Re: ppc64le reliable stack unwinder and scheduled tasks

Joe Lawrence [off-list ref] writes:

On Fri, Jan 11, 2019 at 01:00:38AM +0100, Nicolai Stange wrote:

quoted

Hi Joe,

Joe Lawrence [off-list ref] writes:

quoted

tl;dr: On ppc64le, what is top-most stack frame for scheduled tasks
       about?

If I'm reading the code in _switch() correctly, the first frame is
completely uninitialized except for the pointer back to the caller's
stack frame.

For completeness: _switch() saves the return address, i.e. the link
register into its parent's stack frame, as is mandated by the ABI and
consistent with your findings below: it's always the second stack frame
where the return address into __switch_to() is kept.

Hi Nicolai,

Good, that makes a lot of sense.  I couldn't find any reference
explaining the contents of frame 0, only unwinding code here and there
(as in crash-utility) that stepped over it.

FWIW, I learned about general stack frame usage on ppc from part 4 of
the introductionary series starting at [1]: it's a good reading and I
can definitely recommend it.

Summary:
- Callers of other functions always allocate a stack frame and only
  set the pointer to the previous stack frame (that's the
  'stdu r1, -STACK_FRAME_OVERHEAD(r1)' insn).
- Callees save their stuff into the stack frame allocated by the caller
  if needed. Where "if needed" == callee in turn calls another function.

The insignificance of frame 0's contents follows from this ABI: the
caller might not have called any callee yet, the callee might be a leaf
and so on.

Finally, as I understand it, the only purpose of _switch() creating a
standard stack frame at the bottom of scheduled out tasks is that the
higher ones can be found (for e.g. the backtracing): otherwise
there would be a pt_regs at the bottom of the stack. But I might be
wrong here.

quoted

<snip>

quoted

Example 1 (RHEL-7)
==================

crash> struct task_struct.thread c00000022fd015c0 | grep ksp
    ksp = 0xc0000000288af9c0

crash> rd 0xc0000000288af9c0 -e 0xc0000000288b0000

 - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -

sp[0]:

c0000000288af9c0:  c0000000288afb90 0000000000dd0000   ...(............
c0000000288af9d0:  c000000000002a94 c000000001c60a00   .*..............

        crash> sym c000000000002a94
        c000000000002a94 (T) hardware_interrupt_common+0x114

So that c000000000002a94 certainly wasn't stored by _switch(). I think
what might have happened is that the switching frame aliased with some
prior interrupt frame as setup by hardware_interrupt_common().

The interrupt and switching frames seem to share a common layout as far
as the lower STACK_FRAME_OVERHEAD + sizeof(struct pt_regs) bytes are
concerned.

That address into hardware_interrupt_common() could have been written by
the do_IRQ() called from there.

That was my initial theory, but then when I saw an ordinary scheduled
task with a similarly strange frame 0, I thought that _switch() might
have been doing something clever (or not).  But according your earlier
explanation, it would line up that these values may be inherited from
do_IRQ() or the like.

quoted

quoted

c0000000288af9e0:  c000000001c60a80 0000000000000000   ................
c0000000288af9f0:  c0000000288afbc0 0000000000dd0000   ...(............
c0000000288afa00:  c0000000014322e0 c000000001c60a00   ."C.............
c0000000288afa10:  c0000002303ae380 c0000002303ae380   ..:0......:0....
c0000000288afa20:  7265677368657265 0000000000002200   erehsger."......

        Uh-oh...

        /* Mark stacktraces with exception frames as unreliable. */
        stack[STACK_FRAME_MARKER] == STACK_FRAME_REGS_MARKER

Aliasing of the switching stack frame with some prior interrupt stack
frame would explain why that STACK_FRAME_REGS_MARKER is still found on
the stack, i.e. it's a leftover.

For testing, could you try whether clearing the word at STACK_FRAME_MARKER
from _switch() helps?

I.e. something like (completely untested):

I'll kick off some builds tonight and try to get tests lined up
tomorrow.  Unfortunately these take a bit of time to run setup, schedule
and complete, so perhaps by next week.

Ok, that's probably still a good test for confirmation, but the solution
you proposed below is much better.

quoted

diff --git a/arch/powerpc/kernel/entry_64.S b/arch/powerpc/kernel/entry_64.S
index 435927f549c4..b747d0647ec4 100644
--- a/arch/powerpc/kernel/entry_64.S
+++ b/arch/powerpc/kernel/entry_64.S

@@ -596,6 +596,10 @@ _GLOBAL(_switch)
 	SAVE_8GPRS(14, r1)
 	SAVE_10GPRS(22, r1)
 	std	r0,_NIP(r1)	/* Return to switch caller */
+
+	li	r23,0
+	std	r23,96(r1)	/* 96 == STACK_FRAME_MARKER * sizeof(long) */
+
 	mfcr	r23
 	std	r23,_CCR(r1)
 	std	r1,KSP(r3)	/* Set old stack pointer */

This may be sufficient to avoid the condition, but if the contents of
frame 0 are truely uninitialized (not to be trusted), should the
unwinder be even looking at that frame (for STACK_FRAMES_REGS_MARKER),
aside from the LR and other stack size geometry sanity checks?

That's a very good point: we'll only ever be examining scheduled out tasks
and current (which at that time is running klp_try_complete_transition()).

current won't be in an interrupt/exception when it's walking its own
stack. And scheduled out tasks can't have an exception/interrupt frame
as their frame 0, correct?

Thus, AFAICS, whenever klp_try_complete_transition() finds a
STACK_FRAMES_REGS_MARKER in frame 0, it is known to be garbage, as you
said.

Thanks,

Nicolai

[1] https://www.ibm.com/developerworks/linux/library/l-powasm1/index.html

-- 
SUSE Linux GmbH, GF: Felix Imendörffer, Jane Smithard, Graham Norton,
HRB 21284 (AG Nürnberg)