mm: fix migratetype bug which slowed swapping
[linux-2.6.git] / mm / mincore.c
index 9c1d0a4..7a3436e 100644 (file)
@@ -14,6 +14,7 @@
 #include <linux/syscalls.h>
 #include <linux/swap.h>
 #include <linux/swapops.h>
+#include <linux/hugetlb.h>
 
 #include <asm/uaccess.h>
 #include <asm/pgtable.h>
@@ -33,7 +34,7 @@ static unsigned char mincore_page(struct address_space *mapping, pgoff_t 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 .nopage). So swapped out tmpfs mappings are tested here.
+        * 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.
@@ -72,13 +73,57 @@ 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
+
        /*
         * Calculate how many pages there are left in the last level of the
         * PTE array for our address.
         */
        nr = PTRS_PER_PTE - ((addr >> PAGE_SHIFT) & (PTRS_PER_PTE-1));
-       if (nr > pages)
-               nr = pages;
+
+       /*
+        * Don't overrun this vma
+        */
+       nr = min(nr, (vma->vm_end - addr) >> PAGE_SHIFT);
+
+       /*
+        * Don't return more than the caller asked for
+        */
+       nr = min(nr, pages);
 
        pgd = pgd_offset(vma->vm_mm, addr);
        if (pgd_none_or_clear_bad(pgd))
@@ -169,8 +214,8 @@ none_mapped:
  *             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;