On 03.05.22 14:54, Juergen Gross wrote:
> On 28.04.22 16:50, Jan Beulich wrote:
>> The latest with commit bdd8b6c98239 ("drm/i915: replace X86_FEATURE_PAT
>> with pat_enabled()") pat_enabled() returning false (because of PAT
>> initialization being suppressed in the absence of MTRRs being announced
>> to be available) has become a problem: The i915 driver now fails to
>> initialize when running PV on Xen (i915_gem_object_pin_map() is where I
>> located the induced failure), and its error handling is flaky enough to
>> (at least sometimes) result in a hung system.
>>
>> Yet even beyond that problem the keying of the use of WC mappings to
>> pat_enabled() (see arch_can_pci_mmap_wc()) means that in particular
>> graphics frame buffer accesses would have been quite a bit less
>> performant than possible.
>>
>> Arrange for the function to return true in such environments, without
>> undermining the rest of PAT MSR management logic considering PAT to be
>> disabled: Specifically, no writes to the PAT MSR should occur.
>>
>> For the new boolean to live in .init.data, init_cache_modes() also needs
>> moving to .init.text (where it could/should have lived already before).
>>
>> Signed-off-by: Jan Beulich <jbeulich@suse.com>
> 
> I think this approach isn't the best way to tackle the issue.
> 
> It can be solved rather easily by not deriving the supported caching
> modes via pat_enabled(), but by adding specific functions to query
> the needed caching mode from the PAT translation tables, and to use
> those functions instead of pat_enabled().
> 
> I'm preparing a patch for that purpose.

That attempt was not a complete success.

Especially there are issues with my approach when "nopat" has been
specified as boot parameter, as that would be just ignored.

So right now I can't think of a better approach than the one of Jan's
patch.


Juergen