Re: [RFC] ftrace: Add support to keep some functions out of ftrace

[Alexei Starovoitov]

On Mon, Aug 15, 2022 at 8:02 AM Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
>
> On Mon, Aug 15, 2022 at 07:45:16AM -0700, Alexei Starovoitov wrote:
> > On Mon, Aug 15, 2022 at 7:33 AM Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
>
> > > It is in full control of the 'call __fentry__'. Absolute full NAK on you
> > > trying to make it otherwise.
> >
> > Don't mix 'call fentry' generated by the compiler with nop5 inserted
> > by macroses or JITs.
>
> Looking at:
>
>  https://lore.kernel.org/all/20191211123017.13212-3-bjorn.topel@xxxxxxxxx/
>
> this seems to want to prod at the __fentry__ sites.
>
> > > > Soon we will have nop5 in the middle of the function.
> > > > ftrace must not touch it.
> > >
> > > How are you generating that NOP and what for?
> >
> > We're generating nop5-s in JITed code to further
> > attach to.
>
> Ftrace doesn't know about those; so how can it break that?
>
> Likewise it doesn't know about the static_branch/static_call NOPs and
> nothing is broken.
>
> Ftrace only knows about the __fentry__ sites, and it *does* own them.
> Are you saying ftrace is writing to a code location not a __fentry__
> site?

Let's keep it in one thread:

> It wasn't long before. Yes it landed a few months prior to the
> static_call work, but the whole static_call thing was in progress for a
> long long time.
>
> Anyway, yes it is different. But it's still very much broken. You simply
> cannot step on __fentry__ sites like that.

Ask yourself: should static_call patching logic go through
ftrace infra ? No. Right?
static_call has nothing to do with ftrace (function tracing).
Same thing here. bpf dispatching logic is nothing to do with
function tracing.
In this case bpf_dispatcher_xdp_func is a placeholder written C.
If it was written in asm, fentry recording wouldn't have known about it.
And that's more or less what Jiri patch is doing.
It's hiding a fake function from ftrace, since it's not a function
and ftrace infra shouldn't show it tracing logs.
In other words it's a _notrace_ function with nop5.