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