ocfs2: Re-journal buffers after transaction extend

ocfs2_extend_trans() might call journal_restart() which will commit dirty
buffers and then restart the transaction. This means that any buffers which
still need changes should be passed to journal_access() again. Some paths
during extend weren't doing this right.

Signed-off-by: Mark Fasheh <mark.fasheh@oracle.com>

diff --git a/fs/ocfs2/alloc.c b/fs/ocfs2/alloc.c
index 97f0db5..23c8cda 100644 (file)
--- a/fs/ocfs2/alloc.c
+++ b/fs/ocfs2/alloc.c
@@ -2389,6 +2389,18 @@ static int __ocfs2_rotate_tree_left(struct inode *inode,
                        goto out;
                }
 
+               /*
+                * Caller might still want to make changes to the
+                * tree root, so re-add it to the journal here.
+                */
+               ret = ocfs2_journal_access(handle, inode,
+                                          path_root_bh(left_path),
+                                          OCFS2_JOURNAL_ACCESS_WRITE);
+               if (ret) {
+                       mlog_errno(ret);
+                       goto out;
+               }
+
                ret = ocfs2_rotate_subtree_left(inode, handle, left_path,
                                                right_path, subtree_root,
                                                dealloc, &deleted);
@@ -3289,16 +3301,6 @@ static int ocfs2_insert_path(struct inode *inode,
        int ret, subtree_index;
        struct buffer_head *leaf_bh = path_leaf_bh(right_path);
 
-       /*
-        * Pass both paths to the journal. The majority of inserts
-        * will be touching all components anyway.
-        */
-       ret = ocfs2_journal_access_path(inode, handle, right_path);
-       if (ret < 0) {
-               mlog_errno(ret);
-               goto out;
-       }
-
        if (left_path) {
                int credits = handle->h_buffer_credits;
 
@@ -3323,6 +3325,16 @@ static int ocfs2_insert_path(struct inode *inode,
                }
        }
 
+       /*
+        * Pass both paths to the journal. The majority of inserts
+        * will be touching all components anyway.
+        */
+       ret = ocfs2_journal_access_path(inode, handle, right_path);
+       if (ret < 0) {
+               mlog_errno(ret);
+               goto out;
+       }
+
        if (insert->ins_split != SPLIT_NONE) {
                /*
                 * We could call ocfs2_insert_at_leaf() for some types
@@ -3331,6 +3343,17 @@ static int ocfs2_insert_path(struct inode *inode,
                 */
                ocfs2_split_record(inode, left_path, right_path,
                                   insert_rec, insert->ins_split);
+
+               /*
+                * Split might have modified either leaf and we don't
+                * have a guarantee that the later edge insert will
+                * dirty this for us.
+                */
+               if (left_path)
+                       ret = ocfs2_journal_dirty(handle,
+                                                 path_leaf_bh(left_path));
+                       if (ret)
+                               mlog_errno(ret);
        } else
                ocfs2_insert_at_leaf(insert_rec, path_leaf_el(right_path),
                                     insert, inode);
@@ -3430,6 +3453,17 @@ static int ocfs2_do_insert_extent(struct inode *inode,
                        mlog_errno(ret);
                        goto out;
                }
+
+               /*
+                * ocfs2_rotate_tree_right() might have extended the
+                * transaction without re-journaling our tree root.
+                */
+               ret = ocfs2_journal_access(handle, inode, di_bh,
+                                          OCFS2_JOURNAL_ACCESS_WRITE);
+               if (ret) {
+                       mlog_errno(ret);
+                       goto out;
+               }
        } else if (type->ins_appending == APPEND_TAIL
                   && type->ins_contig != CONTIG_LEFT) {
                ret = ocfs2_append_rec_to_path(inode, handle, insert_rec,
@@ -3941,7 +3975,7 @@ static int __ocfs2_mark_extent_written(struct inode *inode,
 {
        int ret = 0;
        struct ocfs2_extent_list *el = path_leaf_el(path);
-       struct buffer_head *eb_bh, *last_eb_bh = NULL;
+       struct buffer_head *last_eb_bh = NULL;
        struct ocfs2_extent_rec *rec = &el->l_recs[split_index];
        struct ocfs2_merge_ctxt ctxt;
        struct ocfs2_extent_list *rightmost_el;
@@ -3960,14 +3994,6 @@ static int __ocfs2_mark_extent_written(struct inode *inode,
                goto out;
        }
 
-       eb_bh = path_leaf_bh(path);
-       ret = ocfs2_journal_access(handle, inode, eb_bh,
-                                  OCFS2_JOURNAL_ACCESS_WRITE);
-       if (ret) {
-               mlog_errno(ret);
-               goto out;
-       }
-
        ctxt.c_contig_type = ocfs2_figure_merge_contig_type(inode, el,
                                                            split_index,
                                                            split_rec);
@@ -4029,8 +4055,6 @@ static int __ocfs2_mark_extent_written(struct inode *inode,
                        mlog_errno(ret);
        }
 
-       ocfs2_journal_dirty(handle, eb_bh);
-
 out:
        brelse(last_eb_bh);
        return ret;

diff --git a/fs/ocfs2/journal.c b/fs/ocfs2/journal.c
index 0e1250c..8d81f6c 100644 (file)
--- a/fs/ocfs2/journal.c
+++ b/fs/ocfs2/journal.c
@@ -174,6 +174,12 @@ int ocfs2_commit_trans(struct ocfs2_super *osb,
  * transaction. extend_trans will either extend the current handle by
  * nblocks, or commit it and start a new one with nblocks credits.
  *
+ * This might call journal_restart() which will commit dirty buffers
+ * and then restart the transaction. Before calling
+ * ocfs2_extend_trans(), any changed blocks should have been
+ * dirtied. After calling it, all blocks which need to be changed must
+ * go through another set of journal_access/journal_dirty calls.
+ *
  * WARNING: This will not release any semaphores or disk locks taken
  * during the transaction, so make sure they were taken *before*
  * start_trans or we'll have ordering deadlocks.