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 tcf_hashinfo *hinfo = a->ops->hinfo;
  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(&hinfo->lock);
  69
  70         s_i = cb->args[0];
  71
  72         for (i = 0; i < (POL_TAB_MASK + 1); i++) {
  73                 head = &hinfo->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(&hinfo->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 const struct nla_policy police_policy[TCA_POLICE_MAX + 1] = {
 109         [TCA_POLICE_RATE]       = { .len = TC_RTAB_SIZE },
 110         [TCA_POLICE_PEAKRATE]   = { .len = TC_RTAB_SIZE },
 111         [TCA_POLICE_AVRATE]     = { .type = NLA_U32 },
 112         [TCA_POLICE_RESULT]     = { .type = NLA_U32 },
 113 };
 114
 115 static int tcf_act_police_locate(struct net *net, struct nlattr *nla,
 116                                  struct nlattr *est, struct tc_action *a,
 117                                  int ovr, int bind)
 118 {
 119         unsigned int h;
 120         int ret = 0, err;
 121         struct nlattr *tb[TCA_POLICE_MAX + 1];
 122         struct tc_police *parm;
 123         struct tcf_police *police;
 124         struct qdisc_rate_table *R_tab = NULL, *P_tab = NULL;
 125         struct tcf_hashinfo *hinfo = a->ops->hinfo;
 126         int size;
 127
 128         if (nla == NULL)
 129                 return -EINVAL;
 130
 131         err = nla_parse_nested(tb, TCA_POLICE_MAX, nla, police_policy);
 132         if (err < 0)
 133                 return err;
 134
 135         if (tb[TCA_POLICE_TBF] == NULL)
 136                 return -EINVAL;
 137         size = nla_len(tb[TCA_POLICE_TBF]);
 138         if (size != sizeof(*parm) && size != sizeof(struct tc_police_compat))
 139                 return -EINVAL;
 140         parm = nla_data(tb[TCA_POLICE_TBF]);
 141
 142         if (parm->index) {
 143                 struct tcf_common *pc;
 144
 145                 pc = tcf_hash_lookup(parm->index, hinfo);
 146                 if (pc != NULL) {
 147                         a->priv = pc;
 148                         police = to_police(pc);
 149                         if (bind) {
 150                                 police->tcf_bindcnt += 1;
 151                                 police->tcf_refcnt += 1;
 152                                 return 0;
 153                         }
 154                         if (ovr)
 155                                 goto override;
 156                         /* not replacing */
 157                         return -EEXIST;
 158                 }
 159         }
 160
 161         police = kzalloc(sizeof(*police), GFP_KERNEL);
 162         if (police == NULL)
 163                 return -ENOMEM;
 164         ret = ACT_P_CREATED;
 165         police->tcf_refcnt = 1;
 166         spin_lock_init(&police->tcf_lock);
 167         if (bind)
 168                 police->tcf_bindcnt = 1;
 169 override:
 170         if (parm->rate.rate) {
 171                 err = -ENOMEM;
 172                 R_tab = qdisc_get_rtab(&parm->rate, tb[TCA_POLICE_RATE]);
 173                 if (R_tab == NULL)
 174                         goto failure;
 175
 176                 if (parm->peakrate.rate) {
 177                         P_tab = qdisc_get_rtab(&parm->peakrate,
 178                                                tb[TCA_POLICE_PEAKRATE]);
 179                         if (P_tab == NULL)
 180                                 goto failure;
 181                 }
 182         }
 183
 184         spin_lock_bh(&police->tcf_lock);
 185         if (est) {
 186                 err = gen_replace_estimator(&police->tcf_bstats,
 187                                             &police->tcf_rate_est,
 188                                             &police->tcf_lock, est);
 189                 if (err)
 190                         goto failure_unlock;
 191         } else if (tb[TCA_POLICE_AVRATE] &&
 192                    (ret == ACT_P_CREATED ||
 193                     !gen_estimator_active(&police->tcf_bstats,
 194                                           &police->tcf_rate_est))) {
 195                 err = -EINVAL;
 196                 goto failure_unlock;
 197         }
 198
 199         /* No failure allowed after this point */
 200         police->tcfp_mtu = parm->mtu;
 201         if (police->tcfp_mtu == 0) {
 202                 police->tcfp_mtu = ~0;
 203                 if (R_tab)
 204                         police->tcfp_mtu = 255 << R_tab->rate.cell_log;
 205         }
 206         if (R_tab) {
 207                 police->rate_present = true;
 208                 psched_ratecfg_precompute(&police->rate, &R_tab->rate, 0);
 209                 qdisc_put_rtab(R_tab);
 210         } else {
 211                 police->rate_present = false;
 212         }
 213         if (P_tab) {
 214                 police->peak_present = true;
 215                 psched_ratecfg_precompute(&police->peak, &P_tab->rate, 0);
 216                 qdisc_put_rtab(P_tab);
 217         } else {
 218                 police->peak_present = false;
 219         }
 220
 221         if (tb[TCA_POLICE_RESULT])
 222                 police->tcfp_result = nla_get_u32(tb[TCA_POLICE_RESULT]);
 223         police->tcfp_burst = PSCHED_TICKS2NS(parm->burst);
 224         police->tcfp_toks = police->tcfp_burst;
 225         if (police->peak_present) {
 226                 police->tcfp_mtu_ptoks = (s64) psched_l2t_ns(&police->peak,
 227                                                              police->tcfp_mtu);
 228                 police->tcfp_ptoks = police->tcfp_mtu_ptoks;
 229         }
 230         police->tcf_action = parm->action;
 231
 232         if (tb[TCA_POLICE_AVRATE])
 233                 police->tcfp_ewma_rate = nla_get_u32(tb[TCA_POLICE_AVRATE]);
 234
 235         spin_unlock_bh(&police->tcf_lock);
 236         if (ret != ACT_P_CREATED)
 237                 return ret;
 238
 239         police->tcfp_t_c = ktime_to_ns(ktime_get());
 240         police->tcf_index = parm->index ? parm->index :
 241                 tcf_hash_new_index(a->ops->hinfo);
 242         h = tcf_hash(police->tcf_index, POL_TAB_MASK);
 243         spin_lock_bh(&hinfo->lock);
 244         hlist_add_head(&police->tcf_head, &hinfo->htab[h]);
 245         spin_unlock_bh(&hinfo->lock);
 246
 247         a->priv = police;
 248         return ret;
 249
 250 failure_unlock:
 251         spin_unlock_bh(&police->tcf_lock);
 252 failure:
 253         qdisc_put_rtab(P_tab);
 254         qdisc_put_rtab(R_tab);
 255         if (ret == ACT_P_CREATED)
 256                 kfree(police);
 257         return err;
 258 }
 259
 260 static int tcf_act_police_cleanup(struct tc_action *a, int bind)
 261 {
 262         struct tcf_police *p = a->priv;
 263         if (p)
 264                 return tcf_hash_release(&p->common, bind, &police_hash_info);
 265         return 0;
 266 }
 267
 268 static int tcf_act_police(struct sk_buff *skb, const struct tc_action *a,
 269                           struct tcf_result *res)
 270 {
 271         struct tcf_police *police = a->priv;
 272         s64 now;
 273         s64 toks;
 274         s64 ptoks = 0;
 275
 276         spin_lock(&police->tcf_lock);
 277
 278         bstats_update(&police->tcf_bstats, skb);
 279
 280         if (police->tcfp_ewma_rate &&
 281             police->tcf_rate_est.bps >= police->tcfp_ewma_rate) {
 282                 police->tcf_qstats.overlimits++;
 283                 if (police->tcf_action == TC_ACT_SHOT)
 284                         police->tcf_qstats.drops++;
 285                 spin_unlock(&police->tcf_lock);
 286                 return police->tcf_action;
 287         }
 288
 289         if (qdisc_pkt_len(skb) <= police->tcfp_mtu) {
 290                 if (!police->rate_present) {
 291                         spin_unlock(&police->tcf_lock);
 292                         return police->tcfp_result;
 293                 }
 294
 295                 now = ktime_to_ns(ktime_get());
 296                 toks = min_t(s64, now - police->tcfp_t_c,
 297                              police->tcfp_burst);
 298                 if (police->peak_present) {
 299                         ptoks = toks + police->tcfp_ptoks;
 300                         if (ptoks > police->tcfp_mtu_ptoks)
 301                                 ptoks = police->tcfp_mtu_ptoks;
 302                         ptoks -= (s64) psched_l2t_ns(&police->peak,
 303                                                      qdisc_pkt_len(skb));
 304                 }
 305                 toks += police->tcfp_toks;
 306                 if (toks > police->tcfp_burst)
 307                         toks = police->tcfp_burst;
 308                 toks -= (s64) psched_l2t_ns(&police->rate, qdisc_pkt_len(skb));
 309                 if ((toks|ptoks) >= 0) {
 310                         police->tcfp_t_c = now;
 311                         police->tcfp_toks = toks;
 312                         police->tcfp_ptoks = ptoks;
 313                         spin_unlock(&police->tcf_lock);
 314                         return police->tcfp_result;
 315                 }
 316         }
 317
 318         police->tcf_qstats.overlimits++;
 319         if (police->tcf_action == TC_ACT_SHOT)
 320                 police->tcf_qstats.drops++;
 321         spin_unlock(&police->tcf_lock);
 322         return police->tcf_action;
 323 }
 324
 325 static int
 326 tcf_act_police_dump(struct sk_buff *skb, struct tc_action *a, int bind, int ref)
 327 {
 328         unsigned char *b = skb_tail_pointer(skb);
 329         struct tcf_police *police = a->priv;
 330         struct tc_police opt = {
 331                 .index = police->tcf_index,
 332                 .action = police->tcf_action,
 333                 .mtu = police->tcfp_mtu,
 334                 .burst = PSCHED_NS2TICKS(police->tcfp_burst),
 335                 .refcnt = police->tcf_refcnt - ref,
 336                 .bindcnt = police->tcf_bindcnt - bind,
 337         };
 338
 339         if (police->rate_present)
 340                 psched_ratecfg_getrate(&opt.rate, &police->rate);
 341         if (police->peak_present)
 342                 psched_ratecfg_getrate(&opt.peakrate, &police->peak);
 343         if (nla_put(skb, TCA_POLICE_TBF, sizeof(opt), &opt))
 344                 goto nla_put_failure;
 345         if (police->tcfp_result &&
 346             nla_put_u32(skb, TCA_POLICE_RESULT, police->tcfp_result))
 347                 goto nla_put_failure;
 348         if (police->tcfp_ewma_rate &&
 349             nla_put_u32(skb, TCA_POLICE_AVRATE, police->tcfp_ewma_rate))
 350                 goto nla_put_failure;
 351         return skb->len;
 352
 353 nla_put_failure:
 354         nlmsg_trim(skb, b);
 355         return -1;
 356 }
 357
 358 MODULE_AUTHOR("Alexey Kuznetsov");
 359 MODULE_DESCRIPTION("Policing actions");
 360 MODULE_LICENSE("GPL");
 361
 362 static struct tc_action_ops act_police_ops = {
 363         .kind           =       "police",
 364         .hinfo          =       &police_hash_info,
 365         .type           =       TCA_ID_POLICE,
 366         .owner          =       THIS_MODULE,
 367         .act            =       tcf_act_police,
 368         .dump           =       tcf_act_police_dump,
 369         .cleanup        =       tcf_act_police_cleanup,
 370         .init           =       tcf_act_police_locate,
 371         .walk           =       tcf_act_police_walker
 372 };
 373
 374 static int __init
 375 police_init_module(void)
 376 {
 377         int err = tcf_hashinfo_init(&police_hash_info, POL_TAB_MASK);
 378         if (err)
 379                 return err;
 380         err = tcf_register_action(&act_police_ops);
 381         if (err)
 382                 tcf_hashinfo_destroy(&police_hash_info);
 383         return err;
 384 }
 385
 386 static void __exit
 387 police_cleanup_module(void)
 388 {
 389         tcf_hashinfo_destroy(&police_hash_info);
 390         tcf_unregister_action(&act_police_ops);
 391 }
 392
 393 module_init(police_init_module);
 394 module_exit(police_cleanup_module);