xref: /linux/net/sched/sch_cbs.c (revision 2dbc0838bcf24ca59cabc3130cf3b1d6809cdcd4)
1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3  * net/sched/sch_cbs.c	Credit Based Shaper
4  *
5  * Authors:	Vinicius Costa Gomes <vinicius.gomes@intel.com>
6  */
7 
8 /* Credit Based Shaper (CBS)
9  * =========================
10  *
11  * This is a simple rate-limiting shaper aimed at TSN applications on
12  * systems with known traffic workloads.
13  *
14  * Its algorithm is defined by the IEEE 802.1Q-2014 Specification,
15  * Section 8.6.8.2, and explained in more detail in the Annex L of the
16  * same specification.
17  *
18  * There are four tunables to be considered:
19  *
20  *	'idleslope': Idleslope is the rate of credits that is
21  *	accumulated (in kilobits per second) when there is at least
22  *	one packet waiting for transmission. Packets are transmitted
23  *	when the current value of credits is equal or greater than
24  *	zero. When there is no packet to be transmitted the amount of
25  *	credits is set to zero. This is the main tunable of the CBS
26  *	algorithm.
27  *
28  *	'sendslope':
29  *	Sendslope is the rate of credits that is depleted (it should be a
30  *	negative number of kilobits per second) when a transmission is
31  *	ocurring. It can be calculated as follows, (IEEE 802.1Q-2014 Section
32  *	8.6.8.2 item g):
33  *
34  *	sendslope = idleslope - port_transmit_rate
35  *
36  *	'hicredit': Hicredit defines the maximum amount of credits (in
37  *	bytes) that can be accumulated. Hicredit depends on the
38  *	characteristics of interfering traffic,
39  *	'max_interference_size' is the maximum size of any burst of
40  *	traffic that can delay the transmission of a frame that is
41  *	available for transmission for this traffic class, (IEEE
42  *	802.1Q-2014 Annex L, Equation L-3):
43  *
44  *	hicredit = max_interference_size * (idleslope / port_transmit_rate)
45  *
46  *	'locredit': Locredit is the minimum amount of credits that can
47  *	be reached. It is a function of the traffic flowing through
48  *	this qdisc (IEEE 802.1Q-2014 Annex L, Equation L-2):
49  *
50  *	locredit = max_frame_size * (sendslope / port_transmit_rate)
51  */
52 
53 #include <linux/module.h>
54 #include <linux/types.h>
55 #include <linux/kernel.h>
56 #include <linux/string.h>
57 #include <linux/errno.h>
58 #include <linux/skbuff.h>
59 #include <net/netevent.h>
60 #include <net/netlink.h>
61 #include <net/sch_generic.h>
62 #include <net/pkt_sched.h>
63 
64 static LIST_HEAD(cbs_list);
65 static DEFINE_SPINLOCK(cbs_list_lock);
66 
67 #define BYTES_PER_KBIT (1000LL / 8)
68 
69 struct cbs_sched_data {
70 	bool offload;
71 	int queue;
72 	atomic64_t port_rate; /* in bytes/s */
73 	s64 last; /* timestamp in ns */
74 	s64 credits; /* in bytes */
75 	s32 locredit; /* in bytes */
76 	s32 hicredit; /* in bytes */
77 	s64 sendslope; /* in bytes/s */
78 	s64 idleslope; /* in bytes/s */
79 	struct qdisc_watchdog watchdog;
80 	int (*enqueue)(struct sk_buff *skb, struct Qdisc *sch,
81 		       struct sk_buff **to_free);
82 	struct sk_buff *(*dequeue)(struct Qdisc *sch);
83 	struct Qdisc *qdisc;
84 	struct list_head cbs_list;
85 };
86 
87 static int cbs_child_enqueue(struct sk_buff *skb, struct Qdisc *sch,
88 			     struct Qdisc *child,
89 			     struct sk_buff **to_free)
90 {
91 	unsigned int len = qdisc_pkt_len(skb);
92 	int err;
93 
94 	err = child->ops->enqueue(skb, child, to_free);
95 	if (err != NET_XMIT_SUCCESS)
96 		return err;
97 
98 	sch->qstats.backlog += len;
99 	sch->q.qlen++;
100 
101 	return NET_XMIT_SUCCESS;
102 }
103 
104 static int cbs_enqueue_offload(struct sk_buff *skb, struct Qdisc *sch,
105 			       struct sk_buff **to_free)
106 {
107 	struct cbs_sched_data *q = qdisc_priv(sch);
108 	struct Qdisc *qdisc = q->qdisc;
109 
110 	return cbs_child_enqueue(skb, sch, qdisc, to_free);
111 }
112 
113 static int cbs_enqueue_soft(struct sk_buff *skb, struct Qdisc *sch,
114 			    struct sk_buff **to_free)
115 {
116 	struct cbs_sched_data *q = qdisc_priv(sch);
117 	struct Qdisc *qdisc = q->qdisc;
118 
119 	if (sch->q.qlen == 0 && q->credits > 0) {
120 		/* We need to stop accumulating credits when there's
121 		 * no enqueued packets and q->credits is positive.
122 		 */
123 		q->credits = 0;
124 		q->last = ktime_get_ns();
125 	}
126 
127 	return cbs_child_enqueue(skb, sch, qdisc, to_free);
128 }
129 
130 static int cbs_enqueue(struct sk_buff *skb, struct Qdisc *sch,
131 		       struct sk_buff **to_free)
132 {
133 	struct cbs_sched_data *q = qdisc_priv(sch);
134 
135 	return q->enqueue(skb, sch, to_free);
136 }
137 
138 /* timediff is in ns, slope is in bytes/s */
139 static s64 timediff_to_credits(s64 timediff, s64 slope)
140 {
141 	return div64_s64(timediff * slope, NSEC_PER_SEC);
142 }
143 
144 static s64 delay_from_credits(s64 credits, s64 slope)
145 {
146 	if (unlikely(slope == 0))
147 		return S64_MAX;
148 
149 	return div64_s64(-credits * NSEC_PER_SEC, slope);
150 }
151 
152 static s64 credits_from_len(unsigned int len, s64 slope, s64 port_rate)
153 {
154 	if (unlikely(port_rate == 0))
155 		return S64_MAX;
156 
157 	return div64_s64(len * slope, port_rate);
158 }
159 
160 static struct sk_buff *cbs_child_dequeue(struct Qdisc *sch, struct Qdisc *child)
161 {
162 	struct sk_buff *skb;
163 
164 	skb = child->ops->dequeue(child);
165 	if (!skb)
166 		return NULL;
167 
168 	qdisc_qstats_backlog_dec(sch, skb);
169 	qdisc_bstats_update(sch, skb);
170 	sch->q.qlen--;
171 
172 	return skb;
173 }
174 
175 static struct sk_buff *cbs_dequeue_soft(struct Qdisc *sch)
176 {
177 	struct cbs_sched_data *q = qdisc_priv(sch);
178 	struct Qdisc *qdisc = q->qdisc;
179 	s64 now = ktime_get_ns();
180 	struct sk_buff *skb;
181 	s64 credits;
182 	int len;
183 
184 	if (atomic64_read(&q->port_rate) == -1) {
185 		WARN_ONCE(1, "cbs: dequeue() called with unknown port rate.");
186 		return NULL;
187 	}
188 
189 	if (q->credits < 0) {
190 		credits = timediff_to_credits(now - q->last, q->idleslope);
191 
192 		credits = q->credits + credits;
193 		q->credits = min_t(s64, credits, q->hicredit);
194 
195 		if (q->credits < 0) {
196 			s64 delay;
197 
198 			delay = delay_from_credits(q->credits, q->idleslope);
199 			qdisc_watchdog_schedule_ns(&q->watchdog, now + delay);
200 
201 			q->last = now;
202 
203 			return NULL;
204 		}
205 	}
206 	skb = cbs_child_dequeue(sch, qdisc);
207 	if (!skb)
208 		return NULL;
209 
210 	len = qdisc_pkt_len(skb);
211 
212 	/* As sendslope is a negative number, this will decrease the
213 	 * amount of q->credits.
214 	 */
215 	credits = credits_from_len(len, q->sendslope,
216 				   atomic64_read(&q->port_rate));
217 	credits += q->credits;
218 
219 	q->credits = max_t(s64, credits, q->locredit);
220 	q->last = now;
221 
222 	return skb;
223 }
224 
225 static struct sk_buff *cbs_dequeue_offload(struct Qdisc *sch)
226 {
227 	struct cbs_sched_data *q = qdisc_priv(sch);
228 	struct Qdisc *qdisc = q->qdisc;
229 
230 	return cbs_child_dequeue(sch, qdisc);
231 }
232 
233 static struct sk_buff *cbs_dequeue(struct Qdisc *sch)
234 {
235 	struct cbs_sched_data *q = qdisc_priv(sch);
236 
237 	return q->dequeue(sch);
238 }
239 
240 static const struct nla_policy cbs_policy[TCA_CBS_MAX + 1] = {
241 	[TCA_CBS_PARMS]	= { .len = sizeof(struct tc_cbs_qopt) },
242 };
243 
244 static void cbs_disable_offload(struct net_device *dev,
245 				struct cbs_sched_data *q)
246 {
247 	struct tc_cbs_qopt_offload cbs = { };
248 	const struct net_device_ops *ops;
249 	int err;
250 
251 	if (!q->offload)
252 		return;
253 
254 	q->enqueue = cbs_enqueue_soft;
255 	q->dequeue = cbs_dequeue_soft;
256 
257 	ops = dev->netdev_ops;
258 	if (!ops->ndo_setup_tc)
259 		return;
260 
261 	cbs.queue = q->queue;
262 	cbs.enable = 0;
263 
264 	err = ops->ndo_setup_tc(dev, TC_SETUP_QDISC_CBS, &cbs);
265 	if (err < 0)
266 		pr_warn("Couldn't disable CBS offload for queue %d\n",
267 			cbs.queue);
268 }
269 
270 static int cbs_enable_offload(struct net_device *dev, struct cbs_sched_data *q,
271 			      const struct tc_cbs_qopt *opt,
272 			      struct netlink_ext_ack *extack)
273 {
274 	const struct net_device_ops *ops = dev->netdev_ops;
275 	struct tc_cbs_qopt_offload cbs = { };
276 	int err;
277 
278 	if (!ops->ndo_setup_tc) {
279 		NL_SET_ERR_MSG(extack, "Specified device does not support cbs offload");
280 		return -EOPNOTSUPP;
281 	}
282 
283 	cbs.queue = q->queue;
284 
285 	cbs.enable = 1;
286 	cbs.hicredit = opt->hicredit;
287 	cbs.locredit = opt->locredit;
288 	cbs.idleslope = opt->idleslope;
289 	cbs.sendslope = opt->sendslope;
290 
291 	err = ops->ndo_setup_tc(dev, TC_SETUP_QDISC_CBS, &cbs);
292 	if (err < 0) {
293 		NL_SET_ERR_MSG(extack, "Specified device failed to setup cbs hardware offload");
294 		return err;
295 	}
296 
297 	q->enqueue = cbs_enqueue_offload;
298 	q->dequeue = cbs_dequeue_offload;
299 
300 	return 0;
301 }
302 
303 static void cbs_set_port_rate(struct net_device *dev, struct cbs_sched_data *q)
304 {
305 	struct ethtool_link_ksettings ecmd;
306 	int port_rate = -1;
307 
308 	if (!__ethtool_get_link_ksettings(dev, &ecmd) &&
309 	    ecmd.base.speed != SPEED_UNKNOWN)
310 		port_rate = ecmd.base.speed * 1000 * BYTES_PER_KBIT;
311 
312 	atomic64_set(&q->port_rate, port_rate);
313 	netdev_dbg(dev, "cbs: set %s's port_rate to: %lld, linkspeed: %d\n",
314 		   dev->name, (long long)atomic64_read(&q->port_rate),
315 		   ecmd.base.speed);
316 }
317 
318 static int cbs_dev_notifier(struct notifier_block *nb, unsigned long event,
319 			    void *ptr)
320 {
321 	struct net_device *dev = netdev_notifier_info_to_dev(ptr);
322 	struct cbs_sched_data *q;
323 	struct net_device *qdev;
324 	bool found = false;
325 
326 	ASSERT_RTNL();
327 
328 	if (event != NETDEV_UP && event != NETDEV_CHANGE)
329 		return NOTIFY_DONE;
330 
331 	spin_lock(&cbs_list_lock);
332 	list_for_each_entry(q, &cbs_list, cbs_list) {
333 		qdev = qdisc_dev(q->qdisc);
334 		if (qdev == dev) {
335 			found = true;
336 			break;
337 		}
338 	}
339 	spin_unlock(&cbs_list_lock);
340 
341 	if (found)
342 		cbs_set_port_rate(dev, q);
343 
344 	return NOTIFY_DONE;
345 }
346 
347 static int cbs_change(struct Qdisc *sch, struct nlattr *opt,
348 		      struct netlink_ext_ack *extack)
349 {
350 	struct cbs_sched_data *q = qdisc_priv(sch);
351 	struct net_device *dev = qdisc_dev(sch);
352 	struct nlattr *tb[TCA_CBS_MAX + 1];
353 	struct tc_cbs_qopt *qopt;
354 	int err;
355 
356 	err = nla_parse_nested_deprecated(tb, TCA_CBS_MAX, opt, cbs_policy,
357 					  extack);
358 	if (err < 0)
359 		return err;
360 
361 	if (!tb[TCA_CBS_PARMS]) {
362 		NL_SET_ERR_MSG(extack, "Missing CBS parameter which are mandatory");
363 		return -EINVAL;
364 	}
365 
366 	qopt = nla_data(tb[TCA_CBS_PARMS]);
367 
368 	if (!qopt->offload) {
369 		cbs_set_port_rate(dev, q);
370 		cbs_disable_offload(dev, q);
371 	} else {
372 		err = cbs_enable_offload(dev, q, qopt, extack);
373 		if (err < 0)
374 			return err;
375 	}
376 
377 	/* Everything went OK, save the parameters used. */
378 	q->hicredit = qopt->hicredit;
379 	q->locredit = qopt->locredit;
380 	q->idleslope = qopt->idleslope * BYTES_PER_KBIT;
381 	q->sendslope = qopt->sendslope * BYTES_PER_KBIT;
382 	q->offload = qopt->offload;
383 
384 	return 0;
385 }
386 
387 static int cbs_init(struct Qdisc *sch, struct nlattr *opt,
388 		    struct netlink_ext_ack *extack)
389 {
390 	struct cbs_sched_data *q = qdisc_priv(sch);
391 	struct net_device *dev = qdisc_dev(sch);
392 	int err;
393 
394 	if (!opt) {
395 		NL_SET_ERR_MSG(extack, "Missing CBS qdisc options  which are mandatory");
396 		return -EINVAL;
397 	}
398 
399 	q->qdisc = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
400 				     sch->handle, extack);
401 	if (!q->qdisc)
402 		return -ENOMEM;
403 
404 	qdisc_hash_add(q->qdisc, false);
405 
406 	q->queue = sch->dev_queue - netdev_get_tx_queue(dev, 0);
407 
408 	q->enqueue = cbs_enqueue_soft;
409 	q->dequeue = cbs_dequeue_soft;
410 
411 	qdisc_watchdog_init(&q->watchdog, sch);
412 
413 	err = cbs_change(sch, opt, extack);
414 	if (err)
415 		return err;
416 
417 	if (!q->offload) {
418 		spin_lock(&cbs_list_lock);
419 		list_add(&q->cbs_list, &cbs_list);
420 		spin_unlock(&cbs_list_lock);
421 	}
422 
423 	return 0;
424 }
425 
426 static void cbs_destroy(struct Qdisc *sch)
427 {
428 	struct cbs_sched_data *q = qdisc_priv(sch);
429 	struct net_device *dev = qdisc_dev(sch);
430 
431 	spin_lock(&cbs_list_lock);
432 	list_del(&q->cbs_list);
433 	spin_unlock(&cbs_list_lock);
434 
435 	qdisc_watchdog_cancel(&q->watchdog);
436 	cbs_disable_offload(dev, q);
437 
438 	if (q->qdisc)
439 		qdisc_put(q->qdisc);
440 }
441 
442 static int cbs_dump(struct Qdisc *sch, struct sk_buff *skb)
443 {
444 	struct cbs_sched_data *q = qdisc_priv(sch);
445 	struct tc_cbs_qopt opt = { };
446 	struct nlattr *nest;
447 
448 	nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
449 	if (!nest)
450 		goto nla_put_failure;
451 
452 	opt.hicredit = q->hicredit;
453 	opt.locredit = q->locredit;
454 	opt.sendslope = div64_s64(q->sendslope, BYTES_PER_KBIT);
455 	opt.idleslope = div64_s64(q->idleslope, BYTES_PER_KBIT);
456 	opt.offload = q->offload;
457 
458 	if (nla_put(skb, TCA_CBS_PARMS, sizeof(opt), &opt))
459 		goto nla_put_failure;
460 
461 	return nla_nest_end(skb, nest);
462 
463 nla_put_failure:
464 	nla_nest_cancel(skb, nest);
465 	return -1;
466 }
467 
468 static int cbs_dump_class(struct Qdisc *sch, unsigned long cl,
469 			  struct sk_buff *skb, struct tcmsg *tcm)
470 {
471 	struct cbs_sched_data *q = qdisc_priv(sch);
472 
473 	if (cl != 1 || !q->qdisc)	/* only one class */
474 		return -ENOENT;
475 
476 	tcm->tcm_handle |= TC_H_MIN(1);
477 	tcm->tcm_info = q->qdisc->handle;
478 
479 	return 0;
480 }
481 
482 static int cbs_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
483 		     struct Qdisc **old, struct netlink_ext_ack *extack)
484 {
485 	struct cbs_sched_data *q = qdisc_priv(sch);
486 
487 	if (!new) {
488 		new = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
489 					sch->handle, NULL);
490 		if (!new)
491 			new = &noop_qdisc;
492 	}
493 
494 	*old = qdisc_replace(sch, new, &q->qdisc);
495 	return 0;
496 }
497 
498 static struct Qdisc *cbs_leaf(struct Qdisc *sch, unsigned long arg)
499 {
500 	struct cbs_sched_data *q = qdisc_priv(sch);
501 
502 	return q->qdisc;
503 }
504 
505 static unsigned long cbs_find(struct Qdisc *sch, u32 classid)
506 {
507 	return 1;
508 }
509 
510 static void cbs_walk(struct Qdisc *sch, struct qdisc_walker *walker)
511 {
512 	if (!walker->stop) {
513 		if (walker->count >= walker->skip) {
514 			if (walker->fn(sch, 1, walker) < 0) {
515 				walker->stop = 1;
516 				return;
517 			}
518 		}
519 		walker->count++;
520 	}
521 }
522 
523 static const struct Qdisc_class_ops cbs_class_ops = {
524 	.graft		=	cbs_graft,
525 	.leaf		=	cbs_leaf,
526 	.find		=	cbs_find,
527 	.walk		=	cbs_walk,
528 	.dump		=	cbs_dump_class,
529 };
530 
531 static struct Qdisc_ops cbs_qdisc_ops __read_mostly = {
532 	.id		=	"cbs",
533 	.cl_ops		=	&cbs_class_ops,
534 	.priv_size	=	sizeof(struct cbs_sched_data),
535 	.enqueue	=	cbs_enqueue,
536 	.dequeue	=	cbs_dequeue,
537 	.peek		=	qdisc_peek_dequeued,
538 	.init		=	cbs_init,
539 	.reset		=	qdisc_reset_queue,
540 	.destroy	=	cbs_destroy,
541 	.change		=	cbs_change,
542 	.dump		=	cbs_dump,
543 	.owner		=	THIS_MODULE,
544 };
545 
546 static struct notifier_block cbs_device_notifier = {
547 	.notifier_call = cbs_dev_notifier,
548 };
549 
550 static int __init cbs_module_init(void)
551 {
552 	int err;
553 
554 	err = register_netdevice_notifier(&cbs_device_notifier);
555 	if (err)
556 		return err;
557 
558 	err = register_qdisc(&cbs_qdisc_ops);
559 	if (err)
560 		unregister_netdevice_notifier(&cbs_device_notifier);
561 
562 	return err;
563 }
564 
565 static void __exit cbs_module_exit(void)
566 {
567 	unregister_qdisc(&cbs_qdisc_ops);
568 	unregister_netdevice_notifier(&cbs_device_notifier);
569 }
570 module_init(cbs_module_init)
571 module_exit(cbs_module_exit)
572 MODULE_LICENSE("GPL");
573