1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3 * TCP Low Priority (TCP-LP)
4 *
5 * TCP Low Priority is a distributed algorithm whose goal is to utilize only
6 * the excess network bandwidth as compared to the ``fair share`` of
7 * bandwidth as targeted by TCP.
8 *
9 * As of 2.6.13, Linux supports pluggable congestion control algorithms.
10 * Due to the limitation of the API, we take the following changes from
11 * the original TCP-LP implementation:
12 * o We use newReno in most core CA handling. Only add some checking
13 * within cong_avoid.
14 * o Error correcting in remote HZ, therefore remote HZ will be keeped
15 * on checking and updating.
16 * o Handling calculation of One-Way-Delay (OWD) within rtt_sample, since
17 * OWD have a similar meaning as RTT. Also correct the buggy formular.
18 * o Handle reaction for Early Congestion Indication (ECI) within
19 * pkts_acked, as mentioned within pseudo code.
20 * o OWD is handled in relative format, where local time stamp will in
21 * tcp_time_stamp format.
22 *
23 * Original Author:
24 * Aleksandar Kuzmanovic <akuzma@northwestern.edu>
25 * Available from:
26 * http://www.ece.rice.edu/~akuzma/Doc/akuzma/TCP-LP.pdf
27 * Original implementation for 2.4.19:
28 * http://www-ece.rice.edu/networks/TCP-LP/
29 *
30 * 2.6.x module Authors:
31 * Wong Hoi Sing, Edison <hswong3i@gmail.com>
32 * Hung Hing Lun, Mike <hlhung3i@gmail.com>
33 * SourceForge project page:
34 * http://tcp-lp-mod.sourceforge.net/
35 */
36
37 #include <linux/module.h>
38 #include <net/tcp.h>
39
40 /* resolution of owd */
41 #define LP_RESOL TCP_TS_HZ
42
43 /**
44 * enum tcp_lp_state
45 * @LP_VALID_RHZ: is remote HZ valid?
46 * @LP_VALID_OWD: is OWD valid?
47 * @LP_WITHIN_THR: are we within threshold?
48 * @LP_WITHIN_INF: are we within inference?
49 *
50 * TCP-LP's state flags.
51 * We create this set of state flag mainly for debugging.
52 */
53 enum tcp_lp_state {
54 LP_VALID_RHZ = (1 << 0),
55 LP_VALID_OWD = (1 << 1),
56 LP_WITHIN_THR = (1 << 3),
57 LP_WITHIN_INF = (1 << 4),
58 };
59
60 /**
61 * struct lp
62 * @flag: TCP-LP state flag
63 * @sowd: smoothed OWD << 3
64 * @owd_min: min OWD
65 * @owd_max: max OWD
66 * @owd_max_rsv: resrved max owd
67 * @remote_hz: estimated remote HZ
68 * @remote_ref_time: remote reference time
69 * @local_ref_time: local reference time
70 * @last_drop: time for last active drop
71 * @inference: current inference
72 *
73 * TCP-LP's private struct.
74 * We get the idea from original TCP-LP implementation where only left those we
75 * found are really useful.
76 */
77 struct lp {
78 u32 flag;
79 u32 sowd;
80 u32 owd_min;
81 u32 owd_max;
82 u32 owd_max_rsv;
83 u32 remote_hz;
84 u32 remote_ref_time;
85 u32 local_ref_time;
86 u32 last_drop;
87 u32 inference;
88 };
89
90 /**
91 * tcp_lp_init
92 *
93 * Init all required variables.
94 * Clone the handling from Vegas module implementation.
95 */
tcp_lp_init(struct sock * sk)96 static void tcp_lp_init(struct sock *sk)
97 {
98 struct lp *lp = inet_csk_ca(sk);
99
100 lp->flag = 0;
101 lp->sowd = 0;
102 lp->owd_min = 0xffffffff;
103 lp->owd_max = 0;
104 lp->owd_max_rsv = 0;
105 lp->remote_hz = 0;
106 lp->remote_ref_time = 0;
107 lp->local_ref_time = 0;
108 lp->last_drop = 0;
109 lp->inference = 0;
110 }
111
112 /**
113 * tcp_lp_cong_avoid
114 *
115 * Implementation of cong_avoid.
116 * Will only call newReno CA when away from inference.
117 * From TCP-LP's paper, this will be handled in additive increasement.
118 */
tcp_lp_cong_avoid(struct sock * sk,u32 ack,u32 acked)119 static void tcp_lp_cong_avoid(struct sock *sk, u32 ack, u32 acked)
120 {
121 struct lp *lp = inet_csk_ca(sk);
122
123 if (!(lp->flag & LP_WITHIN_INF))
124 tcp_reno_cong_avoid(sk, ack, acked);
125 }
126
127 /**
128 * tcp_lp_remote_hz_estimator
129 *
130 * Estimate remote HZ.
131 * We keep on updating the estimated value, where original TCP-LP
132 * implementation only guest it for once and use forever.
133 */
tcp_lp_remote_hz_estimator(struct sock * sk)134 static u32 tcp_lp_remote_hz_estimator(struct sock *sk)
135 {
136 struct tcp_sock *tp = tcp_sk(sk);
137 struct lp *lp = inet_csk_ca(sk);
138 s64 rhz = lp->remote_hz << 6; /* remote HZ << 6 */
139 s64 m = 0;
140
141 /* not yet record reference time
142 * go away!! record it before come back!! */
143 if (lp->remote_ref_time == 0 || lp->local_ref_time == 0)
144 goto out;
145
146 /* we can't calc remote HZ with no different!! */
147 if (tp->rx_opt.rcv_tsval == lp->remote_ref_time ||
148 tp->rx_opt.rcv_tsecr == lp->local_ref_time)
149 goto out;
150
151 m = TCP_TS_HZ *
152 (tp->rx_opt.rcv_tsval - lp->remote_ref_time) /
153 (tp->rx_opt.rcv_tsecr - lp->local_ref_time);
154 if (m < 0)
155 m = -m;
156
157 if (rhz > 0) {
158 m -= rhz >> 6; /* m is now error in remote HZ est */
159 rhz += m; /* 63/64 old + 1/64 new */
160 } else
161 rhz = m << 6;
162
163 out:
164 /* record time for successful remote HZ calc */
165 if ((rhz >> 6) > 0)
166 lp->flag |= LP_VALID_RHZ;
167 else
168 lp->flag &= ~LP_VALID_RHZ;
169
170 /* record reference time stamp */
171 lp->remote_ref_time = tp->rx_opt.rcv_tsval;
172 lp->local_ref_time = tp->rx_opt.rcv_tsecr;
173
174 return rhz >> 6;
175 }
176
177 /**
178 * tcp_lp_owd_calculator
179 *
180 * Calculate one way delay (in relative format).
181 * Original implement OWD as minus of remote time difference to local time
182 * difference directly. As this time difference just simply equal to RTT, when
183 * the network status is stable, remote RTT will equal to local RTT, and result
184 * OWD into zero.
185 * It seems to be a bug and so we fixed it.
186 */
tcp_lp_owd_calculator(struct sock * sk)187 static u32 tcp_lp_owd_calculator(struct sock *sk)
188 {
189 struct tcp_sock *tp = tcp_sk(sk);
190 struct lp *lp = inet_csk_ca(sk);
191 s64 owd = 0;
192
193 lp->remote_hz = tcp_lp_remote_hz_estimator(sk);
194
195 if (lp->flag & LP_VALID_RHZ) {
196 owd =
197 tp->rx_opt.rcv_tsval * (LP_RESOL / lp->remote_hz) -
198 tp->rx_opt.rcv_tsecr * (LP_RESOL / TCP_TS_HZ);
199 if (owd < 0)
200 owd = -owd;
201 }
202
203 if (owd > 0)
204 lp->flag |= LP_VALID_OWD;
205 else
206 lp->flag &= ~LP_VALID_OWD;
207
208 return owd;
209 }
210
211 /**
212 * tcp_lp_rtt_sample
213 *
214 * Implementation or rtt_sample.
215 * Will take the following action,
216 * 1. calc OWD,
217 * 2. record the min/max OWD,
218 * 3. calc smoothed OWD (SOWD).
219 * Most ideas come from the original TCP-LP implementation.
220 */
tcp_lp_rtt_sample(struct sock * sk,u32 rtt)221 static void tcp_lp_rtt_sample(struct sock *sk, u32 rtt)
222 {
223 struct lp *lp = inet_csk_ca(sk);
224 s64 mowd = tcp_lp_owd_calculator(sk);
225
226 /* sorry that we don't have valid data */
227 if (!(lp->flag & LP_VALID_RHZ) || !(lp->flag & LP_VALID_OWD))
228 return;
229
230 /* record the next min owd */
231 if (mowd < lp->owd_min)
232 lp->owd_min = mowd;
233
234 /* always forget the max of the max
235 * we just set owd_max as one below it */
236 if (mowd > lp->owd_max) {
237 if (mowd > lp->owd_max_rsv) {
238 if (lp->owd_max_rsv == 0)
239 lp->owd_max = mowd;
240 else
241 lp->owd_max = lp->owd_max_rsv;
242 lp->owd_max_rsv = mowd;
243 } else
244 lp->owd_max = mowd;
245 }
246
247 /* calc for smoothed owd */
248 if (lp->sowd != 0) {
249 mowd -= lp->sowd >> 3; /* m is now error in owd est */
250 lp->sowd += mowd; /* owd = 7/8 owd + 1/8 new */
251 } else
252 lp->sowd = mowd << 3; /* take the measured time be owd */
253 }
254
255 /**
256 * tcp_lp_pkts_acked
257 *
258 * Implementation of pkts_acked.
259 * Deal with active drop under Early Congestion Indication.
260 * Only drop to half and 1 will be handle, because we hope to use back
261 * newReno in increase case.
262 * We work it out by following the idea from TCP-LP's paper directly
263 */
tcp_lp_pkts_acked(struct sock * sk,const struct ack_sample * sample)264 static void tcp_lp_pkts_acked(struct sock *sk, const struct ack_sample *sample)
265 {
266 struct tcp_sock *tp = tcp_sk(sk);
267 struct lp *lp = inet_csk_ca(sk);
268 u32 now = tcp_time_stamp(tp);
269 u32 delta;
270
271 if (sample->rtt_us > 0)
272 tcp_lp_rtt_sample(sk, sample->rtt_us);
273
274 /* calc inference */
275 delta = now - tp->rx_opt.rcv_tsecr;
276 if ((s32)delta > 0)
277 lp->inference = 3 * delta;
278
279 /* test if within inference */
280 if (lp->last_drop && (now - lp->last_drop < lp->inference))
281 lp->flag |= LP_WITHIN_INF;
282 else
283 lp->flag &= ~LP_WITHIN_INF;
284
285 /* test if within threshold */
286 if (lp->sowd >> 3 <
287 lp->owd_min + 15 * (lp->owd_max - lp->owd_min) / 100)
288 lp->flag |= LP_WITHIN_THR;
289 else
290 lp->flag &= ~LP_WITHIN_THR;
291
292 pr_debug("TCP-LP: %05o|%5u|%5u|%15u|%15u|%15u\n", lp->flag,
293 tp->snd_cwnd, lp->remote_hz, lp->owd_min, lp->owd_max,
294 lp->sowd >> 3);
295
296 if (lp->flag & LP_WITHIN_THR)
297 return;
298
299 /* FIXME: try to reset owd_min and owd_max here
300 * so decrease the chance the min/max is no longer suitable
301 * and will usually within threshold when whithin inference */
302 lp->owd_min = lp->sowd >> 3;
303 lp->owd_max = lp->sowd >> 2;
304 lp->owd_max_rsv = lp->sowd >> 2;
305
306 /* happened within inference
307 * drop snd_cwnd into 1 */
308 if (lp->flag & LP_WITHIN_INF)
309 tp->snd_cwnd = 1U;
310
311 /* happened after inference
312 * cut snd_cwnd into half */
313 else
314 tp->snd_cwnd = max(tp->snd_cwnd >> 1U, 1U);
315
316 /* record this drop time */
317 lp->last_drop = now;
318 }
319
320 static struct tcp_congestion_ops tcp_lp __read_mostly = {
321 .init = tcp_lp_init,
322 .ssthresh = tcp_reno_ssthresh,
323 .undo_cwnd = tcp_reno_undo_cwnd,
324 .cong_avoid = tcp_lp_cong_avoid,
325 .pkts_acked = tcp_lp_pkts_acked,
326
327 .owner = THIS_MODULE,
328 .name = "lp"
329 };
330
tcp_lp_register(void)331 static int __init tcp_lp_register(void)
332 {
333 BUILD_BUG_ON(sizeof(struct lp) > ICSK_CA_PRIV_SIZE);
334 return tcp_register_congestion_control(&tcp_lp);
335 }
336
tcp_lp_unregister(void)337 static void __exit tcp_lp_unregister(void)
338 {
339 tcp_unregister_congestion_control(&tcp_lp);
340 }
341
342 module_init(tcp_lp_register);
343 module_exit(tcp_lp_unregister);
344
345 MODULE_AUTHOR("Wong Hoi Sing Edison, Hung Hing Lun Mike");
346 MODULE_LICENSE("GPL");
347 MODULE_DESCRIPTION("TCP Low Priority");
348