net/sched/sch_fq.c

   1 /*
   2  * net/sched/sch_fq.c Fair Queue Packet Scheduler (per flow pacing)
   3  *
   4  *  Copyright (C) 2013-2015 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 locally 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 <linux/vmalloc.h>
  51 #include <net/netlink.h>
  52 #include <net/pkt_sched.h>
  53 #include <net/sock.h>
  54 #include <net/tcp_states.h>
  55 #include <net/tcp.h>
  56
  57 /*
  58  * Per flow structure, dynamically allocated
  59  */
  60 struct fq_flow {
  61         struct sk_buff  *head;          /* list of skbs for this flow : first skb */
  62         union {
  63                 struct sk_buff *tail;   /* last skb in the list */
  64                 unsigned long  age;     /* jiffies when flow was emptied, for gc */
  65         };
  66         struct rb_node  fq_node;        /* anchor in fq_root[] trees */
  67         struct sock     *sk;
  68         int             qlen;           /* number of packets in flow queue */
  69         int             credit;
  70         u32             socket_hash;    /* sk_hash */
  71         struct fq_flow *next;           /* next pointer in RR lists, or &detached */
  72
  73         struct rb_node  rate_node;      /* anchor in q->delayed tree */
  74         u64             time_next_packet;
  75 };
  76
  77 struct fq_flow_head {
  78         struct fq_flow *first;
  79         struct fq_flow *last;
  80 };
  81
  82 struct fq_sched_data {
  83         struct fq_flow_head new_flows;
  84
  85         struct fq_flow_head old_flows;
  86
  87         struct rb_root  delayed;        /* for rate limited flows */
  88         u64             time_next_delayed_flow;
  89         unsigned long   unthrottle_latency_ns;
  90
  91         struct fq_flow  internal;       /* for non classified or high prio packets */
  92         u32             quantum;
  93         u32             initial_quantum;
  94         u32             flow_refill_delay;
  95         u32             flow_plimit;    /* max packets per flow */
  96         unsigned long   flow_max_rate;  /* optional max rate per flow */
  97         u32             orphan_mask;    /* mask for orphaned skb */
  98         u32             low_rate_threshold;
  99         struct rb_root  *fq_root;
 100         u8              rate_enable;
 101         u8              fq_trees_log;
 102
 103         u32             flows;
 104         u32             inactive_flows;
 105         u32             throttled_flows;
 106
 107         u64             stat_gc_flows;
 108         u64             stat_internal_packets;
 109         u64             stat_throttled;
 110         u64             stat_flows_plimit;
 111         u64             stat_pkts_too_long;
 112         u64             stat_allocation_errors;
 113         struct qdisc_watchdog watchdog;
 114 };
 115
 116 /* special value to mark a detached flow (not on old/new list) */
 117 static struct fq_flow detached, throttled;
 118
 119 static void fq_flow_set_detached(struct fq_flow *f)
 120 {
 121         f->next = &detached;
 122         f->age = jiffies;
 123 }
 124
 125 static bool fq_flow_is_detached(const struct fq_flow *f)
 126 {
 127         return f->next == &detached;
 128 }
 129
 130 static bool fq_flow_is_throttled(const struct fq_flow *f)
 131 {
 132         return f->next == &throttled;
 133 }
 134
 135 static void fq_flow_add_tail(struct fq_flow_head *head, struct fq_flow *flow)
 136 {
 137         if (head->first)
 138                 head->last->next = flow;
 139         else
 140                 head->first = flow;
 141         head->last = flow;
 142         flow->next = NULL;
 143 }
 144
 145 static void fq_flow_unset_throttled(struct fq_sched_data *q, struct fq_flow *f)
 146 {
 147         rb_erase(&f->rate_node, &q->delayed);
 148         q->throttled_flows--;
 149         fq_flow_add_tail(&q->old_flows, f);
 150 }
 151
 152 static void fq_flow_set_throttled(struct fq_sched_data *q, struct fq_flow *f)
 153 {
 154         struct rb_node **p = &q->delayed.rb_node, *parent = NULL;
 155
 156         while (*p) {
 157                 struct fq_flow *aux;
 158
 159                 parent = *p;
 160                 aux = rb_entry(parent, struct fq_flow, rate_node);
 161                 if (f->time_next_packet >= aux->time_next_packet)
 162                         p = &parent->rb_right;
 163                 else
 164                         p = &parent->rb_left;
 165         }
 166         rb_link_node(&f->rate_node, parent, p);
 167         rb_insert_color(&f->rate_node, &q->delayed);
 168         q->throttled_flows++;
 169         q->stat_throttled++;
 170
 171         f->next = &throttled;
 172         if (q->time_next_delayed_flow > f->time_next_packet)
 173                 q->time_next_delayed_flow = f->time_next_packet;
 174 }
 175
 176
 177 static struct kmem_cache *fq_flow_cachep __read_mostly;
 178
 179
 180 /* limit number of collected flows per round */
 181 #define FQ_GC_MAX 8
 182 #define FQ_GC_AGE (3*HZ)
 183
 184 static bool fq_gc_candidate(const struct fq_flow *f)
 185 {
 186         return fq_flow_is_detached(f) &&
 187                time_after(jiffies, f->age + FQ_GC_AGE);
 188 }
 189
 190 static void fq_gc(struct fq_sched_data *q,
 191                   struct rb_root *root,
 192                   struct sock *sk)
 193 {
 194         struct fq_flow *f, *tofree[FQ_GC_MAX];
 195         struct rb_node **p, *parent;
 196         int fcnt = 0;
 197
 198         p = &root->rb_node;
 199         parent = NULL;
 200         while (*p) {
 201                 parent = *p;
 202
 203                 f = rb_entry(parent, struct fq_flow, fq_node);
 204                 if (f->sk == sk)
 205                         break;
 206
 207                 if (fq_gc_candidate(f)) {
 208                         tofree[fcnt++] = f;
 209                         if (fcnt == FQ_GC_MAX)
 210                                 break;
 211                 }
 212
 213                 if (f->sk > sk)
 214                         p = &parent->rb_right;
 215                 else
 216                         p = &parent->rb_left;
 217         }
 218
 219         q->flows -= fcnt;
 220         q->inactive_flows -= fcnt;
 221         q->stat_gc_flows += fcnt;
 222         while (fcnt) {
 223                 struct fq_flow *f = tofree[--fcnt];
 224
 225                 rb_erase(&f->fq_node, root);
 226                 kmem_cache_free(fq_flow_cachep, f);
 227         }
 228 }
 229
 230 static struct fq_flow *fq_classify(struct sk_buff *skb, struct fq_sched_data *q)
 231 {
 232         struct rb_node **p, *parent;
 233         struct sock *sk = skb->sk;
 234         struct rb_root *root;
 235         struct fq_flow *f;
 236
 237         /* warning: no starvation prevention... */
 238         if (unlikely((skb->priority & TC_PRIO_MAX) == TC_PRIO_CONTROL))
 239                 return &q->internal;
 240
 241         /* SYNACK messages are attached to a TCP_NEW_SYN_RECV request socket
 242          * or a listener (SYNCOOKIE mode)
 243          * 1) request sockets are not full blown,
 244          *    they do not contain sk_pacing_rate
 245          * 2) They are not part of a 'flow' yet
 246          * 3) We do not want to rate limit them (eg SYNFLOOD attack),
 247          *    especially if the listener set SO_MAX_PACING_RATE
 248          * 4) We pretend they are orphaned
 249          */
 250         if (!sk || sk_listener(sk)) {
 251                 unsigned long hash = skb_get_hash(skb) & q->orphan_mask;
 252
 253                 /* By forcing low order bit to 1, we make sure to not
 254                  * collide with a local flow (socket pointers are word aligned)
 255                  */
 256                 sk = (struct sock *)((hash << 1) | 1UL);
 257                 skb_orphan(skb);
 258         }
 259
 260         root = &q->fq_root[hash_ptr(sk, q->fq_trees_log)];
 261
 262         if (q->flows >= (2U << q->fq_trees_log) &&
 263             q->inactive_flows > q->flows/2)
 264                 fq_gc(q, root, sk);
 265
 266         p = &root->rb_node;
 267         parent = NULL;
 268         while (*p) {
 269                 parent = *p;
 270
 271                 f = rb_entry(parent, struct fq_flow, fq_node);
 272                 if (f->sk == sk) {
 273                         /* socket might have been reallocated, so check
 274                          * if its sk_hash is the same.
 275                          * It not, we need to refill credit with
 276                          * initial quantum
 277                          */
 278                         if (unlikely(skb->sk &&
 279                                      f->socket_hash != sk->sk_hash)) {
 280                                 f->credit = q->initial_quantum;
 281                                 f->socket_hash = sk->sk_hash;
 282                                 if (fq_flow_is_throttled(f))
 283                                         fq_flow_unset_throttled(q, f);
 284                                 f->time_next_packet = 0ULL;
 285                         }
 286                         return f;
 287                 }
 288                 if (f->sk > sk)
 289                         p = &parent->rb_right;
 290                 else
 291                         p = &parent->rb_left;
 292         }
 293
 294         f = kmem_cache_zalloc(fq_flow_cachep, GFP_ATOMIC | __GFP_NOWARN);
 295         if (unlikely(!f)) {
 296                 q->stat_allocation_errors++;
 297                 return &q->internal;
 298         }
 299         fq_flow_set_detached(f);
 300         f->sk = sk;
 301         if (skb->sk)
 302                 f->socket_hash = sk->sk_hash;
 303         f->credit = q->initial_quantum;
 304
 305         rb_link_node(&f->fq_node, parent, p);
 306         rb_insert_color(&f->fq_node, root);
 307
 308         q->flows++;
 309         q->inactive_flows++;
 310         return f;
 311 }
 312
 313
 314 /* remove one skb from head of flow queue */
 315 static struct sk_buff *fq_dequeue_head(struct Qdisc *sch, struct fq_flow *flow)
 316 {
 317         struct sk_buff *skb = flow->head;
 318
 319         if (skb) {
 320                 flow->head = skb->next;
 321                 skb_mark_not_on_list(skb);
 322                 flow->qlen--;
 323                 qdisc_qstats_backlog_dec(sch, skb);
 324                 sch->q.qlen--;
 325         }
 326         return skb;
 327 }
 328
 329 static void flow_queue_add(struct fq_flow *flow, struct sk_buff *skb)
 330 {
 331         struct sk_buff *head = flow->head;
 332
 333         skb->next = NULL;
 334         if (!head)
 335                 flow->head = skb;
 336         else
 337                 flow->tail->next = skb;
 338
 339         flow->tail = skb;
 340 }
 341
 342 static int fq_enqueue(struct sk_buff *skb, struct Qdisc *sch,
 343                       struct sk_buff **to_free)
 344 {
 345         struct fq_sched_data *q = qdisc_priv(sch);
 346         struct fq_flow *f;
 347
 348         if (unlikely(sch->q.qlen >= sch->limit))
 349                 return qdisc_drop(skb, sch, to_free);
 350
 351         f = fq_classify(skb, q);
 352         if (unlikely(f->qlen >= q->flow_plimit && f != &q->internal)) {
 353                 q->stat_flows_plimit++;
 354                 return qdisc_drop(skb, sch, to_free);
 355         }
 356
 357         f->qlen++;
 358         qdisc_qstats_backlog_inc(sch, skb);
 359         if (fq_flow_is_detached(f)) {
 360                 struct sock *sk = skb->sk;
 361
 362                 fq_flow_add_tail(&q->new_flows, f);
 363                 if (time_after(jiffies, f->age + q->flow_refill_delay))
 364                         f->credit = max_t(u32, f->credit, q->quantum);
 365                 if (sk && q->rate_enable) {
 366                         if (unlikely(smp_load_acquire(&sk->sk_pacing_status) !=
 367                                      SK_PACING_FQ))
 368                                 smp_store_release(&sk->sk_pacing_status,
 369                                                   SK_PACING_FQ);
 370                 }
 371                 q->inactive_flows--;
 372         }
 373
 374         /* Note: this overwrites f->age */
 375         flow_queue_add(f, skb);
 376
 377         if (unlikely(f == &q->internal)) {
 378                 q->stat_internal_packets++;
 379         }
 380         sch->q.qlen++;
 381
 382         return NET_XMIT_SUCCESS;
 383 }
 384
 385 static void fq_check_throttled(struct fq_sched_data *q, u64 now)
 386 {
 387         unsigned long sample;
 388         struct rb_node *p;
 389
 390         if (q->time_next_delayed_flow > now)
 391                 return;
 392
 393         /* Update unthrottle latency EWMA.
 394          * This is cheap and can help diagnosing timer/latency problems.
 395          */
 396         sample = (unsigned long)(now - q->time_next_delayed_flow);
 397         q->unthrottle_latency_ns -= q->unthrottle_latency_ns >> 3;
 398         q->unthrottle_latency_ns += sample >> 3;
 399
 400         q->time_next_delayed_flow = ~0ULL;
 401         while ((p = rb_first(&q->delayed)) != NULL) {
 402                 struct fq_flow *f = rb_entry(p, struct fq_flow, rate_node);
 403
 404                 if (f->time_next_packet > now) {
 405                         q->time_next_delayed_flow = f->time_next_packet;
 406                         break;
 407                 }
 408                 fq_flow_unset_throttled(q, f);
 409         }
 410 }
 411
 412 static struct sk_buff *fq_dequeue(struct Qdisc *sch)
 413 {
 414         struct fq_sched_data *q = qdisc_priv(sch);
 415         u64 now = ktime_get_ns();
 416         struct fq_flow_head *head;
 417         struct sk_buff *skb;
 418         struct fq_flow *f;
 419         unsigned long rate;
 420         u32 plen;
 421
 422         skb = fq_dequeue_head(sch, &q->internal);
 423         if (skb)
 424                 goto out;
 425         fq_check_throttled(q, now);
 426 begin:
 427         head = &q->new_flows;
 428         if (!head->first) {
 429                 head = &q->old_flows;
 430                 if (!head->first) {
 431                         if (q->time_next_delayed_flow != ~0ULL)
 432                                 qdisc_watchdog_schedule_ns(&q->watchdog,
 433                                                            q->time_next_delayed_flow);
 434                         return NULL;
 435                 }
 436         }
 437         f = head->first;
 438
 439         if (f->credit <= 0) {
 440                 f->credit += q->quantum;
 441                 head->first = f->next;
 442                 fq_flow_add_tail(&q->old_flows, f);
 443                 goto begin;
 444         }
 445
 446         skb = f->head;
 447         if (skb) {
 448                 u64 time_next_packet = max_t(u64, ktime_to_ns(skb->tstamp),
 449                                              f->time_next_packet);
 450
 451                 if (now < time_next_packet) {
 452                         head->first = f->next;
 453                         f->time_next_packet = time_next_packet;
 454                         fq_flow_set_throttled(q, f);
 455                         goto begin;
 456                 }
 457         }
 458
 459         skb = fq_dequeue_head(sch, f);
 460         if (!skb) {
 461                 head->first = f->next;
 462                 /* force a pass through old_flows to prevent starvation */
 463                 if ((head == &q->new_flows) && q->old_flows.first) {
 464                         fq_flow_add_tail(&q->old_flows, f);
 465                 } else {
 466                         fq_flow_set_detached(f);
 467                         q->inactive_flows++;
 468                 }
 469                 goto begin;
 470         }
 471         prefetch(&skb->end);
 472         f->credit -= qdisc_pkt_len(skb);
 473
 474         if (ktime_to_ns(skb->tstamp) || !q->rate_enable)
 475                 goto out;
 476
 477         rate = q->flow_max_rate;
 478         if (skb->sk)
 479                 rate = min(skb->sk->sk_pacing_rate, rate);
 480
 481         if (rate <= q->low_rate_threshold) {
 482                 f->credit = 0;
 483                 plen = qdisc_pkt_len(skb);
 484         } else {
 485                 plen = max(qdisc_pkt_len(skb), q->quantum);
 486                 if (f->credit > 0)
 487                         goto out;
 488         }
 489         if (rate != ~0UL) {
 490                 u64 len = (u64)plen * NSEC_PER_SEC;
 491
 492                 if (likely(rate))
 493                         len = div64_ul(len, rate);
 494                 /* Since socket rate can change later,
 495                  * clamp the delay to 1 second.
 496                  * Really, providers of too big packets should be fixed !
 497                  */
 498                 if (unlikely(len > NSEC_PER_SEC)) {
 499                         len = NSEC_PER_SEC;
 500                         q->stat_pkts_too_long++;
 501                 }
 502                 /* Account for schedule/timers drifts.
 503                  * f->time_next_packet was set when prior packet was sent,
 504                  * and current time (@now) can be too late by tens of us.
 505                  */
 506                 if (f->time_next_packet)
 507                         len -= min(len/2, now - f->time_next_packet);
 508                 f->time_next_packet = now + len;
 509         }
 510 out:
 511         qdisc_bstats_update(sch, skb);
 512         return skb;
 513 }
 514
 515 static void fq_flow_purge(struct fq_flow *flow)
 516 {
 517         rtnl_kfree_skbs(flow->head, flow->tail);
 518         flow->head = NULL;
 519         flow->qlen = 0;
 520 }
 521
 522 static void fq_reset(struct Qdisc *sch)
 523 {
 524         struct fq_sched_data *q = qdisc_priv(sch);
 525         struct rb_root *root;
 526         struct rb_node *p;
 527         struct fq_flow *f;
 528         unsigned int idx;
 529
 530         sch->q.qlen = 0;
 531         sch->qstats.backlog = 0;
 532
 533         fq_flow_purge(&q->internal);
 534
 535         if (!q->fq_root)
 536                 return;
 537
 538         for (idx = 0; idx < (1U << q->fq_trees_log); idx++) {
 539                 root = &q->fq_root[idx];
 540                 while ((p = rb_first(root)) != NULL) {
 541                         f = rb_entry(p, struct fq_flow, fq_node);
 542                         rb_erase(p, root);
 543
 544                         fq_flow_purge(f);
 545
 546                         kmem_cache_free(fq_flow_cachep, f);
 547                 }
 548         }
 549         q->new_flows.first      = NULL;
 550         q->old_flows.first      = NULL;
 551         q->delayed              = RB_ROOT;
 552         q->flows                = 0;
 553         q->inactive_flows       = 0;
 554         q->throttled_flows      = 0;
 555 }
 556
 557 static void fq_rehash(struct fq_sched_data *q,
 558                       struct rb_root *old_array, u32 old_log,
 559                       struct rb_root *new_array, u32 new_log)
 560 {
 561         struct rb_node *op, **np, *parent;
 562         struct rb_root *oroot, *nroot;
 563         struct fq_flow *of, *nf;
 564         int fcnt = 0;
 565         u32 idx;
 566
 567         for (idx = 0; idx < (1U << old_log); idx++) {
 568                 oroot = &old_array[idx];
 569                 while ((op = rb_first(oroot)) != NULL) {
 570                         rb_erase(op, oroot);
 571                         of = rb_entry(op, struct fq_flow, fq_node);
 572                         if (fq_gc_candidate(of)) {
 573                                 fcnt++;
 574                                 kmem_cache_free(fq_flow_cachep, of);
 575                                 continue;
 576                         }
 577                         nroot = &new_array[hash_ptr(of->sk, new_log)];
 578
 579                         np = &nroot->rb_node;
 580                         parent = NULL;
 581                         while (*np) {
 582                                 parent = *np;
 583
 584                                 nf = rb_entry(parent, struct fq_flow, fq_node);
 585                                 BUG_ON(nf->sk == of->sk);
 586
 587                                 if (nf->sk > of->sk)
 588                                         np = &parent->rb_right;
 589                                 else
 590                                         np = &parent->rb_left;
 591                         }
 592
 593                         rb_link_node(&of->fq_node, parent, np);
 594                         rb_insert_color(&of->fq_node, nroot);
 595                 }
 596         }
 597         q->flows -= fcnt;
 598         q->inactive_flows -= fcnt;
 599         q->stat_gc_flows += fcnt;
 600 }
 601
 602 static void fq_free(void *addr)
 603 {
 604         kvfree(addr);
 605 }
 606
 607 static int fq_resize(struct Qdisc *sch, u32 log)
 608 {
 609         struct fq_sched_data *q = qdisc_priv(sch);
 610         struct rb_root *array;
 611         void *old_fq_root;
 612         u32 idx;
 613
 614         if (q->fq_root && log == q->fq_trees_log)
 615                 return 0;
 616
 617         /* If XPS was setup, we can allocate memory on right NUMA node */
 618         array = kvmalloc_node(sizeof(struct rb_root) << log, GFP_KERNEL | __GFP_RETRY_MAYFAIL,
 619                               netdev_queue_numa_node_read(sch->dev_queue));
 620         if (!array)
 621                 return -ENOMEM;
 622
 623         for (idx = 0; idx < (1U << log); idx++)
 624                 array[idx] = RB_ROOT;
 625
 626         sch_tree_lock(sch);
 627
 628         old_fq_root = q->fq_root;
 629         if (old_fq_root)
 630                 fq_rehash(q, old_fq_root, q->fq_trees_log, array, log);
 631
 632         q->fq_root = array;
 633         q->fq_trees_log = log;
 634
 635         sch_tree_unlock(sch);
 636
 637         fq_free(old_fq_root);
 638
 639         return 0;
 640 }
 641
 642 static const struct nla_policy fq_policy[TCA_FQ_MAX + 1] = {
 643         [TCA_FQ_PLIMIT]                 = { .type = NLA_U32 },
 644         [TCA_FQ_FLOW_PLIMIT]            = { .type = NLA_U32 },
 645         [TCA_FQ_QUANTUM]                = { .type = NLA_U32 },
 646         [TCA_FQ_INITIAL_QUANTUM]        = { .type = NLA_U32 },
 647         [TCA_FQ_RATE_ENABLE]            = { .type = NLA_U32 },
 648         [TCA_FQ_FLOW_DEFAULT_RATE]      = { .type = NLA_U32 },
 649         [TCA_FQ_FLOW_MAX_RATE]          = { .type = NLA_U32 },
 650         [TCA_FQ_BUCKETS_LOG]            = { .type = NLA_U32 },
 651         [TCA_FQ_FLOW_REFILL_DELAY]      = { .type = NLA_U32 },
 652         [TCA_FQ_LOW_RATE_THRESHOLD]     = { .type = NLA_U32 },
 653 };
 654
 655 static int fq_change(struct Qdisc *sch, struct nlattr *opt,
 656                      struct netlink_ext_ack *extack)
 657 {
 658         struct fq_sched_data *q = qdisc_priv(sch);
 659         struct nlattr *tb[TCA_FQ_MAX + 1];
 660         int err, drop_count = 0;
 661         unsigned drop_len = 0;
 662         u32 fq_log;
 663
 664         if (!opt)
 665                 return -EINVAL;
 666
 667         err = nla_parse_nested(tb, TCA_FQ_MAX, opt, fq_policy, NULL);
 668         if (err < 0)
 669                 return err;
 670
 671         sch_tree_lock(sch);
 672
 673         fq_log = q->fq_trees_log;
 674
 675         if (tb[TCA_FQ_BUCKETS_LOG]) {
 676                 u32 nval = nla_get_u32(tb[TCA_FQ_BUCKETS_LOG]);
 677
 678                 if (nval >= 1 && nval <= ilog2(256*1024))
 679                         fq_log = nval;
 680                 else
 681                         err = -EINVAL;
 682         }
 683         if (tb[TCA_FQ_PLIMIT])
 684                 sch->limit = nla_get_u32(tb[TCA_FQ_PLIMIT]);
 685
 686         if (tb[TCA_FQ_FLOW_PLIMIT])
 687                 q->flow_plimit = nla_get_u32(tb[TCA_FQ_FLOW_PLIMIT]);
 688
 689         if (tb[TCA_FQ_QUANTUM]) {
 690                 u32 quantum = nla_get_u32(tb[TCA_FQ_QUANTUM]);
 691
 692                 if (quantum > 0)
 693                         q->quantum = quantum;
 694                 else
 695                         err = -EINVAL;
 696         }
 697
 698         if (tb[TCA_FQ_INITIAL_QUANTUM])
 699                 q->initial_quantum = nla_get_u32(tb[TCA_FQ_INITIAL_QUANTUM]);
 700
 701         if (tb[TCA_FQ_FLOW_DEFAULT_RATE])
 702                 pr_warn_ratelimited("sch_fq: defrate %u ignored.\n",
 703                                     nla_get_u32(tb[TCA_FQ_FLOW_DEFAULT_RATE]));
 704
 705         if (tb[TCA_FQ_FLOW_MAX_RATE]) {
 706                 u32 rate = nla_get_u32(tb[TCA_FQ_FLOW_MAX_RATE]);
 707
 708                 q->flow_max_rate = (rate == ~0U) ? ~0UL : rate;
 709         }
 710         if (tb[TCA_FQ_LOW_RATE_THRESHOLD])
 711                 q->low_rate_threshold =
 712                         nla_get_u32(tb[TCA_FQ_LOW_RATE_THRESHOLD]);
 713
 714         if (tb[TCA_FQ_RATE_ENABLE]) {
 715                 u32 enable = nla_get_u32(tb[TCA_FQ_RATE_ENABLE]);
 716
 717                 if (enable <= 1)
 718                         q->rate_enable = enable;
 719                 else
 720                         err = -EINVAL;
 721         }
 722
 723         if (tb[TCA_FQ_FLOW_REFILL_DELAY]) {
 724                 u32 usecs_delay = nla_get_u32(tb[TCA_FQ_FLOW_REFILL_DELAY]) ;
 725
 726                 q->flow_refill_delay = usecs_to_jiffies(usecs_delay);
 727         }
 728
 729         if (tb[TCA_FQ_ORPHAN_MASK])
 730                 q->orphan_mask = nla_get_u32(tb[TCA_FQ_ORPHAN_MASK]);
 731
 732         if (!err) {
 733                 sch_tree_unlock(sch);
 734                 err = fq_resize(sch, fq_log);
 735                 sch_tree_lock(sch);
 736         }
 737         while (sch->q.qlen > sch->limit) {
 738                 struct sk_buff *skb = fq_dequeue(sch);
 739
 740                 if (!skb)
 741                         break;
 742                 drop_len += qdisc_pkt_len(skb);
 743                 rtnl_kfree_skbs(skb, skb);
 744                 drop_count++;
 745         }
 746         qdisc_tree_reduce_backlog(sch, drop_count, drop_len);
 747
 748         sch_tree_unlock(sch);
 749         return err;
 750 }
 751
 752 static void fq_destroy(struct Qdisc *sch)
 753 {
 754         struct fq_sched_data *q = qdisc_priv(sch);
 755
 756         fq_reset(sch);
 757         fq_free(q->fq_root);
 758         qdisc_watchdog_cancel(&q->watchdog);
 759 }
 760
 761 static int fq_init(struct Qdisc *sch, struct nlattr *opt,
 762                    struct netlink_ext_ack *extack)
 763 {
 764         struct fq_sched_data *q = qdisc_priv(sch);
 765         int err;
 766
 767         sch->limit              = 10000;
 768         q->flow_plimit          = 100;
 769         q->quantum              = 2 * psched_mtu(qdisc_dev(sch));
 770         q->initial_quantum      = 10 * psched_mtu(qdisc_dev(sch));
 771         q->flow_refill_delay    = msecs_to_jiffies(40);
 772         q->flow_max_rate        = ~0UL;
 773         q->time_next_delayed_flow = ~0ULL;
 774         q->rate_enable          = 1;
 775         q->new_flows.first      = NULL;
 776         q->old_flows.first      = NULL;
 777         q->delayed              = RB_ROOT;
 778         q->fq_root              = NULL;
 779         q->fq_trees_log         = ilog2(1024);
 780         q->orphan_mask          = 1024 - 1;
 781         q->low_rate_threshold   = 550000 / 8;
 782         qdisc_watchdog_init_clockid(&q->watchdog, sch, CLOCK_MONOTONIC);
 783
 784         if (opt)
 785                 err = fq_change(sch, opt, extack);
 786         else
 787                 err = fq_resize(sch, q->fq_trees_log);
 788
 789         return err;
 790 }
 791
 792 static int fq_dump(struct Qdisc *sch, struct sk_buff *skb)
 793 {
 794         struct fq_sched_data *q = qdisc_priv(sch);
 795         struct nlattr *opts;
 796
 797         opts = nla_nest_start(skb, TCA_OPTIONS);
 798         if (opts == NULL)
 799                 goto nla_put_failure;
 800
 801         /* TCA_FQ_FLOW_DEFAULT_RATE is not used anymore */
 802
 803         if (nla_put_u32(skb, TCA_FQ_PLIMIT, sch->limit) ||
 804             nla_put_u32(skb, TCA_FQ_FLOW_PLIMIT, q->flow_plimit) ||
 805             nla_put_u32(skb, TCA_FQ_QUANTUM, q->quantum) ||
 806             nla_put_u32(skb, TCA_FQ_INITIAL_QUANTUM, q->initial_quantum) ||
 807             nla_put_u32(skb, TCA_FQ_RATE_ENABLE, q->rate_enable) ||
 808             nla_put_u32(skb, TCA_FQ_FLOW_MAX_RATE,
 809                         min_t(unsigned long, q->flow_max_rate, ~0U)) ||
 810             nla_put_u32(skb, TCA_FQ_FLOW_REFILL_DELAY,
 811                         jiffies_to_usecs(q->flow_refill_delay)) ||
 812             nla_put_u32(skb, TCA_FQ_ORPHAN_MASK, q->orphan_mask) ||
 813             nla_put_u32(skb, TCA_FQ_LOW_RATE_THRESHOLD,
 814                         q->low_rate_threshold) ||
 815             nla_put_u32(skb, TCA_FQ_BUCKETS_LOG, q->fq_trees_log))
 816                 goto nla_put_failure;
 817
 818         return nla_nest_end(skb, opts);
 819
 820 nla_put_failure:
 821         return -1;
 822 }
 823
 824 static int fq_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
 825 {
 826         struct fq_sched_data *q = qdisc_priv(sch);
 827         struct tc_fq_qd_stats st;
 828
 829         sch_tree_lock(sch);
 830
 831         st.gc_flows               = q->stat_gc_flows;
 832         st.highprio_packets       = q->stat_internal_packets;
 833         st.tcp_retrans            = 0;
 834         st.throttled              = q->stat_throttled;
 835         st.flows_plimit           = q->stat_flows_plimit;
 836         st.pkts_too_long          = q->stat_pkts_too_long;
 837         st.allocation_errors      = q->stat_allocation_errors;
 838         st.time_next_delayed_flow = q->time_next_delayed_flow - ktime_get_ns();
 839         st.flows                  = q->flows;
 840         st.inactive_flows         = q->inactive_flows;
 841         st.throttled_flows        = q->throttled_flows;
 842         st.unthrottle_latency_ns  = min_t(unsigned long,
 843                                           q->unthrottle_latency_ns, ~0U);
 844         sch_tree_unlock(sch);
 845
 846         return gnet_stats_copy_app(d, &st, sizeof(st));
 847 }
 848
 849 static struct Qdisc_ops fq_qdisc_ops __read_mostly = {
 850         .id             =       "fq",
 851         .priv_size      =       sizeof(struct fq_sched_data),
 852
 853         .enqueue        =       fq_enqueue,
 854         .dequeue        =       fq_dequeue,
 855         .peek           =       qdisc_peek_dequeued,
 856         .init           =       fq_init,
 857         .reset          =       fq_reset,
 858         .destroy        =       fq_destroy,
 859         .change         =       fq_change,
 860         .dump           =       fq_dump,
 861         .dump_stats     =       fq_dump_stats,
 862         .owner          =       THIS_MODULE,
 863 };
 864
 865 static int __init fq_module_init(void)
 866 {
 867         int ret;
 868
 869         fq_flow_cachep = kmem_cache_create("fq_flow_cache",
 870                                            sizeof(struct fq_flow),
 871                                            0, 0, NULL);
 872         if (!fq_flow_cachep)
 873                 return -ENOMEM;
 874
 875         ret = register_qdisc(&fq_qdisc_ops);
 876         if (ret)
 877                 kmem_cache_destroy(fq_flow_cachep);
 878         return ret;
 879 }
 880
 881 static void __exit fq_module_exit(void)
 882 {
 883         unregister_qdisc(&fq_qdisc_ops);
 884         kmem_cache_destroy(fq_flow_cachep);
 885 }
 886
 887 module_init(fq_module_init)
 888 module_exit(fq_module_exit)
 889 MODULE_AUTHOR("Eric Dumazet");
 890 MODULE_LICENSE("GPL");