Re: [PATCH 3/8] Deliver signal in hardware single step

[Pedro Alves <palves@redhat.com>]

On 03/16/2016 10:47 AM, Yao Qi wrote:
> Pedro Alves <palves@redhat.com> writes:
> 
> Hi Pedro,
> 
>> - If there's a signal handler installed, we'll stop at its entry,
>>    but we haven't stepped over the original breakpoint yet.  If we
>>    were already stopped at the entry of the signal handler, and it
>>    nested, we'll find the program stopped at the same PC it had
>>    started at.  But we won't know whether the step-over finished
>>    successfully (could be a "jmp $pc" instruction), or if instead we're
>>    in a new handler invocation, and thus this is a new,
>>    separate breakpoint hit.
> 
> GDBserver doesn't have to tell the program stopped at the same PC
> it had started at when step over is finished.  When step over, GDBserver
> uninsert breakpoint, resumes the lwp, and wait the next event from the
> lwp.  Once the event from the lwp arrives, GDBserver reinsert breakpoint
> and thinks step over is finished.  That is all ...
> 
>>
>>    A signal handler that segfaults in the first instruction
>>    would be the easiest way to reproduce this that I can think of.
>>    (If it's crashing, you'd expect the user might try putting a
>>    breakpoint there.)
> 
> I write a case like this, set the conditional breakpoint on the first
> instruction of handler, and looks breakpoint condition in target side
> still works properly.
> 
> (gdb) disassemble handler
> Dump of assembler code for function handler:
>     0x0000000000400586 <+0>:     movb   $0x0,0x0
>     0x000000000040058e <+8>:     retq
> 
> (gdb) p/x $rsp
> $1 = 0x7fffffffde00
> (gdb) b *handler if $rsp < 0x7fffffffde00 - 3300
> 
> the condition like this can guarantee that the condition will be true
> after several recursions.
> 
>>
>>    Now, if after stepping into the handler, we immediately reinsert the
>>    breakpoint and continue, we'll retrap it, it ends up OK.
> 
> breakpoint is reinserted after stepping into the handler, because
> GDBserver thinks step over is finished, as I said above.
> 
>>    But if we immediately instead start a new step-over for the same
>>    breakpoint address, we will miss the new hit, and nest a new handler,
>>    on and on.
> 
> This won't happen.  After hardware single step with signal delivered,
> the program stops at the entry of handler, and GDBserver gets an event.
> Then, GDBserver thinks step over is finished, and then proceed all lwps.
> However, the thread still stopped at the breakpoint (of different
> stack frames), so it needs step over again. 

> The loop like this will go
> on until the breakpoint condition is true, need_step_over_p decides to
> resume rather than step over.  The program will hit the breakpoint, and
> GDBserver reports the event back to GDB.
> 
> Note in my experiment with the test case above, I don't let GDBserver
> treat SIGSEGV as SIGTRAP, otherwise SIGSEGV will be reported back to GDB.
> 

OK.

We need to think of unconditional tracepoints as well,
not just conditional breakpoints, however.

Here's a likely scenario, that I think will happen:

#1 - tracepoint set at 0xf00.
#2 - Program stops at 0xf00, tracepoint is collected.
#3 - gdbserver starts a step-over
#4 - instead, a signal arrives, step-over cancelled.
#5 - signal should be passed, not reported to gdb.
#6 - gdbserver starts a new step-over, passing signal as well.
#7 - step lands in handler. step over finished.
#8 - gdbserver proceeds/continues
#9 - signal handler returns, again to 0xf00, tracepoint
     traps again and is thus collected again. This is a
     spurious collection.

gdb avoids the spurious double-hit in #9 by remembering that
it was stepping over a breakpoint when a signal arrived, in
order to continue the step over once the handler returns
(tp->step_after_step_resume_breakpoint).  This only works if the handler 
doesn't cause some other unrelated user-visible stop -- in that case,
then the next time the user continues the program, and the
handler returns to the original breakpoint address, we'll still
report a new breakpoint hit.  I think.

Maybe we can just not bother and live with the spurious double
tracepoint hit/collect in gdbserver.  (Likewise dprintf.)
Dunno, but I'm now starting to lean toward that.

Thanks,
Pedro Alves