On 01/13/2011 05:51 PM, Roedel, Joerg wrote:
On Thu, Jan 13, 2011 at 10:42:01AM -0500, Avi Kivity wrote:
> On 01/13/2011 05:22 PM, Joerg Roedel wrote:
> > The vmexit path on SVM needs to restore the KERNEL_GS_BASE
> > MSR in order to savely execute the NMI handler. Otherwise a
> > pending NMI can occur after the STGI instruction and crash
> > the machine.
> > This makes it impossible to run perf and kvm in parallel on
> > an AMD machine in a stable way.
> >
> > Cc: stable@xxxxxxxxxx
> > Signed-off-by: Joerg Roedel<joerg.roedel@xxxxxxx>
> > ---
> > arch/x86/kvm/svm.c | 1 +
> > 1 files changed, 1 insertions(+), 0 deletions(-)
> >
> > diff --git a/arch/x86/kvm/svm.c b/arch/x86/kvm/svm.c
> > index 25bd1bc..8b9bc72 100644
> > --- a/arch/x86/kvm/svm.c
> > +++ b/arch/x86/kvm/svm.c
> > @@ -3637,6 +3637,7 @@ static void svm_vcpu_run(struct kvm_vcpu *vcpu)
> >
> > #ifdef CONFIG_X86_64
> > wrmsrl(MSR_GS_BASE, svm->host.gs_base);
> > + wrmsrl(MSR_KERNEL_GS_BASE, current->thread.gs);
> > #else
> > loadsegment(fs, svm->host.fs);
> > #endif
>
> Why would an NMI crash if MSR_KERNEL_GS_BASE is bad?
>
> I see save_paranoid depends on MSR_GS_BASE (specifically its sign, which
> is bad for the new instructions that allow userspace to write gsbase),
> but not on MSR_KERNEL_GS_BASE.
Thats a good question. I have not idea. I spent some time trying to
figure this out (after I found out that wrong KERNEL_GS_BASE was the
cause of the crashes) but had no luck.
This also doesn't happen every time an NMI is delivered in svm_vcpu_run.
Sometimes it runs perfectly in parallel for a few minutues before the
machine triple-faults.
I also had a look at entry_64.S. The save_paranoid could not be the
cause because MSR_GS_BASE is already negative at this point. But the
re-schedule condition check at the end of the NMI handler code could
also not be the cause because the NMI happens while preemption (and
interrupts) are disabled (a re-schedule should also trigger
preempt-notifiers and restore KERNEL_GS_BASE).
I have it:
ENTRY(native_load_gs_index)
CFI_STARTPROC
pushfq_cfi
DISABLE_INTERRUPTS(CLBR_ANY & ~CLBR_RDI)
SWAPGS
gs_change:
movl %edi,%gs
2: mfence /* workaround */
SWAPGS
popfq_cfi
ret
If an nmi hits between the two SWAPGSs, it sees the guest's
MSR_KERNEL_GS_BASE as the host's MSR_GS_BASE.
An alternative to your fix would be to disable GIF around
load_gs_index() in kvm. I imagine it would be slower than your fix (not
a trivial tradeoff - wrmsr every lightweight exit, vs. clgi/stgi every
heavyweight exit).
Please update the changelog, and add a comment.
--
I have a truly marvellous patch that fixes the bug which this
signature is too narrow to contain.
--
To unsubscribe from this list: send the line "unsubscribe kvm" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
