Re: [PATCH net-next v5 01/13] net: phy: Introduce ethernet link topology representation

[Maxime Chevallier]

Hi Jakub,

On Thu, 4 Jan 2024 15:12:42 -0800
Jakub Kicinski <kuba@xxxxxxxxxx> wrote:

> On Thu, 21 Dec 2023 19:00:34 +0100 Maxime Chevallier wrote:
> > @@ -2441,6 +2442,7 @@ struct net_device {
> >  #if IS_ENABLED(CONFIG_CGROUP_NET_PRIO)
> >  	struct netprio_map __rcu *priomap;
> >  #endif
> > +	struct phy_link_topology	link_topo;  
> 
> Perhaps others would disagree but can we make this a pointer instead?
> Only allocate it on demand, when first PHY gets attached?
> Both saves space and netdevice.h will no longer need to know the
> definition of the struct.
>
> Complete noob question but I thought PHYs get attached at ndo_open
> time for drivers, don't they? We shouldn't want to re-ID in that case.

I'll give it a try, it could be doable to try to keep the index for
_some_ PHYs. I had a first try at this in the very first iteration, but
that was lost when converting to xarray for the index management. Since
the phy_device keeps track of its own index, we can try to re-use it,
but I can see it becoming difficult if we destroy the topology when all
PHYs are detached (after ndo_close).

When re-creating the xarray at the next ndo_open, we would need to know
what was the previous next_index, and we can't know that until PHYs are
all re-attached.

For now it would work because in most cases, we only have one PHY
behind the MAC we can guarantee won't go away, but there are devices
out-there with 2 PHYs connected to the same MAC with a MUX in the
middle (that's one thing this series prepares for), and here it would
be more complex to re-allocate the topology while knowing what's the
next unused index. Hope that makes sense :)

To summarize, I don't think we can easily both dynamically
allocate/destroy the topology based only on the presence of at least
one PHY, and try to keep the PHY index the same for non-hot-pluggable
PHYs.

> 
> >  	struct phy_device	*phydev;
> >  	struct sfp_bus		*sfp_bus;
> >  	struct lock_class_key	*qdisc_tx_busylock;  
> 
> > @@ -10872,6 +10873,8 @@ struct net_device *alloc_netdev_mqs(int sizeof_priv, const char *name,
> >  #ifdef CONFIG_NET_SCHED
> >  	hash_init(dev->qdisc_hash);
> >  #endif
> > +	phy_link_topo_init(&dev->link_topo);
> > +
> >  	dev->priv_flags = IFF_XMIT_DST_RELEASE | IFF_XMIT_DST_RELEASE_PERM;
> >  	setup(dev);
> >    
> 
> I think you're missing a call to xa_destroy() somewhere, no?

Arg you're right... I was under the false assumption that because the
xarray wasn't dynamically allocated, we didn't need to perform any
cleanup, thanks for spotting this.

Maxime