Re: [PATCH v2 2/2] perf tools: add attr->mmap2 support

[Arnaldo Carvalho de Melo]

Em Tue, Sep 10, 2013 at 10:17:28AM -0300, Arnaldo Carvalho de Melo escreveu:
> Em Tue, Sep 10, 2013 at 03:05:03PM +0200, Stephane Eranian escreveu:
> > On Tue, Sep 10, 2013 at 3:00 PM, Arnaldo Carvalho de Melo <acme@xxxxxxxxxxxxxxxxxx> wrote:
> > > Em Mon, Sep 09, 2013 at 04:48:44PM -0300, Arnaldo Carvalho de Melo escreveu:
> > > +++ b/kernel/events/core.c
> > > @@ -5039,6 +5039,7 @@ static void perf_event_mmap_output(struct perf_event *event,
> > >                 mmap_event->event_id.header.size += sizeof(mmap_event->maj);
> > >                 mmap_event->event_id.header.size += sizeof(mmap_event->min);
> > >                 mmap_event->event_id.header.size += sizeof(mmap_event->ino);
> > > +               mmap_event->event_id.header.size += sizeof(mmap_event->ino_generation);
> > >         }
> > >         perf_event_header__init_id(&mmap_event->event_id.header, &sample, event);

> > Arg, yes, this is missing.
> > Do you want to submit the patch?

> I will, in some jiffies, just rebooting with the patched kernel to
> re-run tperf test.
 
> > Thanks for tracking it down.
 
> np :-)

More fallout, this time the resulting tools fail to run on older
kernels, i.e. any kernel that doesn't have perf_event_attr.mmap2,
doesn't matter the event type:

[acme@zoo linux]$ perf record usleep 1
Error:
The sys_perf_event_open() syscall returned with 22 (Invalid argument) for event (cycles).  
/bin/dmesg may provide additional information.
No CONFIG_PERF_EVENTS=y kernel support configured?

usleep: Terminated
[acme@zoo linux]$ perf record -e cpu-clock usleep 1
Error:
The sys_perf_event_open() syscall returned with 22 (Invalid argument) for event (cpu-clock).  
/bin/dmesg may provide additional information.
No CONFIG_PERF_EVENTS=y kernel support configured?

usleep: Terminated
[acme@zoo linux]$

That is because kernel/events/core.c  perf_copy_attr() has this test:

        if (attr->__reserved_1)
                return -EINVAL;

So we must have the handling we have for other things that were added
in the past, like .sample_id_all, .exclude_guest, etc, like is done in
the attached patch, please ack.

Ingo, since you haven't merged my latest perf/urgent pull req, do you
want me to fold this into the patch that makes the tooling side use
PERF_RECORD_MMAP2?

- Arnaldo
diff --git a/tools/perf/util/evsel.c b/tools/perf/util/evsel.c
index ff2098b..404599d 100644
--- a/tools/perf/util/evsel.c
+++ b/tools/perf/util/evsel.c
@@ -27,6 +27,7 @@
 static struct {
 	bool sample_id_all;
 	bool exclude_guest;
+	bool mmap2;
 } perf_missing_features;
 
 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
@@ -677,7 +678,7 @@ void perf_evsel__config(struct perf_evsel *evsel,
 		attr->sample_type	|= PERF_SAMPLE_WEIGHT;
 
 	attr->mmap  = track;
-	attr->mmap2 = track;
+	attr->mmap2 = track && !perf_missing_features.mmap2;
 	attr->comm  = track;
 
 	/*
@@ -1017,6 +1018,8 @@ static int __perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
 	}
 
 fallback_missing_features:
+	if (perf_missing_features.mmap2)
+		evsel->attr.mmap2 = 0;
 	if (perf_missing_features.exclude_guest)
 		evsel->attr.exclude_guest = evsel->attr.exclude_host = 0;
 retry_sample_id:
@@ -1081,8 +1084,11 @@ try_fallback:
 	if (err != -EINVAL || cpu > 0 || thread > 0)
 		goto out_close;
 
-	if (!perf_missing_features.exclude_guest &&
-	    (evsel->attr.exclude_guest || evsel->attr.exclude_host)) {
+	if (!perf_missing_features.mmap2 && evsel->attr.mmap2) {
+		perf_missing_features.mmap2 = true;
+		goto fallback_missing_features;
+	} else if (!perf_missing_features.exclude_guest &&
+		   (evsel->attr.exclude_guest || evsel->attr.exclude_host)) {
 		perf_missing_features.exclude_guest = true;
 		goto fallback_missing_features;
 	} else if (!perf_missing_features.sample_id_all) {