Re: [PATCH v3] ARM64: Dynamically allocate cpumasks and increase supported CPUs to 512

[Catalin Marinas]

On Thu, Mar 14, 2024 at 01:28:40PM +0100, Marek Szyprowski wrote:
> On 14.03.2024 09:39, Catalin Marinas wrote:
> > On Wed, Mar 13, 2024 at 05:13:33PM +0000, Russell King wrote:
> >> So, I wonder whether what you're seeing is a latent bug which is
> >> being tickled by the presence of the CPU masks being off-stack
> >> changing the kernel timing.
> >>
> >> I would suggest the printk debug approach may help here to see when
> >> the OPPs are begun to be parsed, when they're created etc and their
> >> timing relationship to being used. Given the suspicion, it's possible
> >> that the mere addition of printk() may "fix" the problem, which again
> >> would be another semi-useful data point.
> > It might be an init order problem. Passing "initcall_debug" on the
> > cmdline might help a bit.
> >
> > It would also be useful in dev_pm_opp_set_config(), in the WARN_ON
> > block, to print opp_table->opp_list.next to get an idea whether it looks
> > like a valid pointer or memory corruption.
> 
> I've finally found some time to do the step-by-step printk-based 
> debugging of this issue and finally found what's broken!
> 
> Here is the fix:
> 
> diff --git a/drivers/cpufreq/cpufreq-dt.c b/drivers/cpufreq/cpufreq-dt.c
> index 8bd6e5e8f121..2d83bbc65dd0 100644
> --- a/drivers/cpufreq/cpufreq-dt.c
> +++ b/drivers/cpufreq/cpufreq-dt.c
> @@ -208,7 +208,7 @@ static int dt_cpufreq_early_init(struct device *dev, 
> int cpu)
>          if (!priv)
>                  return -ENOMEM;
> 
> -       if (!alloc_cpumask_var(&priv->cpus, GFP_KERNEL))
> +       if (!zalloc_cpumask_var(&priv->cpus, GFP_KERNEL))
>                  return -ENOMEM;
> 
>          cpumask_set_cpu(cpu, priv->cpus);
> 
> 
> It is really surprising that this didn't blow up for anyone else so 
> far... This means that the $subject patch is fine.
> 
> I will send a proper patch fixing this issue in a few minutes.

Nice. Many thanks for tracking this down. I'll revert the revert of the
CPUMASK_OFFSTACK in the second part of the merging window (I already
sent the pull request).

-- 
Catalin