Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
Date:   Mon, 29 Jun 2020 10:03:50 +0200
From:   Michal Hocko <mhocko@kernel.org>
To:     Joonsoo Kim <js1304@gmail.com>
Cc:     Andrew Morton <akpm@linux-foundation.org>,
        Linux Memory Management List <linux-mm@kvack.org>,
        LKML <linux-kernel@vger.kernel.org>, kernel-team@lge.com,
        Vlastimil Babka <vbabka@suse.cz>,
        Christoph Hellwig <hch@infradead.org>,
        Roman Gushchin <guro@fb.com>,
        Mike Kravetz <mike.kravetz@oracle.com>,
        Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>,
        Joonsoo Kim <iamjoonsoo.kim@lge.com>
Subject: Re: [PATCH v3 5/8] mm/migrate: make a standard migration target
 allocation function
Message-ID: <20200629080350.GB32461@dhcp22.suse.cz>
References: <1592892828-1934-1-git-send-email-iamjoonsoo.kim@lge.com>
 <1592892828-1934-6-git-send-email-iamjoonsoo.kim@lge.com>
 <20200625120550.GF1320@dhcp22.suse.cz>
 <CAAmzW4ObN=GAzCLw8betLftTeCEsLs4OihfNXvtg4CaWyWiBCw@mail.gmail.com>
 <20200626073342.GU1320@dhcp22.suse.cz>
 <CAAmzW4PT+M8mUwXMPhQSKLhz9nkPATivL5D-yLyyfC9SOwPnYQ@mail.gmail.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=utf-8
Content-Disposition: inline
Content-Transfer-Encoding: 8bit
In-Reply-To: <CAAmzW4PT+M8mUwXMPhQSKLhz9nkPATivL5D-yLyyfC9SOwPnYQ@mail.gmail.com>
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

On Mon 29-06-20 15:41:37, Joonsoo Kim wrote:
> 2020년 6월 26일 (금) 오후 4:33, Michal Hocko <mhocko@kernel.org>님이 작성:
> >
> > On Fri 26-06-20 14:02:49, Joonsoo Kim wrote:
> > > 2020년 6월 25일 (목) 오후 9:05, Michal Hocko <mhocko@kernel.org>님이 작성:
> > > >
> > > > On Tue 23-06-20 15:13:45, Joonsoo Kim wrote:
> > [...]
> > > > > -struct page *new_page_nodemask(struct page *page,
> > > > > -                             int preferred_nid, nodemask_t *nodemask)
> > > > > +struct page *alloc_migration_target(struct page *page, unsigned long private)
> > > > >  {
> > > > > -     gfp_t gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL;
> > > > > +     struct migration_target_control *mtc;
> > > > > +     gfp_t gfp_mask;
> > > > >       unsigned int order = 0;
> > > > >       struct page *new_page = NULL;
> > > > > +     int zidx;
> > > > > +
> > > > > +     mtc = (struct migration_target_control *)private;
> > > > > +     gfp_mask = mtc->gfp_mask;
> > > > >
> > > > >       if (PageHuge(page)) {
> > > > >               return alloc_huge_page_nodemask(
> > > > > -                             page_hstate(compound_head(page)),
> > > > > -                             preferred_nid, nodemask, 0, false);
> > > > > +                             page_hstate(compound_head(page)), mtc->nid,
> > > > > +                             mtc->nmask, gfp_mask, false);
> > > > >       }
> > > > >
> > > > >       if (PageTransHuge(page)) {
> > > > > +             gfp_mask &= ~__GFP_RECLAIM;
> > > >
> > > > What's up with this gfp_mask modification?
> > >
> > > THP page allocation uses a standard gfp masks, GFP_TRANSHUGE_LIGHT and
> > > GFP_TRANHUGE. __GFP_RECLAIM flags is a big part of this standard mask design.
> > > So, I clear it here so as not to disrupt the THP gfp mask.
> >
> > Why this wasn't really needed before? I guess I must be missing
> > something here. This patch should be mostly mechanical convergence of
> > existing migration callbacks but this change adds a new behavior AFAICS.
> 
> Before this patch, a user cannot specify a gfp_mask and THP allocation
> uses GFP_TRANSHUGE
> statically.

Unless I am misreading there are code paths (e.g.new_page_nodemask) which simply use
add GFP_TRANSHUGE to GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL. And
this goes all the way to thp migration introduction.

> After this patch, a user can specify a gfp_mask and it
> could conflict with GFP_TRANSHUGE.
> This code tries to avoid this conflict.
> 
> > It would effectively drop __GFP_RETRY_MAYFAIL and __GFP_KSWAPD_RECLAIM
> 
> __GFP_RETRY_MAYFAIL isn't dropped. __GFP_RECLAIM is
> "___GFP_DIRECT_RECLAIM|___GFP_KSWAPD_RECLAIM".
> So, __GFP_KSWAPD_RECLAIM would be dropped for THP allocation.
> IIUC, THP allocation doesn't use __GFP_KSWAPD_RECLAIM since it's
> overhead is too large and this overhead should be given to the caller
> rather than system thread (kswapd) and so on.

Yes, there is a reason why KSWAPD is excluded from THP allocations in
the page fault path. Maybe we want to extend that behavior to the
migration as well. I do not have a strong opinion on that because I
haven't seen excessive kswapd reclaim due to THP migrations. They are
likely too rare.

But as I've said in my previous email. Make this a separate patch with
an explanation why we want this.
-- 
Michal Hocko
SUSE Labs