pkt_sched: Use PSCHED_SHIFT in PSCHED time conversion
[safe/jmp/linux-2.6] / net / sched / act_police.c
index 6ffe35d..f8f047b 100644 (file)
  *             J Hadi Salim (action changes)
  */
 
-#include <asm/uaccess.h>
-#include <asm/system.h>
-#include <linux/bitops.h>
 #include <linux/module.h>
 #include <linux/types.h>
 #include <linux/kernel.h>
-#include <linux/sched.h>
 #include <linux/string.h>
-#include <linux/mm.h>
-#include <linux/socket.h>
-#include <linux/sockios.h>
-#include <linux/in.h>
 #include <linux/errno.h>
-#include <linux/interrupt.h>
-#include <linux/netdevice.h>
 #include <linux/skbuff.h>
-#include <linux/module.h>
 #include <linux/rtnetlink.h>
 #include <linux/init.h>
-#include <net/sock.h>
 #include <net/act_api.h>
+#include <net/netlink.h>
 
-#define L2T(p,L)   ((p)->tcfp_R_tab->data[(L)>>(p)->tcfp_R_tab->rate.cell_log])
-#define L2T_P(p,L) ((p)->tcfp_P_tab->data[(L)>>(p)->tcfp_P_tab->rate.cell_log])
+#define L2T(p,L)   qdisc_l2t((p)->tcfp_R_tab, L)
+#define L2T_P(p,L) qdisc_l2t((p)->tcfp_P_tab, L)
 
 #define POL_TAB_MASK     15
 static struct tcf_common *tcf_police_ht[POL_TAB_MASK + 1];
@@ -60,15 +49,14 @@ struct tc_police_compat
 
 /* Each policer is serialized by its individual spinlock */
 
-#ifdef CONFIG_NET_CLS_ACT
 static int tcf_act_police_walker(struct sk_buff *skb, struct netlink_callback *cb,
                              int type, struct tc_action *a)
 {
        struct tcf_common *p;
        int err = 0, index = -1, i = 0, s_i = 0, n_i = 0;
-       struct rtattr *r;
+       struct nlattr *nest;
 
-       read_lock(&police_lock);
+       read_lock_bh(&police_lock);
 
        s_i = cb->args[0];
 
@@ -81,34 +69,34 @@ static int tcf_act_police_walker(struct sk_buff *skb, struct netlink_callback *c
                                continue;
                        a->priv = p;
                        a->order = index;
-                       r = (struct rtattr*) skb->tail;
-                       RTA_PUT(skb, a->order, 0, NULL);
+                       nest = nla_nest_start(skb, a->order);
+                       if (nest == NULL)
+                               goto nla_put_failure;
                        if (type == RTM_DELACTION)
                                err = tcf_action_dump_1(skb, a, 0, 1);
                        else
                                err = tcf_action_dump_1(skb, a, 0, 0);
                        if (err < 0) {
                                index--;
-                               skb_trim(skb, (u8*)r - skb->data);
+                               nla_nest_cancel(skb, nest);
                                goto done;
                        }
-                       r->rta_len = skb->tail - (u8*)r;
+                       nla_nest_end(skb, nest);
                        n_i++;
                }
        }
 done:
-       read_unlock(&police_lock);
+       read_unlock_bh(&police_lock);
        if (n_i)
                cb->args[0] += n_i;
        return n_i;
 
-rtattr_failure:
-       skb_trim(skb, (u8*)r - skb->data);
+nla_put_failure:
+       nla_nest_cancel(skb, nest);
        goto done;
 }
-#endif
 
-void tcf_police_destroy(struct tcf_police *p)
+static void tcf_police_destroy(struct tcf_police *p)
 {
        unsigned int h = tcf_hash(p->tcf_index, POL_TAB_MASK);
        struct tcf_common **p1p;
@@ -118,10 +106,8 @@ void tcf_police_destroy(struct tcf_police *p)
                        write_lock_bh(&police_lock);
                        *p1p = p->tcf_next;
                        write_unlock_bh(&police_lock);
-#ifdef CONFIG_NET_ESTIMATOR
                        gen_kill_estimator(&p->tcf_bstats,
                                           &p->tcf_rate_est);
-#endif
                        if (p->tcfp_R_tab)
                                qdisc_put_rtab(p->tcfp_R_tab);
                        if (p->tcfp_P_tab)
@@ -130,37 +116,40 @@ void tcf_police_destroy(struct tcf_police *p)
                        return;
                }
        }
-       BUG_TRAP(0);
+       WARN_ON(1);
 }
 
-#ifdef CONFIG_NET_CLS_ACT
-static int tcf_act_police_locate(struct rtattr *rta, struct rtattr *est,
+static const struct nla_policy police_policy[TCA_POLICE_MAX + 1] = {
+       [TCA_POLICE_RATE]       = { .len = TC_RTAB_SIZE },
+       [TCA_POLICE_PEAKRATE]   = { .len = TC_RTAB_SIZE },
+       [TCA_POLICE_AVRATE]     = { .type = NLA_U32 },
+       [TCA_POLICE_RESULT]     = { .type = NLA_U32 },
+};
+
+static int tcf_act_police_locate(struct nlattr *nla, struct nlattr *est,
                                 struct tc_action *a, int ovr, int bind)
 {
        unsigned h;
        int ret = 0, err;
-       struct rtattr *tb[TCA_POLICE_MAX];
+       struct nlattr *tb[TCA_POLICE_MAX + 1];
        struct tc_police *parm;
        struct tcf_police *police;
        struct qdisc_rate_table *R_tab = NULL, *P_tab = NULL;
        int size;
 
-       if (rta == NULL || rtattr_parse_nested(tb, TCA_POLICE_MAX, rta) < 0)
+       if (nla == NULL)
                return -EINVAL;
 
-       if (tb[TCA_POLICE_TBF-1] == NULL)
-               return -EINVAL;
-       size = RTA_PAYLOAD(tb[TCA_POLICE_TBF-1]);
-       if (size != sizeof(*parm) && size != sizeof(struct tc_police_compat))
-               return -EINVAL;
-       parm = RTA_DATA(tb[TCA_POLICE_TBF-1]);
+       err = nla_parse_nested(tb, TCA_POLICE_MAX, nla, police_policy);
+       if (err < 0)
+               return err;
 
-       if (tb[TCA_POLICE_RESULT-1] != NULL &&
-           RTA_PAYLOAD(tb[TCA_POLICE_RESULT-1]) != sizeof(u32))
+       if (tb[TCA_POLICE_TBF] == NULL)
                return -EINVAL;
-       if (tb[TCA_POLICE_RESULT-1] != NULL &&
-           RTA_PAYLOAD(tb[TCA_POLICE_RESULT-1]) != sizeof(u32))
+       size = nla_len(tb[TCA_POLICE_TBF]);
+       if (size != sizeof(*parm) && size != sizeof(struct tc_police_compat))
                return -EINVAL;
+       parm = nla_data(tb[TCA_POLICE_TBF]);
 
        if (parm->index) {
                struct tcf_common *pc;
@@ -185,26 +174,39 @@ static int tcf_act_police_locate(struct rtattr *rta, struct rtattr *est,
        ret = ACT_P_CREATED;
        police->tcf_refcnt = 1;
        spin_lock_init(&police->tcf_lock);
-       police->tcf_stats_lock = &police->tcf_lock;
        if (bind)
                police->tcf_bindcnt = 1;
 override:
        if (parm->rate.rate) {
                err = -ENOMEM;
-               R_tab = qdisc_get_rtab(&parm->rate, tb[TCA_POLICE_RATE-1]);
+               R_tab = qdisc_get_rtab(&parm->rate, tb[TCA_POLICE_RATE]);
                if (R_tab == NULL)
                        goto failure;
+
                if (parm->peakrate.rate) {
                        P_tab = qdisc_get_rtab(&parm->peakrate,
-                                              tb[TCA_POLICE_PEAKRATE-1]);
-                       if (P_tab == NULL) {
-                               qdisc_put_rtab(R_tab);
+                                              tb[TCA_POLICE_PEAKRATE]);
+                       if (P_tab == NULL)
                                goto failure;
-                       }
                }
        }
-       /* No failure allowed after this point */
+
        spin_lock_bh(&police->tcf_lock);
+       if (est) {
+               err = gen_replace_estimator(&police->tcf_bstats,
+                                           &police->tcf_rate_est,
+                                           &police->tcf_lock, est);
+               if (err)
+                       goto failure_unlock;
+       } else if (tb[TCA_POLICE_AVRATE] &&
+                  (ret == ACT_P_CREATED ||
+                   !gen_estimator_active(&police->tcf_bstats,
+                                         &police->tcf_rate_est))) {
+               err = -EINVAL;
+               goto failure_unlock;
+       }
+
+       /* No failure allowed after this point */
        if (R_tab != NULL) {
                qdisc_put_rtab(police->tcfp_R_tab);
                police->tcfp_R_tab = R_tab;
@@ -214,8 +216,8 @@ override:
                police->tcfp_P_tab = P_tab;
        }
 
-       if (tb[TCA_POLICE_RESULT-1])
-               police->tcfp_result = *(u32*)RTA_DATA(tb[TCA_POLICE_RESULT-1]);
+       if (tb[TCA_POLICE_RESULT])
+               police->tcfp_result = nla_get_u32(tb[TCA_POLICE_RESULT]);
        police->tcfp_toks = police->tcfp_burst = parm->burst;
        police->tcfp_mtu = parm->mtu;
        if (police->tcfp_mtu == 0) {
@@ -227,21 +229,14 @@ override:
                police->tcfp_ptoks = L2T_P(police, police->tcfp_mtu);
        police->tcf_action = parm->action;
 
-#ifdef CONFIG_NET_ESTIMATOR
-       if (tb[TCA_POLICE_AVRATE-1])
-               police->tcfp_ewma_rate =
-                       *(u32*)RTA_DATA(tb[TCA_POLICE_AVRATE-1]);
-       if (est)
-               gen_replace_estimator(&police->tcf_bstats,
-                                     &police->tcf_rate_est,
-                                     police->tcf_stats_lock, est);
-#endif
+       if (tb[TCA_POLICE_AVRATE])
+               police->tcfp_ewma_rate = nla_get_u32(tb[TCA_POLICE_AVRATE]);
 
        spin_unlock_bh(&police->tcf_lock);
        if (ret != ACT_P_CREATED)
                return ret;
 
-       PSCHED_GET_TIME(police->tcfp_t_c);
+       police->tcfp_t_c = psched_get_time();
        police->tcf_index = parm->index ? parm->index :
                tcf_hash_new_index(&police_idx_gen, &police_hash_info);
        h = tcf_hash(police->tcf_index, POL_TAB_MASK);
@@ -253,7 +248,13 @@ override:
        a->priv = police;
        return ret;
 
+failure_unlock:
+       spin_unlock_bh(&police->tcf_lock);
 failure:
+       if (P_tab)
+               qdisc_put_rtab(P_tab);
+       if (R_tab)
+               qdisc_put_rtab(R_tab);
        if (ret == ACT_P_CREATED)
                kfree(police);
        return err;
@@ -262,10 +263,19 @@ failure:
 static int tcf_act_police_cleanup(struct tc_action *a, int bind)
 {
        struct tcf_police *p = a->priv;
+       int ret = 0;
 
-       if (p != NULL)
-               return tcf_police_release(p, bind);
-       return 0;
+       if (p != NULL) {
+               if (bind)
+                       p->tcf_bindcnt--;
+
+               p->tcf_refcnt--;
+               if (p->tcf_refcnt <= 0 && !p->tcf_bindcnt) {
+                       tcf_police_destroy(p);
+                       ret = 1;
+               }
+       }
+       return ret;
 }
 
 static int tcf_act_police(struct sk_buff *skb, struct tc_action *a,
@@ -278,38 +288,35 @@ static int tcf_act_police(struct sk_buff *skb, struct tc_action *a,
 
        spin_lock(&police->tcf_lock);
 
-       police->tcf_bstats.bytes += skb->len;
+       police->tcf_bstats.bytes += qdisc_pkt_len(skb);
        police->tcf_bstats.packets++;
 
-#ifdef CONFIG_NET_ESTIMATOR
        if (police->tcfp_ewma_rate &&
            police->tcf_rate_est.bps >= police->tcfp_ewma_rate) {
                police->tcf_qstats.overlimits++;
                spin_unlock(&police->tcf_lock);
                return police->tcf_action;
        }
-#endif
 
-       if (skb->len <= police->tcfp_mtu) {
+       if (qdisc_pkt_len(skb) <= police->tcfp_mtu) {
                if (police->tcfp_R_tab == NULL) {
                        spin_unlock(&police->tcf_lock);
                        return police->tcfp_result;
                }
 
-               PSCHED_GET_TIME(now);
-
-               toks = PSCHED_TDIFF_SAFE(now, police->tcfp_t_c,
-                                        police->tcfp_burst);
+               now = psched_get_time();
+               toks = psched_tdiff_bounded(now, police->tcfp_t_c,
+                                           police->tcfp_burst);
                if (police->tcfp_P_tab) {
                        ptoks = toks + police->tcfp_ptoks;
                        if (ptoks > (long)L2T_P(police, police->tcfp_mtu))
                                ptoks = (long)L2T_P(police, police->tcfp_mtu);
-                       ptoks -= L2T_P(police, skb->len);
+                       ptoks -= L2T_P(police, qdisc_pkt_len(skb));
                }
                toks += police->tcfp_toks;
                if (toks > (long)police->tcfp_burst)
                        toks = police->tcfp_burst;
-               toks -= L2T(police, skb->len);
+               toks -= L2T(police, qdisc_pkt_len(skb));
                if ((toks|ptoks) >= 0) {
                        police->tcfp_t_c = now;
                        police->tcfp_toks = toks;
@@ -327,7 +334,7 @@ static int tcf_act_police(struct sk_buff *skb, struct tc_action *a,
 static int
 tcf_act_police_dump(struct sk_buff *skb, struct tc_action *a, int bind, int ref)
 {
-       unsigned char    *b = skb->tail;
+       unsigned char *b = skb_tail_pointer(skb);
        struct tcf_police *police = a->priv;
        struct tc_police opt;
 
@@ -345,18 +352,15 @@ tcf_act_police_dump(struct sk_buff *skb, struct tc_action *a, int bind, int ref)
                opt.peakrate = police->tcfp_P_tab->rate;
        else
                memset(&opt.peakrate, 0, sizeof(opt.peakrate));
-       RTA_PUT(skb, TCA_POLICE_TBF, sizeof(opt), &opt);
+       NLA_PUT(skb, TCA_POLICE_TBF, sizeof(opt), &opt);
        if (police->tcfp_result)
-               RTA_PUT(skb, TCA_POLICE_RESULT, sizeof(int),
-                       &police->tcfp_result);
-#ifdef CONFIG_NET_ESTIMATOR
+               NLA_PUT_U32(skb, TCA_POLICE_RESULT, police->tcfp_result);
        if (police->tcfp_ewma_rate)
-               RTA_PUT(skb, TCA_POLICE_AVRATE, 4, &police->tcfp_ewma_rate);
-#endif
+               NLA_PUT_U32(skb, TCA_POLICE_AVRATE, police->tcfp_ewma_rate);
        return skb->len;
 
-rtattr_failure:
-       skb_trim(skb, b - skb->data);
+nla_put_failure:
+       nlmsg_trim(skb, b);
        return -1;
 }
 
@@ -392,240 +396,3 @@ police_cleanup_module(void)
 
 module_init(police_init_module);
 module_exit(police_cleanup_module);
-
-#else /* CONFIG_NET_CLS_ACT */
-
-static struct tcf_common *tcf_police_lookup(u32 index)
-{
-       struct tcf_hashinfo *hinfo = &police_hash_info;
-       struct tcf_common *p;
-
-       read_lock(hinfo->lock);
-       for (p = hinfo->htab[tcf_hash(index, hinfo->hmask)]; p;
-            p = p->tcfc_next) {
-               if (p->tcfc_index == index)
-                       break;
-       }
-       read_unlock(hinfo->lock);
-
-       return p;
-}
-
-static u32 tcf_police_new_index(void)
-{
-       u32 *idx_gen = &police_idx_gen;
-       u32 val = *idx_gen;
-
-       do {
-               if (++val == 0)
-                       val = 1;
-       } while (tcf_police_lookup(val));
-
-       return (*idx_gen = val);
-}
-
-struct tcf_police *tcf_police_locate(struct rtattr *rta, struct rtattr *est)
-{
-       unsigned int h;
-       struct tcf_police *police;
-       struct rtattr *tb[TCA_POLICE_MAX];
-       struct tc_police *parm;
-       int size;
-
-       if (rtattr_parse_nested(tb, TCA_POLICE_MAX, rta) < 0)
-               return NULL;
-
-       if (tb[TCA_POLICE_TBF-1] == NULL)
-               return NULL;
-       size = RTA_PAYLOAD(tb[TCA_POLICE_TBF-1]);
-       if (size != sizeof(*parm) && size != sizeof(struct tc_police_compat))
-               return NULL;
-
-       parm = RTA_DATA(tb[TCA_POLICE_TBF-1]);
-
-       if (parm->index) {
-               struct tcf_common *pc;
-
-               pc = tcf_police_lookup(parm->index);
-               if (pc) {
-                       police = to_police(pc);
-                       police->tcf_refcnt++;
-                       return police;
-               }
-       }
-       police = kzalloc(sizeof(*police), GFP_KERNEL);
-       if (unlikely(!police))
-               return NULL;
-
-       police->tcf_refcnt = 1;
-       spin_lock_init(&police->tcf_lock);
-       police->tcf_stats_lock = &police->tcf_lock;
-       if (parm->rate.rate) {
-               police->tcfp_R_tab =
-                       qdisc_get_rtab(&parm->rate, tb[TCA_POLICE_RATE-1]);
-               if (police->tcfp_R_tab == NULL)
-                       goto failure;
-               if (parm->peakrate.rate) {
-                       police->tcfp_P_tab =
-                               qdisc_get_rtab(&parm->peakrate,
-                                              tb[TCA_POLICE_PEAKRATE-1]);
-                       if (police->tcfp_P_tab == NULL)
-                               goto failure;
-               }
-       }
-       if (tb[TCA_POLICE_RESULT-1]) {
-               if (RTA_PAYLOAD(tb[TCA_POLICE_RESULT-1]) != sizeof(u32))
-                       goto failure;
-               police->tcfp_result = *(u32*)RTA_DATA(tb[TCA_POLICE_RESULT-1]);
-       }
-#ifdef CONFIG_NET_ESTIMATOR
-       if (tb[TCA_POLICE_AVRATE-1]) {
-               if (RTA_PAYLOAD(tb[TCA_POLICE_AVRATE-1]) != sizeof(u32))
-                       goto failure;
-               police->tcfp_ewma_rate =
-                       *(u32*)RTA_DATA(tb[TCA_POLICE_AVRATE-1]);
-       }
-#endif
-       police->tcfp_toks = police->tcfp_burst = parm->burst;
-       police->tcfp_mtu = parm->mtu;
-       if (police->tcfp_mtu == 0) {
-               police->tcfp_mtu = ~0;
-               if (police->tcfp_R_tab)
-                       police->tcfp_mtu = 255<<police->tcfp_R_tab->rate.cell_log;
-       }
-       if (police->tcfp_P_tab)
-               police->tcfp_ptoks = L2T_P(police, police->tcfp_mtu);
-       PSCHED_GET_TIME(police->tcfp_t_c);
-       police->tcf_index = parm->index ? parm->index :
-               tcf_police_new_index();
-       police->tcf_action = parm->action;
-#ifdef CONFIG_NET_ESTIMATOR
-       if (est)
-               gen_new_estimator(&police->tcf_bstats, &police->tcf_rate_est,
-                                 police->tcf_stats_lock, est);
-#endif
-       h = tcf_hash(police->tcf_index, POL_TAB_MASK);
-       write_lock_bh(&police_lock);
-       police->tcf_next = tcf_police_ht[h];
-       tcf_police_ht[h] = &police->common;
-       write_unlock_bh(&police_lock);
-       return police;
-
-failure:
-       if (police->tcfp_R_tab)
-               qdisc_put_rtab(police->tcfp_R_tab);
-       kfree(police);
-       return NULL;
-}
-
-int tcf_police(struct sk_buff *skb, struct tcf_police *police)
-{
-       psched_time_t now;
-       long toks;
-       long ptoks = 0;
-
-       spin_lock(&police->tcf_lock);
-
-       police->tcf_bstats.bytes += skb->len;
-       police->tcf_bstats.packets++;
-
-#ifdef CONFIG_NET_ESTIMATOR
-       if (police->tcfp_ewma_rate &&
-           police->tcf_rate_est.bps >= police->tcfp_ewma_rate) {
-               police->tcf_qstats.overlimits++;
-               spin_unlock(&police->tcf_lock);
-               return police->tcf_action;
-       }
-#endif
-       if (skb->len <= police->tcfp_mtu) {
-               if (police->tcfp_R_tab == NULL) {
-                       spin_unlock(&police->tcf_lock);
-                       return police->tcfp_result;
-               }
-
-               PSCHED_GET_TIME(now);
-               toks = PSCHED_TDIFF_SAFE(now, police->tcfp_t_c,
-                                        police->tcfp_burst);
-               if (police->tcfp_P_tab) {
-                       ptoks = toks + police->tcfp_ptoks;
-                       if (ptoks > (long)L2T_P(police, police->tcfp_mtu))
-                               ptoks = (long)L2T_P(police, police->tcfp_mtu);
-                       ptoks -= L2T_P(police, skb->len);
-               }
-               toks += police->tcfp_toks;
-               if (toks > (long)police->tcfp_burst)
-                       toks = police->tcfp_burst;
-               toks -= L2T(police, skb->len);
-               if ((toks|ptoks) >= 0) {
-                       police->tcfp_t_c = now;
-                       police->tcfp_toks = toks;
-                       police->tcfp_ptoks = ptoks;
-                       spin_unlock(&police->tcf_lock);
-                       return police->tcfp_result;
-               }
-       }
-
-       police->tcf_qstats.overlimits++;
-       spin_unlock(&police->tcf_lock);
-       return police->tcf_action;
-}
-EXPORT_SYMBOL(tcf_police);
-
-int tcf_police_dump(struct sk_buff *skb, struct tcf_police *police)
-{
-       unsigned char *b = skb->tail;
-       struct tc_police opt;
-
-       opt.index = police->tcf_index;
-       opt.action = police->tcf_action;
-       opt.mtu = police->tcfp_mtu;
-       opt.burst = police->tcfp_burst;
-       if (police->tcfp_R_tab)
-               opt.rate = police->tcfp_R_tab->rate;
-       else
-               memset(&opt.rate, 0, sizeof(opt.rate));
-       if (police->tcfp_P_tab)
-               opt.peakrate = police->tcfp_P_tab->rate;
-       else
-               memset(&opt.peakrate, 0, sizeof(opt.peakrate));
-       RTA_PUT(skb, TCA_POLICE_TBF, sizeof(opt), &opt);
-       if (police->tcfp_result)
-               RTA_PUT(skb, TCA_POLICE_RESULT, sizeof(int),
-                       &police->tcfp_result);
-#ifdef CONFIG_NET_ESTIMATOR
-       if (police->tcfp_ewma_rate)
-               RTA_PUT(skb, TCA_POLICE_AVRATE, 4, &police->tcfp_ewma_rate);
-#endif
-       return skb->len;
-
-rtattr_failure:
-       skb_trim(skb, b - skb->data);
-       return -1;
-}
-
-int tcf_police_dump_stats(struct sk_buff *skb, struct tcf_police *police)
-{
-       struct gnet_dump d;
-
-       if (gnet_stats_start_copy_compat(skb, TCA_STATS2, TCA_STATS,
-                                        TCA_XSTATS, police->tcf_stats_lock,
-                                        &d) < 0)
-               goto errout;
-
-       if (gnet_stats_copy_basic(&d, &police->tcf_bstats) < 0 ||
-#ifdef CONFIG_NET_ESTIMATOR
-           gnet_stats_copy_rate_est(&d, &police->tcf_rate_est) < 0 ||
-#endif
-           gnet_stats_copy_queue(&d, &police->tcf_qstats) < 0)
-               goto errout;
-
-       if (gnet_stats_finish_copy(&d) < 0)
-               goto errout;
-
-       return 0;
-
-errout:
-       return -1;
-}
-
-#endif /* CONFIG_NET_CLS_ACT */