Re: [PATCH net-next 1/4] net: qdisc: add op to run filters/actions before enqueue

[Cong Wang]

(Why not Cc'ing Jamal for net_sched pathes?)

On Tue, Sep 1, 2015 at 9:34 AM, Daniel Borkmann  wrote:
> From: John Fastabend 
>
> Add a new ->preclassify() op to allow multiqueue queuing disciplines
> to call tc_classify() or perform other work before dev_pick_tx().
>
> This helps, for example, with mqprio queueing discipline that has
> offload support by most popular 10G NICs, where the txq effectively
> picks the qdisc.
>
> Once traffic is being directed to a specific queue then hardware TX
> rings may be tuned to support this traffic type. mqprio already
> gives the ability to do this via skb->priority where the ->preclassify()
> provides more control over packet steering, it can classify the skb
> and set the priority, for example, from an eBPF classifier (or action).
>
> Also this allows traffic classifiers to be run without holding the
> qdisc lock and gives one place to attach filters when mqprio is
> in use. ->preclassify() could also be added to other mq qdiscs later
> on: f.e. most classful qdiscs first check major/minor numbers of
> skb->priority before actually consulting a more complex classifier.
>
> For mqprio case today, a filter has to be attached to each txq qdisc
> to have all traffic hit the filter. Since ->preclassify() is currently
> only used by mqprio, the __dev_queue_xmit() fast path is guarded by
> a generic, hidden Kconfig option (NET_CLS_PRECLASSIFY) that is only
> selected by mqprio, otherwise it defaults to off. Also, the Qdisc
> structure size will stay the same, we move __parent, used by cbq only
> into a write-mostly hole. If actions are enabled, __parent is written
> on every enqueue, and only read, rewritten in reshape_fail() phase.
> Therefore, this place in the read-mostly cacheline could be used by
> preclassify, which is written only once.
>

I don't like this approach. Ideally, qdisc layer should be totally
on top of tx queues, which means tx queue selection should
happen after dequeue. I looked at this before, the change is not
trivial at all given the fact that qdisc ties too much with tx queue
probably due to historical reasons, especially the tx softirq part.
But that is really a long-term solution for me.

I have no big objection for this as a short-term solution, however,
once we add these filters before enqueue, we can't remove them
any more. We really need to think twice about it.

Jamal, do you have any better idea?
--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majord...@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: [PATCH net-next 1/4] net: qdisc: add op to run filters/actions before enqueue

[Jamal Hadi Salim]

On 09/02/15 02:22, Cong Wang wrote:

(Why not Cc'ing Jamal for net_sched pathes?)

On Tue, Sep 1, 2015 at 9:34 AM, Daniel Borkmann  wrote:

From: John Fastabend 

Add a new ->preclassify() op to allow multiqueue queuing disciplines
to call tc_classify() or perform other work before dev_pick_tx().

This helps, for example, with mqprio queueing discipline that has
offload support by most popular 10G NICs, where the txq effectively
picks the qdisc.

Once traffic is being directed to a specific queue then hardware TX
rings may be tuned to support this traffic type. mqprio already
gives the ability to do this via skb->priority where the ->preclassify()
provides more control over packet steering, it can classify the skb
and set the priority, for example, from an eBPF classifier (or action).

Also this allows traffic classifiers to be run without holding the
qdisc lock and gives one place to attach filters when mqprio is
in use. ->preclassify() could also be added to other mq qdiscs later
on: f.e. most classful qdiscs first check major/minor numbers of
skb->priority before actually consulting a more complex classifier.

For mqprio case today, a filter has to be attached to each txq qdisc
to have all traffic hit the filter. Since ->preclassify() is currently
only used by mqprio, the __dev_queue_xmit() fast path is guarded by
a generic, hidden Kconfig option (NET_CLS_PRECLASSIFY) that is only
selected by mqprio, otherwise it defaults to off. Also, the Qdisc
structure size will stay the same, we move __parent, used by cbq only
into a write-mostly hole. If actions are enabled, __parent is written
on every enqueue, and only read, rewritten in reshape_fail() phase.
Therefore, this place in the read-mostly cacheline could be used by
preclassify, which is written only once.



I don't like this approach. Ideally, qdisc layer should be totally
on top of tx queues, which means tx queue selection should
happen after dequeue. I looked at this before, the change is not
trivial at all given the fact that qdisc ties too much with tx queue
probably due to historical reasons, especially the tx softirq part.
But that is really a long-term solution for me.

I have no big objection for this as a short-term solution, however,
once we add these filters before enqueue, we can't remove them
any more. We really need to think twice about it.

Jamal, do you have any better idea?



Sorry for the top quote:
Given the rcu-fication of classifiers i believe the idea will
mostly work; expect user will go nuts sticking all kinds of
classifiers and actions that wont work (example, I dont think
connmark action would work nicely here).
Could we strive to do proper offload ala switchdev?

The comment on the patch on reshape_fail + __parent: for the
record, that is an extremely useful feature (allows an inner qdisc
to provide an opportunity for a classful parent qdisc to
reclassify and therefore reschedule).
Yes, CBQ is the only user - but maybe if it was properly documented
more schedulers could put it to good use.

cheers,
jamal

--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majord...@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: [PATCH net-next 1/4] net: qdisc: add op to run filters/actions before enqueue

[Daniel Borkmann]

On 09/02/2015 08:22 AM, Cong Wang wrote:

(Why not Cc'ing Jamal for net_sched pathes?)


Sorry, forgot about it, and thanks for the Cc!
--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majord...@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: [PATCH net-next 1/4] net: qdisc: add op to run filters/actions before enqueue

[Eric Dumazet]

On Tue, 2015-09-01 at 18:34 +0200, Daniel Borkmann wrote:
> From: John Fastabend 
> 
> Add a new ->preclassify() op to allow multiqueue queuing disciplines
> to call tc_classify() or perform other work before dev_pick_tx().
> 
> This helps, for example, with mqprio queueing discipline that has
> offload support by most popular 10G NICs, where the txq effectively
> picks the qdisc.
> 
> Once traffic is being directed to a specific queue then hardware TX
> rings may be tuned to support this traffic type. mqprio already
> gives the ability to do this via skb->priority where the ->preclassify()
> provides more control over packet steering, it can classify the skb
> and set the priority, for example, from an eBPF classifier (or action).
> 
> Also this allows traffic classifiers to be run without holding the
> qdisc lock and gives one place to attach filters when mqprio is
> in use. ->preclassify() could also be added to other mq qdiscs later
> on: f.e. most classful qdiscs first check major/minor numbers of
> skb->priority before actually consulting a more complex classifier.
> 
> For mqprio case today, a filter has to be attached to each txq qdisc
> to have all traffic hit the filter. Since ->preclassify() is currently
> only used by mqprio, the __dev_queue_xmit() fast path is guarded by
> a generic, hidden Kconfig option (NET_CLS_PRECLASSIFY) that is only
> selected by mqprio,


So all distros will select it, basically.

...

> diff --git a/net/core/dev.c b/net/core/dev.c
> index 877c848..b768bca 100644
> --- a/net/core/dev.c
> +++ b/net/core/dev.c
> @@ -3052,6 +3052,23 @@ static int __dev_queue_xmit(struct sk_buff *skb, void 
> *accel_priv)
>   rcu_read_lock_bh();
>  
>   skb_update_prio(skb);
> +#ifdef CONFIG_NET_CLS_PRECLASSIFY
> + q = rcu_dereference_bh(dev->qdisc);
> + if (q && q->preclassify) {
> + switch (q->preclassify(skb, q)) {
> + default:
> + break;
> +#ifdef CONFIG_NET_CLS_ACT
> + case TC_ACT_SHOT:
> + case TC_ACT_STOLEN:
> + case TC_ACT_QUEUED:
> + kfree_skb(skb);
> + rc = NET_XMIT_SUCCESS;
> + goto out;
> +#endif
> + }
> + }
> +#endif
>  

Since its a device attribute after all, why are you storing it in
dev->qdisc->preclassify, adding a cache line miss for moderate load ?

(mqprio/mq root qdisc is normally not fetched in fast path ?)

dev->preclassify would be better IMO, close to dev->_tx



--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majord...@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: [PATCH net-next 1/4] net: qdisc: add op to run filters/actions before enqueue

[Daniel Borkmann]

On 09/01/2015 07:21 PM, Eric Dumazet wrote:

On Tue, 2015-09-01 at 18:34 +0200, Daniel Borkmann wrote:

From: John Fastabend 

Add a new ->preclassify() op to allow multiqueue queuing disciplines
to call tc_classify() or perform other work before dev_pick_tx().

This helps, for example, with mqprio queueing discipline that has
offload support by most popular 10G NICs, where the txq effectively
picks the qdisc.

Once traffic is being directed to a specific queue then hardware TX
rings may be tuned to support this traffic type. mqprio already
gives the ability to do this via skb->priority where the ->preclassify()
provides more control over packet steering, it can classify the skb
and set the priority, for example, from an eBPF classifier (or action).

Also this allows traffic classifiers to be run without holding the
qdisc lock and gives one place to attach filters when mqprio is
in use. ->preclassify() could also be added to other mq qdiscs later
on: f.e. most classful qdiscs first check major/minor numbers of
skb->priority before actually consulting a more complex classifier.

For mqprio case today, a filter has to be attached to each txq qdisc
to have all traffic hit the filter. Since ->preclassify() is currently
only used by mqprio, the __dev_queue_xmit() fast path is guarded by
a generic, hidden Kconfig option (NET_CLS_PRECLASSIFY) that is only
selected by mqprio,


So all distros will select it, basically.

...


diff --git a/net/core/dev.c b/net/core/dev.c
index 877c848..b768bca 100644
--- a/net/core/dev.c
+++ b/net/core/dev.c
@@ -3052,6 +3052,23 @@ static int __dev_queue_xmit(struct sk_buff *skb, void 
*accel_priv)
rcu_read_lock_bh();

skb_update_prio(skb);
+#ifdef CONFIG_NET_CLS_PRECLASSIFY
+   q = rcu_dereference_bh(dev->qdisc);
+   if (q && q->preclassify) {
+   switch (q->preclassify(skb, q)) {
+   default:
+   break;
+#ifdef CONFIG_NET_CLS_ACT
+   case TC_ACT_SHOT:
+   case TC_ACT_STOLEN:
+   case TC_ACT_QUEUED:
+   kfree_skb(skb);
+   rc = NET_XMIT_SUCCESS;
+   goto out;
+#endif
+   }
+   }
+#endif



Since its a device attribute after all, why are you storing it in
dev->qdisc->preclassify, adding a cache line miss for moderate load ?

(mqprio/mq root qdisc is normally not fetched in fast path ?)

dev->preclassify would be better IMO, close to dev->_tx


Yes, makes sense, as this cacheline is accessed anyway few cycles later.
I'll look into how well this approach integrates into tc's configuration
path. I think mqprio/mq/... could just install/remove dev->preclassify()
handler as soon as first filter is attached/detached.

Thanks,
Daniel
--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majord...@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html