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);
1419 req->work.identity->creds = creds;
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;
1449 if (!(req->work.flags & IO_WQ_WORK_FILES) &&
1450 (def->work_flags & IO_WQ_WORK_FILES) &&
1451 !(req->flags & REQ_F_NO_FILE_TABLE)) {
1452 if (id->files != current->files ||
1453 id->nsproxy != current->nsproxy)
1455 atomic_inc(&id->files->count);
1456 get_nsproxy(id->nsproxy);
1457 req->flags |= REQ_F_INFLIGHT;
1459 spin_lock_irq(&ctx->inflight_lock);
1460 list_add(&req->inflight_entry, &ctx->inflight_list);
1461 spin_unlock_irq(&ctx->inflight_lock);
1462 req->work.flags |= IO_WQ_WORK_FILES;
1464 #ifdef CONFIG_BLK_CGROUP
1465 if (!(req->work.flags & IO_WQ_WORK_BLKCG) &&
1466 (def->work_flags & IO_WQ_WORK_BLKCG)) {
1468 if (id->blkcg_css != blkcg_css()) {
1473 * This should be rare, either the cgroup is dying or the task
1474 * is moving cgroups. Just punt to root for the handful of ios.
1476 if (css_tryget_online(id->blkcg_css))
1477 req->work.flags |= IO_WQ_WORK_BLKCG;
1481 if (!(req->work.flags & IO_WQ_WORK_CREDS)) {
1482 if (id->creds != current_cred())
1484 get_cred(id->creds);
1485 req->work.flags |= IO_WQ_WORK_CREDS;
1488 if (!uid_eq(current->loginuid, id->loginuid) ||
1489 current->sessionid != id->sessionid)
1492 if (!(req->work.flags & IO_WQ_WORK_FS) &&
1493 (def->work_flags & IO_WQ_WORK_FS)) {
1494 if (current->fs != id->fs)
1496 spin_lock(&id->fs->lock);
1497 if (!id->fs->in_exec) {
1499 req->work.flags |= IO_WQ_WORK_FS;
1501 req->work.flags |= IO_WQ_WORK_CANCEL;
1503 spin_unlock(¤t->fs->lock);
1509 static void io_prep_async_work(struct io_kiocb *req)
1511 const struct io_op_def *def = &io_op_defs[req->opcode];
1512 struct io_ring_ctx *ctx = req->ctx;
1513 struct io_identity *id;
1515 io_req_init_async(req);
1516 id = req->work.identity;
1518 if (req->flags & REQ_F_FORCE_ASYNC)
1519 req->work.flags |= IO_WQ_WORK_CONCURRENT;
1521 if (req->flags & REQ_F_ISREG) {
1522 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
1523 io_wq_hash_work(&req->work, file_inode(req->file));
1525 if (def->unbound_nonreg_file)
1526 req->work.flags |= IO_WQ_WORK_UNBOUND;
1529 /* ->mm can never change on us */
1530 if (!(req->work.flags & IO_WQ_WORK_MM) &&
1531 (def->work_flags & IO_WQ_WORK_MM)) {
1533 req->work.flags |= IO_WQ_WORK_MM;
1536 /* if we fail grabbing identity, we must COW, regrab, and retry */
1537 if (io_grab_identity(req))
1540 if (!io_identity_cow(req))
1543 /* can't fail at this point */
1544 if (!io_grab_identity(req))
1548 static void io_prep_async_link(struct io_kiocb *req)
1550 struct io_kiocb *cur;
1552 io_for_each_link(cur, req)
1553 io_prep_async_work(cur);
1556 static struct io_kiocb *__io_queue_async_work(struct io_kiocb *req)
1558 struct io_ring_ctx *ctx = req->ctx;
1559 struct io_kiocb *link = io_prep_linked_timeout(req);
1561 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1562 &req->work, req->flags);
1563 io_wq_enqueue(ctx->io_wq, &req->work);
1567 static void io_queue_async_work(struct io_kiocb *req)
1569 struct io_kiocb *link;
1571 /* init ->work of the whole link before punting */
1572 io_prep_async_link(req);
1573 link = __io_queue_async_work(req);
1576 io_queue_linked_timeout(link);
1579 static void io_kill_timeout(struct io_kiocb *req)
1581 struct io_timeout_data *io = req->async_data;
1584 ret = hrtimer_try_to_cancel(&io->timer);
1586 atomic_set(&req->ctx->cq_timeouts,
1587 atomic_read(&req->ctx->cq_timeouts) + 1);
1588 list_del_init(&req->timeout.list);
1589 io_cqring_fill_event(req, 0);
1590 io_put_req_deferred(req, 1);
1595 * Returns true if we found and killed one or more timeouts
1597 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk,
1598 struct files_struct *files)
1600 struct io_kiocb *req, *tmp;
1603 spin_lock_irq(&ctx->completion_lock);
1604 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
1605 if (io_match_task(req, tsk, files)) {
1606 io_kill_timeout(req);
1610 spin_unlock_irq(&ctx->completion_lock);
1611 return canceled != 0;
1614 static void __io_queue_deferred(struct io_ring_ctx *ctx)
1617 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1618 struct io_defer_entry, list);
1619 struct io_kiocb *link;
1621 if (req_need_defer(de->req, de->seq))
1623 list_del_init(&de->list);
1624 /* punt-init is done before queueing for defer */
1625 link = __io_queue_async_work(de->req);
1627 __io_queue_linked_timeout(link);
1628 /* drop submission reference */
1629 io_put_req_deferred(link, 1);
1632 } while (!list_empty(&ctx->defer_list));
1635 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1637 while (!list_empty(&ctx->timeout_list)) {
1638 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1639 struct io_kiocb, timeout.list);
1641 if (io_is_timeout_noseq(req))
1643 if (req->timeout.target_seq != ctx->cached_cq_tail
1644 - atomic_read(&ctx->cq_timeouts))
1647 list_del_init(&req->timeout.list);
1648 io_kill_timeout(req);
1652 static void io_commit_cqring(struct io_ring_ctx *ctx)
1654 io_flush_timeouts(ctx);
1655 __io_commit_cqring(ctx);
1657 if (unlikely(!list_empty(&ctx->defer_list)))
1658 __io_queue_deferred(ctx);
1661 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1663 struct io_rings *r = ctx->rings;
1665 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == r->sq_ring_entries;
1668 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1670 struct io_rings *rings = ctx->rings;
1673 tail = ctx->cached_cq_tail;
1675 * writes to the cq entry need to come after reading head; the
1676 * control dependency is enough as we're using WRITE_ONCE to
1679 if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries)
1682 ctx->cached_cq_tail++;
1683 return &rings->cqes[tail & ctx->cq_mask];
1686 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1690 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1692 if (!ctx->eventfd_async)
1694 return io_wq_current_is_worker();
1697 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1699 if (waitqueue_active(&ctx->wait))
1700 wake_up(&ctx->wait);
1701 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1702 wake_up(&ctx->sq_data->wait);
1703 if (io_should_trigger_evfd(ctx))
1704 eventfd_signal(ctx->cq_ev_fd, 1);
1707 static void io_cqring_mark_overflow(struct io_ring_ctx *ctx)
1709 if (list_empty(&ctx->cq_overflow_list)) {
1710 clear_bit(0, &ctx->sq_check_overflow);
1711 clear_bit(0, &ctx->cq_check_overflow);
1712 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
1716 /* Returns true if there are no backlogged entries after the flush */
1717 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force,
1718 struct task_struct *tsk,
1719 struct files_struct *files)
1721 struct io_rings *rings = ctx->rings;
1722 struct io_kiocb *req, *tmp;
1723 struct io_uring_cqe *cqe;
1724 unsigned long flags;
1728 if (list_empty_careful(&ctx->cq_overflow_list))
1730 if ((ctx->cached_cq_tail - READ_ONCE(rings->cq.head) ==
1731 rings->cq_ring_entries))
1735 spin_lock_irqsave(&ctx->completion_lock, flags);
1737 /* if force is set, the ring is going away. always drop after that */
1739 ctx->cq_overflow_flushed = 1;
1742 list_for_each_entry_safe(req, tmp, &ctx->cq_overflow_list, compl.list) {
1743 if (!io_match_task(req, tsk, files))
1746 cqe = io_get_cqring(ctx);
1750 list_move(&req->compl.list, &list);
1752 WRITE_ONCE(cqe->user_data, req->user_data);
1753 WRITE_ONCE(cqe->res, req->result);
1754 WRITE_ONCE(cqe->flags, req->compl.cflags);
1756 ctx->cached_cq_overflow++;
1757 WRITE_ONCE(ctx->rings->cq_overflow,
1758 ctx->cached_cq_overflow);
1762 io_commit_cqring(ctx);
1763 io_cqring_mark_overflow(ctx);
1765 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1766 io_cqring_ev_posted(ctx);
1768 while (!list_empty(&list)) {
1769 req = list_first_entry(&list, struct io_kiocb, compl.list);
1770 list_del(&req->compl.list);
1777 static void __io_cqring_fill_event(struct io_kiocb *req, long res, long cflags)
1779 struct io_ring_ctx *ctx = req->ctx;
1780 struct io_uring_cqe *cqe;
1782 trace_io_uring_complete(ctx, req->user_data, res);
1785 * If we can't get a cq entry, userspace overflowed the
1786 * submission (by quite a lot). Increment the overflow count in
1789 cqe = io_get_cqring(ctx);
1791 WRITE_ONCE(cqe->user_data, req->user_data);
1792 WRITE_ONCE(cqe->res, res);
1793 WRITE_ONCE(cqe->flags, cflags);
1794 } else if (ctx->cq_overflow_flushed ||
1795 atomic_read(&req->task->io_uring->in_idle)) {
1797 * If we're in ring overflow flush mode, or in task cancel mode,
1798 * then we cannot store the request for later flushing, we need
1799 * to drop it on the floor.
1801 ctx->cached_cq_overflow++;
1802 WRITE_ONCE(ctx->rings->cq_overflow, ctx->cached_cq_overflow);
1804 if (list_empty(&ctx->cq_overflow_list)) {
1805 set_bit(0, &ctx->sq_check_overflow);
1806 set_bit(0, &ctx->cq_check_overflow);
1807 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
1811 req->compl.cflags = cflags;
1812 refcount_inc(&req->refs);
1813 list_add_tail(&req->compl.list, &ctx->cq_overflow_list);
1817 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1819 __io_cqring_fill_event(req, res, 0);
1822 static void io_cqring_add_event(struct io_kiocb *req, long res, long cflags)
1824 struct io_ring_ctx *ctx = req->ctx;
1825 unsigned long flags;
1827 spin_lock_irqsave(&ctx->completion_lock, flags);
1828 __io_cqring_fill_event(req, res, cflags);
1829 io_commit_cqring(ctx);
1830 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1832 io_cqring_ev_posted(ctx);
1835 static void io_submit_flush_completions(struct io_comp_state *cs)
1837 struct io_ring_ctx *ctx = cs->ctx;
1839 spin_lock_irq(&ctx->completion_lock);
1840 while (!list_empty(&cs->list)) {
1841 struct io_kiocb *req;
1843 req = list_first_entry(&cs->list, struct io_kiocb, compl.list);
1844 list_del(&req->compl.list);
1845 __io_cqring_fill_event(req, req->result, req->compl.cflags);
1848 * io_free_req() doesn't care about completion_lock unless one
1849 * of these flags is set. REQ_F_WORK_INITIALIZED is in the list
1850 * because of a potential deadlock with req->work.fs->lock
1852 if (req->flags & (REQ_F_FAIL_LINK|REQ_F_LINK_TIMEOUT
1853 |REQ_F_WORK_INITIALIZED)) {
1854 spin_unlock_irq(&ctx->completion_lock);
1856 spin_lock_irq(&ctx->completion_lock);
1861 io_commit_cqring(ctx);
1862 spin_unlock_irq(&ctx->completion_lock);
1864 io_cqring_ev_posted(ctx);
1868 static void __io_req_complete(struct io_kiocb *req, long res, unsigned cflags,
1869 struct io_comp_state *cs)
1872 io_cqring_add_event(req, res, cflags);
1877 req->compl.cflags = cflags;
1878 list_add_tail(&req->compl.list, &cs->list);
1880 io_submit_flush_completions(cs);
1884 static void io_req_complete(struct io_kiocb *req, long res)
1886 __io_req_complete(req, res, 0, NULL);
1889 static inline bool io_is_fallback_req(struct io_kiocb *req)
1891 return req == (struct io_kiocb *)
1892 ((unsigned long) req->ctx->fallback_req & ~1UL);
1895 static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx)
1897 struct io_kiocb *req;
1899 req = ctx->fallback_req;
1900 if (!test_and_set_bit_lock(0, (unsigned long *) &ctx->fallback_req))
1906 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx,
1907 struct io_submit_state *state)
1909 if (!state->free_reqs) {
1910 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1914 sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs));
1915 ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs);
1918 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1919 * retry single alloc to be on the safe side.
1921 if (unlikely(ret <= 0)) {
1922 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1923 if (!state->reqs[0])
1927 state->free_reqs = ret;
1931 return state->reqs[state->free_reqs];
1933 return io_get_fallback_req(ctx);
1936 static inline void io_put_file(struct io_kiocb *req, struct file *file,
1943 static void io_dismantle_req(struct io_kiocb *req)
1947 if (req->async_data)
1948 kfree(req->async_data);
1950 io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
1951 if (req->fixed_file_refs)
1952 percpu_ref_put(req->fixed_file_refs);
1953 io_req_clean_work(req);
1956 static void __io_free_req(struct io_kiocb *req)
1958 struct io_uring_task *tctx = req->task->io_uring;
1959 struct io_ring_ctx *ctx = req->ctx;
1961 io_dismantle_req(req);
1963 percpu_counter_dec(&tctx->inflight);
1964 if (atomic_read(&tctx->in_idle))
1965 wake_up(&tctx->wait);
1966 put_task_struct(req->task);
1968 if (likely(!io_is_fallback_req(req)))
1969 kmem_cache_free(req_cachep, req);
1971 clear_bit_unlock(0, (unsigned long *) &ctx->fallback_req);
1972 percpu_ref_put(&ctx->refs);
1975 static inline void io_remove_next_linked(struct io_kiocb *req)
1977 struct io_kiocb *nxt = req->link;
1979 req->link = nxt->link;
1983 static void io_kill_linked_timeout(struct io_kiocb *req)
1985 struct io_ring_ctx *ctx = req->ctx;
1986 struct io_kiocb *link;
1987 bool cancelled = false;
1988 unsigned long flags;
1990 spin_lock_irqsave(&ctx->completion_lock, flags);
1994 * Can happen if a linked timeout fired and link had been like
1995 * req -> link t-out -> link t-out [-> ...]
1997 if (link && (link->flags & REQ_F_LTIMEOUT_ACTIVE)) {
1998 struct io_timeout_data *io = link->async_data;
2001 io_remove_next_linked(req);
2002 link->timeout.head = NULL;
2003 ret = hrtimer_try_to_cancel(&io->timer);
2005 io_cqring_fill_event(link, -ECANCELED);
2006 io_commit_cqring(ctx);
2010 req->flags &= ~REQ_F_LINK_TIMEOUT;
2011 spin_unlock_irqrestore(&ctx->completion_lock, flags);
2014 io_cqring_ev_posted(ctx);
2020 static void io_fail_links(struct io_kiocb *req)
2022 struct io_kiocb *link, *nxt;
2023 struct io_ring_ctx *ctx = req->ctx;
2024 unsigned long flags;
2026 spin_lock_irqsave(&ctx->completion_lock, flags);
2034 trace_io_uring_fail_link(req, link);
2035 io_cqring_fill_event(link, -ECANCELED);
2038 * It's ok to free under spinlock as they're not linked anymore,
2039 * but avoid REQ_F_WORK_INITIALIZED because it may deadlock on
2042 if (link->flags & REQ_F_WORK_INITIALIZED)
2043 io_put_req_deferred(link, 2);
2045 io_double_put_req(link);
2048 io_commit_cqring(ctx);
2049 spin_unlock_irqrestore(&ctx->completion_lock, flags);
2051 io_cqring_ev_posted(ctx);
2054 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
2056 if (req->flags & REQ_F_LINK_TIMEOUT)
2057 io_kill_linked_timeout(req);
2060 * If LINK is set, we have dependent requests in this chain. If we
2061 * didn't fail this request, queue the first one up, moving any other
2062 * dependencies to the next request. In case of failure, fail the rest
2065 if (likely(!(req->flags & REQ_F_FAIL_LINK))) {
2066 struct io_kiocb *nxt = req->link;
2075 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
2077 if (likely(!(req->link) && !(req->flags & REQ_F_LINK_TIMEOUT)))
2079 return __io_req_find_next(req);
2082 static int io_req_task_work_add(struct io_kiocb *req, bool twa_signal_ok)
2084 struct task_struct *tsk = req->task;
2085 struct io_ring_ctx *ctx = req->ctx;
2086 enum task_work_notify_mode notify;
2089 if (tsk->flags & PF_EXITING)
2093 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
2094 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
2095 * processing task_work. There's no reliable way to tell if TWA_RESUME
2099 if (!(ctx->flags & IORING_SETUP_SQPOLL) && twa_signal_ok)
2100 notify = TWA_SIGNAL;
2102 ret = task_work_add(tsk, &req->task_work, notify);
2104 wake_up_process(tsk);
2109 static void __io_req_task_cancel(struct io_kiocb *req, int error)
2111 struct io_ring_ctx *ctx = req->ctx;
2113 spin_lock_irq(&ctx->completion_lock);
2114 io_cqring_fill_event(req, error);
2115 io_commit_cqring(ctx);
2116 spin_unlock_irq(&ctx->completion_lock);
2118 io_cqring_ev_posted(ctx);
2119 req_set_fail_links(req);
2120 io_double_put_req(req);
2123 static void io_req_task_cancel(struct callback_head *cb)
2125 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2126 struct io_ring_ctx *ctx = req->ctx;
2128 __io_req_task_cancel(req, -ECANCELED);
2129 percpu_ref_put(&ctx->refs);
2132 static void __io_req_task_submit(struct io_kiocb *req)
2134 struct io_ring_ctx *ctx = req->ctx;
2136 if (!__io_sq_thread_acquire_mm(ctx) &&
2137 !__io_sq_thread_acquire_files(ctx)) {
2138 mutex_lock(&ctx->uring_lock);
2139 __io_queue_sqe(req, NULL);
2140 mutex_unlock(&ctx->uring_lock);
2142 __io_req_task_cancel(req, -EFAULT);
2146 static void io_req_task_submit(struct callback_head *cb)
2148 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2149 struct io_ring_ctx *ctx = req->ctx;
2151 __io_req_task_submit(req);
2152 percpu_ref_put(&ctx->refs);
2155 static void io_req_task_queue(struct io_kiocb *req)
2159 init_task_work(&req->task_work, io_req_task_submit);
2160 percpu_ref_get(&req->ctx->refs);
2162 ret = io_req_task_work_add(req, true);
2163 if (unlikely(ret)) {
2164 struct task_struct *tsk;
2166 init_task_work(&req->task_work, io_req_task_cancel);
2167 tsk = io_wq_get_task(req->ctx->io_wq);
2168 task_work_add(tsk, &req->task_work, TWA_NONE);
2169 wake_up_process(tsk);
2173 static inline void io_queue_next(struct io_kiocb *req)
2175 struct io_kiocb *nxt = io_req_find_next(req);
2178 io_req_task_queue(nxt);
2181 static void io_free_req(struct io_kiocb *req)
2188 void *reqs[IO_IOPOLL_BATCH];
2191 struct task_struct *task;
2195 static inline void io_init_req_batch(struct req_batch *rb)
2202 static void __io_req_free_batch_flush(struct io_ring_ctx *ctx,
2203 struct req_batch *rb)
2205 kmem_cache_free_bulk(req_cachep, rb->to_free, rb->reqs);
2206 percpu_ref_put_many(&ctx->refs, rb->to_free);
2210 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2211 struct req_batch *rb)
2214 __io_req_free_batch_flush(ctx, rb);
2216 struct io_uring_task *tctx = rb->task->io_uring;
2218 percpu_counter_sub(&tctx->inflight, rb->task_refs);
2219 put_task_struct_many(rb->task, rb->task_refs);
2224 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req)
2226 if (unlikely(io_is_fallback_req(req))) {
2232 if (req->task != rb->task) {
2234 struct io_uring_task *tctx = rb->task->io_uring;
2236 percpu_counter_sub(&tctx->inflight, rb->task_refs);
2237 put_task_struct_many(rb->task, rb->task_refs);
2239 rb->task = req->task;
2244 io_dismantle_req(req);
2245 rb->reqs[rb->to_free++] = req;
2246 if (unlikely(rb->to_free == ARRAY_SIZE(rb->reqs)))
2247 __io_req_free_batch_flush(req->ctx, rb);
2251 * Drop reference to request, return next in chain (if there is one) if this
2252 * was the last reference to this request.
2254 static struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2256 struct io_kiocb *nxt = NULL;
2258 if (refcount_dec_and_test(&req->refs)) {
2259 nxt = io_req_find_next(req);
2265 static void io_put_req(struct io_kiocb *req)
2267 if (refcount_dec_and_test(&req->refs))
2271 static void io_put_req_deferred_cb(struct callback_head *cb)
2273 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
2278 static void io_free_req_deferred(struct io_kiocb *req)
2282 init_task_work(&req->task_work, io_put_req_deferred_cb);
2283 ret = io_req_task_work_add(req, true);
2284 if (unlikely(ret)) {
2285 struct task_struct *tsk;
2287 tsk = io_wq_get_task(req->ctx->io_wq);
2288 task_work_add(tsk, &req->task_work, TWA_NONE);
2289 wake_up_process(tsk);
2293 static inline void io_put_req_deferred(struct io_kiocb *req, int refs)
2295 if (refcount_sub_and_test(refs, &req->refs))
2296 io_free_req_deferred(req);
2299 static struct io_wq_work *io_steal_work(struct io_kiocb *req)
2301 struct io_kiocb *nxt;
2304 * A ref is owned by io-wq in which context we're. So, if that's the
2305 * last one, it's safe to steal next work. False negatives are Ok,
2306 * it just will be re-punted async in io_put_work()
2308 if (refcount_read(&req->refs) != 1)
2311 nxt = io_req_find_next(req);
2312 return nxt ? &nxt->work : NULL;
2315 static void io_double_put_req(struct io_kiocb *req)
2317 /* drop both submit and complete references */
2318 if (refcount_sub_and_test(2, &req->refs))
2322 static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush)
2324 struct io_rings *rings = ctx->rings;
2326 if (test_bit(0, &ctx->cq_check_overflow)) {
2328 * noflush == true is from the waitqueue handler, just ensure
2329 * we wake up the task, and the next invocation will flush the
2330 * entries. We cannot safely to it from here.
2332 if (noflush && !list_empty(&ctx->cq_overflow_list))
2335 io_cqring_overflow_flush(ctx, false, NULL, NULL);
2338 /* See comment at the top of this file */
2340 return ctx->cached_cq_tail - READ_ONCE(rings->cq.head);
2343 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2345 struct io_rings *rings = ctx->rings;
2347 /* make sure SQ entry isn't read before tail */
2348 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2351 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2353 unsigned int cflags;
2355 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2356 cflags |= IORING_CQE_F_BUFFER;
2357 req->flags &= ~REQ_F_BUFFER_SELECTED;
2362 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2364 struct io_buffer *kbuf;
2366 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2367 return io_put_kbuf(req, kbuf);
2370 static inline bool io_run_task_work(void)
2373 * Not safe to run on exiting task, and the task_work handling will
2374 * not add work to such a task.
2376 if (unlikely(current->flags & PF_EXITING))
2378 if (current->task_works) {
2379 __set_current_state(TASK_RUNNING);
2387 static void io_iopoll_queue(struct list_head *again)
2389 struct io_kiocb *req;
2392 req = list_first_entry(again, struct io_kiocb, inflight_entry);
2393 list_del(&req->inflight_entry);
2394 __io_complete_rw(req, -EAGAIN, 0, NULL);
2395 } while (!list_empty(again));
2399 * Find and free completed poll iocbs
2401 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2402 struct list_head *done)
2404 struct req_batch rb;
2405 struct io_kiocb *req;
2408 /* order with ->result store in io_complete_rw_iopoll() */
2411 io_init_req_batch(&rb);
2412 while (!list_empty(done)) {
2415 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2416 if (READ_ONCE(req->result) == -EAGAIN) {
2418 req->iopoll_completed = 0;
2419 list_move_tail(&req->inflight_entry, &again);
2422 list_del(&req->inflight_entry);
2424 if (req->flags & REQ_F_BUFFER_SELECTED)
2425 cflags = io_put_rw_kbuf(req);
2427 __io_cqring_fill_event(req, req->result, cflags);
2430 if (refcount_dec_and_test(&req->refs))
2431 io_req_free_batch(&rb, req);
2434 io_commit_cqring(ctx);
2435 if (ctx->flags & IORING_SETUP_SQPOLL)
2436 io_cqring_ev_posted(ctx);
2437 io_req_free_batch_finish(ctx, &rb);
2439 if (!list_empty(&again))
2440 io_iopoll_queue(&again);
2443 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2446 struct io_kiocb *req, *tmp;
2452 * Only spin for completions if we don't have multiple devices hanging
2453 * off our complete list, and we're under the requested amount.
2455 spin = !ctx->poll_multi_file && *nr_events < min;
2458 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2459 struct kiocb *kiocb = &req->rw.kiocb;
2462 * Move completed and retryable entries to our local lists.
2463 * If we find a request that requires polling, break out
2464 * and complete those lists first, if we have entries there.
2466 if (READ_ONCE(req->iopoll_completed)) {
2467 list_move_tail(&req->inflight_entry, &done);
2470 if (!list_empty(&done))
2473 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2477 /* iopoll may have completed current req */
2478 if (READ_ONCE(req->iopoll_completed))
2479 list_move_tail(&req->inflight_entry, &done);
2486 if (!list_empty(&done))
2487 io_iopoll_complete(ctx, nr_events, &done);
2493 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2494 * non-spinning poll check - we'll still enter the driver poll loop, but only
2495 * as a non-spinning completion check.
2497 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
2500 while (!list_empty(&ctx->iopoll_list) && !need_resched()) {
2503 ret = io_do_iopoll(ctx, nr_events, min);
2506 if (*nr_events >= min)
2514 * We can't just wait for polled events to come to us, we have to actively
2515 * find and complete them.
2517 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2519 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2522 mutex_lock(&ctx->uring_lock);
2523 while (!list_empty(&ctx->iopoll_list)) {
2524 unsigned int nr_events = 0;
2526 io_do_iopoll(ctx, &nr_events, 0);
2528 /* let it sleep and repeat later if can't complete a request */
2532 * Ensure we allow local-to-the-cpu processing to take place,
2533 * in this case we need to ensure that we reap all events.
2534 * Also let task_work, etc. to progress by releasing the mutex
2536 if (need_resched()) {
2537 mutex_unlock(&ctx->uring_lock);
2539 mutex_lock(&ctx->uring_lock);
2542 mutex_unlock(&ctx->uring_lock);
2545 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2547 unsigned int nr_events = 0;
2548 int iters = 0, ret = 0;
2551 * We disallow the app entering submit/complete with polling, but we
2552 * still need to lock the ring to prevent racing with polled issue
2553 * that got punted to a workqueue.
2555 mutex_lock(&ctx->uring_lock);
2558 * Don't enter poll loop if we already have events pending.
2559 * If we do, we can potentially be spinning for commands that
2560 * already triggered a CQE (eg in error).
2562 if (io_cqring_events(ctx, false))
2566 * If a submit got punted to a workqueue, we can have the
2567 * application entering polling for a command before it gets
2568 * issued. That app will hold the uring_lock for the duration
2569 * of the poll right here, so we need to take a breather every
2570 * now and then to ensure that the issue has a chance to add
2571 * the poll to the issued list. Otherwise we can spin here
2572 * forever, while the workqueue is stuck trying to acquire the
2575 if (!(++iters & 7)) {
2576 mutex_unlock(&ctx->uring_lock);
2578 mutex_lock(&ctx->uring_lock);
2581 ret = io_iopoll_getevents(ctx, &nr_events, min);
2585 } while (min && !nr_events && !need_resched());
2587 mutex_unlock(&ctx->uring_lock);
2591 static void kiocb_end_write(struct io_kiocb *req)
2594 * Tell lockdep we inherited freeze protection from submission
2597 if (req->flags & REQ_F_ISREG) {
2598 struct inode *inode = file_inode(req->file);
2600 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
2602 file_end_write(req->file);
2605 static void io_complete_rw_common(struct kiocb *kiocb, long res,
2606 struct io_comp_state *cs)
2608 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2611 if (kiocb->ki_flags & IOCB_WRITE)
2612 kiocb_end_write(req);
2614 if (res != req->result)
2615 req_set_fail_links(req);
2616 if (req->flags & REQ_F_BUFFER_SELECTED)
2617 cflags = io_put_rw_kbuf(req);
2618 __io_req_complete(req, res, cflags, cs);
2622 static bool io_resubmit_prep(struct io_kiocb *req, int error)
2624 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2625 ssize_t ret = -ECANCELED;
2626 struct iov_iter iter;
2634 switch (req->opcode) {
2635 case IORING_OP_READV:
2636 case IORING_OP_READ_FIXED:
2637 case IORING_OP_READ:
2640 case IORING_OP_WRITEV:
2641 case IORING_OP_WRITE_FIXED:
2642 case IORING_OP_WRITE:
2646 printk_once(KERN_WARNING "io_uring: bad opcode in resubmit %d\n",
2651 if (!req->async_data) {
2652 ret = io_import_iovec(rw, req, &iovec, &iter, false);
2655 ret = io_setup_async_rw(req, iovec, inline_vecs, &iter, false);
2663 req_set_fail_links(req);
2668 static bool io_rw_reissue(struct io_kiocb *req, long res)
2671 umode_t mode = file_inode(req->file)->i_mode;
2674 if (!S_ISBLK(mode) && !S_ISREG(mode))
2676 if ((res != -EAGAIN && res != -EOPNOTSUPP) || io_wq_current_is_worker())
2679 ret = io_sq_thread_acquire_mm_files(req->ctx, req);
2681 if (io_resubmit_prep(req, ret)) {
2682 refcount_inc(&req->refs);
2683 io_queue_async_work(req);
2691 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2692 struct io_comp_state *cs)
2694 if (!io_rw_reissue(req, res))
2695 io_complete_rw_common(&req->rw.kiocb, res, cs);
2698 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2700 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2702 __io_complete_rw(req, res, res2, NULL);
2705 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2707 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2709 if (kiocb->ki_flags & IOCB_WRITE)
2710 kiocb_end_write(req);
2712 if (res != -EAGAIN && res != req->result)
2713 req_set_fail_links(req);
2715 WRITE_ONCE(req->result, res);
2716 /* order with io_poll_complete() checking ->result */
2718 WRITE_ONCE(req->iopoll_completed, 1);
2722 * After the iocb has been issued, it's safe to be found on the poll list.
2723 * Adding the kiocb to the list AFTER submission ensures that we don't
2724 * find it from a io_iopoll_getevents() thread before the issuer is done
2725 * accessing the kiocb cookie.
2727 static void io_iopoll_req_issued(struct io_kiocb *req, bool in_async)
2729 struct io_ring_ctx *ctx = req->ctx;
2732 * Track whether we have multiple files in our lists. This will impact
2733 * how we do polling eventually, not spinning if we're on potentially
2734 * different devices.
2736 if (list_empty(&ctx->iopoll_list)) {
2737 ctx->poll_multi_file = false;
2738 } else if (!ctx->poll_multi_file) {
2739 struct io_kiocb *list_req;
2741 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2743 if (list_req->file != req->file)
2744 ctx->poll_multi_file = true;
2748 * For fast devices, IO may have already completed. If it has, add
2749 * it to the front so we find it first.
2751 if (READ_ONCE(req->iopoll_completed))
2752 list_add(&req->inflight_entry, &ctx->iopoll_list);
2754 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2757 * If IORING_SETUP_SQPOLL is enabled, sqes are either handled in sq thread
2758 * task context or in io worker task context. If current task context is
2759 * sq thread, we don't need to check whether should wake up sq thread.
2761 if (in_async && (ctx->flags & IORING_SETUP_SQPOLL) &&
2762 wq_has_sleeper(&ctx->sq_data->wait))
2763 wake_up(&ctx->sq_data->wait);
2766 static inline void __io_state_file_put(struct io_submit_state *state)
2768 fput_many(state->file, state->file_refs);
2769 state->file_refs = 0;
2772 static inline void io_state_file_put(struct io_submit_state *state)
2774 if (state->file_refs)
2775 __io_state_file_put(state);
2779 * Get as many references to a file as we have IOs left in this submission,
2780 * assuming most submissions are for one file, or at least that each file
2781 * has more than one submission.
2783 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2788 if (state->file_refs) {
2789 if (state->fd == fd) {
2793 __io_state_file_put(state);
2795 state->file = fget_many(fd, state->ios_left);
2796 if (unlikely(!state->file))
2800 state->file_refs = state->ios_left - 1;
2804 static bool io_bdev_nowait(struct block_device *bdev)
2807 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
2814 * If we tracked the file through the SCM inflight mechanism, we could support
2815 * any file. For now, just ensure that anything potentially problematic is done
2818 static bool io_file_supports_async(struct file *file, int rw)
2820 umode_t mode = file_inode(file)->i_mode;
2822 if (S_ISBLK(mode)) {
2823 if (io_bdev_nowait(file->f_inode->i_bdev))
2827 if (S_ISCHR(mode) || S_ISSOCK(mode))
2829 if (S_ISREG(mode)) {
2830 if (io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2831 file->f_op != &io_uring_fops)
2836 /* any ->read/write should understand O_NONBLOCK */
2837 if (file->f_flags & O_NONBLOCK)
2840 if (!(file->f_mode & FMODE_NOWAIT))
2844 return file->f_op->read_iter != NULL;
2846 return file->f_op->write_iter != NULL;
2849 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2851 struct io_ring_ctx *ctx = req->ctx;
2852 struct kiocb *kiocb = &req->rw.kiocb;
2856 if (S_ISREG(file_inode(req->file)->i_mode))
2857 req->flags |= REQ_F_ISREG;
2859 kiocb->ki_pos = READ_ONCE(sqe->off);
2860 if (kiocb->ki_pos == -1 && !(req->file->f_mode & FMODE_STREAM)) {
2861 req->flags |= REQ_F_CUR_POS;
2862 kiocb->ki_pos = req->file->f_pos;
2864 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2865 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2866 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2870 ioprio = READ_ONCE(sqe->ioprio);
2872 ret = ioprio_check_cap(ioprio);
2876 kiocb->ki_ioprio = ioprio;
2878 kiocb->ki_ioprio = get_current_ioprio();
2880 /* don't allow async punt if RWF_NOWAIT was requested */
2881 if (kiocb->ki_flags & IOCB_NOWAIT)
2882 req->flags |= REQ_F_NOWAIT;
2884 if (ctx->flags & IORING_SETUP_IOPOLL) {
2885 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2886 !kiocb->ki_filp->f_op->iopoll)
2889 kiocb->ki_flags |= IOCB_HIPRI;
2890 kiocb->ki_complete = io_complete_rw_iopoll;
2891 req->iopoll_completed = 0;
2893 if (kiocb->ki_flags & IOCB_HIPRI)
2895 kiocb->ki_complete = io_complete_rw;
2898 req->rw.addr = READ_ONCE(sqe->addr);
2899 req->rw.len = READ_ONCE(sqe->len);
2900 req->buf_index = READ_ONCE(sqe->buf_index);
2904 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2910 case -ERESTARTNOINTR:
2911 case -ERESTARTNOHAND:
2912 case -ERESTART_RESTARTBLOCK:
2914 * We can't just restart the syscall, since previously
2915 * submitted sqes may already be in progress. Just fail this
2921 kiocb->ki_complete(kiocb, ret, 0);
2925 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2926 struct io_comp_state *cs)
2928 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2929 struct io_async_rw *io = req->async_data;
2931 /* add previously done IO, if any */
2932 if (io && io->bytes_done > 0) {
2934 ret = io->bytes_done;
2936 ret += io->bytes_done;
2939 if (req->flags & REQ_F_CUR_POS)
2940 req->file->f_pos = kiocb->ki_pos;
2941 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2942 __io_complete_rw(req, ret, 0, cs);
2944 io_rw_done(kiocb, ret);
2947 static ssize_t io_import_fixed(struct io_kiocb *req, int rw,
2948 struct iov_iter *iter)
2950 struct io_ring_ctx *ctx = req->ctx;
2951 size_t len = req->rw.len;
2952 struct io_mapped_ubuf *imu;
2953 u16 index, buf_index = req->buf_index;
2957 if (unlikely(buf_index >= ctx->nr_user_bufs))
2959 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2960 imu = &ctx->user_bufs[index];
2961 buf_addr = req->rw.addr;
2964 if (buf_addr + len < buf_addr)
2966 /* not inside the mapped region */
2967 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2971 * May not be a start of buffer, set size appropriately
2972 * and advance us to the beginning.
2974 offset = buf_addr - imu->ubuf;
2975 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2979 * Don't use iov_iter_advance() here, as it's really slow for
2980 * using the latter parts of a big fixed buffer - it iterates
2981 * over each segment manually. We can cheat a bit here, because
2984 * 1) it's a BVEC iter, we set it up
2985 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2986 * first and last bvec
2988 * So just find our index, and adjust the iterator afterwards.
2989 * If the offset is within the first bvec (or the whole first
2990 * bvec, just use iov_iter_advance(). This makes it easier
2991 * since we can just skip the first segment, which may not
2992 * be PAGE_SIZE aligned.
2994 const struct bio_vec *bvec = imu->bvec;
2996 if (offset <= bvec->bv_len) {
2997 iov_iter_advance(iter, offset);
2999 unsigned long seg_skip;
3001 /* skip first vec */
3002 offset -= bvec->bv_len;
3003 seg_skip = 1 + (offset >> PAGE_SHIFT);
3005 iter->bvec = bvec + seg_skip;
3006 iter->nr_segs -= seg_skip;
3007 iter->count -= bvec->bv_len + offset;
3008 iter->iov_offset = offset & ~PAGE_MASK;
3015 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
3018 mutex_unlock(&ctx->uring_lock);
3021 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
3024 * "Normal" inline submissions always hold the uring_lock, since we
3025 * grab it from the system call. Same is true for the SQPOLL offload.
3026 * The only exception is when we've detached the request and issue it
3027 * from an async worker thread, grab the lock for that case.
3030 mutex_lock(&ctx->uring_lock);
3033 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
3034 int bgid, struct io_buffer *kbuf,
3037 struct io_buffer *head;
3039 if (req->flags & REQ_F_BUFFER_SELECTED)
3042 io_ring_submit_lock(req->ctx, needs_lock);
3044 lockdep_assert_held(&req->ctx->uring_lock);
3046 head = idr_find(&req->ctx->io_buffer_idr, bgid);
3048 if (!list_empty(&head->list)) {
3049 kbuf = list_last_entry(&head->list, struct io_buffer,
3051 list_del(&kbuf->list);
3054 idr_remove(&req->ctx->io_buffer_idr, bgid);
3056 if (*len > kbuf->len)
3059 kbuf = ERR_PTR(-ENOBUFS);
3062 io_ring_submit_unlock(req->ctx, needs_lock);
3067 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
3070 struct io_buffer *kbuf;
3073 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3074 bgid = req->buf_index;
3075 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
3078 req->rw.addr = (u64) (unsigned long) kbuf;
3079 req->flags |= REQ_F_BUFFER_SELECTED;
3080 return u64_to_user_ptr(kbuf->addr);
3083 #ifdef CONFIG_COMPAT
3084 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
3087 struct compat_iovec __user *uiov;
3088 compat_ssize_t clen;
3092 uiov = u64_to_user_ptr(req->rw.addr);
3093 if (!access_ok(uiov, sizeof(*uiov)))
3095 if (__get_user(clen, &uiov->iov_len))
3101 buf = io_rw_buffer_select(req, &len, needs_lock);
3103 return PTR_ERR(buf);
3104 iov[0].iov_base = buf;
3105 iov[0].iov_len = (compat_size_t) len;
3110 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3113 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
3117 if (copy_from_user(iov, uiov, sizeof(*uiov)))
3120 len = iov[0].iov_len;
3123 buf = io_rw_buffer_select(req, &len, needs_lock);
3125 return PTR_ERR(buf);
3126 iov[0].iov_base = buf;
3127 iov[0].iov_len = len;
3131 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
3134 if (req->flags & REQ_F_BUFFER_SELECTED) {
3135 struct io_buffer *kbuf;
3137 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
3138 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
3139 iov[0].iov_len = kbuf->len;
3144 else if (req->rw.len > 1)
3147 #ifdef CONFIG_COMPAT
3148 if (req->ctx->compat)
3149 return io_compat_import(req, iov, needs_lock);
3152 return __io_iov_buffer_select(req, iov, needs_lock);
3155 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
3156 struct iovec **iovec, struct iov_iter *iter,
3159 void __user *buf = u64_to_user_ptr(req->rw.addr);
3160 size_t sqe_len = req->rw.len;
3164 opcode = req->opcode;
3165 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
3167 return io_import_fixed(req, rw, iter);
3170 /* buffer index only valid with fixed read/write, or buffer select */
3171 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
3174 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
3175 if (req->flags & REQ_F_BUFFER_SELECT) {
3176 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
3178 return PTR_ERR(buf);
3179 req->rw.len = sqe_len;
3182 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
3187 if (req->flags & REQ_F_BUFFER_SELECT) {
3188 ret = io_iov_buffer_select(req, *iovec, needs_lock);
3190 ret = (*iovec)->iov_len;
3191 iov_iter_init(iter, rw, *iovec, 1, ret);
3197 return __import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter,
3201 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3203 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3207 * For files that don't have ->read_iter() and ->write_iter(), handle them
3208 * by looping over ->read() or ->write() manually.
3210 static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)
3212 struct kiocb *kiocb = &req->rw.kiocb;
3213 struct file *file = req->file;
3217 * Don't support polled IO through this interface, and we can't
3218 * support non-blocking either. For the latter, this just causes
3219 * the kiocb to be handled from an async context.
3221 if (kiocb->ki_flags & IOCB_HIPRI)
3223 if (kiocb->ki_flags & IOCB_NOWAIT)
3226 while (iov_iter_count(iter)) {
3230 if (!iov_iter_is_bvec(iter)) {
3231 iovec = iov_iter_iovec(iter);
3233 iovec.iov_base = u64_to_user_ptr(req->rw.addr);
3234 iovec.iov_len = req->rw.len;
3238 nr = file->f_op->read(file, iovec.iov_base,
3239 iovec.iov_len, io_kiocb_ppos(kiocb));
3241 nr = file->f_op->write(file, iovec.iov_base,
3242 iovec.iov_len, io_kiocb_ppos(kiocb));
3251 if (nr != iovec.iov_len)
3255 iov_iter_advance(iter, nr);
3261 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3262 const struct iovec *fast_iov, struct iov_iter *iter)
3264 struct io_async_rw *rw = req->async_data;
3266 memcpy(&rw->iter, iter, sizeof(*iter));
3267 rw->free_iovec = iovec;
3269 /* can only be fixed buffers, no need to do anything */
3270 if (iter->type == ITER_BVEC)
3273 unsigned iov_off = 0;
3275 rw->iter.iov = rw->fast_iov;
3276 if (iter->iov != fast_iov) {
3277 iov_off = iter->iov - fast_iov;
3278 rw->iter.iov += iov_off;
3280 if (rw->fast_iov != fast_iov)
3281 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3282 sizeof(struct iovec) * iter->nr_segs);
3284 req->flags |= REQ_F_NEED_CLEANUP;
3288 static inline int __io_alloc_async_data(struct io_kiocb *req)
3290 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3291 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3292 return req->async_data == NULL;
3295 static int io_alloc_async_data(struct io_kiocb *req)
3297 if (!io_op_defs[req->opcode].needs_async_data)
3300 return __io_alloc_async_data(req);
3303 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3304 const struct iovec *fast_iov,
3305 struct iov_iter *iter, bool force)
3307 if (!force && !io_op_defs[req->opcode].needs_async_data)
3309 if (!req->async_data) {
3310 if (__io_alloc_async_data(req))
3313 io_req_map_rw(req, iovec, fast_iov, iter);
3318 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3320 struct io_async_rw *iorw = req->async_data;
3321 struct iovec *iov = iorw->fast_iov;
3324 ret = io_import_iovec(rw, req, &iov, &iorw->iter, false);
3325 if (unlikely(ret < 0))
3328 iorw->bytes_done = 0;
3329 iorw->free_iovec = iov;
3331 req->flags |= REQ_F_NEED_CLEANUP;
3335 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3339 ret = io_prep_rw(req, sqe);
3343 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3346 /* either don't need iovec imported or already have it */
3347 if (!req->async_data)
3349 return io_rw_prep_async(req, READ);
3353 * This is our waitqueue callback handler, registered through lock_page_async()
3354 * when we initially tried to do the IO with the iocb armed our waitqueue.
3355 * This gets called when the page is unlocked, and we generally expect that to
3356 * happen when the page IO is completed and the page is now uptodate. This will
3357 * queue a task_work based retry of the operation, attempting to copy the data
3358 * again. If the latter fails because the page was NOT uptodate, then we will
3359 * do a thread based blocking retry of the operation. That's the unexpected
3362 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3363 int sync, void *arg)
3365 struct wait_page_queue *wpq;
3366 struct io_kiocb *req = wait->private;
3367 struct wait_page_key *key = arg;
3370 wpq = container_of(wait, struct wait_page_queue, wait);
3372 if (!wake_page_match(wpq, key))
3375 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3376 list_del_init(&wait->entry);
3378 init_task_work(&req->task_work, io_req_task_submit);
3379 percpu_ref_get(&req->ctx->refs);
3381 /* submit ref gets dropped, acquire a new one */
3382 refcount_inc(&req->refs);
3383 ret = io_req_task_work_add(req, true);
3384 if (unlikely(ret)) {
3385 struct task_struct *tsk;
3387 /* queue just for cancelation */
3388 init_task_work(&req->task_work, io_req_task_cancel);
3389 tsk = io_wq_get_task(req->ctx->io_wq);
3390 task_work_add(tsk, &req->task_work, TWA_NONE);
3391 wake_up_process(tsk);
3397 * This controls whether a given IO request should be armed for async page
3398 * based retry. If we return false here, the request is handed to the async
3399 * worker threads for retry. If we're doing buffered reads on a regular file,
3400 * we prepare a private wait_page_queue entry and retry the operation. This
3401 * will either succeed because the page is now uptodate and unlocked, or it
3402 * will register a callback when the page is unlocked at IO completion. Through
3403 * that callback, io_uring uses task_work to setup a retry of the operation.
3404 * That retry will attempt the buffered read again. The retry will generally
3405 * succeed, or in rare cases where it fails, we then fall back to using the
3406 * async worker threads for a blocking retry.
3408 static bool io_rw_should_retry(struct io_kiocb *req)
3410 struct io_async_rw *rw = req->async_data;
3411 struct wait_page_queue *wait = &rw->wpq;
3412 struct kiocb *kiocb = &req->rw.kiocb;
3414 /* never retry for NOWAIT, we just complete with -EAGAIN */
3415 if (req->flags & REQ_F_NOWAIT)
3418 /* Only for buffered IO */
3419 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3423 * just use poll if we can, and don't attempt if the fs doesn't
3424 * support callback based unlocks
3426 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3429 wait->wait.func = io_async_buf_func;
3430 wait->wait.private = req;
3431 wait->wait.flags = 0;
3432 INIT_LIST_HEAD(&wait->wait.entry);
3433 kiocb->ki_flags |= IOCB_WAITQ;
3434 kiocb->ki_flags &= ~IOCB_NOWAIT;
3435 kiocb->ki_waitq = wait;
3439 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3441 if (req->file->f_op->read_iter)
3442 return call_read_iter(req->file, &req->rw.kiocb, iter);
3443 else if (req->file->f_op->read)
3444 return loop_rw_iter(READ, req, iter);
3449 static int io_read(struct io_kiocb *req, bool force_nonblock,
3450 struct io_comp_state *cs)
3452 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3453 struct kiocb *kiocb = &req->rw.kiocb;
3454 struct iov_iter __iter, *iter = &__iter;
3455 struct io_async_rw *rw = req->async_data;
3456 ssize_t io_size, ret, ret2;
3463 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3467 io_size = iov_iter_count(iter);
3468 req->result = io_size;
3471 /* Ensure we clear previously set non-block flag */
3472 if (!force_nonblock)
3473 kiocb->ki_flags &= ~IOCB_NOWAIT;
3475 kiocb->ki_flags |= IOCB_NOWAIT;
3478 /* If the file doesn't support async, just async punt */
3479 no_async = force_nonblock && !io_file_supports_async(req->file, READ);
3483 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), io_size);
3487 ret = io_iter_do_read(req, iter);
3491 } else if (ret == -EIOCBQUEUED) {
3494 } else if (ret == -EAGAIN) {
3495 /* IOPOLL retry should happen for io-wq threads */
3496 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3498 /* no retry on NONBLOCK marked file */
3499 if (req->file->f_flags & O_NONBLOCK)
3501 /* some cases will consume bytes even on error returns */
3502 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3505 } else if (ret < 0) {
3506 /* make sure -ERESTARTSYS -> -EINTR is done */
3510 /* read it all, or we did blocking attempt. no retry. */
3511 if (!iov_iter_count(iter) || !force_nonblock ||
3512 (req->file->f_flags & O_NONBLOCK))
3517 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3524 rw = req->async_data;
3525 /* it's copied and will be cleaned with ->io */
3527 /* now use our persistent iterator, if we aren't already */
3530 rw->bytes_done += ret;
3531 /* if we can retry, do so with the callbacks armed */
3532 if (!io_rw_should_retry(req)) {
3533 kiocb->ki_flags &= ~IOCB_WAITQ;
3538 * Now retry read with the IOCB_WAITQ parts set in the iocb. If we
3539 * get -EIOCBQUEUED, then we'll get a notification when the desired
3540 * page gets unlocked. We can also get a partial read here, and if we
3541 * do, then just retry at the new offset.
3543 ret = io_iter_do_read(req, iter);
3544 if (ret == -EIOCBQUEUED) {
3547 } else if (ret > 0 && ret < io_size) {
3548 /* we got some bytes, but not all. retry. */
3552 kiocb_done(kiocb, ret, cs);
3555 /* it's reportedly faster than delegating the null check to kfree() */
3561 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3565 ret = io_prep_rw(req, sqe);
3569 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3572 /* either don't need iovec imported or already have it */
3573 if (!req->async_data)
3575 return io_rw_prep_async(req, WRITE);
3578 static int io_write(struct io_kiocb *req, bool force_nonblock,
3579 struct io_comp_state *cs)
3581 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3582 struct kiocb *kiocb = &req->rw.kiocb;
3583 struct iov_iter __iter, *iter = &__iter;
3584 struct io_async_rw *rw = req->async_data;
3585 ssize_t ret, ret2, io_size;
3591 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3595 io_size = iov_iter_count(iter);
3596 req->result = io_size;
3598 /* Ensure we clear previously set non-block flag */
3599 if (!force_nonblock)
3600 kiocb->ki_flags &= ~IOCB_NOWAIT;
3602 kiocb->ki_flags |= IOCB_NOWAIT;
3604 /* If the file doesn't support async, just async punt */
3605 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
3608 /* file path doesn't support NOWAIT for non-direct_IO */
3609 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3610 (req->flags & REQ_F_ISREG))
3613 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), io_size);
3618 * Open-code file_start_write here to grab freeze protection,
3619 * which will be released by another thread in
3620 * io_complete_rw(). Fool lockdep by telling it the lock got
3621 * released so that it doesn't complain about the held lock when
3622 * we return to userspace.
3624 if (req->flags & REQ_F_ISREG) {
3625 sb_start_write(file_inode(req->file)->i_sb);
3626 __sb_writers_release(file_inode(req->file)->i_sb,
3629 kiocb->ki_flags |= IOCB_WRITE;
3631 if (req->file->f_op->write_iter)
3632 ret2 = call_write_iter(req->file, kiocb, iter);
3633 else if (req->file->f_op->write)
3634 ret2 = loop_rw_iter(WRITE, req, iter);
3639 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3640 * retry them without IOCB_NOWAIT.
3642 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3644 /* no retry on NONBLOCK marked file */
3645 if (ret2 == -EAGAIN && (req->file->f_flags & O_NONBLOCK))
3647 if (!force_nonblock || ret2 != -EAGAIN) {
3648 /* IOPOLL retry should happen for io-wq threads */
3649 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3652 kiocb_done(kiocb, ret2, cs);
3655 /* some cases will consume bytes even on error returns */
3656 iov_iter_revert(iter, io_size - iov_iter_count(iter));
3657 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3662 /* it's reportedly faster than delegating the null check to kfree() */
3668 static int io_renameat_prep(struct io_kiocb *req,
3669 const struct io_uring_sqe *sqe)
3671 struct io_rename *ren = &req->rename;
3672 const char __user *oldf, *newf;
3674 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3677 ren->old_dfd = READ_ONCE(sqe->fd);
3678 oldf = u64_to_user_ptr(READ_ONCE(sqe->addr));
3679 newf = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3680 ren->new_dfd = READ_ONCE(sqe->len);
3681 ren->flags = READ_ONCE(sqe->rename_flags);
3683 ren->oldpath = getname(oldf);
3684 if (IS_ERR(ren->oldpath))
3685 return PTR_ERR(ren->oldpath);
3687 ren->newpath = getname(newf);
3688 if (IS_ERR(ren->newpath)) {
3689 putname(ren->oldpath);
3690 return PTR_ERR(ren->newpath);
3693 req->flags |= REQ_F_NEED_CLEANUP;
3697 static int io_renameat(struct io_kiocb *req, bool force_nonblock)
3699 struct io_rename *ren = &req->rename;
3705 ret = do_renameat2(ren->old_dfd, ren->oldpath, ren->new_dfd,
3706 ren->newpath, ren->flags);
3708 req->flags &= ~REQ_F_NEED_CLEANUP;
3710 req_set_fail_links(req);
3711 io_req_complete(req, ret);
3715 static int io_unlinkat_prep(struct io_kiocb *req,
3716 const struct io_uring_sqe *sqe)
3718 struct io_unlink *un = &req->unlink;
3719 const char __user *fname;
3721 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3724 un->dfd = READ_ONCE(sqe->fd);
3726 un->flags = READ_ONCE(sqe->unlink_flags);
3727 if (un->flags & ~AT_REMOVEDIR)
3730 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3731 un->filename = getname(fname);
3732 if (IS_ERR(un->filename))
3733 return PTR_ERR(un->filename);
3735 req->flags |= REQ_F_NEED_CLEANUP;
3739 static int io_unlinkat(struct io_kiocb *req, bool force_nonblock)
3741 struct io_unlink *un = &req->unlink;
3747 if (un->flags & AT_REMOVEDIR)
3748 ret = do_rmdir(un->dfd, un->filename);
3750 ret = do_unlinkat(un->dfd, un->filename);
3752 req->flags &= ~REQ_F_NEED_CLEANUP;
3754 req_set_fail_links(req);
3755 io_req_complete(req, ret);
3759 static int io_shutdown_prep(struct io_kiocb *req,
3760 const struct io_uring_sqe *sqe)
3762 #if defined(CONFIG_NET)
3763 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3765 if (sqe->ioprio || sqe->off || sqe->addr || sqe->rw_flags ||
3769 req->shutdown.how = READ_ONCE(sqe->len);
3776 static int io_shutdown(struct io_kiocb *req, bool force_nonblock)
3778 #if defined(CONFIG_NET)
3779 struct socket *sock;
3785 sock = sock_from_file(req->file, &ret);
3786 if (unlikely(!sock))
3789 ret = __sys_shutdown_sock(sock, req->shutdown.how);
3790 io_req_complete(req, ret);
3797 static int __io_splice_prep(struct io_kiocb *req,
3798 const struct io_uring_sqe *sqe)
3800 struct io_splice* sp = &req->splice;
3801 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3803 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3807 sp->len = READ_ONCE(sqe->len);
3808 sp->flags = READ_ONCE(sqe->splice_flags);
3810 if (unlikely(sp->flags & ~valid_flags))
3813 sp->file_in = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in),
3814 (sp->flags & SPLICE_F_FD_IN_FIXED));
3817 req->flags |= REQ_F_NEED_CLEANUP;
3819 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
3821 * Splice operation will be punted aync, and here need to
3822 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3824 io_req_init_async(req);
3825 req->work.flags |= IO_WQ_WORK_UNBOUND;
3831 static int io_tee_prep(struct io_kiocb *req,
3832 const struct io_uring_sqe *sqe)
3834 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3836 return __io_splice_prep(req, sqe);
3839 static int io_tee(struct io_kiocb *req, bool force_nonblock)
3841 struct io_splice *sp = &req->splice;
3842 struct file *in = sp->file_in;
3843 struct file *out = sp->file_out;
3844 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3850 ret = do_tee(in, out, sp->len, flags);
3852 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3853 req->flags &= ~REQ_F_NEED_CLEANUP;
3856 req_set_fail_links(req);
3857 io_req_complete(req, ret);
3861 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3863 struct io_splice* sp = &req->splice;
3865 sp->off_in = READ_ONCE(sqe->splice_off_in);
3866 sp->off_out = READ_ONCE(sqe->off);
3867 return __io_splice_prep(req, sqe);
3870 static int io_splice(struct io_kiocb *req, bool force_nonblock)
3872 struct io_splice *sp = &req->splice;
3873 struct file *in = sp->file_in;
3874 struct file *out = sp->file_out;
3875 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3876 loff_t *poff_in, *poff_out;
3882 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3883 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3886 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3888 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3889 req->flags &= ~REQ_F_NEED_CLEANUP;
3892 req_set_fail_links(req);
3893 io_req_complete(req, ret);
3898 * IORING_OP_NOP just posts a completion event, nothing else.
3900 static int io_nop(struct io_kiocb *req, struct io_comp_state *cs)
3902 struct io_ring_ctx *ctx = req->ctx;
3904 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3907 __io_req_complete(req, 0, 0, cs);
3911 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3913 struct io_ring_ctx *ctx = req->ctx;
3918 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3920 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3923 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3924 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3927 req->sync.off = READ_ONCE(sqe->off);
3928 req->sync.len = READ_ONCE(sqe->len);
3932 static int io_fsync(struct io_kiocb *req, bool force_nonblock)
3934 loff_t end = req->sync.off + req->sync.len;
3937 /* fsync always requires a blocking context */
3941 ret = vfs_fsync_range(req->file, req->sync.off,
3942 end > 0 ? end : LLONG_MAX,
3943 req->sync.flags & IORING_FSYNC_DATASYNC);
3945 req_set_fail_links(req);
3946 io_req_complete(req, ret);
3950 static int io_fallocate_prep(struct io_kiocb *req,
3951 const struct io_uring_sqe *sqe)
3953 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3955 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3958 req->sync.off = READ_ONCE(sqe->off);
3959 req->sync.len = READ_ONCE(sqe->addr);
3960 req->sync.mode = READ_ONCE(sqe->len);
3964 static int io_fallocate(struct io_kiocb *req, bool force_nonblock)
3968 /* fallocate always requiring blocking context */
3971 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3974 req_set_fail_links(req);
3975 io_req_complete(req, ret);
3979 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3981 const char __user *fname;
3984 if (unlikely(sqe->ioprio || sqe->buf_index))
3986 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3989 /* open.how should be already initialised */
3990 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3991 req->open.how.flags |= O_LARGEFILE;
3993 req->open.dfd = READ_ONCE(sqe->fd);
3994 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3995 req->open.filename = getname(fname);
3996 if (IS_ERR(req->open.filename)) {
3997 ret = PTR_ERR(req->open.filename);
3998 req->open.filename = NULL;
4001 req->open.nofile = rlimit(RLIMIT_NOFILE);
4002 req->open.ignore_nonblock = false;
4003 req->flags |= REQ_F_NEED_CLEANUP;
4007 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4011 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4013 mode = READ_ONCE(sqe->len);
4014 flags = READ_ONCE(sqe->open_flags);
4015 req->open.how = build_open_how(flags, mode);
4016 return __io_openat_prep(req, sqe);
4019 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4021 struct open_how __user *how;
4025 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4027 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4028 len = READ_ONCE(sqe->len);
4029 if (len < OPEN_HOW_SIZE_VER0)
4032 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
4037 return __io_openat_prep(req, sqe);
4040 static int io_openat2(struct io_kiocb *req, bool force_nonblock)
4042 struct open_flags op;
4046 if (force_nonblock && !req->open.ignore_nonblock)
4049 ret = build_open_flags(&req->open.how, &op);
4053 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
4057 file = do_filp_open(req->open.dfd, req->open.filename, &op);
4060 ret = PTR_ERR(file);
4062 * A work-around to ensure that /proc/self works that way
4063 * that it should - if we get -EOPNOTSUPP back, then assume
4064 * that proc_self_get_link() failed us because we're in async
4065 * context. We should be safe to retry this from the task
4066 * itself with force_nonblock == false set, as it should not
4067 * block on lookup. Would be nice to know this upfront and
4068 * avoid the async dance, but doesn't seem feasible.
4070 if (ret == -EOPNOTSUPP && io_wq_current_is_worker()) {
4071 req->open.ignore_nonblock = true;
4072 refcount_inc(&req->refs);
4073 io_req_task_queue(req);
4077 fsnotify_open(file);
4078 fd_install(ret, file);
4081 putname(req->open.filename);
4082 req->flags &= ~REQ_F_NEED_CLEANUP;
4084 req_set_fail_links(req);
4085 io_req_complete(req, ret);
4089 static int io_openat(struct io_kiocb *req, bool force_nonblock)
4091 return io_openat2(req, force_nonblock);
4094 static int io_remove_buffers_prep(struct io_kiocb *req,
4095 const struct io_uring_sqe *sqe)
4097 struct io_provide_buf *p = &req->pbuf;
4100 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
4103 tmp = READ_ONCE(sqe->fd);
4104 if (!tmp || tmp > USHRT_MAX)
4107 memset(p, 0, sizeof(*p));
4109 p->bgid = READ_ONCE(sqe->buf_group);
4113 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
4114 int bgid, unsigned nbufs)
4118 /* shouldn't happen */
4122 /* the head kbuf is the list itself */
4123 while (!list_empty(&buf->list)) {
4124 struct io_buffer *nxt;
4126 nxt = list_first_entry(&buf->list, struct io_buffer, list);
4127 list_del(&nxt->list);
4134 idr_remove(&ctx->io_buffer_idr, bgid);
4139 static int io_remove_buffers(struct io_kiocb *req, bool force_nonblock,
4140 struct io_comp_state *cs)
4142 struct io_provide_buf *p = &req->pbuf;
4143 struct io_ring_ctx *ctx = req->ctx;
4144 struct io_buffer *head;
4147 io_ring_submit_lock(ctx, !force_nonblock);
4149 lockdep_assert_held(&ctx->uring_lock);
4152 head = idr_find(&ctx->io_buffer_idr, p->bgid);
4154 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
4156 req_set_fail_links(req);
4158 /* need to hold the lock to complete IOPOLL requests */
4159 if (ctx->flags & IORING_SETUP_IOPOLL) {
4160 __io_req_complete(req, ret, 0, cs);
4161 io_ring_submit_unlock(ctx, !force_nonblock);
4163 io_ring_submit_unlock(ctx, !force_nonblock);
4164 __io_req_complete(req, ret, 0, cs);
4169 static int io_provide_buffers_prep(struct io_kiocb *req,
4170 const struct io_uring_sqe *sqe)
4172 struct io_provide_buf *p = &req->pbuf;
4175 if (sqe->ioprio || sqe->rw_flags)
4178 tmp = READ_ONCE(sqe->fd);
4179 if (!tmp || tmp > USHRT_MAX)
4182 p->addr = READ_ONCE(sqe->addr);
4183 p->len = READ_ONCE(sqe->len);
4185 if (!access_ok(u64_to_user_ptr(p->addr), (p->len * p->nbufs)))
4188 p->bgid = READ_ONCE(sqe->buf_group);
4189 tmp = READ_ONCE(sqe->off);
4190 if (tmp > USHRT_MAX)
4196 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
4198 struct io_buffer *buf;
4199 u64 addr = pbuf->addr;
4200 int i, bid = pbuf->bid;
4202 for (i = 0; i < pbuf->nbufs; i++) {
4203 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
4208 buf->len = pbuf->len;
4213 INIT_LIST_HEAD(&buf->list);
4216 list_add_tail(&buf->list, &(*head)->list);
4220 return i ? i : -ENOMEM;
4223 static int io_provide_buffers(struct io_kiocb *req, bool force_nonblock,
4224 struct io_comp_state *cs)
4226 struct io_provide_buf *p = &req->pbuf;
4227 struct io_ring_ctx *ctx = req->ctx;
4228 struct io_buffer *head, *list;
4231 io_ring_submit_lock(ctx, !force_nonblock);
4233 lockdep_assert_held(&ctx->uring_lock);
4235 list = head = idr_find(&ctx->io_buffer_idr, p->bgid);
4237 ret = io_add_buffers(p, &head);
4242 ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1,
4245 __io_remove_buffers(ctx, head, p->bgid, -1U);
4251 req_set_fail_links(req);
4253 /* need to hold the lock to complete IOPOLL requests */
4254 if (ctx->flags & IORING_SETUP_IOPOLL) {
4255 __io_req_complete(req, ret, 0, cs);
4256 io_ring_submit_unlock(ctx, !force_nonblock);
4258 io_ring_submit_unlock(ctx, !force_nonblock);
4259 __io_req_complete(req, ret, 0, cs);
4264 static int io_epoll_ctl_prep(struct io_kiocb *req,
4265 const struct io_uring_sqe *sqe)
4267 #if defined(CONFIG_EPOLL)
4268 if (sqe->ioprio || sqe->buf_index)
4270 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4273 req->epoll.epfd = READ_ONCE(sqe->fd);
4274 req->epoll.op = READ_ONCE(sqe->len);
4275 req->epoll.fd = READ_ONCE(sqe->off);
4277 if (ep_op_has_event(req->epoll.op)) {
4278 struct epoll_event __user *ev;
4280 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
4281 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
4291 static int io_epoll_ctl(struct io_kiocb *req, bool force_nonblock,
4292 struct io_comp_state *cs)
4294 #if defined(CONFIG_EPOLL)
4295 struct io_epoll *ie = &req->epoll;
4298 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
4299 if (force_nonblock && ret == -EAGAIN)
4303 req_set_fail_links(req);
4304 __io_req_complete(req, ret, 0, cs);
4311 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4313 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4314 if (sqe->ioprio || sqe->buf_index || sqe->off)
4316 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4319 req->madvise.addr = READ_ONCE(sqe->addr);
4320 req->madvise.len = READ_ONCE(sqe->len);
4321 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
4328 static int io_madvise(struct io_kiocb *req, bool force_nonblock)
4330 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
4331 struct io_madvise *ma = &req->madvise;
4337 ret = do_madvise(current->mm, ma->addr, ma->len, ma->advice);
4339 req_set_fail_links(req);
4340 io_req_complete(req, ret);
4347 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4349 if (sqe->ioprio || sqe->buf_index || sqe->addr)
4351 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4354 req->fadvise.offset = READ_ONCE(sqe->off);
4355 req->fadvise.len = READ_ONCE(sqe->len);
4356 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4360 static int io_fadvise(struct io_kiocb *req, bool force_nonblock)
4362 struct io_fadvise *fa = &req->fadvise;
4365 if (force_nonblock) {
4366 switch (fa->advice) {
4367 case POSIX_FADV_NORMAL:
4368 case POSIX_FADV_RANDOM:
4369 case POSIX_FADV_SEQUENTIAL:
4376 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4378 req_set_fail_links(req);
4379 io_req_complete(req, ret);
4383 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4385 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4387 if (sqe->ioprio || sqe->buf_index)
4389 if (req->flags & REQ_F_FIXED_FILE)
4392 req->statx.dfd = READ_ONCE(sqe->fd);
4393 req->statx.mask = READ_ONCE(sqe->len);
4394 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4395 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4396 req->statx.flags = READ_ONCE(sqe->statx_flags);
4401 static int io_statx(struct io_kiocb *req, bool force_nonblock)
4403 struct io_statx *ctx = &req->statx;
4406 if (force_nonblock) {
4407 /* only need file table for an actual valid fd */
4408 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
4409 req->flags |= REQ_F_NO_FILE_TABLE;
4413 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4417 req_set_fail_links(req);
4418 io_req_complete(req, ret);
4422 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4425 * If we queue this for async, it must not be cancellable. That would
4426 * leave the 'file' in an undeterminate state, and here need to modify
4427 * io_wq_work.flags, so initialize io_wq_work firstly.
4429 io_req_init_async(req);
4430 req->work.flags |= IO_WQ_WORK_NO_CANCEL;
4432 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4434 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4435 sqe->rw_flags || sqe->buf_index)
4437 if (req->flags & REQ_F_FIXED_FILE)
4440 req->close.fd = READ_ONCE(sqe->fd);
4441 if ((req->file && req->file->f_op == &io_uring_fops))
4444 req->close.put_file = NULL;
4448 static int io_close(struct io_kiocb *req, bool force_nonblock,
4449 struct io_comp_state *cs)
4451 struct io_close *close = &req->close;
4454 /* might be already done during nonblock submission */
4455 if (!close->put_file) {
4456 ret = __close_fd_get_file(close->fd, &close->put_file);
4458 return (ret == -ENOENT) ? -EBADF : ret;
4461 /* if the file has a flush method, be safe and punt to async */
4462 if (close->put_file->f_op->flush && force_nonblock) {
4463 /* was never set, but play safe */
4464 req->flags &= ~REQ_F_NOWAIT;
4465 /* avoid grabbing files - we don't need the files */
4466 req->flags |= REQ_F_NO_FILE_TABLE;
4470 /* No ->flush() or already async, safely close from here */
4471 ret = filp_close(close->put_file, req->work.identity->files);
4473 req_set_fail_links(req);
4474 fput(close->put_file);
4475 close->put_file = NULL;
4476 __io_req_complete(req, ret, 0, cs);
4480 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4482 struct io_ring_ctx *ctx = req->ctx;
4487 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4489 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4492 req->sync.off = READ_ONCE(sqe->off);
4493 req->sync.len = READ_ONCE(sqe->len);
4494 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4498 static int io_sync_file_range(struct io_kiocb *req, bool force_nonblock)
4502 /* sync_file_range always requires a blocking context */
4506 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4509 req_set_fail_links(req);
4510 io_req_complete(req, ret);
4514 #if defined(CONFIG_NET)
4515 static int io_setup_async_msg(struct io_kiocb *req,
4516 struct io_async_msghdr *kmsg)
4518 struct io_async_msghdr *async_msg = req->async_data;
4522 if (io_alloc_async_data(req)) {
4523 if (kmsg->iov != kmsg->fast_iov)
4527 async_msg = req->async_data;
4528 req->flags |= REQ_F_NEED_CLEANUP;
4529 memcpy(async_msg, kmsg, sizeof(*kmsg));
4533 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4534 struct io_async_msghdr *iomsg)
4536 iomsg->iov = iomsg->fast_iov;
4537 iomsg->msg.msg_name = &iomsg->addr;
4538 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4539 req->sr_msg.msg_flags, &iomsg->iov);
4542 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4544 struct io_async_msghdr *async_msg = req->async_data;
4545 struct io_sr_msg *sr = &req->sr_msg;
4548 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4551 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4552 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4553 sr->len = READ_ONCE(sqe->len);
4555 #ifdef CONFIG_COMPAT
4556 if (req->ctx->compat)
4557 sr->msg_flags |= MSG_CMSG_COMPAT;
4560 if (!async_msg || !io_op_defs[req->opcode].needs_async_data)
4562 ret = io_sendmsg_copy_hdr(req, async_msg);
4564 req->flags |= REQ_F_NEED_CLEANUP;
4568 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4569 struct io_comp_state *cs)
4571 struct io_async_msghdr iomsg, *kmsg;
4572 struct socket *sock;
4576 sock = sock_from_file(req->file, &ret);
4577 if (unlikely(!sock))
4580 if (req->async_data) {
4581 kmsg = req->async_data;
4582 kmsg->msg.msg_name = &kmsg->addr;
4583 /* if iov is set, it's allocated already */
4585 kmsg->iov = kmsg->fast_iov;
4586 kmsg->msg.msg_iter.iov = kmsg->iov;
4588 ret = io_sendmsg_copy_hdr(req, &iomsg);
4594 flags = req->sr_msg.msg_flags;
4595 if (flags & MSG_DONTWAIT)
4596 req->flags |= REQ_F_NOWAIT;
4597 else if (force_nonblock)
4598 flags |= MSG_DONTWAIT;
4600 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4601 if (force_nonblock && ret == -EAGAIN)
4602 return io_setup_async_msg(req, kmsg);
4603 if (ret == -ERESTARTSYS)
4606 if (kmsg->iov != kmsg->fast_iov)
4608 req->flags &= ~REQ_F_NEED_CLEANUP;
4610 req_set_fail_links(req);
4611 __io_req_complete(req, ret, 0, cs);
4615 static int io_send(struct io_kiocb *req, bool force_nonblock,
4616 struct io_comp_state *cs)
4618 struct io_sr_msg *sr = &req->sr_msg;
4621 struct socket *sock;
4625 sock = sock_from_file(req->file, &ret);
4626 if (unlikely(!sock))
4629 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4633 msg.msg_name = NULL;
4634 msg.msg_control = NULL;
4635 msg.msg_controllen = 0;
4636 msg.msg_namelen = 0;
4638 flags = req->sr_msg.msg_flags;
4639 if (flags & MSG_DONTWAIT)
4640 req->flags |= REQ_F_NOWAIT;
4641 else if (force_nonblock)
4642 flags |= MSG_DONTWAIT;
4644 msg.msg_flags = flags;
4645 ret = sock_sendmsg(sock, &msg);
4646 if (force_nonblock && ret == -EAGAIN)
4648 if (ret == -ERESTARTSYS)
4652 req_set_fail_links(req);
4653 __io_req_complete(req, ret, 0, cs);
4657 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4658 struct io_async_msghdr *iomsg)
4660 struct io_sr_msg *sr = &req->sr_msg;
4661 struct iovec __user *uiov;
4665 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4666 &iomsg->uaddr, &uiov, &iov_len);
4670 if (req->flags & REQ_F_BUFFER_SELECT) {
4673 if (copy_from_user(iomsg->iov, uiov, sizeof(*uiov)))
4675 sr->len = iomsg->iov[0].iov_len;
4676 iov_iter_init(&iomsg->msg.msg_iter, READ, iomsg->iov, 1,
4680 ret = __import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4681 &iomsg->iov, &iomsg->msg.msg_iter,
4690 #ifdef CONFIG_COMPAT
4691 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4692 struct io_async_msghdr *iomsg)
4694 struct compat_msghdr __user *msg_compat;
4695 struct io_sr_msg *sr = &req->sr_msg;
4696 struct compat_iovec __user *uiov;
4701 msg_compat = (struct compat_msghdr __user *) sr->umsg;
4702 ret = __get_compat_msghdr(&iomsg->msg, msg_compat, &iomsg->uaddr,
4707 uiov = compat_ptr(ptr);
4708 if (req->flags & REQ_F_BUFFER_SELECT) {
4709 compat_ssize_t clen;
4713 if (!access_ok(uiov, sizeof(*uiov)))
4715 if (__get_user(clen, &uiov->iov_len))
4719 sr->len = iomsg->iov[0].iov_len;
4722 ret = __import_iovec(READ, (struct iovec __user *)uiov, len,
4723 UIO_FASTIOV, &iomsg->iov,
4724 &iomsg->msg.msg_iter, true);
4733 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4734 struct io_async_msghdr *iomsg)
4736 iomsg->msg.msg_name = &iomsg->addr;
4737 iomsg->iov = iomsg->fast_iov;
4739 #ifdef CONFIG_COMPAT
4740 if (req->ctx->compat)
4741 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4744 return __io_recvmsg_copy_hdr(req, iomsg);
4747 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4750 struct io_sr_msg *sr = &req->sr_msg;
4751 struct io_buffer *kbuf;
4753 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4758 req->flags |= REQ_F_BUFFER_SELECTED;
4762 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4764 return io_put_kbuf(req, req->sr_msg.kbuf);
4767 static int io_recvmsg_prep(struct io_kiocb *req,
4768 const struct io_uring_sqe *sqe)
4770 struct io_async_msghdr *async_msg = req->async_data;
4771 struct io_sr_msg *sr = &req->sr_msg;
4774 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4777 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4778 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4779 sr->len = READ_ONCE(sqe->len);
4780 sr->bgid = READ_ONCE(sqe->buf_group);
4782 #ifdef CONFIG_COMPAT
4783 if (req->ctx->compat)
4784 sr->msg_flags |= MSG_CMSG_COMPAT;
4787 if (!async_msg || !io_op_defs[req->opcode].needs_async_data)
4789 ret = io_recvmsg_copy_hdr(req, async_msg);
4791 req->flags |= REQ_F_NEED_CLEANUP;
4795 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4796 struct io_comp_state *cs)
4798 struct io_async_msghdr iomsg, *kmsg;
4799 struct socket *sock;
4800 struct io_buffer *kbuf;
4802 int ret, cflags = 0;
4804 sock = sock_from_file(req->file, &ret);
4805 if (unlikely(!sock))
4808 if (req->async_data) {
4809 kmsg = req->async_data;
4810 kmsg->msg.msg_name = &kmsg->addr;
4811 /* if iov is set, it's allocated already */
4813 kmsg->iov = kmsg->fast_iov;
4814 kmsg->msg.msg_iter.iov = kmsg->iov;
4816 ret = io_recvmsg_copy_hdr(req, &iomsg);
4822 if (req->flags & REQ_F_BUFFER_SELECT) {
4823 kbuf = io_recv_buffer_select(req, !force_nonblock);
4825 return PTR_ERR(kbuf);
4826 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4827 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->iov,
4828 1, req->sr_msg.len);
4831 flags = req->sr_msg.msg_flags;
4832 if (flags & MSG_DONTWAIT)
4833 req->flags |= REQ_F_NOWAIT;
4834 else if (force_nonblock)
4835 flags |= MSG_DONTWAIT;
4837 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4838 kmsg->uaddr, flags);
4839 if (force_nonblock && ret == -EAGAIN)
4840 return io_setup_async_msg(req, kmsg);
4841 if (ret == -ERESTARTSYS)
4844 if (req->flags & REQ_F_BUFFER_SELECTED)
4845 cflags = io_put_recv_kbuf(req);
4846 if (kmsg->iov != kmsg->fast_iov)
4848 req->flags &= ~REQ_F_NEED_CLEANUP;
4850 req_set_fail_links(req);
4851 __io_req_complete(req, ret, cflags, cs);
4855 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4856 struct io_comp_state *cs)
4858 struct io_buffer *kbuf;
4859 struct io_sr_msg *sr = &req->sr_msg;
4861 void __user *buf = sr->buf;
4862 struct socket *sock;
4865 int ret, cflags = 0;
4867 sock = sock_from_file(req->file, &ret);
4868 if (unlikely(!sock))
4871 if (req->flags & REQ_F_BUFFER_SELECT) {
4872 kbuf = io_recv_buffer_select(req, !force_nonblock);
4874 return PTR_ERR(kbuf);
4875 buf = u64_to_user_ptr(kbuf->addr);
4878 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4882 msg.msg_name = NULL;
4883 msg.msg_control = NULL;
4884 msg.msg_controllen = 0;
4885 msg.msg_namelen = 0;
4886 msg.msg_iocb = NULL;
4889 flags = req->sr_msg.msg_flags;
4890 if (flags & MSG_DONTWAIT)
4891 req->flags |= REQ_F_NOWAIT;
4892 else if (force_nonblock)
4893 flags |= MSG_DONTWAIT;
4895 ret = sock_recvmsg(sock, &msg, flags);
4896 if (force_nonblock && ret == -EAGAIN)
4898 if (ret == -ERESTARTSYS)
4901 if (req->flags & REQ_F_BUFFER_SELECTED)
4902 cflags = io_put_recv_kbuf(req);
4904 req_set_fail_links(req);
4905 __io_req_complete(req, ret, cflags, cs);
4909 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4911 struct io_accept *accept = &req->accept;
4913 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4915 if (sqe->ioprio || sqe->len || sqe->buf_index)
4918 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4919 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4920 accept->flags = READ_ONCE(sqe->accept_flags);
4921 accept->nofile = rlimit(RLIMIT_NOFILE);
4925 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4926 struct io_comp_state *cs)
4928 struct io_accept *accept = &req->accept;
4929 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4932 if (req->file->f_flags & O_NONBLOCK)
4933 req->flags |= REQ_F_NOWAIT;
4935 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4936 accept->addr_len, accept->flags,
4938 if (ret == -EAGAIN && force_nonblock)
4941 if (ret == -ERESTARTSYS)
4943 req_set_fail_links(req);
4945 __io_req_complete(req, ret, 0, cs);
4949 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4951 struct io_connect *conn = &req->connect;
4952 struct io_async_connect *io = req->async_data;
4954 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4956 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4959 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4960 conn->addr_len = READ_ONCE(sqe->addr2);
4965 return move_addr_to_kernel(conn->addr, conn->addr_len,
4969 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4970 struct io_comp_state *cs)
4972 struct io_async_connect __io, *io;
4973 unsigned file_flags;
4976 if (req->async_data) {
4977 io = req->async_data;
4979 ret = move_addr_to_kernel(req->connect.addr,
4980 req->connect.addr_len,
4987 file_flags = force_nonblock ? O_NONBLOCK : 0;
4989 ret = __sys_connect_file(req->file, &io->address,
4990 req->connect.addr_len, file_flags);
4991 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4992 if (req->async_data)
4994 if (io_alloc_async_data(req)) {
4998 io = req->async_data;
4999 memcpy(req->async_data, &__io, sizeof(__io));
5002 if (ret == -ERESTARTSYS)
5006 req_set_fail_links(req);
5007 __io_req_complete(req, ret, 0, cs);
5010 #else /* !CONFIG_NET */
5011 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5016 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
5017 struct io_comp_state *cs)
5022 static int io_send(struct io_kiocb *req, bool force_nonblock,
5023 struct io_comp_state *cs)
5028 static int io_recvmsg_prep(struct io_kiocb *req,
5029 const struct io_uring_sqe *sqe)
5034 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
5035 struct io_comp_state *cs)
5040 static int io_recv(struct io_kiocb *req, bool force_nonblock,
5041 struct io_comp_state *cs)
5046 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5051 static int io_accept(struct io_kiocb *req, bool force_nonblock,
5052 struct io_comp_state *cs)
5057 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5062 static int io_connect(struct io_kiocb *req, bool force_nonblock,
5063 struct io_comp_state *cs)
5067 #endif /* CONFIG_NET */
5069 struct io_poll_table {
5070 struct poll_table_struct pt;
5071 struct io_kiocb *req;
5075 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
5076 __poll_t mask, task_work_func_t func)
5081 /* for instances that support it check for an event match first: */
5082 if (mask && !(mask & poll->events))
5085 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
5087 list_del_init(&poll->wait.entry);
5090 init_task_work(&req->task_work, func);
5091 percpu_ref_get(&req->ctx->refs);
5094 * If we using the signalfd wait_queue_head for this wakeup, then
5095 * it's not safe to use TWA_SIGNAL as we could be recursing on the
5096 * tsk->sighand->siglock on doing the wakeup. Should not be needed
5097 * either, as the normal wakeup will suffice.
5099 twa_signal_ok = (poll->head != &req->task->sighand->signalfd_wqh);
5102 * If this fails, then the task is exiting. When a task exits, the
5103 * work gets canceled, so just cancel this request as well instead
5104 * of executing it. We can't safely execute it anyway, as we may not
5105 * have the needed state needed for it anyway.
5107 ret = io_req_task_work_add(req, twa_signal_ok);
5108 if (unlikely(ret)) {
5109 struct task_struct *tsk;
5111 WRITE_ONCE(poll->canceled, true);
5112 tsk = io_wq_get_task(req->ctx->io_wq);
5113 task_work_add(tsk, &req->task_work, TWA_NONE);
5114 wake_up_process(tsk);
5119 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
5120 __acquires(&req->ctx->completion_lock)
5122 struct io_ring_ctx *ctx = req->ctx;
5124 if (!req->result && !READ_ONCE(poll->canceled)) {
5125 struct poll_table_struct pt = { ._key = poll->events };
5127 req->result = vfs_poll(req->file, &pt) & poll->events;
5130 spin_lock_irq(&ctx->completion_lock);
5131 if (!req->result && !READ_ONCE(poll->canceled)) {
5132 add_wait_queue(poll->head, &poll->wait);
5139 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
5141 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
5142 if (req->opcode == IORING_OP_POLL_ADD)
5143 return req->async_data;
5144 return req->apoll->double_poll;
5147 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
5149 if (req->opcode == IORING_OP_POLL_ADD)
5151 return &req->apoll->poll;
5154 static void io_poll_remove_double(struct io_kiocb *req)
5156 struct io_poll_iocb *poll = io_poll_get_double(req);
5158 lockdep_assert_held(&req->ctx->completion_lock);
5160 if (poll && poll->head) {
5161 struct wait_queue_head *head = poll->head;
5163 spin_lock(&head->lock);
5164 list_del_init(&poll->wait.entry);
5165 if (poll->wait.private)
5166 refcount_dec(&req->refs);
5168 spin_unlock(&head->lock);
5172 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
5174 struct io_ring_ctx *ctx = req->ctx;
5176 io_poll_remove_double(req);
5177 req->poll.done = true;
5178 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
5179 io_commit_cqring(ctx);
5182 static void io_poll_task_func(struct callback_head *cb)
5184 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
5185 struct io_ring_ctx *ctx = req->ctx;
5186 struct io_kiocb *nxt;
5188 if (io_poll_rewait(req, &req->poll)) {
5189 spin_unlock_irq(&ctx->completion_lock);
5191 hash_del(&req->hash_node);
5192 io_poll_complete(req, req->result, 0);
5193 spin_unlock_irq(&ctx->completion_lock);
5195 nxt = io_put_req_find_next(req);
5196 io_cqring_ev_posted(ctx);
5198 __io_req_task_submit(nxt);
5201 percpu_ref_put(&ctx->refs);
5204 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
5205 int sync, void *key)
5207 struct io_kiocb *req = wait->private;
5208 struct io_poll_iocb *poll = io_poll_get_single(req);
5209 __poll_t mask = key_to_poll(key);
5211 /* for instances that support it check for an event match first: */
5212 if (mask && !(mask & poll->events))
5215 list_del_init(&wait->entry);
5217 if (poll && poll->head) {
5220 spin_lock(&poll->head->lock);
5221 done = list_empty(&poll->wait.entry);
5223 list_del_init(&poll->wait.entry);
5224 /* make sure double remove sees this as being gone */
5225 wait->private = NULL;
5226 spin_unlock(&poll->head->lock);
5228 /* use wait func handler, so it matches the rq type */
5229 poll->wait.func(&poll->wait, mode, sync, key);
5232 refcount_dec(&req->refs);
5236 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
5237 wait_queue_func_t wake_func)
5241 poll->canceled = false;
5242 poll->events = events;
5243 INIT_LIST_HEAD(&poll->wait.entry);
5244 init_waitqueue_func_entry(&poll->wait, wake_func);
5247 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
5248 struct wait_queue_head *head,
5249 struct io_poll_iocb **poll_ptr)
5251 struct io_kiocb *req = pt->req;
5254 * If poll->head is already set, it's because the file being polled
5255 * uses multiple waitqueues for poll handling (eg one for read, one
5256 * for write). Setup a separate io_poll_iocb if this happens.
5258 if (unlikely(poll->head)) {
5259 struct io_poll_iocb *poll_one = poll;
5261 /* already have a 2nd entry, fail a third attempt */
5263 pt->error = -EINVAL;
5266 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
5268 pt->error = -ENOMEM;
5271 io_init_poll_iocb(poll, poll_one->events, io_poll_double_wake);
5272 refcount_inc(&req->refs);
5273 poll->wait.private = req;
5280 if (poll->events & EPOLLEXCLUSIVE)
5281 add_wait_queue_exclusive(head, &poll->wait);
5283 add_wait_queue(head, &poll->wait);
5286 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
5287 struct poll_table_struct *p)
5289 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5290 struct async_poll *apoll = pt->req->apoll;
5292 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
5295 static void io_async_task_func(struct callback_head *cb)
5297 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
5298 struct async_poll *apoll = req->apoll;
5299 struct io_ring_ctx *ctx = req->ctx;
5301 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
5303 if (io_poll_rewait(req, &apoll->poll)) {
5304 spin_unlock_irq(&ctx->completion_lock);
5305 percpu_ref_put(&ctx->refs);
5309 /* If req is still hashed, it cannot have been canceled. Don't check. */
5310 if (hash_hashed(&req->hash_node))
5311 hash_del(&req->hash_node);
5313 io_poll_remove_double(req);
5314 spin_unlock_irq(&ctx->completion_lock);
5316 if (!READ_ONCE(apoll->poll.canceled))
5317 __io_req_task_submit(req);
5319 __io_req_task_cancel(req, -ECANCELED);
5321 percpu_ref_put(&ctx->refs);
5322 kfree(apoll->double_poll);
5326 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5329 struct io_kiocb *req = wait->private;
5330 struct io_poll_iocb *poll = &req->apoll->poll;
5332 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
5335 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
5338 static void io_poll_req_insert(struct io_kiocb *req)
5340 struct io_ring_ctx *ctx = req->ctx;
5341 struct hlist_head *list;
5343 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
5344 hlist_add_head(&req->hash_node, list);
5347 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
5348 struct io_poll_iocb *poll,
5349 struct io_poll_table *ipt, __poll_t mask,
5350 wait_queue_func_t wake_func)
5351 __acquires(&ctx->completion_lock)
5353 struct io_ring_ctx *ctx = req->ctx;
5354 bool cancel = false;
5356 INIT_HLIST_NODE(&req->hash_node);
5357 io_init_poll_iocb(poll, mask, wake_func);
5358 poll->file = req->file;
5359 poll->wait.private = req;
5361 ipt->pt._key = mask;
5363 ipt->error = -EINVAL;
5365 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5367 spin_lock_irq(&ctx->completion_lock);
5368 if (likely(poll->head)) {
5369 spin_lock(&poll->head->lock);
5370 if (unlikely(list_empty(&poll->wait.entry))) {
5376 if (mask || ipt->error)
5377 list_del_init(&poll->wait.entry);
5379 WRITE_ONCE(poll->canceled, true);
5380 else if (!poll->done) /* actually waiting for an event */
5381 io_poll_req_insert(req);
5382 spin_unlock(&poll->head->lock);
5388 static bool io_arm_poll_handler(struct io_kiocb *req)
5390 const struct io_op_def *def = &io_op_defs[req->opcode];
5391 struct io_ring_ctx *ctx = req->ctx;
5392 struct async_poll *apoll;
5393 struct io_poll_table ipt;
5397 if (!req->file || !file_can_poll(req->file))
5399 if (req->flags & REQ_F_POLLED)
5403 else if (def->pollout)
5407 /* if we can't nonblock try, then no point in arming a poll handler */
5408 if (!io_file_supports_async(req->file, rw))
5411 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5412 if (unlikely(!apoll))
5414 apoll->double_poll = NULL;
5416 req->flags |= REQ_F_POLLED;
5421 mask |= POLLIN | POLLRDNORM;
5423 mask |= POLLOUT | POLLWRNORM;
5425 /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
5426 if ((req->opcode == IORING_OP_RECVMSG) &&
5427 (req->sr_msg.msg_flags & MSG_ERRQUEUE))
5430 mask |= POLLERR | POLLPRI;
5432 ipt.pt._qproc = io_async_queue_proc;
5434 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5436 if (ret || ipt.error) {
5437 io_poll_remove_double(req);
5438 spin_unlock_irq(&ctx->completion_lock);
5439 kfree(apoll->double_poll);
5443 spin_unlock_irq(&ctx->completion_lock);
5444 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
5445 apoll->poll.events);
5449 static bool __io_poll_remove_one(struct io_kiocb *req,
5450 struct io_poll_iocb *poll)
5452 bool do_complete = false;
5454 spin_lock(&poll->head->lock);
5455 WRITE_ONCE(poll->canceled, true);
5456 if (!list_empty(&poll->wait.entry)) {
5457 list_del_init(&poll->wait.entry);
5460 spin_unlock(&poll->head->lock);
5461 hash_del(&req->hash_node);
5465 static bool io_poll_remove_one(struct io_kiocb *req)
5469 io_poll_remove_double(req);
5471 if (req->opcode == IORING_OP_POLL_ADD) {
5472 do_complete = __io_poll_remove_one(req, &req->poll);
5474 struct async_poll *apoll = req->apoll;
5476 /* non-poll requests have submit ref still */
5477 do_complete = __io_poll_remove_one(req, &apoll->poll);
5480 kfree(apoll->double_poll);
5486 io_cqring_fill_event(req, -ECANCELED);
5487 io_commit_cqring(req->ctx);
5488 req_set_fail_links(req);
5489 io_put_req_deferred(req, 1);
5496 * Returns true if we found and killed one or more poll requests
5498 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk,
5499 struct files_struct *files)
5501 struct hlist_node *tmp;
5502 struct io_kiocb *req;
5505 spin_lock_irq(&ctx->completion_lock);
5506 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5507 struct hlist_head *list;
5509 list = &ctx->cancel_hash[i];
5510 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5511 if (io_match_task(req, tsk, files))
5512 posted += io_poll_remove_one(req);
5515 spin_unlock_irq(&ctx->completion_lock);
5518 io_cqring_ev_posted(ctx);
5523 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
5525 struct hlist_head *list;
5526 struct io_kiocb *req;
5528 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5529 hlist_for_each_entry(req, list, hash_node) {
5530 if (sqe_addr != req->user_data)
5532 if (io_poll_remove_one(req))
5540 static int io_poll_remove_prep(struct io_kiocb *req,
5541 const struct io_uring_sqe *sqe)
5543 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5545 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
5549 req->poll_remove.addr = READ_ONCE(sqe->addr);
5554 * Find a running poll command that matches one specified in sqe->addr,
5555 * and remove it if found.
5557 static int io_poll_remove(struct io_kiocb *req)
5559 struct io_ring_ctx *ctx = req->ctx;
5562 spin_lock_irq(&ctx->completion_lock);
5563 ret = io_poll_cancel(ctx, req->poll_remove.addr);
5564 spin_unlock_irq(&ctx->completion_lock);
5567 req_set_fail_links(req);
5568 io_req_complete(req, ret);
5572 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5575 struct io_kiocb *req = wait->private;
5576 struct io_poll_iocb *poll = &req->poll;
5578 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5581 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5582 struct poll_table_struct *p)
5584 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5586 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5589 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5591 struct io_poll_iocb *poll = &req->poll;
5594 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5596 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
5599 events = READ_ONCE(sqe->poll32_events);
5601 events = swahw32(events);
5603 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP |
5604 (events & EPOLLEXCLUSIVE);
5608 static int io_poll_add(struct io_kiocb *req)
5610 struct io_poll_iocb *poll = &req->poll;
5611 struct io_ring_ctx *ctx = req->ctx;
5612 struct io_poll_table ipt;
5615 ipt.pt._qproc = io_poll_queue_proc;
5617 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5620 if (mask) { /* no async, we'd stolen it */
5622 io_poll_complete(req, mask, 0);
5624 spin_unlock_irq(&ctx->completion_lock);
5627 io_cqring_ev_posted(ctx);
5633 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5635 struct io_timeout_data *data = container_of(timer,
5636 struct io_timeout_data, timer);
5637 struct io_kiocb *req = data->req;
5638 struct io_ring_ctx *ctx = req->ctx;
5639 unsigned long flags;
5641 spin_lock_irqsave(&ctx->completion_lock, flags);
5642 list_del_init(&req->timeout.list);
5643 atomic_set(&req->ctx->cq_timeouts,
5644 atomic_read(&req->ctx->cq_timeouts) + 1);
5646 io_cqring_fill_event(req, -ETIME);
5647 io_commit_cqring(ctx);
5648 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5650 io_cqring_ev_posted(ctx);
5651 req_set_fail_links(req);
5653 return HRTIMER_NORESTART;
5656 static struct io_kiocb *io_timeout_extract(struct io_ring_ctx *ctx,
5659 struct io_timeout_data *io;
5660 struct io_kiocb *req;
5663 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5664 if (user_data == req->user_data) {
5671 return ERR_PTR(ret);
5673 io = req->async_data;
5674 ret = hrtimer_try_to_cancel(&io->timer);
5676 return ERR_PTR(-EALREADY);
5677 list_del_init(&req->timeout.list);
5681 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5683 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5686 return PTR_ERR(req);
5688 req_set_fail_links(req);
5689 io_cqring_fill_event(req, -ECANCELED);
5690 io_put_req_deferred(req, 1);
5694 static int io_timeout_update(struct io_ring_ctx *ctx, __u64 user_data,
5695 struct timespec64 *ts, enum hrtimer_mode mode)
5697 struct io_kiocb *req = io_timeout_extract(ctx, user_data);
5698 struct io_timeout_data *data;
5701 return PTR_ERR(req);
5703 req->timeout.off = 0; /* noseq */
5704 data = req->async_data;
5705 list_add_tail(&req->timeout.list, &ctx->timeout_list);
5706 hrtimer_init(&data->timer, CLOCK_MONOTONIC, mode);
5707 data->timer.function = io_timeout_fn;
5708 hrtimer_start(&data->timer, timespec64_to_ktime(*ts), mode);
5712 static int io_timeout_remove_prep(struct io_kiocb *req,
5713 const struct io_uring_sqe *sqe)
5715 struct io_timeout_rem *tr = &req->timeout_rem;
5717 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5719 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5721 if (sqe->ioprio || sqe->buf_index || sqe->len)
5724 tr->addr = READ_ONCE(sqe->addr);
5725 tr->flags = READ_ONCE(sqe->timeout_flags);
5726 if (tr->flags & IORING_TIMEOUT_UPDATE) {
5727 if (tr->flags & ~(IORING_TIMEOUT_UPDATE|IORING_TIMEOUT_ABS))
5729 if (get_timespec64(&tr->ts, u64_to_user_ptr(sqe->addr2)))
5731 } else if (tr->flags) {
5732 /* timeout removal doesn't support flags */
5740 * Remove or update an existing timeout command
5742 static int io_timeout_remove(struct io_kiocb *req)
5744 struct io_timeout_rem *tr = &req->timeout_rem;
5745 struct io_ring_ctx *ctx = req->ctx;
5748 spin_lock_irq(&ctx->completion_lock);
5749 if (req->timeout_rem.flags & IORING_TIMEOUT_UPDATE) {
5750 enum hrtimer_mode mode = (tr->flags & IORING_TIMEOUT_ABS)
5751 ? HRTIMER_MODE_ABS : HRTIMER_MODE_REL;
5753 ret = io_timeout_update(ctx, tr->addr, &tr->ts, mode);
5755 ret = io_timeout_cancel(ctx, tr->addr);
5758 io_cqring_fill_event(req, ret);
5759 io_commit_cqring(ctx);
5760 spin_unlock_irq(&ctx->completion_lock);
5761 io_cqring_ev_posted(ctx);
5763 req_set_fail_links(req);
5768 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5769 bool is_timeout_link)
5771 struct io_timeout_data *data;
5773 u32 off = READ_ONCE(sqe->off);
5775 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5777 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5779 if (off && is_timeout_link)
5781 flags = READ_ONCE(sqe->timeout_flags);
5782 if (flags & ~IORING_TIMEOUT_ABS)
5785 req->timeout.off = off;
5787 if (!req->async_data && io_alloc_async_data(req))
5790 data = req->async_data;
5793 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5796 if (flags & IORING_TIMEOUT_ABS)
5797 data->mode = HRTIMER_MODE_ABS;
5799 data->mode = HRTIMER_MODE_REL;
5801 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5805 static int io_timeout(struct io_kiocb *req)
5807 struct io_ring_ctx *ctx = req->ctx;
5808 struct io_timeout_data *data = req->async_data;
5809 struct list_head *entry;
5810 u32 tail, off = req->timeout.off;
5812 spin_lock_irq(&ctx->completion_lock);
5815 * sqe->off holds how many events that need to occur for this
5816 * timeout event to be satisfied. If it isn't set, then this is
5817 * a pure timeout request, sequence isn't used.
5819 if (io_is_timeout_noseq(req)) {
5820 entry = ctx->timeout_list.prev;
5824 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5825 req->timeout.target_seq = tail + off;
5828 * Insertion sort, ensuring the first entry in the list is always
5829 * the one we need first.
5831 list_for_each_prev(entry, &ctx->timeout_list) {
5832 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5835 if (io_is_timeout_noseq(nxt))
5837 /* nxt.seq is behind @tail, otherwise would've been completed */
5838 if (off >= nxt->timeout.target_seq - tail)
5842 list_add(&req->timeout.list, entry);
5843 data->timer.function = io_timeout_fn;
5844 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5845 spin_unlock_irq(&ctx->completion_lock);
5849 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5851 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5853 return req->user_data == (unsigned long) data;
5856 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
5858 enum io_wq_cancel cancel_ret;
5861 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr, false);
5862 switch (cancel_ret) {
5863 case IO_WQ_CANCEL_OK:
5866 case IO_WQ_CANCEL_RUNNING:
5869 case IO_WQ_CANCEL_NOTFOUND:
5877 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5878 struct io_kiocb *req, __u64 sqe_addr,
5881 unsigned long flags;
5884 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
5885 if (ret != -ENOENT) {
5886 spin_lock_irqsave(&ctx->completion_lock, flags);
5890 spin_lock_irqsave(&ctx->completion_lock, flags);
5891 ret = io_timeout_cancel(ctx, sqe_addr);
5894 ret = io_poll_cancel(ctx, sqe_addr);
5898 io_cqring_fill_event(req, ret);
5899 io_commit_cqring(ctx);
5900 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5901 io_cqring_ev_posted(ctx);
5904 req_set_fail_links(req);
5908 static int io_async_cancel_prep(struct io_kiocb *req,
5909 const struct io_uring_sqe *sqe)
5911 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5913 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5915 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5918 req->cancel.addr = READ_ONCE(sqe->addr);
5922 static int io_async_cancel(struct io_kiocb *req)
5924 struct io_ring_ctx *ctx = req->ctx;
5926 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
5930 static int io_files_update_prep(struct io_kiocb *req,
5931 const struct io_uring_sqe *sqe)
5933 if (unlikely(req->ctx->flags & IORING_SETUP_SQPOLL))
5935 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5937 if (sqe->ioprio || sqe->rw_flags)
5940 req->files_update.offset = READ_ONCE(sqe->off);
5941 req->files_update.nr_args = READ_ONCE(sqe->len);
5942 if (!req->files_update.nr_args)
5944 req->files_update.arg = READ_ONCE(sqe->addr);
5948 static int io_files_update(struct io_kiocb *req, bool force_nonblock,
5949 struct io_comp_state *cs)
5951 struct io_ring_ctx *ctx = req->ctx;
5952 struct io_uring_files_update up;
5958 up.offset = req->files_update.offset;
5959 up.fds = req->files_update.arg;
5961 mutex_lock(&ctx->uring_lock);
5962 ret = __io_sqe_files_update(ctx, &up, req->files_update.nr_args);
5963 mutex_unlock(&ctx->uring_lock);
5966 req_set_fail_links(req);
5967 __io_req_complete(req, ret, 0, cs);
5971 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5973 switch (req->opcode) {
5976 case IORING_OP_READV:
5977 case IORING_OP_READ_FIXED:
5978 case IORING_OP_READ:
5979 return io_read_prep(req, sqe);
5980 case IORING_OP_WRITEV:
5981 case IORING_OP_WRITE_FIXED:
5982 case IORING_OP_WRITE:
5983 return io_write_prep(req, sqe);
5984 case IORING_OP_POLL_ADD:
5985 return io_poll_add_prep(req, sqe);
5986 case IORING_OP_POLL_REMOVE:
5987 return io_poll_remove_prep(req, sqe);
5988 case IORING_OP_FSYNC:
5989 return io_prep_fsync(req, sqe);
5990 case IORING_OP_SYNC_FILE_RANGE:
5991 return io_prep_sfr(req, sqe);
5992 case IORING_OP_SENDMSG:
5993 case IORING_OP_SEND:
5994 return io_sendmsg_prep(req, sqe);
5995 case IORING_OP_RECVMSG:
5996 case IORING_OP_RECV:
5997 return io_recvmsg_prep(req, sqe);
5998 case IORING_OP_CONNECT:
5999 return io_connect_prep(req, sqe);
6000 case IORING_OP_TIMEOUT:
6001 return io_timeout_prep(req, sqe, false);
6002 case IORING_OP_TIMEOUT_REMOVE:
6003 return io_timeout_remove_prep(req, sqe);
6004 case IORING_OP_ASYNC_CANCEL:
6005 return io_async_cancel_prep(req, sqe);
6006 case IORING_OP_LINK_TIMEOUT:
6007 return io_timeout_prep(req, sqe, true);
6008 case IORING_OP_ACCEPT:
6009 return io_accept_prep(req, sqe);
6010 case IORING_OP_FALLOCATE:
6011 return io_fallocate_prep(req, sqe);
6012 case IORING_OP_OPENAT:
6013 return io_openat_prep(req, sqe);
6014 case IORING_OP_CLOSE:
6015 return io_close_prep(req, sqe);
6016 case IORING_OP_FILES_UPDATE:
6017 return io_files_update_prep(req, sqe);
6018 case IORING_OP_STATX:
6019 return io_statx_prep(req, sqe);
6020 case IORING_OP_FADVISE:
6021 return io_fadvise_prep(req, sqe);
6022 case IORING_OP_MADVISE:
6023 return io_madvise_prep(req, sqe);
6024 case IORING_OP_OPENAT2:
6025 return io_openat2_prep(req, sqe);
6026 case IORING_OP_EPOLL_CTL:
6027 return io_epoll_ctl_prep(req, sqe);
6028 case IORING_OP_SPLICE:
6029 return io_splice_prep(req, sqe);
6030 case IORING_OP_PROVIDE_BUFFERS:
6031 return io_provide_buffers_prep(req, sqe);
6032 case IORING_OP_REMOVE_BUFFERS:
6033 return io_remove_buffers_prep(req, sqe);
6035 return io_tee_prep(req, sqe);
6036 case IORING_OP_SHUTDOWN:
6037 return io_shutdown_prep(req, sqe);
6038 case IORING_OP_RENAMEAT:
6039 return io_renameat_prep(req, sqe);
6040 case IORING_OP_UNLINKAT:
6041 return io_unlinkat_prep(req, sqe);
6044 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
6049 static int io_req_defer_prep(struct io_kiocb *req,
6050 const struct io_uring_sqe *sqe)
6054 if (io_alloc_async_data(req))
6056 return io_req_prep(req, sqe);
6059 static u32 io_get_sequence(struct io_kiocb *req)
6061 struct io_kiocb *pos;
6062 struct io_ring_ctx *ctx = req->ctx;
6063 u32 total_submitted, nr_reqs = 0;
6065 io_for_each_link(pos, req)
6068 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
6069 return total_submitted - nr_reqs;
6072 static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe)
6074 struct io_ring_ctx *ctx = req->ctx;
6075 struct io_defer_entry *de;
6079 /* Still need defer if there is pending req in defer list. */
6080 if (likely(list_empty_careful(&ctx->defer_list) &&
6081 !(req->flags & REQ_F_IO_DRAIN)))
6084 seq = io_get_sequence(req);
6085 /* Still a chance to pass the sequence check */
6086 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
6089 if (!req->async_data) {
6090 ret = io_req_defer_prep(req, sqe);
6094 io_prep_async_link(req);
6095 de = kmalloc(sizeof(*de), GFP_KERNEL);
6099 spin_lock_irq(&ctx->completion_lock);
6100 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
6101 spin_unlock_irq(&ctx->completion_lock);
6103 io_queue_async_work(req);
6104 return -EIOCBQUEUED;
6107 trace_io_uring_defer(ctx, req, req->user_data);
6110 list_add_tail(&de->list, &ctx->defer_list);
6111 spin_unlock_irq(&ctx->completion_lock);
6112 return -EIOCBQUEUED;
6115 static void io_req_drop_files(struct io_kiocb *req)
6117 struct io_ring_ctx *ctx = req->ctx;
6118 struct io_uring_task *tctx = req->task->io_uring;
6119 unsigned long flags;
6121 spin_lock_irqsave(&ctx->inflight_lock, flags);
6122 list_del(&req->inflight_entry);
6123 if (atomic_read(&tctx->in_idle))
6124 wake_up(&tctx->wait);
6125 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
6126 req->flags &= ~REQ_F_INFLIGHT;
6127 put_files_struct(req->work.identity->files);
6128 put_nsproxy(req->work.identity->nsproxy);
6129 req->work.flags &= ~IO_WQ_WORK_FILES;
6132 static void __io_clean_op(struct io_kiocb *req)
6134 if (req->flags & REQ_F_BUFFER_SELECTED) {
6135 switch (req->opcode) {
6136 case IORING_OP_READV:
6137 case IORING_OP_READ_FIXED:
6138 case IORING_OP_READ:
6139 kfree((void *)(unsigned long)req->rw.addr);
6141 case IORING_OP_RECVMSG:
6142 case IORING_OP_RECV:
6143 kfree(req->sr_msg.kbuf);
6146 req->flags &= ~REQ_F_BUFFER_SELECTED;
6149 if (req->flags & REQ_F_NEED_CLEANUP) {
6150 switch (req->opcode) {
6151 case IORING_OP_READV:
6152 case IORING_OP_READ_FIXED:
6153 case IORING_OP_READ:
6154 case IORING_OP_WRITEV:
6155 case IORING_OP_WRITE_FIXED:
6156 case IORING_OP_WRITE: {
6157 struct io_async_rw *io = req->async_data;
6159 kfree(io->free_iovec);
6162 case IORING_OP_RECVMSG:
6163 case IORING_OP_SENDMSG: {
6164 struct io_async_msghdr *io = req->async_data;
6165 if (io->iov != io->fast_iov)
6169 case IORING_OP_SPLICE:
6171 io_put_file(req, req->splice.file_in,
6172 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
6174 case IORING_OP_OPENAT:
6175 case IORING_OP_OPENAT2:
6176 if (req->open.filename)
6177 putname(req->open.filename);
6179 case IORING_OP_RENAMEAT:
6180 putname(req->rename.oldpath);
6181 putname(req->rename.newpath);
6183 case IORING_OP_UNLINKAT:
6184 putname(req->unlink.filename);
6187 req->flags &= ~REQ_F_NEED_CLEANUP;
6190 if (req->flags & REQ_F_INFLIGHT)
6191 io_req_drop_files(req);
6194 static int io_issue_sqe(struct io_kiocb *req, bool force_nonblock,
6195 struct io_comp_state *cs)
6197 struct io_ring_ctx *ctx = req->ctx;
6200 switch (req->opcode) {
6202 ret = io_nop(req, cs);
6204 case IORING_OP_READV:
6205 case IORING_OP_READ_FIXED:
6206 case IORING_OP_READ:
6207 ret = io_read(req, force_nonblock, cs);
6209 case IORING_OP_WRITEV:
6210 case IORING_OP_WRITE_FIXED:
6211 case IORING_OP_WRITE:
6212 ret = io_write(req, force_nonblock, cs);
6214 case IORING_OP_FSYNC:
6215 ret = io_fsync(req, force_nonblock);
6217 case IORING_OP_POLL_ADD:
6218 ret = io_poll_add(req);
6220 case IORING_OP_POLL_REMOVE:
6221 ret = io_poll_remove(req);
6223 case IORING_OP_SYNC_FILE_RANGE:
6224 ret = io_sync_file_range(req, force_nonblock);
6226 case IORING_OP_SENDMSG:
6227 ret = io_sendmsg(req, force_nonblock, cs);
6229 case IORING_OP_SEND:
6230 ret = io_send(req, force_nonblock, cs);
6232 case IORING_OP_RECVMSG:
6233 ret = io_recvmsg(req, force_nonblock, cs);
6235 case IORING_OP_RECV:
6236 ret = io_recv(req, force_nonblock, cs);
6238 case IORING_OP_TIMEOUT:
6239 ret = io_timeout(req);
6241 case IORING_OP_TIMEOUT_REMOVE:
6242 ret = io_timeout_remove(req);
6244 case IORING_OP_ACCEPT:
6245 ret = io_accept(req, force_nonblock, cs);
6247 case IORING_OP_CONNECT:
6248 ret = io_connect(req, force_nonblock, cs);
6250 case IORING_OP_ASYNC_CANCEL:
6251 ret = io_async_cancel(req);
6253 case IORING_OP_FALLOCATE:
6254 ret = io_fallocate(req, force_nonblock);
6256 case IORING_OP_OPENAT:
6257 ret = io_openat(req, force_nonblock);
6259 case IORING_OP_CLOSE:
6260 ret = io_close(req, force_nonblock, cs);
6262 case IORING_OP_FILES_UPDATE:
6263 ret = io_files_update(req, force_nonblock, cs);
6265 case IORING_OP_STATX:
6266 ret = io_statx(req, force_nonblock);
6268 case IORING_OP_FADVISE:
6269 ret = io_fadvise(req, force_nonblock);
6271 case IORING_OP_MADVISE:
6272 ret = io_madvise(req, force_nonblock);
6274 case IORING_OP_OPENAT2:
6275 ret = io_openat2(req, force_nonblock);
6277 case IORING_OP_EPOLL_CTL:
6278 ret = io_epoll_ctl(req, force_nonblock, cs);
6280 case IORING_OP_SPLICE:
6281 ret = io_splice(req, force_nonblock);
6283 case IORING_OP_PROVIDE_BUFFERS:
6284 ret = io_provide_buffers(req, force_nonblock, cs);
6286 case IORING_OP_REMOVE_BUFFERS:
6287 ret = io_remove_buffers(req, force_nonblock, cs);
6290 ret = io_tee(req, force_nonblock);
6292 case IORING_OP_SHUTDOWN:
6293 ret = io_shutdown(req, force_nonblock);
6295 case IORING_OP_RENAMEAT:
6296 ret = io_renameat(req, force_nonblock);
6298 case IORING_OP_UNLINKAT:
6299 ret = io_unlinkat(req, force_nonblock);
6309 /* If the op doesn't have a file, we're not polling for it */
6310 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
6311 const bool in_async = io_wq_current_is_worker();
6313 /* workqueue context doesn't hold uring_lock, grab it now */
6315 mutex_lock(&ctx->uring_lock);
6317 io_iopoll_req_issued(req, in_async);
6320 mutex_unlock(&ctx->uring_lock);
6326 static struct io_wq_work *io_wq_submit_work(struct io_wq_work *work)
6328 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6329 struct io_kiocb *timeout;
6332 timeout = io_prep_linked_timeout(req);
6334 io_queue_linked_timeout(timeout);
6336 /* if NO_CANCEL is set, we must still run the work */
6337 if ((work->flags & (IO_WQ_WORK_CANCEL|IO_WQ_WORK_NO_CANCEL)) ==
6338 IO_WQ_WORK_CANCEL) {
6344 ret = io_issue_sqe(req, false, NULL);
6346 * We can get EAGAIN for polled IO even though we're
6347 * forcing a sync submission from here, since we can't
6348 * wait for request slots on the block side.
6358 * io_iopoll_complete() does not hold completion_lock to complete
6359 * polled io, so here for polled io, just mark it done and still let
6360 * io_iopoll_complete() complete it.
6362 if (req->ctx->flags & IORING_SETUP_IOPOLL) {
6363 struct kiocb *kiocb = &req->rw.kiocb;
6365 kiocb_done(kiocb, ret, NULL);
6367 req_set_fail_links(req);
6368 io_req_complete(req, ret);
6372 return io_steal_work(req);
6375 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
6378 struct fixed_file_table *table;
6380 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
6381 return table->files[index & IORING_FILE_TABLE_MASK];
6384 static struct file *io_file_get(struct io_submit_state *state,
6385 struct io_kiocb *req, int fd, bool fixed)
6387 struct io_ring_ctx *ctx = req->ctx;
6391 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
6393 fd = array_index_nospec(fd, ctx->nr_user_files);
6394 file = io_file_from_index(ctx, fd);
6395 io_set_resource_node(req);
6397 trace_io_uring_file_get(ctx, fd);
6398 file = __io_file_get(state, fd);
6404 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6406 struct io_timeout_data *data = container_of(timer,
6407 struct io_timeout_data, timer);
6408 struct io_kiocb *prev, *req = data->req;
6409 struct io_ring_ctx *ctx = req->ctx;
6410 unsigned long flags;
6412 spin_lock_irqsave(&ctx->completion_lock, flags);
6413 prev = req->timeout.head;
6414 req->timeout.head = NULL;
6417 * We don't expect the list to be empty, that will only happen if we
6418 * race with the completion of the linked work.
6420 if (prev && refcount_inc_not_zero(&prev->refs))
6421 io_remove_next_linked(prev);
6424 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6427 req_set_fail_links(prev);
6428 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6431 io_req_complete(req, -ETIME);
6433 return HRTIMER_NORESTART;
6436 static void __io_queue_linked_timeout(struct io_kiocb *req)
6439 * If the back reference is NULL, then our linked request finished
6440 * before we got a chance to setup the timer
6442 if (req->timeout.head) {
6443 struct io_timeout_data *data = req->async_data;
6445 data->timer.function = io_link_timeout_fn;
6446 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6451 static void io_queue_linked_timeout(struct io_kiocb *req)
6453 struct io_ring_ctx *ctx = req->ctx;
6455 spin_lock_irq(&ctx->completion_lock);
6456 __io_queue_linked_timeout(req);
6457 spin_unlock_irq(&ctx->completion_lock);
6459 /* drop submission reference */
6463 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6465 struct io_kiocb *nxt = req->link;
6467 if (!nxt || (req->flags & REQ_F_LINK_TIMEOUT) ||
6468 nxt->opcode != IORING_OP_LINK_TIMEOUT)
6471 nxt->timeout.head = req;
6472 nxt->flags |= REQ_F_LTIMEOUT_ACTIVE;
6473 req->flags |= REQ_F_LINK_TIMEOUT;
6477 static void __io_queue_sqe(struct io_kiocb *req, struct io_comp_state *cs)
6479 struct io_kiocb *linked_timeout;
6480 const struct cred *old_creds = NULL;
6484 linked_timeout = io_prep_linked_timeout(req);
6486 if ((req->flags & REQ_F_WORK_INITIALIZED) &&
6487 (req->work.flags & IO_WQ_WORK_CREDS) &&
6488 req->work.identity->creds != current_cred()) {
6490 revert_creds(old_creds);
6491 if (old_creds == req->work.identity->creds)
6492 old_creds = NULL; /* restored original creds */
6494 old_creds = override_creds(req->work.identity->creds);
6497 ret = io_issue_sqe(req, true, cs);
6500 * We async punt it if the file wasn't marked NOWAIT, or if the file
6501 * doesn't support non-blocking read/write attempts
6503 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6504 if (!io_arm_poll_handler(req)) {
6506 * Queued up for async execution, worker will release
6507 * submit reference when the iocb is actually submitted.
6509 io_queue_async_work(req);
6513 io_queue_linked_timeout(linked_timeout);
6514 } else if (likely(!ret)) {
6515 /* drop submission reference */
6516 req = io_put_req_find_next(req);
6518 io_queue_linked_timeout(linked_timeout);
6521 if (!(req->flags & REQ_F_FORCE_ASYNC))
6523 io_queue_async_work(req);
6526 /* un-prep timeout, so it'll be killed as any other linked */
6527 req->flags &= ~REQ_F_LINK_TIMEOUT;
6528 req_set_fail_links(req);
6530 io_req_complete(req, ret);
6534 revert_creds(old_creds);
6537 static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6538 struct io_comp_state *cs)
6542 ret = io_req_defer(req, sqe);
6544 if (ret != -EIOCBQUEUED) {
6546 req_set_fail_links(req);
6548 io_req_complete(req, ret);
6550 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6551 if (!req->async_data) {
6552 ret = io_req_defer_prep(req, sqe);
6556 io_queue_async_work(req);
6559 ret = io_req_prep(req, sqe);
6563 __io_queue_sqe(req, cs);
6567 static inline void io_queue_link_head(struct io_kiocb *req,
6568 struct io_comp_state *cs)
6570 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
6572 io_req_complete(req, -ECANCELED);
6574 io_queue_sqe(req, NULL, cs);
6577 struct io_submit_link {
6578 struct io_kiocb *head;
6579 struct io_kiocb *last;
6582 static int io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6583 struct io_submit_link *link, struct io_comp_state *cs)
6585 struct io_ring_ctx *ctx = req->ctx;
6589 * If we already have a head request, queue this one for async
6590 * submittal once the head completes. If we don't have a head but
6591 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6592 * submitted sync once the chain is complete. If none of those
6593 * conditions are true (normal request), then just queue it.
6596 struct io_kiocb *head = link->head;
6599 * Taking sequential execution of a link, draining both sides
6600 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6601 * requests in the link. So, it drains the head and the
6602 * next after the link request. The last one is done via
6603 * drain_next flag to persist the effect across calls.
6605 if (req->flags & REQ_F_IO_DRAIN) {
6606 head->flags |= REQ_F_IO_DRAIN;
6607 ctx->drain_next = 1;
6609 ret = io_req_defer_prep(req, sqe);
6610 if (unlikely(ret)) {
6611 /* fail even hard links since we don't submit */
6612 head->flags |= REQ_F_FAIL_LINK;
6615 trace_io_uring_link(ctx, req, head);
6616 link->last->link = req;
6619 /* last request of a link, enqueue the link */
6620 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6621 io_queue_link_head(head, cs);
6625 if (unlikely(ctx->drain_next)) {
6626 req->flags |= REQ_F_IO_DRAIN;
6627 ctx->drain_next = 0;
6629 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6630 ret = io_req_defer_prep(req, sqe);
6632 req->flags |= REQ_F_FAIL_LINK;
6636 io_queue_sqe(req, sqe, cs);
6644 * Batched submission is done, ensure local IO is flushed out.
6646 static void io_submit_state_end(struct io_submit_state *state)
6648 if (!list_empty(&state->comp.list))
6649 io_submit_flush_completions(&state->comp);
6650 if (state->plug_started)
6651 blk_finish_plug(&state->plug);
6652 io_state_file_put(state);
6653 if (state->free_reqs)
6654 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
6658 * Start submission side cache.
6660 static void io_submit_state_start(struct io_submit_state *state,
6661 struct io_ring_ctx *ctx, unsigned int max_ios)
6663 state->plug_started = false;
6665 INIT_LIST_HEAD(&state->comp.list);
6666 state->comp.ctx = ctx;
6667 state->free_reqs = 0;
6668 state->file_refs = 0;
6669 state->ios_left = max_ios;
6672 static void io_commit_sqring(struct io_ring_ctx *ctx)
6674 struct io_rings *rings = ctx->rings;
6677 * Ensure any loads from the SQEs are done at this point,
6678 * since once we write the new head, the application could
6679 * write new data to them.
6681 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6685 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6686 * that is mapped by userspace. This means that care needs to be taken to
6687 * ensure that reads are stable, as we cannot rely on userspace always
6688 * being a good citizen. If members of the sqe are validated and then later
6689 * used, it's important that those reads are done through READ_ONCE() to
6690 * prevent a re-load down the line.
6692 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6694 u32 *sq_array = ctx->sq_array;
6698 * The cached sq head (or cq tail) serves two purposes:
6700 * 1) allows us to batch the cost of updating the user visible
6702 * 2) allows the kernel side to track the head on its own, even
6703 * though the application is the one updating it.
6705 head = READ_ONCE(sq_array[ctx->cached_sq_head & ctx->sq_mask]);
6706 if (likely(head < ctx->sq_entries))
6707 return &ctx->sq_sqes[head];
6709 /* drop invalid entries */
6710 ctx->cached_sq_dropped++;
6711 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6715 static inline void io_consume_sqe(struct io_ring_ctx *ctx)
6717 ctx->cached_sq_head++;
6721 * Check SQE restrictions (opcode and flags).
6723 * Returns 'true' if SQE is allowed, 'false' otherwise.
6725 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6726 struct io_kiocb *req,
6727 unsigned int sqe_flags)
6729 if (!ctx->restricted)
6732 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6735 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6736 ctx->restrictions.sqe_flags_required)
6739 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6740 ctx->restrictions.sqe_flags_required))
6746 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
6747 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
6748 IOSQE_BUFFER_SELECT)
6750 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6751 const struct io_uring_sqe *sqe,
6752 struct io_submit_state *state)
6754 unsigned int sqe_flags;
6757 req->opcode = READ_ONCE(sqe->opcode);
6758 req->user_data = READ_ONCE(sqe->user_data);
6759 req->async_data = NULL;
6764 req->fixed_file_refs = NULL;
6765 /* one is dropped after submission, the other at completion */
6766 refcount_set(&req->refs, 2);
6767 req->task = current;
6770 if (unlikely(req->opcode >= IORING_OP_LAST))
6773 if (unlikely(io_sq_thread_acquire_mm_files(ctx, req)))
6776 sqe_flags = READ_ONCE(sqe->flags);
6777 /* enforce forwards compatibility on users */
6778 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6781 if (unlikely(!io_check_restriction(ctx, req, sqe_flags)))
6784 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6785 !io_op_defs[req->opcode].buffer_select)
6788 id = READ_ONCE(sqe->personality);
6790 struct io_identity *iod;
6792 iod = idr_find(&ctx->personality_idr, id);
6795 refcount_inc(&iod->count);
6797 __io_req_init_async(req);
6798 get_cred(iod->creds);
6799 req->work.identity = iod;
6800 req->work.flags |= IO_WQ_WORK_CREDS;
6803 /* same numerical values with corresponding REQ_F_*, safe to copy */
6804 req->flags |= sqe_flags;
6807 * Plug now if we have more than 1 IO left after this, and the target
6808 * is potentially a read/write to block based storage.
6810 if (!state->plug_started && state->ios_left > 1 &&
6811 io_op_defs[req->opcode].plug) {
6812 blk_start_plug(&state->plug);
6813 state->plug_started = true;
6817 if (io_op_defs[req->opcode].needs_file) {
6818 bool fixed = req->flags & REQ_F_FIXED_FILE;
6820 req->file = io_file_get(state, req, READ_ONCE(sqe->fd), fixed);
6821 if (unlikely(!req->file &&
6822 !io_op_defs[req->opcode].needs_file_no_error))
6830 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6832 struct io_submit_state state;
6833 struct io_submit_link link;
6834 int i, submitted = 0;
6836 /* if we have a backlog and couldn't flush it all, return BUSY */
6837 if (test_bit(0, &ctx->sq_check_overflow)) {
6838 if (!list_empty(&ctx->cq_overflow_list) &&
6839 !io_cqring_overflow_flush(ctx, false, NULL, NULL))
6843 /* make sure SQ entry isn't read before tail */
6844 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6846 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6849 percpu_counter_add(¤t->io_uring->inflight, nr);
6850 refcount_add(nr, ¤t->usage);
6852 io_submit_state_start(&state, ctx, nr);
6855 for (i = 0; i < nr; i++) {
6856 const struct io_uring_sqe *sqe;
6857 struct io_kiocb *req;
6860 sqe = io_get_sqe(ctx);
6861 if (unlikely(!sqe)) {
6862 io_consume_sqe(ctx);
6865 req = io_alloc_req(ctx, &state);
6866 if (unlikely(!req)) {
6868 submitted = -EAGAIN;
6871 io_consume_sqe(ctx);
6872 /* will complete beyond this point, count as submitted */
6875 err = io_init_req(ctx, req, sqe, &state);
6876 if (unlikely(err)) {
6879 io_req_complete(req, err);
6883 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6884 true, io_async_submit(ctx));
6885 err = io_submit_sqe(req, sqe, &link, &state.comp);
6890 if (unlikely(submitted != nr)) {
6891 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6892 struct io_uring_task *tctx = current->io_uring;
6893 int unused = nr - ref_used;
6895 percpu_ref_put_many(&ctx->refs, unused);
6896 percpu_counter_sub(&tctx->inflight, unused);
6897 put_task_struct_many(current, unused);
6900 io_queue_link_head(link.head, &state.comp);
6901 io_submit_state_end(&state);
6903 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6904 io_commit_sqring(ctx);
6909 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6911 /* Tell userspace we may need a wakeup call */
6912 spin_lock_irq(&ctx->completion_lock);
6913 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6914 spin_unlock_irq(&ctx->completion_lock);
6917 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6919 spin_lock_irq(&ctx->completion_lock);
6920 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6921 spin_unlock_irq(&ctx->completion_lock);
6924 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
6926 unsigned int to_submit;
6929 to_submit = io_sqring_entries(ctx);
6930 /* if we're handling multiple rings, cap submit size for fairness */
6931 if (cap_entries && to_submit > 8)
6934 if (!list_empty(&ctx->iopoll_list) || to_submit) {
6935 unsigned nr_events = 0;
6937 mutex_lock(&ctx->uring_lock);
6938 if (!list_empty(&ctx->iopoll_list))
6939 io_do_iopoll(ctx, &nr_events, 0);
6941 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)))
6942 ret = io_submit_sqes(ctx, to_submit);
6943 mutex_unlock(&ctx->uring_lock);
6946 if (!io_sqring_full(ctx) && wq_has_sleeper(&ctx->sqo_sq_wait))
6947 wake_up(&ctx->sqo_sq_wait);
6952 static void io_sqd_update_thread_idle(struct io_sq_data *sqd)
6954 struct io_ring_ctx *ctx;
6955 unsigned sq_thread_idle = 0;
6957 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6958 if (sq_thread_idle < ctx->sq_thread_idle)
6959 sq_thread_idle = ctx->sq_thread_idle;
6962 sqd->sq_thread_idle = sq_thread_idle;
6965 static void io_sqd_init_new(struct io_sq_data *sqd)
6967 struct io_ring_ctx *ctx;
6969 while (!list_empty(&sqd->ctx_new_list)) {
6970 ctx = list_first_entry(&sqd->ctx_new_list, struct io_ring_ctx, sqd_list);
6971 list_move_tail(&ctx->sqd_list, &sqd->ctx_list);
6972 complete(&ctx->sq_thread_comp);
6975 io_sqd_update_thread_idle(sqd);
6978 static int io_sq_thread(void *data)
6980 struct cgroup_subsys_state *cur_css = NULL;
6981 struct files_struct *old_files = current->files;
6982 struct nsproxy *old_nsproxy = current->nsproxy;
6983 const struct cred *old_cred = NULL;
6984 struct io_sq_data *sqd = data;
6985 struct io_ring_ctx *ctx;
6986 unsigned long timeout = 0;
6990 current->files = NULL;
6991 current->nsproxy = NULL;
6992 task_unlock(current);
6994 while (!kthread_should_stop()) {
6996 bool cap_entries, sqt_spin, needs_sched;
6999 * Any changes to the sqd lists are synchronized through the
7000 * kthread parking. This synchronizes the thread vs users,
7001 * the users are synchronized on the sqd->ctx_lock.
7003 if (kthread_should_park()) {
7006 * When sq thread is unparked, in case the previous park operation
7007 * comes from io_put_sq_data(), which means that sq thread is going
7008 * to be stopped, so here needs to have a check.
7010 if (kthread_should_stop())
7014 if (unlikely(!list_empty(&sqd->ctx_new_list))) {
7015 io_sqd_init_new(sqd);
7016 timeout = jiffies + sqd->sq_thread_idle;
7020 cap_entries = !list_is_singular(&sqd->ctx_list);
7021 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
7022 if (current->cred != ctx->creds) {
7024 revert_creds(old_cred);
7025 old_cred = override_creds(ctx->creds);
7027 io_sq_thread_associate_blkcg(ctx, &cur_css);
7029 current->loginuid = ctx->loginuid;
7030 current->sessionid = ctx->sessionid;
7033 ret = __io_sq_thread(ctx, cap_entries);
7034 if (!sqt_spin && (ret > 0 || !list_empty(&ctx->iopoll_list)))
7037 io_sq_thread_drop_mm_files();
7040 if (sqt_spin || !time_after(jiffies, timeout)) {
7044 timeout = jiffies + sqd->sq_thread_idle;
7048 if (kthread_should_park())
7052 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
7053 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
7054 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
7055 !list_empty_careful(&ctx->iopoll_list)) {
7056 needs_sched = false;
7059 if (io_sqring_entries(ctx)) {
7060 needs_sched = false;
7066 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7067 io_ring_set_wakeup_flag(ctx);
7070 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
7071 io_ring_clear_wakeup_flag(ctx);
7074 finish_wait(&sqd->wait, &wait);
7075 timeout = jiffies + sqd->sq_thread_idle;
7081 io_sq_thread_unassociate_blkcg();
7083 revert_creds(old_cred);
7086 current->files = old_files;
7087 current->nsproxy = old_nsproxy;
7088 task_unlock(current);
7095 struct io_wait_queue {
7096 struct wait_queue_entry wq;
7097 struct io_ring_ctx *ctx;
7099 unsigned nr_timeouts;
7102 static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush)
7104 struct io_ring_ctx *ctx = iowq->ctx;
7107 * Wake up if we have enough events, or if a timeout occurred since we
7108 * started waiting. For timeouts, we always want to return to userspace,
7109 * regardless of event count.
7111 return io_cqring_events(ctx, noflush) >= iowq->to_wait ||
7112 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
7115 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
7116 int wake_flags, void *key)
7118 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
7121 /* use noflush == true, as we can't safely rely on locking context */
7122 if (!io_should_wake(iowq, true))
7125 return autoremove_wake_function(curr, mode, wake_flags, key);
7128 static int io_run_task_work_sig(void)
7130 if (io_run_task_work())
7132 if (!signal_pending(current))
7134 if (current->jobctl & JOBCTL_TASK_WORK) {
7135 spin_lock_irq(¤t->sighand->siglock);
7136 current->jobctl &= ~JOBCTL_TASK_WORK;
7137 recalc_sigpending();
7138 spin_unlock_irq(¤t->sighand->siglock);
7145 * Wait until events become available, if we don't already have some. The
7146 * application must reap them itself, as they reside on the shared cq ring.
7148 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
7149 const sigset_t __user *sig, size_t sigsz,
7150 struct __kernel_timespec __user *uts)
7152 struct io_wait_queue iowq = {
7155 .func = io_wake_function,
7156 .entry = LIST_HEAD_INIT(iowq.wq.entry),
7159 .to_wait = min_events,
7161 struct io_rings *rings = ctx->rings;
7162 struct timespec64 ts;
7163 signed long timeout = 0;
7167 if (io_cqring_events(ctx, false) >= min_events)
7169 if (!io_run_task_work())
7174 #ifdef CONFIG_COMPAT
7175 if (in_compat_syscall())
7176 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
7180 ret = set_user_sigmask(sig, sigsz);
7187 if (get_timespec64(&ts, uts))
7189 timeout = timespec64_to_jiffies(&ts);
7192 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
7193 trace_io_uring_cqring_wait(ctx, min_events);
7195 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
7196 TASK_INTERRUPTIBLE);
7197 /* make sure we run task_work before checking for signals */
7198 ret = io_run_task_work_sig();
7203 if (io_should_wake(&iowq, false))
7206 timeout = schedule_timeout(timeout);
7215 finish_wait(&ctx->wait, &iowq.wq);
7217 restore_saved_sigmask_unless(ret == -EINTR);
7219 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
7222 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
7224 #if defined(CONFIG_UNIX)
7225 if (ctx->ring_sock) {
7226 struct sock *sock = ctx->ring_sock->sk;
7227 struct sk_buff *skb;
7229 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
7235 for (i = 0; i < ctx->nr_user_files; i++) {
7238 file = io_file_from_index(ctx, i);
7245 static void io_file_ref_kill(struct percpu_ref *ref)
7247 struct fixed_file_data *data;
7249 data = container_of(ref, struct fixed_file_data, refs);
7250 complete(&data->done);
7253 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
7255 struct fixed_file_data *data = ctx->file_data;
7256 struct fixed_file_ref_node *ref_node = NULL;
7257 unsigned nr_tables, i;
7262 spin_lock_bh(&data->lock);
7263 ref_node = data->node;
7264 spin_unlock_bh(&data->lock);
7266 percpu_ref_kill(&ref_node->refs);
7268 percpu_ref_kill(&data->refs);
7270 /* wait for all refs nodes to complete */
7271 flush_delayed_work(&ctx->file_put_work);
7272 wait_for_completion(&data->done);
7274 __io_sqe_files_unregister(ctx);
7275 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
7276 for (i = 0; i < nr_tables; i++)
7277 kfree(data->table[i].files);
7279 percpu_ref_exit(&data->refs);
7281 ctx->file_data = NULL;
7282 ctx->nr_user_files = 0;
7286 static void io_put_sq_data(struct io_sq_data *sqd)
7288 if (refcount_dec_and_test(&sqd->refs)) {
7290 * The park is a bit of a work-around, without it we get
7291 * warning spews on shutdown with SQPOLL set and affinity
7292 * set to a single CPU.
7295 kthread_park(sqd->thread);
7296 kthread_stop(sqd->thread);
7303 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
7305 struct io_ring_ctx *ctx_attach;
7306 struct io_sq_data *sqd;
7309 f = fdget(p->wq_fd);
7311 return ERR_PTR(-ENXIO);
7312 if (f.file->f_op != &io_uring_fops) {
7314 return ERR_PTR(-EINVAL);
7317 ctx_attach = f.file->private_data;
7318 sqd = ctx_attach->sq_data;
7321 return ERR_PTR(-EINVAL);
7324 refcount_inc(&sqd->refs);
7329 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p)
7331 struct io_sq_data *sqd;
7333 if (p->flags & IORING_SETUP_ATTACH_WQ)
7334 return io_attach_sq_data(p);
7336 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
7338 return ERR_PTR(-ENOMEM);
7340 refcount_set(&sqd->refs, 1);
7341 INIT_LIST_HEAD(&sqd->ctx_list);
7342 INIT_LIST_HEAD(&sqd->ctx_new_list);
7343 mutex_init(&sqd->ctx_lock);
7344 mutex_init(&sqd->lock);
7345 init_waitqueue_head(&sqd->wait);
7349 static void io_sq_thread_unpark(struct io_sq_data *sqd)
7350 __releases(&sqd->lock)
7354 kthread_unpark(sqd->thread);
7355 mutex_unlock(&sqd->lock);
7358 static void io_sq_thread_park(struct io_sq_data *sqd)
7359 __acquires(&sqd->lock)
7363 mutex_lock(&sqd->lock);
7364 kthread_park(sqd->thread);
7367 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
7369 struct io_sq_data *sqd = ctx->sq_data;
7374 * We may arrive here from the error branch in
7375 * io_sq_offload_create() where the kthread is created
7376 * without being waked up, thus wake it up now to make
7377 * sure the wait will complete.
7379 wake_up_process(sqd->thread);
7380 wait_for_completion(&ctx->sq_thread_comp);
7382 io_sq_thread_park(sqd);
7385 mutex_lock(&sqd->ctx_lock);
7386 list_del(&ctx->sqd_list);
7387 io_sqd_update_thread_idle(sqd);
7388 mutex_unlock(&sqd->ctx_lock);
7391 io_sq_thread_unpark(sqd);
7393 io_put_sq_data(sqd);
7394 ctx->sq_data = NULL;
7398 static void io_finish_async(struct io_ring_ctx *ctx)
7400 io_sq_thread_stop(ctx);
7403 io_wq_destroy(ctx->io_wq);
7408 #if defined(CONFIG_UNIX)
7410 * Ensure the UNIX gc is aware of our file set, so we are certain that
7411 * the io_uring can be safely unregistered on process exit, even if we have
7412 * loops in the file referencing.
7414 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7416 struct sock *sk = ctx->ring_sock->sk;
7417 struct scm_fp_list *fpl;
7418 struct sk_buff *skb;
7421 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7425 skb = alloc_skb(0, GFP_KERNEL);
7434 fpl->user = get_uid(ctx->user);
7435 for (i = 0; i < nr; i++) {
7436 struct file *file = io_file_from_index(ctx, i + offset);
7440 fpl->fp[nr_files] = get_file(file);
7441 unix_inflight(fpl->user, fpl->fp[nr_files]);
7446 fpl->max = SCM_MAX_FD;
7447 fpl->count = nr_files;
7448 UNIXCB(skb).fp = fpl;
7449 skb->destructor = unix_destruct_scm;
7450 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7451 skb_queue_head(&sk->sk_receive_queue, skb);
7453 for (i = 0; i < nr_files; i++)
7464 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7465 * causes regular reference counting to break down. We rely on the UNIX
7466 * garbage collection to take care of this problem for us.
7468 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7470 unsigned left, total;
7474 left = ctx->nr_user_files;
7476 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7478 ret = __io_sqe_files_scm(ctx, this_files, total);
7482 total += this_files;
7488 while (total < ctx->nr_user_files) {
7489 struct file *file = io_file_from_index(ctx, total);
7499 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7505 static int io_sqe_alloc_file_tables(struct fixed_file_data *file_data,
7506 unsigned nr_tables, unsigned nr_files)
7510 for (i = 0; i < nr_tables; i++) {
7511 struct fixed_file_table *table = &file_data->table[i];
7512 unsigned this_files;
7514 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
7515 table->files = kcalloc(this_files, sizeof(struct file *),
7519 nr_files -= this_files;
7525 for (i = 0; i < nr_tables; i++) {
7526 struct fixed_file_table *table = &file_data->table[i];
7527 kfree(table->files);
7532 static void io_ring_file_put(struct io_ring_ctx *ctx, struct file *file)
7534 #if defined(CONFIG_UNIX)
7535 struct sock *sock = ctx->ring_sock->sk;
7536 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7537 struct sk_buff *skb;
7540 __skb_queue_head_init(&list);
7543 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7544 * remove this entry and rearrange the file array.
7546 skb = skb_dequeue(head);
7548 struct scm_fp_list *fp;
7550 fp = UNIXCB(skb).fp;
7551 for (i = 0; i < fp->count; i++) {
7554 if (fp->fp[i] != file)
7557 unix_notinflight(fp->user, fp->fp[i]);
7558 left = fp->count - 1 - i;
7560 memmove(&fp->fp[i], &fp->fp[i + 1],
7561 left * sizeof(struct file *));
7568 __skb_queue_tail(&list, skb);
7578 __skb_queue_tail(&list, skb);
7580 skb = skb_dequeue(head);
7583 if (skb_peek(&list)) {
7584 spin_lock_irq(&head->lock);
7585 while ((skb = __skb_dequeue(&list)) != NULL)
7586 __skb_queue_tail(head, skb);
7587 spin_unlock_irq(&head->lock);
7594 struct io_file_put {
7595 struct list_head list;
7599 static void __io_file_put_work(struct fixed_file_ref_node *ref_node)
7601 struct fixed_file_data *file_data = ref_node->file_data;
7602 struct io_ring_ctx *ctx = file_data->ctx;
7603 struct io_file_put *pfile, *tmp;
7605 list_for_each_entry_safe(pfile, tmp, &ref_node->file_list, list) {
7606 list_del(&pfile->list);
7607 io_ring_file_put(ctx, pfile->file);
7611 percpu_ref_exit(&ref_node->refs);
7613 percpu_ref_put(&file_data->refs);
7616 static void io_file_put_work(struct work_struct *work)
7618 struct io_ring_ctx *ctx;
7619 struct llist_node *node;
7621 ctx = container_of(work, struct io_ring_ctx, file_put_work.work);
7622 node = llist_del_all(&ctx->file_put_llist);
7625 struct fixed_file_ref_node *ref_node;
7626 struct llist_node *next = node->next;
7628 ref_node = llist_entry(node, struct fixed_file_ref_node, llist);
7629 __io_file_put_work(ref_node);
7634 static void io_file_data_ref_zero(struct percpu_ref *ref)
7636 struct fixed_file_ref_node *ref_node;
7637 struct fixed_file_data *data;
7638 struct io_ring_ctx *ctx;
7639 bool first_add = false;
7642 ref_node = container_of(ref, struct fixed_file_ref_node, refs);
7643 data = ref_node->file_data;
7646 spin_lock_bh(&data->lock);
7647 ref_node->done = true;
7649 while (!list_empty(&data->ref_list)) {
7650 ref_node = list_first_entry(&data->ref_list,
7651 struct fixed_file_ref_node, node);
7652 /* recycle ref nodes in order */
7653 if (!ref_node->done)
7655 list_del(&ref_node->node);
7656 first_add |= llist_add(&ref_node->llist, &ctx->file_put_llist);
7658 spin_unlock_bh(&data->lock);
7660 if (percpu_ref_is_dying(&data->refs))
7664 mod_delayed_work(system_wq, &ctx->file_put_work, 0);
7666 queue_delayed_work(system_wq, &ctx->file_put_work, delay);
7669 static struct fixed_file_ref_node *alloc_fixed_file_ref_node(
7670 struct io_ring_ctx *ctx)
7672 struct fixed_file_ref_node *ref_node;
7674 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7676 return ERR_PTR(-ENOMEM);
7678 if (percpu_ref_init(&ref_node->refs, io_file_data_ref_zero,
7681 return ERR_PTR(-ENOMEM);
7683 INIT_LIST_HEAD(&ref_node->node);
7684 INIT_LIST_HEAD(&ref_node->file_list);
7685 ref_node->file_data = ctx->file_data;
7686 ref_node->done = false;
7690 static void destroy_fixed_file_ref_node(struct fixed_file_ref_node *ref_node)
7692 percpu_ref_exit(&ref_node->refs);
7696 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7699 __s32 __user *fds = (__s32 __user *) arg;
7700 unsigned nr_tables, i;
7702 int fd, ret = -ENOMEM;
7703 struct fixed_file_ref_node *ref_node;
7704 struct fixed_file_data *file_data;
7710 if (nr_args > IORING_MAX_FIXED_FILES)
7713 file_data = kzalloc(sizeof(*ctx->file_data), GFP_KERNEL);
7716 file_data->ctx = ctx;
7717 init_completion(&file_data->done);
7718 INIT_LIST_HEAD(&file_data->ref_list);
7719 spin_lock_init(&file_data->lock);
7721 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
7722 file_data->table = kcalloc(nr_tables, sizeof(*file_data->table),
7724 if (!file_data->table)
7727 if (percpu_ref_init(&file_data->refs, io_file_ref_kill,
7728 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
7731 if (io_sqe_alloc_file_tables(file_data, nr_tables, nr_args))
7733 ctx->file_data = file_data;
7735 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7736 struct fixed_file_table *table;
7739 if (copy_from_user(&fd, &fds[i], sizeof(fd))) {
7743 /* allow sparse sets */
7753 * Don't allow io_uring instances to be registered. If UNIX
7754 * isn't enabled, then this causes a reference cycle and this
7755 * instance can never get freed. If UNIX is enabled we'll
7756 * handle it just fine, but there's still no point in allowing
7757 * a ring fd as it doesn't support regular read/write anyway.
7759 if (file->f_op == &io_uring_fops) {
7763 table = &file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7764 index = i & IORING_FILE_TABLE_MASK;
7765 table->files[index] = file;
7768 ret = io_sqe_files_scm(ctx);
7770 io_sqe_files_unregister(ctx);
7774 ref_node = alloc_fixed_file_ref_node(ctx);
7775 if (IS_ERR(ref_node)) {
7776 io_sqe_files_unregister(ctx);
7777 return PTR_ERR(ref_node);
7780 file_data->node = ref_node;
7781 spin_lock_bh(&file_data->lock);
7782 list_add_tail(&ref_node->node, &file_data->ref_list);
7783 spin_unlock_bh(&file_data->lock);
7784 percpu_ref_get(&file_data->refs);
7787 for (i = 0; i < ctx->nr_user_files; i++) {
7788 file = io_file_from_index(ctx, i);
7792 for (i = 0; i < nr_tables; i++)
7793 kfree(file_data->table[i].files);
7794 ctx->nr_user_files = 0;
7796 percpu_ref_exit(&file_data->refs);
7798 kfree(file_data->table);
7800 ctx->file_data = NULL;
7804 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7807 #if defined(CONFIG_UNIX)
7808 struct sock *sock = ctx->ring_sock->sk;
7809 struct sk_buff_head *head = &sock->sk_receive_queue;
7810 struct sk_buff *skb;
7813 * See if we can merge this file into an existing skb SCM_RIGHTS
7814 * file set. If there's no room, fall back to allocating a new skb
7815 * and filling it in.
7817 spin_lock_irq(&head->lock);
7818 skb = skb_peek(head);
7820 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7822 if (fpl->count < SCM_MAX_FD) {
7823 __skb_unlink(skb, head);
7824 spin_unlock_irq(&head->lock);
7825 fpl->fp[fpl->count] = get_file(file);
7826 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7828 spin_lock_irq(&head->lock);
7829 __skb_queue_head(head, skb);
7834 spin_unlock_irq(&head->lock);
7841 return __io_sqe_files_scm(ctx, 1, index);
7847 static int io_queue_file_removal(struct fixed_file_data *data,
7850 struct io_file_put *pfile;
7851 struct fixed_file_ref_node *ref_node = data->node;
7853 pfile = kzalloc(sizeof(*pfile), GFP_KERNEL);
7858 list_add(&pfile->list, &ref_node->file_list);
7863 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7864 struct io_uring_files_update *up,
7867 struct fixed_file_data *data = ctx->file_data;
7868 struct fixed_file_ref_node *ref_node;
7873 bool needs_switch = false;
7875 if (check_add_overflow(up->offset, nr_args, &done))
7877 if (done > ctx->nr_user_files)
7880 ref_node = alloc_fixed_file_ref_node(ctx);
7881 if (IS_ERR(ref_node))
7882 return PTR_ERR(ref_node);
7885 fds = u64_to_user_ptr(up->fds);
7887 struct fixed_file_table *table;
7891 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
7895 i = array_index_nospec(up->offset, ctx->nr_user_files);
7896 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7897 index = i & IORING_FILE_TABLE_MASK;
7898 if (table->files[index]) {
7899 file = table->files[index];
7900 err = io_queue_file_removal(data, file);
7903 table->files[index] = NULL;
7904 needs_switch = true;
7913 * Don't allow io_uring instances to be registered. If
7914 * UNIX isn't enabled, then this causes a reference
7915 * cycle and this instance can never get freed. If UNIX
7916 * is enabled we'll handle it just fine, but there's
7917 * still no point in allowing a ring fd as it doesn't
7918 * support regular read/write anyway.
7920 if (file->f_op == &io_uring_fops) {
7925 table->files[index] = file;
7926 err = io_sqe_file_register(ctx, file, i);
7928 table->files[index] = NULL;
7939 percpu_ref_kill(&data->node->refs);
7940 spin_lock_bh(&data->lock);
7941 list_add_tail(&ref_node->node, &data->ref_list);
7942 data->node = ref_node;
7943 spin_unlock_bh(&data->lock);
7944 percpu_ref_get(&ctx->file_data->refs);
7946 destroy_fixed_file_ref_node(ref_node);
7948 return done ? done : err;
7951 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
7954 struct io_uring_files_update up;
7956 if (!ctx->file_data)
7960 if (copy_from_user(&up, arg, sizeof(up)))
7965 return __io_sqe_files_update(ctx, &up, nr_args);
7968 static void io_free_work(struct io_wq_work *work)
7970 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7972 /* Consider that io_steal_work() relies on this ref */
7976 static int io_init_wq_offload(struct io_ring_ctx *ctx,
7977 struct io_uring_params *p)
7979 struct io_wq_data data;
7981 struct io_ring_ctx *ctx_attach;
7982 unsigned int concurrency;
7985 data.user = ctx->user;
7986 data.free_work = io_free_work;
7987 data.do_work = io_wq_submit_work;
7989 if (!(p->flags & IORING_SETUP_ATTACH_WQ)) {
7990 /* Do QD, or 4 * CPUS, whatever is smallest */
7991 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7993 ctx->io_wq = io_wq_create(concurrency, &data);
7994 if (IS_ERR(ctx->io_wq)) {
7995 ret = PTR_ERR(ctx->io_wq);
8001 f = fdget(p->wq_fd);
8005 if (f.file->f_op != &io_uring_fops) {
8010 ctx_attach = f.file->private_data;
8011 /* @io_wq is protected by holding the fd */
8012 if (!io_wq_get(ctx_attach->io_wq, &data)) {
8017 ctx->io_wq = ctx_attach->io_wq;
8023 static int io_uring_alloc_task_context(struct task_struct *task)
8025 struct io_uring_task *tctx;
8028 tctx = kmalloc(sizeof(*tctx), GFP_KERNEL);
8029 if (unlikely(!tctx))
8032 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
8033 if (unlikely(ret)) {
8039 init_waitqueue_head(&tctx->wait);
8041 atomic_set(&tctx->in_idle, 0);
8042 tctx->sqpoll = false;
8043 io_init_identity(&tctx->__identity);
8044 tctx->identity = &tctx->__identity;
8045 task->io_uring = tctx;
8049 void __io_uring_free(struct task_struct *tsk)
8051 struct io_uring_task *tctx = tsk->io_uring;
8053 WARN_ON_ONCE(!xa_empty(&tctx->xa));
8054 WARN_ON_ONCE(refcount_read(&tctx->identity->count) != 1);
8055 if (tctx->identity != &tctx->__identity)
8056 kfree(tctx->identity);
8057 percpu_counter_destroy(&tctx->inflight);
8059 tsk->io_uring = NULL;
8062 static int io_sq_offload_create(struct io_ring_ctx *ctx,
8063 struct io_uring_params *p)
8067 if (ctx->flags & IORING_SETUP_SQPOLL) {
8068 struct io_sq_data *sqd;
8071 if (!capable(CAP_SYS_ADMIN) && !capable(CAP_SYS_NICE))
8074 sqd = io_get_sq_data(p);
8081 io_sq_thread_park(sqd);
8082 mutex_lock(&sqd->ctx_lock);
8083 list_add(&ctx->sqd_list, &sqd->ctx_new_list);
8084 mutex_unlock(&sqd->ctx_lock);
8085 io_sq_thread_unpark(sqd);
8087 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
8088 if (!ctx->sq_thread_idle)
8089 ctx->sq_thread_idle = HZ;
8094 if (p->flags & IORING_SETUP_SQ_AFF) {
8095 int cpu = p->sq_thread_cpu;
8098 if (cpu >= nr_cpu_ids)
8100 if (!cpu_online(cpu))
8103 sqd->thread = kthread_create_on_cpu(io_sq_thread, sqd,
8104 cpu, "io_uring-sq");
8106 sqd->thread = kthread_create(io_sq_thread, sqd,
8109 if (IS_ERR(sqd->thread)) {
8110 ret = PTR_ERR(sqd->thread);
8114 ret = io_uring_alloc_task_context(sqd->thread);
8117 } else if (p->flags & IORING_SETUP_SQ_AFF) {
8118 /* Can't have SQ_AFF without SQPOLL */
8124 ret = io_init_wq_offload(ctx, p);
8130 io_finish_async(ctx);
8134 static void io_sq_offload_start(struct io_ring_ctx *ctx)
8136 struct io_sq_data *sqd = ctx->sq_data;
8138 if ((ctx->flags & IORING_SETUP_SQPOLL) && sqd->thread)
8139 wake_up_process(sqd->thread);
8142 static inline void __io_unaccount_mem(struct user_struct *user,
8143 unsigned long nr_pages)
8145 atomic_long_sub(nr_pages, &user->locked_vm);
8148 static inline int __io_account_mem(struct user_struct *user,
8149 unsigned long nr_pages)
8151 unsigned long page_limit, cur_pages, new_pages;
8153 /* Don't allow more pages than we can safely lock */
8154 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
8157 cur_pages = atomic_long_read(&user->locked_vm);
8158 new_pages = cur_pages + nr_pages;
8159 if (new_pages > page_limit)
8161 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
8162 new_pages) != cur_pages);
8167 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
8168 enum io_mem_account acct)
8171 __io_unaccount_mem(ctx->user, nr_pages);
8173 if (ctx->mm_account) {
8174 if (acct == ACCT_LOCKED)
8175 ctx->mm_account->locked_vm -= nr_pages;
8176 else if (acct == ACCT_PINNED)
8177 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
8181 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
8182 enum io_mem_account acct)
8186 if (ctx->limit_mem) {
8187 ret = __io_account_mem(ctx->user, nr_pages);
8192 if (ctx->mm_account) {
8193 if (acct == ACCT_LOCKED)
8194 ctx->mm_account->locked_vm += nr_pages;
8195 else if (acct == ACCT_PINNED)
8196 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
8202 static void io_mem_free(void *ptr)
8209 page = virt_to_head_page(ptr);
8210 if (put_page_testzero(page))
8211 free_compound_page(page);
8214 static void *io_mem_alloc(size_t size)
8216 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
8219 return (void *) __get_free_pages(gfp_flags, get_order(size));
8222 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
8225 struct io_rings *rings;
8226 size_t off, sq_array_size;
8228 off = struct_size(rings, cqes, cq_entries);
8229 if (off == SIZE_MAX)
8233 off = ALIGN(off, SMP_CACHE_BYTES);
8241 sq_array_size = array_size(sizeof(u32), sq_entries);
8242 if (sq_array_size == SIZE_MAX)
8245 if (check_add_overflow(off, sq_array_size, &off))
8251 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
8255 pages = (size_t)1 << get_order(
8256 rings_size(sq_entries, cq_entries, NULL));
8257 pages += (size_t)1 << get_order(
8258 array_size(sizeof(struct io_uring_sqe), sq_entries));
8263 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
8267 if (!ctx->user_bufs)
8270 for (i = 0; i < ctx->nr_user_bufs; i++) {
8271 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8273 for (j = 0; j < imu->nr_bvecs; j++)
8274 unpin_user_page(imu->bvec[j].bv_page);
8276 if (imu->acct_pages)
8277 io_unaccount_mem(ctx, imu->acct_pages, ACCT_PINNED);
8282 kfree(ctx->user_bufs);
8283 ctx->user_bufs = NULL;
8284 ctx->nr_user_bufs = 0;
8288 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
8289 void __user *arg, unsigned index)
8291 struct iovec __user *src;
8293 #ifdef CONFIG_COMPAT
8295 struct compat_iovec __user *ciovs;
8296 struct compat_iovec ciov;
8298 ciovs = (struct compat_iovec __user *) arg;
8299 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
8302 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
8303 dst->iov_len = ciov.iov_len;
8307 src = (struct iovec __user *) arg;
8308 if (copy_from_user(dst, &src[index], sizeof(*dst)))
8314 * Not super efficient, but this is just a registration time. And we do cache
8315 * the last compound head, so generally we'll only do a full search if we don't
8318 * We check if the given compound head page has already been accounted, to
8319 * avoid double accounting it. This allows us to account the full size of the
8320 * page, not just the constituent pages of a huge page.
8322 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
8323 int nr_pages, struct page *hpage)
8327 /* check current page array */
8328 for (i = 0; i < nr_pages; i++) {
8329 if (!PageCompound(pages[i]))
8331 if (compound_head(pages[i]) == hpage)
8335 /* check previously registered pages */
8336 for (i = 0; i < ctx->nr_user_bufs; i++) {
8337 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8339 for (j = 0; j < imu->nr_bvecs; j++) {
8340 if (!PageCompound(imu->bvec[j].bv_page))
8342 if (compound_head(imu->bvec[j].bv_page) == hpage)
8350 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
8351 int nr_pages, struct io_mapped_ubuf *imu,
8352 struct page **last_hpage)
8356 for (i = 0; i < nr_pages; i++) {
8357 if (!PageCompound(pages[i])) {
8362 hpage = compound_head(pages[i]);
8363 if (hpage == *last_hpage)
8365 *last_hpage = hpage;
8366 if (headpage_already_acct(ctx, pages, i, hpage))
8368 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8372 if (!imu->acct_pages)
8375 ret = io_account_mem(ctx, imu->acct_pages, ACCT_PINNED);
8377 imu->acct_pages = 0;
8381 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
8384 struct vm_area_struct **vmas = NULL;
8385 struct page **pages = NULL;
8386 struct page *last_hpage = NULL;
8387 int i, j, got_pages = 0;
8392 if (!nr_args || nr_args > UIO_MAXIOV)
8395 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
8397 if (!ctx->user_bufs)
8400 for (i = 0; i < nr_args; i++) {
8401 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8402 unsigned long off, start, end, ubuf;
8407 ret = io_copy_iov(ctx, &iov, arg, i);
8412 * Don't impose further limits on the size and buffer
8413 * constraints here, we'll -EINVAL later when IO is
8414 * submitted if they are wrong.
8417 if (!iov.iov_base || !iov.iov_len)
8420 /* arbitrary limit, but we need something */
8421 if (iov.iov_len > SZ_1G)
8424 ubuf = (unsigned long) iov.iov_base;
8425 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8426 start = ubuf >> PAGE_SHIFT;
8427 nr_pages = end - start;
8430 if (!pages || nr_pages > got_pages) {
8433 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
8435 vmas = kvmalloc_array(nr_pages,
8436 sizeof(struct vm_area_struct *),
8438 if (!pages || !vmas) {
8442 got_pages = nr_pages;
8445 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
8452 mmap_read_lock(current->mm);
8453 pret = pin_user_pages(ubuf, nr_pages,
8454 FOLL_WRITE | FOLL_LONGTERM,
8456 if (pret == nr_pages) {
8457 /* don't support file backed memory */
8458 for (j = 0; j < nr_pages; j++) {
8459 struct vm_area_struct *vma = vmas[j];
8462 !is_file_hugepages(vma->vm_file)) {
8468 ret = pret < 0 ? pret : -EFAULT;
8470 mmap_read_unlock(current->mm);
8473 * if we did partial map, or found file backed vmas,
8474 * release any pages we did get
8477 unpin_user_pages(pages, pret);
8482 ret = io_buffer_account_pin(ctx, pages, pret, imu, &last_hpage);
8484 unpin_user_pages(pages, pret);
8489 off = ubuf & ~PAGE_MASK;
8491 for (j = 0; j < nr_pages; j++) {
8494 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8495 imu->bvec[j].bv_page = pages[j];
8496 imu->bvec[j].bv_len = vec_len;
8497 imu->bvec[j].bv_offset = off;
8501 /* store original address for later verification */
8503 imu->len = iov.iov_len;
8504 imu->nr_bvecs = nr_pages;
8506 ctx->nr_user_bufs++;
8514 io_sqe_buffer_unregister(ctx);
8518 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8520 __s32 __user *fds = arg;
8526 if (copy_from_user(&fd, fds, sizeof(*fds)))
8529 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8530 if (IS_ERR(ctx->cq_ev_fd)) {
8531 int ret = PTR_ERR(ctx->cq_ev_fd);
8532 ctx->cq_ev_fd = NULL;
8539 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8541 if (ctx->cq_ev_fd) {
8542 eventfd_ctx_put(ctx->cq_ev_fd);
8543 ctx->cq_ev_fd = NULL;
8550 static int __io_destroy_buffers(int id, void *p, void *data)
8552 struct io_ring_ctx *ctx = data;
8553 struct io_buffer *buf = p;
8555 __io_remove_buffers(ctx, buf, id, -1U);
8559 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8561 idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx);
8562 idr_destroy(&ctx->io_buffer_idr);
8565 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8567 io_finish_async(ctx);
8568 io_sqe_buffer_unregister(ctx);
8570 if (ctx->sqo_task) {
8571 put_task_struct(ctx->sqo_task);
8572 ctx->sqo_task = NULL;
8573 mmdrop(ctx->mm_account);
8574 ctx->mm_account = NULL;
8577 #ifdef CONFIG_BLK_CGROUP
8578 if (ctx->sqo_blkcg_css)
8579 css_put(ctx->sqo_blkcg_css);
8582 io_sqe_files_unregister(ctx);
8583 io_eventfd_unregister(ctx);
8584 io_destroy_buffers(ctx);
8585 idr_destroy(&ctx->personality_idr);
8587 #if defined(CONFIG_UNIX)
8588 if (ctx->ring_sock) {
8589 ctx->ring_sock->file = NULL; /* so that iput() is called */
8590 sock_release(ctx->ring_sock);
8594 io_mem_free(ctx->rings);
8595 io_mem_free(ctx->sq_sqes);
8597 percpu_ref_exit(&ctx->refs);
8598 free_uid(ctx->user);
8599 put_cred(ctx->creds);
8600 kfree(ctx->cancel_hash);
8601 kmem_cache_free(req_cachep, ctx->fallback_req);
8605 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8607 struct io_ring_ctx *ctx = file->private_data;
8610 poll_wait(file, &ctx->cq_wait, wait);
8612 * synchronizes with barrier from wq_has_sleeper call in
8616 if (!io_sqring_full(ctx))
8617 mask |= EPOLLOUT | EPOLLWRNORM;
8618 if (io_cqring_events(ctx, false))
8619 mask |= EPOLLIN | EPOLLRDNORM;
8624 static int io_uring_fasync(int fd, struct file *file, int on)
8626 struct io_ring_ctx *ctx = file->private_data;
8628 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8631 static int io_remove_personalities(int id, void *p, void *data)
8633 struct io_ring_ctx *ctx = data;
8634 struct io_identity *iod;
8636 iod = idr_remove(&ctx->personality_idr, id);
8638 put_cred(iod->creds);
8639 if (refcount_dec_and_test(&iod->count))
8645 static void io_ring_exit_work(struct work_struct *work)
8647 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
8651 * If we're doing polled IO and end up having requests being
8652 * submitted async (out-of-line), then completions can come in while
8653 * we're waiting for refs to drop. We need to reap these manually,
8654 * as nobody else will be looking for them.
8657 io_iopoll_try_reap_events(ctx);
8658 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
8659 io_ring_ctx_free(ctx);
8662 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8664 mutex_lock(&ctx->uring_lock);
8665 percpu_ref_kill(&ctx->refs);
8667 io_cqring_overflow_flush(ctx, true, NULL, NULL);
8668 mutex_unlock(&ctx->uring_lock);
8670 io_kill_timeouts(ctx, NULL, NULL);
8671 io_poll_remove_all(ctx, NULL, NULL);
8674 io_wq_cancel_all(ctx->io_wq);
8676 /* if we failed setting up the ctx, we might not have any rings */
8677 io_iopoll_try_reap_events(ctx);
8678 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
8681 * Do this upfront, so we won't have a grace period where the ring
8682 * is closed but resources aren't reaped yet. This can cause
8683 * spurious failure in setting up a new ring.
8685 io_unaccount_mem(ctx, ring_pages(ctx->sq_entries, ctx->cq_entries),
8688 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8690 * Use system_unbound_wq to avoid spawning tons of event kworkers
8691 * if we're exiting a ton of rings at the same time. It just adds
8692 * noise and overhead, there's no discernable change in runtime
8693 * over using system_wq.
8695 queue_work(system_unbound_wq, &ctx->exit_work);
8698 static int io_uring_release(struct inode *inode, struct file *file)
8700 struct io_ring_ctx *ctx = file->private_data;
8702 file->private_data = NULL;
8703 io_ring_ctx_wait_and_kill(ctx);
8707 struct io_task_cancel {
8708 struct task_struct *task;
8709 struct files_struct *files;
8712 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8714 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8715 struct io_task_cancel *cancel = data;
8718 if (cancel->files && (req->flags & REQ_F_LINK_TIMEOUT)) {
8719 unsigned long flags;
8720 struct io_ring_ctx *ctx = req->ctx;
8722 /* protect against races with linked timeouts */
8723 spin_lock_irqsave(&ctx->completion_lock, flags);
8724 ret = io_match_task(req, cancel->task, cancel->files);
8725 spin_unlock_irqrestore(&ctx->completion_lock, flags);
8727 ret = io_match_task(req, cancel->task, cancel->files);
8732 static void io_cancel_defer_files(struct io_ring_ctx *ctx,
8733 struct task_struct *task,
8734 struct files_struct *files)
8736 struct io_defer_entry *de = NULL;
8739 spin_lock_irq(&ctx->completion_lock);
8740 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8741 if (io_match_task(de->req, task, files)) {
8742 list_cut_position(&list, &ctx->defer_list, &de->list);
8746 spin_unlock_irq(&ctx->completion_lock);
8748 while (!list_empty(&list)) {
8749 de = list_first_entry(&list, struct io_defer_entry, list);
8750 list_del_init(&de->list);
8751 req_set_fail_links(de->req);
8752 io_put_req(de->req);
8753 io_req_complete(de->req, -ECANCELED);
8758 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
8759 struct task_struct *task,
8760 struct files_struct *files)
8762 while (!list_empty_careful(&ctx->inflight_list)) {
8763 struct io_task_cancel cancel = { .task = task, .files = files };
8764 struct io_kiocb *req;
8768 spin_lock_irq(&ctx->inflight_lock);
8769 list_for_each_entry(req, &ctx->inflight_list, inflight_entry) {
8770 if (req->task != task ||
8771 req->work.identity->files != files)
8777 prepare_to_wait(&task->io_uring->wait, &wait,
8778 TASK_UNINTERRUPTIBLE);
8779 spin_unlock_irq(&ctx->inflight_lock);
8781 /* We need to keep going until we don't find a matching req */
8785 io_wq_cancel_cb(ctx->io_wq, io_cancel_task_cb, &cancel, true);
8786 io_poll_remove_all(ctx, task, files);
8787 io_kill_timeouts(ctx, task, files);
8788 /* cancellations _may_ trigger task work */
8791 finish_wait(&task->io_uring->wait, &wait);
8795 static void __io_uring_cancel_task_requests(struct io_ring_ctx *ctx,
8796 struct task_struct *task)
8799 struct io_task_cancel cancel = { .task = task, .files = NULL, };
8800 enum io_wq_cancel cret;
8803 cret = io_wq_cancel_cb(ctx->io_wq, io_cancel_task_cb, &cancel, true);
8804 if (cret != IO_WQ_CANCEL_NOTFOUND)
8807 /* SQPOLL thread does its own polling */
8808 if (!(ctx->flags & IORING_SETUP_SQPOLL)) {
8809 while (!list_empty_careful(&ctx->iopoll_list)) {
8810 io_iopoll_try_reap_events(ctx);
8815 ret |= io_poll_remove_all(ctx, task, NULL);
8816 ret |= io_kill_timeouts(ctx, task, NULL);
8825 * We need to iteratively cancel requests, in case a request has dependent
8826 * hard links. These persist even for failure of cancelations, hence keep
8827 * looping until none are found.
8829 static void io_uring_cancel_task_requests(struct io_ring_ctx *ctx,
8830 struct files_struct *files)
8832 struct task_struct *task = current;
8834 if ((ctx->flags & IORING_SETUP_SQPOLL) && ctx->sq_data) {
8835 task = ctx->sq_data->thread;
8836 atomic_inc(&task->io_uring->in_idle);
8837 io_sq_thread_park(ctx->sq_data);
8840 io_cancel_defer_files(ctx, task, files);
8841 io_ring_submit_lock(ctx, (ctx->flags & IORING_SETUP_IOPOLL));
8842 io_cqring_overflow_flush(ctx, true, task, files);
8843 io_ring_submit_unlock(ctx, (ctx->flags & IORING_SETUP_IOPOLL));
8846 __io_uring_cancel_task_requests(ctx, task);
8848 io_uring_cancel_files(ctx, task, files);
8850 if ((ctx->flags & IORING_SETUP_SQPOLL) && ctx->sq_data) {
8851 atomic_dec(&task->io_uring->in_idle);
8853 * If the files that are going away are the ones in the thread
8854 * identity, clear them out.
8856 if (task->io_uring->identity->files == files)
8857 task->io_uring->identity->files = NULL;
8858 io_sq_thread_unpark(ctx->sq_data);
8863 * Note that this task has used io_uring. We use it for cancelation purposes.
8865 static int io_uring_add_task_file(struct io_ring_ctx *ctx, struct file *file)
8867 struct io_uring_task *tctx = current->io_uring;
8869 if (unlikely(!tctx)) {
8872 ret = io_uring_alloc_task_context(current);
8875 tctx = current->io_uring;
8877 if (tctx->last != file) {
8878 void *old = xa_load(&tctx->xa, (unsigned long)file);
8882 xa_store(&tctx->xa, (unsigned long)file, file, GFP_KERNEL);
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))
9007 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 io_ring_submit_lock(ctx, (ctx->flags & IORING_SETUP_IOPOLL));
9176 if (!list_empty_careful(&ctx->cq_overflow_list))
9177 io_cqring_overflow_flush(ctx, false, NULL, NULL);
9178 io_ring_submit_unlock(ctx, (ctx->flags & IORING_SETUP_IOPOLL));
9179 if (flags & IORING_ENTER_SQ_WAKEUP)
9180 wake_up(&ctx->sq_data->wait);
9181 if (flags & IORING_ENTER_SQ_WAIT)
9182 io_sqpoll_wait_sq(ctx);
9183 submitted = to_submit;
9184 } else if (to_submit) {
9185 ret = io_uring_add_task_file(ctx, f.file);
9188 mutex_lock(&ctx->uring_lock);
9189 submitted = io_submit_sqes(ctx, to_submit);
9190 mutex_unlock(&ctx->uring_lock);
9192 if (submitted != to_submit)
9195 if (flags & IORING_ENTER_GETEVENTS) {
9196 const sigset_t __user *sig;
9197 struct __kernel_timespec __user *ts;
9199 ret = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
9203 min_complete = min(min_complete, ctx->cq_entries);
9206 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
9207 * space applications don't need to do io completion events
9208 * polling again, they can rely on io_sq_thread to do polling
9209 * work, which can reduce cpu usage and uring_lock contention.
9211 if (ctx->flags & IORING_SETUP_IOPOLL &&
9212 !(ctx->flags & IORING_SETUP_SQPOLL)) {
9213 ret = io_iopoll_check(ctx, min_complete);
9215 ret = io_cqring_wait(ctx, min_complete, sig, argsz, ts);
9220 percpu_ref_put(&ctx->refs);
9223 return submitted ? submitted : ret;
9226 #ifdef CONFIG_PROC_FS
9227 static int io_uring_show_cred(int id, void *p, void *data)
9229 struct io_identity *iod = p;
9230 const struct cred *cred = iod->creds;
9231 struct seq_file *m = data;
9232 struct user_namespace *uns = seq_user_ns(m);
9233 struct group_info *gi;
9238 seq_printf(m, "%5d\n", id);
9239 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
9240 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
9241 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
9242 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
9243 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
9244 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
9245 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
9246 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
9247 seq_puts(m, "\n\tGroups:\t");
9248 gi = cred->group_info;
9249 for (g = 0; g < gi->ngroups; g++) {
9250 seq_put_decimal_ull(m, g ? " " : "",
9251 from_kgid_munged(uns, gi->gid[g]));
9253 seq_puts(m, "\n\tCapEff:\t");
9254 cap = cred->cap_effective;
9255 CAP_FOR_EACH_U32(__capi)
9256 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
9261 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
9263 struct io_sq_data *sq = NULL;
9268 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
9269 * since fdinfo case grabs it in the opposite direction of normal use
9270 * cases. If we fail to get the lock, we just don't iterate any
9271 * structures that could be going away outside the io_uring mutex.
9273 has_lock = mutex_trylock(&ctx->uring_lock);
9275 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL))
9278 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
9279 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
9280 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
9281 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
9282 struct fixed_file_table *table;
9285 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
9286 f = table->files[i & IORING_FILE_TABLE_MASK];
9288 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
9290 seq_printf(m, "%5u: <none>\n", i);
9292 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
9293 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
9294 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
9296 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
9297 (unsigned int) buf->len);
9299 if (has_lock && !idr_is_empty(&ctx->personality_idr)) {
9300 seq_printf(m, "Personalities:\n");
9301 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
9303 seq_printf(m, "PollList:\n");
9304 spin_lock_irq(&ctx->completion_lock);
9305 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
9306 struct hlist_head *list = &ctx->cancel_hash[i];
9307 struct io_kiocb *req;
9309 hlist_for_each_entry(req, list, hash_node)
9310 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9311 req->task->task_works != NULL);
9313 spin_unlock_irq(&ctx->completion_lock);
9315 mutex_unlock(&ctx->uring_lock);
9318 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9320 struct io_ring_ctx *ctx = f->private_data;
9322 if (percpu_ref_tryget(&ctx->refs)) {
9323 __io_uring_show_fdinfo(ctx, m);
9324 percpu_ref_put(&ctx->refs);
9329 static const struct file_operations io_uring_fops = {
9330 .release = io_uring_release,
9331 .flush = io_uring_flush,
9332 .mmap = io_uring_mmap,
9334 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9335 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9337 .poll = io_uring_poll,
9338 .fasync = io_uring_fasync,
9339 #ifdef CONFIG_PROC_FS
9340 .show_fdinfo = io_uring_show_fdinfo,
9344 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9345 struct io_uring_params *p)
9347 struct io_rings *rings;
9348 size_t size, sq_array_offset;
9350 /* make sure these are sane, as we already accounted them */
9351 ctx->sq_entries = p->sq_entries;
9352 ctx->cq_entries = p->cq_entries;
9354 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9355 if (size == SIZE_MAX)
9358 rings = io_mem_alloc(size);
9363 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9364 rings->sq_ring_mask = p->sq_entries - 1;
9365 rings->cq_ring_mask = p->cq_entries - 1;
9366 rings->sq_ring_entries = p->sq_entries;
9367 rings->cq_ring_entries = p->cq_entries;
9368 ctx->sq_mask = rings->sq_ring_mask;
9369 ctx->cq_mask = rings->cq_ring_mask;
9371 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9372 if (size == SIZE_MAX) {
9373 io_mem_free(ctx->rings);
9378 ctx->sq_sqes = io_mem_alloc(size);
9379 if (!ctx->sq_sqes) {
9380 io_mem_free(ctx->rings);
9389 * Allocate an anonymous fd, this is what constitutes the application
9390 * visible backing of an io_uring instance. The application mmaps this
9391 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9392 * we have to tie this fd to a socket for file garbage collection purposes.
9394 static int io_uring_get_fd(struct io_ring_ctx *ctx)
9399 #if defined(CONFIG_UNIX)
9400 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9406 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9410 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9411 O_RDWR | O_CLOEXEC);
9415 ret = PTR_ERR(file);
9419 #if defined(CONFIG_UNIX)
9420 ctx->ring_sock->file = file;
9422 if (unlikely(io_uring_add_task_file(ctx, file))) {
9423 file = ERR_PTR(-ENOMEM);
9426 fd_install(ret, file);
9429 #if defined(CONFIG_UNIX)
9430 sock_release(ctx->ring_sock);
9431 ctx->ring_sock = NULL;
9436 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9437 struct io_uring_params __user *params)
9439 struct user_struct *user = NULL;
9440 struct io_ring_ctx *ctx;
9446 if (entries > IORING_MAX_ENTRIES) {
9447 if (!(p->flags & IORING_SETUP_CLAMP))
9449 entries = IORING_MAX_ENTRIES;
9453 * Use twice as many entries for the CQ ring. It's possible for the
9454 * application to drive a higher depth than the size of the SQ ring,
9455 * since the sqes are only used at submission time. This allows for
9456 * some flexibility in overcommitting a bit. If the application has
9457 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9458 * of CQ ring entries manually.
9460 p->sq_entries = roundup_pow_of_two(entries);
9461 if (p->flags & IORING_SETUP_CQSIZE) {
9463 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9464 * to a power-of-two, if it isn't already. We do NOT impose
9465 * any cq vs sq ring sizing.
9467 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9468 if (p->cq_entries < p->sq_entries)
9470 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9471 if (!(p->flags & IORING_SETUP_CLAMP))
9473 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9476 p->cq_entries = 2 * p->sq_entries;
9479 user = get_uid(current_user());
9480 limit_mem = !capable(CAP_IPC_LOCK);
9483 ret = __io_account_mem(user,
9484 ring_pages(p->sq_entries, p->cq_entries));
9491 ctx = io_ring_ctx_alloc(p);
9494 __io_unaccount_mem(user, ring_pages(p->sq_entries,
9499 ctx->compat = in_compat_syscall();
9501 ctx->creds = get_current_cred();
9503 ctx->loginuid = current->loginuid;
9504 ctx->sessionid = current->sessionid;
9506 ctx->sqo_task = get_task_struct(current);
9509 * This is just grabbed for accounting purposes. When a process exits,
9510 * the mm is exited and dropped before the files, hence we need to hang
9511 * on to this mm purely for the purposes of being able to unaccount
9512 * memory (locked/pinned vm). It's not used for anything else.
9514 mmgrab(current->mm);
9515 ctx->mm_account = current->mm;
9517 #ifdef CONFIG_BLK_CGROUP
9519 * The sq thread will belong to the original cgroup it was inited in.
9520 * If the cgroup goes offline (e.g. disabling the io controller), then
9521 * issued bios will be associated with the closest cgroup later in the
9525 ctx->sqo_blkcg_css = blkcg_css();
9526 ret = css_tryget_online(ctx->sqo_blkcg_css);
9529 /* don't init against a dying cgroup, have the user try again */
9530 ctx->sqo_blkcg_css = NULL;
9537 * Account memory _before_ installing the file descriptor. Once
9538 * the descriptor is installed, it can get closed at any time. Also
9539 * do this before hitting the general error path, as ring freeing
9540 * will un-account as well.
9542 io_account_mem(ctx, ring_pages(p->sq_entries, p->cq_entries),
9544 ctx->limit_mem = limit_mem;
9546 ret = io_allocate_scq_urings(ctx, p);
9550 ret = io_sq_offload_create(ctx, p);
9554 if (!(p->flags & IORING_SETUP_R_DISABLED))
9555 io_sq_offload_start(ctx);
9557 memset(&p->sq_off, 0, sizeof(p->sq_off));
9558 p->sq_off.head = offsetof(struct io_rings, sq.head);
9559 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9560 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9561 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9562 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9563 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9564 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9566 memset(&p->cq_off, 0, sizeof(p->cq_off));
9567 p->cq_off.head = offsetof(struct io_rings, cq.head);
9568 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9569 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9570 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9571 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9572 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9573 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9575 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9576 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9577 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9578 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
9579 IORING_FEAT_EXT_ARG;
9581 if (copy_to_user(params, p, sizeof(*p))) {
9587 * Install ring fd as the very last thing, so we don't risk someone
9588 * having closed it before we finish setup
9590 ret = io_uring_get_fd(ctx);
9594 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9597 io_ring_ctx_wait_and_kill(ctx);
9602 * Sets up an aio uring context, and returns the fd. Applications asks for a
9603 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9604 * params structure passed in.
9606 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9608 struct io_uring_params p;
9611 if (copy_from_user(&p, params, sizeof(p)))
9613 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9618 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9619 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9620 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9621 IORING_SETUP_R_DISABLED))
9624 return io_uring_create(entries, &p, params);
9627 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9628 struct io_uring_params __user *, params)
9630 return io_uring_setup(entries, params);
9633 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9635 struct io_uring_probe *p;
9639 size = struct_size(p, ops, nr_args);
9640 if (size == SIZE_MAX)
9642 p = kzalloc(size, GFP_KERNEL);
9647 if (copy_from_user(p, arg, size))
9650 if (memchr_inv(p, 0, size))
9653 p->last_op = IORING_OP_LAST - 1;
9654 if (nr_args > IORING_OP_LAST)
9655 nr_args = IORING_OP_LAST;
9657 for (i = 0; i < nr_args; i++) {
9659 if (!io_op_defs[i].not_supported)
9660 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9665 if (copy_to_user(arg, p, size))
9672 static int io_register_personality(struct io_ring_ctx *ctx)
9674 struct io_identity *id;
9677 id = kmalloc(sizeof(*id), GFP_KERNEL);
9681 io_init_identity(id);
9682 id->creds = get_current_cred();
9684 ret = idr_alloc_cyclic(&ctx->personality_idr, id, 1, USHRT_MAX, GFP_KERNEL);
9686 put_cred(id->creds);
9692 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
9694 struct io_identity *iod;
9696 iod = idr_remove(&ctx->personality_idr, id);
9698 put_cred(iod->creds);
9699 if (refcount_dec_and_test(&iod->count))
9707 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9708 unsigned int nr_args)
9710 struct io_uring_restriction *res;
9714 /* Restrictions allowed only if rings started disabled */
9715 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9718 /* We allow only a single restrictions registration */
9719 if (ctx->restrictions.registered)
9722 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9725 size = array_size(nr_args, sizeof(*res));
9726 if (size == SIZE_MAX)
9729 res = memdup_user(arg, size);
9731 return PTR_ERR(res);
9735 for (i = 0; i < nr_args; i++) {
9736 switch (res[i].opcode) {
9737 case IORING_RESTRICTION_REGISTER_OP:
9738 if (res[i].register_op >= IORING_REGISTER_LAST) {
9743 __set_bit(res[i].register_op,
9744 ctx->restrictions.register_op);
9746 case IORING_RESTRICTION_SQE_OP:
9747 if (res[i].sqe_op >= IORING_OP_LAST) {
9752 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9754 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9755 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9757 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9758 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9767 /* Reset all restrictions if an error happened */
9769 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9771 ctx->restrictions.registered = true;
9777 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9779 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9782 if (ctx->restrictions.registered)
9783 ctx->restricted = 1;
9785 ctx->flags &= ~IORING_SETUP_R_DISABLED;
9787 io_sq_offload_start(ctx);
9792 static bool io_register_op_must_quiesce(int op)
9795 case IORING_UNREGISTER_FILES:
9796 case IORING_REGISTER_FILES_UPDATE:
9797 case IORING_REGISTER_PROBE:
9798 case IORING_REGISTER_PERSONALITY:
9799 case IORING_UNREGISTER_PERSONALITY:
9806 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
9807 void __user *arg, unsigned nr_args)
9808 __releases(ctx->uring_lock)
9809 __acquires(ctx->uring_lock)
9814 * We're inside the ring mutex, if the ref is already dying, then
9815 * someone else killed the ctx or is already going through
9816 * io_uring_register().
9818 if (percpu_ref_is_dying(&ctx->refs))
9821 if (io_register_op_must_quiesce(opcode)) {
9822 percpu_ref_kill(&ctx->refs);
9825 * Drop uring mutex before waiting for references to exit. If
9826 * another thread is currently inside io_uring_enter() it might
9827 * need to grab the uring_lock to make progress. If we hold it
9828 * here across the drain wait, then we can deadlock. It's safe
9829 * to drop the mutex here, since no new references will come in
9830 * after we've killed the percpu ref.
9832 mutex_unlock(&ctx->uring_lock);
9834 ret = wait_for_completion_interruptible(&ctx->ref_comp);
9837 ret = io_run_task_work_sig();
9842 mutex_lock(&ctx->uring_lock);
9845 percpu_ref_resurrect(&ctx->refs);
9850 if (ctx->restricted) {
9851 if (opcode >= IORING_REGISTER_LAST) {
9856 if (!test_bit(opcode, ctx->restrictions.register_op)) {
9863 case IORING_REGISTER_BUFFERS:
9864 ret = io_sqe_buffer_register(ctx, arg, nr_args);
9866 case IORING_UNREGISTER_BUFFERS:
9870 ret = io_sqe_buffer_unregister(ctx);
9872 case IORING_REGISTER_FILES:
9873 ret = io_sqe_files_register(ctx, arg, nr_args);
9875 case IORING_UNREGISTER_FILES:
9879 ret = io_sqe_files_unregister(ctx);
9881 case IORING_REGISTER_FILES_UPDATE:
9882 ret = io_sqe_files_update(ctx, arg, nr_args);
9884 case IORING_REGISTER_EVENTFD:
9885 case IORING_REGISTER_EVENTFD_ASYNC:
9889 ret = io_eventfd_register(ctx, arg);
9892 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
9893 ctx->eventfd_async = 1;
9895 ctx->eventfd_async = 0;
9897 case IORING_UNREGISTER_EVENTFD:
9901 ret = io_eventfd_unregister(ctx);
9903 case IORING_REGISTER_PROBE:
9905 if (!arg || nr_args > 256)
9907 ret = io_probe(ctx, arg, nr_args);
9909 case IORING_REGISTER_PERSONALITY:
9913 ret = io_register_personality(ctx);
9915 case IORING_UNREGISTER_PERSONALITY:
9919 ret = io_unregister_personality(ctx, nr_args);
9921 case IORING_REGISTER_ENABLE_RINGS:
9925 ret = io_register_enable_rings(ctx);
9927 case IORING_REGISTER_RESTRICTIONS:
9928 ret = io_register_restrictions(ctx, arg, nr_args);
9936 if (io_register_op_must_quiesce(opcode)) {
9937 /* bring the ctx back to life */
9938 percpu_ref_reinit(&ctx->refs);
9940 reinit_completion(&ctx->ref_comp);
9945 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
9946 void __user *, arg, unsigned int, nr_args)
9948 struct io_ring_ctx *ctx;
9957 if (f.file->f_op != &io_uring_fops)
9960 ctx = f.file->private_data;
9962 mutex_lock(&ctx->uring_lock);
9963 ret = __io_uring_register(ctx, opcode, arg, nr_args);
9964 mutex_unlock(&ctx->uring_lock);
9965 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
9966 ctx->cq_ev_fd != NULL, ret);
9972 static int __init io_uring_init(void)
9974 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
9975 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
9976 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
9979 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
9980 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
9981 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
9982 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
9983 BUILD_BUG_SQE_ELEM(1, __u8, flags);
9984 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
9985 BUILD_BUG_SQE_ELEM(4, __s32, fd);
9986 BUILD_BUG_SQE_ELEM(8, __u64, off);
9987 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
9988 BUILD_BUG_SQE_ELEM(16, __u64, addr);
9989 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
9990 BUILD_BUG_SQE_ELEM(24, __u32, len);
9991 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
9992 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
9993 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
9994 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
9995 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
9996 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
9997 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
9998 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
9999 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
10000 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
10001 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
10002 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
10003 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
10004 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
10005 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
10006 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
10007 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
10008 BUILD_BUG_SQE_ELEM(42, __u16, personality);
10009 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
10011 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
10012 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
10013 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
10016 __initcall(io_uring_init);