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/module.h>
  20 #include <linux/rtnetlink.h>
  21 #include <linux/init.h>
  22 #include <net/act_api.h>
  23 #include <net/netlink.h>
  24
  25 #define L2T(p,L)   ((p)->tcfp_R_tab->data[(L)>>(p)->tcfp_R_tab->rate.cell_log])
  26 #define L2T_P(p,L) ((p)->tcfp_P_tab->data[(L)>>(p)->tcfp_P_tab->rate.cell_log])
  27
  28 #define POL_TAB_MASK     15
  29 static struct tcf_common *tcf_police_ht[POL_TAB_MASK + 1];
  30 static u32 police_idx_gen;
  31 static DEFINE_RWLOCK(police_lock);
  32
  33 static struct tcf_hashinfo police_hash_info = {
  34         .htab   =       tcf_police_ht,
  35         .hmask  =       POL_TAB_MASK,
  36         .lock   =       &police_lock,
  37 };
  38
  39 /* old policer structure from before tc actions */
  40 struct tc_police_compat
  41 {
  42         u32                     index;
  43         int                     action;
  44         u32                     limit;
  45         u32                     burst;
  46         u32                     mtu;
  47         struct tc_ratespec      rate;
  48         struct tc_ratespec      peakrate;
  49 };
  50
  51 /* Each policer is serialized by its individual spinlock */
  52
  53 static int tcf_act_police_walker(struct sk_buff *skb, struct netlink_callback *cb,
  54                               int type, struct tc_action *a)
  55 {
  56         struct tcf_common *p;
  57         int err = 0, index = -1, i = 0, s_i = 0, n_i = 0;
  58         struct rtattr *r;
  59
  60         read_lock(&police_lock);
  61
  62         s_i = cb->args[0];
  63
  64         for (i = 0; i < (POL_TAB_MASK + 1); i++) {
  65                 p = tcf_police_ht[tcf_hash(i, POL_TAB_MASK)];
  66
  67                 for (; p; p = p->tcfc_next) {
  68                         index++;
  69                         if (index < s_i)
  70                                 continue;
  71                         a->priv = p;
  72                         a->order = index;
  73                         r = (struct rtattr *)skb_tail_pointer(skb);
  74                         RTA_PUT(skb, a->order, 0, NULL);
  75                         if (type == RTM_DELACTION)
  76                                 err = tcf_action_dump_1(skb, a, 0, 1);
  77                         else
  78                                 err = tcf_action_dump_1(skb, a, 0, 0);
  79                         if (err < 0) {
  80                                 index--;
  81                                 nlmsg_trim(skb, r);
  82                                 goto done;
  83                         }
  84                         r->rta_len = skb_tail_pointer(skb) - (u8 *)r;
  85                         n_i++;
  86                 }
  87         }
  88 done:
  89         read_unlock(&police_lock);
  90         if (n_i)
  91                 cb->args[0] += n_i;
  92         return n_i;
  93
  94 rtattr_failure:
  95         nlmsg_trim(skb, r);
  96         goto done;
  97 }
  98
  99 static void tcf_police_destroy(struct tcf_police *p)
 100 {
 101         unsigned int h = tcf_hash(p->tcf_index, POL_TAB_MASK);
 102         struct tcf_common **p1p;
 103
 104         for (p1p = &tcf_police_ht[h]; *p1p; p1p = &(*p1p)->tcfc_next) {
 105                 if (*p1p == &p->common) {
 106                         write_lock_bh(&police_lock);
 107                         *p1p = p->tcf_next;
 108                         write_unlock_bh(&police_lock);
 109                         gen_kill_estimator(&p->tcf_bstats,
 110                                            &p->tcf_rate_est);
 111                         if (p->tcfp_R_tab)
 112                                 qdisc_put_rtab(p->tcfp_R_tab);
 113                         if (p->tcfp_P_tab)
 114                                 qdisc_put_rtab(p->tcfp_P_tab);
 115                         kfree(p);
 116                         return;
 117                 }
 118         }
 119         BUG_TRAP(0);
 120 }
 121
 122 static int tcf_act_police_locate(struct rtattr *rta, struct rtattr *est,
 123                                  struct tc_action *a, int ovr, int bind)
 124 {
 125         unsigned h;
 126         int ret = 0, err;
 127         struct rtattr *tb[TCA_POLICE_MAX];
 128         struct tc_police *parm;
 129         struct tcf_police *police;
 130         struct qdisc_rate_table *R_tab = NULL, *P_tab = NULL;
 131         int size;
 132
 133         if (rta == NULL || rtattr_parse_nested(tb, TCA_POLICE_MAX, rta) < 0)
 134                 return -EINVAL;
 135
 136         if (tb[TCA_POLICE_TBF-1] == NULL)
 137                 return -EINVAL;
 138         size = RTA_PAYLOAD(tb[TCA_POLICE_TBF-1]);
 139         if (size != sizeof(*parm) && size != sizeof(struct tc_police_compat))
 140                 return -EINVAL;
 141         parm = RTA_DATA(tb[TCA_POLICE_TBF-1]);
 142
 143         if (tb[TCA_POLICE_RESULT-1] != NULL &&
 144             RTA_PAYLOAD(tb[TCA_POLICE_RESULT-1]) != sizeof(u32))
 145                 return -EINVAL;
 146         if (tb[TCA_POLICE_RESULT-1] != NULL &&
 147             RTA_PAYLOAD(tb[TCA_POLICE_RESULT-1]) != sizeof(u32))
 148                 return -EINVAL;
 149
 150         if (parm->index) {
 151                 struct tcf_common *pc;
 152
 153                 pc = tcf_hash_lookup(parm->index, &police_hash_info);
 154                 if (pc != NULL) {
 155                         a->priv = pc;
 156                         police = to_police(pc);
 157                         if (bind) {
 158                                 police->tcf_bindcnt += 1;
 159                                 police->tcf_refcnt += 1;
 160                         }
 161                         if (ovr)
 162                                 goto override;
 163                         return ret;
 164                 }
 165         }
 166
 167         police = kzalloc(sizeof(*police), GFP_KERNEL);
 168         if (police == NULL)
 169                 return -ENOMEM;
 170         ret = ACT_P_CREATED;
 171         police->tcf_refcnt = 1;
 172         spin_lock_init(&police->tcf_lock);
 173         if (bind)
 174                 police->tcf_bindcnt = 1;
 175 override:
 176         if (parm->rate.rate) {
 177                 err = -ENOMEM;
 178                 R_tab = qdisc_get_rtab(&parm->rate, tb[TCA_POLICE_RATE-1]);
 179                 if (R_tab == NULL)
 180                         goto failure;
 181                 if (parm->peakrate.rate) {
 182                         P_tab = qdisc_get_rtab(&parm->peakrate,
 183                                                tb[TCA_POLICE_PEAKRATE-1]);
 184                         if (P_tab == NULL) {
 185                                 qdisc_put_rtab(R_tab);
 186                                 goto failure;
 187                         }
 188                 }
 189         }
 190         /* No failure allowed after this point */
 191         spin_lock_bh(&police->tcf_lock);
 192         if (R_tab != NULL) {
 193                 qdisc_put_rtab(police->tcfp_R_tab);
 194                 police->tcfp_R_tab = R_tab;
 195         }
 196         if (P_tab != NULL) {
 197                 qdisc_put_rtab(police->tcfp_P_tab);
 198                 police->tcfp_P_tab = P_tab;
 199         }
 200
 201         if (tb[TCA_POLICE_RESULT-1])
 202                 police->tcfp_result = *(u32*)RTA_DATA(tb[TCA_POLICE_RESULT-1]);
 203         police->tcfp_toks = police->tcfp_burst = parm->burst;
 204         police->tcfp_mtu = parm->mtu;
 205         if (police->tcfp_mtu == 0) {
 206                 police->tcfp_mtu = ~0;
 207                 if (police->tcfp_R_tab)
 208                         police->tcfp_mtu = 255<<police->tcfp_R_tab->rate.cell_log;
 209         }
 210         if (police->tcfp_P_tab)
 211                 police->tcfp_ptoks = L2T_P(police, police->tcfp_mtu);
 212         police->tcf_action = parm->action;
 213
 214         if (tb[TCA_POLICE_AVRATE-1])
 215                 police->tcfp_ewma_rate =
 216                         *(u32*)RTA_DATA(tb[TCA_POLICE_AVRATE-1]);
 217         if (est)
 218                 gen_replace_estimator(&police->tcf_bstats,
 219                                       &police->tcf_rate_est,
 220                                       &police->tcf_lock, est);
 221
 222         spin_unlock_bh(&police->tcf_lock);
 223         if (ret != ACT_P_CREATED)
 224                 return ret;
 225
 226         police->tcfp_t_c = psched_get_time();
 227         police->tcf_index = parm->index ? parm->index :
 228                 tcf_hash_new_index(&police_idx_gen, &police_hash_info);
 229         h = tcf_hash(police->tcf_index, POL_TAB_MASK);
 230         write_lock_bh(&police_lock);
 231         police->tcf_next = tcf_police_ht[h];
 232         tcf_police_ht[h] = &police->common;
 233         write_unlock_bh(&police_lock);
 234
 235         a->priv = police;
 236         return ret;
 237
 238 failure:
 239         if (ret == ACT_P_CREATED)
 240                 kfree(police);
 241         return err;
 242 }
 243
 244 static int tcf_act_police_cleanup(struct tc_action *a, int bind)
 245 {
 246         struct tcf_police *p = a->priv;
 247         int ret = 0;
 248
 249         if (p != NULL) {
 250                 if (bind)
 251                         p->tcf_bindcnt--;
 252
 253                 p->tcf_refcnt--;
 254                 if (p->tcf_refcnt <= 0 && !p->tcf_bindcnt) {
 255                         tcf_police_destroy(p);
 256                         ret = 1;
 257                 }
 258         }
 259         return ret;
 260 }
 261
 262 static int tcf_act_police(struct sk_buff *skb, struct tc_action *a,
 263                           struct tcf_result *res)
 264 {
 265         struct tcf_police *police = a->priv;
 266         psched_time_t now;
 267         long toks;
 268         long ptoks = 0;
 269
 270         spin_lock(&police->tcf_lock);
 271
 272         police->tcf_bstats.bytes += skb->len;
 273         police->tcf_bstats.packets++;
 274
 275         if (police->tcfp_ewma_rate &&
 276             police->tcf_rate_est.bps >= police->tcfp_ewma_rate) {
 277                 police->tcf_qstats.overlimits++;
 278                 spin_unlock(&police->tcf_lock);
 279                 return police->tcf_action;
 280         }
 281
 282         if (skb->len <= police->tcfp_mtu) {
 283                 if (police->tcfp_R_tab == NULL) {
 284                         spin_unlock(&police->tcf_lock);
 285                         return police->tcfp_result;
 286                 }
 287
 288                 now = psched_get_time();
 289                 toks = psched_tdiff_bounded(now, police->tcfp_t_c,
 290                                             police->tcfp_burst);
 291                 if (police->tcfp_P_tab) {
 292                         ptoks = toks + police->tcfp_ptoks;
 293                         if (ptoks > (long)L2T_P(police, police->tcfp_mtu))
 294                                 ptoks = (long)L2T_P(police, police->tcfp_mtu);
 295                         ptoks -= L2T_P(police, skb->len);
 296                 }
 297                 toks += police->tcfp_toks;
 298                 if (toks > (long)police->tcfp_burst)
 299                         toks = police->tcfp_burst;
 300                 toks -= L2T(police, skb->len);
 301                 if ((toks|ptoks) >= 0) {
 302                         police->tcfp_t_c = now;
 303                         police->tcfp_toks = toks;
 304                         police->tcfp_ptoks = ptoks;
 305                         spin_unlock(&police->tcf_lock);
 306                         return police->tcfp_result;
 307                 }
 308         }
 309
 310         police->tcf_qstats.overlimits++;
 311         spin_unlock(&police->tcf_lock);
 312         return police->tcf_action;
 313 }
 314
 315 static int
 316 tcf_act_police_dump(struct sk_buff *skb, struct tc_action *a, int bind, int ref)
 317 {
 318         unsigned char *b = skb_tail_pointer(skb);
 319         struct tcf_police *police = a->priv;
 320         struct tc_police opt;
 321
 322         opt.index = police->tcf_index;
 323         opt.action = police->tcf_action;
 324         opt.mtu = police->tcfp_mtu;
 325         opt.burst = police->tcfp_burst;
 326         opt.refcnt = police->tcf_refcnt - ref;
 327         opt.bindcnt = police->tcf_bindcnt - bind;
 328         if (police->tcfp_R_tab)
 329                 opt.rate = police->tcfp_R_tab->rate;
 330         else
 331                 memset(&opt.rate, 0, sizeof(opt.rate));
 332         if (police->tcfp_P_tab)
 333                 opt.peakrate = police->tcfp_P_tab->rate;
 334         else
 335                 memset(&opt.peakrate, 0, sizeof(opt.peakrate));
 336         RTA_PUT(skb, TCA_POLICE_TBF, sizeof(opt), &opt);
 337         if (police->tcfp_result)
 338                 RTA_PUT(skb, TCA_POLICE_RESULT, sizeof(int),
 339                         &police->tcfp_result);
 340         if (police->tcfp_ewma_rate)
 341                 RTA_PUT(skb, TCA_POLICE_AVRATE, 4, &police->tcfp_ewma_rate);
 342         return skb->len;
 343
 344 rtattr_failure:
 345         nlmsg_trim(skb, b);
 346         return -1;
 347 }
 348
 349 MODULE_AUTHOR("Alexey Kuznetsov");
 350 MODULE_DESCRIPTION("Policing actions");
 351 MODULE_LICENSE("GPL");
 352
 353 static struct tc_action_ops act_police_ops = {
 354         .kind           =       "police",
 355         .hinfo          =       &police_hash_info,
 356         .type           =       TCA_ID_POLICE,
 357         .capab          =       TCA_CAP_NONE,
 358         .owner          =       THIS_MODULE,
 359         .act            =       tcf_act_police,
 360         .dump           =       tcf_act_police_dump,
 361         .cleanup        =       tcf_act_police_cleanup,
 362         .lookup         =       tcf_hash_search,
 363         .init           =       tcf_act_police_locate,
 364         .walk           =       tcf_act_police_walker
 365 };
 366
 367 static int __init
 368 police_init_module(void)
 369 {
 370         return tcf_register_action(&act_police_ops);
 371 }
 372
 373 static void __exit
 374 police_cleanup_module(void)
 375 {
 376         tcf_unregister_action(&act_police_ops);
 377 }
 378
 379 module_init(police_init_module);
 380 module_exit(police_cleanup_module);