Re: [RFC PATCH] linux-user: trap internal SIGABRT's

[Peter Maydell]

On Wed, 9 Feb 2022 at 22:12, Richard Henderson
<richard.henderson@linaro.org> wrote:
>
> On 2/9/22 22:22, Alex Bennée wrote:
> > linux-user wants to trap all signals in case they are related to the
> > guest. This however results in less than helpful core dumps when the
> > error is internal to QEMU. We can detect when an assert failure is in
> > progress by examining __glib_assert_msg and fall through to
> > cpu_abort() which will pretty print something before restoring the
> > default SIGABRT behaviour and dumping core.
> >
> > Signed-off-by: Alex Bennée <alex.bennee@linaro.org>
> > ---
> >   linux-user/signal.c | 6 ++++++
> >   1 file changed, 6 insertions(+)
> >
> > diff --git a/linux-user/signal.c b/linux-user/signal.c
> > index 32854bb375..8ecc1215f7 100644
> > --- a/linux-user/signal.c
> > +++ b/linux-user/signal.c
> > @@ -809,6 +809,8 @@ static inline void rewind_if_in_safe_syscall(void *puc)
> >       }
> >   }
> >
> > +GLIB_VAR char *__glib_assert_msg;
> > +
> >   static void host_signal_handler(int host_sig, siginfo_t *info, void *puc)
> >   {
> >       CPUArchState *env = thread_cpu->env_ptr;
> > @@ -821,6 +823,10 @@ static void host_signal_handler(int host_sig, 
> > siginfo_t *info, void *puc)
> >       uintptr_t pc = 0;
> >       bool sync_sig = false;
> >
> > +    if (__glib_assert_msg) {
> > +        cpu_abort(cpu, "internal QEMU error, aborting...");
> > +    }
>
> I think we should not be trapping SIGABRT.  I think we can preserve all guest 
> behaviour
> wrt SIGABRT by stealing another SIGRTMIN value, and remapping the guest 
> signal number.  We
> can produce the correct result for the system by mapping it back to host 
> SIGABRT in
> core_dump_and_abort().

Stealing signal values is awkward, because you don't know what the guest
(either application or libc) might be trying to do with them. I think
we should prefer not to steal them, especially in this case where the
only reason for taking a signal value for our own use is to deal with
a "should never happen anyway" case. We've already had a few bugs/awkwardnesses
with the current level of signal stealing/rearranging we have to do.

thanks
-- PMM