Sorry - could not find the patch in the thread on LKML -
so could you give some pointers to your corrected nfs client
patch.
Cheers!
On Fri, Jun 01, 2001 at 02:24:02PM +0200, Gergely Tamas wrote:
> Hi!
>
> > > Could you explain a bit more about problems with 2.4.5 nfs clients ?
> > > I'm was about to upgrade quite a few boxes here. (right now 1 is done so
> > > i'll wait a bit and keep an eye on it)
> >
> > Probably Gergely is better person to ask about details.
>
> Ok, here I am. :)
>
> > First he experienced crashes of 2.4.5 nfs client and posted information
> > to the lkml (?) ("BUG at inode.c:486"). Viro and Trond Myklebust made a
> > patch that cured this particular problem.
>
> Here is a short description of the problem by Thrond Myklebust...
>
> " My guess is that is a result of the 'magic nfs path' in iput. It
> appears to be calling clear_inode() without truncating the pages
> first.
>
> The reason we haven't seen this before is that we had 'force_delete'
> that would always set i_nlink = 0. Unfortunately force_delete is toxic
> to mmap(), as it will discard any dirty pages rather than flushing
> them to storage, so it was removed in the 2.4.5-pre series... "
>
> ...
>
> Not normal, but it's a consequence of a bug in nfsd. It's reusing
> inode numbers but not changing the generation count.
>
> In cases where we're the ones doing the deletion, we can fix things
> (see appended patch), but if somebody on the server deletes the file,
> this problem will occur no matter what I do on the client.
>
> ...
>
> It worked in 2.4.4 because we threw away inodes/files whenever we
> finished accessing them. This is what hid the bug.
>
> Now we allow caching of inodes beyond the lifetime of open()/close()
> and so we
>
> a) Improve caching of files that don't change
> b) Don't throw away data that has been changed by mmap() but not
> declared to the NFS system for writing. "
>
> > But then there were crashes during Connectathon benchmark that
> > ceased after switching to 2.4.1 client.
>
> This was only an caused by an accident Thrond's patch. :) We removed this
> part, and I will starting again ...
>
> Gergely
--
Chan Shih-Ping (Richard) <[EMAIL PROTECTED]>
DSO National Laboratories
20 Science Park Drive
Singapore 118230
