[RFC PATCH 09/40] mm: page_alloc: support superpageblock resize for memory hotplug

[Rik van Riel]

setup_superpageblocks() is __init-only and uses memblock_alloc_node(), so
hotplugged memory that extends a zone's span has no superpageblock
coverage.  Pages in those regions would bypass superpageblock steering
entirely.

Add resize_zone_superpageblocks() which is called from
move_pfn_range_to_zone() after the zone span has been updated. It allocates
a new superpageblock array with kvmalloc_node() covering the full zone
span, copies existing superpageblocks (fixing up list head pointers), and
initializes new superpageblocks for the added range.

Use round-up division for partial pageblock counting to match
init_one_superpageblock().

ZONE_DEVICE is excluded since device pages should not participate in anti-
fragmentation steering.

Signed-off-by: Rik van Riel <riel@xxxxxxxxxxx>
Assisted-by: Claude:claude-opus-4.7 syzkaller
---
 include/linux/mmzone.h |   1 +
 mm/internal.h          |   4 ++
 mm/memory_hotplug.c    |   4 ++
 mm/mm_init.c           | 138 +++++++++++++++++++++++++++++++++++++++++
 4 files changed, 147 insertions(+)

diff --git a/include/linux/mmzone.h b/include/linux/mmzone.h
index e3eac971a76a..19190328e0c7 100644
--- a/include/linux/mmzone.h
+++ b/include/linux/mmzone.h
@@ -1057,6 +1057,7 @@ struct zone {
 	struct superpageblock	*superpageblocks;
 	unsigned long		nr_superpageblocks;
 	unsigned long		superpageblock_base_pfn; /* 1GB-aligned base */
+	bool			spb_kvmalloced; /* true if from kvmalloc (hotplug) */
 
 	/* zone_start_pfn == zone_start_paddr >> PAGE_SHIFT */
 	unsigned long		zone_start_pfn;
diff --git a/mm/internal.h b/mm/internal.h
index c8404cb00b08..6a089bc4aa09 100644
--- a/mm/internal.h
+++ b/mm/internal.h
@@ -1101,6 +1101,10 @@ void init_cma_reserved_pageblock(struct page *page);
 
 #endif /* CONFIG_COMPACTION || CONFIG_CMA */
 
+#ifdef CONFIG_MEMORY_HOTPLUG
+void resize_zone_superpageblocks(struct zone *zone);
+#endif
+
 struct cma;
 
 #ifdef CONFIG_CMA
diff --git a/mm/memory_hotplug.c b/mm/memory_hotplug.c
index 2a943ec57c85..b7c30dfdce8e 100644
--- a/mm/memory_hotplug.c
+++ b/mm/memory_hotplug.c
@@ -752,6 +752,10 @@ void move_pfn_range_to_zone(struct zone *zone, unsigned long start_pfn,
 	resize_zone_range(zone, start_pfn, nr_pages);
 	resize_pgdat_range(pgdat, start_pfn, nr_pages);
 
+	/* Grow superpageblock array to cover the new zone span */
+	if (!zone_is_zone_device(zone))
+		resize_zone_superpageblocks(zone);
+
 	/*
 	 * Subsection population requires care in pfn_to_online_page().
 	 * Set the taint to enable the slow path detection of
diff --git a/mm/mm_init.c b/mm/mm_init.c
index de02a6087c21..ad1cbc2b4498 100644
--- a/mm/mm_init.c
+++ b/mm/mm_init.c
@@ -1592,6 +1592,144 @@ static void __init setup_superpageblocks(struct zone *zone)
 					zone_start, zone_end);
 }
 
+#ifdef CONFIG_MEMORY_HOTPLUG
+/**
+ * resize_zone_superpageblocks - grow superpageblock array for memory hotplug
+ * @zone: zone whose span has been extended by hotplug
+ *
+ * Called from move_pfn_range_to_zone() after resize_zone_range() has
+ * updated the zone's span.  Allocates a new superpageblock array covering
+ * the full zone span, copies existing superpageblocks (fixing up list heads),
+ * and initializes new superpageblocks for the added range.
+ *
+ * Must be called under mem_hotplug_lock (write).  No concurrent
+ * allocations can occur since the hotplugged pages are not yet online.
+ */
+void __meminit resize_zone_superpageblocks(struct zone *zone)
+{
+	unsigned long zone_start = zone->zone_start_pfn;
+	unsigned long zone_end = zone_start + zone->spanned_pages;
+	unsigned long new_sb_base, new_nr_sbs;
+	unsigned long old_offset;
+	struct superpageblock *old_sbs;
+	struct superpageblock *new_sbs;
+	bool old_kvmalloced;
+	size_t alloc_size;
+	unsigned long i;
+	int nid = zone_to_nid(zone);
+
+	if (!zone->spanned_pages)
+		return;
+
+	new_sb_base = ALIGN_DOWN(zone_start, SUPERPAGEBLOCK_NR_PAGES);
+	new_nr_sbs = (ALIGN(zone_end, SUPERPAGEBLOCK_NR_PAGES) - new_sb_base) >>
+		     SUPERPAGEBLOCK_ORDER;
+
+	/* Already covered? */
+	if (zone->superpageblocks &&
+	    new_sb_base == zone->superpageblock_base_pfn &&
+	    new_nr_sbs == zone->nr_superpageblocks)
+		return;
+
+	alloc_size = new_nr_sbs * sizeof(struct superpageblock);
+	new_sbs = kvmalloc_node(alloc_size, GFP_KERNEL | __GFP_ZERO, nid);
+	if (!new_sbs) {
+		pr_warn("Failed to allocate %zu bytes for zone %s superpageblocks\n",
+			alloc_size, zone->name);
+		return;
+	}
+
+	/*
+	 * Copy existing superpageblocks to their new position.
+	 * The old array covers [old_base, old_base + old_nr * SB_SIZE).
+	 * The new array covers [new_base, new_base + new_nr * SB_SIZE).
+	 * old_base >= new_base always (zone can only grow).
+	 */
+	if (zone->superpageblocks) {
+		old_offset = (zone->superpageblock_base_pfn - new_sb_base) >>
+			     SUPERPAGEBLOCK_ORDER;
+		memcpy(&new_sbs[old_offset], zone->superpageblocks,
+		       zone->nr_superpageblocks * sizeof(struct superpageblock));
+
+		/*
+		 * Fix up list_head pointers that were self-referencing
+		 * (empty lists) or pointing into the old array.
+		 */
+		for (i = old_offset; i < old_offset + zone->nr_superpageblocks; i++) {
+			struct superpageblock *sb = &new_sbs[i];
+
+			if (list_empty(&sb->list))
+				INIT_LIST_HEAD(&sb->list);
+			else
+				list_replace(&zone->superpageblocks[i - old_offset].list,
+					     &sb->list);
+		}
+	}
+
+	/* Initialize new superpageblocks (slots not covered by old array) */
+	for (i = 0; i < new_nr_sbs; i++) {
+		struct superpageblock *sb = &new_sbs[i];
+		bool is_old = false;
+
+		if (zone->superpageblocks) {
+			old_offset = (zone->superpageblock_base_pfn - new_sb_base) >>
+				     SUPERPAGEBLOCK_ORDER;
+			if (i >= old_offset &&
+			    i < old_offset + zone->nr_superpageblocks)
+				is_old = true;
+		}
+
+		if (is_old)
+			continue;
+
+		init_one_superpageblock(sb, zone,
+					new_sb_base + (i << SUPERPAGEBLOCK_ORDER),
+					zone_start, zone_end);
+	}
+
+	/*
+	 * Update existing superpageblocks whose nr_reserved may have
+	 * increased due to the zone span growing into them.
+	 */
+	if (zone->superpageblocks) {
+		old_offset = (zone->superpageblock_base_pfn - new_sb_base) >>
+			     SUPERPAGEBLOCK_ORDER;
+		for (i = old_offset; i < old_offset + zone->nr_superpageblocks; i++) {
+			struct superpageblock *sb = &new_sbs[i];
+			unsigned long sb_start = sb->start_pfn;
+			unsigned long sb_end = sb_start + SUPERPAGEBLOCK_NR_PAGES;
+			unsigned long pb_start = max(sb_start, zone_start);
+			unsigned long pb_end = min(sb_end, zone_end);
+			u16 new_pbs = (pb_end > pb_start) ?
+				((pb_end - pb_start + pageblock_nr_pages - 1) >>
+				 pageblock_order) : 0;
+			u16 old_pbs = sb->nr_free + sb->nr_unmovable +
+				sb->nr_reclaimable + sb->nr_movable +
+				sb->nr_reserved;
+
+			if (new_pbs > old_pbs)
+				sb->nr_reserved += new_pbs - old_pbs;
+		}
+	}
+
+	/* Swap in the new array */
+	old_sbs = zone->superpageblocks;
+	old_kvmalloced = zone->spb_kvmalloced;
+	zone->superpageblocks = new_sbs;
+	zone->nr_superpageblocks = new_nr_sbs;
+	zone->superpageblock_base_pfn = new_sb_base;
+	zone->spb_kvmalloced = true;
+
+	/*
+	 * The boot-time array was allocated with memblock_alloc, which
+	 * is not individually freeable after boot.  Only kvfree arrays
+	 * from previous hotplug resizes.
+	 */
+	if (old_sbs && old_kvmalloced)
+		kvfree(old_sbs);
+}
+#endif /* CONFIG_MEMORY_HOTPLUG */
+
 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
 
 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
-- 
2.54.0