net_sched: sch_fq: add 3 bands and WRR scheduling

Before Google adopted FQ for its production servers,
we had to ensure AF4 packets would get a higher share
than BE1 ones.

As discussed this week in Netconf 2023 in Paris, it is time
to upstream this for public use.

After this patch FQ can replace pfifo_fast, with the following
differences :

- FQ uses WRR instead of strict prio, to avoid starvation of
  low priority packets.

- We make sure each band/prio tracks its own usage against sch->limit.
  This was done to make sure flood of low priority packets would not
  prevent AF4 packets to be queued. Contributed by Willem.

- priomap can be changed, if needed (default value are the ones
  coming from pfifo_fast).

In this patch, we set default band weights so that :

- high prio (band=0) packets get 90% of the bandwidth
  if they compete with low prio (band=2) packets.

- high prio packets get 75% of the bandwidth
  if they compete with medium prio (band=1) packets.

Following patch in this series adds the possibility to tune
the per-band weights.

As we added many fields in 'struct fq_sched_data', we had
to make sure to have the first cache line read-mostly, and
avoid wasting precious cache lines.

More optimizations are possible but will be sent separately.

Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Dave Taht <dave.taht@gmail.com>
Reviewed-by: Willem de Bruijn <willemb@google.com>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Reviewed-by: Toke Høiland-Jørgensen <toke@redhat.com>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>

2 files changed, 171 insertions, 44 deletions

diff --git a/include/uapi/linux/pkt_sched.h b/include/uapi/linux/pkt_sched.h
index 579f641846b8..ec5ab44d41a2 100644
--- a/include/uapi/linux/pkt_sched.h
+++ b/include/uapi/linux/pkt_sched.h
@@ -941,15 +941,19 @@ enum {
 
 	TCA_FQ_HORIZON_DROP,	/* drop packets beyond horizon, or cap their EDT */
 
+	TCA_FQ_PRIOMAP,		/* prio2band */
+
 	__TCA_FQ_MAX
 };
 
 #define TCA_FQ_MAX	(__TCA_FQ_MAX - 1)
 
+#define FQ_BANDS 3
+
 struct tc_fq_qd_stats {
 	__u64	gc_flows;
-	__u64	highprio_packets;
-	__u64	tcp_retrans;
+	__u64	highprio_packets;	/* obsolete */
+	__u64	tcp_retrans;		/* obsolete */
 	__u64	throttled;
 	__u64	flows_plimit;
 	__u64	pkts_too_long;
@@ -963,6 +967,9 @@ struct tc_fq_qd_stats {
 	__u64	horizon_drops;
 	__u64	horizon_caps;
 	__u64	fastpath_packets;
+	__u64	band_drops[FQ_BANDS];
+	__u32	band_pkt_count[FQ_BANDS];
+	__u32	pad;
 };
 
 /* Heavy-Hitter Filter */
diff --git a/net/sched/sch_fq.c b/net/sched/sch_fq.c
index 818ac786379d..081105801fa6 100644
--- a/net/sched/sch_fq.c
+++ b/net/sched/sch_fq.c
@@ -51,7 +51,8 @@
 #include <net/tcp.h>
 
 struct fq_skb_cb {
-	u64	        time_to_send;
+	u64	time_to_send;
+	u8	band;
 };
 
 static inline struct fq_skb_cb *fq_skb_cb(struct sk_buff *skb)
@@ -84,32 +85,28 @@ struct fq_flow {
 	u32		socket_hash;	/* sk_hash */
 	int		qlen;		/* number of packets in flow queue */
 
-/* Second cache line, used in fq_dequeue() */
+/* Second cache line */
 	int		credit;
-	/* 32bit hole on 64bit arches */
-
+	int		band;
 	struct fq_flow *next;		/* next pointer in RR lists */
 
 	struct rb_node  rate_node;	/* anchor in q->delayed tree */
 	u64		time_next_packet;
-} ____cacheline_aligned_in_smp;
+};
 
 struct fq_flow_head {
 	struct fq_flow *first;
 	struct fq_flow *last;
 };
 
-struct fq_sched_data {
+struct fq_perband_flows {
 	struct fq_flow_head new_flows;
-
 	struct fq_flow_head old_flows;
+	int		    credit;
+	int		    quantum; /* based on band nr : 576KB, 192KB, 64KB */
+};
 
-	struct rb_root	delayed;	/* for rate limited flows */
-	u64		time_next_delayed_flow;
-	unsigned long	unthrottle_latency_ns;
-
-	struct fq_flow	internal;	/* for non classified or high prio packets */
-
+struct fq_sched_data {
 /* Read mostly cache line */
 
 	u32		quantum;
@@ -125,10 +122,21 @@ struct fq_sched_data {
 	u8		rate_enable;
 	u8		fq_trees_log;
 	u8		horizon_drop;
+	u8		prio2band[(TC_PRIO_MAX + 1) >> 2];
 	u32		timer_slack; /* hrtimer slack in ns */
 
 /* Read/Write fields. */
 
+	unsigned int band_nr; /* band being serviced in fq_dequeue() */
+
+	struct fq_perband_flows band_flows[FQ_BANDS];
+
+	struct fq_flow	internal;	/* fastpath queue. */
+	struct rb_root	delayed;	/* for rate limited flows */
+	u64		time_next_delayed_flow;
+	unsigned long	unthrottle_latency_ns;
+
+	u32		band_pkt_count[FQ_BANDS];
 	u32		flows;
 	u32		inactive_flows; /* Flows with no packet to send. */
 	u32		throttled_flows;
@@ -139,7 +147,7 @@ struct fq_sched_data {
 
 /* Seldom used fields. */
 
-	u64		stat_internal_packets; /* aka highprio */
+	u64		stat_band_drops[FQ_BANDS];
 	u64		stat_ce_mark;
 	u64		stat_horizon_drops;
 	u64		stat_horizon_caps;
@@ -148,6 +156,12 @@ struct fq_sched_data {
 	u64		stat_allocation_errors;
 };
 
+/* return the i-th 2-bit value ("crumb") */
+static u8 fq_prio2band(const u8 *prio2band, unsigned int prio)
+{
+	return (prio2band[prio / 4] >> (2 * (prio & 0x3))) & 0x3;
+}
+
 /*
  * f->tail and f->age share the same location.
  * We can use the low order bit to differentiate if this location points
@@ -172,8 +186,19 @@ static bool fq_flow_is_throttled(const struct fq_flow *f)
 	return f->next == &throttled;
 }
 
-static void fq_flow_add_tail(struct fq_flow_head *head, struct fq_flow *flow)
+enum new_flow {
+	NEW_FLOW,
+	OLD_FLOW
+};
+
+static void fq_flow_add_tail(struct fq_sched_data *q, struct fq_flow *flow,
+			     enum new_flow list_sel)
 {
+	struct fq_perband_flows *pband = &q->band_flows[flow->band];
+	struct fq_flow_head *head = (list_sel == NEW_FLOW) ?
+					&pband->new_flows :
+					&pband->old_flows;
+
 	if (head->first)
 		head->last->next = flow;
 	else
@@ -186,7 +211,7 @@ static void fq_flow_unset_throttled(struct fq_sched_data *q, struct fq_flow *f)
 {
 	rb_erase(&f->rate_node, &q->delayed);
 	q->throttled_flows--;
-	fq_flow_add_tail(&q->old_flows, f);
+	fq_flow_add_tail(q, f, OLD_FLOW);
 }
 
 static void fq_flow_set_throttled(struct fq_sched_data *q, struct fq_flow *f)
@@ -326,11 +351,6 @@ static struct fq_flow *fq_classify(struct Qdisc *sch, struct sk_buff *skb,
 	struct rb_root *root;
 	struct fq_flow *f;
 
-	/* warning: no starvation prevention... */
-	if (unlikely((skb->priority & TC_PRIO_MAX) == TC_PRIO_CONTROL)) {
-		q->stat_internal_packets++; /* highprio packet */
-		return &q->internal;
-	}
 	/* SYNACK messages are attached to a TCP_NEW_SYN_RECV request socket
 	 * or a listener (SYNCOOKIE mode)
 	 * 1) request sockets are not full blown,
@@ -509,9 +529,13 @@ static int fq_enqueue(struct sk_buff *skb, struct Qdisc *sch,
 	struct fq_sched_data *q = qdisc_priv(sch);
 	struct fq_flow *f;
 	u64 now;
+	u8 band;
 
-	if (unlikely(sch->q.qlen >= sch->limit))
+	band = fq_prio2band(q->prio2band, skb->priority & TC_PRIO_MAX);
+	if (unlikely(q->band_pkt_count[band] >= sch->limit)) {
+		q->stat_band_drops[band]++;
 		return qdisc_drop(skb, sch, to_free);
+	}
 
 	now = ktime_get_ns();
 	if (!skb->tstamp) {
@@ -538,11 +562,14 @@ static int fq_enqueue(struct sk_buff *skb, struct Qdisc *sch,
 		}
 
 		if (fq_flow_is_detached(f)) {
-			fq_flow_add_tail(&q->new_flows, f);
+			fq_flow_add_tail(q, f, NEW_FLOW);
 			if (time_after(jiffies, f->age + q->flow_refill_delay))
 				f->credit = max_t(u32, f->credit, q->quantum);
 		}
 
+		f->band = band;
+		q->band_pkt_count[band]++;
+		fq_skb_cb(skb)->band = band;
 		if (f->qlen == 0)
 			q->inactive_flows--;
 	}
@@ -584,13 +611,26 @@ static void fq_check_throttled(struct fq_sched_data *q, u64 now)
 	}
 }
 
+static struct fq_flow_head *fq_pband_head_select(struct fq_perband_flows *pband)
+{
+	if (pband->credit <= 0)
+		return NULL;
+
+	if (pband->new_flows.first)
+		return &pband->new_flows;
+
+	return pband->old_flows.first ? &pband->old_flows : NULL;
+}
+
 static struct sk_buff *fq_dequeue(struct Qdisc *sch)
 {
 	struct fq_sched_data *q = qdisc_priv(sch);
+	struct fq_perband_flows *pband;
 	struct fq_flow_head *head;
 	struct sk_buff *skb;
 	struct fq_flow *f;
 	unsigned long rate;
+	int retry;
 	u32 plen;
 	u64 now;
 
@@ -606,24 +646,31 @@ static struct sk_buff *fq_dequeue(struct Qdisc *sch)
 
 	now = ktime_get_ns();
 	fq_check_throttled(q, now);
+	retry = 0;
+	pband = &q->band_flows[q->band_nr];
 begin:
-	head = &q->new_flows;
-	if (!head->first) {
-		head = &q->old_flows;
-		if (!head->first) {
-			if (q->time_next_delayed_flow != ~0ULL)
-				qdisc_watchdog_schedule_range_ns(&q->watchdog,
+	head = fq_pband_head_select(pband);
+	if (!head) {
+		while (++retry < FQ_BANDS) {
+			if (++q->band_nr == FQ_BANDS)
+				q->band_nr = 0;
+			pband = &q->band_flows[q->band_nr];
+			pband->credit = min(pband->credit + pband->quantum,
+					    pband->quantum);
+			goto begin;
+		}
+		if (q->time_next_delayed_flow != ~0ULL)
+			qdisc_watchdog_schedule_range_ns(&q->watchdog,
 							q->time_next_delayed_flow,
 							q->timer_slack);
-			return NULL;
-		}
+		return NULL;
 	}
 	f = head->first;
-
+	retry = 0;
 	if (f->credit <= 0) {
 		f->credit += q->quantum;
 		head->first = f->next;
-		fq_flow_add_tail(&q->old_flows, f);
+		fq_flow_add_tail(q, f, OLD_FLOW);
 		goto begin;
 	}
 
@@ -645,12 +692,13 @@ begin:
 		}
 		if (--f->qlen == 0)
 			q->inactive_flows++;
+		q->band_pkt_count[fq_skb_cb(skb)->band]--;
 		fq_dequeue_skb(sch, f, skb);
 	} else {
 		head->first = f->next;
 		/* force a pass through old_flows to prevent starvation */
-		if ((head == &q->new_flows) && q->old_flows.first) {
-			fq_flow_add_tail(&q->old_flows, f);
+		if (head == &pband->new_flows) {
+			fq_flow_add_tail(q, f, OLD_FLOW);
 		} else {
 			fq_flow_set_detached(f);
 		}
@@ -658,6 +706,7 @@ begin:
 	}
 	plen = qdisc_pkt_len(skb);
 	f->credit -= plen;
+	pband->credit -= plen;
 
 	if (!q->rate_enable)
 		goto out;
@@ -749,8 +798,10 @@ static void fq_reset(struct Qdisc *sch)
 			kmem_cache_free(fq_flow_cachep, f);
 		}
 	}
-	q->new_flows.first	= NULL;
-	q->old_flows.first	= NULL;
+	for (idx = 0; idx < FQ_BANDS; idx++) {
+		q->band_flows[idx].new_flows.first = NULL;
+		q->band_flows[idx].old_flows.first = NULL;
+	}
 	q->delayed		= RB_ROOT;
 	q->flows		= 0;
 	q->inactive_flows	= 0;
@@ -864,8 +915,54 @@ static const struct nla_policy fq_policy[TCA_FQ_MAX + 1] = {
 	[TCA_FQ_TIMER_SLACK]		= { .type = NLA_U32 },
 	[TCA_FQ_HORIZON]		= { .type = NLA_U32 },
 	[TCA_FQ_HORIZON_DROP]		= { .type = NLA_U8 },
+	[TCA_FQ_PRIOMAP]		= {
+			.type = NLA_BINARY,
+			.len = sizeof(struct tc_prio_qopt),
+		},
 };
 
+/* compress a u8 array with all elems <= 3 to an array of 2-bit fields */
+static void fq_prio2band_compress_crumb(const u8 *in, u8 *out)
+{
+	const int num_elems = TC_PRIO_MAX + 1;
+	int i;
+
+	memset(out, 0, num_elems / 4);
+	for (i = 0; i < num_elems; i++)
+		out[i / 4] |= in[i] << (2 * (i & 0x3));
+}
+
+static void fq_prio2band_decompress_crumb(const u8 *in, u8 *out)
+{
+	const int num_elems = TC_PRIO_MAX + 1;
+	int i;
+
+	for (i = 0; i < num_elems; i++)
+		out[i] = fq_prio2band(in, i);
+}
+
+static int fq_load_priomap(struct fq_sched_data *q,
+			   const struct nlattr *attr,
+			   struct netlink_ext_ack *extack)
+{
+	const struct tc_prio_qopt *map = nla_data(attr);
+	int i;
+
+	if (map->bands != FQ_BANDS) {
+		NL_SET_ERR_MSG_MOD(extack, "FQ only supports 3 bands");
+		return -EINVAL;
+	}
+	for (i = 0; i < TC_PRIO_MAX + 1; i++) {
+		if (map->priomap[i] >= FQ_BANDS) {
+			NL_SET_ERR_MSG_FMT_MOD(extack, "FQ priomap field %d maps to a too high band %d",
+					       i, map->priomap[i]);
+			return -EINVAL;
+		}
+	}
+	fq_prio2band_compress_crumb(map->priomap, q->prio2band);
+	return 0;
+}
+
 static int fq_change(struct Qdisc *sch, struct nlattr *opt,
 		     struct netlink_ext_ack *extack)
 {
@@ -940,6 +1037,9 @@ static int fq_change(struct Qdisc *sch, struct nlattr *opt,
 		q->flow_refill_delay = usecs_to_jiffies(usecs_delay);
 	}
 
+	if (!err && tb[TCA_FQ_PRIOMAP])
+		err = fq_load_priomap(q, tb[TCA_FQ_PRIOMAP], extack);
+
 	if (tb[TCA_FQ_ORPHAN_MASK])
 		q->orphan_mask = nla_get_u32(tb[TCA_FQ_ORPHAN_MASK]);
 
@@ -991,7 +1091,7 @@ static int fq_init(struct Qdisc *sch, struct nlattr *opt,
 		   struct netlink_ext_ack *extack)
 {
 	struct fq_sched_data *q = qdisc_priv(sch);
-	int err;
+	int i, err;
 
 	sch->limit		= 10000;
 	q->flow_plimit		= 100;
@@ -1001,8 +1101,13 @@ static int fq_init(struct Qdisc *sch, struct nlattr *opt,
 	q->flow_max_rate	= ~0UL;
 	q->time_next_delayed_flow = ~0ULL;
 	q->rate_enable		= 1;
-	q->new_flows.first	= NULL;
-	q->old_flows.first	= NULL;
+	for (i = 0; i < FQ_BANDS; i++) {
+		q->band_flows[i].new_flows.first = NULL;
+		q->band_flows[i].old_flows.first = NULL;
+	}
+	q->band_flows[0].quantum = 9 << 16;
+	q->band_flows[1].quantum = 3 << 16;
+	q->band_flows[2].quantum = 1 << 16;
 	q->delayed		= RB_ROOT;
 	q->fq_root		= NULL;
 	q->fq_trees_log		= ilog2(1024);
@@ -1017,6 +1122,7 @@ static int fq_init(struct Qdisc *sch, struct nlattr *opt,
 	/* Default ce_threshold of 4294 seconds */
 	q->ce_threshold		= (u64)NSEC_PER_USEC * ~0U;
 
+	fq_prio2band_compress_crumb(sch_default_prio2band, q->prio2band);
 	qdisc_watchdog_init_clockid(&q->watchdog, sch, CLOCK_MONOTONIC);
 
 	if (opt)
@@ -1031,6 +1137,9 @@ static int fq_dump(struct Qdisc *sch, struct sk_buff *skb)
 {
 	struct fq_sched_data *q = qdisc_priv(sch);
 	u64 ce_threshold = q->ce_threshold;
+	struct tc_prio_qopt prio = {
+		.bands = FQ_BANDS,
+	};
 	u64 horizon = q->horizon;
 	struct nlattr *opts;
 
@@ -1062,6 +1171,10 @@ static int fq_dump(struct Qdisc *sch, struct sk_buff *skb)
 	    nla_put_u8(skb, TCA_FQ_HORIZON_DROP, q->horizon_drop))
 		goto nla_put_failure;
 
+	fq_prio2band_decompress_crumb(q->prio2band, prio.priomap);
+	if (nla_put(skb, TCA_FQ_PRIOMAP, sizeof(prio), &prio))
+		goto nla_put_failure;
+
 	return nla_nest_end(skb, opts);
 
 nla_put_failure:
@@ -1072,11 +1185,14 @@ static int fq_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
 {
 	struct fq_sched_data *q = qdisc_priv(sch);
 	struct tc_fq_qd_stats st;
+	int i;
+
+	st.pad = 0;
 
 	sch_tree_lock(sch);
 
 	st.gc_flows		  = q->stat_gc_flows;
-	st.highprio_packets	  = q->stat_internal_packets;
+	st.highprio_packets	  = 0;
 	st.fastpath_packets	  = q->internal.stat_fastpath_packets;
 	st.tcp_retrans		  = 0;
 	st.throttled		  = q->stat_throttled;
@@ -1093,6 +1209,10 @@ static int fq_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
 	st.ce_mark		  = q->stat_ce_mark;
 	st.horizon_drops	  = q->stat_horizon_drops;
 	st.horizon_caps		  = q->stat_horizon_caps;
+	for (i = 0; i < FQ_BANDS; i++) {
+		st.band_drops[i]  = q->stat_band_drops[i];
+		st.band_pkt_count[i] = q->band_pkt_count[i];
+	}
 	sch_tree_unlock(sch);
 
 	return gnet_stats_copy_app(d, &st, sizeof(st));
@@ -1120,7 +1240,7 @@ static int __init fq_module_init(void)
 
 	fq_flow_cachep = kmem_cache_create("fq_flow_cache",
 					   sizeof(struct fq_flow),
-					   0, 0, NULL);
+					   0, SLAB_HWCACHE_ALIGN, NULL);
 	if (!fq_flow_cachep)
 		return -ENOMEM;