On Fri, 2016-07-08 at 13:07 +0200, Thomas Graf wrote:

On 07/07/16 at 10:36pm, Jiri Kosina wrote:

quoted

diff --git a/include/linux/netdevice.h b/include/linux/netdevice.h
index f45929c..630838e 100644
--- a/include/linux/netdevice.h
+++ b/include/linux/netdevice.h

@@ -52,6 +52,7 @@
 #include <uapi/linux/netdevice.h>
 #include <uapi/linux/if_bonding.h>
 #include <uapi/linux/pkt_cls.h>
+#include <linux/hashtable.h>
 
 struct netpoll_info;
 struct device;

@@ -1778,6 +1779,7 @@ struct net_device {
 	unsigned int		num_tx_queues;
 	unsigned int		real_num_tx_queues;
 	struct Qdisc		*qdisc;
+	DECLARE_HASHTABLE	(qdisc_hash, 16);

This blows up net_device to an insane size: 64K * sizeof(struct
hlist_head). Can we allocate this on demand for net_devices where
it is actually needed? The majority of virtual devices won't need
this. Doesn't have to be rhashtable, can still be fixed size but
at least allocate it.

Jiri probably misread the API and should have used :

DECLARE_HASHTABLE	(qdisc_hash, 4);

Google has a very similar patch with 16 buckets, and it is 'good enough',
although we do not hit the qdisc_tree_reduce_backlog() penalty.