net/sched/sch_fq_codel.c

   1 /*
   2  * Fair Queue CoDel discipline
   3  *
   4  *      This program is free software; you can redistribute it and/or
   5  *      modify it under the terms of the GNU General Public License
   6  *      as published by the Free Software Foundation; either version
   7  *      2 of the License, or (at your option) any later version.
   8  *
   9  *  Copyright (C) 2012,2015 Eric Dumazet <edumazet@google.com>
  10  */
  11
  12 #include <linux/module.h>
  13 #include <linux/types.h>
  14 #include <linux/kernel.h>
  15 #include <linux/jiffies.h>
  16 #include <linux/string.h>
  17 #include <linux/in.h>
  18 #include <linux/errno.h>
  19 #include <linux/init.h>
  20 #include <linux/skbuff.h>
  21 #include <linux/jhash.h>
  22 #include <linux/slab.h>
  23 #include <linux/vmalloc.h>
  24 #include <net/netlink.h>
  25 #include <net/pkt_sched.h>
  26 #include <net/pkt_cls.h>
  27 #include <net/codel.h>
  28 #include <net/codel_impl.h>
  29 #include <net/codel_qdisc.h>
  30
  31 /*      Fair Queue CoDel.
  32  *
  33  * Principles :
  34  * Packets are classified (internal classifier or external) on flows.
  35  * This is a Stochastic model (as we use a hash, several flows
  36  *                             might be hashed on same slot)
  37  * Each flow has a CoDel managed queue.
  38  * Flows are linked onto two (Round Robin) lists,
  39  * so that new flows have priority on old ones.
  40  *
  41  * For a given flow, packets are not reordered (CoDel uses a FIFO)
  42  * head drops only.
  43  * ECN capability is on by default.
  44  * Low memory footprint (64 bytes per flow)
  45  */
  46
  47 struct fq_codel_flow {
  48         struct sk_buff    *head;
  49         struct sk_buff    *tail;
  50         struct list_head  flowchain;
  51         int               deficit;
  52         u32               dropped; /* number of drops (or ECN marks) on this flow */
  53         struct codel_vars cvars;
  54 }; /* please try to keep this structure <= 64 bytes */
  55
  56 struct fq_codel_sched_data {
  57         struct tcf_proto __rcu *filter_list; /* optional external classifier */
  58         struct tcf_block *block;
  59         struct fq_codel_flow *flows;    /* Flows table [flows_cnt] */
  60         u32             *backlogs;      /* backlog table [flows_cnt] */
  61         u32             flows_cnt;      /* number of flows */
  62         u32             quantum;        /* psched_mtu(qdisc_dev(sch)); */
  63         u32             drop_batch_size;
  64         u32             memory_limit;
  65         struct codel_params cparams;
  66         struct codel_stats cstats;
  67         u32             memory_usage;
  68         u32             drop_overmemory;
  69         u32             drop_overlimit;
  70         u32             new_flow_count;
  71
  72         struct list_head new_flows;     /* list of new flows */
  73         struct list_head old_flows;     /* list of old flows */
  74 };
  75
  76 static unsigned int fq_codel_hash(const struct fq_codel_sched_data *q,
  77                                   struct sk_buff *skb)
  78 {
  79         return reciprocal_scale(skb_get_hash(skb), q->flows_cnt);
  80 }
  81
  82 static unsigned int fq_codel_classify(struct sk_buff *skb, struct Qdisc *sch,
  83                                       int *qerr)
  84 {
  85         struct fq_codel_sched_data *q = qdisc_priv(sch);
  86         struct tcf_proto *filter;
  87         struct tcf_result res;
  88         int result;
  89
  90         if (TC_H_MAJ(skb->priority) == sch->handle &&
  91             TC_H_MIN(skb->priority) > 0 &&
  92             TC_H_MIN(skb->priority) <= q->flows_cnt)
  93                 return TC_H_MIN(skb->priority);
  94
  95         filter = rcu_dereference_bh(q->filter_list);
  96         if (!filter)
  97                 return fq_codel_hash(q, skb) + 1;
  98
  99         *qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
 100         result = tcf_classify(skb, filter, &res, false);
 101         if (result >= 0) {
 102 #ifdef CONFIG_NET_CLS_ACT
 103                 switch (result) {
 104                 case TC_ACT_STOLEN:
 105                 case TC_ACT_QUEUED:
 106                 case TC_ACT_TRAP:
 107                         *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
 108                         /* fall through */
 109                 case TC_ACT_SHOT:
 110                         return 0;
 111                 }
 112 #endif
 113                 if (TC_H_MIN(res.classid) <= q->flows_cnt)
 114                         return TC_H_MIN(res.classid);
 115         }
 116         return 0;
 117 }
 118
 119 /* helper functions : might be changed when/if skb use a standard list_head */
 120
 121 /* remove one skb from head of slot queue */
 122 static inline struct sk_buff *dequeue_head(struct fq_codel_flow *flow)
 123 {
 124         struct sk_buff *skb = flow->head;
 125
 126         flow->head = skb->next;
 127         skb->next = NULL;
 128         return skb;
 129 }
 130
 131 /* add skb to flow queue (tail add) */
 132 static inline void flow_queue_add(struct fq_codel_flow *flow,
 133                                   struct sk_buff *skb)
 134 {
 135         if (flow->head == NULL)
 136                 flow->head = skb;
 137         else
 138                 flow->tail->next = skb;
 139         flow->tail = skb;
 140         skb->next = NULL;
 141 }
 142
 143 static unsigned int fq_codel_drop(struct Qdisc *sch, unsigned int max_packets,
 144                                   struct sk_buff **to_free)
 145 {
 146         struct fq_codel_sched_data *q = qdisc_priv(sch);
 147         struct sk_buff *skb;
 148         unsigned int maxbacklog = 0, idx = 0, i, len;
 149         struct fq_codel_flow *flow;
 150         unsigned int threshold;
 151         unsigned int mem = 0;
 152
 153         /* Queue is full! Find the fat flow and drop packet(s) from it.
 154          * This might sound expensive, but with 1024 flows, we scan
 155          * 4KB of memory, and we dont need to handle a complex tree
 156          * in fast path (packet queue/enqueue) with many cache misses.
 157          * In stress mode, we'll try to drop 64 packets from the flow,
 158          * amortizing this linear lookup to one cache line per drop.
 159          */
 160         for (i = 0; i < q->flows_cnt; i++) {
 161                 if (q->backlogs[i] > maxbacklog) {
 162                         maxbacklog = q->backlogs[i];
 163                         idx = i;
 164                 }
 165         }
 166
 167         /* Our goal is to drop half of this fat flow backlog */
 168         threshold = maxbacklog >> 1;
 169
 170         flow = &q->flows[idx];
 171         len = 0;
 172         i = 0;
 173         do {
 174                 skb = dequeue_head(flow);
 175                 len += qdisc_pkt_len(skb);
 176                 mem += get_codel_cb(skb)->mem_usage;
 177                 __qdisc_drop(skb, to_free);
 178         } while (++i < max_packets && len < threshold);
 179
 180         flow->dropped += i;
 181         q->backlogs[idx] -= len;
 182         q->memory_usage -= mem;
 183         sch->qstats.drops += i;
 184         sch->qstats.backlog -= len;
 185         sch->q.qlen -= i;
 186         return idx;
 187 }
 188
 189 static int fq_codel_enqueue(struct sk_buff *skb, struct Qdisc *sch,
 190                             struct sk_buff **to_free)
 191 {
 192         struct fq_codel_sched_data *q = qdisc_priv(sch);
 193         unsigned int idx, prev_backlog, prev_qlen;
 194         struct fq_codel_flow *flow;
 195         int uninitialized_var(ret);
 196         unsigned int pkt_len;
 197         bool memory_limited;
 198
 199         idx = fq_codel_classify(skb, sch, &ret);
 200         if (idx == 0) {
 201                 if (ret & __NET_XMIT_BYPASS)
 202                         qdisc_qstats_drop(sch);
 203                 __qdisc_drop(skb, to_free);
 204                 return ret;
 205         }
 206         idx--;
 207
 208         codel_set_enqueue_time(skb);
 209         flow = &q->flows[idx];
 210         flow_queue_add(flow, skb);
 211         q->backlogs[idx] += qdisc_pkt_len(skb);
 212         qdisc_qstats_backlog_inc(sch, skb);
 213
 214         if (list_empty(&flow->flowchain)) {
 215                 list_add_tail(&flow->flowchain, &q->new_flows);
 216                 q->new_flow_count++;
 217                 flow->deficit = q->quantum;
 218                 flow->dropped = 0;
 219         }
 220         get_codel_cb(skb)->mem_usage = skb->truesize;
 221         q->memory_usage += get_codel_cb(skb)->mem_usage;
 222         memory_limited = q->memory_usage > q->memory_limit;
 223         if (++sch->q.qlen <= sch->limit && !memory_limited)
 224                 return NET_XMIT_SUCCESS;
 225
 226         prev_backlog = sch->qstats.backlog;
 227         prev_qlen = sch->q.qlen;
 228
 229         /* save this packet length as it might be dropped by fq_codel_drop() */
 230         pkt_len = qdisc_pkt_len(skb);
 231         /* fq_codel_drop() is quite expensive, as it performs a linear search
 232          * in q->backlogs[] to find a fat flow.
 233          * So instead of dropping a single packet, drop half of its backlog
 234          * with a 64 packets limit to not add a too big cpu spike here.
 235          */
 236         ret = fq_codel_drop(sch, q->drop_batch_size, to_free);
 237
 238         prev_qlen -= sch->q.qlen;
 239         prev_backlog -= sch->qstats.backlog;
 240         q->drop_overlimit += prev_qlen;
 241         if (memory_limited)
 242                 q->drop_overmemory += prev_qlen;
 243
 244         /* As we dropped packet(s), better let upper stack know this.
 245          * If we dropped a packet for this flow, return NET_XMIT_CN,
 246          * but in this case, our parents wont increase their backlogs.
 247          */
 248         if (ret == idx) {
 249                 qdisc_tree_reduce_backlog(sch, prev_qlen - 1,
 250                                           prev_backlog - pkt_len);
 251                 return NET_XMIT_CN;
 252         }
 253         qdisc_tree_reduce_backlog(sch, prev_qlen, prev_backlog);
 254         return NET_XMIT_SUCCESS;
 255 }
 256
 257 /* This is the specific function called from codel_dequeue()
 258  * to dequeue a packet from queue. Note: backlog is handled in
 259  * codel, we dont need to reduce it here.
 260  */
 261 static struct sk_buff *dequeue_func(struct codel_vars *vars, void *ctx)
 262 {
 263         struct Qdisc *sch = ctx;
 264         struct fq_codel_sched_data *q = qdisc_priv(sch);
 265         struct fq_codel_flow *flow;
 266         struct sk_buff *skb = NULL;
 267
 268         flow = container_of(vars, struct fq_codel_flow, cvars);
 269         if (flow->head) {
 270                 skb = dequeue_head(flow);
 271                 q->backlogs[flow - q->flows] -= qdisc_pkt_len(skb);
 272                 q->memory_usage -= get_codel_cb(skb)->mem_usage;
 273                 sch->q.qlen--;
 274                 sch->qstats.backlog -= qdisc_pkt_len(skb);
 275         }
 276         return skb;
 277 }
 278
 279 static void drop_func(struct sk_buff *skb, void *ctx)
 280 {
 281         struct Qdisc *sch = ctx;
 282
 283         kfree_skb(skb);
 284         qdisc_qstats_drop(sch);
 285 }
 286
 287 static struct sk_buff *fq_codel_dequeue(struct Qdisc *sch)
 288 {
 289         struct fq_codel_sched_data *q = qdisc_priv(sch);
 290         struct sk_buff *skb;
 291         struct fq_codel_flow *flow;
 292         struct list_head *head;
 293         u32 prev_drop_count, prev_ecn_mark;
 294
 295 begin:
 296         head = &q->new_flows;
 297         if (list_empty(head)) {
 298                 head = &q->old_flows;
 299                 if (list_empty(head))
 300                         return NULL;
 301         }
 302         flow = list_first_entry(head, struct fq_codel_flow, flowchain);
 303
 304         if (flow->deficit <= 0) {
 305                 flow->deficit += q->quantum;
 306                 list_move_tail(&flow->flowchain, &q->old_flows);
 307                 goto begin;
 308         }
 309
 310         prev_drop_count = q->cstats.drop_count;
 311         prev_ecn_mark = q->cstats.ecn_mark;
 312
 313         skb = codel_dequeue(sch, &sch->qstats.backlog, &q->cparams,
 314                             &flow->cvars, &q->cstats, qdisc_pkt_len,
 315                             codel_get_enqueue_time, drop_func, dequeue_func);
 316
 317         flow->dropped += q->cstats.drop_count - prev_drop_count;
 318         flow->dropped += q->cstats.ecn_mark - prev_ecn_mark;
 319
 320         if (!skb) {
 321                 /* force a pass through old_flows to prevent starvation */
 322                 if ((head == &q->new_flows) && !list_empty(&q->old_flows))
 323                         list_move_tail(&flow->flowchain, &q->old_flows);
 324                 else
 325                         list_del_init(&flow->flowchain);
 326                 goto begin;
 327         }
 328         qdisc_bstats_update(sch, skb);
 329         flow->deficit -= qdisc_pkt_len(skb);
 330         /* We cant call qdisc_tree_reduce_backlog() if our qlen is 0,
 331          * or HTB crashes. Defer it for next round.
 332          */
 333         if (q->cstats.drop_count && sch->q.qlen) {
 334                 qdisc_tree_reduce_backlog(sch, q->cstats.drop_count,
 335                                           q->cstats.drop_len);
 336                 q->cstats.drop_count = 0;
 337                 q->cstats.drop_len = 0;
 338         }
 339         return skb;
 340 }
 341
 342 static void fq_codel_flow_purge(struct fq_codel_flow *flow)
 343 {
 344         rtnl_kfree_skbs(flow->head, flow->tail);
 345         flow->head = NULL;
 346 }
 347
 348 static void fq_codel_reset(struct Qdisc *sch)
 349 {
 350         struct fq_codel_sched_data *q = qdisc_priv(sch);
 351         int i;
 352
 353         INIT_LIST_HEAD(&q->new_flows);
 354         INIT_LIST_HEAD(&q->old_flows);
 355         for (i = 0; i < q->flows_cnt; i++) {
 356                 struct fq_codel_flow *flow = q->flows + i;
 357
 358                 fq_codel_flow_purge(flow);
 359                 INIT_LIST_HEAD(&flow->flowchain);
 360                 codel_vars_init(&flow->cvars);
 361         }
 362         memset(q->backlogs, 0, q->flows_cnt * sizeof(u32));
 363         sch->q.qlen = 0;
 364         sch->qstats.backlog = 0;
 365         q->memory_usage = 0;
 366 }
 367
 368 static const struct nla_policy fq_codel_policy[TCA_FQ_CODEL_MAX + 1] = {
 369         [TCA_FQ_CODEL_TARGET]   = { .type = NLA_U32 },
 370         [TCA_FQ_CODEL_LIMIT]    = { .type = NLA_U32 },
 371         [TCA_FQ_CODEL_INTERVAL] = { .type = NLA_U32 },
 372         [TCA_FQ_CODEL_ECN]      = { .type = NLA_U32 },
 373         [TCA_FQ_CODEL_FLOWS]    = { .type = NLA_U32 },
 374         [TCA_FQ_CODEL_QUANTUM]  = { .type = NLA_U32 },
 375         [TCA_FQ_CODEL_CE_THRESHOLD] = { .type = NLA_U32 },
 376         [TCA_FQ_CODEL_DROP_BATCH_SIZE] = { .type = NLA_U32 },
 377         [TCA_FQ_CODEL_MEMORY_LIMIT] = { .type = NLA_U32 },
 378 };
 379
 380 static int fq_codel_change(struct Qdisc *sch, struct nlattr *opt,
 381                            struct netlink_ext_ack *extack)
 382 {
 383         struct fq_codel_sched_data *q = qdisc_priv(sch);
 384         struct nlattr *tb[TCA_FQ_CODEL_MAX + 1];
 385         int err;
 386
 387         if (!opt)
 388                 return -EINVAL;
 389
 390         err = nla_parse_nested(tb, TCA_FQ_CODEL_MAX, opt, fq_codel_policy,
 391                                NULL);
 392         if (err < 0)
 393                 return err;
 394         if (tb[TCA_FQ_CODEL_FLOWS]) {
 395                 if (q->flows)
 396                         return -EINVAL;
 397                 q->flows_cnt = nla_get_u32(tb[TCA_FQ_CODEL_FLOWS]);
 398                 if (!q->flows_cnt ||
 399                     q->flows_cnt > 65536)
 400                         return -EINVAL;
 401         }
 402         sch_tree_lock(sch);
 403
 404         if (tb[TCA_FQ_CODEL_TARGET]) {
 405                 u64 target = nla_get_u32(tb[TCA_FQ_CODEL_TARGET]);
 406
 407                 q->cparams.target = (target * NSEC_PER_USEC) >> CODEL_SHIFT;
 408         }
 409
 410         if (tb[TCA_FQ_CODEL_CE_THRESHOLD]) {
 411                 u64 val = nla_get_u32(tb[TCA_FQ_CODEL_CE_THRESHOLD]);
 412
 413                 q->cparams.ce_threshold = (val * NSEC_PER_USEC) >> CODEL_SHIFT;
 414         }
 415
 416         if (tb[TCA_FQ_CODEL_INTERVAL]) {
 417                 u64 interval = nla_get_u32(tb[TCA_FQ_CODEL_INTERVAL]);
 418
 419                 q->cparams.interval = (interval * NSEC_PER_USEC) >> CODEL_SHIFT;
 420         }
 421
 422         if (tb[TCA_FQ_CODEL_LIMIT])
 423                 sch->limit = nla_get_u32(tb[TCA_FQ_CODEL_LIMIT]);
 424
 425         if (tb[TCA_FQ_CODEL_ECN])
 426                 q->cparams.ecn = !!nla_get_u32(tb[TCA_FQ_CODEL_ECN]);
 427
 428         if (tb[TCA_FQ_CODEL_QUANTUM])
 429                 q->quantum = max(256U, nla_get_u32(tb[TCA_FQ_CODEL_QUANTUM]));
 430
 431         if (tb[TCA_FQ_CODEL_DROP_BATCH_SIZE])
 432                 q->drop_batch_size = min(1U, nla_get_u32(tb[TCA_FQ_CODEL_DROP_BATCH_SIZE]));
 433
 434         if (tb[TCA_FQ_CODEL_MEMORY_LIMIT])
 435                 q->memory_limit = min(1U << 31, nla_get_u32(tb[TCA_FQ_CODEL_MEMORY_LIMIT]));
 436
 437         while (sch->q.qlen > sch->limit ||
 438                q->memory_usage > q->memory_limit) {
 439                 struct sk_buff *skb = fq_codel_dequeue(sch);
 440
 441                 q->cstats.drop_len += qdisc_pkt_len(skb);
 442                 rtnl_kfree_skbs(skb, skb);
 443                 q->cstats.drop_count++;
 444         }
 445         qdisc_tree_reduce_backlog(sch, q->cstats.drop_count, q->cstats.drop_len);
 446         q->cstats.drop_count = 0;
 447         q->cstats.drop_len = 0;
 448
 449         sch_tree_unlock(sch);
 450         return 0;
 451 }
 452
 453 static void fq_codel_destroy(struct Qdisc *sch)
 454 {
 455         struct fq_codel_sched_data *q = qdisc_priv(sch);
 456
 457         tcf_block_put(q->block);
 458         kvfree(q->backlogs);
 459         kvfree(q->flows);
 460 }
 461
 462 static int fq_codel_init(struct Qdisc *sch, struct nlattr *opt,
 463                          struct netlink_ext_ack *extack)
 464 {
 465         struct fq_codel_sched_data *q = qdisc_priv(sch);
 466         int i;
 467         int err;
 468
 469         sch->limit = 10*1024;
 470         q->flows_cnt = 1024;
 471         q->memory_limit = 32 << 20; /* 32 MBytes */
 472         q->drop_batch_size = 64;
 473         q->quantum = psched_mtu(qdisc_dev(sch));
 474         INIT_LIST_HEAD(&q->new_flows);
 475         INIT_LIST_HEAD(&q->old_flows);
 476         codel_params_init(&q->cparams);
 477         codel_stats_init(&q->cstats);
 478         q->cparams.ecn = true;
 479         q->cparams.mtu = psched_mtu(qdisc_dev(sch));
 480
 481         if (opt) {
 482                 int err = fq_codel_change(sch, opt, extack);
 483                 if (err)
 484                         return err;
 485         }
 486
 487         err = tcf_block_get(&q->block, &q->filter_list, sch, extack);
 488         if (err)
 489                 return err;
 490
 491         if (!q->flows) {
 492                 q->flows = kvcalloc(q->flows_cnt,
 493                                     sizeof(struct fq_codel_flow),
 494                                     GFP_KERNEL);
 495                 if (!q->flows)
 496                         return -ENOMEM;
 497                 q->backlogs = kvcalloc(q->flows_cnt, sizeof(u32), GFP_KERNEL);
 498                 if (!q->backlogs)
 499                         return -ENOMEM;
 500                 for (i = 0; i < q->flows_cnt; i++) {
 501                         struct fq_codel_flow *flow = q->flows + i;
 502
 503                         INIT_LIST_HEAD(&flow->flowchain);
 504                         codel_vars_init(&flow->cvars);
 505                 }
 506         }
 507         if (sch->limit >= 1)
 508                 sch->flags |= TCQ_F_CAN_BYPASS;
 509         else
 510                 sch->flags &= ~TCQ_F_CAN_BYPASS;
 511         return 0;
 512 }
 513
 514 static int fq_codel_dump(struct Qdisc *sch, struct sk_buff *skb)
 515 {
 516         struct fq_codel_sched_data *q = qdisc_priv(sch);
 517         struct nlattr *opts;
 518
 519         opts = nla_nest_start(skb, TCA_OPTIONS);
 520         if (opts == NULL)
 521                 goto nla_put_failure;
 522
 523         if (nla_put_u32(skb, TCA_FQ_CODEL_TARGET,
 524                         codel_time_to_us(q->cparams.target)) ||
 525             nla_put_u32(skb, TCA_FQ_CODEL_LIMIT,
 526                         sch->limit) ||
 527             nla_put_u32(skb, TCA_FQ_CODEL_INTERVAL,
 528                         codel_time_to_us(q->cparams.interval)) ||
 529             nla_put_u32(skb, TCA_FQ_CODEL_ECN,
 530                         q->cparams.ecn) ||
 531             nla_put_u32(skb, TCA_FQ_CODEL_QUANTUM,
 532                         q->quantum) ||
 533             nla_put_u32(skb, TCA_FQ_CODEL_DROP_BATCH_SIZE,
 534                         q->drop_batch_size) ||
 535             nla_put_u32(skb, TCA_FQ_CODEL_MEMORY_LIMIT,
 536                         q->memory_limit) ||
 537             nla_put_u32(skb, TCA_FQ_CODEL_FLOWS,
 538                         q->flows_cnt))
 539                 goto nla_put_failure;
 540
 541         if (q->cparams.ce_threshold != CODEL_DISABLED_THRESHOLD &&
 542             nla_put_u32(skb, TCA_FQ_CODEL_CE_THRESHOLD,
 543                         codel_time_to_us(q->cparams.ce_threshold)))
 544                 goto nla_put_failure;
 545
 546         return nla_nest_end(skb, opts);
 547
 548 nla_put_failure:
 549         return -1;
 550 }
 551
 552 static int fq_codel_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
 553 {
 554         struct fq_codel_sched_data *q = qdisc_priv(sch);
 555         struct tc_fq_codel_xstats st = {
 556                 .type                           = TCA_FQ_CODEL_XSTATS_QDISC,
 557         };
 558         struct list_head *pos;
 559
 560         st.qdisc_stats.maxpacket = q->cstats.maxpacket;
 561         st.qdisc_stats.drop_overlimit = q->drop_overlimit;
 562         st.qdisc_stats.ecn_mark = q->cstats.ecn_mark;
 563         st.qdisc_stats.new_flow_count = q->new_flow_count;
 564         st.qdisc_stats.ce_mark = q->cstats.ce_mark;
 565         st.qdisc_stats.memory_usage  = q->memory_usage;
 566         st.qdisc_stats.drop_overmemory = q->drop_overmemory;
 567
 568         sch_tree_lock(sch);
 569         list_for_each(pos, &q->new_flows)
 570                 st.qdisc_stats.new_flows_len++;
 571
 572         list_for_each(pos, &q->old_flows)
 573                 st.qdisc_stats.old_flows_len++;
 574         sch_tree_unlock(sch);
 575
 576         return gnet_stats_copy_app(d, &st, sizeof(st));
 577 }
 578
 579 static struct Qdisc *fq_codel_leaf(struct Qdisc *sch, unsigned long arg)
 580 {
 581         return NULL;
 582 }
 583
 584 static unsigned long fq_codel_find(struct Qdisc *sch, u32 classid)
 585 {
 586         return 0;
 587 }
 588
 589 static unsigned long fq_codel_bind(struct Qdisc *sch, unsigned long parent,
 590                               u32 classid)
 591 {
 592         /* we cannot bypass queue discipline anymore */
 593         sch->flags &= ~TCQ_F_CAN_BYPASS;
 594         return 0;
 595 }
 596
 597 static void fq_codel_unbind(struct Qdisc *q, unsigned long cl)
 598 {
 599 }
 600
 601 static struct tcf_block *fq_codel_tcf_block(struct Qdisc *sch, unsigned long cl,
 602                                             struct netlink_ext_ack *extack)
 603 {
 604         struct fq_codel_sched_data *q = qdisc_priv(sch);
 605
 606         if (cl)
 607                 return NULL;
 608         return q->block;
 609 }
 610
 611 static int fq_codel_dump_class(struct Qdisc *sch, unsigned long cl,
 612                           struct sk_buff *skb, struct tcmsg *tcm)
 613 {
 614         tcm->tcm_handle |= TC_H_MIN(cl);
 615         return 0;
 616 }
 617
 618 static int fq_codel_dump_class_stats(struct Qdisc *sch, unsigned long cl,
 619                                      struct gnet_dump *d)
 620 {
 621         struct fq_codel_sched_data *q = qdisc_priv(sch);
 622         u32 idx = cl - 1;
 623         struct gnet_stats_queue qs = { 0 };
 624         struct tc_fq_codel_xstats xstats;
 625
 626         if (idx < q->flows_cnt) {
 627                 const struct fq_codel_flow *flow = &q->flows[idx];
 628                 const struct sk_buff *skb;
 629
 630                 memset(&xstats, 0, sizeof(xstats));
 631                 xstats.type = TCA_FQ_CODEL_XSTATS_CLASS;
 632                 xstats.class_stats.deficit = flow->deficit;
 633                 xstats.class_stats.ldelay =
 634                         codel_time_to_us(flow->cvars.ldelay);
 635                 xstats.class_stats.count = flow->cvars.count;
 636                 xstats.class_stats.lastcount = flow->cvars.lastcount;
 637                 xstats.class_stats.dropping = flow->cvars.dropping;
 638                 if (flow->cvars.dropping) {
 639                         codel_tdiff_t delta = flow->cvars.drop_next -
 640                                               codel_get_time();
 641
 642                         xstats.class_stats.drop_next = (delta >= 0) ?
 643                                 codel_time_to_us(delta) :
 644                                 -codel_time_to_us(-delta);
 645                 }
 646                 if (flow->head) {
 647                         sch_tree_lock(sch);
 648                         skb = flow->head;
 649                         while (skb) {
 650                                 qs.qlen++;
 651                                 skb = skb->next;
 652                         }
 653                         sch_tree_unlock(sch);
 654                 }
 655                 qs.backlog = q->backlogs[idx];
 656                 qs.drops = flow->dropped;
 657         }
 658         if (gnet_stats_copy_queue(d, NULL, &qs, qs.qlen) < 0)
 659                 return -1;
 660         if (idx < q->flows_cnt)
 661                 return gnet_stats_copy_app(d, &xstats, sizeof(xstats));
 662         return 0;
 663 }
 664
 665 static void fq_codel_walk(struct Qdisc *sch, struct qdisc_walker *arg)
 666 {
 667         struct fq_codel_sched_data *q = qdisc_priv(sch);
 668         unsigned int i;
 669
 670         if (arg->stop)
 671                 return;
 672
 673         for (i = 0; i < q->flows_cnt; i++) {
 674                 if (list_empty(&q->flows[i].flowchain) ||
 675                     arg->count < arg->skip) {
 676                         arg->count++;
 677                         continue;
 678                 }
 679                 if (arg->fn(sch, i + 1, arg) < 0) {
 680                         arg->stop = 1;
 681                         break;
 682                 }
 683                 arg->count++;
 684         }
 685 }
 686
 687 static const struct Qdisc_class_ops fq_codel_class_ops = {
 688         .leaf           =       fq_codel_leaf,
 689         .find           =       fq_codel_find,
 690         .tcf_block      =       fq_codel_tcf_block,
 691         .bind_tcf       =       fq_codel_bind,
 692         .unbind_tcf     =       fq_codel_unbind,
 693         .dump           =       fq_codel_dump_class,
 694         .dump_stats     =       fq_codel_dump_class_stats,
 695         .walk           =       fq_codel_walk,
 696 };
 697
 698 static struct Qdisc_ops fq_codel_qdisc_ops __read_mostly = {
 699         .cl_ops         =       &fq_codel_class_ops,
 700         .id             =       "fq_codel",
 701         .priv_size      =       sizeof(struct fq_codel_sched_data),
 702         .enqueue        =       fq_codel_enqueue,
 703         .dequeue        =       fq_codel_dequeue,
 704         .peek           =       qdisc_peek_dequeued,
 705         .init           =       fq_codel_init,
 706         .reset          =       fq_codel_reset,
 707         .destroy        =       fq_codel_destroy,
 708         .change         =       fq_codel_change,
 709         .dump           =       fq_codel_dump,
 710         .dump_stats =   fq_codel_dump_stats,
 711         .owner          =       THIS_MODULE,
 712 };
 713
 714 static int __init fq_codel_module_init(void)
 715 {
 716         return register_qdisc(&fq_codel_qdisc_ops);
 717 }
 718
 719 static void __exit fq_codel_module_exit(void)
 720 {
 721         unregister_qdisc(&fq_codel_qdisc_ops);
 722 }
 723
 724 module_init(fq_codel_module_init)
 725 module_exit(fq_codel_module_exit)
 726 MODULE_AUTHOR("Eric Dumazet");
 727 MODULE_LICENSE("GPL");