microblaze: uaccess: fix put_user and get_user macros
[linux-2.6.git] / net / sched / sch_gred.c
index 4ced47b..40408d5 100644 (file)
  *                      from Ren Liu
  *                    - More error checks
  *
- *
- *
- *  For all the glorious comments look at Alexey's sch_red.c
+ *  For all the glorious comments look at include/net/red.h
  */
 
-#include <linux/config.h>
 #include <linux/module.h>
-#include <asm/uaccess.h>
-#include <asm/system.h>
-#include <linux/bitops.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/if_ether.h>
-#include <linux/inet.h>
-#include <linux/netdevice.h>
-#include <linux/etherdevice.h>
-#include <linux/notifier.h>
-#include <net/ip.h>
-#include <net/route.h>
 #include <linux/skbuff.h>
-#include <net/sock.h>
 #include <net/pkt_sched.h>
+#include <net/red.h>
 
-#if 1 /* control */
-#define DPRINTK(format,args...) printk(KERN_DEBUG format,##args)
-#else
-#define DPRINTK(format,args...)
-#endif
-
-#if 0 /* data */
-#define D2PRINTK(format,args...) printk(KERN_DEBUG format,##args)
-#else
-#define D2PRINTK(format,args...)
-#endif
+#define GRED_DEF_PRIO (MAX_DPs / 2)
+#define GRED_VQ_MASK (MAX_DPs - 1)
 
 struct gred_sched_data;
 struct gred_sched;
 
 struct gred_sched_data
 {
-/* Parameters */
        u32             limit;          /* HARD maximal queue length    */
-       u32             qth_min;        /* Min average length threshold: A scaled */
-       u32             qth_max;        /* Max average length threshold: A scaled */
        u32             DP;             /* the drop pramaters */
-       char            Wlog;           /* log(W)               */
-       char            Plog;           /* random number bits   */
-       u32             Scell_max;
-       u32             Rmask;
        u32             bytesin;        /* bytes seen on virtualQ so far*/
        u32             packetsin;      /* packets seen on virtualQ so far*/
        u32             backlog;        /* bytes on the virtualQ */
-       u32             forced; /* packets dropped for exceeding limits */
-       u32             early;  /* packets dropped as a warning */
-       u32             other;  /* packets dropped by invoking drop() */
-       u32             pdrop;  /* packets dropped because we exceeded physical queue limits */
-       char            Scell_log;
-       u8              Stab[256];
-       u8              prio;        /* the prio of this vq */
-
-/* Variables */
-       unsigned long   qave;           /* Average queue length: A scaled */
-       int             qcount;         /* Packets since last random number generation */
-       u32             qR;             /* Cached random number */
-
-       psched_time_t   qidlestart;     /* Start of idle period */
+       u8              prio;           /* the prio of this vq */
+
+       struct red_parms parms;
+       struct red_stats stats;
 };
 
 enum {
        GRED_WRED_MODE = 1,
+       GRED_RIO_MODE,
 };
 
 struct gred_sched
 {
        struct gred_sched_data *tab[MAX_DPs];
        unsigned long   flags;
-       u32             DPs;   
-       u32             def; 
-       u8              initd; 
-       u8              grio; 
+       u32             red_flags;
+       u32             DPs;
+       u32             def;
+       struct red_parms wred_set;
 };
 
 static inline int gred_wred_mode(struct gred_sched *table)
@@ -120,6 +74,21 @@ static inline void gred_disable_wred_mode(struct gred_sched *table)
        __clear_bit(GRED_WRED_MODE, &table->flags);
 }
 
+static inline int gred_rio_mode(struct gred_sched *table)
+{
+       return test_bit(GRED_RIO_MODE, &table->flags);
+}
+
+static inline void gred_enable_rio_mode(struct gred_sched *table)
+{
+       __set_bit(GRED_RIO_MODE, &table->flags);
+}
+
+static inline void gred_disable_rio_mode(struct gred_sched *table)
+{
+       __clear_bit(GRED_RIO_MODE, &table->flags);
+}
+
 static inline int gred_wred_mode_check(struct Qdisc *sch)
 {
        struct gred_sched *table = qdisc_priv(sch);
@@ -142,161 +111,163 @@ static inline int gred_wred_mode_check(struct Qdisc *sch)
        return 0;
 }
 
-static int
-gred_enqueue(struct sk_buff *skb, struct Qdisc* sch)
+static inline unsigned int gred_backlog(struct gred_sched *table,
+                                       struct gred_sched_data *q,
+                                       struct Qdisc *sch)
 {
-       psched_time_t now;
-       struct gred_sched_data *q=NULL;
-       struct gred_sched *t= qdisc_priv(sch);
-       unsigned long   qave=0; 
-       int i=0;
+       if (gred_wred_mode(table))
+               return sch->qstats.backlog;
+       else
+               return q->backlog;
+}
 
-       if (!t->initd && skb_queue_len(&sch->q) < (sch->dev->tx_queue_len ? : 1)) {
-               D2PRINTK("NO GRED Queues setup yet! Enqueued anyway\n");
-               goto do_enqueue;
-       }
+static inline u16 tc_index_to_dp(struct sk_buff *skb)
+{
+       return skb->tc_index & GRED_VQ_MASK;
+}
 
+static inline void gred_load_wred_set(struct gred_sched *table,
+                                     struct gred_sched_data *q)
+{
+       q->parms.qavg = table->wred_set.qavg;
+       q->parms.qidlestart = table->wred_set.qidlestart;
+}
 
-       if ( ((skb->tc_index&0xf) > (t->DPs -1)) || !(q=t->tab[skb->tc_index&0xf])) {
-               printk("GRED: setting to default (%d)\n ",t->def);
-               if (!(q=t->tab[t->def])) {
-                       DPRINTK("GRED: setting to default FAILED! dropping!! "
-                           "(%d)\n ", t->def);
-                       goto drop;
+static inline void gred_store_wred_set(struct gred_sched *table,
+                                      struct gred_sched_data *q)
+{
+       table->wred_set.qavg = q->parms.qavg;
+}
+
+static inline int gred_use_ecn(struct gred_sched *t)
+{
+       return t->red_flags & TC_RED_ECN;
+}
+
+static inline int gred_use_harddrop(struct gred_sched *t)
+{
+       return t->red_flags & TC_RED_HARDDROP;
+}
+
+static int gred_enqueue(struct sk_buff *skb, struct Qdisc* sch)
+{
+       struct gred_sched_data *q=NULL;
+       struct gred_sched *t= qdisc_priv(sch);
+       unsigned long qavg = 0;
+       u16 dp = tc_index_to_dp(skb);
+
+       if (dp >= t->DPs  || (q = t->tab[dp]) == NULL) {
+               dp = t->def;
+
+               if ((q = t->tab[dp]) == NULL) {
+                       /* Pass through packets not assigned to a DP
+                        * if no default DP has been configured. This
+                        * allows for DP flows to be left untouched.
+                        */
+                       if (skb_queue_len(&sch->q) < qdisc_dev(sch)->tx_queue_len)
+                               return qdisc_enqueue_tail(skb, sch);
+                       else
+                               goto drop;
                }
+
                /* fix tc_index? --could be controvesial but needed for
                   requeueing */
-               skb->tc_index=(skb->tc_index&0xfffffff0) | t->def;
+               skb->tc_index = (skb->tc_index & ~GRED_VQ_MASK) | dp;
        }
 
-       D2PRINTK("gred_enqueue virtualQ 0x%x classid %x backlog %d "
-           "general backlog %d\n",skb->tc_index&0xf,sch->handle,q->backlog,
-           sch->qstats.backlog);
-       /* sum up all the qaves of prios <= to ours to get the new qave*/
-       if (!gred_wred_mode(t) && t->grio) {
-               for (i=0;i<t->DPs;i++) {
-                       if ((!t->tab[i]) || (i==q->DP)) 
-                               continue; 
-                               
-                       if ((t->tab[i]->prio < q->prio) && (PSCHED_IS_PASTPERFECT(t->tab[i]->qidlestart)))
-                               qave +=t->tab[i]->qave;
+       /* sum up all the qaves of prios <= to ours to get the new qave */
+       if (!gred_wred_mode(t) && gred_rio_mode(t)) {
+               int i;
+
+               for (i = 0; i < t->DPs; i++) {
+                       if (t->tab[i] && t->tab[i]->prio < q->prio &&
+                           !red_is_idling(&t->tab[i]->parms))
+                               qavg +=t->tab[i]->parms.qavg;
                }
-                       
+
        }
 
        q->packetsin++;
-       q->bytesin+=skb->len;
-
-       if (gred_wred_mode(t)) {
-               qave=0;
-               q->qave=t->tab[t->def]->qave;
-               q->qidlestart=t->tab[t->def]->qidlestart;
-       }
+       q->bytesin += qdisc_pkt_len(skb);
 
-       if (!PSCHED_IS_PASTPERFECT(q->qidlestart)) {
-               long us_idle;
-               PSCHED_GET_TIME(now);
-               us_idle = PSCHED_TDIFF_SAFE(now, q->qidlestart, q->Scell_max);
-               PSCHED_SET_PASTPERFECT(q->qidlestart);
+       if (gred_wred_mode(t))
+               gred_load_wred_set(t, q);
 
-               q->qave >>= q->Stab[(us_idle>>q->Scell_log)&0xFF];
-       } else {
-               if (gred_wred_mode(t)) {
-                       q->qave += sch->qstats.backlog - (q->qave >> q->Wlog);
-               } else {
-                       q->qave += q->backlog - (q->qave >> q->Wlog);
-               }
+       q->parms.qavg = red_calc_qavg(&q->parms, gred_backlog(t, q, sch));
 
-       }
-       
+       if (red_is_idling(&q->parms))
+               red_end_of_idle_period(&q->parms);
 
        if (gred_wred_mode(t))
-               t->tab[t->def]->qave=q->qave;
-
-       if ((q->qave+qave) < q->qth_min) {
-               q->qcount = -1;
-enqueue:
-               if (q->backlog + skb->len <= q->limit) {
-                       q->backlog += skb->len;
-do_enqueue:
-                       __skb_queue_tail(&sch->q, skb);
-                       sch->qstats.backlog += skb->len;
-                       sch->bstats.bytes += skb->len;
-                       sch->bstats.packets++;
-                       return 0;
-               } else {
-                       q->pdrop++;
-               }
+               gred_store_wred_set(t, q);
 
-drop:
-               kfree_skb(skb);
-               sch->qstats.drops++;
-               return NET_XMIT_DROP;
-       }
-       if ((q->qave+qave) >= q->qth_max) {
-               q->qcount = -1;
-               sch->qstats.overlimits++;
-               q->forced++;
-               goto drop;
+       switch (red_action(&q->parms, q->parms.qavg + qavg)) {
+               case RED_DONT_MARK:
+                       break;
+
+               case RED_PROB_MARK:
+                       sch->qstats.overlimits++;
+                       if (!gred_use_ecn(t) || !INET_ECN_set_ce(skb)) {
+                               q->stats.prob_drop++;
+                               goto congestion_drop;
+                       }
+
+                       q->stats.prob_mark++;
+                       break;
+
+               case RED_HARD_MARK:
+                       sch->qstats.overlimits++;
+                       if (gred_use_harddrop(t) || !gred_use_ecn(t) ||
+                           !INET_ECN_set_ce(skb)) {
+                               q->stats.forced_drop++;
+                               goto congestion_drop;
+                       }
+                       q->stats.forced_mark++;
+                       break;
        }
-       if (++q->qcount) {
-               if ((((qave+q->qave) - q->qth_min)>>q->Wlog)*q->qcount < q->qR)
-                       goto enqueue;
-               q->qcount = 0;
-               q->qR = net_random()&q->Rmask;
-               sch->qstats.overlimits++;
-               q->early++;
-               goto drop;
+
+       if (q->backlog + qdisc_pkt_len(skb) <= q->limit) {
+               q->backlog += qdisc_pkt_len(skb);
+               return qdisc_enqueue_tail(skb, sch);
        }
-       q->qR = net_random()&q->Rmask;
-       goto enqueue;
-}
 
-static int
-gred_requeue(struct sk_buff *skb, struct Qdisc* sch)
-{
-       struct gred_sched_data *q;
-       struct gred_sched *t= qdisc_priv(sch);
-       q= t->tab[(skb->tc_index&0xf)];
-/* error checking here -- probably unnecessary */
-       PSCHED_SET_PASTPERFECT(q->qidlestart);
-
-       __skb_queue_head(&sch->q, skb);
-       sch->qstats.backlog += skb->len;
-       sch->qstats.requeues++;
-       q->backlog += skb->len;
-       return 0;
+       q->stats.pdrop++;
+drop:
+       return qdisc_drop(skb, sch);
+
+congestion_drop:
+       qdisc_drop(skb, sch);
+       return NET_XMIT_CN;
 }
 
-static struct sk_buff *
-gred_dequeue(struct Qdisc* sch)
+static struct sk_buff *gred_dequeue(struct Qdisc* sch)
 {
        struct sk_buff *skb;
-       struct gred_sched_data *q;
-       struct gred_sched *t= qdisc_priv(sch);
+       struct gred_sched *t = qdisc_priv(sch);
+
+       skb = qdisc_dequeue_head(sch);
 
-       skb = __skb_dequeue(&sch->q);
        if (skb) {
-               sch->qstats.backlog -= skb->len;
-               q= t->tab[(skb->tc_index&0xf)];
-               if (q) {
-                       q->backlog -= skb->len;
-                       if (!q->backlog && !gred_wred_mode(t))
-                               PSCHED_GET_TIME(q->qidlestart);
+               struct gred_sched_data *q;
+               u16 dp = tc_index_to_dp(skb);
+
+               if (dp >= t->DPs || (q = t->tab[dp]) == NULL) {
+                       if (net_ratelimit())
+                               printk(KERN_WARNING "GRED: Unable to relocate "
+                                      "VQ 0x%x after dequeue, screwing up "
+                                      "backlog.\n", tc_index_to_dp(skb));
                } else {
-                       D2PRINTK("gred_dequeue: skb has bad tcindex %x\n",skb->tc_index&0xf); 
+                       q->backlog -= qdisc_pkt_len(skb);
+
+                       if (!q->backlog && !gred_wred_mode(t))
+                               red_start_of_idle_period(&q->parms);
                }
+
                return skb;
        }
 
-       if (gred_wred_mode(t)) {
-                       q= t->tab[t->def];
-                       if (!q) 
-                               D2PRINTK("no default VQ set: Results will be "
-                                      "screwed up\n");
-                       else
-                               PSCHED_GET_TIME(q->qidlestart);
-       }
+       if (gred_wred_mode(t) && !red_is_idling(&t->wred_set))
+               red_start_of_idle_period(&t->wred_set);
 
        return NULL;
 }
@@ -304,36 +275,34 @@ gred_dequeue(struct Qdisc* sch)
 static unsigned int gred_drop(struct Qdisc* sch)
 {
        struct sk_buff *skb;
+       struct gred_sched *t = qdisc_priv(sch);
 
-       struct gred_sched_data *q;
-       struct gred_sched *t= qdisc_priv(sch);
-
-       skb = __skb_dequeue_tail(&sch->q);
+       skb = qdisc_dequeue_tail(sch);
        if (skb) {
-               unsigned int len = skb->len;
-               sch->qstats.backlog -= len;
-               sch->qstats.drops++;
-               q= t->tab[(skb->tc_index&0xf)];
-               if (q) {
+               unsigned int len = qdisc_pkt_len(skb);
+               struct gred_sched_data *q;
+               u16 dp = tc_index_to_dp(skb);
+
+               if (dp >= t->DPs || (q = t->tab[dp]) == NULL) {
+                       if (net_ratelimit())
+                               printk(KERN_WARNING "GRED: Unable to relocate "
+                                      "VQ 0x%x while dropping, screwing up "
+                                      "backlog.\n", tc_index_to_dp(skb));
+               } else {
                        q->backlog -= len;
-                       q->other++;
+                       q->stats.other++;
+
                        if (!q->backlog && !gred_wred_mode(t))
-                               PSCHED_GET_TIME(q->qidlestart);
-               } else {
-                       D2PRINTK("gred_dequeue: skb has bad tcindex %x\n",skb->tc_index&0xf); 
+                               red_start_of_idle_period(&q->parms);
                }
 
-               kfree_skb(skb);
+               qdisc_drop(skb, sch);
                return len;
        }
 
-       q=t->tab[t->def];
-       if (!q) {
-               D2PRINTK("no default VQ set: Results might be screwed up\n");
-               return 0;
-       }
+       if (gred_wred_mode(t) && !red_is_idling(&t->wred_set))
+               red_start_of_idle_period(&t->wred_set);
 
-       PSCHED_GET_TIME(q->qidlestart);
        return 0;
 
 }
@@ -341,295 +310,260 @@ static unsigned int gred_drop(struct Qdisc* sch)
 static void gred_reset(struct Qdisc* sch)
 {
        int i;
-       struct gred_sched_data *q;
-       struct gred_sched *t= qdisc_priv(sch);
+       struct gred_sched *t = qdisc_priv(sch);
 
-       __skb_queue_purge(&sch->q);
+       qdisc_reset_queue(sch);
 
-       sch->qstats.backlog = 0;
+       for (i = 0; i < t->DPs; i++) {
+               struct gred_sched_data *q = t->tab[i];
 
-        for (i=0;i<t->DPs;i++) {
-               q= t->tab[i];
-               if (!q) 
-                       continue; 
-               PSCHED_SET_PASTPERFECT(q->qidlestart);
-               q->qave = 0;
-               q->qcount = -1;
+               if (!q)
+                       continue;
+
+               red_restart(&q->parms);
                q->backlog = 0;
-               q->other=0;
-               q->forced=0;
-               q->pdrop=0;
-               q->early=0;
        }
 }
 
-static int gred_change(struct Qdisc *sch, struct rtattr *opt)
+static inline void gred_destroy_vq(struct gred_sched_data *q)
+{
+       kfree(q);
+}
+
+static inline int gred_change_table_def(struct Qdisc *sch, struct nlattr *dps)
 {
        struct gred_sched *table = qdisc_priv(sch);
-       struct gred_sched_data *q;
-       struct tc_gred_qopt *ctl;
        struct tc_gred_sopt *sopt;
-       struct rtattr *tb[TCA_GRED_STAB];
-       struct rtattr *tb2[TCA_GRED_DPS];
+       int i;
 
-       if (opt == NULL || rtattr_parse_nested(tb, TCA_GRED_STAB, opt))
+       if (dps == NULL)
                return -EINVAL;
 
-       if (tb[TCA_GRED_PARMS-1] == 0 && tb[TCA_GRED_STAB-1] == 0) {
-               rtattr_parse_nested(tb2, TCA_GRED_DPS, opt);
-
-           if (tb2[TCA_GRED_DPS-1] == 0) 
-                       return -EINVAL;
-
-               sopt = RTA_DATA(tb2[TCA_GRED_DPS-1]);
-               table->DPs=sopt->DPs;   
-               table->def=sopt->def_DP; 
-
-               if (sopt->grio) {
-                       table->grio = 1;
-                       gred_disable_wred_mode(table);
-                       if (gred_wred_mode_check(sch))
-                               gred_enable_wred_mode(table);
-               } else {
-                       table->grio = 0;
-                       gred_disable_wred_mode(table);
-               }
-
-               table->initd=0;
-               /* probably need to clear all the table DP entries as well */
-               return 0;
-           }
+       sopt = nla_data(dps);
 
+       if (sopt->DPs > MAX_DPs || sopt->DPs == 0 || sopt->def_DP >= sopt->DPs)
+               return -EINVAL;
 
-       if (!table->DPs || tb[TCA_GRED_PARMS-1] == 0 || tb[TCA_GRED_STAB-1] == 0 ||
-               RTA_PAYLOAD(tb[TCA_GRED_PARMS-1]) < sizeof(*ctl) ||
-               RTA_PAYLOAD(tb[TCA_GRED_STAB-1]) < 256)
-                       return -EINVAL;
-
-       ctl = RTA_DATA(tb[TCA_GRED_PARMS-1]);
-       if (ctl->DP > MAX_DPs-1 ) {
-               /* misbehaving is punished! Put in the default drop probability */
-               DPRINTK("\nGRED: DP %u not in  the proper range fixed. New DP "
-                       "set to default at %d\n",ctl->DP,table->def);
-               ctl->DP=table->def;
-       }
-       
-       if (table->tab[ctl->DP] == NULL) {
-               table->tab[ctl->DP]=kmalloc(sizeof(struct gred_sched_data),
-                                           GFP_KERNEL);
-               if (NULL == table->tab[ctl->DP])
-                       return -ENOMEM;
-               memset(table->tab[ctl->DP], 0, (sizeof(struct gred_sched_data)));
-       }
-       q= table->tab[ctl->DP]; 
-
-       if (table->grio) {
-               if (ctl->prio <=0) {
-                       if (table->def && table->tab[table->def]) {
-                               DPRINTK("\nGRED: DP %u does not have a prio"
-                                       "setting default to %d\n",ctl->DP,
-                                       table->tab[table->def]->prio);
-                               q->prio=table->tab[table->def]->prio;
-                       } else { 
-                               DPRINTK("\nGRED: DP %u does not have a prio"
-                                       " setting default to 8\n",ctl->DP);
-                               q->prio=8;
-                       }
-               } else {
-                       q->prio=ctl->prio;
-               }
-       } else {
-               q->prio=8;
-       }
+       sch_tree_lock(sch);
+       table->DPs = sopt->DPs;
+       table->def = sopt->def_DP;
+       table->red_flags = sopt->flags;
 
+       /*
+        * Every entry point to GRED is synchronized with the above code
+        * and the DP is checked against DPs, i.e. shadowed VQs can no
+        * longer be found so we can unlock right here.
+        */
+       sch_tree_unlock(sch);
 
-       q->DP=ctl->DP;
-       q->Wlog = ctl->Wlog;
-       q->Plog = ctl->Plog;
-       q->limit = ctl->limit;
-       q->Scell_log = ctl->Scell_log;
-       q->Rmask = ctl->Plog < 32 ? ((1<<ctl->Plog) - 1) : ~0UL;
-       q->Scell_max = (255<<q->Scell_log);
-       q->qth_min = ctl->qth_min<<ctl->Wlog;
-       q->qth_max = ctl->qth_max<<ctl->Wlog;
-       q->qave=0;
-       q->backlog=0;
-       q->qcount = -1;
-       q->other=0;
-       q->forced=0;
-       q->pdrop=0;
-       q->early=0;
-
-       PSCHED_SET_PASTPERFECT(q->qidlestart);
-       memcpy(q->Stab, RTA_DATA(tb[TCA_GRED_STAB-1]), 256);
-
-       if (table->grio) {
+       if (sopt->grio) {
+               gred_enable_rio_mode(table);
                gred_disable_wred_mode(table);
                if (gred_wred_mode_check(sch))
                        gred_enable_wred_mode(table);
+       } else {
+               gred_disable_rio_mode(table);
+               gred_disable_wred_mode(table);
        }
 
-       if (!table->initd) {
-               table->initd=1;
-               /* 
-               the first entry also goes into the default until
-               over-written 
-               */
-
-               if (table->tab[table->def] == NULL) {
-                       table->tab[table->def]=
-                               kmalloc(sizeof(struct gred_sched_data), GFP_KERNEL);
-                       if (NULL == table->tab[table->def])
-                               return -ENOMEM;
-
-                       memset(table->tab[table->def], 0,
-                              (sizeof(struct gred_sched_data)));
+       for (i = table->DPs; i < MAX_DPs; i++) {
+               if (table->tab[i]) {
+                       printk(KERN_WARNING "GRED: Warning: Destroying "
+                              "shadowed VQ 0x%x\n", i);
+                       gred_destroy_vq(table->tab[i]);
+                       table->tab[i] = NULL;
                }
-               q= table->tab[table->def]; 
-               q->DP=table->def;
-               q->Wlog = ctl->Wlog;
-               q->Plog = ctl->Plog;
-               q->limit = ctl->limit;
-               q->Scell_log = ctl->Scell_log;
-               q->Rmask = ctl->Plog < 32 ? ((1<<ctl->Plog) - 1) : ~0UL;
-               q->Scell_max = (255<<q->Scell_log);
-               q->qth_min = ctl->qth_min<<ctl->Wlog;
-               q->qth_max = ctl->qth_max<<ctl->Wlog;
-
-               if (table->grio)
-                       q->prio=table->tab[ctl->DP]->prio;
-               else
-                       q->prio=8;
-
-               q->qcount = -1;
-               PSCHED_SET_PASTPERFECT(q->qidlestart);
-               memcpy(q->Stab, RTA_DATA(tb[TCA_GRED_STAB-1]), 256);
        }
-       return 0;
 
+       return 0;
 }
 
-static int gred_init(struct Qdisc *sch, struct rtattr *opt)
+static inline int gred_change_vq(struct Qdisc *sch, int dp,
+                                struct tc_gred_qopt *ctl, int prio, u8 *stab)
 {
        struct gred_sched *table = qdisc_priv(sch);
-       struct tc_gred_sopt *sopt;
-       struct rtattr *tb[TCA_GRED_STAB];
-       struct rtattr *tb2[TCA_GRED_DPS];
+       struct gred_sched_data *q;
 
-       if (opt == NULL || rtattr_parse_nested(tb, TCA_GRED_STAB, opt))
-               return -EINVAL;
+       if (table->tab[dp] == NULL) {
+               table->tab[dp] = kzalloc(sizeof(*q), GFP_KERNEL);
+               if (table->tab[dp] == NULL)
+                       return -ENOMEM;
+       }
 
-       if (tb[TCA_GRED_PARMS-1] == 0 && tb[TCA_GRED_STAB-1] == 0) {
-               rtattr_parse_nested(tb2, TCA_GRED_DPS, opt);
+       q = table->tab[dp];
+       q->DP = dp;
+       q->prio = prio;
+       q->limit = ctl->limit;
 
-           if (tb2[TCA_GRED_DPS-1] == 0) 
-                       return -EINVAL;
+       if (q->backlog == 0)
+               red_end_of_idle_period(&q->parms);
 
-               sopt = RTA_DATA(tb2[TCA_GRED_DPS-1]);
-               table->DPs=sopt->DPs;   
-               table->def=sopt->def_DP; 
-               table->grio=sopt->grio; 
-               table->initd=0;
-               return 0;
-       }
+       red_set_parms(&q->parms,
+                     ctl->qth_min, ctl->qth_max, ctl->Wlog, ctl->Plog,
+                     ctl->Scell_log, stab);
 
-       DPRINTK("\n GRED_INIT error!\n");
-       return -EINVAL;
+       return 0;
 }
 
-static int gred_dump(struct Qdisc *sch, struct sk_buff *skb)
+static const struct nla_policy gred_policy[TCA_GRED_MAX + 1] = {
+       [TCA_GRED_PARMS]        = { .len = sizeof(struct tc_gred_qopt) },
+       [TCA_GRED_STAB]         = { .len = 256 },
+       [TCA_GRED_DPS]          = { .len = sizeof(struct tc_gred_sopt) },
+};
+
+static int gred_change(struct Qdisc *sch, struct nlattr *opt)
 {
-       unsigned long qave;
-       struct rtattr *rta;
-       struct tc_gred_qopt *opt = NULL ;
-       struct tc_gred_qopt *dst;
        struct gred_sched *table = qdisc_priv(sch);
-       struct gred_sched_data *q;
-       int i;
-       unsigned char    *b = skb->tail;
+       struct tc_gred_qopt *ctl;
+       struct nlattr *tb[TCA_GRED_MAX + 1];
+       int err, prio = GRED_DEF_PRIO;
+       u8 *stab;
+
+       if (opt == NULL)
+               return -EINVAL;
+
+       err = nla_parse_nested(tb, TCA_GRED_MAX, opt, gred_policy);
+       if (err < 0)
+               return err;
+
+       if (tb[TCA_GRED_PARMS] == NULL && tb[TCA_GRED_STAB] == NULL)
+               return gred_change_table_def(sch, opt);
+
+       if (tb[TCA_GRED_PARMS] == NULL ||
+           tb[TCA_GRED_STAB] == NULL)
+               return -EINVAL;
+
+       err = -EINVAL;
+       ctl = nla_data(tb[TCA_GRED_PARMS]);
+       stab = nla_data(tb[TCA_GRED_STAB]);
 
-       rta = (struct rtattr*)b;
-       RTA_PUT(skb, TCA_OPTIONS, 0, NULL);
+       if (ctl->DP >= table->DPs)
+               goto errout;
 
-       opt=kmalloc(sizeof(struct tc_gred_qopt)*MAX_DPs, GFP_KERNEL);
+       if (gred_rio_mode(table)) {
+               if (ctl->prio == 0) {
+                       int def_prio = GRED_DEF_PRIO;
 
-       if (opt  == NULL) {
-               DPRINTK("gred_dump:failed to malloc for %Zd\n",
-                   sizeof(struct tc_gred_qopt)*MAX_DPs);
-               goto rtattr_failure;
+                       if (table->tab[table->def])
+                               def_prio = table->tab[table->def]->prio;
+
+                       printk(KERN_DEBUG "GRED: DP %u does not have a prio "
+                              "setting default to %d\n", ctl->DP, def_prio);
+
+                       prio = def_prio;
+               } else
+                       prio = ctl->prio;
        }
 
-       memset(opt, 0, (sizeof(struct tc_gred_qopt))*table->DPs);
+       sch_tree_lock(sch);
+
+       err = gred_change_vq(sch, ctl->DP, ctl, prio, stab);
+       if (err < 0)
+               goto errout_locked;
 
-       if (!table->initd) {
-               DPRINTK("NO GRED Queues setup!\n");
+       if (gred_rio_mode(table)) {
+               gred_disable_wred_mode(table);
+               if (gred_wred_mode_check(sch))
+                       gred_enable_wred_mode(table);
        }
 
-       for (i=0;i<MAX_DPs;i++) {
-               dst= &opt[i]; 
-               q= table->tab[i]; 
+       err = 0;
+
+errout_locked:
+       sch_tree_unlock(sch);
+errout:
+       return err;
+}
+
+static int gred_init(struct Qdisc *sch, struct nlattr *opt)
+{
+       struct nlattr *tb[TCA_GRED_MAX + 1];
+       int err;
+
+       if (opt == NULL)
+               return -EINVAL;
+
+       err = nla_parse_nested(tb, TCA_GRED_MAX, opt, gred_policy);
+       if (err < 0)
+               return err;
+
+       if (tb[TCA_GRED_PARMS] || tb[TCA_GRED_STAB])
+               return -EINVAL;
+
+       return gred_change_table_def(sch, tb[TCA_GRED_DPS]);
+}
+
+static int gred_dump(struct Qdisc *sch, struct sk_buff *skb)
+{
+       struct gred_sched *table = qdisc_priv(sch);
+       struct nlattr *parms, *opts = NULL;
+       int i;
+       struct tc_gred_sopt sopt = {
+               .DPs    = table->DPs,
+               .def_DP = table->def,
+               .grio   = gred_rio_mode(table),
+               .flags  = table->red_flags,
+       };
+
+       opts = nla_nest_start(skb, TCA_OPTIONS);
+       if (opts == NULL)
+               goto nla_put_failure;
+       NLA_PUT(skb, TCA_GRED_DPS, sizeof(sopt), &sopt);
+       parms = nla_nest_start(skb, TCA_GRED_PARMS);
+       if (parms == NULL)
+               goto nla_put_failure;
+
+       for (i = 0; i < MAX_DPs; i++) {
+               struct gred_sched_data *q = table->tab[i];
+               struct tc_gred_qopt opt;
+
+               memset(&opt, 0, sizeof(opt));
 
                if (!q) {
                        /* hack -- fix at some point with proper message
                           This is how we indicate to tc that there is no VQ
                           at this DP */
 
-                       dst->DP=MAX_DPs+i;
-                       continue;
+                       opt.DP = MAX_DPs + i;
+                       goto append_opt;
                }
 
-               dst->limit=q->limit;
-               dst->qth_min=q->qth_min>>q->Wlog;
-               dst->qth_max=q->qth_max>>q->Wlog;
-               dst->DP=q->DP;
-               dst->backlog=q->backlog;
-               if (q->qave) {
-                       if (gred_wred_mode(table)) {
-                               q->qidlestart=table->tab[table->def]->qidlestart;
-                               q->qave=table->tab[table->def]->qave;
-                       }
-                       if (!PSCHED_IS_PASTPERFECT(q->qidlestart)) {
-                               long idle;
-                               psched_time_t now;
-                               PSCHED_GET_TIME(now);
-                               idle = PSCHED_TDIFF_SAFE(now, q->qidlestart, q->Scell_max);
-                               qave  = q->qave >> q->Stab[(idle>>q->Scell_log)&0xFF];
-                               dst->qave = qave >> q->Wlog;
-
-                       } else {
-                               dst->qave = q->qave >> q->Wlog;
-                       }
-               } else {
-                       dst->qave = 0;
+               opt.limit       = q->limit;
+               opt.DP          = q->DP;
+               opt.backlog     = q->backlog;
+               opt.prio        = q->prio;
+               opt.qth_min     = q->parms.qth_min >> q->parms.Wlog;
+               opt.qth_max     = q->parms.qth_max >> q->parms.Wlog;
+               opt.Wlog        = q->parms.Wlog;
+               opt.Plog        = q->parms.Plog;
+               opt.Scell_log   = q->parms.Scell_log;
+               opt.other       = q->stats.other;
+               opt.early       = q->stats.prob_drop;
+               opt.forced      = q->stats.forced_drop;
+               opt.pdrop       = q->stats.pdrop;
+               opt.packets     = q->packetsin;
+               opt.bytesin     = q->bytesin;
+
+               if (gred_wred_mode(table)) {
+                       q->parms.qidlestart =
+                               table->tab[table->def]->parms.qidlestart;
+                       q->parms.qavg = table->tab[table->def]->parms.qavg;
                }
-               
-
-               dst->Wlog = q->Wlog;
-               dst->Plog = q->Plog;
-               dst->Scell_log = q->Scell_log;
-               dst->other = q->other;
-               dst->forced = q->forced;
-               dst->early = q->early;
-               dst->pdrop = q->pdrop;
-               dst->prio = q->prio;
-               dst->packets=q->packetsin;
-               dst->bytesin=q->bytesin;
-       }
 
-       RTA_PUT(skb, TCA_GRED_PARMS, sizeof(struct tc_gred_qopt)*MAX_DPs, opt);
-       rta->rta_len = skb->tail - b;
+               opt.qave = red_calc_qavg(&q->parms, q->parms.qavg);
+
+append_opt:
+               if (nla_append(skb, sizeof(opt), &opt) < 0)
+                       goto nla_put_failure;
+       }
 
-       kfree(opt);
-       return skb->len;
+       nla_nest_end(skb, parms);
 
-rtattr_failure:
-       if (opt)
-               kfree(opt);
-       DPRINTK("gred_dump: FAILURE!!!!\n");
+       return nla_nest_end(skb, opts);
 
-/* also free the opt struct here */
-       skb_trim(skb, b - skb->data);
-       return -1;
+nla_put_failure:
+       nla_nest_cancel(skb, opts);
+       return -EMSGSIZE;
 }
 
 static void gred_destroy(struct Qdisc *sch)
@@ -637,20 +571,18 @@ static void gred_destroy(struct Qdisc *sch)
        struct gred_sched *table = qdisc_priv(sch);
        int i;
 
-       for (i = 0;i < table->DPs; i++) {
+       for (i = 0; i < table->DPs; i++) {
                if (table->tab[i])
-                       kfree(table->tab[i]);
+                       gred_destroy_vq(table->tab[i]);
        }
 }
 
-static struct Qdisc_ops gred_qdisc_ops = {
-       .next           =       NULL,
-       .cl_ops         =       NULL,
+static struct Qdisc_ops gred_qdisc_ops __read_mostly = {
        .id             =       "gred",
        .priv_size      =       sizeof(struct gred_sched),
        .enqueue        =       gred_enqueue,
        .dequeue        =       gred_dequeue,
-       .requeue        =       gred_requeue,
+       .peek           =       qdisc_peek_head,
        .drop           =       gred_drop,
        .init           =       gred_init,
        .reset          =       gred_reset,
@@ -664,10 +596,13 @@ static int __init gred_module_init(void)
 {
        return register_qdisc(&gred_qdisc_ops);
 }
-static void __exit gred_module_exit(void) 
+
+static void __exit gred_module_exit(void)
 {
        unregister_qdisc(&gred_qdisc_ops);
 }
+
 module_init(gred_module_init)
 module_exit(gred_module_exit)
+
 MODULE_LICENSE("GPL");