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