Hi Elliot,

fdatasync is _exactly_ what we want. The problem is that fdatasync is not
implemented on Linux 2.0 as far as I know (Ted seems to have confirmed
this to me).  Do you know someone who can fill this gap?

fsync is MUCH worse than a raw device since it updates both the file data
_and_ the inode metadata involving a head move of the disk.   A somewhat
larger server ground to a halt doing that here. 

A third option is to use chattr +S which makes writes of attributes and
file data synchronous for just a single file.  Again, we got far too much
head movement. 

- Peter -

On Thu, 9 Oct 1997, Elliot Lee wrote:

> On Thu, 9 Oct 1997, Peter J. Braam wrote:
> 
> > log based scheme is used. On Linux however, we still need a raw character
> > /dev/hda (which does not use the buffer cache) before this is really
> 
> Wouldn't fdatasync() or fsync() achieve the same end result, while
> allowing higher performance? You could still get some buffering in places
> where you don't need things written right away, plus read buffering...
> 
> -- Elliot					http://www.redhat.com/
> "They don't let my code go into shipping products," Gates said. "They
>  haven't done that for eight years." (at the 1997 PDC)
>