aboutsummaryrefslogtreecommitdiffstats
diff options
context:
space:
mode:
authorWong Hoi Sing Edison <hswong3i@gmail.com>2006-06-05 20:27:58 -0400
committerDavid S. Miller <davem@sunset.davemloft.net>2006-06-18 00:29:21 -0400
commit7c106d7e782bd4805f39da30e81018f861b4b8c5 (patch)
treea0bf2f8820c599151ee951a412e86039da3a11ba
parent2f45c340e09242641d4f11498c3be48b35abb926 (diff)
[TCP]: TCP Low Priority congestion control
TCP Low Priority is a distributed algorithm whose goal is to utilize only the excess network bandwidth as compared to the ``fair share`` of bandwidth as targeted by TCP. Available from: http://www.ece.rice.edu/~akuzma/Doc/akuzma/TCP-LP.pdf Original Author: Aleksandar Kuzmanovic <akuzma@northwestern.edu> See http://www-ece.rice.edu/networks/TCP-LP/ for their implementation. As of 2.6.13, Linux supports pluggable congestion control algorithms. Due to the limitation of the API, we take the following changes from the original TCP-LP implementation: o We use newReno in most core CA handling. Only add some checking within cong_avoid. o Error correcting in remote HZ, therefore remote HZ will be keeped on checking and updating. o Handling calculation of One-Way-Delay (OWD) within rtt_sample, sicne OWD have a similar meaning as RTT. Also correct the buggy formular. o Handle reaction for Early Congestion Indication (ECI) within pkts_acked, as mentioned within pseudo code. o OWD is handled in relative format, where local time stamp will in tcp_time_stamp format. Port from 2.4.19 to 2.6.16 as module by: Wong Hoi Sing Edison <hswong3i@gmail.com> Hung Hing Lun <hlhung3i@gmail.com> Signed-off-by: Wong Hoi Sing Edison <hswong3i@gmail.com> Signed-off-by: Stephen Hemminger <shemminger@osdl.org> Signed-off-by: David S. Miller <davem@davemloft.net>
-rw-r--r--net/ipv4/Kconfig10
-rw-r--r--net/ipv4/Makefile1
-rw-r--r--net/ipv4/tcp_lp.c338
3 files changed, 349 insertions, 0 deletions
diff --git a/net/ipv4/Kconfig b/net/ipv4/Kconfig
index 35eb70b1448d..ae85149a0434 100644
--- a/net/ipv4/Kconfig
+++ b/net/ipv4/Kconfig
@@ -550,6 +550,16 @@ config TCP_CONG_SCALABLE
550 properties, though is known to have fairness issues. 550 properties, though is known to have fairness issues.
551 See http://www-lce.eng.cam.ac.uk/~ctk21/scalable/ 551 See http://www-lce.eng.cam.ac.uk/~ctk21/scalable/
552 552
553config TCP_CONG_LP
554 tristate "TCP Low Priority"
555 depends on EXPERIMENTAL
556 default n
557 ---help---
558 TCP Low Priority (TCP-LP), a distributed algorithm whose goal is
559 to utiliza only the excess network bandwidth as compared to the
560 ``fair share`` of bandwidth as targeted by TCP.
561 See http://www-ece.rice.edu/networks/TCP-LP/
562
553endmenu 563endmenu
554 564
555config TCP_CONG_BIC 565config TCP_CONG_BIC
diff --git a/net/ipv4/Makefile b/net/ipv4/Makefile
index 4cc94482eacc..58a7a4634f70 100644
--- a/net/ipv4/Makefile
+++ b/net/ipv4/Makefile
@@ -44,6 +44,7 @@ obj-$(CONFIG_TCP_CONG_HYBLA) += tcp_hybla.o
44obj-$(CONFIG_TCP_CONG_HTCP) += tcp_htcp.o 44obj-$(CONFIG_TCP_CONG_HTCP) += tcp_htcp.o
45obj-$(CONFIG_TCP_CONG_VEGAS) += tcp_vegas.o 45obj-$(CONFIG_TCP_CONG_VEGAS) += tcp_vegas.o
46obj-$(CONFIG_TCP_CONG_SCALABLE) += tcp_scalable.o 46obj-$(CONFIG_TCP_CONG_SCALABLE) += tcp_scalable.o
47obj-$(CONFIG_TCP_CONG_LP) += tcp_lp.o
47 48
48obj-$(CONFIG_XFRM) += xfrm4_policy.o xfrm4_state.o xfrm4_input.o \ 49obj-$(CONFIG_XFRM) += xfrm4_policy.o xfrm4_state.o xfrm4_input.o \
49 xfrm4_output.o 50 xfrm4_output.o
diff --git a/net/ipv4/tcp_lp.c b/net/ipv4/tcp_lp.c
new file mode 100644
index 000000000000..1f977b6ee9a1
--- /dev/null
+++ b/net/ipv4/tcp_lp.c
@@ -0,0 +1,338 @@
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. Available from:
7 * http://www.ece.rice.edu/~akuzma/Doc/akuzma/TCP-LP.pdf
8 *
9 * Original Author:
10 * Aleksandar Kuzmanovic <akuzma@northwestern.edu>
11 *
12 * See http://www-ece.rice.edu/networks/TCP-LP/ for their implementation.
13 * As of 2.6.13, Linux supports pluggable congestion control algorithms.
14 * Due to the limitation of the API, we take the following changes from
15 * the original TCP-LP implementation:
16 * o We use newReno in most core CA handling. Only add some checking
17 * within cong_avoid.
18 * o Error correcting in remote HZ, therefore remote HZ will be keeped
19 * on checking and updating.
20 * o Handling calculation of One-Way-Delay (OWD) within rtt_sample, sicne
21 * OWD have a similar meaning as RTT. Also correct the buggy formular.
22 * o Handle reaction for Early Congestion Indication (ECI) within
23 * pkts_acked, as mentioned within pseudo code.
24 * o OWD is handled in relative format, where local time stamp will in
25 * tcp_time_stamp format.
26 *
27 * Port from 2.4.19 to 2.6.16 as module by:
28 * Wong Hoi Sing Edison <hswong3i@gmail.com>
29 * Hung Hing Lun <hlhung3i@gmail.com>
30 *
31 * Version: $Id: tcp_lp.c,v 1.22 2006-05-02 18:18:19 hswong3i Exp $
32 */
33
34#include <linux/config.h>
35#include <linux/module.h>
36#include <net/tcp.h>
37
38/* resolution of owd */
39#define LP_RESOL 1000
40
41/**
42 * enum tcp_lp_state
43 * @LP_VALID_RHZ: is remote HZ valid?
44 * @LP_VALID_OWD: is OWD valid?
45 * @LP_WITHIN_THR: are we within threshold?
46 * @LP_WITHIN_INF: are we within inference?
47 *
48 * TCP-LP's state flags.
49 * We create this set of state flag mainly for debugging.
50 */
51enum tcp_lp_state {
52 LP_VALID_RHZ = (1 << 0),
53 LP_VALID_OWD = (1 << 1),
54 LP_WITHIN_THR = (1 << 3),
55 LP_WITHIN_INF = (1 << 4),
56};
57
58/**
59 * struct lp
60 * @flag: TCP-LP state flag
61 * @sowd: smoothed OWD << 3
62 * @owd_min: min OWD
63 * @owd_max: max OWD
64 * @owd_max_rsv: resrved max owd
65 * @remote_hz: estimated remote HZ
66 * @remote_ref_time: remote reference time
67 * @local_ref_time: local reference time
68 * @last_drop: time for last active drop
69 * @inference: current inference
70 *
71 * TCP-LP's private struct.
72 * We get the idea from original TCP-LP implementation where only left those we
73 * found are really useful.
74 */
75struct lp {
76 u32 flag;
77 u32 sowd;
78 u32 owd_min;
79 u32 owd_max;
80 u32 owd_max_rsv;
81 u32 remote_hz;
82 u32 remote_ref_time;
83 u32 local_ref_time;
84 u32 last_drop;
85 u32 inference;
86};
87
88/**
89 * tcp_lp_init
90 *
91 * Init all required variables.
92 * Clone the handling from Vegas module implementation.
93 */
94static void tcp_lp_init(struct sock *sk)
95{
96 struct lp *lp = inet_csk_ca(sk);
97
98 lp->flag = 0;
99 lp->sowd = 0;
100 lp->owd_min = 0xffffffff;
101 lp->owd_max = 0;
102 lp->owd_max_rsv = 0;
103 lp->remote_hz = 0;
104 lp->remote_ref_time = 0;
105 lp->local_ref_time = 0;
106 lp->last_drop = 0;
107 lp->inference = 0;
108}
109
110/**
111 * tcp_lp_cong_avoid
112 *
113 * Implementation of cong_avoid.
114 * Will only call newReno CA when away from inference.
115 * From TCP-LP's paper, this will be handled in additive increasement.
116 */
117static void tcp_lp_cong_avoid(struct sock *sk, u32 ack, u32 rtt, u32 in_flight,
118 int flag)
119{
120 struct lp *lp = inet_csk_ca(sk);
121
122 if (!(lp->flag & LP_WITHIN_INF))
123 tcp_reno_cong_avoid(sk, ack, rtt, in_flight, flag);
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 */
133static 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 /* record time for successful remote HZ calc */
163 lp->flag |= LP_VALID_RHZ;
164
165 out:
166 /* record reference time stamp */
167 lp->remote_ref_time = tp->rx_opt.rcv_tsval;
168 lp->local_ref_time = tp->rx_opt.rcv_tsecr;
169
170 return rhz >> 6;
171}
172
173/**
174 * tcp_lp_owd_calculator
175 *
176 * Calculate one way delay (in relative format).
177 * Original implement OWD as minus of remote time difference to local time
178 * difference directly. As this time difference just simply equal to RTT, when
179 * the network status is stable, remote RTT will equal to local RTT, and result
180 * OWD into zero.
181 * It seems to be a bug and so we fixed it.
182 */
183static u32 tcp_lp_owd_calculator(struct sock *sk)
184{
185 struct tcp_sock *tp = tcp_sk(sk);
186 struct lp *lp = inet_csk_ca(sk);
187 s64 owd = 0;
188
189 lp->remote_hz = tcp_lp_remote_hz_estimator(sk);
190
191 if (lp->flag & LP_VALID_RHZ) {
192 owd =
193 tp->rx_opt.rcv_tsval * (LP_RESOL / lp->remote_hz) -
194 tp->rx_opt.rcv_tsecr * (LP_RESOL / HZ);
195 if (owd < 0)
196 owd = -owd;
197 }
198
199 if (owd > 0)
200 lp->flag |= LP_VALID_OWD;
201 else
202 lp->flag &= ~LP_VALID_OWD;
203
204 return owd;
205}
206
207/**
208 * tcp_lp_rtt_sample
209 *
210 * Implementation or rtt_sample.
211 * Will take the following action,
212 * 1. calc OWD,
213 * 2. record the min/max OWD,
214 * 3. calc smoothed OWD (SOWD).
215 * Most ideas come from the original TCP-LP implementation.
216 */
217static void tcp_lp_rtt_sample(struct sock *sk, u32 usrtt)
218{
219 struct lp *lp = inet_csk_ca(sk);
220 s64 mowd = tcp_lp_owd_calculator(sk);
221
222 /* sorry that we don't have valid data */
223 if (!(lp->flag & LP_VALID_RHZ) || !(lp->flag & LP_VALID_OWD))
224 return;
225
226 /* record the next min owd */
227 if (mowd < lp->owd_min)
228 lp->owd_min = mowd;
229
230 /* always forget the max of the max
231 * we just set owd_max as one below it */
232 if (mowd > lp->owd_max) {
233 if (mowd > lp->owd_max_rsv) {
234 if (lp->owd_max_rsv == 0)
235 lp->owd_max = mowd;
236 else
237 lp->owd_max = lp->owd_max_rsv;
238 lp->owd_max_rsv = mowd;
239 } else
240 lp->owd_max = mowd;
241 }
242
243 /* calc for smoothed owd */
244 if (lp->sowd != 0) {
245 mowd -= lp->sowd >> 3; /* m is now error in owd est */
246 lp->sowd += mowd; /* owd = 7/8 owd + 1/8 new */
247 } else
248 lp->sowd = mowd << 3; /* take the measured time be owd */
249}
250
251/**
252 * tcp_lp_pkts_acked
253 *
254 * Implementation of pkts_acked.
255 * Deal with active drop under Early Congestion Indication.
256 * Only drop to half and 1 will be handle, because we hope to use back
257 * newReno in increase case.
258 * We work it out by following the idea from TCP-LP's paper directly
259 */
260static void tcp_lp_pkts_acked(struct sock *sk, u32 num_acked)
261{
262 struct tcp_sock *tp = tcp_sk(sk);
263 struct lp *lp = inet_csk_ca(sk);
264
265 /* calc inference */
266 if (tcp_time_stamp > tp->rx_opt.rcv_tsecr)
267 lp->inference = 3 * (tcp_time_stamp - tp->rx_opt.rcv_tsecr);
268
269 /* test if within inference */
270 if (lp->last_drop && (tcp_time_stamp - lp->last_drop < lp->inference))
271 lp->flag |= LP_WITHIN_INF;
272 else
273 lp->flag &= ~LP_WITHIN_INF;
274
275 /* test if within threshold */
276 if (lp->sowd >> 3 <
277 lp->owd_min + 15 * (lp->owd_max - lp->owd_min) / 100)
278 lp->flag |= LP_WITHIN_THR;
279 else
280 lp->flag &= ~LP_WITHIN_THR;
281
282 pr_debug("TCP-LP: %05o|%5u|%5u|%15u|%15u|%15u\n", lp->flag,
283 tp->snd_cwnd, lp->remote_hz, lp->owd_min, lp->owd_max,
284 lp->sowd >> 3);
285
286 if (lp->flag & LP_WITHIN_THR)
287 return;
288
289 /* FIXME: try to reset owd_min and owd_max here
290 * so decrease the chance the min/max is no longer suitable
291 * and will usually within threshold when whithin inference */
292 lp->owd_min = lp->sowd >> 3;
293 lp->owd_max = lp->sowd >> 2;
294 lp->owd_max_rsv = lp->sowd >> 2;
295
296 /* happened within inference
297 * drop snd_cwnd into 1 */
298 if (lp->flag & LP_WITHIN_INF)
299 tp->snd_cwnd = 1U;
300
301 /* happened after inference
302 * cut snd_cwnd into half */
303 else
304 tp->snd_cwnd = max(tp->snd_cwnd >> 1U, 1U);
305
306 /* record this drop time */
307 lp->last_drop = tcp_time_stamp;
308}
309
310static struct tcp_congestion_ops tcp_lp = {
311 .init = tcp_lp_init,
312 .ssthresh = tcp_reno_ssthresh,
313 .cong_avoid = tcp_lp_cong_avoid,
314 .min_cwnd = tcp_reno_min_cwnd,
315 .rtt_sample = tcp_lp_rtt_sample,
316 .pkts_acked = tcp_lp_pkts_acked,
317
318 .owner = THIS_MODULE,
319 .name = "lp"
320};
321
322static int __init tcp_lp_register(void)
323{
324 BUG_ON(sizeof(struct lp) > ICSK_CA_PRIV_SIZE);
325 return tcp_register_congestion_control(&tcp_lp);
326}
327
328static void __exit tcp_lp_unregister(void)
329{
330 tcp_unregister_congestion_control(&tcp_lp);
331}
332
333module_init(tcp_lp_register);
334module_exit(tcp_lp_unregister);
335
336MODULE_AUTHOR("Wong Hoi Sing Edison, Hung Hing Lun");
337MODULE_LICENSE("GPL");
338MODULE_DESCRIPTION("TCP Low Priority");