wifi: mac80211: fix BA session teardown race

[ Upstream commit 05f136220d17839eb7c155f015ace9152f603225 ]

As previously reported by Alexander, whose commit 69403bad97aa
("wifi: mac80211: sdata can be NULL during AMPDU start") I'm
reverting as part of this commit, there's a race between station
destruction and aggregation setup, where the aggregation setup
can happen while the station is being removed and queue the work
after ieee80211_sta_tear_down_BA_sessions() has already run in
__sta_info_destroy_part1(), and thus the worker will run with a
now freed station. In his case, this manifested in a NULL sdata
pointer, but really there's no guarantee whatsoever.

The real issue seems to be that it's possible at all to have a
situation where this occurs - we want to stop the BA sessions
when doing _part1, but we cannot be sure, and WLAN_STA_BLOCK_BA
isn't necessarily effective since we don't know that the setup
isn't concurrently running and already got past the check.

Simply call ieee80211_sta_tear_down_BA_sessions() again in the
second part of station destruction, since at that point really
nothing else can hold a reference to the station any more.

Also revert the sdata checks since those are just misleading at
this point.

Signed-off-by: Johannes Berg <johannes.berg@intel.com>
Signed-off-by: Sasha Levin <sashal@kernel.org>

3 files changed, 15 insertions, 8 deletions

diff --git a/net/mac80211/agg-tx.c b/net/mac80211/agg-tx.c
index b6b772685881..0a69e47f1c55 100644
--- a/net/mac80211/agg-tx.c
+++ b/net/mac80211/agg-tx.c
@@ -497,7 +497,7 @@ void ieee80211_tx_ba_session_handle_start(struct sta_info *sta, int tid)
 {
 	struct tid_ampdu_tx *tid_tx;
 	struct ieee80211_local *local = sta->local;
-	struct ieee80211_sub_if_data *sdata;
+	struct ieee80211_sub_if_data *sdata = sta->sdata;
 	struct ieee80211_ampdu_params params = {
 		.sta = &sta->sta,
 		.action = IEEE80211_AMPDU_TX_START,
@@ -525,7 +525,6 @@ void ieee80211_tx_ba_session_handle_start(struct sta_info *sta, int tid)
 	 */
 	synchronize_net();
 
-	sdata = sta->sdata;
 	params.ssn = sta->tid_seq[tid] >> 4;
 	ret = drv_ampdu_action(local, sdata, &params);
 	tid_tx->ssn = params.ssn;
@@ -539,9 +538,6 @@ void ieee80211_tx_ba_session_handle_start(struct sta_info *sta, int tid)
 		 */
 		set_bit(HT_AGG_STATE_DRV_READY, &tid_tx->state);
 	} else if (ret) {
-		if (!sdata)
-			return;
-
 		ht_dbg(sdata,
 		       "BA request denied - HW unavailable for %pM tid %d\n",
 		       sta->sta.addr, tid);
diff --git a/net/mac80211/driver-ops.c b/net/mac80211/driver-ops.c
index f8af0c3d405a..d6478fd00bad 100644
--- a/net/mac80211/driver-ops.c
+++ b/net/mac80211/driver-ops.c
@@ -393,9 +393,6 @@ int drv_ampdu_action(struct ieee80211_local *local,
 
 	might_sleep();
 
-	if (!sdata)
-		return -EIO;
-
 	sdata = get_bss_sdata(sdata);
 	if (!check_sdata_in_driver(sdata))
 		return -EIO;
diff --git a/net/mac80211/sta_info.c b/net/mac80211/sta_info.c
index 984f8f67492f..42ba51a9700f 100644
--- a/net/mac80211/sta_info.c
+++ b/net/mac80211/sta_info.c
@@ -1402,6 +1402,20 @@ static void __sta_info_destroy_part2(struct sta_info *sta, bool recalc)
 	 *	 after _part1 and before _part2!
 	 */
 
+	/*
+	 * There's a potential race in _part1 where we set WLAN_STA_BLOCK_BA
+	 * but someone might have just gotten past a check, and not yet into
+	 * queuing the work/creating the data/etc.
+	 *
+	 * Do another round of destruction so that the worker is certainly
+	 * canceled before we later free the station.
+	 *
+	 * Since this is after synchronize_rcu()/synchronize_net() we're now
+	 * certain that nobody can actually hold a reference to the STA and
+	 * be calling e.g. ieee80211_start_tx_ba_session().
+	 */
+	ieee80211_sta_tear_down_BA_sessions(sta, AGG_STOP_DESTROY_STA);
+
 	might_sleep();
 	lockdep_assert_held(&local->sta_mtx);