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) 2013-2015 Eric Dumazet <edumazet@google.com>
   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 };
  56
  57 static inline struct fq_skb_cb *fq_skb_cb(struct sk_buff *skb)
  58 {
  59         qdisc_cb_private_validate(skb, sizeof(struct fq_skb_cb));
  60         return (struct fq_skb_cb *)qdisc_skb_cb(skb)->data;
  61 }
  62
  63 /*
  64  * Per flow structure, dynamically allocated.
  65  * If packets have monotically increasing time_to_send, they are placed in O(1)
  66  * in linear list (head,tail), otherwise are placed in a rbtree (t_root).
  67  */
  68 struct fq_flow {
  69         struct rb_root  t_root;
  70         struct sk_buff  *head;          /* list of skbs for this flow : first skb */
  71         union {
  72                 struct sk_buff *tail;   /* last skb in the list */
  73                 unsigned long  age;     /* jiffies when flow was emptied, for gc */
  74         };
  75         struct rb_node  fq_node;        /* anchor in fq_root[] trees */
  76         struct sock     *sk;
  77         int             qlen;           /* number of packets in flow queue */
  78         int             credit;
  79         u32             socket_hash;    /* sk_hash */
  80         struct fq_flow *next;           /* next pointer in RR lists, or &detached */
  81
  82         struct rb_node  rate_node;      /* anchor in q->delayed tree */
  83         u64             time_next_packet;
  84 };
  85
  86 struct fq_flow_head {
  87         struct fq_flow *first;
  88         struct fq_flow *last;
  89 };
  90
  91 struct fq_sched_data {
  92         struct fq_flow_head new_flows;
  93
  94         struct fq_flow_head old_flows;
  95
  96         struct rb_root  delayed;        /* for rate limited flows */
  97         u64             time_next_delayed_flow;
  98         unsigned long   unthrottle_latency_ns;
  99
 100         struct fq_flow  internal;       /* for non classified or high prio packets */
 101         u32             quantum;
 102         u32             initial_quantum;
 103         u32             flow_refill_delay;
 104         u32             flow_plimit;    /* max packets per flow */
 105         unsigned long   flow_max_rate;  /* optional max rate per flow */
 106         u64             ce_threshold;
 107         u32             orphan_mask;    /* mask for orphaned skb */
 108         u32             low_rate_threshold;
 109         struct rb_root  *fq_root;
 110         u8              rate_enable;
 111         u8              fq_trees_log;
 112
 113         u32             flows;
 114         u32             inactive_flows;
 115         u32             throttled_flows;
 116
 117         u64             stat_gc_flows;
 118         u64             stat_internal_packets;
 119         u64             stat_throttled;
 120         u64             stat_ce_mark;
 121         u64             stat_flows_plimit;
 122         u64             stat_pkts_too_long;
 123         u64             stat_allocation_errors;
 124         struct qdisc_watchdog watchdog;
 125 };
 126
 127 /* special value to mark a detached flow (not on old/new list) */
 128 static struct fq_flow detached, throttled;
 129
 130 static void fq_flow_set_detached(struct fq_flow *f)
 131 {
 132         f->next = &detached;
 133         f->age = jiffies;
 134 }
 135
 136 static bool fq_flow_is_detached(const struct fq_flow *f)
 137 {
 138         return f->next == &detached;
 139 }
 140
 141 static bool fq_flow_is_throttled(const struct fq_flow *f)
 142 {
 143         return f->next == &throttled;
 144 }
 145
 146 static void fq_flow_add_tail(struct fq_flow_head *head, struct fq_flow *flow)
 147 {
 148         if (head->first)
 149                 head->last->next = flow;
 150         else
 151                 head->first = flow;
 152         head->last = flow;
 153         flow->next = NULL;
 154 }
 155
 156 static void fq_flow_unset_throttled(struct fq_sched_data *q, struct fq_flow *f)
 157 {
 158         rb_erase(&f->rate_node, &q->delayed);
 159         q->throttled_flows--;
 160         fq_flow_add_tail(&q->old_flows, f);
 161 }
 162
 163 static void fq_flow_set_throttled(struct fq_sched_data *q, struct fq_flow *f)
 164 {
 165         struct rb_node **p = &q->delayed.rb_node, *parent = NULL;
 166
 167         while (*p) {
 168                 struct fq_flow *aux;
 169
 170                 parent = *p;
 171                 aux = rb_entry(parent, struct fq_flow, rate_node);
 172                 if (f->time_next_packet >= aux->time_next_packet)
 173                         p = &parent->rb_right;
 174                 else
 175                         p = &parent->rb_left;
 176         }
 177         rb_link_node(&f->rate_node, parent, p);
 178         rb_insert_color(&f->rate_node, &q->delayed);
 179         q->throttled_flows++;
 180         q->stat_throttled++;
 181
 182         f->next = &throttled;
 183         if (q->time_next_delayed_flow > f->time_next_packet)
 184                 q->time_next_delayed_flow = f->time_next_packet;
 185 }
 186
 187
 188 static struct kmem_cache *fq_flow_cachep __read_mostly;
 189
 190
 191 /* limit number of collected flows per round */
 192 #define FQ_GC_MAX 8
 193 #define FQ_GC_AGE (3*HZ)
 194
 195 static bool fq_gc_candidate(const struct fq_flow *f)
 196 {
 197         return fq_flow_is_detached(f) &&
 198                time_after(jiffies, f->age + FQ_GC_AGE);
 199 }
 200
 201 static void fq_gc(struct fq_sched_data *q,
 202                   struct rb_root *root,
 203                   struct sock *sk)
 204 {
 205         struct fq_flow *f, *tofree[FQ_GC_MAX];
 206         struct rb_node **p, *parent;
 207         int fcnt = 0;
 208
 209         p = &root->rb_node;
 210         parent = NULL;
 211         while (*p) {
 212                 parent = *p;
 213
 214                 f = rb_entry(parent, struct fq_flow, fq_node);
 215                 if (f->sk == sk)
 216                         break;
 217
 218                 if (fq_gc_candidate(f)) {
 219                         tofree[fcnt++] = f;
 220                         if (fcnt == FQ_GC_MAX)
 221                                 break;
 222                 }
 223
 224                 if (f->sk > sk)
 225                         p = &parent->rb_right;
 226                 else
 227                         p = &parent->rb_left;
 228         }
 229
 230         q->flows -= fcnt;
 231         q->inactive_flows -= fcnt;
 232         q->stat_gc_flows += fcnt;
 233         while (fcnt) {
 234                 struct fq_flow *f = tofree[--fcnt];
 235
 236                 rb_erase(&f->fq_node, root);
 237                 kmem_cache_free(fq_flow_cachep, f);
 238         }
 239 }
 240
 241 static struct fq_flow *fq_classify(struct sk_buff *skb, struct fq_sched_data *q)
 242 {
 243         struct rb_node **p, *parent;
 244         struct sock *sk = skb->sk;
 245         struct rb_root *root;
 246         struct fq_flow *f;
 247
 248         /* warning: no starvation prevention... */
 249         if (unlikely((skb->priority & TC_PRIO_MAX) == TC_PRIO_CONTROL))
 250                 return &q->internal;
 251
 252         /* SYNACK messages are attached to a TCP_NEW_SYN_RECV request socket
 253          * or a listener (SYNCOOKIE mode)
 254          * 1) request sockets are not full blown,
 255          *    they do not contain sk_pacing_rate
 256          * 2) They are not part of a 'flow' yet
 257          * 3) We do not want to rate limit them (eg SYNFLOOD attack),
 258          *    especially if the listener set SO_MAX_PACING_RATE
 259          * 4) We pretend they are orphaned
 260          */
 261         if (!sk || sk_listener(sk)) {
 262                 unsigned long hash = skb_get_hash(skb) & q->orphan_mask;
 263
 264                 /* By forcing low order bit to 1, we make sure to not
 265                  * collide with a local flow (socket pointers are word aligned)
 266                  */
 267                 sk = (struct sock *)((hash << 1) | 1UL);
 268                 skb_orphan(skb);
 269         } else if (sk->sk_state == TCP_CLOSE) {
 270                 unsigned long hash = skb_get_hash(skb) & q->orphan_mask;
 271                 /*
 272                  * Sockets in TCP_CLOSE are non connected.
 273                  * Typical use case is UDP sockets, they can send packets
 274                  * with sendto() to many different destinations.
 275                  * We probably could use a generic bit advertising
 276                  * non connected sockets, instead of sk_state == TCP_CLOSE,
 277                  * if we care enough.
 278                  */
 279                 sk = (struct sock *)((hash << 1) | 1UL);
 280         }
 281
 282         root = &q->fq_root[hash_ptr(sk, q->fq_trees_log)];
 283
 284         if (q->flows >= (2U << q->fq_trees_log) &&
 285             q->inactive_flows > q->flows/2)
 286                 fq_gc(q, root, sk);
 287
 288         p = &root->rb_node;
 289         parent = NULL;
 290         while (*p) {
 291                 parent = *p;
 292
 293                 f = rb_entry(parent, struct fq_flow, fq_node);
 294                 if (f->sk == sk) {
 295                         /* socket might have been reallocated, so check
 296                          * if its sk_hash is the same.
 297                          * It not, we need to refill credit with
 298                          * initial quantum
 299                          */
 300                         if (unlikely(skb->sk == sk &&
 301                                      f->socket_hash != sk->sk_hash)) {
 302                                 f->credit = q->initial_quantum;
 303                                 f->socket_hash = sk->sk_hash;
 304                                 if (q->rate_enable)
 305                                         smp_store_release(&sk->sk_pacing_status,
 306                                                           SK_PACING_FQ);
 307                                 if (fq_flow_is_throttled(f))
 308                                         fq_flow_unset_throttled(q, f);
 309                                 f->time_next_packet = 0ULL;
 310                         }
 311                         return f;
 312                 }
 313                 if (f->sk > sk)
 314                         p = &parent->rb_right;
 315                 else
 316                         p = &parent->rb_left;
 317         }
 318
 319         f = kmem_cache_zalloc(fq_flow_cachep, GFP_ATOMIC | __GFP_NOWARN);
 320         if (unlikely(!f)) {
 321                 q->stat_allocation_errors++;
 322                 return &q->internal;
 323         }
 324         /* f->t_root is already zeroed after kmem_cache_zalloc() */
 325
 326         fq_flow_set_detached(f);
 327         f->sk = sk;
 328         if (skb->sk == sk) {
 329                 f->socket_hash = sk->sk_hash;
 330                 if (q->rate_enable)
 331                         smp_store_release(&sk->sk_pacing_status,
 332                                           SK_PACING_FQ);
 333         }
 334         f->credit = q->initial_quantum;
 335
 336         rb_link_node(&f->fq_node, parent, p);
 337         rb_insert_color(&f->fq_node, root);
 338
 339         q->flows++;
 340         q->inactive_flows++;
 341         return f;
 342 }
 343
 344 static struct sk_buff *fq_peek(struct fq_flow *flow)
 345 {
 346         struct sk_buff *skb = skb_rb_first(&flow->t_root);
 347         struct sk_buff *head = flow->head;
 348
 349         if (!skb)
 350                 return head;
 351
 352         if (!head)
 353                 return skb;
 354
 355         if (fq_skb_cb(skb)->time_to_send < fq_skb_cb(head)->time_to_send)
 356                 return skb;
 357         return head;
 358 }
 359
 360 static void fq_erase_head(struct Qdisc *sch, struct fq_flow *flow,
 361                           struct sk_buff *skb)
 362 {
 363         if (skb == flow->head) {
 364                 flow->head = skb->next;
 365         } else {
 366                 rb_erase(&skb->rbnode, &flow->t_root);
 367                 skb->dev = qdisc_dev(sch);
 368         }
 369 }
 370
 371 /* remove one skb from head of flow queue */
 372 static struct sk_buff *fq_dequeue_head(struct Qdisc *sch, struct fq_flow *flow)
 373 {
 374         struct sk_buff *skb = fq_peek(flow);
 375
 376         if (skb) {
 377                 fq_erase_head(sch, flow, skb);
 378                 skb_mark_not_on_list(skb);
 379                 flow->qlen--;
 380                 qdisc_qstats_backlog_dec(sch, skb);
 381                 sch->q.qlen--;
 382         }
 383         return skb;
 384 }
 385
 386 static void flow_queue_add(struct fq_flow *flow, struct sk_buff *skb)
 387 {
 388         struct rb_node **p, *parent;
 389         struct sk_buff *head, *aux;
 390
 391         fq_skb_cb(skb)->time_to_send = skb->tstamp ?: ktime_get_ns();
 392
 393         head = flow->head;
 394         if (!head ||
 395             fq_skb_cb(skb)->time_to_send >= fq_skb_cb(flow->tail)->time_to_send) {
 396                 if (!head)
 397                         flow->head = skb;
 398                 else
 399                         flow->tail->next = skb;
 400                 flow->tail = skb;
 401                 skb->next = NULL;
 402                 return;
 403         }
 404
 405         p = &flow->t_root.rb_node;
 406         parent = NULL;
 407
 408         while (*p) {
 409                 parent = *p;
 410                 aux = rb_to_skb(parent);
 411                 if (fq_skb_cb(skb)->time_to_send >= fq_skb_cb(aux)->time_to_send)
 412                         p = &parent->rb_right;
 413                 else
 414                         p = &parent->rb_left;
 415         }
 416         rb_link_node(&skb->rbnode, parent, p);
 417         rb_insert_color(&skb->rbnode, &flow->t_root);
 418 }
 419
 420 static int fq_enqueue(struct sk_buff *skb, struct Qdisc *sch,
 421                       struct sk_buff **to_free)
 422 {
 423         struct fq_sched_data *q = qdisc_priv(sch);
 424         struct fq_flow *f;
 425
 426         if (unlikely(sch->q.qlen >= sch->limit))
 427                 return qdisc_drop(skb, sch, to_free);
 428
 429         f = fq_classify(skb, q);
 430         if (unlikely(f->qlen >= q->flow_plimit && f != &q->internal)) {
 431                 q->stat_flows_plimit++;
 432                 return qdisc_drop(skb, sch, to_free);
 433         }
 434
 435         f->qlen++;
 436         qdisc_qstats_backlog_inc(sch, skb);
 437         if (fq_flow_is_detached(f)) {
 438                 fq_flow_add_tail(&q->new_flows, f);
 439                 if (time_after(jiffies, f->age + q->flow_refill_delay))
 440                         f->credit = max_t(u32, f->credit, q->quantum);
 441                 q->inactive_flows--;
 442         }
 443
 444         /* Note: this overwrites f->age */
 445         flow_queue_add(f, skb);
 446
 447         if (unlikely(f == &q->internal)) {
 448                 q->stat_internal_packets++;
 449         }
 450         sch->q.qlen++;
 451
 452         return NET_XMIT_SUCCESS;
 453 }
 454
 455 static void fq_check_throttled(struct fq_sched_data *q, u64 now)
 456 {
 457         unsigned long sample;
 458         struct rb_node *p;
 459
 460         if (q->time_next_delayed_flow > now)
 461                 return;
 462
 463         /* Update unthrottle latency EWMA.
 464          * This is cheap and can help diagnosing timer/latency problems.
 465          */
 466         sample = (unsigned long)(now - q->time_next_delayed_flow);
 467         q->unthrottle_latency_ns -= q->unthrottle_latency_ns >> 3;
 468         q->unthrottle_latency_ns += sample >> 3;
 469
 470         q->time_next_delayed_flow = ~0ULL;
 471         while ((p = rb_first(&q->delayed)) != NULL) {
 472                 struct fq_flow *f = rb_entry(p, struct fq_flow, rate_node);
 473
 474                 if (f->time_next_packet > now) {
 475                         q->time_next_delayed_flow = f->time_next_packet;
 476                         break;
 477                 }
 478                 fq_flow_unset_throttled(q, f);
 479         }
 480 }
 481
 482 static struct sk_buff *fq_dequeue(struct Qdisc *sch)
 483 {
 484         struct fq_sched_data *q = qdisc_priv(sch);
 485         struct fq_flow_head *head;
 486         struct sk_buff *skb;
 487         struct fq_flow *f;
 488         unsigned long rate;
 489         u32 plen;
 490         u64 now;
 491
 492         if (!sch->q.qlen)
 493                 return NULL;
 494
 495         skb = fq_dequeue_head(sch, &q->internal);
 496         if (skb)
 497                 goto out;
 498
 499         now = ktime_get_ns();
 500         fq_check_throttled(q, now);
 501 begin:
 502         head = &q->new_flows;
 503         if (!head->first) {
 504                 head = &q->old_flows;
 505                 if (!head->first) {
 506                         if (q->time_next_delayed_flow != ~0ULL)
 507                                 qdisc_watchdog_schedule_ns(&q->watchdog,
 508                                                            q->time_next_delayed_flow);
 509                         return NULL;
 510                 }
 511         }
 512         f = head->first;
 513
 514         if (f->credit <= 0) {
 515                 f->credit += q->quantum;
 516                 head->first = f->next;
 517                 fq_flow_add_tail(&q->old_flows, f);
 518                 goto begin;
 519         }
 520
 521         skb = fq_peek(f);
 522         if (skb) {
 523                 u64 time_next_packet = max_t(u64, fq_skb_cb(skb)->time_to_send,
 524                                              f->time_next_packet);
 525
 526                 if (now < time_next_packet) {
 527                         head->first = f->next;
 528                         f->time_next_packet = time_next_packet;
 529                         fq_flow_set_throttled(q, f);
 530                         goto begin;
 531                 }
 532                 if ((s64)(now - time_next_packet - q->ce_threshold) > 0) {
 533                         INET_ECN_set_ce(skb);
 534                         q->stat_ce_mark++;
 535                 }
 536         }
 537
 538         skb = fq_dequeue_head(sch, f);
 539         if (!skb) {
 540                 head->first = f->next;
 541                 /* force a pass through old_flows to prevent starvation */
 542                 if ((head == &q->new_flows) && q->old_flows.first) {
 543                         fq_flow_add_tail(&q->old_flows, f);
 544                 } else {
 545                         fq_flow_set_detached(f);
 546                         q->inactive_flows++;
 547                 }
 548                 goto begin;
 549         }
 550         prefetch(&skb->end);
 551         plen = qdisc_pkt_len(skb);
 552         f->credit -= plen;
 553
 554         if (!q->rate_enable)
 555                 goto out;
 556
 557         rate = q->flow_max_rate;
 558
 559         /* If EDT time was provided for this skb, we need to
 560          * update f->time_next_packet only if this qdisc enforces
 561          * a flow max rate.
 562          */
 563         if (!skb->tstamp) {
 564                 if (skb->sk)
 565                         rate = min(skb->sk->sk_pacing_rate, rate);
 566
 567                 if (rate <= q->low_rate_threshold) {
 568                         f->credit = 0;
 569                 } else {
 570                         plen = max(plen, q->quantum);
 571                         if (f->credit > 0)
 572                                 goto out;
 573                 }
 574         }
 575         if (rate != ~0UL) {
 576                 u64 len = (u64)plen * NSEC_PER_SEC;
 577
 578                 if (likely(rate))
 579                         len = div64_ul(len, rate);
 580                 /* Since socket rate can change later,
 581                  * clamp the delay to 1 second.
 582                  * Really, providers of too big packets should be fixed !
 583                  */
 584                 if (unlikely(len > NSEC_PER_SEC)) {
 585                         len = NSEC_PER_SEC;
 586                         q->stat_pkts_too_long++;
 587                 }
 588                 /* Account for schedule/timers drifts.
 589                  * f->time_next_packet was set when prior packet was sent,
 590                  * and current time (@now) can be too late by tens of us.
 591                  */
 592                 if (f->time_next_packet)
 593                         len -= min(len/2, now - f->time_next_packet);
 594                 f->time_next_packet = now + len;
 595         }
 596 out:
 597         qdisc_bstats_update(sch, skb);
 598         return skb;
 599 }
 600
 601 static void fq_flow_purge(struct fq_flow *flow)
 602 {
 603         struct rb_node *p = rb_first(&flow->t_root);
 604
 605         while (p) {
 606                 struct sk_buff *skb = rb_to_skb(p);
 607
 608                 p = rb_next(p);
 609                 rb_erase(&skb->rbnode, &flow->t_root);
 610                 rtnl_kfree_skbs(skb, skb);
 611         }
 612         rtnl_kfree_skbs(flow->head, flow->tail);
 613         flow->head = NULL;
 614         flow->qlen = 0;
 615 }
 616
 617 static void fq_reset(struct Qdisc *sch)
 618 {
 619         struct fq_sched_data *q = qdisc_priv(sch);
 620         struct rb_root *root;
 621         struct rb_node *p;
 622         struct fq_flow *f;
 623         unsigned int idx;
 624
 625         sch->q.qlen = 0;
 626         sch->qstats.backlog = 0;
 627
 628         fq_flow_purge(&q->internal);
 629
 630         if (!q->fq_root)
 631                 return;
 632
 633         for (idx = 0; idx < (1U << q->fq_trees_log); idx++) {
 634                 root = &q->fq_root[idx];
 635                 while ((p = rb_first(root)) != NULL) {
 636                         f = rb_entry(p, struct fq_flow, fq_node);
 637                         rb_erase(p, root);
 638
 639                         fq_flow_purge(f);
 640
 641                         kmem_cache_free(fq_flow_cachep, f);
 642                 }
 643         }
 644         q->new_flows.first      = NULL;
 645         q->old_flows.first      = NULL;
 646         q->delayed              = RB_ROOT;
 647         q->flows                = 0;
 648         q->inactive_flows       = 0;
 649         q->throttled_flows      = 0;
 650 }
 651
 652 static void fq_rehash(struct fq_sched_data *q,
 653                       struct rb_root *old_array, u32 old_log,
 654                       struct rb_root *new_array, u32 new_log)
 655 {
 656         struct rb_node *op, **np, *parent;
 657         struct rb_root *oroot, *nroot;
 658         struct fq_flow *of, *nf;
 659         int fcnt = 0;
 660         u32 idx;
 661
 662         for (idx = 0; idx < (1U << old_log); idx++) {
 663                 oroot = &old_array[idx];
 664                 while ((op = rb_first(oroot)) != NULL) {
 665                         rb_erase(op, oroot);
 666                         of = rb_entry(op, struct fq_flow, fq_node);
 667                         if (fq_gc_candidate(of)) {
 668                                 fcnt++;
 669                                 kmem_cache_free(fq_flow_cachep, of);
 670                                 continue;
 671                         }
 672                         nroot = &new_array[hash_ptr(of->sk, new_log)];
 673
 674                         np = &nroot->rb_node;
 675                         parent = NULL;
 676                         while (*np) {
 677                                 parent = *np;
 678
 679                                 nf = rb_entry(parent, struct fq_flow, fq_node);
 680                                 BUG_ON(nf->sk == of->sk);
 681
 682                                 if (nf->sk > of->sk)
 683                                         np = &parent->rb_right;
 684                                 else
 685                                         np = &parent->rb_left;
 686                         }
 687
 688                         rb_link_node(&of->fq_node, parent, np);
 689                         rb_insert_color(&of->fq_node, nroot);
 690                 }
 691         }
 692         q->flows -= fcnt;
 693         q->inactive_flows -= fcnt;
 694         q->stat_gc_flows += fcnt;
 695 }
 696
 697 static void fq_free(void *addr)
 698 {
 699         kvfree(addr);
 700 }
 701
 702 static int fq_resize(struct Qdisc *sch, u32 log)
 703 {
 704         struct fq_sched_data *q = qdisc_priv(sch);
 705         struct rb_root *array;
 706         void *old_fq_root;
 707         u32 idx;
 708
 709         if (q->fq_root && log == q->fq_trees_log)
 710                 return 0;
 711
 712         /* If XPS was setup, we can allocate memory on right NUMA node */
 713         array = kvmalloc_node(sizeof(struct rb_root) << log, GFP_KERNEL | __GFP_RETRY_MAYFAIL,
 714                               netdev_queue_numa_node_read(sch->dev_queue));
 715         if (!array)
 716                 return -ENOMEM;
 717
 718         for (idx = 0; idx < (1U << log); idx++)
 719                 array[idx] = RB_ROOT;
 720
 721         sch_tree_lock(sch);
 722
 723         old_fq_root = q->fq_root;
 724         if (old_fq_root)
 725                 fq_rehash(q, old_fq_root, q->fq_trees_log, array, log);
 726
 727         q->fq_root = array;
 728         q->fq_trees_log = log;
 729
 730         sch_tree_unlock(sch);
 731
 732         fq_free(old_fq_root);
 733
 734         return 0;
 735 }
 736
 737 static const struct nla_policy fq_policy[TCA_FQ_MAX + 1] = {
 738         [TCA_FQ_PLIMIT]                 = { .type = NLA_U32 },
 739         [TCA_FQ_FLOW_PLIMIT]            = { .type = NLA_U32 },
 740         [TCA_FQ_QUANTUM]                = { .type = NLA_U32 },
 741         [TCA_FQ_INITIAL_QUANTUM]        = { .type = NLA_U32 },
 742         [TCA_FQ_RATE_ENABLE]            = { .type = NLA_U32 },
 743         [TCA_FQ_FLOW_DEFAULT_RATE]      = { .type = NLA_U32 },
 744         [TCA_FQ_FLOW_MAX_RATE]          = { .type = NLA_U32 },
 745         [TCA_FQ_BUCKETS_LOG]            = { .type = NLA_U32 },
 746         [TCA_FQ_FLOW_REFILL_DELAY]      = { .type = NLA_U32 },
 747         [TCA_FQ_LOW_RATE_THRESHOLD]     = { .type = NLA_U32 },
 748         [TCA_FQ_CE_THRESHOLD]           = { .type = NLA_U32 },
 749 };
 750
 751 static int fq_change(struct Qdisc *sch, struct nlattr *opt,
 752                      struct netlink_ext_ack *extack)
 753 {
 754         struct fq_sched_data *q = qdisc_priv(sch);
 755         struct nlattr *tb[TCA_FQ_MAX + 1];
 756         int err, drop_count = 0;
 757         unsigned drop_len = 0;
 758         u32 fq_log;
 759
 760         if (!opt)
 761                 return -EINVAL;
 762
 763         err = nla_parse_nested_deprecated(tb, TCA_FQ_MAX, opt, fq_policy,
 764                                           NULL);
 765         if (err < 0)
 766                 return err;
 767
 768         sch_tree_lock(sch);
 769
 770         fq_log = q->fq_trees_log;
 771
 772         if (tb[TCA_FQ_BUCKETS_LOG]) {
 773                 u32 nval = nla_get_u32(tb[TCA_FQ_BUCKETS_LOG]);
 774
 775                 if (nval >= 1 && nval <= ilog2(256*1024))
 776                         fq_log = nval;
 777                 else
 778                         err = -EINVAL;
 779         }
 780         if (tb[TCA_FQ_PLIMIT])
 781                 sch->limit = nla_get_u32(tb[TCA_FQ_PLIMIT]);
 782
 783         if (tb[TCA_FQ_FLOW_PLIMIT])
 784                 q->flow_plimit = nla_get_u32(tb[TCA_FQ_FLOW_PLIMIT]);
 785
 786         if (tb[TCA_FQ_QUANTUM]) {
 787                 u32 quantum = nla_get_u32(tb[TCA_FQ_QUANTUM]);
 788
 789                 if (quantum > 0 && quantum <= (1 << 20)) {
 790                         q->quantum = quantum;
 791                 } else {
 792                         NL_SET_ERR_MSG_MOD(extack, "invalid quantum");
 793                         err = -EINVAL;
 794                 }
 795         }
 796
 797         if (tb[TCA_FQ_INITIAL_QUANTUM])
 798                 q->initial_quantum = nla_get_u32(tb[TCA_FQ_INITIAL_QUANTUM]);
 799
 800         if (tb[TCA_FQ_FLOW_DEFAULT_RATE])
 801                 pr_warn_ratelimited("sch_fq: defrate %u ignored.\n",
 802                                     nla_get_u32(tb[TCA_FQ_FLOW_DEFAULT_RATE]));
 803
 804         if (tb[TCA_FQ_FLOW_MAX_RATE]) {
 805                 u32 rate = nla_get_u32(tb[TCA_FQ_FLOW_MAX_RATE]);
 806
 807                 q->flow_max_rate = (rate == ~0U) ? ~0UL : rate;
 808         }
 809         if (tb[TCA_FQ_LOW_RATE_THRESHOLD])
 810                 q->low_rate_threshold =
 811                         nla_get_u32(tb[TCA_FQ_LOW_RATE_THRESHOLD]);
 812
 813         if (tb[TCA_FQ_RATE_ENABLE]) {
 814                 u32 enable = nla_get_u32(tb[TCA_FQ_RATE_ENABLE]);
 815
 816                 if (enable <= 1)
 817                         q->rate_enable = enable;
 818                 else
 819                         err = -EINVAL;
 820         }
 821
 822         if (tb[TCA_FQ_FLOW_REFILL_DELAY]) {
 823                 u32 usecs_delay = nla_get_u32(tb[TCA_FQ_FLOW_REFILL_DELAY]) ;
 824
 825                 q->flow_refill_delay = usecs_to_jiffies(usecs_delay);
 826         }
 827
 828         if (tb[TCA_FQ_ORPHAN_MASK])
 829                 q->orphan_mask = nla_get_u32(tb[TCA_FQ_ORPHAN_MASK]);
 830
 831         if (tb[TCA_FQ_CE_THRESHOLD])
 832                 q->ce_threshold = (u64)NSEC_PER_USEC *
 833                                   nla_get_u32(tb[TCA_FQ_CE_THRESHOLD]);
 834
 835         if (!err) {
 836                 sch_tree_unlock(sch);
 837                 err = fq_resize(sch, fq_log);
 838                 sch_tree_lock(sch);
 839         }
 840         while (sch->q.qlen > sch->limit) {
 841                 struct sk_buff *skb = fq_dequeue(sch);
 842
 843                 if (!skb)
 844                         break;
 845                 drop_len += qdisc_pkt_len(skb);
 846                 rtnl_kfree_skbs(skb, skb);
 847                 drop_count++;
 848         }
 849         qdisc_tree_reduce_backlog(sch, drop_count, drop_len);
 850
 851         sch_tree_unlock(sch);
 852         return err;
 853 }
 854
 855 static void fq_destroy(struct Qdisc *sch)
 856 {
 857         struct fq_sched_data *q = qdisc_priv(sch);
 858
 859         fq_reset(sch);
 860         fq_free(q->fq_root);
 861         qdisc_watchdog_cancel(&q->watchdog);
 862 }
 863
 864 static int fq_init(struct Qdisc *sch, struct nlattr *opt,
 865                    struct netlink_ext_ack *extack)
 866 {
 867         struct fq_sched_data *q = qdisc_priv(sch);
 868         int err;
 869
 870         sch->limit              = 10000;
 871         q->flow_plimit          = 100;
 872         q->quantum              = 2 * psched_mtu(qdisc_dev(sch));
 873         q->initial_quantum      = 10 * psched_mtu(qdisc_dev(sch));
 874         q->flow_refill_delay    = msecs_to_jiffies(40);
 875         q->flow_max_rate        = ~0UL;
 876         q->time_next_delayed_flow = ~0ULL;
 877         q->rate_enable          = 1;
 878         q->new_flows.first      = NULL;
 879         q->old_flows.first      = NULL;
 880         q->delayed              = RB_ROOT;
 881         q->fq_root              = NULL;
 882         q->fq_trees_log         = ilog2(1024);
 883         q->orphan_mask          = 1024 - 1;
 884         q->low_rate_threshold   = 550000 / 8;
 885
 886         /* Default ce_threshold of 4294 seconds */
 887         q->ce_threshold         = (u64)NSEC_PER_USEC * ~0U;
 888
 889         qdisc_watchdog_init_clockid(&q->watchdog, sch, CLOCK_MONOTONIC);
 890
 891         if (opt)
 892                 err = fq_change(sch, opt, extack);
 893         else
 894                 err = fq_resize(sch, q->fq_trees_log);
 895
 896         return err;
 897 }
 898
 899 static int fq_dump(struct Qdisc *sch, struct sk_buff *skb)
 900 {
 901         struct fq_sched_data *q = qdisc_priv(sch);
 902         u64 ce_threshold = q->ce_threshold;
 903         struct nlattr *opts;
 904
 905         opts = nla_nest_start_noflag(skb, TCA_OPTIONS);
 906         if (opts == NULL)
 907                 goto nla_put_failure;
 908
 909         /* TCA_FQ_FLOW_DEFAULT_RATE is not used anymore */
 910
 911         do_div(ce_threshold, NSEC_PER_USEC);
 912
 913         if (nla_put_u32(skb, TCA_FQ_PLIMIT, sch->limit) ||
 914             nla_put_u32(skb, TCA_FQ_FLOW_PLIMIT, q->flow_plimit) ||
 915             nla_put_u32(skb, TCA_FQ_QUANTUM, q->quantum) ||
 916             nla_put_u32(skb, TCA_FQ_INITIAL_QUANTUM, q->initial_quantum) ||
 917             nla_put_u32(skb, TCA_FQ_RATE_ENABLE, q->rate_enable) ||
 918             nla_put_u32(skb, TCA_FQ_FLOW_MAX_RATE,
 919                         min_t(unsigned long, q->flow_max_rate, ~0U)) ||
 920             nla_put_u32(skb, TCA_FQ_FLOW_REFILL_DELAY,
 921                         jiffies_to_usecs(q->flow_refill_delay)) ||
 922             nla_put_u32(skb, TCA_FQ_ORPHAN_MASK, q->orphan_mask) ||
 923             nla_put_u32(skb, TCA_FQ_LOW_RATE_THRESHOLD,
 924                         q->low_rate_threshold) ||
 925             nla_put_u32(skb, TCA_FQ_CE_THRESHOLD, (u32)ce_threshold) ||
 926             nla_put_u32(skb, TCA_FQ_BUCKETS_LOG, q->fq_trees_log))
 927                 goto nla_put_failure;
 928
 929         return nla_nest_end(skb, opts);
 930
 931 nla_put_failure:
 932         return -1;
 933 }
 934
 935 static int fq_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
 936 {
 937         struct fq_sched_data *q = qdisc_priv(sch);
 938         struct tc_fq_qd_stats st;
 939
 940         sch_tree_lock(sch);
 941
 942         st.gc_flows               = q->stat_gc_flows;
 943         st.highprio_packets       = q->stat_internal_packets;
 944         st.tcp_retrans            = 0;
 945         st.throttled              = q->stat_throttled;
 946         st.flows_plimit           = q->stat_flows_plimit;
 947         st.pkts_too_long          = q->stat_pkts_too_long;
 948         st.allocation_errors      = q->stat_allocation_errors;
 949         st.time_next_delayed_flow = q->time_next_delayed_flow - ktime_get_ns();
 950         st.flows                  = q->flows;
 951         st.inactive_flows         = q->inactive_flows;
 952         st.throttled_flows        = q->throttled_flows;
 953         st.unthrottle_latency_ns  = min_t(unsigned long,
 954                                           q->unthrottle_latency_ns, ~0U);
 955         st.ce_mark                = q->stat_ce_mark;
 956         sch_tree_unlock(sch);
 957
 958         return gnet_stats_copy_app(d, &st, sizeof(st));
 959 }
 960
 961 static struct Qdisc_ops fq_qdisc_ops __read_mostly = {
 962         .id             =       "fq",
 963         .priv_size      =       sizeof(struct fq_sched_data),
 964
 965         .enqueue        =       fq_enqueue,
 966         .dequeue        =       fq_dequeue,
 967         .peek           =       qdisc_peek_dequeued,
 968         .init           =       fq_init,
 969         .reset          =       fq_reset,
 970         .destroy        =       fq_destroy,
 971         .change         =       fq_change,
 972         .dump           =       fq_dump,
 973         .dump_stats     =       fq_dump_stats,
 974         .owner          =       THIS_MODULE,
 975 };
 976
 977 static int __init fq_module_init(void)
 978 {
 979         int ret;
 980
 981         fq_flow_cachep = kmem_cache_create("fq_flow_cache",
 982                                            sizeof(struct fq_flow),
 983                                            0, 0, NULL);
 984         if (!fq_flow_cachep)
 985                 return -ENOMEM;
 986
 987         ret = register_qdisc(&fq_qdisc_ops);
 988         if (ret)
 989                 kmem_cache_destroy(fq_flow_cachep);
 990         return ret;
 991 }
 992
 993 static void __exit fq_module_exit(void)
 994 {
 995         unregister_qdisc(&fq_qdisc_ops);
 996         kmem_cache_destroy(fq_flow_cachep);
 997 }
 998
 999 module_init(fq_module_init)
1000 module_exit(fq_module_exit)
1001 MODULE_AUTHOR("Eric Dumazet");
1002 MODULE_LICENSE("GPL");