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>
85 #define CREATE_TRACE_POINTS
86 #include <trace/events/io_uring.h>
88 #include <uapi/linux/io_uring.h>
93 #define IORING_MAX_ENTRIES 32768
94 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
97 * Shift of 9 is 512 entries, or exactly one page on 64-bit archs
99 #define IORING_FILE_TABLE_SHIFT 9
100 #define IORING_MAX_FILES_TABLE (1U << IORING_FILE_TABLE_SHIFT)
101 #define IORING_FILE_TABLE_MASK (IORING_MAX_FILES_TABLE - 1)
102 #define IORING_MAX_FIXED_FILES (64 * IORING_MAX_FILES_TABLE)
103 #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
104 IORING_REGISTER_LAST + IORING_OP_LAST)
107 u32 head ____cacheline_aligned_in_smp;
108 u32 tail ____cacheline_aligned_in_smp;
112 * This data is shared with the application through the mmap at offsets
113 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
115 * The offsets to the member fields are published through struct
116 * io_sqring_offsets when calling io_uring_setup.
120 * Head and tail offsets into the ring; the offsets need to be
121 * masked to get valid indices.
123 * The kernel controls head of the sq ring and the tail of the cq ring,
124 * and the application controls tail of the sq ring and the head of the
127 struct io_uring sq, cq;
129 * Bitmasks to apply to head and tail offsets (constant, equals
132 u32 sq_ring_mask, cq_ring_mask;
133 /* Ring sizes (constant, power of 2) */
134 u32 sq_ring_entries, cq_ring_entries;
136 * Number of invalid entries dropped by the kernel due to
137 * invalid index stored in array
139 * Written by the kernel, shouldn't be modified by the
140 * application (i.e. get number of "new events" by comparing to
143 * After a new SQ head value was read by the application this
144 * counter includes all submissions that were dropped reaching
145 * the new SQ head (and possibly more).
151 * Written by the kernel, shouldn't be modified by the
154 * The application needs a full memory barrier before checking
155 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
161 * Written by the application, shouldn't be modified by the
166 * Number of completion events lost because the queue was full;
167 * this should be avoided by the application by making sure
168 * there are not more requests pending than there is space in
169 * the completion queue.
171 * Written by the kernel, shouldn't be modified by the
172 * application (i.e. get number of "new events" by comparing to
175 * As completion events come in out of order this counter is not
176 * ordered with any other data.
180 * Ring buffer of completion events.
182 * The kernel writes completion events fresh every time they are
183 * produced, so the application is allowed to modify pending
186 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
189 struct io_mapped_ubuf {
192 struct bio_vec *bvec;
193 unsigned int nr_bvecs;
194 unsigned long acct_pages;
197 struct fixed_file_table {
201 struct fixed_file_ref_node {
202 struct percpu_ref refs;
203 struct list_head node;
204 struct list_head file_list;
205 struct fixed_file_data *file_data;
206 struct llist_node llist;
209 struct fixed_file_data {
210 struct fixed_file_table *table;
211 struct io_ring_ctx *ctx;
213 struct percpu_ref *cur_refs;
214 struct percpu_ref refs;
215 struct completion done;
216 struct list_head ref_list;
221 struct list_head list;
227 struct io_restriction {
228 DECLARE_BITMAP(register_op, IORING_REGISTER_LAST);
229 DECLARE_BITMAP(sqe_op, IORING_OP_LAST);
230 u8 sqe_flags_allowed;
231 u8 sqe_flags_required;
239 /* ctx's that are using this sqd */
240 struct list_head ctx_list;
241 struct list_head ctx_new_list;
242 struct mutex ctx_lock;
244 struct task_struct *thread;
245 struct wait_queue_head wait;
250 struct percpu_ref refs;
251 } ____cacheline_aligned_in_smp;
255 unsigned int compat: 1;
256 unsigned int limit_mem: 1;
257 unsigned int cq_overflow_flushed: 1;
258 unsigned int drain_next: 1;
259 unsigned int eventfd_async: 1;
260 unsigned int restricted: 1;
263 * Ring buffer of indices into array of io_uring_sqe, which is
264 * mmapped by the application using the IORING_OFF_SQES offset.
266 * This indirection could e.g. be used to assign fixed
267 * io_uring_sqe entries to operations and only submit them to
268 * the queue when needed.
270 * The kernel modifies neither the indices array nor the entries
274 unsigned cached_sq_head;
277 unsigned sq_thread_idle;
278 unsigned cached_sq_dropped;
279 atomic_t cached_cq_overflow;
280 unsigned long sq_check_overflow;
282 struct list_head defer_list;
283 struct list_head timeout_list;
284 struct list_head cq_overflow_list;
286 wait_queue_head_t inflight_wait;
287 struct io_uring_sqe *sq_sqes;
288 } ____cacheline_aligned_in_smp;
290 struct io_rings *rings;
296 * For SQPOLL usage - we hold a reference to the parent task, so we
297 * have access to the ->files
299 struct task_struct *sqo_task;
301 /* Only used for accounting purposes */
302 struct mm_struct *mm_account;
304 #ifdef CONFIG_BLK_CGROUP
305 struct cgroup_subsys_state *sqo_blkcg_css;
308 struct io_sq_data *sq_data; /* if using sq thread polling */
310 struct wait_queue_head sqo_sq_wait;
311 struct wait_queue_entry sqo_wait_entry;
312 struct list_head sqd_list;
315 * If used, fixed file set. Writers must ensure that ->refs is dead,
316 * readers must ensure that ->refs is alive as long as the file* is
317 * used. Only updated through io_uring_register(2).
319 struct fixed_file_data *file_data;
320 unsigned nr_user_files;
322 /* if used, fixed mapped user buffers */
323 unsigned nr_user_bufs;
324 struct io_mapped_ubuf *user_bufs;
326 struct user_struct *user;
328 const struct cred *creds;
330 struct completion ref_comp;
331 struct completion sq_thread_comp;
333 /* if all else fails... */
334 struct io_kiocb *fallback_req;
336 #if defined(CONFIG_UNIX)
337 struct socket *ring_sock;
340 struct idr io_buffer_idr;
342 struct idr personality_idr;
345 unsigned cached_cq_tail;
348 atomic_t cq_timeouts;
349 unsigned long cq_check_overflow;
350 struct wait_queue_head cq_wait;
351 struct fasync_struct *cq_fasync;
352 struct eventfd_ctx *cq_ev_fd;
353 } ____cacheline_aligned_in_smp;
356 struct mutex uring_lock;
357 wait_queue_head_t wait;
358 } ____cacheline_aligned_in_smp;
361 spinlock_t completion_lock;
364 * ->iopoll_list is protected by the ctx->uring_lock for
365 * io_uring instances that don't use IORING_SETUP_SQPOLL.
366 * For SQPOLL, only the single threaded io_sq_thread() will
367 * manipulate the list, hence no extra locking is needed there.
369 struct list_head iopoll_list;
370 struct hlist_head *cancel_hash;
371 unsigned cancel_hash_bits;
372 bool poll_multi_file;
374 spinlock_t inflight_lock;
375 struct list_head inflight_list;
376 } ____cacheline_aligned_in_smp;
378 struct delayed_work file_put_work;
379 struct llist_head file_put_llist;
381 struct work_struct exit_work;
382 struct io_restriction restrictions;
386 * First field must be the file pointer in all the
387 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
389 struct io_poll_iocb {
392 struct wait_queue_head *head;
398 struct wait_queue_entry wait;
403 struct file *put_file;
407 struct io_timeout_data {
408 struct io_kiocb *req;
409 struct hrtimer timer;
410 struct timespec64 ts;
411 enum hrtimer_mode mode;
416 struct sockaddr __user *addr;
417 int __user *addr_len;
419 unsigned long nofile;
441 struct list_head list;
445 /* NOTE: kiocb has the file as the first member, so don't do it here */
453 struct sockaddr __user *addr;
460 struct user_msghdr __user *umsg;
466 struct io_buffer *kbuf;
472 struct filename *filename;
474 unsigned long nofile;
477 struct io_files_update {
503 struct epoll_event event;
507 struct file *file_out;
508 struct file *file_in;
515 struct io_provide_buf {
529 const char __user *filename;
530 struct statx __user *buffer;
533 struct io_completion {
535 struct list_head list;
539 struct io_async_connect {
540 struct sockaddr_storage address;
543 struct io_async_msghdr {
544 struct iovec fast_iov[UIO_FASTIOV];
546 struct sockaddr __user *uaddr;
548 struct sockaddr_storage addr;
552 struct iovec fast_iov[UIO_FASTIOV];
553 const struct iovec *free_iovec;
554 struct iov_iter iter;
556 struct wait_page_queue wpq;
560 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
561 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
562 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
563 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
564 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
565 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
572 REQ_F_LINK_TIMEOUT_BIT,
574 REQ_F_COMP_LOCKED_BIT,
575 REQ_F_NEED_CLEANUP_BIT,
577 REQ_F_BUFFER_SELECTED_BIT,
578 REQ_F_NO_FILE_TABLE_BIT,
579 REQ_F_WORK_INITIALIZED_BIT,
581 /* not a real bit, just to check we're not overflowing the space */
587 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
588 /* drain existing IO first */
589 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
591 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
592 /* doesn't sever on completion < 0 */
593 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
595 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
596 /* IOSQE_BUFFER_SELECT */
597 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
600 REQ_F_LINK_HEAD = BIT(REQ_F_LINK_HEAD_BIT),
601 /* fail rest of links */
602 REQ_F_FAIL_LINK = BIT(REQ_F_FAIL_LINK_BIT),
603 /* on inflight list */
604 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
605 /* read/write uses file position */
606 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
607 /* must not punt to workers */
608 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
609 /* has linked timeout */
610 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
612 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
613 /* completion under lock */
614 REQ_F_COMP_LOCKED = BIT(REQ_F_COMP_LOCKED_BIT),
616 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
617 /* already went through poll handler */
618 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
619 /* buffer already selected */
620 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
621 /* doesn't need file table for this request */
622 REQ_F_NO_FILE_TABLE = BIT(REQ_F_NO_FILE_TABLE_BIT),
623 /* io_wq_work is initialized */
624 REQ_F_WORK_INITIALIZED = BIT(REQ_F_WORK_INITIALIZED_BIT),
628 struct io_poll_iocb poll;
629 struct io_poll_iocb *double_poll;
633 * NOTE! Each of the iocb union members has the file pointer
634 * as the first entry in their struct definition. So you can
635 * access the file pointer through any of the sub-structs,
636 * or directly as just 'ki_filp' in this struct.
642 struct io_poll_iocb poll;
643 struct io_accept accept;
645 struct io_cancel cancel;
646 struct io_timeout timeout;
647 struct io_connect connect;
648 struct io_sr_msg sr_msg;
650 struct io_close close;
651 struct io_files_update files_update;
652 struct io_fadvise fadvise;
653 struct io_madvise madvise;
654 struct io_epoll epoll;
655 struct io_splice splice;
656 struct io_provide_buf pbuf;
657 struct io_statx statx;
658 /* use only after cleaning per-op data, see io_clean_op() */
659 struct io_completion compl;
662 /* opcode allocated if it needs to store data for async defer */
665 /* polled IO has completed */
671 struct io_ring_ctx *ctx;
674 struct task_struct *task;
677 struct list_head link_list;
680 * 1. used with ctx->iopoll_list with reads/writes
681 * 2. to track reqs with ->files (see io_op_def::file_table)
683 struct list_head inflight_entry;
685 struct percpu_ref *fixed_file_refs;
686 struct callback_head task_work;
687 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
688 struct hlist_node hash_node;
689 struct async_poll *apoll;
690 struct io_wq_work work;
693 struct io_defer_entry {
694 struct list_head list;
695 struct io_kiocb *req;
699 #define IO_IOPOLL_BATCH 8
701 struct io_comp_state {
703 struct list_head list;
704 struct io_ring_ctx *ctx;
707 struct io_submit_state {
708 struct blk_plug plug;
711 * io_kiocb alloc cache
713 void *reqs[IO_IOPOLL_BATCH];
714 unsigned int free_reqs;
717 * Batch completion logic
719 struct io_comp_state comp;
722 * File reference cache
726 unsigned int has_refs;
727 unsigned int ios_left;
731 /* needs current->mm setup, does mm access */
732 unsigned needs_mm : 1;
733 /* needs req->file assigned */
734 unsigned needs_file : 1;
735 /* don't fail if file grab fails */
736 unsigned needs_file_no_error : 1;
737 /* hash wq insertion if file is a regular file */
738 unsigned hash_reg_file : 1;
739 /* unbound wq insertion if file is a non-regular file */
740 unsigned unbound_nonreg_file : 1;
741 /* opcode is not supported by this kernel */
742 unsigned not_supported : 1;
743 /* needs file table */
744 unsigned file_table : 1;
746 unsigned needs_fs : 1;
747 /* set if opcode supports polled "wait" */
749 unsigned pollout : 1;
750 /* op supports buffer selection */
751 unsigned buffer_select : 1;
752 /* needs rlimit(RLIMIT_FSIZE) assigned */
753 unsigned needs_fsize : 1;
754 /* must always have async data allocated */
755 unsigned needs_async_data : 1;
756 /* needs blkcg context, issues async io potentially */
757 unsigned needs_blkcg : 1;
758 /* size of async data needed, if any */
759 unsigned short async_size;
762 static const struct io_op_def io_op_defs[] __read_mostly = {
763 [IORING_OP_NOP] = {},
764 [IORING_OP_READV] = {
767 .unbound_nonreg_file = 1,
770 .needs_async_data = 1,
772 .async_size = sizeof(struct io_async_rw),
774 [IORING_OP_WRITEV] = {
778 .unbound_nonreg_file = 1,
781 .needs_async_data = 1,
783 .async_size = sizeof(struct io_async_rw),
785 [IORING_OP_FSYNC] = {
789 [IORING_OP_READ_FIXED] = {
791 .unbound_nonreg_file = 1,
794 .async_size = sizeof(struct io_async_rw),
796 [IORING_OP_WRITE_FIXED] = {
799 .unbound_nonreg_file = 1,
803 .async_size = sizeof(struct io_async_rw),
805 [IORING_OP_POLL_ADD] = {
807 .unbound_nonreg_file = 1,
809 [IORING_OP_POLL_REMOVE] = {},
810 [IORING_OP_SYNC_FILE_RANGE] = {
814 [IORING_OP_SENDMSG] = {
817 .unbound_nonreg_file = 1,
820 .needs_async_data = 1,
822 .async_size = sizeof(struct io_async_msghdr),
824 [IORING_OP_RECVMSG] = {
827 .unbound_nonreg_file = 1,
831 .needs_async_data = 1,
833 .async_size = sizeof(struct io_async_msghdr),
835 [IORING_OP_TIMEOUT] = {
837 .needs_async_data = 1,
838 .async_size = sizeof(struct io_timeout_data),
840 [IORING_OP_TIMEOUT_REMOVE] = {},
841 [IORING_OP_ACCEPT] = {
844 .unbound_nonreg_file = 1,
848 [IORING_OP_ASYNC_CANCEL] = {},
849 [IORING_OP_LINK_TIMEOUT] = {
851 .needs_async_data = 1,
852 .async_size = sizeof(struct io_timeout_data),
854 [IORING_OP_CONNECT] = {
857 .unbound_nonreg_file = 1,
859 .needs_async_data = 1,
860 .async_size = sizeof(struct io_async_connect),
862 [IORING_OP_FALLOCATE] = {
867 [IORING_OP_OPENAT] = {
872 [IORING_OP_CLOSE] = {
874 .needs_file_no_error = 1,
878 [IORING_OP_FILES_UPDATE] = {
882 [IORING_OP_STATX] = {
891 .unbound_nonreg_file = 1,
895 .async_size = sizeof(struct io_async_rw),
897 [IORING_OP_WRITE] = {
900 .unbound_nonreg_file = 1,
904 .async_size = sizeof(struct io_async_rw),
906 [IORING_OP_FADVISE] = {
910 [IORING_OP_MADVISE] = {
917 .unbound_nonreg_file = 1,
924 .unbound_nonreg_file = 1,
929 [IORING_OP_OPENAT2] = {
934 [IORING_OP_EPOLL_CTL] = {
935 .unbound_nonreg_file = 1,
938 [IORING_OP_SPLICE] = {
941 .unbound_nonreg_file = 1,
944 [IORING_OP_PROVIDE_BUFFERS] = {},
945 [IORING_OP_REMOVE_BUFFERS] = {},
949 .unbound_nonreg_file = 1,
953 enum io_mem_account {
958 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
959 struct io_comp_state *cs);
960 static void io_cqring_fill_event(struct io_kiocb *req, long res);
961 static void io_put_req(struct io_kiocb *req);
962 static void io_double_put_req(struct io_kiocb *req);
963 static void __io_double_put_req(struct io_kiocb *req);
964 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
965 static void __io_queue_linked_timeout(struct io_kiocb *req);
966 static void io_queue_linked_timeout(struct io_kiocb *req);
967 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
968 struct io_uring_files_update *ip,
970 static int io_prep_work_files(struct io_kiocb *req);
971 static void __io_clean_op(struct io_kiocb *req);
972 static int io_file_get(struct io_submit_state *state, struct io_kiocb *req,
973 int fd, struct file **out_file, bool fixed);
974 static void __io_queue_sqe(struct io_kiocb *req, struct io_comp_state *cs);
975 static void io_file_put_work(struct work_struct *work);
977 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
978 struct iovec **iovec, struct iov_iter *iter,
980 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
981 const struct iovec *fast_iov,
982 struct iov_iter *iter, bool force);
984 static struct kmem_cache *req_cachep;
986 static const struct file_operations io_uring_fops __read_mostly;
988 struct sock *io_uring_get_socket(struct file *file)
990 #if defined(CONFIG_UNIX)
991 if (file->f_op == &io_uring_fops) {
992 struct io_ring_ctx *ctx = file->private_data;
994 return ctx->ring_sock->sk;
999 EXPORT_SYMBOL(io_uring_get_socket);
1001 static inline void io_clean_op(struct io_kiocb *req)
1003 if (req->flags & (REQ_F_NEED_CLEANUP | REQ_F_BUFFER_SELECTED |
1008 static void io_sq_thread_drop_mm(void)
1010 struct mm_struct *mm = current->mm;
1013 kthread_unuse_mm(mm);
1018 static int __io_sq_thread_acquire_mm(struct io_ring_ctx *ctx)
1021 if (unlikely(!(ctx->flags & IORING_SETUP_SQPOLL) ||
1022 !ctx->sqo_task->mm ||
1023 !mmget_not_zero(ctx->sqo_task->mm)))
1025 kthread_use_mm(ctx->sqo_task->mm);
1031 static int io_sq_thread_acquire_mm(struct io_ring_ctx *ctx,
1032 struct io_kiocb *req)
1034 if (!io_op_defs[req->opcode].needs_mm)
1036 return __io_sq_thread_acquire_mm(ctx);
1039 static void io_sq_thread_associate_blkcg(struct io_ring_ctx *ctx,
1040 struct cgroup_subsys_state **cur_css)
1043 #ifdef CONFIG_BLK_CGROUP
1044 /* puts the old one when swapping */
1045 if (*cur_css != ctx->sqo_blkcg_css) {
1046 kthread_associate_blkcg(ctx->sqo_blkcg_css);
1047 *cur_css = ctx->sqo_blkcg_css;
1052 static void io_sq_thread_unassociate_blkcg(void)
1054 #ifdef CONFIG_BLK_CGROUP
1055 kthread_associate_blkcg(NULL);
1059 static inline void req_set_fail_links(struct io_kiocb *req)
1061 if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
1062 req->flags |= REQ_F_FAIL_LINK;
1066 * Note: must call io_req_init_async() for the first time you
1067 * touch any members of io_wq_work.
1069 static inline void io_req_init_async(struct io_kiocb *req)
1071 if (req->flags & REQ_F_WORK_INITIALIZED)
1074 memset(&req->work, 0, sizeof(req->work));
1075 req->flags |= REQ_F_WORK_INITIALIZED;
1078 static inline bool io_async_submit(struct io_ring_ctx *ctx)
1080 return ctx->flags & IORING_SETUP_SQPOLL;
1083 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
1085 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
1087 complete(&ctx->ref_comp);
1090 static inline bool io_is_timeout_noseq(struct io_kiocb *req)
1092 return !req->timeout.off;
1095 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
1097 struct io_ring_ctx *ctx;
1100 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1104 ctx->fallback_req = kmem_cache_alloc(req_cachep, GFP_KERNEL);
1105 if (!ctx->fallback_req)
1109 * Use 5 bits less than the max cq entries, that should give us around
1110 * 32 entries per hash list if totally full and uniformly spread.
1112 hash_bits = ilog2(p->cq_entries);
1116 ctx->cancel_hash_bits = hash_bits;
1117 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
1119 if (!ctx->cancel_hash)
1121 __hash_init(ctx->cancel_hash, 1U << hash_bits);
1123 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
1124 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
1127 ctx->flags = p->flags;
1128 init_waitqueue_head(&ctx->sqo_sq_wait);
1129 INIT_LIST_HEAD(&ctx->sqd_list);
1130 init_waitqueue_head(&ctx->cq_wait);
1131 INIT_LIST_HEAD(&ctx->cq_overflow_list);
1132 init_completion(&ctx->ref_comp);
1133 init_completion(&ctx->sq_thread_comp);
1134 idr_init(&ctx->io_buffer_idr);
1135 idr_init(&ctx->personality_idr);
1136 mutex_init(&ctx->uring_lock);
1137 init_waitqueue_head(&ctx->wait);
1138 spin_lock_init(&ctx->completion_lock);
1139 INIT_LIST_HEAD(&ctx->iopoll_list);
1140 INIT_LIST_HEAD(&ctx->defer_list);
1141 INIT_LIST_HEAD(&ctx->timeout_list);
1142 init_waitqueue_head(&ctx->inflight_wait);
1143 spin_lock_init(&ctx->inflight_lock);
1144 INIT_LIST_HEAD(&ctx->inflight_list);
1145 INIT_DELAYED_WORK(&ctx->file_put_work, io_file_put_work);
1146 init_llist_head(&ctx->file_put_llist);
1149 if (ctx->fallback_req)
1150 kmem_cache_free(req_cachep, ctx->fallback_req);
1151 kfree(ctx->cancel_hash);
1156 static bool req_need_defer(struct io_kiocb *req, u32 seq)
1158 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
1159 struct io_ring_ctx *ctx = req->ctx;
1161 return seq != ctx->cached_cq_tail
1162 + atomic_read(&ctx->cached_cq_overflow);
1168 static void __io_commit_cqring(struct io_ring_ctx *ctx)
1170 struct io_rings *rings = ctx->rings;
1172 /* order cqe stores with ring update */
1173 smp_store_release(&rings->cq.tail, ctx->cached_cq_tail);
1175 if (wq_has_sleeper(&ctx->cq_wait)) {
1176 wake_up_interruptible(&ctx->cq_wait);
1177 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1182 * Returns true if we need to defer file table putting. This can only happen
1183 * from the error path with REQ_F_COMP_LOCKED set.
1185 static bool io_req_clean_work(struct io_kiocb *req)
1187 if (!(req->flags & REQ_F_WORK_INITIALIZED))
1190 req->flags &= ~REQ_F_WORK_INITIALIZED;
1193 mmdrop(req->work.mm);
1194 req->work.mm = NULL;
1196 #ifdef CONFIG_BLK_CGROUP
1197 if (req->work.blkcg_css)
1198 css_put(req->work.blkcg_css);
1200 if (req->work.creds) {
1201 put_cred(req->work.creds);
1202 req->work.creds = NULL;
1205 struct fs_struct *fs = req->work.fs;
1207 if (req->flags & REQ_F_COMP_LOCKED)
1210 spin_lock(&req->work.fs->lock);
1213 spin_unlock(&req->work.fs->lock);
1216 req->work.fs = NULL;
1222 static void io_prep_async_work(struct io_kiocb *req)
1224 const struct io_op_def *def = &io_op_defs[req->opcode];
1226 io_req_init_async(req);
1228 if (req->flags & REQ_F_ISREG) {
1229 if (def->hash_reg_file || (req->ctx->flags & IORING_SETUP_IOPOLL))
1230 io_wq_hash_work(&req->work, file_inode(req->file));
1232 if (def->unbound_nonreg_file)
1233 req->work.flags |= IO_WQ_WORK_UNBOUND;
1235 if (!req->work.mm && def->needs_mm) {
1236 mmgrab(current->mm);
1237 req->work.mm = current->mm;
1239 #ifdef CONFIG_BLK_CGROUP
1240 if (!req->work.blkcg_css && def->needs_blkcg) {
1242 req->work.blkcg_css = blkcg_css();
1244 * This should be rare, either the cgroup is dying or the task
1245 * is moving cgroups. Just punt to root for the handful of ios.
1247 if (!css_tryget_online(req->work.blkcg_css))
1248 req->work.blkcg_css = NULL;
1252 if (!req->work.creds)
1253 req->work.creds = get_current_cred();
1254 if (!req->work.fs && def->needs_fs) {
1255 spin_lock(¤t->fs->lock);
1256 if (!current->fs->in_exec) {
1257 req->work.fs = current->fs;
1258 req->work.fs->users++;
1260 req->work.flags |= IO_WQ_WORK_CANCEL;
1262 spin_unlock(¤t->fs->lock);
1264 if (def->needs_fsize)
1265 req->work.fsize = rlimit(RLIMIT_FSIZE);
1267 req->work.fsize = RLIM_INFINITY;
1270 static void io_prep_async_link(struct io_kiocb *req)
1272 struct io_kiocb *cur;
1274 io_prep_async_work(req);
1275 if (req->flags & REQ_F_LINK_HEAD)
1276 list_for_each_entry(cur, &req->link_list, link_list)
1277 io_prep_async_work(cur);
1280 static struct io_kiocb *__io_queue_async_work(struct io_kiocb *req)
1282 struct io_ring_ctx *ctx = req->ctx;
1283 struct io_kiocb *link = io_prep_linked_timeout(req);
1285 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1286 &req->work, req->flags);
1287 io_wq_enqueue(ctx->io_wq, &req->work);
1291 static void io_queue_async_work(struct io_kiocb *req)
1293 struct io_kiocb *link;
1295 /* init ->work of the whole link before punting */
1296 io_prep_async_link(req);
1297 link = __io_queue_async_work(req);
1300 io_queue_linked_timeout(link);
1303 static void io_kill_timeout(struct io_kiocb *req)
1305 struct io_timeout_data *io = req->async_data;
1308 ret = hrtimer_try_to_cancel(&io->timer);
1310 atomic_set(&req->ctx->cq_timeouts,
1311 atomic_read(&req->ctx->cq_timeouts) + 1);
1312 list_del_init(&req->timeout.list);
1313 req->flags |= REQ_F_COMP_LOCKED;
1314 io_cqring_fill_event(req, 0);
1319 static bool io_task_match(struct io_kiocb *req, struct task_struct *tsk)
1321 struct io_ring_ctx *ctx = req->ctx;
1323 if (!tsk || req->task == tsk)
1325 if (ctx->flags & IORING_SETUP_SQPOLL) {
1326 if (ctx->sq_data && req->task == ctx->sq_data->thread)
1333 * Returns true if we found and killed one or more timeouts
1335 static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk)
1337 struct io_kiocb *req, *tmp;
1340 spin_lock_irq(&ctx->completion_lock);
1341 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) {
1342 if (io_task_match(req, tsk)) {
1343 io_kill_timeout(req);
1347 spin_unlock_irq(&ctx->completion_lock);
1348 return canceled != 0;
1351 static void __io_queue_deferred(struct io_ring_ctx *ctx)
1354 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
1355 struct io_defer_entry, list);
1356 struct io_kiocb *link;
1358 if (req_need_defer(de->req, de->seq))
1360 list_del_init(&de->list);
1361 /* punt-init is done before queueing for defer */
1362 link = __io_queue_async_work(de->req);
1364 __io_queue_linked_timeout(link);
1365 /* drop submission reference */
1366 link->flags |= REQ_F_COMP_LOCKED;
1370 } while (!list_empty(&ctx->defer_list));
1373 static void io_flush_timeouts(struct io_ring_ctx *ctx)
1375 while (!list_empty(&ctx->timeout_list)) {
1376 struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
1377 struct io_kiocb, timeout.list);
1379 if (io_is_timeout_noseq(req))
1381 if (req->timeout.target_seq != ctx->cached_cq_tail
1382 - atomic_read(&ctx->cq_timeouts))
1385 list_del_init(&req->timeout.list);
1386 io_kill_timeout(req);
1390 static void io_commit_cqring(struct io_ring_ctx *ctx)
1392 io_flush_timeouts(ctx);
1393 __io_commit_cqring(ctx);
1395 if (unlikely(!list_empty(&ctx->defer_list)))
1396 __io_queue_deferred(ctx);
1399 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
1401 struct io_rings *r = ctx->rings;
1403 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == r->sq_ring_entries;
1406 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1408 struct io_rings *rings = ctx->rings;
1411 tail = ctx->cached_cq_tail;
1413 * writes to the cq entry need to come after reading head; the
1414 * control dependency is enough as we're using WRITE_ONCE to
1417 if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries)
1420 ctx->cached_cq_tail++;
1421 return &rings->cqes[tail & ctx->cq_mask];
1424 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1428 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
1430 if (!ctx->eventfd_async)
1432 return io_wq_current_is_worker();
1435 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1437 if (waitqueue_active(&ctx->wait))
1438 wake_up(&ctx->wait);
1439 if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait))
1440 wake_up(&ctx->sq_data->wait);
1441 if (io_should_trigger_evfd(ctx))
1442 eventfd_signal(ctx->cq_ev_fd, 1);
1445 static void io_cqring_mark_overflow(struct io_ring_ctx *ctx)
1447 if (list_empty(&ctx->cq_overflow_list)) {
1448 clear_bit(0, &ctx->sq_check_overflow);
1449 clear_bit(0, &ctx->cq_check_overflow);
1450 ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW;
1454 static inline bool io_match_files(struct io_kiocb *req,
1455 struct files_struct *files)
1459 if (req->flags & REQ_F_WORK_INITIALIZED)
1460 return req->work.files == files;
1464 /* Returns true if there are no backlogged entries after the flush */
1465 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force,
1466 struct task_struct *tsk,
1467 struct files_struct *files)
1469 struct io_rings *rings = ctx->rings;
1470 struct io_kiocb *req, *tmp;
1471 struct io_uring_cqe *cqe;
1472 unsigned long flags;
1476 if (list_empty_careful(&ctx->cq_overflow_list))
1478 if ((ctx->cached_cq_tail - READ_ONCE(rings->cq.head) ==
1479 rings->cq_ring_entries))
1483 spin_lock_irqsave(&ctx->completion_lock, flags);
1485 /* if force is set, the ring is going away. always drop after that */
1487 ctx->cq_overflow_flushed = 1;
1490 list_for_each_entry_safe(req, tmp, &ctx->cq_overflow_list, compl.list) {
1491 if (tsk && req->task != tsk)
1493 if (!io_match_files(req, files))
1496 cqe = io_get_cqring(ctx);
1500 list_move(&req->compl.list, &list);
1502 WRITE_ONCE(cqe->user_data, req->user_data);
1503 WRITE_ONCE(cqe->res, req->result);
1504 WRITE_ONCE(cqe->flags, req->compl.cflags);
1506 WRITE_ONCE(ctx->rings->cq_overflow,
1507 atomic_inc_return(&ctx->cached_cq_overflow));
1511 io_commit_cqring(ctx);
1512 io_cqring_mark_overflow(ctx);
1514 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1515 io_cqring_ev_posted(ctx);
1517 while (!list_empty(&list)) {
1518 req = list_first_entry(&list, struct io_kiocb, compl.list);
1519 list_del(&req->compl.list);
1526 static void __io_cqring_fill_event(struct io_kiocb *req, long res, long cflags)
1528 struct io_ring_ctx *ctx = req->ctx;
1529 struct io_uring_cqe *cqe;
1531 trace_io_uring_complete(ctx, req->user_data, res);
1534 * If we can't get a cq entry, userspace overflowed the
1535 * submission (by quite a lot). Increment the overflow count in
1538 cqe = io_get_cqring(ctx);
1540 WRITE_ONCE(cqe->user_data, req->user_data);
1541 WRITE_ONCE(cqe->res, res);
1542 WRITE_ONCE(cqe->flags, cflags);
1543 } else if (ctx->cq_overflow_flushed || req->task->io_uring->in_idle) {
1545 * If we're in ring overflow flush mode, or in task cancel mode,
1546 * then we cannot store the request for later flushing, we need
1547 * to drop it on the floor.
1549 WRITE_ONCE(ctx->rings->cq_overflow,
1550 atomic_inc_return(&ctx->cached_cq_overflow));
1552 if (list_empty(&ctx->cq_overflow_list)) {
1553 set_bit(0, &ctx->sq_check_overflow);
1554 set_bit(0, &ctx->cq_check_overflow);
1555 ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW;
1559 req->compl.cflags = cflags;
1560 refcount_inc(&req->refs);
1561 list_add_tail(&req->compl.list, &ctx->cq_overflow_list);
1565 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1567 __io_cqring_fill_event(req, res, 0);
1570 static void io_cqring_add_event(struct io_kiocb *req, long res, long cflags)
1572 struct io_ring_ctx *ctx = req->ctx;
1573 unsigned long flags;
1575 spin_lock_irqsave(&ctx->completion_lock, flags);
1576 __io_cqring_fill_event(req, res, cflags);
1577 io_commit_cqring(ctx);
1578 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1580 io_cqring_ev_posted(ctx);
1583 static void io_submit_flush_completions(struct io_comp_state *cs)
1585 struct io_ring_ctx *ctx = cs->ctx;
1587 spin_lock_irq(&ctx->completion_lock);
1588 while (!list_empty(&cs->list)) {
1589 struct io_kiocb *req;
1591 req = list_first_entry(&cs->list, struct io_kiocb, compl.list);
1592 list_del(&req->compl.list);
1593 __io_cqring_fill_event(req, req->result, req->compl.cflags);
1594 if (!(req->flags & REQ_F_LINK_HEAD)) {
1595 req->flags |= REQ_F_COMP_LOCKED;
1598 spin_unlock_irq(&ctx->completion_lock);
1600 spin_lock_irq(&ctx->completion_lock);
1603 io_commit_cqring(ctx);
1604 spin_unlock_irq(&ctx->completion_lock);
1606 io_cqring_ev_posted(ctx);
1610 static void __io_req_complete(struct io_kiocb *req, long res, unsigned cflags,
1611 struct io_comp_state *cs)
1614 io_cqring_add_event(req, res, cflags);
1619 req->compl.cflags = cflags;
1620 list_add_tail(&req->compl.list, &cs->list);
1622 io_submit_flush_completions(cs);
1626 static void io_req_complete(struct io_kiocb *req, long res)
1628 __io_req_complete(req, res, 0, NULL);
1631 static inline bool io_is_fallback_req(struct io_kiocb *req)
1633 return req == (struct io_kiocb *)
1634 ((unsigned long) req->ctx->fallback_req & ~1UL);
1637 static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx)
1639 struct io_kiocb *req;
1641 req = ctx->fallback_req;
1642 if (!test_and_set_bit_lock(0, (unsigned long *) &ctx->fallback_req))
1648 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx,
1649 struct io_submit_state *state)
1651 if (!state->free_reqs) {
1652 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1656 sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs));
1657 ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs);
1660 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1661 * retry single alloc to be on the safe side.
1663 if (unlikely(ret <= 0)) {
1664 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1665 if (!state->reqs[0])
1669 state->free_reqs = ret;
1673 return state->reqs[state->free_reqs];
1675 return io_get_fallback_req(ctx);
1678 static inline void io_put_file(struct io_kiocb *req, struct file *file,
1682 percpu_ref_put(req->fixed_file_refs);
1687 static bool io_dismantle_req(struct io_kiocb *req)
1691 if (req->async_data)
1692 kfree(req->async_data);
1694 io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
1696 return io_req_clean_work(req);
1699 static void __io_free_req_finish(struct io_kiocb *req)
1701 struct io_uring_task *tctx = req->task->io_uring;
1702 struct io_ring_ctx *ctx = req->ctx;
1704 atomic_long_inc(&tctx->req_complete);
1706 wake_up(&tctx->wait);
1707 put_task_struct(req->task);
1709 if (likely(!io_is_fallback_req(req)))
1710 kmem_cache_free(req_cachep, req);
1712 clear_bit_unlock(0, (unsigned long *) &ctx->fallback_req);
1713 percpu_ref_put(&ctx->refs);
1716 static void io_req_task_file_table_put(struct callback_head *cb)
1718 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1719 struct fs_struct *fs = req->work.fs;
1721 spin_lock(&req->work.fs->lock);
1724 spin_unlock(&req->work.fs->lock);
1727 req->work.fs = NULL;
1728 __io_free_req_finish(req);
1731 static void __io_free_req(struct io_kiocb *req)
1733 if (!io_dismantle_req(req)) {
1734 __io_free_req_finish(req);
1738 init_task_work(&req->task_work, io_req_task_file_table_put);
1739 ret = task_work_add(req->task, &req->task_work, TWA_RESUME);
1740 if (unlikely(ret)) {
1741 struct task_struct *tsk;
1743 tsk = io_wq_get_task(req->ctx->io_wq);
1744 task_work_add(tsk, &req->task_work, 0);
1749 static bool io_link_cancel_timeout(struct io_kiocb *req)
1751 struct io_timeout_data *io = req->async_data;
1752 struct io_ring_ctx *ctx = req->ctx;
1755 ret = hrtimer_try_to_cancel(&io->timer);
1757 io_cqring_fill_event(req, -ECANCELED);
1758 io_commit_cqring(ctx);
1759 req->flags &= ~REQ_F_LINK_HEAD;
1767 static bool __io_kill_linked_timeout(struct io_kiocb *req)
1769 struct io_kiocb *link;
1772 if (list_empty(&req->link_list))
1774 link = list_first_entry(&req->link_list, struct io_kiocb, link_list);
1775 if (link->opcode != IORING_OP_LINK_TIMEOUT)
1778 list_del_init(&link->link_list);
1779 link->flags |= REQ_F_COMP_LOCKED;
1780 wake_ev = io_link_cancel_timeout(link);
1781 req->flags &= ~REQ_F_LINK_TIMEOUT;
1785 static void io_kill_linked_timeout(struct io_kiocb *req)
1787 struct io_ring_ctx *ctx = req->ctx;
1790 if (!(req->flags & REQ_F_COMP_LOCKED)) {
1791 unsigned long flags;
1793 spin_lock_irqsave(&ctx->completion_lock, flags);
1794 wake_ev = __io_kill_linked_timeout(req);
1795 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1797 wake_ev = __io_kill_linked_timeout(req);
1801 io_cqring_ev_posted(ctx);
1804 static struct io_kiocb *io_req_link_next(struct io_kiocb *req)
1806 struct io_kiocb *nxt;
1809 * The list should never be empty when we are called here. But could
1810 * potentially happen if the chain is messed up, check to be on the
1813 if (unlikely(list_empty(&req->link_list)))
1816 nxt = list_first_entry(&req->link_list, struct io_kiocb, link_list);
1817 list_del_init(&req->link_list);
1818 if (!list_empty(&nxt->link_list))
1819 nxt->flags |= REQ_F_LINK_HEAD;
1824 * Called if REQ_F_LINK_HEAD is set, and we fail the head request
1826 static void __io_fail_links(struct io_kiocb *req)
1828 struct io_ring_ctx *ctx = req->ctx;
1830 while (!list_empty(&req->link_list)) {
1831 struct io_kiocb *link = list_first_entry(&req->link_list,
1832 struct io_kiocb, link_list);
1834 list_del_init(&link->link_list);
1835 trace_io_uring_fail_link(req, link);
1837 io_cqring_fill_event(link, -ECANCELED);
1838 link->flags |= REQ_F_COMP_LOCKED;
1839 __io_double_put_req(link);
1840 req->flags &= ~REQ_F_LINK_TIMEOUT;
1843 io_commit_cqring(ctx);
1844 io_cqring_ev_posted(ctx);
1847 static void io_fail_links(struct io_kiocb *req)
1849 struct io_ring_ctx *ctx = req->ctx;
1851 if (!(req->flags & REQ_F_COMP_LOCKED)) {
1852 unsigned long flags;
1854 spin_lock_irqsave(&ctx->completion_lock, flags);
1855 __io_fail_links(req);
1856 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1858 __io_fail_links(req);
1861 io_cqring_ev_posted(ctx);
1864 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
1866 req->flags &= ~REQ_F_LINK_HEAD;
1867 if (req->flags & REQ_F_LINK_TIMEOUT)
1868 io_kill_linked_timeout(req);
1871 * If LINK is set, we have dependent requests in this chain. If we
1872 * didn't fail this request, queue the first one up, moving any other
1873 * dependencies to the next request. In case of failure, fail the rest
1876 if (likely(!(req->flags & REQ_F_FAIL_LINK)))
1877 return io_req_link_next(req);
1882 static struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1884 if (likely(!(req->flags & REQ_F_LINK_HEAD)))
1886 return __io_req_find_next(req);
1889 static int io_req_task_work_add(struct io_kiocb *req, bool twa_signal_ok)
1891 struct task_struct *tsk = req->task;
1892 struct io_ring_ctx *ctx = req->ctx;
1895 if (tsk->flags & PF_EXITING)
1899 * SQPOLL kernel thread doesn't need notification, just a wakeup. For
1900 * all other cases, use TWA_SIGNAL unconditionally to ensure we're
1901 * processing task_work. There's no reliable way to tell if TWA_RESUME
1905 if (!(ctx->flags & IORING_SETUP_SQPOLL) && twa_signal_ok)
1906 notify = TWA_SIGNAL;
1908 ret = task_work_add(tsk, &req->task_work, notify);
1910 wake_up_process(tsk);
1915 static void __io_req_task_cancel(struct io_kiocb *req, int error)
1917 struct io_ring_ctx *ctx = req->ctx;
1919 spin_lock_irq(&ctx->completion_lock);
1920 io_cqring_fill_event(req, error);
1921 io_commit_cqring(ctx);
1922 spin_unlock_irq(&ctx->completion_lock);
1924 io_cqring_ev_posted(ctx);
1925 req_set_fail_links(req);
1926 io_double_put_req(req);
1929 static void io_req_task_cancel(struct callback_head *cb)
1931 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1932 struct io_ring_ctx *ctx = req->ctx;
1934 __io_req_task_cancel(req, -ECANCELED);
1935 percpu_ref_put(&ctx->refs);
1938 static void __io_req_task_submit(struct io_kiocb *req)
1940 struct io_ring_ctx *ctx = req->ctx;
1942 if (!__io_sq_thread_acquire_mm(ctx)) {
1943 mutex_lock(&ctx->uring_lock);
1944 __io_queue_sqe(req, NULL);
1945 mutex_unlock(&ctx->uring_lock);
1947 __io_req_task_cancel(req, -EFAULT);
1951 static void io_req_task_submit(struct callback_head *cb)
1953 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
1954 struct io_ring_ctx *ctx = req->ctx;
1956 __io_req_task_submit(req);
1957 percpu_ref_put(&ctx->refs);
1960 static void io_req_task_queue(struct io_kiocb *req)
1964 init_task_work(&req->task_work, io_req_task_submit);
1965 percpu_ref_get(&req->ctx->refs);
1967 ret = io_req_task_work_add(req, true);
1968 if (unlikely(ret)) {
1969 struct task_struct *tsk;
1971 init_task_work(&req->task_work, io_req_task_cancel);
1972 tsk = io_wq_get_task(req->ctx->io_wq);
1973 task_work_add(tsk, &req->task_work, 0);
1974 wake_up_process(tsk);
1978 static void io_queue_next(struct io_kiocb *req)
1980 struct io_kiocb *nxt = io_req_find_next(req);
1983 io_req_task_queue(nxt);
1986 static void io_free_req(struct io_kiocb *req)
1993 void *reqs[IO_IOPOLL_BATCH];
1996 struct task_struct *task;
2000 static inline void io_init_req_batch(struct req_batch *rb)
2007 static void __io_req_free_batch_flush(struct io_ring_ctx *ctx,
2008 struct req_batch *rb)
2010 kmem_cache_free_bulk(req_cachep, rb->to_free, rb->reqs);
2011 percpu_ref_put_many(&ctx->refs, rb->to_free);
2015 static void io_req_free_batch_finish(struct io_ring_ctx *ctx,
2016 struct req_batch *rb)
2019 __io_req_free_batch_flush(ctx, rb);
2021 atomic_long_add(rb->task_refs, &rb->task->io_uring->req_complete);
2022 put_task_struct_many(rb->task, rb->task_refs);
2027 static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req)
2029 if (unlikely(io_is_fallback_req(req))) {
2033 if (req->flags & REQ_F_LINK_HEAD)
2036 if (req->task != rb->task) {
2038 atomic_long_add(rb->task_refs, &rb->task->io_uring->req_complete);
2039 put_task_struct_many(rb->task, rb->task_refs);
2041 rb->task = req->task;
2046 WARN_ON_ONCE(io_dismantle_req(req));
2047 rb->reqs[rb->to_free++] = req;
2048 if (unlikely(rb->to_free == ARRAY_SIZE(rb->reqs)))
2049 __io_req_free_batch_flush(req->ctx, rb);
2053 * Drop reference to request, return next in chain (if there is one) if this
2054 * was the last reference to this request.
2056 static struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
2058 struct io_kiocb *nxt = NULL;
2060 if (refcount_dec_and_test(&req->refs)) {
2061 nxt = io_req_find_next(req);
2067 static void io_put_req(struct io_kiocb *req)
2069 if (refcount_dec_and_test(&req->refs))
2073 static struct io_wq_work *io_steal_work(struct io_kiocb *req)
2075 struct io_kiocb *nxt;
2078 * A ref is owned by io-wq in which context we're. So, if that's the
2079 * last one, it's safe to steal next work. False negatives are Ok,
2080 * it just will be re-punted async in io_put_work()
2082 if (refcount_read(&req->refs) != 1)
2085 nxt = io_req_find_next(req);
2086 return nxt ? &nxt->work : NULL;
2090 * Must only be used if we don't need to care about links, usually from
2091 * within the completion handling itself.
2093 static void __io_double_put_req(struct io_kiocb *req)
2095 /* drop both submit and complete references */
2096 if (refcount_sub_and_test(2, &req->refs))
2100 static void io_double_put_req(struct io_kiocb *req)
2102 /* drop both submit and complete references */
2103 if (refcount_sub_and_test(2, &req->refs))
2107 static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush)
2109 struct io_rings *rings = ctx->rings;
2111 if (test_bit(0, &ctx->cq_check_overflow)) {
2113 * noflush == true is from the waitqueue handler, just ensure
2114 * we wake up the task, and the next invocation will flush the
2115 * entries. We cannot safely to it from here.
2117 if (noflush && !list_empty(&ctx->cq_overflow_list))
2120 io_cqring_overflow_flush(ctx, false, NULL, NULL);
2123 /* See comment at the top of this file */
2125 return ctx->cached_cq_tail - READ_ONCE(rings->cq.head);
2128 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
2130 struct io_rings *rings = ctx->rings;
2132 /* make sure SQ entry isn't read before tail */
2133 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
2136 static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf)
2138 unsigned int cflags;
2140 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
2141 cflags |= IORING_CQE_F_BUFFER;
2142 req->flags &= ~REQ_F_BUFFER_SELECTED;
2147 static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
2149 struct io_buffer *kbuf;
2151 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2152 return io_put_kbuf(req, kbuf);
2155 static inline bool io_run_task_work(void)
2158 * Not safe to run on exiting task, and the task_work handling will
2159 * not add work to such a task.
2161 if (unlikely(current->flags & PF_EXITING))
2163 if (current->task_works) {
2164 __set_current_state(TASK_RUNNING);
2172 static void io_iopoll_queue(struct list_head *again)
2174 struct io_kiocb *req;
2177 req = list_first_entry(again, struct io_kiocb, inflight_entry);
2178 list_del(&req->inflight_entry);
2179 __io_complete_rw(req, -EAGAIN, 0, NULL);
2180 } while (!list_empty(again));
2184 * Find and free completed poll iocbs
2186 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
2187 struct list_head *done)
2189 struct req_batch rb;
2190 struct io_kiocb *req;
2193 /* order with ->result store in io_complete_rw_iopoll() */
2196 io_init_req_batch(&rb);
2197 while (!list_empty(done)) {
2200 req = list_first_entry(done, struct io_kiocb, inflight_entry);
2201 if (READ_ONCE(req->result) == -EAGAIN) {
2203 req->iopoll_completed = 0;
2204 list_move_tail(&req->inflight_entry, &again);
2207 list_del(&req->inflight_entry);
2209 if (req->flags & REQ_F_BUFFER_SELECTED)
2210 cflags = io_put_rw_kbuf(req);
2212 __io_cqring_fill_event(req, req->result, cflags);
2215 if (refcount_dec_and_test(&req->refs))
2216 io_req_free_batch(&rb, req);
2219 io_commit_cqring(ctx);
2220 if (ctx->flags & IORING_SETUP_SQPOLL)
2221 io_cqring_ev_posted(ctx);
2222 io_req_free_batch_finish(ctx, &rb);
2224 if (!list_empty(&again))
2225 io_iopoll_queue(&again);
2228 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
2231 struct io_kiocb *req, *tmp;
2237 * Only spin for completions if we don't have multiple devices hanging
2238 * off our complete list, and we're under the requested amount.
2240 spin = !ctx->poll_multi_file && *nr_events < min;
2243 list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) {
2244 struct kiocb *kiocb = &req->rw.kiocb;
2247 * Move completed and retryable entries to our local lists.
2248 * If we find a request that requires polling, break out
2249 * and complete those lists first, if we have entries there.
2251 if (READ_ONCE(req->iopoll_completed)) {
2252 list_move_tail(&req->inflight_entry, &done);
2255 if (!list_empty(&done))
2258 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
2262 /* iopoll may have completed current req */
2263 if (READ_ONCE(req->iopoll_completed))
2264 list_move_tail(&req->inflight_entry, &done);
2271 if (!list_empty(&done))
2272 io_iopoll_complete(ctx, nr_events, &done);
2278 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
2279 * non-spinning poll check - we'll still enter the driver poll loop, but only
2280 * as a non-spinning completion check.
2282 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
2285 while (!list_empty(&ctx->iopoll_list) && !need_resched()) {
2288 ret = io_do_iopoll(ctx, nr_events, min);
2291 if (*nr_events >= min)
2299 * We can't just wait for polled events to come to us, we have to actively
2300 * find and complete them.
2302 static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
2304 if (!(ctx->flags & IORING_SETUP_IOPOLL))
2307 mutex_lock(&ctx->uring_lock);
2308 while (!list_empty(&ctx->iopoll_list)) {
2309 unsigned int nr_events = 0;
2311 io_do_iopoll(ctx, &nr_events, 0);
2313 /* let it sleep and repeat later if can't complete a request */
2317 * Ensure we allow local-to-the-cpu processing to take place,
2318 * in this case we need to ensure that we reap all events.
2319 * Also let task_work, etc. to progress by releasing the mutex
2321 if (need_resched()) {
2322 mutex_unlock(&ctx->uring_lock);
2324 mutex_lock(&ctx->uring_lock);
2327 mutex_unlock(&ctx->uring_lock);
2330 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
2332 unsigned int nr_events = 0;
2333 int iters = 0, ret = 0;
2336 * We disallow the app entering submit/complete with polling, but we
2337 * still need to lock the ring to prevent racing with polled issue
2338 * that got punted to a workqueue.
2340 mutex_lock(&ctx->uring_lock);
2343 * Don't enter poll loop if we already have events pending.
2344 * If we do, we can potentially be spinning for commands that
2345 * already triggered a CQE (eg in error).
2347 if (io_cqring_events(ctx, false))
2351 * If a submit got punted to a workqueue, we can have the
2352 * application entering polling for a command before it gets
2353 * issued. That app will hold the uring_lock for the duration
2354 * of the poll right here, so we need to take a breather every
2355 * now and then to ensure that the issue has a chance to add
2356 * the poll to the issued list. Otherwise we can spin here
2357 * forever, while the workqueue is stuck trying to acquire the
2360 if (!(++iters & 7)) {
2361 mutex_unlock(&ctx->uring_lock);
2363 mutex_lock(&ctx->uring_lock);
2366 ret = io_iopoll_getevents(ctx, &nr_events, min);
2370 } while (min && !nr_events && !need_resched());
2372 mutex_unlock(&ctx->uring_lock);
2376 static void kiocb_end_write(struct io_kiocb *req)
2379 * Tell lockdep we inherited freeze protection from submission
2382 if (req->flags & REQ_F_ISREG) {
2383 struct inode *inode = file_inode(req->file);
2385 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
2387 file_end_write(req->file);
2390 static void io_complete_rw_common(struct kiocb *kiocb, long res,
2391 struct io_comp_state *cs)
2393 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2396 if (kiocb->ki_flags & IOCB_WRITE)
2397 kiocb_end_write(req);
2399 if (res != req->result)
2400 req_set_fail_links(req);
2401 if (req->flags & REQ_F_BUFFER_SELECTED)
2402 cflags = io_put_rw_kbuf(req);
2403 __io_req_complete(req, res, cflags, cs);
2407 static bool io_resubmit_prep(struct io_kiocb *req, int error)
2409 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2410 ssize_t ret = -ECANCELED;
2411 struct iov_iter iter;
2419 switch (req->opcode) {
2420 case IORING_OP_READV:
2421 case IORING_OP_READ_FIXED:
2422 case IORING_OP_READ:
2425 case IORING_OP_WRITEV:
2426 case IORING_OP_WRITE_FIXED:
2427 case IORING_OP_WRITE:
2431 printk_once(KERN_WARNING "io_uring: bad opcode in resubmit %d\n",
2436 if (!req->async_data) {
2437 ret = io_import_iovec(rw, req, &iovec, &iter, false);
2440 ret = io_setup_async_rw(req, iovec, inline_vecs, &iter, false);
2448 req_set_fail_links(req);
2449 io_req_complete(req, ret);
2454 static bool io_rw_reissue(struct io_kiocb *req, long res)
2457 umode_t mode = file_inode(req->file)->i_mode;
2460 if (!S_ISBLK(mode) && !S_ISREG(mode))
2462 if ((res != -EAGAIN && res != -EOPNOTSUPP) || io_wq_current_is_worker())
2465 ret = io_sq_thread_acquire_mm(req->ctx, req);
2467 if (io_resubmit_prep(req, ret)) {
2468 refcount_inc(&req->refs);
2469 io_queue_async_work(req);
2477 static void __io_complete_rw(struct io_kiocb *req, long res, long res2,
2478 struct io_comp_state *cs)
2480 if (!io_rw_reissue(req, res))
2481 io_complete_rw_common(&req->rw.kiocb, res, cs);
2484 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
2486 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2488 __io_complete_rw(req, res, res2, NULL);
2491 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
2493 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2495 if (kiocb->ki_flags & IOCB_WRITE)
2496 kiocb_end_write(req);
2498 if (res != -EAGAIN && res != req->result)
2499 req_set_fail_links(req);
2501 WRITE_ONCE(req->result, res);
2502 /* order with io_poll_complete() checking ->result */
2504 WRITE_ONCE(req->iopoll_completed, 1);
2508 * After the iocb has been issued, it's safe to be found on the poll list.
2509 * Adding the kiocb to the list AFTER submission ensures that we don't
2510 * find it from a io_iopoll_getevents() thread before the issuer is done
2511 * accessing the kiocb cookie.
2513 static void io_iopoll_req_issued(struct io_kiocb *req)
2515 struct io_ring_ctx *ctx = req->ctx;
2518 * Track whether we have multiple files in our lists. This will impact
2519 * how we do polling eventually, not spinning if we're on potentially
2520 * different devices.
2522 if (list_empty(&ctx->iopoll_list)) {
2523 ctx->poll_multi_file = false;
2524 } else if (!ctx->poll_multi_file) {
2525 struct io_kiocb *list_req;
2527 list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb,
2529 if (list_req->file != req->file)
2530 ctx->poll_multi_file = true;
2534 * For fast devices, IO may have already completed. If it has, add
2535 * it to the front so we find it first.
2537 if (READ_ONCE(req->iopoll_completed))
2538 list_add(&req->inflight_entry, &ctx->iopoll_list);
2540 list_add_tail(&req->inflight_entry, &ctx->iopoll_list);
2542 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2543 wq_has_sleeper(&ctx->sq_data->wait))
2544 wake_up(&ctx->sq_data->wait);
2547 static void __io_state_file_put(struct io_submit_state *state)
2549 if (state->has_refs)
2550 fput_many(state->file, state->has_refs);
2554 static inline void io_state_file_put(struct io_submit_state *state)
2557 __io_state_file_put(state);
2561 * Get as many references to a file as we have IOs left in this submission,
2562 * assuming most submissions are for one file, or at least that each file
2563 * has more than one submission.
2565 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2571 if (state->fd == fd) {
2576 __io_state_file_put(state);
2578 state->file = fget_many(fd, state->ios_left);
2584 state->has_refs = state->ios_left;
2588 static bool io_bdev_nowait(struct block_device *bdev)
2591 return !bdev || queue_is_mq(bdev_get_queue(bdev));
2598 * If we tracked the file through the SCM inflight mechanism, we could support
2599 * any file. For now, just ensure that anything potentially problematic is done
2602 static bool io_file_supports_async(struct file *file, int rw)
2604 umode_t mode = file_inode(file)->i_mode;
2606 if (S_ISBLK(mode)) {
2607 if (io_bdev_nowait(file->f_inode->i_bdev))
2611 if (S_ISCHR(mode) || S_ISSOCK(mode))
2613 if (S_ISREG(mode)) {
2614 if (io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
2615 file->f_op != &io_uring_fops)
2620 /* any ->read/write should understand O_NONBLOCK */
2621 if (file->f_flags & O_NONBLOCK)
2624 if (!(file->f_mode & FMODE_NOWAIT))
2628 return file->f_op->read_iter != NULL;
2630 return file->f_op->write_iter != NULL;
2633 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2635 struct io_ring_ctx *ctx = req->ctx;
2636 struct kiocb *kiocb = &req->rw.kiocb;
2640 if (S_ISREG(file_inode(req->file)->i_mode))
2641 req->flags |= REQ_F_ISREG;
2643 kiocb->ki_pos = READ_ONCE(sqe->off);
2644 if (kiocb->ki_pos == -1 && !(req->file->f_mode & FMODE_STREAM)) {
2645 req->flags |= REQ_F_CUR_POS;
2646 kiocb->ki_pos = req->file->f_pos;
2648 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2649 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2650 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2654 ioprio = READ_ONCE(sqe->ioprio);
2656 ret = ioprio_check_cap(ioprio);
2660 kiocb->ki_ioprio = ioprio;
2662 kiocb->ki_ioprio = get_current_ioprio();
2664 /* don't allow async punt if RWF_NOWAIT was requested */
2665 if (kiocb->ki_flags & IOCB_NOWAIT)
2666 req->flags |= REQ_F_NOWAIT;
2668 if (ctx->flags & IORING_SETUP_IOPOLL) {
2669 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2670 !kiocb->ki_filp->f_op->iopoll)
2673 kiocb->ki_flags |= IOCB_HIPRI;
2674 kiocb->ki_complete = io_complete_rw_iopoll;
2675 req->iopoll_completed = 0;
2677 if (kiocb->ki_flags & IOCB_HIPRI)
2679 kiocb->ki_complete = io_complete_rw;
2682 req->rw.addr = READ_ONCE(sqe->addr);
2683 req->rw.len = READ_ONCE(sqe->len);
2684 req->buf_index = READ_ONCE(sqe->buf_index);
2688 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2694 case -ERESTARTNOINTR:
2695 case -ERESTARTNOHAND:
2696 case -ERESTART_RESTARTBLOCK:
2698 * We can't just restart the syscall, since previously
2699 * submitted sqes may already be in progress. Just fail this
2705 kiocb->ki_complete(kiocb, ret, 0);
2709 static void kiocb_done(struct kiocb *kiocb, ssize_t ret,
2710 struct io_comp_state *cs)
2712 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2713 struct io_async_rw *io = req->async_data;
2715 /* add previously done IO, if any */
2716 if (io && io->bytes_done > 0) {
2718 ret = io->bytes_done;
2720 ret += io->bytes_done;
2723 if (req->flags & REQ_F_CUR_POS)
2724 req->file->f_pos = kiocb->ki_pos;
2725 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2726 __io_complete_rw(req, ret, 0, cs);
2728 io_rw_done(kiocb, ret);
2731 static ssize_t io_import_fixed(struct io_kiocb *req, int rw,
2732 struct iov_iter *iter)
2734 struct io_ring_ctx *ctx = req->ctx;
2735 size_t len = req->rw.len;
2736 struct io_mapped_ubuf *imu;
2737 u16 index, buf_index = req->buf_index;
2741 if (unlikely(buf_index >= ctx->nr_user_bufs))
2743 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2744 imu = &ctx->user_bufs[index];
2745 buf_addr = req->rw.addr;
2748 if (buf_addr + len < buf_addr)
2750 /* not inside the mapped region */
2751 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2755 * May not be a start of buffer, set size appropriately
2756 * and advance us to the beginning.
2758 offset = buf_addr - imu->ubuf;
2759 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2763 * Don't use iov_iter_advance() here, as it's really slow for
2764 * using the latter parts of a big fixed buffer - it iterates
2765 * over each segment manually. We can cheat a bit here, because
2768 * 1) it's a BVEC iter, we set it up
2769 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2770 * first and last bvec
2772 * So just find our index, and adjust the iterator afterwards.
2773 * If the offset is within the first bvec (or the whole first
2774 * bvec, just use iov_iter_advance(). This makes it easier
2775 * since we can just skip the first segment, which may not
2776 * be PAGE_SIZE aligned.
2778 const struct bio_vec *bvec = imu->bvec;
2780 if (offset <= bvec->bv_len) {
2781 iov_iter_advance(iter, offset);
2783 unsigned long seg_skip;
2785 /* skip first vec */
2786 offset -= bvec->bv_len;
2787 seg_skip = 1 + (offset >> PAGE_SHIFT);
2789 iter->bvec = bvec + seg_skip;
2790 iter->nr_segs -= seg_skip;
2791 iter->count -= bvec->bv_len + offset;
2792 iter->iov_offset = offset & ~PAGE_MASK;
2799 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2802 mutex_unlock(&ctx->uring_lock);
2805 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2808 * "Normal" inline submissions always hold the uring_lock, since we
2809 * grab it from the system call. Same is true for the SQPOLL offload.
2810 * The only exception is when we've detached the request and issue it
2811 * from an async worker thread, grab the lock for that case.
2814 mutex_lock(&ctx->uring_lock);
2817 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2818 int bgid, struct io_buffer *kbuf,
2821 struct io_buffer *head;
2823 if (req->flags & REQ_F_BUFFER_SELECTED)
2826 io_ring_submit_lock(req->ctx, needs_lock);
2828 lockdep_assert_held(&req->ctx->uring_lock);
2830 head = idr_find(&req->ctx->io_buffer_idr, bgid);
2832 if (!list_empty(&head->list)) {
2833 kbuf = list_last_entry(&head->list, struct io_buffer,
2835 list_del(&kbuf->list);
2838 idr_remove(&req->ctx->io_buffer_idr, bgid);
2840 if (*len > kbuf->len)
2843 kbuf = ERR_PTR(-ENOBUFS);
2846 io_ring_submit_unlock(req->ctx, needs_lock);
2851 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2854 struct io_buffer *kbuf;
2857 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2858 bgid = req->buf_index;
2859 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2862 req->rw.addr = (u64) (unsigned long) kbuf;
2863 req->flags |= REQ_F_BUFFER_SELECTED;
2864 return u64_to_user_ptr(kbuf->addr);
2867 #ifdef CONFIG_COMPAT
2868 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2871 struct compat_iovec __user *uiov;
2872 compat_ssize_t clen;
2876 uiov = u64_to_user_ptr(req->rw.addr);
2877 if (!access_ok(uiov, sizeof(*uiov)))
2879 if (__get_user(clen, &uiov->iov_len))
2885 buf = io_rw_buffer_select(req, &len, needs_lock);
2887 return PTR_ERR(buf);
2888 iov[0].iov_base = buf;
2889 iov[0].iov_len = (compat_size_t) len;
2894 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2897 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2901 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2904 len = iov[0].iov_len;
2907 buf = io_rw_buffer_select(req, &len, needs_lock);
2909 return PTR_ERR(buf);
2910 iov[0].iov_base = buf;
2911 iov[0].iov_len = len;
2915 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2918 if (req->flags & REQ_F_BUFFER_SELECTED) {
2919 struct io_buffer *kbuf;
2921 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2922 iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
2923 iov[0].iov_len = kbuf->len;
2928 else if (req->rw.len > 1)
2931 #ifdef CONFIG_COMPAT
2932 if (req->ctx->compat)
2933 return io_compat_import(req, iov, needs_lock);
2936 return __io_iov_buffer_select(req, iov, needs_lock);
2939 static ssize_t __io_import_iovec(int rw, struct io_kiocb *req,
2940 struct iovec **iovec, struct iov_iter *iter,
2943 void __user *buf = u64_to_user_ptr(req->rw.addr);
2944 size_t sqe_len = req->rw.len;
2948 opcode = req->opcode;
2949 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
2951 return io_import_fixed(req, rw, iter);
2954 /* buffer index only valid with fixed read/write, or buffer select */
2955 if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT))
2958 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
2959 if (req->flags & REQ_F_BUFFER_SELECT) {
2960 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
2962 return PTR_ERR(buf);
2963 req->rw.len = sqe_len;
2966 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
2968 return ret < 0 ? ret : sqe_len;
2971 if (req->flags & REQ_F_BUFFER_SELECT) {
2972 ret = io_iov_buffer_select(req, *iovec, needs_lock);
2974 ret = (*iovec)->iov_len;
2975 iov_iter_init(iter, rw, *iovec, 1, ret);
2981 #ifdef CONFIG_COMPAT
2982 if (req->ctx->compat)
2983 return compat_import_iovec(rw, buf, sqe_len, UIO_FASTIOV,
2987 return import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter);
2990 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
2991 struct iovec **iovec, struct iov_iter *iter,
2994 struct io_async_rw *iorw = req->async_data;
2997 return __io_import_iovec(rw, req, iovec, iter, needs_lock);
2999 return iov_iter_count(&iorw->iter);
3002 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
3004 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
3008 * For files that don't have ->read_iter() and ->write_iter(), handle them
3009 * by looping over ->read() or ->write() manually.
3011 static ssize_t loop_rw_iter(int rw, struct file *file, struct kiocb *kiocb,
3012 struct iov_iter *iter)
3017 * Don't support polled IO through this interface, and we can't
3018 * support non-blocking either. For the latter, this just causes
3019 * the kiocb to be handled from an async context.
3021 if (kiocb->ki_flags & IOCB_HIPRI)
3023 if (kiocb->ki_flags & IOCB_NOWAIT)
3026 while (iov_iter_count(iter)) {
3030 if (!iov_iter_is_bvec(iter)) {
3031 iovec = iov_iter_iovec(iter);
3033 /* fixed buffers import bvec */
3034 iovec.iov_base = kmap(iter->bvec->bv_page)
3036 iovec.iov_len = min(iter->count,
3037 iter->bvec->bv_len - iter->iov_offset);
3041 nr = file->f_op->read(file, iovec.iov_base,
3042 iovec.iov_len, io_kiocb_ppos(kiocb));
3044 nr = file->f_op->write(file, iovec.iov_base,
3045 iovec.iov_len, io_kiocb_ppos(kiocb));
3048 if (iov_iter_is_bvec(iter))
3049 kunmap(iter->bvec->bv_page);
3057 if (nr != iovec.iov_len)
3059 iov_iter_advance(iter, nr);
3065 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
3066 const struct iovec *fast_iov, struct iov_iter *iter)
3068 struct io_async_rw *rw = req->async_data;
3070 memcpy(&rw->iter, iter, sizeof(*iter));
3071 rw->free_iovec = iovec;
3073 /* can only be fixed buffers, no need to do anything */
3074 if (iter->type == ITER_BVEC)
3077 unsigned iov_off = 0;
3079 rw->iter.iov = rw->fast_iov;
3080 if (iter->iov != fast_iov) {
3081 iov_off = iter->iov - fast_iov;
3082 rw->iter.iov += iov_off;
3084 if (rw->fast_iov != fast_iov)
3085 memcpy(rw->fast_iov + iov_off, fast_iov + iov_off,
3086 sizeof(struct iovec) * iter->nr_segs);
3088 req->flags |= REQ_F_NEED_CLEANUP;
3092 static inline int __io_alloc_async_data(struct io_kiocb *req)
3094 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
3095 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
3096 return req->async_data == NULL;
3099 static int io_alloc_async_data(struct io_kiocb *req)
3101 if (!io_op_defs[req->opcode].needs_async_data)
3104 return __io_alloc_async_data(req);
3107 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
3108 const struct iovec *fast_iov,
3109 struct iov_iter *iter, bool force)
3111 if (!force && !io_op_defs[req->opcode].needs_async_data)
3113 if (!req->async_data) {
3114 if (__io_alloc_async_data(req))
3117 io_req_map_rw(req, iovec, fast_iov, iter);
3122 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
3124 struct io_async_rw *iorw = req->async_data;
3125 struct iovec *iov = iorw->fast_iov;
3128 ret = __io_import_iovec(rw, req, &iov, &iorw->iter, false);
3129 if (unlikely(ret < 0))
3132 iorw->bytes_done = 0;
3133 iorw->free_iovec = iov;
3135 req->flags |= REQ_F_NEED_CLEANUP;
3139 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3143 ret = io_prep_rw(req, sqe);
3147 if (unlikely(!(req->file->f_mode & FMODE_READ)))
3150 /* either don't need iovec imported or already have it */
3151 if (!req->async_data)
3153 return io_rw_prep_async(req, READ);
3157 * This is our waitqueue callback handler, registered through lock_page_async()
3158 * when we initially tried to do the IO with the iocb armed our waitqueue.
3159 * This gets called when the page is unlocked, and we generally expect that to
3160 * happen when the page IO is completed and the page is now uptodate. This will
3161 * queue a task_work based retry of the operation, attempting to copy the data
3162 * again. If the latter fails because the page was NOT uptodate, then we will
3163 * do a thread based blocking retry of the operation. That's the unexpected
3166 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
3167 int sync, void *arg)
3169 struct wait_page_queue *wpq;
3170 struct io_kiocb *req = wait->private;
3171 struct wait_page_key *key = arg;
3174 wpq = container_of(wait, struct wait_page_queue, wait);
3176 if (!wake_page_match(wpq, key))
3179 req->rw.kiocb.ki_flags &= ~IOCB_WAITQ;
3180 list_del_init(&wait->entry);
3182 init_task_work(&req->task_work, io_req_task_submit);
3183 percpu_ref_get(&req->ctx->refs);
3185 /* submit ref gets dropped, acquire a new one */
3186 refcount_inc(&req->refs);
3187 ret = io_req_task_work_add(req, true);
3188 if (unlikely(ret)) {
3189 struct task_struct *tsk;
3191 /* queue just for cancelation */
3192 init_task_work(&req->task_work, io_req_task_cancel);
3193 tsk = io_wq_get_task(req->ctx->io_wq);
3194 task_work_add(tsk, &req->task_work, 0);
3195 wake_up_process(tsk);
3201 * This controls whether a given IO request should be armed for async page
3202 * based retry. If we return false here, the request is handed to the async
3203 * worker threads for retry. If we're doing buffered reads on a regular file,
3204 * we prepare a private wait_page_queue entry and retry the operation. This
3205 * will either succeed because the page is now uptodate and unlocked, or it
3206 * will register a callback when the page is unlocked at IO completion. Through
3207 * that callback, io_uring uses task_work to setup a retry of the operation.
3208 * That retry will attempt the buffered read again. The retry will generally
3209 * succeed, or in rare cases where it fails, we then fall back to using the
3210 * async worker threads for a blocking retry.
3212 static bool io_rw_should_retry(struct io_kiocb *req)
3214 struct io_async_rw *rw = req->async_data;
3215 struct wait_page_queue *wait = &rw->wpq;
3216 struct kiocb *kiocb = &req->rw.kiocb;
3218 /* never retry for NOWAIT, we just complete with -EAGAIN */
3219 if (req->flags & REQ_F_NOWAIT)
3222 /* Only for buffered IO */
3223 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
3227 * just use poll if we can, and don't attempt if the fs doesn't
3228 * support callback based unlocks
3230 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
3233 wait->wait.func = io_async_buf_func;
3234 wait->wait.private = req;
3235 wait->wait.flags = 0;
3236 INIT_LIST_HEAD(&wait->wait.entry);
3237 kiocb->ki_flags |= IOCB_WAITQ;
3238 kiocb->ki_flags &= ~IOCB_NOWAIT;
3239 kiocb->ki_waitq = wait;
3243 static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter)
3245 if (req->file->f_op->read_iter)
3246 return call_read_iter(req->file, &req->rw.kiocb, iter);
3247 else if (req->file->f_op->read)
3248 return loop_rw_iter(READ, req->file, &req->rw.kiocb, iter);
3253 static int io_read(struct io_kiocb *req, bool force_nonblock,
3254 struct io_comp_state *cs)
3256 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3257 struct kiocb *kiocb = &req->rw.kiocb;
3258 struct iov_iter __iter, *iter = &__iter;
3259 struct io_async_rw *rw = req->async_data;
3260 ssize_t io_size, ret, ret2;
3267 ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock);
3270 iov_count = iov_iter_count(iter);
3272 req->result = io_size;
3275 /* Ensure we clear previously set non-block flag */
3276 if (!force_nonblock)
3277 kiocb->ki_flags &= ~IOCB_NOWAIT;
3279 kiocb->ki_flags |= IOCB_NOWAIT;
3282 /* If the file doesn't support async, just async punt */
3283 no_async = force_nonblock && !io_file_supports_async(req->file, READ);
3287 ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), iov_count);
3291 ret = io_iter_do_read(req, iter);
3295 } else if (ret == -EIOCBQUEUED) {
3298 } else if (ret == -EAGAIN) {
3299 /* IOPOLL retry should happen for io-wq threads */
3300 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
3302 /* no retry on NONBLOCK marked file */
3303 if (req->file->f_flags & O_NONBLOCK)
3305 /* some cases will consume bytes even on error returns */
3306 iov_iter_revert(iter, iov_count - iov_iter_count(iter));
3309 } else if (ret < 0) {
3310 /* make sure -ERESTARTSYS -> -EINTR is done */
3314 /* read it all, or we did blocking attempt. no retry. */
3315 if (!iov_iter_count(iter) || !force_nonblock ||
3316 (req->file->f_flags & O_NONBLOCK))
3321 ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true);
3328 rw = req->async_data;
3329 /* it's copied and will be cleaned with ->io */
3331 /* now use our persistent iterator, if we aren't already */
3334 rw->bytes_done += ret;
3335 /* if we can retry, do so with the callbacks armed */
3336 if (!io_rw_should_retry(req)) {
3337 kiocb->ki_flags &= ~IOCB_WAITQ;
3342 * Now retry read with the IOCB_WAITQ parts set in the iocb. If we
3343 * get -EIOCBQUEUED, then we'll get a notification when the desired
3344 * page gets unlocked. We can also get a partial read here, and if we
3345 * do, then just retry at the new offset.
3347 ret = io_iter_do_read(req, iter);
3348 if (ret == -EIOCBQUEUED) {
3351 } else if (ret > 0 && ret < io_size) {
3352 /* we got some bytes, but not all. retry. */
3356 kiocb_done(kiocb, ret, cs);
3359 /* it's reportedly faster than delegating the null check to kfree() */
3365 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3369 ret = io_prep_rw(req, sqe);
3373 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
3376 /* either don't need iovec imported or already have it */
3377 if (!req->async_data)
3379 return io_rw_prep_async(req, WRITE);
3382 static int io_write(struct io_kiocb *req, bool force_nonblock,
3383 struct io_comp_state *cs)
3385 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3386 struct kiocb *kiocb = &req->rw.kiocb;
3387 struct iov_iter __iter, *iter = &__iter;
3388 struct io_async_rw *rw = req->async_data;
3390 ssize_t ret, ret2, io_size;
3395 ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock);
3398 iov_count = iov_iter_count(iter);
3400 req->result = io_size;
3402 /* Ensure we clear previously set non-block flag */
3403 if (!force_nonblock)
3404 kiocb->ki_flags &= ~IOCB_NOWAIT;
3406 kiocb->ki_flags |= IOCB_NOWAIT;
3408 /* If the file doesn't support async, just async punt */
3409 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
3412 /* file path doesn't support NOWAIT for non-direct_IO */
3413 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
3414 (req->flags & REQ_F_ISREG))
3417 ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), iov_count);
3422 * Open-code file_start_write here to grab freeze protection,
3423 * which will be released by another thread in
3424 * io_complete_rw(). Fool lockdep by telling it the lock got
3425 * released so that it doesn't complain about the held lock when
3426 * we return to userspace.
3428 if (req->flags & REQ_F_ISREG) {
3429 __sb_start_write(file_inode(req->file)->i_sb,
3430 SB_FREEZE_WRITE, true);
3431 __sb_writers_release(file_inode(req->file)->i_sb,
3434 kiocb->ki_flags |= IOCB_WRITE;
3436 if (req->file->f_op->write_iter)
3437 ret2 = call_write_iter(req->file, kiocb, iter);
3438 else if (req->file->f_op->write)
3439 ret2 = loop_rw_iter(WRITE, req->file, kiocb, iter);
3444 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
3445 * retry them without IOCB_NOWAIT.
3447 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
3449 /* no retry on NONBLOCK marked file */
3450 if (ret2 == -EAGAIN && (req->file->f_flags & O_NONBLOCK))
3452 if (!force_nonblock || ret2 != -EAGAIN) {
3453 /* IOPOLL retry should happen for io-wq threads */
3454 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN)
3457 kiocb_done(kiocb, ret2, cs);
3460 /* some cases will consume bytes even on error returns */
3461 iov_iter_revert(iter, iov_count - iov_iter_count(iter));
3462 ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false);
3467 /* it's reportedly faster than delegating the null check to kfree() */
3473 static int __io_splice_prep(struct io_kiocb *req,
3474 const struct io_uring_sqe *sqe)
3476 struct io_splice* sp = &req->splice;
3477 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
3480 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3484 sp->len = READ_ONCE(sqe->len);
3485 sp->flags = READ_ONCE(sqe->splice_flags);
3487 if (unlikely(sp->flags & ~valid_flags))
3490 ret = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in), &sp->file_in,
3491 (sp->flags & SPLICE_F_FD_IN_FIXED));
3494 req->flags |= REQ_F_NEED_CLEANUP;
3496 if (!S_ISREG(file_inode(sp->file_in)->i_mode)) {
3498 * Splice operation will be punted aync, and here need to
3499 * modify io_wq_work.flags, so initialize io_wq_work firstly.
3501 io_req_init_async(req);
3502 req->work.flags |= IO_WQ_WORK_UNBOUND;
3508 static int io_tee_prep(struct io_kiocb *req,
3509 const struct io_uring_sqe *sqe)
3511 if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off))
3513 return __io_splice_prep(req, sqe);
3516 static int io_tee(struct io_kiocb *req, bool force_nonblock)
3518 struct io_splice *sp = &req->splice;
3519 struct file *in = sp->file_in;
3520 struct file *out = sp->file_out;
3521 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3527 ret = do_tee(in, out, sp->len, flags);
3529 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3530 req->flags &= ~REQ_F_NEED_CLEANUP;
3533 req_set_fail_links(req);
3534 io_req_complete(req, ret);
3538 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3540 struct io_splice* sp = &req->splice;
3542 sp->off_in = READ_ONCE(sqe->splice_off_in);
3543 sp->off_out = READ_ONCE(sqe->off);
3544 return __io_splice_prep(req, sqe);
3547 static int io_splice(struct io_kiocb *req, bool force_nonblock)
3549 struct io_splice *sp = &req->splice;
3550 struct file *in = sp->file_in;
3551 struct file *out = sp->file_out;
3552 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
3553 loff_t *poff_in, *poff_out;
3559 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
3560 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
3563 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
3565 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
3566 req->flags &= ~REQ_F_NEED_CLEANUP;
3569 req_set_fail_links(req);
3570 io_req_complete(req, ret);
3575 * IORING_OP_NOP just posts a completion event, nothing else.
3577 static int io_nop(struct io_kiocb *req, struct io_comp_state *cs)
3579 struct io_ring_ctx *ctx = req->ctx;
3581 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3584 __io_req_complete(req, 0, 0, cs);
3588 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3590 struct io_ring_ctx *ctx = req->ctx;
3595 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3597 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3600 req->sync.flags = READ_ONCE(sqe->fsync_flags);
3601 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
3604 req->sync.off = READ_ONCE(sqe->off);
3605 req->sync.len = READ_ONCE(sqe->len);
3609 static int io_fsync(struct io_kiocb *req, bool force_nonblock)
3611 loff_t end = req->sync.off + req->sync.len;
3614 /* fsync always requires a blocking context */
3618 ret = vfs_fsync_range(req->file, req->sync.off,
3619 end > 0 ? end : LLONG_MAX,
3620 req->sync.flags & IORING_FSYNC_DATASYNC);
3622 req_set_fail_links(req);
3623 io_req_complete(req, ret);
3627 static int io_fallocate_prep(struct io_kiocb *req,
3628 const struct io_uring_sqe *sqe)
3630 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
3632 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3635 req->sync.off = READ_ONCE(sqe->off);
3636 req->sync.len = READ_ONCE(sqe->addr);
3637 req->sync.mode = READ_ONCE(sqe->len);
3641 static int io_fallocate(struct io_kiocb *req, bool force_nonblock)
3645 /* fallocate always requiring blocking context */
3648 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
3651 req_set_fail_links(req);
3652 io_req_complete(req, ret);
3656 static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3658 const char __user *fname;
3661 if (unlikely(sqe->ioprio || sqe->buf_index))
3663 if (unlikely(req->flags & REQ_F_FIXED_FILE))
3666 /* open.how should be already initialised */
3667 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3668 req->open.how.flags |= O_LARGEFILE;
3670 req->open.dfd = READ_ONCE(sqe->fd);
3671 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3672 req->open.filename = getname(fname);
3673 if (IS_ERR(req->open.filename)) {
3674 ret = PTR_ERR(req->open.filename);
3675 req->open.filename = NULL;
3678 req->open.nofile = rlimit(RLIMIT_NOFILE);
3679 req->flags |= REQ_F_NEED_CLEANUP;
3683 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3687 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3689 mode = READ_ONCE(sqe->len);
3690 flags = READ_ONCE(sqe->open_flags);
3691 req->open.how = build_open_how(flags, mode);
3692 return __io_openat_prep(req, sqe);
3695 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3697 struct open_how __user *how;
3701 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3703 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3704 len = READ_ONCE(sqe->len);
3705 if (len < OPEN_HOW_SIZE_VER0)
3708 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3713 return __io_openat_prep(req, sqe);
3716 static int io_openat2(struct io_kiocb *req, bool force_nonblock)
3718 struct open_flags op;
3725 ret = build_open_flags(&req->open.how, &op);
3729 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3733 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3736 ret = PTR_ERR(file);
3738 fsnotify_open(file);
3739 fd_install(ret, file);
3742 putname(req->open.filename);
3743 req->flags &= ~REQ_F_NEED_CLEANUP;
3745 req_set_fail_links(req);
3746 io_req_complete(req, ret);
3750 static int io_openat(struct io_kiocb *req, bool force_nonblock)
3752 return io_openat2(req, force_nonblock);
3755 static int io_remove_buffers_prep(struct io_kiocb *req,
3756 const struct io_uring_sqe *sqe)
3758 struct io_provide_buf *p = &req->pbuf;
3761 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3764 tmp = READ_ONCE(sqe->fd);
3765 if (!tmp || tmp > USHRT_MAX)
3768 memset(p, 0, sizeof(*p));
3770 p->bgid = READ_ONCE(sqe->buf_group);
3774 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3775 int bgid, unsigned nbufs)
3779 /* shouldn't happen */
3783 /* the head kbuf is the list itself */
3784 while (!list_empty(&buf->list)) {
3785 struct io_buffer *nxt;
3787 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3788 list_del(&nxt->list);
3795 idr_remove(&ctx->io_buffer_idr, bgid);
3800 static int io_remove_buffers(struct io_kiocb *req, bool force_nonblock,
3801 struct io_comp_state *cs)
3803 struct io_provide_buf *p = &req->pbuf;
3804 struct io_ring_ctx *ctx = req->ctx;
3805 struct io_buffer *head;
3808 io_ring_submit_lock(ctx, !force_nonblock);
3810 lockdep_assert_held(&ctx->uring_lock);
3813 head = idr_find(&ctx->io_buffer_idr, p->bgid);
3815 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3817 io_ring_submit_lock(ctx, !force_nonblock);
3819 req_set_fail_links(req);
3820 __io_req_complete(req, ret, 0, cs);
3824 static int io_provide_buffers_prep(struct io_kiocb *req,
3825 const struct io_uring_sqe *sqe)
3827 struct io_provide_buf *p = &req->pbuf;
3830 if (sqe->ioprio || sqe->rw_flags)
3833 tmp = READ_ONCE(sqe->fd);
3834 if (!tmp || tmp > USHRT_MAX)
3837 p->addr = READ_ONCE(sqe->addr);
3838 p->len = READ_ONCE(sqe->len);
3840 if (!access_ok(u64_to_user_ptr(p->addr), (p->len * p->nbufs)))
3843 p->bgid = READ_ONCE(sqe->buf_group);
3844 tmp = READ_ONCE(sqe->off);
3845 if (tmp > USHRT_MAX)
3851 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3853 struct io_buffer *buf;
3854 u64 addr = pbuf->addr;
3855 int i, bid = pbuf->bid;
3857 for (i = 0; i < pbuf->nbufs; i++) {
3858 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
3863 buf->len = pbuf->len;
3868 INIT_LIST_HEAD(&buf->list);
3871 list_add_tail(&buf->list, &(*head)->list);
3875 return i ? i : -ENOMEM;
3878 static int io_provide_buffers(struct io_kiocb *req, bool force_nonblock,
3879 struct io_comp_state *cs)
3881 struct io_provide_buf *p = &req->pbuf;
3882 struct io_ring_ctx *ctx = req->ctx;
3883 struct io_buffer *head, *list;
3886 io_ring_submit_lock(ctx, !force_nonblock);
3888 lockdep_assert_held(&ctx->uring_lock);
3890 list = head = idr_find(&ctx->io_buffer_idr, p->bgid);
3892 ret = io_add_buffers(p, &head);
3897 ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1,
3900 __io_remove_buffers(ctx, head, p->bgid, -1U);
3905 io_ring_submit_unlock(ctx, !force_nonblock);
3907 req_set_fail_links(req);
3908 __io_req_complete(req, ret, 0, cs);
3912 static int io_epoll_ctl_prep(struct io_kiocb *req,
3913 const struct io_uring_sqe *sqe)
3915 #if defined(CONFIG_EPOLL)
3916 if (sqe->ioprio || sqe->buf_index)
3918 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
3921 req->epoll.epfd = READ_ONCE(sqe->fd);
3922 req->epoll.op = READ_ONCE(sqe->len);
3923 req->epoll.fd = READ_ONCE(sqe->off);
3925 if (ep_op_has_event(req->epoll.op)) {
3926 struct epoll_event __user *ev;
3928 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
3929 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
3939 static int io_epoll_ctl(struct io_kiocb *req, bool force_nonblock,
3940 struct io_comp_state *cs)
3942 #if defined(CONFIG_EPOLL)
3943 struct io_epoll *ie = &req->epoll;
3946 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
3947 if (force_nonblock && ret == -EAGAIN)
3951 req_set_fail_links(req);
3952 __io_req_complete(req, ret, 0, cs);
3959 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3961 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3962 if (sqe->ioprio || sqe->buf_index || sqe->off)
3964 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3967 req->madvise.addr = READ_ONCE(sqe->addr);
3968 req->madvise.len = READ_ONCE(sqe->len);
3969 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
3976 static int io_madvise(struct io_kiocb *req, bool force_nonblock)
3978 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3979 struct io_madvise *ma = &req->madvise;
3985 ret = do_madvise(ma->addr, ma->len, ma->advice);
3987 req_set_fail_links(req);
3988 io_req_complete(req, ret);
3995 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3997 if (sqe->ioprio || sqe->buf_index || sqe->addr)
3999 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4002 req->fadvise.offset = READ_ONCE(sqe->off);
4003 req->fadvise.len = READ_ONCE(sqe->len);
4004 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
4008 static int io_fadvise(struct io_kiocb *req, bool force_nonblock)
4010 struct io_fadvise *fa = &req->fadvise;
4013 if (force_nonblock) {
4014 switch (fa->advice) {
4015 case POSIX_FADV_NORMAL:
4016 case POSIX_FADV_RANDOM:
4017 case POSIX_FADV_SEQUENTIAL:
4024 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
4026 req_set_fail_links(req);
4027 io_req_complete(req, ret);
4031 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4033 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL)))
4035 if (sqe->ioprio || sqe->buf_index)
4037 if (req->flags & REQ_F_FIXED_FILE)
4040 req->statx.dfd = READ_ONCE(sqe->fd);
4041 req->statx.mask = READ_ONCE(sqe->len);
4042 req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr));
4043 req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4044 req->statx.flags = READ_ONCE(sqe->statx_flags);
4049 static int io_statx(struct io_kiocb *req, bool force_nonblock)
4051 struct io_statx *ctx = &req->statx;
4054 if (force_nonblock) {
4055 /* only need file table for an actual valid fd */
4056 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
4057 req->flags |= REQ_F_NO_FILE_TABLE;
4061 ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask,
4065 req_set_fail_links(req);
4066 io_req_complete(req, ret);
4070 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4073 * If we queue this for async, it must not be cancellable. That would
4074 * leave the 'file' in an undeterminate state, and here need to modify
4075 * io_wq_work.flags, so initialize io_wq_work firstly.
4077 io_req_init_async(req);
4078 req->work.flags |= IO_WQ_WORK_NO_CANCEL;
4080 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4082 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
4083 sqe->rw_flags || sqe->buf_index)
4085 if (req->flags & REQ_F_FIXED_FILE)
4088 req->close.fd = READ_ONCE(sqe->fd);
4089 if ((req->file && req->file->f_op == &io_uring_fops))
4092 req->close.put_file = NULL;
4096 static int io_close(struct io_kiocb *req, bool force_nonblock,
4097 struct io_comp_state *cs)
4099 struct io_close *close = &req->close;
4102 /* might be already done during nonblock submission */
4103 if (!close->put_file) {
4104 ret = __close_fd_get_file(close->fd, &close->put_file);
4106 return (ret == -ENOENT) ? -EBADF : ret;
4109 /* if the file has a flush method, be safe and punt to async */
4110 if (close->put_file->f_op->flush && force_nonblock) {
4111 /* was never set, but play safe */
4112 req->flags &= ~REQ_F_NOWAIT;
4113 /* avoid grabbing files - we don't need the files */
4114 req->flags |= REQ_F_NO_FILE_TABLE;
4118 /* No ->flush() or already async, safely close from here */
4119 ret = filp_close(close->put_file, req->work.files);
4121 req_set_fail_links(req);
4122 fput(close->put_file);
4123 close->put_file = NULL;
4124 __io_req_complete(req, ret, 0, cs);
4128 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4130 struct io_ring_ctx *ctx = req->ctx;
4135 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
4137 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
4140 req->sync.off = READ_ONCE(sqe->off);
4141 req->sync.len = READ_ONCE(sqe->len);
4142 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
4146 static int io_sync_file_range(struct io_kiocb *req, bool force_nonblock)
4150 /* sync_file_range always requires a blocking context */
4154 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
4157 req_set_fail_links(req);
4158 io_req_complete(req, ret);
4162 #if defined(CONFIG_NET)
4163 static int io_setup_async_msg(struct io_kiocb *req,
4164 struct io_async_msghdr *kmsg)
4166 struct io_async_msghdr *async_msg = req->async_data;
4170 if (io_alloc_async_data(req)) {
4171 if (kmsg->iov != kmsg->fast_iov)
4175 async_msg = req->async_data;
4176 req->flags |= REQ_F_NEED_CLEANUP;
4177 memcpy(async_msg, kmsg, sizeof(*kmsg));
4181 static int io_sendmsg_copy_hdr(struct io_kiocb *req,
4182 struct io_async_msghdr *iomsg)
4184 iomsg->iov = iomsg->fast_iov;
4185 iomsg->msg.msg_name = &iomsg->addr;
4186 return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg,
4187 req->sr_msg.msg_flags, &iomsg->iov);
4190 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4192 struct io_async_msghdr *async_msg = req->async_data;
4193 struct io_sr_msg *sr = &req->sr_msg;
4196 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4199 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4200 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4201 sr->len = READ_ONCE(sqe->len);
4203 #ifdef CONFIG_COMPAT
4204 if (req->ctx->compat)
4205 sr->msg_flags |= MSG_CMSG_COMPAT;
4208 if (!async_msg || !io_op_defs[req->opcode].needs_async_data)
4210 ret = io_sendmsg_copy_hdr(req, async_msg);
4212 req->flags |= REQ_F_NEED_CLEANUP;
4216 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4217 struct io_comp_state *cs)
4219 struct io_async_msghdr iomsg, *kmsg;
4220 struct socket *sock;
4224 sock = sock_from_file(req->file, &ret);
4225 if (unlikely(!sock))
4228 if (req->async_data) {
4229 kmsg = req->async_data;
4230 kmsg->msg.msg_name = &kmsg->addr;
4231 /* if iov is set, it's allocated already */
4233 kmsg->iov = kmsg->fast_iov;
4234 kmsg->msg.msg_iter.iov = kmsg->iov;
4236 ret = io_sendmsg_copy_hdr(req, &iomsg);
4242 flags = req->sr_msg.msg_flags;
4243 if (flags & MSG_DONTWAIT)
4244 req->flags |= REQ_F_NOWAIT;
4245 else if (force_nonblock)
4246 flags |= MSG_DONTWAIT;
4248 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
4249 if (force_nonblock && ret == -EAGAIN)
4250 return io_setup_async_msg(req, kmsg);
4251 if (ret == -ERESTARTSYS)
4254 if (kmsg->iov != kmsg->fast_iov)
4256 req->flags &= ~REQ_F_NEED_CLEANUP;
4258 req_set_fail_links(req);
4259 __io_req_complete(req, ret, 0, cs);
4263 static int io_send(struct io_kiocb *req, bool force_nonblock,
4264 struct io_comp_state *cs)
4266 struct io_sr_msg *sr = &req->sr_msg;
4269 struct socket *sock;
4273 sock = sock_from_file(req->file, &ret);
4274 if (unlikely(!sock))
4277 ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter);
4281 msg.msg_name = NULL;
4282 msg.msg_control = NULL;
4283 msg.msg_controllen = 0;
4284 msg.msg_namelen = 0;
4286 flags = req->sr_msg.msg_flags;
4287 if (flags & MSG_DONTWAIT)
4288 req->flags |= REQ_F_NOWAIT;
4289 else if (force_nonblock)
4290 flags |= MSG_DONTWAIT;
4292 msg.msg_flags = flags;
4293 ret = sock_sendmsg(sock, &msg);
4294 if (force_nonblock && ret == -EAGAIN)
4296 if (ret == -ERESTARTSYS)
4300 req_set_fail_links(req);
4301 __io_req_complete(req, ret, 0, cs);
4305 static int __io_recvmsg_copy_hdr(struct io_kiocb *req,
4306 struct io_async_msghdr *iomsg)
4308 struct io_sr_msg *sr = &req->sr_msg;
4309 struct iovec __user *uiov;
4313 ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg,
4314 &iomsg->uaddr, &uiov, &iov_len);
4318 if (req->flags & REQ_F_BUFFER_SELECT) {
4321 if (copy_from_user(iomsg->iov, uiov, sizeof(*uiov)))
4323 sr->len = iomsg->iov[0].iov_len;
4324 iov_iter_init(&iomsg->msg.msg_iter, READ, iomsg->iov, 1,
4328 ret = import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
4329 &iomsg->iov, &iomsg->msg.msg_iter);
4337 #ifdef CONFIG_COMPAT
4338 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
4339 struct io_async_msghdr *iomsg)
4341 struct compat_msghdr __user *msg_compat;
4342 struct io_sr_msg *sr = &req->sr_msg;
4343 struct compat_iovec __user *uiov;
4348 msg_compat = (struct compat_msghdr __user *) sr->umsg;
4349 ret = __get_compat_msghdr(&iomsg->msg, msg_compat, &iomsg->uaddr,
4354 uiov = compat_ptr(ptr);
4355 if (req->flags & REQ_F_BUFFER_SELECT) {
4356 compat_ssize_t clen;
4360 if (!access_ok(uiov, sizeof(*uiov)))
4362 if (__get_user(clen, &uiov->iov_len))
4366 sr->len = iomsg->iov[0].iov_len;
4369 ret = compat_import_iovec(READ, uiov, len, UIO_FASTIOV,
4371 &iomsg->msg.msg_iter);
4380 static int io_recvmsg_copy_hdr(struct io_kiocb *req,
4381 struct io_async_msghdr *iomsg)
4383 iomsg->msg.msg_name = &iomsg->addr;
4384 iomsg->iov = iomsg->fast_iov;
4386 #ifdef CONFIG_COMPAT
4387 if (req->ctx->compat)
4388 return __io_compat_recvmsg_copy_hdr(req, iomsg);
4391 return __io_recvmsg_copy_hdr(req, iomsg);
4394 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
4397 struct io_sr_msg *sr = &req->sr_msg;
4398 struct io_buffer *kbuf;
4400 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
4405 req->flags |= REQ_F_BUFFER_SELECTED;
4409 static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req)
4411 return io_put_kbuf(req, req->sr_msg.kbuf);
4414 static int io_recvmsg_prep(struct io_kiocb *req,
4415 const struct io_uring_sqe *sqe)
4417 struct io_async_msghdr *async_msg = req->async_data;
4418 struct io_sr_msg *sr = &req->sr_msg;
4421 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4424 sr->msg_flags = READ_ONCE(sqe->msg_flags);
4425 sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr));
4426 sr->len = READ_ONCE(sqe->len);
4427 sr->bgid = READ_ONCE(sqe->buf_group);
4429 #ifdef CONFIG_COMPAT
4430 if (req->ctx->compat)
4431 sr->msg_flags |= MSG_CMSG_COMPAT;
4434 if (!async_msg || !io_op_defs[req->opcode].needs_async_data)
4436 ret = io_recvmsg_copy_hdr(req, async_msg);
4438 req->flags |= REQ_F_NEED_CLEANUP;
4442 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4443 struct io_comp_state *cs)
4445 struct io_async_msghdr iomsg, *kmsg;
4446 struct socket *sock;
4447 struct io_buffer *kbuf;
4449 int ret, cflags = 0;
4451 sock = sock_from_file(req->file, &ret);
4452 if (unlikely(!sock))
4455 if (req->async_data) {
4456 kmsg = req->async_data;
4457 kmsg->msg.msg_name = &kmsg->addr;
4458 /* if iov is set, it's allocated already */
4460 kmsg->iov = kmsg->fast_iov;
4461 kmsg->msg.msg_iter.iov = kmsg->iov;
4463 ret = io_recvmsg_copy_hdr(req, &iomsg);
4469 if (req->flags & REQ_F_BUFFER_SELECT) {
4470 kbuf = io_recv_buffer_select(req, !force_nonblock);
4472 return PTR_ERR(kbuf);
4473 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
4474 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->iov,
4475 1, req->sr_msg.len);
4478 flags = req->sr_msg.msg_flags;
4479 if (flags & MSG_DONTWAIT)
4480 req->flags |= REQ_F_NOWAIT;
4481 else if (force_nonblock)
4482 flags |= MSG_DONTWAIT;
4484 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg,
4485 kmsg->uaddr, flags);
4486 if (force_nonblock && ret == -EAGAIN)
4487 return io_setup_async_msg(req, kmsg);
4488 if (ret == -ERESTARTSYS)
4491 if (req->flags & REQ_F_BUFFER_SELECTED)
4492 cflags = io_put_recv_kbuf(req);
4493 if (kmsg->iov != kmsg->fast_iov)
4495 req->flags &= ~REQ_F_NEED_CLEANUP;
4497 req_set_fail_links(req);
4498 __io_req_complete(req, ret, cflags, cs);
4502 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4503 struct io_comp_state *cs)
4505 struct io_buffer *kbuf;
4506 struct io_sr_msg *sr = &req->sr_msg;
4508 void __user *buf = sr->buf;
4509 struct socket *sock;
4512 int ret, cflags = 0;
4514 sock = sock_from_file(req->file, &ret);
4515 if (unlikely(!sock))
4518 if (req->flags & REQ_F_BUFFER_SELECT) {
4519 kbuf = io_recv_buffer_select(req, !force_nonblock);
4521 return PTR_ERR(kbuf);
4522 buf = u64_to_user_ptr(kbuf->addr);
4525 ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter);
4529 msg.msg_name = NULL;
4530 msg.msg_control = NULL;
4531 msg.msg_controllen = 0;
4532 msg.msg_namelen = 0;
4533 msg.msg_iocb = NULL;
4536 flags = req->sr_msg.msg_flags;
4537 if (flags & MSG_DONTWAIT)
4538 req->flags |= REQ_F_NOWAIT;
4539 else if (force_nonblock)
4540 flags |= MSG_DONTWAIT;
4542 ret = sock_recvmsg(sock, &msg, flags);
4543 if (force_nonblock && ret == -EAGAIN)
4545 if (ret == -ERESTARTSYS)
4548 if (req->flags & REQ_F_BUFFER_SELECTED)
4549 cflags = io_put_recv_kbuf(req);
4551 req_set_fail_links(req);
4552 __io_req_complete(req, ret, cflags, cs);
4556 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4558 struct io_accept *accept = &req->accept;
4560 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4562 if (sqe->ioprio || sqe->len || sqe->buf_index)
4565 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4566 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
4567 accept->flags = READ_ONCE(sqe->accept_flags);
4568 accept->nofile = rlimit(RLIMIT_NOFILE);
4572 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4573 struct io_comp_state *cs)
4575 struct io_accept *accept = &req->accept;
4576 unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0;
4579 if (req->file->f_flags & O_NONBLOCK)
4580 req->flags |= REQ_F_NOWAIT;
4582 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
4583 accept->addr_len, accept->flags,
4585 if (ret == -EAGAIN && force_nonblock)
4588 if (ret == -ERESTARTSYS)
4590 req_set_fail_links(req);
4592 __io_req_complete(req, ret, 0, cs);
4596 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4598 struct io_connect *conn = &req->connect;
4599 struct io_async_connect *io = req->async_data;
4601 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4603 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4606 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4607 conn->addr_len = READ_ONCE(sqe->addr2);
4612 return move_addr_to_kernel(conn->addr, conn->addr_len,
4616 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4617 struct io_comp_state *cs)
4619 struct io_async_connect __io, *io;
4620 unsigned file_flags;
4623 if (req->async_data) {
4624 io = req->async_data;
4626 ret = move_addr_to_kernel(req->connect.addr,
4627 req->connect.addr_len,
4634 file_flags = force_nonblock ? O_NONBLOCK : 0;
4636 ret = __sys_connect_file(req->file, &io->address,
4637 req->connect.addr_len, file_flags);
4638 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4639 if (req->async_data)
4641 if (io_alloc_async_data(req)) {
4645 io = req->async_data;
4646 memcpy(req->async_data, &__io, sizeof(__io));
4649 if (ret == -ERESTARTSYS)
4653 req_set_fail_links(req);
4654 __io_req_complete(req, ret, 0, cs);
4657 #else /* !CONFIG_NET */
4658 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4663 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock,
4664 struct io_comp_state *cs)
4669 static int io_send(struct io_kiocb *req, bool force_nonblock,
4670 struct io_comp_state *cs)
4675 static int io_recvmsg_prep(struct io_kiocb *req,
4676 const struct io_uring_sqe *sqe)
4681 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock,
4682 struct io_comp_state *cs)
4687 static int io_recv(struct io_kiocb *req, bool force_nonblock,
4688 struct io_comp_state *cs)
4693 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4698 static int io_accept(struct io_kiocb *req, bool force_nonblock,
4699 struct io_comp_state *cs)
4704 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4709 static int io_connect(struct io_kiocb *req, bool force_nonblock,
4710 struct io_comp_state *cs)
4714 #endif /* CONFIG_NET */
4716 struct io_poll_table {
4717 struct poll_table_struct pt;
4718 struct io_kiocb *req;
4722 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4723 __poll_t mask, task_work_func_t func)
4728 /* for instances that support it check for an event match first: */
4729 if (mask && !(mask & poll->events))
4732 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4734 list_del_init(&poll->wait.entry);
4737 init_task_work(&req->task_work, func);
4738 percpu_ref_get(&req->ctx->refs);
4741 * If we using the signalfd wait_queue_head for this wakeup, then
4742 * it's not safe to use TWA_SIGNAL as we could be recursing on the
4743 * tsk->sighand->siglock on doing the wakeup. Should not be needed
4744 * either, as the normal wakeup will suffice.
4746 twa_signal_ok = (poll->head != &req->task->sighand->signalfd_wqh);
4749 * If this fails, then the task is exiting. When a task exits, the
4750 * work gets canceled, so just cancel this request as well instead
4751 * of executing it. We can't safely execute it anyway, as we may not
4752 * have the needed state needed for it anyway.
4754 ret = io_req_task_work_add(req, twa_signal_ok);
4755 if (unlikely(ret)) {
4756 struct task_struct *tsk;
4758 WRITE_ONCE(poll->canceled, true);
4759 tsk = io_wq_get_task(req->ctx->io_wq);
4760 task_work_add(tsk, &req->task_work, 0);
4761 wake_up_process(tsk);
4766 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4767 __acquires(&req->ctx->completion_lock)
4769 struct io_ring_ctx *ctx = req->ctx;
4771 if (!req->result && !READ_ONCE(poll->canceled)) {
4772 struct poll_table_struct pt = { ._key = poll->events };
4774 req->result = vfs_poll(req->file, &pt) & poll->events;
4777 spin_lock_irq(&ctx->completion_lock);
4778 if (!req->result && !READ_ONCE(poll->canceled)) {
4779 add_wait_queue(poll->head, &poll->wait);
4786 static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req)
4788 /* pure poll stashes this in ->async_data, poll driven retry elsewhere */
4789 if (req->opcode == IORING_OP_POLL_ADD)
4790 return req->async_data;
4791 return req->apoll->double_poll;
4794 static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req)
4796 if (req->opcode == IORING_OP_POLL_ADD)
4798 return &req->apoll->poll;
4801 static void io_poll_remove_double(struct io_kiocb *req)
4803 struct io_poll_iocb *poll = io_poll_get_double(req);
4805 lockdep_assert_held(&req->ctx->completion_lock);
4807 if (poll && poll->head) {
4808 struct wait_queue_head *head = poll->head;
4810 spin_lock(&head->lock);
4811 list_del_init(&poll->wait.entry);
4812 if (poll->wait.private)
4813 refcount_dec(&req->refs);
4815 spin_unlock(&head->lock);
4819 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
4821 struct io_ring_ctx *ctx = req->ctx;
4823 io_poll_remove_double(req);
4824 req->poll.done = true;
4825 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
4826 io_commit_cqring(ctx);
4829 static void io_poll_task_handler(struct io_kiocb *req, struct io_kiocb **nxt)
4831 struct io_ring_ctx *ctx = req->ctx;
4833 if (io_poll_rewait(req, &req->poll)) {
4834 spin_unlock_irq(&ctx->completion_lock);
4838 hash_del(&req->hash_node);
4839 io_poll_complete(req, req->result, 0);
4840 req->flags |= REQ_F_COMP_LOCKED;
4841 *nxt = io_put_req_find_next(req);
4842 spin_unlock_irq(&ctx->completion_lock);
4844 io_cqring_ev_posted(ctx);
4847 static void io_poll_task_func(struct callback_head *cb)
4849 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4850 struct io_ring_ctx *ctx = req->ctx;
4851 struct io_kiocb *nxt = NULL;
4853 io_poll_task_handler(req, &nxt);
4855 __io_req_task_submit(nxt);
4856 percpu_ref_put(&ctx->refs);
4859 static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode,
4860 int sync, void *key)
4862 struct io_kiocb *req = wait->private;
4863 struct io_poll_iocb *poll = io_poll_get_single(req);
4864 __poll_t mask = key_to_poll(key);
4866 /* for instances that support it check for an event match first: */
4867 if (mask && !(mask & poll->events))
4870 list_del_init(&wait->entry);
4872 if (poll && poll->head) {
4875 spin_lock(&poll->head->lock);
4876 done = list_empty(&poll->wait.entry);
4878 list_del_init(&poll->wait.entry);
4879 /* make sure double remove sees this as being gone */
4880 wait->private = NULL;
4881 spin_unlock(&poll->head->lock);
4883 __io_async_wake(req, poll, mask, io_poll_task_func);
4885 refcount_dec(&req->refs);
4889 static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events,
4890 wait_queue_func_t wake_func)
4894 poll->canceled = false;
4895 poll->events = events;
4896 INIT_LIST_HEAD(&poll->wait.entry);
4897 init_waitqueue_func_entry(&poll->wait, wake_func);
4900 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
4901 struct wait_queue_head *head,
4902 struct io_poll_iocb **poll_ptr)
4904 struct io_kiocb *req = pt->req;
4907 * If poll->head is already set, it's because the file being polled
4908 * uses multiple waitqueues for poll handling (eg one for read, one
4909 * for write). Setup a separate io_poll_iocb if this happens.
4911 if (unlikely(poll->head)) {
4912 /* already have a 2nd entry, fail a third attempt */
4914 pt->error = -EINVAL;
4917 poll = kmalloc(sizeof(*poll), GFP_ATOMIC);
4919 pt->error = -ENOMEM;
4922 io_init_poll_iocb(poll, req->poll.events, io_poll_double_wake);
4923 refcount_inc(&req->refs);
4924 poll->wait.private = req;
4931 if (poll->events & EPOLLEXCLUSIVE)
4932 add_wait_queue_exclusive(head, &poll->wait);
4934 add_wait_queue(head, &poll->wait);
4937 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
4938 struct poll_table_struct *p)
4940 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
4941 struct async_poll *apoll = pt->req->apoll;
4943 __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll);
4946 static void io_async_task_func(struct callback_head *cb)
4948 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4949 struct async_poll *apoll = req->apoll;
4950 struct io_ring_ctx *ctx = req->ctx;
4952 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
4954 if (io_poll_rewait(req, &apoll->poll)) {
4955 spin_unlock_irq(&ctx->completion_lock);
4956 percpu_ref_put(&ctx->refs);
4960 /* If req is still hashed, it cannot have been canceled. Don't check. */
4961 if (hash_hashed(&req->hash_node))
4962 hash_del(&req->hash_node);
4964 io_poll_remove_double(req);
4965 spin_unlock_irq(&ctx->completion_lock);
4967 if (!READ_ONCE(apoll->poll.canceled))
4968 __io_req_task_submit(req);
4970 __io_req_task_cancel(req, -ECANCELED);
4972 percpu_ref_put(&ctx->refs);
4973 kfree(apoll->double_poll);
4977 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
4980 struct io_kiocb *req = wait->private;
4981 struct io_poll_iocb *poll = &req->apoll->poll;
4983 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
4986 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
4989 static void io_poll_req_insert(struct io_kiocb *req)
4991 struct io_ring_ctx *ctx = req->ctx;
4992 struct hlist_head *list;
4994 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
4995 hlist_add_head(&req->hash_node, list);
4998 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
4999 struct io_poll_iocb *poll,
5000 struct io_poll_table *ipt, __poll_t mask,
5001 wait_queue_func_t wake_func)
5002 __acquires(&ctx->completion_lock)
5004 struct io_ring_ctx *ctx = req->ctx;
5005 bool cancel = false;
5007 io_init_poll_iocb(poll, mask, wake_func);
5008 poll->file = req->file;
5009 poll->wait.private = req;
5011 ipt->pt._key = mask;
5013 ipt->error = -EINVAL;
5015 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
5017 spin_lock_irq(&ctx->completion_lock);
5018 if (likely(poll->head)) {
5019 spin_lock(&poll->head->lock);
5020 if (unlikely(list_empty(&poll->wait.entry))) {
5026 if (mask || ipt->error)
5027 list_del_init(&poll->wait.entry);
5029 WRITE_ONCE(poll->canceled, true);
5030 else if (!poll->done) /* actually waiting for an event */
5031 io_poll_req_insert(req);
5032 spin_unlock(&poll->head->lock);
5038 static bool io_arm_poll_handler(struct io_kiocb *req)
5040 const struct io_op_def *def = &io_op_defs[req->opcode];
5041 struct io_ring_ctx *ctx = req->ctx;
5042 struct async_poll *apoll;
5043 struct io_poll_table ipt;
5047 if (!req->file || !file_can_poll(req->file))
5049 if (req->flags & REQ_F_POLLED)
5053 else if (def->pollout)
5057 /* if we can't nonblock try, then no point in arming a poll handler */
5058 if (!io_file_supports_async(req->file, rw))
5061 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
5062 if (unlikely(!apoll))
5064 apoll->double_poll = NULL;
5066 req->flags |= REQ_F_POLLED;
5068 INIT_HLIST_NODE(&req->hash_node);
5072 mask |= POLLIN | POLLRDNORM;
5074 mask |= POLLOUT | POLLWRNORM;
5075 mask |= POLLERR | POLLPRI;
5077 ipt.pt._qproc = io_async_queue_proc;
5079 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
5081 if (ret || ipt.error) {
5082 io_poll_remove_double(req);
5083 spin_unlock_irq(&ctx->completion_lock);
5084 kfree(apoll->double_poll);
5088 spin_unlock_irq(&ctx->completion_lock);
5089 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
5090 apoll->poll.events);
5094 static bool __io_poll_remove_one(struct io_kiocb *req,
5095 struct io_poll_iocb *poll)
5097 bool do_complete = false;
5099 spin_lock(&poll->head->lock);
5100 WRITE_ONCE(poll->canceled, true);
5101 if (!list_empty(&poll->wait.entry)) {
5102 list_del_init(&poll->wait.entry);
5105 spin_unlock(&poll->head->lock);
5106 hash_del(&req->hash_node);
5110 static bool io_poll_remove_one(struct io_kiocb *req)
5114 io_poll_remove_double(req);
5116 if (req->opcode == IORING_OP_POLL_ADD) {
5117 do_complete = __io_poll_remove_one(req, &req->poll);
5119 struct async_poll *apoll = req->apoll;
5121 /* non-poll requests have submit ref still */
5122 do_complete = __io_poll_remove_one(req, &apoll->poll);
5125 kfree(apoll->double_poll);
5131 io_cqring_fill_event(req, -ECANCELED);
5132 io_commit_cqring(req->ctx);
5133 req->flags |= REQ_F_COMP_LOCKED;
5134 req_set_fail_links(req);
5142 * Returns true if we found and killed one or more poll requests
5144 static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk)
5146 struct hlist_node *tmp;
5147 struct io_kiocb *req;
5150 spin_lock_irq(&ctx->completion_lock);
5151 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
5152 struct hlist_head *list;
5154 list = &ctx->cancel_hash[i];
5155 hlist_for_each_entry_safe(req, tmp, list, hash_node) {
5156 if (io_task_match(req, tsk))
5157 posted += io_poll_remove_one(req);
5160 spin_unlock_irq(&ctx->completion_lock);
5163 io_cqring_ev_posted(ctx);
5168 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
5170 struct hlist_head *list;
5171 struct io_kiocb *req;
5173 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
5174 hlist_for_each_entry(req, list, hash_node) {
5175 if (sqe_addr != req->user_data)
5177 if (io_poll_remove_one(req))
5185 static int io_poll_remove_prep(struct io_kiocb *req,
5186 const struct io_uring_sqe *sqe)
5188 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5190 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
5194 req->poll.addr = READ_ONCE(sqe->addr);
5199 * Find a running poll command that matches one specified in sqe->addr,
5200 * and remove it if found.
5202 static int io_poll_remove(struct io_kiocb *req)
5204 struct io_ring_ctx *ctx = req->ctx;
5208 addr = req->poll.addr;
5209 spin_lock_irq(&ctx->completion_lock);
5210 ret = io_poll_cancel(ctx, addr);
5211 spin_unlock_irq(&ctx->completion_lock);
5214 req_set_fail_links(req);
5215 io_req_complete(req, ret);
5219 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
5222 struct io_kiocb *req = wait->private;
5223 struct io_poll_iocb *poll = &req->poll;
5225 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
5228 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
5229 struct poll_table_struct *p)
5231 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
5233 __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->async_data);
5236 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5238 struct io_poll_iocb *poll = &req->poll;
5241 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5243 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
5248 events = READ_ONCE(sqe->poll32_events);
5250 events = swahw32(events);
5252 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP |
5253 (events & EPOLLEXCLUSIVE);
5257 static int io_poll_add(struct io_kiocb *req)
5259 struct io_poll_iocb *poll = &req->poll;
5260 struct io_ring_ctx *ctx = req->ctx;
5261 struct io_poll_table ipt;
5264 INIT_HLIST_NODE(&req->hash_node);
5265 ipt.pt._qproc = io_poll_queue_proc;
5267 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
5270 if (mask) { /* no async, we'd stolen it */
5272 io_poll_complete(req, mask, 0);
5274 spin_unlock_irq(&ctx->completion_lock);
5277 io_cqring_ev_posted(ctx);
5283 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
5285 struct io_timeout_data *data = container_of(timer,
5286 struct io_timeout_data, timer);
5287 struct io_kiocb *req = data->req;
5288 struct io_ring_ctx *ctx = req->ctx;
5289 unsigned long flags;
5291 spin_lock_irqsave(&ctx->completion_lock, flags);
5292 atomic_set(&req->ctx->cq_timeouts,
5293 atomic_read(&req->ctx->cq_timeouts) + 1);
5296 * We could be racing with timeout deletion. If the list is empty,
5297 * then timeout lookup already found it and will be handling it.
5299 if (!list_empty(&req->timeout.list))
5300 list_del_init(&req->timeout.list);
5302 io_cqring_fill_event(req, -ETIME);
5303 io_commit_cqring(ctx);
5304 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5306 io_cqring_ev_posted(ctx);
5307 req_set_fail_links(req);
5309 return HRTIMER_NORESTART;
5312 static int __io_timeout_cancel(struct io_kiocb *req)
5314 struct io_timeout_data *io = req->async_data;
5317 list_del_init(&req->timeout.list);
5319 ret = hrtimer_try_to_cancel(&io->timer);
5323 req_set_fail_links(req);
5324 req->flags |= REQ_F_COMP_LOCKED;
5325 io_cqring_fill_event(req, -ECANCELED);
5330 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
5332 struct io_kiocb *req;
5335 list_for_each_entry(req, &ctx->timeout_list, timeout.list) {
5336 if (user_data == req->user_data) {
5345 return __io_timeout_cancel(req);
5348 static int io_timeout_remove_prep(struct io_kiocb *req,
5349 const struct io_uring_sqe *sqe)
5351 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5353 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5355 if (sqe->ioprio || sqe->buf_index || sqe->len)
5358 req->timeout.addr = READ_ONCE(sqe->addr);
5359 req->timeout.flags = READ_ONCE(sqe->timeout_flags);
5360 if (req->timeout.flags)
5367 * Remove or update an existing timeout command
5369 static int io_timeout_remove(struct io_kiocb *req)
5371 struct io_ring_ctx *ctx = req->ctx;
5374 spin_lock_irq(&ctx->completion_lock);
5375 ret = io_timeout_cancel(ctx, req->timeout.addr);
5377 io_cqring_fill_event(req, ret);
5378 io_commit_cqring(ctx);
5379 spin_unlock_irq(&ctx->completion_lock);
5380 io_cqring_ev_posted(ctx);
5382 req_set_fail_links(req);
5387 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5388 bool is_timeout_link)
5390 struct io_timeout_data *data;
5392 u32 off = READ_ONCE(sqe->off);
5394 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5396 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
5398 if (off && is_timeout_link)
5400 flags = READ_ONCE(sqe->timeout_flags);
5401 if (flags & ~IORING_TIMEOUT_ABS)
5404 req->timeout.off = off;
5406 if (!req->async_data && io_alloc_async_data(req))
5409 data = req->async_data;
5412 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
5415 if (flags & IORING_TIMEOUT_ABS)
5416 data->mode = HRTIMER_MODE_ABS;
5418 data->mode = HRTIMER_MODE_REL;
5420 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
5424 static int io_timeout(struct io_kiocb *req)
5426 struct io_ring_ctx *ctx = req->ctx;
5427 struct io_timeout_data *data = req->async_data;
5428 struct list_head *entry;
5429 u32 tail, off = req->timeout.off;
5431 spin_lock_irq(&ctx->completion_lock);
5434 * sqe->off holds how many events that need to occur for this
5435 * timeout event to be satisfied. If it isn't set, then this is
5436 * a pure timeout request, sequence isn't used.
5438 if (io_is_timeout_noseq(req)) {
5439 entry = ctx->timeout_list.prev;
5443 tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
5444 req->timeout.target_seq = tail + off;
5447 * Insertion sort, ensuring the first entry in the list is always
5448 * the one we need first.
5450 list_for_each_prev(entry, &ctx->timeout_list) {
5451 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb,
5454 if (io_is_timeout_noseq(nxt))
5456 /* nxt.seq is behind @tail, otherwise would've been completed */
5457 if (off >= nxt->timeout.target_seq - tail)
5461 list_add(&req->timeout.list, entry);
5462 data->timer.function = io_timeout_fn;
5463 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
5464 spin_unlock_irq(&ctx->completion_lock);
5468 static bool io_cancel_cb(struct io_wq_work *work, void *data)
5470 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5472 return req->user_data == (unsigned long) data;
5475 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
5477 enum io_wq_cancel cancel_ret;
5480 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr, false);
5481 switch (cancel_ret) {
5482 case IO_WQ_CANCEL_OK:
5485 case IO_WQ_CANCEL_RUNNING:
5488 case IO_WQ_CANCEL_NOTFOUND:
5496 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
5497 struct io_kiocb *req, __u64 sqe_addr,
5500 unsigned long flags;
5503 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
5504 if (ret != -ENOENT) {
5505 spin_lock_irqsave(&ctx->completion_lock, flags);
5509 spin_lock_irqsave(&ctx->completion_lock, flags);
5510 ret = io_timeout_cancel(ctx, sqe_addr);
5513 ret = io_poll_cancel(ctx, sqe_addr);
5517 io_cqring_fill_event(req, ret);
5518 io_commit_cqring(ctx);
5519 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5520 io_cqring_ev_posted(ctx);
5523 req_set_fail_links(req);
5527 static int io_async_cancel_prep(struct io_kiocb *req,
5528 const struct io_uring_sqe *sqe)
5530 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
5532 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5534 if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags)
5537 req->cancel.addr = READ_ONCE(sqe->addr);
5541 static int io_async_cancel(struct io_kiocb *req)
5543 struct io_ring_ctx *ctx = req->ctx;
5545 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
5549 static int io_files_update_prep(struct io_kiocb *req,
5550 const struct io_uring_sqe *sqe)
5552 if (unlikely(req->ctx->flags & IORING_SETUP_SQPOLL))
5554 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT)))
5556 if (sqe->ioprio || sqe->rw_flags)
5559 req->files_update.offset = READ_ONCE(sqe->off);
5560 req->files_update.nr_args = READ_ONCE(sqe->len);
5561 if (!req->files_update.nr_args)
5563 req->files_update.arg = READ_ONCE(sqe->addr);
5567 static int io_files_update(struct io_kiocb *req, bool force_nonblock,
5568 struct io_comp_state *cs)
5570 struct io_ring_ctx *ctx = req->ctx;
5571 struct io_uring_files_update up;
5577 up.offset = req->files_update.offset;
5578 up.fds = req->files_update.arg;
5580 mutex_lock(&ctx->uring_lock);
5581 ret = __io_sqe_files_update(ctx, &up, req->files_update.nr_args);
5582 mutex_unlock(&ctx->uring_lock);
5585 req_set_fail_links(req);
5586 __io_req_complete(req, ret, 0, cs);
5590 static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5592 switch (req->opcode) {
5595 case IORING_OP_READV:
5596 case IORING_OP_READ_FIXED:
5597 case IORING_OP_READ:
5598 return io_read_prep(req, sqe);
5599 case IORING_OP_WRITEV:
5600 case IORING_OP_WRITE_FIXED:
5601 case IORING_OP_WRITE:
5602 return io_write_prep(req, sqe);
5603 case IORING_OP_POLL_ADD:
5604 return io_poll_add_prep(req, sqe);
5605 case IORING_OP_POLL_REMOVE:
5606 return io_poll_remove_prep(req, sqe);
5607 case IORING_OP_FSYNC:
5608 return io_prep_fsync(req, sqe);
5609 case IORING_OP_SYNC_FILE_RANGE:
5610 return io_prep_sfr(req, sqe);
5611 case IORING_OP_SENDMSG:
5612 case IORING_OP_SEND:
5613 return io_sendmsg_prep(req, sqe);
5614 case IORING_OP_RECVMSG:
5615 case IORING_OP_RECV:
5616 return io_recvmsg_prep(req, sqe);
5617 case IORING_OP_CONNECT:
5618 return io_connect_prep(req, sqe);
5619 case IORING_OP_TIMEOUT:
5620 return io_timeout_prep(req, sqe, false);
5621 case IORING_OP_TIMEOUT_REMOVE:
5622 return io_timeout_remove_prep(req, sqe);
5623 case IORING_OP_ASYNC_CANCEL:
5624 return io_async_cancel_prep(req, sqe);
5625 case IORING_OP_LINK_TIMEOUT:
5626 return io_timeout_prep(req, sqe, true);
5627 case IORING_OP_ACCEPT:
5628 return io_accept_prep(req, sqe);
5629 case IORING_OP_FALLOCATE:
5630 return io_fallocate_prep(req, sqe);
5631 case IORING_OP_OPENAT:
5632 return io_openat_prep(req, sqe);
5633 case IORING_OP_CLOSE:
5634 return io_close_prep(req, sqe);
5635 case IORING_OP_FILES_UPDATE:
5636 return io_files_update_prep(req, sqe);
5637 case IORING_OP_STATX:
5638 return io_statx_prep(req, sqe);
5639 case IORING_OP_FADVISE:
5640 return io_fadvise_prep(req, sqe);
5641 case IORING_OP_MADVISE:
5642 return io_madvise_prep(req, sqe);
5643 case IORING_OP_OPENAT2:
5644 return io_openat2_prep(req, sqe);
5645 case IORING_OP_EPOLL_CTL:
5646 return io_epoll_ctl_prep(req, sqe);
5647 case IORING_OP_SPLICE:
5648 return io_splice_prep(req, sqe);
5649 case IORING_OP_PROVIDE_BUFFERS:
5650 return io_provide_buffers_prep(req, sqe);
5651 case IORING_OP_REMOVE_BUFFERS:
5652 return io_remove_buffers_prep(req, sqe);
5654 return io_tee_prep(req, sqe);
5657 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5662 static int io_req_defer_prep(struct io_kiocb *req,
5663 const struct io_uring_sqe *sqe)
5669 if (io_alloc_async_data(req))
5672 ret = io_prep_work_files(req);
5676 io_prep_async_work(req);
5678 return io_req_prep(req, sqe);
5681 static u32 io_get_sequence(struct io_kiocb *req)
5683 struct io_kiocb *pos;
5684 struct io_ring_ctx *ctx = req->ctx;
5685 u32 total_submitted, nr_reqs = 1;
5687 if (req->flags & REQ_F_LINK_HEAD)
5688 list_for_each_entry(pos, &req->link_list, link_list)
5691 total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
5692 return total_submitted - nr_reqs;
5695 static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5697 struct io_ring_ctx *ctx = req->ctx;
5698 struct io_defer_entry *de;
5702 /* Still need defer if there is pending req in defer list. */
5703 if (likely(list_empty_careful(&ctx->defer_list) &&
5704 !(req->flags & REQ_F_IO_DRAIN)))
5707 seq = io_get_sequence(req);
5708 /* Still a chance to pass the sequence check */
5709 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list))
5712 if (!req->async_data) {
5713 ret = io_req_defer_prep(req, sqe);
5717 io_prep_async_link(req);
5718 de = kmalloc(sizeof(*de), GFP_KERNEL);
5722 spin_lock_irq(&ctx->completion_lock);
5723 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
5724 spin_unlock_irq(&ctx->completion_lock);
5726 io_queue_async_work(req);
5727 return -EIOCBQUEUED;
5730 trace_io_uring_defer(ctx, req, req->user_data);
5733 list_add_tail(&de->list, &ctx->defer_list);
5734 spin_unlock_irq(&ctx->completion_lock);
5735 return -EIOCBQUEUED;
5738 static void io_req_drop_files(struct io_kiocb *req)
5740 struct io_ring_ctx *ctx = req->ctx;
5741 unsigned long flags;
5743 spin_lock_irqsave(&ctx->inflight_lock, flags);
5744 list_del(&req->inflight_entry);
5745 if (waitqueue_active(&ctx->inflight_wait))
5746 wake_up(&ctx->inflight_wait);
5747 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
5748 req->flags &= ~REQ_F_INFLIGHT;
5749 put_files_struct(req->work.files);
5750 put_nsproxy(req->work.nsproxy);
5751 req->work.files = NULL;
5754 static void __io_clean_op(struct io_kiocb *req)
5756 if (req->flags & REQ_F_BUFFER_SELECTED) {
5757 switch (req->opcode) {
5758 case IORING_OP_READV:
5759 case IORING_OP_READ_FIXED:
5760 case IORING_OP_READ:
5761 kfree((void *)(unsigned long)req->rw.addr);
5763 case IORING_OP_RECVMSG:
5764 case IORING_OP_RECV:
5765 kfree(req->sr_msg.kbuf);
5768 req->flags &= ~REQ_F_BUFFER_SELECTED;
5771 if (req->flags & REQ_F_NEED_CLEANUP) {
5772 switch (req->opcode) {
5773 case IORING_OP_READV:
5774 case IORING_OP_READ_FIXED:
5775 case IORING_OP_READ:
5776 case IORING_OP_WRITEV:
5777 case IORING_OP_WRITE_FIXED:
5778 case IORING_OP_WRITE: {
5779 struct io_async_rw *io = req->async_data;
5781 kfree(io->free_iovec);
5784 case IORING_OP_RECVMSG:
5785 case IORING_OP_SENDMSG: {
5786 struct io_async_msghdr *io = req->async_data;
5787 if (io->iov != io->fast_iov)
5791 case IORING_OP_SPLICE:
5793 io_put_file(req, req->splice.file_in,
5794 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
5796 case IORING_OP_OPENAT:
5797 case IORING_OP_OPENAT2:
5798 if (req->open.filename)
5799 putname(req->open.filename);
5802 req->flags &= ~REQ_F_NEED_CLEANUP;
5805 if (req->flags & REQ_F_INFLIGHT)
5806 io_req_drop_files(req);
5809 static int io_issue_sqe(struct io_kiocb *req, bool force_nonblock,
5810 struct io_comp_state *cs)
5812 struct io_ring_ctx *ctx = req->ctx;
5815 switch (req->opcode) {
5817 ret = io_nop(req, cs);
5819 case IORING_OP_READV:
5820 case IORING_OP_READ_FIXED:
5821 case IORING_OP_READ:
5822 ret = io_read(req, force_nonblock, cs);
5824 case IORING_OP_WRITEV:
5825 case IORING_OP_WRITE_FIXED:
5826 case IORING_OP_WRITE:
5827 ret = io_write(req, force_nonblock, cs);
5829 case IORING_OP_FSYNC:
5830 ret = io_fsync(req, force_nonblock);
5832 case IORING_OP_POLL_ADD:
5833 ret = io_poll_add(req);
5835 case IORING_OP_POLL_REMOVE:
5836 ret = io_poll_remove(req);
5838 case IORING_OP_SYNC_FILE_RANGE:
5839 ret = io_sync_file_range(req, force_nonblock);
5841 case IORING_OP_SENDMSG:
5842 case IORING_OP_SEND:
5843 if (req->opcode == IORING_OP_SENDMSG)
5844 ret = io_sendmsg(req, force_nonblock, cs);
5846 ret = io_send(req, force_nonblock, cs);
5848 case IORING_OP_RECVMSG:
5849 case IORING_OP_RECV:
5850 if (req->opcode == IORING_OP_RECVMSG)
5851 ret = io_recvmsg(req, force_nonblock, cs);
5853 ret = io_recv(req, force_nonblock, cs);
5855 case IORING_OP_TIMEOUT:
5856 ret = io_timeout(req);
5858 case IORING_OP_TIMEOUT_REMOVE:
5859 ret = io_timeout_remove(req);
5861 case IORING_OP_ACCEPT:
5862 ret = io_accept(req, force_nonblock, cs);
5864 case IORING_OP_CONNECT:
5865 ret = io_connect(req, force_nonblock, cs);
5867 case IORING_OP_ASYNC_CANCEL:
5868 ret = io_async_cancel(req);
5870 case IORING_OP_FALLOCATE:
5871 ret = io_fallocate(req, force_nonblock);
5873 case IORING_OP_OPENAT:
5874 ret = io_openat(req, force_nonblock);
5876 case IORING_OP_CLOSE:
5877 ret = io_close(req, force_nonblock, cs);
5879 case IORING_OP_FILES_UPDATE:
5880 ret = io_files_update(req, force_nonblock, cs);
5882 case IORING_OP_STATX:
5883 ret = io_statx(req, force_nonblock);
5885 case IORING_OP_FADVISE:
5886 ret = io_fadvise(req, force_nonblock);
5888 case IORING_OP_MADVISE:
5889 ret = io_madvise(req, force_nonblock);
5891 case IORING_OP_OPENAT2:
5892 ret = io_openat2(req, force_nonblock);
5894 case IORING_OP_EPOLL_CTL:
5895 ret = io_epoll_ctl(req, force_nonblock, cs);
5897 case IORING_OP_SPLICE:
5898 ret = io_splice(req, force_nonblock);
5900 case IORING_OP_PROVIDE_BUFFERS:
5901 ret = io_provide_buffers(req, force_nonblock, cs);
5903 case IORING_OP_REMOVE_BUFFERS:
5904 ret = io_remove_buffers(req, force_nonblock, cs);
5907 ret = io_tee(req, force_nonblock);
5917 /* If the op doesn't have a file, we're not polling for it */
5918 if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) {
5919 const bool in_async = io_wq_current_is_worker();
5921 /* workqueue context doesn't hold uring_lock, grab it now */
5923 mutex_lock(&ctx->uring_lock);
5925 io_iopoll_req_issued(req);
5928 mutex_unlock(&ctx->uring_lock);
5934 static struct io_wq_work *io_wq_submit_work(struct io_wq_work *work)
5936 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5937 struct io_kiocb *timeout;
5940 timeout = io_prep_linked_timeout(req);
5942 io_queue_linked_timeout(timeout);
5944 /* if NO_CANCEL is set, we must still run the work */
5945 if ((work->flags & (IO_WQ_WORK_CANCEL|IO_WQ_WORK_NO_CANCEL)) ==
5946 IO_WQ_WORK_CANCEL) {
5952 ret = io_issue_sqe(req, false, NULL);
5954 * We can get EAGAIN for polled IO even though we're
5955 * forcing a sync submission from here, since we can't
5956 * wait for request slots on the block side.
5965 req_set_fail_links(req);
5966 io_req_complete(req, ret);
5969 return io_steal_work(req);
5972 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
5975 struct fixed_file_table *table;
5977 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
5978 return table->files[index & IORING_FILE_TABLE_MASK];
5981 static int io_file_get(struct io_submit_state *state, struct io_kiocb *req,
5982 int fd, struct file **out_file, bool fixed)
5984 struct io_ring_ctx *ctx = req->ctx;
5988 if (unlikely(!ctx->file_data ||
5989 (unsigned) fd >= ctx->nr_user_files))
5991 fd = array_index_nospec(fd, ctx->nr_user_files);
5992 file = io_file_from_index(ctx, fd);
5994 req->fixed_file_refs = ctx->file_data->cur_refs;
5995 percpu_ref_get(req->fixed_file_refs);
5998 trace_io_uring_file_get(ctx, fd);
5999 file = __io_file_get(state, fd);
6002 if (file || io_op_defs[req->opcode].needs_file_no_error) {
6009 static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req,
6014 fixed = (req->flags & REQ_F_FIXED_FILE) != 0;
6015 if (unlikely(!fixed && io_async_submit(req->ctx)))
6018 return io_file_get(state, req, fd, &req->file, fixed);
6021 static int io_grab_files(struct io_kiocb *req)
6023 struct io_ring_ctx *ctx = req->ctx;
6025 io_req_init_async(req);
6027 if (req->work.files || (req->flags & REQ_F_NO_FILE_TABLE))
6030 req->work.files = get_files_struct(current);
6031 get_nsproxy(current->nsproxy);
6032 req->work.nsproxy = current->nsproxy;
6033 req->flags |= REQ_F_INFLIGHT;
6035 spin_lock_irq(&ctx->inflight_lock);
6036 list_add(&req->inflight_entry, &ctx->inflight_list);
6037 spin_unlock_irq(&ctx->inflight_lock);
6041 static inline int io_prep_work_files(struct io_kiocb *req)
6043 if (!io_op_defs[req->opcode].file_table)
6045 return io_grab_files(req);
6048 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
6050 struct io_timeout_data *data = container_of(timer,
6051 struct io_timeout_data, timer);
6052 struct io_kiocb *req = data->req;
6053 struct io_ring_ctx *ctx = req->ctx;
6054 struct io_kiocb *prev = NULL;
6055 unsigned long flags;
6057 spin_lock_irqsave(&ctx->completion_lock, flags);
6060 * We don't expect the list to be empty, that will only happen if we
6061 * race with the completion of the linked work.
6063 if (!list_empty(&req->link_list)) {
6064 prev = list_entry(req->link_list.prev, struct io_kiocb,
6066 if (refcount_inc_not_zero(&prev->refs)) {
6067 list_del_init(&req->link_list);
6068 prev->flags &= ~REQ_F_LINK_TIMEOUT;
6073 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6076 req_set_fail_links(prev);
6077 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
6080 io_req_complete(req, -ETIME);
6082 return HRTIMER_NORESTART;
6085 static void __io_queue_linked_timeout(struct io_kiocb *req)
6088 * If the list is now empty, then our linked request finished before
6089 * we got a chance to setup the timer
6091 if (!list_empty(&req->link_list)) {
6092 struct io_timeout_data *data = req->async_data;
6094 data->timer.function = io_link_timeout_fn;
6095 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
6100 static void io_queue_linked_timeout(struct io_kiocb *req)
6102 struct io_ring_ctx *ctx = req->ctx;
6104 spin_lock_irq(&ctx->completion_lock);
6105 __io_queue_linked_timeout(req);
6106 spin_unlock_irq(&ctx->completion_lock);
6108 /* drop submission reference */
6112 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
6114 struct io_kiocb *nxt;
6116 if (!(req->flags & REQ_F_LINK_HEAD))
6118 if (req->flags & REQ_F_LINK_TIMEOUT)
6121 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb,
6123 if (!nxt || nxt->opcode != IORING_OP_LINK_TIMEOUT)
6126 req->flags |= REQ_F_LINK_TIMEOUT;
6130 static void __io_queue_sqe(struct io_kiocb *req, struct io_comp_state *cs)
6132 struct io_kiocb *linked_timeout;
6133 struct io_kiocb *nxt;
6134 const struct cred *old_creds = NULL;
6138 linked_timeout = io_prep_linked_timeout(req);
6140 if ((req->flags & REQ_F_WORK_INITIALIZED) && req->work.creds &&
6141 req->work.creds != current_cred()) {
6143 revert_creds(old_creds);
6144 if (old_creds == req->work.creds)
6145 old_creds = NULL; /* restored original creds */
6147 old_creds = override_creds(req->work.creds);
6150 ret = io_issue_sqe(req, true, cs);
6153 * We async punt it if the file wasn't marked NOWAIT, or if the file
6154 * doesn't support non-blocking read/write attempts
6156 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
6157 if (!io_arm_poll_handler(req)) {
6159 ret = io_prep_work_files(req);
6163 * Queued up for async execution, worker will release
6164 * submit reference when the iocb is actually submitted.
6166 io_queue_async_work(req);
6170 io_queue_linked_timeout(linked_timeout);
6174 if (unlikely(ret)) {
6176 /* un-prep timeout, so it'll be killed as any other linked */
6177 req->flags &= ~REQ_F_LINK_TIMEOUT;
6178 req_set_fail_links(req);
6180 io_req_complete(req, ret);
6184 /* drop submission reference */
6185 nxt = io_put_req_find_next(req);
6187 io_queue_linked_timeout(linked_timeout);
6192 if (req->flags & REQ_F_FORCE_ASYNC)
6198 revert_creds(old_creds);
6201 static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6202 struct io_comp_state *cs)
6206 ret = io_req_defer(req, sqe);
6208 if (ret != -EIOCBQUEUED) {
6210 req_set_fail_links(req);
6212 io_req_complete(req, ret);
6214 } else if (req->flags & REQ_F_FORCE_ASYNC) {
6215 if (!req->async_data) {
6216 ret = io_req_defer_prep(req, sqe);
6222 * Never try inline submit of IOSQE_ASYNC is set, go straight
6223 * to async execution.
6225 io_req_init_async(req);
6226 req->work.flags |= IO_WQ_WORK_CONCURRENT;
6227 io_queue_async_work(req);
6230 ret = io_req_prep(req, sqe);
6234 __io_queue_sqe(req, cs);
6238 static inline void io_queue_link_head(struct io_kiocb *req,
6239 struct io_comp_state *cs)
6241 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
6243 io_req_complete(req, -ECANCELED);
6245 io_queue_sqe(req, NULL, cs);
6248 static int io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
6249 struct io_kiocb **link, struct io_comp_state *cs)
6251 struct io_ring_ctx *ctx = req->ctx;
6255 * If we already have a head request, queue this one for async
6256 * submittal once the head completes. If we don't have a head but
6257 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
6258 * submitted sync once the chain is complete. If none of those
6259 * conditions are true (normal request), then just queue it.
6262 struct io_kiocb *head = *link;
6265 * Taking sequential execution of a link, draining both sides
6266 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
6267 * requests in the link. So, it drains the head and the
6268 * next after the link request. The last one is done via
6269 * drain_next flag to persist the effect across calls.
6271 if (req->flags & REQ_F_IO_DRAIN) {
6272 head->flags |= REQ_F_IO_DRAIN;
6273 ctx->drain_next = 1;
6275 ret = io_req_defer_prep(req, sqe);
6276 if (unlikely(ret)) {
6277 /* fail even hard links since we don't submit */
6278 head->flags |= REQ_F_FAIL_LINK;
6281 trace_io_uring_link(ctx, req, head);
6282 list_add_tail(&req->link_list, &head->link_list);
6284 /* last request of a link, enqueue the link */
6285 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
6286 io_queue_link_head(head, cs);
6290 if (unlikely(ctx->drain_next)) {
6291 req->flags |= REQ_F_IO_DRAIN;
6292 ctx->drain_next = 0;
6294 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
6295 req->flags |= REQ_F_LINK_HEAD;
6296 INIT_LIST_HEAD(&req->link_list);
6298 ret = io_req_defer_prep(req, sqe);
6300 req->flags |= REQ_F_FAIL_LINK;
6303 io_queue_sqe(req, sqe, cs);
6311 * Batched submission is done, ensure local IO is flushed out.
6313 static void io_submit_state_end(struct io_submit_state *state)
6315 if (!list_empty(&state->comp.list))
6316 io_submit_flush_completions(&state->comp);
6317 blk_finish_plug(&state->plug);
6318 io_state_file_put(state);
6319 if (state->free_reqs)
6320 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
6324 * Start submission side cache.
6326 static void io_submit_state_start(struct io_submit_state *state,
6327 struct io_ring_ctx *ctx, unsigned int max_ios)
6329 blk_start_plug(&state->plug);
6331 INIT_LIST_HEAD(&state->comp.list);
6332 state->comp.ctx = ctx;
6333 state->free_reqs = 0;
6335 state->ios_left = max_ios;
6338 static void io_commit_sqring(struct io_ring_ctx *ctx)
6340 struct io_rings *rings = ctx->rings;
6343 * Ensure any loads from the SQEs are done at this point,
6344 * since once we write the new head, the application could
6345 * write new data to them.
6347 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
6351 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
6352 * that is mapped by userspace. This means that care needs to be taken to
6353 * ensure that reads are stable, as we cannot rely on userspace always
6354 * being a good citizen. If members of the sqe are validated and then later
6355 * used, it's important that those reads are done through READ_ONCE() to
6356 * prevent a re-load down the line.
6358 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
6360 u32 *sq_array = ctx->sq_array;
6364 * The cached sq head (or cq tail) serves two purposes:
6366 * 1) allows us to batch the cost of updating the user visible
6368 * 2) allows the kernel side to track the head on its own, even
6369 * though the application is the one updating it.
6371 head = READ_ONCE(sq_array[ctx->cached_sq_head & ctx->sq_mask]);
6372 if (likely(head < ctx->sq_entries))
6373 return &ctx->sq_sqes[head];
6375 /* drop invalid entries */
6376 ctx->cached_sq_dropped++;
6377 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
6381 static inline void io_consume_sqe(struct io_ring_ctx *ctx)
6383 ctx->cached_sq_head++;
6387 * Check SQE restrictions (opcode and flags).
6389 * Returns 'true' if SQE is allowed, 'false' otherwise.
6391 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
6392 struct io_kiocb *req,
6393 unsigned int sqe_flags)
6395 if (!ctx->restricted)
6398 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
6401 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
6402 ctx->restrictions.sqe_flags_required)
6405 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
6406 ctx->restrictions.sqe_flags_required))
6412 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
6413 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
6414 IOSQE_BUFFER_SELECT)
6416 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
6417 const struct io_uring_sqe *sqe,
6418 struct io_submit_state *state)
6420 unsigned int sqe_flags;
6423 req->opcode = READ_ONCE(sqe->opcode);
6424 req->user_data = READ_ONCE(sqe->user_data);
6425 req->async_data = NULL;
6429 /* one is dropped after submission, the other at completion */
6430 refcount_set(&req->refs, 2);
6431 req->task = current;
6434 if (unlikely(req->opcode >= IORING_OP_LAST))
6437 if (unlikely(io_sq_thread_acquire_mm(ctx, req)))
6440 sqe_flags = READ_ONCE(sqe->flags);
6441 /* enforce forwards compatibility on users */
6442 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
6445 if (unlikely(!io_check_restriction(ctx, req, sqe_flags)))
6448 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
6449 !io_op_defs[req->opcode].buffer_select)
6452 id = READ_ONCE(sqe->personality);
6454 io_req_init_async(req);
6455 req->work.creds = idr_find(&ctx->personality_idr, id);
6456 if (unlikely(!req->work.creds))
6458 get_cred(req->work.creds);
6461 /* same numerical values with corresponding REQ_F_*, safe to copy */
6462 req->flags |= sqe_flags;
6464 if (!io_op_defs[req->opcode].needs_file)
6467 return io_req_set_file(state, req, READ_ONCE(sqe->fd));
6470 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
6472 struct io_submit_state state;
6473 struct io_kiocb *link = NULL;
6474 int i, submitted = 0;
6476 /* if we have a backlog and couldn't flush it all, return BUSY */
6477 if (test_bit(0, &ctx->sq_check_overflow)) {
6478 if (!list_empty(&ctx->cq_overflow_list) &&
6479 !io_cqring_overflow_flush(ctx, false, NULL, NULL))
6483 /* make sure SQ entry isn't read before tail */
6484 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
6486 if (!percpu_ref_tryget_many(&ctx->refs, nr))
6489 atomic_long_add(nr, ¤t->io_uring->req_issue);
6490 refcount_add(nr, ¤t->usage);
6492 io_submit_state_start(&state, ctx, nr);
6494 for (i = 0; i < nr; i++) {
6495 const struct io_uring_sqe *sqe;
6496 struct io_kiocb *req;
6499 sqe = io_get_sqe(ctx);
6500 if (unlikely(!sqe)) {
6501 io_consume_sqe(ctx);
6504 req = io_alloc_req(ctx, &state);
6505 if (unlikely(!req)) {
6507 submitted = -EAGAIN;
6511 err = io_init_req(ctx, req, sqe, &state);
6512 io_consume_sqe(ctx);
6513 /* will complete beyond this point, count as submitted */
6516 if (unlikely(err)) {
6519 io_req_complete(req, err);
6523 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
6524 true, io_async_submit(ctx));
6525 err = io_submit_sqe(req, sqe, &link, &state.comp);
6530 if (unlikely(submitted != nr)) {
6531 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
6533 percpu_ref_put_many(&ctx->refs, nr - ref_used);
6534 atomic_long_sub(nr - ref_used, ¤t->io_uring->req_issue);
6535 put_task_struct_many(current, nr - ref_used);
6538 io_queue_link_head(link, &state.comp);
6539 io_submit_state_end(&state);
6541 /* Commit SQ ring head once we've consumed and submitted all SQEs */
6542 io_commit_sqring(ctx);
6547 static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx)
6549 /* Tell userspace we may need a wakeup call */
6550 spin_lock_irq(&ctx->completion_lock);
6551 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6552 spin_unlock_irq(&ctx->completion_lock);
6555 static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx)
6557 spin_lock_irq(&ctx->completion_lock);
6558 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6559 spin_unlock_irq(&ctx->completion_lock);
6562 static int io_sq_wake_function(struct wait_queue_entry *wqe, unsigned mode,
6563 int sync, void *key)
6565 struct io_ring_ctx *ctx = container_of(wqe, struct io_ring_ctx, sqo_wait_entry);
6568 ret = autoremove_wake_function(wqe, mode, sync, key);
6570 unsigned long flags;
6572 spin_lock_irqsave(&ctx->completion_lock, flags);
6573 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6574 spin_unlock_irqrestore(&ctx->completion_lock, flags);
6585 static enum sq_ret __io_sq_thread(struct io_ring_ctx *ctx,
6586 unsigned long start_jiffies, bool cap_entries)
6588 unsigned long timeout = start_jiffies + ctx->sq_thread_idle;
6589 struct io_sq_data *sqd = ctx->sq_data;
6590 unsigned int to_submit;
6594 if (!list_empty(&ctx->iopoll_list)) {
6595 unsigned nr_events = 0;
6597 mutex_lock(&ctx->uring_lock);
6598 if (!list_empty(&ctx->iopoll_list) && !need_resched())
6599 io_do_iopoll(ctx, &nr_events, 0);
6600 mutex_unlock(&ctx->uring_lock);
6603 to_submit = io_sqring_entries(ctx);
6606 * If submit got -EBUSY, flag us as needing the application
6607 * to enter the kernel to reap and flush events.
6609 if (!to_submit || ret == -EBUSY || need_resched()) {
6611 * Drop cur_mm before scheduling, we can't hold it for
6612 * long periods (or over schedule()). Do this before
6613 * adding ourselves to the waitqueue, as the unuse/drop
6616 io_sq_thread_drop_mm();
6619 * We're polling. If we're within the defined idle
6620 * period, then let us spin without work before going
6621 * to sleep. The exception is if we got EBUSY doing
6622 * more IO, we should wait for the application to
6623 * reap events and wake us up.
6625 if (!list_empty(&ctx->iopoll_list) || need_resched() ||
6626 (!time_after(jiffies, timeout) && ret != -EBUSY &&
6627 !percpu_ref_is_dying(&ctx->refs)))
6630 prepare_to_wait(&sqd->wait, &ctx->sqo_wait_entry,
6631 TASK_INTERRUPTIBLE);
6634 * While doing polled IO, before going to sleep, we need
6635 * to check if there are new reqs added to iopoll_list,
6636 * it is because reqs may have been punted to io worker
6637 * and will be added to iopoll_list later, hence check
6638 * the iopoll_list again.
6640 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6641 !list_empty_careful(&ctx->iopoll_list)) {
6642 finish_wait(&sqd->wait, &ctx->sqo_wait_entry);
6646 to_submit = io_sqring_entries(ctx);
6647 if (!to_submit || ret == -EBUSY)
6651 finish_wait(&sqd->wait, &ctx->sqo_wait_entry);
6652 io_ring_clear_wakeup_flag(ctx);
6654 /* if we're handling multiple rings, cap submit size for fairness */
6655 if (cap_entries && to_submit > 8)
6658 mutex_lock(&ctx->uring_lock);
6659 if (likely(!percpu_ref_is_dying(&ctx->refs)))
6660 ret = io_submit_sqes(ctx, to_submit);
6661 mutex_unlock(&ctx->uring_lock);
6663 if (!io_sqring_full(ctx) && wq_has_sleeper(&ctx->sqo_sq_wait))
6664 wake_up(&ctx->sqo_sq_wait);
6666 return SQT_DID_WORK;
6669 static void io_sqd_init_new(struct io_sq_data *sqd)
6671 struct io_ring_ctx *ctx;
6673 while (!list_empty(&sqd->ctx_new_list)) {
6674 ctx = list_first_entry(&sqd->ctx_new_list, struct io_ring_ctx, sqd_list);
6675 init_wait(&ctx->sqo_wait_entry);
6676 ctx->sqo_wait_entry.func = io_sq_wake_function;
6677 list_move_tail(&ctx->sqd_list, &sqd->ctx_list);
6678 complete(&ctx->sq_thread_comp);
6682 static int io_sq_thread(void *data)
6684 struct cgroup_subsys_state *cur_css = NULL;
6685 const struct cred *old_cred = NULL;
6686 struct io_sq_data *sqd = data;
6687 struct io_ring_ctx *ctx;
6688 unsigned long start_jiffies;
6690 start_jiffies = jiffies;
6691 while (!kthread_should_stop()) {
6692 enum sq_ret ret = 0;
6696 * Any changes to the sqd lists are synchronized through the
6697 * kthread parking. This synchronizes the thread vs users,
6698 * the users are synchronized on the sqd->ctx_lock.
6700 if (kthread_should_park())
6703 if (unlikely(!list_empty(&sqd->ctx_new_list)))
6704 io_sqd_init_new(sqd);
6706 cap_entries = !list_is_singular(&sqd->ctx_list);
6708 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
6709 if (current->cred != ctx->creds) {
6711 revert_creds(old_cred);
6712 old_cred = override_creds(ctx->creds);
6714 io_sq_thread_associate_blkcg(ctx, &cur_css);
6716 ret |= __io_sq_thread(ctx, start_jiffies, cap_entries);
6718 io_sq_thread_drop_mm();
6721 if (ret & SQT_SPIN) {
6724 } else if (ret == SQT_IDLE) {
6725 if (kthread_should_park())
6727 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6728 io_ring_set_wakeup_flag(ctx);
6730 start_jiffies = jiffies;
6731 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
6732 io_ring_clear_wakeup_flag(ctx);
6739 io_sq_thread_unassociate_blkcg();
6741 revert_creds(old_cred);
6748 struct io_wait_queue {
6749 struct wait_queue_entry wq;
6750 struct io_ring_ctx *ctx;
6752 unsigned nr_timeouts;
6755 static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush)
6757 struct io_ring_ctx *ctx = iowq->ctx;
6760 * Wake up if we have enough events, or if a timeout occurred since we
6761 * started waiting. For timeouts, we always want to return to userspace,
6762 * regardless of event count.
6764 return io_cqring_events(ctx, noflush) >= iowq->to_wait ||
6765 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6768 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6769 int wake_flags, void *key)
6771 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6774 /* use noflush == true, as we can't safely rely on locking context */
6775 if (!io_should_wake(iowq, true))
6778 return autoremove_wake_function(curr, mode, wake_flags, key);
6781 static int io_run_task_work_sig(void)
6783 if (io_run_task_work())
6785 if (!signal_pending(current))
6787 if (current->jobctl & JOBCTL_TASK_WORK) {
6788 spin_lock_irq(¤t->sighand->siglock);
6789 current->jobctl &= ~JOBCTL_TASK_WORK;
6790 recalc_sigpending();
6791 spin_unlock_irq(¤t->sighand->siglock);
6798 * Wait until events become available, if we don't already have some. The
6799 * application must reap them itself, as they reside on the shared cq ring.
6801 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6802 const sigset_t __user *sig, size_t sigsz)
6804 struct io_wait_queue iowq = {
6807 .func = io_wake_function,
6808 .entry = LIST_HEAD_INIT(iowq.wq.entry),
6811 .to_wait = min_events,
6813 struct io_rings *rings = ctx->rings;
6817 if (io_cqring_events(ctx, false) >= min_events)
6819 if (!io_run_task_work())
6824 #ifdef CONFIG_COMPAT
6825 if (in_compat_syscall())
6826 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
6830 ret = set_user_sigmask(sig, sigsz);
6836 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
6837 trace_io_uring_cqring_wait(ctx, min_events);
6839 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
6840 TASK_INTERRUPTIBLE);
6841 /* make sure we run task_work before checking for signals */
6842 ret = io_run_task_work_sig();
6847 if (io_should_wake(&iowq, false))
6851 finish_wait(&ctx->wait, &iowq.wq);
6853 restore_saved_sigmask_unless(ret == -EINTR);
6855 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
6858 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
6860 #if defined(CONFIG_UNIX)
6861 if (ctx->ring_sock) {
6862 struct sock *sock = ctx->ring_sock->sk;
6863 struct sk_buff *skb;
6865 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
6871 for (i = 0; i < ctx->nr_user_files; i++) {
6874 file = io_file_from_index(ctx, i);
6881 static void io_file_ref_kill(struct percpu_ref *ref)
6883 struct fixed_file_data *data;
6885 data = container_of(ref, struct fixed_file_data, refs);
6886 complete(&data->done);
6889 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
6891 struct fixed_file_data *data = ctx->file_data;
6892 struct fixed_file_ref_node *ref_node = NULL;
6893 unsigned nr_tables, i;
6898 spin_lock(&data->lock);
6899 if (!list_empty(&data->ref_list))
6900 ref_node = list_first_entry(&data->ref_list,
6901 struct fixed_file_ref_node, node);
6902 spin_unlock(&data->lock);
6904 percpu_ref_kill(&ref_node->refs);
6906 percpu_ref_kill(&data->refs);
6908 /* wait for all refs nodes to complete */
6909 flush_delayed_work(&ctx->file_put_work);
6910 wait_for_completion(&data->done);
6912 __io_sqe_files_unregister(ctx);
6913 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
6914 for (i = 0; i < nr_tables; i++)
6915 kfree(data->table[i].files);
6917 percpu_ref_exit(&data->refs);
6919 ctx->file_data = NULL;
6920 ctx->nr_user_files = 0;
6924 static void io_put_sq_data(struct io_sq_data *sqd)
6926 if (refcount_dec_and_test(&sqd->refs)) {
6928 * The park is a bit of a work-around, without it we get
6929 * warning spews on shutdown with SQPOLL set and affinity
6930 * set to a single CPU.
6933 kthread_park(sqd->thread);
6934 kthread_stop(sqd->thread);
6941 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
6943 struct io_ring_ctx *ctx_attach;
6944 struct io_sq_data *sqd;
6947 f = fdget(p->wq_fd);
6949 return ERR_PTR(-ENXIO);
6950 if (f.file->f_op != &io_uring_fops) {
6952 return ERR_PTR(-EINVAL);
6955 ctx_attach = f.file->private_data;
6956 sqd = ctx_attach->sq_data;
6959 return ERR_PTR(-EINVAL);
6962 refcount_inc(&sqd->refs);
6967 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p)
6969 struct io_sq_data *sqd;
6971 if (p->flags & IORING_SETUP_ATTACH_WQ)
6972 return io_attach_sq_data(p);
6974 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
6976 return ERR_PTR(-ENOMEM);
6978 refcount_set(&sqd->refs, 1);
6979 INIT_LIST_HEAD(&sqd->ctx_list);
6980 INIT_LIST_HEAD(&sqd->ctx_new_list);
6981 mutex_init(&sqd->ctx_lock);
6982 mutex_init(&sqd->lock);
6983 init_waitqueue_head(&sqd->wait);
6987 static void io_sq_thread_unpark(struct io_sq_data *sqd)
6988 __releases(&sqd->lock)
6992 kthread_unpark(sqd->thread);
6993 mutex_unlock(&sqd->lock);
6996 static void io_sq_thread_park(struct io_sq_data *sqd)
6997 __acquires(&sqd->lock)
7001 mutex_lock(&sqd->lock);
7002 kthread_park(sqd->thread);
7005 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
7007 struct io_sq_data *sqd = ctx->sq_data;
7012 * We may arrive here from the error branch in
7013 * io_sq_offload_create() where the kthread is created
7014 * without being waked up, thus wake it up now to make
7015 * sure the wait will complete.
7017 wake_up_process(sqd->thread);
7018 wait_for_completion(&ctx->sq_thread_comp);
7020 io_sq_thread_park(sqd);
7023 mutex_lock(&sqd->ctx_lock);
7024 list_del(&ctx->sqd_list);
7025 mutex_unlock(&sqd->ctx_lock);
7028 finish_wait(&sqd->wait, &ctx->sqo_wait_entry);
7029 io_sq_thread_unpark(sqd);
7032 io_put_sq_data(sqd);
7033 ctx->sq_data = NULL;
7037 static void io_finish_async(struct io_ring_ctx *ctx)
7039 io_sq_thread_stop(ctx);
7042 io_wq_destroy(ctx->io_wq);
7047 #if defined(CONFIG_UNIX)
7049 * Ensure the UNIX gc is aware of our file set, so we are certain that
7050 * the io_uring can be safely unregistered on process exit, even if we have
7051 * loops in the file referencing.
7053 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
7055 struct sock *sk = ctx->ring_sock->sk;
7056 struct scm_fp_list *fpl;
7057 struct sk_buff *skb;
7060 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
7064 skb = alloc_skb(0, GFP_KERNEL);
7073 fpl->user = get_uid(ctx->user);
7074 for (i = 0; i < nr; i++) {
7075 struct file *file = io_file_from_index(ctx, i + offset);
7079 fpl->fp[nr_files] = get_file(file);
7080 unix_inflight(fpl->user, fpl->fp[nr_files]);
7085 fpl->max = SCM_MAX_FD;
7086 fpl->count = nr_files;
7087 UNIXCB(skb).fp = fpl;
7088 skb->destructor = unix_destruct_scm;
7089 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
7090 skb_queue_head(&sk->sk_receive_queue, skb);
7092 for (i = 0; i < nr_files; i++)
7103 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
7104 * causes regular reference counting to break down. We rely on the UNIX
7105 * garbage collection to take care of this problem for us.
7107 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7109 unsigned left, total;
7113 left = ctx->nr_user_files;
7115 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
7117 ret = __io_sqe_files_scm(ctx, this_files, total);
7121 total += this_files;
7127 while (total < ctx->nr_user_files) {
7128 struct file *file = io_file_from_index(ctx, total);
7138 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
7144 static int io_sqe_alloc_file_tables(struct io_ring_ctx *ctx, unsigned nr_tables,
7149 for (i = 0; i < nr_tables; i++) {
7150 struct fixed_file_table *table = &ctx->file_data->table[i];
7151 unsigned this_files;
7153 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
7154 table->files = kcalloc(this_files, sizeof(struct file *),
7158 nr_files -= this_files;
7164 for (i = 0; i < nr_tables; i++) {
7165 struct fixed_file_table *table = &ctx->file_data->table[i];
7166 kfree(table->files);
7171 static void io_ring_file_put(struct io_ring_ctx *ctx, struct file *file)
7173 #if defined(CONFIG_UNIX)
7174 struct sock *sock = ctx->ring_sock->sk;
7175 struct sk_buff_head list, *head = &sock->sk_receive_queue;
7176 struct sk_buff *skb;
7179 __skb_queue_head_init(&list);
7182 * Find the skb that holds this file in its SCM_RIGHTS. When found,
7183 * remove this entry and rearrange the file array.
7185 skb = skb_dequeue(head);
7187 struct scm_fp_list *fp;
7189 fp = UNIXCB(skb).fp;
7190 for (i = 0; i < fp->count; i++) {
7193 if (fp->fp[i] != file)
7196 unix_notinflight(fp->user, fp->fp[i]);
7197 left = fp->count - 1 - i;
7199 memmove(&fp->fp[i], &fp->fp[i + 1],
7200 left * sizeof(struct file *));
7207 __skb_queue_tail(&list, skb);
7217 __skb_queue_tail(&list, skb);
7219 skb = skb_dequeue(head);
7222 if (skb_peek(&list)) {
7223 spin_lock_irq(&head->lock);
7224 while ((skb = __skb_dequeue(&list)) != NULL)
7225 __skb_queue_tail(head, skb);
7226 spin_unlock_irq(&head->lock);
7233 struct io_file_put {
7234 struct list_head list;
7238 static void __io_file_put_work(struct fixed_file_ref_node *ref_node)
7240 struct fixed_file_data *file_data = ref_node->file_data;
7241 struct io_ring_ctx *ctx = file_data->ctx;
7242 struct io_file_put *pfile, *tmp;
7244 list_for_each_entry_safe(pfile, tmp, &ref_node->file_list, list) {
7245 list_del(&pfile->list);
7246 io_ring_file_put(ctx, pfile->file);
7250 spin_lock(&file_data->lock);
7251 list_del(&ref_node->node);
7252 spin_unlock(&file_data->lock);
7254 percpu_ref_exit(&ref_node->refs);
7256 percpu_ref_put(&file_data->refs);
7259 static void io_file_put_work(struct work_struct *work)
7261 struct io_ring_ctx *ctx;
7262 struct llist_node *node;
7264 ctx = container_of(work, struct io_ring_ctx, file_put_work.work);
7265 node = llist_del_all(&ctx->file_put_llist);
7268 struct fixed_file_ref_node *ref_node;
7269 struct llist_node *next = node->next;
7271 ref_node = llist_entry(node, struct fixed_file_ref_node, llist);
7272 __io_file_put_work(ref_node);
7277 static void io_file_data_ref_zero(struct percpu_ref *ref)
7279 struct fixed_file_ref_node *ref_node;
7280 struct io_ring_ctx *ctx;
7284 ref_node = container_of(ref, struct fixed_file_ref_node, refs);
7285 ctx = ref_node->file_data->ctx;
7287 if (percpu_ref_is_dying(&ctx->file_data->refs))
7290 first_add = llist_add(&ref_node->llist, &ctx->file_put_llist);
7292 mod_delayed_work(system_wq, &ctx->file_put_work, 0);
7294 queue_delayed_work(system_wq, &ctx->file_put_work, delay);
7297 static struct fixed_file_ref_node *alloc_fixed_file_ref_node(
7298 struct io_ring_ctx *ctx)
7300 struct fixed_file_ref_node *ref_node;
7302 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
7304 return ERR_PTR(-ENOMEM);
7306 if (percpu_ref_init(&ref_node->refs, io_file_data_ref_zero,
7309 return ERR_PTR(-ENOMEM);
7311 INIT_LIST_HEAD(&ref_node->node);
7312 INIT_LIST_HEAD(&ref_node->file_list);
7313 ref_node->file_data = ctx->file_data;
7317 static void destroy_fixed_file_ref_node(struct fixed_file_ref_node *ref_node)
7319 percpu_ref_exit(&ref_node->refs);
7323 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
7326 __s32 __user *fds = (__s32 __user *) arg;
7331 struct fixed_file_ref_node *ref_node;
7337 if (nr_args > IORING_MAX_FIXED_FILES)
7340 ctx->file_data = kzalloc(sizeof(*ctx->file_data), GFP_KERNEL);
7341 if (!ctx->file_data)
7343 ctx->file_data->ctx = ctx;
7344 init_completion(&ctx->file_data->done);
7345 INIT_LIST_HEAD(&ctx->file_data->ref_list);
7346 spin_lock_init(&ctx->file_data->lock);
7348 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
7349 ctx->file_data->table = kcalloc(nr_tables,
7350 sizeof(struct fixed_file_table),
7352 if (!ctx->file_data->table) {
7353 kfree(ctx->file_data);
7354 ctx->file_data = NULL;
7358 if (percpu_ref_init(&ctx->file_data->refs, io_file_ref_kill,
7359 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) {
7360 kfree(ctx->file_data->table);
7361 kfree(ctx->file_data);
7362 ctx->file_data = NULL;
7366 if (io_sqe_alloc_file_tables(ctx, nr_tables, nr_args)) {
7367 percpu_ref_exit(&ctx->file_data->refs);
7368 kfree(ctx->file_data->table);
7369 kfree(ctx->file_data);
7370 ctx->file_data = NULL;
7374 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
7375 struct fixed_file_table *table;
7379 if (copy_from_user(&fd, &fds[i], sizeof(fd)))
7381 /* allow sparse sets */
7387 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7388 index = i & IORING_FILE_TABLE_MASK;
7396 * Don't allow io_uring instances to be registered. If UNIX
7397 * isn't enabled, then this causes a reference cycle and this
7398 * instance can never get freed. If UNIX is enabled we'll
7399 * handle it just fine, but there's still no point in allowing
7400 * a ring fd as it doesn't support regular read/write anyway.
7402 if (file->f_op == &io_uring_fops) {
7407 table->files[index] = file;
7411 for (i = 0; i < ctx->nr_user_files; i++) {
7412 file = io_file_from_index(ctx, i);
7416 for (i = 0; i < nr_tables; i++)
7417 kfree(ctx->file_data->table[i].files);
7419 percpu_ref_exit(&ctx->file_data->refs);
7420 kfree(ctx->file_data->table);
7421 kfree(ctx->file_data);
7422 ctx->file_data = NULL;
7423 ctx->nr_user_files = 0;
7427 ret = io_sqe_files_scm(ctx);
7429 io_sqe_files_unregister(ctx);
7433 ref_node = alloc_fixed_file_ref_node(ctx);
7434 if (IS_ERR(ref_node)) {
7435 io_sqe_files_unregister(ctx);
7436 return PTR_ERR(ref_node);
7439 ctx->file_data->cur_refs = &ref_node->refs;
7440 spin_lock(&ctx->file_data->lock);
7441 list_add(&ref_node->node, &ctx->file_data->ref_list);
7442 spin_unlock(&ctx->file_data->lock);
7443 percpu_ref_get(&ctx->file_data->refs);
7447 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
7450 #if defined(CONFIG_UNIX)
7451 struct sock *sock = ctx->ring_sock->sk;
7452 struct sk_buff_head *head = &sock->sk_receive_queue;
7453 struct sk_buff *skb;
7456 * See if we can merge this file into an existing skb SCM_RIGHTS
7457 * file set. If there's no room, fall back to allocating a new skb
7458 * and filling it in.
7460 spin_lock_irq(&head->lock);
7461 skb = skb_peek(head);
7463 struct scm_fp_list *fpl = UNIXCB(skb).fp;
7465 if (fpl->count < SCM_MAX_FD) {
7466 __skb_unlink(skb, head);
7467 spin_unlock_irq(&head->lock);
7468 fpl->fp[fpl->count] = get_file(file);
7469 unix_inflight(fpl->user, fpl->fp[fpl->count]);
7471 spin_lock_irq(&head->lock);
7472 __skb_queue_head(head, skb);
7477 spin_unlock_irq(&head->lock);
7484 return __io_sqe_files_scm(ctx, 1, index);
7490 static int io_queue_file_removal(struct fixed_file_data *data,
7493 struct io_file_put *pfile;
7494 struct percpu_ref *refs = data->cur_refs;
7495 struct fixed_file_ref_node *ref_node;
7497 pfile = kzalloc(sizeof(*pfile), GFP_KERNEL);
7501 ref_node = container_of(refs, struct fixed_file_ref_node, refs);
7503 list_add(&pfile->list, &ref_node->file_list);
7508 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
7509 struct io_uring_files_update *up,
7512 struct fixed_file_data *data = ctx->file_data;
7513 struct fixed_file_ref_node *ref_node;
7518 bool needs_switch = false;
7520 if (check_add_overflow(up->offset, nr_args, &done))
7522 if (done > ctx->nr_user_files)
7525 ref_node = alloc_fixed_file_ref_node(ctx);
7526 if (IS_ERR(ref_node))
7527 return PTR_ERR(ref_node);
7530 fds = u64_to_user_ptr(up->fds);
7532 struct fixed_file_table *table;
7536 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
7540 i = array_index_nospec(up->offset, ctx->nr_user_files);
7541 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7542 index = i & IORING_FILE_TABLE_MASK;
7543 if (table->files[index]) {
7544 file = table->files[index];
7545 err = io_queue_file_removal(data, file);
7548 table->files[index] = NULL;
7549 needs_switch = true;
7558 * Don't allow io_uring instances to be registered. If
7559 * UNIX isn't enabled, then this causes a reference
7560 * cycle and this instance can never get freed. If UNIX
7561 * is enabled we'll handle it just fine, but there's
7562 * still no point in allowing a ring fd as it doesn't
7563 * support regular read/write anyway.
7565 if (file->f_op == &io_uring_fops) {
7570 table->files[index] = file;
7571 err = io_sqe_file_register(ctx, file, i);
7573 table->files[index] = NULL;
7584 percpu_ref_kill(data->cur_refs);
7585 spin_lock(&data->lock);
7586 list_add(&ref_node->node, &data->ref_list);
7587 data->cur_refs = &ref_node->refs;
7588 spin_unlock(&data->lock);
7589 percpu_ref_get(&ctx->file_data->refs);
7591 destroy_fixed_file_ref_node(ref_node);
7593 return done ? done : err;
7596 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
7599 struct io_uring_files_update up;
7601 if (!ctx->file_data)
7605 if (copy_from_user(&up, arg, sizeof(up)))
7610 return __io_sqe_files_update(ctx, &up, nr_args);
7613 static void io_free_work(struct io_wq_work *work)
7615 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
7617 /* Consider that io_steal_work() relies on this ref */
7621 static int io_init_wq_offload(struct io_ring_ctx *ctx,
7622 struct io_uring_params *p)
7624 struct io_wq_data data;
7626 struct io_ring_ctx *ctx_attach;
7627 unsigned int concurrency;
7630 data.user = ctx->user;
7631 data.free_work = io_free_work;
7632 data.do_work = io_wq_submit_work;
7634 if (!(p->flags & IORING_SETUP_ATTACH_WQ)) {
7635 /* Do QD, or 4 * CPUS, whatever is smallest */
7636 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
7638 ctx->io_wq = io_wq_create(concurrency, &data);
7639 if (IS_ERR(ctx->io_wq)) {
7640 ret = PTR_ERR(ctx->io_wq);
7646 f = fdget(p->wq_fd);
7650 if (f.file->f_op != &io_uring_fops) {
7655 ctx_attach = f.file->private_data;
7656 /* @io_wq is protected by holding the fd */
7657 if (!io_wq_get(ctx_attach->io_wq, &data)) {
7662 ctx->io_wq = ctx_attach->io_wq;
7668 static int io_uring_alloc_task_context(struct task_struct *task)
7670 struct io_uring_task *tctx;
7672 tctx = kmalloc(sizeof(*tctx), GFP_KERNEL);
7673 if (unlikely(!tctx))
7677 init_waitqueue_head(&tctx->wait);
7680 atomic_long_set(&tctx->req_issue, 0);
7681 atomic_long_set(&tctx->req_complete, 0);
7682 task->io_uring = tctx;
7686 void __io_uring_free(struct task_struct *tsk)
7688 struct io_uring_task *tctx = tsk->io_uring;
7690 WARN_ON_ONCE(!xa_empty(&tctx->xa));
7692 tsk->io_uring = NULL;
7695 static int io_sq_offload_create(struct io_ring_ctx *ctx,
7696 struct io_uring_params *p)
7700 if (ctx->flags & IORING_SETUP_SQPOLL) {
7701 struct io_sq_data *sqd;
7704 if (!capable(CAP_SYS_ADMIN))
7707 sqd = io_get_sq_data(p);
7714 io_sq_thread_park(sqd);
7715 mutex_lock(&sqd->ctx_lock);
7716 list_add(&ctx->sqd_list, &sqd->ctx_new_list);
7717 mutex_unlock(&sqd->ctx_lock);
7718 io_sq_thread_unpark(sqd);
7720 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
7721 if (!ctx->sq_thread_idle)
7722 ctx->sq_thread_idle = HZ;
7727 if (p->flags & IORING_SETUP_SQ_AFF) {
7728 int cpu = p->sq_thread_cpu;
7731 if (cpu >= nr_cpu_ids)
7733 if (!cpu_online(cpu))
7736 sqd->thread = kthread_create_on_cpu(io_sq_thread, sqd,
7737 cpu, "io_uring-sq");
7739 sqd->thread = kthread_create(io_sq_thread, sqd,
7742 if (IS_ERR(sqd->thread)) {
7743 ret = PTR_ERR(sqd->thread);
7747 ret = io_uring_alloc_task_context(sqd->thread);
7750 } else if (p->flags & IORING_SETUP_SQ_AFF) {
7751 /* Can't have SQ_AFF without SQPOLL */
7757 ret = io_init_wq_offload(ctx, p);
7763 io_finish_async(ctx);
7767 static void io_sq_offload_start(struct io_ring_ctx *ctx)
7769 struct io_sq_data *sqd = ctx->sq_data;
7771 if ((ctx->flags & IORING_SETUP_SQPOLL) && sqd->thread)
7772 wake_up_process(sqd->thread);
7775 static inline void __io_unaccount_mem(struct user_struct *user,
7776 unsigned long nr_pages)
7778 atomic_long_sub(nr_pages, &user->locked_vm);
7781 static inline int __io_account_mem(struct user_struct *user,
7782 unsigned long nr_pages)
7784 unsigned long page_limit, cur_pages, new_pages;
7786 /* Don't allow more pages than we can safely lock */
7787 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
7790 cur_pages = atomic_long_read(&user->locked_vm);
7791 new_pages = cur_pages + nr_pages;
7792 if (new_pages > page_limit)
7794 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
7795 new_pages) != cur_pages);
7800 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7801 enum io_mem_account acct)
7804 __io_unaccount_mem(ctx->user, nr_pages);
7806 if (ctx->mm_account) {
7807 if (acct == ACCT_LOCKED)
7808 ctx->mm_account->locked_vm -= nr_pages;
7809 else if (acct == ACCT_PINNED)
7810 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm);
7814 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages,
7815 enum io_mem_account acct)
7819 if (ctx->limit_mem) {
7820 ret = __io_account_mem(ctx->user, nr_pages);
7825 if (ctx->mm_account) {
7826 if (acct == ACCT_LOCKED)
7827 ctx->mm_account->locked_vm += nr_pages;
7828 else if (acct == ACCT_PINNED)
7829 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm);
7835 static void io_mem_free(void *ptr)
7842 page = virt_to_head_page(ptr);
7843 if (put_page_testzero(page))
7844 free_compound_page(page);
7847 static void *io_mem_alloc(size_t size)
7849 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
7852 return (void *) __get_free_pages(gfp_flags, get_order(size));
7855 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
7858 struct io_rings *rings;
7859 size_t off, sq_array_size;
7861 off = struct_size(rings, cqes, cq_entries);
7862 if (off == SIZE_MAX)
7866 off = ALIGN(off, SMP_CACHE_BYTES);
7874 sq_array_size = array_size(sizeof(u32), sq_entries);
7875 if (sq_array_size == SIZE_MAX)
7878 if (check_add_overflow(off, sq_array_size, &off))
7884 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
7888 pages = (size_t)1 << get_order(
7889 rings_size(sq_entries, cq_entries, NULL));
7890 pages += (size_t)1 << get_order(
7891 array_size(sizeof(struct io_uring_sqe), sq_entries));
7896 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
7900 if (!ctx->user_bufs)
7903 for (i = 0; i < ctx->nr_user_bufs; i++) {
7904 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7906 for (j = 0; j < imu->nr_bvecs; j++)
7907 unpin_user_page(imu->bvec[j].bv_page);
7909 if (imu->acct_pages)
7910 io_unaccount_mem(ctx, imu->acct_pages, ACCT_PINNED);
7915 kfree(ctx->user_bufs);
7916 ctx->user_bufs = NULL;
7917 ctx->nr_user_bufs = 0;
7921 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
7922 void __user *arg, unsigned index)
7924 struct iovec __user *src;
7926 #ifdef CONFIG_COMPAT
7928 struct compat_iovec __user *ciovs;
7929 struct compat_iovec ciov;
7931 ciovs = (struct compat_iovec __user *) arg;
7932 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
7935 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
7936 dst->iov_len = ciov.iov_len;
7940 src = (struct iovec __user *) arg;
7941 if (copy_from_user(dst, &src[index], sizeof(*dst)))
7947 * Not super efficient, but this is just a registration time. And we do cache
7948 * the last compound head, so generally we'll only do a full search if we don't
7951 * We check if the given compound head page has already been accounted, to
7952 * avoid double accounting it. This allows us to account the full size of the
7953 * page, not just the constituent pages of a huge page.
7955 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages,
7956 int nr_pages, struct page *hpage)
7960 /* check current page array */
7961 for (i = 0; i < nr_pages; i++) {
7962 if (!PageCompound(pages[i]))
7964 if (compound_head(pages[i]) == hpage)
7968 /* check previously registered pages */
7969 for (i = 0; i < ctx->nr_user_bufs; i++) {
7970 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7972 for (j = 0; j < imu->nr_bvecs; j++) {
7973 if (!PageCompound(imu->bvec[j].bv_page))
7975 if (compound_head(imu->bvec[j].bv_page) == hpage)
7983 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages,
7984 int nr_pages, struct io_mapped_ubuf *imu,
7985 struct page **last_hpage)
7989 for (i = 0; i < nr_pages; i++) {
7990 if (!PageCompound(pages[i])) {
7995 hpage = compound_head(pages[i]);
7996 if (hpage == *last_hpage)
7998 *last_hpage = hpage;
7999 if (headpage_already_acct(ctx, pages, i, hpage))
8001 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT;
8005 if (!imu->acct_pages)
8008 ret = io_account_mem(ctx, imu->acct_pages, ACCT_PINNED);
8010 imu->acct_pages = 0;
8014 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
8017 struct vm_area_struct **vmas = NULL;
8018 struct page **pages = NULL;
8019 struct page *last_hpage = NULL;
8020 int i, j, got_pages = 0;
8025 if (!nr_args || nr_args > UIO_MAXIOV)
8028 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
8030 if (!ctx->user_bufs)
8033 for (i = 0; i < nr_args; i++) {
8034 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
8035 unsigned long off, start, end, ubuf;
8040 ret = io_copy_iov(ctx, &iov, arg, i);
8045 * Don't impose further limits on the size and buffer
8046 * constraints here, we'll -EINVAL later when IO is
8047 * submitted if they are wrong.
8050 if (!iov.iov_base || !iov.iov_len)
8053 /* arbitrary limit, but we need something */
8054 if (iov.iov_len > SZ_1G)
8057 ubuf = (unsigned long) iov.iov_base;
8058 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
8059 start = ubuf >> PAGE_SHIFT;
8060 nr_pages = end - start;
8063 if (!pages || nr_pages > got_pages) {
8066 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
8068 vmas = kvmalloc_array(nr_pages,
8069 sizeof(struct vm_area_struct *),
8071 if (!pages || !vmas) {
8075 got_pages = nr_pages;
8078 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
8085 mmap_read_lock(current->mm);
8086 pret = pin_user_pages(ubuf, nr_pages,
8087 FOLL_WRITE | FOLL_LONGTERM,
8089 if (pret == nr_pages) {
8090 /* don't support file backed memory */
8091 for (j = 0; j < nr_pages; j++) {
8092 struct vm_area_struct *vma = vmas[j];
8095 !is_file_hugepages(vma->vm_file)) {
8101 ret = pret < 0 ? pret : -EFAULT;
8103 mmap_read_unlock(current->mm);
8106 * if we did partial map, or found file backed vmas,
8107 * release any pages we did get
8110 unpin_user_pages(pages, pret);
8115 ret = io_buffer_account_pin(ctx, pages, pret, imu, &last_hpage);
8117 unpin_user_pages(pages, pret);
8122 off = ubuf & ~PAGE_MASK;
8124 for (j = 0; j < nr_pages; j++) {
8127 vec_len = min_t(size_t, size, PAGE_SIZE - off);
8128 imu->bvec[j].bv_page = pages[j];
8129 imu->bvec[j].bv_len = vec_len;
8130 imu->bvec[j].bv_offset = off;
8134 /* store original address for later verification */
8136 imu->len = iov.iov_len;
8137 imu->nr_bvecs = nr_pages;
8139 ctx->nr_user_bufs++;
8147 io_sqe_buffer_unregister(ctx);
8151 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
8153 __s32 __user *fds = arg;
8159 if (copy_from_user(&fd, fds, sizeof(*fds)))
8162 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
8163 if (IS_ERR(ctx->cq_ev_fd)) {
8164 int ret = PTR_ERR(ctx->cq_ev_fd);
8165 ctx->cq_ev_fd = NULL;
8172 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
8174 if (ctx->cq_ev_fd) {
8175 eventfd_ctx_put(ctx->cq_ev_fd);
8176 ctx->cq_ev_fd = NULL;
8183 static int __io_destroy_buffers(int id, void *p, void *data)
8185 struct io_ring_ctx *ctx = data;
8186 struct io_buffer *buf = p;
8188 __io_remove_buffers(ctx, buf, id, -1U);
8192 static void io_destroy_buffers(struct io_ring_ctx *ctx)
8194 idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx);
8195 idr_destroy(&ctx->io_buffer_idr);
8198 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
8200 io_finish_async(ctx);
8201 io_sqe_buffer_unregister(ctx);
8203 if (ctx->sqo_task) {
8204 put_task_struct(ctx->sqo_task);
8205 ctx->sqo_task = NULL;
8206 mmdrop(ctx->mm_account);
8207 ctx->mm_account = NULL;
8210 #ifdef CONFIG_BLK_CGROUP
8211 if (ctx->sqo_blkcg_css)
8212 css_put(ctx->sqo_blkcg_css);
8215 io_sqe_files_unregister(ctx);
8216 io_eventfd_unregister(ctx);
8217 io_destroy_buffers(ctx);
8218 idr_destroy(&ctx->personality_idr);
8220 #if defined(CONFIG_UNIX)
8221 if (ctx->ring_sock) {
8222 ctx->ring_sock->file = NULL; /* so that iput() is called */
8223 sock_release(ctx->ring_sock);
8227 io_mem_free(ctx->rings);
8228 io_mem_free(ctx->sq_sqes);
8230 percpu_ref_exit(&ctx->refs);
8231 free_uid(ctx->user);
8232 put_cred(ctx->creds);
8233 kfree(ctx->cancel_hash);
8234 kmem_cache_free(req_cachep, ctx->fallback_req);
8238 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
8240 struct io_ring_ctx *ctx = file->private_data;
8243 poll_wait(file, &ctx->cq_wait, wait);
8245 * synchronizes with barrier from wq_has_sleeper call in
8249 if (!io_sqring_full(ctx))
8250 mask |= EPOLLOUT | EPOLLWRNORM;
8251 if (io_cqring_events(ctx, false))
8252 mask |= EPOLLIN | EPOLLRDNORM;
8257 static int io_uring_fasync(int fd, struct file *file, int on)
8259 struct io_ring_ctx *ctx = file->private_data;
8261 return fasync_helper(fd, file, on, &ctx->cq_fasync);
8264 static int io_remove_personalities(int id, void *p, void *data)
8266 struct io_ring_ctx *ctx = data;
8267 const struct cred *cred;
8269 cred = idr_remove(&ctx->personality_idr, id);
8275 static void io_ring_exit_work(struct work_struct *work)
8277 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
8281 * If we're doing polled IO and end up having requests being
8282 * submitted async (out-of-line), then completions can come in while
8283 * we're waiting for refs to drop. We need to reap these manually,
8284 * as nobody else will be looking for them.
8288 io_cqring_overflow_flush(ctx, true, NULL, NULL);
8289 io_iopoll_try_reap_events(ctx);
8290 } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20));
8291 io_ring_ctx_free(ctx);
8294 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
8296 mutex_lock(&ctx->uring_lock);
8297 percpu_ref_kill(&ctx->refs);
8298 mutex_unlock(&ctx->uring_lock);
8300 io_kill_timeouts(ctx, NULL);
8301 io_poll_remove_all(ctx, NULL);
8304 io_wq_cancel_all(ctx->io_wq);
8306 /* if we failed setting up the ctx, we might not have any rings */
8308 io_cqring_overflow_flush(ctx, true, NULL, NULL);
8309 io_iopoll_try_reap_events(ctx);
8310 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
8313 * Do this upfront, so we won't have a grace period where the ring
8314 * is closed but resources aren't reaped yet. This can cause
8315 * spurious failure in setting up a new ring.
8317 io_unaccount_mem(ctx, ring_pages(ctx->sq_entries, ctx->cq_entries),
8320 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
8322 * Use system_unbound_wq to avoid spawning tons of event kworkers
8323 * if we're exiting a ton of rings at the same time. It just adds
8324 * noise and overhead, there's no discernable change in runtime
8325 * over using system_wq.
8327 queue_work(system_unbound_wq, &ctx->exit_work);
8330 static int io_uring_release(struct inode *inode, struct file *file)
8332 struct io_ring_ctx *ctx = file->private_data;
8334 file->private_data = NULL;
8335 io_ring_ctx_wait_and_kill(ctx);
8339 static bool io_wq_files_match(struct io_wq_work *work, void *data)
8341 struct files_struct *files = data;
8343 return !files || work->files == files;
8347 * Returns true if 'preq' is the link parent of 'req'
8349 static bool io_match_link(struct io_kiocb *preq, struct io_kiocb *req)
8351 struct io_kiocb *link;
8353 if (!(preq->flags & REQ_F_LINK_HEAD))
8356 list_for_each_entry(link, &preq->link_list, link_list) {
8364 static bool io_match_link_files(struct io_kiocb *req,
8365 struct files_struct *files)
8367 struct io_kiocb *link;
8369 if (io_match_files(req, files))
8371 if (req->flags & REQ_F_LINK_HEAD) {
8372 list_for_each_entry(link, &req->link_list, link_list) {
8373 if (io_match_files(link, files))
8381 * We're looking to cancel 'req' because it's holding on to our files, but
8382 * 'req' could be a link to another request. See if it is, and cancel that
8383 * parent request if so.
8385 static bool io_poll_remove_link(struct io_ring_ctx *ctx, struct io_kiocb *req)
8387 struct hlist_node *tmp;
8388 struct io_kiocb *preq;
8392 spin_lock_irq(&ctx->completion_lock);
8393 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
8394 struct hlist_head *list;
8396 list = &ctx->cancel_hash[i];
8397 hlist_for_each_entry_safe(preq, tmp, list, hash_node) {
8398 found = io_match_link(preq, req);
8400 io_poll_remove_one(preq);
8405 spin_unlock_irq(&ctx->completion_lock);
8409 static bool io_timeout_remove_link(struct io_ring_ctx *ctx,
8410 struct io_kiocb *req)
8412 struct io_kiocb *preq;
8415 spin_lock_irq(&ctx->completion_lock);
8416 list_for_each_entry(preq, &ctx->timeout_list, timeout.list) {
8417 found = io_match_link(preq, req);
8419 __io_timeout_cancel(preq);
8423 spin_unlock_irq(&ctx->completion_lock);
8427 static bool io_cancel_link_cb(struct io_wq_work *work, void *data)
8429 return io_match_link(container_of(work, struct io_kiocb, work), data);
8432 static void io_attempt_cancel(struct io_ring_ctx *ctx, struct io_kiocb *req)
8434 enum io_wq_cancel cret;
8436 /* cancel this particular work, if it's running */
8437 cret = io_wq_cancel_work(ctx->io_wq, &req->work);
8438 if (cret != IO_WQ_CANCEL_NOTFOUND)
8441 /* find links that hold this pending, cancel those */
8442 cret = io_wq_cancel_cb(ctx->io_wq, io_cancel_link_cb, req, true);
8443 if (cret != IO_WQ_CANCEL_NOTFOUND)
8446 /* if we have a poll link holding this pending, cancel that */
8447 if (io_poll_remove_link(ctx, req))
8450 /* final option, timeout link is holding this req pending */
8451 io_timeout_remove_link(ctx, req);
8454 static void io_cancel_defer_files(struct io_ring_ctx *ctx,
8455 struct files_struct *files)
8457 struct io_defer_entry *de = NULL;
8460 spin_lock_irq(&ctx->completion_lock);
8461 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
8462 if (io_match_link_files(de->req, files)) {
8463 list_cut_position(&list, &ctx->defer_list, &de->list);
8467 spin_unlock_irq(&ctx->completion_lock);
8469 while (!list_empty(&list)) {
8470 de = list_first_entry(&list, struct io_defer_entry, list);
8471 list_del_init(&de->list);
8472 req_set_fail_links(de->req);
8473 io_put_req(de->req);
8474 io_req_complete(de->req, -ECANCELED);
8480 * Returns true if we found and killed one or more files pinning requests
8482 static bool io_uring_cancel_files(struct io_ring_ctx *ctx,
8483 struct files_struct *files)
8485 if (list_empty_careful(&ctx->inflight_list))
8488 io_cancel_defer_files(ctx, files);
8489 /* cancel all at once, should be faster than doing it one by one*/
8490 io_wq_cancel_cb(ctx->io_wq, io_wq_files_match, files, true);
8492 while (!list_empty_careful(&ctx->inflight_list)) {
8493 struct io_kiocb *cancel_req = NULL, *req;
8496 spin_lock_irq(&ctx->inflight_lock);
8497 list_for_each_entry(req, &ctx->inflight_list, inflight_entry) {
8498 if (files && req->work.files != files)
8500 /* req is being completed, ignore */
8501 if (!refcount_inc_not_zero(&req->refs))
8507 prepare_to_wait(&ctx->inflight_wait, &wait,
8508 TASK_UNINTERRUPTIBLE);
8509 spin_unlock_irq(&ctx->inflight_lock);
8511 /* We need to keep going until we don't find a matching req */
8514 /* cancel this request, or head link requests */
8515 io_attempt_cancel(ctx, cancel_req);
8516 io_put_req(cancel_req);
8517 /* cancellations _may_ trigger task work */
8520 finish_wait(&ctx->inflight_wait, &wait);
8526 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
8528 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
8529 struct task_struct *task = data;
8531 return io_task_match(req, task);
8534 static bool __io_uring_cancel_task_requests(struct io_ring_ctx *ctx,
8535 struct task_struct *task,
8536 struct files_struct *files)
8540 ret = io_uring_cancel_files(ctx, files);
8542 enum io_wq_cancel cret;
8544 cret = io_wq_cancel_cb(ctx->io_wq, io_cancel_task_cb, task, true);
8545 if (cret != IO_WQ_CANCEL_NOTFOUND)
8548 /* SQPOLL thread does its own polling */
8549 if (!(ctx->flags & IORING_SETUP_SQPOLL)) {
8550 while (!list_empty_careful(&ctx->iopoll_list)) {
8551 io_iopoll_try_reap_events(ctx);
8556 ret |= io_poll_remove_all(ctx, task);
8557 ret |= io_kill_timeouts(ctx, task);
8564 * We need to iteratively cancel requests, in case a request has dependent
8565 * hard links. These persist even for failure of cancelations, hence keep
8566 * looping until none are found.
8568 static void io_uring_cancel_task_requests(struct io_ring_ctx *ctx,
8569 struct files_struct *files)
8571 struct task_struct *task = current;
8573 if ((ctx->flags & IORING_SETUP_SQPOLL) && ctx->sq_data)
8574 task = ctx->sq_data->thread;
8576 io_cqring_overflow_flush(ctx, true, task, files);
8578 while (__io_uring_cancel_task_requests(ctx, task, files)) {
8585 * Note that this task has used io_uring. We use it for cancelation purposes.
8587 static int io_uring_add_task_file(struct file *file)
8589 if (unlikely(!current->io_uring)) {
8592 ret = io_uring_alloc_task_context(current);
8596 if (current->io_uring->last != file) {
8597 XA_STATE(xas, ¤t->io_uring->xa, (unsigned long) file);
8601 old = xas_load(&xas);
8605 xas_store(&xas, file);
8609 current->io_uring->last = file;
8616 * Remove this io_uring_file -> task mapping.
8618 static void io_uring_del_task_file(struct file *file)
8620 struct io_uring_task *tctx = current->io_uring;
8621 XA_STATE(xas, &tctx->xa, (unsigned long) file);
8623 if (tctx->last == file)
8627 file = xas_store(&xas, NULL);
8634 static void __io_uring_attempt_task_drop(struct file *file)
8636 XA_STATE(xas, ¤t->io_uring->xa, (unsigned long) file);
8640 old = xas_load(&xas);
8644 io_uring_del_task_file(file);
8648 * Drop task note for this file if we're the only ones that hold it after
8651 static void io_uring_attempt_task_drop(struct file *file, bool exiting)
8653 if (!current->io_uring)
8656 * fput() is pending, will be 2 if the only other ref is our potential
8657 * task file note. If the task is exiting, drop regardless of count.
8659 if (!exiting && atomic_long_read(&file->f_count) != 2)
8662 __io_uring_attempt_task_drop(file);
8665 void __io_uring_files_cancel(struct files_struct *files)
8667 struct io_uring_task *tctx = current->io_uring;
8668 XA_STATE(xas, &tctx->xa, 0);
8670 /* make sure overflow events are dropped */
8671 tctx->in_idle = true;
8674 struct io_ring_ctx *ctx;
8678 file = xas_next_entry(&xas, ULONG_MAX);
8684 ctx = file->private_data;
8686 io_uring_cancel_task_requests(ctx, files);
8688 io_uring_del_task_file(file);
8692 static inline bool io_uring_task_idle(struct io_uring_task *tctx)
8694 return atomic_long_read(&tctx->req_issue) ==
8695 atomic_long_read(&tctx->req_complete);
8699 * Find any io_uring fd that this task has registered or done IO on, and cancel
8702 void __io_uring_task_cancel(void)
8704 struct io_uring_task *tctx = current->io_uring;
8708 /* make sure overflow events are dropped */
8709 tctx->in_idle = true;
8711 while (!io_uring_task_idle(tctx)) {
8712 /* read completions before cancelations */
8713 completions = atomic_long_read(&tctx->req_complete);
8714 __io_uring_files_cancel(NULL);
8716 prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE);
8719 * If we've seen completions, retry. This avoids a race where
8720 * a completion comes in before we did prepare_to_wait().
8722 if (completions != atomic_long_read(&tctx->req_complete))
8724 if (io_uring_task_idle(tctx))
8729 finish_wait(&tctx->wait, &wait);
8730 tctx->in_idle = false;
8733 static int io_uring_flush(struct file *file, void *data)
8735 struct io_ring_ctx *ctx = file->private_data;
8738 * If the task is going away, cancel work it may have pending
8740 if (fatal_signal_pending(current) || (current->flags & PF_EXITING))
8743 io_uring_cancel_task_requests(ctx, data);
8744 io_uring_attempt_task_drop(file, !data);
8748 static void *io_uring_validate_mmap_request(struct file *file,
8749 loff_t pgoff, size_t sz)
8751 struct io_ring_ctx *ctx = file->private_data;
8752 loff_t offset = pgoff << PAGE_SHIFT;
8757 case IORING_OFF_SQ_RING:
8758 case IORING_OFF_CQ_RING:
8761 case IORING_OFF_SQES:
8765 return ERR_PTR(-EINVAL);
8768 page = virt_to_head_page(ptr);
8769 if (sz > page_size(page))
8770 return ERR_PTR(-EINVAL);
8777 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
8779 size_t sz = vma->vm_end - vma->vm_start;
8783 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
8785 return PTR_ERR(ptr);
8787 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
8788 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
8791 #else /* !CONFIG_MMU */
8793 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
8795 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
8798 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
8800 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
8803 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
8804 unsigned long addr, unsigned long len,
8805 unsigned long pgoff, unsigned long flags)
8809 ptr = io_uring_validate_mmap_request(file, pgoff, len);
8811 return PTR_ERR(ptr);
8813 return (unsigned long) ptr;
8816 #endif /* !CONFIG_MMU */
8818 static void io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
8823 if (!io_sqring_full(ctx))
8826 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
8828 if (!io_sqring_full(ctx))
8832 } while (!signal_pending(current));
8834 finish_wait(&ctx->sqo_sq_wait, &wait);
8837 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
8838 u32, min_complete, u32, flags, const sigset_t __user *, sig,
8841 struct io_ring_ctx *ctx;
8848 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
8849 IORING_ENTER_SQ_WAIT))
8857 if (f.file->f_op != &io_uring_fops)
8861 ctx = f.file->private_data;
8862 if (!percpu_ref_tryget(&ctx->refs))
8866 if (ctx->flags & IORING_SETUP_R_DISABLED)
8870 * For SQ polling, the thread will do all submissions and completions.
8871 * Just return the requested submit count, and wake the thread if
8875 if (ctx->flags & IORING_SETUP_SQPOLL) {
8876 if (!list_empty_careful(&ctx->cq_overflow_list))
8877 io_cqring_overflow_flush(ctx, false, NULL, NULL);
8878 if (flags & IORING_ENTER_SQ_WAKEUP)
8879 wake_up(&ctx->sq_data->wait);
8880 if (flags & IORING_ENTER_SQ_WAIT)
8881 io_sqpoll_wait_sq(ctx);
8882 submitted = to_submit;
8883 } else if (to_submit) {
8884 ret = io_uring_add_task_file(f.file);
8887 mutex_lock(&ctx->uring_lock);
8888 submitted = io_submit_sqes(ctx, to_submit);
8889 mutex_unlock(&ctx->uring_lock);
8891 if (submitted != to_submit)
8894 if (flags & IORING_ENTER_GETEVENTS) {
8895 min_complete = min(min_complete, ctx->cq_entries);
8898 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
8899 * space applications don't need to do io completion events
8900 * polling again, they can rely on io_sq_thread to do polling
8901 * work, which can reduce cpu usage and uring_lock contention.
8903 if (ctx->flags & IORING_SETUP_IOPOLL &&
8904 !(ctx->flags & IORING_SETUP_SQPOLL)) {
8905 ret = io_iopoll_check(ctx, min_complete);
8907 ret = io_cqring_wait(ctx, min_complete, sig, sigsz);
8912 percpu_ref_put(&ctx->refs);
8915 return submitted ? submitted : ret;
8918 #ifdef CONFIG_PROC_FS
8919 static int io_uring_show_cred(int id, void *p, void *data)
8921 const struct cred *cred = p;
8922 struct seq_file *m = data;
8923 struct user_namespace *uns = seq_user_ns(m);
8924 struct group_info *gi;
8929 seq_printf(m, "%5d\n", id);
8930 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
8931 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
8932 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
8933 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
8934 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
8935 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
8936 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
8937 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
8938 seq_puts(m, "\n\tGroups:\t");
8939 gi = cred->group_info;
8940 for (g = 0; g < gi->ngroups; g++) {
8941 seq_put_decimal_ull(m, g ? " " : "",
8942 from_kgid_munged(uns, gi->gid[g]));
8944 seq_puts(m, "\n\tCapEff:\t");
8945 cap = cred->cap_effective;
8946 CAP_FOR_EACH_U32(__capi)
8947 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
8952 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
8954 struct io_sq_data *sq = NULL;
8959 * Avoid ABBA deadlock between the seq lock and the io_uring mutex,
8960 * since fdinfo case grabs it in the opposite direction of normal use
8961 * cases. If we fail to get the lock, we just don't iterate any
8962 * structures that could be going away outside the io_uring mutex.
8964 has_lock = mutex_trylock(&ctx->uring_lock);
8966 if (has_lock && (ctx->flags & IORING_SETUP_SQPOLL))
8969 seq_printf(m, "SqThread:\t%d\n", sq ? task_pid_nr(sq->thread) : -1);
8970 seq_printf(m, "SqThreadCpu:\t%d\n", sq ? task_cpu(sq->thread) : -1);
8971 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
8972 for (i = 0; has_lock && i < ctx->nr_user_files; i++) {
8973 struct fixed_file_table *table;
8976 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
8977 f = table->files[i & IORING_FILE_TABLE_MASK];
8979 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
8981 seq_printf(m, "%5u: <none>\n", i);
8983 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
8984 for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) {
8985 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
8987 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
8988 (unsigned int) buf->len);
8990 if (has_lock && !idr_is_empty(&ctx->personality_idr)) {
8991 seq_printf(m, "Personalities:\n");
8992 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
8994 seq_printf(m, "PollList:\n");
8995 spin_lock_irq(&ctx->completion_lock);
8996 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
8997 struct hlist_head *list = &ctx->cancel_hash[i];
8998 struct io_kiocb *req;
9000 hlist_for_each_entry(req, list, hash_node)
9001 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
9002 req->task->task_works != NULL);
9004 spin_unlock_irq(&ctx->completion_lock);
9006 mutex_unlock(&ctx->uring_lock);
9009 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
9011 struct io_ring_ctx *ctx = f->private_data;
9013 if (percpu_ref_tryget(&ctx->refs)) {
9014 __io_uring_show_fdinfo(ctx, m);
9015 percpu_ref_put(&ctx->refs);
9020 static const struct file_operations io_uring_fops = {
9021 .release = io_uring_release,
9022 .flush = io_uring_flush,
9023 .mmap = io_uring_mmap,
9025 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
9026 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
9028 .poll = io_uring_poll,
9029 .fasync = io_uring_fasync,
9030 #ifdef CONFIG_PROC_FS
9031 .show_fdinfo = io_uring_show_fdinfo,
9035 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
9036 struct io_uring_params *p)
9038 struct io_rings *rings;
9039 size_t size, sq_array_offset;
9041 /* make sure these are sane, as we already accounted them */
9042 ctx->sq_entries = p->sq_entries;
9043 ctx->cq_entries = p->cq_entries;
9045 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
9046 if (size == SIZE_MAX)
9049 rings = io_mem_alloc(size);
9054 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
9055 rings->sq_ring_mask = p->sq_entries - 1;
9056 rings->cq_ring_mask = p->cq_entries - 1;
9057 rings->sq_ring_entries = p->sq_entries;
9058 rings->cq_ring_entries = p->cq_entries;
9059 ctx->sq_mask = rings->sq_ring_mask;
9060 ctx->cq_mask = rings->cq_ring_mask;
9062 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
9063 if (size == SIZE_MAX) {
9064 io_mem_free(ctx->rings);
9069 ctx->sq_sqes = io_mem_alloc(size);
9070 if (!ctx->sq_sqes) {
9071 io_mem_free(ctx->rings);
9080 * Allocate an anonymous fd, this is what constitutes the application
9081 * visible backing of an io_uring instance. The application mmaps this
9082 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
9083 * we have to tie this fd to a socket for file garbage collection purposes.
9085 static int io_uring_get_fd(struct io_ring_ctx *ctx)
9090 #if defined(CONFIG_UNIX)
9091 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
9097 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
9101 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
9102 O_RDWR | O_CLOEXEC);
9106 ret = PTR_ERR(file);
9110 #if defined(CONFIG_UNIX)
9111 ctx->ring_sock->file = file;
9113 if (unlikely(io_uring_add_task_file(file))) {
9114 file = ERR_PTR(-ENOMEM);
9117 fd_install(ret, file);
9120 #if defined(CONFIG_UNIX)
9121 sock_release(ctx->ring_sock);
9122 ctx->ring_sock = NULL;
9127 static int io_uring_create(unsigned entries, struct io_uring_params *p,
9128 struct io_uring_params __user *params)
9130 struct user_struct *user = NULL;
9131 struct io_ring_ctx *ctx;
9137 if (entries > IORING_MAX_ENTRIES) {
9138 if (!(p->flags & IORING_SETUP_CLAMP))
9140 entries = IORING_MAX_ENTRIES;
9144 * Use twice as many entries for the CQ ring. It's possible for the
9145 * application to drive a higher depth than the size of the SQ ring,
9146 * since the sqes are only used at submission time. This allows for
9147 * some flexibility in overcommitting a bit. If the application has
9148 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
9149 * of CQ ring entries manually.
9151 p->sq_entries = roundup_pow_of_two(entries);
9152 if (p->flags & IORING_SETUP_CQSIZE) {
9154 * If IORING_SETUP_CQSIZE is set, we do the same roundup
9155 * to a power-of-two, if it isn't already. We do NOT impose
9156 * any cq vs sq ring sizing.
9158 if (p->cq_entries < p->sq_entries)
9160 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
9161 if (!(p->flags & IORING_SETUP_CLAMP))
9163 p->cq_entries = IORING_MAX_CQ_ENTRIES;
9165 p->cq_entries = roundup_pow_of_two(p->cq_entries);
9167 p->cq_entries = 2 * p->sq_entries;
9170 user = get_uid(current_user());
9171 limit_mem = !capable(CAP_IPC_LOCK);
9174 ret = __io_account_mem(user,
9175 ring_pages(p->sq_entries, p->cq_entries));
9182 ctx = io_ring_ctx_alloc(p);
9185 __io_unaccount_mem(user, ring_pages(p->sq_entries,
9190 ctx->compat = in_compat_syscall();
9192 ctx->creds = get_current_cred();
9194 ctx->sqo_task = get_task_struct(current);
9197 * This is just grabbed for accounting purposes. When a process exits,
9198 * the mm is exited and dropped before the files, hence we need to hang
9199 * on to this mm purely for the purposes of being able to unaccount
9200 * memory (locked/pinned vm). It's not used for anything else.
9202 mmgrab(current->mm);
9203 ctx->mm_account = current->mm;
9205 #ifdef CONFIG_BLK_CGROUP
9207 * The sq thread will belong to the original cgroup it was inited in.
9208 * If the cgroup goes offline (e.g. disabling the io controller), then
9209 * issued bios will be associated with the closest cgroup later in the
9213 ctx->sqo_blkcg_css = blkcg_css();
9214 ret = css_tryget_online(ctx->sqo_blkcg_css);
9217 /* don't init against a dying cgroup, have the user try again */
9218 ctx->sqo_blkcg_css = NULL;
9225 * Account memory _before_ installing the file descriptor. Once
9226 * the descriptor is installed, it can get closed at any time. Also
9227 * do this before hitting the general error path, as ring freeing
9228 * will un-account as well.
9230 io_account_mem(ctx, ring_pages(p->sq_entries, p->cq_entries),
9232 ctx->limit_mem = limit_mem;
9234 ret = io_allocate_scq_urings(ctx, p);
9238 ret = io_sq_offload_create(ctx, p);
9242 if (!(p->flags & IORING_SETUP_R_DISABLED))
9243 io_sq_offload_start(ctx);
9245 memset(&p->sq_off, 0, sizeof(p->sq_off));
9246 p->sq_off.head = offsetof(struct io_rings, sq.head);
9247 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
9248 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
9249 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
9250 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
9251 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
9252 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9254 memset(&p->cq_off, 0, sizeof(p->cq_off));
9255 p->cq_off.head = offsetof(struct io_rings, cq.head);
9256 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
9257 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
9258 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
9259 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
9260 p->cq_off.cqes = offsetof(struct io_rings, cqes);
9261 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9263 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
9264 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
9265 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
9266 IORING_FEAT_POLL_32BITS;
9268 if (copy_to_user(params, p, sizeof(*p))) {
9274 * Install ring fd as the very last thing, so we don't risk someone
9275 * having closed it before we finish setup
9277 ret = io_uring_get_fd(ctx);
9281 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
9284 io_ring_ctx_wait_and_kill(ctx);
9289 * Sets up an aio uring context, and returns the fd. Applications asks for a
9290 * ring size, we return the actual sq/cq ring sizes (among other things) in the
9291 * params structure passed in.
9293 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
9295 struct io_uring_params p;
9298 if (copy_from_user(&p, params, sizeof(p)))
9300 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
9305 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
9306 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
9307 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
9308 IORING_SETUP_R_DISABLED))
9311 return io_uring_create(entries, &p, params);
9314 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
9315 struct io_uring_params __user *, params)
9317 return io_uring_setup(entries, params);
9320 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
9322 struct io_uring_probe *p;
9326 size = struct_size(p, ops, nr_args);
9327 if (size == SIZE_MAX)
9329 p = kzalloc(size, GFP_KERNEL);
9334 if (copy_from_user(p, arg, size))
9337 if (memchr_inv(p, 0, size))
9340 p->last_op = IORING_OP_LAST - 1;
9341 if (nr_args > IORING_OP_LAST)
9342 nr_args = IORING_OP_LAST;
9344 for (i = 0; i < nr_args; i++) {
9346 if (!io_op_defs[i].not_supported)
9347 p->ops[i].flags = IO_URING_OP_SUPPORTED;
9352 if (copy_to_user(arg, p, size))
9359 static int io_register_personality(struct io_ring_ctx *ctx)
9361 const struct cred *creds = get_current_cred();
9364 id = idr_alloc_cyclic(&ctx->personality_idr, (void *) creds, 1,
9365 USHRT_MAX, GFP_KERNEL);
9371 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
9373 const struct cred *old_creds;
9375 old_creds = idr_remove(&ctx->personality_idr, id);
9377 put_cred(old_creds);
9384 static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg,
9385 unsigned int nr_args)
9387 struct io_uring_restriction *res;
9391 /* Restrictions allowed only if rings started disabled */
9392 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9395 /* We allow only a single restrictions registration */
9396 if (ctx->restrictions.registered)
9399 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
9402 size = array_size(nr_args, sizeof(*res));
9403 if (size == SIZE_MAX)
9406 res = memdup_user(arg, size);
9408 return PTR_ERR(res);
9412 for (i = 0; i < nr_args; i++) {
9413 switch (res[i].opcode) {
9414 case IORING_RESTRICTION_REGISTER_OP:
9415 if (res[i].register_op >= IORING_REGISTER_LAST) {
9420 __set_bit(res[i].register_op,
9421 ctx->restrictions.register_op);
9423 case IORING_RESTRICTION_SQE_OP:
9424 if (res[i].sqe_op >= IORING_OP_LAST) {
9429 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
9431 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
9432 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
9434 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
9435 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
9444 /* Reset all restrictions if an error happened */
9446 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
9448 ctx->restrictions.registered = true;
9454 static int io_register_enable_rings(struct io_ring_ctx *ctx)
9456 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
9459 if (ctx->restrictions.registered)
9460 ctx->restricted = 1;
9462 ctx->flags &= ~IORING_SETUP_R_DISABLED;
9464 io_sq_offload_start(ctx);
9469 static bool io_register_op_must_quiesce(int op)
9472 case IORING_UNREGISTER_FILES:
9473 case IORING_REGISTER_FILES_UPDATE:
9474 case IORING_REGISTER_PROBE:
9475 case IORING_REGISTER_PERSONALITY:
9476 case IORING_UNREGISTER_PERSONALITY:
9483 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
9484 void __user *arg, unsigned nr_args)
9485 __releases(ctx->uring_lock)
9486 __acquires(ctx->uring_lock)
9491 * We're inside the ring mutex, if the ref is already dying, then
9492 * someone else killed the ctx or is already going through
9493 * io_uring_register().
9495 if (percpu_ref_is_dying(&ctx->refs))
9498 if (io_register_op_must_quiesce(opcode)) {
9499 percpu_ref_kill(&ctx->refs);
9502 * Drop uring mutex before waiting for references to exit. If
9503 * another thread is currently inside io_uring_enter() it might
9504 * need to grab the uring_lock to make progress. If we hold it
9505 * here across the drain wait, then we can deadlock. It's safe
9506 * to drop the mutex here, since no new references will come in
9507 * after we've killed the percpu ref.
9509 mutex_unlock(&ctx->uring_lock);
9511 ret = wait_for_completion_interruptible(&ctx->ref_comp);
9514 if (io_run_task_work_sig() > 0)
9518 mutex_lock(&ctx->uring_lock);
9521 percpu_ref_resurrect(&ctx->refs);
9527 if (ctx->restricted) {
9528 if (opcode >= IORING_REGISTER_LAST) {
9533 if (!test_bit(opcode, ctx->restrictions.register_op)) {
9540 case IORING_REGISTER_BUFFERS:
9541 ret = io_sqe_buffer_register(ctx, arg, nr_args);
9543 case IORING_UNREGISTER_BUFFERS:
9547 ret = io_sqe_buffer_unregister(ctx);
9549 case IORING_REGISTER_FILES:
9550 ret = io_sqe_files_register(ctx, arg, nr_args);
9552 case IORING_UNREGISTER_FILES:
9556 ret = io_sqe_files_unregister(ctx);
9558 case IORING_REGISTER_FILES_UPDATE:
9559 ret = io_sqe_files_update(ctx, arg, nr_args);
9561 case IORING_REGISTER_EVENTFD:
9562 case IORING_REGISTER_EVENTFD_ASYNC:
9566 ret = io_eventfd_register(ctx, arg);
9569 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
9570 ctx->eventfd_async = 1;
9572 ctx->eventfd_async = 0;
9574 case IORING_UNREGISTER_EVENTFD:
9578 ret = io_eventfd_unregister(ctx);
9580 case IORING_REGISTER_PROBE:
9582 if (!arg || nr_args > 256)
9584 ret = io_probe(ctx, arg, nr_args);
9586 case IORING_REGISTER_PERSONALITY:
9590 ret = io_register_personality(ctx);
9592 case IORING_UNREGISTER_PERSONALITY:
9596 ret = io_unregister_personality(ctx, nr_args);
9598 case IORING_REGISTER_ENABLE_RINGS:
9602 ret = io_register_enable_rings(ctx);
9604 case IORING_REGISTER_RESTRICTIONS:
9605 ret = io_register_restrictions(ctx, arg, nr_args);
9613 if (io_register_op_must_quiesce(opcode)) {
9614 /* bring the ctx back to life */
9615 percpu_ref_reinit(&ctx->refs);
9617 reinit_completion(&ctx->ref_comp);
9622 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
9623 void __user *, arg, unsigned int, nr_args)
9625 struct io_ring_ctx *ctx;
9634 if (f.file->f_op != &io_uring_fops)
9637 ctx = f.file->private_data;
9639 mutex_lock(&ctx->uring_lock);
9640 ret = __io_uring_register(ctx, opcode, arg, nr_args);
9641 mutex_unlock(&ctx->uring_lock);
9642 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
9643 ctx->cq_ev_fd != NULL, ret);
9649 static int __init io_uring_init(void)
9651 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
9652 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
9653 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
9656 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
9657 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
9658 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
9659 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
9660 BUILD_BUG_SQE_ELEM(1, __u8, flags);
9661 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
9662 BUILD_BUG_SQE_ELEM(4, __s32, fd);
9663 BUILD_BUG_SQE_ELEM(8, __u64, off);
9664 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
9665 BUILD_BUG_SQE_ELEM(16, __u64, addr);
9666 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
9667 BUILD_BUG_SQE_ELEM(24, __u32, len);
9668 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
9669 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
9670 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
9671 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
9672 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
9673 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
9674 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
9675 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
9676 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
9677 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
9678 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
9679 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
9680 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
9681 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
9682 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
9683 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
9684 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
9685 BUILD_BUG_SQE_ELEM(42, __u16, personality);
9686 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
9688 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
9689 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
9690 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
9693 __initcall(io_uring_init);