Re: [PATCH v3] clocksource: Add driver for the Ingenic JZ47xx OST

[Paul Cercueil]

Hi Daniel,

Le sam., févr. 8, 2020 at 08:09, Daniel Lezcano 
<daniel.lezcano@xxxxxxxxxx> a écrit :
> Hi Paul,
>
> On 15/01/2020 21:58, Paul Cercueil wrote:
> >  Le mer., janv. 15, 2020 at 20:54, Thomas Gleixner 
> > <tglx@xxxxxxxxxxxxx> a
> >  écrit :
> > >  Paul Cercueil <paul@xxxxxxxxxxxxxxx> writes:
> > > >   Le mer., janv. 15, 2020 at 18:48, Maarten ter Huurne
> > > >   <maarten@xxxxxxxxxxxxxx> a écrit :
> > > > >   On Wednesday, 15 January 2020 14:57:01 CET Paul Cercueil wrote:
> > > > > >    Le mer., janv. 15, 2020 at 14:44, Daniel Lezcano
> > > > > >    <daniel.lezcano@xxxxxxxxxx> a écrit :
> > > > > >    > Is the JZ47xx OST really a mfd needing a regmap? (Note 
> > > > > > regmap_read
> > > > > >    > will take a lock).
> > > > > >
> > > > > >    Yes, the TCU_REG_OST_TCSR register is shared with the clocks 
> > > > > > driver.
> > > > >
> > > > >   The TCU_REG_OST_TCSR register is only used in the probe though.
> > > > >
> > > > >   To get the counter value from TCU_REG_OST_CNTL/TCU_REG_OST_CNTH 
> > > > > you
> > > > >   could technically do it by reading the register directly, if
> > > > >   performance
> > > > >   concerns make it necessary to bypass the usual kernel 
> > > > > infrastructure
> > > > >   for
> > > > >   dealing with shared registers.
> > > >
> > > >   In theory yes, in practice there's no easy way to do that (the
> > > >   underlying mmio pointer is not obtainable from the regmap), and
> > > >   besides, the lock is just a spinlock and not a mutex.
> > >
> > >  That lock still a massive contention point as clock readouts can be
> > >  pretty
> > >  frequent depending on workloads. Just think about tracing ...
> > >
> > >  So I really would avoid both the lock and that ugly 64bit readout 
> > > thing.
> >
> >  The 64bit readout thing is gone in V3.
> >
> >  The lock cannot go away unless we have a way to retrieve the 
> > underlying
> >  mmio pointer from the regmap, which the regmap maintainers will 
> > never
> >  accept. So I can't really change that now. Besides,
> >  drivers/clocksource/ingenic-timer.c also registers a clocksource 
> > that's
> >  read with the regmap, and nobody complained.
>
> Is there any progress on this? Having a lock in this code path is very
> impacting.

I have a v4 ready, I'm just waiting for 5.6-rc1 to start to rebase on 
top and send it.

-Paul