sched, lockdep: inline double_unlock_balance()
authorSripathi Kodi <sripathik@in.ibm.com>
Wed, 5 Nov 2008 13:27:14 +0000 (18:27 +0530)
committerIngo Molnar <mingo@elte.hu>
Thu, 6 Nov 2008 21:12:09 +0000 (22:12 +0100)
commitcf7f8690e864c6fe11e77202dd847fa60f483418
tree9f0e3cca10a550698c3761c3ee5de6496ecf1e78
parente113a745f693af196c8081b328bf42def086989b
sched, lockdep: inline double_unlock_balance()

We have a test case which measures the variation in the amount of time
needed to perform a fixed amount of work on the preempt_rt kernel. We
started seeing deterioration in it's performance recently. The test
should never take more than 10 microseconds, but we started 5-10%
failure rate.

Using elimination method, we traced the problem to commit
1b12bbc747560ea68bcc132c3d05699e52271da0 (lockdep: re-annotate
scheduler runqueues).

When LOCKDEP is disabled, this patch only adds an additional function
call to double_unlock_balance(). Hence I inlined double_unlock_balance()
and the problem went away. Here is a patch to make this change.

Signed-off-by: Sripathi Kodi <sripathik@in.ibm.com>
Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
kernel/sched.c
kernel/sched_rt.c