Message-ID: <20231003190011.jcpahazy3uqlz7q4@skbuf>
Date: Tue, 3 Oct 2023 22:00:11 +0300
From: Vladimir Oltean <vladimir.oltean@....com>
To: "Russell King (Oracle)" <linux@...linux.org.uk>
Cc: netdev@...r.kernel.org, devicetree@...r.kernel.org,
	linux-kernel@...r.kernel.org, linux-phy@...ts.infradead.org,
	Heiner Kallweit <hkallweit1@...il.com>,
	Andrew Lunn <andrew@...n.ch>,
	Florian Fainelli <f.fainelli@...il.com>,
	Madalin Bucur <madalin.bucur@....com>,
	Ioana Ciornei <ioana.ciornei@....com>,
	Camelia Groza <camelia.groza@....com>, Li Yang <leoyang.li@....com>,
	Rob Herring <robh+dt@...nel.org>,
	Krzysztof Kozlowski <krzysztof.kozlowski+dt@...aro.org>,
	Conor Dooley <conor@...nel.org>,
	Sean Anderson <sean.anderson@...o.com>,
	Maxime Chevallier <maxime.chevallier@...tlin.com>,
	Vinod Koul <vkoul@...nel.org>,
	Kishon Vijay Abraham I <kishon@...nel.org>
Subject: Re: [RFC PATCH v2 net-next 15/15] net: pcs: lynx: use MTIP AN/LT
 block for copper backplanes

Hi Russell,

On Tue, Oct 03, 2023 at 02:14:41PM +0100, Russell King (Oracle) wrote:
> On Sat, Sep 23, 2023 at 04:49:04PM +0300, Vladimir Oltean wrote:
> > +static int lynx_pcs_parse_fwnode(struct lynx_pcs *lynx)
> > +{
> > +	struct fwnode_handle *node = lynx->mdio->dev.fwnode;
> > +	enum mtip_model model = MTIP_MODEL_AUTODETECT;
> > +	struct device_node *np = to_of_node(node);
> > +	struct mdio_device *mdio = lynx->mdio;
> > +	struct device *dev = &mdio->dev;
> > +	struct phy *phy;
> > +	int i, err;
> > +
> > +	if (!node)
> > +		return 0;
> > +
> > +	lynx->backplane_mode = fwnode_property_present(node, "fsl,backplane-mode");
> > +	if (!lynx->backplane_mode)
> > +		return 0;
> > +
> > +	if (fwnode_device_is_compatible(node, "fsl,lx2160a-lynx-pcs"))
> > +		model = MTIP_MODEL_LX2160A;
> > +
> > +	lynx->num_lanes = of_count_phandle_with_args(np, "phys", "#phy-cells");
> > +	if (lynx->num_lanes < 0)
> > +		return lynx->num_lanes;
> 
> Is it possible for ->num_lanes to be zero at this point? If that is
> possible, then ->anlt[PRIMARY_LANE] will be NULL but ->backplane_mode
> will be set, so won't that cause the mtip_* calls above to pass a
> NULL pointer into those functions? Is that safe? Should we trap that
> case here?

Assuming the dt-bindings as proposed here, that case would be an invalid
device tree ("fsl,backplane-mode" present but "phys" isn't present),
which I indeed failed to catch.

But in my reply to Krzysztof here:
https://lore.kernel.org/netdev/20231002121958.xybzovgjzldfiae2@skbuf/

I said that for v3, I'm looking to move the "phys" property from the PCS
to the MAC (it's already in the MAC for the non-backplane use cases).

On LS1028A (ENETC, Felix), the lynx-pcs is not OF-based (we use
lynx_pcs_create_mdiodev()), but it would be good to support the
1000Base-KX link mode there also. As I'm starting to look beyond
LX2160A, I'm starting to see why adding extra dt-bindings to the
lynx-pcs (both "phys" and "fsl,backplane-mode") will be problematic
if there is no OF node to speak of.

I will leave a separate comment with some new ideas.

> If that's correct, then I don't see any point in storing
> ->backplane_mode, since we can then use ->num_lanes > PRIMARY_LANE
> or similar instead.

Well, in v3, my plan is for the caller of lynx_pcs_create() (aka the MAC)
to always pass an array of phys (the ones from its own OF node). In that
case, we would indeed need the "fsl,backplane-mode" property in the PCS,
because otherwise, with your proposal, the PCS would instantiate the
AN/LT block even when it's not expected.

> > +
> > +	if (WARN_ON(lynx->num_lanes > MAX_NUM_LANES))
> > +		return -EINVAL;
> 
> Do we need to use WARN_ON() here, or would it be better to print a short
> error-level message?

Admittedly I may not have the best intuition here, but I didn't want to
over-complicate the code with error messages that can only be triggered
with invalid device trees.

> > +
> > +	for (i = 0; i < lynx->num_lanes; i++) {
> > +		phy = devm_of_phy_get_by_index(dev, np, i);
> > +		if (IS_ERR(phy))
> > +			return dev_err_probe(dev, PTR_ERR(phy),
> > +					     "Failed to get SerDes PHY %d\n", i);
> > +
> > +		lynx->anlt[i] = mtip_backplane_create(mdio, phy, model);
> > +		if (IS_ERR(lynx->anlt[i])) {
> > +			err = PTR_ERR(lynx->anlt[i]);
> > +
> > +			while (i-- > 0)
> > +				mtip_backplane_destroy(lynx->anlt[i]);
> > +
> > +			return err;
> > +		}
> > +	}
> > +
> > +	for (i = 1; i < lynx->num_lanes; i++) {
> > +		err = mtip_backplane_add_subordinate(lynx->anlt[PRIMARY_LANE],
> > +						     lynx->anlt[i]);
> > +		if (WARN_ON(err)) {
> 
> Again, does this need to be a backtrace-producing WARN_ON()?

mtip_backplane_add_subordinate() will only return -ERANGE if called too
many times (more than MTIP_MAX_NUM_SUBORDINATES times, aka more than
"MAX_NUM_LANES - 1" times).

Given the way that the code is constructed, it is technically impossible
for that to happen, but only because MTIP_MAX_NUM_SUBORDINATES is
hand-crafted to be 3 and MAX_NUM_LANES to be 4. I think that if I define
MTIP_MAX_NUM_SUBORDINATES in terms of MAX_NUM_LANES - 1, I can simply
make mtip_backplane_add_subordinate() return void.

What I want to avoid is to add error handling for errors which cannot
take place. Which is where the WARN_ON() came from.

> > +			/* Too many SerDes lanes in the device tree? */
> > +			for (i = 0; i < lynx->num_lanes; i++)
> > +				mtip_backplane_destroy(lynx->anlt[i]);
> > +			return err;
> > +		}
> > +	}
> > +
> > +	return 0;
> > +}
> > +
> >  static struct phylink_pcs *lynx_pcs_create(struct mdio_device *mdio)
> >  {
> >  	struct lynx_pcs *lynx;
> > +	int err;
> >  
> >  	lynx = kzalloc(sizeof(*lynx), GFP_KERNEL);
> >  	if (!lynx)
> > @@ -327,6 +451,12 @@ static struct phylink_pcs *lynx_pcs_create(struct mdio_device *mdio)
> >  	lynx->pcs.neg_mode = true;
> >  	lynx->pcs.poll = true;
> >  
> > +	err = lynx_pcs_parse_fwnode(lynx);
> > +	if (err) {
> > +		kfree(lynx);
> > +		return ERR_PTR(err);
> > +	}
> > +
> >  	return lynx_to_phylink_pcs(lynx);
> >  }
> >  
> > @@ -392,6 +522,11 @@ EXPORT_SYMBOL_GPL(lynx_pcs_create_fwnode);
> >  void lynx_pcs_destroy(struct phylink_pcs *pcs)
> >  {
> >  	struct lynx_pcs *lynx = phylink_pcs_to_lynx(pcs);
> > +	int i;
> > +
> > +	if (lynx->backplane_mode)
> > +		for (i = 0; i < lynx->num_lanes; i++)
> > +			mtip_backplane_destroy(lynx->anlt[i]);
> 
> Won't ->num_lanes only be non-zero when ->backplane_mode is set, so
> isn't the test for ->backplane_mode redundant here?

I think it won't be redundant anymore once the series reaches a less
"WIP" state.

> >  
> >  	mdio_device_put(lynx->mdio);
> >  	kfree(lynx);
> > -- 
> > 2.34.1
> > 
> > 
> 
> -- 
> RMK's Patch system: https://www.armlinux.org.uk/developer/patches/
> FTTP is here! 80Mbps down 10Mbps up. Decent connectivity at last!

Powered by blists - more mailing lists