[PATCH 2/3] sched/core: Add sched_stack_end_check= boot parameter

[Chris Down]

Hi Peter,

Peter Zijlstra writes:
>I just noticed that there's a number of task_stack_end_corrupted()
>checks that:
>
> - do not also do task_scs_end_corrupted(); should
>   task_scs_end_corrupted() be put inside task_stack_end_corrupted() to
>   make sure they're always done in tandem and the shadow stack check
>   isn't lost?

I had a look at the other callers, and I don't think the shadow stack
check is being lost in any of them. Two are the oops paths in the x86 and
powerpc fault handlers, and neither architecture has
CONFIG_SHADOW_CALL_STACK. So putting it inside task_stack_end_corrupted()
would only change __might_resched() and the stack tracer, and only on
arm64 and riscv.

Both of those are about the task stack in particular. __might_resched()
got its hint in
a8b686b3af44 ("sched/debug: Check for stack overflow in ___might_sleep()")
because at the time a stack overrun could corrupt thread_info and trigger
the warning on its own. The stack tracer checks right after it has
recorded a new maximum depth for the task stack. Neither has anything to
do with the shadow stack as far as I can see.

schedule_debug() would also still want to check the two separately, so
that the panic message keeps saying which stack it was.

> - should these other sites also respect this switch? It seems weird to
>   only have this one site affected.

It is a bit odd, but I think only this one should follow it. None of the
others depend on CONFIG_SCHED_STACK_END_CHECK today, and the switch is
only meant to be the boot time version of that option. They also only
look at the magic once something has already gone wrong and we're
printing a report about it. The stack tracer is the exception, but it
only looks when it's enabled and sees a new maximum. So there's no cost
to avoid by turning them off.

If they did follow the switch, anyone who leaves it off would lose the
"Thread overran stack" hint from their oopses.

So I've left both as they are in v2, and the changelog now says why the
other callers don't follow the switch. Let me know if you'd still rather
have them folded together, though.

Thanks!

Chris