[RFC PATCH] powerpc: Investigate static_call concept

Peter Zijlstra peterz at infradead.org
Sat Aug 28 00:18:47 AEST 2021


On Fri, Aug 27, 2021 at 09:45:37AM +0000, Christophe Leroy wrote:
> This RFC is to validate the concept of static_call on powerpc.
> 
> Highly copied from x86.
> 
> It replaces ppc_md.get_irq() which is called at every IRQ, by
> a static call.

The code looks saner, but does it actually improve performance? I'm
thinking the double branch also isn't free.

> When updating the call, we just replace the instruction at the
> trampoline address by a relative jump to the function.
> 
> For the time being, the case of out-of-range functions is not handled.

The paranoid in me would've made it:

	BUG_ON(patch_branch(...));

just to make sure to notice the target not fitting. Ohh, patch_branch()
doesn't return the create_branch() error, perhaps that wants to be
fixed?

Did you see the arm64 variant that deals with out-of-range functions in
their trampoline?

  https://lore.kernel.org/linux-arm-kernel/20201120082103.4840-1-ardb@kernel.org/

Not exactly 'nice' but supposedly that works.

> +#define ARCH_DEFINE_STATIC_CALL_NULL_TRAMP(name)			\
> +	asm(".pushsection .text, \"ax\"				\n"	\
> +	    ".align 4						\n"	\
> +	    ".globl " STATIC_CALL_TRAMP_STR(name) "		\n"	\
> +	    STATIC_CALL_TRAMP_STR(name) ":			\n"	\
> +	    "	blr						\n"	\
> +	    ".type " STATIC_CALL_TRAMP_STR(name) ", @function	\n"	\
> +	    ".size " STATIC_CALL_TRAMP_STR(name) ", . - " STATIC_CALL_TRAMP_STR(name) " \n" \
> +	    ".popsection					\n")
> +

Since you support CALL_NULL_TRAMP, your patch function below:

> +void arch_static_call_transform(void *site, void *tramp, void *func, bool tail)
> +{
> +	mutex_lock(&text_mutex);
> +
> +	if (tramp)
> +		patch_branch(tramp, (unsigned long)func, 0);
> +
> +	mutex_unlock(&text_mutex);
> +}
> +EXPORT_SYMBOL_GPL(arch_static_call_transform);

Ought to patch in "blr" when !func to be consistent :-)

I'm thinking that your core kernel text all fits in the native range and
only modules need out-of-range ?


More information about the Linuxppc-dev mailing list