Re: pci-gart_64.c: could we get better explanation?

[Pavel Machek <pavel@suse.cz>]

On Tue 2008-05-27 05:51:39, Benjamin Herrenschmidt wrote:
> 
> On Mon, 2008-05-26 at 20:47 +0200, Pavel Machek wrote:
> > Hi!
> > 
> > +++ b/arch/x86/kernel/pci-gart_64.c
> > @@ -774,6 +778,7 @@ #endif
> >  
> >         /*
> >          * Try to workaround a bug (thanks to BenH)
> >          * Set unmapped entries to a scratch page instead of 0.
> >          * Any prefetches that hit unmapped entries won't get an bus abort
> > 
> > Could someone explain? Which bug is being worked around? CPUs
> > prefetching over NULL?
> 
> P2P bridge prefetching on DMA reads.

---

Add better explanation to pci-gart.

Signed-off-by: Pavel Machek <pavel@suse.cz>

diff --git a/arch/x86/kernel/pci-gart_64.c b/arch/x86/kernel/pci-gart_64.c
index 838a1d3..e72e583 100644
--- a/arch/x86/kernel/pci-gart_64.c
+++ b/arch/x86/kernel/pci-gart_64.c
@@ -800,10 +800,10 @@ #endif
 	wbinvd();
 
 	/*
-	 * Try to workaround a bug (thanks to BenH)
+	 * Try to workaround a bug (thanks to BenH):
 	 * Set unmapped entries to a scratch page instead of 0.
 	 * Any prefetches that hit unmapped entries won't get an bus abort
-	 * then.
+	 * then. (P2P bridge may be prefetching on DMA reads).
 	 */
 	scratch = get_zeroed_page(GFP_KERNEL);
 	if (!scratch)

-- 
(english) http://www.livejournal.com/~pavelmachek
(cesky, pictures) http://atrey.karlin.mff.cuni.cz/~pavel/picture/horses/blog.html