Received-SPF: pass (google.com: domain of linux-kernel+bounces-137810-linux.lists.archive=gmail.com@vger.kernel.org designates 2604:1380:45e3:2400::1 as permitted sender) client-ip=2604:1380:45e3:2400::1;
From: "Huang, Ying" <ying.huang@intel.com>
To: Nhat Pham <nphamcs@gmail.com>
Cc: Zhaoyu Liu <liuzhaoyu.zackary@bytedance.com>,  Andrew Morton
 <akpm@linux-foundation.org>,  ryncsn@gmail.com,  songmuchun@bytedance.com,
  david@redhat.com,  chrisl@kernel.org,  guo.ziliang@zte.com.cn,
  yosryahmed@google.com,  linux-kernel@vger.kernel.org,  linux-mm@kvack.org
Subject: Re: [PATCH v2] mm: swap: prejudgement swap_has_cache to avoid page
 allocation
In-Reply-To: <CAKEwX=MBts2mGgTE__VP-ZVMrMFTzQnbTAkMPTJs3KNRQ2QDjg@mail.gmail.com>
	(Nhat Pham's message of "Tue, 9 Apr 2024 10:52:40 -0700")
References: <20240408121439.GA252652@bytedance>
	<20240408132704.f966adc8d3928df4d3b8c0a9@linux-foundation.org>
	<87edbf8hta.fsf@yhuang6-desk2.ccr.corp.intel.com>
	<20240409145740.GA543696@bytedance>
	<CAKEwX=MBts2mGgTE__VP-ZVMrMFTzQnbTAkMPTJs3KNRQ2QDjg@mail.gmail.com>
Date: Wed, 10 Apr 2024 09:46:18 +0800
Message-ID: <8734ru6lcl.fsf@yhuang6-desk2.ccr.corp.intel.com>
User-Agent: Gnus/5.13 (Gnus v5.13)
Precedence: bulk
MIME-Version: 1.0
Content-Type: text/plain; charset=utf-8
Content-Transfer-Encoding: quoted-printable

Nhat Pham <nphamcs@gmail.com> writes:

> On Tue, Apr 9, 2024 at 7:57=E2=80=AFAM Zhaoyu Liu
> <liuzhaoyu.zackary@bytedance.com> wrote:
>>
>> On Tue, Apr 09, 2024 at 09:07:29AM +0800, Huang, Ying wrote:
>> > Andrew Morton <akpm@linux-foundation.org> writes:
>> >
>> > > On Mon, 8 Apr 2024 20:14:39 +0800 Zhaoyu Liu <liuzhaoyu.zackary@byte=
dance.com> wrote:
>> > >
>> > >> Based on qemu arm64 - latest kernel + 100M memory + 1024M swapfile.
>> > >> Create 1G anon mmap and set it to shared, and has two processes
>> > >> randomly access the shared memory. When they are racing on swap cac=
he,
>> > >> on average, each "alloc_pages_mpol + swapcache_prepare + folio_put"
>> > >> took about 1475 us.
>> > >
>> > > And what effect does this patch have upon the measured time?  ANd up=
on
>> > > overall runtime?
>> >
>> > And the patch will cause increased lock contention, please test with
>> > more processes and perhaps HDD swap device too.
>>
>> Hi Ying,
>>
>> Thank you for your suggestion.
>> It may indeed cause some lock contention, as mentioned by Kairui before.
>>
>> If so, is it recommended?
>> ---
>>   unsigned char swap_map, mapcount, hascache;
>>   ...
>>   /* Return raw data of the si->swap_map[offset] */
>>   swap_map =3D __swap_map(si, entry);
>>   mapcount =3D swap_map & ~SWAP_HAS_CACHE;
>>   if (!mapcount && swap_slot_cache_enabled)
>>   ...
>>   hascache =3D swap_map & SWAP_HAS_CACHE;
>>   /* Could judge that it's being added to swap cache with high probabili=
ty */
>>   if (mapcount && hascache)
>>     goto skip_alloc;
>>   ...
>> ---
>> In doing so, there is no additional use of locks.
>>
>
> Hmm so is this a lockless check now? Ummmm... Could someone with more
> expertise in the Linux kernel memory model double check that this is
> even a valid state we're observing here? Looks like we're performing
> an unguarded, unsynchronized, non-atomic read with the possibility of
> concurrent write - is there a chance we might see partial/invalid
> results?
>
> Could you also test with zswap enabled (and perhaps with zswap
> shrinker enabled)?

READ_ONCE() will save us from partial/invalid results.

--
Best Regards,
Huang, Ying