tcp: provide timestamps for partial writes
authorSoheil Hassas Yeganeh <soheil@google.com>
Wed, 4 Jan 2017 16:19:34 +0000 (11:19 -0500)
committerDavid S. Miller <davem@davemloft.net>
Thu, 5 Jan 2017 19:56:16 +0000 (14:56 -0500)
For TCP sockets, TX timestamps are only captured when the user data
is successfully and fully written to the socket. In many cases,
however, TCP writes can be partial for which no timestamp is
collected.

Collect timestamps whenever any user data is (fully or partially)
copied into the socket. Pass tcp_write_queue_tail to tcp_tx_timestamp
instead of the local skb pointer since it can be set to NULL on
the error path.

Note that tcp_write_queue_tail can be NULL, even if bytes have been
copied to the socket. This is because acknowledgements are being
processed in tcp_sendmsg(), and by the time tcp_tx_timestamp is
called tcp_write_queue_tail can be NULL. For such cases, this patch
does not collect any timestamps (i.e., it is best-effort).

This patch is written with suggestions from Willem de Bruijn and
Eric Dumazet.

Change-log V1 -> V2:
- Use sockc.tsflags instead of sk->sk_tsflags.
- Use the same code path for normal writes and errors.

Signed-off-by: Soheil Hassas Yeganeh <soheil@google.com>
Acked-by: Yuchung Cheng <ycheng@google.com>
Cc: Willem de Bruijn <willemb@google.com>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Neal Cardwell <ncardwell@google.com>
Cc: Martin KaFai Lau <kafai@fb.com>
Acked-by: Willem de Bruijn <willemb@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
net/ipv4/tcp.c

index 2e3807d8eba8ba6c37898a057154e67590325903..ec97e4b4a62f44decdcb559991500a737c7d6379 100644 (file)
@@ -429,7 +429,7 @@ EXPORT_SYMBOL(tcp_init_sock);
 
 static void tcp_tx_timestamp(struct sock *sk, u16 tsflags, struct sk_buff *skb)
 {
-       if (tsflags) {
+       if (tsflags && skb) {
                struct skb_shared_info *shinfo = skb_shinfo(skb);
                struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
 
@@ -958,10 +958,8 @@ new_segment:
                copied += copy;
                offset += copy;
                size -= copy;
-               if (!size) {
-                       tcp_tx_timestamp(sk, sk->sk_tsflags, skb);
+               if (!size)
                        goto out;
-               }
 
                if (skb->len < size_goal || (flags & MSG_OOB))
                        continue;
@@ -987,8 +985,11 @@ wait_for_memory:
        }
 
 out:
-       if (copied && !(flags & MSG_SENDPAGE_NOTLAST))
-               tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
+       if (copied) {
+               tcp_tx_timestamp(sk, sk->sk_tsflags, tcp_write_queue_tail(sk));
+               if (!(flags & MSG_SENDPAGE_NOTLAST))
+                       tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
+       }
        return copied;
 
 do_error:
@@ -1281,7 +1282,6 @@ new_segment:
 
                copied += copy;
                if (!msg_data_left(msg)) {
-                       tcp_tx_timestamp(sk, sockc.tsflags, skb);
                        if (unlikely(flags & MSG_EOR))
                                TCP_SKB_CB(skb)->eor = 1;
                        goto out;
@@ -1312,8 +1312,10 @@ wait_for_memory:
        }
 
 out:
-       if (copied)
+       if (copied) {
+               tcp_tx_timestamp(sk, sockc.tsflags, tcp_write_queue_tail(sk));
                tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
+       }
 out_nopush:
        release_sock(sk);
        return copied + copied_syn;