Re: [RFA/commit] Memory leak in on reading frame register

[Pedro Alves <palves@redhat.com>]

On 05/11/2015 09:53 PM, Joel Brobecker wrote:
>>> When using a conditional breakpoint where the condition evaluated
>>> to false a large number of times before the program stopped,
>>> a user reported that GDB's memory consumption was growing very
>>> quickly until it ran out of memory.
>>>
>>> The problem was tracked down to temporary struct values being created
>>> each time the program stops and we evaluate those conditions. This
>>> patch fixes the issue by releasing the temporary values, and adds
>>> a comment explaining why we do that.
>>>
>>> gdb/ChangeLog:
>>>
>>>         Jerome Guitton  <guitton@adacore.com>:
>>> 	* findvar.c (read_frame_register_value): Fix a memory leak.
>>>
>>> Tested on x86_64-linux. No regression.
>>>
>>
>> Not sure about this.
>>
>> How come this in bpstat_check_breakpoint_conditions didn't
>> handle this issue already? :
>>
>> ...
>>       /* We use value_mark and value_free_to_mark because it could
>> 	 be a long time before we return to the command level and
>> 	 call free_all_values.  We can't call free_all_values
>> 	 because we might be in the middle of evaluating a
>> 	 function call.  */
>>       struct value *mark = value_mark ();
>>
>> ...
>>       value_free_to_mark (mark);
> 
> An excellent question, which I will try to research in the next
> couple of days!

Thanks.  I wonder whether the leaks come from constructing the
current frame at each stop, instead of from evaluating
breakpoint conditions.  E.g.., if we do a "step" over:

   while (1);

... are we constantly leaking values until the user does
ctrl-c?

That would suggest to me to that we should be doing
value_mark/value_free_to_mark around each
handle_inferior_event.

> 
> ...
> 
>> Otherwise, what is releasing other kinds of temporary values?
>> Are we leaking them?  E.g., with:
>>
>> int global_val;
>> void foo () {}
>> int main () { while (1) foo (); }
>>
>> and then:
>>
>> (gdb) break foo if global_var == 1
>>
>> an/or:
>>
>> (gdb) break foo if (global_var + 1) == 2
>>
>>
>> Maybe nothing breaks with this patch as its deleting register lval
>> values, but the case above would involve lval_memory values,
>> and if we did something for those like in this patch, I fear
>> that places that want to walk an expression's value chain,
>> like update_watchpoint / can_use_hardware_watchpoint would break.
> 
> But I confess I don't quite understand what you mean by the above.
> Are you saying that the current patch may be OK (because we're

Right, I'm saying that it may not be breaking things just because
(I think) we don't look at lval_register values in the code I
pointed out.

> creating and deleting a value that we know is independent of all
> other values), but that it sets a precendent for other forms where
> it might not be OK?

Right.

Thanks,
Pedro Alves