Question about more headroom in skb

Question about more headroom in skb

[Sharat Masetty]

Hello All,

For my project I need 3 words of headroom in the skb in the network driver level, to add a  custom header to the ethernet packet. I looked into the tcp code and figured out tcp uses sk->sk_prot->max_header for header allocation size. But I was not able to confirm that all other transport protocol use the same mechanism(?) For example in UDP/ICMP I was not able to figure out from the code where the allocation and header reservation happens(Any light here would be really helpful.)

I have also looked at an API in skbuff skb_pad() which does what I want(add either headroom or tailroom), but I want to avoid that for performance reasons(skb_pad does kmalloc and memcpy). I want to figure out a good way(may be tune some parameters) to allocate extra 3 words for any skbuff independant of the transport protocol being used. Any light here would be very much appreciated.

Thanks,
Sharat.


      

Re: Question about more headroom in skb

[Eric Dumazet]

Le lundi 10 mai 2010 à 13:09 -0700, Sharat Masetty a écrit :
> Hello All,

Please dont use too long lines

> 
> For my project I need 3 words of headroom in the skb in the network

>  driver level, to add a  custom header to the ethernet packet. I 

> looked into the tcp code and figured out tcp uses sk->sk_prot->max_header 

> for header allocation size. But I was not able to confirm that all other 

> transport protocol use the same mechanism(?) For example in UDP/ICMP I was 

> not able to figure out from the code where the allocation and header

>  reservation happens(Any light here would be really helpful.)
> 
> I have also looked at an API in skbuff skb_pad() which does what I want

> (add either headroom or tailroom), but I want to avoid that for performance

>  reasons(skb_pad does kmalloc and memcpy). I want to figure out a good way

> (may be tune some parameters) to allocate extra 3 words for any skbuff 

> independant of the transport protocol being used. 

> Any light here would be very much appreciated.

LL_RESERVED_SPACE() is the magic you need.

#define LL_RESERVED_SPACE(dev) \
	((((dev)->hard_header_len+(dev)->needed_headroom)&~(HH_DATA_MOD - 1)) + HH_DATA_MOD)

sendmsg() -> ip_append_data()
...
hh_len = LL_RESERVED_SPACE(rt->u.dst.dev);
fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
...

                if (transhdrlen) {
                        skb = sock_alloc_send_skb(sk,
                                        alloclen + hh_len + 15,
                                        (flags & MSG_DONTWAIT), &err);
                } else {
                        skb = NULL;
                        if (atomic_read(&sk->sk_wmem_alloc) <=
                            2 * sk->sk_sndbuf)
                                skb = sock_wmalloc(sk,
                                                   alloclen + hh_len + 15, 1,
                                                   sk->sk_allocation);