rsync-2.6.2: NFS clients confused after an rsync

[Brian Childs]

We use rsync to update an nfs server.  After an update, we noticed that
a large number of clients didn't see the updated data.

It took me a while to be able to reliably reproduce this problem, but it
happens on old and new versions of rysnc.  It also happens across all
the platforms we use here (sun/linux/netapp).

This shows the problem: [Note my home directory is NFS mounted]

***************************************************
rm -f /tmp/testpath/* $HOME/testpath/*

#Prime it
echo some data > /tmp/testpath/testfile
sleep 1
rsync -a /tmp/testpath/ $HOME/testpath/
ssh $otherbox ls -li $HOME/testpath

#Break it
echo additional data >> /tmp/testpath/testfile
sleep 1
rsync -a -b /tmp/testpath/ $HOME/testpath/
ssh $otherbox ls -li $HOME/testpath

#Fix it
touch $HOME/testpath
ssh $otherbox ls -li $HOME/testpath
***************************************************

Here's the output:
total 0
    483079 -rw-r--r--   1 brian    source        10 May  6 18:05 testfile
total 0
    483079 -rw-r--r--   1 brian    source        10 May  6 18:05 testfile
total 0
    483080 -rw-r--r--   1 brian    source        26 May  6 18:05 testfile
    483079 -rw-r--r--   1 brian    source        10 May  6 18:05 testfile~

The output from the "break it" section matches "prime it", but it should
match the "fix it" output.

What I think is happening:
--------------------------

* testfile is updated in place, so the mtime of testpath
  isn't udpated.
* The nfs client is caching the mtime of "testpath"
* The seconds rsync modifies the contents of testpath, but
  when finished, it resets the mtime on "testpath"
* The client believes it's cache is valid, so I doesn't refresh, and
  therefore it misses the update

NFS Clients tested: sun (solaris 8 & 9), linux (2.4.18 & 2.4.24)
NFS Servers tested: sun, linux, and netapp.

They all behaved the same.

Proposed fix:
-------------
Change rsync so it doesn't reset mtimes on directories if --backup is
in effect.

Thanks,
Brian