xref: /openbmc/linux/net/sched/sch_cbs.c (revision 95298d63)
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 	/* The previous packet is still being sent */
185 	if (now < q->last) {
186 		qdisc_watchdog_schedule_ns(&q->watchdog, q->last);
187 		return NULL;
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 	/* Estimate of the transmission of the last byte of the packet in ns */
221 	if (unlikely(atomic64_read(&q->port_rate) == 0))
222 		q->last = now;
223 	else
224 		q->last = now + div64_s64(len * NSEC_PER_SEC,
225 					  atomic64_read(&q->port_rate));
226 
227 	return skb;
228 }
229 
230 static struct sk_buff *cbs_dequeue_offload(struct Qdisc *sch)
231 {
232 	struct cbs_sched_data *q = qdisc_priv(sch);
233 	struct Qdisc *qdisc = q->qdisc;
234 
235 	return cbs_child_dequeue(sch, qdisc);
236 }
237 
238 static struct sk_buff *cbs_dequeue(struct Qdisc *sch)
239 {
240 	struct cbs_sched_data *q = qdisc_priv(sch);
241 
242 	return q->dequeue(sch);
243 }
244 
245 static const struct nla_policy cbs_policy[TCA_CBS_MAX + 1] = {
246 	[TCA_CBS_PARMS]	= { .len = sizeof(struct tc_cbs_qopt) },
247 };
248 
249 static void cbs_disable_offload(struct net_device *dev,
250 				struct cbs_sched_data *q)
251 {
252 	struct tc_cbs_qopt_offload cbs = { };
253 	const struct net_device_ops *ops;
254 	int err;
255 
256 	if (!q->offload)
257 		return;
258 
259 	q->enqueue = cbs_enqueue_soft;
260 	q->dequeue = cbs_dequeue_soft;
261 
262 	ops = dev->netdev_ops;
263 	if (!ops->ndo_setup_tc)
264 		return;
265 
266 	cbs.queue = q->queue;
267 	cbs.enable = 0;
268 
269 	err = ops->ndo_setup_tc(dev, TC_SETUP_QDISC_CBS, &cbs);
270 	if (err < 0)
271 		pr_warn("Couldn't disable CBS offload for queue %d\n",
272 			cbs.queue);
273 }
274 
275 static int cbs_enable_offload(struct net_device *dev, struct cbs_sched_data *q,
276 			      const struct tc_cbs_qopt *opt,
277 			      struct netlink_ext_ack *extack)
278 {
279 	const struct net_device_ops *ops = dev->netdev_ops;
280 	struct tc_cbs_qopt_offload cbs = { };
281 	int err;
282 
283 	if (!ops->ndo_setup_tc) {
284 		NL_SET_ERR_MSG(extack, "Specified device does not support cbs offload");
285 		return -EOPNOTSUPP;
286 	}
287 
288 	cbs.queue = q->queue;
289 
290 	cbs.enable = 1;
291 	cbs.hicredit = opt->hicredit;
292 	cbs.locredit = opt->locredit;
293 	cbs.idleslope = opt->idleslope;
294 	cbs.sendslope = opt->sendslope;
295 
296 	err = ops->ndo_setup_tc(dev, TC_SETUP_QDISC_CBS, &cbs);
297 	if (err < 0) {
298 		NL_SET_ERR_MSG(extack, "Specified device failed to setup cbs hardware offload");
299 		return err;
300 	}
301 
302 	q->enqueue = cbs_enqueue_offload;
303 	q->dequeue = cbs_dequeue_offload;
304 
305 	return 0;
306 }
307 
308 static void cbs_set_port_rate(struct net_device *dev, struct cbs_sched_data *q)
309 {
310 	struct ethtool_link_ksettings ecmd;
311 	int speed = SPEED_10;
312 	int port_rate;
313 	int err;
314 
315 	err = __ethtool_get_link_ksettings(dev, &ecmd);
316 	if (err < 0)
317 		goto skip;
318 
319 	if (ecmd.base.speed && ecmd.base.speed != SPEED_UNKNOWN)
320 		speed = ecmd.base.speed;
321 
322 skip:
323 	port_rate = speed * 1000 * BYTES_PER_KBIT;
324 
325 	atomic64_set(&q->port_rate, port_rate);
326 	netdev_dbg(dev, "cbs: set %s's port_rate to: %lld, linkspeed: %d\n",
327 		   dev->name, (long long)atomic64_read(&q->port_rate),
328 		   ecmd.base.speed);
329 }
330 
331 static int cbs_dev_notifier(struct notifier_block *nb, unsigned long event,
332 			    void *ptr)
333 {
334 	struct net_device *dev = netdev_notifier_info_to_dev(ptr);
335 	struct cbs_sched_data *q;
336 	struct net_device *qdev;
337 	bool found = false;
338 
339 	ASSERT_RTNL();
340 
341 	if (event != NETDEV_UP && event != NETDEV_CHANGE)
342 		return NOTIFY_DONE;
343 
344 	spin_lock(&cbs_list_lock);
345 	list_for_each_entry(q, &cbs_list, cbs_list) {
346 		qdev = qdisc_dev(q->qdisc);
347 		if (qdev == dev) {
348 			found = true;
349 			break;
350 		}
351 	}
352 	spin_unlock(&cbs_list_lock);
353 
354 	if (found)
355 		cbs_set_port_rate(dev, q);
356 
357 	return NOTIFY_DONE;
358 }
359 
360 static int cbs_change(struct Qdisc *sch, struct nlattr *opt,
361 		      struct netlink_ext_ack *extack)
362 {
363 	struct cbs_sched_data *q = qdisc_priv(sch);
364 	struct net_device *dev = qdisc_dev(sch);
365 	struct nlattr *tb[TCA_CBS_MAX + 1];
366 	struct tc_cbs_qopt *qopt;
367 	int err;
368 
369 	err = nla_parse_nested_deprecated(tb, TCA_CBS_MAX, opt, cbs_policy,
370 					  extack);
371 	if (err < 0)
372 		return err;
373 
374 	if (!tb[TCA_CBS_PARMS]) {
375 		NL_SET_ERR_MSG(extack, "Missing CBS parameter which are mandatory");
376 		return -EINVAL;
377 	}
378 
379 	qopt = nla_data(tb[TCA_CBS_PARMS]);
380 
381 	if (!qopt->offload) {
382 		cbs_set_port_rate(dev, q);
383 		cbs_disable_offload(dev, q);
384 	} else {
385 		err = cbs_enable_offload(dev, q, qopt, extack);
386 		if (err < 0)
387 			return err;
388 	}
389 
390 	/* Everything went OK, save the parameters used. */
391 	q->hicredit = qopt->hicredit;
392 	q->locredit = qopt->locredit;
393 	q->idleslope = qopt->idleslope * BYTES_PER_KBIT;
394 	q->sendslope = qopt->sendslope * BYTES_PER_KBIT;
395 	q->offload = qopt->offload;
396 
397 	return 0;
398 }
399 
400 static int cbs_init(struct Qdisc *sch, struct nlattr *opt,
401 		    struct netlink_ext_ack *extack)
402 {
403 	struct cbs_sched_data *q = qdisc_priv(sch);
404 	struct net_device *dev = qdisc_dev(sch);
405 
406 	if (!opt) {
407 		NL_SET_ERR_MSG(extack, "Missing CBS qdisc options  which are mandatory");
408 		return -EINVAL;
409 	}
410 
411 	q->qdisc = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
412 				     sch->handle, extack);
413 	if (!q->qdisc)
414 		return -ENOMEM;
415 
416 	spin_lock(&cbs_list_lock);
417 	list_add(&q->cbs_list, &cbs_list);
418 	spin_unlock(&cbs_list_lock);
419 
420 	qdisc_hash_add(q->qdisc, false);
421 
422 	q->queue = sch->dev_queue - netdev_get_tx_queue(dev, 0);
423 
424 	q->enqueue = cbs_enqueue_soft;
425 	q->dequeue = cbs_dequeue_soft;
426 
427 	qdisc_watchdog_init(&q->watchdog, sch);
428 
429 	return cbs_change(sch, opt, extack);
430 }
431 
432 static void cbs_destroy(struct Qdisc *sch)
433 {
434 	struct cbs_sched_data *q = qdisc_priv(sch);
435 	struct net_device *dev = qdisc_dev(sch);
436 
437 	/* Nothing to do if we couldn't create the underlying qdisc */
438 	if (!q->qdisc)
439 		return;
440 
441 	qdisc_watchdog_cancel(&q->watchdog);
442 	cbs_disable_offload(dev, q);
443 
444 	spin_lock(&cbs_list_lock);
445 	list_del(&q->cbs_list);
446 	spin_unlock(&cbs_list_lock);
447 
448 	qdisc_put(q->qdisc);
449 }
450 
451 static int cbs_dump(struct Qdisc *sch, struct sk_buff *skb)
452 {
453 	struct cbs_sched_data *q = qdisc_priv(sch);
454 	struct tc_cbs_qopt opt = { };
455 	struct nlattr *nest;
456 
457 	nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
458 	if (!nest)
459 		goto nla_put_failure;
460 
461 	opt.hicredit = q->hicredit;
462 	opt.locredit = q->locredit;
463 	opt.sendslope = div64_s64(q->sendslope, BYTES_PER_KBIT);
464 	opt.idleslope = div64_s64(q->idleslope, BYTES_PER_KBIT);
465 	opt.offload = q->offload;
466 
467 	if (nla_put(skb, TCA_CBS_PARMS, sizeof(opt), &opt))
468 		goto nla_put_failure;
469 
470 	return nla_nest_end(skb, nest);
471 
472 nla_put_failure:
473 	nla_nest_cancel(skb, nest);
474 	return -1;
475 }
476 
477 static int cbs_dump_class(struct Qdisc *sch, unsigned long cl,
478 			  struct sk_buff *skb, struct tcmsg *tcm)
479 {
480 	struct cbs_sched_data *q = qdisc_priv(sch);
481 
482 	if (cl != 1 || !q->qdisc)	/* only one class */
483 		return -ENOENT;
484 
485 	tcm->tcm_handle |= TC_H_MIN(1);
486 	tcm->tcm_info = q->qdisc->handle;
487 
488 	return 0;
489 }
490 
491 static int cbs_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
492 		     struct Qdisc **old, struct netlink_ext_ack *extack)
493 {
494 	struct cbs_sched_data *q = qdisc_priv(sch);
495 
496 	if (!new) {
497 		new = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
498 					sch->handle, NULL);
499 		if (!new)
500 			new = &noop_qdisc;
501 	}
502 
503 	*old = qdisc_replace(sch, new, &q->qdisc);
504 	return 0;
505 }
506 
507 static struct Qdisc *cbs_leaf(struct Qdisc *sch, unsigned long arg)
508 {
509 	struct cbs_sched_data *q = qdisc_priv(sch);
510 
511 	return q->qdisc;
512 }
513 
514 static unsigned long cbs_find(struct Qdisc *sch, u32 classid)
515 {
516 	return 1;
517 }
518 
519 static void cbs_walk(struct Qdisc *sch, struct qdisc_walker *walker)
520 {
521 	if (!walker->stop) {
522 		if (walker->count >= walker->skip) {
523 			if (walker->fn(sch, 1, walker) < 0) {
524 				walker->stop = 1;
525 				return;
526 			}
527 		}
528 		walker->count++;
529 	}
530 }
531 
532 static const struct Qdisc_class_ops cbs_class_ops = {
533 	.graft		=	cbs_graft,
534 	.leaf		=	cbs_leaf,
535 	.find		=	cbs_find,
536 	.walk		=	cbs_walk,
537 	.dump		=	cbs_dump_class,
538 };
539 
540 static struct Qdisc_ops cbs_qdisc_ops __read_mostly = {
541 	.id		=	"cbs",
542 	.cl_ops		=	&cbs_class_ops,
543 	.priv_size	=	sizeof(struct cbs_sched_data),
544 	.enqueue	=	cbs_enqueue,
545 	.dequeue	=	cbs_dequeue,
546 	.peek		=	qdisc_peek_dequeued,
547 	.init		=	cbs_init,
548 	.reset		=	qdisc_reset_queue,
549 	.destroy	=	cbs_destroy,
550 	.change		=	cbs_change,
551 	.dump		=	cbs_dump,
552 	.owner		=	THIS_MODULE,
553 };
554 
555 static struct notifier_block cbs_device_notifier = {
556 	.notifier_call = cbs_dev_notifier,
557 };
558 
559 static int __init cbs_module_init(void)
560 {
561 	int err;
562 
563 	err = register_netdevice_notifier(&cbs_device_notifier);
564 	if (err)
565 		return err;
566 
567 	err = register_qdisc(&cbs_qdisc_ops);
568 	if (err)
569 		unregister_netdevice_notifier(&cbs_device_notifier);
570 
571 	return err;
572 }
573 
574 static void __exit cbs_module_exit(void)
575 {
576 	unregister_qdisc(&cbs_qdisc_ops);
577 	unregister_netdevice_notifier(&cbs_device_notifier);
578 }
579 module_init(cbs_module_init)
580 module_exit(cbs_module_exit)
581 MODULE_LICENSE("GPL");
582