[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]

Re: data corrupting bug in 2.4.20 ext3, data=journal

"Stephen C. Tweedie" wrote:
> ...
> The problem is that ext3 is expecting that truncate_inode_pages() (and
> hence ext3_flushpage) is only called during a truncate.

That's OK, because there shouldn't be any dirty data attached to the
inodes at that time.  But there is, because the commit which write_super()
started hasn't finished yet:

static int do_sync_supers = 0;
        target = log_start_commit(EXT3_SB(sb)->s_journal, NULL);

        if (do_sync_supers) {
                log_wait_commit(EXT3_SB(sb)->s_journal, target);

We need to do a full sync at unmount.

I assume that in other journalling modes the buffers are dirty
anyway, so sync_buffers() gets them.

And indeed enabling do_sync_supers fixes it up in both 2.4 and 2.5 (2.5
doesn't have sync_buffers(), but sync_inodes_sb() gets everything).

But are we sure that ->write_super() will always be called?

int fsync_super(struct super_block *sb)
        if (sb->s_dirt && sb->s_op && sb->s_op->write_super)

I suspect that if s_dirt is not set, and we have dirty inodes,
write_super is not called and nothing will force a commit anywhere
in the unmount process.  Which could explain the similar failure
in 2.4.19-rc1 which Nick reports.

We need to be able to distinguish between a periodic flush of the
superblock and a real sync.  The write_super overload has always
been uncomfortable.

Google for "write_super is not for syncing" ;)  I think Chris's
patch is the way to fix all this up.

[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]