Re: [PATCH v2 1/2] tracing: fprobe: optimization for entry only case

[Menglong Dong <menglong.dong@linux.dev>]

On 2025/10/14 22:51, Masami Hiramatsu wrote:
> Hi Menglong,
> 
> I remember why I haven't implement this.
> 
> On Fri, 10 Oct 2025 11:38:46 +0800
> Menglong Dong <menglong8.dong@gmail.com> wrote:
> 
> > +
> > +static struct ftrace_ops fprobe_ftrace_ops = {
> > +	.func	= fprobe_ftrace_entry,
> > +	.flags	= FTRACE_OPS_FL_SAVE_REGS,
> 
> Actually, this flag is the problem. This can fail fprobe on architecture
> which does not support CONFIG_DYNAMIC_FTRACE_WITH_REGS (e.g. arm64, riscv)
> 
>  * SAVE_REGS - The ftrace_ops wants regs saved at each function called
>  *            and passed to the callback. If this flag is set, but the
>  *            architecture does not support passing regs
>  *            (CONFIG_DYNAMIC_FTRACE_WITH_REGS is not defined), then the
>  *            ftrace_ops will fail to register, unless the next flag
>  *            is set.
> 
> fgraph has a special entry code for saving ftrace_regs.
> So at least we need to fail back to fgraph if arch does not
> support CONFIG_DYNAMIC_FTRACE_WITH_REGS.

Ah, I have be working on x86_64 and didn't notice it. You are
right, we do need fallback if CONFIG_DYNAMIC_FTRACE_WITH_REGS
not supported. I'll send a V4 later.

BTW, is the FTRACE_OPS_FL_SAVE_REGS necessary here? I guess
not all architectures save the function argument regs in
fentry_caller() like x86_64, that's why we need it here :/

Thanks!
Menglong Dong

> 
> Thank you,
> 
>