net/sched/sch_mq.c

   1 /*
   2  * net/sched/sch_mq.c           Classful multiqueue dummy scheduler
   3  *
   4  * Copyright (c) 2009 Patrick McHardy <kaber@trash.net>
   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/export.h>
  15 #include <linux/string.h>
  16 #include <linux/errno.h>
  17 #include <linux/skbuff.h>
  18 #include <net/netlink.h>
  19 #include <net/pkt_cls.h>
  20 #include <net/pkt_sched.h>
  21 #include <net/sch_generic.h>
  22
  23 struct mq_sched {
  24         struct Qdisc            **qdiscs;
  25 };
  26
  27 static int mq_offload(struct Qdisc *sch, enum tc_mq_command cmd)
  28 {
  29         struct net_device *dev = qdisc_dev(sch);
  30         struct tc_mq_qopt_offload opt = {
  31                 .command = cmd,
  32                 .handle = sch->handle,
  33         };
  34
  35         if (!tc_can_offload(dev) || !dev->netdev_ops->ndo_setup_tc)
  36                 return -EOPNOTSUPP;
  37
  38         return dev->netdev_ops->ndo_setup_tc(dev, TC_SETUP_QDISC_MQ, &opt);
  39 }
  40
  41 static void mq_offload_stats(struct Qdisc *sch)
  42 {
  43         struct net_device *dev = qdisc_dev(sch);
  44         struct tc_mq_qopt_offload opt = {
  45                 .command = TC_MQ_STATS,
  46                 .handle = sch->handle,
  47                 .stats = {
  48                         .bstats = &sch->bstats,
  49                         .qstats = &sch->qstats,
  50                 },
  51         };
  52
  53         if (tc_can_offload(dev) && dev->netdev_ops->ndo_setup_tc)
  54                 dev->netdev_ops->ndo_setup_tc(dev, TC_SETUP_QDISC_MQ, &opt);
  55 }
  56
  57 static void mq_destroy(struct Qdisc *sch)
  58 {
  59         struct net_device *dev = qdisc_dev(sch);
  60         struct mq_sched *priv = qdisc_priv(sch);
  61         unsigned int ntx;
  62
  63         mq_offload(sch, TC_MQ_DESTROY);
  64
  65         if (!priv->qdiscs)
  66                 return;
  67         for (ntx = 0; ntx < dev->num_tx_queues && priv->qdiscs[ntx]; ntx++)
  68                 qdisc_put(priv->qdiscs[ntx]);
  69         kfree(priv->qdiscs);
  70 }
  71
  72 static int mq_init(struct Qdisc *sch, struct nlattr *opt,
  73                    struct netlink_ext_ack *extack)
  74 {
  75         struct net_device *dev = qdisc_dev(sch);
  76         struct mq_sched *priv = qdisc_priv(sch);
  77         struct netdev_queue *dev_queue;
  78         struct Qdisc *qdisc;
  79         unsigned int ntx;
  80
  81         if (sch->parent != TC_H_ROOT)
  82                 return -EOPNOTSUPP;
  83
  84         if (!netif_is_multiqueue(dev))
  85                 return -EOPNOTSUPP;
  86
  87         /* pre-allocate qdiscs, attachment can't fail */
  88         priv->qdiscs = kcalloc(dev->num_tx_queues, sizeof(priv->qdiscs[0]),
  89                                GFP_KERNEL);
  90         if (!priv->qdiscs)
  91                 return -ENOMEM;
  92
  93         for (ntx = 0; ntx < dev->num_tx_queues; ntx++) {
  94                 dev_queue = netdev_get_tx_queue(dev, ntx);
  95                 qdisc = qdisc_create_dflt(dev_queue, get_default_qdisc_ops(dev, ntx),
  96                                           TC_H_MAKE(TC_H_MAJ(sch->handle),
  97                                                     TC_H_MIN(ntx + 1)),
  98                                           extack);
  99                 if (!qdisc)
 100                         return -ENOMEM;
 101                 priv->qdiscs[ntx] = qdisc;
 102                 qdisc->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
 103         }
 104
 105         sch->flags |= TCQ_F_MQROOT;
 106
 107         mq_offload(sch, TC_MQ_CREATE);
 108         return 0;
 109 }
 110
 111 static void mq_attach(struct Qdisc *sch)
 112 {
 113         struct net_device *dev = qdisc_dev(sch);
 114         struct mq_sched *priv = qdisc_priv(sch);
 115         struct Qdisc *qdisc, *old;
 116         unsigned int ntx;
 117
 118         for (ntx = 0; ntx < dev->num_tx_queues; ntx++) {
 119                 qdisc = priv->qdiscs[ntx];
 120                 old = dev_graft_qdisc(qdisc->dev_queue, qdisc);
 121                 if (old)
 122                         qdisc_put(old);
 123 #ifdef CONFIG_NET_SCHED
 124                 if (ntx < dev->real_num_tx_queues)
 125                         qdisc_hash_add(qdisc, false);
 126 #endif
 127
 128         }
 129         kfree(priv->qdiscs);
 130         priv->qdiscs = NULL;
 131 }
 132
 133 static int mq_dump(struct Qdisc *sch, struct sk_buff *skb)
 134 {
 135         struct net_device *dev = qdisc_dev(sch);
 136         struct Qdisc *qdisc;
 137         unsigned int ntx;
 138         __u32 qlen = 0;
 139
 140         sch->q.qlen = 0;
 141         memset(&sch->bstats, 0, sizeof(sch->bstats));
 142         memset(&sch->qstats, 0, sizeof(sch->qstats));
 143
 144         /* MQ supports lockless qdiscs. However, statistics accounting needs
 145          * to account for all, none, or a mix of locked and unlocked child
 146          * qdiscs. Percpu stats are added to counters in-band and locking
 147          * qdisc totals are added at end.
 148          */
 149         for (ntx = 0; ntx < dev->num_tx_queues; ntx++) {
 150                 qdisc = netdev_get_tx_queue(dev, ntx)->qdisc_sleeping;
 151                 spin_lock_bh(qdisc_lock(qdisc));
 152
 153                 if (qdisc_is_percpu_stats(qdisc)) {
 154                         qlen = qdisc_qlen_sum(qdisc);
 155                         __gnet_stats_copy_basic(NULL, &sch->bstats,
 156                                                 qdisc->cpu_bstats,
 157                                                 &qdisc->bstats);
 158                         __gnet_stats_copy_queue(&sch->qstats,
 159                                                 qdisc->cpu_qstats,
 160                                                 &qdisc->qstats, qlen);
 161                 } else {
 162                         sch->q.qlen             += qdisc->q.qlen;
 163                         sch->bstats.bytes       += qdisc->bstats.bytes;
 164                         sch->bstats.packets     += qdisc->bstats.packets;
 165                         sch->qstats.qlen        += qdisc->qstats.qlen;
 166                         sch->qstats.backlog     += qdisc->qstats.backlog;
 167                         sch->qstats.drops       += qdisc->qstats.drops;
 168                         sch->qstats.requeues    += qdisc->qstats.requeues;
 169                         sch->qstats.overlimits  += qdisc->qstats.overlimits;
 170                 }
 171
 172                 spin_unlock_bh(qdisc_lock(qdisc));
 173         }
 174         mq_offload_stats(sch);
 175
 176         return 0;
 177 }
 178
 179 static struct netdev_queue *mq_queue_get(struct Qdisc *sch, unsigned long cl)
 180 {
 181         struct net_device *dev = qdisc_dev(sch);
 182         unsigned long ntx = cl - 1;
 183
 184         if (ntx >= dev->num_tx_queues)
 185                 return NULL;
 186         return netdev_get_tx_queue(dev, ntx);
 187 }
 188
 189 static struct netdev_queue *mq_select_queue(struct Qdisc *sch,
 190                                             struct tcmsg *tcm)
 191 {
 192         return mq_queue_get(sch, TC_H_MIN(tcm->tcm_parent));
 193 }
 194
 195 static int mq_graft(struct Qdisc *sch, unsigned long cl, struct Qdisc *new,
 196                     struct Qdisc **old, struct netlink_ext_ack *extack)
 197 {
 198         struct netdev_queue *dev_queue = mq_queue_get(sch, cl);
 199         struct net_device *dev = qdisc_dev(sch);
 200
 201         if (dev->flags & IFF_UP)
 202                 dev_deactivate(dev);
 203
 204         *old = dev_graft_qdisc(dev_queue, new);
 205         if (new)
 206                 new->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
 207         if (dev->flags & IFF_UP)
 208                 dev_activate(dev);
 209         return 0;
 210 }
 211
 212 static struct Qdisc *mq_leaf(struct Qdisc *sch, unsigned long cl)
 213 {
 214         struct netdev_queue *dev_queue = mq_queue_get(sch, cl);
 215
 216         return dev_queue->qdisc_sleeping;
 217 }
 218
 219 static unsigned long mq_find(struct Qdisc *sch, u32 classid)
 220 {
 221         unsigned int ntx = TC_H_MIN(classid);
 222
 223         if (!mq_queue_get(sch, ntx))
 224                 return 0;
 225         return ntx;
 226 }
 227
 228 static int mq_dump_class(struct Qdisc *sch, unsigned long cl,
 229                          struct sk_buff *skb, struct tcmsg *tcm)
 230 {
 231         struct netdev_queue *dev_queue = mq_queue_get(sch, cl);
 232
 233         tcm->tcm_parent = TC_H_ROOT;
 234         tcm->tcm_handle |= TC_H_MIN(cl);
 235         tcm->tcm_info = dev_queue->qdisc_sleeping->handle;
 236         return 0;
 237 }
 238
 239 static int mq_dump_class_stats(struct Qdisc *sch, unsigned long cl,
 240                                struct gnet_dump *d)
 241 {
 242         struct netdev_queue *dev_queue = mq_queue_get(sch, cl);
 243
 244         sch = dev_queue->qdisc_sleeping;
 245         if (gnet_stats_copy_basic(&sch->running, d, NULL, &sch->bstats) < 0 ||
 246             gnet_stats_copy_queue(d, NULL, &sch->qstats, sch->q.qlen) < 0)
 247                 return -1;
 248         return 0;
 249 }
 250
 251 static void mq_walk(struct Qdisc *sch, struct qdisc_walker *arg)
 252 {
 253         struct net_device *dev = qdisc_dev(sch);
 254         unsigned int ntx;
 255
 256         if (arg->stop)
 257                 return;
 258
 259         arg->count = arg->skip;
 260         for (ntx = arg->skip; ntx < dev->num_tx_queues; ntx++) {
 261                 if (arg->fn(sch, ntx + 1, arg) < 0) {
 262                         arg->stop = 1;
 263                         break;
 264                 }
 265                 arg->count++;
 266         }
 267 }
 268
 269 static const struct Qdisc_class_ops mq_class_ops = {
 270         .select_queue   = mq_select_queue,
 271         .graft          = mq_graft,
 272         .leaf           = mq_leaf,
 273         .find           = mq_find,
 274         .walk           = mq_walk,
 275         .dump           = mq_dump_class,
 276         .dump_stats     = mq_dump_class_stats,
 277 };
 278
 279 struct Qdisc_ops mq_qdisc_ops __read_mostly = {
 280         .cl_ops         = &mq_class_ops,
 281         .id             = "mq",
 282         .priv_size      = sizeof(struct mq_sched),
 283         .init           = mq_init,
 284         .destroy        = mq_destroy,
 285         .attach         = mq_attach,
 286         .dump           = mq_dump,
 287         .owner          = THIS_MODULE,
 288 };