Re: [PATCH 0/2] AMD ARAT fixes

[Chuck Ebbert <cebbert@redhat.com>]

On Wed, 18 May 2011 17:50:17 +0200
Borislav Petkov <bp@amd64.org> wrote:

> Hi Greg,
> 
> Ingo just confirmed that the following two fixes went upstream. I
> haven't tagged them for stable so I'd appreciate if you could take them
> for the next cycle. AFAICT, the relevant trees should be .38-stable,
> 32-longterm and 33-longterm.
> 
> There should be no problem cherry-picking them but if there is, please
> let me know and I'll give you rebased versions.
> 
> Here the commit ids again, for reference:
> 
> http://git.kernel.org/tip/14fb57dccb6e1defe9f89a66f548fcb24c374c1d
> http://git.kernel.org/tip/328935e6348c6a7cb34798a68c326f4b8372e68a
> 

This still leaves family 10h model 6 stepping 2 (and possibly others)
broken in -stable as well as 2.6.39.

Looking at -stable, this whole mess was caused by:

  commit b87cf80af3ba4b4c008b4face3c68d604e1715c6
  x86, AMD: Set ARAT feature on AMD processors

That caused stalls on family 0fh and family 10h processors, and then
the (partial) fix for that in 2.6.38.6:

  commit e20a2d205c05cef6b5783df339a7d54adeb50962
  x86, AMD: Fix APIC timer erratum 400 affecting K8 Rev.A-E processors

caused instant crashes on boot on older family 0fh processors.

Now it looks like family 0fh is finally fixed in 2.6.38.7.

But I can't find any reason for the original commit that went in 2.6.38.4
to be there in the first place. It doesn't fix any bug whatsoever and
appears to be just a performance enhancement. So how did it get there?

I came up with this (untested) hack for now to fix the remaining bug,
should something like this go in -stable to fix family 10h until a
better way is found?

--- a/arch/x86/kernel/cpu/amd.c
+++ b/arch/x86/kernel/cpu/amd.c
@@ -724,6 +724,15 @@ bool cpu_has_amd_erratum(const int *erra
 		return false;
 
 	/*
+	 * Temporary workaround for ARAT bug on Sempron.
+	 * The BIOS clears the bit in OSVW, so the check
+	 * fails, then ARAT gets set and when the processor
+	 * uses C3 it hangs. Always return true for that CPU.
+	 */
+	if (cpu->x86 == 0x10 && cpu->x86_model == 6 && cpu->x86_mask == 2)
+		return true;
+
+	/*
 	 * Must match family-model-stepping range first so that the
 	 * range checks will override OSVW checking.
 	 */