Re: [PATCH] intel-iommu: Fix alloc_coherent for pass-through devices

[Alex Williamson <alex.williamson@hp.com>]

On Tue, Nov 3, 2009 at 4:57 PM, Andrew Morton <akpm@linux-foundation.org> wrote:
> On Fri, 23 Oct 2009 00:49:01 +0900
> David Woodhouse <dwmw2@infradead.org> wrote:
>> diff --git a/drivers/pci/intel-iommu.c b/drivers/pci/intel-iommu.c
>> index b1e97e6..773a662 100644
>> --- a/drivers/pci/intel-iommu.c
>> +++ b/drivers/pci/intel-iommu.c
>> @@ -2765,6 +2765,10 @@ static void *intel_alloc_coherent(struct device *hwdev, size_t size,
>>       void *vaddr;
>>       int order;
>>
>> +     if (iommu_no_mapping(hwdev))
>> +             return dma_generic_alloc_coherent(hwdev, size, dma_handle,
>> +                                               flags);
>> +
>>       size = PAGE_ALIGN(size);
>>       order = get_order(size);
>>       flags &= ~(GFP_DMA | GFP_DMA32);
>>
>> (That won't build on IA64)
>>
>
> This is a box-killing post-2.6.31 regression, yes?
>
> Alex appears to have disappeared on us, making it rather hard to fix
> this up promptly.

Nope, I'm still here.  David and I chatted about some of the issues
and I sent out a series of patches last week to address them.  See
this thread:

https://lists.linux-foundation.org/pipermail/iommu/2009-October/001851.html

The first patch, attempting to implement the above was a non-starter
and dropped.  I believe the other patches in the series are valid and
I'd certainly like to see at least the alloc_coherent part fixed for
.32.  I'll be happy to send out a revised set if there's any confusion
of where things stand.  Thanks,

Alex