xref: /openbmc/linux/net/ipv4/tcp_htcp.c (revision 2359ccdd)
1 /*
2  * H-TCP congestion control. The algorithm is detailed in:
3  * R.N.Shorten, D.J.Leith:
4  *   "H-TCP: TCP for high-speed and long-distance networks"
5  *   Proc. PFLDnet, Argonne, 2004.
6  * http://www.hamilton.ie/net/htcp3.pdf
7  */
8 
9 #include <linux/mm.h>
10 #include <linux/module.h>
11 #include <net/tcp.h>
12 
13 #define ALPHA_BASE	(1<<7)	/* 1.0 with shift << 7 */
14 #define BETA_MIN	(1<<6)	/* 0.5 with shift << 7 */
15 #define BETA_MAX	102	/* 0.8 with shift << 7 */
16 
17 static int use_rtt_scaling __read_mostly = 1;
18 module_param(use_rtt_scaling, int, 0644);
19 MODULE_PARM_DESC(use_rtt_scaling, "turn on/off RTT scaling");
20 
21 static int use_bandwidth_switch __read_mostly = 1;
22 module_param(use_bandwidth_switch, int, 0644);
23 MODULE_PARM_DESC(use_bandwidth_switch, "turn on/off bandwidth switcher");
24 
25 struct htcp {
26 	u32	alpha;		/* Fixed point arith, << 7 */
27 	u8	beta;           /* Fixed point arith, << 7 */
28 	u8	modeswitch;	/* Delay modeswitch
29 				   until we had at least one congestion event */
30 	u16	pkts_acked;
31 	u32	packetcount;
32 	u32	minRTT;
33 	u32	maxRTT;
34 	u32	last_cong;	/* Time since last congestion event end */
35 	u32	undo_last_cong;
36 
37 	u32	undo_maxRTT;
38 	u32	undo_old_maxB;
39 
40 	/* Bandwidth estimation */
41 	u32	minB;
42 	u32	maxB;
43 	u32	old_maxB;
44 	u32	Bi;
45 	u32	lasttime;
46 };
47 
48 static inline u32 htcp_cong_time(const struct htcp *ca)
49 {
50 	return jiffies - ca->last_cong;
51 }
52 
53 static inline u32 htcp_ccount(const struct htcp *ca)
54 {
55 	return htcp_cong_time(ca) / ca->minRTT;
56 }
57 
58 static inline void htcp_reset(struct htcp *ca)
59 {
60 	ca->undo_last_cong = ca->last_cong;
61 	ca->undo_maxRTT = ca->maxRTT;
62 	ca->undo_old_maxB = ca->old_maxB;
63 
64 	ca->last_cong = jiffies;
65 }
66 
67 static u32 htcp_cwnd_undo(struct sock *sk)
68 {
69 	struct htcp *ca = inet_csk_ca(sk);
70 
71 	if (ca->undo_last_cong) {
72 		ca->last_cong = ca->undo_last_cong;
73 		ca->maxRTT = ca->undo_maxRTT;
74 		ca->old_maxB = ca->undo_old_maxB;
75 		ca->undo_last_cong = 0;
76 	}
77 
78 	return tcp_reno_undo_cwnd(sk);
79 }
80 
81 static inline void measure_rtt(struct sock *sk, u32 srtt)
82 {
83 	const struct inet_connection_sock *icsk = inet_csk(sk);
84 	struct htcp *ca = inet_csk_ca(sk);
85 
86 	/* keep track of minimum RTT seen so far, minRTT is zero at first */
87 	if (ca->minRTT > srtt || !ca->minRTT)
88 		ca->minRTT = srtt;
89 
90 	/* max RTT */
91 	if (icsk->icsk_ca_state == TCP_CA_Open) {
92 		if (ca->maxRTT < ca->minRTT)
93 			ca->maxRTT = ca->minRTT;
94 		if (ca->maxRTT < srtt &&
95 		    srtt <= ca->maxRTT + msecs_to_jiffies(20))
96 			ca->maxRTT = srtt;
97 	}
98 }
99 
100 static void measure_achieved_throughput(struct sock *sk,
101 					const struct ack_sample *sample)
102 {
103 	const struct inet_connection_sock *icsk = inet_csk(sk);
104 	const struct tcp_sock *tp = tcp_sk(sk);
105 	struct htcp *ca = inet_csk_ca(sk);
106 	u32 now = tcp_jiffies32;
107 
108 	if (icsk->icsk_ca_state == TCP_CA_Open)
109 		ca->pkts_acked = sample->pkts_acked;
110 
111 	if (sample->rtt_us > 0)
112 		measure_rtt(sk, usecs_to_jiffies(sample->rtt_us));
113 
114 	if (!use_bandwidth_switch)
115 		return;
116 
117 	/* achieved throughput calculations */
118 	if (!((1 << icsk->icsk_ca_state) & (TCPF_CA_Open | TCPF_CA_Disorder))) {
119 		ca->packetcount = 0;
120 		ca->lasttime = now;
121 		return;
122 	}
123 
124 	ca->packetcount += sample->pkts_acked;
125 
126 	if (ca->packetcount >= tp->snd_cwnd - (ca->alpha >> 7 ? : 1) &&
127 	    now - ca->lasttime >= ca->minRTT &&
128 	    ca->minRTT > 0) {
129 		__u32 cur_Bi = ca->packetcount * HZ / (now - ca->lasttime);
130 
131 		if (htcp_ccount(ca) <= 3) {
132 			/* just after backoff */
133 			ca->minB = ca->maxB = ca->Bi = cur_Bi;
134 		} else {
135 			ca->Bi = (3 * ca->Bi + cur_Bi) / 4;
136 			if (ca->Bi > ca->maxB)
137 				ca->maxB = ca->Bi;
138 			if (ca->minB > ca->maxB)
139 				ca->minB = ca->maxB;
140 		}
141 		ca->packetcount = 0;
142 		ca->lasttime = now;
143 	}
144 }
145 
146 static inline void htcp_beta_update(struct htcp *ca, u32 minRTT, u32 maxRTT)
147 {
148 	if (use_bandwidth_switch) {
149 		u32 maxB = ca->maxB;
150 		u32 old_maxB = ca->old_maxB;
151 
152 		ca->old_maxB = ca->maxB;
153 		if (!between(5 * maxB, 4 * old_maxB, 6 * old_maxB)) {
154 			ca->beta = BETA_MIN;
155 			ca->modeswitch = 0;
156 			return;
157 		}
158 	}
159 
160 	if (ca->modeswitch && minRTT > msecs_to_jiffies(10) && maxRTT) {
161 		ca->beta = (minRTT << 7) / maxRTT;
162 		if (ca->beta < BETA_MIN)
163 			ca->beta = BETA_MIN;
164 		else if (ca->beta > BETA_MAX)
165 			ca->beta = BETA_MAX;
166 	} else {
167 		ca->beta = BETA_MIN;
168 		ca->modeswitch = 1;
169 	}
170 }
171 
172 static inline void htcp_alpha_update(struct htcp *ca)
173 {
174 	u32 minRTT = ca->minRTT;
175 	u32 factor = 1;
176 	u32 diff = htcp_cong_time(ca);
177 
178 	if (diff > HZ) {
179 		diff -= HZ;
180 		factor = 1 + (10 * diff + ((diff / 2) * (diff / 2) / HZ)) / HZ;
181 	}
182 
183 	if (use_rtt_scaling && minRTT) {
184 		u32 scale = (HZ << 3) / (10 * minRTT);
185 
186 		/* clamping ratio to interval [0.5,10]<<3 */
187 		scale = min(max(scale, 1U << 2), 10U << 3);
188 		factor = (factor << 3) / scale;
189 		if (!factor)
190 			factor = 1;
191 	}
192 
193 	ca->alpha = 2 * factor * ((1 << 7) - ca->beta);
194 	if (!ca->alpha)
195 		ca->alpha = ALPHA_BASE;
196 }
197 
198 /*
199  * After we have the rtt data to calculate beta, we'd still prefer to wait one
200  * rtt before we adjust our beta to ensure we are working from a consistent
201  * data.
202  *
203  * This function should be called when we hit a congestion event since only at
204  * that point do we really have a real sense of maxRTT (the queues en route
205  * were getting just too full now).
206  */
207 static void htcp_param_update(struct sock *sk)
208 {
209 	struct htcp *ca = inet_csk_ca(sk);
210 	u32 minRTT = ca->minRTT;
211 	u32 maxRTT = ca->maxRTT;
212 
213 	htcp_beta_update(ca, minRTT, maxRTT);
214 	htcp_alpha_update(ca);
215 
216 	/* add slowly fading memory for maxRTT to accommodate routing changes */
217 	if (minRTT > 0 && maxRTT > minRTT)
218 		ca->maxRTT = minRTT + ((maxRTT - minRTT) * 95) / 100;
219 }
220 
221 static u32 htcp_recalc_ssthresh(struct sock *sk)
222 {
223 	const struct tcp_sock *tp = tcp_sk(sk);
224 	const struct htcp *ca = inet_csk_ca(sk);
225 
226 	htcp_param_update(sk);
227 	return max((tp->snd_cwnd * ca->beta) >> 7, 2U);
228 }
229 
230 static void htcp_cong_avoid(struct sock *sk, u32 ack, u32 acked)
231 {
232 	struct tcp_sock *tp = tcp_sk(sk);
233 	struct htcp *ca = inet_csk_ca(sk);
234 
235 	if (!tcp_is_cwnd_limited(sk))
236 		return;
237 
238 	if (tcp_in_slow_start(tp))
239 		tcp_slow_start(tp, acked);
240 	else {
241 		/* In dangerous area, increase slowly.
242 		 * In theory this is tp->snd_cwnd += alpha / tp->snd_cwnd
243 		 */
244 		if ((tp->snd_cwnd_cnt * ca->alpha)>>7 >= tp->snd_cwnd) {
245 			if (tp->snd_cwnd < tp->snd_cwnd_clamp)
246 				tp->snd_cwnd++;
247 			tp->snd_cwnd_cnt = 0;
248 			htcp_alpha_update(ca);
249 		} else
250 			tp->snd_cwnd_cnt += ca->pkts_acked;
251 
252 		ca->pkts_acked = 1;
253 	}
254 }
255 
256 static void htcp_init(struct sock *sk)
257 {
258 	struct htcp *ca = inet_csk_ca(sk);
259 
260 	memset(ca, 0, sizeof(struct htcp));
261 	ca->alpha = ALPHA_BASE;
262 	ca->beta = BETA_MIN;
263 	ca->pkts_acked = 1;
264 	ca->last_cong = jiffies;
265 }
266 
267 static void htcp_state(struct sock *sk, u8 new_state)
268 {
269 	switch (new_state) {
270 	case TCP_CA_Open:
271 		{
272 			struct htcp *ca = inet_csk_ca(sk);
273 
274 			if (ca->undo_last_cong) {
275 				ca->last_cong = jiffies;
276 				ca->undo_last_cong = 0;
277 			}
278 		}
279 		break;
280 	case TCP_CA_CWR:
281 	case TCP_CA_Recovery:
282 	case TCP_CA_Loss:
283 		htcp_reset(inet_csk_ca(sk));
284 		break;
285 	}
286 }
287 
288 static struct tcp_congestion_ops htcp __read_mostly = {
289 	.init		= htcp_init,
290 	.ssthresh	= htcp_recalc_ssthresh,
291 	.cong_avoid	= htcp_cong_avoid,
292 	.set_state	= htcp_state,
293 	.undo_cwnd	= htcp_cwnd_undo,
294 	.pkts_acked	= measure_achieved_throughput,
295 	.owner		= THIS_MODULE,
296 	.name		= "htcp",
297 };
298 
299 static int __init htcp_register(void)
300 {
301 	BUILD_BUG_ON(sizeof(struct htcp) > ICSK_CA_PRIV_SIZE);
302 	BUILD_BUG_ON(BETA_MIN >= BETA_MAX);
303 	return tcp_register_congestion_control(&htcp);
304 }
305 
306 static void __exit htcp_unregister(void)
307 {
308 	tcp_unregister_congestion_control(&htcp);
309 }
310 
311 module_init(htcp_register);
312 module_exit(htcp_unregister);
313 
314 MODULE_AUTHOR("Baruch Even");
315 MODULE_LICENSE("GPL");
316 MODULE_DESCRIPTION("H-TCP");
317