Instead of proceeding to have mpiexec fail or hang,
do an up-front check that the MPI version is the same
on all of the nodes in the test.
Signed-off-by: John Spray <john.spray@redhat.com>
Previously this was remounting in the background
which meant that if a workunit was executed immediately
afterwards it could put files in the mount dir before
the FS was mounted, causing the mount to fail.
Signed-off-by: John Spray <john.spray@redhat.com>
A place for any tests which should be run with
standbyreplay MDSs in play. Initially just
the mds_journal_migration task, to cover
the #8811 case.
Signed-off-by: John Spray <john.spray@redhat.com>
Hopefully this makes it easier to digest and maintain
than reading yaml fragments.
Signed-off-by: Josh Durgin <josh.durgin@inktank.com>
Signed-off-by: Yuri Weintstein <yuri.weinstein@inktank.com>
This failure to delete used to be hidden in
a catch-all exception handler that was
removed in 396dc073.
This was causing failures in kcephfs suite
like:
Command failed on plana96 with status 1: 'rm -rf -- /home/ubuntu/cephtest/mnt.0/client.0'
Signed-off-by: John Spray <john.spray@redhat.com>
This was always trying to remove the mountpoint,
and then swallowing the exception and printing
it to the log.
Instead:
* Ensure it's only trying to delete mnt if it created it
* Don't swallow those exceptions: if deleting the dirs
fails then something has gone wrong and it should bubble
up.
Signed-off-by: John Spray <john.spray@redhat.com>
This enables tasks like mds_journal_migration to be
run in an environment with standby-replay MDSs present.
Signed-off-by: John Spray <john.spray@redhat.com>
This has been timing out for weeks to months (hard to say since before that
we were consistently hitting an OSD crash on shutdown). The runs I've looked
at are making progress and seem fine ... just slow because of valgrind and
because of thrashing.
Backport: firefly
Signed-off-by: Sage Weil <sage@redhat.com>
Follow up to a24efaef, remove the mds from
the (mainly singleton) test configs that
explicitly list it.
Signed-off-by: John Spray <john.spray@redhat.com>
The difference is whether we upgrade all at once or each individually.
Also, expect a quorum between a mix of dumpling and firefly monitors.
Signed-off-by: Sage Weil <sage@redhat.com>
Since refactoring ceph_fuse and kclient tasks
to store Mount objects on ctx, the mds_journal_migration
task can use those instead of explicitly instantiating its own.
Signed-off-by: John Spray <john.spray@redhat.com>
No need to install ceph just to get daemon-helper, adjust-ulimits, etc..
These utilities are used by lots of tasks, not just ceph.
Signed-off-by: Sage Weil <sage@redhat.com>
