net/sched/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  */
  11
  12 #include <asm/uaccess.h>
  13 #include <asm/system.h>
  14 #include <asm/bitops.h>
  15 #include <linux/config.h>
  16 #include <linux/types.h>
  17 #include <linux/kernel.h>
  18 #include <linux/sched.h>
  19 #include <linux/string.h>
  20 #include <linux/mm.h>
  21 #include <linux/socket.h>
  22 #include <linux/sockios.h>
  23 #include <linux/in.h>
  24 #include <linux/errno.h>
  25 #include <linux/interrupt.h>
  26 #include <linux/netdevice.h>
  27 #include <linux/skbuff.h>
  28 #include <linux/rtnetlink.h>
  29 #include <linux/init.h>
  30 #include <linux/proc_fs.h>
  31 #include <net/sock.h>
  32 #include <net/pkt_sched.h>
  33
  34 #define L2T(p,L)   ((p)->R_tab->data[(L)>>(p)->R_tab->rate.cell_log])
  35 #define L2T_P(p,L) ((p)->P_tab->data[(L)>>(p)->P_tab->rate.cell_log])
  36
  37 static u32 idx_gen;
  38 static struct tcf_police *tcf_police_ht[16];
  39 /* Policer hash table lock */
  40 static rwlock_t police_lock = RW_LOCK_UNLOCKED;
  41
  42 /* Each policer is serialized by its individual spinlock */
  43
  44 static __inline__ unsigned tcf_police_hash(u32 index)
  45 {
  46         return index&0xF;
  47 }
  48
  49 static __inline__ struct tcf_police * tcf_police_lookup(u32 index)
  50 {
  51         struct tcf_police *p;
  52
  53         read_lock(&police_lock);
  54         for (p = tcf_police_ht[tcf_police_hash(index)]; p; p = p->next) {
  55                 if (p->index == index)
  56                         break;
  57         }
  58         read_unlock(&police_lock);
  59         return p;
  60 }
  61
  62 static __inline__ u32 tcf_police_new_index(void)
  63 {
  64         do {
  65                 if (++idx_gen == 0)
  66                         idx_gen = 1;
  67         } while (tcf_police_lookup(idx_gen));
  68
  69         return idx_gen;
  70 }
  71
  72
  73 void tcf_police_destroy(struct tcf_police *p)
  74 {
  75         unsigned h = tcf_police_hash(p->index);
  76         struct tcf_police **p1p;
  77
  78         for (p1p = &tcf_police_ht[h]; *p1p; p1p = &(*p1p)->next) {
  79                 if (*p1p == p) {
  80                         write_lock_bh(&police_lock);
  81                         *p1p = p->next;
  82                         write_unlock_bh(&police_lock);
  83 #ifdef CONFIG_NET_ESTIMATOR
  84                         qdisc_kill_estimator(&p->stats);
  85 #endif
  86                         if (p->R_tab)
  87                                 qdisc_put_rtab(p->R_tab);
  88                         if (p->P_tab)
  89                                 qdisc_put_rtab(p->P_tab);
  90                         kfree(p);
  91                         return;
  92                 }
  93         }
  94         BUG_TRAP(0);
  95 }
  96
  97 struct tcf_police * tcf_police_locate(struct rtattr *rta, struct rtattr *est)
  98 {
  99         unsigned h;
 100         struct tcf_police *p;
 101         struct rtattr *tb[TCA_POLICE_MAX];
 102         struct tc_police *parm;
 103
 104         if (rtattr_parse(tb, TCA_POLICE_MAX, RTA_DATA(rta), RTA_PAYLOAD(rta)) < 0)
 105                 return NULL;
 106
 107         if (tb[TCA_POLICE_TBF-1] == NULL)
 108                 return NULL;
 109
 110         parm = RTA_DATA(tb[TCA_POLICE_TBF-1]);
 111
 112         if (parm->index && (p = tcf_police_lookup(parm->index)) != NULL) {
 113                 p->refcnt++;
 114                 return p;
 115         }
 116
 117         p = kmalloc(sizeof(*p), GFP_KERNEL);
 118         if (p == NULL)
 119                 return NULL;
 120
 121         memset(p, 0, sizeof(*p));
 122         p->refcnt = 1;
 123         spin_lock_init(&p->lock);
 124         p->stats.lock = &p->lock;
 125         if (parm->rate.rate) {
 126                 if ((p->R_tab = qdisc_get_rtab(&parm->rate, tb[TCA_POLICE_RATE-1])) == NULL)
 127                         goto failure;
 128                 if (parm->peakrate.rate &&
 129                     (p->P_tab = qdisc_get_rtab(&parm->peakrate, tb[TCA_POLICE_PEAKRATE-1])) == NULL)
 130                         goto failure;
 131         }
 132         if (tb[TCA_POLICE_RESULT-1])
 133                 p->result = *(int*)RTA_DATA(tb[TCA_POLICE_RESULT-1]);
 134 #ifdef CONFIG_NET_ESTIMATOR
 135         if (tb[TCA_POLICE_AVRATE-1])
 136                 p->ewma_rate = *(u32*)RTA_DATA(tb[TCA_POLICE_AVRATE-1]);
 137 #endif
 138         p->toks = p->burst = parm->burst;
 139         p->mtu = parm->mtu;
 140         if (p->mtu == 0) {
 141                 p->mtu = ~0;
 142                 if (p->R_tab)
 143                         p->mtu = 255<<p->R_tab->rate.cell_log;
 144         }
 145         if (p->P_tab)
 146                 p->ptoks = L2T_P(p, p->mtu);
 147         PSCHED_GET_TIME(p->t_c);
 148         p->index = parm->index ? : tcf_police_new_index();
 149         p->action = parm->action;
 150 #ifdef CONFIG_NET_ESTIMATOR
 151         if (est)
 152                 qdisc_new_estimator(&p->stats, est);
 153 #endif
 154         h = tcf_police_hash(p->index);
 155         write_lock_bh(&police_lock);
 156         p->next = tcf_police_ht[h];
 157         tcf_police_ht[h] = p;
 158         write_unlock_bh(&police_lock);
 159         return p;
 160
 161 failure:
 162         if (p->R_tab)
 163                 qdisc_put_rtab(p->R_tab);
 164         kfree(p);
 165         return NULL;
 166 }
 167
 168 int tcf_police(struct sk_buff *skb, struct tcf_police *p)
 169 {
 170         psched_time_t now;
 171         long toks;
 172         long ptoks = 0;
 173
 174         spin_lock(&p->lock);
 175
 176         p->stats.bytes += skb->len;
 177         p->stats.packets++;
 178
 179 #ifdef CONFIG_NET_ESTIMATOR
 180         if (p->ewma_rate && p->stats.bps >= p->ewma_rate) {
 181                 p->stats.overlimits++;
 182                 spin_unlock(&p->lock);
 183                 return p->action;
 184         }
 185 #endif
 186
 187         if (skb->len <= p->mtu) {
 188                 if (p->R_tab == NULL) {
 189                         spin_unlock(&p->lock);
 190                         return p->result;
 191                 }
 192
 193                 PSCHED_GET_TIME(now);
 194
 195                 toks = PSCHED_TDIFF_SAFE(now, p->t_c, p->burst, 0);
 196
 197                 if (p->P_tab) {
 198                         ptoks = toks + p->ptoks;
 199                         if (ptoks > (long)L2T_P(p, p->mtu))
 200                                 ptoks = (long)L2T_P(p, p->mtu);
 201                         ptoks -= L2T_P(p, skb->len);
 202                 }
 203                 toks += p->toks;
 204                 if (toks > (long)p->burst)
 205                         toks = p->burst;
 206                 toks -= L2T(p, skb->len);
 207
 208                 if ((toks|ptoks) >= 0) {
 209                         p->t_c = now;
 210                         p->toks = toks;
 211                         p->ptoks = ptoks;
 212                         spin_unlock(&p->lock);
 213                         return p->result;
 214                 }
 215         }
 216
 217         p->stats.overlimits++;
 218         spin_unlock(&p->lock);
 219         return p->action;
 220 }
 221
 222 int tcf_police_dump(struct sk_buff *skb, struct tcf_police *p)
 223 {
 224         unsigned char    *b = skb->tail;
 225         struct tc_police opt;
 226
 227         opt.index = p->index;
 228         opt.action = p->action;
 229         opt.mtu = p->mtu;
 230         opt.burst = p->burst;
 231         if (p->R_tab)
 232                 opt.rate = p->R_tab->rate;
 233         else
 234                 memset(&opt.rate, 0, sizeof(opt.rate));
 235         if (p->P_tab)
 236                 opt.peakrate = p->P_tab->rate;
 237         else
 238                 memset(&opt.peakrate, 0, sizeof(opt.peakrate));
 239         RTA_PUT(skb, TCA_POLICE_TBF, sizeof(opt), &opt);
 240         if (p->result)
 241                 RTA_PUT(skb, TCA_POLICE_RESULT, sizeof(int), &p->result);
 242 #ifdef CONFIG_NET_ESTIMATOR
 243         if (p->ewma_rate)
 244                 RTA_PUT(skb, TCA_POLICE_AVRATE, 4, &p->ewma_rate);
 245 #endif
 246         return skb->len;
 247
 248 rtattr_failure:
 249         skb_trim(skb, b - skb->data);
 250         return -1;
 251 }