Jeremy Fitzhardinge wrote:
> Those are bugs that can occur, but they don't apply in this case. The
> vmi implementation of kmap_atomic_pte() would be:
>
> static void *vmi_kmap_atomic_pte(struct page *page, enum km_type type)
> {
> void *ptep = kmap_atomic(page, type);
> vmi_map_pt_hook(type, ptep, page_to_pfn(page));
> return ptep;
> }
>
> Right? Which is functionally identical to the code in your patch,
> except wrapped up in a new function.
>
Yes, but the hook point now happens before the page table mapping. Not
that it should cause a problem. But we've been testing things the
original way for over a year now, and if we want to get the fix upstream
for 2.6.21, it seems better to upstream a more tested fix rather than a
new way of doing things, even if it is identical in theory.
That said, I have no problems with the approach you propose going
forward. I just don't think it is appropriate for an -rc release,
because it provides no tangible benefit for any of the in-kernel code,
and causes a lot of retesting. I still believe there is almost zero
risk to doing things the way you propose. But I am also a firm believer
in shipping what is tested and working unless there is a compelling
reason to do otherwise. And if Xen is not going to be in 2.6.21, the
compelling reason becomes getting the code working for both of us for
2.6.22 - so let's do that, and keep the patches from Andrew's -mm tree
around to make sure that we have a suitable patch base that can be
applied to 2.6.21 for any distros that are willing to pick up the Xen
paravirt-ops.
Sound reasonable?
Zach
