On 25.06.19 10:25, David Hildenbrand wrote:
> On 25.06.19 09:52, Oscar Salvador wrote:
>> Hi,
>>
>> It has been while since I sent previous version [1].
>>
>> In this version I added some feedback I got back then, like letting
>> the caller decide whether he wants allocating per memory block or
>> per memory range (patch#2), and having the chance to disable vmemmap when
>> users want to expose all hotpluggable memory to userspace (patch#5).
>>
>> [Testing]
>>
>> While I could test last version on powerpc, and Huawei's fellows helped me out
>> testing it on arm64, this time I could only test it on x86_64.
>> The codebase is quite the same, so I would not expect surprises.
>>
>> - x86_64: small and large memblocks (128MB, 1G and 2G)
>> - Kernel module that adds memory spanning multiple memblocks
>> and remove that memory in a different granularity.
>>
>> So far, only acpi memory hotplug uses the new flag.
>> The other callers can be changed depending on their needs.
>>
>> Of course, more testing and feedback is appreciated.
>>
>> [Coverletter]
>>
>> This is another step to make memory hotplug more usable. The primary
>> goal of this patchset is to reduce memory overhead of the hot-added
>> memory (at least for SPARSEMEM_VMEMMAP memory model). The current way we use
>> to populate memmap (struct page array) has two main drawbacks:
>
> Mental note: How will it be handled if a caller specifies "Allocate
> memmap from hotadded memory", but we are running under SPARSEMEM where
> we can't do this.
>
>>
>> a) it consumes an additional memory until the hotadded memory itself is
>> onlined and
>> b) memmap might end up on a different numa node which is especially true
>> for movable_node configuration.
>>
>> a) it is a problem especially for memory hotplug based memory "ballooning"
>> solutions when the delay between physical memory hotplug and the
>> onlining can lead to OOM and that led to introduction of hacks like auto
>> onlining (see 31bc3858ea3e ("memory-hotplug: add automatic onlining
>> policy for the newly added memory")).
>>
>> b) can have performance drawbacks.
>>
>> Another minor case is that I have seen hot-add operations failing on archs
>> because they were running out of order-x pages.
>> E.g On powerpc, in certain configurations, we use order-8 pages,
>> and given 64KB base pagesize, that is 16MB.
>> If we run out of those, we just fail the operation and we cannot add
>> more memory.
>
> At least for SPARSEMEM, we fallback to vmalloc() to work around this
> issue. I haven't looked into the populate_section_memmap() internals
> yet. Can you point me at the code that performs this allocation?
>
>> We could fallback to base pages as x86_64 does, but we can do better.
>>
>> One way to mitigate all these issues is to simply allocate memmap array
>> (which is the largest memory footprint of the physical memory hotplug)
>> from the hot-added memory itself. SPARSEMEM_VMEMMAP memory model allows
>> us to map any pfn range so the memory doesn't need to be online to be
>> usable for the array. See patch 3 for more details.
>> This feature is only usable when CONFIG_SPARSEMEM_VMEMMAP is set.
>>
>> [Overall design]:
>>
>> 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().
>>
>> The flags are either MHP_MEMMAP_DEVICE or MHP_MEMMAP_MEMBLOCK, and only differ
>> in the way they allocate vmemmap pages within the memory blocks.
>>
>> MHP_MEMMAP_MEMBLOCK:
>> - With this flag, we will allocate vmemmap pages in each memory block.
>> This means that if we hot-add a range that spans multiple memory blocks,
>> we will use the beginning of each memory block for the vmemmap pages.
>> This strategy is good for cases where the caller wants the flexiblity
>> to hot-remove memory in a different granularity than when it was added.
>>
>> MHP_MEMMAP_DEVICE:
>> - With this flag, we will store all vmemmap pages at the beginning of
>> hot-added memory.
>>
>> So it is a tradeoff of flexiblity vs contigous memory.
>> More info on the above can be found in patch#2.
>>
>> 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.
>>
>> mark_vmemmap_pages() marks the pages as vmemmap and sets some metadata:
>>
>> The current layout of the Vmemmap pages are:
>>
>> [Head->refcount] : Nr sections used by this altmap
>> [Head->private] : Nr of vmemmap pages
>> [Tail->freelist] : Pointer to the head page
>>
>> This is done to easy the computation we need in some places.
>> E.g:
>>
>> Example 1)
>> We hot-add 1GB on x86_64 (memory block 128MB) using
>> MHP_MEMMAP_DEVICE:
>>
>> head->_refcount = 8 sections
>> head->private = 4096 vmemmap pages
>> tail's->freelist = head
>>
>> Example 2)
>> We hot-add 1GB on x86_64 using MHP_MEMMAP_MEMBLOCK:
>>
>> [at the beginning of each memblock]
>> head->_refcount = 1 section
>> head->private = 512 vmemmap pages
>> tail's->freelist = head
>>
>> We have the refcount because when using MHP_MEMMAP_DEVICE, we need to know
>> how much do we have to defer the call to vmemmap_free().
>> The thing is that the first pages of the hot-added range are used to create
>> the memmap mapping, so we cannot remove those first, otherwise we would blow up
>> when accessing the other pages.
>
> So, assuming we add_memory(1GB, MHP_MEMMAP_DEVICE) and then
> remove_memory(128MB) of the added memory, this will work?
Hmm, I guess this won't work - especially when removing the first 128MB
first, where the memmap resides.
Do we need MHP_MEMMAP_DEVICE at this point or could we start with
MHP_MEMMAP_MEMBLOCK? That "smells" like being the easier case.
--
Thanks,
David / dhildenb
