Blob Blame History Raw
From e044a7def7f6aa61d3bb2a75a1ef597a4724441b Mon Sep 17 00:00:00 2001
From: William Cohen <wcohen@redhat.com>
Date: Mon, 18 Jun 2018 16:47:00 -0400
Subject: [PATCH] Initialize the trans structure fields to values unused by
 perf samples

On newer 4.14 kernel operf was crashing because the kernel commit
1d953111b648e ("perf/core: Don't report zero PIDs for exiting tasks")
made perf samples have TID/PID of -1 for exiting processes.  OProfile
used this same value internally to indicate the that field had not
been initialized.  As a result operf would incorrectly attempt to run
some bookkeeping code due to an invalid match. The initialization of
the trans tgid field is now changed to a value that is unsed by the
perf samples to ensure that these invalid matches do not occur.

Signed-off-by: William Cohen <wcohen@redhat.com>
---
 libperf_events/operf_utils.cpp | 9 ++++++++-
 1 file changed, 8 insertions(+), 1 deletion(-)

diff --git a/libperf_events/operf_utils.cpp b/libperf_events/operf_utils.cpp
index 58c50a6..057692d 100644
--- a/libperf_events/operf_utils.cpp
+++ b/libperf_events/operf_utils.cpp
@@ -73,7 +73,14 @@ static inline void update_trans_last(struct operf_transient * trans)
 
 static inline void clear_trans(struct operf_transient * trans)
 {
-	trans->tgid = ~0U;
+	/* ~0U (-1) could be used by the kernel perf samples
+	 * for the TID/PID of exiting processes. To avoid
+	 * confusing the oprofile logic use ~1U as the
+	 * initialization value for tgid field to avoid
+	 * possibly matching the TID/PID values for exiting
+	 * processes.
+	 */
+	trans->tgid = ~1U;
 	trans->cur_procinfo = NULL;
 }
 
-- 
1.8.3.1