GitLab

ExpandNames holds too much non-locking code (first LU to be converted to
ExpandNames, and we didn't have CheckArguments at that poin), and this
patch moves the checks that are lock-independent to CheckArguments.
Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: Michael Hanselmann <hansmi@google.com>

This fixes an old 'FIXME' entry.
Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: Michael Hanselmann <hansmi@google.com>

This will be used to conditionally enable the watcher node maintenance
feature.
Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: Guido Trotter <ultrotter@google.com>

Signed-off-by: Guido Trotter <ultrotter@google.com>
Reviewed-by: Iustin Pop <iustin@google.com>

Abstract the growable disk types in a ganeti constants, and only run
disk grow, from burnin, on them.
Signed-off-by: Guido Trotter <ultrotter@google.com>
Reviewed-by: Iustin Pop <iustin@google.com>

Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: Guido Trotter <ultrotter@google.com>

Per issue 90, current cluster verify is very very brittle. It's one of
the oldest pieces of code, with only additions without cleanups over the
last years.

Among its problems:

- data initialization interspersed with verification of RPC results,
  leading to non-initialized data for some branches
- due to the above, we order strictly some checks and we have the case
  where a bad node time result will skip checking of node volumes
- many many local variables, with each new check adding a new dict,
  leading to a spaghetti of dicts in the main Exec function
- monolithic code, both Exec() and _NodeVerify() do a lot of
  independent checks

This patch does an imperfect rewrite, but at least we gain:

- a clear infrastructure for adding more checks (the new NodeImage
  class, with it's clear and documented fields), and removal of most
  per-node dicts from the Exec() function
- the new NodeImage object should allow better type safety, e.g. by
  allowing pylint to check the actual object attributes rather than
  strings as dict keys
- a-priori initialization of data fields, eliminating the need to
  introduce dependencies between checks
- per-result-key status field, allowing elimination of duplicate error
  messages (where we want)
- split of most independent checks into separate functions, for greater
  clarity

The new code, being new will probably introduce for the short term more
bugs than it removes. However, it should offer a much better way for
extending cluster verify in the future.
Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: Guido Trotter <ultrotter@google.com>

This is a simple patch that adds the no-install mode for instance
creation, allowing import from foreign source of the actual OS (instead
of requiring the preparation of data in a form expected by the import
scripts).
Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: Michael Hanselmann <hansmi@google.com>

This patch modifies LUSetInstanceParms to allow OS name changes, without
reinstallation, in case an OS gets renamed on-disk.
Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: Michael Hanselmann <hansmi@google.com>

This patch moves the node-has-os checks to a separate function.
Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: Michael Hanselmann <hansmi@google.com>
Reviewed-by: Guido Trotter <ultrotter@google.com>

The current check on whether we require auto_promote or not is wrong, as
we check whether we will have exactly the correct number of master
candidates left. But it is fine if we have more (e.g. when CPS=10 and
mc_remaning=19) than the current number, and in that case we shouldn't
require auto promotion.
Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: Michael Hanselmann <hansmi@google.com>

Currently, the ganeti-confd's HMAC key is called “cluster HMAC key” or
simply “HMAC key” everywhere. With the implementation of inter-cluster
instance moves, another HMAC key will be introduced for signing critical
data. They can not be the same, so this patch clarifies the purpose of the
“cluster HMAC key” by renaming it. The actual file name is not changed.
Signed-off-by: Michael Hanselmann <hansmi@google.com>
Reviewed-by: Iustin Pop <iustin@google.com>

This is much simpler than the opposite, with fewer possibilities of
failures.
Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: Michael Hanselmann <hansmi@google.com>

This patch adds a new mode to instance modify, the changing of the disk
template. For now only plain to drbd conversion is supported, and the
new secondary node must be specified manually (no iallocator support).

The procedure for conversion works as follows:

- a completely new disk template is created, matching the count, size
  and mode of the instance's current disks
- we create manually (not via _CreateDisks) all the missing volumes
- we rename on the primary the LVs to the new name
- we create manually the DRBD devices

Failures during the creation of volumes will leave orphan volumes.
Failure during the rename might leave some disks renamed and some not,
leading to an inconsistent instance.

Once the disks are renamed, we update the instance information and wait
for resync. Any failures of the DRBD sync must be manually handled (like
a normal failure, e.g. by running replace-disks, etc.).
Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: Michael Hanselmann <hansmi@google.com>

Multiple LUs require that an instance is not running while they operate
on the instance (reinstall, rename, modify, recreate disks, deactivate
disks). The code to do this check is duplicate many times, and not very
consistent (some use call_instance_list, some call_instance_info).

The patch moves this check into a separate function that is then reused.
The only drawback is that _SafeShutdowInstanceDisks now raises an
OpPrereqError (even though it is run during Exec()), but this use case
is fine (there are no other modifications in that Exec).
Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: Michael Hanselmann <hansmi@google.com>

Both create instance and grow disk check the free disk space on nodes
using the same, duplicate code. Since we'll need this in other places in
the future, we abstract the check into a new function.

The patch adjusts the error message to be more in-line with the one for
memory checking, and fixes the exception raised for RPC errors.
Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: Michael Hanselmann <hansmi@google.com>

This is a simple check, but we'll need it in multiple places.
Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: Michael Hanselmann <hansmi@google.com>

This new mode, valid only for the plain template disk, allows creation
of an instance based on existing logical volumes (preserving data),
rather than creation of new volumes and OS creation.

The new mode works as follows:

- instead of size, all disks passed in must have an 'adopt' key, which
  signifies the LV name to be used
- all disks must have this key, or neither should
- we check the volume existence, and from the result we fill in the
  actual size
- online (in-use) volumes are not allowed
- 'stealing' of another's instance volumes is prevented via reservation
  of the LV names
- during creation, we rename the logical volumes to the standard Ganeti
  format (based on UUID)
Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: Guido Trotter <ultrotter@google.com>

This way, the parameters are available in CheckArguments too.
Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: Guido Trotter <ultrotter@google.com>

When using pyOpenSSL 0.7 or above, LUClusterVerify will start to show a
warning 30 days before a certificate expires. 7 days before the
certificate expires, the warning becomes an error. Once expired,
LUVerifyCluster will always report an error. The latter is also supported
with pyOpenSSL 0.6.
Signed-off-by: Michael Hanselmann <hansmi@google.com>
Reviewed-by: Iustin Pop <iustin@google.com>

Signed-off-by: Michael Hanselmann <hansmi@google.com>
Reviewed-by: Iustin Pop <iustin@google.com>

In case the hypervisor has issues on one node, currently
backend.VerifyNode will exit via an exception (two exit paths possible,
one via HypervisorError from hypervisor.Verify(), and one via RPCFail
from GetInstanceList). This is bad as it invalidates all other checks of
that node.

This patch catches these two errors and allows the rest of the
VerifyNode function to run. This leads to a more complete verify cluster
run, for example now only real missing LVs are reported, not all of
them.

The cluster verify is not perfect as it will skip some tests even if it
has data, but this will require a more complete rewrite (see issue 90).

Also, the patch fixes and improves some error messages in cmdlib.
Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: Guido Trotter <ultrotter@google.com>

In simulate errors mode, the test "ntime_diff is not None" will be
ignored, and thus a None value will try to be formatted as %.01f. We
workaround this by formatting it before, and then only using %s, which
can format a 'None' value.
Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: Guido Trotter <ultrotter@google.com>

This adds a validation similar to the one for cluster-wide hypervisor
paramters.
Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: René Nussbaumer <rn@google.com>

The current code in LUSetNodeParms regarding the demotion from master
candidate role is complicated and duplicates the code in ConfigWriter,
where such decisions should be made. Furthermore, we still cannot demote
nodes (not even with force), if other regular nodes exist.

This patch adds a new opcode attribute ‘auto_promote’, and changes the
decision tree as follows:

- if the node will be set to offline or drained or explicitly demoted
  from master candidate, and this parameter is set, then we lock all
  nodes in ExpandNames()
- later, in CheckPrereq(), if the node is
  indeed a master candidate, and the future state (as computed via
  GetMasterCandidateStats with the current node in the exception list)
  has fewer nodes than it should, and we didn't lock all nodes, we exit
  with an exception
- in Exec, if we locked all nodes, we do a AdjustCandidatePool() run, to
  ensure nodes are locked as needed (we do it before updating the node
  to remove a warning, and prevent the situation that if the LU fails
  between these, we're not left with an inconsistent state)

Note that in Exec we run the AdjustCP irrespective of any node state
change (just based on lock status), so we might simplify the CheckPrereq
even more by not checking the future state, basically requiring
auto_promote/lock_all for master candidates, since the case where we
have more than needed master candidates is rarer; OTOH, this would prevent
manual promotion ahead of time of another node, which is why I didn't
choose this way.
Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: Guido Trotter <ultrotter@google.com>
Reviewed-by: Michael Hanselmann <hansmi@google.com>

The return from LUVerifyCluster should be True (or equivalent) for pass,
and False (or equivalent) for fail. The HooksCallBack function uses '1'
(= True) when a hook fails, which is exactly the opposite of what we
want - it will make failed hooks to reset the result to success,
overriding actual failures in cluster verify.
Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: René Nussbaumer <rn@google.com>

We need to manually filter out offline nodes before using
rpc.call_upload_file and rpc.call_write_ssconf_files, since these method
are static (they work without a ConfigWriter instance) and thus do not
know which nodes are offline and which are not).

Note that we add a new ConfigWriter._UnlockedGetOnlineNodeList() method
rather than hardcoding the filtering of online nodes in _WriteConfig.
Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: Michael Hanselmann <hansmi@google.com>

This patch implements all modifications to support per-os-hypervisor
parameters in the framework.
Signed-off-by: René Nussbaumer <rn@google.com>
Signed-off-by: Michael Hanselmann <hansmi@google.com>
Reviewed-by: Michael Hanselmann <hansmi@google.com>

This patch adds validation of new names used, i.e. at cluster init time,
node add time, and instance creation.

For instances, especially when using «--no-name-check» (which skips DNS
checks), we should validate the give name, and also normalize it
(otherwise, we could have two instances named inst1 and Inst1).
Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: Michael Hanselmann <hansmi@google.com>

Rationale: the file-based storage backend can add/remove files under a
certain directory. However, the master node is also controlling the
setting of the file-based root directory, so basically it means we can't
prevent arbitrary modifications by the master of the node's filesystem.

In order to mitigate this for setups where the file-based storage is not
used, we introduce a new setting at ./configure time, that controls the
enable/disable of file-based storage. Since this is not modifiable by
the master (over RPC), it is now possible in this case to prevent
unintended modifications of the node's filesystem from the master.

The new setting is used in bdev.py to not expose the file-based storage
at all, and in cmdlib.py to prevent attempts at creation of such
instances.
Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: Michael Hanselmann <hansmi@google.com>

This passes a full burnin with lots of instances, and should be safe as
we mostly to join a known root (various constants) to a run-time
variable.
Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: Michael Hanselmann <hansmi@google.com>

LUQueryConfigValues supports multiple output fields. If the client asked
for the watcher pause status, it would not get a list, but simply the
value.
Signed-off-by: Michael Hanselmann <hansmi@google.com>
Reviewed-by: Iustin Pop <iustin@google.com>

The first argument to _ErrorIf should always be True in this case.
Signed-off-by: Michael Hanselmann <hansmi@google.com>
Reviewed-by: Iustin Pop <iustin@google.com>

Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: Michael Hanselmann <hansmi@google.com>

Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: Michael Hanselmann <hansmi@google.com>

This is a new mode that request a solution for the evacuation of
multiple nodes. The external script will be fed a list of names, and is
expected to return a list of [instance, new_node(s)] lists, detailing
the evacuation path of each instance.
Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: Michael Hanselmann <hansmi@google.com>

This patch switches the default result key from 'nodes' to 'result'. The
old name is still accepted for backwards-compatiblity, and should be
removed in later versions.
Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: Michael Hanselmann <hansmi@google.com>

Currently the 'name' parameter in the constructor is required (as a
non-keyword argument). Since the (to follow) node evac IAllocator mode
doesn't have 'name' as a valid argument, we're moving this one into the
per-request key, leaving the constructor required arguments more
abstract.
Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: Michael Hanselmann <hansmi@google.com>

This moves the setting of the request member on the in_data, of the
request type, and of the branching basef on request type outside of
individual functions and directly into the constructor.

Since the values we're using externally are identical to the
constants.py values, we're also using those directly.
Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: Michael Hanselmann <hansmi@google.com>

This should have been done in the _ExpandNodeName patch.
Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: Michael Hanselmann <hansmi@google.com>