On Tue, 2008-08-05 at 17:16 +1000, Benjamin Herrenschmidt wrote:

On Tue, 2008-08-05 at 16:47 +1200, Paul Collins wrote:

quoted

It's about four years old.  It was in storage for about six months and I
got it repaired a few weeks ago (display cable and inverter).  The sort
of crazy crap I've been reporting certainly smacks of memory corruption.
But on the other hand, 2.6.25 (Debian's) and 2.6.26 (my own) have been
trouble-free.

Any chance you can bisect the problem ?

Ok, so I can reproduce on a few 32 bits configs with ftrace enabled.

Looks like some non volatile GPRs get corrupted. I don't know yet if
ftrace is the culprit though, I couldn't find anything obviously wrong
with the mcount implementation we have.

It looks like the corrupted GPR has been saved/restored on the stack
and that the corruption is due to the stack itself being written
to. It's not clear by whome though and in what circumstances.

We'll have to dig more.

Cheers,
Ben.