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 /* First cache line : used in fq_gc(), fq_enqueue(), fq_dequeue() */
  70         struct rb_root  t_root;
  71         struct sk_buff  *head;          /* list of skbs for this flow : first skb */
  72         union {
  73                 struct sk_buff *tail;   /* last skb in the list */
  74                 unsigned long  age;     /* (jiffies | 1UL) when flow was emptied, for gc */
  75         };
  76         struct rb_node  fq_node;        /* anchor in fq_root[] trees */
  77         struct sock     *sk;
  78         u32             socket_hash;    /* sk_hash */
  79         int             qlen;           /* number of packets in flow queue */
  80
  81 /* Second cache line, used in fq_dequeue() */
  82         int             credit;
  83         /* 32bit hole on 64bit arches */
  84
  85         struct fq_flow *next;           /* next pointer in RR lists */
  86
  87         struct rb_node  rate_node;      /* anchor in q->delayed tree */
  88         u64             time_next_packet;
  89 } ____cacheline_aligned_in_smp;
  90
  91 struct fq_flow_head {
  92         struct fq_flow *first;
  93         struct fq_flow *last;
  94 };
  95
  96 struct fq_sched_data {
  97         struct fq_flow_head new_flows;
  98
  99         struct fq_flow_head old_flows;
 100
 101         struct rb_root  delayed;        /* for rate limited flows */
 102         u64             time_next_delayed_flow;
 103         u64             ktime_cache;    /* copy of last ktime_get_ns() */
 104         unsigned long   unthrottle_latency_ns;
 105
 106         struct fq_flow  internal;       /* for non classified or high prio packets */
 107         u32             quantum;
 108         u32             initial_quantum;
 109         u32             flow_refill_delay;
 110         u32             flow_plimit;    /* max packets per flow */
 111         unsigned long   flow_max_rate;  /* optional max rate per flow */
 112         u64             ce_threshold;
 113         u64             horizon;        /* horizon in ns */
 114         u32             orphan_mask;    /* mask for orphaned skb */
 115         u32             low_rate_threshold;
 116         struct rb_root  *fq_root;
 117         u8              rate_enable;
 118         u8              fq_trees_log;
 119         u8              horizon_drop;
 120         u32             flows;
 121         u32             inactive_flows;
 122         u32             throttled_flows;
 123
 124         u64             stat_gc_flows;
 125         u64             stat_internal_packets;
 126         u64             stat_throttled;
 127         u64             stat_ce_mark;
 128         u64             stat_horizon_drops;
 129         u64             stat_horizon_caps;
 130         u64             stat_flows_plimit;
 131         u64             stat_pkts_too_long;
 132         u64             stat_allocation_errors;
 133
 134         u32             timer_slack; /* hrtimer slack in ns */
 135         struct qdisc_watchdog watchdog;
 136 };
 137
 138 /*
 139  * f->tail and f->age share the same location.
 140  * We can use the low order bit to differentiate if this location points
 141  * to a sk_buff or contains a jiffies value, if we force this value to be odd.
 142  * This assumes f->tail low order bit must be 0 since alignof(struct sk_buff) >= 2
 143  */
 144 static void fq_flow_set_detached(struct fq_flow *f)
 145 {
 146         f->age = jiffies | 1UL;
 147 }
 148
 149 static bool fq_flow_is_detached(const struct fq_flow *f)
 150 {
 151         return !!(f->age & 1UL);
 152 }
 153
 154 /* special value to mark a throttled flow (not on old/new list) */
 155 static struct fq_flow throttled;
 156
 157 static bool fq_flow_is_throttled(const struct fq_flow *f)
 158 {
 159         return f->next == &throttled;
 160 }
 161
 162 static void fq_flow_add_tail(struct fq_flow_head *head, struct fq_flow *flow)
 163 {
 164         if (head->first)
 165                 head->last->next = flow;
 166         else
 167                 head->first = flow;
 168         head->last = flow;
 169         flow->next = NULL;
 170 }
 171
 172 static void fq_flow_unset_throttled(struct fq_sched_data *q, struct fq_flow *f)
 173 {
 174         rb_erase(&f->rate_node, &q->delayed);
 175         q->throttled_flows--;
 176         fq_flow_add_tail(&q->old_flows, f);
 177 }
 178
 179 static void fq_flow_set_throttled(struct fq_sched_data *q, struct fq_flow *f)
 180 {
 181         struct rb_node **p = &q->delayed.rb_node, *parent = NULL;
 182
 183         while (*p) {
 184                 struct fq_flow *aux;
 185
 186                 parent = *p;
 187                 aux = rb_entry(parent, struct fq_flow, rate_node);
 188                 if (f->time_next_packet >= aux->time_next_packet)
 189                         p = &parent->rb_right;
 190                 else
 191                         p = &parent->rb_left;
 192         }
 193         rb_link_node(&f->rate_node, parent, p);
 194         rb_insert_color(&f->rate_node, &q->delayed);
 195         q->throttled_flows++;
 196         q->stat_throttled++;
 197
 198         f->next = &throttled;
 199         if (q->time_next_delayed_flow > f->time_next_packet)
 200                 q->time_next_delayed_flow = f->time_next_packet;
 201 }
 202
 203
 204 static struct kmem_cache *fq_flow_cachep __read_mostly;
 205
 206
 207 /* limit number of collected flows per round */
 208 #define FQ_GC_MAX 8
 209 #define FQ_GC_AGE (3*HZ)
 210
 211 static bool fq_gc_candidate(const struct fq_flow *f)
 212 {
 213         return fq_flow_is_detached(f) &&
 214                time_after(jiffies, f->age + FQ_GC_AGE);
 215 }
 216
 217 static void fq_gc(struct fq_sched_data *q,
 218                   struct rb_root *root,
 219                   struct sock *sk)
 220 {
 221         struct rb_node **p, *parent;
 222         void *tofree[FQ_GC_MAX];
 223         struct fq_flow *f;
 224         int i, fcnt = 0;
 225
 226         p = &root->rb_node;
 227         parent = NULL;
 228         while (*p) {
 229                 parent = *p;
 230
 231                 f = rb_entry(parent, struct fq_flow, fq_node);
 232                 if (f->sk == sk)
 233                         break;
 234
 235                 if (fq_gc_candidate(f)) {
 236                         tofree[fcnt++] = f;
 237                         if (fcnt == FQ_GC_MAX)
 238                                 break;
 239                 }
 240
 241                 if (f->sk > sk)
 242                         p = &parent->rb_right;
 243                 else
 244                         p = &parent->rb_left;
 245         }
 246
 247         if (!fcnt)
 248                 return;
 249
 250         for (i = fcnt; i > 0; ) {
 251                 f = tofree[--i];
 252                 rb_erase(&f->fq_node, root);
 253         }
 254         q->flows -= fcnt;
 255         q->inactive_flows -= fcnt;
 256         q->stat_gc_flows += fcnt;
 257
 258         kmem_cache_free_bulk(fq_flow_cachep, fcnt, tofree);
 259 }
 260
 261 static struct fq_flow *fq_classify(struct sk_buff *skb, struct fq_sched_data *q)
 262 {
 263         struct rb_node **p, *parent;
 264         struct sock *sk = skb->sk;
 265         struct rb_root *root;
 266         struct fq_flow *f;
 267
 268         /* warning: no starvation prevention... */
 269         if (unlikely((skb->priority & TC_PRIO_MAX) == TC_PRIO_CONTROL))
 270                 return &q->internal;
 271
 272         /* SYNACK messages are attached to a TCP_NEW_SYN_RECV request socket
 273          * or a listener (SYNCOOKIE mode)
 274          * 1) request sockets are not full blown,
 275          *    they do not contain sk_pacing_rate
 276          * 2) They are not part of a 'flow' yet
 277          * 3) We do not want to rate limit them (eg SYNFLOOD attack),
 278          *    especially if the listener set SO_MAX_PACING_RATE
 279          * 4) We pretend they are orphaned
 280          */
 281         if (!sk || sk_listener(sk)) {
 282                 unsigned long hash = skb_get_hash(skb) & q->orphan_mask;
 283
 284                 /* By forcing low order bit to 1, we make sure to not
 285                  * collide with a local flow (socket pointers are word aligned)
 286                  */
 287                 sk = (struct sock *)((hash << 1) | 1UL);
 288                 skb_orphan(skb);
 289         } else if (sk->sk_state == TCP_CLOSE) {
 290                 unsigned long hash = skb_get_hash(skb) & q->orphan_mask;
 291                 /*
 292                  * Sockets in TCP_CLOSE are non connected.
 293                  * Typical use case is UDP sockets, they can send packets
 294                  * with sendto() to many different destinations.
 295                  * We probably could use a generic bit advertising
 296                  * non connected sockets, instead of sk_state == TCP_CLOSE,
 297                  * if we care enough.
 298                  */
 299                 sk = (struct sock *)((hash << 1) | 1UL);
 300         }
 301
 302         root = &q->fq_root[hash_ptr(sk, q->fq_trees_log)];
 303
 304         if (q->flows >= (2U << q->fq_trees_log) &&
 305             q->inactive_flows > q->flows/2)
 306                 fq_gc(q, root, sk);
 307
 308         p = &root->rb_node;
 309         parent = NULL;
 310         while (*p) {
 311                 parent = *p;
 312
 313                 f = rb_entry(parent, struct fq_flow, fq_node);
 314                 if (f->sk == sk) {
 315                         /* socket might have been reallocated, so check
 316                          * if its sk_hash is the same.
 317                          * It not, we need to refill credit with
 318                          * initial quantum
 319                          */
 320                         if (unlikely(skb->sk == sk &&
 321                                      f->socket_hash != sk->sk_hash)) {
 322                                 f->credit = q->initial_quantum;
 323                                 f->socket_hash = sk->sk_hash;
 324                                 if (q->rate_enable)
 325                                         smp_store_release(&sk->sk_pacing_status,
 326                                                           SK_PACING_FQ);
 327                                 if (fq_flow_is_throttled(f))
 328                                         fq_flow_unset_throttled(q, f);
 329                                 f->time_next_packet = 0ULL;
 330                         }
 331                         return f;
 332                 }
 333                 if (f->sk > sk)
 334                         p = &parent->rb_right;
 335                 else
 336                         p = &parent->rb_left;
 337         }
 338
 339         f = kmem_cache_zalloc(fq_flow_cachep, GFP_ATOMIC | __GFP_NOWARN);
 340         if (unlikely(!f)) {
 341                 q->stat_allocation_errors++;
 342                 return &q->internal;
 343         }
 344         /* f->t_root is already zeroed after kmem_cache_zalloc() */
 345
 346         fq_flow_set_detached(f);
 347         f->sk = sk;
 348         if (skb->sk == sk) {
 349                 f->socket_hash = sk->sk_hash;
 350                 if (q->rate_enable)
 351                         smp_store_release(&sk->sk_pacing_status,
 352                                           SK_PACING_FQ);
 353         }
 354         f->credit = q->initial_quantum;
 355
 356         rb_link_node(&f->fq_node, parent, p);
 357         rb_insert_color(&f->fq_node, root);
 358
 359         q->flows++;
 360         q->inactive_flows++;
 361         return f;
 362 }
 363
 364 static struct sk_buff *fq_peek(struct fq_flow *flow)
 365 {
 366         struct sk_buff *skb = skb_rb_first(&flow->t_root);
 367         struct sk_buff *head = flow->head;
 368
 369         if (!skb)
 370                 return head;
 371
 372         if (!head)
 373                 return skb;
 374
 375         if (fq_skb_cb(skb)->time_to_send < fq_skb_cb(head)->time_to_send)
 376                 return skb;
 377         return head;
 378 }
 379
 380 static void fq_erase_head(struct Qdisc *sch, struct fq_flow *flow,
 381                           struct sk_buff *skb)
 382 {
 383         if (skb == flow->head) {
 384                 flow->head = skb->next;
 385         } else {
 386                 rb_erase(&skb->rbnode, &flow->t_root);
 387                 skb->dev = qdisc_dev(sch);
 388         }
 389 }
 390
 391 /* Remove one skb from flow queue.
 392  * This skb must be the return value of prior fq_peek().
 393  */
 394 static void fq_dequeue_skb(struct Qdisc *sch, struct fq_flow *flow,
 395                            struct sk_buff *skb)
 396 {
 397         fq_erase_head(sch, flow, skb);
 398         skb_mark_not_on_list(skb);
 399         flow->qlen--;
 400         qdisc_qstats_backlog_dec(sch, skb);
 401         sch->q.qlen--;
 402 }
 403
 404 static void flow_queue_add(struct fq_flow *flow, struct sk_buff *skb)
 405 {
 406         struct rb_node **p, *parent;
 407         struct sk_buff *head, *aux;
 408
 409         head = flow->head;
 410         if (!head ||
 411             fq_skb_cb(skb)->time_to_send >= fq_skb_cb(flow->tail)->time_to_send) {
 412                 if (!head)
 413                         flow->head = skb;
 414                 else
 415                         flow->tail->next = skb;
 416                 flow->tail = skb;
 417                 skb->next = NULL;
 418                 return;
 419         }
 420
 421         p = &flow->t_root.rb_node;
 422         parent = NULL;
 423
 424         while (*p) {
 425                 parent = *p;
 426                 aux = rb_to_skb(parent);
 427                 if (fq_skb_cb(skb)->time_to_send >= fq_skb_cb(aux)->time_to_send)
 428                         p = &parent->rb_right;
 429                 else
 430                         p = &parent->rb_left;
 431         }
 432         rb_link_node(&skb->rbnode, parent, p);
 433         rb_insert_color(&skb->rbnode, &flow->t_root);
 434 }
 435
 436 static bool fq_packet_beyond_horizon(const struct sk_buff *skb,
 437                                     const struct fq_sched_data *q)
 438 {
 439         return unlikely((s64)skb->tstamp > (s64)(q->ktime_cache + q->horizon));
 440 }
 441
 442 static int fq_enqueue(struct sk_buff *skb, struct Qdisc *sch,
 443                       struct sk_buff **to_free)
 444 {
 445         struct fq_sched_data *q = qdisc_priv(sch);
 446         struct fq_flow *f;
 447
 448         if (unlikely(sch->q.qlen >= sch->limit))
 449                 return qdisc_drop(skb, sch, to_free);
 450
 451         if (!skb->tstamp) {
 452                 fq_skb_cb(skb)->time_to_send = q->ktime_cache = ktime_get_ns();
 453         } else {
 454                 /* Check if packet timestamp is too far in the future.
 455                  * Try first if our cached value, to avoid ktime_get_ns()
 456                  * cost in most cases.
 457                  */
 458                 if (fq_packet_beyond_horizon(skb, q)) {
 459                         /* Refresh our cache and check another time */
 460                         q->ktime_cache = ktime_get_ns();
 461                         if (fq_packet_beyond_horizon(skb, q)) {
 462                                 if (q->horizon_drop) {
 463                                         q->stat_horizon_drops++;
 464                                         return qdisc_drop(skb, sch, to_free);
 465                                 }
 466                                 q->stat_horizon_caps++;
 467                                 skb->tstamp = q->ktime_cache + q->horizon;
 468                         }
 469                 }
 470                 fq_skb_cb(skb)->time_to_send = skb->tstamp;
 471         }
 472
 473         f = fq_classify(skb, q);
 474         if (unlikely(f->qlen >= q->flow_plimit && f != &q->internal)) {
 475                 q->stat_flows_plimit++;
 476                 return qdisc_drop(skb, sch, to_free);
 477         }
 478
 479         f->qlen++;
 480         qdisc_qstats_backlog_inc(sch, skb);
 481         if (fq_flow_is_detached(f)) {
 482                 fq_flow_add_tail(&q->new_flows, f);
 483                 if (time_after(jiffies, f->age + q->flow_refill_delay))
 484                         f->credit = max_t(u32, f->credit, q->quantum);
 485                 q->inactive_flows--;
 486         }
 487
 488         /* Note: this overwrites f->age */
 489         flow_queue_add(f, skb);
 490
 491         if (unlikely(f == &q->internal)) {
 492                 q->stat_internal_packets++;
 493         }
 494         sch->q.qlen++;
 495
 496         return NET_XMIT_SUCCESS;
 497 }
 498
 499 static void fq_check_throttled(struct fq_sched_data *q, u64 now)
 500 {
 501         unsigned long sample;
 502         struct rb_node *p;
 503
 504         if (q->time_next_delayed_flow > now)
 505                 return;
 506
 507         /* Update unthrottle latency EWMA.
 508          * This is cheap and can help diagnosing timer/latency problems.
 509          */
 510         sample = (unsigned long)(now - q->time_next_delayed_flow);
 511         q->unthrottle_latency_ns -= q->unthrottle_latency_ns >> 3;
 512         q->unthrottle_latency_ns += sample >> 3;
 513
 514         q->time_next_delayed_flow = ~0ULL;
 515         while ((p = rb_first(&q->delayed)) != NULL) {
 516                 struct fq_flow *f = rb_entry(p, struct fq_flow, rate_node);
 517
 518                 if (f->time_next_packet > now) {
 519                         q->time_next_delayed_flow = f->time_next_packet;
 520                         break;
 521                 }
 522                 fq_flow_unset_throttled(q, f);
 523         }
 524 }
 525
 526 static struct sk_buff *fq_dequeue(struct Qdisc *sch)
 527 {
 528         struct fq_sched_data *q = qdisc_priv(sch);
 529         struct fq_flow_head *head;
 530         struct sk_buff *skb;
 531         struct fq_flow *f;
 532         unsigned long rate;
 533         u32 plen;
 534         u64 now;
 535
 536         if (!sch->q.qlen)
 537                 return NULL;
 538
 539         skb = fq_peek(&q->internal);
 540         if (unlikely(skb)) {
 541                 fq_dequeue_skb(sch, &q->internal, skb);
 542                 goto out;
 543         }
 544
 545         q->ktime_cache = now = ktime_get_ns();
 546         fq_check_throttled(q, now);
 547 begin:
 548         head = &q->new_flows;
 549         if (!head->first) {
 550                 head = &q->old_flows;
 551                 if (!head->first) {
 552                         if (q->time_next_delayed_flow != ~0ULL)
 553                                 qdisc_watchdog_schedule_range_ns(&q->watchdog,
 554                                                         q->time_next_delayed_flow,
 555                                                         q->timer_slack);
 556                         return NULL;
 557                 }
 558         }
 559         f = head->first;
 560
 561         if (f->credit <= 0) {
 562                 f->credit += q->quantum;
 563                 head->first = f->next;
 564                 fq_flow_add_tail(&q->old_flows, f);
 565                 goto begin;
 566         }
 567
 568         skb = fq_peek(f);
 569         if (skb) {
 570                 u64 time_next_packet = max_t(u64, fq_skb_cb(skb)->time_to_send,
 571                                              f->time_next_packet);
 572
 573                 if (now < time_next_packet) {
 574                         head->first = f->next;
 575                         f->time_next_packet = time_next_packet;
 576                         fq_flow_set_throttled(q, f);
 577                         goto begin;
 578                 }
 579                 prefetch(&skb->end);
 580                 if ((s64)(now - time_next_packet - q->ce_threshold) > 0) {
 581                         INET_ECN_set_ce(skb);
 582                         q->stat_ce_mark++;
 583                 }
 584                 fq_dequeue_skb(sch, f, skb);
 585         } else {
 586                 head->first = f->next;
 587                 /* force a pass through old_flows to prevent starvation */
 588                 if ((head == &q->new_flows) && q->old_flows.first) {
 589                         fq_flow_add_tail(&q->old_flows, f);
 590                 } else {
 591                         fq_flow_set_detached(f);
 592                         q->inactive_flows++;
 593                 }
 594                 goto begin;
 595         }
 596         plen = qdisc_pkt_len(skb);
 597         f->credit -= plen;
 598
 599         if (!q->rate_enable)
 600                 goto out;
 601
 602         rate = q->flow_max_rate;
 603
 604         /* If EDT time was provided for this skb, we need to
 605          * update f->time_next_packet only if this qdisc enforces
 606          * a flow max rate.
 607          */
 608         if (!skb->tstamp) {
 609                 if (skb->sk)
 610                         rate = min(skb->sk->sk_pacing_rate, rate);
 611
 612                 if (rate <= q->low_rate_threshold) {
 613                         f->credit = 0;
 614                 } else {
 615                         plen = max(plen, q->quantum);
 616                         if (f->credit > 0)
 617                                 goto out;
 618                 }
 619         }
 620         if (rate != ~0UL) {
 621                 u64 len = (u64)plen * NSEC_PER_SEC;
 622
 623                 if (likely(rate))
 624                         len = div64_ul(len, rate);
 625                 /* Since socket rate can change later,
 626                  * clamp the delay to 1 second.
 627                  * Really, providers of too big packets should be fixed !
 628                  */
 629                 if (unlikely(len > NSEC_PER_SEC)) {
 630                         len = NSEC_PER_SEC;
 631                         q->stat_pkts_too_long++;
 632                 }
 633                 /* Account for schedule/timers drifts.
 634                  * f->time_next_packet was set when prior packet was sent,
 635                  * and current time (@now) can be too late by tens of us.
 636                  */
 637                 if (f->time_next_packet)
 638                         len -= min(len/2, now - f->time_next_packet);
 639                 f->time_next_packet = now + len;
 640         }
 641 out:
 642         qdisc_bstats_update(sch, skb);
 643         return skb;
 644 }
 645
 646 static void fq_flow_purge(struct fq_flow *flow)
 647 {
 648         struct rb_node *p = rb_first(&flow->t_root);
 649
 650         while (p) {
 651                 struct sk_buff *skb = rb_to_skb(p);
 652
 653                 p = rb_next(p);
 654                 rb_erase(&skb->rbnode, &flow->t_root);
 655                 rtnl_kfree_skbs(skb, skb);
 656         }
 657         rtnl_kfree_skbs(flow->head, flow->tail);
 658         flow->head = NULL;
 659         flow->qlen = 0;
 660 }
 661
 662 static void fq_reset(struct Qdisc *sch)
 663 {
 664         struct fq_sched_data *q = qdisc_priv(sch);
 665         struct rb_root *root;
 666         struct rb_node *p;
 667         struct fq_flow *f;
 668         unsigned int idx;
 669
 670         sch->q.qlen = 0;
 671         sch->qstats.backlog = 0;
 672
 673         fq_flow_purge(&q->internal);
 674
 675         if (!q->fq_root)
 676                 return;
 677
 678         for (idx = 0; idx < (1U << q->fq_trees_log); idx++) {
 679                 root = &q->fq_root[idx];
 680                 while ((p = rb_first(root)) != NULL) {
 681                         f = rb_entry(p, struct fq_flow, fq_node);
 682                         rb_erase(p, root);
 683
 684                         fq_flow_purge(f);
 685
 686                         kmem_cache_free(fq_flow_cachep, f);
 687                 }
 688         }
 689         q->new_flows.first      = NULL;
 690         q->old_flows.first      = NULL;
 691         q->delayed              = RB_ROOT;
 692         q->flows                = 0;
 693         q->inactive_flows       = 0;
 694         q->throttled_flows      = 0;
 695 }
 696
 697 static void fq_rehash(struct fq_sched_data *q,
 698                       struct rb_root *old_array, u32 old_log,
 699                       struct rb_root *new_array, u32 new_log)
 700 {
 701         struct rb_node *op, **np, *parent;
 702         struct rb_root *oroot, *nroot;
 703         struct fq_flow *of, *nf;
 704         int fcnt = 0;
 705         u32 idx;
 706
 707         for (idx = 0; idx < (1U << old_log); idx++) {
 708                 oroot = &old_array[idx];
 709                 while ((op = rb_first(oroot)) != NULL) {
 710                         rb_erase(op, oroot);
 711                         of = rb_entry(op, struct fq_flow, fq_node);
 712                         if (fq_gc_candidate(of)) {
 713                                 fcnt++;
 714                                 kmem_cache_free(fq_flow_cachep, of);
 715                                 continue;
 716                         }
 717                         nroot = &new_array[hash_ptr(of->sk, new_log)];
 718
 719                         np = &nroot->rb_node;
 720                         parent = NULL;
 721                         while (*np) {
 722                                 parent = *np;
 723
 724                                 nf = rb_entry(parent, struct fq_flow, fq_node);
 725                                 BUG_ON(nf->sk == of->sk);
 726
 727                                 if (nf->sk > of->sk)
 728                                         np = &parent->rb_right;
 729                                 else
 730                                         np = &parent->rb_left;
 731                         }
 732
 733                         rb_link_node(&of->fq_node, parent, np);
 734                         rb_insert_color(&of->fq_node, nroot);
 735                 }
 736         }
 737         q->flows -= fcnt;
 738         q->inactive_flows -= fcnt;
 739         q->stat_gc_flows += fcnt;
 740 }
 741
 742 static void fq_free(void *addr)
 743 {
 744         kvfree(addr);
 745 }
 746
 747 static int fq_resize(struct Qdisc *sch, u32 log)
 748 {
 749         struct fq_sched_data *q = qdisc_priv(sch);
 750         struct rb_root *array;
 751         void *old_fq_root;
 752         u32 idx;
 753
 754         if (q->fq_root && log == q->fq_trees_log)
 755                 return 0;
 756
 757         /* If XPS was setup, we can allocate memory on right NUMA node */
 758         array = kvmalloc_node(sizeof(struct rb_root) << log, GFP_KERNEL | __GFP_RETRY_MAYFAIL,
 759                               netdev_queue_numa_node_read(sch->dev_queue));
 760         if (!array)
 761                 return -ENOMEM;
 762
 763         for (idx = 0; idx < (1U << log); idx++)
 764                 array[idx] = RB_ROOT;
 765
 766         sch_tree_lock(sch);
 767
 768         old_fq_root = q->fq_root;
 769         if (old_fq_root)
 770                 fq_rehash(q, old_fq_root, q->fq_trees_log, array, log);
 771
 772         q->fq_root = array;
 773         q->fq_trees_log = log;
 774
 775         sch_tree_unlock(sch);
 776
 777         fq_free(old_fq_root);
 778
 779         return 0;
 780 }
 781
 782 static struct netlink_range_validation iq_range = {
 783         .max = INT_MAX,
 784 };
 785
 786 static const struct nla_policy fq_policy[TCA_FQ_MAX + 1] = {
 787         [TCA_FQ_UNSPEC]                 = { .strict_start_type = TCA_FQ_TIMER_SLACK },
 788
 789         [TCA_FQ_PLIMIT]                 = { .type = NLA_U32 },
 790         [TCA_FQ_FLOW_PLIMIT]            = { .type = NLA_U32 },
 791         [TCA_FQ_QUANTUM]                = { .type = NLA_U32 },
 792         [TCA_FQ_INITIAL_QUANTUM]        = NLA_POLICY_FULL_RANGE(NLA_U32, &iq_range),
 793         [TCA_FQ_RATE_ENABLE]            = { .type = NLA_U32 },
 794         [TCA_FQ_FLOW_DEFAULT_RATE]      = { .type = NLA_U32 },
 795         [TCA_FQ_FLOW_MAX_RATE]          = { .type = NLA_U32 },
 796         [TCA_FQ_BUCKETS_LOG]            = { .type = NLA_U32 },
 797         [TCA_FQ_FLOW_REFILL_DELAY]      = { .type = NLA_U32 },
 798         [TCA_FQ_ORPHAN_MASK]            = { .type = NLA_U32 },
 799         [TCA_FQ_LOW_RATE_THRESHOLD]     = { .type = NLA_U32 },
 800         [TCA_FQ_CE_THRESHOLD]           = { .type = NLA_U32 },
 801         [TCA_FQ_TIMER_SLACK]            = { .type = NLA_U32 },
 802         [TCA_FQ_HORIZON]                = { .type = NLA_U32 },
 803         [TCA_FQ_HORIZON_DROP]           = { .type = NLA_U8 },
 804 };
 805
 806 static int fq_change(struct Qdisc *sch, struct nlattr *opt,
 807                      struct netlink_ext_ack *extack)
 808 {
 809         struct fq_sched_data *q = qdisc_priv(sch);
 810         struct nlattr *tb[TCA_FQ_MAX + 1];
 811         int err, drop_count = 0;
 812         unsigned drop_len = 0;
 813         u32 fq_log;
 814
 815         err = nla_parse_nested_deprecated(tb, TCA_FQ_MAX, opt, fq_policy,
 816                                           NULL);
 817         if (err < 0)
 818                 return err;
 819
 820         sch_tree_lock(sch);
 821
 822         fq_log = q->fq_trees_log;
 823
 824         if (tb[TCA_FQ_BUCKETS_LOG]) {
 825                 u32 nval = nla_get_u32(tb[TCA_FQ_BUCKETS_LOG]);
 826
 827                 if (nval >= 1 && nval <= ilog2(256*1024))
 828                         fq_log = nval;
 829                 else
 830                         err = -EINVAL;
 831         }
 832         if (tb[TCA_FQ_PLIMIT])
 833                 sch->limit = nla_get_u32(tb[TCA_FQ_PLIMIT]);
 834
 835         if (tb[TCA_FQ_FLOW_PLIMIT])
 836                 q->flow_plimit = nla_get_u32(tb[TCA_FQ_FLOW_PLIMIT]);
 837
 838         if (tb[TCA_FQ_QUANTUM]) {
 839                 u32 quantum = nla_get_u32(tb[TCA_FQ_QUANTUM]);
 840
 841                 if (quantum > 0 && quantum <= (1 << 20)) {
 842                         q->quantum = quantum;
 843                 } else {
 844                         NL_SET_ERR_MSG_MOD(extack, "invalid quantum");
 845                         err = -EINVAL;
 846                 }
 847         }
 848
 849         if (tb[TCA_FQ_INITIAL_QUANTUM])
 850                 q->initial_quantum = nla_get_u32(tb[TCA_FQ_INITIAL_QUANTUM]);
 851
 852         if (tb[TCA_FQ_FLOW_DEFAULT_RATE])
 853                 pr_warn_ratelimited("sch_fq: defrate %u ignored.\n",
 854                                     nla_get_u32(tb[TCA_FQ_FLOW_DEFAULT_RATE]));
 855
 856         if (tb[TCA_FQ_FLOW_MAX_RATE]) {
 857                 u32 rate = nla_get_u32(tb[TCA_FQ_FLOW_MAX_RATE]);
 858
 859                 q->flow_max_rate = (rate == ~0U) ? ~0UL : rate;
 860         }
 861         if (tb[TCA_FQ_LOW_RATE_THRESHOLD])
 862                 q->low_rate_threshold =
 863                         nla_get_u32(tb[TCA_FQ_LOW_RATE_THRESHOLD]);
 864
 865         if (tb[TCA_FQ_RATE_ENABLE]) {
 866                 u32 enable = nla_get_u32(tb[TCA_FQ_RATE_ENABLE]);
 867
 868                 if (enable <= 1)
 869                         q->rate_enable = enable;
 870                 else
 871                         err = -EINVAL;
 872         }
 873
 874         if (tb[TCA_FQ_FLOW_REFILL_DELAY]) {
 875                 u32 usecs_delay = nla_get_u32(tb[TCA_FQ_FLOW_REFILL_DELAY]) ;
 876
 877                 q->flow_refill_delay = usecs_to_jiffies(usecs_delay);
 878         }
 879
 880         if (tb[TCA_FQ_ORPHAN_MASK])
 881                 q->orphan_mask = nla_get_u32(tb[TCA_FQ_ORPHAN_MASK]);
 882
 883         if (tb[TCA_FQ_CE_THRESHOLD])
 884                 q->ce_threshold = (u64)NSEC_PER_USEC *
 885                                   nla_get_u32(tb[TCA_FQ_CE_THRESHOLD]);
 886
 887         if (tb[TCA_FQ_TIMER_SLACK])
 888                 q->timer_slack = nla_get_u32(tb[TCA_FQ_TIMER_SLACK]);
 889
 890         if (tb[TCA_FQ_HORIZON])
 891                 q->horizon = (u64)NSEC_PER_USEC *
 892                                   nla_get_u32(tb[TCA_FQ_HORIZON]);
 893
 894         if (tb[TCA_FQ_HORIZON_DROP])
 895                 q->horizon_drop = nla_get_u8(tb[TCA_FQ_HORIZON_DROP]);
 896
 897         if (!err) {
 898
 899                 sch_tree_unlock(sch);
 900                 err = fq_resize(sch, fq_log);
 901                 sch_tree_lock(sch);
 902         }
 903         while (sch->q.qlen > sch->limit) {
 904                 struct sk_buff *skb = fq_dequeue(sch);
 905
 906                 if (!skb)
 907                         break;
 908                 drop_len += qdisc_pkt_len(skb);
 909                 rtnl_kfree_skbs(skb, skb);
 910                 drop_count++;
 911         }
 912         qdisc_tree_reduce_backlog(sch, drop_count, drop_len);
 913
 914         sch_tree_unlock(sch);
 915         return err;
 916 }
 917
 918 static void fq_destroy(struct Qdisc *sch)
 919 {
 920         struct fq_sched_data *q = qdisc_priv(sch);
 921
 922         fq_reset(sch);
 923         fq_free(q->fq_root);
 924         qdisc_watchdog_cancel(&q->watchdog);
 925 }
 926
 927 static int fq_init(struct Qdisc *sch, struct nlattr *opt,
 928                    struct netlink_ext_ack *extack)
 929 {
 930         struct fq_sched_data *q = qdisc_priv(sch);
 931         int err;
 932
 933         sch->limit              = 10000;
 934         q->flow_plimit          = 100;
 935         q->quantum              = 2 * psched_mtu(qdisc_dev(sch));
 936         q->initial_quantum      = 10 * psched_mtu(qdisc_dev(sch));
 937         q->flow_refill_delay    = msecs_to_jiffies(40);
 938         q->flow_max_rate        = ~0UL;
 939         q->time_next_delayed_flow = ~0ULL;
 940         q->rate_enable          = 1;
 941         q->new_flows.first      = NULL;
 942         q->old_flows.first      = NULL;
 943         q->delayed              = RB_ROOT;
 944         q->fq_root              = NULL;
 945         q->fq_trees_log         = ilog2(1024);
 946         q->orphan_mask          = 1024 - 1;
 947         q->low_rate_threshold   = 550000 / 8;
 948
 949         q->timer_slack = 10 * NSEC_PER_USEC; /* 10 usec of hrtimer slack */
 950
 951         q->horizon = 10ULL * NSEC_PER_SEC; /* 10 seconds */
 952         q->horizon_drop = 1; /* by default, drop packets beyond horizon */
 953
 954         /* Default ce_threshold of 4294 seconds */
 955         q->ce_threshold         = (u64)NSEC_PER_USEC * ~0U;
 956
 957         qdisc_watchdog_init_clockid(&q->watchdog, sch, CLOCK_MONOTONIC);
 958
 959         if (opt)
 960                 err = fq_change(sch, opt, extack);
 961         else
 962                 err = fq_resize(sch, q->fq_trees_log);
 963
 964         return err;
 965 }
 966
 967 static int fq_dump(struct Qdisc *sch, struct sk_buff *skb)
 968 {
 969         struct fq_sched_data *q = qdisc_priv(sch);
 970         u64 ce_threshold = q->ce_threshold;
 971         u64 horizon = q->horizon;
 972         struct nlattr *opts;
 973
 974         opts = nla_nest_start_noflag(skb, TCA_OPTIONS);
 975         if (opts == NULL)
 976                 goto nla_put_failure;
 977
 978         /* TCA_FQ_FLOW_DEFAULT_RATE is not used anymore */
 979
 980         do_div(ce_threshold, NSEC_PER_USEC);
 981         do_div(horizon, NSEC_PER_USEC);
 982
 983         if (nla_put_u32(skb, TCA_FQ_PLIMIT, sch->limit) ||
 984             nla_put_u32(skb, TCA_FQ_FLOW_PLIMIT, q->flow_plimit) ||
 985             nla_put_u32(skb, TCA_FQ_QUANTUM, q->quantum) ||
 986             nla_put_u32(skb, TCA_FQ_INITIAL_QUANTUM, q->initial_quantum) ||
 987             nla_put_u32(skb, TCA_FQ_RATE_ENABLE, q->rate_enable) ||
 988             nla_put_u32(skb, TCA_FQ_FLOW_MAX_RATE,
 989                         min_t(unsigned long, q->flow_max_rate, ~0U)) ||
 990             nla_put_u32(skb, TCA_FQ_FLOW_REFILL_DELAY,
 991                         jiffies_to_usecs(q->flow_refill_delay)) ||
 992             nla_put_u32(skb, TCA_FQ_ORPHAN_MASK, q->orphan_mask) ||
 993             nla_put_u32(skb, TCA_FQ_LOW_RATE_THRESHOLD,
 994                         q->low_rate_threshold) ||
 995             nla_put_u32(skb, TCA_FQ_CE_THRESHOLD, (u32)ce_threshold) ||
 996             nla_put_u32(skb, TCA_FQ_BUCKETS_LOG, q->fq_trees_log) ||
 997             nla_put_u32(skb, TCA_FQ_TIMER_SLACK, q->timer_slack) ||
 998             nla_put_u32(skb, TCA_FQ_HORIZON, (u32)horizon) ||
 999             nla_put_u8(skb, TCA_FQ_HORIZON_DROP, q->horizon_drop))
1000                 goto nla_put_failure;
1001
1002         return nla_nest_end(skb, opts);
1003
1004 nla_put_failure:
1005         return -1;
1006 }
1007
1008 static int fq_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
1009 {
1010         struct fq_sched_data *q = qdisc_priv(sch);
1011         struct tc_fq_qd_stats st;
1012
1013         sch_tree_lock(sch);
1014
1015         st.gc_flows               = q->stat_gc_flows;
1016         st.highprio_packets       = q->stat_internal_packets;
1017         st.tcp_retrans            = 0;
1018         st.throttled              = q->stat_throttled;
1019         st.flows_plimit           = q->stat_flows_plimit;
1020         st.pkts_too_long          = q->stat_pkts_too_long;
1021         st.allocation_errors      = q->stat_allocation_errors;
1022         st.time_next_delayed_flow = q->time_next_delayed_flow + q->timer_slack -
1023                                     ktime_get_ns();
1024         st.flows                  = q->flows;
1025         st.inactive_flows         = q->inactive_flows;
1026         st.throttled_flows        = q->throttled_flows;
1027         st.unthrottle_latency_ns  = min_t(unsigned long,
1028                                           q->unthrottle_latency_ns, ~0U);
1029         st.ce_mark                = q->stat_ce_mark;
1030         st.horizon_drops          = q->stat_horizon_drops;
1031         st.horizon_caps           = q->stat_horizon_caps;
1032         sch_tree_unlock(sch);
1033
1034         return gnet_stats_copy_app(d, &st, sizeof(st));
1035 }
1036
1037 static struct Qdisc_ops fq_qdisc_ops __read_mostly = {
1038         .id             =       "fq",
1039         .priv_size      =       sizeof(struct fq_sched_data),
1040
1041         .enqueue        =       fq_enqueue,
1042         .dequeue        =       fq_dequeue,
1043         .peek           =       qdisc_peek_dequeued,
1044         .init           =       fq_init,
1045         .reset          =       fq_reset,
1046         .destroy        =       fq_destroy,
1047         .change         =       fq_change,
1048         .dump           =       fq_dump,
1049         .dump_stats     =       fq_dump_stats,
1050         .owner          =       THIS_MODULE,
1051 };
1052
1053 static int __init fq_module_init(void)
1054 {
1055         int ret;
1056
1057         fq_flow_cachep = kmem_cache_create("fq_flow_cache",
1058                                            sizeof(struct fq_flow),
1059                                            0, 0, NULL);
1060         if (!fq_flow_cachep)
1061                 return -ENOMEM;
1062
1063         ret = register_qdisc(&fq_qdisc_ops);
1064         if (ret)
1065                 kmem_cache_destroy(fq_flow_cachep);
1066         return ret;
1067 }
1068
1069 static void __exit fq_module_exit(void)
1070 {
1071         unregister_qdisc(&fq_qdisc_ops);
1072         kmem_cache_destroy(fq_flow_cachep);
1073 }
1074
1075 module_init(fq_module_init)
1076 module_exit(fq_module_exit)
1077 MODULE_AUTHOR("Eric Dumazet");
1078 MODULE_LICENSE("GPL");
1079 MODULE_DESCRIPTION("Fair Queue Packet Scheduler");