2020-03-19 16:57:19

by Heiner Kallweit

[permalink] [raw]
Subject: Re: [PATCH net-next 3/7] net: phy: introduce phy_read_mmd_poll_timeout macro

On 19.03.2020 17:39, Dejin Zheng wrote:
> it is sometimes necessary to poll a phy register by phy_read_mmd()
> function until its value satisfies some condition. introduce
> phy_read_mmd_poll_timeout() macros that do this.
>
> Signed-off-by: Dejin Zheng <[email protected]>
> ---
> include/linux/phy.h | 4 ++++
> 1 file changed, 4 insertions(+)
>
> diff --git a/include/linux/phy.h b/include/linux/phy.h
> index 36d9dea04016..a30e9008647f 100644
> --- a/include/linux/phy.h
> +++ b/include/linux/phy.h
> @@ -24,6 +24,7 @@
> #include <linux/mod_devicetable.h>
> #include <linux/u64_stats_sync.h>
> #include <linux/irqreturn.h>
> +#include <linux/iopoll.h>
>
> #include <linux/atomic.h>
>
> @@ -784,6 +785,9 @@ static inline int __phy_modify_changed(struct phy_device *phydev, u32 regnum,
> */
> int phy_read_mmd(struct phy_device *phydev, int devad, u32 regnum);
>
> +#define phy_read_mmd_poll_timeout(val, cond, sleep_us, timeout_us, args...) \
> + read_poll_timeout(phy_read_mmd, val, cond, sleep_us, timeout_us, args)
> +
> /**
> * __phy_read_mmd - Convenience function for reading a register
> * from an MMD on a given PHY.
>
I'm not fully convinced. Usage of the new macro with its lots of
parameters makes the code quite hard to read. Therefore I'm also
not a big fan of readx_poll_timeout.

Even though I didn't invent it, I prefer the way DECLARE_RTL_COND
is used in the r8169 driver. The resulting code is much better
to read, and in case of a timeout a helpful warning is printed
automatically.


2020-03-19 17:13:08

by Andrew Lunn

[permalink] [raw]
Subject: Re: [PATCH net-next 3/7] net: phy: introduce phy_read_mmd_poll_timeout macro

On Thu, Mar 19, 2020 at 05:56:39PM +0100, Heiner Kallweit wrote:
> On 19.03.2020 17:39, Dejin Zheng wrote:
> > it is sometimes necessary to poll a phy register by phy_read_mmd()
> > function until its value satisfies some condition. introduce
> > phy_read_mmd_poll_timeout() macros that do this.
> >
> > Signed-off-by: Dejin Zheng <[email protected]>
> > ---
> > include/linux/phy.h | 4 ++++
> > 1 file changed, 4 insertions(+)
> >
> > diff --git a/include/linux/phy.h b/include/linux/phy.h
> > index 36d9dea04016..a30e9008647f 100644
> > --- a/include/linux/phy.h
> > +++ b/include/linux/phy.h
> > @@ -24,6 +24,7 @@
> > #include <linux/mod_devicetable.h>
> > #include <linux/u64_stats_sync.h>
> > #include <linux/irqreturn.h>
> > +#include <linux/iopoll.h>
> >
> > #include <linux/atomic.h>
> >
> > @@ -784,6 +785,9 @@ static inline int __phy_modify_changed(struct phy_device *phydev, u32 regnum,
> > */
> > int phy_read_mmd(struct phy_device *phydev, int devad, u32 regnum);
> >
> > +#define phy_read_mmd_poll_timeout(val, cond, sleep_us, timeout_us, args...) \
> > + read_poll_timeout(phy_read_mmd, val, cond, sleep_us, timeout_us, args)
> > +
> > /**
> > * __phy_read_mmd - Convenience function for reading a register
> > * from an MMD on a given PHY.
> >
> I'm not fully convinced. Usage of the new macro with its lots of
> parameters makes the code quite hard to read. Therefore I'm also
> not a big fan of readx_poll_timeout.

One issue is that people often implement polling wrong. They don't
take into account where extra delays can happen, and return -ETIMEOUT
when in fact the operation was successful. readx_poll_timeout gives us
a core which is known to be good.

What i don't like here is phy_read_mmd_poll_timeout() takes args... We
know it should be passed a phydev, and device address and a reg. I
would prefer these to be explicit parameters. We can then get the
compiler to check the correct number of parameters are passed.

Andrew