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