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 <linux/vmalloc.h>
51 #include <net/netlink.h>
52 #include <net/pkt_sched.h>
54 #include <net/tcp_states.h>
57 * Per flow structure, dynamically allocated
60 struct sk_buff *head; /* list of skbs for this flow : first skb */
62 struct sk_buff *tail; /* last skb in the list */
63 unsigned long age; /* jiffies when flow was emptied, for gc */
65 struct rb_node fq_node; /* anchor in fq_root[] trees */
67 int qlen; /* number of packets in flow queue */
69 u32 socket_hash; /* sk_hash */
70 struct fq_flow *next; /* next pointer in RR lists, or &detached */
72 struct rb_node rate_node; /* anchor in q->delayed tree */
77 struct fq_flow *first;
81 struct fq_sched_data {
82 struct fq_flow_head new_flows;
84 struct fq_flow_head old_flows;
86 struct rb_root delayed; /* for rate limited flows */
87 u64 time_next_delayed_flow;
89 struct fq_flow internal; /* for non classified or high prio packets */
92 u32 flow_refill_delay;
93 u32 flow_max_rate; /* optional max rate per flow */
94 u32 flow_plimit; /* max packets per flow */
95 struct rb_root *fq_root;
104 u64 stat_internal_packets;
105 u64 stat_tcp_retrans;
107 u64 stat_flows_plimit;
108 u64 stat_pkts_too_long;
109 u64 stat_allocation_errors;
110 struct qdisc_watchdog watchdog;
113 /* special value to mark a detached flow (not on old/new list) */
114 static struct fq_flow detached, throttled;
116 static void fq_flow_set_detached(struct fq_flow *f)
122 static bool fq_flow_is_detached(const struct fq_flow *f)
124 return f->next == &detached;
127 static void fq_flow_set_throttled(struct fq_sched_data *q, struct fq_flow *f)
129 struct rb_node **p = &q->delayed.rb_node, *parent = NULL;
135 aux = container_of(parent, struct fq_flow, rate_node);
136 if (f->time_next_packet >= aux->time_next_packet)
137 p = &parent->rb_right;
139 p = &parent->rb_left;
141 rb_link_node(&f->rate_node, parent, p);
142 rb_insert_color(&f->rate_node, &q->delayed);
143 q->throttled_flows++;
146 f->next = &throttled;
147 if (q->time_next_delayed_flow > f->time_next_packet)
148 q->time_next_delayed_flow = f->time_next_packet;
152 static struct kmem_cache *fq_flow_cachep __read_mostly;
154 static void fq_flow_add_tail(struct fq_flow_head *head, struct fq_flow *flow)
157 head->last->next = flow;
164 /* limit number of collected flows per round */
166 #define FQ_GC_AGE (3*HZ)
168 static bool fq_gc_candidate(const struct fq_flow *f)
170 return fq_flow_is_detached(f) &&
171 time_after(jiffies, f->age + FQ_GC_AGE);
174 static void fq_gc(struct fq_sched_data *q,
175 struct rb_root *root,
178 struct fq_flow *f, *tofree[FQ_GC_MAX];
179 struct rb_node **p, *parent;
187 f = container_of(parent, struct fq_flow, fq_node);
191 if (fq_gc_candidate(f)) {
193 if (fcnt == FQ_GC_MAX)
198 p = &parent->rb_right;
200 p = &parent->rb_left;
204 q->inactive_flows -= fcnt;
205 q->stat_gc_flows += fcnt;
207 struct fq_flow *f = tofree[--fcnt];
209 rb_erase(&f->fq_node, root);
210 kmem_cache_free(fq_flow_cachep, f);
214 static struct fq_flow *fq_classify(struct sk_buff *skb, struct fq_sched_data *q)
216 struct rb_node **p, *parent;
217 struct sock *sk = skb->sk;
218 struct rb_root *root;
221 /* warning: no starvation prevention... */
222 if (unlikely((skb->priority & TC_PRIO_MAX) == TC_PRIO_CONTROL))
226 /* By forcing low order bit to 1, we make sure to not
227 * collide with a local flow (socket pointers are word aligned)
229 sk = (struct sock *)(skb_get_hash(skb) | 1L);
232 root = &q->fq_root[hash_32((u32)(long)sk, q->fq_trees_log)];
234 if (q->flows >= (2U << q->fq_trees_log) &&
235 q->inactive_flows > q->flows/2)
243 f = container_of(parent, struct fq_flow, fq_node);
245 /* socket might have been reallocated, so check
246 * if its sk_hash is the same.
247 * It not, we need to refill credit with
250 if (unlikely(skb->sk &&
251 f->socket_hash != sk->sk_hash)) {
252 f->credit = q->initial_quantum;
253 f->socket_hash = sk->sk_hash;
254 f->time_next_packet = 0ULL;
259 p = &parent->rb_right;
261 p = &parent->rb_left;
264 f = kmem_cache_zalloc(fq_flow_cachep, GFP_ATOMIC | __GFP_NOWARN);
266 q->stat_allocation_errors++;
269 fq_flow_set_detached(f);
272 f->socket_hash = sk->sk_hash;
273 f->credit = q->initial_quantum;
275 rb_link_node(&f->fq_node, parent, p);
276 rb_insert_color(&f->fq_node, root);
284 /* remove one skb from head of flow queue */
285 static struct sk_buff *fq_dequeue_head(struct Qdisc *sch, struct fq_flow *flow)
287 struct sk_buff *skb = flow->head;
290 flow->head = skb->next;
293 sch->qstats.backlog -= qdisc_pkt_len(skb);
299 /* We might add in the future detection of retransmits
300 * For the time being, just return false
302 static bool skb_is_retransmit(struct sk_buff *skb)
307 /* add skb to flow queue
308 * flow queue is a linked list, kind of FIFO, except for TCP retransmits
309 * We special case tcp retransmits to be transmitted before other packets.
310 * We rely on fact that TCP retransmits are unlikely, so we do not waste
311 * a separate queue or a pointer.
312 * head-> [retrans pkt 1]
317 * tail-> [ normal pkt 4]
319 static void flow_queue_add(struct fq_flow *flow, struct sk_buff *skb)
321 struct sk_buff *prev, *head = flow->head;
329 if (likely(!skb_is_retransmit(skb))) {
330 flow->tail->next = skb;
335 /* This skb is a tcp retransmit,
336 * find the last retrans packet in the queue
339 while (skb_is_retransmit(head)) {
345 if (!prev) { /* no rtx packet in queue, become the new head */
346 skb->next = flow->head;
349 if (prev == flow->tail)
352 skb->next = prev->next;
357 static int fq_enqueue(struct sk_buff *skb, struct Qdisc *sch)
359 struct fq_sched_data *q = qdisc_priv(sch);
362 if (unlikely(sch->q.qlen >= sch->limit))
363 return qdisc_drop(skb, sch);
365 f = fq_classify(skb, q);
366 if (unlikely(f->qlen >= q->flow_plimit && f != &q->internal)) {
367 q->stat_flows_plimit++;
368 return qdisc_drop(skb, sch);
372 if (skb_is_retransmit(skb))
373 q->stat_tcp_retrans++;
374 sch->qstats.backlog += qdisc_pkt_len(skb);
375 if (fq_flow_is_detached(f)) {
376 fq_flow_add_tail(&q->new_flows, f);
377 if (time_after(jiffies, f->age + q->flow_refill_delay))
378 f->credit = max_t(u32, f->credit, q->quantum);
380 qdisc_unthrottled(sch);
383 /* Note: this overwrites f->age */
384 flow_queue_add(f, skb);
386 if (unlikely(f == &q->internal)) {
387 q->stat_internal_packets++;
388 qdisc_unthrottled(sch);
392 return NET_XMIT_SUCCESS;
395 static void fq_check_throttled(struct fq_sched_data *q, u64 now)
399 if (q->time_next_delayed_flow > now)
402 q->time_next_delayed_flow = ~0ULL;
403 while ((p = rb_first(&q->delayed)) != NULL) {
404 struct fq_flow *f = container_of(p, struct fq_flow, rate_node);
406 if (f->time_next_packet > now) {
407 q->time_next_delayed_flow = f->time_next_packet;
410 rb_erase(p, &q->delayed);
411 q->throttled_flows--;
412 fq_flow_add_tail(&q->old_flows, f);
416 static struct sk_buff *fq_dequeue(struct Qdisc *sch)
418 struct fq_sched_data *q = qdisc_priv(sch);
419 u64 now = ktime_to_ns(ktime_get());
420 struct fq_flow_head *head;
425 skb = fq_dequeue_head(sch, &q->internal);
428 fq_check_throttled(q, now);
430 head = &q->new_flows;
432 head = &q->old_flows;
434 if (q->time_next_delayed_flow != ~0ULL)
435 qdisc_watchdog_schedule_ns(&q->watchdog,
436 q->time_next_delayed_flow);
442 if (f->credit <= 0) {
443 f->credit += q->quantum;
444 head->first = f->next;
445 fq_flow_add_tail(&q->old_flows, f);
449 if (unlikely(f->head && now < f->time_next_packet)) {
450 head->first = f->next;
451 fq_flow_set_throttled(q, f);
455 skb = fq_dequeue_head(sch, f);
457 head->first = f->next;
458 /* force a pass through old_flows to prevent starvation */
459 if ((head == &q->new_flows) && q->old_flows.first) {
460 fq_flow_add_tail(&q->old_flows, f);
462 fq_flow_set_detached(f);
468 f->time_next_packet = now;
469 f->credit -= qdisc_pkt_len(skb);
471 if (f->credit > 0 || !q->rate_enable)
474 rate = q->flow_max_rate;
475 if (skb->sk && skb->sk->sk_state != TCP_TIME_WAIT)
476 rate = min(skb->sk->sk_pacing_rate, rate);
479 u32 plen = max(qdisc_pkt_len(skb), q->quantum);
480 u64 len = (u64)plen * NSEC_PER_SEC;
484 /* Since socket rate can change later,
485 * clamp the delay to 125 ms.
486 * TODO: maybe segment the too big skb, as in commit
487 * e43ac79a4bc ("sch_tbf: segment too big GSO packets")
489 if (unlikely(len > 125 * NSEC_PER_MSEC)) {
490 len = 125 * NSEC_PER_MSEC;
491 q->stat_pkts_too_long++;
494 f->time_next_packet = now + len;
497 qdisc_bstats_update(sch, skb);
498 qdisc_unthrottled(sch);
502 static void fq_reset(struct Qdisc *sch)
504 struct fq_sched_data *q = qdisc_priv(sch);
505 struct rb_root *root;
511 while ((skb = fq_dequeue_head(sch, &q->internal)) != NULL)
517 for (idx = 0; idx < (1U << q->fq_trees_log); idx++) {
518 root = &q->fq_root[idx];
519 while ((p = rb_first(root)) != NULL) {
520 f = container_of(p, struct fq_flow, fq_node);
523 while ((skb = fq_dequeue_head(sch, f)) != NULL)
526 kmem_cache_free(fq_flow_cachep, f);
529 q->new_flows.first = NULL;
530 q->old_flows.first = NULL;
531 q->delayed = RB_ROOT;
533 q->inactive_flows = 0;
534 q->throttled_flows = 0;
537 static void fq_rehash(struct fq_sched_data *q,
538 struct rb_root *old_array, u32 old_log,
539 struct rb_root *new_array, u32 new_log)
541 struct rb_node *op, **np, *parent;
542 struct rb_root *oroot, *nroot;
543 struct fq_flow *of, *nf;
547 for (idx = 0; idx < (1U << old_log); idx++) {
548 oroot = &old_array[idx];
549 while ((op = rb_first(oroot)) != NULL) {
551 of = container_of(op, struct fq_flow, fq_node);
552 if (fq_gc_candidate(of)) {
554 kmem_cache_free(fq_flow_cachep, of);
557 nroot = &new_array[hash_32((u32)(long)of->sk, new_log)];
559 np = &nroot->rb_node;
564 nf = container_of(parent, struct fq_flow, fq_node);
565 BUG_ON(nf->sk == of->sk);
568 np = &parent->rb_right;
570 np = &parent->rb_left;
573 rb_link_node(&of->fq_node, parent, np);
574 rb_insert_color(&of->fq_node, nroot);
578 q->inactive_flows -= fcnt;
579 q->stat_gc_flows += fcnt;
582 static void *fq_alloc_node(size_t sz, int node)
586 ptr = kmalloc_node(sz, GFP_KERNEL | __GFP_REPEAT | __GFP_NOWARN, node);
588 ptr = vmalloc_node(sz, node);
592 static void fq_free(void *addr)
594 if (addr && is_vmalloc_addr(addr))
600 static int fq_resize(struct Qdisc *sch, u32 log)
602 struct fq_sched_data *q = qdisc_priv(sch);
603 struct rb_root *array;
607 if (q->fq_root && log == q->fq_trees_log)
610 /* If XPS was setup, we can allocate memory on right NUMA node */
611 array = fq_alloc_node(sizeof(struct rb_root) << log,
612 netdev_queue_numa_node_read(sch->dev_queue));
616 for (idx = 0; idx < (1U << log); idx++)
617 array[idx] = RB_ROOT;
621 old_fq_root = q->fq_root;
623 fq_rehash(q, old_fq_root, q->fq_trees_log, array, log);
626 q->fq_trees_log = log;
628 sch_tree_unlock(sch);
630 fq_free(old_fq_root);
635 static const struct nla_policy fq_policy[TCA_FQ_MAX + 1] = {
636 [TCA_FQ_PLIMIT] = { .type = NLA_U32 },
637 [TCA_FQ_FLOW_PLIMIT] = { .type = NLA_U32 },
638 [TCA_FQ_QUANTUM] = { .type = NLA_U32 },
639 [TCA_FQ_INITIAL_QUANTUM] = { .type = NLA_U32 },
640 [TCA_FQ_RATE_ENABLE] = { .type = NLA_U32 },
641 [TCA_FQ_FLOW_DEFAULT_RATE] = { .type = NLA_U32 },
642 [TCA_FQ_FLOW_MAX_RATE] = { .type = NLA_U32 },
643 [TCA_FQ_BUCKETS_LOG] = { .type = NLA_U32 },
644 [TCA_FQ_FLOW_REFILL_DELAY] = { .type = NLA_U32 },
647 static int fq_change(struct Qdisc *sch, struct nlattr *opt)
649 struct fq_sched_data *q = qdisc_priv(sch);
650 struct nlattr *tb[TCA_FQ_MAX + 1];
651 int err, drop_count = 0;
657 err = nla_parse_nested(tb, TCA_FQ_MAX, opt, fq_policy);
663 fq_log = q->fq_trees_log;
665 if (tb[TCA_FQ_BUCKETS_LOG]) {
666 u32 nval = nla_get_u32(tb[TCA_FQ_BUCKETS_LOG]);
668 if (nval >= 1 && nval <= ilog2(256*1024))
673 if (tb[TCA_FQ_PLIMIT])
674 sch->limit = nla_get_u32(tb[TCA_FQ_PLIMIT]);
676 if (tb[TCA_FQ_FLOW_PLIMIT])
677 q->flow_plimit = nla_get_u32(tb[TCA_FQ_FLOW_PLIMIT]);
679 if (tb[TCA_FQ_QUANTUM])
680 q->quantum = nla_get_u32(tb[TCA_FQ_QUANTUM]);
682 if (tb[TCA_FQ_INITIAL_QUANTUM])
683 q->initial_quantum = nla_get_u32(tb[TCA_FQ_INITIAL_QUANTUM]);
685 if (tb[TCA_FQ_FLOW_DEFAULT_RATE])
686 pr_warn_ratelimited("sch_fq: defrate %u ignored.\n",
687 nla_get_u32(tb[TCA_FQ_FLOW_DEFAULT_RATE]));
689 if (tb[TCA_FQ_FLOW_MAX_RATE])
690 q->flow_max_rate = nla_get_u32(tb[TCA_FQ_FLOW_MAX_RATE]);
692 if (tb[TCA_FQ_RATE_ENABLE]) {
693 u32 enable = nla_get_u32(tb[TCA_FQ_RATE_ENABLE]);
696 q->rate_enable = enable;
701 if (tb[TCA_FQ_FLOW_REFILL_DELAY]) {
702 u32 usecs_delay = nla_get_u32(tb[TCA_FQ_FLOW_REFILL_DELAY]) ;
704 q->flow_refill_delay = usecs_to_jiffies(usecs_delay);
708 sch_tree_unlock(sch);
709 err = fq_resize(sch, fq_log);
712 while (sch->q.qlen > sch->limit) {
713 struct sk_buff *skb = fq_dequeue(sch);
720 qdisc_tree_decrease_qlen(sch, drop_count);
722 sch_tree_unlock(sch);
726 static void fq_destroy(struct Qdisc *sch)
728 struct fq_sched_data *q = qdisc_priv(sch);
732 qdisc_watchdog_cancel(&q->watchdog);
735 static int fq_init(struct Qdisc *sch, struct nlattr *opt)
737 struct fq_sched_data *q = qdisc_priv(sch);
741 q->flow_plimit = 100;
742 q->quantum = 2 * psched_mtu(qdisc_dev(sch));
743 q->initial_quantum = 10 * psched_mtu(qdisc_dev(sch));
744 q->flow_refill_delay = msecs_to_jiffies(40);
745 q->flow_max_rate = ~0U;
747 q->new_flows.first = NULL;
748 q->old_flows.first = NULL;
749 q->delayed = RB_ROOT;
751 q->fq_trees_log = ilog2(1024);
752 qdisc_watchdog_init(&q->watchdog, sch);
755 err = fq_change(sch, opt);
757 err = fq_resize(sch, q->fq_trees_log);
762 static int fq_dump(struct Qdisc *sch, struct sk_buff *skb)
764 struct fq_sched_data *q = qdisc_priv(sch);
767 opts = nla_nest_start(skb, TCA_OPTIONS);
769 goto nla_put_failure;
771 /* TCA_FQ_FLOW_DEFAULT_RATE is not used anymore */
773 if (nla_put_u32(skb, TCA_FQ_PLIMIT, sch->limit) ||
774 nla_put_u32(skb, TCA_FQ_FLOW_PLIMIT, q->flow_plimit) ||
775 nla_put_u32(skb, TCA_FQ_QUANTUM, q->quantum) ||
776 nla_put_u32(skb, TCA_FQ_INITIAL_QUANTUM, q->initial_quantum) ||
777 nla_put_u32(skb, TCA_FQ_RATE_ENABLE, q->rate_enable) ||
778 nla_put_u32(skb, TCA_FQ_FLOW_MAX_RATE, q->flow_max_rate) ||
779 nla_put_u32(skb, TCA_FQ_FLOW_REFILL_DELAY,
780 jiffies_to_usecs(q->flow_refill_delay)) ||
781 nla_put_u32(skb, TCA_FQ_BUCKETS_LOG, q->fq_trees_log))
782 goto nla_put_failure;
784 return nla_nest_end(skb, opts);
790 static int fq_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
792 struct fq_sched_data *q = qdisc_priv(sch);
793 u64 now = ktime_to_ns(ktime_get());
794 struct tc_fq_qd_stats st = {
795 .gc_flows = q->stat_gc_flows,
796 .highprio_packets = q->stat_internal_packets,
797 .tcp_retrans = q->stat_tcp_retrans,
798 .throttled = q->stat_throttled,
799 .flows_plimit = q->stat_flows_plimit,
800 .pkts_too_long = q->stat_pkts_too_long,
801 .allocation_errors = q->stat_allocation_errors,
803 .inactive_flows = q->inactive_flows,
804 .throttled_flows = q->throttled_flows,
805 .time_next_delayed_flow = q->time_next_delayed_flow - now,
808 return gnet_stats_copy_app(d, &st, sizeof(st));
811 static struct Qdisc_ops fq_qdisc_ops __read_mostly = {
813 .priv_size = sizeof(struct fq_sched_data),
815 .enqueue = fq_enqueue,
816 .dequeue = fq_dequeue,
817 .peek = qdisc_peek_dequeued,
820 .destroy = fq_destroy,
823 .dump_stats = fq_dump_stats,
824 .owner = THIS_MODULE,
827 static int __init fq_module_init(void)
831 fq_flow_cachep = kmem_cache_create("fq_flow_cache",
832 sizeof(struct fq_flow),
837 ret = register_qdisc(&fq_qdisc_ops);
839 kmem_cache_destroy(fq_flow_cachep);
843 static void __exit fq_module_exit(void)
845 unregister_qdisc(&fq_qdisc_ops);
846 kmem_cache_destroy(fq_flow_cachep);
849 module_init(fq_module_init)
850 module_exit(fq_module_exit)
851 MODULE_AUTHOR("Eric Dumazet");
852 MODULE_LICENSE("GPL");