net/sched/sch_fq_codel.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /*
   3  * Fair Queue CoDel discipline
   4  *
   5  *  Copyright (C) 2012,2015 Eric Dumazet <[email protected]>
   6  */
   7
   8 #include <linux/module.h>
   9 #include <linux/types.h>
  10 #include <linux/kernel.h>
  11 #include <linux/jiffies.h>
  12 #include <linux/string.h>
  13 #include <linux/in.h>
  14 #include <linux/errno.h>
  15 #include <linux/init.h>
  16 #include <linux/skbuff.h>
  17 #include <linux/slab.h>
  18 #include <linux/vmalloc.h>
  19 #include <net/netlink.h>
  20 #include <net/pkt_sched.h>
  21 #include <net/pkt_cls.h>
  22 #include <net/codel.h>
  23 #include <net/codel_impl.h>
  24 #include <net/codel_qdisc.h>
  25
  26 /*      Fair Queue CoDel.
  27  *
  28  * Principles :
  29  * Packets are classified (internal classifier or external) on flows.
  30  * This is a Stochastic model (as we use a hash, several flows
  31  *                             might be hashed on same slot)
  32  * Each flow has a CoDel managed queue.
  33  * Flows are linked onto two (Round Robin) lists,
  34  * so that new flows have priority on old ones.
  35  *
  36  * For a given flow, packets are not reordered (CoDel uses a FIFO)
  37  * head drops only.
  38  * ECN capability is on by default.
  39  * Low memory footprint (64 bytes per flow)
  40  */
  41
  42 struct fq_codel_flow {
  43         struct sk_buff    *head;
  44         struct sk_buff    *tail;
  45         struct list_head  flowchain;
  46         int               deficit;
  47         struct codel_vars cvars;
  48 }; /* please try to keep this structure <= 64 bytes */
  49
  50 struct fq_codel_sched_data {
  51         struct tcf_proto __rcu *filter_list; /* optional external classifier */
  52         struct tcf_block *block;
  53         struct fq_codel_flow *flows;    /* Flows table [flows_cnt] */
  54         u32             *backlogs;      /* backlog table [flows_cnt] */
  55         u32             flows_cnt;      /* number of flows */
  56         u32             quantum;        /* psched_mtu(qdisc_dev(sch)); */
  57         u32             drop_batch_size;
  58         u32             memory_limit;
  59         struct codel_params cparams;
  60         struct codel_stats cstats;
  61         u32             memory_usage;
  62         u32             drop_overmemory;
  63         u32             drop_overlimit;
  64         u32             new_flow_count;
  65
  66         struct list_head new_flows;     /* list of new flows */
  67         struct list_head old_flows;     /* list of old flows */
  68 };
  69
  70 static unsigned int fq_codel_hash(const struct fq_codel_sched_data *q,
  71                                   struct sk_buff *skb)
  72 {
  73         return reciprocal_scale(skb_get_hash(skb), q->flows_cnt);
  74 }
  75
  76 static unsigned int fq_codel_classify(struct sk_buff *skb, struct Qdisc *sch,
  77                                       int *qerr)
  78 {
  79         struct fq_codel_sched_data *q = qdisc_priv(sch);
  80         struct tcf_proto *filter;
  81         struct tcf_result res;
  82         int result;
  83
  84         if (TC_H_MAJ(skb->priority) == sch->handle &&
  85             TC_H_MIN(skb->priority) > 0 &&
  86             TC_H_MIN(skb->priority) <= q->flows_cnt)
  87                 return TC_H_MIN(skb->priority);
  88
  89         filter = rcu_dereference_bh(q->filter_list);
  90         if (!filter)
  91                 return fq_codel_hash(q, skb) + 1;
  92
  93         *qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
  94         result = tcf_classify(skb, NULL, filter, &res, false);
  95         if (result >= 0) {
  96 #ifdef CONFIG_NET_CLS_ACT
  97                 switch (result) {
  98                 case TC_ACT_STOLEN:
  99                 case TC_ACT_QUEUED:
 100                 case TC_ACT_TRAP:
 101                         *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
 102                         fallthrough;
 103                 case TC_ACT_SHOT:
 104                         return 0;
 105                 }
 106 #endif
 107                 if (TC_H_MIN(res.classid) <= q->flows_cnt)
 108                         return TC_H_MIN(res.classid);
 109         }
 110         return 0;
 111 }
 112
 113 /* helper functions : might be changed when/if skb use a standard list_head */
 114
 115 /* remove one skb from head of slot queue */
 116 static inline struct sk_buff *dequeue_head(struct fq_codel_flow *flow)
 117 {
 118         struct sk_buff *skb = flow->head;
 119
 120         flow->head = skb->next;
 121         skb_mark_not_on_list(skb);
 122         return skb;
 123 }
 124
 125 /* add skb to flow queue (tail add) */
 126 static inline void flow_queue_add(struct fq_codel_flow *flow,
 127                                   struct sk_buff *skb)
 128 {
 129         if (flow->head == NULL)
 130                 flow->head = skb;
 131         else
 132                 flow->tail->next = skb;
 133         flow->tail = skb;
 134         skb->next = NULL;
 135 }
 136
 137 static unsigned int fq_codel_drop(struct Qdisc *sch, unsigned int max_packets,
 138                                   struct sk_buff **to_free)
 139 {
 140         struct fq_codel_sched_data *q = qdisc_priv(sch);
 141         struct sk_buff *skb;
 142         unsigned int maxbacklog = 0, idx = 0, i, len;
 143         struct fq_codel_flow *flow;
 144         unsigned int threshold;
 145         unsigned int mem = 0;
 146
 147         /* Queue is full! Find the fat flow and drop packet(s) from it.
 148          * This might sound expensive, but with 1024 flows, we scan
 149          * 4KB of memory, and we dont need to handle a complex tree
 150          * in fast path (packet queue/enqueue) with many cache misses.
 151          * In stress mode, we'll try to drop 64 packets from the flow,
 152          * amortizing this linear lookup to one cache line per drop.
 153          */
 154         for (i = 0; i < q->flows_cnt; i++) {
 155                 if (q->backlogs[i] > maxbacklog) {
 156                         maxbacklog = q->backlogs[i];
 157                         idx = i;
 158                 }
 159         }
 160
 161         /* Our goal is to drop half of this fat flow backlog */
 162         threshold = maxbacklog >> 1;
 163
 164         flow = &q->flows[idx];
 165         len = 0;
 166         i = 0;
 167         do {
 168                 skb = dequeue_head(flow);
 169                 len += qdisc_pkt_len(skb);
 170                 mem += get_codel_cb(skb)->mem_usage;
 171                 tcf_set_drop_reason(skb, SKB_DROP_REASON_QDISC_OVERLIMIT);
 172                 __qdisc_drop(skb, to_free);
 173         } while (++i < max_packets && len < threshold);
 174
 175         /* Tell codel to increase its signal strength also */
 176         flow->cvars.count += i;
 177         q->backlogs[idx] -= len;
 178         q->memory_usage -= mem;
 179         sch->qstats.drops += i;
 180         sch->qstats.backlog -= len;
 181         sch->q.qlen -= i;
 182         return idx;
 183 }
 184
 185 static int fq_codel_enqueue(struct sk_buff *skb, struct Qdisc *sch,
 186                             struct sk_buff **to_free)
 187 {
 188         struct fq_codel_sched_data *q = qdisc_priv(sch);
 189         unsigned int idx, prev_backlog, prev_qlen;
 190         struct fq_codel_flow *flow;
 191         int ret;
 192         unsigned int pkt_len;
 193         bool memory_limited;
 194
 195         idx = fq_codel_classify(skb, sch, &ret);
 196         if (idx == 0) {
 197                 if (ret & __NET_XMIT_BYPASS)
 198                         qdisc_qstats_drop(sch);
 199                 __qdisc_drop(skb, to_free);
 200                 return ret;
 201         }
 202         idx--;
 203
 204         codel_set_enqueue_time(skb);
 205         flow = &q->flows[idx];
 206         flow_queue_add(flow, skb);
 207         q->backlogs[idx] += qdisc_pkt_len(skb);
 208         qdisc_qstats_backlog_inc(sch, skb);
 209
 210         if (list_empty(&flow->flowchain)) {
 211                 list_add_tail(&flow->flowchain, &q->new_flows);
 212                 q->new_flow_count++;
 213                 flow->deficit = q->quantum;
 214         }
 215         get_codel_cb(skb)->mem_usage = skb->truesize;
 216         q->memory_usage += get_codel_cb(skb)->mem_usage;
 217         memory_limited = q->memory_usage > q->memory_limit;
 218         if (++sch->q.qlen <= sch->limit && !memory_limited)
 219                 return NET_XMIT_SUCCESS;
 220
 221         prev_backlog = sch->qstats.backlog;
 222         prev_qlen = sch->q.qlen;
 223
 224         /* save this packet length as it might be dropped by fq_codel_drop() */
 225         pkt_len = qdisc_pkt_len(skb);
 226         /* fq_codel_drop() is quite expensive, as it performs a linear search
 227          * in q->backlogs[] to find a fat flow.
 228          * So instead of dropping a single packet, drop half of its backlog
 229          * with a 64 packets limit to not add a too big cpu spike here.
 230          */
 231         ret = fq_codel_drop(sch, q->drop_batch_size, to_free);
 232
 233         prev_qlen -= sch->q.qlen;
 234         prev_backlog -= sch->qstats.backlog;
 235         q->drop_overlimit += prev_qlen;
 236         if (memory_limited)
 237                 q->drop_overmemory += prev_qlen;
 238
 239         /* As we dropped packet(s), better let upper stack know this.
 240          * If we dropped a packet for this flow, return NET_XMIT_CN,
 241          * but in this case, our parents wont increase their backlogs.
 242          */
 243         if (ret == idx) {
 244                 qdisc_tree_reduce_backlog(sch, prev_qlen - 1,
 245                                           prev_backlog - pkt_len);
 246                 return NET_XMIT_CN;
 247         }
 248         qdisc_tree_reduce_backlog(sch, prev_qlen, prev_backlog);
 249         return NET_XMIT_SUCCESS;
 250 }
 251
 252 /* This is the specific function called from codel_dequeue()
 253  * to dequeue a packet from queue. Note: backlog is handled in
 254  * codel, we dont need to reduce it here.
 255  */
 256 static struct sk_buff *dequeue_func(struct codel_vars *vars, void *ctx)
 257 {
 258         struct Qdisc *sch = ctx;
 259         struct fq_codel_sched_data *q = qdisc_priv(sch);
 260         struct fq_codel_flow *flow;
 261         struct sk_buff *skb = NULL;
 262
 263         flow = container_of(vars, struct fq_codel_flow, cvars);
 264         if (flow->head) {
 265                 skb = dequeue_head(flow);
 266                 q->backlogs[flow - q->flows] -= qdisc_pkt_len(skb);
 267                 q->memory_usage -= get_codel_cb(skb)->mem_usage;
 268                 sch->q.qlen--;
 269                 sch->qstats.backlog -= qdisc_pkt_len(skb);
 270         }
 271         return skb;
 272 }
 273
 274 static void drop_func(struct sk_buff *skb, void *ctx)
 275 {
 276         struct Qdisc *sch = ctx;
 277
 278         kfree_skb_reason(skb, SKB_DROP_REASON_QDISC_CONGESTED);
 279         qdisc_qstats_drop(sch);
 280 }
 281
 282 static struct sk_buff *fq_codel_dequeue(struct Qdisc *sch)
 283 {
 284         struct fq_codel_sched_data *q = qdisc_priv(sch);
 285         struct sk_buff *skb;
 286         struct fq_codel_flow *flow;
 287         struct list_head *head;
 288
 289 begin:
 290         head = &q->new_flows;
 291         if (list_empty(head)) {
 292                 head = &q->old_flows;
 293                 if (list_empty(head))
 294                         return NULL;
 295         }
 296         flow = list_first_entry(head, struct fq_codel_flow, flowchain);
 297
 298         if (flow->deficit <= 0) {
 299                 flow->deficit += q->quantum;
 300                 list_move_tail(&flow->flowchain, &q->old_flows);
 301                 goto begin;
 302         }
 303
 304         skb = codel_dequeue(sch, &sch->qstats.backlog, &q->cparams,
 305                             &flow->cvars, &q->cstats, qdisc_pkt_len,
 306                             codel_get_enqueue_time, drop_func, dequeue_func);
 307
 308         if (!skb) {
 309                 /* force a pass through old_flows to prevent starvation */
 310                 if ((head == &q->new_flows) && !list_empty(&q->old_flows))
 311                         list_move_tail(&flow->flowchain, &q->old_flows);
 312                 else
 313                         list_del_init(&flow->flowchain);
 314                 goto begin;
 315         }
 316         qdisc_bstats_update(sch, skb);
 317         flow->deficit -= qdisc_pkt_len(skb);
 318         /* We cant call qdisc_tree_reduce_backlog() if our qlen is 0,
 319          * or HTB crashes. Defer it for next round.
 320          */
 321         if (q->cstats.drop_count && sch->q.qlen) {
 322                 qdisc_tree_reduce_backlog(sch, q->cstats.drop_count,
 323                                           q->cstats.drop_len);
 324                 q->cstats.drop_count = 0;
 325                 q->cstats.drop_len = 0;
 326         }
 327         return skb;
 328 }
 329
 330 static void fq_codel_flow_purge(struct fq_codel_flow *flow)
 331 {
 332         rtnl_kfree_skbs(flow->head, flow->tail);
 333         flow->head = NULL;
 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                 fq_codel_flow_purge(flow);
 347                 INIT_LIST_HEAD(&flow->flowchain);
 348                 codel_vars_init(&flow->cvars);
 349         }
 350         memset(q->backlogs, 0, q->flows_cnt * sizeof(u32));
 351         q->memory_usage = 0;
 352 }
 353
 354 static const struct nla_policy fq_codel_policy[TCA_FQ_CODEL_MAX + 1] = {
 355         [TCA_FQ_CODEL_TARGET]   = { .type = NLA_U32 },
 356         [TCA_FQ_CODEL_LIMIT]    = { .type = NLA_U32 },
 357         [TCA_FQ_CODEL_INTERVAL] = { .type = NLA_U32 },
 358         [TCA_FQ_CODEL_ECN]      = { .type = NLA_U32 },
 359         [TCA_FQ_CODEL_FLOWS]    = { .type = NLA_U32 },
 360         [TCA_FQ_CODEL_QUANTUM]  = { .type = NLA_U32 },
 361         [TCA_FQ_CODEL_CE_THRESHOLD] = { .type = NLA_U32 },
 362         [TCA_FQ_CODEL_DROP_BATCH_SIZE] = { .type = NLA_U32 },
 363         [TCA_FQ_CODEL_MEMORY_LIMIT] = { .type = NLA_U32 },
 364         [TCA_FQ_CODEL_CE_THRESHOLD_SELECTOR] = { .type = NLA_U8 },
 365         [TCA_FQ_CODEL_CE_THRESHOLD_MASK] = { .type = NLA_U8 },
 366 };
 367
 368 static int fq_codel_change(struct Qdisc *sch, struct nlattr *opt,
 369                            struct netlink_ext_ack *extack)
 370 {
 371         struct fq_codel_sched_data *q = qdisc_priv(sch);
 372         struct nlattr *tb[TCA_FQ_CODEL_MAX + 1];
 373         u32 quantum = 0;
 374         int err;
 375
 376         err = nla_parse_nested_deprecated(tb, TCA_FQ_CODEL_MAX, opt,
 377                                           fq_codel_policy, NULL);
 378         if (err < 0)
 379                 return err;
 380         if (tb[TCA_FQ_CODEL_FLOWS]) {
 381                 if (q->flows)
 382                         return -EINVAL;
 383                 q->flows_cnt = nla_get_u32(tb[TCA_FQ_CODEL_FLOWS]);
 384                 if (!q->flows_cnt ||
 385                     q->flows_cnt > 65536)
 386                         return -EINVAL;
 387         }
 388         if (tb[TCA_FQ_CODEL_QUANTUM]) {
 389                 quantum = max(256U, nla_get_u32(tb[TCA_FQ_CODEL_QUANTUM]));
 390                 if (quantum > FQ_CODEL_QUANTUM_MAX) {
 391                         NL_SET_ERR_MSG(extack, "Invalid quantum");
 392                         return -EINVAL;
 393                 }
 394         }
 395         sch_tree_lock(sch);
 396
 397         if (tb[TCA_FQ_CODEL_TARGET]) {
 398                 u64 target = nla_get_u32(tb[TCA_FQ_CODEL_TARGET]);
 399
 400                 WRITE_ONCE(q->cparams.target,
 401                            (target * NSEC_PER_USEC) >> CODEL_SHIFT);
 402         }
 403
 404         if (tb[TCA_FQ_CODEL_CE_THRESHOLD]) {
 405                 u64 val = nla_get_u32(tb[TCA_FQ_CODEL_CE_THRESHOLD]);
 406
 407                 WRITE_ONCE(q->cparams.ce_threshold,
 408                            (val * NSEC_PER_USEC) >> CODEL_SHIFT);
 409         }
 410
 411         if (tb[TCA_FQ_CODEL_CE_THRESHOLD_SELECTOR])
 412                 WRITE_ONCE(q->cparams.ce_threshold_selector,
 413                            nla_get_u8(tb[TCA_FQ_CODEL_CE_THRESHOLD_SELECTOR]));
 414         if (tb[TCA_FQ_CODEL_CE_THRESHOLD_MASK])
 415                 WRITE_ONCE(q->cparams.ce_threshold_mask,
 416                            nla_get_u8(tb[TCA_FQ_CODEL_CE_THRESHOLD_MASK]));
 417
 418         if (tb[TCA_FQ_CODEL_INTERVAL]) {
 419                 u64 interval = nla_get_u32(tb[TCA_FQ_CODEL_INTERVAL]);
 420
 421                 WRITE_ONCE(q->cparams.interval,
 422                            (interval * NSEC_PER_USEC) >> CODEL_SHIFT);
 423         }
 424
 425         if (tb[TCA_FQ_CODEL_LIMIT])
 426                 WRITE_ONCE(sch->limit,
 427                            nla_get_u32(tb[TCA_FQ_CODEL_LIMIT]));
 428
 429         if (tb[TCA_FQ_CODEL_ECN])
 430                 WRITE_ONCE(q->cparams.ecn,
 431                            !!nla_get_u32(tb[TCA_FQ_CODEL_ECN]));
 432
 433         if (quantum)
 434                 WRITE_ONCE(q->quantum, quantum);
 435
 436         if (tb[TCA_FQ_CODEL_DROP_BATCH_SIZE])
 437                 WRITE_ONCE(q->drop_batch_size,
 438                            max(1U, nla_get_u32(tb[TCA_FQ_CODEL_DROP_BATCH_SIZE])));
 439
 440         if (tb[TCA_FQ_CODEL_MEMORY_LIMIT])
 441                 WRITE_ONCE(q->memory_limit,
 442                            min(1U << 31, nla_get_u32(tb[TCA_FQ_CODEL_MEMORY_LIMIT])));
 443
 444         while (sch->q.qlen > sch->limit ||
 445                q->memory_usage > q->memory_limit) {
 446                 struct sk_buff *skb = fq_codel_dequeue(sch);
 447
 448                 q->cstats.drop_len += qdisc_pkt_len(skb);
 449                 rtnl_kfree_skbs(skb, skb);
 450                 q->cstats.drop_count++;
 451         }
 452         qdisc_tree_reduce_backlog(sch, q->cstats.drop_count, q->cstats.drop_len);
 453         q->cstats.drop_count = 0;
 454         q->cstats.drop_len = 0;
 455
 456         sch_tree_unlock(sch);
 457         return 0;
 458 }
 459
 460 static void fq_codel_destroy(struct Qdisc *sch)
 461 {
 462         struct fq_codel_sched_data *q = qdisc_priv(sch);
 463
 464         tcf_block_put(q->block);
 465         kvfree(q->backlogs);
 466         kvfree(q->flows);
 467 }
 468
 469 static int fq_codel_init(struct Qdisc *sch, struct nlattr *opt,
 470                          struct netlink_ext_ack *extack)
 471 {
 472         struct fq_codel_sched_data *q = qdisc_priv(sch);
 473         int i;
 474         int err;
 475
 476         sch->limit = 10*1024;
 477         q->flows_cnt = 1024;
 478         q->memory_limit = 32 << 20; /* 32 MBytes */
 479         q->drop_batch_size = 64;
 480         q->quantum = psched_mtu(qdisc_dev(sch));
 481         INIT_LIST_HEAD(&q->new_flows);
 482         INIT_LIST_HEAD(&q->old_flows);
 483         codel_params_init(&q->cparams);
 484         codel_stats_init(&q->cstats);
 485         q->cparams.ecn = true;
 486         q->cparams.mtu = psched_mtu(qdisc_dev(sch));
 487
 488         if (opt) {
 489                 err = fq_codel_change(sch, opt, extack);
 490                 if (err)
 491                         goto init_failure;
 492         }
 493
 494         err = tcf_block_get(&q->block, &q->filter_list, sch, extack);
 495         if (err)
 496                 goto init_failure;
 497
 498         if (!q->flows) {
 499                 q->flows = kvcalloc(q->flows_cnt,
 500                                     sizeof(struct fq_codel_flow),
 501                                     GFP_KERNEL);
 502                 if (!q->flows) {
 503                         err = -ENOMEM;
 504                         goto init_failure;
 505                 }
 506                 q->backlogs = kvcalloc(q->flows_cnt, sizeof(u32), GFP_KERNEL);
 507                 if (!q->backlogs) {
 508                         err = -ENOMEM;
 509                         goto alloc_failure;
 510                 }
 511                 for (i = 0; i < q->flows_cnt; i++) {
 512                         struct fq_codel_flow *flow = q->flows + i;
 513
 514                         INIT_LIST_HEAD(&flow->flowchain);
 515                         codel_vars_init(&flow->cvars);
 516                 }
 517         }
 518         if (sch->limit >= 1)
 519                 sch->flags |= TCQ_F_CAN_BYPASS;
 520         else
 521                 sch->flags &= ~TCQ_F_CAN_BYPASS;
 522         return 0;
 523
 524 alloc_failure:
 525         kvfree(q->flows);
 526         q->flows = NULL;
 527 init_failure:
 528         q->flows_cnt = 0;
 529         return err;
 530 }
 531
 532 static int fq_codel_dump(struct Qdisc *sch, struct sk_buff *skb)
 533 {
 534         struct fq_codel_sched_data *q = qdisc_priv(sch);
 535         codel_time_t ce_threshold;
 536         struct nlattr *opts;
 537
 538         opts = nla_nest_start_noflag(skb, TCA_OPTIONS);
 539         if (opts == NULL)
 540                 goto nla_put_failure;
 541
 542         if (nla_put_u32(skb, TCA_FQ_CODEL_TARGET,
 543                         codel_time_to_us(READ_ONCE(q->cparams.target))) ||
 544             nla_put_u32(skb, TCA_FQ_CODEL_LIMIT,
 545                         READ_ONCE(sch->limit)) ||
 546             nla_put_u32(skb, TCA_FQ_CODEL_INTERVAL,
 547                         codel_time_to_us(READ_ONCE(q->cparams.interval))) ||
 548             nla_put_u32(skb, TCA_FQ_CODEL_ECN,
 549                         READ_ONCE(q->cparams.ecn)) ||
 550             nla_put_u32(skb, TCA_FQ_CODEL_QUANTUM,
 551                         READ_ONCE(q->quantum)) ||
 552             nla_put_u32(skb, TCA_FQ_CODEL_DROP_BATCH_SIZE,
 553                         READ_ONCE(q->drop_batch_size)) ||
 554             nla_put_u32(skb, TCA_FQ_CODEL_MEMORY_LIMIT,
 555                         READ_ONCE(q->memory_limit)) ||
 556             nla_put_u32(skb, TCA_FQ_CODEL_FLOWS,
 557                         READ_ONCE(q->flows_cnt)))
 558                 goto nla_put_failure;
 559
 560         ce_threshold = READ_ONCE(q->cparams.ce_threshold);
 561         if (ce_threshold != CODEL_DISABLED_THRESHOLD) {
 562                 if (nla_put_u32(skb, TCA_FQ_CODEL_CE_THRESHOLD,
 563                                 codel_time_to_us(ce_threshold)))
 564                         goto nla_put_failure;
 565                 if (nla_put_u8(skb, TCA_FQ_CODEL_CE_THRESHOLD_SELECTOR,
 566                                READ_ONCE(q->cparams.ce_threshold_selector)))
 567                         goto nla_put_failure;
 568                 if (nla_put_u8(skb, TCA_FQ_CODEL_CE_THRESHOLD_MASK,
 569                                READ_ONCE(q->cparams.ce_threshold_mask)))
 570                         goto nla_put_failure;
 571         }
 572
 573         return nla_nest_end(skb, opts);
 574
 575 nla_put_failure:
 576         return -1;
 577 }
 578
 579 static int fq_codel_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
 580 {
 581         struct fq_codel_sched_data *q = qdisc_priv(sch);
 582         struct tc_fq_codel_xstats st = {
 583                 .type                           = TCA_FQ_CODEL_XSTATS_QDISC,
 584         };
 585         struct list_head *pos;
 586
 587         st.qdisc_stats.maxpacket = q->cstats.maxpacket;
 588         st.qdisc_stats.drop_overlimit = q->drop_overlimit;
 589         st.qdisc_stats.ecn_mark = q->cstats.ecn_mark;
 590         st.qdisc_stats.new_flow_count = q->new_flow_count;
 591         st.qdisc_stats.ce_mark = q->cstats.ce_mark;
 592         st.qdisc_stats.memory_usage  = q->memory_usage;
 593         st.qdisc_stats.drop_overmemory = q->drop_overmemory;
 594
 595         sch_tree_lock(sch);
 596         list_for_each(pos, &q->new_flows)
 597                 st.qdisc_stats.new_flows_len++;
 598
 599         list_for_each(pos, &q->old_flows)
 600                 st.qdisc_stats.old_flows_len++;
 601         sch_tree_unlock(sch);
 602
 603         return gnet_stats_copy_app(d, &st, sizeof(st));
 604 }
 605
 606 static struct Qdisc *fq_codel_leaf(struct Qdisc *sch, unsigned long arg)
 607 {
 608         return NULL;
 609 }
 610
 611 static unsigned long fq_codel_find(struct Qdisc *sch, u32 classid)
 612 {
 613         return 0;
 614 }
 615
 616 static unsigned long fq_codel_bind(struct Qdisc *sch, unsigned long parent,
 617                               u32 classid)
 618 {
 619         return 0;
 620 }
 621
 622 static void fq_codel_unbind(struct Qdisc *q, unsigned long cl)
 623 {
 624 }
 625
 626 static struct tcf_block *fq_codel_tcf_block(struct Qdisc *sch, unsigned long cl,
 627                                             struct netlink_ext_ack *extack)
 628 {
 629         struct fq_codel_sched_data *q = qdisc_priv(sch);
 630
 631         if (cl)
 632                 return NULL;
 633         return q->block;
 634 }
 635
 636 static int fq_codel_dump_class(struct Qdisc *sch, unsigned long cl,
 637                           struct sk_buff *skb, struct tcmsg *tcm)
 638 {
 639         tcm->tcm_handle |= TC_H_MIN(cl);
 640         return 0;
 641 }
 642
 643 static int fq_codel_dump_class_stats(struct Qdisc *sch, unsigned long cl,
 644                                      struct gnet_dump *d)
 645 {
 646         struct fq_codel_sched_data *q = qdisc_priv(sch);
 647         u32 idx = cl - 1;
 648         struct gnet_stats_queue qs = { 0 };
 649         struct tc_fq_codel_xstats xstats;
 650
 651         if (idx < q->flows_cnt) {
 652                 const struct fq_codel_flow *flow = &q->flows[idx];
 653                 const struct sk_buff *skb;
 654
 655                 memset(&xstats, 0, sizeof(xstats));
 656                 xstats.type = TCA_FQ_CODEL_XSTATS_CLASS;
 657                 xstats.class_stats.deficit = flow->deficit;
 658                 xstats.class_stats.ldelay =
 659                         codel_time_to_us(flow->cvars.ldelay);
 660                 xstats.class_stats.count = flow->cvars.count;
 661                 xstats.class_stats.lastcount = flow->cvars.lastcount;
 662                 xstats.class_stats.dropping = flow->cvars.dropping;
 663                 if (flow->cvars.dropping) {
 664                         codel_tdiff_t delta = flow->cvars.drop_next -
 665                                               codel_get_time();
 666
 667                         xstats.class_stats.drop_next = (delta >= 0) ?
 668                                 codel_time_to_us(delta) :
 669                                 -codel_time_to_us(-delta);
 670                 }
 671                 if (flow->head) {
 672                         sch_tree_lock(sch);
 673                         skb = flow->head;
 674                         while (skb) {
 675                                 qs.qlen++;
 676                                 skb = skb->next;
 677                         }
 678                         sch_tree_unlock(sch);
 679                 }
 680                 qs.backlog = q->backlogs[idx];
 681                 qs.drops = 0;
 682         }
 683         if (gnet_stats_copy_queue(d, NULL, &qs, qs.qlen) < 0)
 684                 return -1;
 685         if (idx < q->flows_cnt)
 686                 return gnet_stats_copy_app(d, &xstats, sizeof(xstats));
 687         return 0;
 688 }
 689
 690 static void fq_codel_walk(struct Qdisc *sch, struct qdisc_walker *arg)
 691 {
 692         struct fq_codel_sched_data *q = qdisc_priv(sch);
 693         unsigned int i;
 694
 695         if (arg->stop)
 696                 return;
 697
 698         for (i = 0; i < q->flows_cnt; i++) {
 699                 if (list_empty(&q->flows[i].flowchain)) {
 700                         arg->count++;
 701                         continue;
 702                 }
 703                 if (!tc_qdisc_stats_dump(sch, i + 1, arg))
 704                         break;
 705         }
 706 }
 707
 708 static const struct Qdisc_class_ops fq_codel_class_ops = {
 709         .leaf           =       fq_codel_leaf,
 710         .find           =       fq_codel_find,
 711         .tcf_block      =       fq_codel_tcf_block,
 712         .bind_tcf       =       fq_codel_bind,
 713         .unbind_tcf     =       fq_codel_unbind,
 714         .dump           =       fq_codel_dump_class,
 715         .dump_stats     =       fq_codel_dump_class_stats,
 716         .walk           =       fq_codel_walk,
 717 };
 718
 719 static struct Qdisc_ops fq_codel_qdisc_ops __read_mostly = {
 720         .cl_ops         =       &fq_codel_class_ops,
 721         .id             =       "fq_codel",
 722         .priv_size      =       sizeof(struct fq_codel_sched_data),
 723         .enqueue        =       fq_codel_enqueue,
 724         .dequeue        =       fq_codel_dequeue,
 725         .peek           =       qdisc_peek_dequeued,
 726         .init           =       fq_codel_init,
 727         .reset          =       fq_codel_reset,
 728         .destroy        =       fq_codel_destroy,
 729         .change         =       fq_codel_change,
 730         .dump           =       fq_codel_dump,
 731         .dump_stats =   fq_codel_dump_stats,
 732         .owner          =       THIS_MODULE,
 733 };
 734 MODULE_ALIAS_NET_SCH("fq_codel");
 735
 736 static int __init fq_codel_module_init(void)
 737 {
 738         return register_qdisc(&fq_codel_qdisc_ops);
 739 }
 740
 741 static void __exit fq_codel_module_exit(void)
 742 {
 743         unregister_qdisc(&fq_codel_qdisc_ops);
 744 }
 745
 746 module_init(fq_codel_module_init)
 747 module_exit(fq_codel_module_exit)
 748 MODULE_AUTHOR("Eric Dumazet");
 749 MODULE_LICENSE("GPL");
 750 MODULE_DESCRIPTION("Fair Queue CoDel discipline");