Re: [PATCH v2 5/7] mm/page_ext: allocate page extensions early for SPARSEMEM

From: Muchun Song

Date: Wed Nov 18 2026 - 10:38:34 EST




On 2026/11/18 22:17, Mike Rapoport wrote:
On Tue, Nov 17, 2026 at 11:00:06PM -0800, Chris Down wrote:
Hi Muchun,

Muchun Song writes:
Mike's suggestion considerably simplifies the code, and early page_ext is
only needed by optional debugging and instrumentation users. I would prefer
to keep this approach.
You're probably right.
I think so :)
The memap_alloc() that essentially allocates the memory map does

memblock_alloc_try_nid_raw(size, align, min_addr,
MEMBLOCK_ALLOC_NOLEAKTRACE, nid);

so we anyway may end up with reserved memory in ZONE_MOVABLE.

Until recently we had memory map allocation before the zones were setup, so
there wasn't much we could do about it, but ...

But can I cheekily suggest something in the middle perhaps? :-)

Given the zones are already set up before page_ext_alloc() runs, perhaps
you can pass the start of the node's ZONE_MOVABLE as max_addr to
memblock_alloc_try_nid() when the node has one. The general idea is:

struct zone *movable = &NODE_DATA(nid)->node_zones[ZONE_MOVABLE];
phys_addr_t max_addr = MEMBLOCK_ALLOC_ACCESSIBLE;

if (!zone_is_empty(movable))
max_addr = PFN_PHYS(movable->zone_start_pfn);
addr = memblock_alloc_try_nid(size, PAGE_SIZE, __pa(MAX_DMA_ADDRESS),
max_addr, nid);
if (!addr && max_addr != MEMBLOCK_ALLOC_ACCESSIBLE)
addr = memblock_alloc_try_nid(size, PAGE_SIZE,
__pa(MAX_DMA_ADDRESS),
MEMBLOCK_ALLOC_ACCESSIBLE, nid);
... with zone extents known we could take it even further :)

page_ext is treated as the memory map in many ways, so we can allocate the
early page_ext sections with memmap_alloc().
And since the latter always runs after zones are set up it could include
this change.

But I think this should come as a separate series that also could move
memmap_boot_pages_add to memmap_alloc().

Agreed. Using memmap_alloc() for early page_ext allocations and
teaching it to avoid ZONE_MOVABLE sounds cleaner. I'll keep this
series as is and handle that in a separate series.

Thanks!

I briefly tried that and it seems to work quite nicely.

But I don't object to it even without that, so feel free to add either way:

Reviewed-by: Chris Down <chris@xxxxxxxxxxxxxx>

Thanks!

Chris