Re: knfsd and system crashes

Bill Hawes (whawes@star.net)
Wed, 12 Nov 1997 18:32:30 -0500


Miguel de Icaza wrote:
> According to the last description of the fixes you are working on
> the knfsd code, I assume that it is not actually possible to have
> client-transparent crashes of the server.
>
> Someone had suggested that an inode->dentry mapping should be
> possible to do, what is the reason to not work on that direction?
> This would allow Linux servers to crash (or be reseted) without
> affecting the client machines.

Actually the approach I'm taking now should allow the same degree of
recoverability supported by other servers. Because it's possible that
the dentry pointer in the filehandle will have become invalid (or
represent a different object), I'm also including the inode number and
parent inode number in the fh.

The intent is to do a fast lookup based on the dentry most of the time,
but to be able to fall back to the unfsd-style volume search if
necessary. So to the extent that inode numbers are a reasonably
permanent representation of a file, following a server crash the old
client filehandles could still be looked up, but more slowly.

But of course Linux-based servers aren't going to crash :-)

Regards,
Bill