Re: USB transfer_buffer allocations on 64bit systems

[Daniel Mack <daniel@caiaq.de>]

On Mon, Apr 12, 2010 at 10:59:57AM +0200, Andi Kleen wrote:
> Alan Stern <stern@rowland.harvard.edu> writes:
> >> 
> >> The fix is to use usb_buffer_alloc() for that purpose which ensures
> >> memory that is suitable for DMA. And on x86_64, this also means that the
> >> upper 32 bits of the address returned are all 0's.
> >
> > That is not a good fix.  usb_buffer_alloc() provides coherent memory, 
> > which is not what we want.  I believe the correct fix is to specify the 
> > GFP_DMA32 flag in the kzalloc() call.
> 
> The traditional way to handle this is to leave it to swiotlb in
> pci_map_*. pci_map_* is needed anyways if you run with a IOMMU.
> 
> Also note at least on x86 systems coherent memory is the same as non coherent
> memory. And GFP_DMA32 is a x86 specific flag, doesn't necessarily 
> do any good anywhere else.
> 
> So if you add x86isms anyways you could as well use dma_alloc_coherent()
> directly which is actually better at this than a simple GFP_DMA32
> and as a bonus handles the IOMMUs correctly too.

Which is exactly what usb_buffer_alloc() does already. So at least for
x86 you say this is the right thing to do? However, we don't necessarily
need coherent memory on other platforms, which is why I hessitate to
enforce that type of memory for all transfer_buffer allocations.

> Or just use GFP_KERNEL and pci_map_* later.

The USB core does this already, but at least on Pedro's machine, this
seems unsufficient. Unfortunately, I can't reproduce the issue yet, even
with more than 4GB of RAM installed.

Daniel