Re: [PATCH 1/2] mac80211: refactor sta_info_insert_rcu to 3 main stages

On Wed, 2011-07-27 at 00:29 +0300, Guy Eilam wrote:

-int sta_info_insert_rcu(struct sta_info *sta) __acquires(RCU)
+static int sta_info_insert_check(struct sta_info *sta)
 {
-	struct ieee80211_local *local = sta->local;
 	struct ieee80211_sub_if_data *sdata = sta->sdata;
-	unsigned long flags;
-	int err = 0;
 
 	/*
 	 * Can't be a WARN_ON because it can be triggered through a race:
 	 * something inserts a STA (on one CPU) without holding the RTNL
 	 * and another CPU turns off the net device.
 	 */
-	if (unlikely(!ieee80211_sdata_running(sdata))) {
-		err = -ENETDOWN;
-		rcu_read_lock();
-		goto out_free;
-	}
+	if (unlikely(!ieee80211_sdata_running(sdata)))
+		return -ENETDOWN;
 
 	if (WARN_ON(compare_ether_addr(sta->sta.addr, sdata->vif.addr) == 0 ||
-		    is_multicast_ether_addr(sta->sta.addr))) {
-		err = -EINVAL;
+		    is_multicast_ether_addr(sta->sta.addr)))
+		return -EINVAL;
+
+	return 0;
+}

That seems nice, even simplifying the code :)

+static int sta_info_insert_adhoc(struct sta_info *sta) __acquires(RCU)
+{
+	struct ieee80211_local *local = sta->local;
+	struct ieee80211_sub_if_data *sdata = sta->sdata;
+	unsigned long flags;
+
+	spin_lock_irqsave(&local->sta_lock, flags);
+	/* check if STA exists already */
+	if (sta_info_get_bss(sdata, sta->sta.addr)) {
+		spin_unlock_irqrestore(&local->sta_lock, flags);
 		rcu_read_lock();
-		goto out_free;
+		__sta_info_free(local, sta);
+		return -EEXIST;
 	}
 
-	/*
-	 * In ad-hoc mode, we sometimes need to insert stations
-	 * from tasklet context from the RX path. To avoid races,
-	 * always do so in that case -- see the comment below.
-	 */
-	if (sdata->vif.type == NL80211_IFTYPE_ADHOC) {
-		spin_lock_irqsave(&local->sta_lock, flags);
-		/* check if STA exists already */
-		if (sta_info_get_bss(sdata, sta->sta.addr)) {
-			spin_unlock_irqrestore(&local->sta_lock, flags);
-			rcu_read_lock();
-			err = -EEXIST;
-			goto out_free;
-		}
-
-		local->num_sta++;
-		local->sta_generation++;
-		smp_mb();
-		sta_info_hash_add(local, sta);
+	local->num_sta++;
+	local->sta_generation++;
+	smp_mb();
+	sta_info_hash_add(local, sta);
 
-		list_add_tail(&sta->list, &local->sta_pending_list);
+	list_add_tail(&sta->list, &local->sta_pending_list);
 
-		rcu_read_lock();
-		spin_unlock_irqrestore(&local->sta_lock, flags);
+	rcu_read_lock();
+	spin_unlock_irqrestore(&local->sta_lock, flags);
 
 #ifdef CONFIG_MAC80211_VERBOSE_DEBUG
-		wiphy_debug(local->hw.wiphy, "Added IBSS STA %pM\n",
-			    sta->sta.addr);
+	wiphy_debug(local->hw.wiphy, "Added IBSS STA %pM\n",
+			sta->sta.addr);
 #endif /* CONFIG_MAC80211_VERBOSE_DEBUG */
 
-		ieee80211_queue_work(&local->hw, &local->sta_finish_work);
+	ieee80211_queue_work(&local->hw, &local->sta_finish_work);
 
-		return 0;
-	}
+	return 0;
+}

Yup, that looks good too, just the patch is a bit odd due to the code
indentation change.

+/*
+ * should be called with sta_mtx locked
+ * this function replaces the mutex lock
+ * with a RCU lock
+ */
+static int sta_info_insert_mgd(struct sta_info *sta) __acquires(RCU)
+{
+	struct ieee80211_local *local = sta->local;
+	struct ieee80211_sub_if_data *sdata = sta->sdata;
+	unsigned long flags;
+	int err = 0;
+
+	lockdep_assert_held(&local->sta_mtx);
 
 	/*
 	 * On first glance, this will look racy, because the code
-	 * below this point, which inserts a station with sleeping,
+	 * in this function, which inserts a station with sleeping,
 	 * unlocks the sta_lock between checking existence in the
 	 * hash table and inserting into it.
 	 *
 	 * However, it is not racy against itself because it keeps
-	 * the mutex locked. It still seems to race against the
-	 * above code that atomically inserts the station... That,
-	 * however, is not true because the above code can only
-	 * be invoked for IBSS interfaces, and the below code will
-	 * not be -- and the two do not race against each other as
-	 * the hash table also keys off the interface.
+	 * the mutex locked.
 	 */
 
-	might_sleep();
-
-	mutex_lock(&local->sta_mtx);
-
 	spin_lock_irqsave(&local->sta_lock, flags);
 	/* check if STA exists already */
 	if (sta_info_get_bss(sdata, sta->sta.addr)) {
 		spin_unlock_irqrestore(&local->sta_lock, flags);
 		mutex_unlock(&local->sta_mtx);
 		rcu_read_lock();
-		err = -EEXIST;
-		goto out_free;
+		return -EEXIST;

Isn't that missing the free now?

and here?

+int sta_info_insert_rcu(struct sta_info *sta) __acquires(RCU)
+{
+	struct ieee80211_local *local = sta->local;
+	struct ieee80211_sub_if_data *sdata = sta->sdata;
+	int err = 0;
+
+	err = sta_info_insert_check(sta);
+	if (err) {
+		rcu_read_lock();
+		goto out_free;
+	}
+
+	/*
+	 * In ad-hoc mode, we sometimes need to insert stations
+	 * from tasklet context from the RX path. To avoid races,
+	 * always do so in that case -- see the comment below.
+	 */
+	if (sdata->vif.type == NL80211_IFTYPE_ADHOC)
+		return sta_info_insert_adhoc(sta);
+
+	/*
+	 * It might seem that the function called below is in race against
+	 * the function call above that atomically inserts the station... That,
+	 * however, is not true because the above code can only
+	 * be invoked for IBSS interfaces, and the below code will
+	 * not be -- and the two do not race against each other as
+	 * the hash table also keys off the interface.
+	 */
+
+	might_sleep();
+
+	mutex_lock(&local->sta_mtx);
+
+	err = sta_info_insert_mgd(sta);
+	if (err)
+		goto out_free;
+
+	return 0;
  out_free:
 	BUG_ON(!err);
 	__sta_info_free(local, sta);

Ah, ok, I guess not.

Hmm. Can we make the ibss vs. non-IBSS code paths more similar wrt.
freeing? The _ibss one will free it, but the _mgd one won't, that seems
a bit strange to me.

Also, I'd prefer to move the might_sleep/locking into _mgd() if
possible?

johannes