Re: [PATCH 09/10] KVM: PPC: Support SC1 hypercalls for PAPR in PR mode

Am 12.08.2011 um 09:43 schrieb David Gibson [off-list ref]:

On Fri, Aug 12, 2011 at 07:35:42AM +0200, Alexander Graf wrote:

quoted

=20
Am 12.08.2011 um 05:33 schrieb David Gibson [off-list ref]=

:

quoted

=20

quoted

On Tue, Aug 09, 2011 at 06:31:47PM +0200, Alexander Graf wrote:

quoted

PAPR defines hypercalls as SC1 instructions. Using these, the guest mod=

ifies

quoted

quoted

quoted

page tables and does other privileged operations that it wouldn't be al=

lowed

quoted

quoted

quoted

to do in supervisor mode.
=20
This patch adds support for PR KVM to trap these instructions and route=

 them

quoted

quoted

quoted

through the same PAPR hypercall interface that we already use for HV st=

yle

quoted

quoted

quoted

KVM.

=20
This will work on a powermac or bare metal host.  Unfortunately, it's
not enough on a pSeries LPAR host - the sc 1 instruction from the
guest problem state will go direct to the hypervisor, which will
return an error rather than trapping to the guest kernel.
=20
The only way around this I can see is for qemu to search for and patch
up sc 1 instructions to something else.  Obviously that would also
need some kernel support, and probably a capability to let it know if
it's necessary.

=20
Well I'd like to keep Qemu out of the patching business, so the
guest kernel would have to patch itself.

=20
Well sure, but guest patching itself means it can't run existing
kernels.  I thought qemu already patched a few things, ugly though
that approach is.

Nope, qemu doesn't patch guest code by itself. The only time the guest kerne=
l doesn't patch itself is the TPR acceleration for Windows - because we can'=
t modify the guest here.

I also don't think it's that important to support older Linux guests if it m=
eans we need to patch the guest from the outside :). If you really need to u=
se PHyP, just run a newer guest kernel or -M mac99.

One thing I agree with though is that we should fail the CAP enable if we ru=
n on broken hypervisors.

Alex

=20