Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
Content-Type: text/plain; charset=utf-8
Content-Transfer-Encoding: quoted-printable
From:   David Hildenbrand <david@redhat.com>
Mime-Version: 1.0 (1.0)
Subject: Re: [PATCH] mm: huge_memory: a new debugfs interface for splitting THP tests.
Date:   Mon, 8 Mar 2021 21:35:58 +0100
Message-Id: <44C62A78-4B37-445D-A9F7-25D1A412A802@redhat.com>
References: <CAHbLzkqjeLZ2ZK=m3+nBqcM366gtq=Ug-=ALiSmnpi+FyZfjZQ@mail.gmail.com>
Cc:     David Hildenbrand <david@redhat.com>, Zi Yan <ziy@nvidia.com>,
        Linux MM <linux-mm@kvack.org>,
        Linux Kernel Mailing List <linux-kernel@vger.kernel.org>,
        linux-kselftest@vger.kernel.org,
        "Kirill A . Shutemov" <kirill.shutemov@linux.intel.com>,
        Andrew Morton <akpm@linux-foundation.org>,
        Shuah Khan <shuah@kernel.org>,
        John Hubbard <jhubbard@nvidia.com>,
        Sandipan Das <sandipan@linux.ibm.com>,
        David Rientjes <rientjes@google.com>,
        Alex Shi <alex.shi@linux.alibaba.com>
In-Reply-To: <CAHbLzkqjeLZ2ZK=m3+nBqcM366gtq=Ug-=ALiSmnpi+FyZfjZQ@mail.gmail.com>
To:     Yang Shi <shy828301@gmail.com>
Precedence: bulk


> Am 08.03.2021 um 21:18 schrieb Yang Shi <shy828301@gmail.com>:
>=20
> =EF=BB=BFOn Mon, Mar 8, 2021 at 11:30 AM David Hildenbrand <david@redhat.c=
om> wrote:
>>=20
>>> On 08.03.21 20:11, Yang Shi wrote:
>>> On Mon, Mar 8, 2021 at 11:01 AM Zi Yan <ziy@nvidia.com> wrote:
>>>>=20
>>>> On 8 Mar 2021, at 13:11, David Hildenbrand wrote:
>>>>=20
>>>>> On 08.03.21 18:49, Zi Yan wrote:
>>>>>> On 8 Mar 2021, at 11:17, David Hildenbrand wrote:
>>>>>>=20
>>>>>>> On 08.03.21 16:22, Zi Yan wrote:
>>>>>>>> From: Zi Yan <ziy@nvidia.com>
>>>>>>>>=20
>>>>>>>> By writing "<pid>,<vaddr_start>,<vaddr_end>" to
>>>>>>>> <debugfs>/split_huge_pages_in_range_pid, THPs in the process with t=
he
>>>>>>>> given pid and virtual address range are split. It is used to test
>>>>>>>> split_huge_page function. In addition, a selftest program is added t=
o
>>>>>>>> tools/testing/selftests/vm to utilize the interface by splitting
>>>>>>>> PMD THPs and PTE-mapped THPs.
>>>>>>>=20
>>>>>>> Won't something like
>>>>>>>=20
>>>>>>> 1. MADV_HUGEPAGE
>>>>>>>=20
>>>>>>> 2. Access memory
>>>>>>>=20
>>>>>>> 3. MADV_NOHUGEPAGE
>>>>>>>=20
>>>>>>> Have a similar effect? What's the benefit of this?
>>>>>>=20
>>>>>> Thanks for checking the patch.
>>>>>>=20
>>>>>> No, MADV_NOHUGEPAGE just replaces VM_HUGEPAGE with VM_NOHUGEPAGE,
>>>>>> nothing else will be done.
>>>>>=20
>>>>> Ah, okay - maybe my memory was tricking me. There is some s390x KVM co=
de that forces MADV_NOHUGEPAGE and force-splits everything.
>>>>>=20
>>>>> I do wonder, though, if this functionality would be worth a proper use=
r interface (e.g., madvise), though. There might be actual benefit in having=
 this as a !debug interface.
>>>>>=20
>>>>> I think you aware of the discussion in https://lkml.kernel.org/r/d098c=
392-273a-36a4-1a29-59731cdf5d3d@google.com
>>>>=20
>>>> Yes. Thanks for bringing this up.
>>>>=20
>>>>>=20
>>>>> If there will be an interface to collapse a THP -- "this memory area i=
s worth extra performance now by collapsing a THP if possible" -- it might a=
lso be helpful to have the opposite functionality -- "this memory area is no=
t worth a THP, rather use that somehwere else".
>>>>>=20
>>>>> MADV_HUGE_COLLAPSE vs. MADV_HUGE_SPLIT
>>>>=20
>>>> I agree that MADV_HUGE_SPLIT would be useful as the opposite of COLLAPS=
E when user might just want PAGESIZE mappings.
>>>> Right now, HUGE_SPLIT is implicit from mapping changes like mprotect or=
 MADV_DONTNEED.
>>>=20
>>> IMHO, it sounds not very useful. MADV_DONTNEED would split PMD for any
>>> partial THP. If the range covers the whole THP, the whole THP is going
>>> to be freed anyway. All other places in kernel which need split THP
>>> have been covered. So I didn't realize any usecase from userspace for
>>> just splitting PMD to PTEs.
>>=20
>> THP are a limited resource. So indicating which virtual memory regions
>> are not performance sensitive right now (e.g., cold pages in a databse)
>> and not worth a THP might be quite valuable, no?
>=20
> Such functionality could be achieved by MADV_COLD or MADV_PAGEOUT,
> right? Then a subsequent call to MADV_NOHUGEPAGE would prevent from
> collapsing or allocating THP for that area.
>=20

I remember these deal with optimizing swapping. Not sure how they interact w=
ith THP, especially on systems without swap - I would guess they don=E2=80=98=
t as of now.

>>=20
>> --
>> Thanks,
>>=20
>> David / dhildenb
>>=20
>=20