Message-ID: <alpine.DEB.2.10.1409030103090.3333@nanos>
Date:	Wed, 3 Sep 2014 02:03:43 +0200 (CEST)
From:	Thomas Gleixner <tglx@...utronix.de>
To:	Daniel Thompson <daniel.thompson@...aro.org>
cc:	Russell King <linux@....linux.org.uk>,
	linaro-kernel@...ts.linaro.org,
	Catalin Marinas <catalin.marinas@....com>, patches@...aro.org,
	kgdb-bugreport@...ts.sourceforge.net,
	Nicolas Pitre <nico@...aro.org>, linux-kernel@...r.kernel.org,
	Frederic Weisbecker <fweisbec@...il.com>,
	Anton Vorontsov <anton.vorontsov@...aro.org>,
	Ben Dooks <ben.dooks@...ethink.co.uk>,
	Fabio Estevam <festevam@...il.com>,
	Colin Cross <ccross@...roid.com>, kernel-team@...roid.com,
	Dave Martin <Dave.Martin@....com>,
	linux-arm-kernel@...ts.infradead.org
Subject: Re: [PATCH v11 01/19] arm: fiq: Add callbacks to manage FIQ
 routings

On Tue, 2 Sep 2014, Daniel Thompson wrote:

> Currently enable_fiq/disable_fiq use a simple offset to convert an IRQ
> virq into a FIQ virq. This is too inflexible for multi-platform kernels
> and makes runtime error checking impossible.

You are missing to explain WHY it is too inflexible ....
 
> We solve this by introducing a flexible mapping that allows interrupt

Flexible in that context reads like random semantics. And that's
definitely nothing we want to see in the context of interrupts/fiqs

> controllers that support FIQ to register those mappings. This, in turn,
> makes it much possible for drivers in DT kernels to install FIQ handlers
> without knowing anything about the interrupt controller.

> +#include <linux/irq.h>
>  #include <asm/ptrace.h>
>  
> +struct fiq_chip {
> +	void (*fiq_enable)(struct irq_data *data);
> +	void (*fiq_disable)(struct irq_data *data);
> +};

So you define an ARM specific extension to the irq infrastructure
which is completely ignoring that there might be other architectures
which have the same issue?

Aside of that you define that structure w/o documenting in any way
what the semantics of the callbacks are and how struct irq_data is
supposed to be (ab)used there.

> +struct fiq_data {
> +	struct fiq_chip *fiq_chip;
> +	struct irq_data *irq_data;
> +};

Nice. Another completely undocumented data structure.

>  static unsigned long no_fiq_insn;
> +static int fiq_start = -1;
> +static RADIX_TREE(fiq_data_tree, GFP_KERNEL);

Along with a radix tree, which again lacks any form of documentation.

> -static int fiq_start;
> +static struct fiq_data *lookup_fiq_data(int fiq)
> +{
> +	struct fiq_data *data;
> +
> +	rcu_read_lock();
> +	data = radix_tree_lookup(&fiq_data_tree, fiq);
> +	rcu_read_unlock();
> +
> +	return data;

That makes a lot of sense. NOT!

What kind of protection is that rcu_read_lock/unlock pair providing
for the return value? Just that it has been valid at lookup time?

What protects against concurrent calls to the various usage sites
which shine by the lack of serialization ...

>  void enable_fiq(int fiq)
>  {
> +	struct fiq_data *data = lookup_fiq_data(fiq);
> +
> +	if (data) {
> +		if (data->fiq_chip->fiq_enable)
> +			data->fiq_chip->fiq_enable(data->irq_data);
> +		enable_irq(fiq);

This is ass backwards, really.

The FIQ target is a property of the interrupt line right?

So why the heck do you need a separate chip/data representation
including a radix tree and whatever? Just because you can?

What is wrong to tell the irq core code via a flag that a particular
interrupt should be targeted at FIQ instead of a regular interrupt?

Nothing as far as I can tell. And that would even allow to retarget an
already enabled interrupt to FIQ or vice versa.

So all your patch does is trying to (ab)use an existing interface
which was created in the last millenium for totally different reasons
and in a totally different context.

But you fail completely to understand what that interface is doing:

    It merily maps the Linux virq to a FIQ via an offset.

While I understand that this is not workable on multiplatform kernels,
I really cannot understand your attempt to change this by adding
complex functionality to a primitive offset mapping function.

You just add some randomly defined new chip/data combination which
abuses the irq_data structure under the hood instead of properly
integrating that irq chip property into irqchip/irqdata itself.

So instead of thinking about the properties of the hardware and
questioning the existing mechanism, you just bolt some crap onto it
and inflict a gazillion of pointless modifications to the irq chip
drivers without any value.

So all it needs is

 1) Adding a property flag to the irq chip which signals that it
    supports FIQ/NMI targets

 2) Adding a interface which allows to flag a particular interrupt
    line targeted to either the regular or the FIQ/NMI which consults
    the flag added by #1

 3) Let the chip driver deal with the target flag at a minimal
    intrusive level w/o registering racy callbacks designed in hell
    and imposing arch specific data structures, storage etc. for no
    value

Thanks,

	tglx

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

Powered by blists - more mailing lists