Received-SPF: pass (google.com: domain of linux-kernel+bounces-52154-linux.lists.archive=gmail.com@vger.kernel.org designates 2604:1380:45e3:2400::1 as permitted sender) client-ip=2604:1380:45e3:2400::1;
Content-Type: text/plain;
	charset=us-ascii
Precedence: bulk
Mime-Version: 1.0
Subject: Re: [PATCH 1/1] hugetlb: fix CONFIG_PADATA dependency for non-SMP
 system
From: Muchun Song <muchun.song@linux.dev>
In-Reply-To: <828f990c-11af-42ad-a030-a66dde97a7f2@linux.dev>
Date: Mon, 5 Feb 2024 15:37:41 +0800
Cc: Gang Li <ligang.bdlg@bytedance.com>,
 Andrew Morton <akpm@linux-foundation.org>,
 linux-kernel@vger.kernel.org,
 linux-mm@kvack.org,
 Randy Dunlap <rdunlap@infradead.org>,
 kernel test robot <lkp@intel.com>
Content-Transfer-Encoding: quoted-printable
Message-Id: <7472563F-5C2D-4DCB-ACD6-F86D7A18BDF2@linux.dev>
References: <20240204072525.1986626-1-gang.li@linux.dev>
 <f05f658a-78fa-45cd-ad07-11d87b824702@linux.dev>
 <a749cb06-9495-40fc-9192-b019b48eca92@linux.dev>
 <828f990c-11af-42ad-a030-a66dde97a7f2@linux.dev>
To: Gang Li <gang.li@linux.dev>


> On Feb 5, 2024, at 14:55, Gang Li <gang.li@linux.dev> wrote:
>=20
>=20
> On 2024/2/4 15:48, Gang Li wrote:
>> On 2024/2/4 15:44, Muchun Song wrote:
>>> I don't think it is a clear way to fix this. If someone want to
>>> use PADATA in a non-SMP system, he should be carefully to handle
>>> the non-SMP case himself. I think the better way is to make PADATA
>>> handle the non-SMP case, I think it should be easy for it, which
>>> could just call ->thread_fn() many times instead of creating many
>>> threads in the non-SMP case.
>>>=20
>>> Thanks.
>>>=20
>> Sounds good, I'll take a look at padata and send a new patch.
>=20
> 1. delete the dependency on SMP
>=20
> PADATA only depends on workqueue and completion. It works well with =
!SMP
> currently but has no performance benefits. What we can do is make =
PADATA
> handle the non-SMP case more elegantly.
>=20
> PADATA has two parts: "Running Multithreaded Jobs" and "Running
> Serialized Jobs".
>=20
> "Running Multithreaded Jobs", which hugetlb parallelization relies on
> can be easily deparallelize through this patch:
>=20
> ```
> @@ -495,7 +495,7 @@ void __init padata_do_multithreaded(struct =
padata_mt_job *job)
>       nworks =3D max(job->size / max(job->min_chunk, job->align), =
1ul);
>       nworks =3D min(nworks, job->max_threads);
>=20
> -      if (nworks =3D=3D 1) {
> +      if (nworks =3D=3D 1 || !IS_ENABLED(CONFIG_SMP)) {
>               /* Single thread, no coordination needed, cut to the =
chase. */
>               job->thread_fn(job->start, job->start + job->size, =
job->fn_arg);
>               return;
> ```
>=20
> However, "Running Serialized Jobs" is more challenging due to its
> various workers queuing each other, making it more complex than =
"Running
> Multithreaded Jobs." I am currently in the process of deciphering the
> code.

Actually, I did not get it. Why the above code cannot work? The above
code already make it serialized in one call, right? What do I miss here?

Thanks.

>=20
> To eliminate kconfig warnings, other methods could be considered:
>=20
> 2. Split hugetlb parallelization into a separate kconfig.
> 3. Wrap hugetlb parallelization with SMP or PADATA macros (already =
ruled out).
> 4. Split PADATA into PADATA_SERIALIZED and PADATA_MULTITHREADED (too =
heavy).
>=20
> Anyway, this is only FYI. I will continue exploring how to =
deparallelize
> "Running Serialized Jobs."