xref: /openbmc/linux/net/sched/sch_mq.c (revision b60a5b8d)
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_cls.h>
20 #include <net/pkt_sched.h>
21 #include <net/sch_generic.h>
22 
23 struct mq_sched {
24 	struct Qdisc		**qdiscs;
25 };
26 
27 static int mq_offload(struct Qdisc *sch, enum tc_mq_command cmd)
28 {
29 	struct net_device *dev = qdisc_dev(sch);
30 	struct tc_mq_qopt_offload opt = {
31 		.command = cmd,
32 		.handle = sch->handle,
33 	};
34 
35 	if (!tc_can_offload(dev) || !dev->netdev_ops->ndo_setup_tc)
36 		return -EOPNOTSUPP;
37 
38 	return dev->netdev_ops->ndo_setup_tc(dev, TC_SETUP_QDISC_MQ, &opt);
39 }
40 
41 static int mq_offload_stats(struct Qdisc *sch)
42 {
43 	struct tc_mq_qopt_offload opt = {
44 		.command = TC_MQ_STATS,
45 		.handle = sch->handle,
46 		.stats = {
47 			.bstats = &sch->bstats,
48 			.qstats = &sch->qstats,
49 		},
50 	};
51 
52 	return qdisc_offload_dump_helper(sch, TC_SETUP_QDISC_MQ, &opt);
53 }
54 
55 static void mq_destroy(struct Qdisc *sch)
56 {
57 	struct net_device *dev = qdisc_dev(sch);
58 	struct mq_sched *priv = qdisc_priv(sch);
59 	unsigned int ntx;
60 
61 	mq_offload(sch, TC_MQ_DESTROY);
62 
63 	if (!priv->qdiscs)
64 		return;
65 	for (ntx = 0; ntx < dev->num_tx_queues && priv->qdiscs[ntx]; ntx++)
66 		qdisc_put(priv->qdiscs[ntx]);
67 	kfree(priv->qdiscs);
68 }
69 
70 static int mq_init(struct Qdisc *sch, struct nlattr *opt,
71 		   struct netlink_ext_ack *extack)
72 {
73 	struct net_device *dev = qdisc_dev(sch);
74 	struct mq_sched *priv = qdisc_priv(sch);
75 	struct netdev_queue *dev_queue;
76 	struct Qdisc *qdisc;
77 	unsigned int ntx;
78 
79 	if (sch->parent != TC_H_ROOT)
80 		return -EOPNOTSUPP;
81 
82 	if (!netif_is_multiqueue(dev))
83 		return -EOPNOTSUPP;
84 
85 	/* pre-allocate qdiscs, attachment can't fail */
86 	priv->qdiscs = kcalloc(dev->num_tx_queues, sizeof(priv->qdiscs[0]),
87 			       GFP_KERNEL);
88 	if (!priv->qdiscs)
89 		return -ENOMEM;
90 
91 	for (ntx = 0; ntx < dev->num_tx_queues; ntx++) {
92 		dev_queue = netdev_get_tx_queue(dev, ntx);
93 		qdisc = qdisc_create_dflt(dev_queue, get_default_qdisc_ops(dev, ntx),
94 					  TC_H_MAKE(TC_H_MAJ(sch->handle),
95 						    TC_H_MIN(ntx + 1)),
96 					  extack);
97 		if (!qdisc)
98 			return -ENOMEM;
99 		priv->qdiscs[ntx] = qdisc;
100 		qdisc->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
101 	}
102 
103 	sch->flags |= TCQ_F_MQROOT;
104 
105 	mq_offload(sch, TC_MQ_CREATE);
106 	return 0;
107 }
108 
109 static void mq_attach(struct Qdisc *sch)
110 {
111 	struct net_device *dev = qdisc_dev(sch);
112 	struct mq_sched *priv = qdisc_priv(sch);
113 	struct Qdisc *qdisc, *old;
114 	unsigned int ntx;
115 
116 	for (ntx = 0; ntx < dev->num_tx_queues; ntx++) {
117 		qdisc = priv->qdiscs[ntx];
118 		old = dev_graft_qdisc(qdisc->dev_queue, qdisc);
119 		if (old)
120 			qdisc_put(old);
121 #ifdef CONFIG_NET_SCHED
122 		if (ntx < dev->real_num_tx_queues)
123 			qdisc_hash_add(qdisc, false);
124 #endif
125 
126 	}
127 	kfree(priv->qdiscs);
128 	priv->qdiscs = NULL;
129 }
130 
131 static int mq_dump(struct Qdisc *sch, struct sk_buff *skb)
132 {
133 	struct net_device *dev = qdisc_dev(sch);
134 	struct Qdisc *qdisc;
135 	unsigned int ntx;
136 	__u32 qlen = 0;
137 
138 	sch->q.qlen = 0;
139 	memset(&sch->bstats, 0, sizeof(sch->bstats));
140 	memset(&sch->qstats, 0, sizeof(sch->qstats));
141 
142 	/* MQ supports lockless qdiscs. However, statistics accounting needs
143 	 * to account for all, none, or a mix of locked and unlocked child
144 	 * qdiscs. Percpu stats are added to counters in-band and locking
145 	 * qdisc totals are added at end.
146 	 */
147 	for (ntx = 0; ntx < dev->num_tx_queues; ntx++) {
148 		qdisc = netdev_get_tx_queue(dev, ntx)->qdisc_sleeping;
149 		spin_lock_bh(qdisc_lock(qdisc));
150 
151 		if (qdisc_is_percpu_stats(qdisc)) {
152 			qlen = qdisc_qlen_sum(qdisc);
153 			__gnet_stats_copy_basic(NULL, &sch->bstats,
154 						qdisc->cpu_bstats,
155 						&qdisc->bstats);
156 			__gnet_stats_copy_queue(&sch->qstats,
157 						qdisc->cpu_qstats,
158 						&qdisc->qstats, qlen);
159 		} else {
160 			sch->q.qlen		+= qdisc->q.qlen;
161 			sch->bstats.bytes	+= qdisc->bstats.bytes;
162 			sch->bstats.packets	+= qdisc->bstats.packets;
163 			sch->qstats.qlen	+= qdisc->qstats.qlen;
164 			sch->qstats.backlog	+= qdisc->qstats.backlog;
165 			sch->qstats.drops	+= qdisc->qstats.drops;
166 			sch->qstats.requeues	+= qdisc->qstats.requeues;
167 			sch->qstats.overlimits	+= qdisc->qstats.overlimits;
168 		}
169 
170 		spin_unlock_bh(qdisc_lock(qdisc));
171 	}
172 
173 	return mq_offload_stats(sch);
174 }
175 
176 static struct netdev_queue *mq_queue_get(struct Qdisc *sch, unsigned long cl)
177 {
178 	struct net_device *dev = qdisc_dev(sch);
179 	unsigned long ntx = cl - 1;
180 
181 	if (ntx >= dev->num_tx_queues)
182 		return NULL;
183 	return netdev_get_tx_queue(dev, ntx);
184 }
185 
186 static struct netdev_queue *mq_select_queue(struct Qdisc *sch,
187 					    struct tcmsg *tcm)
188 {
189 	return mq_queue_get(sch, TC_H_MIN(tcm->tcm_parent));
190 }
191 
192 static int mq_graft(struct Qdisc *sch, unsigned long cl, struct Qdisc *new,
193 		    struct Qdisc **old, struct netlink_ext_ack *extack)
194 {
195 	struct netdev_queue *dev_queue = mq_queue_get(sch, cl);
196 	struct tc_mq_qopt_offload graft_offload;
197 	struct net_device *dev = qdisc_dev(sch);
198 
199 	if (dev->flags & IFF_UP)
200 		dev_deactivate(dev);
201 
202 	*old = dev_graft_qdisc(dev_queue, new);
203 	if (new)
204 		new->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
205 	if (dev->flags & IFF_UP)
206 		dev_activate(dev);
207 
208 	graft_offload.handle = sch->handle;
209 	graft_offload.graft_params.queue = cl - 1;
210 	graft_offload.graft_params.child_handle = new ? new->handle : 0;
211 	graft_offload.command = TC_MQ_GRAFT;
212 
213 	qdisc_offload_graft_helper(qdisc_dev(sch), sch, new, *old,
214 				   TC_SETUP_QDISC_MQ, &graft_offload, extack);
215 	return 0;
216 }
217 
218 static struct Qdisc *mq_leaf(struct Qdisc *sch, unsigned long cl)
219 {
220 	struct netdev_queue *dev_queue = mq_queue_get(sch, cl);
221 
222 	return dev_queue->qdisc_sleeping;
223 }
224 
225 static unsigned long mq_find(struct Qdisc *sch, u32 classid)
226 {
227 	unsigned int ntx = TC_H_MIN(classid);
228 
229 	if (!mq_queue_get(sch, ntx))
230 		return 0;
231 	return ntx;
232 }
233 
234 static int mq_dump_class(struct Qdisc *sch, unsigned long cl,
235 			 struct sk_buff *skb, struct tcmsg *tcm)
236 {
237 	struct netdev_queue *dev_queue = mq_queue_get(sch, cl);
238 
239 	tcm->tcm_parent = TC_H_ROOT;
240 	tcm->tcm_handle |= TC_H_MIN(cl);
241 	tcm->tcm_info = dev_queue->qdisc_sleeping->handle;
242 	return 0;
243 }
244 
245 static int mq_dump_class_stats(struct Qdisc *sch, unsigned long cl,
246 			       struct gnet_dump *d)
247 {
248 	struct netdev_queue *dev_queue = mq_queue_get(sch, cl);
249 
250 	sch = dev_queue->qdisc_sleeping;
251 	if (gnet_stats_copy_basic(&sch->running, d, NULL, &sch->bstats) < 0 ||
252 	    gnet_stats_copy_queue(d, NULL, &sch->qstats, sch->q.qlen) < 0)
253 		return -1;
254 	return 0;
255 }
256 
257 static void mq_walk(struct Qdisc *sch, struct qdisc_walker *arg)
258 {
259 	struct net_device *dev = qdisc_dev(sch);
260 	unsigned int ntx;
261 
262 	if (arg->stop)
263 		return;
264 
265 	arg->count = arg->skip;
266 	for (ntx = arg->skip; ntx < dev->num_tx_queues; ntx++) {
267 		if (arg->fn(sch, ntx + 1, arg) < 0) {
268 			arg->stop = 1;
269 			break;
270 		}
271 		arg->count++;
272 	}
273 }
274 
275 static const struct Qdisc_class_ops mq_class_ops = {
276 	.select_queue	= mq_select_queue,
277 	.graft		= mq_graft,
278 	.leaf		= mq_leaf,
279 	.find		= mq_find,
280 	.walk		= mq_walk,
281 	.dump		= mq_dump_class,
282 	.dump_stats	= mq_dump_class_stats,
283 };
284 
285 struct Qdisc_ops mq_qdisc_ops __read_mostly = {
286 	.cl_ops		= &mq_class_ops,
287 	.id		= "mq",
288 	.priv_size	= sizeof(struct mq_sched),
289 	.init		= mq_init,
290 	.destroy	= mq_destroy,
291 	.attach		= mq_attach,
292 	.dump		= mq_dump,
293 	.owner		= THIS_MODULE,
294 };
295