thp: fix another corner case of munlock() vs. THPs

The following test case triggers BUG() in munlock_vma_pages_range():

	int main(int argc, char *argv[])
	{
		int fd;

		system("mount -t tmpfs -o huge=always none /mnt");
		fd = open("/mnt/test", O_CREAT | O_RDWR);
		ftruncate(fd, 4UL << 20);
		mmap(NULL, 4UL << 20, PROT_READ | PROT_WRITE,
				MAP_SHARED | MAP_FIXED | MAP_LOCKED, fd, 0);
		mmap(NULL, 4096, PROT_READ | PROT_WRITE,
				MAP_SHARED | MAP_LOCKED, fd, 0);
		munlockall();
		return 0;
	}

The second mmap() create PTE-mapping of the first huge page in file.  It
makes kernel munlock the page as we never keep PTE-mapped page mlocked.

On munlockall() when we handle vma created by the first mmap(),
munlock_vma_page() returns page_mask == 0, as the page is not mlocked
anymore.  On next iteration follow_page_mask() return tail page, but
page_mask is HPAGE_NR_PAGES - 1.  It makes us skip to the first tail
page of the next huge page and step on
VM_BUG_ON_PAGE(PageMlocked(page)).

The fix is not use the page_mask from follow_page_mask() at all.  It has
no use for us.

Link: http://lkml.kernel.org/r/20170302150252.34120-1-kirill.shutemov@linux.intel.com
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: <stable@vger.kernel.org>    [4.5+]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

diff --git a/mm/mlock.c b/mm/mlock.c
index 1050511f8b2bdbfbd55a69b8fcfc41ea6cd51b72..02f138244bf5ee9ed384aa3fa244c99da8eb1ced 100644 (file)
--- a/mm/mlock.c
+++ b/mm/mlock.c
@@ -442,7 +442,7 @@ void munlock_vma_pages_range(struct vm_area_struct *vma,
 
        while (start < end) {
                struct page *page;
-               unsigned int page_mask;
+               unsigned int page_mask = 0;
                unsigned long page_increm;
                struct pagevec pvec;
                struct zone *zone;
@@ -456,8 +456,7 @@ void munlock_vma_pages_range(struct vm_area_struct *vma,
                 * suits munlock very well (and if somehow an abnormal page
                 * has sneaked into the range, we won't oops here: great).
                 */
-               page = follow_page_mask(vma, start, FOLL_GET | FOLL_DUMP,
-                               &page_mask);
+               page = follow_page(vma, start, FOLL_GET | FOLL_DUMP);
 
                if (page && !IS_ERR(page)) {
                        if (PageTransTail(page)) {
@@ -468,8 +467,8 @@ void munlock_vma_pages_range(struct vm_area_struct *vma,
                                /*
                                 * Any THP page found by follow_page_mask() may
                                 * have gotten split before reaching
-                                * munlock_vma_page(), so we need to recompute
-                                * the page_mask here.
+                                * munlock_vma_page(), so we need to compute
+                                * the page_mask here instead.
                                 */
                                page_mask = munlock_vma_page(page);
                                unlock_page(page);