Re: [RFC 0/2] Reenable might_sleep() checks for might_fault() when atomic

From: Thomas Gleixner <tglx@linutronix.de>
To: David Hildenbrand <dahi@linux.vnet.ibm.com>
Cc: linux-arch@vger.kernel.org, "Michael S. Tsirkin" <mst@redhat.com>,
	Heiko Carstens <heiko.carstens@de.ibm.com>,
	linux-kernel@vger.kernel.org,
	Christian Borntraeger <borntraeger@de.ibm.com>,
	paulus@samba.org, schwidefsky@de.ibm.com,
	akpm@linux-foundation.org, linuxppc-dev@lists.ozlabs.org,
	mingo@kernel.org
Subject: Re: [RFC 0/2] Reenable might_sleep() checks for might_fault() when atomic
Date: Thu, 27 Nov 2014 22:52:08 +0100 (CET)	[thread overview]
Message-ID: <alpine.DEB.2.11.1411272246110.3961@nanos> (raw)
In-Reply-To: <20141127161905.7c6220ee@thinkpad-w530>

On Thu, 27 Nov 2014, David Hildenbrand wrote:
> > OTOH, there is no reason why we need to disable preemption over that
> > page_fault_disabled() region. There are code pathes which really do
> > not require to disable preemption for that.
> > 
> > We have that seperated in preempt-rt for obvious reasons and IIRC
> > Peter Zijlstra tried to distangle it in mainline some time ago. I
> > forgot why that never got merged.
> > 
> 
> Of course, we can completely separate that in our page fault code by doing
> pagefault_disabled() checks instead of in_atomic() checks (even in add on
> patches later).
> 
> > We tie way too much stuff on the preemption count already, which is a
> > mightmare because we have no clear distinction of protection
> > scopes. 
> 
> Although it might not be optimal, but keeping a separate counter for
> pagefault_disable() as part of the preemption counter seems to be the only
> doable thing right now.

It needs to be seperate, if it should be useful. Otherwise we just
have a extra accounting in preempt_count() which does exactly the same
thing as we have now: disabling preemption.

Now you might say, that we could mask out that part when checking
preempt_count, but that wont work on x86 as x86 has the preempt
counter as a per cpu variable and not as a per thread one.

But if you want to distangle pagefault disable from preempt disable
then you must move it to the thread, because it is a property of the
thread. preempt count is very much a per cpu counter as you can only
go through schedule when it becomes 0.

Btw, I find the x86 representation way more clear, because it
documents that preempt count is a per cpu BKL and not a magic thread
property. And sadly that is how preempt count is used ...

> I am not sure if a completely separated counter is even possible,
> increasing the size of thread_info.

And adding a ulong to thread_info is going to create exactly which
problem?

Thanks,

	tglx