Re: [Xen-ia64-devel] [PATCH] tlb miss handler

[Tristan Gingold <Tristan.Gingold@xxxxxxxx>]

Le Vendredi 27 Janvier 2006 02:56, Isaku Yamahata a écrit :
> Hi. Thanks for your comment.
>
> On Thu, Jan 26, 2006 at 03:31:25PM +0100, Tristan Gingold wrote:
> > Le Jeudi 26 Janvier 2006 13:59, Isaku Yamahata a écrit :
> > > This patch is not for commit.  But for review, commnets.
> > >
> > > Current itlb/dtlb miss handlers don't handle a tlb miss
> > > in xen/ia64 identity mapping area.
> > > Such a miss might occur because xen/ia64 enables vhpt walker
> > > even for region 7. (Please see set_one_rr())
> > > I haven't encountered it yet, I'm guessing so from reading code.
> > >
> > > Vti itlb/dtlb miss handlers do it by checking psr.vm bits.
> > > This patch is not so optimized for review,
> > > but it should be done later.
> >
> > Quiet interesting because I have just worked a little bit in this area.
> >
> > My comments:
> > itlb miss must not occur, since Xen is mapped with a itm.
>
> I agree with you about itlb miss.
> Linux alt_itlb_miss handler handles it because Linux supports
> loadable kernel module.
> On the other hand xen/ia64 doesn't support it (currently),
> so itlb miss in the identity mapping area must not occur.
Ok.

> > dtlb miss should not occur for the same reason, but they occur at least
> > during start-up while Xen reads bootparams and EFI.  This is done before
> > enabling VHPT.
>
> No. dtlb miss in the identity mapping area might occur.
> xen/ia64 also uses the identity mapping to access domain pages.
> Please see map_domain_page() definition in xen/include/xen/domain_page.h
> (note: CONFIG_DOMAIN_PAGE is not defined for xen/ia64)
> map_domain_page() is certainly used.
> E.g. page_scrub_softirq() in xen/commong/page_alloc.c.
> It is invoked when a domain is destructed.
> Fortunately(unfortunately?) xen/ia64 doesn't support domain
> destruction now.
> I guess it's the reason why we haven't met such a dtlb miss
> except start-up.
Thanks.  This simply proves I didn't look far enough!

> > BTW, I think the size of the identity mapped region should be shrinked in
> > order to enabled virtual memory in Xen.  This will be necessary for
> > virtual mem map.  That's the point I have looked.
>
> Why don't we use region 5?
> Linux uses region 5 for virtual mem map
> and xen/ia64 don't use the region 5 for now.
>
> If we restricted the identity mapping region, we would have to
> define CONFIG_DOMAIN_PAGE and implement needed functions.
OK.

Tristan.



_______________________________________________
Xen-ia64-devel mailing list
Xen-ia64-devel@xxxxxxxxxxxxxxxxxxx
http://lists.xensource.com/xen-ia64-devel