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, u8 Scell_log) 172 { 173 if (fls(qth_min) + Wlog > 32) 174 return false; 175 if (fls(qth_max) + Wlog > 32) 176 return false; 177 if (Scell_log >= 32) 178 return false; 179 if (qth_max < qth_min) 180 return false; 181 return true; 182 } 183 184 static inline int red_get_flags(unsigned char qopt_flags, 185 unsigned char historic_mask, 186 struct nlattr *flags_attr, 187 unsigned char supported_mask, 188 struct nla_bitfield32 *p_flags, 189 unsigned char *p_userbits, 190 struct netlink_ext_ack *extack) 191 { 192 struct nla_bitfield32 flags; 193 194 if (qopt_flags && flags_attr) { 195 NL_SET_ERR_MSG_MOD(extack, "flags should be passed either through qopt, or through a dedicated attribute"); 196 return -EINVAL; 197 } 198 199 if (flags_attr) { 200 flags = nla_get_bitfield32(flags_attr); 201 } else { 202 flags.selector = historic_mask; 203 flags.value = qopt_flags & historic_mask; 204 } 205 206 *p_flags = flags; 207 *p_userbits = qopt_flags & ~historic_mask; 208 return 0; 209 } 210 211 static inline int red_validate_flags(unsigned char flags, 212 struct netlink_ext_ack *extack) 213 { 214 if ((flags & TC_RED_NODROP) && !(flags & TC_RED_ECN)) { 215 NL_SET_ERR_MSG_MOD(extack, "nodrop mode is only meaningful with ECN"); 216 return -EINVAL; 217 } 218 219 return 0; 220 } 221 222 static inline void red_set_parms(struct red_parms *p, 223 u32 qth_min, u32 qth_max, u8 Wlog, u8 Plog, 224 u8 Scell_log, u8 *stab, u32 max_P) 225 { 226 int delta = qth_max - qth_min; 227 u32 max_p_delta; 228 229 p->qth_min = qth_min << Wlog; 230 p->qth_max = qth_max << Wlog; 231 p->Wlog = Wlog; 232 p->Plog = Plog; 233 if (delta <= 0) 234 delta = 1; 235 p->qth_delta = delta; 236 if (!max_P) { 237 max_P = red_maxp(Plog); 238 max_P *= delta; /* max_P = (qth_max - qth_min)/2^Plog */ 239 } 240 p->max_P = max_P; 241 max_p_delta = max_P / delta; 242 max_p_delta = max(max_p_delta, 1U); 243 p->max_P_reciprocal = reciprocal_value(max_p_delta); 244 245 /* RED Adaptative target : 246 * [min_th + 0.4*(min_th - max_th), 247 * min_th + 0.6*(min_th - max_th)]. 248 */ 249 delta /= 5; 250 p->target_min = qth_min + 2*delta; 251 p->target_max = qth_min + 3*delta; 252 253 p->Scell_log = Scell_log; 254 p->Scell_max = (255 << Scell_log); 255 256 if (stab) 257 memcpy(p->Stab, stab, sizeof(p->Stab)); 258 } 259 260 static inline int red_is_idling(const struct red_vars *v) 261 { 262 return v->qidlestart != 0; 263 } 264 265 static inline void red_start_of_idle_period(struct red_vars *v) 266 { 267 v->qidlestart = ktime_get(); 268 } 269 270 static inline void red_end_of_idle_period(struct red_vars *v) 271 { 272 v->qidlestart = 0; 273 } 274 275 static inline void red_restart(struct red_vars *v) 276 { 277 red_end_of_idle_period(v); 278 v->qavg = 0; 279 v->qcount = -1; 280 } 281 282 static inline unsigned long red_calc_qavg_from_idle_time(const struct red_parms *p, 283 const struct red_vars *v) 284 { 285 s64 delta = ktime_us_delta(ktime_get(), v->qidlestart); 286 long us_idle = min_t(s64, delta, p->Scell_max); 287 int shift; 288 289 /* 290 * The problem: ideally, average length queue recalcultion should 291 * be done over constant clock intervals. This is too expensive, so 292 * that the calculation is driven by outgoing packets. 293 * When the queue is idle we have to model this clock by hand. 294 * 295 * SF+VJ proposed to "generate": 296 * 297 * m = idletime / (average_pkt_size / bandwidth) 298 * 299 * dummy packets as a burst after idle time, i.e. 300 * 301 * v->qavg *= (1-W)^m 302 * 303 * This is an apparently overcomplicated solution (f.e. we have to 304 * precompute a table to make this calculation in reasonable time) 305 * I believe that a simpler model may be used here, 306 * but it is field for experiments. 307 */ 308 309 shift = p->Stab[(us_idle >> p->Scell_log) & RED_STAB_MASK]; 310 311 if (shift) 312 return v->qavg >> shift; 313 else { 314 /* Approximate initial part of exponent with linear function: 315 * 316 * (1-W)^m ~= 1-mW + ... 317 * 318 * Seems, it is the best solution to 319 * problem of too coarse exponent tabulation. 320 */ 321 us_idle = (v->qavg * (u64)us_idle) >> p->Scell_log; 322 323 if (us_idle < (v->qavg >> 1)) 324 return v->qavg - us_idle; 325 else 326 return v->qavg >> 1; 327 } 328 } 329 330 static inline unsigned long red_calc_qavg_no_idle_time(const struct red_parms *p, 331 const struct red_vars *v, 332 unsigned int backlog) 333 { 334 /* 335 * NOTE: v->qavg is fixed point number with point at Wlog. 336 * The formula below is equvalent to floating point 337 * version: 338 * 339 * qavg = qavg*(1-W) + backlog*W; 340 * 341 * --ANK (980924) 342 */ 343 return v->qavg + (backlog - (v->qavg >> p->Wlog)); 344 } 345 346 static inline unsigned long red_calc_qavg(const struct red_parms *p, 347 const struct red_vars *v, 348 unsigned int backlog) 349 { 350 if (!red_is_idling(v)) 351 return red_calc_qavg_no_idle_time(p, v, backlog); 352 else 353 return red_calc_qavg_from_idle_time(p, v); 354 } 355 356 357 static inline u32 red_random(const struct red_parms *p) 358 { 359 return reciprocal_divide(prandom_u32(), p->max_P_reciprocal); 360 } 361 362 static inline int red_mark_probability(const struct red_parms *p, 363 const struct red_vars *v, 364 unsigned long qavg) 365 { 366 /* The formula used below causes questions. 367 368 OK. qR is random number in the interval 369 (0..1/max_P)*(qth_max-qth_min) 370 i.e. 0..(2^Plog). If we used floating point 371 arithmetics, it would be: (2^Plog)*rnd_num, 372 where rnd_num is less 1. 373 374 Taking into account, that qavg have fixed 375 point at Wlog, two lines 376 below have the following floating point equivalent: 377 378 max_P*(qavg - qth_min)/(qth_max-qth_min) < rnd/qcount 379 380 Any questions? --ANK (980924) 381 */ 382 return !(((qavg - p->qth_min) >> p->Wlog) * v->qcount < v->qR); 383 } 384 385 enum { 386 RED_BELOW_MIN_THRESH, 387 RED_BETWEEN_TRESH, 388 RED_ABOVE_MAX_TRESH, 389 }; 390 391 static inline int red_cmp_thresh(const struct red_parms *p, unsigned long qavg) 392 { 393 if (qavg < p->qth_min) 394 return RED_BELOW_MIN_THRESH; 395 else if (qavg >= p->qth_max) 396 return RED_ABOVE_MAX_TRESH; 397 else 398 return RED_BETWEEN_TRESH; 399 } 400 401 enum { 402 RED_DONT_MARK, 403 RED_PROB_MARK, 404 RED_HARD_MARK, 405 }; 406 407 static inline int red_action(const struct red_parms *p, 408 struct red_vars *v, 409 unsigned long qavg) 410 { 411 switch (red_cmp_thresh(p, qavg)) { 412 case RED_BELOW_MIN_THRESH: 413 v->qcount = -1; 414 return RED_DONT_MARK; 415 416 case RED_BETWEEN_TRESH: 417 if (++v->qcount) { 418 if (red_mark_probability(p, v, qavg)) { 419 v->qcount = 0; 420 v->qR = red_random(p); 421 return RED_PROB_MARK; 422 } 423 } else 424 v->qR = red_random(p); 425 426 return RED_DONT_MARK; 427 428 case RED_ABOVE_MAX_TRESH: 429 v->qcount = -1; 430 return RED_HARD_MARK; 431 } 432 433 BUG(); 434 return RED_DONT_MARK; 435 } 436 437 static inline void red_adaptative_algo(struct red_parms *p, struct red_vars *v) 438 { 439 unsigned long qavg; 440 u32 max_p_delta; 441 442 qavg = v->qavg; 443 if (red_is_idling(v)) 444 qavg = red_calc_qavg_from_idle_time(p, v); 445 446 /* v->qavg is fixed point number with point at Wlog */ 447 qavg >>= p->Wlog; 448 449 if (qavg > p->target_max && p->max_P <= MAX_P_MAX) 450 p->max_P += MAX_P_ALPHA(p->max_P); /* maxp = maxp + alpha */ 451 else if (qavg < p->target_min && p->max_P >= MAX_P_MIN) 452 p->max_P = (p->max_P/10)*9; /* maxp = maxp * Beta */ 453 454 max_p_delta = DIV_ROUND_CLOSEST(p->max_P, p->qth_delta); 455 max_p_delta = max(max_p_delta, 1U); 456 p->max_P_reciprocal = reciprocal_value(max_p_delta); 457 } 458 #endif 459