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