From: Felix Fietkau <nbd@nbd.name>
Date: Sat, 16 Mar 2019 17:57:38 +0100
Subject: [PATCH] mac80211: calculate hash for fq without holding fq->lock
 in itxq enqueue

Reduces lock contention on enqueue/dequeue of iTXQ packets

Signed-off-by: Felix Fietkau <nbd@nbd.name>
---

--- a/include/net/fq_impl.h
+++ b/include/net/fq_impl.h
@@ -107,29 +107,31 @@ begin:
 	return skb;
 }
 
-static struct fq_flow *fq_flow_classify(struct fq *fq,
-					struct fq_tin *tin,
-					struct sk_buff *skb,
-					fq_flow_get_default_t get_default_func)
+static u32 fq_flow_idx(struct fq *fq, struct sk_buff *skb)
 {
-	struct fq_flow *flow;
-	u32 hash;
-	u32 idx;
-
-	lockdep_assert_held(&fq->lock);
-
 #if LINUX_VERSION_IS_GEQ(5,3,10) || \
     LINUX_VERSION_IN_RANGE(4,19,83, 4,20,0) || \
     LINUX_VERSION_IN_RANGE(4,14,153, 4,15,0) || \
     LINUX_VERSION_IN_RANGE(4,9,200, 4,10,0) || \
     LINUX_VERSION_IN_RANGE(4,4,200, 4,5,0)
-	hash = skb_get_hash_perturb(skb, &fq->perturbation);
+	u32 hash = skb_get_hash_perturb(skb, &fq->perturbation);
 #else
-	hash = skb_get_hash_perturb(skb, fq->perturbation);
+	u32 hash = skb_get_hash_perturb(skb, fq->perturbation);
 #endif
-	idx = reciprocal_scale(hash, fq->flows_cnt);
-	flow = &fq->flows[idx];
 
+	return reciprocal_scale(hash, fq->flows_cnt);
+}
+
+static struct fq_flow *fq_flow_classify(struct fq *fq,
+					struct fq_tin *tin, u32 idx,
+					struct sk_buff *skb,
+					fq_flow_get_default_t get_default_func)
+{
+	struct fq_flow *flow;
+
+	lockdep_assert_held(&fq->lock);
+
+	flow = &fq->flows[idx];
 	if (flow->tin && flow->tin != tin) {
 		flow = get_default_func(fq, tin, idx, skb);
 		tin->collisions++;
@@ -161,7 +163,7 @@ static void fq_recalc_backlog(struct fq
 }
 
 static void fq_tin_enqueue(struct fq *fq,
-			   struct fq_tin *tin,
+			   struct fq_tin *tin, u32 idx,
 			   struct sk_buff *skb,
 			   fq_skb_free_t free_func,
 			   fq_flow_get_default_t get_default_func)
@@ -171,7 +173,7 @@ static void fq_tin_enqueue(struct fq *fq
 
 	lockdep_assert_held(&fq->lock);
 
-	flow = fq_flow_classify(fq, tin, skb, get_default_func);
+	flow = fq_flow_classify(fq, tin, idx, skb, get_default_func);
 
 	flow->tin = tin;
 	flow->backlog += skb->len;
--- a/net/mac80211/tx.c
+++ b/net/mac80211/tx.c
@@ -1390,11 +1390,15 @@ static void ieee80211_txq_enqueue(struct
 {
 	struct fq *fq = &local->fq;
 	struct fq_tin *tin = &txqi->tin;
+	u32 flow_idx = fq_flow_idx(fq, skb);
 
 	ieee80211_set_skb_enqueue_time(skb);
-	fq_tin_enqueue(fq, tin, skb,
+
+	spin_lock_bh(&fq->lock);
+	fq_tin_enqueue(fq, tin, flow_idx, skb,
 		       fq_skb_free_func,
 		       fq_flow_get_default_func);
+	spin_unlock_bh(&fq->lock);
 }
 
 static bool fq_vlan_filter_func(struct fq *fq, struct fq_tin *tin,
@@ -1564,7 +1568,6 @@ static bool ieee80211_queue_skb(struct i
 				struct sta_info *sta,
 				struct sk_buff *skb)
 {
-	struct fq *fq = &local->fq;
 	struct ieee80211_vif *vif;
 	struct txq_info *txqi;
 
@@ -1582,9 +1585,7 @@ static bool ieee80211_queue_skb(struct i
 	if (!txqi)
 		return false;
 
-	spin_lock_bh(&fq->lock);
 	ieee80211_txq_enqueue(local, txqi, skb);
-	spin_unlock_bh(&fq->lock);
 
 	schedule_and_wake_txq(local, txqi);
 
@@ -3198,6 +3199,7 @@ static bool ieee80211_amsdu_aggregate(st
 	u8 max_subframes = sta->sta.max_amsdu_subframes;
 	int max_frags = local->hw.max_tx_fragments;
 	int max_amsdu_len = sta->sta.max_amsdu_len;
+	u32 flow_idx;
 	int orig_truesize;
 	__be16 len;
 	void *data;
@@ -3220,6 +3222,8 @@ static bool ieee80211_amsdu_aggregate(st
 		max_amsdu_len = min_t(int, max_amsdu_len,
 				      sta->sta.max_rc_amsdu_len);
 
+	flow_idx = fq_flow_idx(fq, skb);
+
 	spin_lock_bh(&fq->lock);
 
 	/* TODO: Ideally aggregation should be done on dequeue to remain
@@ -3227,7 +3231,8 @@ static bool ieee80211_amsdu_aggregate(st
 	 */
 
 	tin = &txqi->tin;
-	flow = fq_flow_classify(fq, tin, skb, fq_flow_get_default_func);
+	flow = fq_flow_classify(fq, tin, flow_idx, skb,
+				fq_flow_get_default_func);
 	head = skb_peek_tail(&flow->queue);
 	if (!head)
 		goto out;