Re: [PATCH] mm/page_alloc: skip high atomic reservation at or below costly order

["Vlastimil Babka (SUSE)" <vbabka@kernel.org>]

On 5/19/26 22:28, Johannes Weiner wrote:
> On Mon, May 18, 2026 at 06:25:32PM -0700, JP Kobryn (Meta) wrote:
>> We're seeing a pattern in production where 2MB THP order-9 allocations are
>> failing due to fragmentation and triggering reclaim on systems with plenty
>> of free memory. Over time, the success rate of these THP allocations do not
>> increase at all.
>> 
>> Inspecting zone->vm_stat[NR_FREE_PAGES] via kprobe on compaction_suitable()
>> indicated the given zone had sufficient free pages for order-9 allocations,
>> yet they were going unused. Drilling down into the zone and inspecting
>> /proc/pagetypeinfo revealed why. Order-9 blocks were accumulating in the
>> zone's HighAtomic bucket (while zero were present in Movable). THP is
>> unable to draw blocks from HighAtomic since that bucket is not in the
>> fallback list.
>> 
>> The heuristic for reserving pageblocks in HighAtomic is that any atomic
>> allocation greater than order-0 will result in the full pageblock being
>> captured. This means that an order-1 atomic allocation will over-reserve by
>> 256x, a full 512 pageblock.
>> 
>> Gate the reservation on order. Skip for allocations at or below
>> PAGE_ALLOC_COSTLY_ORDER. This prevents smaller atomic allocations from
>> reserving entire pageblocks, and significantly helps when THP is in use on
>> a fragmented but otherwise healthy system.
>> 
>> Testing was performed using an A/B instagram workload receiving prod
>> traffic. Each side had ~60 hosts with 64G memory. The patch resulted in
>> several gains:
>> 
>> Unpatched
>> HighAtomic pageblocks per host: 309-312 (1% of zone or 620MB),
>>   ...all order-9 blocks in HighAtomic
>> THP success rate: 1-6%
>> Compaction success rate: 0-2%
>> pgscan_kswapd (total across ~60 hosts, per minute): ~70.2M
>> Atomic order-4+ allocations: 0
>> 
>> Patched
>> HighAtomic pageblocks per host: 1
>> THP success rate: 44-78%
>> Compaction success rate: 24-47%
>> pgscan_kswapd (total across ~60 hosts, per minute): ~29.9M
>> Atomic order-4+ allocations: 0
> 
> This is an interesting patch. A couple of thoughts:
> 
> 1. You disabled the highatomic reserve for this workload and it didn't
> seem to matter. Presumably <costly orders don't need the protection.
> 
> 2. Maxing out the reserves is odd. ALLOC_HIGHATOMIC allocations will
> try reserved space first,

Hmm, but if the allocation succeeds before entering slowpath,
ALLOC_NON_BLOCK won't be set.
But reserving another block should mean we already exhausted the reserved ones.
Unreserving is only done when direct reclaim made some progress but failed
to produce a page. But if it works, or kswapd does the job, we won't enter it?

> and I'd expect things that are commonly
> highatomic to be short-lived. Why don't we stop with a couple of
> claimed highatomic blocks that get continuously recycled?

Maybe it's some big burst of highatomic allocations that leads to the
reservations and then they stay around "forever"?

If that's the case I think we should be perhaps looking at the unreserving
being done more proactively, rather than limiting things to costly order.

> 3. The impact on THP and compaction success rate is pretty
> extreme. How can 1% of memory throw such a wrench into the gears?

Maybe if ~all free memory is in the highatomic blocks, compaction can't be
effective much. Or some suitability check somewhere in reclaim+compaction
wrongly assumes the highatomic blocks are usable, so it won't do the work.

> Have you tried this with other workloads?