Re: [PATCH] ARM: mm: fix no-MMU ZERO_PAGE() implementation

From: "Arnd Bergmann" <arnd@arndb.de>
To: "Giulio Benetti" <giulio.benetti@benettiengineering.com>,
	linux-arm-kernel@lists.infradead.org,
	linux-kernel@vger.kernel.org
Cc: "Russell King" <linux@armlinux.org.uk>,
	"Anshuman Khandual" <anshuman.khandual@arm.com>,
	"Andrew Morton" <akpm@linux-foundation.org>,
	"Kefeng Wang" <wangkefeng.wang@huawei.com>,
	"Russell King" <rmk+kernel@armlinux.org.uk>,
	"Will Deacon" <will@kernel.org>
Subject: Re: [PATCH] ARM: mm: fix no-MMU ZERO_PAGE() implementation
Date: Wed, 19 Oct 2022 09:00:25 +0200	[thread overview]
Message-ID: <e98b0727-5a5b-4f1d-bbf6-1194e6df299a@app.fastmail.com> (raw)
In-Reply-To: <3fc711c8-4981-26f7-689c-549bdafa40ac@benettiengineering.com>

On Wed, Oct 19, 2022, at 00:32, Giulio Benetti wrote:
> On 18/10/22 20:35, Arnd Bergmann wrote:
>> On Tue, Oct 18, 2022, at 19:44, Giulio Benetti wrote:
>>> On 18/10/22 09:03, Arnd Bergmann wrote:
>>>> In addition to your fix, I see that arm is the only architecture
>>>> that defines 'empty_zero_page' as a pointer to the page, when
>>>> everything else just makes it a pointer to the data itself,
>>>> or an 'extern char empty_zero_page[]' array, which we may want
>>>> to change for consistency.
>>>
>>> I was about doing it, but then I tought to move one piece at a time.
>> 
>> Right, it would definitely be a separate patch, but it
>> can be a series of two patches. We probably wouldn't need to
>> backport the second patch that turns it into a static allocation.
>
> I've sent the patchset of 2:
> https://lore.kernel.org/all/20221018222503.90118-1-giulio.benetti@benettiengineering.com/T/#t
>
> I'm wondering if it makes sense to send a patchset for all those
> architectures that have only one zero page. I've seen that for example
> loongarch has more than one. But for the others I find the array
> approach more linear, with less code all around and a bit faster in term
> of code execution(of course really few, but better than nothing) since
> that array is in .bss, so it will be zeroed earlier during a long
> "memset" where assembly instructions for zeroing 8 bytes at a time are
> used. What about this?

The initial zeroing should not matter at all in terms of performance,
I think the only question is whether one wants a single zero page
to be used everywhere or one per NUMA node to give better locality
for a cache miss.

My guess is that for a system with 4KB pages, all the data
in the zero page are typically available in a CPU cache already,
so it doesn't matter, but it's possible that some machines benefit
from having per-node pages when the page size isn't tiny compared
to the typical cache sizes.

We should probably not touch this for any of the other architectures.

     Arnd