Re: [PATCH 4/5] perf tools: Propagate error info from tp_format

[Arnaldo Carvalho de Melo <acme@kernel.org>]

Em Thu, Sep 10, 2015 at 10:24:52AM +0200, Jiri Olsa escreveu:
> On Wed, Sep 09, 2015 at 05:58:13PM -0300, Arnaldo Carvalho de Melo wrote:
> 
> SNIP
> 
> > This kind of stuff is ok, as evsel is a local variable and you kept the
> > interface for perf_evsel__syscall_newtp(), i.e. it returns NULL if a new
> > evsel can't be instantiated.
> > 
> > Ok, but that is a different interface than the one used by
> > perf_evsel__newtp(), that also instantiates a new evsel.
> > 
> > So when one thinks about "foo__new()" we now need to check which one of
> > the two interfaces it uses, if err.h or if the old NULL based failure
> > reporting one.
> > 
> > Double tricky if it is foo__new() and foo__new_variant(), as
> > perf_evsel__syscall_newtp() and perf_evsel__newtp(), i.e. both will
> > return a "struct perf_evsel" instance, but one using err.h, the other
> > use NULL.
> > 
> > Ok, you marked the ones using a comment, wonder if we couldn't use
> > 'sparse' somehow here, is it used to check IS_ERR() usage in the kernel?
> 
> hum, not sure.. will check ;-)
> 
> at least we could mark related functions with __must_check
> to force the return value check
> 
> > 
> > Ah, but what about this in trace__event_handler() in builtin-trace.c?
> >  
> >         if (evsel->tp_format) {
> >                 event_format__fprintf(evsel->tp_format, sample->cpu,
> >                                       sample->raw_data, sample->raw_size,
> >                                       trace->output);
> >         }
> > 
> > 
> > Don't we have to use IS_ERR() here? Ok, no, because if setting up
> > evsel->tp_format fails, then that evsel will be destroyed and
> > perf_evsel__newtp() will return ERR_PTR(), so it is ok not no use
> > ERR_PTR(evsel->tp_format) because it will only be != NULL when it was
> > successfully set up.
> > 
> > But then, in perf_evsel__newtp_idx if zalloc() fails we will not return
> > ERR_PTR(), but instead NULL, a-ha, this one seems to be a real bug, no?
> 
> hate those allocations in declarations.. never do any good ;-)
> 
> yep, NULL is not an error, so it's real bug, attached patch should fix it
> 
> thanks,
> jirka


Ok continuing, found two more problems in this patch, fixed as follows,
merging.

- Arnaldo

diff --git a/tools/perf/tests/mmap-basic.c b/tools/perf/tests/mmap-basic.c
index 666b67a4df9d..4bb0c5d2059d 100644
--- a/tools/perf/tests/mmap-basic.c
+++ b/tools/perf/tests/mmap-basic.c
@@ -65,7 +65,7 @@ int test__basic_mmap(void)
 
 		snprintf(name, sizeof(name), "sys_enter_%s", syscall_names[i]);
 		evsels[i] = perf_evsel__newtp("syscalls", name);
-		if (evsels[i] == NULL) {
+		if (IS_ERR(evsels[i]) == NULL) {
 			pr_debug("perf_evsel__new\n");
 			goto out_delete_evlist;
 		}
diff --git a/tools/perf/util/evsel.c b/tools/perf/util/evsel.c
index 08c20ee4e27d..6b5d1b509148 100644
--- a/tools/perf/util/evsel.c
+++ b/tools/perf/util/evsel.c
@@ -234,7 +234,9 @@ struct perf_evsel *perf_evsel__newtp_idx(const char *sys, const char *name, int
 	struct perf_evsel *evsel = zalloc(perf_evsel__object.size);
 	int err = -ENOMEM;
 
-	if (evsel != NULL) {
+	if (evsel == NULL) {
+		goto out_err;
+	} else {
 		struct perf_event_attr attr = {
 			.type	       = PERF_TYPE_TRACEPOINT,
 			.sample_type   = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
@@ -261,6 +263,7 @@ struct perf_evsel *perf_evsel__newtp_idx(const char *sys, const char *name, int
 out_free:
 	zfree(&evsel->name);
 	free(evsel);
+out_err:
 	return ERR_PTR(err);
 }
 
diff --git a/tools/perf/util/trace-event.c b/tools/perf/util/trace-event.c
index 8e3a60e3e15f..802bb868d446 100644
--- a/tools/perf/util/trace-event.c
+++ b/tools/perf/util/trace-event.c
@@ -100,7 +100,7 @@ struct event_format*
 trace_event__tp_format(const char *sys, const char *name)
 {
 	if (!tevent_initialized && trace_event__init2())
-		return NULL;
+		return ERR_PTR(-ENOMEM);
 
 	return tp_format(sys, name);
 }