1 files changed, 196 insertions, 184 deletions

diff --git a/fs/jbd/transaction.c b/fs/jbd/transaction.c
index ed886e6db39..1695ba8334a 100644
--- a/fs/jbd/transaction.c
+++ b/fs/jbd/transaction.c
@@ -56,7 +56,8 @@ get_transaction(journal_t *journal, transaction_t *transaction)
 	spin_lock_init(&transaction->t_handle_lock);
 
 	/* Set up the commit timer for the new transaction. */
-	journal->j_commit_timer.expires = round_jiffies(transaction->t_expires);
+	journal->j_commit_timer.expires =
+				round_jiffies_up(transaction->t_expires);
 	add_timer(&journal->j_commit_timer);
 
 	J_ASSERT(journal->j_running_transaction == NULL);
@@ -206,7 +207,7 @@ repeat_locked:
 	 * the committing transaction.  Really, we only need to give it
 	 * committing_transaction->t_outstanding_credits plus "enough" for
 	 * the log control blocks.
-	 * Also, this test is inconsitent with the matching one in
+	 * Also, this test is inconsistent with the matching one in
 	 * journal_extend().
 	 */
 	if (__log_space_left(journal) < jbd_space_needed(journal)) {
@@ -228,6 +229,8 @@ repeat_locked:
 		  __log_space_left(journal));
 	spin_unlock(&transaction->t_handle_lock);
 	spin_unlock(&journal->j_state_lock);
+
+	lock_map_acquire(&handle->h_lockdep_map);
 out:
 	if (unlikely(new_transaction))		/* It's usually NULL */
 		kfree(new_transaction);
@@ -242,7 +245,6 @@ static handle_t *new_handle(int nblocks)
 	handle_t *handle = jbd_alloc_handle(GFP_NOFS);
 	if (!handle)
 		return NULL;
-	memset(handle, 0, sizeof(*handle));
 	handle->h_buffer_credits = nblocks;
 	handle->h_ref = 1;
 
@@ -263,7 +265,8 @@ static handle_t *new_handle(int nblocks)
  * This function is visible to journal users (like ext3fs), so is not
  * called with the journal already locked.
  *
- * Return a pointer to a newly allocated handle, or NULL on failure
+ * Return a pointer to a newly allocated handle, or an ERR_PTR() value
+ * on failure.
  */
 handle_t *journal_start(journal_t *journal, int nblocks)
 {
@@ -290,12 +293,7 @@ handle_t *journal_start(journal_t *journal, int nblocks)
 		jbd_free_handle(handle);
 		current->journal_info = NULL;
 		handle = ERR_PTR(err);
-		goto out;
 	}
-
-	lock_map_acquire(&handle->h_lockdep_map);
-
-out:
 	return handle;
 }
 
@@ -416,6 +414,7 @@ int journal_restart(handle_t *handle, int nblocks)
 	__log_start_commit(journal, transaction->t_tid);
 	spin_unlock(&journal->j_state_lock);
 
+	lock_map_release(&handle->h_lockdep_map);
 	handle->h_buffer_credits = nblocks;
 	ret = start_this_handle(journal, handle);
 	return ret;
@@ -426,17 +425,34 @@ int journal_restart(handle_t *handle, int nblocks)
  * void journal_lock_updates () - establish a transaction barrier.
  * @journal:  Journal to establish a barrier on.
  *
- * This locks out any further updates from being started, and blocks
- * until all existing updates have completed, returning only once the
- * journal is in a quiescent state with no updates running.
+ * This locks out any further updates from being started, and blocks until all
+ * existing updates have completed, returning only once the journal is in a
+ * quiescent state with no updates running.
  *
- * The journal lock should not be held on entry.
+ * We do not use simple mutex for synchronization as there are syscalls which
+ * want to return with filesystem locked and that trips up lockdep. Also
+ * hibernate needs to lock filesystem but locked mutex then blocks hibernation.
+ * Since locking filesystem is rare operation, we use simple counter and
+ * waitqueue for locking.
  */
 void journal_lock_updates(journal_t *journal)
 {
 	DEFINE_WAIT(wait);
 
+wait:
+	/* Wait for previous locked operation to finish */
+	wait_event(journal->j_wait_transaction_locked,
+		   journal->j_barrier_count == 0);
+
 	spin_lock(&journal->j_state_lock);
+	/*
+	 * Check reliably under the lock whether we are the ones winning the race
+	 * and locking the journal
+	 */
+	if (journal->j_barrier_count > 0) {
+		spin_unlock(&journal->j_state_lock);
+		goto wait;
+	}
 	++journal->j_barrier_count;
 
 	/* Wait until there are no running updates */
@@ -460,14 +476,6 @@ void journal_lock_updates(journal_t *journal)
 		spin_lock(&journal->j_state_lock);
 	}
 	spin_unlock(&journal->j_state_lock);
-
-	/*
-	 * We have now established a barrier against other normal updates, but
-	 * we also need to barrier against other journal_lock_updates() calls
-	 * to make sure that we serialise special journal-locked operations
-	 * too.
-	 */
-	mutex_lock(&journal->j_barrier);
 }
 
 /**
@@ -475,48 +483,26 @@ void journal_lock_updates(journal_t *journal)
  * @journal:  Journal to release the barrier on.
  *
  * Release a transaction barrier obtained with journal_lock_updates().
- *
- * Should be called without the journal lock held.
  */
 void journal_unlock_updates (journal_t *journal)
 {
 	J_ASSERT(journal->j_barrier_count != 0);
 
-	mutex_unlock(&journal->j_barrier);
 	spin_lock(&journal->j_state_lock);
 	--journal->j_barrier_count;
 	spin_unlock(&journal->j_state_lock);
 	wake_up(&journal->j_wait_transaction_locked);
 }
 
-/*
- * Report any unexpected dirty buffers which turn up.  Normally those
- * indicate an error, but they can occur if the user is running (say)
- * tune2fs to modify the live filesystem, so we need the option of
- * continuing as gracefully as possible.  #
- *
- * The caller should already hold the journal lock and
- * j_list_lock spinlock: most callers will need those anyway
- * in order to probe the buffer's journaling state safely.
- */
-static void jbd_unexpected_dirty_buffer(struct journal_head *jh)
+static void warn_dirty_buffer(struct buffer_head *bh)
 {
-	int jlist;
-
-	/* If this buffer is one which might reasonably be dirty
-	 * --- ie. data, or not part of this journal --- then
-	 * we're OK to leave it alone, but otherwise we need to
-	 * move the dirty bit to the journal's own internal
-	 * JBDDirty bit. */
-	jlist = jh->b_jlist;
-
-	if (jlist == BJ_Metadata || jlist == BJ_Reserved ||
-	    jlist == BJ_Shadow || jlist == BJ_Forget) {
-		struct buffer_head *bh = jh2bh(jh);
+	char b[BDEVNAME_SIZE];
 
-		if (test_clear_buffer_dirty(bh))
-			set_buffer_jbddirty(bh);
-	}
+	printk(KERN_WARNING
+	       "JBD: Spotted dirty metadata buffer (dev = %s, blocknr = %llu). "
+	       "There's a risk of filesystem corruption in case of system "
+	       "crash.\n",
+	       bdevname(bh->b_bdev, b), (unsigned long long)bh->b_blocknr);
 }
 
 /*
@@ -546,7 +532,7 @@ do_get_write_access(handle_t *handle, struct journal_head *jh,
 	transaction = handle->h_transaction;
 	journal = transaction->t_journal;
 
-	jbd_debug(5, "buffer_head %p, force_copy %d\n", jh, force_copy);
+	jbd_debug(5, "journal_head %p, force_copy %d\n", jh, force_copy);
 
 	JBUFFER_TRACE(jh, "entry");
 repeat:
@@ -583,14 +569,16 @@ repeat:
 			if (jh->b_next_transaction)
 				J_ASSERT_JH(jh, jh->b_next_transaction ==
 							transaction);
+			warn_dirty_buffer(bh);
 		}
 		/*
 		 * In any case we need to clean the dirty flag and we must
 		 * do it under the buffer lock to be sure we don't race
 		 * with running write-out.
 		 */
-		JBUFFER_TRACE(jh, "Unexpected dirty buffer");
-		jbd_unexpected_dirty_buffer(jh);
+		JBUFFER_TRACE(jh, "Journalling dirty buffer");
+		clear_buffer_dirty(bh);
+		set_buffer_jbddirty(bh);
 	}
 
 	unlock_buffer(bh);
@@ -687,7 +675,7 @@ repeat:
 					jbd_alloc(jh2bh(jh)->b_size,
 							 GFP_NOFS);
 				if (!frozen_buffer) {
-					printk(KERN_EMERG
+					printk(KERN_ERR
 					       "%s: OOM for frozen_buffer\n",
 					       __func__);
 					JBUFFER_TRACE(jh, "oom!");
@@ -713,7 +701,6 @@ 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);
 		__journal_file_buffer(jh, transaction, BJ_Reserved);
@@ -729,10 +716,10 @@ done:
 		J_EXPECT_JH(jh, buffer_uptodate(jh2bh(jh)),
 			    "Possible IO failure.\n");
 		page = jh2bh(jh)->b_page;
-		offset = ((unsigned long) jh2bh(jh)->b_data) & ~PAGE_MASK;
-		source = kmap_atomic(page, KM_USER0);
+		offset = offset_in_page(jh2bh(jh)->b_data);
+		source = kmap_atomic(page);
 		memcpy(jh->b_frozen_data, source+offset, jh2bh(jh)->b_size);
-		kunmap_atomic(source, KM_USER0);
+		kunmap_atomic(source);
 	}
 	jbd_unlock_bh_state(bh);
 
@@ -826,7 +813,15 @@ int journal_get_create_access(handle_t *handle, struct buffer_head *bh)
 	J_ASSERT_JH(jh, buffer_locked(jh2bh(jh)));
 
 	if (jh->b_transaction == NULL) {
-		jh->b_transaction = transaction;
+		/*
+		 * Previous journal_forget() could have left the buffer
+		 * with jbddirty bit set because it was being committed. When
+		 * the commit finished, we've filed the buffer for
+		 * checkpointing and marked it dirty. Now we are reallocating
+		 * the buffer so the transaction freeing it must have
+		 * committed and so it's safe to clear the dirty bit.
+		 */
+		clear_buffer_dirty(jh2bh(jh));
 
 		/* first access by this transaction */
 		jh->b_modified = 0;
@@ -852,8 +847,8 @@ int journal_get_create_access(handle_t *handle, struct buffer_head *bh)
 	 */
 	JBUFFER_TRACE(jh, "cancelling revoke");
 	journal_cancel_revoke(handle, jh);
-	journal_put_journal_head(jh);
 out:
+	journal_put_journal_head(jh);
 	return err;
 }
 
@@ -903,7 +898,7 @@ repeat:
 	if (!jh->b_committed_data) {
 		committed_data = jbd_alloc(jh2bh(jh)->b_size, GFP_NOFS);
 		if (!committed_data) {
-			printk(KERN_EMERG "%s: No memory for committed data\n",
+			printk(KERN_ERR "%s: No memory for committed data\n",
 				__func__);
 			err = -ENOMEM;
 			goto out;
@@ -1077,8 +1072,9 @@ int journal_dirty_data(handle_t *handle, struct buffer_head *bh)
 				ret = -EIO;
 				goto no_journal;
 			}
-
-			if (jh->b_transaction != NULL) {
+			/* We might have slept so buffer could be refiled now */
+			if (jh->b_transaction != NULL &&
+			    jh->b_transaction != handle->h_transaction) {
 				JBUFFER_TRACE(jh, "unfile from commit");
 				__journal_temp_unlink_buffer(jh);
 				/* It still points to the committing
@@ -1099,8 +1095,6 @@ int journal_dirty_data(handle_t *handle, struct buffer_head *bh)
 		if (jh->b_jlist != BJ_SyncData && jh->b_jlist != BJ_Locked) {
 			JBUFFER_TRACE(jh, "not on correct data list: unfile");
 			J_ASSERT_JH(jh, jh->b_jlist != BJ_Shadow);
-			__journal_temp_unlink_buffer(jh);
-			jh->b_transaction = handle->h_transaction;
 			JBUFFER_TRACE(jh, "file as data");
 			__journal_file_buffer(jh, handle->h_transaction,
 						BJ_SyncData);
@@ -1264,7 +1258,7 @@ int journal_forget (handle_t *handle, struct buffer_head *bh)
 		goto not_jbd;
 	}
 
-	/* keep track of wether or not this transaction modified us */
+	/* keep track of whether or not this transaction modified us */
 	was_modified = jh->b_modified;
 
 	/*
@@ -1308,8 +1302,6 @@ int journal_forget (handle_t *handle, struct buffer_head *bh)
 			__journal_file_buffer(jh, transaction, BJ_Forget);
 		} else {
 			__journal_unfile_buffer(jh);
-			journal_remove_journal_head(bh);
-			__brelse(bh);
 			if (!buffer_jbd(bh)) {
 				spin_unlock(&journal->j_list_lock);
 				jbd_unlock_bh_state(bh);
@@ -1401,11 +1393,11 @@ int journal_stop(handle_t *handle)
 	 * by 30x or more...
 	 *
 	 * We try and optimize the sleep time against what the underlying disk
-	 * can do, instead of having a static sleep time.  This is usefull for
+	 * can do, instead of having a static sleep time.  This is useful for
 	 * the case where our storage is so fast that it is more optimal to go
 	 * ahead and force a flush and wait for the transaction to be committed
 	 * than it is to wait for an arbitrary amount of time for new writers to
-	 * join the transaction.  We acheive this by measuring how long it takes
+	 * join the transaction.  We achieve this by measuring how long it takes
 	 * to commit a transaction, and compare it with how long this
 	 * transaction has been running, and if run time < commit time then we
 	 * sleep for the delta and commit.  This greatly helps super fast disks
@@ -1440,8 +1432,6 @@ int journal_stop(handle_t *handle)
 		}
 	}
 
-	if (handle->h_sync)
-		transaction->t_synchronous_commit = 1;
 	current->journal_info = NULL;
 	spin_lock(&journal->j_state_lock);
 	spin_lock(&transaction->t_handle_lock);
@@ -1630,19 +1620,32 @@ static void __journal_temp_unlink_buffer(struct journal_head *jh)
 		mark_buffer_dirty(bh);	/* Expose it to the VM */
 }
 
+/*
+ * 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.
+ */
 void __journal_unfile_buffer(struct journal_head *jh)
 {
 	__journal_temp_unlink_buffer(jh);
 	jh->b_transaction = NULL;
+	journal_put_journal_head(jh);
 }
 
 void 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);
 	__journal_unfile_buffer(jh);
 	spin_unlock(&journal->j_list_lock);
-	jbd_unlock_bh_state(jh2bh(jh));
+	jbd_unlock_bh_state(bh);
+	__brelse(bh);
 }
 
 /*
@@ -1669,16 +1672,12 @@ __journal_try_to_free_buffer(journal_t *journal, struct buffer_head *bh)
 			/* A written-back ordered data buffer */
 			JBUFFER_TRACE(jh, "release data");
 			__journal_unfile_buffer(jh);
-			journal_remove_journal_head(bh);
-			__brelse(bh);
 		}
 	} else if (jh->b_cp_transaction != NULL && jh->b_transaction == NULL) {
 		/* written-back checkpointed metadata buffer */
 		if (jh->b_jlist == BJ_None) {
 			JBUFFER_TRACE(jh, "remove from checkpoint list");
 			__journal_remove_checkpoint(jh);
-			journal_remove_journal_head(bh);
-			__brelse(bh);
 		}
 	}
 	spin_unlock(&journal->j_list_lock);
@@ -1686,35 +1685,6 @@ out:
 	return;
 }
 
-/*
- * journal_try_to_free_buffers() could race with journal_commit_transaction()
- * The latter might still hold the a count on buffers when inspecting
- * them on t_syncdata_list or t_locked_list.
- *
- * journal_try_to_free_buffers() will call this function to
- * wait for the current transaction to finish syncing data buffers, before
- * tryinf to free that buffer.
- *
- * Called with journal->j_state_lock held.
- */
-static void journal_wait_for_transaction_sync_data(journal_t *journal)
-{
-	transaction_t *transaction = NULL;
-	tid_t tid;
-
-	spin_lock(&journal->j_state_lock);
-	transaction = journal->j_committing_transaction;
-
-	if (!transaction) {
-		spin_unlock(&journal->j_state_lock);
-		return;
-	}
-
-	tid = transaction->t_tid;
-	spin_unlock(&journal->j_state_lock);
-	log_wait_commit(journal, tid);
-}
-
 /**
  * int journal_try_to_free_buffers() - try to free page buffers.
  * @journal: journal for operation
@@ -1770,7 +1740,7 @@ int 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
-		 * journal_remove_journal_head() and journal_put_journal_head().
+		 * journal_put_journal_head().
 		 */
 		jh = journal_grab_journal_head(bh);
 		if (!jh)
@@ -1786,25 +1756,6 @@ int journal_try_to_free_buffers(journal_t *journal,
 
 	ret = try_to_free_buffers(page);
 
-	/*
-	 * There are a number of places where journal_try_to_free_buffers()
-	 * could race with journal_commit_transaction(), the later still
-	 * holds the reference to the buffers to free while processing them.
-	 * try_to_free_buffers() failed to free those buffers. Some of the
-	 * caller of releasepage() request page buffers to be dropped, otherwise
-	 * treat the fail-to-free as errors (such as generic_file_direct_IO())
-	 *
-	 * So, if the caller of try_to_release_page() wants the synchronous
-	 * behaviour(i.e make sure buffers are dropped upon return),
-	 * let's wait for the current transaction to finish flush of
-	 * dirty data buffers, then try to free those buffers again,
-	 * with the journal locked.
-	 */
-	if (ret == 0 && (gfp_mask & __GFP_WAIT) && (gfp_mask & __GFP_FS)) {
-		journal_wait_for_transaction_sync_data(journal);
-		ret = try_to_free_buffers(page);
-	}
-
 busy:
 	return ret;
 }
@@ -1826,17 +1777,20 @@ static int __dispose_buffer(struct journal_head *jh, transaction_t *transaction)
 	int may_free = 1;
 	struct buffer_head *bh = jh2bh(jh);
 
-	__journal_unfile_buffer(jh);
-
 	if (jh->b_cp_transaction) {
 		JBUFFER_TRACE(jh, "on running+cp transaction");
+		__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
+		 * __journal_file_buffer
+		 */
+		clear_buffer_dirty(bh);
 		__journal_file_buffer(jh, transaction, BJ_Forget);
-		clear_buffer_jbddirty(bh);
 		may_free = 0;
 	} else {
 		JBUFFER_TRACE(jh, "on running transaction");
-		journal_remove_journal_head(bh);
-		__brelse(bh);
+		__journal_unfile_buffer(jh);
 	}
 	return may_free;
 }
@@ -1888,15 +1842,16 @@ static int __dispose_buffer(struct journal_head *jh, transaction_t *transaction)
  * We're outside-transaction here.  Either or both of j_running_transaction
  * and j_committing_transaction may be NULL.
  */
-static int journal_unmap_buffer(journal_t *journal, struct buffer_head *bh)
+static int journal_unmap_buffer(journal_t *journal, struct buffer_head *bh,
+				int partial_page)
 {
 	transaction_t *transaction;
 	struct journal_head *jh;
 	int may_free = 1;
-	int ret;
 
 	BUFFER_TRACE(bh, "entry");
 
+retry:
 	/*
 	 * It is safe to proceed here without the j_list_lock because the
 	 * buffers cannot be stolen by try_to_free_buffers as long as we are
@@ -1914,6 +1869,29 @@ static int journal_unmap_buffer(journal_t *journal, struct buffer_head *bh)
 	if (!jh)
 		goto zap_buffer_no_jh;
 
+	/*
+	 * We cannot remove the buffer from checkpoint lists until the
+	 * transaction adding inode to orphan list (let's call it T)
+	 * is committed.  Otherwise if the transaction changing the
+	 * buffer would be cleaned from the journal before T is
+	 * committed, a crash will cause that the correct contents of
+	 * the buffer will be lost.  On the other hand we have to
+	 * clear the buffer dirty bit at latest at the moment when the
+	 * transaction marking the buffer as freed in the filesystem
+	 * structures is committed because from that moment on the
+	 * block can be reallocated and used by a different page.
+	 * Since the block hasn't been freed yet but the inode has
+	 * already been added to orphan list, it is safe for us to add
+	 * the buffer to BJ_Forget list of the newest transaction.
+	 *
+	 * Also we have to clear buffer_mapped flag of a truncated buffer
+	 * because the buffer_head may be attached to the page straddling
+	 * i_size (can happen only when blocksize < pagesize) and thus the
+	 * buffer_head can be reused when the file is extended again. So we end
+	 * up keeping around invalidated buffers attached to transactions'
+	 * BJ_Forget list just to stop checkpointing code from cleaning up
+	 * the transaction this buffer was modified in.
+	 */
 	transaction = jh->b_transaction;
 	if (transaction == NULL) {
 		/* First case: not on any transaction.  If it
@@ -1939,13 +1917,9 @@ static int journal_unmap_buffer(journal_t *journal, struct buffer_head *bh)
 			 * committed, the buffer won't be needed any
 			 * longer. */
 			JBUFFER_TRACE(jh, "checkpointed: add to BJ_Forget");
-			ret = __dispose_buffer(jh,
+			may_free = __dispose_buffer(jh,
 					journal->j_running_transaction);
-			journal_put_journal_head(jh);
-			spin_unlock(&journal->j_list_lock);
-			jbd_unlock_bh_state(bh);
-			spin_unlock(&journal->j_state_lock);
-			return ret;
+			goto zap_buffer;
 		} else {
 			/* There is no currently-running transaction. So the
 			 * orphan record which we wrote for this file must have
@@ -1953,13 +1927,9 @@ static int journal_unmap_buffer(journal_t *journal, struct buffer_head *bh)
 			 * the committing transaction, if it exists. */
 			if (journal->j_committing_transaction) {
 				JBUFFER_TRACE(jh, "give to committing trans");
-				ret = __dispose_buffer(jh,
+				may_free = __dispose_buffer(jh,
 					journal->j_committing_transaction);
-				journal_put_journal_head(jh);
-				spin_unlock(&journal->j_list_lock);
-				jbd_unlock_bh_state(bh);
-				spin_unlock(&journal->j_state_lock);
-				return ret;
+				goto zap_buffer;
 			} else {
 				/* The orphan record's transaction has
 				 * committed.  We can cleanse this buffer */
@@ -1979,16 +1949,31 @@ static int journal_unmap_buffer(journal_t *journal, struct buffer_head *bh)
 			goto zap_buffer;
 		}
 		/*
-		 * If it is committing, we simply cannot touch it.  We
-		 * can remove it's next_transaction pointer from the
-		 * running transaction if that is set, but nothing
-		 * else. */
-		set_buffer_freed(bh);
-		if (jh->b_next_transaction) {
-			J_ASSERT(jh->b_next_transaction ==
-					journal->j_running_transaction);
-			jh->b_next_transaction = NULL;
+		 * The buffer is committing, we simply cannot touch
+		 * it. If the page is straddling i_size we have to wait
+		 * for commit and try again.
+		 */
+		if (partial_page) {
+			tid_t tid = journal->j_committing_transaction->t_tid;
+
+			journal_put_journal_head(jh);
+			spin_unlock(&journal->j_list_lock);
+			jbd_unlock_bh_state(bh);
+			spin_unlock(&journal->j_state_lock);
+			unlock_buffer(bh);
+			log_wait_commit(journal, tid);
+			lock_buffer(bh);
+			goto retry;
 		}
+		/*
+		 * OK, buffer won't be reachable after truncate. We just set
+		 * j_next_transaction to the running transaction (if there is
+		 * one) and mark buffer as freed so that commit code knows it
+		 * should clear dirty bits when it is done with the buffer.
+		 */
+		set_buffer_freed(bh);
+		if (journal->j_running_transaction && buffer_jbddirty(bh))
+			jh->b_next_transaction = journal->j_running_transaction;
 		journal_put_journal_head(jh);
 		spin_unlock(&journal->j_list_lock);
 		jbd_unlock_bh_state(bh);
@@ -2007,6 +1992,14 @@ static int journal_unmap_buffer(journal_t *journal, struct buffer_head *bh)
 	}
 
 zap_buffer:
+	/*
+	 * This is tricky. Although the buffer is truncated, it may be reused
+	 * if blocksize < pagesize and it is attached to the page straddling
+	 * EOF. Since the buffer might have been added to BJ_Forget list of the
+	 * running transaction, journal_get_write_access() won't clear
+	 * b_modified and credit accounting gets confused. So clear b_modified
+	 * here. */
+	jh->b_modified = 0;
 	journal_put_journal_head(jh);
 zap_buffer_no_jh:
 	spin_unlock(&journal->j_list_lock);
@@ -2026,16 +2019,20 @@ zap_buffer_unlocked:
  * void journal_invalidatepage() - invalidate a journal page
  * @journal: journal to use for flush
  * @page:    page to flush
- * @offset:  length of page to invalidate.
+ * @offset:  offset of the range to invalidate
+ * @length:  length of the range to invalidate
  *
- * Reap page buffers containing data after offset in page.
+ * Reap page buffers containing data in specified range in page.
  */
 void journal_invalidatepage(journal_t *journal,
 		      struct page *page,
-		      unsigned long offset)
+		      unsigned int offset,
+		      unsigned int length)
 {
 	struct buffer_head *head, *bh, *next;
+	unsigned int stop = offset + length;
 	unsigned int curr_off = 0;
+	int partial_page = (offset || length < PAGE_CACHE_SIZE);
 	int may_free = 1;
 
 	if (!PageLocked(page))
@@ -2043,6 +2040,8 @@ void journal_invalidatepage(journal_t *journal,
 	if (!page_has_buffers(page))
 		return;
 
+	BUG_ON(stop > PAGE_CACHE_SIZE || stop < length);
+
 	/* We will potentially be playing with lists other than just the
 	 * data lists (especially for journaled data mode), so be
 	 * cautious in our locking. */
@@ -2052,10 +2051,14 @@ void journal_invalidatepage(journal_t *journal,
 		unsigned int next_off = curr_off + bh->b_size;
 		next = bh->b_this_page;
 
+		if (next_off > stop)
+			return;
+
 		if (offset <= curr_off) {
 			/* This block is wholly outside the truncation point */
 			lock_buffer(bh);
-			may_free &= journal_unmap_buffer(journal, bh);
+			may_free &= journal_unmap_buffer(journal, bh,
+							 partial_page);
 			unlock_buffer(bh);
 		}
 		curr_off = next_off;
@@ -2063,7 +2066,7 @@ void journal_invalidatepage(journal_t *journal,
 
 	} while (bh != head);
 
-	if (!offset) {
+	if (!partial_page) {
 		if (may_free && try_to_free_buffers(page))
 			J_ASSERT(!page_has_buffers(page));
 	}
@@ -2089,12 +2092,17 @@ void __journal_file_buffer(struct journal_head *jh,
 	if (jh->b_transaction && jh->b_jlist == jlist)
 		return;
 
-	/* The following list of buffer states needs to be consistent
-	 * with __jbd_unexpected_dirty_buffer()'s handling of dirty
-	 * state. */
-
 	if (jlist == BJ_Metadata || jlist == BJ_Reserved ||
 	    jlist == BJ_Shadow || jlist == BJ_Forget) {
+		/*
+		 * For metadata buffers, we track dirty bit in buffer_jbddirty
+		 * instead of buffer_dirty. We should not see a dirty bit set
+		 * here because we clear it in do_get_write_access but e.g.
+		 * tune2fs can modify the sb and set the dirty bit at any time
+		 * so we try to gracefully handle that.
+		 */
+		if (buffer_dirty(bh))
+			warn_dirty_buffer(bh);
 		if (test_clear_buffer_dirty(bh) ||
 		    test_clear_buffer_jbddirty(bh))
 			was_dirty = 1;
@@ -2102,6 +2110,8 @@ void __journal_file_buffer(struct journal_head *jh,
 
 	if (jh->b_transaction)
 		__journal_temp_unlink_buffer(jh);
+	else
+		journal_grab_journal_head(bh);
 	jh->b_transaction = transaction;
 
 	switch (jlist) {
@@ -2159,13 +2169,14 @@ void 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 __journal_refile_buffer(struct journal_head *jh)
 {
-	int was_dirty;
+	int was_dirty, jlist;
 	struct buffer_head *bh = jh2bh(jh);
 
 	J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
@@ -2185,10 +2196,20 @@ void __journal_refile_buffer(struct journal_head *jh)
 
 	was_dirty = test_clear_buffer_jbddirty(bh);
 	__journal_temp_unlink_buffer(jh);
+	/*
+	 * We set b_transaction here because b_next_transaction will inherit
+	 * our jh reference and thus __journal_file_buffer() must not take a
+	 * new one.
+	 */
 	jh->b_transaction = jh->b_next_transaction;
 	jh->b_next_transaction = NULL;
-	__journal_file_buffer(jh, jh->b_transaction,
-				jh->b_modified ? BJ_Metadata : BJ_Reserved);
+	if (buffer_freed(bh))
+		jlist = BJ_Forget;
+	else if (jh->b_modified)
+		jlist = BJ_Metadata;
+	else
+		jlist = BJ_Reserved;
+	__journal_file_buffer(jh, jh->b_transaction, jlist);
 	J_ASSERT_JH(jh, jh->b_transaction->t_state == T_RUNNING);
 
 	if (was_dirty)
@@ -2196,30 +2217,21 @@ void __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.
- *
- * __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 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! ***
+ * __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 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);
-
 	__journal_refile_buffer(jh);
 	jbd_unlock_bh_state(bh);
-	journal_remove_journal_head(bh);
-
 	spin_unlock(&journal->j_list_lock);
 	__brelse(bh);
 }