[RFC] ARM: DMA-Mapping: add a new attribute to clear buffer

Message ID 1356656433-2278-1-git-send-email-daeinki@gmail.com
State New
Headers show

Commit Message

Inki Dae Dec. 28, 2012, 1 a.m.
From: Inki Dae <inki.dae@samsung.com>

This patch adds a new attribute, DMA_ATTR_SKIP_BUFFER_CLEAR
to skip buffer clearing. The buffer clearing also flushes CPU cache
so this operation has performance deterioration a little bit.

With this patch, allocated buffer region is cleared as default.
So if you want to skip the buffer clearing, just set this attribute.

But this flag should be used carefully because this use might get
access to some vulnerable content such as security data. So with this
patch, we make sure that all pages will be somehow cleared before
exposing to userspace.

For example, let's say that the security data had been stored
in some memory and freed without clearing it.
And then malicious process allocated the region though some buffer
allocator such as gem and ion without clearing it, and requested blit
operation with cleared another buffer though gpu or other drivers.
At this time, the malicious process could access the security data.

Signed-off-by: Inki Dae <inki.dae@samsung.com>
Signed-off-by: Kyungmin Park <kyungmin.park@samsung.com>
---
 arch/arm/mm/dma-mapping.c |    6 ++++--
 include/linux/dma-attrs.h |    1 +
 2 files changed, 5 insertions(+), 2 deletions(-)

Comments

Subash Patel Jan. 1, 2013, 1:15 a.m. | #1
On Thursday 27 December 2012 05:00 PM, daeinki@gmail.com wrote:
> From: Inki Dae <inki.dae@samsung.com>
>
> This patch adds a new attribute, DMA_ATTR_SKIP_BUFFER_CLEAR
> to skip buffer clearing. The buffer clearing also flushes CPU cache
> so this operation has performance deterioration a little bit.
>
> With this patch, allocated buffer region is cleared as default.
> So if you want to skip the buffer clearing, just set this attribute.
>
> But this flag should be used carefully because this use might get
> access to some vulnerable content such as security data. So with this
> patch, we make sure that all pages will be somehow cleared before
> exposing to userspace.
>
> For example, let's say that the security data had been stored
> in some memory and freed without clearing it.
> And then malicious process allocated the region though some buffer
> allocator such as gem and ion without clearing it, and requested blit
> operation with cleared another buffer though gpu or other drivers.
> At this time, the malicious process could access the security data.

Isnt it always good to use such security related buffers through TZ 
rather than trying to guard them in the non-secure zone?

>
> Signed-off-by: Inki Dae <inki.dae@samsung.com>
> Signed-off-by: Kyungmin Park <kyungmin.park@samsung.com>
> ---
>   arch/arm/mm/dma-mapping.c |    6 ++++--
>   include/linux/dma-attrs.h |    1 +
>   2 files changed, 5 insertions(+), 2 deletions(-)
>
> diff --git a/arch/arm/mm/dma-mapping.c b/arch/arm/mm/dma-mapping.c
> index 6b2fb87..fbe9dff 100644
> --- a/arch/arm/mm/dma-mapping.c
> +++ b/arch/arm/mm/dma-mapping.c
> @@ -1058,7 +1058,8 @@ static struct page **__iommu_alloc_buffer(struct device *dev, size_t size,
>   		if (!page)
>   			goto error;
>
> -		__dma_clear_buffer(page, size);
> +		if (!dma_get_attr(DMA_ATTR_SKIP_BUFFER_CLEAR, attrs))
> +			__dma_clear_buffer(page, size);
>
>   		for (i = 0; i < count; i++)
>   			pages[i] = page + i;
> @@ -1082,7 +1083,8 @@ static struct page **__iommu_alloc_buffer(struct device *dev, size_t size,
>   				pages[i + j] = pages[i] + j;
>   		}
>
> -		__dma_clear_buffer(pages[i], PAGE_SIZE << order);
> +		if (!dma_get_attr(DMA_ATTR_SKIP_BUFFER_CLEAR, attrs))
> +			__dma_clear_buffer(pages[i], PAGE_SIZE << order);
>   		i += 1 << order;
>   		count -= 1 << order;
>   	}
> diff --git a/include/linux/dma-attrs.h b/include/linux/dma-attrs.h
> index c8e1831..2592c05 100644
> --- a/include/linux/dma-attrs.h
> +++ b/include/linux/dma-attrs.h
> @@ -18,6 +18,7 @@ enum dma_attr {
>   	DMA_ATTR_NO_KERNEL_MAPPING,
>   	DMA_ATTR_SKIP_CPU_SYNC,
>   	DMA_ATTR_FORCE_CONTIGUOUS,
> +	DMA_ATTR_SKIP_BUFFER_CLEAR,
>   	DMA_ATTR_MAX,

How is this new macro different from SKIP_CPU_SYNC?

>   };
>
>

Regards,
Subash
Inki Dae Jan. 2, 2013, 1:55 a.m. | #2
2013/1/1 Subash Patel <subashrp@gmail.com>:
>
>
> On Thursday 27 December 2012 05:00 PM, daeinki@gmail.com wrote:
>>
>> From: Inki Dae <inki.dae@samsung.com>
>>
>> This patch adds a new attribute, DMA_ATTR_SKIP_BUFFER_CLEAR
>> to skip buffer clearing. The buffer clearing also flushes CPU cache
>> so this operation has performance deterioration a little bit.
>>
>> With this patch, allocated buffer region is cleared as default.
>> So if you want to skip the buffer clearing, just set this attribute.
>>
>> But this flag should be used carefully because this use might get
>> access to some vulnerable content such as security data. So with this
>> patch, we make sure that all pages will be somehow cleared before
>> exposing to userspace.
>>
>> For example, let's say that the security data had been stored
>> in some memory and freed without clearing it.
>> And then malicious process allocated the region though some buffer
>> allocator such as gem and ion without clearing it, and requested blit
>> operation with cleared another buffer though gpu or other drivers.
>> At this time, the malicious process could access the security data.
>
>
> Isnt it always good to use such security related buffers through TZ rather
> than trying to guard them in the non-secure zone?
>

This is for normal world. We should consider security issue to normal
world and also all cases as possible.

>
>>
>> Signed-off-by: Inki Dae <inki.dae@samsung.com>
>> Signed-off-by: Kyungmin Park <kyungmin.park@samsung.com>
>> ---
>>   arch/arm/mm/dma-mapping.c |    6 ++++--
>>   include/linux/dma-attrs.h |    1 +
>>   2 files changed, 5 insertions(+), 2 deletions(-)
>>
>> diff --git a/arch/arm/mm/dma-mapping.c b/arch/arm/mm/dma-mapping.c
>> index 6b2fb87..fbe9dff 100644
>> --- a/arch/arm/mm/dma-mapping.c
>> +++ b/arch/arm/mm/dma-mapping.c
>> @@ -1058,7 +1058,8 @@ static struct page **__iommu_alloc_buffer(struct
>> device *dev, size_t size,
>>                 if (!page)
>>                         goto error;
>>
>> -               __dma_clear_buffer(page, size);
>> +               if (!dma_get_attr(DMA_ATTR_SKIP_BUFFER_CLEAR, attrs))
>> +                       __dma_clear_buffer(page, size);
>>
>>                 for (i = 0; i < count; i++)
>>                         pages[i] = page + i;
>> @@ -1082,7 +1083,8 @@ static struct page **__iommu_alloc_buffer(struct
>> device *dev, size_t size,
>>                                 pages[i + j] = pages[i] + j;
>>                 }
>>
>> -               __dma_clear_buffer(pages[i], PAGE_SIZE << order);
>> +               if (!dma_get_attr(DMA_ATTR_SKIP_BUFFER_CLEAR, attrs))
>> +                       __dma_clear_buffer(pages[i], PAGE_SIZE << order);
>>                 i += 1 << order;
>>                 count -= 1 << order;
>>         }
>> diff --git a/include/linux/dma-attrs.h b/include/linux/dma-attrs.h
>> index c8e1831..2592c05 100644
>> --- a/include/linux/dma-attrs.h
>> +++ b/include/linux/dma-attrs.h
>> @@ -18,6 +18,7 @@ enum dma_attr {
>>         DMA_ATTR_NO_KERNEL_MAPPING,
>>         DMA_ATTR_SKIP_CPU_SYNC,
>>         DMA_ATTR_FORCE_CONTIGUOUS,
>> +       DMA_ATTR_SKIP_BUFFER_CLEAR,
>>         DMA_ATTR_MAX,
>
>
> How is this new macro different from SKIP_CPU_SYNC?
>

The purpose of this patch is to skip buffer clearing, not to skip
cache opeation.

>>   };
>>
>>
>
> Regards,
> Subash

Patch hide | download patch | download mbox

diff --git a/arch/arm/mm/dma-mapping.c b/arch/arm/mm/dma-mapping.c
index 6b2fb87..fbe9dff 100644
--- a/arch/arm/mm/dma-mapping.c
+++ b/arch/arm/mm/dma-mapping.c
@@ -1058,7 +1058,8 @@  static struct page **__iommu_alloc_buffer(struct device *dev, size_t size,
 		if (!page)
 			goto error;
 
-		__dma_clear_buffer(page, size);
+		if (!dma_get_attr(DMA_ATTR_SKIP_BUFFER_CLEAR, attrs))
+			__dma_clear_buffer(page, size);
 
 		for (i = 0; i < count; i++)
 			pages[i] = page + i;
@@ -1082,7 +1083,8 @@  static struct page **__iommu_alloc_buffer(struct device *dev, size_t size,
 				pages[i + j] = pages[i] + j;
 		}
 
-		__dma_clear_buffer(pages[i], PAGE_SIZE << order);
+		if (!dma_get_attr(DMA_ATTR_SKIP_BUFFER_CLEAR, attrs))
+			__dma_clear_buffer(pages[i], PAGE_SIZE << order);
 		i += 1 << order;
 		count -= 1 << order;
 	}
diff --git a/include/linux/dma-attrs.h b/include/linux/dma-attrs.h
index c8e1831..2592c05 100644
--- a/include/linux/dma-attrs.h
+++ b/include/linux/dma-attrs.h
@@ -18,6 +18,7 @@  enum dma_attr {
 	DMA_ATTR_NO_KERNEL_MAPPING,
 	DMA_ATTR_SKIP_CPU_SYNC,
 	DMA_ATTR_FORCE_CONTIGUOUS,
+	DMA_ATTR_SKIP_BUFFER_CLEAR,
 	DMA_ATTR_MAX,
 };