Re: [PATCH v2 4/5] mm,memory_hotplug: allocate memmap from the added memory range for sparse-vmemmap

[David Hildenbrand]

On 26.06.19 10:13, Oscar Salvador wrote:
> On Tue, Jun 25, 2019 at 10:49:10AM +0200, David Hildenbrand wrote:
>> On 25.06.19 09:52, Oscar Salvador wrote:
>>> Physical memory hotadd has to allocate a memmap (struct page array) for
>>> the newly added memory section. Currently, alloc_pages_node() is used
>>> for those allocations.
>>>
>>> This has some disadvantages:
>>>  a) an existing memory is consumed for that purpose
>>>     (~2MB per 128MB memory section on x86_64)
>>>  b) if the whole node is movable then we have off-node struct pages
>>>     which has performance drawbacks.
>>>
>>> a) has turned out to be a problem for memory hotplug based ballooning
>>>    because the userspace might not react in time to online memory while
>>>    the memory consumed during physical hotadd consumes enough memory to
>>>    push system to OOM. 31bc3858ea3e ("memory-hotplug: add automatic onlining
>>>    policy for the newly added memory") has been added to workaround that
>>>    problem.
>>>
>>> I have also seen hot-add operations failing on powerpc due to the fact
>>> that we try to use order-8 pages. If the base page size is 64KB, this
>>> gives us 16MB, and if we run out of those, we simply fail.
>>> One could arge that we can fall back to basepages as we do in x86_64, but
>>> we can do better when CONFIG_SPARSEMEM_VMEMMAP is enabled.
>>>
>>> Vmemap page tables can map arbitrary memory.
>>> That means that we can simply use the beginning of each memory section and
>>> map struct pages there.
>>> struct pages which back the allocated space then just need to be treated
>>> carefully.
>>>
>>> Implementation wise we reuse vmem_altmap infrastructure to override
>>> the default allocator used by __vmemap_populate. Once the memmap is
>>> allocated we need a way to mark altmap pfns used for the allocation.
>>> If MHP_MEMMAP_{DEVICE,MEMBLOCK} flag was passed, we set up the layout of the
>>> altmap structure at the beginning of __add_pages(), and then we call
>>> mark_vmemmap_pages().
>>>
>>> Depending on which flag is passed (MHP_MEMMAP_DEVICE or MHP_MEMMAP_MEMBLOCK),
>>> mark_vmemmap_pages() gets called at a different stage.
>>> With MHP_MEMMAP_MEMBLOCK, we call it once we have populated the sections
>>> fitting in a single memblock, while with MHP_MEMMAP_DEVICE we wait until all
>>> sections have been populated.
>>
>> So, only MHP_MEMMAP_DEVICE will be used. Would it make sense to only
>> implement one for now (after we decide which one to use), to make things
>> simpler?
>>
>> Or do you have a real user in mind for the other?
> 
> Currently, only MHP_MEMMAP_DEVICE will be used, as we only pass flags from
> acpi memory-hotplug path.
> 
> All the others: hyper-v, Xen,... will have to be evaluated to see which one
> do they want to use.
> 
> Although MHP_MEMMAP_DEVICE is the only one used right now, I introduced
> MHP_MEMMAP_MEMBLOCK to give the callers the choice of using MHP_MEMMAP_MEMBLOCK
> if they think that a strategy where hot-removing works in a different granularity
> makes sense.
> 
> Moreover, since they both use the same API, there is no extra code needed to
> handle it. (Just two lines in __add_pages())
> 
> This arose here [1].
> 
> [1] https://patchwork.kernel.org/project/linux-mm/list/?submitter=137061
> 

Just noting that you can emulate MHP_MEMMAP_MEMBLOCK via
MHP_MEMMAP_DEVICE by adding memory in memory block granularity (which is
what hyper-v and xen do if I am not wrong!).

Not yet convinced that both, MHP_MEMMAP_MEMBLOCK and MHP_MEMMAP_DEVICE
are needed. But we can sort that out later.

-- 

Thanks,

David / dhildenb