Re: [PATCH 11/12] Use reinsert_breakpoint for vCont;s

[Pedro Alves <palves@redhat.com>]

On 06/15/2016 05:41 PM, Yao Qi wrote:
> Pedro Alves <palves@redhat.com> writes:
> 
>> That doesn't tell the reader why we need to stop _all_ threads.  The
>> threads that are about to be resumed are obviously stopped, and
>> thus we could already _access_ inferior memory through them.
> 
> GDB may only resume some threads, and leave other threads running.  In
> order to access inferior memory safely, we must stop all threads.

But what do you mean by "safely" ?  What goes wrong if we
don't stop all threads?

> 
>>
>> I guess this is about flushing instruction caches?
>>
> 
> No, it is not about flushing instruction caches.

Then what is it about?

> 
>>>>> @@ -5176,6 +5241,30 @@ proceed_all_lwps (void)
>>>>>    if (debug_threads)
>>>>>      debug_printf ("Proceeding, no step-over needed\n");
>>>>>  
>>>>> +  /* Re-install the reinsert breakpoints on software single step target
>>>>> +     if the client wants it step.  */
>>>>> +  if (can_software_single_step ())
>>>>
>>>> Not immediately obvious to why is this necessary.  Where were they
>>>> removed in the first place?  I'm it must be necessary, but maybe
>>>> extending the comment helps.
>>>
>>> How about this
>>>
>>>   /* On software single step target, we removed reinsert breakpoints
>>>      after we get any events from the inferior.  
>>
>> Is that all events, even internal events?  From the patch, it seemed
>> like it was only before reporting an event to gdb.
>>
> 
> You are right, I though too much about supporting range-stepping.
> I rewrite the comments in the patch below,
> 
>>> If the client wants
>>>      thread step, re-install these reinsert breakpoints.  */
>>>
>>
>> If we only remove before reporting an event to gdb, then I don't
>> understand this.  We already insert single-step breakpoints when
>> we process the resume request from gdb, no?
> 
> We insert single-step breakpoints when we process the resume requests
> and threads are about to be resumed.  If threads still have pending
> status, single-step breakpoints are not installed, so we need to install
> them in proceed_all_lwps.
> 

> @@ -3518,6 +3521,23 @@ linux_wait_1 (ptid_t ptid,
>        return ignore_event (ourstatus);
>      }
>  
> +  /* Remove reinsert breakpoints ...  */
> +  if (can_software_single_step ()
> +      && has_reinsert_breakpoints (current_thread)
> +      /*... if GDB requests this thread doing resume_step or ...*/
> +      && (current_thread->last_resume_kind == resume_step
> +	  /* GDBserver has already started the step-over for vCont;s,
> +	     but it gets some other signal, like SIGSTOP sent by
> +	     GDBserver for vCont;t or other signal program received.  */
> +	  || !maybe_internal_trap))
> +    {
> +      stop_all_lwps (1, event_child);
> +
> +      delete_reinsert_breakpoints (current_thread);
> +
> +      unstop_all_lwps (1, event_child);
> +    }

I'm re-looking at this and wondering if this is really the
right place to do this.  If the thread hits a breakpoint
that ends up not reported to gdb (e.g., condition evals false),
then we'll remove the reinsert breakpoints here, and then
later reinsert them in proceed_all_lwps.  The extra 
stopping/unstopping everything is best avoided if possible.

Thus, couldn't we move this to after:

  /* We found no reason GDB would want us to stop.  We either hit one
     of our own breakpoints, or finished an internal step GDB
     shouldn't know about.  */
  if (!report_to_gdb)
    {
...
    }
 
?


- Also, if in all-stop mode, if gdb does:

 vCont;s:1;c

thus setting thread 1 stepping, and all others continuing,
and then some other thread but thread 1 hits a breakpoint,
what is removing the reinsert breakpoint of thread 1?

> +
>    /* Note that all addresses are always "out of the step range" when
>       there's no range to begin with.  */
>    in_step_range = lwp_in_step_range (event_child);
> @@ -4281,12 +4301,6 @@ linux_resume_one_lwp_throw (struct lwp_info *lwp,
>  
>        step = maybe_hw_step (thread);
>      }
> -  else
> -    {
> -      /* If the thread isn't doing step-over, there shouldn't be any
> -	 reinsert breakpoints.  */
> -      gdb_assert (!has_reinsert_breakpoints (thread));
> -    }
>  
>    if (fast_tp_collecting == 1)
>      {
> @@ -4841,7 +4855,6 @@ linux_resume_one_thread (struct inferior_list_entry *entry, void *arg)
>  {
>    struct thread_info *thread = (struct thread_info *) entry;
>    struct lwp_info *lwp = get_thread_lwp (thread);
> -  int step;
>    int leave_all_stopped = * (int *) arg;
>    int leave_pending;
>  
> @@ -4910,10 +4923,14 @@ linux_resume_one_thread (struct inferior_list_entry *entry, void *arg)
>  
>    if (!leave_pending)
>      {
> +      int step = 0;
> +
>        if (debug_threads)
>  	debug_printf ("resuming LWP %ld\n", lwpid_of (thread));
>  
> -      step = (lwp->resume->kind == resume_step);
> +      if (lwp->resume->kind == resume_step)
> +	step = maybe_hw_step (thread);
> +
>        linux_resume_one_lwp (lwp, step, lwp->resume->sig, NULL);
>      }
>    else
> @@ -4954,6 +4971,7 @@ linux_resume (struct thread_resume *resume_info, size_t n)
>    struct thread_info *need_step_over = NULL;
>    int any_pending;
>    int leave_all_stopped;
> +  int resume_step_is_handled = 0;
>  
>    if (debug_threads)
>      {
> @@ -4997,12 +5015,55 @@ linux_resume (struct thread_resume *resume_info, size_t n)
>  	debug_printf ("Resuming, no pending status or step over needed\n");
>      }
>  
> +  /* If resume_step is requested by GDB, install reinsert breakpoints
> +     when the thread is about to be actually resumed.  IOW, we don't
> +     insert reinsert breakpoints if any thread has pending status.  */
> +  if (!leave_all_stopped && can_software_single_step ())
> +    {
> +      struct inferior_list_entry *inf, *tmp;
> +
> +      if (debug_threads)
> +	debug_printf ("Handle resume_step.\n");
> +
> +      ALL_INFERIORS (&all_threads, inf, tmp)
> +	{
> +	  struct thread_info *thread = (struct thread_info *) inf;
> +	  struct lwp_info *lwp = get_thread_lwp (thread);
> +
> +	  if (lwp->resume != NULL && lwp->resume->kind == resume_step)
> +	    {
> +	      if (!resume_step_is_handled)
> +		{
> +		  /* We need to access the inferior memory to install
> +		     reinsert breakpoints, so stop all threads.  */
> +		  stop_all_lwps (0, NULL);
> +
> +		  if (debug_threads)
> +		    debug_printf ("Done stopping all threads.\n");
> +
> +		  resume_step_is_handled = 1;
> +		}
> +
> +	      install_software_single_step_breakpoints (lwp);
> +
> +	      if (debug_threads)
> +		debug_printf ("Insert breakpoint for resume_step LWP %ld\n",
> +			      lwpid_of (thread));
> +	    }
> +	}
> +
> +      if (debug_threads)
> +	debug_printf ("Handle resume_step.  Done\n");
> +    }
> +
>    /* Even if we're leaving threads stopped, queue all signals we'd
>       otherwise deliver.  */
>    find_inferior (&all_threads, linux_resume_one_thread, &leave_all_stopped);
>  
>    if (need_step_over)
>      start_step_over (get_thread_lwp (need_step_over));
> +  else if (resume_step_is_handled)
> +    unstop_all_lwps (0, NULL);
>  
>    if (debug_threads)
>      {
> @@ -5098,7 +5159,8 @@ proceed_one_lwp (struct inferior_list_entry *entry, void *except)
>        if (debug_threads)
>  	debug_printf ("   stepping LWP %ld, client wants it stepping\n",
>  		      lwpid_of (thread));
> -      step = 1;
> +
> +      step = maybe_hw_step (thread);
>      }
>    else if (lwp->bp_reinsert != 0)
>      {
> @@ -5164,6 +5226,33 @@ proceed_all_lwps (void)
>    if (debug_threads)
>      debug_printf ("Proceeding, no step-over needed\n");
>  
> +  if (can_software_single_step ())
> +    {
> +      struct inferior_list_entry *inf, *tmp;
> +
> +      ALL_INFERIORS (&all_threads, inf, tmp)
> +	{
> +	  struct thread_info *thread = (struct thread_info *) inf;
> +
> +	  /* On software single step target, we insert reinsert
> +	     breakpoints when the threads are about to be actually
> +	     resumed.  IOW, we don't insert them if any thread has
> +	     pending status.  Before we proceed threads, insert
> +	     reinsert breakpoints if the client wants it step.  */
> +	  if (thread->last_resume_kind == resume_step)
> +	    {
> +	      struct lwp_info *lwp = get_thread_lwp (thread);
> +
> +	      if (!has_reinsert_breakpoints (thread))
> +		install_software_single_step_breakpoints (lwp);
> +
> +	      if (debug_threads)
> +		debug_printf ("Insert breakpoint for resume_step LWP %ld\n",
> +			      lwpid_of (thread));
> +	    }
> +	}
> +    }
> +
>    find_inferior (&all_threads, proceed_one_lwp, NULL);
>  }
>  

Hmm, seeing that we need to handle installing the breakpoints in
both places, I wonder about making linux_resume just handle
setting up the last resume kind and queue signals, and then
end up calling proceed_all_lwps.  If that works, I suspect it
would simplify things a good deal.

Thanks,
Pedro Alves