Re: [PATCH v3 2/2] kvm/arm64: Try stage2 block mapping for host device MMIO

[Keqian Zhu]

Hi Marc,

On 2021/4/15 18:23, Marc Zyngier wrote:
> On Thu, 15 Apr 2021 03:20:52 +0100,
> Keqian Zhu <zhukeqian1@xxxxxxxxxx> wrote:
>>
>> Hi Marc,
>>
>> On 2021/4/14 17:05, Marc Zyngier wrote:
>>> + Santosh, who found some interesting bugs in that area before.
>>>
>>> On Wed, 14 Apr 2021 07:51:09 +0100,
>>> Keqian Zhu <zhukeqian1@xxxxxxxxxx> wrote:
>>>>
>>>> The MMIO region of a device maybe huge (GB level), try to use
>>>> block mapping in stage2 to speedup both map and unmap.
>>>>
>>>> Compared to normal memory mapping, we should consider two more
>>>> points when try block mapping for MMIO region:
>>>>
>>>> 1. For normal memory mapping, the PA(host physical address) and
>>>> HVA have same alignment within PUD_SIZE or PMD_SIZE when we use
>>>> the HVA to request hugepage, so we don't need to consider PA
>>>> alignment when verifing block mapping. But for device memory
>>>> mapping, the PA and HVA may have different alignment.
>>>>
>>>> 2. For normal memory mapping, we are sure hugepage size properly
>>>> fit into vma, so we don't check whether the mapping size exceeds
>>>> the boundary of vma. But for device memory mapping, we should pay
>>>> attention to this.
>>>>
>>>> This adds device_rough_page_shift() to check these two points when
>>>> selecting block mapping size.
>>>>
>>>> Signed-off-by: Keqian Zhu <zhukeqian1@xxxxxxxxxx>
>>>> ---
>>>>  arch/arm64/kvm/mmu.c | 37 +++++++++++++++++++++++++++++++++----
>>>>  1 file changed, 33 insertions(+), 4 deletions(-)
>>>>
>>>> diff --git a/arch/arm64/kvm/mmu.c b/arch/arm64/kvm/mmu.c
>>>> index c59af5ca01b0..1a6d96169d60 100644
>>>> --- a/arch/arm64/kvm/mmu.c
>>>> +++ b/arch/arm64/kvm/mmu.c
>>>> @@ -624,6 +624,31 @@ static void kvm_send_hwpoison_signal(unsigned long address, short lsb)
>>>>  	send_sig_mceerr(BUS_MCEERR_AR, (void __user *)address, lsb, current);
>>>>  }
>>>>  
>>>> +/*
>>>> + * Find a max mapping size that properly insides the vma. And hva and pa must
>>>> + * have the same alignment to this mapping size. It's rough as there are still
>>>> + * other restrictions, will be checked by fault_supports_stage2_huge_mapping().
>>>> + */
>>>> +static short device_rough_page_shift(struct vm_area_struct *vma,
>>>> +				     unsigned long hva)
>>>
>>> My earlier question still stands. Under which circumstances would this
>>> function return something that is *not* the final mapping size? I
>>> really don't see a reason why this would not return the final mapping
>>> size.
>>
>> IIUC, all the restrictions are about alignment and area boundary.
>>
>> That's to say, HVA, IPA and PA must have same alignment within the
>> mapping size.  And the areas are memslot and vma, which means the
>> mapping size must properly fit into the memslot and vma.
>>
>> In this function, we just checked the alignment of HVA and PA, and
>> the boundary of vma.  So we still need to check the alignment of HVA
>> and IPA, and the boundary of memslot.  These will be checked by
>> fault_supports_stage2_huge_mapping().
> 
> But that's no different from what we do with normal memory, is it? So
> it really feels like we should have *one* function that deals with
> establishing the basic mapping size from the VMA (see below for what I
> have in mind).
Right. And it looks better.

> 
>>
>>>
>>>> +{
>>>> +	phys_addr_t pa = (vma->vm_pgoff << PAGE_SHIFT) + (hva - vma->vm_start);
>>>> +
>>>> +#ifndef __PAGETABLE_PMD_FOLDED
>>>> +	if ((hva & (PUD_SIZE - 1)) == (pa & (PUD_SIZE - 1)) &&
>>>> +	    ALIGN_DOWN(hva, PUD_SIZE) >= vma->vm_start &&
>>>> +	    ALIGN(hva, PUD_SIZE) <= vma->vm_end)
>>>> +		return PUD_SHIFT;
>>>> +#endif
>>>> +
>>>> +	if ((hva & (PMD_SIZE - 1)) == (pa & (PMD_SIZE - 1)) &&
>>>> +	    ALIGN_DOWN(hva, PMD_SIZE) >= vma->vm_start &&
>>>> +	    ALIGN(hva, PMD_SIZE) <= vma->vm_end)
>>>> +		return PMD_SHIFT;
>>>> +
>>>> +	return PAGE_SHIFT;
>>>> +}
>>>> +
>>>>  static bool fault_supports_stage2_huge_mapping(struct kvm_memory_slot *memslot,
>>>>  					       unsigned long hva,
>>>>  					       unsigned long map_size)
>>>> @@ -769,7 +794,10 @@ static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa,
>>>>  		return -EFAULT;
>>>>  	}
>>>>  
>>>> -	/* Let's check if we will get back a huge page backed by hugetlbfs */
>>>> +	/*
>>>> +	 * Let's check if we will get back a huge page backed by hugetlbfs, or
>>>> +	 * get block mapping for device MMIO region.
>>>> +	 */
>>>>  	mmap_read_lock(current->mm);
>>>>  	vma = find_vma_intersection(current->mm, hva, hva + 1);
>>>>  	if (unlikely(!vma)) {
>>>> @@ -780,11 +808,12 @@ static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa,
>>>>  
>>>>  	if (is_vm_hugetlb_page(vma))
>>>>  		vma_shift = huge_page_shift(hstate_vma(vma));
>>>> +	else if (vma->vm_flags & VM_PFNMAP)
>>>> +		vma_shift = device_rough_page_shift(vma, hva);
>>>
>>> What prevents a VMA from having both VM_HUGETLB and VM_PFNMAP? This is
>>> pretty unlikely, but I'd like to see this case catered for.
>>>
>> I'm not sure whether VM_HUGETLB and VM_PFNMAP are compatible, and I
>> failed to find a case.
>>
>> VM_PFNMAP is used for page-ranges managed without "struct page",
>> just pure PFN.  IIUC, VM_HUGETLB is used for hugetlbfs, which always
>> has "struct page".  So I think they should not be compatible,
>> otherwise it's a bug of driver.
> 
> For now, maybe. But huge mappings of PFN could land at some point, and
> it'd be hard to catch. I think this case deserves a VM_BUG_ON().
OK.

> 
>>
>>>>  	else
>>>>  		vma_shift = PAGE_SHIFT;
>>>>  
>>>> -	if (logging_active ||
>>>> -	    (vma->vm_flags & VM_PFNMAP)) {
>>>> +	if (logging_active) {
> 
> BTW, don't you introduce a bug here? Logging shouldn't affect device
> mappings.
I think it's not a bug, because for memlsot with VM_PFNMAP, the logging_active is always false.

In kvm_arch_prepare_memory_region(), we make sure KVM_MEM_LOG_DIRTY_PAGES can't be set for a VM_PFNMAP memslot.
Then in __kvm_set_memory_region(), we're sure dirty_bitmap is not allocated for this memslot.
Then memslot_is_logging() will return false for this memslot.

> 
> 
>>>>  		force_pte = true;
>>>>  		vma_shift = PAGE_SHIFT;
>>>>  	}
>>>> @@ -855,7 +884,7 @@ static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa,
>>>>  
>>>>  	if (kvm_is_device_pfn(pfn)) {
>>>>  		device = true;
>>>> -		force_pte = true;
>>>> +		force_pte = (vma_pagesize == PAGE_SIZE);
>>>
>>> Why do we need to set force_pte if we are already dealing with
>>> PAGE_SIZE? I guess you are doing this for the sake of avoiding the
>>> call to transparent_hugepage_adjust(), right?
>> Yes.
>>
>>>
>>> I'd rather you simply don't try to upgrade a device mapping by
>>> explicitly checking for this and keep force_pte for *memory*
>>> exclusively.
>> Agree, that's better.
>>
>>>
>>> Santosh, can you please take a look at this series and try to see if
>>> the problem you fixed in [1] (which ended up as commit 91a2c34b7d6f)
>>> is still OK with this series?
>> I searched the initial version[*], VM_PFNMAP is set when we call
>> gfn_to_pfn_prot()->vma_mmio_fault()->remap_pfn_range().  Then the
>> check of VM_PFNMAP in user_mem_abort() failed, so we will try to
>> call transparent_hugepage_adjust() for device pfn.
>>
>> In that case, our logic of trying block mapping for MMIO is not
>> used. And we still set force_pte for device pfn, so this bugfix is
>> not affected. Santosh, do you agree that?
> 
> But isn't what we just agreed to get rid of just above?
Yes, I agree to get rid of force_pte for device. I'm sure your code
doesn't break the bugfix.

> 
>>
>> I still found that the reason vfio_pci does not have this
>> bug. vfio_pci set VM_PFNMAP for vma when userspace calls mmap().  I
>> will apply this logic for vfio_mdev too, let's see what vfio
>> maintainer think about it.
> 
> I think that'd be good to see what Alex thinks about it...
> 
> Here's the changes I propose. It is completely untested, of course.
> 
> Thanks,
> 
> 	M.
> 
> diff --git a/arch/arm64/kvm/mmu.c b/arch/arm64/kvm/mmu.c
> index 8711894db8c2..f32d956cc199 100644
> --- a/arch/arm64/kvm/mmu.c
> +++ b/arch/arm64/kvm/mmu.c
> @@ -738,6 +738,35 @@ transparent_hugepage_adjust(struct kvm_memory_slot *memslot,
>  	return PAGE_SIZE;
>  }
>  
> +static int get_vma_page_shift(struct vm_area_struct *vma, unsigned long hva)
> +{
> +	unsigned long pa;
> +
> +	if (is_vm_hugetlb_page(vma) && !(vma->vm_flags & VM_PFNMAP))
> +		return huge_page_shift(hstate_vma(vma));
> +
> +	if (!(vma->vm_flags & VM_PFNMAP))
> +		return PAGE_SHIFT;
> +
> +	VM_BUG_ON(is_vm_hugetlb_page(vma));
> +
> +	pa = (vma->vm_pgoff << PAGE_SHIFT) + (hva - vma->vm_start);
> +
> +#ifndef __PAGETABLE_PMD_FOLDED
> +	if ((hva & (PUD_SIZE - 1)) == (pa & (PUD_SIZE - 1)) &&
> +	    ALIGN_DOWN(hva, PUD_SIZE) >= vma->vm_start &&
> +	    ALIGN(hva, PUD_SIZE) <= vma->vm_end)
> +		return PUD_SHIFT;
> +#endif
> +
> +	if ((hva & (PMD_SIZE - 1)) == (pa & (PMD_SIZE - 1)) &&
> +	    ALIGN_DOWN(hva, PMD_SIZE) >= vma->vm_start &&
> +	    ALIGN(hva, PMD_SIZE) <= vma->vm_end)
> +		return PMD_SHIFT;
> +
> +	return PAGE_SHIFT;
> +}
> +
>  static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa,
>  			  struct kvm_memory_slot *memslot, unsigned long hva,
>  			  unsigned long fault_status)
> @@ -778,13 +807,9 @@ static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa,
>  		return -EFAULT;
>  	}
>  
> -	if (is_vm_hugetlb_page(vma))
> -		vma_shift = huge_page_shift(hstate_vma(vma));
> -	else
> -		vma_shift = PAGE_SHIFT;
> +	vma_shift = get_vma_page_shift(vma, hva);
>  
> -	if (logging_active ||
> -	    (vma->vm_flags & VM_PFNMAP)) {
> +	if (logging_active && !(vma->vm_flags & VM_PFNMAP)) {
Maybe we don't need this. I can add some comments to explain it.

>  		force_pte = true;
>  		vma_shift = PAGE_SHIFT;
>  	}
> @@ -854,8 +879,17 @@ static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa,
>  		return -EFAULT;
>  
>  	if (kvm_is_device_pfn(pfn)) {
> +		/*
> +		 * If the page was identified as device early by looking at
> +		 * the VMA flags, vma_pagesize is already representing the
> +		 * largest quantity we can map.  If instead it was mapped
> +		 * via gfn_to_pfn_prot(), vma_pagesize is set to PAGE_SIZE
> +		 * and must not be upgraded.
> +		 *
> +		 * In both cases, we don't let transparent_hugepage_adjust()
> +		 * change things at the last minute.
> +		 */
>  		device = true;
> -		force_pte = true;
>  	} else if (logging_active && !write_fault) {
>  		/*
>  		 * Only actually map the page as writable if this was a write
> @@ -876,7 +910,7 @@ static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa,
>  	 * If we are not forced to use page mapping, check if we are
>  	 * backed by a THP and thus use block mapping if possible.
>  	 */
> -	if (vma_pagesize == PAGE_SIZE && !force_pte)
> +	if (vma_pagesize == PAGE_SIZE && !(force_pte || device))
>  		vma_pagesize = transparent_hugepage_adjust(memslot, hva,
>  							   &pfn, &fault_ipa);
>  	if (writable)
> 
Looks good to me. :)

I will test it. And when I send v4, should I add your Suggested-by or SoB?


Thanks,
Keqian