Josh Poimboeuf [off-list ref] writes:

On Tue, Dec 19, 2017 at 12:28:33PM +0100, Torsten Duwe wrote:

quoted

On Mon, Dec 18, 2017 at 12:56:22PM -0600, Josh Poimboeuf wrote:

quoted

On Mon, Dec 18, 2017 at 03:33:34PM +1000, Nicholas Piggin wrote:

quoted

On Sun, 17 Dec 2017 20:58:54 -0600
Josh Poimboeuf [off-list ref] wrote:

quoted

On Fri, Dec 15, 2017 at 07:40:09PM +1000, Nicholas Piggin wrote:

quoted

On Tue, 12 Dec 2017 08:05:01 -0600
Josh Poimboeuf [off-list ref] wrote:
  

quoted

What about leaf functions?  If a leaf function doesn't establish a stack
frame, and it has inline asm which contains a blr to another function,
this ABI is broken.  

Oops, I meant to say "bl" instead of "blr".

You need to save LR, one way or the other. If gcc thinks it's a leaf function and
does not do it, nor does your asm code, you'll return in an endless loop => bug.

Ah, so the function's return path would be corrupted, and an unreliable
stack trace would be the least of our problems.

That's mostly true.

It is possible to save LR somewhere other than the correct stack slot,
in which case you can return correctly but still confuse the unwinder. A
function can hide its caller that way.

It's stupid and we should never do it, but it's not impossible.

...

So with your proposal, I think I'm convinced that we don't need objtool
for ppc64le.  Does anyone disagree?

I don't disagree, but I'd be happier if we did have objtool support.

Just because it would give us a lot more certainty that we're doing the
right thing everywhere, including in hand-coded asm and inline asm.

It's easy to write powerpc asm such that stack traces are reliable, but
it is *possible* to break them.

There are still a few more things that need to be looked at:

1) With function graph tracing enabled, is the unwinder smart enough to
   get the original function return address, e.g. by calling
   ftrace_graph_ret_addr()?

No I don't think so.

2) Similar question for kretprobes.

3) Any other issues with generated code (e.g., bpf, ftrace trampolines),
   runtime patching (e.g., CPU feature alternatives), kprobes, paravirt,
   etc, that might confuse the unwinder?

We'll have to look, I can't be sure off the top of my head.

4) As a sanity check, it *might* be a good idea for
   save_stack_trace_tsk_reliable() to ensure that it always reaches the
   end of the stack.  There are several ways to do that:

   - If the syscall entry stack frame is always the same size, then the
     "end" would simply mean that the stack pointer is at a certain
     offset from the end of the task stack page.  However this might not
     work for kthreads and idle tasks, unless their stacks also start at
     the same offset.  (On x86 we actually standardized the end of stack
     location for all tasks, both user and kernel.)

Yeah it differs between user and kernel.

   - If the unwinder can get to the syscall frame, it can presumably
     examine regs->msr to check the PR bit to ensure it got all the way
     to syscall entry.  But again this might only work for user tasks,
     depending on how kernel task stacks are set up.

That sounds like a good idea. We could possibly mark the last frame of
kernel tasks somehow.

   - Or a different approach would be to do error checking along the
     way, and reporting an error for any unexpected conditions.

   However, given that backlink/LR corruption doesn't seem possible with
   this architecture, maybe #4 would be overkill.  Personally I would
   feel more comfortable with an "end" check and a WARN() if it doesn't
   reach the end.

Yeah I agree.

cheers