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 tc_action        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
  41 #define to_police(pc) ((struct tcf_police *)pc)
  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 static struct tc_action_ops act_police_ops;
  60
  61 static int tcf_act_police_walker(struct net *net, struct sk_buff *skb,
  62                                  struct netlink_callback *cb, int type,
  63                                  const struct tc_action_ops *ops)
  64 {
  65         struct tc_action_net *tn = net_generic(net, police_net_id);
  66         struct tcf_hashinfo *hinfo = tn->hinfo;
  67         int err = 0, index = -1, i = 0, s_i = 0, n_i = 0;
  68         struct nlattr *nest;
  69
  70         spin_lock_bh(&hinfo->lock);
  71
  72         s_i = cb->args[0];
  73
  74         for (i = 0; i < (POL_TAB_MASK + 1); i++) {
  75                 struct hlist_head *head;
  76                 struct tc_action *p;
  77
  78                 head = &hinfo->htab[tcf_hash(i, POL_TAB_MASK)];
  79
  80                 hlist_for_each_entry_rcu(p, head, tcfa_head) {
  81                         index++;
  82                         if (index < s_i)
  83                                 continue;
  84                         nest = nla_nest_start(skb, index);
  85                         if (nest == NULL)
  86                                 goto nla_put_failure;
  87                         if (type == RTM_DELACTION)
  88                                 err = tcf_action_dump_1(skb, p, 0, 1);
  89                         else
  90                                 err = tcf_action_dump_1(skb, p, 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_check(tn, parm->index, a, bind)) {
 146                         if (ovr)
 147                                 goto override;
 148                         /* not replacing */
 149                         return -EEXIST;
 150                 }
 151         } else {
 152                 ret = tcf_hash_create(tn, parm->index, NULL, a,
 153                                       &act_police_ops, bind, false);
 154                 if (ret)
 155                         return ret;
 156                 ret = ACT_P_CREATED;
 157         }
 158
 159 override:
 160         police = to_police(*a);
 161         if (parm->rate.rate) {
 162                 err = -ENOMEM;
 163                 R_tab = qdisc_get_rtab(&parm->rate, tb[TCA_POLICE_RATE]);
 164                 if (R_tab == NULL)
 165                         goto failure;
 166
 167                 if (parm->peakrate.rate) {
 168                         P_tab = qdisc_get_rtab(&parm->peakrate,
 169                                                tb[TCA_POLICE_PEAKRATE]);
 170                         if (P_tab == NULL)
 171                                 goto failure;
 172                 }
 173         }
 174
 175         spin_lock_bh(&police->tcf_lock);
 176         if (est) {
 177                 err = gen_replace_estimator(&police->tcf_bstats, NULL,
 178                                             &police->tcf_rate_est,
 179                                             &police->tcf_lock,
 180                                             NULL, est);
 181                 if (err)
 182                         goto failure_unlock;
 183         } else if (tb[TCA_POLICE_AVRATE] &&
 184                    (ret == ACT_P_CREATED ||
 185                     !gen_estimator_active(&police->tcf_bstats,
 186                                           &police->tcf_rate_est))) {
 187                 err = -EINVAL;
 188                 goto failure_unlock;
 189         }
 190
 191         /* No failure allowed after this point */
 192         police->tcfp_mtu = parm->mtu;
 193         if (police->tcfp_mtu == 0) {
 194                 police->tcfp_mtu = ~0;
 195                 if (R_tab)
 196                         police->tcfp_mtu = 255 << R_tab->rate.cell_log;
 197         }
 198         if (R_tab) {
 199                 police->rate_present = true;
 200                 psched_ratecfg_precompute(&police->rate, &R_tab->rate, 0);
 201                 qdisc_put_rtab(R_tab);
 202         } else {
 203                 police->rate_present = false;
 204         }
 205         if (P_tab) {
 206                 police->peak_present = true;
 207                 psched_ratecfg_precompute(&police->peak, &P_tab->rate, 0);
 208                 qdisc_put_rtab(P_tab);
 209         } else {
 210                 police->peak_present = false;
 211         }
 212
 213         if (tb[TCA_POLICE_RESULT])
 214                 police->tcfp_result = nla_get_u32(tb[TCA_POLICE_RESULT]);
 215         police->tcfp_burst = PSCHED_TICKS2NS(parm->burst);
 216         police->tcfp_toks = police->tcfp_burst;
 217         if (police->peak_present) {
 218                 police->tcfp_mtu_ptoks = (s64) psched_l2t_ns(&police->peak,
 219                                                              police->tcfp_mtu);
 220                 police->tcfp_ptoks = police->tcfp_mtu_ptoks;
 221         }
 222         police->tcf_action = parm->action;
 223
 224         if (tb[TCA_POLICE_AVRATE])
 225                 police->tcfp_ewma_rate = nla_get_u32(tb[TCA_POLICE_AVRATE]);
 226
 227         spin_unlock_bh(&police->tcf_lock);
 228         if (ret != ACT_P_CREATED)
 229                 return ret;
 230
 231         police->tcfp_t_c = ktime_get_ns();
 232         tcf_hash_insert(tn, *a);
 233
 234         return ret;
 235
 236 failure_unlock:
 237         spin_unlock_bh(&police->tcf_lock);
 238 failure:
 239         qdisc_put_rtab(P_tab);
 240         qdisc_put_rtab(R_tab);
 241         if (ret == ACT_P_CREATED)
 242                 tcf_hash_cleanup(*a, est);
 243         return err;
 244 }
 245
 246 static int tcf_act_police(struct sk_buff *skb, const struct tc_action *a,
 247                           struct tcf_result *res)
 248 {
 249         struct tcf_police *police = to_police(a);
 250         s64 now;
 251         s64 toks;
 252         s64 ptoks = 0;
 253
 254         spin_lock(&police->tcf_lock);
 255
 256         bstats_update(&police->tcf_bstats, skb);
 257         tcf_lastuse_update(&police->tcf_tm);
 258
 259         if (police->tcfp_ewma_rate &&
 260             police->tcf_rate_est.bps >= police->tcfp_ewma_rate) {
 261                 police->tcf_qstats.overlimits++;
 262                 if (police->tcf_action == TC_ACT_SHOT)
 263                         police->tcf_qstats.drops++;
 264                 spin_unlock(&police->tcf_lock);
 265                 return police->tcf_action;
 266         }
 267
 268         if (qdisc_pkt_len(skb) <= police->tcfp_mtu) {
 269                 if (!police->rate_present) {
 270                         spin_unlock(&police->tcf_lock);
 271                         return police->tcfp_result;
 272                 }
 273
 274                 now = ktime_get_ns();
 275                 toks = min_t(s64, now - police->tcfp_t_c,
 276                              police->tcfp_burst);
 277                 if (police->peak_present) {
 278                         ptoks = toks + police->tcfp_ptoks;
 279                         if (ptoks > police->tcfp_mtu_ptoks)
 280                                 ptoks = police->tcfp_mtu_ptoks;
 281                         ptoks -= (s64) psched_l2t_ns(&police->peak,
 282                                                      qdisc_pkt_len(skb));
 283                 }
 284                 toks += police->tcfp_toks;
 285                 if (toks > police->tcfp_burst)
 286                         toks = police->tcfp_burst;
 287                 toks -= (s64) psched_l2t_ns(&police->rate, qdisc_pkt_len(skb));
 288                 if ((toks|ptoks) >= 0) {
 289                         police->tcfp_t_c = now;
 290                         police->tcfp_toks = toks;
 291                         police->tcfp_ptoks = ptoks;
 292                         spin_unlock(&police->tcf_lock);
 293                         return police->tcfp_result;
 294                 }
 295         }
 296
 297         police->tcf_qstats.overlimits++;
 298         if (police->tcf_action == TC_ACT_SHOT)
 299                 police->tcf_qstats.drops++;
 300         spin_unlock(&police->tcf_lock);
 301         return police->tcf_action;
 302 }
 303
 304 static int
 305 tcf_act_police_dump(struct sk_buff *skb, struct tc_action *a, int bind, int ref)
 306 {
 307         unsigned char *b = skb_tail_pointer(skb);
 308         struct tcf_police *police = to_police(a);
 309         struct tc_police opt = {
 310                 .index = police->tcf_index,
 311                 .action = police->tcf_action,
 312                 .mtu = police->tcfp_mtu,
 313                 .burst = PSCHED_NS2TICKS(police->tcfp_burst),
 314                 .refcnt = police->tcf_refcnt - ref,
 315                 .bindcnt = police->tcf_bindcnt - bind,
 316         };
 317         struct tcf_t t;
 318
 319         if (police->rate_present)
 320                 psched_ratecfg_getrate(&opt.rate, &police->rate);
 321         if (police->peak_present)
 322                 psched_ratecfg_getrate(&opt.peakrate, &police->peak);
 323         if (nla_put(skb, TCA_POLICE_TBF, sizeof(opt), &opt))
 324                 goto nla_put_failure;
 325         if (police->tcfp_result &&
 326             nla_put_u32(skb, TCA_POLICE_RESULT, police->tcfp_result))
 327                 goto nla_put_failure;
 328         if (police->tcfp_ewma_rate &&
 329             nla_put_u32(skb, TCA_POLICE_AVRATE, police->tcfp_ewma_rate))
 330                 goto nla_put_failure;
 331
 332         t.install = jiffies_to_clock_t(jiffies - police->tcf_tm.install);
 333         t.lastuse = jiffies_to_clock_t(jiffies - police->tcf_tm.lastuse);
 334         t.firstuse = jiffies_to_clock_t(jiffies - police->tcf_tm.firstuse);
 335         t.expires = jiffies_to_clock_t(police->tcf_tm.expires);
 336         if (nla_put_64bit(skb, TCA_POLICE_TM, sizeof(t), &t, TCA_POLICE_PAD))
 337                 goto nla_put_failure;
 338
 339         return skb->len;
 340
 341 nla_put_failure:
 342         nlmsg_trim(skb, b);
 343         return -1;
 344 }
 345
 346 static int tcf_police_search(struct net *net, struct tc_action **a, u32 index)
 347 {
 348         struct tc_action_net *tn = net_generic(net, police_net_id);
 349
 350         return tcf_hash_search(tn, a, index);
 351 }
 352
 353 MODULE_AUTHOR("Alexey Kuznetsov");
 354 MODULE_DESCRIPTION("Policing actions");
 355 MODULE_LICENSE("GPL");
 356
 357 static struct tc_action_ops act_police_ops = {
 358         .kind           =       "police",
 359         .type           =       TCA_ID_POLICE,
 360         .owner          =       THIS_MODULE,
 361         .act            =       tcf_act_police,
 362         .dump           =       tcf_act_police_dump,
 363         .init           =       tcf_act_police_init,
 364         .walk           =       tcf_act_police_walker,
 365         .lookup         =       tcf_police_search,
 366         .size           =       sizeof(struct tcf_police),
 367 };
 368
 369 static __net_init int police_init_net(struct net *net)
 370 {
 371         struct tc_action_net *tn = net_generic(net, police_net_id);
 372
 373         return tc_action_net_init(tn, &act_police_ops, POL_TAB_MASK);
 374 }
 375
 376 static void __net_exit police_exit_net(struct net *net)
 377 {
 378         struct tc_action_net *tn = net_generic(net, police_net_id);
 379
 380         tc_action_net_exit(tn);
 381 }
 382
 383 static struct pernet_operations police_net_ops = {
 384         .init = police_init_net,
 385         .exit = police_exit_net,
 386         .id   = &police_net_id,
 387         .size = sizeof(struct tc_action_net),
 388 };
 389
 390 static int __init
 391 police_init_module(void)
 392 {
 393         return tcf_register_action(&act_police_ops, &police_net_ops);
 394 }
 395
 396 static void __exit
 397 police_cleanup_module(void)
 398 {
 399         tcf_unregister_action(&act_police_ops, &police_net_ops);
 400 }
 401
 402 module_init(police_init_module);
 403 module_exit(police_cleanup_module);