xref: /openbmc/linux/net/sched/sch_fq.c (revision 827634ad)
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  */
60 struct 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 
77 struct fq_flow_head {
78 	struct fq_flow *first;
79 	struct fq_flow *last;
80 };
81 
82 struct 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) */
116 static struct fq_flow detached, throttled;
117 
118 static void fq_flow_set_detached(struct fq_flow *f)
119 {
120 	f->next = &detached;
121 	f->age = jiffies;
122 }
123 
124 static bool fq_flow_is_detached(const struct fq_flow *f)
125 {
126 	return f->next == &detached;
127 }
128 
129 static 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 
154 static struct kmem_cache *fq_flow_cachep __read_mostly;
155 
156 static 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 
170 static 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 
176 static 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 
216 static 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 listener socket.
228 	 * 1) They are not part of a 'flow' yet
229 	 * 2) We do not want to rate limit them (eg SYNFLOOD attack),
230 	 *    especially if the listener set SO_MAX_PACING_RATE
231 	 * 3) We pretend they are orphaned
232 	 */
233 	if (!sk || sk->sk_state == TCP_LISTEN) {
234 		unsigned long hash = skb_get_hash(skb) & q->orphan_mask;
235 
236 		/* By forcing low order bit to 1, we make sure to not
237 		 * collide with a local flow (socket pointers are word aligned)
238 		 */
239 		sk = (struct sock *)((hash << 1) | 1UL);
240 		skb_orphan(skb);
241 	}
242 
243 	root = &q->fq_root[hash_32((u32)(long)sk, q->fq_trees_log)];
244 
245 	if (q->flows >= (2U << q->fq_trees_log) &&
246 	    q->inactive_flows > q->flows/2)
247 		fq_gc(q, root, sk);
248 
249 	p = &root->rb_node;
250 	parent = NULL;
251 	while (*p) {
252 		parent = *p;
253 
254 		f = container_of(parent, struct fq_flow, fq_node);
255 		if (f->sk == sk) {
256 			/* socket might have been reallocated, so check
257 			 * if its sk_hash is the same.
258 			 * It not, we need to refill credit with
259 			 * initial quantum
260 			 */
261 			if (unlikely(skb->sk &&
262 				     f->socket_hash != sk->sk_hash)) {
263 				f->credit = q->initial_quantum;
264 				f->socket_hash = sk->sk_hash;
265 				f->time_next_packet = 0ULL;
266 			}
267 			return f;
268 		}
269 		if (f->sk > sk)
270 			p = &parent->rb_right;
271 		else
272 			p = &parent->rb_left;
273 	}
274 
275 	f = kmem_cache_zalloc(fq_flow_cachep, GFP_ATOMIC | __GFP_NOWARN);
276 	if (unlikely(!f)) {
277 		q->stat_allocation_errors++;
278 		return &q->internal;
279 	}
280 	fq_flow_set_detached(f);
281 	f->sk = sk;
282 	if (skb->sk)
283 		f->socket_hash = sk->sk_hash;
284 	f->credit = q->initial_quantum;
285 
286 	rb_link_node(&f->fq_node, parent, p);
287 	rb_insert_color(&f->fq_node, root);
288 
289 	q->flows++;
290 	q->inactive_flows++;
291 	return f;
292 }
293 
294 
295 /* remove one skb from head of flow queue */
296 static struct sk_buff *fq_dequeue_head(struct Qdisc *sch, struct fq_flow *flow)
297 {
298 	struct sk_buff *skb = flow->head;
299 
300 	if (skb) {
301 		flow->head = skb->next;
302 		skb->next = NULL;
303 		flow->qlen--;
304 		qdisc_qstats_backlog_dec(sch, skb);
305 		sch->q.qlen--;
306 	}
307 	return skb;
308 }
309 
310 /* We might add in the future detection of retransmits
311  * For the time being, just return false
312  */
313 static bool skb_is_retransmit(struct sk_buff *skb)
314 {
315 	return false;
316 }
317 
318 /* add skb to flow queue
319  * flow queue is a linked list, kind of FIFO, except for TCP retransmits
320  * We special case tcp retransmits to be transmitted before other packets.
321  * We rely on fact that TCP retransmits are unlikely, so we do not waste
322  * a separate queue or a pointer.
323  * head->  [retrans pkt 1]
324  *         [retrans pkt 2]
325  *         [ normal pkt 1]
326  *         [ normal pkt 2]
327  *         [ normal pkt 3]
328  * tail->  [ normal pkt 4]
329  */
330 static void flow_queue_add(struct fq_flow *flow, struct sk_buff *skb)
331 {
332 	struct sk_buff *prev, *head = flow->head;
333 
334 	skb->next = NULL;
335 	if (!head) {
336 		flow->head = skb;
337 		flow->tail = skb;
338 		return;
339 	}
340 	if (likely(!skb_is_retransmit(skb))) {
341 		flow->tail->next = skb;
342 		flow->tail = skb;
343 		return;
344 	}
345 
346 	/* This skb is a tcp retransmit,
347 	 * find the last retrans packet in the queue
348 	 */
349 	prev = NULL;
350 	while (skb_is_retransmit(head)) {
351 		prev = head;
352 		head = head->next;
353 		if (!head)
354 			break;
355 	}
356 	if (!prev) { /* no rtx packet in queue, become the new head */
357 		skb->next = flow->head;
358 		flow->head = skb;
359 	} else {
360 		if (prev == flow->tail)
361 			flow->tail = skb;
362 		else
363 			skb->next = prev->next;
364 		prev->next = skb;
365 	}
366 }
367 
368 static int fq_enqueue(struct sk_buff *skb, struct Qdisc *sch)
369 {
370 	struct fq_sched_data *q = qdisc_priv(sch);
371 	struct fq_flow *f;
372 
373 	if (unlikely(sch->q.qlen >= sch->limit))
374 		return qdisc_drop(skb, sch);
375 
376 	f = fq_classify(skb, q);
377 	if (unlikely(f->qlen >= q->flow_plimit && f != &q->internal)) {
378 		q->stat_flows_plimit++;
379 		return qdisc_drop(skb, sch);
380 	}
381 
382 	f->qlen++;
383 	if (skb_is_retransmit(skb))
384 		q->stat_tcp_retrans++;
385 	qdisc_qstats_backlog_inc(sch, skb);
386 	if (fq_flow_is_detached(f)) {
387 		fq_flow_add_tail(&q->new_flows, f);
388 		if (time_after(jiffies, f->age + q->flow_refill_delay))
389 			f->credit = max_t(u32, f->credit, q->quantum);
390 		q->inactive_flows--;
391 	}
392 
393 	/* Note: this overwrites f->age */
394 	flow_queue_add(f, skb);
395 
396 	if (unlikely(f == &q->internal)) {
397 		q->stat_internal_packets++;
398 	}
399 	sch->q.qlen++;
400 
401 	return NET_XMIT_SUCCESS;
402 }
403 
404 static void fq_check_throttled(struct fq_sched_data *q, u64 now)
405 {
406 	struct rb_node *p;
407 
408 	if (q->time_next_delayed_flow > now)
409 		return;
410 
411 	q->time_next_delayed_flow = ~0ULL;
412 	while ((p = rb_first(&q->delayed)) != NULL) {
413 		struct fq_flow *f = container_of(p, struct fq_flow, rate_node);
414 
415 		if (f->time_next_packet > now) {
416 			q->time_next_delayed_flow = f->time_next_packet;
417 			break;
418 		}
419 		rb_erase(p, &q->delayed);
420 		q->throttled_flows--;
421 		fq_flow_add_tail(&q->old_flows, f);
422 	}
423 }
424 
425 static struct sk_buff *fq_dequeue(struct Qdisc *sch)
426 {
427 	struct fq_sched_data *q = qdisc_priv(sch);
428 	u64 now = ktime_get_ns();
429 	struct fq_flow_head *head;
430 	struct sk_buff *skb;
431 	struct fq_flow *f;
432 	u32 rate;
433 
434 	skb = fq_dequeue_head(sch, &q->internal);
435 	if (skb)
436 		goto out;
437 	fq_check_throttled(q, now);
438 begin:
439 	head = &q->new_flows;
440 	if (!head->first) {
441 		head = &q->old_flows;
442 		if (!head->first) {
443 			if (q->time_next_delayed_flow != ~0ULL)
444 				qdisc_watchdog_schedule_ns(&q->watchdog,
445 							   q->time_next_delayed_flow,
446 							   false);
447 			return NULL;
448 		}
449 	}
450 	f = head->first;
451 
452 	if (f->credit <= 0) {
453 		f->credit += q->quantum;
454 		head->first = f->next;
455 		fq_flow_add_tail(&q->old_flows, f);
456 		goto begin;
457 	}
458 
459 	skb = f->head;
460 	if (unlikely(skb && now < f->time_next_packet &&
461 		     !skb_is_tcp_pure_ack(skb))) {
462 		head->first = f->next;
463 		fq_flow_set_throttled(q, f);
464 		goto begin;
465 	}
466 
467 	skb = fq_dequeue_head(sch, f);
468 	if (!skb) {
469 		head->first = f->next;
470 		/* force a pass through old_flows to prevent starvation */
471 		if ((head == &q->new_flows) && q->old_flows.first) {
472 			fq_flow_add_tail(&q->old_flows, f);
473 		} else {
474 			fq_flow_set_detached(f);
475 			q->inactive_flows++;
476 		}
477 		goto begin;
478 	}
479 	prefetch(&skb->end);
480 	f->credit -= qdisc_pkt_len(skb);
481 
482 	if (f->credit > 0 || !q->rate_enable)
483 		goto out;
484 
485 	/* Do not pace locally generated ack packets */
486 	if (skb_is_tcp_pure_ack(skb))
487 		goto out;
488 
489 	rate = q->flow_max_rate;
490 	if (skb->sk)
491 		rate = min(skb->sk->sk_pacing_rate, rate);
492 
493 	if (rate != ~0U) {
494 		u32 plen = max(qdisc_pkt_len(skb), q->quantum);
495 		u64 len = (u64)plen * NSEC_PER_SEC;
496 
497 		if (likely(rate))
498 			do_div(len, rate);
499 		/* Since socket rate can change later,
500 		 * clamp the delay to 1 second.
501 		 * Really, providers of too big packets should be fixed !
502 		 */
503 		if (unlikely(len > NSEC_PER_SEC)) {
504 			len = NSEC_PER_SEC;
505 			q->stat_pkts_too_long++;
506 		}
507 
508 		f->time_next_packet = now + len;
509 	}
510 out:
511 	qdisc_bstats_update(sch, skb);
512 	return skb;
513 }
514 
515 static void fq_reset(struct Qdisc *sch)
516 {
517 	struct fq_sched_data *q = qdisc_priv(sch);
518 	struct rb_root *root;
519 	struct sk_buff *skb;
520 	struct rb_node *p;
521 	struct fq_flow *f;
522 	unsigned int idx;
523 
524 	while ((skb = fq_dequeue_head(sch, &q->internal)) != NULL)
525 		kfree_skb(skb);
526 
527 	if (!q->fq_root)
528 		return;
529 
530 	for (idx = 0; idx < (1U << q->fq_trees_log); idx++) {
531 		root = &q->fq_root[idx];
532 		while ((p = rb_first(root)) != NULL) {
533 			f = container_of(p, struct fq_flow, fq_node);
534 			rb_erase(p, root);
535 
536 			while ((skb = fq_dequeue_head(sch, f)) != NULL)
537 				kfree_skb(skb);
538 
539 			kmem_cache_free(fq_flow_cachep, f);
540 		}
541 	}
542 	q->new_flows.first	= NULL;
543 	q->old_flows.first	= NULL;
544 	q->delayed		= RB_ROOT;
545 	q->flows		= 0;
546 	q->inactive_flows	= 0;
547 	q->throttled_flows	= 0;
548 }
549 
550 static void fq_rehash(struct fq_sched_data *q,
551 		      struct rb_root *old_array, u32 old_log,
552 		      struct rb_root *new_array, u32 new_log)
553 {
554 	struct rb_node *op, **np, *parent;
555 	struct rb_root *oroot, *nroot;
556 	struct fq_flow *of, *nf;
557 	int fcnt = 0;
558 	u32 idx;
559 
560 	for (idx = 0; idx < (1U << old_log); idx++) {
561 		oroot = &old_array[idx];
562 		while ((op = rb_first(oroot)) != NULL) {
563 			rb_erase(op, oroot);
564 			of = container_of(op, struct fq_flow, fq_node);
565 			if (fq_gc_candidate(of)) {
566 				fcnt++;
567 				kmem_cache_free(fq_flow_cachep, of);
568 				continue;
569 			}
570 			nroot = &new_array[hash_32((u32)(long)of->sk, new_log)];
571 
572 			np = &nroot->rb_node;
573 			parent = NULL;
574 			while (*np) {
575 				parent = *np;
576 
577 				nf = container_of(parent, struct fq_flow, fq_node);
578 				BUG_ON(nf->sk == of->sk);
579 
580 				if (nf->sk > of->sk)
581 					np = &parent->rb_right;
582 				else
583 					np = &parent->rb_left;
584 			}
585 
586 			rb_link_node(&of->fq_node, parent, np);
587 			rb_insert_color(&of->fq_node, nroot);
588 		}
589 	}
590 	q->flows -= fcnt;
591 	q->inactive_flows -= fcnt;
592 	q->stat_gc_flows += fcnt;
593 }
594 
595 static void *fq_alloc_node(size_t sz, int node)
596 {
597 	void *ptr;
598 
599 	ptr = kmalloc_node(sz, GFP_KERNEL | __GFP_REPEAT | __GFP_NOWARN, node);
600 	if (!ptr)
601 		ptr = vmalloc_node(sz, node);
602 	return ptr;
603 }
604 
605 static void fq_free(void *addr)
606 {
607 	kvfree(addr);
608 }
609 
610 static int fq_resize(struct Qdisc *sch, u32 log)
611 {
612 	struct fq_sched_data *q = qdisc_priv(sch);
613 	struct rb_root *array;
614 	void *old_fq_root;
615 	u32 idx;
616 
617 	if (q->fq_root && log == q->fq_trees_log)
618 		return 0;
619 
620 	/* If XPS was setup, we can allocate memory on right NUMA node */
621 	array = fq_alloc_node(sizeof(struct rb_root) << log,
622 			      netdev_queue_numa_node_read(sch->dev_queue));
623 	if (!array)
624 		return -ENOMEM;
625 
626 	for (idx = 0; idx < (1U << log); idx++)
627 		array[idx] = RB_ROOT;
628 
629 	sch_tree_lock(sch);
630 
631 	old_fq_root = q->fq_root;
632 	if (old_fq_root)
633 		fq_rehash(q, old_fq_root, q->fq_trees_log, array, log);
634 
635 	q->fq_root = array;
636 	q->fq_trees_log = log;
637 
638 	sch_tree_unlock(sch);
639 
640 	fq_free(old_fq_root);
641 
642 	return 0;
643 }
644 
645 static const struct nla_policy fq_policy[TCA_FQ_MAX + 1] = {
646 	[TCA_FQ_PLIMIT]			= { .type = NLA_U32 },
647 	[TCA_FQ_FLOW_PLIMIT]		= { .type = NLA_U32 },
648 	[TCA_FQ_QUANTUM]		= { .type = NLA_U32 },
649 	[TCA_FQ_INITIAL_QUANTUM]	= { .type = NLA_U32 },
650 	[TCA_FQ_RATE_ENABLE]		= { .type = NLA_U32 },
651 	[TCA_FQ_FLOW_DEFAULT_RATE]	= { .type = NLA_U32 },
652 	[TCA_FQ_FLOW_MAX_RATE]		= { .type = NLA_U32 },
653 	[TCA_FQ_BUCKETS_LOG]		= { .type = NLA_U32 },
654 	[TCA_FQ_FLOW_REFILL_DELAY]	= { .type = NLA_U32 },
655 };
656 
657 static int fq_change(struct Qdisc *sch, struct nlattr *opt)
658 {
659 	struct fq_sched_data *q = qdisc_priv(sch);
660 	struct nlattr *tb[TCA_FQ_MAX + 1];
661 	int err, drop_count = 0;
662 	u32 fq_log;
663 
664 	if (!opt)
665 		return -EINVAL;
666 
667 	err = nla_parse_nested(tb, TCA_FQ_MAX, opt, fq_policy);
668 	if (err < 0)
669 		return err;
670 
671 	sch_tree_lock(sch);
672 
673 	fq_log = q->fq_trees_log;
674 
675 	if (tb[TCA_FQ_BUCKETS_LOG]) {
676 		u32 nval = nla_get_u32(tb[TCA_FQ_BUCKETS_LOG]);
677 
678 		if (nval >= 1 && nval <= ilog2(256*1024))
679 			fq_log = nval;
680 		else
681 			err = -EINVAL;
682 	}
683 	if (tb[TCA_FQ_PLIMIT])
684 		sch->limit = nla_get_u32(tb[TCA_FQ_PLIMIT]);
685 
686 	if (tb[TCA_FQ_FLOW_PLIMIT])
687 		q->flow_plimit = nla_get_u32(tb[TCA_FQ_FLOW_PLIMIT]);
688 
689 	if (tb[TCA_FQ_QUANTUM]) {
690 		u32 quantum = nla_get_u32(tb[TCA_FQ_QUANTUM]);
691 
692 		if (quantum > 0)
693 			q->quantum = quantum;
694 		else
695 			err = -EINVAL;
696 	}
697 
698 	if (tb[TCA_FQ_INITIAL_QUANTUM])
699 		q->initial_quantum = nla_get_u32(tb[TCA_FQ_INITIAL_QUANTUM]);
700 
701 	if (tb[TCA_FQ_FLOW_DEFAULT_RATE])
702 		pr_warn_ratelimited("sch_fq: defrate %u ignored.\n",
703 				    nla_get_u32(tb[TCA_FQ_FLOW_DEFAULT_RATE]));
704 
705 	if (tb[TCA_FQ_FLOW_MAX_RATE])
706 		q->flow_max_rate = nla_get_u32(tb[TCA_FQ_FLOW_MAX_RATE]);
707 
708 	if (tb[TCA_FQ_RATE_ENABLE]) {
709 		u32 enable = nla_get_u32(tb[TCA_FQ_RATE_ENABLE]);
710 
711 		if (enable <= 1)
712 			q->rate_enable = enable;
713 		else
714 			err = -EINVAL;
715 	}
716 
717 	if (tb[TCA_FQ_FLOW_REFILL_DELAY]) {
718 		u32 usecs_delay = nla_get_u32(tb[TCA_FQ_FLOW_REFILL_DELAY]) ;
719 
720 		q->flow_refill_delay = usecs_to_jiffies(usecs_delay);
721 	}
722 
723 	if (tb[TCA_FQ_ORPHAN_MASK])
724 		q->orphan_mask = nla_get_u32(tb[TCA_FQ_ORPHAN_MASK]);
725 
726 	if (!err) {
727 		sch_tree_unlock(sch);
728 		err = fq_resize(sch, fq_log);
729 		sch_tree_lock(sch);
730 	}
731 	while (sch->q.qlen > sch->limit) {
732 		struct sk_buff *skb = fq_dequeue(sch);
733 
734 		if (!skb)
735 			break;
736 		kfree_skb(skb);
737 		drop_count++;
738 	}
739 	qdisc_tree_decrease_qlen(sch, drop_count);
740 
741 	sch_tree_unlock(sch);
742 	return err;
743 }
744 
745 static void fq_destroy(struct Qdisc *sch)
746 {
747 	struct fq_sched_data *q = qdisc_priv(sch);
748 
749 	fq_reset(sch);
750 	fq_free(q->fq_root);
751 	qdisc_watchdog_cancel(&q->watchdog);
752 }
753 
754 static int fq_init(struct Qdisc *sch, struct nlattr *opt)
755 {
756 	struct fq_sched_data *q = qdisc_priv(sch);
757 	int err;
758 
759 	sch->limit		= 10000;
760 	q->flow_plimit		= 100;
761 	q->quantum		= 2 * psched_mtu(qdisc_dev(sch));
762 	q->initial_quantum	= 10 * psched_mtu(qdisc_dev(sch));
763 	q->flow_refill_delay	= msecs_to_jiffies(40);
764 	q->flow_max_rate	= ~0U;
765 	q->rate_enable		= 1;
766 	q->new_flows.first	= NULL;
767 	q->old_flows.first	= NULL;
768 	q->delayed		= RB_ROOT;
769 	q->fq_root		= NULL;
770 	q->fq_trees_log		= ilog2(1024);
771 	q->orphan_mask		= 1024 - 1;
772 	qdisc_watchdog_init(&q->watchdog, sch);
773 
774 	if (opt)
775 		err = fq_change(sch, opt);
776 	else
777 		err = fq_resize(sch, q->fq_trees_log);
778 
779 	return err;
780 }
781 
782 static int fq_dump(struct Qdisc *sch, struct sk_buff *skb)
783 {
784 	struct fq_sched_data *q = qdisc_priv(sch);
785 	struct nlattr *opts;
786 
787 	opts = nla_nest_start(skb, TCA_OPTIONS);
788 	if (opts == NULL)
789 		goto nla_put_failure;
790 
791 	/* TCA_FQ_FLOW_DEFAULT_RATE is not used anymore */
792 
793 	if (nla_put_u32(skb, TCA_FQ_PLIMIT, sch->limit) ||
794 	    nla_put_u32(skb, TCA_FQ_FLOW_PLIMIT, q->flow_plimit) ||
795 	    nla_put_u32(skb, TCA_FQ_QUANTUM, q->quantum) ||
796 	    nla_put_u32(skb, TCA_FQ_INITIAL_QUANTUM, q->initial_quantum) ||
797 	    nla_put_u32(skb, TCA_FQ_RATE_ENABLE, q->rate_enable) ||
798 	    nla_put_u32(skb, TCA_FQ_FLOW_MAX_RATE, q->flow_max_rate) ||
799 	    nla_put_u32(skb, TCA_FQ_FLOW_REFILL_DELAY,
800 			jiffies_to_usecs(q->flow_refill_delay)) ||
801 	    nla_put_u32(skb, TCA_FQ_ORPHAN_MASK, q->orphan_mask) ||
802 	    nla_put_u32(skb, TCA_FQ_BUCKETS_LOG, q->fq_trees_log))
803 		goto nla_put_failure;
804 
805 	return nla_nest_end(skb, opts);
806 
807 nla_put_failure:
808 	return -1;
809 }
810 
811 static int fq_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
812 {
813 	struct fq_sched_data *q = qdisc_priv(sch);
814 	u64 now = ktime_get_ns();
815 	struct tc_fq_qd_stats st = {
816 		.gc_flows		= q->stat_gc_flows,
817 		.highprio_packets	= q->stat_internal_packets,
818 		.tcp_retrans		= q->stat_tcp_retrans,
819 		.throttled		= q->stat_throttled,
820 		.flows_plimit		= q->stat_flows_plimit,
821 		.pkts_too_long		= q->stat_pkts_too_long,
822 		.allocation_errors	= q->stat_allocation_errors,
823 		.flows			= q->flows,
824 		.inactive_flows		= q->inactive_flows,
825 		.throttled_flows	= q->throttled_flows,
826 		.time_next_delayed_flow	= q->time_next_delayed_flow - now,
827 	};
828 
829 	return gnet_stats_copy_app(d, &st, sizeof(st));
830 }
831 
832 static struct Qdisc_ops fq_qdisc_ops __read_mostly = {
833 	.id		=	"fq",
834 	.priv_size	=	sizeof(struct fq_sched_data),
835 
836 	.enqueue	=	fq_enqueue,
837 	.dequeue	=	fq_dequeue,
838 	.peek		=	qdisc_peek_dequeued,
839 	.init		=	fq_init,
840 	.reset		=	fq_reset,
841 	.destroy	=	fq_destroy,
842 	.change		=	fq_change,
843 	.dump		=	fq_dump,
844 	.dump_stats	=	fq_dump_stats,
845 	.owner		=	THIS_MODULE,
846 };
847 
848 static int __init fq_module_init(void)
849 {
850 	int ret;
851 
852 	fq_flow_cachep = kmem_cache_create("fq_flow_cache",
853 					   sizeof(struct fq_flow),
854 					   0, 0, NULL);
855 	if (!fq_flow_cachep)
856 		return -ENOMEM;
857 
858 	ret = register_qdisc(&fq_qdisc_ops);
859 	if (ret)
860 		kmem_cache_destroy(fq_flow_cachep);
861 	return ret;
862 }
863 
864 static void __exit fq_module_exit(void)
865 {
866 	unregister_qdisc(&fq_qdisc_ops);
867 	kmem_cache_destroy(fq_flow_cachep);
868 }
869 
870 module_init(fq_module_init)
871 module_exit(fq_module_exit)
872 MODULE_AUTHOR("Eric Dumazet");
873 MODULE_LICENSE("GPL");
874