Not coherent cache DMA for G3/G4 CPUs: clarification needed

Gabriel Paubert paubert at iram.es
Fri Apr 21 09:46:40 EST 2006


On Fri, Apr 21, 2006 at 08:39:29AM +1000, Benjamin Herrenschmidt wrote:
> On Thu, 2006-04-20 at 14:13 -0700, Eugene Surovegin wrote:
> > On Fri, Apr 21, 2006 at 07:06:13AM +1000, Benjamin Herrenschmidt wrote:
> > > Unfortunately, he has to do things a bit differently. He can't afford to
> > > have the kernel BAT mapping cover his non-cacheable pages. Thus he needs
> > > a reserved pool. Last I looked at our coherent code, it didn't reserve
> > > memory at all, just address space, thus assuming the CPU can handle
> > > having both a caheable and a non-cacheable mapping of the same pages...
> > > (On 6xx this is deadly even if you don't access those cacheable pages
> > > because the CPU prefetch may do it for you).
> > 
> > Ben, is this _real_ problem on 6xx or just a theory? Does 6xx actually 
> > prefetch beyond page boundary?
> > 
> > So far, all "prefetching" I saw which broke non-coherent DMA was not 
> > due to the CPU doing prefetching, but _software_ prefetching being 
> > too aggressive.
> 
> Not 100% certain... we definitely have a bug with AGP on macs currently
> for that reason though I yet have to isolate a crash caused by it ;)
> (That is, we map AGP pages non-cacheable but they stay in the linear
> mapping).

In this case the problem is double mapping with inconsistent attributes
(through BAT and page tables I assume). 

> 
> On POWER4, 970 and later, the chip guys confirmed that the problem is
> real though. Not only bcs of prefetch but also speculative execution
> which can cause the chip to do a load that will never actually be
> executed. Imagine for example a loop walking an array, the chip might
> speculatively load elements beyond the array by speculatively executing
> beyond the branch that ends the loop.

Even if the page has the guarded bit set?

	Gabriel



More information about the Linuxppc-dev mailing list