mm, thp: Do not make page table dirty unconditionally in touch_p[mu]d()
authorKirill A. Shutemov <kirill.shutemov@linux.intel.com>
Mon, 27 Nov 2017 03:21:25 +0000 (06:21 +0300)
committerLinus Torvalds <torvalds@linux-foundation.org>
Mon, 27 Nov 2017 20:26:29 +0000 (12:26 -0800)
Currently, we unconditionally make page table dirty in touch_pmd().
It may result in false-positive can_follow_write_pmd().

We may avoid the situation, if we would only make the page table entry
dirty if caller asks for write access -- FOLL_WRITE.

The patch also changes touch_pud() in the same way.

Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Hugh Dickins <hughd@google.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
mm/huge_memory.c

index 86fe697e8bfb3c4e8393b23a09f177965450ad93..0e7ded98d114d184877d2fc9bd0f02c3187f2ed5 100644 (file)
@@ -842,20 +842,15 @@ EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud);
 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
 
 static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
-               pmd_t *pmd)
+               pmd_t *pmd, int flags)
 {
        pmd_t _pmd;
 
-       /*
-        * We should set the dirty bit only for FOLL_WRITE but for now
-        * the dirty bit in the pmd is meaningless.  And if the dirty
-        * bit will become meaningful and we'll only set it with
-        * FOLL_WRITE, an atomic set_bit will be required on the pmd to
-        * set the young bit, instead of the current set_pmd_at.
-        */
-       _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
+       _pmd = pmd_mkyoung(*pmd);
+       if (flags & FOLL_WRITE)
+               _pmd = pmd_mkdirty(_pmd);
        if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
-                               pmd, _pmd,  1))
+                               pmd, _pmd, flags & FOLL_WRITE))
                update_mmu_cache_pmd(vma, addr, pmd);
 }
 
@@ -884,7 +879,7 @@ struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
                return NULL;
 
        if (flags & FOLL_TOUCH)
-               touch_pmd(vma, addr, pmd);
+               touch_pmd(vma, addr, pmd, flags);
 
        /*
         * device mapped pages can only be returned if the
@@ -995,20 +990,15 @@ out:
 
 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
 static void touch_pud(struct vm_area_struct *vma, unsigned long addr,
-               pud_t *pud)
+               pud_t *pud, int flags)
 {
        pud_t _pud;
 
-       /*
-        * We should set the dirty bit only for FOLL_WRITE but for now
-        * the dirty bit in the pud is meaningless.  And if the dirty
-        * bit will become meaningful and we'll only set it with
-        * FOLL_WRITE, an atomic set_bit will be required on the pud to
-        * set the young bit, instead of the current set_pud_at.
-        */
-       _pud = pud_mkyoung(pud_mkdirty(*pud));
+       _pud = pud_mkyoung(*pud);
+       if (flags & FOLL_WRITE)
+               _pud = pud_mkdirty(_pud);
        if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK,
-                               pud, _pud,  1))
+                               pud, _pud, flags & FOLL_WRITE))
                update_mmu_cache_pud(vma, addr, pud);
 }
 
@@ -1031,7 +1021,7 @@ struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr,
                return NULL;
 
        if (flags & FOLL_TOUCH)
-               touch_pud(vma, addr, pud);
+               touch_pud(vma, addr, pud, flags);
 
        /*
         * device mapped pages can only be returned if the
@@ -1424,7 +1414,7 @@ struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
        page = pmd_page(*pmd);
        VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page);
        if (flags & FOLL_TOUCH)
-               touch_pmd(vma, addr, pmd);
+               touch_pmd(vma, addr, pmd, flags);
        if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
                /*
                 * We don't mlock() pte-mapped THPs. This way we can avoid