1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  *  linux/fs/ext4/fsync.c
4  *
5  *  Copyright (C) 1993  Stephen Tweedie (sct@redhat.com)
6  *  from
7  *  Copyright (C) 1992  Remy Card (card@masi.ibp.fr)
8  *                      Laboratoire MASI - Institut Blaise Pascal
9  *                      Universite Pierre et Marie Curie (Paris VI)
10  *  from
11  *  linux/fs/minix/truncate.c   Copyright (C) 1991, 1992  Linus Torvalds
12  *
13  *  ext4fs fsync primitive
14  *
15  *  Big-endian to little-endian byte-swapping/bitmaps by
16  *        David S. Miller (davem@caip.rutgers.edu), 1995
17  *
18  *  Removed unnecessary code duplication for little endian machines
19  *  and excessive __inline__s.
20  *        Andi Kleen, 1997
21  *
22  * Major simplications and cleanup - we only need to do the metadata, because
23  * we can depend on generic_block_fdatasync() to sync the data blocks.
24  */
25 
26 #include <linux/time.h>
27 #include <linux/fs.h>
28 #include <linux/sched.h>
29 #include <linux/writeback.h>
30 #include <linux/blkdev.h>
31 
32 #include "ext4.h"
33 #include "ext4_jbd2.h"
34 
35 #include <trace/events/ext4.h>
36 
37 /*
38  * If we're not journaling and this is a just-created file, we have to
39  * sync our parent directory (if it was freshly created) since
40  * otherwise it will only be written by writeback, leaving a huge
41  * window during which a crash may lose the file.  This may apply for
42  * the parent directory's parent as well, and so on recursively, if
43  * they are also freshly created.
44  */
ext4_sync_parent(struct inode * inode)45 static int ext4_sync_parent(struct inode *inode)
46 {
47 	struct dentry *dentry, *next;
48 	int ret = 0;
49 
50 	if (!ext4_test_inode_state(inode, EXT4_STATE_NEWENTRY))
51 		return 0;
52 	dentry = d_find_any_alias(inode);
53 	if (!dentry)
54 		return 0;
55 	while (ext4_test_inode_state(inode, EXT4_STATE_NEWENTRY)) {
56 		ext4_clear_inode_state(inode, EXT4_STATE_NEWENTRY);
57 
58 		next = dget_parent(dentry);
59 		dput(dentry);
60 		dentry = next;
61 		inode = dentry->d_inode;
62 
63 		/*
64 		 * The directory inode may have gone through rmdir by now. But
65 		 * the inode itself and its blocks are still allocated (we hold
66 		 * a reference to the inode via its dentry), so it didn't go
67 		 * through ext4_evict_inode()) and so we are safe to flush
68 		 * metadata blocks and the inode.
69 		 */
70 		ret = sync_mapping_buffers(inode->i_mapping);
71 		if (ret)
72 			break;
73 		ret = sync_inode_metadata(inode, 1);
74 		if (ret)
75 			break;
76 	}
77 	dput(dentry);
78 	return ret;
79 }
80 
ext4_fsync_nojournal(struct inode * inode,bool datasync,bool * needs_barrier)81 static int ext4_fsync_nojournal(struct inode *inode, bool datasync,
82 				bool *needs_barrier)
83 {
84 	int ret, err;
85 
86 	ret = sync_mapping_buffers(inode->i_mapping);
87 	if (!(inode->i_state & I_DIRTY_ALL))
88 		return ret;
89 	if (datasync && !(inode->i_state & I_DIRTY_DATASYNC))
90 		return ret;
91 
92 	err = sync_inode_metadata(inode, 1);
93 	if (!ret)
94 		ret = err;
95 
96 	if (!ret)
97 		ret = ext4_sync_parent(inode);
98 	if (test_opt(inode->i_sb, BARRIER))
99 		*needs_barrier = true;
100 
101 	return ret;
102 }
103 
ext4_fsync_journal(struct inode * inode,bool datasync,bool * needs_barrier)104 static int ext4_fsync_journal(struct inode *inode, bool datasync,
105 			     bool *needs_barrier)
106 {
107 	struct ext4_inode_info *ei = EXT4_I(inode);
108 	journal_t *journal = EXT4_SB(inode->i_sb)->s_journal;
109 	tid_t commit_tid = datasync ? ei->i_datasync_tid : ei->i_sync_tid;
110 
111 	if (journal->j_flags & JBD2_BARRIER &&
112 	    !jbd2_trans_will_send_data_barrier(journal, commit_tid))
113 		*needs_barrier = true;
114 
115 	return ext4_fc_commit(journal, commit_tid);
116 }
117 
118 /*
119  * akpm: A new design for ext4_sync_file().
120  *
121  * This is only called from sys_fsync(), sys_fdatasync() and sys_msync().
122  * There cannot be a transaction open by this task.
123  * Another task could have dirtied this inode.  Its data can be in any
124  * state in the journalling system.
125  *
126  * What we do is just kick off a commit and wait on it.  This will snapshot the
127  * inode to disk.
128  */
ext4_sync_file(struct file * file,loff_t start,loff_t end,int datasync)129 int ext4_sync_file(struct file *file, loff_t start, loff_t end, int datasync)
130 {
131 	int ret = 0, err;
132 	bool needs_barrier = false;
133 	struct inode *inode = file->f_mapping->host;
134 	struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
135 
136 	if (unlikely(ext4_forced_shutdown(sbi)))
137 		return -EIO;
138 
139 	ASSERT(ext4_journal_current_handle() == NULL);
140 
141 	trace_ext4_sync_file_enter(file, datasync);
142 
143 	if (sb_rdonly(inode->i_sb)) {
144 		/* Make sure that we read updated s_mount_flags value */
145 		smp_rmb();
146 		if (ext4_test_mount_flag(inode->i_sb, EXT4_MF_FS_ABORTED))
147 			ret = -EROFS;
148 		goto out;
149 	}
150 
151 	ret = file_write_and_wait_range(file, start, end);
152 	if (ret)
153 		goto out;
154 
155 	/*
156 	 * data=writeback,ordered:
157 	 *  The caller's filemap_fdatawrite()/wait will sync the data.
158 	 *  Metadata is in the journal, we wait for proper transaction to
159 	 *  commit here.
160 	 *
161 	 * data=journal:
162 	 *  filemap_fdatawrite won't do anything (the buffers are clean).
163 	 *  ext4_force_commit will write the file data into the journal and
164 	 *  will wait on that.
165 	 *  filemap_fdatawait() will encounter a ton of newly-dirtied pages
166 	 *  (they were dirtied by commit).  But that's OK - the blocks are
167 	 *  safe in-journal, which is all fsync() needs to ensure.
168 	 */
169 	if (!sbi->s_journal)
170 		ret = ext4_fsync_nojournal(inode, datasync, &needs_barrier);
171 	else if (ext4_should_journal_data(inode))
172 		ret = ext4_force_commit(inode->i_sb);
173 	else
174 		ret = ext4_fsync_journal(inode, datasync, &needs_barrier);
175 
176 	if (needs_barrier) {
177 		err = blkdev_issue_flush(inode->i_sb->s_bdev);
178 		if (!ret)
179 			ret = err;
180 	}
181 out:
182 	err = file_check_and_advance_wb_err(file);
183 	if (ret == 0)
184 		ret = err;
185 	trace_ext4_sync_file_exit(inode, ret);
186 	return ret;
187 }
188