Software bounce buffers Vs allocating DMA buffers in page granularity

Software bounce buffers Vs allocating DMA buffers in page granularity

[Chathura Rajapaksha]

Hi All,

I have a question regarding the IOMMU software bounce buffer feature.
I understand that this feature was introduced as a mitigation to
sub-page vulnerabilities (which were shown to be exploitable by
Thunderclap).

My question is:
Can allocating DMA buffers in page granularity (multiples of 4KB pages)
be a feasible alternative solution to software bounce buffers?

I am trying to understand the challenges and performance impact of
such a solution (compared to SW bounce buffers) if someone were to
proposed it.
In other words, I am trying to find the reasons why SW bounce buffers
were introduced rather than allocating DMA buffers in page granularity.

Thanks in advance.
Regards,
Chathura

Re: Software bounce buffers Vs allocating DMA buffers in page granularity

[Baolu Lu]

On 2023/8/18 4:17, Chathura Rajapaksha wrote:
> Hi All,
> 
> I have a question regarding the IOMMU software bounce buffer feature.
> I understand that this feature was introduced as a mitigation to
> sub-page vulnerabilities (which were shown to be exploitable by
> Thunderclap).
> 
> My question is:
> Can allocating DMA buffers in page granularity (multiples of 4KB pages)
> be a feasible alternative solution to software bounce buffers?

Yes. I remember that we skip software bounce buffer if the buffer size
is one or multiple PAGE_SIZE(s). Any change during recent evolution?

Best regards,
baolu

Re: Software bounce buffers Vs allocating DMA buffers in page granularity

[Robin Murphy]

On 2023-08-17 21:17, Chathura Rajapaksha wrote:
> Hi All,
> 
> I have a question regarding the IOMMU software bounce buffer feature.
> I understand that this feature was introduced as a mitigation to
> sub-page vulnerabilities (which were shown to be exploitable by
> Thunderclap).
> 
> My question is:
> Can allocating DMA buffers in page granularity (multiples of 4KB pages)
> be a feasible alternative solution to software bounce buffers?
> 
> I am trying to understand the challenges and performance impact of
> such a solution (compared to SW bounce buffers) if someone were to
> proposed it.
> In other words, I am trying to find the reasons why SW bounce buffers
> were introduced rather than allocating DMA buffers in page granularity.

Dedicated DMA buffers (dma_alloc_*()) *are* allocated at page 
granularity. SWIOTLB is used by the streaming DMA API (dma_map_*()), 
where the data already exists somewhere in memory, outside our control. 
If a streaming mapping request does happen to be aligned to the IOMMU 
page size already, it isn't bounced.

Robin.