Re: [RFC v2 2/2] mm/memory_hotplug: Shrink spanned pages when offlining memory

From: David Hildenbrand
Date: Wed Aug 22 2018 - 04:11:49 EST


On 22.08.2018 09:50, Oscar Salvador wrote:
> On Tue, Aug 21, 2018 at 03:17:10PM +0200, David Hildenbrand wrote:
>>> add_device_memory is in charge of
>>
>> I wouldn't use the terminology of onlining/offlining here. That applies
>> rather to memory that is exposed to the rest of the system (e.g. buddy
>> allocator, has underlying memory block devices). I guess it is rather a
>> pure setup/teardown of that device memory.
>
> Hi David,
>
> I am not sure if you are referring to:
>
> "
> a) calling either arch_add_memory() or add_pages(), depending on whether
> we want a linear mapping
> b) online the memory sections that correspond to the pfn range
> c) calling move_pfn_range_to_zone() being zone ZONE_DEVICE to
> expand zone/pgdat spanned pages and initialize its pages
> "
>
> Well, that is partialy true.
> I mean, in order to make this work, we need to offline/online the memory
> sections, because shrink_pages will rely on that from now on.
> Is what we do when online/offline pages, but since device memory
> does not go through the "official" channels, we need to do it there
> as well.
>
> Sure I can use another terminology, but since that is what
> offline/online_mem_sections do, I just came up with that.
>

Okay, got it, so it is basically "mark the sections as online/offline".

>> I would really like to see the mem_hotplug_begin/end also getting moved
>> inside add_device_memory()/del_device_memory(). (just like for
>> add/remove_memory)
>>
>> I wonder if kasan_ stuff actually requires this lock, or if it could
>> also be somehow moved inside add_device_memory/del_device_memory.
>
> Yes, that was my first approach, but then I saw that the kasan stuff is being
> handled whithin those locks, so I was not sure and I backed off leaving the
> mem_hotplug_begin/end where they were.
>
> Maybe Jerome can shed some light and, and we can just handle the kasan stuff
> out of the locks.
>
>> Maybe shorten that a bit
>>
>> "HMM/devm memory does not have IORESOURCE_SYSTEM_RAM set. They use
>> devm_request_mem_region/devm_release_mem_region to add/release a
>> resource. Just back off here."
>
> Uhm, fair enough.
>
>> Any reason for these indirections?
>
> I wanted to hide the internals in the memory_hotplug code.
> I thought about removing them, but I finally left them.
> If people think that we are better off without them, I can just
> remove them.

I don't see a need for that. (everyone following the functions has to go
via one indirection that just passes on parameters). It is also not done
for other functions (a.g. add_memory)

>
>> I guess for readability, this patch could be split up into several
>> patches. E.g. factoring out of add_device_memory/del_device_memory,
>> release_mem_region_adjustable change ...
>
> Yes, really true.
> But I wanted first to gather feedback mainly from HMM/devm people to see
> if they saw an outright bug within the series because I am not so
> familiar with that part of the code.
>
> Feedback from Jerome/Dan will be appreciate as well to see if this is a good
> direction.

Yes, they probably know best how this all fits together.

>
> But you are right, in the end, this will have to be slipt up into several
> parts to ease the review.
>
> Thanks for reviewing this David!
> I will try to address your concerns.
>
> Thanks
>


--

Thanks,

David / dhildenb