Loading...
1/*
2 * net/sched/sch_mqprio.c
3 *
4 * Copyright (c) 2010 John Fastabend <john.r.fastabend@intel.com>
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/types.h>
12#include <linux/slab.h>
13#include <linux/kernel.h>
14#include <linux/string.h>
15#include <linux/errno.h>
16#include <linux/skbuff.h>
17#include <linux/module.h>
18#include <net/netlink.h>
19#include <net/pkt_sched.h>
20#include <net/sch_generic.h>
21#include <net/pkt_cls.h>
22
23struct mqprio_sched {
24 struct Qdisc **qdiscs;
25 u16 mode;
26 u16 shaper;
27 int hw_offload;
28 u32 flags;
29 u64 min_rate[TC_QOPT_MAX_QUEUE];
30 u64 max_rate[TC_QOPT_MAX_QUEUE];
31};
32
33static void mqprio_destroy(struct Qdisc *sch)
34{
35 struct net_device *dev = qdisc_dev(sch);
36 struct mqprio_sched *priv = qdisc_priv(sch);
37 unsigned int ntx;
38
39 if (priv->qdiscs) {
40 for (ntx = 0;
41 ntx < dev->num_tx_queues && priv->qdiscs[ntx];
42 ntx++)
43 qdisc_destroy(priv->qdiscs[ntx]);
44 kfree(priv->qdiscs);
45 }
46
47 if (priv->hw_offload && dev->netdev_ops->ndo_setup_tc) {
48 struct tc_mqprio_qopt_offload mqprio = { { 0 } };
49
50 switch (priv->mode) {
51 case TC_MQPRIO_MODE_DCB:
52 case TC_MQPRIO_MODE_CHANNEL:
53 dev->netdev_ops->ndo_setup_tc(dev,
54 TC_SETUP_QDISC_MQPRIO,
55 &mqprio);
56 break;
57 default:
58 return;
59 }
60 } else {
61 netdev_set_num_tc(dev, 0);
62 }
63}
64
65static int mqprio_parse_opt(struct net_device *dev, struct tc_mqprio_qopt *qopt)
66{
67 int i, j;
68
69 /* Verify num_tc is not out of max range */
70 if (qopt->num_tc > TC_MAX_QUEUE)
71 return -EINVAL;
72
73 /* Verify priority mapping uses valid tcs */
74 for (i = 0; i < TC_BITMASK + 1; i++) {
75 if (qopt->prio_tc_map[i] >= qopt->num_tc)
76 return -EINVAL;
77 }
78
79 /* Limit qopt->hw to maximum supported offload value. Drivers have
80 * the option of overriding this later if they don't support the a
81 * given offload type.
82 */
83 if (qopt->hw > TC_MQPRIO_HW_OFFLOAD_MAX)
84 qopt->hw = TC_MQPRIO_HW_OFFLOAD_MAX;
85
86 /* If hardware offload is requested we will leave it to the device
87 * to either populate the queue counts itself or to validate the
88 * provided queue counts. If ndo_setup_tc is not present then
89 * hardware doesn't support offload and we should return an error.
90 */
91 if (qopt->hw)
92 return dev->netdev_ops->ndo_setup_tc ? 0 : -EINVAL;
93
94 for (i = 0; i < qopt->num_tc; i++) {
95 unsigned int last = qopt->offset[i] + qopt->count[i];
96
97 /* Verify the queue count is in tx range being equal to the
98 * real_num_tx_queues indicates the last queue is in use.
99 */
100 if (qopt->offset[i] >= dev->real_num_tx_queues ||
101 !qopt->count[i] ||
102 last > dev->real_num_tx_queues)
103 return -EINVAL;
104
105 /* Verify that the offset and counts do not overlap */
106 for (j = i + 1; j < qopt->num_tc; j++) {
107 if (last > qopt->offset[j])
108 return -EINVAL;
109 }
110 }
111
112 return 0;
113}
114
115static const struct nla_policy mqprio_policy[TCA_MQPRIO_MAX + 1] = {
116 [TCA_MQPRIO_MODE] = { .len = sizeof(u16) },
117 [TCA_MQPRIO_SHAPER] = { .len = sizeof(u16) },
118 [TCA_MQPRIO_MIN_RATE64] = { .type = NLA_NESTED },
119 [TCA_MQPRIO_MAX_RATE64] = { .type = NLA_NESTED },
120};
121
122static int parse_attr(struct nlattr *tb[], int maxtype, struct nlattr *nla,
123 const struct nla_policy *policy, int len)
124{
125 int nested_len = nla_len(nla) - NLA_ALIGN(len);
126
127 if (nested_len >= nla_attr_size(0))
128 return nla_parse(tb, maxtype, nla_data(nla) + NLA_ALIGN(len),
129 nested_len, policy, NULL);
130
131 memset(tb, 0, sizeof(struct nlattr *) * (maxtype + 1));
132 return 0;
133}
134
135static int mqprio_init(struct Qdisc *sch, struct nlattr *opt,
136 struct netlink_ext_ack *extack)
137{
138 struct net_device *dev = qdisc_dev(sch);
139 struct mqprio_sched *priv = qdisc_priv(sch);
140 struct netdev_queue *dev_queue;
141 struct Qdisc *qdisc;
142 int i, err = -EOPNOTSUPP;
143 struct tc_mqprio_qopt *qopt = NULL;
144 struct nlattr *tb[TCA_MQPRIO_MAX + 1];
145 struct nlattr *attr;
146 int rem;
147 int len;
148
149 BUILD_BUG_ON(TC_MAX_QUEUE != TC_QOPT_MAX_QUEUE);
150 BUILD_BUG_ON(TC_BITMASK != TC_QOPT_BITMASK);
151
152 if (sch->parent != TC_H_ROOT)
153 return -EOPNOTSUPP;
154
155 if (!netif_is_multiqueue(dev))
156 return -EOPNOTSUPP;
157
158 /* make certain can allocate enough classids to handle queues */
159 if (dev->num_tx_queues >= TC_H_MIN_PRIORITY)
160 return -ENOMEM;
161
162 if (!opt || nla_len(opt) < sizeof(*qopt))
163 return -EINVAL;
164
165 qopt = nla_data(opt);
166 if (mqprio_parse_opt(dev, qopt))
167 return -EINVAL;
168
169 len = nla_len(opt) - NLA_ALIGN(sizeof(*qopt));
170 if (len > 0) {
171 err = parse_attr(tb, TCA_MQPRIO_MAX, opt, mqprio_policy,
172 sizeof(*qopt));
173 if (err < 0)
174 return err;
175
176 if (!qopt->hw)
177 return -EINVAL;
178
179 if (tb[TCA_MQPRIO_MODE]) {
180 priv->flags |= TC_MQPRIO_F_MODE;
181 priv->mode = *(u16 *)nla_data(tb[TCA_MQPRIO_MODE]);
182 }
183
184 if (tb[TCA_MQPRIO_SHAPER]) {
185 priv->flags |= TC_MQPRIO_F_SHAPER;
186 priv->shaper = *(u16 *)nla_data(tb[TCA_MQPRIO_SHAPER]);
187 }
188
189 if (tb[TCA_MQPRIO_MIN_RATE64]) {
190 if (priv->shaper != TC_MQPRIO_SHAPER_BW_RATE)
191 return -EINVAL;
192 i = 0;
193 nla_for_each_nested(attr, tb[TCA_MQPRIO_MIN_RATE64],
194 rem) {
195 if (nla_type(attr) != TCA_MQPRIO_MIN_RATE64)
196 return -EINVAL;
197 if (i >= qopt->num_tc)
198 break;
199 priv->min_rate[i] = *(u64 *)nla_data(attr);
200 i++;
201 }
202 priv->flags |= TC_MQPRIO_F_MIN_RATE;
203 }
204
205 if (tb[TCA_MQPRIO_MAX_RATE64]) {
206 if (priv->shaper != TC_MQPRIO_SHAPER_BW_RATE)
207 return -EINVAL;
208 i = 0;
209 nla_for_each_nested(attr, tb[TCA_MQPRIO_MAX_RATE64],
210 rem) {
211 if (nla_type(attr) != TCA_MQPRIO_MAX_RATE64)
212 return -EINVAL;
213 if (i >= qopt->num_tc)
214 break;
215 priv->max_rate[i] = *(u64 *)nla_data(attr);
216 i++;
217 }
218 priv->flags |= TC_MQPRIO_F_MAX_RATE;
219 }
220 }
221
222 /* pre-allocate qdisc, attachment can't fail */
223 priv->qdiscs = kcalloc(dev->num_tx_queues, sizeof(priv->qdiscs[0]),
224 GFP_KERNEL);
225 if (!priv->qdiscs)
226 return -ENOMEM;
227
228 for (i = 0; i < dev->num_tx_queues; i++) {
229 dev_queue = netdev_get_tx_queue(dev, i);
230 qdisc = qdisc_create_dflt(dev_queue,
231 get_default_qdisc_ops(dev, i),
232 TC_H_MAKE(TC_H_MAJ(sch->handle),
233 TC_H_MIN(i + 1)), extack);
234 if (!qdisc)
235 return -ENOMEM;
236
237 priv->qdiscs[i] = qdisc;
238 qdisc->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
239 }
240
241 /* If the mqprio options indicate that hardware should own
242 * the queue mapping then run ndo_setup_tc otherwise use the
243 * supplied and verified mapping
244 */
245 if (qopt->hw) {
246 struct tc_mqprio_qopt_offload mqprio = {.qopt = *qopt};
247
248 switch (priv->mode) {
249 case TC_MQPRIO_MODE_DCB:
250 if (priv->shaper != TC_MQPRIO_SHAPER_DCB)
251 return -EINVAL;
252 break;
253 case TC_MQPRIO_MODE_CHANNEL:
254 mqprio.flags = priv->flags;
255 if (priv->flags & TC_MQPRIO_F_MODE)
256 mqprio.mode = priv->mode;
257 if (priv->flags & TC_MQPRIO_F_SHAPER)
258 mqprio.shaper = priv->shaper;
259 if (priv->flags & TC_MQPRIO_F_MIN_RATE)
260 for (i = 0; i < mqprio.qopt.num_tc; i++)
261 mqprio.min_rate[i] = priv->min_rate[i];
262 if (priv->flags & TC_MQPRIO_F_MAX_RATE)
263 for (i = 0; i < mqprio.qopt.num_tc; i++)
264 mqprio.max_rate[i] = priv->max_rate[i];
265 break;
266 default:
267 return -EINVAL;
268 }
269 err = dev->netdev_ops->ndo_setup_tc(dev,
270 TC_SETUP_QDISC_MQPRIO,
271 &mqprio);
272 if (err)
273 return err;
274
275 priv->hw_offload = mqprio.qopt.hw;
276 } else {
277 netdev_set_num_tc(dev, qopt->num_tc);
278 for (i = 0; i < qopt->num_tc; i++)
279 netdev_set_tc_queue(dev, i,
280 qopt->count[i], qopt->offset[i]);
281 }
282
283 /* Always use supplied priority mappings */
284 for (i = 0; i < TC_BITMASK + 1; i++)
285 netdev_set_prio_tc_map(dev, i, qopt->prio_tc_map[i]);
286
287 sch->flags |= TCQ_F_MQROOT;
288 return 0;
289}
290
291static void mqprio_attach(struct Qdisc *sch)
292{
293 struct net_device *dev = qdisc_dev(sch);
294 struct mqprio_sched *priv = qdisc_priv(sch);
295 struct Qdisc *qdisc, *old;
296 unsigned int ntx;
297
298 /* Attach underlying qdisc */
299 for (ntx = 0; ntx < dev->num_tx_queues; ntx++) {
300 qdisc = priv->qdiscs[ntx];
301 old = dev_graft_qdisc(qdisc->dev_queue, qdisc);
302 if (old)
303 qdisc_destroy(old);
304 if (ntx < dev->real_num_tx_queues)
305 qdisc_hash_add(qdisc, false);
306 }
307 kfree(priv->qdiscs);
308 priv->qdiscs = NULL;
309}
310
311static struct netdev_queue *mqprio_queue_get(struct Qdisc *sch,
312 unsigned long cl)
313{
314 struct net_device *dev = qdisc_dev(sch);
315 unsigned long ntx = cl - 1;
316
317 if (ntx >= dev->num_tx_queues)
318 return NULL;
319 return netdev_get_tx_queue(dev, ntx);
320}
321
322static int mqprio_graft(struct Qdisc *sch, unsigned long cl, struct Qdisc *new,
323 struct Qdisc **old, struct netlink_ext_ack *extack)
324{
325 struct net_device *dev = qdisc_dev(sch);
326 struct netdev_queue *dev_queue = mqprio_queue_get(sch, cl);
327
328 if (!dev_queue)
329 return -EINVAL;
330
331 if (dev->flags & IFF_UP)
332 dev_deactivate(dev);
333
334 *old = dev_graft_qdisc(dev_queue, new);
335
336 if (new)
337 new->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
338
339 if (dev->flags & IFF_UP)
340 dev_activate(dev);
341
342 return 0;
343}
344
345static int dump_rates(struct mqprio_sched *priv,
346 struct tc_mqprio_qopt *opt, struct sk_buff *skb)
347{
348 struct nlattr *nest;
349 int i;
350
351 if (priv->flags & TC_MQPRIO_F_MIN_RATE) {
352 nest = nla_nest_start(skb, TCA_MQPRIO_MIN_RATE64);
353 if (!nest)
354 goto nla_put_failure;
355
356 for (i = 0; i < opt->num_tc; i++) {
357 if (nla_put(skb, TCA_MQPRIO_MIN_RATE64,
358 sizeof(priv->min_rate[i]),
359 &priv->min_rate[i]))
360 goto nla_put_failure;
361 }
362 nla_nest_end(skb, nest);
363 }
364
365 if (priv->flags & TC_MQPRIO_F_MAX_RATE) {
366 nest = nla_nest_start(skb, TCA_MQPRIO_MAX_RATE64);
367 if (!nest)
368 goto nla_put_failure;
369
370 for (i = 0; i < opt->num_tc; i++) {
371 if (nla_put(skb, TCA_MQPRIO_MAX_RATE64,
372 sizeof(priv->max_rate[i]),
373 &priv->max_rate[i]))
374 goto nla_put_failure;
375 }
376 nla_nest_end(skb, nest);
377 }
378 return 0;
379
380nla_put_failure:
381 nla_nest_cancel(skb, nest);
382 return -1;
383}
384
385static int mqprio_dump(struct Qdisc *sch, struct sk_buff *skb)
386{
387 struct net_device *dev = qdisc_dev(sch);
388 struct mqprio_sched *priv = qdisc_priv(sch);
389 struct nlattr *nla = (struct nlattr *)skb_tail_pointer(skb);
390 struct tc_mqprio_qopt opt = { 0 };
391 struct Qdisc *qdisc;
392 unsigned int ntx, tc;
393
394 sch->q.qlen = 0;
395 memset(&sch->bstats, 0, sizeof(sch->bstats));
396 memset(&sch->qstats, 0, sizeof(sch->qstats));
397
398 /* MQ supports lockless qdiscs. However, statistics accounting needs
399 * to account for all, none, or a mix of locked and unlocked child
400 * qdiscs. Percpu stats are added to counters in-band and locking
401 * qdisc totals are added at end.
402 */
403 for (ntx = 0; ntx < dev->num_tx_queues; ntx++) {
404 qdisc = netdev_get_tx_queue(dev, ntx)->qdisc_sleeping;
405 spin_lock_bh(qdisc_lock(qdisc));
406
407 if (qdisc_is_percpu_stats(qdisc)) {
408 __u32 qlen = qdisc_qlen_sum(qdisc);
409
410 __gnet_stats_copy_basic(NULL, &sch->bstats,
411 qdisc->cpu_bstats,
412 &qdisc->bstats);
413 __gnet_stats_copy_queue(&sch->qstats,
414 qdisc->cpu_qstats,
415 &qdisc->qstats, qlen);
416 } else {
417 sch->q.qlen += qdisc->q.qlen;
418 sch->bstats.bytes += qdisc->bstats.bytes;
419 sch->bstats.packets += qdisc->bstats.packets;
420 sch->qstats.backlog += qdisc->qstats.backlog;
421 sch->qstats.drops += qdisc->qstats.drops;
422 sch->qstats.requeues += qdisc->qstats.requeues;
423 sch->qstats.overlimits += qdisc->qstats.overlimits;
424 }
425
426 spin_unlock_bh(qdisc_lock(qdisc));
427 }
428
429 opt.num_tc = netdev_get_num_tc(dev);
430 memcpy(opt.prio_tc_map, dev->prio_tc_map, sizeof(opt.prio_tc_map));
431 opt.hw = priv->hw_offload;
432
433 for (tc = 0; tc < netdev_get_num_tc(dev); tc++) {
434 opt.count[tc] = dev->tc_to_txq[tc].count;
435 opt.offset[tc] = dev->tc_to_txq[tc].offset;
436 }
437
438 if (nla_put(skb, TCA_OPTIONS, NLA_ALIGN(sizeof(opt)), &opt))
439 goto nla_put_failure;
440
441 if ((priv->flags & TC_MQPRIO_F_MODE) &&
442 nla_put_u16(skb, TCA_MQPRIO_MODE, priv->mode))
443 goto nla_put_failure;
444
445 if ((priv->flags & TC_MQPRIO_F_SHAPER) &&
446 nla_put_u16(skb, TCA_MQPRIO_SHAPER, priv->shaper))
447 goto nla_put_failure;
448
449 if ((priv->flags & TC_MQPRIO_F_MIN_RATE ||
450 priv->flags & TC_MQPRIO_F_MAX_RATE) &&
451 (dump_rates(priv, &opt, skb) != 0))
452 goto nla_put_failure;
453
454 return nla_nest_end(skb, nla);
455nla_put_failure:
456 nlmsg_trim(skb, nla);
457 return -1;
458}
459
460static struct Qdisc *mqprio_leaf(struct Qdisc *sch, unsigned long cl)
461{
462 struct netdev_queue *dev_queue = mqprio_queue_get(sch, cl);
463
464 if (!dev_queue)
465 return NULL;
466
467 return dev_queue->qdisc_sleeping;
468}
469
470static unsigned long mqprio_find(struct Qdisc *sch, u32 classid)
471{
472 struct net_device *dev = qdisc_dev(sch);
473 unsigned int ntx = TC_H_MIN(classid);
474
475 /* There are essentially two regions here that have valid classid
476 * values. The first region will have a classid value of 1 through
477 * num_tx_queues. All of these are backed by actual Qdiscs.
478 */
479 if (ntx < TC_H_MIN_PRIORITY)
480 return (ntx <= dev->num_tx_queues) ? ntx : 0;
481
482 /* The second region represents the hardware traffic classes. These
483 * are represented by classid values of TC_H_MIN_PRIORITY through
484 * TC_H_MIN_PRIORITY + netdev_get_num_tc - 1
485 */
486 return ((ntx - TC_H_MIN_PRIORITY) < netdev_get_num_tc(dev)) ? ntx : 0;
487}
488
489static int mqprio_dump_class(struct Qdisc *sch, unsigned long cl,
490 struct sk_buff *skb, struct tcmsg *tcm)
491{
492 if (cl < TC_H_MIN_PRIORITY) {
493 struct netdev_queue *dev_queue = mqprio_queue_get(sch, cl);
494 struct net_device *dev = qdisc_dev(sch);
495 int tc = netdev_txq_to_tc(dev, cl - 1);
496
497 tcm->tcm_parent = (tc < 0) ? 0 :
498 TC_H_MAKE(TC_H_MAJ(sch->handle),
499 TC_H_MIN(tc + TC_H_MIN_PRIORITY));
500 tcm->tcm_info = dev_queue->qdisc_sleeping->handle;
501 } else {
502 tcm->tcm_parent = TC_H_ROOT;
503 tcm->tcm_info = 0;
504 }
505 tcm->tcm_handle |= TC_H_MIN(cl);
506 return 0;
507}
508
509static int mqprio_dump_class_stats(struct Qdisc *sch, unsigned long cl,
510 struct gnet_dump *d)
511 __releases(d->lock)
512 __acquires(d->lock)
513{
514 if (cl >= TC_H_MIN_PRIORITY) {
515 int i;
516 __u32 qlen = 0;
517 struct gnet_stats_queue qstats = {0};
518 struct gnet_stats_basic_packed bstats = {0};
519 struct net_device *dev = qdisc_dev(sch);
520 struct netdev_tc_txq tc = dev->tc_to_txq[cl & TC_BITMASK];
521
522 /* Drop lock here it will be reclaimed before touching
523 * statistics this is required because the d->lock we
524 * hold here is the look on dev_queue->qdisc_sleeping
525 * also acquired below.
526 */
527 if (d->lock)
528 spin_unlock_bh(d->lock);
529
530 for (i = tc.offset; i < tc.offset + tc.count; i++) {
531 struct netdev_queue *q = netdev_get_tx_queue(dev, i);
532 struct Qdisc *qdisc = rtnl_dereference(q->qdisc);
533 struct gnet_stats_basic_cpu __percpu *cpu_bstats = NULL;
534 struct gnet_stats_queue __percpu *cpu_qstats = NULL;
535
536 spin_lock_bh(qdisc_lock(qdisc));
537 if (qdisc_is_percpu_stats(qdisc)) {
538 cpu_bstats = qdisc->cpu_bstats;
539 cpu_qstats = qdisc->cpu_qstats;
540 }
541
542 qlen = qdisc_qlen_sum(qdisc);
543 __gnet_stats_copy_basic(NULL, &sch->bstats,
544 cpu_bstats, &qdisc->bstats);
545 __gnet_stats_copy_queue(&sch->qstats,
546 cpu_qstats,
547 &qdisc->qstats,
548 qlen);
549 spin_unlock_bh(qdisc_lock(qdisc));
550 }
551
552 /* Reclaim root sleeping lock before completing stats */
553 if (d->lock)
554 spin_lock_bh(d->lock);
555 if (gnet_stats_copy_basic(NULL, d, NULL, &bstats) < 0 ||
556 gnet_stats_copy_queue(d, NULL, &qstats, qlen) < 0)
557 return -1;
558 } else {
559 struct netdev_queue *dev_queue = mqprio_queue_get(sch, cl);
560
561 sch = dev_queue->qdisc_sleeping;
562 if (gnet_stats_copy_basic(qdisc_root_sleeping_running(sch),
563 d, NULL, &sch->bstats) < 0 ||
564 gnet_stats_copy_queue(d, NULL,
565 &sch->qstats, sch->q.qlen) < 0)
566 return -1;
567 }
568 return 0;
569}
570
571static void mqprio_walk(struct Qdisc *sch, struct qdisc_walker *arg)
572{
573 struct net_device *dev = qdisc_dev(sch);
574 unsigned long ntx;
575
576 if (arg->stop)
577 return;
578
579 /* Walk hierarchy with a virtual class per tc */
580 arg->count = arg->skip;
581 for (ntx = arg->skip; ntx < netdev_get_num_tc(dev); ntx++) {
582 if (arg->fn(sch, ntx + TC_H_MIN_PRIORITY, arg) < 0) {
583 arg->stop = 1;
584 return;
585 }
586 arg->count++;
587 }
588
589 /* Pad the values and skip over unused traffic classes */
590 if (ntx < TC_MAX_QUEUE) {
591 arg->count = TC_MAX_QUEUE;
592 ntx = TC_MAX_QUEUE;
593 }
594
595 /* Reset offset, sort out remaining per-queue qdiscs */
596 for (ntx -= TC_MAX_QUEUE; ntx < dev->num_tx_queues; ntx++) {
597 if (arg->fn(sch, ntx + 1, arg) < 0) {
598 arg->stop = 1;
599 return;
600 }
601 arg->count++;
602 }
603}
604
605static struct netdev_queue *mqprio_select_queue(struct Qdisc *sch,
606 struct tcmsg *tcm)
607{
608 return mqprio_queue_get(sch, TC_H_MIN(tcm->tcm_parent));
609}
610
611static const struct Qdisc_class_ops mqprio_class_ops = {
612 .graft = mqprio_graft,
613 .leaf = mqprio_leaf,
614 .find = mqprio_find,
615 .walk = mqprio_walk,
616 .dump = mqprio_dump_class,
617 .dump_stats = mqprio_dump_class_stats,
618 .select_queue = mqprio_select_queue,
619};
620
621static struct Qdisc_ops mqprio_qdisc_ops __read_mostly = {
622 .cl_ops = &mqprio_class_ops,
623 .id = "mqprio",
624 .priv_size = sizeof(struct mqprio_sched),
625 .init = mqprio_init,
626 .destroy = mqprio_destroy,
627 .attach = mqprio_attach,
628 .dump = mqprio_dump,
629 .owner = THIS_MODULE,
630};
631
632static int __init mqprio_module_init(void)
633{
634 return register_qdisc(&mqprio_qdisc_ops);
635}
636
637static void __exit mqprio_module_exit(void)
638{
639 unregister_qdisc(&mqprio_qdisc_ops);
640}
641
642module_init(mqprio_module_init);
643module_exit(mqprio_module_exit);
644
645MODULE_LICENSE("GPL");
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * net/sched/sch_mqprio.c
4 *
5 * Copyright (c) 2010 John Fastabend <john.r.fastabend@intel.com>
6 */
7
8#include <linux/ethtool_netlink.h>
9#include <linux/types.h>
10#include <linux/slab.h>
11#include <linux/kernel.h>
12#include <linux/string.h>
13#include <linux/errno.h>
14#include <linux/skbuff.h>
15#include <linux/module.h>
16#include <net/netlink.h>
17#include <net/pkt_sched.h>
18#include <net/sch_generic.h>
19#include <net/pkt_cls.h>
20
21#include "sch_mqprio_lib.h"
22
23struct mqprio_sched {
24 struct Qdisc **qdiscs;
25 u16 mode;
26 u16 shaper;
27 int hw_offload;
28 u32 flags;
29 u64 min_rate[TC_QOPT_MAX_QUEUE];
30 u64 max_rate[TC_QOPT_MAX_QUEUE];
31 u32 fp[TC_QOPT_MAX_QUEUE];
32};
33
34static int mqprio_enable_offload(struct Qdisc *sch,
35 const struct tc_mqprio_qopt *qopt,
36 struct netlink_ext_ack *extack)
37{
38 struct mqprio_sched *priv = qdisc_priv(sch);
39 struct net_device *dev = qdisc_dev(sch);
40 struct tc_mqprio_qopt_offload mqprio = {
41 .qopt = *qopt,
42 .extack = extack,
43 };
44 int err, i;
45
46 switch (priv->mode) {
47 case TC_MQPRIO_MODE_DCB:
48 if (priv->shaper != TC_MQPRIO_SHAPER_DCB)
49 return -EINVAL;
50 break;
51 case TC_MQPRIO_MODE_CHANNEL:
52 mqprio.flags = priv->flags;
53 if (priv->flags & TC_MQPRIO_F_MODE)
54 mqprio.mode = priv->mode;
55 if (priv->flags & TC_MQPRIO_F_SHAPER)
56 mqprio.shaper = priv->shaper;
57 if (priv->flags & TC_MQPRIO_F_MIN_RATE)
58 for (i = 0; i < mqprio.qopt.num_tc; i++)
59 mqprio.min_rate[i] = priv->min_rate[i];
60 if (priv->flags & TC_MQPRIO_F_MAX_RATE)
61 for (i = 0; i < mqprio.qopt.num_tc; i++)
62 mqprio.max_rate[i] = priv->max_rate[i];
63 break;
64 default:
65 return -EINVAL;
66 }
67
68 mqprio_fp_to_offload(priv->fp, &mqprio);
69
70 err = dev->netdev_ops->ndo_setup_tc(dev, TC_SETUP_QDISC_MQPRIO,
71 &mqprio);
72 if (err)
73 return err;
74
75 priv->hw_offload = mqprio.qopt.hw;
76
77 return 0;
78}
79
80static void mqprio_disable_offload(struct Qdisc *sch)
81{
82 struct tc_mqprio_qopt_offload mqprio = { { 0 } };
83 struct mqprio_sched *priv = qdisc_priv(sch);
84 struct net_device *dev = qdisc_dev(sch);
85
86 switch (priv->mode) {
87 case TC_MQPRIO_MODE_DCB:
88 case TC_MQPRIO_MODE_CHANNEL:
89 dev->netdev_ops->ndo_setup_tc(dev, TC_SETUP_QDISC_MQPRIO,
90 &mqprio);
91 break;
92 }
93}
94
95static void mqprio_destroy(struct Qdisc *sch)
96{
97 struct net_device *dev = qdisc_dev(sch);
98 struct mqprio_sched *priv = qdisc_priv(sch);
99 unsigned int ntx;
100
101 if (priv->qdiscs) {
102 for (ntx = 0;
103 ntx < dev->num_tx_queues && priv->qdiscs[ntx];
104 ntx++)
105 qdisc_put(priv->qdiscs[ntx]);
106 kfree(priv->qdiscs);
107 }
108
109 if (priv->hw_offload && dev->netdev_ops->ndo_setup_tc)
110 mqprio_disable_offload(sch);
111 else
112 netdev_set_num_tc(dev, 0);
113}
114
115static int mqprio_parse_opt(struct net_device *dev, struct tc_mqprio_qopt *qopt,
116 const struct tc_mqprio_caps *caps,
117 struct netlink_ext_ack *extack)
118{
119 int err;
120
121 /* Limit qopt->hw to maximum supported offload value. Drivers have
122 * the option of overriding this later if they don't support the a
123 * given offload type.
124 */
125 if (qopt->hw > TC_MQPRIO_HW_OFFLOAD_MAX)
126 qopt->hw = TC_MQPRIO_HW_OFFLOAD_MAX;
127
128 /* If hardware offload is requested, we will leave 3 options to the
129 * device driver:
130 * - populate the queue counts itself (and ignore what was requested)
131 * - validate the provided queue counts by itself (and apply them)
132 * - request queue count validation here (and apply them)
133 */
134 err = mqprio_validate_qopt(dev, qopt,
135 !qopt->hw || caps->validate_queue_counts,
136 false, extack);
137 if (err)
138 return err;
139
140 /* If ndo_setup_tc is not present then hardware doesn't support offload
141 * and we should return an error.
142 */
143 if (qopt->hw && !dev->netdev_ops->ndo_setup_tc) {
144 NL_SET_ERR_MSG(extack,
145 "Device does not support hardware offload");
146 return -EINVAL;
147 }
148
149 return 0;
150}
151
152static const struct
153nla_policy mqprio_tc_entry_policy[TCA_MQPRIO_TC_ENTRY_MAX + 1] = {
154 [TCA_MQPRIO_TC_ENTRY_INDEX] = NLA_POLICY_MAX(NLA_U32,
155 TC_QOPT_MAX_QUEUE),
156 [TCA_MQPRIO_TC_ENTRY_FP] = NLA_POLICY_RANGE(NLA_U32,
157 TC_FP_EXPRESS,
158 TC_FP_PREEMPTIBLE),
159};
160
161static const struct nla_policy mqprio_policy[TCA_MQPRIO_MAX + 1] = {
162 [TCA_MQPRIO_MODE] = { .len = sizeof(u16) },
163 [TCA_MQPRIO_SHAPER] = { .len = sizeof(u16) },
164 [TCA_MQPRIO_MIN_RATE64] = { .type = NLA_NESTED },
165 [TCA_MQPRIO_MAX_RATE64] = { .type = NLA_NESTED },
166 [TCA_MQPRIO_TC_ENTRY] = { .type = NLA_NESTED },
167};
168
169static int mqprio_parse_tc_entry(u32 fp[TC_QOPT_MAX_QUEUE],
170 struct nlattr *opt,
171 unsigned long *seen_tcs,
172 struct netlink_ext_ack *extack)
173{
174 struct nlattr *tb[TCA_MQPRIO_TC_ENTRY_MAX + 1];
175 int err, tc;
176
177 err = nla_parse_nested(tb, TCA_MQPRIO_TC_ENTRY_MAX, opt,
178 mqprio_tc_entry_policy, extack);
179 if (err < 0)
180 return err;
181
182 if (NL_REQ_ATTR_CHECK(extack, opt, tb, TCA_MQPRIO_TC_ENTRY_INDEX)) {
183 NL_SET_ERR_MSG(extack, "TC entry index missing");
184 return -EINVAL;
185 }
186
187 tc = nla_get_u32(tb[TCA_MQPRIO_TC_ENTRY_INDEX]);
188 if (*seen_tcs & BIT(tc)) {
189 NL_SET_ERR_MSG_ATTR(extack, tb[TCA_MQPRIO_TC_ENTRY_INDEX],
190 "Duplicate tc entry");
191 return -EINVAL;
192 }
193
194 *seen_tcs |= BIT(tc);
195
196 if (tb[TCA_MQPRIO_TC_ENTRY_FP])
197 fp[tc] = nla_get_u32(tb[TCA_MQPRIO_TC_ENTRY_FP]);
198
199 return 0;
200}
201
202static int mqprio_parse_tc_entries(struct Qdisc *sch, struct nlattr *nlattr_opt,
203 int nlattr_opt_len,
204 struct netlink_ext_ack *extack)
205{
206 struct mqprio_sched *priv = qdisc_priv(sch);
207 struct net_device *dev = qdisc_dev(sch);
208 bool have_preemption = false;
209 unsigned long seen_tcs = 0;
210 u32 fp[TC_QOPT_MAX_QUEUE];
211 struct nlattr *n;
212 int tc, rem;
213 int err = 0;
214
215 for (tc = 0; tc < TC_QOPT_MAX_QUEUE; tc++)
216 fp[tc] = priv->fp[tc];
217
218 nla_for_each_attr_type(n, TCA_MQPRIO_TC_ENTRY, nlattr_opt,
219 nlattr_opt_len, rem) {
220 err = mqprio_parse_tc_entry(fp, n, &seen_tcs, extack);
221 if (err)
222 goto out;
223 }
224
225 for (tc = 0; tc < TC_QOPT_MAX_QUEUE; tc++) {
226 priv->fp[tc] = fp[tc];
227 if (fp[tc] == TC_FP_PREEMPTIBLE)
228 have_preemption = true;
229 }
230
231 if (have_preemption && !ethtool_dev_mm_supported(dev)) {
232 NL_SET_ERR_MSG(extack, "Device does not support preemption");
233 return -EOPNOTSUPP;
234 }
235out:
236 return err;
237}
238
239/* Parse the other netlink attributes that represent the payload of
240 * TCA_OPTIONS, which are appended right after struct tc_mqprio_qopt.
241 */
242static int mqprio_parse_nlattr(struct Qdisc *sch, struct tc_mqprio_qopt *qopt,
243 struct nlattr *opt,
244 struct netlink_ext_ack *extack)
245{
246 struct nlattr *nlattr_opt = nla_data(opt) + NLA_ALIGN(sizeof(*qopt));
247 int nlattr_opt_len = nla_len(opt) - NLA_ALIGN(sizeof(*qopt));
248 struct mqprio_sched *priv = qdisc_priv(sch);
249 struct nlattr *tb[TCA_MQPRIO_MAX + 1] = {};
250 struct nlattr *attr;
251 int i, rem, err;
252
253 if (nlattr_opt_len >= nla_attr_size(0)) {
254 err = nla_parse_deprecated(tb, TCA_MQPRIO_MAX, nlattr_opt,
255 nlattr_opt_len, mqprio_policy,
256 NULL);
257 if (err < 0)
258 return err;
259 }
260
261 if (!qopt->hw) {
262 NL_SET_ERR_MSG(extack,
263 "mqprio TCA_OPTIONS can only contain netlink attributes in hardware mode");
264 return -EINVAL;
265 }
266
267 if (tb[TCA_MQPRIO_MODE]) {
268 priv->flags |= TC_MQPRIO_F_MODE;
269 priv->mode = nla_get_u16(tb[TCA_MQPRIO_MODE]);
270 }
271
272 if (tb[TCA_MQPRIO_SHAPER]) {
273 priv->flags |= TC_MQPRIO_F_SHAPER;
274 priv->shaper = nla_get_u16(tb[TCA_MQPRIO_SHAPER]);
275 }
276
277 if (tb[TCA_MQPRIO_MIN_RATE64]) {
278 if (priv->shaper != TC_MQPRIO_SHAPER_BW_RATE) {
279 NL_SET_ERR_MSG_ATTR(extack, tb[TCA_MQPRIO_MIN_RATE64],
280 "min_rate accepted only when shaper is in bw_rlimit mode");
281 return -EINVAL;
282 }
283 i = 0;
284 nla_for_each_nested(attr, tb[TCA_MQPRIO_MIN_RATE64],
285 rem) {
286 if (nla_type(attr) != TCA_MQPRIO_MIN_RATE64) {
287 NL_SET_ERR_MSG_ATTR(extack, attr,
288 "Attribute type expected to be TCA_MQPRIO_MIN_RATE64");
289 return -EINVAL;
290 }
291
292 if (nla_len(attr) != sizeof(u64)) {
293 NL_SET_ERR_MSG_ATTR(extack, attr,
294 "Attribute TCA_MQPRIO_MIN_RATE64 expected to have 8 bytes length");
295 return -EINVAL;
296 }
297
298 if (i >= qopt->num_tc)
299 break;
300 priv->min_rate[i] = nla_get_u64(attr);
301 i++;
302 }
303 priv->flags |= TC_MQPRIO_F_MIN_RATE;
304 }
305
306 if (tb[TCA_MQPRIO_MAX_RATE64]) {
307 if (priv->shaper != TC_MQPRIO_SHAPER_BW_RATE) {
308 NL_SET_ERR_MSG_ATTR(extack, tb[TCA_MQPRIO_MAX_RATE64],
309 "max_rate accepted only when shaper is in bw_rlimit mode");
310 return -EINVAL;
311 }
312 i = 0;
313 nla_for_each_nested(attr, tb[TCA_MQPRIO_MAX_RATE64],
314 rem) {
315 if (nla_type(attr) != TCA_MQPRIO_MAX_RATE64) {
316 NL_SET_ERR_MSG_ATTR(extack, attr,
317 "Attribute type expected to be TCA_MQPRIO_MAX_RATE64");
318 return -EINVAL;
319 }
320
321 if (nla_len(attr) != sizeof(u64)) {
322 NL_SET_ERR_MSG_ATTR(extack, attr,
323 "Attribute TCA_MQPRIO_MAX_RATE64 expected to have 8 bytes length");
324 return -EINVAL;
325 }
326
327 if (i >= qopt->num_tc)
328 break;
329 priv->max_rate[i] = nla_get_u64(attr);
330 i++;
331 }
332 priv->flags |= TC_MQPRIO_F_MAX_RATE;
333 }
334
335 if (tb[TCA_MQPRIO_TC_ENTRY]) {
336 err = mqprio_parse_tc_entries(sch, nlattr_opt, nlattr_opt_len,
337 extack);
338 if (err)
339 return err;
340 }
341
342 return 0;
343}
344
345static int mqprio_init(struct Qdisc *sch, struct nlattr *opt,
346 struct netlink_ext_ack *extack)
347{
348 struct net_device *dev = qdisc_dev(sch);
349 struct mqprio_sched *priv = qdisc_priv(sch);
350 struct netdev_queue *dev_queue;
351 struct Qdisc *qdisc;
352 int i, err = -EOPNOTSUPP;
353 struct tc_mqprio_qopt *qopt = NULL;
354 struct tc_mqprio_caps caps;
355 int len, tc;
356
357 BUILD_BUG_ON(TC_MAX_QUEUE != TC_QOPT_MAX_QUEUE);
358 BUILD_BUG_ON(TC_BITMASK != TC_QOPT_BITMASK);
359
360 if (sch->parent != TC_H_ROOT)
361 return -EOPNOTSUPP;
362
363 if (!netif_is_multiqueue(dev))
364 return -EOPNOTSUPP;
365
366 /* make certain can allocate enough classids to handle queues */
367 if (dev->num_tx_queues >= TC_H_MIN_PRIORITY)
368 return -ENOMEM;
369
370 if (!opt || nla_len(opt) < sizeof(*qopt))
371 return -EINVAL;
372
373 for (tc = 0; tc < TC_QOPT_MAX_QUEUE; tc++)
374 priv->fp[tc] = TC_FP_EXPRESS;
375
376 qdisc_offload_query_caps(dev, TC_SETUP_QDISC_MQPRIO,
377 &caps, sizeof(caps));
378
379 qopt = nla_data(opt);
380 if (mqprio_parse_opt(dev, qopt, &caps, extack))
381 return -EINVAL;
382
383 len = nla_len(opt) - NLA_ALIGN(sizeof(*qopt));
384 if (len > 0) {
385 err = mqprio_parse_nlattr(sch, qopt, opt, extack);
386 if (err)
387 return err;
388 }
389
390 /* pre-allocate qdisc, attachment can't fail */
391 priv->qdiscs = kcalloc(dev->num_tx_queues, sizeof(priv->qdiscs[0]),
392 GFP_KERNEL);
393 if (!priv->qdiscs)
394 return -ENOMEM;
395
396 for (i = 0; i < dev->num_tx_queues; i++) {
397 dev_queue = netdev_get_tx_queue(dev, i);
398 qdisc = qdisc_create_dflt(dev_queue,
399 get_default_qdisc_ops(dev, i),
400 TC_H_MAKE(TC_H_MAJ(sch->handle),
401 TC_H_MIN(i + 1)), extack);
402 if (!qdisc)
403 return -ENOMEM;
404
405 priv->qdiscs[i] = qdisc;
406 qdisc->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
407 }
408
409 /* If the mqprio options indicate that hardware should own
410 * the queue mapping then run ndo_setup_tc otherwise use the
411 * supplied and verified mapping
412 */
413 if (qopt->hw) {
414 err = mqprio_enable_offload(sch, qopt, extack);
415 if (err)
416 return err;
417 } else {
418 netdev_set_num_tc(dev, qopt->num_tc);
419 for (i = 0; i < qopt->num_tc; i++)
420 netdev_set_tc_queue(dev, i,
421 qopt->count[i], qopt->offset[i]);
422 }
423
424 /* Always use supplied priority mappings */
425 for (i = 0; i < TC_BITMASK + 1; i++)
426 netdev_set_prio_tc_map(dev, i, qopt->prio_tc_map[i]);
427
428 sch->flags |= TCQ_F_MQROOT;
429 return 0;
430}
431
432static void mqprio_attach(struct Qdisc *sch)
433{
434 struct net_device *dev = qdisc_dev(sch);
435 struct mqprio_sched *priv = qdisc_priv(sch);
436 struct Qdisc *qdisc, *old;
437 unsigned int ntx;
438
439 /* Attach underlying qdisc */
440 for (ntx = 0; ntx < dev->num_tx_queues; ntx++) {
441 qdisc = priv->qdiscs[ntx];
442 old = dev_graft_qdisc(qdisc->dev_queue, qdisc);
443 if (old)
444 qdisc_put(old);
445 if (ntx < dev->real_num_tx_queues)
446 qdisc_hash_add(qdisc, false);
447 }
448 kfree(priv->qdiscs);
449 priv->qdiscs = NULL;
450}
451
452static struct netdev_queue *mqprio_queue_get(struct Qdisc *sch,
453 unsigned long cl)
454{
455 struct net_device *dev = qdisc_dev(sch);
456 unsigned long ntx = cl - 1;
457
458 if (ntx >= dev->num_tx_queues)
459 return NULL;
460 return netdev_get_tx_queue(dev, ntx);
461}
462
463static int mqprio_graft(struct Qdisc *sch, unsigned long cl, struct Qdisc *new,
464 struct Qdisc **old, struct netlink_ext_ack *extack)
465{
466 struct net_device *dev = qdisc_dev(sch);
467 struct netdev_queue *dev_queue = mqprio_queue_get(sch, cl);
468
469 if (!dev_queue)
470 return -EINVAL;
471
472 if (dev->flags & IFF_UP)
473 dev_deactivate(dev);
474
475 *old = dev_graft_qdisc(dev_queue, new);
476
477 if (new)
478 new->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
479
480 if (dev->flags & IFF_UP)
481 dev_activate(dev);
482
483 return 0;
484}
485
486static int dump_rates(struct mqprio_sched *priv,
487 struct tc_mqprio_qopt *opt, struct sk_buff *skb)
488{
489 struct nlattr *nest;
490 int i;
491
492 if (priv->flags & TC_MQPRIO_F_MIN_RATE) {
493 nest = nla_nest_start_noflag(skb, TCA_MQPRIO_MIN_RATE64);
494 if (!nest)
495 goto nla_put_failure;
496
497 for (i = 0; i < opt->num_tc; i++) {
498 if (nla_put(skb, TCA_MQPRIO_MIN_RATE64,
499 sizeof(priv->min_rate[i]),
500 &priv->min_rate[i]))
501 goto nla_put_failure;
502 }
503 nla_nest_end(skb, nest);
504 }
505
506 if (priv->flags & TC_MQPRIO_F_MAX_RATE) {
507 nest = nla_nest_start_noflag(skb, TCA_MQPRIO_MAX_RATE64);
508 if (!nest)
509 goto nla_put_failure;
510
511 for (i = 0; i < opt->num_tc; i++) {
512 if (nla_put(skb, TCA_MQPRIO_MAX_RATE64,
513 sizeof(priv->max_rate[i]),
514 &priv->max_rate[i]))
515 goto nla_put_failure;
516 }
517 nla_nest_end(skb, nest);
518 }
519 return 0;
520
521nla_put_failure:
522 nla_nest_cancel(skb, nest);
523 return -1;
524}
525
526static int mqprio_dump_tc_entries(struct mqprio_sched *priv,
527 struct sk_buff *skb)
528{
529 struct nlattr *n;
530 int tc;
531
532 for (tc = 0; tc < TC_QOPT_MAX_QUEUE; tc++) {
533 n = nla_nest_start(skb, TCA_MQPRIO_TC_ENTRY);
534 if (!n)
535 return -EMSGSIZE;
536
537 if (nla_put_u32(skb, TCA_MQPRIO_TC_ENTRY_INDEX, tc))
538 goto nla_put_failure;
539
540 if (nla_put_u32(skb, TCA_MQPRIO_TC_ENTRY_FP, priv->fp[tc]))
541 goto nla_put_failure;
542
543 nla_nest_end(skb, n);
544 }
545
546 return 0;
547
548nla_put_failure:
549 nla_nest_cancel(skb, n);
550 return -EMSGSIZE;
551}
552
553static int mqprio_dump(struct Qdisc *sch, struct sk_buff *skb)
554{
555 struct net_device *dev = qdisc_dev(sch);
556 struct mqprio_sched *priv = qdisc_priv(sch);
557 struct nlattr *nla = (struct nlattr *)skb_tail_pointer(skb);
558 struct tc_mqprio_qopt opt = { 0 };
559 struct Qdisc *qdisc;
560 unsigned int ntx;
561
562 sch->q.qlen = 0;
563 gnet_stats_basic_sync_init(&sch->bstats);
564 memset(&sch->qstats, 0, sizeof(sch->qstats));
565
566 /* MQ supports lockless qdiscs. However, statistics accounting needs
567 * to account for all, none, or a mix of locked and unlocked child
568 * qdiscs. Percpu stats are added to counters in-band and locking
569 * qdisc totals are added at end.
570 */
571 for (ntx = 0; ntx < dev->num_tx_queues; ntx++) {
572 qdisc = rtnl_dereference(netdev_get_tx_queue(dev, ntx)->qdisc_sleeping);
573 spin_lock_bh(qdisc_lock(qdisc));
574
575 gnet_stats_add_basic(&sch->bstats, qdisc->cpu_bstats,
576 &qdisc->bstats, false);
577 gnet_stats_add_queue(&sch->qstats, qdisc->cpu_qstats,
578 &qdisc->qstats);
579 sch->q.qlen += qdisc_qlen(qdisc);
580
581 spin_unlock_bh(qdisc_lock(qdisc));
582 }
583
584 mqprio_qopt_reconstruct(dev, &opt);
585 opt.hw = priv->hw_offload;
586
587 if (nla_put(skb, TCA_OPTIONS, sizeof(opt), &opt))
588 goto nla_put_failure;
589
590 if ((priv->flags & TC_MQPRIO_F_MODE) &&
591 nla_put_u16(skb, TCA_MQPRIO_MODE, priv->mode))
592 goto nla_put_failure;
593
594 if ((priv->flags & TC_MQPRIO_F_SHAPER) &&
595 nla_put_u16(skb, TCA_MQPRIO_SHAPER, priv->shaper))
596 goto nla_put_failure;
597
598 if ((priv->flags & TC_MQPRIO_F_MIN_RATE ||
599 priv->flags & TC_MQPRIO_F_MAX_RATE) &&
600 (dump_rates(priv, &opt, skb) != 0))
601 goto nla_put_failure;
602
603 if (mqprio_dump_tc_entries(priv, skb))
604 goto nla_put_failure;
605
606 return nla_nest_end(skb, nla);
607nla_put_failure:
608 nlmsg_trim(skb, nla);
609 return -1;
610}
611
612static struct Qdisc *mqprio_leaf(struct Qdisc *sch, unsigned long cl)
613{
614 struct netdev_queue *dev_queue = mqprio_queue_get(sch, cl);
615
616 if (!dev_queue)
617 return NULL;
618
619 return rtnl_dereference(dev_queue->qdisc_sleeping);
620}
621
622static unsigned long mqprio_find(struct Qdisc *sch, u32 classid)
623{
624 struct net_device *dev = qdisc_dev(sch);
625 unsigned int ntx = TC_H_MIN(classid);
626
627 /* There are essentially two regions here that have valid classid
628 * values. The first region will have a classid value of 1 through
629 * num_tx_queues. All of these are backed by actual Qdiscs.
630 */
631 if (ntx < TC_H_MIN_PRIORITY)
632 return (ntx <= dev->num_tx_queues) ? ntx : 0;
633
634 /* The second region represents the hardware traffic classes. These
635 * are represented by classid values of TC_H_MIN_PRIORITY through
636 * TC_H_MIN_PRIORITY + netdev_get_num_tc - 1
637 */
638 return ((ntx - TC_H_MIN_PRIORITY) < netdev_get_num_tc(dev)) ? ntx : 0;
639}
640
641static int mqprio_dump_class(struct Qdisc *sch, unsigned long cl,
642 struct sk_buff *skb, struct tcmsg *tcm)
643{
644 if (cl < TC_H_MIN_PRIORITY) {
645 struct netdev_queue *dev_queue = mqprio_queue_get(sch, cl);
646 struct net_device *dev = qdisc_dev(sch);
647 int tc = netdev_txq_to_tc(dev, cl - 1);
648
649 tcm->tcm_parent = (tc < 0) ? 0 :
650 TC_H_MAKE(TC_H_MAJ(sch->handle),
651 TC_H_MIN(tc + TC_H_MIN_PRIORITY));
652 tcm->tcm_info = rtnl_dereference(dev_queue->qdisc_sleeping)->handle;
653 } else {
654 tcm->tcm_parent = TC_H_ROOT;
655 tcm->tcm_info = 0;
656 }
657 tcm->tcm_handle |= TC_H_MIN(cl);
658 return 0;
659}
660
661static int mqprio_dump_class_stats(struct Qdisc *sch, unsigned long cl,
662 struct gnet_dump *d)
663 __releases(d->lock)
664 __acquires(d->lock)
665{
666 if (cl >= TC_H_MIN_PRIORITY) {
667 int i;
668 __u32 qlen;
669 struct gnet_stats_queue qstats = {0};
670 struct gnet_stats_basic_sync bstats;
671 struct net_device *dev = qdisc_dev(sch);
672 struct netdev_tc_txq tc = dev->tc_to_txq[cl & TC_BITMASK];
673
674 gnet_stats_basic_sync_init(&bstats);
675 /* Drop lock here it will be reclaimed before touching
676 * statistics this is required because the d->lock we
677 * hold here is the look on dev_queue->qdisc_sleeping
678 * also acquired below.
679 */
680 if (d->lock)
681 spin_unlock_bh(d->lock);
682
683 for (i = tc.offset; i < tc.offset + tc.count; i++) {
684 struct netdev_queue *q = netdev_get_tx_queue(dev, i);
685 struct Qdisc *qdisc = rtnl_dereference(q->qdisc);
686
687 spin_lock_bh(qdisc_lock(qdisc));
688
689 gnet_stats_add_basic(&bstats, qdisc->cpu_bstats,
690 &qdisc->bstats, false);
691 gnet_stats_add_queue(&qstats, qdisc->cpu_qstats,
692 &qdisc->qstats);
693 sch->q.qlen += qdisc_qlen(qdisc);
694
695 spin_unlock_bh(qdisc_lock(qdisc));
696 }
697 qlen = qdisc_qlen(sch) + qstats.qlen;
698
699 /* Reclaim root sleeping lock before completing stats */
700 if (d->lock)
701 spin_lock_bh(d->lock);
702 if (gnet_stats_copy_basic(d, NULL, &bstats, false) < 0 ||
703 gnet_stats_copy_queue(d, NULL, &qstats, qlen) < 0)
704 return -1;
705 } else {
706 struct netdev_queue *dev_queue = mqprio_queue_get(sch, cl);
707
708 sch = rtnl_dereference(dev_queue->qdisc_sleeping);
709 if (gnet_stats_copy_basic(d, sch->cpu_bstats,
710 &sch->bstats, true) < 0 ||
711 qdisc_qstats_copy(d, sch) < 0)
712 return -1;
713 }
714 return 0;
715}
716
717static void mqprio_walk(struct Qdisc *sch, struct qdisc_walker *arg)
718{
719 struct net_device *dev = qdisc_dev(sch);
720 unsigned long ntx;
721
722 if (arg->stop)
723 return;
724
725 /* Walk hierarchy with a virtual class per tc */
726 arg->count = arg->skip;
727 for (ntx = arg->skip; ntx < netdev_get_num_tc(dev); ntx++) {
728 if (!tc_qdisc_stats_dump(sch, ntx + TC_H_MIN_PRIORITY, arg))
729 return;
730 }
731
732 /* Pad the values and skip over unused traffic classes */
733 if (ntx < TC_MAX_QUEUE) {
734 arg->count = TC_MAX_QUEUE;
735 ntx = TC_MAX_QUEUE;
736 }
737
738 /* Reset offset, sort out remaining per-queue qdiscs */
739 for (ntx -= TC_MAX_QUEUE; ntx < dev->num_tx_queues; ntx++) {
740 if (arg->fn(sch, ntx + 1, arg) < 0) {
741 arg->stop = 1;
742 return;
743 }
744 arg->count++;
745 }
746}
747
748static struct netdev_queue *mqprio_select_queue(struct Qdisc *sch,
749 struct tcmsg *tcm)
750{
751 return mqprio_queue_get(sch, TC_H_MIN(tcm->tcm_parent));
752}
753
754static const struct Qdisc_class_ops mqprio_class_ops = {
755 .graft = mqprio_graft,
756 .leaf = mqprio_leaf,
757 .find = mqprio_find,
758 .walk = mqprio_walk,
759 .dump = mqprio_dump_class,
760 .dump_stats = mqprio_dump_class_stats,
761 .select_queue = mqprio_select_queue,
762};
763
764static struct Qdisc_ops mqprio_qdisc_ops __read_mostly = {
765 .cl_ops = &mqprio_class_ops,
766 .id = "mqprio",
767 .priv_size = sizeof(struct mqprio_sched),
768 .init = mqprio_init,
769 .destroy = mqprio_destroy,
770 .attach = mqprio_attach,
771 .change_real_num_tx = mq_change_real_num_tx,
772 .dump = mqprio_dump,
773 .owner = THIS_MODULE,
774};
775MODULE_ALIAS_NET_SCH("mqprio");
776
777static int __init mqprio_module_init(void)
778{
779 return register_qdisc(&mqprio_qdisc_ops);
780}
781
782static void __exit mqprio_module_exit(void)
783{
784 unregister_qdisc(&mqprio_qdisc_ops);
785}
786
787module_init(mqprio_module_init);
788module_exit(mqprio_module_exit);
789
790MODULE_LICENSE("GPL");
791MODULE_DESCRIPTION("Classful multiqueue prio qdisc");