Loading...
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
37static 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 */
78extern 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 if (list_empty(&ei->i_completed_io_list))
87 return ret;
88
89 dump_completed_IO(inode);
90 spin_lock_irqsave(&ei->i_completed_io_lock, flags);
91 while (!list_empty(&ei->i_completed_io_list)){
92 io = list_entry(ei->i_completed_io_list.next,
93 ext4_io_end_t, list);
94 /*
95 * Calling ext4_end_io_nolock() to convert completed
96 * IO to written.
97 *
98 * When ext4_sync_file() is called, run_queue() may already
99 * about to flush the work corresponding to this io structure.
100 * It will be upset if it founds the io structure related
101 * to the work-to-be schedule is freed.
102 *
103 * Thus we need to keep the io structure still valid here after
104 * conversion finished. The io structure has a flag to
105 * avoid double converting from both fsync and background work
106 * queue work.
107 */
108 spin_unlock_irqrestore(&ei->i_completed_io_lock, flags);
109 ret = ext4_end_io_nolock(io);
110 spin_lock_irqsave(&ei->i_completed_io_lock, flags);
111 if (ret < 0)
112 ret2 = ret;
113 else
114 list_del_init(&io->list);
115 }
116 spin_unlock_irqrestore(&ei->i_completed_io_lock, flags);
117 return (ret2 < 0) ? ret2 : 0;
118}
119
120/*
121 * If we're not journaling and this is a just-created file, we have to
122 * sync our parent directory (if it was freshly created) since
123 * otherwise it will only be written by writeback, leaving a huge
124 * window during which a crash may lose the file. This may apply for
125 * the parent directory's parent as well, and so on recursively, if
126 * they are also freshly created.
127 */
128static int ext4_sync_parent(struct inode *inode)
129{
130 struct writeback_control wbc;
131 struct dentry *dentry = NULL;
132 struct inode *next;
133 int ret = 0;
134
135 if (!ext4_test_inode_state(inode, EXT4_STATE_NEWENTRY))
136 return 0;
137 inode = igrab(inode);
138 while (ext4_test_inode_state(inode, EXT4_STATE_NEWENTRY)) {
139 ext4_clear_inode_state(inode, EXT4_STATE_NEWENTRY);
140 dentry = NULL;
141 spin_lock(&inode->i_lock);
142 if (!list_empty(&inode->i_dentry)) {
143 dentry = list_first_entry(&inode->i_dentry,
144 struct dentry, d_alias);
145 dget(dentry);
146 }
147 spin_unlock(&inode->i_lock);
148 if (!dentry)
149 break;
150 next = igrab(dentry->d_parent->d_inode);
151 dput(dentry);
152 if (!next)
153 break;
154 iput(inode);
155 inode = next;
156 ret = sync_mapping_buffers(inode->i_mapping);
157 if (ret)
158 break;
159 memset(&wbc, 0, sizeof(wbc));
160 wbc.sync_mode = WB_SYNC_ALL;
161 wbc.nr_to_write = 0; /* only write out the inode */
162 ret = sync_inode(inode, &wbc);
163 if (ret)
164 break;
165 }
166 iput(inode);
167 return ret;
168}
169
170/**
171 * __sync_file - generic_file_fsync without the locking and filemap_write
172 * @inode: inode to sync
173 * @datasync: only sync essential metadata if true
174 *
175 * This is just generic_file_fsync without the locking. This is needed for
176 * nojournal mode to make sure this inodes data/metadata makes it to disk
177 * properly. The i_mutex should be held already.
178 */
179static int __sync_inode(struct inode *inode, int datasync)
180{
181 int err;
182 int ret;
183
184 ret = sync_mapping_buffers(inode->i_mapping);
185 if (!(inode->i_state & I_DIRTY))
186 return ret;
187 if (datasync && !(inode->i_state & I_DIRTY_DATASYNC))
188 return ret;
189
190 err = sync_inode_metadata(inode, 1);
191 if (ret == 0)
192 ret = err;
193 return ret;
194}
195
196/*
197 * akpm: A new design for ext4_sync_file().
198 *
199 * This is only called from sys_fsync(), sys_fdatasync() and sys_msync().
200 * There cannot be a transaction open by this task.
201 * Another task could have dirtied this inode. Its data can be in any
202 * state in the journalling system.
203 *
204 * What we do is just kick off a commit and wait on it. This will snapshot the
205 * inode to disk.
206 *
207 * i_mutex lock is held when entering and exiting this function
208 */
209
210int ext4_sync_file(struct file *file, loff_t start, loff_t end, int datasync)
211{
212 struct inode *inode = file->f_mapping->host;
213 struct ext4_inode_info *ei = EXT4_I(inode);
214 journal_t *journal = EXT4_SB(inode->i_sb)->s_journal;
215 int ret;
216 tid_t commit_tid;
217 bool needs_barrier = false;
218
219 J_ASSERT(ext4_journal_current_handle() == NULL);
220
221 trace_ext4_sync_file_enter(file, datasync);
222
223 ret = filemap_write_and_wait_range(inode->i_mapping, start, end);
224 if (ret)
225 return ret;
226 mutex_lock(&inode->i_mutex);
227
228 if (inode->i_sb->s_flags & MS_RDONLY)
229 goto out;
230
231 ret = ext4_flush_completed_IO(inode);
232 if (ret < 0)
233 goto out;
234
235 if (!journal) {
236 ret = __sync_inode(inode, datasync);
237 if (!ret && !list_empty(&inode->i_dentry))
238 ret = ext4_sync_parent(inode);
239 goto out;
240 }
241
242 /*
243 * data=writeback,ordered:
244 * The caller's filemap_fdatawrite()/wait will sync the data.
245 * Metadata is in the journal, we wait for proper transaction to
246 * commit here.
247 *
248 * data=journal:
249 * filemap_fdatawrite won't do anything (the buffers are clean).
250 * ext4_force_commit will write the file data into the journal and
251 * will wait on that.
252 * filemap_fdatawait() will encounter a ton of newly-dirtied pages
253 * (they were dirtied by commit). But that's OK - the blocks are
254 * safe in-journal, which is all fsync() needs to ensure.
255 */
256 if (ext4_should_journal_data(inode)) {
257 ret = ext4_force_commit(inode->i_sb);
258 goto out;
259 }
260
261 commit_tid = datasync ? ei->i_datasync_tid : ei->i_sync_tid;
262 if (journal->j_flags & JBD2_BARRIER &&
263 !jbd2_trans_will_send_data_barrier(journal, commit_tid))
264 needs_barrier = true;
265 jbd2_log_start_commit(journal, commit_tid);
266 ret = jbd2_log_wait_commit(journal, commit_tid);
267 if (needs_barrier)
268 blkdev_issue_flush(inode->i_sb->s_bdev, GFP_KERNEL, NULL);
269 out:
270 mutex_unlock(&inode->i_mutex);
271 trace_ext4_sync_file_exit(inode, ret);
272 return ret;
273}
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 */
45static 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
81static 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
104static 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 */
129int 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 }
181out:
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}