K 10
svn:author
V 5
luigi
K 8
svn:date
V 27
2009-12-27T19:17:03.207078Z
K 7
svn:log
V 1886
Historically, BSD keeps ip_len and ip_off in host format when doing
layer 3 processing.  This often requires to translate the format
back and forth.

I suppose the original motivation was to avoid carrying additional
metadata, and/or save a couple of local variables to store off and len.

My impression is that nowadays there are much bigger disadvantages:
- many unnecessary writes to the mbuf, which make it harder to
  share the data, and also cause cache traffic;
- visual clutter in the code;
- portability problems, either across layers (e.g. we cannot
  use the same code to do deep packet inspection for L2 and L3
  packets) or across operating systems;
- potential source of confusion, because the format is different
  in different parts of the code.

Eventually, I would like these fields to remain in network format
across the lifetime of a packet, but this may take a long time,
so let's see if we can find ways to at least make the switch easier.

Here I define a couple of macros that do the format conversion
back and forth. On BIG_ENDIAN machines, or when we will switch,
they become no-ops.

At the moment I am only using these macros in the ipfw code,
I am not sure this is the best solution across the entire stack.

    #if (BYTE_ORDER == BIG_ENDIAN) || defined(HAVE_NET_IPLEN)
    #define SET_NET_IPLEN(p)       do {} while (0)
    #define SET_HOST_IPLEN(p)      do {} while (0)
    #else
    #define SET_NET_IPLEN(p)       do {         \
        struct ip *h_ip = (p);                  \
        h_ip->ip_len = htons(h_ip->ip_len);     \
        h_ip->ip_off = htons(h_ip->ip_off);     \
        } while (0)

    #define SET_HOST_IPLEN(p)      do {         \
        struct ip *h_ip = (p);                  \
        h_ip->ip_len = ntohs(h_ip->ip_len);     \
        h_ip->ip_off = ntohs(h_ip->ip_off);     \
        } while (0)
    #endif /* !HAVE_NET_IPLEN */

END