Re: Segfaults everywhere on stable/13 after recent upgrade

From: Alfredo_Dal'Ava_Júnior <alfredo_at_freebsd.org>
Date: Mon, 23 Aug 2021 14:56:46 UTC
Hi Julio,

Sorry for the trouble.
The problem was caused by some kind of regression introduced by
LLVM12. The root cause is not well determined and the PPC team is
working on simplifying the test case.
There's a workaround being discussed here: https://reviews.freebsd.org/D31646.

Thanks,
Alfredo


On Sun, Aug 15, 2021 at 3:28 PM Julio Merino <julio@meroh.net> wrote:
>
> Hi all,
>
> I have a G5 running stable/13. A couple of weeks ago, I did the regular
> fetch, rebuild world, and reinstall... and almost all new userland binaries
> started crashing.I thought the crashes would be due to a kernel/userland
> incompatibility, but after rebooting with the new kernel, the problem
> persisted.
>
> It took me some effort to recover the machine into a usable state. I had to
> revert the source tree to right before clang was updated to 12.0.1 on June
> 13th, as I think this is the problematic change, and rebuild + reinstall
> from there.
>
> I just updated the tree again to the most recent stable/13, done another
> buildworld, and I have confirmed that most binaries are still broken in the
> same way:
>
> root@g5:/usr/obj/usr/src/powerpc.powerpc64 # ./bin/echo/echo
> Segmentation fault (core dumped)
> root@g5:/usr/obj/usr/src/powerpc.powerpc64 # gdb ./bin/echo/echo
> ...
> Reading symbols from ./bin/echo/echo...
> Reading symbols from
> /usr/obj/usr/src/powerpc.powerpc64/bin/echo/echo.debug...
> (gdb) run
> Starting program: /usr/obj/usr/src/powerpc.powerpc64/bin/echo/echo
>
> Program received signal SIGSEGV, Segmentation fault.
> 0x0000000010011058 in main (argc=1, argv=0xfffffbfffeab8) at
> /usr/src/bin/echo/echo.c:77
> 77      {
> (gdb)
>
> The stacktraces I get out of the binaries are meaningless. You can see echo
> crashing upon entering main. ls crashes on a local variable assignment. But
> cp gets further along and is able to print its own usage message before
> crashing.
>
> Anybody knows what might be going on?
>
> Thanks!
>
> --
> jmmv.dev