[PATCH 05/12] phylib: Allow reading and writing a mii bus from atomic context.

[Grant Likely]

On Tue, Jun 15, 2010 at 11:08 AM, Richard Cochran
<richardcochran at gmail.com> wrote:
> On Tue, Jun 15, 2010 at 10:43:08AM -0600, Grant Likely wrote:
>> On Tue, Jun 15, 2010 at 10:08 AM, Richard Cochran
>> <richardcochran at gmail.com> wrote:
>> > In order to support hardware time stamping from a PHY, it is necessary to
>> > read from the PHY while running in_interrupt(). This patch allows a mii
>> > bus to operate in an atomic context. An mii_bus driver may declare itself
>> > capable for this mode. Drivers which do not do this will remain with the
>> > default that bus operations may sleep.
>> >
>> > Signed-off-by: Richard Cochran <richard.cochran at omicron.at>
>>
>> Last I checked, the MDIO bus is very slow.  Is this really a good
>> idea?  How much latency does MDIO access have on the hardware you are
>> working with?
>
> Yes, MDIO access is slow, and it can vary (eg bit banging
> implementations). It clear that getting PHY timestamps is costly, but
> for applications that want PTP synchronization, one is willing to pay
> the price.
>
>> I also don't like the idea of taking a spin lock during MDIO
>> operations, and the dual locking mode in the core code.
>
> Originally, the phylib used a spinlock for this. It was replaced with
> a mutex in 35b5f6b1a82b5c586e0b24c711dc6ba944e88ef1 in order to
> accommodate mdio busses that may need to sleep. So, keeping the option
> to use a spinlock is similar to the previous implementation.

That's right, and I fully agree with that change.  To me, going back
to allowing spin locks is a regression because it adds a new source of
scheduling latency.  Using a mutex forces users to take into account
the slow nature of MDIO access.  For existing callers, this isn't a
problem because they already are designed for this characteristic.  A
new user which depends on atomic access should use a different API
which doesn't take the lock with the understanding that it is may
return a failure if it doesn't support it or if it cannot perform the
operation atomically.

That still leaves the troubling MDIO induced latency issue.

g.

-- 
Grant Likely, B.Sc., P.Eng.
Secret Lab Technologies Ltd.