[Qemu-devel] bug in ARM thumb usermode emulation

[Qemu-devel] bug in ARM thumb usermode emulation

[Rafael Espíndola]

Running python 2.4 in qemu sometimes causes this assert

Inconsistency detected by ld.so: rtld.c: 288: _dl_start_final:
Assertion `info->l.l_tls_modid == 0' failed!

The same error is present in all versions tested (0.8.1, 0.8.2 and CVS).

I have been able to create a "small" ARM chroot that contais a test
program and python 2.4. It is available at
http://www.maemo.org.br/platform/rafael/qemu-bug.tar.bz2

The test program sets some environment variables and execs python. The
test can be run with

sudo chroot bug/ ./test

An equivalent test program that skips qemu runs correctly in a real
ARM device, so I thing that the bug is really in qemu.

The bug is very dependent on the environment variables and argv. Small
changes can hide the bug.

Do you have any suggestions on how to debug this?

Thanks,
Rafael

Re: [Qemu-devel] bug in ARM thumb usermode emulation

[Paul Brook]

On Monday 18 September 2006 21:05, Rafael Espíndola wrote:
> Running python 2.4 in qemu sometimes causes this assert
>
> Inconsistency detected by ld.so: rtld.c: 288: _dl_start_final:
> Assertion `info->l.l_tls_modid == 0' failed!

TLS and multithreaded applications are known to not work with usermode 
emulation.

Paul

Re: [Qemu-devel] bug in ARM thumb usermode emulation

[Rafael Espíndola]

> TLS and multithreaded applications are known to not work with usermode
> emulation.

We use python for compiling python modules inside scratchbox. So I
need to fix this  or find a workaround (undef  Py_HAVE_NATIVE_TLS?).

What must be done for implementing TLS + multithread in qemu? Maybe I can help.

> Paul
>

Thanks,
Rafael

Re: [Qemu-devel] bug in ARM thumb usermode emulation

[Paul Brook]

> What must be done for implementing TLS + multithread in qemu? Maybe I can
> help.

* Implement the TLS syscalls (futexes and the architecture specific bits).
There's a patch for clone() here: https://nowt.dyndns.org/patch.qemu_nptl
* Fix all the bits of qemu that aren't threadsafe (mainly the TB cache IIRC).
* Fix guest atomic operations, and make sure guest coherency/ordering 
guaranteed are honoured, even which not guaranteed natively by the host.

Note that this only applies to usermode emulation. Full system emulation works 
fine because this is all handled by the guest kernel, exactly the same as on 
real hardware.

Paul

Re: [Qemu-devel] bug in ARM thumb usermode emulation

[Rafael Espíndola]

On 9/18/06, Paul Brook <paul@codesourcery.com> wrote:
> TLS and multithreaded applications are known to not work with usermode
> emulation.
It looks like this is a memory corruption problem and the assert that
fails is just a coincidence!

I started to remove code from python to find what triggers the assert.
I ended up with an empty main linked only with libc!

I new version of the chroot is available at
http://www.maemo.org.br/platform/rafael/qemu-bug.tar.bz2

Any ideas on how to debug this?

> Paul

Thanks,
Rafael

Re: [Qemu-devel] bug in ARM thumb usermode emulation

[Rafael Espíndola]

> TLS and multithreaded applications are known to not work with usermode
> emulation.

In the end, it was an uninitialized stack allocated variable in libc
that was causing this problem
(http://sources.redhat.com/ml/libc-alpha/2006-09/msg00063.html).

> Paul
>

Thanks,
Rafael