Re: [RFC PATCH] net: don't set __LINK_STATE_START until after dev->open() call

From: Jacob Keller <jacob.e.keller@intel.com>
Date: Mon,  7 Aug 2017 15:24:21 -0700

Fix an issue with relying on netif_running() which could be true during
when dev->open() handler is being called, even if it would exit with
a failure. This ensures the state does not get set and removed with
a narrow race for other callers to read it as open when infact it never
finished opening.

Signed-off-by: Jacob Keller <jacob.e.keller@intel.com>
---
I found this as a result of debugging a race condition in the i40evf
driver, in which we assumed that netif_running() would not be true until
after dev->open() had been called and succeeded. Unfortunately we can't
hold the rtnl_lock() while checking netif_running() because it would
cause a deadlock between our reset task and our ndo_open handler.

I am wondering whether the proposed change is acceptable here, or
whether some ndo_open handlers rely on __LINK_STATE_START being true
prior to their being called?

I think this has the potential to break a bunch of drivers, but I
cannot prove this.

A lot of drivers have several pieces of state setup when they bring
the device up.  And these routines are also invoked from other code
paths like suspend/resume, PCI-E error recovery, etc. and they
probably do netif_running() calls here and there.

This behavior has been this way for a very long time, so the risk is
quite high I think.