cfg80211: avoid holding the RTNL when calling the driver

Currently, _everything_ in cfg80211 holds the RTNL, and if you
have a slow USB device (or a few) you can get some bad lock
contention on that.

Fix that by re-adding a mutex to each wiphy/rdev as we had at
some point, so we have locking for the wireless_dev lists and
all the other things in there, and also so that drivers still
don't have to worry too much about it (they still won't get
parallel calls for a single device).

Then, we can restrict the RTNL to a few cases where we add or
remove interfaces and really need the added protection. Some
of the global list management still also uses the RTNL, since
we need to have it anyway for netdev management, but we only
hold the RTNL for very short periods of time here.

Link: https://lore.kernel.org/r/20210122161942.81df9f5e047a.I4a8e1a60b18863ea8c5e6d3a0faeafb2d45b2f40@changeid
Tested-by: Marek Szyprowski <m.szyprowski@samsung.com> [marvell driver issues]
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
This commit is contained in:
Johannes Berg 2021-01-22 16:19:43 +01:00
parent 2fe8ef1062
commit a05829a722
39 changed files with 886 additions and 515 deletions

View file

@ -67,7 +67,6 @@ static int cfg80211_conn_scan(struct wireless_dev *wdev)
struct cfg80211_scan_request *request;
int n_channels, err;
ASSERT_RTNL();
ASSERT_WDEV_LOCK(wdev);
if (rdev->scan_req || rdev->scan_msg)
@ -233,7 +232,7 @@ void cfg80211_conn_work(struct work_struct *work)
u8 bssid_buf[ETH_ALEN], *bssid = NULL;
enum nl80211_timeout_reason treason;
rtnl_lock();
wiphy_lock(&rdev->wiphy);
list_for_each_entry(wdev, &rdev->wiphy.wdev_list, list) {
if (!wdev->netdev)
@ -266,7 +265,7 @@ void cfg80211_conn_work(struct work_struct *work)
wdev_unlock(wdev);
}
rtnl_unlock();
wiphy_unlock(&rdev->wiphy);
}
/* Returned bss is reference counted and must be cleaned up appropriately. */