Linux Audio

Check our new training course

Linux BSP upgrade and security maintenance

Need help to get security updates for your Linux BSP?
Loading...
v3.5.6
 
  1/*
  2 * Fair Queue CoDel discipline
  3 *
  4 *	This program is free software; you can redistribute it and/or
  5 *	modify it under the terms of the GNU General Public License
  6 *	as published by the Free Software Foundation; either version
  7 *	2 of the License, or (at your option) any later version.
  8 *
  9 *  Copyright (C) 2012 Eric Dumazet <edumazet@google.com>
 10 */
 11
 12#include <linux/module.h>
 13#include <linux/types.h>
 14#include <linux/kernel.h>
 15#include <linux/jiffies.h>
 16#include <linux/string.h>
 17#include <linux/in.h>
 18#include <linux/errno.h>
 19#include <linux/init.h>
 20#include <linux/skbuff.h>
 21#include <linux/jhash.h>
 22#include <linux/slab.h>
 23#include <linux/vmalloc.h>
 24#include <net/netlink.h>
 25#include <net/pkt_sched.h>
 26#include <net/flow_keys.h>
 27#include <net/codel.h>
 
 
 28
 29/*	Fair Queue CoDel.
 30 *
 31 * Principles :
 32 * Packets are classified (internal classifier or external) on flows.
 33 * This is a Stochastic model (as we use a hash, several flows
 34 *			       might be hashed on same slot)
 35 * Each flow has a CoDel managed queue.
 36 * Flows are linked onto two (Round Robin) lists,
 37 * so that new flows have priority on old ones.
 38 *
 39 * For a given flow, packets are not reordered (CoDel uses a FIFO)
 40 * head drops only.
 41 * ECN capability is on by default.
 42 * Low memory footprint (64 bytes per flow)
 43 */
 44
 45struct fq_codel_flow {
 46	struct sk_buff	  *head;
 47	struct sk_buff	  *tail;
 48	struct list_head  flowchain;
 49	int		  deficit;
 50	u32		  dropped; /* number of drops (or ECN marks) on this flow */
 51	struct codel_vars cvars;
 52}; /* please try to keep this structure <= 64 bytes */
 53
 54struct fq_codel_sched_data {
 55	struct tcf_proto *filter_list;	/* optional external classifier */
 
 56	struct fq_codel_flow *flows;	/* Flows table [flows_cnt] */
 57	u32		*backlogs;	/* backlog table [flows_cnt] */
 58	u32		flows_cnt;	/* number of flows */
 59	u32		perturbation;	/* hash perturbation */
 60	u32		quantum;	/* psched_mtu(qdisc_dev(sch)); */
 
 
 61	struct codel_params cparams;
 62	struct codel_stats cstats;
 
 
 63	u32		drop_overlimit;
 64	u32		new_flow_count;
 65
 66	struct list_head new_flows;	/* list of new flows */
 67	struct list_head old_flows;	/* list of old flows */
 68};
 69
 70static unsigned int fq_codel_hash(const struct fq_codel_sched_data *q,
 71				  const struct sk_buff *skb)
 72{
 73	struct flow_keys keys;
 74	unsigned int hash;
 75
 76	skb_flow_dissect(skb, &keys);
 77	hash = jhash_3words((__force u32)keys.dst,
 78			    (__force u32)keys.src ^ keys.ip_proto,
 79			    (__force u32)keys.ports, q->perturbation);
 80	return ((u64)hash * q->flows_cnt) >> 32;
 81}
 82
 83static unsigned int fq_codel_classify(struct sk_buff *skb, struct Qdisc *sch,
 84				      int *qerr)
 85{
 86	struct fq_codel_sched_data *q = qdisc_priv(sch);
 
 87	struct tcf_result res;
 88	int result;
 89
 90	if (TC_H_MAJ(skb->priority) == sch->handle &&
 91	    TC_H_MIN(skb->priority) > 0 &&
 92	    TC_H_MIN(skb->priority) <= q->flows_cnt)
 93		return TC_H_MIN(skb->priority);
 94
 95	if (!q->filter_list)
 
 96		return fq_codel_hash(q, skb) + 1;
 97
 98	*qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
 99	result = tc_classify(skb, q->filter_list, &res);
100	if (result >= 0) {
101#ifdef CONFIG_NET_CLS_ACT
102		switch (result) {
103		case TC_ACT_STOLEN:
104		case TC_ACT_QUEUED:
 
105			*qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
 
106		case TC_ACT_SHOT:
107			return 0;
108		}
109#endif
110		if (TC_H_MIN(res.classid) <= q->flows_cnt)
111			return TC_H_MIN(res.classid);
112	}
113	return 0;
114}
115
116/* helper functions : might be changed when/if skb use a standard list_head */
117
118/* remove one skb from head of slot queue */
119static inline struct sk_buff *dequeue_head(struct fq_codel_flow *flow)
120{
121	struct sk_buff *skb = flow->head;
122
123	flow->head = skb->next;
124	skb->next = NULL;
125	return skb;
126}
127
128/* add skb to flow queue (tail add) */
129static inline void flow_queue_add(struct fq_codel_flow *flow,
130				  struct sk_buff *skb)
131{
132	if (flow->head == NULL)
133		flow->head = skb;
134	else
135		flow->tail->next = skb;
136	flow->tail = skb;
137	skb->next = NULL;
138}
139
140static unsigned int fq_codel_drop(struct Qdisc *sch)
 
141{
142	struct fq_codel_sched_data *q = qdisc_priv(sch);
143	struct sk_buff *skb;
144	unsigned int maxbacklog = 0, idx = 0, i, len;
145	struct fq_codel_flow *flow;
 
 
146
147	/* Queue is full! Find the fat flow and drop packet from it.
148	 * This might sound expensive, but with 1024 flows, we scan
149	 * 4KB of memory, and we dont need to handle a complex tree
150	 * in fast path (packet queue/enqueue) with many cache misses.
 
 
151	 */
152	for (i = 0; i < q->flows_cnt; i++) {
153		if (q->backlogs[i] > maxbacklog) {
154			maxbacklog = q->backlogs[i];
155			idx = i;
156		}
157	}
 
 
 
 
158	flow = &q->flows[idx];
159	skb = dequeue_head(flow);
160	len = qdisc_pkt_len(skb);
 
 
 
 
 
 
 
 
 
161	q->backlogs[idx] -= len;
162	kfree_skb(skb);
163	sch->q.qlen--;
164	sch->qstats.drops++;
165	sch->qstats.backlog -= len;
166	flow->dropped++;
167	return idx;
168}
169
170static int fq_codel_enqueue(struct sk_buff *skb, struct Qdisc *sch)
 
171{
172	struct fq_codel_sched_data *q = qdisc_priv(sch);
173	unsigned int idx;
174	struct fq_codel_flow *flow;
175	int uninitialized_var(ret);
 
 
176
177	idx = fq_codel_classify(skb, sch, &ret);
178	if (idx == 0) {
179		if (ret & __NET_XMIT_BYPASS)
180			sch->qstats.drops++;
181		kfree_skb(skb);
182		return ret;
183	}
184	idx--;
185
186	codel_set_enqueue_time(skb);
187	flow = &q->flows[idx];
188	flow_queue_add(flow, skb);
189	q->backlogs[idx] += qdisc_pkt_len(skb);
190	sch->qstats.backlog += qdisc_pkt_len(skb);
191
192	if (list_empty(&flow->flowchain)) {
193		list_add_tail(&flow->flowchain, &q->new_flows);
194		codel_vars_init(&flow->cvars);
195		q->new_flow_count++;
196		flow->deficit = q->quantum;
197		flow->dropped = 0;
198	}
199	if (++sch->q.qlen < sch->limit)
 
 
 
200		return NET_XMIT_SUCCESS;
201
202	q->drop_overlimit++;
203	/* Return Congestion Notification only if we dropped a packet
204	 * from this flow.
 
 
 
 
 
 
205	 */
206	if (fq_codel_drop(sch) == idx)
207		return NET_XMIT_CN;
208
209	/* As we dropped a packet, better let upper stack know this */
210	qdisc_tree_decrease_qlen(sch, 1);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
211	return NET_XMIT_SUCCESS;
212}
213
214/* This is the specific function called from codel_dequeue()
215 * to dequeue a packet from queue. Note: backlog is handled in
216 * codel, we dont need to reduce it here.
217 */
218static struct sk_buff *dequeue(struct codel_vars *vars, struct Qdisc *sch)
219{
 
220	struct fq_codel_sched_data *q = qdisc_priv(sch);
221	struct fq_codel_flow *flow;
222	struct sk_buff *skb = NULL;
223
224	flow = container_of(vars, struct fq_codel_flow, cvars);
225	if (flow->head) {
226		skb = dequeue_head(flow);
227		q->backlogs[flow - q->flows] -= qdisc_pkt_len(skb);
 
228		sch->q.qlen--;
 
229	}
230	return skb;
231}
232
 
 
 
 
 
 
 
 
233static struct sk_buff *fq_codel_dequeue(struct Qdisc *sch)
234{
235	struct fq_codel_sched_data *q = qdisc_priv(sch);
236	struct sk_buff *skb;
237	struct fq_codel_flow *flow;
238	struct list_head *head;
239	u32 prev_drop_count, prev_ecn_mark;
240
241begin:
242	head = &q->new_flows;
243	if (list_empty(head)) {
244		head = &q->old_flows;
245		if (list_empty(head))
246			return NULL;
247	}
248	flow = list_first_entry(head, struct fq_codel_flow, flowchain);
249
250	if (flow->deficit <= 0) {
251		flow->deficit += q->quantum;
252		list_move_tail(&flow->flowchain, &q->old_flows);
253		goto begin;
254	}
255
256	prev_drop_count = q->cstats.drop_count;
257	prev_ecn_mark = q->cstats.ecn_mark;
258
259	skb = codel_dequeue(sch, &q->cparams, &flow->cvars, &q->cstats,
260			    dequeue);
261
262	flow->dropped += q->cstats.drop_count - prev_drop_count;
263	flow->dropped += q->cstats.ecn_mark - prev_ecn_mark;
264
265	if (!skb) {
266		/* force a pass through old_flows to prevent starvation */
267		if ((head == &q->new_flows) && !list_empty(&q->old_flows))
268			list_move_tail(&flow->flowchain, &q->old_flows);
269		else
270			list_del_init(&flow->flowchain);
271		goto begin;
272	}
273	qdisc_bstats_update(sch, skb);
274	flow->deficit -= qdisc_pkt_len(skb);
275	/* We cant call qdisc_tree_decrease_qlen() if our qlen is 0,
276	 * or HTB crashes. Defer it for next round.
277	 */
278	if (q->cstats.drop_count && sch->q.qlen) {
279		qdisc_tree_decrease_qlen(sch, q->cstats.drop_count);
 
280		q->cstats.drop_count = 0;
 
281	}
282	return skb;
283}
284
 
 
 
 
 
 
285static void fq_codel_reset(struct Qdisc *sch)
286{
287	struct sk_buff *skb;
 
288
289	while ((skb = fq_codel_dequeue(sch)) != NULL)
290		kfree_skb(skb);
 
 
 
 
 
 
 
 
 
 
 
291}
292
293static const struct nla_policy fq_codel_policy[TCA_FQ_CODEL_MAX + 1] = {
294	[TCA_FQ_CODEL_TARGET]	= { .type = NLA_U32 },
295	[TCA_FQ_CODEL_LIMIT]	= { .type = NLA_U32 },
296	[TCA_FQ_CODEL_INTERVAL]	= { .type = NLA_U32 },
297	[TCA_FQ_CODEL_ECN]	= { .type = NLA_U32 },
298	[TCA_FQ_CODEL_FLOWS]	= { .type = NLA_U32 },
299	[TCA_FQ_CODEL_QUANTUM]	= { .type = NLA_U32 },
 
 
 
300};
301
302static int fq_codel_change(struct Qdisc *sch, struct nlattr *opt)
 
303{
304	struct fq_codel_sched_data *q = qdisc_priv(sch);
305	struct nlattr *tb[TCA_FQ_CODEL_MAX + 1];
 
306	int err;
307
308	if (!opt)
309		return -EINVAL;
310
311	err = nla_parse_nested(tb, TCA_FQ_CODEL_MAX, opt, fq_codel_policy);
 
312	if (err < 0)
313		return err;
314	if (tb[TCA_FQ_CODEL_FLOWS]) {
315		if (q->flows)
316			return -EINVAL;
317		q->flows_cnt = nla_get_u32(tb[TCA_FQ_CODEL_FLOWS]);
318		if (!q->flows_cnt ||
319		    q->flows_cnt > 65536)
320			return -EINVAL;
321	}
 
 
 
 
 
 
 
322	sch_tree_lock(sch);
323
324	if (tb[TCA_FQ_CODEL_TARGET]) {
325		u64 target = nla_get_u32(tb[TCA_FQ_CODEL_TARGET]);
326
327		q->cparams.target = (target * NSEC_PER_USEC) >> CODEL_SHIFT;
328	}
329
 
 
 
 
 
 
330	if (tb[TCA_FQ_CODEL_INTERVAL]) {
331		u64 interval = nla_get_u32(tb[TCA_FQ_CODEL_INTERVAL]);
332
333		q->cparams.interval = (interval * NSEC_PER_USEC) >> CODEL_SHIFT;
334	}
335
336	if (tb[TCA_FQ_CODEL_LIMIT])
337		sch->limit = nla_get_u32(tb[TCA_FQ_CODEL_LIMIT]);
338
339	if (tb[TCA_FQ_CODEL_ECN])
340		q->cparams.ecn = !!nla_get_u32(tb[TCA_FQ_CODEL_ECN]);
341
342	if (tb[TCA_FQ_CODEL_QUANTUM])
343		q->quantum = max(256U, nla_get_u32(tb[TCA_FQ_CODEL_QUANTUM]));
 
 
 
 
 
 
344
345	while (sch->q.qlen > sch->limit) {
 
346		struct sk_buff *skb = fq_codel_dequeue(sch);
347
348		kfree_skb(skb);
 
349		q->cstats.drop_count++;
350	}
351	qdisc_tree_decrease_qlen(sch, q->cstats.drop_count);
352	q->cstats.drop_count = 0;
 
353
354	sch_tree_unlock(sch);
355	return 0;
356}
357
358static void *fq_codel_zalloc(size_t sz)
359{
360	void *ptr = kzalloc(sz, GFP_KERNEL | __GFP_NOWARN);
361
362	if (!ptr)
363		ptr = vzalloc(sz);
364	return ptr;
365}
366
367static void fq_codel_free(void *addr)
368{
369	if (addr) {
370		if (is_vmalloc_addr(addr))
371			vfree(addr);
372		else
373			kfree(addr);
374	}
375}
376
377static void fq_codel_destroy(struct Qdisc *sch)
378{
379	struct fq_codel_sched_data *q = qdisc_priv(sch);
380
381	tcf_destroy_chain(&q->filter_list);
382	fq_codel_free(q->backlogs);
383	fq_codel_free(q->flows);
384}
385
386static int fq_codel_init(struct Qdisc *sch, struct nlattr *opt)
 
387{
388	struct fq_codel_sched_data *q = qdisc_priv(sch);
389	int i;
 
390
391	sch->limit = 10*1024;
392	q->flows_cnt = 1024;
 
 
393	q->quantum = psched_mtu(qdisc_dev(sch));
394	q->perturbation = net_random();
395	INIT_LIST_HEAD(&q->new_flows);
396	INIT_LIST_HEAD(&q->old_flows);
397	codel_params_init(&q->cparams);
398	codel_stats_init(&q->cstats);
399	q->cparams.ecn = true;
 
400
401	if (opt) {
402		int err = fq_codel_change(sch, opt);
403		if (err)
404			return err;
405	}
406
 
 
 
 
407	if (!q->flows) {
408		q->flows = fq_codel_zalloc(q->flows_cnt *
409					   sizeof(struct fq_codel_flow));
410		if (!q->flows)
411			return -ENOMEM;
412		q->backlogs = fq_codel_zalloc(q->flows_cnt * sizeof(u32));
 
 
 
413		if (!q->backlogs) {
414			fq_codel_free(q->flows);
415			return -ENOMEM;
416		}
417		for (i = 0; i < q->flows_cnt; i++) {
418			struct fq_codel_flow *flow = q->flows + i;
419
420			INIT_LIST_HEAD(&flow->flowchain);
 
421		}
422	}
423	if (sch->limit >= 1)
424		sch->flags |= TCQ_F_CAN_BYPASS;
425	else
426		sch->flags &= ~TCQ_F_CAN_BYPASS;
427	return 0;
 
 
 
 
 
 
 
428}
429
430static int fq_codel_dump(struct Qdisc *sch, struct sk_buff *skb)
431{
432	struct fq_codel_sched_data *q = qdisc_priv(sch);
433	struct nlattr *opts;
434
435	opts = nla_nest_start(skb, TCA_OPTIONS);
436	if (opts == NULL)
437		goto nla_put_failure;
438
439	if (nla_put_u32(skb, TCA_FQ_CODEL_TARGET,
440			codel_time_to_us(q->cparams.target)) ||
441	    nla_put_u32(skb, TCA_FQ_CODEL_LIMIT,
442			sch->limit) ||
443	    nla_put_u32(skb, TCA_FQ_CODEL_INTERVAL,
444			codel_time_to_us(q->cparams.interval)) ||
445	    nla_put_u32(skb, TCA_FQ_CODEL_ECN,
446			q->cparams.ecn) ||
447	    nla_put_u32(skb, TCA_FQ_CODEL_QUANTUM,
448			q->quantum) ||
 
 
 
 
449	    nla_put_u32(skb, TCA_FQ_CODEL_FLOWS,
450			q->flows_cnt))
451		goto nla_put_failure;
452
453	nla_nest_end(skb, opts);
454	return skb->len;
 
 
 
 
455
456nla_put_failure:
457	return -1;
458}
459
460static int fq_codel_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
461{
462	struct fq_codel_sched_data *q = qdisc_priv(sch);
463	struct tc_fq_codel_xstats st = {
464		.type				= TCA_FQ_CODEL_XSTATS_QDISC,
465	};
466	struct list_head *pos;
467
468	st.qdisc_stats.maxpacket = q->cstats.maxpacket;
469	st.qdisc_stats.drop_overlimit = q->drop_overlimit;
470	st.qdisc_stats.ecn_mark = q->cstats.ecn_mark;
471	st.qdisc_stats.new_flow_count = q->new_flow_count;
 
 
 
472
 
473	list_for_each(pos, &q->new_flows)
474		st.qdisc_stats.new_flows_len++;
475
476	list_for_each(pos, &q->old_flows)
477		st.qdisc_stats.old_flows_len++;
 
478
479	return gnet_stats_copy_app(d, &st, sizeof(st));
480}
481
482static struct Qdisc *fq_codel_leaf(struct Qdisc *sch, unsigned long arg)
483{
484	return NULL;
485}
486
487static unsigned long fq_codel_get(struct Qdisc *sch, u32 classid)
488{
489	return 0;
490}
491
492static unsigned long fq_codel_bind(struct Qdisc *sch, unsigned long parent,
493			      u32 classid)
494{
495	/* we cannot bypass queue discipline anymore */
496	sch->flags &= ~TCQ_F_CAN_BYPASS;
497	return 0;
498}
499
500static void fq_codel_put(struct Qdisc *q, unsigned long cl)
501{
502}
503
504static struct tcf_proto **fq_codel_find_tcf(struct Qdisc *sch, unsigned long cl)
 
505{
506	struct fq_codel_sched_data *q = qdisc_priv(sch);
507
508	if (cl)
509		return NULL;
510	return &q->filter_list;
511}
512
513static int fq_codel_dump_class(struct Qdisc *sch, unsigned long cl,
514			  struct sk_buff *skb, struct tcmsg *tcm)
515{
516	tcm->tcm_handle |= TC_H_MIN(cl);
517	return 0;
518}
519
520static int fq_codel_dump_class_stats(struct Qdisc *sch, unsigned long cl,
521				     struct gnet_dump *d)
522{
523	struct fq_codel_sched_data *q = qdisc_priv(sch);
524	u32 idx = cl - 1;
525	struct gnet_stats_queue qs = { 0 };
526	struct tc_fq_codel_xstats xstats;
527
528	if (idx < q->flows_cnt) {
529		const struct fq_codel_flow *flow = &q->flows[idx];
530		const struct sk_buff *skb = flow->head;
531
532		memset(&xstats, 0, sizeof(xstats));
533		xstats.type = TCA_FQ_CODEL_XSTATS_CLASS;
534		xstats.class_stats.deficit = flow->deficit;
535		xstats.class_stats.ldelay =
536			codel_time_to_us(flow->cvars.ldelay);
537		xstats.class_stats.count = flow->cvars.count;
538		xstats.class_stats.lastcount = flow->cvars.lastcount;
539		xstats.class_stats.dropping = flow->cvars.dropping;
540		if (flow->cvars.dropping) {
541			codel_tdiff_t delta = flow->cvars.drop_next -
542					      codel_get_time();
543
544			xstats.class_stats.drop_next = (delta >= 0) ?
545				codel_time_to_us(delta) :
546				-codel_time_to_us(-delta);
547		}
548		while (skb) {
549			qs.qlen++;
550			skb = skb->next;
 
 
 
 
 
551		}
552		qs.backlog = q->backlogs[idx];
553		qs.drops = flow->dropped;
554	}
555	if (gnet_stats_copy_queue(d, &qs) < 0)
556		return -1;
557	if (idx < q->flows_cnt)
558		return gnet_stats_copy_app(d, &xstats, sizeof(xstats));
559	return 0;
560}
561
562static void fq_codel_walk(struct Qdisc *sch, struct qdisc_walker *arg)
563{
564	struct fq_codel_sched_data *q = qdisc_priv(sch);
565	unsigned int i;
566
567	if (arg->stop)
568		return;
569
570	for (i = 0; i < q->flows_cnt; i++) {
571		if (list_empty(&q->flows[i].flowchain) ||
572		    arg->count < arg->skip) {
573			arg->count++;
574			continue;
575		}
576		if (arg->fn(sch, i + 1, arg) < 0) {
577			arg->stop = 1;
578			break;
579		}
580		arg->count++;
581	}
582}
583
584static const struct Qdisc_class_ops fq_codel_class_ops = {
585	.leaf		=	fq_codel_leaf,
586	.get		=	fq_codel_get,
587	.put		=	fq_codel_put,
588	.tcf_chain	=	fq_codel_find_tcf,
589	.bind_tcf	=	fq_codel_bind,
590	.unbind_tcf	=	fq_codel_put,
591	.dump		=	fq_codel_dump_class,
592	.dump_stats	=	fq_codel_dump_class_stats,
593	.walk		=	fq_codel_walk,
594};
595
596static struct Qdisc_ops fq_codel_qdisc_ops __read_mostly = {
597	.cl_ops		=	&fq_codel_class_ops,
598	.id		=	"fq_codel",
599	.priv_size	=	sizeof(struct fq_codel_sched_data),
600	.enqueue	=	fq_codel_enqueue,
601	.dequeue	=	fq_codel_dequeue,
602	.peek		=	qdisc_peek_dequeued,
603	.drop		=	fq_codel_drop,
604	.init		=	fq_codel_init,
605	.reset		=	fq_codel_reset,
606	.destroy	=	fq_codel_destroy,
607	.change		=	fq_codel_change,
608	.dump		=	fq_codel_dump,
609	.dump_stats =	fq_codel_dump_stats,
610	.owner		=	THIS_MODULE,
611};
612
613static int __init fq_codel_module_init(void)
614{
615	return register_qdisc(&fq_codel_qdisc_ops);
616}
617
618static void __exit fq_codel_module_exit(void)
619{
620	unregister_qdisc(&fq_codel_qdisc_ops);
621}
622
623module_init(fq_codel_module_init)
624module_exit(fq_codel_module_exit)
625MODULE_AUTHOR("Eric Dumazet");
626MODULE_LICENSE("GPL");
v5.14.15
  1// SPDX-License-Identifier: GPL-2.0-or-later
  2/*
  3 * Fair Queue CoDel discipline
  4 *
  5 *  Copyright (C) 2012,2015 Eric Dumazet <edumazet@google.com>
 
 
 
 
 
  6 */
  7
  8#include <linux/module.h>
  9#include <linux/types.h>
 10#include <linux/kernel.h>
 11#include <linux/jiffies.h>
 12#include <linux/string.h>
 13#include <linux/in.h>
 14#include <linux/errno.h>
 15#include <linux/init.h>
 16#include <linux/skbuff.h>
 
 17#include <linux/slab.h>
 18#include <linux/vmalloc.h>
 19#include <net/netlink.h>
 20#include <net/pkt_sched.h>
 21#include <net/pkt_cls.h>
 22#include <net/codel.h>
 23#include <net/codel_impl.h>
 24#include <net/codel_qdisc.h>
 25
 26/*	Fair Queue CoDel.
 27 *
 28 * Principles :
 29 * Packets are classified (internal classifier or external) on flows.
 30 * This is a Stochastic model (as we use a hash, several flows
 31 *			       might be hashed on same slot)
 32 * Each flow has a CoDel managed queue.
 33 * Flows are linked onto two (Round Robin) lists,
 34 * so that new flows have priority on old ones.
 35 *
 36 * For a given flow, packets are not reordered (CoDel uses a FIFO)
 37 * head drops only.
 38 * ECN capability is on by default.
 39 * Low memory footprint (64 bytes per flow)
 40 */
 41
 42struct fq_codel_flow {
 43	struct sk_buff	  *head;
 44	struct sk_buff	  *tail;
 45	struct list_head  flowchain;
 46	int		  deficit;
 
 47	struct codel_vars cvars;
 48}; /* please try to keep this structure <= 64 bytes */
 49
 50struct fq_codel_sched_data {
 51	struct tcf_proto __rcu *filter_list; /* optional external classifier */
 52	struct tcf_block *block;
 53	struct fq_codel_flow *flows;	/* Flows table [flows_cnt] */
 54	u32		*backlogs;	/* backlog table [flows_cnt] */
 55	u32		flows_cnt;	/* number of flows */
 
 56	u32		quantum;	/* psched_mtu(qdisc_dev(sch)); */
 57	u32		drop_batch_size;
 58	u32		memory_limit;
 59	struct codel_params cparams;
 60	struct codel_stats cstats;
 61	u32		memory_usage;
 62	u32		drop_overmemory;
 63	u32		drop_overlimit;
 64	u32		new_flow_count;
 65
 66	struct list_head new_flows;	/* list of new flows */
 67	struct list_head old_flows;	/* list of old flows */
 68};
 69
 70static unsigned int fq_codel_hash(const struct fq_codel_sched_data *q,
 71				  struct sk_buff *skb)
 72{
 73	return reciprocal_scale(skb_get_hash(skb), q->flows_cnt);
 
 
 
 
 
 
 
 74}
 75
 76static unsigned int fq_codel_classify(struct sk_buff *skb, struct Qdisc *sch,
 77				      int *qerr)
 78{
 79	struct fq_codel_sched_data *q = qdisc_priv(sch);
 80	struct tcf_proto *filter;
 81	struct tcf_result res;
 82	int result;
 83
 84	if (TC_H_MAJ(skb->priority) == sch->handle &&
 85	    TC_H_MIN(skb->priority) > 0 &&
 86	    TC_H_MIN(skb->priority) <= q->flows_cnt)
 87		return TC_H_MIN(skb->priority);
 88
 89	filter = rcu_dereference_bh(q->filter_list);
 90	if (!filter)
 91		return fq_codel_hash(q, skb) + 1;
 92
 93	*qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
 94	result = tcf_classify(skb, filter, &res, false);
 95	if (result >= 0) {
 96#ifdef CONFIG_NET_CLS_ACT
 97		switch (result) {
 98		case TC_ACT_STOLEN:
 99		case TC_ACT_QUEUED:
100		case TC_ACT_TRAP:
101			*qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
102			fallthrough;
103		case TC_ACT_SHOT:
104			return 0;
105		}
106#endif
107		if (TC_H_MIN(res.classid) <= q->flows_cnt)
108			return TC_H_MIN(res.classid);
109	}
110	return 0;
111}
112
113/* helper functions : might be changed when/if skb use a standard list_head */
114
115/* remove one skb from head of slot queue */
116static inline struct sk_buff *dequeue_head(struct fq_codel_flow *flow)
117{
118	struct sk_buff *skb = flow->head;
119
120	flow->head = skb->next;
121	skb_mark_not_on_list(skb);
122	return skb;
123}
124
125/* add skb to flow queue (tail add) */
126static inline void flow_queue_add(struct fq_codel_flow *flow,
127				  struct sk_buff *skb)
128{
129	if (flow->head == NULL)
130		flow->head = skb;
131	else
132		flow->tail->next = skb;
133	flow->tail = skb;
134	skb->next = NULL;
135}
136
137static unsigned int fq_codel_drop(struct Qdisc *sch, unsigned int max_packets,
138				  struct sk_buff **to_free)
139{
140	struct fq_codel_sched_data *q = qdisc_priv(sch);
141	struct sk_buff *skb;
142	unsigned int maxbacklog = 0, idx = 0, i, len;
143	struct fq_codel_flow *flow;
144	unsigned int threshold;
145	unsigned int mem = 0;
146
147	/* Queue is full! Find the fat flow and drop packet(s) from it.
148	 * This might sound expensive, but with 1024 flows, we scan
149	 * 4KB of memory, and we dont need to handle a complex tree
150	 * in fast path (packet queue/enqueue) with many cache misses.
151	 * In stress mode, we'll try to drop 64 packets from the flow,
152	 * amortizing this linear lookup to one cache line per drop.
153	 */
154	for (i = 0; i < q->flows_cnt; i++) {
155		if (q->backlogs[i] > maxbacklog) {
156			maxbacklog = q->backlogs[i];
157			idx = i;
158		}
159	}
160
161	/* Our goal is to drop half of this fat flow backlog */
162	threshold = maxbacklog >> 1;
163
164	flow = &q->flows[idx];
165	len = 0;
166	i = 0;
167	do {
168		skb = dequeue_head(flow);
169		len += qdisc_pkt_len(skb);
170		mem += get_codel_cb(skb)->mem_usage;
171		__qdisc_drop(skb, to_free);
172	} while (++i < max_packets && len < threshold);
173
174	/* Tell codel to increase its signal strength also */
175	flow->cvars.count += i;
176	q->backlogs[idx] -= len;
177	q->memory_usage -= mem;
178	sch->qstats.drops += i;
 
179	sch->qstats.backlog -= len;
180	sch->q.qlen -= i;
181	return idx;
182}
183
184static int fq_codel_enqueue(struct sk_buff *skb, struct Qdisc *sch,
185			    struct sk_buff **to_free)
186{
187	struct fq_codel_sched_data *q = qdisc_priv(sch);
188	unsigned int idx, prev_backlog, prev_qlen;
189	struct fq_codel_flow *flow;
190	int ret;
191	unsigned int pkt_len;
192	bool memory_limited;
193
194	idx = fq_codel_classify(skb, sch, &ret);
195	if (idx == 0) {
196		if (ret & __NET_XMIT_BYPASS)
197			qdisc_qstats_drop(sch);
198		__qdisc_drop(skb, to_free);
199		return ret;
200	}
201	idx--;
202
203	codel_set_enqueue_time(skb);
204	flow = &q->flows[idx];
205	flow_queue_add(flow, skb);
206	q->backlogs[idx] += qdisc_pkt_len(skb);
207	qdisc_qstats_backlog_inc(sch, skb);
208
209	if (list_empty(&flow->flowchain)) {
210		list_add_tail(&flow->flowchain, &q->new_flows);
 
211		q->new_flow_count++;
212		flow->deficit = q->quantum;
 
213	}
214	get_codel_cb(skb)->mem_usage = skb->truesize;
215	q->memory_usage += get_codel_cb(skb)->mem_usage;
216	memory_limited = q->memory_usage > q->memory_limit;
217	if (++sch->q.qlen <= sch->limit && !memory_limited)
218		return NET_XMIT_SUCCESS;
219
220	prev_backlog = sch->qstats.backlog;
221	prev_qlen = sch->q.qlen;
222
223	/* save this packet length as it might be dropped by fq_codel_drop() */
224	pkt_len = qdisc_pkt_len(skb);
225	/* fq_codel_drop() is quite expensive, as it performs a linear search
226	 * in q->backlogs[] to find a fat flow.
227	 * So instead of dropping a single packet, drop half of its backlog
228	 * with a 64 packets limit to not add a too big cpu spike here.
229	 */
230	ret = fq_codel_drop(sch, q->drop_batch_size, to_free);
 
231
232	prev_qlen -= sch->q.qlen;
233	prev_backlog -= sch->qstats.backlog;
234	q->drop_overlimit += prev_qlen;
235	if (memory_limited)
236		q->drop_overmemory += prev_qlen;
237
238	/* As we dropped packet(s), better let upper stack know this.
239	 * If we dropped a packet for this flow, return NET_XMIT_CN,
240	 * but in this case, our parents wont increase their backlogs.
241	 */
242	if (ret == idx) {
243		qdisc_tree_reduce_backlog(sch, prev_qlen - 1,
244					  prev_backlog - pkt_len);
245		return NET_XMIT_CN;
246	}
247	qdisc_tree_reduce_backlog(sch, prev_qlen, prev_backlog);
248	return NET_XMIT_SUCCESS;
249}
250
251/* This is the specific function called from codel_dequeue()
252 * to dequeue a packet from queue. Note: backlog is handled in
253 * codel, we dont need to reduce it here.
254 */
255static struct sk_buff *dequeue_func(struct codel_vars *vars, void *ctx)
256{
257	struct Qdisc *sch = ctx;
258	struct fq_codel_sched_data *q = qdisc_priv(sch);
259	struct fq_codel_flow *flow;
260	struct sk_buff *skb = NULL;
261
262	flow = container_of(vars, struct fq_codel_flow, cvars);
263	if (flow->head) {
264		skb = dequeue_head(flow);
265		q->backlogs[flow - q->flows] -= qdisc_pkt_len(skb);
266		q->memory_usage -= get_codel_cb(skb)->mem_usage;
267		sch->q.qlen--;
268		sch->qstats.backlog -= qdisc_pkt_len(skb);
269	}
270	return skb;
271}
272
273static void drop_func(struct sk_buff *skb, void *ctx)
274{
275	struct Qdisc *sch = ctx;
276
277	kfree_skb(skb);
278	qdisc_qstats_drop(sch);
279}
280
281static struct sk_buff *fq_codel_dequeue(struct Qdisc *sch)
282{
283	struct fq_codel_sched_data *q = qdisc_priv(sch);
284	struct sk_buff *skb;
285	struct fq_codel_flow *flow;
286	struct list_head *head;
 
287
288begin:
289	head = &q->new_flows;
290	if (list_empty(head)) {
291		head = &q->old_flows;
292		if (list_empty(head))
293			return NULL;
294	}
295	flow = list_first_entry(head, struct fq_codel_flow, flowchain);
296
297	if (flow->deficit <= 0) {
298		flow->deficit += q->quantum;
299		list_move_tail(&flow->flowchain, &q->old_flows);
300		goto begin;
301	}
302
303	skb = codel_dequeue(sch, &sch->qstats.backlog, &q->cparams,
304			    &flow->cvars, &q->cstats, qdisc_pkt_len,
305			    codel_get_enqueue_time, drop_func, dequeue_func);
 
 
 
 
 
306
307	if (!skb) {
308		/* force a pass through old_flows to prevent starvation */
309		if ((head == &q->new_flows) && !list_empty(&q->old_flows))
310			list_move_tail(&flow->flowchain, &q->old_flows);
311		else
312			list_del_init(&flow->flowchain);
313		goto begin;
314	}
315	qdisc_bstats_update(sch, skb);
316	flow->deficit -= qdisc_pkt_len(skb);
317	/* We cant call qdisc_tree_reduce_backlog() if our qlen is 0,
318	 * or HTB crashes. Defer it for next round.
319	 */
320	if (q->cstats.drop_count && sch->q.qlen) {
321		qdisc_tree_reduce_backlog(sch, q->cstats.drop_count,
322					  q->cstats.drop_len);
323		q->cstats.drop_count = 0;
324		q->cstats.drop_len = 0;
325	}
326	return skb;
327}
328
329static void fq_codel_flow_purge(struct fq_codel_flow *flow)
330{
331	rtnl_kfree_skbs(flow->head, flow->tail);
332	flow->head = NULL;
333}
334
335static void fq_codel_reset(struct Qdisc *sch)
336{
337	struct fq_codel_sched_data *q = qdisc_priv(sch);
338	int i;
339
340	INIT_LIST_HEAD(&q->new_flows);
341	INIT_LIST_HEAD(&q->old_flows);
342	for (i = 0; i < q->flows_cnt; i++) {
343		struct fq_codel_flow *flow = q->flows + i;
344
345		fq_codel_flow_purge(flow);
346		INIT_LIST_HEAD(&flow->flowchain);
347		codel_vars_init(&flow->cvars);
348	}
349	memset(q->backlogs, 0, q->flows_cnt * sizeof(u32));
350	sch->q.qlen = 0;
351	sch->qstats.backlog = 0;
352	q->memory_usage = 0;
353}
354
355static const struct nla_policy fq_codel_policy[TCA_FQ_CODEL_MAX + 1] = {
356	[TCA_FQ_CODEL_TARGET]	= { .type = NLA_U32 },
357	[TCA_FQ_CODEL_LIMIT]	= { .type = NLA_U32 },
358	[TCA_FQ_CODEL_INTERVAL]	= { .type = NLA_U32 },
359	[TCA_FQ_CODEL_ECN]	= { .type = NLA_U32 },
360	[TCA_FQ_CODEL_FLOWS]	= { .type = NLA_U32 },
361	[TCA_FQ_CODEL_QUANTUM]	= { .type = NLA_U32 },
362	[TCA_FQ_CODEL_CE_THRESHOLD] = { .type = NLA_U32 },
363	[TCA_FQ_CODEL_DROP_BATCH_SIZE] = { .type = NLA_U32 },
364	[TCA_FQ_CODEL_MEMORY_LIMIT] = { .type = NLA_U32 },
365};
366
367static int fq_codel_change(struct Qdisc *sch, struct nlattr *opt,
368			   struct netlink_ext_ack *extack)
369{
370	struct fq_codel_sched_data *q = qdisc_priv(sch);
371	struct nlattr *tb[TCA_FQ_CODEL_MAX + 1];
372	u32 quantum = 0;
373	int err;
374
375	if (!opt)
376		return -EINVAL;
377
378	err = nla_parse_nested_deprecated(tb, TCA_FQ_CODEL_MAX, opt,
379					  fq_codel_policy, NULL);
380	if (err < 0)
381		return err;
382	if (tb[TCA_FQ_CODEL_FLOWS]) {
383		if (q->flows)
384			return -EINVAL;
385		q->flows_cnt = nla_get_u32(tb[TCA_FQ_CODEL_FLOWS]);
386		if (!q->flows_cnt ||
387		    q->flows_cnt > 65536)
388			return -EINVAL;
389	}
390	if (tb[TCA_FQ_CODEL_QUANTUM]) {
391		quantum = max(256U, nla_get_u32(tb[TCA_FQ_CODEL_QUANTUM]));
392		if (quantum > FQ_CODEL_QUANTUM_MAX) {
393			NL_SET_ERR_MSG(extack, "Invalid quantum");
394			return -EINVAL;
395		}
396	}
397	sch_tree_lock(sch);
398
399	if (tb[TCA_FQ_CODEL_TARGET]) {
400		u64 target = nla_get_u32(tb[TCA_FQ_CODEL_TARGET]);
401
402		q->cparams.target = (target * NSEC_PER_USEC) >> CODEL_SHIFT;
403	}
404
405	if (tb[TCA_FQ_CODEL_CE_THRESHOLD]) {
406		u64 val = nla_get_u32(tb[TCA_FQ_CODEL_CE_THRESHOLD]);
407
408		q->cparams.ce_threshold = (val * NSEC_PER_USEC) >> CODEL_SHIFT;
409	}
410
411	if (tb[TCA_FQ_CODEL_INTERVAL]) {
412		u64 interval = nla_get_u32(tb[TCA_FQ_CODEL_INTERVAL]);
413
414		q->cparams.interval = (interval * NSEC_PER_USEC) >> CODEL_SHIFT;
415	}
416
417	if (tb[TCA_FQ_CODEL_LIMIT])
418		sch->limit = nla_get_u32(tb[TCA_FQ_CODEL_LIMIT]);
419
420	if (tb[TCA_FQ_CODEL_ECN])
421		q->cparams.ecn = !!nla_get_u32(tb[TCA_FQ_CODEL_ECN]);
422
423	if (quantum)
424		q->quantum = quantum;
425
426	if (tb[TCA_FQ_CODEL_DROP_BATCH_SIZE])
427		q->drop_batch_size = max(1U, nla_get_u32(tb[TCA_FQ_CODEL_DROP_BATCH_SIZE]));
428
429	if (tb[TCA_FQ_CODEL_MEMORY_LIMIT])
430		q->memory_limit = min(1U << 31, nla_get_u32(tb[TCA_FQ_CODEL_MEMORY_LIMIT]));
431
432	while (sch->q.qlen > sch->limit ||
433	       q->memory_usage > q->memory_limit) {
434		struct sk_buff *skb = fq_codel_dequeue(sch);
435
436		q->cstats.drop_len += qdisc_pkt_len(skb);
437		rtnl_kfree_skbs(skb, skb);
438		q->cstats.drop_count++;
439	}
440	qdisc_tree_reduce_backlog(sch, q->cstats.drop_count, q->cstats.drop_len);
441	q->cstats.drop_count = 0;
442	q->cstats.drop_len = 0;
443
444	sch_tree_unlock(sch);
445	return 0;
446}
447
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
448static void fq_codel_destroy(struct Qdisc *sch)
449{
450	struct fq_codel_sched_data *q = qdisc_priv(sch);
451
452	tcf_block_put(q->block);
453	kvfree(q->backlogs);
454	kvfree(q->flows);
455}
456
457static int fq_codel_init(struct Qdisc *sch, struct nlattr *opt,
458			 struct netlink_ext_ack *extack)
459{
460	struct fq_codel_sched_data *q = qdisc_priv(sch);
461	int i;
462	int err;
463
464	sch->limit = 10*1024;
465	q->flows_cnt = 1024;
466	q->memory_limit = 32 << 20; /* 32 MBytes */
467	q->drop_batch_size = 64;
468	q->quantum = psched_mtu(qdisc_dev(sch));
 
469	INIT_LIST_HEAD(&q->new_flows);
470	INIT_LIST_HEAD(&q->old_flows);
471	codel_params_init(&q->cparams);
472	codel_stats_init(&q->cstats);
473	q->cparams.ecn = true;
474	q->cparams.mtu = psched_mtu(qdisc_dev(sch));
475
476	if (opt) {
477		err = fq_codel_change(sch, opt, extack);
478		if (err)
479			goto init_failure;
480	}
481
482	err = tcf_block_get(&q->block, &q->filter_list, sch, extack);
483	if (err)
484		goto init_failure;
485
486	if (!q->flows) {
487		q->flows = kvcalloc(q->flows_cnt,
488				    sizeof(struct fq_codel_flow),
489				    GFP_KERNEL);
490		if (!q->flows) {
491			err = -ENOMEM;
492			goto init_failure;
493		}
494		q->backlogs = kvcalloc(q->flows_cnt, sizeof(u32), GFP_KERNEL);
495		if (!q->backlogs) {
496			err = -ENOMEM;
497			goto alloc_failure;
498		}
499		for (i = 0; i < q->flows_cnt; i++) {
500			struct fq_codel_flow *flow = q->flows + i;
501
502			INIT_LIST_HEAD(&flow->flowchain);
503			codel_vars_init(&flow->cvars);
504		}
505	}
506	if (sch->limit >= 1)
507		sch->flags |= TCQ_F_CAN_BYPASS;
508	else
509		sch->flags &= ~TCQ_F_CAN_BYPASS;
510	return 0;
511
512alloc_failure:
513	kvfree(q->flows);
514	q->flows = NULL;
515init_failure:
516	q->flows_cnt = 0;
517	return err;
518}
519
520static int fq_codel_dump(struct Qdisc *sch, struct sk_buff *skb)
521{
522	struct fq_codel_sched_data *q = qdisc_priv(sch);
523	struct nlattr *opts;
524
525	opts = nla_nest_start_noflag(skb, TCA_OPTIONS);
526	if (opts == NULL)
527		goto nla_put_failure;
528
529	if (nla_put_u32(skb, TCA_FQ_CODEL_TARGET,
530			codel_time_to_us(q->cparams.target)) ||
531	    nla_put_u32(skb, TCA_FQ_CODEL_LIMIT,
532			sch->limit) ||
533	    nla_put_u32(skb, TCA_FQ_CODEL_INTERVAL,
534			codel_time_to_us(q->cparams.interval)) ||
535	    nla_put_u32(skb, TCA_FQ_CODEL_ECN,
536			q->cparams.ecn) ||
537	    nla_put_u32(skb, TCA_FQ_CODEL_QUANTUM,
538			q->quantum) ||
539	    nla_put_u32(skb, TCA_FQ_CODEL_DROP_BATCH_SIZE,
540			q->drop_batch_size) ||
541	    nla_put_u32(skb, TCA_FQ_CODEL_MEMORY_LIMIT,
542			q->memory_limit) ||
543	    nla_put_u32(skb, TCA_FQ_CODEL_FLOWS,
544			q->flows_cnt))
545		goto nla_put_failure;
546
547	if (q->cparams.ce_threshold != CODEL_DISABLED_THRESHOLD &&
548	    nla_put_u32(skb, TCA_FQ_CODEL_CE_THRESHOLD,
549			codel_time_to_us(q->cparams.ce_threshold)))
550		goto nla_put_failure;
551
552	return nla_nest_end(skb, opts);
553
554nla_put_failure:
555	return -1;
556}
557
558static int fq_codel_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
559{
560	struct fq_codel_sched_data *q = qdisc_priv(sch);
561	struct tc_fq_codel_xstats st = {
562		.type				= TCA_FQ_CODEL_XSTATS_QDISC,
563	};
564	struct list_head *pos;
565
566	st.qdisc_stats.maxpacket = q->cstats.maxpacket;
567	st.qdisc_stats.drop_overlimit = q->drop_overlimit;
568	st.qdisc_stats.ecn_mark = q->cstats.ecn_mark;
569	st.qdisc_stats.new_flow_count = q->new_flow_count;
570	st.qdisc_stats.ce_mark = q->cstats.ce_mark;
571	st.qdisc_stats.memory_usage  = q->memory_usage;
572	st.qdisc_stats.drop_overmemory = q->drop_overmemory;
573
574	sch_tree_lock(sch);
575	list_for_each(pos, &q->new_flows)
576		st.qdisc_stats.new_flows_len++;
577
578	list_for_each(pos, &q->old_flows)
579		st.qdisc_stats.old_flows_len++;
580	sch_tree_unlock(sch);
581
582	return gnet_stats_copy_app(d, &st, sizeof(st));
583}
584
585static struct Qdisc *fq_codel_leaf(struct Qdisc *sch, unsigned long arg)
586{
587	return NULL;
588}
589
590static unsigned long fq_codel_find(struct Qdisc *sch, u32 classid)
591{
592	return 0;
593}
594
595static unsigned long fq_codel_bind(struct Qdisc *sch, unsigned long parent,
596			      u32 classid)
597{
 
 
598	return 0;
599}
600
601static void fq_codel_unbind(struct Qdisc *q, unsigned long cl)
602{
603}
604
605static struct tcf_block *fq_codel_tcf_block(struct Qdisc *sch, unsigned long cl,
606					    struct netlink_ext_ack *extack)
607{
608	struct fq_codel_sched_data *q = qdisc_priv(sch);
609
610	if (cl)
611		return NULL;
612	return q->block;
613}
614
615static int fq_codel_dump_class(struct Qdisc *sch, unsigned long cl,
616			  struct sk_buff *skb, struct tcmsg *tcm)
617{
618	tcm->tcm_handle |= TC_H_MIN(cl);
619	return 0;
620}
621
622static int fq_codel_dump_class_stats(struct Qdisc *sch, unsigned long cl,
623				     struct gnet_dump *d)
624{
625	struct fq_codel_sched_data *q = qdisc_priv(sch);
626	u32 idx = cl - 1;
627	struct gnet_stats_queue qs = { 0 };
628	struct tc_fq_codel_xstats xstats;
629
630	if (idx < q->flows_cnt) {
631		const struct fq_codel_flow *flow = &q->flows[idx];
632		const struct sk_buff *skb;
633
634		memset(&xstats, 0, sizeof(xstats));
635		xstats.type = TCA_FQ_CODEL_XSTATS_CLASS;
636		xstats.class_stats.deficit = flow->deficit;
637		xstats.class_stats.ldelay =
638			codel_time_to_us(flow->cvars.ldelay);
639		xstats.class_stats.count = flow->cvars.count;
640		xstats.class_stats.lastcount = flow->cvars.lastcount;
641		xstats.class_stats.dropping = flow->cvars.dropping;
642		if (flow->cvars.dropping) {
643			codel_tdiff_t delta = flow->cvars.drop_next -
644					      codel_get_time();
645
646			xstats.class_stats.drop_next = (delta >= 0) ?
647				codel_time_to_us(delta) :
648				-codel_time_to_us(-delta);
649		}
650		if (flow->head) {
651			sch_tree_lock(sch);
652			skb = flow->head;
653			while (skb) {
654				qs.qlen++;
655				skb = skb->next;
656			}
657			sch_tree_unlock(sch);
658		}
659		qs.backlog = q->backlogs[idx];
660		qs.drops = 0;
661	}
662	if (gnet_stats_copy_queue(d, NULL, &qs, qs.qlen) < 0)
663		return -1;
664	if (idx < q->flows_cnt)
665		return gnet_stats_copy_app(d, &xstats, sizeof(xstats));
666	return 0;
667}
668
669static void fq_codel_walk(struct Qdisc *sch, struct qdisc_walker *arg)
670{
671	struct fq_codel_sched_data *q = qdisc_priv(sch);
672	unsigned int i;
673
674	if (arg->stop)
675		return;
676
677	for (i = 0; i < q->flows_cnt; i++) {
678		if (list_empty(&q->flows[i].flowchain) ||
679		    arg->count < arg->skip) {
680			arg->count++;
681			continue;
682		}
683		if (arg->fn(sch, i + 1, arg) < 0) {
684			arg->stop = 1;
685			break;
686		}
687		arg->count++;
688	}
689}
690
691static const struct Qdisc_class_ops fq_codel_class_ops = {
692	.leaf		=	fq_codel_leaf,
693	.find		=	fq_codel_find,
694	.tcf_block	=	fq_codel_tcf_block,
 
695	.bind_tcf	=	fq_codel_bind,
696	.unbind_tcf	=	fq_codel_unbind,
697	.dump		=	fq_codel_dump_class,
698	.dump_stats	=	fq_codel_dump_class_stats,
699	.walk		=	fq_codel_walk,
700};
701
702static struct Qdisc_ops fq_codel_qdisc_ops __read_mostly = {
703	.cl_ops		=	&fq_codel_class_ops,
704	.id		=	"fq_codel",
705	.priv_size	=	sizeof(struct fq_codel_sched_data),
706	.enqueue	=	fq_codel_enqueue,
707	.dequeue	=	fq_codel_dequeue,
708	.peek		=	qdisc_peek_dequeued,
 
709	.init		=	fq_codel_init,
710	.reset		=	fq_codel_reset,
711	.destroy	=	fq_codel_destroy,
712	.change		=	fq_codel_change,
713	.dump		=	fq_codel_dump,
714	.dump_stats =	fq_codel_dump_stats,
715	.owner		=	THIS_MODULE,
716};
717
718static int __init fq_codel_module_init(void)
719{
720	return register_qdisc(&fq_codel_qdisc_ops);
721}
722
723static void __exit fq_codel_module_exit(void)
724{
725	unregister_qdisc(&fq_codel_qdisc_ops);
726}
727
728module_init(fq_codel_module_init)
729module_exit(fq_codel_module_exit)
730MODULE_AUTHOR("Eric Dumazet");
731MODULE_LICENSE("GPL");
732MODULE_DESCRIPTION("Fair Queue CoDel discipline");