net/ipv4/tcp_dctcp.c

0001 // SPDX-License-Identifier: GPL-2.0-or-later
0002 /* DataCenter TCP (DCTCP) congestion control.
0003  *
0004  * http://simula.stanford.edu/~alizade/Site/DCTCP.html
0005  *
0006  * This is an implementation of DCTCP over Reno, an enhancement to the
0007  * TCP congestion control algorithm designed for data centers. DCTCP
0008  * leverages Explicit Congestion Notification (ECN) in the network to
0009  * provide multi-bit feedback to the end hosts. DCTCP's goal is to meet
0010  * the following three data center transport requirements:
0011  *
0012  *  - High burst tolerance (incast due to partition/aggregate)
0013  *  - Low latency (short flows, queries)
0014  *  - High throughput (continuous data updates, large file transfers)
0015  *    with commodity shallow buffered switches
0016  *
0017  * The algorithm is described in detail in the following two papers:
0018  *
0019  * 1) Mohammad Alizadeh, Albert Greenberg, David A. Maltz, Jitendra Padhye,
0020  *    Parveen Patel, Balaji Prabhakar, Sudipta Sengupta, and Murari Sridharan:
0021  *      "Data Center TCP (DCTCP)", Data Center Networks session
0022  *      Proc. ACM SIGCOMM, New Delhi, 2010.
0023  *   http://simula.stanford.edu/~alizade/Site/DCTCP_files/dctcp-final.pdf
0024  *
0025  * 2) Mohammad Alizadeh, Adel Javanmard, and Balaji Prabhakar:
0026  *      "Analysis of DCTCP: Stability, Convergence, and Fairness"
0027  *      Proc. ACM SIGMETRICS, San Jose, 2011.
0028  *   http://simula.stanford.edu/~alizade/Site/DCTCP_files/dctcp_analysis-full.pdf
0029  *
0030  * Initial prototype from Abdul Kabbani, Masato Yasuda and Mohammad Alizadeh.
0031  *
0032  * Authors:
0033  *
0034  *  Daniel Borkmann <dborkman@redhat.com>
0035  *  Florian Westphal <fw@strlen.de>
0036  *  Glenn Judd <glenn.judd@morganstanley.com>
0037  */
0038
0039 #include <linux/btf.h>
0040 #include <linux/btf_ids.h>
0041 #include <linux/module.h>
0042 #include <linux/mm.h>
0043 #include <net/tcp.h>
0044 #include <linux/inet_diag.h>
0045 #include "tcp_dctcp.h"
0046
0047 #define DCTCP_MAX_ALPHA 1024U
0048
0049 struct dctcp {
0050     u32 old_delivered;
0051     u32 old_delivered_ce;
0052     u32 prior_rcv_nxt;
0053     u32 dctcp_alpha;
0054     u32 next_seq;
0055     u32 ce_state;
0056     u32 loss_cwnd;
0057 };
0058
0059 static unsigned int dctcp_shift_g __read_mostly = 4; /* g = 1/2^4 */
0060 module_param(dctcp_shift_g, uint, 0644);
0061 MODULE_PARM_DESC(dctcp_shift_g, "parameter g for updating dctcp_alpha");
0062
0063 static unsigned int dctcp_alpha_on_init __read_mostly = DCTCP_MAX_ALPHA;
0064 module_param(dctcp_alpha_on_init, uint, 0644);
0065 MODULE_PARM_DESC(dctcp_alpha_on_init, "parameter for initial alpha value");
0066
0067 static struct tcp_congestion_ops dctcp_reno;
0068
0069 static void dctcp_reset(const struct tcp_sock *tp, struct dctcp *ca)
0070 {
0071     ca->next_seq = tp->snd_nxt;
0072
0073     ca->old_delivered = tp->delivered;
0074     ca->old_delivered_ce = tp->delivered_ce;
0075 }
0076
0077 static void dctcp_init(struct sock *sk)
0078 {
0079     const struct tcp_sock *tp = tcp_sk(sk);
0080
0081     if ((tp->ecn_flags & TCP_ECN_OK) ||
0082         (sk->sk_state == TCP_LISTEN ||
0083          sk->sk_state == TCP_CLOSE)) {
0084         struct dctcp *ca = inet_csk_ca(sk);
0085
0086         ca->prior_rcv_nxt = tp->rcv_nxt;
0087
0088         ca->dctcp_alpha = min(dctcp_alpha_on_init, DCTCP_MAX_ALPHA);
0089
0090         ca->loss_cwnd = 0;
0091         ca->ce_state = 0;
0092
0093         dctcp_reset(tp, ca);
0094         return;
0095     }
0096
0097     /* No ECN support? Fall back to Reno. Also need to clear
0098      * ECT from sk since it is set during 3WHS for DCTCP.
0099      */
0100     inet_csk(sk)->icsk_ca_ops = &dctcp_reno;
0101     INET_ECN_dontxmit(sk);
0102 }
0103
0104 static u32 dctcp_ssthresh(struct sock *sk)
0105 {
0106     struct dctcp *ca = inet_csk_ca(sk);
0107     struct tcp_sock *tp = tcp_sk(sk);
0108
0109     ca->loss_cwnd = tcp_snd_cwnd(tp);
0110     return max(tcp_snd_cwnd(tp) - ((tcp_snd_cwnd(tp) * ca->dctcp_alpha) >> 11U), 2U);
0111 }
0112
0113 static void dctcp_update_alpha(struct sock *sk, u32 flags)
0114 {
0115     const struct tcp_sock *tp = tcp_sk(sk);
0116     struct dctcp *ca = inet_csk_ca(sk);
0117
0118     /* Expired RTT */
0119     if (!before(tp->snd_una, ca->next_seq)) {
0120         u32 delivered_ce = tp->delivered_ce - ca->old_delivered_ce;
0121         u32 alpha = ca->dctcp_alpha;
0122
0123         /* alpha = (1 - g) * alpha + g * F */
0124
0125         alpha -= min_not_zero(alpha, alpha >> dctcp_shift_g);
0126         if (delivered_ce) {
0127             u32 delivered = tp->delivered - ca->old_delivered;
0128
0129             /* If dctcp_shift_g == 1, a 32bit value would overflow
0130              * after 8 M packets.
0131              */
0132             delivered_ce <<= (10 - dctcp_shift_g);
0133             delivered_ce /= max(1U, delivered);
0134
0135             alpha = min(alpha + delivered_ce, DCTCP_MAX_ALPHA);
0136         }
0137         /* dctcp_alpha can be read from dctcp_get_info() without
0138          * synchro, so we ask compiler to not use dctcp_alpha
0139          * as a temporary variable in prior operations.
0140          */
0141         WRITE_ONCE(ca->dctcp_alpha, alpha);
0142         dctcp_reset(tp, ca);
0143     }
0144 }
0145
0146 static void dctcp_react_to_loss(struct sock *sk)
0147 {
0148     struct dctcp *ca = inet_csk_ca(sk);
0149     struct tcp_sock *tp = tcp_sk(sk);
0150
0151     ca->loss_cwnd = tcp_snd_cwnd(tp);
0152     tp->snd_ssthresh = max(tcp_snd_cwnd(tp) >> 1U, 2U);
0153 }
0154
0155 static void dctcp_state(struct sock *sk, u8 new_state)
0156 {
0157     if (new_state == TCP_CA_Recovery &&
0158         new_state != inet_csk(sk)->icsk_ca_state)
0159         dctcp_react_to_loss(sk);
0160     /* We handle RTO in dctcp_cwnd_event to ensure that we perform only
0161      * one loss-adjustment per RTT.
0162      */
0163 }
0164
0165 static void dctcp_cwnd_event(struct sock *sk, enum tcp_ca_event ev)
0166 {
0167     struct dctcp *ca = inet_csk_ca(sk);
0168
0169     switch (ev) {
0170     case CA_EVENT_ECN_IS_CE:
0171     case CA_EVENT_ECN_NO_CE:
0172         dctcp_ece_ack_update(sk, ev, &ca->prior_rcv_nxt, &ca->ce_state);
0173         break;
0174     case CA_EVENT_LOSS:
0175         dctcp_react_to_loss(sk);
0176         break;
0177     default:
0178         /* Don't care for the rest. */
0179         break;
0180     }
0181 }
0182
0183 static size_t dctcp_get_info(struct sock *sk, u32 ext, int *attr,
0184                  union tcp_cc_info *info)
0185 {
0186     const struct dctcp *ca = inet_csk_ca(sk);
0187     const struct tcp_sock *tp = tcp_sk(sk);
0188
0189     /* Fill it also in case of VEGASINFO due to req struct limits.
0190      * We can still correctly retrieve it later.
0191      */
0192     if (ext & (1 << (INET_DIAG_DCTCPINFO - 1)) ||
0193         ext & (1 << (INET_DIAG_VEGASINFO - 1))) {
0194         memset(&info->dctcp, 0, sizeof(info->dctcp));
0195         if (inet_csk(sk)->icsk_ca_ops != &dctcp_reno) {
0196             info->dctcp.dctcp_enabled = 1;
0197             info->dctcp.dctcp_ce_state = (u16) ca->ce_state;
0198             info->dctcp.dctcp_alpha = ca->dctcp_alpha;
0199             info->dctcp.dctcp_ab_ecn = tp->mss_cache *
0200                            (tp->delivered_ce - ca->old_delivered_ce);
0201             info->dctcp.dctcp_ab_tot = tp->mss_cache *
0202                            (tp->delivered - ca->old_delivered);
0203         }
0204
0205         *attr = INET_DIAG_DCTCPINFO;
0206         return sizeof(info->dctcp);
0207     }
0208     return 0;
0209 }
0210
0211 static u32 dctcp_cwnd_undo(struct sock *sk)
0212 {
0213     const struct dctcp *ca = inet_csk_ca(sk);
0214     struct tcp_sock *tp = tcp_sk(sk);
0215
0216     return max(tcp_snd_cwnd(tp), ca->loss_cwnd);
0217 }
0218
0219 static struct tcp_congestion_ops dctcp __read_mostly = {
0220     .init       = dctcp_init,
0221     .in_ack_event   = dctcp_update_alpha,
0222     .cwnd_event = dctcp_cwnd_event,
0223     .ssthresh   = dctcp_ssthresh,
0224     .cong_avoid = tcp_reno_cong_avoid,
0225     .undo_cwnd  = dctcp_cwnd_undo,
0226     .set_state  = dctcp_state,
0227     .get_info   = dctcp_get_info,
0228     .flags      = TCP_CONG_NEEDS_ECN,
0229     .owner      = THIS_MODULE,
0230     .name       = "dctcp",
0231 };
0232
0233 static struct tcp_congestion_ops dctcp_reno __read_mostly = {
0234     .ssthresh   = tcp_reno_ssthresh,
0235     .cong_avoid = tcp_reno_cong_avoid,
0236     .undo_cwnd  = tcp_reno_undo_cwnd,
0237     .get_info   = dctcp_get_info,
0238     .owner      = THIS_MODULE,
0239     .name       = "dctcp-reno",
0240 };
0241
0242 BTF_SET8_START(tcp_dctcp_check_kfunc_ids)
0243 #ifdef CONFIG_X86
0244 #ifdef CONFIG_DYNAMIC_FTRACE
0245 BTF_ID_FLAGS(func, dctcp_init)
0246 BTF_ID_FLAGS(func, dctcp_update_alpha)
0247 BTF_ID_FLAGS(func, dctcp_cwnd_event)
0248 BTF_ID_FLAGS(func, dctcp_ssthresh)
0249 BTF_ID_FLAGS(func, dctcp_cwnd_undo)
0250 BTF_ID_FLAGS(func, dctcp_state)
0251 #endif
0252 #endif
0253 BTF_SET8_END(tcp_dctcp_check_kfunc_ids)
0254
0255 static const struct btf_kfunc_id_set tcp_dctcp_kfunc_set = {
0256     .owner = THIS_MODULE,
0257     .set   = &tcp_dctcp_check_kfunc_ids,
0258 };
0259
0260 static int __init dctcp_register(void)
0261 {
0262     int ret;
0263
0264     BUILD_BUG_ON(sizeof(struct dctcp) > ICSK_CA_PRIV_SIZE);
0265
0266     ret = register_btf_kfunc_id_set(BPF_PROG_TYPE_STRUCT_OPS, &tcp_dctcp_kfunc_set);
0267     if (ret < 0)
0268         return ret;
0269     return tcp_register_congestion_control(&dctcp);
0270 }
0271
0272 static void __exit dctcp_unregister(void)
0273 {
0274     tcp_unregister_congestion_control(&dctcp);
0275 }
0276
0277 module_init(dctcp_register);
0278 module_exit(dctcp_unregister);
0279
0280 MODULE_AUTHOR("Daniel Borkmann <dborkman@redhat.com>");
0281 MODULE_AUTHOR("Florian Westphal <fw@strlen.de>");
0282 MODULE_AUTHOR("Glenn Judd <glenn.judd@morganstanley.com>");
0283
0284 MODULE_LICENSE("GPL v2");
0285 MODULE_DESCRIPTION("DataCenter TCP (DCTCP)");