1 #ifndef __NET_SCHED_RED_H 2 #define __NET_SCHED_RED_H 3 4 #include <linux/types.h> 5 #include <linux/bug.h> 6 #include <net/pkt_sched.h> 7 #include <net/inet_ecn.h> 8 #include <net/dsfield.h> 9 #include <linux/reciprocal_div.h> 10 11 /* Random Early Detection (RED) algorithm. 12 ======================================= 13 14 Source: Sally Floyd and Van Jacobson, "Random Early Detection Gateways 15 for Congestion Avoidance", 1993, IEEE/ACM Transactions on Networking. 16 17 This file codes a "divisionless" version of RED algorithm 18 as written down in Fig.17 of the paper. 19 20 Short description. 21 ------------------ 22 23 When a new packet arrives we calculate the average queue length: 24 25 avg = (1-W)*avg + W*current_queue_len, 26 27 W is the filter time constant (chosen as 2^(-Wlog)), it controls 28 the inertia of the algorithm. To allow larger bursts, W should be 29 decreased. 30 31 if (avg > th_max) -> packet marked (dropped). 32 if (avg < th_min) -> packet passes. 33 if (th_min < avg < th_max) we calculate probability: 34 35 Pb = max_P * (avg - th_min)/(th_max-th_min) 36 37 and mark (drop) packet with this probability. 38 Pb changes from 0 (at avg==th_min) to max_P (avg==th_max). 39 max_P should be small (not 1), usually 0.01..0.02 is good value. 40 41 max_P is chosen as a number, so that max_P/(th_max-th_min) 42 is a negative power of two in order arithmetics to contain 43 only shifts. 44 45 46 Parameters, settable by user: 47 ----------------------------- 48 49 qth_min - bytes (should be < qth_max/2) 50 qth_max - bytes (should be at least 2*qth_min and less limit) 51 Wlog - bits (<32) log(1/W). 52 Plog - bits (<32) 53 54 Plog is related to max_P by formula: 55 56 max_P = (qth_max-qth_min)/2^Plog; 57 58 F.e. if qth_max=128K and qth_min=32K, then Plog=22 59 corresponds to max_P=0.02 60 61 Scell_log 62 Stab 63 64 Lookup table for log((1-W)^(t/t_ave). 65 66 67 NOTES: 68 69 Upper bound on W. 70 ----------------- 71 72 If you want to allow bursts of L packets of size S, 73 you should choose W: 74 75 L + 1 - th_min/S < (1-(1-W)^L)/W 76 77 th_min/S = 32 th_min/S = 4 78 79 log(W) L 80 -1 33 81 -2 35 82 -3 39 83 -4 46 84 -5 57 85 -6 75 86 -7 101 87 -8 135 88 -9 190 89 etc. 90 */ 91 92 /* 93 * Adaptative RED : An Algorithm for Increasing the Robustness of RED's AQM 94 * (Sally FLoyd, Ramakrishna Gummadi, and Scott Shenker) August 2001 95 * 96 * Every 500 ms: 97 * if (avg > target and max_p <= 0.5) 98 * increase max_p : max_p += alpha; 99 * else if (avg < target and max_p >= 0.01) 100 * decrease max_p : max_p *= beta; 101 * 102 * target :[qth_min + 0.4*(qth_min - qth_max), 103 * qth_min + 0.6*(qth_min - qth_max)]. 104 * alpha : min(0.01, max_p / 4) 105 * beta : 0.9 106 * max_P is a Q0.32 fixed point number (with 32 bits mantissa) 107 * max_P between 0.01 and 0.5 (1% - 50%) [ Its no longer a negative power of two ] 108 */ 109 #define RED_ONE_PERCENT ((u32)DIV_ROUND_CLOSEST(1ULL<<32, 100)) 110 111 #define MAX_P_MIN (1 * RED_ONE_PERCENT) 112 #define MAX_P_MAX (50 * RED_ONE_PERCENT) 113 #define MAX_P_ALPHA(val) min(MAX_P_MIN, val / 4) 114 115 #define RED_STAB_SIZE 256 116 #define RED_STAB_MASK (RED_STAB_SIZE - 1) 117 118 struct red_stats { 119 u32 prob_drop; /* Early probability drops */ 120 u32 prob_mark; /* Early probability marks */ 121 u32 forced_drop; /* Forced drops, qavg > max_thresh */ 122 u32 forced_mark; /* Forced marks, qavg > max_thresh */ 123 u32 pdrop; /* Drops due to queue limits */ 124 u32 other; /* Drops due to drop() calls */ 125 }; 126 127 struct red_parms { 128 /* Parameters */ 129 u32 qth_min; /* Min avg length threshold: Wlog scaled */ 130 u32 qth_max; /* Max avg length threshold: Wlog scaled */ 131 u32 Scell_max; 132 u32 max_P; /* probability, [0 .. 1.0] 32 scaled */ 133 /* reciprocal_value(max_P / qth_delta) */ 134 struct reciprocal_value max_P_reciprocal; 135 u32 qth_delta; /* max_th - min_th */ 136 u32 target_min; /* min_th + 0.4*(max_th - min_th) */ 137 u32 target_max; /* min_th + 0.6*(max_th - min_th) */ 138 u8 Scell_log; 139 u8 Wlog; /* log(W) */ 140 u8 Plog; /* random number bits */ 141 u8 Stab[RED_STAB_SIZE]; 142 }; 143 144 struct red_vars { 145 /* Variables */ 146 int qcount; /* Number of packets since last random 147 number generation */ 148 u32 qR; /* Cached random number */ 149 150 unsigned long qavg; /* Average queue length: Wlog scaled */ 151 ktime_t qidlestart; /* Start of current idle period */ 152 }; 153 154 static inline u32 red_maxp(u8 Plog) 155 { 156 return Plog < 32 ? (~0U >> Plog) : ~0U; 157 } 158 159 static inline void red_set_vars(struct red_vars *v) 160 { 161 /* Reset average queue length, the value is strictly bound 162 * to the parameters below, reseting hurts a bit but leaving 163 * it might result in an unreasonable qavg for a while. --TGR 164 */ 165 v->qavg = 0; 166 167 v->qcount = -1; 168 } 169 170 static inline void red_set_parms(struct red_parms *p, 171 u32 qth_min, u32 qth_max, u8 Wlog, u8 Plog, 172 u8 Scell_log, u8 *stab, u32 max_P) 173 { 174 int delta = qth_max - qth_min; 175 u32 max_p_delta; 176 177 p->qth_min = qth_min << Wlog; 178 p->qth_max = qth_max << Wlog; 179 p->Wlog = Wlog; 180 p->Plog = Plog; 181 if (delta < 0) 182 delta = 1; 183 p->qth_delta = delta; 184 if (!max_P) { 185 max_P = red_maxp(Plog); 186 max_P *= delta; /* max_P = (qth_max - qth_min)/2^Plog */ 187 } 188 p->max_P = max_P; 189 max_p_delta = max_P / delta; 190 max_p_delta = max(max_p_delta, 1U); 191 p->max_P_reciprocal = reciprocal_value(max_p_delta); 192 193 /* RED Adaptative target : 194 * [min_th + 0.4*(min_th - max_th), 195 * min_th + 0.6*(min_th - max_th)]. 196 */ 197 delta /= 5; 198 p->target_min = qth_min + 2*delta; 199 p->target_max = qth_min + 3*delta; 200 201 p->Scell_log = Scell_log; 202 p->Scell_max = (255 << Scell_log); 203 204 if (stab) 205 memcpy(p->Stab, stab, sizeof(p->Stab)); 206 } 207 208 static inline int red_is_idling(const struct red_vars *v) 209 { 210 return v->qidlestart.tv64 != 0; 211 } 212 213 static inline void red_start_of_idle_period(struct red_vars *v) 214 { 215 v->qidlestart = ktime_get(); 216 } 217 218 static inline void red_end_of_idle_period(struct red_vars *v) 219 { 220 v->qidlestart.tv64 = 0; 221 } 222 223 static inline void red_restart(struct red_vars *v) 224 { 225 red_end_of_idle_period(v); 226 v->qavg = 0; 227 v->qcount = -1; 228 } 229 230 static inline unsigned long red_calc_qavg_from_idle_time(const struct red_parms *p, 231 const struct red_vars *v) 232 { 233 s64 delta = ktime_us_delta(ktime_get(), v->qidlestart); 234 long us_idle = min_t(s64, delta, p->Scell_max); 235 int shift; 236 237 /* 238 * The problem: ideally, average length queue recalcultion should 239 * be done over constant clock intervals. This is too expensive, so 240 * that the calculation is driven by outgoing packets. 241 * When the queue is idle we have to model this clock by hand. 242 * 243 * SF+VJ proposed to "generate": 244 * 245 * m = idletime / (average_pkt_size / bandwidth) 246 * 247 * dummy packets as a burst after idle time, i.e. 248 * 249 * v->qavg *= (1-W)^m 250 * 251 * This is an apparently overcomplicated solution (f.e. we have to 252 * precompute a table to make this calculation in reasonable time) 253 * I believe that a simpler model may be used here, 254 * but it is field for experiments. 255 */ 256 257 shift = p->Stab[(us_idle >> p->Scell_log) & RED_STAB_MASK]; 258 259 if (shift) 260 return v->qavg >> shift; 261 else { 262 /* Approximate initial part of exponent with linear function: 263 * 264 * (1-W)^m ~= 1-mW + ... 265 * 266 * Seems, it is the best solution to 267 * problem of too coarse exponent tabulation. 268 */ 269 us_idle = (v->qavg * (u64)us_idle) >> p->Scell_log; 270 271 if (us_idle < (v->qavg >> 1)) 272 return v->qavg - us_idle; 273 else 274 return v->qavg >> 1; 275 } 276 } 277 278 static inline unsigned long red_calc_qavg_no_idle_time(const struct red_parms *p, 279 const struct red_vars *v, 280 unsigned int backlog) 281 { 282 /* 283 * NOTE: v->qavg is fixed point number with point at Wlog. 284 * The formula below is equvalent to floating point 285 * version: 286 * 287 * qavg = qavg*(1-W) + backlog*W; 288 * 289 * --ANK (980924) 290 */ 291 return v->qavg + (backlog - (v->qavg >> p->Wlog)); 292 } 293 294 static inline unsigned long red_calc_qavg(const struct red_parms *p, 295 const struct red_vars *v, 296 unsigned int backlog) 297 { 298 if (!red_is_idling(v)) 299 return red_calc_qavg_no_idle_time(p, v, backlog); 300 else 301 return red_calc_qavg_from_idle_time(p, v); 302 } 303 304 305 static inline u32 red_random(const struct red_parms *p) 306 { 307 return reciprocal_divide(prandom_u32(), p->max_P_reciprocal); 308 } 309 310 static inline int red_mark_probability(const struct red_parms *p, 311 const struct red_vars *v, 312 unsigned long qavg) 313 { 314 /* The formula used below causes questions. 315 316 OK. qR is random number in the interval 317 (0..1/max_P)*(qth_max-qth_min) 318 i.e. 0..(2^Plog). If we used floating point 319 arithmetics, it would be: (2^Plog)*rnd_num, 320 where rnd_num is less 1. 321 322 Taking into account, that qavg have fixed 323 point at Wlog, two lines 324 below have the following floating point equivalent: 325 326 max_P*(qavg - qth_min)/(qth_max-qth_min) < rnd/qcount 327 328 Any questions? --ANK (980924) 329 */ 330 return !(((qavg - p->qth_min) >> p->Wlog) * v->qcount < v->qR); 331 } 332 333 enum { 334 RED_BELOW_MIN_THRESH, 335 RED_BETWEEN_TRESH, 336 RED_ABOVE_MAX_TRESH, 337 }; 338 339 static inline int red_cmp_thresh(const struct red_parms *p, unsigned long qavg) 340 { 341 if (qavg < p->qth_min) 342 return RED_BELOW_MIN_THRESH; 343 else if (qavg >= p->qth_max) 344 return RED_ABOVE_MAX_TRESH; 345 else 346 return RED_BETWEEN_TRESH; 347 } 348 349 enum { 350 RED_DONT_MARK, 351 RED_PROB_MARK, 352 RED_HARD_MARK, 353 }; 354 355 static inline int red_action(const struct red_parms *p, 356 struct red_vars *v, 357 unsigned long qavg) 358 { 359 switch (red_cmp_thresh(p, qavg)) { 360 case RED_BELOW_MIN_THRESH: 361 v->qcount = -1; 362 return RED_DONT_MARK; 363 364 case RED_BETWEEN_TRESH: 365 if (++v->qcount) { 366 if (red_mark_probability(p, v, qavg)) { 367 v->qcount = 0; 368 v->qR = red_random(p); 369 return RED_PROB_MARK; 370 } 371 } else 372 v->qR = red_random(p); 373 374 return RED_DONT_MARK; 375 376 case RED_ABOVE_MAX_TRESH: 377 v->qcount = -1; 378 return RED_HARD_MARK; 379 } 380 381 BUG(); 382 return RED_DONT_MARK; 383 } 384 385 static inline void red_adaptative_algo(struct red_parms *p, struct red_vars *v) 386 { 387 unsigned long qavg; 388 u32 max_p_delta; 389 390 qavg = v->qavg; 391 if (red_is_idling(v)) 392 qavg = red_calc_qavg_from_idle_time(p, v); 393 394 /* v->qavg is fixed point number with point at Wlog */ 395 qavg >>= p->Wlog; 396 397 if (qavg > p->target_max && p->max_P <= MAX_P_MAX) 398 p->max_P += MAX_P_ALPHA(p->max_P); /* maxp = maxp + alpha */ 399 else if (qavg < p->target_min && p->max_P >= MAX_P_MIN) 400 p->max_P = (p->max_P/10)*9; /* maxp = maxp * Beta */ 401 402 max_p_delta = DIV_ROUND_CLOSEST(p->max_P, p->qth_delta); 403 max_p_delta = max(max_p_delta, 1U); 404 p->max_P_reciprocal = reciprocal_value(max_p_delta); 405 } 406 #endif 407