Re: [PATCH] x86: kvm: hyper-v: deal with buggy TLB flush requests from WS2012

[Sean Christopherson <sean.j.christopherson@intel.com>]

On Thu, Apr 18, 2019 at 06:47:56PM +0200, Paolo Bonzini wrote:
> On 18/04/19 16:17, Sean Christopherson wrote:
> > On Wed, Mar 20, 2019 at 06:43:20PM +0100, Vitaly Kuznetsov wrote:
> >> It was reported that with some special Multi Processor Group configuration,
> >> e.g:
> >>  bcdedit.exe /set groupsize 1
> >>  bcdedit.exe /set maxgroup on
> >>  bcdedit.exe /set groupaware on
> >> for a 16-vCPU guest WS2012 shows BSOD on boot when PV TLB flush mechanism
> >> is in use.
> >>
> >> Tracing kvm_hv_flush_tlb immediately reveals the issue:
> >>
> >>  kvm_hv_flush_tlb: processor_mask 0x0 address_space 0x0 flags 0x2
> >>
> >> The only flag set in this request is HV_FLUSH_ALL_VIRTUAL_ADDRESS_SPACES,
> >> however, processor_mask is 0x0 and no HV_FLUSH_ALL_PROCESSORS is specified.
> >> We don't flush anything and apparently it's not what Windows expects.
> >>
> >> TLFS doesn't say anything about such requests and newer Windows versions
> >> seem to be unaffected. This all feels like a WS2012 bug, which is, however,
> >> easy to workaround in KVM: let's flush everything when we see an empty
> >> flush request, over-flushing doesn't hurt.
> >>
> >> Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
> >> ---
> >>  arch/x86/kvm/hyperv.c | 12 +++++++++++-
> >>  1 file changed, 11 insertions(+), 1 deletion(-)
> >>
> >> diff --git a/arch/x86/kvm/hyperv.c b/arch/x86/kvm/hyperv.c
> >> index 421899f6ad7b..5887f7d22ac6 100644
> >> --- a/arch/x86/kvm/hyperv.c
> >> +++ b/arch/x86/kvm/hyperv.c
> >> @@ -1371,7 +1371,17 @@ static u64 kvm_hv_flush_tlb(struct kvm_vcpu *current_vcpu, u64 ingpa,
> >>  
> >>  		valid_bank_mask = BIT_ULL(0);
> >>  		sparse_banks[0] = flush.processor_mask;
> >> -		all_cpus = flush.flags & HV_FLUSH_ALL_PROCESSORS;
> >> +
> >> +		/*
> >> +		 * WS2012 seems to be buggy, under certain conditions it is
> >> +		 * possible to observe requests with processor_mask = 0x0 and
> >> +		 * no HV_FLUSH_ALL_PROCESSORS flag set. It also seems that
> > 
> > "and no HV_FLUSH_ALL_PROCESSORS flag set" is awkward, and probably
> > extraneous.  The whole comment is a probably a bit more verbose than it
> > needs to be, e.g. most readers won't care how we came to the conclusion
> > that 'processor_mask == 0', and those that care about the background will
> > read the changelog anyways.
> > 
> > Maybe something like this:
> > 
> > 		/*
> > 		 * Some Windows versions, e.g. WS2012, use processor_mask = 0
> > 		 * in lieu of the dedicated flag to flush all processors.
> > 		 */
> 
> Hmm, not really.  "In lieu" seems intentional.  "without" is more accurate.

True, probably isn't exactly by design.

> My take:
> 
>                  * Work around possible WS2012 bug: it sends hypercalls
>                  * with processor_mask = 0x0 and HV_FLUSH_ALL_PROCESSORS clear,
>                  * while also expecting us to flush something and crashing if
>                  * we don't. Let's treat processor_mask == 0 same as
>                  * HV_FLUSH_ALL_PROCESSORS.

Nice.  What about this for the last sentence?  Either way, LGTM.

  For simplicity, flush all CPUs if the guest neglected to set processor_mask.

>                  */
> 
> Paolo
> 
> 
> > 
> > 
> >> +		 * Windows actually expects us to flush something and crashes
> >> +		 * otherwise. Let's treat processor_mask == 0 same as
> >> +		 * HV_FLUSH_ALL_PROCESSORS.
> >> +		 */
> >> +		all_cpus = (flush.flags & HV_FLUSH_ALL_PROCESSORS) ||
> >> +			(flush.processor_mask == 0);
> > 
> > Nits:
> > 
> > Personal preference, but I like '!flush.processor_mask' in this case as it
> > immediately conveys that we're handling the scenario where the guest didn't
> > set a mask.  Then there wouldn't be a visual need for the second set of
> > parentheses.
> > 
> > Aligning its indentation with the first first chunk of the statement would
> > also be nice, but again, personal preference.  :-)
> > 
> >>  	} else {
> >>  		if (unlikely(kvm_read_guest(kvm, ingpa, &flush_ex,
> >>  					    sizeof(flush_ex))))
> >> -- 
> >> 2.20.1
> >>
>