[PATCH 0/1] (Was uprobes/perf: pre-filtering)

[Oleg Nesterov <oleg@redhat.com>]

On 02/06, Oleg Nesterov wrote:
>
> Btw, during the testing I noticed a minor problem (or problems) in tools/perf/,
> I'll try to investigate...

To avoid the confusion, this has nothing to do with perf/uprobe/prefiltering
changes.



Well. I am very new to this code, and I didn't even use perf before,
it is quite possible I missed something/everything.

But I simply can't understand the logic behind perf_evlist__create_maps/etc.

OK, builtin-record.c:record sets .target->uses_mmap == T, this is correct,
we are going to use perf_mmap().

But why do we need it? It is for perf_evlist__create_maps() to ensure we
do not use cpu_map__dummy_new().

But. Even if we use perf_mmap(), "event->cpu == -1 && !event->attr.inherit"
is fine. And indeed, "perf record -e whatever -p1" creates a single counter
with "cpu = -1" and this works. Because, damn, perf_evlist__config_attrs()
notices "evlist->cpus->map[0] < 0" and sets "opts->no_inherit = true".

But at the same time, this does not allow to do, say,
"perf record -e whatever -C0 -p1". -C0 is simply ignored because when
perf_evlist__create_maps() is called perf_target__has_task() == T due to "-p1".

Not to mention, there is no way to trace init and the children it will
fork...

And otoh. "perf record -e whatever -i true" creates a counter for each
cpu due to ->uses_mmap. This is suboptimal, but of course the hack like

	--- a/tools/perf/builtin-record.c
	+++ b/tools/perf/builtin-record.c
	@@ -1081,6 +1078,8 @@ int cmd_record(int argc, const char **argv, const char *prefix __maybe_unused)
		if (!argc && perf_target__none(&rec->opts.target))
			usage_with_options(record_usage, record_options);
	 
	+	rec->opts.target.uses_mmap = !rec->opts.no_inherit;
	+
		if (rec->force && rec->append_file) {
			ui__error("Can't overwrite and append at the same time."
				  " You need to choose between -f and -A");

should not be even discussed.

And why opts->target.system_wide is only set by OPT_BOOLEAN("all-cpus") ?
I meant, why I can't do "perf record -e whatever -C0" to create a "global"
counter on CPU_0? This doesn't work because __cmd_record() sees !.system_wide
and assumes we need perf_event__synthesize_thread_map() which silently fail.

So I am sending a single patch to fix the problem which complicated my
testing. It is trivial but I am not sure it correct, please review.

Oleg.