Skip to contentSkip to conversation

Mapping of Shared Page in e500 Powerpc kvm

DISCUSSION Aashish Mittal 3 participants 3 messages Oct 23, 2011 – Oct 25, 2011

lore
Share & tools
lore
#0 lore raw

Hi

I'm working on powerpc e500 machine (Freescale P2020RDB machine) and using the paravirtualization guest support available inside kvm and wanted
to some bookkeeping of my own on the shared page and from what i
understood is that a single tlb entry is reserved for the shared page (magic page) when running the guest which is never preempted or flushed
but when i tried printing out the dtlb misses on this shared page it show
me a large number of dtlb misses on this shared page which means that this page is being preempted or flushed out in some way . Can someone pleas e clarify this situation if i'm understanding something wrong or is it a bug ?

Thanks
Aashish Mittal

#1 lore raw

Am 23.10.2011 um 15:45 schrieb Aashish Mittal <[email protected]>:

Quoted 11 lines

Hi

I'm working on powerpc e500 machine (Freescale P2020RDB machine) and using the paravirtualization guest support available inside kvm and wanted to some bookkeeping of my own on the shared page and from what i understood is that a single tlb entry is reserved for the shared page (magic page) when running the guest which is never preempted or flushed but when i tried printing out the dtlb misses on this shared page it show me a large number of dtlb misses on this shared page which means that this page is being preempted or flushed out in some way . Can someone pleas e clarify this situation if i'm understanding something wrong or is it a bug ?

We have seen similar effects and concluded it to be a bug, but didn't track down what exactly was causing it yet.

Alex

Quoted 1 line
#2 lore raw
In reply to Aashish Mittal
Quoted 11 lines · Aashish Mittal

Hi

I'm working on powerpc e500 machine (Freescale P2020RDB machine) and using the paravirtualization guest support available inside kvm and wanted
to some bookkeeping of my own on the shared page and from what i
understood is that a single tlb entry is reserved for the shared page (magic page) when running the guest which is never preempted or flushed
but when i tried printing out the dtlb misses on this shared page it show
me a large number of dtlb misses on this shared page which means that this page is being preempted or flushed out in some way . Can someone pleas e clarify this situation if i'm understanding something wrong or is it a bug ?

What kernel are you running, and how many is "a large number"?

The magic page TLB entry will need to be reloaded if the shadow PID gets invalidated, another guest gets scheduled on that CPU, or if something like hugetlbfs steps on that TLB entry.

If you're running a kernel that predates shadow PID support, it will be invalidated (along with everything else in the TLB) whenever you schedule out or change guest execution mode .

-Scott