linux-kernel - Re: [PATCH] mm/memory.c: Add return NUMA_NO_NODE in numa_migrate_check() when folio_nid() and numa_node

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [day] [month] [year] [list]

Message-ID: <45806c9f-b9cc-4917-ae93-98cabf08bf2e@redhat.com>
Date: Thu, 9 Jan 2025 15:40:01 +0100
From: David Hildenbrand <david@...hat.com>
To: Donet Tom <donettom@...ux.ibm.com>,
 Andrew Morton <akpm@...ux-foundation.org>, linux-mm@...ck.org,
 linux-kernel@...r.kernel.org
Cc: Ritesh Harjani <ritesh.list@...il.com>,
 Baolin Wang <baolin.wang@...ux.alibaba.com>,
 "Aneesh Kumar K . V" <aneesh.kumar@...nel.org>,
 Matthew Wilcox <willy@...radead.org>, Zi Yan <ziy@...dia.com>,
 Muchun Song <muchun.song@...ux.dev>
Subject: Re: [PATCH] mm/memory.c: Add return NUMA_NO_NODE in
 numa_migrate_check() when folio_nid() and numa_node_id() are the same.

On 09.01.25 15:29, Donet Tom wrote:
> 
> On 1/9/25 18:43, David Hildenbrand wrote:
>> On 09.01.25 07:46, Donet Tom wrote:
>>> If the folio_nid() and numa_node_id() are the same, it indicates
>>> that the folio is already on the same node as the process. In
>>> this case, there's no need to migrate the pages.
>>>
>>> This patch adds return NUMA_NO_NODE in numa_migrate_check() when
>>> the folio_nid() and numa_node_id() match, preventing the function
>>> from executing the remaining code unnecessarily.
>>>
>>> Signed-off-by: Donet Tom <donettom@...ux.ibm.com>
>>> ---
>>>    mm/memory.c | 1 +
>>>    1 file changed, 1 insertion(+)
>>>
>>> diff --git a/mm/memory.c b/mm/memory.c
>>> index 398c031be9ba..dfd89ff7f639 100644
>>> --- a/mm/memory.c
>>> +++ b/mm/memory.c
>>> @@ -5509,6 +5509,7 @@ int numa_migrate_check(struct folio *folio,
>>> struct vm_fault *vmf,
>>>        if (folio_nid(folio) == numa_node_id()) {
>>>            count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
>>>            *flags |= TNF_FAULT_LOCAL;
>>> +        return NUMA_NO_NODE;
>>
>> Doesn't this just mean that it is a local fault, but not necessarily
>> that we don't want to migrate that folio?
>>
>> mpol_misplaced states: "check whether current folio node is valid in
>> policy"
>>
>> Could we have a different policy set that does not indicate the local
>> node as the target node?
>>
>> Note how mpol_misplaced() obtains the target node to the do
>>
>>
>> int curnid = folio_nid(folio);
>> ...
>> int polnid = NUMA_NO_NODE;
>> int ret = NUMA_NO_NODE
>>
>> ... detect polnid
>>
>> if (curnid != polnid)
>>      ret = polnid;
>> ...
>> return ret;
>>
>>
>> So mpol_misplaced() will return "NUMA_NO_NODE" if already on the
>> correct target node.
> 
> Thank you, David. I understood my patch is wrong.
> 
> I have a small question: Page access latency is lower when the folio is
> on the same NUMA
> node as the process. However, if the policy node is set to a different
> NUMA node and the
> MPOL_F_MORON flag is not set, we migrate the page to the policy node,
> thereby increasing
> access latency.  Could this have an impact on performance? What benefits
> do we gain from this?

I have to admit, that I have not much idea about the details here 
(especially, how MORON comes into play here), so what I write below 
might be wrong :)

Consider the following example: making use of memory on CPU-less NUMA 
nodes (e.g., CXL, NVDIMM, whatsoever). In that case, access latency 
might be higher, but maybe that's exactly what someone wants to achieve 
here (e.g., placing cold data).

Maybe page demotion can come into play here as well. But the 
MPOL_F_MORON is slightly confusing me.

-- 
Cheers,

David / dhildenb