Re: do_sync() and XFSQA test 182 failures....

[Christoph Hellwig <hch@xxxxxxxxxxxxx>]

On Sat, Nov 01, 2008 at 09:03:49AM +1100, Dave Chinner wrote:
> > >   if (!(sb->s_flags & MS_RDONLY))
> > > +         xfs_sync_fsdata(XFS_M(sb), SYNC_WAIT);
> > 
> > Not commenting on the rest yet, but I'm sure you noticed that before
> > your patch we ever called xfs_sync_fsdata with SYNC_WAIT at all.  I
> > think with the current codebase there is nothing to guarantee we ever
> > write the superblock out to disk as xfs_wait_buftarg specificly
> > excludes the superblock (XBF_FS_MANAGED)
> 
> Almost. unmount calls xfs_unmountfs_writesb() which does a
> synchronous write of the superblock. But in the current code sync(1)
> certainly never, ever triggers a superblock write we wait for...

True, missed it due to the opencoding.  With my patch to kill the
shared readonly bits we should be able to make it use xfs_sync_fsdata
for consistency.