xref: /linux/net/sched/sch_multiq.c (revision e2be04c7f9958dde770eeb8b30e829ca969b37bb)
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 	case TC_ACT_SHOT:
58 		return NULL;
59 	}
60 #endif
61 	band = skb_get_queue_mapping(skb);
62 
63 	if (band >= q->bands)
64 		return q->queues[0];
65 
66 	return q->queues[band];
67 }
68 
69 static int
70 multiq_enqueue(struct sk_buff *skb, struct Qdisc *sch,
71 	       struct sk_buff **to_free)
72 {
73 	struct Qdisc *qdisc;
74 	int ret;
75 
76 	qdisc = multiq_classify(skb, sch, &ret);
77 #ifdef CONFIG_NET_CLS_ACT
78 	if (qdisc == NULL) {
79 
80 		if (ret & __NET_XMIT_BYPASS)
81 			qdisc_qstats_drop(sch);
82 		__qdisc_drop(skb, to_free);
83 		return ret;
84 	}
85 #endif
86 
87 	ret = qdisc_enqueue(skb, qdisc, to_free);
88 	if (ret == NET_XMIT_SUCCESS) {
89 		sch->q.qlen++;
90 		return NET_XMIT_SUCCESS;
91 	}
92 	if (net_xmit_drop_count(ret))
93 		qdisc_qstats_drop(sch);
94 	return ret;
95 }
96 
97 static struct sk_buff *multiq_dequeue(struct Qdisc *sch)
98 {
99 	struct multiq_sched_data *q = qdisc_priv(sch);
100 	struct Qdisc *qdisc;
101 	struct sk_buff *skb;
102 	int band;
103 
104 	for (band = 0; band < q->bands; band++) {
105 		/* cycle through bands to ensure fairness */
106 		q->curband++;
107 		if (q->curband >= q->bands)
108 			q->curband = 0;
109 
110 		/* Check that target subqueue is available before
111 		 * pulling an skb to avoid head-of-line blocking.
112 		 */
113 		if (!netif_xmit_stopped(
114 		    netdev_get_tx_queue(qdisc_dev(sch), q->curband))) {
115 			qdisc = q->queues[q->curband];
116 			skb = qdisc->dequeue(qdisc);
117 			if (skb) {
118 				qdisc_bstats_update(sch, skb);
119 				sch->q.qlen--;
120 				return skb;
121 			}
122 		}
123 	}
124 	return NULL;
125 
126 }
127 
128 static struct sk_buff *multiq_peek(struct Qdisc *sch)
129 {
130 	struct multiq_sched_data *q = qdisc_priv(sch);
131 	unsigned int curband = q->curband;
132 	struct Qdisc *qdisc;
133 	struct sk_buff *skb;
134 	int band;
135 
136 	for (band = 0; band < q->bands; band++) {
137 		/* cycle through bands to ensure fairness */
138 		curband++;
139 		if (curband >= q->bands)
140 			curband = 0;
141 
142 		/* Check that target subqueue is available before
143 		 * pulling an skb to avoid head-of-line blocking.
144 		 */
145 		if (!netif_xmit_stopped(
146 		    netdev_get_tx_queue(qdisc_dev(sch), curband))) {
147 			qdisc = q->queues[curband];
148 			skb = qdisc->ops->peek(qdisc);
149 			if (skb)
150 				return skb;
151 		}
152 	}
153 	return NULL;
154 
155 }
156 
157 static void
158 multiq_reset(struct Qdisc *sch)
159 {
160 	u16 band;
161 	struct multiq_sched_data *q = qdisc_priv(sch);
162 
163 	for (band = 0; band < q->bands; band++)
164 		qdisc_reset(q->queues[band]);
165 	sch->q.qlen = 0;
166 	q->curband = 0;
167 }
168 
169 static void
170 multiq_destroy(struct Qdisc *sch)
171 {
172 	int band;
173 	struct multiq_sched_data *q = qdisc_priv(sch);
174 
175 	tcf_block_put(q->block);
176 	for (band = 0; band < q->bands; band++)
177 		qdisc_destroy(q->queues[band]);
178 
179 	kfree(q->queues);
180 }
181 
182 static int multiq_tune(struct Qdisc *sch, struct nlattr *opt)
183 {
184 	struct multiq_sched_data *q = qdisc_priv(sch);
185 	struct tc_multiq_qopt *qopt;
186 	int i;
187 
188 	if (!netif_is_multiqueue(qdisc_dev(sch)))
189 		return -EOPNOTSUPP;
190 	if (nla_len(opt) < sizeof(*qopt))
191 		return -EINVAL;
192 
193 	qopt = nla_data(opt);
194 
195 	qopt->bands = qdisc_dev(sch)->real_num_tx_queues;
196 
197 	sch_tree_lock(sch);
198 	q->bands = qopt->bands;
199 	for (i = q->bands; i < q->max_bands; i++) {
200 		if (q->queues[i] != &noop_qdisc) {
201 			struct Qdisc *child = q->queues[i];
202 			q->queues[i] = &noop_qdisc;
203 			qdisc_tree_reduce_backlog(child, child->q.qlen,
204 						  child->qstats.backlog);
205 			qdisc_destroy(child);
206 		}
207 	}
208 
209 	sch_tree_unlock(sch);
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));
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_tree_reduce_backlog(old,
227 								  old->q.qlen,
228 								  old->qstats.backlog);
229 					qdisc_destroy(old);
230 				}
231 				sch_tree_unlock(sch);
232 			}
233 		}
234 	}
235 	return 0;
236 }
237 
238 static int multiq_init(struct Qdisc *sch, struct nlattr *opt)
239 {
240 	struct multiq_sched_data *q = qdisc_priv(sch);
241 	int i, err;
242 
243 	q->queues = NULL;
244 
245 	if (opt == NULL)
246 		return -EINVAL;
247 
248 	err = tcf_block_get(&q->block, &q->filter_list);
249 	if (err)
250 		return err;
251 
252 	q->max_bands = qdisc_dev(sch)->num_tx_queues;
253 
254 	q->queues = kcalloc(q->max_bands, sizeof(struct Qdisc *), GFP_KERNEL);
255 	if (!q->queues)
256 		return -ENOBUFS;
257 	for (i = 0; i < q->max_bands; i++)
258 		q->queues[i] = &noop_qdisc;
259 
260 	return multiq_tune(sch, opt);
261 }
262 
263 static int multiq_dump(struct Qdisc *sch, struct sk_buff *skb)
264 {
265 	struct multiq_sched_data *q = qdisc_priv(sch);
266 	unsigned char *b = skb_tail_pointer(skb);
267 	struct tc_multiq_qopt opt;
268 
269 	opt.bands = q->bands;
270 	opt.max_bands = q->max_bands;
271 
272 	if (nla_put(skb, TCA_OPTIONS, sizeof(opt), &opt))
273 		goto nla_put_failure;
274 
275 	return skb->len;
276 
277 nla_put_failure:
278 	nlmsg_trim(skb, b);
279 	return -1;
280 }
281 
282 static int multiq_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
283 		      struct Qdisc **old)
284 {
285 	struct multiq_sched_data *q = qdisc_priv(sch);
286 	unsigned long band = arg - 1;
287 
288 	if (new == NULL)
289 		new = &noop_qdisc;
290 
291 	*old = qdisc_replace(sch, new, &q->queues[band]);
292 	return 0;
293 }
294 
295 static struct Qdisc *
296 multiq_leaf(struct Qdisc *sch, unsigned long arg)
297 {
298 	struct multiq_sched_data *q = qdisc_priv(sch);
299 	unsigned long band = arg - 1;
300 
301 	return q->queues[band];
302 }
303 
304 static unsigned long multiq_find(struct Qdisc *sch, u32 classid)
305 {
306 	struct multiq_sched_data *q = qdisc_priv(sch);
307 	unsigned long band = TC_H_MIN(classid);
308 
309 	if (band - 1 >= q->bands)
310 		return 0;
311 	return band;
312 }
313 
314 static unsigned long multiq_bind(struct Qdisc *sch, unsigned long parent,
315 				 u32 classid)
316 {
317 	return multiq_find(sch, classid);
318 }
319 
320 
321 static void multiq_unbind(struct Qdisc *q, unsigned long cl)
322 {
323 }
324 
325 static int multiq_dump_class(struct Qdisc *sch, unsigned long cl,
326 			     struct sk_buff *skb, struct tcmsg *tcm)
327 {
328 	struct multiq_sched_data *q = qdisc_priv(sch);
329 
330 	tcm->tcm_handle |= TC_H_MIN(cl);
331 	tcm->tcm_info = q->queues[cl - 1]->handle;
332 	return 0;
333 }
334 
335 static int multiq_dump_class_stats(struct Qdisc *sch, unsigned long cl,
336 				 struct gnet_dump *d)
337 {
338 	struct multiq_sched_data *q = qdisc_priv(sch);
339 	struct Qdisc *cl_q;
340 
341 	cl_q = q->queues[cl - 1];
342 	if (gnet_stats_copy_basic(qdisc_root_sleeping_running(sch),
343 				  d, NULL, &cl_q->bstats) < 0 ||
344 	    gnet_stats_copy_queue(d, NULL, &cl_q->qstats, cl_q->q.qlen) < 0)
345 		return -1;
346 
347 	return 0;
348 }
349 
350 static void multiq_walk(struct Qdisc *sch, struct qdisc_walker *arg)
351 {
352 	struct multiq_sched_data *q = qdisc_priv(sch);
353 	int band;
354 
355 	if (arg->stop)
356 		return;
357 
358 	for (band = 0; band < q->bands; band++) {
359 		if (arg->count < arg->skip) {
360 			arg->count++;
361 			continue;
362 		}
363 		if (arg->fn(sch, band + 1, arg) < 0) {
364 			arg->stop = 1;
365 			break;
366 		}
367 		arg->count++;
368 	}
369 }
370 
371 static struct tcf_block *multiq_tcf_block(struct Qdisc *sch, unsigned long cl)
372 {
373 	struct multiq_sched_data *q = qdisc_priv(sch);
374 
375 	if (cl)
376 		return NULL;
377 	return q->block;
378 }
379 
380 static const struct Qdisc_class_ops multiq_class_ops = {
381 	.graft		=	multiq_graft,
382 	.leaf		=	multiq_leaf,
383 	.find		=	multiq_find,
384 	.walk		=	multiq_walk,
385 	.tcf_block	=	multiq_tcf_block,
386 	.bind_tcf	=	multiq_bind,
387 	.unbind_tcf	=	multiq_unbind,
388 	.dump		=	multiq_dump_class,
389 	.dump_stats	=	multiq_dump_class_stats,
390 };
391 
392 static struct Qdisc_ops multiq_qdisc_ops __read_mostly = {
393 	.next		=	NULL,
394 	.cl_ops		=	&multiq_class_ops,
395 	.id		=	"multiq",
396 	.priv_size	=	sizeof(struct multiq_sched_data),
397 	.enqueue	=	multiq_enqueue,
398 	.dequeue	=	multiq_dequeue,
399 	.peek		=	multiq_peek,
400 	.init		=	multiq_init,
401 	.reset		=	multiq_reset,
402 	.destroy	=	multiq_destroy,
403 	.change		=	multiq_tune,
404 	.dump		=	multiq_dump,
405 	.owner		=	THIS_MODULE,
406 };
407 
408 static int __init multiq_module_init(void)
409 {
410 	return register_qdisc(&multiq_qdisc_ops);
411 }
412 
413 static void __exit multiq_module_exit(void)
414 {
415 	unregister_qdisc(&multiq_qdisc_ops);
416 }
417 
418 module_init(multiq_module_init)
419 module_exit(multiq_module_exit)
420 
421 MODULE_LICENSE("GPL");
422