Re: [PATCH v2 2/5] net: hpe: Add GXP UMAC MDIO

[Andrew Lunn]

On Wed, Aug 16, 2023 at 12:55:56AM +0000, Hawkins, Nick wrote:
> Greetings Andrew,
> 
> I have a follow up question below:
> 
> >> +static int umac_mdio_read(struct mii_bus *bus, int phy_id, int reg)
> >> +{
> >> + struct umac_mdio_priv *umac_mdio = bus->priv;
> >> + unsigned int status;
> >> + unsigned int value;
> >> + int ret;
> >> +
> >> + status = __raw_readl(umac_mdio->base + UMAC_MII);
> >> +
> >> + status &= ~(UMAC_MII_PHY_ADDR_MASK | UMAC_MII_REG_ADDR_MASK);
> >> + status |= ((phy_id << UMAC_MII_PHY_ADDR_SHIFT) &
> >> + UMAC_MII_PHY_ADDR_MASK);
> >> + status |= (reg & UMAC_MII_REG_ADDR_MASK);
> >> + status |= UMAC_MII_MRNW; /* set bit for read mode */
> >> + __raw_writel(status, umac_mdio->base + UMAC_MII);
> >> +
> >> + status |= UMAC_MII_MOWNER; /* set bit to activate mii transfer */
> >> + __raw_writel(status, umac_mdio->base + UMAC_MII);
> 
> 
> > I assume UMAC_MII_MOWNER must be set in a separate operation? But
> > using __raw_writel() i'm not sure there is any barrier between the two
> > writes.
> 
> Is there a function you would recommend using instead?

writel().

In general, it is best to use writel()/readl() for correctness. In the
hot path, dealing with actually Ethernet frames where every uS counts,
you can then think about using writel_relaxed()/readl_relaxed(). But
for something slow like an MDIO bus driver, i would always avoid the
possibility of having hard to find bugs because of missing barriers.

	Andrew