Re: ppc64le reliable stack unwinder and scheduled tasks
From: Nicolai Stange
Date: Fri Jan 11 2019 - 02:51:59 EST
Joe Lawrence <joe.lawrence@xxxxxxxxxx> writes:
> On Fri, Jan 11, 2019 at 01:00:38AM +0100, Nicolai Stange wrote:
>> Hi Joe,
>>
>> Joe Lawrence <joe.lawrence@xxxxxxxxxx> writes:
>>
>> > 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.
>> <snip>
>>
>> >
>> >
>> > 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.
>
>>
>> > 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.
>>
>> 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 ImendÃrffer, Jane Smithard, Graham Norton,
HRB 21284 (AG NÃrnberg)