2 * net/sched/sch_fq.c Fair Queue Packet Scheduler (per flow pacing)
4 * Copyright (C) 2013-2015 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 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'.
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>
58 * Per flow structure, dynamically allocated
61 struct sk_buff *head; /* list of skbs for this flow : first skb */
63 struct sk_buff *tail; /* last skb in the list */
64 unsigned long age; /* jiffies when flow was emptied, for gc */
66 struct rb_node fq_node; /* anchor in fq_root[] trees */
68 int qlen; /* number of packets in flow queue */
70 u32 socket_hash; /* sk_hash */
71 struct fq_flow *next; /* next pointer in RR lists, or &detached */
73 struct rb_node rate_node; /* anchor in q->delayed tree */
78 struct fq_flow *first;
82 struct fq_sched_data {
83 struct fq_flow_head new_flows;
85 struct fq_flow_head old_flows;
87 struct rb_root delayed; /* for rate limited flows */
88 u64 time_next_delayed_flow;
90 struct fq_flow internal; /* for non classified or high prio packets */
93 u32 flow_refill_delay;
94 u32 flow_max_rate; /* optional max rate per flow */
95 u32 flow_plimit; /* max packets per flow */
96 u32 orphan_mask; /* mask for orphaned skb */
97 struct rb_root *fq_root;
106 u64 stat_internal_packets;
107 u64 stat_tcp_retrans;
109 u64 stat_flows_plimit;
110 u64 stat_pkts_too_long;
111 u64 stat_allocation_errors;
112 struct qdisc_watchdog watchdog;
115 /* special value to mark a detached flow (not on old/new list) */
116 static struct fq_flow detached, throttled;
118 static void fq_flow_set_detached(struct fq_flow *f)
124 static bool fq_flow_is_detached(const struct fq_flow *f)
126 return f->next == &detached;
129 static bool fq_flow_is_throttled(const struct fq_flow *f)
131 return f->next == &throttled;
134 static void fq_flow_add_tail(struct fq_flow_head *head, struct fq_flow *flow)
137 head->last->next = flow;
144 static void fq_flow_unset_throttled(struct fq_sched_data *q, struct fq_flow *f)
146 rb_erase(&f->rate_node, &q->delayed);
147 q->throttled_flows--;
148 fq_flow_add_tail(&q->old_flows, f);
151 static void fq_flow_set_throttled(struct fq_sched_data *q, struct fq_flow *f)
153 struct rb_node **p = &q->delayed.rb_node, *parent = NULL;
159 aux = container_of(parent, struct fq_flow, rate_node);
160 if (f->time_next_packet >= aux->time_next_packet)
161 p = &parent->rb_right;
163 p = &parent->rb_left;
165 rb_link_node(&f->rate_node, parent, p);
166 rb_insert_color(&f->rate_node, &q->delayed);
167 q->throttled_flows++;
170 f->next = &throttled;
171 if (q->time_next_delayed_flow > f->time_next_packet)
172 q->time_next_delayed_flow = f->time_next_packet;
176 static struct kmem_cache *fq_flow_cachep __read_mostly;
179 /* limit number of collected flows per round */
181 #define FQ_GC_AGE (3*HZ)
183 static bool fq_gc_candidate(const struct fq_flow *f)
185 return fq_flow_is_detached(f) &&
186 time_after(jiffies, f->age + FQ_GC_AGE);
189 static void fq_gc(struct fq_sched_data *q,
190 struct rb_root *root,
193 struct fq_flow *f, *tofree[FQ_GC_MAX];
194 struct rb_node **p, *parent;
202 f = container_of(parent, struct fq_flow, fq_node);
206 if (fq_gc_candidate(f)) {
208 if (fcnt == FQ_GC_MAX)
213 p = &parent->rb_right;
215 p = &parent->rb_left;
219 q->inactive_flows -= fcnt;
220 q->stat_gc_flows += fcnt;
222 struct fq_flow *f = tofree[--fcnt];
224 rb_erase(&f->fq_node, root);
225 kmem_cache_free(fq_flow_cachep, f);
229 static struct fq_flow *fq_classify(struct sk_buff *skb, struct fq_sched_data *q)
231 struct rb_node **p, *parent;
232 struct sock *sk = skb->sk;
233 struct rb_root *root;
236 /* warning: no starvation prevention... */
237 if (unlikely((skb->priority & TC_PRIO_MAX) == TC_PRIO_CONTROL))
240 /* SYNACK messages are attached to a TCP_NEW_SYN_RECV request socket
241 * or a listener (SYNCOOKIE mode)
242 * 1) request sockets are not full blown,
243 * they do not contain sk_pacing_rate
244 * 2) They are not part of a 'flow' yet
245 * 3) We do not want to rate limit them (eg SYNFLOOD attack),
246 * especially if the listener set SO_MAX_PACING_RATE
247 * 4) We pretend they are orphaned
249 if (!sk || sk_listener(sk)) {
250 unsigned long hash = skb_get_hash(skb) & q->orphan_mask;
252 /* By forcing low order bit to 1, we make sure to not
253 * collide with a local flow (socket pointers are word aligned)
255 sk = (struct sock *)((hash << 1) | 1UL);
259 root = &q->fq_root[hash_32((u32)(long)sk, q->fq_trees_log)];
261 if (q->flows >= (2U << q->fq_trees_log) &&
262 q->inactive_flows > q->flows/2)
270 f = container_of(parent, struct fq_flow, fq_node);
272 /* socket might have been reallocated, so check
273 * if its sk_hash is the same.
274 * It not, we need to refill credit with
277 if (unlikely(skb->sk &&
278 f->socket_hash != sk->sk_hash)) {
279 f->credit = q->initial_quantum;
280 f->socket_hash = sk->sk_hash;
281 if (fq_flow_is_throttled(f))
282 fq_flow_unset_throttled(q, f);
283 f->time_next_packet = 0ULL;
288 p = &parent->rb_right;
290 p = &parent->rb_left;
293 f = kmem_cache_zalloc(fq_flow_cachep, GFP_ATOMIC | __GFP_NOWARN);
295 q->stat_allocation_errors++;
298 fq_flow_set_detached(f);
301 f->socket_hash = sk->sk_hash;
302 f->credit = q->initial_quantum;
304 rb_link_node(&f->fq_node, parent, p);
305 rb_insert_color(&f->fq_node, root);
313 /* remove one skb from head of flow queue */
314 static struct sk_buff *fq_dequeue_head(struct Qdisc *sch, struct fq_flow *flow)
316 struct sk_buff *skb = flow->head;
319 flow->head = skb->next;
322 qdisc_qstats_backlog_dec(sch, skb);
328 /* We might add in the future detection of retransmits
329 * For the time being, just return false
331 static bool skb_is_retransmit(struct sk_buff *skb)
336 /* add skb to flow queue
337 * flow queue is a linked list, kind of FIFO, except for TCP retransmits
338 * We special case tcp retransmits to be transmitted before other packets.
339 * We rely on fact that TCP retransmits are unlikely, so we do not waste
340 * a separate queue or a pointer.
341 * head-> [retrans pkt 1]
346 * tail-> [ normal pkt 4]
348 static void flow_queue_add(struct fq_flow *flow, struct sk_buff *skb)
350 struct sk_buff *prev, *head = flow->head;
358 if (likely(!skb_is_retransmit(skb))) {
359 flow->tail->next = skb;
364 /* This skb is a tcp retransmit,
365 * find the last retrans packet in the queue
368 while (skb_is_retransmit(head)) {
374 if (!prev) { /* no rtx packet in queue, become the new head */
375 skb->next = flow->head;
378 if (prev == flow->tail)
381 skb->next = prev->next;
386 static int fq_enqueue(struct sk_buff *skb, struct Qdisc *sch)
388 struct fq_sched_data *q = qdisc_priv(sch);
391 if (unlikely(sch->q.qlen >= sch->limit))
392 return qdisc_drop(skb, sch);
394 f = fq_classify(skb, q);
395 if (unlikely(f->qlen >= q->flow_plimit && f != &q->internal)) {
396 q->stat_flows_plimit++;
397 return qdisc_drop(skb, sch);
401 if (skb_is_retransmit(skb))
402 q->stat_tcp_retrans++;
403 qdisc_qstats_backlog_inc(sch, skb);
404 if (fq_flow_is_detached(f)) {
405 fq_flow_add_tail(&q->new_flows, f);
406 if (time_after(jiffies, f->age + q->flow_refill_delay))
407 f->credit = max_t(u32, f->credit, q->quantum);
411 /* Note: this overwrites f->age */
412 flow_queue_add(f, skb);
414 if (unlikely(f == &q->internal)) {
415 q->stat_internal_packets++;
419 return NET_XMIT_SUCCESS;
422 static void fq_check_throttled(struct fq_sched_data *q, u64 now)
426 if (q->time_next_delayed_flow > now)
429 q->time_next_delayed_flow = ~0ULL;
430 while ((p = rb_first(&q->delayed)) != NULL) {
431 struct fq_flow *f = container_of(p, struct fq_flow, rate_node);
433 if (f->time_next_packet > now) {
434 q->time_next_delayed_flow = f->time_next_packet;
437 fq_flow_unset_throttled(q, f);
441 static struct sk_buff *fq_dequeue(struct Qdisc *sch)
443 struct fq_sched_data *q = qdisc_priv(sch);
444 u64 now = ktime_get_ns();
445 struct fq_flow_head *head;
450 skb = fq_dequeue_head(sch, &q->internal);
453 fq_check_throttled(q, now);
455 head = &q->new_flows;
457 head = &q->old_flows;
459 if (q->time_next_delayed_flow != ~0ULL)
460 qdisc_watchdog_schedule_ns(&q->watchdog,
461 q->time_next_delayed_flow,
468 if (f->credit <= 0) {
469 f->credit += q->quantum;
470 head->first = f->next;
471 fq_flow_add_tail(&q->old_flows, f);
476 if (unlikely(skb && now < f->time_next_packet &&
477 !skb_is_tcp_pure_ack(skb))) {
478 head->first = f->next;
479 fq_flow_set_throttled(q, f);
483 skb = fq_dequeue_head(sch, f);
485 head->first = f->next;
486 /* force a pass through old_flows to prevent starvation */
487 if ((head == &q->new_flows) && q->old_flows.first) {
488 fq_flow_add_tail(&q->old_flows, f);
490 fq_flow_set_detached(f);
496 f->credit -= qdisc_pkt_len(skb);
498 if (f->credit > 0 || !q->rate_enable)
501 /* Do not pace locally generated ack packets */
502 if (skb_is_tcp_pure_ack(skb))
505 rate = q->flow_max_rate;
507 rate = min(skb->sk->sk_pacing_rate, rate);
510 u32 plen = max(qdisc_pkt_len(skb), q->quantum);
511 u64 len = (u64)plen * NSEC_PER_SEC;
515 /* Since socket rate can change later,
516 * clamp the delay to 1 second.
517 * Really, providers of too big packets should be fixed !
519 if (unlikely(len > NSEC_PER_SEC)) {
521 q->stat_pkts_too_long++;
524 f->time_next_packet = now + len;
527 qdisc_bstats_update(sch, skb);
531 static void fq_reset(struct Qdisc *sch)
533 struct fq_sched_data *q = qdisc_priv(sch);
534 struct rb_root *root;
540 while ((skb = fq_dequeue_head(sch, &q->internal)) != NULL)
546 for (idx = 0; idx < (1U << q->fq_trees_log); idx++) {
547 root = &q->fq_root[idx];
548 while ((p = rb_first(root)) != NULL) {
549 f = container_of(p, struct fq_flow, fq_node);
552 while ((skb = fq_dequeue_head(sch, f)) != NULL)
555 kmem_cache_free(fq_flow_cachep, f);
558 q->new_flows.first = NULL;
559 q->old_flows.first = NULL;
560 q->delayed = RB_ROOT;
562 q->inactive_flows = 0;
563 q->throttled_flows = 0;
566 static void fq_rehash(struct fq_sched_data *q,
567 struct rb_root *old_array, u32 old_log,
568 struct rb_root *new_array, u32 new_log)
570 struct rb_node *op, **np, *parent;
571 struct rb_root *oroot, *nroot;
572 struct fq_flow *of, *nf;
576 for (idx = 0; idx < (1U << old_log); idx++) {
577 oroot = &old_array[idx];
578 while ((op = rb_first(oroot)) != NULL) {
580 of = container_of(op, struct fq_flow, fq_node);
581 if (fq_gc_candidate(of)) {
583 kmem_cache_free(fq_flow_cachep, of);
586 nroot = &new_array[hash_32((u32)(long)of->sk, new_log)];
588 np = &nroot->rb_node;
593 nf = container_of(parent, struct fq_flow, fq_node);
594 BUG_ON(nf->sk == of->sk);
597 np = &parent->rb_right;
599 np = &parent->rb_left;
602 rb_link_node(&of->fq_node, parent, np);
603 rb_insert_color(&of->fq_node, nroot);
607 q->inactive_flows -= fcnt;
608 q->stat_gc_flows += fcnt;
611 static void *fq_alloc_node(size_t sz, int node)
615 ptr = kmalloc_node(sz, GFP_KERNEL | __GFP_REPEAT | __GFP_NOWARN, node);
617 ptr = vmalloc_node(sz, node);
621 static void fq_free(void *addr)
626 static int fq_resize(struct Qdisc *sch, u32 log)
628 struct fq_sched_data *q = qdisc_priv(sch);
629 struct rb_root *array;
633 if (q->fq_root && log == q->fq_trees_log)
636 /* If XPS was setup, we can allocate memory on right NUMA node */
637 array = fq_alloc_node(sizeof(struct rb_root) << log,
638 netdev_queue_numa_node_read(sch->dev_queue));
642 for (idx = 0; idx < (1U << log); idx++)
643 array[idx] = RB_ROOT;
647 old_fq_root = q->fq_root;
649 fq_rehash(q, old_fq_root, q->fq_trees_log, array, log);
652 q->fq_trees_log = log;
654 sch_tree_unlock(sch);
656 fq_free(old_fq_root);
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_ORPHAN_MASK] = { .type = NLA_U32 },
674 static int fq_change(struct Qdisc *sch, struct nlattr *opt)
676 struct fq_sched_data *q = qdisc_priv(sch);
677 struct nlattr *tb[TCA_FQ_MAX + 1];
678 int err, drop_count = 0;
679 unsigned drop_len = 0;
685 err = nla_parse_nested(tb, TCA_FQ_MAX, opt, fq_policy);
691 fq_log = q->fq_trees_log;
693 if (tb[TCA_FQ_BUCKETS_LOG]) {
694 u32 nval = nla_get_u32(tb[TCA_FQ_BUCKETS_LOG]);
696 if (nval >= 1 && nval <= ilog2(256*1024))
701 if (tb[TCA_FQ_PLIMIT])
702 sch->limit = nla_get_u32(tb[TCA_FQ_PLIMIT]);
704 if (tb[TCA_FQ_FLOW_PLIMIT])
705 q->flow_plimit = nla_get_u32(tb[TCA_FQ_FLOW_PLIMIT]);
707 if (tb[TCA_FQ_QUANTUM]) {
708 u32 quantum = nla_get_u32(tb[TCA_FQ_QUANTUM]);
710 if (quantum > 0 && quantum <= (1 << 20))
711 q->quantum = quantum;
716 if (tb[TCA_FQ_INITIAL_QUANTUM])
717 q->initial_quantum = nla_get_u32(tb[TCA_FQ_INITIAL_QUANTUM]);
719 if (tb[TCA_FQ_FLOW_DEFAULT_RATE])
720 pr_warn_ratelimited("sch_fq: defrate %u ignored.\n",
721 nla_get_u32(tb[TCA_FQ_FLOW_DEFAULT_RATE]));
723 if (tb[TCA_FQ_FLOW_MAX_RATE])
724 q->flow_max_rate = nla_get_u32(tb[TCA_FQ_FLOW_MAX_RATE]);
726 if (tb[TCA_FQ_RATE_ENABLE]) {
727 u32 enable = nla_get_u32(tb[TCA_FQ_RATE_ENABLE]);
730 q->rate_enable = enable;
735 if (tb[TCA_FQ_FLOW_REFILL_DELAY]) {
736 u32 usecs_delay = nla_get_u32(tb[TCA_FQ_FLOW_REFILL_DELAY]) ;
738 q->flow_refill_delay = usecs_to_jiffies(usecs_delay);
741 if (tb[TCA_FQ_ORPHAN_MASK])
742 q->orphan_mask = nla_get_u32(tb[TCA_FQ_ORPHAN_MASK]);
745 sch_tree_unlock(sch);
746 err = fq_resize(sch, fq_log);
749 while (sch->q.qlen > sch->limit) {
750 struct sk_buff *skb = fq_dequeue(sch);
754 drop_len += qdisc_pkt_len(skb);
758 qdisc_tree_reduce_backlog(sch, drop_count, drop_len);
760 sch_tree_unlock(sch);
764 static void fq_destroy(struct Qdisc *sch)
766 struct fq_sched_data *q = qdisc_priv(sch);
770 qdisc_watchdog_cancel(&q->watchdog);
773 static int fq_init(struct Qdisc *sch, struct nlattr *opt)
775 struct fq_sched_data *q = qdisc_priv(sch);
779 q->flow_plimit = 100;
780 q->quantum = 2 * psched_mtu(qdisc_dev(sch));
781 q->initial_quantum = 10 * psched_mtu(qdisc_dev(sch));
782 q->flow_refill_delay = msecs_to_jiffies(40);
783 q->flow_max_rate = ~0U;
785 q->new_flows.first = NULL;
786 q->old_flows.first = NULL;
787 q->delayed = RB_ROOT;
789 q->fq_trees_log = ilog2(1024);
790 q->orphan_mask = 1024 - 1;
791 qdisc_watchdog_init(&q->watchdog, sch);
794 err = fq_change(sch, opt);
796 err = fq_resize(sch, q->fq_trees_log);
801 static int fq_dump(struct Qdisc *sch, struct sk_buff *skb)
803 struct fq_sched_data *q = qdisc_priv(sch);
806 opts = nla_nest_start(skb, TCA_OPTIONS);
808 goto nla_put_failure;
810 /* TCA_FQ_FLOW_DEFAULT_RATE is not used anymore */
812 if (nla_put_u32(skb, TCA_FQ_PLIMIT, sch->limit) ||
813 nla_put_u32(skb, TCA_FQ_FLOW_PLIMIT, q->flow_plimit) ||
814 nla_put_u32(skb, TCA_FQ_QUANTUM, q->quantum) ||
815 nla_put_u32(skb, TCA_FQ_INITIAL_QUANTUM, q->initial_quantum) ||
816 nla_put_u32(skb, TCA_FQ_RATE_ENABLE, q->rate_enable) ||
817 nla_put_u32(skb, TCA_FQ_FLOW_MAX_RATE, q->flow_max_rate) ||
818 nla_put_u32(skb, TCA_FQ_FLOW_REFILL_DELAY,
819 jiffies_to_usecs(q->flow_refill_delay)) ||
820 nla_put_u32(skb, TCA_FQ_ORPHAN_MASK, q->orphan_mask) ||
821 nla_put_u32(skb, TCA_FQ_BUCKETS_LOG, q->fq_trees_log))
822 goto nla_put_failure;
824 return nla_nest_end(skb, opts);
830 static int fq_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
832 struct fq_sched_data *q = qdisc_priv(sch);
833 u64 now = ktime_get_ns();
834 struct tc_fq_qd_stats st = {
835 .gc_flows = q->stat_gc_flows,
836 .highprio_packets = q->stat_internal_packets,
837 .tcp_retrans = q->stat_tcp_retrans,
838 .throttled = q->stat_throttled,
839 .flows_plimit = q->stat_flows_plimit,
840 .pkts_too_long = q->stat_pkts_too_long,
841 .allocation_errors = q->stat_allocation_errors,
843 .inactive_flows = q->inactive_flows,
844 .throttled_flows = q->throttled_flows,
845 .time_next_delayed_flow = q->time_next_delayed_flow - now,
848 return gnet_stats_copy_app(d, &st, sizeof(st));
851 static struct Qdisc_ops fq_qdisc_ops __read_mostly = {
853 .priv_size = sizeof(struct fq_sched_data),
855 .enqueue = fq_enqueue,
856 .dequeue = fq_dequeue,
857 .peek = qdisc_peek_dequeued,
860 .destroy = fq_destroy,
863 .dump_stats = fq_dump_stats,
864 .owner = THIS_MODULE,
867 static int __init fq_module_init(void)
871 fq_flow_cachep = kmem_cache_create("fq_flow_cache",
872 sizeof(struct fq_flow),
877 ret = register_qdisc(&fq_qdisc_ops);
879 kmem_cache_destroy(fq_flow_cachep);
883 static void __exit fq_module_exit(void)
885 unregister_qdisc(&fq_qdisc_ops);
886 kmem_cache_destroy(fq_flow_cachep);
889 module_init(fq_module_init)
890 module_exit(fq_module_exit)
891 MODULE_AUTHOR("Eric Dumazet");
892 MODULE_LICENSE("GPL");