Le 13/12/2021 à 18:15, Steven Rostedt a écrit :
> On Mon, 13 Dec 2021 14:39:15 +0000
> Christophe Leroy <christophe.leroy@xxxxxxxxxx> wrote:
>
>>> Note, you can implement this first, (I looked over the patches and they
>>> seem fine) and then update both ppc64 and ppc32 to implement
>>> DYNAMIC_FTRACE_WITH_ARGS.
>>>
>>
>> I tried to activate DYNAMIC_FTRACE_WITH_ARGS on PPC32.
>>
>> I copied into powerpc the changes from 5740a7c71ab6 ("s390/ftrace: add
>> HAVE_DYNAMIC_FTRACE_WITH_ARGS support")
>>
>> Ftrace selftests tell "Testing tracer function_graph: FAILED!".
>>
>> Is there anything else to do ?
>
> Yes. Because BPF is now hooking into the function callbacks, it causes
> issues with function graph tracer. So what we did was to have function
> graph tracing to now use the function tracer callback as well (this allows
> both the BPF direct trampolines to work with function graph tracer).
>
> As it requires DYNAMIC_FTRACE_WITH_ARGS, and x86 was the only one to
> support that for now, I decided to make all the archs change function graph
> tracing when they implement DYNAMIC_FTRACE_WITH_ARGS too. (It is becoming a
> pain to have too many variants of function tracing between the archs).
>
> The change that did this for x86 was:
>
> 0c0593b45c9b4 ("x86/ftrace: Make function graph use ftrace directly")
>
> This actually simplifies the function graph tracer, as you no longer need
> it's own entry trampoline (still need the trampoline for the return of the
> function).
>
> What you need to do is:
>
> In your arch/*/include/asm/ftrace.h add:
>
> struct ftrace_ops;
>
> #define ftrace_graph_func ftrace_graph_func
> void ftrace_graph_func(unsigned long ip, unsigned long parent_ip,
> struct ftrace_ops *op, struct ftrace_regs *fregs);
>
>
> Where ftrace_graph_func() is now what is called for the function graph
> tracer, directly from the ftrace callbacks (no longer a secondary
> trampoline).
>
> Define the ftrace_graph_func() to be something like:
>
> void ftrace_graph_func(unsigned long ip, unsigned long parent_ip,
> struct ftrace_ops *op, struct ftrace_regs *fregs)
> {
> struct pt_regs *regs = &fregs->regs;
> unsigned long *stack = (unsigned long *)kernel_stack_pointer(regs);
>
> prepare_ftrace_return(ip, (unsigned long *)stack, 0);
> }
>
> This is called by the function tracer code. But because with
> DYNAMIC_FTRACE_WITH_ARGS, we have access to the argument register, we should
> also have access to the link register and the stack. Then you can use that
> to modify the stack and or link register to jump to the the return
> trampoline.
>
> This should all work with powerpc (both 64 and 32) but if it does not, let
> me know. I'm happy to help out.
>
Thanks, I will try that.
I can't find ftrace_graph_func() in s390. Does it mean that s390 doesn't
have a working function tracer anymore ?
I see your commit 0c0593b45c9b4 ("x86/ftrace: Make function graph use
ftrace directly") is dated 8 Oct 2021 while 5740a7c71ab6 ("s390/ftrace:
add HAVE_DYNAMIC_FTRACE_WITH_ARGS support") is 4 Oct 2021.
Christophe
