Message-ID: <2610143f-3274-47c0-9a11-777be673c186@linux.alibaba.com>
Date: Thu, 29 May 2025 16:27:59 +0800
From: Baolin Wang <baolin.wang@...ux.alibaba.com>
To: Nico Pache <npache@...hat.com>
Cc: David Hildenbrand <david@...hat.com>, David Rientjes
 <rientjes@...gle.com>, zokeefe@...gle.com, linux-mm@...ck.org,
 linux-doc@...r.kernel.org, linux-kernel@...r.kernel.org,
 linux-trace-kernel@...r.kernel.org, ziy@...dia.com,
 lorenzo.stoakes@...cle.com, Liam.Howlett@...cle.com, ryan.roberts@....com,
 dev.jain@....com, corbet@....net, rostedt@...dmis.org, mhiramat@...nel.org,
 mathieu.desnoyers@...icios.com, akpm@...ux-foundation.org,
 baohua@...nel.org, willy@...radead.org, peterx@...hat.com,
 wangkefeng.wang@...wei.com, usamaarif642@...il.com, sunnanyong@...wei.com,
 vishal.moola@...il.com, thomas.hellstrom@...ux.intel.com,
 yang@...amperecomputing.com, kirill.shutemov@...ux.intel.com,
 aarcange@...hat.com, raquini@...hat.com, anshuman.khandual@....com,
 catalin.marinas@....com, tiwai@...e.de, will@...nel.org,
 dave.hansen@...ux.intel.com, jack@...e.cz, cl@...two.org,
 jglisse@...gle.com, surenb@...gle.com, hannes@...xchg.org, mhocko@...e.com,
 rdunlap@...radead.org
Subject: Re: [PATCH v7 06/12] khugepaged: introduce khugepaged_scan_bitmap for
 mTHP support



On 2025/5/29 12:02, Nico Pache wrote:
> On Wed, May 28, 2025 at 8:04 AM Baolin Wang
> <baolin.wang@...ux.alibaba.com> wrote:
>>
>>
>>
>> On 2025/5/28 17:26, David Hildenbrand wrote:
>>> On 22.05.25 11:39, Baolin Wang wrote:
>>>>
>>>>
>>>> On 2025/5/21 18:23, Nico Pache wrote:
>>>>> On Tue, May 20, 2025 at 4:09 AM Baolin Wang
>>>>> <baolin.wang@...ux.alibaba.com> wrote:
>>>>>>
>>>>>> Sorry for late reply.
>>>>>>
>>>>>> On 2025/5/17 14:47, Nico Pache wrote:
>>>>>>> On Thu, May 15, 2025 at 9:20 PM Baolin Wang
>>>>>>> <baolin.wang@...ux.alibaba.com> wrote:
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>> On 2025/5/15 11:22, Nico Pache wrote:
>>>>>>>>> khugepaged scans anons PMD ranges for potential collapse to a
>>>>>>>>> hugepage.
>>>>>>>>> To add mTHP support we use this scan to instead record chunks of
>>>>>>>>> utilized
>>>>>>>>> sections of the PMD.
>>>>>>>>>
>>>>>>>>> khugepaged_scan_bitmap uses a stack struct to recursively scan a
>>>>>>>>> bitmap
>>>>>>>>> that represents chunks of utilized regions. We can then determine
>>>>>>>>> what
>>>>>>>>> mTHP size fits best and in the following patch, we set this
>>>>>>>>> bitmap while
>>>>>>>>> scanning the anon PMD. A minimum collapse order of 2 is used as
>>>>>>>>> this is
>>>>>>>>> the lowest order supported by anon memory.
>>>>>>>>>
>>>>>>>>> max_ptes_none is used as a scale to determine how "full" an order
>>>>>>>>> must
>>>>>>>>> be before being considered for collapse.
>>>>>>>>>
>>>>>>>>> When attempting to collapse an order that has its order set to
>>>>>>>>> "always"
>>>>>>>>> lets always collapse to that order in a greedy manner without
>>>>>>>>> considering the number of bits set.
>>>>>>>>>
>>>>>>>>> Signed-off-by: Nico Pache <npache@...hat.com>
>>>>>>>>
>>>>>>>> Sigh. You still haven't addressed or explained the issues I
>>>>>>>> previously
>>>>>>>> raised [1], so I don't know how to review this patch again...
>>>>>>> Can you still reproduce this issue?
>>>>>>
>>>>>> Yes, I can still reproduce this issue with today's (5/20) mm-new
>>>>>> branch.
>>>>>>
>>>>>> I've disabled PMD-sized THP in my system:
>>>>>> [root]# cat /sys/kernel/mm/transparent_hugepage/enabled
>>>>>> always madvise [never]
>>>>>> [root]# cat
>>>>>> /sys/kernel/mm/transparent_hugepage/hugepages-2048kB/enabled
>>>>>> always inherit madvise [never]
>>>>>>
>>>>>> And I tried calling madvise() with MADV_COLLAPSE for anonymous memory,
>>>>>> and I can still see it collapsing to a PMD-sized THP.
>>>>> Hi Baolin ! Thank you for your reply and willingness to test again :)
>>>>>
>>>>> I didn't realize we were talking about madvise collapse-- this makes
>>>>> sense now. I also figured out why I could "reproduce" it before. My
>>>>> script was always enabling the THP settings in two places, and I only
>>>>> commented out one to test this. But this time I was doing more manual
>>>>> testing.
>>>>>
>>>>> The original design of madvise_collapse ignores the sysfs and
>>>>> collapses even if you have an order disabled. I believe this behavior
>>>>> is wrong, but by design. I spent some time playing around with madvise
>>>>> collapses with and w/o my changes. This is not a new thing, I
>>>>> reproduced the issue in 6.11 (Fedora 41), and I think its been
>>>>> possible since the inception of madvise collapse 3 years ago. I
>>>>> noticed a similar behavior on one of my RFC since it was "breaking"
>>>>> selftests, and the fix was to reincorporate this broken sysfs
>>>>> behavior.
>>>>
>>>> OK. Thanks for the explanation.
>>>>
>>>>> 7d8faaf15545 ("mm/madvise: introduce MADV_COLLAPSE sync hugepage
>>>>> collapse")
>>>>> "This call is independent of the system-wide THP sysfs settings, but
>>>>> will fail for memory marked VM_NOHUGEPAGE."
>>>>>
>>>>> The second condition holds true (and fails for VM_NOHUGEPAGE), but I
>>>>> dont know if we actually want madvise_collapse to be independent of
>>>>> the system-wide.
>>>>
>>>> This design principle surprised me a bit, and I failed to find the
>>>> reason in the commit log. I agree that "never should mean never," and we
>>>> should respect the THP/mTHP sysfs setting. Additionally, for the
>>>> 'shmem_enabled' sysfs interface controlled for shmem/tmpfs, THP collapse
>>>> can still be prohibited through the 'deny' configuration. The rules here
>>>> are somewhat confusing.
>>>
>>> I recall that we decided to overwrite "VM_NOHUGEPAGE", because the
>>> assumption is that the same app that triggered MADV_NOHUGEPAGE triggers
>>> the collapse. So the app decides on its own behavior.
>>>
>>> Similarly, allowing for collapsing in a VM without VM_HUGEPAGE in the
>>> "madvise" mode would be fine.
>>>
>>> But in the "never" case, we should just "never" collapse.
>>
>> OK. Let's fix the "never" case first. Thanks.
> Great, I will update that in the next version!

I've sent a patchset to fix the MADV_COLLAPSE issue for anonymous memory 
and shmem [1]. Please have a look.

[1] 
https://lore.kernel.org/all/cover.1748506520.git.baolin.wang@linux.alibaba.com/

Powered by blists - more mailing lists