[Gluster-users] Setting gfid failed on slave geo-rep node

[Aravinda]

Hi ML,

We analyzed the issue. Looks like Changelog is replayed may be because 
of Geo-rep worker crash or Active/Passive switch or both Geo-rep workers 
becoming active.

 From changelogs,

CREATE  logo-login-04.svg.part
RENAME logo-login-04.svg.part logo-login-04.svg

When it is replayed,
CREATE  logo-login-04.svg.part
RENAME logo-login-04.svg.part logo-login-04.svg
CREATE  logo-login-04.svg.part
RENAME logo-login-04.svg.part logo-login-04.svg

During replay backend GFID link is broken and Geo-rep failed to cleanup. 
Milind is working on the patch to fix the same. Patches are in review 
and expected to be available in 3.7.8 release.

http://review.gluster.org/#/c/13316/
http://review.gluster.org/#/c/13189/

Following script can be used to find problematic file in each Brick backend.
https://gist.github.com/aravindavk/29f673f13c2f8963447e

regards
Aravinda

On 02/01/2016 08:45 PM, ML mail wrote:
> Sure, I will just send it to you through an encrypted cloud storage app and send you the password via private mail.
>
> Regards
> ML
>
>
>
> On Monday, February 1, 2016 3:14 PM, Saravanakumar Arumugam <sarumuga at redhat.com> wrote:
>
>
> On 02/01/2016 07:22 PM, ML mail wrote:
>> I just found out I needed to run the getfattr on a mount and not on the glusterfs server directly. So here are the additional output you asked for:
>>
>>
>> # getfattr -n glusterfs.gfid.string  -m .  logo-login-09.svg
>> # file: logo-login-09.svg
>> glusterfs.gfid.string="1c648409-e98b-4544-a7fa-c2aef87f92ad"
>>
>> # grep 1c648409-e98b-4544-a7fa-c2aef87f92ad /data/myvolume/brick/.glusterfs/changelogs -rn
>> Binary file /data/myvolume/brick/.glusterfs/changelogs/CHANGELOG.1454278219 matches
> Great!  Can you share the CHANGELOG ?  ( It contains various fops
> carried out on this gfid)
>
>> Regards
>> ML
>>
>>
>>
>> On Monday, February 1, 2016 1:30 PM, Saravanakumar Arumugam <sarumuga at redhat.com> wrote:
>> Hi,
>>
>> On 02/01/2016 02:14 PM, ML mail wrote:
>>> Hello,
>>>
>>> I just set up distributed geo-replication to a slave on my 2 nodes' replicated volume and noticed quite a few error messages (around 70 of them) in the slave's brick log file:
>>>
>>> The exact log file is: /var/log/glusterfs/bricks/data-myvolume-geo-brick.log
>>>
>>> [2016-01-31 22:19:29.524370] E [MSGID: 113020] [posix.c:1221:posix_mknod] 0-myvolume-geo-posix: setting gfid on /data/myvolume-geo/brick/data/username/files/shared/logo-login-09.svg.ocTransferId1789604916.part failed
>>> [2016-01-31 22:19:29.535478] W [MSGID: 113026] [posix.c:1338:posix_mkdir] 0-myvolume-geo-posix: mkdir (/data/username/files_encryption/keys/files/shared/logo-login-09.svg.ocTransferId1789604916.part): gfid (15bbcec6-a332-4c21-81e4-c52472b1e13d) isalready associated with directory (/data/myvolume-geo/brick/.glusterfs/49/5d/495d6868-4844-4632-8ff9-ad9646a878fe/logo-login-09.svg). Hence,both directories will share same gfid and thiscan lead to inconsistencies.
>> Can you grep for this gfid(of the corresponding files) in changelogs and
>> share those files ?
>>
>> {
>> For example:
>>
>> 1. Get gfid of the files like this:
>>
>> # getfattr -n glusterfs.gfid.string  -m .  /mnt/slave/file456
>> getfattr: Removing leading '/' from absolute path names
>> # file: mnt/slave/file456
>> glusterfs.gfid.string="05b22446-de9e-42df-a63e-399c24d690c4"
>>
>> 2. grep for the corresponding gfid in brick back end like below:
>>
>> [root at gfvm3 changelogs]# grep 05b22446-de9e-42df-a63e-399c24d690c4
>> /opt/volume_test/tv_2/b1/.glusterfs/changelogs/ -rn
>> Binary file
>> /opt/volume_test/tv_2/b1/.glusterfs/changelogs/CHANGELOG.1454135265 matches
>> Binary file
>> /opt/volume_test/tv_2/b1/.glusterfs/changelogs/CHANGELOG.1454135476 matches
>>
>> }
>> This will help in understanding what operations are carried out in
>> master volume, which leads to this inconsistency.
>>
>> Also, get the following:
>> gluster version
>> gluster volume info
>> gluster volume geo-replication status
>>
>>> This doesn't look good at all because the file mentioned in the error message (
>>> logo-login-09.svg.ocTransferId1789604916.part) is left there with 0 kbytes and does not get deleted or cleaned up by glusterfs, leaving my geo-rep slave node in an inconsistent state which does not reflect the reality from the master nodes. The master nodes don't have that file anymore (which is correct). Here below is an "ls" of the concerned file with the correct file on top.
>>>
>>>
>>> -rw-r--r-- 2 www-data www-data   24312 Jan  6  2014 logo-login-09.svg
>>> -rw-r--r-- 1 root     root           0 Jan 31 23:19 logo-login-09.svg.ocTransferId1789604916.part
>> Rename issues in geo-replication are fixed lately. This looks similar to
>>
>> one.
>>
>> Thanks,
>> Saravana
>> _______________________________________________
>> Gluster-users mailing list
>> Gluster-users at gluster.org
>> http://www.gluster.org/mailman/listinfo/gluster-users
> _______________________________________________
> Gluster-users mailing list
> Gluster-users at gluster.org
> http://www.gluster.org/mailman/listinfo/gluster-users