Re: [PATCH] x86/hvm: don't treat MMIO pages as special ones regarding cache attributes

[Roger Pau Monné <roger.pau@xxxxxxxxxx>]

On Thu, Sep 10, 2020 at 11:27:49AM +0200, Jan Beulich wrote:
> On 09.09.2020 16:50, Roger Pau Monne wrote:
> > MMIO regions below the maximum address on the memory map can have a
> > backing page struct that's shared with dom_io (see x86
> > arch_init_memory and it's usage of share_xen_page_with_guest), and
> > thus also fulfill the is_special_page check because the page has the
> > Xen heap bit set.
> > 
> > This is incorrect for MMIO regions when is_special_page is used by
> > epte_get_entry_emt, as it will force direct MMIO regions mapped into
> > the guest p2m to have the cache attributes set to write-back.
> > 
> > Add an extra check in epte_get_entry_emt in order to detect pages
> > shared with dom_io (ie: MMIO regions) and don't force them to
> > write-back cache type on that case.
> 
> Did you consider the alternative of not marking those pages as Xen
> heap ones? In particular when looking at it from this angle I
> consider it at least odd for non-RAM (or more precisely non-heap)
> pages to get marked this way.

I wasn't sure whether this could cause issues in other places of the
code that would rely on this fact and such change seemed more risky
IMO.

> And I can't currently see anything
> requiring them to be marked as such - them being owned by DomIO is
> all that's needed as it seems.

Should those pages then simply be assigned to dom_io and set the
appropriate flags (PGC_allocated | 1), or should
share_xen_page_with_guest be modified to not set the PGC_xen_heap
flag?

I see that such addition was done in a2b4b8c2041, but I'm afraid I
don't fully understand why share_xen_page_with_guest needs to mark
pages as Xen heap.

Thanks, Roger.