hugetlb: fix huge_pmd_unshare address update

The routine huge_pmd_unshare() is passed a pointer to an address
associated with an area which may be unshared.  If unshare is successful
this address is updated to 'optimize' callers iterating over huge page
addresses.  For the optimization to work correctly, address should be
updated to the last huge page in the unmapped/unshared area.  However, in
the common case where the passed address is PUD_SIZE aligned, the address
is incorrectly updated to the address of the preceding huge page.  That
wastes CPU cycles as the unmapped/unshared range is scanned twice.

Link: https://lkml.kernel.org/r/20220524205003.126184-1-mike.kravetz@oracle.com
Fixes: 39dde65c9940 ("shared page table for hugetlb page")
Signed-off-by: Mike Kravetz <mike.kravetz@oracle.com>
Acked-by: Muchun Song <songmuchun@bytedance.com>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>

diff --git a/mm/hugetlb.c b/mm/hugetlb.c
index 3fc721789743e6adb95d3005067392c9d7b53dff..410bbb0aee321cef463f3b26d8fc310313738a5b 100644 (file)
--- a/mm/hugetlb.c
+++ b/mm/hugetlb.c
@@ -6562,7 +6562,14 @@ int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
        pud_clear(pud);
        put_page(virt_to_page(ptep));
        mm_dec_nr_pmds(mm);
-       *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
+       /*
+        * This update of passed address optimizes loops sequentially
+        * processing addresses in increments of huge page size (PMD_SIZE
+        * in this case).  By clearing the pud, a PUD_SIZE area is unmapped.
+        * Update address to the 'last page' in the cleared area so that
+        * calling loop can move to first page past this area.
+        */
+       *addr |= PUD_SIZE - PMD_SIZE;
        return 1;
 }