Re: [PATCH] hugetlb: select PREEMPT_COUNT if HUGETLB_PAGE for in_atomic use

From: Michal Hocko <mhocko@suse.com>
To: Peter Zijlstra <peterz@infradead.org>
Cc: Mike Kravetz <mike.kravetz@oracle.com>,
	linux-mm@kvack.org, linux-kernel@vger.kernel.org,
	"Paul E . McKenney" <paulmck@kernel.org>,
	Shakeel Butt <shakeelb@google.com>,
	tglx@linutronix.de, john.ogness@linutronix.de, urezki@gmail.com,
	ast@fb.com, Eric Dumazet <edumazet@google.com>,
	Mina Almasry <almasrymina@google.com>,
	Andrew Morton <akpm@linux-foundation.org>
Subject: Re: [PATCH] hugetlb: select PREEMPT_COUNT if HUGETLB_PAGE for in_atomic use
Date: Thu, 11 Mar 2021 10:44:56 +0100	[thread overview]
Message-ID: <YEnmmK42kpeB3Ho4@dhcp22.suse.cz> (raw)
In-Reply-To: <YEnjqLpGU2LHaysv@hirez.programming.kicks-ass.net>

On Thu 11-03-21 10:32:24, Peter Zijlstra wrote:
> On Thu, Mar 11, 2021 at 10:01:22AM +0100, Michal Hocko wrote:
> > On Thu 11-03-21 09:46:30, Peter Zijlstra wrote:
> > > On Wed, Mar 10, 2021 at 06:13:21PM -0800, Mike Kravetz wrote:
> > > > from irq context.  Changing the check in the code from !in_task to
> > > > in_atomic would handle the situations when called with irqs disabled.
> > > 
> > > It does not. local_irq_disable() does not change preempt_count().
> > 
> > You are right. Earlier I was suggesting to check of irq_disabled() as
> > well http://lkml.kernel.org/r/YD4I+VPr3UNt063H@dhcp22.suse.cz
> > 
> > back then it was not really clear to me that in fact we do care about
> > spin locks more than irq disabled code. I am not even sure whether we
> > need to care about irq disabled regions without any locks held that
> > wouldn't be covered by in_atomic. But it would be safer to add
> > irq_disabled check as well.
> 
> Safer still is always doing it, replace it with if (true).
> 
> What's the purpose, doing the minimal 'correct', of the maximal safe
> solution?

If we always defer to a WQ context then an admin wouldn't have any
feedback from the syscall when releasing the pool.

> The whole changelog reads like a trainwreck, but akpm already commented
> on that. I picked out a small factual incorrectness, simply because if
> you can't get that right, the whole argument looses weight.

Is there any reason why in_atomic || irq_disabled wouldn't work
universally?

> That said, I don't think you actually need it, if as you write the lock
> should be IRQ-safe, then you're worried about the IRQ recursion
> deadlock:

making hugetlb_lock irqsafe is a long way as explained by Mike
elsewhere. Not only that. The upcoming hugeltb feature to have sparse
vmemmap for hugetlb pages will need to allocate vmemmap when hugetlb
page is to be freed back to the allocator. That cannot happen in any
atomic context so there will be a need to tell those contexts for
special casing.

-- 
Michal Hocko
SUSE Labs