xref: /linux/net/sched/sch_mq.c (revision a38a98821c939e67e5906bddbed1d15af5ca860d)
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 					  extack);
66 		if (!qdisc)
67 			return -ENOMEM;
68 		priv->qdiscs[ntx] = qdisc;
69 		qdisc->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
70 	}
71 
72 	sch->flags |= TCQ_F_MQROOT;
73 	return 0;
74 }
75 
76 static void mq_attach(struct Qdisc *sch)
77 {
78 	struct net_device *dev = qdisc_dev(sch);
79 	struct mq_sched *priv = qdisc_priv(sch);
80 	struct Qdisc *qdisc, *old;
81 	unsigned int ntx;
82 
83 	for (ntx = 0; ntx < dev->num_tx_queues; ntx++) {
84 		qdisc = priv->qdiscs[ntx];
85 		old = dev_graft_qdisc(qdisc->dev_queue, qdisc);
86 		if (old)
87 			qdisc_destroy(old);
88 #ifdef CONFIG_NET_SCHED
89 		if (ntx < dev->real_num_tx_queues)
90 			qdisc_hash_add(qdisc, false);
91 #endif
92 
93 	}
94 	kfree(priv->qdiscs);
95 	priv->qdiscs = NULL;
96 }
97 
98 static int mq_dump(struct Qdisc *sch, struct sk_buff *skb)
99 {
100 	struct net_device *dev = qdisc_dev(sch);
101 	struct Qdisc *qdisc;
102 	unsigned int ntx;
103 	__u32 qlen = 0;
104 
105 	sch->q.qlen = 0;
106 	memset(&sch->bstats, 0, sizeof(sch->bstats));
107 	memset(&sch->qstats, 0, sizeof(sch->qstats));
108 
109 	/* MQ supports lockless qdiscs. However, statistics accounting needs
110 	 * to account for all, none, or a mix of locked and unlocked child
111 	 * qdiscs. Percpu stats are added to counters in-band and locking
112 	 * qdisc totals are added at end.
113 	 */
114 	for (ntx = 0; ntx < dev->num_tx_queues; ntx++) {
115 		qdisc = netdev_get_tx_queue(dev, ntx)->qdisc_sleeping;
116 		spin_lock_bh(qdisc_lock(qdisc));
117 
118 		if (qdisc_is_percpu_stats(qdisc)) {
119 			qlen = qdisc_qlen_sum(qdisc);
120 			__gnet_stats_copy_basic(NULL, &sch->bstats,
121 						qdisc->cpu_bstats,
122 						&qdisc->bstats);
123 			__gnet_stats_copy_queue(&sch->qstats,
124 						qdisc->cpu_qstats,
125 						&qdisc->qstats, qlen);
126 		} else {
127 			sch->q.qlen		+= qdisc->q.qlen;
128 			sch->bstats.bytes	+= qdisc->bstats.bytes;
129 			sch->bstats.packets	+= qdisc->bstats.packets;
130 			sch->qstats.backlog	+= qdisc->qstats.backlog;
131 			sch->qstats.drops	+= qdisc->qstats.drops;
132 			sch->qstats.requeues	+= qdisc->qstats.requeues;
133 			sch->qstats.overlimits	+= qdisc->qstats.overlimits;
134 		}
135 
136 		spin_unlock_bh(qdisc_lock(qdisc));
137 	}
138 
139 	return 0;
140 }
141 
142 static struct netdev_queue *mq_queue_get(struct Qdisc *sch, unsigned long cl)
143 {
144 	struct net_device *dev = qdisc_dev(sch);
145 	unsigned long ntx = cl - 1;
146 
147 	if (ntx >= dev->num_tx_queues)
148 		return NULL;
149 	return netdev_get_tx_queue(dev, ntx);
150 }
151 
152 static struct netdev_queue *mq_select_queue(struct Qdisc *sch,
153 					    struct tcmsg *tcm)
154 {
155 	return mq_queue_get(sch, TC_H_MIN(tcm->tcm_parent));
156 }
157 
158 static int mq_graft(struct Qdisc *sch, unsigned long cl, struct Qdisc *new,
159 		    struct Qdisc **old, struct netlink_ext_ack *extack)
160 {
161 	struct netdev_queue *dev_queue = mq_queue_get(sch, cl);
162 	struct net_device *dev = qdisc_dev(sch);
163 
164 	if (dev->flags & IFF_UP)
165 		dev_deactivate(dev);
166 
167 	*old = dev_graft_qdisc(dev_queue, new);
168 	if (new)
169 		new->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
170 	if (dev->flags & IFF_UP)
171 		dev_activate(dev);
172 	return 0;
173 }
174 
175 static struct Qdisc *mq_leaf(struct Qdisc *sch, unsigned long cl)
176 {
177 	struct netdev_queue *dev_queue = mq_queue_get(sch, cl);
178 
179 	return dev_queue->qdisc_sleeping;
180 }
181 
182 static unsigned long mq_find(struct Qdisc *sch, u32 classid)
183 {
184 	unsigned int ntx = TC_H_MIN(classid);
185 
186 	if (!mq_queue_get(sch, ntx))
187 		return 0;
188 	return ntx;
189 }
190 
191 static int mq_dump_class(struct Qdisc *sch, unsigned long cl,
192 			 struct sk_buff *skb, struct tcmsg *tcm)
193 {
194 	struct netdev_queue *dev_queue = mq_queue_get(sch, cl);
195 
196 	tcm->tcm_parent = TC_H_ROOT;
197 	tcm->tcm_handle |= TC_H_MIN(cl);
198 	tcm->tcm_info = dev_queue->qdisc_sleeping->handle;
199 	return 0;
200 }
201 
202 static int mq_dump_class_stats(struct Qdisc *sch, unsigned long cl,
203 			       struct gnet_dump *d)
204 {
205 	struct netdev_queue *dev_queue = mq_queue_get(sch, cl);
206 
207 	sch = dev_queue->qdisc_sleeping;
208 	if (gnet_stats_copy_basic(&sch->running, d, NULL, &sch->bstats) < 0 ||
209 	    gnet_stats_copy_queue(d, NULL, &sch->qstats, sch->q.qlen) < 0)
210 		return -1;
211 	return 0;
212 }
213 
214 static void mq_walk(struct Qdisc *sch, struct qdisc_walker *arg)
215 {
216 	struct net_device *dev = qdisc_dev(sch);
217 	unsigned int ntx;
218 
219 	if (arg->stop)
220 		return;
221 
222 	arg->count = arg->skip;
223 	for (ntx = arg->skip; ntx < dev->num_tx_queues; ntx++) {
224 		if (arg->fn(sch, ntx + 1, arg) < 0) {
225 			arg->stop = 1;
226 			break;
227 		}
228 		arg->count++;
229 	}
230 }
231 
232 static const struct Qdisc_class_ops mq_class_ops = {
233 	.select_queue	= mq_select_queue,
234 	.graft		= mq_graft,
235 	.leaf		= mq_leaf,
236 	.find		= mq_find,
237 	.walk		= mq_walk,
238 	.dump		= mq_dump_class,
239 	.dump_stats	= mq_dump_class_stats,
240 };
241 
242 struct Qdisc_ops mq_qdisc_ops __read_mostly = {
243 	.cl_ops		= &mq_class_ops,
244 	.id		= "mq",
245 	.priv_size	= sizeof(struct mq_sched),
246 	.init		= mq_init,
247 	.destroy	= mq_destroy,
248 	.attach		= mq_attach,
249 	.dump		= mq_dump,
250 	.owner		= THIS_MODULE,
251 };
252