[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <c877a136-4294-4f00-b0ac-7194fe170452@redhat.com>
Date: Tue, 2 Jul 2024 16:57:44 +0200
From: David Hildenbrand <david@...hat.com>
To: Ryan Roberts <ryan.roberts@....com>,
Andrew Morton <akpm@...ux-foundation.org>, Jonathan Corbet <corbet@....net>,
Barry Song <baohua@...nel.org>, Baolin Wang <baolin.wang@...ux.alibaba.com>,
Lance Yang <ioworker0@...il.com>, Yang Shi <shy828301@...il.com>
Cc: linux-mm@...ck.org, linux-kernel@...r.kernel.org, stable@...r.kernel.org
Subject: Re: [PATCH v1] mm: Fix khugepaged activation policy
On 02.07.24 16:46, Ryan Roberts wrote:
> Since the introduction of mTHP, the docuementation has stated that
> khugepaged would be enabled when any mTHP size is enabled, and disabled
> when all mTHP sizes are disabled. There are 2 problems with this; 1.
> this is not what was implemented by the code and 2. this is not the
> desirable behavior.
>
> Desirable behavior is for khugepaged to be enabled when any PMD-sized
> THP is enabled, anon or file. (Note that file THP is still controlled by
> the top-level control so we must always consider that, as well as the
> PMD-size mTHP control for anon). khugepaged only supports collapsing to
> PMD-sized THP so there is no value in enabling it when PMD-sized THP is
> disabled. So let's change the code and documentation to reflect this
> policy.
>
> Further, per-size enabled control modification events were not
> previously forwarded to khugepaged to give it an opportunity to start or
> stop. Consequently the following was resulting in khugepaged eroneously
> not being activated:
>
> echo never > /sys/kernel/mm/transparent_hugepage/enabled
> echo always > /sys/kernel/mm/transparent_hugepage/hugepages-2048kB/enabled
>
> Signed-off-by: Ryan Roberts <ryan.roberts@....com>
> Fixes: 3485b88390b0 ("mm: thp: introduce multi-size THP sysfs interface")
> Closes: https://lore.kernel.org/linux-mm/7a0bbe69-1e3d-4263-b206-da007791a5c4@redhat.com/
> Cc: stable@...r.kernel.org
> ---
>
> Hi All,
>
> Applies on top of today's mm-unstable (9bb8753acdd8). No regressions observed in
> mm selftests.
>
> When fixing this I also noticed that khugepaged doesn't get (and never has been)
> activated/deactivated by `shmem_enabled=`. I'm not sure if khugepaged knows how
> to collapse shmem - perhaps it should be activated in this case?
>
Call me confused.
khugepaged_scan_mm_slot() and madvise_collapse() only all
hpage_collapse_scan_file() with ... IS_ENABLED(CONFIG_SHMEM) ?
collapse_file() is only called by hpage_collapse_scan_file() ... and
there we check "shmem_file(file)".
So why is the IS_ENABLED(CONFIG_SHMEM) check in there if collapse_file()
seems to "collapse filemap/tmpfs/shmem pages into huge one".
Anyhow, we certainly can collapse shmem (that's how it all started IIUC).
Besides that, khugepaged only seems to collapse !shmem with
VM_BUG_ON(!IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) && !is_shmem);
The thp_vma_allowable_order() check tests if we are allowed to collapse
a PMD_ORDER in that VMA.
--
Cheers,
David / dhildenb
Powered by blists - more mailing lists