2 * net/sched/sch_fq.c Fair Queue Packet Scheduler (per flow pacing)
4 * Copyright (C) 2013 Eric Dumazet <edumazet@google.com>
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.
11 * Meant to be mostly used for localy 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'.
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)
19 * Burst avoidance (aka pacing) capability :
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.
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
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)
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>
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 <net/netlink.h>
51 #include <net/pkt_sched.h>
53 #include <net/tcp_states.h>
56 * Per flow structure, dynamically allocated
59 struct sk_buff *head; /* list of skbs for this flow : first skb */
61 struct sk_buff *tail; /* last skb in the list */
62 unsigned long age; /* jiffies when flow was emptied, for gc */
64 struct rb_node fq_node; /* anchor in fq_root[] trees */
66 int qlen; /* number of packets in flow queue */
68 u32 socket_hash; /* sk_hash */
69 struct fq_flow *next; /* next pointer in RR lists, or &detached */
71 struct rb_node rate_node; /* anchor in q->delayed tree */
76 struct fq_flow *first;
80 struct fq_sched_data {
81 struct fq_flow_head new_flows;
83 struct fq_flow_head old_flows;
85 struct rb_root delayed; /* for rate limited flows */
86 u64 time_next_delayed_flow;
88 struct fq_flow internal; /* for non classified or high prio packets */
91 u32 flow_default_rate;/* rate per flow : bytes per second */
92 u32 flow_max_rate; /* optional max rate per flow */
93 u32 flow_plimit; /* max packets per flow */
94 struct rb_root *fq_root;
103 u64 stat_internal_packets;
104 u64 stat_tcp_retrans;
106 u64 stat_flows_plimit;
107 u64 stat_pkts_too_long;
108 u64 stat_allocation_errors;
109 struct qdisc_watchdog watchdog;
112 /* special value to mark a detached flow (not on old/new list) */
113 static struct fq_flow detached, throttled;
115 static void fq_flow_set_detached(struct fq_flow *f)
120 static bool fq_flow_is_detached(const struct fq_flow *f)
122 return f->next == &detached;
125 static void fq_flow_set_throttled(struct fq_sched_data *q, struct fq_flow *f)
127 struct rb_node **p = &q->delayed.rb_node, *parent = NULL;
133 aux = container_of(parent, struct fq_flow, rate_node);
134 if (f->time_next_packet >= aux->time_next_packet)
135 p = &parent->rb_right;
137 p = &parent->rb_left;
139 rb_link_node(&f->rate_node, parent, p);
140 rb_insert_color(&f->rate_node, &q->delayed);
141 q->throttled_flows++;
144 f->next = &throttled;
145 if (q->time_next_delayed_flow > f->time_next_packet)
146 q->time_next_delayed_flow = f->time_next_packet;
150 static struct kmem_cache *fq_flow_cachep __read_mostly;
152 static void fq_flow_add_tail(struct fq_flow_head *head, struct fq_flow *flow)
155 head->last->next = flow;
162 /* limit number of collected flows per round */
164 #define FQ_GC_AGE (3*HZ)
166 static bool fq_gc_candidate(const struct fq_flow *f)
168 return fq_flow_is_detached(f) &&
169 time_after(jiffies, f->age + FQ_GC_AGE);
172 static void fq_gc(struct fq_sched_data *q,
173 struct rb_root *root,
176 struct fq_flow *f, *tofree[FQ_GC_MAX];
177 struct rb_node **p, *parent;
185 f = container_of(parent, struct fq_flow, fq_node);
189 if (fq_gc_candidate(f)) {
191 if (fcnt == FQ_GC_MAX)
196 p = &parent->rb_right;
198 p = &parent->rb_left;
202 q->inactive_flows -= fcnt;
203 q->stat_gc_flows += fcnt;
205 struct fq_flow *f = tofree[--fcnt];
207 rb_erase(&f->fq_node, root);
208 kmem_cache_free(fq_flow_cachep, f);
212 static const u8 prio2band[TC_PRIO_MAX + 1] = {
213 1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1
216 static struct fq_flow *fq_classify(struct sk_buff *skb, struct fq_sched_data *q)
218 struct rb_node **p, *parent;
219 struct sock *sk = skb->sk;
220 struct rb_root *root;
224 /* warning: no starvation prevention... */
225 band = prio2band[skb->priority & TC_PRIO_MAX];
226 if (unlikely(band == 0))
230 /* By forcing low order bit to 1, we make sure to not
231 * collide with a local flow (socket pointers are word aligned)
233 sk = (struct sock *)(skb_get_rxhash(skb) | 1L);
236 root = &q->fq_root[hash_32((u32)(long)sk, q->fq_trees_log)];
238 if (q->flows >= (2U << q->fq_trees_log) &&
239 q->inactive_flows > q->flows/2)
247 f = container_of(parent, struct fq_flow, fq_node);
249 /* socket might have been reallocated, so check
250 * if its sk_hash is the same.
251 * It not, we need to refill credit with
254 if (unlikely(skb->sk &&
255 f->socket_hash != sk->sk_hash)) {
256 f->credit = q->initial_quantum;
257 f->socket_hash = sk->sk_hash;
258 f->time_next_packet = 0ULL;
263 p = &parent->rb_right;
265 p = &parent->rb_left;
268 f = kmem_cache_zalloc(fq_flow_cachep, GFP_ATOMIC | __GFP_NOWARN);
270 q->stat_allocation_errors++;
273 fq_flow_set_detached(f);
276 f->socket_hash = sk->sk_hash;
277 f->credit = q->initial_quantum;
279 rb_link_node(&f->fq_node, parent, p);
280 rb_insert_color(&f->fq_node, root);
288 /* remove one skb from head of flow queue */
289 static struct sk_buff *fq_dequeue_head(struct Qdisc *sch, struct fq_flow *flow)
291 struct sk_buff *skb = flow->head;
294 flow->head = skb->next;
297 sch->qstats.backlog -= qdisc_pkt_len(skb);
303 /* We might add in the future detection of retransmits
304 * For the time being, just return false
306 static bool skb_is_retransmit(struct sk_buff *skb)
311 /* add skb to flow queue
312 * flow queue is a linked list, kind of FIFO, except for TCP retransmits
313 * We special case tcp retransmits to be transmitted before other packets.
314 * We rely on fact that TCP retransmits are unlikely, so we do not waste
315 * a separate queue or a pointer.
316 * head-> [retrans pkt 1]
321 * tail-> [ normal pkt 4]
323 static void flow_queue_add(struct fq_flow *flow, struct sk_buff *skb)
325 struct sk_buff *prev, *head = flow->head;
333 if (likely(!skb_is_retransmit(skb))) {
334 flow->tail->next = skb;
339 /* This skb is a tcp retransmit,
340 * find the last retrans packet in the queue
343 while (skb_is_retransmit(head)) {
349 if (!prev) { /* no rtx packet in queue, become the new head */
350 skb->next = flow->head;
353 if (prev == flow->tail)
356 skb->next = prev->next;
361 static int fq_enqueue(struct sk_buff *skb, struct Qdisc *sch)
363 struct fq_sched_data *q = qdisc_priv(sch);
366 if (unlikely(sch->q.qlen >= sch->limit))
367 return qdisc_drop(skb, sch);
369 f = fq_classify(skb, q);
370 if (unlikely(f->qlen >= q->flow_plimit && f != &q->internal)) {
371 q->stat_flows_plimit++;
372 return qdisc_drop(skb, sch);
376 flow_queue_add(f, skb);
377 if (skb_is_retransmit(skb))
378 q->stat_tcp_retrans++;
379 sch->qstats.backlog += qdisc_pkt_len(skb);
380 if (fq_flow_is_detached(f)) {
381 fq_flow_add_tail(&q->new_flows, f);
382 if (q->quantum > f->credit)
383 f->credit = q->quantum;
385 qdisc_unthrottled(sch);
387 if (unlikely(f == &q->internal)) {
388 q->stat_internal_packets++;
389 qdisc_unthrottled(sch);
393 return NET_XMIT_SUCCESS;
396 static void fq_check_throttled(struct fq_sched_data *q, u64 now)
400 if (q->time_next_delayed_flow > now)
403 q->time_next_delayed_flow = ~0ULL;
404 while ((p = rb_first(&q->delayed)) != NULL) {
405 struct fq_flow *f = container_of(p, struct fq_flow, rate_node);
407 if (f->time_next_packet > now) {
408 q->time_next_delayed_flow = f->time_next_packet;
411 rb_erase(p, &q->delayed);
412 q->throttled_flows--;
413 fq_flow_add_tail(&q->old_flows, f);
417 static struct sk_buff *fq_dequeue(struct Qdisc *sch)
419 struct fq_sched_data *q = qdisc_priv(sch);
420 u64 now = ktime_to_ns(ktime_get());
421 struct fq_flow_head *head;
426 skb = fq_dequeue_head(sch, &q->internal);
429 fq_check_throttled(q, now);
431 head = &q->new_flows;
433 head = &q->old_flows;
435 if (q->time_next_delayed_flow != ~0ULL)
436 qdisc_watchdog_schedule_ns(&q->watchdog,
437 q->time_next_delayed_flow);
443 if (f->credit <= 0) {
444 f->credit += q->quantum;
445 head->first = f->next;
446 fq_flow_add_tail(&q->old_flows, f);
450 if (unlikely(f->head && now < f->time_next_packet)) {
451 head->first = f->next;
452 fq_flow_set_throttled(q, f);
456 skb = fq_dequeue_head(sch, f);
458 head->first = f->next;
459 /* force a pass through old_flows to prevent starvation */
460 if ((head == &q->new_flows) && q->old_flows.first) {
461 fq_flow_add_tail(&q->old_flows, f);
463 fq_flow_set_detached(f);
470 f->time_next_packet = now;
471 f->credit -= qdisc_pkt_len(skb);
473 if (f->credit > 0 || !q->rate_enable)
476 rate = q->flow_max_rate;
477 if (skb->sk && skb->sk->sk_state != TCP_TIME_WAIT)
478 rate = min(skb->sk->sk_pacing_rate, rate);
481 u32 plen = max(qdisc_pkt_len(skb), q->quantum);
482 u64 len = (u64)plen * NSEC_PER_SEC;
486 /* Since socket rate can change later,
487 * clamp the delay to 125 ms.
488 * TODO: maybe segment the too big skb, as in commit
489 * e43ac79a4bc ("sch_tbf: segment too big GSO packets")
491 if (unlikely(len > 125 * NSEC_PER_MSEC)) {
492 len = 125 * NSEC_PER_MSEC;
493 q->stat_pkts_too_long++;
496 f->time_next_packet = now + len;
499 qdisc_bstats_update(sch, skb);
500 qdisc_unthrottled(sch);
504 static void fq_reset(struct Qdisc *sch)
506 struct fq_sched_data *q = qdisc_priv(sch);
507 struct rb_root *root;
513 while ((skb = fq_dequeue_head(sch, &q->internal)) != NULL)
519 for (idx = 0; idx < (1U << q->fq_trees_log); idx++) {
520 root = &q->fq_root[idx];
521 while ((p = rb_first(root)) != NULL) {
522 f = container_of(p, struct fq_flow, fq_node);
525 while ((skb = fq_dequeue_head(sch, f)) != NULL)
528 kmem_cache_free(fq_flow_cachep, f);
531 q->new_flows.first = NULL;
532 q->old_flows.first = NULL;
533 q->delayed = RB_ROOT;
535 q->inactive_flows = 0;
536 q->throttled_flows = 0;
539 static void fq_rehash(struct fq_sched_data *q,
540 struct rb_root *old_array, u32 old_log,
541 struct rb_root *new_array, u32 new_log)
543 struct rb_node *op, **np, *parent;
544 struct rb_root *oroot, *nroot;
545 struct fq_flow *of, *nf;
549 for (idx = 0; idx < (1U << old_log); idx++) {
550 oroot = &old_array[idx];
551 while ((op = rb_first(oroot)) != NULL) {
553 of = container_of(op, struct fq_flow, fq_node);
554 if (fq_gc_candidate(of)) {
556 kmem_cache_free(fq_flow_cachep, of);
559 nroot = &new_array[hash_32((u32)(long)of->sk, new_log)];
561 np = &nroot->rb_node;
566 nf = container_of(parent, struct fq_flow, fq_node);
567 BUG_ON(nf->sk == of->sk);
570 np = &parent->rb_right;
572 np = &parent->rb_left;
575 rb_link_node(&of->fq_node, parent, np);
576 rb_insert_color(&of->fq_node, nroot);
580 q->inactive_flows -= fcnt;
581 q->stat_gc_flows += fcnt;
584 static int fq_resize(struct fq_sched_data *q, u32 log)
586 struct rb_root *array;
589 if (q->fq_root && log == q->fq_trees_log)
592 array = kmalloc(sizeof(struct rb_root) << log, GFP_KERNEL);
596 for (idx = 0; idx < (1U << log); idx++)
597 array[idx] = RB_ROOT;
600 fq_rehash(q, q->fq_root, q->fq_trees_log, array, log);
604 q->fq_trees_log = log;
609 static const struct nla_policy fq_policy[TCA_FQ_MAX + 1] = {
610 [TCA_FQ_PLIMIT] = { .type = NLA_U32 },
611 [TCA_FQ_FLOW_PLIMIT] = { .type = NLA_U32 },
612 [TCA_FQ_QUANTUM] = { .type = NLA_U32 },
613 [TCA_FQ_INITIAL_QUANTUM] = { .type = NLA_U32 },
614 [TCA_FQ_RATE_ENABLE] = { .type = NLA_U32 },
615 [TCA_FQ_FLOW_DEFAULT_RATE] = { .type = NLA_U32 },
616 [TCA_FQ_FLOW_MAX_RATE] = { .type = NLA_U32 },
617 [TCA_FQ_BUCKETS_LOG] = { .type = NLA_U32 },
620 static int fq_change(struct Qdisc *sch, struct nlattr *opt)
622 struct fq_sched_data *q = qdisc_priv(sch);
623 struct nlattr *tb[TCA_FQ_MAX + 1];
624 int err, drop_count = 0;
630 err = nla_parse_nested(tb, TCA_FQ_MAX, opt, fq_policy);
636 fq_log = q->fq_trees_log;
638 if (tb[TCA_FQ_BUCKETS_LOG]) {
639 u32 nval = nla_get_u32(tb[TCA_FQ_BUCKETS_LOG]);
641 if (nval >= 1 && nval <= ilog2(256*1024))
646 if (tb[TCA_FQ_PLIMIT])
647 sch->limit = nla_get_u32(tb[TCA_FQ_PLIMIT]);
649 if (tb[TCA_FQ_FLOW_PLIMIT])
650 q->flow_plimit = nla_get_u32(tb[TCA_FQ_FLOW_PLIMIT]);
652 if (tb[TCA_FQ_QUANTUM])
653 q->quantum = nla_get_u32(tb[TCA_FQ_QUANTUM]);
655 if (tb[TCA_FQ_INITIAL_QUANTUM])
656 q->initial_quantum = nla_get_u32(tb[TCA_FQ_INITIAL_QUANTUM]);
658 if (tb[TCA_FQ_FLOW_DEFAULT_RATE])
659 q->flow_default_rate = nla_get_u32(tb[TCA_FQ_FLOW_DEFAULT_RATE]);
661 if (tb[TCA_FQ_FLOW_MAX_RATE])
662 q->flow_max_rate = nla_get_u32(tb[TCA_FQ_FLOW_MAX_RATE]);
664 if (tb[TCA_FQ_RATE_ENABLE]) {
665 u32 enable = nla_get_u32(tb[TCA_FQ_RATE_ENABLE]);
668 q->rate_enable = enable;
674 err = fq_resize(q, fq_log);
676 while (sch->q.qlen > sch->limit) {
677 struct sk_buff *skb = fq_dequeue(sch);
684 qdisc_tree_decrease_qlen(sch, drop_count);
686 sch_tree_unlock(sch);
690 static void fq_destroy(struct Qdisc *sch)
692 struct fq_sched_data *q = qdisc_priv(sch);
696 qdisc_watchdog_cancel(&q->watchdog);
699 static int fq_init(struct Qdisc *sch, struct nlattr *opt)
701 struct fq_sched_data *q = qdisc_priv(sch);
705 q->flow_plimit = 100;
706 q->quantum = 2 * psched_mtu(qdisc_dev(sch));
707 q->initial_quantum = 10 * psched_mtu(qdisc_dev(sch));
708 q->flow_default_rate = 0;
709 q->flow_max_rate = ~0U;
711 q->new_flows.first = NULL;
712 q->old_flows.first = NULL;
713 q->delayed = RB_ROOT;
715 q->fq_trees_log = ilog2(1024);
716 qdisc_watchdog_init(&q->watchdog, sch);
719 err = fq_change(sch, opt);
721 err = fq_resize(q, q->fq_trees_log);
726 static int fq_dump(struct Qdisc *sch, struct sk_buff *skb)
728 struct fq_sched_data *q = qdisc_priv(sch);
731 opts = nla_nest_start(skb, TCA_OPTIONS);
733 goto nla_put_failure;
735 /* TCA_FQ_FLOW_DEFAULT_RATE is not used anymore,
736 * do not bother giving its value
738 if (nla_put_u32(skb, TCA_FQ_PLIMIT, sch->limit) ||
739 nla_put_u32(skb, TCA_FQ_FLOW_PLIMIT, q->flow_plimit) ||
740 nla_put_u32(skb, TCA_FQ_QUANTUM, q->quantum) ||
741 nla_put_u32(skb, TCA_FQ_INITIAL_QUANTUM, q->initial_quantum) ||
742 nla_put_u32(skb, TCA_FQ_RATE_ENABLE, q->rate_enable) ||
743 nla_put_u32(skb, TCA_FQ_FLOW_MAX_RATE, q->flow_max_rate) ||
744 nla_put_u32(skb, TCA_FQ_BUCKETS_LOG, q->fq_trees_log))
745 goto nla_put_failure;
747 nla_nest_end(skb, opts);
754 static int fq_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
756 struct fq_sched_data *q = qdisc_priv(sch);
757 u64 now = ktime_to_ns(ktime_get());
758 struct tc_fq_qd_stats st = {
759 .gc_flows = q->stat_gc_flows,
760 .highprio_packets = q->stat_internal_packets,
761 .tcp_retrans = q->stat_tcp_retrans,
762 .throttled = q->stat_throttled,
763 .flows_plimit = q->stat_flows_plimit,
764 .pkts_too_long = q->stat_pkts_too_long,
765 .allocation_errors = q->stat_allocation_errors,
767 .inactive_flows = q->inactive_flows,
768 .throttled_flows = q->throttled_flows,
769 .time_next_delayed_flow = q->time_next_delayed_flow - now,
772 return gnet_stats_copy_app(d, &st, sizeof(st));
775 static struct Qdisc_ops fq_qdisc_ops __read_mostly = {
777 .priv_size = sizeof(struct fq_sched_data),
779 .enqueue = fq_enqueue,
780 .dequeue = fq_dequeue,
781 .peek = qdisc_peek_dequeued,
784 .destroy = fq_destroy,
787 .dump_stats = fq_dump_stats,
788 .owner = THIS_MODULE,
791 static int __init fq_module_init(void)
795 fq_flow_cachep = kmem_cache_create("fq_flow_cache",
796 sizeof(struct fq_flow),
801 ret = register_qdisc(&fq_qdisc_ops);
803 kmem_cache_destroy(fq_flow_cachep);
807 static void __exit fq_module_exit(void)
809 unregister_qdisc(&fq_qdisc_ops);
810 kmem_cache_destroy(fq_flow_cachep);
813 module_init(fq_module_init)
814 module_exit(fq_module_exit)
815 MODULE_AUTHOR("Eric Dumazet");
816 MODULE_LICENSE("GPL");