Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * net/sched/sch_ets.c Enhanced Transmission Selection scheduler
4 *
5 * Description
6 * -----------
7 *
8 * The Enhanced Transmission Selection scheduler is a classful queuing
9 * discipline that merges functionality of PRIO and DRR qdiscs in one scheduler.
10 * ETS makes it easy to configure a set of strict and bandwidth-sharing bands to
11 * implement the transmission selection described in 802.1Qaz.
12 *
13 * Although ETS is technically classful, it's not possible to add and remove
14 * classes at will. Instead one specifies number of classes, how many are
15 * PRIO-like and how many DRR-like, and quanta for the latter.
16 *
17 * Algorithm
18 * ---------
19 *
20 * The strict classes, if any, are tried for traffic first: first band 0, if it
21 * has no traffic then band 1, etc.
22 *
23 * When there is no traffic in any of the strict queues, the bandwidth-sharing
24 * ones are tried next. Each band is assigned a deficit counter, initialized to
25 * "quantum" of that band. ETS maintains a list of active bandwidth-sharing
26 * bands whose qdiscs are non-empty. A packet is dequeued from the band at the
27 * head of the list if the packet size is smaller or equal to the deficit
28 * counter. If the counter is too small, it is increased by "quantum" and the
29 * scheduler moves on to the next band in the active list.
30 */
31
32#include <linux/module.h>
33#include <net/gen_stats.h>
34#include <net/netlink.h>
35#include <net/pkt_cls.h>
36#include <net/pkt_sched.h>
37#include <net/sch_generic.h>
38
39struct ets_class {
40 struct list_head alist; /* In struct ets_sched.active. */
41 struct Qdisc *qdisc;
42 u32 quantum;
43 u32 deficit;
44 struct gnet_stats_basic_sync bstats;
45 struct gnet_stats_queue qstats;
46};
47
48struct ets_sched {
49 struct list_head active;
50 struct tcf_proto __rcu *filter_list;
51 struct tcf_block *block;
52 unsigned int nbands;
53 unsigned int nstrict;
54 u8 prio2band[TC_PRIO_MAX + 1];
55 struct ets_class classes[TCQ_ETS_MAX_BANDS];
56};
57
58static const struct nla_policy ets_policy[TCA_ETS_MAX + 1] = {
59 [TCA_ETS_NBANDS] = { .type = NLA_U8 },
60 [TCA_ETS_NSTRICT] = { .type = NLA_U8 },
61 [TCA_ETS_QUANTA] = { .type = NLA_NESTED },
62 [TCA_ETS_PRIOMAP] = { .type = NLA_NESTED },
63};
64
65static const struct nla_policy ets_priomap_policy[TCA_ETS_MAX + 1] = {
66 [TCA_ETS_PRIOMAP_BAND] = { .type = NLA_U8 },
67};
68
69static const struct nla_policy ets_quanta_policy[TCA_ETS_MAX + 1] = {
70 [TCA_ETS_QUANTA_BAND] = { .type = NLA_U32 },
71};
72
73static const struct nla_policy ets_class_policy[TCA_ETS_MAX + 1] = {
74 [TCA_ETS_QUANTA_BAND] = { .type = NLA_U32 },
75};
76
77static int ets_quantum_parse(struct Qdisc *sch, const struct nlattr *attr,
78 unsigned int *quantum,
79 struct netlink_ext_ack *extack)
80{
81 *quantum = nla_get_u32(attr);
82 if (!*quantum) {
83 NL_SET_ERR_MSG(extack, "ETS quantum cannot be zero");
84 return -EINVAL;
85 }
86 return 0;
87}
88
89static struct ets_class *
90ets_class_from_arg(struct Qdisc *sch, unsigned long arg)
91{
92 struct ets_sched *q = qdisc_priv(sch);
93
94 if (arg == 0 || arg > q->nbands)
95 return NULL;
96 return &q->classes[arg - 1];
97}
98
99static u32 ets_class_id(struct Qdisc *sch, const struct ets_class *cl)
100{
101 struct ets_sched *q = qdisc_priv(sch);
102 int band = cl - q->classes;
103
104 return TC_H_MAKE(sch->handle, band + 1);
105}
106
107static void ets_offload_change(struct Qdisc *sch)
108{
109 struct net_device *dev = qdisc_dev(sch);
110 struct ets_sched *q = qdisc_priv(sch);
111 struct tc_ets_qopt_offload qopt;
112 unsigned int w_psum_prev = 0;
113 unsigned int q_psum = 0;
114 unsigned int q_sum = 0;
115 unsigned int quantum;
116 unsigned int w_psum;
117 unsigned int weight;
118 unsigned int i;
119
120 if (!tc_can_offload(dev) || !dev->netdev_ops->ndo_setup_tc)
121 return;
122
123 qopt.command = TC_ETS_REPLACE;
124 qopt.handle = sch->handle;
125 qopt.parent = sch->parent;
126 qopt.replace_params.bands = q->nbands;
127 qopt.replace_params.qstats = &sch->qstats;
128 memcpy(&qopt.replace_params.priomap,
129 q->prio2band, sizeof(q->prio2band));
130
131 for (i = 0; i < q->nbands; i++)
132 q_sum += q->classes[i].quantum;
133
134 for (i = 0; i < q->nbands; i++) {
135 quantum = q->classes[i].quantum;
136 q_psum += quantum;
137 w_psum = quantum ? q_psum * 100 / q_sum : 0;
138 weight = w_psum - w_psum_prev;
139 w_psum_prev = w_psum;
140
141 qopt.replace_params.quanta[i] = quantum;
142 qopt.replace_params.weights[i] = weight;
143 }
144
145 dev->netdev_ops->ndo_setup_tc(dev, TC_SETUP_QDISC_ETS, &qopt);
146}
147
148static void ets_offload_destroy(struct Qdisc *sch)
149{
150 struct net_device *dev = qdisc_dev(sch);
151 struct tc_ets_qopt_offload qopt;
152
153 if (!tc_can_offload(dev) || !dev->netdev_ops->ndo_setup_tc)
154 return;
155
156 qopt.command = TC_ETS_DESTROY;
157 qopt.handle = sch->handle;
158 qopt.parent = sch->parent;
159 dev->netdev_ops->ndo_setup_tc(dev, TC_SETUP_QDISC_ETS, &qopt);
160}
161
162static void ets_offload_graft(struct Qdisc *sch, struct Qdisc *new,
163 struct Qdisc *old, unsigned long arg,
164 struct netlink_ext_ack *extack)
165{
166 struct net_device *dev = qdisc_dev(sch);
167 struct tc_ets_qopt_offload qopt;
168
169 qopt.command = TC_ETS_GRAFT;
170 qopt.handle = sch->handle;
171 qopt.parent = sch->parent;
172 qopt.graft_params.band = arg - 1;
173 qopt.graft_params.child_handle = new->handle;
174
175 qdisc_offload_graft_helper(dev, sch, new, old, TC_SETUP_QDISC_ETS,
176 &qopt, extack);
177}
178
179static int ets_offload_dump(struct Qdisc *sch)
180{
181 struct tc_ets_qopt_offload qopt;
182
183 qopt.command = TC_ETS_STATS;
184 qopt.handle = sch->handle;
185 qopt.parent = sch->parent;
186 qopt.stats.bstats = &sch->bstats;
187 qopt.stats.qstats = &sch->qstats;
188
189 return qdisc_offload_dump_helper(sch, TC_SETUP_QDISC_ETS, &qopt);
190}
191
192static bool ets_class_is_strict(struct ets_sched *q, const struct ets_class *cl)
193{
194 unsigned int band = cl - q->classes;
195
196 return band < q->nstrict;
197}
198
199static int ets_class_change(struct Qdisc *sch, u32 classid, u32 parentid,
200 struct nlattr **tca, unsigned long *arg,
201 struct netlink_ext_ack *extack)
202{
203 struct ets_class *cl = ets_class_from_arg(sch, *arg);
204 struct ets_sched *q = qdisc_priv(sch);
205 struct nlattr *opt = tca[TCA_OPTIONS];
206 struct nlattr *tb[TCA_ETS_MAX + 1];
207 unsigned int quantum;
208 int err;
209
210 /* Classes can be added and removed only through Qdisc_ops.change
211 * interface.
212 */
213 if (!cl) {
214 NL_SET_ERR_MSG(extack, "Fine-grained class addition and removal is not supported");
215 return -EOPNOTSUPP;
216 }
217
218 if (!opt) {
219 NL_SET_ERR_MSG(extack, "ETS options are required for this operation");
220 return -EINVAL;
221 }
222
223 err = nla_parse_nested(tb, TCA_ETS_MAX, opt, ets_class_policy, extack);
224 if (err < 0)
225 return err;
226
227 if (!tb[TCA_ETS_QUANTA_BAND])
228 /* Nothing to configure. */
229 return 0;
230
231 if (ets_class_is_strict(q, cl)) {
232 NL_SET_ERR_MSG(extack, "Strict bands do not have a configurable quantum");
233 return -EINVAL;
234 }
235
236 err = ets_quantum_parse(sch, tb[TCA_ETS_QUANTA_BAND], &quantum,
237 extack);
238 if (err)
239 return err;
240
241 sch_tree_lock(sch);
242 cl->quantum = quantum;
243 sch_tree_unlock(sch);
244
245 ets_offload_change(sch);
246 return 0;
247}
248
249static int ets_class_graft(struct Qdisc *sch, unsigned long arg,
250 struct Qdisc *new, struct Qdisc **old,
251 struct netlink_ext_ack *extack)
252{
253 struct ets_class *cl = ets_class_from_arg(sch, arg);
254
255 if (!new) {
256 new = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
257 ets_class_id(sch, cl), NULL);
258 if (!new)
259 new = &noop_qdisc;
260 else
261 qdisc_hash_add(new, true);
262 }
263
264 *old = qdisc_replace(sch, new, &cl->qdisc);
265 ets_offload_graft(sch, new, *old, arg, extack);
266 return 0;
267}
268
269static struct Qdisc *ets_class_leaf(struct Qdisc *sch, unsigned long arg)
270{
271 struct ets_class *cl = ets_class_from_arg(sch, arg);
272
273 return cl->qdisc;
274}
275
276static unsigned long ets_class_find(struct Qdisc *sch, u32 classid)
277{
278 unsigned long band = TC_H_MIN(classid);
279 struct ets_sched *q = qdisc_priv(sch);
280
281 if (band - 1 >= q->nbands)
282 return 0;
283 return band;
284}
285
286static void ets_class_qlen_notify(struct Qdisc *sch, unsigned long arg)
287{
288 struct ets_class *cl = ets_class_from_arg(sch, arg);
289 struct ets_sched *q = qdisc_priv(sch);
290
291 /* We get notified about zero-length child Qdiscs as well if they are
292 * offloaded. Those aren't on the active list though, so don't attempt
293 * to remove them.
294 */
295 if (!ets_class_is_strict(q, cl) && sch->q.qlen)
296 list_del(&cl->alist);
297}
298
299static int ets_class_dump(struct Qdisc *sch, unsigned long arg,
300 struct sk_buff *skb, struct tcmsg *tcm)
301{
302 struct ets_class *cl = ets_class_from_arg(sch, arg);
303 struct ets_sched *q = qdisc_priv(sch);
304 struct nlattr *nest;
305
306 tcm->tcm_parent = TC_H_ROOT;
307 tcm->tcm_handle = ets_class_id(sch, cl);
308 tcm->tcm_info = cl->qdisc->handle;
309
310 nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
311 if (!nest)
312 goto nla_put_failure;
313 if (!ets_class_is_strict(q, cl)) {
314 if (nla_put_u32(skb, TCA_ETS_QUANTA_BAND, cl->quantum))
315 goto nla_put_failure;
316 }
317 return nla_nest_end(skb, nest);
318
319nla_put_failure:
320 nla_nest_cancel(skb, nest);
321 return -EMSGSIZE;
322}
323
324static int ets_class_dump_stats(struct Qdisc *sch, unsigned long arg,
325 struct gnet_dump *d)
326{
327 struct ets_class *cl = ets_class_from_arg(sch, arg);
328 struct Qdisc *cl_q = cl->qdisc;
329
330 if (gnet_stats_copy_basic(d, NULL, &cl_q->bstats, true) < 0 ||
331 qdisc_qstats_copy(d, cl_q) < 0)
332 return -1;
333
334 return 0;
335}
336
337static void ets_qdisc_walk(struct Qdisc *sch, struct qdisc_walker *arg)
338{
339 struct ets_sched *q = qdisc_priv(sch);
340 int i;
341
342 if (arg->stop)
343 return;
344
345 for (i = 0; i < q->nbands; i++) {
346 if (!tc_qdisc_stats_dump(sch, i + 1, arg))
347 break;
348 }
349}
350
351static struct tcf_block *
352ets_qdisc_tcf_block(struct Qdisc *sch, unsigned long cl,
353 struct netlink_ext_ack *extack)
354{
355 struct ets_sched *q = qdisc_priv(sch);
356
357 if (cl) {
358 NL_SET_ERR_MSG(extack, "ETS classid must be zero");
359 return NULL;
360 }
361
362 return q->block;
363}
364
365static unsigned long ets_qdisc_bind_tcf(struct Qdisc *sch, unsigned long parent,
366 u32 classid)
367{
368 return ets_class_find(sch, classid);
369}
370
371static void ets_qdisc_unbind_tcf(struct Qdisc *sch, unsigned long arg)
372{
373}
374
375static struct ets_class *ets_classify(struct sk_buff *skb, struct Qdisc *sch,
376 int *qerr)
377{
378 struct ets_sched *q = qdisc_priv(sch);
379 u32 band = skb->priority;
380 struct tcf_result res;
381 struct tcf_proto *fl;
382 int err;
383
384 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
385 if (TC_H_MAJ(skb->priority) != sch->handle) {
386 fl = rcu_dereference_bh(q->filter_list);
387 err = tcf_classify(skb, NULL, fl, &res, false);
388#ifdef CONFIG_NET_CLS_ACT
389 switch (err) {
390 case TC_ACT_STOLEN:
391 case TC_ACT_QUEUED:
392 case TC_ACT_TRAP:
393 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
394 fallthrough;
395 case TC_ACT_SHOT:
396 return NULL;
397 }
398#endif
399 if (!fl || err < 0) {
400 if (TC_H_MAJ(band))
401 band = 0;
402 return &q->classes[q->prio2band[band & TC_PRIO_MAX]];
403 }
404 band = res.classid;
405 }
406 band = TC_H_MIN(band) - 1;
407 if (band >= q->nbands)
408 return &q->classes[q->prio2band[0]];
409 return &q->classes[band];
410}
411
412static int ets_qdisc_enqueue(struct sk_buff *skb, struct Qdisc *sch,
413 struct sk_buff **to_free)
414{
415 unsigned int len = qdisc_pkt_len(skb);
416 struct ets_sched *q = qdisc_priv(sch);
417 struct ets_class *cl;
418 int err = 0;
419 bool first;
420
421 cl = ets_classify(skb, sch, &err);
422 if (!cl) {
423 if (err & __NET_XMIT_BYPASS)
424 qdisc_qstats_drop(sch);
425 __qdisc_drop(skb, to_free);
426 return err;
427 }
428
429 first = !cl->qdisc->q.qlen;
430 err = qdisc_enqueue(skb, cl->qdisc, to_free);
431 if (unlikely(err != NET_XMIT_SUCCESS)) {
432 if (net_xmit_drop_count(err)) {
433 cl->qstats.drops++;
434 qdisc_qstats_drop(sch);
435 }
436 return err;
437 }
438
439 if (first && !ets_class_is_strict(q, cl)) {
440 list_add_tail(&cl->alist, &q->active);
441 cl->deficit = cl->quantum;
442 }
443
444 sch->qstats.backlog += len;
445 sch->q.qlen++;
446 return err;
447}
448
449static struct sk_buff *
450ets_qdisc_dequeue_skb(struct Qdisc *sch, struct sk_buff *skb)
451{
452 qdisc_bstats_update(sch, skb);
453 qdisc_qstats_backlog_dec(sch, skb);
454 sch->q.qlen--;
455 return skb;
456}
457
458static struct sk_buff *ets_qdisc_dequeue(struct Qdisc *sch)
459{
460 struct ets_sched *q = qdisc_priv(sch);
461 struct ets_class *cl;
462 struct sk_buff *skb;
463 unsigned int band;
464 unsigned int len;
465
466 while (1) {
467 for (band = 0; band < q->nstrict; band++) {
468 cl = &q->classes[band];
469 skb = qdisc_dequeue_peeked(cl->qdisc);
470 if (skb)
471 return ets_qdisc_dequeue_skb(sch, skb);
472 }
473
474 if (list_empty(&q->active))
475 goto out;
476
477 cl = list_first_entry(&q->active, struct ets_class, alist);
478 skb = cl->qdisc->ops->peek(cl->qdisc);
479 if (!skb) {
480 qdisc_warn_nonwc(__func__, cl->qdisc);
481 goto out;
482 }
483
484 len = qdisc_pkt_len(skb);
485 if (len <= cl->deficit) {
486 cl->deficit -= len;
487 skb = qdisc_dequeue_peeked(cl->qdisc);
488 if (unlikely(!skb))
489 goto out;
490 if (cl->qdisc->q.qlen == 0)
491 list_del(&cl->alist);
492 return ets_qdisc_dequeue_skb(sch, skb);
493 }
494
495 cl->deficit += cl->quantum;
496 list_move_tail(&cl->alist, &q->active);
497 }
498out:
499 return NULL;
500}
501
502static int ets_qdisc_priomap_parse(struct nlattr *priomap_attr,
503 unsigned int nbands, u8 *priomap,
504 struct netlink_ext_ack *extack)
505{
506 const struct nlattr *attr;
507 int prio = 0;
508 u8 band;
509 int rem;
510 int err;
511
512 err = __nla_validate_nested(priomap_attr, TCA_ETS_MAX,
513 ets_priomap_policy, NL_VALIDATE_STRICT,
514 extack);
515 if (err)
516 return err;
517
518 nla_for_each_nested(attr, priomap_attr, rem) {
519 switch (nla_type(attr)) {
520 case TCA_ETS_PRIOMAP_BAND:
521 if (prio > TC_PRIO_MAX) {
522 NL_SET_ERR_MSG_MOD(extack, "Too many priorities in ETS priomap");
523 return -EINVAL;
524 }
525 band = nla_get_u8(attr);
526 if (band >= nbands) {
527 NL_SET_ERR_MSG_MOD(extack, "Invalid band number in ETS priomap");
528 return -EINVAL;
529 }
530 priomap[prio++] = band;
531 break;
532 default:
533 WARN_ON_ONCE(1); /* Validate should have caught this. */
534 return -EINVAL;
535 }
536 }
537
538 return 0;
539}
540
541static int ets_qdisc_quanta_parse(struct Qdisc *sch, struct nlattr *quanta_attr,
542 unsigned int nbands, unsigned int nstrict,
543 unsigned int *quanta,
544 struct netlink_ext_ack *extack)
545{
546 const struct nlattr *attr;
547 int band = nstrict;
548 int rem;
549 int err;
550
551 err = __nla_validate_nested(quanta_attr, TCA_ETS_MAX,
552 ets_quanta_policy, NL_VALIDATE_STRICT,
553 extack);
554 if (err < 0)
555 return err;
556
557 nla_for_each_nested(attr, quanta_attr, rem) {
558 switch (nla_type(attr)) {
559 case TCA_ETS_QUANTA_BAND:
560 if (band >= nbands) {
561 NL_SET_ERR_MSG_MOD(extack, "ETS quanta has more values than bands");
562 return -EINVAL;
563 }
564 err = ets_quantum_parse(sch, attr, &quanta[band++],
565 extack);
566 if (err)
567 return err;
568 break;
569 default:
570 WARN_ON_ONCE(1); /* Validate should have caught this. */
571 return -EINVAL;
572 }
573 }
574
575 return 0;
576}
577
578static int ets_qdisc_change(struct Qdisc *sch, struct nlattr *opt,
579 struct netlink_ext_ack *extack)
580{
581 unsigned int quanta[TCQ_ETS_MAX_BANDS] = {0};
582 struct Qdisc *queues[TCQ_ETS_MAX_BANDS];
583 struct ets_sched *q = qdisc_priv(sch);
584 struct nlattr *tb[TCA_ETS_MAX + 1];
585 unsigned int oldbands = q->nbands;
586 u8 priomap[TC_PRIO_MAX + 1];
587 unsigned int nstrict = 0;
588 unsigned int nbands;
589 unsigned int i;
590 int err;
591
592 err = nla_parse_nested(tb, TCA_ETS_MAX, opt, ets_policy, extack);
593 if (err < 0)
594 return err;
595
596 if (!tb[TCA_ETS_NBANDS]) {
597 NL_SET_ERR_MSG_MOD(extack, "Number of bands is a required argument");
598 return -EINVAL;
599 }
600 nbands = nla_get_u8(tb[TCA_ETS_NBANDS]);
601 if (nbands < 1 || nbands > TCQ_ETS_MAX_BANDS) {
602 NL_SET_ERR_MSG_MOD(extack, "Invalid number of bands");
603 return -EINVAL;
604 }
605 /* Unless overridden, traffic goes to the last band. */
606 memset(priomap, nbands - 1, sizeof(priomap));
607
608 if (tb[TCA_ETS_NSTRICT]) {
609 nstrict = nla_get_u8(tb[TCA_ETS_NSTRICT]);
610 if (nstrict > nbands) {
611 NL_SET_ERR_MSG_MOD(extack, "Invalid number of strict bands");
612 return -EINVAL;
613 }
614 }
615
616 if (tb[TCA_ETS_PRIOMAP]) {
617 err = ets_qdisc_priomap_parse(tb[TCA_ETS_PRIOMAP],
618 nbands, priomap, extack);
619 if (err)
620 return err;
621 }
622
623 if (tb[TCA_ETS_QUANTA]) {
624 err = ets_qdisc_quanta_parse(sch, tb[TCA_ETS_QUANTA],
625 nbands, nstrict, quanta, extack);
626 if (err)
627 return err;
628 }
629 /* If there are more bands than strict + quanta provided, the remaining
630 * ones are ETS with quantum of MTU. Initialize the missing values here.
631 */
632 for (i = nstrict; i < nbands; i++) {
633 if (!quanta[i])
634 quanta[i] = psched_mtu(qdisc_dev(sch));
635 }
636
637 /* Before commit, make sure we can allocate all new qdiscs */
638 for (i = oldbands; i < nbands; i++) {
639 queues[i] = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
640 ets_class_id(sch, &q->classes[i]),
641 extack);
642 if (!queues[i]) {
643 while (i > oldbands)
644 qdisc_put(queues[--i]);
645 return -ENOMEM;
646 }
647 }
648
649 sch_tree_lock(sch);
650
651 WRITE_ONCE(q->nbands, nbands);
652 for (i = nstrict; i < q->nstrict; i++) {
653 if (q->classes[i].qdisc->q.qlen) {
654 list_add_tail(&q->classes[i].alist, &q->active);
655 q->classes[i].deficit = quanta[i];
656 }
657 }
658 for (i = q->nbands; i < oldbands; i++) {
659 if (i >= q->nstrict && q->classes[i].qdisc->q.qlen)
660 list_del(&q->classes[i].alist);
661 qdisc_tree_flush_backlog(q->classes[i].qdisc);
662 }
663 WRITE_ONCE(q->nstrict, nstrict);
664 memcpy(q->prio2band, priomap, sizeof(priomap));
665
666 for (i = 0; i < q->nbands; i++)
667 WRITE_ONCE(q->classes[i].quantum, quanta[i]);
668
669 for (i = oldbands; i < q->nbands; i++) {
670 q->classes[i].qdisc = queues[i];
671 if (q->classes[i].qdisc != &noop_qdisc)
672 qdisc_hash_add(q->classes[i].qdisc, true);
673 }
674
675 sch_tree_unlock(sch);
676
677 ets_offload_change(sch);
678 for (i = q->nbands; i < oldbands; i++) {
679 qdisc_put(q->classes[i].qdisc);
680 q->classes[i].qdisc = NULL;
681 WRITE_ONCE(q->classes[i].quantum, 0);
682 q->classes[i].deficit = 0;
683 gnet_stats_basic_sync_init(&q->classes[i].bstats);
684 memset(&q->classes[i].qstats, 0, sizeof(q->classes[i].qstats));
685 }
686 return 0;
687}
688
689static int ets_qdisc_init(struct Qdisc *sch, struct nlattr *opt,
690 struct netlink_ext_ack *extack)
691{
692 struct ets_sched *q = qdisc_priv(sch);
693 int err, i;
694
695 if (!opt)
696 return -EINVAL;
697
698 err = tcf_block_get(&q->block, &q->filter_list, sch, extack);
699 if (err)
700 return err;
701
702 INIT_LIST_HEAD(&q->active);
703 for (i = 0; i < TCQ_ETS_MAX_BANDS; i++)
704 INIT_LIST_HEAD(&q->classes[i].alist);
705
706 return ets_qdisc_change(sch, opt, extack);
707}
708
709static void ets_qdisc_reset(struct Qdisc *sch)
710{
711 struct ets_sched *q = qdisc_priv(sch);
712 int band;
713
714 for (band = q->nstrict; band < q->nbands; band++) {
715 if (q->classes[band].qdisc->q.qlen)
716 list_del(&q->classes[band].alist);
717 }
718 for (band = 0; band < q->nbands; band++)
719 qdisc_reset(q->classes[band].qdisc);
720}
721
722static void ets_qdisc_destroy(struct Qdisc *sch)
723{
724 struct ets_sched *q = qdisc_priv(sch);
725 int band;
726
727 ets_offload_destroy(sch);
728 tcf_block_put(q->block);
729 for (band = 0; band < q->nbands; band++)
730 qdisc_put(q->classes[band].qdisc);
731}
732
733static int ets_qdisc_dump(struct Qdisc *sch, struct sk_buff *skb)
734{
735 struct ets_sched *q = qdisc_priv(sch);
736 struct nlattr *opts;
737 struct nlattr *nest;
738 u8 nbands, nstrict;
739 int band;
740 int prio;
741 int err;
742
743 err = ets_offload_dump(sch);
744 if (err)
745 return err;
746
747 opts = nla_nest_start_noflag(skb, TCA_OPTIONS);
748 if (!opts)
749 goto nla_err;
750
751 nbands = READ_ONCE(q->nbands);
752 if (nla_put_u8(skb, TCA_ETS_NBANDS, nbands))
753 goto nla_err;
754
755 nstrict = READ_ONCE(q->nstrict);
756 if (nstrict && nla_put_u8(skb, TCA_ETS_NSTRICT, nstrict))
757 goto nla_err;
758
759 if (nbands > nstrict) {
760 nest = nla_nest_start(skb, TCA_ETS_QUANTA);
761 if (!nest)
762 goto nla_err;
763
764 for (band = nstrict; band < nbands; band++) {
765 if (nla_put_u32(skb, TCA_ETS_QUANTA_BAND,
766 READ_ONCE(q->classes[band].quantum)))
767 goto nla_err;
768 }
769
770 nla_nest_end(skb, nest);
771 }
772
773 nest = nla_nest_start(skb, TCA_ETS_PRIOMAP);
774 if (!nest)
775 goto nla_err;
776
777 for (prio = 0; prio <= TC_PRIO_MAX; prio++) {
778 if (nla_put_u8(skb, TCA_ETS_PRIOMAP_BAND,
779 READ_ONCE(q->prio2band[prio])))
780 goto nla_err;
781 }
782
783 nla_nest_end(skb, nest);
784
785 return nla_nest_end(skb, opts);
786
787nla_err:
788 nla_nest_cancel(skb, opts);
789 return -EMSGSIZE;
790}
791
792static const struct Qdisc_class_ops ets_class_ops = {
793 .change = ets_class_change,
794 .graft = ets_class_graft,
795 .leaf = ets_class_leaf,
796 .find = ets_class_find,
797 .qlen_notify = ets_class_qlen_notify,
798 .dump = ets_class_dump,
799 .dump_stats = ets_class_dump_stats,
800 .walk = ets_qdisc_walk,
801 .tcf_block = ets_qdisc_tcf_block,
802 .bind_tcf = ets_qdisc_bind_tcf,
803 .unbind_tcf = ets_qdisc_unbind_tcf,
804};
805
806static struct Qdisc_ops ets_qdisc_ops __read_mostly = {
807 .cl_ops = &ets_class_ops,
808 .id = "ets",
809 .priv_size = sizeof(struct ets_sched),
810 .enqueue = ets_qdisc_enqueue,
811 .dequeue = ets_qdisc_dequeue,
812 .peek = qdisc_peek_dequeued,
813 .change = ets_qdisc_change,
814 .init = ets_qdisc_init,
815 .reset = ets_qdisc_reset,
816 .destroy = ets_qdisc_destroy,
817 .dump = ets_qdisc_dump,
818 .owner = THIS_MODULE,
819};
820MODULE_ALIAS_NET_SCH("ets");
821
822static int __init ets_init(void)
823{
824 return register_qdisc(&ets_qdisc_ops);
825}
826
827static void __exit ets_exit(void)
828{
829 unregister_qdisc(&ets_qdisc_ops);
830}
831
832module_init(ets_init);
833module_exit(ets_exit);
834MODULE_LICENSE("GPL");
835MODULE_DESCRIPTION("Enhanced Transmission Selection(ETS) scheduler");
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * net/sched/sch_ets.c Enhanced Transmission Selection scheduler
4 *
5 * Description
6 * -----------
7 *
8 * The Enhanced Transmission Selection scheduler is a classful queuing
9 * discipline that merges functionality of PRIO and DRR qdiscs in one scheduler.
10 * ETS makes it easy to configure a set of strict and bandwidth-sharing bands to
11 * implement the transmission selection described in 802.1Qaz.
12 *
13 * Although ETS is technically classful, it's not possible to add and remove
14 * classes at will. Instead one specifies number of classes, how many are
15 * PRIO-like and how many DRR-like, and quanta for the latter.
16 *
17 * Algorithm
18 * ---------
19 *
20 * The strict classes, if any, are tried for traffic first: first band 0, if it
21 * has no traffic then band 1, etc.
22 *
23 * When there is no traffic in any of the strict queues, the bandwidth-sharing
24 * ones are tried next. Each band is assigned a deficit counter, initialized to
25 * "quantum" of that band. ETS maintains a list of active bandwidth-sharing
26 * bands whose qdiscs are non-empty. A packet is dequeued from the band at the
27 * head of the list if the packet size is smaller or equal to the deficit
28 * counter. If the counter is too small, it is increased by "quantum" and the
29 * scheduler moves on to the next band in the active list.
30 */
31
32#include <linux/module.h>
33#include <net/gen_stats.h>
34#include <net/netlink.h>
35#include <net/pkt_cls.h>
36#include <net/pkt_sched.h>
37#include <net/sch_generic.h>
38
39struct ets_class {
40 struct list_head alist; /* In struct ets_sched.active. */
41 struct Qdisc *qdisc;
42 u32 quantum;
43 u32 deficit;
44 struct gnet_stats_basic_sync bstats;
45 struct gnet_stats_queue qstats;
46};
47
48struct ets_sched {
49 struct list_head active;
50 struct tcf_proto __rcu *filter_list;
51 struct tcf_block *block;
52 unsigned int nbands;
53 unsigned int nstrict;
54 u8 prio2band[TC_PRIO_MAX + 1];
55 struct ets_class classes[TCQ_ETS_MAX_BANDS];
56};
57
58static const struct nla_policy ets_policy[TCA_ETS_MAX + 1] = {
59 [TCA_ETS_NBANDS] = { .type = NLA_U8 },
60 [TCA_ETS_NSTRICT] = { .type = NLA_U8 },
61 [TCA_ETS_QUANTA] = { .type = NLA_NESTED },
62 [TCA_ETS_PRIOMAP] = { .type = NLA_NESTED },
63};
64
65static const struct nla_policy ets_priomap_policy[TCA_ETS_MAX + 1] = {
66 [TCA_ETS_PRIOMAP_BAND] = { .type = NLA_U8 },
67};
68
69static const struct nla_policy ets_quanta_policy[TCA_ETS_MAX + 1] = {
70 [TCA_ETS_QUANTA_BAND] = { .type = NLA_U32 },
71};
72
73static const struct nla_policy ets_class_policy[TCA_ETS_MAX + 1] = {
74 [TCA_ETS_QUANTA_BAND] = { .type = NLA_U32 },
75};
76
77static int ets_quantum_parse(struct Qdisc *sch, const struct nlattr *attr,
78 unsigned int *quantum,
79 struct netlink_ext_ack *extack)
80{
81 *quantum = nla_get_u32(attr);
82 if (!*quantum) {
83 NL_SET_ERR_MSG(extack, "ETS quantum cannot be zero");
84 return -EINVAL;
85 }
86 return 0;
87}
88
89static struct ets_class *
90ets_class_from_arg(struct Qdisc *sch, unsigned long arg)
91{
92 struct ets_sched *q = qdisc_priv(sch);
93
94 return &q->classes[arg - 1];
95}
96
97static u32 ets_class_id(struct Qdisc *sch, const struct ets_class *cl)
98{
99 struct ets_sched *q = qdisc_priv(sch);
100 int band = cl - q->classes;
101
102 return TC_H_MAKE(sch->handle, band + 1);
103}
104
105static void ets_offload_change(struct Qdisc *sch)
106{
107 struct net_device *dev = qdisc_dev(sch);
108 struct ets_sched *q = qdisc_priv(sch);
109 struct tc_ets_qopt_offload qopt;
110 unsigned int w_psum_prev = 0;
111 unsigned int q_psum = 0;
112 unsigned int q_sum = 0;
113 unsigned int quantum;
114 unsigned int w_psum;
115 unsigned int weight;
116 unsigned int i;
117
118 if (!tc_can_offload(dev) || !dev->netdev_ops->ndo_setup_tc)
119 return;
120
121 qopt.command = TC_ETS_REPLACE;
122 qopt.handle = sch->handle;
123 qopt.parent = sch->parent;
124 qopt.replace_params.bands = q->nbands;
125 qopt.replace_params.qstats = &sch->qstats;
126 memcpy(&qopt.replace_params.priomap,
127 q->prio2band, sizeof(q->prio2band));
128
129 for (i = 0; i < q->nbands; i++)
130 q_sum += q->classes[i].quantum;
131
132 for (i = 0; i < q->nbands; i++) {
133 quantum = q->classes[i].quantum;
134 q_psum += quantum;
135 w_psum = quantum ? q_psum * 100 / q_sum : 0;
136 weight = w_psum - w_psum_prev;
137 w_psum_prev = w_psum;
138
139 qopt.replace_params.quanta[i] = quantum;
140 qopt.replace_params.weights[i] = weight;
141 }
142
143 dev->netdev_ops->ndo_setup_tc(dev, TC_SETUP_QDISC_ETS, &qopt);
144}
145
146static void ets_offload_destroy(struct Qdisc *sch)
147{
148 struct net_device *dev = qdisc_dev(sch);
149 struct tc_ets_qopt_offload qopt;
150
151 if (!tc_can_offload(dev) || !dev->netdev_ops->ndo_setup_tc)
152 return;
153
154 qopt.command = TC_ETS_DESTROY;
155 qopt.handle = sch->handle;
156 qopt.parent = sch->parent;
157 dev->netdev_ops->ndo_setup_tc(dev, TC_SETUP_QDISC_ETS, &qopt);
158}
159
160static void ets_offload_graft(struct Qdisc *sch, struct Qdisc *new,
161 struct Qdisc *old, unsigned long arg,
162 struct netlink_ext_ack *extack)
163{
164 struct net_device *dev = qdisc_dev(sch);
165 struct tc_ets_qopt_offload qopt;
166
167 qopt.command = TC_ETS_GRAFT;
168 qopt.handle = sch->handle;
169 qopt.parent = sch->parent;
170 qopt.graft_params.band = arg - 1;
171 qopt.graft_params.child_handle = new->handle;
172
173 qdisc_offload_graft_helper(dev, sch, new, old, TC_SETUP_QDISC_ETS,
174 &qopt, extack);
175}
176
177static int ets_offload_dump(struct Qdisc *sch)
178{
179 struct tc_ets_qopt_offload qopt;
180
181 qopt.command = TC_ETS_STATS;
182 qopt.handle = sch->handle;
183 qopt.parent = sch->parent;
184 qopt.stats.bstats = &sch->bstats;
185 qopt.stats.qstats = &sch->qstats;
186
187 return qdisc_offload_dump_helper(sch, TC_SETUP_QDISC_ETS, &qopt);
188}
189
190static bool ets_class_is_strict(struct ets_sched *q, const struct ets_class *cl)
191{
192 unsigned int band = cl - q->classes;
193
194 return band < q->nstrict;
195}
196
197static int ets_class_change(struct Qdisc *sch, u32 classid, u32 parentid,
198 struct nlattr **tca, unsigned long *arg,
199 struct netlink_ext_ack *extack)
200{
201 struct ets_class *cl = ets_class_from_arg(sch, *arg);
202 struct ets_sched *q = qdisc_priv(sch);
203 struct nlattr *opt = tca[TCA_OPTIONS];
204 struct nlattr *tb[TCA_ETS_MAX + 1];
205 unsigned int quantum;
206 int err;
207
208 /* Classes can be added and removed only through Qdisc_ops.change
209 * interface.
210 */
211 if (!cl) {
212 NL_SET_ERR_MSG(extack, "Fine-grained class addition and removal is not supported");
213 return -EOPNOTSUPP;
214 }
215
216 if (!opt) {
217 NL_SET_ERR_MSG(extack, "ETS options are required for this operation");
218 return -EINVAL;
219 }
220
221 err = nla_parse_nested(tb, TCA_ETS_MAX, opt, ets_class_policy, extack);
222 if (err < 0)
223 return err;
224
225 if (!tb[TCA_ETS_QUANTA_BAND])
226 /* Nothing to configure. */
227 return 0;
228
229 if (ets_class_is_strict(q, cl)) {
230 NL_SET_ERR_MSG(extack, "Strict bands do not have a configurable quantum");
231 return -EINVAL;
232 }
233
234 err = ets_quantum_parse(sch, tb[TCA_ETS_QUANTA_BAND], &quantum,
235 extack);
236 if (err)
237 return err;
238
239 sch_tree_lock(sch);
240 cl->quantum = quantum;
241 sch_tree_unlock(sch);
242
243 ets_offload_change(sch);
244 return 0;
245}
246
247static int ets_class_graft(struct Qdisc *sch, unsigned long arg,
248 struct Qdisc *new, struct Qdisc **old,
249 struct netlink_ext_ack *extack)
250{
251 struct ets_class *cl = ets_class_from_arg(sch, arg);
252
253 if (!new) {
254 new = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
255 ets_class_id(sch, cl), NULL);
256 if (!new)
257 new = &noop_qdisc;
258 else
259 qdisc_hash_add(new, true);
260 }
261
262 *old = qdisc_replace(sch, new, &cl->qdisc);
263 ets_offload_graft(sch, new, *old, arg, extack);
264 return 0;
265}
266
267static struct Qdisc *ets_class_leaf(struct Qdisc *sch, unsigned long arg)
268{
269 struct ets_class *cl = ets_class_from_arg(sch, arg);
270
271 return cl->qdisc;
272}
273
274static unsigned long ets_class_find(struct Qdisc *sch, u32 classid)
275{
276 unsigned long band = TC_H_MIN(classid);
277 struct ets_sched *q = qdisc_priv(sch);
278
279 if (band - 1 >= q->nbands)
280 return 0;
281 return band;
282}
283
284static void ets_class_qlen_notify(struct Qdisc *sch, unsigned long arg)
285{
286 struct ets_class *cl = ets_class_from_arg(sch, arg);
287 struct ets_sched *q = qdisc_priv(sch);
288
289 /* We get notified about zero-length child Qdiscs as well if they are
290 * offloaded. Those aren't on the active list though, so don't attempt
291 * to remove them.
292 */
293 if (!ets_class_is_strict(q, cl) && sch->q.qlen)
294 list_del(&cl->alist);
295}
296
297static int ets_class_dump(struct Qdisc *sch, unsigned long arg,
298 struct sk_buff *skb, struct tcmsg *tcm)
299{
300 struct ets_class *cl = ets_class_from_arg(sch, arg);
301 struct ets_sched *q = qdisc_priv(sch);
302 struct nlattr *nest;
303
304 tcm->tcm_parent = TC_H_ROOT;
305 tcm->tcm_handle = ets_class_id(sch, cl);
306 tcm->tcm_info = cl->qdisc->handle;
307
308 nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
309 if (!nest)
310 goto nla_put_failure;
311 if (!ets_class_is_strict(q, cl)) {
312 if (nla_put_u32(skb, TCA_ETS_QUANTA_BAND, cl->quantum))
313 goto nla_put_failure;
314 }
315 return nla_nest_end(skb, nest);
316
317nla_put_failure:
318 nla_nest_cancel(skb, nest);
319 return -EMSGSIZE;
320}
321
322static int ets_class_dump_stats(struct Qdisc *sch, unsigned long arg,
323 struct gnet_dump *d)
324{
325 struct ets_class *cl = ets_class_from_arg(sch, arg);
326 struct Qdisc *cl_q = cl->qdisc;
327
328 if (gnet_stats_copy_basic(d, NULL, &cl_q->bstats, true) < 0 ||
329 qdisc_qstats_copy(d, cl_q) < 0)
330 return -1;
331
332 return 0;
333}
334
335static void ets_qdisc_walk(struct Qdisc *sch, struct qdisc_walker *arg)
336{
337 struct ets_sched *q = qdisc_priv(sch);
338 int i;
339
340 if (arg->stop)
341 return;
342
343 for (i = 0; i < q->nbands; i++) {
344 if (!tc_qdisc_stats_dump(sch, i + 1, arg))
345 break;
346 }
347}
348
349static struct tcf_block *
350ets_qdisc_tcf_block(struct Qdisc *sch, unsigned long cl,
351 struct netlink_ext_ack *extack)
352{
353 struct ets_sched *q = qdisc_priv(sch);
354
355 if (cl) {
356 NL_SET_ERR_MSG(extack, "ETS classid must be zero");
357 return NULL;
358 }
359
360 return q->block;
361}
362
363static unsigned long ets_qdisc_bind_tcf(struct Qdisc *sch, unsigned long parent,
364 u32 classid)
365{
366 return ets_class_find(sch, classid);
367}
368
369static void ets_qdisc_unbind_tcf(struct Qdisc *sch, unsigned long arg)
370{
371}
372
373static struct ets_class *ets_classify(struct sk_buff *skb, struct Qdisc *sch,
374 int *qerr)
375{
376 struct ets_sched *q = qdisc_priv(sch);
377 u32 band = skb->priority;
378 struct tcf_result res;
379 struct tcf_proto *fl;
380 int err;
381
382 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
383 if (TC_H_MAJ(skb->priority) != sch->handle) {
384 fl = rcu_dereference_bh(q->filter_list);
385 err = tcf_classify(skb, NULL, fl, &res, false);
386#ifdef CONFIG_NET_CLS_ACT
387 switch (err) {
388 case TC_ACT_STOLEN:
389 case TC_ACT_QUEUED:
390 case TC_ACT_TRAP:
391 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
392 fallthrough;
393 case TC_ACT_SHOT:
394 return NULL;
395 }
396#endif
397 if (!fl || err < 0) {
398 if (TC_H_MAJ(band))
399 band = 0;
400 return &q->classes[q->prio2band[band & TC_PRIO_MAX]];
401 }
402 band = res.classid;
403 }
404 band = TC_H_MIN(band) - 1;
405 if (band >= q->nbands)
406 return &q->classes[q->prio2band[0]];
407 return &q->classes[band];
408}
409
410static int ets_qdisc_enqueue(struct sk_buff *skb, struct Qdisc *sch,
411 struct sk_buff **to_free)
412{
413 unsigned int len = qdisc_pkt_len(skb);
414 struct ets_sched *q = qdisc_priv(sch);
415 struct ets_class *cl;
416 int err = 0;
417 bool first;
418
419 cl = ets_classify(skb, sch, &err);
420 if (!cl) {
421 if (err & __NET_XMIT_BYPASS)
422 qdisc_qstats_drop(sch);
423 __qdisc_drop(skb, to_free);
424 return err;
425 }
426
427 first = !cl->qdisc->q.qlen;
428 err = qdisc_enqueue(skb, cl->qdisc, to_free);
429 if (unlikely(err != NET_XMIT_SUCCESS)) {
430 if (net_xmit_drop_count(err)) {
431 cl->qstats.drops++;
432 qdisc_qstats_drop(sch);
433 }
434 return err;
435 }
436
437 if (first && !ets_class_is_strict(q, cl)) {
438 list_add_tail(&cl->alist, &q->active);
439 cl->deficit = cl->quantum;
440 }
441
442 sch->qstats.backlog += len;
443 sch->q.qlen++;
444 return err;
445}
446
447static struct sk_buff *
448ets_qdisc_dequeue_skb(struct Qdisc *sch, struct sk_buff *skb)
449{
450 qdisc_bstats_update(sch, skb);
451 qdisc_qstats_backlog_dec(sch, skb);
452 sch->q.qlen--;
453 return skb;
454}
455
456static struct sk_buff *ets_qdisc_dequeue(struct Qdisc *sch)
457{
458 struct ets_sched *q = qdisc_priv(sch);
459 struct ets_class *cl;
460 struct sk_buff *skb;
461 unsigned int band;
462 unsigned int len;
463
464 while (1) {
465 for (band = 0; band < q->nstrict; band++) {
466 cl = &q->classes[band];
467 skb = qdisc_dequeue_peeked(cl->qdisc);
468 if (skb)
469 return ets_qdisc_dequeue_skb(sch, skb);
470 }
471
472 if (list_empty(&q->active))
473 goto out;
474
475 cl = list_first_entry(&q->active, struct ets_class, alist);
476 skb = cl->qdisc->ops->peek(cl->qdisc);
477 if (!skb) {
478 qdisc_warn_nonwc(__func__, cl->qdisc);
479 goto out;
480 }
481
482 len = qdisc_pkt_len(skb);
483 if (len <= cl->deficit) {
484 cl->deficit -= len;
485 skb = qdisc_dequeue_peeked(cl->qdisc);
486 if (unlikely(!skb))
487 goto out;
488 if (cl->qdisc->q.qlen == 0)
489 list_del(&cl->alist);
490 return ets_qdisc_dequeue_skb(sch, skb);
491 }
492
493 cl->deficit += cl->quantum;
494 list_move_tail(&cl->alist, &q->active);
495 }
496out:
497 return NULL;
498}
499
500static int ets_qdisc_priomap_parse(struct nlattr *priomap_attr,
501 unsigned int nbands, u8 *priomap,
502 struct netlink_ext_ack *extack)
503{
504 const struct nlattr *attr;
505 int prio = 0;
506 u8 band;
507 int rem;
508 int err;
509
510 err = __nla_validate_nested(priomap_attr, TCA_ETS_MAX,
511 ets_priomap_policy, NL_VALIDATE_STRICT,
512 extack);
513 if (err)
514 return err;
515
516 nla_for_each_nested(attr, priomap_attr, rem) {
517 switch (nla_type(attr)) {
518 case TCA_ETS_PRIOMAP_BAND:
519 if (prio > TC_PRIO_MAX) {
520 NL_SET_ERR_MSG_MOD(extack, "Too many priorities in ETS priomap");
521 return -EINVAL;
522 }
523 band = nla_get_u8(attr);
524 if (band >= nbands) {
525 NL_SET_ERR_MSG_MOD(extack, "Invalid band number in ETS priomap");
526 return -EINVAL;
527 }
528 priomap[prio++] = band;
529 break;
530 default:
531 WARN_ON_ONCE(1); /* Validate should have caught this. */
532 return -EINVAL;
533 }
534 }
535
536 return 0;
537}
538
539static int ets_qdisc_quanta_parse(struct Qdisc *sch, struct nlattr *quanta_attr,
540 unsigned int nbands, unsigned int nstrict,
541 unsigned int *quanta,
542 struct netlink_ext_ack *extack)
543{
544 const struct nlattr *attr;
545 int band = nstrict;
546 int rem;
547 int err;
548
549 err = __nla_validate_nested(quanta_attr, TCA_ETS_MAX,
550 ets_quanta_policy, NL_VALIDATE_STRICT,
551 extack);
552 if (err < 0)
553 return err;
554
555 nla_for_each_nested(attr, quanta_attr, rem) {
556 switch (nla_type(attr)) {
557 case TCA_ETS_QUANTA_BAND:
558 if (band >= nbands) {
559 NL_SET_ERR_MSG_MOD(extack, "ETS quanta has more values than bands");
560 return -EINVAL;
561 }
562 err = ets_quantum_parse(sch, attr, &quanta[band++],
563 extack);
564 if (err)
565 return err;
566 break;
567 default:
568 WARN_ON_ONCE(1); /* Validate should have caught this. */
569 return -EINVAL;
570 }
571 }
572
573 return 0;
574}
575
576static int ets_qdisc_change(struct Qdisc *sch, struct nlattr *opt,
577 struct netlink_ext_ack *extack)
578{
579 unsigned int quanta[TCQ_ETS_MAX_BANDS] = {0};
580 struct Qdisc *queues[TCQ_ETS_MAX_BANDS];
581 struct ets_sched *q = qdisc_priv(sch);
582 struct nlattr *tb[TCA_ETS_MAX + 1];
583 unsigned int oldbands = q->nbands;
584 u8 priomap[TC_PRIO_MAX + 1];
585 unsigned int nstrict = 0;
586 unsigned int nbands;
587 unsigned int i;
588 int err;
589
590 err = nla_parse_nested(tb, TCA_ETS_MAX, opt, ets_policy, extack);
591 if (err < 0)
592 return err;
593
594 if (!tb[TCA_ETS_NBANDS]) {
595 NL_SET_ERR_MSG_MOD(extack, "Number of bands is a required argument");
596 return -EINVAL;
597 }
598 nbands = nla_get_u8(tb[TCA_ETS_NBANDS]);
599 if (nbands < 1 || nbands > TCQ_ETS_MAX_BANDS) {
600 NL_SET_ERR_MSG_MOD(extack, "Invalid number of bands");
601 return -EINVAL;
602 }
603 /* Unless overridden, traffic goes to the last band. */
604 memset(priomap, nbands - 1, sizeof(priomap));
605
606 if (tb[TCA_ETS_NSTRICT]) {
607 nstrict = nla_get_u8(tb[TCA_ETS_NSTRICT]);
608 if (nstrict > nbands) {
609 NL_SET_ERR_MSG_MOD(extack, "Invalid number of strict bands");
610 return -EINVAL;
611 }
612 }
613
614 if (tb[TCA_ETS_PRIOMAP]) {
615 err = ets_qdisc_priomap_parse(tb[TCA_ETS_PRIOMAP],
616 nbands, priomap, extack);
617 if (err)
618 return err;
619 }
620
621 if (tb[TCA_ETS_QUANTA]) {
622 err = ets_qdisc_quanta_parse(sch, tb[TCA_ETS_QUANTA],
623 nbands, nstrict, quanta, extack);
624 if (err)
625 return err;
626 }
627 /* If there are more bands than strict + quanta provided, the remaining
628 * ones are ETS with quantum of MTU. Initialize the missing values here.
629 */
630 for (i = nstrict; i < nbands; i++) {
631 if (!quanta[i])
632 quanta[i] = psched_mtu(qdisc_dev(sch));
633 }
634
635 /* Before commit, make sure we can allocate all new qdiscs */
636 for (i = oldbands; i < nbands; i++) {
637 queues[i] = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
638 ets_class_id(sch, &q->classes[i]),
639 extack);
640 if (!queues[i]) {
641 while (i > oldbands)
642 qdisc_put(queues[--i]);
643 return -ENOMEM;
644 }
645 }
646
647 sch_tree_lock(sch);
648
649 q->nbands = nbands;
650 for (i = nstrict; i < q->nstrict; i++) {
651 if (q->classes[i].qdisc->q.qlen) {
652 list_add_tail(&q->classes[i].alist, &q->active);
653 q->classes[i].deficit = quanta[i];
654 }
655 }
656 for (i = q->nbands; i < oldbands; i++) {
657 if (i >= q->nstrict && q->classes[i].qdisc->q.qlen)
658 list_del(&q->classes[i].alist);
659 qdisc_tree_flush_backlog(q->classes[i].qdisc);
660 }
661 q->nstrict = nstrict;
662 memcpy(q->prio2band, priomap, sizeof(priomap));
663
664 for (i = 0; i < q->nbands; i++)
665 q->classes[i].quantum = quanta[i];
666
667 for (i = oldbands; i < q->nbands; i++) {
668 q->classes[i].qdisc = queues[i];
669 if (q->classes[i].qdisc != &noop_qdisc)
670 qdisc_hash_add(q->classes[i].qdisc, true);
671 }
672
673 sch_tree_unlock(sch);
674
675 ets_offload_change(sch);
676 for (i = q->nbands; i < oldbands; i++) {
677 qdisc_put(q->classes[i].qdisc);
678 q->classes[i].qdisc = NULL;
679 q->classes[i].quantum = 0;
680 q->classes[i].deficit = 0;
681 gnet_stats_basic_sync_init(&q->classes[i].bstats);
682 memset(&q->classes[i].qstats, 0, sizeof(q->classes[i].qstats));
683 }
684 return 0;
685}
686
687static int ets_qdisc_init(struct Qdisc *sch, struct nlattr *opt,
688 struct netlink_ext_ack *extack)
689{
690 struct ets_sched *q = qdisc_priv(sch);
691 int err, i;
692
693 if (!opt)
694 return -EINVAL;
695
696 err = tcf_block_get(&q->block, &q->filter_list, sch, extack);
697 if (err)
698 return err;
699
700 INIT_LIST_HEAD(&q->active);
701 for (i = 0; i < TCQ_ETS_MAX_BANDS; i++)
702 INIT_LIST_HEAD(&q->classes[i].alist);
703
704 return ets_qdisc_change(sch, opt, extack);
705}
706
707static void ets_qdisc_reset(struct Qdisc *sch)
708{
709 struct ets_sched *q = qdisc_priv(sch);
710 int band;
711
712 for (band = q->nstrict; band < q->nbands; band++) {
713 if (q->classes[band].qdisc->q.qlen)
714 list_del(&q->classes[band].alist);
715 }
716 for (band = 0; band < q->nbands; band++)
717 qdisc_reset(q->classes[band].qdisc);
718}
719
720static void ets_qdisc_destroy(struct Qdisc *sch)
721{
722 struct ets_sched *q = qdisc_priv(sch);
723 int band;
724
725 ets_offload_destroy(sch);
726 tcf_block_put(q->block);
727 for (band = 0; band < q->nbands; band++)
728 qdisc_put(q->classes[band].qdisc);
729}
730
731static int ets_qdisc_dump(struct Qdisc *sch, struct sk_buff *skb)
732{
733 struct ets_sched *q = qdisc_priv(sch);
734 struct nlattr *opts;
735 struct nlattr *nest;
736 int band;
737 int prio;
738 int err;
739
740 err = ets_offload_dump(sch);
741 if (err)
742 return err;
743
744 opts = nla_nest_start_noflag(skb, TCA_OPTIONS);
745 if (!opts)
746 goto nla_err;
747
748 if (nla_put_u8(skb, TCA_ETS_NBANDS, q->nbands))
749 goto nla_err;
750
751 if (q->nstrict &&
752 nla_put_u8(skb, TCA_ETS_NSTRICT, q->nstrict))
753 goto nla_err;
754
755 if (q->nbands > q->nstrict) {
756 nest = nla_nest_start(skb, TCA_ETS_QUANTA);
757 if (!nest)
758 goto nla_err;
759
760 for (band = q->nstrict; band < q->nbands; band++) {
761 if (nla_put_u32(skb, TCA_ETS_QUANTA_BAND,
762 q->classes[band].quantum))
763 goto nla_err;
764 }
765
766 nla_nest_end(skb, nest);
767 }
768
769 nest = nla_nest_start(skb, TCA_ETS_PRIOMAP);
770 if (!nest)
771 goto nla_err;
772
773 for (prio = 0; prio <= TC_PRIO_MAX; prio++) {
774 if (nla_put_u8(skb, TCA_ETS_PRIOMAP_BAND, q->prio2band[prio]))
775 goto nla_err;
776 }
777
778 nla_nest_end(skb, nest);
779
780 return nla_nest_end(skb, opts);
781
782nla_err:
783 nla_nest_cancel(skb, opts);
784 return -EMSGSIZE;
785}
786
787static const struct Qdisc_class_ops ets_class_ops = {
788 .change = ets_class_change,
789 .graft = ets_class_graft,
790 .leaf = ets_class_leaf,
791 .find = ets_class_find,
792 .qlen_notify = ets_class_qlen_notify,
793 .dump = ets_class_dump,
794 .dump_stats = ets_class_dump_stats,
795 .walk = ets_qdisc_walk,
796 .tcf_block = ets_qdisc_tcf_block,
797 .bind_tcf = ets_qdisc_bind_tcf,
798 .unbind_tcf = ets_qdisc_unbind_tcf,
799};
800
801static struct Qdisc_ops ets_qdisc_ops __read_mostly = {
802 .cl_ops = &ets_class_ops,
803 .id = "ets",
804 .priv_size = sizeof(struct ets_sched),
805 .enqueue = ets_qdisc_enqueue,
806 .dequeue = ets_qdisc_dequeue,
807 .peek = qdisc_peek_dequeued,
808 .change = ets_qdisc_change,
809 .init = ets_qdisc_init,
810 .reset = ets_qdisc_reset,
811 .destroy = ets_qdisc_destroy,
812 .dump = ets_qdisc_dump,
813 .owner = THIS_MODULE,
814};
815
816static int __init ets_init(void)
817{
818 return register_qdisc(&ets_qdisc_ops);
819}
820
821static void __exit ets_exit(void)
822{
823 unregister_qdisc(&ets_qdisc_ops);
824}
825
826module_init(ets_init);
827module_exit(ets_exit);
828MODULE_LICENSE("GPL");