Re: [Qemu-devel] QEMU 2.4 for Windows - current status

[Stefan Weil]

Am 06.08.2015 um 10:44 schrieb Kevin Wolf:
> Am 05.08.2015 um 22:30 hat Stefan Weil geschrieben:
> > Am 05.08.2015 um 20:39 schrieb Liviu Ionescu:
> > > > On 05 Aug 2015, at 19:56, Paolo Bonzini <address@hidden> wrote:
> > > >
> > > > ... I am not sure why things break for Stefan...
> > > I confirm Stefan's conclusion, neither in my configuration adding
> > >
> > > #include "qemu-common.h"
> > >
> > > ... in cpu-exec.c makes any difference.
> > >
> > > however adding:
> > >
> > > #if defined(_WIN64)
> > > #ifdef sigsetjmp
> > > #undef sigsetjmp
> > > #endif
> > > #define sigsetjmp(env, savesigs) _setjmp(env, NULL)
> > > #endif
> > >
> > > ... fixes the problem, my custom QEMU happily blinks the LEDs on Win 8.1 
> > > 64-bits (see below).
> > >
> > > perhaps a headers check would be helpful, such mysterious behaviours usually 
> > > back fire at a certain point.
> > >
> > >
> > > regards,
> > >
> > > Liviu
> > http://qemu.weilnetz.de/test/cpu-exec.i shows the result of
> > the C preprocessor:
> >
> > cpu-exec.c gets QEMU's os-win32.h with our definition of setjmp
> > early, but the system header file setjmp.h is included later, and
> > that file re-defines our definitions. Including setjmp.h from
> > os-win32.h would solve the problem, but I think there is a
> > better solution.
> >
> > I am planning to remove the special definitions for _WIN64 from
> > os-win32.h and add them to cpu-exec.c, similar to the code
> > above (which can be shortened a little)
> > but with some comment lines added.
> What happens with the callers outside of cpu-exec.c then? Fixing the
> problem inside os-win32.h seems more robust to me.
>
> Kevin

(sig)setjmp/(sig)longjmp without TCG generated code on the stack
does not need special handling because stack unwinding works
for compiled normal C code.

We only have a problem on 64 bit Windows with code generated
at run time by TCG because stack unwinding does not work there,
so a siglongjmp which tries to exit from that code will crash QEMU
when trying to unwind the call stack.

As far as I see, the only siglongjmp exits from TCG generated code
are in cpu-exec.c (functions cpu_loop_exit and
cpu_resume_from_signal). The matching sigsetjmp is also in
cpu-exec.c, and here we have to disable stack unwinding
by calling _setjmp(cpu->jmp_env, NULL).

Stefan