@@ -1259,6 +1259,7 @@ static int __ip6_append_data(struct sock *sk,
struct ipv6_txoptions *opt = v6_cork->opt;
int csummode = CHECKSUM_NONE;
unsigned int maxnonfragsize, headersize;
+ unsigned int wmem_alloc_delta = 0;
skb = skb_peek_tail(queue);
if (!skb) {
@@ -1411,11 +1412,10 @@ static int __ip6_append_data(struct sock *sk,
(flags & MSG_DONTWAIT), &err);
} else {
skb = NULL;
- if (refcount_read(&sk->sk_wmem_alloc) <=
+ if (refcount_read(&sk->sk_wmem_alloc) + wmem_alloc_delta <=
2 * sk->sk_sndbuf)
- skb = sock_wmalloc(sk,
- alloclen + hh_len, 1,
- sk->sk_allocation);
+ skb = alloc_skb(alloclen + hh_len,
+ sk->sk_allocation);
if (unlikely(!skb))
err = -ENOBUFS;
}
@@ -1474,6 +1474,11 @@ static int __ip6_append_data(struct sock *sk,
/*
* Put the packet on the pending queue
*/
+ if (!skb->destructor) {
+ skb->destructor = sock_wfree;
+ skb->sk = sk;
+ wmem_alloc_delta += skb->truesize;
+ }
__skb_queue_tail(queue, skb);
continue;
}
@@ -1520,12 +1525,13 @@ static int __ip6_append_data(struct sock *sk,
skb->len += copy;
skb->data_len += copy;
skb->truesize += copy;
- refcount_add(copy, &sk->sk_wmem_alloc);
+ wmem_alloc_delta += copy;
}
offset += copy;
length -= copy;
}
+ refcount_add(wmem_alloc_delta, &sk->sk_wmem_alloc);
return 0;
error_efault:
@@ -1533,6 +1539,7 @@ static int __ip6_append_data(struct sock *sk,
error:
cork->length -= length;
IP6_INC_STATS(sock_net(sk), rt->rt6i_idev, IPSTATS_MIB_OUTDISCARDS);
+ refcount_add(wmem_alloc_delta, &sk->sk_wmem_alloc);
return err;
}
While testing my inet defrag changes, I found that the senders could spend ~20% of cpu cycles in skb_set_owner_w() updating sk->sk_wmem_alloc for every fragment they cook, competing with TX completion of prior skbs possibly happening on another cpus. The solution to this problem is to use alloc_skb() instead of sock_wmalloc() and manually perform a single sk_wmem_alloc change. Signed-off-by: Eric Dumazet <edumazet@google.com> --- net/ipv6/ip6_output.c | 17 ++++++++++++----- 1 file changed, 12 insertions(+), 5 deletions(-)