ext4: Rewrite ext4_page_mkwrite() to use generic helpers
Jan Kara [Fri, 24 Jun 2011 18:29:41 +0000 (14:29 -0400)]
Rewrite ext4_page_mkwrite() to use __block_page_mkwrite() helper. This
removes the need of using i_alloc_sem to avoid races with truncate which
seems to be the wrong locking order according to lock ordering documented in
mm/rmap.c. Also calling ext4_da_write_begin() as used by the old code seems to
be problematic because we can decide to flush delay-allocated blocks which
will acquire s_umount semaphore - again creating unpleasant lock dependency
if not directly a deadlock.

Also add a check for frozen filesystem so that we don't busyloop in page fault
when the filesystem is frozen.

Signed-off-by: Jan Kara <jack@suse.cz>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>

fs/ext4/inode.c

index e3126c0..bd30976 100644 (file)
@@ -5843,80 +5843,84 @@ int ext4_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
        struct page *page = vmf->page;
        loff_t size;
        unsigned long len;
-       int ret = -EINVAL;
-       void *fsdata;
+       int ret;
        struct file *file = vma->vm_file;
        struct inode *inode = file->f_path.dentry->d_inode;
        struct address_space *mapping = inode->i_mapping;
+       handle_t *handle;
+       get_block_t *get_block;
+       int retries = 0;
 
        /*
-        * Get i_alloc_sem to stop truncates messing with the inode. We cannot
-        * get i_mutex because we are already holding mmap_sem.
+        * This check is racy but catches the common case. We rely on
+        * __block_page_mkwrite() to do a reliable check.
         */
-       down_read(&inode->i_alloc_sem);
-       size = i_size_read(inode);
-       if (page->mapping != mapping || size <= page_offset(page)
-           || !PageUptodate(page)) {
-               /* page got truncated from under us? */
-               goto out_unlock;
+       vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
+       /* Delalloc case is easy... */
+       if (test_opt(inode->i_sb, DELALLOC) &&
+           !ext4_should_journal_data(inode) &&
+           !ext4_nonda_switch(inode->i_sb)) {
+               do {
+                       ret = __block_page_mkwrite(vma, vmf,
+                                                  ext4_da_get_block_prep);
+               } while (ret == -ENOSPC &&
+                      ext4_should_retry_alloc(inode->i_sb, &retries));
+               goto out_ret;
        }
-       ret = 0;
 
        lock_page(page);
-       wait_on_page_writeback(page);
-       if (PageMappedToDisk(page)) {
-               up_read(&inode->i_alloc_sem);
-               return VM_FAULT_LOCKED;
+       size = i_size_read(inode);
+       /* Page got truncated from under us? */
+       if (page->mapping != mapping || page_offset(page) > size) {
+               unlock_page(page);
+               ret = VM_FAULT_NOPAGE;
+               goto out;
        }
 
        if (page->index == size >> PAGE_CACHE_SHIFT)
                len = size & ~PAGE_CACHE_MASK;
        else
                len = PAGE_CACHE_SIZE;
-
        /*
-        * return if we have all the buffers mapped. This avoid
-        * the need to call write_begin/write_end which does a
-        * journal_start/journal_stop which can block and take
-        * long time
+        * Return if we have all the buffers mapped. This avoids the need to do
+        * journal_start/journal_stop which can block and take a long time
         */
        if (page_has_buffers(page)) {
                if (!walk_page_buffers(NULL, page_buffers(page), 0, len, NULL,
                                        ext4_bh_unmapped)) {
-                       up_read(&inode->i_alloc_sem);
-                       return VM_FAULT_LOCKED;
+                       /* Wait so that we don't change page under IO */
+                       wait_on_page_writeback(page);
+                       ret = VM_FAULT_LOCKED;
+                       goto out;
                }
        }
        unlock_page(page);
-       /*
-        * OK, we need to fill the hole... Do write_begin write_end
-        * to do block allocation/reservation.We are not holding
-        * inode.i__mutex here. That allow * parallel write_begin,
-        * write_end call. lock_page prevent this from happening
-        * on the same page though
-        */
-       ret = mapping->a_ops->write_begin(file, mapping, page_offset(page),
-                       len, AOP_FLAG_UNINTERRUPTIBLE, &page, &fsdata);
-       if (ret < 0)
-               goto out_unlock;
-       ret = mapping->a_ops->write_end(file, mapping, page_offset(page),
-                       len, len, page, fsdata);
-       if (ret < 0)
-               goto out_unlock;
-       ret = 0;
-
-       /*
-        * write_begin/end might have created a dirty page and someone
-        * could wander in and start the IO.  Make sure that hasn't
-        * happened.
-        */
-       lock_page(page);
-       wait_on_page_writeback(page);
-       up_read(&inode->i_alloc_sem);
-       return VM_FAULT_LOCKED;
-out_unlock:
-       if (ret)
+       /* OK, we need to fill the hole... */
+       if (ext4_should_dioread_nolock(inode))
+               get_block = ext4_get_block_write;
+       else
+               get_block = ext4_get_block;
+retry_alloc:
+       handle = ext4_journal_start(inode, ext4_writepage_trans_blocks(inode));
+       if (IS_ERR(handle)) {
                ret = VM_FAULT_SIGBUS;
-       up_read(&inode->i_alloc_sem);
+               goto out;
+       }
+       ret = __block_page_mkwrite(vma, vmf, get_block);
+       if (!ret && ext4_should_journal_data(inode)) {
+               if (walk_page_buffers(handle, page_buffers(page), 0,
+                         PAGE_CACHE_SIZE, NULL, do_journal_get_write_access)) {
+                       unlock_page(page);
+                       ret = VM_FAULT_SIGBUS;
+                       goto out;
+               }
+               ext4_set_inode_state(inode, EXT4_STATE_JDATA);
+       }
+       ext4_journal_stop(handle);
+       if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
+               goto retry_alloc;
+out_ret:
+       ret = block_page_mkwrite_return(ret);
+out:
        return ret;
 }