net/sched/act_police.c

   1 /*
   2  * net/sched/act_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->priv, struct tcf_police, common)
  42
  43 #define POL_TAB_MASK     15
  44
  45 /* old policer structure from before tc actions */
  46 struct tc_police_compat {
  47         u32                     index;
  48         int                     action;
  49         u32                     limit;
  50         u32                     burst;
  51         u32                     mtu;
  52         struct tc_ratespec      rate;
  53         struct tc_ratespec      peakrate;
  54 };
  55
  56 /* Each policer is serialized by its individual spinlock */
  57
  58 static int police_net_id;
  59
  60 static int tcf_act_police_walker(struct net *net, struct sk_buff *skb,
  61                                  struct netlink_callback *cb, int type,
  62                                  struct tc_action *a)
  63 {
  64         struct tc_action_net *tn = net_generic(net, police_net_id);
  65         struct tcf_hashinfo *hinfo = tn->hinfo;
  66         struct hlist_head *head;
  67         struct tcf_common *p;
  68         int err = 0, index = -1, i = 0, s_i = 0, n_i = 0;
  69         struct nlattr *nest;
  70
  71         spin_lock_bh(&hinfo->lock);
  72
  73         s_i = cb->args[0];
  74
  75         for (i = 0; i < (POL_TAB_MASK + 1); i++) {
  76                 head = &hinfo->htab[tcf_hash(i, POL_TAB_MASK)];
  77
  78                 hlist_for_each_entry_rcu(p, head, tcfc_head) {
  79                         index++;
  80                         if (index < s_i)
  81                                 continue;
  82                         a->priv = p;
  83                         a->order = index;
  84                         nest = nla_nest_start(skb, a->order);
  85                         if (nest == NULL)
  86                                 goto nla_put_failure;
  87                         if (type == RTM_DELACTION)
  88                                 err = tcf_action_dump_1(skb, a, 0, 1);
  89                         else
  90                                 err = tcf_action_dump_1(skb, a, 0, 0);
  91                         if (err < 0) {
  92                                 index--;
  93                                 nla_nest_cancel(skb, nest);
  94                                 goto done;
  95                         }
  96                         nla_nest_end(skb, nest);
  97                         n_i++;
  98                 }
  99         }
 100 done:
 101         spin_unlock_bh(&hinfo->lock);
 102         if (n_i)
 103                 cb->args[0] += n_i;
 104         return n_i;
 105
 106 nla_put_failure:
 107         nla_nest_cancel(skb, nest);
 108         goto done;
 109 }
 110
 111 static const struct nla_policy police_policy[TCA_POLICE_MAX + 1] = {
 112         [TCA_POLICE_RATE]       = { .len = TC_RTAB_SIZE },
 113         [TCA_POLICE_PEAKRATE]   = { .len = TC_RTAB_SIZE },
 114         [TCA_POLICE_AVRATE]     = { .type = NLA_U32 },
 115         [TCA_POLICE_RESULT]     = { .type = NLA_U32 },
 116 };
 117
 118 static int tcf_act_police_init(struct net *net, struct nlattr *nla,
 119                                struct nlattr *est, struct tc_action *a,
 120                                int ovr, int bind)
 121 {
 122         int ret = 0, err;
 123         struct nlattr *tb[TCA_POLICE_MAX + 1];
 124         struct tc_police *parm;
 125         struct tcf_police *police;
 126         struct qdisc_rate_table *R_tab = NULL, *P_tab = NULL;
 127         struct tc_action_net *tn = net_generic(net, police_net_id);
 128         int size;
 129
 130         if (nla == NULL)
 131                 return -EINVAL;
 132
 133         err = nla_parse_nested(tb, TCA_POLICE_MAX, nla, police_policy);
 134         if (err < 0)
 135                 return err;
 136
 137         if (tb[TCA_POLICE_TBF] == NULL)
 138                 return -EINVAL;
 139         size = nla_len(tb[TCA_POLICE_TBF]);
 140         if (size != sizeof(*parm) && size != sizeof(struct tc_police_compat))
 141                 return -EINVAL;
 142         parm = nla_data(tb[TCA_POLICE_TBF]);
 143
 144         if (parm->index) {
 145                 if (tcf_hash_search(tn, a, parm->index)) {
 146                         police = to_police(a);
 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         } else {
 158                 ret = tcf_hash_create(tn, parm->index, NULL, a,
 159                                       sizeof(*police), bind, false);
 160                 if (ret)
 161                         return ret;
 162                 ret = ACT_P_CREATED;
 163         }
 164
 165         police = to_police(a);
 166 override:
 167         if (parm->rate.rate) {
 168                 err = -ENOMEM;
 169                 R_tab = qdisc_get_rtab(&parm->rate, tb[TCA_POLICE_RATE]);
 170                 if (R_tab == NULL)
 171                         goto failure;
 172
 173                 if (parm->peakrate.rate) {
 174                         P_tab = qdisc_get_rtab(&parm->peakrate,
 175                                                tb[TCA_POLICE_PEAKRATE]);
 176                         if (P_tab == NULL)
 177                                 goto failure;
 178                 }
 179         }
 180
 181         spin_lock_bh(&police->tcf_lock);
 182         if (est) {
 183                 err = gen_replace_estimator(&police->tcf_bstats, NULL,
 184                                             &police->tcf_rate_est,
 185                                             &police->tcf_lock,
 186                                             NULL, 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_get_ns();
 238         tcf_hash_insert(tn, a);
 239
 240         return ret;
 241
 242 failure_unlock:
 243         spin_unlock_bh(&police->tcf_lock);
 244 failure:
 245         qdisc_put_rtab(P_tab);
 246         qdisc_put_rtab(R_tab);
 247         if (ret == ACT_P_CREATED)
 248                 tcf_hash_cleanup(a, est);
 249         return err;
 250 }
 251
 252 static int tcf_act_police(struct sk_buff *skb, const struct tc_action *a,
 253                           struct tcf_result *res)
 254 {
 255         struct tcf_police *police = a->priv;
 256         s64 now;
 257         s64 toks;
 258         s64 ptoks = 0;
 259
 260         spin_lock(&police->tcf_lock);
 261
 262         bstats_update(&police->tcf_bstats, skb);
 263         tcf_lastuse_update(&police->tcf_tm);
 264
 265         if (police->tcfp_ewma_rate &&
 266             police->tcf_rate_est.bps >= police->tcfp_ewma_rate) {
 267                 police->tcf_qstats.overlimits++;
 268                 if (police->tcf_action == TC_ACT_SHOT)
 269                         police->tcf_qstats.drops++;
 270                 spin_unlock(&police->tcf_lock);
 271                 return police->tcf_action;
 272         }
 273
 274         if (qdisc_pkt_len(skb) <= police->tcfp_mtu) {
 275                 if (!police->rate_present) {
 276                         spin_unlock(&police->tcf_lock);
 277                         return police->tcfp_result;
 278                 }
 279
 280                 now = ktime_get_ns();
 281                 toks = min_t(s64, now - police->tcfp_t_c,
 282                              police->tcfp_burst);
 283                 if (police->peak_present) {
 284                         ptoks = toks + police->tcfp_ptoks;
 285                         if (ptoks > police->tcfp_mtu_ptoks)
 286                                 ptoks = police->tcfp_mtu_ptoks;
 287                         ptoks -= (s64) psched_l2t_ns(&police->peak,
 288                                                      qdisc_pkt_len(skb));
 289                 }
 290                 toks += police->tcfp_toks;
 291                 if (toks > police->tcfp_burst)
 292                         toks = police->tcfp_burst;
 293                 toks -= (s64) psched_l2t_ns(&police->rate, qdisc_pkt_len(skb));
 294                 if ((toks|ptoks) >= 0) {
 295                         police->tcfp_t_c = now;
 296                         police->tcfp_toks = toks;
 297                         police->tcfp_ptoks = ptoks;
 298                         spin_unlock(&police->tcf_lock);
 299                         return police->tcfp_result;
 300                 }
 301         }
 302
 303         police->tcf_qstats.overlimits++;
 304         if (police->tcf_action == TC_ACT_SHOT)
 305                 police->tcf_qstats.drops++;
 306         spin_unlock(&police->tcf_lock);
 307         return police->tcf_action;
 308 }
 309
 310 static int
 311 tcf_act_police_dump(struct sk_buff *skb, struct tc_action *a, int bind, int ref)
 312 {
 313         unsigned char *b = skb_tail_pointer(skb);
 314         struct tcf_police *police = a->priv;
 315         struct tc_police opt = {
 316                 .index = police->tcf_index,
 317                 .action = police->tcf_action,
 318                 .mtu = police->tcfp_mtu,
 319                 .burst = PSCHED_NS2TICKS(police->tcfp_burst),
 320                 .refcnt = police->tcf_refcnt - ref,
 321                 .bindcnt = police->tcf_bindcnt - bind,
 322         };
 323         struct tcf_t t;
 324
 325         if (police->rate_present)
 326                 psched_ratecfg_getrate(&opt.rate, &police->rate);
 327         if (police->peak_present)
 328                 psched_ratecfg_getrate(&opt.peakrate, &police->peak);
 329         if (nla_put(skb, TCA_POLICE_TBF, sizeof(opt), &opt))
 330                 goto nla_put_failure;
 331         if (police->tcfp_result &&
 332             nla_put_u32(skb, TCA_POLICE_RESULT, police->tcfp_result))
 333                 goto nla_put_failure;
 334         if (police->tcfp_ewma_rate &&
 335             nla_put_u32(skb, TCA_POLICE_AVRATE, police->tcfp_ewma_rate))
 336                 goto nla_put_failure;
 337
 338         t.install = jiffies_to_clock_t(jiffies - police->tcf_tm.install);
 339         t.lastuse = jiffies_to_clock_t(jiffies - police->tcf_tm.lastuse);
 340         t.firstuse = jiffies_to_clock_t(jiffies - police->tcf_tm.firstuse);
 341         t.expires = jiffies_to_clock_t(police->tcf_tm.expires);
 342         if (nla_put_64bit(skb, TCA_POLICE_TM, sizeof(t), &t, TCA_POLICE_PAD))
 343                 goto nla_put_failure;
 344
 345         return skb->len;
 346
 347 nla_put_failure:
 348         nlmsg_trim(skb, b);
 349         return -1;
 350 }
 351
 352 static int tcf_police_search(struct net *net, struct tc_action *a, u32 index)
 353 {
 354         struct tc_action_net *tn = net_generic(net, police_net_id);
 355
 356         return tcf_hash_search(tn, a, index);
 357 }
 358
 359 MODULE_AUTHOR("Alexey Kuznetsov");
 360 MODULE_DESCRIPTION("Policing actions");
 361 MODULE_LICENSE("GPL");
 362
 363 static struct tc_action_ops act_police_ops = {
 364         .kind           =       "police",
 365         .type           =       TCA_ID_POLICE,
 366         .owner          =       THIS_MODULE,
 367         .act            =       tcf_act_police,
 368         .dump           =       tcf_act_police_dump,
 369         .init           =       tcf_act_police_init,
 370         .walk           =       tcf_act_police_walker,
 371         .lookup         =       tcf_police_search,
 372 };
 373
 374 static __net_init int police_init_net(struct net *net)
 375 {
 376         struct tc_action_net *tn = net_generic(net, police_net_id);
 377
 378         return tc_action_net_init(tn, &act_police_ops, POL_TAB_MASK);
 379 }
 380
 381 static void __net_exit police_exit_net(struct net *net)
 382 {
 383         struct tc_action_net *tn = net_generic(net, police_net_id);
 384
 385         tc_action_net_exit(tn);
 386 }
 387
 388 static struct pernet_operations police_net_ops = {
 389         .init = police_init_net,
 390         .exit = police_exit_net,
 391         .id   = &police_net_id,
 392         .size = sizeof(struct tc_action_net),
 393 };
 394
 395 static int __init
 396 police_init_module(void)
 397 {
 398         return tcf_register_action(&act_police_ops, &police_net_ops);
 399 }
 400
 401 static void __exit
 402 police_cleanup_module(void)
 403 {
 404         tcf_unregister_action(&act_police_ops, &police_net_ops);
 405 }
 406
 407 module_init(police_init_module);
 408 module_exit(police_cleanup_module);