diff options
-rw-r--r-- | net/ipv4/Kconfig | 14 | ||||
-rw-r--r-- | net/ipv4/Makefile | 1 | ||||
-rw-r--r-- | net/ipv4/tcp_yeah.c | 288 | ||||
-rw-r--r-- | net/ipv4/tcp_yeah.h | 134 |
4 files changed, 437 insertions, 0 deletions
diff --git a/net/ipv4/Kconfig b/net/ipv4/Kconfig index 9e8ef509c51d..dc61e6641624 100644 --- a/net/ipv4/Kconfig +++ b/net/ipv4/Kconfig | |||
@@ -574,6 +574,20 @@ config TCP_CONG_VENO | |||
574 | loss packets. | 574 | loss packets. |
575 | See http://www.ntu.edu.sg/home5/ZHOU0022/papers/CPFu03a.pdf | 575 | See http://www.ntu.edu.sg/home5/ZHOU0022/papers/CPFu03a.pdf |
576 | 576 | ||
577 | config TCP_CONG_YEAH | ||
578 | tristate "YeAH TCP" | ||
579 | depends on EXPERIMENTAL | ||
580 | default n | ||
581 | ---help--- | ||
582 | YeAH-TCP is a sender-side high-speed enabled TCP congestion control | ||
583 | algorithm, which uses a mixed loss/delay approach to compute the | ||
584 | congestion window. It's design goals target high efficiency, | ||
585 | internal, RTT and Reno fairness, resilience to link loss while | ||
586 | keeping network elements load as low as possible. | ||
587 | |||
588 | For further details look here: | ||
589 | http://wil.cs.caltech.edu/pfldnet2007/paper/YeAH_TCP.pdf | ||
590 | |||
577 | choice | 591 | choice |
578 | prompt "Default TCP congestion control" | 592 | prompt "Default TCP congestion control" |
579 | default DEFAULT_CUBIC | 593 | default DEFAULT_CUBIC |
diff --git a/net/ipv4/Makefile b/net/ipv4/Makefile index 7a068626feea..eeb94d5cac96 100644 --- a/net/ipv4/Makefile +++ b/net/ipv4/Makefile | |||
@@ -49,6 +49,7 @@ obj-$(CONFIG_TCP_CONG_VEGAS) += tcp_vegas.o | |||
49 | obj-$(CONFIG_TCP_CONG_VENO) += tcp_veno.o | 49 | obj-$(CONFIG_TCP_CONG_VENO) += tcp_veno.o |
50 | obj-$(CONFIG_TCP_CONG_SCALABLE) += tcp_scalable.o | 50 | obj-$(CONFIG_TCP_CONG_SCALABLE) += tcp_scalable.o |
51 | obj-$(CONFIG_TCP_CONG_LP) += tcp_lp.o | 51 | obj-$(CONFIG_TCP_CONG_LP) += tcp_lp.o |
52 | obj-$(CONFIG_TCP_CONG_YEAH) += tcp_yeah.o | ||
52 | obj-$(CONFIG_NETLABEL) += cipso_ipv4.o | 53 | obj-$(CONFIG_NETLABEL) += cipso_ipv4.o |
53 | 54 | ||
54 | obj-$(CONFIG_XFRM) += xfrm4_policy.o xfrm4_state.o xfrm4_input.o \ | 55 | obj-$(CONFIG_XFRM) += xfrm4_policy.o xfrm4_state.o xfrm4_input.o \ |
diff --git a/net/ipv4/tcp_yeah.c b/net/ipv4/tcp_yeah.c new file mode 100644 index 000000000000..815e020e98fe --- /dev/null +++ b/net/ipv4/tcp_yeah.c | |||
@@ -0,0 +1,288 @@ | |||
1 | /* | ||
2 | * | ||
3 | * YeAH TCP | ||
4 | * | ||
5 | * For further details look at: | ||
6 | * http://wil.cs.caltech.edu/pfldnet2007/paper/YeAH_TCP.pdf | ||
7 | * | ||
8 | */ | ||
9 | |||
10 | #include "tcp_yeah.h" | ||
11 | |||
12 | /* Default values of the Vegas variables, in fixed-point representation | ||
13 | * with V_PARAM_SHIFT bits to the right of the binary point. | ||
14 | */ | ||
15 | #define V_PARAM_SHIFT 1 | ||
16 | |||
17 | #define TCP_YEAH_ALPHA 80 //lin number of packets queued at the bottleneck | ||
18 | #define TCP_YEAH_GAMMA 1 //lin fraction of queue to be removed per rtt | ||
19 | #define TCP_YEAH_DELTA 3 //log minimum fraction of cwnd to be removed on loss | ||
20 | #define TCP_YEAH_EPSILON 1 //log maximum fraction to be removed on early decongestion | ||
21 | #define TCP_YEAH_PHY 8 //lin maximum delta from base | ||
22 | #define TCP_YEAH_RHO 16 //lin minumum number of consecutive rtt to consider competition on loss | ||
23 | #define TCP_YEAH_ZETA 50 //lin minimum number of state switchs to reset reno_count | ||
24 | |||
25 | #define TCP_SCALABLE_AI_CNT 100U | ||
26 | |||
27 | /* YeAH variables */ | ||
28 | struct yeah { | ||
29 | /* Vegas */ | ||
30 | u32 beg_snd_nxt; /* right edge during last RTT */ | ||
31 | u32 beg_snd_una; /* left edge during last RTT */ | ||
32 | u32 beg_snd_cwnd; /* saves the size of the cwnd */ | ||
33 | u8 doing_vegas_now;/* if true, do vegas for this RTT */ | ||
34 | u16 cntRTT; /* # of RTTs measured within last RTT */ | ||
35 | u32 minRTT; /* min of RTTs measured within last RTT (in usec) */ | ||
36 | u32 baseRTT; /* the min of all Vegas RTT measurements seen (in usec) */ | ||
37 | |||
38 | /* YeAH */ | ||
39 | u32 lastQ; | ||
40 | u32 doing_reno_now; | ||
41 | |||
42 | u32 reno_count; | ||
43 | u32 fast_count; | ||
44 | |||
45 | u32 pkts_acked; | ||
46 | }; | ||
47 | |||
48 | static void tcp_yeah_init(struct sock *sk) | ||
49 | { | ||
50 | struct tcp_sock *tp = tcp_sk(sk); | ||
51 | struct yeah *yeah = inet_csk_ca(sk); | ||
52 | |||
53 | tcp_vegas_init(sk); | ||
54 | |||
55 | yeah->doing_reno_now = 0; | ||
56 | yeah->lastQ = 0; | ||
57 | |||
58 | yeah->reno_count = 2; | ||
59 | |||
60 | /* Ensure the MD arithmetic works. This is somewhat pedantic, | ||
61 | * since I don't think we will see a cwnd this large. :) */ | ||
62 | tp->snd_cwnd_clamp = min_t(u32, tp->snd_cwnd_clamp, 0xffffffff/128); | ||
63 | |||
64 | } | ||
65 | |||
66 | |||
67 | static void tcp_yeah_pkts_acked(struct sock *sk, u32 pkts_acked) | ||
68 | { | ||
69 | const struct inet_connection_sock *icsk = inet_csk(sk); | ||
70 | struct yeah *yeah = inet_csk_ca(sk); | ||
71 | |||
72 | if (icsk->icsk_ca_state == TCP_CA_Open) | ||
73 | yeah->pkts_acked = pkts_acked; | ||
74 | } | ||
75 | |||
76 | /* 64bit divisor, dividend and result. dynamic precision */ | ||
77 | static inline u64 div64_64(u64 dividend, u64 divisor) | ||
78 | { | ||
79 | u32 d = divisor; | ||
80 | |||
81 | if (divisor > 0xffffffffULL) { | ||
82 | unsigned int shift = fls(divisor >> 32); | ||
83 | |||
84 | d = divisor >> shift; | ||
85 | dividend >>= shift; | ||
86 | } | ||
87 | |||
88 | /* avoid 64 bit division if possible */ | ||
89 | if (dividend >> 32) | ||
90 | do_div(dividend, d); | ||
91 | else | ||
92 | dividend = (u32) dividend / d; | ||
93 | |||
94 | return dividend; | ||
95 | } | ||
96 | |||
97 | static void tcp_yeah_cong_avoid(struct sock *sk, u32 ack, | ||
98 | u32 seq_rtt, u32 in_flight, int flag) | ||
99 | { | ||
100 | struct tcp_sock *tp = tcp_sk(sk); | ||
101 | struct yeah *yeah = inet_csk_ca(sk); | ||
102 | |||
103 | if (!tcp_is_cwnd_limited(sk, in_flight)) | ||
104 | return; | ||
105 | |||
106 | if (tp->snd_cwnd <= tp->snd_ssthresh) { | ||
107 | tcp_slow_start(tp); | ||
108 | } else if (!yeah->doing_reno_now) { | ||
109 | /* Scalable */ | ||
110 | |||
111 | tp->snd_cwnd_cnt+=yeah->pkts_acked; | ||
112 | if (tp->snd_cwnd_cnt > min(tp->snd_cwnd, TCP_SCALABLE_AI_CNT)){ | ||
113 | if (tp->snd_cwnd < tp->snd_cwnd_clamp) | ||
114 | tp->snd_cwnd++; | ||
115 | tp->snd_cwnd_cnt = 0; | ||
116 | } | ||
117 | |||
118 | yeah->pkts_acked = 1; | ||
119 | |||
120 | } else { | ||
121 | /* Reno */ | ||
122 | |||
123 | if (tp->snd_cwnd_cnt < tp->snd_cwnd) | ||
124 | tp->snd_cwnd_cnt++; | ||
125 | |||
126 | if (tp->snd_cwnd_cnt >= tp->snd_cwnd) { | ||
127 | tp->snd_cwnd++; | ||
128 | tp->snd_cwnd_cnt = 0; | ||
129 | } | ||
130 | } | ||
131 | |||
132 | /* The key players are v_beg_snd_una and v_beg_snd_nxt. | ||
133 | * | ||
134 | * These are so named because they represent the approximate values | ||
135 | * of snd_una and snd_nxt at the beginning of the current RTT. More | ||
136 | * precisely, they represent the amount of data sent during the RTT. | ||
137 | * At the end of the RTT, when we receive an ACK for v_beg_snd_nxt, | ||
138 | * we will calculate that (v_beg_snd_nxt - v_beg_snd_una) outstanding | ||
139 | * bytes of data have been ACKed during the course of the RTT, giving | ||
140 | * an "actual" rate of: | ||
141 | * | ||
142 | * (v_beg_snd_nxt - v_beg_snd_una) / (rtt duration) | ||
143 | * | ||
144 | * Unfortunately, v_beg_snd_una is not exactly equal to snd_una, | ||
145 | * because delayed ACKs can cover more than one segment, so they | ||
146 | * don't line up yeahly with the boundaries of RTTs. | ||
147 | * | ||
148 | * Another unfortunate fact of life is that delayed ACKs delay the | ||
149 | * advance of the left edge of our send window, so that the number | ||
150 | * of bytes we send in an RTT is often less than our cwnd will allow. | ||
151 | * So we keep track of our cwnd separately, in v_beg_snd_cwnd. | ||
152 | */ | ||
153 | |||
154 | if (after(ack, yeah->beg_snd_nxt)) { | ||
155 | |||
156 | /* We do the Vegas calculations only if we got enough RTT | ||
157 | * samples that we can be reasonably sure that we got | ||
158 | * at least one RTT sample that wasn't from a delayed ACK. | ||
159 | * If we only had 2 samples total, | ||
160 | * then that means we're getting only 1 ACK per RTT, which | ||
161 | * means they're almost certainly delayed ACKs. | ||
162 | * If we have 3 samples, we should be OK. | ||
163 | */ | ||
164 | |||
165 | if (yeah->cntRTT > 2) { | ||
166 | u32 rtt; | ||
167 | u32 queue, maxqueue; | ||
168 | |||
169 | /* We have enough RTT samples, so, using the Vegas | ||
170 | * algorithm, we determine if we should increase or | ||
171 | * decrease cwnd, and by how much. | ||
172 | */ | ||
173 | |||
174 | /* Pluck out the RTT we are using for the Vegas | ||
175 | * calculations. This is the min RTT seen during the | ||
176 | * last RTT. Taking the min filters out the effects | ||
177 | * of delayed ACKs, at the cost of noticing congestion | ||
178 | * a bit later. | ||
179 | */ | ||
180 | rtt = yeah->minRTT; | ||
181 | |||
182 | queue = (u32)div64_64((u64)tp->snd_cwnd * (rtt - yeah->baseRTT), rtt); | ||
183 | |||
184 | maxqueue = TCP_YEAH_ALPHA; | ||
185 | |||
186 | if (queue > maxqueue || | ||
187 | rtt - yeah->baseRTT > (yeah->baseRTT / TCP_YEAH_PHY)) { | ||
188 | |||
189 | if (queue > maxqueue && tp->snd_cwnd > yeah->reno_count) { | ||
190 | u32 reduction = min( queue / TCP_YEAH_GAMMA , | ||
191 | tp->snd_cwnd >> TCP_YEAH_EPSILON ); | ||
192 | |||
193 | tp->snd_cwnd -= reduction; | ||
194 | |||
195 | tp->snd_cwnd = max( tp->snd_cwnd, yeah->reno_count); | ||
196 | |||
197 | tp->snd_ssthresh = tp->snd_cwnd; | ||
198 | } | ||
199 | |||
200 | if (yeah->reno_count <= 2) | ||
201 | yeah->reno_count = max( tp->snd_cwnd>>1, 2U); | ||
202 | else | ||
203 | yeah->reno_count++; | ||
204 | |||
205 | yeah->doing_reno_now = | ||
206 | min_t( u32, yeah->doing_reno_now + 1 , 0xffffff); | ||
207 | |||
208 | } else { | ||
209 | yeah->fast_count++; | ||
210 | |||
211 | if (yeah->fast_count > TCP_YEAH_ZETA) { | ||
212 | yeah->reno_count = 2; | ||
213 | yeah->fast_count = 0; | ||
214 | } | ||
215 | |||
216 | yeah->doing_reno_now = 0; | ||
217 | } | ||
218 | |||
219 | yeah->lastQ = queue; | ||
220 | |||
221 | } | ||
222 | |||
223 | /* Save the extent of the current window so we can use this | ||
224 | * at the end of the next RTT. | ||
225 | */ | ||
226 | yeah->beg_snd_una = yeah->beg_snd_nxt; | ||
227 | yeah->beg_snd_nxt = tp->snd_nxt; | ||
228 | yeah->beg_snd_cwnd = tp->snd_cwnd; | ||
229 | |||
230 | /* Wipe the slate clean for the next RTT. */ | ||
231 | yeah->cntRTT = 0; | ||
232 | yeah->minRTT = 0x7fffffff; | ||
233 | } | ||
234 | } | ||
235 | |||
236 | static u32 tcp_yeah_ssthresh(struct sock *sk) { | ||
237 | const struct tcp_sock *tp = tcp_sk(sk); | ||
238 | struct yeah *yeah = inet_csk_ca(sk); | ||
239 | u32 reduction; | ||
240 | |||
241 | if (yeah->doing_reno_now < TCP_YEAH_RHO) { | ||
242 | reduction = yeah->lastQ; | ||
243 | |||
244 | reduction = min( reduction, max(tp->snd_cwnd>>1, 2U) ); | ||
245 | |||
246 | reduction = max( reduction, tp->snd_cwnd >> TCP_YEAH_DELTA); | ||
247 | } else | ||
248 | reduction = max(tp->snd_cwnd>>1,2U); | ||
249 | |||
250 | yeah->fast_count = 0; | ||
251 | yeah->reno_count = max(yeah->reno_count>>1, 2U); | ||
252 | |||
253 | return tp->snd_cwnd - reduction; | ||
254 | } | ||
255 | |||
256 | static struct tcp_congestion_ops tcp_yeah = { | ||
257 | .init = tcp_yeah_init, | ||
258 | .ssthresh = tcp_yeah_ssthresh, | ||
259 | .cong_avoid = tcp_yeah_cong_avoid, | ||
260 | .min_cwnd = tcp_reno_min_cwnd, | ||
261 | .rtt_sample = tcp_vegas_rtt_calc, | ||
262 | .set_state = tcp_vegas_state, | ||
263 | .cwnd_event = tcp_vegas_cwnd_event, | ||
264 | .get_info = tcp_vegas_get_info, | ||
265 | .pkts_acked = tcp_yeah_pkts_acked, | ||
266 | |||
267 | .owner = THIS_MODULE, | ||
268 | .name = "yeah", | ||
269 | }; | ||
270 | |||
271 | static int __init tcp_yeah_register(void) | ||
272 | { | ||
273 | BUG_ON(sizeof(struct yeah) > ICSK_CA_PRIV_SIZE); | ||
274 | tcp_register_congestion_control(&tcp_yeah); | ||
275 | return 0; | ||
276 | } | ||
277 | |||
278 | static void __exit tcp_yeah_unregister(void) | ||
279 | { | ||
280 | tcp_unregister_congestion_control(&tcp_yeah); | ||
281 | } | ||
282 | |||
283 | module_init(tcp_yeah_register); | ||
284 | module_exit(tcp_yeah_unregister); | ||
285 | |||
286 | MODULE_AUTHOR("Angelo P. Castellani"); | ||
287 | MODULE_LICENSE("GPL"); | ||
288 | MODULE_DESCRIPTION("YeAH TCP"); | ||
diff --git a/net/ipv4/tcp_yeah.h b/net/ipv4/tcp_yeah.h new file mode 100644 index 000000000000..b3255dba4e2d --- /dev/null +++ b/net/ipv4/tcp_yeah.h | |||
@@ -0,0 +1,134 @@ | |||
1 | #include <linux/mm.h> | ||
2 | #include <linux/module.h> | ||
3 | #include <linux/skbuff.h> | ||
4 | #include <linux/inet_diag.h> | ||
5 | |||
6 | #include <net/tcp.h> | ||
7 | |||
8 | /* Vegas variables */ | ||
9 | struct vegas { | ||
10 | u32 beg_snd_nxt; /* right edge during last RTT */ | ||
11 | u32 beg_snd_una; /* left edge during last RTT */ | ||
12 | u32 beg_snd_cwnd; /* saves the size of the cwnd */ | ||
13 | u8 doing_vegas_now;/* if true, do vegas for this RTT */ | ||
14 | u16 cntRTT; /* # of RTTs measured within last RTT */ | ||
15 | u32 minRTT; /* min of RTTs measured within last RTT (in usec) */ | ||
16 | u32 baseRTT; /* the min of all Vegas RTT measurements seen (in usec) */ | ||
17 | }; | ||
18 | |||
19 | /* There are several situations when we must "re-start" Vegas: | ||
20 | * | ||
21 | * o when a connection is established | ||
22 | * o after an RTO | ||
23 | * o after fast recovery | ||
24 | * o when we send a packet and there is no outstanding | ||
25 | * unacknowledged data (restarting an idle connection) | ||
26 | * | ||
27 | * In these circumstances we cannot do a Vegas calculation at the | ||
28 | * end of the first RTT, because any calculation we do is using | ||
29 | * stale info -- both the saved cwnd and congestion feedback are | ||
30 | * stale. | ||
31 | * | ||
32 | * Instead we must wait until the completion of an RTT during | ||
33 | * which we actually receive ACKs. | ||
34 | */ | ||
35 | static inline void vegas_enable(struct sock *sk) | ||
36 | { | ||
37 | const struct tcp_sock *tp = tcp_sk(sk); | ||
38 | struct vegas *vegas = inet_csk_ca(sk); | ||
39 | |||
40 | /* Begin taking Vegas samples next time we send something. */ | ||
41 | vegas->doing_vegas_now = 1; | ||
42 | |||
43 | /* Set the beginning of the next send window. */ | ||
44 | vegas->beg_snd_nxt = tp->snd_nxt; | ||
45 | |||
46 | vegas->cntRTT = 0; | ||
47 | vegas->minRTT = 0x7fffffff; | ||
48 | } | ||
49 | |||
50 | /* Stop taking Vegas samples for now. */ | ||
51 | static inline void vegas_disable(struct sock *sk) | ||
52 | { | ||
53 | struct vegas *vegas = inet_csk_ca(sk); | ||
54 | |||
55 | vegas->doing_vegas_now = 0; | ||
56 | } | ||
57 | |||
58 | static void tcp_vegas_init(struct sock *sk) | ||
59 | { | ||
60 | struct vegas *vegas = inet_csk_ca(sk); | ||
61 | |||
62 | vegas->baseRTT = 0x7fffffff; | ||
63 | vegas_enable(sk); | ||
64 | } | ||
65 | |||
66 | static void tcp_vegas_state(struct sock *sk, u8 ca_state) | ||
67 | { | ||
68 | |||
69 | if (ca_state == TCP_CA_Open) | ||
70 | vegas_enable(sk); | ||
71 | else | ||
72 | vegas_disable(sk); | ||
73 | } | ||
74 | |||
75 | /* Do RTT sampling needed for Vegas. | ||
76 | * Basically we: | ||
77 | * o min-filter RTT samples from within an RTT to get the current | ||
78 | * propagation delay + queuing delay (we are min-filtering to try to | ||
79 | * avoid the effects of delayed ACKs) | ||
80 | * o min-filter RTT samples from a much longer window (forever for now) | ||
81 | * to find the propagation delay (baseRTT) | ||
82 | */ | ||
83 | static void tcp_vegas_rtt_calc(struct sock *sk, u32 usrtt) | ||
84 | { | ||
85 | struct vegas *vegas = inet_csk_ca(sk); | ||
86 | u32 vrtt = usrtt + 1; /* Never allow zero rtt or baseRTT */ | ||
87 | |||
88 | /* Filter to find propagation delay: */ | ||
89 | if (vrtt < vegas->baseRTT) | ||
90 | vegas->baseRTT = vrtt; | ||
91 | |||
92 | /* Find the min RTT during the last RTT to find | ||
93 | * the current prop. delay + queuing delay: | ||
94 | */ | ||
95 | vegas->minRTT = min(vegas->minRTT, vrtt); | ||
96 | vegas->cntRTT++; | ||
97 | } | ||
98 | |||
99 | /* | ||
100 | * If the connection is idle and we are restarting, | ||
101 | * then we don't want to do any Vegas calculations | ||
102 | * until we get fresh RTT samples. So when we | ||
103 | * restart, we reset our Vegas state to a clean | ||
104 | * slate. After we get acks for this flight of | ||
105 | * packets, _then_ we can make Vegas calculations | ||
106 | * again. | ||
107 | */ | ||
108 | static void tcp_vegas_cwnd_event(struct sock *sk, enum tcp_ca_event event) | ||
109 | { | ||
110 | if (event == CA_EVENT_CWND_RESTART || | ||
111 | event == CA_EVENT_TX_START) | ||
112 | tcp_vegas_init(sk); | ||
113 | } | ||
114 | |||
115 | /* Extract info for Tcp socket info provided via netlink. */ | ||
116 | static void tcp_vegas_get_info(struct sock *sk, u32 ext, | ||
117 | struct sk_buff *skb) | ||
118 | { | ||
119 | const struct vegas *ca = inet_csk_ca(sk); | ||
120 | if (ext & (1 << (INET_DIAG_VEGASINFO - 1))) { | ||
121 | struct tcpvegas_info *info; | ||
122 | |||
123 | info = RTA_DATA(__RTA_PUT(skb, INET_DIAG_VEGASINFO, | ||
124 | sizeof(*info))); | ||
125 | |||
126 | info->tcpv_enabled = ca->doing_vegas_now; | ||
127 | info->tcpv_rttcnt = ca->cntRTT; | ||
128 | info->tcpv_rtt = ca->baseRTT; | ||
129 | info->tcpv_minrtt = ca->minRTT; | ||
130 | rtattr_failure: ; | ||
131 | } | ||
132 | } | ||
133 | |||
134 | |||