Re: [PATCH 19/24] rcu/tree: Support reclaim for head-less object

From: "Paul E. McKenney" <paulmck@kernel.org>
To: Uladzislau Rezki <urezki@gmail.com>
Cc: Joel Fernandes <joel@joelfernandes.org>,
	LKML <linux-kernel@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>, RCU <rcu@vger.kernel.org>,
	Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Subject: Re: [PATCH 19/24] rcu/tree: Support reclaim for head-less object
Date: Mon, 4 May 2020 08:31:47 -0700	[thread overview]
Message-ID: <20200504153147.GL2869@paulmck-ThinkPad-P72> (raw)
In-Reply-To: <20200504142153.GG17577@pc636>

On Mon, May 04, 2020 at 04:21:53PM +0200, Uladzislau Rezki wrote:
> > > > 
> > > > If we are not doing single-pointer allocation, then that would also eliminate
> > > > entering the low-level page allocator for single-pointer allocations.
> > > > 
> > > > Or did you mean entry into the allocator for the full-page allocations
> > > > related to the pointer array for PREEMPT_RT? Even if we skip entry into the
> > > > allocator for those, we will still have additional caching which further
> > > > reduces chances of getting a full page. In the event of such failure, we can
> > > > simply queue the rcu_head.
> > > > 
> > > > Thoughts?
> > > 
> > > I was just trying to guess why you kept the single-pointer allocation.
> > > It looks like I guessed wrong.  ;-)
> > > 
> > > If, as you say above, you make it go straight to synchronize_rcu()
> > > upon full-page allocation failure, that would be good!
> > 
> > Paul, sounds good. Vlad, are you also Ok with that?
> > 
> OK, let's drop it and keep it simple :)
> 
> BTW, for PREEMPT_RT we still can do a page allocation for single
> argument of kvfree_rcu(). In case of double we just revert everything
> to the rcu_head if no cache.
> 
> For single argument we can drop the lock before the entry to the page
> allocator. Because it follows might_sleep() anotation we avoid of having
> a situation when spinlock(rt mutex) is taken from any atomic context.
> 
> Since the lock is dropped the current context can be interrupted by
> an IRQ which in its turn can also call kvfree_rcu() on current CPU.
> In that case it must be double argument(single is not allowed) kvfree_rcu()
> call. For PREEMPT_RT if no cache everything is reverted to rcu_head usage,
> i.e. the entry to page allocator is bypassed.
> 
> It can be addressed as a separate patch and send out later on if we
> are on the same page.
> 
> Paul, Joel what are your opinions?

I strongly prefer that it be removed from the series.  I do understand
that this is a bit more hassle right now, but this does help avoid
confusion in the future, plus perhaps also avoiding issues with future
bisections.

							Thanx, Paul