xref: /linux/net/sched/sch_ets.c (revision da1d9caf95def6f0320819cf941c9fd1069ba9e1)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * net/sched/sch_ets.c         Enhanced Transmission Selection scheduler
4  *
5  * Description
6  * -----------
7  *
8  * The Enhanced Transmission Selection scheduler is a classful queuing
9  * discipline that merges functionality of PRIO and DRR qdiscs in one scheduler.
10  * ETS makes it easy to configure a set of strict and bandwidth-sharing bands to
11  * implement the transmission selection described in 802.1Qaz.
12  *
13  * Although ETS is technically classful, it's not possible to add and remove
14  * classes at will. Instead one specifies number of classes, how many are
15  * PRIO-like and how many DRR-like, and quanta for the latter.
16  *
17  * Algorithm
18  * ---------
19  *
20  * The strict classes, if any, are tried for traffic first: first band 0, if it
21  * has no traffic then band 1, etc.
22  *
23  * When there is no traffic in any of the strict queues, the bandwidth-sharing
24  * ones are tried next. Each band is assigned a deficit counter, initialized to
25  * "quantum" of that band. ETS maintains a list of active bandwidth-sharing
26  * bands whose qdiscs are non-empty. A packet is dequeued from the band at the
27  * head of the list if the packet size is smaller or equal to the deficit
28  * counter. If the counter is too small, it is increased by "quantum" and the
29  * scheduler moves on to the next band in the active list.
30  */
31 
32 #include <linux/module.h>
33 #include <net/gen_stats.h>
34 #include <net/netlink.h>
35 #include <net/pkt_cls.h>
36 #include <net/pkt_sched.h>
37 #include <net/sch_generic.h>
38 
39 struct ets_class {
40 	struct list_head alist; /* In struct ets_sched.active. */
41 	struct Qdisc *qdisc;
42 	u32 quantum;
43 	u32 deficit;
44 	struct gnet_stats_basic_sync bstats;
45 	struct gnet_stats_queue qstats;
46 };
47 
48 struct ets_sched {
49 	struct list_head active;
50 	struct tcf_proto __rcu *filter_list;
51 	struct tcf_block *block;
52 	unsigned int nbands;
53 	unsigned int nstrict;
54 	u8 prio2band[TC_PRIO_MAX + 1];
55 	struct ets_class classes[TCQ_ETS_MAX_BANDS];
56 };
57 
58 static const struct nla_policy ets_policy[TCA_ETS_MAX + 1] = {
59 	[TCA_ETS_NBANDS] = { .type = NLA_U8 },
60 	[TCA_ETS_NSTRICT] = { .type = NLA_U8 },
61 	[TCA_ETS_QUANTA] = { .type = NLA_NESTED },
62 	[TCA_ETS_PRIOMAP] = { .type = NLA_NESTED },
63 };
64 
65 static const struct nla_policy ets_priomap_policy[TCA_ETS_MAX + 1] = {
66 	[TCA_ETS_PRIOMAP_BAND] = { .type = NLA_U8 },
67 };
68 
69 static const struct nla_policy ets_quanta_policy[TCA_ETS_MAX + 1] = {
70 	[TCA_ETS_QUANTA_BAND] = { .type = NLA_U32 },
71 };
72 
73 static const struct nla_policy ets_class_policy[TCA_ETS_MAX + 1] = {
74 	[TCA_ETS_QUANTA_BAND] = { .type = NLA_U32 },
75 };
76 
77 static int ets_quantum_parse(struct Qdisc *sch, const struct nlattr *attr,
78 			     unsigned int *quantum,
79 			     struct netlink_ext_ack *extack)
80 {
81 	*quantum = nla_get_u32(attr);
82 	if (!*quantum) {
83 		NL_SET_ERR_MSG(extack, "ETS quantum cannot be zero");
84 		return -EINVAL;
85 	}
86 	return 0;
87 }
88 
89 static struct ets_class *
90 ets_class_from_arg(struct Qdisc *sch, unsigned long arg)
91 {
92 	struct ets_sched *q = qdisc_priv(sch);
93 
94 	return &q->classes[arg - 1];
95 }
96 
97 static u32 ets_class_id(struct Qdisc *sch, const struct ets_class *cl)
98 {
99 	struct ets_sched *q = qdisc_priv(sch);
100 	int band = cl - q->classes;
101 
102 	return TC_H_MAKE(sch->handle, band + 1);
103 }
104 
105 static void ets_offload_change(struct Qdisc *sch)
106 {
107 	struct net_device *dev = qdisc_dev(sch);
108 	struct ets_sched *q = qdisc_priv(sch);
109 	struct tc_ets_qopt_offload qopt;
110 	unsigned int w_psum_prev = 0;
111 	unsigned int q_psum = 0;
112 	unsigned int q_sum = 0;
113 	unsigned int quantum;
114 	unsigned int w_psum;
115 	unsigned int weight;
116 	unsigned int i;
117 
118 	if (!tc_can_offload(dev) || !dev->netdev_ops->ndo_setup_tc)
119 		return;
120 
121 	qopt.command = TC_ETS_REPLACE;
122 	qopt.handle = sch->handle;
123 	qopt.parent = sch->parent;
124 	qopt.replace_params.bands = q->nbands;
125 	qopt.replace_params.qstats = &sch->qstats;
126 	memcpy(&qopt.replace_params.priomap,
127 	       q->prio2band, sizeof(q->prio2band));
128 
129 	for (i = 0; i < q->nbands; i++)
130 		q_sum += q->classes[i].quantum;
131 
132 	for (i = 0; i < q->nbands; i++) {
133 		quantum = q->classes[i].quantum;
134 		q_psum += quantum;
135 		w_psum = quantum ? q_psum * 100 / q_sum : 0;
136 		weight = w_psum - w_psum_prev;
137 		w_psum_prev = w_psum;
138 
139 		qopt.replace_params.quanta[i] = quantum;
140 		qopt.replace_params.weights[i] = weight;
141 	}
142 
143 	dev->netdev_ops->ndo_setup_tc(dev, TC_SETUP_QDISC_ETS, &qopt);
144 }
145 
146 static void ets_offload_destroy(struct Qdisc *sch)
147 {
148 	struct net_device *dev = qdisc_dev(sch);
149 	struct tc_ets_qopt_offload qopt;
150 
151 	if (!tc_can_offload(dev) || !dev->netdev_ops->ndo_setup_tc)
152 		return;
153 
154 	qopt.command = TC_ETS_DESTROY;
155 	qopt.handle = sch->handle;
156 	qopt.parent = sch->parent;
157 	dev->netdev_ops->ndo_setup_tc(dev, TC_SETUP_QDISC_ETS, &qopt);
158 }
159 
160 static void ets_offload_graft(struct Qdisc *sch, struct Qdisc *new,
161 			      struct Qdisc *old, unsigned long arg,
162 			      struct netlink_ext_ack *extack)
163 {
164 	struct net_device *dev = qdisc_dev(sch);
165 	struct tc_ets_qopt_offload qopt;
166 
167 	qopt.command = TC_ETS_GRAFT;
168 	qopt.handle = sch->handle;
169 	qopt.parent = sch->parent;
170 	qopt.graft_params.band = arg - 1;
171 	qopt.graft_params.child_handle = new->handle;
172 
173 	qdisc_offload_graft_helper(dev, sch, new, old, TC_SETUP_QDISC_ETS,
174 				   &qopt, extack);
175 }
176 
177 static int ets_offload_dump(struct Qdisc *sch)
178 {
179 	struct tc_ets_qopt_offload qopt;
180 
181 	qopt.command = TC_ETS_STATS;
182 	qopt.handle = sch->handle;
183 	qopt.parent = sch->parent;
184 	qopt.stats.bstats = &sch->bstats;
185 	qopt.stats.qstats = &sch->qstats;
186 
187 	return qdisc_offload_dump_helper(sch, TC_SETUP_QDISC_ETS, &qopt);
188 }
189 
190 static bool ets_class_is_strict(struct ets_sched *q, const struct ets_class *cl)
191 {
192 	unsigned int band = cl - q->classes;
193 
194 	return band < q->nstrict;
195 }
196 
197 static int ets_class_change(struct Qdisc *sch, u32 classid, u32 parentid,
198 			    struct nlattr **tca, unsigned long *arg,
199 			    struct netlink_ext_ack *extack)
200 {
201 	struct ets_class *cl = ets_class_from_arg(sch, *arg);
202 	struct ets_sched *q = qdisc_priv(sch);
203 	struct nlattr *opt = tca[TCA_OPTIONS];
204 	struct nlattr *tb[TCA_ETS_MAX + 1];
205 	unsigned int quantum;
206 	int err;
207 
208 	/* Classes can be added and removed only through Qdisc_ops.change
209 	 * interface.
210 	 */
211 	if (!cl) {
212 		NL_SET_ERR_MSG(extack, "Fine-grained class addition and removal is not supported");
213 		return -EOPNOTSUPP;
214 	}
215 
216 	if (!opt) {
217 		NL_SET_ERR_MSG(extack, "ETS options are required for this operation");
218 		return -EINVAL;
219 	}
220 
221 	err = nla_parse_nested(tb, TCA_ETS_MAX, opt, ets_class_policy, extack);
222 	if (err < 0)
223 		return err;
224 
225 	if (!tb[TCA_ETS_QUANTA_BAND])
226 		/* Nothing to configure. */
227 		return 0;
228 
229 	if (ets_class_is_strict(q, cl)) {
230 		NL_SET_ERR_MSG(extack, "Strict bands do not have a configurable quantum");
231 		return -EINVAL;
232 	}
233 
234 	err = ets_quantum_parse(sch, tb[TCA_ETS_QUANTA_BAND], &quantum,
235 				extack);
236 	if (err)
237 		return err;
238 
239 	sch_tree_lock(sch);
240 	cl->quantum = quantum;
241 	sch_tree_unlock(sch);
242 
243 	ets_offload_change(sch);
244 	return 0;
245 }
246 
247 static int ets_class_graft(struct Qdisc *sch, unsigned long arg,
248 			   struct Qdisc *new, struct Qdisc **old,
249 			   struct netlink_ext_ack *extack)
250 {
251 	struct ets_class *cl = ets_class_from_arg(sch, arg);
252 
253 	if (!new) {
254 		new = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
255 					ets_class_id(sch, cl), NULL);
256 		if (!new)
257 			new = &noop_qdisc;
258 		else
259 			qdisc_hash_add(new, true);
260 	}
261 
262 	*old = qdisc_replace(sch, new, &cl->qdisc);
263 	ets_offload_graft(sch, new, *old, arg, extack);
264 	return 0;
265 }
266 
267 static struct Qdisc *ets_class_leaf(struct Qdisc *sch, unsigned long arg)
268 {
269 	struct ets_class *cl = ets_class_from_arg(sch, arg);
270 
271 	return cl->qdisc;
272 }
273 
274 static unsigned long ets_class_find(struct Qdisc *sch, u32 classid)
275 {
276 	unsigned long band = TC_H_MIN(classid);
277 	struct ets_sched *q = qdisc_priv(sch);
278 
279 	if (band - 1 >= q->nbands)
280 		return 0;
281 	return band;
282 }
283 
284 static void ets_class_qlen_notify(struct Qdisc *sch, unsigned long arg)
285 {
286 	struct ets_class *cl = ets_class_from_arg(sch, arg);
287 	struct ets_sched *q = qdisc_priv(sch);
288 
289 	/* We get notified about zero-length child Qdiscs as well if they are
290 	 * offloaded. Those aren't on the active list though, so don't attempt
291 	 * to remove them.
292 	 */
293 	if (!ets_class_is_strict(q, cl) && sch->q.qlen)
294 		list_del(&cl->alist);
295 }
296 
297 static int ets_class_dump(struct Qdisc *sch, unsigned long arg,
298 			  struct sk_buff *skb, struct tcmsg *tcm)
299 {
300 	struct ets_class *cl = ets_class_from_arg(sch, arg);
301 	struct ets_sched *q = qdisc_priv(sch);
302 	struct nlattr *nest;
303 
304 	tcm->tcm_parent = TC_H_ROOT;
305 	tcm->tcm_handle = ets_class_id(sch, cl);
306 	tcm->tcm_info = cl->qdisc->handle;
307 
308 	nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
309 	if (!nest)
310 		goto nla_put_failure;
311 	if (!ets_class_is_strict(q, cl)) {
312 		if (nla_put_u32(skb, TCA_ETS_QUANTA_BAND, cl->quantum))
313 			goto nla_put_failure;
314 	}
315 	return nla_nest_end(skb, nest);
316 
317 nla_put_failure:
318 	nla_nest_cancel(skb, nest);
319 	return -EMSGSIZE;
320 }
321 
322 static int ets_class_dump_stats(struct Qdisc *sch, unsigned long arg,
323 				struct gnet_dump *d)
324 {
325 	struct ets_class *cl = ets_class_from_arg(sch, arg);
326 	struct Qdisc *cl_q = cl->qdisc;
327 
328 	if (gnet_stats_copy_basic(d, NULL, &cl_q->bstats, true) < 0 ||
329 	    qdisc_qstats_copy(d, cl_q) < 0)
330 		return -1;
331 
332 	return 0;
333 }
334 
335 static void ets_qdisc_walk(struct Qdisc *sch, struct qdisc_walker *arg)
336 {
337 	struct ets_sched *q = qdisc_priv(sch);
338 	int i;
339 
340 	if (arg->stop)
341 		return;
342 
343 	for (i = 0; i < q->nbands; i++) {
344 		if (arg->count < arg->skip) {
345 			arg->count++;
346 			continue;
347 		}
348 		if (arg->fn(sch, i + 1, arg) < 0) {
349 			arg->stop = 1;
350 			break;
351 		}
352 		arg->count++;
353 	}
354 }
355 
356 static struct tcf_block *
357 ets_qdisc_tcf_block(struct Qdisc *sch, unsigned long cl,
358 		    struct netlink_ext_ack *extack)
359 {
360 	struct ets_sched *q = qdisc_priv(sch);
361 
362 	if (cl) {
363 		NL_SET_ERR_MSG(extack, "ETS classid must be zero");
364 		return NULL;
365 	}
366 
367 	return q->block;
368 }
369 
370 static unsigned long ets_qdisc_bind_tcf(struct Qdisc *sch, unsigned long parent,
371 					u32 classid)
372 {
373 	return ets_class_find(sch, classid);
374 }
375 
376 static void ets_qdisc_unbind_tcf(struct Qdisc *sch, unsigned long arg)
377 {
378 }
379 
380 static struct ets_class *ets_classify(struct sk_buff *skb, struct Qdisc *sch,
381 				      int *qerr)
382 {
383 	struct ets_sched *q = qdisc_priv(sch);
384 	u32 band = skb->priority;
385 	struct tcf_result res;
386 	struct tcf_proto *fl;
387 	int err;
388 
389 	*qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
390 	if (TC_H_MAJ(skb->priority) != sch->handle) {
391 		fl = rcu_dereference_bh(q->filter_list);
392 		err = tcf_classify(skb, NULL, fl, &res, false);
393 #ifdef CONFIG_NET_CLS_ACT
394 		switch (err) {
395 		case TC_ACT_STOLEN:
396 		case TC_ACT_QUEUED:
397 		case TC_ACT_TRAP:
398 			*qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
399 			fallthrough;
400 		case TC_ACT_SHOT:
401 			return NULL;
402 		}
403 #endif
404 		if (!fl || err < 0) {
405 			if (TC_H_MAJ(band))
406 				band = 0;
407 			return &q->classes[q->prio2band[band & TC_PRIO_MAX]];
408 		}
409 		band = res.classid;
410 	}
411 	band = TC_H_MIN(band) - 1;
412 	if (band >= q->nbands)
413 		return &q->classes[q->prio2band[0]];
414 	return &q->classes[band];
415 }
416 
417 static int ets_qdisc_enqueue(struct sk_buff *skb, struct Qdisc *sch,
418 			     struct sk_buff **to_free)
419 {
420 	unsigned int len = qdisc_pkt_len(skb);
421 	struct ets_sched *q = qdisc_priv(sch);
422 	struct ets_class *cl;
423 	int err = 0;
424 	bool first;
425 
426 	cl = ets_classify(skb, sch, &err);
427 	if (!cl) {
428 		if (err & __NET_XMIT_BYPASS)
429 			qdisc_qstats_drop(sch);
430 		__qdisc_drop(skb, to_free);
431 		return err;
432 	}
433 
434 	first = !cl->qdisc->q.qlen;
435 	err = qdisc_enqueue(skb, cl->qdisc, to_free);
436 	if (unlikely(err != NET_XMIT_SUCCESS)) {
437 		if (net_xmit_drop_count(err)) {
438 			cl->qstats.drops++;
439 			qdisc_qstats_drop(sch);
440 		}
441 		return err;
442 	}
443 
444 	if (first && !ets_class_is_strict(q, cl)) {
445 		list_add_tail(&cl->alist, &q->active);
446 		cl->deficit = cl->quantum;
447 	}
448 
449 	sch->qstats.backlog += len;
450 	sch->q.qlen++;
451 	return err;
452 }
453 
454 static struct sk_buff *
455 ets_qdisc_dequeue_skb(struct Qdisc *sch, struct sk_buff *skb)
456 {
457 	qdisc_bstats_update(sch, skb);
458 	qdisc_qstats_backlog_dec(sch, skb);
459 	sch->q.qlen--;
460 	return skb;
461 }
462 
463 static struct sk_buff *ets_qdisc_dequeue(struct Qdisc *sch)
464 {
465 	struct ets_sched *q = qdisc_priv(sch);
466 	struct ets_class *cl;
467 	struct sk_buff *skb;
468 	unsigned int band;
469 	unsigned int len;
470 
471 	while (1) {
472 		for (band = 0; band < q->nstrict; band++) {
473 			cl = &q->classes[band];
474 			skb = qdisc_dequeue_peeked(cl->qdisc);
475 			if (skb)
476 				return ets_qdisc_dequeue_skb(sch, skb);
477 		}
478 
479 		if (list_empty(&q->active))
480 			goto out;
481 
482 		cl = list_first_entry(&q->active, struct ets_class, alist);
483 		skb = cl->qdisc->ops->peek(cl->qdisc);
484 		if (!skb) {
485 			qdisc_warn_nonwc(__func__, cl->qdisc);
486 			goto out;
487 		}
488 
489 		len = qdisc_pkt_len(skb);
490 		if (len <= cl->deficit) {
491 			cl->deficit -= len;
492 			skb = qdisc_dequeue_peeked(cl->qdisc);
493 			if (unlikely(!skb))
494 				goto out;
495 			if (cl->qdisc->q.qlen == 0)
496 				list_del(&cl->alist);
497 			return ets_qdisc_dequeue_skb(sch, skb);
498 		}
499 
500 		cl->deficit += cl->quantum;
501 		list_move_tail(&cl->alist, &q->active);
502 	}
503 out:
504 	return NULL;
505 }
506 
507 static int ets_qdisc_priomap_parse(struct nlattr *priomap_attr,
508 				   unsigned int nbands, u8 *priomap,
509 				   struct netlink_ext_ack *extack)
510 {
511 	const struct nlattr *attr;
512 	int prio = 0;
513 	u8 band;
514 	int rem;
515 	int err;
516 
517 	err = __nla_validate_nested(priomap_attr, TCA_ETS_MAX,
518 				    ets_priomap_policy, NL_VALIDATE_STRICT,
519 				    extack);
520 	if (err)
521 		return err;
522 
523 	nla_for_each_nested(attr, priomap_attr, rem) {
524 		switch (nla_type(attr)) {
525 		case TCA_ETS_PRIOMAP_BAND:
526 			if (prio > TC_PRIO_MAX) {
527 				NL_SET_ERR_MSG_MOD(extack, "Too many priorities in ETS priomap");
528 				return -EINVAL;
529 			}
530 			band = nla_get_u8(attr);
531 			if (band >= nbands) {
532 				NL_SET_ERR_MSG_MOD(extack, "Invalid band number in ETS priomap");
533 				return -EINVAL;
534 			}
535 			priomap[prio++] = band;
536 			break;
537 		default:
538 			WARN_ON_ONCE(1); /* Validate should have caught this. */
539 			return -EINVAL;
540 		}
541 	}
542 
543 	return 0;
544 }
545 
546 static int ets_qdisc_quanta_parse(struct Qdisc *sch, struct nlattr *quanta_attr,
547 				  unsigned int nbands, unsigned int nstrict,
548 				  unsigned int *quanta,
549 				  struct netlink_ext_ack *extack)
550 {
551 	const struct nlattr *attr;
552 	int band = nstrict;
553 	int rem;
554 	int err;
555 
556 	err = __nla_validate_nested(quanta_attr, TCA_ETS_MAX,
557 				    ets_quanta_policy, NL_VALIDATE_STRICT,
558 				    extack);
559 	if (err < 0)
560 		return err;
561 
562 	nla_for_each_nested(attr, quanta_attr, rem) {
563 		switch (nla_type(attr)) {
564 		case TCA_ETS_QUANTA_BAND:
565 			if (band >= nbands) {
566 				NL_SET_ERR_MSG_MOD(extack, "ETS quanta has more values than bands");
567 				return -EINVAL;
568 			}
569 			err = ets_quantum_parse(sch, attr, &quanta[band++],
570 						extack);
571 			if (err)
572 				return err;
573 			break;
574 		default:
575 			WARN_ON_ONCE(1); /* Validate should have caught this. */
576 			return -EINVAL;
577 		}
578 	}
579 
580 	return 0;
581 }
582 
583 static int ets_qdisc_change(struct Qdisc *sch, struct nlattr *opt,
584 			    struct netlink_ext_ack *extack)
585 {
586 	unsigned int quanta[TCQ_ETS_MAX_BANDS] = {0};
587 	struct Qdisc *queues[TCQ_ETS_MAX_BANDS];
588 	struct ets_sched *q = qdisc_priv(sch);
589 	struct nlattr *tb[TCA_ETS_MAX + 1];
590 	unsigned int oldbands = q->nbands;
591 	u8 priomap[TC_PRIO_MAX + 1];
592 	unsigned int nstrict = 0;
593 	unsigned int nbands;
594 	unsigned int i;
595 	int err;
596 
597 	if (!opt) {
598 		NL_SET_ERR_MSG(extack, "ETS options are required for this operation");
599 		return -EINVAL;
600 	}
601 
602 	err = nla_parse_nested(tb, TCA_ETS_MAX, opt, ets_policy, extack);
603 	if (err < 0)
604 		return err;
605 
606 	if (!tb[TCA_ETS_NBANDS]) {
607 		NL_SET_ERR_MSG_MOD(extack, "Number of bands is a required argument");
608 		return -EINVAL;
609 	}
610 	nbands = nla_get_u8(tb[TCA_ETS_NBANDS]);
611 	if (nbands < 1 || nbands > TCQ_ETS_MAX_BANDS) {
612 		NL_SET_ERR_MSG_MOD(extack, "Invalid number of bands");
613 		return -EINVAL;
614 	}
615 	/* Unless overridden, traffic goes to the last band. */
616 	memset(priomap, nbands - 1, sizeof(priomap));
617 
618 	if (tb[TCA_ETS_NSTRICT]) {
619 		nstrict = nla_get_u8(tb[TCA_ETS_NSTRICT]);
620 		if (nstrict > nbands) {
621 			NL_SET_ERR_MSG_MOD(extack, "Invalid number of strict bands");
622 			return -EINVAL;
623 		}
624 	}
625 
626 	if (tb[TCA_ETS_PRIOMAP]) {
627 		err = ets_qdisc_priomap_parse(tb[TCA_ETS_PRIOMAP],
628 					      nbands, priomap, extack);
629 		if (err)
630 			return err;
631 	}
632 
633 	if (tb[TCA_ETS_QUANTA]) {
634 		err = ets_qdisc_quanta_parse(sch, tb[TCA_ETS_QUANTA],
635 					     nbands, nstrict, quanta, extack);
636 		if (err)
637 			return err;
638 	}
639 	/* If there are more bands than strict + quanta provided, the remaining
640 	 * ones are ETS with quantum of MTU. Initialize the missing values here.
641 	 */
642 	for (i = nstrict; i < nbands; i++) {
643 		if (!quanta[i])
644 			quanta[i] = psched_mtu(qdisc_dev(sch));
645 	}
646 
647 	/* Before commit, make sure we can allocate all new qdiscs */
648 	for (i = oldbands; i < nbands; i++) {
649 		queues[i] = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
650 					      ets_class_id(sch, &q->classes[i]),
651 					      extack);
652 		if (!queues[i]) {
653 			while (i > oldbands)
654 				qdisc_put(queues[--i]);
655 			return -ENOMEM;
656 		}
657 	}
658 
659 	sch_tree_lock(sch);
660 
661 	q->nbands = nbands;
662 	for (i = nstrict; i < q->nstrict; i++) {
663 		if (q->classes[i].qdisc->q.qlen) {
664 			list_add_tail(&q->classes[i].alist, &q->active);
665 			q->classes[i].deficit = quanta[i];
666 		}
667 	}
668 	for (i = q->nbands; i < oldbands; i++) {
669 		if (i >= q->nstrict && q->classes[i].qdisc->q.qlen)
670 			list_del(&q->classes[i].alist);
671 		qdisc_tree_flush_backlog(q->classes[i].qdisc);
672 	}
673 	q->nstrict = nstrict;
674 	memcpy(q->prio2band, priomap, sizeof(priomap));
675 
676 	for (i = 0; i < q->nbands; i++)
677 		q->classes[i].quantum = quanta[i];
678 
679 	for (i = oldbands; i < q->nbands; i++) {
680 		q->classes[i].qdisc = queues[i];
681 		if (q->classes[i].qdisc != &noop_qdisc)
682 			qdisc_hash_add(q->classes[i].qdisc, true);
683 	}
684 
685 	sch_tree_unlock(sch);
686 
687 	ets_offload_change(sch);
688 	for (i = q->nbands; i < oldbands; i++) {
689 		qdisc_put(q->classes[i].qdisc);
690 		q->classes[i].qdisc = NULL;
691 		q->classes[i].quantum = 0;
692 		q->classes[i].deficit = 0;
693 		gnet_stats_basic_sync_init(&q->classes[i].bstats);
694 		memset(&q->classes[i].qstats, 0, sizeof(q->classes[i].qstats));
695 	}
696 	return 0;
697 }
698 
699 static int ets_qdisc_init(struct Qdisc *sch, struct nlattr *opt,
700 			  struct netlink_ext_ack *extack)
701 {
702 	struct ets_sched *q = qdisc_priv(sch);
703 	int err, i;
704 
705 	if (!opt)
706 		return -EINVAL;
707 
708 	err = tcf_block_get(&q->block, &q->filter_list, sch, extack);
709 	if (err)
710 		return err;
711 
712 	INIT_LIST_HEAD(&q->active);
713 	for (i = 0; i < TCQ_ETS_MAX_BANDS; i++)
714 		INIT_LIST_HEAD(&q->classes[i].alist);
715 
716 	return ets_qdisc_change(sch, opt, extack);
717 }
718 
719 static void ets_qdisc_reset(struct Qdisc *sch)
720 {
721 	struct ets_sched *q = qdisc_priv(sch);
722 	int band;
723 
724 	for (band = q->nstrict; band < q->nbands; band++) {
725 		if (q->classes[band].qdisc->q.qlen)
726 			list_del(&q->classes[band].alist);
727 	}
728 	for (band = 0; band < q->nbands; band++)
729 		qdisc_reset(q->classes[band].qdisc);
730 	sch->qstats.backlog = 0;
731 	sch->q.qlen = 0;
732 }
733 
734 static void ets_qdisc_destroy(struct Qdisc *sch)
735 {
736 	struct ets_sched *q = qdisc_priv(sch);
737 	int band;
738 
739 	ets_offload_destroy(sch);
740 	tcf_block_put(q->block);
741 	for (band = 0; band < q->nbands; band++)
742 		qdisc_put(q->classes[band].qdisc);
743 }
744 
745 static int ets_qdisc_dump(struct Qdisc *sch, struct sk_buff *skb)
746 {
747 	struct ets_sched *q = qdisc_priv(sch);
748 	struct nlattr *opts;
749 	struct nlattr *nest;
750 	int band;
751 	int prio;
752 	int err;
753 
754 	err = ets_offload_dump(sch);
755 	if (err)
756 		return err;
757 
758 	opts = nla_nest_start_noflag(skb, TCA_OPTIONS);
759 	if (!opts)
760 		goto nla_err;
761 
762 	if (nla_put_u8(skb, TCA_ETS_NBANDS, q->nbands))
763 		goto nla_err;
764 
765 	if (q->nstrict &&
766 	    nla_put_u8(skb, TCA_ETS_NSTRICT, q->nstrict))
767 		goto nla_err;
768 
769 	if (q->nbands > q->nstrict) {
770 		nest = nla_nest_start(skb, TCA_ETS_QUANTA);
771 		if (!nest)
772 			goto nla_err;
773 
774 		for (band = q->nstrict; band < q->nbands; band++) {
775 			if (nla_put_u32(skb, TCA_ETS_QUANTA_BAND,
776 					q->classes[band].quantum))
777 				goto nla_err;
778 		}
779 
780 		nla_nest_end(skb, nest);
781 	}
782 
783 	nest = nla_nest_start(skb, TCA_ETS_PRIOMAP);
784 	if (!nest)
785 		goto nla_err;
786 
787 	for (prio = 0; prio <= TC_PRIO_MAX; prio++) {
788 		if (nla_put_u8(skb, TCA_ETS_PRIOMAP_BAND, q->prio2band[prio]))
789 			goto nla_err;
790 	}
791 
792 	nla_nest_end(skb, nest);
793 
794 	return nla_nest_end(skb, opts);
795 
796 nla_err:
797 	nla_nest_cancel(skb, opts);
798 	return -EMSGSIZE;
799 }
800 
801 static const struct Qdisc_class_ops ets_class_ops = {
802 	.change		= ets_class_change,
803 	.graft		= ets_class_graft,
804 	.leaf		= ets_class_leaf,
805 	.find		= ets_class_find,
806 	.qlen_notify	= ets_class_qlen_notify,
807 	.dump		= ets_class_dump,
808 	.dump_stats	= ets_class_dump_stats,
809 	.walk		= ets_qdisc_walk,
810 	.tcf_block	= ets_qdisc_tcf_block,
811 	.bind_tcf	= ets_qdisc_bind_tcf,
812 	.unbind_tcf	= ets_qdisc_unbind_tcf,
813 };
814 
815 static struct Qdisc_ops ets_qdisc_ops __read_mostly = {
816 	.cl_ops		= &ets_class_ops,
817 	.id		= "ets",
818 	.priv_size	= sizeof(struct ets_sched),
819 	.enqueue	= ets_qdisc_enqueue,
820 	.dequeue	= ets_qdisc_dequeue,
821 	.peek		= qdisc_peek_dequeued,
822 	.change		= ets_qdisc_change,
823 	.init		= ets_qdisc_init,
824 	.reset		= ets_qdisc_reset,
825 	.destroy	= ets_qdisc_destroy,
826 	.dump		= ets_qdisc_dump,
827 	.owner		= THIS_MODULE,
828 };
829 
830 static int __init ets_init(void)
831 {
832 	return register_qdisc(&ets_qdisc_ops);
833 }
834 
835 static void __exit ets_exit(void)
836 {
837 	unregister_qdisc(&ets_qdisc_ops);
838 }
839 
840 module_init(ets_init);
841 module_exit(ets_exit);
842 MODULE_LICENSE("GPL");
843