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 struct tcf_hashinfo police_hash_info;
  45
  46 /* old policer structure from before tc actions */
  47 struct tc_police_compat {
  48         u32                     index;
  49         int                     action;
  50         u32                     limit;
  51         u32                     burst;
  52         u32                     mtu;
  53         struct tc_ratespec      rate;
  54         struct tc_ratespec      peakrate;
  55 };
  56
  57 /* Each policer is serialized by its individual spinlock */
  58
  59 static int tcf_act_police_walker(struct sk_buff *skb, struct netlink_callback *cb,
  60                               int type, struct tc_action *a)
  61 {
  62         struct hlist_head *head;
  63         struct tcf_common *p;
  64         int err = 0, index = -1, i = 0, s_i = 0, n_i = 0;
  65         struct nlattr *nest;
  66
  67         spin_lock_bh(&police_hash_info.lock);
  68
  69         s_i = cb->args[0];
  70
  71         for (i = 0; i < (POL_TAB_MASK + 1); i++) {
  72                 head = &police_hash_info.htab[tcf_hash(i, POL_TAB_MASK)];
  73
  74                 hlist_for_each_entry_rcu(p, head, tcfc_head) {
  75                         index++;
  76                         if (index < s_i)
  77                                 continue;
  78                         a->priv = p;
  79                         a->order = index;
  80                         nest = nla_nest_start(skb, a->order);
  81                         if (nest == NULL)
  82                                 goto nla_put_failure;
  83                         if (type == RTM_DELACTION)
  84                                 err = tcf_action_dump_1(skb, a, 0, 1);
  85                         else
  86                                 err = tcf_action_dump_1(skb, a, 0, 0);
  87                         if (err < 0) {
  88                                 index--;
  89                                 nla_nest_cancel(skb, nest);
  90                                 goto done;
  91                         }
  92                         nla_nest_end(skb, nest);
  93                         n_i++;
  94                 }
  95         }
  96 done:
  97         spin_unlock_bh(&police_hash_info.lock);
  98         if (n_i)
  99                 cb->args[0] += n_i;
 100         return n_i;
 101
 102 nla_put_failure:
 103         nla_nest_cancel(skb, nest);
 104         goto done;
 105 }
 106
 107 static void tcf_police_destroy(struct tcf_police *p)
 108 {
 109         spin_lock_bh(&police_hash_info.lock);
 110         hlist_del(&p->tcf_head);
 111         spin_unlock_bh(&police_hash_info.lock);
 112         gen_kill_estimator(&p->tcf_bstats,
 113                            &p->tcf_rate_est);
 114         /*
 115          * gen_estimator est_timer() might access p->tcf_lock
 116          * or bstats, wait a RCU grace period before freeing p
 117          */
 118         kfree_rcu(p, tcf_rcu);
 119 }
 120
 121 static const struct nla_policy police_policy[TCA_POLICE_MAX + 1] = {
 122         [TCA_POLICE_RATE]       = { .len = TC_RTAB_SIZE },
 123         [TCA_POLICE_PEAKRATE]   = { .len = TC_RTAB_SIZE },
 124         [TCA_POLICE_AVRATE]     = { .type = NLA_U32 },
 125         [TCA_POLICE_RESULT]     = { .type = NLA_U32 },
 126 };
 127
 128 static int tcf_act_police_locate(struct net *net, struct nlattr *nla,
 129                                  struct nlattr *est, struct tc_action *a,
 130                                  int ovr, int bind)
 131 {
 132         unsigned int 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                                 return 0;
 165                         }
 166                         if (ovr)
 167                                 goto override;
 168                         /* not replacing */
 169                         return -EEXIST;
 170                 }
 171         }
 172
 173         police = kzalloc(sizeof(*police), GFP_KERNEL);
 174         if (police == NULL)
 175                 return -ENOMEM;
 176         ret = ACT_P_CREATED;
 177         police->tcf_refcnt = 1;
 178         spin_lock_init(&police->tcf_lock);
 179         if (bind)
 180                 police->tcf_bindcnt = 1;
 181 override:
 182         if (parm->rate.rate) {
 183                 err = -ENOMEM;
 184                 R_tab = qdisc_get_rtab(&parm->rate, tb[TCA_POLICE_RATE]);
 185                 if (R_tab == NULL)
 186                         goto failure;
 187
 188                 if (parm->peakrate.rate) {
 189                         P_tab = qdisc_get_rtab(&parm->peakrate,
 190                                                tb[TCA_POLICE_PEAKRATE]);
 191                         if (P_tab == NULL)
 192                                 goto failure;
 193                 }
 194         }
 195
 196         spin_lock_bh(&police->tcf_lock);
 197         if (est) {
 198                 err = gen_replace_estimator(&police->tcf_bstats,
 199                                             &police->tcf_rate_est,
 200                                             &police->tcf_lock, est);
 201                 if (err)
 202                         goto failure_unlock;
 203         } else if (tb[TCA_POLICE_AVRATE] &&
 204                    (ret == ACT_P_CREATED ||
 205                     !gen_estimator_active(&police->tcf_bstats,
 206                                           &police->tcf_rate_est))) {
 207                 err = -EINVAL;
 208                 goto failure_unlock;
 209         }
 210
 211         /* No failure allowed after this point */
 212         police->tcfp_mtu = parm->mtu;
 213         if (police->tcfp_mtu == 0) {
 214                 police->tcfp_mtu = ~0;
 215                 if (R_tab)
 216                         police->tcfp_mtu = 255 << R_tab->rate.cell_log;
 217         }
 218         if (R_tab) {
 219                 police->rate_present = true;
 220                 psched_ratecfg_precompute(&police->rate, &R_tab->rate, 0);
 221                 qdisc_put_rtab(R_tab);
 222         } else {
 223                 police->rate_present = false;
 224         }
 225         if (P_tab) {
 226                 police->peak_present = true;
 227                 psched_ratecfg_precompute(&police->peak, &P_tab->rate, 0);
 228                 qdisc_put_rtab(P_tab);
 229         } else {
 230                 police->peak_present = false;
 231         }
 232
 233         if (tb[TCA_POLICE_RESULT])
 234                 police->tcfp_result = nla_get_u32(tb[TCA_POLICE_RESULT]);
 235         police->tcfp_burst = PSCHED_TICKS2NS(parm->burst);
 236         police->tcfp_toks = police->tcfp_burst;
 237         if (police->peak_present) {
 238                 police->tcfp_mtu_ptoks = (s64) psched_l2t_ns(&police->peak,
 239                                                              police->tcfp_mtu);
 240                 police->tcfp_ptoks = police->tcfp_mtu_ptoks;
 241         }
 242         police->tcf_action = parm->action;
 243
 244         if (tb[TCA_POLICE_AVRATE])
 245                 police->tcfp_ewma_rate = nla_get_u32(tb[TCA_POLICE_AVRATE]);
 246
 247         spin_unlock_bh(&police->tcf_lock);
 248         if (ret != ACT_P_CREATED)
 249                 return ret;
 250
 251         police->tcfp_t_c = ktime_to_ns(ktime_get());
 252         police->tcf_index = parm->index ? parm->index :
 253                 tcf_hash_new_index(&police_hash_info);
 254         h = tcf_hash(police->tcf_index, POL_TAB_MASK);
 255         spin_lock_bh(&police_hash_info.lock);
 256         hlist_add_head(&police->tcf_head, &police_hash_info.htab[h]);
 257         spin_unlock_bh(&police_hash_info.lock);
 258
 259         a->priv = police;
 260         return ret;
 261
 262 failure_unlock:
 263         spin_unlock_bh(&police->tcf_lock);
 264 failure:
 265         qdisc_put_rtab(P_tab);
 266         qdisc_put_rtab(R_tab);
 267         if (ret == ACT_P_CREATED)
 268                 kfree(police);
 269         return err;
 270 }
 271
 272 static int tcf_act_police_cleanup(struct tc_action *a, int bind)
 273 {
 274         struct tcf_police *p = a->priv;
 275         int ret = 0;
 276
 277         if (p != NULL) {
 278                 if (bind)
 279                         p->tcf_bindcnt--;
 280
 281                 p->tcf_refcnt--;
 282                 if (p->tcf_refcnt <= 0 && !p->tcf_bindcnt) {
 283                         tcf_police_destroy(p);
 284                         ret = 1;
 285                 }
 286         }
 287         return ret;
 288 }
 289
 290 static int tcf_act_police(struct sk_buff *skb, const struct tc_action *a,
 291                           struct tcf_result *res)
 292 {
 293         struct tcf_police *police = a->priv;
 294         s64 now;
 295         s64 toks;
 296         s64 ptoks = 0;
 297
 298         spin_lock(&police->tcf_lock);
 299
 300         bstats_update(&police->tcf_bstats, skb);
 301
 302         if (police->tcfp_ewma_rate &&
 303             police->tcf_rate_est.bps >= police->tcfp_ewma_rate) {
 304                 police->tcf_qstats.overlimits++;
 305                 if (police->tcf_action == TC_ACT_SHOT)
 306                         police->tcf_qstats.drops++;
 307                 spin_unlock(&police->tcf_lock);
 308                 return police->tcf_action;
 309         }
 310
 311         if (qdisc_pkt_len(skb) <= police->tcfp_mtu) {
 312                 if (!police->rate_present) {
 313                         spin_unlock(&police->tcf_lock);
 314                         return police->tcfp_result;
 315                 }
 316
 317                 now = ktime_to_ns(ktime_get());
 318                 toks = min_t(s64, now - police->tcfp_t_c,
 319                              police->tcfp_burst);
 320                 if (police->peak_present) {
 321                         ptoks = toks + police->tcfp_ptoks;
 322                         if (ptoks > police->tcfp_mtu_ptoks)
 323                                 ptoks = police->tcfp_mtu_ptoks;
 324                         ptoks -= (s64) psched_l2t_ns(&police->peak,
 325                                                      qdisc_pkt_len(skb));
 326                 }
 327                 toks += police->tcfp_toks;
 328                 if (toks > police->tcfp_burst)
 329                         toks = police->tcfp_burst;
 330                 toks -= (s64) psched_l2t_ns(&police->rate, qdisc_pkt_len(skb));
 331                 if ((toks|ptoks) >= 0) {
 332                         police->tcfp_t_c = now;
 333                         police->tcfp_toks = toks;
 334                         police->tcfp_ptoks = ptoks;
 335                         spin_unlock(&police->tcf_lock);
 336                         return police->tcfp_result;
 337                 }
 338         }
 339
 340         police->tcf_qstats.overlimits++;
 341         if (police->tcf_action == TC_ACT_SHOT)
 342                 police->tcf_qstats.drops++;
 343         spin_unlock(&police->tcf_lock);
 344         return police->tcf_action;
 345 }
 346
 347 static int
 348 tcf_act_police_dump(struct sk_buff *skb, struct tc_action *a, int bind, int ref)
 349 {
 350         unsigned char *b = skb_tail_pointer(skb);
 351         struct tcf_police *police = a->priv;
 352         struct tc_police opt = {
 353                 .index = police->tcf_index,
 354                 .action = police->tcf_action,
 355                 .mtu = police->tcfp_mtu,
 356                 .burst = PSCHED_NS2TICKS(police->tcfp_burst),
 357                 .refcnt = police->tcf_refcnt - ref,
 358                 .bindcnt = police->tcf_bindcnt - bind,
 359         };
 360
 361         if (police->rate_present)
 362                 psched_ratecfg_getrate(&opt.rate, &police->rate);
 363         if (police->peak_present)
 364                 psched_ratecfg_getrate(&opt.peakrate, &police->peak);
 365         if (nla_put(skb, TCA_POLICE_TBF, sizeof(opt), &opt))
 366                 goto nla_put_failure;
 367         if (police->tcfp_result &&
 368             nla_put_u32(skb, TCA_POLICE_RESULT, police->tcfp_result))
 369                 goto nla_put_failure;
 370         if (police->tcfp_ewma_rate &&
 371             nla_put_u32(skb, TCA_POLICE_AVRATE, police->tcfp_ewma_rate))
 372                 goto nla_put_failure;
 373         return skb->len;
 374
 375 nla_put_failure:
 376         nlmsg_trim(skb, b);
 377         return -1;
 378 }
 379
 380 MODULE_AUTHOR("Alexey Kuznetsov");
 381 MODULE_DESCRIPTION("Policing actions");
 382 MODULE_LICENSE("GPL");
 383
 384 static struct tc_action_ops act_police_ops = {
 385         .kind           =       "police",
 386         .hinfo          =       &police_hash_info,
 387         .type           =       TCA_ID_POLICE,
 388         .capab          =       TCA_CAP_NONE,
 389         .owner          =       THIS_MODULE,
 390         .act            =       tcf_act_police,
 391         .dump           =       tcf_act_police_dump,
 392         .cleanup        =       tcf_act_police_cleanup,
 393         .init           =       tcf_act_police_locate,
 394         .walk           =       tcf_act_police_walker
 395 };
 396
 397 static int __init
 398 police_init_module(void)
 399 {
 400         int err = tcf_hashinfo_init(&police_hash_info, POL_TAB_MASK);
 401         if (err)
 402                 return err;
 403         err = tcf_register_action(&act_police_ops);
 404         if (err)
 405                 tcf_hashinfo_destroy(&police_hash_info);
 406         return err;
 407 }
 408
 409 static void __exit
 410 police_cleanup_module(void)
 411 {
 412         tcf_hashinfo_destroy(&police_hash_info);
 413         tcf_unregister_action(&act_police_ops);
 414 }
 415
 416 module_init(police_init_module);
 417 module_exit(police_cleanup_module);