xref: /linux/net/sched/sch_fq_codel.c (revision 0d240e7811c4ec1965760ee4643b5bbc9cfacbb3)
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) 2012,2015 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/codel.h>
27 #include <net/codel_impl.h>
28 #include <net/codel_qdisc.h>
29 
30 /*	Fair Queue CoDel.
31  *
32  * Principles :
33  * Packets are classified (internal classifier or external) on flows.
34  * This is a Stochastic model (as we use a hash, several flows
35  *			       might be hashed on same slot)
36  * Each flow has a CoDel managed queue.
37  * Flows are linked onto two (Round Robin) lists,
38  * so that new flows have priority on old ones.
39  *
40  * For a given flow, packets are not reordered (CoDel uses a FIFO)
41  * head drops only.
42  * ECN capability is on by default.
43  * Low memory footprint (64 bytes per flow)
44  */
45 
46 struct fq_codel_flow {
47 	struct sk_buff	  *head;
48 	struct sk_buff	  *tail;
49 	struct list_head  flowchain;
50 	int		  deficit;
51 	u32		  dropped; /* number of drops (or ECN marks) on this flow */
52 	struct codel_vars cvars;
53 }; /* please try to keep this structure <= 64 bytes */
54 
55 struct fq_codel_sched_data {
56 	struct tcf_proto __rcu *filter_list; /* optional external classifier */
57 	struct fq_codel_flow *flows;	/* Flows table [flows_cnt] */
58 	u32		*backlogs;	/* backlog table [flows_cnt] */
59 	u32		flows_cnt;	/* number of flows */
60 	u32		perturbation;	/* hash perturbation */
61 	u32		quantum;	/* psched_mtu(qdisc_dev(sch)); */
62 	u32		drop_batch_size;
63 	u32		memory_limit;
64 	struct codel_params cparams;
65 	struct codel_stats cstats;
66 	u32		memory_usage;
67 	u32		drop_overmemory;
68 	u32		drop_overlimit;
69 	u32		new_flow_count;
70 
71 	struct list_head new_flows;	/* list of new flows */
72 	struct list_head old_flows;	/* list of old flows */
73 };
74 
75 static unsigned int fq_codel_hash(const struct fq_codel_sched_data *q,
76 				  struct sk_buff *skb)
77 {
78 	u32 hash = skb_get_hash_perturb(skb, q->perturbation);
79 
80 	return reciprocal_scale(hash, q->flows_cnt);
81 }
82 
83 static 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_proto *filter;
88 	struct tcf_result res;
89 	int result;
90 
91 	if (TC_H_MAJ(skb->priority) == sch->handle &&
92 	    TC_H_MIN(skb->priority) > 0 &&
93 	    TC_H_MIN(skb->priority) <= q->flows_cnt)
94 		return TC_H_MIN(skb->priority);
95 
96 	filter = rcu_dereference_bh(q->filter_list);
97 	if (!filter)
98 		return fq_codel_hash(q, skb) + 1;
99 
100 	*qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
101 	result = tc_classify(skb, filter, &res, false);
102 	if (result >= 0) {
103 #ifdef CONFIG_NET_CLS_ACT
104 		switch (result) {
105 		case TC_ACT_STOLEN:
106 		case TC_ACT_QUEUED:
107 			*qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
108 		case TC_ACT_SHOT:
109 			return 0;
110 		}
111 #endif
112 		if (TC_H_MIN(res.classid) <= q->flows_cnt)
113 			return TC_H_MIN(res.classid);
114 	}
115 	return 0;
116 }
117 
118 /* helper functions : might be changed when/if skb use a standard list_head */
119 
120 /* remove one skb from head of slot queue */
121 static inline struct sk_buff *dequeue_head(struct fq_codel_flow *flow)
122 {
123 	struct sk_buff *skb = flow->head;
124 
125 	flow->head = skb->next;
126 	skb->next = NULL;
127 	return skb;
128 }
129 
130 /* add skb to flow queue (tail add) */
131 static inline void flow_queue_add(struct fq_codel_flow *flow,
132 				  struct sk_buff *skb)
133 {
134 	if (flow->head == NULL)
135 		flow->head = skb;
136 	else
137 		flow->tail->next = skb;
138 	flow->tail = skb;
139 	skb->next = NULL;
140 }
141 
142 static unsigned int fq_codel_drop(struct Qdisc *sch, unsigned int max_packets)
143 {
144 	struct fq_codel_sched_data *q = qdisc_priv(sch);
145 	struct sk_buff *skb;
146 	unsigned int maxbacklog = 0, idx = 0, i, len;
147 	struct fq_codel_flow *flow;
148 	unsigned int threshold;
149 	unsigned int mem = 0;
150 
151 	/* Queue is full! Find the fat flow and drop packet(s) from it.
152 	 * This might sound expensive, but with 1024 flows, we scan
153 	 * 4KB of memory, and we dont need to handle a complex tree
154 	 * in fast path (packet queue/enqueue) with many cache misses.
155 	 * In stress mode, we'll try to drop 64 packets from the flow,
156 	 * amortizing this linear lookup to one cache line per drop.
157 	 */
158 	for (i = 0; i < q->flows_cnt; i++) {
159 		if (q->backlogs[i] > maxbacklog) {
160 			maxbacklog = q->backlogs[i];
161 			idx = i;
162 		}
163 	}
164 
165 	/* Our goal is to drop half of this fat flow backlog */
166 	threshold = maxbacklog >> 1;
167 
168 	flow = &q->flows[idx];
169 	len = 0;
170 	i = 0;
171 	do {
172 		skb = dequeue_head(flow);
173 		len += qdisc_pkt_len(skb);
174 		mem += skb->truesize;
175 		kfree_skb(skb);
176 	} while (++i < max_packets && len < threshold);
177 
178 	flow->dropped += i;
179 	q->backlogs[idx] -= len;
180 	q->memory_usage -= mem;
181 	sch->qstats.drops += i;
182 	sch->qstats.backlog -= len;
183 	sch->q.qlen -= i;
184 	return idx;
185 }
186 
187 static int fq_codel_enqueue(struct sk_buff *skb, struct Qdisc *sch)
188 {
189 	struct fq_codel_sched_data *q = qdisc_priv(sch);
190 	unsigned int idx, prev_backlog, prev_qlen;
191 	struct fq_codel_flow *flow;
192 	int uninitialized_var(ret);
193 	unsigned int pkt_len;
194 	bool memory_limited;
195 
196 	idx = fq_codel_classify(skb, sch, &ret);
197 	if (idx == 0) {
198 		if (ret & __NET_XMIT_BYPASS)
199 			qdisc_qstats_drop(sch);
200 		kfree_skb(skb);
201 		return ret;
202 	}
203 	idx--;
204 
205 	codel_set_enqueue_time(skb);
206 	flow = &q->flows[idx];
207 	flow_queue_add(flow, skb);
208 	q->backlogs[idx] += qdisc_pkt_len(skb);
209 	qdisc_qstats_backlog_inc(sch, skb);
210 
211 	if (list_empty(&flow->flowchain)) {
212 		list_add_tail(&flow->flowchain, &q->new_flows);
213 		q->new_flow_count++;
214 		flow->deficit = q->quantum;
215 		flow->dropped = 0;
216 	}
217 	q->memory_usage += skb->truesize;
218 	memory_limited = q->memory_usage > q->memory_limit;
219 	if (++sch->q.qlen <= sch->limit && !memory_limited)
220 		return NET_XMIT_SUCCESS;
221 
222 	prev_backlog = sch->qstats.backlog;
223 	prev_qlen = sch->q.qlen;
224 
225 	/* save this packet length as it might be dropped by fq_codel_drop() */
226 	pkt_len = qdisc_pkt_len(skb);
227 	/* fq_codel_drop() is quite expensive, as it performs a linear search
228 	 * in q->backlogs[] to find a fat flow.
229 	 * So instead of dropping a single packet, drop half of its backlog
230 	 * with a 64 packets limit to not add a too big cpu spike here.
231 	 */
232 	ret = fq_codel_drop(sch, q->drop_batch_size);
233 
234 	prev_qlen -= sch->q.qlen;
235 	prev_backlog -= sch->qstats.backlog;
236 	q->drop_overlimit += prev_qlen;
237 	if (memory_limited)
238 		q->drop_overmemory += prev_qlen;
239 
240 	/* As we dropped packet(s), better let upper stack know this.
241 	 * If we dropped a packet for this flow, return NET_XMIT_CN,
242 	 * but in this case, our parents wont increase their backlogs.
243 	 */
244 	if (ret == idx) {
245 		qdisc_tree_reduce_backlog(sch, prev_qlen - 1,
246 					  prev_backlog - pkt_len);
247 		return NET_XMIT_CN;
248 	}
249 	qdisc_tree_reduce_backlog(sch, prev_qlen, prev_backlog);
250 	return NET_XMIT_SUCCESS;
251 }
252 
253 /* This is the specific function called from codel_dequeue()
254  * to dequeue a packet from queue. Note: backlog is handled in
255  * codel, we dont need to reduce it here.
256  */
257 static struct sk_buff *dequeue_func(struct codel_vars *vars, void *ctx)
258 {
259 	struct Qdisc *sch = ctx;
260 	struct fq_codel_sched_data *q = qdisc_priv(sch);
261 	struct fq_codel_flow *flow;
262 	struct sk_buff *skb = NULL;
263 
264 	flow = container_of(vars, struct fq_codel_flow, cvars);
265 	if (flow->head) {
266 		skb = dequeue_head(flow);
267 		q->backlogs[flow - q->flows] -= qdisc_pkt_len(skb);
268 		q->memory_usage -= skb->truesize;
269 		sch->q.qlen--;
270 		sch->qstats.backlog -= qdisc_pkt_len(skb);
271 	}
272 	return skb;
273 }
274 
275 static void drop_func(struct sk_buff *skb, void *ctx)
276 {
277 	struct Qdisc *sch = ctx;
278 
279 	qdisc_drop(skb, sch);
280 }
281 
282 static struct sk_buff *fq_codel_dequeue(struct Qdisc *sch)
283 {
284 	struct fq_codel_sched_data *q = qdisc_priv(sch);
285 	struct sk_buff *skb;
286 	struct fq_codel_flow *flow;
287 	struct list_head *head;
288 	u32 prev_drop_count, prev_ecn_mark;
289 	unsigned int prev_backlog;
290 
291 begin:
292 	head = &q->new_flows;
293 	if (list_empty(head)) {
294 		head = &q->old_flows;
295 		if (list_empty(head))
296 			return NULL;
297 	}
298 	flow = list_first_entry(head, struct fq_codel_flow, flowchain);
299 
300 	if (flow->deficit <= 0) {
301 		flow->deficit += q->quantum;
302 		list_move_tail(&flow->flowchain, &q->old_flows);
303 		goto begin;
304 	}
305 
306 	prev_drop_count = q->cstats.drop_count;
307 	prev_ecn_mark = q->cstats.ecn_mark;
308 	prev_backlog = sch->qstats.backlog;
309 
310 	skb = codel_dequeue(sch, &sch->qstats.backlog, &q->cparams,
311 			    &flow->cvars, &q->cstats, qdisc_pkt_len,
312 			    codel_get_enqueue_time, drop_func, dequeue_func);
313 
314 	flow->dropped += q->cstats.drop_count - prev_drop_count;
315 	flow->dropped += q->cstats.ecn_mark - prev_ecn_mark;
316 
317 	if (!skb) {
318 		/* force a pass through old_flows to prevent starvation */
319 		if ((head == &q->new_flows) && !list_empty(&q->old_flows))
320 			list_move_tail(&flow->flowchain, &q->old_flows);
321 		else
322 			list_del_init(&flow->flowchain);
323 		goto begin;
324 	}
325 	qdisc_bstats_update(sch, skb);
326 	flow->deficit -= qdisc_pkt_len(skb);
327 	/* We cant call qdisc_tree_reduce_backlog() if our qlen is 0,
328 	 * or HTB crashes. Defer it for next round.
329 	 */
330 	if (q->cstats.drop_count && sch->q.qlen) {
331 		qdisc_tree_reduce_backlog(sch, q->cstats.drop_count,
332 					  q->cstats.drop_len);
333 		q->cstats.drop_count = 0;
334 		q->cstats.drop_len = 0;
335 	}
336 	return skb;
337 }
338 
339 static void fq_codel_flow_purge(struct fq_codel_flow *flow)
340 {
341 	rtnl_kfree_skbs(flow->head, flow->tail);
342 	flow->head = NULL;
343 }
344 
345 static void fq_codel_reset(struct Qdisc *sch)
346 {
347 	struct fq_codel_sched_data *q = qdisc_priv(sch);
348 	int i;
349 
350 	INIT_LIST_HEAD(&q->new_flows);
351 	INIT_LIST_HEAD(&q->old_flows);
352 	for (i = 0; i < q->flows_cnt; i++) {
353 		struct fq_codel_flow *flow = q->flows + i;
354 
355 		fq_codel_flow_purge(flow);
356 		INIT_LIST_HEAD(&flow->flowchain);
357 		codel_vars_init(&flow->cvars);
358 	}
359 	memset(q->backlogs, 0, q->flows_cnt * sizeof(u32));
360 	sch->q.qlen = 0;
361 	sch->qstats.backlog = 0;
362 	q->memory_usage = 0;
363 }
364 
365 static const struct nla_policy fq_codel_policy[TCA_FQ_CODEL_MAX + 1] = {
366 	[TCA_FQ_CODEL_TARGET]	= { .type = NLA_U32 },
367 	[TCA_FQ_CODEL_LIMIT]	= { .type = NLA_U32 },
368 	[TCA_FQ_CODEL_INTERVAL]	= { .type = NLA_U32 },
369 	[TCA_FQ_CODEL_ECN]	= { .type = NLA_U32 },
370 	[TCA_FQ_CODEL_FLOWS]	= { .type = NLA_U32 },
371 	[TCA_FQ_CODEL_QUANTUM]	= { .type = NLA_U32 },
372 	[TCA_FQ_CODEL_CE_THRESHOLD] = { .type = NLA_U32 },
373 	[TCA_FQ_CODEL_DROP_BATCH_SIZE] = { .type = NLA_U32 },
374 	[TCA_FQ_CODEL_MEMORY_LIMIT] = { .type = NLA_U32 },
375 };
376 
377 static int fq_codel_change(struct Qdisc *sch, struct nlattr *opt)
378 {
379 	struct fq_codel_sched_data *q = qdisc_priv(sch);
380 	struct nlattr *tb[TCA_FQ_CODEL_MAX + 1];
381 	int err;
382 
383 	if (!opt)
384 		return -EINVAL;
385 
386 	err = nla_parse_nested(tb, TCA_FQ_CODEL_MAX, opt, fq_codel_policy);
387 	if (err < 0)
388 		return err;
389 	if (tb[TCA_FQ_CODEL_FLOWS]) {
390 		if (q->flows)
391 			return -EINVAL;
392 		q->flows_cnt = nla_get_u32(tb[TCA_FQ_CODEL_FLOWS]);
393 		if (!q->flows_cnt ||
394 		    q->flows_cnt > 65536)
395 			return -EINVAL;
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 (tb[TCA_FQ_CODEL_QUANTUM])
424 		q->quantum = max(256U, nla_get_u32(tb[TCA_FQ_CODEL_QUANTUM]));
425 
426 	if (tb[TCA_FQ_CODEL_DROP_BATCH_SIZE])
427 		q->drop_batch_size = min(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 
448 static void *fq_codel_zalloc(size_t sz)
449 {
450 	void *ptr = kzalloc(sz, GFP_KERNEL | __GFP_NOWARN);
451 
452 	if (!ptr)
453 		ptr = vzalloc(sz);
454 	return ptr;
455 }
456 
457 static void fq_codel_free(void *addr)
458 {
459 	kvfree(addr);
460 }
461 
462 static void fq_codel_destroy(struct Qdisc *sch)
463 {
464 	struct fq_codel_sched_data *q = qdisc_priv(sch);
465 
466 	tcf_destroy_chain(&q->filter_list);
467 	fq_codel_free(q->backlogs);
468 	fq_codel_free(q->flows);
469 }
470 
471 static int fq_codel_init(struct Qdisc *sch, struct nlattr *opt)
472 {
473 	struct fq_codel_sched_data *q = qdisc_priv(sch);
474 	int i;
475 
476 	sch->limit = 10*1024;
477 	q->flows_cnt = 1024;
478 	q->memory_limit = 32 << 20; /* 32 MBytes */
479 	q->drop_batch_size = 64;
480 	q->quantum = psched_mtu(qdisc_dev(sch));
481 	q->perturbation = prandom_u32();
482 	INIT_LIST_HEAD(&q->new_flows);
483 	INIT_LIST_HEAD(&q->old_flows);
484 	codel_params_init(&q->cparams);
485 	codel_stats_init(&q->cstats);
486 	q->cparams.ecn = true;
487 	q->cparams.mtu = psched_mtu(qdisc_dev(sch));
488 
489 	if (opt) {
490 		int err = fq_codel_change(sch, opt);
491 		if (err)
492 			return err;
493 	}
494 
495 	if (!q->flows) {
496 		q->flows = fq_codel_zalloc(q->flows_cnt *
497 					   sizeof(struct fq_codel_flow));
498 		if (!q->flows)
499 			return -ENOMEM;
500 		q->backlogs = fq_codel_zalloc(q->flows_cnt * sizeof(u32));
501 		if (!q->backlogs) {
502 			fq_codel_free(q->flows);
503 			return -ENOMEM;
504 		}
505 		for (i = 0; i < q->flows_cnt; i++) {
506 			struct fq_codel_flow *flow = q->flows + i;
507 
508 			INIT_LIST_HEAD(&flow->flowchain);
509 			codel_vars_init(&flow->cvars);
510 		}
511 	}
512 	if (sch->limit >= 1)
513 		sch->flags |= TCQ_F_CAN_BYPASS;
514 	else
515 		sch->flags &= ~TCQ_F_CAN_BYPASS;
516 	return 0;
517 }
518 
519 static int fq_codel_dump(struct Qdisc *sch, struct sk_buff *skb)
520 {
521 	struct fq_codel_sched_data *q = qdisc_priv(sch);
522 	struct nlattr *opts;
523 
524 	opts = nla_nest_start(skb, TCA_OPTIONS);
525 	if (opts == NULL)
526 		goto nla_put_failure;
527 
528 	if (nla_put_u32(skb, TCA_FQ_CODEL_TARGET,
529 			codel_time_to_us(q->cparams.target)) ||
530 	    nla_put_u32(skb, TCA_FQ_CODEL_LIMIT,
531 			sch->limit) ||
532 	    nla_put_u32(skb, TCA_FQ_CODEL_INTERVAL,
533 			codel_time_to_us(q->cparams.interval)) ||
534 	    nla_put_u32(skb, TCA_FQ_CODEL_ECN,
535 			q->cparams.ecn) ||
536 	    nla_put_u32(skb, TCA_FQ_CODEL_QUANTUM,
537 			q->quantum) ||
538 	    nla_put_u32(skb, TCA_FQ_CODEL_DROP_BATCH_SIZE,
539 			q->drop_batch_size) ||
540 	    nla_put_u32(skb, TCA_FQ_CODEL_MEMORY_LIMIT,
541 			q->memory_limit) ||
542 	    nla_put_u32(skb, TCA_FQ_CODEL_FLOWS,
543 			q->flows_cnt))
544 		goto nla_put_failure;
545 
546 	if (q->cparams.ce_threshold != CODEL_DISABLED_THRESHOLD &&
547 	    nla_put_u32(skb, TCA_FQ_CODEL_CE_THRESHOLD,
548 			codel_time_to_us(q->cparams.ce_threshold)))
549 		goto nla_put_failure;
550 
551 	return nla_nest_end(skb, opts);
552 
553 nla_put_failure:
554 	return -1;
555 }
556 
557 static int fq_codel_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
558 {
559 	struct fq_codel_sched_data *q = qdisc_priv(sch);
560 	struct tc_fq_codel_xstats st = {
561 		.type				= TCA_FQ_CODEL_XSTATS_QDISC,
562 	};
563 	struct list_head *pos;
564 
565 	st.qdisc_stats.maxpacket = q->cstats.maxpacket;
566 	st.qdisc_stats.drop_overlimit = q->drop_overlimit;
567 	st.qdisc_stats.ecn_mark = q->cstats.ecn_mark;
568 	st.qdisc_stats.new_flow_count = q->new_flow_count;
569 	st.qdisc_stats.ce_mark = q->cstats.ce_mark;
570 	st.qdisc_stats.memory_usage  = q->memory_usage;
571 	st.qdisc_stats.drop_overmemory = q->drop_overmemory;
572 
573 	sch_tree_lock(sch);
574 	list_for_each(pos, &q->new_flows)
575 		st.qdisc_stats.new_flows_len++;
576 
577 	list_for_each(pos, &q->old_flows)
578 		st.qdisc_stats.old_flows_len++;
579 	sch_tree_unlock(sch);
580 
581 	return gnet_stats_copy_app(d, &st, sizeof(st));
582 }
583 
584 static struct Qdisc *fq_codel_leaf(struct Qdisc *sch, unsigned long arg)
585 {
586 	return NULL;
587 }
588 
589 static unsigned long fq_codel_get(struct Qdisc *sch, u32 classid)
590 {
591 	return 0;
592 }
593 
594 static unsigned long fq_codel_bind(struct Qdisc *sch, unsigned long parent,
595 			      u32 classid)
596 {
597 	/* we cannot bypass queue discipline anymore */
598 	sch->flags &= ~TCQ_F_CAN_BYPASS;
599 	return 0;
600 }
601 
602 static void fq_codel_put(struct Qdisc *q, unsigned long cl)
603 {
604 }
605 
606 static struct tcf_proto __rcu **fq_codel_find_tcf(struct Qdisc *sch,
607 						  unsigned long cl)
608 {
609 	struct fq_codel_sched_data *q = qdisc_priv(sch);
610 
611 	if (cl)
612 		return NULL;
613 	return &q->filter_list;
614 }
615 
616 static int fq_codel_dump_class(struct Qdisc *sch, unsigned long cl,
617 			  struct sk_buff *skb, struct tcmsg *tcm)
618 {
619 	tcm->tcm_handle |= TC_H_MIN(cl);
620 	return 0;
621 }
622 
623 static int fq_codel_dump_class_stats(struct Qdisc *sch, unsigned long cl,
624 				     struct gnet_dump *d)
625 {
626 	struct fq_codel_sched_data *q = qdisc_priv(sch);
627 	u32 idx = cl - 1;
628 	struct gnet_stats_queue qs = { 0 };
629 	struct tc_fq_codel_xstats xstats;
630 
631 	if (idx < q->flows_cnt) {
632 		const struct fq_codel_flow *flow = &q->flows[idx];
633 		const struct sk_buff *skb;
634 
635 		memset(&xstats, 0, sizeof(xstats));
636 		xstats.type = TCA_FQ_CODEL_XSTATS_CLASS;
637 		xstats.class_stats.deficit = flow->deficit;
638 		xstats.class_stats.ldelay =
639 			codel_time_to_us(flow->cvars.ldelay);
640 		xstats.class_stats.count = flow->cvars.count;
641 		xstats.class_stats.lastcount = flow->cvars.lastcount;
642 		xstats.class_stats.dropping = flow->cvars.dropping;
643 		if (flow->cvars.dropping) {
644 			codel_tdiff_t delta = flow->cvars.drop_next -
645 					      codel_get_time();
646 
647 			xstats.class_stats.drop_next = (delta >= 0) ?
648 				codel_time_to_us(delta) :
649 				-codel_time_to_us(-delta);
650 		}
651 		if (flow->head) {
652 			sch_tree_lock(sch);
653 			skb = flow->head;
654 			while (skb) {
655 				qs.qlen++;
656 				skb = skb->next;
657 			}
658 			sch_tree_unlock(sch);
659 		}
660 		qs.backlog = q->backlogs[idx];
661 		qs.drops = flow->dropped;
662 	}
663 	if (gnet_stats_copy_queue(d, NULL, &qs, qs.qlen) < 0)
664 		return -1;
665 	if (idx < q->flows_cnt)
666 		return gnet_stats_copy_app(d, &xstats, sizeof(xstats));
667 	return 0;
668 }
669 
670 static void fq_codel_walk(struct Qdisc *sch, struct qdisc_walker *arg)
671 {
672 	struct fq_codel_sched_data *q = qdisc_priv(sch);
673 	unsigned int i;
674 
675 	if (arg->stop)
676 		return;
677 
678 	for (i = 0; i < q->flows_cnt; i++) {
679 		if (list_empty(&q->flows[i].flowchain) ||
680 		    arg->count < arg->skip) {
681 			arg->count++;
682 			continue;
683 		}
684 		if (arg->fn(sch, i + 1, arg) < 0) {
685 			arg->stop = 1;
686 			break;
687 		}
688 		arg->count++;
689 	}
690 }
691 
692 static const struct Qdisc_class_ops fq_codel_class_ops = {
693 	.leaf		=	fq_codel_leaf,
694 	.get		=	fq_codel_get,
695 	.put		=	fq_codel_put,
696 	.tcf_chain	=	fq_codel_find_tcf,
697 	.bind_tcf	=	fq_codel_bind,
698 	.unbind_tcf	=	fq_codel_put,
699 	.dump		=	fq_codel_dump_class,
700 	.dump_stats	=	fq_codel_dump_class_stats,
701 	.walk		=	fq_codel_walk,
702 };
703 
704 static struct Qdisc_ops fq_codel_qdisc_ops __read_mostly = {
705 	.cl_ops		=	&fq_codel_class_ops,
706 	.id		=	"fq_codel",
707 	.priv_size	=	sizeof(struct fq_codel_sched_data),
708 	.enqueue	=	fq_codel_enqueue,
709 	.dequeue	=	fq_codel_dequeue,
710 	.peek		=	qdisc_peek_dequeued,
711 	.init		=	fq_codel_init,
712 	.reset		=	fq_codel_reset,
713 	.destroy	=	fq_codel_destroy,
714 	.change		=	fq_codel_change,
715 	.dump		=	fq_codel_dump,
716 	.dump_stats =	fq_codel_dump_stats,
717 	.owner		=	THIS_MODULE,
718 };
719 
720 static int __init fq_codel_module_init(void)
721 {
722 	return register_qdisc(&fq_codel_qdisc_ops);
723 }
724 
725 static void __exit fq_codel_module_exit(void)
726 {
727 	unregister_qdisc(&fq_codel_qdisc_ops);
728 }
729 
730 module_init(fq_codel_module_init)
731 module_exit(fq_codel_module_exit)
732 MODULE_AUTHOR("Eric Dumazet");
733 MODULE_LICENSE("GPL");
734