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 u32 max_P_reciprocal; /* reciprocal_value(max_P / qth_delta) */ 134 u32 qth_delta; /* max_th - min_th */ 135 u32 target_min; /* min_th + 0.4*(max_th - min_th) */ 136 u32 target_max; /* min_th + 0.6*(max_th - min_th) */ 137 u8 Scell_log; 138 u8 Wlog; /* log(W) */ 139 u8 Plog; /* random number bits */ 140 u8 Stab[RED_STAB_SIZE]; 141 }; 142 143 struct red_vars { 144 /* Variables */ 145 int qcount; /* Number of packets since last random 146 number generation */ 147 u32 qR; /* Cached random number */ 148 149 unsigned long qavg; /* Average queue length: Wlog scaled */ 150 ktime_t qidlestart; /* Start of current idle period */ 151 }; 152 153 static inline u32 red_maxp(u8 Plog) 154 { 155 return Plog < 32 ? (~0U >> Plog) : ~0U; 156 } 157 158 static inline void red_set_vars(struct red_vars *v) 159 { 160 /* Reset average queue length, the value is strictly bound 161 * to the parameters below, reseting hurts a bit but leaving 162 * it might result in an unreasonable qavg for a while. --TGR 163 */ 164 v->qavg = 0; 165 166 v->qcount = -1; 167 } 168 169 static inline void red_set_parms(struct red_parms *p, 170 u32 qth_min, u32 qth_max, u8 Wlog, u8 Plog, 171 u8 Scell_log, u8 *stab, u32 max_P) 172 { 173 int delta = qth_max - qth_min; 174 u32 max_p_delta; 175 176 p->qth_min = qth_min << Wlog; 177 p->qth_max = qth_max << Wlog; 178 p->Wlog = Wlog; 179 p->Plog = Plog; 180 if (delta < 0) 181 delta = 1; 182 p->qth_delta = delta; 183 if (!max_P) { 184 max_P = red_maxp(Plog); 185 max_P *= delta; /* max_P = (qth_max - qth_min)/2^Plog */ 186 } 187 p->max_P = max_P; 188 max_p_delta = max_P / delta; 189 max_p_delta = max(max_p_delta, 1U); 190 p->max_P_reciprocal = reciprocal_value(max_p_delta); 191 192 /* RED Adaptative target : 193 * [min_th + 0.4*(min_th - max_th), 194 * min_th + 0.6*(min_th - max_th)]. 195 */ 196 delta /= 5; 197 p->target_min = qth_min + 2*delta; 198 p->target_max = qth_min + 3*delta; 199 200 p->Scell_log = Scell_log; 201 p->Scell_max = (255 << Scell_log); 202 203 if (stab) 204 memcpy(p->Stab, stab, sizeof(p->Stab)); 205 } 206 207 static inline int red_is_idling(const struct red_vars *v) 208 { 209 return v->qidlestart.tv64 != 0; 210 } 211 212 static inline void red_start_of_idle_period(struct red_vars *v) 213 { 214 v->qidlestart = ktime_get(); 215 } 216 217 static inline void red_end_of_idle_period(struct red_vars *v) 218 { 219 v->qidlestart.tv64 = 0; 220 } 221 222 static inline void red_restart(struct red_vars *v) 223 { 224 red_end_of_idle_period(v); 225 v->qavg = 0; 226 v->qcount = -1; 227 } 228 229 static inline unsigned long red_calc_qavg_from_idle_time(const struct red_parms *p, 230 const struct red_vars *v) 231 { 232 s64 delta = ktime_us_delta(ktime_get(), v->qidlestart); 233 long us_idle = min_t(s64, delta, p->Scell_max); 234 int shift; 235 236 /* 237 * The problem: ideally, average length queue recalcultion should 238 * be done over constant clock intervals. This is too expensive, so 239 * that the calculation is driven by outgoing packets. 240 * When the queue is idle we have to model this clock by hand. 241 * 242 * SF+VJ proposed to "generate": 243 * 244 * m = idletime / (average_pkt_size / bandwidth) 245 * 246 * dummy packets as a burst after idle time, i.e. 247 * 248 * v->qavg *= (1-W)^m 249 * 250 * This is an apparently overcomplicated solution (f.e. we have to 251 * precompute a table to make this calculation in reasonable time) 252 * I believe that a simpler model may be used here, 253 * but it is field for experiments. 254 */ 255 256 shift = p->Stab[(us_idle >> p->Scell_log) & RED_STAB_MASK]; 257 258 if (shift) 259 return v->qavg >> shift; 260 else { 261 /* Approximate initial part of exponent with linear function: 262 * 263 * (1-W)^m ~= 1-mW + ... 264 * 265 * Seems, it is the best solution to 266 * problem of too coarse exponent tabulation. 267 */ 268 us_idle = (v->qavg * (u64)us_idle) >> p->Scell_log; 269 270 if (us_idle < (v->qavg >> 1)) 271 return v->qavg - us_idle; 272 else 273 return v->qavg >> 1; 274 } 275 } 276 277 static inline unsigned long red_calc_qavg_no_idle_time(const struct red_parms *p, 278 const struct red_vars *v, 279 unsigned int backlog) 280 { 281 /* 282 * NOTE: v->qavg is fixed point number with point at Wlog. 283 * The formula below is equvalent to floating point 284 * version: 285 * 286 * qavg = qavg*(1-W) + backlog*W; 287 * 288 * --ANK (980924) 289 */ 290 return v->qavg + (backlog - (v->qavg >> p->Wlog)); 291 } 292 293 static inline unsigned long red_calc_qavg(const struct red_parms *p, 294 const struct red_vars *v, 295 unsigned int backlog) 296 { 297 if (!red_is_idling(v)) 298 return red_calc_qavg_no_idle_time(p, v, backlog); 299 else 300 return red_calc_qavg_from_idle_time(p, v); 301 } 302 303 304 static inline u32 red_random(const struct red_parms *p) 305 { 306 return reciprocal_divide(net_random(), p->max_P_reciprocal); 307 } 308 309 static inline int red_mark_probability(const struct red_parms *p, 310 const struct red_vars *v, 311 unsigned long qavg) 312 { 313 /* The formula used below causes questions. 314 315 OK. qR is random number in the interval 316 (0..1/max_P)*(qth_max-qth_min) 317 i.e. 0..(2^Plog). If we used floating point 318 arithmetics, it would be: (2^Plog)*rnd_num, 319 where rnd_num is less 1. 320 321 Taking into account, that qavg have fixed 322 point at Wlog, two lines 323 below have the following floating point equivalent: 324 325 max_P*(qavg - qth_min)/(qth_max-qth_min) < rnd/qcount 326 327 Any questions? --ANK (980924) 328 */ 329 return !(((qavg - p->qth_min) >> p->Wlog) * v->qcount < v->qR); 330 } 331 332 enum { 333 RED_BELOW_MIN_THRESH, 334 RED_BETWEEN_TRESH, 335 RED_ABOVE_MAX_TRESH, 336 }; 337 338 static inline int red_cmp_thresh(const struct red_parms *p, unsigned long qavg) 339 { 340 if (qavg < p->qth_min) 341 return RED_BELOW_MIN_THRESH; 342 else if (qavg >= p->qth_max) 343 return RED_ABOVE_MAX_TRESH; 344 else 345 return RED_BETWEEN_TRESH; 346 } 347 348 enum { 349 RED_DONT_MARK, 350 RED_PROB_MARK, 351 RED_HARD_MARK, 352 }; 353 354 static inline int red_action(const struct red_parms *p, 355 struct red_vars *v, 356 unsigned long qavg) 357 { 358 switch (red_cmp_thresh(p, qavg)) { 359 case RED_BELOW_MIN_THRESH: 360 v->qcount = -1; 361 return RED_DONT_MARK; 362 363 case RED_BETWEEN_TRESH: 364 if (++v->qcount) { 365 if (red_mark_probability(p, v, qavg)) { 366 v->qcount = 0; 367 v->qR = red_random(p); 368 return RED_PROB_MARK; 369 } 370 } else 371 v->qR = red_random(p); 372 373 return RED_DONT_MARK; 374 375 case RED_ABOVE_MAX_TRESH: 376 v->qcount = -1; 377 return RED_HARD_MARK; 378 } 379 380 BUG(); 381 return RED_DONT_MARK; 382 } 383 384 static inline void red_adaptative_algo(struct red_parms *p, struct red_vars *v) 385 { 386 unsigned long qavg; 387 u32 max_p_delta; 388 389 qavg = v->qavg; 390 if (red_is_idling(v)) 391 qavg = red_calc_qavg_from_idle_time(p, v); 392 393 /* v->qavg is fixed point number with point at Wlog */ 394 qavg >>= p->Wlog; 395 396 if (qavg > p->target_max && p->max_P <= MAX_P_MAX) 397 p->max_P += MAX_P_ALPHA(p->max_P); /* maxp = maxp + alpha */ 398 else if (qavg < p->target_min && p->max_P >= MAX_P_MIN) 399 p->max_P = (p->max_P/10)*9; /* maxp = maxp * Beta */ 400 401 max_p_delta = DIV_ROUND_CLOSEST(p->max_P, p->qth_delta); 402 max_p_delta = max(max_p_delta, 1U); 403 p->max_P_reciprocal = reciprocal_value(max_p_delta); 404 } 405 #endif 406