Message-ID: <f366ce31-582c-4f90-bc32-05ddf3e71fa6@gmail.com>
Date: Wed, 25 Jun 2025 12:03:17 +0100
From: Usama Arif <usamaarif642@...il.com>
To: David Hildenbrand <david@...hat.com>,
 Lorenzo Stoakes <lorenzo.stoakes@...cle.com>, Hugh Dickins <hughd@...gle.com>
Cc: Baolin Wang <baolin.wang@...ux.alibaba.com>, akpm@...ux-foundation.org,
 ziy@...dia.com, Liam.Howlett@...cle.com, npache@...hat.com,
 ryan.roberts@....com, dev.jain@....com, baohua@...nel.org,
 zokeefe@...gle.com, shy828301@...il.com, linux-mm@...ck.org,
 linux-kernel@...r.kernel.org, Johannes Weiner <hannes@...xchg.org>
Subject: Re: [PATCH v4 0/2] fix MADV_COLLAPSE issue if THP settings are
 disabled



On 25/06/2025 08:34, David Hildenbrand wrote:
>>>
>>> We would all prefer a less messy world of THP tunables.  I certainly
>>> find plenty to dislike there too; and wish that a less assertive name
>>> than "never" had been chosen originally for the default off position.
>>>
>>> But please don't break the accepted and documented behaviour of
>>> MADV_COLLAPSE now.
>>
>> Again see above, I absolutely disagree this is documented _clearly_. And
>> that's the underlying issue here.
>> > I feel like if you polled 100 system administrators (assuming they knew
>> about THP) as to how you globally disable THP, probably all 100 would say
>> you do it via:
>>
>> # echo never > /sys/kernel/mm/transparent_hugepage/enabled
>>
> 
> Yes. One big problem is that the documentation was not updated.
> 
> Changing the meaning of "entirely disabled" to "entirely disabled automatically (page faults, khugepaged)"
> 
>> So shouldn't 'never break userspace' be based on practical reality rather
>> than a theorised interpretation of documents that sadly are not clear
>> enough?
> 
> I think the problem is that there might indeed be more users out there relying on "never+MADV_COLLPASE" to now place THPs than "never+MADV_COLLPASE" to no place THPs.
> 
> What is the harm when not placing THPs? Performance degradation for some apps?
> 

I think a bigger issue than performance degradation is someone upgrading the kernel and not
seeing MADV_COLLAPSE working as it has since the beginning and not knowing that its due
to a kernel change.

I feel transparent_hugepage/enabled is too messed up, and its difficult to fix it without
breaking it for someone? I still find it weird that we can set transparent_hugepage/enabled
to never and transparent_hugepage/hugepages-2048kB/enabled to madvise and still get hugepages.
(And we actually use this configuration in production for our ARM servers).

Introducing deny for global and page size I feel will over complicate it because of the issue in
the previous paragraph, page size setting overrides global setting. so even if
transparent_hugepage/enabled is deny, we might still get a THP if the page setting is not.
Someone needs to file to deny, which is the same as setting every file to never.

So I just wanted to throw another bad idea in the mix, what if we introduce another sysfs file
(I hate introducing sysfs :)), something like /sys/kernel/mm/thp_allowed (or some other alternate name)
which is default 1.
Once someone sets it to 0, no one can ever get a THP, no matter what future changes we make. Whether its
madv_collapse, bpf THPs, cgroup THPs, prctls, syscalls.. never will mean never.
Notice that its /sys/kernel/mm/thp_allowed and not /sys/kernel/mm/transparent_hugepage/thp_allowed.
Having it one directory above will make it look uglier, but it highlights that whatever you
set in /sys/kernel/mm/transparent_hugepage/ wont matter if /sys/kernel/mm/thp_allowed is set to 0.
Ideally this would be /sys/kernel/mm/transparent_hugepage/enabled=never if we were developing this
from scratch..
Not pushing for this idea, just throwing it out there.

Thanks,
Usama
 

Powered by blists - more mailing lists