TL;DR: Three ZFS tests created ZFS pools on all unmounted devices listed
in /etc/fstab, corrupting their contents. Stop that.
Imagine my surprise when the ESP on my main dev/test VM would "randomly"
become corrupted, making it unbootable. Three tests collect various devices
from the system and try to add them to a test pool. The test expects this
to fail because it _assumes_ these devices are in use and ZFS will correctly
reject the request.
My /etc/fstab has two entries for devices in /dev:
/dev/gpt/swap0 none swap sw,trimonce,late /dev/gpt/esp0 /boot/efi msdosfs rw,noauto
Note the noauto on the ESP. In a remarkable example of irony, I chose
this because it should keep the ESP more protected from corruption;
in fact, mounting it would have protected it from this case.
The tests added all of these devices to a test pool in a _single command_,
expecting the command to fail. The swap device was in use, so the command
correctly failed, but the ESP was added and therefore corrupted. However,
since the command correctly failed, the test didn't notice the ESP problem.
If each device had been added with its own command, the test _might_ have
noticed that one of them incorrectly succeeded. However, two of these
tests would not have noticed:
hotspare_create_001_neg was incorrectly specified as needing the Solaris
dumpadm command, so it was skipped. _Some_ of the test needs that command,
but it checks for its presence and runs fine without it.
Due to bug 241070, zpool_add_005_pos was marked as an expected failure.
Due to the coarse level of integration with ATF, this test would still
"pass" even if it failed for the wrong reason. I wrote bug 267554 to
reconsider the use of atf_expect_fail in these tests.
Let's further consider the use of various devices found around the system.
In addition to devices in /etc/fstab, the tests also used mounted devices
listed by the mount command. If ZFS behaves correctly, it will refuse
to added mounted devices and swap devices to a pool. However, these are
unit tests used by developers to ensure that ZFS still works after they
modify it, so it's reasonable to expect ZFS to do the _wrong_ thing
sometimes. Using random host devices is unsafe.
Fix the root problem by using only the disks provided via the "disks"
variable in kyua.conf. Use one to create a UFS file system and mount it.
Use another as a swap device. Write code to use a third as a dump device,
but leave it disabled due to bug 241070.
While I'm here:
Due to commit 6b6e2954dd65, we can simply add a second dump device and
remove it in cleanup. We no longer need to save, replace, and restore the
pre-existing dump device.
The cleanup_devices function used camcontrol inquiry to distinguish disks
from other devices, such as partitions. That works fine for SCSI, but not
for ATA or VirtIO block. Use geom disk list instead.