Re: [PATCH 04/16] rcu/nocb: Only (re-)initialize segcblist when needed on CPU up

[Frederic Weisbecker <frederic@kernel.org>]

On Thu, Jan 28, 2021 at 11:12:28AM -0800, Paul E. McKenney wrote:
> On Thu, Jan 28, 2021 at 06:12:10PM +0100, Frederic Weisbecker wrote:
> > Simply checking if the segcblist is enabled is enough to know if we
> > need to initialize it or not. It's safe to check within hotplug
> > machine.
> > 
> > Signed-off-by: Frederic Weisbecker <frederic@kernel.org>
> > Cc: Josh Triplett <josh@joshtriplett.org>
> > Cc: Lai Jiangshan <jiangshanlai@gmail.com>
> > Cc: Joel Fernandes <joel@joelfernandes.org>
> > Cc: Neeraj Upadhyay <neeraju@codeaurora.org>
> > Cc: Boqun Feng <boqun.feng@gmail.com>
> 
> Hmmm...
> 
> At the start of a CPU-hotplug operation, an incoming CPU's callback
> list can be in a number of states:
> 
> 1.	Disabled and empty.  This is the case when the boot CPU has
> 	not done call_rcu(), when a non-boot CPU first comes online,
> 	and when a non-offloaded CPU comes back online.  In this case,
> 	it is permissible to initialize ->cblist.  Because either the
> 	CPU is currently running with interrupts disabled (boot CPU)
> 	or is not yet running at all (other CPUs), it is not necessary
> 	to acquire ->nocb_lock.
> 
> 2.	Disabled and non-empty.  This is the case when the boot CPU has
> 	done call_rcu().  It is not permissible to initialize ->cblist
> 	because doing so will leak any callbacks posted by early boot
> 	invocations of call_rcu().

I don't think that's possible. In this case __call_rcu() has called
rcu_segcblist_init() and has enabled the segcblist.

> 
> 	Test for the possibility of leaking by building with
> 	CONFIG_PROVE_RCU=y and booting with rcupdate.rcu_self_test=1.
> 
> 3.	Enabled, whether empty or not.  This is the case when an
> 	offloaded CPU comes back online.  This is the only case where
> 	the ->nocb_lock must be held to modify ->cblist.  However,
> 	it is not necessarily to modify ->cblist because the rcuoc
> 	kthread is on the job.
> 
> So I believe that it is necessary to check for both disabled and empty.
> But don't take my word for it!  Build with CONFIG_PROVE_RCU=y and boot
> with rcupdate.rcu_self_test=1.  ;-)

I'm trying that :-)