[Ocfs2-devel] [PATCH] ocfs2: checkpoint appending truncate log transaction before flushing

[piaojun]

Hi Changwei,

On 2019/2/14 16:53, Changwei Ge wrote:
> Hi Jun,
> 
> Thanks for looking into this :-)
> 
> On 2019/2/14 16:24, piaojun wrote:
>> Hi Changwei,
>>
>> On 2019/2/14 12:03, Changwei Ge wrote:
>>> Appending truncate log(TA) and and flushing truncate log(TF) are
>>> two separated transactions. They can be both committed but not
>>> checkpointed. If crash occurs then, both two transaction will be
>>> replayed with several already released to global bitmap clusters.
>>
>> Do you mean that both the two transactions will release cluster to
>> global bitmap? But I think the TA won't give back clusters to global
>> bitmap.
>>
> 
> No, I don't mean that both TA and TF are releasing clusters to global bitmap.
> 
> But consideration into clusters reclaim , clusters will first be recorded in truncate
> log and then be returned to global bitmap, which involves TA and TF jdb2/transactions.
> 
> TA's job is to append cluster records to truncate log, by which we can overcome a potential space leak problem.
> TF's job is to return clusters to global bitmap.
> 
> It's possible that TA and TF are both committed to JBD but sadly none of them is check-pointed.
> So journal replaying need to replay both TA and TF during next mount.
> Then there is a record residing in truncate log representing the already released cluster
> which has been returned to global bitmap by replaying TF.
> 
> Now the double free shows up.

Do you mean that when mount again, truncate log recovery will find
record residing in truncate log which already released? But after the
TF transaction replayed during mount, truncate log won't be recovered
as tl->tl_used is less than tl->tl_count.

Thanks,
Jun

> 
> 
>>> Then truncate log will be replayed resulting in cluster double free.
>>
>> Does this problem only cause some error log? As below:
>>
>> ocfs2_replay_truncate_records
>>    ocfs2_free_clusters
>>      _ocfs2_free_clusters
>>        _ocfs2_free_suballoc_bits
>>          ocfs2_block_group_clear_bits
>>            "Trying to clear %u bits at offset %u in group descriptor"
>>
> 
> Exactly, when the issue occurs, it will be printed as above.
> 
> Thanks,
> Changwei
> 
>> Thanks,
>> Jun
>>
>>>
>>> To reproduce this issue, just crash the host while punching hole to files.
>>>
>>> Signed-off-by: Changwei Ge <ge.changwei at h3c.com>
>>> ---
>>>   fs/ocfs2/alloc.c | 15 +++++++++++++++
>>>   1 file changed, 15 insertions(+)
>>>
>>> diff --git a/fs/ocfs2/alloc.c b/fs/ocfs2/alloc.c
>>> index d1cbb27..29bc777 100644
>>> --- a/fs/ocfs2/alloc.c
>>> +++ b/fs/ocfs2/alloc.c
>>> @@ -6007,6 +6007,7 @@ int __ocfs2_flush_truncate_log(struct ocfs2_super *osb)
>>>   	struct buffer_head *data_alloc_bh = NULL;
>>>   	struct ocfs2_dinode *di;
>>>   	struct ocfs2_truncate_log *tl;
>>> +	struct ocfs2_journal *journal = osb->journal;
>>>   
>>>   	BUG_ON(inode_trylock(tl_inode));
>>>   
>>> @@ -6027,6 +6028,20 @@ int __ocfs2_flush_truncate_log(struct ocfs2_super *osb)
>>>   		goto out;
>>>   	}
>>>   
>>> +	/* Appending truncate log(TA) and and flushing truncate log(TF) are
>>> +	 * two separated transactions. They can be both committed but not
>>> +	 * checkpointed. If crash occurs then, both two transaction will be
>>> +	 * replayed with several already released to global bitmap clusters.
>>> +	 * Then truncate log will be replayed resulting in cluster double free.
>>> +	 */
>>> +	jbd2_journal_lock_updates(journal->j_journal);
>>> +	status = jbd2_journal_flush(journal->j_journal);
>>> +	jbd2_journal_unlock_updates(journal->j_journal);
>>> +	if (status < 0) {
>>> +		mlog_errno(status);
>>> +		goto out;
>>> +	}
>>> +
>>>   	data_alloc_inode = ocfs2_get_system_file_inode(osb,
>>>   						       GLOBAL_BITMAP_SYSTEM_INODE,
>>>   						       OCFS2_INVALID_SLOT);
>>>
>>
> .
>