Many tape devices will automatically rewind following a poweron/reset.
This can result in data loss as other operations in the driver can write to
the tape when the position is unknown. E.g. MTEOM can write a filemark at
the beginning of the tape. This patch adds code to detect poweron/reset
unit attentions and prevents the driver from writing to the tape when the
position could be unknown.
Customer reported problem description:
We have experienced an issue with the SCSI tape driver (st) which has led
to data loss for us on two separate occasions in production, as well as in
a third case in which we were able to reproduce the failure in our test
environment.
The tape device involved is an Amazon Tape Gateway, a virtual tape library
(VTL) appliance which presents as a series of iSCSI targets (multiple tape
drives and a changer) and is backed by storage in Amazon S3. The problem is
a general one and not limited to any particular SCSI transport or tape
device, though the nature of both iSCSI and the VTL make data loss somewhat
more likely with this combination than with a physical tape drive.
The observed behavior occurs when an error causes the VTL tape gateway
process (on the appliance) to crash and restart. This interrupts the iSCSI
TCP connections and, when it occurs during a write, causes the write to
fail with EIO. However, we then find that the virtual tape in question is
now completely blank. We raised this issue with AWS support, thinking this
must be a bug in the VTL appliance, but that turns out not to be the case.
Per AWS support, when the gateway crashes in this manner, its notion of the
current tape position is reset to the beginning of the tape. It also sets a
unit attention condition, such that the next request results in a CHECK
CONDITION status with sense key UNIT ATTENTION and asc/ascq indicating a
device reset. According to their logs the next command being sent is WRITE
FILEMARK, which results in writing an FM at the beginning of the tape,
effectively discarding its contents.
In fact, once the write fails with EIO, our software attempts to recover by
rewinding and repositioning the tape, then resuming operation. If this
fails, it attempts to rewind and reposition again, write a marker at the
end of the tape, and then unmount. It does not under any circumstances
write either data or filemarks without having successfully positioned the
tape to a known point.
What actually happens is that, since the last operation was a write, the
kernel executes an implied MTWEOF operation (which translate to a Write
Filemarks command) before the rewind that was actually requested. This
seems not entirely unreasonable, provided the tape position is known.
However, once this request fails (due to the unit attention condition), our
next rewind attempt also triggers an implied MTWEOF, which does _not_ fail
(the unit attention condition persists only until the initiator has been
notified); this is the command that unexpectedly erases the tape.
Our analysis is that the st driver is in fact completely ignoring the UNIT
ATTENTION and associated reset notification from the device. This is not a
condition that can be detected in the transport or mid-layer, as it occurs
entirely within the target and is reported only via the UNIT ATTENTION
sense key. The upper driver (i.e. st) needs to detect this indication and
reset its internal model of the device to an unknown state.
Suggested-by: Jeffrey Hutzelman <jhutz@cmu.edu>
Signed-off-by: John Meneghini <jmeneghi@redhat.com>
Link: https://lore.kernel.org/r/20230822181413.1210647-1-jmeneghi@redhat.com
Acked-by: Kai Mäkisara <kai.makisara@kolumbus.fi>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>