Re: [PATCH V6 06/12] perf tools: remove unnecessary callchain validation

[Adrian Hunter <adrian.hunter@intel.com>]

On 16/07/13 15:05, Jiri Olsa wrote:
> On Tue, Jul 16, 2013 at 09:38:12AM +0300, Adrian Hunter wrote:
> 
> SNIP
> 
>>  	}
>>  }
>>  
>> -static int perf_session__preprocess_sample(struct perf_session *session,
>> -					   union perf_event *event, struct perf_sample *sample)
>> -{
>> -	if (event->header.type != PERF_RECORD_SAMPLE ||
>> -	    !sample->callchain)
>> -		return 0;
>> -
>> -	if (!ip_callchain__valid(sample->callchain, event)) {
>> -		pr_debug("call-chain problem with event, skipping it.\n");
>> -		++session->stats.nr_invalid_chains;
>> -		session->stats.total_invalid_chains += sample->period;
> 
> How about the '*invalid_chains' stats here? I dont see
> it incremented in the parsing routine.
> 
> Also the current behaviour is to increments stats for invalid
> callchains, but dont fail. With your changes we fail during the
> parsing.

It would fail during parsing sometimes anyway.  The code was:

	if (type & PERF_SAMPLE_CALLCHAIN) {
		if (sample_overlap(event, array, sizeof(data->callchain->nr)))
			return -EFAULT;

		data->callchain = (struct ip_callchain *)array;

		if (sample_overlap(event, array, data->callchain->nr))
			return -EFAULT;

		array += 1 + data->callchain->nr;
	}

But sample overlap did not handle size being effectively negative
i.e. 'offset + size' overflows

static bool sample_overlap(const union perf_event *event,
			   const void *offset, u64 size)
{
	const void *base = event;

	if (offset + size > base + event->header.size)
		return true;

	return false;
}


> 
> On the other hand.. maybe we should fail ;-) I think that
> invalid callchain data is serious enough to be overlooked
> by not seeing the nr_invalid_chains got incremented.
> 
> let's see other comments and then silently push it :-)
> 
> jirka
> 
>