Patchwork pkt_sched: QFQ Plus: fair-queueing service at DRR cost

login
register
mail settings
Submitter Paolo Valente
Date Nov. 20, 2012, 5:45 p.m.
Message ID <20121120174513.GA21618@paolo-ThinkPad-W520>
Download mbox | patch
Permalink /patch/200491/
State Changes Requested
Delegated to: David Miller
Headers show

Comments

Paolo Valente - Nov. 20, 2012, 5:45 p.m.
This patch turns QFQ into QFQ+, a variant of QFQ that provides the
following two benefits: 1) QFQ+ is faster than QFQ, 2) differently
from QFQ, QFQ+ correctly schedules also non-leaves classes in a
hierarchical setting.  The way how QFQ+ achieves these goals is
discussed briefly below. I also report some performance
measurements. A detailed description of QFQ+ and of these results can
be found in [1].

QFQ+ achieves a higher speed than QFQ by grouping classes into
aggregates, and uses the original QFQ scheduling algorithm to schedule
aggregates instead of single classes. An aggregate is made of at most
M classes, all with the same weight and maximum packet size.  M is
equal to the minimum between tx_queue_len+1 and 8 (value chosen to get
a good trade-off between execution time and service guarantees). QFQ+
associates each aggregate with a budget equal to the maximum packet
size for the classes in the aggregate, multiplied by the number of
classes of the aggregate. Once selected an aggregate for service, QFQ+
dequeues only the packets of its classes, until the aggregate finishes
its budget. Finally, within an aggregate, classes are scheduled with
DRR. In my measurements, described below, the execution time of QFQ+
without TSO/GSO and with M=8 was from 16% to 31% lower than that of
QFQ, and close to that of DRR.

QFQ+ does not use packet lengths for computing aggregate timestamps,
but budgets. Hence it does not need to modify any timestamp if the
head packet of a class changes. As a consequence, differently from
QFQ, which uses head-packet lengths to compute class timestamps, QFQ+
does not need further modifications to correctly schedule also
non-leaf classes and classes with non-FIFO qdiscs.

As for service guarantees, thanks to the way how M is computed, the
service of QFQ+ is close to the one of QFQ. For example, as proved in
[1], under QFQ+ every packet of a given class is guaranteed the same
worst-case completion time as under QFQ, plus an additional delay
equal to the transmission time, at the rate reserved to the class, of
three maximum-size packet. See [1, Section 7.1] for a numerical
comparison among the packet delays guaranteed by QFQ+, QFQ and DRR.

I measured the execution time of QFQ+, DRR and QFQ using the testing
environment [2]. In particular, for each scheduler I measured the
average total execution time of a packet enqueue plus a packet
dequeue, without TSO/GSO. What would happen with TSO/GSO is discussed
at the end of this description. For practical reasons, in the testing
environment each enqueue&dequeue is also charged for the cost of
generating and discarding an empty, fixed-size packet (using a free
list). The following table reports the results with an i7-2760QM,
against four different class sets. Time is measured in nanoseconds,
while each set or subset of classes is denoted as
<num_classes>-w<weight>, where <num_classes> and <weight> are,
respectively, the number of classes and the weight of every class in
the set/subset (for example, 250-w1 stands for 250 classes with weight
1). For QFQ+, the table shows the results for the two extremes for M:
1 and 8 (see [1, Section 7.2] for results with other values of M and
for more information).

 -----------------------------------------------
| Set of  |      QFQ+ (M)     |   DRR      QFQ  |
| classes |    1          8   |                 |
|-----------------------------------------------|     
| 1k-w1   |   89         63   |    56       81  |
|-----------------------------------------------|
| 500-w1, |                   |                 |
| 250-w2, |  102         71   |    87      103  |
| 250-w4  |                   |                 |
|-----------------------------------------------|
| 32k-w1  |  267        225   |   173      257  |
|-----------------------------------------------|
| 16k-w1, |                   |                 |
| 8k-w2,  |  253        187   |   252      257  |
| 8k-w4   |                   |                 |
 -----------------------------------------------

About DRR, it achieves its best performance when all the classes have
the same weight. This is fortunate, because in such scenarios it is
actually pointless to use a fair-queueing scheduler, as the latter
would provide the same quality of service as DRR. In contrast, when
classes have differentiated weights and the better service properties
of QFQ+ make a difference, QFQ+ has better performance than DRR. It
happens mainly because QFQ+ dequeues packets in an order that causes
about 8% less cache misses than DRR. As for the number of
instructions, QFQ+ executes instead about 7% more instructions than
DRR, whereas QFQ executes from 25% to 34% more instructions than DRR.

With TSO/GSO, the number of instructions per packet enqueue/dequeue
executed by QFQ and QFQ+ is the same as without TSO/GSO. In contrast,
for each dequeue, the number of iterations executed by the main loop
of DRR is 64K/quantum times as high as without TSO/GSO.

Paolo

[1] P. Valente, "Reducing the Execution Time of Fair-Queueing Schedulers"
http://algo.ing.unimo.it/people/paolo/agg-sched/agg-sched.pdf

[2] http://algo.ing.unimo.it/people/paolo/agg-sched/test-env.tgz

Signed-off-by: Paolo Valente <paolo.valente@unimore.it>
---
 net/sched/sch_qfq.c |  834 +++++++++++++++++++++++++++++++++++----------------
 1 file changed, 571 insertions(+), 263 deletions(-)
David Miller - Nov. 20, 2012, 6:54 p.m.
From: Paolo Valente <paolo.valente@unimore.it>
Date: Tue, 20 Nov 2012 18:45:13 +0100

> -	struct sk_buff *skb;
> +	struct sk_buff *skb = NULL;

This is not really an improvement, now the compiler can think
that NULL is passed eventually into qdisc_bstats_update().

Please make the logic easier for the compiler to digest.

For example, restructure the top-level logic into something like:

	skb = NULL;
	if (!list_empty(&in_serv_agg->active))
		skb = qfq_peek_skb(in_serv_agg, &cl, &len);
	else
		len = 0; /* no more active classes in the in-service agg */

	if (len == 0 || in_serv_agg->budget < len) {
 ...
		/*
		 * If we get here, there are other aggregates queued:
		 * choose the new aggregate to serve.
		 */
		in_serv_agg = q->in_serv_agg = qfq_choose_next_agg(q);
		skb = qfq_peek_skb(in_serv_agg, &cl, &len);
	}
	if (!skb)
		return NULL;

That way it is clearer, to both humans and the compiler, what is
going on here.

Thanks.
--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Paolo Valente - Nov. 21, 2012, 9:45 a.m.
Il 20/11/2012 19:54, David Miller ha scritto:
> From: Paolo Valente <paolo.valente@unimore.it>
> Date: Tue, 20 Nov 2012 18:45:13 +0100
>
>> -	struct sk_buff *skb;
>> +	struct sk_buff *skb = NULL;
>
> This is not really an improvement,
Sorry for trying this silly short cut
  now the compiler can think
> that NULL is passed eventually into qdisc_bstats_update().
>
> Please make the logic easier for the compiler to digest.
>
> For example, restructure the top-level logic into something like:
>
> 	skb = NULL;
> 	if (!list_empty(&in_serv_agg->active))
> 		skb = qfq_peek_skb(in_serv_agg, &cl, &len);
> 	else
> 		len = 0; /* no more active classes in the in-service agg */
>
> 	if (len == 0 || in_serv_agg->budget < len) {
>   ...
> 		/*
> 		 * If we get here, there are other aggregates queued:
> 		 * choose the new aggregate to serve.
> 		 */
> 		in_serv_agg = q->in_serv_agg = qfq_choose_next_agg(q);
> 		skb = qfq_peek_skb(in_serv_agg, &cl, &len);
> 	}
> 	if (!skb)
> 		return NULL;
>
> That way it is clearer, to both humans and the compiler, what is
> going on here.
>
Got it. Actually, if the first qfq_peek_skb returns NULL, then the 
example version that you are proposing apparently may behave in a 
different way than the original one: in your proposal the scheduler 
tries to switch to a new aggregate and may return a non-NULL value, 
whereas the original version would immediately return NULL. I guess that 
this slightly different behavior is fine as well, and I am preparing a 
new patch that integrates these changes.
> Thanks.
>
David Miller - Nov. 21, 2012, 5:04 p.m.
From: Paolo Valente <paolo.valente@unimore.it>
Date: Wed, 21 Nov 2012 10:45:14 +0100

> Got it. Actually, if the first qfq_peek_skb returns NULL, then the
> example version that you are proposing apparently may behave in a
> different way than the original one: in your proposal the scheduler
> tries to switch to a new aggregate and may return a non-NULL value,
> whereas the original version would immediately return NULL. I guess
> that this slightly different behavior is fine as well, and I am
> preparing a new patch that integrates these changes.

Thanks.
--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Patch

diff --git a/net/sched/sch_qfq.c b/net/sched/sch_qfq.c
index 9687fa1..fe1823a 100644
--- a/net/sched/sch_qfq.c
+++ b/net/sched/sch_qfq.c
@@ -1,7 +1,8 @@ 
 /*
- * net/sched/sch_qfq.c         Quick Fair Queueing Scheduler.
+ * net/sched/sch_qfq.c         Quick Fair Queueing Plus Scheduler.
  *
  * Copyright (c) 2009 Fabio Checconi, Luigi Rizzo, and Paolo Valente.
+ * Copyright (c) 2012 Paolo Valente.
  *
  * This program is free software; you can redistribute it and/or
  * modify it under the terms of the GNU General Public License
@@ -19,12 +20,18 @@ 
 #include <net/pkt_cls.h>
 
 
-/*  Quick Fair Queueing
-    ===================
+/*  Quick Fair Queueing Plus
+    ========================
 
     Sources:
 
-    Fabio Checconi, Luigi Rizzo, and Paolo Valente: "QFQ: Efficient
+    [1] Paolo Valente,
+    "Reducing the Execution Time of Fair-Queueing Schedulers."
+    http://algo.ing.unimo.it/people/paolo/agg-sched/agg-sched.pdf
+
+    Sources for QFQ:
+
+    [2] Fabio Checconi, Luigi Rizzo, and Paolo Valente: "QFQ: Efficient
     Packet Scheduling with Tight Bandwidth Distribution Guarantees."
 
     See also:
@@ -33,6 +40,20 @@ 
 
 /*
 
+  QFQ+ divides classes into aggregates of at most MAX_AGG_CLASSES
+  classes. Each aggregate is timestamped with a virtual start time S
+  and a virtual finish time F, and scheduled according to its
+  timestamps. S and F are computed as a function of a system virtual
+  time function V. The classes within each aggregate are instead
+  scheduled with DRR.
+
+  To speed up operations, QFQ+ divides also aggregates into a limited
+  number of groups. Which group a class belongs to depends on the
+  ratio between the maximum packet length for the class and the weight
+  of the class. Groups have their own S and F. In the end, QFQ+
+  schedules groups, then aggregates within groups, then classes within
+  aggregates. See [1] and [2] for a full description.
+
   Virtual time computations.
 
   S, F and V are all computed in fixed point arithmetic with
@@ -76,27 +97,28 @@ 
 #define QFQ_MAX_SLOTS	32
 
 /*
- * Shifts used for class<->group mapping.  We allow class weights that are
- * in the range [1, 2^MAX_WSHIFT], and we try to map each class i to the
+ * Shifts used for aggregate<->group mapping.  We allow class weights that are
+ * in the range [1, 2^MAX_WSHIFT], and we try to map each aggregate i to the
  * group with the smallest index that can support the L_i / r_i configured
- * for the class.
+ * for the classes in the aggregate.
  *
  * grp->index is the index of the group; and grp->slot_shift
  * is the shift for the corresponding (scaled) sigma_i.
  */
 #define QFQ_MAX_INDEX		24
-#define QFQ_MAX_WSHIFT		12
+#define QFQ_MAX_WSHIFT		10
 
-#define	QFQ_MAX_WEIGHT		(1<<QFQ_MAX_WSHIFT)
-#define QFQ_MAX_WSUM		(16*QFQ_MAX_WEIGHT)
+#define	QFQ_MAX_WEIGHT		(1<<QFQ_MAX_WSHIFT) /* see qfq_slot_insert */
+#define QFQ_MAX_WSUM		(64*QFQ_MAX_WEIGHT)
 
 #define FRAC_BITS		30	/* fixed point arithmetic */
 #define ONE_FP			(1UL << FRAC_BITS)
 #define IWSUM			(ONE_FP/QFQ_MAX_WSUM)
 
 #define QFQ_MTU_SHIFT		16	/* to support TSO/GSO */
-#define QFQ_MIN_SLOT_SHIFT	(FRAC_BITS + QFQ_MTU_SHIFT - QFQ_MAX_INDEX)
-#define QFQ_MIN_LMAX		256	/* min possible lmax for a class */
+#define QFQ_MIN_LMAX		512	/* see qfq_slot_insert */
+
+#define QFQ_MAX_AGG_CLASSES	8 /* max num classes per aggregate allowed */
 
 /*
  * Possible group states.  These values are used as indexes for the bitmaps
@@ -106,6 +128,8 @@  enum qfq_state { ER, IR, EB, IB, QFQ_MAX_STATE };
 
 struct qfq_group;
 
+struct qfq_aggregate;
+
 struct qfq_class {
 	struct Qdisc_class_common common;
 
@@ -116,7 +140,15 @@  struct qfq_class {
 	struct gnet_stats_queue qstats;
 	struct gnet_stats_rate_est rate_est;
 	struct Qdisc *qdisc;
+	struct list_head alist;		/* Link for active-classes list. */
+	struct qfq_aggregate *agg;	/* Parent aggregate. */
+	int deficit;			/* DRR deficit counter. */
+};
 
+/*
+
+ */
+struct qfq_aggregate {
 	struct hlist_node next;	/* Link for the slot list. */
 	u64 S, F;		/* flow timestamps (exact) */
 
@@ -127,8 +159,18 @@  struct qfq_class {
 	struct qfq_group *grp;
 
 	/* these are copied from the flowset. */
-	u32	inv_w;		/* ONE_FP/weight */
-	u32	lmax;		/* Max packet size for this flow. */
+	u32	class_weight; /* Weight of each class in this aggregate. */
+	/* Max pkt size for the classes in this aggregate, DRR quantum. */
+	int	lmax;
+
+	u32	inv_w;	    /* ONE_FP/(sum of weights of classes in aggr.). */
+	u32	budgetmax;  /* Max budget for this aggregate. */
+	u32	initial_budget, budget;     /* Initial and current budget. */
+
+	int		  num_classes;	/* Number of classes in this aggr. */
+	struct list_head  active;	/* DRR queue of active classes. */
+
+	struct hlist_node nonfull_next;	/* See nonfull_aggs in qfq_sched. */
 };
 
 struct qfq_group {
@@ -138,7 +180,7 @@  struct qfq_group {
 	unsigned int front;		/* Index of the front slot. */
 	unsigned long full_slots;	/* non-empty slots */
 
-	/* Array of RR lists of active classes. */
+	/* Array of RR lists of active aggregates. */
 	struct hlist_head slots[QFQ_MAX_SLOTS];
 };
 
@@ -146,13 +188,28 @@  struct qfq_sched {
 	struct tcf_proto *filter_list;
 	struct Qdisc_class_hash clhash;
 
-	u64		V;		/* Precise virtual time. */
-	u32		wsum;		/* weight sum */
+	u64			oldV, V;	/* Precise virtual times. */
+	struct qfq_aggregate	*in_serv_agg;   /* Aggregate being served. */
+	u32			num_active_agg; /* Num. of active aggregates */
+	u32			wsum;		/* weight sum */
 
 	unsigned long bitmaps[QFQ_MAX_STATE];	    /* Group bitmaps. */
 	struct qfq_group groups[QFQ_MAX_INDEX + 1]; /* The groups. */
+	u32 min_slot_shift;	/* Index of the group-0 bit in the bitmaps. */
+
+	u32 max_agg_classes;		/* Max number of classes per aggr. */
+	struct hlist_head nonfull_aggs; /* Aggs with room for more classes. */
 };
 
+/*
+ * Possible reasons why the timestamps of an aggregate are updated
+ * enqueue: the aggregate switches from idle to active and must scheduled
+ *	    for service
+ * requeue: the aggregate finishes its budget, so it stops being served and
+ *	    must be rescheduled for service
+ */
+enum update_reason {enqueue, requeue};
+
 static struct qfq_class *qfq_find_class(struct Qdisc *sch, u32 classid)
 {
 	struct qfq_sched *q = qdisc_priv(sch);
@@ -182,18 +239,18 @@  static const struct nla_policy qfq_policy[TCA_QFQ_MAX + 1] = {
  * index = log_2(maxlen/weight) but we need to apply the scaling.
  * This is used only once at flow creation.
  */
-static int qfq_calc_index(u32 inv_w, unsigned int maxlen)
+static int qfq_calc_index(u32 inv_w, unsigned int maxlen, u32 min_slot_shift)
 {
 	u64 slot_size = (u64)maxlen * inv_w;
 	unsigned long size_map;
 	int index = 0;
 
-	size_map = slot_size >> QFQ_MIN_SLOT_SHIFT;
+	size_map = slot_size >> min_slot_shift;
 	if (!size_map)
 		goto out;
 
 	index = __fls(size_map) + 1;	/* basically a log_2 */
-	index -= !(slot_size - (1ULL << (index + QFQ_MIN_SLOT_SHIFT - 1)));
+	index -= !(slot_size - (1ULL << (index + min_slot_shift - 1)));
 
 	if (index < 0)
 		index = 0;
@@ -204,66 +261,150 @@  out:
 	return index;
 }
 
-/* Length of the next packet (0 if the queue is empty). */
-static unsigned int qdisc_peek_len(struct Qdisc *sch)
+static void qfq_deactivate_agg(struct qfq_sched *, struct qfq_aggregate *);
+static void qfq_activate_agg(struct qfq_sched *, struct qfq_aggregate *,
+			     enum update_reason);
+
+static void qfq_init_agg(struct qfq_sched *q, struct qfq_aggregate *agg,
+			 u32 lmax, u32 weight)
 {
-	struct sk_buff *skb;
+	INIT_LIST_HEAD(&agg->active);
+	hlist_add_head(&agg->nonfull_next, &q->nonfull_aggs);
 
-	skb = sch->ops->peek(sch);
-	return skb ? qdisc_pkt_len(skb) : 0;
+	agg->lmax = lmax;
+	agg->class_weight = weight;
 }
 
-static void qfq_deactivate_class(struct qfq_sched *, struct qfq_class *);
-static void qfq_activate_class(struct qfq_sched *q, struct qfq_class *cl,
-			       unsigned int len);
+static struct qfq_aggregate *qfq_find_agg(struct qfq_sched *q,
+					  u32 lmax, u32 weight)
+{
+	struct qfq_aggregate *agg;
+	struct hlist_node *n;
 
-static void qfq_update_class_params(struct qfq_sched *q, struct qfq_class *cl,
-				    u32 lmax, u32 inv_w, int delta_w)
+	hlist_for_each_entry(agg, n, &q->nonfull_aggs, nonfull_next)
+		if (agg->lmax == lmax && agg->class_weight == weight)
+			return agg;
+
+	return NULL;
+}
+
+
+/* Update aggregate as a function of the new number of classes. */
+static void qfq_update_agg(struct qfq_sched *q, struct qfq_aggregate *agg,
+			   int new_num_classes)
 {
-	int i;
+	u32 new_agg_weight;
+
+	if (new_num_classes == q->max_agg_classes)
+		hlist_del_init(&agg->nonfull_next);
 
-	/* update qfq-specific data */
-	cl->lmax = lmax;
-	cl->inv_w = inv_w;
-	i = qfq_calc_index(cl->inv_w, cl->lmax);
+	if (agg->num_classes > new_num_classes &&
+	    new_num_classes == q->max_agg_classes - 1) /* agg no more full */
+		hlist_add_head(&agg->nonfull_next, &q->nonfull_aggs);
 
-	cl->grp = &q->groups[i];
+	agg->budgetmax = new_num_classes * agg->lmax;
+	new_agg_weight = agg->class_weight * new_num_classes;
+	agg->inv_w = ONE_FP/new_agg_weight;
 
-	q->wsum += delta_w;
+	if (agg->grp == NULL) {
+		int i = qfq_calc_index(agg->inv_w, agg->budgetmax,
+				       q->min_slot_shift);
+		agg->grp = &q->groups[i];
+	}
+
+	q->wsum +=
+		(int) agg->class_weight * (new_num_classes - agg->num_classes);
+
+	agg->num_classes = new_num_classes;
 }
 
-static void qfq_update_reactivate_class(struct qfq_sched *q,
-					struct qfq_class *cl,
-					u32 inv_w, u32 lmax, int delta_w)
+/* Add class to aggregate. */
+static void qfq_add_to_agg(struct qfq_sched *q,
+			   struct qfq_aggregate *agg,
+			   struct qfq_class *cl)
 {
-	bool need_reactivation = false;
-	int i = qfq_calc_index(inv_w, lmax);
+	cl->agg = agg;
+
+	qfq_update_agg(q, agg, agg->num_classes+1);
+	if (cl->qdisc->q.qlen > 0) { /* adding an active class */
+		list_add_tail(&cl->alist, &agg->active);
+		if (list_first_entry(&agg->active, struct qfq_class, alist) ==
+		    cl && q->in_serv_agg != agg) /* agg was inactive */
+			qfq_activate_agg(q, agg, enqueue); /* schedule agg */
+	}
+}
 
-	if (&q->groups[i] != cl->grp && cl->qdisc->q.qlen > 0) {
-		/*
-		 * shift cl->F back, to not charge the
-		 * class for the not-yet-served head
-		 * packet
-		 */
-		cl->F = cl->S;
-		/* remove class from its slot in the old group */
-		qfq_deactivate_class(q, cl);
-		need_reactivation = true;
+static struct qfq_aggregate *qfq_choose_next_agg(struct qfq_sched *);
+
+static void qfq_destroy_agg(struct qfq_sched *q, struct qfq_aggregate *agg)
+{
+	if (!hlist_unhashed(&agg->nonfull_next))
+		hlist_del_init(&agg->nonfull_next);
+	if (q->in_serv_agg == agg)
+		q->in_serv_agg = qfq_choose_next_agg(q);
+	kfree(agg);
+}
+
+/* Deschedule class from within its parent aggregate. */
+static void qfq_deactivate_class(struct qfq_sched *q, struct qfq_class *cl)
+{
+	struct qfq_aggregate *agg = cl->agg;
+
+
+	list_del(&cl->alist); /* remove from RR queue of the aggregate */
+	if (list_empty(&agg->active)) /* agg is now inactive */
+		qfq_deactivate_agg(q, agg);
+}
+
+/* Remove class from its parent aggregate. */
+static void qfq_rm_from_agg(struct qfq_sched *q, struct qfq_class *cl)
+{
+	struct qfq_aggregate *agg = cl->agg;
+
+	cl->agg = NULL;
+	if (agg->num_classes == 1) { /* agg being emptied, destroy it */
+		qfq_destroy_agg(q, agg);
+		return;
 	}
+	qfq_update_agg(q, agg, agg->num_classes-1);
+}
 
-	qfq_update_class_params(q, cl, lmax, inv_w, delta_w);
+/* Deschedule class and remove it from its parent aggregate. */
+static void qfq_deact_rm_from_agg(struct qfq_sched *q, struct qfq_class *cl)
+{
+	if (cl->qdisc->q.qlen > 0) /* class is active */
+		qfq_deactivate_class(q, cl);
 
-	if (need_reactivation) /* activate in new group */
-		qfq_activate_class(q, cl, qdisc_peek_len(cl->qdisc));
+	qfq_rm_from_agg(q, cl);
 }
 
+/* Move class to a new aggregate, matching the new class weight and/or lmax */
+static int qfq_change_agg(struct Qdisc *sch, struct qfq_class *cl, u32 weight,
+			   u32 lmax)
+{
+	struct qfq_sched *q = qdisc_priv(sch);
+	struct qfq_aggregate *new_agg = qfq_find_agg(q, lmax, weight);
+
+	if (new_agg == NULL) { /* create new aggregate */
+		new_agg = kzalloc(sizeof(*new_agg), GFP_ATOMIC);
+		if (new_agg == NULL)
+			return -ENOBUFS;
+		qfq_init_agg(q, new_agg, lmax, weight);
+	}
+	qfq_deact_rm_from_agg(q, cl);
+	qfq_add_to_agg(q, new_agg, cl);
+
+	return 0;
+}
 
 static int qfq_change_class(struct Qdisc *sch, u32 classid, u32 parentid,
 			    struct nlattr **tca, unsigned long *arg)
 {
 	struct qfq_sched *q = qdisc_priv(sch);
 	struct qfq_class *cl = (struct qfq_class *)*arg;
+	bool existing = false;
 	struct nlattr *tb[TCA_QFQ_MAX + 1];
+	struct qfq_aggregate *new_agg = NULL;
 	u32 weight, lmax, inv_w;
 	int err;
 	int delta_w;
@@ -286,15 +427,6 @@  static int qfq_change_class(struct Qdisc *sch, u32 classid, u32 parentid,
 	} else
 		weight = 1;
 
-	inv_w = ONE_FP / weight;
-	weight = ONE_FP / inv_w;
-	delta_w = weight - (cl ? ONE_FP / cl->inv_w : 0);
-	if (q->wsum + delta_w > QFQ_MAX_WSUM) {
-		pr_notice("qfq: total weight out of range (%u + %u)\n",
-			  delta_w, q->wsum);
-		return -EINVAL;
-	}
-
 	if (tb[TCA_QFQ_LMAX]) {
 		lmax = nla_get_u32(tb[TCA_QFQ_LMAX]);
 		if (lmax < QFQ_MIN_LMAX || lmax > (1UL << QFQ_MTU_SHIFT)) {
@@ -304,7 +436,23 @@  static int qfq_change_class(struct Qdisc *sch, u32 classid, u32 parentid,
 	} else
 		lmax = psched_mtu(qdisc_dev(sch));
 
-	if (cl != NULL) {
+	inv_w = ONE_FP / weight;
+	weight = ONE_FP / inv_w;
+
+	if (cl != NULL &&
+	    lmax == cl->agg->lmax &&
+	    weight == cl->agg->class_weight)
+		return 0; /* nothing to change */
+
+	delta_w = weight - (cl ? cl->agg->class_weight : 0);
+
+	if (q->wsum + delta_w > QFQ_MAX_WSUM) {
+		pr_notice("qfq: total weight out of range (%d + %u)\n",
+			  delta_w, q->wsum);
+		return -EINVAL;
+	}
+
+	if (cl != NULL) { /* modify existing class */
 		if (tca[TCA_RATE]) {
 			err = gen_replace_estimator(&cl->bstats, &cl->rate_est,
 						    qdisc_root_sleeping_lock(sch),
@@ -312,25 +460,18 @@  static int qfq_change_class(struct Qdisc *sch, u32 classid, u32 parentid,
 			if (err)
 				return err;
 		}
-
-		if (lmax == cl->lmax && inv_w == cl->inv_w)
-			return 0; /* nothing to update */
-
-		sch_tree_lock(sch);
-		qfq_update_reactivate_class(q, cl, inv_w, lmax, delta_w);
-		sch_tree_unlock(sch);
-
-		return 0;
+		existing = true;
+		goto set_change_agg;
 	}
 
+	/* create and init new class */
 	cl = kzalloc(sizeof(struct qfq_class), GFP_KERNEL);
 	if (cl == NULL)
 		return -ENOBUFS;
 
 	cl->refcnt = 1;
 	cl->common.classid = classid;
-
-	qfq_update_class_params(q, cl, lmax, inv_w, delta_w);
+	cl->deficit = lmax;
 
 	cl->qdisc = qdisc_create_dflt(sch->dev_queue,
 				      &pfifo_qdisc_ops, classid);
@@ -341,11 +482,8 @@  static int qfq_change_class(struct Qdisc *sch, u32 classid, u32 parentid,
 		err = gen_new_estimator(&cl->bstats, &cl->rate_est,
 					qdisc_root_sleeping_lock(sch),
 					tca[TCA_RATE]);
-		if (err) {
-			qdisc_destroy(cl->qdisc);
-			kfree(cl);
-			return err;
-		}
+		if (err)
+			goto destroy_class;
 	}
 
 	sch_tree_lock(sch);
@@ -354,19 +492,39 @@  static int qfq_change_class(struct Qdisc *sch, u32 classid, u32 parentid,
 
 	qdisc_class_hash_grow(sch, &q->clhash);
 
+set_change_agg:
+	sch_tree_lock(sch);
+	new_agg = qfq_find_agg(q, lmax, weight);
+	if (new_agg == NULL) { /* create new aggregate */
+		sch_tree_unlock(sch);
+		new_agg = kzalloc(sizeof(*new_agg), GFP_KERNEL);
+		if (new_agg == NULL) {
+			err = -ENOBUFS;
+			gen_kill_estimator(&cl->bstats, &cl->rate_est);
+			goto destroy_class;
+		}
+		sch_tree_lock(sch);
+		qfq_init_agg(q, new_agg, lmax, weight);
+	}
+	if (existing)
+		qfq_deact_rm_from_agg(q, cl);
+	qfq_add_to_agg(q, new_agg, cl);
+	sch_tree_unlock(sch);
+
 	*arg = (unsigned long)cl;
 	return 0;
+
+destroy_class:
+	qdisc_destroy(cl->qdisc);
+	kfree(cl);
+	return err;
 }
 
 static void qfq_destroy_class(struct Qdisc *sch, struct qfq_class *cl)
 {
 	struct qfq_sched *q = qdisc_priv(sch);
 
-	if (cl->inv_w) {
-		q->wsum -= ONE_FP / cl->inv_w;
-		cl->inv_w = 0;
-	}
-
+	qfq_rm_from_agg(q, cl);
 	gen_kill_estimator(&cl->bstats, &cl->rate_est);
 	qdisc_destroy(cl->qdisc);
 	kfree(cl);
@@ -481,8 +639,8 @@  static int qfq_dump_class(struct Qdisc *sch, unsigned long arg,
 	nest = nla_nest_start(skb, TCA_OPTIONS);
 	if (nest == NULL)
 		goto nla_put_failure;
-	if (nla_put_u32(skb, TCA_QFQ_WEIGHT, ONE_FP/cl->inv_w) ||
-	    nla_put_u32(skb, TCA_QFQ_LMAX, cl->lmax))
+	if (nla_put_u32(skb, TCA_QFQ_WEIGHT, cl->agg->class_weight) ||
+	    nla_put_u32(skb, TCA_QFQ_LMAX, cl->agg->lmax))
 		goto nla_put_failure;
 	return nla_nest_end(skb, nest);
 
@@ -500,8 +658,8 @@  static int qfq_dump_class_stats(struct Qdisc *sch, unsigned long arg,
 	memset(&xstats, 0, sizeof(xstats));
 	cl->qdisc->qstats.qlen = cl->qdisc->q.qlen;
 
-	xstats.weight = ONE_FP/cl->inv_w;
-	xstats.lmax = cl->lmax;
+	xstats.weight = cl->agg->class_weight;
+	xstats.lmax = cl->agg->lmax;
 
 	if (gnet_stats_copy_basic(d, &cl->bstats) < 0 ||
 	    gnet_stats_copy_rate_est(d, &cl->bstats, &cl->rate_est) < 0 ||
@@ -652,16 +810,16 @@  static void qfq_unblock_groups(struct qfq_sched *q, int index, u64 old_F)
  * perhaps
  *
 	old_V ^= q->V;
-	old_V >>= QFQ_MIN_SLOT_SHIFT;
+	old_V >>= q->min_slot_shift;
 	if (old_V) {
 		...
 	}
  *
  */
-static void qfq_make_eligible(struct qfq_sched *q, u64 old_V)
+static void qfq_make_eligible(struct qfq_sched *q)
 {
-	unsigned long vslot = q->V >> QFQ_MIN_SLOT_SHIFT;
-	unsigned long old_vslot = old_V >> QFQ_MIN_SLOT_SHIFT;
+	unsigned long vslot = q->V >> q->min_slot_shift;
+	unsigned long old_vslot = q->oldV >> q->min_slot_shift;
 
 	if (vslot != old_vslot) {
 		unsigned long mask = (1UL << fls(vslot ^ old_vslot)) - 1;
@@ -672,34 +830,38 @@  static void qfq_make_eligible(struct qfq_sched *q, u64 old_V)
 
 
 /*
- * If the weight and lmax (max_pkt_size) of the classes do not change,
- * then QFQ guarantees that the slot index is never higher than
- * 2 + ((1<<QFQ_MTU_SHIFT)/QFQ_MIN_LMAX) * (QFQ_MAX_WEIGHT/QFQ_MAX_WSUM).
+ * The index of the slot in which the aggregate is to be inserted must
+ * not be higher than QFQ_MAX_SLOTS-2. There is a '-2' and not a '-1'
+ * because the start time of the group may be moved backward by one
+ * slot after the aggregate has been inserted, and this would cause
+ * non-empty slots to be right-shifted by one position.
  *
- * With the current values of the above constants, the index is
- * then guaranteed to never be higher than 2 + 256 * (1 / 16) = 18.
+ * If the weight and lmax (max_pkt_size) of the classes do not change,
+ * then QFQ+ does meet the above contraint according to the current
+ * values of its parameters. In fact, if the weight and lmax of the
+ * classes do not change, then, from the theory, QFQ+ guarantees that
+ * the slot index is never higher than
+ * 2 + QFQ_MAX_AGG_CLASSES * ((1<<QFQ_MTU_SHIFT)/QFQ_MIN_LMAX) *
+ * (QFQ_MAX_WEIGHT/QFQ_MAX_WSUM) = 2 + 8 * 128 * (1 / 64) = 18
  *
  * When the weight of a class is increased or the lmax of the class is
- * decreased, a new class with smaller slot size may happen to be
- * activated. The activation of this class should be properly delayed
- * to when the service of the class has finished in the ideal system
- * tracked by QFQ. If the activation of the class is not delayed to
- * this reference time instant, then this class may be unjustly served
- * before other classes waiting for service. This may cause
- * (unfrequently) the above bound to the slot index to be violated for
- * some of these unlucky classes.
+ * decreased, a new aggregate with smaller slot size than the original
+ * parent aggregate of the class may happen to be activated. The
+ * activation of this aggregate should be properly delayed to when the
+ * service of the class has finished in the ideal system tracked by
+ * QFQ+. If the activation of the aggregate is not delayed to this
+ * reference time instant, then this aggregate may be unjustly served
+ * before other aggregates waiting for service. This may cause the
+ * above bound to the slot index to be violated for some of these
+ * unlucky aggregates.
  *
- * Instead of delaying the activation of the new class, which is quite
- * complex, the following inaccurate but simple solution is used: if
- * the slot index is higher than QFQ_MAX_SLOTS-2, then the timestamps
- * of the class are shifted backward so as to let the slot index
- * become equal to QFQ_MAX_SLOTS-2. This threshold is used because, if
- * the slot index is above it, then the data structure implementing
- * the bucket list either gets immediately corrupted or may get
- * corrupted on a possible next packet arrival that causes the start
- * time of the group to be shifted backward.
+ * Instead of delaying the activation of the new aggregate, which is
+ * quite complex, the following inaccurate but simple solution is used:
+ * if the slot index is higher than QFQ_MAX_SLOTS-2, then the
+ * timestamps of the aggregate are shifted backward so as to let the
+ * slot index become equal to QFQ_MAX_SLOTS-2.
  */
-static void qfq_slot_insert(struct qfq_group *grp, struct qfq_class *cl,
+static void qfq_slot_insert(struct qfq_group *grp, struct qfq_aggregate *agg,
 			    u64 roundedS)
 {
 	u64 slot = (roundedS - grp->S) >> grp->slot_shift;
@@ -708,22 +870,22 @@  static void qfq_slot_insert(struct qfq_group *grp, struct qfq_class *cl,
 	if (unlikely(slot > QFQ_MAX_SLOTS - 2)) {
 		u64 deltaS = roundedS - grp->S -
 			((u64)(QFQ_MAX_SLOTS - 2)<<grp->slot_shift);
-		cl->S -= deltaS;
-		cl->F -= deltaS;
+		agg->S -= deltaS;
+		agg->F -= deltaS;
 		slot = QFQ_MAX_SLOTS - 2;
 	}
 
 	i = (grp->front + slot) % QFQ_MAX_SLOTS;
 
-	hlist_add_head(&cl->next, &grp->slots[i]);
+	hlist_add_head(&agg->next, &grp->slots[i]);
 	__set_bit(slot, &grp->full_slots);
 }
 
 /* Maybe introduce hlist_first_entry?? */
-static struct qfq_class *qfq_slot_head(struct qfq_group *grp)
+static struct qfq_aggregate *qfq_slot_head(struct qfq_group *grp)
 {
 	return hlist_entry(grp->slots[grp->front].first,
-			   struct qfq_class, next);
+			   struct qfq_aggregate, next);
 }
 
 /*
@@ -731,20 +893,20 @@  static struct qfq_class *qfq_slot_head(struct qfq_group *grp)
  */
 static void qfq_front_slot_remove(struct qfq_group *grp)
 {
-	struct qfq_class *cl = qfq_slot_head(grp);
+	struct qfq_aggregate *agg = qfq_slot_head(grp);
 
-	BUG_ON(!cl);
-	hlist_del(&cl->next);
+	BUG_ON(!agg);
+	hlist_del(&agg->next);
 	if (hlist_empty(&grp->slots[grp->front]))
 		__clear_bit(0, &grp->full_slots);
 }
 
 /*
- * Returns the first full queue in a group. As a side effect,
- * adjust the bucket list so the first non-empty bucket is at
- * position 0 in full_slots.
+ * Returns the first aggregate in the first non-empty bucket of the
+ * group. As a side effect, adjusts the bucket list so the first
+ * non-empty bucket is at position 0 in full_slots.
  */
-static struct qfq_class *qfq_slot_scan(struct qfq_group *grp)
+static struct qfq_aggregate *qfq_slot_scan(struct qfq_group *grp)
 {
 	unsigned int i;
 
@@ -780,7 +942,7 @@  static void qfq_slot_rotate(struct qfq_group *grp, u64 roundedS)
 	grp->front = (grp->front - i) % QFQ_MAX_SLOTS;
 }
 
-static void qfq_update_eligible(struct qfq_sched *q, u64 old_V)
+static void qfq_update_eligible(struct qfq_sched *q)
 {
 	struct qfq_group *grp;
 	unsigned long ineligible;
@@ -792,137 +954,229 @@  static void qfq_update_eligible(struct qfq_sched *q, u64 old_V)
 			if (qfq_gt(grp->S, q->V))
 				q->V = grp->S;
 		}
-		qfq_make_eligible(q, old_V);
+		qfq_make_eligible(q);
 	}
 }
 
-/*
- * Updates the class, returns true if also the group needs to be updated.
- */
-static bool qfq_update_class(struct qfq_group *grp, struct qfq_class *cl)
+/* Dequeue head packet of the head class in the DRR queue of the aggregate. */
+static void agg_dequeue(struct qfq_aggregate *agg,
+			struct qfq_class *cl, unsigned int len)
 {
-	unsigned int len = qdisc_peek_len(cl->qdisc);
+	qdisc_dequeue_peeked(cl->qdisc);
 
-	cl->S = cl->F;
-	if (!len)
-		qfq_front_slot_remove(grp);	/* queue is empty */
-	else {
-		u64 roundedS;
-
-		cl->F = cl->S + (u64)len * cl->inv_w;
-		roundedS = qfq_round_down(cl->S, grp->slot_shift);
-		if (roundedS == grp->S)
-			return false;
+	cl->deficit -= (int) len;
 
-		qfq_front_slot_remove(grp);
-		qfq_slot_insert(grp, cl, roundedS);
+	if (cl->qdisc->q.qlen == 0) /* no more packets, remove from list */
+		list_del(&cl->alist);
+	else if (cl->deficit < qdisc_pkt_len(cl->qdisc->ops->peek(cl->qdisc))) {
+		cl->deficit += agg->lmax;
+		list_move_tail(&cl->alist, &agg->active);
 	}
+}
+
+static inline struct sk_buff *qfq_peek_skb(struct qfq_aggregate *agg,
+					   struct qfq_class **cl,
+					   unsigned int *len)
+{
+	struct sk_buff *skb;
 
-	return true;
+	*cl = list_first_entry(&agg->active, struct qfq_class, alist);
+	skb = (*cl)->qdisc->ops->peek((*cl)->qdisc);
+	if (skb == NULL)
+		WARN_ONCE(1, "qfq_dequeue: non-workconserving leaf\n");
+	else
+		*len = qdisc_pkt_len(skb);
+
+	return skb;
+}
+
+/* Update F according to the actual service received by the aggregate. */
+static inline void charge_actual_service(struct qfq_aggregate *agg)
+{
+	/* compute the service received by the aggregate */
+	u32 service_received = agg->initial_budget - agg->budget;
+
+	agg->F = agg->S + (u64)service_received * agg->inv_w;
 }
 
 static struct sk_buff *qfq_dequeue(struct Qdisc *sch)
 {
 	struct qfq_sched *q = qdisc_priv(sch);
-	struct qfq_group *grp;
+	struct qfq_aggregate *in_serv_agg = q->in_serv_agg;
 	struct qfq_class *cl;
-	struct sk_buff *skb;
+	struct sk_buff *skb = NULL;
 	unsigned int len;
-	u64 old_V;
 
-	if (!q->bitmaps[ER])
+	if (in_serv_agg == NULL)
 		return NULL;
 
-	grp = qfq_ffs(q, q->bitmaps[ER]);
+	if (!list_empty(&in_serv_agg->active)) {
+		skb = qfq_peek_skb(in_serv_agg, &cl, &len);
+		if (!skb)
+			return NULL;
+	} else
+		len = 0; /* no more active classes in the in-service agg */
 
-	cl = qfq_slot_head(grp);
-	skb = qdisc_dequeue_peeked(cl->qdisc);
-	if (!skb) {
-		WARN_ONCE(1, "qfq_dequeue: non-workconserving leaf\n");
-		return NULL;
+	/*
+	 * If there are no active classes in the in-service aggregate,
+	 * or if the aggregate has not enough budget to serve its next
+	 * class, then choose the next aggregate to serve.
+	 */
+	if (len == 0 || in_serv_agg->budget < len) {
+		charge_actual_service(in_serv_agg);
+
+		/* recharge the budget of the aggregate */
+		in_serv_agg->initial_budget = in_serv_agg->budget =
+			in_serv_agg->budgetmax;
+
+		if (!list_empty(&in_serv_agg->active))
+			/*
+			 * Still active: reschedule for
+			 * service. Possible optimization: if no other
+			 * aggregate is active, then there is no point
+			 * in rescheduling this aggregate, and we can
+			 * just keep it as the in-service one. This
+			 * should be however a corner case, and to
+			 * handle it, we would need to maintain an
+			 * extra num_active_aggs field.
+			*/
+			qfq_activate_agg(q, in_serv_agg, requeue);
+		else if (sch->q.qlen == 0) { /* no aggregate to serve */
+			q->in_serv_agg = NULL;
+			return NULL;
+		}
+
+		/*
+		 * If we get here, there are other aggregates queued:
+		 * choose the new aggregate to serve.
+		 */
+		in_serv_agg = q->in_serv_agg = qfq_choose_next_agg(q);
+		skb = qfq_peek_skb(in_serv_agg, &cl, &len);
+		if (!skb)
+			return NULL;
 	}
 
 	sch->q.qlen--;
 	qdisc_bstats_update(sch, skb);
 
-	old_V = q->V;
-	len = qdisc_pkt_len(skb);
+	agg_dequeue(in_serv_agg, cl, len);
+	in_serv_agg->budget -= len;
 	q->V += (u64)len * IWSUM;
 	pr_debug("qfq dequeue: len %u F %lld now %lld\n",
-		 len, (unsigned long long) cl->F, (unsigned long long) q->V);
+		 len, (unsigned long long) in_serv_agg->F,
+		 (unsigned long long) q->V);
 
-	if (qfq_update_class(grp, cl)) {
-		u64 old_F = grp->F;
+	return skb;
+}
 
-		cl = qfq_slot_scan(grp);
-		if (!cl)
-			__clear_bit(grp->index, &q->bitmaps[ER]);
-		else {
-			u64 roundedS = qfq_round_down(cl->S, grp->slot_shift);
-			unsigned int s;
+static struct qfq_aggregate *qfq_choose_next_agg(struct qfq_sched *q)
+{
+	struct qfq_group *grp;
+	struct qfq_aggregate *agg, *new_front_agg;
+	u64 old_F;
 
-			if (grp->S == roundedS)
-				goto skip_unblock;
-			grp->S = roundedS;
-			grp->F = roundedS + (2ULL << grp->slot_shift);
-			__clear_bit(grp->index, &q->bitmaps[ER]);
-			s = qfq_calc_state(q, grp);
-			__set_bit(grp->index, &q->bitmaps[s]);
-		}
+	qfq_update_eligible(q);
+	q->oldV = q->V;
+
+	if (!q->bitmaps[ER])
+		return NULL;
+
+	grp = qfq_ffs(q, q->bitmaps[ER]);
+	old_F = grp->F;
 
-		qfq_unblock_groups(q, grp->index, old_F);
+	agg = qfq_slot_head(grp);
+
+	/* agg starts to be served, remove it from schedule */
+	qfq_front_slot_remove(grp);
+
+	new_front_agg = qfq_slot_scan(grp);
+
+	if (new_front_agg == NULL) /* group is now inactive, remove from ER */
+		__clear_bit(grp->index, &q->bitmaps[ER]);
+	else {
+		u64 roundedS = qfq_round_down(new_front_agg->S,
+					      grp->slot_shift);
+		unsigned int s;
+
+		if (grp->S == roundedS)
+			return agg;
+		grp->S = roundedS;
+		grp->F = roundedS + (2ULL << grp->slot_shift);
+		__clear_bit(grp->index, &q->bitmaps[ER]);
+		s = qfq_calc_state(q, grp);
+		__set_bit(grp->index, &q->bitmaps[s]);
 	}
 
-skip_unblock:
-	qfq_update_eligible(q, old_V);
+	qfq_unblock_groups(q, grp->index, old_F);
 
-	return skb;
+	return agg;
 }
 
 /*
- * Assign a reasonable start time for a new flow k in group i.
+ * Assign a reasonable start time for a new aggregate in group i.
  * Admissible values for \hat(F) are multiples of \sigma_i
  * no greater than V+\sigma_i . Larger values mean that
  * we had a wraparound so we consider the timestamp to be stale.
  *
  * If F is not stale and F >= V then we set S = F.
  * Otherwise we should assign S = V, but this may violate
- * the ordering in ER. So, if we have groups in ER, set S to
- * the F_j of the first group j which would be blocking us.
+ * the ordering in EB (see [2]). So, if we have groups in ER,
+ * set S to the F_j of the first group j which would be blocking us.
  * We are guaranteed not to move S backward because
  * otherwise our group i would still be blocked.
  */
-static void qfq_update_start(struct qfq_sched *q, struct qfq_class *cl)
+static void qfq_update_start(struct qfq_sched *q, struct qfq_aggregate *agg)
 {
 	unsigned long mask;
 	u64 limit, roundedF;
-	int slot_shift = cl->grp->slot_shift;
+	int slot_shift = agg->grp->slot_shift;
 
-	roundedF = qfq_round_down(cl->F, slot_shift);
+	roundedF = qfq_round_down(agg->F, slot_shift);
 	limit = qfq_round_down(q->V, slot_shift) + (1ULL << slot_shift);
 
-	if (!qfq_gt(cl->F, q->V) || qfq_gt(roundedF, limit)) {
+	if (!qfq_gt(agg->F, q->V) || qfq_gt(roundedF, limit)) {
 		/* timestamp was stale */
-		mask = mask_from(q->bitmaps[ER], cl->grp->index);
+		mask = mask_from(q->bitmaps[ER], agg->grp->index);
 		if (mask) {
 			struct qfq_group *next = qfq_ffs(q, mask);
 			if (qfq_gt(roundedF, next->F)) {
 				if (qfq_gt(limit, next->F))
-					cl->S = next->F;
+					agg->S = next->F;
 				else /* preserve timestamp correctness */
-					cl->S = limit;
+					agg->S = limit;
 				return;
 			}
 		}
-		cl->S = q->V;
+		agg->S = q->V;
 	} else  /* timestamp is not stale */
-		cl->S = cl->F;
+		agg->S = agg->F;
+}
+
+/*
+ * Update the timestamps of agg before scheduling/rescheduling it for
+ * service.  In particular, assign to agg->F its maximum possible
+ * value, i.e., the virtual finish time with which the aggregate
+ * should be labeled if it used all its budget once in service.
+ */
+static inline void
+qfq_update_agg_ts(struct qfq_sched *q,
+		    struct qfq_aggregate *agg, enum update_reason reason)
+{
+	if (reason != requeue)
+		qfq_update_start(q, agg);
+	else /* just charge agg for the service received */
+		agg->S = agg->F;
+
+	agg->F = agg->S + (u64)agg->budgetmax * agg->inv_w;
 }
 
+static void qfq_schedule_agg(struct qfq_sched *, struct qfq_aggregate *);
+
 static int qfq_enqueue(struct sk_buff *skb, struct Qdisc *sch)
 {
 	struct qfq_sched *q = qdisc_priv(sch);
 	struct qfq_class *cl;
+	struct qfq_aggregate *agg;
 	int err = 0;
 
 	cl = qfq_classify(skb, sch, &err);
@@ -934,11 +1188,13 @@  static int qfq_enqueue(struct sk_buff *skb, struct Qdisc *sch)
 	}
 	pr_debug("qfq_enqueue: cl = %x\n", cl->common.classid);
 
-	if (unlikely(cl->lmax < qdisc_pkt_len(skb))) {
+	if (unlikely(cl->agg->lmax < qdisc_pkt_len(skb))) {
 		pr_debug("qfq: increasing maxpkt from %u to %u for class %u",
-			  cl->lmax, qdisc_pkt_len(skb), cl->common.classid);
-		qfq_update_reactivate_class(q, cl, cl->inv_w,
-					    qdisc_pkt_len(skb), 0);
+			 cl->agg->lmax, qdisc_pkt_len(skb), cl->common.classid);
+		err = qfq_change_agg(sch, cl, cl->agg->class_weight,
+				     qdisc_pkt_len(skb));
+		if (err)
+			return err;
 	}
 
 	err = qdisc_enqueue(skb, cl->qdisc);
@@ -954,35 +1210,50 @@  static int qfq_enqueue(struct sk_buff *skb, struct Qdisc *sch)
 	bstats_update(&cl->bstats, skb);
 	++sch->q.qlen;
 
-	/* If the new skb is not the head of queue, then done here. */
-	if (cl->qdisc->q.qlen != 1)
+	agg = cl->agg;
+	/* if the queue was not empty, then done here */
+	if (cl->qdisc->q.qlen != 1) {
+		if (unlikely(skb == cl->qdisc->ops->peek(cl->qdisc)) &&
+		    list_first_entry(&agg->active, struct qfq_class, alist)
+		    == cl && cl->deficit < qdisc_pkt_len(skb))
+			list_move_tail(&cl->alist, &agg->active);
+
 		return err;
+	}
+
+	/* schedule class for service within the aggregate */
+	cl->deficit = agg->lmax;
+	list_add_tail(&cl->alist, &agg->active);
+
+	if (list_first_entry(&agg->active, struct qfq_class, alist) != cl)
+		return err; /* aggregate was not empty, nothing else to do */
+
+	/* recharge budget */
+	agg->initial_budget = agg->budget = agg->budgetmax;
 
-	/* If reach this point, queue q was idle */
-	qfq_activate_class(q, cl, qdisc_pkt_len(skb));
+	qfq_update_agg_ts(q, agg, enqueue);
+	if (q->in_serv_agg == NULL)
+		q->in_serv_agg = agg;
+	else if (agg != q->in_serv_agg)
+		qfq_schedule_agg(q, agg);
 
 	return err;
 }
 
 /*
- * Handle class switch from idle to backlogged.
+ * Schedule aggregate according to its timestamps.
  */
-static void qfq_activate_class(struct qfq_sched *q, struct qfq_class *cl,
-			       unsigned int pkt_len)
+static void qfq_schedule_agg(struct qfq_sched *q, struct qfq_aggregate *agg)
 {
-	struct qfq_group *grp = cl->grp;
+	struct qfq_group *grp = agg->grp;
 	u64 roundedS;
 	int s;
 
-	qfq_update_start(q, cl);
-
-	/* compute new finish time and rounded start. */
-	cl->F = cl->S + (u64)pkt_len * cl->inv_w;
-	roundedS = qfq_round_down(cl->S, grp->slot_shift);
+	roundedS = qfq_round_down(agg->S, grp->slot_shift);
 
 	/*
-	 * insert cl in the correct bucket.
-	 * If cl->S >= grp->S we don't need to adjust the
+	 * Insert agg in the correct bucket.
+	 * If agg->S >= grp->S we don't need to adjust the
 	 * bucket list and simply go to the insertion phase.
 	 * Otherwise grp->S is decreasing, we must make room
 	 * in the bucket list, and also recompute the group state.
@@ -990,10 +1261,10 @@  static void qfq_activate_class(struct qfq_sched *q, struct qfq_class *cl,
 	 * was in ER make sure to adjust V.
 	 */
 	if (grp->full_slots) {
-		if (!qfq_gt(grp->S, cl->S))
+		if (!qfq_gt(grp->S, agg->S))
 			goto skip_update;
 
-		/* create a slot for this cl->S */
+		/* create a slot for this agg->S */
 		qfq_slot_rotate(grp, roundedS);
 		/* group was surely ineligible, remove */
 		__clear_bit(grp->index, &q->bitmaps[IR]);
@@ -1008,46 +1279,61 @@  static void qfq_activate_class(struct qfq_sched *q, struct qfq_class *cl,
 
 	pr_debug("qfq enqueue: new state %d %#lx S %lld F %lld V %lld\n",
 		 s, q->bitmaps[s],
-		 (unsigned long long) cl->S,
-		 (unsigned long long) cl->F,
+		 (unsigned long long) agg->S,
+		 (unsigned long long) agg->F,
 		 (unsigned long long) q->V);
 
 skip_update:
-	qfq_slot_insert(grp, cl, roundedS);
+	qfq_slot_insert(grp, agg, roundedS);
 }
 
 
+/* Update agg ts and schedule agg for service */
+static void qfq_activate_agg(struct qfq_sched *q, struct qfq_aggregate *agg,
+			     enum update_reason reason)
+{
+	qfq_update_agg_ts(q, agg, reason);
+	qfq_schedule_agg(q, agg);
+}
+
 static void qfq_slot_remove(struct qfq_sched *q, struct qfq_group *grp,
-			    struct qfq_class *cl)
+			    struct qfq_aggregate *agg)
 {
 	unsigned int i, offset;
 	u64 roundedS;
 
-	roundedS = qfq_round_down(cl->S, grp->slot_shift);
+	roundedS = qfq_round_down(agg->S, grp->slot_shift);
 	offset = (roundedS - grp->S) >> grp->slot_shift;
+
 	i = (grp->front + offset) % QFQ_MAX_SLOTS;
 
-	hlist_del(&cl->next);
+	hlist_del(&agg->next);
 	if (hlist_empty(&grp->slots[i]))
 		__clear_bit(offset, &grp->full_slots);
 }
 
 /*
- * called to forcibly destroy a queue.
- * If the queue is not in the front bucket, or if it has
- * other queues in the front bucket, we can simply remove
- * the queue with no other side effects.
+ * Called to forcibly deschedule an aggregate.  If the aggregate is
+ * not in the front bucket, or if the latter has other aggregates in
+ * the front bucket, we can simply remove the aggregate with no other
+ * side effects.
  * Otherwise we must propagate the event up.
  */
-static void qfq_deactivate_class(struct qfq_sched *q, struct qfq_class *cl)
+static void qfq_deactivate_agg(struct qfq_sched *q, struct qfq_aggregate *agg)
 {
-	struct qfq_group *grp = cl->grp;
+	struct qfq_group *grp = agg->grp;
 	unsigned long mask;
 	u64 roundedS;
 	int s;
 
-	cl->F = cl->S;
-	qfq_slot_remove(q, grp, cl);
+	if (agg == q->in_serv_agg) {
+		charge_actual_service(agg);
+		q->in_serv_agg = qfq_choose_next_agg(q);
+		return;
+	}
+
+	agg->F = agg->S;
+	qfq_slot_remove(q, grp, agg);
 
 	if (!grp->full_slots) {
 		__clear_bit(grp->index, &q->bitmaps[IR]);
@@ -1066,8 +1352,8 @@  static void qfq_deactivate_class(struct qfq_sched *q, struct qfq_class *cl)
 		}
 		__clear_bit(grp->index, &q->bitmaps[ER]);
 	} else if (hlist_empty(&grp->slots[grp->front])) {
-		cl = qfq_slot_scan(grp);
-		roundedS = qfq_round_down(cl->S, grp->slot_shift);
+		agg = qfq_slot_scan(grp);
+		roundedS = qfq_round_down(agg->S, grp->slot_shift);
 		if (grp->S != roundedS) {
 			__clear_bit(grp->index, &q->bitmaps[ER]);
 			__clear_bit(grp->index, &q->bitmaps[IR]);
@@ -1080,7 +1366,7 @@  static void qfq_deactivate_class(struct qfq_sched *q, struct qfq_class *cl)
 		}
 	}
 
-	qfq_update_eligible(q, q->V);
+	qfq_update_eligible(q);
 }
 
 static void qfq_qlen_notify(struct Qdisc *sch, unsigned long arg)
@@ -1092,6 +1378,32 @@  static void qfq_qlen_notify(struct Qdisc *sch, unsigned long arg)
 		qfq_deactivate_class(q, cl);
 }
 
+static unsigned int qfq_drop_from_slot(struct qfq_sched *q,
+				       struct hlist_head *slot)
+{
+	struct qfq_aggregate *agg;
+	struct hlist_node *n;
+	struct qfq_class *cl;
+	unsigned int len;
+
+	hlist_for_each_entry(agg, n, slot, next) {
+		list_for_each_entry(cl, &agg->active, alist) {
+
+			if (!cl->qdisc->ops->drop)
+				continue;
+
+			len = cl->qdisc->ops->drop(cl->qdisc);
+			if (len > 0) {
+				if (cl->qdisc->q.qlen == 0)
+					qfq_deactivate_class(q, cl);
+
+				return len;
+			}
+		}
+	}
+	return 0;
+}
+
 static unsigned int qfq_drop(struct Qdisc *sch)
 {
 	struct qfq_sched *q = qdisc_priv(sch);
@@ -1101,24 +1413,13 @@  static unsigned int qfq_drop(struct Qdisc *sch)
 	for (i = 0; i <= QFQ_MAX_INDEX; i++) {
 		grp = &q->groups[i];
 		for (j = 0; j < QFQ_MAX_SLOTS; j++) {
-			struct qfq_class *cl;
-			struct hlist_node *n;
-
-			hlist_for_each_entry(cl, n, &grp->slots[j], next) {
-
-				if (!cl->qdisc->ops->drop)
-					continue;
-
-				len = cl->qdisc->ops->drop(cl->qdisc);
-				if (len > 0) {
-					sch->q.qlen--;
-					if (!cl->qdisc->q.qlen)
-						qfq_deactivate_class(q, cl);
-
-					return len;
-				}
+			len = qfq_drop_from_slot(q, &grp->slots[j]);
+			if (len > 0) {
+				sch->q.qlen--;
+				return len;
 			}
 		}
+
 	}
 
 	return 0;
@@ -1129,44 +1430,51 @@  static int qfq_init_qdisc(struct Qdisc *sch, struct nlattr *opt)
 	struct qfq_sched *q = qdisc_priv(sch);
 	struct qfq_group *grp;
 	int i, j, err;
+	u32 max_cl_shift, maxbudg_shift, max_classes;
 
 	err = qdisc_class_hash_init(&q->clhash);
 	if (err < 0)
 		return err;
 
+	if (qdisc_dev(sch)->tx_queue_len + 1 > QFQ_MAX_AGG_CLASSES)
+		max_classes = QFQ_MAX_AGG_CLASSES;
+	else
+		max_classes = qdisc_dev(sch)->tx_queue_len + 1;
+	/* max_cl_shift = floor(log_2(max_classes)) */
+	max_cl_shift = __fls(max_classes);
+	q->max_agg_classes = 1<<max_cl_shift;
+
+	/* maxbudg_shift = log2(max_len * max_classes_per_agg) */
+	maxbudg_shift = QFQ_MTU_SHIFT + max_cl_shift;
+	q->min_slot_shift = FRAC_BITS + maxbudg_shift - QFQ_MAX_INDEX;
+
 	for (i = 0; i <= QFQ_MAX_INDEX; i++) {
 		grp = &q->groups[i];
 		grp->index = i;
-		grp->slot_shift = QFQ_MTU_SHIFT + FRAC_BITS
-				   - (QFQ_MAX_INDEX - i);
+		grp->slot_shift = q->min_slot_shift + i;
 		for (j = 0; j < QFQ_MAX_SLOTS; j++)
 			INIT_HLIST_HEAD(&grp->slots[j]);
 	}
 
+	INIT_HLIST_HEAD(&q->nonfull_aggs);
+
 	return 0;
 }
 
 static void qfq_reset_qdisc(struct Qdisc *sch)
 {
 	struct qfq_sched *q = qdisc_priv(sch);
-	struct qfq_group *grp;
 	struct qfq_class *cl;
-	struct hlist_node *n, *tmp;
-	unsigned int i, j;
+	struct hlist_node *n;
+	unsigned int i;
 
-	for (i = 0; i <= QFQ_MAX_INDEX; i++) {
-		grp = &q->groups[i];
-		for (j = 0; j < QFQ_MAX_SLOTS; j++) {
-			hlist_for_each_entry_safe(cl, n, tmp,
-						  &grp->slots[j], next) {
+	for (i = 0; i < q->clhash.hashsize; i++) {
+		hlist_for_each_entry(cl, n, &q->clhash.hash[i], common.hnode) {
+			if (cl->qdisc->q.qlen > 0)
 				qfq_deactivate_class(q, cl);
-			}
-		}
-	}
 
-	for (i = 0; i < q->clhash.hashsize; i++) {
-		hlist_for_each_entry(cl, n, &q->clhash.hash[i], common.hnode)
 			qdisc_reset(cl->qdisc);
+		}
 	}
 	sch->q.qlen = 0;
 }