xref: /openbmc/linux/net/sched/sch_fq_codel.c (revision 7bcae826)
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 fq_codel_flow *flows;	/* Flows table [flows_cnt] */
59 	u32		*backlogs;	/* backlog table [flows_cnt] */
60 	u32		flows_cnt;	/* number of flows */
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 	return reciprocal_scale(skb_get_hash(skb), q->flows_cnt);
79 }
80 
81 static unsigned int fq_codel_classify(struct sk_buff *skb, struct Qdisc *sch,
82 				      int *qerr)
83 {
84 	struct fq_codel_sched_data *q = qdisc_priv(sch);
85 	struct tcf_proto *filter;
86 	struct tcf_result res;
87 	int result;
88 
89 	if (TC_H_MAJ(skb->priority) == sch->handle &&
90 	    TC_H_MIN(skb->priority) > 0 &&
91 	    TC_H_MIN(skb->priority) <= q->flows_cnt)
92 		return TC_H_MIN(skb->priority);
93 
94 	filter = rcu_dereference_bh(q->filter_list);
95 	if (!filter)
96 		return fq_codel_hash(q, skb) + 1;
97 
98 	*qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
99 	result = tc_classify(skb, filter, &res, false);
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 */
119 static 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) */
129 static 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 
140 static unsigned int fq_codel_drop(struct Qdisc *sch, unsigned int max_packets,
141 				  struct sk_buff **to_free)
142 {
143 	struct fq_codel_sched_data *q = qdisc_priv(sch);
144 	struct sk_buff *skb;
145 	unsigned int maxbacklog = 0, idx = 0, i, len;
146 	struct fq_codel_flow *flow;
147 	unsigned int threshold;
148 	unsigned int mem = 0;
149 
150 	/* Queue is full! Find the fat flow and drop packet(s) from it.
151 	 * This might sound expensive, but with 1024 flows, we scan
152 	 * 4KB of memory, and we dont need to handle a complex tree
153 	 * in fast path (packet queue/enqueue) with many cache misses.
154 	 * In stress mode, we'll try to drop 64 packets from the flow,
155 	 * amortizing this linear lookup to one cache line per drop.
156 	 */
157 	for (i = 0; i < q->flows_cnt; i++) {
158 		if (q->backlogs[i] > maxbacklog) {
159 			maxbacklog = q->backlogs[i];
160 			idx = i;
161 		}
162 	}
163 
164 	/* Our goal is to drop half of this fat flow backlog */
165 	threshold = maxbacklog >> 1;
166 
167 	flow = &q->flows[idx];
168 	len = 0;
169 	i = 0;
170 	do {
171 		skb = dequeue_head(flow);
172 		len += qdisc_pkt_len(skb);
173 		mem += get_codel_cb(skb)->mem_usage;
174 		__qdisc_drop(skb, to_free);
175 	} while (++i < max_packets && len < threshold);
176 
177 	flow->dropped += i;
178 	q->backlogs[idx] -= len;
179 	q->memory_usage -= mem;
180 	sch->qstats.drops += i;
181 	sch->qstats.backlog -= len;
182 	sch->q.qlen -= i;
183 	return idx;
184 }
185 
186 static int fq_codel_enqueue(struct sk_buff *skb, struct Qdisc *sch,
187 			    struct sk_buff **to_free)
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 		__qdisc_drop(skb, to_free);
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 	get_codel_cb(skb)->mem_usage = skb->truesize;
218 	q->memory_usage += get_codel_cb(skb)->mem_usage;
219 	memory_limited = q->memory_usage > q->memory_limit;
220 	if (++sch->q.qlen <= sch->limit && !memory_limited)
221 		return NET_XMIT_SUCCESS;
222 
223 	prev_backlog = sch->qstats.backlog;
224 	prev_qlen = sch->q.qlen;
225 
226 	/* save this packet length as it might be dropped by fq_codel_drop() */
227 	pkt_len = qdisc_pkt_len(skb);
228 	/* fq_codel_drop() is quite expensive, as it performs a linear search
229 	 * in q->backlogs[] to find a fat flow.
230 	 * So instead of dropping a single packet, drop half of its backlog
231 	 * with a 64 packets limit to not add a too big cpu spike here.
232 	 */
233 	ret = fq_codel_drop(sch, q->drop_batch_size, to_free);
234 
235 	prev_qlen -= sch->q.qlen;
236 	prev_backlog -= sch->qstats.backlog;
237 	q->drop_overlimit += prev_qlen;
238 	if (memory_limited)
239 		q->drop_overmemory += prev_qlen;
240 
241 	/* As we dropped packet(s), better let upper stack know this.
242 	 * If we dropped a packet for this flow, return NET_XMIT_CN,
243 	 * but in this case, our parents wont increase their backlogs.
244 	 */
245 	if (ret == idx) {
246 		qdisc_tree_reduce_backlog(sch, prev_qlen - 1,
247 					  prev_backlog - pkt_len);
248 		return NET_XMIT_CN;
249 	}
250 	qdisc_tree_reduce_backlog(sch, prev_qlen, prev_backlog);
251 	return NET_XMIT_SUCCESS;
252 }
253 
254 /* This is the specific function called from codel_dequeue()
255  * to dequeue a packet from queue. Note: backlog is handled in
256  * codel, we dont need to reduce it here.
257  */
258 static struct sk_buff *dequeue_func(struct codel_vars *vars, void *ctx)
259 {
260 	struct Qdisc *sch = ctx;
261 	struct fq_codel_sched_data *q = qdisc_priv(sch);
262 	struct fq_codel_flow *flow;
263 	struct sk_buff *skb = NULL;
264 
265 	flow = container_of(vars, struct fq_codel_flow, cvars);
266 	if (flow->head) {
267 		skb = dequeue_head(flow);
268 		q->backlogs[flow - q->flows] -= qdisc_pkt_len(skb);
269 		q->memory_usage -= get_codel_cb(skb)->mem_usage;
270 		sch->q.qlen--;
271 		sch->qstats.backlog -= qdisc_pkt_len(skb);
272 	}
273 	return skb;
274 }
275 
276 static void drop_func(struct sk_buff *skb, void *ctx)
277 {
278 	struct Qdisc *sch = ctx;
279 
280 	kfree_skb(skb);
281 	qdisc_qstats_drop(sch);
282 }
283 
284 static struct sk_buff *fq_codel_dequeue(struct Qdisc *sch)
285 {
286 	struct fq_codel_sched_data *q = qdisc_priv(sch);
287 	struct sk_buff *skb;
288 	struct fq_codel_flow *flow;
289 	struct list_head *head;
290 	u32 prev_drop_count, prev_ecn_mark;
291 	unsigned int prev_backlog;
292 
293 begin:
294 	head = &q->new_flows;
295 	if (list_empty(head)) {
296 		head = &q->old_flows;
297 		if (list_empty(head))
298 			return NULL;
299 	}
300 	flow = list_first_entry(head, struct fq_codel_flow, flowchain);
301 
302 	if (flow->deficit <= 0) {
303 		flow->deficit += q->quantum;
304 		list_move_tail(&flow->flowchain, &q->old_flows);
305 		goto begin;
306 	}
307 
308 	prev_drop_count = q->cstats.drop_count;
309 	prev_ecn_mark = q->cstats.ecn_mark;
310 	prev_backlog = sch->qstats.backlog;
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 	if (err < 0)
390 		return err;
391 	if (tb[TCA_FQ_CODEL_FLOWS]) {
392 		if (q->flows)
393 			return -EINVAL;
394 		q->flows_cnt = nla_get_u32(tb[TCA_FQ_CODEL_FLOWS]);
395 		if (!q->flows_cnt ||
396 		    q->flows_cnt > 65536)
397 			return -EINVAL;
398 	}
399 	sch_tree_lock(sch);
400 
401 	if (tb[TCA_FQ_CODEL_TARGET]) {
402 		u64 target = nla_get_u32(tb[TCA_FQ_CODEL_TARGET]);
403 
404 		q->cparams.target = (target * NSEC_PER_USEC) >> CODEL_SHIFT;
405 	}
406 
407 	if (tb[TCA_FQ_CODEL_CE_THRESHOLD]) {
408 		u64 val = nla_get_u32(tb[TCA_FQ_CODEL_CE_THRESHOLD]);
409 
410 		q->cparams.ce_threshold = (val * NSEC_PER_USEC) >> CODEL_SHIFT;
411 	}
412 
413 	if (tb[TCA_FQ_CODEL_INTERVAL]) {
414 		u64 interval = nla_get_u32(tb[TCA_FQ_CODEL_INTERVAL]);
415 
416 		q->cparams.interval = (interval * NSEC_PER_USEC) >> CODEL_SHIFT;
417 	}
418 
419 	if (tb[TCA_FQ_CODEL_LIMIT])
420 		sch->limit = nla_get_u32(tb[TCA_FQ_CODEL_LIMIT]);
421 
422 	if (tb[TCA_FQ_CODEL_ECN])
423 		q->cparams.ecn = !!nla_get_u32(tb[TCA_FQ_CODEL_ECN]);
424 
425 	if (tb[TCA_FQ_CODEL_QUANTUM])
426 		q->quantum = max(256U, nla_get_u32(tb[TCA_FQ_CODEL_QUANTUM]));
427 
428 	if (tb[TCA_FQ_CODEL_DROP_BATCH_SIZE])
429 		q->drop_batch_size = min(1U, nla_get_u32(tb[TCA_FQ_CODEL_DROP_BATCH_SIZE]));
430 
431 	if (tb[TCA_FQ_CODEL_MEMORY_LIMIT])
432 		q->memory_limit = min(1U << 31, nla_get_u32(tb[TCA_FQ_CODEL_MEMORY_LIMIT]));
433 
434 	while (sch->q.qlen > sch->limit ||
435 	       q->memory_usage > q->memory_limit) {
436 		struct sk_buff *skb = fq_codel_dequeue(sch);
437 
438 		q->cstats.drop_len += qdisc_pkt_len(skb);
439 		rtnl_kfree_skbs(skb, skb);
440 		q->cstats.drop_count++;
441 	}
442 	qdisc_tree_reduce_backlog(sch, q->cstats.drop_count, q->cstats.drop_len);
443 	q->cstats.drop_count = 0;
444 	q->cstats.drop_len = 0;
445 
446 	sch_tree_unlock(sch);
447 	return 0;
448 }
449 
450 static void *fq_codel_zalloc(size_t sz)
451 {
452 	void *ptr = kzalloc(sz, GFP_KERNEL | __GFP_NOWARN);
453 
454 	if (!ptr)
455 		ptr = vzalloc(sz);
456 	return ptr;
457 }
458 
459 static void fq_codel_free(void *addr)
460 {
461 	kvfree(addr);
462 }
463 
464 static void fq_codel_destroy(struct Qdisc *sch)
465 {
466 	struct fq_codel_sched_data *q = qdisc_priv(sch);
467 
468 	tcf_destroy_chain(&q->filter_list);
469 	fq_codel_free(q->backlogs);
470 	fq_codel_free(q->flows);
471 }
472 
473 static int fq_codel_init(struct Qdisc *sch, struct nlattr *opt)
474 {
475 	struct fq_codel_sched_data *q = qdisc_priv(sch);
476 	int i;
477 
478 	sch->limit = 10*1024;
479 	q->flows_cnt = 1024;
480 	q->memory_limit = 32 << 20; /* 32 MBytes */
481 	q->drop_batch_size = 64;
482 	q->quantum = psched_mtu(qdisc_dev(sch));
483 	INIT_LIST_HEAD(&q->new_flows);
484 	INIT_LIST_HEAD(&q->old_flows);
485 	codel_params_init(&q->cparams);
486 	codel_stats_init(&q->cstats);
487 	q->cparams.ecn = true;
488 	q->cparams.mtu = psched_mtu(qdisc_dev(sch));
489 
490 	if (opt) {
491 		int err = fq_codel_change(sch, opt);
492 		if (err)
493 			return err;
494 	}
495 
496 	if (!q->flows) {
497 		q->flows = fq_codel_zalloc(q->flows_cnt *
498 					   sizeof(struct fq_codel_flow));
499 		if (!q->flows)
500 			return -ENOMEM;
501 		q->backlogs = fq_codel_zalloc(q->flows_cnt * sizeof(u32));
502 		if (!q->backlogs) {
503 			fq_codel_free(q->flows);
504 			return -ENOMEM;
505 		}
506 		for (i = 0; i < q->flows_cnt; i++) {
507 			struct fq_codel_flow *flow = q->flows + i;
508 
509 			INIT_LIST_HEAD(&flow->flowchain);
510 			codel_vars_init(&flow->cvars);
511 		}
512 	}
513 	if (sch->limit >= 1)
514 		sch->flags |= TCQ_F_CAN_BYPASS;
515 	else
516 		sch->flags &= ~TCQ_F_CAN_BYPASS;
517 	return 0;
518 }
519 
520 static 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(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 
554 nla_put_failure:
555 	return -1;
556 }
557 
558 static 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 
585 static struct Qdisc *fq_codel_leaf(struct Qdisc *sch, unsigned long arg)
586 {
587 	return NULL;
588 }
589 
590 static unsigned long fq_codel_get(struct Qdisc *sch, u32 classid)
591 {
592 	return 0;
593 }
594 
595 static unsigned long fq_codel_bind(struct Qdisc *sch, unsigned long parent,
596 			      u32 classid)
597 {
598 	/* we cannot bypass queue discipline anymore */
599 	sch->flags &= ~TCQ_F_CAN_BYPASS;
600 	return 0;
601 }
602 
603 static void fq_codel_put(struct Qdisc *q, unsigned long cl)
604 {
605 }
606 
607 static struct tcf_proto __rcu **fq_codel_find_tcf(struct Qdisc *sch,
608 						  unsigned long cl)
609 {
610 	struct fq_codel_sched_data *q = qdisc_priv(sch);
611 
612 	if (cl)
613 		return NULL;
614 	return &q->filter_list;
615 }
616 
617 static int fq_codel_dump_class(struct Qdisc *sch, unsigned long cl,
618 			  struct sk_buff *skb, struct tcmsg *tcm)
619 {
620 	tcm->tcm_handle |= TC_H_MIN(cl);
621 	return 0;
622 }
623 
624 static int fq_codel_dump_class_stats(struct Qdisc *sch, unsigned long cl,
625 				     struct gnet_dump *d)
626 {
627 	struct fq_codel_sched_data *q = qdisc_priv(sch);
628 	u32 idx = cl - 1;
629 	struct gnet_stats_queue qs = { 0 };
630 	struct tc_fq_codel_xstats xstats;
631 
632 	if (idx < q->flows_cnt) {
633 		const struct fq_codel_flow *flow = &q->flows[idx];
634 		const struct sk_buff *skb;
635 
636 		memset(&xstats, 0, sizeof(xstats));
637 		xstats.type = TCA_FQ_CODEL_XSTATS_CLASS;
638 		xstats.class_stats.deficit = flow->deficit;
639 		xstats.class_stats.ldelay =
640 			codel_time_to_us(flow->cvars.ldelay);
641 		xstats.class_stats.count = flow->cvars.count;
642 		xstats.class_stats.lastcount = flow->cvars.lastcount;
643 		xstats.class_stats.dropping = flow->cvars.dropping;
644 		if (flow->cvars.dropping) {
645 			codel_tdiff_t delta = flow->cvars.drop_next -
646 					      codel_get_time();
647 
648 			xstats.class_stats.drop_next = (delta >= 0) ?
649 				codel_time_to_us(delta) :
650 				-codel_time_to_us(-delta);
651 		}
652 		if (flow->head) {
653 			sch_tree_lock(sch);
654 			skb = flow->head;
655 			while (skb) {
656 				qs.qlen++;
657 				skb = skb->next;
658 			}
659 			sch_tree_unlock(sch);
660 		}
661 		qs.backlog = q->backlogs[idx];
662 		qs.drops = flow->dropped;
663 	}
664 	if (gnet_stats_copy_queue(d, NULL, &qs, qs.qlen) < 0)
665 		return -1;
666 	if (idx < q->flows_cnt)
667 		return gnet_stats_copy_app(d, &xstats, sizeof(xstats));
668 	return 0;
669 }
670 
671 static void fq_codel_walk(struct Qdisc *sch, struct qdisc_walker *arg)
672 {
673 	struct fq_codel_sched_data *q = qdisc_priv(sch);
674 	unsigned int i;
675 
676 	if (arg->stop)
677 		return;
678 
679 	for (i = 0; i < q->flows_cnt; i++) {
680 		if (list_empty(&q->flows[i].flowchain) ||
681 		    arg->count < arg->skip) {
682 			arg->count++;
683 			continue;
684 		}
685 		if (arg->fn(sch, i + 1, arg) < 0) {
686 			arg->stop = 1;
687 			break;
688 		}
689 		arg->count++;
690 	}
691 }
692 
693 static const struct Qdisc_class_ops fq_codel_class_ops = {
694 	.leaf		=	fq_codel_leaf,
695 	.get		=	fq_codel_get,
696 	.put		=	fq_codel_put,
697 	.tcf_chain	=	fq_codel_find_tcf,
698 	.bind_tcf	=	fq_codel_bind,
699 	.unbind_tcf	=	fq_codel_put,
700 	.dump		=	fq_codel_dump_class,
701 	.dump_stats	=	fq_codel_dump_class_stats,
702 	.walk		=	fq_codel_walk,
703 };
704 
705 static struct Qdisc_ops fq_codel_qdisc_ops __read_mostly = {
706 	.cl_ops		=	&fq_codel_class_ops,
707 	.id		=	"fq_codel",
708 	.priv_size	=	sizeof(struct fq_codel_sched_data),
709 	.enqueue	=	fq_codel_enqueue,
710 	.dequeue	=	fq_codel_dequeue,
711 	.peek		=	qdisc_peek_dequeued,
712 	.init		=	fq_codel_init,
713 	.reset		=	fq_codel_reset,
714 	.destroy	=	fq_codel_destroy,
715 	.change		=	fq_codel_change,
716 	.dump		=	fq_codel_dump,
717 	.dump_stats =	fq_codel_dump_stats,
718 	.owner		=	THIS_MODULE,
719 };
720 
721 static int __init fq_codel_module_init(void)
722 {
723 	return register_qdisc(&fq_codel_qdisc_ops);
724 }
725 
726 static void __exit fq_codel_module_exit(void)
727 {
728 	unregister_qdisc(&fq_codel_qdisc_ops);
729 }
730 
731 module_init(fq_codel_module_init)
732 module_exit(fq_codel_module_exit)
733 MODULE_AUTHOR("Eric Dumazet");
734 MODULE_LICENSE("GPL");
735