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_max_rate;  /* optional max rate per flow */
  96         u32             flow_plimit;    /* max packets 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_tcp_retrans;
 110         u64             stat_throttled;
 111         u64             stat_flows_plimit;
 112         u64             stat_pkts_too_long;
 113         u64             stat_allocation_errors;
 114         struct qdisc_watchdog watchdog;
 115 };
 116
 117 /* special value to mark a detached flow (not on old/new list) */
 118 static struct fq_flow detached, throttled;
 119
 120 static void fq_flow_set_detached(struct fq_flow *f)
 121 {
 122         f->next = &detached;
 123         f->age = jiffies;
 124 }
 125
 126 static bool fq_flow_is_detached(const struct fq_flow *f)
 127 {
 128         return f->next == &detached;
 129 }
 130
 131 static bool fq_flow_is_throttled(const struct fq_flow *f)
 132 {
 133         return f->next == &throttled;
 134 }
 135
 136 static void fq_flow_add_tail(struct fq_flow_head *head, struct fq_flow *flow)
 137 {
 138         if (head->first)
 139                 head->last->next = flow;
 140         else
 141                 head->first = flow;
 142         head->last = flow;
 143         flow->next = NULL;
 144 }
 145
 146 static void fq_flow_unset_throttled(struct fq_sched_data *q, struct fq_flow *f)
 147 {
 148         rb_erase(&f->rate_node, &q->delayed);
 149         q->throttled_flows--;
 150         fq_flow_add_tail(&q->old_flows, f);
 151 }
 152
 153 static void fq_flow_set_throttled(struct fq_sched_data *q, struct fq_flow *f)
 154 {
 155         struct rb_node **p = &q->delayed.rb_node, *parent = NULL;
 156
 157         while (*p) {
 158                 struct fq_flow *aux;
 159
 160                 parent = *p;
 161                 aux = container_of(parent, struct fq_flow, rate_node);
 162                 if (f->time_next_packet >= aux->time_next_packet)
 163                         p = &parent->rb_right;
 164                 else
 165                         p = &parent->rb_left;
 166         }
 167         rb_link_node(&f->rate_node, parent, p);
 168         rb_insert_color(&f->rate_node, &q->delayed);
 169         q->throttled_flows++;
 170         q->stat_throttled++;
 171
 172         f->next = &throttled;
 173         if (q->time_next_delayed_flow > f->time_next_packet)
 174                 q->time_next_delayed_flow = f->time_next_packet;
 175 }
 176
 177
 178 static struct kmem_cache *fq_flow_cachep __read_mostly;
 179
 180
 181 /* limit number of collected flows per round */
 182 #define FQ_GC_MAX 8
 183 #define FQ_GC_AGE (3*HZ)
 184
 185 static bool fq_gc_candidate(const struct fq_flow *f)
 186 {
 187         return fq_flow_is_detached(f) &&
 188                time_after(jiffies, f->age + FQ_GC_AGE);
 189 }
 190
 191 static void fq_gc(struct fq_sched_data *q,
 192                   struct rb_root *root,
 193                   struct sock *sk)
 194 {
 195         struct fq_flow *f, *tofree[FQ_GC_MAX];
 196         struct rb_node **p, *parent;
 197         int fcnt = 0;
 198
 199         p = &root->rb_node;
 200         parent = NULL;
 201         while (*p) {
 202                 parent = *p;
 203
 204                 f = container_of(parent, struct fq_flow, fq_node);
 205                 if (f->sk == sk)
 206                         break;
 207
 208                 if (fq_gc_candidate(f)) {
 209                         tofree[fcnt++] = f;
 210                         if (fcnt == FQ_GC_MAX)
 211                                 break;
 212                 }
 213
 214                 if (f->sk > sk)
 215                         p = &parent->rb_right;
 216                 else
 217                         p = &parent->rb_left;
 218         }
 219
 220         q->flows -= fcnt;
 221         q->inactive_flows -= fcnt;
 222         q->stat_gc_flows += fcnt;
 223         while (fcnt) {
 224                 struct fq_flow *f = tofree[--fcnt];
 225
 226                 rb_erase(&f->fq_node, root);
 227                 kmem_cache_free(fq_flow_cachep, f);
 228         }
 229 }
 230
 231 static struct fq_flow *fq_classify(struct sk_buff *skb, struct fq_sched_data *q)
 232 {
 233         struct rb_node **p, *parent;
 234         struct sock *sk = skb->sk;
 235         struct rb_root *root;
 236         struct fq_flow *f;
 237
 238         /* warning: no starvation prevention... */
 239         if (unlikely((skb->priority & TC_PRIO_MAX) == TC_PRIO_CONTROL))
 240                 return &q->internal;
 241
 242         /* SYNACK messages are attached to a TCP_NEW_SYN_RECV request socket
 243          * or a listener (SYNCOOKIE mode)
 244          * 1) request sockets are not full blown,
 245          *    they do not contain sk_pacing_rate
 246          * 2) They are not part of a 'flow' yet
 247          * 3) We do not want to rate limit them (eg SYNFLOOD attack),
 248          *    especially if the listener set SO_MAX_PACING_RATE
 249          * 4) We pretend they are orphaned
 250          */
 251         if (!sk || sk_listener(sk)) {
 252                 unsigned long hash = skb_get_hash(skb) & q->orphan_mask;
 253
 254                 /* By forcing low order bit to 1, we make sure to not
 255                  * collide with a local flow (socket pointers are word aligned)
 256                  */
 257                 sk = (struct sock *)((hash << 1) | 1UL);
 258                 skb_orphan(skb);
 259         }
 260
 261         root = &q->fq_root[hash_32((u32)(long)sk, q->fq_trees_log)];
 262
 263         if (q->flows >= (2U << q->fq_trees_log) &&
 264             q->inactive_flows > q->flows/2)
 265                 fq_gc(q, root, sk);
 266
 267         p = &root->rb_node;
 268         parent = NULL;
 269         while (*p) {
 270                 parent = *p;
 271
 272                 f = container_of(parent, struct fq_flow, fq_node);
 273                 if (f->sk == sk) {
 274                         /* socket might have been reallocated, so check
 275                          * if its sk_hash is the same.
 276                          * It not, we need to refill credit with
 277                          * initial quantum
 278                          */
 279                         if (unlikely(skb->sk &&
 280                                      f->socket_hash != sk->sk_hash)) {
 281                                 f->credit = q->initial_quantum;
 282                                 f->socket_hash = sk->sk_hash;
 283                                 if (fq_flow_is_throttled(f))
 284                                         fq_flow_unset_throttled(q, f);
 285                                 f->time_next_packet = 0ULL;
 286                         }
 287                         return f;
 288                 }
 289                 if (f->sk > sk)
 290                         p = &parent->rb_right;
 291                 else
 292                         p = &parent->rb_left;
 293         }
 294
 295         f = kmem_cache_zalloc(fq_flow_cachep, GFP_ATOMIC | __GFP_NOWARN);
 296         if (unlikely(!f)) {
 297                 q->stat_allocation_errors++;
 298                 return &q->internal;
 299         }
 300         fq_flow_set_detached(f);
 301         f->sk = sk;
 302         if (skb->sk)
 303                 f->socket_hash = sk->sk_hash;
 304         f->credit = q->initial_quantum;
 305
 306         rb_link_node(&f->fq_node, parent, p);
 307         rb_insert_color(&f->fq_node, root);
 308
 309         q->flows++;
 310         q->inactive_flows++;
 311         return f;
 312 }
 313
 314
 315 /* remove one skb from head of flow queue */
 316 static struct sk_buff *fq_dequeue_head(struct Qdisc *sch, struct fq_flow *flow)
 317 {
 318         struct sk_buff *skb = flow->head;
 319
 320         if (skb) {
 321                 flow->head = skb->next;
 322                 skb->next = NULL;
 323                 flow->qlen--;
 324                 qdisc_qstats_backlog_dec(sch, skb);
 325                 sch->q.qlen--;
 326         }
 327         return skb;
 328 }
 329
 330 /* We might add in the future detection of retransmits
 331  * For the time being, just return false
 332  */
 333 static bool skb_is_retransmit(struct sk_buff *skb)
 334 {
 335         return false;
 336 }
 337
 338 /* add skb to flow queue
 339  * flow queue is a linked list, kind of FIFO, except for TCP retransmits
 340  * We special case tcp retransmits to be transmitted before other packets.
 341  * We rely on fact that TCP retransmits are unlikely, so we do not waste
 342  * a separate queue or a pointer.
 343  * head->  [retrans pkt 1]
 344  *         [retrans pkt 2]
 345  *         [ normal pkt 1]
 346  *         [ normal pkt 2]
 347  *         [ normal pkt 3]
 348  * tail->  [ normal pkt 4]
 349  */
 350 static void flow_queue_add(struct fq_flow *flow, struct sk_buff *skb)
 351 {
 352         struct sk_buff *prev, *head = flow->head;
 353
 354         skb->next = NULL;
 355         if (!head) {
 356                 flow->head = skb;
 357                 flow->tail = skb;
 358                 return;
 359         }
 360         if (likely(!skb_is_retransmit(skb))) {
 361                 flow->tail->next = skb;
 362                 flow->tail = skb;
 363                 return;
 364         }
 365
 366         /* This skb is a tcp retransmit,
 367          * find the last retrans packet in the queue
 368          */
 369         prev = NULL;
 370         while (skb_is_retransmit(head)) {
 371                 prev = head;
 372                 head = head->next;
 373                 if (!head)
 374                         break;
 375         }
 376         if (!prev) { /* no rtx packet in queue, become the new head */
 377                 skb->next = flow->head;
 378                 flow->head = skb;
 379         } else {
 380                 if (prev == flow->tail)
 381                         flow->tail = skb;
 382                 else
 383                         skb->next = prev->next;
 384                 prev->next = skb;
 385         }
 386 }
 387
 388 static int fq_enqueue(struct sk_buff *skb, struct Qdisc *sch,
 389                       struct sk_buff **to_free)
 390 {
 391         struct fq_sched_data *q = qdisc_priv(sch);
 392         struct fq_flow *f;
 393
 394         if (unlikely(sch->q.qlen >= sch->limit))
 395                 return qdisc_drop(skb, sch, to_free);
 396
 397         f = fq_classify(skb, q);
 398         if (unlikely(f->qlen >= q->flow_plimit && f != &q->internal)) {
 399                 q->stat_flows_plimit++;
 400                 return qdisc_drop(skb, sch, to_free);
 401         }
 402
 403         f->qlen++;
 404         if (skb_is_retransmit(skb))
 405                 q->stat_tcp_retrans++;
 406         qdisc_qstats_backlog_inc(sch, skb);
 407         if (fq_flow_is_detached(f)) {
 408                 fq_flow_add_tail(&q->new_flows, f);
 409                 if (time_after(jiffies, f->age + q->flow_refill_delay))
 410                         f->credit = max_t(u32, f->credit, q->quantum);
 411                 q->inactive_flows--;
 412         }
 413
 414         /* Note: this overwrites f->age */
 415         flow_queue_add(f, skb);
 416
 417         if (unlikely(f == &q->internal)) {
 418                 q->stat_internal_packets++;
 419         }
 420         sch->q.qlen++;
 421
 422         return NET_XMIT_SUCCESS;
 423 }
 424
 425 static void fq_check_throttled(struct fq_sched_data *q, u64 now)
 426 {
 427         unsigned long sample;
 428         struct rb_node *p;
 429
 430         if (q->time_next_delayed_flow > now)
 431                 return;
 432
 433         /* Update unthrottle latency EWMA.
 434          * This is cheap and can help diagnosing timer/latency problems.
 435          */
 436         sample = (unsigned long)(now - q->time_next_delayed_flow);
 437         q->unthrottle_latency_ns -= q->unthrottle_latency_ns >> 3;
 438         q->unthrottle_latency_ns += sample >> 3;
 439
 440         q->time_next_delayed_flow = ~0ULL;
 441         while ((p = rb_first(&q->delayed)) != NULL) {
 442                 struct fq_flow *f = container_of(p, struct fq_flow, rate_node);
 443
 444                 if (f->time_next_packet > now) {
 445                         q->time_next_delayed_flow = f->time_next_packet;
 446                         break;
 447                 }
 448                 fq_flow_unset_throttled(q, f);
 449         }
 450 }
 451
 452 static struct sk_buff *fq_dequeue(struct Qdisc *sch)
 453 {
 454         struct fq_sched_data *q = qdisc_priv(sch);
 455         u64 now = ktime_get_ns();
 456         struct fq_flow_head *head;
 457         struct sk_buff *skb;
 458         struct fq_flow *f;
 459         u32 rate, plen;
 460
 461         skb = fq_dequeue_head(sch, &q->internal);
 462         if (skb)
 463                 goto out;
 464         fq_check_throttled(q, now);
 465 begin:
 466         head = &q->new_flows;
 467         if (!head->first) {
 468                 head = &q->old_flows;
 469                 if (!head->first) {
 470                         if (q->time_next_delayed_flow != ~0ULL)
 471                                 qdisc_watchdog_schedule_ns(&q->watchdog,
 472                                                            q->time_next_delayed_flow);
 473                         return NULL;
 474                 }
 475         }
 476         f = head->first;
 477
 478         if (f->credit <= 0) {
 479                 f->credit += q->quantum;
 480                 head->first = f->next;
 481                 fq_flow_add_tail(&q->old_flows, f);
 482                 goto begin;
 483         }
 484
 485         skb = f->head;
 486         if (unlikely(skb && now < f->time_next_packet &&
 487                      !skb_is_tcp_pure_ack(skb))) {
 488                 head->first = f->next;
 489                 fq_flow_set_throttled(q, f);
 490                 goto begin;
 491         }
 492
 493         skb = fq_dequeue_head(sch, f);
 494         if (!skb) {
 495                 head->first = f->next;
 496                 /* force a pass through old_flows to prevent starvation */
 497                 if ((head == &q->new_flows) && q->old_flows.first) {
 498                         fq_flow_add_tail(&q->old_flows, f);
 499                 } else {
 500                         fq_flow_set_detached(f);
 501                         q->inactive_flows++;
 502                 }
 503                 goto begin;
 504         }
 505         prefetch(&skb->end);
 506         f->credit -= qdisc_pkt_len(skb);
 507
 508         if (!q->rate_enable)
 509                 goto out;
 510
 511         /* Do not pace locally generated ack packets */
 512         if (skb_is_tcp_pure_ack(skb))
 513                 goto out;
 514
 515         rate = q->flow_max_rate;
 516         if (skb->sk)
 517                 rate = min(skb->sk->sk_pacing_rate, rate);
 518
 519         if (rate <= q->low_rate_threshold) {
 520                 f->credit = 0;
 521                 plen = qdisc_pkt_len(skb);
 522         } else {
 523                 plen = max(qdisc_pkt_len(skb), q->quantum);
 524                 if (f->credit > 0)
 525                         goto out;
 526         }
 527         if (rate != ~0U) {
 528                 u64 len = (u64)plen * NSEC_PER_SEC;
 529
 530                 if (likely(rate))
 531                         do_div(len, rate);
 532                 /* Since socket rate can change later,
 533                  * clamp the delay to 1 second.
 534                  * Really, providers of too big packets should be fixed !
 535                  */
 536                 if (unlikely(len > NSEC_PER_SEC)) {
 537                         len = NSEC_PER_SEC;
 538                         q->stat_pkts_too_long++;
 539                 }
 540                 /* Account for schedule/timers drifts.
 541                  * f->time_next_packet was set when prior packet was sent,
 542                  * and current time (@now) can be too late by tens of us.
 543                  */
 544                 if (f->time_next_packet)
 545                         len -= min(len/2, now - f->time_next_packet);
 546                 f->time_next_packet = now + len;
 547         }
 548 out:
 549         qdisc_bstats_update(sch, skb);
 550         return skb;
 551 }
 552
 553 static void fq_flow_purge(struct fq_flow *flow)
 554 {
 555         rtnl_kfree_skbs(flow->head, flow->tail);
 556         flow->head = NULL;
 557         flow->qlen = 0;
 558 }
 559
 560 static void fq_reset(struct Qdisc *sch)
 561 {
 562         struct fq_sched_data *q = qdisc_priv(sch);
 563         struct rb_root *root;
 564         struct rb_node *p;
 565         struct fq_flow *f;
 566         unsigned int idx;
 567
 568         sch->q.qlen = 0;
 569         sch->qstats.backlog = 0;
 570
 571         fq_flow_purge(&q->internal);
 572
 573         if (!q->fq_root)
 574                 return;
 575
 576         for (idx = 0; idx < (1U << q->fq_trees_log); idx++) {
 577                 root = &q->fq_root[idx];
 578                 while ((p = rb_first(root)) != NULL) {
 579                         f = container_of(p, struct fq_flow, fq_node);
 580                         rb_erase(p, root);
 581
 582                         fq_flow_purge(f);
 583
 584                         kmem_cache_free(fq_flow_cachep, f);
 585                 }
 586         }
 587         q->new_flows.first      = NULL;
 588         q->old_flows.first      = NULL;
 589         q->delayed              = RB_ROOT;
 590         q->flows                = 0;
 591         q->inactive_flows       = 0;
 592         q->throttled_flows      = 0;
 593 }
 594
 595 static void fq_rehash(struct fq_sched_data *q,
 596                       struct rb_root *old_array, u32 old_log,
 597                       struct rb_root *new_array, u32 new_log)
 598 {
 599         struct rb_node *op, **np, *parent;
 600         struct rb_root *oroot, *nroot;
 601         struct fq_flow *of, *nf;
 602         int fcnt = 0;
 603         u32 idx;
 604
 605         for (idx = 0; idx < (1U << old_log); idx++) {
 606                 oroot = &old_array[idx];
 607                 while ((op = rb_first(oroot)) != NULL) {
 608                         rb_erase(op, oroot);
 609                         of = container_of(op, struct fq_flow, fq_node);
 610                         if (fq_gc_candidate(of)) {
 611                                 fcnt++;
 612                                 kmem_cache_free(fq_flow_cachep, of);
 613                                 continue;
 614                         }
 615                         nroot = &new_array[hash_32((u32)(long)of->sk, new_log)];
 616
 617                         np = &nroot->rb_node;
 618                         parent = NULL;
 619                         while (*np) {
 620                                 parent = *np;
 621
 622                                 nf = container_of(parent, struct fq_flow, fq_node);
 623                                 BUG_ON(nf->sk == of->sk);
 624
 625                                 if (nf->sk > of->sk)
 626                                         np = &parent->rb_right;
 627                                 else
 628                                         np = &parent->rb_left;
 629                         }
 630
 631                         rb_link_node(&of->fq_node, parent, np);
 632                         rb_insert_color(&of->fq_node, nroot);
 633                 }
 634         }
 635         q->flows -= fcnt;
 636         q->inactive_flows -= fcnt;
 637         q->stat_gc_flows += fcnt;
 638 }
 639
 640 static void *fq_alloc_node(size_t sz, int node)
 641 {
 642         void *ptr;
 643
 644         ptr = kmalloc_node(sz, GFP_KERNEL | __GFP_REPEAT | __GFP_NOWARN, node);
 645         if (!ptr)
 646                 ptr = vmalloc_node(sz, node);
 647         return ptr;
 648 }
 649
 650 static void fq_free(void *addr)
 651 {
 652         kvfree(addr);
 653 }
 654
 655 static int fq_resize(struct Qdisc *sch, u32 log)
 656 {
 657         struct fq_sched_data *q = qdisc_priv(sch);
 658         struct rb_root *array;
 659         void *old_fq_root;
 660         u32 idx;
 661
 662         if (q->fq_root && log == q->fq_trees_log)
 663                 return 0;
 664
 665         /* If XPS was setup, we can allocate memory on right NUMA node */
 666         array = fq_alloc_node(sizeof(struct rb_root) << log,
 667                               netdev_queue_numa_node_read(sch->dev_queue));
 668         if (!array)
 669                 return -ENOMEM;
 670
 671         for (idx = 0; idx < (1U << log); idx++)
 672                 array[idx] = RB_ROOT;
 673
 674         sch_tree_lock(sch);
 675
 676         old_fq_root = q->fq_root;
 677         if (old_fq_root)
 678                 fq_rehash(q, old_fq_root, q->fq_trees_log, array, log);
 679
 680         q->fq_root = array;
 681         q->fq_trees_log = log;
 682
 683         sch_tree_unlock(sch);
 684
 685         fq_free(old_fq_root);
 686
 687         return 0;
 688 }
 689
 690 static const struct nla_policy fq_policy[TCA_FQ_MAX + 1] = {
 691         [TCA_FQ_PLIMIT]                 = { .type = NLA_U32 },
 692         [TCA_FQ_FLOW_PLIMIT]            = { .type = NLA_U32 },
 693         [TCA_FQ_QUANTUM]                = { .type = NLA_U32 },
 694         [TCA_FQ_INITIAL_QUANTUM]        = { .type = NLA_U32 },
 695         [TCA_FQ_RATE_ENABLE]            = { .type = NLA_U32 },
 696         [TCA_FQ_FLOW_DEFAULT_RATE]      = { .type = NLA_U32 },
 697         [TCA_FQ_FLOW_MAX_RATE]          = { .type = NLA_U32 },
 698         [TCA_FQ_BUCKETS_LOG]            = { .type = NLA_U32 },
 699         [TCA_FQ_FLOW_REFILL_DELAY]      = { .type = NLA_U32 },
 700         [TCA_FQ_ORPHAN_MASK]            = { .type = NLA_U32 },
 701         [TCA_FQ_LOW_RATE_THRESHOLD]     = { .type = NLA_U32 },
 702 };
 703
 704 static int fq_change(struct Qdisc *sch, struct nlattr *opt)
 705 {
 706         struct fq_sched_data *q = qdisc_priv(sch);
 707         struct nlattr *tb[TCA_FQ_MAX + 1];
 708         int err, drop_count = 0;
 709         unsigned drop_len = 0;
 710         u32 fq_log;
 711
 712         if (!opt)
 713                 return -EINVAL;
 714
 715         err = nla_parse_nested(tb, TCA_FQ_MAX, opt, fq_policy);
 716         if (err < 0)
 717                 return err;
 718
 719         sch_tree_lock(sch);
 720
 721         fq_log = q->fq_trees_log;
 722
 723         if (tb[TCA_FQ_BUCKETS_LOG]) {
 724                 u32 nval = nla_get_u32(tb[TCA_FQ_BUCKETS_LOG]);
 725
 726                 if (nval >= 1 && nval <= ilog2(256*1024))
 727                         fq_log = nval;
 728                 else
 729                         err = -EINVAL;
 730         }
 731         if (tb[TCA_FQ_PLIMIT])
 732                 sch->limit = nla_get_u32(tb[TCA_FQ_PLIMIT]);
 733
 734         if (tb[TCA_FQ_FLOW_PLIMIT])
 735                 q->flow_plimit = nla_get_u32(tb[TCA_FQ_FLOW_PLIMIT]);
 736
 737         if (tb[TCA_FQ_QUANTUM]) {
 738                 u32 quantum = nla_get_u32(tb[TCA_FQ_QUANTUM]);
 739
 740                 if (quantum > 0 && quantum <= (1 << 20))
 741                         q->quantum = quantum;
 742                 else
 743                         err = -EINVAL;
 744         }
 745
 746         if (tb[TCA_FQ_INITIAL_QUANTUM])
 747                 q->initial_quantum = nla_get_u32(tb[TCA_FQ_INITIAL_QUANTUM]);
 748
 749         if (tb[TCA_FQ_FLOW_DEFAULT_RATE])
 750                 pr_warn_ratelimited("sch_fq: defrate %u ignored.\n",
 751                                     nla_get_u32(tb[TCA_FQ_FLOW_DEFAULT_RATE]));
 752
 753         if (tb[TCA_FQ_FLOW_MAX_RATE])
 754                 q->flow_max_rate = nla_get_u32(tb[TCA_FQ_FLOW_MAX_RATE]);
 755
 756         if (tb[TCA_FQ_LOW_RATE_THRESHOLD])
 757                 q->low_rate_threshold =
 758                         nla_get_u32(tb[TCA_FQ_LOW_RATE_THRESHOLD]);
 759
 760         if (tb[TCA_FQ_RATE_ENABLE]) {
 761                 u32 enable = nla_get_u32(tb[TCA_FQ_RATE_ENABLE]);
 762
 763                 if (enable <= 1)
 764                         q->rate_enable = enable;
 765                 else
 766                         err = -EINVAL;
 767         }
 768
 769         if (tb[TCA_FQ_FLOW_REFILL_DELAY]) {
 770                 u32 usecs_delay = nla_get_u32(tb[TCA_FQ_FLOW_REFILL_DELAY]) ;
 771
 772                 q->flow_refill_delay = usecs_to_jiffies(usecs_delay);
 773         }
 774
 775         if (tb[TCA_FQ_ORPHAN_MASK])
 776                 q->orphan_mask = nla_get_u32(tb[TCA_FQ_ORPHAN_MASK]);
 777
 778         if (!err) {
 779                 sch_tree_unlock(sch);
 780                 err = fq_resize(sch, fq_log);
 781                 sch_tree_lock(sch);
 782         }
 783         while (sch->q.qlen > sch->limit) {
 784                 struct sk_buff *skb = fq_dequeue(sch);
 785
 786                 if (!skb)
 787                         break;
 788                 drop_len += qdisc_pkt_len(skb);
 789                 rtnl_kfree_skbs(skb, skb);
 790                 drop_count++;
 791         }
 792         qdisc_tree_reduce_backlog(sch, drop_count, drop_len);
 793
 794         sch_tree_unlock(sch);
 795         return err;
 796 }
 797
 798 static void fq_destroy(struct Qdisc *sch)
 799 {
 800         struct fq_sched_data *q = qdisc_priv(sch);
 801
 802         fq_reset(sch);
 803         fq_free(q->fq_root);
 804         qdisc_watchdog_cancel(&q->watchdog);
 805 }
 806
 807 static int fq_init(struct Qdisc *sch, struct nlattr *opt)
 808 {
 809         struct fq_sched_data *q = qdisc_priv(sch);
 810         int err;
 811
 812         sch->limit              = 10000;
 813         q->flow_plimit          = 100;
 814         q->quantum              = 2 * psched_mtu(qdisc_dev(sch));
 815         q->initial_quantum      = 10 * psched_mtu(qdisc_dev(sch));
 816         q->flow_refill_delay    = msecs_to_jiffies(40);
 817         q->flow_max_rate        = ~0U;
 818         q->time_next_delayed_flow = ~0ULL;
 819         q->rate_enable          = 1;
 820         q->new_flows.first      = NULL;
 821         q->old_flows.first      = NULL;
 822         q->delayed              = RB_ROOT;
 823         q->fq_root              = NULL;
 824         q->fq_trees_log         = ilog2(1024);
 825         q->orphan_mask          = 1024 - 1;
 826         q->low_rate_threshold   = 550000 / 8;
 827         qdisc_watchdog_init(&q->watchdog, sch);
 828
 829         if (opt)
 830                 err = fq_change(sch, opt);
 831         else
 832                 err = fq_resize(sch, q->fq_trees_log);
 833
 834         return err;
 835 }
 836
 837 static int fq_dump(struct Qdisc *sch, struct sk_buff *skb)
 838 {
 839         struct fq_sched_data *q = qdisc_priv(sch);
 840         struct nlattr *opts;
 841
 842         opts = nla_nest_start(skb, TCA_OPTIONS);
 843         if (opts == NULL)
 844                 goto nla_put_failure;
 845
 846         /* TCA_FQ_FLOW_DEFAULT_RATE is not used anymore */
 847
 848         if (nla_put_u32(skb, TCA_FQ_PLIMIT, sch->limit) ||
 849             nla_put_u32(skb, TCA_FQ_FLOW_PLIMIT, q->flow_plimit) ||
 850             nla_put_u32(skb, TCA_FQ_QUANTUM, q->quantum) ||
 851             nla_put_u32(skb, TCA_FQ_INITIAL_QUANTUM, q->initial_quantum) ||
 852             nla_put_u32(skb, TCA_FQ_RATE_ENABLE, q->rate_enable) ||
 853             nla_put_u32(skb, TCA_FQ_FLOW_MAX_RATE, q->flow_max_rate) ||
 854             nla_put_u32(skb, TCA_FQ_FLOW_REFILL_DELAY,
 855                         jiffies_to_usecs(q->flow_refill_delay)) ||
 856             nla_put_u32(skb, TCA_FQ_ORPHAN_MASK, q->orphan_mask) ||
 857             nla_put_u32(skb, TCA_FQ_LOW_RATE_THRESHOLD,
 858                         q->low_rate_threshold) ||
 859             nla_put_u32(skb, TCA_FQ_BUCKETS_LOG, q->fq_trees_log))
 860                 goto nla_put_failure;
 861
 862         return nla_nest_end(skb, opts);
 863
 864 nla_put_failure:
 865         return -1;
 866 }
 867
 868 static int fq_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
 869 {
 870         struct fq_sched_data *q = qdisc_priv(sch);
 871         struct tc_fq_qd_stats st;
 872
 873         sch_tree_lock(sch);
 874
 875         st.gc_flows               = q->stat_gc_flows;
 876         st.highprio_packets       = q->stat_internal_packets;
 877         st.tcp_retrans            = q->stat_tcp_retrans;
 878         st.throttled              = q->stat_throttled;
 879         st.flows_plimit           = q->stat_flows_plimit;
 880         st.pkts_too_long          = q->stat_pkts_too_long;
 881         st.allocation_errors      = q->stat_allocation_errors;
 882         st.time_next_delayed_flow = q->time_next_delayed_flow - ktime_get_ns();
 883         st.flows                  = q->flows;
 884         st.inactive_flows         = q->inactive_flows;
 885         st.throttled_flows        = q->throttled_flows;
 886         st.unthrottle_latency_ns  = min_t(unsigned long,
 887                                           q->unthrottle_latency_ns, ~0U);
 888         sch_tree_unlock(sch);
 889
 890         return gnet_stats_copy_app(d, &st, sizeof(st));
 891 }
 892
 893 static struct Qdisc_ops fq_qdisc_ops __read_mostly = {
 894         .id             =       "fq",
 895         .priv_size      =       sizeof(struct fq_sched_data),
 896
 897         .enqueue        =       fq_enqueue,
 898         .dequeue        =       fq_dequeue,
 899         .peek           =       qdisc_peek_dequeued,
 900         .init           =       fq_init,
 901         .reset          =       fq_reset,
 902         .destroy        =       fq_destroy,
 903         .change         =       fq_change,
 904         .dump           =       fq_dump,
 905         .dump_stats     =       fq_dump_stats,
 906         .owner          =       THIS_MODULE,
 907 };
 908
 909 static int __init fq_module_init(void)
 910 {
 911         int ret;
 912
 913         fq_flow_cachep = kmem_cache_create("fq_flow_cache",
 914                                            sizeof(struct fq_flow),
 915                                            0, 0, NULL);
 916         if (!fq_flow_cachep)
 917                 return -ENOMEM;
 918
 919         ret = register_qdisc(&fq_qdisc_ops);
 920         if (ret)
 921                 kmem_cache_destroy(fq_flow_cachep);
 922         return ret;
 923 }
 924
 925 static void __exit fq_module_exit(void)
 926 {
 927         unregister_qdisc(&fq_qdisc_ops);
 928         kmem_cache_destroy(fq_flow_cachep);
 929 }
 930
 931 module_init(fq_module_init)
 932 module_exit(fq_module_exit)
 933 MODULE_AUTHOR("Eric Dumazet");
 934 MODULE_LICENSE("GPL");