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

[Robin Murphy]

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@xxxxxxxxxxxxxx>
> > > >
> > > > [...]
> > > >
> > > > > @@ -146,10 +153,15 @@ bool __init topology_parse_cpu_capacity(struct
> > > > > device_node *cpu_node, int cpu)
> > > > > ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ return false;
> > > > > ÂÂÂÂÂÂÂÂÂÂÂÂÂ }
> > > > > ÂÂÂÂÂÂÂÂÂ }
> > > > > -ÂÂÂÂÂÂÂ capacity_scale = max(cpu_capacity, capacity_scale);
> > > > > ÂÂÂÂÂÂÂÂÂ raw_capacity[cpu] = cpu_capacity;
> > > > > ÂÂÂÂÂÂÂÂÂ pr_debug("cpu_capacity: %pOF cpu_capacity=%u (raw)\n",
> > > > > ÂÂÂÂÂÂÂÂÂÂÂÂÂ cpu_node, raw_capacity[cpu]);
> > > > > +
> > > > > +ÂÂÂÂÂÂÂ cpu_clk = of_clk_get(cpu_node, 0);
> > > > > +ÂÂÂÂÂÂÂ if (!PTR_ERR_OR_ZERO(cpu_clk))
> > > > > +ÂÂÂÂÂÂÂÂÂÂÂ per_cpu(max_freq, cpu) = clk_get_rate(cpu_clk) / 1000;
> > > > > +
> > > > > +ÂÂÂÂÂÂÂ 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.
> > >
> > > ...
> > > [ÂÂÂ 0.084858] CPU1 cpu_clk=-517
> > > [ÂÂÂ 0.087961] CPU2 cpu_clk=-517
> > > [ÂÂÂ 0.091005] CPU0 cpu_clk=-517
> > > [ÂÂÂ 0.094121] CPU3 cpu_clk=-517
> > > [ÂÂÂ 0.097248] CPU4 cpu_clk=-517
> > > [ÂÂÂ 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.