net: introduce __skb_header_release()

While profiling TCP stack, I noticed one useless atomic operation
in tcp_sendmsg(), caused by skb_header_release().

It turns out all current skb_header_release() users have a fresh skb,
that no other user can see, so we can avoid one atomic operation.

Introduce __skb_header_release() to clearly document this.

This gave me a 1.5 % improvement on TCP_RR workload.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

include/linux/skbuff.h

@@ -1083,6 +1083,7 @@ static inline int skb_header_cloned(const struct sk_buff *skb)
  *	Drop a reference to the header part of the buffer.  This is done
  *	by acquiring a payload reference.  You must not read from the header
  *	part of skb->data after this.
+ *	Note : Check if you can use __skb_header_release() instead.
  */
 static inline void skb_header_release(struct sk_buff *skb)
 {
@@ -1091,6 +1092,20 @@ static inline void skb_header_release(struct sk_buff *skb)
 	atomic_add(1 << SKB_DATAREF_SHIFT, &skb_shinfo(skb)->dataref);
 }
 
+/**
+ *	__skb_header_release - release reference to header
+ *	@skb: buffer to operate on
+ *
+ *	Variant of skb_header_release() assuming skb is private to caller.
+ *	We can avoid one atomic operation.
+ */
+static inline void __skb_header_release(struct sk_buff *skb)
+{
+	skb->nohdr = 1;
+	atomic_set(&skb_shinfo(skb)->dataref, 1 + (1 << SKB_DATAREF_SHIFT));
+}
+
+
 /**
  *	skb_shared - is the buffer shared
  *	@skb: buffer to check

net/core/skbuff.c

@@ -3179,7 +3179,7 @@ int skb_gro_receive(struct sk_buff **head, struct sk_buff *skb)
 	skb_shinfo(nskb)->frag_list = p;
 	skb_shinfo(nskb)->gso_size = pinfo->gso_size;
 	pinfo->gso_size = 0;
-	skb_header_release(p);
+	__skb_header_release(p);
 	NAPI_GRO_CB(nskb)->last = p;
 
 	nskb->data_len += p->len;
@@ -3211,7 +3211,7 @@ int skb_gro_receive(struct sk_buff **head, struct sk_buff *skb)
 	else
 		NAPI_GRO_CB(p)->last->next = skb;
 	NAPI_GRO_CB(p)->last = skb;
-	skb_header_release(skb);
+	__skb_header_release(skb);
 	lp = p;
 
 done:

net/ipv4/tcp.c

@@ -609,7 +609,7 @@ static inline bool forced_push(const struct tcp_sock *tp)
 	return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
 }
 
-static inline void skb_entail(struct sock *sk, struct sk_buff *skb)
+static void skb_entail(struct sock *sk, struct sk_buff *skb)
 {
 	struct tcp_sock *tp = tcp_sk(sk);
 	struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
@@ -618,7 +618,7 @@ static inline void skb_entail(struct sock *sk, struct sk_buff *skb)
 	tcb->seq     = tcb->end_seq = tp->write_seq;
 	tcb->tcp_flags = TCPHDR_ACK;
 	tcb->sacked  = 0;
-	skb_header_release(skb);
+	__skb_header_release(skb);
 	tcp_add_write_queue_tail(sk, skb);
 	sk->sk_wmem_queued += skb->truesize;
 	sk_mem_charge(sk, skb->truesize);

net/ipv4/tcp_output.c

@@ -995,7 +995,7 @@ static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
 
 	/* Advance write_seq and place onto the write_queue. */
 	tp->write_seq = TCP_SKB_CB(skb)->end_seq;
-	skb_header_release(skb);
+	__skb_header_release(skb);
 	tcp_add_write_queue_tail(sk, skb);
 	sk->sk_wmem_queued += skb->truesize;
 	sk_mem_charge(sk, skb->truesize);
@@ -1167,7 +1167,7 @@ int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len,
 	}
 
 	/* Link BUFF into the send queue. */
-	skb_header_release(buff);
+	__skb_header_release(buff);
 	tcp_insert_write_queue_after(skb, buff, sk);
 
 	return 0;
@@ -1671,7 +1671,7 @@ static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
 	tcp_set_skb_tso_segs(sk, buff, mss_now);
 
 	/* Link BUFF into the send queue. */
-	skb_header_release(buff);
+	__skb_header_release(buff);
 	tcp_insert_write_queue_after(skb, buff, sk);
 
 	return 0;
@@ -2772,7 +2772,7 @@ int tcp_send_synack(struct sock *sk)
 			if (nskb == NULL)
 				return -ENOMEM;
 			tcp_unlink_write_queue(skb, sk);
-			skb_header_release(nskb);
+			__skb_header_release(nskb);
 			__tcp_add_write_queue_head(sk, nskb);
 			sk_wmem_free_skb(sk, skb);
 			sk->sk_wmem_queued += nskb->truesize;
@@ -2947,7 +2947,7 @@ static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
 	struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
 
 	tcb->end_seq += skb->len;
-	skb_header_release(skb);
+	__skb_header_release(skb);
 	__tcp_add_write_queue_tail(sk, skb);
 	sk->sk_wmem_queued += skb->truesize;
 	sk_mem_charge(sk, skb->truesize);