server/resolver: don't trust inode-table for RESOLVE_NOT

There have been known races between fops which add a dentry (like
lookup, create, mknod etc) and fops that remove a dentry (like rename,
unlink, rmdir etc) due to which stale dentries are left out in inode
table even though the dentry doesn't exist on backend. For eg.,
consider a lookup (parent/bname) and unlink (parent/bname) racing in
the following order:

* lookup hits storage/posix and finds that dentry exists
* unlink removes the dentry on storage/posix
* unlink reaches protocol/server where the dentry (parent/bname) is
  unlinked from the inode
* lookup reaches protocol/server and creates a dentry (parent/bname)
  on the inode

Now we've a stale dentry (parent/bname) associated with the inode in
itable. This situation is bad for fops like link, create etc which
invoke resolver with type RESOLVE_NOT. These fops fail with EEXIST
even though there is no such dentry on backend fs. This issue can be
solved in two ways:

* Enable "dentry fop serializer" xlator [1].
  # gluster volume set features.sdfs on

* Make sure resolver does a lookup on backend when it finds a dentry
  in itable and validates the state of itable.
   - If a dentry is not found, unlink those stale dentries from itable
     and continue with fop
   - If dentry is found, fail the fop with EEXIST

This patch implements second solution as sdfs is not enabled by
default in brick xlator stack. Once sdfs is enabled by default, this
patch can be reverted.

[1] https://github.com/gluster/glusterfs/issues/397

Change-Id: Ia8bb0cf97f97cb0e72639bce8aadb0f6d3f4a34a
updates: bz#1543279
BUG: 1543279
Signed-off-by: Raghavendra G <rgowdapp@redhat.com>

0 files changed, 0 insertions, 0 deletions