On Thu, Jan 5, 2017 at 9:33 AM, David Miller <da...@davemloft.net> wrote: > From: Wang Yufen <wangyu...@huawei.com> > Date: Thu, 5 Jan 2017 22:13:21 +0800 > >> From: Yufen Wang <wangyu...@huawei.com> >> >> A possible NULL pointer dereference in tg3_get_stats64 while doing >> tg3_free_consistent. > ... >> This patch avoids the NULL pointer dereference by using !tg3_flag(tp, >> INIT_COMPLETE) >> instate of !tp->hw_stats. >> >> Signed-off-by: Yufen Wang <wangyu...@huawei.com> >> --- >> drivers/net/ethernet/broadcom/tg3.c | 2 +- >> 1 file changed, 1 insertion(+), 1 deletion(-) >> >> diff --git a/drivers/net/ethernet/broadcom/tg3.c >> b/drivers/net/ethernet/broadcom/tg3.c >> index 185e9e0..012f18d 100644 >> --- a/drivers/net/ethernet/broadcom/tg3.c >> +++ b/drivers/net/ethernet/broadcom/tg3.c >> @@ -14148,7 +14148,7 @@ static struct rtnl_link_stats64 >> *tg3_get_stats64(struct net_device *dev, >> struct tg3 *tp = netdev_priv(dev); >> >> spin_lock_bh(&tp->lock); >> - if (!tp->hw_stats) { >> + if (!tg3_flag(tp, INIT_COMPLETE)) { > > The real issue is the manner and order in which the driver performs > initialization actions relative to netif_device_{attach,detach}(). > > That is what needs to be fixed here, instead of adding more and more > ad-hoc tests to the various methods which can be invoked once the > netif_device_attach() occurs.
Normally, ndo_get_stats64() should be under rtnl lock in the netlink code path and we should be safe. We only free tp->hw_stats under rtnl lock in the close path or ethtool path. But it looks like ndo_get_stats() can be called without rtnl lock from net-procfs.c. So it is possible that we'll read tp->hw_stats after it has been freed. For example, if we are reading /proc/net/dev and closing tg3 at the same time. David, is not taking rtnl_lock in net-procfs.c by design?