xref: /linux/net/sched/sch_fq_codel.c (revision 74ce1896c6c65b2f8cccbf59162d542988835835)
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/pkt_cls.h>
27 #include <net/codel.h>
28 #include <net/codel_impl.h>
29 #include <net/codel_qdisc.h>
30 
31 /*	Fair Queue CoDel.
32  *
33  * Principles :
34  * Packets are classified (internal classifier or external) on flows.
35  * This is a Stochastic model (as we use a hash, several flows
36  *			       might be hashed on same slot)
37  * Each flow has a CoDel managed queue.
38  * Flows are linked onto two (Round Robin) lists,
39  * so that new flows have priority on old ones.
40  *
41  * For a given flow, packets are not reordered (CoDel uses a FIFO)
42  * head drops only.
43  * ECN capability is on by default.
44  * Low memory footprint (64 bytes per flow)
45  */
46 
47 struct fq_codel_flow {
48 	struct sk_buff	  *head;
49 	struct sk_buff	  *tail;
50 	struct list_head  flowchain;
51 	int		  deficit;
52 	u32		  dropped; /* number of drops (or ECN marks) on this flow */
53 	struct codel_vars cvars;
54 }; /* please try to keep this structure <= 64 bytes */
55 
56 struct fq_codel_sched_data {
57 	struct tcf_proto __rcu *filter_list; /* optional external classifier */
58 	struct tcf_block *block;
59 	struct fq_codel_flow *flows;	/* Flows table [flows_cnt] */
60 	u32		*backlogs;	/* backlog table [flows_cnt] */
61 	u32		flows_cnt;	/* number of flows */
62 	u32		quantum;	/* psched_mtu(qdisc_dev(sch)); */
63 	u32		drop_batch_size;
64 	u32		memory_limit;
65 	struct codel_params cparams;
66 	struct codel_stats cstats;
67 	u32		memory_usage;
68 	u32		drop_overmemory;
69 	u32		drop_overlimit;
70 	u32		new_flow_count;
71 
72 	struct list_head new_flows;	/* list of new flows */
73 	struct list_head old_flows;	/* list of old flows */
74 };
75 
76 static unsigned int fq_codel_hash(const struct fq_codel_sched_data *q,
77 				  struct sk_buff *skb)
78 {
79 	return reciprocal_scale(skb_get_hash(skb), q->flows_cnt);
80 }
81 
82 static unsigned int fq_codel_classify(struct sk_buff *skb, struct Qdisc *sch,
83 				      int *qerr)
84 {
85 	struct fq_codel_sched_data *q = qdisc_priv(sch);
86 	struct tcf_proto *filter;
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 	filter = rcu_dereference_bh(q->filter_list);
96 	if (!filter)
97 		return fq_codel_hash(q, skb) + 1;
98 
99 	*qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
100 	result = tcf_classify(skb, filter, &res, false);
101 	if (result >= 0) {
102 #ifdef CONFIG_NET_CLS_ACT
103 		switch (result) {
104 		case TC_ACT_STOLEN:
105 		case TC_ACT_QUEUED:
106 		case TC_ACT_TRAP:
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 				  struct sk_buff **to_free)
144 {
145 	struct fq_codel_sched_data *q = qdisc_priv(sch);
146 	struct sk_buff *skb;
147 	unsigned int maxbacklog = 0, idx = 0, i, len;
148 	struct fq_codel_flow *flow;
149 	unsigned int threshold;
150 	unsigned int mem = 0;
151 
152 	/* Queue is full! Find the fat flow and drop packet(s) from it.
153 	 * This might sound expensive, but with 1024 flows, we scan
154 	 * 4KB of memory, and we dont need to handle a complex tree
155 	 * in fast path (packet queue/enqueue) with many cache misses.
156 	 * In stress mode, we'll try to drop 64 packets from the flow,
157 	 * amortizing this linear lookup to one cache line per drop.
158 	 */
159 	for (i = 0; i < q->flows_cnt; i++) {
160 		if (q->backlogs[i] > maxbacklog) {
161 			maxbacklog = q->backlogs[i];
162 			idx = i;
163 		}
164 	}
165 
166 	/* Our goal is to drop half of this fat flow backlog */
167 	threshold = maxbacklog >> 1;
168 
169 	flow = &q->flows[idx];
170 	len = 0;
171 	i = 0;
172 	do {
173 		skb = dequeue_head(flow);
174 		len += qdisc_pkt_len(skb);
175 		mem += get_codel_cb(skb)->mem_usage;
176 		__qdisc_drop(skb, to_free);
177 	} while (++i < max_packets && len < threshold);
178 
179 	flow->dropped += i;
180 	q->backlogs[idx] -= len;
181 	q->memory_usage -= mem;
182 	sch->qstats.drops += i;
183 	sch->qstats.backlog -= len;
184 	sch->q.qlen -= i;
185 	return idx;
186 }
187 
188 static int fq_codel_enqueue(struct sk_buff *skb, struct Qdisc *sch,
189 			    struct sk_buff **to_free)
190 {
191 	struct fq_codel_sched_data *q = qdisc_priv(sch);
192 	unsigned int idx, prev_backlog, prev_qlen;
193 	struct fq_codel_flow *flow;
194 	int uninitialized_var(ret);
195 	unsigned int pkt_len;
196 	bool memory_limited;
197 
198 	idx = fq_codel_classify(skb, sch, &ret);
199 	if (idx == 0) {
200 		if (ret & __NET_XMIT_BYPASS)
201 			qdisc_qstats_drop(sch);
202 		__qdisc_drop(skb, to_free);
203 		return ret;
204 	}
205 	idx--;
206 
207 	codel_set_enqueue_time(skb);
208 	flow = &q->flows[idx];
209 	flow_queue_add(flow, skb);
210 	q->backlogs[idx] += qdisc_pkt_len(skb);
211 	qdisc_qstats_backlog_inc(sch, skb);
212 
213 	if (list_empty(&flow->flowchain)) {
214 		list_add_tail(&flow->flowchain, &q->new_flows);
215 		q->new_flow_count++;
216 		flow->deficit = q->quantum;
217 		flow->dropped = 0;
218 	}
219 	get_codel_cb(skb)->mem_usage = skb->truesize;
220 	q->memory_usage += get_codel_cb(skb)->mem_usage;
221 	memory_limited = q->memory_usage > q->memory_limit;
222 	if (++sch->q.qlen <= sch->limit && !memory_limited)
223 		return NET_XMIT_SUCCESS;
224 
225 	prev_backlog = sch->qstats.backlog;
226 	prev_qlen = sch->q.qlen;
227 
228 	/* save this packet length as it might be dropped by fq_codel_drop() */
229 	pkt_len = qdisc_pkt_len(skb);
230 	/* fq_codel_drop() is quite expensive, as it performs a linear search
231 	 * in q->backlogs[] to find a fat flow.
232 	 * So instead of dropping a single packet, drop half of its backlog
233 	 * with a 64 packets limit to not add a too big cpu spike here.
234 	 */
235 	ret = fq_codel_drop(sch, q->drop_batch_size, to_free);
236 
237 	prev_qlen -= sch->q.qlen;
238 	prev_backlog -= sch->qstats.backlog;
239 	q->drop_overlimit += prev_qlen;
240 	if (memory_limited)
241 		q->drop_overmemory += prev_qlen;
242 
243 	/* As we dropped packet(s), better let upper stack know this.
244 	 * If we dropped a packet for this flow, return NET_XMIT_CN,
245 	 * but in this case, our parents wont increase their backlogs.
246 	 */
247 	if (ret == idx) {
248 		qdisc_tree_reduce_backlog(sch, prev_qlen - 1,
249 					  prev_backlog - pkt_len);
250 		return NET_XMIT_CN;
251 	}
252 	qdisc_tree_reduce_backlog(sch, prev_qlen, prev_backlog);
253 	return NET_XMIT_SUCCESS;
254 }
255 
256 /* This is the specific function called from codel_dequeue()
257  * to dequeue a packet from queue. Note: backlog is handled in
258  * codel, we dont need to reduce it here.
259  */
260 static struct sk_buff *dequeue_func(struct codel_vars *vars, void *ctx)
261 {
262 	struct Qdisc *sch = ctx;
263 	struct fq_codel_sched_data *q = qdisc_priv(sch);
264 	struct fq_codel_flow *flow;
265 	struct sk_buff *skb = NULL;
266 
267 	flow = container_of(vars, struct fq_codel_flow, cvars);
268 	if (flow->head) {
269 		skb = dequeue_head(flow);
270 		q->backlogs[flow - q->flows] -= qdisc_pkt_len(skb);
271 		q->memory_usage -= get_codel_cb(skb)->mem_usage;
272 		sch->q.qlen--;
273 		sch->qstats.backlog -= qdisc_pkt_len(skb);
274 	}
275 	return skb;
276 }
277 
278 static void drop_func(struct sk_buff *skb, void *ctx)
279 {
280 	struct Qdisc *sch = ctx;
281 
282 	kfree_skb(skb);
283 	qdisc_qstats_drop(sch);
284 }
285 
286 static struct sk_buff *fq_codel_dequeue(struct Qdisc *sch)
287 {
288 	struct fq_codel_sched_data *q = qdisc_priv(sch);
289 	struct sk_buff *skb;
290 	struct fq_codel_flow *flow;
291 	struct list_head *head;
292 	u32 prev_drop_count, prev_ecn_mark;
293 
294 begin:
295 	head = &q->new_flows;
296 	if (list_empty(head)) {
297 		head = &q->old_flows;
298 		if (list_empty(head))
299 			return NULL;
300 	}
301 	flow = list_first_entry(head, struct fq_codel_flow, flowchain);
302 
303 	if (flow->deficit <= 0) {
304 		flow->deficit += q->quantum;
305 		list_move_tail(&flow->flowchain, &q->old_flows);
306 		goto begin;
307 	}
308 
309 	prev_drop_count = q->cstats.drop_count;
310 	prev_ecn_mark = q->cstats.ecn_mark;
311 
312 	skb = codel_dequeue(sch, &sch->qstats.backlog, &q->cparams,
313 			    &flow->cvars, &q->cstats, qdisc_pkt_len,
314 			    codel_get_enqueue_time, drop_func, dequeue_func);
315 
316 	flow->dropped += q->cstats.drop_count - prev_drop_count;
317 	flow->dropped += q->cstats.ecn_mark - prev_ecn_mark;
318 
319 	if (!skb) {
320 		/* force a pass through old_flows to prevent starvation */
321 		if ((head == &q->new_flows) && !list_empty(&q->old_flows))
322 			list_move_tail(&flow->flowchain, &q->old_flows);
323 		else
324 			list_del_init(&flow->flowchain);
325 		goto begin;
326 	}
327 	qdisc_bstats_update(sch, skb);
328 	flow->deficit -= qdisc_pkt_len(skb);
329 	/* We cant call qdisc_tree_reduce_backlog() if our qlen is 0,
330 	 * or HTB crashes. Defer it for next round.
331 	 */
332 	if (q->cstats.drop_count && sch->q.qlen) {
333 		qdisc_tree_reduce_backlog(sch, q->cstats.drop_count,
334 					  q->cstats.drop_len);
335 		q->cstats.drop_count = 0;
336 		q->cstats.drop_len = 0;
337 	}
338 	return skb;
339 }
340 
341 static void fq_codel_flow_purge(struct fq_codel_flow *flow)
342 {
343 	rtnl_kfree_skbs(flow->head, flow->tail);
344 	flow->head = NULL;
345 }
346 
347 static void fq_codel_reset(struct Qdisc *sch)
348 {
349 	struct fq_codel_sched_data *q = qdisc_priv(sch);
350 	int i;
351 
352 	INIT_LIST_HEAD(&q->new_flows);
353 	INIT_LIST_HEAD(&q->old_flows);
354 	for (i = 0; i < q->flows_cnt; i++) {
355 		struct fq_codel_flow *flow = q->flows + i;
356 
357 		fq_codel_flow_purge(flow);
358 		INIT_LIST_HEAD(&flow->flowchain);
359 		codel_vars_init(&flow->cvars);
360 	}
361 	memset(q->backlogs, 0, q->flows_cnt * sizeof(u32));
362 	sch->q.qlen = 0;
363 	sch->qstats.backlog = 0;
364 	q->memory_usage = 0;
365 }
366 
367 static const struct nla_policy fq_codel_policy[TCA_FQ_CODEL_MAX + 1] = {
368 	[TCA_FQ_CODEL_TARGET]	= { .type = NLA_U32 },
369 	[TCA_FQ_CODEL_LIMIT]	= { .type = NLA_U32 },
370 	[TCA_FQ_CODEL_INTERVAL]	= { .type = NLA_U32 },
371 	[TCA_FQ_CODEL_ECN]	= { .type = NLA_U32 },
372 	[TCA_FQ_CODEL_FLOWS]	= { .type = NLA_U32 },
373 	[TCA_FQ_CODEL_QUANTUM]	= { .type = NLA_U32 },
374 	[TCA_FQ_CODEL_CE_THRESHOLD] = { .type = NLA_U32 },
375 	[TCA_FQ_CODEL_DROP_BATCH_SIZE] = { .type = NLA_U32 },
376 	[TCA_FQ_CODEL_MEMORY_LIMIT] = { .type = NLA_U32 },
377 };
378 
379 static int fq_codel_change(struct Qdisc *sch, struct nlattr *opt)
380 {
381 	struct fq_codel_sched_data *q = qdisc_priv(sch);
382 	struct nlattr *tb[TCA_FQ_CODEL_MAX + 1];
383 	int err;
384 
385 	if (!opt)
386 		return -EINVAL;
387 
388 	err = nla_parse_nested(tb, TCA_FQ_CODEL_MAX, opt, fq_codel_policy,
389 			       NULL);
390 	if (err < 0)
391 		return err;
392 	if (tb[TCA_FQ_CODEL_FLOWS]) {
393 		if (q->flows)
394 			return -EINVAL;
395 		q->flows_cnt = nla_get_u32(tb[TCA_FQ_CODEL_FLOWS]);
396 		if (!q->flows_cnt ||
397 		    q->flows_cnt > 65536)
398 			return -EINVAL;
399 	}
400 	sch_tree_lock(sch);
401 
402 	if (tb[TCA_FQ_CODEL_TARGET]) {
403 		u64 target = nla_get_u32(tb[TCA_FQ_CODEL_TARGET]);
404 
405 		q->cparams.target = (target * NSEC_PER_USEC) >> CODEL_SHIFT;
406 	}
407 
408 	if (tb[TCA_FQ_CODEL_CE_THRESHOLD]) {
409 		u64 val = nla_get_u32(tb[TCA_FQ_CODEL_CE_THRESHOLD]);
410 
411 		q->cparams.ce_threshold = (val * NSEC_PER_USEC) >> CODEL_SHIFT;
412 	}
413 
414 	if (tb[TCA_FQ_CODEL_INTERVAL]) {
415 		u64 interval = nla_get_u32(tb[TCA_FQ_CODEL_INTERVAL]);
416 
417 		q->cparams.interval = (interval * NSEC_PER_USEC) >> CODEL_SHIFT;
418 	}
419 
420 	if (tb[TCA_FQ_CODEL_LIMIT])
421 		sch->limit = nla_get_u32(tb[TCA_FQ_CODEL_LIMIT]);
422 
423 	if (tb[TCA_FQ_CODEL_ECN])
424 		q->cparams.ecn = !!nla_get_u32(tb[TCA_FQ_CODEL_ECN]);
425 
426 	if (tb[TCA_FQ_CODEL_QUANTUM])
427 		q->quantum = max(256U, nla_get_u32(tb[TCA_FQ_CODEL_QUANTUM]));
428 
429 	if (tb[TCA_FQ_CODEL_DROP_BATCH_SIZE])
430 		q->drop_batch_size = min(1U, nla_get_u32(tb[TCA_FQ_CODEL_DROP_BATCH_SIZE]));
431 
432 	if (tb[TCA_FQ_CODEL_MEMORY_LIMIT])
433 		q->memory_limit = min(1U << 31, nla_get_u32(tb[TCA_FQ_CODEL_MEMORY_LIMIT]));
434 
435 	while (sch->q.qlen > sch->limit ||
436 	       q->memory_usage > q->memory_limit) {
437 		struct sk_buff *skb = fq_codel_dequeue(sch);
438 
439 		q->cstats.drop_len += qdisc_pkt_len(skb);
440 		rtnl_kfree_skbs(skb, skb);
441 		q->cstats.drop_count++;
442 	}
443 	qdisc_tree_reduce_backlog(sch, q->cstats.drop_count, q->cstats.drop_len);
444 	q->cstats.drop_count = 0;
445 	q->cstats.drop_len = 0;
446 
447 	sch_tree_unlock(sch);
448 	return 0;
449 }
450 
451 static void fq_codel_destroy(struct Qdisc *sch)
452 {
453 	struct fq_codel_sched_data *q = qdisc_priv(sch);
454 
455 	tcf_block_put(q->block);
456 	kvfree(q->backlogs);
457 	kvfree(q->flows);
458 }
459 
460 static int fq_codel_init(struct Qdisc *sch, struct nlattr *opt)
461 {
462 	struct fq_codel_sched_data *q = qdisc_priv(sch);
463 	int i;
464 	int err;
465 
466 	sch->limit = 10*1024;
467 	q->flows_cnt = 1024;
468 	q->memory_limit = 32 << 20; /* 32 MBytes */
469 	q->drop_batch_size = 64;
470 	q->quantum = psched_mtu(qdisc_dev(sch));
471 	INIT_LIST_HEAD(&q->new_flows);
472 	INIT_LIST_HEAD(&q->old_flows);
473 	codel_params_init(&q->cparams);
474 	codel_stats_init(&q->cstats);
475 	q->cparams.ecn = true;
476 	q->cparams.mtu = psched_mtu(qdisc_dev(sch));
477 
478 	if (opt) {
479 		int err = fq_codel_change(sch, opt);
480 		if (err)
481 			return err;
482 	}
483 
484 	err = tcf_block_get(&q->block, &q->filter_list);
485 	if (err)
486 		return err;
487 
488 	if (!q->flows) {
489 		q->flows = kvzalloc(q->flows_cnt *
490 					   sizeof(struct fq_codel_flow), GFP_KERNEL);
491 		if (!q->flows)
492 			return -ENOMEM;
493 		q->backlogs = kvzalloc(q->flows_cnt * sizeof(u32), GFP_KERNEL);
494 		if (!q->backlogs)
495 			return -ENOMEM;
496 		for (i = 0; i < q->flows_cnt; i++) {
497 			struct fq_codel_flow *flow = q->flows + i;
498 
499 			INIT_LIST_HEAD(&flow->flowchain);
500 			codel_vars_init(&flow->cvars);
501 		}
502 	}
503 	if (sch->limit >= 1)
504 		sch->flags |= TCQ_F_CAN_BYPASS;
505 	else
506 		sch->flags &= ~TCQ_F_CAN_BYPASS;
507 	return 0;
508 }
509 
510 static int fq_codel_dump(struct Qdisc *sch, struct sk_buff *skb)
511 {
512 	struct fq_codel_sched_data *q = qdisc_priv(sch);
513 	struct nlattr *opts;
514 
515 	opts = nla_nest_start(skb, TCA_OPTIONS);
516 	if (opts == NULL)
517 		goto nla_put_failure;
518 
519 	if (nla_put_u32(skb, TCA_FQ_CODEL_TARGET,
520 			codel_time_to_us(q->cparams.target)) ||
521 	    nla_put_u32(skb, TCA_FQ_CODEL_LIMIT,
522 			sch->limit) ||
523 	    nla_put_u32(skb, TCA_FQ_CODEL_INTERVAL,
524 			codel_time_to_us(q->cparams.interval)) ||
525 	    nla_put_u32(skb, TCA_FQ_CODEL_ECN,
526 			q->cparams.ecn) ||
527 	    nla_put_u32(skb, TCA_FQ_CODEL_QUANTUM,
528 			q->quantum) ||
529 	    nla_put_u32(skb, TCA_FQ_CODEL_DROP_BATCH_SIZE,
530 			q->drop_batch_size) ||
531 	    nla_put_u32(skb, TCA_FQ_CODEL_MEMORY_LIMIT,
532 			q->memory_limit) ||
533 	    nla_put_u32(skb, TCA_FQ_CODEL_FLOWS,
534 			q->flows_cnt))
535 		goto nla_put_failure;
536 
537 	if (q->cparams.ce_threshold != CODEL_DISABLED_THRESHOLD &&
538 	    nla_put_u32(skb, TCA_FQ_CODEL_CE_THRESHOLD,
539 			codel_time_to_us(q->cparams.ce_threshold)))
540 		goto nla_put_failure;
541 
542 	return nla_nest_end(skb, opts);
543 
544 nla_put_failure:
545 	return -1;
546 }
547 
548 static int fq_codel_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
549 {
550 	struct fq_codel_sched_data *q = qdisc_priv(sch);
551 	struct tc_fq_codel_xstats st = {
552 		.type				= TCA_FQ_CODEL_XSTATS_QDISC,
553 	};
554 	struct list_head *pos;
555 
556 	st.qdisc_stats.maxpacket = q->cstats.maxpacket;
557 	st.qdisc_stats.drop_overlimit = q->drop_overlimit;
558 	st.qdisc_stats.ecn_mark = q->cstats.ecn_mark;
559 	st.qdisc_stats.new_flow_count = q->new_flow_count;
560 	st.qdisc_stats.ce_mark = q->cstats.ce_mark;
561 	st.qdisc_stats.memory_usage  = q->memory_usage;
562 	st.qdisc_stats.drop_overmemory = q->drop_overmemory;
563 
564 	sch_tree_lock(sch);
565 	list_for_each(pos, &q->new_flows)
566 		st.qdisc_stats.new_flows_len++;
567 
568 	list_for_each(pos, &q->old_flows)
569 		st.qdisc_stats.old_flows_len++;
570 	sch_tree_unlock(sch);
571 
572 	return gnet_stats_copy_app(d, &st, sizeof(st));
573 }
574 
575 static struct Qdisc *fq_codel_leaf(struct Qdisc *sch, unsigned long arg)
576 {
577 	return NULL;
578 }
579 
580 static unsigned long fq_codel_find(struct Qdisc *sch, u32 classid)
581 {
582 	return 0;
583 }
584 
585 static unsigned long fq_codel_bind(struct Qdisc *sch, unsigned long parent,
586 			      u32 classid)
587 {
588 	/* we cannot bypass queue discipline anymore */
589 	sch->flags &= ~TCQ_F_CAN_BYPASS;
590 	return 0;
591 }
592 
593 static void fq_codel_unbind(struct Qdisc *q, unsigned long cl)
594 {
595 }
596 
597 static struct tcf_block *fq_codel_tcf_block(struct Qdisc *sch, unsigned long cl)
598 {
599 	struct fq_codel_sched_data *q = qdisc_priv(sch);
600 
601 	if (cl)
602 		return NULL;
603 	return q->block;
604 }
605 
606 static int fq_codel_dump_class(struct Qdisc *sch, unsigned long cl,
607 			  struct sk_buff *skb, struct tcmsg *tcm)
608 {
609 	tcm->tcm_handle |= TC_H_MIN(cl);
610 	return 0;
611 }
612 
613 static int fq_codel_dump_class_stats(struct Qdisc *sch, unsigned long cl,
614 				     struct gnet_dump *d)
615 {
616 	struct fq_codel_sched_data *q = qdisc_priv(sch);
617 	u32 idx = cl - 1;
618 	struct gnet_stats_queue qs = { 0 };
619 	struct tc_fq_codel_xstats xstats;
620 
621 	if (idx < q->flows_cnt) {
622 		const struct fq_codel_flow *flow = &q->flows[idx];
623 		const struct sk_buff *skb;
624 
625 		memset(&xstats, 0, sizeof(xstats));
626 		xstats.type = TCA_FQ_CODEL_XSTATS_CLASS;
627 		xstats.class_stats.deficit = flow->deficit;
628 		xstats.class_stats.ldelay =
629 			codel_time_to_us(flow->cvars.ldelay);
630 		xstats.class_stats.count = flow->cvars.count;
631 		xstats.class_stats.lastcount = flow->cvars.lastcount;
632 		xstats.class_stats.dropping = flow->cvars.dropping;
633 		if (flow->cvars.dropping) {
634 			codel_tdiff_t delta = flow->cvars.drop_next -
635 					      codel_get_time();
636 
637 			xstats.class_stats.drop_next = (delta >= 0) ?
638 				codel_time_to_us(delta) :
639 				-codel_time_to_us(-delta);
640 		}
641 		if (flow->head) {
642 			sch_tree_lock(sch);
643 			skb = flow->head;
644 			while (skb) {
645 				qs.qlen++;
646 				skb = skb->next;
647 			}
648 			sch_tree_unlock(sch);
649 		}
650 		qs.backlog = q->backlogs[idx];
651 		qs.drops = flow->dropped;
652 	}
653 	if (gnet_stats_copy_queue(d, NULL, &qs, qs.qlen) < 0)
654 		return -1;
655 	if (idx < q->flows_cnt)
656 		return gnet_stats_copy_app(d, &xstats, sizeof(xstats));
657 	return 0;
658 }
659 
660 static void fq_codel_walk(struct Qdisc *sch, struct qdisc_walker *arg)
661 {
662 	struct fq_codel_sched_data *q = qdisc_priv(sch);
663 	unsigned int i;
664 
665 	if (arg->stop)
666 		return;
667 
668 	for (i = 0; i < q->flows_cnt; i++) {
669 		if (list_empty(&q->flows[i].flowchain) ||
670 		    arg->count < arg->skip) {
671 			arg->count++;
672 			continue;
673 		}
674 		if (arg->fn(sch, i + 1, arg) < 0) {
675 			arg->stop = 1;
676 			break;
677 		}
678 		arg->count++;
679 	}
680 }
681 
682 static const struct Qdisc_class_ops fq_codel_class_ops = {
683 	.leaf		=	fq_codel_leaf,
684 	.find		=	fq_codel_find,
685 	.tcf_block	=	fq_codel_tcf_block,
686 	.bind_tcf	=	fq_codel_bind,
687 	.unbind_tcf	=	fq_codel_unbind,
688 	.dump		=	fq_codel_dump_class,
689 	.dump_stats	=	fq_codel_dump_class_stats,
690 	.walk		=	fq_codel_walk,
691 };
692 
693 static struct Qdisc_ops fq_codel_qdisc_ops __read_mostly = {
694 	.cl_ops		=	&fq_codel_class_ops,
695 	.id		=	"fq_codel",
696 	.priv_size	=	sizeof(struct fq_codel_sched_data),
697 	.enqueue	=	fq_codel_enqueue,
698 	.dequeue	=	fq_codel_dequeue,
699 	.peek		=	qdisc_peek_dequeued,
700 	.init		=	fq_codel_init,
701 	.reset		=	fq_codel_reset,
702 	.destroy	=	fq_codel_destroy,
703 	.change		=	fq_codel_change,
704 	.dump		=	fq_codel_dump,
705 	.dump_stats =	fq_codel_dump_stats,
706 	.owner		=	THIS_MODULE,
707 };
708 
709 static int __init fq_codel_module_init(void)
710 {
711 	return register_qdisc(&fq_codel_qdisc_ops);
712 }
713 
714 static void __exit fq_codel_module_exit(void)
715 {
716 	unregister_qdisc(&fq_codel_qdisc_ops);
717 }
718 
719 module_init(fq_codel_module_init)
720 module_exit(fq_codel_module_exit)
721 MODULE_AUTHOR("Eric Dumazet");
722 MODULE_LICENSE("GPL");
723