1 // SPDX-License-Identifier: GPL-2.0-or-later 2 /* 3 * ip_vs_est.c: simple rate estimator for IPVS 4 * 5 * Authors: Wensong Zhang <wensong@linuxvirtualserver.org> 6 * 7 * Changes: Hans Schillstrom <hans.schillstrom@ericsson.com> 8 * Network name space (netns) aware. 9 * Global data moved to netns i.e struct netns_ipvs 10 * Affected data: est_list and est_lock. 11 * estimation_timer() runs with timer per netns. 12 * get_stats()) do the per cpu summing. 13 */ 14 15 #define KMSG_COMPONENT "IPVS" 16 #define pr_fmt(fmt) KMSG_COMPONENT ": " fmt 17 18 #include <linux/kernel.h> 19 #include <linux/jiffies.h> 20 #include <linux/types.h> 21 #include <linux/interrupt.h> 22 #include <linux/sysctl.h> 23 #include <linux/list.h> 24 25 #include <net/ip_vs.h> 26 27 /* 28 This code is to estimate rate in a shorter interval (such as 8 29 seconds) for virtual services and real servers. For measure rate in a 30 long interval, it is easy to implement a user level daemon which 31 periodically reads those statistical counters and measure rate. 32 33 Currently, the measurement is activated by slow timer handler. Hope 34 this measurement will not introduce too much load. 35 36 We measure rate during the last 8 seconds every 2 seconds: 37 38 avgrate = avgrate*(1-W) + rate*W 39 40 where W = 2^(-2) 41 42 NOTES. 43 44 * Average bps is scaled by 2^5, while average pps and cps are scaled by 2^10. 45 46 * Netlink users can see 64-bit values but sockopt users are restricted 47 to 32-bit values for conns, packets, bps, cps and pps. 48 49 * A lot of code is taken from net/core/gen_estimator.c 50 */ 51 52 53 /* 54 * Make a summary from each cpu 55 */ 56 static void ip_vs_read_cpu_stats(struct ip_vs_kstats *sum, 57 struct ip_vs_cpu_stats __percpu *stats) 58 { 59 int i; 60 bool add = false; 61 62 for_each_possible_cpu(i) { 63 struct ip_vs_cpu_stats *s = per_cpu_ptr(stats, i); 64 unsigned int start; 65 u64 conns, inpkts, outpkts, inbytes, outbytes; 66 67 if (add) { 68 do { 69 start = u64_stats_fetch_begin(&s->syncp); 70 conns = s->cnt.conns; 71 inpkts = s->cnt.inpkts; 72 outpkts = s->cnt.outpkts; 73 inbytes = s->cnt.inbytes; 74 outbytes = s->cnt.outbytes; 75 } while (u64_stats_fetch_retry(&s->syncp, start)); 76 sum->conns += conns; 77 sum->inpkts += inpkts; 78 sum->outpkts += outpkts; 79 sum->inbytes += inbytes; 80 sum->outbytes += outbytes; 81 } else { 82 add = true; 83 do { 84 start = u64_stats_fetch_begin(&s->syncp); 85 sum->conns = s->cnt.conns; 86 sum->inpkts = s->cnt.inpkts; 87 sum->outpkts = s->cnt.outpkts; 88 sum->inbytes = s->cnt.inbytes; 89 sum->outbytes = s->cnt.outbytes; 90 } while (u64_stats_fetch_retry(&s->syncp, start)); 91 } 92 } 93 } 94 95 96 static void estimation_timer(struct timer_list *t) 97 { 98 struct ip_vs_estimator *e; 99 struct ip_vs_stats *s; 100 u64 rate; 101 struct netns_ipvs *ipvs = from_timer(ipvs, t, est_timer); 102 103 if (!sysctl_run_estimation(ipvs)) 104 goto skip; 105 106 spin_lock(&ipvs->est_lock); 107 list_for_each_entry(e, &ipvs->est_list, list) { 108 s = container_of(e, struct ip_vs_stats, est); 109 110 spin_lock(&s->lock); 111 ip_vs_read_cpu_stats(&s->kstats, s->cpustats); 112 113 /* scaled by 2^10, but divided 2 seconds */ 114 rate = (s->kstats.conns - e->last_conns) << 9; 115 e->last_conns = s->kstats.conns; 116 e->cps += ((s64)rate - (s64)e->cps) >> 2; 117 118 rate = (s->kstats.inpkts - e->last_inpkts) << 9; 119 e->last_inpkts = s->kstats.inpkts; 120 e->inpps += ((s64)rate - (s64)e->inpps) >> 2; 121 122 rate = (s->kstats.outpkts - e->last_outpkts) << 9; 123 e->last_outpkts = s->kstats.outpkts; 124 e->outpps += ((s64)rate - (s64)e->outpps) >> 2; 125 126 /* scaled by 2^5, but divided 2 seconds */ 127 rate = (s->kstats.inbytes - e->last_inbytes) << 4; 128 e->last_inbytes = s->kstats.inbytes; 129 e->inbps += ((s64)rate - (s64)e->inbps) >> 2; 130 131 rate = (s->kstats.outbytes - e->last_outbytes) << 4; 132 e->last_outbytes = s->kstats.outbytes; 133 e->outbps += ((s64)rate - (s64)e->outbps) >> 2; 134 spin_unlock(&s->lock); 135 } 136 spin_unlock(&ipvs->est_lock); 137 138 skip: 139 mod_timer(&ipvs->est_timer, jiffies + 2*HZ); 140 } 141 142 void ip_vs_start_estimator(struct netns_ipvs *ipvs, struct ip_vs_stats *stats) 143 { 144 struct ip_vs_estimator *est = &stats->est; 145 146 INIT_LIST_HEAD(&est->list); 147 148 spin_lock_bh(&ipvs->est_lock); 149 list_add(&est->list, &ipvs->est_list); 150 spin_unlock_bh(&ipvs->est_lock); 151 } 152 153 void ip_vs_stop_estimator(struct netns_ipvs *ipvs, struct ip_vs_stats *stats) 154 { 155 struct ip_vs_estimator *est = &stats->est; 156 157 spin_lock_bh(&ipvs->est_lock); 158 list_del(&est->list); 159 spin_unlock_bh(&ipvs->est_lock); 160 } 161 162 void ip_vs_zero_estimator(struct ip_vs_stats *stats) 163 { 164 struct ip_vs_estimator *est = &stats->est; 165 struct ip_vs_kstats *k = &stats->kstats; 166 167 /* reset counters, caller must hold the stats->lock lock */ 168 est->last_inbytes = k->inbytes; 169 est->last_outbytes = k->outbytes; 170 est->last_conns = k->conns; 171 est->last_inpkts = k->inpkts; 172 est->last_outpkts = k->outpkts; 173 est->cps = 0; 174 est->inpps = 0; 175 est->outpps = 0; 176 est->inbps = 0; 177 est->outbps = 0; 178 } 179 180 /* Get decoded rates */ 181 void ip_vs_read_estimator(struct ip_vs_kstats *dst, struct ip_vs_stats *stats) 182 { 183 struct ip_vs_estimator *e = &stats->est; 184 185 dst->cps = (e->cps + 0x1FF) >> 10; 186 dst->inpps = (e->inpps + 0x1FF) >> 10; 187 dst->outpps = (e->outpps + 0x1FF) >> 10; 188 dst->inbps = (e->inbps + 0xF) >> 5; 189 dst->outbps = (e->outbps + 0xF) >> 5; 190 } 191 192 int __net_init ip_vs_estimator_net_init(struct netns_ipvs *ipvs) 193 { 194 INIT_LIST_HEAD(&ipvs->est_list); 195 spin_lock_init(&ipvs->est_lock); 196 timer_setup(&ipvs->est_timer, estimation_timer, 0); 197 mod_timer(&ipvs->est_timer, jiffies + 2 * HZ); 198 return 0; 199 } 200 201 void __net_exit ip_vs_estimator_net_cleanup(struct netns_ipvs *ipvs) 202 { 203 del_timer_sync(&ipvs->est_timer); 204 } 205