Re: [RFC][PATCH] btrfs: add an autodefrag property

[Josef Bacik <josef@toxicpanda.com>]

On 7/17/20 7:46 PM, Qu Wenruo wrote:
> 
> 
> On 2020/7/18 上午4:42, Josef Bacik wrote:
>> Autodefrag is very useful for somethings, like the 9000 sqllite files
>> that Firefox uses, but is way less useful for virt images.
>> Unfortunately this is only available currently as a whole mount option.
>> Fix this by adding an "autodefrag" property, that users can set on a per
>> file or per directory basis.  Thus allowing them to control where
>> exactly the extra write activity is going to occur.
>>
>> Signed-off-by: Josef Bacik <josef@toxicpanda.com>
>> ---
>> - This is an RFC because I want to make sure we're ok with this before I go and
>>    add btrfs-progs support for this.  I'm not married to the name or the value,
>>    but I think the core goal is valuable.
> 
> The idea looks pretty good to me.
> 
> Although it would be much more convincing to bring some real-world micro
> bench to show the benefit.
> 

The same benefit as what existed originally for autodefrag, firefox isn't 
unusably slow on spinning rust.

> 
> However I still have a concern related to defrag.
> Since it's on-disk flag, thus can be inherited by snapshot, then what
> would happen if an auto-defrag inode get snapshotted.
> 
> Would any write to the auto-defrag inode in new snapshot break the space
> saving?
Sure but that's the case for all defrag, and this is in fact better than mount 
-o autodefrag because you can limit the damage.  Thanks,

Josef