xref: /openbmc/linux/net/sched/sch_multiq.c (revision 01a6e126)
1 /*
2  * Copyright (c) 2008, Intel Corporation.
3  *
4  * This program is free software; you can redistribute it and/or modify it
5  * under the terms and conditions of the GNU General Public License,
6  * version 2, as published by the Free Software Foundation.
7  *
8  * This program is distributed in the hope it will be useful, but WITHOUT
9  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
10  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
11  * more details.
12  *
13  * You should have received a copy of the GNU General Public License along with
14  * this program; if not, see <http://www.gnu.org/licenses/>.
15  *
16  * Author: Alexander Duyck <alexander.h.duyck@intel.com>
17  */
18 
19 #include <linux/module.h>
20 #include <linux/slab.h>
21 #include <linux/types.h>
22 #include <linux/kernel.h>
23 #include <linux/string.h>
24 #include <linux/errno.h>
25 #include <linux/skbuff.h>
26 #include <net/netlink.h>
27 #include <net/pkt_sched.h>
28 #include <net/pkt_cls.h>
29 
30 struct multiq_sched_data {
31 	u16 bands;
32 	u16 max_bands;
33 	u16 curband;
34 	struct tcf_proto __rcu *filter_list;
35 	struct tcf_block *block;
36 	struct Qdisc **queues;
37 };
38 
39 
40 static struct Qdisc *
41 multiq_classify(struct sk_buff *skb, struct Qdisc *sch, int *qerr)
42 {
43 	struct multiq_sched_data *q = qdisc_priv(sch);
44 	u32 band;
45 	struct tcf_result res;
46 	struct tcf_proto *fl = rcu_dereference_bh(q->filter_list);
47 	int err;
48 
49 	*qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
50 	err = tcf_classify(skb, fl, &res, false);
51 #ifdef CONFIG_NET_CLS_ACT
52 	switch (err) {
53 	case TC_ACT_STOLEN:
54 	case TC_ACT_QUEUED:
55 	case TC_ACT_TRAP:
56 		*qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
57 		/* fall through */
58 	case TC_ACT_SHOT:
59 		return NULL;
60 	}
61 #endif
62 	band = skb_get_queue_mapping(skb);
63 
64 	if (band >= q->bands)
65 		return q->queues[0];
66 
67 	return q->queues[band];
68 }
69 
70 static int
71 multiq_enqueue(struct sk_buff *skb, struct Qdisc *sch,
72 	       struct sk_buff **to_free)
73 {
74 	struct Qdisc *qdisc;
75 	int ret;
76 
77 	qdisc = multiq_classify(skb, sch, &ret);
78 #ifdef CONFIG_NET_CLS_ACT
79 	if (qdisc == NULL) {
80 
81 		if (ret & __NET_XMIT_BYPASS)
82 			qdisc_qstats_drop(sch);
83 		__qdisc_drop(skb, to_free);
84 		return ret;
85 	}
86 #endif
87 
88 	ret = qdisc_enqueue(skb, qdisc, to_free);
89 	if (ret == NET_XMIT_SUCCESS) {
90 		sch->q.qlen++;
91 		return NET_XMIT_SUCCESS;
92 	}
93 	if (net_xmit_drop_count(ret))
94 		qdisc_qstats_drop(sch);
95 	return ret;
96 }
97 
98 static struct sk_buff *multiq_dequeue(struct Qdisc *sch)
99 {
100 	struct multiq_sched_data *q = qdisc_priv(sch);
101 	struct Qdisc *qdisc;
102 	struct sk_buff *skb;
103 	int band;
104 
105 	for (band = 0; band < q->bands; band++) {
106 		/* cycle through bands to ensure fairness */
107 		q->curband++;
108 		if (q->curband >= q->bands)
109 			q->curband = 0;
110 
111 		/* Check that target subqueue is available before
112 		 * pulling an skb to avoid head-of-line blocking.
113 		 */
114 		if (!netif_xmit_stopped(
115 		    netdev_get_tx_queue(qdisc_dev(sch), q->curband))) {
116 			qdisc = q->queues[q->curband];
117 			skb = qdisc->dequeue(qdisc);
118 			if (skb) {
119 				qdisc_bstats_update(sch, skb);
120 				sch->q.qlen--;
121 				return skb;
122 			}
123 		}
124 	}
125 	return NULL;
126 
127 }
128 
129 static struct sk_buff *multiq_peek(struct Qdisc *sch)
130 {
131 	struct multiq_sched_data *q = qdisc_priv(sch);
132 	unsigned int curband = q->curband;
133 	struct Qdisc *qdisc;
134 	struct sk_buff *skb;
135 	int band;
136 
137 	for (band = 0; band < q->bands; band++) {
138 		/* cycle through bands to ensure fairness */
139 		curband++;
140 		if (curband >= q->bands)
141 			curband = 0;
142 
143 		/* Check that target subqueue is available before
144 		 * pulling an skb to avoid head-of-line blocking.
145 		 */
146 		if (!netif_xmit_stopped(
147 		    netdev_get_tx_queue(qdisc_dev(sch), curband))) {
148 			qdisc = q->queues[curband];
149 			skb = qdisc->ops->peek(qdisc);
150 			if (skb)
151 				return skb;
152 		}
153 	}
154 	return NULL;
155 
156 }
157 
158 static void
159 multiq_reset(struct Qdisc *sch)
160 {
161 	u16 band;
162 	struct multiq_sched_data *q = qdisc_priv(sch);
163 
164 	for (band = 0; band < q->bands; band++)
165 		qdisc_reset(q->queues[band]);
166 	sch->q.qlen = 0;
167 	q->curband = 0;
168 }
169 
170 static void
171 multiq_destroy(struct Qdisc *sch)
172 {
173 	int band;
174 	struct multiq_sched_data *q = qdisc_priv(sch);
175 
176 	tcf_block_put(q->block);
177 	for (band = 0; band < q->bands; band++)
178 		qdisc_destroy(q->queues[band]);
179 
180 	kfree(q->queues);
181 }
182 
183 static int multiq_tune(struct Qdisc *sch, struct nlattr *opt,
184 		       struct netlink_ext_ack *extack)
185 {
186 	struct multiq_sched_data *q = qdisc_priv(sch);
187 	struct tc_multiq_qopt *qopt;
188 	int i;
189 
190 	if (!netif_is_multiqueue(qdisc_dev(sch)))
191 		return -EOPNOTSUPP;
192 	if (nla_len(opt) < sizeof(*qopt))
193 		return -EINVAL;
194 
195 	qopt = nla_data(opt);
196 
197 	qopt->bands = qdisc_dev(sch)->real_num_tx_queues;
198 
199 	sch_tree_lock(sch);
200 	q->bands = qopt->bands;
201 	for (i = q->bands; i < q->max_bands; i++) {
202 		if (q->queues[i] != &noop_qdisc) {
203 			struct Qdisc *child = q->queues[i];
204 			q->queues[i] = &noop_qdisc;
205 			qdisc_tree_reduce_backlog(child, child->q.qlen,
206 						  child->qstats.backlog);
207 			qdisc_destroy(child);
208 		}
209 	}
210 
211 	sch_tree_unlock(sch);
212 
213 	for (i = 0; i < q->bands; i++) {
214 		if (q->queues[i] == &noop_qdisc) {
215 			struct Qdisc *child, *old;
216 			child = qdisc_create_dflt(sch->dev_queue,
217 						  &pfifo_qdisc_ops,
218 						  TC_H_MAKE(sch->handle,
219 							    i + 1), extack);
220 			if (child) {
221 				sch_tree_lock(sch);
222 				old = q->queues[i];
223 				q->queues[i] = child;
224 				if (child != &noop_qdisc)
225 					qdisc_hash_add(child, true);
226 
227 				if (old != &noop_qdisc) {
228 					qdisc_tree_reduce_backlog(old,
229 								  old->q.qlen,
230 								  old->qstats.backlog);
231 					qdisc_destroy(old);
232 				}
233 				sch_tree_unlock(sch);
234 			}
235 		}
236 	}
237 	return 0;
238 }
239 
240 static int multiq_init(struct Qdisc *sch, struct nlattr *opt,
241 		       struct netlink_ext_ack *extack)
242 {
243 	struct multiq_sched_data *q = qdisc_priv(sch);
244 	int i, err;
245 
246 	q->queues = NULL;
247 
248 	if (!opt)
249 		return -EINVAL;
250 
251 	err = tcf_block_get(&q->block, &q->filter_list, sch, extack);
252 	if (err)
253 		return err;
254 
255 	q->max_bands = qdisc_dev(sch)->num_tx_queues;
256 
257 	q->queues = kcalloc(q->max_bands, sizeof(struct Qdisc *), GFP_KERNEL);
258 	if (!q->queues)
259 		return -ENOBUFS;
260 	for (i = 0; i < q->max_bands; i++)
261 		q->queues[i] = &noop_qdisc;
262 
263 	return multiq_tune(sch, opt, extack);
264 }
265 
266 static int multiq_dump(struct Qdisc *sch, struct sk_buff *skb)
267 {
268 	struct multiq_sched_data *q = qdisc_priv(sch);
269 	unsigned char *b = skb_tail_pointer(skb);
270 	struct tc_multiq_qopt opt;
271 
272 	opt.bands = q->bands;
273 	opt.max_bands = q->max_bands;
274 
275 	if (nla_put(skb, TCA_OPTIONS, sizeof(opt), &opt))
276 		goto nla_put_failure;
277 
278 	return skb->len;
279 
280 nla_put_failure:
281 	nlmsg_trim(skb, b);
282 	return -1;
283 }
284 
285 static int multiq_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
286 			struct Qdisc **old, struct netlink_ext_ack *extack)
287 {
288 	struct multiq_sched_data *q = qdisc_priv(sch);
289 	unsigned long band = arg - 1;
290 
291 	if (new == NULL)
292 		new = &noop_qdisc;
293 
294 	*old = qdisc_replace(sch, new, &q->queues[band]);
295 	return 0;
296 }
297 
298 static struct Qdisc *
299 multiq_leaf(struct Qdisc *sch, unsigned long arg)
300 {
301 	struct multiq_sched_data *q = qdisc_priv(sch);
302 	unsigned long band = arg - 1;
303 
304 	return q->queues[band];
305 }
306 
307 static unsigned long multiq_find(struct Qdisc *sch, u32 classid)
308 {
309 	struct multiq_sched_data *q = qdisc_priv(sch);
310 	unsigned long band = TC_H_MIN(classid);
311 
312 	if (band - 1 >= q->bands)
313 		return 0;
314 	return band;
315 }
316 
317 static unsigned long multiq_bind(struct Qdisc *sch, unsigned long parent,
318 				 u32 classid)
319 {
320 	return multiq_find(sch, classid);
321 }
322 
323 
324 static void multiq_unbind(struct Qdisc *q, unsigned long cl)
325 {
326 }
327 
328 static int multiq_dump_class(struct Qdisc *sch, unsigned long cl,
329 			     struct sk_buff *skb, struct tcmsg *tcm)
330 {
331 	struct multiq_sched_data *q = qdisc_priv(sch);
332 
333 	tcm->tcm_handle |= TC_H_MIN(cl);
334 	tcm->tcm_info = q->queues[cl - 1]->handle;
335 	return 0;
336 }
337 
338 static int multiq_dump_class_stats(struct Qdisc *sch, unsigned long cl,
339 				 struct gnet_dump *d)
340 {
341 	struct multiq_sched_data *q = qdisc_priv(sch);
342 	struct Qdisc *cl_q;
343 
344 	cl_q = q->queues[cl - 1];
345 	if (gnet_stats_copy_basic(qdisc_root_sleeping_running(sch),
346 				  d, NULL, &cl_q->bstats) < 0 ||
347 	    gnet_stats_copy_queue(d, NULL, &cl_q->qstats, cl_q->q.qlen) < 0)
348 		return -1;
349 
350 	return 0;
351 }
352 
353 static void multiq_walk(struct Qdisc *sch, struct qdisc_walker *arg)
354 {
355 	struct multiq_sched_data *q = qdisc_priv(sch);
356 	int band;
357 
358 	if (arg->stop)
359 		return;
360 
361 	for (band = 0; band < q->bands; band++) {
362 		if (arg->count < arg->skip) {
363 			arg->count++;
364 			continue;
365 		}
366 		if (arg->fn(sch, band + 1, arg) < 0) {
367 			arg->stop = 1;
368 			break;
369 		}
370 		arg->count++;
371 	}
372 }
373 
374 static struct tcf_block *multiq_tcf_block(struct Qdisc *sch, unsigned long cl,
375 					  struct netlink_ext_ack *extack)
376 {
377 	struct multiq_sched_data *q = qdisc_priv(sch);
378 
379 	if (cl)
380 		return NULL;
381 	return q->block;
382 }
383 
384 static const struct Qdisc_class_ops multiq_class_ops = {
385 	.graft		=	multiq_graft,
386 	.leaf		=	multiq_leaf,
387 	.find		=	multiq_find,
388 	.walk		=	multiq_walk,
389 	.tcf_block	=	multiq_tcf_block,
390 	.bind_tcf	=	multiq_bind,
391 	.unbind_tcf	=	multiq_unbind,
392 	.dump		=	multiq_dump_class,
393 	.dump_stats	=	multiq_dump_class_stats,
394 };
395 
396 static struct Qdisc_ops multiq_qdisc_ops __read_mostly = {
397 	.next		=	NULL,
398 	.cl_ops		=	&multiq_class_ops,
399 	.id		=	"multiq",
400 	.priv_size	=	sizeof(struct multiq_sched_data),
401 	.enqueue	=	multiq_enqueue,
402 	.dequeue	=	multiq_dequeue,
403 	.peek		=	multiq_peek,
404 	.init		=	multiq_init,
405 	.reset		=	multiq_reset,
406 	.destroy	=	multiq_destroy,
407 	.change		=	multiq_tune,
408 	.dump		=	multiq_dump,
409 	.owner		=	THIS_MODULE,
410 };
411 
412 static int __init multiq_module_init(void)
413 {
414 	return register_qdisc(&multiq_qdisc_ops);
415 }
416 
417 static void __exit multiq_module_exit(void)
418 {
419 	unregister_qdisc(&multiq_qdisc_ops);
420 }
421 
422 module_init(multiq_module_init)
423 module_exit(multiq_module_exit)
424 
425 MODULE_LICENSE("GPL");
426