----- On Feb 18, 2020, at 11:12 AM, paulmck paulmck@xxxxxxxxxx wrote: > On Tue, Feb 18, 2020 at 01:33:35PM +0900, Masami Hiramatsu wrote: >> On Mon, 17 Feb 2020 08:31:12 -0800 >> "Paul E. McKenney" <paulmck@xxxxxxxxxx> wrote: >> > >> > > BTW, if you consider the x86 specific code is in the generic file, >> > > we can move NOKPROBE_SYMBOL() in arch/x86/kernel/traps.c. >> > > (Sorry, I've hit this idea right now) >> > >> > Might this affect other architectures with NMIs and probe-like things? >> > If so, it might make sense to leave it where it is. >> >> Yes, git grep shows that arm64 is using rcu_nmi_enter() in >> debug_exception_enter(). >> OK, let's keep it, but maybe it is good to update the comment for >> arm64 too. What about following? >> >> +/* >> + * All functions in do_int3() on x86, do_debug_exception() on arm64 must be >> + * marked NOKPROBE before kprobes handler is called. >> + * ist_enter() on x86 and debug_exception_enter() on arm64 which is called >> + * before kprobes handle happens to call rcu_nmi_enter() which means >> + * that rcu_nmi_enter() must be marked NOKRPOBE. >> + */ > > Would it work to describe the general problem, then give x86 details > as a specific example, as follows? > > /* > * On some architectures, certain exceptions prohibit use of kprobes until > * the exception code path reaches a certain point. For example, on x86 all > * functions called by do_int3() must be marked NOKPROBE. However, once > * kprobe_int3_handler() is called, kprobing is permitted. Specifically, > * ist_enter() is called in do_int3() before kprobe_int3_handle(). > * Furthermore, ist_enter() calls rcu_nmi_enter(), which means that > * rcu_nmi_enter() must be marked NOKRPOBE. > */ > > That way, I don't feel like I need to update the commment each time > a new architecture adds this capability. ;-) I suspect this kind of comment would belong in a new Documentation/features/kprobes/annotations.txt or something similar. You might want to look at other "features" files to see the expected layout. Thanks, Mathieu -- Mathieu Desnoyers EfficiOS Inc. http://www.efficios.com
