1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * net/sched/sch_mq.c Classful multiqueue dummy scheduler
4 *
5 * Copyright (c) 2009 Patrick McHardy <kaber@trash.net>
6 */
7
8#include <linux/types.h>
9#include <linux/slab.h>
10#include <linux/kernel.h>
11#include <linux/export.h>
12#include <linux/string.h>
13#include <linux/errno.h>
14#include <linux/skbuff.h>
15#include <net/netlink.h>
16#include <net/pkt_cls.h>
17#include <net/pkt_sched.h>
18#include <net/sch_generic.h>
19
20struct mq_sched {
21 struct Qdisc **qdiscs;
22};
23
24static int mq_offload(struct Qdisc *sch, enum tc_mq_command cmd)
25{
26 struct net_device *dev = qdisc_dev(qdisc: sch);
27 struct tc_mq_qopt_offload opt = {
28 .command = cmd,
29 .handle = sch->handle,
30 };
31
32 if (!tc_can_offload(dev) || !dev->netdev_ops->ndo_setup_tc)
33 return -EOPNOTSUPP;
34
35 return dev->netdev_ops->ndo_setup_tc(dev, TC_SETUP_QDISC_MQ, &opt);
36}
37
38static int mq_offload_stats(struct Qdisc *sch)
39{
40 struct tc_mq_qopt_offload opt = {
41 .command = TC_MQ_STATS,
42 .handle = sch->handle,
43 .stats = {
44 .bstats = &sch->bstats,
45 .qstats = &sch->qstats,
46 },
47 };
48
49 return qdisc_offload_dump_helper(q: sch, type: TC_SETUP_QDISC_MQ, type_data: &opt);
50}
51
52static void mq_destroy(struct Qdisc *sch)
53{
54 struct net_device *dev = qdisc_dev(qdisc: sch);
55 struct mq_sched *priv = qdisc_priv(sch);
56 unsigned int ntx;
57
58 mq_offload(sch, cmd: TC_MQ_DESTROY);
59
60 if (!priv->qdiscs)
61 return;
62 for (ntx = 0; ntx < dev->num_tx_queues && priv->qdiscs[ntx]; ntx++)
63 qdisc_put(qdisc: priv->qdiscs[ntx]);
64 kfree(objp: priv->qdiscs);
65}
66
67static int mq_init(struct Qdisc *sch, struct nlattr *opt,
68 struct netlink_ext_ack *extack)
69{
70 struct net_device *dev = qdisc_dev(qdisc: sch);
71 struct mq_sched *priv = qdisc_priv(sch);
72 struct netdev_queue *dev_queue;
73 struct Qdisc *qdisc;
74 unsigned int ntx;
75
76 if (sch->parent != TC_H_ROOT)
77 return -EOPNOTSUPP;
78
79 if (!netif_is_multiqueue(dev))
80 return -EOPNOTSUPP;
81
82 /* pre-allocate qdiscs, attachment can't fail */
83 priv->qdiscs = kcalloc(dev->num_tx_queues, sizeof(priv->qdiscs[0]),
84 GFP_KERNEL);
85 if (!priv->qdiscs)
86 return -ENOMEM;
87
88 for (ntx = 0; ntx < dev->num_tx_queues; ntx++) {
89 dev_queue = netdev_get_tx_queue(dev, index: ntx);
90 qdisc = qdisc_create_dflt(dev_queue, ops: get_default_qdisc_ops(dev, ntx),
91 TC_H_MAKE(TC_H_MAJ(sch->handle),
92 TC_H_MIN(ntx + 1)),
93 extack);
94 if (!qdisc)
95 return -ENOMEM;
96 priv->qdiscs[ntx] = qdisc;
97 qdisc->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
98 }
99
100 sch->flags |= TCQ_F_MQROOT;
101
102 mq_offload(sch, cmd: TC_MQ_CREATE);
103 return 0;
104}
105
106static void mq_attach(struct Qdisc *sch)
107{
108 struct net_device *dev = qdisc_dev(qdisc: sch);
109 struct mq_sched *priv = qdisc_priv(sch);
110 struct Qdisc *qdisc, *old;
111 unsigned int ntx;
112
113 for (ntx = 0; ntx < dev->num_tx_queues; ntx++) {
114 qdisc = priv->qdiscs[ntx];
115 old = dev_graft_qdisc(dev_queue: qdisc->dev_queue, qdisc);
116 if (old)
117 qdisc_put(qdisc: old);
118#ifdef CONFIG_NET_SCHED
119 if (ntx < dev->real_num_tx_queues)
120 qdisc_hash_add(q: qdisc, invisible: false);
121#endif
122
123 }
124 kfree(objp: priv->qdiscs);
125 priv->qdiscs = NULL;
126}
127
128static int mq_dump(struct Qdisc *sch, struct sk_buff *skb)
129{
130 struct net_device *dev = qdisc_dev(qdisc: sch);
131 struct Qdisc *qdisc;
132 unsigned int ntx;
133
134 sch->q.qlen = 0;
135 gnet_stats_basic_sync_init(b: &sch->bstats);
136 memset(s: &sch->qstats, c: 0, n: sizeof(sch->qstats));
137
138 /* MQ supports lockless qdiscs. However, statistics accounting needs
139 * to account for all, none, or a mix of locked and unlocked child
140 * qdiscs. Percpu stats are added to counters in-band and locking
141 * qdisc totals are added at end.
142 */
143 for (ntx = 0; ntx < dev->num_tx_queues; ntx++) {
144 qdisc = rtnl_dereference(netdev_get_tx_queue(dev, ntx)->qdisc_sleeping);
145 spin_lock_bh(lock: qdisc_lock(qdisc));
146
147 gnet_stats_add_basic(bstats: &sch->bstats, cpu: qdisc->cpu_bstats,
148 b: &qdisc->bstats, running: false);
149 gnet_stats_add_queue(qstats: &sch->qstats, cpu_q: qdisc->cpu_qstats,
150 q: &qdisc->qstats);
151 sch->q.qlen += qdisc_qlen(q: qdisc);
152
153 spin_unlock_bh(lock: qdisc_lock(qdisc));
154 }
155
156 return mq_offload_stats(sch);
157}
158
159static struct netdev_queue *mq_queue_get(struct Qdisc *sch, unsigned long cl)
160{
161 struct net_device *dev = qdisc_dev(qdisc: sch);
162 unsigned long ntx = cl - 1;
163
164 if (ntx >= dev->num_tx_queues)
165 return NULL;
166 return netdev_get_tx_queue(dev, index: ntx);
167}
168
169static struct netdev_queue *mq_select_queue(struct Qdisc *sch,
170 struct tcmsg *tcm)
171{
172 return mq_queue_get(sch, TC_H_MIN(tcm->tcm_parent));
173}
174
175static int mq_graft(struct Qdisc *sch, unsigned long cl, struct Qdisc *new,
176 struct Qdisc **old, struct netlink_ext_ack *extack)
177{
178 struct netdev_queue *dev_queue = mq_queue_get(sch, cl);
179 struct tc_mq_qopt_offload graft_offload;
180 struct net_device *dev = qdisc_dev(qdisc: sch);
181
182 if (dev->flags & IFF_UP)
183 dev_deactivate(dev);
184
185 *old = dev_graft_qdisc(dev_queue, qdisc: new);
186 if (new)
187 new->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
188 if (dev->flags & IFF_UP)
189 dev_activate(dev);
190
191 graft_offload.handle = sch->handle;
192 graft_offload.graft_params.queue = cl - 1;
193 graft_offload.graft_params.child_handle = new ? new->handle : 0;
194 graft_offload.command = TC_MQ_GRAFT;
195
196 qdisc_offload_graft_helper(dev: qdisc_dev(qdisc: sch), sch, new, old: *old,
197 type: TC_SETUP_QDISC_MQ, type_data: &graft_offload, extack);
198 return 0;
199}
200
201static struct Qdisc *mq_leaf(struct Qdisc *sch, unsigned long cl)
202{
203 struct netdev_queue *dev_queue = mq_queue_get(sch, cl);
204
205 return rtnl_dereference(dev_queue->qdisc_sleeping);
206}
207
208static unsigned long mq_find(struct Qdisc *sch, u32 classid)
209{
210 unsigned int ntx = TC_H_MIN(classid);
211
212 if (!mq_queue_get(sch, cl: ntx))
213 return 0;
214 return ntx;
215}
216
217static int mq_dump_class(struct Qdisc *sch, unsigned long cl,
218 struct sk_buff *skb, struct tcmsg *tcm)
219{
220 struct netdev_queue *dev_queue = mq_queue_get(sch, cl);
221
222 tcm->tcm_parent = TC_H_ROOT;
223 tcm->tcm_handle |= TC_H_MIN(cl);
224 tcm->tcm_info = rtnl_dereference(dev_queue->qdisc_sleeping)->handle;
225 return 0;
226}
227
228static int mq_dump_class_stats(struct Qdisc *sch, unsigned long cl,
229 struct gnet_dump *d)
230{
231 struct netdev_queue *dev_queue = mq_queue_get(sch, cl);
232
233 sch = rtnl_dereference(dev_queue->qdisc_sleeping);
234 if (gnet_stats_copy_basic(d, cpu: sch->cpu_bstats, b: &sch->bstats, running: true) < 0 ||
235 qdisc_qstats_copy(d, sch) < 0)
236 return -1;
237 return 0;
238}
239
240static void mq_walk(struct Qdisc *sch, struct qdisc_walker *arg)
241{
242 struct net_device *dev = qdisc_dev(qdisc: sch);
243 unsigned int ntx;
244
245 if (arg->stop)
246 return;
247
248 arg->count = arg->skip;
249 for (ntx = arg->skip; ntx < dev->num_tx_queues; ntx++) {
250 if (!tc_qdisc_stats_dump(sch, cl: ntx + 1, arg))
251 break;
252 }
253}
254
255static const struct Qdisc_class_ops mq_class_ops = {
256 .select_queue = mq_select_queue,
257 .graft = mq_graft,
258 .leaf = mq_leaf,
259 .find = mq_find,
260 .walk = mq_walk,
261 .dump = mq_dump_class,
262 .dump_stats = mq_dump_class_stats,
263};
264
265struct Qdisc_ops mq_qdisc_ops __read_mostly = {
266 .cl_ops = &mq_class_ops,
267 .id = "mq",
268 .priv_size = sizeof(struct mq_sched),
269 .init = mq_init,
270 .destroy = mq_destroy,
271 .attach = mq_attach,
272 .change_real_num_tx = mq_change_real_num_tx,
273 .dump = mq_dump,
274 .owner = THIS_MODULE,
275};
276