Re: [PATCH net-next v4] rps: selective flow shedding during softnet overflow

[Jamal Hadi Salim <jhs@mojatatu.com>]

[-- Attachment #1: Type: text/plain, Size: 839 bytes --]

On 13-04-23 09:32 PM, Eric Dumazet wrote:
> On Tue, 2013-04-23 at 21:25 -0400, Jamal Hadi Salim wrote:


> Not sure what you mean. The qdisc stuff would replace the 'cpu backlog',

Aha ;->
So you would have many little backlogs one per ring per cpu, correct?


> not be added to it. Think of having possibility to control backlog using
> standard qdiscs, like fq_codel ;)

Excellent. So this is not as a big surgery as it sounds then.
the backloglets just need to be exposed as netdevs.

> Yes, but the per cpu backlog is shared for all devices. We probably want
> different qdisc for gre tunnel, eth0, ...

Makes sense.

BTW, looking at __skb_get_rxhash(), if i had a driver that sets either
skb->rxhash (picks it off the dma descriptor), could i not use that 
instead of computing the hash? something like attached patch.

cheers,
jamal

[-- Attachment #2: p1 --]
[-- Type: text/plain, Size: 412 bytes --]

diff --git a/net/core/flow_dissector.c b/net/core/flow_dissector.c
index e187bf0..a6abee0 100644
--- a/net/core/flow_dissector.c
+++ b/net/core/flow_dissector.c
@@ -159,8 +159,9 @@ void __skb_get_rxhash(struct sk_buff *skb)
 	struct flow_keys keys;
 	u32 hash;
 
-	if (!skb_flow_dissect(skb, &keys))
+	if (skb->rxhash || !skb_flow_dissect(skb, &keys)) {
 		return;
+	}
 
 	if (keys.ports)
 		skb->l4_rxhash = 1;