mac80211: fix key vs. sta locking problems

Up to now, key manipulation is supposed to run under RTNL to
avoid concurrent manipulations and also allow the set_key()
hardware callback to sleep. This is not feasible because STA
structs are rcu-protected and thus a lot of operations there
cannot take the RTNL. Also, key references are rcu-protected
so we cannot do things atomically.

This patch changes key locking completely:
 * key operations are now atomic
 * hardware crypto offload is enabled and disabled from
   a workqueue, due to that key freeing is also delayed
 * debugfs code is also run from a workqueue
 * keys reference STAs (and vice versa!) so during STA
   unlink the STAs key reference is removed but not the
   keys STA reference, to avoid races key todo work is
   run before STA destruction.
 * fewer STA operations now need the RTNL which was
   required due to key operations

This fixes the locking problems lockdep pointed out and also
makes things more light-weight because the rtnl isn't required
as much.

Note that the key todo lock/key mutex are global locks, this
is not required, of course, they could be per-hardware instead.

Signed-off-by: Johannes Berg <johannes@sipsolutions.net>
Signed-off-by: John W. Linville <linville@tuxdriver.com>

1 files changed, 2 insertions, 2 deletions

diff --git a/net/mac80211/ieee80211_i.h b/net/mac80211/ieee80211_i.h
index 8614c35544bb..7f4e7f93ee47 100644
--- a/net/mac80211/ieee80211_i.h
+++ b/net/mac80211/ieee80211_i.h
@@ -600,8 +600,8 @@ struct ieee80211_local {
         /*
          * The lock only protects the list, hash, timer and counter
          * against manipulation, reads are done in RCU. Additionally,
-         * the lock protects each BSS's TIM bitmap and a few items
-         * in a STA info structure.
+         * the lock protects each BSS's TIM bitmap, a few items in
+         * STA info structures and various key pointers.
          */
         spinlock_t sta_lock;
         unsigned long num_sta;