net/sched/sch_pie.c

   1 /* Copyright (C) 2013 Cisco Systems, Inc, 2013.
   2  *
   3  * This program is free software; you can redistribute it and/or
   4  * modify it under the terms of the GNU General Public License
   5  * as published by the Free Software Foundation; either version 2
   6  * of the License.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  11  * GNU General Public License for more details.
  12  *
  13  * Author: Vijay Subramanian <vijaynsu@cisco.com>
  14  * Author: Mythili Prabhu <mysuryan@cisco.com>
  15  *
  16  * ECN support is added by Naeem Khademi <naeemk@ifi.uio.no>
  17  * University of Oslo, Norway.
  18  */
  19
  20 #include <linux/module.h>
  21 #include <linux/slab.h>
  22 #include <linux/types.h>
  23 #include <linux/kernel.h>
  24 #include <linux/errno.h>
  25 #include <linux/skbuff.h>
  26 #include <net/pkt_sched.h>
  27 #include <net/inet_ecn.h>
  28
  29 #define QUEUE_THRESHOLD 10000
  30 #define DQCOUNT_INVALID -1
  31 #define MAX_PROB  0xffffffff
  32 #define PIE_SCALE 8
  33
  34 /* parameters used */
  35 struct pie_params {
  36         psched_time_t target;   /* user specified target delay in pschedtime */
  37         u32 tupdate;            /* timer frequency (in jiffies) */
  38         u32 limit;              /* number of packets that can be enqueued */
  39         u32 alpha;              /* alpha and beta are between -4 and 4 */
  40         u32 beta;               /* and are used for shift relative to 1 */
  41         bool ecn;               /* true if ecn is enabled */
  42         bool bytemode;          /* to scale drop early prob based on pkt size */
  43 };
  44
  45 /* variables used */
  46 struct pie_vars {
  47         u32 prob;               /* probability but scaled by u32 limit. */
  48         psched_time_t burst_time;
  49         psched_time_t qdelay;
  50         psched_time_t qdelay_old;
  51         u64 dq_count;           /* measured in bytes */
  52         psched_time_t dq_tstamp;        /* drain rate */
  53         u32 avg_dq_rate;        /* bytes per pschedtime tick,scaled */
  54         u32 qlen_old;           /* in bytes */
  55 };
  56
  57 /* statistics gathering */
  58 struct pie_stats {
  59         u32 packets_in;         /* total number of packets enqueued */
  60         u32 dropped;            /* packets dropped due to pie_action */
  61         u32 overlimit;          /* dropped due to lack of space in queue */
  62         u32 maxq;               /* maximum queue size */
  63         u32 ecn_mark;           /* packets marked with ECN */
  64 };
  65
  66 /* private data for the Qdisc */
  67 struct pie_sched_data {
  68         struct pie_params params;
  69         struct pie_vars vars;
  70         struct pie_stats stats;
  71         struct timer_list adapt_timer;
  72 };
  73
  74 static void pie_params_init(struct pie_params *params)
  75 {
  76         params->alpha = 2;
  77         params->beta = 20;
  78         params->tupdate = usecs_to_jiffies(30 * USEC_PER_MSEC); /* 30 ms */
  79         params->limit = 1000;   /* default of 1000 packets */
  80         params->target = PSCHED_NS2TICKS(20 * NSEC_PER_MSEC);   /* 20 ms */
  81         params->ecn = false;
  82         params->bytemode = false;
  83 }
  84
  85 static void pie_vars_init(struct pie_vars *vars)
  86 {
  87         vars->dq_count = DQCOUNT_INVALID;
  88         vars->avg_dq_rate = 0;
  89         /* default of 100 ms in pschedtime */
  90         vars->burst_time = PSCHED_NS2TICKS(100 * NSEC_PER_MSEC);
  91 }
  92
  93 static bool drop_early(struct Qdisc *sch, u32 packet_size)
  94 {
  95         struct pie_sched_data *q = qdisc_priv(sch);
  96         u32 rnd;
  97         u32 local_prob = q->vars.prob;
  98         u32 mtu = psched_mtu(qdisc_dev(sch));
  99
 100         /* If there is still burst allowance left skip random early drop */
 101         if (q->vars.burst_time > 0)
 102                 return false;
 103
 104         /* If current delay is less than half of target, and
 105          * if drop prob is low already, disable early_drop
 106          */
 107         if ((q->vars.qdelay < q->params.target / 2)
 108             && (q->vars.prob < MAX_PROB / 5))
 109                 return false;
 110
 111         /* If we have fewer than 2 mtu-sized packets, disable drop_early,
 112          * similar to min_th in RED
 113          */
 114         if (sch->qstats.backlog < 2 * mtu)
 115                 return false;
 116
 117         /* If bytemode is turned on, use packet size to compute new
 118          * probablity. Smaller packets will have lower drop prob in this case
 119          */
 120         if (q->params.bytemode && packet_size <= mtu)
 121                 local_prob = (local_prob / mtu) * packet_size;
 122         else
 123                 local_prob = q->vars.prob;
 124
 125         rnd = prandom_u32();
 126         if (rnd < local_prob)
 127                 return true;
 128
 129         return false;
 130 }
 131
 132 static int pie_qdisc_enqueue(struct sk_buff *skb, struct Qdisc *sch)
 133 {
 134         struct pie_sched_data *q = qdisc_priv(sch);
 135         bool enqueue = false;
 136
 137         if (unlikely(qdisc_qlen(sch) >= sch->limit)) {
 138                 q->stats.overlimit++;
 139                 goto out;
 140         }
 141
 142         if (!drop_early(sch, skb->len)) {
 143                 enqueue = true;
 144         } else if (q->params.ecn && (q->vars.prob <= MAX_PROB / 10) &&
 145                    INET_ECN_set_ce(skb)) {
 146                 /* If packet is ecn capable, mark it if drop probability
 147                  * is lower than 10%, else drop it.
 148                  */
 149                 q->stats.ecn_mark++;
 150                 enqueue = true;
 151         }
 152
 153         /* we can enqueue the packet */
 154         if (enqueue) {
 155                 q->stats.packets_in++;
 156                 if (qdisc_qlen(sch) > q->stats.maxq)
 157                         q->stats.maxq = qdisc_qlen(sch);
 158
 159                 return qdisc_enqueue_tail(skb, sch);
 160         }
 161
 162 out:
 163         q->stats.dropped++;
 164         return qdisc_drop(skb, sch);
 165 }
 166
 167 static const struct nla_policy pie_policy[TCA_PIE_MAX + 1] = {
 168         [TCA_PIE_TARGET] = {.type = NLA_U32},
 169         [TCA_PIE_LIMIT] = {.type = NLA_U32},
 170         [TCA_PIE_TUPDATE] = {.type = NLA_U32},
 171         [TCA_PIE_ALPHA] = {.type = NLA_U32},
 172         [TCA_PIE_BETA] = {.type = NLA_U32},
 173         [TCA_PIE_ECN] = {.type = NLA_U32},
 174         [TCA_PIE_BYTEMODE] = {.type = NLA_U32},
 175 };
 176
 177 static int pie_change(struct Qdisc *sch, struct nlattr *opt)
 178 {
 179         struct pie_sched_data *q = qdisc_priv(sch);
 180         struct nlattr *tb[TCA_PIE_MAX + 1];
 181         unsigned int qlen;
 182         int err;
 183
 184         if (!opt)
 185                 return -EINVAL;
 186
 187         err = nla_parse_nested(tb, TCA_PIE_MAX, opt, pie_policy);
 188         if (err < 0)
 189                 return err;
 190
 191         sch_tree_lock(sch);
 192
 193         /* convert from microseconds to pschedtime */
 194         if (tb[TCA_PIE_TARGET]) {
 195                 /* target is in us */
 196                 u32 target = nla_get_u32(tb[TCA_PIE_TARGET]);
 197
 198                 /* convert to pschedtime */
 199                 q->params.target = PSCHED_NS2TICKS((u64)target * NSEC_PER_USEC);
 200         }
 201
 202         /* tupdate is in jiffies */
 203         if (tb[TCA_PIE_TUPDATE])
 204                 q->params.tupdate = usecs_to_jiffies(nla_get_u32(tb[TCA_PIE_TUPDATE]));
 205
 206         if (tb[TCA_PIE_LIMIT]) {
 207                 u32 limit = nla_get_u32(tb[TCA_PIE_LIMIT]);
 208
 209                 q->params.limit = limit;
 210                 sch->limit = limit;
 211         }
 212
 213         if (tb[TCA_PIE_ALPHA])
 214                 q->params.alpha = nla_get_u32(tb[TCA_PIE_ALPHA]);
 215
 216         if (tb[TCA_PIE_BETA])
 217                 q->params.beta = nla_get_u32(tb[TCA_PIE_BETA]);
 218
 219         if (tb[TCA_PIE_ECN])
 220                 q->params.ecn = nla_get_u32(tb[TCA_PIE_ECN]);
 221
 222         if (tb[TCA_PIE_BYTEMODE])
 223                 q->params.bytemode = nla_get_u32(tb[TCA_PIE_BYTEMODE]);
 224
 225         /* Drop excess packets if new limit is lower */
 226         qlen = sch->q.qlen;
 227         while (sch->q.qlen > sch->limit) {
 228                 struct sk_buff *skb = __skb_dequeue(&sch->q);
 229
 230                 sch->qstats.backlog -= qdisc_pkt_len(skb);
 231                 qdisc_drop(skb, sch);
 232         }
 233         qdisc_tree_decrease_qlen(sch, qlen - sch->q.qlen);
 234
 235         sch_tree_unlock(sch);
 236         return 0;
 237 }
 238
 239 static void pie_process_dequeue(struct Qdisc *sch, struct sk_buff *skb)
 240 {
 241
 242         struct pie_sched_data *q = qdisc_priv(sch);
 243         int qlen = sch->qstats.backlog; /* current queue size in bytes */
 244
 245         /* If current queue is about 10 packets or more and dq_count is unset
 246          * we have enough packets to calculate the drain rate. Save
 247          * current time as dq_tstamp and start measurement cycle.
 248          */
 249         if (qlen >= QUEUE_THRESHOLD && q->vars.dq_count == DQCOUNT_INVALID) {
 250                 q->vars.dq_tstamp = psched_get_time();
 251                 q->vars.dq_count = 0;
 252         }
 253
 254         /* Calculate the average drain rate from this value.  If queue length
 255          * has receded to a small value viz., <= QUEUE_THRESHOLD bytes,reset
 256          * the dq_count to -1 as we don't have enough packets to calculate the
 257          * drain rate anymore The following if block is entered only when we
 258          * have a substantial queue built up (QUEUE_THRESHOLD bytes or more)
 259          * and we calculate the drain rate for the threshold here.  dq_count is
 260          * in bytes, time difference in psched_time, hence rate is in
 261          * bytes/psched_time.
 262          */
 263         if (q->vars.dq_count != DQCOUNT_INVALID) {
 264                 q->vars.dq_count += skb->len;
 265
 266                 if (q->vars.dq_count >= QUEUE_THRESHOLD) {
 267                         psched_time_t now = psched_get_time();
 268                         u32 dtime = now - q->vars.dq_tstamp;
 269                         u32 count = q->vars.dq_count << PIE_SCALE;
 270
 271                         if (dtime == 0)
 272                                 return;
 273
 274                         count = count / dtime;
 275
 276                         if (q->vars.avg_dq_rate == 0)
 277                                 q->vars.avg_dq_rate = count;
 278                         else
 279                                 q->vars.avg_dq_rate =
 280                                     (q->vars.avg_dq_rate -
 281                                      (q->vars.avg_dq_rate >> 3)) + (count >> 3);
 282
 283                         /* If the queue has receded below the threshold, we hold
 284                          * on to the last drain rate calculated, else we reset
 285                          * dq_count to 0 to re-enter the if block when the next
 286                          * packet is dequeued
 287                          */
 288                         if (qlen < QUEUE_THRESHOLD)
 289                                 q->vars.dq_count = DQCOUNT_INVALID;
 290                         else {
 291                                 q->vars.dq_count = 0;
 292                                 q->vars.dq_tstamp = psched_get_time();
 293                         }
 294
 295                         if (q->vars.burst_time > 0) {
 296                                 if (q->vars.burst_time > dtime)
 297                                         q->vars.burst_time -= dtime;
 298                                 else
 299                                         q->vars.burst_time = 0;
 300                         }
 301                 }
 302         }
 303 }
 304
 305 static void calculate_probability(struct Qdisc *sch)
 306 {
 307         struct pie_sched_data *q = qdisc_priv(sch);
 308         u32 qlen = sch->qstats.backlog; /* queue size in bytes */
 309         psched_time_t qdelay = 0;       /* in pschedtime */
 310         psched_time_t qdelay_old = q->vars.qdelay;      /* in pschedtime */
 311         s32 delta = 0;          /* determines the change in probability */
 312         u32 oldprob;
 313         u32 alpha, beta;
 314         bool update_prob = true;
 315
 316         q->vars.qdelay_old = q->vars.qdelay;
 317
 318         if (q->vars.avg_dq_rate > 0)
 319                 qdelay = (qlen << PIE_SCALE) / q->vars.avg_dq_rate;
 320         else
 321                 qdelay = 0;
 322
 323         /* If qdelay is zero and qlen is not, it means qlen is very small, less
 324          * than dequeue_rate, so we do not update probabilty in this round
 325          */
 326         if (qdelay == 0 && qlen != 0)
 327                 update_prob = false;
 328
 329         /* Add ranges for alpha and beta, more aggressive for high dropping
 330          * mode and gentle steps for light dropping mode
 331          * In light dropping mode, take gentle steps; in medium dropping mode,
 332          * take medium steps; in high dropping mode, take big steps.
 333          */
 334         if (q->vars.prob < MAX_PROB / 100) {
 335                 alpha =
 336                     (q->params.alpha * (MAX_PROB / PSCHED_TICKS_PER_SEC)) >> 7;
 337                 beta =
 338                     (q->params.beta * (MAX_PROB / PSCHED_TICKS_PER_SEC)) >> 7;
 339         } else if (q->vars.prob < MAX_PROB / 10) {
 340                 alpha =
 341                     (q->params.alpha * (MAX_PROB / PSCHED_TICKS_PER_SEC)) >> 5;
 342                 beta =
 343                     (q->params.beta * (MAX_PROB / PSCHED_TICKS_PER_SEC)) >> 5;
 344         } else {
 345                 alpha =
 346                     (q->params.alpha * (MAX_PROB / PSCHED_TICKS_PER_SEC)) >> 4;
 347                 beta =
 348                     (q->params.beta * (MAX_PROB / PSCHED_TICKS_PER_SEC)) >> 4;
 349         }
 350
 351         /* alpha and beta should be between 0 and 32, in multiples of 1/16 */
 352         delta += alpha * ((qdelay - q->params.target));
 353         delta += beta * ((qdelay - qdelay_old));
 354
 355         oldprob = q->vars.prob;
 356
 357         /* to ensure we increase probability in steps of no more than 2% */
 358         if (delta > (s32) (MAX_PROB / (100 / 2)) &&
 359             q->vars.prob >= MAX_PROB / 10)
 360                 delta = (MAX_PROB / 100) * 2;
 361
 362         /* Non-linear drop:
 363          * Tune drop probability to increase quickly for high delays(>= 250ms)
 364          * 250ms is derived through experiments and provides error protection
 365          */
 366
 367         if (qdelay > (PSCHED_NS2TICKS(250 * NSEC_PER_MSEC)))
 368                 delta += MAX_PROB / (100 / 2);
 369
 370         q->vars.prob += delta;
 371
 372         if (delta > 0) {
 373                 /* prevent overflow */
 374                 if (q->vars.prob < oldprob) {
 375                         q->vars.prob = MAX_PROB;
 376                         /* Prevent normalization error. If probability is at
 377                          * maximum value already, we normalize it here, and
 378                          * skip the check to do a non-linear drop in the next
 379                          * section.
 380                          */
 381                         update_prob = false;
 382                 }
 383         } else {
 384                 /* prevent underflow */
 385                 if (q->vars.prob > oldprob)
 386                         q->vars.prob = 0;
 387         }
 388
 389         /* Non-linear drop in probability: Reduce drop probability quickly if
 390          * delay is 0 for 2 consecutive Tupdate periods.
 391          */
 392
 393         if ((qdelay == 0) && (qdelay_old == 0) && update_prob)
 394                 q->vars.prob = (q->vars.prob * 98) / 100;
 395
 396         q->vars.qdelay = qdelay;
 397         q->vars.qlen_old = qlen;
 398
 399         /* We restart the measurement cycle if the following conditions are met
 400          * 1. If the delay has been low for 2 consecutive Tupdate periods
 401          * 2. Calculated drop probability is zero
 402          * 3. We have atleast one estimate for the avg_dq_rate ie.,
 403          *    is a non-zero value
 404          */
 405         if ((q->vars.qdelay < q->params.target / 2) &&
 406             (q->vars.qdelay_old < q->params.target / 2) &&
 407             (q->vars.prob == 0) &&
 408             (q->vars.avg_dq_rate > 0))
 409                 pie_vars_init(&q->vars);
 410 }
 411
 412 static void pie_timer(unsigned long arg)
 413 {
 414         struct Qdisc *sch = (struct Qdisc *)arg;
 415         struct pie_sched_data *q = qdisc_priv(sch);
 416         spinlock_t *root_lock = qdisc_lock(qdisc_root_sleeping(sch));
 417
 418         spin_lock(root_lock);
 419         calculate_probability(sch);
 420
 421         /* reset the timer to fire after 'tupdate'. tupdate is in jiffies. */
 422         if (q->params.tupdate)
 423                 mod_timer(&q->adapt_timer, jiffies + q->params.tupdate);
 424         spin_unlock(root_lock);
 425
 426 }
 427
 428 static int pie_init(struct Qdisc *sch, struct nlattr *opt)
 429 {
 430         struct pie_sched_data *q = qdisc_priv(sch);
 431
 432         pie_params_init(&q->params);
 433         pie_vars_init(&q->vars);
 434         sch->limit = q->params.limit;
 435
 436         setup_timer(&q->adapt_timer, pie_timer, (unsigned long)sch);
 437         mod_timer(&q->adapt_timer, jiffies + HZ / 2);
 438
 439         if (opt) {
 440                 int err = pie_change(sch, opt);
 441
 442                 if (err)
 443                         return err;
 444         }
 445
 446         return 0;
 447 }
 448
 449 static int pie_dump(struct Qdisc *sch, struct sk_buff *skb)
 450 {
 451         struct pie_sched_data *q = qdisc_priv(sch);
 452         struct nlattr *opts;
 453
 454         opts = nla_nest_start(skb, TCA_OPTIONS);
 455         if (opts == NULL)
 456                 goto nla_put_failure;
 457
 458         /* convert target from pschedtime to us */
 459         if (nla_put_u32(skb, TCA_PIE_TARGET,
 460                         ((u32) PSCHED_TICKS2NS(q->params.target)) /
 461                         NSEC_PER_USEC) ||
 462             nla_put_u32(skb, TCA_PIE_LIMIT, sch->limit) ||
 463             nla_put_u32(skb, TCA_PIE_TUPDATE, jiffies_to_usecs(q->params.tupdate)) ||
 464             nla_put_u32(skb, TCA_PIE_ALPHA, q->params.alpha) ||
 465             nla_put_u32(skb, TCA_PIE_BETA, q->params.beta) ||
 466             nla_put_u32(skb, TCA_PIE_ECN, q->params.ecn) ||
 467             nla_put_u32(skb, TCA_PIE_BYTEMODE, q->params.bytemode))
 468                 goto nla_put_failure;
 469
 470         return nla_nest_end(skb, opts);
 471
 472 nla_put_failure:
 473         nla_nest_cancel(skb, opts);
 474         return -1;
 475
 476 }
 477
 478 static int pie_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
 479 {
 480         struct pie_sched_data *q = qdisc_priv(sch);
 481         struct tc_pie_xstats st = {
 482                 .prob           = q->vars.prob,
 483                 .delay          = ((u32) PSCHED_TICKS2NS(q->vars.qdelay)) /
 484                                    NSEC_PER_USEC,
 485                 /* unscale and return dq_rate in bytes per sec */
 486                 .avg_dq_rate    = q->vars.avg_dq_rate *
 487                                   (PSCHED_TICKS_PER_SEC) >> PIE_SCALE,
 488                 .packets_in     = q->stats.packets_in,
 489                 .overlimit      = q->stats.overlimit,
 490                 .maxq           = q->stats.maxq,
 491                 .dropped        = q->stats.dropped,
 492                 .ecn_mark       = q->stats.ecn_mark,
 493         };
 494
 495         return gnet_stats_copy_app(d, &st, sizeof(st));
 496 }
 497
 498 static struct sk_buff *pie_qdisc_dequeue(struct Qdisc *sch)
 499 {
 500         struct sk_buff *skb;
 501         skb = __qdisc_dequeue_head(sch, &sch->q);
 502
 503         if (!skb)
 504                 return NULL;
 505
 506         pie_process_dequeue(sch, skb);
 507         return skb;
 508 }
 509
 510 static void pie_reset(struct Qdisc *sch)
 511 {
 512         struct pie_sched_data *q = qdisc_priv(sch);
 513         qdisc_reset_queue(sch);
 514         pie_vars_init(&q->vars);
 515 }
 516
 517 static void pie_destroy(struct Qdisc *sch)
 518 {
 519         struct pie_sched_data *q = qdisc_priv(sch);
 520         q->params.tupdate = 0;
 521         del_timer_sync(&q->adapt_timer);
 522 }
 523
 524 static struct Qdisc_ops pie_qdisc_ops __read_mostly = {
 525         .id = "pie",
 526         .priv_size      = sizeof(struct pie_sched_data),
 527         .enqueue        = pie_qdisc_enqueue,
 528         .dequeue        = pie_qdisc_dequeue,
 529         .peek           = qdisc_peek_dequeued,
 530         .init           = pie_init,
 531         .destroy        = pie_destroy,
 532         .reset          = pie_reset,
 533         .change         = pie_change,
 534         .dump           = pie_dump,
 535         .dump_stats     = pie_dump_stats,
 536         .owner          = THIS_MODULE,
 537 };
 538
 539 static int __init pie_module_init(void)
 540 {
 541         return register_qdisc(&pie_qdisc_ops);
 542 }
 543
 544 static void __exit pie_module_exit(void)
 545 {
 546         unregister_qdisc(&pie_qdisc_ops);
 547 }
 548
 549 module_init(pie_module_init);
 550 module_exit(pie_module_exit);
 551
 552 MODULE_DESCRIPTION("Proportional Integral controller Enhanced (PIE) scheduler");
 553 MODULE_AUTHOR("Vijay Subramanian");
 554 MODULE_AUTHOR("Mythili Prabhu");
 555 MODULE_LICENSE("GPL");