2 * TCP Low Priority (TCP-LP)
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.
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
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.
23 * Aleksandar Kuzmanovic <akuzma@northwestern.edu>
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/
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/
36 #include <linux/module.h>
39 /* resolution of owd */
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?
49 * TCP-LP's state flags.
50 * We create this set of state flag mainly for debugging.
53 LP_VALID_RHZ
= (1 << 0),
54 LP_VALID_OWD
= (1 << 1),
55 LP_WITHIN_THR
= (1 << 3),
56 LP_WITHIN_INF
= (1 << 4),
61 * @flag: TCP-LP state flag
62 * @sowd: smoothed OWD << 3
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
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.
92 * Init all required variables.
93 * Clone the handling from Vegas module implementation.
95 static void tcp_lp_init(struct sock
*sk
)
97 struct lp
*lp
= inet_csk_ca(sk
);
101 lp
->owd_min
= 0xffffffff;
105 lp
->remote_ref_time
= 0;
106 lp
->local_ref_time
= 0;
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.
118 static void tcp_lp_cong_avoid(struct sock
*sk
, u32 ack
, u32 acked
,
121 struct lp
*lp
= inet_csk_ca(sk
);
123 if (!(lp
->flag
& LP_WITHIN_INF
))
124 tcp_reno_cong_avoid(sk
, ack
, acked
, in_flight
);
128 * tcp_lp_remote_hz_estimator
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.
134 static u32
tcp_lp_remote_hz_estimator(struct sock
*sk
)
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 */
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)
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
)
151 m
= HZ
* (tp
->rx_opt
.rcv_tsval
-
152 lp
->remote_ref_time
) / (tp
->rx_opt
.rcv_tsecr
-
158 m
-= rhz
>> 6; /* m is now error in remote HZ est */
159 rhz
+= m
; /* 63/64 old + 1/64 new */
164 /* record time for successful remote HZ calc */
166 lp
->flag
|= LP_VALID_RHZ
;
168 lp
->flag
&= ~LP_VALID_RHZ
;
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
;
178 * tcp_lp_owd_calculator
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
185 * It seems to be a bug and so we fixed it.
187 static u32
tcp_lp_owd_calculator(struct sock
*sk
)
189 struct tcp_sock
*tp
= tcp_sk(sk
);
190 struct lp
*lp
= inet_csk_ca(sk
);
193 lp
->remote_hz
= tcp_lp_remote_hz_estimator(sk
);
195 if (lp
->flag
& LP_VALID_RHZ
) {
197 tp
->rx_opt
.rcv_tsval
* (LP_RESOL
/ lp
->remote_hz
) -
198 tp
->rx_opt
.rcv_tsecr
* (LP_RESOL
/ HZ
);
204 lp
->flag
|= LP_VALID_OWD
;
206 lp
->flag
&= ~LP_VALID_OWD
;
214 * Implementation or rtt_sample.
215 * Will take the following action,
217 * 2. record the min/max OWD,
218 * 3. calc smoothed OWD (SOWD).
219 * Most ideas come from the original TCP-LP implementation.
221 static void tcp_lp_rtt_sample(struct sock
*sk
, u32 rtt
)
223 struct lp
*lp
= inet_csk_ca(sk
);
224 s64 mowd
= tcp_lp_owd_calculator(sk
);
226 /* sorry that we don't have valid data */
227 if (!(lp
->flag
& LP_VALID_RHZ
) || !(lp
->flag
& LP_VALID_OWD
))
230 /* record the next min owd */
231 if (mowd
< lp
->owd_min
)
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)
241 lp
->owd_max
= lp
->owd_max_rsv
;
242 lp
->owd_max_rsv
= mowd
;
247 /* calc for smoothed owd */
249 mowd
-= lp
->sowd
>> 3; /* m is now error in owd est */
250 lp
->sowd
+= mowd
; /* owd = 7/8 owd + 1/8 new */
252 lp
->sowd
= mowd
<< 3; /* take the measured time be owd */
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
264 static void tcp_lp_pkts_acked(struct sock
*sk
, u32 num_acked
, s32 rtt_us
)
266 struct tcp_sock
*tp
= tcp_sk(sk
);
267 struct lp
*lp
= inet_csk_ca(sk
);
270 tcp_lp_rtt_sample(sk
, rtt_us
);
273 if (tcp_time_stamp
> tp
->rx_opt
.rcv_tsecr
)
274 lp
->inference
= 3 * (tcp_time_stamp
- tp
->rx_opt
.rcv_tsecr
);
276 /* test if within inference */
277 if (lp
->last_drop
&& (tcp_time_stamp
- lp
->last_drop
< lp
->inference
))
278 lp
->flag
|= LP_WITHIN_INF
;
280 lp
->flag
&= ~LP_WITHIN_INF
;
282 /* test if within threshold */
284 lp
->owd_min
+ 15 * (lp
->owd_max
- lp
->owd_min
) / 100)
285 lp
->flag
|= LP_WITHIN_THR
;
287 lp
->flag
&= ~LP_WITHIN_THR
;
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
,
293 if (lp
->flag
& LP_WITHIN_THR
)
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;
303 /* happened within inference
304 * drop snd_cwnd into 1 */
305 if (lp
->flag
& LP_WITHIN_INF
)
308 /* happened after inference
309 * cut snd_cwnd into half */
311 tp
->snd_cwnd
= max(tp
->snd_cwnd
>> 1U, 1U);
313 /* record this drop time */
314 lp
->last_drop
= tcp_time_stamp
;
317 static struct tcp_congestion_ops tcp_lp __read_mostly
= {
318 .flags
= TCP_CONG_RTT_STAMP
,
320 .ssthresh
= tcp_reno_ssthresh
,
321 .cong_avoid
= tcp_lp_cong_avoid
,
322 .min_cwnd
= tcp_reno_min_cwnd
,
323 .pkts_acked
= tcp_lp_pkts_acked
,
325 .owner
= THIS_MODULE
,
329 static int __init
tcp_lp_register(void)
331 BUILD_BUG_ON(sizeof(struct lp
) > ICSK_CA_PRIV_SIZE
);
332 return tcp_register_congestion_control(&tcp_lp
);
335 static void __exit
tcp_lp_unregister(void)
337 tcp_unregister_congestion_control(&tcp_lp
);
340 module_init(tcp_lp_register
);
341 module_exit(tcp_lp_unregister
);
343 MODULE_AUTHOR("Wong Hoi Sing Edison, Hung Hing Lun Mike");
344 MODULE_LICENSE("GPL");
345 MODULE_DESCRIPTION("TCP Low Priority");