Skip to contentSkip to conversation

perf: Adjust perf-inject output data offset for backward compatibility

PATCH Superseded Raul Silvera 3 participants 3 messages Jun 20, 2022 – Jun 21, 2022 perf events

v1 of 2 Superseded Latest is v2 →

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 perf: Adjust perf-inject output data offset for backward compatibility

When perf inject creates a new file, it reuses the data offset from the input file. If there has been a change on the size of the header, as happened in v5.12 -> v5.13, the new offsets will be wrong, resulting in a corrupted output file.

This change adds the function perf_session__data_offset to compute the data offset based on the current header size, and uses that instead of the offset from the original input file.

---
 tools/perf/builtin-inject.c |  2 +-
 tools/perf/util/header.c    | 16 +++++++++++++++-
 tools/perf/util/header.h    |  2 ++
 3 files changed, 18 insertions(+), 2 deletions(-)
tools/perf/builtin-inject.c +1 -1
index a75bf11585b5..1dfdcef36607 100644
@@ -916,7 +916,7 @@ static int __cmd_inject(struct perf_inject *inject)
 		inject->tool.tracing_data = perf_event__repipe_tracing_data;
 	}
 
-	output_data_offset = session->header.data_offset;
+	output_data_offset = perf_session__data_offset(session->evlist);
 
 	if (inject->build_id_all) {
 		inject->tool.mmap	  = perf_event__repipe_buildid_mmap;
tools/perf/util/header.c +15 -1
index 53332da100e8..93706ae827f1 100644
@@ -3660,7 +3660,7 @@ static int perf_session__do_write_header(struct perf_session *session,
 			.size   = evlist->core.nr_entries * sizeof(f_attr),
 		},
 		.data = {
-			.offset = header->data_offset,
+			.offset = attr_offset + evlist->core.nr_entries * sizeof(f_attr),
 			.size	= header->data_size,
 		},
 		/* event_types is ignored, store zeros */
@@ -3686,6 +3686,20 @@ int perf_session__write_header(struct perf_session *session,
 	return perf_session__do_write_header(session, evlist, fd, at_exit, NULL);
 }
 
+size_t perf_session__data_offset(const struct evlist *evlist)
+{
+	struct evsel *evsel;
+	size_t data_offset;
+
+	data_offset = sizeof(struct perf_file_header);
+	evlist__for_each_entry(evlist, evsel) {
+		data_offset += evsel->core.ids * sizeof(u64);
+	}
+	data_offset += evlist->core.nr_entries * sizeof(struct perf_file_attr);
+
+	return data_offset;
+}
+
 int perf_session__inject_header(struct perf_session *session,
 				struct evlist *evlist,
 				int fd,
tools/perf/util/header.h +2 -0
index 08563c1f1bff..56916dabce7b 100644
@@ -136,6 +136,8 @@ int perf_session__inject_header(struct perf_session *session,
 				int fd,
 				struct feat_copier *fc);
 
+size_t perf_session__data_offset(const struct evlist *evlist);
+
 void perf_header__set_feat(struct perf_header *header, int feat);
 void perf_header__clear_feat(struct perf_header *header, int feat);
 bool perf_header__has_feat(const struct perf_header *header, int feat);
Signature · 1 line
2.37.0.rc0.104.g0611611a94-goog
reviewer
#1 lore raw
Quoted 34 lines · Raul Silvera

When perf inject creates a new file, it reuses the data offset from the input file. If there has been a change on the size of the header, as happened in v5.12 -> v5.13, the new offsets will be wrong, resulting in a corrupted output file.

This change adds the function perf_session__data_offset to compute the data offset based on the current header size, and uses that instead of the offset from the original input file.

---
 tools/perf/builtin-inject.c |  2 +-
 tools/perf/util/header.c    | 16 +++++++++++++++-
 tools/perf/util/header.h    |  2 ++
 3 files changed, 18 insertions(+), 2 deletions(-)
tools/perf/builtin-inject.c +1 -1
index a75bf11585b5..1dfdcef36607 100644
@@ -916,7 +916,7 @@ static int __cmd_inject(struct perf_inject *inject)
 		inject->tool.tracing_data = perf_event__repipe_tracing_data;
 	}
 
-	output_data_offset = session->header.data_offset;
+	output_data_offset = perf_session__data_offset(session->evlist);
 
 	if (inject->build_id_all) {
 		inject->tool.mmap	  = perf_event__repipe_buildid_mmap;
tools/perf/util/header.c +0 -0
index 53332da100e8..93706ae827f1 100644

tools/perf/util/header.c static int perf_session__do_write_header(struct perf_session *session, · line 3663

Quoted 5 lines
tools/perf/util/header.c +1 -1
@@ -3660,7 +3660,7 @@ static int perf_session__do_write_header(struct perf_session *session,
 			.size   = evlist->core.nr_entries * sizeof(f_attr),
 		},
 		.data = {
-			.offset = header->data_offset,
+			.offset = attr_offset + evlist->core.nr_entries * sizeof(f_attr),

This looks too late since, although the header comes before the data, it also gets written afterwards (at_exit == true). It would be better if we could rely on header->data_offset at this point.

Quoted 37 lines
tools/perf/util/header.c +14 -0
 			.size	= header->data_size,
 		},
 		/* event_types is ignored, store zeros */
@@ -3686,6 +3686,20 @@ int perf_session__write_header(struct perf_session *session,
 	return perf_session__do_write_header(session, evlist, fd, at_exit, NULL);
 }
 
+size_t perf_session__data_offset(const struct evlist *evlist)
+{
+	struct evsel *evsel;
+	size_t data_offset;
+
+	data_offset = sizeof(struct perf_file_header);
+	evlist__for_each_entry(evlist, evsel) {
+		data_offset += evsel->core.ids * sizeof(u64);
+	}
+	data_offset += evlist->core.nr_entries * sizeof(struct perf_file_attr);
+
+	return data_offset;
+}
+
 int perf_session__inject_header(struct perf_session *session,
 				struct evlist *evlist,
 				int fd,
tools/perf/util/header.h +2 -0
index 08563c1f1bff..56916dabce7b 100644
@@ -136,6 +136,8 @@ int perf_session__inject_header(struct perf_session *session,
 				int fd,
 				struct feat_copier *fc);
 
+size_t perf_session__data_offset(const struct evlist *evlist);
+
 void perf_header__set_feat(struct perf_header *header, int feat);
 void perf_header__clear_feat(struct perf_header *header, int feat);
 bool perf_header__has_feat(const struct perf_header *header, int feat);
#2 lore raw
Quoted 5 lines · Adrian Hunter

This looks too late since, although the header comes before the data, it also gets written afterwards (at_exit == true). It would be better if we could rely on header->data_offset at this point.

Thank you. You're right. We can rely on header->data_offset at this point as it is
adjusted upfront in __cmd_inject. Will send an updated patch.

Raúl E. Silvera
Software Engineer
waymo.com