Re: [GIT PULL] slab fix for 6.18-rc7

[Matthew Wilcox <willy@infradead.org>]

On Thu, Nov 20, 2025 at 10:58:13AM -0800, Linus Torvalds wrote:
> On Thu, 20 Nov 2025 at 10:45, Vlastimil Babka <vbabka@suse.cz> wrote:
> >
> > * Fix mempool poisoning order>0 pages with CONFIG_HIGHMEM (Vlastimil Babka)
> 
> I've pulled this, but honestly, CONFIG_HIGHMEM should be considered a
> dying breed, and I'd have been happier with just not adding extra code
> for that thing.

Would you rather see something like this that hides the fact it's
dealing with HIGHMEM?


diff --git a/include/linux/page-flags.h b/include/linux/page-flags.h
index 0091ad1986bf..68b952e68176 100644
--- a/include/linux/page-flags.h
+++ b/include/linux/page-flags.h
@@ -623,8 +623,10 @@ PAGEFLAG_FALSE(HighMem, highmem)
 /* Does kmap_local_folio() only allow access to one page of the folio? */
 #ifdef CONFIG_DEBUG_KMAP_LOCAL_FORCE_MAP
 #define folio_test_partial_kmap(f)	true
+#define PagePartialKmap(p)		true
 #else
 #define folio_test_partial_kmap(f)	folio_test_highmem(f)
+#define PagePartialKmap(p)		PageHighMem(p)
 #endif
 
 #ifdef CONFIG_SWAP
diff --git a/mm/mempool.c b/mm/mempool.c
index 1c38e873e546..b1f5d81d70c6 100644
--- a/mm/mempool.c
+++ b/mm/mempool.c
@@ -67,11 +67,19 @@ static void check_element(mempool_t *pool, void *element)
 		__check_element(pool, element, kmem_cache_size(pool->pool_data));
 	} else if (pool->free == mempool_free_pages) {
 		/* Mempools backed by page allocator */
-		int order = (int)(long)pool->pool_data;
-		void *addr = kmap_local_page((struct page *)element);
-
-		__check_element(pool, addr, 1UL << (PAGE_SHIFT + order));
-		kunmap_local(addr);
+		size_t len = 1UL << (long)pool->pool_data;
+		do {
+			struct page *page = (struct page *)element;
+			void *addr = kmap_local_page(page);
+			size_t chunk = len;
+
+			if (PagePartialKmap(page) && chunk > PAGE_SIZE)
+				chunk = PAGE_SIZE;
+			__check_element(pool, addr, chunk);
+			kunmap_local(addr);
+			len -= chunk;
+			element += chunk;
+		} while (len > 0);
 	}
 }
 
(not actually tested, but based on memcpy_from_folio())