On 19/04/16 15:50, Eric Dumazet wrote: > The main problem in UDP stack today is having to lock the socket because > of the dumb forward allocation problem. I'm not quite sure what you're referring to here, care to educate me?
> Are you really going to provide > a list of skbs up to _one_ UDP socket ? In principle we should be able to take it that far, yes. AFAICT the socket already has a receive queue that we end up appending the packet to (and which I presume the recvmsg() syscall pulls from), I don't see why we couldn't just splice a list of skbs on the end rather than appending them one by one. Thus amortising looking up the socket.
