Re: [PATCH] swiotlb: sync buffer when mapping FROM_DEVICE

[Robin Murphy]

On 23/05/2019 17:43, Christoph Hellwig wrote:
> On Thu, May 23, 2019 at 07:35:07AM +0200, Marek Szyprowski wrote:
> > Don't we have DMA_BIDIRECTIONAL for such case?
>
> Not sure if it was intended for that case, but it definitively should
> do the right thing for swiotlb, and it should also do the right thing
> in terms of cache maintainance.
>
> > Maybe we should update
> > documentation a bit to point that DMA_FROM_DEVICE expects the whole
> > buffer to be filled by the device?
>
> Probably. Horia, can you try to use DMA_BIDIRECTIONAL?

Yes, in general that should be a viable option. I got rather focused on 
the distinction that a "partial" FROM_DEVICE mapping would still be 
allowed to physically prevent the device from making any reads, whereas 
BIDIRECTIONAL would not, but I suspect any benefit being lost there is 
mostly one of debugging visibility rather than appreciable security, and 
probably not enough to justify additional complexity on its own - I 
couldn't say off-hand how many IOMMUs actually support write-only 
permissions anyway.

Whichever way, I'd certainly have no objection to formalising what seems 
to be the existing behaviour (both SWIOTLB and ARM dmabounce look 
consistent, at least) as something like "for the DMA_FROM_DEVICE 
direction, any parts of the buffer not written to by the device will 
become undefined". The IOMMU bounce page stuff is going to be another 
one in this boat, too.

Robin.