Re: [PATCH v4 03/14] irqchip: mips-gic: Introduce for_each_online_cpu_gic()

[Thomas Gleixner <tglx@linutronix.de>]

On Sat, May 11 2024 at 12:43, Aleksandar Rikalo wrote:
> A few pieces of code in the MIPS GIC driver operate on the GIC local
> register block for each online CPU, accessing each via the GIC's
> other/redirect register block. This patch abstracts the process of
> iterating over online CPUs & configuring the other/redirect region to
> access their registers through a new for_each_online_cpu_gic() macro.
>
> This simplifies users of the new macro slightly, and more importantly
> prepares us for handling multi-cluster systems where the register
> configuration will be done via the CM's GCR_CL_REDIRECT register. By
> abstracting all other/redirect block configuration through this macro,
> and the __gic_with_next_online_cpu() function which backs it, users will
> trivially gain support for multi-cluster when it is implemented in
> __gic_with_next_online_cpu().

Can you please rework the change log according to

  https://www.kernel.org/doc/html/latest/process/maintainer-tip.html#changelog

> Signed-off-by: Paul Burton <paulburton@kernel.org>
> Signed-off-by: Chao-ying Fu <cfu@wavecomp.com>
> Signed-off-by: Dragan Mladjenovic <dragan.mladjenovic@syrmia.com>
> Signed-off-by: Aleksandar Rikalo <aleksandar.rikalo@syrmia.com>
> ---
>  drivers/irqchip/irq-mips-gic.c | 61 +++++++++++++++++++++++++++++-----
>  1 file changed, 52 insertions(+), 9 deletions(-)
>
> diff --git a/drivers/irqchip/irq-mips-gic.c b/drivers/irqchip/irq-mips-gic.c
> index 76253e864f23..9e7182150b5c 100644
> --- a/drivers/irqchip/irq-mips-gic.c
> +++ b/drivers/irqchip/irq-mips-gic.c
> @@ -66,6 +66,52 @@ static struct gic_all_vpes_chip_data {
>  	bool	mask;
>  } gic_all_vpes_chip_data[GIC_NUM_LOCAL_INTRS];
>  
> +static int __gic_with_next_online_cpu(int prev)
> +{
> +	unsigned int cpu;
> +
> +	/* Discover the next online CPU */
> +	cpu = cpumask_next(prev, cpu_online_mask);
> +
> +	/* If there isn't one, we're done */
> +	if (cpu >= nr_cpu_ids)
> +		return cpu;
> +
> +	/*
> +	 * Lock access to the next CPU's GIC local register block.
> +	 *
> +	 * In the single cluster case we simply set GIC_VL_OTHER. The caller
> +	 * holds gic_lock so nothing can clobber the value we write.
> +	 */
> +	write_gic_vl_other(mips_cm_vp_id(cpu));

What unlocks the access? I assume it's magic, but then magic wants to be
documented.

> +
> +	return cpu;
> +}
> +
> +static inline void __lockdep_assert_held(raw_spinlock_t *gic_lock)
> +{
> +	lockdep_assert_held(gic_lock);
> +}

What is exactly the point of this indirection?

> +/**
> + * for_each_online_cpu_gic() - Iterate over online CPUs, access local registers
> + * @cpu: An integer variable to hold the current CPU number
> + * @gic_lock: A pointer to raw spin lock used as a guard
> + *
> + * Iterate over online CPUs & configure the other/redirect register region to
> + * access each CPUs GIC local register block, which can be accessed from the
> + * loop body using read_gic_vo_*() or write_gic_vo_*() accessor functions or
> + * their derivatives.
> + *
> + * The caller must hold gic_lock throughout the loop, such that GIC_VL_OTHER
> + * cannot be clobbered.
> + */
> +#define for_each_online_cpu_gic(cpu, gic_lock)		\
> +	for (__lockdep_assert_held(gic_lock),		\
> +	     (cpu) = __gic_with_next_online_cpu(-1);	\
> +	     (cpu) = __gic_with_next_online_cpu(cpu),	\
> +	     (cpu) < nr_cpu_ids;)

That's broken. It resolves to:

     for (cpu = foo(-1); cpu = foo(cpu), cpu < nr_cpu_ids; )

So on entering the loop:

      cpu = foo(-1);    -> cpu == 0

Now it has to evaluate the loop condition which does:

      cpu = foo(cpu)    -> cpu == 1
      ...

So CPU 0 is skipped unconditionally. No?

Aside of that. Instead of this __lockdep_assert_held() obfuscation you
can simply do:

#define for_each_online_cpu_gic(cpu, gic_lock)		\
	guard(raw_spinlock_irqsave)(gic_lock);		\
	for ((cpu) = __gic_with_next_online_cpu(-1);	\
	     (cpu) < nr_cpu_ids;			\
	     (cpu) = __gic_with_next_online_cpu(cpu);)

which simplifies the callsites even further.

Thanks,

        tglx