[RFC PATCH v2 0/2] kvm/arm64: Try stage2 block mapping for host device MMIO

[RFC PATCH v2 0/2] kvm/arm64: Try stage2 block mapping for host device MMIO

[Keqian Zhu]

Hi all,

We have two pathes to build stage2 mapping for MMIO regions.

Create time's path and stage2 fault path.

Patch#1 removes the creation time's mapping of MMIO regions
Patch#2 tries stage2 block mapping for host device MMIO at fault path

Thanks,
Keqian

Keqian Zhu (2):
  kvm/arm64: Remove the creation time's mapping of MMIO regions
  kvm/arm64: Try stage2 block mapping for host device MMIO

 arch/arm64/kvm/mmu.c | 80 +++++++++++++++++++++++---------------------
 1 file changed, 41 insertions(+), 39 deletions(-)

-- 
2.19.1

[RFC PATCH v2 1/2] kvm/arm64: Remove the creation time's mapping of MMIO regions

[Keqian Zhu]

The MMIO regions may be unmapped for many reasons and can be remapped
by stage2 fault path. Map MMIO regions at creation time becomes a
minor optimization and makes these two mapping path hard to sync.

Remove the mapping code while keep the useful sanity check.

Signed-off-by: Keqian Zhu <zhukeqian1@huawei.com>
---
 arch/arm64/kvm/mmu.c | 38 +++-----------------------------------
 1 file changed, 3 insertions(+), 35 deletions(-)

diff --git a/arch/arm64/kvm/mmu.c b/arch/arm64/kvm/mmu.c
index 8711894db8c2..c59af5ca01b0 100644
--- a/arch/arm64/kvm/mmu.c
+++ b/arch/arm64/kvm/mmu.c
@@ -1301,7 +1301,6 @@ int kvm_arch_prepare_memory_region(struct kvm *kvm,
 {
 	hva_t hva = mem->userspace_addr;
 	hva_t reg_end = hva + mem->memory_size;
-	bool writable = !(mem->flags & KVM_MEM_READONLY);
 	int ret = 0;
 
 	if (change != KVM_MR_CREATE && change != KVM_MR_MOVE &&
@@ -1318,8 +1317,7 @@ int kvm_arch_prepare_memory_region(struct kvm *kvm,
 	mmap_read_lock(current->mm);
 	/*
 	 * A memory region could potentially cover multiple VMAs, and any holes
-	 * between them, so iterate over all of them to find out if we can map
-	 * any of them right now.
+	 * between them, so iterate over all of them.
 	 *
 	 *     +--------------------------------------------+
 	 * +---------------+----------------+   +----------------+
@@ -1330,50 +1328,20 @@ int kvm_arch_prepare_memory_region(struct kvm *kvm,
 	 */
 	do {
 		struct vm_area_struct *vma = find_vma(current->mm, hva);
-		hva_t vm_start, vm_end;
 
 		if (!vma || vma->vm_start >= reg_end)
 			break;
 
-		/*
-		 * Take the intersection of this VMA with the memory region
-		 */
-		vm_start = max(hva, vma->vm_start);
-		vm_end = min(reg_end, vma->vm_end);
-
 		if (vma->vm_flags & VM_PFNMAP) {
-			gpa_t gpa = mem->guest_phys_addr +
-				    (vm_start - mem->userspace_addr);
-			phys_addr_t pa;
-
-			pa = (phys_addr_t)vma->vm_pgoff << PAGE_SHIFT;
-			pa += vm_start - vma->vm_start;
-
 			/* IO region dirty page logging not allowed */
 			if (memslot->flags & KVM_MEM_LOG_DIRTY_PAGES) {
 				ret = -EINVAL;
-				goto out;
-			}
-
-			ret = kvm_phys_addr_ioremap(kvm, gpa, pa,
-						    vm_end - vm_start,
-						    writable);
-			if (ret)
 				break;
+			}
 		}
-		hva = vm_end;
+		hva = min(reg_end, vma->vm_end);
 	} while (hva < reg_end);
 
-	if (change == KVM_MR_FLAGS_ONLY)
-		goto out;
-
-	spin_lock(&kvm->mmu_lock);
-	if (ret)
-		unmap_stage2_range(&kvm->arch.mmu, mem->guest_phys_addr, mem->memory_size);
-	else if (!cpus_have_final_cap(ARM64_HAS_STAGE2_FWB))
-		stage2_flush_memslot(kvm, memslot);
-	spin_unlock(&kvm->mmu_lock);
-out:
 	mmap_read_unlock(current->mm);
 	return ret;
 }
-- 
2.19.1

[RFC PATCH v2 2/2] kvm/arm64: Try stage2 block mapping for host device MMIO

[Keqian Zhu]

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@huawei.com>
---

Mainly for RFC, not fully tested. I will fully test it when the
code logic is well accepted.

---
 arch/arm64/kvm/mmu.c | 42 ++++++++++++++++++++++++++++++++++++++----
 1 file changed, 38 insertions(+), 4 deletions(-)

diff --git a/arch/arm64/kvm/mmu.c b/arch/arm64/kvm/mmu.c
index c59af5ca01b0..224aa15eb4d9 100644
--- a/arch/arm64/kvm/mmu.c
+++ b/arch/arm64/kvm/mmu.c
@@ -624,6 +624,36 @@ static void kvm_send_hwpoison_signal(unsigned long address, short lsb)
 	send_sig_mceerr(BUS_MCEERR_AR, (void __user *)address, lsb, current);
 }
 
+/*
+ * Find a mapping size that properly insides the intersection of vma and
+ * memslot. And hva and pa have the same alignment to this mapping size.
+ * It's rough because there are still other restrictions, which will be
+ * checked by the following fault_supports_stage2_huge_mapping().
+ */
+static short device_rough_page_shift(struct kvm_memory_slot *memslot,
+				     struct vm_area_struct *vma,
+				     unsigned long hva)
+{
+	size_t size = memslot->npages * PAGE_SIZE;
+	hva_t sec_start = max(memslot->userspace_addr, vma->vm_start);
+	hva_t sec_end = min(memslot->userspace_addr + size, vma->vm_end);
+	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) >= sec_start &&
+	    ALIGN(hva, PUD_SIZE) <= sec_end)
+		return PUD_SHIFT;
+#endif
+
+	if ((hva & (PMD_SIZE - 1)) == (pa & (PMD_SIZE - 1)) &&
+	    ALIGN_DOWN(hva, PMD_SIZE) >= sec_start &&
+	    ALIGN(hva, PMD_SIZE) <= sec_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 +799,10 @@ 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 +813,12 @@ 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(memslot, vma, hva);
 	else
 		vma_shift = PAGE_SHIFT;
 
-	if (logging_active ||
-	    (vma->vm_flags & VM_PFNMAP)) {
+	if (logging_active) {
 		force_pte = true;
 		vma_shift = PAGE_SHIFT;
 	}
@@ -855,7 +889,7 @@ 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);
 	} else if (logging_active && !write_fault) {
 		/*
 		 * Only actually map the page as writable if this was a write
-- 
2.19.1

Re: [RFC PATCH v2 0/2] kvm/arm64: Try stage2 block mapping for host device MMIO

[Keqian Zhu]

Kind ping...

On 2021/3/16 21:43, Keqian Zhu wrote:
> Hi all,
> 
> We have two pathes to build stage2 mapping for MMIO regions.
> 
> Create time's path and stage2 fault path.
> 
> Patch#1 removes the creation time's mapping of MMIO regions
> Patch#2 tries stage2 block mapping for host device MMIO at fault path
> 
> Thanks,
> Keqian
> 
> Keqian Zhu (2):
>   kvm/arm64: Remove the creation time's mapping of MMIO regions
>   kvm/arm64: Try stage2 block mapping for host device MMIO
> 
>  arch/arm64/kvm/mmu.c | 80 +++++++++++++++++++++++---------------------
>  1 file changed, 41 insertions(+), 39 deletions(-)
> 

Re: [RFC PATCH v2 2/2] kvm/arm64: Try stage2 block mapping for host device MMIO

[Marc Zyngier]

On Tue, 16 Mar 2021 13:43:38 +0000,
Keqian Zhu <zhukeqian1@huawei.com> 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@huawei.com>
> ---
> 
> Mainly for RFC, not fully tested. I will fully test it when the
> code logic is well accepted.
> 
> ---
>  arch/arm64/kvm/mmu.c | 42 ++++++++++++++++++++++++++++++++++++++----
>  1 file changed, 38 insertions(+), 4 deletions(-)
> 
> diff --git a/arch/arm64/kvm/mmu.c b/arch/arm64/kvm/mmu.c
> index c59af5ca01b0..224aa15eb4d9 100644
> --- a/arch/arm64/kvm/mmu.c
> +++ b/arch/arm64/kvm/mmu.c
> @@ -624,6 +624,36 @@ static void kvm_send_hwpoison_signal(unsigned long address, short lsb)
>  	send_sig_mceerr(BUS_MCEERR_AR, (void __user *)address, lsb, current);
>  }
>  
> +/*
> + * Find a mapping size that properly insides the intersection of vma and
> + * memslot. And hva and pa have the same alignment to this mapping size.
> + * It's rough because there are still other restrictions, which will be
> + * checked by the following fault_supports_stage2_huge_mapping().

I don't think these restrictions make complete sense to me. If this is
a PFNMAP VMA, we should use the biggest mapping size that covers the
VMA, and not more than the VMA.

> + */
> +static short device_rough_page_shift(struct kvm_memory_slot *memslot,
> +				     struct vm_area_struct *vma,
> +				     unsigned long hva)
> +{
> +	size_t size = memslot->npages * PAGE_SIZE;
> +	hva_t sec_start = max(memslot->userspace_addr, vma->vm_start);
> +	hva_t sec_end = min(memslot->userspace_addr + size, vma->vm_end);
> +	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) >= sec_start &&
> +	    ALIGN(hva, PUD_SIZE) <= sec_end)
> +		return PUD_SHIFT;
> +#endif
> +
> +	if ((hva & (PMD_SIZE - 1)) == (pa & (PMD_SIZE - 1)) &&
> +	    ALIGN_DOWN(hva, PMD_SIZE) >= sec_start &&
> +	    ALIGN(hva, PMD_SIZE) <= sec_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 +799,10 @@ 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 +813,12 @@ 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(memslot, vma, hva);
>  	else
>  		vma_shift = PAGE_SHIFT;
>  
> -	if (logging_active ||
> -	    (vma->vm_flags & VM_PFNMAP)) {
> +	if (logging_active) {
>  		force_pte = true;
>  		vma_shift = PAGE_SHIFT;

But why should we downgrade to page-size mappings if logging? This is
a device, and you aren't moving the device around, are you? Or is your
device actually memory with a device mapping that you are trying to
migrate?

>  	}
> @@ -855,7 +889,7 @@ 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);
>  	} else if (logging_active && !write_fault) {
>  		/*
>  		 * Only actually map the page as writable if this was a write
> -- 
> 2.19.1
> 
> 

Thanks,

	M.

-- 
Without deviation from the norm, progress is not possible.

Re: [RFC PATCH v2 2/2] kvm/arm64: Try stage2 block mapping for host device MMIO

[Keqian Zhu]

Hi Marc,

On 2021/4/7 21:18, Marc Zyngier wrote:
> On Tue, 16 Mar 2021 13:43:38 +0000,
> Keqian Zhu <zhukeqian1@huawei.com> 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@huawei.com>
>> ---
>>
>> Mainly for RFC, not fully tested. I will fully test it when the
>> code logic is well accepted.
>>
>> ---
>>  arch/arm64/kvm/mmu.c | 42 ++++++++++++++++++++++++++++++++++++++----
>>  1 file changed, 38 insertions(+), 4 deletions(-)
>>
>> diff --git a/arch/arm64/kvm/mmu.c b/arch/arm64/kvm/mmu.c
>> index c59af5ca01b0..224aa15eb4d9 100644
>> --- a/arch/arm64/kvm/mmu.c
>> +++ b/arch/arm64/kvm/mmu.c
>> @@ -624,6 +624,36 @@ static void kvm_send_hwpoison_signal(unsigned long address, short lsb)
>>  	send_sig_mceerr(BUS_MCEERR_AR, (void __user *)address, lsb, current);
>>  }
>>  
>> +/*
>> + * Find a mapping size that properly insides the intersection of vma and
>> + * memslot. And hva and pa have the same alignment to this mapping size.
>> + * It's rough because there are still other restrictions, which will be
>> + * checked by the following fault_supports_stage2_huge_mapping().
> 
> I don't think these restrictions make complete sense to me. If this is
> a PFNMAP VMA, we should use the biggest mapping size that covers the
> VMA, and not more than the VMA.
But as described by kvm_arch_prepare_memory_region(), the memslot may not fully
cover the VMA. If that's true and we just consider the boundary of the VMA, our
block mapping may beyond the boundary of memslot. Is this a problem?

> 
>> + */
>> +static short device_rough_page_shift(struct kvm_memory_slot *memslot,
>> +				     struct vm_area_struct *vma,
>> +				     unsigned long hva)
>> +{
>> +	size_t size = memslot->npages * PAGE_SIZE;
>> +	hva_t sec_start = max(memslot->userspace_addr, vma->vm_start);
>> +	hva_t sec_end = min(memslot->userspace_addr + size, vma->vm_end);
>> +	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) >= sec_start &&
>> +	    ALIGN(hva, PUD_SIZE) <= sec_end)
>> +		return PUD_SHIFT;
>> +#endif
>> +
>> +	if ((hva & (PMD_SIZE - 1)) == (pa & (PMD_SIZE - 1)) &&
>> +	    ALIGN_DOWN(hva, PMD_SIZE) >= sec_start &&
>> +	    ALIGN(hva, PMD_SIZE) <= sec_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 +799,10 @@ 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 +813,12 @@ 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(memslot, vma, hva);
>>  	else
>>  		vma_shift = PAGE_SHIFT;
>>  
>> -	if (logging_active ||
>> -	    (vma->vm_flags & VM_PFNMAP)) {
>> +	if (logging_active) {
>>  		force_pte = true;
>>  		vma_shift = PAGE_SHIFT;
> 
> But why should we downgrade to page-size mappings if logging? This is
> a device, and you aren't moving the device around, are you? Or is your
> device actually memory with a device mapping that you are trying to
> migrate?
Thanks for the point. We should not move the device around, so we do not
need to consider logging when we build mapping for device.

I found that logging_active is per memslot and we're sure it's always false
for memslot with PFNMAP VMA, because the kvm_arch_prepare_memory_region()
forbids that. Then I think we're OK here.

Thanks,
Keqian

> 
>>  	}
>> @@ -855,7 +889,7 @@ 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);
>>  	} else if (logging_active && !write_fault) {
>>  		/*
>>  		 * Only actually map the page as writable if this was a write
>> -- 
>> 2.19.1
>>
>>
> 
> Thanks,
> 
> 	M.
> 

Re: [RFC PATCH v2 2/2] kvm/arm64: Try stage2 block mapping for host device MMIO

[Keqian Zhu]

Hi Marc,

I think I have fully tested this patch. The next step is to do some restriction on
HVA in vfio module, so we can build block mapping for it with a higher probability.

Is there anything to improve? If not, could you apply it? ^_^

Thanks,
Keqian

On 2021/4/7 21:18, Marc Zyngier wrote:
> On Tue, 16 Mar 2021 13:43:38 +0000,
> Keqian Zhu <zhukeqian1@huawei.com> 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@huawei.com>
>> ---
>>
>> Mainly for RFC, not fully tested. I will fully test it when the
>> code logic is well accepted.
>>
>> ---
>>  arch/arm64/kvm/mmu.c | 42 ++++++++++++++++++++++++++++++++++++++----
>>  1 file changed, 38 insertions(+), 4 deletions(-)
>>
>> diff --git a/arch/arm64/kvm/mmu.c b/arch/arm64/kvm/mmu.c
>> index c59af5ca01b0..224aa15eb4d9 100644
>> --- a/arch/arm64/kvm/mmu.c
>> +++ b/arch/arm64/kvm/mmu.c
>> @@ -624,6 +624,36 @@ static void kvm_send_hwpoison_signal(unsigned long address, short lsb)
>>  	send_sig_mceerr(BUS_MCEERR_AR, (void __user *)address, lsb, current);
>>  }
>>  
>> +/*
>> + * Find a mapping size that properly insides the intersection of vma and
>> + * memslot. And hva and pa have the same alignment to this mapping size.
>> + * It's rough because there are still other restrictions, which will be
>> + * checked by the following fault_supports_stage2_huge_mapping().
> 
> I don't think these restrictions make complete sense to me. If this is
> a PFNMAP VMA, we should use the biggest mapping size that covers the
> VMA, and not more than the VMA.
> 
>> + */
>> +static short device_rough_page_shift(struct kvm_memory_slot *memslot,
>> +				     struct vm_area_struct *vma,
>> +				     unsigned long hva)
>> +{
>> +	size_t size = memslot->npages * PAGE_SIZE;
>> +	hva_t sec_start = max(memslot->userspace_addr, vma->vm_start);
>> +	hva_t sec_end = min(memslot->userspace_addr + size, vma->vm_end);
>> +	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) >= sec_start &&
>> +	    ALIGN(hva, PUD_SIZE) <= sec_end)
>> +		return PUD_SHIFT;
>> +#endif
>> +
>> +	if ((hva & (PMD_SIZE - 1)) == (pa & (PMD_SIZE - 1)) &&
>> +	    ALIGN_DOWN(hva, PMD_SIZE) >= sec_start &&
>> +	    ALIGN(hva, PMD_SIZE) <= sec_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 +799,10 @@ 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 +813,12 @@ 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(memslot, vma, hva);
>>  	else
>>  		vma_shift = PAGE_SHIFT;
>>  
>> -	if (logging_active ||
>> -	    (vma->vm_flags & VM_PFNMAP)) {
>> +	if (logging_active) {
>>  		force_pte = true;
>>  		vma_shift = PAGE_SHIFT;
> 
> But why should we downgrade to page-size mappings if logging? This is
> a device, and you aren't moving the device around, are you? Or is your
> device actually memory with a device mapping that you are trying to
> migrate?
> 
>>  	}
>> @@ -855,7 +889,7 @@ 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);
>>  	} else if (logging_active && !write_fault) {
>>  		/*
>>  		 * Only actually map the page as writable if this was a write
>> -- 
>> 2.19.1
>>
>>
> 
> Thanks,
> 
> 	M.
> 

Re: [RFC PATCH v2 2/2] kvm/arm64: Try stage2 block mapping for host device MMIO

[Keqian Zhu]

Hi Marc,

On 2021/4/8 15:28, Keqian Zhu wrote:
> Hi Marc,
> 
> On 2021/4/7 21:18, Marc Zyngier wrote:
>> On Tue, 16 Mar 2021 13:43:38 +0000,
>> Keqian Zhu <zhukeqian1@huawei.com> wrote:
>>>
[...]

>>>  
>>> +/*
>>> + * Find a mapping size that properly insides the intersection of vma and
>>> + * memslot. And hva and pa have the same alignment to this mapping size.
>>> + * It's rough because there are still other restrictions, which will be
>>> + * checked by the following fault_supports_stage2_huge_mapping().
>>
>> I don't think these restrictions make complete sense to me. If this is
>> a PFNMAP VMA, we should use the biggest mapping size that covers the
>> VMA, and not more than the VMA.
> But as described by kvm_arch_prepare_memory_region(), the memslot may not fully
> cover the VMA. If that's true and we just consider the boundary of the VMA, our
> block mapping may beyond the boundary of memslot. Is this a problem?
emm... Sorry I missed something. The fault_supports_stage2_huge_mapping() will check
the boundary of memslot, so we don't need to check it here. I have send v3, please
check that.

BRs,
Keqian