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
  25 struct tcf_police {
  26         struct tc_action        common;
  27         int                     tcfp_result;
  28         u32                     tcfp_ewma_rate;
  29         s64                     tcfp_burst;
  30         u32                     tcfp_mtu;
  31         s64                     tcfp_toks;
  32         s64                     tcfp_ptoks;
  33         s64                     tcfp_mtu_ptoks;
  34         s64                     tcfp_t_c;
  35         struct psched_ratecfg   rate;
  36         bool                    rate_present;
  37         struct psched_ratecfg   peak;
  38         bool                    peak_present;
  39 };
  40
  41 #define to_police(pc) ((struct tcf_police *)pc)
  42
  43 #define POL_TAB_MASK     15
  44
  45 /* old policer structure from before tc actions */
  46 struct tc_police_compat {
  47         u32                     index;
  48         int                     action;
  49         u32                     limit;
  50         u32                     burst;
  51         u32                     mtu;
  52         struct tc_ratespec      rate;
  53         struct tc_ratespec      peakrate;
  54 };
  55
  56 /* Each policer is serialized by its individual spinlock */
  57
  58 static int police_net_id;
  59 static struct tc_action_ops act_police_ops;
  60
  61 static int tcf_act_police_walker(struct net *net, struct sk_buff *skb,
  62                                  struct netlink_callback *cb, int type,
  63                                  const struct tc_action_ops *ops)
  64 {
  65         struct tc_action_net *tn = net_generic(net, police_net_id);
  66         struct tcf_hashinfo *hinfo = tn->hinfo;
  67         int err = 0, index = -1, i = 0, s_i = 0, n_i = 0;
  68         struct nlattr *nest;
  69
  70         spin_lock_bh(&hinfo->lock);
  71
  72         s_i = cb->args[0];
  73
  74         for (i = 0; i < (POL_TAB_MASK + 1); i++) {
  75                 struct hlist_head *head;
  76                 struct tc_action *p;
  77
  78                 head = &hinfo->htab[tcf_hash(i, POL_TAB_MASK)];
  79
  80                 hlist_for_each_entry_rcu(p, head, tcfa_head) {
  81                         index++;
  82                         if (index < s_i)
  83                                 continue;
  84                         nest = nla_nest_start(skb, index);
  85                         if (nest == NULL)
  86                                 goto nla_put_failure;
  87                         if (type == RTM_DELACTION)
  88                                 err = tcf_action_dump_1(skb, p, 0, 1);
  89                         else
  90                                 err = tcf_action_dump_1(skb, p, 0, 0);
  91                         if (err < 0) {
  92                                 index--;
  93                                 nla_nest_cancel(skb, nest);
  94                                 goto done;
  95                         }
  96                         nla_nest_end(skb, nest);
  97                         n_i++;
  98                 }
  99         }
 100 done:
 101         spin_unlock_bh(&hinfo->lock);
 102         if (n_i)
 103                 cb->args[0] += n_i;
 104         return n_i;
 105
 106 nla_put_failure:
 107         nla_nest_cancel(skb, nest);
 108         goto done;
 109 }
 110
 111 static const struct nla_policy police_policy[TCA_POLICE_MAX + 1] = {
 112         [TCA_POLICE_RATE]       = { .len = TC_RTAB_SIZE },
 113         [TCA_POLICE_PEAKRATE]   = { .len = TC_RTAB_SIZE },
 114         [TCA_POLICE_AVRATE]     = { .type = NLA_U32 },
 115         [TCA_POLICE_RESULT]     = { .type = NLA_U32 },
 116 };
 117
 118 static int tcf_act_police_init(struct net *net, struct nlattr *nla,
 119                                struct nlattr *est, struct tc_action **a,
 120                                int ovr, int bind)
 121 {
 122         int ret = 0, err;
 123         struct nlattr *tb[TCA_POLICE_MAX + 1];
 124         struct tc_police *parm;
 125         struct tcf_police *police;
 126         struct qdisc_rate_table *R_tab = NULL, *P_tab = NULL;
 127         struct tc_action_net *tn = net_generic(net, police_net_id);
 128         bool exists = false;
 129         int size;
 130
 131         if (nla == NULL)
 132                 return -EINVAL;
 133
 134         err = nla_parse_nested(tb, TCA_POLICE_MAX, nla, police_policy);
 135         if (err < 0)
 136                 return err;
 137
 138         if (tb[TCA_POLICE_TBF] == NULL)
 139                 return -EINVAL;
 140         size = nla_len(tb[TCA_POLICE_TBF]);
 141         if (size != sizeof(*parm) && size != sizeof(struct tc_police_compat))
 142                 return -EINVAL;
 143
 144         parm = nla_data(tb[TCA_POLICE_TBF]);
 145         exists = tcf_hash_check(tn, parm->index, a, bind);
 146         if (exists && bind)
 147                 return 0;
 148
 149         if (!exists) {
 150                 ret = tcf_hash_create(tn, parm->index, NULL, a,
 151                                       &act_police_ops, bind, false);
 152                 if (ret)
 153                         return ret;
 154                 ret = ACT_P_CREATED;
 155         } else {
 156                 tcf_hash_release(*a, bind);
 157                 if (!ovr)
 158                         return -EEXIST;
 159         }
 160
 161         police = to_police(*a);
 162         if (parm->rate.rate) {
 163                 err = -ENOMEM;
 164                 R_tab = qdisc_get_rtab(&parm->rate, tb[TCA_POLICE_RATE]);
 165                 if (R_tab == NULL)
 166                         goto failure;
 167
 168                 if (parm->peakrate.rate) {
 169                         P_tab = qdisc_get_rtab(&parm->peakrate,
 170                                                tb[TCA_POLICE_PEAKRATE]);
 171                         if (P_tab == NULL)
 172                                 goto failure;
 173                 }
 174         }
 175
 176         spin_lock_bh(&police->tcf_lock);
 177         if (est) {
 178                 err = gen_replace_estimator(&police->tcf_bstats, NULL,
 179                                             &police->tcf_rate_est,
 180                                             &police->tcf_lock,
 181                                             NULL, est);
 182                 if (err)
 183                         goto failure_unlock;
 184         } else if (tb[TCA_POLICE_AVRATE] &&
 185                    (ret == ACT_P_CREATED ||
 186                     !gen_estimator_active(&police->tcf_bstats,
 187                                           &police->tcf_rate_est))) {
 188                 err = -EINVAL;
 189                 goto failure_unlock;
 190         }
 191
 192         /* No failure allowed after this point */
 193         police->tcfp_mtu = parm->mtu;
 194         if (police->tcfp_mtu == 0) {
 195                 police->tcfp_mtu = ~0;
 196                 if (R_tab)
 197                         police->tcfp_mtu = 255 << R_tab->rate.cell_log;
 198         }
 199         if (R_tab) {
 200                 police->rate_present = true;
 201                 psched_ratecfg_precompute(&police->rate, &R_tab->rate, 0);
 202                 qdisc_put_rtab(R_tab);
 203         } else {
 204                 police->rate_present = false;
 205         }
 206         if (P_tab) {
 207                 police->peak_present = true;
 208                 psched_ratecfg_precompute(&police->peak, &P_tab->rate, 0);
 209                 qdisc_put_rtab(P_tab);
 210         } else {
 211                 police->peak_present = false;
 212         }
 213
 214         if (tb[TCA_POLICE_RESULT])
 215                 police->tcfp_result = nla_get_u32(tb[TCA_POLICE_RESULT]);
 216         police->tcfp_burst = PSCHED_TICKS2NS(parm->burst);
 217         police->tcfp_toks = police->tcfp_burst;
 218         if (police->peak_present) {
 219                 police->tcfp_mtu_ptoks = (s64) psched_l2t_ns(&police->peak,
 220                                                              police->tcfp_mtu);
 221                 police->tcfp_ptoks = police->tcfp_mtu_ptoks;
 222         }
 223         police->tcf_action = parm->action;
 224
 225         if (tb[TCA_POLICE_AVRATE])
 226                 police->tcfp_ewma_rate = nla_get_u32(tb[TCA_POLICE_AVRATE]);
 227
 228         spin_unlock_bh(&police->tcf_lock);
 229         if (ret != ACT_P_CREATED)
 230                 return ret;
 231
 232         police->tcfp_t_c = ktime_get_ns();
 233         tcf_hash_insert(tn, *a);
 234
 235         return ret;
 236
 237 failure_unlock:
 238         spin_unlock_bh(&police->tcf_lock);
 239 failure:
 240         qdisc_put_rtab(P_tab);
 241         qdisc_put_rtab(R_tab);
 242         if (ret == ACT_P_CREATED)
 243                 tcf_hash_cleanup(*a, est);
 244         return err;
 245 }
 246
 247 static int tcf_act_police(struct sk_buff *skb, const struct tc_action *a,
 248                           struct tcf_result *res)
 249 {
 250         struct tcf_police *police = to_police(a);
 251         s64 now;
 252         s64 toks;
 253         s64 ptoks = 0;
 254
 255         spin_lock(&police->tcf_lock);
 256
 257         bstats_update(&police->tcf_bstats, skb);
 258         tcf_lastuse_update(&police->tcf_tm);
 259
 260         if (police->tcfp_ewma_rate &&
 261             police->tcf_rate_est.bps >= police->tcfp_ewma_rate) {
 262                 police->tcf_qstats.overlimits++;
 263                 if (police->tcf_action == TC_ACT_SHOT)
 264                         police->tcf_qstats.drops++;
 265                 spin_unlock(&police->tcf_lock);
 266                 return police->tcf_action;
 267         }
 268
 269         if (qdisc_pkt_len(skb) <= police->tcfp_mtu) {
 270                 if (!police->rate_present) {
 271                         spin_unlock(&police->tcf_lock);
 272                         return police->tcfp_result;
 273                 }
 274
 275                 now = ktime_get_ns();
 276                 toks = min_t(s64, now - police->tcfp_t_c,
 277                              police->tcfp_burst);
 278                 if (police->peak_present) {
 279                         ptoks = toks + police->tcfp_ptoks;
 280                         if (ptoks > police->tcfp_mtu_ptoks)
 281                                 ptoks = police->tcfp_mtu_ptoks;
 282                         ptoks -= (s64) psched_l2t_ns(&police->peak,
 283                                                      qdisc_pkt_len(skb));
 284                 }
 285                 toks += police->tcfp_toks;
 286                 if (toks > police->tcfp_burst)
 287                         toks = police->tcfp_burst;
 288                 toks -= (s64) psched_l2t_ns(&police->rate, qdisc_pkt_len(skb));
 289                 if ((toks|ptoks) >= 0) {
 290                         police->tcfp_t_c = now;
 291                         police->tcfp_toks = toks;
 292                         police->tcfp_ptoks = ptoks;
 293                         spin_unlock(&police->tcf_lock);
 294                         return police->tcfp_result;
 295                 }
 296         }
 297
 298         police->tcf_qstats.overlimits++;
 299         if (police->tcf_action == TC_ACT_SHOT)
 300                 police->tcf_qstats.drops++;
 301         spin_unlock(&police->tcf_lock);
 302         return police->tcf_action;
 303 }
 304
 305 static int
 306 tcf_act_police_dump(struct sk_buff *skb, struct tc_action *a, int bind, int ref)
 307 {
 308         unsigned char *b = skb_tail_pointer(skb);
 309         struct tcf_police *police = to_police(a);
 310         struct tc_police opt = {
 311                 .index = police->tcf_index,
 312                 .action = police->tcf_action,
 313                 .mtu = police->tcfp_mtu,
 314                 .burst = PSCHED_NS2TICKS(police->tcfp_burst),
 315                 .refcnt = police->tcf_refcnt - ref,
 316                 .bindcnt = police->tcf_bindcnt - bind,
 317         };
 318         struct tcf_t t;
 319
 320         if (police->rate_present)
 321                 psched_ratecfg_getrate(&opt.rate, &police->rate);
 322         if (police->peak_present)
 323                 psched_ratecfg_getrate(&opt.peakrate, &police->peak);
 324         if (nla_put(skb, TCA_POLICE_TBF, sizeof(opt), &opt))
 325                 goto nla_put_failure;
 326         if (police->tcfp_result &&
 327             nla_put_u32(skb, TCA_POLICE_RESULT, police->tcfp_result))
 328                 goto nla_put_failure;
 329         if (police->tcfp_ewma_rate &&
 330             nla_put_u32(skb, TCA_POLICE_AVRATE, police->tcfp_ewma_rate))
 331                 goto nla_put_failure;
 332
 333         t.install = jiffies_to_clock_t(jiffies - police->tcf_tm.install);
 334         t.lastuse = jiffies_to_clock_t(jiffies - police->tcf_tm.lastuse);
 335         t.firstuse = jiffies_to_clock_t(jiffies - police->tcf_tm.firstuse);
 336         t.expires = jiffies_to_clock_t(police->tcf_tm.expires);
 337         if (nla_put_64bit(skb, TCA_POLICE_TM, sizeof(t), &t, TCA_POLICE_PAD))
 338                 goto nla_put_failure;
 339
 340         return skb->len;
 341
 342 nla_put_failure:
 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_hash_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         .type           =       TCA_ID_POLICE,
 361         .owner          =       THIS_MODULE,
 362         .act            =       tcf_act_police,
 363         .dump           =       tcf_act_police_dump,
 364         .init           =       tcf_act_police_init,
 365         .walk           =       tcf_act_police_walker,
 366         .lookup         =       tcf_police_search,
 367         .size           =       sizeof(struct tcf_police),
 368 };
 369
 370 static __net_init int police_init_net(struct net *net)
 371 {
 372         struct tc_action_net *tn = net_generic(net, police_net_id);
 373
 374         return tc_action_net_init(tn, &act_police_ops, POL_TAB_MASK);
 375 }
 376
 377 static void __net_exit police_exit_net(struct net *net)
 378 {
 379         struct tc_action_net *tn = net_generic(net, police_net_id);
 380
 381         tc_action_net_exit(tn);
 382 }
 383
 384 static struct pernet_operations police_net_ops = {
 385         .init = police_init_net,
 386         .exit = police_exit_net,
 387         .id   = &police_net_id,
 388         .size = sizeof(struct tc_action_net),
 389 };
 390
 391 static int __init
 392 police_init_module(void)
 393 {
 394         return tcf_register_action(&act_police_ops, &police_net_ops);
 395 }
 396
 397 static void __exit
 398 police_cleanup_module(void)
 399 {
 400         tcf_unregister_action(&act_police_ops, &police_net_ops);
 401 }
 402
 403 module_init(police_init_module);
 404 module_exit(police_cleanup_module);