Re: [RFC] endianness bugs in net/batman-adv/

On Fri, Apr 13, 2012 at 08:46:29 +0100, Al Viro wrote:

	Let's start with net/batman-adv/translation-table.c:send_tt_request(). 
Its tt_crc argument gets stored into skb as-is and skb goes on the wire.
OK, so it's fixed-endian, right?

ok, it's a bug. The tt_crc field must be stored in the skb by using htons().

	That sucker comes straight from the (only) caller - tt_update_orig().
There it gets compared with ->tt_crc of struct orig_node instances.  Fine,
except that just prior to that comparison we assign to ->tt_crc the return
value of tt_global_crc().  Which is built on crc16_byte() and clearly
returns a host-endian value.  Additionally, orig_node ->tt_crc is getting
compared to tt_request->tt_data in send_other_tt_response(), which ultimately
comes from recv_tt_query() where it's flipped from net-endian to host-endian.

We want to do every computation using host-endian data. There is no "fixed
endian" anywhere. As I wrote above, we forgot to use htons() before sending the
tt_crc over the wire.

	It gets even funnier - we have 3 structures with ->tt_crc in them;
one is struct orig_node (see above), another is struct batman_ogv_packet and
then there's the weirdest one - struct bat_priv.  Where ->tt_crc is
atomic_t, of all things.  With exactly two things ever done to it:
        batman_ogm_packet->tt_crc = htons((uint16_t)
                                                atomic_read(&bat_priv->tt_crc));
in bat_iv_ogm_schedule() and
        atomic_set(&bat_priv->tt_crc, tt_local_crc(bat_priv));
in prepare_packet_buffer().  What the hell does that have to do with atomic_t?

Thank you for spotting this. It was defined as atomic_t in the early development
phase of this new tt framework, but, then, I'd say that we forgot to convert it
to uint16_t once atomic_t was not needed anymore.

At least that one is definitely host-endian all along (tt_local_crc() is
the same kind of built-on-crc16_byte() thing).

	And then there's batman_ogv_packet, where we flip the damn field
from net-endian to host-endian and back.  That's where the argument of
tt_update_orig() comes from, AFAICS always in host-endian form.

	IOW, unless I'm misreading that code we have
bat_priv ->tt_crc: host-endian, no need to make it atomic_t

I agree.

orig_node ->tt_crc: host-endian

I agree. As I said before we want to use host-endian everywhere. We just want to
convert the data to net-endian before sending it over the wire (like people
should normally do).

tt_update_orig()/send_tt_request() tt_crc argument: host-endian

I agree.

the value put into the packet in send_tt_request(): broken; should be
net-endian, in reality it's host-endian.  Missing htons() at the very
least.

Exactly. This is the bug I was talking about in my first inline response.

All the problems come from a missing htons() before sending the tt_request
packet.

Thank you very much. I'll fix it.


Best regards,


-- 
Antonio Quartulli

..each of us alone is worth nothing..
Ernesto "Che" Guevara