scsi: scsi_debug: Address races following module load

When scsi_debug is loaded as a module with many (simulated) hosts, targets,
and devices (LUs), modprobe can take a long time to return.  Only a small
amount of this time is spent in the scsi_debug_init(); the rest is other
parts of the kernel reacting to to the appearance of new storage
devices. As soon as scsi_debug_init() has completed the user space may call
'rmmod scsi_debug' and this was found to cause race problems as outlined
here:

    https://bugzilla.kernel.org/show_bug.cgi?id=212337

To reliably generate this race a sysfs parameter called rm_all_hosts was
added and the code was strengthened in this area. The main change was to
make the count of scsi_debug hosts present an atomic.  Then it was found
that the handling of the existing add_host parameter needed the same
strengthening. Further: 'echo -9999 >
/sys/bus/pseudo/drivers/scsi_debug/add_host has the same effect as
rm_all_hosts so rm_all_hosts was not needed.

To inhibit a race between two invocations of writes to add_host, a mutex
was added. Also address a possible race when rmmod is called but LUs are
still being added.

The logic to remove (all) hosts is rather crude: it works backwards down a
linked lists of hosts. Any pending requests are terminated with
DID_NO_CONNECT as are any new requests. In the case where not all hosts are
being removed, the ones that remain may have lost requests as just
outlined. The lowest numbered host (id) hosts will remain.

Cc: Bart Van Assche <bvanassche@acm.org>
Link: https://lore.kernel.org/r/20220109012853.301953-2-dgilbert@interlog.comSigned-off-by: Douglas Gilbert <dgilbert@interlog.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>