Re: [RFC PATCH 1/2] mm/demotion: Expose memory type details via sysfs

From: Aneesh Kumar K V
Date: Fri Aug 26 2022 - 04:06:34 EST


On 8/26/22 1:30 PM, Wei Xu wrote:
> On Thu, Aug 25, 2022 at 8:00 PM Aneesh Kumar K V
> <aneesh.kumar@xxxxxxxxxxxxx> wrote:
>>
>> On 8/26/22 7:20 AM, Huang, Ying wrote:
>>> "Aneesh Kumar K.V" <aneesh.kumar@xxxxxxxxxxxxx> writes:
>>>
>>>> This patch adds /sys/devices/virtual/memtier/ where all memory tier related
>>>> details can be found. All allocated memory types will be listed there as
>>>> /sys/devices/virtual/memtier/memtypeN/
>>>
>>> Another choice is to make memory types and memory tiers system devices.
>>> That is,
>>>
>>> /sys/devices/system/memory_type/memory_typeN
>>> /sys/devices/system/memory_tier/memory_tierN
>>>
>>
>> subsys_system_register() documentation says
>>
>> * Do not use this interface for anything new, it exists for compatibility
>> * with bad ideas only. New subsystems should use plain subsystems; and
>> * add the subsystem-wide attributes should be added to the subsystem
>> * directory itself and not some create fake root-device placed in
>> * /sys/devices/system/<name>.
>>
>> memtier being a virtual device, I was under the impression that /sys/devices/virtual
>> is the recommended place.
>>
>>> That looks more natural to me. Because we already have "node" and
>>> "memory" devices there. Why don't you put memory types and memory tiers
>>> there?
>>>
>>> And, I think we shouldn't put "memory_type" in the "memory_tier"
>>> directory. "memory_type" isn't a part of "memory_tier".
>>>
>>
>> I was looking consolidating both memory tier and memory type into the same sysfs subsystem.
>> Your recommendation imply we create two subsystem memory_tier and memtype. I was
>> trying to avoid that. May be a generic term like "memory_tiering" can help to
>> consolidate all tiering related details there?
>>
>
> A generic term "memory_tiering" sounds good to me.
>
> Given that this will be a user-facing, stable kernel API, I think we'd
> better to only add what is most useful for userspace and don't have to
> mirror the kernel internal data structures in this interface.
>
> My understanding is that we haven't fully settled down on how to
> customize memory tiers from userspace. So we don't have to show
> memory_type yet, which is a kernel data structure at this point.
>
> The userspace does need to know what are the memory tiers and which
> NUMA nodes are included in each memory tier. How about we provide the
> "nodelist" interface for each memory tier as in the original proposal?
>
> The userspace would also like to know which memory tiers/nodes belong
> to the top tiers (the promotion targets). We can provide a "toptiers"
> or "toptiers_nodelist" interface to report that.
>

How about also including abstract distance range of a memory tier?
That will be useful to derive the hierarchy.

> Both should still be useful even if we decide to add memory_type for
> memory tier customization.
>

-aneesh