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_time_stamp;
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, u32 cnt, s32 rtt)
103{
104	struct westwood *w = inet_csk_ca(sk);
105
106	if (rtt > 0)
107		w->rtt = usecs_to_jiffies(rtt);
108}
109
110/*
111 * @westwood_update_window
112 * It updates RTT evaluation window if it is the right moment to do
113 * it. If so it calls filter for evaluating bandwidth.
114 */
115static void westwood_update_window(struct sock *sk)
116{
117	struct westwood *w = inet_csk_ca(sk);
118	s32 delta = tcp_time_stamp - w->rtt_win_sx;
119
120	/* Initialize w->snd_una with the first acked sequence number in order
121	 * to fix mismatch between tp->snd_una and w->snd_una for the first
122	 * bandwidth sample
123	 */
124	if (w->first_ack) {
125		w->snd_una = tcp_sk(sk)->snd_una;
126		w->first_ack = 0;
127	}
128
129	/*
130	 * See if a RTT-window has passed.
131	 * Be careful since if RTT is less than
132	 * 50ms we don't filter but we continue 'building the sample'.
133	 * This minimum limit was chosen since an estimation on small
134	 * time intervals is better to avoid...
135	 * Obviously on a LAN we reasonably will always have
136	 * right_bound = left_bound + WESTWOOD_RTT_MIN
137	 */
138	if (w->rtt && delta > max_t(u32, w->rtt, TCP_WESTWOOD_RTT_MIN)) {
139		westwood_filter(w, delta);
140
141		w->bk = 0;
142		w->rtt_win_sx = tcp_time_stamp;
143	}
144}
145
146static inline void update_rtt_min(struct westwood *w)
147{
148	if (w->reset_rtt_min) {
149		w->rtt_min = w->rtt;
150		w->reset_rtt_min = 0;
151	} else
152		w->rtt_min = min(w->rtt, w->rtt_min);
153}
154
155/*
156 * @westwood_fast_bw
157 * It is called when we are in fast path. In particular it is called when
158 * header prediction is successful. In such case in fact update is
159 * straight forward and doesn't need any particular care.
160 */
161static inline void westwood_fast_bw(struct sock *sk)
162{
163	const struct tcp_sock *tp = tcp_sk(sk);
164	struct westwood *w = inet_csk_ca(sk);
165
166	westwood_update_window(sk);
167
168	w->bk += tp->snd_una - w->snd_una;
169	w->snd_una = tp->snd_una;
170	update_rtt_min(w);
171}
172
173/*
174 * @westwood_acked_count
175 * This function evaluates cumul_ack for evaluating bk in case of
176 * delayed or partial acks.
177 */
178static inline u32 westwood_acked_count(struct sock *sk)
179{
180	const struct tcp_sock *tp = tcp_sk(sk);
181	struct westwood *w = inet_csk_ca(sk);
182
183	w->cumul_ack = tp->snd_una - w->snd_una;
184
185	/* If cumul_ack is 0 this is a dupack since it's not moving
186	 * tp->snd_una.
187	 */
188	if (!w->cumul_ack) {
189		w->accounted += tp->mss_cache;
190		w->cumul_ack = tp->mss_cache;
191	}
192
193	if (w->cumul_ack > tp->mss_cache) {
194		/* Partial or delayed ack */
195		if (w->accounted >= w->cumul_ack) {
196			w->accounted -= w->cumul_ack;
197			w->cumul_ack = tp->mss_cache;
198		} else {
199			w->cumul_ack -= w->accounted;
200			w->accounted = 0;
201		}
202	}
203
204	w->snd_una = tp->snd_una;
205
206	return w->cumul_ack;
207}
208
209/*
210 * TCP Westwood
211 * Here limit is evaluated as Bw estimation*RTTmin (for obtaining it
212 * in packets we use mss_cache). Rttmin is guaranteed to be >= 2
213 * so avoids ever returning 0.
214 */
215static u32 tcp_westwood_bw_rttmin(const struct sock *sk)
216{
217	const struct tcp_sock *tp = tcp_sk(sk);
218	const struct westwood *w = inet_csk_ca(sk);
219
220	return max_t(u32, (w->bw_est * w->rtt_min) / tp->mss_cache, 2);
221}
222
223static void tcp_westwood_ack(struct sock *sk, u32 ack_flags)
224{
225	if (ack_flags & CA_ACK_SLOWPATH) {
226		struct westwood *w = inet_csk_ca(sk);
227
228		westwood_update_window(sk);
229		w->bk += westwood_acked_count(sk);
230
231		update_rtt_min(w);
232		return;
233	}
234
235	westwood_fast_bw(sk);
236}
237
238static void tcp_westwood_event(struct sock *sk, enum tcp_ca_event event)
239{
240	struct tcp_sock *tp = tcp_sk(sk);
241	struct westwood *w = inet_csk_ca(sk);
242
243	switch (event) {
244	case CA_EVENT_COMPLETE_CWR:
245		tp->snd_cwnd = tp->snd_ssthresh = tcp_westwood_bw_rttmin(sk);
246		break;
247	case CA_EVENT_LOSS:
248		tp->snd_ssthresh = tcp_westwood_bw_rttmin(sk);
249		/* Update RTT_min when next ack arrives */
250		w->reset_rtt_min = 1;
251		break;
252	default:
253		/* don't care */
254		break;
255	}
256}
257
258/* Extract info for Tcp socket info provided via netlink. */
259static size_t tcp_westwood_info(struct sock *sk, u32 ext, int *attr,
260				union tcp_cc_info *info)
261{
262	const struct westwood *ca = inet_csk_ca(sk);
263
264	if (ext & (1 << (INET_DIAG_VEGASINFO - 1))) {
265		info->vegas.tcpv_enabled = 1;
266		info->vegas.tcpv_rttcnt	= 0;
267		info->vegas.tcpv_rtt	= jiffies_to_usecs(ca->rtt),
268		info->vegas.tcpv_minrtt	= jiffies_to_usecs(ca->rtt_min),
269
270		*attr = INET_DIAG_VEGASINFO;
271		return sizeof(struct tcpvegas_info);
272	}
273	return 0;
274}
275
276static struct tcp_congestion_ops tcp_westwood __read_mostly = {
277	.init		= tcp_westwood_init,
278	.ssthresh	= tcp_reno_ssthresh,
279	.cong_avoid	= tcp_reno_cong_avoid,
280	.cwnd_event	= tcp_westwood_event,
281	.in_ack_event	= tcp_westwood_ack,
282	.get_info	= tcp_westwood_info,
283	.pkts_acked	= tcp_westwood_pkts_acked,
284
285	.owner		= THIS_MODULE,
286	.name		= "westwood"
287};
288
289static int __init tcp_westwood_register(void)
290{
291	BUILD_BUG_ON(sizeof(struct westwood) > ICSK_CA_PRIV_SIZE);
292	return tcp_register_congestion_control(&tcp_westwood);
293}
294
295static void __exit tcp_westwood_unregister(void)
296{
297	tcp_unregister_congestion_control(&tcp_westwood);
298}
299
300module_init(tcp_westwood_register);
301module_exit(tcp_westwood_unregister);
302
303MODULE_AUTHOR("Stephen Hemminger, Angelo Dell'Aera");
304MODULE_LICENSE("GPL");
305MODULE_DESCRIPTION("TCP Westwood+");
306