Re: Proper way to check for restricted DMA addressing from device driver

[Robin Murphy]

On 26/02/2020 4:04 pm, Lucas Stach wrote:
> On Mi, 2020-02-26 at 15:51 +0000, Robin Murphy wrote:
> > On 26/02/2020 3:44 pm, Lucas Stach wrote:
> > > Hi all,
> > >
> > > I'm currently struggling with how to properly check for restricted DMA
> > > addressing from a device driver side. The basic issue I'm facing is
> > > that I have a embedded GPU, which isn't able to address all system
> > > memory due to interconnect being restricted to 32bit addressing. The
> > > limits are properly described in the system device-tree and thus
> > > SWIOTLB is working.
> > >
> > > However graphics buffers are large and graphics drivers really like to
> > > keep the dma mapping alive for performance reasons, which means I'm
> > > running out of SWIOTLB space pretty easily, aside from the obvious
> > > performance implications of SWIOTLB.
> > >
> > > As 3 out of the maximum 4GB system memory are located in the DMA32 zone
> > > and thus located in the GPU addressable space, I just want to avoid
> > > allocating graphics buffers outside of the DMA32 zone.
> > >
> > > To add the DMA32 restriction to my drivers allocations, I need a
> > > reliable way from the device driver side to check if the GPU is in such
> > > a restricted system. What I'm currently doing in my WIP patch is this:
> > >
> > >    /*
> > >     * If the GPU is part of a system with only 32bit bus addressing
> > >     * capabilities, request pages for our SHM backend buffers from the
> > >     * DMA32 zone to avoid performance killing SWIOTLB bounce buffering.
> > >     */
> > >    if (*gpu->dev->dma_mask < BIT_ULL(32) && !device_iommu_mapped(gpu->dev))
> > >            priv->shm_gfp_mask |= GFP_DMA32;
> > >
> > > However I'm not sure if there are edge cases where this check would
> > > fool me. Is there any better way to check for DMA addressing
> > > restrictions from the device driver side?
> >
> > dma_addressing_limited()?
>
> While amdgpu and radeon do seem to use this to trigger DMA32
> allocations, the bool return value doesn't really tell if DMA32 is
> helpful or not. All it tells is that the system has memory outside of
> the device dma addressing capabilities.

If the limit falls somewhere between 32 bits and "everything" then DMA32 
is the best you can do anyway; if it's *less* than 32 bits then you'd be 
into the realm of "serious problems" already. There is a point at which 
it's reasonable to assume you don't have to support contrived 
pathological cases :)
> Is the return value of this function correct if the device is behind a
> IOMMU? While the device might be on a bus that is address limited a
> IOMMU further up the path to memory might allow to address all system
> memory, no?

It should be OK - see the comment in dma_get_required_mask(). Obviously 
it's theoretically possible to integrate a device with 32+ bits of DMA 
capability to an IOMMU input 32+ bits wide, via an interconnect less 
than 32 bits wide, but any designer silly enough to actually do that is 
welcome to have their SoC never supported by Linux.

Robin.