s390 && user_enable_single_step() (Was: odd utrace testing results on s390x)

[Roland McGrath]

> Oh, I am not sure. But I don't understand TIF_SINGLE_STEP on s390,
> absolutely.
> 
> For example, why do_signal() sets TIF_SINGLE_STEP? Why can't we do

I think we could.  That would be more consistent with other machines.  On
s390, once we set TIF_SINGLE_STEP, we are going to post a SIGTRAP
eventually before going to user mode.  But then tracehook_signal_handler()
also gets stepping=1 and the expected meaning of this is that the arch code
is not itself simulating a single-step for the handler setup.  So the
tracehook (i.e. ptrace/utrace) code does what it does for "need a fake
single-step".  

In ptrace (including utrace-based ptrace), this winds up with sending a
SIGTRAP.  So when we finally do get out of do_signal and TIF_SINGLE_STEP
causes a second SIGTRAP, it's already pending and the second one makes no
difference.

But for the general case of utrace, we'll have the UTRACE_SIGNAL_HANDLER
report, followed by a SIGTRAP that appears to be an authentic single-step
trap, but takes place on the same instruction.  If the resumption after the
UTRACE_SIGNAL_HANDLER report didn't use stepping, then this is an entirely
unexpected extra SIGTRAP.  If we do continue stepping, then we are
expecting the SIGTRAP, but this gets us a spurious and errnoeous report
that looks like the instruction right before the handler's entry point in
memory was just executed.

[Martin:]
> The reason why we set the TIF_SINGLE_STEP bit in do_signal is that we
> want to be able to stop the debugged program before the first
> instruction of the signal handler has been executed. The PER single
> step causes a trap after an instruction has been executed. That first
> instruction can do bad things to the arguments of the signal handler..

That's what tracehook_signal_handler is for.  You're both doing it yourself
in the arch code (by setting TIF_SINGLE_STEP), and then telling the generic
code to do it (by passing stepping=1 to tracehook_signal_handler).


Thanks,
Roland