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 const struct nla_policy police_policy[TCA_POLICE_MAX + 1] = {
 108         [TCA_POLICE_RATE]       = { .len = TC_RTAB_SIZE },
 109         [TCA_POLICE_PEAKRATE]   = { .len = TC_RTAB_SIZE },
 110         [TCA_POLICE_AVRATE]     = { .type = NLA_U32 },
 111         [TCA_POLICE_RESULT]     = { .type = NLA_U32 },
 112 };
 113
 114 static int tcf_act_police_locate(struct net *net, struct nlattr *nla,
 115                                  struct nlattr *est, struct tc_action *a,
 116                                  int ovr, int bind)
 117 {
 118         unsigned int h;
 119         int ret = 0, err;
 120         struct nlattr *tb[TCA_POLICE_MAX + 1];
 121         struct tc_police *parm;
 122         struct tcf_police *police;
 123         struct qdisc_rate_table *R_tab = NULL, *P_tab = NULL;
 124         int size;
 125
 126         if (nla == NULL)
 127                 return -EINVAL;
 128
 129         err = nla_parse_nested(tb, TCA_POLICE_MAX, nla, police_policy);
 130         if (err < 0)
 131                 return err;
 132
 133         if (tb[TCA_POLICE_TBF] == NULL)
 134                 return -EINVAL;
 135         size = nla_len(tb[TCA_POLICE_TBF]);
 136         if (size != sizeof(*parm) && size != sizeof(struct tc_police_compat))
 137                 return -EINVAL;
 138         parm = nla_data(tb[TCA_POLICE_TBF]);
 139
 140         if (parm->index) {
 141                 struct tcf_common *pc;
 142
 143                 pc = tcf_hash_lookup(parm->index, &police_hash_info);
 144                 if (pc != NULL) {
 145                         a->priv = pc;
 146                         police = to_police(pc);
 147                         if (bind) {
 148                                 police->tcf_bindcnt += 1;
 149                                 police->tcf_refcnt += 1;
 150                                 return 0;
 151                         }
 152                         if (ovr)
 153                                 goto override;
 154                         /* not replacing */
 155                         return -EEXIST;
 156                 }
 157         }
 158
 159         police = kzalloc(sizeof(*police), GFP_KERNEL);
 160         if (police == NULL)
 161                 return -ENOMEM;
 162         ret = ACT_P_CREATED;
 163         police->tcf_refcnt = 1;
 164         spin_lock_init(&police->tcf_lock);
 165         if (bind)
 166                 police->tcf_bindcnt = 1;
 167 override:
 168         if (parm->rate.rate) {
 169                 err = -ENOMEM;
 170                 R_tab = qdisc_get_rtab(&parm->rate, tb[TCA_POLICE_RATE]);
 171                 if (R_tab == NULL)
 172                         goto failure;
 173
 174                 if (parm->peakrate.rate) {
 175                         P_tab = qdisc_get_rtab(&parm->peakrate,
 176                                                tb[TCA_POLICE_PEAKRATE]);
 177                         if (P_tab == NULL)
 178                                 goto failure;
 179                 }
 180         }
 181
 182         spin_lock_bh(&police->tcf_lock);
 183         if (est) {
 184                 err = gen_replace_estimator(&police->tcf_bstats,
 185                                             &police->tcf_rate_est,
 186                                             &police->tcf_lock, est);
 187                 if (err)
 188                         goto failure_unlock;
 189         } else if (tb[TCA_POLICE_AVRATE] &&
 190                    (ret == ACT_P_CREATED ||
 191                     !gen_estimator_active(&police->tcf_bstats,
 192                                           &police->tcf_rate_est))) {
 193                 err = -EINVAL;
 194                 goto failure_unlock;
 195         }
 196
 197         /* No failure allowed after this point */
 198         police->tcfp_mtu = parm->mtu;
 199         if (police->tcfp_mtu == 0) {
 200                 police->tcfp_mtu = ~0;
 201                 if (R_tab)
 202                         police->tcfp_mtu = 255 << R_tab->rate.cell_log;
 203         }
 204         if (R_tab) {
 205                 police->rate_present = true;
 206                 psched_ratecfg_precompute(&police->rate, &R_tab->rate, 0);
 207                 qdisc_put_rtab(R_tab);
 208         } else {
 209                 police->rate_present = false;
 210         }
 211         if (P_tab) {
 212                 police->peak_present = true;
 213                 psched_ratecfg_precompute(&police->peak, &P_tab->rate, 0);
 214                 qdisc_put_rtab(P_tab);
 215         } else {
 216                 police->peak_present = false;
 217         }
 218
 219         if (tb[TCA_POLICE_RESULT])
 220                 police->tcfp_result = nla_get_u32(tb[TCA_POLICE_RESULT]);
 221         police->tcfp_burst = PSCHED_TICKS2NS(parm->burst);
 222         police->tcfp_toks = police->tcfp_burst;
 223         if (police->peak_present) {
 224                 police->tcfp_mtu_ptoks = (s64) psched_l2t_ns(&police->peak,
 225                                                              police->tcfp_mtu);
 226                 police->tcfp_ptoks = police->tcfp_mtu_ptoks;
 227         }
 228         police->tcf_action = parm->action;
 229
 230         if (tb[TCA_POLICE_AVRATE])
 231                 police->tcfp_ewma_rate = nla_get_u32(tb[TCA_POLICE_AVRATE]);
 232
 233         spin_unlock_bh(&police->tcf_lock);
 234         if (ret != ACT_P_CREATED)
 235                 return ret;
 236
 237         police->tcfp_t_c = ktime_to_ns(ktime_get());
 238         police->tcf_index = parm->index ? parm->index :
 239                 tcf_hash_new_index(&police_hash_info);
 240         h = tcf_hash(police->tcf_index, POL_TAB_MASK);
 241         spin_lock_bh(&police_hash_info.lock);
 242         hlist_add_head(&police->tcf_head, &police_hash_info.htab[h]);
 243         spin_unlock_bh(&police_hash_info.lock);
 244
 245         a->priv = police;
 246         return ret;
 247
 248 failure_unlock:
 249         spin_unlock_bh(&police->tcf_lock);
 250 failure:
 251         qdisc_put_rtab(P_tab);
 252         qdisc_put_rtab(R_tab);
 253         if (ret == ACT_P_CREATED)
 254                 kfree(police);
 255         return err;
 256 }
 257
 258 static int tcf_act_police_cleanup(struct tc_action *a, int bind)
 259 {
 260         struct tcf_police *p = a->priv;
 261         if (p)
 262                 return tcf_hash_release(&p->common, bind, &police_hash_info);
 263         return 0;
 264 }
 265
 266 static int tcf_act_police(struct sk_buff *skb, const struct tc_action *a,
 267                           struct tcf_result *res)
 268 {
 269         struct tcf_police *police = a->priv;
 270         s64 now;
 271         s64 toks;
 272         s64 ptoks = 0;
 273
 274         spin_lock(&police->tcf_lock);
 275
 276         bstats_update(&police->tcf_bstats, skb);
 277
 278         if (police->tcfp_ewma_rate &&
 279             police->tcf_rate_est.bps >= police->tcfp_ewma_rate) {
 280                 police->tcf_qstats.overlimits++;
 281                 if (police->tcf_action == TC_ACT_SHOT)
 282                         police->tcf_qstats.drops++;
 283                 spin_unlock(&police->tcf_lock);
 284                 return police->tcf_action;
 285         }
 286
 287         if (qdisc_pkt_len(skb) <= police->tcfp_mtu) {
 288                 if (!police->rate_present) {
 289                         spin_unlock(&police->tcf_lock);
 290                         return police->tcfp_result;
 291                 }
 292
 293                 now = ktime_to_ns(ktime_get());
 294                 toks = min_t(s64, now - police->tcfp_t_c,
 295                              police->tcfp_burst);
 296                 if (police->peak_present) {
 297                         ptoks = toks + police->tcfp_ptoks;
 298                         if (ptoks > police->tcfp_mtu_ptoks)
 299                                 ptoks = police->tcfp_mtu_ptoks;
 300                         ptoks -= (s64) psched_l2t_ns(&police->peak,
 301                                                      qdisc_pkt_len(skb));
 302                 }
 303                 toks += police->tcfp_toks;
 304                 if (toks > police->tcfp_burst)
 305                         toks = police->tcfp_burst;
 306                 toks -= (s64) psched_l2t_ns(&police->rate, qdisc_pkt_len(skb));
 307                 if ((toks|ptoks) >= 0) {
 308                         police->tcfp_t_c = now;
 309                         police->tcfp_toks = toks;
 310                         police->tcfp_ptoks = ptoks;
 311                         spin_unlock(&police->tcf_lock);
 312                         return police->tcfp_result;
 313                 }
 314         }
 315
 316         police->tcf_qstats.overlimits++;
 317         if (police->tcf_action == TC_ACT_SHOT)
 318                 police->tcf_qstats.drops++;
 319         spin_unlock(&police->tcf_lock);
 320         return police->tcf_action;
 321 }
 322
 323 static int
 324 tcf_act_police_dump(struct sk_buff *skb, struct tc_action *a, int bind, int ref)
 325 {
 326         unsigned char *b = skb_tail_pointer(skb);
 327         struct tcf_police *police = a->priv;
 328         struct tc_police opt = {
 329                 .index = police->tcf_index,
 330                 .action = police->tcf_action,
 331                 .mtu = police->tcfp_mtu,
 332                 .burst = PSCHED_NS2TICKS(police->tcfp_burst),
 333                 .refcnt = police->tcf_refcnt - ref,
 334                 .bindcnt = police->tcf_bindcnt - bind,
 335         };
 336
 337         if (police->rate_present)
 338                 psched_ratecfg_getrate(&opt.rate, &police->rate);
 339         if (police->peak_present)
 340                 psched_ratecfg_getrate(&opt.peakrate, &police->peak);
 341         if (nla_put(skb, TCA_POLICE_TBF, sizeof(opt), &opt))
 342                 goto nla_put_failure;
 343         if (police->tcfp_result &&
 344             nla_put_u32(skb, TCA_POLICE_RESULT, police->tcfp_result))
 345                 goto nla_put_failure;
 346         if (police->tcfp_ewma_rate &&
 347             nla_put_u32(skb, TCA_POLICE_AVRATE, police->tcfp_ewma_rate))
 348                 goto nla_put_failure;
 349         return skb->len;
 350
 351 nla_put_failure:
 352         nlmsg_trim(skb, b);
 353         return -1;
 354 }
 355
 356 MODULE_AUTHOR("Alexey Kuznetsov");
 357 MODULE_DESCRIPTION("Policing actions");
 358 MODULE_LICENSE("GPL");
 359
 360 static struct tc_action_ops act_police_ops = {
 361         .kind           =       "police",
 362         .hinfo          =       &police_hash_info,
 363         .type           =       TCA_ID_POLICE,
 364         .capab          =       TCA_CAP_NONE,
 365         .owner          =       THIS_MODULE,
 366         .act            =       tcf_act_police,
 367         .dump           =       tcf_act_police_dump,
 368         .cleanup        =       tcf_act_police_cleanup,
 369         .init           =       tcf_act_police_locate,
 370         .walk           =       tcf_act_police_walker
 371 };
 372
 373 static int __init
 374 police_init_module(void)
 375 {
 376         int err = tcf_hashinfo_init(&police_hash_info, POL_TAB_MASK);
 377         if (err)
 378                 return err;
 379         err = tcf_register_action(&act_police_ops);
 380         if (err)
 381                 tcf_hashinfo_destroy(&police_hash_info);
 382         return err;
 383 }
 384
 385 static void __exit
 386 police_cleanup_module(void)
 387 {
 388         tcf_hashinfo_destroy(&police_hash_info);
 389         tcf_unregister_action(&act_police_ops);
 390 }
 391
 392 module_init(police_init_module);
 393 module_exit(police_cleanup_module);