On 24-Jul-11, at 2:42 PM, Carlos O'Donell wrote:
On Fri, Jul 22, 2011 at 5:51 PM, John David Anglin <dave.anglin@xxxxxxxx
> wrote:
Do you have a userspace patch? There was the thread stack
allocation bug
and possibly futex related issues.
I do have a userspace patch, what would you like the patch against?
The current version is unstable or 2.11.2-10.
The "Check Summary" is always 0x8400000000800000. I think Kyle
fixed one of
these last March.
The cause was missing IPC system calls. I'm a bit vague regarding
whether
the fix was installed
or not.
I wouldn't expect a missing syscall to HPMC the machine, it should
return ENOSYS
and userspace should return a failure code... eventually something
might fail to
check the function return and fail.
That's far from an HPMC though.
Yah, I now think it was a matter of luck exposing the missing IPC calls.
I wish we knew how to decode the "Check Summary" but I haven't found
any documentation
on the net. My current theory is the value indicates some kind of
cache malfunction. I suspect
that the corruption may occur during range based flushes due code
containing inequivalent
aliases. I think 0x2000000000000000 indicates a memory timeout.
I did find a small bug in flush_cache_range (sr3 has wrong type). If
I'm correct about the
problem being range based, this bug probably made things better by
calling flush_cache_all
more frequently.
I have a whole collection of hpmcs and almost all have the same check
summary. It very
hard to see much consistency. Many have hpmc's in the idle loop. A
couple have hpmc's
in flush_data_cache. It's always possible that there is a hardware
problem, but the problem
occurs so frequently in the libgomp testsuite that I have to think it
is software triggered.
I haven't been able to trigger by running compilations or tests
manually.
There must be some interconnection between the cores because a fault
in one almost triggers
a TOC hpmc in the other. The "Assist Check" value is a space register
value and it's probably
the context of the process that caused the cache problem.
I'm currently running a test to check operation using mainly
flush_instruction_cache and
flush_data_cache (whole cache flushes). So far, it hasn't hpmc'd in
the libgomp testsuite,
but it's very slooooow.
I plan to rebuild the main packages used in the libgomp testsuite to
see if this helps.
Dave
--
John David Anglin dave.anglin@xxxxxxxx
--
To unsubscribe from this list: send the line "unsubscribe linux-parisc" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
