Autzoo.1568
hacknews
utzoo!henry
Sun Apr 18 02:17:24 1982
crashes, new unix
After a hard-error-in-swap crash, and two attempts to reboot which were
unsuccessful because the same error showed up just before coming up, I
had to kludge.

Fortunately, I'd been thinking about the problem a bit earlier, and had
noticed that it *is* possible to do partial verification of swapout
operations.  Due to the race conditions within the kernel, it is not
possible to compare core to disk, but the controller does have an operation
which simply checks the CRCs on the disk copy.  This does not protect
against cases where the CRCs are right but the data is wrong -- and we
do get such errors sometimes -- but it will help.

We are now running a new unix with a rapidly-kludged SMC driver which
does partial verifies on swapouts.  /chkunix is a tidier version, to
be installed next time we reboot (probably after full backups).  The
partial-verify code has already caught one error.

The rate of these errors is becoming serious;  we should try to lean
on Ampex about them.  Verifies do keep the problem under control, but
at the cost of degraded performance.  Some of my experiences during the
kludging also suggest that we are getting a truly gross rate of cylinder
errors, which the driver copes with, but again at a price in performance.