xref: /openbmc/linux/net/sched/sch_fq.c (revision aac5987a)
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 	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) */
118 static struct fq_flow detached, throttled;
119 
120 static void fq_flow_set_detached(struct fq_flow *f)
121 {
122 	f->next = &detached;
123 	f->age = jiffies;
124 }
125 
126 static bool fq_flow_is_detached(const struct fq_flow *f)
127 {
128 	return f->next == &detached;
129 }
130 
131 static void fq_flow_set_throttled(struct fq_sched_data *q, struct fq_flow *f)
132 {
133 	struct rb_node **p = &q->delayed.rb_node, *parent = NULL;
134 
135 	while (*p) {
136 		struct fq_flow *aux;
137 
138 		parent = *p;
139 		aux = rb_entry(parent, struct fq_flow, rate_node);
140 		if (f->time_next_packet >= aux->time_next_packet)
141 			p = &parent->rb_right;
142 		else
143 			p = &parent->rb_left;
144 	}
145 	rb_link_node(&f->rate_node, parent, p);
146 	rb_insert_color(&f->rate_node, &q->delayed);
147 	q->throttled_flows++;
148 	q->stat_throttled++;
149 
150 	f->next = &throttled;
151 	if (q->time_next_delayed_flow > f->time_next_packet)
152 		q->time_next_delayed_flow = f->time_next_packet;
153 }
154 
155 
156 static struct kmem_cache *fq_flow_cachep __read_mostly;
157 
158 static void fq_flow_add_tail(struct fq_flow_head *head, struct fq_flow *flow)
159 {
160 	if (head->first)
161 		head->last->next = flow;
162 	else
163 		head->first = flow;
164 	head->last = flow;
165 	flow->next = NULL;
166 }
167 
168 /* limit number of collected flows per round */
169 #define FQ_GC_MAX 8
170 #define FQ_GC_AGE (3*HZ)
171 
172 static bool fq_gc_candidate(const struct fq_flow *f)
173 {
174 	return fq_flow_is_detached(f) &&
175 	       time_after(jiffies, f->age + FQ_GC_AGE);
176 }
177 
178 static void fq_gc(struct fq_sched_data *q,
179 		  struct rb_root *root,
180 		  struct sock *sk)
181 {
182 	struct fq_flow *f, *tofree[FQ_GC_MAX];
183 	struct rb_node **p, *parent;
184 	int fcnt = 0;
185 
186 	p = &root->rb_node;
187 	parent = NULL;
188 	while (*p) {
189 		parent = *p;
190 
191 		f = rb_entry(parent, struct fq_flow, fq_node);
192 		if (f->sk == sk)
193 			break;
194 
195 		if (fq_gc_candidate(f)) {
196 			tofree[fcnt++] = f;
197 			if (fcnt == FQ_GC_MAX)
198 				break;
199 		}
200 
201 		if (f->sk > sk)
202 			p = &parent->rb_right;
203 		else
204 			p = &parent->rb_left;
205 	}
206 
207 	q->flows -= fcnt;
208 	q->inactive_flows -= fcnt;
209 	q->stat_gc_flows += fcnt;
210 	while (fcnt) {
211 		struct fq_flow *f = tofree[--fcnt];
212 
213 		rb_erase(&f->fq_node, root);
214 		kmem_cache_free(fq_flow_cachep, f);
215 	}
216 }
217 
218 static struct fq_flow *fq_classify(struct sk_buff *skb, struct fq_sched_data *q)
219 {
220 	struct rb_node **p, *parent;
221 	struct sock *sk = skb->sk;
222 	struct rb_root *root;
223 	struct fq_flow *f;
224 
225 	/* warning: no starvation prevention... */
226 	if (unlikely((skb->priority & TC_PRIO_MAX) == TC_PRIO_CONTROL))
227 		return &q->internal;
228 
229 	/* SYNACK messages are attached to a TCP_NEW_SYN_RECV request socket
230 	 * or a listener (SYNCOOKIE mode)
231 	 * 1) request sockets are not full blown,
232 	 *    they do not contain sk_pacing_rate
233 	 * 2) They are not part of a 'flow' yet
234 	 * 3) We do not want to rate limit them (eg SYNFLOOD attack),
235 	 *    especially if the listener set SO_MAX_PACING_RATE
236 	 * 4) We pretend they are orphaned
237 	 */
238 	if (!sk || sk_listener(sk)) {
239 		unsigned long hash = skb_get_hash(skb) & q->orphan_mask;
240 
241 		/* By forcing low order bit to 1, we make sure to not
242 		 * collide with a local flow (socket pointers are word aligned)
243 		 */
244 		sk = (struct sock *)((hash << 1) | 1UL);
245 		skb_orphan(skb);
246 	}
247 
248 	root = &q->fq_root[hash_ptr(sk, q->fq_trees_log)];
249 
250 	if (q->flows >= (2U << q->fq_trees_log) &&
251 	    q->inactive_flows > q->flows/2)
252 		fq_gc(q, root, sk);
253 
254 	p = &root->rb_node;
255 	parent = NULL;
256 	while (*p) {
257 		parent = *p;
258 
259 		f = rb_entry(parent, struct fq_flow, fq_node);
260 		if (f->sk == sk) {
261 			/* socket might have been reallocated, so check
262 			 * if its sk_hash is the same.
263 			 * It not, we need to refill credit with
264 			 * initial quantum
265 			 */
266 			if (unlikely(skb->sk &&
267 				     f->socket_hash != sk->sk_hash)) {
268 				f->credit = q->initial_quantum;
269 				f->socket_hash = sk->sk_hash;
270 				f->time_next_packet = 0ULL;
271 			}
272 			return f;
273 		}
274 		if (f->sk > sk)
275 			p = &parent->rb_right;
276 		else
277 			p = &parent->rb_left;
278 	}
279 
280 	f = kmem_cache_zalloc(fq_flow_cachep, GFP_ATOMIC | __GFP_NOWARN);
281 	if (unlikely(!f)) {
282 		q->stat_allocation_errors++;
283 		return &q->internal;
284 	}
285 	fq_flow_set_detached(f);
286 	f->sk = sk;
287 	if (skb->sk)
288 		f->socket_hash = sk->sk_hash;
289 	f->credit = q->initial_quantum;
290 
291 	rb_link_node(&f->fq_node, parent, p);
292 	rb_insert_color(&f->fq_node, root);
293 
294 	q->flows++;
295 	q->inactive_flows++;
296 	return f;
297 }
298 
299 
300 /* remove one skb from head of flow queue */
301 static struct sk_buff *fq_dequeue_head(struct Qdisc *sch, struct fq_flow *flow)
302 {
303 	struct sk_buff *skb = flow->head;
304 
305 	if (skb) {
306 		flow->head = skb->next;
307 		skb->next = NULL;
308 		flow->qlen--;
309 		qdisc_qstats_backlog_dec(sch, skb);
310 		sch->q.qlen--;
311 	}
312 	return skb;
313 }
314 
315 /* We might add in the future detection of retransmits
316  * For the time being, just return false
317  */
318 static bool skb_is_retransmit(struct sk_buff *skb)
319 {
320 	return false;
321 }
322 
323 /* add skb to flow queue
324  * flow queue is a linked list, kind of FIFO, except for TCP retransmits
325  * We special case tcp retransmits to be transmitted before other packets.
326  * We rely on fact that TCP retransmits are unlikely, so we do not waste
327  * a separate queue or a pointer.
328  * head->  [retrans pkt 1]
329  *         [retrans pkt 2]
330  *         [ normal pkt 1]
331  *         [ normal pkt 2]
332  *         [ normal pkt 3]
333  * tail->  [ normal pkt 4]
334  */
335 static void flow_queue_add(struct fq_flow *flow, struct sk_buff *skb)
336 {
337 	struct sk_buff *prev, *head = flow->head;
338 
339 	skb->next = NULL;
340 	if (!head) {
341 		flow->head = skb;
342 		flow->tail = skb;
343 		return;
344 	}
345 	if (likely(!skb_is_retransmit(skb))) {
346 		flow->tail->next = skb;
347 		flow->tail = skb;
348 		return;
349 	}
350 
351 	/* This skb is a tcp retransmit,
352 	 * find the last retrans packet in the queue
353 	 */
354 	prev = NULL;
355 	while (skb_is_retransmit(head)) {
356 		prev = head;
357 		head = head->next;
358 		if (!head)
359 			break;
360 	}
361 	if (!prev) { /* no rtx packet in queue, become the new head */
362 		skb->next = flow->head;
363 		flow->head = skb;
364 	} else {
365 		if (prev == flow->tail)
366 			flow->tail = skb;
367 		else
368 			skb->next = prev->next;
369 		prev->next = skb;
370 	}
371 }
372 
373 static int fq_enqueue(struct sk_buff *skb, struct Qdisc *sch,
374 		      struct sk_buff **to_free)
375 {
376 	struct fq_sched_data *q = qdisc_priv(sch);
377 	struct fq_flow *f;
378 
379 	if (unlikely(sch->q.qlen >= sch->limit))
380 		return qdisc_drop(skb, sch, to_free);
381 
382 	f = fq_classify(skb, q);
383 	if (unlikely(f->qlen >= q->flow_plimit && f != &q->internal)) {
384 		q->stat_flows_plimit++;
385 		return qdisc_drop(skb, sch, to_free);
386 	}
387 
388 	f->qlen++;
389 	if (skb_is_retransmit(skb))
390 		q->stat_tcp_retrans++;
391 	qdisc_qstats_backlog_inc(sch, skb);
392 	if (fq_flow_is_detached(f)) {
393 		fq_flow_add_tail(&q->new_flows, f);
394 		if (time_after(jiffies, f->age + q->flow_refill_delay))
395 			f->credit = max_t(u32, f->credit, q->quantum);
396 		q->inactive_flows--;
397 	}
398 
399 	/* Note: this overwrites f->age */
400 	flow_queue_add(f, skb);
401 
402 	if (unlikely(f == &q->internal)) {
403 		q->stat_internal_packets++;
404 	}
405 	sch->q.qlen++;
406 
407 	return NET_XMIT_SUCCESS;
408 }
409 
410 static void fq_check_throttled(struct fq_sched_data *q, u64 now)
411 {
412 	unsigned long sample;
413 	struct rb_node *p;
414 
415 	if (q->time_next_delayed_flow > now)
416 		return;
417 
418 	/* Update unthrottle latency EWMA.
419 	 * This is cheap and can help diagnosing timer/latency problems.
420 	 */
421 	sample = (unsigned long)(now - q->time_next_delayed_flow);
422 	q->unthrottle_latency_ns -= q->unthrottle_latency_ns >> 3;
423 	q->unthrottle_latency_ns += sample >> 3;
424 
425 	q->time_next_delayed_flow = ~0ULL;
426 	while ((p = rb_first(&q->delayed)) != NULL) {
427 		struct fq_flow *f = rb_entry(p, struct fq_flow, rate_node);
428 
429 		if (f->time_next_packet > now) {
430 			q->time_next_delayed_flow = f->time_next_packet;
431 			break;
432 		}
433 		rb_erase(p, &q->delayed);
434 		q->throttled_flows--;
435 		fq_flow_add_tail(&q->old_flows, f);
436 	}
437 }
438 
439 static struct sk_buff *fq_dequeue(struct Qdisc *sch)
440 {
441 	struct fq_sched_data *q = qdisc_priv(sch);
442 	u64 now = ktime_get_ns();
443 	struct fq_flow_head *head;
444 	struct sk_buff *skb;
445 	struct fq_flow *f;
446 	u32 rate, plen;
447 
448 	skb = fq_dequeue_head(sch, &q->internal);
449 	if (skb)
450 		goto out;
451 	fq_check_throttled(q, now);
452 begin:
453 	head = &q->new_flows;
454 	if (!head->first) {
455 		head = &q->old_flows;
456 		if (!head->first) {
457 			if (q->time_next_delayed_flow != ~0ULL)
458 				qdisc_watchdog_schedule_ns(&q->watchdog,
459 							   q->time_next_delayed_flow);
460 			return NULL;
461 		}
462 	}
463 	f = head->first;
464 
465 	if (f->credit <= 0) {
466 		f->credit += q->quantum;
467 		head->first = f->next;
468 		fq_flow_add_tail(&q->old_flows, f);
469 		goto begin;
470 	}
471 
472 	skb = f->head;
473 	if (unlikely(skb && now < f->time_next_packet &&
474 		     !skb_is_tcp_pure_ack(skb))) {
475 		head->first = f->next;
476 		fq_flow_set_throttled(q, f);
477 		goto begin;
478 	}
479 
480 	skb = fq_dequeue_head(sch, f);
481 	if (!skb) {
482 		head->first = f->next;
483 		/* force a pass through old_flows to prevent starvation */
484 		if ((head == &q->new_flows) && q->old_flows.first) {
485 			fq_flow_add_tail(&q->old_flows, f);
486 		} else {
487 			fq_flow_set_detached(f);
488 			q->inactive_flows++;
489 		}
490 		goto begin;
491 	}
492 	prefetch(&skb->end);
493 	f->credit -= qdisc_pkt_len(skb);
494 
495 	if (!q->rate_enable)
496 		goto out;
497 
498 	/* Do not pace locally generated ack packets */
499 	if (skb_is_tcp_pure_ack(skb))
500 		goto out;
501 
502 	rate = q->flow_max_rate;
503 	if (skb->sk)
504 		rate = min(skb->sk->sk_pacing_rate, rate);
505 
506 	if (rate <= q->low_rate_threshold) {
507 		f->credit = 0;
508 		plen = qdisc_pkt_len(skb);
509 	} else {
510 		plen = max(qdisc_pkt_len(skb), q->quantum);
511 		if (f->credit > 0)
512 			goto out;
513 	}
514 	if (rate != ~0U) {
515 		u64 len = (u64)plen * NSEC_PER_SEC;
516 
517 		if (likely(rate))
518 			do_div(len, rate);
519 		/* Since socket rate can change later,
520 		 * clamp the delay to 1 second.
521 		 * Really, providers of too big packets should be fixed !
522 		 */
523 		if (unlikely(len > NSEC_PER_SEC)) {
524 			len = NSEC_PER_SEC;
525 			q->stat_pkts_too_long++;
526 		}
527 		/* Account for schedule/timers drifts.
528 		 * f->time_next_packet was set when prior packet was sent,
529 		 * and current time (@now) can be too late by tens of us.
530 		 */
531 		if (f->time_next_packet)
532 			len -= min(len/2, now - f->time_next_packet);
533 		f->time_next_packet = now + len;
534 	}
535 out:
536 	qdisc_bstats_update(sch, skb);
537 	return skb;
538 }
539 
540 static void fq_flow_purge(struct fq_flow *flow)
541 {
542 	rtnl_kfree_skbs(flow->head, flow->tail);
543 	flow->head = NULL;
544 	flow->qlen = 0;
545 }
546 
547 static void fq_reset(struct Qdisc *sch)
548 {
549 	struct fq_sched_data *q = qdisc_priv(sch);
550 	struct rb_root *root;
551 	struct rb_node *p;
552 	struct fq_flow *f;
553 	unsigned int idx;
554 
555 	sch->q.qlen = 0;
556 	sch->qstats.backlog = 0;
557 
558 	fq_flow_purge(&q->internal);
559 
560 	if (!q->fq_root)
561 		return;
562 
563 	for (idx = 0; idx < (1U << q->fq_trees_log); idx++) {
564 		root = &q->fq_root[idx];
565 		while ((p = rb_first(root)) != NULL) {
566 			f = rb_entry(p, struct fq_flow, fq_node);
567 			rb_erase(p, root);
568 
569 			fq_flow_purge(f);
570 
571 			kmem_cache_free(fq_flow_cachep, f);
572 		}
573 	}
574 	q->new_flows.first	= NULL;
575 	q->old_flows.first	= NULL;
576 	q->delayed		= RB_ROOT;
577 	q->flows		= 0;
578 	q->inactive_flows	= 0;
579 	q->throttled_flows	= 0;
580 }
581 
582 static void fq_rehash(struct fq_sched_data *q,
583 		      struct rb_root *old_array, u32 old_log,
584 		      struct rb_root *new_array, u32 new_log)
585 {
586 	struct rb_node *op, **np, *parent;
587 	struct rb_root *oroot, *nroot;
588 	struct fq_flow *of, *nf;
589 	int fcnt = 0;
590 	u32 idx;
591 
592 	for (idx = 0; idx < (1U << old_log); idx++) {
593 		oroot = &old_array[idx];
594 		while ((op = rb_first(oroot)) != NULL) {
595 			rb_erase(op, oroot);
596 			of = rb_entry(op, struct fq_flow, fq_node);
597 			if (fq_gc_candidate(of)) {
598 				fcnt++;
599 				kmem_cache_free(fq_flow_cachep, of);
600 				continue;
601 			}
602 			nroot = &new_array[hash_ptr(of->sk, new_log)];
603 
604 			np = &nroot->rb_node;
605 			parent = NULL;
606 			while (*np) {
607 				parent = *np;
608 
609 				nf = rb_entry(parent, struct fq_flow, fq_node);
610 				BUG_ON(nf->sk == of->sk);
611 
612 				if (nf->sk > of->sk)
613 					np = &parent->rb_right;
614 				else
615 					np = &parent->rb_left;
616 			}
617 
618 			rb_link_node(&of->fq_node, parent, np);
619 			rb_insert_color(&of->fq_node, nroot);
620 		}
621 	}
622 	q->flows -= fcnt;
623 	q->inactive_flows -= fcnt;
624 	q->stat_gc_flows += fcnt;
625 }
626 
627 static void *fq_alloc_node(size_t sz, int node)
628 {
629 	void *ptr;
630 
631 	ptr = kmalloc_node(sz, GFP_KERNEL | __GFP_REPEAT | __GFP_NOWARN, node);
632 	if (!ptr)
633 		ptr = vmalloc_node(sz, node);
634 	return ptr;
635 }
636 
637 static void fq_free(void *addr)
638 {
639 	kvfree(addr);
640 }
641 
642 static int fq_resize(struct Qdisc *sch, u32 log)
643 {
644 	struct fq_sched_data *q = qdisc_priv(sch);
645 	struct rb_root *array;
646 	void *old_fq_root;
647 	u32 idx;
648 
649 	if (q->fq_root && log == q->fq_trees_log)
650 		return 0;
651 
652 	/* If XPS was setup, we can allocate memory on right NUMA node */
653 	array = fq_alloc_node(sizeof(struct rb_root) << log,
654 			      netdev_queue_numa_node_read(sch->dev_queue));
655 	if (!array)
656 		return -ENOMEM;
657 
658 	for (idx = 0; idx < (1U << log); idx++)
659 		array[idx] = RB_ROOT;
660 
661 	sch_tree_lock(sch);
662 
663 	old_fq_root = q->fq_root;
664 	if (old_fq_root)
665 		fq_rehash(q, old_fq_root, q->fq_trees_log, array, log);
666 
667 	q->fq_root = array;
668 	q->fq_trees_log = log;
669 
670 	sch_tree_unlock(sch);
671 
672 	fq_free(old_fq_root);
673 
674 	return 0;
675 }
676 
677 static const struct nla_policy fq_policy[TCA_FQ_MAX + 1] = {
678 	[TCA_FQ_PLIMIT]			= { .type = NLA_U32 },
679 	[TCA_FQ_FLOW_PLIMIT]		= { .type = NLA_U32 },
680 	[TCA_FQ_QUANTUM]		= { .type = NLA_U32 },
681 	[TCA_FQ_INITIAL_QUANTUM]	= { .type = NLA_U32 },
682 	[TCA_FQ_RATE_ENABLE]		= { .type = NLA_U32 },
683 	[TCA_FQ_FLOW_DEFAULT_RATE]	= { .type = NLA_U32 },
684 	[TCA_FQ_FLOW_MAX_RATE]		= { .type = NLA_U32 },
685 	[TCA_FQ_BUCKETS_LOG]		= { .type = NLA_U32 },
686 	[TCA_FQ_FLOW_REFILL_DELAY]	= { .type = NLA_U32 },
687 	[TCA_FQ_LOW_RATE_THRESHOLD]	= { .type = NLA_U32 },
688 };
689 
690 static int fq_change(struct Qdisc *sch, struct nlattr *opt)
691 {
692 	struct fq_sched_data *q = qdisc_priv(sch);
693 	struct nlattr *tb[TCA_FQ_MAX + 1];
694 	int err, drop_count = 0;
695 	unsigned drop_len = 0;
696 	u32 fq_log;
697 
698 	if (!opt)
699 		return -EINVAL;
700 
701 	err = nla_parse_nested(tb, TCA_FQ_MAX, opt, fq_policy);
702 	if (err < 0)
703 		return err;
704 
705 	sch_tree_lock(sch);
706 
707 	fq_log = q->fq_trees_log;
708 
709 	if (tb[TCA_FQ_BUCKETS_LOG]) {
710 		u32 nval = nla_get_u32(tb[TCA_FQ_BUCKETS_LOG]);
711 
712 		if (nval >= 1 && nval <= ilog2(256*1024))
713 			fq_log = nval;
714 		else
715 			err = -EINVAL;
716 	}
717 	if (tb[TCA_FQ_PLIMIT])
718 		sch->limit = nla_get_u32(tb[TCA_FQ_PLIMIT]);
719 
720 	if (tb[TCA_FQ_FLOW_PLIMIT])
721 		q->flow_plimit = nla_get_u32(tb[TCA_FQ_FLOW_PLIMIT]);
722 
723 	if (tb[TCA_FQ_QUANTUM]) {
724 		u32 quantum = nla_get_u32(tb[TCA_FQ_QUANTUM]);
725 
726 		if (quantum > 0)
727 			q->quantum = quantum;
728 		else
729 			err = -EINVAL;
730 	}
731 
732 	if (tb[TCA_FQ_INITIAL_QUANTUM])
733 		q->initial_quantum = nla_get_u32(tb[TCA_FQ_INITIAL_QUANTUM]);
734 
735 	if (tb[TCA_FQ_FLOW_DEFAULT_RATE])
736 		pr_warn_ratelimited("sch_fq: defrate %u ignored.\n",
737 				    nla_get_u32(tb[TCA_FQ_FLOW_DEFAULT_RATE]));
738 
739 	if (tb[TCA_FQ_FLOW_MAX_RATE])
740 		q->flow_max_rate = nla_get_u32(tb[TCA_FQ_FLOW_MAX_RATE]);
741 
742 	if (tb[TCA_FQ_LOW_RATE_THRESHOLD])
743 		q->low_rate_threshold =
744 			nla_get_u32(tb[TCA_FQ_LOW_RATE_THRESHOLD]);
745 
746 	if (tb[TCA_FQ_RATE_ENABLE]) {
747 		u32 enable = nla_get_u32(tb[TCA_FQ_RATE_ENABLE]);
748 
749 		if (enable <= 1)
750 			q->rate_enable = enable;
751 		else
752 			err = -EINVAL;
753 	}
754 
755 	if (tb[TCA_FQ_FLOW_REFILL_DELAY]) {
756 		u32 usecs_delay = nla_get_u32(tb[TCA_FQ_FLOW_REFILL_DELAY]) ;
757 
758 		q->flow_refill_delay = usecs_to_jiffies(usecs_delay);
759 	}
760 
761 	if (tb[TCA_FQ_ORPHAN_MASK])
762 		q->orphan_mask = nla_get_u32(tb[TCA_FQ_ORPHAN_MASK]);
763 
764 	if (!err) {
765 		sch_tree_unlock(sch);
766 		err = fq_resize(sch, fq_log);
767 		sch_tree_lock(sch);
768 	}
769 	while (sch->q.qlen > sch->limit) {
770 		struct sk_buff *skb = fq_dequeue(sch);
771 
772 		if (!skb)
773 			break;
774 		drop_len += qdisc_pkt_len(skb);
775 		rtnl_kfree_skbs(skb, skb);
776 		drop_count++;
777 	}
778 	qdisc_tree_reduce_backlog(sch, drop_count, drop_len);
779 
780 	sch_tree_unlock(sch);
781 	return err;
782 }
783 
784 static void fq_destroy(struct Qdisc *sch)
785 {
786 	struct fq_sched_data *q = qdisc_priv(sch);
787 
788 	fq_reset(sch);
789 	fq_free(q->fq_root);
790 	qdisc_watchdog_cancel(&q->watchdog);
791 }
792 
793 static int fq_init(struct Qdisc *sch, struct nlattr *opt)
794 {
795 	struct fq_sched_data *q = qdisc_priv(sch);
796 	int err;
797 
798 	sch->limit		= 10000;
799 	q->flow_plimit		= 100;
800 	q->quantum		= 2 * psched_mtu(qdisc_dev(sch));
801 	q->initial_quantum	= 10 * psched_mtu(qdisc_dev(sch));
802 	q->flow_refill_delay	= msecs_to_jiffies(40);
803 	q->flow_max_rate	= ~0U;
804 	q->time_next_delayed_flow = ~0ULL;
805 	q->rate_enable		= 1;
806 	q->new_flows.first	= NULL;
807 	q->old_flows.first	= NULL;
808 	q->delayed		= RB_ROOT;
809 	q->fq_root		= NULL;
810 	q->fq_trees_log		= ilog2(1024);
811 	q->orphan_mask		= 1024 - 1;
812 	q->low_rate_threshold	= 550000 / 8;
813 	qdisc_watchdog_init(&q->watchdog, sch);
814 
815 	if (opt)
816 		err = fq_change(sch, opt);
817 	else
818 		err = fq_resize(sch, q->fq_trees_log);
819 
820 	return err;
821 }
822 
823 static int fq_dump(struct Qdisc *sch, struct sk_buff *skb)
824 {
825 	struct fq_sched_data *q = qdisc_priv(sch);
826 	struct nlattr *opts;
827 
828 	opts = nla_nest_start(skb, TCA_OPTIONS);
829 	if (opts == NULL)
830 		goto nla_put_failure;
831 
832 	/* TCA_FQ_FLOW_DEFAULT_RATE is not used anymore */
833 
834 	if (nla_put_u32(skb, TCA_FQ_PLIMIT, sch->limit) ||
835 	    nla_put_u32(skb, TCA_FQ_FLOW_PLIMIT, q->flow_plimit) ||
836 	    nla_put_u32(skb, TCA_FQ_QUANTUM, q->quantum) ||
837 	    nla_put_u32(skb, TCA_FQ_INITIAL_QUANTUM, q->initial_quantum) ||
838 	    nla_put_u32(skb, TCA_FQ_RATE_ENABLE, q->rate_enable) ||
839 	    nla_put_u32(skb, TCA_FQ_FLOW_MAX_RATE, q->flow_max_rate) ||
840 	    nla_put_u32(skb, TCA_FQ_FLOW_REFILL_DELAY,
841 			jiffies_to_usecs(q->flow_refill_delay)) ||
842 	    nla_put_u32(skb, TCA_FQ_ORPHAN_MASK, q->orphan_mask) ||
843 	    nla_put_u32(skb, TCA_FQ_LOW_RATE_THRESHOLD,
844 			q->low_rate_threshold) ||
845 	    nla_put_u32(skb, TCA_FQ_BUCKETS_LOG, q->fq_trees_log))
846 		goto nla_put_failure;
847 
848 	return nla_nest_end(skb, opts);
849 
850 nla_put_failure:
851 	return -1;
852 }
853 
854 static int fq_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
855 {
856 	struct fq_sched_data *q = qdisc_priv(sch);
857 	struct tc_fq_qd_stats st;
858 
859 	sch_tree_lock(sch);
860 
861 	st.gc_flows		  = q->stat_gc_flows;
862 	st.highprio_packets	  = q->stat_internal_packets;
863 	st.tcp_retrans		  = q->stat_tcp_retrans;
864 	st.throttled		  = q->stat_throttled;
865 	st.flows_plimit		  = q->stat_flows_plimit;
866 	st.pkts_too_long	  = q->stat_pkts_too_long;
867 	st.allocation_errors	  = q->stat_allocation_errors;
868 	st.time_next_delayed_flow = q->time_next_delayed_flow - ktime_get_ns();
869 	st.flows		  = q->flows;
870 	st.inactive_flows	  = q->inactive_flows;
871 	st.throttled_flows	  = q->throttled_flows;
872 	st.unthrottle_latency_ns  = min_t(unsigned long,
873 					  q->unthrottle_latency_ns, ~0U);
874 	sch_tree_unlock(sch);
875 
876 	return gnet_stats_copy_app(d, &st, sizeof(st));
877 }
878 
879 static struct Qdisc_ops fq_qdisc_ops __read_mostly = {
880 	.id		=	"fq",
881 	.priv_size	=	sizeof(struct fq_sched_data),
882 
883 	.enqueue	=	fq_enqueue,
884 	.dequeue	=	fq_dequeue,
885 	.peek		=	qdisc_peek_dequeued,
886 	.init		=	fq_init,
887 	.reset		=	fq_reset,
888 	.destroy	=	fq_destroy,
889 	.change		=	fq_change,
890 	.dump		=	fq_dump,
891 	.dump_stats	=	fq_dump_stats,
892 	.owner		=	THIS_MODULE,
893 };
894 
895 static int __init fq_module_init(void)
896 {
897 	int ret;
898 
899 	fq_flow_cachep = kmem_cache_create("fq_flow_cache",
900 					   sizeof(struct fq_flow),
901 					   0, 0, NULL);
902 	if (!fq_flow_cachep)
903 		return -ENOMEM;
904 
905 	ret = register_qdisc(&fq_qdisc_ops);
906 	if (ret)
907 		kmem_cache_destroy(fq_flow_cachep);
908 	return ret;
909 }
910 
911 static void __exit fq_module_exit(void)
912 {
913 	unregister_qdisc(&fq_qdisc_ops);
914 	kmem_cache_destroy(fq_flow_cachep);
915 }
916 
917 module_init(fq_module_init)
918 module_exit(fq_module_exit)
919 MODULE_AUTHOR("Eric Dumazet");
920 MODULE_LICENSE("GPL");
921