Re: [PATCH v1] arch_topology: Adjust initial CPU capacities with current freq

[Robin Murphy <robin.murphy@arm.com>]

On 2020-01-11 2:51 am, JeffyChen wrote:
> Hi Robin,
> 
> Thanks for the clarification :)
> 
> On 01/10/2020 08:28 PM, Robin Murphy wrote:
>> On 2020-01-10 12:01 pm, Dietmar Eggemann wrote:
>>> On 10/01/2020 12:37, Sudeep Holla wrote:
>>>> On Thu, Jan 09, 2020 at 03:52:14PM +0800, Jeffy Chen wrote:
>>>>> The CPU freqs are not supposed to change before cpufreq policies
>>>>> properly registered, meaning that they should be used to calculate the
>>>>> initial CPU capacities.
>>>>>
>>>>> Doing this helps choosing the best CPU during early boot, especially
>>>>> for the initramfs decompressing.
>>>>>
>>>>> Signed-off-by: Jeffy Chen <jeffy.chen@rock-chips.com>
>>>>
>>>> [...]
>>>>
>>>>> @@ -146,10 +153,15 @@ bool __init topology_parse_cpu_capacity(struct
>>>>> device_node *cpu_node, int cpu)
>>>>>                   return false;
>>>>>               }
>>>>>           }
>>>>> -        capacity_scale = max(cpu_capacity, capacity_scale);
>>>>>           raw_capacity[cpu] = cpu_capacity;
>>>>>           pr_debug("cpu_capacity: %pOF cpu_capacity=%u (raw)\n",
>>>>>               cpu_node, raw_capacity[cpu]);
>>>>> +
>>>>> +        cpu_clk = of_clk_get(cpu_node, 0);
>>>>> +        if (!PTR_ERR_OR_ZERO(cpu_clk))
>>>>> +            per_cpu(max_freq, cpu) = clk_get_rate(cpu_clk) / 1000;
>>>>> +
>>>>> +        clk_put(cpu_clk);
>>>>
>>>> I don't like to assume DVFS to be supplied only using 'clk'. So NACK!
>>>> We have other non-clk mechanism for CPU DVFS and this needs to simply
>>>> use cpufreq APIs to get frequency value if required.
>>>
>>> To support this, it's failing on my Arm64 Juno board.
>>>
>>> ...
>>> [    0.084858] CPU1 cpu_clk=-517
>>> [    0.087961] CPU2 cpu_clk=-517
>>> [    0.091005] CPU0 cpu_clk=-517
>>> [    0.094121] CPU3 cpu_clk=-517
>>> [    0.097248] CPU4 cpu_clk=-517
>>> [    0.100415] CPU5 cpu_clk=-517
> 
> It there any other way to get the initial cpu capacity for this case?
> 
> Or can we just assuming all the cores running at the same freq here?
> 
>>> ...
>>>
>>> Since you're on a big.LITTLE platform, did you specify
>>> 'capacity-dmips-mhz' for CPUs to be able to distinguish big and little
>>> CPUs before CPUfreq kicks in?
>>
>> Indeed, and that's the "problem" - the capacities are there, but with
>> the broken firmware the kernel starts with the little (boot) cluster
>> clocked at either 400 or 200MHz, but the big cluster at just 12MHz. At
>> that speed, a full distro config can take about 3 minutes to get to the
>> point of loading cpufreq as a module, and I've seen at least one distro
>> reverting 97df3aa76b4a to 'fix' the symptom :(
> 
> Right, for the big cluster, the bootrom(maskrom) will init the clock to 
> 24MHz, and if the bootloader(u-boot for example) doesn't bump it, it 
> would become 12MHz after kernel initialized the whole clk tree.
> 
> And in rockchip's BSP 4.4 kernel, there are hacks to bump it to 
> 800MHz(higher freq might require regulator changing) in clk tree 
> initialization, the BSP u-boot also added that recently.
> 
> The chromeos's coreboot looks fine, but upstream u-boot seems missing 
> that part too, i'll try to send a patch for that :)

Actually, last time I looked both the BSP U-Boot and mainline do contain 
equivalent code to initialise both PLLs to (IIRC) 600MHz and apparently 
adjust a couple of other things set by the maskrom. The trap is that 
mainline does it in the SPL - thus the unfortunately common combination 
of using the upstream main stage with the miniloader ends up missing out 
that step entirely. In comparison, I'm now using the full upstream 
TPL/SPL flow on my RK3399 board (NanoPC-T4) and even a full generic 
distro kernel is acceptably quick:

[    2.315378] Trying to unpack rootfs image as initramfs...
[    2.781747] Freeing initrd memory: 7316K
...
[    4.239990] Freeing unused kernel memory: 1984K
[    4.247829] Run /init as init process

Robin.