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/codel.h>
  27 #include <net/codel_impl.h>
  28 #include <net/codel_qdisc.h>
  29
  30 /*      Fair Queue CoDel.
  31  *
  32  * Principles :
  33  * Packets are classified (internal classifier or external) on flows.
  34  * This is a Stochastic model (as we use a hash, several flows
  35  *                             might be hashed on same slot)
  36  * Each flow has a CoDel managed queue.
  37  * Flows are linked onto two (Round Robin) lists,
  38  * so that new flows have priority on old ones.
  39  *
  40  * For a given flow, packets are not reordered (CoDel uses a FIFO)
  41  * head drops only.
  42  * ECN capability is on by default.
  43  * Low memory footprint (64 bytes per flow)
  44  */
  45
  46 struct fq_codel_flow {
  47         struct sk_buff    *head;
  48         struct sk_buff    *tail;
  49         struct list_head  flowchain;
  50         int               deficit;
  51         u32               dropped; /* number of drops (or ECN marks) on this flow */
  52         struct codel_vars cvars;
  53 }; /* please try to keep this structure <= 64 bytes */
  54
  55 struct fq_codel_sched_data {
  56         struct tcf_proto __rcu *filter_list; /* optional external classifier */
  57         struct fq_codel_flow *flows;    /* Flows table [flows_cnt] */
  58         u32             *backlogs;      /* backlog table [flows_cnt] */
  59         u32             flows_cnt;      /* number of flows */
  60         u32             perturbation;   /* hash perturbation */
  61         u32             quantum;        /* psched_mtu(qdisc_dev(sch)); */
  62         u32             drop_batch_size;
  63         u32             memory_limit;
  64         struct codel_params cparams;
  65         struct codel_stats cstats;
  66         u32             memory_usage;
  67         u32             drop_overmemory;
  68         u32             drop_overlimit;
  69         u32             new_flow_count;
  70
  71         struct list_head new_flows;     /* list of new flows */
  72         struct list_head old_flows;     /* list of old flows */
  73 };
  74
  75 static unsigned int fq_codel_hash(const struct fq_codel_sched_data *q,
  76                                   struct sk_buff *skb)
  77 {
  78         u32 hash = skb_get_hash_perturb(skb, q->perturbation);
  79
  80         return reciprocal_scale(hash, q->flows_cnt);
  81 }
  82
  83 static unsigned int fq_codel_classify(struct sk_buff *skb, struct Qdisc *sch,
  84                                       int *qerr)
  85 {
  86         struct fq_codel_sched_data *q = qdisc_priv(sch);
  87         struct tcf_proto *filter;
  88         struct tcf_result res;
  89         int result;
  90
  91         if (TC_H_MAJ(skb->priority) == sch->handle &&
  92             TC_H_MIN(skb->priority) > 0 &&
  93             TC_H_MIN(skb->priority) <= q->flows_cnt)
  94                 return TC_H_MIN(skb->priority);
  95
  96         filter = rcu_dereference_bh(q->filter_list);
  97         if (!filter)
  98                 return fq_codel_hash(q, skb) + 1;
  99
 100         *qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
 101         result = tc_classify(skb, filter, &res, false);
 102         if (result >= 0) {
 103 #ifdef CONFIG_NET_CLS_ACT
 104                 switch (result) {
 105                 case TC_ACT_STOLEN:
 106                 case TC_ACT_QUEUED:
 107                         *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
 108                 case TC_ACT_SHOT:
 109                         return 0;
 110                 }
 111 #endif
 112                 if (TC_H_MIN(res.classid) <= q->flows_cnt)
 113                         return TC_H_MIN(res.classid);
 114         }
 115         return 0;
 116 }
 117
 118 /* helper functions : might be changed when/if skb use a standard list_head */
 119
 120 /* remove one skb from head of slot queue */
 121 static inline struct sk_buff *dequeue_head(struct fq_codel_flow *flow)
 122 {
 123         struct sk_buff *skb = flow->head;
 124
 125         flow->head = skb->next;
 126         skb->next = NULL;
 127         return skb;
 128 }
 129
 130 /* add skb to flow queue (tail add) */
 131 static inline void flow_queue_add(struct fq_codel_flow *flow,
 132                                   struct sk_buff *skb)
 133 {
 134         if (flow->head == NULL)
 135                 flow->head = skb;
 136         else
 137                 flow->tail->next = skb;
 138         flow->tail = skb;
 139         skb->next = NULL;
 140 }
 141
 142 static unsigned int fq_codel_drop(struct Qdisc *sch, unsigned int max_packets)
 143 {
 144         struct fq_codel_sched_data *q = qdisc_priv(sch);
 145         struct sk_buff *skb;
 146         unsigned int maxbacklog = 0, idx = 0, i, len;
 147         struct fq_codel_flow *flow;
 148         unsigned int threshold;
 149         unsigned int mem = 0;
 150
 151         /* Queue is full! Find the fat flow and drop packet(s) from it.
 152          * This might sound expensive, but with 1024 flows, we scan
 153          * 4KB of memory, and we dont need to handle a complex tree
 154          * in fast path (packet queue/enqueue) with many cache misses.
 155          * In stress mode, we'll try to drop 64 packets from the flow,
 156          * amortizing this linear lookup to one cache line per drop.
 157          */
 158         for (i = 0; i < q->flows_cnt; i++) {
 159                 if (q->backlogs[i] > maxbacklog) {
 160                         maxbacklog = q->backlogs[i];
 161                         idx = i;
 162                 }
 163         }
 164
 165         /* Our goal is to drop half of this fat flow backlog */
 166         threshold = maxbacklog >> 1;
 167
 168         flow = &q->flows[idx];
 169         len = 0;
 170         i = 0;
 171         do {
 172                 skb = dequeue_head(flow);
 173                 len += qdisc_pkt_len(skb);
 174                 mem += skb->truesize;
 175                 kfree_skb(skb);
 176         } while (++i < max_packets && len < threshold);
 177
 178         flow->dropped += i;
 179         q->backlogs[idx] -= len;
 180         q->memory_usage -= mem;
 181         sch->qstats.drops += i;
 182         sch->qstats.backlog -= len;
 183         sch->q.qlen -= i;
 184         return idx;
 185 }
 186
 187 static unsigned int fq_codel_qdisc_drop(struct Qdisc *sch)
 188 {
 189         unsigned int prev_backlog;
 190
 191         prev_backlog = sch->qstats.backlog;
 192         fq_codel_drop(sch, 1U);
 193         return prev_backlog - sch->qstats.backlog;
 194 }
 195
 196 static int fq_codel_enqueue(struct sk_buff *skb, struct Qdisc *sch)
 197 {
 198         struct fq_codel_sched_data *q = qdisc_priv(sch);
 199         unsigned int idx, prev_backlog, prev_qlen;
 200         struct fq_codel_flow *flow;
 201         int uninitialized_var(ret);
 202         bool memory_limited;
 203
 204         idx = fq_codel_classify(skb, sch, &ret);
 205         if (idx == 0) {
 206                 if (ret & __NET_XMIT_BYPASS)
 207                         qdisc_qstats_drop(sch);
 208                 kfree_skb(skb);
 209                 return ret;
 210         }
 211         idx--;
 212
 213         codel_set_enqueue_time(skb);
 214         flow = &q->flows[idx];
 215         flow_queue_add(flow, skb);
 216         q->backlogs[idx] += qdisc_pkt_len(skb);
 217         qdisc_qstats_backlog_inc(sch, skb);
 218
 219         if (list_empty(&flow->flowchain)) {
 220                 list_add_tail(&flow->flowchain, &q->new_flows);
 221                 q->new_flow_count++;
 222                 flow->deficit = q->quantum;
 223                 flow->dropped = 0;
 224         }
 225         q->memory_usage += skb->truesize;
 226         memory_limited = q->memory_usage > q->memory_limit;
 227         if (++sch->q.qlen <= sch->limit && !memory_limited)
 228                 return NET_XMIT_SUCCESS;
 229
 230         prev_backlog = sch->qstats.backlog;
 231         prev_qlen = sch->q.qlen;
 232
 233         /* fq_codel_drop() is quite expensive, as it performs a linear search
 234          * in q->backlogs[] to find a fat flow.
 235          * So instead of dropping a single packet, drop half of its backlog
 236          * with a 64 packets limit to not add a too big cpu spike here.
 237          */
 238         ret = fq_codel_drop(sch, q->drop_batch_size);
 239
 240         q->drop_overlimit += prev_qlen - sch->q.qlen;
 241         if (memory_limited)
 242                 q->drop_overmemory += prev_qlen - sch->q.qlen;
 243         /* As we dropped packet(s), better let upper stack know this */
 244         qdisc_tree_reduce_backlog(sch, prev_qlen - sch->q.qlen,
 245                                   prev_backlog - sch->qstats.backlog);
 246
 247         return ret == idx ? NET_XMIT_CN : NET_XMIT_SUCCESS;
 248 }
 249
 250 /* This is the specific function called from codel_dequeue()
 251  * to dequeue a packet from queue. Note: backlog is handled in
 252  * codel, we dont need to reduce it here.
 253  */
 254 static struct sk_buff *dequeue_func(struct codel_vars *vars, void *ctx)
 255 {
 256         struct Qdisc *sch = ctx;
 257         struct fq_codel_sched_data *q = qdisc_priv(sch);
 258         struct fq_codel_flow *flow;
 259         struct sk_buff *skb = NULL;
 260
 261         flow = container_of(vars, struct fq_codel_flow, cvars);
 262         if (flow->head) {
 263                 skb = dequeue_head(flow);
 264                 q->backlogs[flow - q->flows] -= qdisc_pkt_len(skb);
 265                 sch->q.qlen--;
 266                 sch->qstats.backlog -= qdisc_pkt_len(skb);
 267         }
 268         return skb;
 269 }
 270
 271 static void drop_func(struct sk_buff *skb, void *ctx)
 272 {
 273         struct Qdisc *sch = ctx;
 274
 275         qdisc_drop(skb, sch);
 276 }
 277
 278 static struct sk_buff *fq_codel_dequeue(struct Qdisc *sch)
 279 {
 280         struct fq_codel_sched_data *q = qdisc_priv(sch);
 281         struct sk_buff *skb;
 282         struct fq_codel_flow *flow;
 283         struct list_head *head;
 284         u32 prev_drop_count, prev_ecn_mark;
 285         unsigned int prev_backlog;
 286
 287 begin:
 288         head = &q->new_flows;
 289         if (list_empty(head)) {
 290                 head = &q->old_flows;
 291                 if (list_empty(head))
 292                         return NULL;
 293         }
 294         flow = list_first_entry(head, struct fq_codel_flow, flowchain);
 295
 296         if (flow->deficit <= 0) {
 297                 flow->deficit += q->quantum;
 298                 list_move_tail(&flow->flowchain, &q->old_flows);
 299                 goto begin;
 300         }
 301
 302         prev_drop_count = q->cstats.drop_count;
 303         prev_ecn_mark = q->cstats.ecn_mark;
 304         prev_backlog = sch->qstats.backlog;
 305
 306         skb = codel_dequeue(sch, &sch->qstats.backlog, &q->cparams,
 307                             &flow->cvars, &q->cstats, qdisc_pkt_len,
 308                             codel_get_enqueue_time, drop_func, dequeue_func);
 309
 310         flow->dropped += q->cstats.drop_count - prev_drop_count;
 311         flow->dropped += q->cstats.ecn_mark - prev_ecn_mark;
 312
 313         if (!skb) {
 314                 /* force a pass through old_flows to prevent starvation */
 315                 if ((head == &q->new_flows) && !list_empty(&q->old_flows))
 316                         list_move_tail(&flow->flowchain, &q->old_flows);
 317                 else
 318                         list_del_init(&flow->flowchain);
 319                 goto begin;
 320         }
 321         q->memory_usage -= skb->truesize;
 322         qdisc_bstats_update(sch, skb);
 323         flow->deficit -= qdisc_pkt_len(skb);
 324         /* We cant call qdisc_tree_reduce_backlog() if our qlen is 0,
 325          * or HTB crashes. Defer it for next round.
 326          */
 327         if (q->cstats.drop_count && sch->q.qlen) {
 328                 qdisc_tree_reduce_backlog(sch, q->cstats.drop_count,
 329                                           q->cstats.drop_len);
 330                 q->cstats.drop_count = 0;
 331                 q->cstats.drop_len = 0;
 332         }
 333         return skb;
 334 }
 335
 336 static void fq_codel_reset(struct Qdisc *sch)
 337 {
 338         struct fq_codel_sched_data *q = qdisc_priv(sch);
 339         int i;
 340
 341         INIT_LIST_HEAD(&q->new_flows);
 342         INIT_LIST_HEAD(&q->old_flows);
 343         for (i = 0; i < q->flows_cnt; i++) {
 344                 struct fq_codel_flow *flow = q->flows + i;
 345
 346                 while (flow->head) {
 347                         struct sk_buff *skb = dequeue_head(flow);
 348
 349                         qdisc_qstats_backlog_dec(sch, skb);
 350                         kfree_skb(skb);
 351                 }
 352
 353                 INIT_LIST_HEAD(&flow->flowchain);
 354                 codel_vars_init(&flow->cvars);
 355         }
 356         memset(q->backlogs, 0, q->flows_cnt * sizeof(u32));
 357         sch->q.qlen = 0;
 358 }
 359
 360 static const struct nla_policy fq_codel_policy[TCA_FQ_CODEL_MAX + 1] = {
 361         [TCA_FQ_CODEL_TARGET]   = { .type = NLA_U32 },
 362         [TCA_FQ_CODEL_LIMIT]    = { .type = NLA_U32 },
 363         [TCA_FQ_CODEL_INTERVAL] = { .type = NLA_U32 },
 364         [TCA_FQ_CODEL_ECN]      = { .type = NLA_U32 },
 365         [TCA_FQ_CODEL_FLOWS]    = { .type = NLA_U32 },
 366         [TCA_FQ_CODEL_QUANTUM]  = { .type = NLA_U32 },
 367         [TCA_FQ_CODEL_CE_THRESHOLD] = { .type = NLA_U32 },
 368         [TCA_FQ_CODEL_DROP_BATCH_SIZE] = { .type = NLA_U32 },
 369         [TCA_FQ_CODEL_MEMORY_LIMIT] = { .type = NLA_U32 },
 370 };
 371
 372 static int fq_codel_change(struct Qdisc *sch, struct nlattr *opt)
 373 {
 374         struct fq_codel_sched_data *q = qdisc_priv(sch);
 375         struct nlattr *tb[TCA_FQ_CODEL_MAX + 1];
 376         int err;
 377
 378         if (!opt)
 379                 return -EINVAL;
 380
 381         err = nla_parse_nested(tb, TCA_FQ_CODEL_MAX, opt, fq_codel_policy);
 382         if (err < 0)
 383                 return err;
 384         if (tb[TCA_FQ_CODEL_FLOWS]) {
 385                 if (q->flows)
 386                         return -EINVAL;
 387                 q->flows_cnt = nla_get_u32(tb[TCA_FQ_CODEL_FLOWS]);
 388                 if (!q->flows_cnt ||
 389                     q->flows_cnt > 65536)
 390                         return -EINVAL;
 391         }
 392         sch_tree_lock(sch);
 393
 394         if (tb[TCA_FQ_CODEL_TARGET]) {
 395                 u64 target = nla_get_u32(tb[TCA_FQ_CODEL_TARGET]);
 396
 397                 q->cparams.target = (target * NSEC_PER_USEC) >> CODEL_SHIFT;
 398         }
 399
 400         if (tb[TCA_FQ_CODEL_CE_THRESHOLD]) {
 401                 u64 val = nla_get_u32(tb[TCA_FQ_CODEL_CE_THRESHOLD]);
 402
 403                 q->cparams.ce_threshold = (val * NSEC_PER_USEC) >> CODEL_SHIFT;
 404         }
 405
 406         if (tb[TCA_FQ_CODEL_INTERVAL]) {
 407                 u64 interval = nla_get_u32(tb[TCA_FQ_CODEL_INTERVAL]);
 408
 409                 q->cparams.interval = (interval * NSEC_PER_USEC) >> CODEL_SHIFT;
 410         }
 411
 412         if (tb[TCA_FQ_CODEL_LIMIT])
 413                 sch->limit = nla_get_u32(tb[TCA_FQ_CODEL_LIMIT]);
 414
 415         if (tb[TCA_FQ_CODEL_ECN])
 416                 q->cparams.ecn = !!nla_get_u32(tb[TCA_FQ_CODEL_ECN]);
 417
 418         if (tb[TCA_FQ_CODEL_QUANTUM])
 419                 q->quantum = max(256U, nla_get_u32(tb[TCA_FQ_CODEL_QUANTUM]));
 420
 421         if (tb[TCA_FQ_CODEL_DROP_BATCH_SIZE])
 422                 q->drop_batch_size = min(1U, nla_get_u32(tb[TCA_FQ_CODEL_DROP_BATCH_SIZE]));
 423
 424         if (tb[TCA_FQ_CODEL_MEMORY_LIMIT])
 425                 q->memory_limit = min(1U << 31, nla_get_u32(tb[TCA_FQ_CODEL_MEMORY_LIMIT]));
 426
 427         while (sch->q.qlen > sch->limit ||
 428                q->memory_usage > q->memory_limit) {
 429                 struct sk_buff *skb = fq_codel_dequeue(sch);
 430
 431                 q->cstats.drop_len += qdisc_pkt_len(skb);
 432                 kfree_skb(skb);
 433                 q->cstats.drop_count++;
 434         }
 435         qdisc_tree_reduce_backlog(sch, q->cstats.drop_count, q->cstats.drop_len);
 436         q->cstats.drop_count = 0;
 437         q->cstats.drop_len = 0;
 438
 439         sch_tree_unlock(sch);
 440         return 0;
 441 }
 442
 443 static void *fq_codel_zalloc(size_t sz)
 444 {
 445         void *ptr = kzalloc(sz, GFP_KERNEL | __GFP_NOWARN);
 446
 447         if (!ptr)
 448                 ptr = vzalloc(sz);
 449         return ptr;
 450 }
 451
 452 static void fq_codel_free(void *addr)
 453 {
 454         kvfree(addr);
 455 }
 456
 457 static void fq_codel_destroy(struct Qdisc *sch)
 458 {
 459         struct fq_codel_sched_data *q = qdisc_priv(sch);
 460
 461         tcf_destroy_chain(&q->filter_list);
 462         fq_codel_free(q->backlogs);
 463         fq_codel_free(q->flows);
 464 }
 465
 466 static int fq_codel_init(struct Qdisc *sch, struct nlattr *opt)
 467 {
 468         struct fq_codel_sched_data *q = qdisc_priv(sch);
 469         int i;
 470
 471         sch->limit = 10*1024;
 472         q->flows_cnt = 1024;
 473         q->memory_limit = 32 << 20; /* 32 MBytes */
 474         q->drop_batch_size = 64;
 475         q->quantum = psched_mtu(qdisc_dev(sch));
 476         q->perturbation = prandom_u32();
 477         INIT_LIST_HEAD(&q->new_flows);
 478         INIT_LIST_HEAD(&q->old_flows);
 479         codel_params_init(&q->cparams);
 480         codel_stats_init(&q->cstats);
 481         q->cparams.ecn = true;
 482         q->cparams.mtu = psched_mtu(qdisc_dev(sch));
 483
 484         if (opt) {
 485                 int err = fq_codel_change(sch, opt);
 486                 if (err)
 487                         return err;
 488         }
 489
 490         if (!q->flows) {
 491                 q->flows = fq_codel_zalloc(q->flows_cnt *
 492                                            sizeof(struct fq_codel_flow));
 493                 if (!q->flows)
 494                         return -ENOMEM;
 495                 q->backlogs = fq_codel_zalloc(q->flows_cnt * sizeof(u32));
 496                 if (!q->backlogs) {
 497                         fq_codel_free(q->flows);
 498                         return -ENOMEM;
 499                 }
 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         list_for_each(pos, &q->new_flows)
 569                 st.qdisc_stats.new_flows_len++;
 570
 571         list_for_each(pos, &q->old_flows)
 572                 st.qdisc_stats.old_flows_len++;
 573
 574         return gnet_stats_copy_app(d, &st, sizeof(st));
 575 }
 576
 577 static struct Qdisc *fq_codel_leaf(struct Qdisc *sch, unsigned long arg)
 578 {
 579         return NULL;
 580 }
 581
 582 static unsigned long fq_codel_get(struct Qdisc *sch, u32 classid)
 583 {
 584         return 0;
 585 }
 586
 587 static unsigned long fq_codel_bind(struct Qdisc *sch, unsigned long parent,
 588                               u32 classid)
 589 {
 590         /* we cannot bypass queue discipline anymore */
 591         sch->flags &= ~TCQ_F_CAN_BYPASS;
 592         return 0;
 593 }
 594
 595 static void fq_codel_put(struct Qdisc *q, unsigned long cl)
 596 {
 597 }
 598
 599 static struct tcf_proto __rcu **fq_codel_find_tcf(struct Qdisc *sch,
 600                                                   unsigned long cl)
 601 {
 602         struct fq_codel_sched_data *q = qdisc_priv(sch);
 603
 604         if (cl)
 605                 return NULL;
 606         return &q->filter_list;
 607 }
 608
 609 static int fq_codel_dump_class(struct Qdisc *sch, unsigned long cl,
 610                           struct sk_buff *skb, struct tcmsg *tcm)
 611 {
 612         tcm->tcm_handle |= TC_H_MIN(cl);
 613         return 0;
 614 }
 615
 616 static int fq_codel_dump_class_stats(struct Qdisc *sch, unsigned long cl,
 617                                      struct gnet_dump *d)
 618 {
 619         struct fq_codel_sched_data *q = qdisc_priv(sch);
 620         u32 idx = cl - 1;
 621         struct gnet_stats_queue qs = { 0 };
 622         struct tc_fq_codel_xstats xstats;
 623
 624         if (idx < q->flows_cnt) {
 625                 const struct fq_codel_flow *flow = &q->flows[idx];
 626                 const struct sk_buff *skb = flow->head;
 627
 628                 memset(&xstats, 0, sizeof(xstats));
 629                 xstats.type = TCA_FQ_CODEL_XSTATS_CLASS;
 630                 xstats.class_stats.deficit = flow->deficit;
 631                 xstats.class_stats.ldelay =
 632                         codel_time_to_us(flow->cvars.ldelay);
 633                 xstats.class_stats.count = flow->cvars.count;
 634                 xstats.class_stats.lastcount = flow->cvars.lastcount;
 635                 xstats.class_stats.dropping = flow->cvars.dropping;
 636                 if (flow->cvars.dropping) {
 637                         codel_tdiff_t delta = flow->cvars.drop_next -
 638                                               codel_get_time();
 639
 640                         xstats.class_stats.drop_next = (delta >= 0) ?
 641                                 codel_time_to_us(delta) :
 642                                 -codel_time_to_us(-delta);
 643                 }
 644                 while (skb) {
 645                         qs.qlen++;
 646                         skb = skb->next;
 647                 }
 648                 qs.backlog = q->backlogs[idx];
 649                 qs.drops = flow->dropped;
 650         }
 651         if (gnet_stats_copy_queue(d, NULL, &qs, 0) < 0)
 652                 return -1;
 653         if (idx < q->flows_cnt)
 654                 return gnet_stats_copy_app(d, &xstats, sizeof(xstats));
 655         return 0;
 656 }
 657
 658 static void fq_codel_walk(struct Qdisc *sch, struct qdisc_walker *arg)
 659 {
 660         struct fq_codel_sched_data *q = qdisc_priv(sch);
 661         unsigned int i;
 662
 663         if (arg->stop)
 664                 return;
 665
 666         for (i = 0; i < q->flows_cnt; i++) {
 667                 if (list_empty(&q->flows[i].flowchain) ||
 668                     arg->count < arg->skip) {
 669                         arg->count++;
 670                         continue;
 671                 }
 672                 if (arg->fn(sch, i + 1, arg) < 0) {
 673                         arg->stop = 1;
 674                         break;
 675                 }
 676                 arg->count++;
 677         }
 678 }
 679
 680 static const struct Qdisc_class_ops fq_codel_class_ops = {
 681         .leaf           =       fq_codel_leaf,
 682         .get            =       fq_codel_get,
 683         .put            =       fq_codel_put,
 684         .tcf_chain      =       fq_codel_find_tcf,
 685         .bind_tcf       =       fq_codel_bind,
 686         .unbind_tcf     =       fq_codel_put,
 687         .dump           =       fq_codel_dump_class,
 688         .dump_stats     =       fq_codel_dump_class_stats,
 689         .walk           =       fq_codel_walk,
 690 };
 691
 692 static struct Qdisc_ops fq_codel_qdisc_ops __read_mostly = {
 693         .cl_ops         =       &fq_codel_class_ops,
 694         .id             =       "fq_codel",
 695         .priv_size      =       sizeof(struct fq_codel_sched_data),
 696         .enqueue        =       fq_codel_enqueue,
 697         .dequeue        =       fq_codel_dequeue,
 698         .peek           =       qdisc_peek_dequeued,
 699         .drop           =       fq_codel_qdisc_drop,
 700         .init           =       fq_codel_init,
 701         .reset          =       fq_codel_reset,
 702         .destroy        =       fq_codel_destroy,
 703         .change         =       fq_codel_change,
 704         .dump           =       fq_codel_dump,
 705         .dump_stats =   fq_codel_dump_stats,
 706         .owner          =       THIS_MODULE,
 707 };
 708
 709 static int __init fq_codel_module_init(void)
 710 {
 711         return register_qdisc(&fq_codel_qdisc_ops);
 712 }
 713
 714 static void __exit fq_codel_module_exit(void)
 715 {
 716         unregister_qdisc(&fq_codel_qdisc_ops);
 717 }
 718
 719 module_init(fq_codel_module_init)
 720 module_exit(fq_codel_module_exit)
 721 MODULE_AUTHOR("Eric Dumazet");
 722 MODULE_LICENSE("GPL");