Re: kern/38717: sysinst shouldn't create LFS file systems

[Alan Barrett <apb%cequrux.com@localhost>]

On Tue, 27 May 2008, Andrew Doran wrote:
>  I have done a lot of stress testing on the file system code recently. Using
>  a mixture of SGI fsstress, fsx, bonnie and postmark over different versions
>  of NetBSD, my experience is that:
>  
>  - ffs, ext2fs, and tmpfs in -current /seem/ unbreakable
>  - lfs and ffs+softdep can be broken within minutes
>  - lfs in -current breaks within seconds

For me, ffs used to be stable, but the past two or three months have
been terrible.  Heavy file system activity such as rsync between two
disks easily crashes the machine; I have never even managed to get a
backtrace because the hotkey to flip to the console didn't work, so I
haven't sent any PRs.  Twice recently I have noticed files that ended
up with random binary content (of exactly the same size as the original
file) after a crash.  (Both times, they were files that would have been
created or overwritten shortly before the crash.)

ffs with -o async is much worse than ffs without -o async; I suspect
that it uses up all memory with buffers waiting to be written to disk,
and then something fails due to lack of memory.  Again, I haven't been
able to get backtraces.

--apb (Alan Barrett)