kk Blog —— 通用基础


date [-d @int|str] [+%s|"+%F %T"]
netstat -ltunp
sar -n DEV 1

FRTO—虚假超时剖析

http://blog.csdn.net/zhangskd/article/details/7446441

F-RTO：Forward RTO-Recovery，for a TCP sender to recover after a retransmission timeout. F-RTO的主要目的：The main motivation of the algorithm is to recover efficiently from a spurious RTO.

F-RTO的基本思想

The guideline behind F-RTO is, that an RTO either indicates a loss, or it is caused by an excessive delay in packet delivery while there still are outstanding segments in flight. If the RTO was due to delay, i.e. the RTO was spurious, acknowledgements for non-retransmitted segments sent before the RTO should arrive at the sender after the RTO occurred. If no such segments arrive, the RTO is concluded to be non-spurious and the conventional RTO recovery with go-back-N retransmissions should take place at the TCP sender.

To implement the principle described above, an F-RTO sender acts as follows: if the first ACK arriving after a RTO-triggered retransmission advances the window, transmit two new segments instead of continuing retransmissions. If also the second incoming acknowledgement advances the window, RTO is likely to be spurious, because the second ACK is triggered by an originally transmitted segment that has not been retransmitted after the RTO. If either one of the two acknowledgements after RTO is a duplicate ACK, the sender continues retransmissions similarly to the conventional RTO recovery algorithm.

When the retransmission timer expires, the F-RTO algorithm takes the following steps at the TCP sender. In the algorithm description below we use SND.UNA to indicate the first unacknowledged segment.

1.When the retransmission timer expires, retransmit the segment that triggered the timeout. As required by the TCP congestion control specifications, the ssthresh is adjusted to half of the number of currently outstanding segments. However, the congestion window is not yet set to one segment, but the sender waits for the next two acknowledgements before deciding on what to do with the congestion window.

2.When the first acknowledgement after RTO arrives at the sender, the sender chooses the following actions depending on whether the ACK advances the window or whether it is a duplicate ACK.

（a）If the acknowledgement advances SND.UNA, transmit up to two new (previously unsent) segments. This is the main point in which the F-RTO algorithm differs from the conventional way of recovering from RTO. After transmitting the two new segments, the congestion window size is set to have the same value as ssthresh. In effect this reduces the transmission rate of the sender to half of the transmission rate before the RTO. At this point the TCP sender has transmitted a total of three segments after the RTO, similarly to the conventional recovery algorithm. If transmitting two new segments is not possible due to advertised window limitation, or because there is no more data to send, the sender may transmit only one segment. If now new data can be transmitted, the TCP sender follows the conventional RTO recovery algorithm and starts retransmitting the unacknowledged data using slow start.

（b）If the acknowledgement is duplicate ACK, set the congestion window to one segment and proceed with the conventional RTO recovery. Two new segments are not transmitted in this case, because the conventional RTO recovery algorithm would not transmit anything at this point either. Instead, the F-RTO sender continues with slow start and performs similarly to the conventional TCP sender in retransmitting the unacknowledged segments. Step 3 of the F-RTO algorithm is not entered in this case. A common reason for executing this branch is the loss of a segment, in which case the segments injected by the sender before the RTO may still trigger duplicate ACKs that arrive at the sender after the RTO.

3.When the second acknowledgement after the RTO arrives, either continue transmitting new data, or start retransmitting with the slow start algorithm, depending on whether new data was acknowledged.

（a）If the acknowledgement advances SND.UNA, continue transmitting new data following the congestion avoidance algorithm. Because the TCP sender has retransmitted only one segment after the RTO, this acknowledgement indicates that an originally transmitted segment has arrived at the receiver. This is regarded as a strong indication of a suprious RTO. However, since the TCP sender cannot surely know at this point whether the segment that triggered the RTO was actually lost, adjusting the congestion control parameters after the RTO is the conservative action. From this point on, the TCP sender continues as in the normal congestion avoidance.

If this algorithm branch is taken, the TCP sender ignores the send_high variable that indicates the highest sequence number transmitted so far. The send_high variable was proposed as a bugfix for avoiding unnecessary multiple fast retransmits when RTO expires during fast recovery with NewReon TCP. As the sender has not retransmitted other segments but the one that triggered RTO, the problem addressed by the bugfix cannot occur. Therefore, if there are duplicate ACKs arriving at the sender after the RTO, they are likely to indicate a packet loss, hence fast retransmit should bu used to allow efficient recovery. Alternatively, if there are not enough duplicate ACKs arriving at the sender after a packet loss, the retransmission timer expires another time and the sender enters step 1 of this algorithm to detect whether the new RTO is spurious.

（b）If the acknowledgement is duplicate ACK, set the congestion window to three segments, continue with the slow start algorithm retransmitting unacknowledged segments. The duplicate ACK indicates that at least one segment other than the segment that triggered RTO is lost in the last window of data. There is no sufficient evidence that any of the segments was delayed. Therefore the sender proceeds with retransmissions similarly to the conventional RTO recovery algorithm, with the send_high variable stored when the retransmission timer expired to avoid unnecessary fast retransmits.

引起RTO的主要因素：

（1）Sudden delays
The primary motivation of the F-RTO algorithm is to improve the TCP performance when sudden delays cause spurious retransmission timeouts.

（2）Packet losses
These timeouts occur mainly when retransmissions are lost, since lost original packets are usually recovered by fast retransmit.

（3）Bursty losses
Losses of several successive packets can result in a retransmission timeout.

造成虚假RTO的原因还有：

Wireless links may also suffer from link outages that cause persistent data loss for a period of time.
Oher potential reasons for sudden delays that have been reported to trigger spurious RTOs include a delay due to tedious actions required to complete a hand-off or re-routing of packets to the new serving access point after the hand-off, arrival of competing traffic on a shared link with low bandwidth, and a sudden bandwidth degradation due to reduced resources on a wireless channel.

造成真实RTO的原因：

A RTO-triggered retransmission is needed when a retransmission is lost, or when nearly a whole window of data is lost, thus making it impossible for the receiver to generate enough duplicate ACKs for triggering TCP fast retransmit.

虚假RTO的后果

If no segments were lost but the retransmission timer expires spuriously, the segments retransmitted in the slow-start are sent unnecessarily. Particularly, this phenomenon is very possible with the various wireless access network technologies that are prone to sudden delay spikes. The retransmission timer expires because of the delay, spuriously triggering the RTO recovery and unnecessarily retransmission of all unacknowledged segments. This happens because after the delay the ACKs for the original segments arrive at the sender one at the time but too late, because the TCP sender has already entered the RTO recovery. Therefore, each of the ACKs trigger the retransmission of segments for which the original ACKs will arrive after a while. This continues until the whole window of segments is eventually unnecessarily retransmitted. Furthermore, because a full window of retransmitted segments arrive unnecessarily at the receiver, it generates duplicate ACKs for these out-of-order segments. Later on, the duplicate ACKs unnecessarily trigger fast retransmit at the sender.

TCP uses the fast retransmit mechanism to trigger retransmissions after receiving three successive duplicate acknowledgements (ACKs). If for a certain time period TCP sender does not receive ACKs that acknowledge new data, the TCP retransmission timer expires as a backoff mechanism. When the retransmission time expires, the TCP sender retransmits the first unacknowledged segment assuming it was lost in the network. Because a retransmission timeout (RTO) can be an indication of severe congestion in the network, the TCP sender resets its congestion window to one segment and starts increasing it according to the slow start algorithm. However, if the RTO occurs spuriously and there still are segments outstanding in the network, a false slow start is harmful for the potentially congested network as it injects extra segments to the network at increasing rate.

虚假的RTO不仅会降低吞吐量，而且由于丢包后会使用慢启动算法，快速的向网络中注入数据包， 而此时网络中还有原来发送的数据包，这样可能会造成真正的网络拥塞！

How about Reliable link-layer protocol ? Since wireless networks are often subject to high packet loss rate due to corruption or hand-offs, reliable link-layer protocols are widely employed with wireless links. The link-layer receiver often aims to deliver the packets to the upper protocol layers in order, which implies that the later arriving packets are blocked until the head of the queue arrives successfully. Due to the strict link-layer ordering, the communication end point observe a pause in packet delivery that can cause a spurious TCP RTO instead of getting out-of-order packets that could result in a false fast retransmit instead. Either way, interaction between TCP retransmission mechanisms and link-layer recovery can cause poor performance.

DSACK不能解决此问题 If the unnecessary retransmissions occurred due to spurious RTO caused by a sudden delay, the acknowledgements with the DSACK information arrive at the sender only after the acknowledgements of the original segments. Therefore, the unnecessary retransmissions following the spurious RTO cannot be avoided by using DSACK. Instead, the suggested recovery algorithm using DSACK can only revert the congestion control parameters to the state preceding the spurious retransmissions.

F-RTO实现

F-RTO is implemented (mainly) in four functions:
（1）tcp_use_frto() is used to determine if TCP can use F-RTO.

（2）tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is called when tcp_use_frto() showed green light.

（3）tcp_process_frto() handles incoming ACKs during F-RTO algorithm.

（4）tcp_enter_frto_loss() is called if there is not enough evidence to prove that the RTO is indeed spurious. It transfers the control from F-RTO to the conventional RTO recovery.

判断是否可以使用F-RTO

调用时机：当TCP段传送超时后，会引起段的重传，在重传定时器的处理过程中会判断是否可以使用F-RTO算法。

1
2
3
4
5
6
7
8
9
10
11
12

void tcp_retransmit_timer (struct sock *sk)  
{  
	....  
  
	if (tcp_use_frto(sk)) {  
		tcp_enter_frto(sk);  
	} else {  
		tcp_enter_loss(sk);  
	}  
  
	....  
}

能够使用F-RTO的条件：
（1）tcp_frto非零，此为TCP参数
（2）MTU probe没使用，因为它和F-RTO有冲突
（3）a. 如果启用了sackfrto，则可以使用
b. 如果没启用sackfrto，不能重传过除head以外的数据

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45

/* F-RTO can only be used if TCP has never retransmitted anything other than 
 * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here) 
 */  
int tcp_use_frto(struct sock *sk)  
{  
	const struct tcp_sock *tp = tcp_sk(sk);  
	const struct inet_connection_sock *icsk = inet_csk(sk);  
	struct sk_buff *skb;  
  
	if (! sysctl_tcp_frto)  
		return 0;  
  
	/* MTU probe and F-RTO won't really play nicely along currently */  
	if (icsk->icsk_mtup.probe_size)  
		return 0;  
  
	if (tcp_is_sackfrto(tp))  
		return 1;  
  
	/* Avoid expensive walking of rexmit queue if possible */  
	if (tp->retrans_out > 1)  
		return 0; /* 不能重过传除了head以外的数据*/  
  
	skb = tcp_write_queue_head(sk);  
	if (tcp_skb_is_last(sk, skb))  
		return 1;  
	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; /* 不允许处head以外的数据包被重传过 */  
  
		/* Short-circut when first non-SACKed skb has been checked */  
		if (! (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))  
		break;  
	}  
	return 1;  
}  
  
static int tcp_is_sackfrto(const struct tcp_sock *tp)  
{  
	return (sysctl_tcp_frto == 0x2) && ! tcp_is_reno(tp);  
}

进入F-RTO状态

启用F-RTO后，虽然传送超时，但还没进入Loss状态，相反，先进入Disorder状态。减小慢启动阈值，而snd_cwnd暂时保持不变。此时对应head数据包还没重传前。

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65

/* 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. 
 */  
  
void tcp_enter_frto (struct sock *sk)  
{  
	const struct inet_connection_sock *icsk = inet_csk(sk);  
	struct tcp_sock *tp = tcp_sk(sk);  
	struct sk_buff *skb;  
  
	/* Do like tcp_enter_loss() would*/  
	if ((! tp->frto_counter && icsk->icsk_ca_state <= TCP_CA_Disorder) ||  
		tp->snd_una == tp->high_seq ||   
		((icsk->icsk_ca_state == TCP_CA_Loss || tp->frto_counter) &&  
		! icsk->icsk_retransmits)) {  
  
		tp->prior_ssthresh = tcp_current_ssthresh(sk); /* 保存旧阈值*/  
  
		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); /* 减小阈值*/  
		}  
  
		tcp_ca_event(sk, CA_EVENT_FRTO); /* 触发FRTO事件 */  
	}  
  
	tp->undo_marker = tp->snd_una;  
	tp->undo_retrans = 0;  
  
	skb = tcp_write_queue_head(sk);  
	if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)  
		tp->undo_marker = 0;  
  
	/* 清除head与重传相关的标志*/  
	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_verfify_left_out(tp);  
  
	/* Too bad if TCP was application limited */  
	tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1);  
  
	/* Earlier loss recovery underway */  
	if (tcp_is_sackfrto(tp) && (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); /* 设置拥塞状态*/  
	tp->high_seq = tp->snd_nxt;  
	tp->frto_counter = 1; /* 表示刚进入F-RTO状态！*/  
}

F-RTO算法处理

F-RTO算法的处理过程主要发生在重传完超时数据包后。发送方在接收到ACK后，在处理ACK时会检查是否处于F-RTO处理阶段。如果是则会调用tcp_process_frto()进行F-RTO阶段的处理。

1
2
3
4
5
6
7
8
9

static int tcp_ack (struct sock *sk, const struct sk_buff *skb, int flag)  
{  
	....  
  
	if (tp->frto_counter )  
		frto_cwnd = tcp_process_frto(sk, flag);  
  
	....  
}

2.6.20的F-RTO

tcp_process_frto()用于判断RTO是否为虚假的，主要依据为RTO后的两个ACK。

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48

static void tcp_process_frto (struct sock *sk, u32 prior_snd_una)  
{  
	struct tcp_sock *tp = tcp_sk(sk);  
	tcp_sync_left_out(tp);  
  
	/* RTO was caused by loss, start retransmitting in 
	 * go-back-N slow start. 
	 * 包括两种情况： 
	  * （1）此ACK为dupack 
	 * （2）此ACK确认完整个窗口 
	  * 以上两种情况都表示有数据包丢失了，需要采用传统的方法。 
	  */  
	if (tp->snd_una == prior_snd_una ||   
		! before(tp->snd_una, tp->frto_highmark)) {  
  
		tcp_enter_frto_loss(sk);  
		return;  
	}  
  
	/* First ACK after RTO advances the window: allow two new  
	 * segments out. 
	 * frto_counter = 1表示收到第一个有效的ACK，则重新设置 
	 * 拥塞窗口，确保可以在F-RTO处理阶段在输出两个数据包， 
	 * 因为此时还没进入Loss状态，所以可以发送新数据包。 
	 */  
	if (tp->frto_counter == 1) {  
  
		tp->snd_cwnd = tcp_packets_in_flight(tp) + 2;  
  
	} else {  
  
		/* Also the second ACK after RTO advances the window. 
		 * The RTO was likely spurious. Reduce cwnd and continue 
		 * in congestion avoidance. 
		 * 第二个ACK有效，则调整拥塞窗口，直接进入拥塞避免阶段， 
		  * 而不用重传数据包。 
		  * / 
		tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh); 
		tcp_moderate_cwnd(tp); 
	} 
 
	/* F-RTO affects on two new ACKs following RTO. 
	 * At latest on third ACK the TCP behavior is back to normal. 
	 * 如果能连续收到两个确认了新数据的ACK，则说明RTO是虚假的，因此 
	  * 退出F-RTO。 
	  */  
	tp->frto_counter = (tp->frto_counter + 1) % 3;  
}

如果确定RTO为虚假的，则调用tcp_enter_frto_loss()，进入RTO恢复阶段，开始慢启动。

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53

/* Enter Loss state after F-RTO was applied. Dupack arrived after RTO, which 
 * indicates that we should follow the traditional RTO recovery, i.e. mark  
 * erverything lost and do go-back-N retransmission. 
 */  
static void tcp_enter_frto_loss (struct sock *sk)  
{  
	struct tcp_sock *tp = tcp_sk(sk);  
	struct sk_buff *skb;  
	int cnt = 0;  
  
	/* 进入Loss状态后，清零SACK、lost、retrans_out等数据*/  
	tp->sacked_out = 0;  
	tp->lost_out = 0;  
	tp->fackets_out = 0;  
  
	/* 遍历重传队列，重新标志LOST。对于那些在RTO发生后传输 
	 * 的数据不用标志为LOST。 
	 */  
	sk_stream_for_retrans_queue(skb, sk) {  
		cnt += tcp_skb_pcount(skb);  
		TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;  
  
		/* 对于那些没被SACK的数据包，需要把它标志为LOST。*/  
		if (! (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {  
			/* Do not mark those segments lost that were forward 
			 * transmitted after RTO. 
			 */  
			 if (! after(TCP_SKB_CB(skb)->end_seq, tp->frto_highmark))  
			 {  
				TCP_SKB_CB(skb)->sacked |= TCP_LOST;  
				tp->lost_out += tcp_skb_pcount(skb);  
			 }  
  
		} else { /* 对于那些已被sacked的数据包，则不用标志LOST。*/  
			tp->sacked_out += tcp_skb_pcount(skb);  
			tp->fackets_out = cnt;  
		}  
	}  
	tcp_syn_left_out(tp);  
  
	tp->snd_cwnd = tp->frto_counter + tcp_packets_in_flight(tp) + 1;  
	tp->snd_cwnd_cnt = 0;  
	tp->snd_cwnd_stamp = tcp_time_stamp;  
	tp->undo_marker = 0; /* 不需要undo标志*/  
	tp->frto_counter = 0; /* 表示F-RTO结束了*/  
  
	/* 更新乱序队列的最大值*/  
	tp->reordering = min_t(unsigned int, tp->reordering, sysctl_tcp_reordering);  
	tcp_set_ca_state(sk, TCP_CA_Loss); /* 进入loss状态*/  
	tp->high_seq = tp->frto_highmark; /*RTO时的最大序列号*/  
	TCP_ECN_queue_cwr(tp); /* 设置显示拥塞标志*/  
	clear_all_retrans_hints(tp);  
}

3.2.12的F-RTO

F-RTO spurious RTO detection algorithm (RFC4138)
F-RTO affects during two new ACKs following RTO (well, almost, see inline comments). State (ACK number) is kept in frto_counter. When ACK advances window (but not to or beyond highest sequence sent before RTO) :
On First ACK, send two new segments out.
On second ACK, RTO was likely spurious. Do spurious response (response
algorithm is not part of the F-RTO detection algorithm given in RFC4138 but
can be selected separately).

Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss and TCP falls back to conventional RTO recovery. F-RTO allows overriding of Nagle, this is done using frto_counter states 2 and 3, when a new data segment of any size sent during F-RTO, state 2 is upgraded to 3.

Rationale: if the RTO was suprious, 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.

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98

static int tcp_process_frto(struct sock *sk, int flag)  
{  
	struct tcp_sock *tp = tcp_sk(sk);  
	tcp_verify_left_out(tp);  
   
	/* Duplicate the behavior from Loss state (fastretrans_alert) */  
	if (flag & FLAG_DATA_ACKED)  
		inet_csk(sk)->icsk_retransmits = 0; /*重传次数归零*/  
   
	if ((flag & FLAG_NONHEAD_RETRANS_ACKED) ||  
		((tp->frto_counter >= 2) && (flag & FLAG_RETRANS_DATA_ACKED)))  
		tp->undo_marker = 0;  
   
	/* 一个ACK确认完RTO时整个窗口，表示出现了丢包*/  
	if (! before(tp->snd_una, tp->frto_highmark)) {  
		tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3), flag) ;  
		return 1;  
	}  
  
	/* Reno的处理方式 */  
	if (! tcp_is_sackfrto(tp)) {   
		/* RFC4138 shortcoming in step2; should also have case c): 
		 * ACK isn't duplicate nor advances window, e.g., opposite dir 
		 * data, winupdate 
		 */  
		if (! (flag & FLAG_ANY_PROGRESS) && (flag & FLAG_NOT_DUP))  
			return 1; /*不采取任何措施，忽略*/  
  
		if (! (flag & FLAG_DATA_ACKED)) { /* 没有确认新的数据*/  
			tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 0 : 3), flag);  
			return 1;  
		}  
  
	} else { /* SACK的处理方式 */  
		/* Prevent sender of new data. 表示第一个ACK没有确认新数据， 
		 * 这个时候不允许发送新的数据，直接返回。 
		 */  
		if (! (flag & FLAG_DATA_ACKED) & (tp->frto_conter == 1) {  
			tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp));  
			return 1;  
		}  
  
		/* 当第二个ACK也没有确认新的数据时，判定RTO真实，退出F-RTO。*/  
		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) {  
		/* tcp_may_send_now needs to see updated state */  
		tp->snd_cwnd = tcp_packets_in_flight(tp) + 2;  
		tp->frto_counter = 2;  
		  
		if (! tcp_may_send_now(sk))  
			tcp_enter_frto_loss(sk, 2, flag);  
		return 1;  
  
	} else {  
		switch (sysctl_tcp_frto_response) {  
		case 2: /* 比较激进的，恢复到RTO前的窗口和阈值*/  
			tcp_undo_spur_to_response(sk, flag);  
			break;  
  
		case 1: /* 非常保守，阈值减小B，可窗口一再减小，为B/2 */  
			tcp_conservative_spur_to_response(sk);  
			break;  
  
		default:  
			/* 保守*/  
			tcp_ratehalving_spur_to_response(sk);  
			break;  
		}  
  
		tp->frto_counter = 0; /*F-RTO算法结束标志*/  
		tp->undo_marker = 0; /*清零undo标志*/  
		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSPURIOUSRTOS);  
	}  
	return 0;   
}  
  
#define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */  
#define FLAG_NONHEAD_RETRANS_ACKED 0x1000 /* Non-head rexmit data was ACKed. */  
#define FLAG_RETRANS_DATA_ACKED 0x08 /* some of which was retransmitted.*/  
  
#define FLAG_ACKED (FLAG_DATA_ACKED | FLAG_SYN_ACKED)  
#define FLAG_FORWARD_PROGRESS (FLAG_ACKED | FLAG_DATA_SACKED)  
#define FLAG_ANY_PROGRESS (FLAG_RORWARD_PROGRESS | FLAG_SND_UNA_ADVANCED)  
   
#define FLAG_NOT_DUP (FLAG_DATA | FLAG_WIN_UPDATE | FLAG_ACKED)

tcp_frto_response选项

tcp_frto_response表示TCP在检测到虚假的RTO后，采用什么函数来进行阈值和拥塞窗口的调整，它有三种取值：

（1）值为2

表示使用tcp_undo_spur_to_response()，这是一种比较激进的处理方法，它把阈值和拥塞窗口都恢复到RTO前的值。

（2）值为1

表示使用tcp_conservative_spur_to_response()，这是一种很保守的处理方法。
假设减小因子为B，RTO前的窗口为C，那么一般情况下（因为阈值调整算法不同）
此后ssthresh=（1 - B）C，cwnd = （1 -B ）（1- B）C

（3）值为0或其它（默认为0）

表示使用默认的tcp_ratehalving_spur_to_response()，也是一种保守的处理方法。

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32

static void tcp_undo_spur_to_response (struct sock *sk, int flag)  
{  
	/* 如果有显示拥塞标志，则进入CWR状态，最终阈值不变，窗口减半*/  
	if (flag & FLAG_ECE)  
		tcp_ratehalving_spur_to_response(sk);  
	else  
	/* 撤销阈值调整，撤销窗口调整，恢复RTO前的状态*/  
		tcp_undo_cwr(sk, true);  
}  
  
/* 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);  
}  
  
/* A very conservative spurious RTO response algorithm: reduce cwnd 
 * and continue in congestion avoidance. 
 */  
static void tcp_conservative_spur_to_response(struct tcp_sock *tp)  
{  
	tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);  
	tp->snd_cwnd_cnt = 0;  
	tp->bytes_acked = 0;  
	/* 竟然又设置了显示拥塞标志，那窗口就还要减小到阈值的（1-B）！ 
	 * 果然是非常保守。 
	 */  
	TCP_ECN_queue_cwr(tp);   
	tcp_moderate_cwnd(tp);  
}

如果判断RTO是真实的，就调用tcp_enter_frto_loss()来处理。

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71

/* Enter Loss state after F-RTO was applied. Dupack arrived after RTO, 
 * which indicates that we should follow the tradditional 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, int flag)  
{  
	struct tcp_sock *tp = tcp_sk(sk);  
	struct sk_buff *skb;  
  
	tp->lost_out = 0;  
	tp->retrans_out = 0;  
  
	if (tcp_is_reno(tp))  
		tcp_reset_reno_sack(tp);  
  
	tcp_for_write_queue(skb, sk) {  
		if (skb == tcp_send_head(sk))  
			break;  
  
		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)) {  
			/* For some reason this R-bit might get cleared ? */  
			if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)  
				tp->retrans_out += tcp_skb_pcount(skb);  
  
			/* enter this if branch just for the first segment */  
			flag |= FLAG_DATA_ACKED;  
		} else {  
  
			if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)  
				tp->undo_marker = 0;  
			TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;  
		}  
  
		/* Marking forward transmissions that were made after RTO lost can 
		* cause unnecessary retransmissions in some scenarios, 
		* SACK blocks will mitigate that in some but not in all cases. 
		* We used to not mark them but it was casuing break-ups with 
		* receivers that do only in-order receival. 
		*  
		* TODO: we could detect presence of such receiver and select different 
		* behavior per flow. 
		*/  
	   if (! (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {  
		  TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;  
		   tp->lost_out += tcp_skb_pcount(skb);  
		   tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;  
	   }  
	}  
	tcp_verify_left_out(tp);  
  
	/* allowed_segments应该不大于3*/  
	tp->snd_cwnd = tcp_packets_in_flight(tp) + allowed_segments;  
	tp->snd_cwnd_cnt = 0;  
	tp->snd_cwnd_stamp = tcp_time_stamp;  
	tp->frto_counter = 0; /* F-RTO结束了*/  
	tp->bytes_acked = 0;  
  
	/* 更新乱序队列的最大长度*/  
	tp->reordering = min_t(unsigned int, tp->reordering,  
						sysctl_tcp_reordering);  
  
	tcp_set_ca_state(sk, TCP_CA_Loss); /*设置成Loss状态*/  
	tp->high_seq = tp->snd_nxt;  
	TCP_ECN_queue_cwr(tp); /*设置显式拥塞标志*/  
	tcp_clear_all_retrans_hints(tp);  
}

总结

现在内核（3.2.12）是默认使用F-RTO算法的。
其中tcp_frto默认为2，tcp_frto_response默认为0。

TCP拥塞状态机的实现tcp_fastretrans_alert

TCP拥塞状态机的实现（上）
TCP拥塞状态机的实现（中）
TCP拥塞状态机的实现（下）

TCP拥塞状态机的实现（上）

内容：本文主要分析TCP拥塞状态机的实现中，主体函数tcp_fastretrans_alert()的实现。接下来的文章会对其中重要的部分进行更具体的分析。

内核版本：2.6.37

原理

先来看一下涉及到的知识。

拥塞状态：

（1）Open：Normal state, no dubious events, fast path.
（2）Disorder：In all respects it is Open, but requres a bit more attention.
It is entered when we see some SACKs or dupacks. It is split of Open mainly to move some processing from fast path to slow one.
（3）CWR：cwnd was reduced due to some Congestion Notification event.
It can be ECN, ICMP source quench, local device congestion.
（4）Recovery：cwnd was reduced, we are fast-retransmitting.
（5）Loss：cwnd was reduced due to RTO timeout or SACK reneging.

tcp_fastretrans_alert() is entered：

（1）each incoming ACK, if state is not Open
（2）when arrived ACK is unusual, namely:
SACK
Duplicate ACK
ECN ECE

Counting packets in flight is pretty simple.

（1）in_flight = packets_out - left_out + retrans_out
packets_out is SND.NXT - SND.UNA counted in packets.
retrans_out is number of retransmitted segments.
left_out is number of segments left network, but not ACKed yet.

（2）left_out = sacked_out + lost_out
sacked_out：Packets, which arrived to receiver out of order and hence not ACKed. With SACK this number is simply amount of SACKed data. Even without SACKs it is easy to give pretty reliable estimate of this number, counting duplicate ACKs.

（3）lost_out：Packets lost by network. TCP has no explicit loss notification feedback from network(for now). It means that this number can be only guessed. Actually, it is the heuristics to predict lossage that distinguishes different algorithms.
F.e. after RTO, when all the queue is considered as lost, lost_out = packets_out and in_flight = retrans_out.

Essentially, we have now two algorithms counting lost packets.

1）FACK：It is the simplest heuristics. As soon as we decided that something is lost, we decide that all not SACKed packets until the most forward SACK are lost. I.e.
lost_out = fackets_out - sacked_out and left_out = fackets_out
It is absolutely correct estimate, if network does not reorder packets. And it loses any connection to reality when reordering takes place. We use FACK by defaut until reordering is suspected on the path to this destination.

2）NewReno：when Recovery is entered, we assume that one segment is lost (classic Reno). While we are in Recovery and a partial ACK arrives, we assume that one more packet is lost (NewReno).
This heuristics are the same in NewReno and SACK.
Imagine, that’s all! Forget about all this shamanism about CWND inflation deflation etc. CWND is real congestion window, never inflated, changes only according to classic VJ rules.

Really tricky (and requiring careful tuning) part of algorithm is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().

tcp_time_to_recover()

It determines the moment when we should reduce cwnd and, hence, slow down forward transmission. In fact, it determines the moment when we decide that hole is caused by loss, rather than by a reorder.

tcp_xmit_retransmit_queue()

It decides what we should retransmit to fill holes, caused by lost packets.

undo heuristics

And the most logically complicated part of algorithm is undo heuristics. We detect false retransmits due to both too early fast retransmit (reordering) and underestimated RTO, analyzing timestamps and D-SACKs. When we detect that some segments were retransmitted by mistake and CWND reduction was wrong, we undo window reduction and abort recovery phase. This logic is hidden inside several functions named tcp_try_undo_.

主体函数

TCP拥塞状态机主要是在tcp_fastretrans_alert()中实现的，tcp_fastretrans_alert()在tcp_ack()中被调用。

此函数分成几个阶段：
A. FLAG_ECE，收到包含ECE标志的ACK。
B. reneging SACKs，ACK指向已经被SACK的数据段。如果是此原因，进入超时处理，然后返回。
C. state is not Open，发现丢包，需要标志出丢失的包，这样就知道该重传哪些包了。
D. 检查是否有错误( left_out > packets_out)。
E. 各个状态是怎样退出的，当snd_una >= high_seq时候。
F. 各个状态的处理和进入。

下文会围绕这几个阶段进行具体分析。

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211

/* Process an event, which can update packets-in-flight not trivially.
 * Main goal of this function is to calculate new estimate for left_out,
 * taking into account both packets sitting in receiver's buffer and
 * packets lost by network. 
 * 
 * Besides that it does CWND reduction, when packet loss is detected
 * and changes state of machine.
 *
 * It does not decide what to send, it is made in function
 * tcp_xmit_retransmit_queue().
 */

/* 此函数被调用的条件：
 * (1) each incoming ACK, if state is not Open
 * (2) when arrived ACK is unusual, namely:
 *       SACK
 *       Duplicate ACK
 *       ECN ECE
 */

static void tcp_fastretrans_alert(struct sock *sk, int pkts_acked, int flag)
{
	struct inet_connection_sock *icsk = inet_csk(sk);
	struct tcp_sock *tp = tcp_sk(sk);

	/* 判断是不是重复的ACK*/
	int is_dupack = ! (flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));

	/* tcp_fackets_out()返回hole的大小，如果大于reordering，则认为发生丢包.*/
	int do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) && 
				(tcp_fackets_out(tp) > tp->reordering ));

	int fast_rexmit = 0, mib_idx;

	/* 如果packet_out为0，那么不可能有sacked_out */
	if (WARN_ON(!tp->packets_out && tp->sacked_out))
		tp->sacked_out = 0;

	/* fack的计数至少需要依赖一个SACK的段.*/
	if (WARN_ON(!tp->sacked_out && tp->fackets_out))
		tp->fackets_out = 0;
 
	/* Now state machine starts.
	 * A. ECE, hence prohibit cwnd undoing, the reduction is required. 
	 * 禁止拥塞窗口撤销，并开始减小拥塞窗口。
	 */
	if (flag & FLAG_ECE)
		tp->prior_ssthresh = 0;
	
	/* B. In all the states check for reneging SACKs. 
	 * 检查是否为虚假的SACK，即ACK是否确认已经被SACK的数据.
	 */
	if (tcp_check_sack_reneging(sk, flag))
		return;
	 
	/* C. Process data loss notification, provided it is valid. 
	 * 为什么需要这么多个条件？不太理解。
	 * 此时不在Open态，发现丢包，需要标志出丢失的包。
	  */
	if (tcp_is_fack(tp) && (flag & FLAG_DATA_LOSS) &&
		before(tp->snd_una, tp->high_seq) &&
		icsk->icsk_ca_state != TCP_CA_Open &&
		tp->fackets_out > tp->reordering) {
		tcp_mark_head_lost(sk, tp->fackets_out - tp->reordering, 0);
		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSS);
		}

	/* D. Check consistency of the current state. 
	 * 确定left_out < packets_out
	 */
	tcp_verify_left_out(tp); 

	/* E. Check state exit conditions. State can be terminated 
	 * when high_seq is ACKed. */
	if (icsk->icsk_ca_state == TCP_CA_Open) {
		/* 在Open状态，不可能有重传且尚未确认的段*/
		WARN_ON(tp->retrans_out != 0);
		/* 清除上次重传阶段第一个重传段的发送时间*/
		tp->retrans_stamp = 0;

	} else if (!before(tp->snd_una, tp->high_seq) {/* high_seq被确认了*/
		switch(icsk->icsk_ca_state) {
			case TCP_CA_Loss:
				icsk->icsk_retransmits = 0; /*超时重传次数归0*/ 

				/*不管undo成功与否，都会返回Open态，除非没有使用SACK*/
				if (tcp_try_undo_recovery(sk)) 
					return;
				break;
 
			case TCP_CA_CWR:
				/* CWR is to be held someting *above* high_seq is ACKed
				 * for CWR bit to reach receiver.
				 * 需要snd_una > high_seq才能撤销
				   */
				if (tp->snd_una != tp->high_seq) {
					tcp_complete_cwr(sk);
					tcp_set_ca_state(sk, TCP_CA_Open);
				}
				break;

			case TCP_CA_Disorder:
				tcp_try_undo_dsack(sk);
				 /* For SACK case do not Open to allow to undo
				  * catching for all duplicate ACKs.?*/
				if (!tp->undo_marker || tcp_is_reno(tp) || 
					tp->snd_una != tp->high_seq) {
					tp->undo_marker = 0;
					tcp_set_ca_state(sk, TCP_CA_Open);
				}

			case TCP_CA_Recovery:
				if (tcp_is_reno(tp))
					tcp_reset_reno_sack(tp)); /* sacked_out清零*/

				if (tcp_try_undo_recovery(sk))
					return;

				tcp_complete_cwr(sk);
				break;
		}
	}

	/* F. Process state. */
	switch(icsk->icsk_ca_state) {
		case TCP_CA_Recovery:
			if (!(flag & FLAG_SND_UNA_ADVANCED)) {
				if (tcp_is_reno(tp) && is_dupack)
					tcp_add_reno_sack(sk); /* 增加sacked_out ，检查是否出现reorder*/
			} else 
				do_lost = tcp_try_undo_partial(sk, pkts_acked);
			break;

		case TCP_CA_Loss:
			/* 收到partical ack，超时重传的次数归零*/
			if (flag & FLAG_DATA_ACKED)
				icsk->icsk_retransmits = 0;

			if (tcp_is_reno(tp) && flag & FLAG_SND_UNA_ADVANCED)
				tcp_reset_reno_sack(tp); /* sacked_out清零*/

			if (!tcp_try_undo_loss(sk)) { /* 尝试撤销拥塞调整，进入Open态*/
				/* 如果不能撤销，则继续重传标志为丢失的包*/
				tcp_moderate_cwnd(tp);
				tcp_xmit_retransmit_queue(sk); /* 待看*/
			   return;
			}

			if (icsk->icsk_ca_state != TCP_CA_Open)
				return;
 
		/* Loss is undone; fall through to process in Open state.*/
		default:
			if (tcp_is_reno(tp)) {
				if (flag & FLAG_SND_UNA_ADVANCED)
				   tcp_reset_reno_sack(tp);

				if (is_dupack)
				   tcp_add_reno_sack(sk);
			}

			if (icsk->icsk_ca_state == TCP_CA_Disorder)
				tcp_try_undo_dsack(sk); /*D-SACK确认了所有重传的段*/
			 
			/* 判断是否应该进入Recovery状态*/
			if (! tcp_time_to_recover(sk)) {
			   /*此过程中，会判断是否进入Open、Disorder、CWR状态*/
				tcp_try_to_open(sk, flag); 
				return;
			}

			/* MTU probe failure: don't reduce cwnd */
			/* 关于MTU探测部分此处略过！*/
			......

			/* Otherwise enter Recovery state */
			if (tcp_is_reno(tp))
				mib_idx = LINUX_MIB_TCPRENORECOVERY;
			else
				mib_idx = LINUX_MIB_TCPSACKRECOVERY;

			 NET_INC_STATS_BH(sock_net(sk), mib_idx);

			/* 进入Recovery状态前，保存那些用于恢复的数据*/
			tp->high_seq = tp->snd_nxt; /* 用于判断退出时机*/
			tp->prior_ssthresh = 0;
			tp->undo_marker = tp->snd_una;
			tp->undo_retrans=tp->retrans_out;
 
		   if (icsk->icsk_ca_state < TCP_CA_CWR) {
			   if (! (flag & FLAG_ECE))
				   tp->prior_ssthresh = tcp_current_ssthresh(sk); /*保存旧阈值*/
			   tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);/*更新阈值*/
			   TCP_ECN_queue_cwr(tp);
		   }

		   tp->bytes_acked = 0;
		   tp->snd_cwnd_cnt = 0;

		   tcp_set_ca_state(sk, TCP_CA_Recovery); /* 进入Recovery状态*/
		   fast_rexmit = 1; /* 快速重传标志 */
	}

	if (do_lost || (tcp_is_fack(tp) && tcp_head_timeout(sk)))
		/* 更新记分牌，标志丢失和超时的数据包，增加lost_out */
		tcp_update_scoreboard(sk, fast_rexmit); 

	/* 减小snd_cwnd */
	tcp_cwnd_down(sk, flag);
	tcp_xmit_retransmit_queue(sk);
}

flag标志

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34

#define FLAG_DATA 0x01  /* Incoming frame contained data. */  
#define FLAG_WIN_UPDATE 0x02  /* Incoming ACK was a window update. */  
#define FLAG_SND_UNA_ADVANCED 0x400  /* snd_una was changed (!= FLAG_DATA_ACKED) */  
#define FLAG_DATA_SACKED 0x20  /* New SACK. */  
#define FLAG_ECE 0x40  /* ECE in this ACK */  
#define FLAG_SACK_RENEGING 0x2000  /* snd_una advanced to a sacked seq */  
#define FLAG_DATA_LOST  /* SACK detected data lossage. */  
   
#define FLAG_DATA_ACKED 0x04  /* This ACK acknowledged new data. */  
#define FLAG_SYN_ACKED 0x10    /* This ACK acknowledged SYN. */  
#define FLAG_ACKED (FLAG_DATA_ACKED | FLAG_SYN_ACKED)  
   
#define FLAG_NOT_DUP (FLAG_DATA | FLAG_WIN_UPDATE | FLAG_ACKED)  /* 定义非重复ACK*/  
   
#define FLAG_FORWARD_PROGRESS (FLAG_ACKED | FLAG_DATA_SACKED)  
#define FLAG_ANY_PROGRESS (FLAG_FORWARD_PROGRESS | FLAG_SND_UNA_ADVANCED)  
#define FLAG_DSACKING_ACK 0x800  /* SACK blocks contained D-SACK info */  
  
struct tcp_sock {  
	...  
	u32 retrans_out; /*重传还未得到确认的TCP段数目*/  
	u32 retrans_stamp; /* 记录上次重传阶段，第一个段的发送时间，用于判断是否可以进行拥塞调整撤销*/  
  
	struct sk_buff *highest_sack; /* highest skb with SACK received,  
					*(validity guaranteed only if sacked_out > 0)  
					*/  
   ...  
}  
   
struct inet_connection_sock {  
	...  
	__u8 icks_retransmits; /* 记录超时重传的次数*/  
	...  
}

SACK/ RENO/ FACK是否启用

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33

/* These function determine how the currrent flow behaves in respect of SACK 
 * handling. SACK is negotiated with the peer, and therefore it can very between 
 * different flows. 
 * 
 * tcp_is_sack - SACK enabled 
 * tcp_is_reno - No SACK 
 * tcp_is_fack - FACK enabled, implies SACK enabled 
 */  
  
static inline int tcp_is_sack (const struct tcp_sock *tp)  
{  
		return tp->rx_opt.sack_ok; /* SACK seen on SYN packet */  
}  
  
static inline int tcp_is_reno (const struct tcp_sock *tp)  
{  
		return ! tcp_is_sack(tp);  
}  
  
static inline int tcp_is_fack (const struct tcp_sock *tp)  
{  
		return tp->rx_opt.sack_ok & 2;  
}  
   
static inline void tcp_enable_fack(struct tcp_sock *tp)  
{  
		tp->rx_opt.sack_ok |= 2;  
}  
   
static inline int tcp_fackets_out(const struct tcp_sock *tp)  
{  
		return tcp_is_reno(tp) ? tp->sacked_out +1 : tp->fackets_out;  
}

（1）如果启用了FACK，那么fackets_out = left_out
fackets_out = sacked_out + loss_out
所以：loss_out = fackets_out - sacked_out
这是一种比较激进的丢包估算，即FACK。

（2）如果没启用FACK，那么就假设只丢了一个数据包，所以left_out = sacked_out + 1
这是一种较为保守的做法，当出现大量丢包时，这种做法会出现问题。

TCP拥塞状态机的实现（中）

内容：本文主要分析TCP拥塞状态机的实现中，虚假SACK的处理、标志丢失数据包的详细过程。
内核版本：2.6.37

虚假SACK

state B

如果接收的ACK指向已记录的SACK，这说明记录的SACK并没有反应接收方的真实的状态，也就是说接收方现在已经处于严重拥塞的状态或者在处理上有bug，所以接下来就按照超时重传的方式去处理。因为按照正常的逻辑流程，接收的ACK不应该指向已记录的SACK，而应该指向SACK后面未接收的地方。通常情况下，此时接收方已经删除了保存到失序队列中的段。

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56

/* If ACK arrived pointing to a remembered SACK, it means that our remembered 
 * SACKs do not reflect real state of receiver i.e. receiver host is heavily congested 
 * or buggy. 
 * 
 * Do processing similar to RTO timeout. 
 */  
  
static int tcp_check_sack_reneging (struct sock *sk, int flag)  
{  
	if (flag & FLAG_SACK_RENEGING) {  
		struct inet_connection_sock *icsk = inet_csk(sk);  
		/* 记录mib信息，供SNMP使用*/  
		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);  
		  
		/* 进入loss状态，1表示清除SACKED标志*/  
		tcp_enter_loss(sk, 1);  /* 此函数在前面blog中分析过：）*/  
		  
		icsk->icsk_retransmits++; /* 未恢复的RTO加一*/  
   
		/* 重传发送队列中的第一个数据包*/  
		tcp_retransmit_skb(sk, tcp_write_queue_head(sk));   
   
		/* 更新超时重传定时器*/  
		inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,   
						icsk->icsk_rto, TCP_RTO_MAX);  
		return 1;  
	}  
	return 0;  
}  
  
/** 用于返回发送队列中的第一个数据包，或者NULL 
 * skb_peek - peek at the head of an &sk_buff_head 
 * @list_ : list to peek at  
 * 
 * Peek an &sk_buff. Unlike most other operations you must 
 * be careful with this one. A peek leaves the buffer on the 
 * list and someone else may run off with it. You must hold 
 * the appropriate locks or have a private queue to do this. 
 * 
 * Returns %NULL for an empty list or a pointer to the head element. 
 * The reference count is not incremented and the reference is therefore 
 * volatile. Use with caution. 
 */  
  
static inline struct sk_buff *skb_peek (const struct sk_buff_head *list_)  
{  
	struct sk_buff *list = ((const struct sk_buff *) list_)->next;  
	if (list == (struct sk_buff *) list_)  
		list = NULL;  
	return list;  
}  
  
static inline struct sk_buff *tcp_write_queue_head(const struct sock *sk)  
{  
	return skb_peek(&sk->sk_write_queue);  
}

tcp_retransmit_skb()用来重传一个数据包。它最终调用tcp_transmit_skb()来发送一个数据包。这个函数在接下来的blog中会分析。

重设重传定时器

state B

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70

/** inet_connection_sock - INET connection oriented sock 
 * 
 * @icsk_timeout: Timeout 
 * @icsk_retransmit_timer: Resend (no ack) 
 * @icsk_rto: Retransmission timeout 
 * @icsk_ca_ops: Pluggable congestion control hook 
 * @icsk_ca_state: Congestion control state 
 * @icsk_ca_retransmits: Number of unrecovered [RTO] timeouts 
 * @icsk_pending: scheduled timer event 
 * @icsk_ack: Delayed ACK control data 
 */  
  
struct inet_connection_sock {  
	...  
	unsigned long icsk_timeout; /* 数据包超时时间*/  
	struct timer_list icsk_retransmit_timer; /* 重传定时器*/  
	struct timer_list icsk_delack_timer; /* delay ack定时器*/  
	__u32 icsk_rto; /*超时时间*/  
	const struct tcp_congestion ops *icsk_ca_ops; /*拥塞控制算法*/  
	__u8 icsk_ca_state; /*所处拥塞状态*/  
	__u8 icsk_retransmits; /*还没恢复的timeout个数*/  
	__u8 icsk_pending; /* 等待的定时器事件*/  
	...  
	struct {  
	   ...  
		__u8 pending; /* ACK is pending */  
		unsigned long timeout; /* Currently scheduled timeout */  
		...  
	} icsk_ack; /* Delayed ACK的控制模块*/  
	...  
	u32 icsk_ca_priv[16]; /*放置拥塞控制算法的参数*/  
	...  
#define ICSK_CA_PRIV_SIZE (16*sizeof(u32))  
}  
   
#define ICSK_TIME_RETRANS 1 /* Retransmit timer */  
#define ICSK_TIME_DACK 2 /* Delayed ack timer */  
#define ICSK_TIME_PROBE0 3 /* Zero window probe timer */  
  
/* 
 * Reset the retransmissiion timer 
 */  
static inline void inet_csk_reset_xmit_timer(struct sock *sk, const int what,  
						unsigned long when,  
						const unsigned long max_when)  
{  
	struct inet_connection_sock *icsk = inet_csk(sk);  
  
	if (when > max_when) {  
#ifdef INET_CSK_DEBUG  
		pr_debug("reset_xmit_timer: sk=%p %d when=0x%lx, caller=%p\n",  
					sk, what, when, current_text_addr());  
#endif  
		when = max_when;  
	}  
	if (what == ICSK_TIME_RETRANS || what == ICSK_TIME_PROBE0) {  
		icsk->icsk_pending = what;  
		icsk->icsk_timeout = jiffies + when; /*数据包超时时刻*/  
		sk_reset_timer(sk, &icsk->icsk_retransmit_timer, icsk->icsk_timeout);  
	} else if (what == ICSK_TIME_DACK) {  
		icsk->icsk_ack.pending |= ICSK_ACK_TIMER;  
		icsk->icsk_ack.timeout = jiffies + when; /*Delay ACK定时器超时时刻*/  
		sk_reset_timer(sk, &icsk->icsk_delack_timer, icsk->icsk_ack.timeout);  
	}  
#ifdef INET_CSK_DEBUG  
	else {  
		pr_debug("%s", inet_csk_timer_bug_msg);  
	}    
#endif       
}

添加LOST标志

state C

Q: 我们发现有数据包丢失了，怎么知道要重传哪些数据包呢？
A: tcp_mark_head_lost()通过给丢失的数据包标志TCPCB_LOST，就可以表明哪些数据包需要重传。
如果通过SACK发现有段丢失，则需要从重传队首或上次标志丢失段的位置开始，为记分牌为0的段添加LOST标志，直到所有被标志LOST的段数达到packets或者被标志序号超过high_seq为止。

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73

/* Mark head of queue up as lost. With RFC3517 SACK, the packets is against sakced cnt, 
 * otherwise it's against fakced cnt. 
 * packets = fackets_out - reordering，表示sacked_out和lost_out的总和。 
 * 所以，被标志为LOST的段数不能超过packets。 
 * high_seq : 可以标志为LOST的段序号的最大值。 
 * mark_head: 为1表示只需要标志发送队列的第一个段。 
 */  
  
static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)  
{  
	struct tcp_sock *tp = tcp_sk(sk);  
	struct sk_buff *skb;  
	int cnt, oldcnt;  
	int err;  
	unsigned int mss;  
  
	/* 被标志为丢失的段不能超过发送出去的数据段数*/  
	WARN_ON(packets > tp->packets_out);  
  
	/* 如果已经有标识为丢失的段了*/  
	if (tp->lost_skb_hint) {  
		skb = tp->lost_skb_hint; /* 下一个要标志的段 */  
		cnt = tp->lost_cnt_hint; /* 已经标志了多少段 */  
  
		/* Head already handled? 如果发送队列第一个数据包已经标志了，则返回 */  
		if (mark_head && skb != tcp_write_queue_head(sk))  
			return;  
  
	} else {  
		skb = tcp_write_queue_head(sk);  
		cnt = 0;  
	}  
  
	tcp_for_write_queue_from(skb, sk) {  
		if (skb == tcp_send_head(sk))  
			break; /* 如果遍历到snd_nxt，则停止*/  
  
		/* 更新丢失队列信息*/  
		tp->lost_skb_hint = skb;  
		tp->lost_cnt_hint = cnt ;  
  
		/* 标志为LOST的段序号不能超过high_seq */  
		if (after(TCP_SKB_CB(skb)->end_seq, tp->high_seq))  
			break;  
  
		oldcnt = cnt;  
  
		if (tcp_is_fack(tp) || tcp_is_reno(tp) ||   
			(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))  
			cnt += tcp_skb_pcount(skb); /* 此段已经被sacked */  
				 
		/* 主要用于判断退出时机 */  
		if (cnt > packets) {  
			if ((tcp_is_sack(tp) && !tcp_is_fack(tp) ||   
				(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) ||  
				(oldcnt >= pakcets))  
  
				break;  
  
			 mss = skb_shinfo(skb)->gso_size;  
			 err = tcp_fragment(sk, skb, (packets - oldcnt) * mss, mss);  
			 if (err < 0)  
				 break;  
			 cnt = packets;  
		}  
  
		/* 标志动作：标志一个段为LOST*/  
		tcp_skb_mark_lost(tp, skb);  
		if (mark_head)  
			break;  
	}  
	tcp_verify_left_out(tp);  
}

涉及变量

1
2
3
4
5
6
7
8
9

struct tcp_sock {  
	/* 在重传队列中，缓存下次要标志的段，为了加速对重传队列的标志操作 */  
	struct sk_buff *lost_skb_hint; /* 下一次要标志的段 */  
	int lost_cnt_hint; /* 已经标志了多少个段 */  
  
	struct sk_buff *retransmit_skb_hint; /* 表示将要重传的起始包*/  
	u32 retransmit_high; /*重传队列的最大序列号*/  
	struct sk_buff *scoreboard_skb_hint; /* 记录超时的数据包，序号最大*/  
}

TCP分片函数tcp_fragment

1
2
3
4
5
6
7
8

/* Function to create two new TCP segments. shrinks the given segment 
 * to the specified size and appends a new segment with the rest of the 
 * packet to the list. This won't be called frequently, I hope. 
 * Remember, these are still headerless SKBs at this point. 
 */  
  
int tcp_fragment (struct sock *sk, struct sk_buff *skb, u32 len,  
				unsigned int mss_now) {}

给一个段添加一个LOST标志

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21

static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb)  
{  
	if (! (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST | TCPCB_SACKED_ACKED))) {  
		tcp_verify_retransmit_hint(tp, skb); /* 更新重传队列*/  
		tp->lost_out += tcp_skb_pcount(skb); /*增加LOST的段数*/  
		TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; /* 添加LOST标志*/  
	}  
}  
  
/* This must be called before lost_out is incremented */  
static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)  
{  
	if ((tp->retransmit_skb_hint == NULL) ||  
		 before(TCP_SKB_CB(skb)->seq,  
					   TCP_SKB_CB(tp->retransmit_skb_hint)->seq))  
	tp->retransmit_skb_hint = skb;   
   
	if (! tp->lost_out ||  
		after(TCP_SKB_CB(skb)->end_seq, tp->retransmit_high))  
		tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;  
}

TCP拥塞状态机的实现（下）

内容：本文主要分析TCP拥塞状态机的实现中，各个拥塞状态的进入、处理和退出的详细过程。
内核版本：2.6.37

各状态的退出

state E

各状态的退出时机：tp->snd_una >= tp->high_seq

（1） Open

因为Open态是正常态，所以无所谓退出，保持原样。

（2）Loss

icsk->icsk_retransmits = 0; /超时重传次数归0/
tcp_try_undo_recovery(sk);

检查是否需要undo，不管undo成功与否，都返回Open态。

（3）CWR

If seq number greater than high_seq is acked, it indicates that the CWR indication has reached the peer TCP, call tcp_complete_cwr() to bring down the cwnd to ssthresh value.

tcp_complete_cwr(sk)中：
tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);

（4）Disorder

启用sack，则tcp_try_undo_dsack(sk)，交给它处理。否则，tp->undo_marker = 0;

（5）Recovery

tcp_try_undo_recovery(sk);
在tcp_complete_cwr(sk)中：
tp->snd_cwnd = tp->snd_ssthresh;

1
2
3
4
5
6
7
8
9

/*cwr状态或Recovery状态结束时调用，减小cwnd*/   
  
static inline void tcp_complete_cwr(struct sock *sk)  
{  
	struct tcp_sock *tp = tcp_sk(sk);  
	tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);  
	tp->snd_cwnd_stamp = tcp_time_stamp;  
	tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);  
}

Recovery状态处理

state F

（1）收到dupack

如果收到的ACK并没有使snd_una前进、是重复的ACK，并且没有使用SACK，则：
sacked_out++，增加sacked数据包的个数。
检查是否有reordering，如果有reordering则：
纠正sacked_out
禁用FACK(画外音：这实际上是多此一举，没有使用SACK，哪来的FACK？)
更新tp->reordering

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41

/* Emulate SACKs for SACKless connection: account for a new dupack.*/  
static void tcp_add_reno_sack(struct sock *sk)  
{  
	struct tcp_sock *tp = tcp_sk(sk);  
	tp->sacked_out++; /* 增加sacked数据包个数*/  
	tcp_check_reno_reordering(sk, 0); /*检查是否有reordering*/  
	tcp_verify_left_out(tp);  
}  
   
/* If we receive more dupacks than we expected counting segments in  
 * assumption of absent reordering, interpret this as reordering. 
 * The only another reason could be bug in receiver TCP. 
 * tcp_limit_reno_sack()是判断是否有reordering的函数。 
 */  
static void tcp_check_reno_reordering(struct sock *sk, const int addend)  
{  
	struct tcp_sock *tp = tcp_sk(sk);  
	if (tcp_limit_reno_sack(tp)) /* 检查sack是否过多*/  
		/* 如果是reordering则更新reordering信息*/  
		tcp_update_reordering(sk, tp->packets_out + addend, 0);  
}  
   
/* Limit sacked_out so that sum with lost_out isn't ever larger than packets_out. 
 * Returns zero if sacked_out adjustment wasn't necessary. 
 * 检查sacked_out是否过多，过多则限制，且返回1说明出现reordering了。 
 * Q: 怎么判断是否有reordering呢？ 
 * A: 我们知道dupack可能由lost引起，也有可能由reorder引起，那么如果 
 *    sacked_out + lost_out > packets_out，则说明sacked_out偏大了，因为它错误的把由reorder 
 *    引起的dupack当客户端的sack了。 
 */  
static int tcp_limit_reno_sacked(struct tcp_sock *tp)  
{  
	u32 holes;  
	holes = max(tp->lost_out, 1U);  
	holes = min(holes, tp->packets_out);  
	if ((tp->sacked_out + holes) > tp->packets_out) {  
		tp->sacked_out = tp->packets_out - holes;  
		return 1;  
	}  
	return 0;  
}

更新reordering信息

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40

static void tcp_update_reordering(struct sock *sk, const int metric,  
					const int ts)  
{  
	struct tcp_sock *tp = tcp_sk(sk);  
  
	if (metric > tp->reordering) {  
		int mib_idx;  
		/* 更新reordering的值，取其小者*/  
		tp->reordering = min(TCP_MAX_REORDERING, metric);  
		  
		if (ts)  
			mib_idx = LINUX_MIB_TCPTSREORDER;  
		else if (tcp_is_reno(tp))  
			mib_idx = LINUX_MIB_TCPRENOREORDER;  
		else if (tcp_is_fack(tp))  
			mib_idx = LINUX_MIB_TCPFACKREORDER;  
		else   
			mib_idx = LINUX_MIB_TCPSACKREORDER;  
  
		NET_INC_STATS_BH(sock_net(sk), mib_idx);  
#if FASTRETRANS_DEBUG > 1  
		printk(KERN_DEBUG "Disorder%d %d %u f%u s%u rr%d\n",  
				   tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,  
				   tp->reordering, tp->fackets_out, tp->sacked_out,  
				   tp->undo_marker ? tp->undo_retrans : 0);  
#endif  
		tcp_disable_fack(tp); /* 出现了reorder，再用fack就太激进了*/  
	}  
}  
/* Packet counting of FACK is based on in-order assumptions, therefore 
 * TCP disables it when reordering is detected. 
 */  
  
static void tcp_disable_fack(struct tcp_sock *tp)  
{  
	/* RFC3517 uses different metric in lost marker => reset on change */  
	if (tcp_is_fack(tp))  
		tp->lost_skb_hint = NULL;  
	tp->rx_opt.sack_ok &= ~2; /* 取消FACK选项*/  
}
（2）收到partical ack

do_lost = tcp_try_undo_partical(sk, pkts_acked);
一般情况下do_lost都会为真，除非需要undo。
具体可以看前面blog《TCP拥塞窗口调整撤销剖析》。

（3）跳出F state，标志丢失的数据段

执行完(1)或(2)后，就跳出F state。
如果有丢失的数据包，或者发送队列的第一个数据包超时，则调用tcp_update_scoreboard()来更新记分牌，给丢失的段加TCPCB_LOST标志，增加lost_out。

检查发送队列的第一个数据包是否超时。

1
2
3
4
5
6
7
8
9
10
11
12
13
14

/* 检验发送队列的第一个数据包是否超时*/  
static inline int tcp_head_timeout(const struct sock *sk)  
{  
	const struct tcp_sock *tp = tcp_sk(sk);  
	return tp->packets_out &&   
				tcp_skb_timeout(sk, tcp_write_queue_head(sk));  
}  
  
/* 检验发送队列的某个数据包是否超时*/  
static inline int tcp_skb_timeout(const struct sock *sk,  
				 const struct sk_buff *skb)  
{  
	return tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto;  
}

为确定丢失的段更新记分牌，记分牌指的是tcp_skb_cb结构中的sacked，保存该数据包的状态信息。
(1) 没有使用SACK，每次收到dupack或partical ack时，只能标志一个包为丢失。

(2) 使用FACK，每次收到dupack或partical ack时，分两种情况：
如果lost = fackets_out - reordering <= 0，这时虽然不能排除是由乱序引起的，但是fack的思想较为激进，所以也标志一个包为丢失。
如果lost >0，就可以肯定有丢包，一次性可以标志lost个包为丢失。

(3) 使用SACK，但是没有使用FACK。
如果sacked_upto = sacked_out - reordering，这是不能排除是由乱序引起的，除非快速重传标志fast_rexmit为真，才标志一个包为丢失。
如果sacked_upto > 0，就可以肯定有丢包，一次性可以标志sacked_upto个包为丢失。

内核默认使用的是(2)。

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30

/* Account newly detected lost packet(s) */  
  
 static void tcp_update_scoreboard (struct sock *sk, int fast_rexmit)  
{  
	struct tcp_sock *tp = tcp_sk(sk);  
	if (tcp_is_reno(tp)) {  
		/* 只标志第一个数据包为丢失，reno一次性只标志一个包*/  
		tcp_mark_head_lost(sk, 1, 1);  
  
	} else if (tcp_is_fack(tp)) {  
		/* 还是考虑到乱序的，对于可能是由乱序引起的部分，一次标志一个包*/  
		int lost = tp->fackets_out - tp->reordering;  
		if (lost <= 0)  
			lost = 1;  
  
		/* 因为使用了FACK，可以标志多个数据包丢失*/  
		tcp_mark_head_lost(sk, lost, 0);  
  
	} else {  
		int sacked_upto = tp->sacked_out - tp->reordering;  
		if (sacked_upto >= 0)  
			tcp_mark_head_lost(sk, sacked_upto, 0);  
  
		else if (fast_rexmit)  
			tcp_mark_head_lost(sk, 1, 1);  
	}  
  
	/* 检查发送队列中的数据包是否超时，如果超时则标志为丢失*/  
	tcp_timeout_skbs(sk);  
}

检查发送队列中哪些数据包超时，并标志为丢失

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26

static void tcp_timeout_skbs(struct sock *sk)  
{  
	struct tcp_sock *tp = tcp_sk(sk);  
	struct sk_buff *skb;  
  
	if (! tcp_is_fack(tp) || !tcp_head_timeout(sk))  
		return;  
  
	skb = tp->scoreboard_skb_hint;  
  
	if (tp->scoreboard_skb_hint == NULL)  
		skb = tcp_write_queue_head(sk));  
  
	tcp_for_write_queue_from(skb, sk) {  
		if (skb == tcp_send_head(sk)) /*遇到snd_nxt则停止*/  
			break;  
  
		if (!tcp_skb_timeout(sk, skb)) /* 数据包不超时则停止*/  
			break;  
  
		tcp_skb_mark_lost(tp, skb); /* 标志为LOST，并增加lost_out */  
	}  
  
	tp->scoreboard_skb_hint = skb;  
	tcp_verify_left_out(tp);  
}
（4）减小snd_cwnd

拥塞窗口每隔一个确认段减小一个段，即每收到2个确认将拥塞窗口减1，直到拥塞窗口等于慢启动阈值为止。

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30

/* Decrease cwnd each second ack. */  
static void tcp_cwnd_down (struct sock *sk, int flag)  
{  
	struct tcp_sock *tp = tcp_sk(sk);  
	int decr = tp->snd_cwnd_cnt + 1;  
  
	if ((flag & (FLAG_ANY_PROGRESS | FLAG_DSACKING_ACK )) ||  
		(tcp_is_reno(tp) && ! (flag & FLAG_NOT_DUP))) {  
		tp->snd_cwnd_cnt = decr & 1; /* 0=>1,1=>0 */  
  
		decr >>= 1; /*与上个snd_cwnd_cnt相同，0或1*/  
  
		/* 减小cwnd */  
		if (decr && tp->snd_cwnd > tcp_cwnd_min(sk))  
			tp->snd_cwnd -= decr;  
			  
		/* 注：不太理解这句的用意。*/  
		tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) +1);  
		tp->snd_cwnd_stamp = tcp_time_stamp;  
	}  
}  
  
/* Lower bound on congestion window is slow start threshold 
 * unless congestion avoidance choice decides to override it. 
 */  
static inline u32 tcp_cwnd_min(const struct sock *tp)  
{  
	const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;  
	return ca_ops->min_cwnd ? ca_ops->min_cwnd(sk) : tcp_sk(sk)->snd_ssthresh;  
}
（5）重传标志为丢失的段
1
2
3
4
5
6
7
8

/* This gets called after a retransmit timeout, and the initially retransmitted data is  
 * acknowledged. It tries to continue resending the rest of the retransmit queue, until  
 * either we've sent it all or the congestion window limit is reached. If doing SACK,  
 * the first ACK which comes back for a timeout based retransmit packet might feed us  
 * FACK information again. If so, we use it to avoid unnecessarily retransmissions. 
 */  
  
void tcp_xmit_retransmit_queue (struct sock *sk) {}

这个函数决定着发送哪些包，比较复杂，会在之后的blog单独分析。

（6）什么时候进入Recovery状态

tcp_time_to_recover()是一个重要函数，决定什么时候进入Recovery状态。

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61

/* This function decides, when we should leave Disordered state and enter Recovery 
 * phase, reducing congestion window. 
 * 决定什么时候离开Disorder状态，进入Recovery状态。 
 */  
  
static int tcp_time_to_recover(struct sock *sk)  
{  
	struct tcp_sock *tp = tcp_sk(sk);  
	__u32 packets_out;  
  
	/* Do not perform any recovery during F-RTO algorithm 
	 * 这说明Recovery状态不能打断Loss状态。 
	 */  
	if (tp->frto_counter)  
		return 0;  
  
	/* Trick#1: The loss is proven.  
	 * 如果传输过程中存在丢失段，则可以进入Recovery状态。 
	 */  
	if (tp->lost_out)  
		return 1;  
   
	/* Not-A-Trick#2 : Classic rule... 
	 * 如果收到重复的ACK大于乱序的阈值，表示有数据包丢失了， 
	 * 可以进入到Recovery状态。 
	 */  
	if (tcp_dupack_heuristics(tp) > tp->reordering)  
		return 1;  
   
	/* Trick#3 : when we use RFC2988 timer restart, fast 
	 * retransmit can be triggered by timeout of queue head. 
	 * 如果发送队列的第一个数据包超时，则进入Recovery状态。 
	 */  
	  if (tcp_is_fack(tp) && tcp_head_timeout(sk))  
		 return 1;  
  
	/* Trick#4 : It is still not OK... But will it be useful to delay recovery more? 
	 * 如果此时由于应用程序或接收窗口的限制而不能发包，且接收到很多的重复ACK。那么不能再等下去了， 
	 * 推测发生了丢包，且马上进入Recovery状态。 
	 */  
	if (packets_out <= tp->reordering &&  
		tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering)  
		&& ! tcp_may_send_now(sk)  ) {  
		/* We have nothing to send. This connection is limited 
		 * either by receiver window or by application. 
		 */  
		return 1;  
	}  
  
	/* If a thin stream is detected, retransmit after first received 
	 * dupack. Employ only if SACK is supported in order to avoid  
	 * possible corner-case series of spurious retransmissions 
	 * Use only if there are no unsent data. 
	 */  
	if ((tp->thin_dupack || sysctl_tcp_thin_dupack) &&  
		 tcp_stream_is_thin(tp) && tcp_dupack_heuristics(tp) > 1 &&  
		 tcp_is_sack(tp) && ! tcp_send_head(sk))  
		 return 1;  
  
	return 0; /*表示为假*/  
}
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33

/* Heurestics to calculate number of duplicate ACKs. There's no  
 * dupACKs counter when SACK is enabled (without SACK, sacked_out 
 * is used for that purpose). 
 * Instead, with FACK TCP uses fackets_out that includes both SACKed 
 * segments up to the highest received SACK block so far and holes in 
 * between them. 
 * 
 * With reordering, holes may still be in filght, so RFC3517 recovery uses 
 * pure sacked_out (total number of SACKed segment) even though it 
 * violates the RFC that uses duplicate ACKs, often these are equal but 
 * when e.g. out-of-window ACKs or packet duplication occurs, they differ. 
 * Since neither occurs due to loss, TCP shuld really ignore them. 
 */  
static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)  
{  
	return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1;  
}  
  
  
/* Determines whether this is a thin stream (which may suffer from increased 
 * latency). Used to trigger latency-reducing mechanisms. 
 */  
static inline unsigned int tcp_stream_is_thin(struct tcp_sock *tp)  
{  
	return tp->packets_out < 4 && ! tcp_in_initial_slowstart(tp);  
}  
  
#define TCP_INFINITE_SSTHRESH 0x7fffffff  
  
static inline bool tcp_in_initial_slowstart(const struct tcp_sock *tp)  
{  
	return tp->snd_ssthresh >= TCP_INFINITE_SSTHRESH;  
}

This function examines various parameters (like number of packet lost) for TCP connection to decide whether it is the right time to move to Recovery state. It’s time to recover when TCP heuristics suggest a strong possibility of packet loss in the network, the following checks are made.

总的来说，一旦确定有丢包，或者很可能丢包，就可以进入Recovery状态恢复丢包了。

可以进入Recovery状态的条件包括：
(1) some packets are lost (lost_out is non zero)。发现有丢包。

(2) SACK is an acknowledgement for out of order packets. If number of packets Sacked is greater than the
reordering metrics of the network, then loss is assumed to have happened.
被fack数据或收到的重复ACK，大于乱序的阈值，表明很可能发生丢包。

(3) If the first packet waiting to be acked (head of the write Queue) has waited for time equivalent to retransmission
timeout, the packet is assumed to have been lost. 发送队列的第一个数据段超时，表明它可能丢失了。

(4) If the following three conditions are true, TCP sender is in a state where no more data can be transmitted
and number of packets acked is big enough to assume that rest of the packets are lost in the network:
A: If packets in flight is less than the reordering metrics.
B: More than half of the packets in flight have been sacked by the receiver or number of packets sacked is more
than the Fast Retransmit thresh. (Fast Retransmit thresh is the number of dupacks that sender awaits before
fast retransmission)
C: The sender can not send any more packets because either it is bound by the sliding window or the application
has not delivered any more data to it in anticipation of ACK for already provided data.
我们收到很多的重复ACK，那么很可能有数据段丢失了。如果此时由于接收窗口或应用程序的限制而不能发送数据，那么我们不打算再等下去，直接进入Recovery状态。

(5) 当检测到当前流量很小时（packets_out < 4），如果还满足以下条件：
A: tp->thin_dupack == 1 / Fast retransmit on first dupack /
或者sysctl_tcp_thin_dupack为1，表明允许在收到第一个重复的包时就重传。
B: 启用SACK，且FACK或SACK的数据量大于1。
C: 没有未发送的数据，tcp_send_head(sk) == NULL。
这是一种特殊的情况，只有当流量非常小的时候才采用。

（7）刚进入Recovery时的设置
保存那些用于undo的数据：
tp->prior_ssthresh = tp->snd_ssthresh; / 保存旧阈值/
tp->undo_marker = tp->snd_una; / tracking retrans started here./
tp->undo_retrans = tp->retrans_out; / Retransmitted packets out /

保存退出点：
tp->high_seq = tp->snd_nxt;

重置变量：
tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
tp->bytes_acked = 0;
tp->snd_cwnd_cnt = 0;

进入Recovery状态：
tcp_set_ca_state(sk, TCP_CA_Recovery);

Loss状态处理

state F

（1）收到partical ack

icsk->icsk_retransmits = 0; / 超时重传的次数归零/
如果使用的是reno，没有使用sack，则归零tp->sacked_out。

（2）尝试undo

调用tcp_try_undo_loss()，当使用时间戳检测到一个不必要的重传时：
移除记分牌中所有段的Loss标志，从而发送新的数据而不再重传。
调用tcp_undo_cwr()来撤销拥塞窗口和阈值的调整。

否则：
tcp_moderate_cwnd()调整拥塞窗口，防止爆发式重传。
tcp_xmit_retransmit_queue()继续重传丢失的数据段。

其它状态处理

state F

如果tcp_time_to_recover(sk)返回值为假，也就是说不能进入Recovery状态，则进行CWR、Disorder或Open状态的处理。

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35

static void tcp_try_to_open (struct sock *sk, int flag)  
{  
	struct tcp_sock *tp = tcp_sk(sk);  
	tcp_verify_left_out(tp);  
  
	if (!tp->frto_conter && !tcp_any_retrans_done(sk))  
		tp->retrans_stamp = 0; /* 归零，因为不需要undo了*/  
  
	/* 判断是否需要进入CWR状态*/  
	if (flag & FLAG_ECE)  
		tcp_enter_cwr(sk, 1);  
   
	if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) { /*没进入CWR*/  
		tcp_try_keep_open(sk); /* 尝试保持Open状态*/  
		tcp_moderate_cwnd(tp);  
  
	} else { /* 说明进入CWR状态*/  
		tcp_cwnd_down(sk, flag);/* 每2个ACK减小cwnd*/  
	}  
}  
  
static void tcp_try_keep_open(struct sock *sk)  
{  
	struct tcp_sock *tp = tcp_sk(sk);  
	int state = TCP_CA_Open;  
	  
	/* 是否需要进入Disorder状态*/  
	if (tcp_left_out(tp) || tcp_any_retrans_done(sk) || tp->undo_marker)  
		state = TCP_CA_Disorder;  
  
	if (inet_csk(sk)->icsk_ca_state != state) {  
		tcp_set_ca_state(sk, state);  
		tp->high_seq = tp->snd_nxt;  
	}  
}
（1）CWR状态

Q: 什么时候进入CWR状态？
A: 如果检测到ACK包含ECE标志，表示接收方通知发送法进行显示拥塞控制。

1
2
3

 @tcp_try_to_open():
 if (flag & FLAG_ECE)
	 tcp_enter_cwr(sk, 1);

tcp_enter_cwr()函数分析可见前面blog《TCP拥塞状态变迁》。
它主要做了：
1. 重新设置慢启动阈值。
2. 清除undo需要的标志，不允许undo。
3. 记录此时的最高序号(high_seq = snd_nxt)，用于判断退出时机。
4. 添加CWR标志，用于通知接收方它已经做出反应。
5. 设置此时的状态为TCP_CA_CWR。

Q: 在CWR期间采取什么措施？
A: 拥塞窗口每隔一个确认段减小一个段，即每收到2个确认将拥塞窗口减1，直到拥塞窗口等于慢启动阈值为止。
调用tcp_cwnd_down()。

（2）Disorder状态

Q: 什么时候进入Disorder状态？
A: 如果检测到有被sacked的数据包，或者有重传的数据包，则进入Disorder状态。
当然，之前已经确认不能进入Loss或Recovery状态了。
判断条件： sacked_out、lost_out、retrans_out、undo_marker不为0。

Q: 在Disorder期间采取什么措施？
A: 1. 设置CA状态为TCP_CA_Disorder。
2. 记录此时的最高序号(high_seq = snd_nxt)，用于判断退出时机。
3. 微调拥塞窗口，防止爆发式传输。

In Disorder state TCP is still unsure of genuiness of loss, after receiving acks with sack there may be a clearing ack which acks many packets non dubiously in one go. Such a clearing ack may cause a packet burst in the network, to avoid this cwnd size is reduced to allow no more than max_burst (usually 3) number of packets.

（3）Open状态

因为Open状态是正常的状态，是状态处理的最终目的，所以不需要进行额外处理。

TCP接收缓存大小的动态调整

http://blog.csdn.net/zhangskd/article/details/8200048

引言

TCP中有拥塞控制，也有流控制，它们各自有什么作用呢？

拥塞控制(Congestion Control) — A mechanism to prevent a TCP sender from overwhelming the network.
流控制(Flow Control) — A mechanism to prevent a TCP sender from overwhelming a TCP receiver.

下面是一段关于流控制原理的简要描述。
“The basic flow control algorithm works as follows: The receiver communicates to the sender the maximum amount of data it can accept using the rwnd protocol field. This is called the receive window. The TCP sender then sends no more than this amount of data across the network. The TCP sender then stops and waits for acknowledgements back from the receiver. When acknowledgement of the previously sent data is returned to the sender, the sender then resumes sending new data. It’s essentially the old maxim hurry up and wait. ”

由于发送速度可能大于接收速度、接收端的应用程序未能及时从接收缓冲区读取数据、接收缓冲区不够大不能缓存所有接收到的报文等原因，TCP接收端的接收缓冲区很快就会被塞满，从而导致不能接收后续的数据，发送端此后发送数据是无效的，因此需要流控制。TCP流控制主要用于匹配发送端和接收端的速度，即根据接收端当前的接收能力来调整发送端的发送速度。

TCP流控制中一个很重要的地方就是，TCP接收缓存大小是如何动态调整的，即TCP确认窗口上限是如何动态调整的？

本文主要分析TCP接收缓存大小动态调整的原理和实现。

原理

早期的TCP实现中，TCP接收缓存的大小是固定的。随着网络的发展，固定的TCP接收缓存值就不适应了，成为TCP性能的瓶颈之一。这时候就需要手动去调整，因为不同的网络需要不同大小的TCP接收缓存，手动调整不仅费时费力，还会引起一些问题。TCP接收缓存设置小了，就不能充分利用网络。而TCP缓存设置大了，又浪费了内存。

如果把TCP接收缓存设置为无穷大，那就更糟糕了，因为某些应用可能会耗尽内存，使其它应用的连接陷入饥饿。所以TCP接收缓存的大小需要动态调整，才能达到最佳的效果。

动态调整TCP接收缓存大小，就是使TCP接收缓存按需分配，同时要确保TCP接收缓存大小不会成为传输的限制。

linux采用Dynamic Right-Sizing方法来动态调整TCP的接收缓存大小，其基本思想就是：通过估算发送方的拥塞窗口的大小，来动态设置TCP接收缓存的大小。

It has been demomstrated that this method can successfully grow the receiver’s advertised window at a pace sufficient to avoid constraining the sender’s throughput. As a result, systems can avoid the network performance problems that result from either the under-utilization or over-utilization of buffer space.

实现

下文代码基于3.2.12内核，主要源文件为：net/ipv4/tcp_input.c。

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35

struct tcp_sock {  
	...  
	u32 rcv_nxt; /* What we want to receive next，希望接收的下一个序列号 */  
	u32 rcv_wnd; /* Current receiver window，当前接收窗口的大小*/  
	u32 copied_seq; /* Head of yet unread data，应用程序下次从这里复制数据 */  
	u16 advmss; /* Advertised MSS，接收端通告的MSS */  
	u32 window_clamp; /* Maximal window to advertise，通告窗口的上限*/  
  
	/* Receiver side RTT estimation */  
	struct {  
		u32 rtt;  
		u32 seq;  
		u32 time;  
	} rcv_rtt_est; /* 用于接收端的RTT测量*/  
  
	/* Receiver queue space */  
	struct {  
		int space;  
		u32 seq;  
		u32 time;  
	} rcvq_space; /* 用于调整接收缓冲区和接收窗口*/  
  
	/* Options received (usually on last packet, some only on SYN packets). */  
	struct tcp_options_received rx_opt; /* TCP选项*/  
	...  
};  
  
struct sock {  
	...  
	int sk_rcvbuf; /* TCP接收缓冲区的大小*/  
	int sk_sndbuf; /* TCP发送缓冲区大小*/  
	unsigned int ...  
		sk_userlocks : 4, /*TCP接收缓冲区的锁标志*/  
	...  
};

RTT测量

在发送端有两种RTT的测量方法(具体可见前面blog)，但是因为TCP流控制是在接收端进行的，所以接收端也需要有测量RTT的方法。

（1）没有时间戳时的测量方法
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17

static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)  
{  
	/* 第一次接收到数据时，需要对相关变量初始化*/  
	if (tp->rcv_rtt_est.time == 0)  
		goto new_measure;  
  
	/* 收到指定的序列号后，才能获取一个RTT测量样本*/  
	if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))  
		return;  
  
	/* RTT的样本：jiffies - tp->rcv_rtt_est.time */  
	tcp_rcv_rtt_update(tp, jiffies - tp->rcv_rtt_est.time, 1);  
  
new_measure:  
	tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd; /* 收到此序列号的ack时，一个RTT样本的计时结束*/  
	tp->rcv_rtt_est.time = tcp_time_stamp; /* 一个RTT样本开始计时*/  
}

此函数在接收到带有负载的数据段时被调用。

此函数的原理：我们知道发送端不可能在一个RTT期间发送大于一个通告窗口的数据量。那么接收端可以把接收一个确认窗口的数据量(rcv_wnd)所用的时间作为RTT。接收端收到一个数据段，然后发送确认(确认号为rcv_nxt，通告窗口为rcv_wnd)，开始计时，RTT就是收到序号为rcv_nxt + rcv_wnd的数据段所用的时间。很显然，这种假设并不准确，测量所得的RTT会偏大一些。所以这种方法只有当没有采用时间戳选项时才使用，而内核默认是采用时间戳选项的(tcp_timestamps为1)。

下面是一段对此方法的评价：
If the sender is being throttled by the network, this estimate will be valid. However, if the sending application did not have any data to send, the measured time could be much larger than the actual round-trip time. Thus this measurement acts only as an upper-bound on the round-trip time.

（2）采用时间戳选项时的测量方法
1
2
3
4
5
6
7
8
9

static inline void tcp_rcv_rtt_measure_ts(struct sock *sk, const struct sk_buff *skb)  
{  
	struct tcp_sock *tp = tcp_sk(sk);  
	/* 启用了Timestamps选项，并且流量稳定*/  
	if (tp->rx_opt.rcv_tsecr && (TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq >=  
		inet_csk(sk)->icsk_ack.rcv_mss))  
		/* RTT = 当前时间 - 回显时间*/  
		tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);  
}

虽然此种方法是默认方法，但是在流量小的时候，通过时间戳采样得到的RTT的值会偏大，此时就会采用没有时间戳时的RTT测量方法。

（3）采样处理

不管是没有使用时间戳选项的RTT采样，还是使用时间戳选项的RTT采样，都是获得一个RTT样本。之后还需要对获得的RTT样本进行处理，以得到最终的RTT。

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34

/* win_dep表示是否对RTT采样进行微调，1为不进行微调，0为进行微调。*/  
static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)  
{  
	u32 new_sample = tp->rcv_rtt_est.rtt;  
	long m = sample;  
  
	if (m == 0)  
		m = 1; /* 时延最小为1ms*/  
  
	if (new_sample != 0) { /* 不是第一次获得样本*/  
		/* If we sample in larger samples in the non-timestamp case, we could grossly 
		 * overestimate the RTT especially with chatty applications or bulk transfer apps 
		 * which are stalled on filesystem I/O. 
		 * 
		 * Also, since we are only going for a minimum in the non-timestamp case, we do 
		 * not smooth things out else with timestamps disabled convergence takes too long. 
		 */  
		/* 对RTT采样进行微调，新的RTT样本只占最终RTT的1/8 */  
		if (! win_dep) {   
			m -= (new_sample >> 3);  
			new_sample += m;  
  
		} else if (m < new_sample)  
			/* 不对RTT采样进行微调，直接取最小值，原因可见上面那段注释*/  
			new_sample = m << 3;   
  
	} else {   
		/* No previous measure. 第一次获得样本*/  
		new_sample = m << 3;  
	}  
  
	if (tp->rcv_rtt_est.rtt != new_sample)  
		tp->rcv_rtt_est.rtt = new_sample; /* 更新RTT*/  
}

对于没有使用时间戳选项的RTT测量方法，不进行微调。因为用此种方法获得的RTT采样值已经偏高而且收敛很慢。直接选择最小RTT样本作为最终的RTT测量值。
对于使用时间戳选项的RTT测量方法，进行微调，新样本占最终RTT的1/8，即rtt = 7/8 old + 1/8 new。

调整接收缓存

当数据从TCP接收缓存复制到用户空间之后，会调用tcp_rcv_space_adjust()来调整TCP接收缓存和接收窗口上限的大小。

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82

/*  
 * This function should be called every time data is copied to user space. 
 * It calculates the appropriate TCP receive buffer space. 
 */  
void tcp_rcv_space_adjust(struct sock *sk)  
{  
	struct tcp_sock *tp = tcp_sk(sk);  
	int time;  
	int space;  
  
	/* 第一次调整*/  
	if (tp->rcvq_space.time == 0)  
		goto new_measure;  
  
	time = tcp_time_stamp - tp->rcvq_space.time; /*计算上次调整到现在的时间*/  
  
	/* 调整至少每隔一个RTT才进行一次，RTT的作用在这里！*/  
	if (time < (tp->rcv_rtt_est.rtt >> 3) || tp->rcv_rtt_est.rtt == 0)  
		return;  
  
	/* 一个RTT内接收方应用程序接收并复制到用户空间的数据量的2倍*/  
	space = 2 * (tp->copied_seq - tp->rcvq_space.seq);  
	space = max(tp->rcvq_space.space, space);  
  
	/* 如果这次的space比上次的大*/  
	if (tp->rcvq_space.space != space) {  
		int rcvmem;  
		tp->rcvq_space.space = space; /*更新rcvq_space.space*/  
  
		/* 启用自动调节接收缓冲区大小，并且接收缓冲区没有上锁*/  
		if (sysctl_tcp_moderate_rcvbuf && ! (sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {  
			int new_clamp = space;  
			/* Receive space grows, normalize in order to take into account packet headers and 
			 * sk_buff structure overhead. 
			 */  
			 space /= tp->advmss; /* 接收缓冲区可以缓存数据包的个数*/  
  
			 if (!space)  
				space = 1;  
  
			/* 一个数据包耗费的总内存包括： 
			   * 应用层数据：tp->advmss， 
			   * 协议头：MAX_TCP_HEADER， 
			   * sk_buff结构， 
			   * skb_shared_info结构。 
			   */  
			 rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER);  
  
			 /* 对rcvmem进行微调*/  
			 while(tcp_win_from_space(rcvmem) < tp->advmss)  
				 rcvmem += 128;  
  
			 space *= rcvmem;  
			 space = min(space, sysctl_tcp_rmem[2]); /*不能超过允许的最大接收缓冲区大小*/  
  
			 if (space > sk->sk_rcvbuf) {  
				 sk->sk_rcvbuf = space; /* 调整接收缓冲区的大小*/  
				 /* Make the window clamp follow along. */  
				 tp->window_clamp = new_clamp; /*调整接收窗口的上限*/  
			 }  
		}  
	}  
  
new_measure:  
	 /*此序号之前的数据已复制到用户空间，下次复制将从这里开始*/  
	tp->rcvq_space.seq = tp->copied_seq;  
	tp->rcvq_space.time = tcp_time_stamp; /*记录这次调整的时间*/  
}  
  
  
/* return minimum truesize of the skb containing X bytes of data */  
#define SKB_TRUESIZE(X) ((X) +              \  
	SKB_DATA_ALIGN(sizeof(struct sk_buff)) +        \  
	SKB_DATA_ALIGN(sizeof(struct skb_shared_info)))  
  
  
static inline int tcp_win_from_space(int space)  
{  
	return sysctl_tcp_adv_win_scale <= 0 ?  
			  (space >> (-sysctl_tcp_adv_win_scale)) :  
			   space - (space >> sysctl_tcp_adv_win_scale);  
}

tp->rcvq_space.space表示当前接收缓存的大小（只包括应用层数据，单位为字节）。
sk->sk_rcvbuf表示当前接收缓存的大小（包括应用层数据、TCP协议头、sk_buff和skb_shared_info结构，tcp_adv_win_scale微调，单位为字节）。

系统参数

(1) tcp_moderate_rcvbuf

是否自动调节TCP接收缓冲区的大小，默认值为1。

(2) tcp_adv_win_scale

在tcp_moderate_rcvbuf启用的情况下，用来对计算接收缓冲区和接收窗口的参数进行微调，默认值为2。
This means that the application buffer is ¼th of the total buffer space specified in the tcp_rmem variable.

(3) tcp_rmem

包括三个参数：min default max。
tcp_rmem[1] — default ：接收缓冲区长度的初始值，用来初始化sock的sk_rcvbuf，默认为87380字节。
tcp_rmem[2] — max：接收缓冲区长度的最大值，用来调整sock的sk_rcvbuf，默认为4194304，一般是2000多个数据包。

小结

接收端的接收窗口上限和接收缓冲区大小，是接收方应用程序在上个RTT内接收并复制到用户空间的数据量的2倍，并且接收窗口上限和接收缓冲区大小是递增的。

（1）为什么是2倍呢？

In order to keep pace with the growth of the sender’s congestion window during slow-start, the receiver should use the same doubling factor. Thus the receiver should advertise a window that is twice the size of the last measured window size.

这样就能保证接收窗口上限的增长速度不小于拥塞窗口的增长速度，避免接收窗口成为传输瓶颈。

（2）收到乱序包时有什么影响？

Packets that are received out of order may have lowered the goodput during this measurement, but will increase the goodput of the following measurement which, if larger, will supercede this measurement.

乱序包会使本次的吞吐量测量值偏小，使下次的吞吐量测量值偏大。

Reference

[1] Mike Fisk, Wu-chun Feng, “Dynamic Right-Sizing in TCP”.