Re: [PATCH v3 0/5] Allocate memmap from hotadded memory

[David Hildenbrand <david@redhat.com>]

On 25.07.19 18:02, Oscar Salvador wrote:
> Here we go with v3.
> 
> v3 -> v2:
>         * Rewrite about vmemmap pages handling.
>           Prior to this version, I was (ab)using hugepages fields
>           from struct page, while here I am officially adding a new
>           sub-page type with the fields I need.
> 
>         * Drop MHP_MEMMAP_{MEMBLOCK,DEVICE} in favor of MHP_MEMMAP_ON_MEMORY.
>           While I am still not 100% if this the right decision, and while I
>           still see some gaining in having MHP_MEMMAP_{MEMBLOCK,DEVICE},
>           having only one flag ease the code.
>           If the user wants to allocate memmaps per memblock, it'll
>           have to call add_memory() variants with memory-block granularity.
> 
>           If we happen to have a more clear usecase MHP_MEMMAP_MEMBLOCK
>           flag in the future, so user does not have to bother about the way
>           it calls add_memory() variants, but only pass a flag, we can add it.
>           Actually, I already had the code, so add it in the future is going to be
>           easy.
> 
>         * Granularity check when hot-removing memory.
>           Just checking that the granularity is the same.
> 
> [Testing]
> 
>  - x86_64: small and large memblocks (128MB, 1G and 2G)
> 
> So far, only acpi memory hotplug uses the new flag.
> The other callers can be changed depending on their needs.
> 
> [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:
> 
> 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.
> 
> 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_ON_MEMORY 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().
> 
> MHP_MEMMAP_ON_MEMORY flag parameter will specify to allocate memmaps
> from the hot-added range.
> If callers wants memmaps to be allocated per memory block, it will
> have to call add_memory() variants in memory-block granularity
> spanning the whole range, while if it wants to allocate memmaps
> per whole memory range, just one call will do.
> 
> Want to add 384MB (3 sections, 3 memory-blocks)
> e.g:
> 
> add_memory(0x1000, size_memory_block);
> add_memory(0x2000, size_memory_block);
> add_memory(0x3000, size_memory_block);
> 

Some more thoughts:

1. It can happen that pfn_online() for a vmemmap page returns either
true or false, depending on the state of the section. It could be that
the memory block holding the vmemmap is offline while another memory
block making use of it is online.

I guess this isn't bad (I assume it is similar for the altmap), however
it could be that makedumpfile will exclude the vmemmap from dumps (as it
will usually only dump pages in sections marked online if I am not wrong
- maybe it special cases vmemmaps already). Also, could be that it is
not saved/restored during hibernation. We'll have to verify.


2. memmap access when adding/removing memory

The memmap is initialized when onlining memory. We still have to clean
up accessing the memmap in remove_memory(). You seem to introduce new
users - which is bad. Especially when removing memory we never onlined.

When removing memory, you shouldn't have to worry about any orders -
nobody should touch the memmap. I am aware that we still query the zone
- are there other users that touch the memmap when removing memory?


3. isolation/compaction

I am not sure if simply unconditionally skipping over Vmemmap pages is a
good idea. I would have guessed it is better to hinder callers from even
triggering this.

E.g., Only online the pieces that don't contain the vmemmap. When
offlining a memory block, only actually try to offline the pieces that
were onlined - excluding the vmemmap.

Might require some smaller reworks but shouldn't be too hard as far as I
can tell.


4. mhp_flags and altmap with __add_pages()

I have hoped that we can  handle the specific of MMAP_ON_MEMORY
completely in add_memory() - nobody else needs MMAP_ON_MEMORY (we have
the generic altmap concept already).

So, setup the struct vmem_altmap; in add_memory() and pass it directly.
During arch_add_memory(), nobody should be touching the vmemmap either
way, as it is completely uninitialized.

When we return from arch_add_memory() in add_memory(), we could then
initialize the memmap for the vmemmap pages (e.g., set them to
PageVmemmap) - via mhp_mark_vmemmap_pages() or such.

What exactly speaks against this approach? (moving the MMAP_ON_MEMORY
handling completely out of __add_pages())? Am I missing some access the
could be evil while the pages are not mapped?

(I'd love to see __add_pages() only eat an altmap again, and keep the
MMAP_ON_MEMORY thingy specific to add_memory())

-- 

Thanks,

David / dhildenb