Re: [PATCH] Mark i2o config broken on 64-bit platforms.

[Miquel van Smoorenburg <miquels@cistron.nl>]

On Wed, 2008-07-09 at 14:59 -0500, James Bottomley wrote:
> On Wed, 2008-07-09 at 21:22 +0200, Miquel van Smoorenburg wrote:
> > On Wed, 2008-07-09 at 14:13 -0500, James Bottomley wrote:
> > > On Wed, 2008-07-09 at 21:06 +0200, Miquel van Smoorenburg wrote:
> > > > On Wed, 2008-07-09 at 16:49 +0100, Alan Cox wrote:
> > > > > On Wed, 09 Jul 2008 15:15:14 +0100
> > > > > David Howells <dhowells@redhat.com> wrote:
> > > > > 
> > > > > > Miquel van Smoorenburg <miquels@cistron.nl> wrote:
> > > > > > Maybe, but have you looked at i2o_cfg_passthru()?  Take this, for example:
> > > > > > 
> > > > > > 			/* Allocate memory for the transfer */
> > > > > > 			p = kmalloc(sg_size, GFP_KERNEL);
> > > > > > 			...
> > > > > > 			//TODO 64bit fix
> > > > > > 			sg[i].addr_bus = virt_to_bus(p);
> > > > > > 
> > > > > > That looks distinctly dodgy.  virt_to_bus() returns a 64-bit address, and as
> > > > > 
> > > > > Agreed - stick | GFP_DMA32 on the end then ;)
> > > > 
> > > > GFP_DMA32 doesn't work with kmalloc(), you need dma_alloc_coherent() or
> > > > pci_alloc_consistent() [here, i2o_dma_alloc() ]
> > > 
> > > Yes it does ... it was specifically designed for it.  GFP_DMA32 was
> > > introduced to allow this type of thing to happen (in the old days most
> > > drivers were allowed to assume kmalloc would return memory whose
> > > physical address was < 4GB; GFP_DMA32 allows that to continue while
> > > allowing kmalloc to stray beyond 4GB physical).
> > 
> > If you use alloc_pages(), yes. But not for kmalloc(). There are no
> > general GFP_DMA32 slabs.
> 
> No ... it's platform specific.  Platforms whose ZONE_NORMAL covers only
> up to 4GB need do nothing.  However, x86_64 definitely implements
> ZONE_DMA32 for precisely this.  Several other platforms (like ia64)
> should but don't.

I tried  using it before (for dpt_i2o.c) and it didn't work.

I re-read quite a bit of kernel code before I posted this, and I could
not find GFP_DMA32 slabs. There are also no calls to kmalloc(size,
GFP_...|GFP_DMA32) in the recent kernels I checked. And google tells me
things like
http://lkml.org/lkml/2007/5/17/267

So I don't think GFP_DMA32 works with kmalloc. GFP_DMA does though.

Mike.