Re: 32-bit PTI with THP = userspace corruption

From: Joerg Roedel
Date: Fri Aug 31 2018 - 03:07:26 EST

On Fri, Aug 31, 2018 at 07:12:44AM +0300, Meelis Roos wrote:
> > Thanks for the report! I'll try to reproduce the problem tomorrow and
> > investigate it. Can you please check if any of the kernel configurations
> > that show the bug has CONFIG_X86_PAE set? If not, can you please test
> > if enabling this option still triggers the problem?
> Will check, but out of my memery there were 2 G3 HP Proliants that did
> not fit into the pattern (problem did not appear). I have more than 4G
> RAM in those and HIGHMEM_4G there, maybe that's it?

Yeah, I thought a bit about it, and for legacy paging the PMD paging
level is the root-level where we do the mirroring between kernel and
user page-table for PTI. This means we also need to collect A/D bits
from both entries, which we don't do yet.

But that all means it shouldn't happen with CONFIG_X86_PAE=y.

I'll try to reproduce and work on a fix.