Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
Date:   Wed, 25 May 2022 08:30:15 -0400
From:   Johannes Weiner <hannes@cmpxchg.org>
To:     Muchun Song <songmuchun@bytedance.com>
Cc:     mhocko@kernel.org, roman.gushchin@linux.dev, shakeelb@google.com,
        cgroups@vger.kernel.org, linux-mm@kvack.org,
        linux-kernel@vger.kernel.org, duanxiongchun@bytedance.com,
        longman@redhat.com
Subject: Re: [PATCH v4 03/11] mm: memcontrol: make lruvec lock safe when LRU
 pages are reparented
Message-ID: <Yo4hVw7B+bUlMzLX@cmpxchg.org>
References: <20220524060551.80037-1-songmuchun@bytedance.com>
 <20220524060551.80037-4-songmuchun@bytedance.com>
 <Yo0xmKOkBkhRy+bq@cmpxchg.org>
 <Yo38mlkMBFz2h+yP@FVFYT0MHHV2J.googleapis.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <Yo38mlkMBFz2h+yP@FVFYT0MHHV2J.googleapis.com>
Precedence: bulk

On Wed, May 25, 2022 at 05:53:30PM +0800, Muchun Song wrote:
> On Tue, May 24, 2022 at 03:27:20PM -0400, Johannes Weiner wrote:
> > On Tue, May 24, 2022 at 02:05:43PM +0800, Muchun Song wrote:
> > > The diagram below shows how to make the folio lruvec lock safe when LRU
> > > pages are reparented.
> > > 
> > > folio_lruvec_lock(folio)
> > >     retry:
> > > 	lruvec = folio_lruvec(folio);
> > > 
> > >         // The folio is reparented at this time.
> > >         spin_lock(&lruvec->lru_lock);
> > > 
> > >         if (unlikely(lruvec_memcg(lruvec) != folio_memcg(folio)))
> > >             // Acquired the wrong lruvec lock and need to retry.
> > >             // Because this folio is on the parent memcg lruvec list.
> > >             goto retry;
> > > 
> > >         // If we reach here, it means that folio_memcg(folio) is stable.
> > > 
> > > memcg_reparent_objcgs(memcg)
> > >     // lruvec belongs to memcg and lruvec_parent belongs to parent memcg.
> > >     spin_lock(&lruvec->lru_lock);
> > >     spin_lock(&lruvec_parent->lru_lock);
> > > 
> > >     // Move all the pages from the lruvec list to the parent lruvec list.
> > > 
> > >     spin_unlock(&lruvec_parent->lru_lock);
> > >     spin_unlock(&lruvec->lru_lock);
> > > 
> > > After we acquire the lruvec lock, we need to check whether the folio is
> > > reparented. If so, we need to reacquire the new lruvec lock. On the
> > > routine of the LRU pages reparenting, we will also acquire the lruvec
> > > lock (will be implemented in the later patch). So folio_memcg() cannot
> > > be changed when we hold the lruvec lock.
> > > 
> > > Since lruvec_memcg(lruvec) is always equal to folio_memcg(folio) after
> > > we hold the lruvec lock, lruvec_memcg_debug() check is pointless. So
> > > remove it.
> > > 
> > > This is a preparation for reparenting the LRU pages.
> > > 
> > > Signed-off-by: Muchun Song <songmuchun@bytedance.com>
> > 
> > This looks good to me. Just one question:
> > 
> > > @@ -1230,10 +1213,23 @@ void lruvec_memcg_debug(struct lruvec *lruvec, struct folio *folio)
> > >   */
> > >  struct lruvec *folio_lruvec_lock(struct folio *folio)
> > >  {
> > > -	struct lruvec *lruvec = folio_lruvec(folio);
> > > +	struct lruvec *lruvec;
> > >  
> > > +	rcu_read_lock();
> > > +retry:
> > > +	lruvec = folio_lruvec(folio);
> > >  	spin_lock(&lruvec->lru_lock);
> > > -	lruvec_memcg_debug(lruvec, folio);
> > > +
> > > +	if (unlikely(lruvec_memcg(lruvec) != folio_memcg(folio))) {
> > > +		spin_unlock(&lruvec->lru_lock);
> > > +		goto retry;
> > > +	}
> > > +
> > > +	/*
> > > +	 * Preemption is disabled in the internal of spin_lock, which can serve
> > > +	 * as RCU read-side critical sections.
> > > +	 */
> > > +	rcu_read_unlock();
> > 
> > The code looks right to me, but I don't understand the comment: why do
> > we care that the rcu read-side continues? With the lru_lock held,
> > reparenting is on hold and the lruvec cannot be rcu-freed anyway, no?
> >
> 
> Right. We could hold rcu read lock until end of reparting.  So you mean
> we do rcu_read_unlock in folio_lruvec_lock()?

The comment seems to suggest that disabling preemption is what keeps
the lruvec alive. But it's the lru_lock that keeps it alive. The
cgroup destruction path tries to take the lru_lock long before it even
gets to synchronize_rcu(). Once you hold the lru_lock, having an
implied read-side critical section as well doesn't seem to matter.

Should the comment be deleted?