1 /* 2 * net/sched/sch_fq.c Fair Queue Packet Scheduler (per flow pacing) 3 * 4 * Copyright (C) 2013-2015 Eric Dumazet <edumazet@google.com> 5 * 6 * This program is free software; you can redistribute it and/or 7 * modify it under the terms of the GNU General Public License 8 * as published by the Free Software Foundation; either version 9 * 2 of the License, or (at your option) any later version. 10 * 11 * Meant to be mostly used for locally generated traffic : 12 * Fast classification depends on skb->sk being set before reaching us. 13 * If not, (router workload), we use rxhash as fallback, with 32 bits wide hash. 14 * All packets belonging to a socket are considered as a 'flow'. 15 * 16 * Flows are dynamically allocated and stored in a hash table of RB trees 17 * They are also part of one Round Robin 'queues' (new or old flows) 18 * 19 * Burst avoidance (aka pacing) capability : 20 * 21 * Transport (eg TCP) can set in sk->sk_pacing_rate a rate, enqueue a 22 * bunch of packets, and this packet scheduler adds delay between 23 * packets to respect rate limitation. 24 * 25 * enqueue() : 26 * - lookup one RB tree (out of 1024 or more) to find the flow. 27 * If non existent flow, create it, add it to the tree. 28 * Add skb to the per flow list of skb (fifo). 29 * - Use a special fifo for high prio packets 30 * 31 * dequeue() : serves flows in Round Robin 32 * Note : When a flow becomes empty, we do not immediately remove it from 33 * rb trees, for performance reasons (its expected to send additional packets, 34 * or SLAB cache will reuse socket for another flow) 35 */ 36 37 #include <linux/module.h> 38 #include <linux/types.h> 39 #include <linux/kernel.h> 40 #include <linux/jiffies.h> 41 #include <linux/string.h> 42 #include <linux/in.h> 43 #include <linux/errno.h> 44 #include <linux/init.h> 45 #include <linux/skbuff.h> 46 #include <linux/slab.h> 47 #include <linux/rbtree.h> 48 #include <linux/hash.h> 49 #include <linux/prefetch.h> 50 #include <linux/vmalloc.h> 51 #include <net/netlink.h> 52 #include <net/pkt_sched.h> 53 #include <net/sock.h> 54 #include <net/tcp_states.h> 55 #include <net/tcp.h> 56 57 /* 58 * Per flow structure, dynamically allocated 59 */ 60 struct fq_flow { 61 struct sk_buff *head; /* list of skbs for this flow : first skb */ 62 union { 63 struct sk_buff *tail; /* last skb in the list */ 64 unsigned long age; /* jiffies when flow was emptied, for gc */ 65 }; 66 struct rb_node fq_node; /* anchor in fq_root[] trees */ 67 struct sock *sk; 68 int qlen; /* number of packets in flow queue */ 69 int credit; 70 u32 socket_hash; /* sk_hash */ 71 struct fq_flow *next; /* next pointer in RR lists, or &detached */ 72 73 struct rb_node rate_node; /* anchor in q->delayed tree */ 74 u64 time_next_packet; 75 }; 76 77 struct fq_flow_head { 78 struct fq_flow *first; 79 struct fq_flow *last; 80 }; 81 82 struct fq_sched_data { 83 struct fq_flow_head new_flows; 84 85 struct fq_flow_head old_flows; 86 87 struct rb_root delayed; /* for rate limited flows */ 88 u64 time_next_delayed_flow; 89 unsigned long unthrottle_latency_ns; 90 91 struct fq_flow internal; /* for non classified or high prio packets */ 92 u32 quantum; 93 u32 initial_quantum; 94 u32 flow_refill_delay; 95 u32 flow_plimit; /* max packets per flow */ 96 unsigned long flow_max_rate; /* optional max rate per flow */ 97 u64 ce_threshold; 98 u32 orphan_mask; /* mask for orphaned skb */ 99 u32 low_rate_threshold; 100 struct rb_root *fq_root; 101 u8 rate_enable; 102 u8 fq_trees_log; 103 104 u32 flows; 105 u32 inactive_flows; 106 u32 throttled_flows; 107 108 u64 stat_gc_flows; 109 u64 stat_internal_packets; 110 u64 stat_throttled; 111 u64 stat_ce_mark; 112 u64 stat_flows_plimit; 113 u64 stat_pkts_too_long; 114 u64 stat_allocation_errors; 115 struct qdisc_watchdog watchdog; 116 }; 117 118 /* special value to mark a detached flow (not on old/new list) */ 119 static struct fq_flow detached, throttled; 120 121 static void fq_flow_set_detached(struct fq_flow *f) 122 { 123 f->next = &detached; 124 f->age = jiffies; 125 } 126 127 static bool fq_flow_is_detached(const struct fq_flow *f) 128 { 129 return f->next == &detached; 130 } 131 132 static bool fq_flow_is_throttled(const struct fq_flow *f) 133 { 134 return f->next == &throttled; 135 } 136 137 static void fq_flow_add_tail(struct fq_flow_head *head, struct fq_flow *flow) 138 { 139 if (head->first) 140 head->last->next = flow; 141 else 142 head->first = flow; 143 head->last = flow; 144 flow->next = NULL; 145 } 146 147 static void fq_flow_unset_throttled(struct fq_sched_data *q, struct fq_flow *f) 148 { 149 rb_erase(&f->rate_node, &q->delayed); 150 q->throttled_flows--; 151 fq_flow_add_tail(&q->old_flows, f); 152 } 153 154 static void fq_flow_set_throttled(struct fq_sched_data *q, struct fq_flow *f) 155 { 156 struct rb_node **p = &q->delayed.rb_node, *parent = NULL; 157 158 while (*p) { 159 struct fq_flow *aux; 160 161 parent = *p; 162 aux = rb_entry(parent, struct fq_flow, rate_node); 163 if (f->time_next_packet >= aux->time_next_packet) 164 p = &parent->rb_right; 165 else 166 p = &parent->rb_left; 167 } 168 rb_link_node(&f->rate_node, parent, p); 169 rb_insert_color(&f->rate_node, &q->delayed); 170 q->throttled_flows++; 171 q->stat_throttled++; 172 173 f->next = &throttled; 174 if (q->time_next_delayed_flow > f->time_next_packet) 175 q->time_next_delayed_flow = f->time_next_packet; 176 } 177 178 179 static struct kmem_cache *fq_flow_cachep __read_mostly; 180 181 182 /* limit number of collected flows per round */ 183 #define FQ_GC_MAX 8 184 #define FQ_GC_AGE (3*HZ) 185 186 static bool fq_gc_candidate(const struct fq_flow *f) 187 { 188 return fq_flow_is_detached(f) && 189 time_after(jiffies, f->age + FQ_GC_AGE); 190 } 191 192 static void fq_gc(struct fq_sched_data *q, 193 struct rb_root *root, 194 struct sock *sk) 195 { 196 struct fq_flow *f, *tofree[FQ_GC_MAX]; 197 struct rb_node **p, *parent; 198 int fcnt = 0; 199 200 p = &root->rb_node; 201 parent = NULL; 202 while (*p) { 203 parent = *p; 204 205 f = rb_entry(parent, struct fq_flow, fq_node); 206 if (f->sk == sk) 207 break; 208 209 if (fq_gc_candidate(f)) { 210 tofree[fcnt++] = f; 211 if (fcnt == FQ_GC_MAX) 212 break; 213 } 214 215 if (f->sk > sk) 216 p = &parent->rb_right; 217 else 218 p = &parent->rb_left; 219 } 220 221 q->flows -= fcnt; 222 q->inactive_flows -= fcnt; 223 q->stat_gc_flows += fcnt; 224 while (fcnt) { 225 struct fq_flow *f = tofree[--fcnt]; 226 227 rb_erase(&f->fq_node, root); 228 kmem_cache_free(fq_flow_cachep, f); 229 } 230 } 231 232 static struct fq_flow *fq_classify(struct sk_buff *skb, struct fq_sched_data *q) 233 { 234 struct rb_node **p, *parent; 235 struct sock *sk = skb->sk; 236 struct rb_root *root; 237 struct fq_flow *f; 238 239 /* warning: no starvation prevention... */ 240 if (unlikely((skb->priority & TC_PRIO_MAX) == TC_PRIO_CONTROL)) 241 return &q->internal; 242 243 /* SYNACK messages are attached to a TCP_NEW_SYN_RECV request socket 244 * or a listener (SYNCOOKIE mode) 245 * 1) request sockets are not full blown, 246 * they do not contain sk_pacing_rate 247 * 2) They are not part of a 'flow' yet 248 * 3) We do not want to rate limit them (eg SYNFLOOD attack), 249 * especially if the listener set SO_MAX_PACING_RATE 250 * 4) We pretend they are orphaned 251 */ 252 if (!sk || sk_listener(sk)) { 253 unsigned long hash = skb_get_hash(skb) & q->orphan_mask; 254 255 /* By forcing low order bit to 1, we make sure to not 256 * collide with a local flow (socket pointers are word aligned) 257 */ 258 sk = (struct sock *)((hash << 1) | 1UL); 259 skb_orphan(skb); 260 } 261 262 root = &q->fq_root[hash_ptr(sk, q->fq_trees_log)]; 263 264 if (q->flows >= (2U << q->fq_trees_log) && 265 q->inactive_flows > q->flows/2) 266 fq_gc(q, root, sk); 267 268 p = &root->rb_node; 269 parent = NULL; 270 while (*p) { 271 parent = *p; 272 273 f = rb_entry(parent, struct fq_flow, fq_node); 274 if (f->sk == sk) { 275 /* socket might have been reallocated, so check 276 * if its sk_hash is the same. 277 * It not, we need to refill credit with 278 * initial quantum 279 */ 280 if (unlikely(skb->sk && 281 f->socket_hash != sk->sk_hash)) { 282 f->credit = q->initial_quantum; 283 f->socket_hash = sk->sk_hash; 284 if (fq_flow_is_throttled(f)) 285 fq_flow_unset_throttled(q, f); 286 f->time_next_packet = 0ULL; 287 } 288 return f; 289 } 290 if (f->sk > sk) 291 p = &parent->rb_right; 292 else 293 p = &parent->rb_left; 294 } 295 296 f = kmem_cache_zalloc(fq_flow_cachep, GFP_ATOMIC | __GFP_NOWARN); 297 if (unlikely(!f)) { 298 q->stat_allocation_errors++; 299 return &q->internal; 300 } 301 fq_flow_set_detached(f); 302 f->sk = sk; 303 if (skb->sk) 304 f->socket_hash = sk->sk_hash; 305 f->credit = q->initial_quantum; 306 307 rb_link_node(&f->fq_node, parent, p); 308 rb_insert_color(&f->fq_node, root); 309 310 q->flows++; 311 q->inactive_flows++; 312 return f; 313 } 314 315 316 /* remove one skb from head of flow queue */ 317 static struct sk_buff *fq_dequeue_head(struct Qdisc *sch, struct fq_flow *flow) 318 { 319 struct sk_buff *skb = flow->head; 320 321 if (skb) { 322 flow->head = skb->next; 323 skb_mark_not_on_list(skb); 324 flow->qlen--; 325 qdisc_qstats_backlog_dec(sch, skb); 326 sch->q.qlen--; 327 } 328 return skb; 329 } 330 331 static void flow_queue_add(struct fq_flow *flow, struct sk_buff *skb) 332 { 333 struct sk_buff *head = flow->head; 334 335 skb->next = NULL; 336 if (!head) 337 flow->head = skb; 338 else 339 flow->tail->next = skb; 340 341 flow->tail = skb; 342 } 343 344 static int fq_enqueue(struct sk_buff *skb, struct Qdisc *sch, 345 struct sk_buff **to_free) 346 { 347 struct fq_sched_data *q = qdisc_priv(sch); 348 struct fq_flow *f; 349 350 if (unlikely(sch->q.qlen >= sch->limit)) 351 return qdisc_drop(skb, sch, to_free); 352 353 f = fq_classify(skb, q); 354 if (unlikely(f->qlen >= q->flow_plimit && f != &q->internal)) { 355 q->stat_flows_plimit++; 356 return qdisc_drop(skb, sch, to_free); 357 } 358 359 f->qlen++; 360 qdisc_qstats_backlog_inc(sch, skb); 361 if (fq_flow_is_detached(f)) { 362 struct sock *sk = skb->sk; 363 364 fq_flow_add_tail(&q->new_flows, f); 365 if (time_after(jiffies, f->age + q->flow_refill_delay)) 366 f->credit = max_t(u32, f->credit, q->quantum); 367 if (sk && q->rate_enable) { 368 if (unlikely(smp_load_acquire(&sk->sk_pacing_status) != 369 SK_PACING_FQ)) 370 smp_store_release(&sk->sk_pacing_status, 371 SK_PACING_FQ); 372 } 373 q->inactive_flows--; 374 } 375 376 /* Note: this overwrites f->age */ 377 flow_queue_add(f, skb); 378 379 if (unlikely(f == &q->internal)) { 380 q->stat_internal_packets++; 381 } 382 sch->q.qlen++; 383 384 return NET_XMIT_SUCCESS; 385 } 386 387 static void fq_check_throttled(struct fq_sched_data *q, u64 now) 388 { 389 unsigned long sample; 390 struct rb_node *p; 391 392 if (q->time_next_delayed_flow > now) 393 return; 394 395 /* Update unthrottle latency EWMA. 396 * This is cheap and can help diagnosing timer/latency problems. 397 */ 398 sample = (unsigned long)(now - q->time_next_delayed_flow); 399 q->unthrottle_latency_ns -= q->unthrottle_latency_ns >> 3; 400 q->unthrottle_latency_ns += sample >> 3; 401 402 q->time_next_delayed_flow = ~0ULL; 403 while ((p = rb_first(&q->delayed)) != NULL) { 404 struct fq_flow *f = rb_entry(p, struct fq_flow, rate_node); 405 406 if (f->time_next_packet > now) { 407 q->time_next_delayed_flow = f->time_next_packet; 408 break; 409 } 410 fq_flow_unset_throttled(q, f); 411 } 412 } 413 414 static struct sk_buff *fq_dequeue(struct Qdisc *sch) 415 { 416 struct fq_sched_data *q = qdisc_priv(sch); 417 struct fq_flow_head *head; 418 struct sk_buff *skb; 419 struct fq_flow *f; 420 unsigned long rate; 421 u32 plen; 422 u64 now; 423 424 if (!sch->q.qlen) 425 return NULL; 426 427 skb = fq_dequeue_head(sch, &q->internal); 428 if (skb) 429 goto out; 430 431 now = ktime_get_ns(); 432 fq_check_throttled(q, now); 433 begin: 434 head = &q->new_flows; 435 if (!head->first) { 436 head = &q->old_flows; 437 if (!head->first) { 438 if (q->time_next_delayed_flow != ~0ULL) 439 qdisc_watchdog_schedule_ns(&q->watchdog, 440 q->time_next_delayed_flow); 441 return NULL; 442 } 443 } 444 f = head->first; 445 446 if (f->credit <= 0) { 447 f->credit += q->quantum; 448 head->first = f->next; 449 fq_flow_add_tail(&q->old_flows, f); 450 goto begin; 451 } 452 453 skb = f->head; 454 if (skb) { 455 u64 time_next_packet = max_t(u64, ktime_to_ns(skb->tstamp), 456 f->time_next_packet); 457 458 if (now < time_next_packet) { 459 head->first = f->next; 460 f->time_next_packet = time_next_packet; 461 fq_flow_set_throttled(q, f); 462 goto begin; 463 } 464 if (time_next_packet && 465 (s64)(now - time_next_packet - q->ce_threshold) > 0) { 466 INET_ECN_set_ce(skb); 467 q->stat_ce_mark++; 468 } 469 } 470 471 skb = fq_dequeue_head(sch, f); 472 if (!skb) { 473 head->first = f->next; 474 /* force a pass through old_flows to prevent starvation */ 475 if ((head == &q->new_flows) && q->old_flows.first) { 476 fq_flow_add_tail(&q->old_flows, f); 477 } else { 478 fq_flow_set_detached(f); 479 q->inactive_flows++; 480 } 481 goto begin; 482 } 483 prefetch(&skb->end); 484 plen = qdisc_pkt_len(skb); 485 f->credit -= plen; 486 487 if (!q->rate_enable) 488 goto out; 489 490 rate = q->flow_max_rate; 491 492 /* If EDT time was provided for this skb, we need to 493 * update f->time_next_packet only if this qdisc enforces 494 * a flow max rate. 495 */ 496 if (!skb->tstamp) { 497 if (skb->sk) 498 rate = min(skb->sk->sk_pacing_rate, rate); 499 500 if (rate <= q->low_rate_threshold) { 501 f->credit = 0; 502 } else { 503 plen = max(plen, q->quantum); 504 if (f->credit > 0) 505 goto out; 506 } 507 } 508 if (rate != ~0UL) { 509 u64 len = (u64)plen * NSEC_PER_SEC; 510 511 if (likely(rate)) 512 len = div64_ul(len, rate); 513 /* Since socket rate can change later, 514 * clamp the delay to 1 second. 515 * Really, providers of too big packets should be fixed ! 516 */ 517 if (unlikely(len > NSEC_PER_SEC)) { 518 len = NSEC_PER_SEC; 519 q->stat_pkts_too_long++; 520 } 521 /* Account for schedule/timers drifts. 522 * f->time_next_packet was set when prior packet was sent, 523 * and current time (@now) can be too late by tens of us. 524 */ 525 if (f->time_next_packet) 526 len -= min(len/2, now - f->time_next_packet); 527 f->time_next_packet = now + len; 528 } 529 out: 530 qdisc_bstats_update(sch, skb); 531 return skb; 532 } 533 534 static void fq_flow_purge(struct fq_flow *flow) 535 { 536 rtnl_kfree_skbs(flow->head, flow->tail); 537 flow->head = NULL; 538 flow->qlen = 0; 539 } 540 541 static void fq_reset(struct Qdisc *sch) 542 { 543 struct fq_sched_data *q = qdisc_priv(sch); 544 struct rb_root *root; 545 struct rb_node *p; 546 struct fq_flow *f; 547 unsigned int idx; 548 549 sch->q.qlen = 0; 550 sch->qstats.backlog = 0; 551 552 fq_flow_purge(&q->internal); 553 554 if (!q->fq_root) 555 return; 556 557 for (idx = 0; idx < (1U << q->fq_trees_log); idx++) { 558 root = &q->fq_root[idx]; 559 while ((p = rb_first(root)) != NULL) { 560 f = rb_entry(p, struct fq_flow, fq_node); 561 rb_erase(p, root); 562 563 fq_flow_purge(f); 564 565 kmem_cache_free(fq_flow_cachep, f); 566 } 567 } 568 q->new_flows.first = NULL; 569 q->old_flows.first = NULL; 570 q->delayed = RB_ROOT; 571 q->flows = 0; 572 q->inactive_flows = 0; 573 q->throttled_flows = 0; 574 } 575 576 static void fq_rehash(struct fq_sched_data *q, 577 struct rb_root *old_array, u32 old_log, 578 struct rb_root *new_array, u32 new_log) 579 { 580 struct rb_node *op, **np, *parent; 581 struct rb_root *oroot, *nroot; 582 struct fq_flow *of, *nf; 583 int fcnt = 0; 584 u32 idx; 585 586 for (idx = 0; idx < (1U << old_log); idx++) { 587 oroot = &old_array[idx]; 588 while ((op = rb_first(oroot)) != NULL) { 589 rb_erase(op, oroot); 590 of = rb_entry(op, struct fq_flow, fq_node); 591 if (fq_gc_candidate(of)) { 592 fcnt++; 593 kmem_cache_free(fq_flow_cachep, of); 594 continue; 595 } 596 nroot = &new_array[hash_ptr(of->sk, new_log)]; 597 598 np = &nroot->rb_node; 599 parent = NULL; 600 while (*np) { 601 parent = *np; 602 603 nf = rb_entry(parent, struct fq_flow, fq_node); 604 BUG_ON(nf->sk == of->sk); 605 606 if (nf->sk > of->sk) 607 np = &parent->rb_right; 608 else 609 np = &parent->rb_left; 610 } 611 612 rb_link_node(&of->fq_node, parent, np); 613 rb_insert_color(&of->fq_node, nroot); 614 } 615 } 616 q->flows -= fcnt; 617 q->inactive_flows -= fcnt; 618 q->stat_gc_flows += fcnt; 619 } 620 621 static void fq_free(void *addr) 622 { 623 kvfree(addr); 624 } 625 626 static int fq_resize(struct Qdisc *sch, u32 log) 627 { 628 struct fq_sched_data *q = qdisc_priv(sch); 629 struct rb_root *array; 630 void *old_fq_root; 631 u32 idx; 632 633 if (q->fq_root && log == q->fq_trees_log) 634 return 0; 635 636 /* If XPS was setup, we can allocate memory on right NUMA node */ 637 array = kvmalloc_node(sizeof(struct rb_root) << log, GFP_KERNEL | __GFP_RETRY_MAYFAIL, 638 netdev_queue_numa_node_read(sch->dev_queue)); 639 if (!array) 640 return -ENOMEM; 641 642 for (idx = 0; idx < (1U << log); idx++) 643 array[idx] = RB_ROOT; 644 645 sch_tree_lock(sch); 646 647 old_fq_root = q->fq_root; 648 if (old_fq_root) 649 fq_rehash(q, old_fq_root, q->fq_trees_log, array, log); 650 651 q->fq_root = array; 652 q->fq_trees_log = log; 653 654 sch_tree_unlock(sch); 655 656 fq_free(old_fq_root); 657 658 return 0; 659 } 660 661 static const struct nla_policy fq_policy[TCA_FQ_MAX + 1] = { 662 [TCA_FQ_PLIMIT] = { .type = NLA_U32 }, 663 [TCA_FQ_FLOW_PLIMIT] = { .type = NLA_U32 }, 664 [TCA_FQ_QUANTUM] = { .type = NLA_U32 }, 665 [TCA_FQ_INITIAL_QUANTUM] = { .type = NLA_U32 }, 666 [TCA_FQ_RATE_ENABLE] = { .type = NLA_U32 }, 667 [TCA_FQ_FLOW_DEFAULT_RATE] = { .type = NLA_U32 }, 668 [TCA_FQ_FLOW_MAX_RATE] = { .type = NLA_U32 }, 669 [TCA_FQ_BUCKETS_LOG] = { .type = NLA_U32 }, 670 [TCA_FQ_FLOW_REFILL_DELAY] = { .type = NLA_U32 }, 671 [TCA_FQ_LOW_RATE_THRESHOLD] = { .type = NLA_U32 }, 672 [TCA_FQ_CE_THRESHOLD] = { .type = NLA_U32 }, 673 }; 674 675 static int fq_change(struct Qdisc *sch, struct nlattr *opt, 676 struct netlink_ext_ack *extack) 677 { 678 struct fq_sched_data *q = qdisc_priv(sch); 679 struct nlattr *tb[TCA_FQ_MAX + 1]; 680 int err, drop_count = 0; 681 unsigned drop_len = 0; 682 u32 fq_log; 683 684 if (!opt) 685 return -EINVAL; 686 687 err = nla_parse_nested(tb, TCA_FQ_MAX, opt, fq_policy, NULL); 688 if (err < 0) 689 return err; 690 691 sch_tree_lock(sch); 692 693 fq_log = q->fq_trees_log; 694 695 if (tb[TCA_FQ_BUCKETS_LOG]) { 696 u32 nval = nla_get_u32(tb[TCA_FQ_BUCKETS_LOG]); 697 698 if (nval >= 1 && nval <= ilog2(256*1024)) 699 fq_log = nval; 700 else 701 err = -EINVAL; 702 } 703 if (tb[TCA_FQ_PLIMIT]) 704 sch->limit = nla_get_u32(tb[TCA_FQ_PLIMIT]); 705 706 if (tb[TCA_FQ_FLOW_PLIMIT]) 707 q->flow_plimit = nla_get_u32(tb[TCA_FQ_FLOW_PLIMIT]); 708 709 if (tb[TCA_FQ_QUANTUM]) { 710 u32 quantum = nla_get_u32(tb[TCA_FQ_QUANTUM]); 711 712 if (quantum > 0) 713 q->quantum = quantum; 714 else 715 err = -EINVAL; 716 } 717 718 if (tb[TCA_FQ_INITIAL_QUANTUM]) 719 q->initial_quantum = nla_get_u32(tb[TCA_FQ_INITIAL_QUANTUM]); 720 721 if (tb[TCA_FQ_FLOW_DEFAULT_RATE]) 722 pr_warn_ratelimited("sch_fq: defrate %u ignored.\n", 723 nla_get_u32(tb[TCA_FQ_FLOW_DEFAULT_RATE])); 724 725 if (tb[TCA_FQ_FLOW_MAX_RATE]) { 726 u32 rate = nla_get_u32(tb[TCA_FQ_FLOW_MAX_RATE]); 727 728 q->flow_max_rate = (rate == ~0U) ? ~0UL : rate; 729 } 730 if (tb[TCA_FQ_LOW_RATE_THRESHOLD]) 731 q->low_rate_threshold = 732 nla_get_u32(tb[TCA_FQ_LOW_RATE_THRESHOLD]); 733 734 if (tb[TCA_FQ_RATE_ENABLE]) { 735 u32 enable = nla_get_u32(tb[TCA_FQ_RATE_ENABLE]); 736 737 if (enable <= 1) 738 q->rate_enable = enable; 739 else 740 err = -EINVAL; 741 } 742 743 if (tb[TCA_FQ_FLOW_REFILL_DELAY]) { 744 u32 usecs_delay = nla_get_u32(tb[TCA_FQ_FLOW_REFILL_DELAY]) ; 745 746 q->flow_refill_delay = usecs_to_jiffies(usecs_delay); 747 } 748 749 if (tb[TCA_FQ_ORPHAN_MASK]) 750 q->orphan_mask = nla_get_u32(tb[TCA_FQ_ORPHAN_MASK]); 751 752 if (tb[TCA_FQ_CE_THRESHOLD]) 753 q->ce_threshold = (u64)NSEC_PER_USEC * 754 nla_get_u32(tb[TCA_FQ_CE_THRESHOLD]); 755 756 if (!err) { 757 sch_tree_unlock(sch); 758 err = fq_resize(sch, fq_log); 759 sch_tree_lock(sch); 760 } 761 while (sch->q.qlen > sch->limit) { 762 struct sk_buff *skb = fq_dequeue(sch); 763 764 if (!skb) 765 break; 766 drop_len += qdisc_pkt_len(skb); 767 rtnl_kfree_skbs(skb, skb); 768 drop_count++; 769 } 770 qdisc_tree_reduce_backlog(sch, drop_count, drop_len); 771 772 sch_tree_unlock(sch); 773 return err; 774 } 775 776 static void fq_destroy(struct Qdisc *sch) 777 { 778 struct fq_sched_data *q = qdisc_priv(sch); 779 780 fq_reset(sch); 781 fq_free(q->fq_root); 782 qdisc_watchdog_cancel(&q->watchdog); 783 } 784 785 static int fq_init(struct Qdisc *sch, struct nlattr *opt, 786 struct netlink_ext_ack *extack) 787 { 788 struct fq_sched_data *q = qdisc_priv(sch); 789 int err; 790 791 sch->limit = 10000; 792 q->flow_plimit = 100; 793 q->quantum = 2 * psched_mtu(qdisc_dev(sch)); 794 q->initial_quantum = 10 * psched_mtu(qdisc_dev(sch)); 795 q->flow_refill_delay = msecs_to_jiffies(40); 796 q->flow_max_rate = ~0UL; 797 q->time_next_delayed_flow = ~0ULL; 798 q->rate_enable = 1; 799 q->new_flows.first = NULL; 800 q->old_flows.first = NULL; 801 q->delayed = RB_ROOT; 802 q->fq_root = NULL; 803 q->fq_trees_log = ilog2(1024); 804 q->orphan_mask = 1024 - 1; 805 q->low_rate_threshold = 550000 / 8; 806 807 /* Default ce_threshold of 4294 seconds */ 808 q->ce_threshold = (u64)NSEC_PER_USEC * ~0U; 809 810 qdisc_watchdog_init_clockid(&q->watchdog, sch, CLOCK_MONOTONIC); 811 812 if (opt) 813 err = fq_change(sch, opt, extack); 814 else 815 err = fq_resize(sch, q->fq_trees_log); 816 817 return err; 818 } 819 820 static int fq_dump(struct Qdisc *sch, struct sk_buff *skb) 821 { 822 struct fq_sched_data *q = qdisc_priv(sch); 823 u64 ce_threshold = q->ce_threshold; 824 struct nlattr *opts; 825 826 opts = nla_nest_start(skb, TCA_OPTIONS); 827 if (opts == NULL) 828 goto nla_put_failure; 829 830 /* TCA_FQ_FLOW_DEFAULT_RATE is not used anymore */ 831 832 do_div(ce_threshold, NSEC_PER_USEC); 833 834 if (nla_put_u32(skb, TCA_FQ_PLIMIT, sch->limit) || 835 nla_put_u32(skb, TCA_FQ_FLOW_PLIMIT, q->flow_plimit) || 836 nla_put_u32(skb, TCA_FQ_QUANTUM, q->quantum) || 837 nla_put_u32(skb, TCA_FQ_INITIAL_QUANTUM, q->initial_quantum) || 838 nla_put_u32(skb, TCA_FQ_RATE_ENABLE, q->rate_enable) || 839 nla_put_u32(skb, TCA_FQ_FLOW_MAX_RATE, 840 min_t(unsigned long, q->flow_max_rate, ~0U)) || 841 nla_put_u32(skb, TCA_FQ_FLOW_REFILL_DELAY, 842 jiffies_to_usecs(q->flow_refill_delay)) || 843 nla_put_u32(skb, TCA_FQ_ORPHAN_MASK, q->orphan_mask) || 844 nla_put_u32(skb, TCA_FQ_LOW_RATE_THRESHOLD, 845 q->low_rate_threshold) || 846 nla_put_u32(skb, TCA_FQ_CE_THRESHOLD, (u32)ce_threshold) || 847 nla_put_u32(skb, TCA_FQ_BUCKETS_LOG, q->fq_trees_log)) 848 goto nla_put_failure; 849 850 return nla_nest_end(skb, opts); 851 852 nla_put_failure: 853 return -1; 854 } 855 856 static int fq_dump_stats(struct Qdisc *sch, struct gnet_dump *d) 857 { 858 struct fq_sched_data *q = qdisc_priv(sch); 859 struct tc_fq_qd_stats st; 860 861 sch_tree_lock(sch); 862 863 st.gc_flows = q->stat_gc_flows; 864 st.highprio_packets = q->stat_internal_packets; 865 st.tcp_retrans = 0; 866 st.throttled = q->stat_throttled; 867 st.flows_plimit = q->stat_flows_plimit; 868 st.pkts_too_long = q->stat_pkts_too_long; 869 st.allocation_errors = q->stat_allocation_errors; 870 st.time_next_delayed_flow = q->time_next_delayed_flow - ktime_get_ns(); 871 st.flows = q->flows; 872 st.inactive_flows = q->inactive_flows; 873 st.throttled_flows = q->throttled_flows; 874 st.unthrottle_latency_ns = min_t(unsigned long, 875 q->unthrottle_latency_ns, ~0U); 876 st.ce_mark = q->stat_ce_mark; 877 sch_tree_unlock(sch); 878 879 return gnet_stats_copy_app(d, &st, sizeof(st)); 880 } 881 882 static struct Qdisc_ops fq_qdisc_ops __read_mostly = { 883 .id = "fq", 884 .priv_size = sizeof(struct fq_sched_data), 885 886 .enqueue = fq_enqueue, 887 .dequeue = fq_dequeue, 888 .peek = qdisc_peek_dequeued, 889 .init = fq_init, 890 .reset = fq_reset, 891 .destroy = fq_destroy, 892 .change = fq_change, 893 .dump = fq_dump, 894 .dump_stats = fq_dump_stats, 895 .owner = THIS_MODULE, 896 }; 897 898 static int __init fq_module_init(void) 899 { 900 int ret; 901 902 fq_flow_cachep = kmem_cache_create("fq_flow_cache", 903 sizeof(struct fq_flow), 904 0, 0, NULL); 905 if (!fq_flow_cachep) 906 return -ENOMEM; 907 908 ret = register_qdisc(&fq_qdisc_ops); 909 if (ret) 910 kmem_cache_destroy(fq_flow_cachep); 911 return ret; 912 } 913 914 static void __exit fq_module_exit(void) 915 { 916 unregister_qdisc(&fq_qdisc_ops); 917 kmem_cache_destroy(fq_flow_cachep); 918 } 919 920 module_init(fq_module_init) 921 module_exit(fq_module_exit) 922 MODULE_AUTHOR("Eric Dumazet"); 923 MODULE_LICENSE("GPL"); 924