Re: [PATCHv4 09/14] mm/hugetlb: Remove fake head pages

[Kiryl Shutsemau <kas@kernel.org>]

On Wed, Jan 28, 2026 at 10:43:13AM +0800, Muchun Song wrote:
> 
> 
> > On Jan 27, 2026, at 22:51, Kiryl Shutsemau <kas@kernel.org> wrote:
> > 
> > On Thu, Jan 22, 2026 at 03:00:03PM +0800, Muchun Song wrote:
> >>> + if (pfn)
> >>> + 	return pfn_to_page(pfn);
> >>> +
> >>> + tail = alloc_pages_node(node, GFP_KERNEL | __GFP_ZERO, 0);
> >>> + if (!tail)
> >>> + 	return NULL;
> >>> +
> >>> + p = page_to_virt(tail);
> >>> + for (int i = 0; i < PAGE_SIZE / sizeof(struct page); i++)
> >>> + 	prep_compound_tail(p + i, NULL, order);
> >>> +
> >>> + spin_lock(&hugetlb_lock);
> >> 
> >> hugetlb_lock is considered a contended lock, better not to abuse it.
> >> cmpxchg() is enought in this case.
> > 
> > We hit the lock once per node (excluding races). Its contribution to the
> > lock contention is negligible. spin_lock() is easier to follow. I will
> > keep it.
> 
> I don't think cmpxchg() is hard to follow. It’s precisely because of
> your abuse that interrupts still have to be disabled here—hugetlb_lock
> must be an irq-off lock. Are you really going to use spin_lock_irq just
> because “it feels simpler” to you?

I looked again at it and reconsidered. I will use cmpxchg(), but mostly
because hugetlb_lock is a bad fit to protect anything in pg_data_t.
vmemmap_tails can be used by code outside hugetlb.

Here's the fixup.

diff --git a/mm/hugetlb_vmemmap.c b/mm/hugetlb_vmemmap.c
index 29e9bbb43178..63e7ca85c8c9 100644
--- a/mm/hugetlb_vmemmap.c
+++ b/mm/hugetlb_vmemmap.c
@@ -512,18 +512,11 @@ static struct page *vmemmap_get_tail(unsigned int order, int node)
 	for (int i = 0; i < PAGE_SIZE / sizeof(struct page); i++)
 		prep_compound_tail(p + i, NULL, order);
 
-	spin_lock(&hugetlb_lock);
-	if (!NODE_DATA(node)->vmemmap_tails[idx]) {
-		pfn = PHYS_PFN(virt_to_phys(p));
-		NODE_DATA(node)->vmemmap_tails[idx] = pfn;
-		tail = NULL;
-	} else {
-		pfn = NODE_DATA(node)->vmemmap_tails[idx];
-	}
-	spin_unlock(&hugetlb_lock);
-
-	if (tail)
+	pfn = PHYS_PFN(virt_to_phys(p));
+	if (cmpxchg(&NODE_DATA(node)->vmemmap_tails[idx], 0, pfn)) {
 		__free_page(tail);
+		pfn = READ_ONCE(NODE_DATA(node)->vmemmap_tails[idx]);
+	}
 
 	return pfn_to_page(pfn);
 }
-- 
  Kiryl Shutsemau / Kirill A. Shutemov