| blob | 36a0e5cc3d776337f361f5930b3b225e87aa2ac3 |
1 ext4: handle layout changes to pinned DAX mappings
3 From: Ross Zwisler <ross.zwisler@linux.intel.com>
5 Follow the lead of xfs_break_dax_layouts() and add synchronization between
6 operations in ext4 which remove blocks from an inode (hole punch, truncate
7 down, etc.) and pages which are pinned due to DAX DMA operations.
9 Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com>
10 Signed-off-by: Theodore Ts'o <tytso@mit.edu>
11 Reviewed-by: Jan Kara <jack@suse.cz>
12 Reviewed-by: Lukas Czerner <lczerner@redhat.com>
13 ---
14 fs/ext4/ext4.h | 1 +
15 fs/ext4/extents.c | 17 +++++++++++++++++
16 fs/ext4/inode.c | 46 ++++++++++++++++++++++++++++++++++++++++++++++
17 fs/ext4/truncate.h | 4 ++++
18 4 files changed, 68 insertions(+)
20 diff --git a/fs/ext4/ext4.h b/fs/ext4/ext4.h
21 index 0b127853c584..34bccd64d83d 100644
22 --- a/fs/ext4/ext4.h
23 +++ b/fs/ext4/ext4.h
24 @@ -2460,6 +2460,7 @@ extern int ext4_get_inode_loc(struct inode *, struct ext4_iloc *);
25 extern int ext4_inode_attach_jinode(struct inode *inode);
26 extern int ext4_can_truncate(struct inode *inode);
27 extern int ext4_truncate(struct inode *);
28 +extern int ext4_break_layouts(struct inode *);
29 extern int ext4_punch_hole(struct inode *inode, loff_t offset, loff_t length);
30 extern int ext4_truncate_restart_trans(handle_t *, struct inode *, int nblocks);
31 extern void ext4_set_inode_flags(struct inode *);
32 diff --git a/fs/ext4/extents.c b/fs/ext4/extents.c
33 index 0057fe3f248d..b8161e6b88d1 100644
34 --- a/fs/ext4/extents.c
35 +++ b/fs/ext4/extents.c
36 @@ -4820,6 +4820,13 @@ static long ext4_zero_range(struct file *file, loff_t offset,
37 * released from page cache.
38 */
39 down_write(&EXT4_I(inode)->i_mmap_sem);
40 +
41 + ret = ext4_break_layouts(inode);
42 + if (ret) {
43 + up_write(&EXT4_I(inode)->i_mmap_sem);
44 + goto out_mutex;
45 + }
46 +
47 ret = ext4_update_disksize_before_punch(inode, offset, len);
48 if (ret) {
49 up_write(&EXT4_I(inode)->i_mmap_sem);
50 @@ -5493,6 +5500,11 @@ int ext4_collapse_range(struct inode *inode, loff_t offset, loff_t len)
51 * page cache.
52 */
53 down_write(&EXT4_I(inode)->i_mmap_sem);
54 +
55 + ret = ext4_break_layouts(inode);
56 + if (ret)
57 + goto out_mmap;
58 +
59 /*
60 * Need to round down offset to be aligned with page size boundary
61 * for page size > block size.
62 @@ -5641,6 +5653,11 @@ int ext4_insert_range(struct inode *inode, loff_t offset, loff_t len)
63 * page cache.
64 */
65 down_write(&EXT4_I(inode)->i_mmap_sem);
66 +
67 + ret = ext4_break_layouts(inode);
68 + if (ret)
69 + goto out_mmap;
70 +
71 /*
72 * Need to round down to align start offset to page size boundary
73 * for page size > block size.
74 diff --git a/fs/ext4/inode.c b/fs/ext4/inode.c
75 index 2ea07efbe016..fadb8ecacb1e 100644
76 --- a/fs/ext4/inode.c
77 +++ b/fs/ext4/inode.c
78 @@ -4193,6 +4193,39 @@ int ext4_update_disksize_before_punch(struct inode *inode, loff_t offset,
79 return 0;
80 }
82 +static void ext4_wait_dax_page(struct ext4_inode_info *ei, bool *did_unlock)
83 +{
84 + *did_unlock = true;
85 + up_write(&ei->i_mmap_sem);
86 + schedule();
87 + down_write(&ei->i_mmap_sem);
88 +}
89 +
90 +int ext4_break_layouts(struct inode *inode)
91 +{
92 + struct ext4_inode_info *ei = EXT4_I(inode);
93 + struct page *page;
94 + bool retry;
95 + int error;
96 +
97 + if (WARN_ON_ONCE(!rwsem_is_locked(&ei->i_mmap_sem)))
98 + return -EINVAL;
99 +
100 + do {
101 + retry = false;
102 + page = dax_layout_busy_page(inode->i_mapping);
103 + if (!page)
104 + return 0;
105 +
106 + error = ___wait_var_event(&page->_refcount,
107 + atomic_read(&page->_refcount) == 1,
108 + TASK_INTERRUPTIBLE, 0, 0,
109 + ext4_wait_dax_page(ei, &retry));
110 + } while (error == 0 && retry);
111 +
112 + return error;
113 +}
114 +
115 /*
116 * ext4_punch_hole: punches a hole in a file by releasing the blocks
117 * associated with the given offset and length
118 @@ -4266,6 +4299,11 @@ int ext4_punch_hole(struct inode *inode, loff_t offset, loff_t length)
119 * page cache.
120 */
121 down_write(&EXT4_I(inode)->i_mmap_sem);
122 +
123 + ret = ext4_break_layouts(inode);
124 + if (ret)
125 + goto out_dio;
126 +
127 first_block_offset = round_up(offset, sb->s_blocksize);
128 last_block_offset = round_down((offset + length), sb->s_blocksize) - 1;
130 @@ -5554,6 +5592,14 @@ int ext4_setattr(struct dentry *dentry, struct iattr *attr)
131 ext4_wait_for_tail_page_commit(inode);
132 }
133 down_write(&EXT4_I(inode)->i_mmap_sem);
134 +
135 + rc = ext4_break_layouts(inode);
136 + if (rc) {
137 + up_write(&EXT4_I(inode)->i_mmap_sem);
138 + error = rc;
139 + goto err_out;
140 + }
141 +
142 /*
143 * Truncate pagecache after we've waited for commit
144 * in data=journal mode to make pages freeable.
145 diff --git a/fs/ext4/truncate.h b/fs/ext4/truncate.h
146 index 0cb13badf473..bcbe3668c1d4 100644
147 --- a/fs/ext4/truncate.h
148 +++ b/fs/ext4/truncate.h
149 @@ -11,6 +11,10 @@
150 */
151 static inline void ext4_truncate_failed_write(struct inode *inode)
152 {
153 + /*
154 + * We don't need to call ext4_break_layouts() because the blocks we
155 + * are truncating were never visible to userspace.
156 + */
157 down_write(&EXT4_I(inode)->i_mmap_sem);
158 truncate_inode_pages(inode->i_mapping, inode->i_size);
159 ext4_truncate(inode);
160 --
161 2.14.4