Loading...
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 */
60struct 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
77struct fq_flow_head {
78 struct fq_flow *first;
79 struct fq_flow *last;
80};
81
82struct 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
90 struct fq_flow internal; /* for non classified or high prio packets */
91 u32 quantum;
92 u32 initial_quantum;
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;
98 u8 rate_enable;
99 u8 fq_trees_log;
100
101 u32 flows;
102 u32 inactive_flows;
103 u32 throttled_flows;
104
105 u64 stat_gc_flows;
106 u64 stat_internal_packets;
107 u64 stat_tcp_retrans;
108 u64 stat_throttled;
109 u64 stat_flows_plimit;
110 u64 stat_pkts_too_long;
111 u64 stat_allocation_errors;
112 struct qdisc_watchdog watchdog;
113};
114
115/* special value to mark a detached flow (not on old/new list) */
116static struct fq_flow detached, throttled;
117
118static void fq_flow_set_detached(struct fq_flow *f)
119{
120 f->next = &detached;
121 f->age = jiffies;
122}
123
124static bool fq_flow_is_detached(const struct fq_flow *f)
125{
126 return f->next == &detached;
127}
128
129static void fq_flow_set_throttled(struct fq_sched_data *q, struct fq_flow *f)
130{
131 struct rb_node **p = &q->delayed.rb_node, *parent = NULL;
132
133 while (*p) {
134 struct fq_flow *aux;
135
136 parent = *p;
137 aux = container_of(parent, struct fq_flow, rate_node);
138 if (f->time_next_packet >= aux->time_next_packet)
139 p = &parent->rb_right;
140 else
141 p = &parent->rb_left;
142 }
143 rb_link_node(&f->rate_node, parent, p);
144 rb_insert_color(&f->rate_node, &q->delayed);
145 q->throttled_flows++;
146 q->stat_throttled++;
147
148 f->next = &throttled;
149 if (q->time_next_delayed_flow > f->time_next_packet)
150 q->time_next_delayed_flow = f->time_next_packet;
151}
152
153
154static struct kmem_cache *fq_flow_cachep __read_mostly;
155
156static void fq_flow_add_tail(struct fq_flow_head *head, struct fq_flow *flow)
157{
158 if (head->first)
159 head->last->next = flow;
160 else
161 head->first = flow;
162 head->last = flow;
163 flow->next = NULL;
164}
165
166/* limit number of collected flows per round */
167#define FQ_GC_MAX 8
168#define FQ_GC_AGE (3*HZ)
169
170static bool fq_gc_candidate(const struct fq_flow *f)
171{
172 return fq_flow_is_detached(f) &&
173 time_after(jiffies, f->age + FQ_GC_AGE);
174}
175
176static void fq_gc(struct fq_sched_data *q,
177 struct rb_root *root,
178 struct sock *sk)
179{
180 struct fq_flow *f, *tofree[FQ_GC_MAX];
181 struct rb_node **p, *parent;
182 int fcnt = 0;
183
184 p = &root->rb_node;
185 parent = NULL;
186 while (*p) {
187 parent = *p;
188
189 f = container_of(parent, struct fq_flow, fq_node);
190 if (f->sk == sk)
191 break;
192
193 if (fq_gc_candidate(f)) {
194 tofree[fcnt++] = f;
195 if (fcnt == FQ_GC_MAX)
196 break;
197 }
198
199 if (f->sk > sk)
200 p = &parent->rb_right;
201 else
202 p = &parent->rb_left;
203 }
204
205 q->flows -= fcnt;
206 q->inactive_flows -= fcnt;
207 q->stat_gc_flows += fcnt;
208 while (fcnt) {
209 struct fq_flow *f = tofree[--fcnt];
210
211 rb_erase(&f->fq_node, root);
212 kmem_cache_free(fq_flow_cachep, f);
213 }
214}
215
216static struct fq_flow *fq_classify(struct sk_buff *skb, struct fq_sched_data *q)
217{
218 struct rb_node **p, *parent;
219 struct sock *sk = skb->sk;
220 struct rb_root *root;
221 struct fq_flow *f;
222
223 /* warning: no starvation prevention... */
224 if (unlikely((skb->priority & TC_PRIO_MAX) == TC_PRIO_CONTROL))
225 return &q->internal;
226
227 /* SYNACK messages are attached to a TCP_NEW_SYN_RECV request socket
228 * or a listener (SYNCOOKIE mode)
229 * 1) request sockets are not full blown,
230 * they do not contain sk_pacing_rate
231 * 2) They are not part of a 'flow' yet
232 * 3) We do not want to rate limit them (eg SYNFLOOD attack),
233 * especially if the listener set SO_MAX_PACING_RATE
234 * 4) We pretend they are orphaned
235 */
236 if (!sk || sk_listener(sk)) {
237 unsigned long hash = skb_get_hash(skb) & q->orphan_mask;
238
239 /* By forcing low order bit to 1, we make sure to not
240 * collide with a local flow (socket pointers are word aligned)
241 */
242 sk = (struct sock *)((hash << 1) | 1UL);
243 skb_orphan(skb);
244 }
245
246 root = &q->fq_root[hash_32((u32)(long)sk, q->fq_trees_log)];
247
248 if (q->flows >= (2U << q->fq_trees_log) &&
249 q->inactive_flows > q->flows/2)
250 fq_gc(q, root, sk);
251
252 p = &root->rb_node;
253 parent = NULL;
254 while (*p) {
255 parent = *p;
256
257 f = container_of(parent, struct fq_flow, fq_node);
258 if (f->sk == sk) {
259 /* socket might have been reallocated, so check
260 * if its sk_hash is the same.
261 * It not, we need to refill credit with
262 * initial quantum
263 */
264 if (unlikely(skb->sk &&
265 f->socket_hash != sk->sk_hash)) {
266 f->credit = q->initial_quantum;
267 f->socket_hash = sk->sk_hash;
268 f->time_next_packet = 0ULL;
269 }
270 return f;
271 }
272 if (f->sk > sk)
273 p = &parent->rb_right;
274 else
275 p = &parent->rb_left;
276 }
277
278 f = kmem_cache_zalloc(fq_flow_cachep, GFP_ATOMIC | __GFP_NOWARN);
279 if (unlikely(!f)) {
280 q->stat_allocation_errors++;
281 return &q->internal;
282 }
283 fq_flow_set_detached(f);
284 f->sk = sk;
285 if (skb->sk)
286 f->socket_hash = sk->sk_hash;
287 f->credit = q->initial_quantum;
288
289 rb_link_node(&f->fq_node, parent, p);
290 rb_insert_color(&f->fq_node, root);
291
292 q->flows++;
293 q->inactive_flows++;
294 return f;
295}
296
297
298/* remove one skb from head of flow queue */
299static struct sk_buff *fq_dequeue_head(struct Qdisc *sch, struct fq_flow *flow)
300{
301 struct sk_buff *skb = flow->head;
302
303 if (skb) {
304 flow->head = skb->next;
305 skb->next = NULL;
306 flow->qlen--;
307 qdisc_qstats_backlog_dec(sch, skb);
308 sch->q.qlen--;
309 }
310 return skb;
311}
312
313/* We might add in the future detection of retransmits
314 * For the time being, just return false
315 */
316static bool skb_is_retransmit(struct sk_buff *skb)
317{
318 return false;
319}
320
321/* add skb to flow queue
322 * flow queue is a linked list, kind of FIFO, except for TCP retransmits
323 * We special case tcp retransmits to be transmitted before other packets.
324 * We rely on fact that TCP retransmits are unlikely, so we do not waste
325 * a separate queue or a pointer.
326 * head-> [retrans pkt 1]
327 * [retrans pkt 2]
328 * [ normal pkt 1]
329 * [ normal pkt 2]
330 * [ normal pkt 3]
331 * tail-> [ normal pkt 4]
332 */
333static void flow_queue_add(struct fq_flow *flow, struct sk_buff *skb)
334{
335 struct sk_buff *prev, *head = flow->head;
336
337 skb->next = NULL;
338 if (!head) {
339 flow->head = skb;
340 flow->tail = skb;
341 return;
342 }
343 if (likely(!skb_is_retransmit(skb))) {
344 flow->tail->next = skb;
345 flow->tail = skb;
346 return;
347 }
348
349 /* This skb is a tcp retransmit,
350 * find the last retrans packet in the queue
351 */
352 prev = NULL;
353 while (skb_is_retransmit(head)) {
354 prev = head;
355 head = head->next;
356 if (!head)
357 break;
358 }
359 if (!prev) { /* no rtx packet in queue, become the new head */
360 skb->next = flow->head;
361 flow->head = skb;
362 } else {
363 if (prev == flow->tail)
364 flow->tail = skb;
365 else
366 skb->next = prev->next;
367 prev->next = skb;
368 }
369}
370
371static int fq_enqueue(struct sk_buff *skb, struct Qdisc *sch)
372{
373 struct fq_sched_data *q = qdisc_priv(sch);
374 struct fq_flow *f;
375
376 if (unlikely(sch->q.qlen >= sch->limit))
377 return qdisc_drop(skb, sch);
378
379 f = fq_classify(skb, q);
380 if (unlikely(f->qlen >= q->flow_plimit && f != &q->internal)) {
381 q->stat_flows_plimit++;
382 return qdisc_drop(skb, sch);
383 }
384
385 f->qlen++;
386 if (skb_is_retransmit(skb))
387 q->stat_tcp_retrans++;
388 qdisc_qstats_backlog_inc(sch, skb);
389 if (fq_flow_is_detached(f)) {
390 fq_flow_add_tail(&q->new_flows, f);
391 if (time_after(jiffies, f->age + q->flow_refill_delay))
392 f->credit = max_t(u32, f->credit, q->quantum);
393 q->inactive_flows--;
394 }
395
396 /* Note: this overwrites f->age */
397 flow_queue_add(f, skb);
398
399 if (unlikely(f == &q->internal)) {
400 q->stat_internal_packets++;
401 }
402 sch->q.qlen++;
403
404 return NET_XMIT_SUCCESS;
405}
406
407static void fq_check_throttled(struct fq_sched_data *q, u64 now)
408{
409 struct rb_node *p;
410
411 if (q->time_next_delayed_flow > now)
412 return;
413
414 q->time_next_delayed_flow = ~0ULL;
415 while ((p = rb_first(&q->delayed)) != NULL) {
416 struct fq_flow *f = container_of(p, struct fq_flow, rate_node);
417
418 if (f->time_next_packet > now) {
419 q->time_next_delayed_flow = f->time_next_packet;
420 break;
421 }
422 rb_erase(p, &q->delayed);
423 q->throttled_flows--;
424 fq_flow_add_tail(&q->old_flows, f);
425 }
426}
427
428static struct sk_buff *fq_dequeue(struct Qdisc *sch)
429{
430 struct fq_sched_data *q = qdisc_priv(sch);
431 u64 now = ktime_get_ns();
432 struct fq_flow_head *head;
433 struct sk_buff *skb;
434 struct fq_flow *f;
435 u32 rate;
436
437 skb = fq_dequeue_head(sch, &q->internal);
438 if (skb)
439 goto out;
440 fq_check_throttled(q, now);
441begin:
442 head = &q->new_flows;
443 if (!head->first) {
444 head = &q->old_flows;
445 if (!head->first) {
446 if (q->time_next_delayed_flow != ~0ULL)
447 qdisc_watchdog_schedule_ns(&q->watchdog,
448 q->time_next_delayed_flow,
449 false);
450 return NULL;
451 }
452 }
453 f = head->first;
454
455 if (f->credit <= 0) {
456 f->credit += q->quantum;
457 head->first = f->next;
458 fq_flow_add_tail(&q->old_flows, f);
459 goto begin;
460 }
461
462 skb = f->head;
463 if (unlikely(skb && now < f->time_next_packet &&
464 !skb_is_tcp_pure_ack(skb))) {
465 head->first = f->next;
466 fq_flow_set_throttled(q, f);
467 goto begin;
468 }
469
470 skb = fq_dequeue_head(sch, f);
471 if (!skb) {
472 head->first = f->next;
473 /* force a pass through old_flows to prevent starvation */
474 if ((head == &q->new_flows) && q->old_flows.first) {
475 fq_flow_add_tail(&q->old_flows, f);
476 } else {
477 fq_flow_set_detached(f);
478 q->inactive_flows++;
479 }
480 goto begin;
481 }
482 prefetch(&skb->end);
483 f->credit -= qdisc_pkt_len(skb);
484
485 if (f->credit > 0 || !q->rate_enable)
486 goto out;
487
488 /* Do not pace locally generated ack packets */
489 if (skb_is_tcp_pure_ack(skb))
490 goto out;
491
492 rate = q->flow_max_rate;
493 if (skb->sk)
494 rate = min(skb->sk->sk_pacing_rate, rate);
495
496 if (rate != ~0U) {
497 u32 plen = max(qdisc_pkt_len(skb), q->quantum);
498 u64 len = (u64)plen * NSEC_PER_SEC;
499
500 if (likely(rate))
501 do_div(len, rate);
502 /* Since socket rate can change later,
503 * clamp the delay to 1 second.
504 * Really, providers of too big packets should be fixed !
505 */
506 if (unlikely(len > NSEC_PER_SEC)) {
507 len = NSEC_PER_SEC;
508 q->stat_pkts_too_long++;
509 }
510
511 f->time_next_packet = now + len;
512 }
513out:
514 qdisc_bstats_update(sch, skb);
515 return skb;
516}
517
518static void fq_reset(struct Qdisc *sch)
519{
520 struct fq_sched_data *q = qdisc_priv(sch);
521 struct rb_root *root;
522 struct sk_buff *skb;
523 struct rb_node *p;
524 struct fq_flow *f;
525 unsigned int idx;
526
527 while ((skb = fq_dequeue_head(sch, &q->internal)) != NULL)
528 kfree_skb(skb);
529
530 if (!q->fq_root)
531 return;
532
533 for (idx = 0; idx < (1U << q->fq_trees_log); idx++) {
534 root = &q->fq_root[idx];
535 while ((p = rb_first(root)) != NULL) {
536 f = container_of(p, struct fq_flow, fq_node);
537 rb_erase(p, root);
538
539 while ((skb = fq_dequeue_head(sch, f)) != NULL)
540 kfree_skb(skb);
541
542 kmem_cache_free(fq_flow_cachep, f);
543 }
544 }
545 q->new_flows.first = NULL;
546 q->old_flows.first = NULL;
547 q->delayed = RB_ROOT;
548 q->flows = 0;
549 q->inactive_flows = 0;
550 q->throttled_flows = 0;
551}
552
553static void fq_rehash(struct fq_sched_data *q,
554 struct rb_root *old_array, u32 old_log,
555 struct rb_root *new_array, u32 new_log)
556{
557 struct rb_node *op, **np, *parent;
558 struct rb_root *oroot, *nroot;
559 struct fq_flow *of, *nf;
560 int fcnt = 0;
561 u32 idx;
562
563 for (idx = 0; idx < (1U << old_log); idx++) {
564 oroot = &old_array[idx];
565 while ((op = rb_first(oroot)) != NULL) {
566 rb_erase(op, oroot);
567 of = container_of(op, struct fq_flow, fq_node);
568 if (fq_gc_candidate(of)) {
569 fcnt++;
570 kmem_cache_free(fq_flow_cachep, of);
571 continue;
572 }
573 nroot = &new_array[hash_32((u32)(long)of->sk, new_log)];
574
575 np = &nroot->rb_node;
576 parent = NULL;
577 while (*np) {
578 parent = *np;
579
580 nf = container_of(parent, struct fq_flow, fq_node);
581 BUG_ON(nf->sk == of->sk);
582
583 if (nf->sk > of->sk)
584 np = &parent->rb_right;
585 else
586 np = &parent->rb_left;
587 }
588
589 rb_link_node(&of->fq_node, parent, np);
590 rb_insert_color(&of->fq_node, nroot);
591 }
592 }
593 q->flows -= fcnt;
594 q->inactive_flows -= fcnt;
595 q->stat_gc_flows += fcnt;
596}
597
598static void *fq_alloc_node(size_t sz, int node)
599{
600 void *ptr;
601
602 ptr = kmalloc_node(sz, GFP_KERNEL | __GFP_REPEAT | __GFP_NOWARN, node);
603 if (!ptr)
604 ptr = vmalloc_node(sz, node);
605 return ptr;
606}
607
608static void fq_free(void *addr)
609{
610 kvfree(addr);
611}
612
613static int fq_resize(struct Qdisc *sch, u32 log)
614{
615 struct fq_sched_data *q = qdisc_priv(sch);
616 struct rb_root *array;
617 void *old_fq_root;
618 u32 idx;
619
620 if (q->fq_root && log == q->fq_trees_log)
621 return 0;
622
623 /* If XPS was setup, we can allocate memory on right NUMA node */
624 array = fq_alloc_node(sizeof(struct rb_root) << log,
625 netdev_queue_numa_node_read(sch->dev_queue));
626 if (!array)
627 return -ENOMEM;
628
629 for (idx = 0; idx < (1U << log); idx++)
630 array[idx] = RB_ROOT;
631
632 sch_tree_lock(sch);
633
634 old_fq_root = q->fq_root;
635 if (old_fq_root)
636 fq_rehash(q, old_fq_root, q->fq_trees_log, array, log);
637
638 q->fq_root = array;
639 q->fq_trees_log = log;
640
641 sch_tree_unlock(sch);
642
643 fq_free(old_fq_root);
644
645 return 0;
646}
647
648static const struct nla_policy fq_policy[TCA_FQ_MAX + 1] = {
649 [TCA_FQ_PLIMIT] = { .type = NLA_U32 },
650 [TCA_FQ_FLOW_PLIMIT] = { .type = NLA_U32 },
651 [TCA_FQ_QUANTUM] = { .type = NLA_U32 },
652 [TCA_FQ_INITIAL_QUANTUM] = { .type = NLA_U32 },
653 [TCA_FQ_RATE_ENABLE] = { .type = NLA_U32 },
654 [TCA_FQ_FLOW_DEFAULT_RATE] = { .type = NLA_U32 },
655 [TCA_FQ_FLOW_MAX_RATE] = { .type = NLA_U32 },
656 [TCA_FQ_BUCKETS_LOG] = { .type = NLA_U32 },
657 [TCA_FQ_FLOW_REFILL_DELAY] = { .type = NLA_U32 },
658};
659
660static int fq_change(struct Qdisc *sch, struct nlattr *opt)
661{
662 struct fq_sched_data *q = qdisc_priv(sch);
663 struct nlattr *tb[TCA_FQ_MAX + 1];
664 int err, drop_count = 0;
665 unsigned drop_len = 0;
666 u32 fq_log;
667
668 if (!opt)
669 return -EINVAL;
670
671 err = nla_parse_nested(tb, TCA_FQ_MAX, opt, fq_policy);
672 if (err < 0)
673 return err;
674
675 sch_tree_lock(sch);
676
677 fq_log = q->fq_trees_log;
678
679 if (tb[TCA_FQ_BUCKETS_LOG]) {
680 u32 nval = nla_get_u32(tb[TCA_FQ_BUCKETS_LOG]);
681
682 if (nval >= 1 && nval <= ilog2(256*1024))
683 fq_log = nval;
684 else
685 err = -EINVAL;
686 }
687 if (tb[TCA_FQ_PLIMIT])
688 sch->limit = nla_get_u32(tb[TCA_FQ_PLIMIT]);
689
690 if (tb[TCA_FQ_FLOW_PLIMIT])
691 q->flow_plimit = nla_get_u32(tb[TCA_FQ_FLOW_PLIMIT]);
692
693 if (tb[TCA_FQ_QUANTUM]) {
694 u32 quantum = nla_get_u32(tb[TCA_FQ_QUANTUM]);
695
696 if (quantum > 0)
697 q->quantum = quantum;
698 else
699 err = -EINVAL;
700 }
701
702 if (tb[TCA_FQ_INITIAL_QUANTUM])
703 q->initial_quantum = nla_get_u32(tb[TCA_FQ_INITIAL_QUANTUM]);
704
705 if (tb[TCA_FQ_FLOW_DEFAULT_RATE])
706 pr_warn_ratelimited("sch_fq: defrate %u ignored.\n",
707 nla_get_u32(tb[TCA_FQ_FLOW_DEFAULT_RATE]));
708
709 if (tb[TCA_FQ_FLOW_MAX_RATE])
710 q->flow_max_rate = nla_get_u32(tb[TCA_FQ_FLOW_MAX_RATE]);
711
712 if (tb[TCA_FQ_RATE_ENABLE]) {
713 u32 enable = nla_get_u32(tb[TCA_FQ_RATE_ENABLE]);
714
715 if (enable <= 1)
716 q->rate_enable = enable;
717 else
718 err = -EINVAL;
719 }
720
721 if (tb[TCA_FQ_FLOW_REFILL_DELAY]) {
722 u32 usecs_delay = nla_get_u32(tb[TCA_FQ_FLOW_REFILL_DELAY]) ;
723
724 q->flow_refill_delay = usecs_to_jiffies(usecs_delay);
725 }
726
727 if (tb[TCA_FQ_ORPHAN_MASK])
728 q->orphan_mask = nla_get_u32(tb[TCA_FQ_ORPHAN_MASK]);
729
730 if (!err) {
731 sch_tree_unlock(sch);
732 err = fq_resize(sch, fq_log);
733 sch_tree_lock(sch);
734 }
735 while (sch->q.qlen > sch->limit) {
736 struct sk_buff *skb = fq_dequeue(sch);
737
738 if (!skb)
739 break;
740 drop_len += qdisc_pkt_len(skb);
741 kfree_skb(skb);
742 drop_count++;
743 }
744 qdisc_tree_reduce_backlog(sch, drop_count, drop_len);
745
746 sch_tree_unlock(sch);
747 return err;
748}
749
750static void fq_destroy(struct Qdisc *sch)
751{
752 struct fq_sched_data *q = qdisc_priv(sch);
753
754 fq_reset(sch);
755 fq_free(q->fq_root);
756 qdisc_watchdog_cancel(&q->watchdog);
757}
758
759static int fq_init(struct Qdisc *sch, struct nlattr *opt)
760{
761 struct fq_sched_data *q = qdisc_priv(sch);
762 int err;
763
764 sch->limit = 10000;
765 q->flow_plimit = 100;
766 q->quantum = 2 * psched_mtu(qdisc_dev(sch));
767 q->initial_quantum = 10 * psched_mtu(qdisc_dev(sch));
768 q->flow_refill_delay = msecs_to_jiffies(40);
769 q->flow_max_rate = ~0U;
770 q->rate_enable = 1;
771 q->new_flows.first = NULL;
772 q->old_flows.first = NULL;
773 q->delayed = RB_ROOT;
774 q->fq_root = NULL;
775 q->fq_trees_log = ilog2(1024);
776 q->orphan_mask = 1024 - 1;
777 qdisc_watchdog_init(&q->watchdog, sch);
778
779 if (opt)
780 err = fq_change(sch, opt);
781 else
782 err = fq_resize(sch, q->fq_trees_log);
783
784 return err;
785}
786
787static int fq_dump(struct Qdisc *sch, struct sk_buff *skb)
788{
789 struct fq_sched_data *q = qdisc_priv(sch);
790 struct nlattr *opts;
791
792 opts = nla_nest_start(skb, TCA_OPTIONS);
793 if (opts == NULL)
794 goto nla_put_failure;
795
796 /* TCA_FQ_FLOW_DEFAULT_RATE is not used anymore */
797
798 if (nla_put_u32(skb, TCA_FQ_PLIMIT, sch->limit) ||
799 nla_put_u32(skb, TCA_FQ_FLOW_PLIMIT, q->flow_plimit) ||
800 nla_put_u32(skb, TCA_FQ_QUANTUM, q->quantum) ||
801 nla_put_u32(skb, TCA_FQ_INITIAL_QUANTUM, q->initial_quantum) ||
802 nla_put_u32(skb, TCA_FQ_RATE_ENABLE, q->rate_enable) ||
803 nla_put_u32(skb, TCA_FQ_FLOW_MAX_RATE, q->flow_max_rate) ||
804 nla_put_u32(skb, TCA_FQ_FLOW_REFILL_DELAY,
805 jiffies_to_usecs(q->flow_refill_delay)) ||
806 nla_put_u32(skb, TCA_FQ_ORPHAN_MASK, q->orphan_mask) ||
807 nla_put_u32(skb, TCA_FQ_BUCKETS_LOG, q->fq_trees_log))
808 goto nla_put_failure;
809
810 return nla_nest_end(skb, opts);
811
812nla_put_failure:
813 return -1;
814}
815
816static int fq_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
817{
818 struct fq_sched_data *q = qdisc_priv(sch);
819 u64 now = ktime_get_ns();
820 struct tc_fq_qd_stats st = {
821 .gc_flows = q->stat_gc_flows,
822 .highprio_packets = q->stat_internal_packets,
823 .tcp_retrans = q->stat_tcp_retrans,
824 .throttled = q->stat_throttled,
825 .flows_plimit = q->stat_flows_plimit,
826 .pkts_too_long = q->stat_pkts_too_long,
827 .allocation_errors = q->stat_allocation_errors,
828 .flows = q->flows,
829 .inactive_flows = q->inactive_flows,
830 .throttled_flows = q->throttled_flows,
831 .time_next_delayed_flow = q->time_next_delayed_flow - now,
832 };
833
834 return gnet_stats_copy_app(d, &st, sizeof(st));
835}
836
837static struct Qdisc_ops fq_qdisc_ops __read_mostly = {
838 .id = "fq",
839 .priv_size = sizeof(struct fq_sched_data),
840
841 .enqueue = fq_enqueue,
842 .dequeue = fq_dequeue,
843 .peek = qdisc_peek_dequeued,
844 .init = fq_init,
845 .reset = fq_reset,
846 .destroy = fq_destroy,
847 .change = fq_change,
848 .dump = fq_dump,
849 .dump_stats = fq_dump_stats,
850 .owner = THIS_MODULE,
851};
852
853static int __init fq_module_init(void)
854{
855 int ret;
856
857 fq_flow_cachep = kmem_cache_create("fq_flow_cache",
858 sizeof(struct fq_flow),
859 0, 0, NULL);
860 if (!fq_flow_cachep)
861 return -ENOMEM;
862
863 ret = register_qdisc(&fq_qdisc_ops);
864 if (ret)
865 kmem_cache_destroy(fq_flow_cachep);
866 return ret;
867}
868
869static void __exit fq_module_exit(void)
870{
871 unregister_qdisc(&fq_qdisc_ops);
872 kmem_cache_destroy(fq_flow_cachep);
873}
874
875module_init(fq_module_init)
876module_exit(fq_module_exit)
877MODULE_AUTHOR("Eric Dumazet");
878MODULE_LICENSE("GPL");
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 */
60struct 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
77struct fq_flow_head {
78 struct fq_flow *first;
79 struct fq_flow *last;
80};
81
82struct 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_max_rate; /* optional max rate per flow */
96 u32 flow_plimit; /* max packets per flow */
97 u32 orphan_mask; /* mask for orphaned skb */
98 u32 low_rate_threshold;
99 struct rb_root *fq_root;
100 u8 rate_enable;
101 u8 fq_trees_log;
102
103 u32 flows;
104 u32 inactive_flows;
105 u32 throttled_flows;
106
107 u64 stat_gc_flows;
108 u64 stat_internal_packets;
109 u64 stat_tcp_retrans;
110 u64 stat_throttled;
111 u64 stat_flows_plimit;
112 u64 stat_pkts_too_long;
113 u64 stat_allocation_errors;
114 struct qdisc_watchdog watchdog;
115};
116
117/* special value to mark a detached flow (not on old/new list) */
118static struct fq_flow detached, throttled;
119
120static void fq_flow_set_detached(struct fq_flow *f)
121{
122 f->next = &detached;
123 f->age = jiffies;
124}
125
126static bool fq_flow_is_detached(const struct fq_flow *f)
127{
128 return f->next == &detached;
129}
130
131static bool fq_flow_is_throttled(const struct fq_flow *f)
132{
133 return f->next == &throttled;
134}
135
136static void fq_flow_add_tail(struct fq_flow_head *head, struct fq_flow *flow)
137{
138 if (head->first)
139 head->last->next = flow;
140 else
141 head->first = flow;
142 head->last = flow;
143 flow->next = NULL;
144}
145
146static void fq_flow_unset_throttled(struct fq_sched_data *q, struct fq_flow *f)
147{
148 rb_erase(&f->rate_node, &q->delayed);
149 q->throttled_flows--;
150 fq_flow_add_tail(&q->old_flows, f);
151}
152
153static void fq_flow_set_throttled(struct fq_sched_data *q, struct fq_flow *f)
154{
155 struct rb_node **p = &q->delayed.rb_node, *parent = NULL;
156
157 while (*p) {
158 struct fq_flow *aux;
159
160 parent = *p;
161 aux = rb_entry(parent, struct fq_flow, rate_node);
162 if (f->time_next_packet >= aux->time_next_packet)
163 p = &parent->rb_right;
164 else
165 p = &parent->rb_left;
166 }
167 rb_link_node(&f->rate_node, parent, p);
168 rb_insert_color(&f->rate_node, &q->delayed);
169 q->throttled_flows++;
170 q->stat_throttled++;
171
172 f->next = &throttled;
173 if (q->time_next_delayed_flow > f->time_next_packet)
174 q->time_next_delayed_flow = f->time_next_packet;
175}
176
177
178static struct kmem_cache *fq_flow_cachep __read_mostly;
179
180
181/* limit number of collected flows per round */
182#define FQ_GC_MAX 8
183#define FQ_GC_AGE (3*HZ)
184
185static bool fq_gc_candidate(const struct fq_flow *f)
186{
187 return fq_flow_is_detached(f) &&
188 time_after(jiffies, f->age + FQ_GC_AGE);
189}
190
191static void fq_gc(struct fq_sched_data *q,
192 struct rb_root *root,
193 struct sock *sk)
194{
195 struct fq_flow *f, *tofree[FQ_GC_MAX];
196 struct rb_node **p, *parent;
197 int fcnt = 0;
198
199 p = &root->rb_node;
200 parent = NULL;
201 while (*p) {
202 parent = *p;
203
204 f = rb_entry(parent, struct fq_flow, fq_node);
205 if (f->sk == sk)
206 break;
207
208 if (fq_gc_candidate(f)) {
209 tofree[fcnt++] = f;
210 if (fcnt == FQ_GC_MAX)
211 break;
212 }
213
214 if (f->sk > sk)
215 p = &parent->rb_right;
216 else
217 p = &parent->rb_left;
218 }
219
220 q->flows -= fcnt;
221 q->inactive_flows -= fcnt;
222 q->stat_gc_flows += fcnt;
223 while (fcnt) {
224 struct fq_flow *f = tofree[--fcnt];
225
226 rb_erase(&f->fq_node, root);
227 kmem_cache_free(fq_flow_cachep, f);
228 }
229}
230
231static struct fq_flow *fq_classify(struct sk_buff *skb, struct fq_sched_data *q)
232{
233 struct rb_node **p, *parent;
234 struct sock *sk = skb->sk;
235 struct rb_root *root;
236 struct fq_flow *f;
237
238 /* warning: no starvation prevention... */
239 if (unlikely((skb->priority & TC_PRIO_MAX) == TC_PRIO_CONTROL))
240 return &q->internal;
241
242 /* SYNACK messages are attached to a TCP_NEW_SYN_RECV request socket
243 * or a listener (SYNCOOKIE mode)
244 * 1) request sockets are not full blown,
245 * they do not contain sk_pacing_rate
246 * 2) They are not part of a 'flow' yet
247 * 3) We do not want to rate limit them (eg SYNFLOOD attack),
248 * especially if the listener set SO_MAX_PACING_RATE
249 * 4) We pretend they are orphaned
250 */
251 if (!sk || sk_listener(sk)) {
252 unsigned long hash = skb_get_hash(skb) & q->orphan_mask;
253
254 /* By forcing low order bit to 1, we make sure to not
255 * collide with a local flow (socket pointers are word aligned)
256 */
257 sk = (struct sock *)((hash << 1) | 1UL);
258 skb_orphan(skb);
259 }
260
261 root = &q->fq_root[hash_ptr(sk, q->fq_trees_log)];
262
263 if (q->flows >= (2U << q->fq_trees_log) &&
264 q->inactive_flows > q->flows/2)
265 fq_gc(q, root, sk);
266
267 p = &root->rb_node;
268 parent = NULL;
269 while (*p) {
270 parent = *p;
271
272 f = rb_entry(parent, struct fq_flow, fq_node);
273 if (f->sk == sk) {
274 /* socket might have been reallocated, so check
275 * if its sk_hash is the same.
276 * It not, we need to refill credit with
277 * initial quantum
278 */
279 if (unlikely(skb->sk &&
280 f->socket_hash != sk->sk_hash)) {
281 f->credit = q->initial_quantum;
282 f->socket_hash = sk->sk_hash;
283 if (fq_flow_is_throttled(f))
284 fq_flow_unset_throttled(q, f);
285 f->time_next_packet = 0ULL;
286 }
287 return f;
288 }
289 if (f->sk > sk)
290 p = &parent->rb_right;
291 else
292 p = &parent->rb_left;
293 }
294
295 f = kmem_cache_zalloc(fq_flow_cachep, GFP_ATOMIC | __GFP_NOWARN);
296 if (unlikely(!f)) {
297 q->stat_allocation_errors++;
298 return &q->internal;
299 }
300 fq_flow_set_detached(f);
301 f->sk = sk;
302 if (skb->sk)
303 f->socket_hash = sk->sk_hash;
304 f->credit = q->initial_quantum;
305
306 rb_link_node(&f->fq_node, parent, p);
307 rb_insert_color(&f->fq_node, root);
308
309 q->flows++;
310 q->inactive_flows++;
311 return f;
312}
313
314
315/* remove one skb from head of flow queue */
316static struct sk_buff *fq_dequeue_head(struct Qdisc *sch, struct fq_flow *flow)
317{
318 struct sk_buff *skb = flow->head;
319
320 if (skb) {
321 flow->head = skb->next;
322 skb->next = NULL;
323 flow->qlen--;
324 qdisc_qstats_backlog_dec(sch, skb);
325 sch->q.qlen--;
326 }
327 return skb;
328}
329
330/* We might add in the future detection of retransmits
331 * For the time being, just return false
332 */
333static bool skb_is_retransmit(struct sk_buff *skb)
334{
335 return false;
336}
337
338/* add skb to flow queue
339 * flow queue is a linked list, kind of FIFO, except for TCP retransmits
340 * We special case tcp retransmits to be transmitted before other packets.
341 * We rely on fact that TCP retransmits are unlikely, so we do not waste
342 * a separate queue or a pointer.
343 * head-> [retrans pkt 1]
344 * [retrans pkt 2]
345 * [ normal pkt 1]
346 * [ normal pkt 2]
347 * [ normal pkt 3]
348 * tail-> [ normal pkt 4]
349 */
350static void flow_queue_add(struct fq_flow *flow, struct sk_buff *skb)
351{
352 struct sk_buff *prev, *head = flow->head;
353
354 skb->next = NULL;
355 if (!head) {
356 flow->head = skb;
357 flow->tail = skb;
358 return;
359 }
360 if (likely(!skb_is_retransmit(skb))) {
361 flow->tail->next = skb;
362 flow->tail = skb;
363 return;
364 }
365
366 /* This skb is a tcp retransmit,
367 * find the last retrans packet in the queue
368 */
369 prev = NULL;
370 while (skb_is_retransmit(head)) {
371 prev = head;
372 head = head->next;
373 if (!head)
374 break;
375 }
376 if (!prev) { /* no rtx packet in queue, become the new head */
377 skb->next = flow->head;
378 flow->head = skb;
379 } else {
380 if (prev == flow->tail)
381 flow->tail = skb;
382 else
383 skb->next = prev->next;
384 prev->next = skb;
385 }
386}
387
388static int fq_enqueue(struct sk_buff *skb, struct Qdisc *sch,
389 struct sk_buff **to_free)
390{
391 struct fq_sched_data *q = qdisc_priv(sch);
392 struct fq_flow *f;
393
394 if (unlikely(sch->q.qlen >= sch->limit))
395 return qdisc_drop(skb, sch, to_free);
396
397 f = fq_classify(skb, q);
398 if (unlikely(f->qlen >= q->flow_plimit && f != &q->internal)) {
399 q->stat_flows_plimit++;
400 return qdisc_drop(skb, sch, to_free);
401 }
402
403 f->qlen++;
404 if (skb_is_retransmit(skb))
405 q->stat_tcp_retrans++;
406 qdisc_qstats_backlog_inc(sch, skb);
407 if (fq_flow_is_detached(f)) {
408 struct sock *sk = skb->sk;
409
410 fq_flow_add_tail(&q->new_flows, f);
411 if (time_after(jiffies, f->age + q->flow_refill_delay))
412 f->credit = max_t(u32, f->credit, q->quantum);
413 if (sk && q->rate_enable) {
414 if (unlikely(smp_load_acquire(&sk->sk_pacing_status) !=
415 SK_PACING_FQ))
416 smp_store_release(&sk->sk_pacing_status,
417 SK_PACING_FQ);
418 }
419 q->inactive_flows--;
420 }
421
422 /* Note: this overwrites f->age */
423 flow_queue_add(f, skb);
424
425 if (unlikely(f == &q->internal)) {
426 q->stat_internal_packets++;
427 }
428 sch->q.qlen++;
429
430 return NET_XMIT_SUCCESS;
431}
432
433static void fq_check_throttled(struct fq_sched_data *q, u64 now)
434{
435 unsigned long sample;
436 struct rb_node *p;
437
438 if (q->time_next_delayed_flow > now)
439 return;
440
441 /* Update unthrottle latency EWMA.
442 * This is cheap and can help diagnosing timer/latency problems.
443 */
444 sample = (unsigned long)(now - q->time_next_delayed_flow);
445 q->unthrottle_latency_ns -= q->unthrottle_latency_ns >> 3;
446 q->unthrottle_latency_ns += sample >> 3;
447
448 q->time_next_delayed_flow = ~0ULL;
449 while ((p = rb_first(&q->delayed)) != NULL) {
450 struct fq_flow *f = rb_entry(p, struct fq_flow, rate_node);
451
452 if (f->time_next_packet > now) {
453 q->time_next_delayed_flow = f->time_next_packet;
454 break;
455 }
456 fq_flow_unset_throttled(q, f);
457 }
458}
459
460static struct sk_buff *fq_dequeue(struct Qdisc *sch)
461{
462 struct fq_sched_data *q = qdisc_priv(sch);
463 u64 now = ktime_get_ns();
464 struct fq_flow_head *head;
465 struct sk_buff *skb;
466 struct fq_flow *f;
467 u32 rate, plen;
468
469 skb = fq_dequeue_head(sch, &q->internal);
470 if (skb)
471 goto out;
472 fq_check_throttled(q, now);
473begin:
474 head = &q->new_flows;
475 if (!head->first) {
476 head = &q->old_flows;
477 if (!head->first) {
478 if (q->time_next_delayed_flow != ~0ULL)
479 qdisc_watchdog_schedule_ns(&q->watchdog,
480 q->time_next_delayed_flow);
481 return NULL;
482 }
483 }
484 f = head->first;
485
486 if (f->credit <= 0) {
487 f->credit += q->quantum;
488 head->first = f->next;
489 fq_flow_add_tail(&q->old_flows, f);
490 goto begin;
491 }
492
493 skb = f->head;
494 if (unlikely(skb && now < f->time_next_packet &&
495 !skb_is_tcp_pure_ack(skb))) {
496 head->first = f->next;
497 fq_flow_set_throttled(q, f);
498 goto begin;
499 }
500
501 skb = fq_dequeue_head(sch, f);
502 if (!skb) {
503 head->first = f->next;
504 /* force a pass through old_flows to prevent starvation */
505 if ((head == &q->new_flows) && q->old_flows.first) {
506 fq_flow_add_tail(&q->old_flows, f);
507 } else {
508 fq_flow_set_detached(f);
509 q->inactive_flows++;
510 }
511 goto begin;
512 }
513 prefetch(&skb->end);
514 f->credit -= qdisc_pkt_len(skb);
515
516 if (!q->rate_enable)
517 goto out;
518
519 /* Do not pace locally generated ack packets */
520 if (skb_is_tcp_pure_ack(skb))
521 goto out;
522
523 rate = q->flow_max_rate;
524 if (skb->sk)
525 rate = min(skb->sk->sk_pacing_rate, rate);
526
527 if (rate <= q->low_rate_threshold) {
528 f->credit = 0;
529 plen = qdisc_pkt_len(skb);
530 } else {
531 plen = max(qdisc_pkt_len(skb), q->quantum);
532 if (f->credit > 0)
533 goto out;
534 }
535 if (rate != ~0U) {
536 u64 len = (u64)plen * NSEC_PER_SEC;
537
538 if (likely(rate))
539 do_div(len, rate);
540 /* Since socket rate can change later,
541 * clamp the delay to 1 second.
542 * Really, providers of too big packets should be fixed !
543 */
544 if (unlikely(len > NSEC_PER_SEC)) {
545 len = NSEC_PER_SEC;
546 q->stat_pkts_too_long++;
547 }
548 /* Account for schedule/timers drifts.
549 * f->time_next_packet was set when prior packet was sent,
550 * and current time (@now) can be too late by tens of us.
551 */
552 if (f->time_next_packet)
553 len -= min(len/2, now - f->time_next_packet);
554 f->time_next_packet = now + len;
555 }
556out:
557 qdisc_bstats_update(sch, skb);
558 return skb;
559}
560
561static void fq_flow_purge(struct fq_flow *flow)
562{
563 rtnl_kfree_skbs(flow->head, flow->tail);
564 flow->head = NULL;
565 flow->qlen = 0;
566}
567
568static void fq_reset(struct Qdisc *sch)
569{
570 struct fq_sched_data *q = qdisc_priv(sch);
571 struct rb_root *root;
572 struct rb_node *p;
573 struct fq_flow *f;
574 unsigned int idx;
575
576 sch->q.qlen = 0;
577 sch->qstats.backlog = 0;
578
579 fq_flow_purge(&q->internal);
580
581 if (!q->fq_root)
582 return;
583
584 for (idx = 0; idx < (1U << q->fq_trees_log); idx++) {
585 root = &q->fq_root[idx];
586 while ((p = rb_first(root)) != NULL) {
587 f = rb_entry(p, struct fq_flow, fq_node);
588 rb_erase(p, root);
589
590 fq_flow_purge(f);
591
592 kmem_cache_free(fq_flow_cachep, f);
593 }
594 }
595 q->new_flows.first = NULL;
596 q->old_flows.first = NULL;
597 q->delayed = RB_ROOT;
598 q->flows = 0;
599 q->inactive_flows = 0;
600 q->throttled_flows = 0;
601}
602
603static void fq_rehash(struct fq_sched_data *q,
604 struct rb_root *old_array, u32 old_log,
605 struct rb_root *new_array, u32 new_log)
606{
607 struct rb_node *op, **np, *parent;
608 struct rb_root *oroot, *nroot;
609 struct fq_flow *of, *nf;
610 int fcnt = 0;
611 u32 idx;
612
613 for (idx = 0; idx < (1U << old_log); idx++) {
614 oroot = &old_array[idx];
615 while ((op = rb_first(oroot)) != NULL) {
616 rb_erase(op, oroot);
617 of = rb_entry(op, struct fq_flow, fq_node);
618 if (fq_gc_candidate(of)) {
619 fcnt++;
620 kmem_cache_free(fq_flow_cachep, of);
621 continue;
622 }
623 nroot = &new_array[hash_ptr(of->sk, new_log)];
624
625 np = &nroot->rb_node;
626 parent = NULL;
627 while (*np) {
628 parent = *np;
629
630 nf = rb_entry(parent, struct fq_flow, fq_node);
631 BUG_ON(nf->sk == of->sk);
632
633 if (nf->sk > of->sk)
634 np = &parent->rb_right;
635 else
636 np = &parent->rb_left;
637 }
638
639 rb_link_node(&of->fq_node, parent, np);
640 rb_insert_color(&of->fq_node, nroot);
641 }
642 }
643 q->flows -= fcnt;
644 q->inactive_flows -= fcnt;
645 q->stat_gc_flows += fcnt;
646}
647
648static void fq_free(void *addr)
649{
650 kvfree(addr);
651}
652
653static int fq_resize(struct Qdisc *sch, u32 log)
654{
655 struct fq_sched_data *q = qdisc_priv(sch);
656 struct rb_root *array;
657 void *old_fq_root;
658 u32 idx;
659
660 if (q->fq_root && log == q->fq_trees_log)
661 return 0;
662
663 /* If XPS was setup, we can allocate memory on right NUMA node */
664 array = kvmalloc_node(sizeof(struct rb_root) << log, GFP_KERNEL | __GFP_RETRY_MAYFAIL,
665 netdev_queue_numa_node_read(sch->dev_queue));
666 if (!array)
667 return -ENOMEM;
668
669 for (idx = 0; idx < (1U << log); idx++)
670 array[idx] = RB_ROOT;
671
672 sch_tree_lock(sch);
673
674 old_fq_root = q->fq_root;
675 if (old_fq_root)
676 fq_rehash(q, old_fq_root, q->fq_trees_log, array, log);
677
678 q->fq_root = array;
679 q->fq_trees_log = log;
680
681 sch_tree_unlock(sch);
682
683 fq_free(old_fq_root);
684
685 return 0;
686}
687
688static const struct nla_policy fq_policy[TCA_FQ_MAX + 1] = {
689 [TCA_FQ_PLIMIT] = { .type = NLA_U32 },
690 [TCA_FQ_FLOW_PLIMIT] = { .type = NLA_U32 },
691 [TCA_FQ_QUANTUM] = { .type = NLA_U32 },
692 [TCA_FQ_INITIAL_QUANTUM] = { .type = NLA_U32 },
693 [TCA_FQ_RATE_ENABLE] = { .type = NLA_U32 },
694 [TCA_FQ_FLOW_DEFAULT_RATE] = { .type = NLA_U32 },
695 [TCA_FQ_FLOW_MAX_RATE] = { .type = NLA_U32 },
696 [TCA_FQ_BUCKETS_LOG] = { .type = NLA_U32 },
697 [TCA_FQ_FLOW_REFILL_DELAY] = { .type = NLA_U32 },
698 [TCA_FQ_LOW_RATE_THRESHOLD] = { .type = NLA_U32 },
699};
700
701static int fq_change(struct Qdisc *sch, struct nlattr *opt,
702 struct netlink_ext_ack *extack)
703{
704 struct fq_sched_data *q = qdisc_priv(sch);
705 struct nlattr *tb[TCA_FQ_MAX + 1];
706 int err, drop_count = 0;
707 unsigned drop_len = 0;
708 u32 fq_log;
709
710 if (!opt)
711 return -EINVAL;
712
713 err = nla_parse_nested(tb, TCA_FQ_MAX, opt, fq_policy, NULL);
714 if (err < 0)
715 return err;
716
717 sch_tree_lock(sch);
718
719 fq_log = q->fq_trees_log;
720
721 if (tb[TCA_FQ_BUCKETS_LOG]) {
722 u32 nval = nla_get_u32(tb[TCA_FQ_BUCKETS_LOG]);
723
724 if (nval >= 1 && nval <= ilog2(256*1024))
725 fq_log = nval;
726 else
727 err = -EINVAL;
728 }
729 if (tb[TCA_FQ_PLIMIT])
730 sch->limit = nla_get_u32(tb[TCA_FQ_PLIMIT]);
731
732 if (tb[TCA_FQ_FLOW_PLIMIT])
733 q->flow_plimit = nla_get_u32(tb[TCA_FQ_FLOW_PLIMIT]);
734
735 if (tb[TCA_FQ_QUANTUM]) {
736 u32 quantum = nla_get_u32(tb[TCA_FQ_QUANTUM]);
737
738 if (quantum > 0)
739 q->quantum = quantum;
740 else
741 err = -EINVAL;
742 }
743
744 if (tb[TCA_FQ_INITIAL_QUANTUM])
745 q->initial_quantum = nla_get_u32(tb[TCA_FQ_INITIAL_QUANTUM]);
746
747 if (tb[TCA_FQ_FLOW_DEFAULT_RATE])
748 pr_warn_ratelimited("sch_fq: defrate %u ignored.\n",
749 nla_get_u32(tb[TCA_FQ_FLOW_DEFAULT_RATE]));
750
751 if (tb[TCA_FQ_FLOW_MAX_RATE])
752 q->flow_max_rate = nla_get_u32(tb[TCA_FQ_FLOW_MAX_RATE]);
753
754 if (tb[TCA_FQ_LOW_RATE_THRESHOLD])
755 q->low_rate_threshold =
756 nla_get_u32(tb[TCA_FQ_LOW_RATE_THRESHOLD]);
757
758 if (tb[TCA_FQ_RATE_ENABLE]) {
759 u32 enable = nla_get_u32(tb[TCA_FQ_RATE_ENABLE]);
760
761 if (enable <= 1)
762 q->rate_enable = enable;
763 else
764 err = -EINVAL;
765 }
766
767 if (tb[TCA_FQ_FLOW_REFILL_DELAY]) {
768 u32 usecs_delay = nla_get_u32(tb[TCA_FQ_FLOW_REFILL_DELAY]) ;
769
770 q->flow_refill_delay = usecs_to_jiffies(usecs_delay);
771 }
772
773 if (tb[TCA_FQ_ORPHAN_MASK])
774 q->orphan_mask = nla_get_u32(tb[TCA_FQ_ORPHAN_MASK]);
775
776 if (!err) {
777 sch_tree_unlock(sch);
778 err = fq_resize(sch, fq_log);
779 sch_tree_lock(sch);
780 }
781 while (sch->q.qlen > sch->limit) {
782 struct sk_buff *skb = fq_dequeue(sch);
783
784 if (!skb)
785 break;
786 drop_len += qdisc_pkt_len(skb);
787 rtnl_kfree_skbs(skb, skb);
788 drop_count++;
789 }
790 qdisc_tree_reduce_backlog(sch, drop_count, drop_len);
791
792 sch_tree_unlock(sch);
793 return err;
794}
795
796static void fq_destroy(struct Qdisc *sch)
797{
798 struct fq_sched_data *q = qdisc_priv(sch);
799
800 fq_reset(sch);
801 fq_free(q->fq_root);
802 qdisc_watchdog_cancel(&q->watchdog);
803}
804
805static int fq_init(struct Qdisc *sch, struct nlattr *opt,
806 struct netlink_ext_ack *extack)
807{
808 struct fq_sched_data *q = qdisc_priv(sch);
809 int err;
810
811 sch->limit = 10000;
812 q->flow_plimit = 100;
813 q->quantum = 2 * psched_mtu(qdisc_dev(sch));
814 q->initial_quantum = 10 * psched_mtu(qdisc_dev(sch));
815 q->flow_refill_delay = msecs_to_jiffies(40);
816 q->flow_max_rate = ~0U;
817 q->time_next_delayed_flow = ~0ULL;
818 q->rate_enable = 1;
819 q->new_flows.first = NULL;
820 q->old_flows.first = NULL;
821 q->delayed = RB_ROOT;
822 q->fq_root = NULL;
823 q->fq_trees_log = ilog2(1024);
824 q->orphan_mask = 1024 - 1;
825 q->low_rate_threshold = 550000 / 8;
826 qdisc_watchdog_init(&q->watchdog, sch);
827
828 if (opt)
829 err = fq_change(sch, opt, extack);
830 else
831 err = fq_resize(sch, q->fq_trees_log);
832
833 return err;
834}
835
836static int fq_dump(struct Qdisc *sch, struct sk_buff *skb)
837{
838 struct fq_sched_data *q = qdisc_priv(sch);
839 struct nlattr *opts;
840
841 opts = nla_nest_start(skb, TCA_OPTIONS);
842 if (opts == NULL)
843 goto nla_put_failure;
844
845 /* TCA_FQ_FLOW_DEFAULT_RATE is not used anymore */
846
847 if (nla_put_u32(skb, TCA_FQ_PLIMIT, sch->limit) ||
848 nla_put_u32(skb, TCA_FQ_FLOW_PLIMIT, q->flow_plimit) ||
849 nla_put_u32(skb, TCA_FQ_QUANTUM, q->quantum) ||
850 nla_put_u32(skb, TCA_FQ_INITIAL_QUANTUM, q->initial_quantum) ||
851 nla_put_u32(skb, TCA_FQ_RATE_ENABLE, q->rate_enable) ||
852 nla_put_u32(skb, TCA_FQ_FLOW_MAX_RATE, q->flow_max_rate) ||
853 nla_put_u32(skb, TCA_FQ_FLOW_REFILL_DELAY,
854 jiffies_to_usecs(q->flow_refill_delay)) ||
855 nla_put_u32(skb, TCA_FQ_ORPHAN_MASK, q->orphan_mask) ||
856 nla_put_u32(skb, TCA_FQ_LOW_RATE_THRESHOLD,
857 q->low_rate_threshold) ||
858 nla_put_u32(skb, TCA_FQ_BUCKETS_LOG, q->fq_trees_log))
859 goto nla_put_failure;
860
861 return nla_nest_end(skb, opts);
862
863nla_put_failure:
864 return -1;
865}
866
867static int fq_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
868{
869 struct fq_sched_data *q = qdisc_priv(sch);
870 struct tc_fq_qd_stats st;
871
872 sch_tree_lock(sch);
873
874 st.gc_flows = q->stat_gc_flows;
875 st.highprio_packets = q->stat_internal_packets;
876 st.tcp_retrans = q->stat_tcp_retrans;
877 st.throttled = q->stat_throttled;
878 st.flows_plimit = q->stat_flows_plimit;
879 st.pkts_too_long = q->stat_pkts_too_long;
880 st.allocation_errors = q->stat_allocation_errors;
881 st.time_next_delayed_flow = q->time_next_delayed_flow - ktime_get_ns();
882 st.flows = q->flows;
883 st.inactive_flows = q->inactive_flows;
884 st.throttled_flows = q->throttled_flows;
885 st.unthrottle_latency_ns = min_t(unsigned long,
886 q->unthrottle_latency_ns, ~0U);
887 sch_tree_unlock(sch);
888
889 return gnet_stats_copy_app(d, &st, sizeof(st));
890}
891
892static struct Qdisc_ops fq_qdisc_ops __read_mostly = {
893 .id = "fq",
894 .priv_size = sizeof(struct fq_sched_data),
895
896 .enqueue = fq_enqueue,
897 .dequeue = fq_dequeue,
898 .peek = qdisc_peek_dequeued,
899 .init = fq_init,
900 .reset = fq_reset,
901 .destroy = fq_destroy,
902 .change = fq_change,
903 .dump = fq_dump,
904 .dump_stats = fq_dump_stats,
905 .owner = THIS_MODULE,
906};
907
908static int __init fq_module_init(void)
909{
910 int ret;
911
912 fq_flow_cachep = kmem_cache_create("fq_flow_cache",
913 sizeof(struct fq_flow),
914 0, 0, NULL);
915 if (!fq_flow_cachep)
916 return -ENOMEM;
917
918 ret = register_qdisc(&fq_qdisc_ops);
919 if (ret)
920 kmem_cache_destroy(fq_flow_cachep);
921 return ret;
922}
923
924static void __exit fq_module_exit(void)
925{
926 unregister_qdisc(&fq_qdisc_ops);
927 kmem_cache_destroy(fq_flow_cachep);
928}
929
930module_init(fq_module_init)
931module_exit(fq_module_exit)
932MODULE_AUTHOR("Eric Dumazet");
933MODULE_LICENSE("GPL");