tcp: bug fix in proportional rate reduction.

This patch is a fix for a bug triggering newly_acked_sacked < 0
in tcp_ack(.).

The bug is triggered by sacked_out decreasing relative to prior_sacked,
but packets_out remaining the same as pior_packets. This is because the
snapshot of prior_packets is taken after tcp_sacktag_write_queue() while
prior_sacked is captured before tcp_sacktag_write_queue(). The problem
is: tcp_sacktag_write_queue (tcp_match_skb_to_sack() -> tcp_fragment)
adjusts the pcount for packets_out and sacked_out (MSS change or other
reason). As a result, this delta in pcount is reflected in
(prior_sacked - sacked_out) but not in (prior_packets - packets_out).

This patch does the following:
1) initializes prior_packets at the start of tcp_ack() so as to
capture the delta in packets_out created by tcp_fragment.
2) introduces a new "previous_packets_out" variable that snapshots
packets_out right before tcp_clean_rtx_queue, so pkts_acked can be
correctly computed as before.
3) Computes pkts_acked using previous_packets_out, and computes
newly_acked_sacked using prior_packets.

Signed-off-by: Nandita Dukkipati <nanditad@google.com>
Acked-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

diff --git a/net/ipv4/tcp_input.c b/net/ipv4/tcp_input.c
index 08bbe609652891f593ccc2f48ec9975bbe2f335d..9c6225780bd5aafdb175a762d45c704e82416582 100644 (file)
--- a/net/ipv4/tcp_input.c
+++ b/net/ipv4/tcp_input.c
@@ -2743,8 +2743,8 @@ static void tcp_process_loss(struct sock *sk, int flag, bool is_dupack)
  * tcp_xmit_retransmit_queue().
  */
 static void tcp_fastretrans_alert(struct sock *sk, int pkts_acked,
-                                 int prior_sacked, bool is_dupack,
-                                 int flag)
+                                 int prior_sacked, int prior_packets,
+                                 bool is_dupack, int flag)
 {
        struct inet_connection_sock *icsk = inet_csk(sk);
        struct tcp_sock *tp = tcp_sk(sk);
@@ -2804,7 +2804,8 @@ static void tcp_fastretrans_alert(struct sock *sk, int pkts_acked,
                                tcp_add_reno_sack(sk);
                } else
                        do_lost = tcp_try_undo_partial(sk, pkts_acked);
-               newly_acked_sacked = pkts_acked + tp->sacked_out - prior_sacked;
+               newly_acked_sacked = prior_packets - tp->packets_out +
+                                    tp->sacked_out - prior_sacked;
                break;
        case TCP_CA_Loss:
                tcp_process_loss(sk, flag, is_dupack);
@@ -2818,7 +2819,8 @@ static void tcp_fastretrans_alert(struct sock *sk, int pkts_acked,
                        if (is_dupack)
                                tcp_add_reno_sack(sk);
                }
-               newly_acked_sacked = pkts_acked + tp->sacked_out - prior_sacked;
+               newly_acked_sacked = prior_packets - tp->packets_out +
+                                    tp->sacked_out - prior_sacked;
 
                if (icsk->icsk_ca_state <= TCP_CA_Disorder)
                        tcp_try_undo_dsack(sk);
@@ -3330,9 +3332,10 @@ static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
        bool is_dupack = false;
        u32 prior_in_flight;
        u32 prior_fackets;
-       int prior_packets;
+       int prior_packets = tp->packets_out;
        int prior_sacked = tp->sacked_out;
        int pkts_acked = 0;
+       int previous_packets_out = 0;
 
        /* If the ack is older than previous acks
         * then we can probably ignore it.
@@ -3403,14 +3406,14 @@ static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
        sk->sk_err_soft = 0;
        icsk->icsk_probes_out = 0;
        tp->rcv_tstamp = tcp_time_stamp;
-       prior_packets = tp->packets_out;
        if (!prior_packets)
                goto no_queue;
 
        /* See if we can take anything off of the retransmit queue. */
+       previous_packets_out = tp->packets_out;
        flag |= tcp_clean_rtx_queue(sk, prior_fackets, prior_snd_una);
 
-       pkts_acked = prior_packets - tp->packets_out;
+       pkts_acked = previous_packets_out - tp->packets_out;
 
        if (tcp_ack_is_dubious(sk, flag)) {
                /* Advance CWND, if state allows this. */
@@ -3418,7 +3421,7 @@ static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
                        tcp_cong_avoid(sk, ack, prior_in_flight);
                is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));
                tcp_fastretrans_alert(sk, pkts_acked, prior_sacked,
-                                     is_dupack, flag);
+                                     prior_packets, is_dupack, flag);
        } else {
                if (flag & FLAG_DATA_ACKED)
                        tcp_cong_avoid(sk, ack, prior_in_flight);
@@ -3441,7 +3444,7 @@ no_queue:
        /* If data was DSACKed, see if we can undo a cwnd reduction. */
        if (flag & FLAG_DSACKING_ACK)
                tcp_fastretrans_alert(sk, pkts_acked, prior_sacked,
-                                     is_dupack, flag);
+                                     prior_packets, is_dupack, flag);
        /* If this ack opens up a zero window, clear backoff.  It was
         * being used to time the probes, and is probably far higher than
         * it needs to be for normal retransmission.
@@ -3464,7 +3467,7 @@ old_ack:
        if (TCP_SKB_CB(skb)->sacked) {
                flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
                tcp_fastretrans_alert(sk, pkts_acked, prior_sacked,
-                                     is_dupack, flag);
+                                     prior_packets, is_dupack, flag);
        }
 
        SOCK_DEBUG(sk, "Ack %u before %u:%u\n", ack, tp->snd_una, tp->snd_nxt);