include/net/codel.h

   1 #ifndef __NET_SCHED_CODEL_H
   2 #define __NET_SCHED_CODEL_H
   3
   4 /*
   5  * Codel - The Controlled-Delay Active Queue Management algorithm
   6  *
   7  *  Copyright (C) 2011-2012 Kathleen Nichols <nichols@pollere.com>
   8  *  Copyright (C) 2011-2012 Van Jacobson <van@pollere.net>
   9  *  Copyright (C) 2012 Michael D. Taht <dave.taht@bufferbloat.net>
  10  *  Copyright (C) 2012 Eric Dumazet <edumazet@google.com>
  11  *
  12  * Redistribution and use in source and binary forms, with or without
  13  * modification, are permitted provided that the following conditions
  14  * are met:
  15  * 1. Redistributions of source code must retain the above copyright
  16  *    notice, this list of conditions, and the following disclaimer,
  17  *    without modification.
  18  * 2. Redistributions in binary form must reproduce the above copyright
  19  *    notice, this list of conditions and the following disclaimer in the
  20  *    documentation and/or other materials provided with the distribution.
  21  * 3. The names of the authors may not be used to endorse or promote products
  22  *    derived from this software without specific prior written permission.
  23  *
  24  * Alternatively, provided that this notice is retained in full, this
  25  * software may be distributed under the terms of the GNU General
  26  * Public License ("GPL") version 2, in which case the provisions of the
  27  * GPL apply INSTEAD OF those given above.
  28  *
  29  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  30  * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  31  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  32  * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  33  * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  34  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  35  * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  36  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  37  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  38  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  39  * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH
  40  * DAMAGE.
  41  *
  42  */
  43
  44 #include <linux/types.h>
  45 #include <linux/ktime.h>
  46 #include <linux/skbuff.h>
  47 #include <net/pkt_sched.h>
  48 #include <net/inet_ecn.h>
  49 #include <linux/reciprocal_div.h>
  50
  51 /* Controlling Queue Delay (CoDel) algorithm
  52  * =========================================
  53  * Source : Kathleen Nichols and Van Jacobson
  54  * http://queue.acm.org/detail.cfm?id=2209336
  55  *
  56  * Implemented on linux by Dave Taht and Eric Dumazet
  57  */
  58
  59
  60 /* CoDel uses a 1024 nsec clock, encoded in u32
  61  * This gives a range of 2199 seconds, because of signed compares
  62  */
  63 typedef u32 codel_time_t;
  64 typedef s32 codel_tdiff_t;
  65 #define CODEL_SHIFT 10
  66 #define MS2TIME(a) ((a * NSEC_PER_MSEC) >> CODEL_SHIFT)
  67
  68 static inline codel_time_t codel_get_time(void)
  69 {
  70         u64 ns = ktime_to_ns(ktime_get());
  71
  72         return ns >> CODEL_SHIFT;
  73 }
  74
  75 /* Dealing with timer wrapping, according to RFC 1982, as desc in wikipedia:
  76  *  https://en.wikipedia.org/wiki/Serial_number_arithmetic#General_Solution
  77  * codel_time_after(a,b) returns true if the time a is after time b.
  78  */
  79 #define codel_time_after(a, b)                                          \
  80         (typecheck(codel_time_t, a) &&                                  \
  81          typecheck(codel_time_t, b) &&                                  \
  82          ((s32)((a) - (b)) > 0))
  83 #define codel_time_before(a, b)         codel_time_after(b, a)
  84
  85 #define codel_time_after_eq(a, b)                                       \
  86         (typecheck(codel_time_t, a) &&                                  \
  87          typecheck(codel_time_t, b) &&                                  \
  88          ((s32)((a) - (b)) >= 0))
  89 #define codel_time_before_eq(a, b)      codel_time_after_eq(b, a)
  90
  91 /* Qdiscs using codel plugin must use codel_skb_cb in their own cb[] */
  92 struct codel_skb_cb {
  93         codel_time_t enqueue_time;
  94 };
  95
  96 static struct codel_skb_cb *get_codel_cb(const struct sk_buff *skb)
  97 {
  98         qdisc_cb_private_validate(skb, sizeof(struct codel_skb_cb));
  99         return (struct codel_skb_cb *)qdisc_skb_cb(skb)->data;
 100 }
 101
 102 static codel_time_t codel_get_enqueue_time(const struct sk_buff *skb)
 103 {
 104         return get_codel_cb(skb)->enqueue_time;
 105 }
 106
 107 static void codel_set_enqueue_time(struct sk_buff *skb)
 108 {
 109         get_codel_cb(skb)->enqueue_time = codel_get_time();
 110 }
 111
 112 static inline u32 codel_time_to_us(codel_time_t val)
 113 {
 114         u64 valns = ((u64)val << CODEL_SHIFT);
 115
 116         do_div(valns, NSEC_PER_USEC);
 117         return (u32)valns;
 118 }
 119
 120 /**
 121  * struct codel_params - contains codel parameters
 122  * @target:     target queue size (in time units)
 123  * @interval:   width of moving time window
 124  * @ecn:        is Explicit Congestion Notification enabled
 125  */
 126 struct codel_params {
 127         codel_time_t    target;
 128         codel_time_t    interval;
 129         bool            ecn;
 130 };
 131
 132 /**
 133  * struct codel_vars - contains codel variables
 134  * @count:              how many drops we've done since the last time we
 135  *                      entered dropping state
 136  * @lastcount:          count at entry to dropping state
 137  * @dropping:           set to true if in dropping state
 138  * @rec_inv_sqrt:       reciprocal value of sqrt(count) >> 1
 139  * @first_above_time:   when we went (or will go) continuously above target
 140  *                      for interval
 141  * @drop_next:          time to drop next packet, or when we dropped last
 142  * @ldelay:             sojourn time of last dequeued packet
 143  */
 144 struct codel_vars {
 145         u32             count;
 146         u32             lastcount;
 147         bool            dropping;
 148         u16             rec_inv_sqrt;
 149         codel_time_t    first_above_time;
 150         codel_time_t    drop_next;
 151         codel_time_t    ldelay;
 152 };
 153
 154 #define REC_INV_SQRT_BITS (8 * sizeof(u16)) /* or sizeof_in_bits(rec_inv_sqrt) */
 155 /* needed shift to get a Q0.32 number from rec_inv_sqrt */
 156 #define REC_INV_SQRT_SHIFT (32 - REC_INV_SQRT_BITS)
 157
 158 /**
 159  * struct codel_stats - contains codel shared variables and stats
 160  * @maxpacket:  largest packet we've seen so far
 161  * @drop_count: temp count of dropped packets in dequeue()
 162  * ecn_mark:    number of packets we ECN marked instead of dropping
 163  */
 164 struct codel_stats {
 165         u32             maxpacket;
 166         u32             drop_count;
 167         u32             ecn_mark;
 168 };
 169
 170 static void codel_params_init(struct codel_params *params)
 171 {
 172         params->interval = MS2TIME(100);
 173         params->target = MS2TIME(5);
 174         params->ecn = false;
 175 }
 176
 177 static void codel_vars_init(struct codel_vars *vars)
 178 {
 179         memset(vars, 0, sizeof(*vars));
 180 }
 181
 182 static void codel_stats_init(struct codel_stats *stats)
 183 {
 184         stats->maxpacket = 256;
 185 }
 186
 187 /*
 188  * http://en.wikipedia.org/wiki/Methods_of_computing_square_roots#Iterative_methods_for_reciprocal_square_roots
 189  * new_invsqrt = (invsqrt / 2) * (3 - count * invsqrt^2)
 190  *
 191  * Here, invsqrt is a fixed point number (< 1.0), 32bit mantissa, aka Q0.32
 192  */
 193 static void codel_Newton_step(struct codel_vars *vars)
 194 {
 195         u32 invsqrt = ((u32)vars->rec_inv_sqrt) << REC_INV_SQRT_SHIFT;
 196         u32 invsqrt2 = ((u64)invsqrt * invsqrt) >> 32;
 197         u64 val = (3LL << 32) - ((u64)vars->count * invsqrt2);
 198
 199         val >>= 2; /* avoid overflow in following multiply */
 200         val = (val * invsqrt) >> (32 - 2 + 1);
 201
 202         vars->rec_inv_sqrt = val >> REC_INV_SQRT_SHIFT;
 203 }
 204
 205 /*
 206  * CoDel control_law is t + interval/sqrt(count)
 207  * We maintain in rec_inv_sqrt the reciprocal value of sqrt(count) to avoid
 208  * both sqrt() and divide operation.
 209  */
 210 static codel_time_t codel_control_law(codel_time_t t,
 211                                       codel_time_t interval,
 212                                       u32 rec_inv_sqrt)
 213 {
 214         return t + reciprocal_divide(interval, rec_inv_sqrt << REC_INV_SQRT_SHIFT);
 215 }
 216
 217
 218 static bool codel_should_drop(const struct sk_buff *skb,
 219                               struct Qdisc *sch,
 220                               struct codel_vars *vars,
 221                               struct codel_params *params,
 222                               struct codel_stats *stats,
 223                               codel_time_t now)
 224 {
 225         bool ok_to_drop;
 226
 227         if (!skb) {
 228                 vars->first_above_time = 0;
 229                 return false;
 230         }
 231
 232         vars->ldelay = now - codel_get_enqueue_time(skb);
 233         sch->qstats.backlog -= qdisc_pkt_len(skb);
 234
 235         if (unlikely(qdisc_pkt_len(skb) > stats->maxpacket))
 236                 stats->maxpacket = qdisc_pkt_len(skb);
 237
 238         if (codel_time_before(vars->ldelay, params->target) ||
 239             sch->qstats.backlog <= stats->maxpacket) {
 240                 /* went below - stay below for at least interval */
 241                 vars->first_above_time = 0;
 242                 return false;
 243         }
 244         ok_to_drop = false;
 245         if (vars->first_above_time == 0) {
 246                 /* just went above from below. If we stay above
 247                  * for at least interval we'll say it's ok to drop
 248                  */
 249                 vars->first_above_time = now + params->interval;
 250         } else if (codel_time_after(now, vars->first_above_time)) {
 251                 ok_to_drop = true;
 252         }
 253         return ok_to_drop;
 254 }
 255
 256 typedef struct sk_buff * (*codel_skb_dequeue_t)(struct codel_vars *vars,
 257                                                 struct Qdisc *sch);
 258
 259 static struct sk_buff *codel_dequeue(struct Qdisc *sch,
 260                                      struct codel_params *params,
 261                                      struct codel_vars *vars,
 262                                      struct codel_stats *stats,
 263                                      codel_skb_dequeue_t dequeue_func)
 264 {
 265         struct sk_buff *skb = dequeue_func(vars, sch);
 266         codel_time_t now;
 267         bool drop;
 268
 269         if (!skb) {
 270                 vars->dropping = false;
 271                 return skb;
 272         }
 273         now = codel_get_time();
 274         drop = codel_should_drop(skb, sch, vars, params, stats, now);
 275         if (vars->dropping) {
 276                 if (!drop) {
 277                         /* sojourn time below target - leave dropping state */
 278                         vars->dropping = false;
 279                 } else if (codel_time_after_eq(now, vars->drop_next)) {
 280                         /* It's time for the next drop. Drop the current
 281                          * packet and dequeue the next. The dequeue might
 282                          * take us out of dropping state.
 283                          * If not, schedule the next drop.
 284                          * A large backlog might result in drop rates so high
 285                          * that the next drop should happen now,
 286                          * hence the while loop.
 287                          */
 288                         while (vars->dropping &&
 289                                codel_time_after_eq(now, vars->drop_next)) {
 290                                 vars->count++; /* dont care of possible wrap
 291                                                 * since there is no more divide
 292                                                 */
 293                                 codel_Newton_step(vars);
 294                                 if (params->ecn && INET_ECN_set_ce(skb)) {
 295                                         stats->ecn_mark++;
 296                                         vars->drop_next =
 297                                                 codel_control_law(vars->drop_next,
 298                                                                   params->interval,
 299                                                                   vars->rec_inv_sqrt);
 300                                         goto end;
 301                                 }
 302                                 qdisc_drop(skb, sch);
 303                                 stats->drop_count++;
 304                                 skb = dequeue_func(vars, sch);
 305                                 if (!codel_should_drop(skb, sch,
 306                                                        vars, params, stats, now)) {
 307                                         /* leave dropping state */
 308                                         vars->dropping = false;
 309                                 } else {
 310                                         /* and schedule the next drop */
 311                                         vars->drop_next =
 312                                                 codel_control_law(vars->drop_next,
 313                                                                   params->interval,
 314                                                                   vars->rec_inv_sqrt);
 315                                 }
 316                         }
 317                 }
 318         } else if (drop) {
 319                 u32 delta;
 320
 321                 if (params->ecn && INET_ECN_set_ce(skb)) {
 322                         stats->ecn_mark++;
 323                 } else {
 324                         qdisc_drop(skb, sch);
 325                         stats->drop_count++;
 326
 327                         skb = dequeue_func(vars, sch);
 328                         drop = codel_should_drop(skb, sch, vars, params,
 329                                                  stats, now);
 330                 }
 331                 vars->dropping = true;
 332                 /* if min went above target close to when we last went below it
 333                  * assume that the drop rate that controlled the queue on the
 334                  * last cycle is a good starting point to control it now.
 335                  */
 336                 delta = vars->count - vars->lastcount;
 337                 if (delta > 1 &&
 338                     codel_time_before(now - vars->drop_next,
 339                                       16 * params->interval)) {
 340                         vars->count = delta;
 341                         /* we dont care if rec_inv_sqrt approximation
 342                          * is not very precise :
 343                          * Next Newton steps will correct it quadratically.
 344                          */
 345                         codel_Newton_step(vars);
 346                 } else {
 347                         vars->count = 1;
 348                         vars->rec_inv_sqrt = ~0U >> REC_INV_SQRT_SHIFT;
 349                 }
 350                 vars->lastcount = vars->count;
 351                 vars->drop_next = codel_control_law(now, params->interval,
 352                                                     vars->rec_inv_sqrt);
 353         }
 354 end:
 355         return skb;
 356 }
 357 #endif