Re: [PATCH v6 00/16] x86/mtrr: fix handling with PAT but without MTRR

[Borislav Petkov]

On Wed, May 10, 2023 at 01:36:41AM +0200, Borislav Petkov wrote:
> More staring at this tomorrow, on a clear head.

Yeah, I'm going to leave it as is. Tried doing a union with bitfields
but doesn't get any prettier.

Next crapola:

The Intel box says now:

[    8.138683] sgx: EPC section 0x80200000-0x85ffffff
[    8.204838] pmd_set_huge: Cannot satisfy [mem 0x80200000-0x80400000] with a huge-page mapping due to MTRR override, uniform: 0

(I've extended the debug output).

and that happens because

[    8.174229] mtrr_type_lookup: mtrr_state_set: 1
[    8.178909] mtrr_type_lookup: start: 0x80200000, cache_map[3].start: 0x88800000

that's

	 if (start < cache_map[i].start) {

in mtrr_type_lookup(). I fail to see how that check would work for the
range 0x80200000-0x80400000 and the MTRR map is:

[    0.000587] MTRR map: 4 entries (3 fixed + 1 variable; max 23), built from 10 variable MTRRs
[    0.000588]   0: 0000000000000000-000000000009ffff write-back
[    0.000589]   1: 00000000000a0000-00000000000bffff uncachable
[    0.000590]   2: 00000000000c0000-00000000000fffff write-protect
[    0.000591]   3: 0000000088800000-00000000ffffffff uncachable

so the UC range comes after this one we request.

[    8.186372] mtrr_type_lookup: type: 0x6, cache_map[3].type: 0x0

now the next type merging happens and the 3rd region's type is UC, ofc.

[    8.192433] type_merge: type: 0x6, new_type: 0x0, effective_type: 0x0, clear uniform

we clear uniform and we fail:

[    8.200331] mtrr_type_lookup: ret, uniform: 0

So this map lookup thing is wrong in this case.

-- 
Regards/Gruss,
    Boris.

https://people.kernel.org/tglx/notes-about-netiquette