jbd2: Fix oops in jbd2_journal_remove_journal_head()

jbd2_journal_remove_journal_head() can oops when trying to access
journal_head returned by bh2jh(). This is caused for example by the
following race:

	TASK1					TASK2
  jbd2_journal_commit_transaction()
    ...
    processing t_forget list
      __jbd2_journal_refile_buffer(jh);
      if (!jh->b_transaction) {
        jbd_unlock_bh_state(bh);
					jbd2_journal_try_to_free_buffers()
					  jbd2_journal_grab_journal_head(bh)
					  jbd_lock_bh_state(bh)
					  __journal_try_to_free_buffer()
					  jbd2_journal_put_journal_head(jh)
        jbd2_journal_remove_journal_head(bh);

jbd2_journal_put_journal_head() in TASK2 sees that b_jcount == 0 and
buffer is not part of any transaction and thus frees journal_head
before TASK1 gets to doing so. Note that even buffer_head can be
released by try_to_free_buffers() after
jbd2_journal_put_journal_head() which adds even larger opportunity for
oops (but I didn't see this happen in reality).

Fix the problem by making transactions hold their own journal_head
reference (in b_jcount). That way we don't have to remove journal_head
explicitely via jbd2_journal_remove_journal_head() and instead just
remove journal_head when b_jcount drops to zero. The result of this is
that [__]jbd2_journal_refile_buffer(),
[__]jbd2_journal_unfile_buffer(), and
__jdb2_journal_remove_checkpoint() can free journal_head which needs
modification of a few callers. Also we have to be careful because once
journal_head is removed, buffer_head might be freed as well. So we
have to get our own buffer_head reference where it matters.

Signed-off-by: Jan Kara <jack@suse.cz>
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>

1 files changed, 35 insertions, 32 deletions

diff --git a/fs/jbd2/transaction.c b/fs/jbd2/transaction.c
index 547b101049e..2d7109414cd 100644
--- a/fs/jbd2/transaction.c
+++ b/fs/jbd2/transaction.c
@@ -30,6 +30,7 @@
 #include <linux/module.h>
 
 static void __jbd2_journal_temp_unlink_buffer(struct journal_head *jh);
+static void __jbd2_journal_unfile_buffer(struct journal_head *jh);
 
 /*
  * jbd2_get_transaction: obtain a new transaction_t object.
@@ -764,7 +765,6 @@ repeat:
 	if (!jh->b_transaction) {
 		JBUFFER_TRACE(jh, "no transaction");
 		J_ASSERT_JH(jh, !jh->b_next_transaction);
-		jh->b_transaction = transaction;
 		JBUFFER_TRACE(jh, "file as BJ_Reserved");
 		spin_lock(&journal->j_list_lock);
 		__jbd2_journal_file_buffer(jh, transaction, BJ_Reserved);
@@ -895,8 +895,6 @@ int jbd2_journal_get_create_access(handle_t *handle, struct buffer_head *bh)
 		 * committed and so it's safe to clear the dirty bit.
 		 */
 		clear_buffer_dirty(jh2bh(jh));
-		jh->b_transaction = transaction;
-
 		/* first access by this transaction */
 		jh->b_modified = 0;
 
@@ -1230,8 +1228,6 @@ int jbd2_journal_forget (handle_t *handle, struct buffer_head *bh)
 			__jbd2_journal_file_buffer(jh, transaction, BJ_Forget);
 		} else {
 			__jbd2_journal_unfile_buffer(jh);
-			jbd2_journal_remove_journal_head(bh);
-			__brelse(bh);
 			if (!buffer_jbd(bh)) {
 				spin_unlock(&journal->j_list_lock);
 				jbd_unlock_bh_state(bh);
@@ -1554,19 +1550,32 @@ void __jbd2_journal_temp_unlink_buffer(struct journal_head *jh)
 		mark_buffer_dirty(bh);	/* Expose it to the VM */
 }
 
-void __jbd2_journal_unfile_buffer(struct journal_head *jh)
+/*
+ * Remove buffer from all transactions.
+ *
+ * Called with bh_state lock and j_list_lock
+ *
+ * jh and bh may be already freed when this function returns.
+ */
+static void __jbd2_journal_unfile_buffer(struct journal_head *jh)
 {
 	__jbd2_journal_temp_unlink_buffer(jh);
 	jh->b_transaction = NULL;
+	jbd2_journal_put_journal_head(jh);
 }
 
 void jbd2_journal_unfile_buffer(journal_t *journal, struct journal_head *jh)
 {
-	jbd_lock_bh_state(jh2bh(jh));
+	struct buffer_head *bh = jh2bh(jh);
+
+	/* Get reference so that buffer cannot be freed before we unlock it */
+	get_bh(bh);
+	jbd_lock_bh_state(bh);
 	spin_lock(&journal->j_list_lock);
 	__jbd2_journal_unfile_buffer(jh);
 	spin_unlock(&journal->j_list_lock);
-	jbd_unlock_bh_state(jh2bh(jh));
+	jbd_unlock_bh_state(bh);
+	__brelse(bh);
 }
 
 /*
@@ -1593,8 +1602,6 @@ __journal_try_to_free_buffer(journal_t *journal, struct buffer_head *bh)
 		if (jh->b_jlist == BJ_None) {
 			JBUFFER_TRACE(jh, "remove from checkpoint list");
 			__jbd2_journal_remove_checkpoint(jh);
-			jbd2_journal_remove_journal_head(bh);
-			__brelse(bh);
 		}
 	}
 	spin_unlock(&journal->j_list_lock);
@@ -1657,7 +1664,6 @@ int jbd2_journal_try_to_free_buffers(journal_t *journal,
 		/*
 		 * We take our own ref against the journal_head here to avoid
 		 * having to add tons of locking around each instance of
-		 * jbd2_journal_remove_journal_head() and
 		 * jbd2_journal_put_journal_head().
 		 */
 		jh = jbd2_journal_grab_journal_head(bh);
@@ -1695,10 +1701,9 @@ static int __dispose_buffer(struct journal_head *jh, transaction_t *transaction)
 	int may_free = 1;
 	struct buffer_head *bh = jh2bh(jh);
 
-	__jbd2_journal_unfile_buffer(jh);
-
 	if (jh->b_cp_transaction) {
 		JBUFFER_TRACE(jh, "on running+cp transaction");
+		__jbd2_journal_temp_unlink_buffer(jh);
 		/*
 		 * We don't want to write the buffer anymore, clear the
 		 * bit so that we don't confuse checks in
@@ -1709,8 +1714,7 @@ static int __dispose_buffer(struct journal_head *jh, transaction_t *transaction)
 		may_free = 0;
 	} else {
 		JBUFFER_TRACE(jh, "on running transaction");
-		jbd2_journal_remove_journal_head(bh);
-		__brelse(bh);
+		__jbd2_journal_unfile_buffer(jh);
 	}
 	return may_free;
 }
@@ -1988,6 +1992,8 @@ void __jbd2_journal_file_buffer(struct journal_head *jh,
 
 	if (jh->b_transaction)
 		__jbd2_journal_temp_unlink_buffer(jh);
+	else
+		jbd2_journal_grab_journal_head(bh);
 	jh->b_transaction = transaction;
 
 	switch (jlist) {
@@ -2039,9 +2045,10 @@ void jbd2_journal_file_buffer(struct journal_head *jh,
  * already started to be used by a subsequent transaction, refile the
  * buffer on that transaction's metadata list.
  *
- * Called under journal->j_list_lock
- *
+ * Called under j_list_lock
  * Called under jbd_lock_bh_state(jh2bh(jh))
+ *
+ * jh and bh may be already free when this function returns
  */
 void __jbd2_journal_refile_buffer(struct journal_head *jh)
 {
@@ -2065,6 +2072,11 @@ void __jbd2_journal_refile_buffer(struct journal_head *jh)
 
 	was_dirty = test_clear_buffer_jbddirty(bh);
 	__jbd2_journal_temp_unlink_buffer(jh);
+	/*
+	 * We set b_transaction here because b_next_transaction will inherit
+	 * our jh reference and thus __jbd2_journal_file_buffer() must not
+	 * take a new one.
+	 */
 	jh->b_transaction = jh->b_next_transaction;
 	jh->b_next_transaction = NULL;
 	if (buffer_freed(bh))
@@ -2081,30 +2093,21 @@ void __jbd2_journal_refile_buffer(struct journal_head *jh)
 }
 
 /*
- * For the unlocked version of this call, also make sure that any
- * hanging journal_head is cleaned up if necessary.
- *
- * __jbd2_journal_refile_buffer is usually called as part of a single locked
- * operation on a buffer_head, in which the caller is probably going to
- * be hooking the journal_head onto other lists.  In that case it is up
- * to the caller to remove the journal_head if necessary.  For the
- * unlocked jbd2_journal_refile_buffer call, the caller isn't going to be
- * doing anything else to the buffer so we need to do the cleanup
- * ourselves to avoid a jh leak.
- *
- * *** The journal_head may be freed by this call! ***
+ * __jbd2_journal_refile_buffer() with necessary locking added. We take our
+ * bh reference so that we can safely unlock bh.
+ *
+ * The jh and bh may be freed by this call.
  */
 void jbd2_journal_refile_buffer(journal_t *journal, struct journal_head *jh)
 {
 	struct buffer_head *bh = jh2bh(jh);
 
+	/* Get reference so that buffer cannot be freed before we unlock it */
+	get_bh(bh);
 	jbd_lock_bh_state(bh);
 	spin_lock(&journal->j_list_lock);
-
 	__jbd2_journal_refile_buffer(jh);
 	jbd_unlock_bh_state(bh);
-	jbd2_journal_remove_journal_head(bh);
-
 	spin_unlock(&journal->j_list_lock);
 	__brelse(bh);
 }