Re: [PATCH] phy: cadence: Sierra: Add support for skipping configuration

[Dan Carpenter <dan.carpenter@oracle.com>]

On Thu, Jan 27, 2022 at 02:26:58PM +0530, Aswath Govindraju wrote:
> Skip the phy configuration if the required configurations were done in an
> earlier boot stage.
> 

Why are you doing this?  Could you please put in the commit message if
the user will see an improvement from this change.

> Signed-off-by: Aswath Govindraju <a-govindraju@ti.com>
> ---

[ snip ]

> @@ -1382,16 +1401,24 @@ static int cdns_sierra_phy_probe(struct platform_device *pdev)
>  	if (ret)
>  		return ret;
>  
> -	ret = cdns_sierra_phy_get_resets(sp, dev);
> -	if (ret)
> -		goto unregister_clk;
> -
>  	ret = cdns_sierra_phy_enable_clocks(sp);
>  	if (ret)
>  		goto unregister_clk;
>  
> -	/* Enable APB */
> -	reset_control_deassert(sp->apb_rst);
> +	regmap_field_read(sp->pma_cmn_ready, &sp->already_configured);
> +
> +	if (!(sp->already_configured)) {

Delete extra parens.

> +		ret = cdns_sierra_phy_clk(sp);
> +		if (ret)
> +			goto unregister_clk;

The goto should release the most recent successful allocation which is
cdns_sierra_phy_enable_clocks().  So this should be goto clk_disable.
Except that will also call reset_control_assert() which is wrong...  The
rules are generally that error handling should be in the reverse order
from how we allocated it.  If allocation is optional the cleanup should
be optional.  The allocation and unwind code should mirror each other.

> +
> +		ret = cdns_sierra_phy_get_resets(sp, dev);
> +		if (ret)
> +			goto unregister_clk;

goto clk_disable;

> +
> +		/* Enable APB */
> +		reset_control_deassert(sp->apb_rst);

Since this is now optional it should be optional in the cleanup.

> +	}
>  

Since the order of allocations has changed, the other gotos need to be
updated to free the most recent allocation as well.  Then the error
handling looks like this:

	return 0;

put_control:
	while (--node >= 0)
		reset_control_put(sp->phys[node].lnk_rst);
ctrl_assert:
	if (!sp->already_configured)
		reset_control_assert(sp->apb_rst);
clk_disable:
	cdns_sierra_phy_disable_clocks(sp);
unregister_clk:
	cdns_sierra_clk_unregister(sp);
	return ret;


>  	/* Check that PHY is present */
>  	regmap_field_read(sp->macro_id_type, &id_value);
> @@ -1433,8 +1460,10 @@ static int cdns_sierra_phy_probe(struct platform_device *pdev)
>  
>  		sp->num_lanes += sp->phys[node].num_lanes;
>  
> -		gphy = devm_phy_create(dev, child, &ops);
> -
> +		if (!(sp->already_configured))

Delete parens.

> +			gphy = devm_phy_create(dev, child, &ops);
> +		else
> +			gphy = devm_phy_create(dev, child, &noop_ops);
>  		if (IS_ERR(gphy)) {
>  			ret = PTR_ERR(gphy);
>  			of_node_put(child);
> @@ -1455,7 +1484,7 @@ static int cdns_sierra_phy_probe(struct platform_device *pdev)
>  	}
>  
>  	/* If more than one subnode, configure the PHY as multilink */
> -	if (!sp->autoconf && sp->nsubnodes > 1) {
> +	if (!(sp->already_configured && sp->autoconf) && sp->nsubnodes > 1) {

It's normally easier to understand conditions when you push the ! as
far in as possible:

	if ((!sp->already_configured || !sp->autoconf) &&
	    sp->nsubnodes > 1) {

Is this condition right?  Shouldn't it be:

	if (!sp->already_configured && !sp->autoconf && sp->nsubnodes > 1) {

The ->already_configured is set/stored in firmware so I don't know when
that happens.  Please, add that information to the commit message when
you resend.

regards,
dan carpenter