When do files syncronize?

[Kevin Oberman]

> From: Michael Collette <metrol at metrol.net>
> Date: Tue, 27 May 2003 12:35:06 -0700
> Sender: owner-freebsd-mobile at freebsd.org
> 
> Here on this Thinkpad T23 I normally just Fn-F4 to put it to sleep when moving 
> it from work to home.  This had been working pretty well, with an occasional 
> lock up still happening.  That much is a pretty well known problem.
> 
> This last Friday I did the usual Fn-F4 and brought my lappy on home.  What I 
> didn't do was plug it back in until late Sunday.  Bad move, as my battery is 
> starting to get on the weak side here.  Sure enough, lost power while it was 
> in sleep mode.
> 
> Powering back up the system went through the usual post-crash fsck.  
> Everything looked cool until I started up KMail.  Lost the entirety of my 
> mail settings, which were anything but trivial.  I wrote up a pretty bitter 
> bug report over on KDE's site.  I really wasn't in a good mood when this 
> happened.
> 
> http://bugs.kde.org/show_bug.cgi?id=58954
> 
> I then mentioned this problem on the KDE-FreeBSD mailing list.  I got some 
> advice there to turn off all write cacheing via loader.conf.  Now that's 
> done, my system seems to be a little bogged, but not too horribly bad.
> 
> With my sob story out of the way, here's my questions...
> 
> When this here Thinkpad goes into a sleep mode, is there any chit chat with 
> the OS to be sure and syncronize any outstanding writes?
> 
> Would it be wiser to initiate going into sleep via software, such as KDE's 
> laptop app or some other method?  Is there any difference?
> 
> Is there some way to force outstanding writes to clear their buffers before I 
> put this box into sleep mode?
> 
> 
> At some point I'd like to turn on write cacheing again.  I REALLY don't feel 
> comfortable about it at this time though.  Is there some way to get a feeling 
> of comfort back with this feature?

You now know exactly why there is a strong recommendation that write
cache be disabled on boxes that are subject to power failure. Files are
synced just fine, but that only gets updates to the write cache. There
is no way of being sure just when the data will actually get written to
the disk and, if the system loses power, the answer is never. This also
screws up write ordering needed by soft updates to make sure that the
metadata stays consistent.

You case is a "worst case" in that you had probably just updated this
file and the metadata for it was still queued up. You sent it off to
sleep, which spin down the disk. Since the metadata was never updated,
the file ceased to exist (even though all of the data had probably made
it to the disk).

The cost of turning off write cache is high. I have demonstrated a factor
of 8 slow-down for dd with write cache disabled.

A possible alternative is to make sure that you wait a minute after
terminating use of the system before suspending, Another is to do a
hibernate instead of a suspend. This will write all of you memory to
disk before turning off. (This also cuts battery use as the system is
REALLY off when you hibernate.) You will need to create a hibernation
slice on your hard disk to do this. A utility to create it is available
from the IBM web site. http://www.pc.ibm.com/qtechinfo/MIGR-4PESMK.html

R. Kevin Oberman, Network Engineer
Energy Sciences Network (ESnet)
Ernest O. Lawrence Berkeley National Laboratory (Berkeley Lab)
E-mail: oberman at es.net			Phone: +1 510 486-8634