Re: [PATCH] ext3: ext3_commit_super should always mark super uptodate

[Phillip Susi <psusi@cfl.rr.com>]

Jeff Mahoney wrote:
>  After a superblock write failure, the buffer_head is marked !uptodate.
>  Since the superblock is something of an exception -- read once and
>  a reference kept for the duration of the mount -- it is by definition
>  always uptodate.

Not if it has been modified and not yet written back to disk.  Then it 
is dirty by definition.

>  This is somewhat academic for the most part until we encounter error
>  conditions. For example, if a disk goes away in a SAN environment, the
>  write failure will occur and it will be followed by others. ext3_error
>  wants to mark the superblock dirty via ext3_commit_super, but once
>  the first write fails, the subsequent mark_buffer_dirty calls will
>  issue warnings because the buffer is not uptodate.

Why on earth is a warning issued for marking an already dirty buffer 
dirty again?  Surly buffers can be modified again and thus, marked as 
dirty a second time before pdflush gets around to syncing them all the time.

> +	set_buffer_uptodate(sbh);
>  	mark_buffer_dirty(sbh);
>  	if (sync)
>  		sync_dirty_buffer(sbh);

Are you sure this only matters when there is an IO error?  It looks 
there like if !sync then the sb can continue to sit in memory marked as 
dirty for some time before someone tries to flush it to disk. 
Eventually can't we end up back through this code patch and calling 
mark_buffer_dirty again before it was ever flushed in the first place, 
even without an IO error?