Re: [PATCH 5/6] KVM: PPC: BOOK3S: Use hpte_update_in_progress to track invalid hpte during an hpte update

["Aneesh Kumar K.V" <aneesh.kumar@linux.vnet.ibm.com>]

Paul Mackerras <paulus@samba.org> writes:

> On Sun, Jun 29, 2014 at 04:47:34PM +0530, Aneesh Kumar K.V wrote:
>> As per ISA, we first need to mark hpte invalid (V=0) before we update
>> the hpte lower half bits. With virtual page class key protection mechanism we want
>> to send any fault other than key fault to guest directly without
>> searching the hash page table. But then we can get NO_HPTE fault while
>> we are updating the hpte. To track that add a vm specific atomic
>> variable that we check in the fault path to always send the fault
>> to host.
>
> [...]
>
>> @@ -750,13 +751,15 @@ int kvmppc_book3s_hv_page_fault(struct kvm_run *run, struct kvm_vcpu *vcpu,
>>  	r &= rcbits | ~(HPTE_R_R | HPTE_R_C);
>>  
>>  	if (be64_to_cpu(hptep[0]) & HPTE_V_VALID) {
>> -		/* HPTE was previously valid, so we need to invalidate it */
>> +		/*
>> +		 * If we had mapped this hpte before, we now need to
>> +		 * invalidate that.
>> +		 */
>>  		unlock_rmap(rmap);
>> -		/* Always mark HPTE_V_ABSENT before invalidating */
>> -		kvmppc_unmap_host_hpte(kvm, hptep);
>>  		kvmppc_invalidate_hpte(kvm, hptep, index);
>>  		/* don't lose previous R and C bits */
>>  		r |= be64_to_cpu(hptep[1]) & (HPTE_R_R | HPTE_R_C);
>> +		hpte_invalidated = true;
>
> So now we're not setting the ABSENT bit before invalidating the HPTE.
> That means that another guest vcpu could do an H_ENTER which could
> think that this HPTE is free and use it for another unrelated guest
> HPTE, which would be bad...

But henter looks at HPTE_V_HVLOCK, and we keep that set through out. But
I will double the code again to make sure it is safe in the above
scenario.

>
>> @@ -1144,8 +1149,8 @@ static int kvm_test_clear_dirty_npages(struct kvm *kvm, unsigned long *rmapp)
>>  				npages_dirty = n;
>>  			eieio();
>>  		}
>> -		kvmppc_map_host_hpte(kvm, &v, &r);
>> -		hptep[0] = cpu_to_be64(v & ~HPTE_V_HVLOCK);
>> +		hptep[0] = cpu_to_be64(v & ~HPTE_V_LOCK);
>> +		atomic_dec(&kvm->arch.hpte_update_in_progress);
>
> Why are we using LOCK rather than HVLOCK now?  (And why didn't you
> mention this change and its rationale in the patch description?)

Sorry, that is a typo. I intend to use HPTE_V_HVLOCK.

-aneesh