Re: [Xen-ia64-devel] RE: [PATCH]: disable handling of legacy privified insns

[Tristan Gingold <Tristan.Gingold@xxxxxxxx>]

Le Lundi 20 Mars 2006 17:08, Magenheimer, Dan (HP Labs Fort Collins) a écrit :
Hi,

[...]
> Is there a reason you are proposing to change the design
> of hyperprivops? 
No, I don't want to change the design, I am just trying to understand it 
fully.

> I realize that the few hyperprivops you
> just implemented are a bit ugly because it is necessary
> to save/restore virtual psr.ic and virtual psr.i, but
> these were never previously implemented as hyperprivops
> because they were relatively low frequency.  The virtual
> psr.ic trick works nicely for the high frequency ones.
Sure.

> Do you think the pl=2 approach will yield a significant
> performance advantage?  Or is there some other reason?
I don't see immediat performance advantage.

I was just thinking about privileged operations.
We all agree these are slow and have a bad impact on D cache.
Their handling also suppose that I=D, because the bundle is read using ld8.
We agree this is not a problem with para-virtualized linux; this is more or 
less a theorical issue.  But it makes vcpu_translate strange.

Now suppose we'd like to fix this issue.  We can either fetch the opcode using 
a manually translated mpa.  This would be slow.  Maybe we could make 
measures.  Or we can replace the pivop by a break on the fly.  To do this, we 
must be able to catch break insn and have enough room in the iim space.

This was just a thinking, I have written it down.

Tristan.



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