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

From: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
To: Uladzislau Rezki <urezki@gmail.com>
Cc: Joel Fernandes <joel@joelfernandes.org>,
	LKML <linux-kernel@vger.kernel.org>, RCU <rcu@vger.kernel.org>,
	linux-mm <linux-mm@kvack.org>,
	"Paul E . McKenney" <paulmck@kernel.org>,
	Andrew Morton <akpm@linux-foundation.org>,
	"Theodore Y . Ts'o" <tytso@mit.edu>,
	Matthew Wilcox <willy@infradead.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 17:04:14 +0200	[thread overview]
Message-ID: <20200716150414.iqpyby6nrww4zbyk@linutronix.de> (raw)
In-Reply-To: <20200716144728.GA31046@pc636>

On 2020-07-16 16:47:28 [+0200], Uladzislau Rezki wrote:
> On Thu, Jul 16, 2020 at 04:25:37PM +0200, Sebastian Andrzej Siewior wrote:
> > On 2020-07-16 11:19:13 [+0200], Uladzislau Rezki wrote:
> > > Sebastian, could you please confirm that if that patch that is in
> > > question fixes it?
> > > 
> > > It would be appreciated!
> > 
> > So that preempt disable should in terms any warnings. However I don't
> > think that it is strictly needed and from scheduling point of view you
> > forbid a CPU migration which might be good otherwise.
> >
> Please elaborate your point regarding "i do not think it is strictly needed".
> 
> Actually i can rework the patch to remove even such preempt_enable/disable
> to stay on the same CPU, but i do not see the point of doing it.
> 
> Do you see the point?

You disable preemption for what reason? It is not documented, it is not
obvious - why is it required?

> As for scheduling point of view. Well, there are many places when there
> is a demand in memory or pages from atomic context. Also, getting a page
> is not considered as a hot path in the kfree_rcu(). 

If you disable preemption than you assume that you wouldn't be atomic
otherwise. You say that at this point it is not a hot path so if this is
not *that* important why not allow preemption and allow the schedule to
place you somewhere else if the scheduler decides that it is a good idea.

> > Also if interrupts and everything is enabled then someone else might
> > invoke kfree_rcu() from BH context for instance.
> > 
> And what? What is a problem here, please elaborate if you see any
> issues.

That the kfree_rcu() caller from BH context will end up here as well,
asking for a page.

Sebastian