Re: [Xen-devel] [PATCH] x86 fixes for 3.3 impacting distros (v1).

["H. Peter Anvin" <hpa@zytor.com>]

On 06/28/2012 07:28 AM, Konrad Rzeszutek Wilk wrote:
> 
> Peter mentioned to me had some ideas about software PAT table lookup. I am not
> exactly sure what he meant by that.
> 

I could see the kernel have programmable PAT values rather than fixed if
and only if it can be showed to have no measurable performance impact.

> Just to summarize, there were two ways proposed to fix this:
> 
>  1). Make __page_change_attr_set_clr use a new wrapper: pte_attr, that calls
>      pte_val (pvops call) instead of pte_flag (native). Here is the patch:
>      http://git.kernel.org/?p=linux/kernel/git/konrad/xen.git;a=commitdiff;h=4f93aa02acd0e34806d4ac9c3a700bb5d040eab6
>      (no perf regressions across all platforms)
> 
>  2). Introduce a new pvops call - pte_flags, which would make pte_flags
>      (which currently is doing just a bit mask) be pvops-fied.
>      http://darnok.org/results/baseline_pte_flags_pte_attrs/0001-x86-paravirt-xen-Introduce-pte_flags.patch
>      http://darnok.org/results/baseline_pte_flags_pte_attrs/0002-x86-paravirt-xen-Optimize-pte_flags-by-marking-it-as.patch
>      (weird results on AMD, other platforms had no perf degradations)
> 
>   3). (not posted), was to do 2), but alter the alternative_asm and instead use asm_goto to
>      make the compiler use less registers and hopefully reduce the code:
>      http://git.kernel.org/?p=linux/kernel/git/konrad/xen.git;a=shortlog;h=refs/heads/devel/mmu-perf
>      But the results I got showed worst performance on baremetal.. which was weird?
>      Perhaps it is compiler related - never got to follow up on it.
> 

OK, let me be blunt: I will unconditionally veto any of these.

> 
> I also chatted with the core Xen hypervisor folks about adding in the context switch code
> to alter the PAT layout - but they were not keen a about it - and I am not sure how much
> CPU cycles one loses by doing a wrmsr to the PAT register on every guest context switch
> (worst case when on has a pvops kernel and a old-style one - where the WC bit would differ)?
> 

And you're comparing that to a bunch of new pvops calls?  The discussion
shouldn't even have started until you had ruled out this solution and
had data to show it.

	-hpa