Re: [PATCH 03/14] drivers: net: mdio: mdio-ip8064: improve busy wait delay

[Andrew Lunn]

On Fri, Apr 23, 2021 at 03:47:29AM +0200, Ansuel Smith wrote:
> With the use of the qca8k dsa driver, some problem arised related to
> port status detection. With a load on a specific port (for example a
> simple speed test), the driver starts to bheave in a strange way and
> garbage data is produced. To address this, enlarge the sleep delay and
> address a bug for the reg offset 31 that require additional delay for
> this specific reg.

>  struct ipq8064_mdio {
>  	struct regmap *base; /* NSS_GMAC0_BASE */
> @@ -65,7 +66,7 @@ ipq8064_mdio_read(struct mii_bus *bus, int phy_addr, int reg_offset)
>  		   ((reg_offset << MII_REG_SHIFT) & MII_REG_MASK);
>  
>  	regmap_write(priv->base, MII_ADDR_REG_ADDR, miiaddr);
> -	usleep_range(8, 10);
> +	usleep_range(10, 13);
>  
>  	err = ipq8064_mdio_wait_busy(priv);
>  	if (err)
> @@ -91,7 +92,14 @@ ipq8064_mdio_write(struct mii_bus *bus, int phy_addr, int reg_offset, u16 data)
>  		   ((reg_offset << MII_REG_SHIFT) & MII_REG_MASK);
>  
>  	regmap_write(priv->base, MII_ADDR_REG_ADDR, miiaddr);
> -	usleep_range(8, 10);
> +
> +	/* For the specific reg 31 extra time is needed or the next
> +	 * read will produce grabage data.
> +	 */
> +	if (reg_offset == 31)
> +		usleep_range(30, 43);
> +	else
> +		usleep_range(10, 13);
>  
>  	return ipq8064_mdio_wait_busy(priv);

Is there any documentation as to what register 31 does?

Maybe the real problem is in ipq8064_mdio_wait_busy()? Have you looked
at how long you typically end up waiting? If you know the MDIO bus
speed, you can work out how long a transaction should take. If it is
taking less, something is broken.

Are you sure regmap caching is disabled, so that
ipq8064_mdio_wait_busy() really is reading the register, not some
old cached value?

      Andrew





>  }
> -- 
> 2.30.2
>