net/sched/sch_mqprio.c

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