do_mremap() untangling, part 3
[safe/jmp/linux-2.6] / mm / memory.c
index 9bdbd10..6ab19dd 100644 (file)
@@ -109,6 +109,7 @@ static int __init disable_randmaps(char *s)
 __setup("norandmaps", disable_randmaps);
 
 unsigned long zero_pfn __read_mostly;
+unsigned long highest_memmap_pfn __read_mostly;
 
 /*
  * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
@@ -296,7 +297,8 @@ void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
                unsigned long addr = vma->vm_start;
 
                /*
-                * Hide vma from rmap and vmtruncate before freeing pgtables
+                * Hide vma from rmap and truncate_pagecache before freeing
+                * pgtables
                 */
                anon_vma_unlink(vma);
                unlink_file_vma(vma);
@@ -639,6 +641,7 @@ static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
                pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
                unsigned long addr, unsigned long end)
 {
+       pte_t *orig_src_pte, *orig_dst_pte;
        pte_t *src_pte, *dst_pte;
        spinlock_t *src_ptl, *dst_ptl;
        int progress = 0;
@@ -652,6 +655,8 @@ again:
        src_pte = pte_offset_map_nested(src_pmd, addr);
        src_ptl = pte_lockptr(src_mm, src_pmd);
        spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
+       orig_src_pte = src_pte;
+       orig_dst_pte = dst_pte;
        arch_enter_lazy_mmu_mode();
 
        do {
@@ -675,9 +680,9 @@ again:
 
        arch_leave_lazy_mmu_mode();
        spin_unlock(src_ptl);
-       pte_unmap_nested(src_pte - 1);
+       pte_unmap_nested(orig_src_pte);
        add_mm_rss(dst_mm, rss[0], rss[1]);
-       pte_unmap_unlock(dst_pte - 1, dst_ptl);
+       pte_unmap_unlock(orig_dst_pte, dst_ptl);
        cond_resched();
        if (addr != end)
                goto again;
@@ -1324,7 +1329,8 @@ int __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
                                if (ret & VM_FAULT_ERROR) {
                                        if (ret & VM_FAULT_OOM)
                                                return i ? i : -ENOMEM;
-                                       else if (ret & VM_FAULT_SIGBUS)
+                                       if (ret &
+                                           (VM_FAULT_HWPOISON|VM_FAULT_SIGBUS))
                                                return i ? i : -EFAULT;
                                        BUG();
                                }
@@ -1817,10 +1823,10 @@ static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
        token = pmd_pgtable(*pmd);
 
        do {
-               err = fn(pte, token, addr, data);
+               err = fn(pte++, token, addr, data);
                if (err)
                        break;
-       } while (pte++, addr += PAGE_SIZE, addr != end);
+       } while (addr += PAGE_SIZE, addr != end);
 
        arch_leave_lazy_mmu_mode();
 
@@ -2406,7 +2412,7 @@ restart:
  * @mapping: the address space containing mmaps to be unmapped.
  * @holebegin: byte in first page to unmap, relative to the start of
  * the underlying file.  This will be rounded down to a PAGE_SIZE
- * boundary.  Note that this is different from vmtruncate(), which
+ * boundary.  Note that this is different from truncate_pagecache(), which
  * must keep the partial page.  In contrast, we must get rid of
  * partial pages.
  * @holelen: size of prospective hole in bytes.  This will be rounded
@@ -2457,63 +2463,6 @@ void unmap_mapping_range(struct address_space *mapping,
 }
 EXPORT_SYMBOL(unmap_mapping_range);
 
-/**
- * vmtruncate - unmap mappings "freed" by truncate() syscall
- * @inode: inode of the file used
- * @offset: file offset to start truncating
- *
- * NOTE! We have to be ready to update the memory sharing
- * between the file and the memory map for a potential last
- * incomplete page.  Ugly, but necessary.
- */
-int vmtruncate(struct inode * inode, loff_t offset)
-{
-       if (inode->i_size < offset) {
-               unsigned long limit;
-
-               limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
-               if (limit != RLIM_INFINITY && offset > limit)
-                       goto out_sig;
-               if (offset > inode->i_sb->s_maxbytes)
-                       goto out_big;
-               i_size_write(inode, offset);
-       } else {
-               struct address_space *mapping = inode->i_mapping;
-
-               /*
-                * truncation of in-use swapfiles is disallowed - it would
-                * cause subsequent swapout to scribble on the now-freed
-                * blocks.
-                */
-               if (IS_SWAPFILE(inode))
-                       return -ETXTBSY;
-               i_size_write(inode, offset);
-
-               /*
-                * unmap_mapping_range is called twice, first simply for
-                * efficiency so that truncate_inode_pages does fewer
-                * single-page unmaps.  However after this first call, and
-                * before truncate_inode_pages finishes, it is possible for
-                * private pages to be COWed, which remain after
-                * truncate_inode_pages finishes, hence the second
-                * unmap_mapping_range call must be made for correctness.
-                */
-               unmap_mapping_range(mapping, offset + PAGE_SIZE - 1, 0, 1);
-               truncate_inode_pages(mapping, offset);
-               unmap_mapping_range(mapping, offset + PAGE_SIZE - 1, 0, 1);
-       }
-
-       if (inode->i_op->truncate)
-               inode->i_op->truncate(inode);
-       return 0;
-
-out_sig:
-       send_sig(SIGXFSZ, current, 0);
-out_big:
-       return -EFBIG;
-}
-EXPORT_SYMBOL(vmtruncate);
-
 int vmtruncate_range(struct inode *inode, loff_t offset, loff_t end)
 {
        struct address_space *mapping = inode->i_mapping;
@@ -2558,8 +2507,15 @@ static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
                goto out;
 
        entry = pte_to_swp_entry(orig_pte);
-       if (is_migration_entry(entry)) {
-               migration_entry_wait(mm, pmd, address);
+       if (unlikely(non_swap_entry(entry))) {
+               if (is_migration_entry(entry)) {
+                       migration_entry_wait(mm, pmd, address);
+               } else if (is_hwpoison_entry(entry)) {
+                       ret = VM_FAULT_HWPOISON;
+               } else {
+                       print_bad_pte(vma, address, orig_pte, NULL);
+                       ret = VM_FAULT_OOM;
+               }
                goto out;
        }
        delayacct_set_flag(DELAYACCT_PF_SWAPIN);
@@ -2583,6 +2539,10 @@ static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
                /* Had to read the page from swap area: Major fault */
                ret = VM_FAULT_MAJOR;
                count_vm_event(PGMAJFAULT);
+       } else if (PageHWPoison(page)) {
+               ret = VM_FAULT_HWPOISON;
+               delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
+               goto out_release;
        }
 
        lock_page(page);
@@ -2654,6 +2614,7 @@ out_nomap:
        pte_unmap_unlock(page_table, ptl);
 out_page:
        unlock_page(page);
+out_release:
        page_cache_release(page);
        return ret;
 }
@@ -2759,6 +2720,12 @@ static int __do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
        if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
                return ret;
 
+       if (unlikely(PageHWPoison(vmf.page))) {
+               if (ret & VM_FAULT_LOCKED)
+                       unlock_page(vmf.page);
+               return VM_FAULT_HWPOISON;
+       }
+
        /*
         * For consistency in subsequent calls, make the faulted page always
         * locked.