--- a/net80211/ieee80211_node.c
+++ b/net80211/ieee80211_node.c
@@ -123,6 +123,9 @@ static void ieee80211_node_table_cleanup
 static void ieee80211_node_table_reset(struct ieee80211_node_table *,
 	struct ieee80211vap *);
 
+static struct ieee80211_node *
+lookup_rxnode(struct ieee80211com *ic, struct ieee80211vap *vap, const u_int8_t *addr);
+
 MALLOC_DEFINE(M_80211_NODE, "80211node", "802.11 node state");
 
 void
@@ -697,7 +700,7 @@ ieee80211_sta_join(struct ieee80211vap *
 	struct ieee80211com *ic = vap->iv_ic;
 	struct ieee80211_node *ni;
 
-	ni = ieee80211_find_node(&ic->ic_sta, se->se_macaddr);
+	ni = lookup_rxnode(ic, vap, se->se_macaddr);
 	if (ni == NULL) {
 		ni = ieee80211_alloc_node_table(vap, se->se_macaddr);
 		IEEE80211_DPRINTF(vap, IEEE80211_MSG_ASSOC,
@@ -1394,6 +1397,53 @@ ieee80211_add_neighbor(struct ieee80211v
 	return ni;
 }
 
+struct ieee80211vap *
+ieee80211_find_rxvap(struct ieee80211com *ic, const u_int8_t *mac)
+{
+	struct ieee80211vap *vap;
+
+	TAILQ_FOREACH(vap, &ic->ic_vaps, iv_next) {
+		if (IEEE80211_ADDR_EQ(vap->iv_myaddr, mac))
+			return vap;
+	}
+	return NULL;
+}
+EXPORT_SYMBOL(ieee80211_find_rxvap);
+
+static struct ieee80211_node *
+lookup_rxnode(struct ieee80211com *ic, struct ieee80211vap *vap,
+	const u_int8_t *addr)
+{
+	struct ieee80211_node_table *nt;
+	struct ieee80211_node *ni = NULL;
+	int use_bss = 0;
+	int hash;
+
+	nt = &ic->ic_sta;
+	IEEE80211_NODE_TABLE_LOCK_IRQ(nt);
+	hash = IEEE80211_NODE_HASH(addr);
+	LIST_FOREACH(ni, &nt->nt_hash[hash], ni_hash) {
+		if (IEEE80211_ADDR_EQ(ni->ni_macaddr, addr)) {
+			/* allow multiple nodes on different vaps */
+			if (vap && (ni->ni_vap != vap))
+				continue;
+
+			ieee80211_ref_node(ni);
+			goto out;
+		}
+	}
+
+	/* no match found */
+	ni = NULL;
+
+out:
+	IEEE80211_NODE_TABLE_UNLOCK_IRQ(nt);
+	return ni;
+#undef IS_PSPOLL
+#undef IS_CTL
+}
+
+
 /*
  * Return the node for the sender of a frame; if the sender is unknown return 
  * NULL. The caller is expected to deal with this. (The frame is sent to all 
@@ -1403,10 +1453,10 @@ ieee80211_add_neighbor(struct ieee80211v
  */
 struct ieee80211_node *
 #ifdef IEEE80211_DEBUG_REFCNT
-ieee80211_find_rxnode_debug(struct ieee80211com *ic,
+ieee80211_find_rxnode_debug(struct ieee80211com *ic, struct ieee80211vap *vap,
 	const struct ieee80211_frame_min *wh, const char *func, int line)
 #else
-ieee80211_find_rxnode(struct ieee80211com *ic,
+ieee80211_find_rxnode(struct ieee80211com *ic, struct ieee80211vap *vap,
 	const struct ieee80211_frame_min *wh)
 #endif
 {
@@ -1414,9 +1464,8 @@ ieee80211_find_rxnode(struct ieee80211co
 	((wh->i_fc[0] & IEEE80211_FC0_TYPE_MASK) == IEEE80211_FC0_TYPE_CTL)
 #define	IS_PSPOLL(wh) \
 	((wh->i_fc[0] & IEEE80211_FC0_SUBTYPE_MASK) == IEEE80211_FC0_SUBTYPE_PS_POLL)
-	struct ieee80211_node_table *nt;
-	struct ieee80211_node *ni;
-	struct ieee80211vap *vap, *avp;
+	struct ieee80211_node *ni = NULL;
+	struct ieee80211vap *avp;
 	const u_int8_t *addr;
 
 	if (IS_CTL(wh) && !IS_PSPOLL(wh) /*&& !IS_RTS(ah)*/)
@@ -1429,32 +1478,25 @@ ieee80211_find_rxnode(struct ieee80211co
 
 	/* XXX check ic_bss first in station mode */
 	/* XXX 4-address frames? */
-	nt = &ic->ic_sta;
-	IEEE80211_NODE_TABLE_LOCK_IRQ(nt);
 	if ((wh->i_fc[1] & IEEE80211_FC1_DIR_MASK) == IEEE80211_FC1_DIR_DSTODS) {
-		TAILQ_FOREACH(vap, &ic->ic_vaps, iv_next) {
+		if (vap) { /* assume unicast if vap is set, mcast not supported for wds */
 			TAILQ_FOREACH(avp, &vap->iv_wdslinks, iv_wdsnext) {
-				if (!IEEE80211_ADDR_EQ(addr, avp->wds_mac))
+				if (!IEEE80211_ADDR_EQ(addr, avp->wds_mac) ||
+					!IEEE80211_ADDR_EQ(wh->i_addr1, avp->iv_myaddr))
 					continue;
 
 				if (avp->iv_wdsnode)
-					return ieee80211_ref_node(avp->iv_wdsnode);
-				else
-					return NULL;
+					ni = ieee80211_ref_node(avp->iv_wdsnode);
+				return ni;
 			}
+			if (!(vap->iv_flags_ext & IEEE80211_FEXT_WDS))
+				return NULL;
+		} else {
+			return NULL;
 		}
 	}
 
-#ifdef IEEE80211_DEBUG_REFCNT
-	ni = ieee80211_find_node_locked_debug(nt, addr, func, line);
-#else
-	ni = ieee80211_find_node_locked(nt, addr);
-#endif
-	IEEE80211_NODE_TABLE_UNLOCK_IRQ(nt);
-
-	return ni;
-#undef IS_PSPOLL
-#undef IS_CTL
+	return lookup_rxnode(ic, vap, addr);
 }
 #ifdef IEEE80211_DEBUG_REFCNT
 EXPORT_SYMBOL(ieee80211_find_rxnode_debug);
@@ -1479,15 +1521,14 @@ ieee80211_find_txnode(struct ieee80211va
 	struct ieee80211com *ic = vap->iv_ic;
 	struct ieee80211_node_table *nt;
 	struct ieee80211_node *ni = NULL;
+	int hash;
 
-	IEEE80211_LOCK_IRQ(ic);
 	if (vap->iv_opmode == IEEE80211_M_WDS) {
 		if (vap->iv_wdsnode && (vap->iv_state == IEEE80211_S_RUN))
 			return ieee80211_ref_node(vap->iv_wdsnode);
 		else
 			return NULL;
 	}
-	IEEE80211_UNLOCK_IRQ(ic);
 
 	/*
 	 * The destination address should be in the node table
@@ -1505,11 +1546,22 @@ ieee80211_find_txnode(struct ieee80211va
 	/* XXX: Can't hold lock across dup_bss due to recursive locking. */
 	nt = &vap->iv_ic->ic_sta;
 	IEEE80211_NODE_TABLE_LOCK_IRQ(nt);
+	hash = IEEE80211_NODE_HASH(mac);
+	LIST_FOREACH(ni, &nt->nt_hash[hash], ni_hash) {
+		if (ni->ni_vap != vap)
+			continue;
+
+		if (IEEE80211_ADDR_EQ(ni->ni_macaddr, mac)) {
 #ifdef IEEE80211_DEBUG_REFCNT
-	ni = ieee80211_find_node_locked_debug(nt, mac, func, line);
+			ieee80211_ref_node_debug(ni, func, line);
 #else
-	ni = ieee80211_find_node_locked(nt, mac);
+			ieee80211_ref_node(ni);
 #endif
+			goto found;
+		}
+	}
+	ni = NULL;
+found:
 	IEEE80211_NODE_TABLE_UNLOCK_IRQ(nt);
 
 	if (ni == NULL) {
@@ -1964,13 +2016,32 @@ remove_worse_nodes(void *arg, struct iee
 		}
 }
 
+static void
+remove_duplicate_nodes(void *arg, struct ieee80211_node *ni)
+{
+	struct ieee80211_node *rni = arg;
+
+	if (ni == rni)
+		return;
+
+	if (ni->ni_vap == rni->ni_vap)
+		return;
+
+	if (!IEEE80211_ADDR_EQ(rni->ni_macaddr, ni->ni_macaddr))
+		return;
+
+	ieee80211_node_leave(ni);
+}
+
 void
 ieee80211_node_join(struct ieee80211_node *ni, int resp)
 {
 	struct ieee80211com *ic = ni->ni_ic;
 	struct ieee80211vap *vap = ni->ni_vap;
+	struct ieee80211_node *tni;
 	int newassoc;
 
+	ieee80211_iterate_nodes(&ic->ic_sta, remove_duplicate_nodes, ni);
 	if (ni->ni_associd == 0) {
 		u_int16_t aid;
 
--- a/net80211/ieee80211_input.c
+++ b/net80211/ieee80211_input.c
@@ -216,16 +216,14 @@ ieee80211_input(struct ieee80211vap * va
 
 	type = -1;			/* undefined */
 
-	if (!vap)
-		goto out;
+	if (!vap || !vap->iv_bss || !vap->iv_dev || !vap->iv_ic)
+		goto discard;
 
 	ic = vap->iv_ic;
-	if (!ic)
-		goto out;
-
 	dev = vap->iv_dev;
-	if (!dev)
-		goto out;
+
+	if ((vap->iv_dev->flags & (IFF_UP|IFF_RUNNING)) != (IFF_UP|IFF_RUNNING))
+		goto discard;
 
 	/* initialize ni as in the previous API */
 	if (ni_or_null == NULL) {
@@ -233,9 +231,10 @@ ieee80211_input(struct ieee80211vap * va
                 * guarantee its existence during the following call, hence
                 * briefly grab our own reference. */
 		ni = ieee80211_ref_node(vap->iv_bss);
+		KASSERT(ni != NULL, ("null node"));
+	} else {
+		ni->ni_inact = ni->ni_inact_reload;
 	}
-	KASSERT(ni != NULL, ("null node"));
-	ni->ni_inact = ni->ni_inact_reload;
 
 	KASSERT(skb->len >= sizeof(struct ieee80211_frame_min),
 		("frame length too short: %u", skb->len));
@@ -844,10 +843,11 @@ ieee80211_input(struct ieee80211vap * va
 err:
 	vap->iv_devstats.rx_errors++;
 out:
-	if (skb != NULL)
-		ieee80211_dev_kfree_skb(&skb);
 	if (ni_or_null == NULL)
 		ieee80211_unref_node(&ni);
+discard:
+	if (skb != NULL)
+		ieee80211_dev_kfree_skb(&skb);
 	return type;
 #undef HAS_SEQ
 }
@@ -929,16 +929,23 @@ int
 ieee80211_input_all(struct ieee80211com *ic,
 	struct sk_buff *skb, int rssi, u_int64_t rtsf)
 {
+	struct ieee80211_frame_min *wh = (struct ieee80211_frame_min *) skb->data;
 	struct ieee80211vap *vap;
 	int type = -1;
 
 	/* XXX locking */
 	TAILQ_FOREACH(vap, &ic->ic_vaps, iv_next) {
+		struct ieee80211_node *ni = NULL;
 		struct sk_buff *skb1;
 
 		if ((vap->iv_dev->flags & (IFF_UP|IFF_RUNNING)) != (IFF_UP|IFF_RUNNING))
 			continue;
 
+		if ((vap->iv_opmode == IEEE80211_M_HOSTAP) &&
+			!IEEE80211_IS_MULTICAST(wh->i_addr1))
+			continue;
+
+		ni = ieee80211_find_rxnode(ic, vap, wh);
 		if (TAILQ_NEXT(vap, iv_next) != NULL) {
 			skb1 = skb_copy(skb, GFP_ATOMIC);
 			if (skb1 == NULL) {
@@ -950,8 +957,10 @@ ieee80211_input_all(struct ieee80211com 
 			skb1 = skb;
 			skb = NULL;
 		}
-		type = ieee80211_input(vap, NULL, skb1, rssi, rtsf);
+		type = ieee80211_input(vap, ni, skb1, rssi, rtsf);
 	}
+
+out:
 	if (skb != NULL)		/* no vaps, reclaim skb */
 		ieee80211_dev_kfree_skb(&skb);
 	return type;
@@ -1146,11 +1155,9 @@ ieee80211_deliver_data(struct ieee80211_
 			 * sending it will not work; just let it be
 			 * delivered normally.
 			 */
-			struct ieee80211_node *ni1 = ieee80211_find_node(
-				&vap->iv_ic->ic_sta, eh->ether_dhost);
+			struct ieee80211_node *ni1 = ieee80211_find_txnode(vap, eh->ether_dhost);
 			if (ni1 != NULL) {
-				if (ni1->ni_vap == vap &&
-				    ieee80211_node_is_authorized(ni1) &&
+				if (ieee80211_node_is_authorized(ni1) &&
 					!ni1->ni_subif &&
 				    ni1 != vap->iv_bss) {
 
--- a/ath/if_ath.c
+++ b/ath/if_ath.c
@@ -6589,9 +6589,8 @@ ath_recv_mgmt(struct ieee80211vap * vap,
 
 	sc->sc_recv_mgmt(vap, ni_or_null, skb, subtype, rssi, rtsf);
 
-
 	/* Lookup the new node if any (this grabs a reference to it) */
-	ni = ieee80211_find_rxnode(vap->iv_ic,
+	ni = ieee80211_find_rxnode(vap->iv_ic, vap,
 	         (const struct ieee80211_frame_min *)skb->data);
 	if (ni == NULL) {
 		DPRINTF(sc, ATH_DEBUG_BEACON, "Dropping; node unknown.\n");
@@ -6746,7 +6745,9 @@ ath_rx_poll(struct net_device *dev, int 
 	struct ath_desc *ds;
 	struct ath_rx_status *rs;
 	struct sk_buff *skb = NULL;
+	struct ieee80211vap *vap;
 	struct ieee80211_node *ni;
+	const struct ieee80211_frame_min *wh;
 	unsigned int len;
 	int type;
 	u_int phyerr;
@@ -6901,12 +6902,15 @@ rx_accept:
 		skb_trim(skb, skb->len - IEEE80211_CRC_LEN);
 
 		if (mic_fail) {
+			wh = (const struct ieee80211_frame_min *) skb->data;
+
 			/* Ignore control frames which are reported with mic error */
-		    if ((((struct ieee80211_frame *)skb->data)->i_fc[0] &
+		    if ((wh->i_fc[0] &
 					IEEE80211_FC0_TYPE_MASK) == IEEE80211_FC0_TYPE_CTL)
 				goto drop_micfail;
 
-			ni = ieee80211_find_rxnode(ic, (const struct ieee80211_frame_min *) skb->data);
+			vap = ieee80211_find_rxvap(ic, wh->i_addr1);
+			ni = ieee80211_find_rxnode(ic, vap, wh);
 
 			if (ni && ni->ni_table) {
 				ieee80211_check_mic(ni, skb);
@@ -6968,11 +6972,24 @@ drop_micfail:
 		 * for its use.  If the sender is unknown spam the
 		 * frame; it'll be dropped where it's not wanted.
 		 */
-		if (rs->rs_keyix != HAL_RXKEYIX_INVALID &&
+		wh = (const struct ieee80211_frame_min *) skb->data;
+		if ((rs->rs_keyix != HAL_RXKEYIX_INVALID) &&
 		    (ni = sc->sc_keyixmap[rs->rs_keyix]) != NULL) {
 			/* Fast path: node is present in the key map;
 			 * grab a reference for processing the frame. */
-			ni = ieee80211_ref_node(ni);
+			ieee80211_ref_node(ni);
+			if ((ATH_GET_VAP_ID(wh->i_addr1) !=
+			     ATH_GET_VAP_ID(ni->ni_vap->iv_myaddr)) ||
+				((wh->i_fc[1] & IEEE80211_FC1_DIR_MASK) ==
+				 IEEE80211_FC1_DIR_DSTODS)) {
+				/* key cache node lookup is fast, but it can
+				 * lead to problems in multi-bss (foreign vap
+				 * node reference) or wds (wdsap node ref instead
+				 * of base ap node ref).
+				 * use slowpath lookup in both cases
+				 */
+				goto lookup_slowpath;
+			}
 			ATH_RSSI_LPF(ATH_NODE(ni)->an_avgrssi, rs->rs_rssi);
 			type = ieee80211_input(ni->ni_vap, ni, skb, rs->rs_rssi, bf->bf_tsf);
 			ieee80211_unref_node(&ni);
@@ -6981,24 +6998,35 @@ drop_micfail:
 			 * No key index or no entry, do a lookup and
 			 * add the node to the mapping table if possible.
 			 */
-			ni = ieee80211_find_rxnode(ic,
-				(const struct ieee80211_frame_min *)skb->data);
+
+lookup_slowpath:
+			vap = ieee80211_find_rxvap(ic, wh->i_addr1);
+			if (vap)
+				ni = ieee80211_find_rxnode(ic, vap, wh);
+			else
+				ni = NULL;
+
 			if (ni != NULL) {
 				ieee80211_keyix_t keyix;
 
 				ATH_RSSI_LPF(ATH_NODE(ni)->an_avgrssi, rs->rs_rssi);
-				type = ieee80211_input(ni->ni_vap, ni, skb, rs->rs_rssi, bf->bf_tsf);
+				type = ieee80211_input(vap, ni, skb, rs->rs_rssi, bf->bf_tsf);
 				/*
 				 * If the station has a key cache slot assigned
 				 * update the key->node mapping table.
 				 */
 				keyix = ni->ni_ucastkey.wk_keyix;
 				if (keyix != IEEE80211_KEYIX_NONE &&
-				    sc->sc_keyixmap[keyix] == NULL)
+				    sc->sc_keyixmap[keyix] == NULL) {
 					sc->sc_keyixmap[keyix] = ieee80211_ref_node(ni);
+				}
 				ieee80211_unref_node(&ni);
-			} else
-				type = ieee80211_input_all(ic, skb, rs->rs_rssi, bf->bf_tsf);
+			} else {
+				if (vap)
+					type = ieee80211_input(vap, NULL, skb, rs->rs_rssi, bf->bf_tsf);
+				else
+					type = ieee80211_input_all(ic, skb, rs->rs_rssi, bf->bf_tsf);
+			}
 		}
 
 		if (sc->sc_diversity) {
--- a/net80211/ieee80211_node.h
+++ b/net80211/ieee80211_node.h
@@ -286,15 +286,18 @@ struct ieee80211_node *ieee80211_find_no
 	const u_int8_t *);
 #endif /* #ifdef IEEE80211_DEBUG_REFCNT */
 
+struct ieee80211vap *
+ieee80211_find_rxvap(struct ieee80211com *ic, const u_int8_t *mac);
+
 /* Returns a ieee80211_node* with refcount incremented, if found */
 #ifdef IEEE80211_DEBUG_REFCNT
-#define	ieee80211_find_rxnode(_nt, _wh) \
-	ieee80211_find_rxnode_debug(_nt, _wh, __func__, __LINE__)
+#define	ieee80211_find_rxnode(_nt, _vap, _wh) \
+	ieee80211_find_rxnode_debug(_nt, _vap, _wh, __func__, __LINE__)
 struct ieee80211_node *ieee80211_find_rxnode_debug(struct ieee80211com *,
-	const struct ieee80211_frame_min *, const char *, int);
+	struct ieee80211vap *, const struct ieee80211_frame_min *, const char *, int);
 #else
 struct ieee80211_node *ieee80211_find_rxnode(struct ieee80211com *,
-	const struct ieee80211_frame_min *);
+	struct ieee80211vap *, const struct ieee80211_frame_min *);
 #endif /* #ifdef IEEE80211_DEBUG_REFCNT */
 
 /* Returns a ieee80211_node* with refcount incremented, if found */