[lkml]   [2001]   [Jul]   [16]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
SubjectRE: *very* strange heisenbug (kernel? libc?)
It sounds like you are having the same problem I've been having for 
the past several weeks. Check out or
for my description of the problem. The basic idea is, a multithreaded
program on an SMP box seems to run fine except when 'exec'd from
a shell script where I get intermittant crashes and hangs. Removing
CPU's down to one or booting a uniprocessor kernel makes the
problem go away.

I've been playing with linuxthreads and the kernel trying to figure out
what's going on. Linuxthreads makes clone calls from time to time
to create new threads. Occasionally, it seems that the cloned
process is getting a SEGV as it is being scheduled for the very first
time. As far as I can tell this is happening in the switch_to() macro
call in schedule() (kernel/sched.c). This happens before the original
cloning process is even done with the do_fork() call which is a little

I'm not sure if the kernel or linuxthreads is at fault. I will continue to
poke around at this problem. Suggestions encouraged.


----- Message from David Madore <> -----

Date: Sat, 14 Jul 2001 05:46:41 +0200
From: David Madore <>
Subject: *very* strange heisenbug (kernel? libc?)
User-Agent: Mutt/1.2.5i

Hello all.

I have encountered a very, very strange bug, which has caused me a
good number of hours pulling my hair trying to figure out what was
going on. I don't even know if it's a kernel bug, a glibc bug, or a
misunderstanding on my part.

The basic program is simple enough - see All it
does is open the Linux TUN/TAP device (/dev/misc/net/tun), register a
tap device named "test0" (the actual name is unimportant), and then
create a thread which runs an idle loop before quitting immediately.
This seems sane enough. The code to open the TUN/TAP device (function
init_tap()) is practically copied from the sample code in the
documentation, and the rest is trivial. I can hardly imagine where it
might contain an error.

Yet I have gotten it to dump core. See for the core,
which is produced by the binary at URL: This is
all on i386 (detailed system characteristics follow below).

Now that's not all. There are several intriguing things about this

First of all, the place where the segfault happens: in
pthread_create(), of all places! To be precise, the function
__pthread_create_1_2() (defined in linuxthreads/pthread.c in the glibc
sources) calls thread_self() (defined in linuxthreads/internals.h),
which is inlined, and which in turn calls the macro THREAD_SELF,
defined in linuxthreads/sysdeps/i386/useldt.h (around line 57). It is
this macro which segfaults. In fact, the libc tries to determine the
thread's identity by looking up a place in memory accessed by %gs:0,
where %gs is supposed to be a selector for the LDT (normally, 0x7, in
fact); but here it happens that %gs was not initialized, and still
refers to the "flat" memory model (GDT selector 0x2b), so the program
tries to access memory page 0 and fails.

So this would look like a glibc bug (under some weird circumstances,
%gs is not properly initialized by INIT_THREAD_SELF before THREAD_SELF
is called). Except that I've heard rumors that core files were not
reliable for threaded programs. So maybe the segfault is not where I
think it is. (Except that in any case it seems that %gs is wrong at
the place where it is.)

The second funny thing is the circumstances under which I can
reproduce this bug. If I run the program directly, it doesn't
segfault. I have only been able to reproduce it by launching it with
the following ridiculous script:

#! /bin/sh
if echo FOO | egrep -q FOO ; then
exec /tmp/weirdbug
exec /tmp/weirdbug

(wow!). Even then, it is unpredictable. It doesn't always work, and
I've only ever gotten anything on my dual-processor machine. Adding a
&quot;sleep&quot; here or there will get rid of the bug. Of course, trying to
do anything like strace or gdb will make the bug go away - therefore
it's a heisenbug.

Also, sometimes, instead of dumping core, it simply hangs and does

So what say the wise? Is this a glibc bug in the linuxthreads
implementation? A kernel bug in the TUN/TAP device with odd
consequences? A kernel bug in the scheduler leading to a very rare
race condition? A bug in my program? A karma problem?

Detailed system characteristics: RedHat-7.1 system: glibc-2.2.2 (RPM:
glibc-2.2.2-10). Kernel is stock 2.4.6 + international crypto patch
(, compiled for SMP, and with devfsd. Filesystem in /tmp is
ReiserFS. Hardware is dual Intel Pentium II (Deschutes) at 450MHz.
Compiler for kernel is stock egcs-1.1.2. Compiler for program is
RedHat's 2.96-81. Program is compiled against actual kernel headers.

Session log (I'm willing to bet nobody can reproduce it, anyway):

vega root /tmp # /sbin/modprobe tun
vega root /tmp # ls -la /dev/misc/net/tun
crw-r----- 1 root root 10, 200 Jan 1 1970 /dev/misc/net/tun
vega root /tmp # md5sum weirdbug.c
0cbeaa9de585868b3d5dfd55f6f50397 weirdbug.c
vega root /tmp # gcc -o weirdbug weirdbug.c -O6 -lpthread
vega root /tmp # gcc -v
Reading specs from /usr/lib/gcc-lib/i386-redhat-linux/2.96/specs
gcc version 2.96 20000731 (Red Hat Linux 7.1 2.96-81)
vega root /tmp # md5sum weirdbug
7977421cf8b60a63659063d7c165c58b weirdbug
vega root /tmp # cat weirdbug-launcher
#! /bin/sh
if echo FOO | egrep -q FOO ; then
exec /tmp/weirdbug
exec /tmp/weirdbug
vega root /tmp # md5sum weirdbug-launcher
460666262466f12d89f33900e4dc74ec weirdbug-launcher
vega root /tmp # ./weirdbug
vega root /tmp # ./weirdbug-launcher
zsh: segmentation fault (core dumped) ./weirdbug-launcher

David A. Madore
(, )

----- End message -----

To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to
More majordomo info at
Please read the FAQ at

 \ /
  Last update: 2005-03-22 12:57    [W:0.033 / U:0.292 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site