1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * Floating proportions with flexible aging period 4 * 5 * Copyright (C) 2011, SUSE, Jan Kara <jack@suse.cz> 6 * 7 * The goal of this code is: Given different types of event, measure proportion 8 * of each type of event over time. The proportions are measured with 9 * exponentially decaying history to give smooth transitions. A formula 10 * expressing proportion of event of type 'j' is: 11 * 12 * p_{j} = (\Sum_{i>=0} x_{i,j}/2^{i+1})/(\Sum_{i>=0} x_i/2^{i+1}) 13 * 14 * Where x_{i,j} is j's number of events in i-th last time period and x_i is 15 * total number of events in i-th last time period. 16 * 17 * Note that p_{j}'s are normalised, i.e. 18 * 19 * \Sum_{j} p_{j} = 1, 20 * 21 * This formula can be straightforwardly computed by maintaining denominator 22 * (let's call it 'd') and for each event type its numerator (let's call it 23 * 'n_j'). When an event of type 'j' happens, we simply need to do: 24 * n_j++; d++; 25 * 26 * When a new period is declared, we could do: 27 * d /= 2 28 * for each j 29 * n_j /= 2 30 * 31 * To avoid iteration over all event types, we instead shift numerator of event 32 * j lazily when someone asks for a proportion of event j or when event j 33 * occurs. This can bit trivially implemented by remembering last period in 34 * which something happened with proportion of type j. 35 */ 36 #include <linux/flex_proportions.h> 37 38 int fprop_global_init(struct fprop_global *p, gfp_t gfp) 39 { 40 int err; 41 42 p->period = 0; 43 /* Use 1 to avoid dealing with periods with 0 events... */ 44 err = percpu_counter_init(&p->events, 1, gfp); 45 if (err) 46 return err; 47 seqcount_init(&p->sequence); 48 return 0; 49 } 50 51 void fprop_global_destroy(struct fprop_global *p) 52 { 53 percpu_counter_destroy(&p->events); 54 } 55 56 /* 57 * Declare @periods new periods. It is upto the caller to make sure period 58 * transitions cannot happen in parallel. 59 * 60 * The function returns true if the proportions are still defined and false 61 * if aging zeroed out all events. This can be used to detect whether declaring 62 * further periods has any effect. 63 */ 64 bool fprop_new_period(struct fprop_global *p, int periods) 65 { 66 s64 events; 67 unsigned long flags; 68 69 local_irq_save(flags); 70 events = percpu_counter_sum(&p->events); 71 /* 72 * Don't do anything if there are no events. 73 */ 74 if (events <= 1) { 75 local_irq_restore(flags); 76 return false; 77 } 78 write_seqcount_begin(&p->sequence); 79 if (periods < 64) 80 events -= events >> periods; 81 /* Use addition to avoid losing events happening between sum and set */ 82 percpu_counter_add(&p->events, -events); 83 p->period += periods; 84 write_seqcount_end(&p->sequence); 85 local_irq_restore(flags); 86 87 return true; 88 } 89 90 /* 91 * ---- SINGLE ---- 92 */ 93 94 int fprop_local_init_single(struct fprop_local_single *pl) 95 { 96 pl->events = 0; 97 pl->period = 0; 98 raw_spin_lock_init(&pl->lock); 99 return 0; 100 } 101 102 void fprop_local_destroy_single(struct fprop_local_single *pl) 103 { 104 } 105 106 static void fprop_reflect_period_single(struct fprop_global *p, 107 struct fprop_local_single *pl) 108 { 109 unsigned int period = p->period; 110 unsigned long flags; 111 112 /* Fast path - period didn't change */ 113 if (pl->period == period) 114 return; 115 raw_spin_lock_irqsave(&pl->lock, flags); 116 /* Someone updated pl->period while we were spinning? */ 117 if (pl->period >= period) { 118 raw_spin_unlock_irqrestore(&pl->lock, flags); 119 return; 120 } 121 /* Aging zeroed our fraction? */ 122 if (period - pl->period < BITS_PER_LONG) 123 pl->events >>= period - pl->period; 124 else 125 pl->events = 0; 126 pl->period = period; 127 raw_spin_unlock_irqrestore(&pl->lock, flags); 128 } 129 130 /* Event of type pl happened */ 131 void __fprop_inc_single(struct fprop_global *p, struct fprop_local_single *pl) 132 { 133 fprop_reflect_period_single(p, pl); 134 pl->events++; 135 percpu_counter_add(&p->events, 1); 136 } 137 138 /* Return fraction of events of type pl */ 139 void fprop_fraction_single(struct fprop_global *p, 140 struct fprop_local_single *pl, 141 unsigned long *numerator, unsigned long *denominator) 142 { 143 unsigned int seq; 144 s64 num, den; 145 146 do { 147 seq = read_seqcount_begin(&p->sequence); 148 fprop_reflect_period_single(p, pl); 149 num = pl->events; 150 den = percpu_counter_read_positive(&p->events); 151 } while (read_seqcount_retry(&p->sequence, seq)); 152 153 /* 154 * Make fraction <= 1 and denominator > 0 even in presence of percpu 155 * counter errors 156 */ 157 if (den <= num) { 158 if (num) 159 den = num; 160 else 161 den = 1; 162 } 163 *denominator = den; 164 *numerator = num; 165 } 166 167 /* 168 * ---- PERCPU ---- 169 */ 170 #define PROP_BATCH (8*(1+ilog2(nr_cpu_ids))) 171 172 int fprop_local_init_percpu(struct fprop_local_percpu *pl, gfp_t gfp) 173 { 174 int err; 175 176 err = percpu_counter_init(&pl->events, 0, gfp); 177 if (err) 178 return err; 179 pl->period = 0; 180 raw_spin_lock_init(&pl->lock); 181 return 0; 182 } 183 184 void fprop_local_destroy_percpu(struct fprop_local_percpu *pl) 185 { 186 percpu_counter_destroy(&pl->events); 187 } 188 189 static void fprop_reflect_period_percpu(struct fprop_global *p, 190 struct fprop_local_percpu *pl) 191 { 192 unsigned int period = p->period; 193 unsigned long flags; 194 195 /* Fast path - period didn't change */ 196 if (pl->period == period) 197 return; 198 raw_spin_lock_irqsave(&pl->lock, flags); 199 /* Someone updated pl->period while we were spinning? */ 200 if (pl->period >= period) { 201 raw_spin_unlock_irqrestore(&pl->lock, flags); 202 return; 203 } 204 /* Aging zeroed our fraction? */ 205 if (period - pl->period < BITS_PER_LONG) { 206 s64 val = percpu_counter_read(&pl->events); 207 208 if (val < (nr_cpu_ids * PROP_BATCH)) 209 val = percpu_counter_sum(&pl->events); 210 211 percpu_counter_add_batch(&pl->events, 212 -val + (val >> (period-pl->period)), PROP_BATCH); 213 } else 214 percpu_counter_set(&pl->events, 0); 215 pl->period = period; 216 raw_spin_unlock_irqrestore(&pl->lock, flags); 217 } 218 219 /* Event of type pl happened */ 220 void __fprop_add_percpu(struct fprop_global *p, struct fprop_local_percpu *pl, 221 long nr) 222 { 223 fprop_reflect_period_percpu(p, pl); 224 percpu_counter_add_batch(&pl->events, nr, PROP_BATCH); 225 percpu_counter_add(&p->events, nr); 226 } 227 228 void fprop_fraction_percpu(struct fprop_global *p, 229 struct fprop_local_percpu *pl, 230 unsigned long *numerator, unsigned long *denominator) 231 { 232 unsigned int seq; 233 s64 num, den; 234 235 do { 236 seq = read_seqcount_begin(&p->sequence); 237 fprop_reflect_period_percpu(p, pl); 238 num = percpu_counter_read_positive(&pl->events); 239 den = percpu_counter_read_positive(&p->events); 240 } while (read_seqcount_retry(&p->sequence, seq)); 241 242 /* 243 * Make fraction <= 1 and denominator > 0 even in presence of percpu 244 * counter errors 245 */ 246 if (den <= num) { 247 if (num) 248 den = num; 249 else 250 den = 1; 251 } 252 *denominator = den; 253 *numerator = num; 254 } 255 256 /* 257 * Like __fprop_add_percpu() except that event is counted only if the given 258 * type has fraction smaller than @max_frac/FPROP_FRAC_BASE 259 */ 260 void __fprop_add_percpu_max(struct fprop_global *p, 261 struct fprop_local_percpu *pl, int max_frac, long nr) 262 { 263 if (unlikely(max_frac < FPROP_FRAC_BASE)) { 264 unsigned long numerator, denominator; 265 s64 tmp; 266 267 fprop_fraction_percpu(p, pl, &numerator, &denominator); 268 /* Adding 'nr' to fraction exceeds max_frac/FPROP_FRAC_BASE? */ 269 tmp = (u64)denominator * max_frac - 270 ((u64)numerator << FPROP_FRAC_SHIFT); 271 if (tmp < 0) { 272 /* Maximum fraction already exceeded? */ 273 return; 274 } else if (tmp < nr * (FPROP_FRAC_BASE - max_frac)) { 275 /* Add just enough for the fraction to saturate */ 276 nr = div_u64(tmp + FPROP_FRAC_BASE - max_frac - 1, 277 FPROP_FRAC_BASE - max_frac); 278 } 279 } 280 281 __fprop_add_percpu(p, pl, nr); 282 } 283