Re: [patch] Fix leak of bp_jit_event breakpoints

[Paul Pluzhnikov <ppluzhnikov@google.com>]

On Wed, Jan 19, 2011 at 1:09 PM, Pedro Alves <pedro@codesourcery.com> wrote:
> On Wednesday 19 January 2011 13:43:15, Paul Pluzhnikov wrote:
>> @@ -10588,11 +10588,12 @@ breakpoint_re_set_one (void *bint)
>>        printf_filtered (_("Deleting unknown breakpoint type %d\n"),
>> b->type); /* fall through */
>>        /* Delete overlay event and longjmp master breakpoints; they will be
>> -        reset later by breakpoint_re_set.  */
>> +        reset later by breakpoint_re_set.  Likewise for jit_event.  */
>>      case bp_overlay_event:
>>      case bp_longjmp_master:
>>      case bp_std_terminate_master:
>>      case bp_exception_master:
>> +    case bp_jit_event:
>>        delete_breakpoint (b);
>>        break;
>>
>> @@ -10619,7 +10620,6 @@ breakpoint_re_set_one (void *bint)
>>      case bp_longjmp_resume:
>>      case bp_exception:
>>      case bp_exception_resume:
>> -    case bp_jit_event:
>>        break;
>>      }
>
> This part doesn't look quite right.  In non-stop (breakpoint always-inserted)
> mode, this looks it will create a race window where you delete the jit event
> breakpoint whenever a random thread loads/unloads a DSO, meaning
> you could miss a jit registration done by some other thread still running.

Good point.

We could go with Yao's solution of searching for a matching existing
bp_jit_event at the given address; but I don't like it for two
reasons:
- yet another linear loop
- if a binary is reloaded, and &__jit_debug_register_code changes,
we'll still leak a breakpoint on rerun

Would recording the current jit_event breakpoint in inferior (via
register_inferior_data_with_cleanup ()) be a good solution?

Thanks,
-- 
Paul Pluzhnikov