Re: [PATCH v1] drivers/base/node.c: Simplify unregister_memory_block_under_nodes()

[David Hildenbrand]

On 19.07.19 11:09, Michal Hocko wrote:
> On Fri 19-07-19 10:48:19, David Hildenbrand wrote:
>> On 19.07.19 10:42, Michal Hocko wrote:
>>> On Thu 18-07-19 16:22:39, David Hildenbrand wrote:
>>>> We don't allow to offline memory block devices that belong to multiple
>>>> numa nodes. Therefore, such devices can never get removed. It is
>>>> sufficient to process a single node when removing the memory block.
>>>>
>>>> Remember for each memory block if it belongs to no, a single, or mixed
>>>> nodes, so we can use that information to skip unregistering or print a
>>>> warning (essentially a safety net to catch BUGs).
>>>
>>> I do not really like NUMA_NO_NODE - 1 thing. This is yet another invalid
>>> node that is magic. Why should we even care? In other words why is this
>>> patch an improvement?
>>
>> I mean we can of course go ahead and drop the "NUMA_NO_NODE - 1" thingy
>> from the patch. A memory block with multiple nodes would (as of now)
>> only indicate one of the nodes.
> 
> Yes and that seemed to work reasonably well so far. Sure there is a
> potential confusion but platforms with interleaved nodes are rare enough
> to somebody to even notice so far.

Let's hope there are no BUGs related to that and we just didn't catch
them yet because it's barely used :)

> 
>> Then there is simply no way to WARN_ON_ONCE() in case unexpected things
>> would happen. (I mean it really shouldn't happen or we have a BUG
>> somewhere else)
> 
> I do not really see much point to warn here. What can user potentially
> do?

We could detect this while testing and see that some other code seems to
do unexpected things (remove such memory blocks although not allowed).

> 
>> Alternative: Add "bool mixed_nids;" to "struct memory block".
> 
> That would be certainly possible but do we actually care?

Only if we want to warn. And I am fine with dropping this part.

-- 

Thanks,

David / dhildenb