[mythtv-users] Playback freeze/lockup + kernel panic
f-myth-users at media.mit.edu
f-myth-users at media.mit.edu
Wed Feb 3 08:59:26 UTC 2010
> Date: Sun, 31 Jan 2010 09:19:51 -0700
> From: Brian Wood <beww at beww.org>
> On Sunday 31 January 2010 09:12:55 am steve wrote:
> > So I guess it is a memory problem. The thing is the memtest86 didn't turn
> > up anything.
> How long did you run it? Sometimes you need to run it for a day or more.
memtest86+ is also not a magic bullet. You can easily have memory
problems that it cannot detect.
Two examples from my recent personal experience; lots more details
upon request (though I may have sent same in some previous message):
(a) Motherboard that flips bits (with low probability) reading -from-
memory -only if- CPU is throttled down. Screws machine if being
used, say, to rsync some files (gets a few flipped bits every ten
gig or so), but undetectable by memtest86+ because the latter runs
at full speed and CPU is never throttled down. Problem was data-
pattern-dependent as well. Certain problematic files produce the
same md5sum every time (once file was in RAM cache) if run in a
tight loop, but -different- sums with 10-second sleeps in between
each run (each of which only took a few hundred ms and hence
didn't cause the CPU to throttle up and stay that way. Once I
detected the problem, I could very rapidly rule out the filesystem
(crypto FS -> flipped bits would have trashed entire swaths) and
all I/O (same behavior via IDE, SATA, and even USB). [Solution:
disable throttling on that motherboard and run at full speed always.
Biggest issue there was verifying the integrity of a couple of
months of previous operation; what a mess but ultimately okay.]
(b) Three supposedly-identical motherboards (same production batch and
BIOS rev, even) with differing but always not-quite-right behavior
whenever there is more than one stick of a supposedly-compatible
memory installed. One won't run in dual-channel mode without
errors. One -must- run in dual-channel mode. Etc. For certain
memory voltage/timings, memtest86+ (sometimes) detected errors.
For many more, it never detected a thing---but copying 1-5 gig via
NFS would cause one to hang or provoke a kernel oops (or both).
[Solution: lock FSB ratio 1:1 instead of the motherboard's choice
of 6:5. Problem instantly solved, no matter where the memory
sticks are. That one took months, off and on, to finally resolve.]
Cases (a) and (b) were not even the same brand of motherboard and were
purchased about 5 years apart. All were using vendor-recommended RAM.
High-quality power supplies. Problem did not follow sticks. Single
stick -always- okay. Etc.
In neither case was memtest86+ remotely helpful, even on multi-day
runs, but memory errors were very clearly happening. Trying various
versions of memtest86+ did not improve matters.
So if you -do- get an error from memtest86+, you have a problem.
If you -don't- get an error, you actually don't know much of anything.
More information about the mythtv-users