[PATCH RESEND v15 07/10] KVM: arm: page logging 2nd stage fault handling

[Christoffer Dall]

On Tue, Jan 13, 2015 at 09:42:34AM -0800, Mario Smarduch wrote:
> On 01/12/2015 11:43 AM, Christoffer Dall wrote:
> > On Mon, Jan 12, 2015 at 11:04:45AM -0800, Mario Smarduch wrote:
> > 
> > [...]
> > 
> >>>>>> @@ -1059,12 +1104,35 @@ static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa,
> >>>>>>  	if (is_error_pfn(pfn))
> >>>>>>  		return -EFAULT;
> >>>>>>  
> >>>>>> -	if (kvm_is_device_pfn(pfn))
> >>>>>> +	if (kvm_is_device_pfn(pfn)) {
> >>>>>>  		mem_type = PAGE_S2_DEVICE;
> >>>>>> +		set_pte_flags = KVM_S2PTE_FLAG_IS_IOMAP;
> >>>>>> +	}
> >>>>>>  
> >>>>>>  	spin_lock(&kvm->mmu_lock);
> >>>>>>  	if (mmu_notifier_retry(kvm, mmu_seq))
> >>>>>>  		goto out_unlock;
> >>>>>> +
> >>>>>> +	/*
> >>>>>> +	 * When logging is enabled general page fault handling changes:
> >>>>>> +	 * -  Writable huge pages are dissolved on a read or write fault.
> >>>>>
> >>>>> why dissolve huge pages on a read fault?
> >>>>
> >>>> What I noticed on write you would dissolve, on read you
> >>>> rebuild THPs, flip back and forth like that, performance
> >>>> & convergence was really bad.
> >>>
> >>> ah, that makes sense, we should probably indicate that reasoning
> >>> somehow.  In fact, what threw me off was the use of the word "dissolve
> >>> huge pages" which is not really what you're doing on a read fault, there
> >>> you are just never adjusting to huge pages.
> >>>
> >>> I'm wondering why that would slow things down much though, the only cost
> >>> would be the extra tlb invalidation and replacing the PMD on a
> >>> subsequent write fault, but I trust your numbers nevertheless.
> >>
> >> If I understand correctly -
> >> you do few writes, dissolve a huge page insert pte TLB entries,
> >> then a read page fault installs a pmd clears the TLB cache
> >> for that range, and it repeats over. Appears like you
> >> need to constantly re-fault pte TLBs on writes to huge
> >> page range.
> > 
> > that makes good sense, thanks for the explanation.
> > 
> > [...]
> > 
> >>>>>  	} else {
> >>>>> +		unsigned long flags = 0;
> >>>>>  		pte_t new_pte = pfn_pte(pfn, mem_type);
> >>>>> +
> >>>>>  		if (writable) {
> >>>>>  			kvm_set_s2pte_writable(&new_pte);
> >>>>>  			kvm_set_pfn_dirty(pfn);
> >>>>>  		}
> >>>>>  		coherent_cache_guest_page(vcpu, hva, PAGE_SIZE,
> >>>>>  					  fault_ipa_uncached);
> >>>>> -		ret = stage2_set_pte(kvm, memcache, fault_ipa, &new_pte,
> >>>>> -			pgprot_val(mem_type) == pgprot_val(PAGE_S2_DEVICE));
> >>>>> +
> >>>>> +		if (pgprot_val(mem_type) == pgprot_val(PAGE_S2_DEVICE))
> >>>>> +			flags |= KVM_S2PTE_FLAG_IS_IOMAP;
> >>>>> +
> >>>>> +		if (memslot_is_logging(memslot))
> >>>>> +			flags |= KVM_S2_FLAG_LOGGING_ACTIVE;
> >>>> Now that it either IOMAP or LOGGING_ACTIVE do we need to acumulate flags?
> >>>> Although we don't know if device mappings will be handled here.
> >>>>
> >>>
> >>> so forget all I said about this in the past, I confused the code
> >>> checking for !cache with the iomap flag.
> >>>
> >>> So, I think you can always safeful assume that stage2_get_pmd() gives you
> >>> something valid back when you have the LOGGING flag set, because you
> >>> always call the function with a valid cache when the LOGGING flag is
> >>> set.  It could be worth adding the following to stage2_set_pte():
> >>>
> >>> VM_BUG_ON((flags & KVM_S2_FLAG_LOGGING_ACTIVE) && !cache)
> >>
> >> I see ok, thanks for clearing that up.
> >>
> >>>
> >>> As for this code, the IOMAP flag's only effect is that we return -EFAULT
> >>> if we are seeing an existing PTE for the faulting address.  This would
> >>> no longer be valid if we allow logging dirty device memory pages, so we
> >> Sorry, do you mean allow or disallow?
> > 
> > if we (by these patches) allow logging dirty pages for device memory,
> > then we...
> > 
> >>
> >>> really need to think about if there's any conceivable use case for this?
> 
> No I can't think of any use case to log Device address space.
> 
> So I could move forward - drop the IOMAP flag, and add the
> VM_BUG_ON to stage2_set_pte().
> 
add the VM_BUG_ON, but keep the IOMAP flag as a separate thing from page
logging (assuming we all agree they are orthogonal events), see other
mail thread.

-Christoffer