Re: [RFC PATCH 0/3] hugetlb: add demote/split page functionality

["Zi Yan" <ziy@nvidia.com>]

[-- Attachment #1: Type: text/plain, Size: 4042 bytes --]

On 10 Mar 2021, at 12:05, Michal Hocko wrote:

> On Wed 10-03-21 11:46:57, Zi Yan wrote:
>> On 10 Mar 2021, at 11:23, Michal Hocko wrote:
>>
>>> On Mon 08-03-21 16:18:52, Mike Kravetz wrote:
>>> [...]
>>>> Converting larger to smaller hugetlb pages can be accomplished today by
>>>> first freeing the larger page to the buddy allocator and then allocating
>>>> the smaller pages.  However, there are two issues with this approach:
>>>> 1) This process can take quite some time, especially if allocation of
>>>>    the smaller pages is not immediate and requires migration/compaction.
>>>> 2) There is no guarantee that the total size of smaller pages allocated
>>>>    will match the size of the larger page which was freed.  This is
>>>>    because the area freed by the larger page could quickly be
>>>>    fragmented.
>>>
>>> I will likely not surprise to show some level of reservation. While your
>>> concerns about reconfiguration by existing interfaces are quite real is
>>> this really a problem in practice? How often do you need such a
>>> reconfiguration?
>>>
>>> Is this all really worth the additional code to something as tricky as
>>> hugetlb code base?
>>>
>>>>  include/linux/hugetlb.h |   8 ++
>>>>  mm/hugetlb.c            | 199 +++++++++++++++++++++++++++++++++++++++-
>>>>  2 files changed, 204 insertions(+), 3 deletions(-)
>>>>
>>>> -- 
>>>> 2.29.2
>>>>
>>
>> The high level goal of this patchset seems to enable flexible huge page
>> allocation from a single pool, when multiple huge page sizes are available
>> to use. The limitation of existing mechanism is that user has to specify
>> how many huge pages he/she wants and how many gigantic pages he/she wants
>> before the actual use.
>
> I believe I have understood this part. And I am not questioning that.
> This seems useful. I am mostly asking whether we need such a
> flexibility. Mostly because of the additional code and future
> maintenance complexity which has turned to be a problem for a long time.
> Each new feature tends to just add on top of the existing complexity.

I totally agree. This patchset looks to me like a partial functional
replication of splitting high order free pages to lower order ones in buddy
allocator. That is why I had the crazy idea below.

>
>> I just want to throw an idea here, please ignore if it is too crazy.
>> Could we have a variant buddy allocator for huge page allocations,
>> which only has available huge page orders in the free list? For example,
>> if user wants 2MB and 1GB pages, the allocator will only have order-9 and
>> order-19 pages; when order-9 pages run out, we can split order-19 pages;
>> if possible, adjacent order-9 pages can be merged back to order-19 pages.
>
> I assume you mean to remove those pages from the allocator when they
> are reserved rather than really used, right? I am not really sure how

No. The allocator maintains all the reserved pages for huge page allocations,
replacing existing cma_alloc or alloc_contig_pages. The kernel builds
the free list when pages are reserved either during boot time or runtime.

> you want to deal with lower orders consuming/splitting too much from
> higher orders which then makes those unusable for the use even though
> they were preallocated for a specific workload. Another worry is that a
> gap between 2MB and 1GB pages is just too big so a single 2MB request
> from 1G pool will make the whole 1GB page unusable even when the smaller
> pool needs few pages.

Yeah, the gap between 2MB and 1GB is large. The fragmentation will be
a problem. Maybe we do not need it right now, since this patchset does not
propose promoting/merging pages. Or we can reuse the existing
anti fragmentation mechanisms but with pageblock set to gigantic page size
in this pool.

I admit my idea is a much intrusive change, but I feel that more
functionality replications of core mm are added to hugetlb code, then why
not reuse the core mm code.


—
Best Regards,
Yan Zi

[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 854 bytes --]