Red Hat Bugzilla – Bug 84530
Bad stacks on PowerPC
Last modified: 2007-04-18 12:51:21 EDT
Description of problem:
Enabling CYGFUN_KERNEL_THREADS_STACK_MEASUREMENT causes the stack frames
to confuse GDB.
Version-Release number of selected component (if applicable):
Steps to Reproduce:
1. Build a PowerPC application with CYGFUN_KERNEL_THREADS_STACK_MEASUREMENT
2. Stop at a breakpoint or ^C
3. GDB 'bt' command will fail
From further discussion on ecos-discuss:
Gary D. Thomas wrote:
> On Tue, 2003-02-18 at 09:27, Jonathan Larmour wrote:
>> I know you've identified the stack checking as the reason, but it's not
really the cause (other than GDB's stack guesswork), so there's still another
interesting issue with this: the stub shouldn't barf. There's meant to be code
to catch this (see cyg_hal_exception_handler() in
hal/powerpc/arch/current/src/hal_misc.c - look at the stuff at the top about
>> If it correctly caught the exception the backtrace would likely work using
GDB's admittedly ramshackle heuristics. *That's* the problem, not the stack
>> I'll add this to the bug.
> I agree that GDB should not barf - certainly that's a problem.
> However, I think that the stack checking code is breaking the
> EABI and putting stuff on the stack that is, in essence, invalid.
> The PowerPC ABIs really suck (thanks, IBM) in that there needs
> to be some space (I don't recall how much) *above* the stack
> that belongs to the current frame. This means that the last
> stack frame (the ending one, hopefully) can't have those
> 0xDEADBEEF markers immediately adjacent (IIRC).
> We should probably look at the ABI and make sure that what we
> do with the stack for checking is legal.
Well if there's a problem it's with the HAL. HAL_THREAD_INIT_CONTEXT is called
_after_ all the padding is added, and it's that which should add any relevant
space. For the powerpc, it already leaves CYGARC_PPC_STACK_FRAME_SIZE which is a
whole 56 bytes (which looking at the eABI I saw, looks like overkill). However I
also note that area isn't initialized to anything. But then we'd see it without
the stack frame checking enabled either.
This bug has moved to http://bugs.ecos.sourceware.org/show_bug.cgi?id=84530