Re: [PATCH v4 14/18] static_call: Add static_cond_call()

[Peter Zijlstra]

On Wed, May 06, 2020 at 12:24:55PM -0500, Josh Poimboeuf wrote:
> On Fri, May 01, 2020 at 10:29:03PM +0200, Peter Zijlstra wrote:
> > +++ b/arch/x86/include/asm/static_call.h
> > @@ -30,4 +30,14 @@
> >  	    ".size " STATIC_CALL_TRAMP_STR(name) ", . - " STATIC_CALL_TRAMP_STR(name) " \n" \
> >  	    ".popsection					\n")
> >  
> > +#define ARCH_DEFINE_STATIC_CALL_RETTRAMP(name)				\
> > +	asm(".pushsection .static_call.text, \"ax\"		\n"	\
> > +	    ".align 4						\n"	\
> > +	    ".globl " STATIC_CALL_TRAMP_STR(name) "		\n"	\
> > +	    STATIC_CALL_TRAMP_STR(name) ":			\n"	\
> > +	    "	ret; nop; nop; nop; nop;			\n"	\
> > +	    ".type " STATIC_CALL_TRAMP_STR(name) ", @function	\n"	\
> > +	    ".size " STATIC_CALL_TRAMP_STR(name) ", . - " STATIC_CALL_TRAMP_STR(name) " \n" \
> > +	    ".popsection					\n")
> > +
> 
> The boilerplate in these two trampoline macros is identical except for
> the actual instructions, maybe there could be a shared
> __ARCH_DEFINE_STATIC_CALL_TRAMP(name, insns) macro which does most of
> the dirty work.

I'm afraid that'll just make it less readable :/

> >  #endif /* _ASM_STATIC_CALL_H */
> > --- a/arch/x86/kernel/static_call.c
> > +++ b/arch/x86/kernel/static_call.c
> > @@ -4,19 +4,41 @@
> >  #include <linux/bug.h>
> >  #include <asm/text-patching.h>
> >  
> > -static void __static_call_transform(void *insn, u8 opcode, void *func)
> > +enum insn_type {
> > +	call = 0, /* site call */
> > +	nop = 1,  /* site cond-call */
> > +	jmp = 2,  /* tramp / site tail-call */
> > +	ret = 3,  /* tramp / site cond-tail-call */
> > +};
> 
> The lowercase enums threw me for a loop, I thought they were variables a
> few times.  Starting a new enum trend? :-)

I can UPPERCASE them I suppose, not sure where this came from.

> >  void arch_static_call_transform(void *site, void *tramp, void *func)
> > @@ -24,10 +46,10 @@ void arch_static_call_transform(void *si
> >  	mutex_lock(&text_mutex);
> >  
> >  	if (tramp)
> > -		__static_call_transform(tramp, JMP32_INSN_OPCODE, func);
> > +		__static_call_transform(tramp, jmp + !func, func);
> >  
> >  	if (IS_ENABLED(CONFIG_HAVE_STATIC_CALL_INLINE) && site)
> > -		__static_call_transform(site, CALL_INSN_OPCODE, func);
> > +		__static_call_transform(site, !func, func);
> 
> Clever enum math, but probably more robust to be ignorant of the values:
> 
> 	if (tramp)
> 		__static_call_transform(tramp, func ? jmp : ret, func);
> 
>   	if (IS_ENABLED(CONFIG_HAVE_STATIC_CALL_INLINE) && site)
> 		__static_call_transform(site, func ? call : nop, func);
> 

That is more readable, and I checked, GCC is clever enough to not
actually emit branches for that, so w00t.

> > +++ b/include/linux/static_call.h
> > @@ -16,7 +16,9 @@
> >   *
> >   *   DECLARE_STATIC_CALL(name, func);
> >   *   DEFINE_STATIC_CALL(name, func);
> > + *   DEFINE_STATIC_COND_CALL(name, typename);
> >   *   static_call(name)(args...);
> > + *   static_cond_call(name)(args...)
> >   *   static_call_update(name, func);
> 
> Missing semicolon, also an updated description/example would be useful.

Yes, I already promised Rasmus more documentation.

> On that note, what do you think about tweaking the naming from
> 
>   DEFINE_STATIC_COND_CALL(name, typename);
>   static_cond_call(name)(args...);
> 
> to
> 
>   DEFINE_STATIC_CALL_NO_FUNC(name, typename);
>   static_call_if_func(name)(args...);
> 
> ?
> 
> Seems clearer to me.  They're still STATIC_CALLs, so it seems logical to
> keep those two words together.  And NO_FUNC clarifies the initialized
> value.
> 
> Similarly RETTRAMP could be ARCH_DEFINE_STATIC_CALL_NO_FUNC.

What can I say, I'm sorta used to the old naming by now, but sure, any
other opinions before I edit things?