net/sched/sch_fq.c

   1 /*
   2  * net/sched/sch_fq.c Fair Queue Packet Scheduler (per flow pacing)
   3  *
   4  *  Copyright (C) 2013 Eric Dumazet <edumazet@google.com>
   5  *
   6  *      This program is free software; you can redistribute it and/or
   7  *      modify it under the terms of the GNU General Public License
   8  *      as published by the Free Software Foundation; either version
   9  *      2 of the License, or (at your option) any later version.
  10  *
  11  *  Meant to be mostly used for localy generated traffic :
  12  *  Fast classification depends on skb->sk being set before reaching us.
  13  *  If not, (router workload), we use rxhash as fallback, with 32 bits wide hash.
  14  *  All packets belonging to a socket are considered as a 'flow'.
  15  *
  16  *  Flows are dynamically allocated and stored in a hash table of RB trees
  17  *  They are also part of one Round Robin 'queues' (new or old flows)
  18  *
  19  *  Burst avoidance (aka pacing) capability :
  20  *
  21  *  Transport (eg TCP) can set in sk->sk_pacing_rate a rate, enqueue a
  22  *  bunch of packets, and this packet scheduler adds delay between
  23  *  packets to respect rate limitation.
  24  *
  25  *  enqueue() :
  26  *   - lookup one RB tree (out of 1024 or more) to find the flow.
  27  *     If non existent flow, create it, add it to the tree.
  28  *     Add skb to the per flow list of skb (fifo).
  29  *   - Use a special fifo for high prio packets
  30  *
  31  *  dequeue() : serves flows in Round Robin
  32  *  Note : When a flow becomes empty, we do not immediately remove it from
  33  *  rb trees, for performance reasons (its expected to send additional packets,
  34  *  or SLAB cache will reuse socket for another flow)
  35  */
  36
  37 #include <linux/module.h>
  38 #include <linux/types.h>
  39 #include <linux/kernel.h>
  40 #include <linux/jiffies.h>
  41 #include <linux/string.h>
  42 #include <linux/in.h>
  43 #include <linux/errno.h>
  44 #include <linux/init.h>
  45 #include <linux/skbuff.h>
  46 #include <linux/slab.h>
  47 #include <linux/rbtree.h>
  48 #include <linux/hash.h>
  49 #include <linux/prefetch.h>
  50 #include <net/netlink.h>
  51 #include <net/pkt_sched.h>
  52 #include <net/sock.h>
  53 #include <net/tcp_states.h>
  54
  55 /*
  56  * Per flow structure, dynamically allocated
  57  */
  58 struct fq_flow {
  59         struct sk_buff  *head;          /* list of skbs for this flow : first skb */
  60         union {
  61                 struct sk_buff *tail;   /* last skb in the list */
  62                 unsigned long  age;     /* jiffies when flow was emptied, for gc */
  63         };
  64         struct rb_node  fq_node;        /* anchor in fq_root[] trees */
  65         struct sock     *sk;
  66         int             qlen;           /* number of packets in flow queue */
  67         int             credit;
  68         u32             socket_hash;    /* sk_hash */
  69         struct fq_flow *next;           /* next pointer in RR lists, or &detached */
  70
  71         struct rb_node  rate_node;      /* anchor in q->delayed tree */
  72         u64             time_next_packet;
  73 };
  74
  75 struct fq_flow_head {
  76         struct fq_flow *first;
  77         struct fq_flow *last;
  78 };
  79
  80 struct fq_sched_data {
  81         struct fq_flow_head new_flows;
  82
  83         struct fq_flow_head old_flows;
  84
  85         struct rb_root  delayed;        /* for rate limited flows */
  86         u64             time_next_delayed_flow;
  87
  88         struct fq_flow  internal;       /* for non classified or high prio packets */
  89         u32             quantum;
  90         u32             initial_quantum;
  91         u32             flow_default_rate;/* rate per flow : bytes per second */
  92         u32             flow_max_rate;  /* optional max rate per flow */
  93         u32             flow_plimit;    /* max packets per flow */
  94         struct rb_root  *fq_root;
  95         u8              rate_enable;
  96         u8              fq_trees_log;
  97
  98         u32             flows;
  99         u32             inactive_flows;
 100         u32             throttled_flows;
 101
 102         u64             stat_gc_flows;
 103         u64             stat_internal_packets;
 104         u64             stat_tcp_retrans;
 105         u64             stat_throttled;
 106         u64             stat_flows_plimit;
 107         u64             stat_pkts_too_long;
 108         u64             stat_allocation_errors;
 109         struct qdisc_watchdog watchdog;
 110 };
 111
 112 /* special value to mark a detached flow (not on old/new list) */
 113 static struct fq_flow detached, throttled;
 114
 115 static void fq_flow_set_detached(struct fq_flow *f)
 116 {
 117         f->next = &detached;
 118 }
 119
 120 static bool fq_flow_is_detached(const struct fq_flow *f)
 121 {
 122         return f->next == &detached;
 123 }
 124
 125 static void fq_flow_set_throttled(struct fq_sched_data *q, struct fq_flow *f)
 126 {
 127         struct rb_node **p = &q->delayed.rb_node, *parent = NULL;
 128
 129         while (*p) {
 130                 struct fq_flow *aux;
 131
 132                 parent = *p;
 133                 aux = container_of(parent, struct fq_flow, rate_node);
 134                 if (f->time_next_packet >= aux->time_next_packet)
 135                         p = &parent->rb_right;
 136                 else
 137                         p = &parent->rb_left;
 138         }
 139         rb_link_node(&f->rate_node, parent, p);
 140         rb_insert_color(&f->rate_node, &q->delayed);
 141         q->throttled_flows++;
 142         q->stat_throttled++;
 143
 144         f->next = &throttled;
 145         if (q->time_next_delayed_flow > f->time_next_packet)
 146                 q->time_next_delayed_flow = f->time_next_packet;
 147 }
 148
 149
 150 static struct kmem_cache *fq_flow_cachep __read_mostly;
 151
 152 static void fq_flow_add_tail(struct fq_flow_head *head, struct fq_flow *flow)
 153 {
 154         if (head->first)
 155                 head->last->next = flow;
 156         else
 157                 head->first = flow;
 158         head->last = flow;
 159         flow->next = NULL;
 160 }
 161
 162 /* limit number of collected flows per round */
 163 #define FQ_GC_MAX 8
 164 #define FQ_GC_AGE (3*HZ)
 165
 166 static bool fq_gc_candidate(const struct fq_flow *f)
 167 {
 168         return fq_flow_is_detached(f) &&
 169                time_after(jiffies, f->age + FQ_GC_AGE);
 170 }
 171
 172 static void fq_gc(struct fq_sched_data *q,
 173                   struct rb_root *root,
 174                   struct sock *sk)
 175 {
 176         struct fq_flow *f, *tofree[FQ_GC_MAX];
 177         struct rb_node **p, *parent;
 178         int fcnt = 0;
 179
 180         p = &root->rb_node;
 181         parent = NULL;
 182         while (*p) {
 183                 parent = *p;
 184
 185                 f = container_of(parent, struct fq_flow, fq_node);
 186                 if (f->sk == sk)
 187                         break;
 188
 189                 if (fq_gc_candidate(f)) {
 190                         tofree[fcnt++] = f;
 191                         if (fcnt == FQ_GC_MAX)
 192                                 break;
 193                 }
 194
 195                 if (f->sk > sk)
 196                         p = &parent->rb_right;
 197                 else
 198                         p = &parent->rb_left;
 199         }
 200
 201         q->flows -= fcnt;
 202         q->inactive_flows -= fcnt;
 203         q->stat_gc_flows += fcnt;
 204         while (fcnt) {
 205                 struct fq_flow *f = tofree[--fcnt];
 206
 207                 rb_erase(&f->fq_node, root);
 208                 kmem_cache_free(fq_flow_cachep, f);
 209         }
 210 }
 211
 212 static const u8 prio2band[TC_PRIO_MAX + 1] = {
 213         1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1
 214 };
 215
 216 static struct fq_flow *fq_classify(struct sk_buff *skb, struct fq_sched_data *q)
 217 {
 218         struct rb_node **p, *parent;
 219         struct sock *sk = skb->sk;
 220         struct rb_root *root;
 221         struct fq_flow *f;
 222         int band;
 223
 224         /* warning: no starvation prevention... */
 225         band = prio2band[skb->priority & TC_PRIO_MAX];
 226         if (unlikely(band == 0))
 227                 return &q->internal;
 228
 229         if (unlikely(!sk)) {
 230                 /* By forcing low order bit to 1, we make sure to not
 231                  * collide with a local flow (socket pointers are word aligned)
 232                  */
 233                 sk = (struct sock *)(skb_get_rxhash(skb) | 1L);
 234         }
 235
 236         root = &q->fq_root[hash_32((u32)(long)sk, q->fq_trees_log)];
 237
 238         if (q->flows >= (2U << q->fq_trees_log) &&
 239             q->inactive_flows > q->flows/2)
 240                 fq_gc(q, root, sk);
 241
 242         p = &root->rb_node;
 243         parent = NULL;
 244         while (*p) {
 245                 parent = *p;
 246
 247                 f = container_of(parent, struct fq_flow, fq_node);
 248                 if (f->sk == sk) {
 249                         /* socket might have been reallocated, so check
 250                          * if its sk_hash is the same.
 251                          * It not, we need to refill credit with
 252                          * initial quantum
 253                          */
 254                         if (unlikely(skb->sk &&
 255                                      f->socket_hash != sk->sk_hash)) {
 256                                 f->credit = q->initial_quantum;
 257                                 f->socket_hash = sk->sk_hash;
 258                                 f->time_next_packet = 0ULL;
 259                         }
 260                         return f;
 261                 }
 262                 if (f->sk > sk)
 263                         p = &parent->rb_right;
 264                 else
 265                         p = &parent->rb_left;
 266         }
 267
 268         f = kmem_cache_zalloc(fq_flow_cachep, GFP_ATOMIC | __GFP_NOWARN);
 269         if (unlikely(!f)) {
 270                 q->stat_allocation_errors++;
 271                 return &q->internal;
 272         }
 273         fq_flow_set_detached(f);
 274         f->sk = sk;
 275         if (skb->sk)
 276                 f->socket_hash = sk->sk_hash;
 277         f->credit = q->initial_quantum;
 278
 279         rb_link_node(&f->fq_node, parent, p);
 280         rb_insert_color(&f->fq_node, root);
 281
 282         q->flows++;
 283         q->inactive_flows++;
 284         return f;
 285 }
 286
 287
 288 /* remove one skb from head of flow queue */
 289 static struct sk_buff *fq_dequeue_head(struct Qdisc *sch, struct fq_flow *flow)
 290 {
 291         struct sk_buff *skb = flow->head;
 292
 293         if (skb) {
 294                 flow->head = skb->next;
 295                 skb->next = NULL;
 296                 flow->qlen--;
 297                 sch->qstats.backlog -= qdisc_pkt_len(skb);
 298                 sch->q.qlen--;
 299         }
 300         return skb;
 301 }
 302
 303 /* We might add in the future detection of retransmits
 304  * For the time being, just return false
 305  */
 306 static bool skb_is_retransmit(struct sk_buff *skb)
 307 {
 308         return false;
 309 }
 310
 311 /* add skb to flow queue
 312  * flow queue is a linked list, kind of FIFO, except for TCP retransmits
 313  * We special case tcp retransmits to be transmitted before other packets.
 314  * We rely on fact that TCP retransmits are unlikely, so we do not waste
 315  * a separate queue or a pointer.
 316  * head->  [retrans pkt 1]
 317  *         [retrans pkt 2]
 318  *         [ normal pkt 1]
 319  *         [ normal pkt 2]
 320  *         [ normal pkt 3]
 321  * tail->  [ normal pkt 4]
 322  */
 323 static void flow_queue_add(struct fq_flow *flow, struct sk_buff *skb)
 324 {
 325         struct sk_buff *prev, *head = flow->head;
 326
 327         skb->next = NULL;
 328         if (!head) {
 329                 flow->head = skb;
 330                 flow->tail = skb;
 331                 return;
 332         }
 333         if (likely(!skb_is_retransmit(skb))) {
 334                 flow->tail->next = skb;
 335                 flow->tail = skb;
 336                 return;
 337         }
 338
 339         /* This skb is a tcp retransmit,
 340          * find the last retrans packet in the queue
 341          */
 342         prev = NULL;
 343         while (skb_is_retransmit(head)) {
 344                 prev = head;
 345                 head = head->next;
 346                 if (!head)
 347                         break;
 348         }
 349         if (!prev) { /* no rtx packet in queue, become the new head */
 350                 skb->next = flow->head;
 351                 flow->head = skb;
 352         } else {
 353                 if (prev == flow->tail)
 354                         flow->tail = skb;
 355                 else
 356                         skb->next = prev->next;
 357                 prev->next = skb;
 358         }
 359 }
 360
 361 static int fq_enqueue(struct sk_buff *skb, struct Qdisc *sch)
 362 {
 363         struct fq_sched_data *q = qdisc_priv(sch);
 364         struct fq_flow *f;
 365
 366         if (unlikely(sch->q.qlen >= sch->limit))
 367                 return qdisc_drop(skb, sch);
 368
 369         f = fq_classify(skb, q);
 370         if (unlikely(f->qlen >= q->flow_plimit && f != &q->internal)) {
 371                 q->stat_flows_plimit++;
 372                 return qdisc_drop(skb, sch);
 373         }
 374
 375         f->qlen++;
 376         flow_queue_add(f, skb);
 377         if (skb_is_retransmit(skb))
 378                 q->stat_tcp_retrans++;
 379         sch->qstats.backlog += qdisc_pkt_len(skb);
 380         if (fq_flow_is_detached(f)) {
 381                 fq_flow_add_tail(&q->new_flows, f);
 382                 if (q->quantum > f->credit)
 383                         f->credit = q->quantum;
 384                 q->inactive_flows--;
 385                 qdisc_unthrottled(sch);
 386         }
 387         if (unlikely(f == &q->internal)) {
 388                 q->stat_internal_packets++;
 389                 qdisc_unthrottled(sch);
 390         }
 391         sch->q.qlen++;
 392
 393         return NET_XMIT_SUCCESS;
 394 }
 395
 396 static void fq_check_throttled(struct fq_sched_data *q, u64 now)
 397 {
 398         struct rb_node *p;
 399
 400         if (q->time_next_delayed_flow > now)
 401                 return;
 402
 403         q->time_next_delayed_flow = ~0ULL;
 404         while ((p = rb_first(&q->delayed)) != NULL) {
 405                 struct fq_flow *f = container_of(p, struct fq_flow, rate_node);
 406
 407                 if (f->time_next_packet > now) {
 408                         q->time_next_delayed_flow = f->time_next_packet;
 409                         break;
 410                 }
 411                 rb_erase(p, &q->delayed);
 412                 q->throttled_flows--;
 413                 fq_flow_add_tail(&q->old_flows, f);
 414         }
 415 }
 416
 417 static struct sk_buff *fq_dequeue(struct Qdisc *sch)
 418 {
 419         struct fq_sched_data *q = qdisc_priv(sch);
 420         u64 now = ktime_to_ns(ktime_get());
 421         struct fq_flow_head *head;
 422         struct sk_buff *skb;
 423         struct fq_flow *f;
 424         u32 rate;
 425
 426         skb = fq_dequeue_head(sch, &q->internal);
 427         if (skb)
 428                 goto out;
 429         fq_check_throttled(q, now);
 430 begin:
 431         head = &q->new_flows;
 432         if (!head->first) {
 433                 head = &q->old_flows;
 434                 if (!head->first) {
 435                         if (q->time_next_delayed_flow != ~0ULL)
 436                                 qdisc_watchdog_schedule_ns(&q->watchdog,
 437                                                            q->time_next_delayed_flow);
 438                         return NULL;
 439                 }
 440         }
 441         f = head->first;
 442
 443         if (f->credit <= 0) {
 444                 f->credit += q->quantum;
 445                 head->first = f->next;
 446                 fq_flow_add_tail(&q->old_flows, f);
 447                 goto begin;
 448         }
 449
 450         if (unlikely(f->head && now < f->time_next_packet)) {
 451                 head->first = f->next;
 452                 fq_flow_set_throttled(q, f);
 453                 goto begin;
 454         }
 455
 456         skb = fq_dequeue_head(sch, f);
 457         if (!skb) {
 458                 head->first = f->next;
 459                 /* force a pass through old_flows to prevent starvation */
 460                 if ((head == &q->new_flows) && q->old_flows.first) {
 461                         fq_flow_add_tail(&q->old_flows, f);
 462                 } else {
 463                         fq_flow_set_detached(f);
 464                         f->age = jiffies;
 465                         q->inactive_flows++;
 466                 }
 467                 goto begin;
 468         }
 469         prefetch(&skb->end);
 470         f->time_next_packet = now;
 471         f->credit -= qdisc_pkt_len(skb);
 472
 473         if (f->credit > 0 || !q->rate_enable)
 474                 goto out;
 475
 476         rate = q->flow_max_rate;
 477         if (skb->sk && skb->sk->sk_state != TCP_TIME_WAIT)
 478                 rate = min(skb->sk->sk_pacing_rate, rate);
 479
 480         if (rate != ~0U) {
 481                 u32 plen = max(qdisc_pkt_len(skb), q->quantum);
 482                 u64 len = (u64)plen * NSEC_PER_SEC;
 483
 484                 if (likely(rate))
 485                         do_div(len, rate);
 486                 /* Since socket rate can change later,
 487                  * clamp the delay to 125 ms.
 488                  * TODO: maybe segment the too big skb, as in commit
 489                  * e43ac79a4bc ("sch_tbf: segment too big GSO packets")
 490                  */
 491                 if (unlikely(len > 125 * NSEC_PER_MSEC)) {
 492                         len = 125 * NSEC_PER_MSEC;
 493                         q->stat_pkts_too_long++;
 494                 }
 495
 496                 f->time_next_packet = now + len;
 497         }
 498 out:
 499         qdisc_bstats_update(sch, skb);
 500         qdisc_unthrottled(sch);
 501         return skb;
 502 }
 503
 504 static void fq_reset(struct Qdisc *sch)
 505 {
 506         struct fq_sched_data *q = qdisc_priv(sch);
 507         struct rb_root *root;
 508         struct sk_buff *skb;
 509         struct rb_node *p;
 510         struct fq_flow *f;
 511         unsigned int idx;
 512
 513         while ((skb = fq_dequeue_head(sch, &q->internal)) != NULL)
 514                 kfree_skb(skb);
 515
 516         if (!q->fq_root)
 517                 return;
 518
 519         for (idx = 0; idx < (1U << q->fq_trees_log); idx++) {
 520                 root = &q->fq_root[idx];
 521                 while ((p = rb_first(root)) != NULL) {
 522                         f = container_of(p, struct fq_flow, fq_node);
 523                         rb_erase(p, root);
 524
 525                         while ((skb = fq_dequeue_head(sch, f)) != NULL)
 526                                 kfree_skb(skb);
 527
 528                         kmem_cache_free(fq_flow_cachep, f);
 529                 }
 530         }
 531         q->new_flows.first      = NULL;
 532         q->old_flows.first      = NULL;
 533         q->delayed              = RB_ROOT;
 534         q->flows                = 0;
 535         q->inactive_flows       = 0;
 536         q->throttled_flows      = 0;
 537 }
 538
 539 static void fq_rehash(struct fq_sched_data *q,
 540                       struct rb_root *old_array, u32 old_log,
 541                       struct rb_root *new_array, u32 new_log)
 542 {
 543         struct rb_node *op, **np, *parent;
 544         struct rb_root *oroot, *nroot;
 545         struct fq_flow *of, *nf;
 546         int fcnt = 0;
 547         u32 idx;
 548
 549         for (idx = 0; idx < (1U << old_log); idx++) {
 550                 oroot = &old_array[idx];
 551                 while ((op = rb_first(oroot)) != NULL) {
 552                         rb_erase(op, oroot);
 553                         of = container_of(op, struct fq_flow, fq_node);
 554                         if (fq_gc_candidate(of)) {
 555                                 fcnt++;
 556                                 kmem_cache_free(fq_flow_cachep, of);
 557                                 continue;
 558                         }
 559                         nroot = &new_array[hash_32((u32)(long)of->sk, new_log)];
 560
 561                         np = &nroot->rb_node;
 562                         parent = NULL;
 563                         while (*np) {
 564                                 parent = *np;
 565
 566                                 nf = container_of(parent, struct fq_flow, fq_node);
 567                                 BUG_ON(nf->sk == of->sk);
 568
 569                                 if (nf->sk > of->sk)
 570                                         np = &parent->rb_right;
 571                                 else
 572                                         np = &parent->rb_left;
 573                         }
 574
 575                         rb_link_node(&of->fq_node, parent, np);
 576                         rb_insert_color(&of->fq_node, nroot);
 577                 }
 578         }
 579         q->flows -= fcnt;
 580         q->inactive_flows -= fcnt;
 581         q->stat_gc_flows += fcnt;
 582 }
 583
 584 static int fq_resize(struct fq_sched_data *q, u32 log)
 585 {
 586         struct rb_root *array;
 587         u32 idx;
 588
 589         if (q->fq_root && log == q->fq_trees_log)
 590                 return 0;
 591
 592         array = kmalloc(sizeof(struct rb_root) << log, GFP_KERNEL);
 593         if (!array)
 594                 return -ENOMEM;
 595
 596         for (idx = 0; idx < (1U << log); idx++)
 597                 array[idx] = RB_ROOT;
 598
 599         if (q->fq_root) {
 600                 fq_rehash(q, q->fq_root, q->fq_trees_log, array, log);
 601                 kfree(q->fq_root);
 602         }
 603         q->fq_root = array;
 604         q->fq_trees_log = log;
 605
 606         return 0;
 607 }
 608
 609 static const struct nla_policy fq_policy[TCA_FQ_MAX + 1] = {
 610         [TCA_FQ_PLIMIT]                 = { .type = NLA_U32 },
 611         [TCA_FQ_FLOW_PLIMIT]            = { .type = NLA_U32 },
 612         [TCA_FQ_QUANTUM]                = { .type = NLA_U32 },
 613         [TCA_FQ_INITIAL_QUANTUM]        = { .type = NLA_U32 },
 614         [TCA_FQ_RATE_ENABLE]            = { .type = NLA_U32 },
 615         [TCA_FQ_FLOW_DEFAULT_RATE]      = { .type = NLA_U32 },
 616         [TCA_FQ_FLOW_MAX_RATE]          = { .type = NLA_U32 },
 617         [TCA_FQ_BUCKETS_LOG]            = { .type = NLA_U32 },
 618 };
 619
 620 static int fq_change(struct Qdisc *sch, struct nlattr *opt)
 621 {
 622         struct fq_sched_data *q = qdisc_priv(sch);
 623         struct nlattr *tb[TCA_FQ_MAX + 1];
 624         int err, drop_count = 0;
 625         u32 fq_log;
 626
 627         if (!opt)
 628                 return -EINVAL;
 629
 630         err = nla_parse_nested(tb, TCA_FQ_MAX, opt, fq_policy);
 631         if (err < 0)
 632                 return err;
 633
 634         sch_tree_lock(sch);
 635
 636         fq_log = q->fq_trees_log;
 637
 638         if (tb[TCA_FQ_BUCKETS_LOG]) {
 639                 u32 nval = nla_get_u32(tb[TCA_FQ_BUCKETS_LOG]);
 640
 641                 if (nval >= 1 && nval <= ilog2(256*1024))
 642                         fq_log = nval;
 643                 else
 644                         err = -EINVAL;
 645         }
 646         if (tb[TCA_FQ_PLIMIT])
 647                 sch->limit = nla_get_u32(tb[TCA_FQ_PLIMIT]);
 648
 649         if (tb[TCA_FQ_FLOW_PLIMIT])
 650                 q->flow_plimit = nla_get_u32(tb[TCA_FQ_FLOW_PLIMIT]);
 651
 652         if (tb[TCA_FQ_QUANTUM])
 653                 q->quantum = nla_get_u32(tb[TCA_FQ_QUANTUM]);
 654
 655         if (tb[TCA_FQ_INITIAL_QUANTUM])
 656                 q->initial_quantum = nla_get_u32(tb[TCA_FQ_INITIAL_QUANTUM]);
 657
 658         if (tb[TCA_FQ_FLOW_DEFAULT_RATE])
 659                 q->flow_default_rate = nla_get_u32(tb[TCA_FQ_FLOW_DEFAULT_RATE]);
 660
 661         if (tb[TCA_FQ_FLOW_MAX_RATE])
 662                 q->flow_max_rate = nla_get_u32(tb[TCA_FQ_FLOW_MAX_RATE]);
 663
 664         if (tb[TCA_FQ_RATE_ENABLE]) {
 665                 u32 enable = nla_get_u32(tb[TCA_FQ_RATE_ENABLE]);
 666
 667                 if (enable <= 1)
 668                         q->rate_enable = enable;
 669                 else
 670                         err = -EINVAL;
 671         }
 672
 673         if (!err)
 674                 err = fq_resize(q, fq_log);
 675
 676         while (sch->q.qlen > sch->limit) {
 677                 struct sk_buff *skb = fq_dequeue(sch);
 678
 679                 if (!skb)
 680                         break;
 681                 kfree_skb(skb);
 682                 drop_count++;
 683         }
 684         qdisc_tree_decrease_qlen(sch, drop_count);
 685
 686         sch_tree_unlock(sch);
 687         return err;
 688 }
 689
 690 static void fq_destroy(struct Qdisc *sch)
 691 {
 692         struct fq_sched_data *q = qdisc_priv(sch);
 693
 694         fq_reset(sch);
 695         kfree(q->fq_root);
 696         qdisc_watchdog_cancel(&q->watchdog);
 697 }
 698
 699 static int fq_init(struct Qdisc *sch, struct nlattr *opt)
 700 {
 701         struct fq_sched_data *q = qdisc_priv(sch);
 702         int err;
 703
 704         sch->limit              = 10000;
 705         q->flow_plimit          = 100;
 706         q->quantum              = 2 * psched_mtu(qdisc_dev(sch));
 707         q->initial_quantum      = 10 * psched_mtu(qdisc_dev(sch));
 708         q->flow_default_rate    = 0;
 709         q->flow_max_rate        = ~0U;
 710         q->rate_enable          = 1;
 711         q->new_flows.first      = NULL;
 712         q->old_flows.first      = NULL;
 713         q->delayed              = RB_ROOT;
 714         q->fq_root              = NULL;
 715         q->fq_trees_log         = ilog2(1024);
 716         qdisc_watchdog_init(&q->watchdog, sch);
 717
 718         if (opt)
 719                 err = fq_change(sch, opt);
 720         else
 721                 err = fq_resize(q, q->fq_trees_log);
 722
 723         return err;
 724 }
 725
 726 static int fq_dump(struct Qdisc *sch, struct sk_buff *skb)
 727 {
 728         struct fq_sched_data *q = qdisc_priv(sch);
 729         struct nlattr *opts;
 730
 731         opts = nla_nest_start(skb, TCA_OPTIONS);
 732         if (opts == NULL)
 733                 goto nla_put_failure;
 734
 735         /* TCA_FQ_FLOW_DEFAULT_RATE is not used anymore,
 736          * do not bother giving its value
 737          */
 738         if (nla_put_u32(skb, TCA_FQ_PLIMIT, sch->limit) ||
 739             nla_put_u32(skb, TCA_FQ_FLOW_PLIMIT, q->flow_plimit) ||
 740             nla_put_u32(skb, TCA_FQ_QUANTUM, q->quantum) ||
 741             nla_put_u32(skb, TCA_FQ_INITIAL_QUANTUM, q->initial_quantum) ||
 742             nla_put_u32(skb, TCA_FQ_RATE_ENABLE, q->rate_enable) ||
 743             nla_put_u32(skb, TCA_FQ_FLOW_MAX_RATE, q->flow_max_rate) ||
 744             nla_put_u32(skb, TCA_FQ_BUCKETS_LOG, q->fq_trees_log))
 745                 goto nla_put_failure;
 746
 747         nla_nest_end(skb, opts);
 748         return skb->len;
 749
 750 nla_put_failure:
 751         return -1;
 752 }
 753
 754 static int fq_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
 755 {
 756         struct fq_sched_data *q = qdisc_priv(sch);
 757         u64 now = ktime_to_ns(ktime_get());
 758         struct tc_fq_qd_stats st = {
 759                 .gc_flows               = q->stat_gc_flows,
 760                 .highprio_packets       = q->stat_internal_packets,
 761                 .tcp_retrans            = q->stat_tcp_retrans,
 762                 .throttled              = q->stat_throttled,
 763                 .flows_plimit           = q->stat_flows_plimit,
 764                 .pkts_too_long          = q->stat_pkts_too_long,
 765                 .allocation_errors      = q->stat_allocation_errors,
 766                 .flows                  = q->flows,
 767                 .inactive_flows         = q->inactive_flows,
 768                 .throttled_flows        = q->throttled_flows,
 769                 .time_next_delayed_flow = q->time_next_delayed_flow - now,
 770         };
 771
 772         return gnet_stats_copy_app(d, &st, sizeof(st));
 773 }
 774
 775 static struct Qdisc_ops fq_qdisc_ops __read_mostly = {
 776         .id             =       "fq",
 777         .priv_size      =       sizeof(struct fq_sched_data),
 778
 779         .enqueue        =       fq_enqueue,
 780         .dequeue        =       fq_dequeue,
 781         .peek           =       qdisc_peek_dequeued,
 782         .init           =       fq_init,
 783         .reset          =       fq_reset,
 784         .destroy        =       fq_destroy,
 785         .change         =       fq_change,
 786         .dump           =       fq_dump,
 787         .dump_stats     =       fq_dump_stats,
 788         .owner          =       THIS_MODULE,
 789 };
 790
 791 static int __init fq_module_init(void)
 792 {
 793         int ret;
 794
 795         fq_flow_cachep = kmem_cache_create("fq_flow_cache",
 796                                            sizeof(struct fq_flow),
 797                                            0, 0, NULL);
 798         if (!fq_flow_cachep)
 799                 return -ENOMEM;
 800
 801         ret = register_qdisc(&fq_qdisc_ops);
 802         if (ret)
 803                 kmem_cache_destroy(fq_flow_cachep);
 804         return ret;
 805 }
 806
 807 static void __exit fq_module_exit(void)
 808 {
 809         unregister_qdisc(&fq_qdisc_ops);
 810         kmem_cache_destroy(fq_flow_cachep);
 811 }
 812
 813 module_init(fq_module_init)
 814 module_exit(fq_module_exit)
 815 MODULE_AUTHOR("Eric Dumazet");
 816 MODULE_LICENSE("GPL");