安卓cpu优化 tcp拥塞算法cubic和reno怎么选择
Posted
tags:
篇首语：本文由小常识网(cha138.com)小编为大家整理，主要介绍了安卓cpu优化 tcp拥塞算法cubic和reno怎么选择相关的知识，希望对你有一定的参考价值。
上述具体的论文可以参考：CUBIC: A New TCP-Friendly High-Speed TCP Variant
1. tcp cubic数学模型
CUBIC在设计上简化了BIC-TCP的窗口调整算法，在BIC-TCP的窗口调整中会出现一个凹和凸(这里的凹和凸指的是数学意义上的凹和凸，凹函数/凸函数)的增长曲线，CUBIC使用了一个三次函数(即一个立方函数)，在三次函数曲线中同样存在一个凹和凸的部分，该曲线形状和BIC-TCP的曲线图十分相似，于是该部分取代BIC-TCP的增长曲线。另外，CUBIC中最关键的点在于它的窗口增长函数仅仅取决于连续的两次拥塞事件的时间间隔值，从而窗口增长完全独立于网络的时延RTT，之前讲述过的HSTCP存在严重的RTT不公平性，而CUBIC的RTT独立性质使得CUBIC能够在多条共享瓶颈链路的TCP连接之间保持良好的RRTT公平性。
来看下具体细节：当某次拥塞事件发生时，Wmax设置为此时发生拥塞时的窗口值，然后把窗口进行乘法减小，乘法减小因子设为β，当从快速恢复阶段退出然后进入到拥塞避免阶段，此时CUBIC的窗口增长开始按照“凹”式增长曲线进行增长，该过程一直持续直到窗口再次增长到Wmax，紧接着，该函数转入“凸”式增长阶段。该方式的增长可以使得窗口一直维持在Wmax附近，从而可以达到网络带宽的高利用率和协议本身的稳定性。
窗口的增长函数如下：
W(t) = C * (t-K)3 + Wmax, 其中C和β为常量。
t为当前时间距上一次窗口减小的时间差，而K就代表该函数从W增长到Wmax的时间周期，。
当收到ACK后，CUBIC计算利用该算法计算下一个RTT内的窗口增长速度，即计算W(t+RTT)，该值将作为cwnd的目标值，根据cwnd的大小，CUBIC将进入三种不同模式，如果cwnd会小于在标准TCP下经过上次拥塞之后的时刻t窗口将会达到的值(该值是通过标准TCP的窗口增长函数计算出来的)，那么CUBIC就处于标准TCP模式，如果小于Wmax，那么位于凹阶段的，如果大于Wmax，那么处于凸阶段。
tcp cubic 内核源代码调用逻辑
CUBIC整体架构调用的逻辑如下：
1. 连接每收到一个ack，则调用tcp_ack
2. tcp_ack会调用bictcp_acked，用来更新cnt和delayed_ack（用来消除delay包的影响）
3. tcp_ack会调用bictcp_cong_avoid，这是分两种情况：
（1）snd_cwnd小于慢启动阈值，处于慢启动阶段，则调用tcp_slow_start
（2）snd_cwnd大于慢启动阈值，处于拥塞避免阶段，则调用bictcp_update来更新bictcp，再调用tcp_cong_avoid_ai
4. tcp_ack中如果检测到丢包，进入拥塞处理阶段，则调用bictcp_recalc_ssthresh来更新慢启动阈值
5. tcp_ack中完成丢包重传后，退出拥塞处理阶段，则调用bictcp_undo_cwnd来更新
快速重传：tcp_ack中的丢包检测，即检测到连续3个重复ACK。
快速恢复：bictcp_undo_cwnd，直接把snd_cwnd更新为max(snd_cwnd，last_max_cwnd)，和掉包前相差不大。
我对TCP CDG拥塞控制算法的改进和优化
其实这不是我的优化，我是借用了BBR之力。借了什么力呢？这是我一再强调的，BBR最大的共享不是为Linux贡献了一个TCP拥塞控制算法(它同时在也BSD上被实现...)，而是它重构了Linux TCP的实现！借助BBR对Linux TCP实现的重构，很多之前做不到的事情，现在可以做到了。
简而言之，BBR算法对Linux TCP实现的重构中，将以下三件事完全分离：
1.重传哪些包；
2.传输多少包；
3.实际传输。
拥塞控制算法侧重解决上述第2点问题。
-----------------------------------
CDG必须要拥塞窗口的背后默默维护一个”自己的窗口“，称为shadow_wnd，该窗口只受”实际拥塞情况“的影响，而不受”Linux TCP拥塞状态机“的影响。所以说，即便在丢包重传的Recovery时期，也必须动态维护这个shadow_wnd，使其按照Reno方式增长(或者按照CUBIC方式，随便什么方式都可以)。
然则这在BBR之前的Linux 4.8版本之前的内核中是无法做到的。因为tcp_congestion_ops机构体中没有一个回调函数是在Recovery阶段可以被调用的到的，而你所能控制的拥塞算法只能通过tcp_congestion_ops结构体的回调来实现。
BBR将以下的逻辑引入到了Linux：
static void tcp_cong_control(struct sock *sk, u32 ack, u32 acked_sacked,
int flag, const struct rate_sample *rs)
{
const struct inet_connection_sock *icsk = inet_csk(sk);
if (icsk->icsk_ca_ops->cong_control) {
icsk->icsk_ca_ops->cong_control(sk, rs);
return;
}
if (tcp_in_cwnd_reduction(sk)) {
/* Reduce cwnd if state mandates */
tcp_cwnd_reduction(sk, acked_sacked, flag);
} else if (tcp_may_raise_cwnd(sk, flag)) {
/* Advance cwnd if state allows */
tcp_cong_avoid(sk, ack, acked_sacked);
}
tcp_update_pacing_rate(sk);
}只要实现了cong_control回调，那就就不会再调用标准的PRR算法和拥塞避免tcp_cong_avoid函数，无论在任何阶段，均调用cong_control回调。因此，我的方法是，在Recovery或者Loss状态调用cong_control回调即可！在该回调中维护CDG的shadow窗口。
这谈何容易！BBR引入的逻辑非常粗糙，只要实现了cong_control，该函数就无条件返回。事实上正确的做法是cong_control回调有个返回值，当满足一定条件时返回，否则继续下面的逻辑。但是BBR并没有引入这些。
-----------------------------------
但是，我将其引入了。
请看，我将tcp_input.c中的tcp_cong_control改成了下面的样子：
static void tcp_cong_control(struct sock *sk, u32 ack, u32 prior_in_flight, u32 acked_sacked,
int flag, const struct rate_sample *rs)
{
const struct inet_connection_sock *icsk = inet_csk(sk);
#ifdef BBR
if (icsk->icsk_ca_ops->cong_control) {
icsk->icsk_ca_ops->cong_control(sk, rs);
#ifdef CDG
// 以下是我添加的判断，新增了rs的flag字段，一旦置位就继续而不返回。
if (!(rs->flag & CDG_CONT))
return;
#endif
}
#endif
if (tcp_in_cwnd_reduction(sk)) {
/* Reduce cwnd if state mandates */
tcp_cwnd_reduction(sk, acked_sacked, 1);
} else if (tcp_may_raise_cwnd(sk, flag)) {
/* Advance cwnd if state allows */
tcp_cong_avoid(sk, ack, prior_in_flight);
}
tcp_update_pacing_rate(sk);
}我添加了个判断。其实我的目的很简单，就是在Recovery状态下也能调用到CDG的逻辑，就这么简单个逻辑在不懂的人眼里显得如此高大上，在懂的人眼里显得如此傻逼...不管怎样，我做了。-----------------------------------
以下的代码只是我对标准Linux 4.3内核CDG算法的differ，想理解代码细节的，请先阅读标准CDG代码，我虽然是个传说中有求必应的人，但那只是传说...请注意，我的目标内核是3.10内核，在我移植CDG之前，我已经移植了BBR，所以说，你最好以4.9内核为准，然而这样一来，又会对3.10内核的一些接口表示费解..这里不就不多解释了，我要说的是，想彻底逃离学院派，就必须把所有这些代码都搞清楚！不然的话，首先，你根本什么都看不懂，其次，即便你有想法，你也做不来。完整的代码我会附在本文最后。
以下是patch中几个重要函数的说明：
1.CDG的cong_control回调函数cdg_main：
static void cdg_main(struct sock *sk, struct rate_sample *rs)
{
struct inet_connection_sock *icsk = inet_csk(sk);
struct tcp_sock *tp = tcp_sk(sk);
struct cdg *ca = inet_csk_ca(sk);
if (!shadow_grow) {
rs->flag |= CDG_CONT;
return;
}
if (icsk->icsk_ca_state != TCP_CA_Open) {
// 在重传阶段，依然要采集rtt，因为链路不问包类型，重传包也会影响网络可用容量。
if (rs->rtt_us) {
// 感谢BBR增加了rs结构体，从中可以取rtt_us
ca->rtt.min = min_not_zero(ca->rtt.min, (s32)rs->rtt_us);
ca->rtt.max = max(ca->rtt.max, (s32)rs->rtt_us);
}
if (ca->state == CDG_NONFULL && use_tolerance) {
if (!shadow_fast && (ca->ack_sack_cnt < 0 || ca->ack_sack_cnt == 0) && ca->rtt.v64) {
s32 grad = 0;
if (ca->rtt_prev.v64)
grad = tcp_cdg_grad(ca);
ca->rtt_prev = ca->rtt;
ca->ack_sack_cnt = tcp_packets_in_flight(tp);
ca->rtt.v64 = 0;
}
ca->ack_sack_cnt -= rs->acked_sacked;
if (ca->state == CDG_NONFULL || shadow_fast) {
// 如果链路未完全拥塞，那么shadow窗口便默默地帮助实际窗口占据空间，等到快速恢复结束，便可以由实际窗口可用。
tcp_cong_avoid_ai_shadow(sk, ca->shadow_wnd, rs->acked_sacked);
tp->snd_cwnd = ca->shadow_wnd;
}
rs->flag |= CDG_CONT;
}
} else {
// 为了让执行流继续，增加CDG_CONT标志。
rs->flag |= CDG_CONT;
}
}2.状态设置回调函数cdg_state：
static void cdg_state(struct sock *sk, u8 new_state)
{
struct cdg *ca = inet_csk_ca(sk);
struct tcp_sock *tp = tcp_sk(sk);
if (!recovery_restore)
return;
if (new_state == TCP_CA_Open)
// 进入Open状态时，直接接管shadow窗口，这里可能会有突发问题。
tp->snd_cwnd = max(max(tp->snd_cwnd, ca->shadow_wnd), 2U);
if (new_state == TCP_CA_Loss) {
// 进入Loss状态，判断是否是噪声丢包
if (ca->state == CDG_NONFULL && use_tolerance) {
// 如果是噪声丢包，那么便恢复窗口。
tp->snd_cwnd = ca->shadow_wnd;
printk("#### cwnd:%u \n", tp->snd_cwnd);
if (loss_push)
// 如果是噪声丢包，那么在窗口内继续发送数据。
tcp_push_pending_frames(sk);
}
// 如果是拥塞丢包，那么执行原有流程。
}
}3.UNDO函数tcp_cdg_undo_cwnd：
static u32 tcp_cdg_undo_cwnd(struct sock *sk)
{
struct cdg *ca = inet_csk_ca(sk);
struct tcp_sock *tp = tcp_sk(sk);
// undo到shadow窗口
return max3(2U, ca->shadow_wnd, max(tp->snd_cwnd, ca->undo_cwnd));
}4.RTT梯度计算函数tcp_cdg_grad：
static s32 tcp_cdg_grad(struct cdg *ca)
{
// rtt在pkts_acked回调和cong_control中被采样值更新
s32 gmin = ca->rtt.min - ca->rtt_prev.min;
s32 gmax = ca->rtt.max - ca->rtt_prev.max;
s32 grad;
if (ca->gradients) {
ca->gsum.min += gmin - ca->gradients[ca->tail].min;
ca->gsum.max += gmax - ca->gradients[ca->tail].max;
ca->gradients[ca->tail].min = gmin;
ca->gradients[ca->tail].max = gmax;
ca->tail = (ca->tail + 1) & (window - 1);
gmin = ca->gsum.min;
gmax = ca->gsum.max;
}
......
/* Backoff was effectual: */
if (gmin <= -32 || gmax <= -32)
ca->backoff_cnt = 0;
if (use_tolerance) {
/* Reduce small variations to zero: */
gmin = DIV_ROUND_CLOSEST(gmin, 64);
gmax = DIV_ROUND_CLOSEST(gmax, 64);
// 注意看上一篇文章CDG模型图示的边沿触发条件。
if (gmin > 0 && gmax <= 0)
ca->state = CDG_FULL;
else if ((gmin > 0 && gmax > 0) || gmax < 0)
ca->state = CDG_NONFULL;
}
return grad;
}我首先盲测了一下原生的CDG，Oh NO！太垃圾，比CUBIC好，高丢包率下竟然与Westwood相当，在所有这一切中，BBR始终是另类，遥不可及，在我看了Paper之后，迅速自己实现了一版，感谢BBR对Linux TCP的重构！我承认我自己只懂Reno，BIC，CUBIC，Vegas，BBR这几种算法，其它HTCP，Westwood这些我并没有详细分析过，但是无论我怎么测，我发现我的CDG(应该是我改过的CDG)，一直跟BBR接近。
CDG是什么？CDG实际上就是传统基于丢包的算法加上了一个抗噪声机制，本来基于丢包的算法就是以不断填充缓存为手段，直到填满缓存发生丢包进行减窗，然而有的时候并非拥塞的原因也会发生丢包，此时按照算法来看依然会减窗，这就大大降低了带宽的利用率。加上了这个CDG的RTT梯度抗噪声机制后，网络带宽的利用率大大提高了。然而可能会加重拥塞，所以CDG内置了backoff算法，这里就不赘述了。
-----------------------------------
tcp_cdg.c代码：
#include <linux/kernel.h>
#include <linux/random.h>
#include <linux/module.h>
#include <net/tcp.h>
#define HYSTART_ACK_TRAIN 1
#define HYSTART_DELAY 2
static int window __read_mostly = 8;
static unsigned int backoff_beta __read_mostly = 0.7071 * 1024; /* sqrt 0.5 */
static unsigned int backoff_factor __read_mostly = 42;
static unsigned int hystart_detect __read_mostly = 3;
static unsigned int use_ineff __read_mostly = 5;
static unsigned int use_shadow __read_mostly = 1;
static unsigned int backoff __read_mostly = 0;
static unsigned int use_tolerance __read_mostly = 1;
static unsigned int shadow_fast __read_mostly = 1;
static unsigned int shadow_grow __read_mostly = 1;
static unsigned int recovery_restore __read_mostly = 1;
static unsigned int loss_push __read_mostly = 1;
module_param(window, int, 0444);
MODULE_PARM_DESC(window, "gradient window size (power of two <= 256)");
module_param(backoff_beta, uint, 0644);
MODULE_PARM_DESC(backoff_beta, "backoff beta (0-1024)");
module_param(backoff_factor, uint, 0644);
MODULE_PARM_DESC(backoff_factor, "backoff probability scale factor");
module_param(hystart_detect, uint, 0644);
MODULE_PARM_DESC(hystart_detect, "use Hybrid Slow start "
"(0: disabled, 1: ACK train, 2: delay threshold, 3: both)");
module_param(use_ineff, uint, 0644);
MODULE_PARM_DESC(use_ineff, "use ineffectual backoff detection (threshold)");
module_param(use_shadow, uint, 0644);
MODULE_PARM_DESC(use_shadow, "use shadow window heuristic");
module_param(backoff, uint, 0644);
MODULE_PARM_DESC(backoff, "back");
module_param(use_tolerance, uint, 0644);
MODULE_PARM_DESC(use_tolerance, "use loss tolerance heuristic");
module_param(shadow_fast, uint, 0644);
MODULE_PARM_DESC(shadow_fast, "back");
module_param(shadow_grow, uint, 0644);
MODULE_PARM_DESC(shadow_grow, "back");
module_param(recovery_restore, uint, 0644);
MODULE_PARM_DESC(recovery_restore, "back");
module_param(loss_push, uint, 0644);
MODULE_PARM_DESC(loss_push, "back");
struct cdg_minmax {
union {
struct {
s32 min;
s32 max;
};
u64 v64;
};
};
enum cdg_state {
CDG_UNKNOWN = 0,
CDG_NONFULL = 1,
CDG_FULL = 2,
CDG_BACKOFF = 3,
};
struct cdg {
struct cdg_minmax rtt;
struct cdg_minmax rtt_prev;
struct cdg_minmax *gradients;
struct cdg_minmax gsum;
bool gfilled;
u8 tail;
u8 state;
u8 delack;
u32 rtt_seq;
u32 undo_cwnd;
u32 shadow_wnd;
u32 snd_cwnd_cnt;
u16 backoff_cnt;
u16 sample_cnt;
s32 delay_min;
s32 ack_sack_cnt;
u32 last_ack;
u32 round_start;
};
/**
* nexp_u32 - negative base-e exponential
* @ux: x in units of micro
*
* Returns exp(ux * -1e-6) * U32_MAX.
*/
static u32 __pure nexp_u32(u32 ux)
{
static const u16 v[] = {
/* exp(-x)*65536-1 for x = 0, 0.000256, 0.000512, ... */
65535,
65518, 65501, 65468, 65401, 65267, 65001, 64470, 63422,
61378, 57484, 50423, 38795, 22965, 8047, 987, 14,
};
u32 msb = ux >> 8;
u32 res;
int i;
/* Cut off when ux >= 2^24 (actual result is <= 222/U32_MAX). */
if (msb > U16_MAX)
return 0;
/* Scale first eight bits linearly: */
res = U32_MAX - (ux & 0xff) * (U32_MAX / 1000000);
/* Obtain e^(x + y + ...) by computing e^x * e^y * ...: */
for (i = 1; msb; i++, msb >>= 1) {
u32 y = v[i & -(msb & 1)] + U32_C(1);
res = ((u64)res * y) >> 16;
}
return res;
}
/* Based on the HyStart algorithm (by Ha et al.) that is implemented in
* tcp_cubic. Differences/experimental changes:
* o Using Hayes‘ delayed ACK filter.
* o Using a usec clock for the ACK train.
* o Reset ACK train when application limited.
* o Invoked at any cwnd (i.e. also when cwnd < 16).
* o Invoked only when cwnd < ssthresh (i.e. not when cwnd == ssthresh).
*/
static void tcp_cdg_hystart_update(struct sock *sk)
{
struct cdg *ca = inet_csk_ca(sk);
struct tcp_sock *tp = tcp_sk(sk);
ca->delay_min = min_not_zero(ca->delay_min, ca->rtt.min);
if (ca->delay_min == 0)
return;
if (hystart_detect & HYSTART_ACK_TRAIN) {
u32 now_us = div_u64(local_clock(), NSEC_PER_USEC);
if (ca->last_ack == 0 || !tcp_is_cwnd_limited(sk, tcp_packets_in_flight(tp))) {
ca->last_ack = now_us;
ca->round_start = now_us;
} else if (before(now_us, ca->last_ack + 3000)) {
u32 base_owd = max(ca->delay_min / 2U, 125U);
ca->last_ack = now_us;
if (after(now_us, ca->round_start + base_owd)) {
tp->snd_ssthresh = tp->snd_cwnd;
return;
}
}
}
if (hystart_detect & HYSTART_DELAY) {
if (ca->sample_cnt < 8) {
ca->sample_cnt++;
} else {
s32 thresh = max(ca->delay_min + ca->delay_min / 8U,
125U);
if (ca->rtt.min > thresh) {
tp->snd_ssthresh = tp->snd_cwnd;
}
}
}
}
static s32 tcp_cdg_grad(struct cdg *ca)
{
s32 gmin = ca->rtt.min - ca->rtt_prev.min;
s32 gmax = ca->rtt.max - ca->rtt_prev.max;
s32 grad;
if (ca->gradients) {
ca->gsum.min += gmin - ca->gradients[ca->tail].min;
ca->gsum.max += gmax - ca->gradients[ca->tail].max;
ca->gradients[ca->tail].min = gmin;
ca->gradients[ca->tail].max = gmax;
ca->tail = (ca->tail + 1) & (window - 1);
gmin = ca->gsum.min;
gmax = ca->gsum.max;
}
/* We keep sums to ignore gradients during cwnd reductions;
* the paper‘s smoothed gradients otherwise simplify to:
* (rtt_latest - rtt_oldest) / window.
*
* We also drop division by window here.
*/
grad = gmin > 0 ? gmin : gmax;
/* Extrapolate missing values in gradient window: */
if (!ca->gfilled) {
if (!ca->gradients && window > 1)
grad *= window; /* Memory allocation failed. */
else if (ca->tail == 0)
ca->gfilled = true;
else
grad = (grad * window) / (int)ca->tail;
}
/* Backoff was effectual: */
if (gmin <= -32 || gmax <= -32)
ca->backoff_cnt = 0;
if (use_tolerance) {
/* Reduce small variations to zero: */
gmin = DIV_ROUND_CLOSEST(gmin, 64);
gmax = DIV_ROUND_CLOSEST(gmax, 64);
if (gmin > 0 && gmax <= 0)
ca->state = CDG_FULL;
else if ((gmin > 0 && gmax > 0) || gmax < 0)
ca->state = CDG_NONFULL;
}
return grad;
}
void tcp_enter_cwr_1(struct sock *sk)
{
struct tcp_sock *tp = tcp_sk(sk);
tp->prior_ssthresh = 0;
if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
tp->undo_marker = 0;
tp->high_seq = tp->snd_nxt;
tp->tlp_high_seq = 0;
tp->snd_cwnd_cnt = 0;
tp->prior_cwnd = tp->snd_cwnd;
tp->prr_delivered = 0;
tp->prr_out = 0;
tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
if (tp->ecn_flags & TCP_ECN_OK)
tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
tcp_set_ca_state(sk, TCP_CA_CWR);
}
}
static bool tcp_cdg_backoff(struct sock *sk, u32 grad)
{
struct cdg *ca = inet_csk_ca(sk);
struct tcp_sock *tp = tcp_sk(sk);
if (prandom_u32() <= nexp_u32(grad * backoff_factor))
return false;
if (use_ineff) {
ca->backoff_cnt++;
if (ca->backoff_cnt > use_ineff)
return false;
}
ca->shadow_wnd = max(ca->shadow_wnd, tp->snd_cwnd);
ca->state = CDG_BACKOFF;
tcp_enter_cwr_1(sk);
return true;
}
void tcp_cong_avoid_ai_shadow(struct sock *sk, u32 w, u32 acked)
{
struct tcp_sock *tp = tcp_sk(sk);
struct cdg *ca = inet_csk_ca(sk);
if (ca->snd_cwnd_cnt >= w) {
ca->snd_cwnd_cnt = 0;
ca->shadow_wnd ++;
}
ca->snd_cwnd_cnt += acked;
if (ca->snd_cwnd_cnt >= w) {
u32 delta = ca->snd_cwnd_cnt / w;
ca->snd_cwnd_cnt -= delta * w;
ca->shadow_wnd += delta;
}
ca->shadow_wnd = min(ca->shadow_wnd, tp->snd_cwnd_clamp);
}
/* Not called in CWR or Recovery state. */
static void tcp_cdg_cong_avoid(struct sock *sk, u32 ack, u32 acked)
{
struct cdg *ca = inet_csk_ca(sk);
struct tcp_sock *tp = tcp_sk(sk);
u32 prior_snd_cwnd;
u32 incr;
if (tp->snd_cwnd <= tp->snd_ssthresh && hystart_detect)
tcp_cdg_hystart_update(sk);
if (after(ack, ca->rtt_seq) && ca->rtt.v64) {
s32 grad = 0;
if (ca->rtt_prev.v64)
grad = tcp_cdg_grad(ca);
ca->rtt_seq = tp->snd_nxt;
ca->rtt_prev = ca->rtt;
ca->rtt.v64 = 0;
ca->last_ack = 0;
ca->sample_cnt = 0;
if (backoff && grad > 0 && tcp_cdg_backoff(sk, grad))
return;
}
if (!tcp_is_cwnd_limited(sk, tcp_packets_in_flight(tp))) {
ca->shadow_wnd = min(ca->shadow_wnd, tp->snd_cwnd);
return;
}
prior_snd_cwnd = tp->snd_cwnd;
tcp_reno_cong_avoid(sk, ack, acked);
incr = tp->snd_cwnd - prior_snd_cwnd;
ca->shadow_wnd = max(ca->shadow_wnd, ca->shadow_wnd + incr);
}
static void tcp_cdg_acked(struct sock *sk, u32 num_acked, s32 rtt_us)
{
struct cdg *ca = inet_csk_ca(sk);
struct tcp_sock *tp = tcp_sk(sk);
if (rtt_us <= 0)
return;
/* A heuristic for filtering delayed ACKs, adapted from:
* D.A. Hayes. "Timing enhancements to the FreeBSD kernel to support
* delay and rate based TCP mechanisms." TR 100219A. CAIA, 2010.
*/
if (tp->sacked_out == 0) {
if (num_acked == 1 && ca->delack) {
/* A delayed ACK is only used for the minimum if it is
* provenly lower than an existing non-zero minimum.
*/
ca->rtt.min = min(ca->rtt.min, rtt_us);
ca->delack--;
return;
} else if (num_acked > 1 && ca->delack < 5) {
ca->delack++;
}
}
ca->rtt.min = min_not_zero(ca->rtt.min, rtt_us);
ca->rtt.max = max(ca->rtt.max, rtt_us);
}
static u32 tcp_cdg_ssthresh(struct sock *sk)
{
struct cdg *ca = inet_csk_ca(sk);
struct tcp_sock *tp = tcp_sk(sk);
ca->undo_cwnd = tp->snd_cwnd;
ca->snd_cwnd_cnt = 0;
ca->ack_sack_cnt = tcp_packets_in_flight(tp);
if (ca->state == CDG_BACKOFF)
return max(2U, (tp->snd_cwnd * min(1024U, backoff_beta)) >> 10);
if (ca->state == CDG_NONFULL && use_tolerance)
return tp->snd_cwnd;
ca->shadow_wnd = max(min(ca->shadow_wnd >> 1, tp->snd_cwnd), 2U);
if (use_shadow)
return max3(2U, ca->shadow_wnd, tp->snd_cwnd >> 1);
return max(2U, tp->snd_cwnd >> 1);
}
static u32 tcp_cdg_undo_cwnd(struct sock *sk)
{
struct cdg *ca = inet_csk_ca(sk);
struct tcp_sock *tp = tcp_sk(sk);
return max3(2U, ca->shadow_wnd, max(tp->snd_cwnd, ca->undo_cwnd));
}
static void tcp_cdg_cwnd_event(struct sock *sk, const enum tcp_ca_event ev)
{
struct cdg *ca = inet_csk_ca(sk);
struct tcp_sock *tp = tcp_sk(sk);
struct cdg_minmax *gradients;
switch (ev) {
case CA_EVENT_CWND_RESTART:
gradients = ca->gradients;
if (gradients)
memset(gradients, 0, window * sizeof(gradients[0]));
memset(ca, 0, sizeof(*ca));
ca->gradients = gradients;
ca->rtt_seq = tp->snd_nxt;
ca->shadow_wnd = tp->snd_cwnd;
break;
case CA_EVENT_COMPLETE_CWR:
ca->state = CDG_UNKNOWN;
ca->rtt_seq = tp->snd_nxt;
ca->rtt_prev = ca->rtt;
ca->rtt.v64 = 0;
break;
default:
break;
}
}
static void tcp_cdg_init(struct sock *sk)
{
struct cdg *ca = inet_csk_ca(sk);
struct tcp_sock *tp = tcp_sk(sk);
/* We silently fall back to window = 1 if allocation fails. */
if (window > 1)
ca->gradients = kcalloc(window, sizeof(ca->gradients[0]),
GFP_NOWAIT | __GFP_NOWARN);
ca->rtt_seq = tp->snd_nxt;
ca->shadow_wnd = tp->snd_cwnd;
ca->ack_sack_cnt = 0;
}
static void tcp_cdg_release(struct sock *sk)
{
struct cdg *ca = inet_csk_ca(sk);
kfree(ca->gradients);
}
static void cdg_main(struct sock *sk, struct rate_sample *rs)
{
struct inet_connection_sock *icsk = inet_csk(sk);
struct tcp_sock *tp = tcp_sk(sk);
struct cdg *ca = inet_csk_ca(sk);
if (!shadow_grow) {
rs->flag |= CDG_CONT;
return;
}
if (icsk->icsk_ca_state != TCP_CA_Open) {
if (rs->rtt_us) {
ca->rtt.min = min_not_zero(ca->rtt.min, (s32)rs->rtt_us);
ca->rtt.max = max(ca->rtt.max, (s32)rs->rtt_us);
}
if (ca->state == CDG_NONFULL && use_tolerance) {
if (!shadow_fast && (ca->ack_sack_cnt < 0 || ca->ack_sack_cnt == 0) && ca->rtt.v64) {
s32 grad = 0;
if (ca->rtt_prev.v64)
grad = tcp_cdg_grad(ca);
ca->rtt_prev = ca->rtt;
ca->ack_sack_cnt = tcp_packets_in_flight(tp);
ca->rtt.v64 = 0;
}
ca->ack_sack_cnt -= rs->acked_sacked;
if (ca->state == CDG_NONFULL || shadow_fast) {
tcp_cong_avoid_ai_shadow(sk, ca->shadow_wnd, rs->acked_sacked);
tp->snd_cwnd = ca->shadow_wnd;
}
rs->flag |= CDG_CONT;
}
} else {
rs->flag |= CDG_CONT;
}
}
static void cdg_state(struct sock *sk, u8 new_state)
{
struct cdg *ca = inet_csk_ca(sk);
struct tcp_sock *tp = tcp_sk(sk);
if (!recovery_restore)
return;
if (new_state == TCP_CA_Open)
tp->snd_cwnd = max(max(tp->snd_cwnd, ca->shadow_wnd), 2U);
if (new_state == TCP_CA_Loss) {
if (ca->state == CDG_NONFULL && use_tolerance) {
tp->snd_cwnd = ca->shadow_wnd;
if (loss_push)
tcp_push_pending_frames(sk);
}
}
}
struct tcp_congestion_ops tcp_cdg __read_mostly = {
.cong_avoid = tcp_cdg_cong_avoid,
.cong_control = cdg_main,
.set_state = cdg_state,
.cwnd_event = tcp_cdg_cwnd_event,
.pkts_acked = tcp_cdg_acked,
.undo_cwnd = tcp_cdg_undo_cwnd,
.ssthresh = tcp_cdg_ssthresh,
.release = tcp_cdg_release,
.init = tcp_cdg_init,
.owner = THIS_MODULE,
.name = "cdg",
};
static int __init tcp_cdg_register(void)
{
if (backoff_beta > 1024 || window < 1 || window > 256)
return -ERANGE;
if (!is_power_of_2(window))
return -EINVAL;
BUILD_BUG_ON(sizeof(struct cdg) > ICSK_CA_PRIV_SIZE);
tcp_register_congestion_control(&tcp_cdg);
return 0;
}
static void __exit tcp_cdg_unregister(void)
{
tcp_unregister_congestion_control(&tcp_cdg);
}
module_init(tcp_cdg_register);
module_exit(tcp_cdg_unregister);
MODULE_AUTHOR("...");
MODULE_LICENSE("GPL");
MODULE_DESCRIPTION("TCP CDG");
以上是关于安卓cpu优化 tcp拥塞算法cubic和reno怎么选择的主要内容，如果未能解决你的问题，请参考以下文章