windows mingw compiled qemu-system-x86_64 crash on startup

[humeafo]

qemu-1.0-rc2/cpu-exec.c:37 longjmp(env->jmp_env, 1); it seems that env->jmp_env destroyed, (gdb) p env->jmp_env
$3 = {0, 0, 0, 36249608, 41418280, 5303318, 41418664, 0, 0, 0, 0, 0, 0, 0, 0, 0}

[humeafo (humeafo)]

it's compiled on windows 2003 and using mingw gcc version 4.6.1

[Stefan Hajnoczi (stefanha)]

On Wed, Nov 16, 2011 at 7:01 AM, humeafo <email address hidden> wrote:
> Public bug reported:
>
> qemu-1.0-rc2/cpu-exec.c:37 longjmp(env->jmp_env, 1); it seems that env->jmp_env destroyed, (gdb) p env->jmp_env
> $3 = {0, 0, 0, 36249608, 41418280, 5303318, 41418664, 0, 0, 0, 0, 0, 0, 0, 0, 0}

Kevin: Is this similar to the issue you found with your mingw cross-compiler?

Stefan

[Kevin Wolf (kwolf-redhat)]

Am 16.11.2011 11:35, schrieb Stefan Hajnoczi:
> On Wed, Nov 16, 2011 at 7:01 AM, humeafo <email address hidden> wrote:
>> Public bug reported:
>>
>> qemu-1.0-rc2/cpu-exec.c:37 longjmp(env->jmp_env, 1); it seems that env->jmp_env destroyed, (gdb) p env->jmp_env
>> $3 = {0, 0, 0, 36249608, 41418280, 5303318, 41418664, 0, 0, 0, 0, 0, 0, 0, 0, 0}
>
> Kevin: Is this similar to the issue you found with your mingw cross-compiler?

The symptoms were different. I didn't get a broken TCG state but some
internals of the Fiber used for coroutines must have been corrupted
(SwitchFiber() crashed when dereferencing a null pointer, but the
externally visible pointer that qemu passed to it was still ok).

Maybe both could be symptoms of the same kind of memory corruption.

Kevin

[humeafo (humeafo)]

maybe it's caused by mingw/gcc? the same binary runs well on win7-x64, but not on win2003-32 bit I'll do more test, if I've time, i'd debug it and try to find the reason

[humeafo (humeafo)]

after some debugging I confirmed that this is caused by a mingw gcc 4.6.1-2 optiomization bug, gcc generated optimized code that used ebp to store some results , while later ebp is used in setjmp and longjmp, so a beiju occurred. mingw gcc 4.5.2works well. the bug should be closed.

[Thomas Huth (th-huth)]

Closing according to comment #5.