Re: [PATCH] tcp: gso: do not generate out of order packets

[Eric Dumazet <eric.dumazet@gmail.com>]

On Wed, 2013-05-15 at 20:58 -0700, Maciej Żenczykowski wrote:
> I'd be worried that calling the destructor that many times would cause
> performance problems
> (and only call the destructor and do memory accounting on the last segment).
> 
> Could we instead move the queue mapping into the skb somehow instead?

There is no performance problem, because all these packets are going to
be freed at the same time from TX completion handler.

The socket cache lines we use (sk->sk_state, sk->sk_wmem_alloc) are hot.

Anyway its way cleaner propagating socket information, as it might be
needed by netfilter or classifiers.

Some months ago I tested a variant of sock_wfree() not testing
sk>sk_state as TCP sockets set SOCK_USE_WRITE_QUEUE flag, and I had no
change in performance. sk_state & sk_wmem_alloc were on separate cache
lines :

offsetof(struct sock, sk_flags) = 0xe8
offsetof(struct sock, sk_wmem_alloc) = 0x104

void sock_fast_wfree(struct sk_buff *skb)
{
        struct sock *sk = skb->sk;

        if (atomic_sub_and_test(skb->truesize, &sk->sk_wmem_alloc))
                __sk_free(sk);
}