Re: Xen 4.3 development update

[George Dunlap <george.dunlap@eu.citrix.com>]

On 07/05/13 14:56, Pasi Kärkkäinen wrote:
> On Mon, Apr 29, 2013 at 03:21:34PM +0200, Peter Maloney wrote:
>> On 04/04/2013 07:05 PM, Tim Deegan wrote:
>>> At 16:23 +0100 on 04 Apr (1365092601), Tim Deegan wrote:
>>>> At 11:34 -0400 on 03 Apr (1364988853), Andres Lagar-Cavilla wrote:
>>>>> On Apr 3, 2013, at 6:53 AM, George Dunlap <george.dunlap@eu.citrix.com> wrote:
>>>> Yes, 4.2 is definitely slower.  A compile test on a 4-vcpu VM that takes
>>>> about 12 minutes before this locking change takes more than 20 minutes
>>>> on the current tip of xen-unstable (I gave up at 22 minutes and rebooted
>>>> to test something else).
>>> I did a bit of prodding at this, but messed up my measurements in a
>>> bunch of different ways over the afternoon. :(  I'm going to be away
>>> from my test boxes for a couple of weeks now, so all I can say is, if
>>> you're investigating this bug, beware that:
>>>
>>>   - the revision before this change still has the RTC bugs that were
>>>     fixed last week, so don't measure performance based on guest
>>>     wallclock time, or your 'before' perf will look too good.
>>>   - the current unstable tip has test code to exercise the new
>>>     map_domain_page(), which will badly affect all the many memory
>>>     accesses done in HVM emulation, so make sure you use debug=n builds
>>>     for measurement.
>>>
>>> Also, if there is still a bad slowdown, caused by the p2m lookups, this
>>> might help a little bit:
>>>
>>> diff --git a/xen/arch/x86/hvm/hvm.c b/xen/arch/x86/hvm/hvm.c
>>> index 38e87ce..7bd8646 100644
>>> --- a/xen/arch/x86/hvm/hvm.c
>>> +++ b/xen/arch/x86/hvm/hvm.c
>>> @@ -1361,6 +1361,18 @@ int hvm_hap_nested_page_fault(paddr_t gpa,
>>>           }
>>>       }
>>>   
>>> +
>>> +    /* For the benefit of 32-bit WinXP (& older Windows) on AMD CPUs,
>>> +     * a fast path for LAPIC accesses, skipping the p2m lookup. */
>>> +    if ( !nestedhvm_vcpu_in_guestmode(v)
>>> +         && gfn == vlapic_base_address(vcpu_vlapic(current)) >> PAGE_SHIFT )
>>> +    {
>>> +        if ( !handle_mmio() )
>>> +            hvm_inject_hw_exception(TRAP_gp_fault, 0);
>>> +        rc = 1;
>>> +        goto out;
>>> +    }
>>> +
>>>       p2m = p2m_get_hostp2m(v->domain);
>>>       mfn = get_gfn_type_access(p2m, gfn, &p2mt, &p2ma,
>>>                                 P2M_ALLOC | (access_w ? P2M_UNSHARE : 0), NULL);
>> This patch (applied to 4.2.2) has a very large improvement on my box
>> (AMD FX-8150) and WinXP 32 bit.
>>
>> It only took about 2.5 minutes to log in and see task manager. It takes
>> about 6 minutes without the patch. And 2.5 minutes is still terrible,
>> but obviously better.
>>
> So is the problem only on WinXP with "booting up / logging in to windows",
> or do you see performance regressions on some actual benchmark tools aswell
> (after windows has started up) ?

For the sake of people watching this thread: The last 4-5 mails I've 
sent to Peter Maloney have bounced with "Mailbox Full" messages; so it's 
possible he's not actually hearing this part of the discussion...

  -George