On Fri, 2017-03-31 at 08:09 -0700, Eric Dumazet wrote: > On Fri, 2017-03-31 at 16:33 +0200, Paolo Abeni wrote: > > > I did the above to avoid increasing the udp_sock struct size; this will > > costs more than a whole cacheline. > > Yes, but who cares :) > > Also note that we discussed about having a secondary receive queue in > the future, to decouple the fact that producers/consumer have to grab a > contended spinlock for every enqueued and dequeued packet. > > With a secondary queue, the consumer can transfer one queue into another > in one batch. > > Or simply use ptr_ring / skb_array now these infras are available thanks > to Michael. > > So we will likely increase UDP socket size in a near future... > > > > > I did not hit others false sharing issues because: > > - gro_receive/gro_complete are touched only for packets coming from > > devices with udp tunnel offload enabled, that hit the tunnel offload > > path on the nic; such packets will most probably land in the udp tunnel > > and will not use 'forward_deficit' > > > > - encap_destroy is touched only socket shutdown > > - encap_rcv is protected by the 'udp_encap_needed' static key > > > > I think this latter is problematic, so I'm ok with the patch you > > suggested. > > > > The above change could still make sense, the udp code is already > > checking for udplite sockets with either pcflag and protocol; > > testing always the same data will make the code more cleaner. > > Where are we testing sk->sk_prototocol in receive path ?
Sorry, I was ambiguous: sk->sk_protocol is not used yet; before the socket lockup, __udp4_lib_rcv() and __udp6_lib_rcv() use the protocol number provided by the caller to properly account udp vs udplite stats. Cheers, Paolo