xref: /openbmc/linux/net/sched/sch_fq_codel.c (revision 3fc41476)
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 			/* fall through */
109 		case TC_ACT_SHOT:
110 			return 0;
111 		}
112 #endif
113 		if (TC_H_MIN(res.classid) <= q->flows_cnt)
114 			return TC_H_MIN(res.classid);
115 	}
116 	return 0;
117 }
118 
119 /* helper functions : might be changed when/if skb use a standard list_head */
120 
121 /* remove one skb from head of slot queue */
122 static inline struct sk_buff *dequeue_head(struct fq_codel_flow *flow)
123 {
124 	struct sk_buff *skb = flow->head;
125 
126 	flow->head = skb->next;
127 	skb_mark_not_on_list(skb);
128 	return skb;
129 }
130 
131 /* add skb to flow queue (tail add) */
132 static inline void flow_queue_add(struct fq_codel_flow *flow,
133 				  struct sk_buff *skb)
134 {
135 	if (flow->head == NULL)
136 		flow->head = skb;
137 	else
138 		flow->tail->next = skb;
139 	flow->tail = skb;
140 	skb->next = NULL;
141 }
142 
143 static unsigned int fq_codel_drop(struct Qdisc *sch, unsigned int max_packets,
144 				  struct sk_buff **to_free)
145 {
146 	struct fq_codel_sched_data *q = qdisc_priv(sch);
147 	struct sk_buff *skb;
148 	unsigned int maxbacklog = 0, idx = 0, i, len;
149 	struct fq_codel_flow *flow;
150 	unsigned int threshold;
151 	unsigned int mem = 0;
152 
153 	/* Queue is full! Find the fat flow and drop packet(s) from it.
154 	 * This might sound expensive, but with 1024 flows, we scan
155 	 * 4KB of memory, and we dont need to handle a complex tree
156 	 * in fast path (packet queue/enqueue) with many cache misses.
157 	 * In stress mode, we'll try to drop 64 packets from the flow,
158 	 * amortizing this linear lookup to one cache line per drop.
159 	 */
160 	for (i = 0; i < q->flows_cnt; i++) {
161 		if (q->backlogs[i] > maxbacklog) {
162 			maxbacklog = q->backlogs[i];
163 			idx = i;
164 		}
165 	}
166 
167 	/* Our goal is to drop half of this fat flow backlog */
168 	threshold = maxbacklog >> 1;
169 
170 	flow = &q->flows[idx];
171 	len = 0;
172 	i = 0;
173 	do {
174 		skb = dequeue_head(flow);
175 		len += qdisc_pkt_len(skb);
176 		mem += get_codel_cb(skb)->mem_usage;
177 		__qdisc_drop(skb, to_free);
178 	} while (++i < max_packets && len < threshold);
179 
180 	flow->dropped += i;
181 	q->backlogs[idx] -= len;
182 	q->memory_usage -= mem;
183 	sch->qstats.drops += i;
184 	sch->qstats.backlog -= len;
185 	sch->q.qlen -= i;
186 	return idx;
187 }
188 
189 static int fq_codel_enqueue(struct sk_buff *skb, struct Qdisc *sch,
190 			    struct sk_buff **to_free)
191 {
192 	struct fq_codel_sched_data *q = qdisc_priv(sch);
193 	unsigned int idx, prev_backlog, prev_qlen;
194 	struct fq_codel_flow *flow;
195 	int uninitialized_var(ret);
196 	unsigned int pkt_len;
197 	bool memory_limited;
198 
199 	idx = fq_codel_classify(skb, sch, &ret);
200 	if (idx == 0) {
201 		if (ret & __NET_XMIT_BYPASS)
202 			qdisc_qstats_drop(sch);
203 		__qdisc_drop(skb, to_free);
204 		return ret;
205 	}
206 	idx--;
207 
208 	codel_set_enqueue_time(skb);
209 	flow = &q->flows[idx];
210 	flow_queue_add(flow, skb);
211 	q->backlogs[idx] += qdisc_pkt_len(skb);
212 	qdisc_qstats_backlog_inc(sch, skb);
213 
214 	if (list_empty(&flow->flowchain)) {
215 		list_add_tail(&flow->flowchain, &q->new_flows);
216 		q->new_flow_count++;
217 		flow->deficit = q->quantum;
218 		flow->dropped = 0;
219 	}
220 	get_codel_cb(skb)->mem_usage = skb->truesize;
221 	q->memory_usage += get_codel_cb(skb)->mem_usage;
222 	memory_limited = q->memory_usage > q->memory_limit;
223 	if (++sch->q.qlen <= sch->limit && !memory_limited)
224 		return NET_XMIT_SUCCESS;
225 
226 	prev_backlog = sch->qstats.backlog;
227 	prev_qlen = sch->q.qlen;
228 
229 	/* save this packet length as it might be dropped by fq_codel_drop() */
230 	pkt_len = qdisc_pkt_len(skb);
231 	/* fq_codel_drop() is quite expensive, as it performs a linear search
232 	 * in q->backlogs[] to find a fat flow.
233 	 * So instead of dropping a single packet, drop half of its backlog
234 	 * with a 64 packets limit to not add a too big cpu spike here.
235 	 */
236 	ret = fq_codel_drop(sch, q->drop_batch_size, to_free);
237 
238 	prev_qlen -= sch->q.qlen;
239 	prev_backlog -= sch->qstats.backlog;
240 	q->drop_overlimit += prev_qlen;
241 	if (memory_limited)
242 		q->drop_overmemory += prev_qlen;
243 
244 	/* As we dropped packet(s), better let upper stack know this.
245 	 * If we dropped a packet for this flow, return NET_XMIT_CN,
246 	 * but in this case, our parents wont increase their backlogs.
247 	 */
248 	if (ret == idx) {
249 		qdisc_tree_reduce_backlog(sch, prev_qlen - 1,
250 					  prev_backlog - pkt_len);
251 		return NET_XMIT_CN;
252 	}
253 	qdisc_tree_reduce_backlog(sch, prev_qlen, prev_backlog);
254 	return NET_XMIT_SUCCESS;
255 }
256 
257 /* This is the specific function called from codel_dequeue()
258  * to dequeue a packet from queue. Note: backlog is handled in
259  * codel, we dont need to reduce it here.
260  */
261 static struct sk_buff *dequeue_func(struct codel_vars *vars, void *ctx)
262 {
263 	struct Qdisc *sch = ctx;
264 	struct fq_codel_sched_data *q = qdisc_priv(sch);
265 	struct fq_codel_flow *flow;
266 	struct sk_buff *skb = NULL;
267 
268 	flow = container_of(vars, struct fq_codel_flow, cvars);
269 	if (flow->head) {
270 		skb = dequeue_head(flow);
271 		q->backlogs[flow - q->flows] -= qdisc_pkt_len(skb);
272 		q->memory_usage -= get_codel_cb(skb)->mem_usage;
273 		sch->q.qlen--;
274 		sch->qstats.backlog -= qdisc_pkt_len(skb);
275 	}
276 	return skb;
277 }
278 
279 static void drop_func(struct sk_buff *skb, void *ctx)
280 {
281 	struct Qdisc *sch = ctx;
282 
283 	kfree_skb(skb);
284 	qdisc_qstats_drop(sch);
285 }
286 
287 static struct sk_buff *fq_codel_dequeue(struct Qdisc *sch)
288 {
289 	struct fq_codel_sched_data *q = qdisc_priv(sch);
290 	struct sk_buff *skb;
291 	struct fq_codel_flow *flow;
292 	struct list_head *head;
293 	u32 prev_drop_count, prev_ecn_mark;
294 
295 begin:
296 	head = &q->new_flows;
297 	if (list_empty(head)) {
298 		head = &q->old_flows;
299 		if (list_empty(head))
300 			return NULL;
301 	}
302 	flow = list_first_entry(head, struct fq_codel_flow, flowchain);
303 
304 	if (flow->deficit <= 0) {
305 		flow->deficit += q->quantum;
306 		list_move_tail(&flow->flowchain, &q->old_flows);
307 		goto begin;
308 	}
309 
310 	prev_drop_count = q->cstats.drop_count;
311 	prev_ecn_mark = q->cstats.ecn_mark;
312 
313 	skb = codel_dequeue(sch, &sch->qstats.backlog, &q->cparams,
314 			    &flow->cvars, &q->cstats, qdisc_pkt_len,
315 			    codel_get_enqueue_time, drop_func, dequeue_func);
316 
317 	flow->dropped += q->cstats.drop_count - prev_drop_count;
318 	flow->dropped += q->cstats.ecn_mark - prev_ecn_mark;
319 
320 	if (!skb) {
321 		/* force a pass through old_flows to prevent starvation */
322 		if ((head == &q->new_flows) && !list_empty(&q->old_flows))
323 			list_move_tail(&flow->flowchain, &q->old_flows);
324 		else
325 			list_del_init(&flow->flowchain);
326 		goto begin;
327 	}
328 	qdisc_bstats_update(sch, skb);
329 	flow->deficit -= qdisc_pkt_len(skb);
330 	/* We cant call qdisc_tree_reduce_backlog() if our qlen is 0,
331 	 * or HTB crashes. Defer it for next round.
332 	 */
333 	if (q->cstats.drop_count && sch->q.qlen) {
334 		qdisc_tree_reduce_backlog(sch, q->cstats.drop_count,
335 					  q->cstats.drop_len);
336 		q->cstats.drop_count = 0;
337 		q->cstats.drop_len = 0;
338 	}
339 	return skb;
340 }
341 
342 static void fq_codel_flow_purge(struct fq_codel_flow *flow)
343 {
344 	rtnl_kfree_skbs(flow->head, flow->tail);
345 	flow->head = NULL;
346 }
347 
348 static void fq_codel_reset(struct Qdisc *sch)
349 {
350 	struct fq_codel_sched_data *q = qdisc_priv(sch);
351 	int i;
352 
353 	INIT_LIST_HEAD(&q->new_flows);
354 	INIT_LIST_HEAD(&q->old_flows);
355 	for (i = 0; i < q->flows_cnt; i++) {
356 		struct fq_codel_flow *flow = q->flows + i;
357 
358 		fq_codel_flow_purge(flow);
359 		INIT_LIST_HEAD(&flow->flowchain);
360 		codel_vars_init(&flow->cvars);
361 	}
362 	memset(q->backlogs, 0, q->flows_cnt * sizeof(u32));
363 	sch->q.qlen = 0;
364 	sch->qstats.backlog = 0;
365 	q->memory_usage = 0;
366 }
367 
368 static const struct nla_policy fq_codel_policy[TCA_FQ_CODEL_MAX + 1] = {
369 	[TCA_FQ_CODEL_TARGET]	= { .type = NLA_U32 },
370 	[TCA_FQ_CODEL_LIMIT]	= { .type = NLA_U32 },
371 	[TCA_FQ_CODEL_INTERVAL]	= { .type = NLA_U32 },
372 	[TCA_FQ_CODEL_ECN]	= { .type = NLA_U32 },
373 	[TCA_FQ_CODEL_FLOWS]	= { .type = NLA_U32 },
374 	[TCA_FQ_CODEL_QUANTUM]	= { .type = NLA_U32 },
375 	[TCA_FQ_CODEL_CE_THRESHOLD] = { .type = NLA_U32 },
376 	[TCA_FQ_CODEL_DROP_BATCH_SIZE] = { .type = NLA_U32 },
377 	[TCA_FQ_CODEL_MEMORY_LIMIT] = { .type = NLA_U32 },
378 };
379 
380 static int fq_codel_change(struct Qdisc *sch, struct nlattr *opt,
381 			   struct netlink_ext_ack *extack)
382 {
383 	struct fq_codel_sched_data *q = qdisc_priv(sch);
384 	struct nlattr *tb[TCA_FQ_CODEL_MAX + 1];
385 	int err;
386 
387 	if (!opt)
388 		return -EINVAL;
389 
390 	err = nla_parse_nested_deprecated(tb, TCA_FQ_CODEL_MAX, opt,
391 					  fq_codel_policy, NULL);
392 	if (err < 0)
393 		return err;
394 	if (tb[TCA_FQ_CODEL_FLOWS]) {
395 		if (q->flows)
396 			return -EINVAL;
397 		q->flows_cnt = nla_get_u32(tb[TCA_FQ_CODEL_FLOWS]);
398 		if (!q->flows_cnt ||
399 		    q->flows_cnt > 65536)
400 			return -EINVAL;
401 	}
402 	sch_tree_lock(sch);
403 
404 	if (tb[TCA_FQ_CODEL_TARGET]) {
405 		u64 target = nla_get_u32(tb[TCA_FQ_CODEL_TARGET]);
406 
407 		q->cparams.target = (target * NSEC_PER_USEC) >> CODEL_SHIFT;
408 	}
409 
410 	if (tb[TCA_FQ_CODEL_CE_THRESHOLD]) {
411 		u64 val = nla_get_u32(tb[TCA_FQ_CODEL_CE_THRESHOLD]);
412 
413 		q->cparams.ce_threshold = (val * NSEC_PER_USEC) >> CODEL_SHIFT;
414 	}
415 
416 	if (tb[TCA_FQ_CODEL_INTERVAL]) {
417 		u64 interval = nla_get_u32(tb[TCA_FQ_CODEL_INTERVAL]);
418 
419 		q->cparams.interval = (interval * NSEC_PER_USEC) >> CODEL_SHIFT;
420 	}
421 
422 	if (tb[TCA_FQ_CODEL_LIMIT])
423 		sch->limit = nla_get_u32(tb[TCA_FQ_CODEL_LIMIT]);
424 
425 	if (tb[TCA_FQ_CODEL_ECN])
426 		q->cparams.ecn = !!nla_get_u32(tb[TCA_FQ_CODEL_ECN]);
427 
428 	if (tb[TCA_FQ_CODEL_QUANTUM])
429 		q->quantum = max(256U, nla_get_u32(tb[TCA_FQ_CODEL_QUANTUM]));
430 
431 	if (tb[TCA_FQ_CODEL_DROP_BATCH_SIZE])
432 		q->drop_batch_size = min(1U, nla_get_u32(tb[TCA_FQ_CODEL_DROP_BATCH_SIZE]));
433 
434 	if (tb[TCA_FQ_CODEL_MEMORY_LIMIT])
435 		q->memory_limit = min(1U << 31, nla_get_u32(tb[TCA_FQ_CODEL_MEMORY_LIMIT]));
436 
437 	while (sch->q.qlen > sch->limit ||
438 	       q->memory_usage > q->memory_limit) {
439 		struct sk_buff *skb = fq_codel_dequeue(sch);
440 
441 		q->cstats.drop_len += qdisc_pkt_len(skb);
442 		rtnl_kfree_skbs(skb, skb);
443 		q->cstats.drop_count++;
444 	}
445 	qdisc_tree_reduce_backlog(sch, q->cstats.drop_count, q->cstats.drop_len);
446 	q->cstats.drop_count = 0;
447 	q->cstats.drop_len = 0;
448 
449 	sch_tree_unlock(sch);
450 	return 0;
451 }
452 
453 static void fq_codel_destroy(struct Qdisc *sch)
454 {
455 	struct fq_codel_sched_data *q = qdisc_priv(sch);
456 
457 	tcf_block_put(q->block);
458 	kvfree(q->backlogs);
459 	kvfree(q->flows);
460 }
461 
462 static int fq_codel_init(struct Qdisc *sch, struct nlattr *opt,
463 			 struct netlink_ext_ack *extack)
464 {
465 	struct fq_codel_sched_data *q = qdisc_priv(sch);
466 	int i;
467 	int err;
468 
469 	sch->limit = 10*1024;
470 	q->flows_cnt = 1024;
471 	q->memory_limit = 32 << 20; /* 32 MBytes */
472 	q->drop_batch_size = 64;
473 	q->quantum = psched_mtu(qdisc_dev(sch));
474 	INIT_LIST_HEAD(&q->new_flows);
475 	INIT_LIST_HEAD(&q->old_flows);
476 	codel_params_init(&q->cparams);
477 	codel_stats_init(&q->cstats);
478 	q->cparams.ecn = true;
479 	q->cparams.mtu = psched_mtu(qdisc_dev(sch));
480 
481 	if (opt) {
482 		err = fq_codel_change(sch, opt, extack);
483 		if (err)
484 			goto init_failure;
485 	}
486 
487 	err = tcf_block_get(&q->block, &q->filter_list, sch, extack);
488 	if (err)
489 		goto init_failure;
490 
491 	if (!q->flows) {
492 		q->flows = kvcalloc(q->flows_cnt,
493 				    sizeof(struct fq_codel_flow),
494 				    GFP_KERNEL);
495 		if (!q->flows) {
496 			err = -ENOMEM;
497 			goto init_failure;
498 		}
499 		q->backlogs = kvcalloc(q->flows_cnt, sizeof(u32), GFP_KERNEL);
500 		if (!q->backlogs) {
501 			err = -ENOMEM;
502 			goto alloc_failure;
503 		}
504 		for (i = 0; i < q->flows_cnt; i++) {
505 			struct fq_codel_flow *flow = q->flows + i;
506 
507 			INIT_LIST_HEAD(&flow->flowchain);
508 			codel_vars_init(&flow->cvars);
509 		}
510 	}
511 	if (sch->limit >= 1)
512 		sch->flags |= TCQ_F_CAN_BYPASS;
513 	else
514 		sch->flags &= ~TCQ_F_CAN_BYPASS;
515 	return 0;
516 
517 alloc_failure:
518 	kvfree(q->flows);
519 	q->flows = NULL;
520 init_failure:
521 	q->flows_cnt = 0;
522 	return err;
523 }
524 
525 static int fq_codel_dump(struct Qdisc *sch, struct sk_buff *skb)
526 {
527 	struct fq_codel_sched_data *q = qdisc_priv(sch);
528 	struct nlattr *opts;
529 
530 	opts = nla_nest_start_noflag(skb, TCA_OPTIONS);
531 	if (opts == NULL)
532 		goto nla_put_failure;
533 
534 	if (nla_put_u32(skb, TCA_FQ_CODEL_TARGET,
535 			codel_time_to_us(q->cparams.target)) ||
536 	    nla_put_u32(skb, TCA_FQ_CODEL_LIMIT,
537 			sch->limit) ||
538 	    nla_put_u32(skb, TCA_FQ_CODEL_INTERVAL,
539 			codel_time_to_us(q->cparams.interval)) ||
540 	    nla_put_u32(skb, TCA_FQ_CODEL_ECN,
541 			q->cparams.ecn) ||
542 	    nla_put_u32(skb, TCA_FQ_CODEL_QUANTUM,
543 			q->quantum) ||
544 	    nla_put_u32(skb, TCA_FQ_CODEL_DROP_BATCH_SIZE,
545 			q->drop_batch_size) ||
546 	    nla_put_u32(skb, TCA_FQ_CODEL_MEMORY_LIMIT,
547 			q->memory_limit) ||
548 	    nla_put_u32(skb, TCA_FQ_CODEL_FLOWS,
549 			q->flows_cnt))
550 		goto nla_put_failure;
551 
552 	if (q->cparams.ce_threshold != CODEL_DISABLED_THRESHOLD &&
553 	    nla_put_u32(skb, TCA_FQ_CODEL_CE_THRESHOLD,
554 			codel_time_to_us(q->cparams.ce_threshold)))
555 		goto nla_put_failure;
556 
557 	return nla_nest_end(skb, opts);
558 
559 nla_put_failure:
560 	return -1;
561 }
562 
563 static int fq_codel_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
564 {
565 	struct fq_codel_sched_data *q = qdisc_priv(sch);
566 	struct tc_fq_codel_xstats st = {
567 		.type				= TCA_FQ_CODEL_XSTATS_QDISC,
568 	};
569 	struct list_head *pos;
570 
571 	st.qdisc_stats.maxpacket = q->cstats.maxpacket;
572 	st.qdisc_stats.drop_overlimit = q->drop_overlimit;
573 	st.qdisc_stats.ecn_mark = q->cstats.ecn_mark;
574 	st.qdisc_stats.new_flow_count = q->new_flow_count;
575 	st.qdisc_stats.ce_mark = q->cstats.ce_mark;
576 	st.qdisc_stats.memory_usage  = q->memory_usage;
577 	st.qdisc_stats.drop_overmemory = q->drop_overmemory;
578 
579 	sch_tree_lock(sch);
580 	list_for_each(pos, &q->new_flows)
581 		st.qdisc_stats.new_flows_len++;
582 
583 	list_for_each(pos, &q->old_flows)
584 		st.qdisc_stats.old_flows_len++;
585 	sch_tree_unlock(sch);
586 
587 	return gnet_stats_copy_app(d, &st, sizeof(st));
588 }
589 
590 static struct Qdisc *fq_codel_leaf(struct Qdisc *sch, unsigned long arg)
591 {
592 	return NULL;
593 }
594 
595 static unsigned long fq_codel_find(struct Qdisc *sch, u32 classid)
596 {
597 	return 0;
598 }
599 
600 static unsigned long fq_codel_bind(struct Qdisc *sch, unsigned long parent,
601 			      u32 classid)
602 {
603 	/* we cannot bypass queue discipline anymore */
604 	sch->flags &= ~TCQ_F_CAN_BYPASS;
605 	return 0;
606 }
607 
608 static void fq_codel_unbind(struct Qdisc *q, unsigned long cl)
609 {
610 }
611 
612 static struct tcf_block *fq_codel_tcf_block(struct Qdisc *sch, unsigned long cl,
613 					    struct netlink_ext_ack *extack)
614 {
615 	struct fq_codel_sched_data *q = qdisc_priv(sch);
616 
617 	if (cl)
618 		return NULL;
619 	return q->block;
620 }
621 
622 static int fq_codel_dump_class(struct Qdisc *sch, unsigned long cl,
623 			  struct sk_buff *skb, struct tcmsg *tcm)
624 {
625 	tcm->tcm_handle |= TC_H_MIN(cl);
626 	return 0;
627 }
628 
629 static int fq_codel_dump_class_stats(struct Qdisc *sch, unsigned long cl,
630 				     struct gnet_dump *d)
631 {
632 	struct fq_codel_sched_data *q = qdisc_priv(sch);
633 	u32 idx = cl - 1;
634 	struct gnet_stats_queue qs = { 0 };
635 	struct tc_fq_codel_xstats xstats;
636 
637 	if (idx < q->flows_cnt) {
638 		const struct fq_codel_flow *flow = &q->flows[idx];
639 		const struct sk_buff *skb;
640 
641 		memset(&xstats, 0, sizeof(xstats));
642 		xstats.type = TCA_FQ_CODEL_XSTATS_CLASS;
643 		xstats.class_stats.deficit = flow->deficit;
644 		xstats.class_stats.ldelay =
645 			codel_time_to_us(flow->cvars.ldelay);
646 		xstats.class_stats.count = flow->cvars.count;
647 		xstats.class_stats.lastcount = flow->cvars.lastcount;
648 		xstats.class_stats.dropping = flow->cvars.dropping;
649 		if (flow->cvars.dropping) {
650 			codel_tdiff_t delta = flow->cvars.drop_next -
651 					      codel_get_time();
652 
653 			xstats.class_stats.drop_next = (delta >= 0) ?
654 				codel_time_to_us(delta) :
655 				-codel_time_to_us(-delta);
656 		}
657 		if (flow->head) {
658 			sch_tree_lock(sch);
659 			skb = flow->head;
660 			while (skb) {
661 				qs.qlen++;
662 				skb = skb->next;
663 			}
664 			sch_tree_unlock(sch);
665 		}
666 		qs.backlog = q->backlogs[idx];
667 		qs.drops = flow->dropped;
668 	}
669 	if (gnet_stats_copy_queue(d, NULL, &qs, qs.qlen) < 0)
670 		return -1;
671 	if (idx < q->flows_cnt)
672 		return gnet_stats_copy_app(d, &xstats, sizeof(xstats));
673 	return 0;
674 }
675 
676 static void fq_codel_walk(struct Qdisc *sch, struct qdisc_walker *arg)
677 {
678 	struct fq_codel_sched_data *q = qdisc_priv(sch);
679 	unsigned int i;
680 
681 	if (arg->stop)
682 		return;
683 
684 	for (i = 0; i < q->flows_cnt; i++) {
685 		if (list_empty(&q->flows[i].flowchain) ||
686 		    arg->count < arg->skip) {
687 			arg->count++;
688 			continue;
689 		}
690 		if (arg->fn(sch, i + 1, arg) < 0) {
691 			arg->stop = 1;
692 			break;
693 		}
694 		arg->count++;
695 	}
696 }
697 
698 static const struct Qdisc_class_ops fq_codel_class_ops = {
699 	.leaf		=	fq_codel_leaf,
700 	.find		=	fq_codel_find,
701 	.tcf_block	=	fq_codel_tcf_block,
702 	.bind_tcf	=	fq_codel_bind,
703 	.unbind_tcf	=	fq_codel_unbind,
704 	.dump		=	fq_codel_dump_class,
705 	.dump_stats	=	fq_codel_dump_class_stats,
706 	.walk		=	fq_codel_walk,
707 };
708 
709 static struct Qdisc_ops fq_codel_qdisc_ops __read_mostly = {
710 	.cl_ops		=	&fq_codel_class_ops,
711 	.id		=	"fq_codel",
712 	.priv_size	=	sizeof(struct fq_codel_sched_data),
713 	.enqueue	=	fq_codel_enqueue,
714 	.dequeue	=	fq_codel_dequeue,
715 	.peek		=	qdisc_peek_dequeued,
716 	.init		=	fq_codel_init,
717 	.reset		=	fq_codel_reset,
718 	.destroy	=	fq_codel_destroy,
719 	.change		=	fq_codel_change,
720 	.dump		=	fq_codel_dump,
721 	.dump_stats =	fq_codel_dump_stats,
722 	.owner		=	THIS_MODULE,
723 };
724 
725 static int __init fq_codel_module_init(void)
726 {
727 	return register_qdisc(&fq_codel_qdisc_ops);
728 }
729 
730 static void __exit fq_codel_module_exit(void)
731 {
732 	unregister_qdisc(&fq_codel_qdisc_ops);
733 }
734 
735 module_init(fq_codel_module_init)
736 module_exit(fq_codel_module_exit)
737 MODULE_AUTHOR("Eric Dumazet");
738 MODULE_LICENSE("GPL");
739