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 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 <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
  90         struct fq_flow  internal;       /* for non classified or high prio packets */
  91         u32             quantum;
  92         u32             initial_quantum;
  93         u32             flow_refill_delay;
  94         u32             flow_max_rate;  /* optional max rate per flow */
  95         u32             flow_plimit;    /* max packets per flow */
  96         struct rb_root  *fq_root;
  97         u8              rate_enable;
  98         u8              fq_trees_log;
  99
 100         u32             flows;
 101         u32             inactive_flows;
 102         u32             throttled_flows;
 103
 104         u64             stat_gc_flows;
 105         u64             stat_internal_packets;
 106         u64             stat_tcp_retrans;
 107         u64             stat_throttled;
 108         u64             stat_flows_plimit;
 109         u64             stat_pkts_too_long;
 110         u64             stat_allocation_errors;
 111         struct qdisc_watchdog watchdog;
 112 };
 113
 114 /* special value to mark a detached flow (not on old/new list) */
 115 static struct fq_flow detached, throttled;
 116
 117 static void fq_flow_set_detached(struct fq_flow *f)
 118 {
 119         f->next = &detached;
 120         f->age = jiffies;
 121 }
 122
 123 static bool fq_flow_is_detached(const struct fq_flow *f)
 124 {
 125         return f->next == &detached;
 126 }
 127
 128 static void fq_flow_set_throttled(struct fq_sched_data *q, struct fq_flow *f)
 129 {
 130         struct rb_node **p = &q->delayed.rb_node, *parent = NULL;
 131
 132         while (*p) {
 133                 struct fq_flow *aux;
 134
 135                 parent = *p;
 136                 aux = container_of(parent, struct fq_flow, rate_node);
 137                 if (f->time_next_packet >= aux->time_next_packet)
 138                         p = &parent->rb_right;
 139                 else
 140                         p = &parent->rb_left;
 141         }
 142         rb_link_node(&f->rate_node, parent, p);
 143         rb_insert_color(&f->rate_node, &q->delayed);
 144         q->throttled_flows++;
 145         q->stat_throttled++;
 146
 147         f->next = &throttled;
 148         if (q->time_next_delayed_flow > f->time_next_packet)
 149                 q->time_next_delayed_flow = f->time_next_packet;
 150 }
 151
 152
 153 static struct kmem_cache *fq_flow_cachep __read_mostly;
 154
 155 static void fq_flow_add_tail(struct fq_flow_head *head, struct fq_flow *flow)
 156 {
 157         if (head->first)
 158                 head->last->next = flow;
 159         else
 160                 head->first = flow;
 161         head->last = flow;
 162         flow->next = NULL;
 163 }
 164
 165 /* limit number of collected flows per round */
 166 #define FQ_GC_MAX 8
 167 #define FQ_GC_AGE (3*HZ)
 168
 169 static bool fq_gc_candidate(const struct fq_flow *f)
 170 {
 171         return fq_flow_is_detached(f) &&
 172                time_after(jiffies, f->age + FQ_GC_AGE);
 173 }
 174
 175 static void fq_gc(struct fq_sched_data *q,
 176                   struct rb_root *root,
 177                   struct sock *sk)
 178 {
 179         struct fq_flow *f, *tofree[FQ_GC_MAX];
 180         struct rb_node **p, *parent;
 181         int fcnt = 0;
 182
 183         p = &root->rb_node;
 184         parent = NULL;
 185         while (*p) {
 186                 parent = *p;
 187
 188                 f = container_of(parent, struct fq_flow, fq_node);
 189                 if (f->sk == sk)
 190                         break;
 191
 192                 if (fq_gc_candidate(f)) {
 193                         tofree[fcnt++] = f;
 194                         if (fcnt == FQ_GC_MAX)
 195                                 break;
 196                 }
 197
 198                 if (f->sk > sk)
 199                         p = &parent->rb_right;
 200                 else
 201                         p = &parent->rb_left;
 202         }
 203
 204         q->flows -= fcnt;
 205         q->inactive_flows -= fcnt;
 206         q->stat_gc_flows += fcnt;
 207         while (fcnt) {
 208                 struct fq_flow *f = tofree[--fcnt];
 209
 210                 rb_erase(&f->fq_node, root);
 211                 kmem_cache_free(fq_flow_cachep, f);
 212         }
 213 }
 214
 215 static struct fq_flow *fq_classify(struct sk_buff *skb, struct fq_sched_data *q)
 216 {
 217         struct rb_node **p, *parent;
 218         struct sock *sk = skb->sk;
 219         struct rb_root *root;
 220         struct fq_flow *f;
 221
 222         /* warning: no starvation prevention... */
 223         if (unlikely((skb->priority & TC_PRIO_MAX) == TC_PRIO_CONTROL))
 224                 return &q->internal;
 225
 226         if (unlikely(!sk)) {
 227                 /* By forcing low order bit to 1, we make sure to not
 228                  * collide with a local flow (socket pointers are word aligned)
 229                  */
 230                 sk = (struct sock *)(skb_get_hash(skb) | 1L);
 231         }
 232
 233         root = &q->fq_root[hash_32((u32)(long)sk, q->fq_trees_log)];
 234
 235         if (q->flows >= (2U << q->fq_trees_log) &&
 236             q->inactive_flows > q->flows/2)
 237                 fq_gc(q, root, sk);
 238
 239         p = &root->rb_node;
 240         parent = NULL;
 241         while (*p) {
 242                 parent = *p;
 243
 244                 f = container_of(parent, struct fq_flow, fq_node);
 245                 if (f->sk == sk) {
 246                         /* socket might have been reallocated, so check
 247                          * if its sk_hash is the same.
 248                          * It not, we need to refill credit with
 249                          * initial quantum
 250                          */
 251                         if (unlikely(skb->sk &&
 252                                      f->socket_hash != sk->sk_hash)) {
 253                                 f->credit = q->initial_quantum;
 254                                 f->socket_hash = sk->sk_hash;
 255                                 f->time_next_packet = 0ULL;
 256                         }
 257                         return f;
 258                 }
 259                 if (f->sk > sk)
 260                         p = &parent->rb_right;
 261                 else
 262                         p = &parent->rb_left;
 263         }
 264
 265         f = kmem_cache_zalloc(fq_flow_cachep, GFP_ATOMIC | __GFP_NOWARN);
 266         if (unlikely(!f)) {
 267                 q->stat_allocation_errors++;
 268                 return &q->internal;
 269         }
 270         fq_flow_set_detached(f);
 271         f->sk = sk;
 272         if (skb->sk)
 273                 f->socket_hash = sk->sk_hash;
 274         f->credit = q->initial_quantum;
 275
 276         rb_link_node(&f->fq_node, parent, p);
 277         rb_insert_color(&f->fq_node, root);
 278
 279         q->flows++;
 280         q->inactive_flows++;
 281         return f;
 282 }
 283
 284
 285 /* remove one skb from head of flow queue */
 286 static struct sk_buff *fq_dequeue_head(struct Qdisc *sch, struct fq_flow *flow)
 287 {
 288         struct sk_buff *skb = flow->head;
 289
 290         if (skb) {
 291                 flow->head = skb->next;
 292                 skb->next = NULL;
 293                 flow->qlen--;
 294                 qdisc_qstats_backlog_dec(sch, skb);
 295                 sch->q.qlen--;
 296         }
 297         return skb;
 298 }
 299
 300 /* We might add in the future detection of retransmits
 301  * For the time being, just return false
 302  */
 303 static bool skb_is_retransmit(struct sk_buff *skb)
 304 {
 305         return false;
 306 }
 307
 308 /* add skb to flow queue
 309  * flow queue is a linked list, kind of FIFO, except for TCP retransmits
 310  * We special case tcp retransmits to be transmitted before other packets.
 311  * We rely on fact that TCP retransmits are unlikely, so we do not waste
 312  * a separate queue or a pointer.
 313  * head->  [retrans pkt 1]
 314  *         [retrans pkt 2]
 315  *         [ normal pkt 1]
 316  *         [ normal pkt 2]
 317  *         [ normal pkt 3]
 318  * tail->  [ normal pkt 4]
 319  */
 320 static void flow_queue_add(struct fq_flow *flow, struct sk_buff *skb)
 321 {
 322         struct sk_buff *prev, *head = flow->head;
 323
 324         skb->next = NULL;
 325         if (!head) {
 326                 flow->head = skb;
 327                 flow->tail = skb;
 328                 return;
 329         }
 330         if (likely(!skb_is_retransmit(skb))) {
 331                 flow->tail->next = skb;
 332                 flow->tail = skb;
 333                 return;
 334         }
 335
 336         /* This skb is a tcp retransmit,
 337          * find the last retrans packet in the queue
 338          */
 339         prev = NULL;
 340         while (skb_is_retransmit(head)) {
 341                 prev = head;
 342                 head = head->next;
 343                 if (!head)
 344                         break;
 345         }
 346         if (!prev) { /* no rtx packet in queue, become the new head */
 347                 skb->next = flow->head;
 348                 flow->head = skb;
 349         } else {
 350                 if (prev == flow->tail)
 351                         flow->tail = skb;
 352                 else
 353                         skb->next = prev->next;
 354                 prev->next = skb;
 355         }
 356 }
 357
 358 static int fq_enqueue(struct sk_buff *skb, struct Qdisc *sch)
 359 {
 360         struct fq_sched_data *q = qdisc_priv(sch);
 361         struct fq_flow *f;
 362
 363         if (unlikely(sch->q.qlen >= sch->limit))
 364                 return qdisc_drop(skb, sch);
 365
 366         f = fq_classify(skb, q);
 367         if (unlikely(f->qlen >= q->flow_plimit && f != &q->internal)) {
 368                 q->stat_flows_plimit++;
 369                 return qdisc_drop(skb, sch);
 370         }
 371
 372         f->qlen++;
 373         if (skb_is_retransmit(skb))
 374                 q->stat_tcp_retrans++;
 375         qdisc_qstats_backlog_inc(sch, skb);
 376         if (fq_flow_is_detached(f)) {
 377                 fq_flow_add_tail(&q->new_flows, f);
 378                 if (time_after(jiffies, f->age + q->flow_refill_delay))
 379                         f->credit = max_t(u32, f->credit, q->quantum);
 380                 q->inactive_flows--;
 381         }
 382
 383         /* Note: this overwrites f->age */
 384         flow_queue_add(f, skb);
 385
 386         if (unlikely(f == &q->internal)) {
 387                 q->stat_internal_packets++;
 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_get_ns();
 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                                                            false);
 437                         return NULL;
 438                 }
 439         }
 440         f = head->first;
 441
 442         if (f->credit <= 0) {
 443                 f->credit += q->quantum;
 444                 head->first = f->next;
 445                 fq_flow_add_tail(&q->old_flows, f);
 446                 goto begin;
 447         }
 448
 449         skb = f->head;
 450         if (unlikely(skb && now < f->time_next_packet &&
 451                      !skb_is_tcp_pure_ack(skb))) {
 452                 head->first = f->next;
 453                 fq_flow_set_throttled(q, f);
 454                 goto begin;
 455         }
 456
 457         skb = fq_dequeue_head(sch, f);
 458         if (!skb) {
 459                 head->first = f->next;
 460                 /* force a pass through old_flows to prevent starvation */
 461                 if ((head == &q->new_flows) && q->old_flows.first) {
 462                         fq_flow_add_tail(&q->old_flows, f);
 463                 } else {
 464                         fq_flow_set_detached(f);
 465                         q->inactive_flows++;
 466                 }
 467                 goto begin;
 468         }
 469         prefetch(&skb->end);
 470         f->credit -= qdisc_pkt_len(skb);
 471
 472         if (f->credit > 0 || !q->rate_enable)
 473                 goto out;
 474
 475         /* Do not pace locally generated ack packets */
 476         if (skb_is_tcp_pure_ack(skb))
 477                 goto out;
 478
 479         rate = q->flow_max_rate;
 480         if (skb->sk)
 481                 rate = min(skb->sk->sk_pacing_rate, rate);
 482
 483         if (rate != ~0U) {
 484                 u32 plen = max(qdisc_pkt_len(skb), q->quantum);
 485                 u64 len = (u64)plen * NSEC_PER_SEC;
 486
 487                 if (likely(rate))
 488                         do_div(len, rate);
 489                 /* Since socket rate can change later,
 490                  * clamp the delay to 1 second.
 491                  * Really, providers of too big packets should be fixed !
 492                  */
 493                 if (unlikely(len > NSEC_PER_SEC)) {
 494                         len = NSEC_PER_SEC;
 495                         q->stat_pkts_too_long++;
 496                 }
 497
 498                 f->time_next_packet = now + len;
 499         }
 500 out:
 501         qdisc_bstats_update(sch, skb);
 502         return skb;
 503 }
 504
 505 static void fq_reset(struct Qdisc *sch)
 506 {
 507         struct fq_sched_data *q = qdisc_priv(sch);
 508         struct rb_root *root;
 509         struct sk_buff *skb;
 510         struct rb_node *p;
 511         struct fq_flow *f;
 512         unsigned int idx;
 513
 514         while ((skb = fq_dequeue_head(sch, &q->internal)) != NULL)
 515                 kfree_skb(skb);
 516
 517         if (!q->fq_root)
 518                 return;
 519
 520         for (idx = 0; idx < (1U << q->fq_trees_log); idx++) {
 521                 root = &q->fq_root[idx];
 522                 while ((p = rb_first(root)) != NULL) {
 523                         f = container_of(p, struct fq_flow, fq_node);
 524                         rb_erase(p, root);
 525
 526                         while ((skb = fq_dequeue_head(sch, f)) != NULL)
 527                                 kfree_skb(skb);
 528
 529                         kmem_cache_free(fq_flow_cachep, f);
 530                 }
 531         }
 532         q->new_flows.first      = NULL;
 533         q->old_flows.first      = NULL;
 534         q->delayed              = RB_ROOT;
 535         q->flows                = 0;
 536         q->inactive_flows       = 0;
 537         q->throttled_flows      = 0;
 538 }
 539
 540 static void fq_rehash(struct fq_sched_data *q,
 541                       struct rb_root *old_array, u32 old_log,
 542                       struct rb_root *new_array, u32 new_log)
 543 {
 544         struct rb_node *op, **np, *parent;
 545         struct rb_root *oroot, *nroot;
 546         struct fq_flow *of, *nf;
 547         int fcnt = 0;
 548         u32 idx;
 549
 550         for (idx = 0; idx < (1U << old_log); idx++) {
 551                 oroot = &old_array[idx];
 552                 while ((op = rb_first(oroot)) != NULL) {
 553                         rb_erase(op, oroot);
 554                         of = container_of(op, struct fq_flow, fq_node);
 555                         if (fq_gc_candidate(of)) {
 556                                 fcnt++;
 557                                 kmem_cache_free(fq_flow_cachep, of);
 558                                 continue;
 559                         }
 560                         nroot = &new_array[hash_32((u32)(long)of->sk, new_log)];
 561
 562                         np = &nroot->rb_node;
 563                         parent = NULL;
 564                         while (*np) {
 565                                 parent = *np;
 566
 567                                 nf = container_of(parent, struct fq_flow, fq_node);
 568                                 BUG_ON(nf->sk == of->sk);
 569
 570                                 if (nf->sk > of->sk)
 571                                         np = &parent->rb_right;
 572                                 else
 573                                         np = &parent->rb_left;
 574                         }
 575
 576                         rb_link_node(&of->fq_node, parent, np);
 577                         rb_insert_color(&of->fq_node, nroot);
 578                 }
 579         }
 580         q->flows -= fcnt;
 581         q->inactive_flows -= fcnt;
 582         q->stat_gc_flows += fcnt;
 583 }
 584
 585 static void *fq_alloc_node(size_t sz, int node)
 586 {
 587         void *ptr;
 588
 589         ptr = kmalloc_node(sz, GFP_KERNEL | __GFP_REPEAT | __GFP_NOWARN, node);
 590         if (!ptr)
 591                 ptr = vmalloc_node(sz, node);
 592         return ptr;
 593 }
 594
 595 static void fq_free(void *addr)
 596 {
 597         kvfree(addr);
 598 }
 599
 600 static int fq_resize(struct Qdisc *sch, u32 log)
 601 {
 602         struct fq_sched_data *q = qdisc_priv(sch);
 603         struct rb_root *array;
 604         void *old_fq_root;
 605         u32 idx;
 606
 607         if (q->fq_root && log == q->fq_trees_log)
 608                 return 0;
 609
 610         /* If XPS was setup, we can allocate memory on right NUMA node */
 611         array = fq_alloc_node(sizeof(struct rb_root) << log,
 612                               netdev_queue_numa_node_read(sch->dev_queue));
 613         if (!array)
 614                 return -ENOMEM;
 615
 616         for (idx = 0; idx < (1U << log); idx++)
 617                 array[idx] = RB_ROOT;
 618
 619         sch_tree_lock(sch);
 620
 621         old_fq_root = q->fq_root;
 622         if (old_fq_root)
 623                 fq_rehash(q, old_fq_root, q->fq_trees_log, array, log);
 624
 625         q->fq_root = array;
 626         q->fq_trees_log = log;
 627
 628         sch_tree_unlock(sch);
 629
 630         fq_free(old_fq_root);
 631
 632         return 0;
 633 }
 634
 635 static const struct nla_policy fq_policy[TCA_FQ_MAX + 1] = {
 636         [TCA_FQ_PLIMIT]                 = { .type = NLA_U32 },
 637         [TCA_FQ_FLOW_PLIMIT]            = { .type = NLA_U32 },
 638         [TCA_FQ_QUANTUM]                = { .type = NLA_U32 },
 639         [TCA_FQ_INITIAL_QUANTUM]        = { .type = NLA_U32 },
 640         [TCA_FQ_RATE_ENABLE]            = { .type = NLA_U32 },
 641         [TCA_FQ_FLOW_DEFAULT_RATE]      = { .type = NLA_U32 },
 642         [TCA_FQ_FLOW_MAX_RATE]          = { .type = NLA_U32 },
 643         [TCA_FQ_BUCKETS_LOG]            = { .type = NLA_U32 },
 644         [TCA_FQ_FLOW_REFILL_DELAY]      = { .type = NLA_U32 },
 645 };
 646
 647 static int fq_change(struct Qdisc *sch, struct nlattr *opt)
 648 {
 649         struct fq_sched_data *q = qdisc_priv(sch);
 650         struct nlattr *tb[TCA_FQ_MAX + 1];
 651         int err, drop_count = 0;
 652         u32 fq_log;
 653
 654         if (!opt)
 655                 return -EINVAL;
 656
 657         err = nla_parse_nested(tb, TCA_FQ_MAX, opt, fq_policy);
 658         if (err < 0)
 659                 return err;
 660
 661         sch_tree_lock(sch);
 662
 663         fq_log = q->fq_trees_log;
 664
 665         if (tb[TCA_FQ_BUCKETS_LOG]) {
 666                 u32 nval = nla_get_u32(tb[TCA_FQ_BUCKETS_LOG]);
 667
 668                 if (nval >= 1 && nval <= ilog2(256*1024))
 669                         fq_log = nval;
 670                 else
 671                         err = -EINVAL;
 672         }
 673         if (tb[TCA_FQ_PLIMIT])
 674                 sch->limit = nla_get_u32(tb[TCA_FQ_PLIMIT]);
 675
 676         if (tb[TCA_FQ_FLOW_PLIMIT])
 677                 q->flow_plimit = nla_get_u32(tb[TCA_FQ_FLOW_PLIMIT]);
 678
 679         if (tb[TCA_FQ_QUANTUM])
 680                 q->quantum = nla_get_u32(tb[TCA_FQ_QUANTUM]);
 681
 682         if (tb[TCA_FQ_INITIAL_QUANTUM])
 683                 q->initial_quantum = nla_get_u32(tb[TCA_FQ_INITIAL_QUANTUM]);
 684
 685         if (tb[TCA_FQ_FLOW_DEFAULT_RATE])
 686                 pr_warn_ratelimited("sch_fq: defrate %u ignored.\n",
 687                                     nla_get_u32(tb[TCA_FQ_FLOW_DEFAULT_RATE]));
 688
 689         if (tb[TCA_FQ_FLOW_MAX_RATE])
 690                 q->flow_max_rate = nla_get_u32(tb[TCA_FQ_FLOW_MAX_RATE]);
 691
 692         if (tb[TCA_FQ_RATE_ENABLE]) {
 693                 u32 enable = nla_get_u32(tb[TCA_FQ_RATE_ENABLE]);
 694
 695                 if (enable <= 1)
 696                         q->rate_enable = enable;
 697                 else
 698                         err = -EINVAL;
 699         }
 700
 701         if (tb[TCA_FQ_FLOW_REFILL_DELAY]) {
 702                 u32 usecs_delay = nla_get_u32(tb[TCA_FQ_FLOW_REFILL_DELAY]) ;
 703
 704                 q->flow_refill_delay = usecs_to_jiffies(usecs_delay);
 705         }
 706
 707         if (!err) {
 708                 sch_tree_unlock(sch);
 709                 err = fq_resize(sch, fq_log);
 710                 sch_tree_lock(sch);
 711         }
 712         while (sch->q.qlen > sch->limit) {
 713                 struct sk_buff *skb = fq_dequeue(sch);
 714
 715                 if (!skb)
 716                         break;
 717                 kfree_skb(skb);
 718                 drop_count++;
 719         }
 720         qdisc_tree_decrease_qlen(sch, drop_count);
 721
 722         sch_tree_unlock(sch);
 723         return err;
 724 }
 725
 726 static void fq_destroy(struct Qdisc *sch)
 727 {
 728         struct fq_sched_data *q = qdisc_priv(sch);
 729
 730         fq_reset(sch);
 731         fq_free(q->fq_root);
 732         qdisc_watchdog_cancel(&q->watchdog);
 733 }
 734
 735 static int fq_init(struct Qdisc *sch, struct nlattr *opt)
 736 {
 737         struct fq_sched_data *q = qdisc_priv(sch);
 738         int err;
 739
 740         sch->limit              = 10000;
 741         q->flow_plimit          = 100;
 742         q->quantum              = 2 * psched_mtu(qdisc_dev(sch));
 743         q->initial_quantum      = 10 * psched_mtu(qdisc_dev(sch));
 744         q->flow_refill_delay    = msecs_to_jiffies(40);
 745         q->flow_max_rate        = ~0U;
 746         q->rate_enable          = 1;
 747         q->new_flows.first      = NULL;
 748         q->old_flows.first      = NULL;
 749         q->delayed              = RB_ROOT;
 750         q->fq_root              = NULL;
 751         q->fq_trees_log         = ilog2(1024);
 752         qdisc_watchdog_init(&q->watchdog, sch);
 753
 754         if (opt)
 755                 err = fq_change(sch, opt);
 756         else
 757                 err = fq_resize(sch, q->fq_trees_log);
 758
 759         return err;
 760 }
 761
 762 static int fq_dump(struct Qdisc *sch, struct sk_buff *skb)
 763 {
 764         struct fq_sched_data *q = qdisc_priv(sch);
 765         struct nlattr *opts;
 766
 767         opts = nla_nest_start(skb, TCA_OPTIONS);
 768         if (opts == NULL)
 769                 goto nla_put_failure;
 770
 771         /* TCA_FQ_FLOW_DEFAULT_RATE is not used anymore */
 772
 773         if (nla_put_u32(skb, TCA_FQ_PLIMIT, sch->limit) ||
 774             nla_put_u32(skb, TCA_FQ_FLOW_PLIMIT, q->flow_plimit) ||
 775             nla_put_u32(skb, TCA_FQ_QUANTUM, q->quantum) ||
 776             nla_put_u32(skb, TCA_FQ_INITIAL_QUANTUM, q->initial_quantum) ||
 777             nla_put_u32(skb, TCA_FQ_RATE_ENABLE, q->rate_enable) ||
 778             nla_put_u32(skb, TCA_FQ_FLOW_MAX_RATE, q->flow_max_rate) ||
 779             nla_put_u32(skb, TCA_FQ_FLOW_REFILL_DELAY,
 780                         jiffies_to_usecs(q->flow_refill_delay)) ||
 781             nla_put_u32(skb, TCA_FQ_BUCKETS_LOG, q->fq_trees_log))
 782                 goto nla_put_failure;
 783
 784         return nla_nest_end(skb, opts);
 785
 786 nla_put_failure:
 787         return -1;
 788 }
 789
 790 static int fq_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
 791 {
 792         struct fq_sched_data *q = qdisc_priv(sch);
 793         u64 now = ktime_get_ns();
 794         struct tc_fq_qd_stats st = {
 795                 .gc_flows               = q->stat_gc_flows,
 796                 .highprio_packets       = q->stat_internal_packets,
 797                 .tcp_retrans            = q->stat_tcp_retrans,
 798                 .throttled              = q->stat_throttled,
 799                 .flows_plimit           = q->stat_flows_plimit,
 800                 .pkts_too_long          = q->stat_pkts_too_long,
 801                 .allocation_errors      = q->stat_allocation_errors,
 802                 .flows                  = q->flows,
 803                 .inactive_flows         = q->inactive_flows,
 804                 .throttled_flows        = q->throttled_flows,
 805                 .time_next_delayed_flow = q->time_next_delayed_flow - now,
 806         };
 807
 808         return gnet_stats_copy_app(d, &st, sizeof(st));
 809 }
 810
 811 static struct Qdisc_ops fq_qdisc_ops __read_mostly = {
 812         .id             =       "fq",
 813         .priv_size      =       sizeof(struct fq_sched_data),
 814
 815         .enqueue        =       fq_enqueue,
 816         .dequeue        =       fq_dequeue,
 817         .peek           =       qdisc_peek_dequeued,
 818         .init           =       fq_init,
 819         .reset          =       fq_reset,
 820         .destroy        =       fq_destroy,
 821         .change         =       fq_change,
 822         .dump           =       fq_dump,
 823         .dump_stats     =       fq_dump_stats,
 824         .owner          =       THIS_MODULE,
 825 };
 826
 827 static int __init fq_module_init(void)
 828 {
 829         int ret;
 830
 831         fq_flow_cachep = kmem_cache_create("fq_flow_cache",
 832                                            sizeof(struct fq_flow),
 833                                            0, 0, NULL);
 834         if (!fq_flow_cachep)
 835                 return -ENOMEM;
 836
 837         ret = register_qdisc(&fq_qdisc_ops);
 838         if (ret)
 839                 kmem_cache_destroy(fq_flow_cachep);
 840         return ret;
 841 }
 842
 843 static void __exit fq_module_exit(void)
 844 {
 845         unregister_qdisc(&fq_qdisc_ops);
 846         kmem_cache_destroy(fq_flow_cachep);
 847 }
 848
 849 module_init(fq_module_init)
 850 module_exit(fq_module_exit)
 851 MODULE_AUTHOR("Eric Dumazet");
 852 MODULE_LICENSE("GPL");