xref: /linux/net/sched/sch_fq.c (revision 3932b9ca55b0be314a36d3e84faff3e823c081f5)
1 /*
2  * net/sched/sch_fq.c Fair Queue Packet Scheduler (per flow pacing)
3  *
4  *  Copyright (C) 2013 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 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'.
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 
56 /*
57  * Per flow structure, dynamically allocated
58  */
59 struct fq_flow {
60 	struct sk_buff	*head;		/* list of skbs for this flow : first skb */
61 	union {
62 		struct sk_buff *tail;	/* last skb in the list */
63 		unsigned long  age;	/* jiffies when flow was emptied, for gc */
64 	};
65 	struct rb_node	fq_node; 	/* anchor in fq_root[] trees */
66 	struct sock	*sk;
67 	int		qlen;		/* number of packets in flow queue */
68 	int		credit;
69 	u32		socket_hash;	/* sk_hash */
70 	struct fq_flow *next;		/* next pointer in RR lists, or &detached */
71 
72 	struct rb_node  rate_node;	/* anchor in q->delayed tree */
73 	u64		time_next_packet;
74 };
75 
76 struct fq_flow_head {
77 	struct fq_flow *first;
78 	struct fq_flow *last;
79 };
80 
81 struct fq_sched_data {
82 	struct fq_flow_head new_flows;
83 
84 	struct fq_flow_head old_flows;
85 
86 	struct rb_root	delayed;	/* for rate limited flows */
87 	u64		time_next_delayed_flow;
88 
89 	struct fq_flow	internal;	/* for non classified or high prio packets */
90 	u32		quantum;
91 	u32		initial_quantum;
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;
96 	u8		rate_enable;
97 	u8		fq_trees_log;
98 
99 	u32		flows;
100 	u32		inactive_flows;
101 	u32		throttled_flows;
102 
103 	u64		stat_gc_flows;
104 	u64		stat_internal_packets;
105 	u64		stat_tcp_retrans;
106 	u64		stat_throttled;
107 	u64		stat_flows_plimit;
108 	u64		stat_pkts_too_long;
109 	u64		stat_allocation_errors;
110 	struct qdisc_watchdog watchdog;
111 };
112 
113 /* special value to mark a detached flow (not on old/new list) */
114 static struct fq_flow detached, throttled;
115 
116 static void fq_flow_set_detached(struct fq_flow *f)
117 {
118 	f->next = &detached;
119 	f->age = jiffies;
120 }
121 
122 static bool fq_flow_is_detached(const struct fq_flow *f)
123 {
124 	return f->next == &detached;
125 }
126 
127 static void fq_flow_set_throttled(struct fq_sched_data *q, struct fq_flow *f)
128 {
129 	struct rb_node **p = &q->delayed.rb_node, *parent = NULL;
130 
131 	while (*p) {
132 		struct fq_flow *aux;
133 
134 		parent = *p;
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;
138 		else
139 			p = &parent->rb_left;
140 	}
141 	rb_link_node(&f->rate_node, parent, p);
142 	rb_insert_color(&f->rate_node, &q->delayed);
143 	q->throttled_flows++;
144 	q->stat_throttled++;
145 
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;
149 }
150 
151 
152 static struct kmem_cache *fq_flow_cachep __read_mostly;
153 
154 static void fq_flow_add_tail(struct fq_flow_head *head, struct fq_flow *flow)
155 {
156 	if (head->first)
157 		head->last->next = flow;
158 	else
159 		head->first = flow;
160 	head->last = flow;
161 	flow->next = NULL;
162 }
163 
164 /* limit number of collected flows per round */
165 #define FQ_GC_MAX 8
166 #define FQ_GC_AGE (3*HZ)
167 
168 static bool fq_gc_candidate(const struct fq_flow *f)
169 {
170 	return fq_flow_is_detached(f) &&
171 	       time_after(jiffies, f->age + FQ_GC_AGE);
172 }
173 
174 static void fq_gc(struct fq_sched_data *q,
175 		  struct rb_root *root,
176 		  struct sock *sk)
177 {
178 	struct fq_flow *f, *tofree[FQ_GC_MAX];
179 	struct rb_node **p, *parent;
180 	int fcnt = 0;
181 
182 	p = &root->rb_node;
183 	parent = NULL;
184 	while (*p) {
185 		parent = *p;
186 
187 		f = container_of(parent, struct fq_flow, fq_node);
188 		if (f->sk == sk)
189 			break;
190 
191 		if (fq_gc_candidate(f)) {
192 			tofree[fcnt++] = f;
193 			if (fcnt == FQ_GC_MAX)
194 				break;
195 		}
196 
197 		if (f->sk > sk)
198 			p = &parent->rb_right;
199 		else
200 			p = &parent->rb_left;
201 	}
202 
203 	q->flows -= fcnt;
204 	q->inactive_flows -= fcnt;
205 	q->stat_gc_flows += fcnt;
206 	while (fcnt) {
207 		struct fq_flow *f = tofree[--fcnt];
208 
209 		rb_erase(&f->fq_node, root);
210 		kmem_cache_free(fq_flow_cachep, f);
211 	}
212 }
213 
214 static struct fq_flow *fq_classify(struct sk_buff *skb, struct fq_sched_data *q)
215 {
216 	struct rb_node **p, *parent;
217 	struct sock *sk = skb->sk;
218 	struct rb_root *root;
219 	struct fq_flow *f;
220 
221 	/* warning: no starvation prevention... */
222 	if (unlikely((skb->priority & TC_PRIO_MAX) == TC_PRIO_CONTROL))
223 		return &q->internal;
224 
225 	if (unlikely(!sk)) {
226 		/* By forcing low order bit to 1, we make sure to not
227 		 * collide with a local flow (socket pointers are word aligned)
228 		 */
229 		sk = (struct sock *)(skb_get_hash(skb) | 1L);
230 	}
231 
232 	root = &q->fq_root[hash_32((u32)(long)sk, q->fq_trees_log)];
233 
234 	if (q->flows >= (2U << q->fq_trees_log) &&
235 	    q->inactive_flows > q->flows/2)
236 		fq_gc(q, root, sk);
237 
238 	p = &root->rb_node;
239 	parent = NULL;
240 	while (*p) {
241 		parent = *p;
242 
243 		f = container_of(parent, struct fq_flow, fq_node);
244 		if (f->sk == sk) {
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
248 			 * initial quantum
249 			 */
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;
255 			}
256 			return f;
257 		}
258 		if (f->sk > sk)
259 			p = &parent->rb_right;
260 		else
261 			p = &parent->rb_left;
262 	}
263 
264 	f = kmem_cache_zalloc(fq_flow_cachep, GFP_ATOMIC | __GFP_NOWARN);
265 	if (unlikely(!f)) {
266 		q->stat_allocation_errors++;
267 		return &q->internal;
268 	}
269 	fq_flow_set_detached(f);
270 	f->sk = sk;
271 	if (skb->sk)
272 		f->socket_hash = sk->sk_hash;
273 	f->credit = q->initial_quantum;
274 
275 	rb_link_node(&f->fq_node, parent, p);
276 	rb_insert_color(&f->fq_node, root);
277 
278 	q->flows++;
279 	q->inactive_flows++;
280 	return f;
281 }
282 
283 
284 /* remove one skb from head of flow queue */
285 static struct sk_buff *fq_dequeue_head(struct Qdisc *sch, struct fq_flow *flow)
286 {
287 	struct sk_buff *skb = flow->head;
288 
289 	if (skb) {
290 		flow->head = skb->next;
291 		skb->next = NULL;
292 		flow->qlen--;
293 		sch->qstats.backlog -= qdisc_pkt_len(skb);
294 		sch->q.qlen--;
295 	}
296 	return skb;
297 }
298 
299 /* We might add in the future detection of retransmits
300  * For the time being, just return false
301  */
302 static bool skb_is_retransmit(struct sk_buff *skb)
303 {
304 	return false;
305 }
306 
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]
313  *         [retrans pkt 2]
314  *         [ normal pkt 1]
315  *         [ normal pkt 2]
316  *         [ normal pkt 3]
317  * tail->  [ normal pkt 4]
318  */
319 static void flow_queue_add(struct fq_flow *flow, struct sk_buff *skb)
320 {
321 	struct sk_buff *prev, *head = flow->head;
322 
323 	skb->next = NULL;
324 	if (!head) {
325 		flow->head = skb;
326 		flow->tail = skb;
327 		return;
328 	}
329 	if (likely(!skb_is_retransmit(skb))) {
330 		flow->tail->next = skb;
331 		flow->tail = skb;
332 		return;
333 	}
334 
335 	/* This skb is a tcp retransmit,
336 	 * find the last retrans packet in the queue
337 	 */
338 	prev = NULL;
339 	while (skb_is_retransmit(head)) {
340 		prev = head;
341 		head = head->next;
342 		if (!head)
343 			break;
344 	}
345 	if (!prev) { /* no rtx packet in queue, become the new head */
346 		skb->next = flow->head;
347 		flow->head = skb;
348 	} else {
349 		if (prev == flow->tail)
350 			flow->tail = skb;
351 		else
352 			skb->next = prev->next;
353 		prev->next = skb;
354 	}
355 }
356 
357 static int fq_enqueue(struct sk_buff *skb, struct Qdisc *sch)
358 {
359 	struct fq_sched_data *q = qdisc_priv(sch);
360 	struct fq_flow *f;
361 
362 	if (unlikely(sch->q.qlen >= sch->limit))
363 		return qdisc_drop(skb, sch);
364 
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);
369 	}
370 
371 	f->qlen++;
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);
379 		q->inactive_flows--;
380 		qdisc_unthrottled(sch);
381 	}
382 
383 	/* Note: this overwrites f->age */
384 	flow_queue_add(f, skb);
385 
386 	if (unlikely(f == &q->internal)) {
387 		q->stat_internal_packets++;
388 		qdisc_unthrottled(sch);
389 	}
390 	sch->q.qlen++;
391 
392 	return NET_XMIT_SUCCESS;
393 }
394 
395 static void fq_check_throttled(struct fq_sched_data *q, u64 now)
396 {
397 	struct rb_node *p;
398 
399 	if (q->time_next_delayed_flow > now)
400 		return;
401 
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);
405 
406 		if (f->time_next_packet > now) {
407 			q->time_next_delayed_flow = f->time_next_packet;
408 			break;
409 		}
410 		rb_erase(p, &q->delayed);
411 		q->throttled_flows--;
412 		fq_flow_add_tail(&q->old_flows, f);
413 	}
414 }
415 
416 static struct sk_buff *fq_dequeue(struct Qdisc *sch)
417 {
418 	struct fq_sched_data *q = qdisc_priv(sch);
419 	u64 now = ktime_to_ns(ktime_get());
420 	struct fq_flow_head *head;
421 	struct sk_buff *skb;
422 	struct fq_flow *f;
423 	u32 rate;
424 
425 	skb = fq_dequeue_head(sch, &q->internal);
426 	if (skb)
427 		goto out;
428 	fq_check_throttled(q, now);
429 begin:
430 	head = &q->new_flows;
431 	if (!head->first) {
432 		head = &q->old_flows;
433 		if (!head->first) {
434 			if (q->time_next_delayed_flow != ~0ULL)
435 				qdisc_watchdog_schedule_ns(&q->watchdog,
436 							   q->time_next_delayed_flow);
437 			return NULL;
438 		}
439 	}
440 	f = head->first;
441 
442 	if (f->credit <= 0) {
443 		f->credit += q->quantum;
444 		head->first = f->next;
445 		fq_flow_add_tail(&q->old_flows, f);
446 		goto begin;
447 	}
448 
449 	if (unlikely(f->head && now < f->time_next_packet)) {
450 		head->first = f->next;
451 		fq_flow_set_throttled(q, f);
452 		goto begin;
453 	}
454 
455 	skb = fq_dequeue_head(sch, f);
456 	if (!skb) {
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);
461 		} else {
462 			fq_flow_set_detached(f);
463 			q->inactive_flows++;
464 		}
465 		goto begin;
466 	}
467 	prefetch(&skb->end);
468 	f->time_next_packet = now;
469 	f->credit -= qdisc_pkt_len(skb);
470 
471 	if (f->credit > 0 || !q->rate_enable)
472 		goto out;
473 
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);
477 
478 	if (rate != ~0U) {
479 		u32 plen = max(qdisc_pkt_len(skb), q->quantum);
480 		u64 len = (u64)plen * NSEC_PER_SEC;
481 
482 		if (likely(rate))
483 			do_div(len, rate);
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")
488 		 */
489 		if (unlikely(len > 125 * NSEC_PER_MSEC)) {
490 			len = 125 * NSEC_PER_MSEC;
491 			q->stat_pkts_too_long++;
492 		}
493 
494 		f->time_next_packet = now + len;
495 	}
496 out:
497 	qdisc_bstats_update(sch, skb);
498 	qdisc_unthrottled(sch);
499 	return skb;
500 }
501 
502 static void fq_reset(struct Qdisc *sch)
503 {
504 	struct fq_sched_data *q = qdisc_priv(sch);
505 	struct rb_root *root;
506 	struct sk_buff *skb;
507 	struct rb_node *p;
508 	struct fq_flow *f;
509 	unsigned int idx;
510 
511 	while ((skb = fq_dequeue_head(sch, &q->internal)) != NULL)
512 		kfree_skb(skb);
513 
514 	if (!q->fq_root)
515 		return;
516 
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);
521 			rb_erase(p, root);
522 
523 			while ((skb = fq_dequeue_head(sch, f)) != NULL)
524 				kfree_skb(skb);
525 
526 			kmem_cache_free(fq_flow_cachep, f);
527 		}
528 	}
529 	q->new_flows.first	= NULL;
530 	q->old_flows.first	= NULL;
531 	q->delayed		= RB_ROOT;
532 	q->flows		= 0;
533 	q->inactive_flows	= 0;
534 	q->throttled_flows	= 0;
535 }
536 
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)
540 {
541 	struct rb_node *op, **np, *parent;
542 	struct rb_root *oroot, *nroot;
543 	struct fq_flow *of, *nf;
544 	int fcnt = 0;
545 	u32 idx;
546 
547 	for (idx = 0; idx < (1U << old_log); idx++) {
548 		oroot = &old_array[idx];
549 		while ((op = rb_first(oroot)) != NULL) {
550 			rb_erase(op, oroot);
551 			of = container_of(op, struct fq_flow, fq_node);
552 			if (fq_gc_candidate(of)) {
553 				fcnt++;
554 				kmem_cache_free(fq_flow_cachep, of);
555 				continue;
556 			}
557 			nroot = &new_array[hash_32((u32)(long)of->sk, new_log)];
558 
559 			np = &nroot->rb_node;
560 			parent = NULL;
561 			while (*np) {
562 				parent = *np;
563 
564 				nf = container_of(parent, struct fq_flow, fq_node);
565 				BUG_ON(nf->sk == of->sk);
566 
567 				if (nf->sk > of->sk)
568 					np = &parent->rb_right;
569 				else
570 					np = &parent->rb_left;
571 			}
572 
573 			rb_link_node(&of->fq_node, parent, np);
574 			rb_insert_color(&of->fq_node, nroot);
575 		}
576 	}
577 	q->flows -= fcnt;
578 	q->inactive_flows -= fcnt;
579 	q->stat_gc_flows += fcnt;
580 }
581 
582 static void *fq_alloc_node(size_t sz, int node)
583 {
584 	void *ptr;
585 
586 	ptr = kmalloc_node(sz, GFP_KERNEL | __GFP_REPEAT | __GFP_NOWARN, node);
587 	if (!ptr)
588 		ptr = vmalloc_node(sz, node);
589 	return ptr;
590 }
591 
592 static void fq_free(void *addr)
593 {
594 	kvfree(addr);
595 }
596 
597 static int fq_resize(struct Qdisc *sch, u32 log)
598 {
599 	struct fq_sched_data *q = qdisc_priv(sch);
600 	struct rb_root *array;
601 	void *old_fq_root;
602 	u32 idx;
603 
604 	if (q->fq_root && log == q->fq_trees_log)
605 		return 0;
606 
607 	/* If XPS was setup, we can allocate memory on right NUMA node */
608 	array = fq_alloc_node(sizeof(struct rb_root) << log,
609 			      netdev_queue_numa_node_read(sch->dev_queue));
610 	if (!array)
611 		return -ENOMEM;
612 
613 	for (idx = 0; idx < (1U << log); idx++)
614 		array[idx] = RB_ROOT;
615 
616 	sch_tree_lock(sch);
617 
618 	old_fq_root = q->fq_root;
619 	if (old_fq_root)
620 		fq_rehash(q, old_fq_root, q->fq_trees_log, array, log);
621 
622 	q->fq_root = array;
623 	q->fq_trees_log = log;
624 
625 	sch_tree_unlock(sch);
626 
627 	fq_free(old_fq_root);
628 
629 	return 0;
630 }
631 
632 static const struct nla_policy fq_policy[TCA_FQ_MAX + 1] = {
633 	[TCA_FQ_PLIMIT]			= { .type = NLA_U32 },
634 	[TCA_FQ_FLOW_PLIMIT]		= { .type = NLA_U32 },
635 	[TCA_FQ_QUANTUM]		= { .type = NLA_U32 },
636 	[TCA_FQ_INITIAL_QUANTUM]	= { .type = NLA_U32 },
637 	[TCA_FQ_RATE_ENABLE]		= { .type = NLA_U32 },
638 	[TCA_FQ_FLOW_DEFAULT_RATE]	= { .type = NLA_U32 },
639 	[TCA_FQ_FLOW_MAX_RATE]		= { .type = NLA_U32 },
640 	[TCA_FQ_BUCKETS_LOG]		= { .type = NLA_U32 },
641 	[TCA_FQ_FLOW_REFILL_DELAY]	= { .type = NLA_U32 },
642 };
643 
644 static int fq_change(struct Qdisc *sch, struct nlattr *opt)
645 {
646 	struct fq_sched_data *q = qdisc_priv(sch);
647 	struct nlattr *tb[TCA_FQ_MAX + 1];
648 	int err, drop_count = 0;
649 	u32 fq_log;
650 
651 	if (!opt)
652 		return -EINVAL;
653 
654 	err = nla_parse_nested(tb, TCA_FQ_MAX, opt, fq_policy);
655 	if (err < 0)
656 		return err;
657 
658 	sch_tree_lock(sch);
659 
660 	fq_log = q->fq_trees_log;
661 
662 	if (tb[TCA_FQ_BUCKETS_LOG]) {
663 		u32 nval = nla_get_u32(tb[TCA_FQ_BUCKETS_LOG]);
664 
665 		if (nval >= 1 && nval <= ilog2(256*1024))
666 			fq_log = nval;
667 		else
668 			err = -EINVAL;
669 	}
670 	if (tb[TCA_FQ_PLIMIT])
671 		sch->limit = nla_get_u32(tb[TCA_FQ_PLIMIT]);
672 
673 	if (tb[TCA_FQ_FLOW_PLIMIT])
674 		q->flow_plimit = nla_get_u32(tb[TCA_FQ_FLOW_PLIMIT]);
675 
676 	if (tb[TCA_FQ_QUANTUM])
677 		q->quantum = nla_get_u32(tb[TCA_FQ_QUANTUM]);
678 
679 	if (tb[TCA_FQ_INITIAL_QUANTUM])
680 		q->initial_quantum = nla_get_u32(tb[TCA_FQ_INITIAL_QUANTUM]);
681 
682 	if (tb[TCA_FQ_FLOW_DEFAULT_RATE])
683 		pr_warn_ratelimited("sch_fq: defrate %u ignored.\n",
684 				    nla_get_u32(tb[TCA_FQ_FLOW_DEFAULT_RATE]));
685 
686 	if (tb[TCA_FQ_FLOW_MAX_RATE])
687 		q->flow_max_rate = nla_get_u32(tb[TCA_FQ_FLOW_MAX_RATE]);
688 
689 	if (tb[TCA_FQ_RATE_ENABLE]) {
690 		u32 enable = nla_get_u32(tb[TCA_FQ_RATE_ENABLE]);
691 
692 		if (enable <= 1)
693 			q->rate_enable = enable;
694 		else
695 			err = -EINVAL;
696 	}
697 
698 	if (tb[TCA_FQ_FLOW_REFILL_DELAY]) {
699 		u32 usecs_delay = nla_get_u32(tb[TCA_FQ_FLOW_REFILL_DELAY]) ;
700 
701 		q->flow_refill_delay = usecs_to_jiffies(usecs_delay);
702 	}
703 
704 	if (!err) {
705 		sch_tree_unlock(sch);
706 		err = fq_resize(sch, fq_log);
707 		sch_tree_lock(sch);
708 	}
709 	while (sch->q.qlen > sch->limit) {
710 		struct sk_buff *skb = fq_dequeue(sch);
711 
712 		if (!skb)
713 			break;
714 		kfree_skb(skb);
715 		drop_count++;
716 	}
717 	qdisc_tree_decrease_qlen(sch, drop_count);
718 
719 	sch_tree_unlock(sch);
720 	return err;
721 }
722 
723 static void fq_destroy(struct Qdisc *sch)
724 {
725 	struct fq_sched_data *q = qdisc_priv(sch);
726 
727 	fq_reset(sch);
728 	fq_free(q->fq_root);
729 	qdisc_watchdog_cancel(&q->watchdog);
730 }
731 
732 static int fq_init(struct Qdisc *sch, struct nlattr *opt)
733 {
734 	struct fq_sched_data *q = qdisc_priv(sch);
735 	int err;
736 
737 	sch->limit		= 10000;
738 	q->flow_plimit		= 100;
739 	q->quantum		= 2 * psched_mtu(qdisc_dev(sch));
740 	q->initial_quantum	= 10 * psched_mtu(qdisc_dev(sch));
741 	q->flow_refill_delay	= msecs_to_jiffies(40);
742 	q->flow_max_rate	= ~0U;
743 	q->rate_enable		= 1;
744 	q->new_flows.first	= NULL;
745 	q->old_flows.first	= NULL;
746 	q->delayed		= RB_ROOT;
747 	q->fq_root		= NULL;
748 	q->fq_trees_log		= ilog2(1024);
749 	qdisc_watchdog_init(&q->watchdog, sch);
750 
751 	if (opt)
752 		err = fq_change(sch, opt);
753 	else
754 		err = fq_resize(sch, q->fq_trees_log);
755 
756 	return err;
757 }
758 
759 static int fq_dump(struct Qdisc *sch, struct sk_buff *skb)
760 {
761 	struct fq_sched_data *q = qdisc_priv(sch);
762 	struct nlattr *opts;
763 
764 	opts = nla_nest_start(skb, TCA_OPTIONS);
765 	if (opts == NULL)
766 		goto nla_put_failure;
767 
768 	/* TCA_FQ_FLOW_DEFAULT_RATE is not used anymore */
769 
770 	if (nla_put_u32(skb, TCA_FQ_PLIMIT, sch->limit) ||
771 	    nla_put_u32(skb, TCA_FQ_FLOW_PLIMIT, q->flow_plimit) ||
772 	    nla_put_u32(skb, TCA_FQ_QUANTUM, q->quantum) ||
773 	    nla_put_u32(skb, TCA_FQ_INITIAL_QUANTUM, q->initial_quantum) ||
774 	    nla_put_u32(skb, TCA_FQ_RATE_ENABLE, q->rate_enable) ||
775 	    nla_put_u32(skb, TCA_FQ_FLOW_MAX_RATE, q->flow_max_rate) ||
776 	    nla_put_u32(skb, TCA_FQ_FLOW_REFILL_DELAY,
777 			jiffies_to_usecs(q->flow_refill_delay)) ||
778 	    nla_put_u32(skb, TCA_FQ_BUCKETS_LOG, q->fq_trees_log))
779 		goto nla_put_failure;
780 
781 	return nla_nest_end(skb, opts);
782 
783 nla_put_failure:
784 	return -1;
785 }
786 
787 static int fq_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
788 {
789 	struct fq_sched_data *q = qdisc_priv(sch);
790 	u64 now = ktime_to_ns(ktime_get());
791 	struct tc_fq_qd_stats st = {
792 		.gc_flows		= q->stat_gc_flows,
793 		.highprio_packets	= q->stat_internal_packets,
794 		.tcp_retrans		= q->stat_tcp_retrans,
795 		.throttled		= q->stat_throttled,
796 		.flows_plimit		= q->stat_flows_plimit,
797 		.pkts_too_long		= q->stat_pkts_too_long,
798 		.allocation_errors	= q->stat_allocation_errors,
799 		.flows			= q->flows,
800 		.inactive_flows		= q->inactive_flows,
801 		.throttled_flows	= q->throttled_flows,
802 		.time_next_delayed_flow	= q->time_next_delayed_flow - now,
803 	};
804 
805 	return gnet_stats_copy_app(d, &st, sizeof(st));
806 }
807 
808 static struct Qdisc_ops fq_qdisc_ops __read_mostly = {
809 	.id		=	"fq",
810 	.priv_size	=	sizeof(struct fq_sched_data),
811 
812 	.enqueue	=	fq_enqueue,
813 	.dequeue	=	fq_dequeue,
814 	.peek		=	qdisc_peek_dequeued,
815 	.init		=	fq_init,
816 	.reset		=	fq_reset,
817 	.destroy	=	fq_destroy,
818 	.change		=	fq_change,
819 	.dump		=	fq_dump,
820 	.dump_stats	=	fq_dump_stats,
821 	.owner		=	THIS_MODULE,
822 };
823 
824 static int __init fq_module_init(void)
825 {
826 	int ret;
827 
828 	fq_flow_cachep = kmem_cache_create("fq_flow_cache",
829 					   sizeof(struct fq_flow),
830 					   0, 0, NULL);
831 	if (!fq_flow_cachep)
832 		return -ENOMEM;
833 
834 	ret = register_qdisc(&fq_qdisc_ops);
835 	if (ret)
836 		kmem_cache_destroy(fq_flow_cachep);
837 	return ret;
838 }
839 
840 static void __exit fq_module_exit(void)
841 {
842 	unregister_qdisc(&fq_qdisc_ops);
843 	kmem_cache_destroy(fq_flow_cachep);
844 }
845 
846 module_init(fq_module_init)
847 module_exit(fq_module_exit)
848 MODULE_AUTHOR("Eric Dumazet");
849 MODULE_LICENSE("GPL");
850