1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * TCP Westwood+: end-to-end bandwidth estimation for TCP
4 *
5 *      Angelo Dell'Aera: author of the first version of TCP Westwood+ in Linux 2.4
6 *
7 * Support at http://c3lab.poliba.it/index.php/Westwood
8 * Main references in literature:
9 *
10 * - Mascolo S, Casetti, M. Gerla et al.
11 *   "TCP Westwood: bandwidth estimation for TCP" Proc. ACM Mobicom 2001
12 *
13 * - A. Grieco, s. Mascolo
14 *   "Performance evaluation of New Reno, Vegas, Westwood+ TCP" ACM Computer
15 *     Comm. Review, 2004
16 *
17 * - A. Dell'Aera, L. Grieco, S. Mascolo.
18 *   "Linux 2.4 Implementation of Westwood+ TCP with Rate-Halving :
19 *    A Performance Evaluation Over the Internet" (ICC 2004), Paris, June 2004
20 *
21 * Westwood+ employs end-to-end bandwidth measurement to set cwnd and
22 * ssthresh after packet loss. The probing phase is as the original Reno.
23 */
24
25#include <linux/mm.h>
26#include <linux/module.h>
27#include <linux/skbuff.h>
28#include <linux/inet_diag.h>
29#include <net/tcp.h>
30
31/* TCP Westwood structure */
32struct westwood {
33	u32    bw_ns_est;        /* first bandwidth estimation..not too smoothed 8) */
34	u32    bw_est;           /* bandwidth estimate */
35	u32    rtt_win_sx;       /* here starts a new evaluation... */
36	u32    bk;
37	u32    snd_una;          /* used for evaluating the number of acked bytes */
38	u32    cumul_ack;
39	u32    accounted;
40	u32    rtt;
41	u32    rtt_min;          /* minimum observed RTT */
42	u8     first_ack;        /* flag which infers that this is the first ack */
43	u8     reset_rtt_min;    /* Reset RTT min to next RTT sample*/
44};
45
46/* TCP Westwood functions and constants */
47#define TCP_WESTWOOD_RTT_MIN   (HZ/20)	/* 50ms */
48#define TCP_WESTWOOD_INIT_RTT  (20*HZ)	/* maybe too conservative?! */
49
50/*
51 * @tcp_westwood_create
52 * This function initializes fields used in TCP Westwood+,
53 * it is called after the initial SYN, so the sequence numbers
54 * are correct but new passive connections we have no
55 * information about RTTmin at this time so we simply set it to
56 * TCP_WESTWOOD_INIT_RTT. This value was chosen to be too conservative
57 * since in this way we're sure it will be updated in a consistent
58 * way as soon as possible. It will reasonably happen within the first
59 * RTT period of the connection lifetime.
60 */
61static void tcp_westwood_init(struct sock *sk)
62{
63	struct westwood *w = inet_csk_ca(sk);
64
65	w->bk = 0;
66	w->bw_ns_est = 0;
67	w->bw_est = 0;
68	w->accounted = 0;
69	w->cumul_ack = 0;
70	w->reset_rtt_min = 1;
71	w->rtt_min = w->rtt = TCP_WESTWOOD_INIT_RTT;
72	w->rtt_win_sx = tcp_jiffies32;
73	w->snd_una = tcp_sk(sk)->snd_una;
74	w->first_ack = 1;
75}
76
77/*
78 * @westwood_do_filter
79 * Low-pass filter. Implemented using constant coefficients.
80 */
81static inline u32 westwood_do_filter(u32 a, u32 b)
82{
83	return ((7 * a) + b) >> 3;
84}
85
86static void westwood_filter(struct westwood *w, u32 delta)
87{
88	/* If the filter is empty fill it with the first sample of bandwidth  */
89	if (w->bw_ns_est == 0 && w->bw_est == 0) {
90		w->bw_ns_est = w->bk / delta;
91		w->bw_est = w->bw_ns_est;
92	} else {
93		w->bw_ns_est = westwood_do_filter(w->bw_ns_est, w->bk / delta);
94		w->bw_est = westwood_do_filter(w->bw_est, w->bw_ns_est);
95	}
96}
97
98/*
99 * @westwood_pkts_acked
100 * Called after processing group of packets.
101 * but all westwood needs is the last sample of srtt.
102 */
103static void tcp_westwood_pkts_acked(struct sock *sk,
104				    const struct ack_sample *sample)
105{
106	struct westwood *w = inet_csk_ca(sk);
107
108	if (sample->rtt_us > 0)
109		w->rtt = usecs_to_jiffies(sample->rtt_us);
110}
111
112/*
113 * @westwood_update_window
114 * It updates RTT evaluation window if it is the right moment to do
115 * it. If so it calls filter for evaluating bandwidth.
116 */
117static void westwood_update_window(struct sock *sk)
118{
119	struct westwood *w = inet_csk_ca(sk);
120	s32 delta = tcp_jiffies32 - w->rtt_win_sx;
121
122	/* Initialize w->snd_una with the first acked sequence number in order
123	 * to fix mismatch between tp->snd_una and w->snd_una for the first
124	 * bandwidth sample
125	 */
126	if (w->first_ack) {
127		w->snd_una = tcp_sk(sk)->snd_una;
128		w->first_ack = 0;
129	}
130
131	/*
132	 * See if a RTT-window has passed.
133	 * Be careful since if RTT is less than
134	 * 50ms we don't filter but we continue 'building the sample'.
135	 * This minimum limit was chosen since an estimation on small
136	 * time intervals is better to avoid...
137	 * Obviously on a LAN we reasonably will always have
138	 * right_bound = left_bound + WESTWOOD_RTT_MIN
139	 */
140	if (w->rtt && delta > max_t(u32, w->rtt, TCP_WESTWOOD_RTT_MIN)) {
141		westwood_filter(w, delta);
142
143		w->bk = 0;
144		w->rtt_win_sx = tcp_jiffies32;
145	}
146}
147
148static inline void update_rtt_min(struct westwood *w)
149{
150	if (w->reset_rtt_min) {
151		w->rtt_min = w->rtt;
152		w->reset_rtt_min = 0;
153	} else
154		w->rtt_min = min(w->rtt, w->rtt_min);
155}
156
157/*
158 * @westwood_fast_bw
159 * It is called when we are in fast path. In particular it is called when
160 * header prediction is successful. In such case in fact update is
161 * straight forward and doesn't need any particular care.
162 */
163static inline void westwood_fast_bw(struct sock *sk)
164{
165	const struct tcp_sock *tp = tcp_sk(sk);
166	struct westwood *w = inet_csk_ca(sk);
167
168	westwood_update_window(sk);
169
170	w->bk += tp->snd_una - w->snd_una;
171	w->snd_una = tp->snd_una;
172	update_rtt_min(w);
173}
174
175/*
176 * @westwood_acked_count
177 * This function evaluates cumul_ack for evaluating bk in case of
178 * delayed or partial acks.
179 */
180static inline u32 westwood_acked_count(struct sock *sk)
181{
182	const struct tcp_sock *tp = tcp_sk(sk);
183	struct westwood *w = inet_csk_ca(sk);
184
185	w->cumul_ack = tp->snd_una - w->snd_una;
186
187	/* If cumul_ack is 0 this is a dupack since it's not moving
188	 * tp->snd_una.
189	 */
190	if (!w->cumul_ack) {
191		w->accounted += tp->mss_cache;
192		w->cumul_ack = tp->mss_cache;
193	}
194
195	if (w->cumul_ack > tp->mss_cache) {
196		/* Partial or delayed ack */
197		if (w->accounted >= w->cumul_ack) {
198			w->accounted -= w->cumul_ack;
199			w->cumul_ack = tp->mss_cache;
200		} else {
201			w->cumul_ack -= w->accounted;
202			w->accounted = 0;
203		}
204	}
205
206	w->snd_una = tp->snd_una;
207
208	return w->cumul_ack;
209}
210
211/*
212 * TCP Westwood
213 * Here limit is evaluated as Bw estimation*RTTmin (for obtaining it
214 * in packets we use mss_cache). Rttmin is guaranteed to be >= 2
215 * so avoids ever returning 0.
216 */
217static u32 tcp_westwood_bw_rttmin(const struct sock *sk)
218{
219	const struct tcp_sock *tp = tcp_sk(sk);
220	const struct westwood *w = inet_csk_ca(sk);
221
222	return max_t(u32, (w->bw_est * w->rtt_min) / tp->mss_cache, 2);
223}
224
225static void tcp_westwood_ack(struct sock *sk, u32 ack_flags)
226{
227	if (ack_flags & CA_ACK_SLOWPATH) {
228		struct westwood *w = inet_csk_ca(sk);
229
230		westwood_update_window(sk);
231		w->bk += westwood_acked_count(sk);
232
233		update_rtt_min(w);
234		return;
235	}
236
237	westwood_fast_bw(sk);
238}
239
240static void tcp_westwood_event(struct sock *sk, enum tcp_ca_event event)
241{
242	struct tcp_sock *tp = tcp_sk(sk);
243	struct westwood *w = inet_csk_ca(sk);
244
245	switch (event) {
246	case CA_EVENT_COMPLETE_CWR:
247		tp->snd_ssthresh = tcp_westwood_bw_rttmin(sk);
248		tcp_snd_cwnd_set(tp, tp->snd_ssthresh);
249		break;
250	case CA_EVENT_LOSS:
251		tp->snd_ssthresh = tcp_westwood_bw_rttmin(sk);
252		/* Update RTT_min when next ack arrives */
253		w->reset_rtt_min = 1;
254		break;
255	default:
256		/* don't care */
257		break;
258	}
259}
260
261/* Extract info for Tcp socket info provided via netlink. */
262static size_t tcp_westwood_info(struct sock *sk, u32 ext, int *attr,
263				union tcp_cc_info *info)
264{
265	const struct westwood *ca = inet_csk_ca(sk);
266
267	if (ext & (1 << (INET_DIAG_VEGASINFO - 1))) {
268		info->vegas.tcpv_enabled = 1;
269		info->vegas.tcpv_rttcnt	= 0;
270		info->vegas.tcpv_rtt	= jiffies_to_usecs(ca->rtt);
271		info->vegas.tcpv_minrtt	= jiffies_to_usecs(ca->rtt_min);
272
273		*attr = INET_DIAG_VEGASINFO;
274		return sizeof(struct tcpvegas_info);
275	}
276	return 0;
277}
278
279static struct tcp_congestion_ops tcp_westwood __read_mostly = {
280	.init		= tcp_westwood_init,
281	.ssthresh	= tcp_reno_ssthresh,
282	.cong_avoid	= tcp_reno_cong_avoid,
283	.undo_cwnd      = tcp_reno_undo_cwnd,
284	.cwnd_event	= tcp_westwood_event,
285	.in_ack_event	= tcp_westwood_ack,
286	.get_info	= tcp_westwood_info,
287	.pkts_acked	= tcp_westwood_pkts_acked,
288
289	.owner		= THIS_MODULE,
290	.name		= "westwood"
291};
292
293static int __init tcp_westwood_register(void)
294{
295	BUILD_BUG_ON(sizeof(struct westwood) > ICSK_CA_PRIV_SIZE);
296	return tcp_register_congestion_control(&tcp_westwood);
297}
298
299static void __exit tcp_westwood_unregister(void)
300{
301	tcp_unregister_congestion_control(&tcp_westwood);
302}
303
304module_init(tcp_westwood_register);
305module_exit(tcp_westwood_unregister);
306
307MODULE_AUTHOR("Stephen Hemminger, Angelo Dell'Aera");
308MODULE_LICENSE("GPL");
309MODULE_DESCRIPTION("TCP Westwood+");
310