Re: ext4 regression panic

On Thu, Jan 21, 2021 at 06:15:47PM +0800, Murphy Zhou wrote:

Hi Jack,

A panic was introduced by this commit. It's easy and reliable to
reproduce.

commit 2d01ddc86606564fb08c56e3bc93a0693895f710
Author: Jan Kara [off-list ref]
Date:   Wed Dec 16 11:18:40 2020 +0100

    ext4: save error info to sb through journal if available

Hi Murphy,

Thanks for the bug report.  What's happening is that we haven't yet
initialized mballoc yet --- that happens in line 4943 of
fs/ext4/super.c, in ext4_fill_super().

But in line 4903 (in the case of the BZ #199275 reproducer), we
attempt to fetch the root inode, which is fails because it is
unallocated.  That then triggers a call to ext4_error(), which now
results in a journalled change, since the journal is initialized
starting in line 4793, and in line 4838, we set up the
j_commit_callback, which is what ends up calling
ext4_process_freed_data(), but since the multiblock allocator hasn't
been set up yet, that causes the NULL pointer dereference.

So what we need to do is to *not* set up the callback until after the
call to ext4_mb_init().

We should probably create an ext4-specific test in xfstests which
tries mounting a small, deliberately corrupted file system, to make
sure we handle this case correctly in the future.

						- Ted

commit 6c2f9a8247273cf1108ff71c99680b7457f48318
Author: Theodore Ts'o [off-list ref]
Date:   Thu Jan 21 12:33:20 2021 -0500

    ext4: don't try to processed freed blocks until mballoc is initialized
    
    If we try to make any changes via the journal between when the journal
    is initialized, but before the multi-block allocated is initialized,
    we will end up deferencing a NULL pointer when the journal commit
    callback function calls ext4_process_freed_data().
    
    The proximate cause of this failure was commit 2d01ddc86606 ("ext4:
    save error info to sb through journal if available") since file system
    corruption problems detected before the call to ext4_mb_init() would
    result in a journal commit before we aborted the mount of the file
    system.... and we would then trigger the NULL pointer deref.
    
    Cc: Jan Kara [off-list ref]
    Reported by: Murphy Zhou [off-list ref]
    Signed-off-by: Theodore Ts'o [off-list ref]

diff --git a/fs/ext4/super.c b/fs/ext4/super.c
index 0f0db49031dc..802ef55f0a55 100644
--- a/fs/ext4/super.c
+++ b/fs/ext4/super.c

@@ -4876,7 +4876,6 @@ static int ext4_fill_super(struct super_block *sb, void *data, int silent)
 
 	set_task_ioprio(sbi->s_journal->j_task, journal_ioprio);
 
-	sbi->s_journal->j_commit_callback = ext4_journal_commit_callback;
 	sbi->s_journal->j_submit_inode_data_buffers =
 		ext4_journal_submit_inode_data_buffers;
 	sbi->s_journal->j_finish_inode_data_buffers =

@@ -4993,6 +4992,14 @@ static int ext4_fill_super(struct super_block *sb, void *data, int silent)
 		goto failed_mount5;
 	}
 
+	/*
+	 * We can only set up the journal commit callback once
+	 * mballoc is initialized
+	 */
+	if (sbi->s_journal)
+		sbi->s_journal->j_commit_callback =
+			ext4_journal_commit_callback;
+
 	block = ext4_count_free_clusters(sb);
 	ext4_free_blocks_count_set(sbi->s_es, 
 				   EXT4_C2B(sbi, block));