jbd2: Fix a race between checkpointing code and journal_get_write_access()

The following race can happen:

CPU1 CPU2
checkpointing code checks the buffer, adds
it to an array for writeback
do_get_write_access()
...
lock_buffer()
unlock_buffer()
flush_batch() submits the buffer for IO
__jbd2_journal_file_buffer()

So a buffer under writeout is returned from
do_get_write_access(). Since the filesystem code relies on the fact
that journaled buffers cannot be written out, it does not take the
buffer lock and so it can modify buffer while it is under
writeout. That can lead to a filesystem corruption if we crash at the
right moment.

We fix the problem by clearing the buffer dirty bit under buffer_lock
even if the buffer is on BJ_None list. Actually, we clear the dirty
bit regardless the list the buffer is in and warn about the fact if
the buffer is already journalled.

Thanks for spotting the problem goes to dingdinghua <[e-mail]>.

Reported-by: dingdinghua <[e-mail]>
Signed-off-by: Jan Kara <[e-mail]>
Signed-off-by: "Theodore Ts'o" <[e-mail]>

fs/jbd2/transaction.c

···
 	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;
+	char b[BDEVNAME_SIZE];
 
-	/* 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);
-
-		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);
 }
 
 /*
···
 			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);
···
 	J_ASSERT_JH(jh, buffer_locked(jh2bh(jh)));
 
 	if (jh->b_transaction == NULL) {
+		/*
+		 * Previous jbd2_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));
 		jh->b_transaction = transaction;
 
 		/* first access by this transaction */
···
 
 	if (jh->b_cp_transaction) {
 		JBUFFER_TRACE(jh, "on running+cp transaction");
+		/*
+		 * 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);
 		__jbd2_journal_file_buffer(jh, transaction, BJ_Forget);
-		clear_buffer_jbddirty(bh);
 		may_free = 0;
 	} else {
 		JBUFFER_TRACE(jh, "on running transaction");
···
 	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;