1/*
2 * TCP Westwood+: end-to-end bandwidth estimation for TCP
3 *
4 * Angelo Dell'Aera: author of the first version of TCP Westwood+ in Linux 2.4
5 *
6 * Support at http://c3lab.poliba.it/index.php/Westwood
7 * Main references in literature:
8 *
9 * - Mascolo S, Casetti, M. Gerla et al.
10 * "TCP Westwood: bandwidth estimation for TCP" Proc. ACM Mobicom 2001
11 *
12 * - A. Grieco, s. Mascolo
13 * "Performance evaluation of New Reno, Vegas, Westwood+ TCP" ACM Computer
14 * Comm. Review, 2004
15 *
16 * - A. Dell'Aera, L. Grieco, S. Mascolo.
17 * "Linux 2.4 Implementation of Westwood+ TCP with Rate-Halving :
18 * A Performance Evaluation Over the Internet" (ICC 2004), Paris, June 2004
19 *
20 * Westwood+ employs end-to-end bandwidth measurement to set cwnd and
21 * ssthresh after packet loss. The probing phase is as the original Reno.
22 */
23
24#include <linux/mm.h>
25#include <linux/module.h>
26#include <linux/skbuff.h>
27#include <linux/inet_diag.h>
28#include <net/tcp.h>
29
30/* TCP Westwood structure */
31struct westwood {
32 u32 bw_ns_est; /* first bandwidth estimation..not too smoothed 8) */
33 u32 bw_est; /* bandwidth estimate */
34 u32 rtt_win_sx; /* here starts a new evaluation... */
35 u32 bk;
36 u32 snd_una; /* used for evaluating the number of acked bytes */
37 u32 cumul_ack;
38 u32 accounted;
39 u32 rtt;
40 u32 rtt_min; /* minimum observed RTT */
41 u8 first_ack; /* flag which infers that this is the first ack */
42 u8 reset_rtt_min; /* Reset RTT min to next RTT sample*/
43};
44
45/* TCP Westwood functions and constants */
46#define TCP_WESTWOOD_RTT_MIN (HZ/20) /* 50ms */
47#define TCP_WESTWOOD_INIT_RTT (20*HZ) /* maybe too conservative?! */
48
49/*
50 * @tcp_westwood_create
51 * This function initializes fields used in TCP Westwood+,
52 * it is called after the initial SYN, so the sequence numbers
53 * are correct but new passive connections we have no
54 * information about RTTmin at this time so we simply set it to
55 * TCP_WESTWOOD_INIT_RTT. This value was chosen to be too conservative
56 * since in this way we're sure it will be updated in a consistent
57 * way as soon as possible. It will reasonably happen within the first
58 * RTT period of the connection lifetime.
59 */
60static void tcp_westwood_init(struct sock *sk)
61{
62 struct westwood *w = inet_csk_ca(sk);
63
64 w->bk = 0;
65 w->bw_ns_est = 0;
66 w->bw_est = 0;
67 w->accounted = 0;
68 w->cumul_ack = 0;
69 w->reset_rtt_min = 1;
70 w->rtt_min = w->rtt = TCP_WESTWOOD_INIT_RTT;
71 w->rtt_win_sx = tcp_jiffies32;
72 w->snd_una = tcp_sk(sk)->snd_una;
73 w->first_ack = 1;
74}
75
76/*
77 * @westwood_do_filter
78 * Low-pass filter. Implemented using constant coefficients.
79 */
80static inline u32 westwood_do_filter(u32 a, u32 b)
81{
82 return ((7 * a) + b) >> 3;
83}
84
85static void westwood_filter(struct westwood *w, u32 delta)
86{
87 /* If the filter is empty fill it with the first sample of bandwidth */
88 if (w->bw_ns_est == 0 && w->bw_est == 0) {
89 w->bw_ns_est = w->bk / delta;
90 w->bw_est = w->bw_ns_est;
91 } else {
92 w->bw_ns_est = westwood_do_filter(w->bw_ns_est, w->bk / delta);
93 w->bw_est = westwood_do_filter(w->bw_est, w->bw_ns_est);
94 }
95}
96
97/*
98 * @westwood_pkts_acked
99 * Called after processing group of packets.
100 * but all westwood needs is the last sample of srtt.
101 */
102static void tcp_westwood_pkts_acked(struct sock *sk,
103 const struct ack_sample *sample)
104{
105 struct westwood *w = inet_csk_ca(sk);
106
107 if (sample->rtt_us > 0)
108 w->rtt = usecs_to_jiffies(sample->rtt_us);
109}
110
111/*
112 * @westwood_update_window
113 * It updates RTT evaluation window if it is the right moment to do
114 * it. If so it calls filter for evaluating bandwidth.
115 */
116static void westwood_update_window(struct sock *sk)
117{
118 struct westwood *w = inet_csk_ca(sk);
119 s32 delta = tcp_jiffies32 - w->rtt_win_sx;
120
121 /* Initialize w->snd_una with the first acked sequence number in order
122 * to fix mismatch between tp->snd_una and w->snd_una for the first
123 * bandwidth sample
124 */
125 if (w->first_ack) {
126 w->snd_una = tcp_sk(sk)->snd_una;
127 w->first_ack = 0;
128 }
129
130 /*
131 * See if a RTT-window has passed.
132 * Be careful since if RTT is less than
133 * 50ms we don't filter but we continue 'building the sample'.
134 * This minimum limit was chosen since an estimation on small
135 * time intervals is better to avoid...
136 * Obviously on a LAN we reasonably will always have
137 * right_bound = left_bound + WESTWOOD_RTT_MIN
138 */
139 if (w->rtt && delta > max_t(u32, w->rtt, TCP_WESTWOOD_RTT_MIN)) {
140 westwood_filter(w, delta);
141
142 w->bk = 0;
143 w->rtt_win_sx = tcp_jiffies32;
144 }
145}
146
147static inline void update_rtt_min(struct westwood *w)
148{
149 if (w->reset_rtt_min) {
150 w->rtt_min = w->rtt;
151 w->reset_rtt_min = 0;
152 } else
153 w->rtt_min = min(w->rtt, w->rtt_min);
154}
155
156/*
157 * @westwood_fast_bw
158 * It is called when we are in fast path. In particular it is called when
159 * header prediction is successful. In such case in fact update is
160 * straight forward and doesn't need any particular care.
161 */
162static inline void westwood_fast_bw(struct sock *sk)
163{
164 const struct tcp_sock *tp = tcp_sk(sk);
165 struct westwood *w = inet_csk_ca(sk);
166
167 westwood_update_window(sk);
168
169 w->bk += tp->snd_una - w->snd_una;
170 w->snd_una = tp->snd_una;
171 update_rtt_min(w);
172}
173
174/*
175 * @westwood_acked_count
176 * This function evaluates cumul_ack for evaluating bk in case of
177 * delayed or partial acks.
178 */
179static inline u32 westwood_acked_count(struct sock *sk)
180{
181 const struct tcp_sock *tp = tcp_sk(sk);
182 struct westwood *w = inet_csk_ca(sk);
183
184 w->cumul_ack = tp->snd_una - w->snd_una;
185
186 /* If cumul_ack is 0 this is a dupack since it's not moving
187 * tp->snd_una.
188 */
189 if (!w->cumul_ack) {
190 w->accounted += tp->mss_cache;
191 w->cumul_ack = tp->mss_cache;
192 }
193
194 if (w->cumul_ack > tp->mss_cache) {
195 /* Partial or delayed ack */
196 if (w->accounted >= w->cumul_ack) {
197 w->accounted -= w->cumul_ack;
198 w->cumul_ack = tp->mss_cache;
199 } else {
200 w->cumul_ack -= w->accounted;
201 w->accounted = 0;
202 }
203 }
204
205 w->snd_una = tp->snd_una;
206
207 return w->cumul_ack;
208}
209
210/*
211 * TCP Westwood
212 * Here limit is evaluated as Bw estimation*RTTmin (for obtaining it
213 * in packets we use mss_cache). Rttmin is guaranteed to be >= 2
214 * so avoids ever returning 0.
215 */
216static u32 tcp_westwood_bw_rttmin(const struct sock *sk)
217{
218 const struct tcp_sock *tp = tcp_sk(sk);
219 const struct westwood *w = inet_csk_ca(sk);
220
221 return max_t(u32, (w->bw_est * w->rtt_min) / tp->mss_cache, 2);
222}
223
224static void tcp_westwood_ack(struct sock *sk, u32 ack_flags)
225{
226 if (ack_flags & CA_ACK_SLOWPATH) {
227 struct westwood *w = inet_csk_ca(sk);
228
229 westwood_update_window(sk);
230 w->bk += westwood_acked_count(sk);
231
232 update_rtt_min(w);
233 return;
234 }
235
236 westwood_fast_bw(sk);
237}
238
239static void tcp_westwood_event(struct sock *sk, enum tcp_ca_event event)
240{
241 struct tcp_sock *tp = tcp_sk(sk);
242 struct westwood *w = inet_csk_ca(sk);
243
244 switch (event) {
245 case CA_EVENT_COMPLETE_CWR:
246 tp->snd_cwnd = tp->snd_ssthresh = tcp_westwood_bw_rttmin(sk);
247 break;
248 case CA_EVENT_LOSS:
249 tp->snd_ssthresh = tcp_westwood_bw_rttmin(sk);
250 /* Update RTT_min when next ack arrives */
251 w->reset_rtt_min = 1;
252 break;
253 default:
254 /* don't care */
255 break;
256 }
257}
258
259/* Extract info for Tcp socket info provided via netlink. */
260static size_t tcp_westwood_info(struct sock *sk, u32 ext, int *attr,
261 union tcp_cc_info *info)
262{
263 const struct westwood *ca = inet_csk_ca(sk);
264
265 if (ext & (1 << (INET_DIAG_VEGASINFO - 1))) {
266 info->vegas.tcpv_enabled = 1;
267 info->vegas.tcpv_rttcnt = 0;
268 info->vegas.tcpv_rtt = jiffies_to_usecs(ca->rtt);
269 info->vegas.tcpv_minrtt = jiffies_to_usecs(ca->rtt_min);
270
271 *attr = INET_DIAG_VEGASINFO;
272 return sizeof(struct tcpvegas_info);
273 }
274 return 0;
275}
276
277static struct tcp_congestion_ops tcp_westwood __read_mostly = {
278 .init = tcp_westwood_init,
279 .ssthresh = tcp_reno_ssthresh,
280 .cong_avoid = tcp_reno_cong_avoid,
281 .undo_cwnd = tcp_reno_undo_cwnd,
282 .cwnd_event = tcp_westwood_event,
283 .in_ack_event = tcp_westwood_ack,
284 .get_info = tcp_westwood_info,
285 .pkts_acked = tcp_westwood_pkts_acked,
286
287 .owner = THIS_MODULE,
288 .name = "westwood"
289};
290
291static int __init tcp_westwood_register(void)
292{
293 BUILD_BUG_ON(sizeof(struct westwood) > ICSK_CA_PRIV_SIZE);
294 return tcp_register_congestion_control(&tcp_westwood);
295}
296
297static void __exit tcp_westwood_unregister(void)
298{
299 tcp_unregister_congestion_control(&tcp_westwood);
300}
301
302module_init(tcp_westwood_register);
303module_exit(tcp_westwood_unregister);
304
305MODULE_AUTHOR("Stephen Hemminger, Angelo Dell'Aera");
306MODULE_LICENSE("GPL");
307MODULE_DESCRIPTION("TCP Westwood+");
308