1 /*
2  *  linux/fs/ext4/fsync.c
3  *
4  *  Copyright (C) 1993  Stephen Tweedie (sct@redhat.com)
5  *  from
6  *  Copyright (C) 1992  Remy Card (card@masi.ibp.fr)
7  *                      Laboratoire MASI - Institut Blaise Pascal
8  *                      Universite Pierre et Marie Curie (Paris VI)
9  *  from
10  *  linux/fs/minix/truncate.c   Copyright (C) 1991, 1992  Linus Torvalds
11  *
12  *  ext4fs fsync primitive
13  *
14  *  Big-endian to little-endian byte-swapping/bitmaps by
15  *        David S. Miller (davem@caip.rutgers.edu), 1995
16  *
17  *  Removed unnecessary code duplication for little endian machines
18  *  and excessive __inline__s.
19  *        Andi Kleen, 1997
20  *
21  * Major simplications and cleanup - we only need to do the metadata, because
22  * we can depend on generic_block_fdatasync() to sync the data blocks.
23  */
24 
25 #include <linux/time.h>
26 #include <linux/fs.h>
27 #include <linux/sched.h>
28 #include <linux/writeback.h>
29 #include <linux/jbd2.h>
30 #include <linux/blkdev.h>
31 
32 #include "ext4.h"
33 #include "ext4_jbd2.h"
34 
35 #include <trace/events/ext4.h>
36 
dump_completed_IO(struct inode * inode)37 static void dump_completed_IO(struct inode * inode)
38 {
39 #ifdef	EXT4FS_DEBUG
40 	struct list_head *cur, *before, *after;
41 	ext4_io_end_t *io, *io0, *io1;
42 	unsigned long flags;
43 
44 	if (list_empty(&EXT4_I(inode)->i_completed_io_list)){
45 		ext4_debug("inode %lu completed_io list is empty\n", inode->i_ino);
46 		return;
47 	}
48 
49 	ext4_debug("Dump inode %lu completed_io list \n", inode->i_ino);
50 	spin_lock_irqsave(&EXT4_I(inode)->i_completed_io_lock, flags);
51 	list_for_each_entry(io, &EXT4_I(inode)->i_completed_io_list, list){
52 		cur = &io->list;
53 		before = cur->prev;
54 		io0 = container_of(before, ext4_io_end_t, list);
55 		after = cur->next;
56 		io1 = container_of(after, ext4_io_end_t, list);
57 
58 		ext4_debug("io 0x%p from inode %lu,prev 0x%p,next 0x%p\n",
59 			    io, inode->i_ino, io0, io1);
60 	}
61 	spin_unlock_irqrestore(&EXT4_I(inode)->i_completed_io_lock, flags);
62 #endif
63 }
64 
65 /*
66  * This function is called from ext4_sync_file().
67  *
68  * When IO is completed, the work to convert unwritten extents to
69  * written is queued on workqueue but may not get immediately
70  * scheduled. When fsync is called, we need to ensure the
71  * conversion is complete before fsync returns.
72  * The inode keeps track of a list of pending/completed IO that
73  * might needs to do the conversion. This function walks through
74  * the list and convert the related unwritten extents for completed IO
75  * to written.
76  * The function return the number of pending IOs on success.
77  */
ext4_flush_completed_IO(struct inode * inode)78 int ext4_flush_completed_IO(struct inode *inode)
79 {
80 	ext4_io_end_t *io;
81 	struct ext4_inode_info *ei = EXT4_I(inode);
82 	unsigned long flags;
83 	int ret = 0;
84 	int ret2 = 0;
85 
86 	dump_completed_IO(inode);
87 	spin_lock_irqsave(&ei->i_completed_io_lock, flags);
88 	while (!list_empty(&ei->i_completed_io_list)){
89 		io = list_entry(ei->i_completed_io_list.next,
90 				ext4_io_end_t, list);
91 		list_del_init(&io->list);
92 		io->flag |= EXT4_IO_END_IN_FSYNC;
93 		/*
94 		 * Calling ext4_end_io_nolock() to convert completed
95 		 * IO to written.
96 		 *
97 		 * When ext4_sync_file() is called, run_queue() may already
98 		 * about to flush the work corresponding to this io structure.
99 		 * It will be upset if it founds the io structure related
100 		 * to the work-to-be schedule is freed.
101 		 *
102 		 * Thus we need to keep the io structure still valid here after
103 		 * conversion finished. The io structure has a flag to
104 		 * avoid double converting from both fsync and background work
105 		 * queue work.
106 		 */
107 		spin_unlock_irqrestore(&ei->i_completed_io_lock, flags);
108 		ret = ext4_end_io_nolock(io);
109 		if (ret < 0)
110 			ret2 = ret;
111 		spin_lock_irqsave(&ei->i_completed_io_lock, flags);
112 		io->flag &= ~EXT4_IO_END_IN_FSYNC;
113 	}
114 	spin_unlock_irqrestore(&ei->i_completed_io_lock, flags);
115 	return (ret2 < 0) ? ret2 : 0;
116 }
117 
118 /*
119  * If we're not journaling and this is a just-created file, we have to
120  * sync our parent directory (if it was freshly created) since
121  * otherwise it will only be written by writeback, leaving a huge
122  * window during which a crash may lose the file.  This may apply for
123  * the parent directory's parent as well, and so on recursively, if
124  * they are also freshly created.
125  */
ext4_sync_parent(struct inode * inode)126 static int ext4_sync_parent(struct inode *inode)
127 {
128 	struct writeback_control wbc;
129 	struct dentry *dentry = NULL;
130 	struct inode *next;
131 	int ret = 0;
132 
133 	if (!ext4_test_inode_state(inode, EXT4_STATE_NEWENTRY))
134 		return 0;
135 	inode = igrab(inode);
136 	while (ext4_test_inode_state(inode, EXT4_STATE_NEWENTRY)) {
137 		ext4_clear_inode_state(inode, EXT4_STATE_NEWENTRY);
138 		dentry = NULL;
139 		spin_lock(&inode->i_lock);
140 		if (!list_empty(&inode->i_dentry)) {
141 			dentry = list_first_entry(&inode->i_dentry,
142 						  struct dentry, d_alias);
143 			dget(dentry);
144 		}
145 		spin_unlock(&inode->i_lock);
146 		if (!dentry)
147 			break;
148 		next = igrab(dentry->d_parent->d_inode);
149 		dput(dentry);
150 		if (!next)
151 			break;
152 		iput(inode);
153 		inode = next;
154 		ret = sync_mapping_buffers(inode->i_mapping);
155 		if (ret)
156 			break;
157 		memset(&wbc, 0, sizeof(wbc));
158 		wbc.sync_mode = WB_SYNC_ALL;
159 		wbc.nr_to_write = 0;         /* only write out the inode */
160 		ret = sync_inode(inode, &wbc);
161 		if (ret)
162 			break;
163 	}
164 	iput(inode);
165 	return ret;
166 }
167 
168 /**
169  * __sync_file - generic_file_fsync without the locking and filemap_write
170  * @inode:	inode to sync
171  * @datasync:	only sync essential metadata if true
172  *
173  * This is just generic_file_fsync without the locking.  This is needed for
174  * nojournal mode to make sure this inodes data/metadata makes it to disk
175  * properly.  The i_mutex should be held already.
176  */
__sync_inode(struct inode * inode,int datasync)177 static int __sync_inode(struct inode *inode, int datasync)
178 {
179 	int err;
180 	int ret;
181 
182 	ret = sync_mapping_buffers(inode->i_mapping);
183 	if (!(inode->i_state & I_DIRTY))
184 		return ret;
185 	if (datasync && !(inode->i_state & I_DIRTY_DATASYNC))
186 		return ret;
187 
188 	err = sync_inode_metadata(inode, 1);
189 	if (ret == 0)
190 		ret = err;
191 	return ret;
192 }
193 
194 /*
195  * akpm: A new design for ext4_sync_file().
196  *
197  * This is only called from sys_fsync(), sys_fdatasync() and sys_msync().
198  * There cannot be a transaction open by this task.
199  * Another task could have dirtied this inode.  Its data can be in any
200  * state in the journalling system.
201  *
202  * What we do is just kick off a commit and wait on it.  This will snapshot the
203  * inode to disk.
204  *
205  * i_mutex lock is held when entering and exiting this function
206  */
207 
ext4_sync_file(struct file * file,loff_t start,loff_t end,int datasync)208 int ext4_sync_file(struct file *file, loff_t start, loff_t end, int datasync)
209 {
210 	struct inode *inode = file->f_mapping->host;
211 	struct ext4_inode_info *ei = EXT4_I(inode);
212 	journal_t *journal = EXT4_SB(inode->i_sb)->s_journal;
213 	int ret;
214 	tid_t commit_tid;
215 	bool needs_barrier = false;
216 
217 	J_ASSERT(ext4_journal_current_handle() == NULL);
218 
219 	trace_ext4_sync_file_enter(file, datasync);
220 
221 	ret = filemap_write_and_wait_range(inode->i_mapping, start, end);
222 	if (ret)
223 		return ret;
224 	mutex_lock(&inode->i_mutex);
225 
226 	if (inode->i_sb->s_flags & MS_RDONLY)
227 		goto out;
228 
229 	ret = ext4_flush_completed_IO(inode);
230 	if (ret < 0)
231 		goto out;
232 
233 	if (!journal) {
234 		ret = __sync_inode(inode, datasync);
235 		if (!ret && !list_empty(&inode->i_dentry))
236 			ret = ext4_sync_parent(inode);
237 		goto out;
238 	}
239 
240 	/*
241 	 * data=writeback,ordered:
242 	 *  The caller's filemap_fdatawrite()/wait will sync the data.
243 	 *  Metadata is in the journal, we wait for proper transaction to
244 	 *  commit here.
245 	 *
246 	 * data=journal:
247 	 *  filemap_fdatawrite won't do anything (the buffers are clean).
248 	 *  ext4_force_commit will write the file data into the journal and
249 	 *  will wait on that.
250 	 *  filemap_fdatawait() will encounter a ton of newly-dirtied pages
251 	 *  (they were dirtied by commit).  But that's OK - the blocks are
252 	 *  safe in-journal, which is all fsync() needs to ensure.
253 	 */
254 	if (ext4_should_journal_data(inode)) {
255 		ret = ext4_force_commit(inode->i_sb);
256 		goto out;
257 	}
258 
259 	commit_tid = datasync ? ei->i_datasync_tid : ei->i_sync_tid;
260 	if (journal->j_flags & JBD2_BARRIER &&
261 	    !jbd2_trans_will_send_data_barrier(journal, commit_tid))
262 		needs_barrier = true;
263 	ret = jbd2_complete_transaction(journal, commit_tid);
264 	if (needs_barrier)
265 		blkdev_issue_flush(inode->i_sb->s_bdev, GFP_KERNEL, NULL);
266  out:
267 	mutex_unlock(&inode->i_mutex);
268 	trace_ext4_sync_file_exit(inode, ret);
269 	return ret;
270 }
271