Re: [PATCH net-next] bonding: allow bond in mode balance-alb to work properly in bridge -try4.2

[Eric Dumazet <dada1@cosmosbay.com>]

Jiri Pirko a écrit :
> [PATCH net-next] bonding: allow bond in mode balance-alb to work properly in bridge -try4.2
> 
> (updated)
> changes v4.1 -> v4.2
> - use skb->pkt_type == PACKET_HOST compare rather then comparing skb dest addr
>   against skb->dev->dev_addr
> 
> Hi all.
> 
> The problem is described in following bugzilla:
> https://bugzilla.redhat.com/show_bug.cgi?id=487763
> 
> Basically here's what's going on. In every mode, bonding interface uses the same
> mac address for all enslaved devices (except fail_over_mac). Only balance-alb
> will simultaneously use multiple MAC addresses across different slaves. When you
> put this kind of bond device into a bridge it will only add one of mac adresses
> into a hash list of mac addresses, say X. This mac address is marked as local.
> But this bonding interface also has mac address Y. Now then packet arrives with
> destination address Y, this address is not marked as local and the packed looks
> like it needs to be forwarded. This packet is then lost which is wrong.
> 
> Notice that interfaces can be added and removed from bond while it is in bridge.
> 
> ***
> When the multiple addresses for bridge port approach failed to solve this issue
> due to STP I started to think other way to solve this. I returned to previous
> solution but tweaked one.
> 
> This patch solves the situation in the bonding without touching bridge code.
> For every incoming frame to bonding the destination address is compared to
> current address of the slave device from which tha packet came. If these two
> match destination address is replaced by mac address of the master. This address
> is known by bridge so it is delivered properly. Note that the comparsion is not
> made directly, it's used skb->pkt_type == PACKET_HOST instead. This is "set"
> previously in eth_type_trans().
> 
> I experimentally tried that this works as good as searching through the slave
> list (v4 of this patch).
> 
> I was forced to create a new header because I need to use
> compare_ether_addr_64bits() (defined in linux/etherdevice.h) in
> linux/netdevice.h. I've hit some cross include issues. I think that it's good
> to have skb_bond_should_drop() in a separate file anyway.
> 
> Jirka
> 
> 
> Signed-off-by: Jiri Pirko <jpirko@redhat.com>
> 
> diff --git a/include/linux/bonding.h b/include/linux/bonding.h
> new file mode 100644
> index 0000000..e50939d
> --- /dev/null
> +++ b/include/linux/bonding.h
> @@ -0,0 +1,78 @@
> +/*
> + * include/linux/bonding.h
> + *
> + * Copyright (C) 2009 Jiri Pirko <jpirko@redhat.com>
> + *
> + * This program is free software; you can redistribute it and/or modify
> + * it under the terms of the GNU General Public License version 2
> + * as published by the Free Software Foundation.
> + *
> + * Bonding device helpers.
> + */
> +
> +#ifndef _LINUX_BONDING_H
> +#define _LINUX_BONDING_H
> +
> +#ifdef __KERNEL__
> +
> +#include <linux/skbuff.h>
> +#include <linux/netdevice.h>
> +#include <linux/if.h>
> +#include <linux/etherdevice.h>
> +#include <linux/if_ether.h>
> +#include <linux/if_packet.h>
> +
> +static inline void skb_bond_set_mac_by_master(struct sk_buff *skb,
> +					      struct net_device *master)
> +{
> +	unsigned char *dest = eth_hdr(skb)->h_dest;
> +
> +	if (compare_ether_addr_64bits(dest, master->dev_addr) &&
> +	    (skb->pkt_type == PACKET_HOST))
> +		memcpy(dest, master->dev_addr, ETH_ALEN);

Just overwriting the dest would be faster, and avoids
to include <linux/etherdevice.h>, maybe a new include file
could be avoided ?

If it is already the master->dev_addr, then memcpy() is a no-op
If it wasnt the master->dev_addr, then memcpy() does what you wanted.

You can also give a hint to gcc as h_dest is guaranteed to be 16 bit aligned

static inline void skb_bond_set_mac_by_master(struct sk_buff *skb,
					      struct net_device *master)
{
	if (skb->pkt_type == PACKET_HOST) {
		u16 *dest = (u16 *)eth_hdr(skb)->h_dest;

		memcpy(dest, master->dev_addr, ETH_ALEN);
	}
}

Compiler will emit better code for memcpy() on some arches.
(not on x86, as it already does one 32bit and one 16bit move)