Re: [PATCH v4 16/27] tracing: Remove regular RCU context for _rcuidle tracepoints (again)

[Alexei Starovoitov]

On Fri, Mar 6, 2020 at 12:55 PM Mathieu Desnoyers
<mathieu.desnoyers@xxxxxxxxxxxx> wrote:
>
> ----- On Mar 6, 2020, at 3:45 PM, rostedt rostedt@xxxxxxxxxxx wrote:
>
> > On Fri, 6 Mar 2020 15:22:46 -0500 (EST)
> > Mathieu Desnoyers <mathieu.desnoyers@xxxxxxxxxxxx> wrote:
> >
> >> I agree with the overall approach. Just a bit of nitpicking on the API:
> >>
> >> I understand that the "prio" argument is a separate argument because it can take
> >> many values. However, "rcu" is just a boolean, so I wonder if we should not
> >> rather
> >> introduce a "int flags" with a bitmask enum, e.g.
> >
> > I thought about this approach, but thought it was a bit overkill. As the
> > kernel doesn't have an internal API, I figured we can switch this over to
> > flags when we get another flag to add. Unless you can think of one in the
> > near future.
>
> The additional feature I have in mind for near future would be to register
> a probe which can take a page fault to a "sleepable" tracepoint. This would
> require preemption to be enabled and use of SRCU.

I'm working on sleepable bpf as well and this extra flag for tracepoints
would come very handy, so I would go with flags approach right away.
We wouldn't need to touch the same protos multiple times,
less conflicts for us all, etc.