On Wed, Dec 03, 2014 at 14:56 +0100, Mike Belopuhov wrote:
> bpf aligns data following the datalink header (e.g. ethernet)
> on the BPF_ALIGNMENT boundary. Since rev1.41 of bpf.h it's
> uint32_t instead of a long. And also since then almost all
> packets become "unaligned" from the tcpdump perspective and
> require costly copies into the internal buffer. Neither IP
> header (struct ip) nor IPv6 (struct ip6_hdr) have fields
> larger than 32 bits and therefore alignment requirements for
> them are at most 32 bit.
>
> I've successfully tested this diff on sparc64 and amd64.
>
> OK?
>
Still looking for OKs.
> diff --git usr.sbin/tcpdump/print-ip.c usr.sbin/tcpdump/print-ip.c
> index 1839869..60946a1 100644
> --- usr.sbin/tcpdump/print-ip.c
> +++ usr.sbin/tcpdump/print-ip.c
> @@ -368,11 +368,11 @@ ip_print(register const u_char *bp, register u_int
> length)
> /*
> * If the IP header is not aligned, copy into abuf.
> * This will never happen with BPF. It does happen with raw packet
> * dumps from -r.
> */
> - if ((intptr_t)ip & (sizeof(long)-1)) {
> + if ((intptr_t)ip & (sizeof(u_int32_t)-1)) {
> static u_char *abuf = NULL;
> static int didwarn = 0;
> int clen = snapend - bp;
>
> if (clen > snaplen)
> diff --git usr.sbin/tcpdump/print-ip6.c usr.sbin/tcpdump/print-ip6.c
> index cca8495..ea3a9b3 100644
> --- usr.sbin/tcpdump/print-ip6.c
> +++ usr.sbin/tcpdump/print-ip6.c
> @@ -70,11 +70,11 @@ ip6_print(register const u_char *bp, register u_int
> length)
> /*
> * The IP header is not word aligned, so copy into abuf.
> * This will never happen with BPF. It does happen with
> * raw packet dumps from -r.
> */
> - if ((intptr_t)ip6 & (sizeof(long)-1)) {
> + if ((intptr_t)ip6 & (sizeof(u_int32_t)-1)) {
> static u_char *abuf = NULL;
> static int didwarn = 0;
> int clen = snapend - bp;
> if (clen > snaplen)
> clen = snaplen;
