r355056 added making the outstanding commands busy before completing them. This
got around the panic from the state machine being incorrect. However, it lead to
double completions of the commands. After we sent MPI2_SAS_OP_REMOVE_DEVICE, the
IOC would send the normal completion records for them back, so we were racing
that process. The race produced later panics with the command in the wrong state
(usually trying to allcoate a busy command).
Instead, move the fail-safe completion of the outstanding commands from just
after we send MPI2_SAS_OP_REMOVE_DEVICE to that commands successful completion
routine. This eliminates the race.
With this change, I'm now able to turn off the PHY on very active disks without
it leading to a panic. Prior to this change, a moderate amount of traffic to the
device was all that was required to trigger a panic all the time.