Message-ID: <20210531073252.GC56979@shbuild999.sh.intel.com>
Date:   Mon, 31 May 2021 15:32:52 +0800
From:   Feng Tang <feng.tang@...el.com>
To:     Michal Hocko <mhocko@...e.com>
Cc:     linux-mm@...ck.org, Andrew Morton <akpm@...ux-foundation.org>,
        David Rientjes <rientjes@...gle.com>,
        Dave Hansen <dave.hansen@...el.com>,
        Ben Widawsky <ben.widawsky@...el.com>,
        linux-kernel@...r.kernel.org,
        Andrea Arcangeli <aarcange@...hat.com>,
        Mel Gorman <mgorman@...hsingularity.net>,
        Mike Kravetz <mike.kravetz@...cle.com>,
        Randy Dunlap <rdunlap@...radead.org>,
        Vlastimil Babka <vbabka@...e.cz>,
        Andi Kleen <ak@...ux.intel.com>,
        Dan Williams <dan.j.williams@...el.com>, ying.huang@...el.com
Subject: Re: [PATCH v1 4/4] mm/mempolicy: kill MPOL_F_LOCAL bit

On Mon, May 31, 2021 at 09:00:25AM +0200, Michal Hocko wrote:
> On Fri 28-05-21 12:39:54, Feng Tang wrote:
> > On Thu, May 27, 2021 at 05:34:56PM +0200, Michal Hocko wrote:
> > > On Thu 27-05-21 21:34:36, Feng Tang wrote:
> > > > On Thu, May 27, 2021 at 02:26:24PM +0200, Michal Hocko wrote:
> > > > > On Thu 27-05-21 20:10:41, Feng Tang wrote:
> > > > > > On Thu, May 27, 2021 at 10:20:08AM +0200, Michal Hocko wrote:
> > > > > > > On Wed 26-05-21 13:01:42, Feng Tang wrote:
> > > > > > > > Now the only remaining case of a real 'local' policy faked by
> > > > > > > > 'prefer' policy plus MPOL_F_LOCAL bit is:
> > > > > > > > 
> > > > > > > > A valid 'prefer' policy with a valid 'preferred' node is 'rebind'
> > > > > > > > to a nodemask which doesn't contains the 'preferred' node, then it
> > > > > > > > will handle allocation with 'local' policy.
> > > > > > > > 
> > > > > > > > Add a new 'MPOL_F_LOCAL_TEMP' bit for this case, and kill the
> > > > > > > > MPOL_F_LOCAL bit, which could simplify the code much.
> > > > > > > 
> > > > > > > As I've pointed out in the reply to the previous patch. It would have
> > > > > > > been much better if most of the MPOL_F_LOCAL usage was gone by this
> > > > > > > patch.
> > > > > > > 
> > > > > > > I also dislike a new MPOL_F_LOCAL_TEMP. This smells like sneaking the
> > > > > > > hack back in after you have painstakingly removed it. So this looks like
> > > > > > > a step backwards to me. I also do not understand why do we need the
> > > > > > > rebind callback for local policy at all. There is no node mask for local
> > > > > > > so what is going on here?
> > > > > > 
> > > > > > This is the special case 4 for 'perfer' policy with MPOL_F_STATIC_NODES
> > > > > > flag set, say it prefer node 1, when it is later 'refind' to a new
> > > > > > nodemask node 2-3, according to current code it will be add the
> > > > > > MPOL_F_LOCAL bit and performs 'local' policy acctually. And in future
> > > > > > it is 'rebind' again with a nodemask 1-2, it will be restored back
> > > > > > to 'prefer' policy with preferred node 1.
> > > > > 
> > > > > Honestly I still do not follow the actual problem. 
> > > > 
> > > > I was confused too, and don't know the original thought behind it. This
> > > > case 4 was just imagined by reading the code.
> > > > 
> > > > > A preferred node is a
> > > > > _hint_. If you rebind the task to a different cpuset then why should we
> > > > > actually care? The allocator will fallback to the closest node according
> > > > > to the distance metric. Maybe the original code was trying to handle
> > > > > that in some way but I really do fail to understand that code and I
> > > > > strongly suspect it is more likely to overengineered rather than backed
> > > > > by a real usecase. I might be wrong here but then this is an excellent
> > > > > opportunity to clarify all those subtleties.
> > > > 
> > > > From the code, the original special handling may be needed in 3 cases:
> > > >     get_policy_nodemask()
> > > >     policy_node()
> > > >     mempolicy_slab_node()
> > > > to not return the preset prefer_nid.
> > > 
> > > I am sorry but I do not follow. What is actually wrong if the preferred
> > > node is outside of the cpuset nodemask?
> > 
> > Sorry, I didn't make it clear. With current code logic, it will perform
> > as 'local' policy, but its mode is kept as 'prefer', so the code still
> > has these tricky bit checking when these APIs are called for this policy.
> > I agree with you that these ping-pong rebind() may be over engineering,
> > so for this case can we just change the policy from 'prefer' to 'local',
> > and drop the tricky bit manipulation, as the 'prefer' is just a hint,
> > if these rebind misses the target node, there is no need to stick with
> > the 'prefer' policy?
> 
> Again. I really do not understand why we should rebind or mark as local
> anything here. Is this a documented/expected behavior? What if somebody
> just changes the cpuset to include the preferred node again. Is it
> expected to have local preference now?

Good point! Marking 'local' doesn't solve the whole issue. And I didn't 
find any document defining the semantics.

> I can see you have posted a newer version which I haven't seen yet but
> this is really better to get resolved before building up more on top.
> And let me be explicit. I do believe that rebinding preferred policy is
> just bogus and it should be dropped altogether on the ground that a 
> preference is a mere hint from userspace where to start the allocation. 

Yes, the current mpol_rebind_preferred()'s logic is confusing. Let me
try to understand it correctly, are you suggesting to do nothing for
'prefer's rebinding regarding MPOL_F_STATIC_NODES and MPOL_F_RELATIVE_NODES,
while just setting 'pol->w.cpuset_mems_allowed' to the new nodemask?

Thanks,
Feng

> Unless I am missing something cpusets will be always authoritative for
> the final placement. The preferred node just acts as a starting point
> and it should be really preserved when cpusets changes. Otherwise we
> have a very subtle behavior corner cases.
> -- 
> Michal Hocko
> SUSE Labs

Powered by blists - more mailing lists