Re: Was: boot failure, "DWARF2 unwinder stuck at 0xc0100199"

[Keith Owens <kaos@ocs.com.au>]

Andi Kleen (on Thu, 31 Aug 2006 09:41:40 +0200) wrote:
>P.S.: Badari, we worked out your kernel_math_context trace too:
>that one is actually a gcc bug related to dubious unwind tables generated
>for noreturn calls (in your case do_exit). We were still discussing the best 
>workaround for that one.

Lots of luck.  I logged a bug several years ago against gcc for ia64
with noreturn calls.  When gcc sees a call to a function marked
noreturn (like do_exit or panic), gcc has been known to discard all
code past that point.  The unwind code has to assume that the return
address is pointing into the previous function.  Where does the return
address point after a noreturn call compiled with the gcc bug?  - at
the start of the next function.  Goodbye unwind.

I asked that gcc always insert at least one instruction after a call to
a noreturn function.  That would keep the return address inside the
right function and the unwind code would work.  Ideally that
instruction would cause an error if it was ever executed (break 0 on
ia64, ud2 on i386/x86_64) but even a no-op would be good enough.  Most
of the ia64 list thought it was a good idea, the gcc team disagreed.
AFAIK the bug is still outstanding.