Re: [PATCH 10/15] static_call: Add basic static call infrastructure

[Josh Poimboeuf]

On Wed, Oct 02, 2019 at 03:54:17PM +0200, Peter Zijlstra wrote:
> On Fri, Jun 07, 2019 at 10:28:51AM +0200, Peter Zijlstra wrote:
> > On Thu, Jun 06, 2019 at 10:44:23PM +0000, Nadav Amit wrote:
> > > > + * Usage example:
> > > > + *
> > > > + *   # Start with the following functions (with identical prototypes):
> > > > + *   int func_a(int arg1, int arg2);
> > > > + *   int func_b(int arg1, int arg2);
> > > > + *
> > > > + *   # Define a 'my_key' reference, associated with func_a() by default
> > > > + *   DEFINE_STATIC_CALL(my_key, func_a);
> > > > + *
> > > > + *   # Call func_a()
> > > > + *   static_call(my_key, arg1, arg2);
> > > > + *
> > > > + *   # Update 'my_key' to point to func_b()
> > > > + *   static_call_update(my_key, func_b);
> > > > + *
> > > > + *   # Call func_b()
> > > > + *   static_call(my_key, arg1, arg2);
> > > 
> > > I think that this calling interface is not very intuitive.
> > 
> > Yeah, it is somewhat unfortunate..
> > 
> > > I understand that
> > > the macros/objtool cannot allow the calling interface to be completely
> > > transparent (as compiler plugin could). But, can the macros be used to
> > > paste the key with the âstatic_callâ? I think that having something like:
> > > 
> > >   static_call__func(arg1, arg2)
> > > 
> > > Is more readable than
> > > 
> > >   static_call(func, arg1, arg2)
> > 
> > Doesn't really make it much better for me; I think I'd prefer to switch
> > to the GCC plugin scheme over this.  ISTR there being some propotypes
> > there, but I couldn't quickly locate them.
> 
> How about something like:
> 
> 	static_call(key)(arg1, arg2);
> 
> which is very close to the regular indirect call syntax.

Looks ok to me.

> Furthermore, how about we put the trampolines in .static_call.text
> instead of relying on prefixes?

Yeah, that would probably be better.

> Also, I think I can shrink static_call_key by half:
> 
>  - we can do away with static_call_key::tramp; there are only two usage
>    sites:
> 
>      o __static_call_update, the static_call() macro can provide the
>        address of STATIC_CALL_TRAMP(key) directly
> 
>      o static_call_add_module(), has two cases:
> 
>        * the trampoline is from outside the module; in this case
>          it will already have been updated by a previous call to
> 	 __static_call_update.
>        * the trampoline is from inside the module; in this case
>          it will have the default value and it doesn't need an
> 	 update.
> 
>        so in no case does static_call_add_module() need to modify a
>        trampoline.

Sounds plausible.

>   - we can change static_call_key::site_mods into a single next pointer,
>     just like jump_label's static_key.

Yep.

> But so far all the schemes I've come up with require 'key' to be a name,
> it cannot be an actual 'struct static_call_key *' value. And therefore
> usage from within structures isn't allowed.

Is that something we need?  At least we were able to work around this
limitation with tracepoints' usage of static calls.  But I could see how
it could be useful.

One way to solve that would be a completely different implementation:
have a global trampoline which detects the call site of the caller,
associates it with the given key, schedules some work to patch the call
site later, and then jumps to key->func.  So the first call would
trigger the patching.

Then we might not even need objtool :-)  But it might be tricky to pull
off.

-- 
Josh