Re: [PATCH v6 3/4] clocksource/drivers: Add a goldfish-timer clocksource

[Laurent Vivier]

Le 14/01/2022 à 11:46, Arnd Bergmann a écrit :
> On Thu, Jan 13, 2022 at 9:19 PM Laurent Vivier <laurent@xxxxxxxxx> wrote:
> > Add a clocksource based on the goldfish-rtc device.
> >
> > Signed-off-by: Laurent Vivier <laurent@xxxxxxxxx>
> > ---
> >   drivers/clocksource/Kconfig          |   7 ++
> >   drivers/clocksource/Makefile         |   1 +
> >   drivers/clocksource/timer-goldfish.c | 130 +++++++++++++++++++++++++++
> >   include/clocksource/timer-goldfish.h |  12 +++
> >   4 files changed, 150 insertions(+)
> >   create mode 100644 drivers/clocksource/timer-goldfish.c
> >   create mode 100644 include/clocksource/timer-goldfish.h
> >
> > diff --git a/drivers/clocksource/Kconfig b/drivers/clocksource/Kconfig
> > index f65e31bab9ae..6ca9bb78407d 100644
> > --- a/drivers/clocksource/Kconfig
> > +++ b/drivers/clocksource/Kconfig
> > @@ -711,4 +711,11 @@ config MICROCHIP_PIT64B
> >            modes and high resolution. It is used as a clocksource
> >            and a clockevent.
> >
> > +config GOLDFISH_TIMER
> > +       bool "Clocksource using goldfish-rtc"
> > +       select RTC_CLASS
> > +       select RTC_DRV_GOLDFISH
>
> This should probably be
>
>            depends on M68K || COMPILE_TEST
>            depends on RTC_DRV_GOLDFISH
>
> A driver should never 'select' another user-selectable subsystem

ok

> > +static int goldfish_timer_set_oneshot(struct clock_event_device *evt)
> > +{
> > +       struct goldfish_timer *timerdrv = ced_to_gf(evt);
> > +       void __iomem *base = timerdrv->base;
> > +
> > +       __raw_writel(0, base + TIMER_ALARM_HIGH);
> > +       __raw_writel(0, base + TIMER_ALARM_LOW);
> > +       __raw_writel(1, base + TIMER_IRQ_ENABLED);
>
> As mentioned elsewhere, the __raw_* accessors are not portable, please
> use readl()/writel() here, or possibly ioread32_be()/iowrite32_be() for
> the big-endian variant.

We can't use readl()/writel() here because it's supposed to read from a little endian device, and 
goldfish endianness depends on the endianness of the machine.

For goldfish, readl()/writel() works fine on little-endian machine but not on big-endian machine.

On m68k, you have:

#define readl(addr)      in_le32(addr)
#define writel(val,addr) out_le32((addr),(val))

and with goldfish it's wrong as the device is not little-endian, it is big-endian like the machine.

same comment with ioread32_be()/iowrite32_be(): it will work on big-endian machine not on little-endian.

We need an accessor that doesn't byteswap the value, that accesses it natively, and in all other 
parts of the kernel __raw_writel() and __raw_readl() are used.

Thanks,
Laurent