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
22struct mqprio_sched {
23	struct Qdisc		**qdiscs;
24	int hw_owned;
25};
26
27static 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
47static 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
92static 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, default_qdisc_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		qdisc->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
136	}
137
138	/* If the mqprio options indicate that hardware should own
139	 * the queue mapping then run ndo_setup_tc otherwise use the
140	 * supplied and verified mapping
141	 */
142	if (qopt->hw) {
143		priv->hw_owned = 1;
144		err = dev->netdev_ops->ndo_setup_tc(dev, qopt->num_tc);
145		if (err)
146			goto err;
147	} else {
148		netdev_set_num_tc(dev, qopt->num_tc);
149		for (i = 0; i < qopt->num_tc; i++)
150			netdev_set_tc_queue(dev, i,
151					    qopt->count[i], qopt->offset[i]);
152	}
153
154	/* Always use supplied priority mappings */
155	for (i = 0; i < TC_BITMASK + 1; i++)
156		netdev_set_prio_tc_map(dev, i, qopt->prio_tc_map[i]);
157
158	sch->flags |= TCQ_F_MQROOT;
159	return 0;
160
161err:
162	mqprio_destroy(sch);
163	return err;
164}
165
166static void mqprio_attach(struct Qdisc *sch)
167{
168	struct net_device *dev = qdisc_dev(sch);
169	struct mqprio_sched *priv = qdisc_priv(sch);
170	struct Qdisc *qdisc, *old;
171	unsigned int ntx;
172
173	/* Attach underlying qdisc */
174	for (ntx = 0; ntx < dev->num_tx_queues; ntx++) {
175		qdisc = priv->qdiscs[ntx];
176		old = dev_graft_qdisc(qdisc->dev_queue, qdisc);
177		if (old)
178			qdisc_destroy(old);
179		if (ntx < dev->real_num_tx_queues)
180			qdisc_list_add(qdisc);
181	}
182	kfree(priv->qdiscs);
183	priv->qdiscs = NULL;
184}
185
186static struct netdev_queue *mqprio_queue_get(struct Qdisc *sch,
187					     unsigned long cl)
188{
189	struct net_device *dev = qdisc_dev(sch);
190	unsigned long ntx = cl - 1 - netdev_get_num_tc(dev);
191
192	if (ntx >= dev->num_tx_queues)
193		return NULL;
194	return netdev_get_tx_queue(dev, ntx);
195}
196
197static int mqprio_graft(struct Qdisc *sch, unsigned long cl, struct Qdisc *new,
198		    struct Qdisc **old)
199{
200	struct net_device *dev = qdisc_dev(sch);
201	struct netdev_queue *dev_queue = mqprio_queue_get(sch, cl);
202
203	if (!dev_queue)
204		return -EINVAL;
205
206	if (dev->flags & IFF_UP)
207		dev_deactivate(dev);
208
209	*old = dev_graft_qdisc(dev_queue, new);
210
211	if (new)
212		new->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
213
214	if (dev->flags & IFF_UP)
215		dev_activate(dev);
216
217	return 0;
218}
219
220static int mqprio_dump(struct Qdisc *sch, struct sk_buff *skb)
221{
222	struct net_device *dev = qdisc_dev(sch);
223	struct mqprio_sched *priv = qdisc_priv(sch);
224	unsigned char *b = skb_tail_pointer(skb);
225	struct tc_mqprio_qopt opt = { 0 };
226	struct Qdisc *qdisc;
227	unsigned int i;
228
229	sch->q.qlen = 0;
230	memset(&sch->bstats, 0, sizeof(sch->bstats));
231	memset(&sch->qstats, 0, sizeof(sch->qstats));
232
233	for (i = 0; i < dev->num_tx_queues; i++) {
234		qdisc = rtnl_dereference(netdev_get_tx_queue(dev, i)->qdisc);
235		spin_lock_bh(qdisc_lock(qdisc));
236		sch->q.qlen		+= qdisc->q.qlen;
237		sch->bstats.bytes	+= qdisc->bstats.bytes;
238		sch->bstats.packets	+= qdisc->bstats.packets;
239		sch->qstats.backlog	+= qdisc->qstats.backlog;
240		sch->qstats.drops	+= qdisc->qstats.drops;
241		sch->qstats.requeues	+= qdisc->qstats.requeues;
242		sch->qstats.overlimits	+= qdisc->qstats.overlimits;
243		spin_unlock_bh(qdisc_lock(qdisc));
244	}
245
246	opt.num_tc = netdev_get_num_tc(dev);
247	memcpy(opt.prio_tc_map, dev->prio_tc_map, sizeof(opt.prio_tc_map));
248	opt.hw = priv->hw_owned;
249
250	for (i = 0; i < netdev_get_num_tc(dev); i++) {
251		opt.count[i] = dev->tc_to_txq[i].count;
252		opt.offset[i] = dev->tc_to_txq[i].offset;
253	}
254
255	if (nla_put(skb, TCA_OPTIONS, sizeof(opt), &opt))
256		goto nla_put_failure;
257
258	return skb->len;
259nla_put_failure:
260	nlmsg_trim(skb, b);
261	return -1;
262}
263
264static struct Qdisc *mqprio_leaf(struct Qdisc *sch, unsigned long cl)
265{
266	struct netdev_queue *dev_queue = mqprio_queue_get(sch, cl);
267
268	if (!dev_queue)
269		return NULL;
270
271	return dev_queue->qdisc_sleeping;
272}
273
274static unsigned long mqprio_get(struct Qdisc *sch, u32 classid)
275{
276	struct net_device *dev = qdisc_dev(sch);
277	unsigned int ntx = TC_H_MIN(classid);
278
279	if (ntx > dev->num_tx_queues + netdev_get_num_tc(dev))
280		return 0;
281	return ntx;
282}
283
284static void mqprio_put(struct Qdisc *sch, unsigned long cl)
285{
286}
287
288static int mqprio_dump_class(struct Qdisc *sch, unsigned long cl,
289			 struct sk_buff *skb, struct tcmsg *tcm)
290{
291	struct net_device *dev = qdisc_dev(sch);
292
293	if (cl <= netdev_get_num_tc(dev)) {
294		tcm->tcm_parent = TC_H_ROOT;
295		tcm->tcm_info = 0;
296	} else {
297		int i;
298		struct netdev_queue *dev_queue;
299
300		dev_queue = mqprio_queue_get(sch, cl);
301		tcm->tcm_parent = 0;
302		for (i = 0; i < netdev_get_num_tc(dev); i++) {
303			struct netdev_tc_txq tc = dev->tc_to_txq[i];
304			int q_idx = cl - netdev_get_num_tc(dev);
305
306			if (q_idx > tc.offset &&
307			    q_idx <= tc.offset + tc.count) {
308				tcm->tcm_parent =
309					TC_H_MAKE(TC_H_MAJ(sch->handle),
310						  TC_H_MIN(i + 1));
311				break;
312			}
313		}
314		tcm->tcm_info = dev_queue->qdisc_sleeping->handle;
315	}
316	tcm->tcm_handle |= TC_H_MIN(cl);
317	return 0;
318}
319
320static int mqprio_dump_class_stats(struct Qdisc *sch, unsigned long cl,
321				   struct gnet_dump *d)
322	__releases(d->lock)
323	__acquires(d->lock)
324{
325	struct net_device *dev = qdisc_dev(sch);
326
327	if (cl <= netdev_get_num_tc(dev)) {
328		int i;
329		__u32 qlen = 0;
330		struct Qdisc *qdisc;
331		struct gnet_stats_queue qstats = {0};
332		struct gnet_stats_basic_packed bstats = {0};
333		struct netdev_tc_txq tc = dev->tc_to_txq[cl - 1];
334
335		/* Drop lock here it will be reclaimed before touching
336		 * statistics this is required because the d->lock we
337		 * hold here is the look on dev_queue->qdisc_sleeping
338		 * also acquired below.
339		 */
340		spin_unlock_bh(d->lock);
341
342		for (i = tc.offset; i < tc.offset + tc.count; i++) {
343			struct netdev_queue *q = netdev_get_tx_queue(dev, i);
344
345			qdisc = rtnl_dereference(q->qdisc);
346			spin_lock_bh(qdisc_lock(qdisc));
347			qlen		  += qdisc->q.qlen;
348			bstats.bytes      += qdisc->bstats.bytes;
349			bstats.packets    += qdisc->bstats.packets;
350			qstats.backlog    += qdisc->qstats.backlog;
351			qstats.drops      += qdisc->qstats.drops;
352			qstats.requeues   += qdisc->qstats.requeues;
353			qstats.overlimits += qdisc->qstats.overlimits;
354			spin_unlock_bh(qdisc_lock(qdisc));
355		}
356		/* Reclaim root sleeping lock before completing stats */
357		spin_lock_bh(d->lock);
358		if (gnet_stats_copy_basic(d, NULL, &bstats) < 0 ||
359		    gnet_stats_copy_queue(d, NULL, &qstats, qlen) < 0)
360			return -1;
361	} else {
362		struct netdev_queue *dev_queue = mqprio_queue_get(sch, cl);
363
364		sch = dev_queue->qdisc_sleeping;
365		if (gnet_stats_copy_basic(d, NULL, &sch->bstats) < 0 ||
366		    gnet_stats_copy_queue(d, NULL,
367					  &sch->qstats, sch->q.qlen) < 0)
368			return -1;
369	}
370	return 0;
371}
372
373static void mqprio_walk(struct Qdisc *sch, struct qdisc_walker *arg)
374{
375	struct net_device *dev = qdisc_dev(sch);
376	unsigned long ntx;
377
378	if (arg->stop)
379		return;
380
381	/* Walk hierarchy with a virtual class per tc */
382	arg->count = arg->skip;
383	for (ntx = arg->skip;
384	     ntx < dev->num_tx_queues + netdev_get_num_tc(dev);
385	     ntx++) {
386		if (arg->fn(sch, ntx + 1, arg) < 0) {
387			arg->stop = 1;
388			break;
389		}
390		arg->count++;
391	}
392}
393
394static const struct Qdisc_class_ops mqprio_class_ops = {
395	.graft		= mqprio_graft,
396	.leaf		= mqprio_leaf,
397	.get		= mqprio_get,
398	.put		= mqprio_put,
399	.walk		= mqprio_walk,
400	.dump		= mqprio_dump_class,
401	.dump_stats	= mqprio_dump_class_stats,
402};
403
404static struct Qdisc_ops mqprio_qdisc_ops __read_mostly = {
405	.cl_ops		= &mqprio_class_ops,
406	.id		= "mqprio",
407	.priv_size	= sizeof(struct mqprio_sched),
408	.init		= mqprio_init,
409	.destroy	= mqprio_destroy,
410	.attach		= mqprio_attach,
411	.dump		= mqprio_dump,
412	.owner		= THIS_MODULE,
413};
414
415static int __init mqprio_module_init(void)
416{
417	return register_qdisc(&mqprio_qdisc_ops);
418}
419
420static void __exit mqprio_module_exit(void)
421{
422	unregister_qdisc(&mqprio_qdisc_ops);
423}
424
425module_init(mqprio_module_init);
426module_exit(mqprio_module_exit);
427
428MODULE_LICENSE("GPL");
429