提交 13f172ff 编写于 作者: N Neil Horman 提交者: David S. Miller

netconsole: fix deadlock when removing net driver that netconsole is using (v2)

A deadlock was reported to me recently that occured when netconsole was being
used in a virtual guest.  If the virtio_net driver was removed while netconsole
was setup to use an interface that was driven by that driver, the guest
deadlocked.  No backtrace was provided because netconsole was the only console
configured, but it became clear pretty quickly what the problem was.  In
netconsole_netdev_event, if we get an unregister event, we call
__netpoll_cleanup with the target_list_lock held and irqs disabled.
__netpoll_cleanup can, if pending netpoll packets are waiting call
cancel_delayed_work_sync, which is a sleeping path.  the might_sleep call in
that path gets triggered, causing a console warning to be issued.  The
netconsole write handler of course tries to take the target_list_lock again,
which we already hold, causing deadlock.

The fix is pretty striaghtforward.  Simply drop the target_list_lock and
re-enable irqs prior to calling __netpoll_cleanup, the re-acquire the lock, and
restart the loop.  Confirmed by myself to fix the problem reported.
Signed-off-by: NNeil Horman <nhorman@tuxdriver.com>
CC: "David S. Miller" <davem@davemloft.net>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>
上级 1ed3aad1
......@@ -671,6 +671,7 @@ static int netconsole_netdev_event(struct notifier_block *this,
goto done;
spin_lock_irqsave(&target_list_lock, flags);
restart:
list_for_each_entry(nt, &target_list, list) {
netconsole_target_get(nt);
if (nt->np.dev == dev) {
......@@ -683,9 +684,16 @@ static int netconsole_netdev_event(struct notifier_block *this,
* rtnl_lock already held
*/
if (nt->np.dev) {
spin_unlock_irqrestore(
&target_list_lock,
flags);
__netpoll_cleanup(&nt->np);
spin_lock_irqsave(&target_list_lock,
flags);
dev_put(nt->np.dev);
nt->np.dev = NULL;
netconsole_target_put(nt);
goto restart;
}
/* Fall through */
case NETDEV_GOING_DOWN:
......
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册