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