hbm_kern.h (5837B)
1/* SPDX-License-Identifier: GPL-2.0 2 * 3 * Copyright (c) 2019 Facebook 4 * 5 * This program is free software; you can redistribute it and/or 6 * modify it under the terms of version 2 of the GNU General Public 7 * License as published by the Free Software Foundation. 8 * 9 * Include file for sample Host Bandwidth Manager (HBM) BPF programs 10 */ 11#define KBUILD_MODNAME "foo" 12#include <uapi/linux/bpf.h> 13#include <uapi/linux/if_ether.h> 14#include <uapi/linux/if_packet.h> 15#include <uapi/linux/ip.h> 16#include <uapi/linux/ipv6.h> 17#include <uapi/linux/in.h> 18#include <uapi/linux/tcp.h> 19#include <uapi/linux/filter.h> 20#include <uapi/linux/pkt_cls.h> 21#include <net/ipv6.h> 22#include <net/inet_ecn.h> 23#include <bpf/bpf_endian.h> 24#include <bpf/bpf_helpers.h> 25#include "hbm.h" 26 27#define DROP_PKT 0 28#define ALLOW_PKT 1 29#define TCP_ECN_OK 1 30#define CWR 2 31 32#ifndef HBM_DEBUG // Define HBM_DEBUG to enable debugging 33#undef bpf_printk 34#define bpf_printk(fmt, ...) 35#endif 36 37#define INITIAL_CREDIT_PACKETS 100 38#define MAX_BYTES_PER_PACKET 1500 39#define MARK_THRESH (40 * MAX_BYTES_PER_PACKET) 40#define DROP_THRESH (80 * 5 * MAX_BYTES_PER_PACKET) 41#define LARGE_PKT_DROP_THRESH (DROP_THRESH - (15 * MAX_BYTES_PER_PACKET)) 42#define MARK_REGION_SIZE (LARGE_PKT_DROP_THRESH - MARK_THRESH) 43#define LARGE_PKT_THRESH 120 44#define MAX_CREDIT (100 * MAX_BYTES_PER_PACKET) 45#define INIT_CREDIT (INITIAL_CREDIT_PACKETS * MAX_BYTES_PER_PACKET) 46 47// Time base accounting for fq's EDT 48#define BURST_SIZE_NS 100000 // 100us 49#define MARK_THRESH_NS 50000 // 50us 50#define DROP_THRESH_NS 500000 // 500us 51// Reserve 20us of queuing for small packets (less than 120 bytes) 52#define LARGE_PKT_DROP_THRESH_NS (DROP_THRESH_NS - 20000) 53#define MARK_REGION_SIZE_NS (LARGE_PKT_DROP_THRESH_NS - MARK_THRESH_NS) 54 55// rate in bytes per ns << 20 56#define CREDIT_PER_NS(delta, rate) ((((u64)(delta)) * (rate)) >> 20) 57#define BYTES_PER_NS(delta, rate) ((((u64)(delta)) * (rate)) >> 20) 58#define BYTES_TO_NS(bytes, rate) div64_u64(((u64)(bytes)) << 20, (u64)(rate)) 59 60struct { 61 __uint(type, BPF_MAP_TYPE_CGROUP_STORAGE); 62 __type(key, struct bpf_cgroup_storage_key); 63 __type(value, struct hbm_vqueue); 64} queue_state SEC(".maps"); 65 66struct { 67 __uint(type, BPF_MAP_TYPE_ARRAY); 68 __uint(max_entries, 1); 69 __type(key, u32); 70 __type(value, struct hbm_queue_stats); 71} queue_stats SEC(".maps"); 72 73struct hbm_pkt_info { 74 int cwnd; 75 int rtt; 76 int packets_out; 77 bool is_ip; 78 bool is_tcp; 79 short ecn; 80}; 81 82static int get_tcp_info(struct __sk_buff *skb, struct hbm_pkt_info *pkti) 83{ 84 struct bpf_sock *sk; 85 struct bpf_tcp_sock *tp; 86 87 sk = skb->sk; 88 if (sk) { 89 sk = bpf_sk_fullsock(sk); 90 if (sk) { 91 if (sk->protocol == IPPROTO_TCP) { 92 tp = bpf_tcp_sock(sk); 93 if (tp) { 94 pkti->cwnd = tp->snd_cwnd; 95 pkti->rtt = tp->srtt_us >> 3; 96 pkti->packets_out = tp->packets_out; 97 return 0; 98 } 99 } 100 } 101 } 102 pkti->cwnd = 0; 103 pkti->rtt = 0; 104 pkti->packets_out = 0; 105 return 1; 106} 107 108static void hbm_get_pkt_info(struct __sk_buff *skb, 109 struct hbm_pkt_info *pkti) 110{ 111 struct iphdr iph; 112 struct ipv6hdr *ip6h; 113 114 pkti->cwnd = 0; 115 pkti->rtt = 0; 116 bpf_skb_load_bytes(skb, 0, &iph, 12); 117 if (iph.version == 6) { 118 ip6h = (struct ipv6hdr *)&iph; 119 pkti->is_ip = true; 120 pkti->is_tcp = (ip6h->nexthdr == 6); 121 pkti->ecn = (ip6h->flow_lbl[0] >> 4) & INET_ECN_MASK; 122 } else if (iph.version == 4) { 123 pkti->is_ip = true; 124 pkti->is_tcp = (iph.protocol == 6); 125 pkti->ecn = iph.tos & INET_ECN_MASK; 126 } else { 127 pkti->is_ip = false; 128 pkti->is_tcp = false; 129 pkti->ecn = 0; 130 } 131 if (pkti->is_tcp) 132 get_tcp_info(skb, pkti); 133} 134 135static __always_inline void hbm_init_vqueue(struct hbm_vqueue *qdp, int rate) 136{ 137 bpf_printk("Initializing queue_state, rate:%d\n", rate * 128); 138 qdp->lasttime = bpf_ktime_get_ns(); 139 qdp->credit = INIT_CREDIT; 140 qdp->rate = rate * 128; 141} 142 143static __always_inline void hbm_init_edt_vqueue(struct hbm_vqueue *qdp, 144 int rate) 145{ 146 unsigned long long curtime; 147 148 curtime = bpf_ktime_get_ns(); 149 bpf_printk("Initializing queue_state, rate:%d\n", rate * 128); 150 qdp->lasttime = curtime - BURST_SIZE_NS; // support initial burst 151 qdp->credit = 0; // not used 152 qdp->rate = rate * 128; 153} 154 155static __always_inline void hbm_update_stats(struct hbm_queue_stats *qsp, 156 int len, 157 unsigned long long curtime, 158 bool congestion_flag, 159 bool drop_flag, 160 bool cwr_flag, 161 bool ecn_ce_flag, 162 struct hbm_pkt_info *pkti, 163 int credit) 164{ 165 int rv = ALLOW_PKT; 166 167 if (qsp != NULL) { 168 // Following is needed for work conserving 169 __sync_add_and_fetch(&(qsp->bytes_total), len); 170 if (qsp->stats) { 171 // Optionally update statistics 172 if (qsp->firstPacketTime == 0) 173 qsp->firstPacketTime = curtime; 174 qsp->lastPacketTime = curtime; 175 __sync_add_and_fetch(&(qsp->pkts_total), 1); 176 if (congestion_flag) { 177 __sync_add_and_fetch(&(qsp->pkts_marked), 1); 178 __sync_add_and_fetch(&(qsp->bytes_marked), len); 179 } 180 if (drop_flag) { 181 __sync_add_and_fetch(&(qsp->pkts_dropped), 1); 182 __sync_add_and_fetch(&(qsp->bytes_dropped), 183 len); 184 } 185 if (ecn_ce_flag) 186 __sync_add_and_fetch(&(qsp->pkts_ecn_ce), 1); 187 if (pkti->cwnd) { 188 __sync_add_and_fetch(&(qsp->sum_cwnd), 189 pkti->cwnd); 190 __sync_add_and_fetch(&(qsp->sum_cwnd_cnt), 1); 191 } 192 if (pkti->rtt) 193 __sync_add_and_fetch(&(qsp->sum_rtt), 194 pkti->rtt); 195 __sync_add_and_fetch(&(qsp->sum_credit), credit); 196 197 if (drop_flag) 198 rv = DROP_PKT; 199 if (cwr_flag) 200 rv |= 2; 201 if (rv == DROP_PKT) 202 __sync_add_and_fetch(&(qsp->returnValCount[0]), 203 1); 204 else if (rv == ALLOW_PKT) 205 __sync_add_and_fetch(&(qsp->returnValCount[1]), 206 1); 207 else if (rv == 2) 208 __sync_add_and_fetch(&(qsp->returnValCount[2]), 209 1); 210 else if (rv == 3) 211 __sync_add_and_fetch(&(qsp->returnValCount[3]), 212 1); 213 } 214 } 215}