net/sched/act_police.c

   1 /*
   2  * net/sched/act_police.c       Input police filter
   3  *
   4  *              This program is free software; you can redistribute it and/or
   5  *              modify it under the terms of the GNU General Public License
   6  *              as published by the Free Software Foundation; either version
   7  *              2 of the License, or (at your option) any later version.
   8  *
   9  * Authors:     Alexey Kuznetsov, <kuznet@ms2.inr.ac.ru>
  10  *              J Hadi Salim (action changes)
  11  */
  12
  13 #include <linux/module.h>
  14 #include <linux/types.h>
  15 #include <linux/kernel.h>
  16 #include <linux/string.h>
  17 #include <linux/errno.h>
  18 #include <linux/skbuff.h>
  19 #include <linux/rtnetlink.h>
  20 #include <linux/init.h>
  21 #include <linux/slab.h>
  22 #include <net/act_api.h>
  23 #include <net/netlink.h>
  24 #include <net/pkt_cls.h>
  25 #include <net/tc_act/tc_police.h>
  26
  27 /* Each policer is serialized by its individual spinlock */
  28
  29 static unsigned int police_net_id;
  30 static struct tc_action_ops act_police_ops;
  31
  32 static int tcf_police_walker(struct net *net, struct sk_buff *skb,
  33                                  struct netlink_callback *cb, int type,
  34                                  const struct tc_action_ops *ops,
  35                                  struct netlink_ext_ack *extack)
  36 {
  37         struct tc_action_net *tn = net_generic(net, police_net_id);
  38
  39         return tcf_generic_walker(tn, skb, cb, type, ops, extack);
  40 }
  41
  42 static const struct nla_policy police_policy[TCA_POLICE_MAX + 1] = {
  43         [TCA_POLICE_RATE]       = { .len = TC_RTAB_SIZE },
  44         [TCA_POLICE_PEAKRATE]   = { .len = TC_RTAB_SIZE },
  45         [TCA_POLICE_AVRATE]     = { .type = NLA_U32 },
  46         [TCA_POLICE_RESULT]     = { .type = NLA_U32 },
  47 };
  48
  49 static int tcf_police_init(struct net *net, struct nlattr *nla,
  50                                struct nlattr *est, struct tc_action **a,
  51                                int ovr, int bind, bool rtnl_held,
  52                                struct tcf_proto *tp,
  53                                struct netlink_ext_ack *extack)
  54 {
  55         int ret = 0, tcfp_result = TC_ACT_OK, err, size;
  56         struct nlattr *tb[TCA_POLICE_MAX + 1];
  57         struct tcf_chain *goto_ch = NULL;
  58         struct tc_police *parm;
  59         struct tcf_police *police;
  60         struct qdisc_rate_table *R_tab = NULL, *P_tab = NULL;
  61         struct tc_action_net *tn = net_generic(net, police_net_id);
  62         struct tcf_police_params *new;
  63         bool exists = false;
  64
  65         if (nla == NULL)
  66                 return -EINVAL;
  67
  68         err = nla_parse_nested_deprecated(tb, TCA_POLICE_MAX, nla,
  69                                           police_policy, NULL);
  70         if (err < 0)
  71                 return err;
  72
  73         if (tb[TCA_POLICE_TBF] == NULL)
  74                 return -EINVAL;
  75         size = nla_len(tb[TCA_POLICE_TBF]);
  76         if (size != sizeof(*parm) && size != sizeof(struct tc_police_compat))
  77                 return -EINVAL;
  78
  79         parm = nla_data(tb[TCA_POLICE_TBF]);
  80         err = tcf_idr_check_alloc(tn, &parm->index, a, bind);
  81         if (err < 0)
  82                 return err;
  83         exists = err;
  84         if (exists && bind)
  85                 return 0;
  86
  87         if (!exists) {
  88                 ret = tcf_idr_create(tn, parm->index, NULL, a,
  89                                      &act_police_ops, bind, true);
  90                 if (ret) {
  91                         tcf_idr_cleanup(tn, parm->index);
  92                         return ret;
  93                 }
  94                 ret = ACT_P_CREATED;
  95                 spin_lock_init(&(to_police(*a)->tcfp_lock));
  96         } else if (!ovr) {
  97                 tcf_idr_release(*a, bind);
  98                 return -EEXIST;
  99         }
 100         err = tcf_action_check_ctrlact(parm->action, tp, &goto_ch, extack);
 101         if (err < 0)
 102                 goto release_idr;
 103
 104         police = to_police(*a);
 105         if (parm->rate.rate) {
 106                 err = -ENOMEM;
 107                 R_tab = qdisc_get_rtab(&parm->rate, tb[TCA_POLICE_RATE], NULL);
 108                 if (R_tab == NULL)
 109                         goto failure;
 110
 111                 if (parm->peakrate.rate) {
 112                         P_tab = qdisc_get_rtab(&parm->peakrate,
 113                                                tb[TCA_POLICE_PEAKRATE], NULL);
 114                         if (P_tab == NULL)
 115                                 goto failure;
 116                 }
 117         }
 118
 119         if (est) {
 120                 err = gen_replace_estimator(&police->tcf_bstats,
 121                                             police->common.cpu_bstats,
 122                                             &police->tcf_rate_est,
 123                                             &police->tcf_lock,
 124                                             NULL, est);
 125                 if (err)
 126                         goto failure;
 127         } else if (tb[TCA_POLICE_AVRATE] &&
 128                    (ret == ACT_P_CREATED ||
 129                     !gen_estimator_active(&police->tcf_rate_est))) {
 130                 err = -EINVAL;
 131                 goto failure;
 132         }
 133
 134         if (tb[TCA_POLICE_RESULT]) {
 135                 tcfp_result = nla_get_u32(tb[TCA_POLICE_RESULT]);
 136                 if (TC_ACT_EXT_CMP(tcfp_result, TC_ACT_GOTO_CHAIN)) {
 137                         NL_SET_ERR_MSG(extack,
 138                                        "goto chain not allowed on fallback");
 139                         err = -EINVAL;
 140                         goto failure;
 141                 }
 142         }
 143
 144         new = kzalloc(sizeof(*new), GFP_KERNEL);
 145         if (unlikely(!new)) {
 146                 err = -ENOMEM;
 147                 goto failure;
 148         }
 149
 150         /* No failure allowed after this point */
 151         new->tcfp_result = tcfp_result;
 152         new->tcfp_mtu = parm->mtu;
 153         if (!new->tcfp_mtu) {
 154                 new->tcfp_mtu = ~0;
 155                 if (R_tab)
 156                         new->tcfp_mtu = 255 << R_tab->rate.cell_log;
 157         }
 158         if (R_tab) {
 159                 new->rate_present = true;
 160                 psched_ratecfg_precompute(&new->rate, &R_tab->rate, 0);
 161                 qdisc_put_rtab(R_tab);
 162         } else {
 163                 new->rate_present = false;
 164         }
 165         if (P_tab) {
 166                 new->peak_present = true;
 167                 psched_ratecfg_precompute(&new->peak, &P_tab->rate, 0);
 168                 qdisc_put_rtab(P_tab);
 169         } else {
 170                 new->peak_present = false;
 171         }
 172
 173         new->tcfp_burst = PSCHED_TICKS2NS(parm->burst);
 174         if (new->peak_present)
 175                 new->tcfp_mtu_ptoks = (s64)psched_l2t_ns(&new->peak,
 176                                                          new->tcfp_mtu);
 177
 178         if (tb[TCA_POLICE_AVRATE])
 179                 new->tcfp_ewma_rate = nla_get_u32(tb[TCA_POLICE_AVRATE]);
 180
 181         spin_lock_bh(&police->tcf_lock);
 182         spin_lock_bh(&police->tcfp_lock);
 183         police->tcfp_t_c = ktime_get_ns();
 184         police->tcfp_toks = new->tcfp_burst;
 185         if (new->peak_present)
 186                 police->tcfp_ptoks = new->tcfp_mtu_ptoks;
 187         spin_unlock_bh(&police->tcfp_lock);
 188         goto_ch = tcf_action_set_ctrlact(*a, parm->action, goto_ch);
 189         rcu_swap_protected(police->params,
 190                            new,
 191                            lockdep_is_held(&police->tcf_lock));
 192         spin_unlock_bh(&police->tcf_lock);
 193
 194         if (goto_ch)
 195                 tcf_chain_put_by_act(goto_ch);
 196         if (new)
 197                 kfree_rcu(new, rcu);
 198
 199         if (ret == ACT_P_CREATED)
 200                 tcf_idr_insert(tn, *a);
 201         return ret;
 202
 203 failure:
 204         qdisc_put_rtab(P_tab);
 205         qdisc_put_rtab(R_tab);
 206         if (goto_ch)
 207                 tcf_chain_put_by_act(goto_ch);
 208 release_idr:
 209         tcf_idr_release(*a, bind);
 210         return err;
 211 }
 212
 213 static int tcf_police_act(struct sk_buff *skb, const struct tc_action *a,
 214                           struct tcf_result *res)
 215 {
 216         struct tcf_police *police = to_police(a);
 217         struct tcf_police_params *p;
 218         s64 now, toks, ptoks = 0;
 219         int ret;
 220
 221         tcf_lastuse_update(&police->tcf_tm);
 222         bstats_cpu_update(this_cpu_ptr(police->common.cpu_bstats), skb);
 223
 224         ret = READ_ONCE(police->tcf_action);
 225         p = rcu_dereference_bh(police->params);
 226
 227         if (p->tcfp_ewma_rate) {
 228                 struct gnet_stats_rate_est64 sample;
 229
 230                 if (!gen_estimator_read(&police->tcf_rate_est, &sample) ||
 231                     sample.bps >= p->tcfp_ewma_rate)
 232                         goto inc_overlimits;
 233         }
 234
 235         if (qdisc_pkt_len(skb) <= p->tcfp_mtu) {
 236                 if (!p->rate_present) {
 237                         ret = p->tcfp_result;
 238                         goto end;
 239                 }
 240
 241                 now = ktime_get_ns();
 242                 spin_lock_bh(&police->tcfp_lock);
 243                 toks = min_t(s64, now - police->tcfp_t_c, p->tcfp_burst);
 244                 if (p->peak_present) {
 245                         ptoks = toks + police->tcfp_ptoks;
 246                         if (ptoks > p->tcfp_mtu_ptoks)
 247                                 ptoks = p->tcfp_mtu_ptoks;
 248                         ptoks -= (s64)psched_l2t_ns(&p->peak,
 249                                                     qdisc_pkt_len(skb));
 250                 }
 251                 toks += police->tcfp_toks;
 252                 if (toks > p->tcfp_burst)
 253                         toks = p->tcfp_burst;
 254                 toks -= (s64)psched_l2t_ns(&p->rate, qdisc_pkt_len(skb));
 255                 if ((toks|ptoks) >= 0) {
 256                         police->tcfp_t_c = now;
 257                         police->tcfp_toks = toks;
 258                         police->tcfp_ptoks = ptoks;
 259                         spin_unlock_bh(&police->tcfp_lock);
 260                         ret = p->tcfp_result;
 261                         goto inc_drops;
 262                 }
 263                 spin_unlock_bh(&police->tcfp_lock);
 264         }
 265
 266 inc_overlimits:
 267         qstats_overlimit_inc(this_cpu_ptr(police->common.cpu_qstats));
 268 inc_drops:
 269         if (ret == TC_ACT_SHOT)
 270                 qstats_drop_inc(this_cpu_ptr(police->common.cpu_qstats));
 271 end:
 272         return ret;
 273 }
 274
 275 static void tcf_police_cleanup(struct tc_action *a)
 276 {
 277         struct tcf_police *police = to_police(a);
 278         struct tcf_police_params *p;
 279
 280         p = rcu_dereference_protected(police->params, 1);
 281         if (p)
 282                 kfree_rcu(p, rcu);
 283 }
 284
 285 static void tcf_police_stats_update(struct tc_action *a,
 286                                     u64 bytes, u32 packets,
 287                                     u64 lastuse, bool hw)
 288 {
 289         struct tcf_police *police = to_police(a);
 290         struct tcf_t *tm = &police->tcf_tm;
 291
 292         _bstats_cpu_update(this_cpu_ptr(a->cpu_bstats), bytes, packets);
 293         if (hw)
 294                 _bstats_cpu_update(this_cpu_ptr(a->cpu_bstats_hw),
 295                                    bytes, packets);
 296         tm->lastuse = max_t(u64, tm->lastuse, lastuse);
 297 }
 298
 299 static int tcf_police_dump(struct sk_buff *skb, struct tc_action *a,
 300                                int bind, int ref)
 301 {
 302         unsigned char *b = skb_tail_pointer(skb);
 303         struct tcf_police *police = to_police(a);
 304         struct tcf_police_params *p;
 305         struct tc_police opt = {
 306                 .index = police->tcf_index,
 307                 .refcnt = refcount_read(&police->tcf_refcnt) - ref,
 308                 .bindcnt = atomic_read(&police->tcf_bindcnt) - bind,
 309         };
 310         struct tcf_t t;
 311
 312         spin_lock_bh(&police->tcf_lock);
 313         opt.action = police->tcf_action;
 314         p = rcu_dereference_protected(police->params,
 315                                       lockdep_is_held(&police->tcf_lock));
 316         opt.mtu = p->tcfp_mtu;
 317         opt.burst = PSCHED_NS2TICKS(p->tcfp_burst);
 318         if (p->rate_present)
 319                 psched_ratecfg_getrate(&opt.rate, &p->rate);
 320         if (p->peak_present)
 321                 psched_ratecfg_getrate(&opt.peakrate, &p->peak);
 322         if (nla_put(skb, TCA_POLICE_TBF, sizeof(opt), &opt))
 323                 goto nla_put_failure;
 324         if (p->tcfp_result &&
 325             nla_put_u32(skb, TCA_POLICE_RESULT, p->tcfp_result))
 326                 goto nla_put_failure;
 327         if (p->tcfp_ewma_rate &&
 328             nla_put_u32(skb, TCA_POLICE_AVRATE, p->tcfp_ewma_rate))
 329                 goto nla_put_failure;
 330
 331         t.install = jiffies_to_clock_t(jiffies - police->tcf_tm.install);
 332         t.lastuse = jiffies_to_clock_t(jiffies - police->tcf_tm.lastuse);
 333         t.firstuse = jiffies_to_clock_t(jiffies - police->tcf_tm.firstuse);
 334         t.expires = jiffies_to_clock_t(police->tcf_tm.expires);
 335         if (nla_put_64bit(skb, TCA_POLICE_TM, sizeof(t), &t, TCA_POLICE_PAD))
 336                 goto nla_put_failure;
 337         spin_unlock_bh(&police->tcf_lock);
 338
 339         return skb->len;
 340
 341 nla_put_failure:
 342         spin_unlock_bh(&police->tcf_lock);
 343         nlmsg_trim(skb, b);
 344         return -1;
 345 }
 346
 347 static int tcf_police_search(struct net *net, struct tc_action **a, u32 index)
 348 {
 349         struct tc_action_net *tn = net_generic(net, police_net_id);
 350
 351         return tcf_idr_search(tn, a, index);
 352 }
 353
 354 MODULE_AUTHOR("Alexey Kuznetsov");
 355 MODULE_DESCRIPTION("Policing actions");
 356 MODULE_LICENSE("GPL");
 357
 358 static struct tc_action_ops act_police_ops = {
 359         .kind           =       "police",
 360         .id             =       TCA_ID_POLICE,
 361         .owner          =       THIS_MODULE,
 362         .stats_update   =       tcf_police_stats_update,
 363         .act            =       tcf_police_act,
 364         .dump           =       tcf_police_dump,
 365         .init           =       tcf_police_init,
 366         .walk           =       tcf_police_walker,
 367         .lookup         =       tcf_police_search,
 368         .cleanup        =       tcf_police_cleanup,
 369         .size           =       sizeof(struct tcf_police),
 370 };
 371
 372 static __net_init int police_init_net(struct net *net)
 373 {
 374         struct tc_action_net *tn = net_generic(net, police_net_id);
 375
 376         return tc_action_net_init(tn, &act_police_ops);
 377 }
 378
 379 static void __net_exit police_exit_net(struct list_head *net_list)
 380 {
 381         tc_action_net_exit(net_list, police_net_id);
 382 }
 383
 384 static struct pernet_operations police_net_ops = {
 385         .init = police_init_net,
 386         .exit_batch = police_exit_net,
 387         .id   = &police_net_id,
 388         .size = sizeof(struct tc_action_net),
 389 };
 390
 391 static int __init police_init_module(void)
 392 {
 393         return tcf_register_action(&act_police_ops, &police_net_ops);
 394 }
 395
 396 static void __exit police_cleanup_module(void)
 397 {
 398         tcf_unregister_action(&act_police_ops, &police_net_ops);
 399 }
 400
 401 module_init(police_init_module);
 402 module_exit(police_cleanup_module);