[OpenAFS] mmap issue?

[Simon Wilkinson]

On 23 Mar 2012, at 18:52, Eric Chris Garrison wrote:
>=20
> I've searched for this, and it appears mmap operations used to be a
> problem.  At what version of OpenAFS (server and client) should we be =
at
> for this to not be a problem?

There's no difference on the server between files that are written with =
mmap, and those that are written with normal write() operations. All of =
the complexity of mmap is handled in the client. Historically there were =
all sorts of exciting issues with mmap but these were all a) in Linux =
and b) fixed a while ago. So, it's possible that you are seeing a Mac OS =
X mmap problem (1.5.77 is pretty old, though, and 1.5 was never =
"stable")

Another possibility, as this appears to be some kind of database that he =
is running, is that the database is using byte range locks to =
synchronise between multiple processes or threads. These don't work in =
AFS at the best of times, and Mac OS X doesn't have the bodge to make =
them work between processes on the local machine - so this is a =
potential cause of data corruption. If byte range locks are in play, =
there should be a load of warning messages in dmesg.

Hope all that helps,

Simon.