On Thu, 17 Aug 2023 10:57:50 +0200 Florent Revest <revest@xxxxxxxxxxxx> wrote: > On Sat, Aug 12, 2023 at 7:37 AM Masami Hiramatsu (Google) > <mhiramat@xxxxxxxxxx> wrote: > > > > diff --git a/kernel/trace/Kconfig b/kernel/trace/Kconfig > > index d56304276318..6fb4ecf8767d 100644 > > --- a/kernel/trace/Kconfig > > +++ b/kernel/trace/Kconfig > > @@ -679,7 +679,6 @@ config FPROBE_EVENTS > > select TRACING > > select PROBE_EVENTS > > select DYNAMIC_EVENTS > > - depends on DYNAMIC_FTRACE_WITH_REGS > > I believe that, in practice, fprobe events still rely on WITH_REGS: > > > diff --git a/kernel/trace/trace_fprobe.c b/kernel/trace/trace_fprobe.c > > index f440c97e050f..94c01dc061ec 100644 > > --- a/kernel/trace/trace_fprobe.c > > +++ b/kernel/trace/trace_fprobe.c > > @@ -327,14 +328,15 @@ static int fentry_dispatcher(struct fprobe *fp, unsigned long entry_ip, > > struct pt_regs *regs = ftrace_get_regs(fregs); > > Because here you require the entry handler needs ftrace_regs that are > full pt_regs. Ah, that is for perf events. Yes, that is the problematic point. Since perf's interfaces are depending on the pt_regs (especially stacktrace) I can not remove this part. This is the next issue to be solved. Maybe we can use partial pt_regs for stack tracing, so we can swap the order of the patches to introduce ftrace_partial_regs() before this and use it for perf event. > > > int ret = 0; > > > > + if (trace_probe_test_flag(&tf->tp, TP_FLAG_TRACE)) > > + fentry_trace_func(tf, entry_ip, fregs); > > + > > +#ifdef CONFIG_PERF_EVENTS > > if (!regs) > > return 0; > > > > - if (trace_probe_test_flag(&tf->tp, TP_FLAG_TRACE)) > > - fentry_trace_func(tf, entry_ip, regs); > > -#ifdef CONFIG_PERF_EVENTS > > if (trace_probe_test_flag(&tf->tp, TP_FLAG_PROFILE)) > > - ret = fentry_perf_func(tf, entry_ip, regs); > > + ret = fentry_perf_func(tf, entry_ip, fregs, regs); > > #endif > > return ret; > > } > > @@ -347,14 +349,15 @@ static void fexit_dispatcher(struct fprobe *fp, unsigned long entry_ip, > > struct trace_fprobe *tf = container_of(fp, struct trace_fprobe, fp); > > struct pt_regs *regs = ftrace_get_regs(fregs); > > And same here with the return handler > > I think fprobe events would need the same sort of refactoring as > kprobe_multi bpf: using ftrace_partial_regs so they work on build > !WITH_REGS. Actually, kprobe_multi is using fprobe directly, so this is not related to bpf part. Thank you, -- Masami Hiramatsu (Google) <mhiramat@xxxxxxxxxx>
