Re: [PATCH v4 00/10] linux-user: Detect and report host crashes

[Michael Tokarev]

18.09.2023 17:05, Helge Deller wrote:
> On 9/12/23 12:34, Michael Tokarev wrote:
> > 12.09.2023 12:45, Helge Deller:
> >
> > > /usr/bin/ld: 
> > > /usr/lib/gcc/x86_64-redhat-linux/13/../../../../lib64/libc.a(abort.o): in 
> > > function `abort':
> > > (.text.unlikely+0x0): multiple definition of `abort'; 
> > > libqemu-aarch64-linux-user.fa.p/linux-user_signal.c.o:/srv/_build/../../home/cvs/qemu/qemu/linux-user/signal.c:723: first defined here
> >
> > [PATCH v4 03/10] linux-user: Use die_with_signal with abort
> >
> > Sigh.
> >
> > I'd be double-cautious when overriding system functions like this, - it's
> > almost always a bad idea.
>
> Richard, I'm not sure, but with that change:
>
> -void abort(void)
> +void  __attribute__((weak)) abort(void)
>
> it will at least successfully link the binary. Not sure which effects it has,
> but probably not worse than before your patch series...

With weak here, ld should peak abort() from glibc, basically reverting this
patch:

linux-user: Use die_with_signal with abort
+/*
+ * The system abort() will raise SIGABRT, which will get caught and deferred
+ * by host_signal_handler.  Returning into system abort will try harder.
+ * Eventually, on x86, it will execute HLT, which raises SIGSEGV.  This goes
+ * back into host_signal_handler, through a different path which may longjmp
+ * back to the main loop.  This often explodes.
+ */
+void abort(void)
+{
+    die_with_signal(SIGABRT);
+}
+

I think it's better to revert it now.

Probably the right solution is to use qemu_abort() (and qemu_assert() etc),
and maybe #define abort(x) qemu_abort(x).  Even if some way to redefine
abort like the above will work on glibc, it does not mean it will work
on *bsd and in other contexts.

Yes, providing its own abort() is tempting due to its simplicity.
But it is a grey area at best.

Thanks,

/mjt