Hi, Keir
Please see my comments embedded.
From: Keir Fraser
[mailto:Keir.Fraser@xxxxxxxxxxxx]
Sent: 2007年12月11日 18:38
To: Shan, Haitao
Cc: xen-devel@xxxxxxxxxxxxxxxxxxx;
Jiang, Yunhong
Subject: Re: [PATCH] Enable Core 2
Duo Performance Counters in HVM guest
* Don’t
define new MSR macros – use the existing ones (or fixup the eixsting ones to
have better names if you really think that is necessary, although I doubt it is
in this case).
I will
update.
* Some of the indentation is rather wacky.
I will
update.
* Poking the LVTPC as a one-off event only when the VCPU writes its
virtual LVTPC is not going to fly. What if the VCPU is subsequently migrated to
a different physical CPU? In any case the LVTPC is a shared resource that needs
synchronised access via e.g., context switching or mutual exclusion (see the
next point below).
I think
in the patch I will save/load LVTPC during context switch.
* How does this interact with, for example, xenoprof? Ideally we should
be able to run xenoprof and this concurrently over disjoint subsets of domains.
At the very least xenoprof and this patch should avoid stomping on each other
by implementing some form of mutual exclusion. Probably some resource sharing
(e.g., the PMC interrupt vector) also is possible. I expect this will require
some design thought, but unfortunately that is the price for being second into
the tree.
Yes.
How to share PMU resources will need careful design. But I really don’t think
it possible to run both concurrently unless we do some sort of PMC’s partition.
Of course, it is silly of me not to implement a mechanism for mutual exclusion.
I will implement one.
* Impact on save/restore to/from Core-2 processors: has this been tested
at all?
I will
try doing the test. I think whether it should support save/restore can be argued.
Does anyone want to run VTune/Oprofile during save/restore? This can be hardly
a good usage model. But at least, I update it to ensure it does not break current
save/restore.
-- Keir
On 11/12/07 10:06, "Shan, Haitao" <haitao.shan@xxxxxxxxx>
wrote:
Hi, Keir,
Currently, HVM guests do not have access to performance counters. So it
is not possible to use performance analyzer software such as vtune in HVM guest
to analyze programme performance. Other usage of performance counters , for
example, the NMI watchdog, won't function either.
This patch will enable performance counters in HVM guest. Currently,
only Core 2 Duo is implemented. Tests are carried out using Intel Vtune
Performance Analyzer in Windows XP Professional with 2 vcpus.
There are some notes at the beginning of the patch, describing
implementation choices in this patch.
Signed-off-by: Haitao Shan
<haitao.shan@xxxxxxxxx>
<<pmu.patch>>
Best Regards
Haitao Shan
