git.oblomov.eu Git - linux-2.6/blob - 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 <net/act_api.h>
  22 #include <net/netlink.h>
  23
  24 #define L2T(p,L)   qdisc_l2t((p)->tcfp_R_tab, L)
  25 #define L2T_P(p,L) qdisc_l2t((p)->tcfp_P_tab, L)
  26
  27 #define POL_TAB_MASK     15
  28 static struct tcf_common *tcf_police_ht[POL_TAB_MASK + 1];
  29 static u32 police_idx_gen;
  30 static DEFINE_RWLOCK(police_lock);
  31
  32 static struct tcf_hashinfo police_hash_info = {
  33         .htab   =       tcf_police_ht,
  34         .hmask  =       POL_TAB_MASK,
  35         .lock   =       &police_lock,
  36 };
  37
  38 /* old policer structure from before tc actions */
  39 struct tc_police_compat
  40 {
  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                         kfree(p);
 116                         return;
 117                 }
 118         }
 119         WARN_ON(1);
 120 }
 121
 122 static const struct nla_policy police_policy[TCA_POLICE_MAX + 1] = {
 123         [TCA_POLICE_RATE]       = { .len = TC_RTAB_SIZE },
 124         [TCA_POLICE_PEAKRATE]   = { .len = TC_RTAB_SIZE },
 125         [TCA_POLICE_AVRATE]     = { .type = NLA_U32 },
 126         [TCA_POLICE_RESULT]     = { .type = NLA_U32 },
 127 };
 128
 129 static int tcf_act_police_locate(struct nlattr *nla, struct nlattr *est,
 130                                  struct tc_action *a, int ovr, int bind)
 131 {
 132         unsigned h;
 133         int ret = 0, err;
 134         struct nlattr *tb[TCA_POLICE_MAX + 1];
 135         struct tc_police *parm;
 136         struct tcf_police *police;
 137         struct qdisc_rate_table *R_tab = NULL, *P_tab = NULL;
 138         int size;
 139
 140         if (nla == NULL)
 141                 return -EINVAL;
 142
 143         err = nla_parse_nested(tb, TCA_POLICE_MAX, nla, police_policy);
 144         if (err < 0)
 145                 return err;
 146
 147         if (tb[TCA_POLICE_TBF] == NULL)
 148                 return -EINVAL;
 149         size = nla_len(tb[TCA_POLICE_TBF]);
 150         if (size != sizeof(*parm) && size != sizeof(struct tc_police_compat))
 151                 return -EINVAL;
 152         parm = nla_data(tb[TCA_POLICE_TBF]);
 153
 154         if (parm->index) {
 155                 struct tcf_common *pc;
 156
 157                 pc = tcf_hash_lookup(parm->index, &police_hash_info);
 158                 if (pc != NULL) {
 159                         a->priv = pc;
 160                         police = to_police(pc);
 161                         if (bind) {
 162                                 police->tcf_bindcnt += 1;
 163                                 police->tcf_refcnt += 1;
 164                         }
 165                         if (ovr)
 166                                 goto override;
 167                         return ret;
 168                 }
 169         }
 170
 171         police = kzalloc(sizeof(*police), GFP_KERNEL);
 172         if (police == NULL)
 173                 return -ENOMEM;
 174         ret = ACT_P_CREATED;
 175         police->tcf_refcnt = 1;
 176         spin_lock_init(&police->tcf_lock);
 177         if (bind)
 178                 police->tcf_bindcnt = 1;
 179 override:
 180         if (parm->rate.rate) {
 181                 err = -ENOMEM;
 182                 R_tab = qdisc_get_rtab(&parm->rate, tb[TCA_POLICE_RATE]);
 183                 if (R_tab == NULL)
 184                         goto failure;
 185
 186                 if (!est && (ret == ACT_P_CREATED ||
 187                              !gen_estimator_active(&police->tcf_bstats,
 188                                                    &police->tcf_rate_est))) {
 189                         err = -EINVAL;
 190                         goto failure;
 191                 }
 192
 193                 if (parm->peakrate.rate) {
 194                         P_tab = qdisc_get_rtab(&parm->peakrate,
 195                                                tb[TCA_POLICE_PEAKRATE]);
 196                         if (P_tab == NULL)
 197                                 goto failure;
 198                 }
 199         }
 200
 201         spin_lock_bh(&police->tcf_lock);
 202         if (est) {
 203                 err = gen_replace_estimator(&police->tcf_bstats,
 204                                             &police->tcf_rate_est,
 205                                             &police->tcf_lock, est);
 206                 if (err)
 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                 spin_unlock(&police->tcf_lock);
 299                 return police->tcf_action;
 300         }
 301
 302         if (qdisc_pkt_len(skb) <= police->tcfp_mtu) {
 303                 if (police->tcfp_R_tab == NULL) {
 304                         spin_unlock(&police->tcf_lock);
 305                         return police->tcfp_result;
 306                 }
 307
 308                 now = psched_get_time();
 309                 toks = psched_tdiff_bounded(now, police->tcfp_t_c,
 310                                             police->tcfp_burst);
 311                 if (police->tcfp_P_tab) {
 312                         ptoks = toks + police->tcfp_ptoks;
 313                         if (ptoks > (long)L2T_P(police, police->tcfp_mtu))
 314                                 ptoks = (long)L2T_P(police, police->tcfp_mtu);
 315                         ptoks -= L2T_P(police, qdisc_pkt_len(skb));
 316                 }
 317                 toks += police->tcfp_toks;
 318                 if (toks > (long)police->tcfp_burst)
 319                         toks = police->tcfp_burst;
 320                 toks -= L2T(police, qdisc_pkt_len(skb));
 321                 if ((toks|ptoks) >= 0) {
 322                         police->tcfp_t_c = now;
 323                         police->tcfp_toks = toks;
 324                         police->tcfp_ptoks = ptoks;
 325                         spin_unlock(&police->tcf_lock);
 326                         return police->tcfp_result;
 327                 }
 328         }
 329
 330         police->tcf_qstats.overlimits++;
 331         spin_unlock(&police->tcf_lock);
 332         return police->tcf_action;
 333 }
 334
 335 static int
 336 tcf_act_police_dump(struct sk_buff *skb, struct tc_action *a, int bind, int ref)
 337 {
 338         unsigned char *b = skb_tail_pointer(skb);
 339         struct tcf_police *police = a->priv;
 340         struct tc_police opt;
 341
 342         opt.index = police->tcf_index;
 343         opt.action = police->tcf_action;
 344         opt.mtu = police->tcfp_mtu;
 345         opt.burst = police->tcfp_burst;
 346         opt.refcnt = police->tcf_refcnt - ref;
 347         opt.bindcnt = police->tcf_bindcnt - bind;
 348         if (police->tcfp_R_tab)
 349                 opt.rate = police->tcfp_R_tab->rate;
 350         else
 351                 memset(&opt.rate, 0, sizeof(opt.rate));
 352         if (police->tcfp_P_tab)
 353                 opt.peakrate = police->tcfp_P_tab->rate;
 354         else
 355                 memset(&opt.peakrate, 0, sizeof(opt.peakrate));
 356         NLA_PUT(skb, TCA_POLICE_TBF, sizeof(opt), &opt);
 357         if (police->tcfp_result)
 358                 NLA_PUT_U32(skb, TCA_POLICE_RESULT, police->tcfp_result);
 359         if (police->tcfp_ewma_rate)
 360                 NLA_PUT_U32(skb, TCA_POLICE_AVRATE, police->tcfp_ewma_rate);
 361         return skb->len;
 362
 363 nla_put_failure:
 364         nlmsg_trim(skb, b);
 365         return -1;
 366 }
 367
 368 MODULE_AUTHOR("Alexey Kuznetsov");
 369 MODULE_DESCRIPTION("Policing actions");
 370 MODULE_LICENSE("GPL");
 371
 372 static struct tc_action_ops act_police_ops = {
 373         .kind           =       "police",
 374         .hinfo          =       &police_hash_info,
 375         .type           =       TCA_ID_POLICE,
 376         .capab          =       TCA_CAP_NONE,
 377         .owner          =       THIS_MODULE,
 378         .act            =       tcf_act_police,
 379         .dump           =       tcf_act_police_dump,
 380         .cleanup        =       tcf_act_police_cleanup,
 381         .lookup         =       tcf_hash_search,
 382         .init           =       tcf_act_police_locate,
 383         .walk           =       tcf_act_police_walker
 384 };
 385
 386 static int __init
 387 police_init_module(void)
 388 {
 389         return tcf_register_action(&act_police_ops);
 390 }
 391
 392 static void __exit
 393 police_cleanup_module(void)
 394 {
 395         tcf_unregister_action(&act_police_ops);
 396 }
 397
 398 module_init(police_init_module);
 399 module_exit(police_cleanup_module);