Re: [PATCH v5 2/2] kvm: Use huge pages for DAX-backed files

[Liran Alon]

> On 13 Dec 2019, at 19:19, Sean Christopherson <sean.j.christopherson@xxxxxxxxx> wrote:
> 
> On Fri, Dec 13, 2019 at 03:07:31AM +0200, Liran Alon wrote:
>> 
>>> On 12 Dec 2019, at 21:55, Barret Rhoden <brho@xxxxxxxxxx> wrote:
>>> 
>>>>>> Note that KVM already faulted in the page (or huge page) in the host's
>>>>>> page table, and we hold the KVM mmu spinlock.  We grabbed that lock in
>>>>>> kvm_mmu_notifier_invalidate_range_end, before checking the mmu seq.
>>>>>> 
>>>>>> Signed-off-by: Barret Rhoden <brho@xxxxxxxxxx>
>>>>> 
>>>>> I donât think the right place to change for this functionality is
>>>>> transparent_hugepage_adjust() which is meant to handle PFNs that are
>>>>> mapped as part of a transparent huge-page.
>>>>> 
>>>>> For example, this would prevent mapping DAX-backed file page as 1GB.  As
>>>>> transparent_hugepage_adjust() only handles the case (level ==
>>>>> PT_PAGE_TABLE_LEVEL).
> 
> Teaching thp_adjust() how to handle 1GB wouldn't be a bad thing.  It's
> unlikely THP itself will support 1GB pages any time soon, but having the
> logic there wouldn't hurt anything.

I agree.

> 
>>>>> As you are parsing the page-tables to discover the page-size the PFN is
>>>>> mapped in, I think you should instead modify kvm_host_page_size() to
>>>>> parse page-tables instead of rely on vma_kernel_pagesize() (Which relies
>>>>> on vma->vm_ops->pagesize()) in case of is_zone_device_page().
>>>>> 
>>>>> The main complication though of doing this is that at this point you
>>>>> donât yet have the PFN that is retrieved by try_async_pf(). So maybe you
>>>>> should consider modifying the order of calls in tdp_page_fault() &
>>>>> FNAME(page_fault)().
>>>>> 
>>>>> -Liran
>>>> Or alternatively when thinking about it more, maybe just rename
>>>> transparent_hugepage_adjust() to not be specific to THP and better handle
>>>> the case of parsing page-tables changing mapping-level to 1GB.
>>>> That is probably easier and more elegant.
> 
> Agreed.
> 
>>> I can rename it to hugepage_adjust(), since it's not just THP anymore.
> 
> Or maybe allowed_hugepage_adjust()?  To pair with disallowed_hugepage_adjust(),
> which adjusts KVM's page size in the opposite direction to avoid the iTLB
> multi-hit issue.
> 
>> 
>> Sounds good.
>> 
>>> 
>>> I was a little hesitant to change the this to handle 1 GB pages with this
>>> patchset at first.  I didn't want to break the non-DAX case stuff by doing
>>> so.
>> 
>> Why would it affect non-DAX case?
>> Your patch should just make hugepage_adjust() to parse page-tables only in case is_zone_device_page(). Otherwise, page tables shouldnât be parsed.
>> i.e. THP merged pages should still be detected by PageTransCompoundMap().
> 
> I think what Barret is saying is that teaching thp_adjust() how to do 1gb
> mappings would naturally affect the code path for THP pages.  But I agree
> that it would be superficial.
> 
>>> Specifically, can a THP page be 1 GB, and if so, how can you tell?  If you
>>> can't tell easily, I could walk the page table for all cases, instead of
>>> just zone_device().
> 
> No, THP doesn't currently support 1gb pages.  Expliciting returning
> PMD_SIZE on PageTransCompoundMap() would be a good thing from a readability
> perspective.

Right.

> 
>> I prefer to walk page-tables only for is_zone_device_page().
>> 
>>> 
>>> I'd also have to drop the "level == PT_PAGE_TABLE_LEVEL" check, I think,
>>> which would open this up to hugetlbfs pages (based on the comments).  Is
>>> there any reason why that would be a bad idea?
> 
> No, the "level == PT_PAGE_TABLE_LEVEL" check is to filter out the case
> where KVM is already planning on using a large page, e.g. when the memory
> is backed by hugetlbs.

Right.

> 
>> KVM already supports mapping 1GB hugetlbfs pages. As level is set to
>> PUD-level by
>> tdp_page_fault()->mapping_level()->host_mapping_level()->kvm_host_page_size()->vma_kernel_pagesize().
>> As VMA which is mmap of hugetlbfs sets vma->vm_ops to hugetlb_vm_ops() where
>> hugetlb_vm_op_pagesize() will return appropriate page-size.
>> 
>> Specifically, I donât think THP ever merges small pages to 1GB pages. I think
>> this is why transparent_hugepage_adjust() checks PageTransCompoundMap() only
>> in case level == PT_PAGE_TABLE_LEVEL. I think you should keep this check in
>> the case of !is_zone_device_page().
> 
> I would add 1gb support for DAX as a third patch in this series.  To pave
> the way in patch 2/2, change it to replace "bool pfn_is_huge_mapped()" with
> "int host_pfn_mapping_level()", and maybe also renaming host_mapping_level()
> to host_vma_mapping_level() to avoid confusion.

I agree.
So also rename kvm_host_page_size() to kvm_host_vma_page_size() :)

> 
> Then allowed_hugepage_adjust() would look something like:
> 
> static void allowed_hugepage_adjust(struct kvm_vcpu *vcpu, gfn_t gfn,
> 				    kvm_pfn_t *pfnp, int *levelp, int max_level)
> {
> 	kvm_pfn_t pfn = *pfnp;
> 	int level = *levelp;	
> 	unsigned long mask;
> 
> 	if (is_error_noslot_pfn(pfn) || !kvm_is_reserved_pfn(pfn) ||
> 	    level == PT_PAGE_TABLE_LEVEL)
> 		return;
> 
> 	/*
> 	 * mmu_notifier_retry() was successful and mmu_lock is held, so
> 	 * the pmd/pud can't be split from under us.
> 	 */
> 	level = host_pfn_mapping_level(vcpu->kvm, gfn, pfn);
> 
> 	*levelp = level = min(level, max_level);
> 	mask = KVM_PAGES_PER_HPAGE(level) - 1;
> 	VM_BUG_ON((gfn & mask) != (pfn & mask));
> 	*pfnp = pfn & ~mask;

Why donât you still need to kvm_release_pfn_clean() for original pfn and kvm_get_pfn() for new huge-page start pfn?

> }

Yep. This is similar to what I had in mind.

Then just put logic of parsing page-tables in case itâs is_zone_device_page() or returning PMD_SIZE in case itâs PageTransCompoundMap() inside host_pfn_mapping_level(). This make code very straight-forward.

-Liran