[TCP]: whitespace cleanup
[powerpc.git] / net / ipv4 / tcp_input.c
index 924b2e6..fb02560 100644 (file)
@@ -86,6 +86,7 @@ int sysctl_tcp_stdurg __read_mostly;
 int sysctl_tcp_rfc1337 __read_mostly;
 int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
 int sysctl_tcp_frto __read_mostly;
+int sysctl_tcp_frto_response __read_mostly;
 int sysctl_tcp_nometrics_save __read_mostly;
 
 int sysctl_tcp_moderate_rcvbuf __read_mostly = 1;
@@ -100,6 +101,7 @@ int sysctl_tcp_abc __read_mostly;
 #define FLAG_ECE               0x40 /* ECE in this ACK                         */
 #define FLAG_DATA_LOST         0x80 /* SACK detected data lossage.             */
 #define FLAG_SLOWPATH          0x100 /* Do not skip RFC checks for window update.*/
+#define FLAG_ONLY_ORIG_SACKED  0x200 /* SACKs only non-rexmit sent before RTO */
 
 #define FLAG_ACKED             (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
 #define FLAG_NOT_DUP           (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
@@ -110,6 +112,8 @@ int sysctl_tcp_abc __read_mostly;
 #define IsFack(tp) ((tp)->rx_opt.sack_ok & 2)
 #define IsDSack(tp) ((tp)->rx_opt.sack_ok & 4)
 
+#define IsSackFrto() (sysctl_tcp_frto == 0x2)
+
 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
 
 /* Adapt the MSS value used to make delayed ack decision to the
@@ -574,7 +578,7 @@ static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt)
         * does not matter how to _calculate_ it. Seems, it was trap
         * that VJ failed to avoid. 8)
         */
-       if(m == 0)
+       if (m == 0)
                m = 1;
        if (tp->srtt != 0) {
                m -= (tp->srtt >> 3);   /* m is now error in rtt est */
@@ -759,15 +763,17 @@ __u32 tcp_init_cwnd(struct tcp_sock *tp, struct dst_entry *dst)
 }
 
 /* Set slow start threshold and cwnd not falling to slow start */
-void tcp_enter_cwr(struct sock *sk)
+void tcp_enter_cwr(struct sock *sk, const int set_ssthresh)
 {
        struct tcp_sock *tp = tcp_sk(sk);
+       const struct inet_connection_sock *icsk = inet_csk(sk);
 
        tp->prior_ssthresh = 0;
        tp->bytes_acked = 0;
-       if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
+       if (icsk->icsk_ca_state < TCP_CA_CWR) {
                tp->undo_marker = 0;
-               tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
+               if (set_ssthresh)
+                       tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
                tp->snd_cwnd = min(tp->snd_cwnd,
                                   tcp_packets_in_flight(tp) + 1U);
                tp->snd_cwnd_cnt = 0;
@@ -1038,7 +1044,7 @@ tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb, u32 prior_snd_
        cached_skb = tp->fastpath_skb_hint;
        cached_fack_count = tp->fastpath_cnt_hint;
        if (!cached_skb) {
-               cached_skb = sk->sk_write_queue.next;
+               cached_skb = tcp_write_queue_head(sk);
                cached_fack_count = 0;
        }
 
@@ -1055,10 +1061,13 @@ tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb, u32 prior_snd_
                if (after(end_seq, tp->high_seq))
                        flag |= FLAG_DATA_LOST;
 
-               sk_stream_for_retrans_queue_from(skb, sk) {
+               tcp_for_write_queue_from(skb, sk) {
                        int in_sack, pcount;
                        u8 sacked;
 
+                       if (skb == tcp_send_head(sk))
+                               break;
+
                        cached_skb = skb;
                        cached_fack_count = fack_count;
                        if (i == first_sack_index) {
@@ -1159,6 +1168,18 @@ tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb, u32 prior_snd_
                                                /* clear lost hint */
                                                tp->retransmit_skb_hint = NULL;
                                        }
+                                       /* SACK enhanced F-RTO detection.
+                                        * Set flag if and only if non-rexmitted
+                                        * segments below frto_highmark are
+                                        * SACKed (RFC4138; Appendix B).
+                                        * Clearing correct due to in-order walk
+                                        */
+                                       if (after(end_seq, tp->frto_highmark)) {
+                                               flag &= ~FLAG_ONLY_ORIG_SACKED;
+                                       } else {
+                                               if (!(sacked & TCPCB_RETRANS))
+                                                       flag |= FLAG_ONLY_ORIG_SACKED;
+                                       }
                                }
 
                                TCP_SKB_CB(skb)->sacked |= TCPCB_SACKED_ACKED;
@@ -1195,7 +1216,9 @@ tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb, u32 prior_snd_
        if (lost_retrans && icsk->icsk_ca_state == TCP_CA_Recovery) {
                struct sk_buff *skb;
 
-               sk_stream_for_retrans_queue(skb, sk) {
+               tcp_for_write_queue(skb, sk) {
+                       if (skb == tcp_send_head(sk))
+                               break;
                        if (after(TCP_SKB_CB(skb)->seq, lost_retrans))
                                break;
                        if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
@@ -1224,7 +1247,8 @@ tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb, u32 prior_snd_
 
        tp->left_out = tp->sacked_out + tp->lost_out;
 
-       if ((reord < tp->fackets_out) && icsk->icsk_ca_state != TCP_CA_Loss)
+       if ((reord < tp->fackets_out) && icsk->icsk_ca_state != TCP_CA_Loss &&
+           (!tp->frto_highmark || after(tp->snd_una, tp->frto_highmark)))
                tcp_update_reordering(sk, ((tp->fackets_out + 1) - reord), 0);
 
 #if FASTRETRANS_DEBUG > 0
@@ -1239,19 +1263,47 @@ tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb, u32 prior_snd_
 /* F-RTO can only be used if these conditions are satisfied:
  *  - there must be some unsent new data
  *  - the advertised window should allow sending it
+ *  - TCP has never retransmitted anything other than head (SACK enhanced
+ *    variant from Appendix B of RFC4138 is more robust here)
  */
-int tcp_use_frto(const struct sock *sk)
+int tcp_use_frto(struct sock *sk)
 {
        const struct tcp_sock *tp = tcp_sk(sk);
+       struct sk_buff *skb;
 
-       return (sysctl_tcp_frto && sk->sk_send_head &&
-               !after(TCP_SKB_CB(sk->sk_send_head)->end_seq,
-                      tp->snd_una + tp->snd_wnd));
+       if (!sysctl_tcp_frto || !tcp_send_head(sk) ||
+               after(TCP_SKB_CB(tcp_send_head(sk))->end_seq,
+                     tp->snd_una + tp->snd_wnd))
+               return 0;
+
+       if (IsSackFrto())
+               return 1;
+
+       /* Avoid expensive walking of rexmit queue if possible */
+       if (tp->retrans_out > 1)
+               return 0;
+
+       skb = tcp_write_queue_head(sk);
+       skb = tcp_write_queue_next(sk, skb);    /* Skips head */
+       tcp_for_write_queue_from(skb, sk) {
+               if (skb == tcp_send_head(sk))
+                       break;
+               if (TCP_SKB_CB(skb)->sacked&TCPCB_RETRANS)
+                       return 0;
+               /* Short-circuit when first non-SACKed skb has been checked */
+               if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED))
+                       break;
+       }
+       return 1;
 }
 
 /* RTO occurred, but do not yet enter Loss state. Instead, defer RTO
  * recovery a bit and use heuristics in tcp_process_frto() to detect if
- * the RTO was spurious.
+ * the RTO was spurious. Only clear SACKED_RETRANS of the head here to
+ * keep retrans_out counting accurate (with SACK F-RTO, other than head
+ * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS
+ * bits are handled if the Loss state is really to be entered (in
+ * tcp_enter_frto_loss).
  *
  * Do like tcp_enter_loss() would; when RTO expires the second time it
  * does:
@@ -1268,27 +1320,56 @@ void tcp_enter_frto(struct sock *sk)
            ((icsk->icsk_ca_state == TCP_CA_Loss || tp->frto_counter) &&
             !icsk->icsk_retransmits)) {
                tp->prior_ssthresh = tcp_current_ssthresh(sk);
-               tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
+               /* Our state is too optimistic in ssthresh() call because cwnd
+                * is not reduced until tcp_enter_frto_loss() when previous FRTO
+                * recovery has not yet completed. Pattern would be this: RTO,
+                * Cumulative ACK, RTO (2xRTO for the same segment does not end
+                * up here twice).
+                * RFC4138 should be more specific on what to do, even though
+                * RTO is quite unlikely to occur after the first Cumulative ACK
+                * due to back-off and complexity of triggering events ...
+                */
+               if (tp->frto_counter) {
+                       u32 stored_cwnd;
+                       stored_cwnd = tp->snd_cwnd;
+                       tp->snd_cwnd = 2;
+                       tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
+                       tp->snd_cwnd = stored_cwnd;
+               } else {
+                       tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
+               }
+               /* ... in theory, cong.control module could do "any tricks" in
+                * ssthresh(), which means that ca_state, lost bits and lost_out
+                * counter would have to be faked before the call occurs. We
+                * consider that too expensive, unlikely and hacky, so modules
+                * using these in ssthresh() must deal these incompatibility
+                * issues if they receives CA_EVENT_FRTO and frto_counter != 0
+                */
                tcp_ca_event(sk, CA_EVENT_FRTO);
        }
 
-       /* Have to clear retransmission markers here to keep the bookkeeping
-        * in shape, even though we are not yet in Loss state.
-        * If something was really lost, it is eventually caught up
-        * in tcp_enter_frto_loss.
-        */
-       tp->retrans_out = 0;
        tp->undo_marker = tp->snd_una;
        tp->undo_retrans = 0;
 
-       sk_stream_for_retrans_queue(skb, sk) {
+       skb = tcp_write_queue_head(sk);
+       if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
                TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
+               tp->retrans_out -= tcp_skb_pcount(skb);
        }
        tcp_sync_left_out(tp);
 
+       /* Earlier loss recovery underway (see RFC4138; Appendix B).
+        * The last condition is necessary at least in tp->frto_counter case.
+        */
+       if (IsSackFrto() && (tp->frto_counter ||
+           ((1 << icsk->icsk_ca_state) & (TCPF_CA_Recovery|TCPF_CA_Loss))) &&
+           after(tp->high_seq, tp->snd_una)) {
+               tp->frto_highmark = tp->high_seq;
+       } else {
+               tp->frto_highmark = tp->snd_nxt;
+       }
        tcp_set_ca_state(sk, TCP_CA_Disorder);
        tp->high_seq = tp->snd_nxt;
-       tp->frto_highmark = tp->snd_nxt;
        tp->frto_counter = 1;
 }
 
@@ -1296,7 +1377,7 @@ void tcp_enter_frto(struct sock *sk)
  * which indicates that we should follow the traditional RTO recovery,
  * i.e. mark everything lost and do go-back-N retransmission.
  */
-static void tcp_enter_frto_loss(struct sock *sk, int allowed_segments)
+static void tcp_enter_frto_loss(struct sock *sk, int allowed_segments, int flag)
 {
        struct tcp_sock *tp = tcp_sk(sk);
        struct sk_buff *skb;
@@ -1305,10 +1386,23 @@ static void tcp_enter_frto_loss(struct sock *sk, int allowed_segments)
        tp->sacked_out = 0;
        tp->lost_out = 0;
        tp->fackets_out = 0;
+       tp->retrans_out = 0;
 
-       sk_stream_for_retrans_queue(skb, sk) {
+       tcp_for_write_queue(skb, sk) {
+               if (skb == tcp_send_head(sk))
+                       break;
                cnt += tcp_skb_pcount(skb);
-               TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
+               /*
+                * Count the retransmission made on RTO correctly (only when
+                * waiting for the first ACK and did not get it)...
+                */
+               if ((tp->frto_counter == 1) && !(flag&FLAG_DATA_ACKED)) {
+                       tp->retrans_out += tcp_skb_pcount(skb);
+                       /* ...enter this if branch just for the first segment */
+                       flag |= FLAG_DATA_ACKED;
+               } else {
+                       TCP_SKB_CB(skb)->sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
+               }
                if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED)) {
 
                        /* Do not mark those segments lost that were
@@ -1384,7 +1478,9 @@ void tcp_enter_loss(struct sock *sk, int how)
        if (!how)
                tp->undo_marker = tp->snd_una;
 
-       sk_stream_for_retrans_queue(skb, sk) {
+       tcp_for_write_queue(skb, sk) {
+               if (skb == tcp_send_head(sk))
+                       break;
                cnt += tcp_skb_pcount(skb);
                if (TCP_SKB_CB(skb)->sacked&TCPCB_RETRANS)
                        tp->undo_marker = 0;
@@ -1419,14 +1515,14 @@ static int tcp_check_sack_reneging(struct sock *sk)
         * receiver _host_ is heavily congested (or buggy).
         * Do processing similar to RTO timeout.
         */
-       if ((skb = skb_peek(&sk->sk_write_queue)) != NULL &&
+       if ((skb = tcp_write_queue_head(sk)) != NULL &&
            (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
                struct inet_connection_sock *icsk = inet_csk(sk);
                NET_INC_STATS_BH(LINUX_MIB_TCPSACKRENEGING);
 
                tcp_enter_loss(sk, 1);
                icsk->icsk_retransmits++;
-               tcp_retransmit_skb(sk, skb_peek(&sk->sk_write_queue));
+               tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
                inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
                                          icsk->icsk_rto, TCP_RTO_MAX);
                return 1;
@@ -1447,7 +1543,7 @@ static inline int tcp_skb_timedout(struct sock *sk, struct sk_buff *skb)
 static inline int tcp_head_timedout(struct sock *sk, struct tcp_sock *tp)
 {
        return tp->packets_out &&
-              tcp_skb_timedout(sk, skb_peek(&sk->sk_write_queue));
+              tcp_skb_timedout(sk, tcp_write_queue_head(sk));
 }
 
 /* Linux NewReno/SACK/FACK/ECN state machine.
@@ -1547,6 +1643,10 @@ static int tcp_time_to_recover(struct sock *sk, struct tcp_sock *tp)
 {
        __u32 packets_out;
 
+       /* Do not perform any recovery during FRTO algorithm */
+       if (tp->frto_counter)
+               return 0;
+
        /* Trick#1: The loss is proven. */
        if (tp->lost_out)
                return 1;
@@ -1638,11 +1738,13 @@ static void tcp_mark_head_lost(struct sock *sk, struct tcp_sock *tp,
                skb = tp->lost_skb_hint;
                cnt = tp->lost_cnt_hint;
        } else {
-               skb = sk->sk_write_queue.next;
+               skb = tcp_write_queue_head(sk);
                cnt = 0;
        }
 
-       sk_stream_for_retrans_queue_from(skb, sk) {
+       tcp_for_write_queue_from(skb, sk) {
+               if (skb == tcp_send_head(sk))
+                       break;
                /* TODO: do this better */
                /* this is not the most efficient way to do this... */
                tp->lost_skb_hint = skb;
@@ -1656,12 +1758,11 @@ static void tcp_mark_head_lost(struct sock *sk, struct tcp_sock *tp,
 
                        /* clear xmit_retransmit_queue hints
                         *  if this is beyond hint */
-                       if(tp->retransmit_skb_hint != NULL &&
-                          before(TCP_SKB_CB(skb)->seq,
-                                 TCP_SKB_CB(tp->retransmit_skb_hint)->seq)) {
-
+                       if (tp->retransmit_skb_hint != NULL &&
+                           before(TCP_SKB_CB(skb)->seq,
+                                  TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
                                tp->retransmit_skb_hint = NULL;
-                       }
+
                }
        }
        tcp_sync_left_out(tp);
@@ -1689,9 +1790,11 @@ static void tcp_update_scoreboard(struct sock *sk, struct tcp_sock *tp)
                struct sk_buff *skb;
 
                skb = tp->scoreboard_skb_hint ? tp->scoreboard_skb_hint
-                       : sk->sk_write_queue.next;
+                       : tcp_write_queue_head(sk);
 
-               sk_stream_for_retrans_queue_from(skb, sk) {
+               tcp_for_write_queue_from(skb, sk) {
+                       if (skb == tcp_send_head(sk))
+                               break;
                        if (!tcp_skb_timedout(sk, skb))
                                break;
 
@@ -1882,7 +1985,9 @@ static int tcp_try_undo_loss(struct sock *sk, struct tcp_sock *tp)
 {
        if (tcp_may_undo(tp)) {
                struct sk_buff *skb;
-               sk_stream_for_retrans_queue(skb, sk) {
+               tcp_for_write_queue(skb, sk) {
+                       if (skb == tcp_send_head(sk))
+                               break;
                        TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
                }
 
@@ -1918,7 +2023,7 @@ static void tcp_try_to_open(struct sock *sk, struct tcp_sock *tp, int flag)
                tp->retrans_stamp = 0;
 
        if (flag&FLAG_ECE)
-               tcp_enter_cwr(sk);
+               tcp_enter_cwr(sk, 1);
 
        if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
                int state = TCP_CA_Open;
@@ -2294,8 +2399,8 @@ static int tcp_clean_rtx_queue(struct sock *sk, __s32 *seq_rtt_p)
                = icsk->icsk_ca_ops->rtt_sample;
        struct timeval tv = { .tv_sec = 0, .tv_usec = 0 };
 
-       while ((skb = skb_peek(&sk->sk_write_queue)) &&
-              skb != sk->sk_send_head) {
+       while ((skb = tcp_write_queue_head(sk)) &&
+              skb != tcp_send_head(sk)) {
                struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
                __u8 sacked = scb->sacked;
 
@@ -2335,7 +2440,7 @@ static int tcp_clean_rtx_queue(struct sock *sk, __s32 *seq_rtt_p)
 
                if (sacked) {
                        if (sacked & TCPCB_RETRANS) {
-                               if(sacked & TCPCB_SACKED_RETRANS)
+                               if (sacked & TCPCB_SACKED_RETRANS)
                                        tp->retrans_out -= tcp_skb_pcount(skb);
                                acked |= FLAG_RETRANS_DATA_ACKED;
                                seq_rtt = -1;
@@ -2358,7 +2463,7 @@ static int tcp_clean_rtx_queue(struct sock *sk, __s32 *seq_rtt_p)
                }
                tcp_dec_pcount_approx(&tp->fackets_out, skb);
                tcp_packets_out_dec(tp, skb);
-               __skb_unlink(skb, &sk->sk_write_queue);
+               tcp_unlink_write_queue(skb, sk);
                sk_stream_free_skb(sk, skb);
                clear_all_retrans_hints(tp);
        }
@@ -2407,7 +2512,7 @@ static void tcp_ack_probe(struct sock *sk)
 
        /* Was it a usable window open? */
 
-       if (!after(TCP_SKB_CB(sk->sk_send_head)->end_seq,
+       if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq,
                   tp->snd_una + tp->snd_wnd)) {
                icsk->icsk_backoff = 0;
                inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
@@ -2494,6 +2599,22 @@ static void tcp_conservative_spur_to_response(struct tcp_sock *tp)
        tcp_moderate_cwnd(tp);
 }
 
+/* A conservative spurious RTO response algorithm: reduce cwnd using
+ * rate halving and continue in congestion avoidance.
+ */
+static void tcp_ratehalving_spur_to_response(struct sock *sk)
+{
+       tcp_enter_cwr(sk, 0);
+}
+
+static void tcp_undo_spur_to_response(struct sock *sk, int flag)
+{
+       if (flag&FLAG_ECE)
+               tcp_ratehalving_spur_to_response(sk);
+       else
+               tcp_undo_cwr(sk, 1);
+}
+
 /* F-RTO spurious RTO detection algorithm (RFC4138)
  *
  * F-RTO affects during two new ACKs following RTO (well, almost, see inline
@@ -2509,6 +2630,10 @@ static void tcp_conservative_spur_to_response(struct tcp_sock *tp)
  * Rationale: if the RTO was spurious, new ACKs should arrive from the
  * original window even after we transmit two new data segments.
  *
+ * SACK version:
+ *   on first step, wait until first cumulative ACK arrives, then move to
+ *   the second step. In second step, the next ACK decides.
+ *
  * F-RTO is implemented (mainly) in four functions:
  *   - tcp_use_frto() is used to determine if TCP is can use F-RTO
  *   - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is
@@ -2518,7 +2643,7 @@ static void tcp_conservative_spur_to_response(struct tcp_sock *tp)
  *     to prove that the RTO is indeed spurious. It transfers the control
  *     from F-RTO to the conventional RTO recovery
  */
-static void tcp_process_frto(struct sock *sk, u32 prior_snd_una, int flag)
+static int tcp_process_frto(struct sock *sk, u32 prior_snd_una, int flag)
 {
        struct tcp_sock *tp = tcp_sk(sk);
 
@@ -2529,29 +2654,63 @@ static void tcp_process_frto(struct sock *sk, u32 prior_snd_una, int flag)
                inet_csk(sk)->icsk_retransmits = 0;
 
        if (!before(tp->snd_una, tp->frto_highmark)) {
-               tcp_enter_frto_loss(sk, tp->frto_counter + 1);
-               return;
+               tcp_enter_frto_loss(sk, tp->frto_counter + 1, flag);
+               return 1;
        }
 
-       /* RFC4138 shortcoming in step 2; should also have case c): ACK isn't
-        * duplicate nor advances window, e.g., opposite dir data, winupdate
-        */
-       if ((tp->snd_una == prior_snd_una) && (flag&FLAG_NOT_DUP) &&
-           !(flag&FLAG_FORWARD_PROGRESS))
-               return;
+       if (!IsSackFrto() || IsReno(tp)) {
+               /* RFC4138 shortcoming in step 2; should also have case c):
+                * ACK isn't duplicate nor advances window, e.g., opposite dir
+                * data, winupdate
+                */
+               if ((tp->snd_una == prior_snd_una) && (flag&FLAG_NOT_DUP) &&
+                   !(flag&FLAG_FORWARD_PROGRESS))
+                       return 1;
 
-       if (!(flag&FLAG_DATA_ACKED)) {
-               tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 0 : 3));
-               return;
+               if (!(flag&FLAG_DATA_ACKED)) {
+                       tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 0 : 3),
+                                           flag);
+                       return 1;
+               }
+       } else {
+               if (!(flag&FLAG_DATA_ACKED) && (tp->frto_counter == 1)) {
+                       /* Prevent sending of new data. */
+                       tp->snd_cwnd = min(tp->snd_cwnd,
+                                          tcp_packets_in_flight(tp));
+                       return 1;
+               }
+
+               if ((tp->frto_counter == 2) &&
+                   (!(flag&FLAG_FORWARD_PROGRESS) ||
+                    ((flag&FLAG_DATA_SACKED) && !(flag&FLAG_ONLY_ORIG_SACKED)))) {
+                       /* RFC4138 shortcoming (see comment above) */
+                       if (!(flag&FLAG_FORWARD_PROGRESS) && (flag&FLAG_NOT_DUP))
+                               return 1;
+
+                       tcp_enter_frto_loss(sk, 3, flag);
+                       return 1;
+               }
        }
 
        if (tp->frto_counter == 1) {
                tp->snd_cwnd = tcp_packets_in_flight(tp) + 2;
+               tp->frto_counter = 2;
+               return 1;
        } else /* frto_counter == 2 */ {
-               tcp_conservative_spur_to_response(tp);
+               switch (sysctl_tcp_frto_response) {
+               case 2:
+                       tcp_undo_spur_to_response(sk, flag);
+                       break;
+               case 1:
+                       tcp_conservative_spur_to_response(tp);
+                       break;
+               default:
+                       tcp_ratehalving_spur_to_response(sk);
+                       break;
+               };
+               tp->frto_counter = 0;
        }
-
-       tp->frto_counter = (tp->frto_counter + 1) % 3;
+       return 0;
 }
 
 /* This routine deals with incoming acks, but not outgoing ones. */
@@ -2565,6 +2724,7 @@ static int tcp_ack(struct sock *sk, struct sk_buff *skb, int flag)
        u32 prior_in_flight;
        s32 seq_rtt;
        int prior_packets;
+       int frto_cwnd = 0;
 
        /* If the ack is newer than sent or older than previous acks
         * then we can probably ignore it.
@@ -2627,15 +2787,16 @@ static int tcp_ack(struct sock *sk, struct sk_buff *skb, int flag)
        flag |= tcp_clean_rtx_queue(sk, &seq_rtt);
 
        if (tp->frto_counter)
-               tcp_process_frto(sk, prior_snd_una, flag);
+               frto_cwnd = tcp_process_frto(sk, prior_snd_una, flag);
 
        if (tcp_ack_is_dubious(sk, flag)) {
                /* Advance CWND, if state allows this. */
-               if ((flag & FLAG_DATA_ACKED) && tcp_may_raise_cwnd(sk, flag))
+               if ((flag & FLAG_DATA_ACKED) && !frto_cwnd &&
+                   tcp_may_raise_cwnd(sk, flag))
                        tcp_cong_avoid(sk, ack,  seq_rtt, prior_in_flight, 0);
                tcp_fastretrans_alert(sk, prior_snd_una, prior_packets, flag);
        } else {
-               if ((flag & FLAG_DATA_ACKED))
+               if ((flag & FLAG_DATA_ACKED) && !frto_cwnd)
                        tcp_cong_avoid(sk, ack, seq_rtt, prior_in_flight, 1);
        }
 
@@ -2651,7 +2812,7 @@ no_queue:
         * being used to time the probes, and is probably far higher than
         * it needs to be for normal retransmission.
         */
-       if (sk->sk_send_head)
+       if (tcp_send_head(sk))
                tcp_ack_probe(sk);
        return 1;
 
@@ -2678,7 +2839,7 @@ void tcp_parse_options(struct sk_buff *skb, struct tcp_options_received *opt_rx,
        ptr = (unsigned char *)(th + 1);
        opt_rx->saw_tstamp = 0;
 
-       while(length>0) {
+       while (length > 0) {
                int opcode=*ptr++;
                int opsize;
 
@@ -2694,9 +2855,9 @@ void tcp_parse_options(struct sk_buff *skb, struct tcp_options_received *opt_rx,
                                        return;
                                if (opsize > length)
                                        return; /* don't parse partial options */
-                               switch(opcode) {
+                               switch (opcode) {
                                case TCPOPT_MSS:
-                                       if(opsize==TCPOLEN_MSS && th->syn && !estab) {
+                                       if (opsize==TCPOLEN_MSS && th->syn && !estab) {
                                                u16 in_mss = ntohs(get_unaligned((__be16 *)ptr));
                                                if (in_mss) {
                                                        if (opt_rx->user_mss && opt_rx->user_mss < in_mss)
@@ -2706,12 +2867,12 @@ void tcp_parse_options(struct sk_buff *skb, struct tcp_options_received *opt_rx,
                                        }
                                        break;
                                case TCPOPT_WINDOW:
-                                       if(opsize==TCPOLEN_WINDOW && th->syn && !estab)
+                                       if (opsize==TCPOLEN_WINDOW && th->syn && !estab)
                                                if (sysctl_tcp_window_scaling) {
                                                        __u8 snd_wscale = *(__u8 *) ptr;
                                                        opt_rx->wscale_ok = 1;
                                                        if (snd_wscale > 14) {
-                                                               if(net_ratelimit())
+                                                               if (net_ratelimit())
                                                                        printk(KERN_INFO "tcp_parse_options: Illegal window "
                                                                               "scaling value %d >14 received.\n",
                                                                               snd_wscale);
@@ -2721,7 +2882,7 @@ void tcp_parse_options(struct sk_buff *skb, struct tcp_options_received *opt_rx,
                                                }
                                        break;
                                case TCPOPT_TIMESTAMP:
-                                       if(opsize==TCPOLEN_TIMESTAMP) {
+                                       if (opsize==TCPOLEN_TIMESTAMP) {
                                                if ((estab && opt_rx->tstamp_ok) ||
                                                    (!estab && sysctl_tcp_timestamps)) {
                                                        opt_rx->saw_tstamp = 1;
@@ -2731,7 +2892,7 @@ void tcp_parse_options(struct sk_buff *skb, struct tcp_options_received *opt_rx,
                                        }
                                        break;
                                case TCPOPT_SACK_PERM:
-                                       if(opsize==TCPOLEN_SACK_PERM && th->syn && !estab) {
+                                       if (opsize==TCPOLEN_SACK_PERM && th->syn && !estab) {
                                                if (sysctl_tcp_sack) {
                                                        opt_rx->sack_ok = 1;
                                                        tcp_sack_reset(opt_rx);
@@ -2740,7 +2901,7 @@ void tcp_parse_options(struct sk_buff *skb, struct tcp_options_received *opt_rx,
                                        break;
 
                                case TCPOPT_SACK:
-                                       if((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
+                                       if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
                                           !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
                                           opt_rx->sack_ok) {
                                                TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
@@ -2789,7 +2950,7 @@ static int tcp_fast_parse_options(struct sk_buff *skb, struct tcphdr *th,
 static inline void tcp_store_ts_recent(struct tcp_sock *tp)
 {
        tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
-       tp->rx_opt.ts_recent_stamp = xtime.tv_sec;
+       tp->rx_opt.ts_recent_stamp = get_seconds();
 }
 
 static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
@@ -2802,8 +2963,8 @@ static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
                 * Not only, also it occurs for expired timestamps.
                 */
 
-               if((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) >= 0 ||
-                  xtime.tv_sec >= tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS)
+               if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) >= 0 ||
+                  get_seconds() >= tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS)
                        tcp_store_ts_recent(tp);
        }
 }
@@ -2855,7 +3016,7 @@ static inline int tcp_paws_discard(const struct sock *sk, const struct sk_buff *
 {
        const struct tcp_sock *tp = tcp_sk(sk);
        return ((s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) > TCP_PAWS_WINDOW &&
-               xtime.tv_sec < tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS &&
+               get_seconds() < tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS &&
                !tcp_disordered_ack(sk, skb));
 }
 
@@ -3061,7 +3222,7 @@ static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
                         */
                        tp->rx_opt.num_sacks--;
                        tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
-                       for(i=this_sack; i < tp->rx_opt.num_sacks; i++)
+                       for (i=this_sack; i < tp->rx_opt.num_sacks; i++)
                                sp[i] = sp[i+1];
                        continue;
                }
@@ -3114,7 +3275,7 @@ static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
                tp->rx_opt.num_sacks--;
                sp--;
        }
-       for(; this_sack > 0; this_sack--, sp--)
+       for (; this_sack > 0; this_sack--, sp--)
                *sp = *(sp-1);
 
 new_sack:
@@ -3140,7 +3301,7 @@ static void tcp_sack_remove(struct tcp_sock *tp)
                return;
        }
 
-       for(this_sack = 0; this_sack < num_sacks; ) {
+       for (this_sack = 0; this_sack < num_sacks; ) {
                /* Check if the start of the sack is covered by RCV.NXT. */
                if (!before(tp->rcv_nxt, sp->start_seq)) {
                        int i;
@@ -3196,7 +3357,7 @@ static void tcp_ofo_queue(struct sock *sk)
                __skb_unlink(skb, &tp->out_of_order_queue);
                __skb_queue_tail(&sk->sk_receive_queue, skb);
                tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
-               if(skb->h.th->fin)
+               if (skb->h.th->fin)
                        tcp_fin(skb, sk, skb->h.th);
        }
 }
@@ -3262,9 +3423,9 @@ queue_and_out:
                        __skb_queue_tail(&sk->sk_receive_queue, skb);
                }
                tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
-               if(skb->len)
+               if (skb->len)
                        tcp_event_data_recv(sk, tp, skb);
-               if(th->fin)
+               if (th->fin)
                        tcp_fin(skb, sk, th);
 
                if (!skb_queue_empty(&tp->out_of_order_queue)) {
@@ -4161,7 +4322,7 @@ slow_path:
                goto discard;
        }
 
-       if(th->rst) {
+       if (th->rst) {
                tcp_reset(sk);
                goto discard;
        }
@@ -4176,7 +4337,7 @@ slow_path:
        }
 
 step5:
-       if(th->ack)
+       if (th->ack)
                tcp_ack(sk, skb, FLAG_SLOWPATH);
 
        tcp_rcv_rtt_measure_ts(sk, skb);
@@ -4464,13 +4625,13 @@ int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
                goto discard;
 
        case TCP_LISTEN:
-               if(th->ack)
+               if (th->ack)
                        return 1;
 
-               if(th->rst)
+               if (th->rst)
                        goto discard;
 
-               if(th->syn) {
+               if (th->syn) {
                        if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
                                return 1;
 
@@ -4526,7 +4687,7 @@ int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
        }
 
        /* step 2: check RST bit */
-       if(th->rst) {
+       if (th->rst) {
                tcp_reset(sk);
                goto discard;
        }
@@ -4549,7 +4710,7 @@ int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
        if (th->ack) {
                int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH);
 
-               switch(sk->sk_state) {
+               switch (sk->sk_state) {
                case TCP_SYN_RECV:
                        if (acceptable) {
                                tp->copied_seq = tp->rcv_nxt;