PageRenderTime 107ms CodeModel.GetById 17ms RepoModel.GetById 1ms app.codeStats 0ms

/kernel-source/net/ipv4/tcp_cubic.c

https://gitlab.com/karrei/inel-imx6-kernel
C | 497 lines | 313 code | 76 blank | 108 comment | 57 complexity | 90cab540de1b273ba97b0f14f39a2cdc MD5 | raw file
  1. /*
  2. * TCP CUBIC: Binary Increase Congestion control for TCP v2.3
  3. * Home page:
  4. * http://netsrv.csc.ncsu.edu/twiki/bin/view/Main/BIC
  5. * This is from the implementation of CUBIC TCP in
  6. * Sangtae Ha, Injong Rhee and Lisong Xu,
  7. * "CUBIC: A New TCP-Friendly High-Speed TCP Variant"
  8. * in ACM SIGOPS Operating System Review, July 2008.
  9. * Available from:
  10. * http://netsrv.csc.ncsu.edu/export/cubic_a_new_tcp_2008.pdf
  11. *
  12. * CUBIC integrates a new slow start algorithm, called HyStart.
  13. * The details of HyStart are presented in
  14. * Sangtae Ha and Injong Rhee,
  15. * "Taming the Elephants: New TCP Slow Start", NCSU TechReport 2008.
  16. * Available from:
  17. * http://netsrv.csc.ncsu.edu/export/hystart_techreport_2008.pdf
  18. *
  19. * All testing results are available from:
  20. * http://netsrv.csc.ncsu.edu/wiki/index.php/TCP_Testing
  21. *
  22. * Unless CUBIC is enabled and congestion window is large
  23. * this behaves the same as the original Reno.
  24. */
  25. #include <linux/mm.h>
  26. #include <linux/module.h>
  27. #include <linux/math64.h>
  28. #include <net/tcp.h>
  29. #define BICTCP_BETA_SCALE 1024 /* Scale factor beta calculation
  30. * max_cwnd = snd_cwnd * beta
  31. */
  32. #define BICTCP_HZ 10 /* BIC HZ 2^10 = 1024 */
  33. /* Two methods of hybrid slow start */
  34. #define HYSTART_ACK_TRAIN 0x1
  35. #define HYSTART_DELAY 0x2
  36. /* Number of delay samples for detecting the increase of delay */
  37. #define HYSTART_MIN_SAMPLES 8
  38. #define HYSTART_DELAY_MIN (4U<<3)
  39. #define HYSTART_DELAY_MAX (16U<<3)
  40. #define HYSTART_DELAY_THRESH(x) clamp(x, HYSTART_DELAY_MIN, HYSTART_DELAY_MAX)
  41. static int fast_convergence __read_mostly = 1;
  42. static int beta __read_mostly = 717; /* = 717/1024 (BICTCP_BETA_SCALE) */
  43. static int initial_ssthresh __read_mostly;
  44. static int bic_scale __read_mostly = 41;
  45. static int tcp_friendliness __read_mostly = 1;
  46. static int hystart __read_mostly = 1;
  47. static int hystart_detect __read_mostly = HYSTART_ACK_TRAIN | HYSTART_DELAY;
  48. static int hystart_low_window __read_mostly = 16;
  49. static int hystart_ack_delta __read_mostly = 2;
  50. static u32 cube_rtt_scale __read_mostly;
  51. static u32 beta_scale __read_mostly;
  52. static u64 cube_factor __read_mostly;
  53. /* Note parameters that are used for precomputing scale factors are read-only */
  54. module_param(fast_convergence, int, 0644);
  55. MODULE_PARM_DESC(fast_convergence, "turn on/off fast convergence");
  56. module_param(beta, int, 0644);
  57. MODULE_PARM_DESC(beta, "beta for multiplicative increase");
  58. module_param(initial_ssthresh, int, 0644);
  59. MODULE_PARM_DESC(initial_ssthresh, "initial value of slow start threshold");
  60. module_param(bic_scale, int, 0444);
  61. MODULE_PARM_DESC(bic_scale, "scale (scaled by 1024) value for bic function (bic_scale/1024)");
  62. module_param(tcp_friendliness, int, 0644);
  63. MODULE_PARM_DESC(tcp_friendliness, "turn on/off tcp friendliness");
  64. module_param(hystart, int, 0644);
  65. MODULE_PARM_DESC(hystart, "turn on/off hybrid slow start algorithm");
  66. module_param(hystart_detect, int, 0644);
  67. MODULE_PARM_DESC(hystart_detect, "hyrbrid slow start detection mechanisms"
  68. " 1: packet-train 2: delay 3: both packet-train and delay");
  69. module_param(hystart_low_window, int, 0644);
  70. MODULE_PARM_DESC(hystart_low_window, "lower bound cwnd for hybrid slow start");
  71. module_param(hystart_ack_delta, int, 0644);
  72. MODULE_PARM_DESC(hystart_ack_delta, "spacing between ack's indicating train (msecs)");
  73. /* BIC TCP Parameters */
  74. struct bictcp {
  75. u32 cnt; /* increase cwnd by 1 after ACKs */
  76. u32 last_max_cwnd; /* last maximum snd_cwnd */
  77. u32 loss_cwnd; /* congestion window at last loss */
  78. u32 last_cwnd; /* the last snd_cwnd */
  79. u32 last_time; /* time when updated last_cwnd */
  80. u32 bic_origin_point;/* origin point of bic function */
  81. u32 bic_K; /* time to origin point from the beginning of the current epoch */
  82. u32 delay_min; /* min delay (msec << 3) */
  83. u32 epoch_start; /* beginning of an epoch */
  84. u32 ack_cnt; /* number of acks */
  85. u32 tcp_cwnd; /* estimated tcp cwnd */
  86. #define ACK_RATIO_SHIFT 4
  87. #define ACK_RATIO_LIMIT (32u << ACK_RATIO_SHIFT)
  88. u16 delayed_ack; /* estimate the ratio of Packets/ACKs << 4 */
  89. u8 sample_cnt; /* number of samples to decide curr_rtt */
  90. u8 found; /* the exit point is found? */
  91. u32 round_start; /* beginning of each round */
  92. u32 end_seq; /* end_seq of the round */
  93. u32 last_ack; /* last time when the ACK spacing is close */
  94. u32 curr_rtt; /* the minimum rtt of current round */
  95. };
  96. static inline void bictcp_reset(struct bictcp *ca)
  97. {
  98. ca->cnt = 0;
  99. ca->last_max_cwnd = 0;
  100. ca->last_cwnd = 0;
  101. ca->last_time = 0;
  102. ca->bic_origin_point = 0;
  103. ca->bic_K = 0;
  104. ca->delay_min = 0;
  105. ca->epoch_start = 0;
  106. ca->delayed_ack = 2 << ACK_RATIO_SHIFT;
  107. ca->ack_cnt = 0;
  108. ca->tcp_cwnd = 0;
  109. ca->found = 0;
  110. }
  111. static inline u32 bictcp_clock(void)
  112. {
  113. #if HZ < 1000
  114. return ktime_to_ms(ktime_get_real());
  115. #else
  116. return jiffies_to_msecs(jiffies);
  117. #endif
  118. }
  119. static inline void bictcp_hystart_reset(struct sock *sk)
  120. {
  121. struct tcp_sock *tp = tcp_sk(sk);
  122. struct bictcp *ca = inet_csk_ca(sk);
  123. ca->round_start = ca->last_ack = bictcp_clock();
  124. ca->end_seq = tp->snd_nxt;
  125. ca->curr_rtt = 0;
  126. ca->sample_cnt = 0;
  127. }
  128. static void bictcp_init(struct sock *sk)
  129. {
  130. struct bictcp *ca = inet_csk_ca(sk);
  131. bictcp_reset(ca);
  132. ca->loss_cwnd = 0;
  133. if (hystart)
  134. bictcp_hystart_reset(sk);
  135. if (!hystart && initial_ssthresh)
  136. tcp_sk(sk)->snd_ssthresh = initial_ssthresh;
  137. }
  138. /* calculate the cubic root of x using a table lookup followed by one
  139. * Newton-Raphson iteration.
  140. * Avg err ~= 0.195%
  141. */
  142. static u32 cubic_root(u64 a)
  143. {
  144. u32 x, b, shift;
  145. /*
  146. * cbrt(x) MSB values for x MSB values in [0..63].
  147. * Precomputed then refined by hand - Willy Tarreau
  148. *
  149. * For x in [0..63],
  150. * v = cbrt(x << 18) - 1
  151. * cbrt(x) = (v[x] + 10) >> 6
  152. */
  153. static const u8 v[] = {
  154. /* 0x00 */ 0, 54, 54, 54, 118, 118, 118, 118,
  155. /* 0x08 */ 123, 129, 134, 138, 143, 147, 151, 156,
  156. /* 0x10 */ 157, 161, 164, 168, 170, 173, 176, 179,
  157. /* 0x18 */ 181, 185, 187, 190, 192, 194, 197, 199,
  158. /* 0x20 */ 200, 202, 204, 206, 209, 211, 213, 215,
  159. /* 0x28 */ 217, 219, 221, 222, 224, 225, 227, 229,
  160. /* 0x30 */ 231, 232, 234, 236, 237, 239, 240, 242,
  161. /* 0x38 */ 244, 245, 246, 248, 250, 251, 252, 254,
  162. };
  163. b = fls64(a);
  164. if (b < 7) {
  165. /* a in [0..63] */
  166. return ((u32)v[(u32)a] + 35) >> 6;
  167. }
  168. b = ((b * 84) >> 8) - 1;
  169. shift = (a >> (b * 3));
  170. x = ((u32)(((u32)v[shift] + 10) << b)) >> 6;
  171. /*
  172. * Newton-Raphson iteration
  173. * 2
  174. * x = ( 2 * x + a / x ) / 3
  175. * k+1 k k
  176. */
  177. x = (2 * x + (u32)div64_u64(a, (u64)x * (u64)(x - 1)));
  178. x = ((x * 341) >> 10);
  179. return x;
  180. }
  181. /*
  182. * Compute congestion window to use.
  183. */
  184. static inline void bictcp_update(struct bictcp *ca, u32 cwnd)
  185. {
  186. u32 delta, bic_target, max_cnt;
  187. u64 offs, t;
  188. ca->ack_cnt++; /* count the number of ACKs */
  189. if (ca->last_cwnd == cwnd &&
  190. (s32)(tcp_time_stamp - ca->last_time) <= HZ / 32)
  191. return;
  192. ca->last_cwnd = cwnd;
  193. ca->last_time = tcp_time_stamp;
  194. if (ca->epoch_start == 0) {
  195. ca->epoch_start = tcp_time_stamp; /* record the beginning of an epoch */
  196. ca->ack_cnt = 1; /* start counting */
  197. ca->tcp_cwnd = cwnd; /* syn with cubic */
  198. if (ca->last_max_cwnd <= cwnd) {
  199. ca->bic_K = 0;
  200. ca->bic_origin_point = cwnd;
  201. } else {
  202. /* Compute new K based on
  203. * (wmax-cwnd) * (srtt>>3 / HZ) / c * 2^(3*bictcp_HZ)
  204. */
  205. ca->bic_K = cubic_root(cube_factor
  206. * (ca->last_max_cwnd - cwnd));
  207. ca->bic_origin_point = ca->last_max_cwnd;
  208. }
  209. }
  210. /* cubic function - calc*/
  211. /* calculate c * time^3 / rtt,
  212. * while considering overflow in calculation of time^3
  213. * (so time^3 is done by using 64 bit)
  214. * and without the support of division of 64bit numbers
  215. * (so all divisions are done by using 32 bit)
  216. * also NOTE the unit of those veriables
  217. * time = (t - K) / 2^bictcp_HZ
  218. * c = bic_scale >> 10
  219. * rtt = (srtt >> 3) / HZ
  220. * !!! The following code does not have overflow problems,
  221. * if the cwnd < 1 million packets !!!
  222. */
  223. t = (s32)(tcp_time_stamp - ca->epoch_start);
  224. t += msecs_to_jiffies(ca->delay_min >> 3);
  225. /* change the unit from HZ to bictcp_HZ */
  226. t <<= BICTCP_HZ;
  227. do_div(t, HZ);
  228. if (t < ca->bic_K) /* t - K */
  229. offs = ca->bic_K - t;
  230. else
  231. offs = t - ca->bic_K;
  232. /* c/rtt * (t-K)^3 */
  233. delta = (cube_rtt_scale * offs * offs * offs) >> (10+3*BICTCP_HZ);
  234. if (t < ca->bic_K) /* below origin*/
  235. bic_target = ca->bic_origin_point - delta;
  236. else /* above origin*/
  237. bic_target = ca->bic_origin_point + delta;
  238. /* cubic function - calc bictcp_cnt*/
  239. if (bic_target > cwnd) {
  240. ca->cnt = cwnd / (bic_target - cwnd);
  241. } else {
  242. ca->cnt = 100 * cwnd; /* very small increment*/
  243. }
  244. /*
  245. * The initial growth of cubic function may be too conservative
  246. * when the available bandwidth is still unknown.
  247. */
  248. if (ca->last_max_cwnd == 0 && ca->cnt > 20)
  249. ca->cnt = 20; /* increase cwnd 5% per RTT */
  250. /* TCP Friendly */
  251. if (tcp_friendliness) {
  252. u32 scale = beta_scale;
  253. delta = (cwnd * scale) >> 3;
  254. while (ca->ack_cnt > delta) { /* update tcp cwnd */
  255. ca->ack_cnt -= delta;
  256. ca->tcp_cwnd++;
  257. }
  258. if (ca->tcp_cwnd > cwnd){ /* if bic is slower than tcp */
  259. delta = ca->tcp_cwnd - cwnd;
  260. max_cnt = cwnd / delta;
  261. if (ca->cnt > max_cnt)
  262. ca->cnt = max_cnt;
  263. }
  264. }
  265. ca->cnt = (ca->cnt << ACK_RATIO_SHIFT) / ca->delayed_ack;
  266. if (ca->cnt == 0) /* cannot be zero */
  267. ca->cnt = 1;
  268. }
  269. static void bictcp_cong_avoid(struct sock *sk, u32 ack, u32 acked,
  270. u32 in_flight)
  271. {
  272. struct tcp_sock *tp = tcp_sk(sk);
  273. struct bictcp *ca = inet_csk_ca(sk);
  274. if (!tcp_is_cwnd_limited(sk, in_flight))
  275. return;
  276. if (tp->snd_cwnd <= tp->snd_ssthresh) {
  277. if (hystart && after(ack, ca->end_seq))
  278. bictcp_hystart_reset(sk);
  279. tcp_slow_start(tp, acked);
  280. } else {
  281. bictcp_update(ca, tp->snd_cwnd);
  282. tcp_cong_avoid_ai(tp, ca->cnt);
  283. }
  284. }
  285. static u32 bictcp_recalc_ssthresh(struct sock *sk)
  286. {
  287. const struct tcp_sock *tp = tcp_sk(sk);
  288. struct bictcp *ca = inet_csk_ca(sk);
  289. ca->epoch_start = 0; /* end of epoch */
  290. /* Wmax and fast convergence */
  291. if (tp->snd_cwnd < ca->last_max_cwnd && fast_convergence)
  292. ca->last_max_cwnd = (tp->snd_cwnd * (BICTCP_BETA_SCALE + beta))
  293. / (2 * BICTCP_BETA_SCALE);
  294. else
  295. ca->last_max_cwnd = tp->snd_cwnd;
  296. ca->loss_cwnd = tp->snd_cwnd;
  297. return max((tp->snd_cwnd * beta) / BICTCP_BETA_SCALE, 2U);
  298. }
  299. static u32 bictcp_undo_cwnd(struct sock *sk)
  300. {
  301. struct bictcp *ca = inet_csk_ca(sk);
  302. return max(tcp_sk(sk)->snd_cwnd, ca->loss_cwnd);
  303. }
  304. static void bictcp_state(struct sock *sk, u8 new_state)
  305. {
  306. if (new_state == TCP_CA_Loss) {
  307. bictcp_reset(inet_csk_ca(sk));
  308. bictcp_hystart_reset(sk);
  309. }
  310. }
  311. static void hystart_update(struct sock *sk, u32 delay)
  312. {
  313. struct tcp_sock *tp = tcp_sk(sk);
  314. struct bictcp *ca = inet_csk_ca(sk);
  315. if (!(ca->found & hystart_detect)) {
  316. u32 now = bictcp_clock();
  317. /* first detection parameter - ack-train detection */
  318. if ((s32)(now - ca->last_ack) <= hystart_ack_delta) {
  319. ca->last_ack = now;
  320. if ((s32)(now - ca->round_start) > ca->delay_min >> 4)
  321. ca->found |= HYSTART_ACK_TRAIN;
  322. }
  323. /* obtain the minimum delay of more than sampling packets */
  324. if (ca->sample_cnt < HYSTART_MIN_SAMPLES) {
  325. if (ca->curr_rtt == 0 || ca->curr_rtt > delay)
  326. ca->curr_rtt = delay;
  327. ca->sample_cnt++;
  328. } else {
  329. if (ca->curr_rtt > ca->delay_min +
  330. HYSTART_DELAY_THRESH(ca->delay_min>>4))
  331. ca->found |= HYSTART_DELAY;
  332. }
  333. /*
  334. * Either one of two conditions are met,
  335. * we exit from slow start immediately.
  336. */
  337. if (ca->found & hystart_detect)
  338. tp->snd_ssthresh = tp->snd_cwnd;
  339. }
  340. }
  341. /* Track delayed acknowledgment ratio using sliding window
  342. * ratio = (15*ratio + sample) / 16
  343. */
  344. static void bictcp_acked(struct sock *sk, u32 cnt, s32 rtt_us)
  345. {
  346. const struct inet_connection_sock *icsk = inet_csk(sk);
  347. const struct tcp_sock *tp = tcp_sk(sk);
  348. struct bictcp *ca = inet_csk_ca(sk);
  349. u32 delay;
  350. if (icsk->icsk_ca_state == TCP_CA_Open) {
  351. u32 ratio = ca->delayed_ack;
  352. ratio -= ca->delayed_ack >> ACK_RATIO_SHIFT;
  353. ratio += cnt;
  354. ca->delayed_ack = clamp(ratio, 1U, ACK_RATIO_LIMIT);
  355. }
  356. /* Some calls are for duplicates without timetamps */
  357. if (rtt_us < 0)
  358. return;
  359. /* Discard delay samples right after fast recovery */
  360. if (ca->epoch_start && (s32)(tcp_time_stamp - ca->epoch_start) < HZ)
  361. return;
  362. delay = (rtt_us << 3) / USEC_PER_MSEC;
  363. if (delay == 0)
  364. delay = 1;
  365. /* first time call or link delay decreases */
  366. if (ca->delay_min == 0 || ca->delay_min > delay)
  367. ca->delay_min = delay;
  368. /* hystart triggers when cwnd is larger than some threshold */
  369. if (hystart && tp->snd_cwnd <= tp->snd_ssthresh &&
  370. tp->snd_cwnd >= hystart_low_window)
  371. hystart_update(sk, delay);
  372. }
  373. static struct tcp_congestion_ops cubictcp __read_mostly = {
  374. .init = bictcp_init,
  375. .ssthresh = bictcp_recalc_ssthresh,
  376. .cong_avoid = bictcp_cong_avoid,
  377. .set_state = bictcp_state,
  378. .undo_cwnd = bictcp_undo_cwnd,
  379. .pkts_acked = bictcp_acked,
  380. .owner = THIS_MODULE,
  381. .name = "cubic",
  382. };
  383. static int __init cubictcp_register(void)
  384. {
  385. BUILD_BUG_ON(sizeof(struct bictcp) > ICSK_CA_PRIV_SIZE);
  386. /* Precompute a bunch of the scaling factors that are used per-packet
  387. * based on SRTT of 100ms
  388. */
  389. beta_scale = 8*(BICTCP_BETA_SCALE+beta)/ 3 / (BICTCP_BETA_SCALE - beta);
  390. cube_rtt_scale = (bic_scale * 10); /* 1024*c/rtt */
  391. /* calculate the "K" for (wmax-cwnd) = c/rtt * K^3
  392. * so K = cubic_root( (wmax-cwnd)*rtt/c )
  393. * the unit of K is bictcp_HZ=2^10, not HZ
  394. *
  395. * c = bic_scale >> 10
  396. * rtt = 100ms
  397. *
  398. * the following code has been designed and tested for
  399. * cwnd < 1 million packets
  400. * RTT < 100 seconds
  401. * HZ < 1,000,00 (corresponding to 10 nano-second)
  402. */
  403. /* 1/c * 2^2*bictcp_HZ * srtt */
  404. cube_factor = 1ull << (10+3*BICTCP_HZ); /* 2^40 */
  405. /* divide by bic_scale and by constant Srtt (100ms) */
  406. do_div(cube_factor, bic_scale * 10);
  407. /* hystart needs ms clock resolution */
  408. if (hystart && HZ < 1000)
  409. cubictcp.flags |= TCP_CONG_RTT_STAMP;
  410. return tcp_register_congestion_control(&cubictcp);
  411. }
  412. static void __exit cubictcp_unregister(void)
  413. {
  414. tcp_unregister_congestion_control(&cubictcp);
  415. }
  416. module_init(cubictcp_register);
  417. module_exit(cubictcp_unregister);
  418. MODULE_AUTHOR("Sangtae Ha, Stephen Hemminger");
  419. MODULE_LICENSE("GPL");
  420. MODULE_DESCRIPTION("CUBIC TCP");
  421. MODULE_VERSION("2.3");