[BISECTED REGRESSION] b43legacy broken on G4 PowerBook

[Benjamin Herrenschmidt]

On Tue, 2019-06-11 at 20:52 -0500, Larry Finger wrote:
> On 6/11/19 5:46 PM, Benjamin Herrenschmidt wrote:
> > On Tue, 2019-06-11 at 17:20 -0500, Larry Finger wrote:
> > > b43-pci-bridge 0001:11:00.0: dma_direct_supported: failed (mask =
> > > 0x3fffffff,
> > > min_mask = 0x5ffff000/0x5ffff000, dma bits = 0x1f
> > 
> > Ugh ? A mask with holes in it ? That's very wrong... That min_mask is
> > bogus.
> 
> I agree, but that is not likely serious as most systems will have enough memory 
> that the max_pfn term will be much larger than the initial min_mask, and 
> min_mask will be unchanged by the min function. 

Well no... it's too much memory that is the problem. If min_mask is
bogus though it will cause problem later too, so one should look into
it.

> In addition, min_mask is not 
> used beyond this routine, and then only to decide if direct dma is supported. 
> The following patch generates masks with no holes, but I cannot see that it is 
> needed.

The right fix is to round up max_pfn to a power of 2, something like

min_mask = min_t(u64, min_mask, (roundup_pow_of_two(max_pfn - 1)) <<
		PAGE_SHIFT) 

> diff --git a/kernel/dma/direct.c b/kernel/dma/direct.c
> index 2c2772e9702a..e3edd4f29e80 100644
> --- a/kernel/dma/direct.c
> +++ b/kernel/dma/direct.c
> @@ -384,7 +384,8 @@ int dma_direct_supported(struct device *dev, u64 mask)
>          else
>                  min_mask = DMA_BIT_MASK(32);
> 
> -       min_mask = min_t(u64, min_mask, (max_pfn - 1) << PAGE_SHIFT);
> +       min_mask = min_t(u64, min_mask, ((max_pfn - 1) << PAGE_SHIFT) |
> +                                        DMA_BIT_MASK(PAGE_SHIFT));
> 
>          /*
>           * This check needs to be against the actual bit mask value, so
> 
> 
> Larry