mm: reliable 1GB page allocation
The goal of this series is to keep gigantic (1GB) pages allocatable after weeks of uptime, by concentrating non-movable pageblocks in a small number of 1GiB-aligned ranges.
Design document: https://linux-mm.org/GigaBlocks
Code: https://github.com/rikvanriel/linux/tree/riel/gigablock-2026-10-03
A 1GB page needs a 1GB-aligned range in which every page is free or movable; one unmovable page disqualifies the range. Over time a small amount of kernel memory spreads across nearly every range. On a 765GB node after 53 days of build and filesystem work, none of 754 1GB ranges was free of non-movable pageblocks. On a 253GB test host the same class of workload mixed 251 of 254 ranges in 9.2 hours, and a request for eight 1GB pages returned none.
This series uses the separate free lists for each migrate type in zone->free_area[] in the fast path. The idea is to satisfy kernel allocations from the unmovable and reclaimable free lists, while movable allocations come from the movable free lists.
For each gigablock, the system tracks whether it contains non-movable content, or only free and movable page blocks. A gigablock can contain a mix of non-movable and movable data.
When the kernel free lists run low, the slow path moves pages and page blocks from inside mixed gigablocks onto the kernel free lists.
Creating kernel free memory inside mixed superblocks is done by claiming page blocks in a mixed superblock, when the kernel free lists are exhausted, and a non-movable allocation has to fall back, or by compaction evacuating movable content from previously claimed kernel pageblocks that have movable content inside.
Claiming a new pageblock in a mixed superblock is done from the allocator path, using a bounded scan, only once the kernel free list is unable to satisfy an allocation. Evacuation is done by kcompactd, and kicked off when the kernel free list falls below a watermark. The goal is for the asynchronous path keep the page allocator on the fast path.
Movable allocations can fall back to the kernel free lists. That content can always be moved later.
Region size is fixed at 1GB rather than PUD size, geometry is frozen at boot, and a hotplug span change disables tracking for the zone rather than resizing metadata.
The diffstat shows that this version adds too much code to page_alloc.c, which is already too large, and should probably be split up somehow.
One candidate in this series is to split the evacuation code and gigablock slow path into its own file, moving that out of both page_alloc.c and compaction.c.
Duplicating a small amount of compaction logic may be preferable to complicating the existing logic at this point.
This code is not ready to be merged yet. The full design document has a list of things that still need to be done.
I am posting this in the hopes of getting feedback on the basic design, so that can be incorporated in the next rounds of tuning and cleanups.
Specifically feedback on design concepts like:
- Using only zone->free_area by migrate type in the fast path
- Moving pages around in the slow path, to keep the kernel free lists stocked.
- Some amount of scanning in the page allocation code if the kernel free lists cannot satisfy an allocation, in order to claim a block in a mixed gigablock.
- Using compaction to evacuate movable content from pageblocks claimed for kernel use.
- The policy of restricting how much kernel allocations can claim new pageblocks anywhere.
- The policy of letting movable allocations fall back to the kernel free lists once the movable free list is exhausted
Documentation/admin-guide/kernel-parameters.txt | 8 include/linux/mmzone.h | 83 include/linux/pageblock-flags.h | 2 include/linux/sched.h | 4 include/linux/vm_event_item.h | 48 mm/compaction.c | 170 mm/folio.c | 5 mm/internal.h | 147 mm/memory_hotplug.c | 3 mm/mm_init.c | 150 mm/page_alloc.c | 5553 +++++++++++++++++------- mm/page_alloc.h | 6 mm/vmstat.c | 273 + 13 files changed, 5015 insertions(+), 1437 deletions(-)
Quoted 21 lines · Rik van Riel
The goal of this series is to keep gigantic (1GB) pages allocatable after weeks of uptime, by concentrating non-movable pageblocks in a small number of 1GiB-aligned ranges.
Design document: https://linux-mm.org/GigaBlocks
Code: https://github.com/rikvanriel/linux/tree/riel/gigablock-2026-10-03A 1GB page needs a 1GB-aligned range in which every page is free or movable; one unmovable page disqualifies the range. Over time a small amount of kernel memory spreads across nearly every range. On a 765GB node after 53 days of build and filesystem work, none of 754 1GB ranges was free of non-movable pageblocks. On a 253GB test host the same class of workload mixed 251 of 254 ranges in 9.2 hours, and a request for eight 1GB pages returned none.
This series uses the separate free lists for each migrate type in zone->free_area[] in the fast path. The idea is to satisfy kernel allocations from the unmovable and reclaimable free lists, while movable allocations come from the movable free lists.
Quoted 6 lines
For each gigablock, the system tracks whether it contains non-movable content, or only free and movable page blocks. A gigablock can contain a mix of non-movable and movable data.
When the kernel free lists run low, the slow path moves pages and page blocks from inside mixed gigablocks onto the kernel free lists.
Did you consider a design to introduce MIGRATE_POLLUTED: simply have another migratetype which stores polluted pageblocks. This idea can work in general, not sure how relevant it is for gigablock concept.
Quoted 65 lines
Creating kernel free memory inside mixed superblocks is done by claiming page blocks in a mixed superblock, when the kernel free lists are exhausted, and a non-movable allocation has to fall back, or by compaction evacuating movable content from previously claimed kernel pageblocks that have movable content inside.
Claiming a new pageblock in a mixed superblock is done from the allocator path, using a bounded scan, only once the kernel free list is unable to satisfy an allocation. Evacuation is done by kcompactd, and kicked off when the kernel free list falls below a watermark. The goal is for the asynchronous path keep the page allocator on the fast path.
Movable allocations can fall back to the kernel free lists. That content can always be moved later.
Region size is fixed at 1GB rather than PUD size, geometry is frozen at boot, and a hotplug span change disables tracking for the zone rather than resizing metadata.
The diffstat shows that this version adds too much code to
page_alloc.c, which is already too large, and should probably be split up somehow.One candidate in this series is to split the evacuation code and gigablock slow path into its own file, moving that out of both
page_alloc.candcompaction.c.Duplicating a small amount of compaction logic may be preferable to complicating the existing logic at this point.
This code is not ready to be merged yet. The full design document has a list of things that still need to be done.
I am posting this in the hopes of getting feedback on the basic design, so that can be incorporated in the next rounds of tuning and cleanups.
Specifically feedback on design concepts like:
- Using onlyzone->free_areaby migrate type in the fast path
- Moving pages around in the slow path, to keep the kernel free lists stocked.
- Some amount of scanning in the page allocation code if the kernel free lists cannot satisfy an allocation, in order to claim a block in a mixed gigablock.
- Using compaction to evacuate movable content from pageblocks claimed for kernel use.
- The policy of restricting how much kernel allocations can claim new pageblocks anywhere.
- The policy of letting movable allocations fall back to the kernel free lists once the movable free list is exhaustedDocumentation/admin-guide/kernel-parameters.txt | 8 include/linux/mmzone.h | 83 include/linux/pageblock-flags.h | 2 include/linux/sched.h | 4 include/linux/vm_event_item.h | 48 mm/compaction.c | 170 mm/folio.c | 5 mm/internal.h | 147 mm/memory_hotplug.c | 3 mm/mm_init.c | 150 mm/page_alloc.c | 5553 +++++++++++++++++------- mm/page_alloc.h | 6 mm/vmstat.c | 273 + 13 files changed, 5015 insertions(+), 1437 deletions(-)
Quoted 4 lines · Dev Jain
Quoted 1 line · Rik van Riel
Quoted 9 lines
Quoted 3 lines
When the kernel free lists run low, the slow path moves pages and page
blocks from inside mixed gigablocks onto the kernel free lists.Did you consider a design to introduce
MIGRATE_POLLUTED: simply have another migratetype which stores polluted pageblocks. This idea can work in general, not sure how relevant it is for gigablock concept.
Don't the first half a dozen or so changes from this prototype do that already?
What differences do you have in mind?
Signature · 1 line
All Rights Reversed.
Quoted 14 lines · Rik van Riel
Quoted 4 lines · Dev Jain
Quoted 1 line · Rik van Riel
Quoted 9 lines
Quoted 3 lines
When the kernel free lists run low, the slow path moves pages and page
blocks from inside mixed gigablocks onto the kernel free lists.Did you consider a design to introduce
MIGRATE_POLLUTED: simply have another migratetype which stores polluted pageblocks. This idea can work in general, not sure how relevant it is for gigablock concept.
Quoted 4 lines
Don't the first half a dozen or so changes from this prototype do that already?
What differences do you have in mind?
Sorry, it was just a drive-by comment. I will eventually try to review your design document.