On Tue, 4 Jan 2011, Ian Campbell wrote:
> On Tue, 2011-01-04 at 18:46 +0000, Konrad Rzeszutek Wilk wrote:
> > On Tue, Jan 04, 2011 at 05:24:38PM +0000, Ian Campbell wrote:
> > > On Thu, 2010-12-30 at 19:48 +0000, Konrad Rzeszutek Wilk wrote:
> > > > Only enabled if XEN_DEBUG_FS is enabled.
> > >
> > > Bit of an odd configuration option to use. Perhaps co-opt
> > > CONFIG_PARAVIRT_DEBUG instead?
> > >
> > > Or maybe we need a new XEN_DEBUG option? or just make it a developer
> > > only EXTRA_CFLAGS +=-DDEBUG thing?
> > I think the 'XEN_DEBUG' works better.
> > >
> > > Is this only temporary until the need for _PAGE_IOMAP is removed anyway?
> > I was thinking to leave it as a way to weed out bugs, but I could as well
> > just leave it in my "debug" branch and not propose it upstream.
> > I am not sure how to remove the _PAGE_IOMAP fully. We need some _flag_ to
> > signal 'xen_pte_val' that the PTE should not be looked up in the M2P.
> > Otherwise, for identity mappings, the value is 0xfffff.. (or 0x55555..
> > sometimes)
> > and the PTE ends up being messed up. Instead there is a check to see if
> > _PAGE_IOMAP is part of the PTE, and if so, no M2P lookup is done.
> > I guess we could do the M2P irregardless and just see if it is 0xfffff...
> > and
> > if so just return the PTE without any changes.
> Perhaps this ties in with the m2p overlay which Stefano+Jeremy have been
> working on to deal with granted foreign pages? I/O pages are a bit like
> foreign memory (if you squint enough)...
In theory the m2p overlay could be used for this purpose but in practice
the current m2p overlay API needs a struct page, also it might end up
stressing the hashtable too much.
Besides I think Konrad's solution might be simpler: if the m2p returns
one of the two special values we just return mfn from pte_mfn_to_pfn.
Keir, could you confirm that the m2p entries of DOM_IO pages are always
0xffffff or 0x55555?
Xen-devel mailing list