kernel - Enhance the FAIRQ ALTQ to use relative weightings

* Previously the FAIRQ altq specification guaranteed a bandwidth minimum
  for each queue but then used a fixed priority model for any excess
  bandwidth.

* Now the FAIRQ uses the minimum bandwidth specifications to also calculate
  relative weightings for any excess bandwidth, and no longer uses a fixed
  priority model.

* This results in a better apportionment of excess bandwidth and allows
  bandwidth specifications to also be used as a pure weighting specification
  if desired.

* Fix bandwidth calculation.  The decay was being executed on every packet
  instead of once a second.

diff --git a/sys/net/altq/altq_fairq.c b/sys/net/altq/altq_fairq.c
index 79283ad..bd197d5 100644 (file)
--- a/sys/net/altq/altq_fairq.c
+++ b/sys/net/altq/altq_fairq.c
@@ -76,11 +76,15 @@
  * EXAMPLE:
  *
  *  altq on em0 fairq bandwidth 650Kb queue { std, bulk }
- *  queue std  priority 3 bandwidth 400Kb \
+ *  queue std  priority 3 bandwidth 200Kb \
  *     fairq (buckets 64, default, hogs 1Kb) qlimit 50
  *  queue bulk priority 2 bandwidth 100Kb \
  *     fairq (buckets 64, hogs 1Kb) qlimit 50
  *
+ *     NOTE: When the aggregate bandwidth is less than the link bandwidth
+ *           any remaining bandwidth is dynamically assigned using the
+ *           existing bandwidth specs as weightings.
+ *
  *  pass out on em0 from any to any keep state queue std
  *  pass out on em0 inet proto tcp ..... port ... keep state queue bulk
  */
@@ -118,9 +122,11 @@
 static int     fairq_clear_interface(struct fairq_if *);
 static int     fairq_request(struct ifaltq *, int, void *);
 static void    fairq_purge(struct fairq_if *);
-static struct fairq_class *fairq_class_create(struct fairq_if *, int, int, u_int, struct fairq_opts *, int);
+static struct fairq_class *fairq_class_create(struct fairq_if *, int,
+                                       int, u_int, struct fairq_opts *, int);
 static int     fairq_class_destroy(struct fairq_class *);
-static int     fairq_enqueue(struct ifaltq *, struct mbuf *, struct altq_pktattr *);
+static int     fairq_enqueue(struct ifaltq *, struct mbuf *,
+                                       struct altq_pktattr *);
 static struct mbuf *fairq_dequeue(struct ifaltq *, struct mbuf *, int);
 
 static int     fairq_addq(struct fairq_class *, struct mbuf *, int hash);
@@ -129,14 +135,16 @@ static struct mbuf *fairq_pollq(struct fairq_class *, uint64_t, int *);
 static fairq_bucket_t *fairq_selectq(struct fairq_class *, int);
 static void    fairq_purgeq(struct fairq_class *);
 
-static void    get_class_stats(struct fairq_classstats *, struct fairq_class *);
+static void    get_class_stats(struct fairq_classstats *,
+                                       struct fairq_class *);
 static struct fairq_class *clh_to_clp(struct fairq_if *, uint32_t);
 
 int
 fairq_pfattach(struct pf_altq *a, struct ifaltq *ifq)
 {
        return altq_attach(ifq, ALTQT_FAIRQ, a->altq_disc,
-           fairq_enqueue, fairq_dequeue, fairq_request, NULL, NULL);
+                          fairq_enqueue, fairq_dequeue,
+                          fairq_request, NULL, NULL);
 }
 
 int
@@ -395,7 +403,7 @@ fairq_class_create(struct fairq_if *pif, int pri, int qlimit,
        for (i = 0; i < cl->cl_nbuckets; ++i) {
                qlimit(&cl->cl_buckets[i].queue) = qlimit;
        }
-       cl->cl_bandwidth = bandwidth / 8;
+       cl->cl_bandwidth = bandwidth / 8;       /* cvt to bytes per second */
        cl->cl_qtype = Q_DROPTAIL;
        cl->cl_flags = flags & FARF_USERFLAGS;
        cl->cl_pri = pri;
@@ -405,6 +413,7 @@ fairq_class_create(struct fairq_if *pif, int pri, int qlimit,
        cl->cl_handle = qid;
        cl->cl_hogs_m1 = opts->hogs_m1 / 8;
        cl->cl_lssc_m1 = opts->lssc_m1 / 8;     /* NOT YET USED */
+       cl->cl_bw_current = 0;
 
 #ifdef ALTQ_RED
        if (flags & (FARF_RED|FARF_RIO)) {
@@ -565,6 +574,8 @@ fairq_dequeue(struct ifaltq *ifq, struct mbuf *mpolled, int op)
        struct mbuf *best_m;
        struct mbuf *m;
        uint64_t cur_time = read_machclk();
+       u_int best_scale;
+       u_int scale;
        int pri;
        int hit_limit;
 
@@ -586,6 +597,7 @@ fairq_dequeue(struct ifaltq *ifq, struct mbuf *mpolled, int op)
        } else {
                best_cl = NULL;
                best_m = NULL;
+               best_scale = 0xFFFFFFFFU;
 
                for (pri = pif->pif_maxpri;  pri >= 0; pri--) {
                        if ((cl = pif->pif_classes[pri]) == NULL)
@@ -599,22 +611,29 @@ fairq_dequeue(struct ifaltq *ifq, struct mbuf *mpolled, int op)
                        }
 
                        /*
-                        * Only override the best choice if we are under
-                        * the BW limit.
+                        * We can halt the search immediately if the queue
+                        * did not hit its bandwidth limit.
                         */
-                       if (hit_limit == 0 || best_cl == NULL) {
+                       if (hit_limit == 0) {
                                best_cl = cl;
                                best_m = m;
+                               break;
                        }
 
                        /*
-                        * Remember the highest priority mbuf in case we
-                        * do not find any lower priority mbufs.
+                        * Otherwise calculate the scale factor and select
+                        * the queue with the lowest scale factor.  This
+                        * apportions any unused bandwidth weighted by
+                        * the relative bandwidth specification.
                         */
-                       if (hit_limit)
-                               continue;
-                       break;
+                       scale = cl->cl_bw_current * 100 / cl->cl_bandwidth;
+                       if (scale < best_scale) {
+                               best_cl = cl;
+                               best_m = m;
+                               best_scale = scale;
+                       }
                }
+
                if (op == ALTDQ_POLL) {
                        pif->pif_poll_cache = best_cl;
                        m = best_m;
@@ -736,8 +755,10 @@ fairq_getq(struct fairq_class *cl, uint64_t cur_time)
                cl->cl_bw_delta += delta;
                cl->cl_bw_bytes += m->m_pkthdr.len;
                cl->cl_last_time = cur_time;
-               cl->cl_bw_delta -= cl->cl_bw_delta >> 3;
-               cl->cl_bw_bytes -= cl->cl_bw_bytes >> 3;
+               if (cl->cl_bw_delta > machclk_freq) {
+                       cl->cl_bw_delta -= cl->cl_bw_delta >> 2;
+                       cl->cl_bw_bytes -= cl->cl_bw_bytes >> 2;
+               }
 
                /*
                 * Per-bucket bandwidth calculation
@@ -748,8 +769,10 @@ fairq_getq(struct fairq_class *cl, uint64_t cur_time)
                b->bw_delta += delta;
                b->bw_bytes += m->m_pkthdr.len;
                b->last_time = cur_time;
-               b->bw_delta -= b->bw_delta >> 3;
-               b->bw_bytes -= b->bw_bytes >> 3;
+               if (b->bw_delta > machclk_freq) {
+                       b->bw_delta -= b->bw_delta >> 2;
+                       b->bw_bytes -= b->bw_bytes >> 2;
+               }
        }
        return(m);
 }
@@ -789,6 +812,7 @@ fairq_pollq(struct fairq_class *cl, uint64_t cur_time, int *hit_limit)
 
                if (bw > cl->cl_bandwidth)
                        *hit_limit = 1;
+               cl->cl_bw_current = bw;
 #if 0
                kprintf("BW %6lld relative to %6u %d queue %p\n",
                        bw, cl->cl_bandwidth, *hit_limit, b);

diff --git a/sys/net/altq/altq_fairq.h b/sys/net/altq/altq_fairq.h
index f163b03..94cb9c1 100644 (file)
--- a/sys/net/altq/altq_fairq.h
+++ b/sys/net/altq/altq_fairq.h
@@ -101,6 +101,7 @@ struct fairq_class {
        u_int           cl_hogs_m1;
        u_int           cl_lssc_m1;
        u_int           cl_bandwidth;
+       u_int           cl_bw_current;
        uint64_t        cl_bw_bytes;
        uint64_t        cl_bw_delta;
        uint64_t        cl_last_time;

diff --git a/usr.sbin/pfctl/pf.conf.5 b/usr.sbin/pfctl/pf.conf.5
index c3fe076..0fdac9c 100644 (file)
--- a/usr.sbin/pfctl/pf.conf.5
+++ b/usr.sbin/pfctl/pf.conf.5
@@ -898,13 +898,18 @@ or
 .Pp
 Packet selection operates as follows:
 The queues are scanned from highest priority to lowest priority.
-If a queue has pending packets and has not reached its bandwidth limit the
+If a queue has pending packets and is under its bandwidth minimum the
 scan stops and a packet is selected from that queue.
-If a queue has reached its bandwidth limit the scan continues searching for
-other, lower priority queues which have not.
-If no queue is found to be
-suitable then the highest priority queue with pending packets is used
-regardless of whether it has reached its bandwidth limit or not.
+If all queues have reached their bandwidth minimum a scale factor based
+on each queue's bandwidth minimum verses that queue's current bandwidth
+usage is calculated and the queue with the lowest scale factor is selected.
+This effectively uses the minimum bandwidth specification as a relative
+weighting for apportioning any remaining bandwidth on the link.
+.Pp
+The priority mechanic is only applicable in cases where the aggregate
+minimum bandwidth guarantees exceed the link bandwidth, and also has
+a small effect on queue selection when prioritizing between equal scale
+calculations.
 .Pp
 A
 .Ar fairq
@@ -926,6 +931,12 @@ rules not characterized by
 .Ar keep state
 is supported.
 Such a queue serves as a basic priority queue with a bandwidth specification.
+.Pp
+Also note that when specifying rules it is always a good idea to specify
+a secondary queue for any tcp rules.
+The secondary queue is selected for pure ACKs without payloads and should
+generally be dedicated to that purpose with a minimum bandwidth specification
+sufficient to max-out the bandwidth for your incoming traffic.
 .El
 .Pp
 The interfaces on which queueing should be activated are declared using