mempolicy: apply page table walker on queue_pages_range()
authorNaoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Wed, 11 Feb 2015 23:28:03 +0000 (15:28 -0800)
committerLinus Torvalds <torvalds@linux-foundation.org>
Thu, 12 Feb 2015 01:06:06 +0000 (17:06 -0800)
queue_pages_range() does page table walking in its own way now, but there
is some code duplicate.  This patch applies page table walker to reduce
lines of code.

queue_pages_range() has to do some precheck to determine whether we really
walk over the vma or just skip it.  Now we have test_walk() callback in
mm_walk for this purpose, so we can do this replacement cleanly.
queue_pages_test_walk() depends on not only the current vma but also the
previous one, so queue_pages->prev is introduced to remember it.

Signed-off-by: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Cyrill Gorcunov <gorcunov@openvz.org>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Pavel Emelyanov <xemul@parallels.com>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
mm/mempolicy.c

index acbbf4c..b1dcd11 100644 (file)
@@ -471,24 +471,34 @@ static const struct mempolicy_operations mpol_ops[MPOL_MAX] = {
 static void migrate_page_add(struct page *page, struct list_head *pagelist,
                                unsigned long flags);
 
+struct queue_pages {
+       struct list_head *pagelist;
+       unsigned long flags;
+       nodemask_t *nmask;
+       struct vm_area_struct *prev;
+};
+
 /*
  * Scan through pages checking if pages follow certain conditions,
  * and move them to the pagelist if they do.
  */
-static int queue_pages_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
-               unsigned long addr, unsigned long end,
-               const nodemask_t *nodes, unsigned long flags,
-               void *private)
+static int queue_pages_pte_range(pmd_t *pmd, unsigned long addr,
+                       unsigned long end, struct mm_walk *walk)
 {
-       pte_t *orig_pte;
+       struct vm_area_struct *vma = walk->vma;
+       struct page *page;
+       struct queue_pages *qp = walk->private;
+       unsigned long flags = qp->flags;
+       int nid;
        pte_t *pte;
        spinlock_t *ptl;
 
-       orig_pte = pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
-       do {
-               struct page *page;
-               int nid;
+       split_huge_page_pmd(vma, addr, pmd);
+       if (pmd_trans_unstable(pmd))
+               return 0;
 
+       pte = pte_offset_map_lock(walk->mm, pmd, addr, &ptl);
+       for (; addr != end; pte++, addr += PAGE_SIZE) {
                if (!pte_present(*pte))
                        continue;
                page = vm_normal_page(vma, addr, *pte);
@@ -501,114 +511,46 @@ static int queue_pages_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
                if (PageReserved(page))
                        continue;
                nid = page_to_nid(page);
-               if (node_isset(nid, *nodes) == !!(flags & MPOL_MF_INVERT))
+               if (node_isset(nid, *qp->nmask) == !!(flags & MPOL_MF_INVERT))
                        continue;
 
                if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
-                       migrate_page_add(page, private, flags);
-               else
-                       break;
-       } while (pte++, addr += PAGE_SIZE, addr != end);
-       pte_unmap_unlock(orig_pte, ptl);
-       return addr != end;
+                       migrate_page_add(page, qp->pagelist, flags);
+       }
+       pte_unmap_unlock(pte - 1, ptl);
+       cond_resched();
+       return 0;
 }
 
-static void queue_pages_hugetlb_pmd_range(struct vm_area_struct *vma,
-               pmd_t *pmd, const nodemask_t *nodes, unsigned long flags,
-                                   void *private)
+static int queue_pages_hugetlb(pte_t *pte, unsigned long hmask,
+                              unsigned long addr, unsigned long end,
+                              struct mm_walk *walk)
 {
 #ifdef CONFIG_HUGETLB_PAGE
+       struct queue_pages *qp = walk->private;
+       unsigned long flags = qp->flags;
        int nid;
        struct page *page;
        spinlock_t *ptl;
        pte_t entry;
 
-       ptl = huge_pte_lock(hstate_vma(vma), vma->vm_mm, (pte_t *)pmd);
-       entry = huge_ptep_get((pte_t *)pmd);
+       ptl = huge_pte_lock(hstate_vma(walk->vma), walk->mm, pte);
+       entry = huge_ptep_get(pte);
        if (!pte_present(entry))
                goto unlock;
        page = pte_page(entry);
        nid = page_to_nid(page);
-       if (node_isset(nid, *nodes) == !!(flags & MPOL_MF_INVERT))
+       if (node_isset(nid, *qp->nmask) == !!(flags & MPOL_MF_INVERT))
                goto unlock;
        /* With MPOL_MF_MOVE, we migrate only unshared hugepage. */
        if (flags & (MPOL_MF_MOVE_ALL) ||
            (flags & MPOL_MF_MOVE && page_mapcount(page) == 1))
-               isolate_huge_page(page, private);
+               isolate_huge_page(page, qp->pagelist);
 unlock:
        spin_unlock(ptl);
 #else
        BUG();
 #endif
-}
-
-static inline int queue_pages_pmd_range(struct vm_area_struct *vma, pud_t *pud,
-               unsigned long addr, unsigned long end,
-               const nodemask_t *nodes, unsigned long flags,
-               void *private)
-{
-       pmd_t *pmd;
-       unsigned long next;
-
-       pmd = pmd_offset(pud, addr);
-       do {
-               next = pmd_addr_end(addr, end);
-               if (!pmd_present(*pmd))
-                       continue;
-               if (pmd_huge(*pmd) && is_vm_hugetlb_page(vma)) {
-                       queue_pages_hugetlb_pmd_range(vma, pmd, nodes,
-                                               flags, private);
-                       continue;
-               }
-               split_huge_page_pmd(vma, addr, pmd);
-               if (pmd_none_or_trans_huge_or_clear_bad(pmd))
-                       continue;
-               if (queue_pages_pte_range(vma, pmd, addr, next, nodes,
-                                   flags, private))
-                       return -EIO;
-       } while (pmd++, addr = next, addr != end);
-       return 0;
-}
-
-static inline int queue_pages_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
-               unsigned long addr, unsigned long end,
-               const nodemask_t *nodes, unsigned long flags,
-               void *private)
-{
-       pud_t *pud;
-       unsigned long next;
-
-       pud = pud_offset(pgd, addr);
-       do {
-               next = pud_addr_end(addr, end);
-               if (pud_huge(*pud) && is_vm_hugetlb_page(vma))
-                       continue;
-               if (pud_none_or_clear_bad(pud))
-                       continue;
-               if (queue_pages_pmd_range(vma, pud, addr, next, nodes,
-                                   flags, private))
-                       return -EIO;
-       } while (pud++, addr = next, addr != end);
-       return 0;
-}
-
-static inline int queue_pages_pgd_range(struct vm_area_struct *vma,
-               unsigned long addr, unsigned long end,
-               const nodemask_t *nodes, unsigned long flags,
-               void *private)
-{
-       pgd_t *pgd;
-       unsigned long next;
-
-       pgd = pgd_offset(vma->vm_mm, addr);
-       do {
-               next = pgd_addr_end(addr, end);
-               if (pgd_none_or_clear_bad(pgd))
-                       continue;
-               if (queue_pages_pud_range(vma, pgd, addr, next, nodes,
-                                   flags, private))
-                       return -EIO;
-       } while (pgd++, addr = next, addr != end);
        return 0;
 }
 
@@ -641,6 +583,46 @@ static unsigned long change_prot_numa(struct vm_area_struct *vma,
 }
 #endif /* CONFIG_NUMA_BALANCING */
 
+static int queue_pages_test_walk(unsigned long start, unsigned long end,
+                               struct mm_walk *walk)
+{
+       struct vm_area_struct *vma = walk->vma;
+       struct queue_pages *qp = walk->private;
+       unsigned long endvma = vma->vm_end;
+       unsigned long flags = qp->flags;
+
+       if (endvma > end)
+               endvma = end;
+       if (vma->vm_start > start)
+               start = vma->vm_start;
+
+       if (!(flags & MPOL_MF_DISCONTIG_OK)) {
+               if (!vma->vm_next && vma->vm_end < end)
+                       return -EFAULT;
+               if (qp->prev && qp->prev->vm_end < vma->vm_start)
+                       return -EFAULT;
+       }
+
+       qp->prev = vma;
+
+       if (vma->vm_flags & VM_PFNMAP)
+               return 1;
+
+       if (flags & MPOL_MF_LAZY) {
+               /* Similar to task_numa_work, skip inaccessible VMAs */
+               if (vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE))
+                       change_prot_numa(vma, start, endvma);
+               return 1;
+       }
+
+       if ((flags & MPOL_MF_STRICT) ||
+           ((flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) &&
+            vma_migratable(vma)))
+               /* queue pages from current vma */
+               return 0;
+       return 1;
+}
+
 /*
  * Walk through page tables and collect pages to be migrated.
  *
@@ -650,50 +632,24 @@ static unsigned long change_prot_numa(struct vm_area_struct *vma,
  */
 static int
 queue_pages_range(struct mm_struct *mm, unsigned long start, unsigned long end,
-               const nodemask_t *nodes, unsigned long flags, void *private)
-{
-       int err = 0;
-       struct vm_area_struct *vma, *prev;
-
-       vma = find_vma(mm, start);
-       if (!vma)
-               return -EFAULT;
-       prev = NULL;
-       for (; vma && vma->vm_start < end; vma = vma->vm_next) {
-               unsigned long endvma = vma->vm_end;
-
-               if (endvma > end)
-                       endvma = end;
-               if (vma->vm_start > start)
-                       start = vma->vm_start;
-
-               if (!(flags & MPOL_MF_DISCONTIG_OK)) {
-                       if (!vma->vm_next && vma->vm_end < end)
-                               return -EFAULT;
-                       if (prev && prev->vm_end < vma->vm_start)
-                               return -EFAULT;
-               }
-
-               if (flags & MPOL_MF_LAZY) {
-                       /* Similar to task_numa_work, skip inaccessible VMAs */
-                       if (vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE))
-                               change_prot_numa(vma, start, endvma);
-                       goto next;
-               }
-
-               if ((flags & MPOL_MF_STRICT) ||
-                    ((flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) &&
-                     vma_migratable(vma))) {
-
-                       err = queue_pages_pgd_range(vma, start, endvma, nodes,
-                                               flags, private);
-                       if (err)
-                               break;
-               }
-next:
-               prev = vma;
-       }
-       return err;
+               nodemask_t *nodes, unsigned long flags,
+               struct list_head *pagelist)
+{
+       struct queue_pages qp = {
+               .pagelist = pagelist,
+               .flags = flags,
+               .nmask = nodes,
+               .prev = NULL,
+       };
+       struct mm_walk queue_pages_walk = {
+               .hugetlb_entry = queue_pages_hugetlb,
+               .pmd_entry = queue_pages_pte_range,
+               .test_walk = queue_pages_test_walk,
+               .mm = mm,
+               .private = &qp,
+       };
+
+       return walk_page_range(start, end, &queue_pages_walk);
 }
 
 /*