Line data Source code
1 : /*
2 : * Copyright (c) 2018-2019 Cisco and/or its affiliates.
3 : * Licensed under the Apache License, Version 2.0 (the "License");
4 : * you may not use this file except in compliance with the License.
5 : * You may obtain a copy of the License at:
6 : *
7 : * http://www.apache.org/licenses/LICENSE-2.0
8 : *
9 : * Unless required by applicable law or agreed to in writing, software
10 : * distributed under the License is distributed on an "AS IS" BASIS,
11 : * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12 : * See the License for the specific language governing permissions and
13 : * limitations under the License.
14 : */
15 :
16 : #include <vnet/tcp/tcp.h>
17 : #include <vnet/tcp/tcp_inlines.h>
18 : #include <math.h>
19 :
20 : #define beta_cubic 0.7
21 : #define cubic_c 0.4
22 : #define west_const (3 * (1 - beta_cubic) / (1 + beta_cubic))
23 :
24 : typedef struct cubic_cfg_
25 : {
26 : u8 fast_convergence;
27 : u32 ssthresh;
28 : } cubic_cfg_t;
29 :
30 : static cubic_cfg_t cubic_cfg = {
31 : .fast_convergence = 1,
32 : .ssthresh = 0x7FFFFFFFU,
33 : };
34 :
35 : typedef struct cubic_data_
36 : {
37 : /** time period (in seconds) needed to increase the current window
38 : * size to W_max if there are no further congestion events */
39 : f64 K;
40 :
41 : /** time (in sec) since the start of current congestion avoidance */
42 : f64 t_start;
43 :
44 : /** Inflection point of the cubic function (in snd_mss segments) */
45 : u32 w_max;
46 :
47 : } __clib_packed cubic_data_t;
48 :
49 : STATIC_ASSERT (sizeof (cubic_data_t) <= TCP_CC_DATA_SZ, "cubic data len");
50 :
51 : static inline f64
52 42194 : cubic_time (u32 thread_index)
53 : {
54 42194 : return tcp_time_now_us (thread_index);
55 : }
56 :
57 : /**
58 : * RFC 8312 Eq. 1
59 : *
60 : * CUBIC window increase function. Time and K need to be provided in seconds.
61 : */
62 : static inline u64
63 0 : W_cubic (cubic_data_t * cd, f64 t)
64 : {
65 0 : f64 diff = t - cd->K;
66 :
67 : /* W_cubic(t) = C*(t-K)^3 + W_max */
68 0 : return cubic_c * diff * diff * diff + cd->w_max;
69 : }
70 :
71 : /**
72 : * RFC 8312 Eq. 2
73 : */
74 : static inline f64
75 0 : K_cubic (cubic_data_t * cd, u32 wnd)
76 : {
77 : /* K = cubic_root(W_max*(1-beta_cubic)/C)
78 : * Because the current window may be less than W_max * beta_cubic because
79 : * of fast convergence, we pass it as parameter */
80 0 : return pow ((f64) (cd->w_max - wnd) / cubic_c, 1 / 3.0);
81 : }
82 :
83 : /**
84 : * RFC 8312 Eq. 4
85 : *
86 : * Estimates the window size of AIMD(alpha_aimd, beta_aimd) for
87 : * alpha_aimd=3*(1-beta_cubic)/(1+beta_cubic) and beta_aimd=beta_cubic.
88 : * Time (t) and rtt should be provided in seconds
89 : */
90 : static inline u32
91 0 : W_est (cubic_data_t * cd, f64 t, f64 rtt)
92 : {
93 : /* W_est(t) = W_max*beta_cubic+[3*(1-beta_cubic)/(1+beta_cubic)]*(t/RTT) */
94 0 : return cd->w_max * beta_cubic + west_const * (t / rtt);
95 : }
96 :
97 : static void
98 0 : cubic_congestion (tcp_connection_t * tc)
99 : {
100 0 : cubic_data_t *cd = (cubic_data_t *) tcp_cc_data (tc);
101 : u32 w_max;
102 :
103 0 : w_max = tc->cwnd / tc->snd_mss;
104 0 : if (cubic_cfg.fast_convergence && w_max < cd->w_max)
105 0 : w_max = w_max * ((1.0 + beta_cubic) / 2.0);
106 :
107 0 : cd->w_max = w_max;
108 0 : tc->ssthresh = clib_max (tc->cwnd * beta_cubic, 2 * tc->snd_mss);
109 0 : tc->cwnd = tc->ssthresh;
110 0 : }
111 :
112 : static void
113 0 : cubic_loss (tcp_connection_t * tc)
114 : {
115 0 : cubic_data_t *cd = (cubic_data_t *) tcp_cc_data (tc);
116 :
117 0 : tc->cwnd = tcp_loss_wnd (tc);
118 0 : cd->t_start = cubic_time (tc->c_thread_index);
119 0 : cd->K = 0;
120 0 : cd->w_max = tc->cwnd / tc->snd_mss;
121 0 : }
122 :
123 : static void
124 0 : cubic_recovered (tcp_connection_t * tc)
125 : {
126 0 : cubic_data_t *cd = (cubic_data_t *) tcp_cc_data (tc);
127 0 : cd->t_start = cubic_time (tc->c_thread_index);
128 0 : tc->cwnd = tc->ssthresh;
129 0 : cd->K = K_cubic (cd, tc->cwnd / tc->snd_mss);
130 0 : }
131 :
132 : static void
133 0 : cubic_cwnd_accumulate (tcp_connection_t * tc, u32 thresh, u32 bytes_acked)
134 : {
135 : /* We just updated the threshold and don't know how large the previous
136 : * one was. Still, optimistically increase cwnd by one segment and
137 : * clear the accumulated bytes. */
138 0 : if (tc->cwnd_acc_bytes > thresh)
139 : {
140 0 : tc->cwnd += tc->snd_mss;
141 0 : tc->cwnd_acc_bytes = 0;
142 : }
143 :
144 0 : tcp_cwnd_accumulate (tc, thresh, bytes_acked);
145 0 : }
146 :
147 : static void
148 1061680 : cubic_rcv_ack (tcp_connection_t * tc, tcp_rate_sample_t * rs)
149 : {
150 1061680 : cubic_data_t *cd = (cubic_data_t *) tcp_cc_data (tc);
151 : u64 w_cubic, w_aimd;
152 : f64 t, rtt_sec;
153 : u32 thresh;
154 :
155 : /* Constrained by tx fifo, can't grow further */
156 1061680 : if (tc->cwnd >= tc->tx_fifo_size)
157 74090 : return;
158 :
159 987590 : if (tcp_in_slowstart (tc))
160 : {
161 987590 : tc->cwnd += rs->delivered;
162 987590 : return;
163 : }
164 :
165 0 : t = cubic_time (tc->c_thread_index) - cd->t_start;
166 0 : rtt_sec = clib_min (tc->mrtt_us, (f64) tc->srtt * TCP_TICK);
167 :
168 0 : w_cubic = W_cubic (cd, t + rtt_sec) * tc->snd_mss;
169 0 : w_aimd = (u64) W_est (cd, t, rtt_sec) * tc->snd_mss;
170 0 : if (w_cubic < w_aimd)
171 : {
172 0 : cubic_cwnd_accumulate (tc, tc->cwnd, rs->delivered);
173 : }
174 : else
175 : {
176 0 : if (w_cubic > tc->cwnd)
177 : {
178 : /* For NewReno and slow start, we increment cwnd based on the
179 : * number of bytes acked, not the number of acks received. In
180 : * particular, for NewReno we increment the cwnd by 1 snd_mss
181 : * only after we accumulate 1 cwnd of acked bytes (RFC 3465).
182 : *
183 : * For Cubic, as per RFC 8312 we should increment cwnd by
184 : * (w_cubic - cwnd)/cwnd for each ack. Instead of using that,
185 : * we compute the number of packets that need to be acked
186 : * before adding snd_mss to cwnd and compute the threshold
187 : */
188 0 : thresh = (tc->snd_mss * tc->cwnd) / (w_cubic - tc->cwnd);
189 :
190 : /* Make sure we don't increase cwnd more often than every segment */
191 0 : thresh = clib_max (thresh, tc->snd_mss);
192 : }
193 : else
194 : {
195 : /* Practically we can't increment so just inflate threshold */
196 0 : thresh = 50 * tc->cwnd;
197 : }
198 0 : cubic_cwnd_accumulate (tc, thresh, rs->delivered);
199 : }
200 : }
201 :
202 : static void
203 267 : cubic_conn_init (tcp_connection_t * tc)
204 : {
205 267 : cubic_data_t *cd = (cubic_data_t *) tcp_cc_data (tc);
206 267 : tc->ssthresh = cubic_cfg.ssthresh;
207 267 : tc->cwnd = tcp_initial_cwnd (tc);
208 267 : cd->w_max = 0;
209 267 : cd->K = 0;
210 267 : cd->t_start = cubic_time (tc->c_thread_index);
211 267 : }
212 :
213 : static uword
214 0 : cubic_unformat_config (unformat_input_t * input)
215 : {
216 0 : u32 ssthresh = 0x7FFFFFFFU;
217 :
218 0 : if (!input)
219 0 : return 0;
220 :
221 0 : unformat_skip_white_space (input);
222 :
223 0 : while (unformat_check_input (input) != UNFORMAT_END_OF_INPUT)
224 : {
225 0 : if (unformat (input, "no-fast-convergence"))
226 0 : cubic_cfg.fast_convergence = 0;
227 0 : else if (unformat (input, "ssthresh %u", &ssthresh))
228 0 : cubic_cfg.ssthresh = ssthresh;
229 : else
230 0 : return 0;
231 : }
232 0 : return 1;
233 : }
234 :
235 : void
236 41927 : cubic_event (tcp_connection_t *tc, tcp_cc_event_t evt)
237 : {
238 : cubic_data_t *cd;
239 : f64 now;
240 :
241 41927 : if (evt != TCP_CC_EVT_START_TX)
242 0 : return;
243 :
244 : /* App was idle so update t_start to avoid artificially
245 : * inflating cwnd if nothing recently sent and acked */
246 41927 : cd = (cubic_data_t *) tcp_cc_data (tc);
247 41927 : now = cubic_time (tc->c_thread_index);
248 41927 : if (now > tc->mrtt_us + 1)
249 41927 : cd->t_start = now;
250 : }
251 :
252 : const static tcp_cc_algorithm_t tcp_cubic = {
253 : .name = "cubic",
254 : .unformat_cfg = cubic_unformat_config,
255 : .congestion = cubic_congestion,
256 : .loss = cubic_loss,
257 : .recovered = cubic_recovered,
258 : .rcv_ack = cubic_rcv_ack,
259 : .rcv_cong_ack = newreno_rcv_cong_ack,
260 : .event = cubic_event,
261 : .init = cubic_conn_init,
262 : };
263 :
264 : clib_error_t *
265 575 : cubic_init (vlib_main_t * vm)
266 : {
267 575 : clib_error_t *error = 0;
268 :
269 575 : tcp_cc_algo_register (TCP_CC_CUBIC, &tcp_cubic);
270 :
271 575 : return error;
272 : }
273 :
274 59327 : VLIB_INIT_FUNCTION (cubic_init);
275 :
276 : /*
277 : * fd.io coding-style-patch-verification: ON
278 : *
279 : * Local Variables:
280 : * eval: (c-set-style "gnu")
281 : * End:
282 : */
|
