xref: /openbmc/linux/net/sched/sch_mqprio.c (revision c54876cd5961ce0f8e74807f79a6739cd6b35ddf)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * net/sched/sch_mqprio.c
4  *
5  * Copyright (c) 2010 John Fastabend <john.r.fastabend@intel.com>
6  */
7 
8 #include <linux/types.h>
9 #include <linux/slab.h>
10 #include <linux/kernel.h>
11 #include <linux/string.h>
12 #include <linux/errno.h>
13 #include <linux/skbuff.h>
14 #include <linux/module.h>
15 #include <net/netlink.h>
16 #include <net/pkt_sched.h>
17 #include <net/sch_generic.h>
18 #include <net/pkt_cls.h>
19 
20 #include "sch_mqprio_lib.h"
21 
22 struct mqprio_sched {
23 	struct Qdisc		**qdiscs;
24 	u16 mode;
25 	u16 shaper;
26 	int hw_offload;
27 	u32 flags;
28 	u64 min_rate[TC_QOPT_MAX_QUEUE];
29 	u64 max_rate[TC_QOPT_MAX_QUEUE];
30 };
31 
32 static int mqprio_enable_offload(struct Qdisc *sch,
33 				 const struct tc_mqprio_qopt *qopt,
34 				 struct netlink_ext_ack *extack)
35 {
36 	struct mqprio_sched *priv = qdisc_priv(sch);
37 	struct net_device *dev = qdisc_dev(sch);
38 	struct tc_mqprio_qopt_offload mqprio = {
39 		.qopt = *qopt,
40 		.extack = extack,
41 	};
42 	int err, i;
43 
44 	switch (priv->mode) {
45 	case TC_MQPRIO_MODE_DCB:
46 		if (priv->shaper != TC_MQPRIO_SHAPER_DCB)
47 			return -EINVAL;
48 		break;
49 	case TC_MQPRIO_MODE_CHANNEL:
50 		mqprio.flags = priv->flags;
51 		if (priv->flags & TC_MQPRIO_F_MODE)
52 			mqprio.mode = priv->mode;
53 		if (priv->flags & TC_MQPRIO_F_SHAPER)
54 			mqprio.shaper = priv->shaper;
55 		if (priv->flags & TC_MQPRIO_F_MIN_RATE)
56 			for (i = 0; i < mqprio.qopt.num_tc; i++)
57 				mqprio.min_rate[i] = priv->min_rate[i];
58 		if (priv->flags & TC_MQPRIO_F_MAX_RATE)
59 			for (i = 0; i < mqprio.qopt.num_tc; i++)
60 				mqprio.max_rate[i] = priv->max_rate[i];
61 		break;
62 	default:
63 		return -EINVAL;
64 	}
65 
66 	err = dev->netdev_ops->ndo_setup_tc(dev, TC_SETUP_QDISC_MQPRIO,
67 					    &mqprio);
68 	if (err)
69 		return err;
70 
71 	priv->hw_offload = mqprio.qopt.hw;
72 
73 	return 0;
74 }
75 
76 static void mqprio_disable_offload(struct Qdisc *sch)
77 {
78 	struct tc_mqprio_qopt_offload mqprio = { { 0 } };
79 	struct mqprio_sched *priv = qdisc_priv(sch);
80 	struct net_device *dev = qdisc_dev(sch);
81 
82 	switch (priv->mode) {
83 	case TC_MQPRIO_MODE_DCB:
84 	case TC_MQPRIO_MODE_CHANNEL:
85 		dev->netdev_ops->ndo_setup_tc(dev, TC_SETUP_QDISC_MQPRIO,
86 					      &mqprio);
87 		break;
88 	}
89 }
90 
91 static void mqprio_destroy(struct Qdisc *sch)
92 {
93 	struct net_device *dev = qdisc_dev(sch);
94 	struct mqprio_sched *priv = qdisc_priv(sch);
95 	unsigned int ntx;
96 
97 	if (priv->qdiscs) {
98 		for (ntx = 0;
99 		     ntx < dev->num_tx_queues && priv->qdiscs[ntx];
100 		     ntx++)
101 			qdisc_put(priv->qdiscs[ntx]);
102 		kfree(priv->qdiscs);
103 	}
104 
105 	if (priv->hw_offload && dev->netdev_ops->ndo_setup_tc)
106 		mqprio_disable_offload(sch);
107 	else
108 		netdev_set_num_tc(dev, 0);
109 }
110 
111 static int mqprio_parse_opt(struct net_device *dev, struct tc_mqprio_qopt *qopt,
112 			    const struct tc_mqprio_caps *caps,
113 			    struct netlink_ext_ack *extack)
114 {
115 	int err;
116 
117 	/* Limit qopt->hw to maximum supported offload value.  Drivers have
118 	 * the option of overriding this later if they don't support the a
119 	 * given offload type.
120 	 */
121 	if (qopt->hw > TC_MQPRIO_HW_OFFLOAD_MAX)
122 		qopt->hw = TC_MQPRIO_HW_OFFLOAD_MAX;
123 
124 	/* If hardware offload is requested, we will leave 3 options to the
125 	 * device driver:
126 	 * - populate the queue counts itself (and ignore what was requested)
127 	 * - validate the provided queue counts by itself (and apply them)
128 	 * - request queue count validation here (and apply them)
129 	 */
130 	err = mqprio_validate_qopt(dev, qopt,
131 				   !qopt->hw || caps->validate_queue_counts,
132 				   false, extack);
133 	if (err)
134 		return err;
135 
136 	/* If ndo_setup_tc is not present then hardware doesn't support offload
137 	 * and we should return an error.
138 	 */
139 	if (qopt->hw && !dev->netdev_ops->ndo_setup_tc) {
140 		NL_SET_ERR_MSG(extack,
141 			       "Device does not support hardware offload");
142 		return -EINVAL;
143 	}
144 
145 	return 0;
146 }
147 
148 static const struct nla_policy mqprio_policy[TCA_MQPRIO_MAX + 1] = {
149 	[TCA_MQPRIO_MODE]	= { .len = sizeof(u16) },
150 	[TCA_MQPRIO_SHAPER]	= { .len = sizeof(u16) },
151 	[TCA_MQPRIO_MIN_RATE64]	= { .type = NLA_NESTED },
152 	[TCA_MQPRIO_MAX_RATE64]	= { .type = NLA_NESTED },
153 };
154 
155 /* Parse the other netlink attributes that represent the payload of
156  * TCA_OPTIONS, which are appended right after struct tc_mqprio_qopt.
157  */
158 static int mqprio_parse_nlattr(struct Qdisc *sch, struct tc_mqprio_qopt *qopt,
159 			       struct nlattr *opt,
160 			       struct netlink_ext_ack *extack)
161 {
162 	struct nlattr *nlattr_opt = nla_data(opt) + NLA_ALIGN(sizeof(*qopt));
163 	int nlattr_opt_len = nla_len(opt) - NLA_ALIGN(sizeof(*qopt));
164 	struct mqprio_sched *priv = qdisc_priv(sch);
165 	struct nlattr *tb[TCA_MQPRIO_MAX + 1] = {};
166 	struct nlattr *attr;
167 	int i, rem, err;
168 
169 	if (nlattr_opt_len >= nla_attr_size(0)) {
170 		err = nla_parse_deprecated(tb, TCA_MQPRIO_MAX, nlattr_opt,
171 					   nlattr_opt_len, mqprio_policy,
172 					   NULL);
173 		if (err < 0)
174 			return err;
175 	}
176 
177 	if (!qopt->hw) {
178 		NL_SET_ERR_MSG(extack,
179 			       "mqprio TCA_OPTIONS can only contain netlink attributes in hardware mode");
180 		return -EINVAL;
181 	}
182 
183 	if (tb[TCA_MQPRIO_MODE]) {
184 		priv->flags |= TC_MQPRIO_F_MODE;
185 		priv->mode = nla_get_u16(tb[TCA_MQPRIO_MODE]);
186 	}
187 
188 	if (tb[TCA_MQPRIO_SHAPER]) {
189 		priv->flags |= TC_MQPRIO_F_SHAPER;
190 		priv->shaper = nla_get_u16(tb[TCA_MQPRIO_SHAPER]);
191 	}
192 
193 	if (tb[TCA_MQPRIO_MIN_RATE64]) {
194 		if (priv->shaper != TC_MQPRIO_SHAPER_BW_RATE) {
195 			NL_SET_ERR_MSG_ATTR(extack, tb[TCA_MQPRIO_MIN_RATE64],
196 					    "min_rate accepted only when shaper is in bw_rlimit mode");
197 			return -EINVAL;
198 		}
199 		i = 0;
200 		nla_for_each_nested(attr, tb[TCA_MQPRIO_MIN_RATE64],
201 				    rem) {
202 			if (nla_type(attr) != TCA_MQPRIO_MIN_RATE64) {
203 				NL_SET_ERR_MSG_ATTR(extack, attr,
204 						    "Attribute type expected to be TCA_MQPRIO_MIN_RATE64");
205 				return -EINVAL;
206 			}
207 			if (i >= qopt->num_tc)
208 				break;
209 			priv->min_rate[i] = nla_get_u64(attr);
210 			i++;
211 		}
212 		priv->flags |= TC_MQPRIO_F_MIN_RATE;
213 	}
214 
215 	if (tb[TCA_MQPRIO_MAX_RATE64]) {
216 		if (priv->shaper != TC_MQPRIO_SHAPER_BW_RATE) {
217 			NL_SET_ERR_MSG_ATTR(extack, tb[TCA_MQPRIO_MAX_RATE64],
218 					    "max_rate accepted only when shaper is in bw_rlimit mode");
219 			return -EINVAL;
220 		}
221 		i = 0;
222 		nla_for_each_nested(attr, tb[TCA_MQPRIO_MAX_RATE64],
223 				    rem) {
224 			if (nla_type(attr) != TCA_MQPRIO_MAX_RATE64) {
225 				NL_SET_ERR_MSG_ATTR(extack, attr,
226 						    "Attribute type expected to be TCA_MQPRIO_MAX_RATE64");
227 				return -EINVAL;
228 			}
229 			if (i >= qopt->num_tc)
230 				break;
231 			priv->max_rate[i] = nla_get_u64(attr);
232 			i++;
233 		}
234 		priv->flags |= TC_MQPRIO_F_MAX_RATE;
235 	}
236 
237 	return 0;
238 }
239 
240 static int mqprio_init(struct Qdisc *sch, struct nlattr *opt,
241 		       struct netlink_ext_ack *extack)
242 {
243 	struct net_device *dev = qdisc_dev(sch);
244 	struct mqprio_sched *priv = qdisc_priv(sch);
245 	struct netdev_queue *dev_queue;
246 	struct Qdisc *qdisc;
247 	int i, err = -EOPNOTSUPP;
248 	struct tc_mqprio_qopt *qopt = NULL;
249 	struct tc_mqprio_caps caps;
250 	int len;
251 
252 	BUILD_BUG_ON(TC_MAX_QUEUE != TC_QOPT_MAX_QUEUE);
253 	BUILD_BUG_ON(TC_BITMASK != TC_QOPT_BITMASK);
254 
255 	if (sch->parent != TC_H_ROOT)
256 		return -EOPNOTSUPP;
257 
258 	if (!netif_is_multiqueue(dev))
259 		return -EOPNOTSUPP;
260 
261 	/* make certain can allocate enough classids to handle queues */
262 	if (dev->num_tx_queues >= TC_H_MIN_PRIORITY)
263 		return -ENOMEM;
264 
265 	if (!opt || nla_len(opt) < sizeof(*qopt))
266 		return -EINVAL;
267 
268 	qdisc_offload_query_caps(dev, TC_SETUP_QDISC_MQPRIO,
269 				 &caps, sizeof(caps));
270 
271 	qopt = nla_data(opt);
272 	if (mqprio_parse_opt(dev, qopt, &caps, extack))
273 		return -EINVAL;
274 
275 	len = nla_len(opt) - NLA_ALIGN(sizeof(*qopt));
276 	if (len > 0) {
277 		err = mqprio_parse_nlattr(sch, qopt, opt, extack);
278 		if (err)
279 			return err;
280 	}
281 
282 	/* pre-allocate qdisc, attachment can't fail */
283 	priv->qdiscs = kcalloc(dev->num_tx_queues, sizeof(priv->qdiscs[0]),
284 			       GFP_KERNEL);
285 	if (!priv->qdiscs)
286 		return -ENOMEM;
287 
288 	for (i = 0; i < dev->num_tx_queues; i++) {
289 		dev_queue = netdev_get_tx_queue(dev, i);
290 		qdisc = qdisc_create_dflt(dev_queue,
291 					  get_default_qdisc_ops(dev, i),
292 					  TC_H_MAKE(TC_H_MAJ(sch->handle),
293 						    TC_H_MIN(i + 1)), extack);
294 		if (!qdisc)
295 			return -ENOMEM;
296 
297 		priv->qdiscs[i] = qdisc;
298 		qdisc->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
299 	}
300 
301 	/* If the mqprio options indicate that hardware should own
302 	 * the queue mapping then run ndo_setup_tc otherwise use the
303 	 * supplied and verified mapping
304 	 */
305 	if (qopt->hw) {
306 		err = mqprio_enable_offload(sch, qopt, extack);
307 		if (err)
308 			return err;
309 	} else {
310 		netdev_set_num_tc(dev, qopt->num_tc);
311 		for (i = 0; i < qopt->num_tc; i++)
312 			netdev_set_tc_queue(dev, i,
313 					    qopt->count[i], qopt->offset[i]);
314 	}
315 
316 	/* Always use supplied priority mappings */
317 	for (i = 0; i < TC_BITMASK + 1; i++)
318 		netdev_set_prio_tc_map(dev, i, qopt->prio_tc_map[i]);
319 
320 	sch->flags |= TCQ_F_MQROOT;
321 	return 0;
322 }
323 
324 static void mqprio_attach(struct Qdisc *sch)
325 {
326 	struct net_device *dev = qdisc_dev(sch);
327 	struct mqprio_sched *priv = qdisc_priv(sch);
328 	struct Qdisc *qdisc, *old;
329 	unsigned int ntx;
330 
331 	/* Attach underlying qdisc */
332 	for (ntx = 0; ntx < dev->num_tx_queues; ntx++) {
333 		qdisc = priv->qdiscs[ntx];
334 		old = dev_graft_qdisc(qdisc->dev_queue, qdisc);
335 		if (old)
336 			qdisc_put(old);
337 		if (ntx < dev->real_num_tx_queues)
338 			qdisc_hash_add(qdisc, false);
339 	}
340 	kfree(priv->qdiscs);
341 	priv->qdiscs = NULL;
342 }
343 
344 static struct netdev_queue *mqprio_queue_get(struct Qdisc *sch,
345 					     unsigned long cl)
346 {
347 	struct net_device *dev = qdisc_dev(sch);
348 	unsigned long ntx = cl - 1;
349 
350 	if (ntx >= dev->num_tx_queues)
351 		return NULL;
352 	return netdev_get_tx_queue(dev, ntx);
353 }
354 
355 static int mqprio_graft(struct Qdisc *sch, unsigned long cl, struct Qdisc *new,
356 			struct Qdisc **old, struct netlink_ext_ack *extack)
357 {
358 	struct net_device *dev = qdisc_dev(sch);
359 	struct netdev_queue *dev_queue = mqprio_queue_get(sch, cl);
360 
361 	if (!dev_queue)
362 		return -EINVAL;
363 
364 	if (dev->flags & IFF_UP)
365 		dev_deactivate(dev);
366 
367 	*old = dev_graft_qdisc(dev_queue, new);
368 
369 	if (new)
370 		new->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
371 
372 	if (dev->flags & IFF_UP)
373 		dev_activate(dev);
374 
375 	return 0;
376 }
377 
378 static int dump_rates(struct mqprio_sched *priv,
379 		      struct tc_mqprio_qopt *opt, struct sk_buff *skb)
380 {
381 	struct nlattr *nest;
382 	int i;
383 
384 	if (priv->flags & TC_MQPRIO_F_MIN_RATE) {
385 		nest = nla_nest_start_noflag(skb, TCA_MQPRIO_MIN_RATE64);
386 		if (!nest)
387 			goto nla_put_failure;
388 
389 		for (i = 0; i < opt->num_tc; i++) {
390 			if (nla_put(skb, TCA_MQPRIO_MIN_RATE64,
391 				    sizeof(priv->min_rate[i]),
392 				    &priv->min_rate[i]))
393 				goto nla_put_failure;
394 		}
395 		nla_nest_end(skb, nest);
396 	}
397 
398 	if (priv->flags & TC_MQPRIO_F_MAX_RATE) {
399 		nest = nla_nest_start_noflag(skb, TCA_MQPRIO_MAX_RATE64);
400 		if (!nest)
401 			goto nla_put_failure;
402 
403 		for (i = 0; i < opt->num_tc; i++) {
404 			if (nla_put(skb, TCA_MQPRIO_MAX_RATE64,
405 				    sizeof(priv->max_rate[i]),
406 				    &priv->max_rate[i]))
407 				goto nla_put_failure;
408 		}
409 		nla_nest_end(skb, nest);
410 	}
411 	return 0;
412 
413 nla_put_failure:
414 	nla_nest_cancel(skb, nest);
415 	return -1;
416 }
417 
418 static int mqprio_dump(struct Qdisc *sch, struct sk_buff *skb)
419 {
420 	struct net_device *dev = qdisc_dev(sch);
421 	struct mqprio_sched *priv = qdisc_priv(sch);
422 	struct nlattr *nla = (struct nlattr *)skb_tail_pointer(skb);
423 	struct tc_mqprio_qopt opt = { 0 };
424 	struct Qdisc *qdisc;
425 	unsigned int ntx;
426 
427 	sch->q.qlen = 0;
428 	gnet_stats_basic_sync_init(&sch->bstats);
429 	memset(&sch->qstats, 0, sizeof(sch->qstats));
430 
431 	/* MQ supports lockless qdiscs. However, statistics accounting needs
432 	 * to account for all, none, or a mix of locked and unlocked child
433 	 * qdiscs. Percpu stats are added to counters in-band and locking
434 	 * qdisc totals are added at end.
435 	 */
436 	for (ntx = 0; ntx < dev->num_tx_queues; ntx++) {
437 		qdisc = netdev_get_tx_queue(dev, ntx)->qdisc_sleeping;
438 		spin_lock_bh(qdisc_lock(qdisc));
439 
440 		gnet_stats_add_basic(&sch->bstats, qdisc->cpu_bstats,
441 				     &qdisc->bstats, false);
442 		gnet_stats_add_queue(&sch->qstats, qdisc->cpu_qstats,
443 				     &qdisc->qstats);
444 		sch->q.qlen += qdisc_qlen(qdisc);
445 
446 		spin_unlock_bh(qdisc_lock(qdisc));
447 	}
448 
449 	mqprio_qopt_reconstruct(dev, &opt);
450 	opt.hw = priv->hw_offload;
451 
452 	if (nla_put(skb, TCA_OPTIONS, sizeof(opt), &opt))
453 		goto nla_put_failure;
454 
455 	if ((priv->flags & TC_MQPRIO_F_MODE) &&
456 	    nla_put_u16(skb, TCA_MQPRIO_MODE, priv->mode))
457 		goto nla_put_failure;
458 
459 	if ((priv->flags & TC_MQPRIO_F_SHAPER) &&
460 	    nla_put_u16(skb, TCA_MQPRIO_SHAPER, priv->shaper))
461 		goto nla_put_failure;
462 
463 	if ((priv->flags & TC_MQPRIO_F_MIN_RATE ||
464 	     priv->flags & TC_MQPRIO_F_MAX_RATE) &&
465 	    (dump_rates(priv, &opt, skb) != 0))
466 		goto nla_put_failure;
467 
468 	return nla_nest_end(skb, nla);
469 nla_put_failure:
470 	nlmsg_trim(skb, nla);
471 	return -1;
472 }
473 
474 static struct Qdisc *mqprio_leaf(struct Qdisc *sch, unsigned long cl)
475 {
476 	struct netdev_queue *dev_queue = mqprio_queue_get(sch, cl);
477 
478 	if (!dev_queue)
479 		return NULL;
480 
481 	return dev_queue->qdisc_sleeping;
482 }
483 
484 static unsigned long mqprio_find(struct Qdisc *sch, u32 classid)
485 {
486 	struct net_device *dev = qdisc_dev(sch);
487 	unsigned int ntx = TC_H_MIN(classid);
488 
489 	/* There are essentially two regions here that have valid classid
490 	 * values. The first region will have a classid value of 1 through
491 	 * num_tx_queues. All of these are backed by actual Qdiscs.
492 	 */
493 	if (ntx < TC_H_MIN_PRIORITY)
494 		return (ntx <= dev->num_tx_queues) ? ntx : 0;
495 
496 	/* The second region represents the hardware traffic classes. These
497 	 * are represented by classid values of TC_H_MIN_PRIORITY through
498 	 * TC_H_MIN_PRIORITY + netdev_get_num_tc - 1
499 	 */
500 	return ((ntx - TC_H_MIN_PRIORITY) < netdev_get_num_tc(dev)) ? ntx : 0;
501 }
502 
503 static int mqprio_dump_class(struct Qdisc *sch, unsigned long cl,
504 			 struct sk_buff *skb, struct tcmsg *tcm)
505 {
506 	if (cl < TC_H_MIN_PRIORITY) {
507 		struct netdev_queue *dev_queue = mqprio_queue_get(sch, cl);
508 		struct net_device *dev = qdisc_dev(sch);
509 		int tc = netdev_txq_to_tc(dev, cl - 1);
510 
511 		tcm->tcm_parent = (tc < 0) ? 0 :
512 			TC_H_MAKE(TC_H_MAJ(sch->handle),
513 				  TC_H_MIN(tc + TC_H_MIN_PRIORITY));
514 		tcm->tcm_info = dev_queue->qdisc_sleeping->handle;
515 	} else {
516 		tcm->tcm_parent = TC_H_ROOT;
517 		tcm->tcm_info = 0;
518 	}
519 	tcm->tcm_handle |= TC_H_MIN(cl);
520 	return 0;
521 }
522 
523 static int mqprio_dump_class_stats(struct Qdisc *sch, unsigned long cl,
524 				   struct gnet_dump *d)
525 	__releases(d->lock)
526 	__acquires(d->lock)
527 {
528 	if (cl >= TC_H_MIN_PRIORITY) {
529 		int i;
530 		__u32 qlen;
531 		struct gnet_stats_queue qstats = {0};
532 		struct gnet_stats_basic_sync bstats;
533 		struct net_device *dev = qdisc_dev(sch);
534 		struct netdev_tc_txq tc = dev->tc_to_txq[cl & TC_BITMASK];
535 
536 		gnet_stats_basic_sync_init(&bstats);
537 		/* Drop lock here it will be reclaimed before touching
538 		 * statistics this is required because the d->lock we
539 		 * hold here is the look on dev_queue->qdisc_sleeping
540 		 * also acquired below.
541 		 */
542 		if (d->lock)
543 			spin_unlock_bh(d->lock);
544 
545 		for (i = tc.offset; i < tc.offset + tc.count; i++) {
546 			struct netdev_queue *q = netdev_get_tx_queue(dev, i);
547 			struct Qdisc *qdisc = rtnl_dereference(q->qdisc);
548 
549 			spin_lock_bh(qdisc_lock(qdisc));
550 
551 			gnet_stats_add_basic(&bstats, qdisc->cpu_bstats,
552 					     &qdisc->bstats, false);
553 			gnet_stats_add_queue(&qstats, qdisc->cpu_qstats,
554 					     &qdisc->qstats);
555 			sch->q.qlen += qdisc_qlen(qdisc);
556 
557 			spin_unlock_bh(qdisc_lock(qdisc));
558 		}
559 		qlen = qdisc_qlen(sch) + qstats.qlen;
560 
561 		/* Reclaim root sleeping lock before completing stats */
562 		if (d->lock)
563 			spin_lock_bh(d->lock);
564 		if (gnet_stats_copy_basic(d, NULL, &bstats, false) < 0 ||
565 		    gnet_stats_copy_queue(d, NULL, &qstats, qlen) < 0)
566 			return -1;
567 	} else {
568 		struct netdev_queue *dev_queue = mqprio_queue_get(sch, cl);
569 
570 		sch = dev_queue->qdisc_sleeping;
571 		if (gnet_stats_copy_basic(d, sch->cpu_bstats,
572 					  &sch->bstats, true) < 0 ||
573 		    qdisc_qstats_copy(d, sch) < 0)
574 			return -1;
575 	}
576 	return 0;
577 }
578 
579 static void mqprio_walk(struct Qdisc *sch, struct qdisc_walker *arg)
580 {
581 	struct net_device *dev = qdisc_dev(sch);
582 	unsigned long ntx;
583 
584 	if (arg->stop)
585 		return;
586 
587 	/* Walk hierarchy with a virtual class per tc */
588 	arg->count = arg->skip;
589 	for (ntx = arg->skip; ntx < netdev_get_num_tc(dev); ntx++) {
590 		if (!tc_qdisc_stats_dump(sch, ntx + TC_H_MIN_PRIORITY, arg))
591 			return;
592 	}
593 
594 	/* Pad the values and skip over unused traffic classes */
595 	if (ntx < TC_MAX_QUEUE) {
596 		arg->count = TC_MAX_QUEUE;
597 		ntx = TC_MAX_QUEUE;
598 	}
599 
600 	/* Reset offset, sort out remaining per-queue qdiscs */
601 	for (ntx -= TC_MAX_QUEUE; ntx < dev->num_tx_queues; ntx++) {
602 		if (arg->fn(sch, ntx + 1, arg) < 0) {
603 			arg->stop = 1;
604 			return;
605 		}
606 		arg->count++;
607 	}
608 }
609 
610 static struct netdev_queue *mqprio_select_queue(struct Qdisc *sch,
611 						struct tcmsg *tcm)
612 {
613 	return mqprio_queue_get(sch, TC_H_MIN(tcm->tcm_parent));
614 }
615 
616 static const struct Qdisc_class_ops mqprio_class_ops = {
617 	.graft		= mqprio_graft,
618 	.leaf		= mqprio_leaf,
619 	.find		= mqprio_find,
620 	.walk		= mqprio_walk,
621 	.dump		= mqprio_dump_class,
622 	.dump_stats	= mqprio_dump_class_stats,
623 	.select_queue	= mqprio_select_queue,
624 };
625 
626 static struct Qdisc_ops mqprio_qdisc_ops __read_mostly = {
627 	.cl_ops		= &mqprio_class_ops,
628 	.id		= "mqprio",
629 	.priv_size	= sizeof(struct mqprio_sched),
630 	.init		= mqprio_init,
631 	.destroy	= mqprio_destroy,
632 	.attach		= mqprio_attach,
633 	.change_real_num_tx = mq_change_real_num_tx,
634 	.dump		= mqprio_dump,
635 	.owner		= THIS_MODULE,
636 };
637 
638 static int __init mqprio_module_init(void)
639 {
640 	return register_qdisc(&mqprio_qdisc_ops);
641 }
642 
643 static void __exit mqprio_module_exit(void)
644 {
645 	unregister_qdisc(&mqprio_qdisc_ops);
646 }
647 
648 module_init(mqprio_module_init);
649 module_exit(mqprio_module_exit);
650 
651 MODULE_LICENSE("GPL");
652