Re: [PATCH net-next 19/20] net: plip: slight optimization of addr compare

[Joe Perches <joe@perches.com>]

On Fri, 2013-12-27 at 21:12 +0100, Julia Lawall wrote:
> On Fri, 27 Dec 2013, Joe Perches wrote:
> 
> > On Fri, 2013-12-27 at 07:48 -0800, Eric Dumazet wrote:
> > > On Fri, 2013-12-27 at 14:49 +0800, Ding Tianhong wrote:
> > > > Use possibly more efficient ether_addr_equal
> > > > to instead of memcmp.
> > []
> > > > diff --git a/drivers/net/plip/plip.c b/drivers/net/plip/plip.c
> > []
> > > > @@ -549,7 +549,7 @@ static __be16 plip_type_trans(struct sk_buff *skb, struct net_device *dev)
> > > >  
> > > >  	if(*eth->h_dest&1)
> > > >  	{
> > > > -		if(memcmp(eth->h_dest,dev->broadcast, ETH_ALEN)==0)
> > > > +		if(ether_addr_equal(eth->h_dest, dev->broadcast))
> > > >  			skb->pkt_type=PACKET_BROADCAST;
> > > >  		else
> > > >  			skb->pkt_type=PACKET_MULTICAST;
> > > 
> > > What about :
> > > 
> > >         if (is_multicast_ether_addr(eth->h_dest)) {
> > >                 if (ether_addr_equal_64bits(eth->h_dest, dev->broadcast))
> > >                         skb->pkt_type = PACKET_BROADCAST;
> > >                 else
> > >                         skb->pkt_type = PACKET_MULTICAST;
> > >         }
> > 
> > That is better though I wonder how many systems are
> > still using laplink via parallel null-printer cables.
> > 
> > No matter, better is better.
> > 
> > The same optimization using ether_addr_equal_64bits
> > may be possible to do in other places given other
> > structs too.
> > 
> > Perhaps it's a possible spatch/coccinelle conversion,
> > 
> > I don't know spatch well enough to know if a
> > mechanism to check if structure members have other
> > fields that follow them in the structure or if the
> > structure member is an array of a minimum size.
> > 
> > Maybe Julia does.  (cc'd)
> 
> I'm not sure to competely understand the issues.  Could you explain more?

Hi Julia.

Maybe this explanation is helpful?

ethernet addresses are u8[6] (48 bits)

ether_addr_equal_64bits gets passed a pointer to u8[8]
and is more efficient on 64 bit architectures than
ether_addr_equal because the test can be done with a
single compare and shift.

The idea is not to access past the end of the ethernet
address as appropriate (think pointer to eeprom or other
such end-of-addressable memory conditions)

If a struct containing an ethernet address has additional
members after the ethernet address, or the u8[6] address
passed to ether_addr_equal is not going to access past
the end of memory or the structure, then
ether_addr_equal_64bits should be used in lieu of
ether_addr_equal.

Joe