xref: /openbmc/linux/net/sched/sch_drr.c (revision 96c63fa7393d0a346acfe5a91e0c7d4c7782641b) 
1 /*
2  * net/sched/sch_drr.c         Deficit Round Robin scheduler
3  *
4  * Copyright (c) 2008 Patrick McHardy <kaber@trash.net>
5  *
6  * This program is free software; you can redistribute it and/or
7  * modify it under the terms of the GNU General Public License
8  * version 2 as published by the Free Software Foundation.
9  */
10 
11 #include <linux/module.h>
12 #include <linux/slab.h>
13 #include <linux/init.h>
14 #include <linux/errno.h>
15 #include <linux/netdevice.h>
16 #include <linux/pkt_sched.h>
17 #include <net/sch_generic.h>
18 #include <net/pkt_sched.h>
19 #include <net/pkt_cls.h>
20 
21 struct drr_class {
22 	struct Qdisc_class_common	common;
23 	unsigned int			refcnt;
24 	unsigned int			filter_cnt;
25 
26 	struct gnet_stats_basic_packed		bstats;
27 	struct gnet_stats_queue		qstats;
28 	struct gnet_stats_rate_est64	rate_est;
29 	struct list_head		alist;
30 	struct Qdisc			*qdisc;
31 
32 	u32				quantum;
33 	u32				deficit;
34 };
35 
36 struct drr_sched {
37 	struct list_head		active;
38 	struct tcf_proto __rcu		*filter_list;
39 	struct Qdisc_class_hash		clhash;
40 };
41 
42 static struct drr_class *drr_find_class(struct Qdisc *sch, u32 classid)
43 {
44 	struct drr_sched *q = qdisc_priv(sch);
45 	struct Qdisc_class_common *clc;
46 
47 	clc = qdisc_class_find(&q->clhash, classid);
48 	if (clc == NULL)
49 		return NULL;
50 	return container_of(clc, struct drr_class, common);
51 }
52 
53 static void drr_purge_queue(struct drr_class *cl)
54 {
55 	unsigned int len = cl->qdisc->q.qlen;
56 	unsigned int backlog = cl->qdisc->qstats.backlog;
57 
58 	qdisc_reset(cl->qdisc);
59 	qdisc_tree_reduce_backlog(cl->qdisc, len, backlog);
60 }
61 
62 static const struct nla_policy drr_policy[TCA_DRR_MAX + 1] = {
63 	[TCA_DRR_QUANTUM]	= { .type = NLA_U32 },
64 };
65 
66 static int drr_change_class(struct Qdisc *sch, u32 classid, u32 parentid,
67 			    struct nlattr **tca, unsigned long *arg)
68 {
69 	struct drr_sched *q = qdisc_priv(sch);
70 	struct drr_class *cl = (struct drr_class *)*arg;
71 	struct nlattr *opt = tca[TCA_OPTIONS];
72 	struct nlattr *tb[TCA_DRR_MAX + 1];
73 	u32 quantum;
74 	int err;
75 
76 	if (!opt)
77 		return -EINVAL;
78 
79 	err = nla_parse_nested(tb, TCA_DRR_MAX, opt, drr_policy);
80 	if (err < 0)
81 		return err;
82 
83 	if (tb[TCA_DRR_QUANTUM]) {
84 		quantum = nla_get_u32(tb[TCA_DRR_QUANTUM]);
85 		if (quantum == 0)
86 			return -EINVAL;
87 	} else
88 		quantum = psched_mtu(qdisc_dev(sch));
89 
90 	if (cl != NULL) {
91 		if (tca[TCA_RATE]) {
92 			err = gen_replace_estimator(&cl->bstats, NULL,
93 						    &cl->rate_est,
94 						    NULL,
95 						    qdisc_root_sleeping_running(sch),
96 						    tca[TCA_RATE]);
97 			if (err)
98 				return err;
99 		}
100 
101 		sch_tree_lock(sch);
102 		if (tb[TCA_DRR_QUANTUM])
103 			cl->quantum = quantum;
104 		sch_tree_unlock(sch);
105 
106 		return 0;
107 	}
108 
109 	cl = kzalloc(sizeof(struct drr_class), GFP_KERNEL);
110 	if (cl == NULL)
111 		return -ENOBUFS;
112 
113 	cl->refcnt	   = 1;
114 	cl->common.classid = classid;
115 	cl->quantum	   = quantum;
116 	cl->qdisc	   = qdisc_create_dflt(sch->dev_queue,
117 					       &pfifo_qdisc_ops, classid);
118 	if (cl->qdisc == NULL)
119 		cl->qdisc = &noop_qdisc;
120 
121 	if (tca[TCA_RATE]) {
122 		err = gen_replace_estimator(&cl->bstats, NULL, &cl->rate_est,
123 					    NULL,
124 					    qdisc_root_sleeping_running(sch),
125 					    tca[TCA_RATE]);
126 		if (err) {
127 			qdisc_destroy(cl->qdisc);
128 			kfree(cl);
129 			return err;
130 		}
131 	}
132 
133 	sch_tree_lock(sch);
134 	qdisc_class_hash_insert(&q->clhash, &cl->common);
135 	sch_tree_unlock(sch);
136 
137 	qdisc_class_hash_grow(sch, &q->clhash);
138 
139 	*arg = (unsigned long)cl;
140 	return 0;
141 }
142 
143 static void drr_destroy_class(struct Qdisc *sch, struct drr_class *cl)
144 {
145 	gen_kill_estimator(&cl->bstats, &cl->rate_est);
146 	qdisc_destroy(cl->qdisc);
147 	kfree(cl);
148 }
149 
150 static int drr_delete_class(struct Qdisc *sch, unsigned long arg)
151 {
152 	struct drr_sched *q = qdisc_priv(sch);
153 	struct drr_class *cl = (struct drr_class *)arg;
154 
155 	if (cl->filter_cnt > 0)
156 		return -EBUSY;
157 
158 	sch_tree_lock(sch);
159 
160 	drr_purge_queue(cl);
161 	qdisc_class_hash_remove(&q->clhash, &cl->common);
162 
163 	BUG_ON(--cl->refcnt == 0);
164 	/*
165 	 * This shouldn't happen: we "hold" one cops->get() when called
166 	 * from tc_ctl_tclass; the destroy method is done from cops->put().
167 	 */
168 
169 	sch_tree_unlock(sch);
170 	return 0;
171 }
172 
173 static unsigned long drr_get_class(struct Qdisc *sch, u32 classid)
174 {
175 	struct drr_class *cl = drr_find_class(sch, classid);
176 
177 	if (cl != NULL)
178 		cl->refcnt++;
179 
180 	return (unsigned long)cl;
181 }
182 
183 static void drr_put_class(struct Qdisc *sch, unsigned long arg)
184 {
185 	struct drr_class *cl = (struct drr_class *)arg;
186 
187 	if (--cl->refcnt == 0)
188 		drr_destroy_class(sch, cl);
189 }
190 
191 static struct tcf_proto __rcu **drr_tcf_chain(struct Qdisc *sch,
192 					      unsigned long cl)
193 {
194 	struct drr_sched *q = qdisc_priv(sch);
195 
196 	if (cl)
197 		return NULL;
198 
199 	return &q->filter_list;
200 }
201 
202 static unsigned long drr_bind_tcf(struct Qdisc *sch, unsigned long parent,
203 				  u32 classid)
204 {
205 	struct drr_class *cl = drr_find_class(sch, classid);
206 
207 	if (cl != NULL)
208 		cl->filter_cnt++;
209 
210 	return (unsigned long)cl;
211 }
212 
213 static void drr_unbind_tcf(struct Qdisc *sch, unsigned long arg)
214 {
215 	struct drr_class *cl = (struct drr_class *)arg;
216 
217 	cl->filter_cnt--;
218 }
219 
220 static int drr_graft_class(struct Qdisc *sch, unsigned long arg,
221 			   struct Qdisc *new, struct Qdisc **old)
222 {
223 	struct drr_class *cl = (struct drr_class *)arg;
224 
225 	if (new == NULL) {
226 		new = qdisc_create_dflt(sch->dev_queue,
227 					&pfifo_qdisc_ops, cl->common.classid);
228 		if (new == NULL)
229 			new = &noop_qdisc;
230 	}
231 
232 	*old = qdisc_replace(sch, new, &cl->qdisc);
233 	return 0;
234 }
235 
236 static struct Qdisc *drr_class_leaf(struct Qdisc *sch, unsigned long arg)
237 {
238 	struct drr_class *cl = (struct drr_class *)arg;
239 
240 	return cl->qdisc;
241 }
242 
243 static void drr_qlen_notify(struct Qdisc *csh, unsigned long arg)
244 {
245 	struct drr_class *cl = (struct drr_class *)arg;
246 
247 	if (cl->qdisc->q.qlen == 0)
248 		list_del(&cl->alist);
249 }
250 
251 static int drr_dump_class(struct Qdisc *sch, unsigned long arg,
252 			  struct sk_buff *skb, struct tcmsg *tcm)
253 {
254 	struct drr_class *cl = (struct drr_class *)arg;
255 	struct nlattr *nest;
256 
257 	tcm->tcm_parent	= TC_H_ROOT;
258 	tcm->tcm_handle	= cl->common.classid;
259 	tcm->tcm_info	= cl->qdisc->handle;
260 
261 	nest = nla_nest_start(skb, TCA_OPTIONS);
262 	if (nest == NULL)
263 		goto nla_put_failure;
264 	if (nla_put_u32(skb, TCA_DRR_QUANTUM, cl->quantum))
265 		goto nla_put_failure;
266 	return nla_nest_end(skb, nest);
267 
268 nla_put_failure:
269 	nla_nest_cancel(skb, nest);
270 	return -EMSGSIZE;
271 }
272 
273 static int drr_dump_class_stats(struct Qdisc *sch, unsigned long arg,
274 				struct gnet_dump *d)
275 {
276 	struct drr_class *cl = (struct drr_class *)arg;
277 	__u32 qlen = cl->qdisc->q.qlen;
278 	struct tc_drr_stats xstats;
279 
280 	memset(&xstats, 0, sizeof(xstats));
281 	if (qlen)
282 		xstats.deficit = cl->deficit;
283 
284 	if (gnet_stats_copy_basic(qdisc_root_sleeping_running(sch),
285 				  d, NULL, &cl->bstats) < 0 ||
286 	    gnet_stats_copy_rate_est(d, &cl->bstats, &cl->rate_est) < 0 ||
287 	    gnet_stats_copy_queue(d, NULL, &cl->qdisc->qstats, qlen) < 0)
288 		return -1;
289 
290 	return gnet_stats_copy_app(d, &xstats, sizeof(xstats));
291 }
292 
293 static void drr_walk(struct Qdisc *sch, struct qdisc_walker *arg)
294 {
295 	struct drr_sched *q = qdisc_priv(sch);
296 	struct drr_class *cl;
297 	unsigned int i;
298 
299 	if (arg->stop)
300 		return;
301 
302 	for (i = 0; i < q->clhash.hashsize; i++) {
303 		hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode) {
304 			if (arg->count < arg->skip) {
305 				arg->count++;
306 				continue;
307 			}
308 			if (arg->fn(sch, (unsigned long)cl, arg) < 0) {
309 				arg->stop = 1;
310 				return;
311 			}
312 			arg->count++;
313 		}
314 	}
315 }
316 
317 static struct drr_class *drr_classify(struct sk_buff *skb, struct Qdisc *sch,
318 				      int *qerr)
319 {
320 	struct drr_sched *q = qdisc_priv(sch);
321 	struct drr_class *cl;
322 	struct tcf_result res;
323 	struct tcf_proto *fl;
324 	int result;
325 
326 	if (TC_H_MAJ(skb->priority ^ sch->handle) == 0) {
327 		cl = drr_find_class(sch, skb->priority);
328 		if (cl != NULL)
329 			return cl;
330 	}
331 
332 	*qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
333 	fl = rcu_dereference_bh(q->filter_list);
334 	result = tc_classify(skb, fl, &res, false);
335 	if (result >= 0) {
336 #ifdef CONFIG_NET_CLS_ACT
337 		switch (result) {
338 		case TC_ACT_QUEUED:
339 		case TC_ACT_STOLEN:
340 			*qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
341 		case TC_ACT_SHOT:
342 			return NULL;
343 		}
344 #endif
345 		cl = (struct drr_class *)res.class;
346 		if (cl == NULL)
347 			cl = drr_find_class(sch, res.classid);
348 		return cl;
349 	}
350 	return NULL;
351 }
352 
353 static int drr_enqueue(struct sk_buff *skb, struct Qdisc *sch)
354 {
355 	struct drr_sched *q = qdisc_priv(sch);
356 	struct drr_class *cl;
357 	int err = 0;
358 
359 	cl = drr_classify(skb, sch, &err);
360 	if (cl == NULL) {
361 		if (err & __NET_XMIT_BYPASS)
362 			qdisc_qstats_drop(sch);
363 		kfree_skb(skb);
364 		return err;
365 	}
366 
367 	err = qdisc_enqueue(skb, cl->qdisc);
368 	if (unlikely(err != NET_XMIT_SUCCESS)) {
369 		if (net_xmit_drop_count(err)) {
370 			cl->qstats.drops++;
371 			qdisc_qstats_drop(sch);
372 		}
373 		return err;
374 	}
375 
376 	if (cl->qdisc->q.qlen == 1) {
377 		list_add_tail(&cl->alist, &q->active);
378 		cl->deficit = cl->quantum;
379 	}
380 
381 	sch->q.qlen++;
382 	return err;
383 }
384 
385 static struct sk_buff *drr_dequeue(struct Qdisc *sch)
386 {
387 	struct drr_sched *q = qdisc_priv(sch);
388 	struct drr_class *cl;
389 	struct sk_buff *skb;
390 	unsigned int len;
391 
392 	if (list_empty(&q->active))
393 		goto out;
394 	while (1) {
395 		cl = list_first_entry(&q->active, struct drr_class, alist);
396 		skb = cl->qdisc->ops->peek(cl->qdisc);
397 		if (skb == NULL) {
398 			qdisc_warn_nonwc(__func__, cl->qdisc);
399 			goto out;
400 		}
401 
402 		len = qdisc_pkt_len(skb);
403 		if (len <= cl->deficit) {
404 			cl->deficit -= len;
405 			skb = qdisc_dequeue_peeked(cl->qdisc);
406 			if (unlikely(skb == NULL))
407 				goto out;
408 			if (cl->qdisc->q.qlen == 0)
409 				list_del(&cl->alist);
410 
411 			bstats_update(&cl->bstats, skb);
412 			qdisc_bstats_update(sch, skb);
413 			sch->q.qlen--;
414 			return skb;
415 		}
416 
417 		cl->deficit += cl->quantum;
418 		list_move_tail(&cl->alist, &q->active);
419 	}
420 out:
421 	return NULL;
422 }
423 
424 static unsigned int drr_drop(struct Qdisc *sch)
425 {
426 	struct drr_sched *q = qdisc_priv(sch);
427 	struct drr_class *cl;
428 	unsigned int len;
429 
430 	list_for_each_entry(cl, &q->active, alist) {
431 		if (cl->qdisc->ops->drop) {
432 			len = cl->qdisc->ops->drop(cl->qdisc);
433 			if (len > 0) {
434 				sch->q.qlen--;
435 				if (cl->qdisc->q.qlen == 0)
436 					list_del(&cl->alist);
437 				return len;
438 			}
439 		}
440 	}
441 	return 0;
442 }
443 
444 static int drr_init_qdisc(struct Qdisc *sch, struct nlattr *opt)
445 {
446 	struct drr_sched *q = qdisc_priv(sch);
447 	int err;
448 
449 	err = qdisc_class_hash_init(&q->clhash);
450 	if (err < 0)
451 		return err;
452 	INIT_LIST_HEAD(&q->active);
453 	return 0;
454 }
455 
456 static void drr_reset_qdisc(struct Qdisc *sch)
457 {
458 	struct drr_sched *q = qdisc_priv(sch);
459 	struct drr_class *cl;
460 	unsigned int i;
461 
462 	for (i = 0; i < q->clhash.hashsize; i++) {
463 		hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode) {
464 			if (cl->qdisc->q.qlen)
465 				list_del(&cl->alist);
466 			qdisc_reset(cl->qdisc);
467 		}
468 	}
469 	sch->q.qlen = 0;
470 }
471 
472 static void drr_destroy_qdisc(struct Qdisc *sch)
473 {
474 	struct drr_sched *q = qdisc_priv(sch);
475 	struct drr_class *cl;
476 	struct hlist_node *next;
477 	unsigned int i;
478 
479 	tcf_destroy_chain(&q->filter_list);
480 
481 	for (i = 0; i < q->clhash.hashsize; i++) {
482 		hlist_for_each_entry_safe(cl, next, &q->clhash.hash[i],
483 					  common.hnode)
484 			drr_destroy_class(sch, cl);
485 	}
486 	qdisc_class_hash_destroy(&q->clhash);
487 }
488 
489 static const struct Qdisc_class_ops drr_class_ops = {
490 	.change		= drr_change_class,
491 	.delete		= drr_delete_class,
492 	.get		= drr_get_class,
493 	.put		= drr_put_class,
494 	.tcf_chain	= drr_tcf_chain,
495 	.bind_tcf	= drr_bind_tcf,
496 	.unbind_tcf	= drr_unbind_tcf,
497 	.graft		= drr_graft_class,
498 	.leaf		= drr_class_leaf,
499 	.qlen_notify	= drr_qlen_notify,
500 	.dump		= drr_dump_class,
501 	.dump_stats	= drr_dump_class_stats,
502 	.walk		= drr_walk,
503 };
504 
505 static struct Qdisc_ops drr_qdisc_ops __read_mostly = {
506 	.cl_ops		= &drr_class_ops,
507 	.id		= "drr",
508 	.priv_size	= sizeof(struct drr_sched),
509 	.enqueue	= drr_enqueue,
510 	.dequeue	= drr_dequeue,
511 	.peek		= qdisc_peek_dequeued,
512 	.drop		= drr_drop,
513 	.init		= drr_init_qdisc,
514 	.reset		= drr_reset_qdisc,
515 	.destroy	= drr_destroy_qdisc,
516 	.owner		= THIS_MODULE,
517 };
518 
519 static int __init drr_init(void)
520 {
521 	return register_qdisc(&drr_qdisc_ops);
522 }
523 
524 static void __exit drr_exit(void)
525 {
526 	unregister_qdisc(&drr_qdisc_ops);
527 }
528 
529 module_init(drr_init);
530 module_exit(drr_exit);
531 MODULE_LICENSE("GPL");
532