mm: fix migratetype bug which slowed swapping
[linux-2.6.git] / mm / mincore.c
index 8aca6f7..7a3436e 100644 (file)
@@ -12,6 +12,9 @@
 #include <linux/mm.h>
 #include <linux/mman.h>
 #include <linux/syscalls.h>
+#include <linux/swap.h>
+#include <linux/swapops.h>
+#include <linux/hugetlb.h>
 
 #include <asm/uaccess.h>
 #include <asm/pgtable.h>
  * and is up to date; i.e. that no page-in operation would be required
  * at this time if an application were to map and access this page.
  */
-static unsigned char mincore_page(struct vm_area_struct * vma,
-       unsigned long pgoff)
+static unsigned char mincore_page(struct address_space *mapping, pgoff_t pgoff)
 {
        unsigned char present = 0;
-       struct address_space * as = vma->vm_file->f_mapping;
-       struct page * page;
+       struct page *page;
 
-       page = find_get_page(as, pgoff);
+       /*
+        * When tmpfs swaps out a page from a file, any process mapping that
+        * file will not get a swp_entry_t in its pte, but rather it is like
+        * any other file mapping (ie. marked !present and faulted in with
+        * tmpfs's .fault). So swapped out tmpfs mappings are tested here.
+        *
+        * However when tmpfs moves the page from pagecache and into swapcache,
+        * it is still in core, but the find_get_page below won't find it.
+        * No big deal, but make a note of it.
+        */
+       page = find_get_page(mapping, pgoff);
        if (page) {
                present = PageUptodate(page);
                page_cache_release(page);
@@ -45,7 +56,14 @@ static unsigned char mincore_page(struct vm_area_struct * vma,
  */
 static long do_mincore(unsigned long addr, unsigned char *vec, unsigned long pages)
 {
-       unsigned long i, nr, pgoff;
+       pgd_t *pgd;
+       pud_t *pud;
+       pmd_t *pmd;
+       pte_t *ptep;
+       spinlock_t *ptl;
+       unsigned long nr;
+       int i;
+       pgoff_t pgoff;
        struct vm_area_struct *vma = find_vma(current->mm, addr);
 
        /*
@@ -55,32 +73,119 @@ static long do_mincore(unsigned long addr, unsigned char *vec, unsigned long pag
        if (!vma || addr < vma->vm_start)
                return -ENOMEM;
 
+#ifdef CONFIG_HUGETLB_PAGE
+       if (is_vm_hugetlb_page(vma)) {
+               struct hstate *h;
+               unsigned long nr_huge;
+               unsigned char present;
+
+               i = 0;
+               nr = min(pages, (vma->vm_end - addr) >> PAGE_SHIFT);
+               h = hstate_vma(vma);
+               nr_huge = ((addr + pages * PAGE_SIZE - 1) >> huge_page_shift(h))
+                         - (addr >> huge_page_shift(h)) + 1;
+               nr_huge = min(nr_huge,
+                             (vma->vm_end - addr) >> huge_page_shift(h));
+               while (1) {
+                       /* hugepage always in RAM for now,
+                        * but generally it needs to be check */
+                       ptep = huge_pte_offset(current->mm,
+                                              addr & huge_page_mask(h));
+                       present = !!(ptep &&
+                                    !huge_pte_none(huge_ptep_get(ptep)));
+                       while (1) {
+                               vec[i++] = present;
+                               addr += PAGE_SIZE;
+                               /* reach buffer limit */
+                               if (i == nr)
+                                       return nr;
+                               /* check hugepage border */
+                               if (!((addr & ~huge_page_mask(h))
+                                     >> PAGE_SHIFT))
+                                       break;
+                       }
+               }
+               return nr;
+       }
+#endif
+
        /*
-        * Ok, got it. But check whether it's a segment we support
-        * mincore() on. Right now, we don't do any anonymous mappings.
-        *
-        * FIXME: This is just stupid. And returning ENOMEM is 
-        * stupid too. We should just look at the page tables. But
-        * this is what we've traditionally done, so we'll just
-        * continue doing it.
+        * Calculate how many pages there are left in the last level of the
+        * PTE array for our address.
         */
-       if (!vma->vm_file)
-               return -ENOMEM;
+       nr = PTRS_PER_PTE - ((addr >> PAGE_SHIFT) & (PTRS_PER_PTE-1));
+
+       /*
+        * Don't overrun this vma
+        */
+       nr = min(nr, (vma->vm_end - addr) >> PAGE_SHIFT);
 
        /*
-        * Calculate how many pages there are left in the vma, and
-        * what the pgoff is for our address.
+        * Don't return more than the caller asked for
         */
-       nr = (vma->vm_end - addr) >> PAGE_SHIFT;
-       if (nr > pages)
-               nr = pages;
+       nr = min(nr, pages);
 
-       pgoff = (addr - vma->vm_start) >> PAGE_SHIFT;
-       pgoff += vma->vm_pgoff;
+       pgd = pgd_offset(vma->vm_mm, addr);
+       if (pgd_none_or_clear_bad(pgd))
+               goto none_mapped;
+       pud = pud_offset(pgd, addr);
+       if (pud_none_or_clear_bad(pud))
+               goto none_mapped;
+       pmd = pmd_offset(pud, addr);
+       if (pmd_none_or_clear_bad(pmd))
+               goto none_mapped;
 
-       /* And then we just fill the sucker in.. */
-       for (i = 0 ; i < nr; i++, pgoff++)
-               vec[i] = mincore_page(vma, pgoff);
+       ptep = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
+       for (i = 0; i < nr; i++, ptep++, addr += PAGE_SIZE) {
+               unsigned char present;
+               pte_t pte = *ptep;
+
+               if (pte_present(pte)) {
+                       present = 1;
+
+               } else if (pte_none(pte)) {
+                       if (vma->vm_file) {
+                               pgoff = linear_page_index(vma, addr);
+                               present = mincore_page(vma->vm_file->f_mapping,
+                                                       pgoff);
+                       } else
+                               present = 0;
+
+               } else if (pte_file(pte)) {
+                       pgoff = pte_to_pgoff(pte);
+                       present = mincore_page(vma->vm_file->f_mapping, pgoff);
+
+               } else { /* pte is a swap entry */
+                       swp_entry_t entry = pte_to_swp_entry(pte);
+                       if (is_migration_entry(entry)) {
+                               /* migration entries are always uptodate */
+                               present = 1;
+                       } else {
+#ifdef CONFIG_SWAP
+                               pgoff = entry.val;
+                               present = mincore_page(&swapper_space, pgoff);
+#else
+                               WARN_ON(1);
+                               present = 1;
+#endif
+                       }
+               }
+
+               vec[i] = present;
+       }
+       pte_unmap_unlock(ptep-1, ptl);
+
+       return nr;
+
+none_mapped:
+       if (vma->vm_file) {
+               pgoff = linear_page_index(vma, addr);
+               for (i = 0; i < nr; i++, pgoff++)
+                       vec[i] = mincore_page(vma->vm_file->f_mapping, pgoff);
+       } else {
+               for (i = 0; i < nr; i++)
+                       vec[i] = 0;
+       }
 
        return nr;
 }
@@ -109,8 +214,8 @@ static long do_mincore(unsigned long addr, unsigned char *vec, unsigned long pag
  *             mapped
  *  -EAGAIN - A kernel resource was temporarily unavailable.
  */
-asmlinkage long sys_mincore(unsigned long start, size_t len,
-       unsigned char __user * vec)
+SYSCALL_DEFINE3(mincore, unsigned long, start, size_t, len,
+               unsigned char __user *, vec)
 {
        long retval;
        unsigned long pages;