xref: /linux/net/sched/sch_mq.c (revision 653d6fd68d8e5b43d496ca8a1d38331d515a226b)
1 /*
2  * net/sched/sch_mq.c		Classful multiqueue dummy scheduler
3  *
4  * Copyright (c) 2009 Patrick McHardy <kaber@trash.net>
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/export.h>
15 #include <linux/string.h>
16 #include <linux/errno.h>
17 #include <linux/skbuff.h>
18 #include <net/netlink.h>
19 #include <net/pkt_sched.h>
20 #include <net/sch_generic.h>
21 
22 struct mq_sched {
23 	struct Qdisc		**qdiscs;
24 };
25 
26 static void mq_destroy(struct Qdisc *sch)
27 {
28 	struct net_device *dev = qdisc_dev(sch);
29 	struct mq_sched *priv = qdisc_priv(sch);
30 	unsigned int ntx;
31 
32 	if (!priv->qdiscs)
33 		return;
34 	for (ntx = 0; ntx < dev->num_tx_queues && priv->qdiscs[ntx]; ntx++)
35 		qdisc_destroy(priv->qdiscs[ntx]);
36 	kfree(priv->qdiscs);
37 }
38 
39 static int mq_init(struct Qdisc *sch, struct nlattr *opt,
40 		   struct netlink_ext_ack *extack)
41 {
42 	struct net_device *dev = qdisc_dev(sch);
43 	struct mq_sched *priv = qdisc_priv(sch);
44 	struct netdev_queue *dev_queue;
45 	struct Qdisc *qdisc;
46 	unsigned int ntx;
47 
48 	if (sch->parent != TC_H_ROOT)
49 		return -EOPNOTSUPP;
50 
51 	if (!netif_is_multiqueue(dev))
52 		return -EOPNOTSUPP;
53 
54 	/* pre-allocate qdiscs, attachment can't fail */
55 	priv->qdiscs = kcalloc(dev->num_tx_queues, sizeof(priv->qdiscs[0]),
56 			       GFP_KERNEL);
57 	if (!priv->qdiscs)
58 		return -ENOMEM;
59 
60 	for (ntx = 0; ntx < dev->num_tx_queues; ntx++) {
61 		dev_queue = netdev_get_tx_queue(dev, ntx);
62 		qdisc = qdisc_create_dflt(dev_queue, get_default_qdisc_ops(dev, ntx),
63 					  TC_H_MAKE(TC_H_MAJ(sch->handle),
64 						    TC_H_MIN(ntx + 1)));
65 		if (!qdisc)
66 			return -ENOMEM;
67 		priv->qdiscs[ntx] = qdisc;
68 		qdisc->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
69 	}
70 
71 	sch->flags |= TCQ_F_MQROOT;
72 	return 0;
73 }
74 
75 static void mq_attach(struct Qdisc *sch)
76 {
77 	struct net_device *dev = qdisc_dev(sch);
78 	struct mq_sched *priv = qdisc_priv(sch);
79 	struct Qdisc *qdisc, *old;
80 	unsigned int ntx;
81 
82 	for (ntx = 0; ntx < dev->num_tx_queues; ntx++) {
83 		qdisc = priv->qdiscs[ntx];
84 		old = dev_graft_qdisc(qdisc->dev_queue, qdisc);
85 		if (old)
86 			qdisc_destroy(old);
87 #ifdef CONFIG_NET_SCHED
88 		if (ntx < dev->real_num_tx_queues)
89 			qdisc_hash_add(qdisc, false);
90 #endif
91 
92 	}
93 	kfree(priv->qdiscs);
94 	priv->qdiscs = NULL;
95 }
96 
97 static int mq_dump(struct Qdisc *sch, struct sk_buff *skb)
98 {
99 	struct net_device *dev = qdisc_dev(sch);
100 	struct Qdisc *qdisc;
101 	unsigned int ntx;
102 	__u32 qlen = 0;
103 
104 	sch->q.qlen = 0;
105 	memset(&sch->bstats, 0, sizeof(sch->bstats));
106 	memset(&sch->qstats, 0, sizeof(sch->qstats));
107 
108 	/* MQ supports lockless qdiscs. However, statistics accounting needs
109 	 * to account for all, none, or a mix of locked and unlocked child
110 	 * qdiscs. Percpu stats are added to counters in-band and locking
111 	 * qdisc totals are added at end.
112 	 */
113 	for (ntx = 0; ntx < dev->num_tx_queues; ntx++) {
114 		qdisc = netdev_get_tx_queue(dev, ntx)->qdisc_sleeping;
115 		spin_lock_bh(qdisc_lock(qdisc));
116 
117 		if (qdisc_is_percpu_stats(qdisc)) {
118 			qlen = qdisc_qlen_sum(qdisc);
119 			__gnet_stats_copy_basic(NULL, &sch->bstats,
120 						qdisc->cpu_bstats,
121 						&qdisc->bstats);
122 			__gnet_stats_copy_queue(&sch->qstats,
123 						qdisc->cpu_qstats,
124 						&qdisc->qstats, qlen);
125 		} else {
126 			sch->q.qlen		+= qdisc->q.qlen;
127 			sch->bstats.bytes	+= qdisc->bstats.bytes;
128 			sch->bstats.packets	+= qdisc->bstats.packets;
129 			sch->qstats.backlog	+= qdisc->qstats.backlog;
130 			sch->qstats.drops	+= qdisc->qstats.drops;
131 			sch->qstats.requeues	+= qdisc->qstats.requeues;
132 			sch->qstats.overlimits	+= qdisc->qstats.overlimits;
133 		}
134 
135 		spin_unlock_bh(qdisc_lock(qdisc));
136 	}
137 
138 	return 0;
139 }
140 
141 static struct netdev_queue *mq_queue_get(struct Qdisc *sch, unsigned long cl)
142 {
143 	struct net_device *dev = qdisc_dev(sch);
144 	unsigned long ntx = cl - 1;
145 
146 	if (ntx >= dev->num_tx_queues)
147 		return NULL;
148 	return netdev_get_tx_queue(dev, ntx);
149 }
150 
151 static struct netdev_queue *mq_select_queue(struct Qdisc *sch,
152 					    struct tcmsg *tcm)
153 {
154 	return mq_queue_get(sch, TC_H_MIN(tcm->tcm_parent));
155 }
156 
157 static int mq_graft(struct Qdisc *sch, unsigned long cl, struct Qdisc *new,
158 		    struct Qdisc **old, struct netlink_ext_ack *extack)
159 {
160 	struct netdev_queue *dev_queue = mq_queue_get(sch, cl);
161 	struct net_device *dev = qdisc_dev(sch);
162 
163 	if (dev->flags & IFF_UP)
164 		dev_deactivate(dev);
165 
166 	*old = dev_graft_qdisc(dev_queue, new);
167 	if (new)
168 		new->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
169 	if (dev->flags & IFF_UP)
170 		dev_activate(dev);
171 	return 0;
172 }
173 
174 static struct Qdisc *mq_leaf(struct Qdisc *sch, unsigned long cl)
175 {
176 	struct netdev_queue *dev_queue = mq_queue_get(sch, cl);
177 
178 	return dev_queue->qdisc_sleeping;
179 }
180 
181 static unsigned long mq_find(struct Qdisc *sch, u32 classid)
182 {
183 	unsigned int ntx = TC_H_MIN(classid);
184 
185 	if (!mq_queue_get(sch, ntx))
186 		return 0;
187 	return ntx;
188 }
189 
190 static int mq_dump_class(struct Qdisc *sch, unsigned long cl,
191 			 struct sk_buff *skb, struct tcmsg *tcm)
192 {
193 	struct netdev_queue *dev_queue = mq_queue_get(sch, cl);
194 
195 	tcm->tcm_parent = TC_H_ROOT;
196 	tcm->tcm_handle |= TC_H_MIN(cl);
197 	tcm->tcm_info = dev_queue->qdisc_sleeping->handle;
198 	return 0;
199 }
200 
201 static int mq_dump_class_stats(struct Qdisc *sch, unsigned long cl,
202 			       struct gnet_dump *d)
203 {
204 	struct netdev_queue *dev_queue = mq_queue_get(sch, cl);
205 
206 	sch = dev_queue->qdisc_sleeping;
207 	if (gnet_stats_copy_basic(&sch->running, d, NULL, &sch->bstats) < 0 ||
208 	    gnet_stats_copy_queue(d, NULL, &sch->qstats, sch->q.qlen) < 0)
209 		return -1;
210 	return 0;
211 }
212 
213 static void mq_walk(struct Qdisc *sch, struct qdisc_walker *arg)
214 {
215 	struct net_device *dev = qdisc_dev(sch);
216 	unsigned int ntx;
217 
218 	if (arg->stop)
219 		return;
220 
221 	arg->count = arg->skip;
222 	for (ntx = arg->skip; ntx < dev->num_tx_queues; ntx++) {
223 		if (arg->fn(sch, ntx + 1, arg) < 0) {
224 			arg->stop = 1;
225 			break;
226 		}
227 		arg->count++;
228 	}
229 }
230 
231 static const struct Qdisc_class_ops mq_class_ops = {
232 	.select_queue	= mq_select_queue,
233 	.graft		= mq_graft,
234 	.leaf		= mq_leaf,
235 	.find		= mq_find,
236 	.walk		= mq_walk,
237 	.dump		= mq_dump_class,
238 	.dump_stats	= mq_dump_class_stats,
239 };
240 
241 struct Qdisc_ops mq_qdisc_ops __read_mostly = {
242 	.cl_ops		= &mq_class_ops,
243 	.id		= "mq",
244 	.priv_size	= sizeof(struct mq_sched),
245 	.init		= mq_init,
246 	.destroy	= mq_destroy,
247 	.attach		= mq_attach,
248 	.dump		= mq_dump,
249 	.owner		= THIS_MODULE,
250 };
251