Re: [PATCH 1/3] tracepoints: Do not use call_rcu_sched() before early_initcall()

["Paul E. McKenney" <paulmck@linux.vnet.ibm.com>]

On Sun, Dec 14, 2014 at 12:44:31PM -0500, Steven Rostedt wrote:
> On Sun, 14 Dec 2014 17:29:28 +0000 (UTC)
> Mathieu Desnoyers <mathieu.desnoyers@efficios.com> wrote:
> 
> > Given that your reason for having this RCU-specific logic in tracepoint.c
> > rather than within call_rcu*() is not slowing down a fast-path, how about
> > creating a new call_rcu_early() and call_rcu_sched_early() which can be
> > called in normal operation and at early boot ?
> 
> That's a possibility.
> 
> > 
> > This would allow us to keep the RCU logic within the RCU implementation
> > rather than strongly coupling it with the tracepoint code.
> > 
> 
> It's not that strong of a coupling to RCU. It's more coupled to being
> called really early (which needs special care).
> 
> It just happened that RCU failed for being called that early. Other
> things could possible fail too (if added to the tracepoint logic).
> Maybe I should rename the variable to "tracepoint_earlyboot" instead.

But you do have to call this quite early.  After rcu_init() is invoked,
things should work fine.

> But as RCU is the only thing that failed (so far in my testing), I'll
> think about adding a call_rcu_sched_early(). But then, this does make
> things more complex, and me more nervous about adding it.

I really am nervous about a call_rcu_sched_early() that immediately
invokes the specified callback.  That is just begging for someone to
invoke it while traversing a list in an RCU read-side critical section,
which will break.

My thought is to make the compiler initialize the pieces of RCU that
are needed.  That said, this initialization includes per-CPU variables,
so the question then becomes "when do per-CPU variables get initialized?"

							Thanx, Paul