include/net/red.h

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