net/sched/sch_fq_pie.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /* Flow Queue PIE discipline
   3  *
   4  * Copyright (C) 2019 Mohit P. Tahiliani <tahiliani@nitk.edu.in>
   5  * Copyright (C) 2019 Sachin D. Patil <sdp.sachin@gmail.com>
   6  * Copyright (C) 2019 V. Saicharan <vsaicharan1998@gmail.com>
   7  * Copyright (C) 2019 Mohit Bhasi <mohitbhasi1998@gmail.com>
   8  * Copyright (C) 2019 Leslie Monis <lesliemonis@gmail.com>
   9  * Copyright (C) 2019 Gautam Ramakrishnan <gautamramk@gmail.com>
  10  */
  11
  12 #include <linux/jhash.h>
  13 #include <linux/sizes.h>
  14 #include <linux/vmalloc.h>
  15 #include <net/pkt_cls.h>
  16 #include <net/pie.h>
  17
  18 /* Flow Queue PIE
  19  *
  20  * Principles:
  21  *   - Packets are classified on flows.
  22  *   - This is a Stochastic model (as we use a hash, several flows might
  23  *                                 be hashed to the same slot)
  24  *   - Each flow has a PIE managed queue.
  25  *   - Flows are linked onto two (Round Robin) lists,
  26  *     so that new flows have priority on old ones.
  27  *   - For a given flow, packets are not reordered.
  28  *   - Drops during enqueue only.
  29  *   - ECN capability is off by default.
  30  *   - ECN threshold (if ECN is enabled) is at 10% by default.
  31  *   - Uses timestamps to calculate queue delay by default.
  32  */
  33
  34 /**
  35  * struct fq_pie_flow - contains data for each flow
  36  * @vars:       pie vars associated with the flow
  37  * @deficit:    number of remaining byte credits
  38  * @backlog:    size of data in the flow
  39  * @qlen:       number of packets in the flow
  40  * @flowchain:  flowchain for the flow
  41  * @head:       first packet in the flow
  42  * @tail:       last packet in the flow
  43  */
  44 struct fq_pie_flow {
  45         struct pie_vars vars;
  46         s32 deficit;
  47         u32 backlog;
  48         u32 qlen;
  49         struct list_head flowchain;
  50         struct sk_buff *head;
  51         struct sk_buff *tail;
  52 };
  53
  54 struct fq_pie_sched_data {
  55         struct tcf_proto __rcu *filter_list; /* optional external classifier */
  56         struct tcf_block *block;
  57         struct fq_pie_flow *flows;
  58         struct Qdisc *sch;
  59         struct list_head old_flows;
  60         struct list_head new_flows;
  61         struct pie_params p_params;
  62         u32 ecn_prob;
  63         u32 flows_cnt;
  64         u32 quantum;
  65         u32 memory_limit;
  66         u32 new_flow_count;
  67         u32 memory_usage;
  68         u32 overmemory;
  69         struct pie_stats stats;
  70         struct timer_list adapt_timer;
  71 };
  72
  73 static unsigned int fq_pie_hash(const struct fq_pie_sched_data *q,
  74                                 struct sk_buff *skb)
  75 {
  76         return reciprocal_scale(skb_get_hash(skb), q->flows_cnt);
  77 }
  78
  79 static unsigned int fq_pie_classify(struct sk_buff *skb, struct Qdisc *sch,
  80                                     int *qerr)
  81 {
  82         struct fq_pie_sched_data *q = qdisc_priv(sch);
  83         struct tcf_proto *filter;
  84         struct tcf_result res;
  85         int result;
  86
  87         if (TC_H_MAJ(skb->priority) == sch->handle &&
  88             TC_H_MIN(skb->priority) > 0 &&
  89             TC_H_MIN(skb->priority) <= q->flows_cnt)
  90                 return TC_H_MIN(skb->priority);
  91
  92         filter = rcu_dereference_bh(q->filter_list);
  93         if (!filter)
  94                 return fq_pie_hash(q, skb) + 1;
  95
  96         *qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
  97         result = tcf_classify(skb, NULL, filter, &res, false);
  98         if (result >= 0) {
  99 #ifdef CONFIG_NET_CLS_ACT
 100                 switch (result) {
 101                 case TC_ACT_STOLEN:
 102                 case TC_ACT_QUEUED:
 103                 case TC_ACT_TRAP:
 104                         *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
 105                         fallthrough;
 106                 case TC_ACT_SHOT:
 107                         return 0;
 108                 }
 109 #endif
 110                 if (TC_H_MIN(res.classid) <= q->flows_cnt)
 111                         return TC_H_MIN(res.classid);
 112         }
 113         return 0;
 114 }
 115
 116 /* add skb to flow queue (tail add) */
 117 static inline void flow_queue_add(struct fq_pie_flow *flow,
 118                                   struct sk_buff *skb)
 119 {
 120         if (!flow->head)
 121                 flow->head = skb;
 122         else
 123                 flow->tail->next = skb;
 124         flow->tail = skb;
 125         skb->next = NULL;
 126 }
 127
 128 static int fq_pie_qdisc_enqueue(struct sk_buff *skb, struct Qdisc *sch,
 129                                 struct sk_buff **to_free)
 130 {
 131         struct fq_pie_sched_data *q = qdisc_priv(sch);
 132         struct fq_pie_flow *sel_flow;
 133         int ret;
 134         u8 memory_limited = false;
 135         u8 enqueue = false;
 136         u32 pkt_len;
 137         u32 idx;
 138
 139         /* Classifies packet into corresponding flow */
 140         idx = fq_pie_classify(skb, sch, &ret);
 141         if (idx == 0) {
 142                 if (ret & __NET_XMIT_BYPASS)
 143                         qdisc_qstats_drop(sch);
 144                 __qdisc_drop(skb, to_free);
 145                 return ret;
 146         }
 147         idx--;
 148
 149         sel_flow = &q->flows[idx];
 150         /* Checks whether adding a new packet would exceed memory limit */
 151         get_pie_cb(skb)->mem_usage = skb->truesize;
 152         memory_limited = q->memory_usage > q->memory_limit + skb->truesize;
 153
 154         /* Checks if the qdisc is full */
 155         if (unlikely(qdisc_qlen(sch) >= sch->limit)) {
 156                 q->stats.overlimit++;
 157                 goto out;
 158         } else if (unlikely(memory_limited)) {
 159                 q->overmemory++;
 160         }
 161
 162         if (!pie_drop_early(sch, &q->p_params, &sel_flow->vars,
 163                             sel_flow->backlog, skb->len)) {
 164                 enqueue = true;
 165         } else if (q->p_params.ecn &&
 166                    sel_flow->vars.prob <= (MAX_PROB / 100) * q->ecn_prob &&
 167                    INET_ECN_set_ce(skb)) {
 168                 /* If packet is ecn capable, mark it if drop probability
 169                  * is lower than the parameter ecn_prob, else drop it.
 170                  */
 171                 q->stats.ecn_mark++;
 172                 enqueue = true;
 173         }
 174         if (enqueue) {
 175                 /* Set enqueue time only when dq_rate_estimator is disabled. */
 176                 if (!q->p_params.dq_rate_estimator)
 177                         pie_set_enqueue_time(skb);
 178
 179                 pkt_len = qdisc_pkt_len(skb);
 180                 q->stats.packets_in++;
 181                 q->memory_usage += skb->truesize;
 182                 sch->qstats.backlog += pkt_len;
 183                 sch->q.qlen++;
 184                 flow_queue_add(sel_flow, skb);
 185                 if (list_empty(&sel_flow->flowchain)) {
 186                         list_add_tail(&sel_flow->flowchain, &q->new_flows);
 187                         q->new_flow_count++;
 188                         sel_flow->deficit = q->quantum;
 189                         sel_flow->qlen = 0;
 190                         sel_flow->backlog = 0;
 191                 }
 192                 sel_flow->qlen++;
 193                 sel_flow->backlog += pkt_len;
 194                 return NET_XMIT_SUCCESS;
 195         }
 196 out:
 197         q->stats.dropped++;
 198         sel_flow->vars.accu_prob = 0;
 199         __qdisc_drop(skb, to_free);
 200         qdisc_qstats_drop(sch);
 201         return NET_XMIT_CN;
 202 }
 203
 204 static const struct nla_policy fq_pie_policy[TCA_FQ_PIE_MAX + 1] = {
 205         [TCA_FQ_PIE_LIMIT]              = {.type = NLA_U32},
 206         [TCA_FQ_PIE_FLOWS]              = {.type = NLA_U32},
 207         [TCA_FQ_PIE_TARGET]             = {.type = NLA_U32},
 208         [TCA_FQ_PIE_TUPDATE]            = {.type = NLA_U32},
 209         [TCA_FQ_PIE_ALPHA]              = {.type = NLA_U32},
 210         [TCA_FQ_PIE_BETA]               = {.type = NLA_U32},
 211         [TCA_FQ_PIE_QUANTUM]            = {.type = NLA_U32},
 212         [TCA_FQ_PIE_MEMORY_LIMIT]       = {.type = NLA_U32},
 213         [TCA_FQ_PIE_ECN_PROB]           = {.type = NLA_U32},
 214         [TCA_FQ_PIE_ECN]                = {.type = NLA_U32},
 215         [TCA_FQ_PIE_BYTEMODE]           = {.type = NLA_U32},
 216         [TCA_FQ_PIE_DQ_RATE_ESTIMATOR]  = {.type = NLA_U32},
 217 };
 218
 219 static inline struct sk_buff *dequeue_head(struct fq_pie_flow *flow)
 220 {
 221         struct sk_buff *skb = flow->head;
 222
 223         flow->head = skb->next;
 224         skb->next = NULL;
 225         return skb;
 226 }
 227
 228 static struct sk_buff *fq_pie_qdisc_dequeue(struct Qdisc *sch)
 229 {
 230         struct fq_pie_sched_data *q = qdisc_priv(sch);
 231         struct sk_buff *skb = NULL;
 232         struct fq_pie_flow *flow;
 233         struct list_head *head;
 234         u32 pkt_len;
 235
 236 begin:
 237         head = &q->new_flows;
 238         if (list_empty(head)) {
 239                 head = &q->old_flows;
 240                 if (list_empty(head))
 241                         return NULL;
 242         }
 243
 244         flow = list_first_entry(head, struct fq_pie_flow, flowchain);
 245         /* Flow has exhausted all its credits */
 246         if (flow->deficit <= 0) {
 247                 flow->deficit += q->quantum;
 248                 list_move_tail(&flow->flowchain, &q->old_flows);
 249                 goto begin;
 250         }
 251
 252         if (flow->head) {
 253                 skb = dequeue_head(flow);
 254                 pkt_len = qdisc_pkt_len(skb);
 255                 sch->qstats.backlog -= pkt_len;
 256                 sch->q.qlen--;
 257                 qdisc_bstats_update(sch, skb);
 258         }
 259
 260         if (!skb) {
 261                 /* force a pass through old_flows to prevent starvation */
 262                 if (head == &q->new_flows && !list_empty(&q->old_flows))
 263                         list_move_tail(&flow->flowchain, &q->old_flows);
 264                 else
 265                         list_del_init(&flow->flowchain);
 266                 goto begin;
 267         }
 268
 269         flow->qlen--;
 270         flow->deficit -= pkt_len;
 271         flow->backlog -= pkt_len;
 272         q->memory_usage -= get_pie_cb(skb)->mem_usage;
 273         pie_process_dequeue(skb, &q->p_params, &flow->vars, flow->backlog);
 274         return skb;
 275 }
 276
 277 static int fq_pie_change(struct Qdisc *sch, struct nlattr *opt,
 278                          struct netlink_ext_ack *extack)
 279 {
 280         struct fq_pie_sched_data *q = qdisc_priv(sch);
 281         struct nlattr *tb[TCA_FQ_PIE_MAX + 1];
 282         unsigned int len_dropped = 0;
 283         unsigned int num_dropped = 0;
 284         int err;
 285
 286         if (!opt)
 287                 return -EINVAL;
 288
 289         err = nla_parse_nested(tb, TCA_FQ_PIE_MAX, opt, fq_pie_policy, extack);
 290         if (err < 0)
 291                 return err;
 292
 293         sch_tree_lock(sch);
 294         if (tb[TCA_FQ_PIE_LIMIT]) {
 295                 u32 limit = nla_get_u32(tb[TCA_FQ_PIE_LIMIT]);
 296
 297                 q->p_params.limit = limit;
 298                 sch->limit = limit;
 299         }
 300         if (tb[TCA_FQ_PIE_FLOWS]) {
 301                 if (q->flows) {
 302                         NL_SET_ERR_MSG_MOD(extack,
 303                                            "Number of flows cannot be changed");
 304                         goto flow_error;
 305                 }
 306                 q->flows_cnt = nla_get_u32(tb[TCA_FQ_PIE_FLOWS]);
 307                 if (!q->flows_cnt || q->flows_cnt > 65536) {
 308                         NL_SET_ERR_MSG_MOD(extack,
 309                                            "Number of flows must range in [1..65536]");
 310                         goto flow_error;
 311                 }
 312         }
 313
 314         /* convert from microseconds to pschedtime */
 315         if (tb[TCA_FQ_PIE_TARGET]) {
 316                 /* target is in us */
 317                 u32 target = nla_get_u32(tb[TCA_FQ_PIE_TARGET]);
 318
 319                 /* convert to pschedtime */
 320                 q->p_params.target =
 321                         PSCHED_NS2TICKS((u64)target * NSEC_PER_USEC);
 322         }
 323
 324         /* tupdate is in jiffies */
 325         if (tb[TCA_FQ_PIE_TUPDATE])
 326                 q->p_params.tupdate =
 327                         usecs_to_jiffies(nla_get_u32(tb[TCA_FQ_PIE_TUPDATE]));
 328
 329         if (tb[TCA_FQ_PIE_ALPHA])
 330                 q->p_params.alpha = nla_get_u32(tb[TCA_FQ_PIE_ALPHA]);
 331
 332         if (tb[TCA_FQ_PIE_BETA])
 333                 q->p_params.beta = nla_get_u32(tb[TCA_FQ_PIE_BETA]);
 334
 335         if (tb[TCA_FQ_PIE_QUANTUM])
 336                 q->quantum = nla_get_u32(tb[TCA_FQ_PIE_QUANTUM]);
 337
 338         if (tb[TCA_FQ_PIE_MEMORY_LIMIT])
 339                 q->memory_limit = nla_get_u32(tb[TCA_FQ_PIE_MEMORY_LIMIT]);
 340
 341         if (tb[TCA_FQ_PIE_ECN_PROB])
 342                 q->ecn_prob = nla_get_u32(tb[TCA_FQ_PIE_ECN_PROB]);
 343
 344         if (tb[TCA_FQ_PIE_ECN])
 345                 q->p_params.ecn = nla_get_u32(tb[TCA_FQ_PIE_ECN]);
 346
 347         if (tb[TCA_FQ_PIE_BYTEMODE])
 348                 q->p_params.bytemode = nla_get_u32(tb[TCA_FQ_PIE_BYTEMODE]);
 349
 350         if (tb[TCA_FQ_PIE_DQ_RATE_ESTIMATOR])
 351                 q->p_params.dq_rate_estimator =
 352                         nla_get_u32(tb[TCA_FQ_PIE_DQ_RATE_ESTIMATOR]);
 353
 354         /* Drop excess packets if new limit is lower */
 355         while (sch->q.qlen > sch->limit) {
 356                 struct sk_buff *skb = fq_pie_qdisc_dequeue(sch);
 357
 358                 len_dropped += qdisc_pkt_len(skb);
 359                 num_dropped += 1;
 360                 rtnl_kfree_skbs(skb, skb);
 361         }
 362         qdisc_tree_reduce_backlog(sch, num_dropped, len_dropped);
 363
 364         sch_tree_unlock(sch);
 365         return 0;
 366
 367 flow_error:
 368         sch_tree_unlock(sch);
 369         return -EINVAL;
 370 }
 371
 372 static void fq_pie_timer(struct timer_list *t)
 373 {
 374         struct fq_pie_sched_data *q = from_timer(q, t, adapt_timer);
 375         struct Qdisc *sch = q->sch;
 376         spinlock_t *root_lock; /* to lock qdisc for probability calculations */
 377         u32 idx;
 378
 379         root_lock = qdisc_lock(qdisc_root_sleeping(sch));
 380         spin_lock(root_lock);
 381
 382         for (idx = 0; idx < q->flows_cnt; idx++)
 383                 pie_calculate_probability(&q->p_params, &q->flows[idx].vars,
 384                                           q->flows[idx].backlog);
 385
 386         /* reset the timer to fire after 'tupdate' jiffies. */
 387         if (q->p_params.tupdate)
 388                 mod_timer(&q->adapt_timer, jiffies + q->p_params.tupdate);
 389
 390         spin_unlock(root_lock);
 391 }
 392
 393 static int fq_pie_init(struct Qdisc *sch, struct nlattr *opt,
 394                        struct netlink_ext_ack *extack)
 395 {
 396         struct fq_pie_sched_data *q = qdisc_priv(sch);
 397         int err;
 398         u32 idx;
 399
 400         pie_params_init(&q->p_params);
 401         sch->limit = 10 * 1024;
 402         q->p_params.limit = sch->limit;
 403         q->quantum = psched_mtu(qdisc_dev(sch));
 404         q->sch = sch;
 405         q->ecn_prob = 10;
 406         q->flows_cnt = 1024;
 407         q->memory_limit = SZ_32M;
 408
 409         INIT_LIST_HEAD(&q->new_flows);
 410         INIT_LIST_HEAD(&q->old_flows);
 411         timer_setup(&q->adapt_timer, fq_pie_timer, 0);
 412
 413         if (opt) {
 414                 err = fq_pie_change(sch, opt, extack);
 415
 416                 if (err)
 417                         return err;
 418         }
 419
 420         err = tcf_block_get(&q->block, &q->filter_list, sch, extack);
 421         if (err)
 422                 goto init_failure;
 423
 424         q->flows = kvcalloc(q->flows_cnt, sizeof(struct fq_pie_flow),
 425                             GFP_KERNEL);
 426         if (!q->flows) {
 427                 err = -ENOMEM;
 428                 goto init_failure;
 429         }
 430         for (idx = 0; idx < q->flows_cnt; idx++) {
 431                 struct fq_pie_flow *flow = q->flows + idx;
 432
 433                 INIT_LIST_HEAD(&flow->flowchain);
 434                 pie_vars_init(&flow->vars);
 435         }
 436
 437         mod_timer(&q->adapt_timer, jiffies + HZ / 2);
 438
 439         return 0;
 440
 441 init_failure:
 442         q->flows_cnt = 0;
 443
 444         return err;
 445 }
 446
 447 static int fq_pie_dump(struct Qdisc *sch, struct sk_buff *skb)
 448 {
 449         struct fq_pie_sched_data *q = qdisc_priv(sch);
 450         struct nlattr *opts;
 451
 452         opts = nla_nest_start(skb, TCA_OPTIONS);
 453         if (!opts)
 454                 return -EMSGSIZE;
 455
 456         /* convert target from pschedtime to us */
 457         if (nla_put_u32(skb, TCA_FQ_PIE_LIMIT, sch->limit) ||
 458             nla_put_u32(skb, TCA_FQ_PIE_FLOWS, q->flows_cnt) ||
 459             nla_put_u32(skb, TCA_FQ_PIE_TARGET,
 460                         ((u32)PSCHED_TICKS2NS(q->p_params.target)) /
 461                         NSEC_PER_USEC) ||
 462             nla_put_u32(skb, TCA_FQ_PIE_TUPDATE,
 463                         jiffies_to_usecs(q->p_params.tupdate)) ||
 464             nla_put_u32(skb, TCA_FQ_PIE_ALPHA, q->p_params.alpha) ||
 465             nla_put_u32(skb, TCA_FQ_PIE_BETA, q->p_params.beta) ||
 466             nla_put_u32(skb, TCA_FQ_PIE_QUANTUM, q->quantum) ||
 467             nla_put_u32(skb, TCA_FQ_PIE_MEMORY_LIMIT, q->memory_limit) ||
 468             nla_put_u32(skb, TCA_FQ_PIE_ECN_PROB, q->ecn_prob) ||
 469             nla_put_u32(skb, TCA_FQ_PIE_ECN, q->p_params.ecn) ||
 470             nla_put_u32(skb, TCA_FQ_PIE_BYTEMODE, q->p_params.bytemode) ||
 471             nla_put_u32(skb, TCA_FQ_PIE_DQ_RATE_ESTIMATOR,
 472                         q->p_params.dq_rate_estimator))
 473                 goto nla_put_failure;
 474
 475         return nla_nest_end(skb, opts);
 476
 477 nla_put_failure:
 478         nla_nest_cancel(skb, opts);
 479         return -EMSGSIZE;
 480 }
 481
 482 static int fq_pie_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
 483 {
 484         struct fq_pie_sched_data *q = qdisc_priv(sch);
 485         struct tc_fq_pie_xstats st = {
 486                 .packets_in     = q->stats.packets_in,
 487                 .overlimit      = q->stats.overlimit,
 488                 .overmemory     = q->overmemory,
 489                 .dropped        = q->stats.dropped,
 490                 .ecn_mark       = q->stats.ecn_mark,
 491                 .new_flow_count = q->new_flow_count,
 492                 .memory_usage   = q->memory_usage,
 493         };
 494         struct list_head *pos;
 495
 496         sch_tree_lock(sch);
 497         list_for_each(pos, &q->new_flows)
 498                 st.new_flows_len++;
 499
 500         list_for_each(pos, &q->old_flows)
 501                 st.old_flows_len++;
 502         sch_tree_unlock(sch);
 503
 504         return gnet_stats_copy_app(d, &st, sizeof(st));
 505 }
 506
 507 static void fq_pie_reset(struct Qdisc *sch)
 508 {
 509         struct fq_pie_sched_data *q = qdisc_priv(sch);
 510         u32 idx;
 511
 512         INIT_LIST_HEAD(&q->new_flows);
 513         INIT_LIST_HEAD(&q->old_flows);
 514         for (idx = 0; idx < q->flows_cnt; idx++) {
 515                 struct fq_pie_flow *flow = q->flows + idx;
 516
 517                 /* Removes all packets from flow */
 518                 rtnl_kfree_skbs(flow->head, flow->tail);
 519                 flow->head = NULL;
 520
 521                 INIT_LIST_HEAD(&flow->flowchain);
 522                 pie_vars_init(&flow->vars);
 523         }
 524
 525         sch->q.qlen = 0;
 526         sch->qstats.backlog = 0;
 527 }
 528
 529 static void fq_pie_destroy(struct Qdisc *sch)
 530 {
 531         struct fq_pie_sched_data *q = qdisc_priv(sch);
 532
 533         tcf_block_put(q->block);
 534         q->p_params.tupdate = 0;
 535         del_timer_sync(&q->adapt_timer);
 536         kvfree(q->flows);
 537 }
 538
 539 static struct Qdisc_ops fq_pie_qdisc_ops __read_mostly = {
 540         .id             = "fq_pie",
 541         .priv_size      = sizeof(struct fq_pie_sched_data),
 542         .enqueue        = fq_pie_qdisc_enqueue,
 543         .dequeue        = fq_pie_qdisc_dequeue,
 544         .peek           = qdisc_peek_dequeued,
 545         .init           = fq_pie_init,
 546         .destroy        = fq_pie_destroy,
 547         .reset          = fq_pie_reset,
 548         .change         = fq_pie_change,
 549         .dump           = fq_pie_dump,
 550         .dump_stats     = fq_pie_dump_stats,
 551         .owner          = THIS_MODULE,
 552 };
 553
 554 static int __init fq_pie_module_init(void)
 555 {
 556         return register_qdisc(&fq_pie_qdisc_ops);
 557 }
 558
 559 static void __exit fq_pie_module_exit(void)
 560 {
 561         unregister_qdisc(&fq_pie_qdisc_ops);
 562 }
 563
 564 module_init(fq_pie_module_init);
 565 module_exit(fq_pie_module_exit);
 566
 567 MODULE_DESCRIPTION("Flow Queue Proportional Integral controller Enhanced (FQ-PIE)");
 568 MODULE_AUTHOR("Mohit P. Tahiliani");
 569 MODULE_LICENSE("GPL");