mm/page_alloc: reduce potential fragmentation in make_alloc_exact()
authorLiam R. Howlett <Liam.Howlett@Oracle.com>
Tue, 31 May 2022 13:20:51 +0000 (09:20 -0400)
committerAndrew Morton <akpm@linux-foundation.org>
Fri, 21 Oct 2022 04:27:23 +0000 (21:27 -0700)
Try to avoid using the left over split page on the next request for a page
by calling __free_pages_ok() with FPI_TO_TAIL.  This increases the
potential of defragmenting memory when it's used for a short period of
time.

Link: https://lkml.kernel.org/r/20220531185626.yvlmymbxyoe5vags@revolver
Signed-off-by: Liam R. Howlett <Liam.Howlett@oracle.com>
Suggested-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
mm/page_alloc.c

index e20ade8..b5a6c81 100644 (file)
@@ -5784,14 +5784,18 @@ static void *make_alloc_exact(unsigned long addr, unsigned int order,
                size_t size)
 {
        if (addr) {
-               unsigned long alloc_end = addr + (PAGE_SIZE << order);
-               unsigned long used = addr + PAGE_ALIGN(size);
-
-               split_page(virt_to_page((void *)addr), order);
-               while (used < alloc_end) {
-                       free_page(used);
-                       used += PAGE_SIZE;
-               }
+               unsigned long nr = DIV_ROUND_UP(size, PAGE_SIZE);
+               struct page *page = virt_to_page((void *)addr);
+               struct page *last = page + nr;
+
+               split_page_owner(page, 1 << order);
+               split_page_memcg(page, 1 << order);
+               while (page < --last)
+                       set_page_refcounted(last);
+
+               last = page + (1UL << order);
+               for (page += nr; page < last; page++)
+                       __free_pages_ok(page, 0, FPI_TO_TAIL);
        }
        return (void *)addr;
 }