On Sat, 14 Nov 2026 at 05:16, Richard Henderson
<[email protected]> wrote:
>
> On 11/12/26 11:02, Peter Maydell wrote:
> > In net_checksum_calculate() we calculate the position of the TCP or
> > UDP header in the buffer with:
> >    tcp_header *tcp = (tcp_header *)(ip + 1);
> > where ip is a 'struct ip_header' (and similarly for udp).
> >
> > This only gives the correct location if the IP packet has no options
> > and its header is the minimum length.  Otherwise we will corrupt some
> > part of the packet in writing the checksum word.
> >
> > Fortunately packets with IP options are rare (and often get blocked
> > by routers or firewalls), so this probably didn't often cause
> > problems, and the incorrect offset we write to is always inside the
> > packet, not out-of-bounds.
> >
> > Use the IP header length value from the packet to locate the TCP or
> > UDP header, instead of the fixed minimum size.  While we're adjusting
> > this code, make sure that we don't assign to the tcp_header or
> > udp_header pointer until after we've made the length check that the
> > buffer is big enough (the previous ordering of operations was
> > harmless as we never used the resulting pointer, but not ideal).
> >
> > This was mentioned in passing in issue #4791 (which is otherwise
> > a duplicate of #4720).
> >
> > Cc: [email protected]
> > Fixes: 50dbce6538d1 ("net: improve UDP/TCP checksum computation.")
> > Reported-by: Alex Meiburg <https://gitlab.com/timeroot.alex>
> > Signed-off-by: Peter Maydell <[email protected]>
> > ---
> >   net/checksum.c | 14 ++++++++++----
> >   1 file changed, 10 insertions(+), 4 deletions(-)
> >
> > diff --git a/net/checksum.c b/net/checksum.c
> > index 997a0dd2350..277b2e652d9 100644
> > --- a/net/checksum.c
> > +++ b/net/checksum.c
> > @@ -62,6 +62,7 @@ void net_checksum_calculate(void *data, int length, int 
> > csum_flag)
> >       int mac_hdr_len, ip_len, ip_hdr_len;
> >       struct ip_header *ip;
> >       uint16_t csum;
> > +    uint8_t *ip_payload;
>
> Why uint8_t*z and not void*?

Mostly because it is a byte buffer that we're dealing with.
But yes, a void* would avoid the cast.

> This could also plausibly be reordered to
>
>          case IP_PROTO_TCP:
>              if ((csum & CSUM_TCP) && ip_len >= sizeof(tcg_header)) {
>                  tcp_header *tcp = ip_payload;
>                  ...
>              }

I think I prefer the "early return" style here personally.

thanks
-- PMM

Reply via email to