Xen devel - Feb 2008 - [PATCH 3/4] Enable VirtualPC 2007 run on top of XEN: REPS prefix emulation

REPS prefix emulation for INS and OUTS is already in xen. Can we enable
REPS for "stos" and "movs" in paging mode? Although it
doesn''t benefit a
lot for normal hvm guest, in virtualpc case, I have to do that.
VirtualPC will call "reps stos"(ecx=0x400, eax=0) a lot, about 1000
times per second. It seems to clear a L1 page table page, it can''t be
fast unshadowed in check_for_early_unshadow and it''s even ReadOnly in
guest page table(CR0.wp = 0). So we get 0x400 * 1000 times page faults
per second, so CPU cycle is almost wasted on it.


Reviewed-by: Kevin Tian <kevin.tian@intel.com>
Signed-off-by: Disheng Su <disheng.su.com>


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

Hi, 

At 14:24 +0800 on 01 Feb (1201875864), Su, Disheng
wrote:
> REPS prefix emulation for INS and OUTS is already in xen. Can we enable
> REPS for "stos" and "movs" in paging mode? Although it
doesn''t benefit a
> lot for normal hvm guest, in virtualpc case, I have to do that.
> VirtualPC will call "reps stos"(ecx=0x400, eax=0) a lot, about
1000
> times per second. It seems to clear a L1 page table page, it can''t
be
> fast unshadowed in check_for_early_unshadow and it''s even ReadOnly
in
> guest page table(CR0.wp = 0). So we get 0x400 * 1000 times page faults
> per second, so CPU cycle is almost wasted on it.
This patch looks fine to me; I''ll let Keir decide whether the extra 200
lines or so of shadow/emulate code is worthwhile for running VPC under Xen. 

Keir, even if we don''t take this, please do take this hunk:

diff -r 05b3bdb3b7fa xen/arch/x86/mm/shadow/multi.c
--- a/xen/arch/x86/mm/shadow/multi.c	Wed Jan 30 00:09:03 2008 +0800
+++ b/xen/arch/x86/mm/shadow/multi.c	Wed Jan 30 00:09:04 2008 +0800
@@ -4076,8 +4091,8 @@ static void *emulate_map_dest(struct vcp
         sh_ctxt->mfn2 = emulate_gva_to_mfn(v, (vaddr + bytes - 1) &
PAGE_MASK,
                                            sh_ctxt);
         if ( !mfn_valid(sh_ctxt->mfn2) ) 
-            return ((mfn_x(sh_ctxt->mfn1) == BAD_GVA_TO_GFN) ?
+            return ((mfn_x(sh_ctxt->mfn2) == BAD_GVA_TO_GFN) ?
                     MAPPING_EXCEPTION : MAPPING_UNHANDLEABLE);
 
         /* Cross-page writes mean probably not a pagetable */
         sh_remove_shadows(v, sh_ctxt->mfn2, 0, 0 /* Slow, can fail */ );

Cheers,

Tim.

-- 
Tim Deegan <Tim.Deegan@citrix.com>
Principal Software Engineer, Citrix Systems (R&D) Ltd.
[Company #02300071, SL9 0DZ, UK.]

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

>>> On 2/1/2008 at 2:36 AM, Tim Deegan <Tim.Deegan@citrix.com>
wrote:
> Hi, 
> 
> At 14:24 +0800 on 01 Feb (1201875864), Su, Disheng wrote:
>> REPS prefix emulation for INS and OUTS is already in xen. Can we enable
>> REPS for "stos" and "movs" in paging mode? Although
it doesn''t benefit a
>> lot for normal hvm guest, in virtualpc case, I have to do that.
>> VirtualPC will call "reps stos"(ecx=0x400, eax=0) a lot,
about 1000
>> times per second. It seems to clear a L1 page table page, it
can''t be
>> fast unshadowed in check_for_early_unshadow and it''s even
ReadOnly in
>> guest page table(CR0.wp = 0). So we get 0x400 * 1000 times page faults
>> per second, so CPU cycle is almost wasted on it.
> 
> This patch looks fine to me; I''ll let Keir decide whether the
extra 200
> lines or so of shadow/emulate code is worthwhile for running VPC under Xen.
> 
> Keir, even if we don''t take this, please do take this hunk:
> 
> diff -r 05b3bdb3b7fa xen/arch/x86/mm/shadow/multi.c
> --- a/xen/arch/x86/mm/shadow/multi.c	Wed Jan 30 00:09:03 2008 +0800
> +++ b/xen/arch/x86/mm/shadow/multi.c	Wed Jan 30 00:09:04 2008 +0800
> @@ -4076,8 +4091,8 @@ static void *emulate_map_dest(struct vcp
>          sh_ctxt->mfn2 = emulate_gva_to_mfn(v, (vaddr + bytes - 1) &
PAGE_MASK,
>                                             sh_ctxt);
>          if ( !mfn_valid(sh_ctxt->mfn2) ) 
> -            return ((mfn_x(sh_ctxt->mfn1) == BAD_GVA_TO_GFN) ?
> +            return ((mfn_x(sh_ctxt->mfn2) == BAD_GVA_TO_GFN) ?
>                      MAPPING_EXCEPTION : MAPPING_UNHANDLEABLE);
>  
>          /* Cross-page writes mean probably not a pagetable */
>          sh_remove_shadows(v, sh_ctxt->mfn2, 0, 0 /* Slow, can fail */
);
> 
> Cheers,
> 
> Tim.

If I could chime in here, I''m all for better support of CR0.WP = 0
guests.  NetWare has traditionally run in this mode, and although we''ve
recently made it capable of running with WP bit set as part of the
paravirtualization effort, there would be benefit to running older versions as
HVM guests.  I haven''t had the cycles myself to work on that, but with
all the other changes to real mode support, it looks like we''re really
close to having NetWare run fully virtualized. (Another gottcha is that we write
the debug registers in real mode, but that should be a fairly simple hypervisor
fix.)

Thanks to everyone who has worked on getting HVM support more robust.

- Bruce


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