Re: [RFC 2/3] use reinsert breakpoint for vCont;s

[Antoine Tremblay <antoine dot tremblay at ericsson dot com>]

Yao Qi writes:

> Yao Qi <qiyaoltc@gmail.com> writes:
>
>> +
>> +      if (!step_over_finished && !can_hardware_single_step ())
>> +	{
>> +	  /* If the thread resumed by resume_step hits the reinsert
>> +	     breakpoint, delete the reinsert breakpoint for it.  */
>> +	  if (current_thread->last_resume_kind == resume_step)
>> +	    delete_reinsert_breakpoints (current_thread);
>> +	  else
>> +	    {
>> +	      /* If the thread resumed by other kind, like
>> +		 resume_continue, hits the breakpoint (either
>> +		 reinsert breakpoint or GDB breakpoint), delete
>> +		 all reinsert breakpoints if it hits non-reinsert
>> +		 breakpoints, otherwise, leave reinsert breakpoint there
>> +		 and step over it.  */
>> +	      if (non_reinsert_breakpoint_inserted_here (event_child->stop_pc))
>> +		delete_reinsert_breakpoints (NULL);
>> +	    }
>> +	}
>>      }
>>    else
>>      {
>>        /* We have some other signal, possibly a step-over dance was in
>>  	 progress, and it should be cancelled too.  */
>>        step_over_finished = finish_step_over (event_child);
>> +
>> +      if (!step_over_finished && !can_hardware_single_step ())
>> +	delete_reinsert_breakpoints (NULL);
>>      }
>>  
>>    /* We have all the data we need.  Either report the event to GDB, or
>> @@ -3568,6 +3590,8 @@ linux_wait_1 (ptid_t ptid,
>>  
>>    /* Alright, we're going to report a stop.  */
>>  
>> +  delete_reinsert_breakpoints (NULL);
>> +
>
> The SIGILL is caused by removing these reinsert breakpoints when threads
> are still running.  I adjust the code removing reinsert breakpoints when
> threads stop, the SIGILL goes away.

I think the insertion of the breakpoints may also be unsafe in non-stop
mode, since correct me if am wrong but in linux_resume we can't assume
that all threads are stopped and thus when we call single_step from:
linux_resume->linux_resume_one_lwp->single_step another thread could
hit the memory we're writing to.

We should stop all threads before the breakpoint insertion like done in
start_step_over.

Actually I think we should have a function like
start_software_vCont that does stops all threads, insert the
breakpoints, resume all threads...

I have not pinpointed the design of this however as I'd rather not call
it from the linux_resume_one_lwp callback. It would be weird to mess
with the thread running state there.

I'm thinking, maybe again close to what step over is doing having a:

if (software_single_step && !hardware_single_step)
   find_inferior (&all_threads, need_software_vCont..)

if (need_software_vCont)
  start_software_vCont
        - stop all threads
        - call single_step to insert the breakpoints
        - resume all threads

And I guess we can have a stop_software_vCont to match it.

I have not gone through the stop scenarios enough yet to tell where however...