nfs_wb_all

nfs_wb_all is the only caller of sync_inode which does not set nr_to_write to 0, that is the only one actually using it to write out file data.

Now looking at the callers I'm not sure this actually is intended:

- nfs_msync_inode does a filemap_fdatawrite before and a filemap_fdatawait after the call to nfs_wb_all, o the data writing happens twice here, although it probably really should be a filemap_write_and_wait before the call to sync_inode
- nfs_rename doesn't write data next to it. Then again writing out data here from sync_inode probably isn't intentional either.
- nfs_do_fsync already has the data written back from vfs_fsync if called though ->fsync although a data write is needed for the call from ->flush and O_SYNC handling, although the latter should probably be switched to use generic_write_sync
- nfs_sync_mapping seems to actually want the data writeback
- nfs_setattr already does a filemap_write_and_wait just before the call to nfs_wb_all

The intention is that nfs_wb_all() should do a full data integrity sync.

nfs_msync_inode() should just be replaced with a single call to nfs_wb_all().

The call to nfs_wb_all() in nfs_rename() should probably be removed. It was put in there because some (broken!) nfs servers don't preserve filehandles across renames. On a compliant nfs server, we shouldn't have to worry about that.

There is one special case that generic_write_sync() won't catch: when the writeback code hits an error condition, we try to switch the writes to synchronous mode until the error clears. It is mainly in order to deal with temporary problems such as ENOSPC on the server.

So we can probably replace that call to nfs_wb_all() with a call to nfs_commit_inode(), and then replace the calls to nfs_do_fsync() in nfs_file_write() and nfs_file_splice_write() with calls to vfs_fsync().

Again, we should just get rid of the filemap_write_and_wait().

Cheers
Trond