lkml.org 
[lkml]   [2020]   [Sep]   [23]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH RFC 0/4] mm: place pages to the freelist tail when onling and undoing isolation
On Wed, Sep 23, 2020 at 04:31:25PM +0200, Vlastimil Babka wrote:
>On 9/16/20 9:31 PM, David Hildenbrand wrote:
>>
>>
>>> Am 16.09.2020 um 20:50 schrieb osalvador@suse.de:
>>>
>>> On 2020-09-16 20:34, David Hildenbrand wrote:
>>>> When adding separate memory blocks via add_memory*() and onlining them
>>>> immediately, the metadata (especially the memmap) of the next block will be
>>>> placed onto one of the just added+onlined block. This creates a chain
>>>> of unmovable allocations: If the last memory block cannot get
>>>> offlined+removed() so will all dependant ones. We directly have unmovable
>>>> allocations all over the place.
>>>> This can be observed quite easily using virtio-mem, however, it can also
>>>> be observed when using DIMMs. The freshly onlined pages will usually be
>>>> placed to the head of the freelists, meaning they will be allocated next,
>>>> turning the just-added memory usually immediately un-removable. The
>>>> fresh pages are cold, prefering to allocate others (that might be hot)
>>>> also feels to be the natural thing to do.
>>>> It also applies to the hyper-v balloon xen-balloon, and ppc64 dlpar: when
>>>> adding separate, successive memory blocks, each memory block will have
>>>> unmovable allocations on them - for example gigantic pages will fail to
>>>> allocate.
>>>> While the ZONE_NORMAL doesn't provide any guarantees that memory can get
>>>> offlined+removed again (any kind of fragmentation with unmovable
>>>> allocations is possible), there are many scenarios (hotplugging a lot of
>>>> memory, running workload, hotunplug some memory/as much as possible) where
>>>> we can offline+remove quite a lot with this patchset.
>>>
>>> Hi David,
>>>
>>
>> Hi Oscar.
>>
>>> I did not read through the patchset yet, so sorry if the question is nonsense, but is this not trying to fix the same issue the vmemmap patches did? [1]
>>
>> Not nonesense at all. It only helps to some degree, though. It solves the dependencies due to the memmap. However, it‘s not completely ideal, especially for single memory blocks.
>>
>> With single memory blocks (virtio-mem, xen-balloon, hv balloon, ppc dlpar) you still have unmovable (vmemmap chunks) all over the physical address space. Consider the gigantic page example after hotplug. You directly fragmented all hotplugged memory.
>>
>> Of course, there might be (less extreme) dependencies due page tables for the identity mapping, extended struct pages and similar.
>>
>> Having that said, there are other benefits when preferring other memory over just hotplugged memory. Think about adding+onlining memory during boot (dimms under QEMU, virtio-mem), once the system is up you will have most (all) of that memory completely untouched.
>>
>> So while vmemmap on hotplugged memory would tackle some part of the issue, there are cases where this approach is better, and there are even benefits when combining both.
>
>I see the point, but I don't think the head/tail mechanism is great for this. It
>might sort of work, but with other interfering activity there are no guarantees
>and it relies on a subtle implementation detail. There are better mechanisms
>possible I think, such as preparing a larger MIGRATE_UNMOVABLE area in the
>existing memory before we allocate those long-term management structures. Or
>onlining a bunch of blocks as zone_movable first and only later convert to
>zone_normal in a controlled way when existing normal zone becomes depeted?
>

To be honest, David's approach is easy to understand for me.

And I don't see some negative effect.

>I guess it's an issue that the e.g. 128M block onlines are so disconnected from
>each other it's hard to employ a strategy that works best for e.g. a whole bunch
>of GB onlined at once. But I noticed some effort towards new API, so maybe that
>will be solved there too?
>
>> Thanks!
>>
>> David
>>
>>>
>>> I was about to give it a new respin now that thw hwpoison stuff has been settled.
>>>
>>> [1] https://patchwork.kernel.org/cover/11059175/
>>>
>>

--
Wei Yang
Help you, Help me

\
 
 \ /
  Last update: 2020-09-24 03:58    [W:0.096 / U:4.732 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site