Re: [RFC PATCH v2 0/7] Use pageblock_order for cma and alloc_contig_range alignment.

On 10 Dec 2021, at 13:36, David Hildenbrand wrote:

On 10.12.21 00:04, Zi Yan wrote:

quoted

From: Zi Yan <ziy@nvidia.com>

Hi all,

Hi,

thanks for working on that!

quoted

This patchset tries to remove the MAX_ORDER - 1 alignment requirement for CMA
and alloc_contig_range(). It prepares for my upcoming changes to make MAX_ORDER
adjustable at boot time[1].

The MAX_ORDER - 1 alignment requirement comes from that alloc_contig_range()
isolates pageblocks to remove free memory from buddy allocator but isolating
only a subset of pageblocks within a page spanning across multiple pageblocks
causes free page accounting issues. Isolated page might not be put into the
right free list, since the code assumes the migratetype of the first pageblock
as the whole free page migratetype. This is based on the discussion at [2].

To remove the requirement, this patchset:
1. still isolates pageblocks at MAX_ORDER - 1 granularity;
2. but saves the pageblock migratetypes outside the specified range of
   alloc_contig_range() and restores them after all pages within the range
   become free after __alloc_contig_migrate_range();
3. splits free pages spanning multiple pageblocks at the beginning and the end
   of the range and puts the split pages to the right migratetype free lists
   based on the pageblock migratetypes;
4. returns pages not in the range as it did before this patch.

Isolation needs to happen at MAX_ORDER - 1 granularity, because otherwise
1) extra code is needed to detect pages (free, PageHuge, THP, or PageCompound)
to make sure all pageblocks belonging to a single page are isolated together
and later pageblocks outside the range need to have their migratetypes restored;
or 2) extra logic will need to be added during page free time to split a free
page with multi-migratetype pageblocks.

Two optimizations might come later:
1. only check unmovable pages within the range instead of MAX_ORDER - 1 aligned
   range during isolation to increase successful rate of alloc_contig_range().

The issue with virtio-mem is that we'll need that as soon as we change
the granularity to pageblocks, because otherwise, you can heavily
degrade unplug reliably in sane setups:

Previous:
* Try unplug free 4M range (2 pageblocks): succeeds

Now:
* Try unplug 2M range (first pageblock): succeeds.
* Try unplug next 2M range (second pageblock): fails because first
contains unmovable allcoations.

OK. Make sense. I will add it in the next version.


--
Best Regards,
Yan, Zi