Re: [PATCH 1/2] x86/mm/pat: Change pat_disable() to emulate PAT table

[Toshi Kani <toshi.kani@hpe.com>]

On Sat, 2016-03-12 at 12:55 +0100, Borislav Petkov wrote:
> Ok, let's start again.
> 
> So what's wrong with this simpler, cleaner version below?
> 
> * It makes sure to run the rendezvous handler on the BP on init.
> 
> * It disables PAT otherwise.
> 
> * Oh, and it boots fine with Paul's reproducer, X is there (even though
> vncviewer dies when X starts with "Rect too large: 640x480 at (0, 0)"
> and I have to reconnect to see the X window).
> 
> In dmesg I see "x86/PAT: Disabled by MTRR".

Your patch is a simplified version of mine.  So, yes, it fixes the Paul's
issue, but it does not address other issues that my patchset also
addressed.  In specific, I think your patch has the following issues.

- pat_disable() is now callable from other modules. So, it needs to check
with boot_cpu_done. We cannot disable PAT once it is initialized.

- mtrr_bp_init() needs to check with mtrr_enabled() when it
calls mtrr_pat_setup_bp(). Otherwise, PAT is left initialized on BSP only
when MTRR is disabled by its MSR. In your patch, mtrr_bp_init() calls
pat_setup() again, but it does not help since boot_cpu_done is set.

- When PAT is disabled in CPU feature, pat_bsp_init() calls pat_disable()
and returns. However, it does not initialize a PAT table by calling
pat_init_cache_modes().

- When CONFIG_MTRR is unset, it does not call pat_setup().

Thanks,
-Toshi