xref: /openbmc/linux/net/sched/sch_etf.c (revision a17922de) 
1 // SPDX-License-Identifier: GPL-2.0
2 
3 /* net/sched/sch_etf.c  Earliest TxTime First queueing discipline.
4  *
5  * Authors:	Jesus Sanchez-Palencia <jesus.sanchez-palencia@intel.com>
6  *		Vinicius Costa Gomes <vinicius.gomes@intel.com>
7  */
8 
9 #include <linux/module.h>
10 #include <linux/types.h>
11 #include <linux/kernel.h>
12 #include <linux/string.h>
13 #include <linux/errno.h>
14 #include <linux/errqueue.h>
15 #include <linux/rbtree.h>
16 #include <linux/skbuff.h>
17 #include <linux/posix-timers.h>
18 #include <net/netlink.h>
19 #include <net/sch_generic.h>
20 #include <net/pkt_sched.h>
21 #include <net/sock.h>
22 
23 #define DEADLINE_MODE_IS_ON(x) ((x)->flags & TC_ETF_DEADLINE_MODE_ON)
24 #define OFFLOAD_IS_ON(x) ((x)->flags & TC_ETF_OFFLOAD_ON)
25 
26 struct etf_sched_data {
27 	bool offload;
28 	bool deadline_mode;
29 	int clockid;
30 	int queue;
31 	s32 delta; /* in ns */
32 	ktime_t last; /* The txtime of the last skb sent to the netdevice. */
33 	struct rb_root head;
34 	struct qdisc_watchdog watchdog;
35 	ktime_t (*get_time)(void);
36 };
37 
38 static const struct nla_policy etf_policy[TCA_ETF_MAX + 1] = {
39 	[TCA_ETF_PARMS]	= { .len = sizeof(struct tc_etf_qopt) },
40 };
41 
42 static inline int validate_input_params(struct tc_etf_qopt *qopt,
43 					struct netlink_ext_ack *extack)
44 {
45 	/* Check if params comply to the following rules:
46 	 *	* Clockid and delta must be valid.
47 	 *
48 	 *	* Dynamic clockids are not supported.
49 	 *
50 	 *	* Delta must be a positive integer.
51 	 *
52 	 * Also note that for the HW offload case, we must
53 	 * expect that system clocks have been synchronized to PHC.
54 	 */
55 	if (qopt->clockid < 0) {
56 		NL_SET_ERR_MSG(extack, "Dynamic clockids are not supported");
57 		return -ENOTSUPP;
58 	}
59 
60 	if (qopt->clockid != CLOCK_TAI) {
61 		NL_SET_ERR_MSG(extack, "Invalid clockid. CLOCK_TAI must be used");
62 		return -EINVAL;
63 	}
64 
65 	if (qopt->delta < 0) {
66 		NL_SET_ERR_MSG(extack, "Delta must be positive");
67 		return -EINVAL;
68 	}
69 
70 	return 0;
71 }
72 
73 static bool is_packet_valid(struct Qdisc *sch, struct sk_buff *nskb)
74 {
75 	struct etf_sched_data *q = qdisc_priv(sch);
76 	ktime_t txtime = nskb->tstamp;
77 	struct sock *sk = nskb->sk;
78 	ktime_t now;
79 
80 	if (!sk)
81 		return false;
82 
83 	if (!sock_flag(sk, SOCK_TXTIME))
84 		return false;
85 
86 	/* We don't perform crosstimestamping.
87 	 * Drop if packet's clockid differs from qdisc's.
88 	 */
89 	if (sk->sk_clockid != q->clockid)
90 		return false;
91 
92 	if (sk->sk_txtime_deadline_mode != q->deadline_mode)
93 		return false;
94 
95 	now = q->get_time();
96 	if (ktime_before(txtime, now) || ktime_before(txtime, q->last))
97 		return false;
98 
99 	return true;
100 }
101 
102 static struct sk_buff *etf_peek_timesortedlist(struct Qdisc *sch)
103 {
104 	struct etf_sched_data *q = qdisc_priv(sch);
105 	struct rb_node *p;
106 
107 	p = rb_first(&q->head);
108 	if (!p)
109 		return NULL;
110 
111 	return rb_to_skb(p);
112 }
113 
114 static void reset_watchdog(struct Qdisc *sch)
115 {
116 	struct etf_sched_data *q = qdisc_priv(sch);
117 	struct sk_buff *skb = etf_peek_timesortedlist(sch);
118 	ktime_t next;
119 
120 	if (!skb)
121 		return;
122 
123 	next = ktime_sub_ns(skb->tstamp, q->delta);
124 	qdisc_watchdog_schedule_ns(&q->watchdog, ktime_to_ns(next));
125 }
126 
127 static void report_sock_error(struct sk_buff *skb, u32 err, u8 code)
128 {
129 	struct sock_exterr_skb *serr;
130 	struct sk_buff *clone;
131 	ktime_t txtime = skb->tstamp;
132 
133 	if (!skb->sk || !(skb->sk->sk_txtime_report_errors))
134 		return;
135 
136 	clone = skb_clone(skb, GFP_ATOMIC);
137 	if (!clone)
138 		return;
139 
140 	serr = SKB_EXT_ERR(clone);
141 	serr->ee.ee_errno = err;
142 	serr->ee.ee_origin = SO_EE_ORIGIN_TXTIME;
143 	serr->ee.ee_type = 0;
144 	serr->ee.ee_code = code;
145 	serr->ee.ee_pad = 0;
146 	serr->ee.ee_data = (txtime >> 32); /* high part of tstamp */
147 	serr->ee.ee_info = txtime; /* low part of tstamp */
148 
149 	if (sock_queue_err_skb(skb->sk, clone))
150 		kfree_skb(clone);
151 }
152 
153 static int etf_enqueue_timesortedlist(struct sk_buff *nskb, struct Qdisc *sch,
154 				      struct sk_buff **to_free)
155 {
156 	struct etf_sched_data *q = qdisc_priv(sch);
157 	struct rb_node **p = &q->head.rb_node, *parent = NULL;
158 	ktime_t txtime = nskb->tstamp;
159 
160 	if (!is_packet_valid(sch, nskb)) {
161 		report_sock_error(nskb, EINVAL,
162 				  SO_EE_CODE_TXTIME_INVALID_PARAM);
163 		return qdisc_drop(nskb, sch, to_free);
164 	}
165 
166 	while (*p) {
167 		struct sk_buff *skb;
168 
169 		parent = *p;
170 		skb = rb_to_skb(parent);
171 		if (ktime_after(txtime, skb->tstamp))
172 			p = &parent->rb_right;
173 		else
174 			p = &parent->rb_left;
175 	}
176 	rb_link_node(&nskb->rbnode, parent, p);
177 	rb_insert_color(&nskb->rbnode, &q->head);
178 
179 	qdisc_qstats_backlog_inc(sch, nskb);
180 	sch->q.qlen++;
181 
182 	/* Now we may need to re-arm the qdisc watchdog for the next packet. */
183 	reset_watchdog(sch);
184 
185 	return NET_XMIT_SUCCESS;
186 }
187 
188 static void timesortedlist_erase(struct Qdisc *sch, struct sk_buff *skb,
189 				 bool drop)
190 {
191 	struct etf_sched_data *q = qdisc_priv(sch);
192 
193 	rb_erase(&skb->rbnode, &q->head);
194 
195 	/* The rbnode field in the skb re-uses these fields, now that
196 	 * we are done with the rbnode, reset them.
197 	 */
198 	skb->next = NULL;
199 	skb->prev = NULL;
200 	skb->dev = qdisc_dev(sch);
201 
202 	qdisc_qstats_backlog_dec(sch, skb);
203 
204 	if (drop) {
205 		struct sk_buff *to_free = NULL;
206 
207 		report_sock_error(skb, ECANCELED, SO_EE_CODE_TXTIME_MISSED);
208 
209 		qdisc_drop(skb, sch, &to_free);
210 		kfree_skb_list(to_free);
211 		qdisc_qstats_overlimit(sch);
212 	} else {
213 		qdisc_bstats_update(sch, skb);
214 
215 		q->last = skb->tstamp;
216 	}
217 
218 	sch->q.qlen--;
219 }
220 
221 static struct sk_buff *etf_dequeue_timesortedlist(struct Qdisc *sch)
222 {
223 	struct etf_sched_data *q = qdisc_priv(sch);
224 	struct sk_buff *skb;
225 	ktime_t now, next;
226 
227 	skb = etf_peek_timesortedlist(sch);
228 	if (!skb)
229 		return NULL;
230 
231 	now = q->get_time();
232 
233 	/* Drop if packet has expired while in queue. */
234 	if (ktime_before(skb->tstamp, now)) {
235 		timesortedlist_erase(sch, skb, true);
236 		skb = NULL;
237 		goto out;
238 	}
239 
240 	/* When in deadline mode, dequeue as soon as possible and change the
241 	 * txtime from deadline to (now + delta).
242 	 */
243 	if (q->deadline_mode) {
244 		timesortedlist_erase(sch, skb, false);
245 		skb->tstamp = now;
246 		goto out;
247 	}
248 
249 	next = ktime_sub_ns(skb->tstamp, q->delta);
250 
251 	/* Dequeue only if now is within the [txtime - delta, txtime] range. */
252 	if (ktime_after(now, next))
253 		timesortedlist_erase(sch, skb, false);
254 	else
255 		skb = NULL;
256 
257 out:
258 	/* Now we may need to re-arm the qdisc watchdog for the next packet. */
259 	reset_watchdog(sch);
260 
261 	return skb;
262 }
263 
264 static void etf_disable_offload(struct net_device *dev,
265 				struct etf_sched_data *q)
266 {
267 	struct tc_etf_qopt_offload etf = { };
268 	const struct net_device_ops *ops;
269 	int err;
270 
271 	if (!q->offload)
272 		return;
273 
274 	ops = dev->netdev_ops;
275 	if (!ops->ndo_setup_tc)
276 		return;
277 
278 	etf.queue = q->queue;
279 	etf.enable = 0;
280 
281 	err = ops->ndo_setup_tc(dev, TC_SETUP_QDISC_ETF, &etf);
282 	if (err < 0)
283 		pr_warn("Couldn't disable ETF offload for queue %d\n",
284 			etf.queue);
285 }
286 
287 static int etf_enable_offload(struct net_device *dev, struct etf_sched_data *q,
288 			      struct netlink_ext_ack *extack)
289 {
290 	const struct net_device_ops *ops = dev->netdev_ops;
291 	struct tc_etf_qopt_offload etf = { };
292 	int err;
293 
294 	if (q->offload)
295 		return 0;
296 
297 	if (!ops->ndo_setup_tc) {
298 		NL_SET_ERR_MSG(extack, "Specified device does not support ETF offload");
299 		return -EOPNOTSUPP;
300 	}
301 
302 	etf.queue = q->queue;
303 	etf.enable = 1;
304 
305 	err = ops->ndo_setup_tc(dev, TC_SETUP_QDISC_ETF, &etf);
306 	if (err < 0) {
307 		NL_SET_ERR_MSG(extack, "Specified device failed to setup ETF hardware offload");
308 		return err;
309 	}
310 
311 	return 0;
312 }
313 
314 static int etf_init(struct Qdisc *sch, struct nlattr *opt,
315 		    struct netlink_ext_ack *extack)
316 {
317 	struct etf_sched_data *q = qdisc_priv(sch);
318 	struct net_device *dev = qdisc_dev(sch);
319 	struct nlattr *tb[TCA_ETF_MAX + 1];
320 	struct tc_etf_qopt *qopt;
321 	int err;
322 
323 	if (!opt) {
324 		NL_SET_ERR_MSG(extack,
325 			       "Missing ETF qdisc options which are mandatory");
326 		return -EINVAL;
327 	}
328 
329 	err = nla_parse_nested(tb, TCA_ETF_MAX, opt, etf_policy, extack);
330 	if (err < 0)
331 		return err;
332 
333 	if (!tb[TCA_ETF_PARMS]) {
334 		NL_SET_ERR_MSG(extack, "Missing mandatory ETF parameters");
335 		return -EINVAL;
336 	}
337 
338 	qopt = nla_data(tb[TCA_ETF_PARMS]);
339 
340 	pr_debug("delta %d clockid %d offload %s deadline %s\n",
341 		 qopt->delta, qopt->clockid,
342 		 OFFLOAD_IS_ON(qopt) ? "on" : "off",
343 		 DEADLINE_MODE_IS_ON(qopt) ? "on" : "off");
344 
345 	err = validate_input_params(qopt, extack);
346 	if (err < 0)
347 		return err;
348 
349 	q->queue = sch->dev_queue - netdev_get_tx_queue(dev, 0);
350 
351 	if (OFFLOAD_IS_ON(qopt)) {
352 		err = etf_enable_offload(dev, q, extack);
353 		if (err < 0)
354 			return err;
355 	}
356 
357 	/* Everything went OK, save the parameters used. */
358 	q->delta = qopt->delta;
359 	q->clockid = qopt->clockid;
360 	q->offload = OFFLOAD_IS_ON(qopt);
361 	q->deadline_mode = DEADLINE_MODE_IS_ON(qopt);
362 
363 	switch (q->clockid) {
364 	case CLOCK_REALTIME:
365 		q->get_time = ktime_get_real;
366 		break;
367 	case CLOCK_MONOTONIC:
368 		q->get_time = ktime_get;
369 		break;
370 	case CLOCK_BOOTTIME:
371 		q->get_time = ktime_get_boottime;
372 		break;
373 	case CLOCK_TAI:
374 		q->get_time = ktime_get_clocktai;
375 		break;
376 	default:
377 		NL_SET_ERR_MSG(extack, "Clockid is not supported");
378 		return -ENOTSUPP;
379 	}
380 
381 	qdisc_watchdog_init_clockid(&q->watchdog, sch, q->clockid);
382 
383 	return 0;
384 }
385 
386 static void timesortedlist_clear(struct Qdisc *sch)
387 {
388 	struct etf_sched_data *q = qdisc_priv(sch);
389 	struct rb_node *p = rb_first(&q->head);
390 
391 	while (p) {
392 		struct sk_buff *skb = rb_to_skb(p);
393 
394 		p = rb_next(p);
395 
396 		rb_erase(&skb->rbnode, &q->head);
397 		rtnl_kfree_skbs(skb, skb);
398 		sch->q.qlen--;
399 	}
400 }
401 
402 static void etf_reset(struct Qdisc *sch)
403 {
404 	struct etf_sched_data *q = qdisc_priv(sch);
405 
406 	/* Only cancel watchdog if it's been initialized. */
407 	if (q->watchdog.qdisc == sch)
408 		qdisc_watchdog_cancel(&q->watchdog);
409 
410 	/* No matter which mode we are on, it's safe to clear both lists. */
411 	timesortedlist_clear(sch);
412 	__qdisc_reset_queue(&sch->q);
413 
414 	sch->qstats.backlog = 0;
415 	sch->q.qlen = 0;
416 
417 	q->last = 0;
418 }
419 
420 static void etf_destroy(struct Qdisc *sch)
421 {
422 	struct etf_sched_data *q = qdisc_priv(sch);
423 	struct net_device *dev = qdisc_dev(sch);
424 
425 	/* Only cancel watchdog if it's been initialized. */
426 	if (q->watchdog.qdisc == sch)
427 		qdisc_watchdog_cancel(&q->watchdog);
428 
429 	etf_disable_offload(dev, q);
430 }
431 
432 static int etf_dump(struct Qdisc *sch, struct sk_buff *skb)
433 {
434 	struct etf_sched_data *q = qdisc_priv(sch);
435 	struct tc_etf_qopt opt = { };
436 	struct nlattr *nest;
437 
438 	nest = nla_nest_start(skb, TCA_OPTIONS);
439 	if (!nest)
440 		goto nla_put_failure;
441 
442 	opt.delta = q->delta;
443 	opt.clockid = q->clockid;
444 	if (q->offload)
445 		opt.flags |= TC_ETF_OFFLOAD_ON;
446 
447 	if (q->deadline_mode)
448 		opt.flags |= TC_ETF_DEADLINE_MODE_ON;
449 
450 	if (nla_put(skb, TCA_ETF_PARMS, sizeof(opt), &opt))
451 		goto nla_put_failure;
452 
453 	return nla_nest_end(skb, nest);
454 
455 nla_put_failure:
456 	nla_nest_cancel(skb, nest);
457 	return -1;
458 }
459 
460 static struct Qdisc_ops etf_qdisc_ops __read_mostly = {
461 	.id		=	"etf",
462 	.priv_size	=	sizeof(struct etf_sched_data),
463 	.enqueue	=	etf_enqueue_timesortedlist,
464 	.dequeue	=	etf_dequeue_timesortedlist,
465 	.peek		=	etf_peek_timesortedlist,
466 	.init		=	etf_init,
467 	.reset		=	etf_reset,
468 	.destroy	=	etf_destroy,
469 	.dump		=	etf_dump,
470 	.owner		=	THIS_MODULE,
471 };
472 
473 static int __init etf_module_init(void)
474 {
475 	return register_qdisc(&etf_qdisc_ops);
476 }
477 
478 static void __exit etf_module_exit(void)
479 {
480 	unregister_qdisc(&etf_qdisc_ops);
481 }
482 module_init(etf_module_init)
483 module_exit(etf_module_exit)
484 MODULE_LICENSE("GPL");
485