[PATCH] mm: hugetlb truncation fixes
authorHugh Dickins <hugh@veritas.com>
Thu, 20 Oct 2005 04:23:43 +0000 (21:23 -0700)
committerLinus Torvalds <torvalds@g5.osdl.org>
Thu, 20 Oct 2005 06:04:30 +0000 (23:04 -0700)
hugetlbfs allows truncation of its files (should it?), but hugetlb.c often
forgets that: crashes and misaccounting ensue.

copy_hugetlb_page_range better grab the src page_table_lock since we don't
want to guess what happens if concurrently truncated.  unmap_hugepage_range
rss accounting must not assume the full range was mapped.  follow_hugetlb_page
must guard with page_table_lock and be prepared to exit early.

Restyle copy_hugetlb_page_range with a for loop like the others there.

Signed-off-by: Hugh Dickins <hugh@veritas.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
mm/hugetlb.c

index 901ac523a1c39fc17cc15d626ffb9cc8aa7f083d..a1b30d45459e827249da8c6226ccf080f15d65a6 100644 (file)
@@ -274,21 +274,22 @@ int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
 {
        pte_t *src_pte, *dst_pte, entry;
        struct page *ptepage;
-       unsigned long addr = vma->vm_start;
-       unsigned long end = vma->vm_end;
+       unsigned long addr;
 
-       while (addr < end) {
+       for (addr = vma->vm_start; addr < vma->vm_end; addr += HPAGE_SIZE) {
                dst_pte = huge_pte_alloc(dst, addr);
                if (!dst_pte)
                        goto nomem;
+               spin_lock(&src->page_table_lock);
                src_pte = huge_pte_offset(src, addr);
-               BUG_ON(!src_pte || pte_none(*src_pte)); /* prefaulted */
-               entry = *src_pte;
-               ptepage = pte_page(entry);
-               get_page(ptepage);
-               add_mm_counter(dst, rss, HPAGE_SIZE / PAGE_SIZE);
-               set_huge_pte_at(dst, addr, dst_pte, entry);
-               addr += HPAGE_SIZE;
+               if (src_pte && !pte_none(*src_pte)) {
+                       entry = *src_pte;
+                       ptepage = pte_page(entry);
+                       get_page(ptepage);
+                       add_mm_counter(dst, rss, HPAGE_SIZE / PAGE_SIZE);
+                       set_huge_pte_at(dst, addr, dst_pte, entry);
+               }
+               spin_unlock(&src->page_table_lock);
        }
        return 0;
 
@@ -323,8 +324,8 @@ void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
 
                page = pte_page(pte);
                put_page(page);
+               add_mm_counter(mm, rss,  - (HPAGE_SIZE / PAGE_SIZE));
        }
-       add_mm_counter(mm, rss,  -((end - start) >> PAGE_SHIFT));
        flush_tlb_range(vma, start, end);
 }
 
@@ -403,6 +404,7 @@ int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
        BUG_ON(!is_vm_hugetlb_page(vma));
 
        vpfn = vaddr/PAGE_SIZE;
+       spin_lock(&mm->page_table_lock);
        while (vaddr < vma->vm_end && remainder) {
 
                if (pages) {
@@ -415,8 +417,13 @@ int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
                         * indexing below to work. */
                        pte = huge_pte_offset(mm, vaddr & HPAGE_MASK);
 
-                       /* hugetlb should be locked, and hence, prefaulted */
-                       WARN_ON(!pte || pte_none(*pte));
+                       /* the hugetlb file might have been truncated */
+                       if (!pte || pte_none(*pte)) {
+                               remainder = 0;
+                               if (!i)
+                                       i = -EFAULT;
+                               break;
+                       }
 
                        page = &pte_page(*pte)[vpfn % (HPAGE_SIZE/PAGE_SIZE)];
 
@@ -434,7 +441,7 @@ int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
                --remainder;
                ++i;
        }
-
+       spin_unlock(&mm->page_table_lock);
        *length = remainder;
        *position = vaddr;