Otto Moerbeek <[email protected]> wrote:

> > right. So I have tried my patch (without the memcpy dance) on sparc64
> > over udp and tcp and I have also tracked this down in the code. This
> > should be fine as is. ar->ar_data comes directly out of malloc
> > (reallocarray) in ensure_ibuf() and the struct is defined thusly:
> > 
> > struct asr_dns_header {
> >         uint16_t        id;
> >         uint16_t        flags;
> >         uint16_t        qdcount;
> >         uint16_t        ancount;
> >         uint16_t        nscount;
> >         uint16_t        arcount;
> > };
> > 
> 
> So that is indeed safe as long as nobody starts allocating packet
> buffers in different ways,

The common problem is in a libpcap parser, finding it in a packet.
There, the 14-byte ethernet header comes into play, offsetting everything
by 2 bytes. However, this struct has 2-byte alignment, so the compiler will
generate safe alignment accesses.

The problem really shows itself when you find per-byte packed encapsulation
headers, due to an insufficient number of axe murderers lurking the halls
of IETF.  That would place these 2-byte objects at 1-byte alignment, and the
compiler generated 2-byte instructions bomb.  Such encapsulations are far
more rare.  The major libpcap parser (tcpdump) avoids this by always considering
packets byte aligned..

Reply via email to