep we inherited freeze protection from submission
2589 if (req->flags & REQ_F_ISREG) {
2590 struct inode *inode = file_inode(req->file);
2592 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
2594 file_end_write(req->file);
2597 static void io_complete_rw_common(struct kiocb *kiocb, long res,
2598 struct io_comp_state *cs)
2600 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2603 if (kiocb->ki_flags & IOCB_WRITE)
2604 kiocb_end_write(req);
2606 if (res != req->result)
2607 req_set_fail_links(req);
2608 if (req->flags & REQ_F_BUFFER_SELECTED)
2609 cflags = io_put_rw_kbuf(req);
2610 __io_req_complete(req, res, cflags, cs);
2614 static bool io_resubmit_prep(struct io_kiocb *req, int error)
2616 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2617 ssize_t ret = -ECANCELED;
2618 struct iov_iter iter;
2626 switch (req->opcode) {
2627 case IORING_OP_READV:
2628 case IORING_OP_READ_FIXED:
2629 case IORING_OP_READ:
2632 case IORING_OP_WRITEV:
2633 case IORING_OP_WRITE_FIXED:
2634 case IORING_OP_WRITE:
2638 printk_once(KERN_WARNING "io_uring: bad opcode in resubmit %d\n",
2643 if (!req->async_data) {
2644 ret = io_import_iovec(rw, req, &iovec, &iter, false);
2647 ret = io_setup_async_rw(req, iovec, inline_vecs, &iter, false);
2655 req_set_fail_links(req);
2660 static bool io_rw_reissue(struct io_kiocb *req, long res)
2663 umode_t mode = file_inode(req->file)->i_mode;
2666 if (!S_ISBLK(mode) && !S_ISREG(mode))
2668 if ((res != -EAGAIN && res != -EOPNOTSUPP) || io_wq_current_is_worker())
2671 ret = io_sq_thread_acquire_mm_files(req->ctx, req);
2673 if (io_resubmit_prep(req, ret)) {
2674 refcount_inc(&req->refs);
2675 io_queue_async_work(req);
2683 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2684 struct io_comp_state *cs)
2686 if (!io_rw_reissue(req, res))
2687 io_complete_rw_common(&req->rw.kiocb, res, cs);
2690 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2692 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2694 __io_complete_rw(req, res, res2, NULL);
2697 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2699 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2701 if (kiocb->ki_flags & IOCB_WRITE)
2702 kiocb_end_write(req);
2704 if (res != -EAGAIN && res != req->result)
2705 req_set_fail_links(req);
2707 WRITE_ONCE(req->result, res);
2708 /* order with io_poll_complete() checking ->result */
2710 WRITE_ONCE(req->iopoll_completed, 1);
2714 * After the iocb has been issued, it's safe to be found on the poll list.
2715 * Adding the kiocb to the list AFTER submission ensures that we don't
2716 * find it from a io_iopoll_getevents() thread before the issuer is done
2717 * accessing the kiocb cookie.
2719 static void io_iopoll_req_issued(struct io_kiocb *req, bool in_async)
2721 struct io_ring_ctx *ctx = req->ctx;
2724 * Track whether we have multiple files in our lists. This will impact
2725 * how we do polling eventually, not spinning if we're on potentially
2726 * different devices.
2728 if (list_empty(&ctx->iopoll_list)) {
2729 ctx->poll_multi_file = false;
2730 } else if (!ctx->poll_multi_file) {
2731 struct io_kiocb *list_req;
2733 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2735 if (list_req->file != req->file)
2736 ctx->poll_multi_file = true;
2740 * For fast devices, IO may have already completed. If it has, add
2741 * it to the front so we find it first.
2743 if (READ_ONCE(req->iopoll_completed))
2744 list_add(&req->inflight_entry, &ctx->iopoll_list);
2746 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2749 * If IORING_SETUP_SQPOLL is enabled, sqes are either handled in sq thread
2750 * task context or in io worker task context. If current task context is
2751 * sq thread, we don't need to check whether should wake up sq thread.
2753 if (in_async && (ctx->flags & IORING_SETUP_SQPOLL) &&
2754 wq_has_sleeper(&ctx->sq_data->wait))
2755 wake_up(&ctx->sq_data->wait);
2758 static inline void __io_state_file_put(struct io_submit_state *state)
2760 fput_many(state->file, state->file_refs);
2761 state->file_refs = 0;
2764 static inline void io_state_file_put(struct io_submit_state *state)
2766 if (state->file_refs)
2767 __io_state_file_put(state);
2771 * Get as many references to a file as we have IOs left in this submission,
2772 * assuming most submissions are for one file, or at least that each file
2773 * has more than one submission.
2775 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2780 if (state->file_refs) {
2781 if (state->fd == fd) {
2785 __io_state_file_put(state);
2787 state->file = fget_many(fd, state->ios_left);
2788 if (unlikely(!state->file))
2792 state->file_refs = state->ios_left - 1;
2796 static bool io_bdev_nowait(struct block_device *bdev)
2798 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2802 * If we tracked the file through the SCM inflight mechanism, we could support
2803 * any file. For now, just ensure that anything potentially problematic is done
2806 static bool io_file_supports_async(struct file *file, int rw)
2808 umode_t mode = file_inode(file)->i_mode;
2810 if (S_ISBLK(mode)) {
2811 if (IS_ENABLED(CONFIG_BLOCK) &&
2812 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2816 if (S_ISCHR(mode) || S_ISSOCK(mode))
2818 if (S_ISREG(mode)) {
2819 if (IS_ENABLED(CONFIG_BLOCK) &&
2820 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2821 file->f_op != &io_uring_fops)
2826 /* any ->read/write should understand O_NONBLOCK */
2827 if (file->f_flags & O_NONBLOCK)
2830 if (!(file->f_mode & FMODE_NOWAIT))
2834 return file->f_op->read_iter != NULL;
2836 return file->f_op->write_iter != NULL;
2839 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2841 struct io_ring_ctx *ctx = req->ctx;
2842 struct kiocb *kiocb = &req->rw.kiocb;
2846 if (S_ISREG(file_inode(req->file)->i_mode))
2847 req->flags |= REQ_F_ISREG;
2849 kiocb->ki_pos = READ_ONCE(sqe->off);
2850 if (kiocb->ki_pos == -1 && !(req->file->f_mode & FMODE_STREAM)) {
2851 req->flags |= REQ_F_CUR_POS;
2852 kiocb->ki_pos = req->file->f_pos;
2854 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2855 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2856 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2860 ioprio = READ_ONCE(sqe->ioprio);
2862 ret = ioprio_check_cap(ioprio);
2866 kiocb->ki_ioprio = ioprio;
2868 kiocb->ki_ioprio = get_current_ioprio();
2870 /* don't allow async punt if RWF_NOWAIT was requested */
2871 if (kiocb->ki_flags & IOCB_NOWAIT)
2872 req->flags |= REQ_F_NOWAIT;
2874 if (ctx->flags & IORING_SETUP_IOPOLL) {
2875 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2876 !kiocb->ki_filp->f_op->iopoll)
2879 kiocb->ki_flags |= IOCB_HIPRI;
2880 kiocb->ki_complete = io_complete_rw_iopoll;
2881 req->iopoll_completed = 0;
2883 if (kiocb->ki_flags & IOCB_HIPRI)
2885 kiocb->ki_complete = io_complete_rw;
2888 req->rw.addr = READ_ONCE(sqe->addr);
2889 req->rw.len = READ_ONCE(sqe->len);
2890 req->buf_index = READ_ONCE(sqe->buf_index);
2894 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2900 case -ERESTARTNOINTR:
2901 case -ERESTARTNOHAND:
2902 case -ERESTART_RESTARTBLOCK:
2904 * We can't just restart the syscall, since previously
2905 * submitted sqes may already be in progress. Just fail this
2911 kiocb->ki_complete(kiocb, ret, 0);
2915 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2916 struct io_comp_state *cs)
2918 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2919 struct io_async_rw *io = req->async_data;
2921 /* add previously done IO, if any */
2922 if (io && io->bytes_done > 0) {
2924 ret = io->bytes_done;
2926 ret += io->bytes_done;
2929 if (req->flags & REQ_F_CUR_POS)
2930 req->file->f_pos = kiocb->ki_pos;
2931 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2932 __io_complete_rw(req, ret, 0, cs);
2934 io_rw_done(kiocb, ret);
2937 static ssize_t io_import_fixed(struct io_kiocb *req, int rw,
2938 struct iov_iter *iter)
2940 struct io_ring_ctx *ctx = req->ctx;
2941 size_t len = req->rw.len;
2942 struct io_mapped_ubuf *imu;
2943 u16 index, buf_index = req->buf_index;
2947 if (unlikely(buf_index >= ctx->nr_user_bufs))
2949 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2950 imu = &ctx->user_bufs[index];
2951 buf_addr = req->rw.addr;
2954 if (buf_addr + len < buf_addr)
2956 /* not inside the mapped region */
2957 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2961 * May not be a start of buffer, set size appropriately
2962 * and advance us to the beginning.
2964 offset = buf_addr - imu->ubuf;
2965 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2969 * Don't use iov_iter_advance() here, as it's really slow for
2970 * using the latter parts of a big fixed buffer - it iterates
2971 * over each segment manually. We can cheat a bit here, because
2974 * 1) it's a BVEC iter, we set it up
2975 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2976 * first and last bvec
2978 * So just find our index, and adjust the iterator afterwards.
2979 * If the offset is within the first bvec (or the whole first
2980 * bvec, just use iov_iter_advance(). This makes it easier
2981 * since we can just skip the first segment, which may not
2982 * be PAGE_SIZE aligned.
2984 const struct bio_vec *bvec = imu->bvec;
2986 if (offset <= bvec->bv_len) {
2987 iov_iter_advance(iter, offset);
2989 unsigned long seg_skip;
2991 /* skip first vec */
2992 offset -= bvec->bv_len;
2993 seg_skip = 1 + (offset >> PAGE_SHIFT);
2995 iter->bvec = bvec + seg_skip;
2996 iter->nr_segs -= seg_skip;
2997 iter->count -= bvec->bv_len + offset;
2998 iter->iov_offset = offset & ~PAGE_MASK;
3005 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
3008 mutex_unlock(&ctx->uring_lock);
3011 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
3014 * "Normal" inline submissions always hold the uring_lock, since we
3015 * grab it from the system call. Same is true for the SQPOLL offload.
3016 * The only exception is when we've detached the request and issue it
3017 * from an async worker thread, grab the lock for that case.
3020 mutex_lock(&ctx->uring_lock);
3023 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
3024 int bgid, struct io_buffer *kbuf,
3027 struct io_buffer *head;
3029 if (req->flags & REQ_F_BUFFER_SELECTED)
3032 io_ring_submit_lock(req->ctx, needs_lock);
3034 lockdep_assert_held(&req->ctx->uring_lock);
3036 head = idr_find(&req->ctx->io_buffer_idr, bgid);
3038 if (!list_empty(&head->list)) {
3039 kbuf = list_last_entry(&head->list, struct io_buffer,
3041 list_del(&kbuf->list);
3044 idr_remove(&req->ctx->io_buffer_idr, bgid);
3046 if (*len > kbuf->len)
3049 kbuf = ERR_PTR(-ENOBUFS);
3052 io_ring_submit_unlock(req->ctx, needs_lock);
3057 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
3060 struct io_buffer *kbuf;
3063 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3064 bgid = req->buf_index;
3065 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
3068 req->rw.addr = (u64) (unsigned long) kbuf;
3069 req->flags |= REQ_F_BUFFER_SELECTED;
3070 return u64_to_user_ptr(kbuf->addr);
3073 #ifdef CONFIG_COMPAT
3074 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
3077 struct compat_iovec __user *uiov;
3078 compat_ssize_t clen;
3082 uiov = u64_to_user_ptr(req->rw.addr);
3083 if (!access_ok(uiov, sizeof(*uiov)))
3085 if (__get_user(clen, &uiov->iov_len))
3091 buf = io_rw_buffer_select(req, &len, needs_lock);
3093 return PTR_ERR(buf);
3094 iov[0].iov_base = buf;
3095 iov[0].iov_len = (compat_size_t) len;
3100 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3103 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
3107 if (copy_from_user(iov, uiov, sizeof(*uiov)))
3110 len = iov[0].iov_len;
3113 buf = io_rw_buffer_select(req, &len, needs_lock);
3115 return PTR_ERR(buf);
3116 iov[0].iov_base = buf;
3117 iov[0].iov_len = len;
3121 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3124 if (req->flags & REQ_F_BUFFER_SELECTED) {
3125 struct io_buffer *kbuf;
3127 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3128 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
3129 iov[0].iov_len = kbuf->len;
3134 else if (req->rw.len > 1)
3137 #ifdef CONFIG_COMPAT
3138 if (req->ctx->compat)
3139 return io_compat_import(req, iov, needs_lock);
3142 return __io_iov_buffer_select(req, iov, needs_lock);
3145 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
3146 struct iovec **iovec, struct iov_iter *iter,
3149 void __user *buf = u64_to_user_ptr(req->rw.addr);
3150 size_t sqe_len = req->rw.len;
3154 opcode = req->opcode;
3155 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3157 return io_import_fixed(req, rw, iter);
3160 /* buffer index only valid with fixed read/write, or buffer select */
3161 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3164 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3165 if (req->flags & REQ_F_BUFFER_SELECT) {
3166 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3168 return PTR_ERR(buf);
3169 req->rw.len = sqe_len;
3172 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3177 if (req->flags & REQ_F_BUFFER_SELECT) {
3178 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3180 ret = (*iovec)->iov_len;
3181 iov_iter_init(iter, rw, *iovec, 1, ret);
3187 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3191 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3193 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3197 * For files that don't have ->read_iter() and ->write_iter(), handle them
3198 * by looping over ->read() or ->write() manually.
3200 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3202 struct kiocb *kiocb = &req->rw.kiocb;
3203 struct file *file = req->file;
3207 * Don't support polled IO through this interface, and we can't
3208 * support non-blocking either. For the latter, this just causes
3209 * the kiocb to be handled from an async context.
3211 if (kiocb->ki_flags & IOCB_HIPRI)
3213 if (kiocb->ki_flags & IOCB_NOWAIT)
3216 while (iov_iter_count(iter)) {
3220 if (!iov_iter_is_bvec(iter)) {
3221 iovec = iov_iter_iovec(iter);
3223 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3224 iovec.iov_len = req->rw.len;
3228 nr = file->f_op->read(file, iovec.iov_base,
3229 iovec.iov_len, io_kiocb_ppos(kiocb));
3231 nr = file->f_op->write(file, iovec.iov_base,
3232 iovec.iov_len, io_kiocb_ppos(kiocb));
3241 if (nr != iovec.iov_len)
3245 iov_iter_advance(iter, nr);
3251 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3252 const struct iovec *fast_iov, struct iov_iter *iter)
3254 struct io_async_rw *rw = req->async_data;
3256 memcpy(&rw->iter, iter, sizeof(*iter));
3257 rw->free_iovec = iovec;
3259 /* can only be fixed buffers, no need to do anything */
3260 if (iov_iter_is_bvec(iter))
3263 unsigned iov_off = 0;
3265 rw->iter.iov = rw->fast_iov;
3266 if (iter->iov != fast_iov) {
3267 iov_off = iter->iov - fast_iov;
3268 rw->iter.iov += iov_off;
3270 if (rw->fast_iov != fast_iov)
3271 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3272 sizeof(struct iovec) * iter->nr_segs);
3274 req->flags |= REQ_F_NEED_CLEANUP;
3278 static inline int __io_alloc_async_data(struct io_kiocb *req)
3280 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3281 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3282 return req->async_data == NULL;
3285 static int io_alloc_async_data(struct io_kiocb *req)
3287 if (!io_op_defs[req->opcode].needs_async_data)
3290 return __io_alloc_async_data(req);
3293 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3294 const struct iovec *fast_iov,
3295 struct iov_iter *iter, bool force)
3297 if (!force && !io_op_defs[req->opcode].needs_async_data)
3299 if (!req->async_data) {
3300 if (__io_alloc_async_data(req))
3303 io_req_map_rw(req, iovec, fast_iov, iter);
3308 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3310 struct io_async_rw *iorw = req->async_data;
3311 struct iovec *iov = iorw->fast_iov;
3314 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3315 if (unlikely(ret < 0))
3318 iorw->bytes_done = 0;
3319 iorw->free_iovec = iov;
3321 req->flags |= REQ_F_NEED_CLEANUP;
3325 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3329 ret = io_prep_rw(req, sqe);
3333 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3336 /* either don't need iovec imported or already have it */
3337 if (!req->async_data)
3339 return io_rw_prep_async(req, READ);
3343 * This is our waitqueue callback handler, registered through lock_page_async()
3344 * when we initially tried to do the IO with the iocb armed our waitqueue.
3345 * This gets called when the page is unlocked, and we generally expect that to
3346 * happen when the page IO is completed and the page is now uptodate. This will
3347 * queue a task_work based retry of the operation, attempting to copy the data
3348 * again. If the latter fails because the page was NOT uptodate, then we will
3349 * do a thread based blocking retry of the operation. That's the unexpected
3352 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3353 int sync, void *arg)
3355 struct wait_page_queue *wpq;
3356 struct io_kiocb *req = wait->private;
3357 struct wait_page_key *key = arg;
3360 wpq = container_of(wait, struct wait_page_queue, wait);
3362 if (!wake_page_match(wpq, key))
3365 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3366 list_del_init(&wait->entry);
3368 init_task_work(&req->task_work, io_req_task_submit);
3369 percpu_ref_get(&req->ctx->refs);
3371 /* submit ref gets dropped, acquire a new one */
3372 refcount_inc(&req->refs);
3373 ret = io_req_task_work_add(req);
3374 if (unlikely(ret)) {
3375 struct task_struct *tsk;
3377 /* queue just for cancelation */
3378 init_task_work(&req->task_work, io_req_task_cancel);
3379 tsk = io_wq_get_task(req->ctx->io_wq);
3380 task_work_add(tsk, &req->task_work, TWA_NONE);
3381 wake_up_process(tsk);
3387 * This controls whether a given IO request should be armed for async page
3388 * based retry. If we return false here, the request is handed to the async
3389 * worker threads for retry. If we're doing buffered reads on a regular file,
3390 * we prepare a private wait_page_queue entry and retry the operation. This
3391 * will either succeed because the page is now uptodate and unlocked, or it
3392 * will register a callback when the page is unlocked at IO completion. Through
3393 * that callback, io_uring uses task_work to setup a retry of the operation.
3394 * That retry will attempt the buffered read again. The retry will generally
3395 * succeed, or in rare cases where it fails, we then fall back to using the
3396 * async worker threads for a blocking retry.
3398 static bool io_rw_should_retry(struct io_kiocb *req)
3400 struct io_async_rw *rw = req->async_data;
3401 struct wait_page_queue *wait = &rw->wpq;
3402 struct kiocb *kiocb = &req->rw.kiocb;
3404 /* never retry for NOWAIT, we just complete with -EAGAIN */
3405 if (req->flags & REQ_F_NOWAIT)
3408 /* Only for buffered IO */
3409 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3413 * just use poll if we can, and don't attempt if the fs doesn't
3414 * support callback based unlocks
3416 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3419 wait->wait.func = io_async_buf_func;
3420 wait->wait.private = req;
3421 wait->wait.flags = 0;
3422 INIT_LIST_HEAD(&wait->wait.entry);
3423 kiocb->ki_flags |= IOCB_WAITQ;
3424 kiocb->ki_flags &= ~IOCB_NOWAIT;
3425 kiocb->ki_waitq = wait;
3429 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3431 if (req->file->f_op->read_iter)
3432 return call_read_iter(req->file, &req->rw.kiocb, iter);
3433 else if (req->file->f_op->read)
3434 return loop_rw_iter(READ, req, iter);
3439 static int io_read(struct io_kiocb *req, bool force_nonblock,
3440 struct io_comp_state *cs)
3442 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3443 struct kiocb *kiocb = &req->rw.kiocb;
3444 struct iov_iter __iter, *iter = &__iter;
3445 struct io_async_rw *rw = req->async_data;
3446 ssize_t io_size, ret, ret2;
3453 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3457 io_size = iov_iter_count(iter);
3458 req->result = io_size;
3461 /* Ensure we clear previously set non-block flag */
3462 if (!force_nonblock)
3463 kiocb->ki_flags &= ~IOCB_NOWAIT;
3465 kiocb->ki_flags |= IOCB_NOWAIT;
3468 /* If the file doesn't support async, just async punt */
3469 no_async = force_nonblock && !io_file_supports_async(req->file, READ);
3473 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size);
3477 ret = io_iter_do_read(req, iter);
3481 } else if (ret == -EIOCBQUEUED) {
3484 } else if (ret == -EAGAIN) {
3485 /* IOPOLL retry should happen for io-wq threads */
3486 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3488 /* no retry on NONBLOCK marked file */
3489 if (req->file->f_flags & O_NONBLOCK)
3491 /* some cases will consume bytes even on error returns */
3492 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3495 } else if (ret < 0) {
3496 /* make sure -ERESTARTSYS -> -EINTR is done */
3500 /* read it all, or we did blocking attempt. no retry. */
3501 if (!iov_iter_count(iter) || !force_nonblock ||
3502 (req->file->f_flags & O_NONBLOCK))
3507 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3514 rw = req->async_data;
3515 /* it's copied and will be cleaned with ->io */
3517 /* now use our persistent iterator, if we aren't already */
3520 rw->bytes_done += ret;
3521 /* if we can retry, do so with the callbacks armed */
3522 if (!io_rw_should_retry(req)) {
3523 kiocb->ki_flags &= ~IOCB_WAITQ;
3528 * Now retry read with the IOCB_WAITQ parts set in the iocb. If we
3529 * get -EIOCBQUEUED, then we'll get a notification when the desired
3530 * page gets unlocked. We can also get a partial read here, and if we
3531 * do, then just retry at the new offset.
3533 ret = io_iter_do_read(req, iter);
3534 if (ret == -EIOCBQUEUED) {
3537 } else if (ret > 0 && ret < io_size) {
3538 /* we got some bytes, but not all. retry. */
3542 kiocb_done(kiocb, ret, cs);
3545 /* it's reportedly faster than delegating the null check to kfree() */
3551 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3555 ret = io_prep_rw(req, sqe);
3559 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3562 /* either don't need iovec imported or already have it */
3563 if (!req->async_data)
3565 return io_rw_prep_async(req, WRITE);
3568 static int io_write(struct io_kiocb *req, bool force_nonblock,
3569 struct io_comp_state *cs)
3571 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3572 struct kiocb *kiocb = &req->rw.kiocb;
3573 struct iov_iter __iter, *iter = &__iter;
3574 struct io_async_rw *rw = req->async_data;
3575 ssize_t ret, ret2, io_size;
3581 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3585 io_size = iov_iter_count(iter);
3586 req->result = io_size;
3588 /* Ensure we clear previously set non-block flag */
3589 if (!force_nonblock)
3590 kiocb->ki_flags &= ~IOCB_NOWAIT;
3592 kiocb->ki_flags |= IOCB_NOWAIT;
3594 /* If the file doesn't support async, just async punt */
3595 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
3598 /* file path doesn't support NOWAIT for non-direct_IO */
3599 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3600 (req->flags & REQ_F_ISREG))
3603 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size);
3608 * Open-code file_start_write here to grab freeze protection,
3609 * which will be released by another thread in
3610 * io_complete_rw(). Fool lockdep by telling it the lock got
3611 * released so that it doesn't complain about the held lock when
3612 * we return to userspace.
3614 if (req->flags & REQ_F_ISREG) {
3615 sb_start_write(file_inode(req->file)->i_sb);
3616 __sb_writers_release(file_inode(req->file)->i_sb,
3619 kiocb->ki_flags |= IOCB_WRITE;
3621 if (req->file->f_op->write_iter)
3622 ret2 = call_write_iter(req->file, kiocb, iter);
3623 else if (req->file->f_op->write)
3624 ret2 = loop_rw_iter(WRITE, req, iter);
3629 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3630 * retry them without IOCB_NOWAIT.
3632 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3634 /* no retry on NONBLOCK marked file */
3635 if (ret2 == -EAGAIN && (req->file->f_flags & O_NONBLOCK))
3637 if (!force_nonblock || ret2 != -EAGAIN) {
3638 /* IOPOLL retry should happen for io-wq threads */
3639 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3642 kiocb_done(kiocb, ret2, cs);
3645 /* some cases will consume bytes even on error returns */
3646 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3647 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3652 /* it's reportedly faster than delegating the null check to kfree() */
3658 static int io_renameat_prep(struct io_kiocb *req,
3659 const struct io_uring_sqe *sqe)
3661 struct io_rename *ren = &req->rename;
3662 const char __user *oldf, *newf;
3664 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3667 ren->old_dfd = READ_ONCE(sqe->fd);
3668 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3669 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3670 ren->new_dfd = READ_ONCE(sqe->len);
3671 ren->flags = READ_ONCE(sqe->rename_flags);
3673 ren->oldpath = getname(oldf);
3674 if (IS_ERR(ren->oldpath))
3675 return PTR_ERR(ren->oldpath);
3677 ren->newpath = getname(newf);
3678 if (IS_ERR(ren->newpath)) {
3679 putname(ren->oldpath);
3680 return PTR_ERR(ren->newpath);
3683 req->flags |= REQ_F_NEED_CLEANUP;
3687 static int io_renameat(struct io_kiocb *req, bool force_nonblock)
3689 struct io_rename *ren = &req->rename;
3695 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3696 ren->newpath, ren->flags);
3698 req->flags &= ~REQ_F_NEED_CLEANUP;
3700 req_set_fail_links(req);
3701 io_req_complete(req, ret);
3705 static int io_unlinkat_prep(struct io_kiocb *req,
3706 const struct io_uring_sqe *sqe)
3708 struct io_unlink *un = &req->unlink;
3709 const char __user *fname;
3711 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3714 un->dfd = READ_ONCE(sqe->fd);
3716 un->flags = READ_ONCE(sqe->unlink_flags);
3717 if (un->flags & ~AT_REMOVEDIR)
3720 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3721 un->filename = getname(fname);
3722 if (IS_ERR(un->filename))
3723 return PTR_ERR(un->filename);
3725 req->flags |= REQ_F_NEED_CLEANUP;
3729 static int io_unlinkat(struct io_kiocb *req, bool force_nonblock)
3731 struct io_unlink *un = &req->unlink;
3737 if (un->flags & AT_REMOVEDIR)
3738 ret = do_rmdir(un->dfd, un->filename);
3740 ret = do_unlinkat(un->dfd, un->filename);
3742 req->flags &= ~REQ_F_NEED_CLEANUP;
3744 req_set_fail_links(req);
3745 io_req_complete(req, ret);
3749 static int io_shutdown_prep(struct io_kiocb *req,
3750 const struct io_uring_sqe *sqe)
3752 #if defined(CONFIG_NET)
3753 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3755 if (sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3759 req->shutdown.how = READ_ONCE(sqe->len);
3766 static int io_shutdown(struct io_kiocb *req, bool force_nonblock)
3768 #if defined(CONFIG_NET)
3769 struct socket *sock;
3775 sock = sock_from_file(req->file);
3776 if (unlikely(!sock))
3779 ret = __sys_shutdown_sock(sock, req->shutdown.how);
3781 req_set_fail_links(req);
3782 io_req_complete(req, ret);
3789 static int __io_splice_prep(struct io_kiocb *req,
3790 const struct io_uring_sqe *sqe)
3792 struct io_splice* sp = &req->splice;
3793 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3795 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3799 sp->len = READ_ONCE(sqe->len);
3800 sp->flags = READ_ONCE(sqe->splice_flags);
3802 if (unlikely(sp->flags & ~valid_flags))
3805 sp->file_in = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in),
3806 (sp->flags & SPLICE_F_FD_IN_FIXED));
3809 req->flags |= REQ_F_NEED_CLEANUP;
3811 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
3813 * Splice operation will be punted aync, and here need to
3814 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3816 io_req_init_async(req);
3817 req->work.flags |= IO_WQ_WORK_UNBOUND;
3823 static int io_tee_prep(struct io_kiocb *req,
3824 const struct io_uring_sqe *sqe)
3826 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3828 return __io_splice_prep(req, sqe);
3831 static int io_tee(struct io_kiocb *req, bool force_nonblock)
3833 struct io_splice *sp = &req->splice;
3834 struct file *in = sp->file_in;
3835 struct file *out = sp->file_out;
3836 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3842 ret = do_tee(in, out, sp->len, flags);
3844 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3845 req->flags &= ~REQ_F_NEED_CLEANUP;
3848 req_set_fail_links(req);
3849 io_req_complete(req, ret);
3853 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3855 struct io_splice* sp = &req->splice;
3857 sp->off_in = READ_ONCE(sqe->splice_off_in);
3858 sp->off_out = READ_ONCE(sqe->off);
3859 return __io_splice_prep(req, sqe);
3862 static int io_splice(struct io_kiocb *req, bool force_nonblock)
3864 struct io_splice *sp = &req->splice;
3865 struct file *in = sp->file_in;
3866 struct file *out = sp->file_out;
3867 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3868 loff_t *poff_in, *poff_out;
3874 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3875 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3878 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3880 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3881 req->flags &= ~REQ_F_NEED_CLEANUP;
3884 req_set_fail_links(req);
3885 io_req_complete(req, ret);
3890 * IORING_OP_NOP just posts a completion event, nothing else.
3892 static int io_nop(struct io_kiocb *req, struct io_comp_state *cs)
3894 struct io_ring_ctx *ctx = req->ctx;
3896 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3899 __io_req_complete(req, 0, 0, cs);
3903 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3905 struct io_ring_ctx *ctx = req->ctx;
3910 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3912 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3915 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3916 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3919 req->sync.off = READ_ONCE(sqe->off);
3920 req->sync.len = READ_ONCE(sqe->len);
3924 static int io_fsync(struct io_kiocb *req, bool force_nonblock)
3926 loff_t end = req->sync.off + req->sync.len;
3929 /* fsync always requires a blocking context */
3933 ret = vfs_fsync_range(req->file, req->sync.off,
3934 end > 0 ? end : LLONG_MAX,
3935 req->sync.flags & IORING_FSYNC_DATASYNC);
3937 req_set_fail_links(req);
3938 io_req_complete(req, ret);
3942 static int io_fallocate_prep(struct io_kiocb *req,
3943 const struct io_uring_sqe *sqe)
3945 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3947 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3950 req->sync.off = READ_ONCE(sqe->off);
3951 req->sync.len = READ_ONCE(sqe->addr);
3952 req->sync.mode = READ_ONCE(sqe->len);
3956 static int io_fallocate(struct io_kiocb *req, bool force_nonblock)
3960 /* fallocate always requiring blocking context */
3963 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3966 req_set_fail_links(req);
3967 io_req_complete(req, ret);
3971 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3973 const char __user *fname;
3976 if (unlikely(sqe->ioprio || sqe->buf_index))
3978 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3981 /* open.how should be already initialised */
3982 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3983 req->open.how.flags |= O_LARGEFILE;
3985 req->open.dfd = READ_ONCE(sqe->fd);
3986 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3987 req->open.filename = getname(fname);
3988 if (IS_ERR(req->open.filename)) {
3989 ret = PTR_ERR(req->open.filename);
3990 req->open.filename = NULL;
3993 req->open.nofile = rlimit(RLIMIT_NOFILE);
3994 req->open.ignore_nonblock = false;
3995 req->flags |= REQ_F_NEED_CLEANUP;
3999 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4003 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4005 mode = READ_ONCE(sqe->len);
4006 flags = READ_ONCE(sqe->open_flags);
4007 req->open.how = build_open_how(flags, mode);
4008 return __io_openat_prep(req, sqe);
4011 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4013 struct open_how __user *how;
4017 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4019 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4020 len = READ_ONCE(sqe->len);
4021 if (len < OPEN_HOW_SIZE_VER0)
4024 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
4029 return __io_openat_prep(req, sqe);
4032 static int io_openat2(struct io_kiocb *req, bool force_nonblock)
4034 struct open_flags op;
4038 if (force_nonblock && !req->open.ignore_nonblock)
4041 ret = build_open_flags(&req->open.how, &op);
4045 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
4049 file = do_filp_open(req->open.dfd, req->open.filename, &op);
4052 ret = PTR_ERR(file);
4054 * A work-around to ensure that /proc/self works that way
4055 * that it should - if we get -EOPNOTSUPP back, then assume
4056 * that proc_self_get_link() failed us because we're in async
4057 * context. We should be safe to retry this from the task
4058 * itself with force_nonblock == false set, as it should not
4059 * block on lookup. Would be nice to know this upfront and
4060 * avoid the async dance, but doesn't seem feasible.
4062 if (ret == -EOPNOTSUPP && io_wq_current_is_worker()) {
4063 req->open.ignore_nonblock = true;
4064 refcount_inc(&req->refs);
4065 io_req_task_queue(req);
4069 fsnotify_open(file);
4070 fd_install(ret, file);
4073 putname(req->open.filename);
4074 req->flags &= ~REQ_F_NEED_CLEANUP;
4076 req_set_fail_links(req);
4077 io_req_complete(req, ret);
4081 static int io_openat(struct io_kiocb *req, bool force_nonblock)
4083 return io_openat2(req, force_nonblock);
4086 static int io_remove_buffers_prep(struct io_kiocb *req,
4087 const struct io_uring_sqe *sqe)
4089 struct io_provide_buf *p = &req->pbuf;
4092 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
4095 tmp = READ_ONCE(sqe->fd);
4096 if (!tmp || tmp > USHRT_MAX)
4099 memset(p, 0, sizeof(*p));
4101 p->bgid = READ_ONCE(sqe->buf_group);
4105 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
4106 int bgid, unsigned nbufs)
4110 /* shouldn't happen */
4114 /* the head kbuf is the list itself */
4115 while (!list_empty(&buf->list)) {
4116 struct io_buffer *nxt;
4118 nxt = list_first_entry(&buf->list, struct io_buffer, list);
4119 list_del(&nxt->list);
4126 idr_remove(&ctx->io_buffer_idr, bgid);
4131 static int io_remove_buffers(struct io_kiocb *req, bool force_nonblock,
4132 struct io_comp_state *cs)
4134 struct io_provide_buf *p = &req->pbuf;
4135 struct io_ring_ctx *ctx = req->ctx;
4136 struct io_buffer *head;
4139 io_ring_submit_lock(ctx, !force_nonblock);
4141 lockdep_assert_held(&ctx->uring_lock);
4144 head = idr_find(&ctx->io_buffer_idr, p->bgid);
4146 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
4148 req_set_fail_links(req);
4150 /* need to hold the lock to complete IOPOLL requests */
4151 if (ctx->flags & IORING_SETUP_IOPOLL) {
4152 __io_req_complete(req, ret, 0, cs);
4153 io_ring_submit_unlock(ctx, !force_nonblock);
4155 io_ring_submit_unlock(ctx, !force_nonblock);
4156 __io_req_complete(req, ret, 0, cs);
4161 static int io_provide_buffers_prep(struct io_kiocb *req,
4162 const struct io_uring_sqe *sqe)
4164 struct io_provide_buf *p = &req->pbuf;
4167 if (sqe->ioprio || sqe->rw_flags)
4170 tmp = READ_ONCE(sqe->fd);
4171 if (!tmp || tmp > USHRT_MAX)
4174 p->addr = READ_ONCE(sqe->addr);
4175 p->len = READ_ONCE(sqe->len);
4177 if (!access_ok(u64_to_user_ptr(p->addr), (p->len * p->nbufs)))
4180 p->bgid = READ_ONCE(sqe->buf_group);
4181 tmp = READ_ONCE(sqe->off);
4182 if (tmp > USHRT_MAX)
4188 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
4190 struct io_buffer *buf;
4191 u64 addr = pbuf->addr;
4192 int i, bid = pbuf->bid;
4194 for (i = 0; i < pbuf->nbufs; i++) {
4195 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
4200 buf->len = pbuf->len;
4205 INIT_LIST_HEAD(&buf->list);
4208 list_add_tail(&buf->list, &(*head)->list);
4212 return i ? i : -ENOMEM;
4215 static int io_provide_buffers(struct io_kiocb *req, bool force_nonblock,
4216 struct io_comp_state *cs)
4218 struct io_provide_buf *p = &req->pbuf;
4219 struct io_ring_ctx *ctx = req->ctx;
4220 struct io_buffer *head, *list;
4223 io_ring_submit_lock(ctx, !force_nonblock);
4225 lockdep_assert_held(&ctx->uring_lock);
4227 list = head = idr_find(&ctx->io_buffer_idr, p->bgid);
4229 ret = io_add_buffers(p, &head);
4234 ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1,
4237 __io_remove_buffers(ctx, head, p->bgid, -1U);
4243 req_set_fail_links(req);
4245 /* need to hold the lock to complete IOPOLL requests */
4246 if (ctx->flags & IORING_SETUP_IOPOLL) {
4247 __io_req_complete(req, ret, 0, cs);
4248 io_ring_submit_unlock(ctx, !force_nonblock);
4250 io_ring_submit_unlock(ctx, !force_nonblock);
4251 __io_req_complete(req, ret, 0, cs);
4256 static int io_epoll_ctl_prep(struct io_kiocb *req,
4257 const struct io_uring_sqe *sqe)
4259 #if defined(CONFIG_EPOLL)
4260 if (sqe->ioprio || sqe->buf_index)
4262 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4265 req->epoll.epfd = READ_ONCE(sqe->fd);
4266 req->epoll.op = READ_ONCE(sqe->len);
4267 req->epoll.fd = READ_ONCE(sqe->off);
4269 if (ep_op_has_event(req->epoll.op)) {
4270 struct epoll_event __user *ev;
4272 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4273 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4283 static int io_epoll_ctl(struct io_kiocb *req, bool force_nonblock,
4284 struct io_comp_state *cs)
4286 #if defined(CONFIG_EPOLL)
4287 struct io_epoll *ie = &req->epoll;
4290 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4291 if (force_nonblock && ret == -EAGAIN)
4295 req_set_fail_links(req);
4296 __io_req_complete(req, ret, 0, cs);
4303 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4305 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4306 if (sqe->ioprio || sqe->buf_index || sqe->off)
4308 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4311 req->madvise.addr = READ_ONCE(sqe->addr);
4312 req->madvise.len = READ_ONCE(sqe->len);
4313 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4320 static int io_madvise(struct io_kiocb *req, bool force_nonblock)
4322 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4323 struct io_madvise *ma = &req->madvise;
4329 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4331 req_set_fail_links(req);
4332 io_req_complete(req, ret);
4339 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4341 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4343 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4346 req->fadvise.offset = READ_ONCE(sqe->off);
4347 req->fadvise.len = READ_ONCE(sqe->len);
4348 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4352 static int io_fadvise(struct io_kiocb *req, bool force_nonblock)
4354 struct io_fadvise *fa = &req->fadvise;
4357 if (force_nonblock) {
4358 switch (fa->advice) {
4359 case POSIX_FADV_NORMAL:
4360 case POSIX_FADV_RANDOM:
4361 case POSIX_FADV_SEQUENTIAL:
4368 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4370 req_set_fail_links(req);
4371 io_req_complete(req, ret);
4375 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4377 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4379 if (sqe->ioprio || sqe->buf_index)
4381 if (req->flags & REQ_F_FIXED_FILE)
4384 req->statx.dfd = READ_ONCE(sqe->fd);
4385 req->statx.mask = READ_ONCE(sqe->len);
4386 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4387 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4388 req->statx.flags = READ_ONCE(sqe->statx_flags);
4393 static int io_statx(struct io_kiocb *req, bool force_nonblock)
4395 struct io_statx *ctx = &req->statx;
4398 if (force_nonblock) {
4399 /* only need file table for an actual valid fd */
4400 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
4401 req->flags |= REQ_F_NO_FILE_TABLE;
4405 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4409 req_set_fail_links(req);
4410 io_req_complete(req, ret);
4414 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4417 * If we queue this for async, it must not be cancellable. That would
4418 * leave the 'file' in an undeterminate state, and here need to modify
4419 * io_wq_work.flags, so initialize io_wq_work firstly.
4421 io_req_init_async(req);
4422 req->work.flags |= IO_WQ_WORK_NO_CANCEL;
4424 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4426 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4427 sqe->rw_flags || sqe->buf_index)
4429 if (req->flags & REQ_F_FIXED_FILE)
4432 req->close.fd = READ_ONCE(sqe->fd);
4433 if ((req->file && req->file->f_op == &io_uring_fops))
4436 req->close.put_file = NULL;
4440 static int io_close(struct io_kiocb *req, bool force_nonblock,
4441 struct io_comp_state *cs)
4443 struct io_close *close = &req->close;
4446 /* might be already done during nonblock submission */
4447 if (!close->put_file) {
4448 ret = close_fd_get_file(close->fd, &close->put_file);
4450 return (ret == -ENOENT) ? -EBADF : ret;
4453 /* if the file has a flush method, be safe and punt to async */
4454 if (close->put_file->f_op->flush && force_nonblock) {
4455 /* was never set, but play safe */
4456 req->flags &= ~REQ_F_NOWAIT;
4457 /* avoid grabbing files - we don't need the files */
4458 req->flags |= REQ_F_NO_FILE_TABLE;
4462 /* No ->flush() or already async, safely close from here */
4463 ret = filp_close(close->put_file, req->work.identity->files);
4465 req_set_fail_links(req);
4466 fput(close->put_file);
4467 close->put_file = NULL;
4468 __io_req_complete(req, ret, 0, cs);
4472 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4474 struct io_ring_ctx *ctx = req->ctx;
4479 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4481 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4484 req->sync.off = READ_ONCE(sqe->off);
4485 req->sync.len = READ_ONCE(sqe->len);
4486 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4490 static int io_sync_file_range(struct io_kiocb *req, bool force_nonblock)
4494 /* sync_file_range always requires a blocking context */
4498 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4501 req_set_fail_links(req);
4502 io_req_complete(req, ret);
4506 #if defined(CONFIG_NET)
4507 static int io_setup_async_msg(struct io_kiocb *req,
4508 struct io_async_msghdr *kmsg)
4510 struct io_async_msghdr *async_msg = req->async_data;
4514 if (io_alloc_async_data(req)) {
4515 if (kmsg->iov != kmsg->fast_iov)
4519 async_msg = req->async_data;
4520 req->flags |= REQ_F_NEED_CLEANUP;
4521 memcpy(async_msg, kmsg, sizeof(*kmsg));
4525 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4526 struct io_async_msghdr *iomsg)
4528 iomsg->iov = iomsg->fast_iov;
4529 iomsg->msg.msg_name = &iomsg->addr;
4530 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4531 req->sr_msg.msg_flags, &iomsg->iov);
4534 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4536 struct io_async_msghdr *async_msg = req->async_data;
4537 struct io_sr_msg *sr = &req->sr_msg;
4540 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4543 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4544 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4545 sr->len = READ_ONCE(sqe->len);
4547 #ifdef CONFIG_COMPAT
4548 if (req->ctx->compat)
4549 sr->msg_flags |= MSG_CMSG_COMPAT;
4552 if (!async_msg || !io_op_defs[req->opcode].needs_async_data)
4554 ret = io_sendmsg_copy_hdr(req, async_msg);
4556 req->flags |= REQ_F_NEED_CLEANUP;
4560 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4561 struct io_comp_state *cs)
4563 struct io_async_msghdr iomsg, *kmsg;
4564 struct socket *sock;
4568 sock = sock_from_file(req->file);
4569 if (unlikely(!sock))
4572 if (req->async_data) {
4573 kmsg = req->async_data;
4574 kmsg->msg.msg_name = &kmsg->addr;
4575 /* if iov is set, it's allocated already */
4577 kmsg->iov = kmsg->fast_iov;
4578 kmsg->msg.msg_iter.iov = kmsg->iov;
4580 ret = io_sendmsg_copy_hdr(req, &iomsg);
4586 flags = req->sr_msg.msg_flags;
4587 if (flags & MSG_DONTWAIT)
4588 req->flags |= REQ_F_NOWAIT;
4589 else if (force_nonblock)
4590 flags |= MSG_DONTWAIT;
4592 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4593 if (force_nonblock && ret == -EAGAIN)
4594 return io_setup_async_msg(req, kmsg);
4595 if (ret == -ERESTARTSYS)
4598 if (kmsg->iov != kmsg->fast_iov)
4600 req->flags &= ~REQ_F_NEED_CLEANUP;
4602 req_set_fail_links(req);
4603 __io_req_complete(req, ret, 0, cs);
4607 static int io_send(struct io_kiocb *req, bool force_nonblock,
4608 struct io_comp_state *cs)
4610 struct io_sr_msg *sr = &req->sr_msg;
4613 struct socket *sock;
4617 sock = sock_from_file(req->file);
4618 if (unlikely(!sock))
4621 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4625 msg.msg_name = NULL;
4626 msg.msg_control = NULL;
4627 msg.msg_controllen = 0;
4628 msg.msg_namelen = 0;
4630 flags = req->sr_msg.msg_flags;
4631 if (flags & MSG_DONTWAIT)
4632 req->flags |= REQ_F_NOWAIT;
4633 else if (force_nonblock)
4634 flags |= MSG_DONTWAIT;
4636 msg.msg_flags = flags;
4637 ret = sock_sendmsg(sock, &msg);
4638 if (force_nonblock && ret == -EAGAIN)
4640 if (ret == -ERESTARTSYS)
4644 req_set_fail_links(req);
4645 __io_req_complete(req, ret, 0, cs);
4649 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4650 struct io_async_msghdr *iomsg)
4652 struct io_sr_msg *sr = &req->sr_msg;
4653 struct iovec __user *uiov;
4657 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4658 &iomsg->uaddr, &uiov, &iov_len);
4662 if (req->flags & REQ_F_BUFFER_SELECT) {
4665 if (copy_from_user(iomsg->iov, uiov, sizeof(*uiov)))
4667 sr->len = iomsg->iov[0].iov_len;
4668 iov_iter_init(&iomsg->msg.msg_iter, READ, iomsg->iov, 1,
4672 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4673 &iomsg->iov, &iomsg->msg.msg_iter,
4682 #ifdef CONFIG_COMPAT
4683 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4684 struct io_async_msghdr *iomsg)
4686 struct compat_msghdr __user *msg_compat;
4687 struct io_sr_msg *sr = &req->sr_msg;
4688 struct compat_iovec __user *uiov;
4693 msg_compat = (struct compat_msghdr __user *) sr->umsg;
4694 ret = __get_compat_msghdr(&iomsg->msg, msg_compat, &iomsg->uaddr,
4699 uiov = compat_ptr(ptr);
4700 if (req->flags & REQ_F_BUFFER_SELECT) {
4701 compat_ssize_t clen;
4705 if (!access_ok(uiov, sizeof(*uiov)))
4707 if (__get_user(clen, &uiov->iov_len))
4712 iomsg->iov[0].iov_len = clen;
4715 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4716 UIO_FASTIOV, &iomsg->iov,
4717 &iomsg->msg.msg_iter, true);
4726 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4727 struct io_async_msghdr *iomsg)
4729 iomsg->msg.msg_name = &iomsg->addr;
4730 iomsg->iov = iomsg->fast_iov;
4732 #ifdef CONFIG_COMPAT
4733 if (req->ctx->compat)
4734 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4737 return __io_recvmsg_copy_hdr(req, iomsg);
4740 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4743 struct io_sr_msg *sr = &req->sr_msg;
4744 struct io_buffer *kbuf;
4746 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4751 req->flags |= REQ_F_BUFFER_SELECTED;
4755 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4757 return io_put_kbuf(req, req->sr_msg.kbuf);
4760 static int io_recvmsg_prep(struct io_kiocb *req,
4761 const struct io_uring_sqe *sqe)
4763 struct io_async_msghdr *async_msg = req->async_data;
4764 struct io_sr_msg *sr = &req->sr_msg;
4767 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4770 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4771 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4772 sr->len = READ_ONCE(sqe->len);
4773 sr->bgid = READ_ONCE(sqe->buf_group);
4775 #ifdef CONFIG_COMPAT
4776 if (req->ctx->compat)
4777 sr->msg_flags |= MSG_CMSG_COMPAT;
4780 if (!async_msg || !io_op_defs[req->opcode].needs_async_data)
4782 ret = io_recvmsg_copy_hdr(req, async_msg);
4784 req->flags |= REQ_F_NEED_CLEANUP;
4788 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4789 struct io_comp_state *cs)
4791 struct io_async_msghdr iomsg, *kmsg;
4792 struct socket *sock;
4793 struct io_buffer *kbuf;
4795 int ret, cflags = 0;
4797 sock = sock_from_file(req->file);
4798 if (unlikely(!sock))
4801 if (req->async_data) {
4802 kmsg = req->async_data;
4803 kmsg->msg.msg_name = &kmsg->addr;
4804 /* if iov is set, it's allocated already */
4806 kmsg->iov = kmsg->fast_iov;
4807 kmsg->msg.msg_iter.iov = kmsg->iov;
4809 ret = io_recvmsg_copy_hdr(req, &iomsg);
4815 if (req->flags & REQ_F_BUFFER_SELECT) {
4816 kbuf = io_recv_buffer_select(req, !force_nonblock);
4818 return PTR_ERR(kbuf);
4819 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4820 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->iov,
4821 1, req->sr_msg.len);
4824 flags = req->sr_msg.msg_flags;
4825 if (flags & MSG_DONTWAIT)
4826 req->flags |= REQ_F_NOWAIT;
4827 else if (force_nonblock)
4828 flags |= MSG_DONTWAIT;
4830 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4831 kmsg->uaddr, flags);
4832 if (force_nonblock && ret == -EAGAIN)
4833 return io_setup_async_msg(req, kmsg);
4834 if (ret == -ERESTARTSYS)
4837 if (req->flags & REQ_F_BUFFER_SELECTED)
4838 cflags = io_put_recv_kbuf(req);
4839 if (kmsg->iov != kmsg->fast_iov)
4841 req->flags &= ~REQ_F_NEED_CLEANUP;
4843 req_set_fail_links(req);
4844 __io_req_complete(req, ret, cflags, cs);
4848 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4849 struct io_comp_state *cs)
4851 struct io_buffer *kbuf;
4852 struct io_sr_msg *sr = &req->sr_msg;
4854 void __user *buf = sr->buf;
4855 struct socket *sock;
4858 int ret, cflags = 0;
4860 sock = sock_from_file(req->file);
4861 if (unlikely(!sock))
4864 if (req->flags & REQ_F_BUFFER_SELECT) {
4865 kbuf = io_recv_buffer_select(req, !force_nonblock);
4867 return PTR_ERR(kbuf);
4868 buf = u64_to_user_ptr(kbuf->addr);
4871 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4875 msg.msg_name = NULL;
4876 msg.msg_control = NULL;
4877 msg.msg_controllen = 0;
4878 msg.msg_namelen = 0;
4879 msg.msg_iocb = NULL;
4882 flags = req->sr_msg.msg_flags;
4883 if (flags & MSG_DONTWAIT)
4884 req->flags |= REQ_F_NOWAIT;
4885 else if (force_nonblock)
4886 flags |= MSG_DONTWAIT;
4888 ret = sock_recvmsg(sock, &msg, flags);
4889 if (force_nonblock && ret == -EAGAIN)
4891 if (ret == -ERESTARTSYS)
4894 if (req->flags & REQ_F_BUFFER_SELECTED)
4895 cflags = io_put_recv_kbuf(req);
4897 req_set_fail_links(req);
4898 __io_req_complete(req, ret, cflags, cs);
4902 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4904 struct io_accept *accept = &req->accept;
4906 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4908 if (sqe->ioprio || sqe->len || sqe->buf_index)
4911 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4912 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4913 accept->flags = READ_ONCE(sqe->accept_flags);
4914 accept->nofile = rlimit(RLIMIT_NOFILE);
4918 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4919 struct io_comp_state *cs)
4921 struct io_accept *accept = &req->accept;
4922 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4925 if (req->file->f_flags & O_NONBLOCK)
4926 req->flags |= REQ_F_NOWAIT;
4928 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4929 accept->addr_len, accept->flags,
4931 if (ret == -EAGAIN && force_nonblock)
4934 if (ret == -ERESTARTSYS)
4936 req_set_fail_links(req);
4938 __io_req_complete(req, ret, 0, cs);
4942 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4944 struct io_connect *conn = &req->connect;
4945 struct io_async_connect *io = req->async_data;
4947 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4949 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4952 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4953 conn->addr_len = READ_ONCE(sqe->addr2);
4958 return move_addr_to_kernel(conn->addr, conn->addr_len,
4962 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4963 struct io_comp_state *cs)
4965 struct io_async_connect __io, *io;
4966 unsigned file_flags;
4969 if (req->async_data) {
4970 io = req->async_data;
4972 ret = move_addr_to_kernel(req->connect.addr,
4973 req->connect.addr_len,
4980 file_flags = force_nonblock ? O_NONBLOCK : 0;
4982 ret = __sys_connect_file(req->file, &io->address,
4983 req->connect.addr_len, file_flags);
4984 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4985 if (req->async_data)
4987 if (io_alloc_async_data(req)) {
4991 io = req->async_data;
4992 memcpy(req->async_data, &__io, sizeof(__io));
4995 if (ret == -ERESTARTSYS)
4999 req_set_fail_links(req);
5000 __io_req_complete(req, ret, 0, cs);
5003 #else /* !CONFIG_NET */
5004 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5009 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
5010 struct io_comp_state *cs)
5015 static int io_send(struct io_kiocb *req, bool force_nonblock,
5016 struct io_comp_state *cs)
5021 static int io_recvmsg_prep(struct io_kiocb *req,
5022 const struct io_uring_sqe *sqe)
5027 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
5028 struct io_comp_state *cs)
5033 static int io_recv(struct io_kiocb *req, bool force_nonblock,
5034 struct io_comp_state *cs)
5039 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5044 static int io_accept(struct io_kiocb *req, bool force_nonblock,
5045 struct io_comp_state *cs)
5050 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5055 static int io_connect(struct io_kiocb *req, bool force_nonblock,
5056 struct io_comp_state *cs)
5060 #endif /* CONFIG_NET */
5062 struct io_poll_table {
5063 struct poll_table_struct pt;
5064 struct io_kiocb *req;
5068 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
5069 __poll_t mask, task_work_func_t func)
5073 /* for instances that support it check for an event match first: */
5074 if (mask && !(mask & poll->events))
5077 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
5079 list_del_init(&poll->wait.entry);
5082 init_task_work(&req->task_work, func);
5083 percpu_ref_get(&req->ctx->refs);
5086 * If this fails, then the task is exiting. When a task exits, the
5087 * work gets canceled, so just cancel this request as well instead
5088 * of executing it. We can't safely execute it anyway, as we may not
5089 * have the needed state needed for it anyway.
5091 ret = io_req_task_work_add(req);
5092 if (unlikely(ret)) {
5093 struct task_struct *tsk;
5095 WRITE_ONCE(poll->canceled, true);
5096 tsk = io_wq_get_task(req->ctx->io_wq);
5097 task_work_add(tsk, &req->task_work, TWA_NONE);
5098 wake_up_process(tsk);
5103 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
5104 __acquires(&req->ctx->completion_lock)
5106 struct io_ring_ctx *ctx = req->ctx;
5108 if (!req->result && !READ_ONCE(poll->canceled)) {
5109 struct poll_table_struct pt = { ._key = poll->events };
5111 req->result = vfs_poll(req->file, &pt) & poll->events;
5114 spin_lock_irq(&ctx->completion_lock);
5115 if (!req->result && !READ_ONCE(poll->canceled)) {
5116 add_wait_queue(poll->head, &poll->wait);
5123 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
5125 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
5126 if (req->opcode == IORING_OP_POLL_ADD)
5127 return req->async_data;
5128 return req->apoll->double_poll;
5131 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
5133 if (req->opcode == IORING_OP_POLL_ADD)
5135 return &req->apoll->poll;
5138 static void io_poll_remove_double(struct io_kiocb *req)
5140 struct io_poll_iocb *poll = io_poll_get_double(req);
5142 lockdep_assert_held(&req->ctx->completion_lock);
5144 if (poll && poll->head) {
5145 struct wait_queue_head *head = poll->head;
5147 spin_lock(&head->lock);
5148 list_del_init(&poll->wait.entry);
5149 if (poll->wait.private)
5150 refcount_dec(&req->refs);
5152 spin_unlock(&head->lock);
5156 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
5158 struct io_ring_ctx *ctx = req->ctx;
5160 io_poll_remove_double(req);
5161 req->poll.done = true;
5162 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
5163 io_commit_cqring(ctx);
5166 static void io_poll_task_func(struct callback_head *cb)
5168 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
5169 struct io_ring_ctx *ctx = req->ctx;
5170 struct io_kiocb *nxt;
5172 if (io_poll_rewait(req, &req->poll)) {
5173 spin_unlock_irq(&ctx->completion_lock);
5175 hash_del(&req->hash_node);
5176 io_poll_complete(req, req->result, 0);
5177 spin_unlock_irq(&ctx->completion_lock);
5179 nxt = io_put_req_find_next(req);
5180 io_cqring_ev_posted(ctx);
5182 __io_req_task_submit(nxt);
5185 percpu_ref_put(&ctx->refs);
5188 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
5189 int sync, void *key)
5191 struct io_kiocb *req = wait->private;
5192 struct io_poll_iocb *poll = io_poll_get_single(req);
5193 __poll_t mask = key_to_poll(key);
5195 /* for instances that support it check for an event match first: */
5196 if (mask && !(mask & poll->events))
5199 list_del_init(&wait->entry);
5201 if (poll && poll->head) {
5204 spin_lock(&poll->head->lock);
5205 done = list_empty(&poll->wait.entry);
5207 list_del_init(&poll->wait.entry);
5208 /* make sure double remove sees this as being gone */
5209 wait->private = NULL;
5210 spin_unlock(&poll->head->lock);
5212 /* use wait func handler, so it matches the rq type */
5213 poll->wait.func(&poll->wait, mode, sync, key);
5216 refcount_dec(&req->refs);
5220 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
5221 wait_queue_func_t wake_func)
5225 poll->canceled = false;
5226 poll->events = events;
5227 INIT_LIST_HEAD(&poll->wait.entry);
5228 init_waitqueue_func_entry(&poll->wait, wake_func);
5231 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5232 struct wait_queue_head *head,
5233 struct io_poll_iocb **poll_ptr)
5235 struct io_kiocb *req = pt->req;
5238 * If poll->head is already set, it's because the file being polled
5239 * uses multiple waitqueues for poll handling (eg one for read, one
5240 * for write). Setup a separate io_poll_iocb if this happens.
5242 if (unlikely(poll->head)) {
5243 struct io_poll_iocb *poll_one = poll;
5245 /* already have a 2nd entry, fail a third attempt */
5247 pt->error = -EINVAL;
5250 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5252 pt->error = -ENOMEM;
5255 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5256 refcount_inc(&req->refs);
5257 poll->wait.private = req;
5264 if (poll->events & EPOLLEXCLUSIVE)
5265 add_wait_queue_exclusive(head, &poll->wait);
5267 add_wait_queue(head, &poll->wait);
5270 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5271 struct poll_table_struct *p)
5273 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5274 struct async_poll *apoll = pt->req->apoll;
5276 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5279 static void io_async_task_func(struct callback_head *cb)
5281 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
5282 struct async_poll *apoll = req->apoll;
5283 struct io_ring_ctx *ctx = req->ctx;
5285 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
5287 if (io_poll_rewait(req, &apoll->poll)) {
5288 spin_unlock_irq(&ctx->completion_lock);
5289 percpu_ref_put(&ctx->refs);
5293 /* If req is still hashed, it cannot have been canceled. Don't check. */
5294 if (hash_hashed(&req->hash_node))
5295 hash_del(&req->hash_node);
5297 io_poll_remove_double(req);
5298 spin_unlock_irq(&ctx->completion_lock);
5300 if (!READ_ONCE(apoll->poll.canceled))
5301 __io_req_task_submit(req);
5303 __io_req_task_cancel(req, -ECANCELED);
5305 percpu_ref_put(&ctx->refs);
5306 kfree(apoll->double_poll);
5310 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5313 struct io_kiocb *req = wait->private;
5314 struct io_poll_iocb *poll = &req->apoll->poll;
5316 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5319 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5322 static void io_poll_req_insert(struct io_kiocb *req)
5324 struct io_ring_ctx *ctx = req->ctx;
5325 struct hlist_head *list;
5327 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5328 hlist_add_head(&req->hash_node, list);
5331 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5332 struct io_poll_iocb *poll,
5333 struct io_poll_table *ipt, __poll_t mask,
5334 wait_queue_func_t wake_func)
5335 __acquires(&ctx->completion_lock)
5337 struct io_ring_ctx *ctx = req->ctx;
5338 bool cancel = false;
5340 INIT_HLIST_NODE(&req->hash_node);
5341 io_init_poll_iocb(poll, mask, wake_func);
5342 poll->file = req->file;
5343 poll->wait.private = req;
5345 ipt->pt._key = mask;
5347 ipt->error = -EINVAL;
5349 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5351 spin_lock_irq(&ctx->completion_lock);
5352 if (likely(poll->head)) {
5353 spin_lock(&poll->head->lock);
5354 if (unlikely(list_empty(&poll->wait.entry))) {
5360 if (mask || ipt->error)
5361 list_del_init(&poll->wait.entry);
5363 WRITE_ONCE(poll->canceled, true);
5364 else if (!poll->done) /* actually waiting for an event */
5365 io_poll_req_insert(req);
5366 spin_unlock(&poll->head->lock);
5372 static bool io_arm_poll_handler(struct io_kiocb *req)
5374 const struct io_op_def *def = &io_op_defs[req->opcode];
5375 struct io_ring_ctx *ctx = req->ctx;
5376 struct async_poll *apoll;
5377 struct io_poll_table ipt;
5381 if (!req->file || !file_can_poll(req->file))
5383 if (req->flags & REQ_F_POLLED)
5387 else if (def->pollout)
5391 /* if we can't nonblock try, then no point in arming a poll handler */
5392 if (!io_file_supports_async(req->file, rw))
5395 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5396 if (unlikely(!apoll))
5398 apoll->double_poll = NULL;
5400 req->flags |= REQ_F_POLLED;
5405 mask |= POLLIN | POLLRDNORM;
5407 mask |= POLLOUT | POLLWRNORM;
5409 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5410 if ((req->opcode == IORING_OP_RECVMSG) &&
5411 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5414 mask |= POLLERR | POLLPRI;
5416 ipt.pt._qproc = io_async_queue_proc;
5418 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5420 if (ret || ipt.error) {
5421 io_poll_remove_double(req);
5422 spin_unlock_irq(&ctx->completion_lock);
5423 kfree(apoll->double_poll);
5427 spin_unlock_irq(&ctx->completion_lock);
5428 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
5429 apoll->poll.events);
5433 static bool __io_poll_remove_one(struct io_kiocb *req,
5434 struct io_poll_iocb *poll)
5436 bool do_complete = false;
5438 spin_lock(&poll->head->lock);
5439 WRITE_ONCE(poll->canceled, true);
5440 if (!list_empty(&poll->wait.entry)) {
5441 list_del_init(&poll->wait.entry);
5444 spin_unlock(&poll->head->lock);
5445 hash_del(&req->hash_node);
5449 static bool io_poll_remove_one(struct io_kiocb *req)
5453 io_poll_remove_double(req);
5455 if (req->opcode == IORING_OP_POLL_ADD) {
5456 do_complete = __io_poll_remove_one(req, &req->poll);
5458 struct async_poll *apoll = req->apoll;
5460 /* non-poll requests have submit ref still */
5461 do_complete = __io_poll_remove_one(req, &apoll->poll);
5464 kfree(apoll->double_poll);
5470 io_cqring_fill_event(req, -ECANCELED);
5471 io_commit_cqring(req->ctx);
5472 req_set_fail_links(req);
5473 io_put_req_deferred(req, 1);
5480 * Returns true if we found and killed one or more poll requests
5482 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5483 struct files_struct *files)
5485 struct hlist_node *tmp;
5486 struct io_kiocb *req;
5489 spin_lock_irq(&ctx->completion_lock);
5490 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5491 struct hlist_head *list;
5493 list = &ctx->cancel_hash[i];
5494 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5495 if (io_match_task(req, tsk, files))
5496 posted += io_poll_remove_one(req);
5499 spin_unlock_irq(&ctx->completion_lock);
5502 io_cqring_ev_posted(ctx);
5507 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
5509 struct hlist_head *list;
5510 struct io_kiocb *req;
5512 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5513 hlist_for_each_entry(req, list, hash_node) {
5514 if (sqe_addr != req->user_data)
5516 if (io_poll_remove_one(req))
5524 static int io_poll_remove_prep(struct io_kiocb *req,
5525 const struct io_uring_sqe *sqe)
5527 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5529 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
5533 req->poll_remove.addr = READ_ONCE(sqe->addr);
5538 * Find a running poll command that matches one specified in sqe->addr,
5539 * and remove it if found.
5541 static int io_poll_remove(struct io_kiocb *req)
5543 struct io_ring_ctx *ctx = req->ctx;
5546 spin_lock_irq(&ctx->completion_lock);
5547 ret = io_poll_cancel(ctx, req->poll_remove.addr);
5548 spin_unlock_irq(&ctx->completion_lock);
5551 req_set_fail_links(req);
5552 io_req_complete(req, ret);
5556 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5559 struct io_kiocb *req = wait->private;
5560 struct io_poll_iocb *poll = &req->poll;
5562 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5565 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5566 struct poll_table_struct *p)
5568 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5570 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5573 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5575 struct io_poll_iocb *poll = &req->poll;
5578 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5580 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
5583 events = READ_ONCE(sqe->poll32_events);
5585 events = swahw32(events);
5587 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP |
5588 (events & EPOLLEXCLUSIVE);
5592 static int io_poll_add(struct io_kiocb *req)
5594 struct io_poll_iocb *poll = &req->poll;
5595 struct io_ring_ctx *ctx = req->ctx;
5596 struct io_poll_table ipt;
5599 ipt.pt._qproc = io_poll_queue_proc;
5601 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5604 if (mask) { /* no async, we'd stolen it */
5606 io_poll_complete(req, mask, 0);
5608 spin_unlock_irq(&ctx->completion_lock);
5611 io_cqring_ev_posted(ctx);
5617 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5619 struct io_timeout_data *data = container_of(timer,
5620 struct io_timeout_data, timer);
5621 struct io_kiocb *req = data->req;
5622 struct io_ring_ctx *ctx = req->ctx;
5623 unsigned long flags;
5625 spin_lock_irqsave(&ctx->completion_lock, flags);
5626 list_del_init(&req->timeout.list);
5627 atomic_set(&req->ctx->cq_timeouts,
5628 atomic_read(&req->ctx->cq_timeouts) + 1);
5630 io_cqring_fill_event(req, -ETIME);
5631 io_commit_cqring(ctx);
5632 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5634 io_cqring_ev_posted(ctx);
5635 req_set_fail_links(req);
5637 return HRTIMER_NORESTART;
5640 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5643 struct io_timeout_data *io;
5644 struct io_kiocb *req;
5647 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5648 if (user_data == req->user_data) {
5655 return ERR_PTR(ret);
5657 io = req->async_data;
5658 ret = hrtimer_try_to_cancel(&io->timer);
5660 return ERR_PTR(-EALREADY);
5661 list_del_init(&req->timeout.list);
5665 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5667 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5670 return PTR_ERR(req);
5672 req_set_fail_links(req);
5673 io_cqring_fill_event(req, -ECANCELED);
5674 io_put_req_deferred(req, 1);
5678 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5679 struct timespec64 *ts, enum hrtimer_mode mode)
5681 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5682 struct io_timeout_data *data;
5685 return PTR_ERR(req);
5687 req->timeout.off = 0; /* noseq */
5688 data = req->async_data;
5689 list_add_tail(&req->timeout.list, &ctx->timeout_list);
5690 hrtimer_init(&data->timer, CLOCK_MONOTONIC, mode);
5691 data->timer.function = io_timeout_fn;
5692 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
5696 static int io_timeout_remove_prep(struct io_kiocb *req,
5697 const struct io_uring_sqe *sqe)
5699 struct io_timeout_rem *tr = &req->timeout_rem;
5701 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5703 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5705 if (sqe->ioprio || sqe->buf_index || sqe->len)
5708 tr->addr = READ_ONCE(sqe->addr);
5709 tr->flags = READ_ONCE(sqe->timeout_flags);
5710 if (tr->flags & IORING_TIMEOUT_UPDATE) {
5711 if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS))
5713 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
5715 } else if (tr->flags) {
5716 /* timeout removal doesn't support flags */
5724 * Remove or update an existing timeout command
5726 static int io_timeout_remove(struct io_kiocb *req)
5728 struct io_timeout_rem *tr = &req->timeout_rem;
5729 struct io_ring_ctx *ctx = req->ctx;
5732 spin_lock_irq(&ctx->completion_lock);
5733 if (req->timeout_rem.flags & IORING_TIMEOUT_UPDATE) {
5734 enum hrtimer_mode mode = (tr->flags & IORING_TIMEOUT_ABS)
5735 ? HRTIMER_MODE_ABS : HRTIMER_MODE_REL;
5737 ret = io_timeout_update(ctx, tr->addr, &tr->ts, mode);
5739 ret = io_timeout_cancel(ctx, tr->addr);
5742 io_cqring_fill_event(req, ret);
5743 io_commit_cqring(ctx);
5744 spin_unlock_irq(&ctx->completion_lock);
5745 io_cqring_ev_posted(ctx);
5747 req_set_fail_links(req);
5752 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5753 bool is_timeout_link)
5755 struct io_timeout_data *data;
5757 u32 off = READ_ONCE(sqe->off);
5759 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5761 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5763 if (off && is_timeout_link)
5765 flags = READ_ONCE(sqe->timeout_flags);
5766 if (flags & ~IORING_TIMEOUT_ABS)
5769 req->timeout.off = off;
5771 if (!req->async_data && io_alloc_async_data(req))
5774 data = req->async_data;
5777 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5780 if (flags & IORING_TIMEOUT_ABS)
5781 data->mode = HRTIMER_MODE_ABS;
5783 data->mode = HRTIMER_MODE_REL;
5785 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5789 static int io_timeout(struct io_kiocb *req)
5791 struct io_ring_ctx *ctx = req->ctx;
5792 struct io_timeout_data *data = req->async_data;
5793 struct list_head *entry;
5794 u32 tail, off = req->timeout.off;
5796 spin_lock_irq(&ctx->completion_lock);
5799 * sqe->off holds how many events that need to occur for this
5800 * timeout event to be satisfied. If it isn't set, then this is
5801 * a pure timeout request, sequence isn't used.
5803 if (io_is_timeout_noseq(req)) {
5804 entry = ctx->timeout_list.prev;
5808 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5809 req->timeout.target_seq = tail + off;
5812 * Insertion sort, ensuring the first entry in the list is always
5813 * the one we need first.
5815 list_for_each_prev(entry, &ctx->timeout_list) {
5816 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5819 if (io_is_timeout_noseq(nxt))
5821 /* nxt.seq is behind @tail, otherwise would've been completed */
5822 if (off >= nxt->timeout.target_seq - tail)
5826 list_add(&req->timeout.list, entry);
5827 data->timer.function = io_timeout_fn;
5828 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5829 spin_unlock_irq(&ctx->completion_lock);
5833 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5835 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5837 return req->user_data == (unsigned long) data;
5840 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
5842 enum io_wq_cancel cancel_ret;
5845 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr, false);
5846 switch (cancel_ret) {
5847 case IO_WQ_CANCEL_OK:
5850 case IO_WQ_CANCEL_RUNNING:
5853 case IO_WQ_CANCEL_NOTFOUND:
5861 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5862 struct io_kiocb *req, __u64 sqe_addr,
5865 unsigned long flags;
5868 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
5869 if (ret != -ENOENT) {
5870 spin_lock_irqsave(&ctx->completion_lock, flags);
5874 spin_lock_irqsave(&ctx->completion_lock, flags);
5875 ret = io_timeout_cancel(ctx, sqe_addr);
5878 ret = io_poll_cancel(ctx, sqe_addr);
5882 io_cqring_fill_event(req, ret);
5883 io_commit_cqring(ctx);
5884 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5885 io_cqring_ev_posted(ctx);
5888 req_set_fail_links(req);
5892 static int io_async_cancel_prep(struct io_kiocb *req,
5893 const struct io_uring_sqe *sqe)
5895 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5897 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5899 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5902 req->cancel.addr = READ_ONCE(sqe->addr);
5906 static int io_async_cancel(struct io_kiocb *req)
5908 struct io_ring_ctx *ctx = req->ctx;
5910 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
5914 static int io_files_update_prep(struct io_kiocb *req,
5915 const struct io_uring_sqe *sqe)
5917 if (unlikely(req->ctx->flags & IORING_SETUP_SQPOLL))
5919 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5921 if (sqe->ioprio || sqe->rw_flags)
5924 req->files_update.offset = READ_ONCE(sqe->off);
5925 req->files_update.nr_args = READ_ONCE(sqe->len);
5926 if (!req->files_update.nr_args)
5928 req->files_update.arg = READ_ONCE(sqe->addr);
5932 static int io_files_update(struct io_kiocb *req, bool force_nonblock,
5933 struct io_comp_state *cs)
5935 struct io_ring_ctx *ctx = req->ctx;
5936 struct io_uring_files_update up;
5942 up.offset = req->files_update.offset;
5943 up.fds = req->files_update.arg;
5945 mutex_lock(&ctx->uring_lock);
5946 ret = __io_sqe_files_update(ctx, &up, req->files_update.nr_args);
5947 mutex_unlock(&ctx->uring_lock);
5950 req_set_fail_links(req);
5951 __io_req_complete(req, ret, 0, cs);
5955 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5957 switch (req->opcode) {
5960 case IORING_OP_READV:
5961 case IORING_OP_READ_FIXED:
5962 case IORING_OP_READ:
5963 return io_read_prep(req, sqe);
5964 case IORING_OP_WRITEV:
5965 case IORING_OP_WRITE_FIXED:
5966 case IORING_OP_WRITE:
5967 return io_write_prep(req, sqe);
5968 case IORING_OP_POLL_ADD:
5969 return io_poll_add_prep(req, sqe);
5970 case IORING_OP_POLL_REMOVE:
5971 return io_poll_remove_prep(req, sqe);
5972 case IORING_OP_FSYNC:
5973 return io_prep_fsync(req, sqe);
5974 case IORING_OP_SYNC_FILE_RANGE:
5975 return io_prep_sfr(req, sqe);
5976 case IORING_OP_SENDMSG:
5977 case IORING_OP_SEND:
5978 return io_sendmsg_prep(req, sqe);
5979 case IORING_OP_RECVMSG:
5980 case IORING_OP_RECV:
5981 return io_recvmsg_prep(req, sqe);
5982 case IORING_OP_CONNECT:
5983 return io_connect_prep(req, sqe);
5984 case IORING_OP_TIMEOUT:
5985 return io_timeout_prep(req, sqe, false);
5986 case IORING_OP_TIMEOUT_REMOVE:
5987 return io_timeout_remove_prep(req, sqe);
5988 case IORING_OP_ASYNC_CANCEL:
5989 return io_async_cancel_prep(req, sqe);
5990 case IORING_OP_LINK_TIMEOUT:
5991 return io_timeout_prep(req, sqe, true);
5992 case IORING_OP_ACCEPT:
5993 return io_accept_prep(req, sqe);
5994 case IORING_OP_FALLOCATE:
5995 return io_fallocate_prep(req, sqe);
5996 case IORING_OP_OPENAT:
5997 return io_openat_prep(req, sqe);
5998 case IORING_OP_CLOSE:
5999 return io_close_prep(req, sqe);
6000 case IORING_OP_FILES_UPDATE:
6001 return io_files_update_prep(req, sqe);
6002 case IORING_OP_STATX:
6003 return io_statx_prep(req, sqe);
6004 case IORING_OP_FADVISE:
6005 return io_fadvise_prep(req, sqe);
6006 case IORING_OP_MADVISE:
6007 return io_madvise_prep(req, sqe);
6008 case IORING_OP_OPENAT2:
6009 return io_openat2_prep(req, sqe);
6010 case IORING_OP_EPOLL_CTL:
6011 return io_epoll_ctl_prep(req, sqe);
6012 case IORING_OP_SPLICE:
6013 return io_splice_prep(req, sqe);
6014 case IORING_OP_PROVIDE_BUFFERS:
6015 return io_provide_buffers_prep(req, sqe);
6016 case IORING_OP_REMOVE_BUFFERS:
6017 return io_remove_buffers_prep(req, sqe);
6019 return io_tee_prep(req, sqe);
6020 case IORING_OP_SHUTDOWN:
6021 return io_shutdown_prep(req, sqe);
6022 case IORING_OP_RENAMEAT:
6023 return io_renameat_prep(req, sqe);
6024 case IORING_OP_UNLINKAT:
6025 return io_unlinkat_prep(req, sqe);
6028 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
6033 static int io_req_defer_prep(struct io_kiocb *req,
6034 const struct io_uring_sqe *sqe)
6038 if (io_alloc_async_data(req))
6040 return io_req_prep(req, sqe);
6043 static u32 io_get_sequence(struct io_kiocb *req)
6045 struct io_kiocb *pos;
6046 struct io_ring_ctx *ctx = req->ctx;
6047 u32 total_submitted, nr_reqs = 0;
6049 io_for_each_link(pos, req)
6052 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
6053 return total_submitted - nr_reqs;
6056 static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe)
6058 struct io_ring_ctx *ctx = req->ctx;
6059 struct io_defer_entry *de;
6063 /* Still need defer if there is pending req in defer list. */
6064 if (likely(list_empty_careful(&ctx->defer_list) &&
6065 !(req->flags & REQ_F_IO_DRAIN)))
6068 seq = io_get_sequence(req);
6069 /* Still a chance to pass the sequence check */
6070 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
6073 if (!req->async_data) {
6074 ret = io_req_defer_prep(req, sqe);
6078 io_prep_async_link(req);
6079 de = kmalloc(sizeof(*de), GFP_KERNEL);
6083 spin_lock_irq(&ctx->completion_lock);
6084 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
6085 spin_unlock_irq(&ctx->completion_lock);
6087 io_queue_async_work(req);
6088 return -EIOCBQUEUED;
6091 trace_io_uring_defer(ctx, req, req->user_data);
6094 list_add_tail(&de->list, &ctx->defer_list);
6095 spin_unlock_irq(&ctx->completion_lock);
6096 return -EIOCBQUEUED;
6099 static void io_req_drop_files(struct io_kiocb *req)
6101 struct io_ring_ctx *ctx = req->ctx;
6102 struct io_uring_task *tctx = req->task->io_uring;
6103 unsigned long flags;
6105 spin_lock_irqsave(&ctx->inflight_lock, flags);
6106 list_del(&req->inflight_entry);
6107 if (atomic_read(&tctx->in_idle))
6108 wake_up(&tctx->wait);
6109 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
6110 req->flags &= ~REQ_F_INFLIGHT;
6111 put_files_struct(req->work.identity->files);
6112 put_nsproxy(req->work.identity->nsproxy);
6113 req->work.flags &= ~IO_WQ_WORK_FILES;
6116 static void __io_clean_op(struct io_kiocb *req)
6118 if (req->flags & REQ_F_BUFFER_SELECTED) {
6119 switch (req->opcode) {
6120 case IORING_OP_READV:
6121 case IORING_OP_READ_FIXED:
6122 case IORING_OP_READ:
6123 kfree((void *)(unsigned long)req->rw.addr);
6125 case IORING_OP_RECVMSG:
6126 case IORING_OP_RECV:
6127 kfree(req->sr_msg.kbuf);
6130 req->flags &= ~REQ_F_BUFFER_SELECTED;
6133 if (req->flags & REQ_F_NEED_CLEANUP) {
6134 switch (req->opcode) {
6135 case IORING_OP_READV:
6136 case IORING_OP_READ_FIXED:
6137 case IORING_OP_READ:
6138 case IORING_OP_WRITEV:
6139 case IORING_OP_WRITE_FIXED:
6140 case IORING_OP_WRITE: {
6141 struct io_async_rw *io = req->async_data;
6143 kfree(io->free_iovec);
6146 case IORING_OP_RECVMSG:
6147 case IORING_OP_SENDMSG: {
6148 struct io_async_msghdr *io = req->async_data;
6149 if (io->iov != io->fast_iov)
6153 case IORING_OP_SPLICE:
6155 io_put_file(req, req->splice.file_in,
6156 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
6158 case IORING_OP_OPENAT:
6159 case IORING_OP_OPENAT2:
6160 if (req->open.filename)
6161 putname(req->open.filename);
6163 case IORING_OP_RENAMEAT:
6164 putname(req->rename.oldpath);
6165 putname(req->rename.newpath);
6167 case IORING_OP_UNLINKAT:
6168 putname(req->unlink.filename);
6171 req->flags &= ~REQ_F_NEED_CLEANUP;
6174 if (req->flags & REQ_F_INFLIGHT)
6175 io_req_drop_files(req);
6178 static int io_issue_sqe(struct io_kiocb *req, bool force_nonblock,
6179 struct io_comp_state *cs)
6181 struct io_ring_ctx *ctx = req->ctx;
6184 switch (req->opcode) {
6186 ret = io_nop(req, cs);
6188 case IORING_OP_READV:
6189 case IORING_OP_READ_FIXED:
6190 case IORING_OP_READ:
6191 ret = io_read(req, force_nonblock, cs);
6193 case IORING_OP_WRITEV:
6194 case IORING_OP_WRITE_FIXED:
6195 case IORING_OP_WRITE:
6196 ret = io_write(req, force_nonblock, cs);
6198 case IORING_OP_FSYNC:
6199 ret = io_fsync(req, force_nonblock);
6201 case IORING_OP_POLL_ADD:
6202 ret = io_poll_add(req);
6204 case IORING_OP_POLL_REMOVE:
6205 ret = io_poll_remove(req);
6207 case IORING_OP_SYNC_FILE_RANGE:
6208 ret = io_sync_file_range(req, force_nonblock);
6210 case IORING_OP_SENDMSG:
6211 ret = io_sendmsg(req, force_nonblock, cs);
6213 case IORING_OP_SEND:
6214 ret = io_send(req, force_nonblock, cs);
6216 case IORING_OP_RECVMSG:
6217 ret = io_recvmsg(req, force_nonblock, cs);
6219 case IORING_OP_RECV:
6220 ret = io_recv(req, force_nonblock, cs);
6222 case IORING_OP_TIMEOUT:
6223 ret = io_timeout(req);
6225 case IORING_OP_TIMEOUT_REMOVE:
6226 ret = io_timeout_remove(req);
6228 case IORING_OP_ACCEPT:
6229 ret = io_accept(req, force_nonblock, cs);
6231 case IORING_OP_CONNECT:
6232 ret = io_connect(req, force_nonblock, cs);
6234 case IORING_OP_ASYNC_CANCEL:
6235 ret = io_async_cancel(req);
6237 case IORING_OP_FALLOCATE:
6238 ret = io_fallocate(req, force_nonblock);
6240 case IORING_OP_OPENAT:
6241 ret = io_openat(req, force_nonblock);
6243 case IORING_OP_CLOSE:
6244 ret = io_close(req, force_nonblock, cs);
6246 case IORING_OP_FILES_UPDATE:
6247 ret = io_files_update(req, force_nonblock, cs);
6249 case IORING_OP_STATX:
6250 ret = io_statx(req, force_nonblock);
6252 case IORING_OP_FADVISE:
6253 ret = io_fadvise(req, force_nonblock);
6255 case IORING_OP_MADVISE:
6256 ret = io_madvise(req, force_nonblock);
6258 case IORING_OP_OPENAT2:
6259 ret = io_openat2(req, force_nonblock);
6261 case IORING_OP_EPOLL_CTL:
6262 ret = io_epoll_ctl(req, force_nonblock, cs);
6264 case IORING_OP_SPLICE:
6265 ret = io_splice(req, force_nonblock);
6267 case IORING_OP_PROVIDE_BUFFERS:
6268 ret = io_provide_buffers(req, force_nonblock, cs);
6270 case IORING_OP_REMOVE_BUFFERS:
6271 ret = io_remove_buffers(req, force_nonblock, cs);
6274 ret = io_tee(req, force_nonblock);
6276 case IORING_OP_SHUTDOWN:
6277 ret = io_shutdown(req, force_nonblock);
6279 case IORING_OP_RENAMEAT:
6280 ret = io_renameat(req, force_nonblock);
6282 case IORING_OP_UNLINKAT:
6283 ret = io_unlinkat(req, force_nonblock);
6293 /* If the op doesn't have a file, we're not polling for it */
6294 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
6295 const bool in_async = io_wq_current_is_worker();
6297 /* workqueue context doesn't hold uring_lock, grab it now */
6299 mutex_lock(&ctx->uring_lock);
6301 io_iopoll_req_issued(req, in_async);
6304 mutex_unlock(&ctx->uring_lock);
6310 static struct io_wq_work *io_wq_submit_work(struct io_wq_work *work)
6312 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6313 struct io_kiocb *timeout;
6316 timeout = io_prep_linked_timeout(req);
6318 io_queue_linked_timeout(timeout);
6320 /* if NO_CANCEL is set, we must still run the work */
6321 if ((work->flags & (IO_WQ_WORK_CANCEL|IO_WQ_WORK_NO_CANCEL)) ==
6322 IO_WQ_WORK_CANCEL) {
6328 ret = io_issue_sqe(req, false, NULL);
6330 * We can get EAGAIN for polled IO even though we're
6331 * forcing a sync submission from here, since we can't
6332 * wait for request slots on the block side.
6342 * io_iopoll_complete() does not hold completion_lock to complete
6343 * polled io, so here for polled io, just mark it done and still let
6344 * io_iopoll_complete() complete it.
6346 if (req->ctx->flags & IORING_SETUP_IOPOLL) {
6347 struct kiocb *kiocb = &req->rw.kiocb;
6349 kiocb_done(kiocb, ret, NULL);
6351 req_set_fail_links(req);
6352 io_req_complete(req, ret);
6356 return io_steal_work(req);
6359 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6362 struct fixed_file_table *table;
6364 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
6365 return table->files[index & IORING_FILE_TABLE_MASK];
6368 static struct file *io_file_get(struct io_submit_state *state,
6369 struct io_kiocb *req, int fd, bool fixed)
6371 struct io_ring_ctx *ctx = req->ctx;
6375 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6377 fd = array_index_nospec(fd, ctx->nr_user_files);
6378 file = io_file_from_index(ctx, fd);
6379 io_set_resource_node(req);
6381 trace_io_uring_file_get(ctx, fd);
6382 file = __io_file_get(state, fd);
6388 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6390 struct io_timeout_data *data = container_of(timer,
6391 struct io_timeout_data, timer);
6392 struct io_kiocb *prev, *req = data->req;
6393 struct io_ring_ctx *ctx = req->ctx;
6394 unsigned long flags;
6396 spin_lock_irqsave(&ctx->completion_lock, flags);
6397 prev = req->timeout.head;
6398 req->timeout.head = NULL;
6401 * We don't expect the list to be empty, that will only happen if we
6402 * race with the completion of the linked work.
6404 if (prev && refcount_inc_not_zero(&prev->refs))
6405 io_remove_next_linked(prev);
6408 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6411 req_set_fail_links(prev);
6412 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6415 io_req_complete(req, -ETIME);
6417 return HRTIMER_NORESTART;
6420 static void __io_queue_linked_timeout(struct io_kiocb *req)
6423 * If the back reference is NULL, then our linked request finished
6424 * before we got a chance to setup the timer
6426 if (req->timeout.head) {
6427 struct io_timeout_data *data = req->async_data;
6429 data->timer.function = io_link_timeout_fn;
6430 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6435 static void io_queue_linked_timeout(struct io_kiocb *req)
6437 struct io_ring_ctx *ctx = req->ctx;
6439 spin_lock_irq(&ctx->completion_lock);
6440 __io_queue_linked_timeout(req);
6441 spin_unlock_irq(&ctx->completion_lock);
6443 /* drop submission reference */
6447 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6449 struct io_kiocb *nxt = req->link;
6451 if (!nxt || (req->flags & REQ_F_LINK_TIMEOUT) ||
6452 nxt->opcode != IORING_OP_LINK_TIMEOUT)
6455 nxt->timeout.head = req;
6456 nxt->flags |= REQ_F_LTIMEOUT_ACTIVE;
6457 req->flags |= REQ_F_LINK_TIMEOUT;
6461 static void __io_queue_sqe(struct io_kiocb *req, struct io_comp_state *cs)
6463 struct io_kiocb *linked_timeout;
6464 const struct cred *old_creds = NULL;
6468 linked_timeout = io_prep_linked_timeout(req);
6470 if ((req->flags & REQ_F_WORK_INITIALIZED) &&
6471 (req->work.flags & IO_WQ_WORK_CREDS) &&
6472 req->work.identity->creds != current_cred()) {
6474 revert_creds(old_creds);
6475 if (old_creds == req->work.identity->creds)
6476 old_creds = NULL; /* restored original creds */
6478 old_creds = override_creds(req->work.identity->creds);
6481 ret = io_issue_sqe(req, true, cs);
6484 * We async punt it if the file wasn't marked NOWAIT, or if the file
6485 * doesn't support non-blocking read/write attempts
6487 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6488 if (!io_arm_poll_handler(req)) {
6490 * Queued up for async execution, worker will release
6491 * submit reference when the iocb is actually submitted.
6493 io_queue_async_work(req);
6497 io_queue_linked_timeout(linked_timeout);
6498 } else if (likely(!ret)) {
6499 /* drop submission reference */
6500 req = io_put_req_find_next(req);
6502 io_queue_linked_timeout(linked_timeout);
6505 if (!(req->flags & REQ_F_FORCE_ASYNC))
6507 io_queue_async_work(req);
6510 /* un-prep timeout, so it'll be killed as any other linked */
6511 req->flags &= ~REQ_F_LINK_TIMEOUT;
6512 req_set_fail_links(req);
6514 io_req_complete(req, ret);
6518 revert_creds(old_creds);
6521 static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6522 struct io_comp_state *cs)
6526 ret = io_req_defer(req, sqe);
6528 if (ret != -EIOCBQUEUED) {
6530 req_set_fail_links(req);
6532 io_req_complete(req, ret);
6534 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6535 if (!req->async_data) {
6536 ret = io_req_defer_prep(req, sqe);
6540 io_queue_async_work(req);
6543 ret = io_req_prep(req, sqe);
6547 __io_queue_sqe(req, cs);
6551 static inline void io_queue_link_head(struct io_kiocb *req,
6552 struct io_comp_state *cs)
6554 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
6556 io_req_complete(req, -ECANCELED);
6558 io_queue_sqe(req, NULL, cs);
6561 struct io_submit_link {
6562 struct io_kiocb *head;
6563 struct io_kiocb *last;
6566 static int io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6567 struct io_submit_link *link, struct io_comp_state *cs)
6569 struct io_ring_ctx *ctx = req->ctx;
6573 * If we already have a head request, queue this one for async
6574 * submittal once the head completes. If we don't have a head but
6575 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6576 * submitted sync once the chain is complete. If none of those
6577 * conditions are true (normal request), then just queue it.
6580 struct io_kiocb *head = link->head;
6583 * Taking sequential execution of a link, draining both sides
6584 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6585 * requests in the link. So, it drains the head and the
6586 * next after the link request. The last one is done via
6587 * drain_next flag to persist the effect across calls.
6589 if (req->flags & REQ_F_IO_DRAIN) {
6590 head->flags |= REQ_F_IO_DRAIN;
6591 ctx->drain_next = 1;
6593 ret = io_req_defer_prep(req, sqe);
6594 if (unlikely(ret)) {
6595 /* fail even hard links since we don't submit */
6596 head->flags |= REQ_F_FAIL_LINK;
6599 trace_io_uring_link(ctx, req, head);
6600 link->last->link = req;
6603 /* last request of a link, enqueue the link */
6604 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6605 io_queue_link_head(head, cs);
6609 if (unlikely(ctx->drain_next)) {
6610 req->flags |= REQ_F_IO_DRAIN;
6611 ctx->drain_next = 0;
6613 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6614 ret = io_req_defer_prep(req, sqe);
6616 req->flags |= REQ_F_FAIL_LINK;
6620 io_queue_sqe(req, sqe, cs);
6628 * Batched submission is done, ensure local IO is flushed out.
6630 static void io_submit_state_end(struct io_submit_state *state)
6632 if (!list_empty(&state->comp.list))
6633 io_submit_flush_completions(&state->comp);
6634 if (state->plug_started)
6635 blk_finish_plug(&state->plug);
6636 io_state_file_put(state);
6637 if (state->free_reqs)
6638 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
6642 * Start submission side cache.
6644 static void io_submit_state_start(struct io_submit_state *state,
6645 struct io_ring_ctx *ctx, unsigned int max_ios)
6647 state->plug_started = false;
6649 INIT_LIST_HEAD(&state->comp.list);
6650 state->comp.ctx = ctx;
6651 state->free_reqs = 0;
6652 state->file_refs = 0;
6653 state->ios_left = max_ios;
6656 static void io_commit_sqring(struct io_ring_ctx *ctx)
6658 struct io_rings *rings = ctx->rings;
6661 * Ensure any loads from the SQEs are done at this point,
6662 * since once we write the new head, the application could
6663 * write new data to them.
6665 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6669 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6670 * that is mapped by userspace. This means that care needs to be taken to
6671 * ensure that reads are stable, as we cannot rely on userspace always
6672 * being a good citizen. If members of the sqe are validated and then later
6673 * used, it's important that those reads are done through READ_ONCE() to
6674 * prevent a re-load down the line.
6676 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6678 u32 *sq_array = ctx->sq_array;
6682 * The cached sq head (or cq tail) serves two purposes:
6684 * 1) allows us to batch the cost of updating the user visible
6686 * 2) allows the kernel side to track the head on its own, even
6687 * though the application is the one updating it.
6689 head = READ_ONCE(sq_array[ctx->cached_sq_head & ctx->sq_mask]);
6690 if (likely(head < ctx->sq_entries))
6691 return &ctx->sq_sqes[head];
6693 /* drop invalid entries */
6694 ctx->cached_sq_dropped++;
6695 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6699 static inline void io_consume_sqe(struct io_ring_ctx *ctx)
6701 ctx->cached_sq_head++;
6705 * Check SQE restrictions (opcode and flags).
6707 * Returns 'true' if SQE is allowed, 'false' otherwise.
6709 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6710 struct io_kiocb *req,
6711 unsigned int sqe_flags)
6713 if (!ctx->restricted)
6716 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6719 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6720 ctx->restrictions.sqe_flags_required)
6723 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6724 ctx->restrictions.sqe_flags_required))
6730 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
6731 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
6732 IOSQE_BUFFER_SELECT)
6734 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6735 const struct io_uring_sqe *sqe,
6736 struct io_submit_state *state)
6738 unsigned int sqe_flags;
6741 req->opcode = READ_ONCE(sqe->opcode);
6742 req->user_data = READ_ONCE(sqe->user_data);
6743 req->async_data = NULL;
6748 req->fixed_file_refs = NULL;
6749 /* one is dropped after submission, the other at completion */
6750 refcount_set(&req->refs, 2);
6751 req->task = current;
6754 if (unlikely(req->opcode >= IORING_OP_LAST))
6757 if (unlikely(io_sq_thread_acquire_mm_files(ctx, req)))
6760 sqe_flags = READ_ONCE(sqe->flags);
6761 /* enforce forwards compatibility on users */
6762 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6765 if (unlikely(!io_check_restriction(ctx, req, sqe_flags)))
6768 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6769 !io_op_defs[req->opcode].buffer_select)
6772 id = READ_ONCE(sqe->personality);
6774 struct io_identity *iod;
6776 iod = idr_find(&ctx->personality_idr, id);
6779 refcount_inc(&iod->count);
6781 __io_req_init_async(req);
6782 get_cred(iod->creds);
6783 req->work.identity = iod;
6784 req->work.flags |= IO_WQ_WORK_CREDS;
6787 /* same numerical values with corresponding REQ_F_*, safe to copy */
6788 req->flags |= sqe_flags;
6791 * Plug now if we have more than 1 IO left after this, and the target
6792 * is potentially a read/write to block based storage.
6794 if (!state->plug_started && state->ios_left > 1 &&
6795 io_op_defs[req->opcode].plug) {
6796 blk_start_plug(&state->plug);
6797 state->plug_started = true;
6801 if (io_op_defs[req->opcode].needs_file) {
6802 bool fixed = req->flags & REQ_F_FIXED_FILE;
6804 req->file = io_file_get(state, req, READ_ONCE(sqe->fd), fixed);
6805 if (unlikely(!req->file &&
6806 !io_op_defs[req->opcode].needs_file_no_error))
6814 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6816 struct io_submit_state state;
6817 struct io_submit_link link;
6818 int i, submitted = 0;
6820 /* if we have a backlog and couldn't flush it all, return BUSY */
6821 if (test_bit(0, &ctx->sq_check_overflow)) {
6822 if (!io_cqring_overflow_flush(ctx, false, NULL, NULL))
6826 /* make sure SQ entry isn't read before tail */
6827 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6829 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6832 percpu_counter_add(¤t->io_uring->inflight, nr);
6833 refcount_add(nr, ¤t->usage);
6835 io_submit_state_start(&state, ctx, nr);
6838 for (i = 0; i < nr; i++) {
6839 const struct io_uring_sqe *sqe;
6840 struct io_kiocb *req;
6843 sqe = io_get_sqe(ctx);
6844 if (unlikely(!sqe)) {
6845 io_consume_sqe(ctx);
6848 req = io_alloc_req(ctx, &state);
6849 if (unlikely(!req)) {
6851 submitted = -EAGAIN;
6854 io_consume_sqe(ctx);
6855 /* will complete beyond this point, count as submitted */
6858 err = io_init_req(ctx, req, sqe, &state);
6859 if (unlikely(err)) {
6862 io_req_complete(req, err);
6866 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6867 true, io_async_submit(ctx));
6868 err = io_submit_sqe(req, sqe, &link, &state.comp);
6873 if (unlikely(submitted != nr)) {
6874 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6875 struct io_uring_task *tctx = current->io_uring;
6876 int unused = nr - ref_used;
6878 percpu_ref_put_many(&ctx->refs, unused);
6879 percpu_counter_sub(&tctx->inflight, unused);
6880 put_task_struct_many(current, unused);
6883 io_queue_link_head(link.head, &state.comp);
6884 io_submit_state_end(&state);
6886 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6887 io_commit_sqring(ctx);
6892 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6894 /* Tell userspace we may need a wakeup call */
6895 spin_lock_irq(&ctx->completion_lock);
6896 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6897 spin_unlock_irq(&ctx->completion_lock);
6900 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6902 spin_lock_irq(&ctx->completion_lock);
6903 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6904 spin_unlock_irq(&ctx->completion_lock);
6907 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
6909 unsigned int to_submit;
6912 to_submit = io_sqring_entries(ctx);
6913 /* if we're handling multiple rings, cap submit size for fairness */
6914 if (cap_entries && to_submit > 8)
6917 if (!list_empty(&ctx->iopoll_list) || to_submit) {
6918 unsigned nr_events = 0;
6920 mutex_lock(&ctx->uring_lock);
6921 if (!list_empty(&ctx->iopoll_list))
6922 io_do_iopoll(ctx, &nr_events, 0);
6924 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)))
6925 ret = io_submit_sqes(ctx, to_submit);
6926 mutex_unlock(&ctx->uring_lock);
6929 if (!io_sqring_full(ctx) && wq_has_sleeper(&ctx->sqo_sq_wait))
6930 wake_up(&ctx->sqo_sq_wait);
6935 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
6937 struct io_ring_ctx *ctx;
6938 unsigned sq_thread_idle = 0;
6940 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6941 if (sq_thread_idle < ctx->sq_thread_idle)
6942 sq_thread_idle = ctx->sq_thread_idle;
6945 sqd->sq_thread_idle = sq_thread_idle;
6948 static void io_sqd_init_new(struct io_sq_data *sqd)
6950 struct io_ring_ctx *ctx;
6952 while (!list_empty(&sqd->ctx_new_list)) {
6953 ctx = list_first_entry(&sqd->ctx_new_list, struct io_ring_ctx, sqd_list);
6954 list_move_tail(&ctx->sqd_list, &sqd->ctx_list);
6955 complete(&ctx->sq_thread_comp);
6958 io_sqd_update_thread_idle(sqd);
6961 static int io_sq_thread(void *data)
6963 struct cgroup_subsys_state *cur_css = NULL;
6964 struct files_struct *old_files = current->files;
6965 struct nsproxy *old_nsproxy = current->nsproxy;
6966 const struct cred *old_cred = NULL;
6967 struct io_sq_data *sqd = data;
6968 struct io_ring_ctx *ctx;
6969 unsigned long timeout = 0;
6973 current->files = NULL;
6974 current->nsproxy = NULL;
6975 task_unlock(current);
6977 while (!kthread_should_stop()) {
6979 bool cap_entries, sqt_spin, needs_sched;
6982 * Any changes to the sqd lists are synchronized through the
6983 * kthread parking. This synchronizes the thread vs users,
6984 * the users are synchronized on the sqd->ctx_lock.
6986 if (kthread_should_park()) {
6989 * When sq thread is unparked, in case the previous park operation
6990 * comes from io_put_sq_data(), which means that sq thread is going
6991 * to be stopped, so here needs to have a check.
6993 if (kthread_should_stop())
6997 if (unlikely(!list_empty(&sqd->ctx_new_list))) {
6998 io_sqd_init_new(sqd);
6999 timeout = jiffies + sqd->sq_thread_idle;
7003 cap_entries = !list_is_singular(&sqd->ctx_list);
7004 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
7005 if (current->cred != ctx->creds) {
7007 revert_creds(old_cred);
7008 old_cred = override_creds(ctx->creds);
7010 io_sq_thread_associate_blkcg(ctx, &cur_css);
7012 current->loginuid = ctx->loginuid;
7013 current->sessionid = ctx->sessionid;
7016 ret = __io_sq_thread(ctx, cap_entries);
7017 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
7020 io_sq_thread_drop_mm_files();
7023 if (sqt_spin || !time_after(jiffies, timeout)) {
7027 timeout = jiffies + sqd->sq_thread_idle;
7031 if (kthread_should_park())
7035 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
7036 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
7037 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
7038 !list_empty_careful(&ctx->iopoll_list)) {
7039 needs_sched = false;
7042 if (io_sqring_entries(ctx)) {
7043 needs_sched = false;
7049 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7050 io_ring_set_wakeup_flag(ctx);
7053 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7054 io_ring_clear_wakeup_flag(ctx);
7057 finish_wait(&sqd->wait, &wait);
7058 timeout = jiffies + sqd->sq_thread_idle;
7064 io_sq_thread_unassociate_blkcg();
7066 revert_creds(old_cred);
7069 current->files = old_files;
7070 current->nsproxy = old_nsproxy;
7071 task_unlock(current);
7078 struct io_wait_queue {
7079 struct wait_queue_entry wq;
7080 struct io_ring_ctx *ctx;
7082 unsigned nr_timeouts;
7085 static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush)
7087 struct io_ring_ctx *ctx = iowq->ctx;
7090 * Wake up if we have enough events, or if a timeout occurred since we
7091 * started waiting. For timeouts, we always want to return to userspace,
7092 * regardless of event count.
7094 return io_cqring_events(ctx, noflush) >= iowq->to_wait ||
7095 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
7098 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
7099 int wake_flags, void *key)
7101 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
7104 /* use noflush == true, as we can't safely rely on locking context */
7105 if (!io_should_wake(iowq, true))
7108 return autoremove_wake_function(curr, mode, wake_flags, key);
7111 static int io_run_task_work_sig(void)
7113 if (io_run_task_work())
7115 if (!signal_pending(current))
7117 if (test_tsk_thread_flag(current, TIF_NOTIFY_SIGNAL))
7118 return -ERESTARTSYS;
7123 * Wait until events become available, if we don't already have some. The
7124 * application must reap them itself, as they reside on the shared cq ring.
7126 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
7127 const sigset_t __user *sig, size_t sigsz,
7128 struct __kernel_timespec __user *uts)
7130 struct io_wait_queue iowq = {
7133 .func = io_wake_function,
7134 .entry = LIST_HEAD_INIT(iowq.wq.entry),
7137 .to_wait = min_events,
7139 struct io_rings *rings = ctx->rings;
7140 struct timespec64 ts;
7141 signed long timeout = 0;
7145 if (io_cqring_events(ctx, false) >= min_events)
7147 if (!io_run_task_work())
7152 #ifdef CONFIG_COMPAT
7153 if (in_compat_syscall())
7154 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
7158 ret = set_user_sigmask(sig, sigsz);
7165 if (get_timespec64(&ts, uts))
7167 timeout = timespec64_to_jiffies(&ts);
7170 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
7171 trace_io_uring_cqring_wait(ctx, min_events);
7173 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
7174 TASK_INTERRUPTIBLE);
7175 /* make sure we run task_work before checking for signals */
7176 ret = io_run_task_work_sig();
7181 if (io_should_wake(&iowq, false))
7184 timeout = schedule_timeout(timeout);
7193 finish_wait(&ctx->wait, &iowq.wq);
7195 restore_saved_sigmask_unless(ret == -EINTR);
7197 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
7200 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
7202 #if defined(CONFIG_UNIX)
7203 if (ctx->ring_sock) {
7204 struct sock *sock = ctx->ring_sock->sk;
7205 struct sk_buff *skb;
7207 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
7213 for (i = 0; i < ctx->nr_user_files; i++) {
7216 file = io_file_from_index(ctx, i);
7223 static void io_file_ref_kill(struct percpu_ref *ref)
7225 struct fixed_file_data *data;
7227 data = container_of(ref, struct fixed_file_data, refs);
7228 complete(&data->done);
7231 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7233 struct fixed_file_data *data = ctx->file_data;
7234 struct fixed_file_ref_node *ref_node = NULL;
7235 unsigned nr_tables, i;
7240 spin_lock_bh(&data->lock);
7241 ref_node = data->node;
7242 spin_unlock_bh(&data->lock);
7244 percpu_ref_kill(&ref_node->refs);
7246 percpu_ref_kill(&data->refs);
7248 /* wait for all refs nodes to complete */
7249 flush_delayed_work(&ctx->file_put_work);
7250 wait_for_completion(&data->done);
7252 __io_sqe_files_unregister(ctx);
7253 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
7254 for (i = 0; i < nr_tables; i++)
7255 kfree(data->table[i].files);
7257 percpu_ref_exit(&data->refs);
7259 ctx->file_data = NULL;
7260 ctx->nr_user_files = 0;
7264 static void io_put_sq_data(struct io_sq_data *sqd)
7266 if (refcount_dec_and_test(&sqd->refs)) {
7268 * The park is a bit of a work-around, without it we get
7269 * warning spews on shutdown with SQPOLL set and affinity
7270 * set to a single CPU.
7273 kthread_park(sqd->thread);
7274 kthread_stop(sqd->thread);
7281 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7283 struct io_ring_ctx *ctx_attach;
7284 struct io_sq_data *sqd;
7287 f = fdget(p->wq_fd);
7289 return ERR_PTR(-ENXIO);
7290 if (f.file->f_op != &io_uring_fops) {
7292 return ERR_PTR(-EINVAL);
7295 ctx_attach = f.file->private_data;
7296 sqd = ctx_attach->sq_data;
7299 return ERR_PTR(-EINVAL);
7302 refcount_inc(&sqd->refs);
7307 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p)
7309 struct io_sq_data *sqd;
7311 if (p->flags & IORING_SETUP_ATTACH_WQ)
7312 return io_attach_sq_data(p);
7314 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7316 return ERR_PTR(-ENOMEM);
7318 refcount_set(&sqd->refs, 1);
7319 INIT_LIST_HEAD(&sqd->ctx_list);
7320 INIT_LIST_HEAD(&sqd->ctx_new_list);
7321 mutex_init(&sqd->ctx_lock);
7322 mutex_init(&sqd->lock);
7323 init_waitqueue_head(&sqd->wait);
7327 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7328 __releases(&sqd->lock)
7332 kthread_unpark(sqd->thread);
7333 mutex_unlock(&sqd->lock);
7336 static void io_sq_thread_park(struct io_sq_data *sqd)
7337 __acquires(&sqd->lock)
7341 mutex_lock(&sqd->lock);
7342 kthread_park(sqd->thread);
7345 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
7347 struct io_sq_data *sqd = ctx->sq_data;
7352 * We may arrive here from the error branch in
7353 * io_sq_offload_create() where the kthread is created
7354 * without being waked up, thus wake it up now to make
7355 * sure the wait will complete.
7357 wake_up_process(sqd->thread);
7358 wait_for_completion(&ctx->sq_thread_comp);
7360 io_sq_thread_park(sqd);
7363 mutex_lock(&sqd->ctx_lock);
7364 list_del(&ctx->sqd_list);
7365 io_sqd_update_thread_idle(sqd);
7366 mutex_unlock(&sqd->ctx_lock);
7369 io_sq_thread_unpark(sqd);
7371 io_put_sq_data(sqd);
7372 ctx->sq_data = NULL;
7376 static void io_finish_async(struct io_ring_ctx *ctx)
7378 io_sq_thread_stop(ctx);
7381 io_wq_destroy(ctx->io_wq);
7386 #if defined(CONFIG_UNIX)
7388 * Ensure the UNIX gc is aware of our file set, so we are certain that
7389 * the io_uring can be safely unregistered on process exit, even if we have
7390 * loops in the file referencing.
7392 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7394 struct sock *sk = ctx->ring_sock->sk;
7395 struct scm_fp_list *fpl;
7396 struct sk_buff *skb;
7399 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7403 skb = alloc_skb(0, GFP_KERNEL);
7412 fpl->user = get_uid(ctx->user);
7413 for (i = 0; i < nr; i++) {
7414 struct file *file = io_file_from_index(ctx, i + offset);
7418 fpl->fp[nr_files] = get_file(file);
7419 unix_inflight(fpl->user, fpl->fp[nr_files]);
7424 fpl->max = SCM_MAX_FD;
7425 fpl->count = nr_files;
7426 UNIXCB(skb).fp = fpl;
7427 skb->destructor = unix_destruct_scm;
7428 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7429 skb_queue_head(&sk->sk_receive_queue, skb);
7431 for (i = 0; i < nr_files; i++)
7442 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7443 * causes regular reference counting to break down. We rely on the UNIX
7444 * garbage collection to take care of this problem for us.
7446 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7448 unsigned left, total;
7452 left = ctx->nr_user_files;
7454 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7456 ret = __io_sqe_files_scm(ctx, this_files, total);
7460 total += this_files;
7466 while (total < ctx->nr_user_files) {
7467 struct file *file = io_file_from_index(ctx, total);
7477 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7483 static int io_sqe_alloc_file_tables(struct fixed_file_data *file_data,
7484 unsigned nr_tables, unsigned nr_files)
7488 for (i = 0; i < nr_tables; i++) {
7489 struct fixed_file_table *table = &file_data->table[i];
7490 unsigned this_files;
7492 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
7493 table->files = kcalloc(this_files, sizeof(struct file *),
7497 nr_files -= this_files;
7503 for (i = 0; i < nr_tables; i++) {
7504 struct fixed_file_table *table = &file_data->table[i];
7505 kfree(table->files);
7510 static void io_ring_file_put(struct io_ring_ctx *ctx, struct file *file)
7512 #if defined(CONFIG_UNIX)
7513 struct sock *sock = ctx->ring_sock->sk;
7514 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7515 struct sk_buff *skb;
7518 __skb_queue_head_init(&list);
7521 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7522 * remove this entry and rearrange the file array.
7524 skb = skb_dequeue(head);
7526 struct scm_fp_list *fp;
7528 fp = UNIXCB(skb).fp;
7529 for (i = 0; i < fp->count; i++) {
7532 if (fp->fp[i] != file)
7535 unix_notinflight(fp->user, fp->fp[i]);
7536 left = fp->count - 1 - i;
7538 memmove(&fp->fp[i], &fp->fp[i + 1],
7539 left * sizeof(struct file *));
7546 __skb_queue_tail(&list, skb);
7556 __skb_queue_tail(&list, skb);
7558 skb = skb_dequeue(head);
7561 if (skb_peek(&list)) {
7562 spin_lock_irq(&head->lock);
7563 while ((skb = __skb_dequeue(&list)) != NULL)
7564 __skb_queue_tail(head, skb);
7565 spin_unlock_irq(&head->lock);
7572 struct io_file_put {
7573 struct list_head list;
7577 static void __io_file_put_work(struct fixed_file_ref_node *ref_node)
7579 struct fixed_file_data *file_data = ref_node->file_data;
7580 struct io_ring_ctx *ctx = file_data->ctx;
7581 struct io_file_put *pfile, *tmp;
7583 list_for_each_entry_safe(pfile, tmp, &ref_node->file_list, list) {
7584 list_del(&pfile->list);
7585 io_ring_file_put(ctx, pfile->file);
7589 percpu_ref_exit(&ref_node->refs);
7591 percpu_ref_put(&file_data->refs);
7594 static void io_file_put_work(struct work_struct *work)
7596 struct io_ring_ctx *ctx;
7597 struct llist_node *node;
7599 ctx = container_of(work, struct io_ring_ctx, file_put_work.work);
7600 node = llist_del_all(&ctx->file_put_llist);
7603 struct fixed_file_ref_node *ref_node;
7604 struct llist_node *next = node->next;
7606 ref_node = llist_entry(node, struct fixed_file_ref_node, llist);
7607 __io_file_put_work(ref_node);
7612 static void io_file_data_ref_zero(struct percpu_ref *ref)
7614 struct fixed_file_ref_node *ref_node;
7615 struct fixed_file_data *data;
7616 struct io_ring_ctx *ctx;
7617 bool first_add = false;
7620 ref_node = container_of(ref, struct fixed_file_ref_node, refs);
7621 data = ref_node->file_data;
7624 spin_lock_bh(&data->lock);
7625 ref_node->done = true;
7627 while (!list_empty(&data->ref_list)) {
7628 ref_node = list_first_entry(&data->ref_list,
7629 struct fixed_file_ref_node, node);
7630 /* recycle ref nodes in order */
7631 if (!ref_node->done)
7633 list_del(&ref_node->node);
7634 first_add |= llist_add(&ref_node->llist, &ctx->file_put_llist);
7636 spin_unlock_bh(&data->lock);
7638 if (percpu_ref_is_dying(&data->refs))
7642 mod_delayed_work(system_wq, &ctx->file_put_work, 0);
7644 queue_delayed_work(system_wq, &ctx->file_put_work, delay);
7647 static struct fixed_file_ref_node *alloc_fixed_file_ref_node(
7648 struct io_ring_ctx *ctx)
7650 struct fixed_file_ref_node *ref_node;
7652 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7654 return ERR_PTR(-ENOMEM);
7656 if (percpu_ref_init(&ref_node->refs, io_file_data_ref_zero,
7659 return ERR_PTR(-ENOMEM);
7661 INIT_LIST_HEAD(&ref_node->node);
7662 INIT_LIST_HEAD(&ref_node->file_list);
7663 ref_node->file_data = ctx->file_data;
7664 ref_node->done = false;
7668 static void destroy_fixed_file_ref_node(struct fixed_file_ref_node *ref_node)
7670 percpu_ref_exit(&ref_node->refs);
7674 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7677 __s32 __user *fds = (__s32 __user *) arg;
7678 unsigned nr_tables, i;
7680 int fd, ret = -ENOMEM;
7681 struct fixed_file_ref_node *ref_node;
7682 struct fixed_file_data *file_data;
7688 if (nr_args > IORING_MAX_FIXED_FILES)
7691 file_data = kzalloc(sizeof(*ctx->file_data), GFP_KERNEL);
7694 file_data->ctx = ctx;
7695 init_completion(&file_data->done);
7696 INIT_LIST_HEAD(&file_data->ref_list);
7697 spin_lock_init(&file_data->lock);
7699 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
7700 file_data->table = kcalloc(nr_tables, sizeof(*file_data->table),
7702 if (!file_data->table)
7705 if (percpu_ref_init(&file_data->refs, io_file_ref_kill,
7706 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
7709 if (io_sqe_alloc_file_tables(file_data, nr_tables, nr_args))
7711 ctx->file_data = file_data;
7713 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7714 struct fixed_file_table *table;
7717 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7721 /* allow sparse sets */
7731 * Don't allow io_uring instances to be registered. If UNIX
7732 * isn't enabled, then this causes a reference cycle and this
7733 * instance can never get freed. If UNIX is enabled we'll
7734 * handle it just fine, but there's still no point in allowing
7735 * a ring fd as it doesn't support regular read/write anyway.
7737 if (file->f_op == &io_uring_fops) {
7741 table = &file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7742 index = i & IORING_FILE_TABLE_MASK;
7743 table->files[index] = file;
7746 ret = io_sqe_files_scm(ctx);
7748 io_sqe_files_unregister(ctx);
7752 ref_node = alloc_fixed_file_ref_node(ctx);
7753 if (IS_ERR(ref_node)) {
7754 io_sqe_files_unregister(ctx);
7755 return PTR_ERR(ref_node);
7758 file_data->node = ref_node;
7759 spin_lock_bh(&file_data->lock);
7760 list_add_tail(&ref_node->node, &file_data->ref_list);
7761 spin_unlock_bh(&file_data->lock);
7762 percpu_ref_get(&file_data->refs);
7765 for (i = 0; i < ctx->nr_user_files; i++) {
7766 file = io_file_from_index(ctx, i);
7770 for (i = 0; i < nr_tables; i++)
7771 kfree(file_data->table[i].files);
7772 ctx->nr_user_files = 0;
7774 percpu_ref_exit(&file_data->refs);
7776 kfree(file_data->table);
7778 ctx->file_data = NULL;
7782 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7785 #if defined(CONFIG_UNIX)
7786 struct sock *sock = ctx->ring_sock->sk;
7787 struct sk_buff_head *head = &sock->sk_receive_queue;
7788 struct sk_buff *skb;
7791 * See if we can merge this file into an existing skb SCM_RIGHTS
7792 * file set. If there's no room, fall back to allocating a new skb
7793 * and filling it in.
7795 spin_lock_irq(&head->lock);
7796 skb = skb_peek(head);
7798 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7800 if (fpl->count < SCM_MAX_FD) {
7801 __skb_unlink(skb, head);
7802 spin_unlock_irq(&head->lock);
7803 fpl->fp[fpl->count] = get_file(file);
7804 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7806 spin_lock_irq(&head->lock);
7807 __skb_queue_head(head, skb);
7812 spin_unlock_irq(&head->lock);
7819 return __io_sqe_files_scm(ctx, 1, index);
7825 static int io_queue_file_removal(struct fixed_file_data *data,
7828 struct io_file_put *pfile;
7829 struct fixed_file_ref_node *ref_node = data->node;
7831 pfile = kzalloc(sizeof(*pfile), GFP_KERNEL);
7836 list_add(&pfile->list, &ref_node->file_list);
7841 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7842 struct io_uring_files_update *up,
7845 struct fixed_file_data *data = ctx->file_data;
7846 struct fixed_file_ref_node *ref_node;
7851 bool needs_switch = false;
7853 if (check_add_overflow(up->offset, nr_args, &done))
7855 if (done > ctx->nr_user_files)
7858 ref_node = alloc_fixed_file_ref_node(ctx);
7859 if (IS_ERR(ref_node))
7860 return PTR_ERR(ref_node);
7863 fds = u64_to_user_ptr(up->fds);
7865 struct fixed_file_table *table;
7869 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
7873 i = array_index_nospec(up->offset, ctx->nr_user_files);
7874 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7875 index = i & IORING_FILE_TABLE_MASK;
7876 if (table->files[index]) {
7877 file = table->files[index];
7878 err = io_queue_file_removal(data, file);
7881 table->files[index] = NULL;
7882 needs_switch = true;
7891 * Don't allow io_uring instances to be registered. If
7892 * UNIX isn't enabled, then this causes a reference
7893 * cycle and this instance can never get freed. If UNIX
7894 * is enabled we'll handle it just fine, but there's
7895 * still no point in allowing a ring fd as it doesn't
7896 * support regular read/write anyway.
7898 if (file->f_op == &io_uring_fops) {
7903 table->files[index] = file;
7904 err = io_sqe_file_register(ctx, file, i);
7906 table->files[index] = NULL;
7917 percpu_ref_kill(&data->node->refs);
7918 spin_lock_bh(&data->lock);
7919 list_add_tail(&ref_node->node, &data->ref_list);
7920 data->node = ref_node;
7921 spin_unlock_bh(&data->lock);
7922 percpu_ref_get(&ctx->file_data->refs);
7924 destroy_fixed_file_ref_node(ref_node);
7926 return done ? done : err;
7929 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
7932 struct io_uring_files_update up;
7934 if (!ctx->file_data)
7938 if (copy_from_user(&up, arg, sizeof(up)))
7943 return __io_sqe_files_update(ctx, &up, nr_args);
7946 static void io_free_work(struct io_wq_work *work)
7948 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7950 /* Consider that io_steal_work() relies on this ref */
7954 static int io_init_wq_offload(struct io_ring_ctx *ctx,
7955 struct io_uring_params *p)
7957 struct io_wq_data data;
7959 struct io_ring_ctx *ctx_attach;
7960 unsigned int concurrency;
7963 data.user = ctx->user;
7964 data.free_work = io_free_work;
7965 data.do_work = io_wq_submit_work;
7967 if (!(p->flags & IORING_SETUP_ATTACH_WQ)) {
7968 /* Do QD, or 4 * CPUS, whatever is smallest */
7969 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7971 ctx->io_wq = io_wq_create(concurrency, &data);
7972 if (IS_ERR(ctx->io_wq)) {
7973 ret = PTR_ERR(ctx->io_wq);
7979 f = fdget(p->wq_fd);
7983 if (f.file->f_op != &io_uring_fops) {
7988 ctx_attach = f.file->private_data;
7989 /* @io_wq is protected by holding the fd */
7990 if (!io_wq_get(ctx_attach->io_wq, &data)) {
7995 ctx->io_wq = ctx_attach->io_wq;
8001 static int io_uring_alloc_task_context(struct task_struct *task)
8003 struct io_uring_task *tctx;
8006 tctx = kmalloc(sizeof(*tctx), GFP_KERNEL);
8007 if (unlikely(!tctx))
8010 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
8011 if (unlikely(ret)) {
8017 init_waitqueue_head(&tctx->wait);
8019 atomic_set(&tctx->in_idle, 0);
8020 tctx->sqpoll = false;
8021 io_init_identity(&tctx->__identity);
8022 tctx->identity = &tctx->__identity;
8023 task->io_uring = tctx;
8027 void __io_uring_free(struct task_struct *tsk)
8029 struct io_uring_task *tctx = tsk->io_uring;
8031 WARN_ON_ONCE(!xa_empty(&tctx->xa));
8032 WARN_ON_ONCE(refcount_read(&tctx->identity->count) != 1);
8033 if (tctx->identity != &tctx->__identity)
8034 kfree(tctx->identity);
8035 percpu_counter_destroy(&tctx->inflight);
8037 tsk->io_uring = NULL;
8040 static int io_sq_offload_create(struct io_ring_ctx *ctx,
8041 struct io_uring_params *p)
8045 if (ctx->flags & IORING_SETUP_SQPOLL) {
8046 struct io_sq_data *sqd;
8049 if (!capable(CAP_SYS_ADMIN) && !capable(CAP_SYS_NICE))
8052 sqd = io_get_sq_data(p);
8059 io_sq_thread_park(sqd);
8060 mutex_lock(&sqd->ctx_lock);
8061 list_add(&ctx->sqd_list, &sqd->ctx_new_list);
8062 mutex_unlock(&sqd->ctx_lock);
8063 io_sq_thread_unpark(sqd);
8065 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
8066 if (!ctx->sq_thread_idle)
8067 ctx->sq_thread_idle = HZ;
8072 if (p->flags & IORING_SETUP_SQ_AFF) {
8073 int cpu = p->sq_thread_cpu;
8076 if (cpu >= nr_cpu_ids)
8078 if (!cpu_online(cpu))
8081 sqd->thread = kthread_create_on_cpu(io_sq_thread, sqd,
8082 cpu, "io_uring-sq");
8084 sqd->thread = kthread_create(io_sq_thread, sqd,
8087 if (IS_ERR(sqd->thread)) {
8088 ret = PTR_ERR(sqd->thread);
8092 ret = io_uring_alloc_task_context(sqd->thread);
8095 } else if (p->flags & IORING_SETUP_SQ_AFF) {
8096 /* Can't have SQ_AFF without SQPOLL */
8102 ret = io_init_wq_offload(ctx, p);
8108 io_finish_async(ctx);
8112 static void io_sq_offload_start(struct io_ring_ctx *ctx)
8114 struct io_sq_data *sqd = ctx->sq_data;
8116 if ((ctx->flags & IORING_SETUP_SQPOLL) && sqd->thread)
8117 wake_up_process(sqd->thread);
8120 static inline void __io_unaccount_mem(struct user_struct *user,
8121 unsigned long nr_pages)
8123 atomic_long_sub(nr_pages, &user->locked_vm);
8126 static inline int __io_account_mem(struct user_struct *user,
8127 unsigned long nr_pages)
8129 unsigned long page_limit, cur_pages, new_pages;
8131 /* Don't allow more pages than we can safely lock */
8132 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
8135 cur_pages = atomic_long_read(&user->locked_vm);
8136 new_pages = cur_pages + nr_pages;
8137 if (new_pages > page_limit)
8139 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
8140 new_pages) != cur_pages);
8145 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
8146 enum io_mem_account acct)
8149 __io_unaccount_mem(ctx->user, nr_pages);
8151 if (ctx->mm_account) {
8152 if (acct == ACCT_LOCKED) {
8153 mmap_write_lock(ctx->mm_account);
8154 ctx->mm_account->locked_vm -= nr_pages;
8155 mmap_write_unlock(ctx->mm_account);
8156 }else if (acct == ACCT_PINNED) {
8157 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
8162 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
8163 enum io_mem_account acct)
8167 if (ctx->limit_mem) {
8168 ret = __io_account_mem(ctx->user, nr_pages);
8173 if (ctx->mm_account) {
8174 if (acct == ACCT_LOCKED) {
8175 mmap_write_lock(ctx->mm_account);
8176 ctx->mm_account->locked_vm += nr_pages;
8177 mmap_write_unlock(ctx->mm_account);
8178 } else if (acct == ACCT_PINNED) {
8179 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
8186 static void io_mem_free(void *ptr)
8193 page = virt_to_head_page(ptr);
8194 if (put_page_testzero(page))
8195 free_compound_page(page);
8198 static void *io_mem_alloc(size_t size)
8200 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
8203 return (void *) __get_free_pages(gfp_flags, get_order(size));
8206 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8209 struct io_rings *rings;
8210 size_t off, sq_array_size;
8212 off = struct_size(rings, cqes, cq_entries);
8213 if (off == SIZE_MAX)
8217 off = ALIGN(off, SMP_CACHE_BYTES);
8225 sq_array_size = array_size(sizeof(u32), sq_entries);
8226 if (sq_array_size == SIZE_MAX)
8229 if (check_add_overflow(off, sq_array_size, &off))
8235 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
8239 pages = (size_t)1 << get_order(
8240 rings_size(sq_entries, cq_entries, NULL));
8241 pages += (size_t)1 << get_order(
8242 array_size(sizeof(struct io_uring_sqe), sq_entries));
8247 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
8251 if (!ctx->user_bufs)
8254 for (i = 0; i < ctx->nr_user_bufs; i++) {
8255 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8257 for (j = 0; j < imu->nr_bvecs; j++)
8258 unpin_user_page(imu->bvec[j].bv_page);
8260 if (imu->acct_pages)
8261 io_unaccount_mem(ctx, imu->acct_pages, ACCT_PINNED);
8266 kfree(ctx->user_bufs);
8267 ctx->user_bufs = NULL;
8268 ctx->nr_user_bufs = 0;
8272 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8273 void __user *arg, unsigned index)
8275 struct iovec __user *src;
8277 #ifdef CONFIG_COMPAT
8279 struct compat_iovec __user *ciovs;
8280 struct compat_iovec ciov;
8282 ciovs = (struct compat_iovec __user *) arg;
8283 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8286 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8287 dst->iov_len = ciov.iov_len;
8291 src = (struct iovec __user *) arg;
8292 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8298 * Not super efficient, but this is just a registration time. And we do cache
8299 * the last compound head, so generally we'll only do a full search if we don't
8302 * We check if the given compound head page has already been accounted, to
8303 * avoid double accounting it. This allows us to account the full size of the
8304 * page, not just the constituent pages of a huge page.
8306 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8307 int nr_pages, struct page *hpage)
8311 /* check current page array */
8312 for (i = 0; i < nr_pages; i++) {
8313 if (!PageCompound(pages[i]))
8315 if (compound_head(pages[i]) == hpage)
8319 /* check previously registered pages */
8320 for (i = 0; i < ctx->nr_user_bufs; i++) {
8321 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8323 for (j = 0; j < imu->nr_bvecs; j++) {
8324 if (!PageCompound(imu->bvec[j].bv_page))
8326 if (compound_head(imu->bvec[j].bv_page) == hpage)
8334 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8335 int nr_pages, struct io_mapped_ubuf *imu,
8336 struct page **last_hpage)
8340 for (i = 0; i < nr_pages; i++) {
8341 if (!PageCompound(pages[i])) {
8346 hpage = compound_head(pages[i]);
8347 if (hpage == *last_hpage)
8349 *last_hpage = hpage;
8350 if (headpage_already_acct(ctx, pages, i, hpage))
8352 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8356 if (!imu->acct_pages)
8359 ret = io_account_mem(ctx, imu->acct_pages, ACCT_PINNED);
8361 imu->acct_pages = 0;
8365 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
8368 struct vm_area_struct **vmas = NULL;
8369 struct page **pages = NULL;
8370 struct page *last_hpage = NULL;
8371 int i, j, got_pages = 0;
8376 if (!nr_args || nr_args > UIO_MAXIOV)
8379 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
8381 if (!ctx->user_bufs)
8384 for (i = 0; i < nr_args; i++) {
8385 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8386 unsigned long off, start, end, ubuf;
8391 ret = io_copy_iov(ctx, &iov, arg, i);
8396 * Don't impose further limits on the size and buffer
8397 * constraints here, we'll -EINVAL later when IO is
8398 * submitted if they are wrong.
8401 if (!iov.iov_base || !iov.iov_len)
8404 /* arbitrary limit, but we need something */
8405 if (iov.iov_len > SZ_1G)
8408 ubuf = (unsigned long) iov.iov_base;
8409 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8410 start = ubuf >> PAGE_SHIFT;
8411 nr_pages = end - start;
8414 if (!pages || nr_pages > got_pages) {
8417 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
8419 vmas = kvmalloc_array(nr_pages,
8420 sizeof(struct vm_area_struct *),
8422 if (!pages || !vmas) {
8426 got_pages = nr_pages;
8429 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
8436 mmap_read_lock(current->mm);
8437 pret = pin_user_pages(ubuf, nr_pages,
8438 FOLL_WRITE | FOLL_LONGTERM,
8440 if (pret == nr_pages) {
8441 /* don't support file backed memory */
8442 for (j = 0; j < nr_pages; j++) {
8443 struct vm_area_struct *vma = vmas[j];
8446 !is_file_hugepages(vma->vm_file)) {
8452 ret = pret < 0 ? pret : -EFAULT;
8454 mmap_read_unlock(current->mm);
8457 * if we did partial map, or found file backed vmas,
8458 * release any pages we did get
8461 unpin_user_pages(pages, pret);
8466 ret = io_buffer_account_pin(ctx, pages, pret, imu, &last_hpage);
8468 unpin_user_pages(pages, pret);
8473 off = ubuf & ~PAGE_MASK;
8475 for (j = 0; j < nr_pages; j++) {
8478 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8479 imu->bvec[j].bv_page = pages[j];
8480 imu->bvec[j].bv_len = vec_len;
8481 imu->bvec[j].bv_offset = off;
8485 /* store original address for later verification */
8487 imu->len = iov.iov_len;
8488 imu->nr_bvecs = nr_pages;
8490 ctx->nr_user_bufs++;
8498 io_sqe_buffer_unregister(ctx);
8502 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8504 __s32 __user *fds = arg;
8510 if (copy_from_user(&fd, fds, sizeof(*fds)))
8513 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8514 if (IS_ERR(ctx->cq_ev_fd)) {
8515 int ret = PTR_ERR(ctx->cq_ev_fd);
8516 ctx->cq_ev_fd = NULL;
8523 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8525 if (ctx->cq_ev_fd) {
8526 eventfd_ctx_put(ctx->cq_ev_fd);
8527 ctx->cq_ev_fd = NULL;
8534 static int __io_destroy_buffers(int id, void *p, void *data)
8536 struct io_ring_ctx *ctx = data;
8537 struct io_buffer *buf = p;
8539 __io_remove_buffers(ctx, buf, id, -1U);
8543 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8545 idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx);
8546 idr_destroy(&ctx->io_buffer_idr);
8549 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8551 io_finish_async(ctx);
8552 io_sqe_buffer_unregister(ctx);
8554 if (ctx->sqo_task) {
8555 put_task_struct(ctx->sqo_task);
8556 ctx->sqo_task = NULL;
8557 mmdrop(ctx->mm_account);
8558 ctx->mm_account = NULL;
8561 #ifdef CONFIG_BLK_CGROUP
8562 if (ctx->sqo_blkcg_css)
8563 css_put(ctx->sqo_blkcg_css);
8566 io_sqe_files_unregister(ctx);
8567 io_eventfd_unregister(ctx);
8568 io_destroy_buffers(ctx);
8569 idr_destroy(&ctx->personality_idr);
8571 #if defined(CONFIG_UNIX)
8572 if (ctx->ring_sock) {
8573 ctx->ring_sock->file = NULL; /* so that iput() is called */
8574 sock_release(ctx->ring_sock);
8578 io_mem_free(ctx->rings);
8579 io_mem_free(ctx->sq_sqes);
8581 percpu_ref_exit(&ctx->refs);
8582 free_uid(ctx->user);
8583 put_cred(ctx->creds);
8584 kfree(ctx->cancel_hash);
8585 kmem_cache_free(req_cachep, ctx->fallback_req);
8589 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8591 struct io_ring_ctx *ctx = file->private_data;
8594 poll_wait(file, &ctx->cq_wait, wait);
8596 * synchronizes with barrier from wq_has_sleeper call in
8600 if (!io_sqring_full(ctx))
8601 mask |= EPOLLOUT | EPOLLWRNORM;
8602 if (io_cqring_events(ctx, false))
8603 mask |= EPOLLIN | EPOLLRDNORM;
8608 static int io_uring_fasync(int fd, struct file *file, int on)
8610 struct io_ring_ctx *ctx = file->private_data;
8612 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8615 static int io_remove_personalities(int id, void *p, void *data)
8617 struct io_ring_ctx *ctx = data;
8618 struct io_identity *iod;
8620 iod = idr_remove(&ctx->personality_idr, id);
8622 put_cred(iod->creds);
8623 if (refcount_dec_and_test(&iod->count))
8629 static void io_ring_exit_work(struct work_struct *work)
8631 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
8635 * If we're doing polled IO and end up having requests being
8636 * submitted async (out-of-line), then completions can come in while
8637 * we're waiting for refs to drop. We need to reap these manually,
8638 * as nobody else will be looking for them.
8641 io_iopoll_try_reap_events(ctx);
8642 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
8643 io_ring_ctx_free(ctx);
8646 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8648 mutex_lock(&ctx->uring_lock);
8649 percpu_ref_kill(&ctx->refs);
8650 /* if force is set, the ring is going away. always drop after that */
8651 ctx->cq_overflow_flushed = 1;
8653 io_cqring_overflow_flush(ctx, true, NULL, NULL);
8654 mutex_unlock(&ctx->uring_lock);
8656 io_kill_timeouts(ctx, NULL, NULL);
8657 io_poll_remove_all(ctx, NULL, NULL);
8660 io_wq_cancel_all(ctx->io_wq);
8662 /* if we failed setting up the ctx, we might not have any rings */
8663 io_iopoll_try_reap_events(ctx);
8664 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
8667 * Do this upfront, so we won't have a grace period where the ring
8668 * is closed but resources aren't reaped yet. This can cause
8669 * spurious failure in setting up a new ring.
8671 io_unaccount_mem(ctx, ring_pages(ctx->sq_entries, ctx->cq_entries),
8674 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8676 * Use system_unbound_wq to avoid spawning tons of event kworkers
8677 * if we're exiting a ton of rings at the same time. It just adds
8678 * noise and overhead, there's no discernable change in runtime
8679 * over using system_wq.
8681 queue_work(system_unbound_wq, &ctx->exit_work);
8684 static int io_uring_release(struct inode *inode, struct file *file)
8686 struct io_ring_ctx *ctx = file->private_data;
8688 file->private_data = NULL;
8689 io_ring_ctx_wait_and_kill(ctx);
8693 struct io_task_cancel {
8694 struct task_struct *task;
8695 struct files_struct *files;
8698 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8700 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8701 struct io_task_cancel *cancel = data;
8704 if (cancel->files && (req->flags & REQ_F_LINK_TIMEOUT)) {
8705 unsigned long flags;
8706 struct io_ring_ctx *ctx = req->ctx;
8708 /* protect against races with linked timeouts */
8709 spin_lock_irqsave(&ctx->completion_lock, flags);
8710 ret = io_match_task(req, cancel->task, cancel->files);
8711 spin_unlock_irqrestore(&ctx->completion_lock, flags);
8713 ret = io_match_task(req, cancel->task, cancel->files);
8718 static void io_cancel_defer_files(struct io_ring_ctx *ctx,
8719 struct task_struct *task,
8720 struct files_struct *files)
8722 struct io_defer_entry *de = NULL;
8725 spin_lock_irq(&ctx->completion_lock);
8726 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8727 if (io_match_task(de->req, task, files)) {
8728 list_cut_position(&list, &ctx->defer_list, &de->list);
8732 spin_unlock_irq(&ctx->completion_lock);
8734 while (!list_empty(&list)) {
8735 de = list_first_entry(&list, struct io_defer_entry, list);
8736 list_del_init(&de->list);
8737 req_set_fail_links(de->req);
8738 io_put_req(de->req);
8739 io_req_complete(de->req, -ECANCELED);
8744 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
8745 struct task_struct *task,
8746 struct files_struct *files)
8748 while (!list_empty_careful(&ctx->inflight_list)) {
8749 struct io_task_cancel cancel = { .task = task, .files = files };
8750 struct io_kiocb *req;
8754 spin_lock_irq(&ctx->inflight_lock);
8755 list_for_each_entry(req, &ctx->inflight_list, inflight_entry) {
8756 if (req->task != task ||
8757 req->work.identity->files != files)
8763 prepare_to_wait(&task->io_uring->wait, &wait,
8764 TASK_UNINTERRUPTIBLE);
8765 spin_unlock_irq(&ctx->inflight_lock);
8767 /* We need to keep going until we don't find a matching req */
8771 io_wq_cancel_cb(ctx->io_wq, io_cancel_task_cb, &cancel, true);
8772 io_poll_remove_all(ctx, task, files);
8773 io_kill_timeouts(ctx, task, files);
8774 /* cancellations _may_ trigger task work */
8777 finish_wait(&task->io_uring->wait, &wait);
8781 static void __io_uring_cancel_task_requests(struct io_ring_ctx *ctx,
8782 struct task_struct *task)
8785 struct io_task_cancel cancel = { .task = task, .files = NULL, };
8786 enum io_wq_cancel cret;
8789 cret = io_wq_cancel_cb(ctx->io_wq, io_cancel_task_cb, &cancel, true);
8790 if (cret != IO_WQ_CANCEL_NOTFOUND)
8793 /* SQPOLL thread does its own polling */
8794 if (!(ctx->flags & IORING_SETUP_SQPOLL)) {
8795 while (!list_empty_careful(&ctx->iopoll_list)) {
8796 io_iopoll_try_reap_events(ctx);
8801 ret |= io_poll_remove_all(ctx, task, NULL);
8802 ret |= io_kill_timeouts(ctx, task, NULL);
8811 * We need to iteratively cancel requests, in case a request has dependent
8812 * hard links. These persist even for failure of cancelations, hence keep
8813 * looping until none are found.
8815 static void io_uring_cancel_task_requests(struct io_ring_ctx *ctx,
8816 struct files_struct *files)
8818 struct task_struct *task = current;
8820 if ((ctx->flags & IORING_SETUP_SQPOLL) && ctx->sq_data) {
8821 task = ctx->sq_data->thread;
8822 atomic_inc(&task->io_uring->in_idle);
8823 io_sq_thread_park(ctx->sq_data);
8826 io_cancel_defer_files(ctx, task, files);
8827 io_ring_submit_lock(ctx, (ctx->flags & IORING_SETUP_IOPOLL));
8828 io_cqring_overflow_flush(ctx, true, task, files);
8829 io_ring_submit_unlock(ctx, (ctx->flags & IORING_SETUP_IOPOLL));
8832 __io_uring_cancel_task_requests(ctx, task);
8834 io_uring_cancel_files(ctx, task, files);
8836 if ((ctx->flags & IORING_SETUP_SQPOLL) && ctx->sq_data) {
8837 atomic_dec(&task->io_uring->in_idle);
8839 * If the files that are going away are the ones in the thread
8840 * identity, clear them out.
8842 if (task->io_uring->identity->files == files)
8843 task->io_uring->identity->files = NULL;
8844 io_sq_thread_unpark(ctx->sq_data);
8849 * Note that this task has used io_uring. We use it for cancelation purposes.
8851 static int io_uring_add_task_file(struct io_ring_ctx *ctx, struct file *file)
8853 struct io_uring_task *tctx = current->io_uring;
8855 if (unlikely(!tctx)) {
8858 ret = io_uring_alloc_task_context(current);
8861 tctx = current->io_uring;
8863 if (tctx->last != file) {
8864 void *old = xa_load(&tctx->xa, (unsigned long)file);
8868 xa_store(&tctx->xa, (unsigned long)file, file, GFP_KERNEL);
8874 * This is race safe in that the task itself is doing this, hence it
8875 * cannot be going through the exit/cancel paths at the same time.
8876 * This cannot be modified while exit/cancel is running.
8878 if (!tctx->sqpoll && (ctx->flags & IORING_SETUP_SQPOLL))
8879 tctx->sqpoll = true;
8885 * Remove this io_uring_file -> task mapping.
8887 static void io_uring_del_task_file(struct file *file)
8889 struct io_uring_task *tctx = current->io_uring;
8891 if (tctx->last == file)
8893 file = xa_erase(&tctx->xa, (unsigned long)file);
8899 * Drop task note for this file if we're the only ones that hold it after
8902 static void io_uring_attempt_task_drop(struct file *file)
8904 if (!current->io_uring)
8907 * fput() is pending, will be 2 if the only other ref is our potential
8908 * task file note. If the task is exiting, drop regardless of count.
8910 if (fatal_signal_pending(current) || (current->flags & PF_EXITING) ||
8911 atomic_long_read(&file->f_count) == 2)
8912 io_uring_del_task_file(file);
8915 void __io_uring_files_cancel(struct files_struct *files)
8917 struct io_uring_task *tctx = current->io_uring;
8919 unsigned long index;
8921 /* make sure overflow events are dropped */
8922 atomic_inc(&tctx->in_idle);
8924 xa_for_each(&tctx->xa, index, file) {
8925 struct io_ring_ctx *ctx = file->private_data;
8927 io_uring_cancel_task_requests(ctx, files);
8929 io_uring_del_task_file(file);
8932 atomic_dec(&tctx->in_idle);
8935 static s64 tctx_inflight(struct io_uring_task *tctx)
8937 unsigned long index;
8941 inflight = percpu_counter_sum(&tctx->inflight);
8946 * If we have SQPOLL rings, then we need to iterate and find them, and
8947 * add the pending count for those.
8949 xa_for_each(&tctx->xa, index, file) {
8950 struct io_ring_ctx *ctx = file->private_data;
8952 if (ctx->flags & IORING_SETUP_SQPOLL) {
8953 struct io_uring_task *__tctx = ctx->sqo_task->io_uring;
8955 inflight += percpu_counter_sum(&__tctx->inflight);
8963 * Find any io_uring fd that this task has registered or done IO on, and cancel
8966 void __io_uring_task_cancel(void)
8968 struct io_uring_task *tctx = current->io_uring;
8972 /* make sure overflow events are dropped */
8973 atomic_inc(&tctx->in_idle);
8976 /* read completions before cancelations */
8977 inflight = tctx_inflight(tctx);
8980 __io_uring_files_cancel(NULL);
8982 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
8985 * If we've seen completions, retry. This avoids a race where
8986 * a completion comes in before we did prepare_to_wait().
8988 if (inflight != tctx_inflight(tctx))
8993 finish_wait(&tctx->wait, &wait);
8994 atomic_dec(&tctx->in_idle);
8997 static int io_uring_flush(struct file *file, void *data)
8999 io_uring_attempt_task_drop(file);
9003 static void *io_uring_validate_mmap_request(struct file *file,
9004 loff_t pgoff, size_t sz)
9006 struct io_ring_ctx *ctx = file->private_data;
9007 loff_t offset = pgoff << PAGE_SHIFT;
9012 case IORING_OFF_SQ_RING:
9013 case IORING_OFF_CQ_RING:
9016 case IORING_OFF_SQES:
9020 return ERR_PTR(-EINVAL);
9023 page = virt_to_head_page(ptr);
9024 if (sz > page_size(page))
9025 return ERR_PTR(-EINVAL);
9032 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9034 size_t sz = vma->vm_end - vma->vm_start;
9038 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
9040 return PTR_ERR(ptr);
9042 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
9043 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
9046 #else /* !CONFIG_MMU */
9048 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9050 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
9053 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
9055 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
9058 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
9059 unsigned long addr, unsigned long len,
9060 unsigned long pgoff, unsigned long flags)
9064 ptr = io_uring_validate_mmap_request(file, pgoff, len);
9066 return PTR_ERR(ptr);
9068 return (unsigned long) ptr;
9071 #endif /* !CONFIG_MMU */
9073 static void io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
9078 if (!io_sqring_full(ctx))
9081 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9083 if (!io_sqring_full(ctx))
9087 } while (!signal_pending(current));
9089 finish_wait(&ctx->sqo_sq_wait, &wait);
9092 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9093 struct __kernel_timespec __user **ts,
9094 const sigset_t __user **sig)
9096 struct io_uring_getevents_arg arg;
9099 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9100 * is just a pointer to the sigset_t.
9102 if (!(flags & IORING_ENTER_EXT_ARG)) {
9103 *sig = (const sigset_t __user *) argp;
9109 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9110 * timespec and sigset_t pointers if good.
9112 if (*argsz != sizeof(arg))
9114 if (copy_from_user(&arg, argp, sizeof(arg)))
9116 *sig = u64_to_user_ptr(arg.sigmask);
9117 *argsz = arg.sigmask_sz;
9118 *ts = u64_to_user_ptr(arg.ts);
9122 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9123 u32, min_complete, u32, flags, const void __user *, argp,
9126 struct io_ring_ctx *ctx;
9133 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9134 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG))
9142 if (f.file->f_op != &io_uring_fops)
9146 ctx = f.file->private_data;
9147 if (!percpu_ref_tryget(&ctx->refs))
9151 if (ctx->flags & IORING_SETUP_R_DISABLED)
9155 * For SQ polling, the thread will do all submissions and completions.
9156 * Just return the requested submit count, and wake the thread if
9160 if (ctx->flags & IORING_SETUP_SQPOLL) {
9161 io_ring_submit_lock(ctx, (ctx->flags & IORING_SETUP_IOPOLL));
9162 if (!list_empty_careful(&ctx->cq_overflow_list))
9163 io_cqring_overflow_flush(ctx, false, NULL, NULL);
9164 io_ring_submit_unlock(ctx, (ctx->flags & IORING_SETUP_IOPOLL));
9165 if (flags & IORING_ENTER_SQ_WAKEUP)
9166 wake_up(&ctx->sq_data->wait);
9167 if (flags & IORING_ENTER_SQ_WAIT)
9168 io_sqpoll_wait_sq(ctx);
9169 submitted = to_submit;
9170 } else if (to_submit) {
9171 ret = io_uring_add_task_file(ctx, f.file);
9174 mutex_lock(&ctx->uring_lock);
9175 submitted = io_submit_sqes(ctx, to_submit);
9176 mutex_unlock(&ctx->uring_lock);
9178 if (submitted != to_submit)
9181 if (flags & IORING_ENTER_GETEVENTS) {
9182 const sigset_t __user *sig;
9183 struct __kernel_timespec __user *ts;
9185 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
9189 min_complete = min(min_complete, ctx->cq_entries);
9192 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9193 * space applications don't need to do io completion events
9194 * polling again, they can rely on io_sq_thread to do polling
9195 * work, which can reduce cpu usage and uring_lock contention.
9197 if (ctx->flags & IORING_SETUP_IOPOLL &&
9198 !(ctx->flags & IORING_SETUP_SQPOLL)) {
9199 ret = io_iopoll_check(ctx, min_complete);
9201 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
9206 percpu_ref_put(&ctx->refs);
9209 return submitted ? submitted : ret;
9212 #ifdef CONFIG_PROC_FS
9213 static int io_uring_show_cred(int id, void *p, void *data)
9215 struct io_identity *iod = p;
9216 const struct cred *cred = iod->creds;
9217 struct seq_file *m = data;
9218 struct user_namespace *uns = seq_user_ns(m);
9219 struct group_info *gi;
9224 seq_printf(m, "%5d\n", id);
9225 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
9226 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
9227 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
9228 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
9229 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
9230 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
9231 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
9232 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
9233 seq_puts(m, "\n\tGroups:\t");
9234 gi = cred->group_info;
9235 for (g = 0; g < gi->ngroups; g++) {
9236 seq_put_decimal_ull(m, g ? " " : "",
9237 from_kgid_munged(uns, gi->gid[g]));
9239 seq_puts(m, "\n\tCapEff:\t");
9240 cap = cred->cap_effective;
9241 CAP_FOR_EACH_U32(__capi)
9242 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
9247 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
9249 struct io_sq_data *sq = NULL;
9254 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9255 * since fdinfo case grabs it in the opposite direction of normal use
9256 * cases. If we fail to get the lock, we just don't iterate any
9257 * structures that could be going away outside the io_uring mutex.
9259 has_lock = mutex_trylock(&ctx->uring_lock);
9261 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL))
9264 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9265 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9266 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9267 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9268 struct fixed_file_table *table;
9271 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
9272 f = table->files[i & IORING_FILE_TABLE_MASK];
9274 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9276 seq_printf(m, "%5u: <none>\n", i);
9278 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9279 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9280 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
9282 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
9283 (unsigned int) buf->len);
9285 if (has_lock && !idr_is_empty(&ctx->personality_idr)) {
9286 seq_printf(m, "Personalities:\n");
9287 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
9289 seq_printf(m, "PollList:\n");
9290 spin_lock_irq(&ctx->completion_lock);
9291 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9292 struct hlist_head *list = &ctx->cancel_hash[i];
9293 struct io_kiocb *req;
9295 hlist_for_each_entry(req, list, hash_node)
9296 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9297 req->task->task_works != NULL);
9299 spin_unlock_irq(&ctx->completion_lock);
9301 mutex_unlock(&ctx->uring_lock);
9304 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9306 struct io_ring_ctx *ctx = f->private_data;
9308 if (percpu_ref_tryget(&ctx->refs)) {
9309 __io_uring_show_fdinfo(ctx, m);
9310 percpu_ref_put(&ctx->refs);
9315 static const struct file_operations io_uring_fops = {
9316 .release = io_uring_release,
9317 .flush = io_uring_flush,
9318 .mmap = io_uring_mmap,
9320 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9321 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9323 .poll = io_uring_poll,
9324 .fasync = io_uring_fasync,
9325 #ifdef CONFIG_PROC_FS
9326 .show_fdinfo = io_uring_show_fdinfo,
9330 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9331 struct io_uring_params *p)
9333 struct io_rings *rings;
9334 size_t size, sq_array_offset;
9336 /* make sure these are sane, as we already accounted them */
9337 ctx->sq_entries = p->sq_entries;
9338 ctx->cq_entries = p->cq_entries;
9340 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9341 if (size == SIZE_MAX)
9344 rings = io_mem_alloc(size);
9349 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9350 rings->sq_ring_mask = p->sq_entries - 1;
9351 rings->cq_ring_mask = p->cq_entries - 1;
9352 rings->sq_ring_entries = p->sq_entries;
9353 rings->cq_ring_entries = p->cq_entries;
9354 ctx->sq_mask = rings->sq_ring_mask;
9355 ctx->cq_mask = rings->cq_ring_mask;
9357 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9358 if (size == SIZE_MAX) {
9359 io_mem_free(ctx->rings);
9364 ctx->sq_sqes = io_mem_alloc(size);
9365 if (!ctx->sq_sqes) {
9366 io_mem_free(ctx->rings);
9375 * Allocate an anonymous fd, this is what constitutes the application
9376 * visible backing of an io_uring instance. The application mmaps this
9377 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9378 * we have to tie this fd to a socket for file garbage collection purposes.
9380 static int io_uring_get_fd(struct io_ring_ctx *ctx)
9386 #if defined(CONFIG_UNIX)
9387 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9393 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9398 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9399 O_RDWR | O_CLOEXEC);
9402 ret = PTR_ERR(file);
9406 #if defined(CONFIG_UNIX)
9407 ctx->ring_sock->file = file;
9409 ret = io_uring_add_task_file(ctx, file);
9415 fd_install(fd, file);
9418 #if defined(CONFIG_UNIX)
9419 sock_release(ctx->ring_sock);
9420 ctx->ring_sock = NULL;
9425 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9426 struct io_uring_params __user *params)
9428 struct user_struct *user = NULL;
9429 struct io_ring_ctx *ctx;
9435 if (entries > IORING_MAX_ENTRIES) {
9436 if (!(p->flags & IORING_SETUP_CLAMP))
9438 entries = IORING_MAX_ENTRIES;
9442 * Use twice as many entries for the CQ ring. It's possible for the
9443 * application to drive a higher depth than the size of the SQ ring,
9444 * since the sqes are only used at submission time. This allows for
9445 * some flexibility in overcommitting a bit. If the application has
9446 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9447 * of CQ ring entries manually.
9449 p->sq_entries = roundup_pow_of_two(entries);
9450 if (p->flags & IORING_SETUP_CQSIZE) {
9452 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9453 * to a power-of-two, if it isn't already. We do NOT impose
9454 * any cq vs sq ring sizing.
9458 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9459 if (!(p->flags & IORING_SETUP_CLAMP))
9461 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9463 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9464 if (p->cq_entries < p->sq_entries)
9467 p->cq_entries = 2 * p->sq_entries;
9470 user = get_uid(current_user());
9471 limit_mem = !capable(CAP_IPC_LOCK);
9474 ret = __io_account_mem(user,
9475 ring_pages(p->sq_entries, p->cq_entries));
9482 ctx = io_ring_ctx_alloc(p);
9485 __io_unaccount_mem(user, ring_pages(p->sq_entries,
9490 ctx->compat = in_compat_syscall();
9492 ctx->creds = get_current_cred();
9494 ctx->loginuid = current->loginuid;
9495 ctx->sessionid = current->sessionid;
9497 ctx->sqo_task = get_task_struct(current);
9500 * This is just grabbed for accounting purposes. When a process exits,
9501 * the mm is exited and dropped before the files, hence we need to hang
9502 * on to this mm purely for the purposes of being able to unaccount
9503 * memory (locked/pinned vm). It's not used for anything else.
9505 mmgrab(current->mm);
9506 ctx->mm_account = current->mm;
9508 #ifdef CONFIG_BLK_CGROUP
9510 * The sq thread will belong to the original cgroup it was inited in.
9511 * If the cgroup goes offline (e.g. disabling the io controller), then
9512 * issued bios will be associated with the closest cgroup later in the
9516 ctx->sqo_blkcg_css = blkcg_css();
9517 ret = css_tryget_online(ctx->sqo_blkcg_css);
9520 /* don't init against a dying cgroup, have the user try again */
9521 ctx->sqo_blkcg_css = NULL;
9528 * Account memory _before_ installing the file descriptor. Once
9529 * the descriptor is installed, it can get closed at any time. Also
9530 * do this before hitting the general error path, as ring freeing
9531 * will un-account as well.
9533 io_account_mem(ctx, ring_pages(p->sq_entries, p->cq_entries),
9535 ctx->limit_mem = limit_mem;
9537 ret = io_allocate_scq_urings(ctx, p);
9541 ret = io_sq_offload_create(ctx, p);
9545 if (!(p->flags & IORING_SETUP_R_DISABLED))
9546 io_sq_offload_start(ctx);
9548 memset(&p->sq_off, 0, sizeof(p->sq_off));
9549 p->sq_off.head = offsetof(struct io_rings, sq.head);
9550 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9551 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9552 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9553 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9554 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9555 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9557 memset(&p->cq_off, 0, sizeof(p->cq_off));
9558 p->cq_off.head = offsetof(struct io_rings, cq.head);
9559 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9560 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9561 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9562 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9563 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9564 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9566 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9567 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9568 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9569 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
9570 IORING_FEAT_EXT_ARG;
9572 if (copy_to_user(params, p, sizeof(*p))) {
9578 * Install ring fd as the very last thing, so we don't risk someone
9579 * having closed it before we finish setup
9581 ret = io_uring_get_fd(ctx);
9585 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9588 io_ring_ctx_wait_and_kill(ctx);
9593 * Sets up an aio uring context, and returns the fd. Applications asks for a
9594 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9595 * params structure passed in.
9597 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9599 struct io_uring_params p;
9602 if (copy_from_user(&p, params, sizeof(p)))
9604 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9609 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9610 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9611 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9612 IORING_SETUP_R_DISABLED))
9615 return io_uring_create(entries, &p, params);
9618 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9619 struct io_uring_params __user *, params)
9621 return io_uring_setup(entries, params);
9624 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9626 struct io_uring_probe *p;
9630 size = struct_size(p, ops, nr_args);
9631 if (size == SIZE_MAX)
9633 p = kzalloc(size, GFP_KERNEL);
9638 if (copy_from_user(p, arg, size))
9641 if (memchr_inv(p, 0, size))
9644 p->last_op = IORING_OP_LAST - 1;
9645 if (nr_args > IORING_OP_LAST)
9646 nr_args = IORING_OP_LAST;
9648 for (i = 0; i < nr_args; i++) {
9650 if (!io_op_defs[i].not_supported)
9651 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9656 if (copy_to_user(arg, p, size))
9663 static int io_register_personality(struct io_ring_ctx *ctx)
9665 struct io_identity *id;
9668 id = kmalloc(sizeof(*id), GFP_KERNEL);
9672 io_init_identity(id);
9673 id->creds = get_current_cred();
9675 ret = idr_alloc_cyclic(&ctx->personality_idr, id, 1, USHRT_MAX, GFP_KERNEL);
9677 put_cred(id->creds);
9683 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
9685 struct io_identity *iod;
9687 iod = idr_remove(&ctx->personality_idr, id);
9689 put_cred(iod->creds);
9690 if (refcount_dec_and_test(&iod->count))
9698 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9699 unsigned int nr_args)
9701 struct io_uring_restriction *res;
9705 /* Restrictions allowed only if rings started disabled */
9706 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9709 /* We allow only a single restrictions registration */
9710 if (ctx->restrictions.registered)
9713 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9716 size = array_size(nr_args, sizeof(*res));
9717 if (size == SIZE_MAX)
9720 res = memdup_user(arg, size);
9722 return PTR_ERR(res);
9726 for (i = 0; i < nr_args; i++) {
9727 switch (res[i].opcode) {
9728 case IORING_RESTRICTION_REGISTER_OP:
9729 if (res[i].register_op >= IORING_REGISTER_LAST) {
9734 __set_bit(res[i].register_op,
9735 ctx->restrictions.register_op);
9737 case IORING_RESTRICTION_SQE_OP:
9738 if (res[i].sqe_op >= IORING_OP_LAST) {
9743 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9745 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9746 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9748 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9749 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9758 /* Reset all restrictions if an error happened */
9760 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9762 ctx->restrictions.registered = true;
9768 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9770 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9773 if (ctx->restrictions.registered)
9774 ctx->restricted = 1;
9776 ctx->flags &= ~IORING_SETUP_R_DISABLED;
9778 io_sq_offload_start(ctx);
9783 static bool io_register_op_must_quiesce(int op)
9786 case IORING_UNREGISTER_FILES:
9787 case IORING_REGISTER_FILES_UPDATE:
9788 case IORING_REGISTER_PROBE:
9789 case IORING_REGISTER_PERSONALITY:
9790 case IORING_UNREGISTER_PERSONALITY:
9797 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
9798 void __user *arg, unsigned nr_args)
9799 __releases(ctx->uring_lock)
9800 __acquires(ctx->uring_lock)
9805 * We're inside the ring mutex, if the ref is already dying, then
9806 * someone else killed the ctx or is already going through
9807 * io_uring_register().
9809 if (percpu_ref_is_dying(&ctx->refs))
9812 if (io_register_op_must_quiesce(opcode)) {
9813 percpu_ref_kill(&ctx->refs);
9816 * Drop uring mutex before waiting for references to exit. If
9817 * another thread is currently inside io_uring_enter() it might
9818 * need to grab the uring_lock to make progress. If we hold it
9819 * here across the drain wait, then we can deadlock. It's safe
9820 * to drop the mutex here, since no new references will come in
9821 * after we've killed the percpu ref.
9823 mutex_unlock(&ctx->uring_lock);
9825 ret = wait_for_completion_interruptible(&ctx->ref_comp);
9828 ret = io_run_task_work_sig();
9833 mutex_lock(&ctx->uring_lock);
9836 percpu_ref_resurrect(&ctx->refs);
9841 if (ctx->restricted) {
9842 if (opcode >= IORING_REGISTER_LAST) {
9847 if (!test_bit(opcode, ctx->restrictions.register_op)) {
9854 case IORING_REGISTER_BUFFERS:
9855 ret = io_sqe_buffer_register(ctx, arg, nr_args);
9857 case IORING_UNREGISTER_BUFFERS:
9861 ret = io_sqe_buffer_unregister(ctx);
9863 case IORING_REGISTER_FILES:
9864 ret = io_sqe_files_register(ctx, arg, nr_args);
9866 case IORING_UNREGISTER_FILES:
9870 ret = io_sqe_files_unregister(ctx);
9872 case IORING_REGISTER_FILES_UPDATE:
9873 ret = io_sqe_files_update(ctx, arg, nr_args);
9875 case IORING_REGISTER_EVENTFD:
9876 case IORING_REGISTER_EVENTFD_ASYNC:
9880 ret = io_eventfd_register(ctx, arg);
9883 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
9884 ctx->eventfd_async = 1;
9886 ctx->eventfd_async = 0;
9888 case IORING_UNREGISTER_EVENTFD:
9892 ret = io_eventfd_unregister(ctx);
9894 case IORING_REGISTER_PROBE:
9896 if (!arg || nr_args > 256)
9898 ret = io_probe(ctx, arg, nr_args);
9900 case IORING_REGISTER_PERSONALITY:
9904 ret = io_register_personality(ctx);
9906 case IORING_UNREGISTER_PERSONALITY:
9910 ret = io_unregister_personality(ctx, nr_args);
9912 case IORING_REGISTER_ENABLE_RINGS:
9916 ret = io_register_enable_rings(ctx);
9918 case IORING_REGISTER_RESTRICTIONS:
9919 ret = io_register_restrictions(ctx, arg, nr_args);
9927 if (io_register_op_must_quiesce(opcode)) {
9928 /* bring the ctx back to life */
9929 percpu_ref_reinit(&ctx->refs);
9931 reinit_completion(&ctx->ref_comp);
9936 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
9937 void __user *, arg, unsigned int, nr_args)
9939 struct io_ring_ctx *ctx;
9948 if (f.file->f_op != &io_uring_fops)
9951 ctx = f.file->private_data;
9953 mutex_lock(&ctx->uring_lock);
9954 ret = __io_uring_register(ctx, opcode, arg, nr_args);
9955 mutex_unlock(&ctx->uring_lock);
9956 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
9957 ctx->cq_ev_fd != NULL, ret);
9963 static int __init io_uring_init(void)
9965 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
9966 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
9967 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
9970 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
9971 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
9972 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
9973 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
9974 BUILD_BUG_SQE_ELEM(1, __u8, flags);
9975 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
9976 BUILD_BUG_SQE_ELEM(4, __s32, fd);
9977 BUILD_BUG_SQE_ELEM(8, __u64, off);
9978 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
9979 BUILD_BUG_SQE_ELEM(16, __u64, addr);
9980 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
9981 BUILD_BUG_SQE_ELEM(24, __u32, len);
9982 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
9983 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
9984 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
9985 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
9986 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
9987 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
9988 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
9989 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
9990 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
9991 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
9992 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
9993 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
9994 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
9995 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
9996 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
9997 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
9998 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
9999 BUILD_BUG_SQE_ELEM(42, __u16, personality);
10000 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
10002 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
10003 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
10004 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
10007 __initcall(io_uring_init);