From: KOSAKI Motohiro <kosaki.motohiro@gmail.com>
Date: Tue, 24 Jun 2014 02:17:36 +0000 (-0400)
Subject: workaround for /proc/stat go backward
X-Git-Tag: v11.1.1~13^2
X-Git-Url: https://granicus.if.org/sourcecode?a=commitdiff_plain;h=1655f22d0c6d1f23fe3d167936ae54eb2a50cff1;p=sysstat

workaround for /proc/stat go backward

Since dyn-tick feature was introduced (at kernel 2.6.21),
/proc/stat is unreliable and unstable. Stopped cpu can't provide
reliable stat. kernel try to provide a maximum guess. But it is
not 100% accurate.

Unfortunately, the fields are unsigned. So, stepping backward
can make overflow and some command (e.g. mpstat) show very big
number.

This patch provide a workaround.

Signed-off-by: KOSAKI Motohiro <kosaki.motohiro@gmail.com>
---

diff --git a/common.c b/common.c
index 0562993..e96c881 100644
--- a/common.c
+++ b/common.c
@@ -504,13 +504,23 @@ void get_HZ(void)
 double ll_sp_value(unsigned long long value1, unsigned long long value2,
 		   unsigned long long itv)
 {
-	return SP_VALUE(value1, value2, itv);
+	/* Workaround: dyn-tick kernel has a race issue and /proc/stat values
+	   could be backward. */
+	if (value2 < value1)
+		return 0;
+	else
+		return SP_VALUE(value1, value2, itv);
 }
 
 double ll_s_value(unsigned long long value1, unsigned long long value2,
 		  unsigned long long itv)
 {
-	return S_VALUE(value1, value2, itv);
+	/* Workaround: dyn-tick kernel has a race issue and /proc/stat values
+	   could be backward. */
+	if (value2 < value1)
+		return 0;
+	else
+		return S_VALUE(value1, value2, itv);
 }
 
 /*