henning, all, > > > It's a matter of fact that qmail isn't as reliable on Linux as on *BSD as it > > > relies on some FFS semantics ext2fs just doen not fulfill. > > If that is the case then there is a bug in qmail - the software should > > be correct to the system underneath it, not apply semantics that aren't > > supported and say "it's their problem, not ours...!" when it doesn't > > work properly. > > No, it's a bug in ext2fs. that's crap, of course. if it's a bug in ext2 then it's a bug in softupdates and large numbers of other filesystems that don't provide syncrhonous metadata updates by default (although all of them do and can). there has just been a long discussion about exactly this issue on the linux kernel mailing list, specifically whether the SUS (single unix specification) requires than an fsync also synch a directory tree up to the root before returning. the conclusion of many people on the lkml (including those advocating the behavior) is that SUS does not require this behavior although it would be nice to offer an ext2 or vfs option for supporting it. people also noted that linux uses the semantics of fsync(dir) to synchronize the name metadata for a file and that that is fast and efficient (as others have noted, ext2 also offers the option of mounting filesystems and directories in synchronous mode, but that slows down everything of course). a patch to make qmail fsync the parent directory has been around for a long time and causes qmail to behave reliably (and fast!) on linux. the bottom line is that FFS semantics are not universal and they haven't been for a long time, even among bsds. assuming that they are is a mistake. todd underwood vice president & chief technology officer oso grande technologies, inc. [EMAIL PROTECTED]
