Re: linux-next: build warnings after merge of the tip tree

[Masami Hiramatsu <mhiramat@kernel.org>]

On Tue, 22 Mar 2022 15:35:54 +0100
Peter Zijlstra <peterz@infradead.org> wrote:

> On Tue, Mar 22, 2022 at 09:12:42AM -0400, Steven Rostedt wrote:
> 
> > > Suppose:
> > > 
> > > notrace func_B()
> > > {
> > > 	...
> > > }
> > > 
> > > func_A()
> > > {
> > > 	...
> > > 	return func_B();
> > > }
> > > 
> > > then inhibiting tail calls would end up looking like:
> > 
> > If we inhibit tail calls, then we do not need to make func_B notrace.
> 
> Dude, you're arguing in circles :-( the notrace was a given.
> 
> > > func_A:
> > > 	call __fentry__
> > > 	...
> > > 	call func_B
> > > 	call __fexit__
> > > 	ret
> > > 
> > > Then A is fully traced, B is invisible, as per spec. What is the
> > > problem?
> > 
> > The above is fine, but then func_B is not a tail call and can also be
> > traced.
> 
> Again, B is notrace as a given. This was all about how to deal with
> notrace functions.
> 
> I suggested inhibiting tail-call to notrace, you said no. You now seem to
> agree that solves it.
> 
> > > The problem you initially had, of doing a tail-call into a notrace, was
> > > that the __fexit__ call went missing, because notrace will obviously not
> > > have that. But that's avoided by inhibiting all tail-calls between
> > > notrace and !notrace functions (note that notrace must also not
> > > tail-call !notrace).
> > 
> > I'm confused by the above. Why can't a notrace tail call a !notrace?
> > If we tail call to a
> > 
> > func_B:
> > 	call __fentry__
> > 	...
> > 	call __fexit__
> > 	ret
> > 
> > then the fentry and fexit show a perfectly valid trace of func_B.
> 
> Bah; I thought I had a case this morning, but now I can't seem to recall
> :/
> 
> > > Your worry seems to stem about loosing visiblilty of !notrace functions,
> > > but AFAICT that doesn't happen.
> > 
> > My worry is:
> > 
> > func_A:
> > 	call __fentry__
> > 	...
> > 	jmp func_B
> > 
> > Where do we do the call __fexit__ ?
> 
> In B (or wherever if B again does a tail-call).
> 
> > That was the original concern, and I think the proposed solutions have
> > convoluted our thoughts about what we are trying to fix. So let's go back
> > to the beginning, and see how to deal with it.
> > 
> > That is, we have:
> > 
> > func_C:
> > 	call __fenty__
> > 	...
> > 	call func_A:
> > 	...
> > 	call func_B:
> > 	...
> > 	call __fexit__
> > 	ret
> > 
> > func_A:
> > 	call __fentry__
> > 	...
> 	call __ftail__
> > 	jmp func_B
> > 
> > func_B:
> > 	call __fentry__
> > 	...
> > 	call __fexit__
> > 	ret
> > 
> > Where the above is C calling A and B as normal functions, A calling B as a
> > tail call and B just being a normal function called by both A and C (and
> > many other functions).
> 
> We need the __ftail__ thing to mark the trace-stack entry of func_A as
> complete, then any future __fexit__ will be able to pop all completed
> entries.
> 
> In recap:
> 
> 	__fentry__ -- push on trace-stack
> 	__ftail__  -- mark top-most entry complete
> 	__fexit__  -- mark top-most entry complete;
> 	              pop all completed entries
> 
> inhibit tail-calls to notrace.
> 
> > And note, I do not want to limit function tracing (which does not rely on
> > __fexit__) just because we can't figure out how to handle __fexit__.
> 
> I'm not following. Regular function tracing needs none of this.
> 
> It's function graph tracing, kretprobes and whatever else this rethook
> stuff is about that needs this because return trampolines will stop
> working somewhere in the not too distant future.

I see the __fexit__ is needed, but why __ftail__ is needed? I guess because
func_B is notrace, in that case the __fexit__ will not be in the func_B.
Am I correct?

Thank you,

-- 
Masami Hiramatsu <mhiramat@kernel.org>