1216115Slstewart/*- 2216115Slstewart * Copyright (c) 2007-2008 3216115Slstewart * Swinburne University of Technology, Melbourne, Australia 4216115Slstewart * Copyright (c) 2009-2010 Lawrence Stewart <lstewart@freebsd.org> 5216115Slstewart * Copyright (c) 2010 The FreeBSD Foundation 6216115Slstewart * All rights reserved. 7216115Slstewart * 8216115Slstewart * This software was developed at the Centre for Advanced Internet 9220560Slstewart * Architectures, Swinburne University of Technology, by Lawrence Stewart and 10220560Slstewart * James Healy, made possible in part by a grant from the Cisco University 11220560Slstewart * Research Program Fund at Community Foundation Silicon Valley. 12216115Slstewart * 13216115Slstewart * Portions of this software were developed at the Centre for Advanced 14216115Slstewart * Internet Architectures, Swinburne University of Technology, Melbourne, 15216115Slstewart * Australia by David Hayes under sponsorship from the FreeBSD Foundation. 16216115Slstewart * 17216115Slstewart * Redistribution and use in source and binary forms, with or without 18216115Slstewart * modification, are permitted provided that the following conditions 19216115Slstewart * are met: 20216115Slstewart * 1. Redistributions of source code must retain the above copyright 21216115Slstewart * notice, this list of conditions and the following disclaimer. 22216115Slstewart * 2. Redistributions in binary form must reproduce the above copyright 23216115Slstewart * notice, this list of conditions and the following disclaimer in the 24216115Slstewart * documentation and/or other materials provided with the distribution. 25216115Slstewart * 26216115Slstewart * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 27216115Slstewart * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 28216115Slstewart * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 29216115Slstewart * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 30216115Slstewart * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 31216115Slstewart * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 32216115Slstewart * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 33216115Slstewart * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 34216115Slstewart * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 35216115Slstewart * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 36216115Slstewart * SUCH DAMAGE. 37216115Slstewart */ 38216115Slstewart 39216115Slstewart/* 40216115Slstewart * An implementation of the H-TCP congestion control algorithm for FreeBSD, 41216115Slstewart * based on the Internet Draft "draft-leith-tcp-htcp-06.txt" by Leith and 42216115Slstewart * Shorten. Originally released as part of the NewTCP research project at 43220560Slstewart * Swinburne University of Technology's Centre for Advanced Internet 44220560Slstewart * Architectures, Melbourne, Australia, which was made possible in part by a 45220560Slstewart * grant from the Cisco University Research Program Fund at Community Foundation 46220560Slstewart * Silicon Valley. More details are available at: 47216115Slstewart * http://caia.swin.edu.au/urp/newtcp/ 48216115Slstewart */ 49216115Slstewart 50216115Slstewart#include <sys/cdefs.h> 51216115Slstewart__FBSDID("$FreeBSD: releng/10.3/sys/netinet/cc/cc_htcp.c 220592 2011-04-13 11:28:46Z pluknet $"); 52216115Slstewart 53216115Slstewart#include <sys/param.h> 54216115Slstewart#include <sys/kernel.h> 55216115Slstewart#include <sys/limits.h> 56216115Slstewart#include <sys/malloc.h> 57216115Slstewart#include <sys/module.h> 58216115Slstewart#include <sys/socket.h> 59216115Slstewart#include <sys/socketvar.h> 60216115Slstewart#include <sys/sysctl.h> 61216115Slstewart#include <sys/systm.h> 62216115Slstewart 63216115Slstewart#include <net/vnet.h> 64216115Slstewart 65216115Slstewart#include <netinet/cc.h> 66216115Slstewart#include <netinet/tcp_seq.h> 67216115Slstewart#include <netinet/tcp_timer.h> 68216115Slstewart#include <netinet/tcp_var.h> 69216115Slstewart 70216115Slstewart#include <netinet/cc/cc_module.h> 71216115Slstewart 72216115Slstewart/* Fixed point math shifts. */ 73216115Slstewart#define HTCP_SHIFT 8 74216115Slstewart#define HTCP_ALPHA_INC_SHIFT 4 75216115Slstewart 76216115Slstewart#define HTCP_INIT_ALPHA 1 77216115Slstewart#define HTCP_DELTA_L hz /* 1 sec in ticks. */ 78216115Slstewart#define HTCP_MINBETA 128 /* 0.5 << HTCP_SHIFT. */ 79216115Slstewart#define HTCP_MAXBETA 204 /* ~0.8 << HTCP_SHIFT. */ 80216115Slstewart#define HTCP_MINROWE 26 /* ~0.1 << HTCP_SHIFT. */ 81216115Slstewart#define HTCP_MAXROWE 512 /* 2 << HTCP_SHIFT. */ 82216115Slstewart 83216115Slstewart/* RTT_ref (ms) used in the calculation of alpha if RTT scaling is enabled. */ 84216115Slstewart#define HTCP_RTT_REF 100 85216115Slstewart 86216115Slstewart/* Don't trust SRTT until this many samples have been taken. */ 87216115Slstewart#define HTCP_MIN_RTT_SAMPLES 8 88216115Slstewart 89216115Slstewart/* 90216115Slstewart * HTCP_CALC_ALPHA performs a fixed point math calculation to determine the 91216115Slstewart * value of alpha, based on the function defined in the HTCP spec. 92216115Slstewart * 93216115Slstewart * i.e. 1 + 10(delta - delta_l) + ((delta - delta_l) / 2) ^ 2 94216115Slstewart * 95216115Slstewart * "diff" is passed in to the macro as "delta - delta_l" and is expected to be 96216115Slstewart * in units of ticks. 97216115Slstewart * 98216115Slstewart * The joyousnous of fixed point maths means our function implementation looks a 99216115Slstewart * little funky... 100216115Slstewart * 101216115Slstewart * In order to maintain some precision in the calculations, a fixed point shift 102216115Slstewart * HTCP_ALPHA_INC_SHIFT is used to ensure the integer divisions don't 103216115Slstewart * truncate the results too badly. 104216115Slstewart * 105216115Slstewart * The "16" value is the "1" term in the alpha function shifted up by 106216115Slstewart * HTCP_ALPHA_INC_SHIFT 107216115Slstewart * 108216115Slstewart * The "160" value is the "10" multiplier in the alpha function multiplied by 109216115Slstewart * 2^HTCP_ALPHA_INC_SHIFT 110216115Slstewart * 111216115Slstewart * Specifying these as constants reduces the computations required. After 112216115Slstewart * up-shifting all the terms in the function and performing the required 113216115Slstewart * calculations, we down-shift the final result by HTCP_ALPHA_INC_SHIFT to 114216115Slstewart * ensure it is back in the correct range. 115216115Slstewart * 116216115Slstewart * The "hz" terms are required as kernels can be configured to run with 117216115Slstewart * different tick timers, which we have to adjust for in the alpha calculation 118216115Slstewart * (which originally was defined in terms of seconds). 119216115Slstewart * 120216115Slstewart * We also have to be careful to constrain the value of diff such that it won't 121216115Slstewart * overflow whilst performing the calculation. The middle term i.e. (160 * diff) 122216115Slstewart * / hz is the limiting factor in the calculation. We must constrain diff to be 123216115Slstewart * less than the max size of an int divided by the constant 160 figure 124216115Slstewart * i.e. diff < INT_MAX / 160 125216115Slstewart * 126216115Slstewart * NB: Changing HTCP_ALPHA_INC_SHIFT will require you to MANUALLY update the 127216115Slstewart * constants used in this function! 128216115Slstewart */ 129216115Slstewart#define HTCP_CALC_ALPHA(diff) \ 130216115Slstewart((\ 131216115Slstewart (16) + \ 132216115Slstewart ((160 * (diff)) / hz) + \ 133216115Slstewart (((diff) / hz) * (((diff) << HTCP_ALPHA_INC_SHIFT) / (4 * hz))) \ 134216115Slstewart) >> HTCP_ALPHA_INC_SHIFT) 135216115Slstewart 136216115Slstewartstatic void htcp_ack_received(struct cc_var *ccv, uint16_t type); 137216115Slstewartstatic void htcp_cb_destroy(struct cc_var *ccv); 138216115Slstewartstatic int htcp_cb_init(struct cc_var *ccv); 139216115Slstewartstatic void htcp_cong_signal(struct cc_var *ccv, uint32_t type); 140216115Slstewartstatic int htcp_mod_init(void); 141216115Slstewartstatic void htcp_post_recovery(struct cc_var *ccv); 142216115Slstewartstatic void htcp_recalc_alpha(struct cc_var *ccv); 143216115Slstewartstatic void htcp_recalc_beta(struct cc_var *ccv); 144216115Slstewartstatic void htcp_record_rtt(struct cc_var *ccv); 145216115Slstewartstatic void htcp_ssthresh_update(struct cc_var *ccv); 146216115Slstewart 147216115Slstewartstruct htcp { 148216115Slstewart /* cwnd before entering cong recovery. */ 149216115Slstewart unsigned long prev_cwnd; 150216115Slstewart /* cwnd additive increase parameter. */ 151216115Slstewart int alpha; 152216115Slstewart /* cwnd multiplicative decrease parameter. */ 153216115Slstewart int beta; 154216115Slstewart /* Largest rtt seen for the flow. */ 155216115Slstewart int maxrtt; 156216115Slstewart /* Shortest rtt seen for the flow. */ 157216115Slstewart int minrtt; 158216115Slstewart /* Time of last congestion event in ticks. */ 159216115Slstewart int t_last_cong; 160216115Slstewart}; 161216115Slstewart 162216115Slstewartstatic int htcp_rtt_ref; 163216115Slstewart/* 164216115Slstewart * The maximum number of ticks the value of diff can reach in 165216115Slstewart * htcp_recalc_alpha() before alpha will stop increasing due to overflow. 166216115Slstewart * See comment above HTCP_CALC_ALPHA for more info. 167216115Slstewart */ 168216115Slstewartstatic int htcp_max_diff = INT_MAX / ((1 << HTCP_ALPHA_INC_SHIFT) * 10); 169216115Slstewart 170216115Slstewart/* Per-netstack vars. */ 171217322Smdfstatic VNET_DEFINE(u_int, htcp_adaptive_backoff) = 0; 172217322Smdfstatic VNET_DEFINE(u_int, htcp_rtt_scaling) = 0; 173216115Slstewart#define V_htcp_adaptive_backoff VNET(htcp_adaptive_backoff) 174216115Slstewart#define V_htcp_rtt_scaling VNET(htcp_rtt_scaling) 175216115Slstewart 176220592Spluknetstatic MALLOC_DEFINE(M_HTCP, "htcp data", 177216115Slstewart "Per connection data required for the HTCP congestion control algorithm"); 178216115Slstewart 179216115Slstewartstruct cc_algo htcp_cc_algo = { 180216115Slstewart .name = "htcp", 181216115Slstewart .ack_received = htcp_ack_received, 182216115Slstewart .cb_destroy = htcp_cb_destroy, 183216115Slstewart .cb_init = htcp_cb_init, 184216115Slstewart .cong_signal = htcp_cong_signal, 185216115Slstewart .mod_init = htcp_mod_init, 186216115Slstewart .post_recovery = htcp_post_recovery, 187216115Slstewart}; 188216115Slstewart 189216115Slstewartstatic void 190216115Slstewarthtcp_ack_received(struct cc_var *ccv, uint16_t type) 191216115Slstewart{ 192216115Slstewart struct htcp *htcp_data; 193216115Slstewart 194216115Slstewart htcp_data = ccv->cc_data; 195216115Slstewart htcp_record_rtt(ccv); 196216115Slstewart 197216115Slstewart /* 198216115Slstewart * Regular ACK and we're not in cong/fast recovery and we're cwnd 199216115Slstewart * limited and we're either not doing ABC or are slow starting or are 200216115Slstewart * doing ABC and we've sent a cwnd's worth of bytes. 201216115Slstewart */ 202216115Slstewart if (type == CC_ACK && !IN_RECOVERY(CCV(ccv, t_flags)) && 203216115Slstewart (ccv->flags & CCF_CWND_LIMITED) && (!V_tcp_do_rfc3465 || 204216115Slstewart CCV(ccv, snd_cwnd) <= CCV(ccv, snd_ssthresh) || 205216115Slstewart (V_tcp_do_rfc3465 && ccv->flags & CCF_ABC_SENTAWND))) { 206216115Slstewart htcp_recalc_beta(ccv); 207216115Slstewart htcp_recalc_alpha(ccv); 208216115Slstewart /* 209216115Slstewart * Use the logic in NewReno ack_received() for slow start and 210216115Slstewart * for the first HTCP_DELTA_L ticks after either the flow starts 211216115Slstewart * or a congestion event (when alpha equals 1). 212216115Slstewart */ 213216115Slstewart if (htcp_data->alpha == 1 || 214216115Slstewart CCV(ccv, snd_cwnd) <= CCV(ccv, snd_ssthresh)) 215216115Slstewart newreno_cc_algo.ack_received(ccv, type); 216216115Slstewart else { 217216115Slstewart if (V_tcp_do_rfc3465) { 218216115Slstewart /* Increment cwnd by alpha segments. */ 219216115Slstewart CCV(ccv, snd_cwnd) += htcp_data->alpha * 220216115Slstewart CCV(ccv, t_maxseg); 221216115Slstewart ccv->flags &= ~CCF_ABC_SENTAWND; 222216115Slstewart } else 223216115Slstewart /* 224216115Slstewart * Increment cwnd by alpha/cwnd segments to 225216115Slstewart * approximate an increase of alpha segments 226216115Slstewart * per RTT. 227216115Slstewart */ 228216115Slstewart CCV(ccv, snd_cwnd) += (((htcp_data->alpha << 229216115Slstewart HTCP_SHIFT) / (CCV(ccv, snd_cwnd) / 230216115Slstewart CCV(ccv, t_maxseg))) * CCV(ccv, t_maxseg)) 231216115Slstewart >> HTCP_SHIFT; 232216115Slstewart } 233216115Slstewart } 234216115Slstewart} 235216115Slstewart 236216115Slstewartstatic void 237216115Slstewarthtcp_cb_destroy(struct cc_var *ccv) 238216115Slstewart{ 239216115Slstewart 240216115Slstewart if (ccv->cc_data != NULL) 241216115Slstewart free(ccv->cc_data, M_HTCP); 242216115Slstewart} 243216115Slstewart 244216115Slstewartstatic int 245216115Slstewarthtcp_cb_init(struct cc_var *ccv) 246216115Slstewart{ 247216115Slstewart struct htcp *htcp_data; 248216115Slstewart 249216115Slstewart htcp_data = malloc(sizeof(struct htcp), M_HTCP, M_NOWAIT); 250216115Slstewart 251216115Slstewart if (htcp_data == NULL) 252216115Slstewart return (ENOMEM); 253216115Slstewart 254216115Slstewart /* Init some key variables with sensible defaults. */ 255216115Slstewart htcp_data->alpha = HTCP_INIT_ALPHA; 256216115Slstewart htcp_data->beta = HTCP_MINBETA; 257216115Slstewart htcp_data->maxrtt = TCPTV_SRTTBASE; 258216115Slstewart htcp_data->minrtt = TCPTV_SRTTBASE; 259216115Slstewart htcp_data->prev_cwnd = 0; 260216115Slstewart htcp_data->t_last_cong = ticks; 261216115Slstewart 262216115Slstewart ccv->cc_data = htcp_data; 263216115Slstewart 264216115Slstewart return (0); 265216115Slstewart} 266216115Slstewart 267216115Slstewart/* 268216115Slstewart * Perform any necessary tasks before we enter congestion recovery. 269216115Slstewart */ 270216115Slstewartstatic void 271216115Slstewarthtcp_cong_signal(struct cc_var *ccv, uint32_t type) 272216115Slstewart{ 273216115Slstewart struct htcp *htcp_data; 274216115Slstewart 275216115Slstewart htcp_data = ccv->cc_data; 276216115Slstewart 277216115Slstewart switch (type) { 278216115Slstewart case CC_NDUPACK: 279216115Slstewart if (!IN_FASTRECOVERY(CCV(ccv, t_flags))) { 280216115Slstewart if (!IN_CONGRECOVERY(CCV(ccv, t_flags))) { 281216115Slstewart /* 282216115Slstewart * Apply hysteresis to maxrtt to ensure 283216115Slstewart * reductions in the RTT are reflected in our 284216115Slstewart * measurements. 285216115Slstewart */ 286216115Slstewart htcp_data->maxrtt = (htcp_data->minrtt + 287216115Slstewart (htcp_data->maxrtt - htcp_data->minrtt) * 288216115Slstewart 95) / 100; 289216115Slstewart htcp_ssthresh_update(ccv); 290216115Slstewart htcp_data->t_last_cong = ticks; 291216115Slstewart htcp_data->prev_cwnd = CCV(ccv, snd_cwnd); 292216115Slstewart } 293216115Slstewart ENTER_RECOVERY(CCV(ccv, t_flags)); 294216115Slstewart } 295216115Slstewart break; 296216115Slstewart 297216115Slstewart case CC_ECN: 298216115Slstewart if (!IN_CONGRECOVERY(CCV(ccv, t_flags))) { 299216115Slstewart /* 300216115Slstewart * Apply hysteresis to maxrtt to ensure reductions in 301216115Slstewart * the RTT are reflected in our measurements. 302216115Slstewart */ 303216115Slstewart htcp_data->maxrtt = (htcp_data->minrtt + (htcp_data->maxrtt - 304216115Slstewart htcp_data->minrtt) * 95) / 100; 305216115Slstewart htcp_ssthresh_update(ccv); 306216115Slstewart CCV(ccv, snd_cwnd) = CCV(ccv, snd_ssthresh); 307216115Slstewart htcp_data->t_last_cong = ticks; 308216115Slstewart htcp_data->prev_cwnd = CCV(ccv, snd_cwnd); 309216115Slstewart ENTER_CONGRECOVERY(CCV(ccv, t_flags)); 310216115Slstewart } 311216115Slstewart break; 312216115Slstewart 313216115Slstewart case CC_RTO: 314216115Slstewart /* 315216115Slstewart * Grab the current time and record it so we know when the 316216115Slstewart * most recent congestion event was. Only record it when the 317216115Slstewart * timeout has fired more than once, as there is a reasonable 318216115Slstewart * chance the first one is a false alarm and may not indicate 319216115Slstewart * congestion. 320216115Slstewart */ 321216115Slstewart if (CCV(ccv, t_rxtshift) >= 2) 322216115Slstewart htcp_data->t_last_cong = ticks; 323216115Slstewart break; 324216115Slstewart } 325216115Slstewart} 326216115Slstewart 327216115Slstewartstatic int 328216115Slstewarthtcp_mod_init(void) 329216115Slstewart{ 330216115Slstewart 331216115Slstewart htcp_cc_algo.after_idle = newreno_cc_algo.after_idle; 332216115Slstewart 333216115Slstewart /* 334216115Slstewart * HTCP_RTT_REF is defined in ms, and t_srtt in the tcpcb is stored in 335216115Slstewart * units of TCP_RTT_SCALE*hz. Scale HTCP_RTT_REF to be in the same units 336216115Slstewart * as t_srtt. 337216115Slstewart */ 338216115Slstewart htcp_rtt_ref = (HTCP_RTT_REF * TCP_RTT_SCALE * hz) / 1000; 339216115Slstewart 340216115Slstewart return (0); 341216115Slstewart} 342216115Slstewart 343216115Slstewart/* 344216115Slstewart * Perform any necessary tasks before we exit congestion recovery. 345216115Slstewart */ 346216115Slstewartstatic void 347216115Slstewarthtcp_post_recovery(struct cc_var *ccv) 348216115Slstewart{ 349216115Slstewart struct htcp *htcp_data; 350216115Slstewart 351216115Slstewart htcp_data = ccv->cc_data; 352216115Slstewart 353216115Slstewart if (IN_FASTRECOVERY(CCV(ccv, t_flags))) { 354216115Slstewart /* 355216115Slstewart * If inflight data is less than ssthresh, set cwnd 356216115Slstewart * conservatively to avoid a burst of data, as suggested in the 357216115Slstewart * NewReno RFC. Otherwise, use the HTCP method. 358216115Slstewart * 359216115Slstewart * XXXLAS: Find a way to do this without needing curack 360216115Slstewart */ 361216115Slstewart if (SEQ_GT(ccv->curack + CCV(ccv, snd_ssthresh), 362216115Slstewart CCV(ccv, snd_max))) 363216115Slstewart CCV(ccv, snd_cwnd) = CCV(ccv, snd_max) - ccv->curack + 364216115Slstewart CCV(ccv, t_maxseg); 365216115Slstewart else 366216115Slstewart CCV(ccv, snd_cwnd) = max(1, ((htcp_data->beta * 367216115Slstewart htcp_data->prev_cwnd / CCV(ccv, t_maxseg)) 368216115Slstewart >> HTCP_SHIFT)) * CCV(ccv, t_maxseg); 369216115Slstewart } 370216115Slstewart} 371216115Slstewart 372216115Slstewartstatic void 373216115Slstewarthtcp_recalc_alpha(struct cc_var *ccv) 374216115Slstewart{ 375216115Slstewart struct htcp *htcp_data; 376216115Slstewart int alpha, diff, now; 377216115Slstewart 378216115Slstewart htcp_data = ccv->cc_data; 379216115Slstewart now = ticks; 380216115Slstewart 381216115Slstewart /* 382216115Slstewart * If ticks has wrapped around (will happen approximately once every 49 383216115Slstewart * days on a machine with the default kern.hz=1000) and a flow straddles 384216115Slstewart * the wrap point, our alpha calcs will be completely wrong. We cut our 385216115Slstewart * losses and restart alpha from scratch by setting t_last_cong = now - 386216115Slstewart * HTCP_DELTA_L. 387216115Slstewart * 388216115Slstewart * This does not deflate our cwnd at all. It simply slows the rate cwnd 389216115Slstewart * is growing by until alpha regains the value it held prior to taking 390216115Slstewart * this drastic measure. 391216115Slstewart */ 392216115Slstewart if (now < htcp_data->t_last_cong) 393216115Slstewart htcp_data->t_last_cong = now - HTCP_DELTA_L; 394216115Slstewart 395216115Slstewart diff = now - htcp_data->t_last_cong - HTCP_DELTA_L; 396216115Slstewart 397216115Slstewart /* Cap alpha if the value of diff would overflow HTCP_CALC_ALPHA(). */ 398216115Slstewart if (diff < htcp_max_diff) { 399216115Slstewart /* 400216115Slstewart * If it has been more than HTCP_DELTA_L ticks since congestion, 401216115Slstewart * increase alpha according to the function defined in the spec. 402216115Slstewart */ 403216115Slstewart if (diff > 0) { 404216115Slstewart alpha = HTCP_CALC_ALPHA(diff); 405216115Slstewart 406216115Slstewart /* 407216115Slstewart * Adaptive backoff fairness adjustment: 408216115Slstewart * 2 * (1 - beta) * alpha_raw 409216115Slstewart */ 410216115Slstewart if (V_htcp_adaptive_backoff) 411216115Slstewart alpha = max(1, (2 * ((1 << HTCP_SHIFT) - 412216115Slstewart htcp_data->beta) * alpha) >> HTCP_SHIFT); 413216115Slstewart 414216115Slstewart /* 415216115Slstewart * RTT scaling: (RTT / RTT_ref) * alpha 416216115Slstewart * alpha will be the raw value from HTCP_CALC_ALPHA() if 417216115Slstewart * adaptive backoff is off, or the adjusted value if 418216115Slstewart * adaptive backoff is on. 419216115Slstewart */ 420216115Slstewart if (V_htcp_rtt_scaling) 421216115Slstewart alpha = max(1, (min(max(HTCP_MINROWE, 422216115Slstewart (CCV(ccv, t_srtt) << HTCP_SHIFT) / 423216115Slstewart htcp_rtt_ref), HTCP_MAXROWE) * alpha) 424216115Slstewart >> HTCP_SHIFT); 425216115Slstewart 426216115Slstewart } else 427216115Slstewart alpha = 1; 428216115Slstewart 429216115Slstewart htcp_data->alpha = alpha; 430216115Slstewart } 431216115Slstewart} 432216115Slstewart 433216115Slstewartstatic void 434216115Slstewarthtcp_recalc_beta(struct cc_var *ccv) 435216115Slstewart{ 436216115Slstewart struct htcp *htcp_data; 437216115Slstewart 438216115Slstewart htcp_data = ccv->cc_data; 439216115Slstewart 440216115Slstewart /* 441216115Slstewart * TCPTV_SRTTBASE is the initialised value of each connection's SRTT, so 442216115Slstewart * we only calc beta if the connection's SRTT has been changed from its 443216115Slstewart * inital value. beta is bounded to ensure it is always between 444216115Slstewart * HTCP_MINBETA and HTCP_MAXBETA. 445216115Slstewart */ 446216115Slstewart if (V_htcp_adaptive_backoff && htcp_data->minrtt != TCPTV_SRTTBASE && 447216115Slstewart htcp_data->maxrtt != TCPTV_SRTTBASE) 448216115Slstewart htcp_data->beta = min(max(HTCP_MINBETA, 449216115Slstewart (htcp_data->minrtt << HTCP_SHIFT) / htcp_data->maxrtt), 450216115Slstewart HTCP_MAXBETA); 451216115Slstewart else 452216115Slstewart htcp_data->beta = HTCP_MINBETA; 453216115Slstewart} 454216115Slstewart 455216115Slstewart/* 456216115Slstewart * Record the minimum and maximum RTT seen for the connection. These are used in 457216115Slstewart * the calculation of beta if adaptive backoff is enabled. 458216115Slstewart */ 459216115Slstewartstatic void 460216115Slstewarthtcp_record_rtt(struct cc_var *ccv) 461216115Slstewart{ 462216115Slstewart struct htcp *htcp_data; 463216115Slstewart 464216115Slstewart htcp_data = ccv->cc_data; 465216115Slstewart 466216115Slstewart /* XXXLAS: Should there be some hysteresis for minrtt? */ 467216115Slstewart 468216115Slstewart /* 469216115Slstewart * Record the current SRTT as our minrtt if it's the smallest we've seen 470216115Slstewart * or minrtt is currently equal to its initialised value. Ignore SRTT 471216115Slstewart * until a min number of samples have been taken. 472216115Slstewart */ 473216115Slstewart if ((CCV(ccv, t_srtt) < htcp_data->minrtt || 474216115Slstewart htcp_data->minrtt == TCPTV_SRTTBASE) && 475216115Slstewart (CCV(ccv, t_rttupdated) >= HTCP_MIN_RTT_SAMPLES)) 476216115Slstewart htcp_data->minrtt = CCV(ccv, t_srtt); 477216115Slstewart 478216115Slstewart /* 479216115Slstewart * Record the current SRTT as our maxrtt if it's the largest we've 480216115Slstewart * seen. Ignore SRTT until a min number of samples have been taken. 481216115Slstewart */ 482216115Slstewart if (CCV(ccv, t_srtt) > htcp_data->maxrtt 483216115Slstewart && CCV(ccv, t_rttupdated) >= HTCP_MIN_RTT_SAMPLES) 484216115Slstewart htcp_data->maxrtt = CCV(ccv, t_srtt); 485216115Slstewart} 486216115Slstewart 487216115Slstewart/* 488216115Slstewart * Update the ssthresh in the event of congestion. 489216115Slstewart */ 490216115Slstewartstatic void 491216115Slstewarthtcp_ssthresh_update(struct cc_var *ccv) 492216115Slstewart{ 493216115Slstewart struct htcp *htcp_data; 494216115Slstewart 495216115Slstewart htcp_data = ccv->cc_data; 496216115Slstewart 497216115Slstewart /* 498216115Slstewart * On the first congestion event, set ssthresh to cwnd * 0.5, on 499216115Slstewart * subsequent congestion events, set it to cwnd * beta. 500216115Slstewart */ 501216115Slstewart if (CCV(ccv, snd_ssthresh) == TCP_MAXWIN << TCP_MAX_WINSHIFT) 502216115Slstewart CCV(ccv, snd_ssthresh) = (CCV(ccv, snd_cwnd) * HTCP_MINBETA) 503216115Slstewart >> HTCP_SHIFT; 504216115Slstewart else { 505216115Slstewart htcp_recalc_beta(ccv); 506216115Slstewart CCV(ccv, snd_ssthresh) = (CCV(ccv, snd_cwnd) * htcp_data->beta) 507216115Slstewart >> HTCP_SHIFT; 508216115Slstewart } 509216115Slstewart} 510216115Slstewart 511216115Slstewart 512216115SlstewartSYSCTL_DECL(_net_inet_tcp_cc_htcp); 513216115SlstewartSYSCTL_NODE(_net_inet_tcp_cc, OID_AUTO, htcp, CTLFLAG_RW, 514216115Slstewart NULL, "H-TCP related settings"); 515216115SlstewartSYSCTL_VNET_UINT(_net_inet_tcp_cc_htcp, OID_AUTO, adaptive_backoff, CTLFLAG_RW, 516216115Slstewart &VNET_NAME(htcp_adaptive_backoff), 0, "enable H-TCP adaptive backoff"); 517216115SlstewartSYSCTL_VNET_UINT(_net_inet_tcp_cc_htcp, OID_AUTO, rtt_scaling, CTLFLAG_RW, 518216115Slstewart &VNET_NAME(htcp_rtt_scaling), 0, "enable H-TCP RTT scaling"); 519216115Slstewart 520216115SlstewartDECLARE_CC_MODULE(htcp, &htcp_cc_algo); 521