rcu classic: new algorithm for callbacks-processing(v2)

This is v2, it's a little deference from v1 that I
had send to lkml.
use ACCESS_ONCE
use rcu_batch_after/rcu_batch_before for batch # comparison.

rcutorture test result:
(hotplugs: do cpu-online/offline once per second)

No CONFIG_NO_HZ:           OK, 12hours
No CONFIG_NO_HZ, hotplugs: OK, 12hours
CONFIG_NO_HZ=y:            OK, 24hours
CONFIG_NO_HZ=y, hotplugs:  Failed.
(Failed also without my patch applied, exactly the same bug occurred,
http://lkml.org/lkml/2008/7/3/24)

v1's email thread:
http://lkml.org/lkml/2008/6/2/539

v1's description:

The code/algorithm of the implement of current callbacks-processing
is very efficient and technical. But when I studied it and I found
a disadvantage:

In multi-CPU systems, when a new RCU callback is being
queued(call_rcu[_bh]), this callback will be invoked after the grace
period for the batch with batch number = rcp->cur+2 has completed
very very likely in current implement. Actually, this callback can be
invoked after the grace period for the batch with
batch number = rcp->cur+1 has completed. The delay of invocation means
that latency of synchronize_rcu() is extended. But more important thing
is that the callbacks usually free memory, and these works are delayed
too! it's necessary for reclaimer to free memory as soon as
possible when left memory is few.

A very simple way can solve this problem:
a field(struct rcu_head::batch) is added to record the batch number for
the RCU callback. And when a new RCU callback is being queued, we
determine the batch number for this callback(head->batch = rcp->cur+1)
and we move this callback to rdp->donelist if we find
that head->batch <= rcp->completed when we process callbacks.
This simple way reduces the wait time for invocation a lot. (about
2.5Grace Period -> 1.5Grace Period in average in multi-CPU systems)

This is my algorithm. But I do not add any field for struct rcu_head
in my implement. We just need to memorize the last 2 batches and
their batch number, because these 2 batches include all entries that
for whom the grace period hasn't completed. So we use a special
linked-list rather than add a field.
Please see the comment of struct rcu_data.

Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com>
Cc: "Paul E. McKenney" <paulmck@linux.vnet.ibm.com>
Cc: Dipankar Sarma <dipankar@in.ibm.com>
Cc: Gautham Shenoy <ego@in.ibm.com>
Cc: Dhaval Giani <dhaval@linux.vnet.ibm.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

1 files changed, 17 insertions, 9 deletions

diff --git a/include/linux/rcuclassic.h b/include/linux/rcuclassic.h
index c847e59c6006..04c728147be0 100644
--- a/include/linux/rcuclassic.h
+++ b/include/linux/rcuclassic.h
@@ -66,11 +66,7 @@ static inline int rcu_batch_after(long a, long b)
 	return (a - b) > 0;
 }
 
-/*
- * Per-CPU data for Read-Copy UPdate.
- * nxtlist - new callbacks are added here
- * curlist - current batch for which quiescent cycle started if any
- */
+/* Per-CPU data for Read-Copy UPdate. */
 struct rcu_data {
 	/* 1) quiescent state handling : */
 	long		quiescbatch;     /* Batch # for grace period */
@@ -78,12 +74,24 @@ struct rcu_data {
 	int		qs_pending;	 /* core waits for quiesc state */
 
 	/* 2) batch handling */
-	long  	       	batch;           /* Batch # for current RCU batch */
+	/*
+	 * if nxtlist is not NULL, then:
+	 * batch:
+	 *	The batch # for the last entry of nxtlist
+	 * [*nxttail[1], NULL = *nxttail[2]):
+	 *	Entries that batch # <= batch
+	 * [*nxttail[0], *nxttail[1]):
+	 *	Entries that batch # <= batch - 1
+	 * [nxtlist, *nxttail[0]):
+	 *	Entries that batch # <= batch - 2
+	 *	The grace period for these entries has completed, and
+	 *	the other grace-period-completed entries may be moved
+	 *	here temporarily in rcu_process_callbacks().
+	 */
+	long  	       	batch;
 	struct rcu_head *nxtlist;
-	struct rcu_head **nxttail;
+	struct rcu_head **nxttail[3];
 	long            qlen; 	 	 /* # of queued callbacks */
-	struct rcu_head *curlist;
-	struct rcu_head **curtail;
 	struct rcu_head *donelist;
 	struct rcu_head **donetail;
 	long		blimit;		 /* Upper limit on a processed batch */