net/sched/act_police.c

   1 /*
   2  * net/sched/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, struct tcf_police, common)
  42
  43 #define POL_TAB_MASK     15
  44 static struct tcf_common *tcf_police_ht[POL_TAB_MASK + 1];
  45 static u32 police_idx_gen;
  46 static DEFINE_RWLOCK(police_lock);
  47
  48 static struct tcf_hashinfo police_hash_info = {
  49         .htab   =       tcf_police_ht,
  50         .hmask  =       POL_TAB_MASK,
  51         .lock   =       &police_lock,
  52 };
  53
  54 /* old policer structure from before tc actions */
  55 struct tc_police_compat {
  56         u32                     index;
  57         int                     action;
  58         u32                     limit;
  59         u32                     burst;
  60         u32                     mtu;
  61         struct tc_ratespec      rate;
  62         struct tc_ratespec      peakrate;
  63 };
  64
  65 /* Each policer is serialized by its individual spinlock */
  66
  67 static int tcf_act_police_walker(struct sk_buff *skb, struct netlink_callback *cb,
  68                               int type, struct tc_action *a)
  69 {
  70         struct tcf_common *p;
  71         int err = 0, index = -1, i = 0, s_i = 0, n_i = 0;
  72         struct nlattr *nest;
  73
  74         read_lock_bh(&police_lock);
  75
  76         s_i = cb->args[0];
  77
  78         for (i = 0; i < (POL_TAB_MASK + 1); i++) {
  79                 p = tcf_police_ht[tcf_hash(i, POL_TAB_MASK)];
  80
  81                 for (; p; p = p->tcfc_next) {
  82                         index++;
  83                         if (index < s_i)
  84                                 continue;
  85                         a->priv = p;
  86                         a->order = index;
  87                         nest = nla_nest_start(skb, a->order);
  88                         if (nest == NULL)
  89                                 goto nla_put_failure;
  90                         if (type == RTM_DELACTION)
  91                                 err = tcf_action_dump_1(skb, a, 0, 1);
  92                         else
  93                                 err = tcf_action_dump_1(skb, a, 0, 0);
  94                         if (err < 0) {
  95                                 index--;
  96                                 nla_nest_cancel(skb, nest);
  97                                 goto done;
  98                         }
  99                         nla_nest_end(skb, nest);
 100                         n_i++;
 101                 }
 102         }
 103 done:
 104         read_unlock_bh(&police_lock);
 105         if (n_i)
 106                 cb->args[0] += n_i;
 107         return n_i;
 108
 109 nla_put_failure:
 110         nla_nest_cancel(skb, nest);
 111         goto done;
 112 }
 113
 114 static void tcf_police_destroy(struct tcf_police *p)
 115 {
 116         unsigned int h = tcf_hash(p->tcf_index, POL_TAB_MASK);
 117         struct tcf_common **p1p;
 118
 119         for (p1p = &tcf_police_ht[h]; *p1p; p1p = &(*p1p)->tcfc_next) {
 120                 if (*p1p == &p->common) {
 121                         write_lock_bh(&police_lock);
 122                         *p1p = p->tcf_next;
 123                         write_unlock_bh(&police_lock);
 124                         gen_kill_estimator(&p->tcf_bstats,
 125                                            &p->tcf_rate_est);
 126                         /*
 127                          * gen_estimator est_timer() might access p->tcf_lock
 128                          * or bstats, wait a RCU grace period before freeing p
 129                          */
 130                         kfree_rcu(p, tcf_rcu);
 131                         return;
 132                 }
 133         }
 134         WARN_ON(1);
 135 }
 136
 137 static const struct nla_policy police_policy[TCA_POLICE_MAX + 1] = {
 138         [TCA_POLICE_RATE]       = { .len = TC_RTAB_SIZE },
 139         [TCA_POLICE_PEAKRATE]   = { .len = TC_RTAB_SIZE },
 140         [TCA_POLICE_AVRATE]     = { .type = NLA_U32 },
 141         [TCA_POLICE_RESULT]     = { .type = NLA_U32 },
 142 };
 143
 144 static int tcf_act_police_locate(struct net *net, struct nlattr *nla,
 145                                  struct nlattr *est, struct tc_action *a,
 146                                  int ovr, int bind)
 147 {
 148         unsigned int h;
 149         int ret = 0, err;
 150         struct nlattr *tb[TCA_POLICE_MAX + 1];
 151         struct tc_police *parm;
 152         struct tcf_police *police;
 153         struct qdisc_rate_table *R_tab = NULL, *P_tab = NULL;
 154         int size;
 155
 156         if (nla == NULL)
 157                 return -EINVAL;
 158
 159         err = nla_parse_nested(tb, TCA_POLICE_MAX, nla, police_policy);
 160         if (err < 0)
 161                 return err;
 162
 163         if (tb[TCA_POLICE_TBF] == NULL)
 164                 return -EINVAL;
 165         size = nla_len(tb[TCA_POLICE_TBF]);
 166         if (size != sizeof(*parm) && size != sizeof(struct tc_police_compat))
 167                 return -EINVAL;
 168         parm = nla_data(tb[TCA_POLICE_TBF]);
 169
 170         if (parm->index) {
 171                 struct tcf_common *pc;
 172
 173                 pc = tcf_hash_lookup(parm->index, &police_hash_info);
 174                 if (pc != NULL) {
 175                         a->priv = pc;
 176                         police = to_police(pc);
 177                         if (bind) {
 178                                 police->tcf_bindcnt += 1;
 179                                 police->tcf_refcnt += 1;
 180                                 return 0;
 181                         }
 182                         if (ovr)
 183                                 goto override;
 184                         /* not replacing */
 185                         return -EEXIST;
 186                 }
 187         }
 188
 189         police = kzalloc(sizeof(*police), GFP_KERNEL);
 190         if (police == NULL)
 191                 return -ENOMEM;
 192         ret = ACT_P_CREATED;
 193         police->tcf_refcnt = 1;
 194         spin_lock_init(&police->tcf_lock);
 195         if (bind)
 196                 police->tcf_bindcnt = 1;
 197 override:
 198         if (parm->rate.rate) {
 199                 err = -ENOMEM;
 200                 R_tab = qdisc_get_rtab(&parm->rate, tb[TCA_POLICE_RATE]);
 201                 if (R_tab == NULL)
 202                         goto failure;
 203
 204                 if (parm->peakrate.rate) {
 205                         P_tab = qdisc_get_rtab(&parm->peakrate,
 206                                                tb[TCA_POLICE_PEAKRATE]);
 207                         if (P_tab == NULL)
 208                                 goto failure;
 209                 }
 210         }
 211
 212         spin_lock_bh(&police->tcf_lock);
 213         if (est) {
 214                 err = gen_replace_estimator(&police->tcf_bstats,
 215                                             &police->tcf_rate_est,
 216                                             &police->tcf_lock, est);
 217                 if (err)
 218                         goto failure_unlock;
 219         } else if (tb[TCA_POLICE_AVRATE] &&
 220                    (ret == ACT_P_CREATED ||
 221                     !gen_estimator_active(&police->tcf_bstats,
 222                                           &police->tcf_rate_est))) {
 223                 err = -EINVAL;
 224                 goto failure_unlock;
 225         }
 226
 227         /* No failure allowed after this point */
 228         police->tcfp_mtu = parm->mtu;
 229         if (police->tcfp_mtu == 0) {
 230                 police->tcfp_mtu = ~0;
 231                 if (R_tab)
 232                         police->tcfp_mtu = 255 << R_tab->rate.cell_log;
 233         }
 234         if (R_tab) {
 235                 police->rate_present = true;
 236                 psched_ratecfg_precompute(&police->rate, &R_tab->rate, 0);
 237                 qdisc_put_rtab(R_tab);
 238         } else {
 239                 police->rate_present = false;
 240         }
 241         if (P_tab) {
 242                 police->peak_present = true;
 243                 psched_ratecfg_precompute(&police->peak, &P_tab->rate, 0);
 244                 qdisc_put_rtab(P_tab);
 245         } else {
 246                 police->peak_present = false;
 247         }
 248
 249         if (tb[TCA_POLICE_RESULT])
 250                 police->tcfp_result = nla_get_u32(tb[TCA_POLICE_RESULT]);
 251         police->tcfp_burst = PSCHED_TICKS2NS(parm->burst);
 252         police->tcfp_toks = police->tcfp_burst;
 253         if (police->peak_present) {
 254                 police->tcfp_mtu_ptoks = (s64) psched_l2t_ns(&police->peak,
 255                                                              police->tcfp_mtu);
 256                 police->tcfp_ptoks = police->tcfp_mtu_ptoks;
 257         }
 258         police->tcf_action = parm->action;
 259
 260         if (tb[TCA_POLICE_AVRATE])
 261                 police->tcfp_ewma_rate = nla_get_u32(tb[TCA_POLICE_AVRATE]);
 262
 263         spin_unlock_bh(&police->tcf_lock);
 264         if (ret != ACT_P_CREATED)
 265                 return ret;
 266
 267         police->tcfp_t_c = ktime_to_ns(ktime_get());
 268         police->tcf_index = parm->index ? parm->index :
 269                 tcf_hash_new_index(&police_idx_gen, &police_hash_info);
 270         h = tcf_hash(police->tcf_index, POL_TAB_MASK);
 271         write_lock_bh(&police_lock);
 272         police->tcf_next = tcf_police_ht[h];
 273         tcf_police_ht[h] = &police->common;
 274         write_unlock_bh(&police_lock);
 275
 276         a->priv = police;
 277         return ret;
 278
 279 failure_unlock:
 280         spin_unlock_bh(&police->tcf_lock);
 281 failure:
 282         if (P_tab)
 283                 qdisc_put_rtab(P_tab);
 284         if (R_tab)
 285                 qdisc_put_rtab(R_tab);
 286         if (ret == ACT_P_CREATED)
 287                 kfree(police);
 288         return err;
 289 }
 290
 291 static int tcf_act_police_cleanup(struct tc_action *a, int bind)
 292 {
 293         struct tcf_police *p = a->priv;
 294         int ret = 0;
 295
 296         if (p != NULL) {
 297                 if (bind)
 298                         p->tcf_bindcnt--;
 299
 300                 p->tcf_refcnt--;
 301                 if (p->tcf_refcnt <= 0 && !p->tcf_bindcnt) {
 302                         tcf_police_destroy(p);
 303                         ret = 1;
 304                 }
 305         }
 306         return ret;
 307 }
 308
 309 static int tcf_act_police(struct sk_buff *skb, const struct tc_action *a,
 310                           struct tcf_result *res)
 311 {
 312         struct tcf_police *police = a->priv;
 313         s64 now;
 314         s64 toks;
 315         s64 ptoks = 0;
 316
 317         spin_lock(&police->tcf_lock);
 318
 319         bstats_update(&police->tcf_bstats, skb);
 320
 321         if (police->tcfp_ewma_rate &&
 322             police->tcf_rate_est.bps >= police->tcfp_ewma_rate) {
 323                 police->tcf_qstats.overlimits++;
 324                 if (police->tcf_action == TC_ACT_SHOT)
 325                         police->tcf_qstats.drops++;
 326                 spin_unlock(&police->tcf_lock);
 327                 return police->tcf_action;
 328         }
 329
 330         if (qdisc_pkt_len(skb) <= police->tcfp_mtu) {
 331                 if (!police->rate_present) {
 332                         spin_unlock(&police->tcf_lock);
 333                         return police->tcfp_result;
 334                 }
 335
 336                 now = ktime_to_ns(ktime_get());
 337                 toks = min_t(s64, now - police->tcfp_t_c,
 338                              police->tcfp_burst);
 339                 if (police->peak_present) {
 340                         ptoks = toks + police->tcfp_ptoks;
 341                         if (ptoks > police->tcfp_mtu_ptoks)
 342                                 ptoks = police->tcfp_mtu_ptoks;
 343                         ptoks -= (s64) psched_l2t_ns(&police->peak,
 344                                                      qdisc_pkt_len(skb));
 345                 }
 346                 toks += police->tcfp_toks;
 347                 if (toks > police->tcfp_burst)
 348                         toks = police->tcfp_burst;
 349                 toks -= (s64) psched_l2t_ns(&police->rate, qdisc_pkt_len(skb));
 350                 if ((toks|ptoks) >= 0) {
 351                         police->tcfp_t_c = now;
 352                         police->tcfp_toks = toks;
 353                         police->tcfp_ptoks = ptoks;
 354                         spin_unlock(&police->tcf_lock);
 355                         return police->tcfp_result;
 356                 }
 357         }
 358
 359         police->tcf_qstats.overlimits++;
 360         if (police->tcf_action == TC_ACT_SHOT)
 361                 police->tcf_qstats.drops++;
 362         spin_unlock(&police->tcf_lock);
 363         return police->tcf_action;
 364 }
 365
 366 static int
 367 tcf_act_police_dump(struct sk_buff *skb, struct tc_action *a, int bind, int ref)
 368 {
 369         unsigned char *b = skb_tail_pointer(skb);
 370         struct tcf_police *police = a->priv;
 371         struct tc_police opt = {
 372                 .index = police->tcf_index,
 373                 .action = police->tcf_action,
 374                 .mtu = police->tcfp_mtu,
 375                 .burst = PSCHED_NS2TICKS(police->tcfp_burst),
 376                 .refcnt = police->tcf_refcnt - ref,
 377                 .bindcnt = police->tcf_bindcnt - bind,
 378         };
 379
 380         if (police->rate_present)
 381                 psched_ratecfg_getrate(&opt.rate, &police->rate);
 382         if (police->peak_present)
 383                 psched_ratecfg_getrate(&opt.peakrate, &police->peak);
 384         if (nla_put(skb, TCA_POLICE_TBF, sizeof(opt), &opt))
 385                 goto nla_put_failure;
 386         if (police->tcfp_result &&
 387             nla_put_u32(skb, TCA_POLICE_RESULT, police->tcfp_result))
 388                 goto nla_put_failure;
 389         if (police->tcfp_ewma_rate &&
 390             nla_put_u32(skb, TCA_POLICE_AVRATE, police->tcfp_ewma_rate))
 391                 goto nla_put_failure;
 392         return skb->len;
 393
 394 nla_put_failure:
 395         nlmsg_trim(skb, b);
 396         return -1;
 397 }
 398
 399 MODULE_AUTHOR("Alexey Kuznetsov");
 400 MODULE_DESCRIPTION("Policing actions");
 401 MODULE_LICENSE("GPL");
 402
 403 static struct tc_action_ops act_police_ops = {
 404         .kind           =       "police",
 405         .hinfo          =       &police_hash_info,
 406         .type           =       TCA_ID_POLICE,
 407         .capab          =       TCA_CAP_NONE,
 408         .owner          =       THIS_MODULE,
 409         .act            =       tcf_act_police,
 410         .dump           =       tcf_act_police_dump,
 411         .cleanup        =       tcf_act_police_cleanup,
 412         .init           =       tcf_act_police_locate,
 413         .walk           =       tcf_act_police_walker
 414 };
 415
 416 static int __init
 417 police_init_module(void)
 418 {
 419         return tcf_register_action(&act_police_ops);
 420 }
 421
 422 static void __exit
 423 police_cleanup_module(void)
 424 {
 425         tcf_unregister_action(&act_police_ops);
 426 }
 427
 428 module_init(police_init_module);
 429 module_exit(police_cleanup_module);