1 | // SPDX-License-Identifier: GPL-2.0-only |
2 | /* |
3 | * TCP Vegas congestion control |
4 | * |
5 | * This is based on the congestion detection/avoidance scheme described in |
6 | * Lawrence S. Brakmo and Larry L. Peterson. |
7 | * "TCP Vegas: End to end congestion avoidance on a global internet." |
8 | * IEEE Journal on Selected Areas in Communication, 13(8):1465--1480, |
9 | * October 1995. Available from: |
10 | * ftp://ftp.cs.arizona.edu/xkernel/Papers/jsac.ps |
11 | * |
12 | * See http://www.cs.arizona.edu/xkernel/ for their implementation. |
13 | * The main aspects that distinguish this implementation from the |
14 | * Arizona Vegas implementation are: |
15 | * o We do not change the loss detection or recovery mechanisms of |
16 | * Linux in any way. Linux already recovers from losses quite well, |
17 | * using fine-grained timers, NewReno, and FACK. |
18 | * o To avoid the performance penalty imposed by increasing cwnd |
19 | * only every-other RTT during slow start, we increase during |
20 | * every RTT during slow start, just like Reno. |
21 | * o Largely to allow continuous cwnd growth during slow start, |
22 | * we use the rate at which ACKs come back as the "actual" |
23 | * rate, rather than the rate at which data is sent. |
24 | * o To speed convergence to the right rate, we set the cwnd |
25 | * to achieve the right ("actual") rate when we exit slow start. |
26 | * o To filter out the noise caused by delayed ACKs, we use the |
27 | * minimum RTT sample observed during the last RTT to calculate |
28 | * the actual rate. |
29 | * o When the sender re-starts from idle, it waits until it has |
30 | * received ACKs for an entire flight of new data before making |
31 | * a cwnd adjustment decision. The original Vegas implementation |
32 | * assumed senders never went idle. |
33 | */ |
34 | |
35 | #include <linux/mm.h> |
36 | #include <linux/module.h> |
37 | #include <linux/skbuff.h> |
38 | #include <linux/inet_diag.h> |
39 | |
40 | #include <net/tcp.h> |
41 | |
42 | #include "tcp_vegas.h" |
43 | |
44 | static int alpha = 2; |
45 | static int beta = 4; |
46 | static int gamma = 1; |
47 | |
48 | module_param(alpha, int, 0644); |
49 | MODULE_PARM_DESC(alpha, "lower bound of packets in network" ); |
50 | module_param(beta, int, 0644); |
51 | MODULE_PARM_DESC(beta, "upper bound of packets in network" ); |
52 | module_param(gamma, int, 0644); |
53 | MODULE_PARM_DESC(gamma, "limit on increase (scale by 2)" ); |
54 | |
55 | /* There are several situations when we must "re-start" Vegas: |
56 | * |
57 | * o when a connection is established |
58 | * o after an RTO |
59 | * o after fast recovery |
60 | * o when we send a packet and there is no outstanding |
61 | * unacknowledged data (restarting an idle connection) |
62 | * |
63 | * In these circumstances we cannot do a Vegas calculation at the |
64 | * end of the first RTT, because any calculation we do is using |
65 | * stale info -- both the saved cwnd and congestion feedback are |
66 | * stale. |
67 | * |
68 | * Instead we must wait until the completion of an RTT during |
69 | * which we actually receive ACKs. |
70 | */ |
71 | static void vegas_enable(struct sock *sk) |
72 | { |
73 | const struct tcp_sock *tp = tcp_sk(sk); |
74 | struct vegas *vegas = inet_csk_ca(sk); |
75 | |
76 | /* Begin taking Vegas samples next time we send something. */ |
77 | vegas->doing_vegas_now = 1; |
78 | |
79 | /* Set the beginning of the next send window. */ |
80 | vegas->beg_snd_nxt = tp->snd_nxt; |
81 | |
82 | vegas->cntRTT = 0; |
83 | vegas->minRTT = 0x7fffffff; |
84 | } |
85 | |
86 | /* Stop taking Vegas samples for now. */ |
87 | static inline void vegas_disable(struct sock *sk) |
88 | { |
89 | struct vegas *vegas = inet_csk_ca(sk); |
90 | |
91 | vegas->doing_vegas_now = 0; |
92 | } |
93 | |
94 | void tcp_vegas_init(struct sock *sk) |
95 | { |
96 | struct vegas *vegas = inet_csk_ca(sk); |
97 | |
98 | vegas->baseRTT = 0x7fffffff; |
99 | vegas_enable(sk); |
100 | } |
101 | EXPORT_SYMBOL_GPL(tcp_vegas_init); |
102 | |
103 | /* Do RTT sampling needed for Vegas. |
104 | * Basically we: |
105 | * o min-filter RTT samples from within an RTT to get the current |
106 | * propagation delay + queuing delay (we are min-filtering to try to |
107 | * avoid the effects of delayed ACKs) |
108 | * o min-filter RTT samples from a much longer window (forever for now) |
109 | * to find the propagation delay (baseRTT) |
110 | */ |
111 | void tcp_vegas_pkts_acked(struct sock *sk, const struct ack_sample *sample) |
112 | { |
113 | struct vegas *vegas = inet_csk_ca(sk); |
114 | u32 vrtt; |
115 | |
116 | if (sample->rtt_us < 0) |
117 | return; |
118 | |
119 | /* Never allow zero rtt or baseRTT */ |
120 | vrtt = sample->rtt_us + 1; |
121 | |
122 | /* Filter to find propagation delay: */ |
123 | if (vrtt < vegas->baseRTT) |
124 | vegas->baseRTT = vrtt; |
125 | |
126 | /* Find the min RTT during the last RTT to find |
127 | * the current prop. delay + queuing delay: |
128 | */ |
129 | vegas->minRTT = min(vegas->minRTT, vrtt); |
130 | vegas->cntRTT++; |
131 | } |
132 | EXPORT_SYMBOL_GPL(tcp_vegas_pkts_acked); |
133 | |
134 | void tcp_vegas_state(struct sock *sk, u8 ca_state) |
135 | { |
136 | if (ca_state == TCP_CA_Open) |
137 | vegas_enable(sk); |
138 | else |
139 | vegas_disable(sk); |
140 | } |
141 | EXPORT_SYMBOL_GPL(tcp_vegas_state); |
142 | |
143 | /* |
144 | * If the connection is idle and we are restarting, |
145 | * then we don't want to do any Vegas calculations |
146 | * until we get fresh RTT samples. So when we |
147 | * restart, we reset our Vegas state to a clean |
148 | * slate. After we get acks for this flight of |
149 | * packets, _then_ we can make Vegas calculations |
150 | * again. |
151 | */ |
152 | void tcp_vegas_cwnd_event(struct sock *sk, enum tcp_ca_event event) |
153 | { |
154 | if (event == CA_EVENT_CWND_RESTART || |
155 | event == CA_EVENT_TX_START) |
156 | tcp_vegas_init(sk); |
157 | } |
158 | EXPORT_SYMBOL_GPL(tcp_vegas_cwnd_event); |
159 | |
160 | static inline u32 tcp_vegas_ssthresh(struct tcp_sock *tp) |
161 | { |
162 | return min(tp->snd_ssthresh, tcp_snd_cwnd(tp)); |
163 | } |
164 | |
165 | static void tcp_vegas_cong_avoid(struct sock *sk, u32 ack, u32 acked) |
166 | { |
167 | struct tcp_sock *tp = tcp_sk(sk); |
168 | struct vegas *vegas = inet_csk_ca(sk); |
169 | |
170 | if (!vegas->doing_vegas_now) { |
171 | tcp_reno_cong_avoid(sk, ack, acked); |
172 | return; |
173 | } |
174 | |
175 | if (after(ack, vegas->beg_snd_nxt)) { |
176 | /* Do the Vegas once-per-RTT cwnd adjustment. */ |
177 | |
178 | /* Save the extent of the current window so we can use this |
179 | * at the end of the next RTT. |
180 | */ |
181 | vegas->beg_snd_nxt = tp->snd_nxt; |
182 | |
183 | /* We do the Vegas calculations only if we got enough RTT |
184 | * samples that we can be reasonably sure that we got |
185 | * at least one RTT sample that wasn't from a delayed ACK. |
186 | * If we only had 2 samples total, |
187 | * then that means we're getting only 1 ACK per RTT, which |
188 | * means they're almost certainly delayed ACKs. |
189 | * If we have 3 samples, we should be OK. |
190 | */ |
191 | |
192 | if (vegas->cntRTT <= 2) { |
193 | /* We don't have enough RTT samples to do the Vegas |
194 | * calculation, so we'll behave like Reno. |
195 | */ |
196 | tcp_reno_cong_avoid(sk, ack, acked); |
197 | } else { |
198 | u32 rtt, diff; |
199 | u64 target_cwnd; |
200 | |
201 | /* We have enough RTT samples, so, using the Vegas |
202 | * algorithm, we determine if we should increase or |
203 | * decrease cwnd, and by how much. |
204 | */ |
205 | |
206 | /* Pluck out the RTT we are using for the Vegas |
207 | * calculations. This is the min RTT seen during the |
208 | * last RTT. Taking the min filters out the effects |
209 | * of delayed ACKs, at the cost of noticing congestion |
210 | * a bit later. |
211 | */ |
212 | rtt = vegas->minRTT; |
213 | |
214 | /* Calculate the cwnd we should have, if we weren't |
215 | * going too fast. |
216 | * |
217 | * This is: |
218 | * (actual rate in segments) * baseRTT |
219 | */ |
220 | target_cwnd = (u64)tcp_snd_cwnd(tp) * vegas->baseRTT; |
221 | do_div(target_cwnd, rtt); |
222 | |
223 | /* Calculate the difference between the window we had, |
224 | * and the window we would like to have. This quantity |
225 | * is the "Diff" from the Arizona Vegas papers. |
226 | */ |
227 | diff = tcp_snd_cwnd(tp) * (rtt-vegas->baseRTT) / vegas->baseRTT; |
228 | |
229 | if (diff > gamma && tcp_in_slow_start(tp)) { |
230 | /* Going too fast. Time to slow down |
231 | * and switch to congestion avoidance. |
232 | */ |
233 | |
234 | /* Set cwnd to match the actual rate |
235 | * exactly: |
236 | * cwnd = (actual rate) * baseRTT |
237 | * Then we add 1 because the integer |
238 | * truncation robs us of full link |
239 | * utilization. |
240 | */ |
241 | tcp_snd_cwnd_set(tp, min(tcp_snd_cwnd(tp), |
242 | (u32)target_cwnd + 1)); |
243 | tp->snd_ssthresh = tcp_vegas_ssthresh(tp); |
244 | |
245 | } else if (tcp_in_slow_start(tp)) { |
246 | /* Slow start. */ |
247 | tcp_slow_start(tp, acked); |
248 | } else { |
249 | /* Congestion avoidance. */ |
250 | |
251 | /* Figure out where we would like cwnd |
252 | * to be. |
253 | */ |
254 | if (diff > beta) { |
255 | /* The old window was too fast, so |
256 | * we slow down. |
257 | */ |
258 | tcp_snd_cwnd_set(tp, val: tcp_snd_cwnd(tp) - 1); |
259 | tp->snd_ssthresh |
260 | = tcp_vegas_ssthresh(tp); |
261 | } else if (diff < alpha) { |
262 | /* We don't have enough extra packets |
263 | * in the network, so speed up. |
264 | */ |
265 | tcp_snd_cwnd_set(tp, val: tcp_snd_cwnd(tp) + 1); |
266 | } else { |
267 | /* Sending just as fast as we |
268 | * should be. |
269 | */ |
270 | } |
271 | } |
272 | |
273 | if (tcp_snd_cwnd(tp) < 2) |
274 | tcp_snd_cwnd_set(tp, val: 2); |
275 | else if (tcp_snd_cwnd(tp) > tp->snd_cwnd_clamp) |
276 | tcp_snd_cwnd_set(tp, val: tp->snd_cwnd_clamp); |
277 | |
278 | tp->snd_ssthresh = tcp_current_ssthresh(sk); |
279 | } |
280 | |
281 | /* Wipe the slate clean for the next RTT. */ |
282 | vegas->cntRTT = 0; |
283 | vegas->minRTT = 0x7fffffff; |
284 | } |
285 | /* Use normal slow start */ |
286 | else if (tcp_in_slow_start(tp)) |
287 | tcp_slow_start(tp, acked); |
288 | } |
289 | |
290 | /* Extract info for Tcp socket info provided via netlink. */ |
291 | size_t tcp_vegas_get_info(struct sock *sk, u32 ext, int *attr, |
292 | union tcp_cc_info *info) |
293 | { |
294 | const struct vegas *ca = inet_csk_ca(sk); |
295 | |
296 | if (ext & (1 << (INET_DIAG_VEGASINFO - 1))) { |
297 | info->vegas.tcpv_enabled = ca->doing_vegas_now; |
298 | info->vegas.tcpv_rttcnt = ca->cntRTT; |
299 | info->vegas.tcpv_rtt = ca->baseRTT; |
300 | info->vegas.tcpv_minrtt = ca->minRTT; |
301 | |
302 | *attr = INET_DIAG_VEGASINFO; |
303 | return sizeof(struct tcpvegas_info); |
304 | } |
305 | return 0; |
306 | } |
307 | EXPORT_SYMBOL_GPL(tcp_vegas_get_info); |
308 | |
309 | static struct tcp_congestion_ops tcp_vegas __read_mostly = { |
310 | .init = tcp_vegas_init, |
311 | .ssthresh = tcp_reno_ssthresh, |
312 | .undo_cwnd = tcp_reno_undo_cwnd, |
313 | .cong_avoid = tcp_vegas_cong_avoid, |
314 | .pkts_acked = tcp_vegas_pkts_acked, |
315 | .set_state = tcp_vegas_state, |
316 | .cwnd_event = tcp_vegas_cwnd_event, |
317 | .get_info = tcp_vegas_get_info, |
318 | |
319 | .owner = THIS_MODULE, |
320 | .name = "vegas" , |
321 | }; |
322 | |
323 | static int __init tcp_vegas_register(void) |
324 | { |
325 | BUILD_BUG_ON(sizeof(struct vegas) > ICSK_CA_PRIV_SIZE); |
326 | tcp_register_congestion_control(type: &tcp_vegas); |
327 | return 0; |
328 | } |
329 | |
330 | static void __exit tcp_vegas_unregister(void) |
331 | { |
332 | tcp_unregister_congestion_control(type: &tcp_vegas); |
333 | } |
334 | |
335 | module_init(tcp_vegas_register); |
336 | module_exit(tcp_vegas_unregister); |
337 | |
338 | MODULE_AUTHOR("Stephen Hemminger" ); |
339 | MODULE_LICENSE("GPL" ); |
340 | MODULE_DESCRIPTION("TCP Vegas" ); |
341 | |