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