Re: [PATCH] kvm: arm: Skip stage2 huge mappings for unaligned ipa backed by THP

[Suzuki K Poulose <suzuki.poulose@arm.com>]

On 10/04/2019 03:20, Zenghui Yu wrote:
> 
> On 2019/4/9 22:59, Suzuki K Poulose wrote:
>> Hi Zenghui
>>
>> On 04/09/2019 09:05 AM, Zenghui Yu wrote:
>>>
>>>
>>> On 2019/4/9 2:40, Suzuki K Poulose wrote:
>>>> Hi Zenhui,
>>>>
>>>> On 04/08/2019 04:11 PM, Zenghui Yu wrote:
>>>>> Hi Suzuki,
>>>>>
>>>>> Thanks for the reply.
>>>>>
>>>>
>>>> ...
>>>>
>>>>>>> Hi Suzuki,
>>>>>>>
>>>>>>> Why not making use of fault_supports_stage2_huge_mapping()?  Let
>>>>>>> it do
>>>>>>> some checks for us.
>>>>>>>
>>>>>>> fault_supports_stage2_huge_mapping() was intended to do a *two-step*
>>>>>>> check to tell us that can we create stage2 huge block mappings,
>>>>>>> and this
>>>>>>> check is both for hugetlbfs and THP.  With commit
>>>>>>> a80868f398554842b14,
>>>>>>> we pass PAGE_SIZE as "map_size" for normal size pages (which
>>>>>>> turned out
>>>>>>> to be almost meaningless), and unfortunately the THP check no longer
>>>>>>> works.
>>>>>>
>>>>>> Thats correct.
>>>>>>
>>>>>>>
>>>>>>> So we want to rework *THP* check process.  Your patch fixes the first
>>>>>>> checking-step, but the second is still missed, am I wrong?
>>>>>>
>>>>>> It fixes the step explicitly for the THP by making sure that the
>>>>>> GPA and
>>>>>> the HVA are aligned to the map size.
>>>>>
>>>>> Yes, I understand how your patch had fixed the issue.  But what I'm
>>>>> really concerned about here is the *second* checking-step in
>>>>> fault_supports_stage2_huge_mapping().
>>>>>
>>>>> We have to check if we are mapping a non-block aligned or non-block
>>>>> sized memslot, if so, we can not create block mappings for the
>>>>> beginning
>>>>> and end of this memslot.  This is what the second part of
>>>>> fault_supports_stage2_huge_mapping() had done.
>>>>>
>>>>> I haven't seen this checking-step in your patch, did I miss something?
>>>>>
>>>>
>>>> I see.
>>>>
>>>>>> I don't think this calls for a VM_BUG_ON(). It is simply a case where
>>>>>> the GPA is not aligned to HVA, but for normal VMA that could be
>>>>>> made THP.
>>>>>>
>>>>>> We had this VM_BUG_ON(), which would have never hit because we would
>>>>>> have set force_pte if they were not aligned.
>>>>>
>>>>> Yes, I agree.
>>>>>
>>>>>>>> +        /* Skip memslots with unaligned IPA and user address */
>>>>>>>> +        if ((gfn & mask) != (pfn & mask))
>>>>>>>> +            return false;
>>>>>>>>            if (pfn & mask) {
>>>>>>>>                *ipap &= PMD_MASK;
>>>>>>>>                kvm_release_pfn_clean(pfn);
>>>>>>>>
>>>>>>>
>>>>>>> ---8>---
>>>>>>>
>>>>>>> Rework fault_supports_stage2_huge_mapping(), let it check THP again.
>>>>>>>
>>>>>>> Signed-off-by: Zenghui Yu <yuzenghui@huawei.com>
>>>>>>> ---
>>>>>>>    virt/kvm/arm/mmu.c | 11 ++++++++++-
>>>>>>>    1 file changed, 10 insertions(+), 1 deletion(-)
>>>>>>>
>>>>>>> diff --git a/virt/kvm/arm/mmu.c b/virt/kvm/arm/mmu.c
>>>>>>> index 27c9583..5e1b258 100644
>>>>>>> --- a/virt/kvm/arm/mmu.c
>>>>>>> +++ b/virt/kvm/arm/mmu.c
>>>>>>> @@ -1632,6 +1632,15 @@ static bool
>>>>>>> fault_supports_stage2_huge_mapping(struct kvm_memory_slot *memslot,
>>>>>>>        uaddr_end = uaddr_start + size;
>>>>>>>
>>>>>>>        /*
>>>>>>> +     * If the memslot is _not_ backed by hugetlbfs, then check if it
>>>>>>> +     * can be backed by transparent hugepages.
>>>>>>> +     *
>>>>>>> +     * Currently only PMD_SIZE THPs are supported, revisit it later.
>>>>>>> +     */
>>>>>>> +    if (map_size == PAGE_SIZE)
>>>>>>> +        map_size = PMD_SIZE;
>>>>>>> +
>>>>>>
>>>>>> This looks hackish. What is we support PUD_SIZE huge page in the
>>>>>> future
>>>>>> ?
>>>>>
>>>>> Yes, this might make the code a little difficult to understand. But by
>>>>> doing so, we follow the same logic before commit a80868f398554842b14,
>>>>> that said, we do the two-step checking for normal size pages in
>>>>> fault_supports_stage2_huge_mapping(), to decide if we can create THP
>>>>> mappings for these pages.
>>>>>
>>>>> As for PUD_SIZE THPs, to be honest, I have no idea now :(
>>>>
>>>> How about the following diff ?
>>>>
>>>> diff --git a/virt/kvm/arm/mmu.c b/virt/kvm/arm/mmu.c
>>>> index 97b5417..98e5cec 100644
>>>> --- a/virt/kvm/arm/mmu.c
>>>> +++ b/virt/kvm/arm/mmu.c
>>>> @@ -1791,7 +1791,8 @@ static int user_mem_abort(struct kvm_vcpu
>>>> *vcpu, phys_addr_t fault_ipa,
>>>>             * currently supported. This code will need to be
>>>>             * updated to support other THP sizes.
>>>>             */
>>>> -        if (transparent_hugepage_adjust(&pfn, &fault_ipa))
>>>> +        if (fault_supports_stage2_huge_mappings(memslot, hva,
>>>> PMD_SIZE) &&
>>>> +            transparent_hugepage_adjust(&pfn, &fault_ipa))
>>>>                vma_pagesize = PMD_SIZE;
>>>>        }
>>>
>>> I think this is good enough for the issue.
>>>
>>> (One minor concern: With this change, it seems that we no longer need
>>> "force_pte" and can just use "logging_active" instead. But this is not
>>> much related to what we're fixing.)
>>
>> I would still leave the force_pte there to avoid checking for a THP case
>> in a situation where we forced to PTE level mapping on a hugepage backed
>> VMA. It would serve to avoid another check.
> 
> Hi Suzuki,
> 
> Yes, I agree, thanks.

Cool, I have a patch to fix this properly and two other patches to clean up
and unify the way we handle the THP backed hugepages. Will send them out after
a bit of testing, later today.

Cheers
Suzuki