Re: Missing backtrace information on aarch64

[Stephen Hemminger]

On Thu, 19 Mar 2020 09:09:36 +0800
Phil Reid <address@hidden> wrote:

> On 19/03/2020 08:57, Stephen Hemminger wrote:
> > On Thu, 19 Mar 2020 08:34:35 +0800
> > Phil Reid <address@hidden> wrote:
> >   
> >> On 19/03/2020 01:45, Stephen Hemminger wrote:  
> >>> Trying to use libunwind on Arm64 and discovered that it is not working
> >>> as expected. Only one function is printed.
> >>>
> >>> On x86 Debian 10 (1.2.1-9)
> >>> $ ./bt
> >>> ^\Signal Quit recevied
> >>> #0 0x5640deb47308 (sigquit+0x29)
> >>> #1 0x7f4337b8587f (killpg+0x40)
> >>> #2 0x7f4337c146f4 (nanosleep+0x14)
> >>> #3 0x7f4337c1462a (sleep+0x3a)
> >>> #4 0x5640deb4732c (main+0x1f)
> >>> #5 0x7f4337b7209b (__libc_start_main+0xeb)
> >>> #6 0x564
> >>>
> >>>
> >>> On Arm64
> >>>
> >>> # /tmp/bt
> >>> ^\Signal Quit recevied
> >>> #0 0xab9f2f927bf0 (sigquit+0x34)
> >>> Aborted (core dumped)
> >>>      
> >>
> >> I patched gcc libs with the following for an 32bit arm system to get more 
> >> complete backtraces.
> >>
> >>
> >> diff -Naur a/libsupc++/eh_terminate.cc b/libsupc++/eh_terminate.cc
> >> --- a/libstdc++-v3/libsupc++/eh_terminate.cc       2019-02-28 
> >> 14:31:50.951261183 +0800
> >> +++ b/libstdc++-v3/libsupc++/eh_terminate.cc       2019-02-28 
> >> 14:32:42.040396849 +0800
> >> @@ -40,7 +40,7 @@
> >>    using namespace __cxxabiv1;
> >>
> >>    void
> >> -__cxxabiv1::__terminate (std::terminate_handler handler) throw ()
> >> +__cxxabiv1::__terminate (std::terminate_handler handler)
> >>    {
> >>      __try
> >>        {
> >> diff -Naur a/libsupc++/unwind-cxx.h b/libsupc++/unwind-cxx.h
> >> --- a/libstdc++-v3/libsupc++/unwind-cxx.h  2019-02-28 14:31:50.949253953 
> >> +0800
> >> +++ b/libstdc++-v3/libsupc++/unwind-cxx.h  2019-02-28 14:32:53.073426806 
> >> +0800
> >> @@ -189,8 +189,7 @@
> >>
> >>    // Invokes given handler, dying appropriately if the user handler was
> >>    // so inconsiderate as to return.
> >> -extern void __terminate(std::terminate_handler) throw ()
> >> -  __attribute__((__noreturn__));
> >> +extern void __terminate(std::terminate_handler);
> >>    extern void __unexpected(std::unexpected_handler)
> >>      __attribute__((__noreturn__));  
> > 
> > But gcc backtrace already works on this platform, I am confused.
> >   
> 
> Do you me gdb backtrace or the gcc builtin functions?
> gdb worked for me ok as well before the patch.
> 
> Any which way that's what I had to do.
> I think the noreutrn attribute results in something getting
> missed thats required for libunwind.

Did some testing, the problem seems to be caused by more agressive inlining
on the Aarch64 gcc.  If done with lower optimization, then it is ok