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

[Joerg Roedel <joerg.roedel@amd.com>]

On Thu, Sep 11, 2008 at 02:15:22AM +0900, FUJITA Tomonori wrote:
> On Wed, 10 Sep 2008 19:05:04 +0200
> Joerg Roedel <joerg.roedel@amd.com> wrote:
> 
> > On Thu, Sep 11, 2008 at 01:29:54AM +0900, FUJITA Tomonori wrote:
> > > On Wed, 10 Sep 2008 17:29:18 +0200
> > > Joerg Roedel <joerg.roedel@amd.com> wrote:
> > > 
> > > > What hacks do you need with __get_free_pages? The memory it returns is
> > > > _always_ aligned at its size.
> > > 
> > > Is it guaranteed (documented somewhere) ?
> > 
> > I don't know if there is a formal definition for it. It is documented in
> > some books about the Linux kernel (I read this in some book the first
> > time). This alignment results from the buddy algorithm the page alloctor
> > uses. You can definitly rely on that.
> 
> I meant, if it's not documented as a guaranteed feature (not just the
> characteristic of the current code), it could change any time.

It will never change. I am sure about this. A lot of code relys on
this like the HugeTLBfs implementations for example. In fact, it may be
possible that the requirement to the dma_alloc_coherent function derive
from this behavior of the Linux page allocator (to not break anything
when the dma api was introduced).

Joerg

-- 
           |           AMD Saxony Limited Liability Company & Co. KG
 Operating |         Wilschdorfer Landstr. 101, 01109 Dresden, Germany
 System    |                  Register Court Dresden: HRA 4896
 Research  |              General Partner authorized to represent:
 Center    |             AMD Saxony LLC (Wilmington, Delaware, US)
           | General Manager of AMD Saxony LLC: Dr. Hans-R. Deppe, Thomas McCoy