Re: [patch net-next-2.6 V3] net: convert bonding to use rx_handler

["Nicolas de Pesloüan" <nicolas.2p.debian@gmail.com>]

Le 19/02/2011 12:28, Jiri Pirko a écrit :
> Sat, Feb 19, 2011 at 12:08:31PM CET, jpirko@redhat.com wrote:
>> Sat, Feb 19, 2011 at 11:56:23AM CET, nicolas.2p.debian@gmail.com wrote:
>>> Le 19/02/2011 09:05, Jiri Pirko a écrit :
>>>> This patch converts bonding to use rx_handler. Results in cleaner
>>>> __netif_receive_skb() with much less exceptions needed. Also
>>>> bond-specific work is moved into bond code.
>>>>
>>>> Signed-off-by: Jiri Pirko<jpirko@redhat.com>
>>>>
>>>> v1->v2:
>>>>          using skb_iif instead of new input_dev to remember original
>>>> 	device
>>>> v2->v3:
>>>> 	set orig_dev = skb->dev if skb_iif is set
>>>>
>>>
>>> Why do we need to let the rx_handlers call netif_rx() or __netif_receive_skb()?
>>>
>>> Bonding used to be handled with very few overhead, simply replacing
>>> skb->dev with skb->dev->master. Time has passed and we eventually
>>> added many special processing for bonding into __netif_receive_skb(),
>>> but the overhead remained very light.
>>>
>>> Calling netif_rx() (or __netif_receive_skb()) to allow nesting would probably lead to some overhead.
>>>
>>> Can't we, instead, loop inside __netif_receive_skb(), and deliver
>>> whatever need to be delivered, to whoever need, inside the loop ?
>>>
>>> rx_handler = rcu_dereference(skb->dev->rx_handler);
>>> while (rx_handler) {
>>> 	/* ...  */
>>> 	orig_dev = skb->dev;
>>> 	skb = rx_handler(skb);
>>> 	/* ... */
>>> 	rx_handler = (skb->dev != orig_dev) ? rcu_dereference(skb->dev->rx_handler) : NULL;
>>> }
>>>
>>> This would reduce the overhead, while still allowing nesting: vlan on
>>> top on bonding, bridge on top on bonding, ...
>>
>> I see your point. Makes sense to me. But the loop would have to include
>> at least processing of ptype_all too. I'm going to cook a follow-up
>> patch.
>>
>
> DRAFT (doesn't modify rx_handlers):
>
> diff --git a/net/core/dev.c b/net/core/dev.c
> index 4ebf7fe..e5dba47 100644
> --- a/net/core/dev.c
> +++ b/net/core/dev.c
> @@ -3115,6 +3115,7 @@ static int __netif_receive_skb(struct sk_buff *skb)
>   {
>   	struct packet_type *ptype, *pt_prev;
>   	rx_handler_func_t *rx_handler;
> +	struct net_device *dev;
>   	struct net_device *orig_dev;
>   	struct net_device *null_or_dev;
>   	int ret = NET_RX_DROP;
> @@ -3129,7 +3130,9 @@ static int __netif_receive_skb(struct sk_buff *skb)
>   	if (netpoll_receive_skb(skb))
>   		return NET_RX_DROP;
>
> -	__this_cpu_inc(softnet_data.processed);
> +	skb->skb_iif = skb->dev->ifindex;
> +	orig_dev = skb->dev;

orig_dev should be set inside the loop, to reflect "previously crossed device", while following the 
path:

eth0 -> bond0 -> br0.

First step inside loop:

orig_dev = eth0
skb->dev = bond0 (at the end of the loop).

Second step inside loop:

orig_dev = bond0
skb->dev = br0 (et the end of the loop).

This would allow for exact match delivery to bond0 if someone bind there.

> +
>   	skb_reset_network_header(skb);
>   	skb_reset_transport_header(skb);
>   	skb->mac_len = skb->network_header - skb->mac_header;
> @@ -3138,12 +3141,9 @@ static int __netif_receive_skb(struct sk_buff *skb)
>
>   	rcu_read_lock();
>
> -	if (!skb->skb_iif) {
> -		skb->skb_iif = skb->dev->ifindex;
> -		orig_dev = skb->dev;
> -	} else {
> -		orig_dev = dev_get_by_index_rcu(dev_net(skb->dev), skb->skb_iif);
> -	}

I like the fact that it removes the above part.

> +another_round:
> +	__this_cpu_inc(softnet_data.processed);
> +	dev = skb->dev;
>
>   #ifdef CONFIG_NET_CLS_ACT
>   	if (skb->tc_verd&  TC_NCLS) {
> @@ -3153,7 +3153,7 @@ static int __netif_receive_skb(struct sk_buff *skb)
>   #endif
>
>   	list_for_each_entry_rcu(ptype,&ptype_all, list) {
> -		if (!ptype->dev || ptype->dev == skb->dev) {
> +		if (!ptype->dev || ptype->dev == dev) {
>   			if (pt_prev)
>   				ret = deliver_skb(skb, pt_prev, orig_dev);
>   			pt_prev = ptype;

Inside the loop, we should only do exact match delivery, for &ptype_all and for 
&ptype_base[ntohs(type) & PTYPE_HASH_MASK]:

         list_for_each_entry_rcu(ptype, &ptype_all, list) {
-               if (!ptype->dev || ptype->dev == dev) {
+               if (ptype->dev == dev) {
                         if (pt_prev)
                                 ret = deliver_skb(skb, pt_prev, orig_dev);
                         pt_prev = ptype;
                 }
         }


         list_for_each_entry_rcu(ptype,
                         &ptype_base[ntohs(type) & PTYPE_HASH_MASK], list) {
                 if (ptype->type == type &&
-                   (ptype->dev == null_or_dev || ptype->dev == skb->dev)) {
+                   (ptype->dev == skb->dev)) {
                         if (pt_prev)
                                 ret = deliver_skb(skb, pt_prev, orig_dev);
                         pt_prev = ptype;
                 }
         }

After leaving the loop, we can do wilcard delivery, if skb is not NULL.

         list_for_each_entry_rcu(ptype, &ptype_all, list) {
-               if (!ptype->dev || ptype->dev == dev) {
+               if (!ptype->dev) {
                         if (pt_prev)
                                 ret = deliver_skb(skb, pt_prev, orig_dev);
                         pt_prev = ptype;
                }
         }


         list_for_each_entry_rcu(ptype,
                         &ptype_base[ntohs(type) & PTYPE_HASH_MASK], list) {
-               if (ptype->type == type &&
-                   (ptype->dev == null_or_dev || ptype->dev == skb->dev)) {
+		if (ptype->type == type && !ptype->dev) {
                         if (pt_prev)
                                 ret = deliver_skb(skb, pt_prev, orig_dev);
                         pt_prev = ptype;
                 }
         }

This would reduce the number of tests inside the list_for_each_entry_rcu() loops. And because we 
match only ptype->dev == dev inside the loop and !ptype->dev outside the loop, this should avoid 
duplicate delivery.

Also, for performance reason, exact match protocol handler lists might be moved from ptype_base or 
ptype_all to a per net_device list. That way, the list_for_each_entry_rcu() inside the loop could be 
empty if no protocol handler bind on the current dev.

inside loop:

         list_for_each_entry_rcu(ptype, dev->ptype_all, list) {
                 if (pt_prev)
			ret = deliver_skb(skb, pt_prev, orig_dev);
                 pt_prev = ptype;
         }

         list_for_each_entry_rcu(ptype,
                         dev->ptype_base[ntohs(type) & PTYPE_HASH_MASK], list) {
                 if (ptype->type == type) {
                         if (pt_prev)
                                 ret = deliver_skb(skb, pt_prev, orig_dev);
                         pt_prev = ptype;
                 }
         }

Outside loop :

         list_for_each_entry_rcu(ptype, &ptype_all, list) {
                 if (pt_prev)
                         ret = deliver_skb(skb, pt_prev, orig_dev);
                 pt_prev = ptype;
         }


         list_for_each_entry_rcu(ptype,
                         &ptype_base[ntohs(type) & PTYPE_HASH_MASK], list) {
		if (ptype->type == type) {
                         if (pt_prev)
                                 ret = deliver_skb(skb, pt_prev, orig_dev);
                         pt_prev = ptype;
                 }
         }

This would require several changes into ptype_all and ptype_base handling, but should be faster.

> @@ -3167,7 +3167,7 @@ static int __netif_receive_skb(struct sk_buff *skb)
>   ncls:
>   #endif
>
> -	rx_handler = rcu_dereference(skb->dev->rx_handler);
> +	rx_handler = rcu_dereference(dev->rx_handler);
>   	if (rx_handler) {
>   		if (pt_prev) {
>   			ret = deliver_skb(skb, pt_prev, orig_dev);
> @@ -3176,6 +3176,8 @@ ncls:
>   		skb = rx_handler(skb);
>   		if (!skb)
>   			goto out;
> +		if (dev != skb->dev)

I would use "if (skb->dev != dev)" for clarity, because skb->dev is expected to have changed, not dev.

> +			goto another_round;
>   	}
>
>   	if (vlan_tx_tag_present(skb)) {
>

	Nicolas.