Hello, On 28/11/18 - 19:15:12, Eric Dumazet wrote: > On Wed, Nov 28, 2018 at 7:09 PM Jana Iyengar <jri.i...@gmail.com> wrote: > > > > > > On Wed, Nov 28, 2018 at 6:19 PM Eric Dumazet <eduma...@google.com> wrote: > >> > >> On Wed, Nov 28, 2018 at 5:57 PM Christoph Paasch <cpaa...@apple.com> wrote: > >> > > >> > There are use-cases where a host wants to use a UDP socket with a > >> > specific 4-tuple. The way to do this is to bind() and then connect() the > >> > socket. However, after the bind(), the socket starts receiving data even > >> > if it does not match the intended 4-tuple. That is because after the > >> > bind() UDP-socket will match in the lookup for all incoming UDP-traffic > >> > that has the specific IP/port. > >> > > >> > This patch prevents any incoming traffic until the connect() system-call > >> > is called whenever the app sets the UDP socket-option > >> > UDP_WAIT_FOR_CONNECT. > >> > >> Please do not add something that could mislead applications writers to > >> think UDP stack can scale. > > > > > >> UDP stack does not have a full hash on 4-tuples, it means that > >> incoming traffic on a 'shared port' has > >> to scan a list of XXX sockets to find the best match ... > > > > > >> Also you add another cache line miss in UDP lookup to access > >> udp_sk()->wait_for_connect.
Fair enough. We could add the socket later to the hash (see below). > >> > >> recvfrom() can be used to filter whatever frame that came before the > >> connect() > > > > > > I don't think I understand that argument -- connect() is supported for UDP > > sockets, and UDP sockets are being used for serving QUIC traffic. Are you > > suggesting that connect() never be used? > > If the source port is not shared, Christoph patch is not needed. > > If it is shared, then a whole can of worm is opened. > > Trying to hack UDP stack while it is not fully 4-tuple ready is not > going to fly. Indeed, the UDP-stack is not fully 4-tuple ready. What are your thoughts on getting it there? Should be doable by simply using a similar approach as TCP, no? Any caveats you see with that? Then, when one sets the "wait-for-connect"-flag we would add the socket to the hash-table only at connect()-time also addressing the cache-line miss you mentioned above. Thanks, Christoph