RE: [PATCH v3 3/3] timer: riscv_aclint_timer: add timer_get_boot_us for BOOTSTAGE

["Chanho Park" <chanho61.park@samsung.com>]

Hi,

> -----Original Message-----
> From: Heinrich Schuchardt <heinrich.schuchardt@canonical.com>
> Sent: Thursday, September 7, 2023 2:27 AM
> To: Chanho Park <chanho61.park@samsung.com>; Simon Glass
> <sjg@chromium.org>
> Cc: u-boot@lists.denx.de; Rick Chen <rick@andestech.com>; Leo
> <ycliang@andestech.com>
> Subject: Re: [PATCH v3 3/3] timer: riscv_aclint_timer: add
> timer_get_boot_us for BOOTSTAGE
> 
> On 06.09.23 07:18, Chanho Park wrote:
> > timer_get_boot_us function is required to record the boot stages as
> > us-based timestamp.
> > To get a micro-second time from a timer tick, this converts the
> > formula like below to avoid zero result of (tick / rate) part.
> >
> > From: time(us) = (tick / rate) * 10000000
Still typo 10000000 -> 1000000
> 
> Where is the old implementation that you refer to?

I referred it from timer_get_boot_us function of lib/time.c

lib/time.c
55         else if (timer_rate > 1000000)
56                 return lldiv(count, timer_rate / 1000000);
57         else
58                 return (unsigned long long)count * 1000000 / timer_rate;

> 
> > To  : time(us) = (tick * 1000) / (rate / 1000)
> >
> > Signed-off-by: Chanho Park <chanho61.park@samsung.com>
> > ---
> >   drivers/timer/riscv_aclint_timer.c | 23 +++++++++++++++++++++++
> >   1 file changed, 23 insertions(+)
> >
> > diff --git a/drivers/timer/riscv_aclint_timer.c
> b/drivers/timer/riscv_aclint_timer.c
> > index e29d527c8d77..73fb87912851 100644
> > --- a/drivers/timer/riscv_aclint_timer.c
> > +++ b/drivers/timer/riscv_aclint_timer.c
> > @@ -6,6 +6,7 @@
> >
> >   #include <common.h>
> >   #include <clk.h>
> > +#include <div64.h>
> >   #include <dm.h>
> >   #include <timer.h>
> >   #include <asm/io.h>
> > @@ -44,6 +45,28 @@ u64 notrace timer_early_get_count(void)
> >   }
> >   #endif
> >
> > +#if CONFIG_IS_ENABLED(RISCV_MMODE) && CONFIG_IS_ENABLED(BOOTSTAGE)
> > +ulong timer_get_boot_us(void)
> > +{
> > +	int ret;
> > +	u64 ticks = 0;
> > +	u32 rate;
> > +
> > +	ret = dm_timer_init();
> > +	if (!ret) {
> > +		rate = timer_get_rate(gd->timer);
> > +		timer_get_count(gd->timer, &ticks);
> > +	} else {
> > +		rate = RISCV_MMODE_TIMER_FREQ;
> > +		ticks = readq((void __iomem
> *)MTIME_REG(RISCV_MMODE_TIMERBASE,
> > +							RISCV_MMODE_TIMEROFF));
> > +	}
> > +
> > +	/* Below is converted from time(us) = (tick / rate) * 10000000 */
> > +	return lldiv(ticks * 1000, (rate / 1000));
> 
> I found similar code in drivers/timer/cadence-ttc.c and
> drivers/timer/omap-timer.c with
> 
>      us = (ticks * 1000) / rate;
>      return us.
> 
> Either their code or yours must be wrong.
> 
> What I am missing in include/timer.h is a documentation that defines if
> timer_dev_priv.clock_rate and timer_get_rate() yield the frequency in Hz
> or kHz.

'rate' seems to be Hz not kHz. So, I think they need to be corrected.

> 
> Once we have added the missing information in the include we can start
> reviewing this patch.
> 
> I really dislike that we have code per architecture and don't update and
> use a implementation in lib/time.c (where we also have an
> implementation) or drivers/timer/timer-uclass.c. Can't we have a single
> implementation which is driver model based and eliminate all others?

Actually, I tried to use lib/time.c's implementation or make a generic function in timer-uclass.c as you mentioned.
However, there are different implementations in cadence-ttc.c, rockchip_timer.c, tegra-timer.c, omap-timer.c and tsc_timer.c.
The basic codes seems to be almost identical if we can get a DM timer successfully but fallback codes look different.
If timer_get_boot_us can be implement in the timer-uclass.c, we still need per-driver callback function for supporting these different fallback codes.

Best Regards,
Chanho Park