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 struct fq_skb_cb {
  58         u64             time_to_send;
  59 };
  60
  61 static inline struct fq_skb_cb *fq_skb_cb(struct sk_buff *skb)
  62 {
  63         qdisc_cb_private_validate(skb, sizeof(struct fq_skb_cb));
  64         return (struct fq_skb_cb *)qdisc_skb_cb(skb)->data;
  65 }
  66
  67 /*
  68  * Per flow structure, dynamically allocated.
  69  * If packets have monotically increasing time_to_send, they are placed in O(1)
  70  * in linear list (head,tail), otherwise are placed in a rbtree (t_root).
  71  */
  72 struct fq_flow {
  73         struct rb_root  t_root;
  74         struct sk_buff  *head;          /* list of skbs for this flow : first skb */
  75         union {
  76                 struct sk_buff *tail;   /* last skb in the list */
  77                 unsigned long  age;     /* jiffies when flow was emptied, for gc */
  78         };
  79         struct rb_node  fq_node;        /* anchor in fq_root[] trees */
  80         struct sock     *sk;
  81         int             qlen;           /* number of packets in flow queue */
  82         int             credit;
  83         u32             socket_hash;    /* sk_hash */
  84         struct fq_flow *next;           /* next pointer in RR lists, or &detached */
  85
  86         struct rb_node  rate_node;      /* anchor in q->delayed tree */
  87         u64             time_next_packet;
  88 };
  89
  90 struct fq_flow_head {
  91         struct fq_flow *first;
  92         struct fq_flow *last;
  93 };
  94
  95 struct fq_sched_data {
  96         struct fq_flow_head new_flows;
  97
  98         struct fq_flow_head old_flows;
  99
 100         struct rb_root  delayed;        /* for rate limited flows */
 101         u64             time_next_delayed_flow;
 102         unsigned long   unthrottle_latency_ns;
 103
 104         struct fq_flow  internal;       /* for non classified or high prio packets */
 105         u32             quantum;
 106         u32             initial_quantum;
 107         u32             flow_refill_delay;
 108         u32             flow_plimit;    /* max packets per flow */
 109         unsigned long   flow_max_rate;  /* optional max rate per flow */
 110         u64             ce_threshold;
 111         u32             orphan_mask;    /* mask for orphaned skb */
 112         u32             low_rate_threshold;
 113         struct rb_root  *fq_root;
 114         u8              rate_enable;
 115         u8              fq_trees_log;
 116
 117         u32             flows;
 118         u32             inactive_flows;
 119         u32             throttled_flows;
 120
 121         u64             stat_gc_flows;
 122         u64             stat_internal_packets;
 123         u64             stat_throttled;
 124         u64             stat_ce_mark;
 125         u64             stat_flows_plimit;
 126         u64             stat_pkts_too_long;
 127         u64             stat_allocation_errors;
 128         struct qdisc_watchdog watchdog;
 129 };
 130
 131 /* special value to mark a detached flow (not on old/new list) */
 132 static struct fq_flow detached, throttled;
 133
 134 static void fq_flow_set_detached(struct fq_flow *f)
 135 {
 136         f->next = &detached;
 137         f->age = jiffies;
 138 }
 139
 140 static bool fq_flow_is_detached(const struct fq_flow *f)
 141 {
 142         return f->next == &detached;
 143 }
 144
 145 static bool fq_flow_is_throttled(const struct fq_flow *f)
 146 {
 147         return f->next == &throttled;
 148 }
 149
 150 static void fq_flow_add_tail(struct fq_flow_head *head, struct fq_flow *flow)
 151 {
 152         if (head->first)
 153                 head->last->next = flow;
 154         else
 155                 head->first = flow;
 156         head->last = flow;
 157         flow->next = NULL;
 158 }
 159
 160 static void fq_flow_unset_throttled(struct fq_sched_data *q, struct fq_flow *f)
 161 {
 162         rb_erase(&f->rate_node, &q->delayed);
 163         q->throttled_flows--;
 164         fq_flow_add_tail(&q->old_flows, f);
 165 }
 166
 167 static void fq_flow_set_throttled(struct fq_sched_data *q, struct fq_flow *f)
 168 {
 169         struct rb_node **p = &q->delayed.rb_node, *parent = NULL;
 170
 171         while (*p) {
 172                 struct fq_flow *aux;
 173
 174                 parent = *p;
 175                 aux = rb_entry(parent, struct fq_flow, rate_node);
 176                 if (f->time_next_packet >= aux->time_next_packet)
 177                         p = &parent->rb_right;
 178                 else
 179                         p = &parent->rb_left;
 180         }
 181         rb_link_node(&f->rate_node, parent, p);
 182         rb_insert_color(&f->rate_node, &q->delayed);
 183         q->throttled_flows++;
 184         q->stat_throttled++;
 185
 186         f->next = &throttled;
 187         if (q->time_next_delayed_flow > f->time_next_packet)
 188                 q->time_next_delayed_flow = f->time_next_packet;
 189 }
 190
 191
 192 static struct kmem_cache *fq_flow_cachep __read_mostly;
 193
 194
 195 /* limit number of collected flows per round */
 196 #define FQ_GC_MAX 8
 197 #define FQ_GC_AGE (3*HZ)
 198
 199 static bool fq_gc_candidate(const struct fq_flow *f)
 200 {
 201         return fq_flow_is_detached(f) &&
 202                time_after(jiffies, f->age + FQ_GC_AGE);
 203 }
 204
 205 static void fq_gc(struct fq_sched_data *q,
 206                   struct rb_root *root,
 207                   struct sock *sk)
 208 {
 209         struct fq_flow *f, *tofree[FQ_GC_MAX];
 210         struct rb_node **p, *parent;
 211         int fcnt = 0;
 212
 213         p = &root->rb_node;
 214         parent = NULL;
 215         while (*p) {
 216                 parent = *p;
 217
 218                 f = rb_entry(parent, struct fq_flow, fq_node);
 219                 if (f->sk == sk)
 220                         break;
 221
 222                 if (fq_gc_candidate(f)) {
 223                         tofree[fcnt++] = f;
 224                         if (fcnt == FQ_GC_MAX)
 225                                 break;
 226                 }
 227
 228                 if (f->sk > sk)
 229                         p = &parent->rb_right;
 230                 else
 231                         p = &parent->rb_left;
 232         }
 233
 234         q->flows -= fcnt;
 235         q->inactive_flows -= fcnt;
 236         q->stat_gc_flows += fcnt;
 237         while (fcnt) {
 238                 struct fq_flow *f = tofree[--fcnt];
 239
 240                 rb_erase(&f->fq_node, root);
 241                 kmem_cache_free(fq_flow_cachep, f);
 242         }
 243 }
 244
 245 static struct fq_flow *fq_classify(struct sk_buff *skb, struct fq_sched_data *q)
 246 {
 247         struct rb_node **p, *parent;
 248         struct sock *sk = skb->sk;
 249         struct rb_root *root;
 250         struct fq_flow *f;
 251
 252         /* warning: no starvation prevention... */
 253         if (unlikely((skb->priority & TC_PRIO_MAX) == TC_PRIO_CONTROL))
 254                 return &q->internal;
 255
 256         /* SYNACK messages are attached to a TCP_NEW_SYN_RECV request socket
 257          * or a listener (SYNCOOKIE mode)
 258          * 1) request sockets are not full blown,
 259          *    they do not contain sk_pacing_rate
 260          * 2) They are not part of a 'flow' yet
 261          * 3) We do not want to rate limit them (eg SYNFLOOD attack),
 262          *    especially if the listener set SO_MAX_PACING_RATE
 263          * 4) We pretend they are orphaned
 264          */
 265         if (!sk || sk_listener(sk)) {
 266                 unsigned long hash = skb_get_hash(skb) & q->orphan_mask;
 267
 268                 /* By forcing low order bit to 1, we make sure to not
 269                  * collide with a local flow (socket pointers are word aligned)
 270                  */
 271                 sk = (struct sock *)((hash << 1) | 1UL);
 272                 skb_orphan(skb);
 273         } else if (sk->sk_state == TCP_CLOSE) {
 274                 unsigned long hash = skb_get_hash(skb) & q->orphan_mask;
 275                 /*
 276                  * Sockets in TCP_CLOSE are non connected.
 277                  * Typical use case is UDP sockets, they can send packets
 278                  * with sendto() to many different destinations.
 279                  * We probably could use a generic bit advertising
 280                  * non connected sockets, instead of sk_state == TCP_CLOSE,
 281                  * if we care enough.
 282                  */
 283                 sk = (struct sock *)((hash << 1) | 1UL);
 284         }
 285
 286         root = &q->fq_root[hash_ptr(sk, q->fq_trees_log)];
 287
 288         if (q->flows >= (2U << q->fq_trees_log) &&
 289             q->inactive_flows > q->flows/2)
 290                 fq_gc(q, root, sk);
 291
 292         p = &root->rb_node;
 293         parent = NULL;
 294         while (*p) {
 295                 parent = *p;
 296
 297                 f = rb_entry(parent, struct fq_flow, fq_node);
 298                 if (f->sk == sk) {
 299                         /* socket might have been reallocated, so check
 300                          * if its sk_hash is the same.
 301                          * It not, we need to refill credit with
 302                          * initial quantum
 303                          */
 304                         if (unlikely(skb->sk == sk &&
 305                                      f->socket_hash != sk->sk_hash)) {
 306                                 f->credit = q->initial_quantum;
 307                                 f->socket_hash = sk->sk_hash;
 308                                 if (fq_flow_is_throttled(f))
 309                                         fq_flow_unset_throttled(q, f);
 310                                 f->time_next_packet = 0ULL;
 311                         }
 312                         return f;
 313                 }
 314                 if (f->sk > sk)
 315                         p = &parent->rb_right;
 316                 else
 317                         p = &parent->rb_left;
 318         }
 319
 320         f = kmem_cache_zalloc(fq_flow_cachep, GFP_ATOMIC | __GFP_NOWARN);
 321         if (unlikely(!f)) {
 322                 q->stat_allocation_errors++;
 323                 return &q->internal;
 324         }
 325         /* f->t_root is already zeroed after kmem_cache_zalloc() */
 326
 327         fq_flow_set_detached(f);
 328         f->sk = sk;
 329         if (skb->sk == sk)
 330                 f->socket_hash = sk->sk_hash;
 331         f->credit = q->initial_quantum;
 332
 333         rb_link_node(&f->fq_node, parent, p);
 334         rb_insert_color(&f->fq_node, root);
 335
 336         q->flows++;
 337         q->inactive_flows++;
 338         return f;
 339 }
 340
 341 static struct sk_buff *fq_peek(struct fq_flow *flow)
 342 {
 343         struct sk_buff *skb = skb_rb_first(&flow->t_root);
 344         struct sk_buff *head = flow->head;
 345
 346         if (!skb)
 347                 return head;
 348
 349         if (!head)
 350                 return skb;
 351
 352         if (fq_skb_cb(skb)->time_to_send < fq_skb_cb(head)->time_to_send)
 353                 return skb;
 354         return head;
 355 }
 356
 357 static void fq_erase_head(struct Qdisc *sch, struct fq_flow *flow,
 358                           struct sk_buff *skb)
 359 {
 360         if (skb == flow->head) {
 361                 flow->head = skb->next;
 362         } else {
 363                 rb_erase(&skb->rbnode, &flow->t_root);
 364                 skb->dev = qdisc_dev(sch);
 365         }
 366 }
 367
 368 /* remove one skb from head of flow queue */
 369 static struct sk_buff *fq_dequeue_head(struct Qdisc *sch, struct fq_flow *flow)
 370 {
 371         struct sk_buff *skb = fq_peek(flow);
 372
 373         if (skb) {
 374                 fq_erase_head(sch, flow, skb);
 375                 skb_mark_not_on_list(skb);
 376                 flow->qlen--;
 377                 qdisc_qstats_backlog_dec(sch, skb);
 378                 sch->q.qlen--;
 379         }
 380         return skb;
 381 }
 382
 383 static void flow_queue_add(struct fq_flow *flow, struct sk_buff *skb)
 384 {
 385         struct rb_node **p, *parent;
 386         struct sk_buff *head, *aux;
 387
 388         fq_skb_cb(skb)->time_to_send = skb->tstamp ?: ktime_get_ns();
 389
 390         head = flow->head;
 391         if (!head ||
 392             fq_skb_cb(skb)->time_to_send >= fq_skb_cb(flow->tail)->time_to_send) {
 393                 if (!head)
 394                         flow->head = skb;
 395                 else
 396                         flow->tail->next = skb;
 397                 flow->tail = skb;
 398                 skb->next = NULL;
 399                 return;
 400         }
 401
 402         p = &flow->t_root.rb_node;
 403         parent = NULL;
 404
 405         while (*p) {
 406                 parent = *p;
 407                 aux = rb_to_skb(parent);
 408                 if (fq_skb_cb(skb)->time_to_send >= fq_skb_cb(aux)->time_to_send)
 409                         p = &parent->rb_right;
 410                 else
 411                         p = &parent->rb_left;
 412         }
 413         rb_link_node(&skb->rbnode, parent, p);
 414         rb_insert_color(&skb->rbnode, &flow->t_root);
 415 }
 416
 417 static int fq_enqueue(struct sk_buff *skb, struct Qdisc *sch,
 418                       struct sk_buff **to_free)
 419 {
 420         struct fq_sched_data *q = qdisc_priv(sch);
 421         struct fq_flow *f;
 422
 423         if (unlikely(sch->q.qlen >= sch->limit))
 424                 return qdisc_drop(skb, sch, to_free);
 425
 426         f = fq_classify(skb, q);
 427         if (unlikely(f->qlen >= q->flow_plimit && f != &q->internal)) {
 428                 q->stat_flows_plimit++;
 429                 return qdisc_drop(skb, sch, to_free);
 430         }
 431
 432         f->qlen++;
 433         qdisc_qstats_backlog_inc(sch, skb);
 434         if (fq_flow_is_detached(f)) {
 435                 struct sock *sk = skb->sk;
 436
 437                 fq_flow_add_tail(&q->new_flows, f);
 438                 if (time_after(jiffies, f->age + q->flow_refill_delay))
 439                         f->credit = max_t(u32, f->credit, q->quantum);
 440                 if (sk && q->rate_enable) {
 441                         if (unlikely(smp_load_acquire(&sk->sk_pacing_status) !=
 442                                      SK_PACING_FQ))
 443                                 smp_store_release(&sk->sk_pacing_status,
 444                                                   SK_PACING_FQ);
 445                 }
 446                 q->inactive_flows--;
 447         }
 448
 449         /* Note: this overwrites f->age */
 450         flow_queue_add(f, skb);
 451
 452         if (unlikely(f == &q->internal)) {
 453                 q->stat_internal_packets++;
 454         }
 455         sch->q.qlen++;
 456
 457         return NET_XMIT_SUCCESS;
 458 }
 459
 460 static void fq_check_throttled(struct fq_sched_data *q, u64 now)
 461 {
 462         unsigned long sample;
 463         struct rb_node *p;
 464
 465         if (q->time_next_delayed_flow > now)
 466                 return;
 467
 468         /* Update unthrottle latency EWMA.
 469          * This is cheap and can help diagnosing timer/latency problems.
 470          */
 471         sample = (unsigned long)(now - q->time_next_delayed_flow);
 472         q->unthrottle_latency_ns -= q->unthrottle_latency_ns >> 3;
 473         q->unthrottle_latency_ns += sample >> 3;
 474
 475         q->time_next_delayed_flow = ~0ULL;
 476         while ((p = rb_first(&q->delayed)) != NULL) {
 477                 struct fq_flow *f = rb_entry(p, struct fq_flow, rate_node);
 478
 479                 if (f->time_next_packet > now) {
 480                         q->time_next_delayed_flow = f->time_next_packet;
 481                         break;
 482                 }
 483                 fq_flow_unset_throttled(q, f);
 484         }
 485 }
 486
 487 static struct sk_buff *fq_dequeue(struct Qdisc *sch)
 488 {
 489         struct fq_sched_data *q = qdisc_priv(sch);
 490         struct fq_flow_head *head;
 491         struct sk_buff *skb;
 492         struct fq_flow *f;
 493         unsigned long rate;
 494         u32 plen;
 495         u64 now;
 496
 497         if (!sch->q.qlen)
 498                 return NULL;
 499
 500         skb = fq_dequeue_head(sch, &q->internal);
 501         if (skb)
 502                 goto out;
 503
 504         now = ktime_get_ns();
 505         fq_check_throttled(q, now);
 506 begin:
 507         head = &q->new_flows;
 508         if (!head->first) {
 509                 head = &q->old_flows;
 510                 if (!head->first) {
 511                         if (q->time_next_delayed_flow != ~0ULL)
 512                                 qdisc_watchdog_schedule_ns(&q->watchdog,
 513                                                            q->time_next_delayed_flow);
 514                         return NULL;
 515                 }
 516         }
 517         f = head->first;
 518
 519         if (f->credit <= 0) {
 520                 f->credit += q->quantum;
 521                 head->first = f->next;
 522                 fq_flow_add_tail(&q->old_flows, f);
 523                 goto begin;
 524         }
 525
 526         skb = fq_peek(f);
 527         if (skb) {
 528                 u64 time_next_packet = max_t(u64, fq_skb_cb(skb)->time_to_send,
 529                                              f->time_next_packet);
 530
 531                 if (now < time_next_packet) {
 532                         head->first = f->next;
 533                         f->time_next_packet = time_next_packet;
 534                         fq_flow_set_throttled(q, f);
 535                         goto begin;
 536                 }
 537                 if (time_next_packet &&
 538                     (s64)(now - time_next_packet - q->ce_threshold) > 0) {
 539                         INET_ECN_set_ce(skb);
 540                         q->stat_ce_mark++;
 541                 }
 542         }
 543
 544         skb = fq_dequeue_head(sch, f);
 545         if (!skb) {
 546                 head->first = f->next;
 547                 /* force a pass through old_flows to prevent starvation */
 548                 if ((head == &q->new_flows) && q->old_flows.first) {
 549                         fq_flow_add_tail(&q->old_flows, f);
 550                 } else {
 551                         fq_flow_set_detached(f);
 552                         q->inactive_flows++;
 553                 }
 554                 goto begin;
 555         }
 556         prefetch(&skb->end);
 557         plen = qdisc_pkt_len(skb);
 558         f->credit -= plen;
 559
 560         if (!q->rate_enable)
 561                 goto out;
 562
 563         rate = q->flow_max_rate;
 564
 565         /* If EDT time was provided for this skb, we need to
 566          * update f->time_next_packet only if this qdisc enforces
 567          * a flow max rate.
 568          */
 569         if (!skb->tstamp) {
 570                 if (skb->sk)
 571                         rate = min(skb->sk->sk_pacing_rate, rate);
 572
 573                 if (rate <= q->low_rate_threshold) {
 574                         f->credit = 0;
 575                 } else {
 576                         plen = max(plen, q->quantum);
 577                         if (f->credit > 0)
 578                                 goto out;
 579                 }
 580         }
 581         if (rate != ~0UL) {
 582                 u64 len = (u64)plen * NSEC_PER_SEC;
 583
 584                 if (likely(rate))
 585                         len = div64_ul(len, rate);
 586                 /* Since socket rate can change later,
 587                  * clamp the delay to 1 second.
 588                  * Really, providers of too big packets should be fixed !
 589                  */
 590                 if (unlikely(len > NSEC_PER_SEC)) {
 591                         len = NSEC_PER_SEC;
 592                         q->stat_pkts_too_long++;
 593                 }
 594                 /* Account for schedule/timers drifts.
 595                  * f->time_next_packet was set when prior packet was sent,
 596                  * and current time (@now) can be too late by tens of us.
 597                  */
 598                 if (f->time_next_packet)
 599                         len -= min(len/2, now - f->time_next_packet);
 600                 f->time_next_packet = now + len;
 601         }
 602 out:
 603         qdisc_bstats_update(sch, skb);
 604         return skb;
 605 }
 606
 607 static void fq_flow_purge(struct fq_flow *flow)
 608 {
 609         struct rb_node *p = rb_first(&flow->t_root);
 610
 611         while (p) {
 612                 struct sk_buff *skb = rb_to_skb(p);
 613
 614                 p = rb_next(p);
 615                 rb_erase(&skb->rbnode, &flow->t_root);
 616                 rtnl_kfree_skbs(skb, skb);
 617         }
 618         rtnl_kfree_skbs(flow->head, flow->tail);
 619         flow->head = NULL;
 620         flow->qlen = 0;
 621 }
 622
 623 static void fq_reset(struct Qdisc *sch)
 624 {
 625         struct fq_sched_data *q = qdisc_priv(sch);
 626         struct rb_root *root;
 627         struct rb_node *p;
 628         struct fq_flow *f;
 629         unsigned int idx;
 630
 631         sch->q.qlen = 0;
 632         sch->qstats.backlog = 0;
 633
 634         fq_flow_purge(&q->internal);
 635
 636         if (!q->fq_root)
 637                 return;
 638
 639         for (idx = 0; idx < (1U << q->fq_trees_log); idx++) {
 640                 root = &q->fq_root[idx];
 641                 while ((p = rb_first(root)) != NULL) {
 642                         f = rb_entry(p, struct fq_flow, fq_node);
 643                         rb_erase(p, root);
 644
 645                         fq_flow_purge(f);
 646
 647                         kmem_cache_free(fq_flow_cachep, f);
 648                 }
 649         }
 650         q->new_flows.first      = NULL;
 651         q->old_flows.first      = NULL;
 652         q->delayed              = RB_ROOT;
 653         q->flows                = 0;
 654         q->inactive_flows       = 0;
 655         q->throttled_flows      = 0;
 656 }
 657
 658 static void fq_rehash(struct fq_sched_data *q,
 659                       struct rb_root *old_array, u32 old_log,
 660                       struct rb_root *new_array, u32 new_log)
 661 {
 662         struct rb_node *op, **np, *parent;
 663         struct rb_root *oroot, *nroot;
 664         struct fq_flow *of, *nf;
 665         int fcnt = 0;
 666         u32 idx;
 667
 668         for (idx = 0; idx < (1U << old_log); idx++) {
 669                 oroot = &old_array[idx];
 670                 while ((op = rb_first(oroot)) != NULL) {
 671                         rb_erase(op, oroot);
 672                         of = rb_entry(op, struct fq_flow, fq_node);
 673                         if (fq_gc_candidate(of)) {
 674                                 fcnt++;
 675                                 kmem_cache_free(fq_flow_cachep, of);
 676                                 continue;
 677                         }
 678                         nroot = &new_array[hash_ptr(of->sk, new_log)];
 679
 680                         np = &nroot->rb_node;
 681                         parent = NULL;
 682                         while (*np) {
 683                                 parent = *np;
 684
 685                                 nf = rb_entry(parent, struct fq_flow, fq_node);
 686                                 BUG_ON(nf->sk == of->sk);
 687
 688                                 if (nf->sk > of->sk)
 689                                         np = &parent->rb_right;
 690                                 else
 691                                         np = &parent->rb_left;
 692                         }
 693
 694                         rb_link_node(&of->fq_node, parent, np);
 695                         rb_insert_color(&of->fq_node, nroot);
 696                 }
 697         }
 698         q->flows -= fcnt;
 699         q->inactive_flows -= fcnt;
 700         q->stat_gc_flows += fcnt;
 701 }
 702
 703 static void fq_free(void *addr)
 704 {
 705         kvfree(addr);
 706 }
 707
 708 static int fq_resize(struct Qdisc *sch, u32 log)
 709 {
 710         struct fq_sched_data *q = qdisc_priv(sch);
 711         struct rb_root *array;
 712         void *old_fq_root;
 713         u32 idx;
 714
 715         if (q->fq_root && log == q->fq_trees_log)
 716                 return 0;
 717
 718         /* If XPS was setup, we can allocate memory on right NUMA node */
 719         array = kvmalloc_node(sizeof(struct rb_root) << log, GFP_KERNEL | __GFP_RETRY_MAYFAIL,
 720                               netdev_queue_numa_node_read(sch->dev_queue));
 721         if (!array)
 722                 return -ENOMEM;
 723
 724         for (idx = 0; idx < (1U << log); idx++)
 725                 array[idx] = RB_ROOT;
 726
 727         sch_tree_lock(sch);
 728
 729         old_fq_root = q->fq_root;
 730         if (old_fq_root)
 731                 fq_rehash(q, old_fq_root, q->fq_trees_log, array, log);
 732
 733         q->fq_root = array;
 734         q->fq_trees_log = log;
 735
 736         sch_tree_unlock(sch);
 737
 738         fq_free(old_fq_root);
 739
 740         return 0;
 741 }
 742
 743 static const struct nla_policy fq_policy[TCA_FQ_MAX + 1] = {
 744         [TCA_FQ_PLIMIT]                 = { .type = NLA_U32 },
 745         [TCA_FQ_FLOW_PLIMIT]            = { .type = NLA_U32 },
 746         [TCA_FQ_QUANTUM]                = { .type = NLA_U32 },
 747         [TCA_FQ_INITIAL_QUANTUM]        = { .type = NLA_U32 },
 748         [TCA_FQ_RATE_ENABLE]            = { .type = NLA_U32 },
 749         [TCA_FQ_FLOW_DEFAULT_RATE]      = { .type = NLA_U32 },
 750         [TCA_FQ_FLOW_MAX_RATE]          = { .type = NLA_U32 },
 751         [TCA_FQ_BUCKETS_LOG]            = { .type = NLA_U32 },
 752         [TCA_FQ_FLOW_REFILL_DELAY]      = { .type = NLA_U32 },
 753         [TCA_FQ_LOW_RATE_THRESHOLD]     = { .type = NLA_U32 },
 754         [TCA_FQ_CE_THRESHOLD]           = { .type = NLA_U32 },
 755 };
 756
 757 static int fq_change(struct Qdisc *sch, struct nlattr *opt,
 758                      struct netlink_ext_ack *extack)
 759 {
 760         struct fq_sched_data *q = qdisc_priv(sch);
 761         struct nlattr *tb[TCA_FQ_MAX + 1];
 762         int err, drop_count = 0;
 763         unsigned drop_len = 0;
 764         u32 fq_log;
 765
 766         if (!opt)
 767                 return -EINVAL;
 768
 769         err = nla_parse_nested_deprecated(tb, TCA_FQ_MAX, opt, fq_policy,
 770                                           NULL);
 771         if (err < 0)
 772                 return err;
 773
 774         sch_tree_lock(sch);
 775
 776         fq_log = q->fq_trees_log;
 777
 778         if (tb[TCA_FQ_BUCKETS_LOG]) {
 779                 u32 nval = nla_get_u32(tb[TCA_FQ_BUCKETS_LOG]);
 780
 781                 if (nval >= 1 && nval <= ilog2(256*1024))
 782                         fq_log = nval;
 783                 else
 784                         err = -EINVAL;
 785         }
 786         if (tb[TCA_FQ_PLIMIT])
 787                 sch->limit = nla_get_u32(tb[TCA_FQ_PLIMIT]);
 788
 789         if (tb[TCA_FQ_FLOW_PLIMIT])
 790                 q->flow_plimit = nla_get_u32(tb[TCA_FQ_FLOW_PLIMIT]);
 791
 792         if (tb[TCA_FQ_QUANTUM]) {
 793                 u32 quantum = nla_get_u32(tb[TCA_FQ_QUANTUM]);
 794
 795                 if (quantum > 0)
 796                         q->quantum = quantum;
 797                 else
 798                         err = -EINVAL;
 799         }
 800
 801         if (tb[TCA_FQ_INITIAL_QUANTUM])
 802                 q->initial_quantum = nla_get_u32(tb[TCA_FQ_INITIAL_QUANTUM]);
 803
 804         if (tb[TCA_FQ_FLOW_DEFAULT_RATE])
 805                 pr_warn_ratelimited("sch_fq: defrate %u ignored.\n",
 806                                     nla_get_u32(tb[TCA_FQ_FLOW_DEFAULT_RATE]));
 807
 808         if (tb[TCA_FQ_FLOW_MAX_RATE]) {
 809                 u32 rate = nla_get_u32(tb[TCA_FQ_FLOW_MAX_RATE]);
 810
 811                 q->flow_max_rate = (rate == ~0U) ? ~0UL : rate;
 812         }
 813         if (tb[TCA_FQ_LOW_RATE_THRESHOLD])
 814                 q->low_rate_threshold =
 815                         nla_get_u32(tb[TCA_FQ_LOW_RATE_THRESHOLD]);
 816
 817         if (tb[TCA_FQ_RATE_ENABLE]) {
 818                 u32 enable = nla_get_u32(tb[TCA_FQ_RATE_ENABLE]);
 819
 820                 if (enable <= 1)
 821                         q->rate_enable = enable;
 822                 else
 823                         err = -EINVAL;
 824         }
 825
 826         if (tb[TCA_FQ_FLOW_REFILL_DELAY]) {
 827                 u32 usecs_delay = nla_get_u32(tb[TCA_FQ_FLOW_REFILL_DELAY]) ;
 828
 829                 q->flow_refill_delay = usecs_to_jiffies(usecs_delay);
 830         }
 831
 832         if (tb[TCA_FQ_ORPHAN_MASK])
 833                 q->orphan_mask = nla_get_u32(tb[TCA_FQ_ORPHAN_MASK]);
 834
 835         if (tb[TCA_FQ_CE_THRESHOLD])
 836                 q->ce_threshold = (u64)NSEC_PER_USEC *
 837                                   nla_get_u32(tb[TCA_FQ_CE_THRESHOLD]);
 838
 839         if (!err) {
 840                 sch_tree_unlock(sch);
 841                 err = fq_resize(sch, fq_log);
 842                 sch_tree_lock(sch);
 843         }
 844         while (sch->q.qlen > sch->limit) {
 845                 struct sk_buff *skb = fq_dequeue(sch);
 846
 847                 if (!skb)
 848                         break;
 849                 drop_len += qdisc_pkt_len(skb);
 850                 rtnl_kfree_skbs(skb, skb);
 851                 drop_count++;
 852         }
 853         qdisc_tree_reduce_backlog(sch, drop_count, drop_len);
 854
 855         sch_tree_unlock(sch);
 856         return err;
 857 }
 858
 859 static void fq_destroy(struct Qdisc *sch)
 860 {
 861         struct fq_sched_data *q = qdisc_priv(sch);
 862
 863         fq_reset(sch);
 864         fq_free(q->fq_root);
 865         qdisc_watchdog_cancel(&q->watchdog);
 866 }
 867
 868 static int fq_init(struct Qdisc *sch, struct nlattr *opt,
 869                    struct netlink_ext_ack *extack)
 870 {
 871         struct fq_sched_data *q = qdisc_priv(sch);
 872         int err;
 873
 874         sch->limit              = 10000;
 875         q->flow_plimit          = 100;
 876         q->quantum              = 2 * psched_mtu(qdisc_dev(sch));
 877         q->initial_quantum      = 10 * psched_mtu(qdisc_dev(sch));
 878         q->flow_refill_delay    = msecs_to_jiffies(40);
 879         q->flow_max_rate        = ~0UL;
 880         q->time_next_delayed_flow = ~0ULL;
 881         q->rate_enable          = 1;
 882         q->new_flows.first      = NULL;
 883         q->old_flows.first      = NULL;
 884         q->delayed              = RB_ROOT;
 885         q->fq_root              = NULL;
 886         q->fq_trees_log         = ilog2(1024);
 887         q->orphan_mask          = 1024 - 1;
 888         q->low_rate_threshold   = 550000 / 8;
 889
 890         /* Default ce_threshold of 4294 seconds */
 891         q->ce_threshold         = (u64)NSEC_PER_USEC * ~0U;
 892
 893         qdisc_watchdog_init_clockid(&q->watchdog, sch, CLOCK_MONOTONIC);
 894
 895         if (opt)
 896                 err = fq_change(sch, opt, extack);
 897         else
 898                 err = fq_resize(sch, q->fq_trees_log);
 899
 900         return err;
 901 }
 902
 903 static int fq_dump(struct Qdisc *sch, struct sk_buff *skb)
 904 {
 905         struct fq_sched_data *q = qdisc_priv(sch);
 906         u64 ce_threshold = q->ce_threshold;
 907         struct nlattr *opts;
 908
 909         opts = nla_nest_start_noflag(skb, TCA_OPTIONS);
 910         if (opts == NULL)
 911                 goto nla_put_failure;
 912
 913         /* TCA_FQ_FLOW_DEFAULT_RATE is not used anymore */
 914
 915         do_div(ce_threshold, NSEC_PER_USEC);
 916
 917         if (nla_put_u32(skb, TCA_FQ_PLIMIT, sch->limit) ||
 918             nla_put_u32(skb, TCA_FQ_FLOW_PLIMIT, q->flow_plimit) ||
 919             nla_put_u32(skb, TCA_FQ_QUANTUM, q->quantum) ||
 920             nla_put_u32(skb, TCA_FQ_INITIAL_QUANTUM, q->initial_quantum) ||
 921             nla_put_u32(skb, TCA_FQ_RATE_ENABLE, q->rate_enable) ||
 922             nla_put_u32(skb, TCA_FQ_FLOW_MAX_RATE,
 923                         min_t(unsigned long, q->flow_max_rate, ~0U)) ||
 924             nla_put_u32(skb, TCA_FQ_FLOW_REFILL_DELAY,
 925                         jiffies_to_usecs(q->flow_refill_delay)) ||
 926             nla_put_u32(skb, TCA_FQ_ORPHAN_MASK, q->orphan_mask) ||
 927             nla_put_u32(skb, TCA_FQ_LOW_RATE_THRESHOLD,
 928                         q->low_rate_threshold) ||
 929             nla_put_u32(skb, TCA_FQ_CE_THRESHOLD, (u32)ce_threshold) ||
 930             nla_put_u32(skb, TCA_FQ_BUCKETS_LOG, q->fq_trees_log))
 931                 goto nla_put_failure;
 932
 933         return nla_nest_end(skb, opts);
 934
 935 nla_put_failure:
 936         return -1;
 937 }
 938
 939 static int fq_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
 940 {
 941         struct fq_sched_data *q = qdisc_priv(sch);
 942         struct tc_fq_qd_stats st;
 943
 944         sch_tree_lock(sch);
 945
 946         st.gc_flows               = q->stat_gc_flows;
 947         st.highprio_packets       = q->stat_internal_packets;
 948         st.tcp_retrans            = 0;
 949         st.throttled              = q->stat_throttled;
 950         st.flows_plimit           = q->stat_flows_plimit;
 951         st.pkts_too_long          = q->stat_pkts_too_long;
 952         st.allocation_errors      = q->stat_allocation_errors;
 953         st.time_next_delayed_flow = q->time_next_delayed_flow - ktime_get_ns();
 954         st.flows                  = q->flows;
 955         st.inactive_flows         = q->inactive_flows;
 956         st.throttled_flows        = q->throttled_flows;
 957         st.unthrottle_latency_ns  = min_t(unsigned long,
 958                                           q->unthrottle_latency_ns, ~0U);
 959         st.ce_mark                = q->stat_ce_mark;
 960         sch_tree_unlock(sch);
 961
 962         return gnet_stats_copy_app(d, &st, sizeof(st));
 963 }
 964
 965 static struct Qdisc_ops fq_qdisc_ops __read_mostly = {
 966         .id             =       "fq",
 967         .priv_size      =       sizeof(struct fq_sched_data),
 968
 969         .enqueue        =       fq_enqueue,
 970         .dequeue        =       fq_dequeue,
 971         .peek           =       qdisc_peek_dequeued,
 972         .init           =       fq_init,
 973         .reset          =       fq_reset,
 974         .destroy        =       fq_destroy,
 975         .change         =       fq_change,
 976         .dump           =       fq_dump,
 977         .dump_stats     =       fq_dump_stats,
 978         .owner          =       THIS_MODULE,
 979 };
 980
 981 static int __init fq_module_init(void)
 982 {
 983         int ret;
 984
 985         fq_flow_cachep = kmem_cache_create("fq_flow_cache",
 986                                            sizeof(struct fq_flow),
 987                                            0, 0, NULL);
 988         if (!fq_flow_cachep)
 989                 return -ENOMEM;
 990
 991         ret = register_qdisc(&fq_qdisc_ops);
 992         if (ret)
 993                 kmem_cache_destroy(fq_flow_cachep);
 994         return ret;
 995 }
 996
 997 static void __exit fq_module_exit(void)
 998 {
 999         unregister_qdisc(&fq_qdisc_ops);
1000         kmem_cache_destroy(fq_flow_cachep);
1001 }
1002
1003 module_init(fq_module_init)
1004 module_exit(fq_module_exit)
1005 MODULE_AUTHOR("Eric Dumazet");
1006 MODULE_LICENSE("GPL");