Re: [Gluster-devel] Need inputs for fixing 866908

[Vijay Bellur]

On 12/12/2012 02:09 PM, Pranith Kumar Karampuri wrote:
> hi,
>    https://bugzilla.redhat.com/show_bug.cgi?id=866908.
> gluster volume heal <vol_name> info" command outputs entries which are in ".glusterfs" 
> and ".landfill" directory
>
> Reason for the issue:
> Whenever a brick is restarted and self-heal has to remove a directory 'dir' it sets a 
> flag indicating in rmdir fop that it is 'rm -rf'. The 'dir' is moved to 
> '.glusterfs/landfill' by posix because of the flags in rmdir fop, Janitor thread 
> removes it asynchronously. gfid-handles exists for 'dir' and files under 'dir' until 
> the janitor thread deletes them from landfill. Janitor thread is run every 10 
> minutes. Self-heald depends on gfid-handle presence to determine that a file exists 
> in the filesystem. Because of this behaviour, stale entries show up in the output of 
> 'gluster volume heal <volname> info'. The issue is transient, after 10 minutes 
> the stale entries won't show up in the output. One way to fix it is by waking up 
> Janitor thread as soon as an rmdir fop comes with a special flag indicating that rm 
> -rf needs to happen inside landfill. We can even close the issue documenting the 
> behaviour as known issue. Let me know your inputs.

The problem as reported seems to be with self-heal-daemon informing the 
CLI that there are entries to be healed and some of them not being 
relevant as they are part of .landfill. I cannot think of a foolproof 
solution for this right away as a rm -rf might happen after 
self-heal-daemon has sent out a response that contains entries needing 
self-heal. I think a better way out here is to educate users about 
entries seen from .landfill in the output of heal * commands as 
transient and not being relevant for self-healing.

-Vijay