2 * TCP CUBIC: Binary Increase Congestion control for TCP v2.0
4 * This is from the implementation of CUBIC TCP in
5 * Injong Rhee, Lisong Xu.
6 * "CUBIC: A New TCP-Friendly High-Speed TCP Variant
9 * http://www.csc.ncsu.edu/faculty/rhee/export/bitcp/cubic-paper.pdf
11 * Unless CUBIC is enabled and congestion window is large
12 * this behaves the same as the original Reno.
16 #include <linux/module.h>
18 #include <asm/div64.h>
20 #define BICTCP_BETA_SCALE 1024 /* Scale factor beta calculation
21 * max_cwnd = snd_cwnd * beta
25 * go to point (max+min)/N
27 #define BICTCP_HZ 10 /* BIC HZ 2^10 = 1024 */
29 static int fast_convergence __read_mostly = 1;
30 static int max_increment __read_mostly = 16;
31 static int beta __read_mostly = 819; /* = 819/1024 (BICTCP_BETA_SCALE) */
32 static int initial_ssthresh __read_mostly = 100;
33 static int bic_scale __read_mostly = 41;
34 static int tcp_friendliness __read_mostly = 1;
36 static u32 cube_rtt_scale __read_mostly;
37 static u32 beta_scale __read_mostly;
38 static u64 cube_factor __read_mostly;
40 /* Note parameters that are used for precomputing scale factors are read-only */
41 module_param(fast_convergence, int, 0644);
42 MODULE_PARM_DESC(fast_convergence, "turn on/off fast convergence");
43 module_param(max_increment, int, 0644);
44 MODULE_PARM_DESC(max_increment, "Limit on increment allowed during binary search");
45 module_param(beta, int, 0444);
46 MODULE_PARM_DESC(beta, "beta for multiplicative increase");
47 module_param(initial_ssthresh, int, 0644);
48 MODULE_PARM_DESC(initial_ssthresh, "initial value of slow start threshold");
49 module_param(bic_scale, int, 0444);
50 MODULE_PARM_DESC(bic_scale, "scale (scaled by 1024) value for bic function (bic_scale/1024)");
51 module_param(tcp_friendliness, int, 0644);
52 MODULE_PARM_DESC(tcp_friendliness, "turn on/off tcp friendliness");
54 /* BIC TCP Parameters */
56 u32 cnt; /* increase cwnd by 1 after ACKs */
57 u32 last_max_cwnd; /* last maximum snd_cwnd */
58 u32 loss_cwnd; /* congestion window at last loss */
59 u32 last_cwnd; /* the last snd_cwnd */
60 u32 last_time; /* time when updated last_cwnd */
61 u32 bic_origin_point;/* origin point of bic function */
62 u32 bic_K; /* time to origin point from the beginning of the current epoch */
63 u32 delay_min; /* min delay */
64 u32 epoch_start; /* beginning of an epoch */
65 u32 ack_cnt; /* number of acks */
66 u32 tcp_cwnd; /* estimated tcp cwnd */
67 #define ACK_RATIO_SHIFT 4
68 u32 delayed_ack; /* estimate the ratio of Packets/ACKs << 4 */
71 static inline void bictcp_reset(struct bictcp *ca)
74 ca->last_max_cwnd = 0;
78 ca->bic_origin_point = 0;
82 ca->delayed_ack = 2 << ACK_RATIO_SHIFT;
87 static void bictcp_init(struct sock *sk)
89 bictcp_reset(inet_csk_ca(sk));
91 tcp_sk(sk)->snd_ssthresh = initial_ssthresh;
95 * calculate the cubic root of x using Newton-Raphson
97 static u32 cubic_root(u64 a)
101 /* Initial estimate is based on:
102 * cbrt(x) = exp(log(x) / 3)
104 x = 1u << (fls64(a)/3);
107 * Iteration based on:
109 * x = ( 2 * x + a / x ) / 3
114 x = (2 * x + (uint32_t) div64_64(a, x*x)) / 3;
115 } while (abs(x1 - x) > 1);
121 * Compute congestion window to use.
123 static inline void bictcp_update(struct bictcp *ca, u32 cwnd)
126 u32 delta, t, bic_target, min_cnt, max_cnt;
128 ca->ack_cnt++; /* count the number of ACKs */
130 if (ca->last_cwnd == cwnd &&
131 (s32)(tcp_time_stamp - ca->last_time) <= HZ / 32)
134 ca->last_cwnd = cwnd;
135 ca->last_time = tcp_time_stamp;
137 if (ca->epoch_start == 0) {
138 ca->epoch_start = tcp_time_stamp; /* record the beginning of an epoch */
139 ca->ack_cnt = 1; /* start counting */
140 ca->tcp_cwnd = cwnd; /* syn with cubic */
142 if (ca->last_max_cwnd <= cwnd) {
144 ca->bic_origin_point = cwnd;
146 /* Compute new K based on
147 * (wmax-cwnd) * (srtt>>3 / HZ) / c * 2^(3*bictcp_HZ)
149 ca->bic_K = cubic_root(cube_factor
150 * (ca->last_max_cwnd - cwnd));
151 ca->bic_origin_point = ca->last_max_cwnd;
155 /* cubic function - calc*/
156 /* calculate c * time^3 / rtt,
157 * while considering overflow in calculation of time^3
158 * (so time^3 is done by using 64 bit)
159 * and without the support of division of 64bit numbers
160 * (so all divisions are done by using 32 bit)
161 * also NOTE the unit of those veriables
162 * time = (t - K) / 2^bictcp_HZ
163 * c = bic_scale >> 10
164 * rtt = (srtt >> 3) / HZ
165 * !!! The following code does not have overflow problems,
166 * if the cwnd < 1 million packets !!!
169 /* change the unit from HZ to bictcp_HZ */
170 t = ((tcp_time_stamp + (ca->delay_min>>3) - ca->epoch_start)
173 if (t < ca->bic_K) /* t - K */
174 offs = ca->bic_K - t;
176 offs = t - ca->bic_K;
178 /* c/rtt * (t-K)^3 */
179 delta = (cube_rtt_scale * offs * offs * offs) >> (10+3*BICTCP_HZ);
180 if (t < ca->bic_K) /* below origin*/
181 bic_target = ca->bic_origin_point - delta;
182 else /* above origin*/
183 bic_target = ca->bic_origin_point + delta;
185 /* cubic function - calc bictcp_cnt*/
186 if (bic_target > cwnd) {
187 ca->cnt = cwnd / (bic_target - cwnd);
189 ca->cnt = 100 * cwnd; /* very small increment*/
192 if (ca->delay_min > 0) {
193 /* max increment = Smax * rtt / 0.1 */
194 min_cnt = (cwnd * HZ * 8)/(10 * max_increment * ca->delay_min);
195 if (ca->cnt < min_cnt)
199 /* slow start and low utilization */
200 if (ca->loss_cwnd == 0) /* could be aggressive in slow start */
204 if (tcp_friendliness) {
205 u32 scale = beta_scale;
206 delta = (cwnd * scale) >> 3;
207 while (ca->ack_cnt > delta) { /* update tcp cwnd */
208 ca->ack_cnt -= delta;
212 if (ca->tcp_cwnd > cwnd){ /* if bic is slower than tcp */
213 delta = ca->tcp_cwnd - cwnd;
214 max_cnt = cwnd / delta;
215 if (ca->cnt > max_cnt)
220 ca->cnt = (ca->cnt << ACK_RATIO_SHIFT) / ca->delayed_ack;
221 if (ca->cnt == 0) /* cannot be zero */
226 /* Keep track of minimum rtt */
227 static inline void measure_delay(struct sock *sk)
229 const struct tcp_sock *tp = tcp_sk(sk);
230 struct bictcp *ca = inet_csk_ca(sk);
234 if (!(tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr) ||
235 /* Discard delay samples right after fast recovery */
236 (s32)(tcp_time_stamp - ca->epoch_start) < HZ)
239 delay = (tcp_time_stamp - tp->rx_opt.rcv_tsecr)<<3;
243 /* first time call or link delay decreases */
244 if (ca->delay_min == 0 || ca->delay_min > delay)
245 ca->delay_min = delay;
248 static void bictcp_cong_avoid(struct sock *sk, u32 ack,
249 u32 seq_rtt, u32 in_flight, int data_acked)
251 struct tcp_sock *tp = tcp_sk(sk);
252 struct bictcp *ca = inet_csk_ca(sk);
257 if (!tcp_is_cwnd_limited(sk, in_flight))
260 if (tp->snd_cwnd <= tp->snd_ssthresh)
263 bictcp_update(ca, tp->snd_cwnd);
265 /* In dangerous area, increase slowly.
266 * In theory this is tp->snd_cwnd += 1 / tp->snd_cwnd
268 if (tp->snd_cwnd_cnt >= ca->cnt) {
269 if (tp->snd_cwnd < tp->snd_cwnd_clamp)
271 tp->snd_cwnd_cnt = 0;
278 static u32 bictcp_recalc_ssthresh(struct sock *sk)
280 const struct tcp_sock *tp = tcp_sk(sk);
281 struct bictcp *ca = inet_csk_ca(sk);
283 ca->epoch_start = 0; /* end of epoch */
285 /* Wmax and fast convergence */
286 if (tp->snd_cwnd < ca->last_max_cwnd && fast_convergence)
287 ca->last_max_cwnd = (tp->snd_cwnd * (BICTCP_BETA_SCALE + beta))
288 / (2 * BICTCP_BETA_SCALE);
290 ca->last_max_cwnd = tp->snd_cwnd;
292 ca->loss_cwnd = tp->snd_cwnd;
294 return max((tp->snd_cwnd * beta) / BICTCP_BETA_SCALE, 2U);
297 static u32 bictcp_undo_cwnd(struct sock *sk)
299 struct bictcp *ca = inet_csk_ca(sk);
301 return max(tcp_sk(sk)->snd_cwnd, ca->last_max_cwnd);
304 static void bictcp_state(struct sock *sk, u8 new_state)
306 if (new_state == TCP_CA_Loss)
307 bictcp_reset(inet_csk_ca(sk));
310 /* Track delayed acknowledgment ratio using sliding window
311 * ratio = (15*ratio + sample) / 16
313 static void bictcp_acked(struct sock *sk, u32 cnt)
315 const struct inet_connection_sock *icsk = inet_csk(sk);
317 if (cnt > 0 && icsk->icsk_ca_state == TCP_CA_Open) {
318 struct bictcp *ca = inet_csk_ca(sk);
319 cnt -= ca->delayed_ack >> ACK_RATIO_SHIFT;
320 ca->delayed_ack += cnt;
325 static struct tcp_congestion_ops cubictcp = {
327 .ssthresh = bictcp_recalc_ssthresh,
328 .cong_avoid = bictcp_cong_avoid,
329 .set_state = bictcp_state,
330 .undo_cwnd = bictcp_undo_cwnd,
331 .pkts_acked = bictcp_acked,
332 .owner = THIS_MODULE,
336 static int __init cubictcp_register(void)
338 BUILD_BUG_ON(sizeof(struct bictcp) > ICSK_CA_PRIV_SIZE);
340 /* Precompute a bunch of the scaling factors that are used per-packet
341 * based on SRTT of 100ms
344 beta_scale = 8*(BICTCP_BETA_SCALE+beta)/ 3 / (BICTCP_BETA_SCALE - beta);
346 cube_rtt_scale = (bic_scale * 10); /* 1024*c/rtt */
348 /* calculate the "K" for (wmax-cwnd) = c/rtt * K^3
349 * so K = cubic_root( (wmax-cwnd)*rtt/c )
350 * the unit of K is bictcp_HZ=2^10, not HZ
352 * c = bic_scale >> 10
355 * the following code has been designed and tested for
356 * cwnd < 1 million packets
358 * HZ < 1,000,00 (corresponding to 10 nano-second)
361 /* 1/c * 2^2*bictcp_HZ * srtt */
362 cube_factor = 1ull << (10+3*BICTCP_HZ); /* 2^40 */
364 /* divide by bic_scale and by constant Srtt (100ms) */
365 do_div(cube_factor, bic_scale * 10);
367 return tcp_register_congestion_control(&cubictcp);
370 static void __exit cubictcp_unregister(void)
372 tcp_unregister_congestion_control(&cubictcp);
375 module_init(cubictcp_register);
376 module_exit(cubictcp_unregister);
378 MODULE_AUTHOR("Sangtae Ha, Stephen Hemminger");
379 MODULE_LICENSE("GPL");
380 MODULE_DESCRIPTION("CUBIC TCP");
381 MODULE_VERSION("2.0");