On Thu, Dec 8, 2016 at 10:38 AM, Eric Dumazet <eduma...@google.com> wrote: > On Thu, Dec 8, 2016 at 10:36 AM, Eric Dumazet <eduma...@google.com> wrote: >> On Thu, Dec 8, 2016 at 10:24 AM, Paolo Abeni <pab...@redhat.com> wrote: >> >>> Nice one! This sounds like a relevant improvement! >>> >>> I'm wondering if it may cause regressions with small value of >>> sk_rcvbuf ?!? e.g. with: >>> >>> netperf -t UDP_STREAM -H 127.0.0.1 -- -s 1280 -S 1280 -m 1024 -M 1024 >>> >> >> Possibly, then simply we can refine the test to : >> >> size = up->forward_deficit; >> if (size < (sk->sk_rcvbuf >> 2) && !skb_queue_empty(sk->sk_receive_buf)) >> return; >
I will also add this patch : This really makes sure our changes to sk_forward_alloc wont be slowed because producers see the change to sk_rmem_alloc too soon. diff --git a/net/ipv4/udp.c b/net/ipv4/udp.c index 8400d6954558..6bdcbe103390 100644 --- a/net/ipv4/udp.c +++ b/net/ipv4/udp.c @@ -1191,13 +1191,14 @@ static void udp_rmem_release(struct sock *sk, int size, int partial) } up->forward_deficit = 0; - atomic_sub(size, &sk->sk_rmem_alloc); sk->sk_forward_alloc += size; amt = (sk->sk_forward_alloc - partial) & ~(SK_MEM_QUANTUM - 1); sk->sk_forward_alloc -= amt; if (amt) __sk_mem_reduce_allocated(sk, amt >> SK_MEM_QUANTUM_SHIFT); + + atomic_sub(size, &sk->sk_rmem_alloc); } /* Note: called with sk_receive_queue.lock held.