jbd: Fix assertion failure in fs/jbd/checkpoint.c
authorJan Kara <jack@suse.cz>
Wed, 5 Dec 2007 07:45:27 +0000 (23:45 -0800)
committerLinus Torvalds <torvalds@woody.linux-foundation.org>
Wed, 5 Dec 2007 17:21:20 +0000 (09:21 -0800)
Before we start committing a transaction, we call
__journal_clean_checkpoint_list() to cleanup transaction's written-back
buffers.

If this call happens to remove all of them (and there were already some
buffers), __journal_remove_checkpoint() will decide to free the transaction
because it isn't (yet) a committing transaction and soon we fail some
assertion - the transaction really isn't ready to be freed :).

We change the check in __journal_remove_checkpoint() to free only a
transaction in T_FINISHED state.  The locking there is subtle though (as
everywhere in JBD ;().  We use j_list_lock to protect the check and a
subsequent call to __journal_drop_transaction() and do the same in the end
of journal_commit_transaction() which is the only place where a transaction
can get to T_FINISHED state.

Probably I'm too paranoid here and such locking is not really necessary -
checkpoint lists are processed only from log_do_checkpoint() where a
transaction must be already committed to be processed or from
__journal_clean_checkpoint_list() where kjournald itself calls it and thus
transaction cannot change state either.  Better be safe if something
changes in future...

Signed-off-by: Jan Kara <jack@suse.cz>
Cc: <linux-ext4@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
fs/jbd/checkpoint.c
fs/jbd/commit.c
include/linux/jbd.h

index 47552d4a6324effeb085c603c91dfcf945ed46f3..0f69c416eebc6adef7d05eeaf3283083adb9e700 100644 (file)
@@ -602,15 +602,15 @@ int __journal_remove_checkpoint(struct journal_head *jh)
 
        /*
         * There is one special case to worry about: if we have just pulled the
-        * buffer off a committing transaction's forget list, then even if the
-        * checkpoint list is empty, the transaction obviously cannot be
-        * dropped!
+        * buffer off a running or committing transaction's checkpoing list,
+        * then even if the checkpoint list is empty, the transaction obviously
+        * cannot be dropped!
         *
-        * The locking here around j_committing_transaction is a bit sleazy.
+        * The locking here around t_state is a bit sleazy.
         * See the comment at the end of journal_commit_transaction().
         */
-       if (transaction == journal->j_committing_transaction) {
-               JBUFFER_TRACE(jh, "belongs to committing transaction");
+       if (transaction->t_state != T_FINISHED) {
+               JBUFFER_TRACE(jh, "belongs to running/committing transaction");
                goto out;
        }
 
index 8f1f2aa5fb39b2df5b744c2f7e495145cf1b2df8..610264b99a8e8bb357ed46dc9a08735443ab2659 100644 (file)
@@ -858,10 +858,10 @@ restart_loop:
        }
        spin_unlock(&journal->j_list_lock);
        /*
-        * This is a bit sleazy.  We borrow j_list_lock to protect
-        * journal->j_committing_transaction in __journal_remove_checkpoint.
-        * Really, __journal_remove_checkpoint should be using j_state_lock but
-        * it's a bit hassle to hold that across __journal_remove_checkpoint
+        * This is a bit sleazy.  We use j_list_lock to protect transition
+        * of a transaction into T_FINISHED state and calling
+        * __journal_drop_transaction(). Otherwise we could race with
+        * other checkpointing code processing the transaction...
         */
        spin_lock(&journal->j_state_lock);
        spin_lock(&journal->j_list_lock);
index 16e7ed855a18d9a1bc7fafd58d70f01e0753ae19..d9ecd13393b0859ffa98034ec9cb77a3bd8f1912 100644 (file)
@@ -439,6 +439,8 @@ struct transaction_s
        /*
         * Transaction's current state
         * [no locking - only kjournald alters this]
+        * [j_list_lock] guards transition of a transaction into T_FINISHED
+        * state and subsequent call of __journal_drop_transaction()
         * FIXME: needs barriers
         * KLUDGE: [use j_state_lock]
         */