net/sched/act_police.c

   1 /*
   2  * net/sched/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 #define L2T(p, L)   qdisc_l2t((p)->tcfp_R_tab, L)
  26 #define L2T_P(p, L) qdisc_l2t((p)->tcfp_P_tab, L)
  27
  28 #define POL_TAB_MASK     15
  29 static struct tcf_common *tcf_police_ht[POL_TAB_MASK + 1];
  30 static u32 police_idx_gen;
  31 static DEFINE_RWLOCK(police_lock);
  32
  33 static struct tcf_hashinfo police_hash_info = {
  34         .htab   =       tcf_police_ht,
  35         .hmask  =       POL_TAB_MASK,
  36         .lock   =       &police_lock,
  37 };
  38
  39 /* old policer structure from before tc actions */
  40 struct tc_police_compat {
  41         u32                     index;
  42         int                     action;
  43         u32                     limit;
  44         u32                     burst;
  45         u32                     mtu;
  46         struct tc_ratespec      rate;
  47         struct tc_ratespec      peakrate;
  48 };
  49
  50 /* Each policer is serialized by its individual spinlock */
  51
  52 static int tcf_act_police_walker(struct sk_buff *skb, struct netlink_callback *cb,
  53                               int type, struct tc_action *a)
  54 {
  55         struct tcf_common *p;
  56         int err = 0, index = -1, i = 0, s_i = 0, n_i = 0;
  57         struct nlattr *nest;
  58
  59         read_lock_bh(&police_lock);
  60
  61         s_i = cb->args[0];
  62
  63         for (i = 0; i < (POL_TAB_MASK + 1); i++) {
  64                 p = tcf_police_ht[tcf_hash(i, POL_TAB_MASK)];
  65
  66                 for (; p; p = p->tcfc_next) {
  67                         index++;
  68                         if (index < s_i)
  69                                 continue;
  70                         a->priv = p;
  71                         a->order = index;
  72                         nest = nla_nest_start(skb, a->order);
  73                         if (nest == NULL)
  74                                 goto nla_put_failure;
  75                         if (type == RTM_DELACTION)
  76                                 err = tcf_action_dump_1(skb, a, 0, 1);
  77                         else
  78                                 err = tcf_action_dump_1(skb, a, 0, 0);
  79                         if (err < 0) {
  80                                 index--;
  81                                 nla_nest_cancel(skb, nest);
  82                                 goto done;
  83                         }
  84                         nla_nest_end(skb, nest);
  85                         n_i++;
  86                 }
  87         }
  88 done:
  89         read_unlock_bh(&police_lock);
  90         if (n_i)
  91                 cb->args[0] += n_i;
  92         return n_i;
  93
  94 nla_put_failure:
  95         nla_nest_cancel(skb, nest);
  96         goto done;
  97 }
  98
  99 static void tcf_police_free_rcu(struct rcu_head *head)
 100 {
 101         kfree(container_of(head, struct tcf_police, tcf_rcu));
 102 }
 103
 104 static void tcf_police_destroy(struct tcf_police *p)
 105 {
 106         unsigned int h = tcf_hash(p->tcf_index, POL_TAB_MASK);
 107         struct tcf_common **p1p;
 108
 109         for (p1p = &tcf_police_ht[h]; *p1p; p1p = &(*p1p)->tcfc_next) {
 110                 if (*p1p == &p->common) {
 111                         write_lock_bh(&police_lock);
 112                         *p1p = p->tcf_next;
 113                         write_unlock_bh(&police_lock);
 114                         gen_kill_estimator(&p->tcf_bstats,
 115                                            &p->tcf_rate_est);
 116                         if (p->tcfp_R_tab)
 117                                 qdisc_put_rtab(p->tcfp_R_tab);
 118                         if (p->tcfp_P_tab)
 119                                 qdisc_put_rtab(p->tcfp_P_tab);
 120                         /*
 121                          * gen_estimator est_timer() might access p->tcf_lock
 122                          * or bstats, wait a RCU grace period before freeing p
 123                          */
 124                         call_rcu(&p->tcf_rcu, tcf_police_free_rcu);
 125                         return;
 126                 }
 127         }
 128         WARN_ON(1);
 129 }
 130
 131 static const struct nla_policy police_policy[TCA_POLICE_MAX + 1] = {
 132         [TCA_POLICE_RATE]       = { .len = TC_RTAB_SIZE },
 133         [TCA_POLICE_PEAKRATE]   = { .len = TC_RTAB_SIZE },
 134         [TCA_POLICE_AVRATE]     = { .type = NLA_U32 },
 135         [TCA_POLICE_RESULT]     = { .type = NLA_U32 },
 136 };
 137
 138 static int tcf_act_police_locate(struct nlattr *nla, struct nlattr *est,
 139                                  struct tc_action *a, int ovr, int bind)
 140 {
 141         unsigned int h;
 142         int ret = 0, err;
 143         struct nlattr *tb[TCA_POLICE_MAX + 1];
 144         struct tc_police *parm;
 145         struct tcf_police *police;
 146         struct qdisc_rate_table *R_tab = NULL, *P_tab = NULL;
 147         int size;
 148
 149         if (nla == NULL)
 150                 return -EINVAL;
 151
 152         err = nla_parse_nested(tb, TCA_POLICE_MAX, nla, police_policy);
 153         if (err < 0)
 154                 return err;
 155
 156         if (tb[TCA_POLICE_TBF] == NULL)
 157                 return -EINVAL;
 158         size = nla_len(tb[TCA_POLICE_TBF]);
 159         if (size != sizeof(*parm) && size != sizeof(struct tc_police_compat))
 160                 return -EINVAL;
 161         parm = nla_data(tb[TCA_POLICE_TBF]);
 162
 163         if (parm->index) {
 164                 struct tcf_common *pc;
 165
 166                 pc = tcf_hash_lookup(parm->index, &police_hash_info);
 167                 if (pc != NULL) {
 168                         a->priv = pc;
 169                         police = to_police(pc);
 170                         if (bind) {
 171                                 police->tcf_bindcnt += 1;
 172                                 police->tcf_refcnt += 1;
 173                         }
 174                         if (ovr)
 175                                 goto override;
 176                         return ret;
 177                 }
 178         }
 179
 180         police = kzalloc(sizeof(*police), GFP_KERNEL);
 181         if (police == NULL)
 182                 return -ENOMEM;
 183         ret = ACT_P_CREATED;
 184         police->tcf_refcnt = 1;
 185         spin_lock_init(&police->tcf_lock);
 186         if (bind)
 187                 police->tcf_bindcnt = 1;
 188 override:
 189         if (parm->rate.rate) {
 190                 err = -ENOMEM;
 191                 R_tab = qdisc_get_rtab(&parm->rate, tb[TCA_POLICE_RATE]);
 192                 if (R_tab == NULL)
 193                         goto failure;
 194
 195                 if (parm->peakrate.rate) {
 196                         P_tab = qdisc_get_rtab(&parm->peakrate,
 197                                                tb[TCA_POLICE_PEAKRATE]);
 198                         if (P_tab == NULL)
 199                                 goto failure;
 200                 }
 201         }
 202
 203         spin_lock_bh(&police->tcf_lock);
 204         if (est) {
 205                 err = gen_replace_estimator(&police->tcf_bstats,
 206                                             &police->tcf_rate_est,
 207                                             &police->tcf_lock, est);
 208                 if (err)
 209                         goto failure_unlock;
 210         } else if (tb[TCA_POLICE_AVRATE] &&
 211                    (ret == ACT_P_CREATED ||
 212                     !gen_estimator_active(&police->tcf_bstats,
 213                                           &police->tcf_rate_est))) {
 214                 err = -EINVAL;
 215                 goto failure_unlock;
 216         }
 217
 218         /* No failure allowed after this point */
 219         if (R_tab != NULL) {
 220                 qdisc_put_rtab(police->tcfp_R_tab);
 221                 police->tcfp_R_tab = R_tab;
 222         }
 223         if (P_tab != NULL) {
 224                 qdisc_put_rtab(police->tcfp_P_tab);
 225                 police->tcfp_P_tab = P_tab;
 226         }
 227
 228         if (tb[TCA_POLICE_RESULT])
 229                 police->tcfp_result = nla_get_u32(tb[TCA_POLICE_RESULT]);
 230         police->tcfp_toks = police->tcfp_burst = parm->burst;
 231         police->tcfp_mtu = parm->mtu;
 232         if (police->tcfp_mtu == 0) {
 233                 police->tcfp_mtu = ~0;
 234                 if (police->tcfp_R_tab)
 235                         police->tcfp_mtu = 255<<police->tcfp_R_tab->rate.cell_log;
 236         }
 237         if (police->tcfp_P_tab)
 238                 police->tcfp_ptoks = L2T_P(police, police->tcfp_mtu);
 239         police->tcf_action = parm->action;
 240
 241         if (tb[TCA_POLICE_AVRATE])
 242                 police->tcfp_ewma_rate = nla_get_u32(tb[TCA_POLICE_AVRATE]);
 243
 244         spin_unlock_bh(&police->tcf_lock);
 245         if (ret != ACT_P_CREATED)
 246                 return ret;
 247
 248         police->tcfp_t_c = psched_get_time();
 249         police->tcf_index = parm->index ? parm->index :
 250                 tcf_hash_new_index(&police_idx_gen, &police_hash_info);
 251         h = tcf_hash(police->tcf_index, POL_TAB_MASK);
 252         write_lock_bh(&police_lock);
 253         police->tcf_next = tcf_police_ht[h];
 254         tcf_police_ht[h] = &police->common;
 255         write_unlock_bh(&police_lock);
 256
 257         a->priv = police;
 258         return ret;
 259
 260 failure_unlock:
 261         spin_unlock_bh(&police->tcf_lock);
 262 failure:
 263         if (P_tab)
 264                 qdisc_put_rtab(P_tab);
 265         if (R_tab)
 266                 qdisc_put_rtab(R_tab);
 267         if (ret == ACT_P_CREATED)
 268                 kfree(police);
 269         return err;
 270 }
 271
 272 static int tcf_act_police_cleanup(struct tc_action *a, int bind)
 273 {
 274         struct tcf_police *p = a->priv;
 275         int ret = 0;
 276
 277         if (p != NULL) {
 278                 if (bind)
 279                         p->tcf_bindcnt--;
 280
 281                 p->tcf_refcnt--;
 282                 if (p->tcf_refcnt <= 0 && !p->tcf_bindcnt) {
 283                         tcf_police_destroy(p);
 284                         ret = 1;
 285                 }
 286         }
 287         return ret;
 288 }
 289
 290 static int tcf_act_police(struct sk_buff *skb, struct tc_action *a,
 291                           struct tcf_result *res)
 292 {
 293         struct tcf_police *police = a->priv;
 294         psched_time_t now;
 295         long toks;
 296         long ptoks = 0;
 297
 298         spin_lock(&police->tcf_lock);
 299
 300         bstats_update(&police->tcf_bstats, skb);
 301
 302         if (police->tcfp_ewma_rate &&
 303             police->tcf_rate_est.bps >= police->tcfp_ewma_rate) {
 304                 police->tcf_qstats.overlimits++;
 305                 if (police->tcf_action == TC_ACT_SHOT)
 306                         police->tcf_qstats.drops++;
 307                 spin_unlock(&police->tcf_lock);
 308                 return police->tcf_action;
 309         }
 310
 311         if (qdisc_pkt_len(skb) <= police->tcfp_mtu) {
 312                 if (police->tcfp_R_tab == NULL) {
 313                         spin_unlock(&police->tcf_lock);
 314                         return police->tcfp_result;
 315                 }
 316
 317                 now = psched_get_time();
 318                 toks = psched_tdiff_bounded(now, police->tcfp_t_c,
 319                                             police->tcfp_burst);
 320                 if (police->tcfp_P_tab) {
 321                         ptoks = toks + police->tcfp_ptoks;
 322                         if (ptoks > (long)L2T_P(police, police->tcfp_mtu))
 323                                 ptoks = (long)L2T_P(police, police->tcfp_mtu);
 324                         ptoks -= L2T_P(police, qdisc_pkt_len(skb));
 325                 }
 326                 toks += police->tcfp_toks;
 327                 if (toks > (long)police->tcfp_burst)
 328                         toks = police->tcfp_burst;
 329                 toks -= L2T(police, qdisc_pkt_len(skb));
 330                 if ((toks|ptoks) >= 0) {
 331                         police->tcfp_t_c = now;
 332                         police->tcfp_toks = toks;
 333                         police->tcfp_ptoks = ptoks;
 334                         spin_unlock(&police->tcf_lock);
 335                         return police->tcfp_result;
 336                 }
 337         }
 338
 339         police->tcf_qstats.overlimits++;
 340         if (police->tcf_action == TC_ACT_SHOT)
 341                 police->tcf_qstats.drops++;
 342         spin_unlock(&police->tcf_lock);
 343         return police->tcf_action;
 344 }
 345
 346 static int
 347 tcf_act_police_dump(struct sk_buff *skb, struct tc_action *a, int bind, int ref)
 348 {
 349         unsigned char *b = skb_tail_pointer(skb);
 350         struct tcf_police *police = a->priv;
 351         struct tc_police opt = {
 352                 .index = police->tcf_index,
 353                 .action = police->tcf_action,
 354                 .mtu = police->tcfp_mtu,
 355                 .burst = police->tcfp_burst,
 356                 .refcnt = police->tcf_refcnt - ref,
 357                 .bindcnt = police->tcf_bindcnt - bind,
 358         };
 359
 360         if (police->tcfp_R_tab)
 361                 opt.rate = police->tcfp_R_tab->rate;
 362         if (police->tcfp_P_tab)
 363                 opt.peakrate = police->tcfp_P_tab->rate;
 364         NLA_PUT(skb, TCA_POLICE_TBF, sizeof(opt), &opt);
 365         if (police->tcfp_result)
 366                 NLA_PUT_U32(skb, TCA_POLICE_RESULT, police->tcfp_result);
 367         if (police->tcfp_ewma_rate)
 368                 NLA_PUT_U32(skb, TCA_POLICE_AVRATE, police->tcfp_ewma_rate);
 369         return skb->len;
 370
 371 nla_put_failure:
 372         nlmsg_trim(skb, b);
 373         return -1;
 374 }
 375
 376 MODULE_AUTHOR("Alexey Kuznetsov");
 377 MODULE_DESCRIPTION("Policing actions");
 378 MODULE_LICENSE("GPL");
 379
 380 static struct tc_action_ops act_police_ops = {
 381         .kind           =       "police",
 382         .hinfo          =       &police_hash_info,
 383         .type           =       TCA_ID_POLICE,
 384         .capab          =       TCA_CAP_NONE,
 385         .owner          =       THIS_MODULE,
 386         .act            =       tcf_act_police,
 387         .dump           =       tcf_act_police_dump,
 388         .cleanup        =       tcf_act_police_cleanup,
 389         .lookup         =       tcf_hash_search,
 390         .init           =       tcf_act_police_locate,
 391         .walk           =       tcf_act_police_walker
 392 };
 393
 394 static int __init
 395 police_init_module(void)
 396 {
 397         return tcf_register_action(&act_police_ops);
 398 }
 399
 400 static void __exit
 401 police_cleanup_module(void)
 402 {
 403         tcf_unregister_action(&act_police_ops);
 404         rcu_barrier(); /* Wait for completion of call_rcu()'s (tcf_police_free_rcu) */
 405 }
 406
 407 module_init(police_init_module);
 408 module_exit(police_cleanup_module);