net/sched/sch_fq.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /*
   3  * net/sched/sch_fq.c Fair Queue Packet Scheduler (per flow pacing)
   4  *
   5  *  Copyright (C) 2013-2023 Eric Dumazet <[email protected]>
   6  *
   7  *  Meant to be mostly used for locally generated traffic :
   8  *  Fast classification depends on skb->sk being set before reaching us.
   9  *  If not, (router workload), we use rxhash as fallback, with 32 bits wide hash.
  10  *  All packets belonging to a socket are considered as a 'flow'.
  11  *
  12  *  Flows are dynamically allocated and stored in a hash table of RB trees
  13  *  They are also part of one Round Robin 'queues' (new or old flows)
  14  *
  15  *  Burst avoidance (aka pacing) capability :
  16  *
  17  *  Transport (eg TCP) can set in sk->sk_pacing_rate a rate, enqueue a
  18  *  bunch of packets, and this packet scheduler adds delay between
  19  *  packets to respect rate limitation.
  20  *
  21  *  enqueue() :
  22  *   - lookup one RB tree (out of 1024 or more) to find the flow.
  23  *     If non existent flow, create it, add it to the tree.
  24  *     Add skb to the per flow list of skb (fifo).
  25  *   - Use a special fifo for high prio packets
  26  *
  27  *  dequeue() : serves flows in Round Robin
  28  *  Note : When a flow becomes empty, we do not immediately remove it from
  29  *  rb trees, for performance reasons (its expected to send additional packets,
  30  *  or SLAB cache will reuse socket for another flow)
  31  */
  32
  33 #include <linux/module.h>
  34 #include <linux/types.h>
  35 #include <linux/kernel.h>
  36 #include <linux/jiffies.h>
  37 #include <linux/string.h>
  38 #include <linux/in.h>
  39 #include <linux/errno.h>
  40 #include <linux/init.h>
  41 #include <linux/skbuff.h>
  42 #include <linux/slab.h>
  43 #include <linux/rbtree.h>
  44 #include <linux/hash.h>
  45 #include <linux/prefetch.h>
  46 #include <linux/vmalloc.h>
  47 #include <net/netlink.h>
  48 #include <net/pkt_sched.h>
  49 #include <net/sock.h>
  50 #include <net/tcp_states.h>
  51 #include <net/tcp.h>
  52
  53 struct fq_skb_cb {
  54         u64     time_to_send;
  55         u8      band;
  56 };
  57
  58 static inline struct fq_skb_cb *fq_skb_cb(struct sk_buff *skb)
  59 {
  60         qdisc_cb_private_validate(skb, sizeof(struct fq_skb_cb));
  61         return (struct fq_skb_cb *)qdisc_skb_cb(skb)->data;
  62 }
  63
  64 /*
  65  * Per flow structure, dynamically allocated.
  66  * If packets have monotically increasing time_to_send, they are placed in O(1)
  67  * in linear list (head,tail), otherwise are placed in a rbtree (t_root).
  68  */
  69 struct fq_flow {
  70 /* First cache line : used in fq_gc(), fq_enqueue(), fq_dequeue() */
  71         struct rb_root  t_root;
  72         struct sk_buff  *head;          /* list of skbs for this flow : first skb */
  73         union {
  74                 struct sk_buff *tail;   /* last skb in the list */
  75                 unsigned long  age;     /* (jiffies | 1UL) when flow was emptied, for gc */
  76         };
  77         union {
  78                 struct rb_node  fq_node;        /* anchor in fq_root[] trees */
  79                 /* Following field is only used for q->internal,
  80                  * because q->internal is not hashed in fq_root[]
  81                  */
  82                 u64             stat_fastpath_packets;
  83         };
  84         struct sock     *sk;
  85         u32             socket_hash;    /* sk_hash */
  86         int             qlen;           /* number of packets in flow queue */
  87
  88 /* Second cache line */
  89         int             credit;
  90         int             band;
  91         struct fq_flow *next;           /* next pointer in RR lists */
  92
  93         struct rb_node  rate_node;      /* anchor in q->delayed tree */
  94         u64             time_next_packet;
  95 };
  96
  97 struct fq_flow_head {
  98         struct fq_flow *first;
  99         struct fq_flow *last;
 100 };
 101
 102 struct fq_perband_flows {
 103         struct fq_flow_head new_flows;
 104         struct fq_flow_head old_flows;
 105         int                 credit;
 106         int                 quantum; /* based on band nr : 576KB, 192KB, 64KB */
 107 };
 108
 109 struct fq_sched_data {
 110 /* Read mostly cache line */
 111
 112         u32             quantum;
 113         u32             initial_quantum;
 114         u32             flow_refill_delay;
 115         u32             flow_plimit;    /* max packets per flow */
 116         unsigned long   flow_max_rate;  /* optional max rate per flow */
 117         u64             ce_threshold;
 118         u64             horizon;        /* horizon in ns */
 119         u32             orphan_mask;    /* mask for orphaned skb */
 120         u32             low_rate_threshold;
 121         struct rb_root  *fq_root;
 122         u8              rate_enable;
 123         u8              fq_trees_log;
 124         u8              horizon_drop;
 125         u8              prio2band[(TC_PRIO_MAX + 1) >> 2];
 126         u32             timer_slack; /* hrtimer slack in ns */
 127
 128 /* Read/Write fields. */
 129
 130         unsigned int band_nr; /* band being serviced in fq_dequeue() */
 131
 132         struct fq_perband_flows band_flows[FQ_BANDS];
 133
 134         struct fq_flow  internal;       /* fastpath queue. */
 135         struct rb_root  delayed;        /* for rate limited flows */
 136         u64             time_next_delayed_flow;
 137         unsigned long   unthrottle_latency_ns;
 138
 139         u32             band_pkt_count[FQ_BANDS];
 140         u32             flows;
 141         u32             inactive_flows; /* Flows with no packet to send. */
 142         u32             throttled_flows;
 143
 144         u64             stat_throttled;
 145         struct qdisc_watchdog watchdog;
 146         u64             stat_gc_flows;
 147
 148 /* Seldom used fields. */
 149
 150         u64             stat_band_drops[FQ_BANDS];
 151         u64             stat_ce_mark;
 152         u64             stat_horizon_drops;
 153         u64             stat_horizon_caps;
 154         u64             stat_flows_plimit;
 155         u64             stat_pkts_too_long;
 156         u64             stat_allocation_errors;
 157 };
 158
 159 /* return the i-th 2-bit value ("crumb") */
 160 static u8 fq_prio2band(const u8 *prio2band, unsigned int prio)
 161 {
 162         return (prio2band[prio / 4] >> (2 * (prio & 0x3))) & 0x3;
 163 }
 164
 165 /*
 166  * f->tail and f->age share the same location.
 167  * We can use the low order bit to differentiate if this location points
 168  * to a sk_buff or contains a jiffies value, if we force this value to be odd.
 169  * This assumes f->tail low order bit must be 0 since alignof(struct sk_buff) >= 2
 170  */
 171 static void fq_flow_set_detached(struct fq_flow *f)
 172 {
 173         f->age = jiffies | 1UL;
 174 }
 175
 176 static bool fq_flow_is_detached(const struct fq_flow *f)
 177 {
 178         return !!(f->age & 1UL);
 179 }
 180
 181 /* special value to mark a throttled flow (not on old/new list) */
 182 static struct fq_flow throttled;
 183
 184 static bool fq_flow_is_throttled(const struct fq_flow *f)
 185 {
 186         return f->next == &throttled;
 187 }
 188
 189 enum new_flow {
 190         NEW_FLOW,
 191         OLD_FLOW
 192 };
 193
 194 static void fq_flow_add_tail(struct fq_sched_data *q, struct fq_flow *flow,
 195                              enum new_flow list_sel)
 196 {
 197         struct fq_perband_flows *pband = &q->band_flows[flow->band];
 198         struct fq_flow_head *head = (list_sel == NEW_FLOW) ?
 199                                         &pband->new_flows :
 200                                         &pband->old_flows;
 201
 202         if (head->first)
 203                 head->last->next = flow;
 204         else
 205                 head->first = flow;
 206         head->last = flow;
 207         flow->next = NULL;
 208 }
 209
 210 static void fq_flow_unset_throttled(struct fq_sched_data *q, struct fq_flow *f)
 211 {
 212         rb_erase(&f->rate_node, &q->delayed);
 213         q->throttled_flows--;
 214         fq_flow_add_tail(q, f, OLD_FLOW);
 215 }
 216
 217 static void fq_flow_set_throttled(struct fq_sched_data *q, struct fq_flow *f)
 218 {
 219         struct rb_node **p = &q->delayed.rb_node, *parent = NULL;
 220
 221         while (*p) {
 222                 struct fq_flow *aux;
 223
 224                 parent = *p;
 225                 aux = rb_entry(parent, struct fq_flow, rate_node);
 226                 if (f->time_next_packet >= aux->time_next_packet)
 227                         p = &parent->rb_right;
 228                 else
 229                         p = &parent->rb_left;
 230         }
 231         rb_link_node(&f->rate_node, parent, p);
 232         rb_insert_color(&f->rate_node, &q->delayed);
 233         q->throttled_flows++;
 234         q->stat_throttled++;
 235
 236         f->next = &throttled;
 237         if (q->time_next_delayed_flow > f->time_next_packet)
 238                 q->time_next_delayed_flow = f->time_next_packet;
 239 }
 240
 241
 242 static struct kmem_cache *fq_flow_cachep __read_mostly;
 243
 244
 245 /* limit number of collected flows per round */
 246 #define FQ_GC_MAX 8
 247 #define FQ_GC_AGE (3*HZ)
 248
 249 static bool fq_gc_candidate(const struct fq_flow *f)
 250 {
 251         return fq_flow_is_detached(f) &&
 252                time_after(jiffies, f->age + FQ_GC_AGE);
 253 }
 254
 255 static void fq_gc(struct fq_sched_data *q,
 256                   struct rb_root *root,
 257                   struct sock *sk)
 258 {
 259         struct rb_node **p, *parent;
 260         void *tofree[FQ_GC_MAX];
 261         struct fq_flow *f;
 262         int i, fcnt = 0;
 263
 264         p = &root->rb_node;
 265         parent = NULL;
 266         while (*p) {
 267                 parent = *p;
 268
 269                 f = rb_entry(parent, struct fq_flow, fq_node);
 270                 if (f->sk == sk)
 271                         break;
 272
 273                 if (fq_gc_candidate(f)) {
 274                         tofree[fcnt++] = f;
 275                         if (fcnt == FQ_GC_MAX)
 276                                 break;
 277                 }
 278
 279                 if (f->sk > sk)
 280                         p = &parent->rb_right;
 281                 else
 282                         p = &parent->rb_left;
 283         }
 284
 285         if (!fcnt)
 286                 return;
 287
 288         for (i = fcnt; i > 0; ) {
 289                 f = tofree[--i];
 290                 rb_erase(&f->fq_node, root);
 291         }
 292         q->flows -= fcnt;
 293         q->inactive_flows -= fcnt;
 294         q->stat_gc_flows += fcnt;
 295
 296         kmem_cache_free_bulk(fq_flow_cachep, fcnt, tofree);
 297 }
 298
 299 /* Fast path can be used if :
 300  * 1) Packet tstamp is in the past.
 301  * 2) FQ qlen == 0   OR
 302  *   (no flow is currently eligible for transmit,
 303  *    AND fast path queue has less than 8 packets)
 304  * 3) No SO_MAX_PACING_RATE on the socket (if any).
 305  * 4) No @maxrate attribute on this qdisc,
 306  *
 307  * FQ can not use generic TCQ_F_CAN_BYPASS infrastructure.
 308  */
 309 static bool fq_fastpath_check(const struct Qdisc *sch, struct sk_buff *skb,
 310                               u64 now)
 311 {
 312         const struct fq_sched_data *q = qdisc_priv(sch);
 313         const struct sock *sk;
 314
 315         if (fq_skb_cb(skb)->time_to_send > now)
 316                 return false;
 317
 318         if (sch->q.qlen != 0) {
 319                 /* Even if some packets are stored in this qdisc,
 320                  * we can still enable fast path if all of them are
 321                  * scheduled in the future (ie no flows are eligible)
 322                  * or in the fast path queue.
 323                  */
 324                 if (q->flows != q->inactive_flows + q->throttled_flows)
 325                         return false;
 326
 327                 /* Do not allow fast path queue to explode, we want Fair Queue mode
 328                  * under pressure.
 329                  */
 330                 if (q->internal.qlen >= 8)
 331                         return false;
 332         }
 333
 334         sk = skb->sk;
 335         if (sk && sk_fullsock(sk) && !sk_is_tcp(sk) &&
 336             sk->sk_max_pacing_rate != ~0UL)
 337                 return false;
 338
 339         if (q->flow_max_rate != ~0UL)
 340                 return false;
 341
 342         return true;
 343 }
 344
 345 static struct fq_flow *fq_classify(struct Qdisc *sch, struct sk_buff *skb,
 346                                    u64 now)
 347 {
 348         struct fq_sched_data *q = qdisc_priv(sch);
 349         struct rb_node **p, *parent;
 350         struct sock *sk = skb->sk;
 351         struct rb_root *root;
 352         struct fq_flow *f;
 353
 354         /* SYNACK messages are attached to a TCP_NEW_SYN_RECV request socket
 355          * or a listener (SYNCOOKIE mode)
 356          * 1) request sockets are not full blown,
 357          *    they do not contain sk_pacing_rate
 358          * 2) They are not part of a 'flow' yet
 359          * 3) We do not want to rate limit them (eg SYNFLOOD attack),
 360          *    especially if the listener set SO_MAX_PACING_RATE
 361          * 4) We pretend they are orphaned
 362          */
 363         if (!sk || sk_listener(sk)) {
 364                 unsigned long hash = skb_get_hash(skb) & q->orphan_mask;
 365
 366                 /* By forcing low order bit to 1, we make sure to not
 367                  * collide with a local flow (socket pointers are word aligned)
 368                  */
 369                 sk = (struct sock *)((hash << 1) | 1UL);
 370                 skb_orphan(skb);
 371         } else if (sk->sk_state == TCP_CLOSE) {
 372                 unsigned long hash = skb_get_hash(skb) & q->orphan_mask;
 373                 /*
 374                  * Sockets in TCP_CLOSE are non connected.
 375                  * Typical use case is UDP sockets, they can send packets
 376                  * with sendto() to many different destinations.
 377                  * We probably could use a generic bit advertising
 378                  * non connected sockets, instead of sk_state == TCP_CLOSE,
 379                  * if we care enough.
 380                  */
 381                 sk = (struct sock *)((hash << 1) | 1UL);
 382         }
 383
 384         if (fq_fastpath_check(sch, skb, now)) {
 385                 q->internal.stat_fastpath_packets++;
 386                 if (skb->sk == sk && q->rate_enable &&
 387                     READ_ONCE(sk->sk_pacing_status) != SK_PACING_FQ)
 388                         smp_store_release(&sk->sk_pacing_status,
 389                                           SK_PACING_FQ);
 390                 return &q->internal;
 391         }
 392
 393         root = &q->fq_root[hash_ptr(sk, q->fq_trees_log)];
 394
 395         fq_gc(q, root, sk);
 396
 397         p = &root->rb_node;
 398         parent = NULL;
 399         while (*p) {
 400                 parent = *p;
 401
 402                 f = rb_entry(parent, struct fq_flow, fq_node);
 403                 if (f->sk == sk) {
 404                         /* socket might have been reallocated, so check
 405                          * if its sk_hash is the same.
 406                          * It not, we need to refill credit with
 407                          * initial quantum
 408                          */
 409                         if (unlikely(skb->sk == sk &&
 410                                      f->socket_hash != sk->sk_hash)) {
 411                                 f->credit = q->initial_quantum;
 412                                 f->socket_hash = sk->sk_hash;
 413                                 if (q->rate_enable)
 414                                         smp_store_release(&sk->sk_pacing_status,
 415                                                           SK_PACING_FQ);
 416                                 if (fq_flow_is_throttled(f))
 417                                         fq_flow_unset_throttled(q, f);
 418                                 f->time_next_packet = 0ULL;
 419                         }
 420                         return f;
 421                 }
 422                 if (f->sk > sk)
 423                         p = &parent->rb_right;
 424                 else
 425                         p = &parent->rb_left;
 426         }
 427
 428         f = kmem_cache_zalloc(fq_flow_cachep, GFP_ATOMIC | __GFP_NOWARN);
 429         if (unlikely(!f)) {
 430                 q->stat_allocation_errors++;
 431                 return &q->internal;
 432         }
 433         /* f->t_root is already zeroed after kmem_cache_zalloc() */
 434
 435         fq_flow_set_detached(f);
 436         f->sk = sk;
 437         if (skb->sk == sk) {
 438                 f->socket_hash = sk->sk_hash;
 439                 if (q->rate_enable)
 440                         smp_store_release(&sk->sk_pacing_status,
 441                                           SK_PACING_FQ);
 442         }
 443         f->credit = q->initial_quantum;
 444
 445         rb_link_node(&f->fq_node, parent, p);
 446         rb_insert_color(&f->fq_node, root);
 447
 448         q->flows++;
 449         q->inactive_flows++;
 450         return f;
 451 }
 452
 453 static struct sk_buff *fq_peek(struct fq_flow *flow)
 454 {
 455         struct sk_buff *skb = skb_rb_first(&flow->t_root);
 456         struct sk_buff *head = flow->head;
 457
 458         if (!skb)
 459                 return head;
 460
 461         if (!head)
 462                 return skb;
 463
 464         if (fq_skb_cb(skb)->time_to_send < fq_skb_cb(head)->time_to_send)
 465                 return skb;
 466         return head;
 467 }
 468
 469 static void fq_erase_head(struct Qdisc *sch, struct fq_flow *flow,
 470                           struct sk_buff *skb)
 471 {
 472         if (skb == flow->head) {
 473                 flow->head = skb->next;
 474         } else {
 475                 rb_erase(&skb->rbnode, &flow->t_root);
 476                 skb->dev = qdisc_dev(sch);
 477         }
 478 }
 479
 480 /* Remove one skb from flow queue.
 481  * This skb must be the return value of prior fq_peek().
 482  */
 483 static void fq_dequeue_skb(struct Qdisc *sch, struct fq_flow *flow,
 484                            struct sk_buff *skb)
 485 {
 486         fq_erase_head(sch, flow, skb);
 487         skb_mark_not_on_list(skb);
 488         qdisc_qstats_backlog_dec(sch, skb);
 489         sch->q.qlen--;
 490 }
 491
 492 static void flow_queue_add(struct fq_flow *flow, struct sk_buff *skb)
 493 {
 494         struct rb_node **p, *parent;
 495         struct sk_buff *head, *aux;
 496
 497         head = flow->head;
 498         if (!head ||
 499             fq_skb_cb(skb)->time_to_send >= fq_skb_cb(flow->tail)->time_to_send) {
 500                 if (!head)
 501                         flow->head = skb;
 502                 else
 503                         flow->tail->next = skb;
 504                 flow->tail = skb;
 505                 skb->next = NULL;
 506                 return;
 507         }
 508
 509         p = &flow->t_root.rb_node;
 510         parent = NULL;
 511
 512         while (*p) {
 513                 parent = *p;
 514                 aux = rb_to_skb(parent);
 515                 if (fq_skb_cb(skb)->time_to_send >= fq_skb_cb(aux)->time_to_send)
 516                         p = &parent->rb_right;
 517                 else
 518                         p = &parent->rb_left;
 519         }
 520         rb_link_node(&skb->rbnode, parent, p);
 521         rb_insert_color(&skb->rbnode, &flow->t_root);
 522 }
 523
 524 static bool fq_packet_beyond_horizon(const struct sk_buff *skb,
 525                                      const struct fq_sched_data *q, u64 now)
 526 {
 527         return unlikely((s64)skb->tstamp > (s64)(now + q->horizon));
 528 }
 529
 530 static int fq_enqueue(struct sk_buff *skb, struct Qdisc *sch,
 531                       struct sk_buff **to_free)
 532 {
 533         struct fq_sched_data *q = qdisc_priv(sch);
 534         struct fq_flow *f;
 535         u64 now;
 536         u8 band;
 537
 538         band = fq_prio2band(q->prio2band, skb->priority & TC_PRIO_MAX);
 539         if (unlikely(q->band_pkt_count[band] >= sch->limit)) {
 540                 q->stat_band_drops[band]++;
 541                 return qdisc_drop(skb, sch, to_free);
 542         }
 543
 544         now = ktime_get_ns();
 545         if (!skb->tstamp) {
 546                 fq_skb_cb(skb)->time_to_send = now;
 547         } else {
 548                 /* Check if packet timestamp is too far in the future. */
 549                 if (fq_packet_beyond_horizon(skb, q, now)) {
 550                         if (q->horizon_drop) {
 551                                         q->stat_horizon_drops++;
 552                                         return qdisc_drop(skb, sch, to_free);
 553                         }
 554                         q->stat_horizon_caps++;
 555                         skb->tstamp = now + q->horizon;
 556                 }
 557                 fq_skb_cb(skb)->time_to_send = skb->tstamp;
 558         }
 559
 560         f = fq_classify(sch, skb, now);
 561
 562         if (f != &q->internal) {
 563                 if (unlikely(f->qlen >= q->flow_plimit)) {
 564                         q->stat_flows_plimit++;
 565                         return qdisc_drop(skb, sch, to_free);
 566                 }
 567
 568                 if (fq_flow_is_detached(f)) {
 569                         fq_flow_add_tail(q, f, NEW_FLOW);
 570                         if (time_after(jiffies, f->age + q->flow_refill_delay))
 571                                 f->credit = max_t(u32, f->credit, q->quantum);
 572                 }
 573
 574                 f->band = band;
 575                 q->band_pkt_count[band]++;
 576                 fq_skb_cb(skb)->band = band;
 577                 if (f->qlen == 0)
 578                         q->inactive_flows--;
 579         }
 580
 581         f->qlen++;
 582         /* Note: this overwrites f->age */
 583         flow_queue_add(f, skb);
 584
 585         qdisc_qstats_backlog_inc(sch, skb);
 586         sch->q.qlen++;
 587
 588         return NET_XMIT_SUCCESS;
 589 }
 590
 591 static void fq_check_throttled(struct fq_sched_data *q, u64 now)
 592 {
 593         unsigned long sample;
 594         struct rb_node *p;
 595
 596         if (q->time_next_delayed_flow > now)
 597                 return;
 598
 599         /* Update unthrottle latency EWMA.
 600          * This is cheap and can help diagnosing timer/latency problems.
 601          */
 602         sample = (unsigned long)(now - q->time_next_delayed_flow);
 603         q->unthrottle_latency_ns -= q->unthrottle_latency_ns >> 3;
 604         q->unthrottle_latency_ns += sample >> 3;
 605
 606         q->time_next_delayed_flow = ~0ULL;
 607         while ((p = rb_first(&q->delayed)) != NULL) {
 608                 struct fq_flow *f = rb_entry(p, struct fq_flow, rate_node);
 609
 610                 if (f->time_next_packet > now) {
 611                         q->time_next_delayed_flow = f->time_next_packet;
 612                         break;
 613                 }
 614                 fq_flow_unset_throttled(q, f);
 615         }
 616 }
 617
 618 static struct fq_flow_head *fq_pband_head_select(struct fq_perband_flows *pband)
 619 {
 620         if (pband->credit <= 0)
 621                 return NULL;
 622
 623         if (pband->new_flows.first)
 624                 return &pband->new_flows;
 625
 626         return pband->old_flows.first ? &pband->old_flows : NULL;
 627 }
 628
 629 static struct sk_buff *fq_dequeue(struct Qdisc *sch)
 630 {
 631         struct fq_sched_data *q = qdisc_priv(sch);
 632         struct fq_perband_flows *pband;
 633         struct fq_flow_head *head;
 634         struct sk_buff *skb;
 635         struct fq_flow *f;
 636         unsigned long rate;
 637         int retry;
 638         u32 plen;
 639         u64 now;
 640
 641         if (!sch->q.qlen)
 642                 return NULL;
 643
 644         skb = fq_peek(&q->internal);
 645         if (unlikely(skb)) {
 646                 q->internal.qlen--;
 647                 fq_dequeue_skb(sch, &q->internal, skb);
 648                 goto out;
 649         }
 650
 651         now = ktime_get_ns();
 652         fq_check_throttled(q, now);
 653         retry = 0;
 654         pband = &q->band_flows[q->band_nr];
 655 begin:
 656         head = fq_pband_head_select(pband);
 657         if (!head) {
 658                 while (++retry <= FQ_BANDS) {
 659                         if (++q->band_nr == FQ_BANDS)
 660                                 q->band_nr = 0;
 661                         pband = &q->band_flows[q->band_nr];
 662                         pband->credit = min(pband->credit + pband->quantum,
 663                                             pband->quantum);
 664                         goto begin;
 665                 }
 666                 if (q->time_next_delayed_flow != ~0ULL)
 667                         qdisc_watchdog_schedule_range_ns(&q->watchdog,
 668                                                         q->time_next_delayed_flow,
 669                                                         q->timer_slack);
 670                 return NULL;
 671         }
 672         f = head->first;
 673         retry = 0;
 674         if (f->credit <= 0) {
 675                 f->credit += q->quantum;
 676                 head->first = f->next;
 677                 fq_flow_add_tail(q, f, OLD_FLOW);
 678                 goto begin;
 679         }
 680
 681         skb = fq_peek(f);
 682         if (skb) {
 683                 u64 time_next_packet = max_t(u64, fq_skb_cb(skb)->time_to_send,
 684                                              f->time_next_packet);
 685
 686                 if (now < time_next_packet) {
 687                         head->first = f->next;
 688                         f->time_next_packet = time_next_packet;
 689                         fq_flow_set_throttled(q, f);
 690                         goto begin;
 691                 }
 692                 prefetch(&skb->end);
 693                 if ((s64)(now - time_next_packet - q->ce_threshold) > 0) {
 694                         INET_ECN_set_ce(skb);
 695                         q->stat_ce_mark++;
 696                 }
 697                 if (--f->qlen == 0)
 698                         q->inactive_flows++;
 699                 q->band_pkt_count[fq_skb_cb(skb)->band]--;
 700                 fq_dequeue_skb(sch, f, skb);
 701         } else {
 702                 head->first = f->next;
 703                 /* force a pass through old_flows to prevent starvation */
 704                 if (head == &pband->new_flows) {
 705                         fq_flow_add_tail(q, f, OLD_FLOW);
 706                 } else {
 707                         fq_flow_set_detached(f);
 708                 }
 709                 goto begin;
 710         }
 711         plen = qdisc_pkt_len(skb);
 712         f->credit -= plen;
 713         pband->credit -= plen;
 714
 715         if (!q->rate_enable)
 716                 goto out;
 717
 718         rate = q->flow_max_rate;
 719
 720         /* If EDT time was provided for this skb, we need to
 721          * update f->time_next_packet only if this qdisc enforces
 722          * a flow max rate.
 723          */
 724         if (!skb->tstamp) {
 725                 if (skb->sk)
 726                         rate = min(READ_ONCE(skb->sk->sk_pacing_rate), rate);
 727
 728                 if (rate <= q->low_rate_threshold) {
 729                         f->credit = 0;
 730                 } else {
 731                         plen = max(plen, q->quantum);
 732                         if (f->credit > 0)
 733                                 goto out;
 734                 }
 735         }
 736         if (rate != ~0UL) {
 737                 u64 len = (u64)plen * NSEC_PER_SEC;
 738
 739                 if (likely(rate))
 740                         len = div64_ul(len, rate);
 741                 /* Since socket rate can change later,
 742                  * clamp the delay to 1 second.
 743                  * Really, providers of too big packets should be fixed !
 744                  */
 745                 if (unlikely(len > NSEC_PER_SEC)) {
 746                         len = NSEC_PER_SEC;
 747                         q->stat_pkts_too_long++;
 748                 }
 749                 /* Account for schedule/timers drifts.
 750                  * f->time_next_packet was set when prior packet was sent,
 751                  * and current time (@now) can be too late by tens of us.
 752                  */
 753                 if (f->time_next_packet)
 754                         len -= min(len/2, now - f->time_next_packet);
 755                 f->time_next_packet = now + len;
 756         }
 757 out:
 758         qdisc_bstats_update(sch, skb);
 759         return skb;
 760 }
 761
 762 static void fq_flow_purge(struct fq_flow *flow)
 763 {
 764         struct rb_node *p = rb_first(&flow->t_root);
 765
 766         while (p) {
 767                 struct sk_buff *skb = rb_to_skb(p);
 768
 769                 p = rb_next(p);
 770                 rb_erase(&skb->rbnode, &flow->t_root);
 771                 rtnl_kfree_skbs(skb, skb);
 772         }
 773         rtnl_kfree_skbs(flow->head, flow->tail);
 774         flow->head = NULL;
 775         flow->qlen = 0;
 776 }
 777
 778 static void fq_reset(struct Qdisc *sch)
 779 {
 780         struct fq_sched_data *q = qdisc_priv(sch);
 781         struct rb_root *root;
 782         struct rb_node *p;
 783         struct fq_flow *f;
 784         unsigned int idx;
 785
 786         sch->q.qlen = 0;
 787         sch->qstats.backlog = 0;
 788
 789         fq_flow_purge(&q->internal);
 790
 791         if (!q->fq_root)
 792                 return;
 793
 794         for (idx = 0; idx < (1U << q->fq_trees_log); idx++) {
 795                 root = &q->fq_root[idx];
 796                 while ((p = rb_first(root)) != NULL) {
 797                         f = rb_entry(p, struct fq_flow, fq_node);
 798                         rb_erase(p, root);
 799
 800                         fq_flow_purge(f);
 801
 802                         kmem_cache_free(fq_flow_cachep, f);
 803                 }
 804         }
 805         for (idx = 0; idx < FQ_BANDS; idx++) {
 806                 q->band_flows[idx].new_flows.first = NULL;
 807                 q->band_flows[idx].old_flows.first = NULL;
 808         }
 809         q->delayed              = RB_ROOT;
 810         q->flows                = 0;
 811         q->inactive_flows       = 0;
 812         q->throttled_flows      = 0;
 813 }
 814
 815 static void fq_rehash(struct fq_sched_data *q,
 816                       struct rb_root *old_array, u32 old_log,
 817                       struct rb_root *new_array, u32 new_log)
 818 {
 819         struct rb_node *op, **np, *parent;
 820         struct rb_root *oroot, *nroot;
 821         struct fq_flow *of, *nf;
 822         int fcnt = 0;
 823         u32 idx;
 824
 825         for (idx = 0; idx < (1U << old_log); idx++) {
 826                 oroot = &old_array[idx];
 827                 while ((op = rb_first(oroot)) != NULL) {
 828                         rb_erase(op, oroot);
 829                         of = rb_entry(op, struct fq_flow, fq_node);
 830                         if (fq_gc_candidate(of)) {
 831                                 fcnt++;
 832                                 kmem_cache_free(fq_flow_cachep, of);
 833                                 continue;
 834                         }
 835                         nroot = &new_array[hash_ptr(of->sk, new_log)];
 836
 837                         np = &nroot->rb_node;
 838                         parent = NULL;
 839                         while (*np) {
 840                                 parent = *np;
 841
 842                                 nf = rb_entry(parent, struct fq_flow, fq_node);
 843                                 BUG_ON(nf->sk == of->sk);
 844
 845                                 if (nf->sk > of->sk)
 846                                         np = &parent->rb_right;
 847                                 else
 848                                         np = &parent->rb_left;
 849                         }
 850
 851                         rb_link_node(&of->fq_node, parent, np);
 852                         rb_insert_color(&of->fq_node, nroot);
 853                 }
 854         }
 855         q->flows -= fcnt;
 856         q->inactive_flows -= fcnt;
 857         q->stat_gc_flows += fcnt;
 858 }
 859
 860 static void fq_free(void *addr)
 861 {
 862         kvfree(addr);
 863 }
 864
 865 static int fq_resize(struct Qdisc *sch, u32 log)
 866 {
 867         struct fq_sched_data *q = qdisc_priv(sch);
 868         struct rb_root *array;
 869         void *old_fq_root;
 870         u32 idx;
 871
 872         if (q->fq_root && log == q->fq_trees_log)
 873                 return 0;
 874
 875         /* If XPS was setup, we can allocate memory on right NUMA node */
 876         array = kvmalloc_node(sizeof(struct rb_root) << log, GFP_KERNEL | __GFP_RETRY_MAYFAIL,
 877                               netdev_queue_numa_node_read(sch->dev_queue));
 878         if (!array)
 879                 return -ENOMEM;
 880
 881         for (idx = 0; idx < (1U << log); idx++)
 882                 array[idx] = RB_ROOT;
 883
 884         sch_tree_lock(sch);
 885
 886         old_fq_root = q->fq_root;
 887         if (old_fq_root)
 888                 fq_rehash(q, old_fq_root, q->fq_trees_log, array, log);
 889
 890         q->fq_root = array;
 891         q->fq_trees_log = log;
 892
 893         sch_tree_unlock(sch);
 894
 895         fq_free(old_fq_root);
 896
 897         return 0;
 898 }
 899
 900 static const struct netlink_range_validation iq_range = {
 901         .max = INT_MAX,
 902 };
 903
 904 static const struct nla_policy fq_policy[TCA_FQ_MAX + 1] = {
 905         [TCA_FQ_UNSPEC]                 = { .strict_start_type = TCA_FQ_TIMER_SLACK },
 906
 907         [TCA_FQ_PLIMIT]                 = { .type = NLA_U32 },
 908         [TCA_FQ_FLOW_PLIMIT]            = { .type = NLA_U32 },
 909         [TCA_FQ_QUANTUM]                = { .type = NLA_U32 },
 910         [TCA_FQ_INITIAL_QUANTUM]        = NLA_POLICY_FULL_RANGE(NLA_U32, &iq_range),
 911         [TCA_FQ_RATE_ENABLE]            = { .type = NLA_U32 },
 912         [TCA_FQ_FLOW_DEFAULT_RATE]      = { .type = NLA_U32 },
 913         [TCA_FQ_FLOW_MAX_RATE]          = { .type = NLA_U32 },
 914         [TCA_FQ_BUCKETS_LOG]            = { .type = NLA_U32 },
 915         [TCA_FQ_FLOW_REFILL_DELAY]      = { .type = NLA_U32 },
 916         [TCA_FQ_ORPHAN_MASK]            = { .type = NLA_U32 },
 917         [TCA_FQ_LOW_RATE_THRESHOLD]     = { .type = NLA_U32 },
 918         [TCA_FQ_CE_THRESHOLD]           = { .type = NLA_U32 },
 919         [TCA_FQ_TIMER_SLACK]            = { .type = NLA_U32 },
 920         [TCA_FQ_HORIZON]                = { .type = NLA_U32 },
 921         [TCA_FQ_HORIZON_DROP]           = { .type = NLA_U8 },
 922         [TCA_FQ_PRIOMAP]                = NLA_POLICY_EXACT_LEN(sizeof(struct tc_prio_qopt)),
 923         [TCA_FQ_WEIGHTS]                = NLA_POLICY_EXACT_LEN(FQ_BANDS * sizeof(s32)),
 924 };
 925
 926 /* compress a u8 array with all elems <= 3 to an array of 2-bit fields */
 927 static void fq_prio2band_compress_crumb(const u8 *in, u8 *out)
 928 {
 929         const int num_elems = TC_PRIO_MAX + 1;
 930         int i;
 931
 932         memset(out, 0, num_elems / 4);
 933         for (i = 0; i < num_elems; i++)
 934                 out[i / 4] |= in[i] << (2 * (i & 0x3));
 935 }
 936
 937 static void fq_prio2band_decompress_crumb(const u8 *in, u8 *out)
 938 {
 939         const int num_elems = TC_PRIO_MAX + 1;
 940         int i;
 941
 942         for (i = 0; i < num_elems; i++)
 943                 out[i] = fq_prio2band(in, i);
 944 }
 945
 946 static int fq_load_weights(struct fq_sched_data *q,
 947                            const struct nlattr *attr,
 948                            struct netlink_ext_ack *extack)
 949 {
 950         s32 *weights = nla_data(attr);
 951         int i;
 952
 953         for (i = 0; i < FQ_BANDS; i++) {
 954                 if (weights[i] < FQ_MIN_WEIGHT) {
 955                         NL_SET_ERR_MSG_FMT_MOD(extack, "Weight %d less that minimum allowed %d",
 956                                                weights[i], FQ_MIN_WEIGHT);
 957                         return -EINVAL;
 958                 }
 959         }
 960         for (i = 0; i < FQ_BANDS; i++)
 961                 q->band_flows[i].quantum = weights[i];
 962         return 0;
 963 }
 964
 965 static int fq_load_priomap(struct fq_sched_data *q,
 966                            const struct nlattr *attr,
 967                            struct netlink_ext_ack *extack)
 968 {
 969         const struct tc_prio_qopt *map = nla_data(attr);
 970         int i;
 971
 972         if (map->bands != FQ_BANDS) {
 973                 NL_SET_ERR_MSG_MOD(extack, "FQ only supports 3 bands");
 974                 return -EINVAL;
 975         }
 976         for (i = 0; i < TC_PRIO_MAX + 1; i++) {
 977                 if (map->priomap[i] >= FQ_BANDS) {
 978                         NL_SET_ERR_MSG_FMT_MOD(extack, "FQ priomap field %d maps to a too high band %d",
 979                                                i, map->priomap[i]);
 980                         return -EINVAL;
 981                 }
 982         }
 983         fq_prio2band_compress_crumb(map->priomap, q->prio2band);
 984         return 0;
 985 }
 986
 987 static int fq_change(struct Qdisc *sch, struct nlattr *opt,
 988                      struct netlink_ext_ack *extack)
 989 {
 990         struct fq_sched_data *q = qdisc_priv(sch);
 991         struct nlattr *tb[TCA_FQ_MAX + 1];
 992         int err, drop_count = 0;
 993         unsigned drop_len = 0;
 994         u32 fq_log;
 995
 996         err = nla_parse_nested_deprecated(tb, TCA_FQ_MAX, opt, fq_policy,
 997                                           NULL);
 998         if (err < 0)
 999                 return err;
1000
1001         sch_tree_lock(sch);
1002
1003         fq_log = q->fq_trees_log;
1004
1005         if (tb[TCA_FQ_BUCKETS_LOG]) {
1006                 u32 nval = nla_get_u32(tb[TCA_FQ_BUCKETS_LOG]);
1007
1008                 if (nval >= 1 && nval <= ilog2(256*1024))
1009                         fq_log = nval;
1010                 else
1011                         err = -EINVAL;
1012         }
1013         if (tb[TCA_FQ_PLIMIT])
1014                 sch->limit = nla_get_u32(tb[TCA_FQ_PLIMIT]);
1015
1016         if (tb[TCA_FQ_FLOW_PLIMIT])
1017                 q->flow_plimit = nla_get_u32(tb[TCA_FQ_FLOW_PLIMIT]);
1018
1019         if (tb[TCA_FQ_QUANTUM]) {
1020                 u32 quantum = nla_get_u32(tb[TCA_FQ_QUANTUM]);
1021
1022                 if (quantum > 0 && quantum <= (1 << 20)) {
1023                         q->quantum = quantum;
1024                 } else {
1025                         NL_SET_ERR_MSG_MOD(extack, "invalid quantum");
1026                         err = -EINVAL;
1027                 }
1028         }
1029
1030         if (tb[TCA_FQ_INITIAL_QUANTUM])
1031                 q->initial_quantum = nla_get_u32(tb[TCA_FQ_INITIAL_QUANTUM]);
1032
1033         if (tb[TCA_FQ_FLOW_DEFAULT_RATE])
1034                 pr_warn_ratelimited("sch_fq: defrate %u ignored.\n",
1035                                     nla_get_u32(tb[TCA_FQ_FLOW_DEFAULT_RATE]));
1036
1037         if (tb[TCA_FQ_FLOW_MAX_RATE]) {
1038                 u32 rate = nla_get_u32(tb[TCA_FQ_FLOW_MAX_RATE]);
1039
1040                 q->flow_max_rate = (rate == ~0U) ? ~0UL : rate;
1041         }
1042         if (tb[TCA_FQ_LOW_RATE_THRESHOLD])
1043                 q->low_rate_threshold =
1044                         nla_get_u32(tb[TCA_FQ_LOW_RATE_THRESHOLD]);
1045
1046         if (tb[TCA_FQ_RATE_ENABLE]) {
1047                 u32 enable = nla_get_u32(tb[TCA_FQ_RATE_ENABLE]);
1048
1049                 if (enable <= 1)
1050                         q->rate_enable = enable;
1051                 else
1052                         err = -EINVAL;
1053         }
1054
1055         if (tb[TCA_FQ_FLOW_REFILL_DELAY]) {
1056                 u32 usecs_delay = nla_get_u32(tb[TCA_FQ_FLOW_REFILL_DELAY]) ;
1057
1058                 q->flow_refill_delay = usecs_to_jiffies(usecs_delay);
1059         }
1060
1061         if (!err && tb[TCA_FQ_PRIOMAP])
1062                 err = fq_load_priomap(q, tb[TCA_FQ_PRIOMAP], extack);
1063
1064         if (!err && tb[TCA_FQ_WEIGHTS])
1065                 err = fq_load_weights(q, tb[TCA_FQ_WEIGHTS], extack);
1066
1067         if (tb[TCA_FQ_ORPHAN_MASK])
1068                 q->orphan_mask = nla_get_u32(tb[TCA_FQ_ORPHAN_MASK]);
1069
1070         if (tb[TCA_FQ_CE_THRESHOLD])
1071                 q->ce_threshold = (u64)NSEC_PER_USEC *
1072                                   nla_get_u32(tb[TCA_FQ_CE_THRESHOLD]);
1073
1074         if (tb[TCA_FQ_TIMER_SLACK])
1075                 q->timer_slack = nla_get_u32(tb[TCA_FQ_TIMER_SLACK]);
1076
1077         if (tb[TCA_FQ_HORIZON])
1078                 q->horizon = (u64)NSEC_PER_USEC *
1079                                   nla_get_u32(tb[TCA_FQ_HORIZON]);
1080
1081         if (tb[TCA_FQ_HORIZON_DROP])
1082                 q->horizon_drop = nla_get_u8(tb[TCA_FQ_HORIZON_DROP]);
1083
1084         if (!err) {
1085
1086                 sch_tree_unlock(sch);
1087                 err = fq_resize(sch, fq_log);
1088                 sch_tree_lock(sch);
1089         }
1090         while (sch->q.qlen > sch->limit) {
1091                 struct sk_buff *skb = fq_dequeue(sch);
1092
1093                 if (!skb)
1094                         break;
1095                 drop_len += qdisc_pkt_len(skb);
1096                 rtnl_kfree_skbs(skb, skb);
1097                 drop_count++;
1098         }
1099         qdisc_tree_reduce_backlog(sch, drop_count, drop_len);
1100
1101         sch_tree_unlock(sch);
1102         return err;
1103 }
1104
1105 static void fq_destroy(struct Qdisc *sch)
1106 {
1107         struct fq_sched_data *q = qdisc_priv(sch);
1108
1109         fq_reset(sch);
1110         fq_free(q->fq_root);
1111         qdisc_watchdog_cancel(&q->watchdog);
1112 }
1113
1114 static int fq_init(struct Qdisc *sch, struct nlattr *opt,
1115                    struct netlink_ext_ack *extack)
1116 {
1117         struct fq_sched_data *q = qdisc_priv(sch);
1118         int i, err;
1119
1120         sch->limit              = 10000;
1121         q->flow_plimit          = 100;
1122         q->quantum              = 2 * psched_mtu(qdisc_dev(sch));
1123         q->initial_quantum      = 10 * psched_mtu(qdisc_dev(sch));
1124         q->flow_refill_delay    = msecs_to_jiffies(40);
1125         q->flow_max_rate        = ~0UL;
1126         q->time_next_delayed_flow = ~0ULL;
1127         q->rate_enable          = 1;
1128         for (i = 0; i < FQ_BANDS; i++) {
1129                 q->band_flows[i].new_flows.first = NULL;
1130                 q->band_flows[i].old_flows.first = NULL;
1131         }
1132         q->band_flows[0].quantum = 9 << 16;
1133         q->band_flows[1].quantum = 3 << 16;
1134         q->band_flows[2].quantum = 1 << 16;
1135         q->delayed              = RB_ROOT;
1136         q->fq_root              = NULL;
1137         q->fq_trees_log         = ilog2(1024);
1138         q->orphan_mask          = 1024 - 1;
1139         q->low_rate_threshold   = 550000 / 8;
1140
1141         q->timer_slack = 10 * NSEC_PER_USEC; /* 10 usec of hrtimer slack */
1142
1143         q->horizon = 10ULL * NSEC_PER_SEC; /* 10 seconds */
1144         q->horizon_drop = 1; /* by default, drop packets beyond horizon */
1145
1146         /* Default ce_threshold of 4294 seconds */
1147         q->ce_threshold         = (u64)NSEC_PER_USEC * ~0U;
1148
1149         fq_prio2band_compress_crumb(sch_default_prio2band, q->prio2band);
1150         qdisc_watchdog_init_clockid(&q->watchdog, sch, CLOCK_MONOTONIC);
1151
1152         if (opt)
1153                 err = fq_change(sch, opt, extack);
1154         else
1155                 err = fq_resize(sch, q->fq_trees_log);
1156
1157         return err;
1158 }
1159
1160 static int fq_dump(struct Qdisc *sch, struct sk_buff *skb)
1161 {
1162         struct fq_sched_data *q = qdisc_priv(sch);
1163         u64 ce_threshold = q->ce_threshold;
1164         struct tc_prio_qopt prio = {
1165                 .bands = FQ_BANDS,
1166         };
1167         u64 horizon = q->horizon;
1168         struct nlattr *opts;
1169         s32 weights[3];
1170
1171         opts = nla_nest_start_noflag(skb, TCA_OPTIONS);
1172         if (opts == NULL)
1173                 goto nla_put_failure;
1174
1175         /* TCA_FQ_FLOW_DEFAULT_RATE is not used anymore */
1176
1177         do_div(ce_threshold, NSEC_PER_USEC);
1178         do_div(horizon, NSEC_PER_USEC);
1179
1180         if (nla_put_u32(skb, TCA_FQ_PLIMIT, sch->limit) ||
1181             nla_put_u32(skb, TCA_FQ_FLOW_PLIMIT, q->flow_plimit) ||
1182             nla_put_u32(skb, TCA_FQ_QUANTUM, q->quantum) ||
1183             nla_put_u32(skb, TCA_FQ_INITIAL_QUANTUM, q->initial_quantum) ||
1184             nla_put_u32(skb, TCA_FQ_RATE_ENABLE, q->rate_enable) ||
1185             nla_put_u32(skb, TCA_FQ_FLOW_MAX_RATE,
1186                         min_t(unsigned long, q->flow_max_rate, ~0U)) ||
1187             nla_put_u32(skb, TCA_FQ_FLOW_REFILL_DELAY,
1188                         jiffies_to_usecs(q->flow_refill_delay)) ||
1189             nla_put_u32(skb, TCA_FQ_ORPHAN_MASK, q->orphan_mask) ||
1190             nla_put_u32(skb, TCA_FQ_LOW_RATE_THRESHOLD,
1191                         q->low_rate_threshold) ||
1192             nla_put_u32(skb, TCA_FQ_CE_THRESHOLD, (u32)ce_threshold) ||
1193             nla_put_u32(skb, TCA_FQ_BUCKETS_LOG, q->fq_trees_log) ||
1194             nla_put_u32(skb, TCA_FQ_TIMER_SLACK, q->timer_slack) ||
1195             nla_put_u32(skb, TCA_FQ_HORIZON, (u32)horizon) ||
1196             nla_put_u8(skb, TCA_FQ_HORIZON_DROP, q->horizon_drop))
1197                 goto nla_put_failure;
1198
1199         fq_prio2band_decompress_crumb(q->prio2band, prio.priomap);
1200         if (nla_put(skb, TCA_FQ_PRIOMAP, sizeof(prio), &prio))
1201                 goto nla_put_failure;
1202
1203         weights[0] = q->band_flows[0].quantum;
1204         weights[1] = q->band_flows[1].quantum;
1205         weights[2] = q->band_flows[2].quantum;
1206         if (nla_put(skb, TCA_FQ_WEIGHTS, sizeof(weights), &weights))
1207                 goto nla_put_failure;
1208
1209         return nla_nest_end(skb, opts);
1210
1211 nla_put_failure:
1212         return -1;
1213 }
1214
1215 static int fq_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
1216 {
1217         struct fq_sched_data *q = qdisc_priv(sch);
1218         struct tc_fq_qd_stats st;
1219         int i;
1220
1221         st.pad = 0;
1222
1223         sch_tree_lock(sch);
1224
1225         st.gc_flows               = q->stat_gc_flows;
1226         st.highprio_packets       = 0;
1227         st.fastpath_packets       = q->internal.stat_fastpath_packets;
1228         st.tcp_retrans            = 0;
1229         st.throttled              = q->stat_throttled;
1230         st.flows_plimit           = q->stat_flows_plimit;
1231         st.pkts_too_long          = q->stat_pkts_too_long;
1232         st.allocation_errors      = q->stat_allocation_errors;
1233         st.time_next_delayed_flow = q->time_next_delayed_flow + q->timer_slack -
1234                                     ktime_get_ns();
1235         st.flows                  = q->flows;
1236         st.inactive_flows         = q->inactive_flows;
1237         st.throttled_flows        = q->throttled_flows;
1238         st.unthrottle_latency_ns  = min_t(unsigned long,
1239                                           q->unthrottle_latency_ns, ~0U);
1240         st.ce_mark                = q->stat_ce_mark;
1241         st.horizon_drops          = q->stat_horizon_drops;
1242         st.horizon_caps           = q->stat_horizon_caps;
1243         for (i = 0; i < FQ_BANDS; i++) {
1244                 st.band_drops[i]  = q->stat_band_drops[i];
1245                 st.band_pkt_count[i] = q->band_pkt_count[i];
1246         }
1247         sch_tree_unlock(sch);
1248
1249         return gnet_stats_copy_app(d, &st, sizeof(st));
1250 }
1251
1252 static struct Qdisc_ops fq_qdisc_ops __read_mostly = {
1253         .id             =       "fq",
1254         .priv_size      =       sizeof(struct fq_sched_data),
1255
1256         .enqueue        =       fq_enqueue,
1257         .dequeue        =       fq_dequeue,
1258         .peek           =       qdisc_peek_dequeued,
1259         .init           =       fq_init,
1260         .reset          =       fq_reset,
1261         .destroy        =       fq_destroy,
1262         .change         =       fq_change,
1263         .dump           =       fq_dump,
1264         .dump_stats     =       fq_dump_stats,
1265         .owner          =       THIS_MODULE,
1266 };
1267 MODULE_ALIAS_NET_SCH("fq");
1268
1269 static int __init fq_module_init(void)
1270 {
1271         int ret;
1272
1273         fq_flow_cachep = kmem_cache_create("fq_flow_cache",
1274                                            sizeof(struct fq_flow),
1275                                            0, SLAB_HWCACHE_ALIGN, NULL);
1276         if (!fq_flow_cachep)
1277                 return -ENOMEM;
1278
1279         ret = register_qdisc(&fq_qdisc_ops);
1280         if (ret)
1281                 kmem_cache_destroy(fq_flow_cachep);
1282         return ret;
1283 }
1284
1285 static void __exit fq_module_exit(void)
1286 {
1287         unregister_qdisc(&fq_qdisc_ops);
1288         kmem_cache_destroy(fq_flow_cachep);
1289 }
1290
1291 module_init(fq_module_init)
1292 module_exit(fq_module_exit)
1293 MODULE_AUTHOR("Eric Dumazet");
1294 MODULE_LICENSE("GPL");
1295 MODULE_DESCRIPTION("Fair Queue Packet Scheduler");