net-qdisc-hhf: Heavy-Hitter Filter (HHF) qdisc

This patch implements the first size-based qdisc that attempts to
differentiate between small flows and heavy-hitters.  The goal is to
catch the heavy-hitters and move them to a separate queue with less
priority so that bulk traffic does not affect the latency of critical
traffic.  Currently "less priority" means less weight (2:1 in
particular) in a Weighted Deficit Round Robin (WDRR) scheduler.

In essence, this patch addresses the "delay-bloat" problem due to
bloated buffers. In some systems, large queues may be necessary for
obtaining CPU efficiency, or due to the presence of unresponsive
traffic like UDP, or just a large number of connections with each
having a small amount of outstanding traffic. In these circumstances,
HHF aims to reduce the HoL blocking for latency sensitive traffic,
while not impacting the queues built up by bulk traffic.  HHF can also
be used in conjunction with other AQM mechanisms such as CoDel.

To capture heavy-hitters, we implement the "multi-stage filter" design
in the following paper:
C. Estan and G. Varghese, "New Directions in Traffic Measurement and
Accounting", in ACM SIGCOMM, 2002.

Some configurable qdisc settings through 'tc':
- hhf_reset_timeout: period to reset counter values in the multi-stage
                     filter (default 40ms)
- hhf_admit_bytes:   threshold to classify heavy-hitters
                     (default 128KB)
- hhf_evict_timeout: threshold to evict idle heavy-hitters
                     (default 1s)
- hhf_non_hh_weight: Weighted Deficit Round Robin (WDRR) weight for
                     non-heavy-hitters (default 2)
- hh_flows_limit:    max number of heavy-hitter flow entries
                     (default 2048)

Note that the ratio between hhf_admit_bytes and hhf_reset_timeout
reflects the bandwidth of heavy-hitters that we attempt to capture
(25Mbps with the above default settings).

The false negative rate (heavy-hitter flows getting away unclassified)
is zero by the design of the multi-stage filter algorithm.
With 100 heavy-hitter flows, using four hashes and 4000 counters yields
a false positive rate (non-heavy-hitters mistakenly classified as
heavy-hitters) of less than 1e-4.
Signed-off-by: Terry Lam <vtlam@google.com>
Acked-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

include/uapi/linux/pkt_sched.h

@@ -790,4 +790,29 @@ struct tc_fq_qd_stats {
 	__u32	throttled_flows;
 	__u32	pad;
 };
+
+/* Heavy-Hitter Filter */
+
+enum {
+	TCA_HHF_UNSPEC,
+	TCA_HHF_BACKLOG_LIMIT,
+	TCA_HHF_QUANTUM,
+	TCA_HHF_HH_FLOWS_LIMIT,
+	TCA_HHF_RESET_TIMEOUT,
+	TCA_HHF_ADMIT_BYTES,
+	TCA_HHF_EVICT_TIMEOUT,
+	TCA_HHF_NON_HH_WEIGHT,
+	__TCA_HHF_MAX
+};
+
+#define TCA_HHF_MAX	(__TCA_HHF_MAX - 1)
+
+struct tc_hhf_xstats {
+	__u32	drop_overlimit; /* number of times max qdisc packet limit
+				 * was hit
+				 */
+	__u32	hh_overlimit;   /* number of times max heavy-hitters was hit */
+	__u32	hh_tot_count;   /* number of captured heavy-hitters so far */
+	__u32	hh_cur_count;   /* number of current heavy-hitters */
+};
 #endif

net/sched/Kconfig

@@ -286,6 +286,15 @@ config NET_SCH_FQ
 
 	  If unsure, say N.
 
+config NET_SCH_HHF
+	tristate "Heavy-Hitter Filter (HHF)"
+	help
+	  Say Y here if you want to use the Heavy-Hitter Filter (HHF)
+	  packet scheduling algorithm.
+
+	  To compile this driver as a module, choose M here: the module
+	  will be called sch_hhf.
+
 config NET_SCH_INGRESS
 	tristate "Ingress Qdisc"
 	depends on NET_CLS_ACT

net/sched/Makefile

@@ -40,6 +40,7 @@ obj-$(CONFIG_NET_SCH_QFQ)	+= sch_qfq.o
 obj-$(CONFIG_NET_SCH_CODEL)	+= sch_codel.o
 obj-$(CONFIG_NET_SCH_FQ_CODEL)	+= sch_fq_codel.o
 obj-$(CONFIG_NET_SCH_FQ)	+= sch_fq.o
+obj-$(CONFIG_NET_SCH_HHF)	+= sch_hhf.o
 
 obj-$(CONFIG_NET_CLS_U32)	+= cls_u32.o
 obj-$(CONFIG_NET_CLS_ROUTE4)	+= cls_route.o

net/sched/sch_hhf.c

+/* net/sched/sch_hhf.c		Heavy-Hitter Filter (HHF)
+ *
+ * Copyright (C) 2013 Terry Lam <vtlam@google.com>
+ * Copyright (C) 2013 Nandita Dukkipati <nanditad@google.com>
+ */
+
+#include <linux/jhash.h>
+#include <linux/jiffies.h>
+#include <linux/module.h>
+#include <linux/skbuff.h>
+#include <linux/vmalloc.h>
+#include <net/flow_keys.h>
+#include <net/pkt_sched.h>
+#include <net/sock.h>
+
+/*	Heavy-Hitter Filter (HHF)
+ *
+ * Principles :
+ * Flows are classified into two buckets: non-heavy-hitter and heavy-hitter
+ * buckets. Initially, a new flow starts as non-heavy-hitter. Once classified
+ * as heavy-hitter, it is immediately switched to the heavy-hitter bucket.
+ * The buckets are dequeued by a Weighted Deficit Round Robin (WDRR) scheduler,
+ * in which the heavy-hitter bucket is served with less weight.
+ * In other words, non-heavy-hitters (e.g., short bursts of critical traffic)
+ * are isolated from heavy-hitters (e.g., persistent bulk traffic) and also have
+ * higher share of bandwidth.
+ *
+ * To capture heavy-hitters, we use the "multi-stage filter" algorithm in the
+ * following paper:
+ * [EV02] C. Estan and G. Varghese, "New Directions in Traffic Measurement and
+ * Accounting", in ACM SIGCOMM, 2002.
+ *
+ * Conceptually, a multi-stage filter comprises k independent hash functions
+ * and k counter arrays. Packets are indexed into k counter arrays by k hash
+ * functions, respectively. The counters are then increased by the packet sizes.
+ * Therefore,
+ *    - For a heavy-hitter flow: *all* of its k array counters must be large.
+ *    - For a non-heavy-hitter flow: some of its k array counters can be large
+ *      due to hash collision with other small flows; however, with high
+ *      probability, not *all* k counters are large.
+ *
+ * By the design of the multi-stage filter algorithm, the false negative rate
+ * (heavy-hitters getting away uncaptured) is zero. However, the algorithm is
+ * susceptible to false positives (non-heavy-hitters mistakenly classified as
+ * heavy-hitters).
+ * Therefore, we also implement the following optimizations to reduce false
+ * positives by avoiding unnecessary increment of the counter values:
+ *    - Optimization O1: once a heavy-hitter is identified, its bytes are not
+ *        accounted in the array counters. This technique is called "shielding"
+ *        in Section 3.3.1 of [EV02].
+ *    - Optimization O2: conservative update of counters
+ *                       (Section 3.3.2 of [EV02]),
+ *        New counter value = max {old counter value,
+ *                                 smallest counter value + packet bytes}
+ *
+ * Finally, we refresh the counters periodically since otherwise the counter
+ * values will keep accumulating.
+ *
+ * Once a flow is classified as heavy-hitter, we also save its per-flow state
+ * in an exact-matching flow table so that its subsequent packets can be
+ * dispatched to the heavy-hitter bucket accordingly.
+ *
+ *
+ * At a high level, this qdisc works as follows:
+ * Given a packet p:
+ *   - If the flow-id of p (e.g., TCP 5-tuple) is already in the exact-matching
+ *     heavy-hitter flow table, denoted table T, then send p to the heavy-hitter
+ *     bucket.
+ *   - Otherwise, forward p to the multi-stage filter, denoted filter F
+ *        + If F decides that p belongs to a non-heavy-hitter flow, then send p
+ *          to the non-heavy-hitter bucket.
+ *        + Otherwise, if F decides that p belongs to a new heavy-hitter flow,
+ *          then set up a new flow entry for the flow-id of p in the table T and
+ *          send p to the heavy-hitter bucket.
+ *
+ * In this implementation:
+ *   - T is a fixed-size hash-table with 1024 entries. Hash collision is
+ *     resolved by linked-list chaining.
+ *   - F has four counter arrays, each array containing 1024 32-bit counters.
+ *     That means 4 * 1024 * 32 bits = 16KB of memory.
+ *   - Since each array in F contains 1024 counters, 10 bits are sufficient to
+ *     index into each array.
+ *     Hence, instead of having four hash functions, we chop the 32-bit
+ *     skb-hash into three 10-bit chunks, and the remaining 10-bit chunk is
+ *     computed as XOR sum of those three chunks.
+ *   - We need to clear the counter arrays periodically; however, directly
+ *     memsetting 16KB of memory can lead to cache eviction and unwanted delay.
+ *     So by representing each counter by a valid bit, we only need to reset
+ *     4K of 1 bit (i.e. 512 bytes) instead of 16KB of memory.
+ *   - The Deficit Round Robin engine is taken from fq_codel implementation
+ *     (net/sched/sch_fq_codel.c). Note that wdrr_bucket corresponds to
+ *     fq_codel_flow in fq_codel implementation.
+ *
+ */
+
+/* Non-configurable parameters */
+#define HH_FLOWS_CNT	 1024  /* number of entries in exact-matching table T */
+#define HHF_ARRAYS_CNT	 4     /* number of arrays in multi-stage filter F */
+#define HHF_ARRAYS_LEN	 1024  /* number of counters in each array of F */
+#define HHF_BIT_MASK_LEN 10    /* masking 10 bits */
+#define HHF_BIT_MASK	 0x3FF /* bitmask of 10 bits */
+
+#define WDRR_BUCKET_CNT  2     /* two buckets for Weighted DRR */
+enum wdrr_bucket_idx {
+	WDRR_BUCKET_FOR_HH	= 0, /* bucket id for heavy-hitters */
+	WDRR_BUCKET_FOR_NON_HH	= 1  /* bucket id for non-heavy-hitters */
+};
+
+#define hhf_time_before(a, b)	\
+	(typecheck(u32, a) && typecheck(u32, b) && ((s32)((a) - (b)) < 0))
+
+/* Heavy-hitter per-flow state */
+struct hh_flow_state {
+	u32		 hash_id;	/* hash of flow-id (e.g. TCP 5-tuple) */
+	u32		 hit_timestamp;	/* last time heavy-hitter was seen */
+	struct list_head flowchain;	/* chaining under hash collision */
+};
+
+/* Weighted Deficit Round Robin (WDRR) scheduler */
+struct wdrr_bucket {
+	struct sk_buff	  *head;
+	struct sk_buff	  *tail;
+	struct list_head  bucketchain;
+	int		  deficit;
+};
+
+struct hhf_sched_data {
+	struct wdrr_bucket buckets[WDRR_BUCKET_CNT];
+	u32		   perturbation;   /* hash perturbation */
+	u32		   quantum;        /* psched_mtu(qdisc_dev(sch)); */
+	u32		   drop_overlimit; /* number of times max qdisc packet
+					    * limit was hit
+					    */
+	struct list_head   *hh_flows;       /* table T (currently active HHs) */
+	u32		   hh_flows_limit;            /* max active HH allocs */
+	u32		   hh_flows_overlimit; /* num of disallowed HH allocs */
+	u32		   hh_flows_total_cnt;          /* total admitted HHs */
+	u32		   hh_flows_current_cnt;        /* total current HHs  */
+	u32		   *hhf_arrays[HHF_ARRAYS_CNT]; /* HH filter F */
+	u32		   hhf_arrays_reset_timestamp;  /* last time hhf_arrays
+							 * was reset
+							 */
+	unsigned long	   *hhf_valid_bits[HHF_ARRAYS_CNT]; /* shadow valid bits
+							     * of hhf_arrays
+							     */
+	/* Similar to the "new_flows" vs. "old_flows" concept in fq_codel DRR */
+	struct list_head   new_buckets; /* list of new buckets */
+	struct list_head   old_buckets; /* list of old buckets */
+
+	/* Configurable HHF parameters */
+	u32		   hhf_reset_timeout; /* interval to reset counter
+					       * arrays in filter F
+					       * (default 40ms)
+					       */
+	u32		   hhf_admit_bytes;   /* counter thresh to classify as
+					       * HH (default 128KB).
+					       * With these default values,
+					       * 128KB / 40ms = 25 Mbps
+					       * i.e., we expect to capture HHs
+					       * sending > 25 Mbps.
+					       */
+	u32		   hhf_evict_timeout; /* aging threshold to evict idle
+					       * HHs out of table T. This should
+					       * be large enough to avoid
+					       * reordering during HH eviction.
+					       * (default 1s)
+					       */
+	u32		   hhf_non_hh_weight; /* WDRR weight for non-HHs
+					       * (default 2,
+					       *  i.e., non-HH : HH = 2 : 1)
+					       */
+};
+
+static u32 hhf_time_stamp(void)
+{
+	return jiffies;
+}
+
+static unsigned int skb_hash(const struct hhf_sched_data *q,
+			     const struct sk_buff *skb)
+{
+	struct flow_keys keys;
+	unsigned int hash;
+
+	if (skb->sk && skb->sk->sk_hash)
+		return skb->sk->sk_hash;
+
+	skb_flow_dissect(skb, &keys);
+	hash = jhash_3words((__force u32)keys.dst,
+			    (__force u32)keys.src ^ keys.ip_proto,
+			    (__force u32)keys.ports, q->perturbation);
+	return hash;
+}
+
+/* Looks up a heavy-hitter flow in a chaining list of table T. */
+static struct hh_flow_state *seek_list(const u32 hash,
+				       struct list_head *head,
+				       struct hhf_sched_data *q)
+{
+	struct hh_flow_state *flow, *next;
+	u32 now = hhf_time_stamp();
+
+	if (list_empty(head))
+		return NULL;
+
+	list_for_each_entry_safe(flow, next, head, flowchain) {
+		u32 prev = flow->hit_timestamp + q->hhf_evict_timeout;
+
+		if (hhf_time_before(prev, now)) {
+			/* Delete expired heavy-hitters, but preserve one entry
+			 * to avoid kzalloc() when next time this slot is hit.
+			 */
+			if (list_is_last(&flow->flowchain, head))
+				return NULL;
+			list_del(&flow->flowchain);
+			kfree(flow);
+			q->hh_flows_current_cnt--;
+		} else if (flow->hash_id == hash) {
+			return flow;
+		}
+	}
+	return NULL;
+}
+
+/* Returns a flow state entry for a new heavy-hitter.  Either reuses an expired
+ * entry or dynamically alloc a new entry.
+ */
+static struct hh_flow_state *alloc_new_hh(struct list_head *head,
+					  struct hhf_sched_data *q)
+{
+	struct hh_flow_state *flow;
+	u32 now = hhf_time_stamp();
+
+	if (!list_empty(head)) {
+		/* Find an expired heavy-hitter flow entry. */
+		list_for_each_entry(flow, head, flowchain) {
+			u32 prev = flow->hit_timestamp + q->hhf_evict_timeout;
+
+			if (hhf_time_before(prev, now))
+				return flow;
+		}
+	}
+
+	if (q->hh_flows_current_cnt >= q->hh_flows_limit) {
+		q->hh_flows_overlimit++;
+		return NULL;
+	}
+	/* Create new entry. */
+	flow = kzalloc(sizeof(struct hh_flow_state), GFP_ATOMIC);
+	if (!flow)
+		return NULL;
+
+	q->hh_flows_current_cnt++;
+	INIT_LIST_HEAD(&flow->flowchain);
+	list_add_tail(&flow->flowchain, head);
+
+	return flow;
+}
+
+/* Assigns packets to WDRR buckets.  Implements a multi-stage filter to
+ * classify heavy-hitters.
+ */
+static enum wdrr_bucket_idx hhf_classify(struct sk_buff *skb, struct Qdisc *sch)
+{
+	struct hhf_sched_data *q = qdisc_priv(sch);
+	u32 tmp_hash, hash;
+	u32 xorsum, filter_pos[HHF_ARRAYS_CNT], flow_pos;
+	struct hh_flow_state *flow;
+	u32 pkt_len, min_hhf_val;
+	int i;
+	u32 prev;
+	u32 now = hhf_time_stamp();
+
+	/* Reset the HHF counter arrays if this is the right time. */
+	prev = q->hhf_arrays_reset_timestamp + q->hhf_reset_timeout;
+	if (hhf_time_before(prev, now)) {
+		for (i = 0; i < HHF_ARRAYS_CNT; i++)
+			bitmap_zero(q->hhf_valid_bits[i], HHF_ARRAYS_LEN);
+		q->hhf_arrays_reset_timestamp = now;
+	}
+
+	/* Get hashed flow-id of the skb. */
+	hash = skb_hash(q, skb);
+
+	/* Check if this packet belongs to an already established HH flow. */
+	flow_pos = hash & HHF_BIT_MASK;
+	flow = seek_list(hash, &q->hh_flows[flow_pos], q);
+	if (flow) { /* found its HH flow */
+		flow->hit_timestamp = now;
+		return WDRR_BUCKET_FOR_HH;
+	}
+
+	/* Now pass the packet through the multi-stage filter. */
+	tmp_hash = hash;
+	xorsum = 0;
+	for (i = 0; i < HHF_ARRAYS_CNT - 1; i++) {
+		/* Split the skb_hash into three 10-bit chunks. */
+		filter_pos[i] = tmp_hash & HHF_BIT_MASK;
+		xorsum ^= filter_pos[i];
+		tmp_hash >>= HHF_BIT_MASK_LEN;
+	}
+	/* The last chunk is computed as XOR sum of other chunks. */
+	filter_pos[HHF_ARRAYS_CNT - 1] = xorsum ^ tmp_hash;
+
+	pkt_len = qdisc_pkt_len(skb);
+	min_hhf_val = ~0U;
+	for (i = 0; i < HHF_ARRAYS_CNT; i++) {
+		u32 val;
+
+		if (!test_bit(filter_pos[i], q->hhf_valid_bits[i])) {
+			q->hhf_arrays[i][filter_pos[i]] = 0;
+			__set_bit(filter_pos[i], q->hhf_valid_bits[i]);
+		}
+
+		val = q->hhf_arrays[i][filter_pos[i]] + pkt_len;
+		if (min_hhf_val > val)
+			min_hhf_val = val;
+	}
+
+	/* Found a new HH iff all counter values > HH admit threshold. */
+	if (min_hhf_val > q->hhf_admit_bytes) {
+		/* Just captured a new heavy-hitter. */
+		flow = alloc_new_hh(&q->hh_flows[flow_pos], q);
+		if (!flow) /* memory alloc problem */
+			return WDRR_BUCKET_FOR_NON_HH;
+		flow->hash_id = hash;
+		flow->hit_timestamp = now;
+		q->hh_flows_total_cnt++;
+
+		/* By returning without updating counters in q->hhf_arrays,
+		 * we implicitly implement "shielding" (see Optimization O1).
+		 */
+		return WDRR_BUCKET_FOR_HH;
+	}
+
+	/* Conservative update of HHF arrays (see Optimization O2). */
+	for (i = 0; i < HHF_ARRAYS_CNT; i++) {
+		if (q->hhf_arrays[i][filter_pos[i]] < min_hhf_val)
+			q->hhf_arrays[i][filter_pos[i]] = min_hhf_val;
+	}
+	return WDRR_BUCKET_FOR_NON_HH;
+}
+
+/* Removes one skb from head of bucket. */
+static struct sk_buff *dequeue_head(struct wdrr_bucket *bucket)
+{
+	struct sk_buff *skb = bucket->head;
+
+	bucket->head = skb->next;
+	skb->next = NULL;
+	return skb;
+}
+
+/* Tail-adds skb to bucket. */
+static void bucket_add(struct wdrr_bucket *bucket, struct sk_buff *skb)
+{
+	if (bucket->head == NULL)
+		bucket->head = skb;
+	else
+		bucket->tail->next = skb;
+	bucket->tail = skb;
+	skb->next = NULL;
+}
+
+static unsigned int hhf_drop(struct Qdisc *sch)
+{
+	struct hhf_sched_data *q = qdisc_priv(sch);
+	struct wdrr_bucket *bucket;
+
+	/* Always try to drop from heavy-hitters first. */
+	bucket = &q->buckets[WDRR_BUCKET_FOR_HH];
+	if (!bucket->head)
+		bucket = &q->buckets[WDRR_BUCKET_FOR_NON_HH];
+
+	if (bucket->head) {
+		struct sk_buff *skb = dequeue_head(bucket);
+
+		sch->q.qlen--;
+		sch->qstats.drops++;
+		sch->qstats.backlog -= qdisc_pkt_len(skb);
+		kfree_skb(skb);
+	}
+
+	/* Return id of the bucket from which the packet was dropped. */
+	return bucket - q->buckets;
+}
+
+static int hhf_enqueue(struct sk_buff *skb, struct Qdisc *sch)
+{
+	struct hhf_sched_data *q = qdisc_priv(sch);
+	enum wdrr_bucket_idx idx;
+	struct wdrr_bucket *bucket;
+
+	idx = hhf_classify(skb, sch);
+
+	bucket = &q->buckets[idx];
+	bucket_add(bucket, skb);
+	sch->qstats.backlog += qdisc_pkt_len(skb);
+
+	if (list_empty(&bucket->bucketchain)) {
+		unsigned int weight;
+
+		/* The logic of new_buckets vs. old_buckets is the same as
+		 * new_flows vs. old_flows in the implementation of fq_codel,
+		 * i.e., short bursts of non-HHs should have strict priority.
+		 */
+		if (idx == WDRR_BUCKET_FOR_HH) {
+			/* Always move heavy-hitters to old bucket. */
+			weight = 1;
+			list_add_tail(&bucket->bucketchain, &q->old_buckets);
+		} else {
+			weight = q->hhf_non_hh_weight;
+			list_add_tail(&bucket->bucketchain, &q->new_buckets);
+		}
+		bucket->deficit = weight * q->quantum;
+	}
+	if (++sch->q.qlen < sch->limit)
+		return NET_XMIT_SUCCESS;
+
+	q->drop_overlimit++;
+	/* Return Congestion Notification only if we dropped a packet from this
+	 * bucket.
+	 */
+	if (hhf_drop(sch) == idx)
+		return NET_XMIT_CN;
+
+	/* As we dropped a packet, better let upper stack know this. */
+	qdisc_tree_decrease_qlen(sch, 1);
+	return NET_XMIT_SUCCESS;
+}
+
+static struct sk_buff *hhf_dequeue(struct Qdisc *sch)
+{
+	struct hhf_sched_data *q = qdisc_priv(sch);
+	struct sk_buff *skb = NULL;
+	struct wdrr_bucket *bucket;
+	struct list_head *head;
+
+begin:
+	head = &q->new_buckets;
+	if (list_empty(head)) {
+		head = &q->old_buckets;
+		if (list_empty(head))
+			return NULL;
+	}
+	bucket = list_first_entry(head, struct wdrr_bucket, bucketchain);
+
+	if (bucket->deficit <= 0) {
+		int weight = (bucket - q->buckets == WDRR_BUCKET_FOR_HH) ?
+			      1 : q->hhf_non_hh_weight;
+
+		bucket->deficit += weight * q->quantum;
+		list_move_tail(&bucket->bucketchain, &q->old_buckets);
+		goto begin;
+	}
+
+	if (bucket->head) {
+		skb = dequeue_head(bucket);
+		sch->q.qlen--;
+		sch->qstats.backlog -= qdisc_pkt_len(skb);
+	}
+
+	if (!skb) {
+		/* Force a pass through old_buckets to prevent starvation. */
+		if ((head == &q->new_buckets) && !list_empty(&q->old_buckets))
+			list_move_tail(&bucket->bucketchain, &q->old_buckets);
+		else
+			list_del_init(&bucket->bucketchain);
+		goto begin;
+	}
+	qdisc_bstats_update(sch, skb);
+	bucket->deficit -= qdisc_pkt_len(skb);
+
+	return skb;
+}
+
+static void hhf_reset(struct Qdisc *sch)
+{
+	struct sk_buff *skb;
+
+	while ((skb = hhf_dequeue(sch)) != NULL)
+		kfree_skb(skb);
+}
+
+static void *hhf_zalloc(size_t sz)
+{
+	void *ptr = kzalloc(sz, GFP_KERNEL | __GFP_NOWARN);
+
+	if (!ptr)
+		ptr = vzalloc(sz);
+
+	return ptr;
+}
+
+static void hhf_free(void *addr)
+{
+	if (addr) {
+		if (is_vmalloc_addr(addr))
+			vfree(addr);
+		else
+			kfree(addr);
+	}
+}
+
+static void hhf_destroy(struct Qdisc *sch)
+{
+	int i;
+	struct hhf_sched_data *q = qdisc_priv(sch);
+
+	for (i = 0; i < HHF_ARRAYS_CNT; i++) {
+		hhf_free(q->hhf_arrays[i]);
+		hhf_free(q->hhf_valid_bits[i]);
+	}
+
+	for (i = 0; i < HH_FLOWS_CNT; i++) {
+		struct hh_flow_state *flow, *next;
+		struct list_head *head = &q->hh_flows[i];
+
+		if (list_empty(head))
+			continue;
+		list_for_each_entry_safe(flow, next, head, flowchain) {
+			list_del(&flow->flowchain);
+			kfree(flow);
+		}
+	}
+	hhf_free(q->hh_flows);
+}
+
+static const struct nla_policy hhf_policy[TCA_HHF_MAX + 1] = {
+	[TCA_HHF_BACKLOG_LIMIT]	 = { .type = NLA_U32 },
+	[TCA_HHF_QUANTUM]	 = { .type = NLA_U32 },
+	[TCA_HHF_HH_FLOWS_LIMIT] = { .type = NLA_U32 },
+	[TCA_HHF_RESET_TIMEOUT]	 = { .type = NLA_U32 },
+	[TCA_HHF_ADMIT_BYTES]	 = { .type = NLA_U32 },
+	[TCA_HHF_EVICT_TIMEOUT]	 = { .type = NLA_U32 },
+	[TCA_HHF_NON_HH_WEIGHT]	 = { .type = NLA_U32 },
+};
+
+static int hhf_change(struct Qdisc *sch, struct nlattr *opt)
+{
+	struct hhf_sched_data *q = qdisc_priv(sch);
+	struct nlattr *tb[TCA_HHF_MAX + 1];
+	unsigned int qlen;
+	int err;
+	u64 non_hh_quantum;
+	u32 new_quantum = q->quantum;
+	u32 new_hhf_non_hh_weight = q->hhf_non_hh_weight;
+
+	if (!opt)
+		return -EINVAL;
+
+	err = nla_parse_nested(tb, TCA_HHF_MAX, opt, hhf_policy);
+	if (err < 0)
+		return err;
+
+	sch_tree_lock(sch);
+
+	if (tb[TCA_HHF_BACKLOG_LIMIT])
+		sch->limit = nla_get_u32(tb[TCA_HHF_BACKLOG_LIMIT]);
+
+	if (tb[TCA_HHF_QUANTUM])
+		new_quantum = nla_get_u32(tb[TCA_HHF_QUANTUM]);
+
+	if (tb[TCA_HHF_NON_HH_WEIGHT])
+		new_hhf_non_hh_weight = nla_get_u32(tb[TCA_HHF_NON_HH_WEIGHT]);
+
+	non_hh_quantum = (u64)new_quantum * new_hhf_non_hh_weight;
+	if (non_hh_quantum > INT_MAX)
+		return -EINVAL;
+	q->quantum = new_quantum;
+	q->hhf_non_hh_weight = new_hhf_non_hh_weight;
+
+	if (tb[TCA_HHF_HH_FLOWS_LIMIT])
+		q->hh_flows_limit = nla_get_u32(tb[TCA_HHF_HH_FLOWS_LIMIT]);
+
+	if (tb[TCA_HHF_RESET_TIMEOUT]) {
+		u32 ms = nla_get_u32(tb[TCA_HHF_RESET_TIMEOUT]);
+
+		q->hhf_reset_timeout = msecs_to_jiffies(ms);
+	}
+
+	if (tb[TCA_HHF_ADMIT_BYTES])
+		q->hhf_admit_bytes = nla_get_u32(tb[TCA_HHF_ADMIT_BYTES]);
+
+	if (tb[TCA_HHF_EVICT_TIMEOUT]) {
+		u32 ms = nla_get_u32(tb[TCA_HHF_EVICT_TIMEOUT]);
+
+		q->hhf_evict_timeout = msecs_to_jiffies(ms);
+	}
+
+	qlen = sch->q.qlen;
+	while (sch->q.qlen > sch->limit) {
+		struct sk_buff *skb = hhf_dequeue(sch);
+
+		kfree_skb(skb);
+	}
+	qdisc_tree_decrease_qlen(sch, qlen - sch->q.qlen);
+
+	sch_tree_unlock(sch);
+	return 0;
+}
+
+static int hhf_init(struct Qdisc *sch, struct nlattr *opt)
+{
+	struct hhf_sched_data *q = qdisc_priv(sch);
+	int i;
+
+	sch->limit = 1000;
+	q->quantum = psched_mtu(qdisc_dev(sch));
+	q->perturbation = net_random();
+	INIT_LIST_HEAD(&q->new_buckets);
+	INIT_LIST_HEAD(&q->old_buckets);
+
+	/* Configurable HHF parameters */
+	q->hhf_reset_timeout = HZ / 25; /* 40  ms */
+	q->hhf_admit_bytes = 131072;    /* 128 KB */
+	q->hhf_evict_timeout = HZ;      /* 1  sec */
+	q->hhf_non_hh_weight = 2;
+
+	if (opt) {
+		int err = hhf_change(sch, opt);
+
+		if (err)
+			return err;
+	}
+
+	if (!q->hh_flows) {
+		/* Initialize heavy-hitter flow table. */
+		q->hh_flows = hhf_zalloc(HH_FLOWS_CNT *
+					 sizeof(struct list_head));
+		if (!q->hh_flows)
+			return -ENOMEM;
+		for (i = 0; i < HH_FLOWS_CNT; i++)
+			INIT_LIST_HEAD(&q->hh_flows[i]);
+
+		/* Cap max active HHs at twice len of hh_flows table. */
+		q->hh_flows_limit = 2 * HH_FLOWS_CNT;
+		q->hh_flows_overlimit = 0;
+		q->hh_flows_total_cnt = 0;
+		q->hh_flows_current_cnt = 0;
+
+		/* Initialize heavy-hitter filter arrays. */
+		for (i = 0; i < HHF_ARRAYS_CNT; i++) {
+			q->hhf_arrays[i] = hhf_zalloc(HHF_ARRAYS_LEN *
+						      sizeof(u32));
+			if (!q->hhf_arrays[i]) {
+				hhf_destroy(sch);
+				return -ENOMEM;
+			}
+		}
+		q->hhf_arrays_reset_timestamp = hhf_time_stamp();
+
+		/* Initialize valid bits of heavy-hitter filter arrays. */
+		for (i = 0; i < HHF_ARRAYS_CNT; i++) {
+			q->hhf_valid_bits[i] = hhf_zalloc(HHF_ARRAYS_LEN /
+							  BITS_PER_BYTE);
+			if (!q->hhf_valid_bits[i]) {
+				hhf_destroy(sch);
+				return -ENOMEM;
+			}
+		}
+
+		/* Initialize Weighted DRR buckets. */
+		for (i = 0; i < WDRR_BUCKET_CNT; i++) {
+			struct wdrr_bucket *bucket = q->buckets + i;
+
+			INIT_LIST_HEAD(&bucket->bucketchain);
+		}
+	}
+
+	return 0;
+}
+
+static int hhf_dump(struct Qdisc *sch, struct sk_buff *skb)
+{
+	struct hhf_sched_data *q = qdisc_priv(sch);
+	struct nlattr *opts;
+
+	opts = nla_nest_start(skb, TCA_OPTIONS);
+	if (opts == NULL)
+		goto nla_put_failure;
+
+	if (nla_put_u32(skb, TCA_HHF_BACKLOG_LIMIT, sch->limit) ||
+	    nla_put_u32(skb, TCA_HHF_QUANTUM, q->quantum) ||
+	    nla_put_u32(skb, TCA_HHF_HH_FLOWS_LIMIT, q->hh_flows_limit) ||
+	    nla_put_u32(skb, TCA_HHF_RESET_TIMEOUT,
+			jiffies_to_msecs(q->hhf_reset_timeout)) ||
+	    nla_put_u32(skb, TCA_HHF_ADMIT_BYTES, q->hhf_admit_bytes) ||
+	    nla_put_u32(skb, TCA_HHF_EVICT_TIMEOUT,
+			jiffies_to_msecs(q->hhf_evict_timeout)) ||
+	    nla_put_u32(skb, TCA_HHF_NON_HH_WEIGHT, q->hhf_non_hh_weight))
+		goto nla_put_failure;
+
+	nla_nest_end(skb, opts);
+	return skb->len;
+
+nla_put_failure:
+	return -1;
+}
+
+static int hhf_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
+{
+	struct hhf_sched_data *q = qdisc_priv(sch);
+	struct tc_hhf_xstats st = {
+		.drop_overlimit = q->drop_overlimit,
+		.hh_overlimit	= q->hh_flows_overlimit,
+		.hh_tot_count	= q->hh_flows_total_cnt,
+		.hh_cur_count	= q->hh_flows_current_cnt,
+	};
+
+	return gnet_stats_copy_app(d, &st, sizeof(st));
+}
+
+struct Qdisc_ops hhf_qdisc_ops __read_mostly = {
+	.id		=	"hhf",
+	.priv_size	=	sizeof(struct hhf_sched_data),
+
+	.enqueue	=	hhf_enqueue,
+	.dequeue	=	hhf_dequeue,
+	.peek		=	qdisc_peek_dequeued,
+	.drop		=	hhf_drop,
+	.init		=	hhf_init,
+	.reset		=	hhf_reset,
+	.destroy	=	hhf_destroy,
+	.change		=	hhf_change,
+	.dump		=	hhf_dump,
+	.dump_stats	=	hhf_dump_stats,
+	.owner		=	THIS_MODULE,
+};
+EXPORT_SYMBOL(hhf_qdisc_ops);
+
+static int __init hhf_module_init(void)
+{
+	return register_qdisc(&hhf_qdisc_ops);
+}
+
+static void __exit hhf_module_exit(void)
+{
+	unregister_qdisc(&hhf_qdisc_ops);
+}
+
+module_init(hhf_module_init)
+module_exit(hhf_module_exit)
+MODULE_AUTHOR("Terry Lam");
+MODULE_AUTHOR("Nandita Dukkipati");
+MODULE_LICENSE("GPL");