Merge ssh://linux-ntfs@bkbits.net/ntfs-2.6-devel

into cantab.net:/home/src/ntfs-2.6-devel

fs/ntfs/ChangeLog

@@ -118,6 +118,16 @@ ToDo/Notes:
 	  inode semaphore around the code thst sets ni->itype.index.bmp_ino to
 	  NULL and reorganize the code to optimize it a bit.  (Thanks to
 	  Christoph Hellwig for spotting this.)
+	- Modify fs/ntfs/aops.c::mark_ntfs_record_dirty() to no longer take the
+	  ntfs inode as a parameter as this is confusing and misleading and the
+	  needed ntfs inode is available via NTFS_I(page->mapping->host).
+	  Adapt all callers to this change.
+	- Modify fs/ntfs/mft.c::write_mft_record_nolock() and
+	  fs/ntfs/aops.c::ntfs_write_mst_block() to only check the dirty state
+	  of the first buffer in a record and to take this as the ntfs record
+	  dirty state.  We cannot look at the dirty state for subsequent
+	  buffers because we might be racing with
+	  fs/ntfs/aops.c::mark_ntfs_record_dirty().
 
 2.1.20 - Fix two stupid bugs introduced in 2.1.18 release.
 

fs/ntfs/aops.c

@@ -868,23 +868,24 @@ static int ntfs_write_mst_block(struct writeback_control *wbc,
 			clear_buffer_dirty(bh);
 			continue;
 		}
-		if (rec_block == block) {
+		if (likely(block < rec_block)) {
+			/*
+			 * This block is not the first one in the record.  We
+			 * ignore the buffer's dirty state because we could
+			 * have raced with a parallel mark_ntfs_record_dirty().
+			 */
+			if (!rec_is_dirty)
+				continue;
+		} else /* if (block == rec_block) */ {
+			BUG_ON(block > rec_block);
 			/* This block is the first one in the record. */
 			rec_block += bhs_per_rec;
 			if (!buffer_dirty(bh)) {
-				/* Clean buffers are not written out. */
+				/* Clean records are not written out. */
 				rec_is_dirty = FALSE;
 				continue;
 			}
 			rec_is_dirty = TRUE;
-		} else {
-			/* This block is not the first one in the record. */
-			if (!buffer_dirty(bh)) {
-				/* Clean buffers are not written out. */
-				BUG_ON(rec_is_dirty);
-				continue;
-			}
-			BUG_ON(!rec_is_dirty);
 		}
 		BUG_ON(!buffer_mapped(bh));
 		BUG_ON(!buffer_uptodate(bh));
@@ -973,10 +974,8 @@ static int ntfs_write_mst_block(struct writeback_control *wbc,
 			continue;
 		if (unlikely(test_set_buffer_locked(tbh)))
 			BUG();
-		if (unlikely(!test_clear_buffer_dirty(tbh))) {
-			unlock_buffer(tbh);
-			continue;
-		}
+		/* The buffer dirty state is now irrelevant, just clean it. */
+		clear_buffer_dirty(tbh);
 		BUG_ON(!buffer_uptodate(tbh));
 		BUG_ON(!buffer_mapped(tbh));
 		get_bh(tbh);
@@ -2132,9 +2131,8 @@ struct address_space_operations ntfs_mst_aops = {
 
 /**
  * mark_ntfs_record_dirty - mark an ntfs record dirty
- * @ni:		ntfs inode containing the ntfs record to be marked dirty
  * @page:	page containing the ntfs record to mark dirty
- * @rec_start:	byte offset within @page at which the ntfs record begins
+ * @ofs:	byte offset within @page at which the ntfs record begins
  *
  * If the ntfs record is the same size as the page cache page @page, set all
  * buffers in the page dirty.  Otherwise, set only the buffers in which the
@@ -2143,26 +2141,29 @@ struct address_space_operations ntfs_mst_aops = {
  * Also, set the page containing the ntfs record dirty, which also marks the
  * vfs inode the ntfs record belongs to dirty (I_DIRTY_PAGES).
  */
-void mark_ntfs_record_dirty(ntfs_inode *ni, struct page *page,
-		unsigned int rec_start) {
+void mark_ntfs_record_dirty(struct page *page, const unsigned int ofs) {
+	ntfs_inode *ni;
 	struct buffer_head *bh, *head;
-	unsigned int rec_end, bh_size, bh_start, bh_end;
+	unsigned int end, bh_size, bh_ofs;
 
 	BUG_ON(!page);
 	BUG_ON(!page_has_buffers(page));
+	ni = NTFS_I(page->mapping->host);
+	BUG_ON(!ni);
 	if (ni->itype.index.block_size == PAGE_CACHE_SIZE) {
 		__set_page_dirty_buffers(page);
 		return;
 	}
-	rec_end = rec_start + ni->itype.index.block_size;
+	end = ofs + ni->itype.index.block_size;
 	bh_size = ni->vol->sb->s_blocksize;
-	bh_start = 0;
 	bh = head = page_buffers(page);
 	do {
-		bh_end = bh_start + bh_size;
-		if ((bh_start >= rec_start) && (bh_end <= rec_end))
-			set_buffer_dirty(bh);
-		bh_start = bh_end;
+		bh_ofs = bh_offset(bh);
+		if (bh_ofs + bh_size <= ofs)
+			continue;
+		if (unlikely(bh_ofs >= end))
+			break;
+		set_buffer_dirty(bh);
 	} while ((bh = bh->b_this_page) != head);
 	__set_page_dirty_nobuffers(page);
 }

fs/ntfs/aops.h

@@ -95,8 +95,7 @@ static inline struct page *ntfs_map_page(struct address_space *mapping,
 
 #ifdef NTFS_RW
 
-extern void mark_ntfs_record_dirty(ntfs_inode *ni, struct page *page,
-		unsigned int rec_start);
+extern void mark_ntfs_record_dirty(struct page *page, const unsigned int ofs);
 
 #endif /* NTFS_RW */
 

fs/ntfs/index.h

@@ -139,8 +139,8 @@ static inline void ntfs_index_entry_mark_dirty(ntfs_index_context *ictx)
 	if (ictx->is_in_root)
 		mark_mft_record_dirty(ictx->actx->ntfs_ino);
 	else
-		mark_ntfs_record_dirty(ictx->idx_ni, ictx->page,
-			(u8*)ictx->ia - (u8*)page_address(ictx->page));
+		mark_ntfs_record_dirty(ictx->page,
+				(u8*)ictx->ia - (u8*)page_address(ictx->page));
 }
 
 #endif /* NTFS_RW */

fs/ntfs/inode.c

@@ -2513,8 +2513,8 @@ int ntfs_write_inode(struct inode *vi, int sync)
 	 * this function returns.
 	 */
 	if (modified && !NInoTestSetDirty(ctx->ntfs_ino))
-		mark_ntfs_record_dirty(NTFS_I(ni->vol->mft_ino),
-				ctx->ntfs_ino->page, ctx->ntfs_ino->page_ofs);
+		mark_ntfs_record_dirty(ctx->ntfs_ino->page,
+				ctx->ntfs_ino->page_ofs);
 	ntfs_attr_put_search_ctx(ctx);
 	/* Now the access times are updated, write the base mft record. */
 	if (NInoDirty(ni))

fs/ntfs/mft.c

@@ -380,8 +380,7 @@ void __mark_mft_record_dirty(ntfs_inode *ni)
 
 	ntfs_debug("Entering for inode 0x%lx.", ni->mft_no);
 	BUG_ON(NInoAttr(ni));
-	mark_ntfs_record_dirty(NTFS_I(ni->vol->mft_ino), ni->page,
-			ni->page_ofs);
+	mark_ntfs_record_dirty(ni->page, ni->page_ofs);
 	/* Determine the base vfs inode and mark it dirty, too. */
 	down(&ni->extent_lock);
 	if (likely(ni->nr_extents >= 0))
@@ -573,8 +572,10 @@ int ntfs_sync_mft_mirror(ntfs_volume *vol, const unsigned long mft_no,
  * ntfs inode @ni to backing store.  If the mft record @m has a counterpart in
  * the mft mirror, that is also updated.
  *
- * We only write the mft record if the ntfs inode @ni is dirty and the buffers
- * belonging to its mft record are dirty, too.
+ * We only write the mft record if the ntfs inode @ni is dirty and the first
+ * buffer belonging to its mft record is dirty, too.  We ignore the dirty state
+ * of subsequent buffers because we could have raced with
+ * fs/ntfs/aops.c::mark_ntfs_record_dirty().
  *
  * On success, clean the mft record and return 0.  On error, leave the mft
  * record dirty and return -errno.  The caller should call make_bad_inode() on
@@ -633,21 +634,23 @@ int write_mft_record_nolock(ntfs_inode *ni, MFT_RECORD *m, int sync)
 			continue;
 		if (unlikely(block_start >= m_end))
 			break;
-		/*
-		 * If the buffer is clean and it is the first buffer of the mft
-		 * record, it was written out by other means already so we are
-		 * done.  For safety we make sure all the other buffers are
-		 * clean also.  If it is clean but not the first buffer and the
-		 * first buffer was dirty it is a bug.
-		 */
-		if (!buffer_dirty(bh)) {
-			if (block_start == m_start)
+		if (block_start == m_start) {
+			/* This block is the first one in the record. */
+			if (!buffer_dirty(bh)) {
+				/* Clean records are not written out. */
 				rec_is_dirty = FALSE;
-			else
-				BUG_ON(rec_is_dirty);
-			continue;
+				continue;
+			}
+			rec_is_dirty = TRUE;
+		} else {
+			/*
+			 * This block is not the first one in the record.  We
+			 * ignore the buffer's dirty state because we could
+			 * have raced with a parallel mark_ntfs_record_dirty().
+			 */
+			if (!rec_is_dirty)
+				continue;
 		}
-		BUG_ON(!rec_is_dirty);
 		BUG_ON(!buffer_mapped(bh));
 		BUG_ON(!buffer_uptodate(bh));
 		BUG_ON(!nr_bhs && (m_start != block_start));