Re: [uml-user] 2.4.27-bs1 guest on 2.6.5 SKAS enabled host...

[Blaisorblade <bla...@ya...>]

On Friday 03 December 2004 08:22, Anthony Brock wrote:
> I'm attempting to build a 2.4 guest that will run and function on our 2.6.5
> SKAS enabled host (stock SuSE kernel). I moved the compile to a SuSE 9.0
> installation when I experienced problems (redefinition errors)
Something about igmp.c? I remember something like that on SuSE 9.1 (never 
diagnosed - host headers problem, probably).

> compiling 
> the kernel on the SuSE 9.1 Professional distribution. Both the kernel and
> image worked fine on the older development machine (once I changed the
> compiler optimization to '-O1').

> However, I just copied the kernel and image back over to our SuSE 9.1
> machine. While the kernel will "launch", it appears to freeze or hang
> immediately. I see the following:


> $ /bin/linux-2.4.27-bs1-32bit umid=test-1 mem=128M con=pts con0=fd:0,fd:1
> ubd0=/opt/images/instances/test_root.img
> ubd1=/opt/images/instances/test_swap.img
> eth0=tuntap,test-1-0,fe:fd:c6:6b:03:47
> Checking for the skas3 patch in the host...found
> Checking for /proc/mm...found
> Checking PROT_EXEC mmap in /tmp...OK
>
> [1]+  Stopped                 /bin/linux-2.4.27-bs1-32bit umid=test-1
> mem=128M con=pts con0=fd:0,fd:1 ubd0=/opt/images/instances/test_root.img
> ubd1=/opt/images/instances/test_swap.img
> eth0=tuntap,test-1-0,fe:fd:c6:6b:03:47
> $
>
> I was quite shocked that the process would "stop" by itself. I then
> attempted to attach with gdb using:

Yes, this is a long-known problem we fixed very recently in 2.6 (I'll backport 
the fix to next 2.4 release, *when* I'll find time).

It is due to a glibc bug in NPTL (see below for explaination), and to 
workaround it there are various roads, all leading to UML not being run with 
libraries from /lib/tls (the NPTL version) but the one in /lib/ (using 
LinuxThreads, which does not show this bug). NPTL evidently is not in SuSE 
9.0 (which runs by default on a 2.4 kernel, so it could not support NPTL), so 
you don't see the problem.

1) running "export LD_ASSUME_KERNEL=2.4.1" before launching UML - the linker 
will then run libraries compatible with a 2.4.1 kernel (i.e., not NPTL).

2) compile UML by also enabling CONFIG_MODE_TT during compilation (you can 
then run it on SKAS mode). This implies that it's statically linked, and 
static linking on normal distros happens against non-NPTL libs (Gentoo breaks 
this rule, but not your case). This is way this bug is not seen often...

3) The same result would be given by enabling CONFIG_STATIC_LINK during 
compilation.

4) Moving on the host /lib/tls to /lib/tls.off - this will however slow your 
host down for heavily threaded programs, so it's just the last resort.

In that point, UML is going to send a SIGSTOP to a child it fork()ed through 
the clone(2) syscall, but it gets the pid to kill() through a getpid() call 
executed by the child.

Now, getpid() is reading its pid from a cache which is not flushed - so it is 
getting the pid value of the father, i.e. the pid of the main uml thread. And 
sending SIGSTOP to it, after which it booms (it will later segfault because 
the flow of execution is completely different from the expected one).

> Attaching to program: /bin/linux-2.4.27-bs1-32bit, process 25303

> At this point, gdb "freezes" as well.
it's waiting for the child to proceed and execute a new instruction...
> I can't do anything until I close the 
> shell in which I originally launched the 2.4.27 kernel. Even "kill -9
> 25303" does not affect the process.

Probably because it's ptraced...

> The only way I've been able to produce 
> a backtrace was:

Thanks for the description.

Do you mind posting what I explained on
http://uml.harlowhill.com/ on the troubleshooting section? It's a Wiki, i.e. 
freely editable - it's the modern way to manage docs, and it's a pity that 
the UML one is so little.

I'd really like to spread it over everybody here!
Bye
-- 
Paolo Giarrusso, aka Blaisorblade
Linux registered user n. 292729
http://www.user-mode-linux.org/~blaisorblade