mac80211: fix a race condition that could lead to a use-after-free on a timer
authorFelix Fietkau <nbd@nbd.name>
Mon, 20 Nov 2017 19:05:19 +0000 (20:05 +0100)
committerFelix Fietkau <nbd@nbd.name>
Mon, 20 Nov 2017 19:05:19 +0000 (20:05 +0100)
Signed-off-by: Felix Fietkau <nbd@nbd.name>
package/kernel/mac80211/patches/320-mac80211-properly-free-requested-but-not-started-TX-.patch [new file with mode: 0644]

diff --git a/package/kernel/mac80211/patches/320-mac80211-properly-free-requested-but-not-started-TX-.patch b/package/kernel/mac80211/patches/320-mac80211-properly-free-requested-but-not-started-TX-.patch
new file mode 100644 (file)
index 0000000..ecc5e49
--- /dev/null
@@ -0,0 +1,37 @@
+From: Johannes Berg <johannes.berg@intel.com>
+Date: Mon, 20 Nov 2017 17:01:44 +0100
+Subject: [PATCH] mac80211: properly free requested-but-not-started TX agg
+ sessions
+
+When deleting a station or otherwise tearing down all aggregation
+sessions, make sure to delete requested but not yet started ones,
+to avoid the following scenario:
+
+ * session is requested, added to tid_start_tx[]
+ * ieee80211_ba_session_work() runs, gets past BLOCK_BA check
+ * ieee80211_sta_tear_down_BA_sessions() runs, locks &sta->ampdu_mlme.mtx,
+   e.g. while deleting the station - deleting all active sessions
+ * ieee80211_ba_session_work() continues since tear down flushes it, and
+   calls ieee80211_tx_ba_session_handle_start() for the new session, arms
+   the timer for it
+ * station deletion continues to __cleanup_single_sta() and frees the
+   session struct, while the timer is armed
+
+Reported-by: Fengguang Wu <fengguang.wu@intel.com>
+Signed-off-by: Johannes Berg <johannes.berg@intel.com>
+---
+
+--- a/net/mac80211/agg-tx.c
++++ b/net/mac80211/agg-tx.c
+@@ -330,6 +330,11 @@ int ___ieee80211_stop_tx_ba_session(stru
+       spin_lock_bh(&sta->lock);
++      /* free struct pending for start, if present */
++      tid_tx = sta->ampdu_mlme.tid_start_tx[tid];
++      kfree(tid_tx);
++      sta->ampdu_mlme.tid_start_tx[tid] = NULL;
++
+       tid_tx = rcu_dereference_protected_tid_tx(sta, tid);
+       if (!tid_tx) {
+               spin_unlock_bh(&sta->lock);