On Fri, Jun 16, 2000 at 06:40:10PM -0700, Ethan Solomita wrote:
> Thor Lancelot Simon wrote:
> > 
> > This is all really much ado over nothing.
> > 
> 	I agree. The warning that started this whole discussion warned against
> using *any* async filesystem, where soft updates was given as an
> example. If someone is really motivated, I guess you could use
> ktrace/kdump to see if it opens all its files O_SYNC. Or you could put
> it on its own partition and mount it sync.

I think once again we're running into a terminology problem.

When you say "async filesystem" I think of Linux ext2fs, or FFS mounted
-o async: *no* synchronous writes are done except at unmount time.  Such
a filesystem *cannot* provide guarantees about the consistency of its
metadata.

That's distinctly different from both FFS with soft updates, FFS without
soft updates, and even LFS.  There, the filesystem guarantees that the
metadata will be consistent -- and, generally, that they will be
consistent with the file data, though the reordering of directory
and file operations in softdep that's just been discussed here worries
me a bit -- and that you can get "synchronous" writes by inserting
explicit barriers, e.g. with fsync().  It's easy to use fsync() to
guarantee consistency at critical points *without* taking the hit of 
doing synchronous writes all the time.

It's also quite different from what I think of when you say "sync
filesystem", where no asynchronous writes are ever done.  FFS
mounted -o sync should be such a filesystem, though currently
some async writes seem to slip through.

-- 
Thor Lancelot Simon	                                      tls@rek.tjls.com
	"And where do all these highways go, now that we are free?"