Pileus Git - ~andy/linux/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 <linux/slab.h>
  22 #include <net/act_api.h>
  23 #include <net/netlink.h>
  24
  25 struct tcf_police {
  26         struct tcf_common       common;
  27         int                     tcfp_result;
  28         u32                     tcfp_ewma_rate;
  29         s64                     tcfp_burst;
  30         u32                     tcfp_mtu;
  31         s64                     tcfp_toks;
  32         s64                     tcfp_ptoks;
  33         s64                     tcfp_mtu_ptoks;
  34         s64                     tcfp_t_c;
  35         struct psched_ratecfg   rate;
  36         bool                    rate_present;
  37         struct psched_ratecfg   peak;
  38         bool                    peak_present;
  39 };
  40 #define to_police(pc)   \
  41         container_of(pc, struct tcf_police, common)
  42
  43 #define POL_TAB_MASK     15
  44 static u32 police_idx_gen;
  45 static struct tcf_hashinfo police_hash_info;
  46
  47 /* old policer structure from before tc actions */
  48 struct tc_police_compat {
  49         u32                     index;
  50         int                     action;
  51         u32                     limit;
  52         u32                     burst;
  53         u32                     mtu;
  54         struct tc_ratespec      rate;
  55         struct tc_ratespec      peakrate;
  56 };
  57
  58 /* Each policer is serialized by its individual spinlock */
  59
  60 static int tcf_act_police_walker(struct sk_buff *skb, struct netlink_callback *cb,
  61                               int type, struct tc_action *a)
  62 {
  63         struct hlist_head *head;
  64         struct tcf_common *p;
  65         int err = 0, index = -1, i = 0, s_i = 0, n_i = 0;
  66         struct nlattr *nest;
  67
  68         spin_lock_bh(&police_hash_info.lock);
  69
  70         s_i = cb->args[0];
  71
  72         for (i = 0; i < (POL_TAB_MASK + 1); i++) {
  73                 head = &police_hash_info.htab[tcf_hash(i, POL_TAB_MASK)];
  74
  75                 hlist_for_each_entry_rcu(p, head, tcfc_head) {
  76                         index++;
  77                         if (index < s_i)
  78                                 continue;
  79                         a->priv = p;
  80                         a->order = index;
  81                         nest = nla_nest_start(skb, a->order);
  82                         if (nest == NULL)
  83                                 goto nla_put_failure;
  84                         if (type == RTM_DELACTION)
  85                                 err = tcf_action_dump_1(skb, a, 0, 1);
  86                         else
  87                                 err = tcf_action_dump_1(skb, a, 0, 0);
  88                         if (err < 0) {
  89                                 index--;
  90                                 nla_nest_cancel(skb, nest);
  91                                 goto done;
  92                         }
  93                         nla_nest_end(skb, nest);
  94                         n_i++;
  95                 }
  96         }
  97 done:
  98         spin_unlock_bh(&police_hash_info.lock);
  99         if (n_i)
 100                 cb->args[0] += n_i;
 101         return n_i;
 102
 103 nla_put_failure:
 104         nla_nest_cancel(skb, nest);
 105         goto done;
 106 }
 107
 108 static void tcf_police_destroy(struct tcf_police *p)
 109 {
 110         spin_lock_bh(&police_hash_info.lock);
 111         hlist_del(&p->tcf_head);
 112         spin_unlock_bh(&police_hash_info.lock);
 113         gen_kill_estimator(&p->tcf_bstats,
 114                            &p->tcf_rate_est);
 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 }
 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 net *net, struct nlattr *nla,
 130                                  struct nlattr *est, struct tc_action *a,
 131                                  int ovr, int bind)
 132 {
 133         unsigned int 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                                 return 0;
 166                         }
 167                         if (ovr)
 168                                 goto override;
 169                         /* not replacing */
 170                         return -EEXIST;
 171                 }
 172         }
 173
 174         police = kzalloc(sizeof(*police), GFP_KERNEL);
 175         if (police == NULL)
 176                 return -ENOMEM;
 177         ret = ACT_P_CREATED;
 178         police->tcf_refcnt = 1;
 179         spin_lock_init(&police->tcf_lock);
 180         if (bind)
 181                 police->tcf_bindcnt = 1;
 182 override:
 183         if (parm->rate.rate) {
 184                 err = -ENOMEM;
 185                 R_tab = qdisc_get_rtab(&parm->rate, tb[TCA_POLICE_RATE]);
 186                 if (R_tab == NULL)
 187                         goto failure;
 188
 189                 if (parm->peakrate.rate) {
 190                         P_tab = qdisc_get_rtab(&parm->peakrate,
 191                                                tb[TCA_POLICE_PEAKRATE]);
 192                         if (P_tab == NULL)
 193                                 goto failure;
 194                 }
 195         }
 196
 197         spin_lock_bh(&police->tcf_lock);
 198         if (est) {
 199                 err = gen_replace_estimator(&police->tcf_bstats,
 200                                             &police->tcf_rate_est,
 201                                             &police->tcf_lock, est);
 202                 if (err)
 203                         goto failure_unlock;
 204         } else if (tb[TCA_POLICE_AVRATE] &&
 205                    (ret == ACT_P_CREATED ||
 206                     !gen_estimator_active(&police->tcf_bstats,
 207                                           &police->tcf_rate_est))) {
 208                 err = -EINVAL;
 209                 goto failure_unlock;
 210         }
 211
 212         /* No failure allowed after this point */
 213         police->tcfp_mtu = parm->mtu;
 214         if (police->tcfp_mtu == 0) {
 215                 police->tcfp_mtu = ~0;
 216                 if (R_tab)
 217                         police->tcfp_mtu = 255 << R_tab->rate.cell_log;
 218         }
 219         if (R_tab) {
 220                 police->rate_present = true;
 221                 psched_ratecfg_precompute(&police->rate, &R_tab->rate, 0);
 222                 qdisc_put_rtab(R_tab);
 223         } else {
 224                 police->rate_present = false;
 225         }
 226         if (P_tab) {
 227                 police->peak_present = true;
 228                 psched_ratecfg_precompute(&police->peak, &P_tab->rate, 0);
 229                 qdisc_put_rtab(P_tab);
 230         } else {
 231                 police->peak_present = false;
 232         }
 233
 234         if (tb[TCA_POLICE_RESULT])
 235                 police->tcfp_result = nla_get_u32(tb[TCA_POLICE_RESULT]);
 236         police->tcfp_burst = PSCHED_TICKS2NS(parm->burst);
 237         police->tcfp_toks = police->tcfp_burst;
 238         if (police->peak_present) {
 239                 police->tcfp_mtu_ptoks = (s64) psched_l2t_ns(&police->peak,
 240                                                              police->tcfp_mtu);
 241                 police->tcfp_ptoks = police->tcfp_mtu_ptoks;
 242         }
 243         police->tcf_action = parm->action;
 244
 245         if (tb[TCA_POLICE_AVRATE])
 246                 police->tcfp_ewma_rate = nla_get_u32(tb[TCA_POLICE_AVRATE]);
 247
 248         spin_unlock_bh(&police->tcf_lock);
 249         if (ret != ACT_P_CREATED)
 250                 return ret;
 251
 252         police->tcfp_t_c = ktime_to_ns(ktime_get());
 253         police->tcf_index = parm->index ? parm->index :
 254                 tcf_hash_new_index(&police_idx_gen, &police_hash_info);
 255         h = tcf_hash(police->tcf_index, POL_TAB_MASK);
 256         spin_lock_bh(&police_hash_info.lock);
 257         hlist_add_head(&police->tcf_head, &police_hash_info.htab[h]);
 258         spin_unlock_bh(&police_hash_info.lock);
 259
 260         a->priv = police;
 261         return ret;
 262
 263 failure_unlock:
 264         spin_unlock_bh(&police->tcf_lock);
 265 failure:
 266         qdisc_put_rtab(P_tab);
 267         qdisc_put_rtab(R_tab);
 268         if (ret == ACT_P_CREATED)
 269                 kfree(police);
 270         return err;
 271 }
 272
 273 static int tcf_act_police_cleanup(struct tc_action *a, int bind)
 274 {
 275         struct tcf_police *p = a->priv;
 276         int ret = 0;
 277
 278         if (p != NULL) {
 279                 if (bind)
 280                         p->tcf_bindcnt--;
 281
 282                 p->tcf_refcnt--;
 283                 if (p->tcf_refcnt <= 0 && !p->tcf_bindcnt) {
 284                         tcf_police_destroy(p);
 285                         ret = 1;
 286                 }
 287         }
 288         return ret;
 289 }
 290
 291 static int tcf_act_police(struct sk_buff *skb, const struct tc_action *a,
 292                           struct tcf_result *res)
 293 {
 294         struct tcf_police *police = a->priv;
 295         s64 now;
 296         s64 toks;
 297         s64 ptoks = 0;
 298
 299         spin_lock(&police->tcf_lock);
 300
 301         bstats_update(&police->tcf_bstats, skb);
 302
 303         if (police->tcfp_ewma_rate &&
 304             police->tcf_rate_est.bps >= police->tcfp_ewma_rate) {
 305                 police->tcf_qstats.overlimits++;
 306                 if (police->tcf_action == TC_ACT_SHOT)
 307                         police->tcf_qstats.drops++;
 308                 spin_unlock(&police->tcf_lock);
 309                 return police->tcf_action;
 310         }
 311
 312         if (qdisc_pkt_len(skb) <= police->tcfp_mtu) {
 313                 if (!police->rate_present) {
 314                         spin_unlock(&police->tcf_lock);
 315                         return police->tcfp_result;
 316                 }
 317
 318                 now = ktime_to_ns(ktime_get());
 319                 toks = min_t(s64, now - police->tcfp_t_c,
 320                              police->tcfp_burst);
 321                 if (police->peak_present) {
 322                         ptoks = toks + police->tcfp_ptoks;
 323                         if (ptoks > police->tcfp_mtu_ptoks)
 324                                 ptoks = police->tcfp_mtu_ptoks;
 325                         ptoks -= (s64) psched_l2t_ns(&police->peak,
 326                                                      qdisc_pkt_len(skb));
 327                 }
 328                 toks += police->tcfp_toks;
 329                 if (toks > police->tcfp_burst)
 330                         toks = police->tcfp_burst;
 331                 toks -= (s64) psched_l2t_ns(&police->rate, qdisc_pkt_len(skb));
 332                 if ((toks|ptoks) >= 0) {
 333                         police->tcfp_t_c = now;
 334                         police->tcfp_toks = toks;
 335                         police->tcfp_ptoks = ptoks;
 336                         spin_unlock(&police->tcf_lock);
 337                         return police->tcfp_result;
 338                 }
 339         }
 340
 341         police->tcf_qstats.overlimits++;
 342         if (police->tcf_action == TC_ACT_SHOT)
 343                 police->tcf_qstats.drops++;
 344         spin_unlock(&police->tcf_lock);
 345         return police->tcf_action;
 346 }
 347
 348 static int
 349 tcf_act_police_dump(struct sk_buff *skb, struct tc_action *a, int bind, int ref)
 350 {
 351         unsigned char *b = skb_tail_pointer(skb);
 352         struct tcf_police *police = a->priv;
 353         struct tc_police opt = {
 354                 .index = police->tcf_index,
 355                 .action = police->tcf_action,
 356                 .mtu = police->tcfp_mtu,
 357                 .burst = PSCHED_NS2TICKS(police->tcfp_burst),
 358                 .refcnt = police->tcf_refcnt - ref,
 359                 .bindcnt = police->tcf_bindcnt - bind,
 360         };
 361
 362         if (police->rate_present)
 363                 psched_ratecfg_getrate(&opt.rate, &police->rate);
 364         if (police->peak_present)
 365                 psched_ratecfg_getrate(&opt.peakrate, &police->peak);
 366         if (nla_put(skb, TCA_POLICE_TBF, sizeof(opt), &opt))
 367                 goto nla_put_failure;
 368         if (police->tcfp_result &&
 369             nla_put_u32(skb, TCA_POLICE_RESULT, police->tcfp_result))
 370                 goto nla_put_failure;
 371         if (police->tcfp_ewma_rate &&
 372             nla_put_u32(skb, TCA_POLICE_AVRATE, police->tcfp_ewma_rate))
 373                 goto nla_put_failure;
 374         return skb->len;
 375
 376 nla_put_failure:
 377         nlmsg_trim(skb, b);
 378         return -1;
 379 }
 380
 381 MODULE_AUTHOR("Alexey Kuznetsov");
 382 MODULE_DESCRIPTION("Policing actions");
 383 MODULE_LICENSE("GPL");
 384
 385 static struct tc_action_ops act_police_ops = {
 386         .kind           =       "police",
 387         .hinfo          =       &police_hash_info,
 388         .type           =       TCA_ID_POLICE,
 389         .capab          =       TCA_CAP_NONE,
 390         .owner          =       THIS_MODULE,
 391         .act            =       tcf_act_police,
 392         .dump           =       tcf_act_police_dump,
 393         .cleanup        =       tcf_act_police_cleanup,
 394         .init           =       tcf_act_police_locate,
 395         .walk           =       tcf_act_police_walker
 396 };
 397
 398 static int __init
 399 police_init_module(void)
 400 {
 401         int err = tcf_hashinfo_init(&police_hash_info, POL_TAB_MASK);
 402         if (err)
 403                 return err;
 404         err = tcf_register_action(&act_police_ops);
 405         if (err)
 406                 tcf_hashinfo_destroy(&police_hash_info);
 407         return err;
 408 }
 409
 410 static void __exit
 411 police_cleanup_module(void)
 412 {
 413         tcf_hashinfo_destroy(&police_hash_info);
 414         tcf_unregister_action(&act_police_ops);
 415 }
 416
 417 module_init(police_init_module);
 418 module_exit(police_cleanup_module);