Re: x86-64 bad pmds in 2.6.11.6 II

[Dave Jones <davej@redhat.com>]

On Fri, Apr 15, 2005 at 06:58:20PM +0100, Hugh Dickins wrote:

 > > > If there was a fix for the bad pmd problem it might be a candidate
 > > > for stable, but so far we dont know what causes it yet.
 > > If I figure a way to trigger here, I'll report back.
 > 
 > Dave, earlier on you were quite able to reproduce the problem on 2.6.11,
 > finding it happened the first time you ran X.  Do you have any time to
 > reverify that, then try to reproduce with the load_cr3 in leave_mm patch?
 > 
 > But please don't waste your time on this unless you think it's plausible.

I used to be able to reproduce it 100% by doing this on an vanilla
upstream kernel. Then it changed behaviour so I only saw it happening
on the Fedora kernel.  For the latest Fedora update kernel I backported
this change..
- x86_64: Only free PMDs and PUDs after other CPUs have been flushed
as a 'try it and see'.  At first I thought it killed the bug, but
a day or so later, it started doing it again.

In the Fedora kernel we have a patch which restricts /dev/mem reading,
so I got suspicious about this interacting with any of the changes
that had happened to drivers/char/mem.c
Out of curiousity, I backported the 3-4 patches from .12rc to
the Fedora .11 kernel, and haven't seen the problem since.

The bizarre thing is I can't explain why any of those patches would
make such a difference.  Given the bug seems to be coming and going
for me, its possible its just masked the problem.

		Dave