Re: AMD64 dma_alloc_coherent crashes on non PCI device (was SATA open bugs)

[James Bottomley <James.Bottomley@SteelEye.com>]

On Thu, 2007-08-09 at 18:53 +0100, Alan Cox wrote:
> > Someone must ask that caller library to DMA to/from that device
> > in the first place. Whoever it is it is wrong. 
> 
> No I disagree.

I'm with Andi here ... you're fortunate that parisc has no working
IDE/SATA interface (or rather we have so few running boxes with PCMCIA),
but if it did this call sequence would explode on that platform for a
different reason.

> > Or perhaps you got the wrong device here? For ISA devices we 
> > traditionally used NULL. Or if you set up your own ISA devices
> > (which I can't see a reason for but there might be one I'm missing) 
> > at least give them a dma mask. Then it should probably work on x86-64.
> 
> The libata code currently (and this seems to work for all but x86-64)
> does the following if it is setting up a potentially DMA capable device
> 
> 	- Allocate a dma_coherent buffer

You cannot allocate a dma_coherent buffer without passing in the correct
underlying device ... on parisc we'd explode trying to find the iommu to
map through (some of our hw has more than one).

> 	- If it is refused then turn off DMA and use PIO
> 
> It has no idea whether a pcmcia, isa, platform or even PCI device happens
> to be DMA capable, and there are platforms with PCI but very limited DMA
> for example in the embedded space. In fact it has no idea this level
> whether it is working a PCI, ISA, PCMCIA, SBUS or some other bus device.
> It's supposed to be generic code.
> 
> Obviously doing anything other than dma_alloc_coherent if the allocation
> fails is stupid but trying to allocate dma memory to find out if the
> device can be used with DMA on a given platform is quite logical and
> sensible in some cases.
> 
> Thus I think dma_alloc_coherent() for an ISA, PCMCIA or other class
> (platform particularly) shouldn't explode on AMD64 but simply return
> NULL. Its a sane request to make when you don't in your library know what
> dev is.

PCMCIA should be correctly plumbed into the generic model.  It's really
just a bridge ... how that bridge is wired critically affects how
coherent memory is allocated.

James