Date:	Mon, 15 Jun 2015 10:32:16 -0400
From:	Jerome Glisse <j.glisse@...il.com>
To:	Mark Hairgrove <mhairgrove@...dia.com>
Cc:	"akpm@...ux-foundation.org" <akpm@...ux-foundation.org>,
	"linux-kernel@...r.kernel.org" <linux-kernel@...r.kernel.org>,
	"linux-mm@...ck.org" <linux-mm@...ck.org>,
	Linus Torvalds <torvalds@...ux-foundation.org>,
	"joro@...tes.org" <joro@...tes.org>, Mel Gorman <mgorman@...e.de>,
	"H. Peter Anvin" <hpa@...or.com>,
	Peter Zijlstra <peterz@...radead.org>,
	Andrea Arcangeli <aarcange@...hat.com>,
	Johannes Weiner <jweiner@...hat.com>,
	Larry Woodman <lwoodman@...hat.com>,
	Rik van Riel <riel@...hat.com>,
	Dave Airlie <airlied@...hat.com>,
	Brendan Conoboy <blc@...hat.com>,
	Joe Donohue <jdonohue@...hat.com>,
	Duncan Poole <dpoole@...dia.com>,
	Sherry Cheung <SCheung@...dia.com>,
	Subhash Gutti <sgutti@...dia.com>,
	John Hubbard <jhubbard@...dia.com>,
	Lucien Dunning <ldunning@...dia.com>,
	Cameron Buschardt <cabuschardt@...dia.com>,
	Arvind Gopalakrishnan <arvindg@...dia.com>,
	Haggai Eran <haggaie@...lanox.com>,
	Shachar Raindel <raindel@...lanox.com>,
	Liran Liss <liranl@...lanox.com>,
	Roland Dreier <roland@...estorage.com>,
	Ben Sander <ben.sander@....com>,
	Greg Stoner <Greg.Stoner@....com>,
	John Bridgman <John.Bridgman@....com>,
	Michael Mantor <Michael.Mantor@....com>,
	Paul Blinzer <Paul.Blinzer@....com>,
	Laurent Morichetti <Laurent.Morichetti@....com>,
	Alexander Deucher <Alexander.Deucher@....com>,
	Oded Gabbay <Oded.Gabbay@....com>,
	Jérôme Glisse <jglisse@...hat.com>,
	Jatin Kumar <jakumar@...dia.com>,
	"linux-rdma@...r.kernel.org" <linux-rdma@...r.kernel.org>
Subject: Re: [PATCH 05/36] HMM: introduce heterogeneous memory management v3.

On Thu, Jun 11, 2015 at 03:26:46PM -0700, Mark Hairgrove wrote:
> On Thu, 11 Jun 2015, Jerome Glisse wrote:
> > On Wed, Jun 10, 2015 at 06:15:08PM -0700, Mark Hairgrove wrote:

[...]
> > Ok i see the race you are afraid of and really it is an unlikely one
> > __mutex_unlock_common_slowpath() take a spinlock right after allowing
> > other to take the mutex, when we are in your scenario there is no
> > contention on that spinlock so it is taken right away and as there
> > is no one in the mutex wait list then it goes directly to unlock the
> > spinlock and return. You can ignore the debug function as if debugging
> > is enabled than the mutex_lock() would need to also take the spinlock
> > and thus you would have proper synchronization btw 2 thread thanks to
> > the mutex.wait_lock.
> > 
> > So basicly while CPU1 is going :
> > spin_lock(mutex.wait_lock)
> > if (!list_empty(mutex.wait_list)) {
> >   // wait_list is empty so branch not taken
> > }
> > spin_unlock(mutex.wait_lock)
> > 
> > CPU2 would have to test the mirror list and mutex_unlock and return
> > before the spin_unlock() of CPU1. This is a tight race, i can add a
> > synchronize_rcu() to device_unregister after the mutex_unlock() so
> > that we also add a grace period before the device is potentialy freed
> > which should make that race completely unlikely.
> > 
> > Moreover for something really bad to happen it would need that the
> > freed memory to be reallocated right away by some other thread. Which
> > really sound unlikely unless CPU1 is the slowest of all :)
> > 
> > Cheers,
> > Jérôme
> > 
> 
> But CPU1 could get preempted between the atomic_set and the 
> spin_lock_mutex, and then it doesn't matter whether or not a grace period 
> has elapsed before CPU2 proceeds.
> 
> Making race conditions less likely just makes them harder to pinpoint when 
> they inevitably appear in the wild. I don't think it makes sense to spend 
> any effort in making a race condition less likely, and that thread I 
> referenced (https://lkml.org/lkml/2013/12/2/997) is fairly strong evidence 
> that fixing this race actually matters. So, I think this race condition 
> really needs to be fixed.
> 
> One fix is for hmm_mirror_unregister to wait for hmm_notifier_release 
> completion between hmm_mirror_kill and hmm_mirror_unref. It can do this by 
> calling synchronize_srcu() on the mmu_notifier's srcu. This has the 
> benefit that the driver is guaranteed not to get the "mm is dead" callback 
> after hmm_mirror_unregister returns.
> 
> In fact, are there any callbacks on the mirror that can arrive after 
> hmm_mirror_unregister? If so, how will hmm_device_unregister solve them?
> 
> From a general standpoint, hmm_device_unregister must perform some kind of 
> synchronization to be sure that all mirrors are completely released and 
> done and no new callbacks will trigger. Since that has to be true, can't 
> that synchronization be moved into hmm_mirror_unregister instead?
> 
> If that happens there's no need for a "mirror can be freed" ->release 
> callback at all because the driver is guaranteed that a mirror is done 
> after hmm_mirror_unregister.

Well there is no need or 2 callback (relase|stop , free) just one, the
release|stop that is needed. I kind of went halfway last week on this.
I will probably rework that a little to keep just one call and rely on
driver to call hmm_mirror_unregister()

Cheers,
Jérôme
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

Powered by blists - more mailing lists