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 /* Returns true if there are no backlogged entries after the flush */
1712 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force,
1713 struct task_struct *tsk,
1714 struct files_struct *files)
1716 struct io_rings *rings = ctx->rings;
1717 struct io_kiocb *req, *tmp;
1718 struct io_uring_cqe *cqe;
1719 unsigned long flags;
1723 if (!force && __io_cqring_events(ctx) == rings->cq_ring_entries)
1726 spin_lock_irqsave(&ctx->completion_lock, flags);
1727 list_for_each_entry_safe(req, tmp, &ctx->cq_overflow_list, compl.list) {
1728 if (!io_match_task(req, tsk, files))
1731 cqe = io_get_cqring(ctx);
1735 list_move(&req->compl.list, &list);
1737 WRITE_ONCE(cqe->user_data, req->user_data);
1738 WRITE_ONCE(cqe->res, req->result);
1739 WRITE_ONCE(cqe->flags, req->compl.cflags);
1741 ctx->cached_cq_overflow++;
1742 WRITE_ONCE(ctx->rings->cq_overflow,
1743 ctx->cached_cq_overflow);
1747 all_flushed = list_empty(&ctx->cq_overflow_list);
1749 clear_bit(0, &ctx->sq_check_overflow);
1750 clear_bit(0, &ctx->cq_check_overflow);
1751 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
1754 io_commit_cqring(ctx);
1755 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1756 io_cqring_ev_posted(ctx);
1758 while (!list_empty(&list)) {
1759 req = list_first_entry(&list, struct io_kiocb, compl.list);
1760 list_del(&req->compl.list);
1767 static void __io_cqring_fill_event(struct io_kiocb *req, long res, long cflags)
1769 struct io_ring_ctx *ctx = req->ctx;
1770 struct io_uring_cqe *cqe;
1772 trace_io_uring_complete(ctx, req->user_data, res);
1775 * If we can't get a cq entry, userspace overflowed the
1776 * submission (by quite a lot). Increment the overflow count in
1779 cqe = io_get_cqring(ctx);
1781 WRITE_ONCE(cqe->user_data, req->user_data);
1782 WRITE_ONCE(cqe->res, res);
1783 WRITE_ONCE(cqe->flags, cflags);
1784 } else if (ctx->cq_overflow_flushed ||
1785 atomic_read(&req->task->io_uring->in_idle)) {
1787 * If we're in ring overflow flush mode, or in task cancel mode,
1788 * then we cannot store the request for later flushing, we need
1789 * to drop it on the floor.
1791 ctx->cached_cq_overflow++;
1792 WRITE_ONCE(ctx->rings->cq_overflow, ctx->cached_cq_overflow);
1794 if (list_empty(&ctx->cq_overflow_list)) {
1795 set_bit(0, &ctx->sq_check_overflow);
1796 set_bit(0, &ctx->cq_check_overflow);
1797 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
1801 req->compl.cflags = cflags;
1802 refcount_inc(&req->refs);
1803 list_add_tail(&req->compl.list, &ctx->cq_overflow_list);
1807 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1809 __io_cqring_fill_event(req, res, 0);
1812 static void io_cqring_add_event(struct io_kiocb *req, long res, long cflags)
1814 struct io_ring_ctx *ctx = req->ctx;
1815 unsigned long flags;
1817 spin_lock_irqsave(&ctx->completion_lock, flags);
1818 __io_cqring_fill_event(req, res, cflags);
1819 io_commit_cqring(ctx);
1820 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1822 io_cqring_ev_posted(ctx);
1825 static void io_submit_flush_completions(struct io_comp_state *cs)
1827 struct io_ring_ctx *ctx = cs->ctx;
1829 spin_lock_irq(&ctx->completion_lock);
1830 while (!list_empty(&cs->list)) {
1831 struct io_kiocb *req;
1833 req = list_first_entry(&cs->list, struct io_kiocb, compl.list);
1834 list_del(&req->compl.list);
1835 __io_cqring_fill_event(req, req->result, req->compl.cflags);
1838 * io_free_req() doesn't care about completion_lock unless one
1839 * of these flags is set. REQ_F_WORK_INITIALIZED is in the list
1840 * because of a potential deadlock with req->work.fs->lock
1842 if (req->flags & (REQ_F_FAIL_LINK|REQ_F_LINK_TIMEOUT
1843 |REQ_F_WORK_INITIALIZED)) {
1844 spin_unlock_irq(&ctx->completion_lock);
1846 spin_lock_irq(&ctx->completion_lock);
1851 io_commit_cqring(ctx);
1852 spin_unlock_irq(&ctx->completion_lock);
1854 io_cqring_ev_posted(ctx);
1858 static void __io_req_complete(struct io_kiocb *req, long res, unsigned cflags,
1859 struct io_comp_state *cs)
1862 io_cqring_add_event(req, res, cflags);
1867 req->compl.cflags = cflags;
1868 list_add_tail(&req->compl.list, &cs->list);
1870 io_submit_flush_completions(cs);
1874 static void io_req_complete(struct io_kiocb *req, long res)
1876 __io_req_complete(req, res, 0, NULL);
1879 static inline bool io_is_fallback_req(struct io_kiocb *req)
1881 return req == (struct io_kiocb *)
1882 ((unsigned long) req->ctx->fallback_req & ~1UL);
1885 static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx)
1887 struct io_kiocb *req;
1889 req = ctx->fallback_req;
1890 if (!test_and_set_bit_lock(0, (unsigned long *) &ctx->fallback_req))
1896 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx,
1897 struct io_submit_state *state)
1899 if (!state->free_reqs) {
1900 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1904 sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs));
1905 ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs);
1908 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1909 * retry single alloc to be on the safe side.
1911 if (unlikely(ret <= 0)) {
1912 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1913 if (!state->reqs[0])
1917 state->free_reqs = ret;
1921 return state->reqs[state->free_reqs];
1923 return io_get_fallback_req(ctx);
1926 static inline void io_put_file(struct io_kiocb *req, struct file *file,
1933 static void io_dismantle_req(struct io_kiocb *req)
1937 if (req->async_data)
1938 kfree(req->async_data);
1940 io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
1941 if (req->fixed_file_refs)
1942 percpu_ref_put(req->fixed_file_refs);
1943 io_req_clean_work(req);
1946 static void __io_free_req(struct io_kiocb *req)
1948 struct io_uring_task *tctx = req->task->io_uring;
1949 struct io_ring_ctx *ctx = req->ctx;
1951 io_dismantle_req(req);
1953 percpu_counter_dec(&tctx->inflight);
1954 if (atomic_read(&tctx->in_idle))
1955 wake_up(&tctx->wait);
1956 put_task_struct(req->task);
1958 if (likely(!io_is_fallback_req(req)))
1959 kmem_cache_free(req_cachep, req);
1961 clear_bit_unlock(0, (unsigned long *) &ctx->fallback_req);
1962 percpu_ref_put(&ctx->refs);
1965 static inline void io_remove_next_linked(struct io_kiocb *req)
1967 struct io_kiocb *nxt = req->link;
1969 req->link = nxt->link;
1973 static void io_kill_linked_timeout(struct io_kiocb *req)
1975 struct io_ring_ctx *ctx = req->ctx;
1976 struct io_kiocb *link;
1977 bool cancelled = false;
1978 unsigned long flags;
1980 spin_lock_irqsave(&ctx->completion_lock, flags);
1984 * Can happen if a linked timeout fired and link had been like
1985 * req -> link t-out -> link t-out [-> ...]
1987 if (link && (link->flags & REQ_F_LTIMEOUT_ACTIVE)) {
1988 struct io_timeout_data *io = link->async_data;
1991 io_remove_next_linked(req);
1992 link->timeout.head = NULL;
1993 ret = hrtimer_try_to_cancel(&io->timer);
1995 io_cqring_fill_event(link, -ECANCELED);
1996 io_commit_cqring(ctx);
2000 req->flags &= ~REQ_F_LINK_TIMEOUT;
2001 spin_unlock_irqrestore(&ctx->completion_lock, flags);
2004 io_cqring_ev_posted(ctx);
2010 static void io_fail_links(struct io_kiocb *req)
2012 struct io_kiocb *link, *nxt;
2013 struct io_ring_ctx *ctx = req->ctx;
2014 unsigned long flags;
2016 spin_lock_irqsave(&ctx->completion_lock, flags);
2024 trace_io_uring_fail_link(req, link);
2025 io_cqring_fill_event(link, -ECANCELED);
2028 * It's ok to free under spinlock as they're not linked anymore,
2029 * but avoid REQ_F_WORK_INITIALIZED because it may deadlock on
2032 if (link->flags & REQ_F_WORK_INITIALIZED)
2033 io_put_req_deferred(link, 2);
2035 io_double_put_req(link);
2038 io_commit_cqring(ctx);
2039 spin_unlock_irqrestore(&ctx->completion_lock, flags);
2041 io_cqring_ev_posted(ctx);
2044 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
2046 if (req->flags & REQ_F_LINK_TIMEOUT)
2047 io_kill_linked_timeout(req);
2050 * If LINK is set, we have dependent requests in this chain. If we
2051 * didn't fail this request, queue the first one up, moving any other
2052 * dependencies to the next request. In case of failure, fail the rest
2055 if (likely(!(req->flags & REQ_F_FAIL_LINK))) {
2056 struct io_kiocb *nxt = req->link;
2065 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
2067 if (likely(!(req->link) && !(req->flags & REQ_F_LINK_TIMEOUT)))
2069 return __io_req_find_next(req);
2072 static int io_req_task_work_add(struct io_kiocb *req)
2074 struct task_struct *tsk = req->task;
2075 struct io_ring_ctx *ctx = req->ctx;
2076 enum task_work_notify_mode notify;
2079 if (tsk->flags & PF_EXITING)
2083 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
2084 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
2085 * processing task_work. There's no reliable way to tell if TWA_RESUME
2089 if (!(ctx->flags & IORING_SETUP_SQPOLL))
2090 notify = TWA_SIGNAL;
2092 ret = task_work_add(tsk, &req->task_work, notify);
2094 wake_up_process(tsk);
2099 static void __io_req_task_cancel(struct io_kiocb *req, int error)
2101 struct io_ring_ctx *ctx = req->ctx;
2103 spin_lock_irq(&ctx->completion_lock);
2104 io_cqring_fill_event(req, error);
2105 io_commit_cqring(ctx);
2106 spin_unlock_irq(&ctx->completion_lock);
2108 io_cqring_ev_posted(ctx);
2109 req_set_fail_links(req);
2110 io_double_put_req(req);
2113 static void io_req_task_cancel(struct callback_head *cb)
2115 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2116 struct io_ring_ctx *ctx = req->ctx;
2118 __io_req_task_cancel(req, -ECANCELED);
2119 percpu_ref_put(&ctx->refs);
2122 static void __io_req_task_submit(struct io_kiocb *req)
2124 struct io_ring_ctx *ctx = req->ctx;
2126 if (!__io_sq_thread_acquire_mm(ctx) &&
2127 !__io_sq_thread_acquire_files(ctx)) {
2128 mutex_lock(&ctx->uring_lock);
2129 __io_queue_sqe(req, NULL);
2130 mutex_unlock(&ctx->uring_lock);
2132 __io_req_task_cancel(req, -EFAULT);
2136 static void io_req_task_submit(struct callback_head *cb)
2138 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2139 struct io_ring_ctx *ctx = req->ctx;
2141 __io_req_task_submit(req);
2142 percpu_ref_put(&ctx->refs);
2145 static void io_req_task_queue(struct io_kiocb *req)
2149 init_task_work(&req->task_work, io_req_task_submit);
2150 percpu_ref_get(&req->ctx->refs);
2152 ret = io_req_task_work_add(req);
2153 if (unlikely(ret)) {
2154 struct task_struct *tsk;
2156 init_task_work(&req->task_work, io_req_task_cancel);
2157 tsk = io_wq_get_task(req->ctx->io_wq);
2158 task_work_add(tsk, &req->task_work, TWA_NONE);
2159 wake_up_process(tsk);
2163 static inline void io_queue_next(struct io_kiocb *req)
2165 struct io_kiocb *nxt = io_req_find_next(req);
2168 io_req_task_queue(nxt);
2171 static void io_free_req(struct io_kiocb *req)
2178 void *reqs[IO_IOPOLL_BATCH];
2181 struct task_struct *task;
2185 static inline void io_init_req_batch(struct req_batch *rb)
2192 static void __io_req_free_batch_flush(struct io_ring_ctx *ctx,
2193 struct req_batch *rb)
2195 kmem_cache_free_bulk(req_cachep, rb->to_free, rb->reqs);
2196 percpu_ref_put_many(&ctx->refs, rb->to_free);
2200 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2201 struct req_batch *rb)
2204 __io_req_free_batch_flush(ctx, rb);
2206 struct io_uring_task *tctx = rb->task->io_uring;
2208 percpu_counter_sub(&tctx->inflight, rb->task_refs);
2209 put_task_struct_many(rb->task, rb->task_refs);
2214 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req)
2216 if (unlikely(io_is_fallback_req(req))) {
2222 if (req->task != rb->task) {
2224 struct io_uring_task *tctx = rb->task->io_uring;
2226 percpu_counter_sub(&tctx->inflight, rb->task_refs);
2227 put_task_struct_many(rb->task, rb->task_refs);
2229 rb->task = req->task;
2234 io_dismantle_req(req);
2235 rb->reqs[rb->to_free++] = req;
2236 if (unlikely(rb->to_free == ARRAY_SIZE(rb->reqs)))
2237 __io_req_free_batch_flush(req->ctx, rb);
2241 * Drop reference to request, return next in chain (if there is one) if this
2242 * was the last reference to this request.
2244 static struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2246 struct io_kiocb *nxt = NULL;
2248 if (refcount_dec_and_test(&req->refs)) {
2249 nxt = io_req_find_next(req);
2255 static void io_put_req(struct io_kiocb *req)
2257 if (refcount_dec_and_test(&req->refs))
2261 static void io_put_req_deferred_cb(struct callback_head *cb)
2263 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2268 static void io_free_req_deferred(struct io_kiocb *req)
2272 init_task_work(&req->task_work, io_put_req_deferred_cb);
2273 ret = io_req_task_work_add(req);
2274 if (unlikely(ret)) {
2275 struct task_struct *tsk;
2277 tsk = io_wq_get_task(req->ctx->io_wq);
2278 task_work_add(tsk, &req->task_work, TWA_NONE);
2279 wake_up_process(tsk);
2283 static inline void io_put_req_deferred(struct io_kiocb *req, int refs)
2285 if (refcount_sub_and_test(refs, &req->refs))
2286 io_free_req_deferred(req);
2289 static struct io_wq_work *io_steal_work(struct io_kiocb *req)
2291 struct io_kiocb *nxt;
2294 * A ref is owned by io-wq in which context we're. So, if that's the
2295 * last one, it's safe to steal next work. False negatives are Ok,
2296 * it just will be re-punted async in io_put_work()
2298 if (refcount_read(&req->refs) != 1)
2301 nxt = io_req_find_next(req);
2302 return nxt ? &nxt->work : NULL;
2305 static void io_double_put_req(struct io_kiocb *req)
2307 /* drop both submit and complete references */
2308 if (refcount_sub_and_test(2, &req->refs))
2312 static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush)
2314 if (test_bit(0, &ctx->cq_check_overflow)) {
2316 * noflush == true is from the waitqueue handler, just ensure
2317 * we wake up the task, and the next invocation will flush the
2318 * entries. We cannot safely to it from here.
2323 io_cqring_overflow_flush(ctx, false, NULL, NULL);
2326 /* See comment at the top of this file */
2328 return __io_cqring_events(ctx);
2331 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2333 struct io_rings *rings = ctx->rings;
2335 /* make sure SQ entry isn't read before tail */
2336 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2339 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2341 unsigned int cflags;
2343 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2344 cflags |= IORING_CQE_F_BUFFER;
2345 req->flags &= ~REQ_F_BUFFER_SELECTED;
2350 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2352 struct io_buffer *kbuf;
2354 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2355 return io_put_kbuf(req, kbuf);
2358 static inline bool io_run_task_work(void)
2361 * Not safe to run on exiting task, and the task_work handling will
2362 * not add work to such a task.
2364 if (unlikely(current->flags & PF_EXITING))
2366 if (current->task_works) {
2367 __set_current_state(TASK_RUNNING);
2375 static void io_iopoll_queue(struct list_head *again)
2377 struct io_kiocb *req;
2380 req = list_first_entry(again, struct io_kiocb, inflight_entry);
2381 list_del(&req->inflight_entry);
2382 __io_complete_rw(req, -EAGAIN, 0, NULL);
2383 } while (!list_empty(again));
2387 * Find and free completed poll iocbs
2389 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2390 struct list_head *done)
2392 struct req_batch rb;
2393 struct io_kiocb *req;
2396 /* order with ->result store in io_complete_rw_iopoll() */
2399 io_init_req_batch(&rb);
2400 while (!list_empty(done)) {
2403 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2404 if (READ_ONCE(req->result) == -EAGAIN) {
2406 req->iopoll_completed = 0;
2407 list_move_tail(&req->inflight_entry, &again);
2410 list_del(&req->inflight_entry);
2412 if (req->flags & REQ_F_BUFFER_SELECTED)
2413 cflags = io_put_rw_kbuf(req);
2415 __io_cqring_fill_event(req, req->result, cflags);
2418 if (refcount_dec_and_test(&req->refs))
2419 io_req_free_batch(&rb, req);
2422 io_commit_cqring(ctx);
2423 if (ctx->flags & IORING_SETUP_SQPOLL)
2424 io_cqring_ev_posted(ctx);
2425 io_req_free_batch_finish(ctx, &rb);
2427 if (!list_empty(&again))
2428 io_iopoll_queue(&again);
2431 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2434 struct io_kiocb *req, *tmp;
2440 * Only spin for completions if we don't have multiple devices hanging
2441 * off our complete list, and we're under the requested amount.
2443 spin = !ctx->poll_multi_file && *nr_events < min;
2446 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2447 struct kiocb *kiocb = &req->rw.kiocb;
2450 * Move completed and retryable entries to our local lists.
2451 * If we find a request that requires polling, break out
2452 * and complete those lists first, if we have entries there.
2454 if (READ_ONCE(req->iopoll_completed)) {
2455 list_move_tail(&req->inflight_entry, &done);
2458 if (!list_empty(&done))
2461 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2465 /* iopoll may have completed current req */
2466 if (READ_ONCE(req->iopoll_completed))
2467 list_move_tail(&req->inflight_entry, &done);
2474 if (!list_empty(&done))
2475 io_iopoll_complete(ctx, nr_events, &done);
2481 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2482 * non-spinning poll check - we'll still enter the driver poll loop, but only
2483 * as a non-spinning completion check.
2485 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
2488 while (!list_empty(&ctx->iopoll_list) && !need_resched()) {
2491 ret = io_do_iopoll(ctx, nr_events, min);
2494 if (*nr_events >= min)
2502 * We can't just wait for polled events to come to us, we have to actively
2503 * find and complete them.
2505 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2507 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2510 mutex_lock(&ctx->uring_lock);
2511 while (!list_empty(&ctx->iopoll_list)) {
2512 unsigned int nr_events = 0;
2514 io_do_iopoll(ctx, &nr_events, 0);
2516 /* let it sleep and repeat later if can't complete a request */
2520 * Ensure we allow local-to-the-cpu processing to take place,
2521 * in this case we need to ensure that we reap all events.
2522 * Also let task_work, etc. to progress by releasing the mutex
2524 if (need_resched()) {
2525 mutex_unlock(&ctx->uring_lock);
2527 mutex_lock(&ctx->uring_lock);
2530 mutex_unlock(&ctx->uring_lock);
2533 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2535 unsigned int nr_events = 0;
2536 int iters = 0, ret = 0;
2539 * We disallow the app entering submit/complete with polling, but we
2540 * still need to lock the ring to prevent racing with polled issue
2541 * that got punted to a workqueue.
2543 mutex_lock(&ctx->uring_lock);
2546 * Don't enter poll loop if we already have events pending.
2547 * If we do, we can potentially be spinning for commands that
2548 * already triggered a CQE (eg in error).
2550 if (io_cqring_events(ctx, false))
2554 * If a submit got punted to a workqueue, we can have the
2555 * application entering polling for a command before it gets
2556 * issued. That app will hold the uring_lock for the duration
2557 * of the poll right here, so we need to take a breather every
2558 * now and then to ensure that the issue has a chance to add
2559 * the poll to the issued list. Otherwise we can spin here
2560 * forever, while the workqueue is stuck trying to acquire the
2563 if (!(++iters & 7)) {
2564 mutex_unlock(&ctx->uring_lock);
2566 mutex_lock(&ctx->uring_lock);
2569 ret = io_iopoll_getevents(ctx, &nr_events, min);
2573 } while (min && !nr_events && !need_resched());
2575 mutex_unlock(&ctx->uring_lock);
2579 static void kiocb_end_write(struct io_kiocb *req)
2582 * Tell lockdep we inherited freeze protection from submission
2585 if (req->flags & REQ_F_ISREG) {
2586 struct inode *inode = file_inode(req->file);
2588 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
2590 file_end_write(req->file);
2593 static void io_complete_rw_common(struct kiocb *kiocb, long res,
2594 struct io_comp_state *cs)
2596 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2599 if (kiocb->ki_flags & IOCB_WRITE)
2600 kiocb_end_write(req);
2602 if (res != req->result)
2603 req_set_fail_links(req);
2604 if (req->flags & REQ_F_BUFFER_SELECTED)
2605 cflags = io_put_rw_kbuf(req);
2606 __io_req_complete(req, res, cflags, cs);
2610 static bool io_resubmit_prep(struct io_kiocb *req, int error)
2612 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2613 ssize_t ret = -ECANCELED;
2614 struct iov_iter iter;
2622 switch (req->opcode) {
2623 case IORING_OP_READV:
2624 case IORING_OP_READ_FIXED:
2625 case IORING_OP_READ:
2628 case IORING_OP_WRITEV:
2629 case IORING_OP_WRITE_FIXED:
2630 case IORING_OP_WRITE:
2634 printk_once(KERN_WARNING "io_uring: bad opcode in resubmit %d\n",
2639 if (!req->async_data) {
2640 ret = io_import_iovec(rw, req, &iovec, &iter, false);
2643 ret = io_setup_async_rw(req, iovec, inline_vecs, &iter, false);
2651 req_set_fail_links(req);
2656 static bool io_rw_reissue(struct io_kiocb *req, long res)
2659 umode_t mode = file_inode(req->file)->i_mode;
2662 if (!S_ISBLK(mode) && !S_ISREG(mode))
2664 if ((res != -EAGAIN && res != -EOPNOTSUPP) || io_wq_current_is_worker())
2667 ret = io_sq_thread_acquire_mm_files(req->ctx, req);
2669 if (io_resubmit_prep(req, ret)) {
2670 refcount_inc(&req->refs);
2671 io_queue_async_work(req);
2679 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2680 struct io_comp_state *cs)
2682 if (!io_rw_reissue(req, res))
2683 io_complete_rw_common(&req->rw.kiocb, res, cs);
2686 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2688 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2690 __io_complete_rw(req, res, res2, NULL);
2693 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2695 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2697 if (kiocb->ki_flags & IOCB_WRITE)
2698 kiocb_end_write(req);
2700 if (res != -EAGAIN && res != req->result)
2701 req_set_fail_links(req);
2703 WRITE_ONCE(req->result, res);
2704 /* order with io_poll_complete() checking ->result */
2706 WRITE_ONCE(req->iopoll_completed, 1);
2710 * After the iocb has been issued, it's safe to be found on the poll list.
2711 * Adding the kiocb to the list AFTER submission ensures that we don't
2712 * find it from a io_iopoll_getevents() thread before the issuer is done
2713 * accessing the kiocb cookie.
2715 static void io_iopoll_req_issued(struct io_kiocb *req, bool in_async)
2717 struct io_ring_ctx *ctx = req->ctx;
2720 * Track whether we have multiple files in our lists. This will impact
2721 * how we do polling eventually, not spinning if we're on potentially
2722 * different devices.
2724 if (list_empty(&ctx->iopoll_list)) {
2725 ctx->poll_multi_file = false;
2726 } else if (!ctx->poll_multi_file) {
2727 struct io_kiocb *list_req;
2729 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2731 if (list_req->file != req->file)
2732 ctx->poll_multi_file = true;
2736 * For fast devices, IO may have already completed. If it has, add
2737 * it to the front so we find it first.
2739 if (READ_ONCE(req->iopoll_completed))
2740 list_add(&req->inflight_entry, &ctx->iopoll_list);
2742 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2745 * If IORING_SETUP_SQPOLL is enabled, sqes are either handled in sq thread
2746 * task context or in io worker task context. If current task context is
2747 * sq thread, we don't need to check whether should wake up sq thread.
2749 if (in_async && (ctx->flags & IORING_SETUP_SQPOLL) &&
2750 wq_has_sleeper(&ctx->sq_data->wait))
2751 wake_up(&ctx->sq_data->wait);
2754 static inline void __io_state_file_put(struct io_submit_state *state)
2756 fput_many(state->file, state->file_refs);
2757 state->file_refs = 0;
2760 static inline void io_state_file_put(struct io_submit_state *state)
2762 if (state->file_refs)
2763 __io_state_file_put(state);
2767 * Get as many references to a file as we have IOs left in this submission,
2768 * assuming most submissions are for one file, or at least that each file
2769 * has more than one submission.
2771 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2776 if (state->file_refs) {
2777 if (state->fd == fd) {
2781 __io_state_file_put(state);
2783 state->file = fget_many(fd, state->ios_left);
2784 if (unlikely(!state->file))
2788 state->file_refs = state->ios_left - 1;
2792 static bool io_bdev_nowait(struct block_device *bdev)
2794 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2798 * If we tracked the file through the SCM inflight mechanism, we could support
2799 * any file. For now, just ensure that anything potentially problematic is done
2802 static bool io_file_supports_async(struct file *file, int rw)
2804 umode_t mode = file_inode(file)->i_mode;
2806 if (S_ISBLK(mode)) {
2807 if (IS_ENABLED(CONFIG_BLOCK) &&
2808 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
2812 if (S_ISCHR(mode) || S_ISSOCK(mode))
2814 if (S_ISREG(mode)) {
2815 if (IS_ENABLED(CONFIG_BLOCK) &&
2816 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2817 file->f_op != &io_uring_fops)
2822 /* any ->read/write should understand O_NONBLOCK */
2823 if (file->f_flags & O_NONBLOCK)
2826 if (!(file->f_mode & FMODE_NOWAIT))
2830 return file->f_op->read_iter != NULL;
2832 return file->f_op->write_iter != NULL;
2835 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2837 struct io_ring_ctx *ctx = req->ctx;
2838 struct kiocb *kiocb = &req->rw.kiocb;
2842 if (S_ISREG(file_inode(req->file)->i_mode))
2843 req->flags |= REQ_F_ISREG;
2845 kiocb->ki_pos = READ_ONCE(sqe->off);
2846 if (kiocb->ki_pos == -1 && !(req->file->f_mode & FMODE_STREAM)) {
2847 req->flags |= REQ_F_CUR_POS;
2848 kiocb->ki_pos = req->file->f_pos;
2850 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2851 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2852 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2856 ioprio = READ_ONCE(sqe->ioprio);
2858 ret = ioprio_check_cap(ioprio);
2862 kiocb->ki_ioprio = ioprio;
2864 kiocb->ki_ioprio = get_current_ioprio();
2866 /* don't allow async punt if RWF_NOWAIT was requested */
2867 if (kiocb->ki_flags & IOCB_NOWAIT)
2868 req->flags |= REQ_F_NOWAIT;
2870 if (ctx->flags & IORING_SETUP_IOPOLL) {
2871 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2872 !kiocb->ki_filp->f_op->iopoll)
2875 kiocb->ki_flags |= IOCB_HIPRI;
2876 kiocb->ki_complete = io_complete_rw_iopoll;
2877 req->iopoll_completed = 0;
2879 if (kiocb->ki_flags & IOCB_HIPRI)
2881 kiocb->ki_complete = io_complete_rw;
2884 req->rw.addr = READ_ONCE(sqe->addr);
2885 req->rw.len = READ_ONCE(sqe->len);
2886 req->buf_index = READ_ONCE(sqe->buf_index);
2890 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2896 case -ERESTARTNOINTR:
2897 case -ERESTARTNOHAND:
2898 case -ERESTART_RESTARTBLOCK:
2900 * We can't just restart the syscall, since previously
2901 * submitted sqes may already be in progress. Just fail this
2907 kiocb->ki_complete(kiocb, ret, 0);
2911 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2912 struct io_comp_state *cs)
2914 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2915 struct io_async_rw *io = req->async_data;
2917 /* add previously done IO, if any */
2918 if (io && io->bytes_done > 0) {
2920 ret = io->bytes_done;
2922 ret += io->bytes_done;
2925 if (req->flags & REQ_F_CUR_POS)
2926 req->file->f_pos = kiocb->ki_pos;
2927 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2928 __io_complete_rw(req, ret, 0, cs);
2930 io_rw_done(kiocb, ret);
2933 static ssize_t io_import_fixed(struct io_kiocb *req, int rw,
2934 struct iov_iter *iter)
2936 struct io_ring_ctx *ctx = req->ctx;
2937 size_t len = req->rw.len;
2938 struct io_mapped_ubuf *imu;
2939 u16 index, buf_index = req->buf_index;
2943 if (unlikely(buf_index >= ctx->nr_user_bufs))
2945 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2946 imu = &ctx->user_bufs[index];
2947 buf_addr = req->rw.addr;
2950 if (buf_addr + len < buf_addr)
2952 /* not inside the mapped region */
2953 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2957 * May not be a start of buffer, set size appropriately
2958 * and advance us to the beginning.
2960 offset = buf_addr - imu->ubuf;
2961 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2965 * Don't use iov_iter_advance() here, as it's really slow for
2966 * using the latter parts of a big fixed buffer - it iterates
2967 * over each segment manually. We can cheat a bit here, because
2970 * 1) it's a BVEC iter, we set it up
2971 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2972 * first and last bvec
2974 * So just find our index, and adjust the iterator afterwards.
2975 * If the offset is within the first bvec (or the whole first
2976 * bvec, just use iov_iter_advance(). This makes it easier
2977 * since we can just skip the first segment, which may not
2978 * be PAGE_SIZE aligned.
2980 const struct bio_vec *bvec = imu->bvec;
2982 if (offset <= bvec->bv_len) {
2983 iov_iter_advance(iter, offset);
2985 unsigned long seg_skip;
2987 /* skip first vec */
2988 offset -= bvec->bv_len;
2989 seg_skip = 1 + (offset >> PAGE_SHIFT);
2991 iter->bvec = bvec + seg_skip;
2992 iter->nr_segs -= seg_skip;
2993 iter->count -= bvec->bv_len + offset;
2994 iter->iov_offset = offset & ~PAGE_MASK;
3001 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
3004 mutex_unlock(&ctx->uring_lock);
3007 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
3010 * "Normal" inline submissions always hold the uring_lock, since we
3011 * grab it from the system call. Same is true for the SQPOLL offload.
3012 * The only exception is when we've detached the request and issue it
3013 * from an async worker thread, grab the lock for that case.
3016 mutex_lock(&ctx->uring_lock);
3019 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
3020 int bgid, struct io_buffer *kbuf,
3023 struct io_buffer *head;
3025 if (req->flags & REQ_F_BUFFER_SELECTED)
3028 io_ring_submit_lock(req->ctx, needs_lock);
3030 lockdep_assert_held(&req->ctx->uring_lock);
3032 head = idr_find(&req->ctx->io_buffer_idr, bgid);
3034 if (!list_empty(&head->list)) {
3035 kbuf = list_last_entry(&head->list, struct io_buffer,
3037 list_del(&kbuf->list);
3040 idr_remove(&req->ctx->io_buffer_idr, bgid);
3042 if (*len > kbuf->len)
3045 kbuf = ERR_PTR(-ENOBUFS);
3048 io_ring_submit_unlock(req->ctx, needs_lock);
3053 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
3056 struct io_buffer *kbuf;
3059 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3060 bgid = req->buf_index;
3061 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
3064 req->rw.addr = (u64) (unsigned long) kbuf;
3065 req->flags |= REQ_F_BUFFER_SELECTED;
3066 return u64_to_user_ptr(kbuf->addr);
3069 #ifdef CONFIG_COMPAT
3070 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
3073 struct compat_iovec __user *uiov;
3074 compat_ssize_t clen;
3078 uiov = u64_to_user_ptr(req->rw.addr);
3079 if (!access_ok(uiov, sizeof(*uiov)))
3081 if (__get_user(clen, &uiov->iov_len))
3087 buf = io_rw_buffer_select(req, &len, needs_lock);
3089 return PTR_ERR(buf);
3090 iov[0].iov_base = buf;
3091 iov[0].iov_len = (compat_size_t) len;
3096 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3099 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
3103 if (copy_from_user(iov, uiov, sizeof(*uiov)))
3106 len = iov[0].iov_len;
3109 buf = io_rw_buffer_select(req, &len, needs_lock);
3111 return PTR_ERR(buf);
3112 iov[0].iov_base = buf;
3113 iov[0].iov_len = len;
3117 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3120 if (req->flags & REQ_F_BUFFER_SELECTED) {
3121 struct io_buffer *kbuf;
3123 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3124 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
3125 iov[0].iov_len = kbuf->len;
3128 if (req->rw.len != 1)
3131 #ifdef CONFIG_COMPAT
3132 if (req->ctx->compat)
3133 return io_compat_import(req, iov, needs_lock);
3136 return __io_iov_buffer_select(req, iov, needs_lock);
3139 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
3140 struct iovec **iovec, struct iov_iter *iter,
3143 void __user *buf = u64_to_user_ptr(req->rw.addr);
3144 size_t sqe_len = req->rw.len;
3148 opcode = req->opcode;
3149 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3151 return io_import_fixed(req, rw, iter);
3154 /* buffer index only valid with fixed read/write, or buffer select */
3155 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3158 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3159 if (req->flags & REQ_F_BUFFER_SELECT) {
3160 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3162 return PTR_ERR(buf);
3163 req->rw.len = sqe_len;
3166 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3171 if (req->flags & REQ_F_BUFFER_SELECT) {
3172 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3174 ret = (*iovec)->iov_len;
3175 iov_iter_init(iter, rw, *iovec, 1, ret);
3181 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3185 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3187 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3191 * For files that don't have ->read_iter() and ->write_iter(), handle them
3192 * by looping over ->read() or ->write() manually.
3194 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3196 struct kiocb *kiocb = &req->rw.kiocb;
3197 struct file *file = req->file;
3201 * Don't support polled IO through this interface, and we can't
3202 * support non-blocking either. For the latter, this just causes
3203 * the kiocb to be handled from an async context.
3205 if (kiocb->ki_flags & IOCB_HIPRI)
3207 if (kiocb->ki_flags & IOCB_NOWAIT)
3210 while (iov_iter_count(iter)) {
3214 if (!iov_iter_is_bvec(iter)) {
3215 iovec = iov_iter_iovec(iter);
3217 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3218 iovec.iov_len = req->rw.len;
3222 nr = file->f_op->read(file, iovec.iov_base,
3223 iovec.iov_len, io_kiocb_ppos(kiocb));
3225 nr = file->f_op->write(file, iovec.iov_base,
3226 iovec.iov_len, io_kiocb_ppos(kiocb));
3235 if (nr != iovec.iov_len)
3239 iov_iter_advance(iter, nr);
3245 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3246 const struct iovec *fast_iov, struct iov_iter *iter)
3248 struct io_async_rw *rw = req->async_data;
3250 memcpy(&rw->iter, iter, sizeof(*iter));
3251 rw->free_iovec = iovec;
3253 /* can only be fixed buffers, no need to do anything */
3254 if (iov_iter_is_bvec(iter))
3257 unsigned iov_off = 0;
3259 rw->iter.iov = rw->fast_iov;
3260 if (iter->iov != fast_iov) {
3261 iov_off = iter->iov - fast_iov;
3262 rw->iter.iov += iov_off;
3264 if (rw->fast_iov != fast_iov)
3265 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3266 sizeof(struct iovec) * iter->nr_segs);
3268 req->flags |= REQ_F_NEED_CLEANUP;
3272 static inline int __io_alloc_async_data(struct io_kiocb *req)
3274 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3275 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3276 return req->async_data == NULL;
3279 static int io_alloc_async_data(struct io_kiocb *req)
3281 if (!io_op_defs[req->opcode].needs_async_data)
3284 return __io_alloc_async_data(req);
3287 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3288 const struct iovec *fast_iov,
3289 struct iov_iter *iter, bool force)
3291 if (!force && !io_op_defs[req->opcode].needs_async_data)
3293 if (!req->async_data) {
3294 if (__io_alloc_async_data(req))
3297 io_req_map_rw(req, iovec, fast_iov, iter);
3302 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3304 struct io_async_rw *iorw = req->async_data;
3305 struct iovec *iov = iorw->fast_iov;
3308 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3309 if (unlikely(ret < 0))
3312 iorw->bytes_done = 0;
3313 iorw->free_iovec = iov;
3315 req->flags |= REQ_F_NEED_CLEANUP;
3319 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3323 ret = io_prep_rw(req, sqe);
3327 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3330 /* either don't need iovec imported or already have it */
3331 if (!req->async_data)
3333 return io_rw_prep_async(req, READ);
3337 * This is our waitqueue callback handler, registered through lock_page_async()
3338 * when we initially tried to do the IO with the iocb armed our waitqueue.
3339 * This gets called when the page is unlocked, and we generally expect that to
3340 * happen when the page IO is completed and the page is now uptodate. This will
3341 * queue a task_work based retry of the operation, attempting to copy the data
3342 * again. If the latter fails because the page was NOT uptodate, then we will
3343 * do a thread based blocking retry of the operation. That's the unexpected
3346 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3347 int sync, void *arg)
3349 struct wait_page_queue *wpq;
3350 struct io_kiocb *req = wait->private;
3351 struct wait_page_key *key = arg;
3354 wpq = container_of(wait, struct wait_page_queue, wait);
3356 if (!wake_page_match(wpq, key))
3359 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3360 list_del_init(&wait->entry);
3362 init_task_work(&req->task_work, io_req_task_submit);
3363 percpu_ref_get(&req->ctx->refs);
3365 /* submit ref gets dropped, acquire a new one */
3366 refcount_inc(&req->refs);
3367 ret = io_req_task_work_add(req);
3368 if (unlikely(ret)) {
3369 struct task_struct *tsk;
3371 /* queue just for cancelation */
3372 init_task_work(&req->task_work, io_req_task_cancel);
3373 tsk = io_wq_get_task(req->ctx->io_wq);
3374 task_work_add(tsk, &req->task_work, TWA_NONE);
3375 wake_up_process(tsk);
3381 * This controls whether a given IO request should be armed for async page
3382 * based retry. If we return false here, the request is handed to the async
3383 * worker threads for retry. If we're doing buffered reads on a regular file,
3384 * we prepare a private wait_page_queue entry and retry the operation. This
3385 * will either succeed because the page is now uptodate and unlocked, or it
3386 * will register a callback when the page is unlocked at IO completion. Through
3387 * that callback, io_uring uses task_work to setup a retry of the operation.
3388 * That retry will attempt the buffered read again. The retry will generally
3389 * succeed, or in rare cases where it fails, we then fall back to using the
3390 * async worker threads for a blocking retry.
3392 static bool io_rw_should_retry(struct io_kiocb *req)
3394 struct io_async_rw *rw = req->async_data;
3395 struct wait_page_queue *wait = &rw->wpq;
3396 struct kiocb *kiocb = &req->rw.kiocb;
3398 /* never retry for NOWAIT, we just complete with -EAGAIN */
3399 if (req->flags & REQ_F_NOWAIT)
3402 /* Only for buffered IO */
3403 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3407 * just use poll if we can, and don't attempt if the fs doesn't
3408 * support callback based unlocks
3410 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3413 wait->wait.func = io_async_buf_func;
3414 wait->wait.private = req;
3415 wait->wait.flags = 0;
3416 INIT_LIST_HEAD(&wait->wait.entry);
3417 kiocb->ki_flags |= IOCB_WAITQ;
3418 kiocb->ki_flags &= ~IOCB_NOWAIT;
3419 kiocb->ki_waitq = wait;
3423 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3425 if (req->file->f_op->read_iter)
3426 return call_read_iter(req->file, &req->rw.kiocb, iter);
3427 else if (req->file->f_op->read)
3428 return loop_rw_iter(READ, req, iter);
3433 static int io_read(struct io_kiocb *req, bool force_nonblock,
3434 struct io_comp_state *cs)
3436 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3437 struct kiocb *kiocb = &req->rw.kiocb;
3438 struct iov_iter __iter, *iter = &__iter;
3439 struct io_async_rw *rw = req->async_data;
3440 ssize_t io_size, ret, ret2;
3447 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3451 io_size = iov_iter_count(iter);
3452 req->result = io_size;
3455 /* Ensure we clear previously set non-block flag */
3456 if (!force_nonblock)
3457 kiocb->ki_flags &= ~IOCB_NOWAIT;
3459 kiocb->ki_flags |= IOCB_NOWAIT;
3462 /* If the file doesn't support async, just async punt */
3463 no_async = force_nonblock && !io_file_supports_async(req->file, READ);
3467 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size);
3471 ret = io_iter_do_read(req, iter);
3475 } else if (ret == -EIOCBQUEUED) {
3478 } else if (ret == -EAGAIN) {
3479 /* IOPOLL retry should happen for io-wq threads */
3480 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3482 /* no retry on NONBLOCK marked file */
3483 if (req->file->f_flags & O_NONBLOCK)
3485 /* some cases will consume bytes even on error returns */
3486 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3489 } else if (ret < 0) {
3490 /* make sure -ERESTARTSYS -> -EINTR is done */
3494 /* read it all, or we did blocking attempt. no retry. */
3495 if (!iov_iter_count(iter) || !force_nonblock ||
3496 (req->file->f_flags & O_NONBLOCK))
3501 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3508 rw = req->async_data;
3509 /* it's copied and will be cleaned with ->io */
3511 /* now use our persistent iterator, if we aren't already */
3514 rw->bytes_done += ret;
3515 /* if we can retry, do so with the callbacks armed */
3516 if (!io_rw_should_retry(req)) {
3517 kiocb->ki_flags &= ~IOCB_WAITQ;
3522 * Now retry read with the IOCB_WAITQ parts set in the iocb. If we
3523 * get -EIOCBQUEUED, then we'll get a notification when the desired
3524 * page gets unlocked. We can also get a partial read here, and if we
3525 * do, then just retry at the new offset.
3527 ret = io_iter_do_read(req, iter);
3528 if (ret == -EIOCBQUEUED) {
3531 } else if (ret > 0 && ret < io_size) {
3532 /* we got some bytes, but not all. retry. */
3536 kiocb_done(kiocb, ret, cs);
3539 /* it's reportedly faster than delegating the null check to kfree() */
3545 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3549 ret = io_prep_rw(req, sqe);
3553 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3556 /* either don't need iovec imported or already have it */
3557 if (!req->async_data)
3559 return io_rw_prep_async(req, WRITE);
3562 static int io_write(struct io_kiocb *req, bool force_nonblock,
3563 struct io_comp_state *cs)
3565 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3566 struct kiocb *kiocb = &req->rw.kiocb;
3567 struct iov_iter __iter, *iter = &__iter;
3568 struct io_async_rw *rw = req->async_data;
3569 ssize_t ret, ret2, io_size;
3575 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3579 io_size = iov_iter_count(iter);
3580 req->result = io_size;
3582 /* Ensure we clear previously set non-block flag */
3583 if (!force_nonblock)
3584 kiocb->ki_flags &= ~IOCB_NOWAIT;
3586 kiocb->ki_flags |= IOCB_NOWAIT;
3588 /* If the file doesn't support async, just async punt */
3589 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
3592 /* file path doesn't support NOWAIT for non-direct_IO */
3593 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3594 (req->flags & REQ_F_ISREG))
3597 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size);
3602 * Open-code file_start_write here to grab freeze protection,
3603 * which will be released by another thread in
3604 * io_complete_rw(). Fool lockdep by telling it the lock got
3605 * released so that it doesn't complain about the held lock when
3606 * we return to userspace.
3608 if (req->flags & REQ_F_ISREG) {
3609 sb_start_write(file_inode(req->file)->i_sb);
3610 __sb_writers_release(file_inode(req->file)->i_sb,
3613 kiocb->ki_flags |= IOCB_WRITE;
3615 if (req->file->f_op->write_iter)
3616 ret2 = call_write_iter(req->file, kiocb, iter);
3617 else if (req->file->f_op->write)
3618 ret2 = loop_rw_iter(WRITE, req, iter);
3623 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3624 * retry them without IOCB_NOWAIT.
3626 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3628 /* no retry on NONBLOCK marked file */
3629 if (ret2 == -EAGAIN && (req->file->f_flags & O_NONBLOCK))
3631 if (!force_nonblock || ret2 != -EAGAIN) {
3632 /* IOPOLL retry should happen for io-wq threads */
3633 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3636 kiocb_done(kiocb, ret2, cs);
3639 /* some cases will consume bytes even on error returns */
3640 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3641 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3646 /* it's reportedly faster than delegating the null check to kfree() */
3652 static int io_renameat_prep(struct io_kiocb *req,
3653 const struct io_uring_sqe *sqe)
3655 struct io_rename *ren = &req->rename;
3656 const char __user *oldf, *newf;
3658 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3661 ren->old_dfd = READ_ONCE(sqe->fd);
3662 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3663 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3664 ren->new_dfd = READ_ONCE(sqe->len);
3665 ren->flags = READ_ONCE(sqe->rename_flags);
3667 ren->oldpath = getname(oldf);
3668 if (IS_ERR(ren->oldpath))
3669 return PTR_ERR(ren->oldpath);
3671 ren->newpath = getname(newf);
3672 if (IS_ERR(ren->newpath)) {
3673 putname(ren->oldpath);
3674 return PTR_ERR(ren->newpath);
3677 req->flags |= REQ_F_NEED_CLEANUP;
3681 static int io_renameat(struct io_kiocb *req, bool force_nonblock)
3683 struct io_rename *ren = &req->rename;
3689 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3690 ren->newpath, ren->flags);
3692 req->flags &= ~REQ_F_NEED_CLEANUP;
3694 req_set_fail_links(req);
3695 io_req_complete(req, ret);
3699 static int io_unlinkat_prep(struct io_kiocb *req,
3700 const struct io_uring_sqe *sqe)
3702 struct io_unlink *un = &req->unlink;
3703 const char __user *fname;
3705 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3708 un->dfd = READ_ONCE(sqe->fd);
3710 un->flags = READ_ONCE(sqe->unlink_flags);
3711 if (un->flags & ~AT_REMOVEDIR)
3714 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3715 un->filename = getname(fname);
3716 if (IS_ERR(un->filename))
3717 return PTR_ERR(un->filename);
3719 req->flags |= REQ_F_NEED_CLEANUP;
3723 static int io_unlinkat(struct io_kiocb *req, bool force_nonblock)
3725 struct io_unlink *un = &req->unlink;
3731 if (un->flags & AT_REMOVEDIR)
3732 ret = do_rmdir(un->dfd, un->filename);
3734 ret = do_unlinkat(un->dfd, un->filename);
3736 req->flags &= ~REQ_F_NEED_CLEANUP;
3738 req_set_fail_links(req);
3739 io_req_complete(req, ret);
3743 static int io_shutdown_prep(struct io_kiocb *req,
3744 const struct io_uring_sqe *sqe)
3746 #if defined(CONFIG_NET)
3747 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3749 if (sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3753 req->shutdown.how = READ_ONCE(sqe->len);
3760 static int io_shutdown(struct io_kiocb *req, bool force_nonblock)
3762 #if defined(CONFIG_NET)
3763 struct socket *sock;
3769 sock = sock_from_file(req->file);
3770 if (unlikely(!sock))
3773 ret = __sys_shutdown_sock(sock, req->shutdown.how);
3775 req_set_fail_links(req);
3776 io_req_complete(req, ret);
3783 static int __io_splice_prep(struct io_kiocb *req,
3784 const struct io_uring_sqe *sqe)
3786 struct io_splice* sp = &req->splice;
3787 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3789 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3793 sp->len = READ_ONCE(sqe->len);
3794 sp->flags = READ_ONCE(sqe->splice_flags);
3796 if (unlikely(sp->flags & ~valid_flags))
3799 sp->file_in = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in),
3800 (sp->flags & SPLICE_F_FD_IN_FIXED));
3803 req->flags |= REQ_F_NEED_CLEANUP;
3805 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
3807 * Splice operation will be punted aync, and here need to
3808 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3810 io_req_init_async(req);
3811 req->work.flags |= IO_WQ_WORK_UNBOUND;
3817 static int io_tee_prep(struct io_kiocb *req,
3818 const struct io_uring_sqe *sqe)
3820 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3822 return __io_splice_prep(req, sqe);
3825 static int io_tee(struct io_kiocb *req, bool force_nonblock)
3827 struct io_splice *sp = &req->splice;
3828 struct file *in = sp->file_in;
3829 struct file *out = sp->file_out;
3830 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3836 ret = do_tee(in, out, sp->len, flags);
3838 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3839 req->flags &= ~REQ_F_NEED_CLEANUP;
3842 req_set_fail_links(req);
3843 io_req_complete(req, ret);
3847 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3849 struct io_splice* sp = &req->splice;
3851 sp->off_in = READ_ONCE(sqe->splice_off_in);
3852 sp->off_out = READ_ONCE(sqe->off);
3853 return __io_splice_prep(req, sqe);
3856 static int io_splice(struct io_kiocb *req, bool force_nonblock)
3858 struct io_splice *sp = &req->splice;
3859 struct file *in = sp->file_in;
3860 struct file *out = sp->file_out;
3861 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3862 loff_t *poff_in, *poff_out;
3868 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3869 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3872 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3874 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3875 req->flags &= ~REQ_F_NEED_CLEANUP;
3878 req_set_fail_links(req);
3879 io_req_complete(req, ret);
3884 * IORING_OP_NOP just posts a completion event, nothing else.
3886 static int io_nop(struct io_kiocb *req, struct io_comp_state *cs)
3888 struct io_ring_ctx *ctx = req->ctx;
3890 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3893 __io_req_complete(req, 0, 0, cs);
3897 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3899 struct io_ring_ctx *ctx = req->ctx;
3904 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3906 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3909 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3910 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3913 req->sync.off = READ_ONCE(sqe->off);
3914 req->sync.len = READ_ONCE(sqe->len);
3918 static int io_fsync(struct io_kiocb *req, bool force_nonblock)
3920 loff_t end = req->sync.off + req->sync.len;
3923 /* fsync always requires a blocking context */
3927 ret = vfs_fsync_range(req->file, req->sync.off,
3928 end > 0 ? end : LLONG_MAX,
3929 req->sync.flags & IORING_FSYNC_DATASYNC);
3931 req_set_fail_links(req);
3932 io_req_complete(req, ret);
3936 static int io_fallocate_prep(struct io_kiocb *req,
3937 const struct io_uring_sqe *sqe)
3939 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3941 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3944 req->sync.off = READ_ONCE(sqe->off);
3945 req->sync.len = READ_ONCE(sqe->addr);
3946 req->sync.mode = READ_ONCE(sqe->len);
3950 static int io_fallocate(struct io_kiocb *req, bool force_nonblock)
3954 /* fallocate always requiring blocking context */
3957 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3960 req_set_fail_links(req);
3961 io_req_complete(req, ret);
3965 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3967 const char __user *fname;
3970 if (unlikely(sqe->ioprio || sqe->buf_index))
3972 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3975 /* open.how should be already initialised */
3976 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3977 req->open.how.flags |= O_LARGEFILE;
3979 req->open.dfd = READ_ONCE(sqe->fd);
3980 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3981 req->open.filename = getname(fname);
3982 if (IS_ERR(req->open.filename)) {
3983 ret = PTR_ERR(req->open.filename);
3984 req->open.filename = NULL;
3987 req->open.nofile = rlimit(RLIMIT_NOFILE);
3988 req->open.ignore_nonblock = false;
3989 req->flags |= REQ_F_NEED_CLEANUP;
3993 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3997 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3999 mode = READ_ONCE(sqe->len);
4000 flags = READ_ONCE(sqe->open_flags);
4001 req->open.how = build_open_how(flags, mode);
4002 return __io_openat_prep(req, sqe);
4005 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4007 struct open_how __user *how;
4011 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4013 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4014 len = READ_ONCE(sqe->len);
4015 if (len < OPEN_HOW_SIZE_VER0)
4018 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
4023 return __io_openat_prep(req, sqe);
4026 static int io_openat2(struct io_kiocb *req, bool force_nonblock)
4028 struct open_flags op;
4032 if (force_nonblock && !req->open.ignore_nonblock)
4035 ret = build_open_flags(&req->open.how, &op);
4039 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
4043 file = do_filp_open(req->open.dfd, req->open.filename, &op);
4046 ret = PTR_ERR(file);
4048 * A work-around to ensure that /proc/self works that way
4049 * that it should - if we get -EOPNOTSUPP back, then assume
4050 * that proc_self_get_link() failed us because we're in async
4051 * context. We should be safe to retry this from the task
4052 * itself with force_nonblock == false set, as it should not
4053 * block on lookup. Would be nice to know this upfront and
4054 * avoid the async dance, but doesn't seem feasible.
4056 if (ret == -EOPNOTSUPP && io_wq_current_is_worker()) {
4057 req->open.ignore_nonblock = true;
4058 refcount_inc(&req->refs);
4059 io_req_task_queue(req);
4063 fsnotify_open(file);
4064 fd_install(ret, file);
4067 putname(req->open.filename);
4068 req->flags &= ~REQ_F_NEED_CLEANUP;
4070 req_set_fail_links(req);
4071 io_req_complete(req, ret);
4075 static int io_openat(struct io_kiocb *req, bool force_nonblock)
4077 return io_openat2(req, force_nonblock);
4080 static int io_remove_buffers_prep(struct io_kiocb *req,
4081 const struct io_uring_sqe *sqe)
4083 struct io_provide_buf *p = &req->pbuf;
4086 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
4089 tmp = READ_ONCE(sqe->fd);
4090 if (!tmp || tmp > USHRT_MAX)
4093 memset(p, 0, sizeof(*p));
4095 p->bgid = READ_ONCE(sqe->buf_group);
4099 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
4100 int bgid, unsigned nbufs)
4104 /* shouldn't happen */
4108 /* the head kbuf is the list itself */
4109 while (!list_empty(&buf->list)) {
4110 struct io_buffer *nxt;
4112 nxt = list_first_entry(&buf->list, struct io_buffer, list);
4113 list_del(&nxt->list);
4120 idr_remove(&ctx->io_buffer_idr, bgid);
4125 static int io_remove_buffers(struct io_kiocb *req, bool force_nonblock,
4126 struct io_comp_state *cs)
4128 struct io_provide_buf *p = &req->pbuf;
4129 struct io_ring_ctx *ctx = req->ctx;
4130 struct io_buffer *head;
4133 io_ring_submit_lock(ctx, !force_nonblock);
4135 lockdep_assert_held(&ctx->uring_lock);
4138 head = idr_find(&ctx->io_buffer_idr, p->bgid);
4140 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
4142 req_set_fail_links(req);
4144 /* need to hold the lock to complete IOPOLL requests */
4145 if (ctx->flags & IORING_SETUP_IOPOLL) {
4146 __io_req_complete(req, ret, 0, cs);
4147 io_ring_submit_unlock(ctx, !force_nonblock);
4149 io_ring_submit_unlock(ctx, !force_nonblock);
4150 __io_req_complete(req, ret, 0, cs);
4155 static int io_provide_buffers_prep(struct io_kiocb *req,
4156 const struct io_uring_sqe *sqe)
4158 struct io_provide_buf *p = &req->pbuf;
4161 if (sqe->ioprio || sqe->rw_flags)
4164 tmp = READ_ONCE(sqe->fd);
4165 if (!tmp || tmp > USHRT_MAX)
4168 p->addr = READ_ONCE(sqe->addr);
4169 p->len = READ_ONCE(sqe->len);
4171 if (!access_ok(u64_to_user_ptr(p->addr), (p->len * p->nbufs)))
4174 p->bgid = READ_ONCE(sqe->buf_group);
4175 tmp = READ_ONCE(sqe->off);
4176 if (tmp > USHRT_MAX)
4182 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
4184 struct io_buffer *buf;
4185 u64 addr = pbuf->addr;
4186 int i, bid = pbuf->bid;
4188 for (i = 0; i < pbuf->nbufs; i++) {
4189 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
4194 buf->len = pbuf->len;
4199 INIT_LIST_HEAD(&buf->list);
4202 list_add_tail(&buf->list, &(*head)->list);
4206 return i ? i : -ENOMEM;
4209 static int io_provide_buffers(struct io_kiocb *req, bool force_nonblock,
4210 struct io_comp_state *cs)
4212 struct io_provide_buf *p = &req->pbuf;
4213 struct io_ring_ctx *ctx = req->ctx;
4214 struct io_buffer *head, *list;
4217 io_ring_submit_lock(ctx, !force_nonblock);
4219 lockdep_assert_held(&ctx->uring_lock);
4221 list = head = idr_find(&ctx->io_buffer_idr, p->bgid);
4223 ret = io_add_buffers(p, &head);
4228 ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1,
4231 __io_remove_buffers(ctx, head, p->bgid, -1U);
4237 req_set_fail_links(req);
4239 /* need to hold the lock to complete IOPOLL requests */
4240 if (ctx->flags & IORING_SETUP_IOPOLL) {
4241 __io_req_complete(req, ret, 0, cs);
4242 io_ring_submit_unlock(ctx, !force_nonblock);
4244 io_ring_submit_unlock(ctx, !force_nonblock);
4245 __io_req_complete(req, ret, 0, cs);
4250 static int io_epoll_ctl_prep(struct io_kiocb *req,
4251 const struct io_uring_sqe *sqe)
4253 #if defined(CONFIG_EPOLL)
4254 if (sqe->ioprio || sqe->buf_index)
4256 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4259 req->epoll.epfd = READ_ONCE(sqe->fd);
4260 req->epoll.op = READ_ONCE(sqe->len);
4261 req->epoll.fd = READ_ONCE(sqe->off);
4263 if (ep_op_has_event(req->epoll.op)) {
4264 struct epoll_event __user *ev;
4266 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4267 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4277 static int io_epoll_ctl(struct io_kiocb *req, bool force_nonblock,
4278 struct io_comp_state *cs)
4280 #if defined(CONFIG_EPOLL)
4281 struct io_epoll *ie = &req->epoll;
4284 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4285 if (force_nonblock && ret == -EAGAIN)
4289 req_set_fail_links(req);
4290 __io_req_complete(req, ret, 0, cs);
4297 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4299 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4300 if (sqe->ioprio || sqe->buf_index || sqe->off)
4302 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4305 req->madvise.addr = READ_ONCE(sqe->addr);
4306 req->madvise.len = READ_ONCE(sqe->len);
4307 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4314 static int io_madvise(struct io_kiocb *req, bool force_nonblock)
4316 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4317 struct io_madvise *ma = &req->madvise;
4323 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4325 req_set_fail_links(req);
4326 io_req_complete(req, ret);
4333 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4335 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4337 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4340 req->fadvise.offset = READ_ONCE(sqe->off);
4341 req->fadvise.len = READ_ONCE(sqe->len);
4342 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4346 static int io_fadvise(struct io_kiocb *req, bool force_nonblock)
4348 struct io_fadvise *fa = &req->fadvise;
4351 if (force_nonblock) {
4352 switch (fa->advice) {
4353 case POSIX_FADV_NORMAL:
4354 case POSIX_FADV_RANDOM:
4355 case POSIX_FADV_SEQUENTIAL:
4362 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4364 req_set_fail_links(req);
4365 io_req_complete(req, ret);
4369 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4371 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4373 if (sqe->ioprio || sqe->buf_index)
4375 if (req->flags & REQ_F_FIXED_FILE)
4378 req->statx.dfd = READ_ONCE(sqe->fd);
4379 req->statx.mask = READ_ONCE(sqe->len);
4380 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4381 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4382 req->statx.flags = READ_ONCE(sqe->statx_flags);
4387 static int io_statx(struct io_kiocb *req, bool force_nonblock)
4389 struct io_statx *ctx = &req->statx;
4392 if (force_nonblock) {
4393 /* only need file table for an actual valid fd */
4394 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
4395 req->flags |= REQ_F_NO_FILE_TABLE;
4399 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4403 req_set_fail_links(req);
4404 io_req_complete(req, ret);
4408 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4411 * If we queue this for async, it must not be cancellable. That would
4412 * leave the 'file' in an undeterminate state, and here need to modify
4413 * io_wq_work.flags, so initialize io_wq_work firstly.
4415 io_req_init_async(req);
4416 req->work.flags |= IO_WQ_WORK_NO_CANCEL;
4418 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4420 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4421 sqe->rw_flags || sqe->buf_index)
4423 if (req->flags & REQ_F_FIXED_FILE)
4426 req->close.fd = READ_ONCE(sqe->fd);
4427 if ((req->file && req->file->f_op == &io_uring_fops))
4430 req->close.put_file = NULL;
4434 static int io_close(struct io_kiocb *req, bool force_nonblock,
4435 struct io_comp_state *cs)
4437 struct io_close *close = &req->close;
4440 /* might be already done during nonblock submission */
4441 if (!close->put_file) {
4442 ret = close_fd_get_file(close->fd, &close->put_file);
4444 return (ret == -ENOENT) ? -EBADF : ret;
4447 /* if the file has a flush method, be safe and punt to async */
4448 if (close->put_file->f_op->flush && force_nonblock) {
4449 /* was never set, but play safe */
4450 req->flags &= ~REQ_F_NOWAIT;
4451 /* avoid grabbing files - we don't need the files */
4452 req->flags |= REQ_F_NO_FILE_TABLE;
4456 /* No ->flush() or already async, safely close from here */
4457 ret = filp_close(close->put_file, req->work.identity->files);
4459 req_set_fail_links(req);
4460 fput(close->put_file);
4461 close->put_file = NULL;
4462 __io_req_complete(req, ret, 0, cs);
4466 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4468 struct io_ring_ctx *ctx = req->ctx;
4473 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4475 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4478 req->sync.off = READ_ONCE(sqe->off);
4479 req->sync.len = READ_ONCE(sqe->len);
4480 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4484 static int io_sync_file_range(struct io_kiocb *req, bool force_nonblock)
4488 /* sync_file_range always requires a blocking context */
4492 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4495 req_set_fail_links(req);
4496 io_req_complete(req, ret);
4500 #if defined(CONFIG_NET)
4501 static int io_setup_async_msg(struct io_kiocb *req,
4502 struct io_async_msghdr *kmsg)
4504 struct io_async_msghdr *async_msg = req->async_data;
4508 if (io_alloc_async_data(req)) {
4509 if (kmsg->iov != kmsg->fast_iov)
4513 async_msg = req->async_data;
4514 req->flags |= REQ_F_NEED_CLEANUP;
4515 memcpy(async_msg, kmsg, sizeof(*kmsg));
4519 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4520 struct io_async_msghdr *iomsg)
4522 iomsg->iov = iomsg->fast_iov;
4523 iomsg->msg.msg_name = &iomsg->addr;
4524 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4525 req->sr_msg.msg_flags, &iomsg->iov);
4528 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4530 struct io_async_msghdr *async_msg = req->async_data;
4531 struct io_sr_msg *sr = &req->sr_msg;
4534 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4537 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4538 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4539 sr->len = READ_ONCE(sqe->len);
4541 #ifdef CONFIG_COMPAT
4542 if (req->ctx->compat)
4543 sr->msg_flags |= MSG_CMSG_COMPAT;
4546 if (!async_msg || !io_op_defs[req->opcode].needs_async_data)
4548 ret = io_sendmsg_copy_hdr(req, async_msg);
4550 req->flags |= REQ_F_NEED_CLEANUP;
4554 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4555 struct io_comp_state *cs)
4557 struct io_async_msghdr iomsg, *kmsg;
4558 struct socket *sock;
4562 sock = sock_from_file(req->file);
4563 if (unlikely(!sock))
4566 if (req->async_data) {
4567 kmsg = req->async_data;
4568 kmsg->msg.msg_name = &kmsg->addr;
4569 /* if iov is set, it's allocated already */
4571 kmsg->iov = kmsg->fast_iov;
4572 kmsg->msg.msg_iter.iov = kmsg->iov;
4574 ret = io_sendmsg_copy_hdr(req, &iomsg);
4580 flags = req->sr_msg.msg_flags;
4581 if (flags & MSG_DONTWAIT)
4582 req->flags |= REQ_F_NOWAIT;
4583 else if (force_nonblock)
4584 flags |= MSG_DONTWAIT;
4586 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4587 if (force_nonblock && ret == -EAGAIN)
4588 return io_setup_async_msg(req, kmsg);
4589 if (ret == -ERESTARTSYS)
4592 if (kmsg->iov != kmsg->fast_iov)
4594 req->flags &= ~REQ_F_NEED_CLEANUP;
4596 req_set_fail_links(req);
4597 __io_req_complete(req, ret, 0, cs);
4601 static int io_send(struct io_kiocb *req, bool force_nonblock,
4602 struct io_comp_state *cs)
4604 struct io_sr_msg *sr = &req->sr_msg;
4607 struct socket *sock;
4611 sock = sock_from_file(req->file);
4612 if (unlikely(!sock))
4615 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4619 msg.msg_name = NULL;
4620 msg.msg_control = NULL;
4621 msg.msg_controllen = 0;
4622 msg.msg_namelen = 0;
4624 flags = req->sr_msg.msg_flags;
4625 if (flags & MSG_DONTWAIT)
4626 req->flags |= REQ_F_NOWAIT;
4627 else if (force_nonblock)
4628 flags |= MSG_DONTWAIT;
4630 msg.msg_flags = flags;
4631 ret = sock_sendmsg(sock, &msg);
4632 if (force_nonblock && ret == -EAGAIN)
4634 if (ret == -ERESTARTSYS)
4638 req_set_fail_links(req);
4639 __io_req_complete(req, ret, 0, cs);
4643 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4644 struct io_async_msghdr *iomsg)
4646 struct io_sr_msg *sr = &req->sr_msg;
4647 struct iovec __user *uiov;
4651 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4652 &iomsg->uaddr, &uiov, &iov_len);
4656 if (req->flags & REQ_F_BUFFER_SELECT) {
4659 if (copy_from_user(iomsg->iov, uiov, sizeof(*uiov)))
4661 sr->len = iomsg->iov[0].iov_len;
4662 iov_iter_init(&iomsg->msg.msg_iter, READ, iomsg->iov, 1,
4666 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4667 &iomsg->iov, &iomsg->msg.msg_iter,
4676 #ifdef CONFIG_COMPAT
4677 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4678 struct io_async_msghdr *iomsg)
4680 struct compat_msghdr __user *msg_compat;
4681 struct io_sr_msg *sr = &req->sr_msg;
4682 struct compat_iovec __user *uiov;
4687 msg_compat = (struct compat_msghdr __user *) sr->umsg;
4688 ret = __get_compat_msghdr(&iomsg->msg, msg_compat, &iomsg->uaddr,
4693 uiov = compat_ptr(ptr);
4694 if (req->flags & REQ_F_BUFFER_SELECT) {
4695 compat_ssize_t clen;
4699 if (!access_ok(uiov, sizeof(*uiov)))
4701 if (__get_user(clen, &uiov->iov_len))
4706 iomsg->iov[0].iov_len = clen;
4709 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4710 UIO_FASTIOV, &iomsg->iov,
4711 &iomsg->msg.msg_iter, true);
4720 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4721 struct io_async_msghdr *iomsg)
4723 iomsg->msg.msg_name = &iomsg->addr;
4724 iomsg->iov = iomsg->fast_iov;
4726 #ifdef CONFIG_COMPAT
4727 if (req->ctx->compat)
4728 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4731 return __io_recvmsg_copy_hdr(req, iomsg);
4734 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4737 struct io_sr_msg *sr = &req->sr_msg;
4738 struct io_buffer *kbuf;
4740 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4745 req->flags |= REQ_F_BUFFER_SELECTED;
4749 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4751 return io_put_kbuf(req, req->sr_msg.kbuf);
4754 static int io_recvmsg_prep(struct io_kiocb *req,
4755 const struct io_uring_sqe *sqe)
4757 struct io_async_msghdr *async_msg = req->async_data;
4758 struct io_sr_msg *sr = &req->sr_msg;
4761 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4764 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4765 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4766 sr->len = READ_ONCE(sqe->len);
4767 sr->bgid = READ_ONCE(sqe->buf_group);
4769 #ifdef CONFIG_COMPAT
4770 if (req->ctx->compat)
4771 sr->msg_flags |= MSG_CMSG_COMPAT;
4774 if (!async_msg || !io_op_defs[req->opcode].needs_async_data)
4776 ret = io_recvmsg_copy_hdr(req, async_msg);
4778 req->flags |= REQ_F_NEED_CLEANUP;
4782 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4783 struct io_comp_state *cs)
4785 struct io_async_msghdr iomsg, *kmsg;
4786 struct socket *sock;
4787 struct io_buffer *kbuf;
4789 int ret, cflags = 0;
4791 sock = sock_from_file(req->file);
4792 if (unlikely(!sock))
4795 if (req->async_data) {
4796 kmsg = req->async_data;
4797 kmsg->msg.msg_name = &kmsg->addr;
4798 /* if iov is set, it's allocated already */
4800 kmsg->iov = kmsg->fast_iov;
4801 kmsg->msg.msg_iter.iov = kmsg->iov;
4803 ret = io_recvmsg_copy_hdr(req, &iomsg);
4809 if (req->flags & REQ_F_BUFFER_SELECT) {
4810 kbuf = io_recv_buffer_select(req, !force_nonblock);
4812 return PTR_ERR(kbuf);
4813 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4814 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->iov,
4815 1, req->sr_msg.len);
4818 flags = req->sr_msg.msg_flags;
4819 if (flags & MSG_DONTWAIT)
4820 req->flags |= REQ_F_NOWAIT;
4821 else if (force_nonblock)
4822 flags |= MSG_DONTWAIT;
4824 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4825 kmsg->uaddr, flags);
4826 if (force_nonblock && ret == -EAGAIN)
4827 return io_setup_async_msg(req, kmsg);
4828 if (ret == -ERESTARTSYS)
4831 if (req->flags & REQ_F_BUFFER_SELECTED)
4832 cflags = io_put_recv_kbuf(req);
4833 if (kmsg->iov != kmsg->fast_iov)
4835 req->flags &= ~REQ_F_NEED_CLEANUP;
4837 req_set_fail_links(req);
4838 __io_req_complete(req, ret, cflags, cs);
4842 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4843 struct io_comp_state *cs)
4845 struct io_buffer *kbuf;
4846 struct io_sr_msg *sr = &req->sr_msg;
4848 void __user *buf = sr->buf;
4849 struct socket *sock;
4852 int ret, cflags = 0;
4854 sock = sock_from_file(req->file);
4855 if (unlikely(!sock))
4858 if (req->flags & REQ_F_BUFFER_SELECT) {
4859 kbuf = io_recv_buffer_select(req, !force_nonblock);
4861 return PTR_ERR(kbuf);
4862 buf = u64_to_user_ptr(kbuf->addr);
4865 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4869 msg.msg_name = NULL;
4870 msg.msg_control = NULL;
4871 msg.msg_controllen = 0;
4872 msg.msg_namelen = 0;
4873 msg.msg_iocb = NULL;
4876 flags = req->sr_msg.msg_flags;
4877 if (flags & MSG_DONTWAIT)
4878 req->flags |= REQ_F_NOWAIT;
4879 else if (force_nonblock)
4880 flags |= MSG_DONTWAIT;
4882 ret = sock_recvmsg(sock, &msg, flags);
4883 if (force_nonblock && ret == -EAGAIN)
4885 if (ret == -ERESTARTSYS)
4888 if (req->flags & REQ_F_BUFFER_SELECTED)
4889 cflags = io_put_recv_kbuf(req);
4891 req_set_fail_links(req);
4892 __io_req_complete(req, ret, cflags, cs);
4896 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4898 struct io_accept *accept = &req->accept;
4900 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4902 if (sqe->ioprio || sqe->len || sqe->buf_index)
4905 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4906 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4907 accept->flags = READ_ONCE(sqe->accept_flags);
4908 accept->nofile = rlimit(RLIMIT_NOFILE);
4912 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4913 struct io_comp_state *cs)
4915 struct io_accept *accept = &req->accept;
4916 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4919 if (req->file->f_flags & O_NONBLOCK)
4920 req->flags |= REQ_F_NOWAIT;
4922 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4923 accept->addr_len, accept->flags,
4925 if (ret == -EAGAIN && force_nonblock)
4928 if (ret == -ERESTARTSYS)
4930 req_set_fail_links(req);
4932 __io_req_complete(req, ret, 0, cs);
4936 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4938 struct io_connect *conn = &req->connect;
4939 struct io_async_connect *io = req->async_data;
4941 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4943 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4946 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4947 conn->addr_len = READ_ONCE(sqe->addr2);
4952 return move_addr_to_kernel(conn->addr, conn->addr_len,
4956 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4957 struct io_comp_state *cs)
4959 struct io_async_connect __io, *io;
4960 unsigned file_flags;
4963 if (req->async_data) {
4964 io = req->async_data;
4966 ret = move_addr_to_kernel(req->connect.addr,
4967 req->connect.addr_len,
4974 file_flags = force_nonblock ? O_NONBLOCK : 0;
4976 ret = __sys_connect_file(req->file, &io->address,
4977 req->connect.addr_len, file_flags);
4978 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4979 if (req->async_data)
4981 if (io_alloc_async_data(req)) {
4985 io = req->async_data;
4986 memcpy(req->async_data, &__io, sizeof(__io));
4989 if (ret == -ERESTARTSYS)
4993 req_set_fail_links(req);
4994 __io_req_complete(req, ret, 0, cs);
4997 #else /* !CONFIG_NET */
4998 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5003 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
5004 struct io_comp_state *cs)
5009 static int io_send(struct io_kiocb *req, bool force_nonblock,
5010 struct io_comp_state *cs)
5015 static int io_recvmsg_prep(struct io_kiocb *req,
5016 const struct io_uring_sqe *sqe)
5021 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
5022 struct io_comp_state *cs)
5027 static int io_recv(struct io_kiocb *req, bool force_nonblock,
5028 struct io_comp_state *cs)
5033 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5038 static int io_accept(struct io_kiocb *req, bool force_nonblock,
5039 struct io_comp_state *cs)
5044 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5049 static int io_connect(struct io_kiocb *req, bool force_nonblock,
5050 struct io_comp_state *cs)
5054 #endif /* CONFIG_NET */
5056 struct io_poll_table {
5057 struct poll_table_struct pt;
5058 struct io_kiocb *req;
5062 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
5063 __poll_t mask, task_work_func_t func)
5067 /* for instances that support it check for an event match first: */
5068 if (mask && !(mask & poll->events))
5071 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
5073 list_del_init(&poll->wait.entry);
5076 init_task_work(&req->task_work, func);
5077 percpu_ref_get(&req->ctx->refs);
5080 * If this fails, then the task is exiting. When a task exits, the
5081 * work gets canceled, so just cancel this request as well instead
5082 * of executing it. We can't safely execute it anyway, as we may not
5083 * have the needed state needed for it anyway.
5085 ret = io_req_task_work_add(req);
5086 if (unlikely(ret)) {
5087 struct task_struct *tsk;
5089 WRITE_ONCE(poll->canceled, true);
5090 tsk = io_wq_get_task(req->ctx->io_wq);
5091 task_work_add(tsk, &req->task_work, TWA_NONE);
5092 wake_up_process(tsk);
5097 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
5098 __acquires(&req->ctx->completion_lock)
5100 struct io_ring_ctx *ctx = req->ctx;
5102 if (!req->result && !READ_ONCE(poll->canceled)) {
5103 struct poll_table_struct pt = { ._key = poll->events };
5105 req->result = vfs_poll(req->file, &pt) & poll->events;
5108 spin_lock_irq(&ctx->completion_lock);
5109 if (!req->result && !READ_ONCE(poll->canceled)) {
5110 add_wait_queue(poll->head, &poll->wait);
5117 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
5119 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
5120 if (req->opcode == IORING_OP_POLL_ADD)
5121 return req->async_data;
5122 return req->apoll->double_poll;
5125 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
5127 if (req->opcode == IORING_OP_POLL_ADD)
5129 return &req->apoll->poll;
5132 static void io_poll_remove_double(struct io_kiocb *req)
5134 struct io_poll_iocb *poll = io_poll_get_double(req);
5136 lockdep_assert_held(&req->ctx->completion_lock);
5138 if (poll && poll->head) {
5139 struct wait_queue_head *head = poll->head;
5141 spin_lock(&head->lock);
5142 list_del_init(&poll->wait.entry);
5143 if (poll->wait.private)
5144 refcount_dec(&req->refs);
5146 spin_unlock(&head->lock);
5150 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
5152 struct io_ring_ctx *ctx = req->ctx;
5154 io_poll_remove_double(req);
5155 req->poll.done = true;
5156 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
5157 io_commit_cqring(ctx);
5160 static void io_poll_task_func(struct callback_head *cb)
5162 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
5163 struct io_ring_ctx *ctx = req->ctx;
5164 struct io_kiocb *nxt;
5166 if (io_poll_rewait(req, &req->poll)) {
5167 spin_unlock_irq(&ctx->completion_lock);
5169 hash_del(&req->hash_node);
5170 io_poll_complete(req, req->result, 0);
5171 spin_unlock_irq(&ctx->completion_lock);
5173 nxt = io_put_req_find_next(req);
5174 io_cqring_ev_posted(ctx);
5176 __io_req_task_submit(nxt);
5179 percpu_ref_put(&ctx->refs);
5182 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
5183 int sync, void *key)
5185 struct io_kiocb *req = wait->private;
5186 struct io_poll_iocb *poll = io_poll_get_single(req);
5187 __poll_t mask = key_to_poll(key);
5189 /* for instances that support it check for an event match first: */
5190 if (mask && !(mask & poll->events))
5193 list_del_init(&wait->entry);
5195 if (poll && poll->head) {
5198 spin_lock(&poll->head->lock);
5199 done = list_empty(&poll->wait.entry);
5201 list_del_init(&poll->wait.entry);
5202 /* make sure double remove sees this as being gone */
5203 wait->private = NULL;
5204 spin_unlock(&poll->head->lock);
5206 /* use wait func handler, so it matches the rq type */
5207 poll->wait.func(&poll->wait, mode, sync, key);
5210 refcount_dec(&req->refs);
5214 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
5215 wait_queue_func_t wake_func)
5219 poll->canceled = false;
5220 poll->events = events;
5221 INIT_LIST_HEAD(&poll->wait.entry);
5222 init_waitqueue_func_entry(&poll->wait, wake_func);
5225 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5226 struct wait_queue_head *head,
5227 struct io_poll_iocb **poll_ptr)
5229 struct io_kiocb *req = pt->req;
5232 * If poll->head is already set, it's because the file being polled
5233 * uses multiple waitqueues for poll handling (eg one for read, one
5234 * for write). Setup a separate io_poll_iocb if this happens.
5236 if (unlikely(poll->head)) {
5237 struct io_poll_iocb *poll_one = poll;
5239 /* already have a 2nd entry, fail a third attempt */
5241 pt->error = -EINVAL;
5244 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5246 pt->error = -ENOMEM;
5249 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5250 refcount_inc(&req->refs);
5251 poll->wait.private = req;
5258 if (poll->events & EPOLLEXCLUSIVE)
5259 add_wait_queue_exclusive(head, &poll->wait);
5261 add_wait_queue(head, &poll->wait);
5264 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5265 struct poll_table_struct *p)
5267 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5268 struct async_poll *apoll = pt->req->apoll;
5270 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5273 static void io_async_task_func(struct callback_head *cb)
5275 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
5276 struct async_poll *apoll = req->apoll;
5277 struct io_ring_ctx *ctx = req->ctx;
5279 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
5281 if (io_poll_rewait(req, &apoll->poll)) {
5282 spin_unlock_irq(&ctx->completion_lock);
5283 percpu_ref_put(&ctx->refs);
5287 /* If req is still hashed, it cannot have been canceled. Don't check. */
5288 if (hash_hashed(&req->hash_node))
5289 hash_del(&req->hash_node);
5291 io_poll_remove_double(req);
5292 spin_unlock_irq(&ctx->completion_lock);
5294 if (!READ_ONCE(apoll->poll.canceled))
5295 __io_req_task_submit(req);
5297 __io_req_task_cancel(req, -ECANCELED);
5299 percpu_ref_put(&ctx->refs);
5300 kfree(apoll->double_poll);
5304 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5307 struct io_kiocb *req = wait->private;
5308 struct io_poll_iocb *poll = &req->apoll->poll;
5310 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5313 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5316 static void io_poll_req_insert(struct io_kiocb *req)
5318 struct io_ring_ctx *ctx = req->ctx;
5319 struct hlist_head *list;
5321 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5322 hlist_add_head(&req->hash_node, list);
5325 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5326 struct io_poll_iocb *poll,
5327 struct io_poll_table *ipt, __poll_t mask,
5328 wait_queue_func_t wake_func)
5329 __acquires(&ctx->completion_lock)
5331 struct io_ring_ctx *ctx = req->ctx;
5332 bool cancel = false;
5334 INIT_HLIST_NODE(&req->hash_node);
5335 io_init_poll_iocb(poll, mask, wake_func);
5336 poll->file = req->file;
5337 poll->wait.private = req;
5339 ipt->pt._key = mask;
5341 ipt->error = -EINVAL;
5343 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5345 spin_lock_irq(&ctx->completion_lock);
5346 if (likely(poll->head)) {
5347 spin_lock(&poll->head->lock);
5348 if (unlikely(list_empty(&poll->wait.entry))) {
5354 if (mask || ipt->error)
5355 list_del_init(&poll->wait.entry);
5357 WRITE_ONCE(poll->canceled, true);
5358 else if (!poll->done) /* actually waiting for an event */
5359 io_poll_req_insert(req);
5360 spin_unlock(&poll->head->lock);
5366 static bool io_arm_poll_handler(struct io_kiocb *req)
5368 const struct io_op_def *def = &io_op_defs[req->opcode];
5369 struct io_ring_ctx *ctx = req->ctx;
5370 struct async_poll *apoll;
5371 struct io_poll_table ipt;
5375 if (!req->file || !file_can_poll(req->file))
5377 if (req->flags & REQ_F_POLLED)
5381 else if (def->pollout)
5385 /* if we can't nonblock try, then no point in arming a poll handler */
5386 if (!io_file_supports_async(req->file, rw))
5389 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5390 if (unlikely(!apoll))
5392 apoll->double_poll = NULL;
5394 req->flags |= REQ_F_POLLED;
5399 mask |= POLLIN | POLLRDNORM;
5401 mask |= POLLOUT | POLLWRNORM;
5403 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5404 if ((req->opcode == IORING_OP_RECVMSG) &&
5405 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5408 mask |= POLLERR | POLLPRI;
5410 ipt.pt._qproc = io_async_queue_proc;
5412 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5414 if (ret || ipt.error) {
5415 io_poll_remove_double(req);
5416 spin_unlock_irq(&ctx->completion_lock);
5417 kfree(apoll->double_poll);
5421 spin_unlock_irq(&ctx->completion_lock);
5422 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
5423 apoll->poll.events);
5427 static bool __io_poll_remove_one(struct io_kiocb *req,
5428 struct io_poll_iocb *poll)
5430 bool do_complete = false;
5432 spin_lock(&poll->head->lock);
5433 WRITE_ONCE(poll->canceled, true);
5434 if (!list_empty(&poll->wait.entry)) {
5435 list_del_init(&poll->wait.entry);
5438 spin_unlock(&poll->head->lock);
5439 hash_del(&req->hash_node);
5443 static bool io_poll_remove_one(struct io_kiocb *req)
5447 io_poll_remove_double(req);
5449 if (req->opcode == IORING_OP_POLL_ADD) {
5450 do_complete = __io_poll_remove_one(req, &req->poll);
5452 struct async_poll *apoll = req->apoll;
5454 /* non-poll requests have submit ref still */
5455 do_complete = __io_poll_remove_one(req, &apoll->poll);
5458 kfree(apoll->double_poll);
5464 io_cqring_fill_event(req, -ECANCELED);
5465 io_commit_cqring(req->ctx);
5466 req_set_fail_links(req);
5467 io_put_req_deferred(req, 1);
5474 * Returns true if we found and killed one or more poll requests
5476 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5477 struct files_struct *files)
5479 struct hlist_node *tmp;
5480 struct io_kiocb *req;
5483 spin_lock_irq(&ctx->completion_lock);
5484 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5485 struct hlist_head *list;
5487 list = &ctx->cancel_hash[i];
5488 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5489 if (io_match_task(req, tsk, files))
5490 posted += io_poll_remove_one(req);
5493 spin_unlock_irq(&ctx->completion_lock);
5496 io_cqring_ev_posted(ctx);
5501 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
5503 struct hlist_head *list;
5504 struct io_kiocb *req;
5506 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5507 hlist_for_each_entry(req, list, hash_node) {
5508 if (sqe_addr != req->user_data)
5510 if (io_poll_remove_one(req))
5518 static int io_poll_remove_prep(struct io_kiocb *req,
5519 const struct io_uring_sqe *sqe)
5521 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5523 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
5527 req->poll_remove.addr = READ_ONCE(sqe->addr);
5532 * Find a running poll command that matches one specified in sqe->addr,
5533 * and remove it if found.
5535 static int io_poll_remove(struct io_kiocb *req)
5537 struct io_ring_ctx *ctx = req->ctx;
5540 spin_lock_irq(&ctx->completion_lock);
5541 ret = io_poll_cancel(ctx, req->poll_remove.addr);
5542 spin_unlock_irq(&ctx->completion_lock);
5545 req_set_fail_links(req);
5546 io_req_complete(req, ret);
5550 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5553 struct io_kiocb *req = wait->private;
5554 struct io_poll_iocb *poll = &req->poll;
5556 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5559 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5560 struct poll_table_struct *p)
5562 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5564 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5567 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5569 struct io_poll_iocb *poll = &req->poll;
5572 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5574 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
5577 events = READ_ONCE(sqe->poll32_events);
5579 events = swahw32(events);
5581 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP |
5582 (events & EPOLLEXCLUSIVE);
5586 static int io_poll_add(struct io_kiocb *req)
5588 struct io_poll_iocb *poll = &req->poll;
5589 struct io_ring_ctx *ctx = req->ctx;
5590 struct io_poll_table ipt;
5593 ipt.pt._qproc = io_poll_queue_proc;
5595 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5598 if (mask) { /* no async, we'd stolen it */
5600 io_poll_complete(req, mask, 0);
5602 spin_unlock_irq(&ctx->completion_lock);
5605 io_cqring_ev_posted(ctx);
5611 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5613 struct io_timeout_data *data = container_of(timer,
5614 struct io_timeout_data, timer);
5615 struct io_kiocb *req = data->req;
5616 struct io_ring_ctx *ctx = req->ctx;
5617 unsigned long flags;
5619 spin_lock_irqsave(&ctx->completion_lock, flags);
5620 list_del_init(&req->timeout.list);
5621 atomic_set(&req->ctx->cq_timeouts,
5622 atomic_read(&req->ctx->cq_timeouts) + 1);
5624 io_cqring_fill_event(req, -ETIME);
5625 io_commit_cqring(ctx);
5626 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5628 io_cqring_ev_posted(ctx);
5629 req_set_fail_links(req);
5631 return HRTIMER_NORESTART;
5634 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5637 struct io_timeout_data *io;
5638 struct io_kiocb *req;
5641 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5642 if (user_data == req->user_data) {
5649 return ERR_PTR(ret);
5651 io = req->async_data;
5652 ret = hrtimer_try_to_cancel(&io->timer);
5654 return ERR_PTR(-EALREADY);
5655 list_del_init(&req->timeout.list);
5659 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5661 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5664 return PTR_ERR(req);
5666 req_set_fail_links(req);
5667 io_cqring_fill_event(req, -ECANCELED);
5668 io_put_req_deferred(req, 1);
5672 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5673 struct timespec64 *ts, enum hrtimer_mode mode)
5675 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5676 struct io_timeout_data *data;
5679 return PTR_ERR(req);
5681 req->timeout.off = 0; /* noseq */
5682 data = req->async_data;
5683 list_add_tail(&req->timeout.list, &ctx->timeout_list);
5684 hrtimer_init(&data->timer, CLOCK_MONOTONIC, mode);
5685 data->timer.function = io_timeout_fn;
5686 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
5690 static int io_timeout_remove_prep(struct io_kiocb *req,
5691 const struct io_uring_sqe *sqe)
5693 struct io_timeout_rem *tr = &req->timeout_rem;
5695 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5697 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5699 if (sqe->ioprio || sqe->buf_index || sqe->len)
5702 tr->addr = READ_ONCE(sqe->addr);
5703 tr->flags = READ_ONCE(sqe->timeout_flags);
5704 if (tr->flags & IORING_TIMEOUT_UPDATE) {
5705 if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS))
5707 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
5709 } else if (tr->flags) {
5710 /* timeout removal doesn't support flags */
5718 * Remove or update an existing timeout command
5720 static int io_timeout_remove(struct io_kiocb *req)
5722 struct io_timeout_rem *tr = &req->timeout_rem;
5723 struct io_ring_ctx *ctx = req->ctx;
5726 spin_lock_irq(&ctx->completion_lock);
5727 if (req->timeout_rem.flags & IORING_TIMEOUT_UPDATE) {
5728 enum hrtimer_mode mode = (tr->flags & IORING_TIMEOUT_ABS)
5729 ? HRTIMER_MODE_ABS : HRTIMER_MODE_REL;
5731 ret = io_timeout_update(ctx, tr->addr, &tr->ts, mode);
5733 ret = io_timeout_cancel(ctx, tr->addr);
5736 io_cqring_fill_event(req, ret);
5737 io_commit_cqring(ctx);
5738 spin_unlock_irq(&ctx->completion_lock);
5739 io_cqring_ev_posted(ctx);
5741 req_set_fail_links(req);
5746 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5747 bool is_timeout_link)
5749 struct io_timeout_data *data;
5751 u32 off = READ_ONCE(sqe->off);
5753 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5755 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5757 if (off && is_timeout_link)
5759 flags = READ_ONCE(sqe->timeout_flags);
5760 if (flags & ~IORING_TIMEOUT_ABS)
5763 req->timeout.off = off;
5765 if (!req->async_data && io_alloc_async_data(req))
5768 data = req->async_data;
5771 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5774 if (flags & IORING_TIMEOUT_ABS)
5775 data->mode = HRTIMER_MODE_ABS;
5777 data->mode = HRTIMER_MODE_REL;
5779 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5783 static int io_timeout(struct io_kiocb *req)
5785 struct io_ring_ctx *ctx = req->ctx;
5786 struct io_timeout_data *data = req->async_data;
5787 struct list_head *entry;
5788 u32 tail, off = req->timeout.off;
5790 spin_lock_irq(&ctx->completion_lock);
5793 * sqe->off holds how many events that need to occur for this
5794 * timeout event to be satisfied. If it isn't set, then this is
5795 * a pure timeout request, sequence isn't used.
5797 if (io_is_timeout_noseq(req)) {
5798 entry = ctx->timeout_list.prev;
5802 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5803 req->timeout.target_seq = tail + off;
5806 * Insertion sort, ensuring the first entry in the list is always
5807 * the one we need first.
5809 list_for_each_prev(entry, &ctx->timeout_list) {
5810 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5813 if (io_is_timeout_noseq(nxt))
5815 /* nxt.seq is behind @tail, otherwise would've been completed */
5816 if (off >= nxt->timeout.target_seq - tail)
5820 list_add(&req->timeout.list, entry);
5821 data->timer.function = io_timeout_fn;
5822 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5823 spin_unlock_irq(&ctx->completion_lock);
5827 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5829 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5831 return req->user_data == (unsigned long) data;
5834 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
5836 enum io_wq_cancel cancel_ret;
5839 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr, false);
5840 switch (cancel_ret) {
5841 case IO_WQ_CANCEL_OK:
5844 case IO_WQ_CANCEL_RUNNING:
5847 case IO_WQ_CANCEL_NOTFOUND:
5855 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5856 struct io_kiocb *req, __u64 sqe_addr,
5859 unsigned long flags;
5862 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
5863 if (ret != -ENOENT) {
5864 spin_lock_irqsave(&ctx->completion_lock, flags);
5868 spin_lock_irqsave(&ctx->completion_lock, flags);
5869 ret = io_timeout_cancel(ctx, sqe_addr);
5872 ret = io_poll_cancel(ctx, sqe_addr);
5876 io_cqring_fill_event(req, ret);
5877 io_commit_cqring(ctx);
5878 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5879 io_cqring_ev_posted(ctx);
5882 req_set_fail_links(req);
5886 static int io_async_cancel_prep(struct io_kiocb *req,
5887 const struct io_uring_sqe *sqe)
5889 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5891 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5893 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5896 req->cancel.addr = READ_ONCE(sqe->addr);
5900 static int io_async_cancel(struct io_kiocb *req)
5902 struct io_ring_ctx *ctx = req->ctx;
5904 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
5908 static int io_files_update_prep(struct io_kiocb *req,
5909 const struct io_uring_sqe *sqe)
5911 if (unlikely(req->ctx->flags & IORING_SETUP_SQPOLL))
5913 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5915 if (sqe->ioprio || sqe->rw_flags)
5918 req->files_update.offset = READ_ONCE(sqe->off);
5919 req->files_update.nr_args = READ_ONCE(sqe->len);
5920 if (!req->files_update.nr_args)
5922 req->files_update.arg = READ_ONCE(sqe->addr);
5926 static int io_files_update(struct io_kiocb *req, bool force_nonblock,
5927 struct io_comp_state *cs)
5929 struct io_ring_ctx *ctx = req->ctx;
5930 struct io_uring_files_update up;
5936 up.offset = req->files_update.offset;
5937 up.fds = req->files_update.arg;
5939 mutex_lock(&ctx->uring_lock);
5940 ret = __io_sqe_files_update(ctx, &up, req->files_update.nr_args);
5941 mutex_unlock(&ctx->uring_lock);
5944 req_set_fail_links(req);
5945 __io_req_complete(req, ret, 0, cs);
5949 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5951 switch (req->opcode) {
5954 case IORING_OP_READV:
5955 case IORING_OP_READ_FIXED:
5956 case IORING_OP_READ:
5957 return io_read_prep(req, sqe);
5958 case IORING_OP_WRITEV:
5959 case IORING_OP_WRITE_FIXED:
5960 case IORING_OP_WRITE:
5961 return io_write_prep(req, sqe);
5962 case IORING_OP_POLL_ADD:
5963 return io_poll_add_prep(req, sqe);
5964 case IORING_OP_POLL_REMOVE:
5965 return io_poll_remove_prep(req, sqe);
5966 case IORING_OP_FSYNC:
5967 return io_prep_fsync(req, sqe);
5968 case IORING_OP_SYNC_FILE_RANGE:
5969 return io_prep_sfr(req, sqe);
5970 case IORING_OP_SENDMSG:
5971 case IORING_OP_SEND:
5972 return io_sendmsg_prep(req, sqe);
5973 case IORING_OP_RECVMSG:
5974 case IORING_OP_RECV:
5975 return io_recvmsg_prep(req, sqe);
5976 case IORING_OP_CONNECT:
5977 return io_connect_prep(req, sqe);
5978 case IORING_OP_TIMEOUT:
5979 return io_timeout_prep(req, sqe, false);
5980 case IORING_OP_TIMEOUT_REMOVE:
5981 return io_timeout_remove_prep(req, sqe);
5982 case IORING_OP_ASYNC_CANCEL:
5983 return io_async_cancel_prep(req, sqe);
5984 case IORING_OP_LINK_TIMEOUT:
5985 return io_timeout_prep(req, sqe, true);
5986 case IORING_OP_ACCEPT:
5987 return io_accept_prep(req, sqe);
5988 case IORING_OP_FALLOCATE:
5989 return io_fallocate_prep(req, sqe);
5990 case IORING_OP_OPENAT:
5991 return io_openat_prep(req, sqe);
5992 case IORING_OP_CLOSE:
5993 return io_close_prep(req, sqe);
5994 case IORING_OP_FILES_UPDATE:
5995 return io_files_update_prep(req, sqe);
5996 case IORING_OP_STATX:
5997 return io_statx_prep(req, sqe);
5998 case IORING_OP_FADVISE:
5999 return io_fadvise_prep(req, sqe);
6000 case IORING_OP_MADVISE:
6001 return io_madvise_prep(req, sqe);
6002 case IORING_OP_OPENAT2:
6003 return io_openat2_prep(req, sqe);
6004 case IORING_OP_EPOLL_CTL:
6005 return io_epoll_ctl_prep(req, sqe);
6006 case IORING_OP_SPLICE:
6007 return io_splice_prep(req, sqe);
6008 case IORING_OP_PROVIDE_BUFFERS:
6009 return io_provide_buffers_prep(req, sqe);
6010 case IORING_OP_REMOVE_BUFFERS:
6011 return io_remove_buffers_prep(req, sqe);
6013 return io_tee_prep(req, sqe);
6014 case IORING_OP_SHUTDOWN:
6015 return io_shutdown_prep(req, sqe);
6016 case IORING_OP_RENAMEAT:
6017 return io_renameat_prep(req, sqe);
6018 case IORING_OP_UNLINKAT:
6019 return io_unlinkat_prep(req, sqe);
6022 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
6027 static int io_req_defer_prep(struct io_kiocb *req,
6028 const struct io_uring_sqe *sqe)
6032 if (io_alloc_async_data(req))
6034 return io_req_prep(req, sqe);
6037 static u32 io_get_sequence(struct io_kiocb *req)
6039 struct io_kiocb *pos;
6040 struct io_ring_ctx *ctx = req->ctx;
6041 u32 total_submitted, nr_reqs = 0;
6043 io_for_each_link(pos, req)
6046 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
6047 return total_submitted - nr_reqs;
6050 static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe)
6052 struct io_ring_ctx *ctx = req->ctx;
6053 struct io_defer_entry *de;
6057 /* Still need defer if there is pending req in defer list. */
6058 if (likely(list_empty_careful(&ctx->defer_list) &&
6059 !(req->flags & REQ_F_IO_DRAIN)))
6062 seq = io_get_sequence(req);
6063 /* Still a chance to pass the sequence check */
6064 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
6067 if (!req->async_data) {
6068 ret = io_req_defer_prep(req, sqe);
6072 io_prep_async_link(req);
6073 de = kmalloc(sizeof(*de), GFP_KERNEL);
6077 spin_lock_irq(&ctx->completion_lock);
6078 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
6079 spin_unlock_irq(&ctx->completion_lock);
6081 io_queue_async_work(req);
6082 return -EIOCBQUEUED;
6085 trace_io_uring_defer(ctx, req, req->user_data);
6088 list_add_tail(&de->list, &ctx->defer_list);
6089 spin_unlock_irq(&ctx->completion_lock);
6090 return -EIOCBQUEUED;
6093 static void io_req_drop_files(struct io_kiocb *req)
6095 struct io_ring_ctx *ctx = req->ctx;
6096 struct io_uring_task *tctx = req->task->io_uring;
6097 unsigned long flags;
6099 put_files_struct(req->work.identity->files);
6100 put_nsproxy(req->work.identity->nsproxy);
6101 spin_lock_irqsave(&ctx->inflight_lock, flags);
6102 list_del(&req->inflight_entry);
6103 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
6104 req->flags &= ~REQ_F_INFLIGHT;
6105 req->work.flags &= ~IO_WQ_WORK_FILES;
6106 if (atomic_read(&tctx->in_idle))
6107 wake_up(&tctx->wait);
6110 static void __io_clean_op(struct io_kiocb *req)
6112 if (req->flags & REQ_F_BUFFER_SELECTED) {
6113 switch (req->opcode) {
6114 case IORING_OP_READV:
6115 case IORING_OP_READ_FIXED:
6116 case IORING_OP_READ:
6117 kfree((void *)(unsigned long)req->rw.addr);
6119 case IORING_OP_RECVMSG:
6120 case IORING_OP_RECV:
6121 kfree(req->sr_msg.kbuf);
6124 req->flags &= ~REQ_F_BUFFER_SELECTED;
6127 if (req->flags & REQ_F_NEED_CLEANUP) {
6128 switch (req->opcode) {
6129 case IORING_OP_READV:
6130 case IORING_OP_READ_FIXED:
6131 case IORING_OP_READ:
6132 case IORING_OP_WRITEV:
6133 case IORING_OP_WRITE_FIXED:
6134 case IORING_OP_WRITE: {
6135 struct io_async_rw *io = req->async_data;
6137 kfree(io->free_iovec);
6140 case IORING_OP_RECVMSG:
6141 case IORING_OP_SENDMSG: {
6142 struct io_async_msghdr *io = req->async_data;
6143 if (io->iov != io->fast_iov)
6147 case IORING_OP_SPLICE:
6149 io_put_file(req, req->splice.file_in,
6150 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
6152 case IORING_OP_OPENAT:
6153 case IORING_OP_OPENAT2:
6154 if (req->open.filename)
6155 putname(req->open.filename);
6157 case IORING_OP_RENAMEAT:
6158 putname(req->rename.oldpath);
6159 putname(req->rename.newpath);
6161 case IORING_OP_UNLINKAT:
6162 putname(req->unlink.filename);
6165 req->flags &= ~REQ_F_NEED_CLEANUP;
6168 if (req->flags & REQ_F_INFLIGHT)
6169 io_req_drop_files(req);
6172 static int io_issue_sqe(struct io_kiocb *req, bool force_nonblock,
6173 struct io_comp_state *cs)
6175 struct io_ring_ctx *ctx = req->ctx;
6178 switch (req->opcode) {
6180 ret = io_nop(req, cs);
6182 case IORING_OP_READV:
6183 case IORING_OP_READ_FIXED:
6184 case IORING_OP_READ:
6185 ret = io_read(req, force_nonblock, cs);
6187 case IORING_OP_WRITEV:
6188 case IORING_OP_WRITE_FIXED:
6189 case IORING_OP_WRITE:
6190 ret = io_write(req, force_nonblock, cs);
6192 case IORING_OP_FSYNC:
6193 ret = io_fsync(req, force_nonblock);
6195 case IORING_OP_POLL_ADD:
6196 ret = io_poll_add(req);
6198 case IORING_OP_POLL_REMOVE:
6199 ret = io_poll_remove(req);
6201 case IORING_OP_SYNC_FILE_RANGE:
6202 ret = io_sync_file_range(req, force_nonblock);
6204 case IORING_OP_SENDMSG:
6205 ret = io_sendmsg(req, force_nonblock, cs);
6207 case IORING_OP_SEND:
6208 ret = io_send(req, force_nonblock, cs);
6210 case IORING_OP_RECVMSG:
6211 ret = io_recvmsg(req, force_nonblock, cs);
6213 case IORING_OP_RECV:
6214 ret = io_recv(req, force_nonblock, cs);
6216 case IORING_OP_TIMEOUT:
6217 ret = io_timeout(req);
6219 case IORING_OP_TIMEOUT_REMOVE:
6220 ret = io_timeout_remove(req);
6222 case IORING_OP_ACCEPT:
6223 ret = io_accept(req, force_nonblock, cs);
6225 case IORING_OP_CONNECT:
6226 ret = io_connect(req, force_nonblock, cs);
6228 case IORING_OP_ASYNC_CANCEL:
6229 ret = io_async_cancel(req);
6231 case IORING_OP_FALLOCATE:
6232 ret = io_fallocate(req, force_nonblock);
6234 case IORING_OP_OPENAT:
6235 ret = io_openat(req, force_nonblock);
6237 case IORING_OP_CLOSE:
6238 ret = io_close(req, force_nonblock, cs);
6240 case IORING_OP_FILES_UPDATE:
6241 ret = io_files_update(req, force_nonblock, cs);
6243 case IORING_OP_STATX:
6244 ret = io_statx(req, force_nonblock);
6246 case IORING_OP_FADVISE:
6247 ret = io_fadvise(req, force_nonblock);
6249 case IORING_OP_MADVISE:
6250 ret = io_madvise(req, force_nonblock);
6252 case IORING_OP_OPENAT2:
6253 ret = io_openat2(req, force_nonblock);
6255 case IORING_OP_EPOLL_CTL:
6256 ret = io_epoll_ctl(req, force_nonblock, cs);
6258 case IORING_OP_SPLICE:
6259 ret = io_splice(req, force_nonblock);
6261 case IORING_OP_PROVIDE_BUFFERS:
6262 ret = io_provide_buffers(req, force_nonblock, cs);
6264 case IORING_OP_REMOVE_BUFFERS:
6265 ret = io_remove_buffers(req, force_nonblock, cs);
6268 ret = io_tee(req, force_nonblock);
6270 case IORING_OP_SHUTDOWN:
6271 ret = io_shutdown(req, force_nonblock);
6273 case IORING_OP_RENAMEAT:
6274 ret = io_renameat(req, force_nonblock);
6276 case IORING_OP_UNLINKAT:
6277 ret = io_unlinkat(req, force_nonblock);
6287 /* If the op doesn't have a file, we're not polling for it */
6288 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
6289 const bool in_async = io_wq_current_is_worker();
6291 /* workqueue context doesn't hold uring_lock, grab it now */
6293 mutex_lock(&ctx->uring_lock);
6295 io_iopoll_req_issued(req, in_async);
6298 mutex_unlock(&ctx->uring_lock);
6304 static struct io_wq_work *io_wq_submit_work(struct io_wq_work *work)
6306 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6307 struct io_kiocb *timeout;
6310 timeout = io_prep_linked_timeout(req);
6312 io_queue_linked_timeout(timeout);
6314 /* if NO_CANCEL is set, we must still run the work */
6315 if ((work->flags & (IO_WQ_WORK_CANCEL|IO_WQ_WORK_NO_CANCEL)) ==
6316 IO_WQ_WORK_CANCEL) {
6322 ret = io_issue_sqe(req, false, NULL);
6324 * We can get EAGAIN for polled IO even though we're
6325 * forcing a sync submission from here, since we can't
6326 * wait for request slots on the block side.
6335 struct io_ring_ctx *lock_ctx = NULL;
6337 if (req->ctx->flags & IORING_SETUP_IOPOLL)
6338 lock_ctx = req->ctx;
6341 * io_iopoll_complete() does not hold completion_lock to
6342 * complete polled io, so here for polled io, we can not call
6343 * io_req_complete() directly, otherwise there maybe concurrent
6344 * access to cqring, defer_list, etc, which is not safe. Given
6345 * that io_iopoll_complete() is always called under uring_lock,
6346 * so here for polled io, we also get uring_lock to complete
6350 mutex_lock(&lock_ctx->uring_lock);
6352 req_set_fail_links(req);
6353 io_req_complete(req, ret);
6356 mutex_unlock(&lock_ctx->uring_lock);
6359 return io_steal_work(req);
6362 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6365 struct fixed_file_table *table;
6367 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
6368 return table->files[index & IORING_FILE_TABLE_MASK];
6371 static struct file *io_file_get(struct io_submit_state *state,
6372 struct io_kiocb *req, int fd, bool fixed)
6374 struct io_ring_ctx *ctx = req->ctx;
6378 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6380 fd = array_index_nospec(fd, ctx->nr_user_files);
6381 file = io_file_from_index(ctx, fd);
6382 io_set_resource_node(req);
6384 trace_io_uring_file_get(ctx, fd);
6385 file = __io_file_get(state, fd);
6391 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6393 struct io_timeout_data *data = container_of(timer,
6394 struct io_timeout_data, timer);
6395 struct io_kiocb *prev, *req = data->req;
6396 struct io_ring_ctx *ctx = req->ctx;
6397 unsigned long flags;
6399 spin_lock_irqsave(&ctx->completion_lock, flags);
6400 prev = req->timeout.head;
6401 req->timeout.head = NULL;
6404 * We don't expect the list to be empty, that will only happen if we
6405 * race with the completion of the linked work.
6407 if (prev && refcount_inc_not_zero(&prev->refs))
6408 io_remove_next_linked(prev);
6411 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6414 req_set_fail_links(prev);
6415 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6418 io_req_complete(req, -ETIME);
6420 return HRTIMER_NORESTART;
6423 static void __io_queue_linked_timeout(struct io_kiocb *req)
6426 * If the back reference is NULL, then our linked request finished
6427 * before we got a chance to setup the timer
6429 if (req->timeout.head) {
6430 struct io_timeout_data *data = req->async_data;
6432 data->timer.function = io_link_timeout_fn;
6433 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6438 static void io_queue_linked_timeout(struct io_kiocb *req)
6440 struct io_ring_ctx *ctx = req->ctx;
6442 spin_lock_irq(&ctx->completion_lock);
6443 __io_queue_linked_timeout(req);
6444 spin_unlock_irq(&ctx->completion_lock);
6446 /* drop submission reference */
6450 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6452 struct io_kiocb *nxt = req->link;
6454 if (!nxt || (req->flags & REQ_F_LINK_TIMEOUT) ||
6455 nxt->opcode != IORING_OP_LINK_TIMEOUT)
6458 nxt->timeout.head = req;
6459 nxt->flags |= REQ_F_LTIMEOUT_ACTIVE;
6460 req->flags |= REQ_F_LINK_TIMEOUT;
6464 static void __io_queue_sqe(struct io_kiocb *req, struct io_comp_state *cs)
6466 struct io_kiocb *linked_timeout;
6467 const struct cred *old_creds = NULL;
6471 linked_timeout = io_prep_linked_timeout(req);
6473 if ((req->flags & REQ_F_WORK_INITIALIZED) &&
6474 (req->work.flags & IO_WQ_WORK_CREDS) &&
6475 req->work.identity->creds != current_cred()) {
6477 revert_creds(old_creds);
6478 if (old_creds == req->work.identity->creds)
6479 old_creds = NULL; /* restored original creds */
6481 old_creds = override_creds(req->work.identity->creds);
6484 ret = io_issue_sqe(req, true, cs);
6487 * We async punt it if the file wasn't marked NOWAIT, or if the file
6488 * doesn't support non-blocking read/write attempts
6490 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6491 if (!io_arm_poll_handler(req)) {
6493 * Queued up for async execution, worker will release
6494 * submit reference when the iocb is actually submitted.
6496 io_queue_async_work(req);
6500 io_queue_linked_timeout(linked_timeout);
6501 } else if (likely(!ret)) {
6502 /* drop submission reference */
6503 req = io_put_req_find_next(req);
6505 io_queue_linked_timeout(linked_timeout);
6508 if (!(req->flags & REQ_F_FORCE_ASYNC))
6510 io_queue_async_work(req);
6513 /* un-prep timeout, so it'll be killed as any other linked */
6514 req->flags &= ~REQ_F_LINK_TIMEOUT;
6515 req_set_fail_links(req);
6517 io_req_complete(req, ret);
6521 revert_creds(old_creds);
6524 static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6525 struct io_comp_state *cs)
6529 ret = io_req_defer(req, sqe);
6531 if (ret != -EIOCBQUEUED) {
6533 req_set_fail_links(req);
6535 io_req_complete(req, ret);
6537 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6538 if (!req->async_data) {
6539 ret = io_req_defer_prep(req, sqe);
6543 io_queue_async_work(req);
6546 ret = io_req_prep(req, sqe);
6550 __io_queue_sqe(req, cs);
6554 static inline void io_queue_link_head(struct io_kiocb *req,
6555 struct io_comp_state *cs)
6557 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
6559 io_req_complete(req, -ECANCELED);
6561 io_queue_sqe(req, NULL, cs);
6564 struct io_submit_link {
6565 struct io_kiocb *head;
6566 struct io_kiocb *last;
6569 static int io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6570 struct io_submit_link *link, struct io_comp_state *cs)
6572 struct io_ring_ctx *ctx = req->ctx;
6576 * If we already have a head request, queue this one for async
6577 * submittal once the head completes. If we don't have a head but
6578 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6579 * submitted sync once the chain is complete. If none of those
6580 * conditions are true (normal request), then just queue it.
6583 struct io_kiocb *head = link->head;
6586 * Taking sequential execution of a link, draining both sides
6587 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6588 * requests in the link. So, it drains the head and the
6589 * next after the link request. The last one is done via
6590 * drain_next flag to persist the effect across calls.
6592 if (req->flags & REQ_F_IO_DRAIN) {
6593 head->flags |= REQ_F_IO_DRAIN;
6594 ctx->drain_next = 1;
6596 ret = io_req_defer_prep(req, sqe);
6597 if (unlikely(ret)) {
6598 /* fail even hard links since we don't submit */
6599 head->flags |= REQ_F_FAIL_LINK;
6602 trace_io_uring_link(ctx, req, head);
6603 link->last->link = req;
6606 /* last request of a link, enqueue the link */
6607 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6608 io_queue_link_head(head, cs);
6612 if (unlikely(ctx->drain_next)) {
6613 req->flags |= REQ_F_IO_DRAIN;
6614 ctx->drain_next = 0;
6616 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6617 ret = io_req_defer_prep(req, sqe);
6619 req->flags |= REQ_F_FAIL_LINK;
6623 io_queue_sqe(req, sqe, cs);
6631 * Batched submission is done, ensure local IO is flushed out.
6633 static void io_submit_state_end(struct io_submit_state *state)
6635 if (!list_empty(&state->comp.list))
6636 io_submit_flush_completions(&state->comp);
6637 if (state->plug_started)
6638 blk_finish_plug(&state->plug);
6639 io_state_file_put(state);
6640 if (state->free_reqs)
6641 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
6645 * Start submission side cache.
6647 static void io_submit_state_start(struct io_submit_state *state,
6648 struct io_ring_ctx *ctx, unsigned int max_ios)
6650 state->plug_started = false;
6652 INIT_LIST_HEAD(&state->comp.list);
6653 state->comp.ctx = ctx;
6654 state->free_reqs = 0;
6655 state->file_refs = 0;
6656 state->ios_left = max_ios;
6659 static void io_commit_sqring(struct io_ring_ctx *ctx)
6661 struct io_rings *rings = ctx->rings;
6664 * Ensure any loads from the SQEs are done at this point,
6665 * since once we write the new head, the application could
6666 * write new data to them.
6668 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6672 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6673 * that is mapped by userspace. This means that care needs to be taken to
6674 * ensure that reads are stable, as we cannot rely on userspace always
6675 * being a good citizen. If members of the sqe are validated and then later
6676 * used, it's important that those reads are done through READ_ONCE() to
6677 * prevent a re-load down the line.
6679 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6681 u32 *sq_array = ctx->sq_array;
6685 * The cached sq head (or cq tail) serves two purposes:
6687 * 1) allows us to batch the cost of updating the user visible
6689 * 2) allows the kernel side to track the head on its own, even
6690 * though the application is the one updating it.
6692 head = READ_ONCE(sq_array[ctx->cached_sq_head & ctx->sq_mask]);
6693 if (likely(head < ctx->sq_entries))
6694 return &ctx->sq_sqes[head];
6696 /* drop invalid entries */
6697 ctx->cached_sq_dropped++;
6698 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6702 static inline void io_consume_sqe(struct io_ring_ctx *ctx)
6704 ctx->cached_sq_head++;
6708 * Check SQE restrictions (opcode and flags).
6710 * Returns 'true' if SQE is allowed, 'false' otherwise.
6712 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6713 struct io_kiocb *req,
6714 unsigned int sqe_flags)
6716 if (!ctx->restricted)
6719 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6722 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6723 ctx->restrictions.sqe_flags_required)
6726 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6727 ctx->restrictions.sqe_flags_required))
6733 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
6734 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
6735 IOSQE_BUFFER_SELECT)
6737 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6738 const struct io_uring_sqe *sqe,
6739 struct io_submit_state *state)
6741 unsigned int sqe_flags;
6744 req->opcode = READ_ONCE(sqe->opcode);
6745 req->user_data = READ_ONCE(sqe->user_data);
6746 req->async_data = NULL;
6751 req->fixed_file_refs = NULL;
6752 /* one is dropped after submission, the other at completion */
6753 refcount_set(&req->refs, 2);
6754 req->task = current;
6757 if (unlikely(req->opcode >= IORING_OP_LAST))
6760 if (unlikely(io_sq_thread_acquire_mm_files(ctx, req)))
6763 sqe_flags = READ_ONCE(sqe->flags);
6764 /* enforce forwards compatibility on users */
6765 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6768 if (unlikely(!io_check_restriction(ctx, req, sqe_flags)))
6771 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6772 !io_op_defs[req->opcode].buffer_select)
6775 id = READ_ONCE(sqe->personality);
6777 struct io_identity *iod;
6779 iod = idr_find(&ctx->personality_idr, id);
6782 refcount_inc(&iod->count);
6784 __io_req_init_async(req);
6785 get_cred(iod->creds);
6786 req->work.identity = iod;
6787 req->work.flags |= IO_WQ_WORK_CREDS;
6790 /* same numerical values with corresponding REQ_F_*, safe to copy */
6791 req->flags |= sqe_flags;
6794 * Plug now if we have more than 1 IO left after this, and the target
6795 * is potentially a read/write to block based storage.
6797 if (!state->plug_started && state->ios_left > 1 &&
6798 io_op_defs[req->opcode].plug) {
6799 blk_start_plug(&state->plug);
6800 state->plug_started = true;
6804 if (io_op_defs[req->opcode].needs_file) {
6805 bool fixed = req->flags & REQ_F_FIXED_FILE;
6807 req->file = io_file_get(state, req, READ_ONCE(sqe->fd), fixed);
6808 if (unlikely(!req->file &&
6809 !io_op_defs[req->opcode].needs_file_no_error))
6817 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6819 struct io_submit_state state;
6820 struct io_submit_link link;
6821 int i, submitted = 0;
6823 /* if we have a backlog and couldn't flush it all, return BUSY */
6824 if (test_bit(0, &ctx->sq_check_overflow)) {
6825 if (!io_cqring_overflow_flush(ctx, false, NULL, NULL))
6829 /* make sure SQ entry isn't read before tail */
6830 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6832 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6835 percpu_counter_add(¤t->io_uring->inflight, nr);
6836 refcount_add(nr, ¤t->usage);
6838 io_submit_state_start(&state, ctx, nr);
6841 for (i = 0; i < nr; i++) {
6842 const struct io_uring_sqe *sqe;
6843 struct io_kiocb *req;
6846 sqe = io_get_sqe(ctx);
6847 if (unlikely(!sqe)) {
6848 io_consume_sqe(ctx);
6851 req = io_alloc_req(ctx, &state);
6852 if (unlikely(!req)) {
6854 submitted = -EAGAIN;
6857 io_consume_sqe(ctx);
6858 /* will complete beyond this point, count as submitted */
6861 err = io_init_req(ctx, req, sqe, &state);
6862 if (unlikely(err)) {
6865 io_req_complete(req, err);
6869 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6870 true, io_async_submit(ctx));
6871 err = io_submit_sqe(req, sqe, &link, &state.comp);
6876 if (unlikely(submitted != nr)) {
6877 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6878 struct io_uring_task *tctx = current->io_uring;
6879 int unused = nr - ref_used;
6881 percpu_ref_put_many(&ctx->refs, unused);
6882 percpu_counter_sub(&tctx->inflight, unused);
6883 put_task_struct_many(current, unused);
6886 io_queue_link_head(link.head, &state.comp);
6887 io_submit_state_end(&state);
6889 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6890 io_commit_sqring(ctx);
6895 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6897 /* Tell userspace we may need a wakeup call */
6898 spin_lock_irq(&ctx->completion_lock);
6899 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6900 spin_unlock_irq(&ctx->completion_lock);
6903 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6905 spin_lock_irq(&ctx->completion_lock);
6906 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6907 spin_unlock_irq(&ctx->completion_lock);
6910 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
6912 unsigned int to_submit;
6915 to_submit = io_sqring_entries(ctx);
6916 /* if we're handling multiple rings, cap submit size for fairness */
6917 if (cap_entries && to_submit > 8)
6920 if (!list_empty(&ctx->iopoll_list) || to_submit) {
6921 unsigned nr_events = 0;
6923 mutex_lock(&ctx->uring_lock);
6924 if (!list_empty(&ctx->iopoll_list))
6925 io_do_iopoll(ctx, &nr_events, 0);
6927 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)))
6928 ret = io_submit_sqes(ctx, to_submit);
6929 mutex_unlock(&ctx->uring_lock);
6932 if (!io_sqring_full(ctx) && wq_has_sleeper(&ctx->sqo_sq_wait))
6933 wake_up(&ctx->sqo_sq_wait);
6938 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
6940 struct io_ring_ctx *ctx;
6941 unsigned sq_thread_idle = 0;
6943 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6944 if (sq_thread_idle < ctx->sq_thread_idle)
6945 sq_thread_idle = ctx->sq_thread_idle;
6948 sqd->sq_thread_idle = sq_thread_idle;
6951 static void io_sqd_init_new(struct io_sq_data *sqd)
6953 struct io_ring_ctx *ctx;
6955 while (!list_empty(&sqd->ctx_new_list)) {
6956 ctx = list_first_entry(&sqd->ctx_new_list, struct io_ring_ctx, sqd_list);
6957 list_move_tail(&ctx->sqd_list, &sqd->ctx_list);
6958 complete(&ctx->sq_thread_comp);
6961 io_sqd_update_thread_idle(sqd);
6964 static int io_sq_thread(void *data)
6966 struct cgroup_subsys_state *cur_css = NULL;
6967 struct files_struct *old_files = current->files;
6968 struct nsproxy *old_nsproxy = current->nsproxy;
6969 const struct cred *old_cred = NULL;
6970 struct io_sq_data *sqd = data;
6971 struct io_ring_ctx *ctx;
6972 unsigned long timeout = 0;
6976 current->files = NULL;
6977 current->nsproxy = NULL;
6978 task_unlock(current);
6980 while (!kthread_should_stop()) {
6982 bool cap_entries, sqt_spin, needs_sched;
6985 * Any changes to the sqd lists are synchronized through the
6986 * kthread parking. This synchronizes the thread vs users,
6987 * the users are synchronized on the sqd->ctx_lock.
6989 if (kthread_should_park()) {
6992 * When sq thread is unparked, in case the previous park operation
6993 * comes from io_put_sq_data(), which means that sq thread is going
6994 * to be stopped, so here needs to have a check.
6996 if (kthread_should_stop())
7000 if (unlikely(!list_empty(&sqd->ctx_new_list))) {
7001 io_sqd_init_new(sqd);
7002 timeout = jiffies + sqd->sq_thread_idle;
7006 cap_entries = !list_is_singular(&sqd->ctx_list);
7007 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
7008 if (current->cred != ctx->creds) {
7010 revert_creds(old_cred);
7011 old_cred = override_creds(ctx->creds);
7013 io_sq_thread_associate_blkcg(ctx, &cur_css);
7015 current->loginuid = ctx->loginuid;
7016 current->sessionid = ctx->sessionid;
7019 ret = __io_sq_thread(ctx, cap_entries);
7020 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
7023 io_sq_thread_drop_mm_files();
7026 if (sqt_spin || !time_after(jiffies, timeout)) {
7030 timeout = jiffies + sqd->sq_thread_idle;
7034 if (kthread_should_park())
7038 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
7039 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
7040 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
7041 !list_empty_careful(&ctx->iopoll_list)) {
7042 needs_sched = false;
7045 if (io_sqring_entries(ctx)) {
7046 needs_sched = false;
7052 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7053 io_ring_set_wakeup_flag(ctx);
7056 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7057 io_ring_clear_wakeup_flag(ctx);
7060 finish_wait(&sqd->wait, &wait);
7061 timeout = jiffies + sqd->sq_thread_idle;
7067 io_sq_thread_unassociate_blkcg();
7069 revert_creds(old_cred);
7072 current->files = old_files;
7073 current->nsproxy = old_nsproxy;
7074 task_unlock(current);
7081 struct io_wait_queue {
7082 struct wait_queue_entry wq;
7083 struct io_ring_ctx *ctx;
7085 unsigned nr_timeouts;
7088 static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush)
7090 struct io_ring_ctx *ctx = iowq->ctx;
7093 * Wake up if we have enough events, or if a timeout occurred since we
7094 * started waiting. For timeouts, we always want to return to userspace,
7095 * regardless of event count.
7097 return io_cqring_events(ctx, noflush) >= iowq->to_wait ||
7098 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
7101 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
7102 int wake_flags, void *key)
7104 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
7107 /* use noflush == true, as we can't safely rely on locking context */
7108 if (!io_should_wake(iowq, true))
7111 return autoremove_wake_function(curr, mode, wake_flags, key);
7114 static int io_run_task_work_sig(void)
7116 if (io_run_task_work())
7118 if (!signal_pending(current))
7120 if (test_tsk_thread_flag(current, TIF_NOTIFY_SIGNAL))
7121 return -ERESTARTSYS;
7126 * Wait until events become available, if we don't already have some. The
7127 * application must reap them itself, as they reside on the shared cq ring.
7129 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
7130 const sigset_t __user *sig, size_t sigsz,
7131 struct __kernel_timespec __user *uts)
7133 struct io_wait_queue iowq = {
7136 .func = io_wake_function,
7137 .entry = LIST_HEAD_INIT(iowq.wq.entry),
7140 .to_wait = min_events,
7142 struct io_rings *rings = ctx->rings;
7143 struct timespec64 ts;
7144 signed long timeout = 0;
7148 if (io_cqring_events(ctx, false) >= min_events)
7150 if (!io_run_task_work())
7155 #ifdef CONFIG_COMPAT
7156 if (in_compat_syscall())
7157 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
7161 ret = set_user_sigmask(sig, sigsz);
7168 if (get_timespec64(&ts, uts))
7170 timeout = timespec64_to_jiffies(&ts);
7173 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
7174 trace_io_uring_cqring_wait(ctx, min_events);
7176 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
7177 TASK_INTERRUPTIBLE);
7178 /* make sure we run task_work before checking for signals */
7179 ret = io_run_task_work_sig();
7184 if (io_should_wake(&iowq, false))
7187 timeout = schedule_timeout(timeout);
7196 finish_wait(&ctx->wait, &iowq.wq);
7198 restore_saved_sigmask_unless(ret == -EINTR);
7200 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
7203 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
7205 #if defined(CONFIG_UNIX)
7206 if (ctx->ring_sock) {
7207 struct sock *sock = ctx->ring_sock->sk;
7208 struct sk_buff *skb;
7210 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
7216 for (i = 0; i < ctx->nr_user_files; i++) {
7219 file = io_file_from_index(ctx, i);
7226 static void io_file_ref_kill(struct percpu_ref *ref)
7228 struct fixed_file_data *data;
7230 data = container_of(ref, struct fixed_file_data, refs);
7231 complete(&data->done);
7234 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7236 struct fixed_file_data *data = ctx->file_data;
7237 struct fixed_file_ref_node *ref_node = NULL;
7238 unsigned nr_tables, i;
7243 spin_lock_bh(&data->lock);
7244 ref_node = data->node;
7245 spin_unlock_bh(&data->lock);
7247 percpu_ref_kill(&ref_node->refs);
7249 percpu_ref_kill(&data->refs);
7251 /* wait for all refs nodes to complete */
7252 flush_delayed_work(&ctx->file_put_work);
7253 wait_for_completion(&data->done);
7255 __io_sqe_files_unregister(ctx);
7256 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
7257 for (i = 0; i < nr_tables; i++)
7258 kfree(data->table[i].files);
7260 percpu_ref_exit(&data->refs);
7262 ctx->file_data = NULL;
7263 ctx->nr_user_files = 0;
7267 static void io_put_sq_data(struct io_sq_data *sqd)
7269 if (refcount_dec_and_test(&sqd->refs)) {
7271 * The park is a bit of a work-around, without it we get
7272 * warning spews on shutdown with SQPOLL set and affinity
7273 * set to a single CPU.
7276 kthread_park(sqd->thread);
7277 kthread_stop(sqd->thread);
7284 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7286 struct io_ring_ctx *ctx_attach;
7287 struct io_sq_data *sqd;
7290 f = fdget(p->wq_fd);
7292 return ERR_PTR(-ENXIO);
7293 if (f.file->f_op != &io_uring_fops) {
7295 return ERR_PTR(-EINVAL);
7298 ctx_attach = f.file->private_data;
7299 sqd = ctx_attach->sq_data;
7302 return ERR_PTR(-EINVAL);
7305 refcount_inc(&sqd->refs);
7310 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p)
7312 struct io_sq_data *sqd;
7314 if (p->flags & IORING_SETUP_ATTACH_WQ)
7315 return io_attach_sq_data(p);
7317 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7319 return ERR_PTR(-ENOMEM);
7321 refcount_set(&sqd->refs, 1);
7322 INIT_LIST_HEAD(&sqd->ctx_list);
7323 INIT_LIST_HEAD(&sqd->ctx_new_list);
7324 mutex_init(&sqd->ctx_lock);
7325 mutex_init(&sqd->lock);
7326 init_waitqueue_head(&sqd->wait);
7330 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7331 __releases(&sqd->lock)
7335 kthread_unpark(sqd->thread);
7336 mutex_unlock(&sqd->lock);
7339 static void io_sq_thread_park(struct io_sq_data *sqd)
7340 __acquires(&sqd->lock)
7344 mutex_lock(&sqd->lock);
7345 kthread_park(sqd->thread);
7348 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
7350 struct io_sq_data *sqd = ctx->sq_data;
7355 * We may arrive here from the error branch in
7356 * io_sq_offload_create() where the kthread is created
7357 * without being waked up, thus wake it up now to make
7358 * sure the wait will complete.
7360 wake_up_process(sqd->thread);
7361 wait_for_completion(&ctx->sq_thread_comp);
7363 io_sq_thread_park(sqd);
7366 mutex_lock(&sqd->ctx_lock);
7367 list_del(&ctx->sqd_list);
7368 io_sqd_update_thread_idle(sqd);
7369 mutex_unlock(&sqd->ctx_lock);
7372 io_sq_thread_unpark(sqd);
7374 io_put_sq_data(sqd);
7375 ctx->sq_data = NULL;
7379 static void io_finish_async(struct io_ring_ctx *ctx)
7381 io_sq_thread_stop(ctx);
7384 io_wq_destroy(ctx->io_wq);
7389 #if defined(CONFIG_UNIX)
7391 * Ensure the UNIX gc is aware of our file set, so we are certain that
7392 * the io_uring can be safely unregistered on process exit, even if we have
7393 * loops in the file referencing.
7395 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7397 struct sock *sk = ctx->ring_sock->sk;
7398 struct scm_fp_list *fpl;
7399 struct sk_buff *skb;
7402 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7406 skb = alloc_skb(0, GFP_KERNEL);
7415 fpl->user = get_uid(ctx->user);
7416 for (i = 0; i < nr; i++) {
7417 struct file *file = io_file_from_index(ctx, i + offset);
7421 fpl->fp[nr_files] = get_file(file);
7422 unix_inflight(fpl->user, fpl->fp[nr_files]);
7427 fpl->max = SCM_MAX_FD;
7428 fpl->count = nr_files;
7429 UNIXCB(skb).fp = fpl;
7430 skb->destructor = unix_destruct_scm;
7431 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7432 skb_queue_head(&sk->sk_receive_queue, skb);
7434 for (i = 0; i < nr_files; i++)
7445 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7446 * causes regular reference counting to break down. We rely on the UNIX
7447 * garbage collection to take care of this problem for us.
7449 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7451 unsigned left, total;
7455 left = ctx->nr_user_files;
7457 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7459 ret = __io_sqe_files_scm(ctx, this_files, total);
7463 total += this_files;
7469 while (total < ctx->nr_user_files) {
7470 struct file *file = io_file_from_index(ctx, total);
7480 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7486 static int io_sqe_alloc_file_tables(struct fixed_file_data *file_data,
7487 unsigned nr_tables, unsigned nr_files)
7491 for (i = 0; i < nr_tables; i++) {
7492 struct fixed_file_table *table = &file_data->table[i];
7493 unsigned this_files;
7495 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
7496 table->files = kcalloc(this_files, sizeof(struct file *),
7500 nr_files -= this_files;
7506 for (i = 0; i < nr_tables; i++) {
7507 struct fixed_file_table *table = &file_data->table[i];
7508 kfree(table->files);
7513 static void io_ring_file_put(struct io_ring_ctx *ctx, struct file *file)
7515 #if defined(CONFIG_UNIX)
7516 struct sock *sock = ctx->ring_sock->sk;
7517 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7518 struct sk_buff *skb;
7521 __skb_queue_head_init(&list);
7524 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7525 * remove this entry and rearrange the file array.
7527 skb = skb_dequeue(head);
7529 struct scm_fp_list *fp;
7531 fp = UNIXCB(skb).fp;
7532 for (i = 0; i < fp->count; i++) {
7535 if (fp->fp[i] != file)
7538 unix_notinflight(fp->user, fp->fp[i]);
7539 left = fp->count - 1 - i;
7541 memmove(&fp->fp[i], &fp->fp[i + 1],
7542 left * sizeof(struct file *));
7549 __skb_queue_tail(&list, skb);
7559 __skb_queue_tail(&list, skb);
7561 skb = skb_dequeue(head);
7564 if (skb_peek(&list)) {
7565 spin_lock_irq(&head->lock);
7566 while ((skb = __skb_dequeue(&list)) != NULL)
7567 __skb_queue_tail(head, skb);
7568 spin_unlock_irq(&head->lock);
7575 struct io_file_put {
7576 struct list_head list;
7580 static void __io_file_put_work(struct fixed_file_ref_node *ref_node)
7582 struct fixed_file_data *file_data = ref_node->file_data;
7583 struct io_ring_ctx *ctx = file_data->ctx;
7584 struct io_file_put *pfile, *tmp;
7586 list_for_each_entry_safe(pfile, tmp, &ref_node->file_list, list) {
7587 list_del(&pfile->list);
7588 io_ring_file_put(ctx, pfile->file);
7592 percpu_ref_exit(&ref_node->refs);
7594 percpu_ref_put(&file_data->refs);
7597 static void io_file_put_work(struct work_struct *work)
7599 struct io_ring_ctx *ctx;
7600 struct llist_node *node;
7602 ctx = container_of(work, struct io_ring_ctx, file_put_work.work);
7603 node = llist_del_all(&ctx->file_put_llist);
7606 struct fixed_file_ref_node *ref_node;
7607 struct llist_node *next = node->next;
7609 ref_node = llist_entry(node, struct fixed_file_ref_node, llist);
7610 __io_file_put_work(ref_node);
7615 static void io_file_data_ref_zero(struct percpu_ref *ref)
7617 struct fixed_file_ref_node *ref_node;
7618 struct fixed_file_data *data;
7619 struct io_ring_ctx *ctx;
7620 bool first_add = false;
7623 ref_node = container_of(ref, struct fixed_file_ref_node, refs);
7624 data = ref_node->file_data;
7627 spin_lock_bh(&data->lock);
7628 ref_node->done = true;
7630 while (!list_empty(&data->ref_list)) {
7631 ref_node = list_first_entry(&data->ref_list,
7632 struct fixed_file_ref_node, node);
7633 /* recycle ref nodes in order */
7634 if (!ref_node->done)
7636 list_del(&ref_node->node);
7637 first_add |= llist_add(&ref_node->llist, &ctx->file_put_llist);
7639 spin_unlock_bh(&data->lock);
7641 if (percpu_ref_is_dying(&data->refs))
7645 mod_delayed_work(system_wq, &ctx->file_put_work, 0);
7647 queue_delayed_work(system_wq, &ctx->file_put_work, delay);
7650 static struct fixed_file_ref_node *alloc_fixed_file_ref_node(
7651 struct io_ring_ctx *ctx)
7653 struct fixed_file_ref_node *ref_node;
7655 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7657 return ERR_PTR(-ENOMEM);
7659 if (percpu_ref_init(&ref_node->refs, io_file_data_ref_zero,
7662 return ERR_PTR(-ENOMEM);
7664 INIT_LIST_HEAD(&ref_node->node);
7665 INIT_LIST_HEAD(&ref_node->file_list);
7666 ref_node->file_data = ctx->file_data;
7667 ref_node->done = false;
7671 static void destroy_fixed_file_ref_node(struct fixed_file_ref_node *ref_node)
7673 percpu_ref_exit(&ref_node->refs);
7677 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7680 __s32 __user *fds = (__s32 __user *) arg;
7681 unsigned nr_tables, i;
7683 int fd, ret = -ENOMEM;
7684 struct fixed_file_ref_node *ref_node;
7685 struct fixed_file_data *file_data;
7691 if (nr_args > IORING_MAX_FIXED_FILES)
7694 file_data = kzalloc(sizeof(*ctx->file_data), GFP_KERNEL);
7697 file_data->ctx = ctx;
7698 init_completion(&file_data->done);
7699 INIT_LIST_HEAD(&file_data->ref_list);
7700 spin_lock_init(&file_data->lock);
7702 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
7703 file_data->table = kcalloc(nr_tables, sizeof(*file_data->table),
7705 if (!file_data->table)
7708 if (percpu_ref_init(&file_data->refs, io_file_ref_kill,
7709 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
7712 if (io_sqe_alloc_file_tables(file_data, nr_tables, nr_args))
7714 ctx->file_data = file_data;
7716 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7717 struct fixed_file_table *table;
7720 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7724 /* allow sparse sets */
7734 * Don't allow io_uring instances to be registered. If UNIX
7735 * isn't enabled, then this causes a reference cycle and this
7736 * instance can never get freed. If UNIX is enabled we'll
7737 * handle it just fine, but there's still no point in allowing
7738 * a ring fd as it doesn't support regular read/write anyway.
7740 if (file->f_op == &io_uring_fops) {
7744 table = &file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7745 index = i & IORING_FILE_TABLE_MASK;
7746 table->files[index] = file;
7749 ret = io_sqe_files_scm(ctx);
7751 io_sqe_files_unregister(ctx);
7755 ref_node = alloc_fixed_file_ref_node(ctx);
7756 if (IS_ERR(ref_node)) {
7757 io_sqe_files_unregister(ctx);
7758 return PTR_ERR(ref_node);
7761 file_data->node = ref_node;
7762 spin_lock_bh(&file_data->lock);
7763 list_add_tail(&ref_node->node, &file_data->ref_list);
7764 spin_unlock_bh(&file_data->lock);
7765 percpu_ref_get(&file_data->refs);
7768 for (i = 0; i < ctx->nr_user_files; i++) {
7769 file = io_file_from_index(ctx, i);
7773 for (i = 0; i < nr_tables; i++)
7774 kfree(file_data->table[i].files);
7775 ctx->nr_user_files = 0;
7777 percpu_ref_exit(&file_data->refs);
7779 kfree(file_data->table);
7781 ctx->file_data = NULL;
7785 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7788 #if defined(CONFIG_UNIX)
7789 struct sock *sock = ctx->ring_sock->sk;
7790 struct sk_buff_head *head = &sock->sk_receive_queue;
7791 struct sk_buff *skb;
7794 * See if we can merge this file into an existing skb SCM_RIGHTS
7795 * file set. If there's no room, fall back to allocating a new skb
7796 * and filling it in.
7798 spin_lock_irq(&head->lock);
7799 skb = skb_peek(head);
7801 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7803 if (fpl->count < SCM_MAX_FD) {
7804 __skb_unlink(skb, head);
7805 spin_unlock_irq(&head->lock);
7806 fpl->fp[fpl->count] = get_file(file);
7807 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7809 spin_lock_irq(&head->lock);
7810 __skb_queue_head(head, skb);
7815 spin_unlock_irq(&head->lock);
7822 return __io_sqe_files_scm(ctx, 1, index);
7828 static int io_queue_file_removal(struct fixed_file_data *data,
7831 struct io_file_put *pfile;
7832 struct fixed_file_ref_node *ref_node = data->node;
7834 pfile = kzalloc(sizeof(*pfile), GFP_KERNEL);
7839 list_add(&pfile->list, &ref_node->file_list);
7844 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7845 struct io_uring_files_update *up,
7848 struct fixed_file_data *data = ctx->file_data;
7849 struct fixed_file_ref_node *ref_node;
7854 bool needs_switch = false;
7856 if (check_add_overflow(up->offset, nr_args, &done))
7858 if (done > ctx->nr_user_files)
7861 ref_node = alloc_fixed_file_ref_node(ctx);
7862 if (IS_ERR(ref_node))
7863 return PTR_ERR(ref_node);
7866 fds = u64_to_user_ptr(up->fds);
7868 struct fixed_file_table *table;
7872 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
7876 i = array_index_nospec(up->offset, ctx->nr_user_files);
7877 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7878 index = i & IORING_FILE_TABLE_MASK;
7879 if (table->files[index]) {
7880 file = table->files[index];
7881 err = io_queue_file_removal(data, file);
7884 table->files[index] = NULL;
7885 needs_switch = true;
7894 * Don't allow io_uring instances to be registered. If
7895 * UNIX isn't enabled, then this causes a reference
7896 * cycle and this instance can never get freed. If UNIX
7897 * is enabled we'll handle it just fine, but there's
7898 * still no point in allowing a ring fd as it doesn't
7899 * support regular read/write anyway.
7901 if (file->f_op == &io_uring_fops) {
7906 table->files[index] = file;
7907 err = io_sqe_file_register(ctx, file, i);
7909 table->files[index] = NULL;
7920 percpu_ref_kill(&data->node->refs);
7921 spin_lock_bh(&data->lock);
7922 list_add_tail(&ref_node->node, &data->ref_list);
7923 data->node = ref_node;
7924 spin_unlock_bh(&data->lock);
7925 percpu_ref_get(&ctx->file_data->refs);
7927 destroy_fixed_file_ref_node(ref_node);
7929 return done ? done : err;
7932 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
7935 struct io_uring_files_update up;
7937 if (!ctx->file_data)
7941 if (copy_from_user(&up, arg, sizeof(up)))
7946 return __io_sqe_files_update(ctx, &up, nr_args);
7949 static void io_free_work(struct io_wq_work *work)
7951 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7953 /* Consider that io_steal_work() relies on this ref */
7957 static int io_init_wq_offload(struct io_ring_ctx *ctx,
7958 struct io_uring_params *p)
7960 struct io_wq_data data;
7962 struct io_ring_ctx *ctx_attach;
7963 unsigned int concurrency;
7966 data.user = ctx->user;
7967 data.free_work = io_free_work;
7968 data.do_work = io_wq_submit_work;
7970 if (!(p->flags & IORING_SETUP_ATTACH_WQ)) {
7971 /* Do QD, or 4 * CPUS, whatever is smallest */
7972 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7974 ctx->io_wq = io_wq_create(concurrency, &data);
7975 if (IS_ERR(ctx->io_wq)) {
7976 ret = PTR_ERR(ctx->io_wq);
7982 f = fdget(p->wq_fd);
7986 if (f.file->f_op != &io_uring_fops) {
7991 ctx_attach = f.file->private_data;
7992 /* @io_wq is protected by holding the fd */
7993 if (!io_wq_get(ctx_attach->io_wq, &data)) {
7998 ctx->io_wq = ctx_attach->io_wq;
8004 static int io_uring_alloc_task_context(struct task_struct *task)
8006 struct io_uring_task *tctx;
8009 tctx = kmalloc(sizeof(*tctx), GFP_KERNEL);
8010 if (unlikely(!tctx))
8013 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
8014 if (unlikely(ret)) {
8020 init_waitqueue_head(&tctx->wait);
8022 atomic_set(&tctx->in_idle, 0);
8023 tctx->sqpoll = false;
8024 io_init_identity(&tctx->__identity);
8025 tctx->identity = &tctx->__identity;
8026 task->io_uring = tctx;
8030 void __io_uring_free(struct task_struct *tsk)
8032 struct io_uring_task *tctx = tsk->io_uring;
8034 WARN_ON_ONCE(!xa_empty(&tctx->xa));
8035 WARN_ON_ONCE(refcount_read(&tctx->identity->count) != 1);
8036 if (tctx->identity != &tctx->__identity)
8037 kfree(tctx->identity);
8038 percpu_counter_destroy(&tctx->inflight);
8040 tsk->io_uring = NULL;
8043 static int io_sq_offload_create(struct io_ring_ctx *ctx,
8044 struct io_uring_params *p)
8048 if (ctx->flags & IORING_SETUP_SQPOLL) {
8049 struct io_sq_data *sqd;
8052 if (!capable(CAP_SYS_ADMIN) && !capable(CAP_SYS_NICE))
8055 sqd = io_get_sq_data(p);
8062 io_sq_thread_park(sqd);
8063 mutex_lock(&sqd->ctx_lock);
8064 list_add(&ctx->sqd_list, &sqd->ctx_new_list);
8065 mutex_unlock(&sqd->ctx_lock);
8066 io_sq_thread_unpark(sqd);
8068 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
8069 if (!ctx->sq_thread_idle)
8070 ctx->sq_thread_idle = HZ;
8075 if (p->flags & IORING_SETUP_SQ_AFF) {
8076 int cpu = p->sq_thread_cpu;
8079 if (cpu >= nr_cpu_ids)
8081 if (!cpu_online(cpu))
8084 sqd->thread = kthread_create_on_cpu(io_sq_thread, sqd,
8085 cpu, "io_uring-sq");
8087 sqd->thread = kthread_create(io_sq_thread, sqd,
8090 if (IS_ERR(sqd->thread)) {
8091 ret = PTR_ERR(sqd->thread);
8095 ret = io_uring_alloc_task_context(sqd->thread);
8098 } else if (p->flags & IORING_SETUP_SQ_AFF) {
8099 /* Can't have SQ_AFF without SQPOLL */
8105 ret = io_init_wq_offload(ctx, p);
8111 io_finish_async(ctx);
8115 static void io_sq_offload_start(struct io_ring_ctx *ctx)
8117 struct io_sq_data *sqd = ctx->sq_data;
8119 if ((ctx->flags & IORING_SETUP_SQPOLL) && sqd->thread)
8120 wake_up_process(sqd->thread);
8123 static inline void __io_unaccount_mem(struct user_struct *user,
8124 unsigned long nr_pages)
8126 atomic_long_sub(nr_pages, &user->locked_vm);
8129 static inline int __io_account_mem(struct user_struct *user,
8130 unsigned long nr_pages)
8132 unsigned long page_limit, cur_pages, new_pages;
8134 /* Don't allow more pages than we can safely lock */
8135 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
8138 cur_pages = atomic_long_read(&user->locked_vm);
8139 new_pages = cur_pages + nr_pages;
8140 if (new_pages > page_limit)
8142 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
8143 new_pages) != cur_pages);
8148 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
8149 enum io_mem_account acct)
8152 __io_unaccount_mem(ctx->user, nr_pages);
8154 if (ctx->mm_account) {
8155 if (acct == ACCT_LOCKED) {
8156 mmap_write_lock(ctx->mm_account);
8157 ctx->mm_account->locked_vm -= nr_pages;
8158 mmap_write_unlock(ctx->mm_account);
8159 }else if (acct == ACCT_PINNED) {
8160 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
8165 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
8166 enum io_mem_account acct)
8170 if (ctx->limit_mem) {
8171 ret = __io_account_mem(ctx->user, nr_pages);
8176 if (ctx->mm_account) {
8177 if (acct == ACCT_LOCKED) {
8178 mmap_write_lock(ctx->mm_account);
8179 ctx->mm_account->locked_vm += nr_pages;
8180 mmap_write_unlock(ctx->mm_account);
8181 } else if (acct == ACCT_PINNED) {
8182 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
8189 static void io_mem_free(void *ptr)
8196 page = virt_to_head_page(ptr);
8197 if (put_page_testzero(page))
8198 free_compound_page(page);
8201 static void *io_mem_alloc(size_t size)
8203 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
8206 return (void *) __get_free_pages(gfp_flags, get_order(size));
8209 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8212 struct io_rings *rings;
8213 size_t off, sq_array_size;
8215 off = struct_size(rings, cqes, cq_entries);
8216 if (off == SIZE_MAX)
8220 off = ALIGN(off, SMP_CACHE_BYTES);
8228 sq_array_size = array_size(sizeof(u32), sq_entries);
8229 if (sq_array_size == SIZE_MAX)
8232 if (check_add_overflow(off, sq_array_size, &off))
8238 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
8242 pages = (size_t)1 << get_order(
8243 rings_size(sq_entries, cq_entries, NULL));
8244 pages += (size_t)1 << get_order(
8245 array_size(sizeof(struct io_uring_sqe), sq_entries));
8250 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
8254 if (!ctx->user_bufs)
8257 for (i = 0; i < ctx->nr_user_bufs; i++) {
8258 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8260 for (j = 0; j < imu->nr_bvecs; j++)
8261 unpin_user_page(imu->bvec[j].bv_page);
8263 if (imu->acct_pages)
8264 io_unaccount_mem(ctx, imu->acct_pages, ACCT_PINNED);
8269 kfree(ctx->user_bufs);
8270 ctx->user_bufs = NULL;
8271 ctx->nr_user_bufs = 0;
8275 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8276 void __user *arg, unsigned index)
8278 struct iovec __user *src;
8280 #ifdef CONFIG_COMPAT
8282 struct compat_iovec __user *ciovs;
8283 struct compat_iovec ciov;
8285 ciovs = (struct compat_iovec __user *) arg;
8286 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8289 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8290 dst->iov_len = ciov.iov_len;
8294 src = (struct iovec __user *) arg;
8295 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8301 * Not super efficient, but this is just a registration time. And we do cache
8302 * the last compound head, so generally we'll only do a full search if we don't
8305 * We check if the given compound head page has already been accounted, to
8306 * avoid double accounting it. This allows us to account the full size of the
8307 * page, not just the constituent pages of a huge page.
8309 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8310 int nr_pages, struct page *hpage)
8314 /* check current page array */
8315 for (i = 0; i < nr_pages; i++) {
8316 if (!PageCompound(pages[i]))
8318 if (compound_head(pages[i]) == hpage)
8322 /* check previously registered pages */
8323 for (i = 0; i < ctx->nr_user_bufs; i++) {
8324 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8326 for (j = 0; j < imu->nr_bvecs; j++) {
8327 if (!PageCompound(imu->bvec[j].bv_page))
8329 if (compound_head(imu->bvec[j].bv_page) == hpage)
8337 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8338 int nr_pages, struct io_mapped_ubuf *imu,
8339 struct page **last_hpage)
8343 for (i = 0; i < nr_pages; i++) {
8344 if (!PageCompound(pages[i])) {
8349 hpage = compound_head(pages[i]);
8350 if (hpage == *last_hpage)
8352 *last_hpage = hpage;
8353 if (headpage_already_acct(ctx, pages, i, hpage))
8355 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8359 if (!imu->acct_pages)
8362 ret = io_account_mem(ctx, imu->acct_pages, ACCT_PINNED);
8364 imu->acct_pages = 0;
8368 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
8371 struct vm_area_struct **vmas = NULL;
8372 struct page **pages = NULL;
8373 struct page *last_hpage = NULL;
8374 int i, j, got_pages = 0;
8379 if (!nr_args || nr_args > UIO_MAXIOV)
8382 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
8384 if (!ctx->user_bufs)
8387 for (i = 0; i < nr_args; i++) {
8388 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8389 unsigned long off, start, end, ubuf;
8394 ret = io_copy_iov(ctx, &iov, arg, i);
8399 * Don't impose further limits on the size and buffer
8400 * constraints here, we'll -EINVAL later when IO is
8401 * submitted if they are wrong.
8404 if (!iov.iov_base || !iov.iov_len)
8407 /* arbitrary limit, but we need something */
8408 if (iov.iov_len > SZ_1G)
8411 ubuf = (unsigned long) iov.iov_base;
8412 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8413 start = ubuf >> PAGE_SHIFT;
8414 nr_pages = end - start;
8417 if (!pages || nr_pages > got_pages) {
8420 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
8422 vmas = kvmalloc_array(nr_pages,
8423 sizeof(struct vm_area_struct *),
8425 if (!pages || !vmas) {
8429 got_pages = nr_pages;
8432 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
8439 mmap_read_lock(current->mm);
8440 pret = pin_user_pages(ubuf, nr_pages,
8441 FOLL_WRITE | FOLL_LONGTERM,
8443 if (pret == nr_pages) {
8444 /* don't support file backed memory */
8445 for (j = 0; j < nr_pages; j++) {
8446 struct vm_area_struct *vma = vmas[j];
8449 !is_file_hugepages(vma->vm_file)) {
8455 ret = pret < 0 ? pret : -EFAULT;
8457 mmap_read_unlock(current->mm);
8460 * if we did partial map, or found file backed vmas,
8461 * release any pages we did get
8464 unpin_user_pages(pages, pret);
8469 ret = io_buffer_account_pin(ctx, pages, pret, imu, &last_hpage);
8471 unpin_user_pages(pages, pret);
8476 off = ubuf & ~PAGE_MASK;
8478 for (j = 0; j < nr_pages; j++) {
8481 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8482 imu->bvec[j].bv_page = pages[j];
8483 imu->bvec[j].bv_len = vec_len;
8484 imu->bvec[j].bv_offset = off;
8488 /* store original address for later verification */
8490 imu->len = iov.iov_len;
8491 imu->nr_bvecs = nr_pages;
8493 ctx->nr_user_bufs++;
8501 io_sqe_buffer_unregister(ctx);
8505 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8507 __s32 __user *fds = arg;
8513 if (copy_from_user(&fd, fds, sizeof(*fds)))
8516 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8517 if (IS_ERR(ctx->cq_ev_fd)) {
8518 int ret = PTR_ERR(ctx->cq_ev_fd);
8519 ctx->cq_ev_fd = NULL;
8526 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8528 if (ctx->cq_ev_fd) {
8529 eventfd_ctx_put(ctx->cq_ev_fd);
8530 ctx->cq_ev_fd = NULL;
8537 static int __io_destroy_buffers(int id, void *p, void *data)
8539 struct io_ring_ctx *ctx = data;
8540 struct io_buffer *buf = p;
8542 __io_remove_buffers(ctx, buf, id, -1U);
8546 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8548 idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx);
8549 idr_destroy(&ctx->io_buffer_idr);
8552 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8554 io_finish_async(ctx);
8555 io_sqe_buffer_unregister(ctx);
8557 if (ctx->sqo_task) {
8558 put_task_struct(ctx->sqo_task);
8559 ctx->sqo_task = NULL;
8560 mmdrop(ctx->mm_account);
8561 ctx->mm_account = NULL;
8564 #ifdef CONFIG_BLK_CGROUP
8565 if (ctx->sqo_blkcg_css)
8566 css_put(ctx->sqo_blkcg_css);
8569 io_sqe_files_unregister(ctx);
8570 io_eventfd_unregister(ctx);
8571 io_destroy_buffers(ctx);
8572 idr_destroy(&ctx->personality_idr);
8574 #if defined(CONFIG_UNIX)
8575 if (ctx->ring_sock) {
8576 ctx->ring_sock->file = NULL; /* so that iput() is called */
8577 sock_release(ctx->ring_sock);
8581 io_mem_free(ctx->rings);
8582 io_mem_free(ctx->sq_sqes);
8584 percpu_ref_exit(&ctx->refs);
8585 free_uid(ctx->user);
8586 put_cred(ctx->creds);
8587 kfree(ctx->cancel_hash);
8588 kmem_cache_free(req_cachep, ctx->fallback_req);
8592 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8594 struct io_ring_ctx *ctx = file->private_data;
8597 poll_wait(file, &ctx->cq_wait, wait);
8599 * synchronizes with barrier from wq_has_sleeper call in
8603 if (!io_sqring_full(ctx))
8604 mask |= EPOLLOUT | EPOLLWRNORM;
8605 if (io_cqring_events(ctx, false))
8606 mask |= EPOLLIN | EPOLLRDNORM;
8611 static int io_uring_fasync(int fd, struct file *file, int on)
8613 struct io_ring_ctx *ctx = file->private_data;
8615 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8618 static int io_remove_personalities(int id, void *p, void *data)
8620 struct io_ring_ctx *ctx = data;
8621 struct io_identity *iod;
8623 iod = idr_remove(&ctx->personality_idr, id);
8625 put_cred(iod->creds);
8626 if (refcount_dec_and_test(&iod->count))
8632 static void io_ring_exit_work(struct work_struct *work)
8634 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
8638 * If we're doing polled IO and end up having requests being
8639 * submitted async (out-of-line), then completions can come in while
8640 * we're waiting for refs to drop. We need to reap these manually,
8641 * as nobody else will be looking for them.
8644 io_iopoll_try_reap_events(ctx);
8645 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
8646 io_ring_ctx_free(ctx);
8649 static bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
8651 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8653 return req->ctx == data;
8656 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8658 mutex_lock(&ctx->uring_lock);
8659 percpu_ref_kill(&ctx->refs);
8660 /* if force is set, the ring is going away. always drop after that */
8661 ctx->cq_overflow_flushed = 1;
8663 io_cqring_overflow_flush(ctx, true, NULL, NULL);
8664 mutex_unlock(&ctx->uring_lock);
8666 io_kill_timeouts(ctx, NULL, NULL);
8667 io_poll_remove_all(ctx, NULL, NULL);
8670 io_wq_cancel_cb(ctx->io_wq, io_cancel_ctx_cb, ctx, true);
8672 /* if we failed setting up the ctx, we might not have any rings */
8673 io_iopoll_try_reap_events(ctx);
8674 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
8677 * Do this upfront, so we won't have a grace period where the ring
8678 * is closed but resources aren't reaped yet. This can cause
8679 * spurious failure in setting up a new ring.
8681 io_unaccount_mem(ctx, ring_pages(ctx->sq_entries, ctx->cq_entries),
8684 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8686 * Use system_unbound_wq to avoid spawning tons of event kworkers
8687 * if we're exiting a ton of rings at the same time. It just adds
8688 * noise and overhead, there's no discernable change in runtime
8689 * over using system_wq.
8691 queue_work(system_unbound_wq, &ctx->exit_work);
8694 static int io_uring_release(struct inode *inode, struct file *file)
8696 struct io_ring_ctx *ctx = file->private_data;
8698 file->private_data = NULL;
8699 io_ring_ctx_wait_and_kill(ctx);
8703 struct io_task_cancel {
8704 struct task_struct *task;
8705 struct files_struct *files;
8708 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8710 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8711 struct io_task_cancel *cancel = data;
8714 if (cancel->files && (req->flags & REQ_F_LINK_TIMEOUT)) {
8715 unsigned long flags;
8716 struct io_ring_ctx *ctx = req->ctx;
8718 /* protect against races with linked timeouts */
8719 spin_lock_irqsave(&ctx->completion_lock, flags);
8720 ret = io_match_task(req, cancel->task, cancel->files);
8721 spin_unlock_irqrestore(&ctx->completion_lock, flags);
8723 ret = io_match_task(req, cancel->task, cancel->files);
8728 static void io_cancel_defer_files(struct io_ring_ctx *ctx,
8729 struct task_struct *task,
8730 struct files_struct *files)
8732 struct io_defer_entry *de = NULL;
8735 spin_lock_irq(&ctx->completion_lock);
8736 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8737 if (io_match_task(de->req, task, files)) {
8738 list_cut_position(&list, &ctx->defer_list, &de->list);
8742 spin_unlock_irq(&ctx->completion_lock);
8744 while (!list_empty(&list)) {
8745 de = list_first_entry(&list, struct io_defer_entry, list);
8746 list_del_init(&de->list);
8747 req_set_fail_links(de->req);
8748 io_put_req(de->req);
8749 io_req_complete(de->req, -ECANCELED);
8754 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
8755 struct task_struct *task,
8756 struct files_struct *files)
8758 while (!list_empty_careful(&ctx->inflight_list)) {
8759 struct io_task_cancel cancel = { .task = task, .files = files };
8760 struct io_kiocb *req;
8764 spin_lock_irq(&ctx->inflight_lock);
8765 list_for_each_entry(req, &ctx->inflight_list, inflight_entry) {
8766 if (req->task != task ||
8767 req->work.identity->files != files)
8773 prepare_to_wait(&task->io_uring->wait, &wait,
8774 TASK_UNINTERRUPTIBLE);
8775 spin_unlock_irq(&ctx->inflight_lock);
8777 /* We need to keep going until we don't find a matching req */
8781 io_wq_cancel_cb(ctx->io_wq, io_cancel_task_cb, &cancel, true);
8782 io_poll_remove_all(ctx, task, files);
8783 io_kill_timeouts(ctx, task, files);
8784 /* cancellations _may_ trigger task work */
8787 finish_wait(&task->io_uring->wait, &wait);
8791 static void __io_uring_cancel_task_requests(struct io_ring_ctx *ctx,
8792 struct task_struct *task)
8795 struct io_task_cancel cancel = { .task = task, .files = NULL, };
8796 enum io_wq_cancel cret;
8799 cret = io_wq_cancel_cb(ctx->io_wq, io_cancel_task_cb, &cancel, true);
8800 if (cret != IO_WQ_CANCEL_NOTFOUND)
8803 /* SQPOLL thread does its own polling */
8804 if (!(ctx->flags & IORING_SETUP_SQPOLL)) {
8805 while (!list_empty_careful(&ctx->iopoll_list)) {
8806 io_iopoll_try_reap_events(ctx);
8811 ret |= io_poll_remove_all(ctx, task, NULL);
8812 ret |= io_kill_timeouts(ctx, task, NULL);
8813 ret |= io_run_task_work();
8821 * We need to iteratively cancel requests, in case a request has dependent
8822 * hard links. These persist even for failure of cancelations, hence keep
8823 * looping until none are found.
8825 static void io_uring_cancel_task_requests(struct io_ring_ctx *ctx,
8826 struct files_struct *files)
8828 struct task_struct *task = current;
8830 if ((ctx->flags & IORING_SETUP_SQPOLL) && ctx->sq_data) {
8831 task = ctx->sq_data->thread;
8832 atomic_inc(&task->io_uring->in_idle);
8833 io_sq_thread_park(ctx->sq_data);
8836 io_cancel_defer_files(ctx, task, files);
8837 io_ring_submit_lock(ctx, (ctx->flags & IORING_SETUP_IOPOLL));
8838 io_cqring_overflow_flush(ctx, true, task, files);
8839 io_ring_submit_unlock(ctx, (ctx->flags & IORING_SETUP_IOPOLL));
8842 __io_uring_cancel_task_requests(ctx, task);
8844 io_uring_cancel_files(ctx, task, files);
8846 if ((ctx->flags & IORING_SETUP_SQPOLL) && ctx->sq_data) {
8847 atomic_dec(&task->io_uring->in_idle);
8849 * If the files that are going away are the ones in the thread
8850 * identity, clear them out.
8852 if (task->io_uring->identity->files == files)
8853 task->io_uring->identity->files = NULL;
8854 io_sq_thread_unpark(ctx->sq_data);
8859 * Note that this task has used io_uring. We use it for cancelation purposes.
8861 static int io_uring_add_task_file(struct io_ring_ctx *ctx, struct file *file)
8863 struct io_uring_task *tctx = current->io_uring;
8866 if (unlikely(!tctx)) {
8867 ret = io_uring_alloc_task_context(current);
8870 tctx = current->io_uring;
8872 if (tctx->last != file) {
8873 void *old = xa_load(&tctx->xa, (unsigned long)file);
8877 ret = xa_err(xa_store(&tctx->xa, (unsigned long)file,
8888 * This is race safe in that the task itself is doing this, hence it
8889 * cannot be going through the exit/cancel paths at the same time.
8890 * This cannot be modified while exit/cancel is running.
8892 if (!tctx->sqpoll && (ctx->flags & IORING_SETUP_SQPOLL))
8893 tctx->sqpoll = true;
8899 * Remove this io_uring_file -> task mapping.
8901 static void io_uring_del_task_file(struct file *file)
8903 struct io_uring_task *tctx = current->io_uring;
8905 if (tctx->last == file)
8907 file = xa_erase(&tctx->xa, (unsigned long)file);
8913 * Drop task note for this file if we're the only ones that hold it after
8916 static void io_uring_attempt_task_drop(struct file *file)
8918 if (!current->io_uring)
8921 * fput() is pending, will be 2 if the only other ref is our potential
8922 * task file note. If the task is exiting, drop regardless of count.
8924 if (fatal_signal_pending(current) || (current->flags & PF_EXITING) ||
8925 atomic_long_read(&file->f_count) == 2)
8926 io_uring_del_task_file(file);
8929 void __io_uring_files_cancel(struct files_struct *files)
8931 struct io_uring_task *tctx = current->io_uring;
8933 unsigned long index;
8935 /* make sure overflow events are dropped */
8936 atomic_inc(&tctx->in_idle);
8938 xa_for_each(&tctx->xa, index, file) {
8939 struct io_ring_ctx *ctx = file->private_data;
8941 io_uring_cancel_task_requests(ctx, files);
8943 io_uring_del_task_file(file);
8946 atomic_dec(&tctx->in_idle);
8949 static s64 tctx_inflight(struct io_uring_task *tctx)
8951 unsigned long index;
8955 inflight = percpu_counter_sum(&tctx->inflight);
8960 * If we have SQPOLL rings, then we need to iterate and find them, and
8961 * add the pending count for those.
8963 xa_for_each(&tctx->xa, index, file) {
8964 struct io_ring_ctx *ctx = file->private_data;
8966 if (ctx->flags & IORING_SETUP_SQPOLL) {
8967 struct io_uring_task *__tctx = ctx->sqo_task->io_uring;
8969 inflight += percpu_counter_sum(&__tctx->inflight);
8977 * Find any io_uring fd that this task has registered or done IO on, and cancel
8980 void __io_uring_task_cancel(void)
8982 struct io_uring_task *tctx = current->io_uring;
8986 /* make sure overflow events are dropped */
8987 atomic_inc(&tctx->in_idle);
8990 /* read completions before cancelations */
8991 inflight = tctx_inflight(tctx);
8994 __io_uring_files_cancel(NULL);
8996 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
8999 * If we've seen completions, retry. This avoids a race where
9000 * a completion comes in before we did prepare_to_wait().
9002 if (inflight != tctx_inflight(tctx))
9005 finish_wait(&tctx->wait, &wait);
9008 atomic_dec(&tctx->in_idle);
9011 static int io_uring_flush(struct file *file, void *data)
9013 io_uring_attempt_task_drop(file);
9017 static void *io_uring_validate_mmap_request(struct file *file,
9018 loff_t pgoff, size_t sz)
9020 struct io_ring_ctx *ctx = file->private_data;
9021 loff_t offset = pgoff << PAGE_SHIFT;
9026 case IORING_OFF_SQ_RING:
9027 case IORING_OFF_CQ_RING:
9030 case IORING_OFF_SQES:
9034 return ERR_PTR(-EINVAL);
9037 page = virt_to_head_page(ptr);
9038 if (sz > page_size(page))
9039 return ERR_PTR(-EINVAL);
9046 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9048 size_t sz = vma->vm_end - vma->vm_start;
9052 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
9054 return PTR_ERR(ptr);
9056 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
9057 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
9060 #else /* !CONFIG_MMU */
9062 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
9064 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
9067 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
9069 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
9072 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
9073 unsigned long addr, unsigned long len,
9074 unsigned long pgoff, unsigned long flags)
9078 ptr = io_uring_validate_mmap_request(file, pgoff, len);
9080 return PTR_ERR(ptr);
9082 return (unsigned long) ptr;
9085 #endif /* !CONFIG_MMU */
9087 static void io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
9092 if (!io_sqring_full(ctx))
9095 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
9097 if (!io_sqring_full(ctx))
9101 } while (!signal_pending(current));
9103 finish_wait(&ctx->sqo_sq_wait, &wait);
9106 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
9107 struct __kernel_timespec __user **ts,
9108 const sigset_t __user **sig)
9110 struct io_uring_getevents_arg arg;
9113 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
9114 * is just a pointer to the sigset_t.
9116 if (!(flags & IORING_ENTER_EXT_ARG)) {
9117 *sig = (const sigset_t __user *) argp;
9123 * EXT_ARG is set - ensure we agree on the size of it and copy in our
9124 * timespec and sigset_t pointers if good.
9126 if (*argsz != sizeof(arg))
9128 if (copy_from_user(&arg, argp, sizeof(arg)))
9130 *sig = u64_to_user_ptr(arg.sigmask);
9131 *argsz = arg.sigmask_sz;
9132 *ts = u64_to_user_ptr(arg.ts);
9136 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
9137 u32, min_complete, u32, flags, const void __user *, argp,
9140 struct io_ring_ctx *ctx;
9147 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
9148 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG))
9156 if (f.file->f_op != &io_uring_fops)
9160 ctx = f.file->private_data;
9161 if (!percpu_ref_tryget(&ctx->refs))
9165 if (ctx->flags & IORING_SETUP_R_DISABLED)
9169 * For SQ polling, the thread will do all submissions and completions.
9170 * Just return the requested submit count, and wake the thread if
9174 if (ctx->flags & IORING_SETUP_SQPOLL) {
9175 if (!list_empty_careful(&ctx->cq_overflow_list)) {
9176 bool needs_lock = ctx->flags & IORING_SETUP_IOPOLL;
9178 io_ring_submit_lock(ctx, needs_lock);
9179 io_cqring_overflow_flush(ctx, false, NULL, NULL);
9180 io_ring_submit_unlock(ctx, needs_lock);
9182 if (flags & IORING_ENTER_SQ_WAKEUP)
9183 wake_up(&ctx->sq_data->wait);
9184 if (flags & IORING_ENTER_SQ_WAIT)
9185 io_sqpoll_wait_sq(ctx);
9186 submitted = to_submit;
9187 } else if (to_submit) {
9188 ret = io_uring_add_task_file(ctx, f.file);
9191 mutex_lock(&ctx->uring_lock);
9192 submitted = io_submit_sqes(ctx, to_submit);
9193 mutex_unlock(&ctx->uring_lock);
9195 if (submitted != to_submit)
9198 if (flags & IORING_ENTER_GETEVENTS) {
9199 const sigset_t __user *sig;
9200 struct __kernel_timespec __user *ts;
9202 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
9206 min_complete = min(min_complete, ctx->cq_entries);
9209 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9210 * space applications don't need to do io completion events
9211 * polling again, they can rely on io_sq_thread to do polling
9212 * work, which can reduce cpu usage and uring_lock contention.
9214 if (ctx->flags & IORING_SETUP_IOPOLL &&
9215 !(ctx->flags & IORING_SETUP_SQPOLL)) {
9216 ret = io_iopoll_check(ctx, min_complete);
9218 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
9223 percpu_ref_put(&ctx->refs);
9226 return submitted ? submitted : ret;
9229 #ifdef CONFIG_PROC_FS
9230 static int io_uring_show_cred(int id, void *p, void *data)
9232 struct io_identity *iod = p;
9233 const struct cred *cred = iod->creds;
9234 struct seq_file *m = data;
9235 struct user_namespace *uns = seq_user_ns(m);
9236 struct group_info *gi;
9241 seq_printf(m, "%5d\n", id);
9242 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
9243 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
9244 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
9245 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
9246 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
9247 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
9248 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
9249 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
9250 seq_puts(m, "\n\tGroups:\t");
9251 gi = cred->group_info;
9252 for (g = 0; g < gi->ngroups; g++) {
9253 seq_put_decimal_ull(m, g ? " " : "",
9254 from_kgid_munged(uns, gi->gid[g]));
9256 seq_puts(m, "\n\tCapEff:\t");
9257 cap = cred->cap_effective;
9258 CAP_FOR_EACH_U32(__capi)
9259 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
9264 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
9266 struct io_sq_data *sq = NULL;
9271 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9272 * since fdinfo case grabs it in the opposite direction of normal use
9273 * cases. If we fail to get the lock, we just don't iterate any
9274 * structures that could be going away outside the io_uring mutex.
9276 has_lock = mutex_trylock(&ctx->uring_lock);
9278 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL))
9281 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9282 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9283 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9284 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9285 struct fixed_file_table *table;
9288 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
9289 f = table->files[i & IORING_FILE_TABLE_MASK];
9291 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9293 seq_printf(m, "%5u: <none>\n", i);
9295 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9296 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9297 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
9299 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
9300 (unsigned int) buf->len);
9302 if (has_lock && !idr_is_empty(&ctx->personality_idr)) {
9303 seq_printf(m, "Personalities:\n");
9304 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
9306 seq_printf(m, "PollList:\n");
9307 spin_lock_irq(&ctx->completion_lock);
9308 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9309 struct hlist_head *list = &ctx->cancel_hash[i];
9310 struct io_kiocb *req;
9312 hlist_for_each_entry(req, list, hash_node)
9313 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9314 req->task->task_works != NULL);
9316 spin_unlock_irq(&ctx->completion_lock);
9318 mutex_unlock(&ctx->uring_lock);
9321 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9323 struct io_ring_ctx *ctx = f->private_data;
9325 if (percpu_ref_tryget(&ctx->refs)) {
9326 __io_uring_show_fdinfo(ctx, m);
9327 percpu_ref_put(&ctx->refs);
9332 static const struct file_operations io_uring_fops = {
9333 .release = io_uring_release,
9334 .flush = io_uring_flush,
9335 .mmap = io_uring_mmap,
9337 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9338 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9340 .poll = io_uring_poll,
9341 .fasync = io_uring_fasync,
9342 #ifdef CONFIG_PROC_FS
9343 .show_fdinfo = io_uring_show_fdinfo,
9347 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9348 struct io_uring_params *p)
9350 struct io_rings *rings;
9351 size_t size, sq_array_offset;
9353 /* make sure these are sane, as we already accounted them */
9354 ctx->sq_entries = p->sq_entries;
9355 ctx->cq_entries = p->cq_entries;
9357 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9358 if (size == SIZE_MAX)
9361 rings = io_mem_alloc(size);
9366 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9367 rings->sq_ring_mask = p->sq_entries - 1;
9368 rings->cq_ring_mask = p->cq_entries - 1;
9369 rings->sq_ring_entries = p->sq_entries;
9370 rings->cq_ring_entries = p->cq_entries;
9371 ctx->sq_mask = rings->sq_ring_mask;
9372 ctx->cq_mask = rings->cq_ring_mask;
9374 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9375 if (size == SIZE_MAX) {
9376 io_mem_free(ctx->rings);
9381 ctx->sq_sqes = io_mem_alloc(size);
9382 if (!ctx->sq_sqes) {
9383 io_mem_free(ctx->rings);
9391 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
9395 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9399 ret = io_uring_add_task_file(ctx, file);
9404 fd_install(fd, file);
9409 * Allocate an anonymous fd, this is what constitutes the application
9410 * visible backing of an io_uring instance. The application mmaps this
9411 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9412 * we have to tie this fd to a socket for file garbage collection purposes.
9414 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
9417 #if defined(CONFIG_UNIX)
9420 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9423 return ERR_PTR(ret);
9426 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9427 O_RDWR | O_CLOEXEC);
9428 #if defined(CONFIG_UNIX)
9430 sock_release(ctx->ring_sock);
9431 ctx->ring_sock = NULL;
9433 ctx->ring_sock->file = file;
9439 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9440 struct io_uring_params __user *params)
9442 struct user_struct *user = NULL;
9443 struct io_ring_ctx *ctx;
9450 if (entries > IORING_MAX_ENTRIES) {
9451 if (!(p->flags & IORING_SETUP_CLAMP))
9453 entries = IORING_MAX_ENTRIES;
9457 * Use twice as many entries for the CQ ring. It's possible for the
9458 * application to drive a higher depth than the size of the SQ ring,
9459 * since the sqes are only used at submission time. This allows for
9460 * some flexibility in overcommitting a bit. If the application has
9461 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9462 * of CQ ring entries manually.
9464 p->sq_entries = roundup_pow_of_two(entries);
9465 if (p->flags & IORING_SETUP_CQSIZE) {
9467 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9468 * to a power-of-two, if it isn't already. We do NOT impose
9469 * any cq vs sq ring sizing.
9473 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9474 if (!(p->flags & IORING_SETUP_CLAMP))
9476 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9478 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9479 if (p->cq_entries < p->sq_entries)
9482 p->cq_entries = 2 * p->sq_entries;
9485 user = get_uid(current_user());
9486 limit_mem = !capable(CAP_IPC_LOCK);
9489 ret = __io_account_mem(user,
9490 ring_pages(p->sq_entries, p->cq_entries));
9497 ctx = io_ring_ctx_alloc(p);
9500 __io_unaccount_mem(user, ring_pages(p->sq_entries,
9505 ctx->compat = in_compat_syscall();
9507 ctx->creds = get_current_cred();
9509 ctx->loginuid = current->loginuid;
9510 ctx->sessionid = current->sessionid;
9512 ctx->sqo_task = get_task_struct(current);
9515 * This is just grabbed for accounting purposes. When a process exits,
9516 * the mm is exited and dropped before the files, hence we need to hang
9517 * on to this mm purely for the purposes of being able to unaccount
9518 * memory (locked/pinned vm). It's not used for anything else.
9520 mmgrab(current->mm);
9521 ctx->mm_account = current->mm;
9523 #ifdef CONFIG_BLK_CGROUP
9525 * The sq thread will belong to the original cgroup it was inited in.
9526 * If the cgroup goes offline (e.g. disabling the io controller), then
9527 * issued bios will be associated with the closest cgroup later in the
9531 ctx->sqo_blkcg_css = blkcg_css();
9532 ret = css_tryget_online(ctx->sqo_blkcg_css);
9535 /* don't init against a dying cgroup, have the user try again */
9536 ctx->sqo_blkcg_css = NULL;
9543 * Account memory _before_ installing the file descriptor. Once
9544 * the descriptor is installed, it can get closed at any time. Also
9545 * do this before hitting the general error path, as ring freeing
9546 * will un-account as well.
9548 io_account_mem(ctx, ring_pages(p->sq_entries, p->cq_entries),
9550 ctx->limit_mem = limit_mem;
9552 ret = io_allocate_scq_urings(ctx, p);
9556 ret = io_sq_offload_create(ctx, p);
9560 if (!(p->flags & IORING_SETUP_R_DISABLED))
9561 io_sq_offload_start(ctx);
9563 memset(&p->sq_off, 0, sizeof(p->sq_off));
9564 p->sq_off.head = offsetof(struct io_rings, sq.head);
9565 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9566 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9567 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9568 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9569 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9570 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9572 memset(&p->cq_off, 0, sizeof(p->cq_off));
9573 p->cq_off.head = offsetof(struct io_rings, cq.head);
9574 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9575 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9576 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9577 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9578 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9579 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9581 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9582 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9583 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9584 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
9585 IORING_FEAT_EXT_ARG;
9587 if (copy_to_user(params, p, sizeof(*p))) {
9592 file = io_uring_get_file(ctx);
9594 ret = PTR_ERR(file);
9599 * Install ring fd as the very last thing, so we don't risk someone
9600 * having closed it before we finish setup
9602 ret = io_uring_install_fd(ctx, file);
9604 /* fput will clean it up */
9609 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9612 io_ring_ctx_wait_and_kill(ctx);
9617 * Sets up an aio uring context, and returns the fd. Applications asks for a
9618 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9619 * params structure passed in.
9621 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9623 struct io_uring_params p;
9626 if (copy_from_user(&p, params, sizeof(p)))
9628 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9633 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9634 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9635 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9636 IORING_SETUP_R_DISABLED))
9639 return io_uring_create(entries, &p, params);
9642 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9643 struct io_uring_params __user *, params)
9645 return io_uring_setup(entries, params);
9648 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9650 struct io_uring_probe *p;
9654 size = struct_size(p, ops, nr_args);
9655 if (size == SIZE_MAX)
9657 p = kzalloc(size, GFP_KERNEL);
9662 if (copy_from_user(p, arg, size))
9665 if (memchr_inv(p, 0, size))
9668 p->last_op = IORING_OP_LAST - 1;
9669 if (nr_args > IORING_OP_LAST)
9670 nr_args = IORING_OP_LAST;
9672 for (i = 0; i < nr_args; i++) {
9674 if (!io_op_defs[i].not_supported)
9675 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9680 if (copy_to_user(arg, p, size))
9687 static int io_register_personality(struct io_ring_ctx *ctx)
9689 struct io_identity *id;
9692 id = kmalloc(sizeof(*id), GFP_KERNEL);
9696 io_init_identity(id);
9697 id->creds = get_current_cred();
9699 ret = idr_alloc_cyclic(&ctx->personality_idr, id, 1, USHRT_MAX, GFP_KERNEL);
9701 put_cred(id->creds);
9707 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
9709 struct io_identity *iod;
9711 iod = idr_remove(&ctx->personality_idr, id);
9713 put_cred(iod->creds);
9714 if (refcount_dec_and_test(&iod->count))
9722 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9723 unsigned int nr_args)
9725 struct io_uring_restriction *res;
9729 /* Restrictions allowed only if rings started disabled */
9730 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9733 /* We allow only a single restrictions registration */
9734 if (ctx->restrictions.registered)
9737 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9740 size = array_size(nr_args, sizeof(*res));
9741 if (size == SIZE_MAX)
9744 res = memdup_user(arg, size);
9746 return PTR_ERR(res);
9750 for (i = 0; i < nr_args; i++) {
9751 switch (res[i].opcode) {
9752 case IORING_RESTRICTION_REGISTER_OP:
9753 if (res[i].register_op >= IORING_REGISTER_LAST) {
9758 __set_bit(res[i].register_op,
9759 ctx->restrictions.register_op);
9761 case IORING_RESTRICTION_SQE_OP:
9762 if (res[i].sqe_op >= IORING_OP_LAST) {
9767 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9769 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9770 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9772 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9773 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9782 /* Reset all restrictions if an error happened */
9784 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9786 ctx->restrictions.registered = true;
9792 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9794 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9797 if (ctx->restrictions.registered)
9798 ctx->restricted = 1;
9800 ctx->flags &= ~IORING_SETUP_R_DISABLED;
9802 io_sq_offload_start(ctx);
9807 static bool io_register_op_must_quiesce(int op)
9810 case IORING_UNREGISTER_FILES:
9811 case IORING_REGISTER_FILES_UPDATE:
9812 case IORING_REGISTER_PROBE:
9813 case IORING_REGISTER_PERSONALITY:
9814 case IORING_UNREGISTER_PERSONALITY:
9821 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
9822 void __user *arg, unsigned nr_args)
9823 __releases(ctx->uring_lock)
9824 __acquires(ctx->uring_lock)
9829 * We're inside the ring mutex, if the ref is already dying, then
9830 * someone else killed the ctx or is already going through
9831 * io_uring_register().
9833 if (percpu_ref_is_dying(&ctx->refs))
9836 if (io_register_op_must_quiesce(opcode)) {
9837 percpu_ref_kill(&ctx->refs);
9840 * Drop uring mutex before waiting for references to exit. If
9841 * another thread is currently inside io_uring_enter() it might
9842 * need to grab the uring_lock to make progress. If we hold it
9843 * here across the drain wait, then we can deadlock. It's safe
9844 * to drop the mutex here, since no new references will come in
9845 * after we've killed the percpu ref.
9847 mutex_unlock(&ctx->uring_lock);
9849 ret = wait_for_completion_interruptible(&ctx->ref_comp);
9852 ret = io_run_task_work_sig();
9857 mutex_lock(&ctx->uring_lock);
9860 percpu_ref_resurrect(&ctx->refs);
9865 if (ctx->restricted) {
9866 if (opcode >= IORING_REGISTER_LAST) {
9871 if (!test_bit(opcode, ctx->restrictions.register_op)) {
9878 case IORING_REGISTER_BUFFERS:
9879 ret = io_sqe_buffer_register(ctx, arg, nr_args);
9881 case IORING_UNREGISTER_BUFFERS:
9885 ret = io_sqe_buffer_unregister(ctx);
9887 case IORING_REGISTER_FILES:
9888 ret = io_sqe_files_register(ctx, arg, nr_args);
9890 case IORING_UNREGISTER_FILES:
9894 ret = io_sqe_files_unregister(ctx);
9896 case IORING_REGISTER_FILES_UPDATE:
9897 ret = io_sqe_files_update(ctx, arg, nr_args);
9899 case IORING_REGISTER_EVENTFD:
9900 case IORING_REGISTER_EVENTFD_ASYNC:
9904 ret = io_eventfd_register(ctx, arg);
9907 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
9908 ctx->eventfd_async = 1;
9910 ctx->eventfd_async = 0;
9912 case IORING_UNREGISTER_EVENTFD:
9916 ret = io_eventfd_unregister(ctx);
9918 case IORING_REGISTER_PROBE:
9920 if (!arg || nr_args > 256)
9922 ret = io_probe(ctx, arg, nr_args);
9924 case IORING_REGISTER_PERSONALITY:
9928 ret = io_register_personality(ctx);
9930 case IORING_UNREGISTER_PERSONALITY:
9934 ret = io_unregister_personality(ctx, nr_args);
9936 case IORING_REGISTER_ENABLE_RINGS:
9940 ret = io_register_enable_rings(ctx);
9942 case IORING_REGISTER_RESTRICTIONS:
9943 ret = io_register_restrictions(ctx, arg, nr_args);
9951 if (io_register_op_must_quiesce(opcode)) {
9952 /* bring the ctx back to life */
9953 percpu_ref_reinit(&ctx->refs);
9955 reinit_completion(&ctx->ref_comp);
9960 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
9961 void __user *, arg, unsigned int, nr_args)
9963 struct io_ring_ctx *ctx;
9972 if (f.file->f_op != &io_uring_fops)
9975 ctx = f.file->private_data;
9977 mutex_lock(&ctx->uring_lock);
9978 ret = __io_uring_register(ctx, opcode, arg, nr_args);
9979 mutex_unlock(&ctx->uring_lock);
9980 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
9981 ctx->cq_ev_fd != NULL, ret);
9987 static int __init io_uring_init(void)
9989 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
9990 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
9991 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
9994 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
9995 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
9996 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
9997 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
9998 BUILD_BUG_SQE_ELEM(1, __u8, flags);
9999 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
10000 BUILD_BUG_SQE_ELEM(4, __s32, fd);
10001 BUILD_BUG_SQE_ELEM(8, __u64, off);
10002 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
10003 BUILD_BUG_SQE_ELEM(16, __u64, addr);
10004 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
10005 BUILD_BUG_SQE_ELEM(24, __u32, len);
10006 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
10007 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
10008 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
10009 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
10010 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
10011 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
10012 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
10013 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
10014 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
10015 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
10016 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
10017 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
10018 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
10019 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
10020 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
10021 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
10022 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
10023 BUILD_BUG_SQE_ELEM(42, __u16, personality);
10024 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
10026 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
10027 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
10028 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
10031 __initcall(io_uring_init);