本文分析linux-4.14.69代码的bic拥塞算法
首先回顾下基础的慢启动和拥塞避免函数,慢启动阶段(tcp_slow_start)更新窗口的速度是加acked,acked就是这个ack包对应确认的包个数;拥塞避免阶段(tcp_cong_avoid_ai)更新窗口的速度是加acked/w, w一般情况下就是当前的窗口大小.(在bic中,相当与一个权重,bic会根据当前窗口到最大窗口的差距动态设置这个w.)
u32 tcp_slow_start(struct tcp_sock *tp, u32 acked)
{
u32 cwnd = min(tp->snd_cwnd + acked, tp->snd_ssthresh);
acked -= cwnd - tp->snd_cwnd;
tp->snd_cwnd = min(cwnd, tp->snd_cwnd_clamp);
return acked;
}
EXPORT_SYMBOL_GPL(tcp_slow_start);
void tcp_cong_avoid_ai(struct tcp_sock *tp, u32 w, u32 acked)
{
/* If credits accumulated at a higher w, apply them gently now. */
if (tp->snd_cwnd_cnt >= w) {
tp->snd_cwnd_cnt = 0;
tp->snd_cwnd++;
}
tp->snd_cwnd_cnt += acked;
if (tp->snd_cwnd_cnt >= w) {
u32 delta = tp->snd_cwnd_cnt / w;
tp->snd_cwnd_cnt -= delta * w;
tp->snd_cwnd += delta;
}
tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_cwnd_clamp);
}
EXPORT_SYMBOL_GPL(tcp_cong_avoid_ai);
再看看bic的拥塞避免函数.
static void bictcp_cong_avoid(struct sock *sk, u32 ack, u32 acked)
{
struct tcp_sock *tp = tcp_sk(sk);
struct bictcp *ca = inet_csk_ca(sk);
if (!tcp_is_cwnd_limited(sk))
return;
if (tcp_in_slow_start(tp))
tcp_slow_start(tp, acked);
else {
bictcp_update(ca, tp->snd_cwnd);
tcp_cong_avoid_ai(tp, ca->cnt, 1);
}
}
慢启动就是调用默认的慢启动函数, 拥塞避免阶段则多了一个更新窗口的处理,并将tcp_cong_avoid_ai的第三个参数acked设置固定值1.更新函数就是在更新ca->cnt, 从而控制了拥塞避免阶段增长速度。其中涉及到很多变量,下面一一列出
bictcp结构体
tcp_jiffies32
low_window
BICTCP_B
static int max_increment
smooth_part
/* BIC TCP Parameters */
struct bictcp {
u32 cnt; /* increase cwnd by 1 after ACKs */
u32 last_max_cwnd; /* last maximum snd_cwnd */
u32 last_cwnd; /* the last snd_cwnd */
u32 last_time; /* time when updated last_cwnd */
u32 epoch_start; /* beginning of an epoch */
#define ACK_RATIO_SHIFT 4
u32 delayed_ack; /* estimate the ratio of Packets/ACKs << 4 */
};
/* TCP uses 32bit jiffies to save some space.
* Note that this is different from tcp_time_stamp, which
* historically has been the same until linux-4.13.
*/
#define tcp_jiffies32 ((u32)jiffies)static int low_window = 14;
#define BICTCP_B 4 /*
* In binary search,
* go to point (max+min)/N
*/
static int max_increment = 16; /* Limit on increment allowed during binary search */
static int smooth_part = 20; /* log(B/(B*Smin))/log(B/(B-1))+B, # of RTT from Wmax-B to Wmax */
再来看bictcp_update的流程.
1.判断是否要更新(如果窗口没变,而且与上次更新的时间小于HZ/32,即31.25ms就不用更新,直接跳出).
if (ca->last_cwnd == cwnd &&
(s32)(tcp_jiffies32 - ca->last_time) <= HZ / 32)
return;
2.更新last_cwnd和last_time,(如果是初次,epoch_start为0,需要设置下epoch_start为当前时间tcp_jiffies32),如果当前发送窗口小于low_window,ca->cnt设置为当前发送窗口,即普通拥塞避免模式.
ca->last_cwnd = cwnd;
ca->last_time = tcp_jiffies32;
if (ca->epoch_start == 0) /* record the beginning of an epoch */
ca->epoch_start = tcp_jiffies32;
3.二分查找.
3.1 cwnd < last_max_cwnd
3.1.1当max_cwnd - cwnd大于BICTCP_B*max_increment, cnt =cwnd/max_increment。离最大窗口很远的时候,快速增长,一个rtt增max_increment个窗口,经过BICTCP_B个rtt达到最大窗口
3.1.2当max_cwnd - cwnd小于BICTCP_B, cnt = cwnd *smooth_part / BICTCP_B,一个rtt增(BICTCP_B/smooth_part )个窗口. 接近最大窗口时候非常缓慢的增长,大约经过smooth_part个rtt才增大到最大窗口。
3.1.3当max_cwnd - cwnd在[BICTCP_B, BICTCP_B*max_increment], cnt = cwnd / (max_cwnd - cwnd),一个rtt增max_cwnd-cwnd个窗口,也就是一个rtt达到最大窗口.
3.2 cwnd >= last_max_cwnd
3.2.1当cwnd - max_cwnd小于BICTCP_B, 超过的不多,cnt = cwnd *smooth_part / BICTCP_B,缓慢长,同3.1.2,一个rtt增(BICTCP_B/smooth_part )个窗口
3.2.2当cwnd - max_cwnd在[BICTCP_B, max_increment*(BICTCP_B-1)], cnt = (cwnd * (BICTCP_B-1)) / (cwnd-last_max_cwnd); 一个rtt增大(cwnd-last_max_cwnd)/(BICTCP_B-1)个窗口.速度会越来越快,因为cwnd不断增大.根据if条件范围可以算出1个rtt增加的窗口范围在[BICTCP_B/(BICTCP_B-1), max_increment]
3.2.3当cwnd - max_cwnd大于max_increment,cnt = cwnd / max_increment; ,同3.1.1,一个rtt增max_increment个窗口
/* binary increase */
if (cwnd < ca->last_max_cwnd) {
__u32 dist = (ca->last_max_cwnd - cwnd)
/ BICTCP_B;
if (dist > max_increment)
/* linear increase */
ca->cnt = cwnd / max_increment;
else if (dist <= 1U)
/* binary search increase */
ca->cnt = (cwnd * smooth_part) / BICTCP_B;
else
/* binary search increase */
ca->cnt = cwnd / dist;
} else {
/* slow start AMD linear increase */
if (cwnd < ca->last_max_cwnd + BICTCP_B)
/* slow start */
ca->cnt = (cwnd * smooth_part) / BICTCP_B;
else if (cwnd < ca->last_max_cwnd + max_increment*(BICTCP_B-1))
/* slow start */
ca->cnt = (cwnd * (BICTCP_B-1))
/ (cwnd - ca->last_max_cwnd);
else
/* linear increase */
ca->cnt = cwnd / max_increment;
}
4. 当last_max_cwnd == 0。即在慢启动开始或出现拥塞的时候,控制cnt不超过20.
/* if in slow start or link utilization is very low */
if (ca->last_max_cwnd == 0) {
if (ca->cnt > 20) /* increase cwnd 5% per RTT */
ca->cnt = 20;
}
5.对延迟确认的处理。延迟确认的时候,一个ack不止是确认一个报文,作者的意思是,根据延迟ack的比例1<<ACK_RATIO_SHIFT/delayed_ack,增大窗口的时候需要扩大delayed_ack/2^ACK_RATIO_SHIFT倍。默认延迟比例为50%(ACK_RATIO_SHIFT为常数4, delayed_ack默认值为2<<ACK_RATIO_SHIFT)
ca->cnt = (ca->cnt << ACK_RATIO_SHIFT) / ca->delayed_ack;
if (ca->cnt == 0) /* cannot be zero */
ca->cnt = 1;
ACK_RATIO_SHIFT为常数4, delayed_ack的设置涉及到bictcp_reset函数和bictcp_acked函数.
bictcp_reset对bictcp结构体做初始化,可以看到初始化的时候delayed_ack设置为32,即默认延迟比例为50%。 bictcp_reset在两种情况下被调用: 初始化时(bictcp_init )、进入拥塞处理时(bictcp_state 状态为TCP_CA_Loss)
static inline void bictcp_reset(struct bictcp *ca)
{
ca->cnt = 0;
ca->last_max_cwnd = 0;
ca->last_cwnd = 0;
ca->last_time = 0;
ca->epoch_start = 0;
ca->delayed_ack = 2 << ACK_RATIO_SHIFT;
}
bictcp_acked计算延迟ack的比例delayed_ack ,公式没有看懂,sample是什么东西?
/* Track delayed acknowledgment ratio using sliding window
* ratio = (15*ratio + sample) / 16
*/
static void bictcp_acked(struct sock *sk, const struct ack_sample *sample)
{
const struct inet_connection_sock *icsk = inet_csk(sk);
if (icsk->icsk_ca_state == TCP_CA_Open) {
struct bictcp *ca = inet_csk_ca(sk);
ca->delayed_ack += sample->pkts_acked -
(ca->delayed_ack >> ACK_RATIO_SHIFT);
}
}
最后要关注的就是bic算法对ssthresh的计算了。
每收到一个ack,就会调用tcp_ack(这个函数有点复杂后面慢慢看)。tcp_ack会调用.pkts_acked和.cong_avoid,pkts_acked对应bictcp_acked, cong_avoid对应bictcp_cong_avoid。 tcp_ack中如果检测到丢包,则进入拥塞阶段,调用.ssthresh,对应bic的bictcp_recalc_ssthresh函数,tcp_ack完成重传后,退回到拥塞阶段,调用.undo_cwnd函数,即tcp_reno_undo_cwnd。
bictcp_recalc_ssthresh用于拥塞后计算慢启动阈值ssthresh。 里面有几个参数要注意下
#define BICTCP_BETA_SCALE 1024 /* Scale factor beta calculation
static int beta = 819; /* = 819/1024 (BICTCP_BETA_SCALE) */
1.设置最大窗口
当 snd_cwnd < last_max_cwnd && fast_convergence==1. last_max_cwnd=snd_cwnd*(819+1024)/(2*1024)=0.9*snd_cwnd
当 snd_cwnd >= last_max_cwnd . last_max_cwnd=snd_cwnd
2.设置发送窗口
当snd_cwnd <= low_window。snd_cwnd = max(snd_cwnd/2, 2)
当snd_cwnd > low_window. snd_cwnd= max(snd_cwnd*beta/BICTCP_BETA_SCALE, 2)=max(snd_cwnd*0.8, 2)
/*
* behave like Reno until low_window is reached,
* then increase congestion window slowly
*/
static u32 bictcp_recalc_ssthresh(struct sock *sk)
{
const struct tcp_sock *tp = tcp_sk(sk);
struct bictcp *ca = inet_csk_ca(sk);
ca->epoch_start = 0; /* end of epoch */
/* Wmax and fast convergence */
if (tp->snd_cwnd < ca->last_max_cwnd && fast_convergence)
ca->last_max_cwnd = (tp->snd_cwnd * (BICTCP_BETA_SCALE + beta))
/ (2 * BICTCP_BETA_SCALE);
else
ca->last_max_cwnd = tp->snd_cwnd;
if (tp->snd_cwnd <= low_window)
return max(tp->snd_cwnd >> 1U, 2U);
else
return max((tp->snd_cwnd * beta) / BICTCP_BETA_SCALE, 2U);
}
要深入了解需要了解收到ack后的相关处理细节