Anand Buddhdev <[EMAIL PROTECTED]> wrote:
>
> Dan does not recommend the use of a softupdates file system for the
> queue. If FFS+softupdates = standard FFS+faster, then there should be
> no harm in using it.
This is something I _can_ comment on. In fact, I already did, in
message 42203 on this list:
"Len Budney" <[EMAIL PROTECTED]> wrote:
> [EMAIL PROTECTED] wrote:
> > Using softupdates under *BSD gives you the reliability of sync
> > (somewhat more, actually), with nearly the speed of async.
>
> In October of 1999, that wasn't quite true...
>
> ``In particular, if you put a mail queue on a softupdates filesystem,
> you can lose mail when the power goes out, just as if you were using
> Linux.'' -- Dan Bernstein, <http://x38.deja.com/getdoc.xp?AN=539496358>
>
> From Dan's other comments in that thread, it appears that the rename()
> call returns success before the rename was committed to disk. Dan said
> ``mail programs'' rely on rename() to tell the truth. A look at the
> source code suggests that qmail is not one of them--qmail-queue doesn't
> use the rename() call.
>
> If link() is subject to the same complaint as rename(), then qmail
> probably _does_ have the same reliability problem on a softupdates
> filesystem as on an async filesystem.
Len.
--
Some people are suffering from the delusion that program modularity is
incompatible with good performance.
-- Dan Bernstein
