Re: [PATCH 1/2] alpha/ptrace: Record and handle the absence of switch_stack

[Linus Torvalds]

On Sun, Jun 20, 2021 at 8:18 PM Michael Schmitz <schmitzmic@xxxxxxxxx> wrote:
>
> I hope that makes more sense?

So the problem is in your debug patch: you don't set that
TIS_SWITCH_STACK in nearly enough places.

In this particular example, I think it's that you don't set it in
do_trace_exit, so when you strace the process, the system call exit -
which is where the return value will be picked up - gets that warning.

You did set TIS_SWITCH_STACK on trace_entry, but then it's cleared
again during the system call, and not set at the trace_exit path.
Oddly, your debug patch also _clears_ it on the exit path, but it
doesn't set it when do_trace_exit does the SAVE_SWITCH_STACK.

You oddly also set it for __sys_exit, but not all the other special
system calls that also do that SAVE_SWITCH_STACK.

Really, pretty much every single case of SAVE_SWITCH_STACK would need
to set it. Not just do_trace_enter/exit

It's why I didn't like Eric's debug patch either. It's quite expensive
to do, partly because you look up that curptr thing. All very nasty.

It would be *much* better to make the flag be part of the stack frame,
but sadly at least on alpha we had exported the format of that stack
frame to user space.

Anyway, I think these debug patches are not just expensive but the
m68k one most definitely is also very incomplete.

             Linus