On Thursday 19 December 2002 01:03 pm, Dan Melomedman wrote:
> Ross Patterson wrote:
> > To sum up, has it been definitively established that softdeps are safe? 
>
> The way I understand it is they were designed to guarantee metadata
> integrity, not data integrity. 

If I read McKusick & Ganger's papers correctly, softdeps were designed to 
provide the same degree of metadata integrity as synchronous writes with the 
improved runtime performance of asynchronous writes and the improved 
dirty-mount performance of journaled filesystems.  The sacrifice was to 
accept the lower degree of file-data integrity that asynchronous writes 
provide.

So if softdeps are working correctly, you gain performance but you lose some 
file-data integrity on open files that haven't been fsync()'ed.  That's a 
trade-off decision I (or most other folks) can make on a case-by-case basis.  
I gather that the NetBSD implementation of softdeps does *not* include the 
background-fsck that allows for mounting dirty filesystems, though, so we can 
get the almost-asynch write performance but not the boot-time speedup.  
Still, that's much better than without softdeps.
-- 
Ross A. Patterson
CatchFIRE Systems, Inc.
5885 Trinity Parkway, Suite 220
Centreville, VA  20120
(703) 563-4164