[RFC PATCH 2/4] cpu: advertise CPU features over udev in a generic way

[Dave Martin]

On Thu, Nov 07, 2013 at 06:17:35PM +0100, Ard Biesheuvel wrote:
> This patch implements a generic modalias 'cpu:feature:...' which
> enables CPU feature flag based module loading in a generic way.
> All the arch needs to do is enable CONFIG_ARCH_HAS_CPU_AUTOPROBE
> and export a u32 called 'cpu_features'. (What each bit actually
> means is irrelevant on this level.)

There seems to be an assumption here that a module is either a pure CPU
accelerator, or it is completely independent of CPU features.

I'm not sure that this is true.  A CPU feature isn't a "device".
Rather, it's a property of code (which might be a driver for something
different -- maybe we have a hardware crypto accelerator where key
scheduling must be done in software.  It's still a driver for the
crypto engine, but we might have different implementations of the key
scheduling, based on the CPU features avaiable).

It's also not obvious why we should blindly load all CPU-feature-
dependent helper modules on bgoot, regardless of whether the module(s)
that use them are being loaded.

Maybe the amount of CPU feature dependent code is small enough that
we don't really care about such subtleties, though.


It's also not clear how different optimised modules for the same
thing would get prioritised.  Suppose there we have v5E and NEON
optimised AES helper modules?  Both those CPU features are avaiable,
but which module should we load?

If all candidate modules get loaded, which one actually gets used?
Does the load order matter?

> Signed-off-by: Ard Biesheuvel <ard.biesheuvel at linaro.org>
> ---
>  drivers/base/cpu.c | 24 ++++++++++++++++++++++++
>  1 file changed, 24 insertions(+)
> 
> diff --git a/drivers/base/cpu.c b/drivers/base/cpu.c
> index 49c6f4b..a661d31 100644
> --- a/drivers/base/cpu.c
> +++ b/drivers/base/cpu.c
> @@ -276,6 +276,30 @@ static void cpu_device_release(struct device *dev)
>  }
>  
>  #ifdef CONFIG_ARCH_HAS_CPU_AUTOPROBE
> +ssize_t print_cpu_modalias(struct device *dev,
> +			   struct device_attribute *attr,
> +			   char *buf)
> +{
> +	extern u32 __weak cpu_features;

Why is this __weak?  Surely CONFIG_ARCH_HAS_CPU_AUTOPROBE=y makes no
sense if the arch code does not define either cpu_features or
arch_print_cpu_modalias()?  The build should be made to fail in that
case...

> +	ssize_t n;
> +	int i;
> +	u32 f;
> +
> +	/*
> +	 * With 32 features maximum (taking 3 bytes each to print), we don't
> +	 * need to worry about overrunning the PAGE_SIZE sized buffer.
> +	 */
> +	n = sprintf(buf, "cpu:feature:");
> +	for (f = cpu_features, i = 0; f; f >>= 1, i++)
> +		if (f & 1)
> +			n += sprintf(&buf[n], ",%02X", i);

Why can't this overflow buf?

modalias matching is pretty much based on string matching, so I wonder
whether we could use the human-readable feature strings instead.
Those are already a stable ABI.  Relying on numbers unnecessarily
encrypts the udev/modprobe config.

Otherwise, "%02X" seems to place an arbitrary limit of 256 features.
I'm not sure that padding these numbers to a particular width is
advantageous for the parser.

> +	buf[n++] = '\n';
> +	return n;
> +}
> +
> +ssize_t __attribute__((weak, alias("print_cpu_modalias")))
> +arch_print_cpu_modalias(struct device *, struct device_attribute *, char *);
> +

If an implementation of arch_print_cpu_modalias() is linked with this,
won't that result in the print_cpu_modalias() defined here just being
included as dead code?

i.e., we knowingly link into the kernel some code that the build-time
configuration tells us is dead.

Maybe I'm misunderstanding things here, but I think this weak-symbol
stuff is mainly useful when shipping a binary blob in which people can
override certain symbols at link time.

We build vmlinux in one go, so I'm not sure that's appropriate here (?)

>  static int cpu_uevent(struct device *dev, struct kobj_uevent_env *env)
>  {
>  	char *buf = kzalloc(PAGE_SIZE, GFP_KERNEL);
> -- 
> 1.8.3.2
> 
> 
> _______________________________________________
> linux-arm-kernel mailing list
> linux-arm-kernel at lists.infradead.org
> http://lists.infradead.org/mailman/listinfo/linux-arm-kernel