udp: elide zerocopy operation in hot path

With MSG_ZEROCOPY, each skb holds a reference to a struct ubuf_info.
Release of its last reference triggers a completion notification.

The TCP stack in tcp_sendmsg_locked holds an extra ref independent of
the skbs, because it can build, send and free skbs within its loop,
possibly reaching refcount zero and freeing the ubuf_info too soon.

The UDP stack currently also takes this extra ref, but does not need
it as all skbs are sent after return from __ip(6)_append_data.

Avoid the extra refcount_inc and refcount_dec_and_test, and generally
the sock_zerocopy_put in the common path, by passing the initial
reference to the first skb.

This approach is taken instead of initializing the refcount to 0, as
that would generate error "refcount_t: increment on 0" on the
next skb_zcopy_set.

Changes
  v3 -> v4
    - Move skb_zcopy_set below the only kfree_skb that might cause
      a premature uarg destroy before skb_zerocopy_put_abort
      - Move the entire skb_shinfo assignment block, to keep that
        cacheline access in one place

Signed-off-by: Willem de Bruijn <willemb@google.com>
Acked-by: Paolo Abeni <pabeni@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

5 files changed, 36 insertions, 31 deletions

diff --git a/include/linux/skbuff.h b/include/linux/skbuff.h
index 04f52e719571..75d50ab7997c 100644
--- a/include/linux/skbuff.h
+++ b/include/linux/skbuff.h
@@ -481,7 +481,7 @@ static inline void sock_zerocopy_get(struct ubuf_info *uarg)
 }
 
 void sock_zerocopy_put(struct ubuf_info *uarg);
-void sock_zerocopy_put_abort(struct ubuf_info *uarg);
+void sock_zerocopy_put_abort(struct ubuf_info *uarg, bool have_uref);
 
 void sock_zerocopy_callback(struct ubuf_info *uarg, bool success);
 
@@ -1326,10 +1326,14 @@ static inline struct ubuf_info *skb_zcopy(struct sk_buff *skb)
 	return is_zcopy ? skb_uarg(skb) : NULL;
 }
 
-static inline void skb_zcopy_set(struct sk_buff *skb, struct ubuf_info *uarg)
+static inline void skb_zcopy_set(struct sk_buff *skb, struct ubuf_info *uarg,
+				 bool *have_ref)
 {
 	if (skb && uarg && !skb_zcopy(skb)) {
-		sock_zerocopy_get(uarg);
+		if (unlikely(have_ref && *have_ref))
+			*have_ref = false;
+		else
+			sock_zerocopy_get(uarg);
 		skb_shinfo(skb)->destructor_arg = uarg;
 		skb_shinfo(skb)->tx_flags |= SKBTX_ZEROCOPY_FRAG;
 	}
@@ -1374,7 +1378,7 @@ static inline void skb_zcopy_abort(struct sk_buff *skb)
 	struct ubuf_info *uarg = skb_zcopy(skb);
 
 	if (uarg) {
-		sock_zerocopy_put_abort(uarg);
+		sock_zerocopy_put_abort(uarg, false);
 		skb_shinfo(skb)->tx_flags &= ~SKBTX_ZEROCOPY_FRAG;
 	}
 }
diff --git a/net/core/skbuff.c b/net/core/skbuff.c
index 1350901c5cb8..c78ce114537e 100644
--- a/net/core/skbuff.c
+++ b/net/core/skbuff.c
@@ -1089,7 +1089,7 @@ void sock_zerocopy_put(struct ubuf_info *uarg)
 }
 EXPORT_SYMBOL_GPL(sock_zerocopy_put);
 
-void sock_zerocopy_put_abort(struct ubuf_info *uarg)
+void sock_zerocopy_put_abort(struct ubuf_info *uarg, bool have_uref)
 {
 	if (uarg) {
 		struct sock *sk = skb_from_uarg(uarg)->sk;
@@ -1097,7 +1097,8 @@ void sock_zerocopy_put_abort(struct ubuf_info *uarg)
 		atomic_dec(&sk->sk_zckey);
 		uarg->len--;
 
-		sock_zerocopy_put(uarg);
+		if (have_uref)
+			sock_zerocopy_put(uarg);
 	}
 }
 EXPORT_SYMBOL_GPL(sock_zerocopy_put_abort);
@@ -1137,7 +1138,7 @@ int skb_zerocopy_iter_stream(struct sock *sk, struct sk_buff *skb,
 		return err;
 	}
 
-	skb_zcopy_set(skb, uarg);
+	skb_zcopy_set(skb, uarg, NULL);
 	return skb->len - orig_len;
 }
 EXPORT_SYMBOL_GPL(skb_zerocopy_iter_stream);
@@ -1157,7 +1158,7 @@ static int skb_zerocopy_clone(struct sk_buff *nskb, struct sk_buff *orig,
 			if (skb_copy_ubufs(nskb, GFP_ATOMIC))
 				return -EIO;
 		}
-		skb_zcopy_set(nskb, skb_uarg(orig));
+		skb_zcopy_set(nskb, skb_uarg(orig), NULL);
 	}
 	return 0;
 }
diff --git a/net/ipv4/ip_output.c b/net/ipv4/ip_output.c
index 6f843aff628c..78f028bdad30 100644
--- a/net/ipv4/ip_output.c
+++ b/net/ipv4/ip_output.c
@@ -881,8 +881,8 @@ static int __ip_append_data(struct sock *sk,
 	int csummode = CHECKSUM_NONE;
 	struct rtable *rt = (struct rtable *)cork->dst;
 	unsigned int wmem_alloc_delta = 0;
+	bool paged, extra_uref;
 	u32 tskey = 0;
-	bool paged;
 
 	skb = skb_peek_tail(queue);
 
@@ -921,12 +921,13 @@ static int __ip_append_data(struct sock *sk,
 		uarg = sock_zerocopy_realloc(sk, length, skb_zcopy(skb));
 		if (!uarg)
 			return -ENOBUFS;
+		extra_uref = true;
 		if (rt->dst.dev->features & NETIF_F_SG &&
 		    csummode == CHECKSUM_PARTIAL) {
 			paged = true;
 		} else {
 			uarg->zerocopy = 0;
-			skb_zcopy_set(skb, uarg);
+			skb_zcopy_set(skb, uarg, &extra_uref);
 		}
 	}
 
@@ -1015,13 +1016,6 @@ alloc_new_skb:
 			skb->csum = 0;
 			skb_reserve(skb, hh_len);
 
-			/* only the initial fragment is time stamped */
-			skb_shinfo(skb)->tx_flags = cork->tx_flags;
-			cork->tx_flags = 0;
-			skb_shinfo(skb)->tskey = tskey;
-			tskey = 0;
-			skb_zcopy_set(skb, uarg);
-
 			/*
 			 *	Find where to start putting bytes.
 			 */
@@ -1054,6 +1048,13 @@ alloc_new_skb:
 			exthdrlen = 0;
 			csummode = CHECKSUM_NONE;
 
+			/* only the initial fragment is time stamped */
+			skb_shinfo(skb)->tx_flags = cork->tx_flags;
+			cork->tx_flags = 0;
+			skb_shinfo(skb)->tskey = tskey;
+			tskey = 0;
+			skb_zcopy_set(skb, uarg, &extra_uref);
+
 			if ((flags & MSG_CONFIRM) && !skb_prev)
 				skb_set_dst_pending_confirm(skb, 1);
 
@@ -1124,13 +1125,12 @@ alloc_new_skb:
 
 	if (wmem_alloc_delta)
 		refcount_add(wmem_alloc_delta, &sk->sk_wmem_alloc);
-	sock_zerocopy_put(uarg);
 	return 0;
 
 error_efault:
 	err = -EFAULT;
 error:
-	sock_zerocopy_put_abort(uarg);
+	sock_zerocopy_put_abort(uarg, extra_uref);
 	cork->length -= length;
 	IP_INC_STATS(sock_net(sk), IPSTATS_MIB_OUTDISCARDS);
 	refcount_add(wmem_alloc_delta, &sk->sk_wmem_alloc);
diff --git a/net/ipv4/tcp.c b/net/ipv4/tcp.c
index 215e4d3b3616..dc68c408bba0 100644
--- a/net/ipv4/tcp.c
+++ b/net/ipv4/tcp.c
@@ -1423,7 +1423,7 @@ do_error:
 	if (copied + copied_syn)
 		goto out;
 out_err:
-	sock_zerocopy_put_abort(uarg);
+	sock_zerocopy_put_abort(uarg, true);
 	err = sk_stream_error(sk, flags, err);
 	/* make sure we wake any epoll edge trigger waiter */
 	if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 &&
diff --git a/net/ipv6/ip6_output.c b/net/ipv6/ip6_output.c
index 7df04d20a91f..ec8c235ea891 100644
--- a/net/ipv6/ip6_output.c
+++ b/net/ipv6/ip6_output.c
@@ -1258,7 +1258,7 @@ static int __ip6_append_data(struct sock *sk,
 	int csummode = CHECKSUM_NONE;
 	unsigned int maxnonfragsize, headersize;
 	unsigned int wmem_alloc_delta = 0;
-	bool paged;
+	bool paged, extra_uref;
 
 	skb = skb_peek_tail(queue);
 	if (!skb) {
@@ -1327,12 +1327,13 @@ emsgsize:
 		uarg = sock_zerocopy_realloc(sk, length, skb_zcopy(skb));
 		if (!uarg)
 			return -ENOBUFS;
+		extra_uref = true;
 		if (rt->dst.dev->features & NETIF_F_SG &&
 		    csummode == CHECKSUM_PARTIAL) {
 			paged = true;
 		} else {
 			uarg->zerocopy = 0;
-			skb_zcopy_set(skb, uarg);
+			skb_zcopy_set(skb, uarg, &extra_uref);
 		}
 	}
 
@@ -1454,13 +1455,6 @@ alloc_new_skb:
 			skb_reserve(skb, hh_len + sizeof(struct frag_hdr) +
 				    dst_exthdrlen);
 
-			/* Only the initial fragment is time stamped */
-			skb_shinfo(skb)->tx_flags = cork->tx_flags;
-			cork->tx_flags = 0;
-			skb_shinfo(skb)->tskey = tskey;
-			tskey = 0;
-			skb_zcopy_set(skb, uarg);
-
 			/*
 			 *	Find where to start putting bytes
 			 */
@@ -1492,6 +1486,13 @@ alloc_new_skb:
 			exthdrlen = 0;
 			dst_exthdrlen = 0;
 
+			/* Only the initial fragment is time stamped */
+			skb_shinfo(skb)->tx_flags = cork->tx_flags;
+			cork->tx_flags = 0;
+			skb_shinfo(skb)->tskey = tskey;
+			tskey = 0;
+			skb_zcopy_set(skb, uarg, &extra_uref);
+
 			if ((flags & MSG_CONFIRM) && !skb_prev)
 				skb_set_dst_pending_confirm(skb, 1);
 
@@ -1562,13 +1563,12 @@ alloc_new_skb:
 
 	if (wmem_alloc_delta)
 		refcount_add(wmem_alloc_delta, &sk->sk_wmem_alloc);
-	sock_zerocopy_put(uarg);
 	return 0;
 
 error_efault:
 	err = -EFAULT;
 error:
-	sock_zerocopy_put_abort(uarg);
+	sock_zerocopy_put_abort(uarg, extra_uref);
 	cork->length -= length;
 	IP6_INC_STATS(sock_net(sk), rt->rt6i_idev, IPSTATS_MIB_OUTDISCARDS);
 	refcount_add(wmem_alloc_delta, &sk->sk_wmem_alloc);