Re: [patch 8/8] fs: add i_op->sync_inode

[Ted Ts'o <tytso@mit.edu>]

On Sat, Dec 18, 2010 at 12:46:42PM +1100, Nick Piggin wrote:
> --- linux-2.6.orig/Documentation/filesystems/porting	2010-12-18 00:32:44.000000000 +1100
> +++ linux-2.6/Documentation/filesystems/porting	2010-12-18 03:51:40.000000000 +1100
> +--
> +[mandatory]
> +	Inode writeback and syncing has undergone an overhaul.
> +write_inode_now doesn't sync (unless ->write_inode *always* syncs)
> +

Could you perhaps give a bit more detail in the documentation about
what sync_inode() is supposed to do?  I gather that it is supposed to
(based on the flags), do the specified combination of (a) do writeback
for the pages corresponding to the specified range (if INODE_SYNC_DATA
is given), (b) write out the inode if necessary depending on
INODE_SYNC_DATA_METADATA (for fdatasync) and INODE_SYNC_DATA (for
fsync).

When sync_inode() returns, what is guaranteed is that the appropriate
buffer_heads are dirtied (for both the metadata and the data if the
file system is still using the buffer cache for data I/O) and/or the
data has been pushed to the block I/O system, but the sync_inode()
method does not wait force the dirtied buffers out to disk, and it
does not wait for any initiated I/O to complete (either at the block
I/O layer or guaranteeing that the data has hit the platters by
issuing a barrier request).

Do I have that right?  If so, could we please add that to the
documentation, just so it's clear?  Thanks!!

						- Ted