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 #include <net/pkt_cls.h>
  25
  26 struct tcf_police_params {
  27         int                     tcfp_result;
  28         u32                     tcfp_ewma_rate;
  29         s64                     tcfp_burst;
  30         u32                     tcfp_mtu;
  31         s64                     tcfp_mtu_ptoks;
  32         struct psched_ratecfg   rate;
  33         bool                    rate_present;
  34         struct psched_ratecfg   peak;
  35         bool                    peak_present;
  36         struct rcu_head rcu;
  37 };
  38
  39 struct tcf_police {
  40         struct tc_action        common;
  41         struct tcf_police_params __rcu *params;
  42
  43         spinlock_t              tcfp_lock ____cacheline_aligned_in_smp;
  44         s64                     tcfp_toks;
  45         s64                     tcfp_ptoks;
  46         s64                     tcfp_t_c;
  47 };
  48
  49 #define to_police(pc) ((struct tcf_police *)pc)
  50
  51 /* old policer structure from before tc actions */
  52 struct tc_police_compat {
  53         u32                     index;
  54         int                     action;
  55         u32                     limit;
  56         u32                     burst;
  57         u32                     mtu;
  58         struct tc_ratespec      rate;
  59         struct tc_ratespec      peakrate;
  60 };
  61
  62 /* Each policer is serialized by its individual spinlock */
  63
  64 static unsigned int police_net_id;
  65 static struct tc_action_ops act_police_ops;
  66
  67 static int tcf_police_walker(struct net *net, struct sk_buff *skb,
  68                                  struct netlink_callback *cb, int type,
  69                                  const struct tc_action_ops *ops,
  70                                  struct netlink_ext_ack *extack)
  71 {
  72         struct tc_action_net *tn = net_generic(net, police_net_id);
  73
  74         return tcf_generic_walker(tn, skb, cb, type, ops, extack);
  75 }
  76
  77 static const struct nla_policy police_policy[TCA_POLICE_MAX + 1] = {
  78         [TCA_POLICE_RATE]       = { .len = TC_RTAB_SIZE },
  79         [TCA_POLICE_PEAKRATE]   = { .len = TC_RTAB_SIZE },
  80         [TCA_POLICE_AVRATE]     = { .type = NLA_U32 },
  81         [TCA_POLICE_RESULT]     = { .type = NLA_U32 },
  82 };
  83
  84 static int tcf_police_init(struct net *net, struct nlattr *nla,
  85                                struct nlattr *est, struct tc_action **a,
  86                                int ovr, int bind, bool rtnl_held,
  87                                struct tcf_proto *tp,
  88                                struct netlink_ext_ack *extack)
  89 {
  90         int ret = 0, tcfp_result = TC_ACT_OK, err, size;
  91         struct nlattr *tb[TCA_POLICE_MAX + 1];
  92         struct tcf_chain *goto_ch = NULL;
  93         struct tc_police *parm;
  94         struct tcf_police *police;
  95         struct qdisc_rate_table *R_tab = NULL, *P_tab = NULL;
  96         struct tc_action_net *tn = net_generic(net, police_net_id);
  97         struct tcf_police_params *new;
  98         bool exists = false;
  99
 100         if (nla == NULL)
 101                 return -EINVAL;
 102
 103         err = nla_parse_nested(tb, TCA_POLICE_MAX, nla, police_policy, NULL);
 104         if (err < 0)
 105                 return err;
 106
 107         if (tb[TCA_POLICE_TBF] == NULL)
 108                 return -EINVAL;
 109         size = nla_len(tb[TCA_POLICE_TBF]);
 110         if (size != sizeof(*parm) && size != sizeof(struct tc_police_compat))
 111                 return -EINVAL;
 112
 113         parm = nla_data(tb[TCA_POLICE_TBF]);
 114         err = tcf_idr_check_alloc(tn, &parm->index, a, bind);
 115         if (err < 0)
 116                 return err;
 117         exists = err;
 118         if (exists && bind)
 119                 return 0;
 120
 121         if (!exists) {
 122                 ret = tcf_idr_create(tn, parm->index, NULL, a,
 123                                      &act_police_ops, bind, true);
 124                 if (ret) {
 125                         tcf_idr_cleanup(tn, parm->index);
 126                         return ret;
 127                 }
 128                 ret = ACT_P_CREATED;
 129                 spin_lock_init(&(to_police(*a)->tcfp_lock));
 130         } else if (!ovr) {
 131                 tcf_idr_release(*a, bind);
 132                 return -EEXIST;
 133         }
 134         err = tcf_action_check_ctrlact(parm->action, tp, &goto_ch, extack);
 135         if (err < 0)
 136                 goto release_idr;
 137
 138         police = to_police(*a);
 139         if (parm->rate.rate) {
 140                 err = -ENOMEM;
 141                 R_tab = qdisc_get_rtab(&parm->rate, tb[TCA_POLICE_RATE], NULL);
 142                 if (R_tab == NULL)
 143                         goto failure;
 144
 145                 if (parm->peakrate.rate) {
 146                         P_tab = qdisc_get_rtab(&parm->peakrate,
 147                                                tb[TCA_POLICE_PEAKRATE], NULL);
 148                         if (P_tab == NULL)
 149                                 goto failure;
 150                 }
 151         }
 152
 153         if (est) {
 154                 err = gen_replace_estimator(&police->tcf_bstats,
 155                                             police->common.cpu_bstats,
 156                                             &police->tcf_rate_est,
 157                                             &police->tcf_lock,
 158                                             NULL, est);
 159                 if (err)
 160                         goto failure;
 161         } else if (tb[TCA_POLICE_AVRATE] &&
 162                    (ret == ACT_P_CREATED ||
 163                     !gen_estimator_active(&police->tcf_rate_est))) {
 164                 err = -EINVAL;
 165                 goto failure;
 166         }
 167
 168         if (tb[TCA_POLICE_RESULT]) {
 169                 tcfp_result = nla_get_u32(tb[TCA_POLICE_RESULT]);
 170                 if (TC_ACT_EXT_CMP(tcfp_result, TC_ACT_GOTO_CHAIN)) {
 171                         NL_SET_ERR_MSG(extack,
 172                                        "goto chain not allowed on fallback");
 173                         err = -EINVAL;
 174                         goto failure;
 175                 }
 176         }
 177
 178         new = kzalloc(sizeof(*new), GFP_KERNEL);
 179         if (unlikely(!new)) {
 180                 err = -ENOMEM;
 181                 goto failure;
 182         }
 183
 184         /* No failure allowed after this point */
 185         new->tcfp_result = tcfp_result;
 186         new->tcfp_mtu = parm->mtu;
 187         if (!new->tcfp_mtu) {
 188                 new->tcfp_mtu = ~0;
 189                 if (R_tab)
 190                         new->tcfp_mtu = 255 << R_tab->rate.cell_log;
 191         }
 192         if (R_tab) {
 193                 new->rate_present = true;
 194                 psched_ratecfg_precompute(&new->rate, &R_tab->rate, 0);
 195                 qdisc_put_rtab(R_tab);
 196         } else {
 197                 new->rate_present = false;
 198         }
 199         if (P_tab) {
 200                 new->peak_present = true;
 201                 psched_ratecfg_precompute(&new->peak, &P_tab->rate, 0);
 202                 qdisc_put_rtab(P_tab);
 203         } else {
 204                 new->peak_present = false;
 205         }
 206
 207         new->tcfp_burst = PSCHED_TICKS2NS(parm->burst);
 208         if (new->peak_present)
 209                 new->tcfp_mtu_ptoks = (s64)psched_l2t_ns(&new->peak,
 210                                                          new->tcfp_mtu);
 211
 212         if (tb[TCA_POLICE_AVRATE])
 213                 new->tcfp_ewma_rate = nla_get_u32(tb[TCA_POLICE_AVRATE]);
 214
 215         spin_lock_bh(&police->tcf_lock);
 216         spin_lock_bh(&police->tcfp_lock);
 217         police->tcfp_t_c = ktime_get_ns();
 218         police->tcfp_toks = new->tcfp_burst;
 219         if (new->peak_present)
 220                 police->tcfp_ptoks = new->tcfp_mtu_ptoks;
 221         spin_unlock_bh(&police->tcfp_lock);
 222         goto_ch = tcf_action_set_ctrlact(*a, parm->action, goto_ch);
 223         rcu_swap_protected(police->params,
 224                            new,
 225                            lockdep_is_held(&police->tcf_lock));
 226         spin_unlock_bh(&police->tcf_lock);
 227
 228         if (goto_ch)
 229                 tcf_chain_put_by_act(goto_ch);
 230         if (new)
 231                 kfree_rcu(new, rcu);
 232
 233         if (ret == ACT_P_CREATED)
 234                 tcf_idr_insert(tn, *a);
 235         return ret;
 236
 237 failure:
 238         qdisc_put_rtab(P_tab);
 239         qdisc_put_rtab(R_tab);
 240         if (goto_ch)
 241                 tcf_chain_put_by_act(goto_ch);
 242 release_idr:
 243         tcf_idr_release(*a, bind);
 244         return err;
 245 }
 246
 247 static int tcf_police_act(struct sk_buff *skb, const struct tc_action *a,
 248                           struct tcf_result *res)
 249 {
 250         struct tcf_police *police = to_police(a);
 251         struct tcf_police_params *p;
 252         s64 now, toks, ptoks = 0;
 253         int ret;
 254
 255         tcf_lastuse_update(&police->tcf_tm);
 256         bstats_cpu_update(this_cpu_ptr(police->common.cpu_bstats), skb);
 257
 258         ret = READ_ONCE(police->tcf_action);
 259         p = rcu_dereference_bh(police->params);
 260
 261         if (p->tcfp_ewma_rate) {
 262                 struct gnet_stats_rate_est64 sample;
 263
 264                 if (!gen_estimator_read(&police->tcf_rate_est, &sample) ||
 265                     sample.bps >= p->tcfp_ewma_rate)
 266                         goto inc_overlimits;
 267         }
 268
 269         if (qdisc_pkt_len(skb) <= p->tcfp_mtu) {
 270                 if (!p->rate_present) {
 271                         ret = p->tcfp_result;
 272                         goto end;
 273                 }
 274
 275                 now = ktime_get_ns();
 276                 spin_lock_bh(&police->tcfp_lock);
 277                 toks = min_t(s64, now - police->tcfp_t_c, p->tcfp_burst);
 278                 if (p->peak_present) {
 279                         ptoks = toks + police->tcfp_ptoks;
 280                         if (ptoks > p->tcfp_mtu_ptoks)
 281                                 ptoks = p->tcfp_mtu_ptoks;
 282                         ptoks -= (s64)psched_l2t_ns(&p->peak,
 283                                                     qdisc_pkt_len(skb));
 284                 }
 285                 toks += police->tcfp_toks;
 286                 if (toks > p->tcfp_burst)
 287                         toks = p->tcfp_burst;
 288                 toks -= (s64)psched_l2t_ns(&p->rate, qdisc_pkt_len(skb));
 289                 if ((toks|ptoks) >= 0) {
 290                         police->tcfp_t_c = now;
 291                         police->tcfp_toks = toks;
 292                         police->tcfp_ptoks = ptoks;
 293                         spin_unlock_bh(&police->tcfp_lock);
 294                         ret = p->tcfp_result;
 295                         goto inc_drops;
 296                 }
 297                 spin_unlock_bh(&police->tcfp_lock);
 298         }
 299
 300 inc_overlimits:
 301         qstats_overlimit_inc(this_cpu_ptr(police->common.cpu_qstats));
 302 inc_drops:
 303         if (ret == TC_ACT_SHOT)
 304                 qstats_drop_inc(this_cpu_ptr(police->common.cpu_qstats));
 305 end:
 306         return ret;
 307 }
 308
 309 static void tcf_police_cleanup(struct tc_action *a)
 310 {
 311         struct tcf_police *police = to_police(a);
 312         struct tcf_police_params *p;
 313
 314         p = rcu_dereference_protected(police->params, 1);
 315         if (p)
 316                 kfree_rcu(p, rcu);
 317 }
 318
 319 static int tcf_police_dump(struct sk_buff *skb, struct tc_action *a,
 320                                int bind, int ref)
 321 {
 322         unsigned char *b = skb_tail_pointer(skb);
 323         struct tcf_police *police = to_police(a);
 324         struct tcf_police_params *p;
 325         struct tc_police opt = {
 326                 .index = police->tcf_index,
 327                 .refcnt = refcount_read(&police->tcf_refcnt) - ref,
 328                 .bindcnt = atomic_read(&police->tcf_bindcnt) - bind,
 329         };
 330         struct tcf_t t;
 331
 332         spin_lock_bh(&police->tcf_lock);
 333         opt.action = police->tcf_action;
 334         p = rcu_dereference_protected(police->params,
 335                                       lockdep_is_held(&police->tcf_lock));
 336         opt.mtu = p->tcfp_mtu;
 337         opt.burst = PSCHED_NS2TICKS(p->tcfp_burst);
 338         if (p->rate_present)
 339                 psched_ratecfg_getrate(&opt.rate, &p->rate);
 340         if (p->peak_present)
 341                 psched_ratecfg_getrate(&opt.peakrate, &p->peak);
 342         if (nla_put(skb, TCA_POLICE_TBF, sizeof(opt), &opt))
 343                 goto nla_put_failure;
 344         if (p->tcfp_result &&
 345             nla_put_u32(skb, TCA_POLICE_RESULT, p->tcfp_result))
 346                 goto nla_put_failure;
 347         if (p->tcfp_ewma_rate &&
 348             nla_put_u32(skb, TCA_POLICE_AVRATE, p->tcfp_ewma_rate))
 349                 goto nla_put_failure;
 350
 351         t.install = jiffies_to_clock_t(jiffies - police->tcf_tm.install);
 352         t.lastuse = jiffies_to_clock_t(jiffies - police->tcf_tm.lastuse);
 353         t.firstuse = jiffies_to_clock_t(jiffies - police->tcf_tm.firstuse);
 354         t.expires = jiffies_to_clock_t(police->tcf_tm.expires);
 355         if (nla_put_64bit(skb, TCA_POLICE_TM, sizeof(t), &t, TCA_POLICE_PAD))
 356                 goto nla_put_failure;
 357         spin_unlock_bh(&police->tcf_lock);
 358
 359         return skb->len;
 360
 361 nla_put_failure:
 362         spin_unlock_bh(&police->tcf_lock);
 363         nlmsg_trim(skb, b);
 364         return -1;
 365 }
 366
 367 static int tcf_police_search(struct net *net, struct tc_action **a, u32 index)
 368 {
 369         struct tc_action_net *tn = net_generic(net, police_net_id);
 370
 371         return tcf_idr_search(tn, a, index);
 372 }
 373
 374 MODULE_AUTHOR("Alexey Kuznetsov");
 375 MODULE_DESCRIPTION("Policing actions");
 376 MODULE_LICENSE("GPL");
 377
 378 static struct tc_action_ops act_police_ops = {
 379         .kind           =       "police",
 380         .id             =       TCA_ID_POLICE,
 381         .owner          =       THIS_MODULE,
 382         .act            =       tcf_police_act,
 383         .dump           =       tcf_police_dump,
 384         .init           =       tcf_police_init,
 385         .walk           =       tcf_police_walker,
 386         .lookup         =       tcf_police_search,
 387         .cleanup        =       tcf_police_cleanup,
 388         .size           =       sizeof(struct tcf_police),
 389 };
 390
 391 static __net_init int police_init_net(struct net *net)
 392 {
 393         struct tc_action_net *tn = net_generic(net, police_net_id);
 394
 395         return tc_action_net_init(tn, &act_police_ops);
 396 }
 397
 398 static void __net_exit police_exit_net(struct list_head *net_list)
 399 {
 400         tc_action_net_exit(net_list, police_net_id);
 401 }
 402
 403 static struct pernet_operations police_net_ops = {
 404         .init = police_init_net,
 405         .exit_batch = police_exit_net,
 406         .id   = &police_net_id,
 407         .size = sizeof(struct tc_action_net),
 408 };
 409
 410 static int __init police_init_module(void)
 411 {
 412         return tcf_register_action(&act_police_ops, &police_net_ops);
 413 }
 414
 415 static void __exit police_cleanup_module(void)
 416 {
 417         tcf_unregister_action(&act_police_ops, &police_net_ops);
 418 }
 419
 420 module_init(police_init_module);
 421 module_exit(police_cleanup_module);