Re: [PATCH] x86: avoid unnecessary low zone allocation in AMDIOMMU's alloc_coherent

From: FUJITA Tomonori
Date: Wed Sep 10 2008 - 10:39:34 EST

On Wed, 10 Sep 2008 23:24:57 +0900
FUJITA Tomonori <fujita.tomonori@xxxxxxxxxxxxx> wrote:

> On Wed, 10 Sep 2008 15:53:47 +0200
> Joerg Roedel <joerg.roedel@xxxxxxx> wrote:
> > On Wed, Sep 10, 2008 at 10:37:45PM +0900, FUJITA Tomonori wrote:
> > > On Wed, 10 Sep 2008 15:10:32 +0200
> > > Joerg Roedel <joerg.roedel@xxxxxxx> wrote:
> > > > Can't we simply make the gfp hacks depend on
> > > > dma_ops->is_phys and avoid further gfp hacks in the hardware iommu
> > > > implementations?
> > >
> > > I thought about it but adding a new dma_ops->we_don't_want_gfp_flag
> > > hook doesn't make the code simpler much. Currently, we have the gfp
> > > setting hack in just one place. It's not bad. Adding such new hook
> > > means adding more lines than we can remove.
> >
> > The is_phys flas is already in place and its meaning is "the dma_ops
> > return bus addresses equal to physical addresses". This is exactly the
> > case when we need the gfp hacks. So I don't see a problem in just
> > skipping the gfp rewrite if is_phys is zero. I don't see a point in
> > adding gfp flags in dma_alloc_coherent and remove them again
> > dma_ops->alloc_coherent code. Specially in this case where we already
> > know in dma_alloc_coherent if we really need the flag rewrite.
> dma_ops->is_phys doesn't work well for GART and Intel IOMMU, that do
> virtual mappings for some devices and doesn't for some.
> We need to a hook that can pass a point to a device to IOMMUs like:
> dma_ops->is_phys(struct device *dev)
> Because they need to look at a device to know if they will do virtual
> mappings or not for it.

btw, in tip/x86/iommu, GART's alloc_coherent always does virtual
mappings to allocate a size-aligned memory (as DMA-mapping.txt

Because someone strongly insisted, I modified GART's alloc_coherent to
do so but as I said again and again, it's completely meaningless (only
POWER IOMMU does it and drivers don't depend on such requirement).

I guess that it would be better to do virtual mappings only when
necessary as the current mainline does since GART I/O space is
precious in some systems. But I don't care much. What's your opinion
(as a AMD developer)?
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at
Please read the FAQ at