Re: [PATCH v3 1/2] ext4: commit super block if fs record error when journal record without error

["yebin (H)" <yebin10@huawei.com>]

On 2023/2/17 1:31, Jan Kara wrote:
> On Tue 14-02-23 10:29:04, Ye Bin wrote:
>> From: Ye Bin <yebin10@huawei.com>
>>
>> Now, 'es->s_state' maybe covered by recover journal. And journal errno
>> maybe not recorded in journal sb as IO error. ext4_update_super() only
>> update error information when 'sbi->s_add_error_count' large than zero.
>> Then 'EXT4_ERROR_FS' flag maybe lost.
>> To solve above issue commit error information after recover journal.
>>
>> Signed-off-by: Ye Bin <yebin10@huawei.com>
>> ---
>>   fs/ext4/super.c | 12 ++++++++++++
>>   1 file changed, 12 insertions(+)
>>
>> diff --git a/fs/ext4/super.c b/fs/ext4/super.c
>> index dc3907dff13a..b94754ba8556 100644
>> --- a/fs/ext4/super.c
>> +++ b/fs/ext4/super.c
>> @@ -5932,6 +5932,18 @@ static int ext4_load_journal(struct super_block *sb,
>>   		goto err_out;
>>   	}
>>   
>> +	if (unlikely(es->s_error_count && !jbd2_journal_errno(journal) &&
>> +		     !(le16_to_cpu(es->s_state) & EXT4_ERROR_FS))) {
>> +		EXT4_SB(sb)->s_mount_state |= EXT4_ERROR_FS;
>> +		es->s_state |= cpu_to_le16(EXT4_ERROR_FS);
>> +		err = ext4_commit_super(sb);
>> +		if (err) {
>> +			ext4_msg(sb, KERN_ERR,
>> +				 "Failed to commit error information, please repair fs force!");
>> +			goto err_out;
>> +		}
>> +	}
>> +
> Hum, I'm not sure I follow here. If journal replay has overwritten the
> superblock (and thus the stored error info), then I'd expect
> es->s_error_count got overwritten (possibly to 0) as well. And this is
> actually relatively realistic scenario with errors=remount-ro behavior when
> the first fs error happens.
>
> What I intended in my original suggestion was to save es->s_error_count,
> es->s_state & EXT4_ERROR_FS, es->s_first_error_*, es->s_last_error_* before
> doing journal replay in ext4_load_journal() and then after journal replay
> merge this info back to the superblock
Actually，commit 1c13d5c08728 ("ext4: Save error information to the 
superblock for analysis")
already merged error info back to the superblock after journal replay 
except 'es->s_state'.
The problem I have now is that the error flag in the journal superblock 
was not recorded,
but the error message was recorded in the superblock. So it leads to 
ext4_clear_journal_err()
does not detect errors and marks the file system as an error. Because 
ext4_update_super() is
only set error flag  when 'sbi->s_add_error_count  > 0'. Although 
'sbi->s_mount_state' is
written to the super block when umount, but it is also conditional.
So I handle the scenario "es->s_error_count && 
!jbd2_journal_errno(journal) &&
!(le16_to_cpu(es->s_state) & EXT4_ERROR_FS)". Maybe we can just store
'EXT4_SB(sb)->s_mount_state & EXT4_ERROR_FS' back to the superblock. But i
prefer to mark fs as error if it contain detail error info without 
EXT4_ERROR_FS.
> - if EXT4_ERROR_FS was set, set it
> now as well, take max of old and new s_error_count, set s_first_error_* if
> it is now unset, set s_last_error_* if stored timestamp is newer than
> current timestamp.
>
> Or am I overengineering it now? :)
>
> 								Honza