cfg80211: don't "leak" uncompleted scans

___cfg80211_scan_done() can be called in some cases
(e.g. on NETDEV_DOWN) before the low level driver
notified scan completion (which is indicated by
passing leak=true).

Clearing rdev->scan_req in this case is buggy, as
scan_done_wk might have already being queued/running
(and can't be flushed as it takes rtnl()).

If a new scan will be requested at this stage, the
scan_done_wk will try freeing it (instead of the
previous scan), and this will later result in
a use after free.

Simply remove the "leak" option, and replace it with
a standard WARN_ON.

An example backtrace after such crash:
Unable to handle kernel paging request at virtual address fffffee5
pgd = c0004000
[fffffee5] *pgd=9fdf6821, *pte=00000000, *ppte=00000000
Internal error: Oops: 17 [#1] SMP ARM
PC is at cfg80211_scan_done+0x28/0xc4 [cfg80211]
LR is at __ieee80211_scan_completed+0xe4/0x2dc [mac80211]
[<bf0077b0>] (cfg80211_scan_done+0x28/0xc4 [cfg80211])
[<bf0973d4>] (__ieee80211_scan_completed+0xe4/0x2dc [mac80211])
[<bf0982cc>] (ieee80211_scan_work+0x94/0x4f0 [mac80211])
[<c005fd10>] (process_one_work+0x1b0/0x4a8)
[<c0060404>] (worker_thread+0x138/0x37c)
[<c0066d70>] (kthread+0xa4/0xb0)

Signed-off-by: Eliad Peller <eliad@wizery.com>
Signed-off-by: Johannes Berg <johannes.berg@intel.com>

1 files changed, 3 insertions, 13 deletions

diff --git a/net/wireless/scan.c b/net/wireless/scan.c
index d4397eb..a32d52a 100644
--- a/net/wireless/scan.c
+++ b/net/wireless/scan.c
@@ -161,7 +161,7 @@ static void __cfg80211_bss_expire(struct cfg80211_registered_device *dev,
 		dev->bss_generation++;
 }
 
-void ___cfg80211_scan_done(struct cfg80211_registered_device *rdev, bool leak)
+void ___cfg80211_scan_done(struct cfg80211_registered_device *rdev)
 {
 	struct cfg80211_scan_request *request;
 	struct wireless_dev *wdev;
@@ -210,17 +210,7 @@ void ___cfg80211_scan_done(struct cfg80211_registered_device *rdev, bool leak)
 		dev_put(wdev->netdev);
 
 	rdev->scan_req = NULL;
-
-	/*
-	 * OK. If this is invoked with "leak" then we can't
-	 * free this ... but we've cleaned it up anyway. The
-	 * driver failed to call the scan_done callback, so
-	 * all bets are off, it might still be trying to use
-	 * the scan request or not ... if it accesses the dev
-	 * in there (it shouldn't anyway) then it may crash.
-	 */
-	if (!leak)
-		kfree(request);
+	kfree(request);
 }
 
 void __cfg80211_scan_done(struct work_struct *wk)
@@ -231,7 +221,7 @@ void __cfg80211_scan_done(struct work_struct *wk)
 			    scan_done_wk);
 
 	rtnl_lock();
-	___cfg80211_scan_done(rdev, false);
+	___cfg80211_scan_done(rdev);
 	rtnl_unlock();
 }