xref: /openbmc/linux/net/ipv4/tcp_lp.c (revision 9cfc5c90)
1 /*
2  * TCP Low Priority (TCP-LP)
3  *
4  * TCP Low Priority is a distributed algorithm whose goal is to utilize only
5  *   the excess network bandwidth as compared to the ``fair share`` of
6  *   bandwidth as targeted by TCP.
7  *
8  * As of 2.6.13, Linux supports pluggable congestion control algorithms.
9  * Due to the limitation of the API, we take the following changes from
10  * the original TCP-LP implementation:
11  *   o We use newReno in most core CA handling. Only add some checking
12  *     within cong_avoid.
13  *   o Error correcting in remote HZ, therefore remote HZ will be keeped
14  *     on checking and updating.
15  *   o Handling calculation of One-Way-Delay (OWD) within rtt_sample, since
16  *     OWD have a similar meaning as RTT. Also correct the buggy formular.
17  *   o Handle reaction for Early Congestion Indication (ECI) within
18  *     pkts_acked, as mentioned within pseudo code.
19  *   o OWD is handled in relative format, where local time stamp will in
20  *     tcp_time_stamp format.
21  *
22  * Original Author:
23  *   Aleksandar Kuzmanovic <akuzma@northwestern.edu>
24  * Available from:
25  *   http://www.ece.rice.edu/~akuzma/Doc/akuzma/TCP-LP.pdf
26  * Original implementation for 2.4.19:
27  *   http://www-ece.rice.edu/networks/TCP-LP/
28  *
29  * 2.6.x module Authors:
30  *   Wong Hoi Sing, Edison <hswong3i@gmail.com>
31  *   Hung Hing Lun, Mike <hlhung3i@gmail.com>
32  * SourceForge project page:
33  *   http://tcp-lp-mod.sourceforge.net/
34  */
35 
36 #include <linux/module.h>
37 #include <net/tcp.h>
38 
39 /* resolution of owd */
40 #define LP_RESOL       1000
41 
42 /**
43  * enum tcp_lp_state
44  * @LP_VALID_RHZ: is remote HZ valid?
45  * @LP_VALID_OWD: is OWD valid?
46  * @LP_WITHIN_THR: are we within threshold?
47  * @LP_WITHIN_INF: are we within inference?
48  *
49  * TCP-LP's state flags.
50  * We create this set of state flag mainly for debugging.
51  */
52 enum tcp_lp_state {
53 	LP_VALID_RHZ = (1 << 0),
54 	LP_VALID_OWD = (1 << 1),
55 	LP_WITHIN_THR = (1 << 3),
56 	LP_WITHIN_INF = (1 << 4),
57 };
58 
59 /**
60  * struct lp
61  * @flag: TCP-LP state flag
62  * @sowd: smoothed OWD << 3
63  * @owd_min: min OWD
64  * @owd_max: max OWD
65  * @owd_max_rsv: resrved max owd
66  * @remote_hz: estimated remote HZ
67  * @remote_ref_time: remote reference time
68  * @local_ref_time: local reference time
69  * @last_drop: time for last active drop
70  * @inference: current inference
71  *
72  * TCP-LP's private struct.
73  * We get the idea from original TCP-LP implementation where only left those we
74  * found are really useful.
75  */
76 struct lp {
77 	u32 flag;
78 	u32 sowd;
79 	u32 owd_min;
80 	u32 owd_max;
81 	u32 owd_max_rsv;
82 	u32 remote_hz;
83 	u32 remote_ref_time;
84 	u32 local_ref_time;
85 	u32 last_drop;
86 	u32 inference;
87 };
88 
89 /**
90  * tcp_lp_init
91  *
92  * Init all required variables.
93  * Clone the handling from Vegas module implementation.
94  */
95 static void tcp_lp_init(struct sock *sk)
96 {
97 	struct lp *lp = inet_csk_ca(sk);
98 
99 	lp->flag = 0;
100 	lp->sowd = 0;
101 	lp->owd_min = 0xffffffff;
102 	lp->owd_max = 0;
103 	lp->owd_max_rsv = 0;
104 	lp->remote_hz = 0;
105 	lp->remote_ref_time = 0;
106 	lp->local_ref_time = 0;
107 	lp->last_drop = 0;
108 	lp->inference = 0;
109 }
110 
111 /**
112  * tcp_lp_cong_avoid
113  *
114  * Implementation of cong_avoid.
115  * Will only call newReno CA when away from inference.
116  * From TCP-LP's paper, this will be handled in additive increasement.
117  */
118 static void tcp_lp_cong_avoid(struct sock *sk, u32 ack, u32 acked)
119 {
120 	struct lp *lp = inet_csk_ca(sk);
121 
122 	if (!(lp->flag & LP_WITHIN_INF))
123 		tcp_reno_cong_avoid(sk, ack, acked);
124 }
125 
126 /**
127  * tcp_lp_remote_hz_estimator
128  *
129  * Estimate remote HZ.
130  * We keep on updating the estimated value, where original TCP-LP
131  * implementation only guest it for once and use forever.
132  */
133 static u32 tcp_lp_remote_hz_estimator(struct sock *sk)
134 {
135 	struct tcp_sock *tp = tcp_sk(sk);
136 	struct lp *lp = inet_csk_ca(sk);
137 	s64 rhz = lp->remote_hz << 6;	/* remote HZ << 6 */
138 	s64 m = 0;
139 
140 	/* not yet record reference time
141 	 * go away!! record it before come back!! */
142 	if (lp->remote_ref_time == 0 || lp->local_ref_time == 0)
143 		goto out;
144 
145 	/* we can't calc remote HZ with no different!! */
146 	if (tp->rx_opt.rcv_tsval == lp->remote_ref_time ||
147 	    tp->rx_opt.rcv_tsecr == lp->local_ref_time)
148 		goto out;
149 
150 	m = HZ * (tp->rx_opt.rcv_tsval -
151 		  lp->remote_ref_time) / (tp->rx_opt.rcv_tsecr -
152 					  lp->local_ref_time);
153 	if (m < 0)
154 		m = -m;
155 
156 	if (rhz > 0) {
157 		m -= rhz >> 6;	/* m is now error in remote HZ est */
158 		rhz += m;	/* 63/64 old + 1/64 new */
159 	} else
160 		rhz = m << 6;
161 
162  out:
163 	/* record time for successful remote HZ calc */
164 	if ((rhz >> 6) > 0)
165 		lp->flag |= LP_VALID_RHZ;
166 	else
167 		lp->flag &= ~LP_VALID_RHZ;
168 
169 	/* record reference time stamp */
170 	lp->remote_ref_time = tp->rx_opt.rcv_tsval;
171 	lp->local_ref_time = tp->rx_opt.rcv_tsecr;
172 
173 	return rhz >> 6;
174 }
175 
176 /**
177  * tcp_lp_owd_calculator
178  *
179  * Calculate one way delay (in relative format).
180  * Original implement OWD as minus of remote time difference to local time
181  * difference directly. As this time difference just simply equal to RTT, when
182  * the network status is stable, remote RTT will equal to local RTT, and result
183  * OWD into zero.
184  * It seems to be a bug and so we fixed it.
185  */
186 static u32 tcp_lp_owd_calculator(struct sock *sk)
187 {
188 	struct tcp_sock *tp = tcp_sk(sk);
189 	struct lp *lp = inet_csk_ca(sk);
190 	s64 owd = 0;
191 
192 	lp->remote_hz = tcp_lp_remote_hz_estimator(sk);
193 
194 	if (lp->flag & LP_VALID_RHZ) {
195 		owd =
196 		    tp->rx_opt.rcv_tsval * (LP_RESOL / lp->remote_hz) -
197 		    tp->rx_opt.rcv_tsecr * (LP_RESOL / HZ);
198 		if (owd < 0)
199 			owd = -owd;
200 	}
201 
202 	if (owd > 0)
203 		lp->flag |= LP_VALID_OWD;
204 	else
205 		lp->flag &= ~LP_VALID_OWD;
206 
207 	return owd;
208 }
209 
210 /**
211  * tcp_lp_rtt_sample
212  *
213  * Implementation or rtt_sample.
214  * Will take the following action,
215  *   1. calc OWD,
216  *   2. record the min/max OWD,
217  *   3. calc smoothed OWD (SOWD).
218  * Most ideas come from the original TCP-LP implementation.
219  */
220 static void tcp_lp_rtt_sample(struct sock *sk, u32 rtt)
221 {
222 	struct lp *lp = inet_csk_ca(sk);
223 	s64 mowd = tcp_lp_owd_calculator(sk);
224 
225 	/* sorry that we don't have valid data */
226 	if (!(lp->flag & LP_VALID_RHZ) || !(lp->flag & LP_VALID_OWD))
227 		return;
228 
229 	/* record the next min owd */
230 	if (mowd < lp->owd_min)
231 		lp->owd_min = mowd;
232 
233 	/* always forget the max of the max
234 	 * we just set owd_max as one below it */
235 	if (mowd > lp->owd_max) {
236 		if (mowd > lp->owd_max_rsv) {
237 			if (lp->owd_max_rsv == 0)
238 				lp->owd_max = mowd;
239 			else
240 				lp->owd_max = lp->owd_max_rsv;
241 			lp->owd_max_rsv = mowd;
242 		} else
243 			lp->owd_max = mowd;
244 	}
245 
246 	/* calc for smoothed owd */
247 	if (lp->sowd != 0) {
248 		mowd -= lp->sowd >> 3;	/* m is now error in owd est */
249 		lp->sowd += mowd;	/* owd = 7/8 owd + 1/8 new */
250 	} else
251 		lp->sowd = mowd << 3;	/* take the measured time be owd */
252 }
253 
254 /**
255  * tcp_lp_pkts_acked
256  *
257  * Implementation of pkts_acked.
258  * Deal with active drop under Early Congestion Indication.
259  * Only drop to half and 1 will be handle, because we hope to use back
260  * newReno in increase case.
261  * We work it out by following the idea from TCP-LP's paper directly
262  */
263 static void tcp_lp_pkts_acked(struct sock *sk, u32 num_acked, s32 rtt_us)
264 {
265 	struct tcp_sock *tp = tcp_sk(sk);
266 	struct lp *lp = inet_csk_ca(sk);
267 
268 	if (rtt_us > 0)
269 		tcp_lp_rtt_sample(sk, rtt_us);
270 
271 	/* calc inference */
272 	if (tcp_time_stamp > tp->rx_opt.rcv_tsecr)
273 		lp->inference = 3 * (tcp_time_stamp - tp->rx_opt.rcv_tsecr);
274 
275 	/* test if within inference */
276 	if (lp->last_drop && (tcp_time_stamp - lp->last_drop < lp->inference))
277 		lp->flag |= LP_WITHIN_INF;
278 	else
279 		lp->flag &= ~LP_WITHIN_INF;
280 
281 	/* test if within threshold */
282 	if (lp->sowd >> 3 <
283 	    lp->owd_min + 15 * (lp->owd_max - lp->owd_min) / 100)
284 		lp->flag |= LP_WITHIN_THR;
285 	else
286 		lp->flag &= ~LP_WITHIN_THR;
287 
288 	pr_debug("TCP-LP: %05o|%5u|%5u|%15u|%15u|%15u\n", lp->flag,
289 		 tp->snd_cwnd, lp->remote_hz, lp->owd_min, lp->owd_max,
290 		 lp->sowd >> 3);
291 
292 	if (lp->flag & LP_WITHIN_THR)
293 		return;
294 
295 	/* FIXME: try to reset owd_min and owd_max here
296 	 * so decrease the chance the min/max is no longer suitable
297 	 * and will usually within threshold when whithin inference */
298 	lp->owd_min = lp->sowd >> 3;
299 	lp->owd_max = lp->sowd >> 2;
300 	lp->owd_max_rsv = lp->sowd >> 2;
301 
302 	/* happened within inference
303 	 * drop snd_cwnd into 1 */
304 	if (lp->flag & LP_WITHIN_INF)
305 		tp->snd_cwnd = 1U;
306 
307 	/* happened after inference
308 	 * cut snd_cwnd into half */
309 	else
310 		tp->snd_cwnd = max(tp->snd_cwnd >> 1U, 1U);
311 
312 	/* record this drop time */
313 	lp->last_drop = tcp_time_stamp;
314 }
315 
316 static struct tcp_congestion_ops tcp_lp __read_mostly = {
317 	.init = tcp_lp_init,
318 	.ssthresh = tcp_reno_ssthresh,
319 	.cong_avoid = tcp_lp_cong_avoid,
320 	.pkts_acked = tcp_lp_pkts_acked,
321 
322 	.owner = THIS_MODULE,
323 	.name = "lp"
324 };
325 
326 static int __init tcp_lp_register(void)
327 {
328 	BUILD_BUG_ON(sizeof(struct lp) > ICSK_CA_PRIV_SIZE);
329 	return tcp_register_congestion_control(&tcp_lp);
330 }
331 
332 static void __exit tcp_lp_unregister(void)
333 {
334 	tcp_unregister_congestion_control(&tcp_lp);
335 }
336 
337 module_init(tcp_lp_register);
338 module_exit(tcp_lp_unregister);
339 
340 MODULE_AUTHOR("Wong Hoi Sing Edison, Hung Hing Lun Mike");
341 MODULE_LICENSE("GPL");
342 MODULE_DESCRIPTION("TCP Low Priority");
343