On Wed, 2008-04-23 at 11:47 +0400, Pavel Emelyanov wrote:
quoted text
> The put routine first decrements the users counter and then
> (if it is zero) locks the sta_bss_lock and removes one from
> the list and the hash.
>=20
> Thus, any of ieee80211_sta_config_auth, ieee80211_rx_bss_get
> or ieee80211_rx_mesh_bss_get can race with it by finding a
> bss that is about to get kfree-ed.
>=20
> Using atomic_dec_and_lock in ieee80211_rx_bss_put takes care
> of this race.

Good catch, thanks.

quoted text
> Signed-off-by: Pavel Emelyanov <xemul@openvz.org>

Acked-by: Johannes Berg <johannes@sipsolutions.net>

I think we will need this in -stable (different file, same contents)

quoted text
> ---
>=20
> diff --git a/net/mac80211/mlme.c b/net/mac80211/mlme.c
> index 6b75cb6..dac3c2a 100644
> --- a/net/mac80211/mlme.c
> +++ b/net/mac80211/mlme.c
> @@ -2248,10 +2248,13 @@ static void ieee80211_rx_bss_put(struct net_devic=
e *dev,
quoted text
>  				 struct ieee80211_sta_bss *bss)
>  {
>  	struct ieee80211_local *local =3D wdev_priv(dev->ieee80211_ptr);
> -	if (!atomic_dec_and_test(&bss->users))
> +
> +	local_bh_disable();
> +	if (!atomic_dec_and_lock(&bss->users, &local->sta_bss_lock)) {
> +		local_bh_enable();
>  		return;
> +	}
> =20
> -	spin_lock_bh(&local->sta_bss_lock);
>  	__ieee80211_rx_bss_hash_del(dev, bss);
>  	list_del(&bss->list);
>  	spin_unlock_bh(&local->sta_bss_lock);
>=20