Re: [RFC/PATCH] remove unneeded check in bcm43xx

[Benjamin Herrenschmidt <benh@kernel.crashing.org>]

> Yes, I know they hit the message, that's from a message in some forum
> that i got interested in the issue. It probably comes from an allocation
> from:
> http://www.linux-m32r.org/lxr/http/source/arch/powerpc/kernel/pci_direct_iommu.c#L32
> 
> Either the ppc code is wrong (it doesn't enforce dma_mask) either the
> driver still works without the check.
> 
> Maybe ppc should do the same thing as i386:
> 
> 47         if (dev == NULL || (dev->coherent_dma_mask < 0xffffffff))
> 48                 gfp |= GFP_DMA;

PPC doesn't have a ZONE_DMA... PPC assumes that the whole memory is
DMA'able... What happens here is that it works by "chance" on x86 ;)
ZONE_DMA is a thing of the past that represents ISA DMA (below 16M iirc
or something like that). Thus x86 historically maintained a separate
allocation zone down there for ISA devices. What happens here is that
the x86 code sort-of detects that the DMA mask isn't the full 32 bits
and kicks everything down into ZONE_DMA. Makes things work, though the
pressure on ZONE_DMA can often be high enough that you'll have a lot of
failed allocations ...

PPC doesn't have ZONE_DMA, thus can't provide memory guaranteed to be
below that limit.

Now, for ppc32, it should still sort-of work because all of lowmem is
below 1Gb and people generally don't hack their lowmem size (well, I do
but heh, that doesn't count :) and I don't think you'll get skb's in
highmem. But ppc64 hits the problem and at this point, there is nothing
I can do other than either implementing a split zone allocation mecanism
in the ppc64 architecture for the sole sake of bcm43xx (ick !) or doing
some trick with the iommu...

Ben.


-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/