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_refill_delay;
  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         f->age = jiffies;
 119 }
 120
 121 static bool fq_flow_is_detached(const struct fq_flow *f)
 122 {
 123         return f->next == &detached;
 124 }
 125
 126 static void fq_flow_set_throttled(struct fq_sched_data *q, struct fq_flow *f)
 127 {
 128         struct rb_node **p = &q->delayed.rb_node, *parent = NULL;
 129
 130         while (*p) {
 131                 struct fq_flow *aux;
 132
 133                 parent = *p;
 134                 aux = container_of(parent, struct fq_flow, rate_node);
 135                 if (f->time_next_packet >= aux->time_next_packet)
 136                         p = &parent->rb_right;
 137                 else
 138                         p = &parent->rb_left;
 139         }
 140         rb_link_node(&f->rate_node, parent, p);
 141         rb_insert_color(&f->rate_node, &q->delayed);
 142         q->throttled_flows++;
 143         q->stat_throttled++;
 144
 145         f->next = &throttled;
 146         if (q->time_next_delayed_flow > f->time_next_packet)
 147                 q->time_next_delayed_flow = f->time_next_packet;
 148 }
 149
 150
 151 static struct kmem_cache *fq_flow_cachep __read_mostly;
 152
 153 static void fq_flow_add_tail(struct fq_flow_head *head, struct fq_flow *flow)
 154 {
 155         if (head->first)
 156                 head->last->next = flow;
 157         else
 158                 head->first = flow;
 159         head->last = flow;
 160         flow->next = NULL;
 161 }
 162
 163 /* limit number of collected flows per round */
 164 #define FQ_GC_MAX 8
 165 #define FQ_GC_AGE (3*HZ)
 166
 167 static bool fq_gc_candidate(const struct fq_flow *f)
 168 {
 169         return fq_flow_is_detached(f) &&
 170                time_after(jiffies, f->age + FQ_GC_AGE);
 171 }
 172
 173 static void fq_gc(struct fq_sched_data *q,
 174                   struct rb_root *root,
 175                   struct sock *sk)
 176 {
 177         struct fq_flow *f, *tofree[FQ_GC_MAX];
 178         struct rb_node **p, *parent;
 179         int fcnt = 0;
 180
 181         p = &root->rb_node;
 182         parent = NULL;
 183         while (*p) {
 184                 parent = *p;
 185
 186                 f = container_of(parent, struct fq_flow, fq_node);
 187                 if (f->sk == sk)
 188                         break;
 189
 190                 if (fq_gc_candidate(f)) {
 191                         tofree[fcnt++] = f;
 192                         if (fcnt == FQ_GC_MAX)
 193                                 break;
 194                 }
 195
 196                 if (f->sk > sk)
 197                         p = &parent->rb_right;
 198                 else
 199                         p = &parent->rb_left;
 200         }
 201
 202         q->flows -= fcnt;
 203         q->inactive_flows -= fcnt;
 204         q->stat_gc_flows += fcnt;
 205         while (fcnt) {
 206                 struct fq_flow *f = tofree[--fcnt];
 207
 208                 rb_erase(&f->fq_node, root);
 209                 kmem_cache_free(fq_flow_cachep, f);
 210         }
 211 }
 212
 213 static struct fq_flow *fq_classify(struct sk_buff *skb, struct fq_sched_data *q)
 214 {
 215         struct rb_node **p, *parent;
 216         struct sock *sk = skb->sk;
 217         struct rb_root *root;
 218         struct fq_flow *f;
 219
 220         /* warning: no starvation prevention... */
 221         if (unlikely((skb->priority & TC_PRIO_MAX) == TC_PRIO_CONTROL))
 222                 return &q->internal;
 223
 224         if (unlikely(!sk)) {
 225                 /* By forcing low order bit to 1, we make sure to not
 226                  * collide with a local flow (socket pointers are word aligned)
 227                  */
 228                 sk = (struct sock *)(skb_get_rxhash(skb) | 1L);
 229         }
 230
 231         root = &q->fq_root[hash_32((u32)(long)sk, q->fq_trees_log)];
 232
 233         if (q->flows >= (2U << q->fq_trees_log) &&
 234             q->inactive_flows > q->flows/2)
 235                 fq_gc(q, root, sk);
 236
 237         p = &root->rb_node;
 238         parent = NULL;
 239         while (*p) {
 240                 parent = *p;
 241
 242                 f = container_of(parent, struct fq_flow, fq_node);
 243                 if (f->sk == sk) {
 244                         /* socket might have been reallocated, so check
 245                          * if its sk_hash is the same.
 246                          * It not, we need to refill credit with
 247                          * initial quantum
 248                          */
 249                         if (unlikely(skb->sk &&
 250                                      f->socket_hash != sk->sk_hash)) {
 251                                 f->credit = q->initial_quantum;
 252                                 f->socket_hash = sk->sk_hash;
 253                                 f->time_next_packet = 0ULL;
 254                         }
 255                         return f;
 256                 }
 257                 if (f->sk > sk)
 258                         p = &parent->rb_right;
 259                 else
 260                         p = &parent->rb_left;
 261         }
 262
 263         f = kmem_cache_zalloc(fq_flow_cachep, GFP_ATOMIC | __GFP_NOWARN);
 264         if (unlikely(!f)) {
 265                 q->stat_allocation_errors++;
 266                 return &q->internal;
 267         }
 268         fq_flow_set_detached(f);
 269         f->sk = sk;
 270         if (skb->sk)
 271                 f->socket_hash = sk->sk_hash;
 272         f->credit = q->initial_quantum;
 273
 274         rb_link_node(&f->fq_node, parent, p);
 275         rb_insert_color(&f->fq_node, root);
 276
 277         q->flows++;
 278         q->inactive_flows++;
 279         return f;
 280 }
 281
 282
 283 /* remove one skb from head of flow queue */
 284 static struct sk_buff *fq_dequeue_head(struct Qdisc *sch, struct fq_flow *flow)
 285 {
 286         struct sk_buff *skb = flow->head;
 287
 288         if (skb) {
 289                 flow->head = skb->next;
 290                 skb->next = NULL;
 291                 flow->qlen--;
 292                 sch->qstats.backlog -= qdisc_pkt_len(skb);
 293                 sch->q.qlen--;
 294         }
 295         return skb;
 296 }
 297
 298 /* We might add in the future detection of retransmits
 299  * For the time being, just return false
 300  */
 301 static bool skb_is_retransmit(struct sk_buff *skb)
 302 {
 303         return false;
 304 }
 305
 306 /* add skb to flow queue
 307  * flow queue is a linked list, kind of FIFO, except for TCP retransmits
 308  * We special case tcp retransmits to be transmitted before other packets.
 309  * We rely on fact that TCP retransmits are unlikely, so we do not waste
 310  * a separate queue or a pointer.
 311  * head->  [retrans pkt 1]
 312  *         [retrans pkt 2]
 313  *         [ normal pkt 1]
 314  *         [ normal pkt 2]
 315  *         [ normal pkt 3]
 316  * tail->  [ normal pkt 4]
 317  */
 318 static void flow_queue_add(struct fq_flow *flow, struct sk_buff *skb)
 319 {
 320         struct sk_buff *prev, *head = flow->head;
 321
 322         skb->next = NULL;
 323         if (!head) {
 324                 flow->head = skb;
 325                 flow->tail = skb;
 326                 return;
 327         }
 328         if (likely(!skb_is_retransmit(skb))) {
 329                 flow->tail->next = skb;
 330                 flow->tail = skb;
 331                 return;
 332         }
 333
 334         /* This skb is a tcp retransmit,
 335          * find the last retrans packet in the queue
 336          */
 337         prev = NULL;
 338         while (skb_is_retransmit(head)) {
 339                 prev = head;
 340                 head = head->next;
 341                 if (!head)
 342                         break;
 343         }
 344         if (!prev) { /* no rtx packet in queue, become the new head */
 345                 skb->next = flow->head;
 346                 flow->head = skb;
 347         } else {
 348                 if (prev == flow->tail)
 349                         flow->tail = skb;
 350                 else
 351                         skb->next = prev->next;
 352                 prev->next = skb;
 353         }
 354 }
 355
 356 static int fq_enqueue(struct sk_buff *skb, struct Qdisc *sch)
 357 {
 358         struct fq_sched_data *q = qdisc_priv(sch);
 359         struct fq_flow *f;
 360
 361         if (unlikely(sch->q.qlen >= sch->limit))
 362                 return qdisc_drop(skb, sch);
 363
 364         f = fq_classify(skb, q);
 365         if (unlikely(f->qlen >= q->flow_plimit && f != &q->internal)) {
 366                 q->stat_flows_plimit++;
 367                 return qdisc_drop(skb, sch);
 368         }
 369
 370         f->qlen++;
 371         if (skb_is_retransmit(skb))
 372                 q->stat_tcp_retrans++;
 373         sch->qstats.backlog += qdisc_pkt_len(skb);
 374         if (fq_flow_is_detached(f)) {
 375                 fq_flow_add_tail(&q->new_flows, f);
 376                 if (time_after(jiffies, f->age + q->flow_refill_delay))
 377                         f->credit = max_t(u32, f->credit, q->quantum);
 378                 q->inactive_flows--;
 379                 qdisc_unthrottled(sch);
 380         }
 381
 382         /* Note: this overwrites f->age */
 383         flow_queue_add(f, skb);
 384
 385         if (unlikely(f == &q->internal)) {
 386                 q->stat_internal_packets++;
 387                 qdisc_unthrottled(sch);
 388         }
 389         sch->q.qlen++;
 390
 391         return NET_XMIT_SUCCESS;
 392 }
 393
 394 static void fq_check_throttled(struct fq_sched_data *q, u64 now)
 395 {
 396         struct rb_node *p;
 397
 398         if (q->time_next_delayed_flow > now)
 399                 return;
 400
 401         q->time_next_delayed_flow = ~0ULL;
 402         while ((p = rb_first(&q->delayed)) != NULL) {
 403                 struct fq_flow *f = container_of(p, struct fq_flow, rate_node);
 404
 405                 if (f->time_next_packet > now) {
 406                         q->time_next_delayed_flow = f->time_next_packet;
 407                         break;
 408                 }
 409                 rb_erase(p, &q->delayed);
 410                 q->throttled_flows--;
 411                 fq_flow_add_tail(&q->old_flows, f);
 412         }
 413 }
 414
 415 static struct sk_buff *fq_dequeue(struct Qdisc *sch)
 416 {
 417         struct fq_sched_data *q = qdisc_priv(sch);
 418         u64 now = ktime_to_ns(ktime_get());
 419         struct fq_flow_head *head;
 420         struct sk_buff *skb;
 421         struct fq_flow *f;
 422         u32 rate;
 423
 424         skb = fq_dequeue_head(sch, &q->internal);
 425         if (skb)
 426                 goto out;
 427         fq_check_throttled(q, now);
 428 begin:
 429         head = &q->new_flows;
 430         if (!head->first) {
 431                 head = &q->old_flows;
 432                 if (!head->first) {
 433                         if (q->time_next_delayed_flow != ~0ULL)
 434                                 qdisc_watchdog_schedule_ns(&q->watchdog,
 435                                                            q->time_next_delayed_flow);
 436                         return NULL;
 437                 }
 438         }
 439         f = head->first;
 440
 441         if (f->credit <= 0) {
 442                 f->credit += q->quantum;
 443                 head->first = f->next;
 444                 fq_flow_add_tail(&q->old_flows, f);
 445                 goto begin;
 446         }
 447
 448         if (unlikely(f->head && now < f->time_next_packet)) {
 449                 head->first = f->next;
 450                 fq_flow_set_throttled(q, f);
 451                 goto begin;
 452         }
 453
 454         skb = fq_dequeue_head(sch, f);
 455         if (!skb) {
 456                 head->first = f->next;
 457                 /* force a pass through old_flows to prevent starvation */
 458                 if ((head == &q->new_flows) && q->old_flows.first) {
 459                         fq_flow_add_tail(&q->old_flows, f);
 460                 } else {
 461                         fq_flow_set_detached(f);
 462                         q->inactive_flows++;
 463                 }
 464                 goto begin;
 465         }
 466         prefetch(&skb->end);
 467         f->time_next_packet = now;
 468         f->credit -= qdisc_pkt_len(skb);
 469
 470         if (f->credit > 0 || !q->rate_enable)
 471                 goto out;
 472
 473         rate = q->flow_max_rate;
 474         if (skb->sk && skb->sk->sk_state != TCP_TIME_WAIT)
 475                 rate = min(skb->sk->sk_pacing_rate, rate);
 476
 477         if (rate != ~0U) {
 478                 u32 plen = max(qdisc_pkt_len(skb), q->quantum);
 479                 u64 len = (u64)plen * NSEC_PER_SEC;
 480
 481                 if (likely(rate))
 482                         do_div(len, rate);
 483                 /* Since socket rate can change later,
 484                  * clamp the delay to 125 ms.
 485                  * TODO: maybe segment the too big skb, as in commit
 486                  * e43ac79a4bc ("sch_tbf: segment too big GSO packets")
 487                  */
 488                 if (unlikely(len > 125 * NSEC_PER_MSEC)) {
 489                         len = 125 * NSEC_PER_MSEC;
 490                         q->stat_pkts_too_long++;
 491                 }
 492
 493                 f->time_next_packet = now + len;
 494         }
 495 out:
 496         qdisc_bstats_update(sch, skb);
 497         qdisc_unthrottled(sch);
 498         return skb;
 499 }
 500
 501 static void fq_reset(struct Qdisc *sch)
 502 {
 503         struct fq_sched_data *q = qdisc_priv(sch);
 504         struct rb_root *root;
 505         struct sk_buff *skb;
 506         struct rb_node *p;
 507         struct fq_flow *f;
 508         unsigned int idx;
 509
 510         while ((skb = fq_dequeue_head(sch, &q->internal)) != NULL)
 511                 kfree_skb(skb);
 512
 513         if (!q->fq_root)
 514                 return;
 515
 516         for (idx = 0; idx < (1U << q->fq_trees_log); idx++) {
 517                 root = &q->fq_root[idx];
 518                 while ((p = rb_first(root)) != NULL) {
 519                         f = container_of(p, struct fq_flow, fq_node);
 520                         rb_erase(p, root);
 521
 522                         while ((skb = fq_dequeue_head(sch, f)) != NULL)
 523                                 kfree_skb(skb);
 524
 525                         kmem_cache_free(fq_flow_cachep, f);
 526                 }
 527         }
 528         q->new_flows.first      = NULL;
 529         q->old_flows.first      = NULL;
 530         q->delayed              = RB_ROOT;
 531         q->flows                = 0;
 532         q->inactive_flows       = 0;
 533         q->throttled_flows      = 0;
 534 }
 535
 536 static void fq_rehash(struct fq_sched_data *q,
 537                       struct rb_root *old_array, u32 old_log,
 538                       struct rb_root *new_array, u32 new_log)
 539 {
 540         struct rb_node *op, **np, *parent;
 541         struct rb_root *oroot, *nroot;
 542         struct fq_flow *of, *nf;
 543         int fcnt = 0;
 544         u32 idx;
 545
 546         for (idx = 0; idx < (1U << old_log); idx++) {
 547                 oroot = &old_array[idx];
 548                 while ((op = rb_first(oroot)) != NULL) {
 549                         rb_erase(op, oroot);
 550                         of = container_of(op, struct fq_flow, fq_node);
 551                         if (fq_gc_candidate(of)) {
 552                                 fcnt++;
 553                                 kmem_cache_free(fq_flow_cachep, of);
 554                                 continue;
 555                         }
 556                         nroot = &new_array[hash_32((u32)(long)of->sk, new_log)];
 557
 558                         np = &nroot->rb_node;
 559                         parent = NULL;
 560                         while (*np) {
 561                                 parent = *np;
 562
 563                                 nf = container_of(parent, struct fq_flow, fq_node);
 564                                 BUG_ON(nf->sk == of->sk);
 565
 566                                 if (nf->sk > of->sk)
 567                                         np = &parent->rb_right;
 568                                 else
 569                                         np = &parent->rb_left;
 570                         }
 571
 572                         rb_link_node(&of->fq_node, parent, np);
 573                         rb_insert_color(&of->fq_node, nroot);
 574                 }
 575         }
 576         q->flows -= fcnt;
 577         q->inactive_flows -= fcnt;
 578         q->stat_gc_flows += fcnt;
 579 }
 580
 581 static int fq_resize(struct fq_sched_data *q, u32 log)
 582 {
 583         struct rb_root *array;
 584         u32 idx;
 585
 586         if (q->fq_root && log == q->fq_trees_log)
 587                 return 0;
 588
 589         array = kmalloc(sizeof(struct rb_root) << log, GFP_KERNEL);
 590         if (!array)
 591                 return -ENOMEM;
 592
 593         for (idx = 0; idx < (1U << log); idx++)
 594                 array[idx] = RB_ROOT;
 595
 596         if (q->fq_root) {
 597                 fq_rehash(q, q->fq_root, q->fq_trees_log, array, log);
 598                 kfree(q->fq_root);
 599         }
 600         q->fq_root = array;
 601         q->fq_trees_log = log;
 602
 603         return 0;
 604 }
 605
 606 static const struct nla_policy fq_policy[TCA_FQ_MAX + 1] = {
 607         [TCA_FQ_PLIMIT]                 = { .type = NLA_U32 },
 608         [TCA_FQ_FLOW_PLIMIT]            = { .type = NLA_U32 },
 609         [TCA_FQ_QUANTUM]                = { .type = NLA_U32 },
 610         [TCA_FQ_INITIAL_QUANTUM]        = { .type = NLA_U32 },
 611         [TCA_FQ_RATE_ENABLE]            = { .type = NLA_U32 },
 612         [TCA_FQ_FLOW_DEFAULT_RATE]      = { .type = NLA_U32 },
 613         [TCA_FQ_FLOW_MAX_RATE]          = { .type = NLA_U32 },
 614         [TCA_FQ_BUCKETS_LOG]            = { .type = NLA_U32 },
 615         [TCA_FQ_FLOW_REFILL_DELAY]      = { .type = NLA_U32 },
 616 };
 617
 618 static int fq_change(struct Qdisc *sch, struct nlattr *opt)
 619 {
 620         struct fq_sched_data *q = qdisc_priv(sch);
 621         struct nlattr *tb[TCA_FQ_MAX + 1];
 622         int err, drop_count = 0;
 623         u32 fq_log;
 624
 625         if (!opt)
 626                 return -EINVAL;
 627
 628         err = nla_parse_nested(tb, TCA_FQ_MAX, opt, fq_policy);
 629         if (err < 0)
 630                 return err;
 631
 632         sch_tree_lock(sch);
 633
 634         fq_log = q->fq_trees_log;
 635
 636         if (tb[TCA_FQ_BUCKETS_LOG]) {
 637                 u32 nval = nla_get_u32(tb[TCA_FQ_BUCKETS_LOG]);
 638
 639                 if (nval >= 1 && nval <= ilog2(256*1024))
 640                         fq_log = nval;
 641                 else
 642                         err = -EINVAL;
 643         }
 644         if (tb[TCA_FQ_PLIMIT])
 645                 sch->limit = nla_get_u32(tb[TCA_FQ_PLIMIT]);
 646
 647         if (tb[TCA_FQ_FLOW_PLIMIT])
 648                 q->flow_plimit = nla_get_u32(tb[TCA_FQ_FLOW_PLIMIT]);
 649
 650         if (tb[TCA_FQ_QUANTUM])
 651                 q->quantum = nla_get_u32(tb[TCA_FQ_QUANTUM]);
 652
 653         if (tb[TCA_FQ_INITIAL_QUANTUM])
 654                 q->initial_quantum = nla_get_u32(tb[TCA_FQ_INITIAL_QUANTUM]);
 655
 656         if (tb[TCA_FQ_FLOW_DEFAULT_RATE])
 657                 pr_warn_ratelimited("sch_fq: defrate %u ignored.\n",
 658                                     nla_get_u32(tb[TCA_FQ_FLOW_DEFAULT_RATE]));
 659
 660         if (tb[TCA_FQ_FLOW_MAX_RATE])
 661                 q->flow_max_rate = nla_get_u32(tb[TCA_FQ_FLOW_MAX_RATE]);
 662
 663         if (tb[TCA_FQ_RATE_ENABLE]) {
 664                 u32 enable = nla_get_u32(tb[TCA_FQ_RATE_ENABLE]);
 665
 666                 if (enable <= 1)
 667                         q->rate_enable = enable;
 668                 else
 669                         err = -EINVAL;
 670         }
 671
 672         if (tb[TCA_FQ_FLOW_REFILL_DELAY]) {
 673                 u32 usecs_delay = nla_get_u32(tb[TCA_FQ_FLOW_REFILL_DELAY]) ;
 674
 675                 q->flow_refill_delay = usecs_to_jiffies(usecs_delay);
 676         }
 677
 678         if (!err)
 679                 err = fq_resize(q, fq_log);
 680
 681         while (sch->q.qlen > sch->limit) {
 682                 struct sk_buff *skb = fq_dequeue(sch);
 683
 684                 if (!skb)
 685                         break;
 686                 kfree_skb(skb);
 687                 drop_count++;
 688         }
 689         qdisc_tree_decrease_qlen(sch, drop_count);
 690
 691         sch_tree_unlock(sch);
 692         return err;
 693 }
 694
 695 static void fq_destroy(struct Qdisc *sch)
 696 {
 697         struct fq_sched_data *q = qdisc_priv(sch);
 698
 699         fq_reset(sch);
 700         kfree(q->fq_root);
 701         qdisc_watchdog_cancel(&q->watchdog);
 702 }
 703
 704 static int fq_init(struct Qdisc *sch, struct nlattr *opt)
 705 {
 706         struct fq_sched_data *q = qdisc_priv(sch);
 707         int err;
 708
 709         sch->limit              = 10000;
 710         q->flow_plimit          = 100;
 711         q->quantum              = 2 * psched_mtu(qdisc_dev(sch));
 712         q->initial_quantum      = 10 * psched_mtu(qdisc_dev(sch));
 713         q->flow_refill_delay    = msecs_to_jiffies(40);
 714         q->flow_max_rate        = ~0U;
 715         q->rate_enable          = 1;
 716         q->new_flows.first      = NULL;
 717         q->old_flows.first      = NULL;
 718         q->delayed              = RB_ROOT;
 719         q->fq_root              = NULL;
 720         q->fq_trees_log         = ilog2(1024);
 721         qdisc_watchdog_init(&q->watchdog, sch);
 722
 723         if (opt)
 724                 err = fq_change(sch, opt);
 725         else
 726                 err = fq_resize(q, q->fq_trees_log);
 727
 728         return err;
 729 }
 730
 731 static int fq_dump(struct Qdisc *sch, struct sk_buff *skb)
 732 {
 733         struct fq_sched_data *q = qdisc_priv(sch);
 734         struct nlattr *opts;
 735
 736         opts = nla_nest_start(skb, TCA_OPTIONS);
 737         if (opts == NULL)
 738                 goto nla_put_failure;
 739
 740         /* TCA_FQ_FLOW_DEFAULT_RATE is not used anymore */
 741
 742         if (nla_put_u32(skb, TCA_FQ_PLIMIT, sch->limit) ||
 743             nla_put_u32(skb, TCA_FQ_FLOW_PLIMIT, q->flow_plimit) ||
 744             nla_put_u32(skb, TCA_FQ_QUANTUM, q->quantum) ||
 745             nla_put_u32(skb, TCA_FQ_INITIAL_QUANTUM, q->initial_quantum) ||
 746             nla_put_u32(skb, TCA_FQ_RATE_ENABLE, q->rate_enable) ||
 747             nla_put_u32(skb, TCA_FQ_FLOW_MAX_RATE, q->flow_max_rate) ||
 748             nla_put_u32(skb, TCA_FQ_FLOW_REFILL_DELAY,
 749                         jiffies_to_usecs(q->flow_refill_delay)) ||
 750             nla_put_u32(skb, TCA_FQ_BUCKETS_LOG, q->fq_trees_log))
 751                 goto nla_put_failure;
 752
 753         nla_nest_end(skb, opts);
 754         return skb->len;
 755
 756 nla_put_failure:
 757         return -1;
 758 }
 759
 760 static int fq_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
 761 {
 762         struct fq_sched_data *q = qdisc_priv(sch);
 763         u64 now = ktime_to_ns(ktime_get());
 764         struct tc_fq_qd_stats st = {
 765                 .gc_flows               = q->stat_gc_flows,
 766                 .highprio_packets       = q->stat_internal_packets,
 767                 .tcp_retrans            = q->stat_tcp_retrans,
 768                 .throttled              = q->stat_throttled,
 769                 .flows_plimit           = q->stat_flows_plimit,
 770                 .pkts_too_long          = q->stat_pkts_too_long,
 771                 .allocation_errors      = q->stat_allocation_errors,
 772                 .flows                  = q->flows,
 773                 .inactive_flows         = q->inactive_flows,
 774                 .throttled_flows        = q->throttled_flows,
 775                 .time_next_delayed_flow = q->time_next_delayed_flow - now,
 776         };
 777
 778         return gnet_stats_copy_app(d, &st, sizeof(st));
 779 }
 780
 781 static struct Qdisc_ops fq_qdisc_ops __read_mostly = {
 782         .id             =       "fq",
 783         .priv_size      =       sizeof(struct fq_sched_data),
 784
 785         .enqueue        =       fq_enqueue,
 786         .dequeue        =       fq_dequeue,
 787         .peek           =       qdisc_peek_dequeued,
 788         .init           =       fq_init,
 789         .reset          =       fq_reset,
 790         .destroy        =       fq_destroy,
 791         .change         =       fq_change,
 792         .dump           =       fq_dump,
 793         .dump_stats     =       fq_dump_stats,
 794         .owner          =       THIS_MODULE,
 795 };
 796
 797 static int __init fq_module_init(void)
 798 {
 799         int ret;
 800
 801         fq_flow_cachep = kmem_cache_create("fq_flow_cache",
 802                                            sizeof(struct fq_flow),
 803                                            0, 0, NULL);
 804         if (!fq_flow_cachep)
 805                 return -ENOMEM;
 806
 807         ret = register_qdisc(&fq_qdisc_ops);
 808         if (ret)
 809                 kmem_cache_destroy(fq_flow_cachep);
 810         return ret;
 811 }
 812
 813 static void __exit fq_module_exit(void)
 814 {
 815         unregister_qdisc(&fq_qdisc_ops);
 816         kmem_cache_destroy(fq_flow_cachep);
 817 }
 818
 819 module_init(fq_module_init)
 820 module_exit(fq_module_exit)
 821 MODULE_AUTHOR("Eric Dumazet");
 822 MODULE_LICENSE("GPL");