blob: 991d62a2f9bb3abb1d155d075c202c1ea558ca7f [file] [log] [blame]
Wong Hoi Sing Edison7c106d72006-06-05 17:27:58 -07001/*
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
Wong Hoi Sing Edison3795da42006-09-13 20:30:30 -07006 * bandwidth as targeted by TCP.
Wong Hoi Sing Edison7c106d72006-06-05 17:27:58 -07007 *
Wong Hoi Sing Edison7c106d72006-06-05 17:27:58 -07008 * 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.
Lucas De Marchi25985ed2011-03-30 22:57:33 -030015 * o Handling calculation of One-Way-Delay (OWD) within rtt_sample, since
Wong Hoi Sing Edison7c106d72006-06-05 17:27:58 -070016 * 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 *
Wong Hoi Sing Edison3795da42006-09-13 20:30:30 -070022 * 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/
Wong Hoi Sing Edison7c106d72006-06-05 17:27:58 -070028 *
Wong Hoi Sing Edison3795da42006-09-13 20:30:30 -070029 * 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/
Wong Hoi Sing Edison7c106d72006-06-05 17:27:58 -070034 */
35
Wong Hoi Sing Edison7c106d72006-06-05 17:27:58 -070036#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 */
52enum 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 */
76struct 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 */
95static 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 */
Yuchung Cheng9f9843a72013-10-31 11:07:31 -0700118static void tcp_lp_cong_avoid(struct sock *sk, u32 ack, u32 acked,
119 u32 in_flight)
Wong Hoi Sing Edison7c106d72006-06-05 17:27:58 -0700120{
121 struct lp *lp = inet_csk_ca(sk);
122
123 if (!(lp->flag & LP_WITHIN_INF))
Yuchung Cheng9f9843a72013-10-31 11:07:31 -0700124 tcp_reno_cong_avoid(sk, ack, acked, in_flight);
Wong Hoi Sing Edison7c106d72006-06-05 17:27:58 -0700125}
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 */
134static 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!! */
Joe Perches9d4fb272009-11-23 10:41:23 -0800147 if (tp->rx_opt.rcv_tsval == lp->remote_ref_time ||
148 tp->rx_opt.rcv_tsecr == lp->local_ref_time)
Wong Hoi Sing Edison7c106d72006-06-05 17:27:58 -0700149 goto out;
150
151 m = HZ * (tp->rx_opt.rcv_tsval -
152 lp->remote_ref_time) / (tp->rx_opt.rcv_tsecr -
153 lp->local_ref_time);
154 if (m < 0)
155 m = -m;
156
Wong Hoi Sing Edison3795da42006-09-13 20:30:30 -0700157 if (rhz > 0) {
Wong Hoi Sing Edison7c106d72006-06-05 17:27:58 -0700158 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
Wong Hoi Sing Edison7c106d72006-06-05 17:27:58 -0700163 out:
Wong Hoi Sing Edison3795da42006-09-13 20:30:30 -0700164 /* record time for successful remote HZ calc */
Wong Hoi Sing Edisonbfbea8a2006-09-28 14:40:21 -0700165 if ((rhz >> 6) > 0)
Wong Hoi Sing Edison3795da42006-09-13 20:30:30 -0700166 lp->flag |= LP_VALID_RHZ;
167 else
168 lp->flag &= ~LP_VALID_RHZ;
169
Wong Hoi Sing Edison7c106d72006-06-05 17:27:58 -0700170 /* 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 */
187static 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 / 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 */
Stephen Hemminger164891a2007-04-23 22:26:16 -0700221static void tcp_lp_rtt_sample(struct sock *sk, u32 rtt)
Wong Hoi Sing Edison7c106d72006-06-05 17:27:58 -0700222{
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 */
Stephen Hemminger30cfd0b2007-07-25 23:49:34 -0700264static void tcp_lp_pkts_acked(struct sock *sk, u32 num_acked, s32 rtt_us)
Wong Hoi Sing Edison7c106d72006-06-05 17:27:58 -0700265{
266 struct tcp_sock *tp = tcp_sk(sk);
267 struct lp *lp = inet_csk_ca(sk);
268
Stephen Hemminger30cfd0b2007-07-25 23:49:34 -0700269 if (rtt_us > 0)
270 tcp_lp_rtt_sample(sk, rtt_us);
Stephen Hemminger164891a2007-04-23 22:26:16 -0700271
Wong Hoi Sing Edison7c106d72006-06-05 17:27:58 -0700272 /* calc inference */
273 if (tcp_time_stamp > tp->rx_opt.rcv_tsecr)
274 lp->inference = 3 * (tcp_time_stamp - tp->rx_opt.rcv_tsecr);
275
276 /* test if within inference */
277 if (lp->last_drop && (tcp_time_stamp - lp->last_drop < lp->inference))
278 lp->flag |= LP_WITHIN_INF;
279 else
280 lp->flag &= ~LP_WITHIN_INF;
281
282 /* test if within threshold */
283 if (lp->sowd >> 3 <
284 lp->owd_min + 15 * (lp->owd_max - lp->owd_min) / 100)
285 lp->flag |= LP_WITHIN_THR;
286 else
287 lp->flag &= ~LP_WITHIN_THR;
288
289 pr_debug("TCP-LP: %05o|%5u|%5u|%15u|%15u|%15u\n", lp->flag,
290 tp->snd_cwnd, lp->remote_hz, lp->owd_min, lp->owd_max,
291 lp->sowd >> 3);
292
293 if (lp->flag & LP_WITHIN_THR)
294 return;
295
296 /* FIXME: try to reset owd_min and owd_max here
297 * so decrease the chance the min/max is no longer suitable
298 * and will usually within threshold when whithin inference */
299 lp->owd_min = lp->sowd >> 3;
300 lp->owd_max = lp->sowd >> 2;
301 lp->owd_max_rsv = lp->sowd >> 2;
302
303 /* happened within inference
304 * drop snd_cwnd into 1 */
305 if (lp->flag & LP_WITHIN_INF)
306 tp->snd_cwnd = 1U;
307
308 /* happened after inference
309 * cut snd_cwnd into half */
310 else
311 tp->snd_cwnd = max(tp->snd_cwnd >> 1U, 1U);
312
313 /* record this drop time */
314 lp->last_drop = tcp_time_stamp;
315}
316
Stephen Hemmingera252beb2011-03-10 00:40:17 -0800317static struct tcp_congestion_ops tcp_lp __read_mostly = {
Stephen Hemminger164891a2007-04-23 22:26:16 -0700318 .flags = TCP_CONG_RTT_STAMP,
Wong Hoi Sing Edison7c106d72006-06-05 17:27:58 -0700319 .init = tcp_lp_init,
320 .ssthresh = tcp_reno_ssthresh,
321 .cong_avoid = tcp_lp_cong_avoid,
322 .min_cwnd = tcp_reno_min_cwnd,
Wong Hoi Sing Edison7c106d72006-06-05 17:27:58 -0700323 .pkts_acked = tcp_lp_pkts_acked,
324
325 .owner = THIS_MODULE,
326 .name = "lp"
327};
328
329static int __init tcp_lp_register(void)
330{
Alexey Dobriyan29896972006-08-25 00:37:24 -0700331 BUILD_BUG_ON(sizeof(struct lp) > ICSK_CA_PRIV_SIZE);
Wong Hoi Sing Edison7c106d72006-06-05 17:27:58 -0700332 return tcp_register_congestion_control(&tcp_lp);
333}
334
335static void __exit tcp_lp_unregister(void)
336{
337 tcp_unregister_congestion_control(&tcp_lp);
338}
339
340module_init(tcp_lp_register);
341module_exit(tcp_lp_unregister);
342
Wong Hoi Sing Edison3795da42006-09-13 20:30:30 -0700343MODULE_AUTHOR("Wong Hoi Sing Edison, Hung Hing Lun Mike");
Wong Hoi Sing Edison7c106d72006-06-05 17:27:58 -0700344MODULE_LICENSE("GPL");
345MODULE_DESCRIPTION("TCP Low Priority");