Re: [PATCH 2/5] x86: PAT: fix ambiguous paranoia check in pat_init()

[Andreas Herrmann <andreas.herrmann3@amd.com>]

On Wed, Jun 11, 2008 at 02:58:49PM +0200, Rene Herman wrote:
> On 11-06-08 11:41, Andreas Herrmann wrote:
>
>> As I had no Transmeta or Centaur CPU at hand I just cleared the PAT
>> flag in the CPU identification code to simulate the case that all CPUs
>> of a Vendor are whitelisted (even those w/o PAT support).  The first
>> time pat_init() is entered you get
>>   PAT enabled, but CPU feature cleared     (=> which is wrong as no flag 
>> was cleared)
>
> Again, you are misreading this. Please just replace the message mentally by 
> "PAT enabled, but CPU claims to not support PAT". "cleared" here does not 
> signify that we ourselves cleared anything, just that flag IS clear 
> (unset). Yes, maybe the wording could be better but it's not wrong.

Well, wording might not be best. But I don't care anymore.
(Just wondering which CPUs are out there that support PAT but don't
advertise it with any feature flag.)

>>   x86 PAT enabled: cpu 0, old 0x7040600070406, new 0x7010600070106
>>     (=> which is wrong as PAT shouldn't be enabled on such CPUs)
>
> Again not wrong, or at least by design. Thomas Gleixner did it this way and 
> with that "paranoia check" explicitly bombing out only for SMP this 
> wouldn't have been by accident. He no doubt knows why he did so (and he's 
> in CC so if we're real lucky we might also now...)

I guess at the time Thomas' patch was commited this was just fine.

But with the recent Transmeta/Centaur patch, validate_pat_support()
returns w/o disabling PAT even for such vendor's CPUs that don't
support PAT,

To prevent this, validate_pat_support() should check for cpu_has_pat
in addition to any other white-black-or-whatsoever-listing.


Regards,

Andreas