Online defragmentation

Online defragmentation

[Aneesh Kumar K.V]

Hi Takashi,

I was looking at online defrag code and found that the tmp_inode is 
created with tmp_inode->i_nlink equal to zero. Now i am not sure whether 
i understand the code correctly, but AFAIU we allocate contiguous block 
using this tmp_inode. That means tmp_inode have extent details 
corresponding to the blocks. Now we are mapping the file data found in 
the original inode to this new blocks. Towards the end we does a iput. 
In iput since we have i_nlink as zero it will go ahead and call 
generic_delete_inode which will cause these data blocks to be marked 
free (right ?)


I haven't tested the defrag code. This came up when i was doing the 
online migration. With large file having large number of fragmented 
blocks I was getting the below error during the iput of the temporary 
inode.

"ext4_free_blocks Freeing blocks not in datazone ...."

I was able to fix that by setting i_nlink = 1 which will cause the 
generic_forget_inode to be called


-aneesh

Re: Online defragmentation

[Aneesh Kumar K.V]

Aneesh Kumar K.V wrote:
> Hi Takashi,
> 
> I was looking at online defrag code and found that the tmp_inode is 
> created with tmp_inode->i_nlink equal to zero. Now i am not sure whether 
> i understand the code correctly, but AFAIU we allocate contiguous block 
> using this tmp_inode. That means tmp_inode have extent details 
> corresponding to the blocks. Now we are mapping the file data found in 
> the original inode to this new blocks. Towards the end we does a iput. 
> In iput since we have i_nlink as zero it will go ahead and call 
> generic_delete_inode which will cause these data blocks to be marked 
> free (right ?)
> 

Looking at the code again i guess for defragmentation it is okey. I 
guess what actually happens is the blocks that is corresponding to the 
original inode get accounted under tmp_inode. (it actually does a swap 
of blocks ) So doing a iput with i_nlink = 0 is the correct approach.

Correct me if i am wrong.

Sorry for the noise.

-aneesh

Re: Online defragmentation

[Andreas Dilger]

On May 30, 2007  12:34 +0530, Aneesh Kumar K.V wrote:
> I haven't tested the defrag code. This came up when i was doing the 
> online migration. With large file having large number of fragmented 
> blocks I was getting the below error during the iput of the temporary 
> inode.
> 
> "ext4_free_blocks Freeing blocks not in datazone ...."

This shouldn't happen regardless of whether you are doing defrag or not,
since it will cause the filesystem to go read-only.

Cheers, Andreas
--
Andreas Dilger
Principal Software Engineer
Cluster File Systems, Inc.

Re: Online defragmentation

[Takashi Sato]

Hi,

>> I was looking at online defrag code and found that the tmp_inode is created with 
>> tmp_inode->i_nlink equal to zero. Now i am not sure whether i understand the code 
>> correctly, but AFAIU we allocate contiguous block using this tmp_inode. That means 
>> tmp_inode have extent details corresponding to the blocks. Now we are mapping the file 
>> data found in the original inode to this new blocks. Towards the end we does a iput. In 
>> iput since we have i_nlink as zero it will go ahead and call generic_delete_inode which 
>> will cause these data blocks to be marked free (right ?)
>>
>
> Looking at the code again i guess for defragmentation it is okey. I guess what actually 
> happens is the blocks that is corresponding to the original inode get accounted under 
> tmp_inode. (it actually does a swap of blocks ) So doing a iput with i_nlink = 0 is the 
> correct approach.
>
> Correct me if i am wrong.

Your understanding is right.
The iput() is called to free the old blocks which were in the original
inode.

Cheers, Takashi