On Wed, Apr 18, 2001 at 03:47:09PM +0200, Garry Glendown wrote:
> Jan Harkes wrote:
> > On Wed, Apr 18, 2001 at 03:05:39PM +0200, Garry Glendown wrote:
> > > dig:/coda # cp /tmp/htdig.tgz .
> > > cp: writing `./htdig.tgz': Invalid argument
> > > dig:/coda # ls -l /tmp/htdig.tgz htdig.tgz
> > > -rw-r--r--    1 root     root      4298648 Apr  2 13:11 /tmp/htdig.tgz
> > > -rw-r--r--    1 codadmin nogroup   1998848 Apr 18 14:51 htdig.tgz
> > > (exiting root)
> > 
> > What filesystem is used for /usr/coda/venus.cache? The Coda kernel
> > module redirects all read/write operations directly to the underlying
> > containerfile. These errors are possibly not generated by Coda, but by
> > the filesystem.
> 
> Outsch ... for recovery reasons, the system is running on reiserfs for
> system and data drive. Could this be the reason?

Yes, that is definitely the reason.

ReiserFS journals the metadata when the filehandle is released. Coda's
cachemanager opens the containerfile in reiserfs, grabs the device/inode
numbers and passes them down to the Coda kernel module. Then the
cachemanager has no more use for the filehandle and closes it. Reiser
then logs the size of the containerfile, and the subsequent read/write
operations start failing (or not, depending on the version of reiserfs).

I've got a patch for linux-2.4.2 where the Coda kernel module keeps a
reference to the filehandle until the file in /coda is released. This
allows us to use any filesystem that has special semantics on file_open
and file_release but still uses generic_read/write_file. This should
cover most UNIX like filesystems including ramfs and shmfs.

The patch now needs to be cleaned up and forward ported to the latest
2.4.4-pre kernel before I can send it off to Linus. A 2.2.x version
should be pretty straightforward, although the existing patch won't work
as the Coda kernel module in 2.2 and 2.4 are internally very different.

Jan