Re: Problem with delayed allocation

[Theodore Tso <tytso@mit.edu>]

On Mon, Aug 04, 2008 at 08:46:52AM +0530, Aneesh Kumar K.V wrote:
> __fsync_super use filemap_fdatawait(mapping) for waiting on writeback
> pages. But all the dirty pages of the inode are not in writeback because
> we might have had block allocation failures.

Yes, but that should only happen if the filesystem is full or a user's
quota is overrun, correct?

> Also with the current code base I am seeing buffer_heads which are
> unmapped, non delay and dirty That means writepages won't allocate
> block for them and writepage cannot write them.

I thought all writes went through the page cache?  Are you saying that
the *pages* are clean but the buffer_heads are marked dirty?  In that
case, ext4_da_writepages, if wbc.sync_mode is not WB_SYNC_NONE, *must*
wait on them and not return until the buffers are safely on disk,
since filemap_fdatawait(mapping) won't in __sync_single_inode() won't
do the waiting for the writepages routine.

							- Ted