Stress testing the UFS2 filesystem

[Robert Watson]

On Wed, 3 May 2006, Pavel Merdin wrote:

>> Actually the filesystems mounts without any problems if fsck is run first
>
> That's not a bug in this case. It's a feature. And there is nothing new in 
> this. Background fsck helps saving start time, but it's risky as kernel can 
> panic if system accesses problematic sector before fsck. So background fsck 
> should be turned off if one needs reliability (e.g. on servers).

Well, the feature and problem are that bgfsck relies on invariants holding 
true for data written to the disk, so corrects only a narrow set of expected 
failure modes.  I.e., that soft updates really does sequence changes out to 
the disk such that certain invariants regarding meta-data hold.  This allows 
access to the file system before the checks are complete, but places 
assumptions on what will be found on disk.  Those invariants might fail to 
hold for a few reasons -- bugs in UFS, hardware failure, and design breakage 
in the hardware are among the most common.  One known problem is that the 
guarantees provided by recent ATA disks are really very weak with respect to 
the expectations of currently file systems.

Robert N M Watson