Re: [PATCH 1/2] mm: remove pfn_valid_within() and CONFIG_HOLES_IN_ZONE

From: Zi Yan <ziy@nvidia.com>
To: David Hildenbrand <david@redhat.com>
Cc: Mike Rapoport <rppt@kernel.org>,
	Andrew Morton <akpm@linux-foundation.org>,
	Greg Kroah-Hartman <gregkh@linuxfoundation.org>,
	Mike Rapoport <rppt@linux.ibm.com>,
	"Rafael J. Wysocki" <rafael@kernel.org>,
	linux-kernel@vger.kernel.org, linux-mm@kvack.org
Subject: Re: [PATCH 1/2] mm: remove pfn_valid_within() and CONFIG_HOLES_IN_ZONE
Date: Tue, 13 Jul 2021 16:02:24 -0400	[thread overview]
Message-ID: <CE216FEB-1C75-45E1-8175-3DFA72411E3A@nvidia.com> (raw)
In-Reply-To: <4367d5e4-4f03-6d99-f19b-9d32b71f227d@redhat.com>

[-- Attachment #1: Type: text/plain, Size: 2086 bytes --]

On 13 Jul 2021, at 6:24, David Hildenbrand wrote:

> On 13.07.21 12:22, Mike Rapoport wrote:
>> On Tue, Jul 13, 2021 at 11:51:46AM +0200, David Hildenbrand wrote:
>>> On 13.07.21 10:00, Mike Rapoport wrote:
>>>> From: Mike Rapoport <rppt@linux.ibm.com>
>>>>
>>>> After recent changes in freeing of the unused parts of the memory map and
>>>> rework of pfn_valid() in arm and arm64 there are no architectures that can
>>>> have holes in the memory map within a pageblock and so nothing can enable
>>>> CONFIG_HOLES_IN_ZONE which guards non trivial implementation of
>>>> pfn_valid_within().
>>>>
>>>> With that, pfn_valid_within() is always hardwired to 1 and can be
>>>> completely removed.
>>>>
>>>> Remove calls to pfn_valid_within() and CONFIG_HOLES_IN_ZONE.
>>>>
>>>> Signed-off-by: Mike Rapoport <rppt@linux.ibm.com>
>>>
>>> There is currently the discussion to increase MAX_ORDER, for example, to
>>> cover 1GiB instead of 4MiB on x86-64. This would mean that we could
>>> suddenly, again, have holes insides MAX_ORDER - 1 pages.
>>>
>>> So I assume if we ever go down that path, we'll need something like this
>>> again.
>>
>> It depends whether pageblock_order will be also increased. PFN walkers rely
>> on continuity of pageblocks rather than MAX_ORDER chunks, so if
>> pageblock_order won't change, there won't be need to check for pfn_valid()
>> inside a pageblock.
>
> I'm pushing for letting pageblocks stay untouched, so good to know!
>
> (we still have this crazy special case of pageblocks > MAX_ORDER - 1 right now, which I think we should just eliminate)

I am working on this right now. After I increase MAX_ORDER above SECTION_SIZE_BITS,
holes can appear in a zone even on x86_64 and pfn_valid_within(), which is currently
set to 1, is needed to be pfn_valid(). That seems to be sufficient to make increasing
MAX_ORDER work.

For my use case, I might revive pfn_valid_within() if MAX_ORDER is configured to be
above SECTION_SIZE_BITS. Let me know if there is any better alternative. Thanks.

—
Best Regards,
Yan, Zi

[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 854 bytes --]