On Fri, 22 Jul 2022 13:08:11 +0200 Jiri Olsa <jolsa@xxxxxxxxxx> wrote: > hi, > we recently hit bug where ftrace update raced with bpf_dispatcher_update > that calls directly bpf_arch_text_poke [1]. > > The bpf_dispatcher_update creates special trampoline and attaches it to > designated bpf_dispatcher_xdp_func function, which is run for xdp bpf > programs from several places. > > After discussion with Alexei we'd rather keep this code update out of > ftrace, because it's already slow and had troubles with CI because of > that. > > This patch is presenting the idea to allow some functions not to be > managed by ftrace by marking them with NOFTRACE_SYMBOL macro and > such symbols will not be added to ftrace_pages on the kernel start. NACK on any generic way to hide mcount/fentry functions from ftrace. There's a lot of infrastructure to see what functions are being modified, as the user should know. (See tracefs/enabled_functions). There's no need for a generic way to hide functions. Once that happens, it will grow and then it will be more confusing to why some functions are traced while others are not. > > Please note it's RFC so I did not bother with some fast search for > is_noftrace_function function. > > Perhaps we could use existing NOKPROBE_SYMBOL for this? but I'm not > sure you can (or want) to run function trace on such symbols. I trace those functions all the time. Yes, I want to continue doing so. -- Steve
