Re: [PATCH] ftrace,kcfi: Separate ftrace_stub() and ftrace_stub_graph()

[Peter Zijlstra <peterz@infradead.org>]

On Tue, Oct 18, 2022 at 10:21:00AM -0400, Steven Rostedt wrote:
> On Tue, 18 Oct 2022 14:35:15 +0200
> Peter Zijlstra <peterz@infradead.org> wrote:
> 
> > Different function signatures means they needs to be different
> > functions; otherwise CFI gets upset.
> 
> This is due to this commit:
> 
> commit b83b43ffc6e4b514ca034a0fbdee01322e2f7022
> Author: Steven Rostedt (VMware) <rostedt@goodmis.org>
> Date:   Tue Oct 15 09:00:55 2019 -0400
> 
>     fgraph: Fix function type mismatches of ftrace_graph_return using ftrace_stub
>     
>     The C compiler is allowing more checks to make sure that function pointers
>     are assigned to the correct prototype function. Unfortunately, the function
>     graph tracer uses a special name with its assigned ftrace_graph_return
>     function pointer that maps to a stub function used by the function tracer
>     (ftrace_stub). The ftrace_graph_return variable is compared to the
>     ftrace_stub in some archs to know if the function graph tracer is enabled or
>     not. This means we can not just simply create a new function stub that
>     compares it without modifying all the archs.
>     
>     Instead, have the linker script create a function_graph_stub that maps to
>     ftrace_stub, and this way we can define the prototype for it to match the
>     prototype of ftrace_graph_return, and make the compiler checks all happy!
> 
> 
> Perhaps its time to just modify all the archs and get rid of that hack.

Ideally yes, but given kCFI is in Linus' tree now, I didn't feel like
fixing up all archs in a hurry.

Mark mentioned that most archs can probably move to a common empty C
function for each stub.