]>
Commit | Line | Data |
---|---|---|
a7868ea6 BE |
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 | ||
a7868ea6 BE |
9 | #include <linux/mm.h> |
10 | #include <linux/module.h> | |
11 | #include <net/tcp.h> | |
12 | ||
9121c777 SH |
13 | #define ALPHA_BASE (1<<7) /* 1.0 with shift << 7 */ |
14 | #define BETA_MIN (1<<6) /* 0.5 with shift << 7 */ | |
a7868ea6 BE |
15 | #define BETA_MAX 102 /* 0.8 with shift << 7 */ |
16 | ||
9121c777 | 17 | static int use_rtt_scaling __read_mostly = 1; |
a7868ea6 BE |
18 | module_param(use_rtt_scaling, int, 0644); |
19 | MODULE_PARM_DESC(use_rtt_scaling, "turn on/off RTT scaling"); | |
20 | ||
9121c777 | 21 | static int use_bandwidth_switch __read_mostly = 1; |
a7868ea6 BE |
22 | module_param(use_bandwidth_switch, int, 0644); |
23 | MODULE_PARM_DESC(use_bandwidth_switch, "turn on/off bandwidth switcher"); | |
24 | ||
25 | struct htcp { | |
2a272f98 | 26 | u32 alpha; /* Fixed point arith, << 7 */ |
a7868ea6 | 27 | u8 beta; /* Fixed point arith, << 7 */ |
9121c777 SH |
28 | u8 modeswitch; /* Delay modeswitch |
29 | until we had at least one congestion event */ | |
0bc6d90b BE |
30 | u16 pkts_acked; |
31 | u32 packetcount; | |
a7868ea6 BE |
32 | u32 minRTT; |
33 | u32 maxRTT; | |
2404043a DM |
34 | u32 last_cong; /* Time since last congestion event end */ |
35 | u32 undo_last_cong; | |
a7868ea6 BE |
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 | ||
9121c777 | 48 | static inline u32 htcp_cong_time(const struct htcp *ca) |
50bf3e22 BE |
49 | { |
50 | return jiffies - ca->last_cong; | |
51 | } | |
52 | ||
9121c777 | 53 | static inline u32 htcp_ccount(const struct htcp *ca) |
50bf3e22 | 54 | { |
9121c777 | 55 | return htcp_cong_time(ca) / ca->minRTT; |
50bf3e22 BE |
56 | } |
57 | ||
a7868ea6 BE |
58 | static inline void htcp_reset(struct htcp *ca) |
59 | { | |
50bf3e22 | 60 | ca->undo_last_cong = ca->last_cong; |
a7868ea6 BE |
61 | ca->undo_maxRTT = ca->maxRTT; |
62 | ca->undo_old_maxB = ca->old_maxB; | |
63 | ||
50bf3e22 | 64 | ca->last_cong = jiffies; |
a7868ea6 BE |
65 | } |
66 | ||
6687e988 | 67 | static u32 htcp_cwnd_undo(struct sock *sk) |
a7868ea6 | 68 | { |
6687e988 ACM |
69 | const struct tcp_sock *tp = tcp_sk(sk); |
70 | struct htcp *ca = inet_csk_ca(sk); | |
9121c777 | 71 | |
8f65b535 DL |
72 | if (ca->undo_last_cong) { |
73 | ca->last_cong = ca->undo_last_cong; | |
74 | ca->maxRTT = ca->undo_maxRTT; | |
75 | ca->old_maxB = ca->undo_old_maxB; | |
76 | ca->undo_last_cong = 0; | |
77 | } | |
9121c777 SH |
78 | |
79 | return max(tp->snd_cwnd, (tp->snd_ssthresh << 7) / ca->beta); | |
a7868ea6 BE |
80 | } |
81 | ||
113bbbd8 | 82 | static inline void measure_rtt(struct sock *sk, u32 srtt) |
a7868ea6 | 83 | { |
6687e988 | 84 | const struct inet_connection_sock *icsk = inet_csk(sk); |
6687e988 | 85 | struct htcp *ca = inet_csk_ca(sk); |
a7868ea6 BE |
86 | |
87 | /* keep track of minimum RTT seen so far, minRTT is zero at first */ | |
88 | if (ca->minRTT > srtt || !ca->minRTT) | |
89 | ca->minRTT = srtt; | |
90 | ||
91 | /* max RTT */ | |
f34d1955 | 92 | if (icsk->icsk_ca_state == TCP_CA_Open) { |
a7868ea6 BE |
93 | if (ca->maxRTT < ca->minRTT) |
94 | ca->maxRTT = ca->minRTT; | |
9d4fb27d JP |
95 | if (ca->maxRTT < srtt && |
96 | srtt <= ca->maxRTT + msecs_to_jiffies(20)) | |
a7868ea6 BE |
97 | ca->maxRTT = srtt; |
98 | } | |
99 | } | |
100 | ||
688d1945 | 101 | static void measure_achieved_throughput(struct sock *sk, |
102 | u32 pkts_acked, s32 rtt) | |
a7868ea6 | 103 | { |
6687e988 ACM |
104 | const struct inet_connection_sock *icsk = inet_csk(sk); |
105 | const struct tcp_sock *tp = tcp_sk(sk); | |
106 | struct htcp *ca = inet_csk_ca(sk); | |
a7868ea6 BE |
107 | u32 now = tcp_time_stamp; |
108 | ||
0bc6d90b BE |
109 | if (icsk->icsk_ca_state == TCP_CA_Open) |
110 | ca->pkts_acked = pkts_acked; | |
111 | ||
113bbbd8 SH |
112 | if (rtt > 0) |
113 | measure_rtt(sk, usecs_to_jiffies(rtt)); | |
114 | ||
0bc6d90b BE |
115 | if (!use_bandwidth_switch) |
116 | return; | |
117 | ||
a7868ea6 | 118 | /* achieved throughput calculations */ |
571a5dd8 | 119 | if (!((1 << icsk->icsk_ca_state) & (TCPF_CA_Open | TCPF_CA_Disorder))) { |
a7868ea6 BE |
120 | ca->packetcount = 0; |
121 | ca->lasttime = now; | |
122 | return; | |
123 | } | |
124 | ||
125 | ca->packetcount += pkts_acked; | |
126 | ||
9d4fb27d JP |
127 | if (ca->packetcount >= tp->snd_cwnd - (ca->alpha >> 7 ? : 1) && |
128 | now - ca->lasttime >= ca->minRTT && | |
129 | ca->minRTT > 0) { | |
9121c777 SH |
130 | __u32 cur_Bi = ca->packetcount * HZ / (now - ca->lasttime); |
131 | ||
50bf3e22 | 132 | if (htcp_ccount(ca) <= 3) { |
a7868ea6 BE |
133 | /* just after backoff */ |
134 | ca->minB = ca->maxB = ca->Bi = cur_Bi; | |
135 | } else { | |
9121c777 | 136 | ca->Bi = (3 * ca->Bi + cur_Bi) / 4; |
a7868ea6 BE |
137 | if (ca->Bi > ca->maxB) |
138 | ca->maxB = ca->Bi; | |
139 | if (ca->minB > ca->maxB) | |
140 | ca->minB = ca->maxB; | |
141 | } | |
142 | ca->packetcount = 0; | |
143 | ca->lasttime = now; | |
144 | } | |
145 | } | |
146 | ||
147 | static inline void htcp_beta_update(struct htcp *ca, u32 minRTT, u32 maxRTT) | |
148 | { | |
149 | if (use_bandwidth_switch) { | |
150 | u32 maxB = ca->maxB; | |
151 | u32 old_maxB = ca->old_maxB; | |
a7868ea6 | 152 | |
688d1945 | 153 | ca->old_maxB = ca->maxB; |
9121c777 | 154 | if (!between(5 * maxB, 4 * old_maxB, 6 * old_maxB)) { |
a7868ea6 BE |
155 | ca->beta = BETA_MIN; |
156 | ca->modeswitch = 0; | |
157 | return; | |
158 | } | |
159 | } | |
160 | ||
c33ad6e4 | 161 | if (ca->modeswitch && minRTT > msecs_to_jiffies(10) && maxRTT) { |
9121c777 | 162 | ca->beta = (minRTT << 7) / maxRTT; |
a7868ea6 BE |
163 | if (ca->beta < BETA_MIN) |
164 | ca->beta = BETA_MIN; | |
165 | else if (ca->beta > BETA_MAX) | |
166 | ca->beta = BETA_MAX; | |
167 | } else { | |
168 | ca->beta = BETA_MIN; | |
169 | ca->modeswitch = 1; | |
170 | } | |
171 | } | |
172 | ||
173 | static inline void htcp_alpha_update(struct htcp *ca) | |
174 | { | |
175 | u32 minRTT = ca->minRTT; | |
176 | u32 factor = 1; | |
50bf3e22 | 177 | u32 diff = htcp_cong_time(ca); |
a7868ea6 BE |
178 | |
179 | if (diff > HZ) { | |
180 | diff -= HZ; | |
9121c777 | 181 | factor = 1 + (10 * diff + ((diff / 2) * (diff / 2) / HZ)) / HZ; |
a7868ea6 BE |
182 | } |
183 | ||
184 | if (use_rtt_scaling && minRTT) { | |
9121c777 SH |
185 | u32 scale = (HZ << 3) / (10 * minRTT); |
186 | ||
187 | /* clamping ratio to interval [0.5,10]<<3 */ | |
188 | scale = min(max(scale, 1U << 2), 10U << 3); | |
189 | factor = (factor << 3) / scale; | |
a7868ea6 BE |
190 | if (!factor) |
191 | factor = 1; | |
192 | } | |
193 | ||
9121c777 | 194 | ca->alpha = 2 * factor * ((1 << 7) - ca->beta); |
a7868ea6 BE |
195 | if (!ca->alpha) |
196 | ca->alpha = ALPHA_BASE; | |
197 | } | |
198 | ||
9121c777 SH |
199 | /* |
200 | * After we have the rtt data to calculate beta, we'd still prefer to wait one | |
a7868ea6 BE |
201 | * rtt before we adjust our beta to ensure we are working from a consistent |
202 | * data. | |
203 | * | |
204 | * This function should be called when we hit a congestion event since only at | |
205 | * that point do we really have a real sense of maxRTT (the queues en route | |
206 | * were getting just too full now). | |
207 | */ | |
6687e988 | 208 | static void htcp_param_update(struct sock *sk) |
a7868ea6 | 209 | { |
6687e988 | 210 | struct htcp *ca = inet_csk_ca(sk); |
a7868ea6 BE |
211 | u32 minRTT = ca->minRTT; |
212 | u32 maxRTT = ca->maxRTT; | |
213 | ||
214 | htcp_beta_update(ca, minRTT, maxRTT); | |
215 | htcp_alpha_update(ca); | |
216 | ||
9121c777 | 217 | /* add slowly fading memory for maxRTT to accommodate routing changes */ |
a7868ea6 | 218 | if (minRTT > 0 && maxRTT > minRTT) |
9121c777 | 219 | ca->maxRTT = minRTT + ((maxRTT - minRTT) * 95) / 100; |
a7868ea6 BE |
220 | } |
221 | ||
6687e988 | 222 | static u32 htcp_recalc_ssthresh(struct sock *sk) |
a7868ea6 | 223 | { |
6687e988 ACM |
224 | const struct tcp_sock *tp = tcp_sk(sk); |
225 | const struct htcp *ca = inet_csk_ca(sk); | |
9121c777 | 226 | |
6687e988 | 227 | htcp_param_update(sk); |
a7868ea6 BE |
228 | return max((tp->snd_cwnd * ca->beta) >> 7, 2U); |
229 | } | |
230 | ||
24901551 | 231 | static void htcp_cong_avoid(struct sock *sk, u32 ack, u32 acked) |
a7868ea6 | 232 | { |
6687e988 ACM |
233 | struct tcp_sock *tp = tcp_sk(sk); |
234 | struct htcp *ca = inet_csk_ca(sk); | |
a7868ea6 | 235 | |
24901551 | 236 | if (!tcp_is_cwnd_limited(sk)) |
a7868ea6 BE |
237 | return; |
238 | ||
071d5080 | 239 | if (tcp_in_slow_start(tp)) |
9f9843a7 | 240 | tcp_slow_start(tp, acked); |
7faffa1c | 241 | else { |
7faffa1c | 242 | /* In dangerous area, increase slowly. |
a7868ea6 BE |
243 | * In theory this is tp->snd_cwnd += alpha / tp->snd_cwnd |
244 | */ | |
0bc6d90b | 245 | if ((tp->snd_cwnd_cnt * ca->alpha)>>7 >= tp->snd_cwnd) { |
a7868ea6 BE |
246 | if (tp->snd_cwnd < tp->snd_cwnd_clamp) |
247 | tp->snd_cwnd++; | |
248 | tp->snd_cwnd_cnt = 0; | |
50bf3e22 | 249 | htcp_alpha_update(ca); |
0bc6d90b BE |
250 | } else |
251 | tp->snd_cwnd_cnt += ca->pkts_acked; | |
252 | ||
253 | ca->pkts_acked = 1; | |
a7868ea6 BE |
254 | } |
255 | } | |
256 | ||
6687e988 | 257 | static void htcp_init(struct sock *sk) |
a7868ea6 | 258 | { |
6687e988 | 259 | struct htcp *ca = inet_csk_ca(sk); |
a7868ea6 BE |
260 | |
261 | memset(ca, 0, sizeof(struct htcp)); | |
262 | ca->alpha = ALPHA_BASE; | |
263 | ca->beta = BETA_MIN; | |
0bc6d90b | 264 | ca->pkts_acked = 1; |
50bf3e22 | 265 | ca->last_cong = jiffies; |
a7868ea6 BE |
266 | } |
267 | ||
6687e988 | 268 | static void htcp_state(struct sock *sk, u8 new_state) |
a7868ea6 BE |
269 | { |
270 | switch (new_state) { | |
50bf3e22 BE |
271 | case TCP_CA_Open: |
272 | { | |
273 | struct htcp *ca = inet_csk_ca(sk); | |
688d1945 | 274 | |
8f65b535 DL |
275 | if (ca->undo_last_cong) { |
276 | ca->last_cong = jiffies; | |
277 | ca->undo_last_cong = 0; | |
278 | } | |
50bf3e22 BE |
279 | } |
280 | break; | |
a7868ea6 BE |
281 | case TCP_CA_CWR: |
282 | case TCP_CA_Recovery: | |
283 | case TCP_CA_Loss: | |
6687e988 | 284 | htcp_reset(inet_csk_ca(sk)); |
a7868ea6 BE |
285 | break; |
286 | } | |
287 | } | |
288 | ||
a252bebe | 289 | static struct tcp_congestion_ops htcp __read_mostly = { |
a7868ea6 BE |
290 | .init = htcp_init, |
291 | .ssthresh = htcp_recalc_ssthresh, | |
a7868ea6 BE |
292 | .cong_avoid = htcp_cong_avoid, |
293 | .set_state = htcp_state, | |
294 | .undo_cwnd = htcp_cwnd_undo, | |
295 | .pkts_acked = measure_achieved_throughput, | |
296 | .owner = THIS_MODULE, | |
297 | .name = "htcp", | |
298 | }; | |
299 | ||
300 | static int __init htcp_register(void) | |
301 | { | |
74975d40 | 302 | BUILD_BUG_ON(sizeof(struct htcp) > ICSK_CA_PRIV_SIZE); |
a7868ea6 | 303 | BUILD_BUG_ON(BETA_MIN >= BETA_MAX); |
a7868ea6 BE |
304 | return tcp_register_congestion_control(&htcp); |
305 | } | |
306 | ||
307 | static void __exit htcp_unregister(void) | |
308 | { | |
309 | tcp_unregister_congestion_control(&htcp); | |
310 | } | |
311 | ||
312 | module_init(htcp_register); | |
313 | module_exit(htcp_unregister); | |
314 | ||
315 | MODULE_AUTHOR("Baruch Even"); | |
316 | MODULE_LICENSE("GPL"); | |
317 | MODULE_DESCRIPTION("H-TCP"); |