net/sched/sch_mqprio.c

   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
  22 struct mqprio_sched {
  23         struct Qdisc            **qdiscs;
  24         int hw_owned;
  25 };
  26
  27 static void mqprio_destroy(struct Qdisc *sch)
  28 {
  29         struct net_device *dev = qdisc_dev(sch);
  30         struct mqprio_sched *priv = qdisc_priv(sch);
  31         unsigned int ntx;
  32
  33         if (priv->qdiscs) {
  34                 for (ntx = 0;
  35                      ntx < dev->num_tx_queues && priv->qdiscs[ntx];
  36                      ntx++)
  37                         qdisc_destroy(priv->qdiscs[ntx]);
  38                 kfree(priv->qdiscs);
  39         }
  40
  41         if (priv->hw_owned && dev->netdev_ops->ndo_setup_tc)
  42                 dev->netdev_ops->ndo_setup_tc(dev, 0);
  43         else
  44                 netdev_set_num_tc(dev, 0);
  45 }
  46
  47 static int mqprio_parse_opt(struct net_device *dev, struct tc_mqprio_qopt *qopt)
  48 {
  49         int i, j;
  50
  51         /* Verify num_tc is not out of max range */
  52         if (qopt->num_tc > TC_MAX_QUEUE)
  53                 return -EINVAL;
  54
  55         /* Verify priority mapping uses valid tcs */
  56         for (i = 0; i < TC_BITMASK + 1; i++) {
  57                 if (qopt->prio_tc_map[i] >= qopt->num_tc)
  58                         return -EINVAL;
  59         }
  60
  61         /* net_device does not support requested operation */
  62         if (qopt->hw && !dev->netdev_ops->ndo_setup_tc)
  63                 return -EINVAL;
  64
  65         /* if hw owned qcount and qoffset are taken from LLD so
  66          * no reason to verify them here
  67          */
  68         if (qopt->hw)
  69                 return 0;
  70
  71         for (i = 0; i < qopt->num_tc; i++) {
  72                 unsigned int last = qopt->offset[i] + qopt->count[i];
  73
  74                 /* Verify the queue count is in tx range being equal to the
  75                  * real_num_tx_queues indicates the last queue is in use.
  76                  */
  77                 if (qopt->offset[i] >= dev->real_num_tx_queues ||
  78                     !qopt->count[i] ||
  79                     last > dev->real_num_tx_queues)
  80                         return -EINVAL;
  81
  82                 /* Verify that the offset and counts do not overlap */
  83                 for (j = i + 1; j < qopt->num_tc; j++) {
  84                         if (last > qopt->offset[j])
  85                                 return -EINVAL;
  86                 }
  87         }
  88
  89         return 0;
  90 }
  91
  92 static int mqprio_init(struct Qdisc *sch, struct nlattr *opt)
  93 {
  94         struct net_device *dev = qdisc_dev(sch);
  95         struct mqprio_sched *priv = qdisc_priv(sch);
  96         struct netdev_queue *dev_queue;
  97         struct Qdisc *qdisc;
  98         int i, err = -EOPNOTSUPP;
  99         struct tc_mqprio_qopt *qopt = NULL;
 100
 101         BUILD_BUG_ON(TC_MAX_QUEUE != TC_QOPT_MAX_QUEUE);
 102         BUILD_BUG_ON(TC_BITMASK != TC_QOPT_BITMASK);
 103
 104         if (sch->parent != TC_H_ROOT)
 105                 return -EOPNOTSUPP;
 106
 107         if (!netif_is_multiqueue(dev))
 108                 return -EOPNOTSUPP;
 109
 110         if (!opt || nla_len(opt) < sizeof(*qopt))
 111                 return -EINVAL;
 112
 113         qopt = nla_data(opt);
 114         if (mqprio_parse_opt(dev, qopt))
 115                 return -EINVAL;
 116
 117         /* pre-allocate qdisc, attachment can't fail */
 118         priv->qdiscs = kcalloc(dev->num_tx_queues, sizeof(priv->qdiscs[0]),
 119                                GFP_KERNEL);
 120         if (!priv->qdiscs)
 121                 return -ENOMEM;
 122
 123         for (i = 0; i < dev->num_tx_queues; i++) {
 124                 dev_queue = netdev_get_tx_queue(dev, i);
 125                 qdisc = qdisc_create_dflt(dev_queue, default_qdisc_ops,
 126                                           TC_H_MAKE(TC_H_MAJ(sch->handle),
 127                                                     TC_H_MIN(i + 1)));
 128                 if (!qdisc)
 129                         return -ENOMEM;
 130
 131                 priv->qdiscs[i] = qdisc;
 132                 qdisc->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
 133         }
 134
 135         /* If the mqprio options indicate that hardware should own
 136          * the queue mapping then run ndo_setup_tc otherwise use the
 137          * supplied and verified mapping
 138          */
 139         if (qopt->hw) {
 140                 priv->hw_owned = 1;
 141                 err = dev->netdev_ops->ndo_setup_tc(dev, qopt->num_tc);
 142                 if (err)
 143                         return err;
 144         } else {
 145                 netdev_set_num_tc(dev, qopt->num_tc);
 146                 for (i = 0; i < qopt->num_tc; i++)
 147                         netdev_set_tc_queue(dev, i,
 148                                             qopt->count[i], qopt->offset[i]);
 149         }
 150
 151         /* Always use supplied priority mappings */
 152         for (i = 0; i < TC_BITMASK + 1; i++)
 153                 netdev_set_prio_tc_map(dev, i, qopt->prio_tc_map[i]);
 154
 155         sch->flags |= TCQ_F_MQROOT;
 156         return 0;
 157 }
 158
 159 static void mqprio_attach(struct Qdisc *sch)
 160 {
 161         struct net_device *dev = qdisc_dev(sch);
 162         struct mqprio_sched *priv = qdisc_priv(sch);
 163         struct Qdisc *qdisc, *old;
 164         unsigned int ntx;
 165
 166         /* Attach underlying qdisc */
 167         for (ntx = 0; ntx < dev->num_tx_queues; ntx++) {
 168                 qdisc = priv->qdiscs[ntx];
 169                 old = dev_graft_qdisc(qdisc->dev_queue, qdisc);
 170                 if (old)
 171                         qdisc_destroy(old);
 172                 if (ntx < dev->real_num_tx_queues)
 173                         qdisc_list_add(qdisc);
 174         }
 175         kfree(priv->qdiscs);
 176         priv->qdiscs = NULL;
 177 }
 178
 179 static struct netdev_queue *mqprio_queue_get(struct Qdisc *sch,
 180                                              unsigned long cl)
 181 {
 182         struct net_device *dev = qdisc_dev(sch);
 183         unsigned long ntx = cl - 1 - netdev_get_num_tc(dev);
 184
 185         if (ntx >= dev->num_tx_queues)
 186                 return NULL;
 187         return netdev_get_tx_queue(dev, ntx);
 188 }
 189
 190 static int mqprio_graft(struct Qdisc *sch, unsigned long cl, struct Qdisc *new,
 191                     struct Qdisc **old)
 192 {
 193         struct net_device *dev = qdisc_dev(sch);
 194         struct netdev_queue *dev_queue = mqprio_queue_get(sch, cl);
 195
 196         if (!dev_queue)
 197                 return -EINVAL;
 198
 199         if (dev->flags & IFF_UP)
 200                 dev_deactivate(dev);
 201
 202         *old = dev_graft_qdisc(dev_queue, new);
 203
 204         if (new)
 205                 new->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
 206
 207         if (dev->flags & IFF_UP)
 208                 dev_activate(dev);
 209
 210         return 0;
 211 }
 212
 213 static int mqprio_dump(struct Qdisc *sch, struct sk_buff *skb)
 214 {
 215         struct net_device *dev = qdisc_dev(sch);
 216         struct mqprio_sched *priv = qdisc_priv(sch);
 217         unsigned char *b = skb_tail_pointer(skb);
 218         struct tc_mqprio_qopt opt = { 0 };
 219         struct Qdisc *qdisc;
 220         unsigned int i;
 221
 222         sch->q.qlen = 0;
 223         memset(&sch->bstats, 0, sizeof(sch->bstats));
 224         memset(&sch->qstats, 0, sizeof(sch->qstats));
 225
 226         for (i = 0; i < dev->num_tx_queues; i++) {
 227                 qdisc = rtnl_dereference(netdev_get_tx_queue(dev, i)->qdisc);
 228                 spin_lock_bh(qdisc_lock(qdisc));
 229                 sch->q.qlen             += qdisc->q.qlen;
 230                 sch->bstats.bytes       += qdisc->bstats.bytes;
 231                 sch->bstats.packets     += qdisc->bstats.packets;
 232                 sch->qstats.backlog     += qdisc->qstats.backlog;
 233                 sch->qstats.drops       += qdisc->qstats.drops;
 234                 sch->qstats.requeues    += qdisc->qstats.requeues;
 235                 sch->qstats.overlimits  += qdisc->qstats.overlimits;
 236                 spin_unlock_bh(qdisc_lock(qdisc));
 237         }
 238
 239         opt.num_tc = netdev_get_num_tc(dev);
 240         memcpy(opt.prio_tc_map, dev->prio_tc_map, sizeof(opt.prio_tc_map));
 241         opt.hw = priv->hw_owned;
 242
 243         for (i = 0; i < netdev_get_num_tc(dev); i++) {
 244                 opt.count[i] = dev->tc_to_txq[i].count;
 245                 opt.offset[i] = dev->tc_to_txq[i].offset;
 246         }
 247
 248         if (nla_put(skb, TCA_OPTIONS, sizeof(opt), &opt))
 249                 goto nla_put_failure;
 250
 251         return skb->len;
 252 nla_put_failure:
 253         nlmsg_trim(skb, b);
 254         return -1;
 255 }
 256
 257 static struct Qdisc *mqprio_leaf(struct Qdisc *sch, unsigned long cl)
 258 {
 259         struct netdev_queue *dev_queue = mqprio_queue_get(sch, cl);
 260
 261         if (!dev_queue)
 262                 return NULL;
 263
 264         return dev_queue->qdisc_sleeping;
 265 }
 266
 267 static unsigned long mqprio_get(struct Qdisc *sch, u32 classid)
 268 {
 269         struct net_device *dev = qdisc_dev(sch);
 270         unsigned int ntx = TC_H_MIN(classid);
 271
 272         if (ntx > dev->num_tx_queues + netdev_get_num_tc(dev))
 273                 return 0;
 274         return ntx;
 275 }
 276
 277 static void mqprio_put(struct Qdisc *sch, unsigned long cl)
 278 {
 279 }
 280
 281 static int mqprio_dump_class(struct Qdisc *sch, unsigned long cl,
 282                          struct sk_buff *skb, struct tcmsg *tcm)
 283 {
 284         struct net_device *dev = qdisc_dev(sch);
 285
 286         if (cl <= netdev_get_num_tc(dev)) {
 287                 tcm->tcm_parent = TC_H_ROOT;
 288                 tcm->tcm_info = 0;
 289         } else {
 290                 int i;
 291                 struct netdev_queue *dev_queue;
 292
 293                 dev_queue = mqprio_queue_get(sch, cl);
 294                 tcm->tcm_parent = 0;
 295                 for (i = 0; i < netdev_get_num_tc(dev); i++) {
 296                         struct netdev_tc_txq tc = dev->tc_to_txq[i];
 297                         int q_idx = cl - netdev_get_num_tc(dev);
 298
 299                         if (q_idx > tc.offset &&
 300                             q_idx <= tc.offset + tc.count) {
 301                                 tcm->tcm_parent =
 302                                         TC_H_MAKE(TC_H_MAJ(sch->handle),
 303                                                   TC_H_MIN(i + 1));
 304                                 break;
 305                         }
 306                 }
 307                 tcm->tcm_info = dev_queue->qdisc_sleeping->handle;
 308         }
 309         tcm->tcm_handle |= TC_H_MIN(cl);
 310         return 0;
 311 }
 312
 313 static int mqprio_dump_class_stats(struct Qdisc *sch, unsigned long cl,
 314                                    struct gnet_dump *d)
 315         __releases(d->lock)
 316         __acquires(d->lock)
 317 {
 318         struct net_device *dev = qdisc_dev(sch);
 319
 320         if (cl <= netdev_get_num_tc(dev)) {
 321                 int i;
 322                 __u32 qlen = 0;
 323                 struct Qdisc *qdisc;
 324                 struct gnet_stats_queue qstats = {0};
 325                 struct gnet_stats_basic_packed bstats = {0};
 326                 struct netdev_tc_txq tc = dev->tc_to_txq[cl - 1];
 327
 328                 /* Drop lock here it will be reclaimed before touching
 329                  * statistics this is required because the d->lock we
 330                  * hold here is the look on dev_queue->qdisc_sleeping
 331                  * also acquired below.
 332                  */
 333                 spin_unlock_bh(d->lock);
 334
 335                 for (i = tc.offset; i < tc.offset + tc.count; i++) {
 336                         struct netdev_queue *q = netdev_get_tx_queue(dev, i);
 337
 338                         qdisc = rtnl_dereference(q->qdisc);
 339                         spin_lock_bh(qdisc_lock(qdisc));
 340                         qlen              += qdisc->q.qlen;
 341                         bstats.bytes      += qdisc->bstats.bytes;
 342                         bstats.packets    += qdisc->bstats.packets;
 343                         qstats.backlog    += qdisc->qstats.backlog;
 344                         qstats.drops      += qdisc->qstats.drops;
 345                         qstats.requeues   += qdisc->qstats.requeues;
 346                         qstats.overlimits += qdisc->qstats.overlimits;
 347                         spin_unlock_bh(qdisc_lock(qdisc));
 348                 }
 349                 /* Reclaim root sleeping lock before completing stats */
 350                 spin_lock_bh(d->lock);
 351                 if (gnet_stats_copy_basic(d, NULL, &bstats) < 0 ||
 352                     gnet_stats_copy_queue(d, NULL, &qstats, qlen) < 0)
 353                         return -1;
 354         } else {
 355                 struct netdev_queue *dev_queue = mqprio_queue_get(sch, cl);
 356
 357                 sch = dev_queue->qdisc_sleeping;
 358                 if (gnet_stats_copy_basic(d, NULL, &sch->bstats) < 0 ||
 359                     gnet_stats_copy_queue(d, NULL,
 360                                           &sch->qstats, sch->q.qlen) < 0)
 361                         return -1;
 362         }
 363         return 0;
 364 }
 365
 366 static void mqprio_walk(struct Qdisc *sch, struct qdisc_walker *arg)
 367 {
 368         struct net_device *dev = qdisc_dev(sch);
 369         unsigned long ntx;
 370
 371         if (arg->stop)
 372                 return;
 373
 374         /* Walk hierarchy with a virtual class per tc */
 375         arg->count = arg->skip;
 376         for (ntx = arg->skip;
 377              ntx < dev->num_tx_queues + netdev_get_num_tc(dev);
 378              ntx++) {
 379                 if (arg->fn(sch, ntx + 1, arg) < 0) {
 380                         arg->stop = 1;
 381                         break;
 382                 }
 383                 arg->count++;
 384         }
 385 }
 386
 387 static const struct Qdisc_class_ops mqprio_class_ops = {
 388         .graft          = mqprio_graft,
 389         .leaf           = mqprio_leaf,
 390         .get            = mqprio_get,
 391         .put            = mqprio_put,
 392         .walk           = mqprio_walk,
 393         .dump           = mqprio_dump_class,
 394         .dump_stats     = mqprio_dump_class_stats,
 395 };
 396
 397 static struct Qdisc_ops mqprio_qdisc_ops __read_mostly = {
 398         .cl_ops         = &mqprio_class_ops,
 399         .id             = "mqprio",
 400         .priv_size      = sizeof(struct mqprio_sched),
 401         .init           = mqprio_init,
 402         .destroy        = mqprio_destroy,
 403         .attach         = mqprio_attach,
 404         .dump           = mqprio_dump,
 405         .owner          = THIS_MODULE,
 406 };
 407
 408 static int __init mqprio_module_init(void)
 409 {
 410         return register_qdisc(&mqprio_qdisc_ops);
 411 }
 412
 413 static void __exit mqprio_module_exit(void)
 414 {
 415         unregister_qdisc(&mqprio_qdisc_ops);
 416 }
 417
 418 module_init(mqprio_module_init);
 419 module_exit(mqprio_module_exit);
 420
 421 MODULE_LICENSE("GPL");