Re: [PATCH 3/3] x86/MCE/AMD: Get address from already initialized block

[Borislav Petkov]

On Thu, May 17, 2018 at 08:49:31AM +0200, Johannes Hirte wrote:
> Maybe I'm missing something, but those RDMSR IPSs don't happen on
> pre-SMCA systems, right? So the caching should be avoided here, cause
> the whole lookup looks more expensive to me than the simple switch-block
> in get_block_address.

Yeah, and we should simply cache all those MSR values as I suggested then.

The patch at the end should fix your issue.

Yazen, so I'm working on the assumption that all addresses are the same
on every core - I dumped them and it looks like this:

    128 bank: 00, block: 0 : 0xc0002003
    128 bank: 00, block: 1 : 0x0
    128 bank: 01, block: 0 : 0xc0002013
    128 bank: 01, block: 1 : 0x0
    128 bank: 02, block: 0 : 0xc0002023
    128 bank: 02, block: 1 : 0x0
    128 bank: 03, block: 0 : 0xc0002033
    128 bank: 03, block: 1 : 0x0
    128 bank: 04, block: 0 : 0x0
    128 bank: 05, block: 0 : 0xc0002053
    128 bank: 05, block: 1 : 0x0
    128 bank: 06, block: 0 : 0xc0002063
    128 bank: 06, block: 1 : 0x0
    128 bank: 07, block: 0 : 0xc0002073
    128 bank: 07, block: 1 : 0x0
    128 bank: 08, block: 0 : 0xc0002083
    128 bank: 08, block: 1 : 0x0
    128 bank: 09, block: 0 : 0xc0002093
    128 bank: 09, block: 1 : 0x0
    128 bank: 10, block: 0 : 0xc00020a3
    128 bank: 10, block: 1 : 0x0
    128 bank: 11, block: 0 : 0xc00020b3
    128 bank: 11, block: 1 : 0x0
    128 bank: 12, block: 0 : 0xc00020c3
    128 bank: 12, block: 1 : 0x0
    128 bank: 13, block: 0 : 0xc00020d3
    128 bank: 13, block: 1 : 0x0
    128 bank: 14, block: 0 : 0xc00020e3
    128 bank: 14, block: 1 : 0x0
    128 bank: 15, block: 0 : 0xc00020f3
    128 bank: 15, block: 1 : 0x0
    128 bank: 16, block: 0 : 0xc0002103
    128 bank: 16, block: 1 : 0x0
    128 bank: 17, block: 0 : 0xc0002113
    128 bank: 17, block: 1 : 0x0
    128 bank: 18, block: 0 : 0xc0002123
    128 bank: 18, block: 1 : 0x0
    128 bank: 19, block: 0 : 0xc0002133
    128 bank: 19, block: 1 : 0x0
    128 bank: 20, block: 0 : 0xc0002143
    128 bank: 20, block: 1 : 0x0
    128 bank: 21, block: 0 : 0xc0002153
    128 bank: 21, block: 1 : 0x0
    128 bank: 22, block: 0 : 0xc0002163
    128 bank: 22, block: 1 : 0x0

so you have 128 times the same address on a 128 core Zen box.

If that is correct, then we can use this nice simplification and cache
them globally and not per-CPU. Seems to work here. Scream if something's
amiss.

Thx.

---