From 6ad402ac5ca7d3ebcf79a52592523f5c70ced90b Mon Sep 17 00:00:00 2001 From: Luis Claudio R. Goncalves Date: Fri, 3 Jul 2009 08:30:21 -0500 Subject: [PATCH] sched: Fix spurious system load spikes in /proc/loadavgrt commit 6c24cddbab29f64c31256444d083328936e9be18 in tip. Hello, The values in /proc/loadavgrt are sometimes the real load and sometimes garbage. As you can see in th tests below, it occurs from in 2.6.21.5-rt20 to 2.6.23-rc2-rt2. The code for calc_load(), in kernel/timer.c has not changed much in -rt patches. [lclaudio@lab sandbox]$ ls /proc/loadavg* /proc/loadavg /proc/loadavgrt [lclaudio@lab sandbox]$ uname -a Linux lab.casa 2.6.21-34.el5rt #1 SMP PREEMPT RT Thu Jul 12 15:26:48 EDT 2007 x86_64 x86_64 x86_64 GNU/Linux [lclaudio@lab sandbox]$ cat /proc/loadavg* 4.57 4.90 4.16 3/146 23499 0.44 0.98 1.78 0/146 23499 ... [lclaudio@lab sandbox]$ cat /proc/loadavg* 4.65 4.80 4.75 5/144 20720 23896.04 -898421.23 383170.94 2/144 20720 [root@neverland ~]# uname -a Linux neverland.casa 2.6.21.5-rt20 #2 SMP PREEMPT RT Fri Jul 1318:31:38 BRT 2007 i686 athlon i386 GNU/Linux [root@neverland ~]# cat /proc/loadavg* 0.16 0.16 0.15 1/184 11240 344.65 0.38 311.71 0/184 11240 [williams@torg ~]$ uname -a Linux torg 2.6.23-rc2-rt2 #14 SMP PREEMPT RT Tue Aug 7 20:07:31 CDT 2007 x86_64 x86_64 x86_64 GNU/Linux [williams@torg ~]$ cat /proc/loadavg* 0.88 0.76 0.57 1/257 7267 122947.70 103790.53 -564712.87 0/257 7267 ----------> Fixes spurious system load spikes observed in /proc/loadavgrt, as described in: Bug 253103: /proc/loadavgrt issues weird results https://bugzilla.redhat.com/bugzilla/show_bug.cgi?id=253103 Signed-off-by: Luis Claudio R. Goncalves Signed-off-by: Ingo Molnar Signed-off-by: Thomas Gleixner Signed-off-by: Paul Gortmaker --- kernel/sched_rt.c | 7 +++++++ 1 files changed, 7 insertions(+), 0 deletions(-) diff --git a/kernel/sched_rt.c b/kernel/sched_rt.c index 1996902..cbdaa6e 100644 --- a/kernel/sched_rt.c +++ b/kernel/sched_rt.c @@ -918,6 +918,13 @@ unsigned long rt_nr_uninterruptible(void) for_each_online_cpu(i) sum += cpu_rq(i)->rt.rt_nr_uninterruptible; + /* + * Since we read the counters lockless, it might be slightly + * inaccurate. Do not allow it to go below zero though: + */ + if (unlikely((long)sum < 0)) + sum = 0; + return sum; } -- 1.7.0.4