Bug 1448914

+++ This bug was initially created as a clone of Bug #1426952 +++

Description of problem:
=======================

Have observed a scenario where extended attributes are not syncing to slave via rsync/fuse if the entry and setxattr is done within 15 seconds. 

Master:
=======

[root@dj master]# touch r.1 ; setfattr  -n user.rtest1 -v 100112 r.1 
[root@dj master]# getfattr -d -e hex -m . r.1
# file: r.1
security.selinux=0x73797374656d5f753a6f626a6563745f723a6675736566735f743a733000
user.rtest1=0x313030313132

[root@dj master]# touch r.2 ; sleep 20 ; setfattr  -n user.rtest1 -v 100113 r.2
[root@dj master]# touch r.3 ; sleep 20 ; setfattr  -n user.rtest1 -v 100113 r.3
[root@dj master]# touch r.4 ; setfattr  -n user.rtest1 -v 100112 r.4
[root@dj master]# 


Slave:
======

[root@dj slave]# getfattr -d -e hex -m . r.1
# file: r.1
security.selinux=0x73797374656d5f753a6f626a6563745f723a6675736566735f743a733000

#root@dj slave]# 
[root@dj slave]# getfattr -d -e hex -m . r.2
# file: r.2
security.selinux=0x73797374656d5f753a6f626a6563745f723a6675736566735f743a733000

[root@dj slave]# 
[root@dj slave]# 
[root@dj slave]# getfattr -d -e hex -m . r.2
# file: r.2
security.selinux=0x73797374656d5f753a6f626a6563745f723a6675736566735f743a733000
user.rtest1=0x313030313133

[root@dj slave]# getfattr -d -e hex -m . r.1
# file: r.1
security.selinux=0x73797374656d5f753a6f626a6563745f723a6675736566735f743a733000

[root@dj slave]# getfattr -d -e hex -m . r.3
# file: r.3
security.selinux=0x73797374656d5f753a6f626a6563745f723a6675736566735f743a733000
user.rtest1=0x313030313133

[root@dj slave]# getfattr -d -e hex -m . r.4
# file: r.4
security.selinux=0x73797374656d5f753a6f626a6563745f723a6675736566735f743a733000

[root@dj slave]# 
[root@dj slave]# getfattr -d -e hex -m . r.4
# file: r.4
security.selinux=0x73797374656d5f753a6f626a6563745f723a6675736566735f743a733000

[root@dj slave]# 

Looked into the .processed directory, it doesn't contain any information yet. Looked into the changelogs and .processing directory. When extended attribute are applied after 15 sec, it is rightly logged into the new changelog which is in processing directory with (SETXATTR). If it is set with 15 secs, it is not recorded in the changelog but only Entry and SETATTR is only recorded and no record for SETXATTR.

Version-Release number of selected component (if applicable):
=============================================================

glusterfs-geo-replication-3.8.4-15.el7rhgs.x86_64


How reproducible:
=================

Always

Steps to Reproduce:
===================
1. Create geo-rep session between master and slave volume
2. Mount the volumes 
3. Create a file and wait for 20 secs
4. Set the extended attribute on the file. It will properly sync to slave.
5. Create another file and immediately set the extended attribute on the file

Actual results:
===============

extended attributes are not synced to the slave


Expected results:
=================

extended attributes should be synced to the slave



--- Additional comment from Rahul Hinduja on 2017-02-26 13:38:35 EST ---



I dont have data to prove if this is a regression. But looking at the design this doesn't seem to be a regression, would definitely need Arvinda/kotresh comments on this. 



--- Additional comment from Rahul Hinduja on 2017-02-26 13:40:10 EST ---

> I dont have data to prove if this is a regression. But looking at the design
> this doesn't seem to be a regression, would definitely need Arvinda/kotresh
> comments on this. 
+ Also, the testing of extended attributes were done in earlier releases but not within 15 sec. So do not have testing data.

--- Additional comment from Kotresh HR on 2017-02-27 04:01:08 EST ---

Analysis:

This is not a regression. This is a day one bug. There is an optimization in changelog to not record multiple METADATA ops on single gfid with in a changelog.
I think this was because there was no extended attribute support during that time and only supported SETATTR. Now since we also support SETXATTR, if SETATTR is already recorded for gfid1, we would miss recording SETXATTR on gfid1. Hence
setxattr would not get synced. This bug will affect the user set extended attributes if any.

The solution would be to remove the optimization for METADATA fops and keep it only for DATA fops.