still nfs problems [Was: Linux 2.6.37-rc8]

[Russell King - ARM Linux]

On Wed, Jan 05, 2011 at 01:12:25PM -0500, Trond Myklebust wrote:
> On Wed, 2011-01-05 at 17:26 +0000, Russell King - ARM Linux wrote: 
> > On Wed, Jan 05, 2011 at 12:17:27PM -0500, Trond Myklebust wrote:
> > > We should already be flushing the kernel direct mapping after writing by
> > > means of the calls to flush_dcache_page() in xdr_partial_copy_from_skb()
> > > and all the helpers in net/sunrpc/xdr.c.
> > 
> > Hmm, we're getting into the realms of what flush_dcache_page() is supposed
> > to do and what it's not supposed to do.
> > 
> > Is this page an associated with a mapping (iow, page_mapping(page) is non-
> > NULL)?  If not, flush_dcache_page() won't do anything, and from my
> > understanding, its flush_anon_page() which you want to be using there
> > instead.
> 
> Actually, none of these pages are ever mapped into userspace, nor are
> they mapped into the page cache.
> 
> They are allocated directly using alloc_page() by the thread that called
> the readdir() syscall, so afaics there should be no incoherent mappings
> other than the kernel direct mapping and the one created by
> vm_map_ram().
> 
> So, yes, you are right that we don't need the flush_dcache_page() here.

I do still think you need _something_ there, otherwise data can remain
in the direct map alias and not be visible via the vmap alias.  I don't
see that we have anything in place to handle this at present though.

jejb mentioned something about making kunmap_atomic() always flush the
cache, even for lowmem pages, but I think that's going to be exceedingly
painful, to the extent that I believe it will prevent our PIO-only MMC
drivers working - or we need a scatterlist API that will let drivers
iterate over the scatterlist without needing to continually kmap_atomic
and kunmap_atomic each page.