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 bool red_check_params(u32 qth_min, u32 qth_max, u8 Wlog) 172 { 173 if (fls(qth_min) + Wlog > 32) 174 return false; 175 if (fls(qth_max) + Wlog > 32) 176 return false; 177 if (qth_max < qth_min) 178 return false; 179 return true; 180 } 181 182 static inline void red_set_parms(struct red_parms *p, 183 u32 qth_min, u32 qth_max, u8 Wlog, u8 Plog, 184 u8 Scell_log, u8 *stab, u32 max_P) 185 { 186 int delta = qth_max - qth_min; 187 u32 max_p_delta; 188 189 p->qth_min = qth_min << Wlog; 190 p->qth_max = qth_max << Wlog; 191 p->Wlog = Wlog; 192 p->Plog = Plog; 193 if (delta <= 0) 194 delta = 1; 195 p->qth_delta = delta; 196 if (!max_P) { 197 max_P = red_maxp(Plog); 198 max_P *= delta; /* max_P = (qth_max - qth_min)/2^Plog */ 199 } 200 p->max_P = max_P; 201 max_p_delta = max_P / delta; 202 max_p_delta = max(max_p_delta, 1U); 203 p->max_P_reciprocal = reciprocal_value(max_p_delta); 204 205 /* RED Adaptative target : 206 * [min_th + 0.4*(min_th - max_th), 207 * min_th + 0.6*(min_th - max_th)]. 208 */ 209 delta /= 5; 210 p->target_min = qth_min + 2*delta; 211 p->target_max = qth_min + 3*delta; 212 213 p->Scell_log = Scell_log; 214 p->Scell_max = (255 << Scell_log); 215 216 if (stab) 217 memcpy(p->Stab, stab, sizeof(p->Stab)); 218 } 219 220 static inline int red_is_idling(const struct red_vars *v) 221 { 222 return v->qidlestart != 0; 223 } 224 225 static inline void red_start_of_idle_period(struct red_vars *v) 226 { 227 v->qidlestart = ktime_get(); 228 } 229 230 static inline void red_end_of_idle_period(struct red_vars *v) 231 { 232 v->qidlestart = 0; 233 } 234 235 static inline void red_restart(struct red_vars *v) 236 { 237 red_end_of_idle_period(v); 238 v->qavg = 0; 239 v->qcount = -1; 240 } 241 242 static inline unsigned long red_calc_qavg_from_idle_time(const struct red_parms *p, 243 const struct red_vars *v) 244 { 245 s64 delta = ktime_us_delta(ktime_get(), v->qidlestart); 246 long us_idle = min_t(s64, delta, p->Scell_max); 247 int shift; 248 249 /* 250 * The problem: ideally, average length queue recalcultion should 251 * be done over constant clock intervals. This is too expensive, so 252 * that the calculation is driven by outgoing packets. 253 * When the queue is idle we have to model this clock by hand. 254 * 255 * SF+VJ proposed to "generate": 256 * 257 * m = idletime / (average_pkt_size / bandwidth) 258 * 259 * dummy packets as a burst after idle time, i.e. 260 * 261 * v->qavg *= (1-W)^m 262 * 263 * This is an apparently overcomplicated solution (f.e. we have to 264 * precompute a table to make this calculation in reasonable time) 265 * I believe that a simpler model may be used here, 266 * but it is field for experiments. 267 */ 268 269 shift = p->Stab[(us_idle >> p->Scell_log) & RED_STAB_MASK]; 270 271 if (shift) 272 return v->qavg >> shift; 273 else { 274 /* Approximate initial part of exponent with linear function: 275 * 276 * (1-W)^m ~= 1-mW + ... 277 * 278 * Seems, it is the best solution to 279 * problem of too coarse exponent tabulation. 280 */ 281 us_idle = (v->qavg * (u64)us_idle) >> p->Scell_log; 282 283 if (us_idle < (v->qavg >> 1)) 284 return v->qavg - us_idle; 285 else 286 return v->qavg >> 1; 287 } 288 } 289 290 static inline unsigned long red_calc_qavg_no_idle_time(const struct red_parms *p, 291 const struct red_vars *v, 292 unsigned int backlog) 293 { 294 /* 295 * NOTE: v->qavg is fixed point number with point at Wlog. 296 * The formula below is equvalent to floating point 297 * version: 298 * 299 * qavg = qavg*(1-W) + backlog*W; 300 * 301 * --ANK (980924) 302 */ 303 return v->qavg + (backlog - (v->qavg >> p->Wlog)); 304 } 305 306 static inline unsigned long red_calc_qavg(const struct red_parms *p, 307 const struct red_vars *v, 308 unsigned int backlog) 309 { 310 if (!red_is_idling(v)) 311 return red_calc_qavg_no_idle_time(p, v, backlog); 312 else 313 return red_calc_qavg_from_idle_time(p, v); 314 } 315 316 317 static inline u32 red_random(const struct red_parms *p) 318 { 319 return reciprocal_divide(prandom_u32(), p->max_P_reciprocal); 320 } 321 322 static inline int red_mark_probability(const struct red_parms *p, 323 const struct red_vars *v, 324 unsigned long qavg) 325 { 326 /* The formula used below causes questions. 327 328 OK. qR is random number in the interval 329 (0..1/max_P)*(qth_max-qth_min) 330 i.e. 0..(2^Plog). If we used floating point 331 arithmetics, it would be: (2^Plog)*rnd_num, 332 where rnd_num is less 1. 333 334 Taking into account, that qavg have fixed 335 point at Wlog, two lines 336 below have the following floating point equivalent: 337 338 max_P*(qavg - qth_min)/(qth_max-qth_min) < rnd/qcount 339 340 Any questions? --ANK (980924) 341 */ 342 return !(((qavg - p->qth_min) >> p->Wlog) * v->qcount < v->qR); 343 } 344 345 enum { 346 RED_BELOW_MIN_THRESH, 347 RED_BETWEEN_TRESH, 348 RED_ABOVE_MAX_TRESH, 349 }; 350 351 static inline int red_cmp_thresh(const struct red_parms *p, unsigned long qavg) 352 { 353 if (qavg < p->qth_min) 354 return RED_BELOW_MIN_THRESH; 355 else if (qavg >= p->qth_max) 356 return RED_ABOVE_MAX_TRESH; 357 else 358 return RED_BETWEEN_TRESH; 359 } 360 361 enum { 362 RED_DONT_MARK, 363 RED_PROB_MARK, 364 RED_HARD_MARK, 365 }; 366 367 static inline int red_action(const struct red_parms *p, 368 struct red_vars *v, 369 unsigned long qavg) 370 { 371 switch (red_cmp_thresh(p, qavg)) { 372 case RED_BELOW_MIN_THRESH: 373 v->qcount = -1; 374 return RED_DONT_MARK; 375 376 case RED_BETWEEN_TRESH: 377 if (++v->qcount) { 378 if (red_mark_probability(p, v, qavg)) { 379 v->qcount = 0; 380 v->qR = red_random(p); 381 return RED_PROB_MARK; 382 } 383 } else 384 v->qR = red_random(p); 385 386 return RED_DONT_MARK; 387 388 case RED_ABOVE_MAX_TRESH: 389 v->qcount = -1; 390 return RED_HARD_MARK; 391 } 392 393 BUG(); 394 return RED_DONT_MARK; 395 } 396 397 static inline void red_adaptative_algo(struct red_parms *p, struct red_vars *v) 398 { 399 unsigned long qavg; 400 u32 max_p_delta; 401 402 qavg = v->qavg; 403 if (red_is_idling(v)) 404 qavg = red_calc_qavg_from_idle_time(p, v); 405 406 /* v->qavg is fixed point number with point at Wlog */ 407 qavg >>= p->Wlog; 408 409 if (qavg > p->target_max && p->max_P <= MAX_P_MAX) 410 p->max_P += MAX_P_ALPHA(p->max_P); /* maxp = maxp + alpha */ 411 else if (qavg < p->target_min && p->max_P >= MAX_P_MIN) 412 p->max_P = (p->max_P/10)*9; /* maxp = maxp * Beta */ 413 414 max_p_delta = DIV_ROUND_CLOSEST(p->max_P, p->qth_delta); 415 max_p_delta = max(max_p_delta, 1U); 416 p->max_P_reciprocal = reciprocal_value(max_p_delta); 417 } 418 #endif 419