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
  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         u32             orphan_mask;    /* mask for orphaned skb */
  97         struct rb_root  *fq_root;
  98         u8              rate_enable;
  99         u8              fq_trees_log;
 100
 101         u32             flows;
 102         u32             inactive_flows;
 103         u32             throttled_flows;
 104
 105         u64             stat_gc_flows;
 106         u64             stat_internal_packets;
 107         u64             stat_tcp_retrans;
 108         u64             stat_throttled;
 109         u64             stat_flows_plimit;
 110         u64             stat_pkts_too_long;
 111         u64             stat_allocation_errors;
 112         struct qdisc_watchdog watchdog;
 113 };
 114
 115 /* special value to mark a detached flow (not on old/new list) */
 116 static struct fq_flow detached, throttled;
 117
 118 static void fq_flow_set_detached(struct fq_flow *f)
 119 {
 120         f->next = &detached;
 121         f->age = jiffies;
 122 }
 123
 124 static bool fq_flow_is_detached(const struct fq_flow *f)
 125 {
 126         return f->next == &detached;
 127 }
 128
 129 static bool fq_flow_is_throttled(const struct fq_flow *f)
 130 {
 131         return f->next == &throttled;
 132 }
 133
 134 static void fq_flow_add_tail(struct fq_flow_head *head, struct fq_flow *flow)
 135 {
 136         if (head->first)
 137                 head->last->next = flow;
 138         else
 139                 head->first = flow;
 140         head->last = flow;
 141         flow->next = NULL;
 142 }
 143
 144 static void fq_flow_unset_throttled(struct fq_sched_data *q, struct fq_flow *f)
 145 {
 146         rb_erase(&f->rate_node, &q->delayed);
 147         q->throttled_flows--;
 148         fq_flow_add_tail(&q->old_flows, f);
 149 }
 150
 151 static void fq_flow_set_throttled(struct fq_sched_data *q, struct fq_flow *f)
 152 {
 153         struct rb_node **p = &q->delayed.rb_node, *parent = NULL;
 154
 155         while (*p) {
 156                 struct fq_flow *aux;
 157
 158                 parent = *p;
 159                 aux = container_of(parent, struct fq_flow, rate_node);
 160                 if (f->time_next_packet >= aux->time_next_packet)
 161                         p = &parent->rb_right;
 162                 else
 163                         p = &parent->rb_left;
 164         }
 165         rb_link_node(&f->rate_node, parent, p);
 166         rb_insert_color(&f->rate_node, &q->delayed);
 167         q->throttled_flows++;
 168         q->stat_throttled++;
 169
 170         f->next = &throttled;
 171         if (q->time_next_delayed_flow > f->time_next_packet)
 172                 q->time_next_delayed_flow = f->time_next_packet;
 173 }
 174
 175
 176 static struct kmem_cache *fq_flow_cachep __read_mostly;
 177
 178
 179 /* limit number of collected flows per round */
 180 #define FQ_GC_MAX 8
 181 #define FQ_GC_AGE (3*HZ)
 182
 183 static bool fq_gc_candidate(const struct fq_flow *f)
 184 {
 185         return fq_flow_is_detached(f) &&
 186                time_after(jiffies, f->age + FQ_GC_AGE);
 187 }
 188
 189 static void fq_gc(struct fq_sched_data *q,
 190                   struct rb_root *root,
 191                   struct sock *sk)
 192 {
 193         struct fq_flow *f, *tofree[FQ_GC_MAX];
 194         struct rb_node **p, *parent;
 195         int fcnt = 0;
 196
 197         p = &root->rb_node;
 198         parent = NULL;
 199         while (*p) {
 200                 parent = *p;
 201
 202                 f = container_of(parent, struct fq_flow, fq_node);
 203                 if (f->sk == sk)
 204                         break;
 205
 206                 if (fq_gc_candidate(f)) {
 207                         tofree[fcnt++] = f;
 208                         if (fcnt == FQ_GC_MAX)
 209                                 break;
 210                 }
 211
 212                 if (f->sk > sk)
 213                         p = &parent->rb_right;
 214                 else
 215                         p = &parent->rb_left;
 216         }
 217
 218         q->flows -= fcnt;
 219         q->inactive_flows -= fcnt;
 220         q->stat_gc_flows += fcnt;
 221         while (fcnt) {
 222                 struct fq_flow *f = tofree[--fcnt];
 223
 224                 rb_erase(&f->fq_node, root);
 225                 kmem_cache_free(fq_flow_cachep, f);
 226         }
 227 }
 228
 229 static struct fq_flow *fq_classify(struct sk_buff *skb, struct fq_sched_data *q)
 230 {
 231         struct rb_node **p, *parent;
 232         struct sock *sk = skb->sk;
 233         struct rb_root *root;
 234         struct fq_flow *f;
 235
 236         /* warning: no starvation prevention... */
 237         if (unlikely((skb->priority & TC_PRIO_MAX) == TC_PRIO_CONTROL))
 238                 return &q->internal;
 239
 240         /* SYNACK messages are attached to a TCP_NEW_SYN_RECV request socket
 241          * or a listener (SYNCOOKIE mode)
 242          * 1) request sockets are not full blown,
 243          *    they do not contain sk_pacing_rate
 244          * 2) They are not part of a 'flow' yet
 245          * 3) We do not want to rate limit them (eg SYNFLOOD attack),
 246          *    especially if the listener set SO_MAX_PACING_RATE
 247          * 4) We pretend they are orphaned
 248          */
 249         if (!sk || sk_listener(sk)) {
 250                 unsigned long hash = skb_get_hash(skb) & q->orphan_mask;
 251
 252                 /* By forcing low order bit to 1, we make sure to not
 253                  * collide with a local flow (socket pointers are word aligned)
 254                  */
 255                 sk = (struct sock *)((hash << 1) | 1UL);
 256                 skb_orphan(skb);
 257         }
 258
 259         root = &q->fq_root[hash_32((u32)(long)sk, q->fq_trees_log)];
 260
 261         if (q->flows >= (2U << q->fq_trees_log) &&
 262             q->inactive_flows > q->flows/2)
 263                 fq_gc(q, root, sk);
 264
 265         p = &root->rb_node;
 266         parent = NULL;
 267         while (*p) {
 268                 parent = *p;
 269
 270                 f = container_of(parent, struct fq_flow, fq_node);
 271                 if (f->sk == sk) {
 272                         /* socket might have been reallocated, so check
 273                          * if its sk_hash is the same.
 274                          * It not, we need to refill credit with
 275                          * initial quantum
 276                          */
 277                         if (unlikely(skb->sk &&
 278                                      f->socket_hash != sk->sk_hash)) {
 279                                 f->credit = q->initial_quantum;
 280                                 f->socket_hash = sk->sk_hash;
 281                                 if (fq_flow_is_throttled(f))
 282                                         fq_flow_unset_throttled(q, f);
 283                                 f->time_next_packet = 0ULL;
 284                         }
 285                         return f;
 286                 }
 287                 if (f->sk > sk)
 288                         p = &parent->rb_right;
 289                 else
 290                         p = &parent->rb_left;
 291         }
 292
 293         f = kmem_cache_zalloc(fq_flow_cachep, GFP_ATOMIC | __GFP_NOWARN);
 294         if (unlikely(!f)) {
 295                 q->stat_allocation_errors++;
 296                 return &q->internal;
 297         }
 298         fq_flow_set_detached(f);
 299         f->sk = sk;
 300         if (skb->sk)
 301                 f->socket_hash = sk->sk_hash;
 302         f->credit = q->initial_quantum;
 303
 304         rb_link_node(&f->fq_node, parent, p);
 305         rb_insert_color(&f->fq_node, root);
 306
 307         q->flows++;
 308         q->inactive_flows++;
 309         return f;
 310 }
 311
 312
 313 /* remove one skb from head of flow queue */
 314 static struct sk_buff *fq_dequeue_head(struct Qdisc *sch, struct fq_flow *flow)
 315 {
 316         struct sk_buff *skb = flow->head;
 317
 318         if (skb) {
 319                 flow->head = skb->next;
 320                 skb->next = NULL;
 321                 flow->qlen--;
 322                 qdisc_qstats_backlog_dec(sch, skb);
 323                 sch->q.qlen--;
 324         }
 325         return skb;
 326 }
 327
 328 /* We might add in the future detection of retransmits
 329  * For the time being, just return false
 330  */
 331 static bool skb_is_retransmit(struct sk_buff *skb)
 332 {
 333         return false;
 334 }
 335
 336 /* add skb to flow queue
 337  * flow queue is a linked list, kind of FIFO, except for TCP retransmits
 338  * We special case tcp retransmits to be transmitted before other packets.
 339  * We rely on fact that TCP retransmits are unlikely, so we do not waste
 340  * a separate queue or a pointer.
 341  * head->  [retrans pkt 1]
 342  *         [retrans pkt 2]
 343  *         [ normal pkt 1]
 344  *         [ normal pkt 2]
 345  *         [ normal pkt 3]
 346  * tail->  [ normal pkt 4]
 347  */
 348 static void flow_queue_add(struct fq_flow *flow, struct sk_buff *skb)
 349 {
 350         struct sk_buff *prev, *head = flow->head;
 351
 352         skb->next = NULL;
 353         if (!head) {
 354                 flow->head = skb;
 355                 flow->tail = skb;
 356                 return;
 357         }
 358         if (likely(!skb_is_retransmit(skb))) {
 359                 flow->tail->next = skb;
 360                 flow->tail = skb;
 361                 return;
 362         }
 363
 364         /* This skb is a tcp retransmit,
 365          * find the last retrans packet in the queue
 366          */
 367         prev = NULL;
 368         while (skb_is_retransmit(head)) {
 369                 prev = head;
 370                 head = head->next;
 371                 if (!head)
 372                         break;
 373         }
 374         if (!prev) { /* no rtx packet in queue, become the new head */
 375                 skb->next = flow->head;
 376                 flow->head = skb;
 377         } else {
 378                 if (prev == flow->tail)
 379                         flow->tail = skb;
 380                 else
 381                         skb->next = prev->next;
 382                 prev->next = skb;
 383         }
 384 }
 385
 386 static int fq_enqueue(struct sk_buff *skb, struct Qdisc *sch)
 387 {
 388         struct fq_sched_data *q = qdisc_priv(sch);
 389         struct fq_flow *f;
 390
 391         if (unlikely(sch->q.qlen >= sch->limit))
 392                 return qdisc_drop(skb, sch);
 393
 394         f = fq_classify(skb, q);
 395         if (unlikely(f->qlen >= q->flow_plimit && f != &q->internal)) {
 396                 q->stat_flows_plimit++;
 397                 return qdisc_drop(skb, sch);
 398         }
 399
 400         f->qlen++;
 401         if (skb_is_retransmit(skb))
 402                 q->stat_tcp_retrans++;
 403         qdisc_qstats_backlog_inc(sch, skb);
 404         if (fq_flow_is_detached(f)) {
 405                 fq_flow_add_tail(&q->new_flows, f);
 406                 if (time_after(jiffies, f->age + q->flow_refill_delay))
 407                         f->credit = max_t(u32, f->credit, q->quantum);
 408                 q->inactive_flows--;
 409         }
 410
 411         /* Note: this overwrites f->age */
 412         flow_queue_add(f, skb);
 413
 414         if (unlikely(f == &q->internal)) {
 415                 q->stat_internal_packets++;
 416         }
 417         sch->q.qlen++;
 418
 419         return NET_XMIT_SUCCESS;
 420 }
 421
 422 static void fq_check_throttled(struct fq_sched_data *q, u64 now)
 423 {
 424         struct rb_node *p;
 425
 426         if (q->time_next_delayed_flow > now)
 427                 return;
 428
 429         q->time_next_delayed_flow = ~0ULL;
 430         while ((p = rb_first(&q->delayed)) != NULL) {
 431                 struct fq_flow *f = container_of(p, struct fq_flow, rate_node);
 432
 433                 if (f->time_next_packet > now) {
 434                         q->time_next_delayed_flow = f->time_next_packet;
 435                         break;
 436                 }
 437                 fq_flow_unset_throttled(q, f);
 438         }
 439 }
 440
 441 static struct sk_buff *fq_dequeue(struct Qdisc *sch)
 442 {
 443         struct fq_sched_data *q = qdisc_priv(sch);
 444         u64 now = ktime_get_ns();
 445         struct fq_flow_head *head;
 446         struct sk_buff *skb;
 447         struct fq_flow *f;
 448         u32 rate;
 449
 450         skb = fq_dequeue_head(sch, &q->internal);
 451         if (skb)
 452                 goto out;
 453         fq_check_throttled(q, now);
 454 begin:
 455         head = &q->new_flows;
 456         if (!head->first) {
 457                 head = &q->old_flows;
 458                 if (!head->first) {
 459                         if (q->time_next_delayed_flow != ~0ULL)
 460                                 qdisc_watchdog_schedule_ns(&q->watchdog,
 461                                                            q->time_next_delayed_flow,
 462                                                            false);
 463                         return NULL;
 464                 }
 465         }
 466         f = head->first;
 467
 468         if (f->credit <= 0) {
 469                 f->credit += q->quantum;
 470                 head->first = f->next;
 471                 fq_flow_add_tail(&q->old_flows, f);
 472                 goto begin;
 473         }
 474
 475         skb = f->head;
 476         if (unlikely(skb && now < f->time_next_packet &&
 477                      !skb_is_tcp_pure_ack(skb))) {
 478                 head->first = f->next;
 479                 fq_flow_set_throttled(q, f);
 480                 goto begin;
 481         }
 482
 483         skb = fq_dequeue_head(sch, f);
 484         if (!skb) {
 485                 head->first = f->next;
 486                 /* force a pass through old_flows to prevent starvation */
 487                 if ((head == &q->new_flows) && q->old_flows.first) {
 488                         fq_flow_add_tail(&q->old_flows, f);
 489                 } else {
 490                         fq_flow_set_detached(f);
 491                         q->inactive_flows++;
 492                 }
 493                 goto begin;
 494         }
 495         prefetch(&skb->end);
 496         f->credit -= qdisc_pkt_len(skb);
 497
 498         if (f->credit > 0 || !q->rate_enable)
 499                 goto out;
 500
 501         /* Do not pace locally generated ack packets */
 502         if (skb_is_tcp_pure_ack(skb))
 503                 goto out;
 504
 505         rate = q->flow_max_rate;
 506         if (skb->sk)
 507                 rate = min(skb->sk->sk_pacing_rate, rate);
 508
 509         if (rate != ~0U) {
 510                 u32 plen = max(qdisc_pkt_len(skb), q->quantum);
 511                 u64 len = (u64)plen * NSEC_PER_SEC;
 512
 513                 if (likely(rate))
 514                         do_div(len, rate);
 515                 /* Since socket rate can change later,
 516                  * clamp the delay to 1 second.
 517                  * Really, providers of too big packets should be fixed !
 518                  */
 519                 if (unlikely(len > NSEC_PER_SEC)) {
 520                         len = NSEC_PER_SEC;
 521                         q->stat_pkts_too_long++;
 522                 }
 523
 524                 f->time_next_packet = now + len;
 525         }
 526 out:
 527         qdisc_bstats_update(sch, skb);
 528         return skb;
 529 }
 530
 531 static void fq_reset(struct Qdisc *sch)
 532 {
 533         struct fq_sched_data *q = qdisc_priv(sch);
 534         struct rb_root *root;
 535         struct sk_buff *skb;
 536         struct rb_node *p;
 537         struct fq_flow *f;
 538         unsigned int idx;
 539
 540         while ((skb = fq_dequeue_head(sch, &q->internal)) != NULL)
 541                 kfree_skb(skb);
 542
 543         if (!q->fq_root)
 544                 return;
 545
 546         for (idx = 0; idx < (1U << q->fq_trees_log); idx++) {
 547                 root = &q->fq_root[idx];
 548                 while ((p = rb_first(root)) != NULL) {
 549                         f = container_of(p, struct fq_flow, fq_node);
 550                         rb_erase(p, root);
 551
 552                         while ((skb = fq_dequeue_head(sch, f)) != NULL)
 553                                 kfree_skb(skb);
 554
 555                         kmem_cache_free(fq_flow_cachep, f);
 556                 }
 557         }
 558         q->new_flows.first      = NULL;
 559         q->old_flows.first      = NULL;
 560         q->delayed              = RB_ROOT;
 561         q->flows                = 0;
 562         q->inactive_flows       = 0;
 563         q->throttled_flows      = 0;
 564 }
 565
 566 static void fq_rehash(struct fq_sched_data *q,
 567                       struct rb_root *old_array, u32 old_log,
 568                       struct rb_root *new_array, u32 new_log)
 569 {
 570         struct rb_node *op, **np, *parent;
 571         struct rb_root *oroot, *nroot;
 572         struct fq_flow *of, *nf;
 573         int fcnt = 0;
 574         u32 idx;
 575
 576         for (idx = 0; idx < (1U << old_log); idx++) {
 577                 oroot = &old_array[idx];
 578                 while ((op = rb_first(oroot)) != NULL) {
 579                         rb_erase(op, oroot);
 580                         of = container_of(op, struct fq_flow, fq_node);
 581                         if (fq_gc_candidate(of)) {
 582                                 fcnt++;
 583                                 kmem_cache_free(fq_flow_cachep, of);
 584                                 continue;
 585                         }
 586                         nroot = &new_array[hash_32((u32)(long)of->sk, new_log)];
 587
 588                         np = &nroot->rb_node;
 589                         parent = NULL;
 590                         while (*np) {
 591                                 parent = *np;
 592
 593                                 nf = container_of(parent, struct fq_flow, fq_node);
 594                                 BUG_ON(nf->sk == of->sk);
 595
 596                                 if (nf->sk > of->sk)
 597                                         np = &parent->rb_right;
 598                                 else
 599                                         np = &parent->rb_left;
 600                         }
 601
 602                         rb_link_node(&of->fq_node, parent, np);
 603                         rb_insert_color(&of->fq_node, nroot);
 604                 }
 605         }
 606         q->flows -= fcnt;
 607         q->inactive_flows -= fcnt;
 608         q->stat_gc_flows += fcnt;
 609 }
 610
 611 static void *fq_alloc_node(size_t sz, int node)
 612 {
 613         void *ptr;
 614
 615         ptr = kmalloc_node(sz, GFP_KERNEL | __GFP_REPEAT | __GFP_NOWARN, node);
 616         if (!ptr)
 617                 ptr = vmalloc_node(sz, node);
 618         return ptr;
 619 }
 620
 621 static void fq_free(void *addr)
 622 {
 623         kvfree(addr);
 624 }
 625
 626 static int fq_resize(struct Qdisc *sch, u32 log)
 627 {
 628         struct fq_sched_data *q = qdisc_priv(sch);
 629         struct rb_root *array;
 630         void *old_fq_root;
 631         u32 idx;
 632
 633         if (q->fq_root && log == q->fq_trees_log)
 634                 return 0;
 635
 636         /* If XPS was setup, we can allocate memory on right NUMA node */
 637         array = fq_alloc_node(sizeof(struct rb_root) << log,
 638                               netdev_queue_numa_node_read(sch->dev_queue));
 639         if (!array)
 640                 return -ENOMEM;
 641
 642         for (idx = 0; idx < (1U << log); idx++)
 643                 array[idx] = RB_ROOT;
 644
 645         sch_tree_lock(sch);
 646
 647         old_fq_root = q->fq_root;
 648         if (old_fq_root)
 649                 fq_rehash(q, old_fq_root, q->fq_trees_log, array, log);
 650
 651         q->fq_root = array;
 652         q->fq_trees_log = log;
 653
 654         sch_tree_unlock(sch);
 655
 656         fq_free(old_fq_root);
 657
 658         return 0;
 659 }
 660
 661 static const struct nla_policy fq_policy[TCA_FQ_MAX + 1] = {
 662         [TCA_FQ_PLIMIT]                 = { .type = NLA_U32 },
 663         [TCA_FQ_FLOW_PLIMIT]            = { .type = NLA_U32 },
 664         [TCA_FQ_QUANTUM]                = { .type = NLA_U32 },
 665         [TCA_FQ_INITIAL_QUANTUM]        = { .type = NLA_U32 },
 666         [TCA_FQ_RATE_ENABLE]            = { .type = NLA_U32 },
 667         [TCA_FQ_FLOW_DEFAULT_RATE]      = { .type = NLA_U32 },
 668         [TCA_FQ_FLOW_MAX_RATE]          = { .type = NLA_U32 },
 669         [TCA_FQ_BUCKETS_LOG]            = { .type = NLA_U32 },
 670         [TCA_FQ_FLOW_REFILL_DELAY]      = { .type = NLA_U32 },
 671         [TCA_FQ_ORPHAN_MASK]            = { .type = NLA_U32 },
 672 };
 673
 674 static int fq_change(struct Qdisc *sch, struct nlattr *opt)
 675 {
 676         struct fq_sched_data *q = qdisc_priv(sch);
 677         struct nlattr *tb[TCA_FQ_MAX + 1];
 678         int err, drop_count = 0;
 679         unsigned drop_len = 0;
 680         u32 fq_log;
 681
 682         if (!opt)
 683                 return -EINVAL;
 684
 685         err = nla_parse_nested(tb, TCA_FQ_MAX, opt, fq_policy);
 686         if (err < 0)
 687                 return err;
 688
 689         sch_tree_lock(sch);
 690
 691         fq_log = q->fq_trees_log;
 692
 693         if (tb[TCA_FQ_BUCKETS_LOG]) {
 694                 u32 nval = nla_get_u32(tb[TCA_FQ_BUCKETS_LOG]);
 695
 696                 if (nval >= 1 && nval <= ilog2(256*1024))
 697                         fq_log = nval;
 698                 else
 699                         err = -EINVAL;
 700         }
 701         if (tb[TCA_FQ_PLIMIT])
 702                 sch->limit = nla_get_u32(tb[TCA_FQ_PLIMIT]);
 703
 704         if (tb[TCA_FQ_FLOW_PLIMIT])
 705                 q->flow_plimit = nla_get_u32(tb[TCA_FQ_FLOW_PLIMIT]);
 706
 707         if (tb[TCA_FQ_QUANTUM]) {
 708                 u32 quantum = nla_get_u32(tb[TCA_FQ_QUANTUM]);
 709
 710                 if (quantum > 0 && quantum <= (1 << 20))
 711                         q->quantum = quantum;
 712                 else
 713                         err = -EINVAL;
 714         }
 715
 716         if (tb[TCA_FQ_INITIAL_QUANTUM])
 717                 q->initial_quantum = nla_get_u32(tb[TCA_FQ_INITIAL_QUANTUM]);
 718
 719         if (tb[TCA_FQ_FLOW_DEFAULT_RATE])
 720                 pr_warn_ratelimited("sch_fq: defrate %u ignored.\n",
 721                                     nla_get_u32(tb[TCA_FQ_FLOW_DEFAULT_RATE]));
 722
 723         if (tb[TCA_FQ_FLOW_MAX_RATE])
 724                 q->flow_max_rate = nla_get_u32(tb[TCA_FQ_FLOW_MAX_RATE]);
 725
 726         if (tb[TCA_FQ_RATE_ENABLE]) {
 727                 u32 enable = nla_get_u32(tb[TCA_FQ_RATE_ENABLE]);
 728
 729                 if (enable <= 1)
 730                         q->rate_enable = enable;
 731                 else
 732                         err = -EINVAL;
 733         }
 734
 735         if (tb[TCA_FQ_FLOW_REFILL_DELAY]) {
 736                 u32 usecs_delay = nla_get_u32(tb[TCA_FQ_FLOW_REFILL_DELAY]) ;
 737
 738                 q->flow_refill_delay = usecs_to_jiffies(usecs_delay);
 739         }
 740
 741         if (tb[TCA_FQ_ORPHAN_MASK])
 742                 q->orphan_mask = nla_get_u32(tb[TCA_FQ_ORPHAN_MASK]);
 743
 744         if (!err) {
 745                 sch_tree_unlock(sch);
 746                 err = fq_resize(sch, fq_log);
 747                 sch_tree_lock(sch);
 748         }
 749         while (sch->q.qlen > sch->limit) {
 750                 struct sk_buff *skb = fq_dequeue(sch);
 751
 752                 if (!skb)
 753                         break;
 754                 drop_len += qdisc_pkt_len(skb);
 755                 kfree_skb(skb);
 756                 drop_count++;
 757         }
 758         qdisc_tree_reduce_backlog(sch, drop_count, drop_len);
 759
 760         sch_tree_unlock(sch);
 761         return err;
 762 }
 763
 764 static void fq_destroy(struct Qdisc *sch)
 765 {
 766         struct fq_sched_data *q = qdisc_priv(sch);
 767
 768         fq_reset(sch);
 769         fq_free(q->fq_root);
 770         qdisc_watchdog_cancel(&q->watchdog);
 771 }
 772
 773 static int fq_init(struct Qdisc *sch, struct nlattr *opt)
 774 {
 775         struct fq_sched_data *q = qdisc_priv(sch);
 776         int err;
 777
 778         sch->limit              = 10000;
 779         q->flow_plimit          = 100;
 780         q->quantum              = 2 * psched_mtu(qdisc_dev(sch));
 781         q->initial_quantum      = 10 * psched_mtu(qdisc_dev(sch));
 782         q->flow_refill_delay    = msecs_to_jiffies(40);
 783         q->flow_max_rate        = ~0U;
 784         q->rate_enable          = 1;
 785         q->new_flows.first      = NULL;
 786         q->old_flows.first      = NULL;
 787         q->delayed              = RB_ROOT;
 788         q->fq_root              = NULL;
 789         q->fq_trees_log         = ilog2(1024);
 790         q->orphan_mask          = 1024 - 1;
 791         qdisc_watchdog_init(&q->watchdog, sch);
 792
 793         if (opt)
 794                 err = fq_change(sch, opt);
 795         else
 796                 err = fq_resize(sch, q->fq_trees_log);
 797
 798         return err;
 799 }
 800
 801 static int fq_dump(struct Qdisc *sch, struct sk_buff *skb)
 802 {
 803         struct fq_sched_data *q = qdisc_priv(sch);
 804         struct nlattr *opts;
 805
 806         opts = nla_nest_start(skb, TCA_OPTIONS);
 807         if (opts == NULL)
 808                 goto nla_put_failure;
 809
 810         /* TCA_FQ_FLOW_DEFAULT_RATE is not used anymore */
 811
 812         if (nla_put_u32(skb, TCA_FQ_PLIMIT, sch->limit) ||
 813             nla_put_u32(skb, TCA_FQ_FLOW_PLIMIT, q->flow_plimit) ||
 814             nla_put_u32(skb, TCA_FQ_QUANTUM, q->quantum) ||
 815             nla_put_u32(skb, TCA_FQ_INITIAL_QUANTUM, q->initial_quantum) ||
 816             nla_put_u32(skb, TCA_FQ_RATE_ENABLE, q->rate_enable) ||
 817             nla_put_u32(skb, TCA_FQ_FLOW_MAX_RATE, q->flow_max_rate) ||
 818             nla_put_u32(skb, TCA_FQ_FLOW_REFILL_DELAY,
 819                         jiffies_to_usecs(q->flow_refill_delay)) ||
 820             nla_put_u32(skb, TCA_FQ_ORPHAN_MASK, q->orphan_mask) ||
 821             nla_put_u32(skb, TCA_FQ_BUCKETS_LOG, q->fq_trees_log))
 822                 goto nla_put_failure;
 823
 824         return nla_nest_end(skb, opts);
 825
 826 nla_put_failure:
 827         return -1;
 828 }
 829
 830 static int fq_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
 831 {
 832         struct fq_sched_data *q = qdisc_priv(sch);
 833         u64 now = ktime_get_ns();
 834         struct tc_fq_qd_stats st = {
 835                 .gc_flows               = q->stat_gc_flows,
 836                 .highprio_packets       = q->stat_internal_packets,
 837                 .tcp_retrans            = q->stat_tcp_retrans,
 838                 .throttled              = q->stat_throttled,
 839                 .flows_plimit           = q->stat_flows_plimit,
 840                 .pkts_too_long          = q->stat_pkts_too_long,
 841                 .allocation_errors      = q->stat_allocation_errors,
 842                 .flows                  = q->flows,
 843                 .inactive_flows         = q->inactive_flows,
 844                 .throttled_flows        = q->throttled_flows,
 845                 .time_next_delayed_flow = q->time_next_delayed_flow - now,
 846         };
 847
 848         return gnet_stats_copy_app(d, &st, sizeof(st));
 849 }
 850
 851 static struct Qdisc_ops fq_qdisc_ops __read_mostly = {
 852         .id             =       "fq",
 853         .priv_size      =       sizeof(struct fq_sched_data),
 854
 855         .enqueue        =       fq_enqueue,
 856         .dequeue        =       fq_dequeue,
 857         .peek           =       qdisc_peek_dequeued,
 858         .init           =       fq_init,
 859         .reset          =       fq_reset,
 860         .destroy        =       fq_destroy,
 861         .change         =       fq_change,
 862         .dump           =       fq_dump,
 863         .dump_stats     =       fq_dump_stats,
 864         .owner          =       THIS_MODULE,
 865 };
 866
 867 static int __init fq_module_init(void)
 868 {
 869         int ret;
 870
 871         fq_flow_cachep = kmem_cache_create("fq_flow_cache",
 872                                            sizeof(struct fq_flow),
 873                                            0, 0, NULL);
 874         if (!fq_flow_cachep)
 875                 return -ENOMEM;
 876
 877         ret = register_qdisc(&fq_qdisc_ops);
 878         if (ret)
 879                 kmem_cache_destroy(fq_flow_cachep);
 880         return ret;
 881 }
 882
 883 static void __exit fq_module_exit(void)
 884 {
 885         unregister_qdisc(&fq_qdisc_ops);
 886         kmem_cache_destroy(fq_flow_cachep);
 887 }
 888
 889 module_init(fq_module_init)
 890 module_exit(fq_module_exit)
 891 MODULE_AUTHOR("Eric Dumazet");
 892 MODULE_LICENSE("GPL");