Re: [RFC KVM 06/27] KVM: x86: Exit KVM isolation on IRQ entry

[Alexandre Chartre <alexandre.chartre@oracle.com>]

On 5/14/19 9:07 AM, Peter Zijlstra wrote:
> On Mon, May 13, 2019 at 11:13:34AM -0700, Andy Lutomirski wrote:
>> On Mon, May 13, 2019 at 9:28 AM Alexandre Chartre
>> <alexandre.chartre@oracle.com> wrote:
> 
>>> Actually, I am not sure this is effectively useful because the IRQ
>>> handler is probably faulting before it tries to exit isolation, so
>>> the isolation exit will be done by the kvm page fault handler. I need
>>> to check that.
>>>
>>
>> The whole idea of having #PF exit with a different CR3 than was loaded
>> on entry seems questionable to me.  I'd be a lot more comfortable with
>> the whole idea if a page fault due to accessing the wrong data was an
>> OOPS and the code instead just did the right thing directly.
> 
> So I've ran into this idea before; it basically allows a lazy approach
> to things.
> 
> I'm somewhat conflicted on things, on the one hand, changing CR3 from
> #PF is a natural extention in that #PF already changes page-tables (for
> userspace / vmalloc etc..), on the other hand, there's a thin line
> between being lazy and being sloppy.
> 
> If we're going down this route; I think we need a very coherent design
> and strong rules.
> 

Right. We should particularly ensure that the KVM page-table remains a
subset of the kernel page-table, in particular page-table changes (e.g.
for vmalloc etc...) should happen in the kernel page-table and not in
the kvm page-table.

So we should probably enforce switching to the kernel page-table when
doing operation like vmalloc. The current code doesn't enforce it, but
I can see it faulting, when doing any allocation (because the kvm page
table doesn't have all structures used during an allocation).

alex.