[SCHED]: Qdisc changes and sch_rr added for multiqueue
Peter P Waskiewicz Jr [Fri, 29 Jun 2007 04:04:31 +0000 (21:04 -0700)]
Add the new sch_rr qdisc for multiqueue network device support.  Allow
sch_prio and sch_rr to be compiled with or without multiqueue hardware
support.

sch_rr is part of sch_prio, and is referenced from MODULE_ALIAS.  This
was done since sch_prio and sch_rr only differ in their dequeue
routine.

Signed-off-by: Peter P Waskiewicz Jr <peter.p.waskiewicz.jr@intel.com>
Signed-off-by: Patrick McHardy <kaber@trash.net>
Signed-off-by: David S. Miller <davem@davemloft.net>

include/linux/pkt_sched.h
net/sched/Kconfig
net/sched/sch_prio.c

index d10f353..268c515 100644 (file)
@@ -101,6 +101,15 @@ struct tc_prio_qopt
        __u8    priomap[TC_PRIO_MAX+1]; /* Map: logical priority -> PRIO band */
 };
 
+enum
+{
+       TCA_PRIO_UNSPEC,
+       TCA_PRIO_MQ,
+       __TCA_PRIO_MAX
+};
+
+#define TCA_PRIO_MAX    (__TCA_PRIO_MAX - 1)
+
 /* TBF section */
 
 struct tc_tbf_qopt
index 475df84..f321794 100644 (file)
@@ -111,6 +111,17 @@ config NET_SCH_PRIO
          To compile this code as a module, choose M here: the
          module will be called sch_prio.
 
+config NET_SCH_RR
+       tristate "Multi Band Round Robin Queuing (RR)"
+       select NET_SCH_PRIO
+       ---help---
+         Say Y here if you want to use an n-band round robin packet
+         scheduler.
+
+         The module uses sch_prio for its framework and is aliased as
+         sch_rr, so it will load sch_prio, although it is referred
+         to using sch_rr.
+
 config NET_SCH_RED
        tristate "Random Early Detection (RED)"
        ---help---
index 6d7542c..4045220 100644 (file)
 struct prio_sched_data
 {
        int bands;
+       int curband; /* for round-robin */
        struct tcf_proto *filter_list;
        u8  prio2band[TC_PRIO_MAX+1];
        struct Qdisc *queues[TCQ_PRIO_BANDS];
+       int mq;
 };
 
 
@@ -70,14 +72,17 @@ prio_classify(struct sk_buff *skb, struct Qdisc *sch, int *qerr)
 #endif
                        if (TC_H_MAJ(band))
                                band = 0;
-                       return q->queues[q->prio2band[band&TC_PRIO_MAX]];
+                       band = q->prio2band[band&TC_PRIO_MAX];
+                       goto out;
                }
                band = res.classid;
        }
        band = TC_H_MIN(band) - 1;
        if (band >= q->bands)
-               return q->queues[q->prio2band[0]];
-
+               band = q->prio2band[0];
+out:
+       if (q->mq)
+               skb_set_queue_mapping(skb, band);
        return q->queues[band];
 }
 
@@ -144,17 +149,58 @@ prio_dequeue(struct Qdisc* sch)
        struct Qdisc *qdisc;
 
        for (prio = 0; prio < q->bands; prio++) {
-               qdisc = q->queues[prio];
-               skb = qdisc->dequeue(qdisc);
-               if (skb) {
-                       sch->q.qlen--;
-                       return skb;
+               /* Check if the target subqueue is available before
+                * pulling an skb.  This way we avoid excessive requeues
+                * for slower queues.
+                */
+               if (!netif_subqueue_stopped(sch->dev, (q->mq ? prio : 0))) {
+                       qdisc = q->queues[prio];
+                       skb = qdisc->dequeue(qdisc);
+                       if (skb) {
+                               sch->q.qlen--;
+                               return skb;
+                       }
                }
        }
        return NULL;
 
 }
 
+static struct sk_buff *rr_dequeue(struct Qdisc* sch)
+{
+       struct sk_buff *skb;
+       struct prio_sched_data *q = qdisc_priv(sch);
+       struct Qdisc *qdisc;
+       int bandcount;
+
+       /* Only take one pass through the queues.  If nothing is available,
+        * return nothing.
+        */
+       for (bandcount = 0; bandcount < q->bands; bandcount++) {
+               /* Check if the target subqueue is available before
+                * pulling an skb.  This way we avoid excessive requeues
+                * for slower queues.  If the queue is stopped, try the
+                * next queue.
+                */
+               if (!netif_subqueue_stopped(sch->dev,
+                                           (q->mq ? q->curband : 0))) {
+                       qdisc = q->queues[q->curband];
+                       skb = qdisc->dequeue(qdisc);
+                       if (skb) {
+                               sch->q.qlen--;
+                               q->curband++;
+                               if (q->curband >= q->bands)
+                                       q->curband = 0;
+                               return skb;
+                       }
+               }
+               q->curband++;
+               if (q->curband >= q->bands)
+                       q->curband = 0;
+       }
+       return NULL;
+}
+
 static unsigned int prio_drop(struct Qdisc* sch)
 {
        struct prio_sched_data *q = qdisc_priv(sch);
@@ -198,21 +244,41 @@ prio_destroy(struct Qdisc* sch)
 static int prio_tune(struct Qdisc *sch, struct rtattr *opt)
 {
        struct prio_sched_data *q = qdisc_priv(sch);
-       struct tc_prio_qopt *qopt = RTA_DATA(opt);
+       struct tc_prio_qopt *qopt;
+       struct rtattr *tb[TCA_PRIO_MAX];
        int i;
 
-       if (opt->rta_len < RTA_LENGTH(sizeof(*qopt)))
+       if (rtattr_parse_nested_compat(tb, TCA_PRIO_MAX, opt, qopt,
+                                      sizeof(*qopt)))
                return -EINVAL;
-       if (qopt->bands > TCQ_PRIO_BANDS || qopt->bands < 2)
+       q->bands = qopt->bands;
+       /* If we're multiqueue, make sure the number of incoming bands
+        * matches the number of queues on the device we're associating with.
+        * If the number of bands requested is zero, then set q->bands to
+        * dev->egress_subqueue_count.
+        */
+       q->mq = RTA_GET_FLAG(tb[TCA_PRIO_MQ - 1]);
+       if (q->mq) {
+               if (sch->handle != TC_H_ROOT)
+                       return -EINVAL;
+               if (netif_is_multiqueue(sch->dev)) {
+                       if (q->bands == 0)
+                               q->bands = sch->dev->egress_subqueue_count;
+                       else if (q->bands != sch->dev->egress_subqueue_count)
+                               return -EINVAL;
+               } else
+                       return -EOPNOTSUPP;
+       }
+
+       if (q->bands > TCQ_PRIO_BANDS || q->bands < 2)
                return -EINVAL;
 
        for (i=0; i<=TC_PRIO_MAX; i++) {
-               if (qopt->priomap[i] >= qopt->bands)
+               if (qopt->priomap[i] >= q->bands)
                        return -EINVAL;
        }
 
        sch_tree_lock(sch);
-       q->bands = qopt->bands;
        memcpy(q->prio2band, qopt->priomap, TC_PRIO_MAX+1);
 
        for (i=q->bands; i<TCQ_PRIO_BANDS; i++) {
@@ -268,11 +334,17 @@ static int prio_dump(struct Qdisc *sch, struct sk_buff *skb)
 {
        struct prio_sched_data *q = qdisc_priv(sch);
        unsigned char *b = skb_tail_pointer(skb);
+       struct rtattr *nest;
        struct tc_prio_qopt opt;
 
        opt.bands = q->bands;
        memcpy(&opt.priomap, q->prio2band, TC_PRIO_MAX+1);
-       RTA_PUT(skb, TCA_OPTIONS, sizeof(opt), &opt);
+
+       nest = RTA_NEST_COMPAT(skb, TCA_OPTIONS, sizeof(opt), &opt);
+       if (q->mq)
+               RTA_PUT_FLAG(skb, TCA_PRIO_MQ);
+       RTA_NEST_COMPAT_END(skb, nest);
+
        return skb->len;
 
 rtattr_failure:
@@ -443,17 +515,44 @@ static struct Qdisc_ops prio_qdisc_ops = {
        .owner          =       THIS_MODULE,
 };
 
+static struct Qdisc_ops rr_qdisc_ops = {
+       .next           =       NULL,
+       .cl_ops         =       &prio_class_ops,
+       .id             =       "rr",
+       .priv_size      =       sizeof(struct prio_sched_data),
+       .enqueue        =       prio_enqueue,
+       .dequeue        =       rr_dequeue,
+       .requeue        =       prio_requeue,
+       .drop           =       prio_drop,
+       .init           =       prio_init,
+       .reset          =       prio_reset,
+       .destroy        =       prio_destroy,
+       .change         =       prio_tune,
+       .dump           =       prio_dump,
+       .owner          =       THIS_MODULE,
+};
+
 static int __init prio_module_init(void)
 {
-       return register_qdisc(&prio_qdisc_ops);
+       int err;
+
+       err = register_qdisc(&prio_qdisc_ops);
+       if (err < 0)
+               return err;
+       err = register_qdisc(&rr_qdisc_ops);
+       if (err < 0)
+               unregister_qdisc(&prio_qdisc_ops);
+       return err;
 }
 
 static void __exit prio_module_exit(void)
 {
        unregister_qdisc(&prio_qdisc_ops);
+       unregister_qdisc(&rr_qdisc_ops);
 }
 
 module_init(prio_module_init)
 module_exit(prio_module_exit)
 
 MODULE_LICENSE("GPL");
+MODULE_ALIAS("sch_rr");