If udp_recvmsg() constantly releases sk_rmem_alloc
for every read packet, it gives opportunity for
producers to immediately grab spinlocks and desperatly
try adding another packet, causing false sharing.
We can add a simple heuristic to give the signal
by batches of ~25 % of the queue capacity.
This patch considerably increases performance under
flood by about 50 %, since the thread draining the queue
is no longer slowed by false sharing.
Signed-off-by: Eric Dumazet <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
int (*gro_complete)(struct sock *sk,
struct sk_buff *skb,
int nhoff);
+
+ /* This field is dirtied by udp_recvmsg() */
+ int forward_deficit;
};
static inline struct udp_sock *udp_sk(const struct sock *sk)
/* fully reclaim rmem/fwd memory allocated for skb */
static void udp_rmem_release(struct sock *sk, int size, int partial)
{
+ struct udp_sock *up = udp_sk(sk);
int amt;
+ if (likely(partial)) {
+ up->forward_deficit += size;
+ size = up->forward_deficit;
+ if (size < (sk->sk_rcvbuf >> 2) &&
+ !skb_queue_empty(&sk->sk_receive_queue))
+ return;
+ } else {
+ size += up->forward_deficit;
+ }
+ 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);