Re: [PATCH 1/1] rcu/tree: Drop the lock before entering to page allocator

From: "Paul E. McKenney" <paulmck@kernel.org>
To: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Cc: Uladzislau Rezki <urezki@gmail.com>,
	LKML <linux-kernel@vger.kernel.org>, RCU <rcu@vger.kernel.org>,
	linux-mm@kvack.org, Andrew Morton <akpm@linux-foundation.org>,
	"Theodore Y . Ts'o" <tytso@mit.edu>,
	Matthew Wilcox <willy@infradead.org>,
	Joel Fernandes <joel@joelfernandes.org>,
	Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Subject: Re: [PATCH 1/1] rcu/tree: Drop the lock before entering to page allocator
Date: Thu, 16 Jul 2020 09:36:51 -0700	[thread overview]
Message-ID: <20200716163651.GT9247@paulmck-ThinkPad-P72> (raw)
In-Reply-To: <20200716153638.gfh6dzp2h35ygfaa@linutronix.de>

On Thu, Jul 16, 2020 at 05:36:38PM +0200, Sebastian Andrzej Siewior wrote:
> On 2020-07-16 08:20:27 [-0700], Paul E. McKenney wrote:
> > You lost me on this one.  I am instead concerned that something like this
> > might be needed on short notice:
> > 
> > 	raw_spin_lock(&some_lock);
> > 	kfree_rcu(some_pointer, some_field_offset);
> > 
> > In contrast, single-argument kfree_rcu() cannot be invoked from any
> > environment where synchronize_rcu() cannot be invoked.
> 
> I see. We don't have any kfree() in that context as far as I remember.
> We had a few cases in "resize" where you allocate memory, copy content
> and free old memory while under the lock but they are gone.

True, but we also didn't have any calls to call_rcu() prior to the call
to rcu_init() until suddenly we did.  (Yeah, I could have put my foot
down and prohibited that practice, but the workarounds were quite a bit
more complicated than just making call_rcu() work during very early boot.)

And last I checked, there really were calls to call_rcu() under raw
spinlocks, so the potential or calls to double-argument kfree_rcu()
clearly exists and is very real.

> > > > Yes, dropping to a plain spinlock would be simple in the here and now,
> > > > but experience indicates that it is only a matter of time, and that when
> > > > that time comes it will come as an emergency.
> > > 
> > > Hmmm.
> > 
> > I point out the call_rcu() experience.
> > 
> > > > One approach would be to replace the "IS_ENABLED(CONFIG_PREEMPT_RT)"
> > > > with some sort of check for being in a context where spinlock acquisition
> > > > is not legal.  What could be done along those lines?
> > > 
> > > I would rethink the whole concept how this is implemented now and give
> > > it another try. The code does not look pretty and is looking
> > > complicated. The RT covering of this part then just added a simple
> > > return because nothing else seemed to be possible. This patch here
> > > looks like another duct tape attempt to avoid a warning.
> > 
> > In addition to the possibility of invocation from BH?
> 
> Invocation from BH should be possible because network would probably be
> the first user. I don't remember anything wrong with BH if I remember
> correctly.

OK, that is reassuring.  Here is hoping!

							Thanx, Paul