Skip to contentSkip to conversation

perf callchains: Ask for PERF_RECORD_MMAP for data mmaps for DWARF unwinding

tip:perf/core tip-bot for Arnaldo Carvalho de Melo 1 participant 1 message Jan 17, 2018 perf events

lore
Share & tools
lore

tip-bot for Arnaldo Carvalho de Melo

author merge bot
#0 lore raw
Commit-ID:  0d3dcc0ef13d9d78132d7d6ae068c17ed858e65e
Gitweb:     https://git.kernel.org/tip/0d3dcc0ef13d9d78132d7d6ae068c17ed858e65e
Author:     Arnaldo Carvalho de Melo <[email protected]>
AuthorDate: Tue, 16 Jan 2018 11:16:25 -0300
Committer:  Arnaldo Carvalho de Melo <[email protected]>
CommitDate: Wed, 17 Jan 2018 10:23:33 -0300

perf callchains: Ask for PERF_RECORD_MMAP for data mmaps for DWARF unwinding

When we use a global DWARF setting as in:

	perf record --call-graph dwarf

According to 5c0cf22477ea ("perf record: Store data mmaps for dwarf unwind") we need to set up some extra perf_event_attr bits.

But when we instead do a per event dwarf setting:

perf record -e cycles/call-graph=dwarf/

This was not being done, make them equivalent.

This didn't produce any output changes in my tests while fixing up loose ends in the per-event settings, I found it just by comparing the perf_event_attr fields trying to find an explanation for those problems.

Cc Adrian Hunter <[email protected]>
Cc David Ahern <[email protected]>
Cc Hendrick Brueckner <[email protected]>
Cc Jiri Olsa <[email protected]>
Cc Namhyung Kim <[email protected]>
Cc Noel Grandin <[email protected]>
Cc Thomas Richter <[email protected]>
Cc Wang Nan <[email protected]>
Link https://lkml.kernel.org/n/[email protected] open here
Signed-off-by Arnaldo Carvalho de Melo <[email protected]>
---
 tools/perf/util/evsel.c | 16 +++++++++++++---
 1 file changed, 13 insertions(+), 3 deletions(-)
tools/perf/util/evsel.c +13 -3
index 8f971a2..85eb84d 100644
@@ -726,7 +726,7 @@ perf_evsel__reset_callgraph(struct perf_evsel *evsel,
 }
 
 static void apply_config_terms(struct perf_evsel *evsel,
-			       struct record_opts *opts)
+			       struct record_opts *opts, bool track)
 {
 	struct perf_evsel_config_term *term;
 	struct list_head *config_terms = &evsel->config_terms;

@@ -797,6 +797,8 @@ static void apply_config_terms(struct perf_evsel *evsel,
 
 	/* User explicitly set per-event callgraph, clear the old setting and reset. */
 	if ((callgraph_buf != NULL) || (dump_size > 0) || max_stack) {
+		bool sample_address = false;
+
 		if (max_stack) {
 			param.max_stack = max_stack;
 			if (callgraph_buf == NULL)

@@ -816,6 +818,8 @@ static void apply_config_terms(struct perf_evsel *evsel,
 					       evsel->name);
 					return;
 				}
+				if (param.record_mode == CALLCHAIN_DWARF)
+					sample_address = true;
 			}
 		}
 		if (dump_size > 0) {

@@ -828,8 +832,14 @@ static void apply_config_terms(struct perf_evsel *evsel,
 			perf_evsel__reset_callgraph(evsel, &callchain_param);
 
 		/* set perf-event callgraph */
-		if (param.enabled)
+		if (param.enabled) {
+			if (sample_address) {
+				perf_evsel__set_sample_bit(evsel, ADDR);
+				perf_evsel__set_sample_bit(evsel, DATA_SRC);
+				evsel->attr.mmap_data = track;
+			}
 			perf_evsel__config_callchain(evsel, opts, &param);
+		}
 	}
 }
 

@@ -1060,7 +1070,7 @@ void perf_evsel__config(struct perf_evsel *evsel, struct record_opts *opts,
 	 * Apply event specific term settings,
 	 * it overloads any global configuration.
 	 */
-	apply_config_terms(evsel, opts);
+	apply_config_terms(evsel, opts, track);
 
 	evsel->ignore_missing_thread = opts->ignore_missing_thread;
 }