bonding: set correct tx_queue_len for bond_dev

When we set bandwidth on bond_dev with htb, it's not precise. Because the length of skb_buff queue is bigger than sch->limit whose value is 0, some packets are dropped.
This leads to bond device can't make full use of the bandwidth and influence the accuracy.

With htb, sch->limit is set from bond_dev->tx_queue_len which is initialized to 0 in bond_setup.

To fix this issue, set the bond_dev's tx_queue_len to slave_dev's.

Example:
tc qdisc add dev bond0 root handle 1: htb default 1
tc class add dev bond0 parent 1:0 classid 1:1 htb rate 1Gbit ceil 1Gbit iperf -c host -t 30 -i 10

With old bonding:
[ ID] Interval Transfer Bandwidth

[  3]  0.0-10.0 sec   579 MBytes   486 Mbits/sec
[  3] 10.0-20.0 sec   604 MBytes   507 Mbits/sec
[  3] 20.0-30.0 sec   551 MBytes   462 Mbits/sec
[  3]  0.0-30.0 sec  1.69 GBytes   485 Mbits/sec

With new bonding:

[ ID] Interval       Transfer     Bandwidth
[  3]  0.0-10.0 sec  1.11 GBytes   955 Mbits/sec
[  3] 10.0-20.0 sec  1.11 GBytes   955 Mbits/sec
[  3] 20.0-30.0 sec  1.11 GBytes   955 Mbits/sec
[  3]  0.0-30.0 sec  3.33 GBytes   955 Mbits/sec

---
 drivers/net/bonding/bond_main.c | 6 ++++++
 1 file changed, 6 insertions(+)

index dbbea0e..2536426 100644
@@ -1860,6 +1860,12 @@ int bond_enslave(struct net_device *bond_dev, struct net_device *slave_dev)
 		goto err_dest_symlinks;
 	}
 
+	if (!bond_dev->tx_queue_len)
+		bond_dev->tx_queue_len = slave_dev->tx_queue_len;
+	else
+		bond_dev->tx_queue_len = min(slave_dev->tx_queue_len,
+				bond_dev->tx_queue_len);
+
 	pr_info("%s: enslaving %s as a%s interface with a%s link.\n",
 		bond_dev->name, slave_dev->name,
 		bond_is_active_slave(new_slave) ? "n active" : " backup",

1.7.12



.

From: Yang Yingliang <[email protected]>
Date: Tue, 30 Jul 2013 15:20:42 +0800

This is intentional, software devices should not queue TX packets.

You can only use the packet scheduler with hardware devices.

This makes software devices with htb qdisc can hold only one packet, others are dropped.

Maybe sch->limit's default value which is 1 is too small. Instead of this patch, I'll send another patch to fix this value for fixing this issue.