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 <net/netlink.h>
  18 #include <net/pkt_sched.h>
  19 #include <net/sch_generic.h>
  20
  21 struct mqprio_sched {
  22         struct Qdisc            **qdiscs;
  23         int hw_owned;
  24 };
  25
  26 static void mqprio_destroy(struct Qdisc *sch)
  27 {
  28         struct net_device *dev = qdisc_dev(sch);
  29         struct mqprio_sched *priv = qdisc_priv(sch);
  30         unsigned int ntx;
  31
  32         if (priv->qdiscs) {
  33                 for (ntx = 0;
  34                      ntx < dev->num_tx_queues && priv->qdiscs[ntx];
  35                      ntx++)
  36                         qdisc_destroy(priv->qdiscs[ntx]);
  37                 kfree(priv->qdiscs);
  38         }
  39
  40         if (priv->hw_owned && dev->netdev_ops->ndo_setup_tc)
  41                 dev->netdev_ops->ndo_setup_tc(dev, 0);
  42         else
  43                 netdev_set_num_tc(dev, 0);
  44 }
  45
  46 static int mqprio_parse_opt(struct net_device *dev, struct tc_mqprio_qopt *qopt)
  47 {
  48         int i, j;
  49
  50         /* Verify num_tc is not out of max range */
  51         if (qopt->num_tc > TC_MAX_QUEUE)
  52                 return -EINVAL;
  53
  54         /* Verify priority mapping uses valid tcs */
  55         for (i = 0; i < TC_BITMASK + 1; i++) {
  56                 if (qopt->prio_tc_map[i] >= qopt->num_tc)
  57                         return -EINVAL;
  58         }
  59
  60         /* net_device does not support requested operation */
  61         if (qopt->hw && !dev->netdev_ops->ndo_setup_tc)
  62                 return -EINVAL;
  63
  64         /* if hw owned qcount and qoffset are taken from LLD so
  65          * no reason to verify them here
  66          */
  67         if (qopt->hw)
  68                 return 0;
  69
  70         for (i = 0; i < qopt->num_tc; i++) {
  71                 unsigned int last = qopt->offset[i] + qopt->count[i];
  72
  73                 /* Verify the queue count is in tx range being equal to the
  74                  * real_num_tx_queues indicates the last queue is in use.
  75                  */
  76                 if (qopt->offset[i] >= dev->real_num_tx_queues ||
  77                     !qopt->count[i] ||
  78                     last > dev->real_num_tx_queues)
  79                         return -EINVAL;
  80
  81                 /* Verify that the offset and counts do not overlap */
  82                 for (j = i + 1; j < qopt->num_tc; j++) {
  83                         if (last > qopt->offset[j])
  84                                 return -EINVAL;
  85                 }
  86         }
  87
  88         return 0;
  89 }
  90
  91 static int mqprio_init(struct Qdisc *sch, struct nlattr *opt)
  92 {
  93         struct net_device *dev = qdisc_dev(sch);
  94         struct mqprio_sched *priv = qdisc_priv(sch);
  95         struct netdev_queue *dev_queue;
  96         struct Qdisc *qdisc;
  97         int i, err = -EOPNOTSUPP;
  98         struct tc_mqprio_qopt *qopt = NULL;
  99
 100         BUILD_BUG_ON(TC_MAX_QUEUE != TC_QOPT_MAX_QUEUE);
 101         BUILD_BUG_ON(TC_BITMASK != TC_QOPT_BITMASK);
 102
 103         if (sch->parent != TC_H_ROOT)
 104                 return -EOPNOTSUPP;
 105
 106         if (!netif_is_multiqueue(dev))
 107                 return -EOPNOTSUPP;
 108
 109         if (!opt || nla_len(opt) < sizeof(*qopt))
 110                 return -EINVAL;
 111
 112         qopt = nla_data(opt);
 113         if (mqprio_parse_opt(dev, qopt))
 114                 return -EINVAL;
 115
 116         /* pre-allocate qdisc, attachment can't fail */
 117         priv->qdiscs = kcalloc(dev->num_tx_queues, sizeof(priv->qdiscs[0]),
 118                                GFP_KERNEL);
 119         if (priv->qdiscs == NULL) {
 120                 err = -ENOMEM;
 121                 goto err;
 122         }
 123
 124         for (i = 0; i < dev->num_tx_queues; i++) {
 125                 dev_queue = netdev_get_tx_queue(dev, i);
 126                 qdisc = qdisc_create_dflt(dev_queue, &pfifo_fast_ops,
 127                                           TC_H_MAKE(TC_H_MAJ(sch->handle),
 128                                                     TC_H_MIN(i + 1)));
 129                 if (qdisc == NULL) {
 130                         err = -ENOMEM;
 131                         goto err;
 132                 }
 133                 priv->qdiscs[i] = qdisc;
 134         }
 135
 136         /* If the mqprio options indicate that hardware should own
 137          * the queue mapping then run ndo_setup_tc otherwise use the
 138          * supplied and verified mapping
 139          */
 140         if (qopt->hw) {
 141                 priv->hw_owned = 1;
 142                 err = dev->netdev_ops->ndo_setup_tc(dev, qopt->num_tc);
 143                 if (err)
 144                         goto err;
 145         } else {
 146                 netdev_set_num_tc(dev, qopt->num_tc);
 147                 for (i = 0; i < qopt->num_tc; i++)
 148                         netdev_set_tc_queue(dev, i,
 149                                             qopt->count[i], qopt->offset[i]);
 150         }
 151
 152         /* Always use supplied priority mappings */
 153         for (i = 0; i < TC_BITMASK + 1; i++)
 154                 netdev_set_prio_tc_map(dev, i, qopt->prio_tc_map[i]);
 155
 156         sch->flags |= TCQ_F_MQROOT;
 157         return 0;
 158
 159 err:
 160         mqprio_destroy(sch);
 161         return err;
 162 }
 163
 164 static void mqprio_attach(struct Qdisc *sch)
 165 {
 166         struct net_device *dev = qdisc_dev(sch);
 167         struct mqprio_sched *priv = qdisc_priv(sch);
 168         struct Qdisc *qdisc;
 169         unsigned int ntx;
 170
 171         /* Attach underlying qdisc */
 172         for (ntx = 0; ntx < dev->num_tx_queues; ntx++) {
 173                 qdisc = priv->qdiscs[ntx];
 174                 qdisc = dev_graft_qdisc(qdisc->dev_queue, qdisc);
 175                 if (qdisc)
 176                         qdisc_destroy(qdisc);
 177         }
 178         kfree(priv->qdiscs);
 179         priv->qdiscs = NULL;
 180 }
 181
 182 static struct netdev_queue *mqprio_queue_get(struct Qdisc *sch,
 183                                              unsigned long cl)
 184 {
 185         struct net_device *dev = qdisc_dev(sch);
 186         unsigned long ntx = cl - 1 - netdev_get_num_tc(dev);
 187
 188         if (ntx >= dev->num_tx_queues)
 189                 return NULL;
 190         return netdev_get_tx_queue(dev, ntx);
 191 }
 192
 193 static int mqprio_graft(struct Qdisc *sch, unsigned long cl, struct Qdisc *new,
 194                     struct Qdisc **old)
 195 {
 196         struct net_device *dev = qdisc_dev(sch);
 197         struct netdev_queue *dev_queue = mqprio_queue_get(sch, cl);
 198
 199         if (!dev_queue)
 200                 return -EINVAL;
 201
 202         if (dev->flags & IFF_UP)
 203                 dev_deactivate(dev);
 204
 205         *old = dev_graft_qdisc(dev_queue, new);
 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 = 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.qlen        += qdisc->qstats.qlen;
 233                 sch->qstats.backlog     += qdisc->qstats.backlog;
 234                 sch->qstats.drops       += qdisc->qstats.drops;
 235                 sch->qstats.requeues    += qdisc->qstats.requeues;
 236                 sch->qstats.overlimits  += qdisc->qstats.overlimits;
 237                 spin_unlock_bh(qdisc_lock(qdisc));
 238         }
 239
 240         opt.num_tc = netdev_get_num_tc(dev);
 241         memcpy(opt.prio_tc_map, dev->prio_tc_map, sizeof(opt.prio_tc_map));
 242         opt.hw = priv->hw_owned;
 243
 244         for (i = 0; i < netdev_get_num_tc(dev); i++) {
 245                 opt.count[i] = dev->tc_to_txq[i].count;
 246                 opt.offset[i] = dev->tc_to_txq[i].offset;
 247         }
 248
 249         NLA_PUT(skb, TCA_OPTIONS, sizeof(opt), &opt);
 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                 struct Qdisc *qdisc;
 323                 struct gnet_stats_queue qstats = {0};
 324                 struct gnet_stats_basic_packed bstats = {0};
 325                 struct netdev_tc_txq tc = dev->tc_to_txq[cl - 1];
 326
 327                 /* Drop lock here it will be reclaimed before touching
 328                  * statistics this is required because the d->lock we
 329                  * hold here is the look on dev_queue->qdisc_sleeping
 330                  * also acquired below.
 331                  */
 332                 spin_unlock_bh(d->lock);
 333
 334                 for (i = tc.offset; i < tc.offset + tc.count; i++) {
 335                         qdisc = netdev_get_tx_queue(dev, i)->qdisc;
 336                         spin_lock_bh(qdisc_lock(qdisc));
 337                         bstats.bytes      += qdisc->bstats.bytes;
 338                         bstats.packets    += qdisc->bstats.packets;
 339                         qstats.qlen       += qdisc->qstats.qlen;
 340                         qstats.backlog    += qdisc->qstats.backlog;
 341                         qstats.drops      += qdisc->qstats.drops;
 342                         qstats.requeues   += qdisc->qstats.requeues;
 343                         qstats.overlimits += qdisc->qstats.overlimits;
 344                         spin_unlock_bh(qdisc_lock(qdisc));
 345                 }
 346                 /* Reclaim root sleeping lock before completing stats */
 347                 spin_lock_bh(d->lock);
 348                 if (gnet_stats_copy_basic(d, &bstats) < 0 ||
 349                     gnet_stats_copy_queue(d, &qstats) < 0)
 350                         return -1;
 351         } else {
 352                 struct netdev_queue *dev_queue = mqprio_queue_get(sch, cl);
 353
 354                 sch = dev_queue->qdisc_sleeping;
 355                 sch->qstats.qlen = sch->q.qlen;
 356                 if (gnet_stats_copy_basic(d, &sch->bstats) < 0 ||
 357                     gnet_stats_copy_queue(d, &sch->qstats) < 0)
 358                         return -1;
 359         }
 360         return 0;
 361 }
 362
 363 static void mqprio_walk(struct Qdisc *sch, struct qdisc_walker *arg)
 364 {
 365         struct net_device *dev = qdisc_dev(sch);
 366         unsigned long ntx;
 367
 368         if (arg->stop)
 369                 return;
 370
 371         /* Walk hierarchy with a virtual class per tc */
 372         arg->count = arg->skip;
 373         for (ntx = arg->skip;
 374              ntx < dev->num_tx_queues + netdev_get_num_tc(dev);
 375              ntx++) {
 376                 if (arg->fn(sch, ntx + 1, arg) < 0) {
 377                         arg->stop = 1;
 378                         break;
 379                 }
 380                 arg->count++;
 381         }
 382 }
 383
 384 static const struct Qdisc_class_ops mqprio_class_ops = {
 385         .graft          = mqprio_graft,
 386         .leaf           = mqprio_leaf,
 387         .get            = mqprio_get,
 388         .put            = mqprio_put,
 389         .walk           = mqprio_walk,
 390         .dump           = mqprio_dump_class,
 391         .dump_stats     = mqprio_dump_class_stats,
 392 };
 393
 394 static struct Qdisc_ops mqprio_qdisc_ops __read_mostly = {
 395         .cl_ops         = &mqprio_class_ops,
 396         .id             = "mqprio",
 397         .priv_size      = sizeof(struct mqprio_sched),
 398         .init           = mqprio_init,
 399         .destroy        = mqprio_destroy,
 400         .attach         = mqprio_attach,
 401         .dump           = mqprio_dump,
 402         .owner          = THIS_MODULE,
 403 };
 404
 405 static int __init mqprio_module_init(void)
 406 {
 407         return register_qdisc(&mqprio_qdisc_ops);
 408 }
 409
 410 static void __exit mqprio_module_exit(void)
 411 {
 412         unregister_qdisc(&mqprio_qdisc_ops);
 413 }
 414
 415 module_init(mqprio_module_init);
 416 module_exit(mqprio_module_exit);
 417
 418 MODULE_LICENSE("GPL");