Currently boto3 is not part of requirements but stevedore always tries
to import it and shows error in case boto3 is missing. This is not
a real error unless users actually enable s3 backends and can be quite
confusing.
This makes the driver code ignore ImportError and actually fail only
if users try to enable s3 backend without installing boto3.
Closes-Bug: #2007924
Change-Id: Ia94dd1d12a3d723f6263bdfb0966d416dfbae1af
This ensures glance_store checks whether the dependent libraries (rbd
and rados) when adding rbd store and disables all rados stores in case
these libraries are not available.
This avoid ugly error (eg. AttributeError) caused by missing libraries.
Change-Id: I673473aa8b540c084ce3a4934e8be7069f13e77f
Change Iec6ac7020f66f9f9728d6cf4452127b8d55a8b2d made cinderclient
optional but it did not add proper validation while initializing
a cinder store. This can cause ugly errors in case users tries to
enable cinder stores without installing cinderclient.
This ensures glance_store checks whether cinderclient is available and
disables all cinder stores in case cinderclient is not available.
Change-Id: Idb8abfdab9639b5b70ddddc3a76b1f1d24dd035d
While creating an image, if we want to extend the volume
(to accomodate the image), we need to first detach the volume,
perform the extend and attach it again. This is inefficient
for backends that support extending attached volumes since we
are performing 3 cinder operations (the attachment call
includes 3 API calls which sum to a total of 5 API calls for 3
operations) instead of directly extending it which requires only
1 API call to cinder.
The support for extending attached volumes was added in cinder
microversion 3.42.
This patch adds a new boolean config option
``cinder_do_extend_attached`` which allows operators to specify
if the cinder backend they are using supports extending
attached (in-use) volumes. By default this will be ``false``.
Based on the parameter, we will perform the extend operation,
online (if cinder_do_extend_attached is true) and offline otherwise.
Spec: https://review.opendev.org/c/openstack/glance-specs/+/868901
Depends-On: https://review.opendev.org/c/openstack/glance/+/869021
Depends-On: https://review.opendev.org/c/openstack/cinder/+/869051
Change-Id: I5e70824e9abc5277ea25ba95704b358fe3686037
This is an effort to decouple some of the cinder backend specific
code (like nfs, scaleio) from the generic logic.
The purpose is to make the code modular and any change
for a particular cinder backend should not affect the code path
of other backends thereby reducing regression.
This is also required for another use case of supporting extend
of attached volumes added in [1].
Following are the major changes done in this patch:
1) Move cinder store to a new directory 'cinder' and rename
'cinder.py' to 'store.py' (similar to swift)
2) Create new files for nfs and scaleio backends for moving code
specific to these backends into their own separate file.
This also fixes one bug when using sparse files in nfs and
we wait for file size to be equal to volume size (initially done
for scaleio/powerflex backend) but this will never happen for
nfs sparse files. See bug: 2000584
3) Move cinder tests to 'tests/unit/cinder' directory and add tests
for base, nfs and scaleio files.
4) Modify/fix existing tests
Closes-Bug: #2000584
[1] https://review.opendev.org/c/openstack/glance_store/+/868742
Depends-On: https://review.opendev.org/c/openstack/glance/+/869021
Change-Id: I26c272b6c503e98fbbafca411d3eec47283bd6fc
Add an option to let the operator choose a proper region_name.
Some operator are not using S3 from amazon but from other provider.
In such case, the get_s3_location cannot give good region_name.
Adding an option from config seems a better option that relying on
hardcoded stuff in the code.
Signed-off-by: Arnaud Morin <arnaud.morin@ovhcloud.com>
Change-Id: I71352e75f6415494709d8fb62c8b13f3ea7c6016
Change I59b1579dc9877668b82d4195431c1 added tests that mock an
input buffer, but the mocked return_value for __len__ exceeds the
allowable maximum value for an index in the CPython implementation
on 32-bit systems. Fix this by using half the value and allowing
the buffer to be called twice during the tests.
Change-Id: Iaba20c01fb4d994cd738a570197f8f6f04f22606
Closes-bug: #1991406
The *_domain_id parmaeters should not have any default. Otherwise
keystoneauth ignores the *_domain_name parameters and it requires
only *_domain_id parameters are used.
Closes-Bug: #1620999
Change-Id: I1f8c9184761313f9fc5fda2f257e52233e0196d1
when fetching images with http driver, the redirect URL can have
mandatory query in it, which must be kept intact to successfully
fetch the image.
Change-Id: I2a9d4d026b935ea6c5e5a3a46c86f70ce1e39ae7
Closes-Bug: #1633860
When debugging issues related to glance cinder store, there are
several calls to cinder and it becomes hard to determine which step
we are currently executing without going through cinder logs.
This patch adds some useful debug logs for the new attachment's code
to understand which stage of attachment we are on.
Change-Id: I491b7292a511c47c1d6148dab69ae04269e50c85
As a new feature relevant for the Cinder store, os-brick now supports
setting the location of file locks in a different location from the
locks of the service.
The functionality is intended for HCI deployments and hosts that are
running Cinder and Glance using Cinder backend. In those scenarios the
service can use a service specific location for its file locks while
only sharing the location of os-brick with the other services.
To leverage this functionality the new os-brick code is needed and
method ``os_brick.setup`` needs to be called once the service
configuration options have been loaded.
The default value of the os-brick ``lock_path`` is the one set in
``oslo_concurrency``.
This patch adds support for this new feature in a backward compatible
way so code works even if using an os-brick version that doesn't have
this feature. That is the case for many CI jobs that install os-brick
from PyPi instead of the code present in master.
Change-Id: Ib11d63e3c388e12f145f40247d17030a566b8c5e
Python 2 has been deprecated for almost two years, and has not been
guaranteed to work with glance_store for a while. This patch removes all
traces of six, unicode strings and Python 2 tweaks.
Co-Authored-By: Cyril Roelandt <cyril@redhat.com>
Change-Id: Ifa78924d7ecf4f2d9a54c677888ab2926530c487
There is a syntax error in the exception logging when attaching a
volume fails in attachment state manager. This patch corrects it
and adds a test to guard against similar changes in future.
Closes-Bug: #1970698
Change-Id: I43c407046a49bb37631113e2ea65d05450f9365d
When we try to simultaneously attach same volume multiple times
(multiattach), there are multiple attachments created, suppose
attachA and attachB. If attachA marks the volume "reserved" then
attachB can't proceed until the volume is in "in-use" or "available"
state. We retry until we reach any of these states for which we use
the retrying library.
The retrying library defaults to 1 second retry[1] (5 times) which causes
the original failure as the volume takes time to transition between
"reserved" -> "in-use" state. This patch corrects it by adding an
exponential retry time and max exponential retry i.e.
1: 2 ** 1 = 2 seconds
2: 2 ** 2 = 4 seconds
3: 2 ** 3 = 8 seconds
4: 2 ** 4 = 16 seconds (but max wait time is 10 seconds)
5: 2 ** 5 = 32 seconds (but max wait time is 10 seconds)
[1] https://github.com/rholder/retrying/blob/master/retrying.py#L84
Closes-Bug: #1969373
Change-Id: I0094b044085d7f92b07ea86236de3b6efd7d67ea
This patch adds test coverage of code paths like extend volume,
exception blocks, when image_size is zero, volume delete call etc
Change-Id: I59b1579dc9877668b82d4195431c14cc41cfe892
This patch adds coverage for the various exceptions handled and
raised in get, get_size and delete methods.
It also corrects the behavior of _test_cinder_get_size method where
client.volumes returned dictionary instead of a MagicMock and it
worked due to the existing method "get" in dictionary which has
same name as "get" method in cinderclient. The dictionary object
is replaced with the appropriate MagicMock object in this patch.
Change-Id: If63a6c810b5aab992e54857bc81f5052c2c593c4
This patch adds coverage for some parts in get_cinderclient
method like ``cinder_endpoint_template`` option and keystone
exception case when catalog info for cinder could not be found.
Also it adds tests for complete coverage of _check_context
method.
Change-Id: I18b1e5e8fd818824a2dda2ad14d9456190fe9ff4
The cinder volume create call was moved to the cinder_utils
module to keep the driver code cleaner with commit[1].
A decorator was also added to handle NotFound exception which
made the existing try/except block redundant.
Also the current try/except block has misleading messages:
1) "Invalid volume type configured"
2) "Failed to create image-volume due to invalid
`cinder_volume_type` configured."
A volume create could fail because of many reasons and "invalid
volume type" is just one of them. These messages will confuse
the operators even if the configured volume type is correct.
This patch removes them since the NotFound exception is already
handled at[2] and the messages are misleading.
[1] 1178f113c4
[2] c0c4969a4a/glance_store/common/cinder_utils.py (L37)
Change-Id: I39c9caec85f574fc717e60f11e5681860a20ed30
This patch adds code coverage (UTs) for methods in StoreLocation class:
1) process_specs
2) get_uri
3) parse_uri
It also adds a missing method (_set_url_prefix) coverage in
Multistore cinder tests (test_multistore_cinder).
Change-Id: I8ced5af11669fb131c665ce53be57143f2c7b518
Add test coverage for two config options used to create a keystone
session object for cinderclient namely ``cinder_api_insecure`` and
``cinder_ca_certificates_file``.
Change-Id: Idb6db9f87ef7290441dfe60ca2eb9a59c9ee7bd3
six has been used to maintain compatibility between python2 and
python3 code. Since current openstack development branches doesn't
support python2, we don't need the six library and can use python3
code directly.
This patch removes six usage from cinder store tests.
Change-Id: I4deb193b44394ae9f99f57e31fb81fe04394e247
This patch aims at a refactoring effort that would remove
duplicate tests (current and future) by moving them into a
common base class which is called via both single and multi store
test modules with their specific configurations.
This has a lot of benefits:
1) Removes duplicate code
2) Makes addition of new tests easier and cleaner
3) Ensuring a new method/code path added is tested in both
single and multi store configurations
4) Fixing issues detected while refactoring methods
(Eg: tests for add method in test_multistore_cinder were not
passing the hashing_algo parameter which is currently handled
by the backward compat code (back_compat_add decorator) but
those tests will break when we remove backward compatibility)
Change-Id: I12569af5623f1cd7803c00a6c3b9eb211f15b6fd
This patch aims at a refactoring effort that would remove
duplicate tests (current and future) by moving them into a
common base class which is called via both single and multi store
test modules with their specific configurations.
This has a lot of benefits:
1) Removes duplicate code
2) Makes addition of new tests easier and cleaner
3) Ensuring a new method/code path added is tested in both
single and multi store configurations
4) Fixing issues detected while refactoring methods
(Eg: fake_chown accepted an optional backend parameter which
does not match the signature of original method temporary_chown)
Change-Id: Iaacb3e117cac2a661750bdb21bd0a7496078ea26
FakeObject is used to set dynamic attributes to a fake class which
works in most cases but fails when an attribute is accessed which
is not defined/passed explicitly.
This causes failure of tests whenever a new attribute is accessed
and it's hard to maintain the set of attributes for every object
used.
MagicMock provides the same functionality with additional features
and handles the case effectively when an attribute is accessed
which is not explicitly defined in the tests.
This change will help safeguard against cases seen earlier in[1].
This patch replaces FakeObject with MagicMock in test_multistore_cinder.py
[1] https://review.opendev.org/c/openstack/glance/+/805974
Change-Id: I193bb03d174e6c010132702dedef1dea05f40e41
FakeObject is used to set dynamic attributes to a fake class which
works in most cases but fails when an attribute is accessed which
is not defined/passed explicitly.
This causes failure of tests whenever a new attribute is accessed
and it's hard to maintain the set of attributes for every object
used.
MagicMock provides the same functionality with additional features
and handles the case effectively when an attribute is accessed
which is not explicitly defined in the tests.
This change will help safeguard against cases seen earlier in[1].
This patch replaces FakeObject with MagicMock in test_cinder_store.py
[1] https://review.opendev.org/c/openstack/glance/+/805974
Change-Id: Ifd0905b139cc90c2bd7444bc9e7638f3af6879a7
When we have an image with size > 1 GB, we follow the following steps
to accomodate the image:
1) Detach the volume
2) extend the volume
3) Attach the volume
4) Open the volume device as a file and resume writing the image
Sometimes due to several reasons (mostly network related), the
size of the device file could mismatch with the actual volume size
(or the backend LUN size). This can happen if the extend was performed
(i.e. the control path) but it takes the time to reflect that into
the mapped device (i.e. the data path). This mismatch can cause the
issue "IOError: [Errno 28] No space left on device".
To avoid this scenario, we check if the device size is less than the
volume size, we wait for the extended LUN to show up in mapped device
and then continue the image writing operation.
Closes-Bug: #1959913
Change-Id: I206580f6be615ebc5e15b546b9c23728d4116a5d
Cinder's attachment_complete API is called to mark the volume from
"attaching" state to "in-use" state stating that the volume is
ready to use.
In the current code, attachment_complete is called before the
brick connection is complete. This should be done after the
connection to brick is successful and we have a volume path
to write the image into. This patch corrects the behavior.
The wrong sequence of execution won't have any issues functionally
since we open the volume after all steps are completed but flow
wise it is incorrect.
This change won't have any end user impact.
Change-Id: Ia9652a4ff6d7efbabb58511f0ce93a87b3a4dfa8
While fetching connector information the cinder driver of
glance-store is passing 'hostname' as IP address whereas it
actually needs IP address.
This path fetches the IPV4 or IPV6 address from the available
'hostname' and passes it to os-brick.
This worked for other cinder backends which don't rely specifically
on IP address for initializing connection like fibre channel, nfs etc
and even for some iscsi backends since, in some environments,
the hostname is same as the ip address. This is not the case always
and we should pass the correct IPv4/IPv6 address of the host to
avoid this issue.
Closes-Bug: #1955668
Change-Id: Ic79815972e654a8bfe2775f57c68cfa0bf115e2f
Recently cinder has utilized Ceph clone v2 support for its
RBD backend, since then if you attempt to delete an image from
glance that has a dependent volume, all future uses of that
image will fail in error state. Despite the fact that image
itself is still inside of Ceph/Glance. This issue is reproducible
if you are using ceph client version greater than 'luminous'
To resolve this issue glance RBD driver now checks whether snapshot
of original image has any external references before deleting/removing
it's snapshot and returns 409 Conflict response if it has any.
NOTE: To check this dependency glance osd needs 'read' access to
cinder and nova side RBD pool.
Closes-Bug: #1954883
Depends-on: https://review.opendev.org/c/openstack/devstack-plugin-ceph/+/819476
Change-Id: If30b7bd7acac148b6f89ce46abbe128c678c90e7
Zuul