Re: [PATCH]: 1st step to remove skb_linearize() in ip6_tables.c and optimization

[Patrick McHardy <kaber@trash.net>]

Pablo Neira wrote:
> Hi Patrick,
> 
> Patrick McHardy wrote:
> 
>> skb_copy_bits is taking about twice as much cycles as tcp_match
>> and 20% of total cycles.
> 
> AFAIK we are using skb_copy_bits because it walks through fragment list 
> in a safe way, I read an old thread (Harald and DaveM discussion) about 
> this. If I'm not wrong, that's why this was the method selected when 
> conntrack started handling fragments in 2.6 in a different way.

skb_copy_bits does three things: it copies memory from the linear part
between skb->head and skb->tail, it copies memory from the paged part
in skb->frags[], and it walks skb->frag_list for defragmented packets.

> 
> Well, I wonder if we could check if current skb is fragmented (something 
> like skb->data_len!=0), in that case use skb_copy_bits, if not walk 
> through skb payload directly as we use to do in 2.4 and save that memcpy.

I don't think we should do that, the number of copies would still be
dependant on the ruleset. I see two possibilities: We could linearize
the transport protocol headers or we could copy it to skb->cb.

Linearizing has three cases:
- data is already linear: nothing to do
- parts of data are in paged part in skb->frags[]: append data of frags
   to skb->end until we have enough
- parts of data are in frag_list: walk frag_list until enough linear
   data is in first skb, copy linear data, handle paged data like
   described above

I don't know if this turns out to be much more efficient than the old
call to skb_linearize. Copying to skb->cb is simpler, but the frequent
case of already linear data is handled less efficient than with
linearizing.

Regards
Patrick