Message-Id: <20171213032725.GJ7829@linux.vnet.ibm.com>
Date:   Tue, 12 Dec 2017 19:27:25 -0800
From:   "Paul E. McKenney" <paulmck@...ux.vnet.ibm.com>
To:     "Huang, Ying" <ying.huang@...el.com>
Cc:     Andrew Morton <akpm@...ux-foundation.org>,
        Minchan Kim <minchan@...nel.org>, linux-mm@...ck.org,
        linux-kernel@...r.kernel.org, Hugh Dickins <hughd@...gle.com>,
        Johannes Weiner <hannes@...xchg.org>,
        Tim Chen <tim.c.chen@...ux.intel.com>,
        Shaohua Li <shli@...com>,
        Mel Gorman <mgorman@...hsingularity.net>,
        J�r�me Glisse <jglisse@...hat.com>,
        Michal Hocko <mhocko@...e.com>,
        Andrea Arcangeli <aarcange@...hat.com>,
        David Rientjes <rientjes@...gle.com>,
        Rik van Riel <riel@...hat.com>, Jan Kara <jack@...e.cz>,
        Dave Jiang <dave.jiang@...el.com>,
        Aaron Lu <aaron.lu@...el.com>
Subject: Re: [PATCH -mm] mm, swap: Fix race between swapoff and some swap
 operations

On Wed, Dec 13, 2017 at 10:17:41AM +0800, Huang, Ying wrote:
> "Paul E. McKenney" <paulmck@...ux.vnet.ibm.com> writes:
> 
> > On Tue, Dec 12, 2017 at 09:12:20AM +0800, Huang, Ying wrote:
> >> Hi, Pual,
> >> 
> >> "Paul E. McKenney" <paulmck@...ux.vnet.ibm.com> writes:
> >> 
> >> > On Mon, Dec 11, 2017 at 01:30:03PM +0800, Huang, Ying wrote:
> >> >> Andrew Morton <akpm@...ux-foundation.org> writes:
> >> >> 
> >> >> > On Fri, 08 Dec 2017 16:41:38 +0800 "Huang\, Ying" <ying.huang@...el.com> wrote:
> >> >> >
> >> >> >> > Why do we need srcu here? Is it enough with rcu like below?
> >> >> >> >
> >> >> >> > It might have a bug/room to be optimized about performance/naming.
> >> >> >> > I just wanted to show my intention.
> >> >> >> 
> >> >> >> Yes.  rcu should work too.  But if we use rcu, it may need to be called
> >> >> >> several times to make sure the swap device under us doesn't go away, for
> >> >> >> example, when checking si->max in __swp_swapcount() and
> >> >> >> add_swap_count_continuation().  And I found we need rcu to protect swap
> >> >> >> cache radix tree array too.  So I think it may be better to use one
> >> >> >> calling to srcu_read_lock/unlock() instead of multiple callings to
> >> >> >> rcu_read_lock/unlock().
> >> >> >
> >> >> > Or use stop_machine() ;)  It's very crude but it sure is simple.  Does
> >> >> > anyone have a swapoff-intensive workload?
> >> >> 
> >> >> Sorry, I don't know how to solve the problem with stop_machine().
> >> >> 
> >> >> The problem we try to resolved is that, we have a swap entry, but that
> >> >> swap entry can become invalid because of swappoff between we check it
> >> >> and we use it.  So we need to prevent swapoff to be run between checking
> >> >> and using.
> >> >> 
> >> >> I don't know how to use stop_machine() in swapoff to wait for all users
> >> >> of swap entry to finish.  Anyone can help me on this?
> >> >
> >> > You can think of stop_machine() as being sort of like a reader-writer
> >> > lock.  The readers can be any section of code with preemption disabled,
> >> > and the writer is the function passed to stop_machine().
> >> >
> >> > Users running real-time applications on Linux don't tend to like
> >> > stop_machine() much, but perhaps it is nevertheless the right tool
> >> > for this particular job.
> >> 
> >> Thanks a lot for explanation!  Now I understand this.
> >> 
> >> Another question, for this specific problem, I think both stop_machine()
> >> based solution and rcu_read_lock/unlock() + synchronize_rcu() based
> >> solution work.  If so, what is the difference between them?  I guess rcu
> >> based solution will be a little better for real-time applications?  So
> >> what is the advantage of stop_machine() based solution?
> >
> > The stop_machine() solution places similar restrictions on readers as
> > does rcu_read_lock/unlock() + synchronize_rcu(), if that is what you
> > are asking.
> >
> > More precisely, the stop_machine() solution places exactly the
> > same restrictions on readers as does preempt_disable/enable() and
> > synchronize_sched().
> >
> > I would expect stop_machine() to be faster than either synchronize_rcu()
> > synchronize_sched(), or synchronize_srcu(), but stop_machine() operates
> > by making each CPU spin with interrupts until all the other CPUs arrive.
> > This normally does not make real-time people happy.
> >
> > An compromise position is available in the form of
> > synchronize_rcu_expedited() and synchronize_sched_expedited().  These
> > are faster than their non-expedited counterparts, and only momentarily
> > disturb each CPU, rather than spinning with interrupts disabled.  However,
> > stop_machine() is probably a bit faster.
> >
> > Finally, syncrhonize_srcu_expedited() is reasonably fast, but
> > avoids disturbing other CPUs.  Last I checked, not quite as fast as
> > synchronize_rcu_expedited() and synchronize_sched_expedited(), though.
> >
> > You asked!  ;-)
> 
> Thanks a lot Paul!  That exceeds my expectation!
> 
> The performance of swapoff() isn't very important, probably it's not
> necessary to accelerate it at the cost of realtime.  I think it is
> better to use a rcu or srcu based solution.  I think the cost at reader
> side should be almost same between rcu and srcu?  To use srcu, we need
> to select CONFIG_SRCU when CONFIG_SWAP is enabled in Kconfig.  I think
> that should be OK?

The thing to do is to try SRCU and see if you can see significant
performance degradation.  Given that there is swapping involved, I
would be surprised if the added read-side overhead of SRCU was even
measurable, but then again I have been surprised before.

And yes, just select CONFIG_SRCU when you need it.

							Thanx, Paul

Powered by blists - more mailing lists