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_destroy(struct tcf_police *p)
 100 {
 101         unsigned int h = tcf_hash(p->tcf_index, POL_TAB_MASK);
 102         struct tcf_common **p1p;
 103
 104         for (p1p = &tcf_police_ht[h]; *p1p; p1p = &(*p1p)->tcfc_next) {
 105                 if (*p1p == &p->common) {
 106                         write_lock_bh(&police_lock);
 107                         *p1p = p->tcf_next;
 108                         write_unlock_bh(&police_lock);
 109                         gen_kill_estimator(&p->tcf_bstats,
 110                                            &p->tcf_rate_est);
 111                         if (p->tcfp_R_tab)
 112                                 qdisc_put_rtab(p->tcfp_R_tab);
 113                         if (p->tcfp_P_tab)
 114                                 qdisc_put_rtab(p->tcfp_P_tab);
 115                         /*
 116                          * gen_estimator est_timer() might access p->tcf_lock
 117                          * or bstats, wait a RCU grace period before freeing p
 118                          */
 119                         kfree_rcu(p, tcf_rcu);
 120                         return;
 121                 }
 122         }
 123         WARN_ON(1);
 124 }
 125
 126 static const struct nla_policy police_policy[TCA_POLICE_MAX + 1] = {
 127         [TCA_POLICE_RATE]       = { .len = TC_RTAB_SIZE },
 128         [TCA_POLICE_PEAKRATE]   = { .len = TC_RTAB_SIZE },
 129         [TCA_POLICE_AVRATE]     = { .type = NLA_U32 },
 130         [TCA_POLICE_RESULT]     = { .type = NLA_U32 },
 131 };
 132
 133 static int tcf_act_police_locate(struct net *net, struct nlattr *nla,
 134                                  struct nlattr *est, struct tc_action *a,
 135                                  int ovr, int bind)
 136 {
 137         unsigned int h;
 138         int ret = 0, err;
 139         struct nlattr *tb[TCA_POLICE_MAX + 1];
 140         struct tc_police *parm;
 141         struct tcf_police *police;
 142         struct qdisc_rate_table *R_tab = NULL, *P_tab = NULL;
 143         int size;
 144
 145         if (nla == NULL)
 146                 return -EINVAL;
 147
 148         err = nla_parse_nested(tb, TCA_POLICE_MAX, nla, police_policy);
 149         if (err < 0)
 150                 return err;
 151
 152         if (tb[TCA_POLICE_TBF] == NULL)
 153                 return -EINVAL;
 154         size = nla_len(tb[TCA_POLICE_TBF]);
 155         if (size != sizeof(*parm) && size != sizeof(struct tc_police_compat))
 156                 return -EINVAL;
 157         parm = nla_data(tb[TCA_POLICE_TBF]);
 158
 159         if (parm->index) {
 160                 struct tcf_common *pc;
 161
 162                 pc = tcf_hash_lookup(parm->index, &police_hash_info);
 163                 if (pc != NULL) {
 164                         a->priv = pc;
 165                         police = to_police(pc);
 166                         if (bind) {
 167                                 police->tcf_bindcnt += 1;
 168                                 police->tcf_refcnt += 1;
 169                         }
 170                         if (ovr)
 171                                 goto override;
 172                         return ret;
 173                 }
 174         }
 175
 176         police = kzalloc(sizeof(*police), GFP_KERNEL);
 177         if (police == NULL)
 178                 return -ENOMEM;
 179         ret = ACT_P_CREATED;
 180         police->tcf_refcnt = 1;
 181         spin_lock_init(&police->tcf_lock);
 182         if (bind)
 183                 police->tcf_bindcnt = 1;
 184 override:
 185         if (parm->rate.rate) {
 186                 err = -ENOMEM;
 187                 R_tab = qdisc_get_rtab(&parm->rate, tb[TCA_POLICE_RATE]);
 188                 if (R_tab == NULL)
 189                         goto failure;
 190
 191                 if (parm->peakrate.rate) {
 192                         P_tab = qdisc_get_rtab(&parm->peakrate,
 193                                                tb[TCA_POLICE_PEAKRATE]);
 194                         if (P_tab == NULL)
 195                                 goto failure;
 196                 }
 197         }
 198
 199         spin_lock_bh(&police->tcf_lock);
 200         if (est) {
 201                 err = gen_replace_estimator(&police->tcf_bstats,
 202                                             &police->tcf_rate_est,
 203                                             &police->tcf_lock, est);
 204                 if (err)
 205                         goto failure_unlock;
 206         } else if (tb[TCA_POLICE_AVRATE] &&
 207                    (ret == ACT_P_CREATED ||
 208                     !gen_estimator_active(&police->tcf_bstats,
 209                                           &police->tcf_rate_est))) {
 210                 err = -EINVAL;
 211                 goto failure_unlock;
 212         }
 213
 214         /* No failure allowed after this point */
 215         if (R_tab != NULL) {
 216                 qdisc_put_rtab(police->tcfp_R_tab);
 217                 police->tcfp_R_tab = R_tab;
 218         }
 219         if (P_tab != NULL) {
 220                 qdisc_put_rtab(police->tcfp_P_tab);
 221                 police->tcfp_P_tab = P_tab;
 222         }
 223
 224         if (tb[TCA_POLICE_RESULT])
 225                 police->tcfp_result = nla_get_u32(tb[TCA_POLICE_RESULT]);
 226         police->tcfp_toks = police->tcfp_burst = parm->burst;
 227         police->tcfp_mtu = parm->mtu;
 228         if (police->tcfp_mtu == 0) {
 229                 police->tcfp_mtu = ~0;
 230                 if (police->tcfp_R_tab)
 231                         police->tcfp_mtu = 255<<police->tcfp_R_tab->rate.cell_log;
 232         }
 233         if (police->tcfp_P_tab)
 234                 police->tcfp_ptoks = L2T_P(police, police->tcfp_mtu);
 235         police->tcf_action = parm->action;
 236
 237         if (tb[TCA_POLICE_AVRATE])
 238                 police->tcfp_ewma_rate = nla_get_u32(tb[TCA_POLICE_AVRATE]);
 239
 240         spin_unlock_bh(&police->tcf_lock);
 241         if (ret != ACT_P_CREATED)
 242                 return ret;
 243
 244         police->tcfp_t_c = psched_get_time();
 245         police->tcf_index = parm->index ? parm->index :
 246                 tcf_hash_new_index(&police_idx_gen, &police_hash_info);
 247         h = tcf_hash(police->tcf_index, POL_TAB_MASK);
 248         write_lock_bh(&police_lock);
 249         police->tcf_next = tcf_police_ht[h];
 250         tcf_police_ht[h] = &police->common;
 251         write_unlock_bh(&police_lock);
 252
 253         a->priv = police;
 254         return ret;
 255
 256 failure_unlock:
 257         spin_unlock_bh(&police->tcf_lock);
 258 failure:
 259         if (P_tab)
 260                 qdisc_put_rtab(P_tab);
 261         if (R_tab)
 262                 qdisc_put_rtab(R_tab);
 263         if (ret == ACT_P_CREATED)
 264                 kfree(police);
 265         return err;
 266 }
 267
 268 static int tcf_act_police_cleanup(struct tc_action *a, int bind)
 269 {
 270         struct tcf_police *p = a->priv;
 271         int ret = 0;
 272
 273         if (p != NULL) {
 274                 if (bind)
 275                         p->tcf_bindcnt--;
 276
 277                 p->tcf_refcnt--;
 278                 if (p->tcf_refcnt <= 0 && !p->tcf_bindcnt) {
 279                         tcf_police_destroy(p);
 280                         ret = 1;
 281                 }
 282         }
 283         return ret;
 284 }
 285
 286 static int tcf_act_police(struct sk_buff *skb, const struct tc_action *a,
 287                           struct tcf_result *res)
 288 {
 289         struct tcf_police *police = a->priv;
 290         psched_time_t now;
 291         long toks;
 292         long ptoks = 0;
 293
 294         spin_lock(&police->tcf_lock);
 295
 296         bstats_update(&police->tcf_bstats, skb);
 297
 298         if (police->tcfp_ewma_rate &&
 299             police->tcf_rate_est.bps >= police->tcfp_ewma_rate) {
 300                 police->tcf_qstats.overlimits++;
 301                 if (police->tcf_action == TC_ACT_SHOT)
 302                         police->tcf_qstats.drops++;
 303                 spin_unlock(&police->tcf_lock);
 304                 return police->tcf_action;
 305         }
 306
 307         if (qdisc_pkt_len(skb) <= police->tcfp_mtu) {
 308                 if (police->tcfp_R_tab == NULL) {
 309                         spin_unlock(&police->tcf_lock);
 310                         return police->tcfp_result;
 311                 }
 312
 313                 now = psched_get_time();
 314                 toks = psched_tdiff_bounded(now, police->tcfp_t_c,
 315                                             police->tcfp_burst);
 316                 if (police->tcfp_P_tab) {
 317                         ptoks = toks + police->tcfp_ptoks;
 318                         if (ptoks > (long)L2T_P(police, police->tcfp_mtu))
 319                                 ptoks = (long)L2T_P(police, police->tcfp_mtu);
 320                         ptoks -= L2T_P(police, qdisc_pkt_len(skb));
 321                 }
 322                 toks += police->tcfp_toks;
 323                 if (toks > (long)police->tcfp_burst)
 324                         toks = police->tcfp_burst;
 325                 toks -= L2T(police, qdisc_pkt_len(skb));
 326                 if ((toks|ptoks) >= 0) {
 327                         police->tcfp_t_c = now;
 328                         police->tcfp_toks = toks;
 329                         police->tcfp_ptoks = ptoks;
 330                         spin_unlock(&police->tcf_lock);
 331                         return police->tcfp_result;
 332                 }
 333         }
 334
 335         police->tcf_qstats.overlimits++;
 336         if (police->tcf_action == TC_ACT_SHOT)
 337                 police->tcf_qstats.drops++;
 338         spin_unlock(&police->tcf_lock);
 339         return police->tcf_action;
 340 }
 341
 342 static int
 343 tcf_act_police_dump(struct sk_buff *skb, struct tc_action *a, int bind, int ref)
 344 {
 345         unsigned char *b = skb_tail_pointer(skb);
 346         struct tcf_police *police = a->priv;
 347         struct tc_police opt = {
 348                 .index = police->tcf_index,
 349                 .action = police->tcf_action,
 350                 .mtu = police->tcfp_mtu,
 351                 .burst = police->tcfp_burst,
 352                 .refcnt = police->tcf_refcnt - ref,
 353                 .bindcnt = police->tcf_bindcnt - bind,
 354         };
 355
 356         if (police->tcfp_R_tab)
 357                 opt.rate = police->tcfp_R_tab->rate;
 358         if (police->tcfp_P_tab)
 359                 opt.peakrate = police->tcfp_P_tab->rate;
 360         if (nla_put(skb, TCA_POLICE_TBF, sizeof(opt), &opt))
 361                 goto nla_put_failure;
 362         if (police->tcfp_result &&
 363             nla_put_u32(skb, TCA_POLICE_RESULT, police->tcfp_result))
 364                 goto nla_put_failure;
 365         if (police->tcfp_ewma_rate &&
 366             nla_put_u32(skb, TCA_POLICE_AVRATE, police->tcfp_ewma_rate))
 367                 goto nla_put_failure;
 368         return skb->len;
 369
 370 nla_put_failure:
 371         nlmsg_trim(skb, b);
 372         return -1;
 373 }
 374
 375 MODULE_AUTHOR("Alexey Kuznetsov");
 376 MODULE_DESCRIPTION("Policing actions");
 377 MODULE_LICENSE("GPL");
 378
 379 static struct tc_action_ops act_police_ops = {
 380         .kind           =       "police",
 381         .hinfo          =       &police_hash_info,
 382         .type           =       TCA_ID_POLICE,
 383         .capab          =       TCA_CAP_NONE,
 384         .owner          =       THIS_MODULE,
 385         .act            =       tcf_act_police,
 386         .dump           =       tcf_act_police_dump,
 387         .cleanup        =       tcf_act_police_cleanup,
 388         .lookup         =       tcf_hash_search,
 389         .init           =       tcf_act_police_locate,
 390         .walk           =       tcf_act_police_walker
 391 };
 392
 393 static int __init
 394 police_init_module(void)
 395 {
 396         return tcf_register_action(&act_police_ops);
 397 }
 398
 399 static void __exit
 400 police_cleanup_module(void)
 401 {
 402         tcf_unregister_action(&act_police_ops);
 403 }
 404
 405 module_init(police_init_module);
 406 module_exit(police_cleanup_module);