xref: /linux/net/sched/sch_mqprio.c (revision f49f4ab95c301dbccad0efe85296d908b8ae7ad4)
1 /*
2  * net/sched/sch_mqprio.c
3  *
4  * Copyright (c) 2010 John Fastabend <john.r.fastabend@intel.com>
5  *
6  * This program is free software; you can redistribute it and/or
7  * modify it under the terms of the GNU General Public License
8  * version 2 as published by the Free Software Foundation.
9  */
10 
11 #include <linux/types.h>
12 #include <linux/slab.h>
13 #include <linux/kernel.h>
14 #include <linux/string.h>
15 #include <linux/errno.h>
16 #include <linux/skbuff.h>
17 #include <linux/module.h>
18 #include <net/netlink.h>
19 #include <net/pkt_sched.h>
20 #include <net/sch_generic.h>
21 
22 struct mqprio_sched {
23 	struct Qdisc		**qdiscs;
24 	int hw_owned;
25 };
26 
27 static void mqprio_destroy(struct Qdisc *sch)
28 {
29 	struct net_device *dev = qdisc_dev(sch);
30 	struct mqprio_sched *priv = qdisc_priv(sch);
31 	unsigned int ntx;
32 
33 	if (priv->qdiscs) {
34 		for (ntx = 0;
35 		     ntx < dev->num_tx_queues && priv->qdiscs[ntx];
36 		     ntx++)
37 			qdisc_destroy(priv->qdiscs[ntx]);
38 		kfree(priv->qdiscs);
39 	}
40 
41 	if (priv->hw_owned && dev->netdev_ops->ndo_setup_tc)
42 		dev->netdev_ops->ndo_setup_tc(dev, 0);
43 	else
44 		netdev_set_num_tc(dev, 0);
45 }
46 
47 static int mqprio_parse_opt(struct net_device *dev, struct tc_mqprio_qopt *qopt)
48 {
49 	int i, j;
50 
51 	/* Verify num_tc is not out of max range */
52 	if (qopt->num_tc > TC_MAX_QUEUE)
53 		return -EINVAL;
54 
55 	/* Verify priority mapping uses valid tcs */
56 	for (i = 0; i < TC_BITMASK + 1; i++) {
57 		if (qopt->prio_tc_map[i] >= qopt->num_tc)
58 			return -EINVAL;
59 	}
60 
61 	/* net_device does not support requested operation */
62 	if (qopt->hw && !dev->netdev_ops->ndo_setup_tc)
63 		return -EINVAL;
64 
65 	/* if hw owned qcount and qoffset are taken from LLD so
66 	 * no reason to verify them here
67 	 */
68 	if (qopt->hw)
69 		return 0;
70 
71 	for (i = 0; i < qopt->num_tc; i++) {
72 		unsigned int last = qopt->offset[i] + qopt->count[i];
73 
74 		/* Verify the queue count is in tx range being equal to the
75 		 * real_num_tx_queues indicates the last queue is in use.
76 		 */
77 		if (qopt->offset[i] >= dev->real_num_tx_queues ||
78 		    !qopt->count[i] ||
79 		    last > dev->real_num_tx_queues)
80 			return -EINVAL;
81 
82 		/* Verify that the offset and counts do not overlap */
83 		for (j = i + 1; j < qopt->num_tc; j++) {
84 			if (last > qopt->offset[j])
85 				return -EINVAL;
86 		}
87 	}
88 
89 	return 0;
90 }
91 
92 static int mqprio_init(struct Qdisc *sch, struct nlattr *opt)
93 {
94 	struct net_device *dev = qdisc_dev(sch);
95 	struct mqprio_sched *priv = qdisc_priv(sch);
96 	struct netdev_queue *dev_queue;
97 	struct Qdisc *qdisc;
98 	int i, err = -EOPNOTSUPP;
99 	struct tc_mqprio_qopt *qopt = NULL;
100 
101 	BUILD_BUG_ON(TC_MAX_QUEUE != TC_QOPT_MAX_QUEUE);
102 	BUILD_BUG_ON(TC_BITMASK != TC_QOPT_BITMASK);
103 
104 	if (sch->parent != TC_H_ROOT)
105 		return -EOPNOTSUPP;
106 
107 	if (!netif_is_multiqueue(dev))
108 		return -EOPNOTSUPP;
109 
110 	if (!opt || nla_len(opt) < sizeof(*qopt))
111 		return -EINVAL;
112 
113 	qopt = nla_data(opt);
114 	if (mqprio_parse_opt(dev, qopt))
115 		return -EINVAL;
116 
117 	/* pre-allocate qdisc, attachment can't fail */
118 	priv->qdiscs = kcalloc(dev->num_tx_queues, sizeof(priv->qdiscs[0]),
119 			       GFP_KERNEL);
120 	if (priv->qdiscs == NULL) {
121 		err = -ENOMEM;
122 		goto err;
123 	}
124 
125 	for (i = 0; i < dev->num_tx_queues; i++) {
126 		dev_queue = netdev_get_tx_queue(dev, i);
127 		qdisc = qdisc_create_dflt(dev_queue, &pfifo_fast_ops,
128 					  TC_H_MAKE(TC_H_MAJ(sch->handle),
129 						    TC_H_MIN(i + 1)));
130 		if (qdisc == NULL) {
131 			err = -ENOMEM;
132 			goto err;
133 		}
134 		priv->qdiscs[i] = qdisc;
135 	}
136 
137 	/* If the mqprio options indicate that hardware should own
138 	 * the queue mapping then run ndo_setup_tc otherwise use the
139 	 * supplied and verified mapping
140 	 */
141 	if (qopt->hw) {
142 		priv->hw_owned = 1;
143 		err = dev->netdev_ops->ndo_setup_tc(dev, qopt->num_tc);
144 		if (err)
145 			goto err;
146 	} else {
147 		netdev_set_num_tc(dev, qopt->num_tc);
148 		for (i = 0; i < qopt->num_tc; i++)
149 			netdev_set_tc_queue(dev, i,
150 					    qopt->count[i], qopt->offset[i]);
151 	}
152 
153 	/* Always use supplied priority mappings */
154 	for (i = 0; i < TC_BITMASK + 1; i++)
155 		netdev_set_prio_tc_map(dev, i, qopt->prio_tc_map[i]);
156 
157 	sch->flags |= TCQ_F_MQROOT;
158 	return 0;
159 
160 err:
161 	mqprio_destroy(sch);
162 	return err;
163 }
164 
165 static void mqprio_attach(struct Qdisc *sch)
166 {
167 	struct net_device *dev = qdisc_dev(sch);
168 	struct mqprio_sched *priv = qdisc_priv(sch);
169 	struct Qdisc *qdisc;
170 	unsigned int ntx;
171 
172 	/* Attach underlying qdisc */
173 	for (ntx = 0; ntx < dev->num_tx_queues; ntx++) {
174 		qdisc = priv->qdiscs[ntx];
175 		qdisc = dev_graft_qdisc(qdisc->dev_queue, qdisc);
176 		if (qdisc)
177 			qdisc_destroy(qdisc);
178 	}
179 	kfree(priv->qdiscs);
180 	priv->qdiscs = NULL;
181 }
182 
183 static struct netdev_queue *mqprio_queue_get(struct Qdisc *sch,
184 					     unsigned long cl)
185 {
186 	struct net_device *dev = qdisc_dev(sch);
187 	unsigned long ntx = cl - 1 - netdev_get_num_tc(dev);
188 
189 	if (ntx >= dev->num_tx_queues)
190 		return NULL;
191 	return netdev_get_tx_queue(dev, ntx);
192 }
193 
194 static int mqprio_graft(struct Qdisc *sch, unsigned long cl, struct Qdisc *new,
195 		    struct Qdisc **old)
196 {
197 	struct net_device *dev = qdisc_dev(sch);
198 	struct netdev_queue *dev_queue = mqprio_queue_get(sch, cl);
199 
200 	if (!dev_queue)
201 		return -EINVAL;
202 
203 	if (dev->flags & IFF_UP)
204 		dev_deactivate(dev);
205 
206 	*old = dev_graft_qdisc(dev_queue, new);
207 
208 	if (dev->flags & IFF_UP)
209 		dev_activate(dev);
210 
211 	return 0;
212 }
213 
214 static int mqprio_dump(struct Qdisc *sch, struct sk_buff *skb)
215 {
216 	struct net_device *dev = qdisc_dev(sch);
217 	struct mqprio_sched *priv = qdisc_priv(sch);
218 	unsigned char *b = skb_tail_pointer(skb);
219 	struct tc_mqprio_qopt opt = { 0 };
220 	struct Qdisc *qdisc;
221 	unsigned int i;
222 
223 	sch->q.qlen = 0;
224 	memset(&sch->bstats, 0, sizeof(sch->bstats));
225 	memset(&sch->qstats, 0, sizeof(sch->qstats));
226 
227 	for (i = 0; i < dev->num_tx_queues; i++) {
228 		qdisc = netdev_get_tx_queue(dev, i)->qdisc;
229 		spin_lock_bh(qdisc_lock(qdisc));
230 		sch->q.qlen		+= qdisc->q.qlen;
231 		sch->bstats.bytes	+= qdisc->bstats.bytes;
232 		sch->bstats.packets	+= qdisc->bstats.packets;
233 		sch->qstats.qlen	+= qdisc->qstats.qlen;
234 		sch->qstats.backlog	+= qdisc->qstats.backlog;
235 		sch->qstats.drops	+= qdisc->qstats.drops;
236 		sch->qstats.requeues	+= qdisc->qstats.requeues;
237 		sch->qstats.overlimits	+= qdisc->qstats.overlimits;
238 		spin_unlock_bh(qdisc_lock(qdisc));
239 	}
240 
241 	opt.num_tc = netdev_get_num_tc(dev);
242 	memcpy(opt.prio_tc_map, dev->prio_tc_map, sizeof(opt.prio_tc_map));
243 	opt.hw = priv->hw_owned;
244 
245 	for (i = 0; i < netdev_get_num_tc(dev); i++) {
246 		opt.count[i] = dev->tc_to_txq[i].count;
247 		opt.offset[i] = dev->tc_to_txq[i].offset;
248 	}
249 
250 	if (nla_put(skb, TCA_OPTIONS, sizeof(opt), &opt))
251 		goto nla_put_failure;
252 
253 	return skb->len;
254 nla_put_failure:
255 	nlmsg_trim(skb, b);
256 	return -1;
257 }
258 
259 static struct Qdisc *mqprio_leaf(struct Qdisc *sch, unsigned long cl)
260 {
261 	struct netdev_queue *dev_queue = mqprio_queue_get(sch, cl);
262 
263 	if (!dev_queue)
264 		return NULL;
265 
266 	return dev_queue->qdisc_sleeping;
267 }
268 
269 static unsigned long mqprio_get(struct Qdisc *sch, u32 classid)
270 {
271 	struct net_device *dev = qdisc_dev(sch);
272 	unsigned int ntx = TC_H_MIN(classid);
273 
274 	if (ntx > dev->num_tx_queues + netdev_get_num_tc(dev))
275 		return 0;
276 	return ntx;
277 }
278 
279 static void mqprio_put(struct Qdisc *sch, unsigned long cl)
280 {
281 }
282 
283 static int mqprio_dump_class(struct Qdisc *sch, unsigned long cl,
284 			 struct sk_buff *skb, struct tcmsg *tcm)
285 {
286 	struct net_device *dev = qdisc_dev(sch);
287 
288 	if (cl <= netdev_get_num_tc(dev)) {
289 		tcm->tcm_parent = TC_H_ROOT;
290 		tcm->tcm_info = 0;
291 	} else {
292 		int i;
293 		struct netdev_queue *dev_queue;
294 
295 		dev_queue = mqprio_queue_get(sch, cl);
296 		tcm->tcm_parent = 0;
297 		for (i = 0; i < netdev_get_num_tc(dev); i++) {
298 			struct netdev_tc_txq tc = dev->tc_to_txq[i];
299 			int q_idx = cl - netdev_get_num_tc(dev);
300 
301 			if (q_idx > tc.offset &&
302 			    q_idx <= tc.offset + tc.count) {
303 				tcm->tcm_parent =
304 					TC_H_MAKE(TC_H_MAJ(sch->handle),
305 						  TC_H_MIN(i + 1));
306 				break;
307 			}
308 		}
309 		tcm->tcm_info = dev_queue->qdisc_sleeping->handle;
310 	}
311 	tcm->tcm_handle |= TC_H_MIN(cl);
312 	return 0;
313 }
314 
315 static int mqprio_dump_class_stats(struct Qdisc *sch, unsigned long cl,
316 				   struct gnet_dump *d)
317 	__releases(d->lock)
318 	__acquires(d->lock)
319 {
320 	struct net_device *dev = qdisc_dev(sch);
321 
322 	if (cl <= netdev_get_num_tc(dev)) {
323 		int i;
324 		struct Qdisc *qdisc;
325 		struct gnet_stats_queue qstats = {0};
326 		struct gnet_stats_basic_packed bstats = {0};
327 		struct netdev_tc_txq tc = dev->tc_to_txq[cl - 1];
328 
329 		/* Drop lock here it will be reclaimed before touching
330 		 * statistics this is required because the d->lock we
331 		 * hold here is the look on dev_queue->qdisc_sleeping
332 		 * also acquired below.
333 		 */
334 		spin_unlock_bh(d->lock);
335 
336 		for (i = tc.offset; i < tc.offset + tc.count; i++) {
337 			qdisc = netdev_get_tx_queue(dev, i)->qdisc;
338 			spin_lock_bh(qdisc_lock(qdisc));
339 			bstats.bytes      += qdisc->bstats.bytes;
340 			bstats.packets    += qdisc->bstats.packets;
341 			qstats.qlen       += qdisc->qstats.qlen;
342 			qstats.backlog    += qdisc->qstats.backlog;
343 			qstats.drops      += qdisc->qstats.drops;
344 			qstats.requeues   += qdisc->qstats.requeues;
345 			qstats.overlimits += qdisc->qstats.overlimits;
346 			spin_unlock_bh(qdisc_lock(qdisc));
347 		}
348 		/* Reclaim root sleeping lock before completing stats */
349 		spin_lock_bh(d->lock);
350 		if (gnet_stats_copy_basic(d, &bstats) < 0 ||
351 		    gnet_stats_copy_queue(d, &qstats) < 0)
352 			return -1;
353 	} else {
354 		struct netdev_queue *dev_queue = mqprio_queue_get(sch, cl);
355 
356 		sch = dev_queue->qdisc_sleeping;
357 		sch->qstats.qlen = sch->q.qlen;
358 		if (gnet_stats_copy_basic(d, &sch->bstats) < 0 ||
359 		    gnet_stats_copy_queue(d, &sch->qstats) < 0)
360 			return -1;
361 	}
362 	return 0;
363 }
364 
365 static void mqprio_walk(struct Qdisc *sch, struct qdisc_walker *arg)
366 {
367 	struct net_device *dev = qdisc_dev(sch);
368 	unsigned long ntx;
369 
370 	if (arg->stop)
371 		return;
372 
373 	/* Walk hierarchy with a virtual class per tc */
374 	arg->count = arg->skip;
375 	for (ntx = arg->skip;
376 	     ntx < dev->num_tx_queues + netdev_get_num_tc(dev);
377 	     ntx++) {
378 		if (arg->fn(sch, ntx + 1, arg) < 0) {
379 			arg->stop = 1;
380 			break;
381 		}
382 		arg->count++;
383 	}
384 }
385 
386 static const struct Qdisc_class_ops mqprio_class_ops = {
387 	.graft		= mqprio_graft,
388 	.leaf		= mqprio_leaf,
389 	.get		= mqprio_get,
390 	.put		= mqprio_put,
391 	.walk		= mqprio_walk,
392 	.dump		= mqprio_dump_class,
393 	.dump_stats	= mqprio_dump_class_stats,
394 };
395 
396 static struct Qdisc_ops mqprio_qdisc_ops __read_mostly = {
397 	.cl_ops		= &mqprio_class_ops,
398 	.id		= "mqprio",
399 	.priv_size	= sizeof(struct mqprio_sched),
400 	.init		= mqprio_init,
401 	.destroy	= mqprio_destroy,
402 	.attach		= mqprio_attach,
403 	.dump		= mqprio_dump,
404 	.owner		= THIS_MODULE,
405 };
406 
407 static int __init mqprio_module_init(void)
408 {
409 	return register_qdisc(&mqprio_qdisc_ops);
410 }
411 
412 static void __exit mqprio_module_exit(void)
413 {
414 	unregister_qdisc(&mqprio_qdisc_ops);
415 }
416 
417 module_init(mqprio_module_init);
418 module_exit(mqprio_module_exit);
419 
420 MODULE_LICENSE("GPL");
421