Re: [PATCH]: PCI: GART iommu alignment fixes [v2]

[Prarit Bhargava <prarit@redhat.com>]

Muli Ben-Yehuda wrote:
> On Thu, Aug 07, 2008 at 01:41:40PM -0400, Prarit Bhargava wrote:
>
>   
>> As for Calgary, I'm looking into it ATM.  I think I can get my hands
>> on one.
>>     
>
> Feel free to ping me if Calgary testing is needed.
>   

Muli -- I just ran tests on an IBM system with a Calgary iommu that Ed 
Pollard pointed me at.

dma_ops_alloc_addresses() does not have the option to return 
size-aligned values.  This means that 
pci_alloc_consistent()/dma_alloc_coherent() will return unaligned values 
to callers when the lower 4G of memory not available.

Additionally, a quick test shows that in dma_ops_alloc_addresses()

        boundary_size = ALIGN(dma_get_seg_boundary(dev) + 1,
                        PAGE_SIZE) >> PAGE_SHIFT;

may return 0 in the same manner I've been pointing out -- if 
dma_get_seg_boundary(dev) returns 0xffffffff and 1 is added to that 
result, boundary_size = 0.  Then you BUG() in the iommu-helper code.

Jesse pointed out to me that my fix on that line is incorrect.  _If_ 
this is not a compiler issue (I've emailed jakub privately and cc'd him 
on this email) then a better fix would be to do (sorry for the 
cut-and-paste):

--- a/include/linux/dma-mapping.h
+++ b/include/linux/dma-mapping.h
@@ -78,7 +78,7 @@ static inline unsigned int dma_set_max_seg_size(struct 
device
 static inline unsigned long dma_get_seg_boundary(struct device *dev)
 {
        return dev->dma_parms ?
-               dev->dma_parms->segment_boundary_mask : 0xffffffff;
+               dev->dma_parms->segment_boundary_mask : 0xffffffffUL;
 }
 
However, I'm still waiting for clarification from jakub before 
submitting again with that chunk.

P.


> Cheers,
> Muli
>