Re: [PATCH net-next v2] net: phy: Fix suspicious rcu_dereference usage

On Sun, Jan 19, 2025 at 01:45:18PM +0100, Kory Maincent wrote:

On Fri, 17 Jan 2025 19:07:20 -0800
Jakub Kicinski [off-list ref] wrote:

quoted

On Fri, 17 Jan 2025 23:16:59 +0100 Kory Maincent wrote:

quoted

quoted

If not protected by RTNL, what prevents two threads from calling this
function at the same time,
thus attempting to kfree_rcu() the same pointer twice ?    

I don't think this function can be called simultaneously from two threads,
if this were the case we would have already seen several issues with the
phydev pointer. But maybe I am wrong.

The rcu_lock here is to prevent concurrent dev->hwprov pointer modification
done under rtnl_lock in net/ethtool/tsconfig.c.  

I could also be wrong, but I don't recall being told that suspend path
can't race with anything else. So I think ravb should probably take
rtnl_lock or some such when its shutting itself down.. ?

Should we add an ASSERT_RTNL call in the phy_detach function? (Maybe
also in phy_attach to be consistent)
Even thought, I think it may raise lots of warning from other NIT drivers.

How many drivers use phy_detach() ?

The answer is... phylink, bcm genet and xgbe.

Of the phylink ones:

1. phylink_connect_phy() - for use by drivers. This had better be
   called _before_ the netdev is registered (without rtnl) or
   from .ndo_open that holds the RTNL.

2. phylink_fwnode_phy_connect() - same as above.

3. phylink_sfp_config_phy(), called from the SFP code, and its state
   machines. It will be holding RTNL, because it is only safe to
   attach and detach PHYs from a registered netdev while holding RTNL.

I haven't looked any further.

-- 
RMK's Patch system: https://www.armlinux.org.uk/developer/patches/
FTTP is here! 80Mbps down 10Mbps up. Decent connectivity at last!