Re: Unix Socket buffer attribution

[Yannick Koehler <yannick@koehler.name>]

I did some more research, I found out that netlink and
sock_queue_err_skb does the same trick that I claim to be missing
under the net/unix/af_unix.c.  After adding the code, I got a problem
since af_unix.c "_recvmsg()" functions assume that the skb->sk is
holding the peer socket not the current one related to the receive
skb.  It extract the sun_path name from it.  Since with UDP each
packet may have a different peer, the only solution I found was to use
the skb control block to hold the peer name.  The problem is that this
cb member is 48 bytes in length and sun_path is 108 bytes in size.  So
I had to increase it from 48 to 160 bytes.

This obviously increase the cost of an SKB struct, so I do not really
like this solution. But at least it seems to prove my point and now,
with this I can have my clients working except the one mis-behaving.
I am attaching a patch.

I also saw this thread that would be somehow related:

http://www.mail-archive.com/netdev@vger.kernel.org/msg20195.html

Basically since we now count the buffer size into the socket who
receives it we can safely use this check instead of looking at the
receive queue size:

   if (atomic_read(&other->sk_rmem_alloc) + skb->truesize >=
(unsigned)other->sk_rcvbuf) {

So anyway I sent my patch, and I am awaiting for comments on how to improve it.

--
Yannick Koehler