Bug 1330567

+++ This bug was initially created as a clone of Bug #1322247 +++

Description of problem:
On a windows client when a tiered volume is mounted and microsoft office files are created with some data in it, the size of the file is still remains 0KB.

For a distribute volume it reflects the actual size for the same files.
Even on "cifs" & "fuse" the actual size is reflecting.

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

mainline

How reproducible:
Always

Steps to Reproduce:

1.Mount a tier volume (cold tier : Distributed-Disperse | 2 x (8 + 4) | Hot Tier Type : Distributed-Replicate 4 x 2 = 8) on windows client

2. Create docx files with data in it.
3.do ls in smbclient

Actual results:

File size is 0KB

Expected results:

Actual size should reflect

Additional info:

--- Additional comment from Vivek Das on 2016-03-30 03:00:20 EDT ---


--- Additional comment from Mohammed Rafi KC on 2016-04-15 10:27:18 EDT ---

Myself together with Anoop CS reproduced the issue using both windows machine and smbclient

set-up details : gluster - upstream master
                samba - 4.4

The problem turned out to be the same problem with T file issue in tiering which is described in #bug 1303298 .

RCA:

For tiered volume , we send readdirp only to cold tier for performance improvement and other functional issues. Since cold tier is the default hashed subvol, every file or corresponding linkfile should be there in cold tier. For files in hot tier, there will be a linkfile and readdirp from cold tier for such files will not have proper attributes. In normal client stack, tier xlators will set inode=null for such entries to force a lookup in case of fuse mount and for nfs , we will set a flag to let nfs client know about such an entry as a stale.

In case of samba client, They will just print what they got, means the stat will have only sticky bit set.

More details:

We brought this fix to increase the readdir performance and to fix an issue which can potentially result in data loss (#bug 1278384). So reverting the changes will result in  performance drop in readdir and we need additional patch to solve the issue in #bug 1278384.

Otherwise we can fix this problem from gf_api layer or from tier layer by doing a stat to hot tier for entries in hot tier, which will also give a worst performance if there are two or more entries in hot tier. We can reduce the performance drop by introducing bulk lookup (one lookup for all entries since files can be either in hot or cold).

Or otherwise if there is any way to let samba client about a stale entry in a readdirp, we could try that also, so that samba can do a stat before printing it.