On 2022/11/29 6:28, Theodore Ts'o wrote:
On Tue, Nov 08, 2022 at 02:06:29PM +0000, Luís Henriques wrote:
What makes you believe that? Look at how s_group_info is initialized in ext4_mb_alloc_groupinfo() in fs/ext4/mballoc.c. It's pretty careful to make sure this is not the case.
Right. I may be missing something, but I don't think we get that far. __ext4_fill_super() will first call ext4_setup_system_zone() (which is where this bug occurs) and only after that ext4_mb_init() will be invoked (which is where ext4_mb_alloc_groupinfo() will eventually be called).
I finally got around to taking a closer look at this, and I have a much better understandign of what is going on. For more details, and a suggested fix, please see:
https://bugzilla.kernel.org/show_bug.cgi?id=216541#c1 - Ted
Hi Theodore,
In my opinion, the s_journal_inum should not be modified when the file system is mounted, especially after we have successfully loaded and replayed the journal with the current s_journal_inum. Even if the s_journal_inumon the disk is modified, we should use the current one. This is how journal_devnum is handled in ext4_load_journal():
if (!really_read_only && journal_devnum && journal_devnum != le32_to_cpu(es->s_journal_dev)) { es->s_journal_dev = cpu_to_le32(journal_devnum);
/* Make sure we flush the recovery flag to disk. */ ext4_commit_super(sb); }
We can avoid this problem by adding a similar check for journal_inum in ext4_load_journal().