diff --git a/net/mac80211/key.c b/net/mac80211/key.c
index 6eb4888..99e9f6a 100644
--- a/net/mac80211/key.c
+++ b/net/mac80211/key.c
@@ -397,7 +397,8 @@
 	return key;
 }
 
-static void __ieee80211_key_destroy(struct ieee80211_key *key)
+static void __ieee80211_key_destroy(struct ieee80211_key *key,
+				    bool delay_tailroom)
 {
 	if (!key)
 		return;
@@ -416,8 +417,18 @@
 	if (key->conf.cipher == WLAN_CIPHER_SUITE_AES_CMAC)
 		ieee80211_aes_cmac_key_free(key->u.aes_cmac.tfm);
 	if (key->local) {
+		struct ieee80211_sub_if_data *sdata = key->sdata;
+
 		ieee80211_debugfs_key_remove(key);
-		key->sdata->crypto_tx_tailroom_needed_cnt--;
+
+		if (delay_tailroom) {
+			/* see ieee80211_delayed_tailroom_dec */
+			sdata->crypto_tx_tailroom_pending_dec++;
+			schedule_delayed_work(&sdata->dec_tailroom_needed_wk,
+					      HZ/2);
+		} else {
+			sdata->crypto_tx_tailroom_needed_cnt--;
+		}
 	}
 
 	kfree(key);
@@ -452,7 +463,7 @@
 	increment_tailroom_need_count(sdata);
 
 	__ieee80211_key_replace(sdata, sta, pairwise, old_key, key);
-	__ieee80211_key_destroy(old_key);
+	__ieee80211_key_destroy(old_key, true);
 
 	ieee80211_debugfs_key_add(key);
 
@@ -463,7 +474,7 @@
 	return ret;
 }
 
-void __ieee80211_key_free(struct ieee80211_key *key)
+void __ieee80211_key_free(struct ieee80211_key *key, bool delay_tailroom)
 {
 	if (!key)
 		return;
@@ -475,14 +486,14 @@
 		__ieee80211_key_replace(key->sdata, key->sta,
 				key->conf.flags & IEEE80211_KEY_FLAG_PAIRWISE,
 				key, NULL);
-	__ieee80211_key_destroy(key);
+	__ieee80211_key_destroy(key, delay_tailroom);
 }
 
 void ieee80211_key_free(struct ieee80211_local *local,
 			struct ieee80211_key *key)
 {
 	mutex_lock(&local->key_mtx);
-	__ieee80211_key_free(key);
+	__ieee80211_key_free(key, true);
 	mutex_unlock(&local->key_mtx);
 }
 
@@ -544,18 +555,56 @@
 {
 	struct ieee80211_key *key, *tmp;
 
+	cancel_delayed_work_sync(&sdata->dec_tailroom_needed_wk);
+
 	mutex_lock(&sdata->local->key_mtx);
 
+	sdata->crypto_tx_tailroom_needed_cnt -=
+		sdata->crypto_tx_tailroom_pending_dec;
+	sdata->crypto_tx_tailroom_pending_dec = 0;
+
 	ieee80211_debugfs_key_remove_mgmt_default(sdata);
 
 	list_for_each_entry_safe(key, tmp, &sdata->key_list, list)
-		__ieee80211_key_free(key);
+		__ieee80211_key_free(key, false);
 
 	ieee80211_debugfs_key_update_default(sdata);
 
+	WARN_ON_ONCE(sdata->crypto_tx_tailroom_needed_cnt ||
+		     sdata->crypto_tx_tailroom_pending_dec);
+
 	mutex_unlock(&sdata->local->key_mtx);
 }
 
+void ieee80211_delayed_tailroom_dec(struct work_struct *wk)
+{
+	struct ieee80211_sub_if_data *sdata;
+
+	sdata = container_of(wk, struct ieee80211_sub_if_data,
+			     dec_tailroom_needed_wk.work);
+
+	/*
+	 * The reason for the delayed tailroom needed decrementing is to
+	 * make roaming faster: during roaming, all keys are first deleted
+	 * and then new keys are installed. The first new key causes the
+	 * crypto_tx_tailroom_needed_cnt to go from 0 to 1, which invokes
+	 * the cost of synchronize_net() (which can be slow). Avoid this
+	 * by deferring the crypto_tx_tailroom_needed_cnt decrementing on
+	 * key removal for a while, so if we roam the value is larger than
+	 * zero and no 0->1 transition happens.
+	 *
+	 * The cost is that if the AP switching was from an AP with keys
+	 * to one without, we still allocate tailroom while it would no
+	 * longer be needed. However, in the typical (fast) roaming case
+	 * within an ESS this usually won't happen.
+	 */
+
+	mutex_lock(&sdata->local->key_mtx);
+	sdata->crypto_tx_tailroom_needed_cnt -=
+		sdata->crypto_tx_tailroom_pending_dec;
+	sdata->crypto_tx_tailroom_pending_dec = 0;
+	mutex_unlock(&sdata->local->key_mtx);
+}
 
 void ieee80211_gtk_rekey_notify(struct ieee80211_vif *vif, const u8 *bssid,
 				const u8 *replay_ctr, gfp_t gfp)