Re: still nfs problems [Was: Linux 2.6.37-rc8]

[James Bottomley <James.Bottomley@HansenPartnership.com>]

On Fri, 2011-01-07 at 19:02 +0000, Russell King - ARM Linux wrote:
> On Fri, Jan 07, 2011 at 01:53:25PM -0500, Trond Myklebust wrote:
> > I'd still like to keep the existing code for those architectures that
> > don't have problems, since that allows us to send 32k READDIR requests
> > instead of being limited to 4k. For large directories, that is a clear
> > win.
> > For the NOMMU case we will just go back to using a single page for
> > storage (and 4k READDIR requests only). Should I just do the same for
> > architectures like ARM and PARISC?
> 
> I think you said that readdir reads via the vmalloc mapping of the
> group of pages, but XDR writes to the individual pages.

Actually it's the other way around, but the point still stands.

> As I understand NFS, you receive a packet, you then have to use XDR
> to unpack the data, which you presumably write into the set of
> struct page *'s using kmap?
> 
> Isn't a solution to have XDR write directly into the vmalloc mapping
> rather than using struct page * and kmap?

So, unfortuantely, I looked at doing this and we can't.  the ->readdir()
call takes an array of pages, not a kernel virtual address of the pages,
so there's no way to tell it to use a different mapping from the usual
kernel one on them.

On the other hand, the xdr routines, since they take the pages anyway,
could use a scatterlist approach to writing through the kernel mapping
instead of using vmap ... we have all the machinery for this in
lib/scatterlist.c ... it's not designed for this case, since it's
designed to allow arbitrary linear reads and writes on a block
scatterlist, but the principle is the same ... it looks like it would be
rather a big patch, though ... 

James