1 /* SPDX-License-Identifier: GPL-2.0 */ 2 #ifndef __NET_SCHED_RED_H 3 #define __NET_SCHED_RED_H 4 5 #include <linux/types.h> 6 #include <linux/bug.h> 7 #include <net/pkt_sched.h> 8 #include <net/inet_ecn.h> 9 #include <net/dsfield.h> 10 #include <linux/reciprocal_div.h> 11 12 /* Random Early Detection (RED) algorithm. 13 ======================================= 14 15 Source: Sally Floyd and Van Jacobson, "Random Early Detection Gateways 16 for Congestion Avoidance", 1993, IEEE/ACM Transactions on Networking. 17 18 This file codes a "divisionless" version of RED algorithm 19 as written down in Fig.17 of the paper. 20 21 Short description. 22 ------------------ 23 24 When a new packet arrives we calculate the average queue length: 25 26 avg = (1-W)*avg + W*current_queue_len, 27 28 W is the filter time constant (chosen as 2^(-Wlog)), it controls 29 the inertia of the algorithm. To allow larger bursts, W should be 30 decreased. 31 32 if (avg > th_max) -> packet marked (dropped). 33 if (avg < th_min) -> packet passes. 34 if (th_min < avg < th_max) we calculate probability: 35 36 Pb = max_P * (avg - th_min)/(th_max-th_min) 37 38 and mark (drop) packet with this probability. 39 Pb changes from 0 (at avg==th_min) to max_P (avg==th_max). 40 max_P should be small (not 1), usually 0.01..0.02 is good value. 41 42 max_P is chosen as a number, so that max_P/(th_max-th_min) 43 is a negative power of two in order arithmetics to contain 44 only shifts. 45 46 47 Parameters, settable by user: 48 ----------------------------- 49 50 qth_min - bytes (should be < qth_max/2) 51 qth_max - bytes (should be at least 2*qth_min and less limit) 52 Wlog - bits (<32) log(1/W). 53 Plog - bits (<32) 54 55 Plog is related to max_P by formula: 56 57 max_P = (qth_max-qth_min)/2^Plog; 58 59 F.e. if qth_max=128K and qth_min=32K, then Plog=22 60 corresponds to max_P=0.02 61 62 Scell_log 63 Stab 64 65 Lookup table for log((1-W)^(t/t_ave). 66 67 68 NOTES: 69 70 Upper bound on W. 71 ----------------- 72 73 If you want to allow bursts of L packets of size S, 74 you should choose W: 75 76 L + 1 - th_min/S < (1-(1-W)^L)/W 77 78 th_min/S = 32 th_min/S = 4 79 80 log(W) L 81 -1 33 82 -2 35 83 -3 39 84 -4 46 85 -5 57 86 -6 75 87 -7 101 88 -8 135 89 -9 190 90 etc. 91 */ 92 93 /* 94 * Adaptative RED : An Algorithm for Increasing the Robustness of RED's AQM 95 * (Sally FLoyd, Ramakrishna Gummadi, and Scott Shenker) August 2001 96 * 97 * Every 500 ms: 98 * if (avg > target and max_p <= 0.5) 99 * increase max_p : max_p += alpha; 100 * else if (avg < target and max_p >= 0.01) 101 * decrease max_p : max_p *= beta; 102 * 103 * target :[qth_min + 0.4*(qth_min - qth_max), 104 * qth_min + 0.6*(qth_min - qth_max)]. 105 * alpha : min(0.01, max_p / 4) 106 * beta : 0.9 107 * max_P is a Q0.32 fixed point number (with 32 bits mantissa) 108 * max_P between 0.01 and 0.5 (1% - 50%) [ Its no longer a negative power of two ] 109 */ 110 #define RED_ONE_PERCENT ((u32)DIV_ROUND_CLOSEST(1ULL<<32, 100)) 111 112 #define MAX_P_MIN (1 * RED_ONE_PERCENT) 113 #define MAX_P_MAX (50 * RED_ONE_PERCENT) 114 #define MAX_P_ALPHA(val) min(MAX_P_MIN, val / 4) 115 116 #define RED_STAB_SIZE 256 117 #define RED_STAB_MASK (RED_STAB_SIZE - 1) 118 119 struct red_stats { 120 u32 prob_drop; /* Early probability drops */ 121 u32 prob_mark; /* Early probability marks */ 122 u32 forced_drop; /* Forced drops, qavg > max_thresh */ 123 u32 forced_mark; /* Forced marks, qavg > max_thresh */ 124 u32 pdrop; /* Drops due to queue limits */ 125 u32 other; /* Drops due to drop() calls */ 126 }; 127 128 struct red_parms { 129 /* Parameters */ 130 u32 qth_min; /* Min avg length threshold: Wlog scaled */ 131 u32 qth_max; /* Max avg length threshold: Wlog scaled */ 132 u32 Scell_max; 133 u32 max_P; /* probability, [0 .. 1.0] 32 scaled */ 134 /* reciprocal_value(max_P / qth_delta) */ 135 struct reciprocal_value max_P_reciprocal; 136 u32 qth_delta; /* max_th - min_th */ 137 u32 target_min; /* min_th + 0.4*(max_th - min_th) */ 138 u32 target_max; /* min_th + 0.6*(max_th - min_th) */ 139 u8 Scell_log; 140 u8 Wlog; /* log(W) */ 141 u8 Plog; /* random number bits */ 142 u8 Stab[RED_STAB_SIZE]; 143 }; 144 145 struct red_vars { 146 /* Variables */ 147 int qcount; /* Number of packets since last random 148 number generation */ 149 u32 qR; /* Cached random number */ 150 151 unsigned long qavg; /* Average queue length: Wlog scaled */ 152 ktime_t qidlestart; /* Start of current idle period */ 153 }; 154 155 static inline u32 red_maxp(u8 Plog) 156 { 157 return Plog < 32 ? (~0U >> Plog) : ~0U; 158 } 159 160 static inline void red_set_vars(struct red_vars *v) 161 { 162 /* Reset average queue length, the value is strictly bound 163 * to the parameters below, reseting hurts a bit but leaving 164 * it might result in an unreasonable qavg for a while. --TGR 165 */ 166 v->qavg = 0; 167 168 v->qcount = -1; 169 } 170 171 static inline void red_set_parms(struct red_parms *p, 172 u32 qth_min, u32 qth_max, u8 Wlog, u8 Plog, 173 u8 Scell_log, u8 *stab, u32 max_P) 174 { 175 int delta = qth_max - qth_min; 176 u32 max_p_delta; 177 178 p->qth_min = qth_min << Wlog; 179 p->qth_max = qth_max << Wlog; 180 p->Wlog = Wlog; 181 p->Plog = Plog; 182 if (delta < 0) 183 delta = 1; 184 p->qth_delta = delta; 185 if (!max_P) { 186 max_P = red_maxp(Plog); 187 max_P *= delta; /* max_P = (qth_max - qth_min)/2^Plog */ 188 } 189 p->max_P = max_P; 190 max_p_delta = max_P / delta; 191 max_p_delta = max(max_p_delta, 1U); 192 p->max_P_reciprocal = reciprocal_value(max_p_delta); 193 194 /* RED Adaptative target : 195 * [min_th + 0.4*(min_th - max_th), 196 * min_th + 0.6*(min_th - max_th)]. 197 */ 198 delta /= 5; 199 p->target_min = qth_min + 2*delta; 200 p->target_max = qth_min + 3*delta; 201 202 p->Scell_log = Scell_log; 203 p->Scell_max = (255 << Scell_log); 204 205 if (stab) 206 memcpy(p->Stab, stab, sizeof(p->Stab)); 207 } 208 209 static inline int red_is_idling(const struct red_vars *v) 210 { 211 return v->qidlestart != 0; 212 } 213 214 static inline void red_start_of_idle_period(struct red_vars *v) 215 { 216 v->qidlestart = ktime_get(); 217 } 218 219 static inline void red_end_of_idle_period(struct red_vars *v) 220 { 221 v->qidlestart = 0; 222 } 223 224 static inline void red_restart(struct red_vars *v) 225 { 226 red_end_of_idle_period(v); 227 v->qavg = 0; 228 v->qcount = -1; 229 } 230 231 static inline unsigned long red_calc_qavg_from_idle_time(const struct red_parms *p, 232 const struct red_vars *v) 233 { 234 s64 delta = ktime_us_delta(ktime_get(), v->qidlestart); 235 long us_idle = min_t(s64, delta, p->Scell_max); 236 int shift; 237 238 /* 239 * The problem: ideally, average length queue recalcultion should 240 * be done over constant clock intervals. This is too expensive, so 241 * that the calculation is driven by outgoing packets. 242 * When the queue is idle we have to model this clock by hand. 243 * 244 * SF+VJ proposed to "generate": 245 * 246 * m = idletime / (average_pkt_size / bandwidth) 247 * 248 * dummy packets as a burst after idle time, i.e. 249 * 250 * v->qavg *= (1-W)^m 251 * 252 * This is an apparently overcomplicated solution (f.e. we have to 253 * precompute a table to make this calculation in reasonable time) 254 * I believe that a simpler model may be used here, 255 * but it is field for experiments. 256 */ 257 258 shift = p->Stab[(us_idle >> p->Scell_log) & RED_STAB_MASK]; 259 260 if (shift) 261 return v->qavg >> shift; 262 else { 263 /* Approximate initial part of exponent with linear function: 264 * 265 * (1-W)^m ~= 1-mW + ... 266 * 267 * Seems, it is the best solution to 268 * problem of too coarse exponent tabulation. 269 */ 270 us_idle = (v->qavg * (u64)us_idle) >> p->Scell_log; 271 272 if (us_idle < (v->qavg >> 1)) 273 return v->qavg - us_idle; 274 else 275 return v->qavg >> 1; 276 } 277 } 278 279 static inline unsigned long red_calc_qavg_no_idle_time(const struct red_parms *p, 280 const struct red_vars *v, 281 unsigned int backlog) 282 { 283 /* 284 * NOTE: v->qavg is fixed point number with point at Wlog. 285 * The formula below is equvalent to floating point 286 * version: 287 * 288 * qavg = qavg*(1-W) + backlog*W; 289 * 290 * --ANK (980924) 291 */ 292 return v->qavg + (backlog - (v->qavg >> p->Wlog)); 293 } 294 295 static inline unsigned long red_calc_qavg(const struct red_parms *p, 296 const struct red_vars *v, 297 unsigned int backlog) 298 { 299 if (!red_is_idling(v)) 300 return red_calc_qavg_no_idle_time(p, v, backlog); 301 else 302 return red_calc_qavg_from_idle_time(p, v); 303 } 304 305 306 static inline u32 red_random(const struct red_parms *p) 307 { 308 return reciprocal_divide(prandom_u32(), p->max_P_reciprocal); 309 } 310 311 static inline int red_mark_probability(const struct red_parms *p, 312 const struct red_vars *v, 313 unsigned long qavg) 314 { 315 /* The formula used below causes questions. 316 317 OK. qR is random number in the interval 318 (0..1/max_P)*(qth_max-qth_min) 319 i.e. 0..(2^Plog). If we used floating point 320 arithmetics, it would be: (2^Plog)*rnd_num, 321 where rnd_num is less 1. 322 323 Taking into account, that qavg have fixed 324 point at Wlog, two lines 325 below have the following floating point equivalent: 326 327 max_P*(qavg - qth_min)/(qth_max-qth_min) < rnd/qcount 328 329 Any questions? --ANK (980924) 330 */ 331 return !(((qavg - p->qth_min) >> p->Wlog) * v->qcount < v->qR); 332 } 333 334 enum { 335 RED_BELOW_MIN_THRESH, 336 RED_BETWEEN_TRESH, 337 RED_ABOVE_MAX_TRESH, 338 }; 339 340 static inline int red_cmp_thresh(const struct red_parms *p, unsigned long qavg) 341 { 342 if (qavg < p->qth_min) 343 return RED_BELOW_MIN_THRESH; 344 else if (qavg >= p->qth_max) 345 return RED_ABOVE_MAX_TRESH; 346 else 347 return RED_BETWEEN_TRESH; 348 } 349 350 enum { 351 RED_DONT_MARK, 352 RED_PROB_MARK, 353 RED_HARD_MARK, 354 }; 355 356 static inline int red_action(const struct red_parms *p, 357 struct red_vars *v, 358 unsigned long qavg) 359 { 360 switch (red_cmp_thresh(p, qavg)) { 361 case RED_BELOW_MIN_THRESH: 362 v->qcount = -1; 363 return RED_DONT_MARK; 364 365 case RED_BETWEEN_TRESH: 366 if (++v->qcount) { 367 if (red_mark_probability(p, v, qavg)) { 368 v->qcount = 0; 369 v->qR = red_random(p); 370 return RED_PROB_MARK; 371 } 372 } else 373 v->qR = red_random(p); 374 375 return RED_DONT_MARK; 376 377 case RED_ABOVE_MAX_TRESH: 378 v->qcount = -1; 379 return RED_HARD_MARK; 380 } 381 382 BUG(); 383 return RED_DONT_MARK; 384 } 385 386 static inline void red_adaptative_algo(struct red_parms *p, struct red_vars *v) 387 { 388 unsigned long qavg; 389 u32 max_p_delta; 390 391 qavg = v->qavg; 392 if (red_is_idling(v)) 393 qavg = red_calc_qavg_from_idle_time(p, v); 394 395 /* v->qavg is fixed point number with point at Wlog */ 396 qavg >>= p->Wlog; 397 398 if (qavg > p->target_max && p->max_P <= MAX_P_MAX) 399 p->max_P += MAX_P_ALPHA(p->max_P); /* maxp = maxp + alpha */ 400 else if (qavg < p->target_min && p->max_P >= MAX_P_MIN) 401 p->max_P = (p->max_P/10)*9; /* maxp = maxp * Beta */ 402 403 max_p_delta = DIV_ROUND_CLOSEST(p->max_P, p->qth_delta); 404 max_p_delta = max(max_p_delta, 1U); 405 p->max_P_reciprocal = reciprocal_value(max_p_delta); 406 } 407 #endif 408