167fef78bSLawrence Stewart /*- 2*4d846d26SWarner Losh * SPDX-License-Identifier: BSD-2-Clause 3fe267a55SPedro F. Giffuni * 467fef78bSLawrence Stewart * Copyright (c) 2008-2010 Lawrence Stewart <lstewart@freebsd.org> 567fef78bSLawrence Stewart * Copyright (c) 2010 The FreeBSD Foundation 667fef78bSLawrence Stewart * All rights reserved. 767fef78bSLawrence Stewart * 867fef78bSLawrence Stewart * This software was developed by Lawrence Stewart while studying at the Centre 9891b8ed4SLawrence Stewart * for Advanced Internet Architectures, Swinburne University of Technology, made 10891b8ed4SLawrence Stewart * possible in part by a grant from the Cisco University Research Program Fund 11891b8ed4SLawrence Stewart * at Community Foundation Silicon Valley. 1267fef78bSLawrence Stewart * 1367fef78bSLawrence Stewart * Portions of this software were developed at the Centre for Advanced 1467fef78bSLawrence Stewart * Internet Architectures, Swinburne University of Technology, Melbourne, 1567fef78bSLawrence Stewart * Australia by David Hayes under sponsorship from the FreeBSD Foundation. 1667fef78bSLawrence Stewart * 1767fef78bSLawrence Stewart * Redistribution and use in source and binary forms, with or without 1867fef78bSLawrence Stewart * modification, are permitted provided that the following conditions 1967fef78bSLawrence Stewart * are met: 2067fef78bSLawrence Stewart * 1. Redistributions of source code must retain the above copyright 2167fef78bSLawrence Stewart * notice, this list of conditions and the following disclaimer. 2267fef78bSLawrence Stewart * 2. Redistributions in binary form must reproduce the above copyright 2367fef78bSLawrence Stewart * notice, this list of conditions and the following disclaimer in the 2467fef78bSLawrence Stewart * documentation and/or other materials provided with the distribution. 2567fef78bSLawrence Stewart * 2667fef78bSLawrence Stewart * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 2767fef78bSLawrence Stewart * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 2867fef78bSLawrence Stewart * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 2967fef78bSLawrence Stewart * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 3067fef78bSLawrence Stewart * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 3167fef78bSLawrence Stewart * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 3267fef78bSLawrence Stewart * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 3367fef78bSLawrence Stewart * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 3467fef78bSLawrence Stewart * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 3567fef78bSLawrence Stewart * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 3667fef78bSLawrence Stewart * SUCH DAMAGE. 3767fef78bSLawrence Stewart * 3867fef78bSLawrence Stewart * $FreeBSD$ 3967fef78bSLawrence Stewart */ 4067fef78bSLawrence Stewart 4167fef78bSLawrence Stewart #ifndef _NETINET_CC_CUBIC_H_ 4267fef78bSLawrence Stewart #define _NETINET_CC_CUBIC_H_ 4367fef78bSLawrence Stewart 44c968c769SMichael Tuexen #include <sys/limits.h> 45c968c769SMichael Tuexen 4667fef78bSLawrence Stewart /* Number of bits of precision for fixed point math calcs. */ 4767fef78bSLawrence Stewart #define CUBIC_SHIFT 8 4867fef78bSLawrence Stewart 4967fef78bSLawrence Stewart #define CUBIC_SHIFT_4 32 5067fef78bSLawrence Stewart 5167fef78bSLawrence Stewart /* 0.5 << CUBIC_SHIFT. */ 5267fef78bSLawrence Stewart #define RENO_BETA 128 5367fef78bSLawrence Stewart 5468ff29afSSean Bruno /* ~0.7 << CUBIC_SHIFT. */ 5568ff29afSSean Bruno #define CUBIC_BETA 179 5667fef78bSLawrence Stewart 5768ff29afSSean Bruno /* ~0.3 << CUBIC_SHIFT. */ 5868ff29afSSean Bruno #define ONE_SUB_CUBIC_BETA 77 5967fef78bSLawrence Stewart 6067fef78bSLawrence Stewart /* 3 * ONE_SUB_CUBIC_BETA. */ 6168ff29afSSean Bruno #define THREE_X_PT3 231 6267fef78bSLawrence Stewart 6367fef78bSLawrence Stewart /* (2 << CUBIC_SHIFT) - ONE_SUB_CUBIC_BETA. */ 6468ff29afSSean Bruno #define TWO_SUB_PT3 435 6567fef78bSLawrence Stewart 6667fef78bSLawrence Stewart /* ~0.4 << CUBIC_SHIFT. */ 6767fef78bSLawrence Stewart #define CUBIC_C_FACTOR 102 6867fef78bSLawrence Stewart 6968ff29afSSean Bruno /* CUBIC fast convergence factor: (1+beta_cubic)/2. */ 7068ff29afSSean Bruno #define CUBIC_FC_FACTOR 217 7167fef78bSLawrence Stewart 7267fef78bSLawrence Stewart /* Don't trust s_rtt until this many rtt samples have been taken. */ 7367fef78bSLawrence Stewart #define CUBIC_MIN_RTT_SAMPLES 8 7467fef78bSLawrence Stewart 75c968c769SMichael Tuexen /* 76c968c769SMichael Tuexen * (2^21)^3 is long max. Dividing (2^63) by Cubic_C_factor 77c968c769SMichael Tuexen * and taking cube-root yields 448845 as the effective useful limit 78c968c769SMichael Tuexen */ 79c968c769SMichael Tuexen #define CUBED_ROOT_MAX_ULONG 448845 80c968c769SMichael Tuexen 81a9696510SRandall Stewart /* Flags used in the cubic structure */ 82a9696510SRandall Stewart #define CUBICFLAG_CONG_EVENT 0x00000001 /* congestion experienced */ 83a9696510SRandall Stewart #define CUBICFLAG_IN_SLOWSTART 0x00000002 /* in slow start */ 84a9696510SRandall Stewart #define CUBICFLAG_IN_APPLIMIT 0x00000004 /* application limited */ 85a9696510SRandall Stewart #define CUBICFLAG_RTO_EVENT 0x00000008 /* RTO experienced */ 86a9696510SRandall Stewart #define CUBICFLAG_HYSTART_ENABLED 0x00000010 /* Hystart++ is enabled */ 87a9696510SRandall Stewart #define CUBICFLAG_HYSTART_IN_CSS 0x00000020 /* We are in Hystart++ CSS */ 88a9696510SRandall Stewart 89a9696510SRandall Stewart /* Kernel only bits */ 90a9696510SRandall Stewart #ifdef _KERNEL 91a9696510SRandall Stewart struct cubic { 92ea6d0de2SRichard Scheffenegger /* CUBIC K in fixed point form with CUBIC_SHIFT worth of precision. */ 93a9696510SRandall Stewart int64_t K; 94a9696510SRandall Stewart /* Sum of RTT samples across an epoch in ticks. */ 95a9696510SRandall Stewart int64_t sum_rtt_ticks; 96a9696510SRandall Stewart /* cwnd at the most recent congestion event. */ 97a9696510SRandall Stewart unsigned long max_cwnd; 98a9696510SRandall Stewart /* cwnd at the previous congestion event. */ 99a9696510SRandall Stewart unsigned long prev_max_cwnd; 100a9696510SRandall Stewart /* A copy of prev_max_cwnd. Used for CC_RTO_ERR */ 101a9696510SRandall Stewart unsigned long prev_max_cwnd_cp; 102a9696510SRandall Stewart /* various flags */ 103a9696510SRandall Stewart uint32_t flags; 104a9696510SRandall Stewart /* Minimum observed rtt in ticks. */ 105a9696510SRandall Stewart int min_rtt_ticks; 106a9696510SRandall Stewart /* Mean observed rtt between congestion epochs. */ 107a9696510SRandall Stewart int mean_rtt_ticks; 108a9696510SRandall Stewart /* ACKs since last congestion event. */ 109a9696510SRandall Stewart int epoch_ack_count; 110a9696510SRandall Stewart /* Timestamp (in ticks) of arriving in congestion avoidance from last 111a9696510SRandall Stewart * congestion event. 112a9696510SRandall Stewart */ 113a9696510SRandall Stewart int t_last_cong; 114a9696510SRandall Stewart /* Timestamp (in ticks) of a previous congestion event. Used for 115a9696510SRandall Stewart * CC_RTO_ERR. 116a9696510SRandall Stewart */ 117a9696510SRandall Stewart int t_last_cong_prev; 118a9696510SRandall Stewart uint32_t css_baseline_minrtt; 119a9696510SRandall Stewart uint32_t css_current_round_minrtt; 120a9696510SRandall Stewart uint32_t css_lastround_minrtt; 121a9696510SRandall Stewart uint32_t css_rttsample_count; 122a9696510SRandall Stewart uint32_t css_entered_at_round; 123a9696510SRandall Stewart uint32_t css_current_round; 124a9696510SRandall Stewart uint32_t css_fas_at_css_entry; 125a9696510SRandall Stewart uint32_t css_lowrtt_fas; 126a9696510SRandall Stewart uint32_t css_last_fas; 127a9696510SRandall Stewart }; 128a9696510SRandall Stewart #endif 129a9696510SRandall Stewart 13067fef78bSLawrence Stewart /* Userland only bits. */ 13167fef78bSLawrence Stewart #ifndef _KERNEL 13267fef78bSLawrence Stewart 13367fef78bSLawrence Stewart extern int hz; 13467fef78bSLawrence Stewart 13567fef78bSLawrence Stewart /* 13667fef78bSLawrence Stewart * Implementation based on the formulae found in the CUBIC Internet Draft 13768ff29afSSean Bruno * "draft-ietf-tcpm-cubic-04". 13867fef78bSLawrence Stewart * 13967fef78bSLawrence Stewart */ 14067fef78bSLawrence Stewart 14167fef78bSLawrence Stewart static __inline float 14267fef78bSLawrence Stewart theoretical_cubic_k(double wmax_pkts) 14367fef78bSLawrence Stewart { 14467fef78bSLawrence Stewart double C; 14567fef78bSLawrence Stewart 14667fef78bSLawrence Stewart C = 0.4; 14767fef78bSLawrence Stewart 14868ff29afSSean Bruno return (pow((wmax_pkts * 0.3) / C, (1.0 / 3.0)) * pow(2, CUBIC_SHIFT)); 14967fef78bSLawrence Stewart } 15067fef78bSLawrence Stewart 15167fef78bSLawrence Stewart static __inline unsigned long 15267fef78bSLawrence Stewart theoretical_cubic_cwnd(int ticks_since_cong, unsigned long wmax, uint32_t smss) 15367fef78bSLawrence Stewart { 15467fef78bSLawrence Stewart double C, wmax_pkts; 15567fef78bSLawrence Stewart 15667fef78bSLawrence Stewart C = 0.4; 15767fef78bSLawrence Stewart wmax_pkts = wmax / (double)smss; 15867fef78bSLawrence Stewart 15967fef78bSLawrence Stewart return (smss * (wmax_pkts + 16067fef78bSLawrence Stewart (C * pow(ticks_since_cong / (double)hz - 16167fef78bSLawrence Stewart theoretical_cubic_k(wmax_pkts) / pow(2, CUBIC_SHIFT), 3.0)))); 16267fef78bSLawrence Stewart } 16367fef78bSLawrence Stewart 16467fef78bSLawrence Stewart static __inline unsigned long 16567fef78bSLawrence Stewart theoretical_reno_cwnd(int ticks_since_cong, int rtt_ticks, unsigned long wmax, 16667fef78bSLawrence Stewart uint32_t smss) 16767fef78bSLawrence Stewart { 16867fef78bSLawrence Stewart 16967fef78bSLawrence Stewart return ((wmax * 0.5) + ((ticks_since_cong / (float)rtt_ticks) * smss)); 17067fef78bSLawrence Stewart } 17167fef78bSLawrence Stewart 17267fef78bSLawrence Stewart static __inline unsigned long 17367fef78bSLawrence Stewart theoretical_tf_cwnd(int ticks_since_cong, int rtt_ticks, unsigned long wmax, 17467fef78bSLawrence Stewart uint32_t smss) 17567fef78bSLawrence Stewart { 17667fef78bSLawrence Stewart 17768ff29afSSean Bruno return ((wmax * 0.7) + ((3 * 0.3) / (2 - 0.3) * 17867fef78bSLawrence Stewart (ticks_since_cong / (float)rtt_ticks) * smss)); 17967fef78bSLawrence Stewart } 18067fef78bSLawrence Stewart 18167fef78bSLawrence Stewart #endif /* !_KERNEL */ 18267fef78bSLawrence Stewart 18367fef78bSLawrence Stewart /* 18467fef78bSLawrence Stewart * Compute the CUBIC K value used in the cwnd calculation, using an 18567fef78bSLawrence Stewart * implementation of eqn 2 in the I-D. The method used 18667fef78bSLawrence Stewart * here is adapted from Apple Computer Technical Report #KT-32. 18767fef78bSLawrence Stewart */ 18867fef78bSLawrence Stewart static __inline int64_t 18967fef78bSLawrence Stewart cubic_k(unsigned long wmax_pkts) 19067fef78bSLawrence Stewart { 19167fef78bSLawrence Stewart int64_t s, K; 19267fef78bSLawrence Stewart uint16_t p; 19367fef78bSLawrence Stewart 19467fef78bSLawrence Stewart K = s = 0; 19567fef78bSLawrence Stewart p = 0; 19667fef78bSLawrence Stewart 19767fef78bSLawrence Stewart /* (wmax * beta)/C with CUBIC_SHIFT worth of precision. */ 19867fef78bSLawrence Stewart s = ((wmax_pkts * ONE_SUB_CUBIC_BETA) << CUBIC_SHIFT) / CUBIC_C_FACTOR; 19967fef78bSLawrence Stewart 20067fef78bSLawrence Stewart /* Rebase s to be between 1 and 1/8 with a shift of CUBIC_SHIFT. */ 20167fef78bSLawrence Stewart while (s >= 256) { 20267fef78bSLawrence Stewart s >>= 3; 20367fef78bSLawrence Stewart p++; 20467fef78bSLawrence Stewart } 20567fef78bSLawrence Stewart 20667fef78bSLawrence Stewart /* 20767fef78bSLawrence Stewart * Some magic constants taken from the Apple TR with appropriate 20867fef78bSLawrence Stewart * shifts: 275 == 1.072302 << CUBIC_SHIFT, 98 == 0.3812513 << 20967fef78bSLawrence Stewart * CUBIC_SHIFT, 120 == 0.46946116 << CUBIC_SHIFT. 21067fef78bSLawrence Stewart */ 21167fef78bSLawrence Stewart K = (((s * 275) >> CUBIC_SHIFT) + 98) - 21267fef78bSLawrence Stewart (((s * s * 120) >> CUBIC_SHIFT) >> CUBIC_SHIFT); 21367fef78bSLawrence Stewart 21467fef78bSLawrence Stewart /* Multiply by 2^p to undo the rebasing of s from above. */ 21567fef78bSLawrence Stewart return (K <<= p); 21667fef78bSLawrence Stewart } 21767fef78bSLawrence Stewart 21867fef78bSLawrence Stewart /* 21967fef78bSLawrence Stewart * Compute the new cwnd value using an implementation of eqn 1 from the I-D. 22067fef78bSLawrence Stewart * Thanks to Kip Macy for help debugging this function. 22151e712f8SHiren Panchasara * 22251e712f8SHiren Panchasara * XXXLAS: Characterise bounds for overflow. 22367fef78bSLawrence Stewart */ 22467fef78bSLawrence Stewart static __inline unsigned long 22567fef78bSLawrence Stewart cubic_cwnd(int ticks_since_cong, unsigned long wmax, uint32_t smss, int64_t K) 22667fef78bSLawrence Stewart { 22767fef78bSLawrence Stewart int64_t cwnd; 22867fef78bSLawrence Stewart 22967fef78bSLawrence Stewart /* K is in fixed point form with CUBIC_SHIFT worth of precision. */ 23067fef78bSLawrence Stewart 23167fef78bSLawrence Stewart /* t - K, with CUBIC_SHIFT worth of precision. */ 232c968c769SMichael Tuexen cwnd = (((int64_t)ticks_since_cong << CUBIC_SHIFT) - (K * hz)) / hz; 233c968c769SMichael Tuexen 234c968c769SMichael Tuexen if (cwnd > CUBED_ROOT_MAX_ULONG) 235c968c769SMichael Tuexen return INT_MAX; 236c968c769SMichael Tuexen if (cwnd < -CUBED_ROOT_MAX_ULONG) 237c968c769SMichael Tuexen return 0; 23867fef78bSLawrence Stewart 23967fef78bSLawrence Stewart /* (t - K)^3, with CUBIC_SHIFT^3 worth of precision. */ 24067fef78bSLawrence Stewart cwnd *= (cwnd * cwnd); 24167fef78bSLawrence Stewart 24267fef78bSLawrence Stewart /* 24367fef78bSLawrence Stewart * C(t - K)^3 + wmax 24467fef78bSLawrence Stewart * The down shift by CUBIC_SHIFT_4 is because cwnd has 4 lots of 24567fef78bSLawrence Stewart * CUBIC_SHIFT included in the value. 3 from the cubing of cwnd above, 24667fef78bSLawrence Stewart * and an extra from multiplying through by CUBIC_C_FACTOR. 24767fef78bSLawrence Stewart */ 24867fef78bSLawrence Stewart 249c968c769SMichael Tuexen cwnd = ((cwnd * CUBIC_C_FACTOR) >> CUBIC_SHIFT_4) * smss + wmax; 250c968c769SMichael Tuexen 251c968c769SMichael Tuexen /* 252c968c769SMichael Tuexen * for negative cwnd, limiting to zero as lower bound 253c968c769SMichael Tuexen */ 254c968c769SMichael Tuexen return (lmax(0,cwnd)); 25567fef78bSLawrence Stewart } 25667fef78bSLawrence Stewart 25767fef78bSLawrence Stewart /* 25867fef78bSLawrence Stewart * Compute an approximation of the NewReno cwnd some number of ticks after a 25967fef78bSLawrence Stewart * congestion event. RTT should be the average RTT estimate for the path 26067fef78bSLawrence Stewart * measured over the previous congestion epoch and wmax is the value of cwnd at 26167fef78bSLawrence Stewart * the last congestion event. The "TCP friendly" concept in the CUBIC I-D is 26267fef78bSLawrence Stewart * rather tricky to understand and it turns out this function is not required. 26367fef78bSLawrence Stewart * It is left here for reference. 26467fef78bSLawrence Stewart */ 26567fef78bSLawrence Stewart static __inline unsigned long 26667fef78bSLawrence Stewart reno_cwnd(int ticks_since_cong, int rtt_ticks, unsigned long wmax, 26767fef78bSLawrence Stewart uint32_t smss) 26867fef78bSLawrence Stewart { 26967fef78bSLawrence Stewart 27067fef78bSLawrence Stewart /* 27167fef78bSLawrence Stewart * For NewReno, beta = 0.5, therefore: W_tcp(t) = wmax*0.5 + t/RTT 27267fef78bSLawrence Stewart * W_tcp(t) deals with cwnd/wmax in pkts, so because our cwnd is in 27367fef78bSLawrence Stewart * bytes, we have to multiply by smss. 27467fef78bSLawrence Stewart */ 27567fef78bSLawrence Stewart return (((wmax * RENO_BETA) + (((ticks_since_cong * smss) 27667fef78bSLawrence Stewart << CUBIC_SHIFT) / rtt_ticks)) >> CUBIC_SHIFT); 27767fef78bSLawrence Stewart } 27867fef78bSLawrence Stewart 27967fef78bSLawrence Stewart /* 28067fef78bSLawrence Stewart * Compute an approximation of the "TCP friendly" cwnd some number of ticks 28167fef78bSLawrence Stewart * after a congestion event that is designed to yield the same average cwnd as 28268ff29afSSean Bruno * NewReno while using CUBIC's beta of 0.7. RTT should be the average RTT 28367fef78bSLawrence Stewart * estimate for the path measured over the previous congestion epoch and wmax is 28467fef78bSLawrence Stewart * the value of cwnd at the last congestion event. 28567fef78bSLawrence Stewart */ 28667fef78bSLawrence Stewart static __inline unsigned long 28767fef78bSLawrence Stewart tf_cwnd(int ticks_since_cong, int rtt_ticks, unsigned long wmax, 28867fef78bSLawrence Stewart uint32_t smss) 28967fef78bSLawrence Stewart { 29067fef78bSLawrence Stewart 29167fef78bSLawrence Stewart /* Equation 4 of I-D. */ 292c968c769SMichael Tuexen return (((wmax * CUBIC_BETA) + 293c968c769SMichael Tuexen (((THREE_X_PT3 * (unsigned long)ticks_since_cong * 294c968c769SMichael Tuexen (unsigned long)smss) << CUBIC_SHIFT) / (TWO_SUB_PT3 * rtt_ticks))) 295c968c769SMichael Tuexen >> CUBIC_SHIFT); 29667fef78bSLawrence Stewart } 29767fef78bSLawrence Stewart 29867fef78bSLawrence Stewart #endif /* _NETINET_CC_CUBIC_H_ */ 299