Description of problem:

On a distributed-disperse volume, if you try to rename a symbolic link, the fuse process dies.

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


How reproducible:

Often, depends on the names used as original and renamed names

Steps to Reproduce:
1. gluster volume create test disperse 6 redundancy 2 host:/bricks/test_{1..12}
2. gluster volume start test
3. mount -t glusterfs host:/test /mnt
4. ln -s /dummy /mnt/symlink
5. mv /mnt/symlink /mnt/symlink.2

Actual results:

Fuse mount crashes

Expected results:

The rename should not cause a crash

Additional info:

The problem appears at the creation of the linkto file when the new name doesn't go to the same brick the symlink was residing before the rename. In this case DHT creates the symlink and tries to set some attributes calling setattr fop.

When setattr is called, the passed inode references the "real" file (i.e. a symlink). However, when ec receives the answers from bricks, statpre and statpost say that the inode is a regular file (because linkto files are regular files on the bricks).

This makes ec assume that it's a file, so it must have a defined size, but symlinks do not have size, which causes an assertion to fail.

We shouldn't use the same inode structure to represent two different things.