Re: [PATCH] arm64: vdso: Include .eh_frame in debug ELF

[Vincent Whitchurch]

On Fri, Nov 25, 2022 at 01:27:14PM +0100, Szabolcs Nagy wrote:
> The 11/25/2022 11:37, Vincent Whitchurch wrote:
> > We currently strip out .eh_frame to work around crashes in libgcc when
> > it tries to unwind out of signal handlers, see commit 87676cfca141
> > ("arm64: vdso: Disable dwarf unwinding through the sigreturn
> > trampoline").
> > 
> > The .eh_frame does however have correct unwind information for the
> > functions implemented in C in vgettimeofday.c, but currently this
> > information is not available even for offline unwinding using the
> > vdso.so.dbg.  As a result of this, perf built with libdw is unable to
> > unwind the stack when the PC is inside one of these functions.
> > 
> > To fix this, strip the .eh_frame section only from the vdso.so and not
> > from the vdso.so.dbg.  This can be used by offline unwinders with access
> > to the debug symbols, and will not affect libgcc since the section will
> > still not be present in the normal vDSO.
> 
> adding eh_frame to vdso.so.dbg makes sense.
> 
> but if libdw correctly unwinds across a signal handler
> then libgcc should be able to do so too.

I have not tested if libdw can unwind across a signal handler.  It is
unlikely to work since all the CFI directives in __kernel_rt_sigreturn
are commented out. The CFI in .eh_frame only covers the C functions and
unwinding those works with libdw.

 $ aarch64-linux-gnu-objdump --dwarf=frames arch/arm64/kernel/vdso/vdso.so.dbg | grep pc
 00000014 000000000000001c 00000018 FDE cie=00000000 pc=0000000000000330..00000000000005cc
 00000034 0000000000000014 00000038 FDE cie=00000000 pc=00000000000005d0..0000000000000784
 0000004c 0000000000000018 00000050 FDE cie=00000000 pc=0000000000000784..00000000000007fc
 $ nm -n arch/arm64/kernel/vdso/vdso.so.dbg
 0000000000000330 T __kernel_clock_gettime
 00000000000005d0 T __kernel_gettimeofday
 0000000000000784 T __kernel_clock_getres
 0000000000000820 T __kernel_rt_sigreturn

> so maybe eh_frame should be added back to vdso.so just
> without frame info for __kernel_rt_sigreturn+NOP to
> ensure unwinders use heuristics for sigreturn.
> (i dont know if this was considered back when eh_frame
> was dropped from the vdso)

I don't know either why 87676cfca141 had to both remove the CFI
directives from __kernel_rt_sigreturn and remove the .eh_frame section
completely from vdso.so instead of only doing the former, but I assume
there was a good reason for that.