Currently, both the Python and Haskell code return the internal
instance list unsorted, which means the output can vary depending on
the phase of the moon (well, the Haskell code actually uses internally
a tree, sorted by the instance name, but it's implementation detail).

By adding an explicit sort step, we guarantee both stability of the
output and consistency between the two query paths.

Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: Helga Velroyen <helgav@google.com>

Currently, the 'node' field is declared as a simple config field, so
when only selecting this fields, the runtime gathering is no longer
run and it's presumed that all nodes have a backup. So the output is
not truthful (instead of just listing the nodes with at least one
export, all are listed).

By simply making it a runtime field, filtering works properly and we
get the same output as the masterd code.

Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: Helga Velroyen <helgav@google.com>

And also use it for simple network fields. Currently the Python
library returns jsnull, which is formatted as 'None' by the Python
client, which (IMHO) is not nice, since these fields are strings (some
of them) or "other". An undeclared gateway is easier to diagnose when
properly reported by the server as missing, rather than "existing but
null".

Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: Helga Velroyen <helgav@google.com>

The headers/type/descriptions had some differences from the Python
code, when checked for exact equivalence.

Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: Helga Velroyen <helgav@google.com>

This patch updates the design doc "Design correct reporting of storage
free space". The modifications were chosen to not conflict with any
future changes of Ganeti regarding storage pools.

Signed-off-by: Helga Velroyen <helgav@google.com>
Reviewed-by: Guido Trotter <ultrotter@google.com>

Diskless instances cannot really be reinstalled, not even through RAPI.

Signed-off-by: Bernardo Dal Seno <bdalseno@google.com>
Reviewed-by: Iustin Pop <iustin@google.com>

Signed-off-by: Michele Tartara <mtartara@google.com>
Reviewed-by: Guido Trotter <ultrotter@google.com>

* devel-2.7
  Add QA for instance creation with policy violation
  Add QA for policy-instance interactions
  Add QA for cluster policies
  Unit tests for objects.InstancePolicy + a fix
  Unit tests for objects.FillIPolicy() + small fix
  Fix upgrade of policy in objects.Cluster
  Fix instance policy checks for default back-end parameters
  Fix restoring default instance specs in group policies
  Unit tests for cmdlib._GetUpdatedIPolicy()
  Fix policy check for disk templates
  Fix merge 8e09e801 that resulted in duplicated code
  GanetiRapiClient: fix the no_remember option

Conflicts:
	qa/qa_cluster.py
	qa/qa_instance.py

Conflicts are due to QA config in master using objects instead of
dictionaries. Also updated some new QA code in devel for the same reason.

Signed-off-by: Bernardo Dal Seno <bdalseno@google.com>
Reviewed-by: Guido Trotter <ultrotter@google.com>

The first line of a function should be blank, unless it is able to contain the
whole function.

Signed-off-by: Michele Tartara <mtartara@google.com>
Reviewed-by: Guido Trotter <ultrotter@google.com>

When instance policy is violated, creation fails.

Signed-off-by: Bernardo Dal Seno <bdalseno@google.com>
Reviewed-by: Guido Trotter <ultrotter@google.com>

Violations on policy changes are checked.

Signed-off-by: Bernardo Dal Seno <bdalseno@google.com>
Reviewed-by: Guido Trotter <ultrotter@google.com>

qa_cluster.TestClusterSetISpecs() is exported as it will be used in future
tests.

Signed-off-by: Bernardo Dal Seno <bdalseno@google.com>
Reviewed-by: Guido Trotter <ultrotter@google.com>

Tests for:
  objects.InstancePolicy.CheckParameterSyntax()
  objects.InstancePolicy.CheckDiskTemplates()
  objects.InstancePolicy.CheckISpecSyntax()

Instance policies with an empty disk-template list now are reported.

Signed-off-by: Bernardo Dal Seno <bdalseno@google.com>
Reviewed-by: Guido Trotter <ultrotter@google.com>

IPOLICY_DEFAULTS is now a legal policy (the disk-templates entry was a set
instead of a list, before).

Signed-off-by: Bernardo Dal Seno <bdalseno@google.com>
Reviewed-by: Guido Trotter <ultrotter@google.com>

Unknown elements were silently removed on startup. This means that a
software upgrade could result in lost configuration information if
cfgupgrade wasn't run promptly.

Added unit test for Cluster.UpgradeConfig() to cover this case.

Signed-off-by: Bernardo Dal Seno <bdalseno@google.com>
Reviewed-by: Guido Trotter <ultrotter@google.com>

Policy violations of back-end parameters that used the cluster default
value were not reported in cluster-verify.

Signed-off-by: Bernardo Dal Seno <bdalseno@google.com>
Reviewed-by: Guido Trotter <ultrotter@google.com>

"default" was not accepted as a valid input value for instance specs in
group policies, due to a bug introduced in 2cc673a3 (and released with
2.6.0). Added QA for this and another similar case.

Signed-off-by: Bernardo Dal Seno <bdalseno@google.com>
Reviewed-by: Guido Trotter <ultrotter@google.com>

Not 100% coverage, though.

Signed-off-by: Bernardo Dal Seno <bdalseno@google.com>
Reviewed-by: Guido Trotter <ultrotter@google.com>

Instance disk template is checked against the policy, and diskless
instances aren't checked for the number of disks.

Signed-off-by: Bernardo Dal Seno <bdalseno@google.com>
Reviewed-by: Guido Trotter <ultrotter@google.com>

A fragment in LUInstanceCreate.CheckPrereq() removed in commit ba147ff8 was
reintroduced in merge 8e09e801 due to a change in df28c49b.

Signed-off-by: Bernardo Dal Seno <bdalseno@google.com>
Reviewed-by: Guido Trotter <ultrotter@google.com>

There was a typo which prevented the correct option from being passed to
RAPI

Signed-off-by: Daniel Krambrock <danielk_lists@z9d.de>
Signed-off-by: Guido Trotter <ultrotter@google.com>
Reviewed-by: Guido Trotter <ultrotter@google.com>

* devel-2.7:
  Change hbal behaviour in case of early exit
  Fix build/sphinx_ext.py with tuple defaults for op params
  Fix bug in man build rule
  Fix hscolour style sheet building

Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: Michele Tartara <mtartara@google.com>

Currently, hbal exits with status 1 if early exit is requested, even
when all jobs are successful. This is counter-intuitive behaviour, so
let's fix it (Issue 386).

Note that the man page had conflicting information already, so it's a
good thing to clean this up.

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

The "reason" opcode parameter can now apply to multiple opcodes, not just the
one for rebooting instances.

Signed-off-by: Michele Tartara <mtartara@google.com>
Reviewed-by: Iustin Pop <iustin@google.com>

When an OpCode's parameter has a tuple as default value, this code
will break:

  buf.write("defaults to ``%s``" % default)

The patch fixes this and other potential cases by always passing a
tuple to '%'.

Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: Michele Tartara <mtartara@google.com>

Currently the man page build rule is run without ‘-e’, so the actual
checks in there are just warnings (check-man-warnings,
check-man-dashes).

The patch fixes this and also fixes the man pages which made me see the
problem.

Additionally, check-man-dashes is now verbose, otherwise it's hard to
find out actually where in the page the error is.

Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: Michele Tartara <mtartara@google.com>

Currently we only symlink this into two hard-coded directories, which
breaks the source code viewing for all the others.

The patch replaces this with automatic building; as playing with
relative symlinks is non-trivial, I just generate it many times (the
file is small enough (<250) that it shouldn't matter, size-wise, and
it takes 2ms to do it).

Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: Michele Tartara <mtartara@google.com>

The "reason" attribute will be common to many opcodes.
This patch factors it out so that it will be possible for the next commits
to use it without code duplication.

Signed-off-by: Michele Tartara <mtartara@google.com>
Reviewed-by: Guido Trotter <ultrotter@google.com>

* devel-2.7:
  Update required pylint version
  Constants.hs.in: improve Haddock markup in the template
  convert-constants: generate better Haddock markup
  Expand Haddock to run over test files as well

Conflicts:
        Makefile.am (curl changes and new hs directories)

Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: Michele Tartara <mtartara@google.com>

This should help a lot of users, since pylint 0.21.1 was quite an old
version…

The tree is "clean" w.r.t. the new pylint version.

Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: Michele Tartara <mtartara@google.com>

Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: Michele Tartara <mtartara@google.com>

Currently the Python source constants are used as-is, which is not
nice on the eyes (especially for unqualified names, it was hard to
separate them visually).

Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: Michele Tartara <mtartara@google.com>

This patch does multiple enhancements to the way we build the Haddock
docs, motivated by the fact that if we don't run Haddock over test
files as well, bad formatting can be submitted and will accumulate
over time (without any checks).

The patch does:

- replace manual built ALL_APIDOC_HS_DIRS with automatically built one
  (from HS_DIRS)
- change Haddock so that it runs from the top directory (instead of
  from src/)
- change HsColour target file to be built via bash parameter
  substitution, rather than sed (I don't know how to do it in one go,
  so I use 2 intermediate variables)
- change 'hs-apidoc' target so that it depends on the real target
  file; in case no source file has been modified, running 'make
  hs-apidoc' twice will not result in two runs
- run HsColour/Haddock under en_US.UTF-8 locale, otherwise they can't
  parse correctly the Unicode chars in the test files

Additionally, wrong formatting (oops) in a test file has been
corrected.

Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: Michele Tartara <mtartara@google.com>

* devel-2.7: (23 commits)
  Add export lists for files which didn't have them
  Fix Haskell compatibility tests with disabled file storage
  Fix QA with disabled file storage
  Fix convert-constants handling of booleans
  Fix handling of disabled (shared) file storage
  Fix low verbosity levels in htools
  Fix improperly formatted docstring
  Allow iallocator to work without LVM storage
  Allow rpc.MakeLegacyNodeInfo to parse non-LVM results
  Fix LUTestAllocator with instance alloc
  Fix confd issue regarding --no-lvm-storage
  Fix networks in _PrepareNicModifications()
  Fix sphinx label namespace
  Clarify use of move-instance with self-signed certificates
  Remove early returns in network LUs
  Fix HooksDict() in case of no tags
  Add networks to _AllIDs()
  Fix locking in LUNetworkConnect()
  Fix networks in LUInstanceSetParams()
  Fix another docstring typo
  ...

Conflicts:
        src/Ganeti/Query/Node.hs (trivial; function rename in master
                                  and 'vgs' change in devel-2.7)
        also exported new entity from TestCommon.hs (used in master)

Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: Helga Velroyen <helgav@google.com>

This further improves the comparison for "non-trivial"
numbers. Without this patch, there are still cases where the absolute
error is too big, and we need to switch to relative error.

Concept has been taken from
<http://randomascii.wordpress.com/2012/02/25/comparing-floating-point-numbers-2012-edition/

>.

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

According to the documentation, “This function is almost ten times
faster than rational, but is slightly less accurate. For 94.2% of
numbers, this function and rational give identical results, but for
the remaining 5.8%, this function loses precision around the 15th
decimal place. For 0.001% of numbers, this function will lose
precision at the 13th or 14th decimal place.”. What happens is that
for our tests, it can happen that “Attoparsec.double (show a_double)”
is quite different from “a_double”, such that we have much more than
1e-12 absolute difference.

Since our xm lists should not be too big, I think switching to
rational is better. Next patch also changes the way we compare
doubles, so maybe this patch is not really needed…

Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: Michele Tartara <mtartara@google.com>

Currently, the test uses a frequency of 5 string/5 double/1 list for
generating Arbitrary instances of ListConfig. However, the list case
has simply a "choose (1, 20)" `vectorOf` arbitrary, which means it
could recurse forever.

Manually running only this test gives runtime as such:

- ~100-200ms: very often
- ~1-2s: often
- ~5s: rare
- ~20s: very rare (but I hit this when running < 30 times the test,
  so…)

On average, this makes this test one of the slowest ones, which is
annoying.

By changing to a sized generator, we can control the depth of the
recursion, ensuring that we have a consistent runtime: out of 100
runs, one is 229ms, one is 164ms, the other are 80-120ms.

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

Currently, this tests and its helper function 'isAlmostEqual' uses
plain booleans to signify failures, which means you can't really debug
a failed test. The patch changes the call chain to use annotated
properties all through, which results in messages like:

  Failing almost equal check
  Delta 3.725290298461914e-9 not smaller than 1e-12
  expected: 2.147785408767952e7
   but got: 2.1477854087679517e7

which (IMHO) it's much more readable.

The patch also replaces a 'fail' with 'failTest' in another test.

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

This implements QffHostname and fixes the node listing (as well as
export listing when filtering on node name).

This bug was hidden by the fact that node listing with "gnt-node list
aa" works if you don't have live queries (as it was originally), as
the choosing of wanted nodes out of the config based on short names
works. What didn't work was later post-filtering based on such short
names (kind of duplicate, but that's how the code path is).

By implementing QffHostname, we can have custom equality checks, like
in the Python code. What I don't like is how convoluted the testing on
various left/right combinations is, but I didn't find an easier way.

The included unittest tests the partial filtering behaviour, and fails
if the node name flag is not set to QffHostname.

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

Since we use the primitive string type for group UUIDs, the group
fields have a bug where we pass the group name as filter for node
tests, whereas the nodes themselves use the group UUID. This results
in zero node count, empty node list, and no instances being reported
as assigned to groups.

The patch fixes this and adds a test for the node count. It does some
test generation improvement, and also cleans up whitespace issues in
Test/G/Q/Query.hs (the functions case_queryNode_allfields,
prop_queryGroup_noUnknown and case_queryGroup_allfields are unchanged
but simply have indentation fixed).

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