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:   Tue, 15 Dec 2020 17:18:09 -0700
From:   Yu Zhao <yuzhao@google.com>
To:     Hugh Dickins <hughd@google.com>
Cc:     Andrew Morton <akpm@linux-foundation.org>,
        Alex Shi <alex.shi@linux.alibaba.com>,
        Michal Hocko <mhocko@kernel.org>,
        Johannes Weiner <hannes@cmpxchg.org>,
        Vladimir Davydov <vdavydov.dev@gmail.com>,
        Roman Gushchin <guro@fb.com>, Vlastimil Babka <vbabka@suse.cz>,
        Matthew Wilcox <willy@infradead.org>, linux-mm@kvack.org,
        linux-kernel@vger.kernel.org
Subject: Re: [PATCH 11/11] mm: enlarge the "int nr_pages" parameter of
 update_lru_size()
Message-ID: <X9lSQfI5SjRI/sz0@google.com>
References: <20201207220949.830352-1-yuzhao@google.com>
 <20201207220949.830352-12-yuzhao@google.com>
 <alpine.LSU.2.11.2012141317351.1925@eggly.anvils>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <alpine.LSU.2.11.2012141317351.1925@eggly.anvils>
Precedence: bulk

On Mon, Dec 14, 2020 at 01:50:16PM -0800, Hugh Dickins wrote:
> On Mon, 7 Dec 2020, Yu Zhao wrote:
> 
> > update_lru_sizes() defines an unsigned long argument and passes it as
> > nr_pages to update_lru_size(). Though this isn't causing any overflows
> > I'm aware of, it's a bad idea to go through the demotion given that we
> > have recently stumbled on a related type promotion problem fixed by
> > commit 2da9f6305f30 ("mm/vmscan: fix NR_ISOLATED_FILE corruption on 64-bit")
> > 
> > Note that the underlying counters are already in long. This is another
> > reason we shouldn't have the demotion.
> > 
> > This patch enlarges all relevant parameters on the path to the final
> > underlying counters:
> > 	update_lru_size(int -> long)
> > 		if memcg:
> > 			__mod_lruvec_state(int -> long)
> > 				if smp:
> > 					__mod_node_page_state(long)
> > 				else:
> > 					__mod_node_page_state(int -> long)
> > 			__mod_memcg_lruvec_state(int -> long)
> > 				__mod_memcg_state(int -> long)
> > 		else:
> > 			__mod_lruvec_state(int -> long)
> > 				if smp:
> > 					__mod_node_page_state(long)
> > 				else:
> > 					__mod_node_page_state(int -> long)
> > 
> > 		__mod_zone_page_state(long)
> > 
> > 		if memcg:
> > 			mem_cgroup_update_lru_size(int -> long)
> > 
> > Note that __mod_node_page_state() for the smp case and
> > __mod_zone_page_state() already use long. So this change also fixes
> > the inconsistency.
> > 
> > Signed-off-by: Yu Zhao <yuzhao@google.com>
> 
> NAK from me to this 11/11: I'm running happily with your 1-10 on top of
> mmotm (I'll review them n a few days, but currently more concerned with
> Rik's shmem huge gfp_mask), but had to leave this one out.
> 
> You think you are future-proofing with this, but it is present-breaking.
> 
> It looks plausible (though seems random: why these particular functions
> use long but others not? why __mod_memcg_state() long, mod_memcg_state()
> int?), and I was fooled; but fortunately was still testing with memcg
> moving, for Alex's patchset.

My apologies. The patch was fully tested on 4.15. Apparently I didn't
pay enough attention to what's changed in mem_cgroup_move_account() nor
had any test coverage on it when rebasing this patch.

> Soon got stuck waiting in balance_dirty_pages(), /proc/vmstat showing
> nr_anon_pages 2263142822377729
> nr_mapped 125095217474159
> nr_file_pages 225421358649526
> nr_dirty 8589934592
> nr_writeback 1202590842920
> nr_shmem 40501541678768
> nr_anon_transparent_hugepages 51539607554
> 
> That last (anon THPs) nothing to do with this patch, but illustrates
> what Muchun is fixing in his 1/7 "mm: memcontrol: fix NR_ANON_THPS
> accounting in charge moving".
> 
> The rest of them could be fixed by changing mem_cgroup_move_account()'s
> "unsigned int nr_pages" to "long nr_pages" in this patch, but I think
> it's safer just to drop the patch: the promotion of "unsigned int" to
> "long" does not work as you would like it to.
> 
> I see that mm/vmscan.c contains several "unsigned int" counts of pages,
> everything works fine at present so far as I know, and those appeared
> to work even with your patch; but I am not confident in my test coverage,
> and not confident in us being able to outlaw unsigned int page counts in
> future.

I'll just drop this one. Thanks.