Re: [PATCH v15 4/7] mm: Introduce Reported pages

From: Mel Gorman <mgorman@techsingularity.net>
To: Alexander Duyck <alexander.h.duyck@linux.intel.com>
Cc: Nitesh Narayan Lal <nitesh@redhat.com>,
	Alexander Duyck <alexander.duyck@gmail.com>,
	kvm@vger.kernel.org, mst@redhat.com,
	linux-kernel@vger.kernel.org, willy@infradead.org,
	mhocko@kernel.org, linux-mm@kvack.org, akpm@linux-foundation.org,
	vbabka@suse.cz, yang.zhang.wz@gmail.com, konrad.wilk@oracle.com,
	david@redhat.com, pagupta@redhat.com, riel@surriel.com,
	lcapitulino@redhat.com, dave.hansen@intel.com,
	wei.w.wang@intel.com, aarcange@redhat.com, pbonzini@redhat.com,
	dan.j.williams@intel.com, osalvador@suse.de
Subject: Re: [PATCH v15 4/7] mm: Introduce Reported pages
Date: Wed, 18 Dec 2019 07:31:39 +0000	[thread overview]
Message-ID: <20191218073139.GE3178@techsingularity.net> (raw)
In-Reply-To: <03e1e95c2cc8d6e3206212df48a971e9696d3b20.camel@linux.intel.com>

On Tue, Dec 17, 2019 at 08:31:59AM -0800, Alexander Duyck wrote:
> > > > I think you recently switched to using an atomic variable for maintaining page
> > > > reporting status as I was doing in v12.
> > > > Which is good, as we will not have a disagreement on it now.
> > >
> > > There is still some differences between our approaches if I am not
> > > mistaken. Specifically I have code in place so that any requests to report
> > > while we are actively working on reporting will trigger another pass being
> > > scheduled after we completed. I still believe you were lacking any logic
> > > like that as I recall.
> > > 
> > 
> > Yes, I was specifically referring to the atomic state variable.
> > Though I am wondering if having an atomic variable to track page reporting state
> > is better than having a page reporting specific unsigned long flag, which we can
> > manipulate via __set_bit() and __clear_bit().
> 
> So the reason for using an atomic state variable is because I only really
> have 3 possible states; idle, active, and requested. It allows for a
> pretty simple state machine as any transition from idle indicates that we
> need to schedule the worker, transition from requested to active when the
> worker starts, and if at the end of a pass if we are still in the active
> state it means we can transition back to idle, otherwise we reschedule the
> worker.
> 
> In order to do the same sort of thing using the bitops would require at
> least 2 bits. In addition with the requirement that I cannot use the zone
> lock for protection of the state I cannot use the non-atomic versions of
> things such as __set_bit and __clear_bit so they would require additional
> locking protections.
> 

I completely agree with this. I had pointed out in an earlier review
that expanding the scope of the zone lock was inappropriate, the
non-atomic operations on separate flags potentially misses updates and
in general, I prefer the atomic variable approach a lot more than the
previous zone->flag based approach.

-- 
Mel Gorman
SUSE Labs