Signed-off-by: René Nussbaumer <rn@google.com>
Reviewed-by: Iustin Pop <iustin@google.com>

Signed-off-by: René Nussbaumer <rn@google.com>
Reviewed-by: Iustin Pop <iustin@google.com>

Learning, learning. It turns out that for objects with even just five
fields, generating them and filtering can be costly, especially when
the filtering is nested (as it was in this case, arbitrary1 →
arbitrary2 + filter on arbitrary1 → arbitrary plus filter on
aribitrary2).

Changing the code to generate directly the good values speeds things
up significantly. Additionally, we also generate correctly unique
lists of disk templates (again, this should be a set, not a list,
but…).

Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: René Nussbaumer <rn@google.com>

It could happen that the node is generated with 0 disk available, in
which case the operation will fail due to disk issues, not CPU. Fix
this by using the recently-introduced genOnlineNode helper.

Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: René Nussbaumer <rn@google.com>

As a followup from "Remove extraneous check in policy creation", there
are more places where we build an ipolicy, and then manually check for
its validity. This is very bad style, as it duplicates the
verification code across many places.

This patch removes all such explicit checks (except for one in
cmdlib.py which is correct), and instead does a bit more validation in
the builder functions or in the actual dedicated verification
functions. It also fixes cluster init which used the wrong,
non-completed ipolicy (this was not detected before as we did call
check on it, but otherwise we ignored it), and fixes a too-strong
assert (due to the call chain, we first create the ipolicy from
cmdline params, and only then we fill it).

Finally, it removes an extraneous logging.info which I forgot from
debugging.

Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: René Nussbaumer <rn@google.com>

Another "good" seed value manages to generate nodes with identical
names, which of course makes the sets be of smaller value than the
lists. Avoid this issue by generating unique node names.

Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: René Nussbaumer <rn@google.com>

While wondering why Cluster.tryReloc is not unit-tested, I realised
that this is no longer uses. Apparently, last summer's new IAllocator
modes and related changes have silently deprecated this function.

Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: René Nussbaumer <rn@google.com>

This patch fixes two ugly but non-critical issues:

- the getFQDN wrongly generated only the first component as
  limited-length name, the rest could be any length, and that makes
  the generated name bad for display purposes
- the random generator had real issues in finding an offline instance
  with the right properties when starting from certain seeds (found
  accidentally); this meant the test took 1 minute, instead of ~300ms,
  for the node_addSecOffline

Furthermore, for the same evil seed, the Container_findByName test
wasn't able to generate enough good data, so we change its
implementation to directly generate an uniques list, instead of
generating random lists and filtering the unique ones.

Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: René Nussbaumer <rn@google.com>

Currently they only check for change secondary, via iallocator's
relocate method. We add two separate unittests that checks all the
node evacuate modes (ChangeAll, ChangePrimary) and also the group
change request. All of these are done only for DRBD instances.

The patch also enhances an error message in Cluster.findBestAllocGroup
for invalid input data (guess how this was detected…).

Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: René Nussbaumer <rn@google.com>

… instead of always using "printTestCase … False".

Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: René Nussbaumer <rn@google.com>

Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: René Nussbaumer <rn@google.com>

Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: René Nussbaumer <rn@google.com>

This adds a new optional (so we're backwards-compat) section to the
cluster data so that we can reliably save/restore the policy data.

Note that different from the other sections, we save all policies
together and separately from the group data. I thought this is cleaner
than extending adding the policy data near each group, and a new
section just for the cluster policy.

This change also allows simplification of the unittest for cluster
create/serialize, although we would need a more complete test (right
now we only dump/load the defIPolicy or nullIPolicy).

Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: René Nussbaumer <rn@google.com>

This is unfortunately a complex test that checks if:

- starting from a empty cluster
- allocating a few instances on it
- serialising it via the Text backend
- loading it back into internal data structures

we end up with the same thing. More precisely, it checks that the
cluster after dump/load is the same as after allocation.

While this is in the Text unittests, as the docstring saying, this
check more stuff than just the dump/load. It increases the overall
coverage by about 3%.

There were a few changes needed for this unittest: adding some Eq
instances for some data types, and fixing a cosmetic bug in
Cluster.iterateAlloc, where the instance indices were growing by 2
instead of by 1 in each round.

Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: René Nussbaumer <rn@google.com>

Nothing breaks otherwise in the cluster routines, but serialisation
to/from text gets confused otherwise.

Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: René Nussbaumer <rn@google.com>

This abstracts the concept of "give an sane, online node" in a single
place, rather than spread around the tests. We also change the
signature for isNodeBig for easier usage.

Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: René Nussbaumer <rn@google.com>

The addition of the offline state for instances has introduced a few
bugs related to their handling. The current code adds a helper
function for making the check whether an instances uses reserved
memory on the secondary (and uses it instead of just autoBalance),
and furthermore uses consistently the state for primary node
operations.

Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: René Nussbaumer <rn@google.com>

Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: René Nussbaumer <rn@google.com>

Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: René Nussbaumer <rn@google.com>

Currently, in the human-readable mode, hspace doesn't show any
messages when no instances can be allocated in tiered-alloc mode. This
is confusing, so let's add an explicit message.

Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: René Nussbaumer <rn@google.com>

This currently checks the minimum/maximum sizes and the disk
template. Unittests will come later.

Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: René Nussbaumer <rn@google.com>

In order to be able to use the group policy during allocation, we have
a few choices:

- pass the group policy as a separate argument to all involved
  functions, or
- add the policy as an attribute to nodes, or
- add the group (not only the group index) as an attribute to nodes

The first and last one would change more than wanted for this patch,
so I went with the second one; it will allow easily checking the
target policy at runtime (based on the instance's primary node), with
the only downside that without a node, we can't check the policy.

Note that the first option also would complicate the allocation
routines, which examine (big) lists of node/node pairs, so adding the
group to this would complicate things even more.

Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: René Nussbaumer <rn@google.com>

During loading, we can eliminate some intermediate data
structures. This only matters for big clusters, and even there it's
less than 1% less memory allocated, but still, it's a no-brainer.

Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: René Nussbaumer <rn@google.com>

This is not in the hot-path, but still, no sense in doing O(n) instead
of O(1).

Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: René Nussbaumer <rn@google.com>

This is straightforward; as before, so far the text and simu backends
have no way to use a non-default ipolicy.

Furthermore, we do some unit-test preparation for future changes
(these seem a no-op for now, but they are needed for later patches).

Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: René Nussbaumer <rn@google.com>

This is needed for the next patch.

Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: René Nussbaumer <rn@google.com>

The default disk template is the first enabled template.

Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: René Nussbaumer <rn@google.com>

The default value is badly defined (hardcoded defaults)…

Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: René Nussbaumer <rn@google.com>

This is needed since we'll need the DiskTemplate definition in the
IPolicy one.

Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: René Nussbaumer <rn@google.com>

This is a bit more complex patch, as it requires changing the
assumption that all keys in the policy dict points to values that are
themselves dicts. Right now we introduce an assumption that any
non-dicts are lists, we'll see in the future if this holds or whether
we need more complex type checking (manual, yay Python).

The patch also does some trivial style changes.

Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: René Nussbaumer <rn@google.com>

The values are already checked in CreateIPolicy, no need to manually
check them again.

Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: René Nussbaumer <rn@google.com>

Currently these are defined twice, instead of a single time in
cli.py. Also adds the new disk_templates option to the common block,
even though it's not yet used.

Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: René Nussbaumer <rn@google.com>

Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: René Nussbaumer <rn@google.com>

This simply splits the value in the option parser, instead of needing
to do it in the client code.

Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: René Nussbaumer <rn@google.com>

This code raises a configuration error, but we need to transform it
into a prereq error (or possibly exec error), depending on when we
call this function.

Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: René Nussbaumer <rn@google.com>

This changes from the current hardcoded defaults to the cluster
policy. The command line options now override the defaults from the
cluster, and the tiered spec mode is always enabled.

Also fixes a tiny typo in the man page (together with the man page
updates).

Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: René Nussbaumer <rn@google.com>

The rspec type is simpler than the ispec one; most likely it should be
deprecated later.

Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: René Nussbaumer <rn@google.com>

Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: René Nussbaumer <rn@google.com>

This requires changing from querying the /tags resource to the /info
resource.

Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: René Nussbaumer <rn@google.com>

Recent changes to the instance beparams have replaced memory with
maxmem in Rapi bulk queries. Until this is either reverted (for
backwards compat) or we decide to go ahead with only maxmem, we change
the backend to read both; this only affects the "instance down" code
path (where ``oper_ram`` is missing).

Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: René Nussbaumer <rn@google.com>