[Gluster-users] Fixing split brain

[Pranith Kumar K]

hi Emmanuel,
      I see that with master branch code, the test case works fine for 
me. The information that the file went into a split brain situation is 
stored in inode
You can check the code of the function afr_set_split_brain (xlator_t 
*this, inode_t *inode, gf_boolean_t set)

Pranith
pranith @ /mnt/client2
13:06:00 :) $ ls abc
ls: cannot access abc: Input/output error

pranith @ /mnt/client2
13:06:35 :( $ sudo getfattr -d -m . /tmp/{1,2}/abc
getfattr: Removing leading '/' from absolute path names
# file: tmp/1/abc
trusted.gfid=0s1aS1Nib4QrmwS0gnCdIP4g==

# file: tmp/2/abc
trusted.gfid=0sq3um4Gh2TN+zkO9ULMpgkw==


pranith @ /mnt/client2
13:06:53 :) $ sudo rm -f /tmp/1/abc

pranith @ /mnt/client2
13:07:03 :) $ ls abc
abc

pranith @ /mnt/client2
13:07:08 :) $ sudo getfattr -d -m . /tmp/{1,2}/abc
getfattr: Removing leading '/' from absolute path names
# file: tmp/1/abc
trusted.afr.vol-client-0=0sAAAAAAAAAAAAAAAA
trusted.afr.vol-client-1=0sAAAAAAAAAAAAAAAA
trusted.gfid=0sq3um4Gh2TN+zkO9ULMpgkw==

# file: tmp/2/abc
trusted.afr.vol-client-0=0sAAAAAAAAAAAAAAAA
trusted.afr.vol-client-1=0sAAAAAP////8AAAAA
trusted.gfid=0sq3um4Gh2TN+zkO9ULMpgkw==

Pranith

On 08/21/2011 05:23 PM, Emmanuel Dreyfus wrote:
> Hi
>
> Consider the typical spit brain situation: reading from file gets EIO,
> logs say:
> [2011-08-21 13:38:54.607590] W [afr-open.c:168:afr_open]
>     0-gfs-replicate-0: failed to open as split brain seen, returning EIO
> [2011-08-21 13:38:54.607895] W [fuse-bridge.c:585:fuse_fd_cbk]
>     0-glusterfs-fuse: 1371456: OPEN()
>    /manu/netbsd/usr/src/gnu/dist/groff/doc/Makefile.sub =>  -1
>    (Input/output error)
>
> On the backend I have two versions, one with size 0, the other with a
> decent size. I removed the one with size zero, ran ls -l on the client
> to trigger self heal: the log does not say it started, and I still get
> EIO accessing the file. Obviously that answer is cached somewhere.
>
> Is it possible to handle that situation without unmounting and
> remounting the volume, that is, without restarting glusterfs client?
>