Re: [PATCH v6 5/5] MIPS: ranchu: Add Ranchu as a new generic-based board

[James Hogan]

On Thu, Nov 02, 2017 at 12:47:27PM +0000, Miodrag Dinic wrote:
> > > +static __init uint64_t read_rtc_time(void __iomem *base)
> > > +{
> > > +     u64 time_low;
> > > +     u64 time_high;
> > > +
> > > +     time_low = readl(base + GOLDFISH_TIMER_LOW);
> > > +     time_high = readl(base + GOLDFISH_TIMER_HIGH);
> > > +
> > > +     return (time_high << 32) | time_low;
> > 
> > What if high changes while reading this?
> > 
> > E.g.
> > TIMER_LOW        0x00000000 *0xffffffff*
> > TIMER_HIGH      *0x00000001* 0x00000000
> > 
> > You'd presumably get 0x00000001ffffffff.
> > 
> > Perhaps it should read HIGH before too, and retry if it has changed.
> 
> This was already discussed in some earlier posts. (https://patchwork.linux-mips.org/patch/16628/)
> Reading the low value first actually latches the high value,
> so it is safe to leave it this way. Here is the relevant RTC device
> implementation in QEMU:
> 
> static uint64_t goldfish_rtc_read(void *opaque, hwaddr offset, unsigned size)
> {
>     struct rtc_state *s = (struct rtc_state *)opaque;
>     switch(offset) {
>         case TIMER_TIME_LOW:
>             s->now_ns = qemu_clock_get_ns(QEMU_CLOCK_VIRTUAL) + s->time_base;
>             return s->now_ns;
>         case TIMER_TIME_HIGH:
>             return s->now_ns >> 32;

Ah okay. In that case the side effect of reading low clearly isn't
obvious enough reading the driver code and it needs a comment to clarify
the hardware behaviour :-)

Cheers
James