On Sun, 13 Jun 2004 18:02:12 -0400, Chiheb Djabri wrote:
> I tried to compile kernel 2.6.6 for the 4th time but it failed:
>
> include/asm/io.h: In function `isa_check_signature': include/asm/io.h:237:
> internal compiler error: Segmentation fault
Does it fail at exactly the same point each time? If not; read on:
Segfaults during intense IO+CPU activity are classic symptoms of faulty
ram. An apparently stable system can work for months without incident,
only to exhibit random errors during peak workloads. A Kernel compile is
one of those types of activities that is, in fact, so good at producing
this kind of condition, that it is very often specifically used as a means
of stress testing, benchmarking, and doing post-assembly burn-ins.
I should add that (graphic) installs of Linux (pretty much any distro)
also tend to expose memory problems, depending on whether there is a large
package set to install ("everything" type installs).
Somebody should add this to a FAQ, since this has been said a few times in
the past, and is one of those obscure problems that is often difficult to
diagnose, and very often wrongly attributed to software bugs (thus
unnecessarily filling up Bugzilla).
I've literally lost count of how many systems I've seen do this, only to
be magically cured by replacement ram. Don't buy second hand, and buy ECC
if you can use/afford it. Oh, and triple-check your working environment
for static whilst installing the new modules (anti-stat gloves, bare feet,
no carpet, etc.) Don't skim over that; it *is* important. Other computer
components will quite happily survive rough handling, where memory modules
will not - and the damage is imperceptible (you can't see static damage).
These kinds of memory problems are rarely exposed while running Windows,
due to the fact that Linux uses physical memory more aggressively, and
Windows favours swap space (even when it doesn't need it).
I've also had poor results using Memtest86, which almost always fails to
expose memory problems on the same system where a kernel compile will.
I've speculated before that this might be related to the way main bus
memory is accessed during DMA transfers, which IIRC Memtest86 doesn't test
(either at all, or under sufficient load). It has been a while since I
used it though.
-
K.
