[RFC] [PATCH] Don't run __qdisc_run() on a stopped TX queue

[Krishna Kumar <krkumar2@in.ibm.com>]

From: Krishna Kumar <krkumar2@in.ibm.com>

Driver sends an skb and stops the queue if no more space is
available on the device, and returns OK. When dev_queue_xmit
runs for the next skb, it enqueue's the skb & calls qdisc_run.
qdisc_run dequeue's the skb and finds the queue is stopped
after getting the HARD_LOCK_TX, and puts the skb back on
gso_skb (the next iteration fails faster at dequeue_skb).

This patch avoids calling __qdisc_run if the queue is stopped.
This also lets us remove the queue_stopped check in dequeue_skb
and in qdisc_restart. When the queue is re-enabled, it runs with
one less queue_stopped() check for every skb on the queue (we
still need to own the __QDISC_STATE_RUNNING bit to catch the
stopped condition correctly since stopped cannot be set without
holding this bit).

Results for 3 iteration testing for 1, 4, 8, 12 netperf sessions
running for 1 minute each:

-----------------------------------------------------------------
	       System-X			         P6
-----------------------------------------------------------------
Size	ORG BW		NEW BW		ORG BW		NEW BW
-----------------------------------------------------------------
16K	72183.05	74417.79	60775.76	63413.17
128K	69097.87	72447.12	61692.16	62251.07
256K	60456.21	61415.81	59694.03	61641.81
-----------------------------------------------------------------

Signed-off-by: Krishna Kumar <krkumar2@in.ibm.com>
---

 include/net/pkt_sched.h |    9 +++++++--
 net/sched/sch_generic.c |    9 ++-------
 2 files changed, 9 insertions(+), 9 deletions(-)

diff -ruNp org/include/net/pkt_sched.h new/include/net/pkt_sched.h
--- org/include/net/pkt_sched.h	2009-06-22 11:40:31.000000000 +0530
+++ new/include/net/pkt_sched.h	2009-07-20 13:09:18.000000000 +0530
@@ -92,8 +92,13 @@ extern void __qdisc_run(struct Qdisc *q)
 
 static inline void qdisc_run(struct Qdisc *q)
 {
-	if (!test_and_set_bit(__QDISC_STATE_RUNNING, &q->state))
-		__qdisc_run(q);
+	if (!test_and_set_bit(__QDISC_STATE_RUNNING, &q->state)) {
+		if (!netif_tx_queue_stopped(q->dev_queue))
+			__qdisc_run(q);
+		else
+			/* driver will wake us up anyway, just clear */
+			clear_bit(__QDISC_STATE_RUNNING, &q->state);
+	}
 }
 
 extern int tc_classify_compat(struct sk_buff *skb, struct tcf_proto *tp,
diff -ruNp org/net/sched/sch_generic.c new/net/sched/sch_generic.c
--- org/net/sched/sch_generic.c	2009-05-25 07:48:07.000000000 +0530
+++ new/net/sched/sch_generic.c	2009-07-20 13:09:18.000000000 +0530
@@ -56,12 +56,8 @@ static inline struct sk_buff *dequeue_sk
 	struct sk_buff *skb = q->gso_skb;
 
 	if (unlikely(skb)) {
-		struct net_device *dev = qdisc_dev(q);
-		struct netdev_queue *txq;
-
 		/* check the reason of requeuing without tx lock first */
-		txq = netdev_get_tx_queue(dev, skb_get_queue_mapping(skb));
-		if (!netif_tx_queue_stopped(txq) && !netif_tx_queue_frozen(txq))
+		if (!netif_tx_queue_frozen(q->dev_queue))
 			q->gso_skb = NULL;
 		else
 			skb = NULL;
@@ -142,8 +138,7 @@ static inline int qdisc_restart(struct Q
 	txq = netdev_get_tx_queue(dev, skb_get_queue_mapping(skb));
 
 	HARD_TX_LOCK(dev, txq, smp_processor_id());
-	if (!netif_tx_queue_stopped(txq) &&
-	    !netif_tx_queue_frozen(txq))
+	if (!netif_tx_queue_frozen(txq))
 		ret = dev_hard_start_xmit(skb, dev, txq);
 	HARD_TX_UNLOCK(dev, txq);