[PATCH v2 3/8] timer: orion-timer: Add timer_get_boot_us() for BOOTSTAGE support

[Michael Walle]

Am 2022-09-04 02:02, schrieb Tony Dinh:
> Hi Stefan,
> 
> Sorry, that message was prematurely sent (fat finger). Please see the
> continuation below.
> 
> On Sat, Sep 3, 2022 at 4:43 PM Tony Dinh <mibodhi at gmail.com> wrote:
>> 
>> Hi Stefan,
>> 
>> On Sat, Sep 3, 2022 at 3:44 AM Stefan Roese <sr at denx.de> wrote:
>> >
>> > Hi Tony,
>> >
>> > On 03.09.22 11:44, Tony Dinh wrote:
>> > > Hi Stefan,
>> > >
>> > > On Thu, Sep 1, 2022 at 11:25 PM Stefan Roese <sr at denx.de> wrote:
>> > >>
>> > >> Add timer_get_boot_us() to support boards, that have CONFIG_BOOTSTAGE
>> > >> enabled, like pogo_v4.
>> > >>
>> > >> Signed-off-by: Stefan Roese <sr at denx.de>
>> > >> ---
>> > >> v2:
>> > >> - Change timer_get_boot_us() to use the timer_early functions
>> > >> - Remove #if CONFIG_IS_ENABLED(BOOTSTAGE)
>> > >>
>> > >> Simon, I'm currently looking into this timer_get_boot_us() to using
>> > >> timer_early_get_count() etc consolidation.
>> > >
>> > > Indeed, as you've mentioned above, I think timer_early_get_count() and
>> > > timer_early_get_rate() do need to take into consideration  what the
>> > > input_clock_type is for Kirkwood boards with CONFIG_BOOTSTAGE such as
>> > > the Pogo V4.
>> > >
>> > > I'm seeing on the Pogo V4 test, the timer command reports time about 6
>> > > times slower than it should. It does seem to jive with the fact that
>> > > the Pogo V4 CONFIG_SYS_TCLK is 166Mhz, versus MVEBU 25MHz clock rate.
>> >
>> > Ah, I've missing updating the early functions to also differentiate
>> > between fixed clocks and TCLK timer.
>> >
>> > Please give the attached patch a try - should be applied on top of this
>> > latest patchset.
>> 
> 
> That looks promising, but I think we are still missing something.
> After applying the attached patch, I ran the test again and it behaved
> the same way (clock rate 6 times slower). So I did another test.
> 
> --- Test 1
> Pogo_V4> timer start; sleep 60; timer get; sleep 30; timer get
> 60.000
> 90.000
> 
> So apparently the sleep cmd has reset the correct clock rate.
> 
> --- Test 2
> 
> Pogo_V4> timer start; sleep 30; timer get; sleep 30; timer get
> 30.000
> 60.000
> 
> And then wait for 30 seconds, do another "timer get" (I expected to
> see about 90 to 95 seconds).
> 
> Pogo_V4> timer get
> 66.237

I've did the same test and can confirm it. But I think that is
a drawback from the timer command. With a 32bit timer and a
TCLK of 166MHz (on my board), the timer will wrap around about
every 26s. So if you don't do a get_timer() call for that whole
period, the overflow won't be counted in timer_conv_64().

For the sleep command it works correctly, because it will
poll get_timer() every 100us.

In summary, you cannot expect a "timer get" on the console
to work if you cannot make sure, the you call it at least
every "2^32 / TCLK" seconds.

-michael