xref: /openbmc/linux/net/sched/sch_prio.c (revision f9a82c48) 
1 /*
2  * net/sched/sch_prio.c	Simple 3-band priority "scheduler".
3  *
4  *		This program is free software; you can redistribute it and/or
5  *		modify it under the terms of the GNU General Public License
6  *		as published by the Free Software Foundation; either version
7  *		2 of the License, or (at your option) any later version.
8  *
9  * Authors:	Alexey Kuznetsov, <kuznet@ms2.inr.ac.ru>
10  * Fixes:       19990609: J Hadi Salim <hadi@nortelnetworks.com>:
11  *              Init --  EINVAL when opt undefined
12  */
13 
14 #include <linux/module.h>
15 #include <linux/slab.h>
16 #include <linux/types.h>
17 #include <linux/kernel.h>
18 #include <linux/string.h>
19 #include <linux/errno.h>
20 #include <linux/skbuff.h>
21 #include <net/netlink.h>
22 #include <net/pkt_sched.h>
23 #include <net/pkt_cls.h>
24 
25 struct prio_sched_data {
26 	int bands;
27 	struct tcf_proto __rcu *filter_list;
28 	struct tcf_block *block;
29 	u8  prio2band[TC_PRIO_MAX+1];
30 	struct Qdisc *queues[TCQ_PRIO_BANDS];
31 };
32 
33 
34 static struct Qdisc *
35 prio_classify(struct sk_buff *skb, struct Qdisc *sch, int *qerr)
36 {
37 	struct prio_sched_data *q = qdisc_priv(sch);
38 	u32 band = skb->priority;
39 	struct tcf_result res;
40 	struct tcf_proto *fl;
41 	int err;
42 
43 	*qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
44 	if (TC_H_MAJ(skb->priority) != sch->handle) {
45 		fl = rcu_dereference_bh(q->filter_list);
46 		err = tcf_classify(skb, fl, &res, false);
47 #ifdef CONFIG_NET_CLS_ACT
48 		switch (err) {
49 		case TC_ACT_STOLEN:
50 		case TC_ACT_QUEUED:
51 		case TC_ACT_TRAP:
52 			*qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
53 			/* fall through */
54 		case TC_ACT_SHOT:
55 			return NULL;
56 		}
57 #endif
58 		if (!fl || err < 0) {
59 			if (TC_H_MAJ(band))
60 				band = 0;
61 			return q->queues[q->prio2band[band & TC_PRIO_MAX]];
62 		}
63 		band = res.classid;
64 	}
65 	band = TC_H_MIN(band) - 1;
66 	if (band >= q->bands)
67 		return q->queues[q->prio2band[0]];
68 
69 	return q->queues[band];
70 }
71 
72 static int
73 prio_enqueue(struct sk_buff *skb, struct Qdisc *sch, struct sk_buff **to_free)
74 {
75 	unsigned int len = qdisc_pkt_len(skb);
76 	struct Qdisc *qdisc;
77 	int ret;
78 
79 	qdisc = prio_classify(skb, sch, &ret);
80 #ifdef CONFIG_NET_CLS_ACT
81 	if (qdisc == NULL) {
82 
83 		if (ret & __NET_XMIT_BYPASS)
84 			qdisc_qstats_drop(sch);
85 		__qdisc_drop(skb, to_free);
86 		return ret;
87 	}
88 #endif
89 
90 	ret = qdisc_enqueue(skb, qdisc, to_free);
91 	if (ret == NET_XMIT_SUCCESS) {
92 		sch->qstats.backlog += len;
93 		sch->q.qlen++;
94 		return NET_XMIT_SUCCESS;
95 	}
96 	if (net_xmit_drop_count(ret))
97 		qdisc_qstats_drop(sch);
98 	return ret;
99 }
100 
101 static struct sk_buff *prio_peek(struct Qdisc *sch)
102 {
103 	struct prio_sched_data *q = qdisc_priv(sch);
104 	int prio;
105 
106 	for (prio = 0; prio < q->bands; prio++) {
107 		struct Qdisc *qdisc = q->queues[prio];
108 		struct sk_buff *skb = qdisc->ops->peek(qdisc);
109 		if (skb)
110 			return skb;
111 	}
112 	return NULL;
113 }
114 
115 static struct sk_buff *prio_dequeue(struct Qdisc *sch)
116 {
117 	struct prio_sched_data *q = qdisc_priv(sch);
118 	int prio;
119 
120 	for (prio = 0; prio < q->bands; prio++) {
121 		struct Qdisc *qdisc = q->queues[prio];
122 		struct sk_buff *skb = qdisc_dequeue_peeked(qdisc);
123 		if (skb) {
124 			qdisc_bstats_update(sch, skb);
125 			qdisc_qstats_backlog_dec(sch, skb);
126 			sch->q.qlen--;
127 			return skb;
128 		}
129 	}
130 	return NULL;
131 
132 }
133 
134 static void
135 prio_reset(struct Qdisc *sch)
136 {
137 	int prio;
138 	struct prio_sched_data *q = qdisc_priv(sch);
139 
140 	for (prio = 0; prio < q->bands; prio++)
141 		qdisc_reset(q->queues[prio]);
142 	sch->qstats.backlog = 0;
143 	sch->q.qlen = 0;
144 }
145 
146 static int prio_offload(struct Qdisc *sch, struct tc_prio_qopt *qopt)
147 {
148 	struct net_device *dev = qdisc_dev(sch);
149 	struct tc_prio_qopt_offload opt = {
150 		.handle = sch->handle,
151 		.parent = sch->parent,
152 	};
153 
154 	if (!tc_can_offload(dev) || !dev->netdev_ops->ndo_setup_tc)
155 		return -EOPNOTSUPP;
156 
157 	if (qopt) {
158 		opt.command = TC_PRIO_REPLACE;
159 		opt.replace_params.bands = qopt->bands;
160 		memcpy(&opt.replace_params.priomap, qopt->priomap,
161 		       TC_PRIO_MAX + 1);
162 		opt.replace_params.qstats = &sch->qstats;
163 	} else {
164 		opt.command = TC_PRIO_DESTROY;
165 	}
166 
167 	return dev->netdev_ops->ndo_setup_tc(dev, TC_SETUP_QDISC_PRIO, &opt);
168 }
169 
170 static void
171 prio_destroy(struct Qdisc *sch)
172 {
173 	int prio;
174 	struct prio_sched_data *q = qdisc_priv(sch);
175 
176 	tcf_block_put(q->block);
177 	prio_offload(sch, NULL);
178 	for (prio = 0; prio < q->bands; prio++)
179 		qdisc_put(q->queues[prio]);
180 }
181 
182 static int prio_tune(struct Qdisc *sch, struct nlattr *opt,
183 		     struct netlink_ext_ack *extack)
184 {
185 	struct prio_sched_data *q = qdisc_priv(sch);
186 	struct Qdisc *queues[TCQ_PRIO_BANDS];
187 	int oldbands = q->bands, i;
188 	struct tc_prio_qopt *qopt;
189 
190 	if (nla_len(opt) < sizeof(*qopt))
191 		return -EINVAL;
192 	qopt = nla_data(opt);
193 
194 	if (qopt->bands > TCQ_PRIO_BANDS || qopt->bands < 2)
195 		return -EINVAL;
196 
197 	for (i = 0; i <= TC_PRIO_MAX; i++) {
198 		if (qopt->priomap[i] >= qopt->bands)
199 			return -EINVAL;
200 	}
201 
202 	/* Before commit, make sure we can allocate all new qdiscs */
203 	for (i = oldbands; i < qopt->bands; i++) {
204 		queues[i] = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
205 					      TC_H_MAKE(sch->handle, i + 1),
206 					      extack);
207 		if (!queues[i]) {
208 			while (i > oldbands)
209 				qdisc_put(queues[--i]);
210 			return -ENOMEM;
211 		}
212 	}
213 
214 	prio_offload(sch, qopt);
215 	sch_tree_lock(sch);
216 	q->bands = qopt->bands;
217 	memcpy(q->prio2band, qopt->priomap, TC_PRIO_MAX+1);
218 
219 	for (i = q->bands; i < oldbands; i++) {
220 		struct Qdisc *child = q->queues[i];
221 
222 		qdisc_tree_reduce_backlog(child, child->q.qlen,
223 					  child->qstats.backlog);
224 	}
225 
226 	for (i = oldbands; i < q->bands; i++) {
227 		q->queues[i] = queues[i];
228 		if (q->queues[i] != &noop_qdisc)
229 			qdisc_hash_add(q->queues[i], true);
230 	}
231 
232 	sch_tree_unlock(sch);
233 
234 	for (i = q->bands; i < oldbands; i++)
235 		qdisc_put(q->queues[i]);
236 	return 0;
237 }
238 
239 static int prio_init(struct Qdisc *sch, struct nlattr *opt,
240 		     struct netlink_ext_ack *extack)
241 {
242 	struct prio_sched_data *q = qdisc_priv(sch);
243 	int err;
244 
245 	if (!opt)
246 		return -EINVAL;
247 
248 	err = tcf_block_get(&q->block, &q->filter_list, sch, extack);
249 	if (err)
250 		return err;
251 
252 	return prio_tune(sch, opt, extack);
253 }
254 
255 static int prio_dump_offload(struct Qdisc *sch)
256 {
257 	struct tc_prio_qopt_offload hw_stats = {
258 		.command = TC_PRIO_STATS,
259 		.handle = sch->handle,
260 		.parent = sch->parent,
261 		{
262 			.stats = {
263 				.bstats = &sch->bstats,
264 				.qstats = &sch->qstats,
265 			},
266 		},
267 	};
268 
269 	return qdisc_offload_dump_helper(sch, TC_SETUP_QDISC_PRIO, &hw_stats);
270 }
271 
272 static int prio_dump(struct Qdisc *sch, struct sk_buff *skb)
273 {
274 	struct prio_sched_data *q = qdisc_priv(sch);
275 	unsigned char *b = skb_tail_pointer(skb);
276 	struct tc_prio_qopt opt;
277 	int err;
278 
279 	opt.bands = q->bands;
280 	memcpy(&opt.priomap, q->prio2band, TC_PRIO_MAX + 1);
281 
282 	err = prio_dump_offload(sch);
283 	if (err)
284 		goto nla_put_failure;
285 
286 	if (nla_put(skb, TCA_OPTIONS, sizeof(opt), &opt))
287 		goto nla_put_failure;
288 
289 	return skb->len;
290 
291 nla_put_failure:
292 	nlmsg_trim(skb, b);
293 	return -1;
294 }
295 
296 static int prio_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
297 		      struct Qdisc **old, struct netlink_ext_ack *extack)
298 {
299 	struct prio_sched_data *q = qdisc_priv(sch);
300 	struct tc_prio_qopt_offload graft_offload;
301 	unsigned long band = arg - 1;
302 
303 	if (new == NULL)
304 		new = &noop_qdisc;
305 
306 	*old = qdisc_replace(sch, new, &q->queues[band]);
307 
308 	graft_offload.handle = sch->handle;
309 	graft_offload.parent = sch->parent;
310 	graft_offload.graft_params.band = band;
311 	graft_offload.graft_params.child_handle = new->handle;
312 	graft_offload.command = TC_PRIO_GRAFT;
313 
314 	qdisc_offload_graft_helper(qdisc_dev(sch), sch, new, *old,
315 				   TC_SETUP_QDISC_PRIO, &graft_offload,
316 				   extack);
317 	return 0;
318 }
319 
320 static struct Qdisc *
321 prio_leaf(struct Qdisc *sch, unsigned long arg)
322 {
323 	struct prio_sched_data *q = qdisc_priv(sch);
324 	unsigned long band = arg - 1;
325 
326 	return q->queues[band];
327 }
328 
329 static unsigned long prio_find(struct Qdisc *sch, u32 classid)
330 {
331 	struct prio_sched_data *q = qdisc_priv(sch);
332 	unsigned long band = TC_H_MIN(classid);
333 
334 	if (band - 1 >= q->bands)
335 		return 0;
336 	return band;
337 }
338 
339 static unsigned long prio_bind(struct Qdisc *sch, unsigned long parent, u32 classid)
340 {
341 	return prio_find(sch, classid);
342 }
343 
344 
345 static void prio_unbind(struct Qdisc *q, unsigned long cl)
346 {
347 }
348 
349 static int prio_dump_class(struct Qdisc *sch, unsigned long cl, struct sk_buff *skb,
350 			   struct tcmsg *tcm)
351 {
352 	struct prio_sched_data *q = qdisc_priv(sch);
353 
354 	tcm->tcm_handle |= TC_H_MIN(cl);
355 	tcm->tcm_info = q->queues[cl-1]->handle;
356 	return 0;
357 }
358 
359 static int prio_dump_class_stats(struct Qdisc *sch, unsigned long cl,
360 				 struct gnet_dump *d)
361 {
362 	struct prio_sched_data *q = qdisc_priv(sch);
363 	struct Qdisc *cl_q;
364 
365 	cl_q = q->queues[cl - 1];
366 	if (gnet_stats_copy_basic(qdisc_root_sleeping_running(sch),
367 				  d, NULL, &cl_q->bstats) < 0 ||
368 	    gnet_stats_copy_queue(d, NULL, &cl_q->qstats, cl_q->q.qlen) < 0)
369 		return -1;
370 
371 	return 0;
372 }
373 
374 static void prio_walk(struct Qdisc *sch, struct qdisc_walker *arg)
375 {
376 	struct prio_sched_data *q = qdisc_priv(sch);
377 	int prio;
378 
379 	if (arg->stop)
380 		return;
381 
382 	for (prio = 0; prio < q->bands; prio++) {
383 		if (arg->count < arg->skip) {
384 			arg->count++;
385 			continue;
386 		}
387 		if (arg->fn(sch, prio + 1, arg) < 0) {
388 			arg->stop = 1;
389 			break;
390 		}
391 		arg->count++;
392 	}
393 }
394 
395 static struct tcf_block *prio_tcf_block(struct Qdisc *sch, unsigned long cl,
396 					struct netlink_ext_ack *extack)
397 {
398 	struct prio_sched_data *q = qdisc_priv(sch);
399 
400 	if (cl)
401 		return NULL;
402 	return q->block;
403 }
404 
405 static const struct Qdisc_class_ops prio_class_ops = {
406 	.graft		=	prio_graft,
407 	.leaf		=	prio_leaf,
408 	.find		=	prio_find,
409 	.walk		=	prio_walk,
410 	.tcf_block	=	prio_tcf_block,
411 	.bind_tcf	=	prio_bind,
412 	.unbind_tcf	=	prio_unbind,
413 	.dump		=	prio_dump_class,
414 	.dump_stats	=	prio_dump_class_stats,
415 };
416 
417 static struct Qdisc_ops prio_qdisc_ops __read_mostly = {
418 	.next		=	NULL,
419 	.cl_ops		=	&prio_class_ops,
420 	.id		=	"prio",
421 	.priv_size	=	sizeof(struct prio_sched_data),
422 	.enqueue	=	prio_enqueue,
423 	.dequeue	=	prio_dequeue,
424 	.peek		=	prio_peek,
425 	.init		=	prio_init,
426 	.reset		=	prio_reset,
427 	.destroy	=	prio_destroy,
428 	.change		=	prio_tune,
429 	.dump		=	prio_dump,
430 	.owner		=	THIS_MODULE,
431 };
432 
433 static int __init prio_module_init(void)
434 {
435 	return register_qdisc(&prio_qdisc_ops);
436 }
437 
438 static void __exit prio_module_exit(void)
439 {
440 	unregister_qdisc(&prio_qdisc_ops);
441 }
442 
443 module_init(prio_module_init)
444 module_exit(prio_module_exit)
445 
446 MODULE_LICENSE("GPL");
447