Re: [RFC PATCH 5/7] perf tools: Support setting different slots in a BPF map separately

[Arnaldo Carvalho de Melo]

Em Fri, Nov 20, 2015 at 09:25:36PM +0800, Wangnan (F) escreveu:
> >+		case BPF_MAP_PRIV_KEY_INDICS:
> >+			for (i = 0; i < priv->key.indics.nr_indics; i++) {
> >+				u64 _idx = priv->key.indics.indics[i];
> >+				unsigned int idx = (unsigned int)(_idx);
> >+
> >+				err = (*func)(name, map_fd, &def,
> >+					      priv, &idx, arg);
> >+				if (err) {
> >+					pr_debug("ERROR: failed to insert value to %s[%u]\n",
> >+						 name, idx);
> >+					return err;
> >+				}
> >+			}
> 
> This for-loop has a potential problem that, if perf's user want to
> set a very big array using indices, for example:
> 
>  # perf record -e
> mybpf.c/maps:mymap:values[1,2,3,10-100000,200000-400000]=3/
> mybpf.c/maps:mymap:values[100000-200000]=3/ ...
> 
> Perf would alloc nearly 300000 slots for indices array, consume too much
> memory.
> 
> I will fix this problem by reinterprete indices array, makes negative
> value represent range start and use next slot to store range size. For
> example, the above perf cmdline can be converted to:
> 
> {1,2,3,-10, 99991,-200000,200001} and {-100000,100001}.

Why is that changing the way you specify what entries should be set to
a value will make it not allocate too much memory?

I found the first form of representing ( start-end ) to be better than (
-start, size ), but I would use what the C language uses for expressing
ranges in switch case ranges, which is familiar and doesn't reuses the
minus arithmetic operator to express a range, i.e.:

 # perf record -e \
   mybpf.c/maps:mymap:values[1,2,3,10..100000,200000..400000]=3/

 # perf record -e \
   mybpf.c/maps:mymap:values[100000..200000]=3/ ...

- Arnaldo
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/