1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Copyright (c) 2008, Intel Corporation.
4  *
5  * Author: Alexander Duyck <alexander.h.duyck@intel.com>
6  */
7 
8 #include <linux/module.h>
9 #include <linux/slab.h>
10 #include <linux/types.h>
11 #include <linux/kernel.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_sched.h>
17 #include <net/pkt_cls.h>
18 
19 struct multiq_sched_data {
20 	u16 bands;
21 	u16 max_bands;
22 	u16 curband;
23 	struct tcf_proto __rcu *filter_list;
24 	struct tcf_block *block;
25 	struct Qdisc **queues;
26 };
27 
28 
29 static struct Qdisc *
multiq_classify(struct sk_buff * skb,struct Qdisc * sch,int * qerr)30 multiq_classify(struct sk_buff *skb, struct Qdisc *sch, int *qerr)
31 {
32 	struct multiq_sched_data *q = qdisc_priv(sch);
33 	u32 band;
34 	struct tcf_result res;
35 	struct tcf_proto *fl = rcu_dereference_bh(q->filter_list);
36 	int err;
37 
38 	*qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
39 	err = tcf_classify(skb, NULL, fl, &res, false);
40 #ifdef CONFIG_NET_CLS_ACT
41 	switch (err) {
42 	case TC_ACT_STOLEN:
43 	case TC_ACT_QUEUED:
44 	case TC_ACT_TRAP:
45 		*qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
46 		fallthrough;
47 	case TC_ACT_SHOT:
48 		return NULL;
49 	}
50 #endif
51 	band = skb_get_queue_mapping(skb);
52 
53 	if (band >= q->bands)
54 		return q->queues[0];
55 
56 	return q->queues[band];
57 }
58 
59 static int
multiq_enqueue(struct sk_buff * skb,struct Qdisc * sch,struct sk_buff ** to_free)60 multiq_enqueue(struct sk_buff *skb, struct Qdisc *sch,
61 	       struct sk_buff **to_free)
62 {
63 	struct Qdisc *qdisc;
64 	int ret;
65 
66 	qdisc = multiq_classify(skb, sch, &ret);
67 #ifdef CONFIG_NET_CLS_ACT
68 	if (qdisc == NULL) {
69 
70 		if (ret & __NET_XMIT_BYPASS)
71 			qdisc_qstats_drop(sch);
72 		__qdisc_drop(skb, to_free);
73 		return ret;
74 	}
75 #endif
76 
77 	ret = qdisc_enqueue(skb, qdisc, to_free);
78 	if (ret == NET_XMIT_SUCCESS) {
79 		sch->q.qlen++;
80 		return NET_XMIT_SUCCESS;
81 	}
82 	if (net_xmit_drop_count(ret))
83 		qdisc_qstats_drop(sch);
84 	return ret;
85 }
86 
multiq_dequeue(struct Qdisc * sch)87 static struct sk_buff *multiq_dequeue(struct Qdisc *sch)
88 {
89 	struct multiq_sched_data *q = qdisc_priv(sch);
90 	struct Qdisc *qdisc;
91 	struct sk_buff *skb;
92 	int band;
93 
94 	for (band = 0; band < q->bands; band++) {
95 		/* cycle through bands to ensure fairness */
96 		q->curband++;
97 		if (q->curband >= q->bands)
98 			q->curband = 0;
99 
100 		/* Check that target subqueue is available before
101 		 * pulling an skb to avoid head-of-line blocking.
102 		 */
103 		if (!netif_xmit_stopped(
104 		    netdev_get_tx_queue(qdisc_dev(sch), q->curband))) {
105 			qdisc = q->queues[q->curband];
106 			skb = qdisc->dequeue(qdisc);
107 			if (skb) {
108 				qdisc_bstats_update(sch, skb);
109 				sch->q.qlen--;
110 				return skb;
111 			}
112 		}
113 	}
114 	return NULL;
115 
116 }
117 
multiq_peek(struct Qdisc * sch)118 static struct sk_buff *multiq_peek(struct Qdisc *sch)
119 {
120 	struct multiq_sched_data *q = qdisc_priv(sch);
121 	unsigned int curband = q->curband;
122 	struct Qdisc *qdisc;
123 	struct sk_buff *skb;
124 	int band;
125 
126 	for (band = 0; band < q->bands; band++) {
127 		/* cycle through bands to ensure fairness */
128 		curband++;
129 		if (curband >= q->bands)
130 			curband = 0;
131 
132 		/* Check that target subqueue is available before
133 		 * pulling an skb to avoid head-of-line blocking.
134 		 */
135 		if (!netif_xmit_stopped(
136 		    netdev_get_tx_queue(qdisc_dev(sch), curband))) {
137 			qdisc = q->queues[curband];
138 			skb = qdisc->ops->peek(qdisc);
139 			if (skb)
140 				return skb;
141 		}
142 	}
143 	return NULL;
144 
145 }
146 
147 static void
multiq_reset(struct Qdisc * sch)148 multiq_reset(struct Qdisc *sch)
149 {
150 	u16 band;
151 	struct multiq_sched_data *q = qdisc_priv(sch);
152 
153 	for (band = 0; band < q->bands; band++)
154 		qdisc_reset(q->queues[band]);
155 	q->curband = 0;
156 }
157 
158 static void
multiq_destroy(struct Qdisc * sch)159 multiq_destroy(struct Qdisc *sch)
160 {
161 	int band;
162 	struct multiq_sched_data *q = qdisc_priv(sch);
163 
164 	tcf_block_put(q->block);
165 	for (band = 0; band < q->bands; band++)
166 		qdisc_put(q->queues[band]);
167 
168 	kfree(q->queues);
169 }
170 
multiq_tune(struct Qdisc * sch,struct nlattr * opt,struct netlink_ext_ack * extack)171 static int multiq_tune(struct Qdisc *sch, struct nlattr *opt,
172 		       struct netlink_ext_ack *extack)
173 {
174 	struct multiq_sched_data *q = qdisc_priv(sch);
175 	struct tc_multiq_qopt *qopt;
176 	struct Qdisc **removed;
177 	int i, n_removed = 0;
178 
179 	if (!netif_is_multiqueue(qdisc_dev(sch)))
180 		return -EOPNOTSUPP;
181 	if (nla_len(opt) < sizeof(*qopt))
182 		return -EINVAL;
183 
184 	qopt = nla_data(opt);
185 
186 	qopt->bands = qdisc_dev(sch)->real_num_tx_queues;
187 
188 	removed = kmalloc(sizeof(*removed) * (q->max_bands - q->bands),
189 			  GFP_KERNEL);
190 	if (!removed)
191 		return -ENOMEM;
192 
193 	sch_tree_lock(sch);
194 	q->bands = qopt->bands;
195 	for (i = q->bands; i < q->max_bands; i++) {
196 		if (q->queues[i] != &noop_qdisc) {
197 			struct Qdisc *child = q->queues[i];
198 
199 			q->queues[i] = &noop_qdisc;
200 			qdisc_purge_queue(child);
201 			removed[n_removed++] = child;
202 		}
203 	}
204 
205 	sch_tree_unlock(sch);
206 
207 	for (i = 0; i < n_removed; i++)
208 		qdisc_put(removed[i]);
209 	kfree(removed);
210 
211 	for (i = 0; i < q->bands; i++) {
212 		if (q->queues[i] == &noop_qdisc) {
213 			struct Qdisc *child, *old;
214 			child = qdisc_create_dflt(sch->dev_queue,
215 						  &pfifo_qdisc_ops,
216 						  TC_H_MAKE(sch->handle,
217 							    i + 1), extack);
218 			if (child) {
219 				sch_tree_lock(sch);
220 				old = q->queues[i];
221 				q->queues[i] = child;
222 				if (child != &noop_qdisc)
223 					qdisc_hash_add(child, true);
224 
225 				if (old != &noop_qdisc)
226 					qdisc_purge_queue(old);
227 				sch_tree_unlock(sch);
228 				qdisc_put(old);
229 			}
230 		}
231 	}
232 	return 0;
233 }
234 
multiq_init(struct Qdisc * sch,struct nlattr * opt,struct netlink_ext_ack * extack)235 static int multiq_init(struct Qdisc *sch, struct nlattr *opt,
236 		       struct netlink_ext_ack *extack)
237 {
238 	struct multiq_sched_data *q = qdisc_priv(sch);
239 	int i, err;
240 
241 	q->queues = NULL;
242 
243 	if (!opt)
244 		return -EINVAL;
245 
246 	err = tcf_block_get(&q->block, &q->filter_list, sch, extack);
247 	if (err)
248 		return err;
249 
250 	q->max_bands = qdisc_dev(sch)->num_tx_queues;
251 
252 	q->queues = kcalloc(q->max_bands, sizeof(struct Qdisc *), GFP_KERNEL);
253 	if (!q->queues)
254 		return -ENOBUFS;
255 	for (i = 0; i < q->max_bands; i++)
256 		q->queues[i] = &noop_qdisc;
257 
258 	return multiq_tune(sch, opt, extack);
259 }
260 
multiq_dump(struct Qdisc * sch,struct sk_buff * skb)261 static int multiq_dump(struct Qdisc *sch, struct sk_buff *skb)
262 {
263 	struct multiq_sched_data *q = qdisc_priv(sch);
264 	unsigned char *b = skb_tail_pointer(skb);
265 	struct tc_multiq_qopt opt;
266 
267 	opt.bands = q->bands;
268 	opt.max_bands = q->max_bands;
269 
270 	if (nla_put(skb, TCA_OPTIONS, sizeof(opt), &opt))
271 		goto nla_put_failure;
272 
273 	return skb->len;
274 
275 nla_put_failure:
276 	nlmsg_trim(skb, b);
277 	return -1;
278 }
279 
multiq_graft(struct Qdisc * sch,unsigned long arg,struct Qdisc * new,struct Qdisc ** old,struct netlink_ext_ack * extack)280 static int multiq_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
281 			struct Qdisc **old, struct netlink_ext_ack *extack)
282 {
283 	struct multiq_sched_data *q = qdisc_priv(sch);
284 	unsigned long band = arg - 1;
285 
286 	if (new == NULL)
287 		new = &noop_qdisc;
288 
289 	*old = qdisc_replace(sch, new, &q->queues[band]);
290 	return 0;
291 }
292 
293 static struct Qdisc *
multiq_leaf(struct Qdisc * sch,unsigned long arg)294 multiq_leaf(struct Qdisc *sch, unsigned long arg)
295 {
296 	struct multiq_sched_data *q = qdisc_priv(sch);
297 	unsigned long band = arg - 1;
298 
299 	return q->queues[band];
300 }
301 
multiq_find(struct Qdisc * sch,u32 classid)302 static unsigned long multiq_find(struct Qdisc *sch, u32 classid)
303 {
304 	struct multiq_sched_data *q = qdisc_priv(sch);
305 	unsigned long band = TC_H_MIN(classid);
306 
307 	if (band - 1 >= q->bands)
308 		return 0;
309 	return band;
310 }
311 
multiq_bind(struct Qdisc * sch,unsigned long parent,u32 classid)312 static unsigned long multiq_bind(struct Qdisc *sch, unsigned long parent,
313 				 u32 classid)
314 {
315 	return multiq_find(sch, classid);
316 }
317 
318 
multiq_unbind(struct Qdisc * q,unsigned long cl)319 static void multiq_unbind(struct Qdisc *q, unsigned long cl)
320 {
321 }
322 
multiq_dump_class(struct Qdisc * sch,unsigned long cl,struct sk_buff * skb,struct tcmsg * tcm)323 static int multiq_dump_class(struct Qdisc *sch, unsigned long cl,
324 			     struct sk_buff *skb, struct tcmsg *tcm)
325 {
326 	struct multiq_sched_data *q = qdisc_priv(sch);
327 
328 	tcm->tcm_handle |= TC_H_MIN(cl);
329 	tcm->tcm_info = q->queues[cl - 1]->handle;
330 	return 0;
331 }
332 
multiq_dump_class_stats(struct Qdisc * sch,unsigned long cl,struct gnet_dump * d)333 static int multiq_dump_class_stats(struct Qdisc *sch, unsigned long cl,
334 				 struct gnet_dump *d)
335 {
336 	struct multiq_sched_data *q = qdisc_priv(sch);
337 	struct Qdisc *cl_q;
338 
339 	cl_q = q->queues[cl - 1];
340 	if (gnet_stats_copy_basic(d, cl_q->cpu_bstats, &cl_q->bstats, true) < 0 ||
341 	    qdisc_qstats_copy(d, cl_q) < 0)
342 		return -1;
343 
344 	return 0;
345 }
346 
multiq_walk(struct Qdisc * sch,struct qdisc_walker * arg)347 static void multiq_walk(struct Qdisc *sch, struct qdisc_walker *arg)
348 {
349 	struct multiq_sched_data *q = qdisc_priv(sch);
350 	int band;
351 
352 	if (arg->stop)
353 		return;
354 
355 	for (band = 0; band < q->bands; band++) {
356 		if (!tc_qdisc_stats_dump(sch, band + 1, arg))
357 			break;
358 	}
359 }
360 
multiq_tcf_block(struct Qdisc * sch,unsigned long cl,struct netlink_ext_ack * extack)361 static struct tcf_block *multiq_tcf_block(struct Qdisc *sch, unsigned long cl,
362 					  struct netlink_ext_ack *extack)
363 {
364 	struct multiq_sched_data *q = qdisc_priv(sch);
365 
366 	if (cl)
367 		return NULL;
368 	return q->block;
369 }
370 
371 static const struct Qdisc_class_ops multiq_class_ops = {
372 	.graft		=	multiq_graft,
373 	.leaf		=	multiq_leaf,
374 	.find		=	multiq_find,
375 	.walk		=	multiq_walk,
376 	.tcf_block	=	multiq_tcf_block,
377 	.bind_tcf	=	multiq_bind,
378 	.unbind_tcf	=	multiq_unbind,
379 	.dump		=	multiq_dump_class,
380 	.dump_stats	=	multiq_dump_class_stats,
381 };
382 
383 static struct Qdisc_ops multiq_qdisc_ops __read_mostly = {
384 	.next		=	NULL,
385 	.cl_ops		=	&multiq_class_ops,
386 	.id		=	"multiq",
387 	.priv_size	=	sizeof(struct multiq_sched_data),
388 	.enqueue	=	multiq_enqueue,
389 	.dequeue	=	multiq_dequeue,
390 	.peek		=	multiq_peek,
391 	.init		=	multiq_init,
392 	.reset		=	multiq_reset,
393 	.destroy	=	multiq_destroy,
394 	.change		=	multiq_tune,
395 	.dump		=	multiq_dump,
396 	.owner		=	THIS_MODULE,
397 };
398 
multiq_module_init(void)399 static int __init multiq_module_init(void)
400 {
401 	return register_qdisc(&multiq_qdisc_ops);
402 }
403 
multiq_module_exit(void)404 static void __exit multiq_module_exit(void)
405 {
406 	unregister_qdisc(&multiq_qdisc_ops);
407 }
408 
409 module_init(multiq_module_init)
410 module_exit(multiq_module_exit)
411 
412 MODULE_LICENSE("GPL");
413