| Age | Commit message (Collapse) | Author |
|---|
|
DataSourceAzure previously writes the preprovisioning
reported ready marker file before it goes through the
report ready workflow. On certain VM instances, the
marker file is successfully written but then reporting
ready fails.
Upon rare VM reboots by the platform, cloud-init sees
that the report ready marker file already exists.
The existence of this marker file tells cloud-init
not to report ready again (because it mistakenly
assumes that it already reported ready in
preprovisioning).
In this scenario, cloud-init instead erroneously
takes the reprovisioning workflow instead of
reporting ready again.
|
|
This moves logging into `report_diagnostic_event`, to clean up its callsites.
|
|
enumeration of physical network devices (#591)
|
|
fails (#549)
Azure datasource's `parse_network_config` throws a fatal uncaught exception when an exception is raised during generation of network config from IMDS metadata. This happens when IMDS metadata is invalid/corrupted (such as when it is missing network or interface metadata). This causes the rest of provisioning to fail.
This changes `parse_network_config` to be a non-fatal implementation. Additionally, when generating network config from IMDS metadata fails, fall back on generating fallback network config (`_generate_network_config_from_fallback_config`).
This also changes fallback network config generation (`_generate_network_config_from_fallback_config`) to blacklist an additional driver: `mlx5_core`.
|
|
* pull ssh keys from imds first and fall back to ovf if unavailable
* refactor log and diagnostic messages
* refactor the OpenSSLManager instantiation and certificate usage
* fix unit test where exception was being silenced for generate cert
* fix tests now that certificate is not always generated
* add documentation for ssh key retrieval
* add ability to check if http client has security enabled
* refactor certificate logic to GoalState
|
|
Changes:
tox: bump the pylint version to 2.6.0 in the default run
Fix pylint 2.6.0 W0707 warnings (raise-missing-from)
|
|
This fixes a long delay during boot of some instances. For Azure instance types using SR-IOV via the Hyper-V netvsc network driver, two network interfaces are created that share the same MAC, but only the virtual device should be configured and used. Updating the netplan configuration to filter on the hv_netvsc driver prevents netplan from trying to figure both devices.
LP: #1830740
|
|
Push the cloud-init.log file (Up to 500KB at once) to the KVP before reporting ready to the Azure platform.
Based on the analysis done on a large sample of cloud-init.log files, Here's the statistics collected on the log file size:
P50 P90 P95 P99 P99.9 P99.99
137K 423K 537K 3.5MB 6MB 16MB
This change limits the size of cloud-init.log file data that gets dumped to KVP to 500KB. So for ~95% of the cases, the whole log file will be dumped and for the remaining ~5%, we will get the last 500KB of the cloud-init.log file.
To asses the performance of the 500KB limit, 250 VM were deployed with a 500KB cloud-init.log file and the time taken to compress, encode and dump the entries to KVP was measured. Here's the time in milliseconds percentiles:
P50 P99 P999
75.705 232.701 1169.636
Another 250 VMs were deployed with this logic dumping their normal cloud-init.log file to KVP, the same timing was measured as above. Here's the time in milliseconds percentiles:
P50 P99 P999
1.88 5.277 6.992
Added excluded_handlers to the report_event function to be able to opt-out from reporting the events of the compressed cloud-init.log file to the cloud-init.log file.
The KVP break_down logic had a bug, where it will reuse the same key for all the split chunks of KVP which results in overwriting the split KVPs by the last one when consumed by Hyper-V. I added the split chunk index as a differentiator to the KVP key.
The Hyper-V consumes the KVPs from the KVP file as chunks whose key is 512KB and value is 2048KB but the Azure platform expects the value to be 1024KB, thus I introduced the Azure value limit.
|
|
DataSourceAzure: Gracefully handle the case of set hostname failure during provisioning
|
|
JSONDecodeError is only available in Python 3.5+. When it isn't available (i.e. on Python 3.4, which cloud-init still supports) use the more generic ValueError.
|
|
* cloudinit: remove global disable of pylint W0107 and fix errors
This includes removing a test class which contained no tests but wasn't
detected as empty because of an errant pass statement.
* .pylintrc: update disable comment to match arguments
|
|
This includes a fix to a test that had a string concatenation issue, and
so was only testing a prefix of what was intended.
|
|
This introduces a way to log the dhclient error stream, and uses it for the Azure datasource (where we have a specific requirement for this data to be logged).
|
|
|
|
This was painful, but it finishes a TODO from cloudinit/subp.py.
It moves the following from util to subp:
ProcessExecutionError
subp
which
target_path
I moved subp_blob_in_tempfile into cc_chef, which is its only caller.
That saved us from having to deal with it using write_file
and temp_utils from subp (which does not import any cloudinit things now).
It is arguable that 'target_path' could be moved to a 'path_utils' or
something, but in order to use it from subp and also from utils,
we had to get it out of utils.
|
|
Improving the debugability of this code path by logging the thrown exception details for the non 404 exceptions.
Retry IMDS on HTTP Error 404 and 410, re-run DHCP on other exceptions.
|
|
* cc_ssh: fix capitalisation of SSH
* doc: fix capitalisation of SSH
* cc_keys_to_console: fix capitalisation of SSH
* ssh_util: fix capitalisation of SSH
* DataSourceIBMCloud: fix capitalisation of SSH
* DataSourceAzure: fix capitalisation of SSH
* cs_utils: fix capitalisation of SSH
* distros/__init__: fix capitalisation of SSH
* cc_set_passwords: fix capitalisation of SSH
* cc_ssh_import_id: fix capitalisation of SSH
* cc_users_groups: fix capitalisation of SSH
* cc_ssh_authkey_fingerprints: fix capitalisation of SSH
|
|
Azure stores the instance ID with an incorrect byte ordering for the
first three hyphen delimited parts. This results in invalid
is_new_instance checks forcing Azure datasource to recrawl the metadata
service.
When persisting instance-id from the metadata service, swap the
instance-id string byte order such that it is consistent with
that returned by dmi information. Check whether the instance-id
string is a byte-swapped match when determining correctly whether
the Azure platform instance-id has actually changed.
|
|
Azure's Instance Metadata Service (IMDS) reports multiple IPv6
addresses, via the http://169.254.169.254/metadata/instance/network
route. Any additional values after the first in 'ipAddresses' under the
'ipv6' interface key are extracted and configured as static IPs on
the interface.
|
|
Network v2 configuration for Azure will set both dhcp4 and
dhcp6 to False by default.
When IPv6 privateIpAddresses are present for an interface in Azure's
Instance Metadata Service (IMDS), set dhcp6: True and provide a
route-metric value that will match the corresponding dhcp4 route-metric.
The route-metric value will increase by 100 for each additional
interface present to ensure the primary interface has a route to IMDS.
Also fix dhcp route-metric rendering for eni and sysconfig distros.
LP: #1850308
|
|
After initial boot ovf-env.xml is copied to agent dir
(/var/lib/waagent/) with REDACTED password.
On subsequent boots DataSourceAzure loads with a configuration where the
user specified in /var/lib/waagent/ovf-env.xml is locked.
If instance id changes, cc_users_groups action will lock the user.
Fix this behavior by not locking the user if its password is REDACTED.
LP: #1849677
|
|
Collect and record the following information through KVP:
+ timestamps related to kernel initialization and systemd activation
of cloud-init services
+ system information including cloud-init version, kernel version,
distro version, and python version
+ diagnostic events for the most common provisioning error issues
such as empty dhcp lease, corrupted ovf-env.xml, etc.
+ increasing the log frequency of polling IMDS during reprovision.
|
|
The function generate_fallback_config is used by Azure by default when
not consuming IMDS configuration data. This function is also used by any
datasource which does not implement it's own network config. This simple
fallback configuration sets up dhcp on the most likely NIC. It will now
emit network v2 instead of network v1.
This is a step toward moving all components talking in v2 and allows us
to avoid costly conversions between v1 and v2 for newer distributions
which rely on netplan.
|
|
This allows cloud-init query region to show valid region data for Azure
|
|
If the IMDS primary server is not available, falling back to the
secondary server takes about 1s. The net result is that the
expected E2E time is slightly more than 1s. This change increases
the timeout to 2s to prevent the infinite loop of timeouts.
|
|
On FreeBSD, mount_cd9660 does not accept the sync option that is enabled
by default. In addition, the sync is only useful with the `rw` mode.
However the `rw` mode was never used.
This patch removes the `rw` and `sync` parameter of `mount_cb` to
simplify the code base and resolve the FreeBSD issue.
LP: #1645824
|
|
When the Azure datasource persists all of its metadata to the
instance directory, it deliberately sets the self.network_config
value to be the sources.UNSET value. The goal is to ensure that
each time the system boots, fresh network configuration data is
fetched from the cloud platform so that any control plane changes
will take effect. When a VM is first created, there's no pickled
instance to restore, so self._network_config is None, resulting
in self.network_config() properly building a new config. Azure
suffered from LP: #1801364 which prevented ds from being stored
in obj.pkl in the instance directory, so subsequent reboots always
regenerated their network configuration.
Commit 0dc3a77f41f4544e4cb5a41637af7693410d4cdf introduced a
new bug in which self.network_config() assumed the
self._network_config value was either None or trustable; when
the config was unpickled, that value was _unset, thus breaking
the assumption.
LP: #1823084
|
|
Create an Azure logging decorator and use additional ReportEventStack
context managers to provide additional logging details.
|
|
The Azure platform surfaces random bytes into /sys via Hyper-V.
Python 2.7 json.dump() raises an exception if asked to convert
a str with non-character content, and python 3.0 json.dump()
won't serialize a "bytes" value. As a result, c-i instance
data is often not written by Azure, making reboots slower (c-i
has to repeat work).
The random data is base64-encoded and then decoded into a string
(str or unicode depending on the version of Python in use). The
base64 string has just as many bits of entropy, so we're not
throwing away useful "information", but we can be certain
json.dump() will correctly serialize the bits.
|
|
The Azure data source is expected to expose a list of
ssh keys for the user-to-be-provisioned in the crawled
metadata. When configured to use the __builtin__ agent
this list is built by the WALinuxAgentShim. The shim
retrieves the full set of certificates and public keys
exposed to the VM from the wireserver, extracts any
ssh keys it can, and returns that list.
This fix reduces that list of ssh keys to just the
ones whose fingerprints appear in the "administrative
user" section of the ovf-env.xml file. The Azure
control plane exposes other ssh keys to the VM for
other reasons, but those should not be added to the
authorized_keys file for the provisioned user.
|
|
Testing startup of large numbers of VMs (of varying distros) in Azure
shows that 3 retries results in a small percentage of failed VMs.
Increasing that by a few dramatically decreases the occurrence of
provisioning timeout errors. The initial choice of "3 retries" was
uninformed by heavy testing. Also, the alternate provisioning
mechanism for Azure (waagent) retries the Wireserver crawl without
limit. 10 retries seems a more reasonable choice.
|
|
The tip-pylint tox target correctly reported the invalid use of
string formatting. The change here is to:
a.) Fix the error that was caught.
b.) move to pylint 2.2.2 for the default 'pylint' target.
|
|
Replace Azure pre-provision polling on IMDS with a blocking call
which watches for netlink link state change messages. The media
change event happens when a pre-provisioned VM has been activated
and is connected to the users virtual network and cloud-init can
then resume operation to complete image instantiation.
|
|
Check the appropriate variables based on code review. Correcting what
seems to be a copy/paste mistake for the error handling from a few lines
above.
|
|
Upon URL timeout, _poll_imds is expected to re-dhcp to get updated
IP configuration. We don't want to indefinitely retry because the
instance likely has invalid IP configuration.
LP: #1803598
|
|
There is an infrequent race when the booting instance can hit the IMDS
service before it is fully available. This results in a
requests.ConnectTimeout being raised.
Azure's retry_callback logic now retries on either 404s or Timeouts.
LP:1800223
|
|
If Azure detects an ntfs filesystem type during mount attempt, it should
still report the resource device as reformattable. There are slight
differences in error message format on RedHat and SuSE. This patch
simplifies the expected error match to work on both distributions.
LP: #1799338
|
|
In commitish 9073951 azure datasource tried to leverage stale DHCP
information obtained from EphemeralDHCPv4 context manager to report
updated provisioning status to the fabric earlier in the boot process.
Unfortunately the stale ephemeral network configuration had already been
torn down in preparation to bring up IMDS network config so the report
attempt failed on timeout.
This branch introduces obtain_lease and clean_network public methods on
EphemeralDHCPv4 to allow for setup and teardown of ephemeral network
configuration without using a context manager. Azure datasource now uses
this to persist ephemeral network configuration across multiple contexts
during provisioning to avoid multiple DHCP roundtrips.
|
|
There was a typo in the seeded filename s/azure-hotplug/hotplug-azure/.
|
|
When reusing a preprovisioned VM, report ready to Azure fabric as soon as
we get the reprovision data and the goal state so that we are not delayed
by the cloud-init stage switch, saving 2-3 seconds. Also reduce logging
when polling IMDS for reprovision data.
LP: #1799594
|
|
Azure generates network configuration from the IMDS service and removes
any preexisting hotplug network scripts which exist in Azure cloud images.
Add a datasource configuration option which allows for writing a default
network configuration which sets up dhcp on eth0 and leave the hotplug
handling to the cloud-image scripts.
To disable network-config from Azure IMDS, add the following to
/etc/cloud/cloud.cfg.d/99-azure-no-imds-network.cfg:
datasource:
Azure:
apply_network_config: False
LP: #1798424
|
|
Add the following instance-data.json standardized keys:
* v1._beta_keys: List any v1 keys in beta development,
e.g. ['subplatform'].
* v1.public_ssh_keys: List of any cloud-provided ssh keys for the
instance.
* v1.platform: String representing the cloud platform api supporting the
datasource. For example: 'ec2' for aws, aliyun and brightbox cloud
names.
* v1.subplatform: String with more details about the source of the
metadata consumed. For example, metadata uri, config drive device path
or seed directory.
To support the new platform and subplatform standardized instance-data,
DataSource and its subclasses grew platform and subplatform attributes.
The platform attribute defaults to the lowercase string datasource name at
self.dsname. This method is overridden in NoCloud, Ec2 and ConfigDrive
datasources.
The subplatform attribute calls a _get_subplatform method which will
return a string containing a simple slug for subplatform type such as
metadata, seed-dir or config-drive followed by a detailed uri, device or
directory path where the datasource consumed its configuration.
As part of this work, DatasourceEC2 methods _get_data and _crawl_metadata
have been refactored for a few reasons:
- crawl_metadata is now a read-only operation, persisting no attributes on
the datasource instance and returns a dictionary of consumed metadata.
- crawl_metadata now closely represents the raw stucture of the ec2
metadata consumed, so that end-users can leverage public ec2 metadata
documentation where possible.
- crawl_metadata adds a '_metadata_api_version' key to the crawled
ds.metadata to advertise what version of EC2's api was consumed by
cloud-init.
- _get_data now does all the processing of crawl_metadata and saves
datasource instance attributes userdata_raw, metadata etc.
Additional drive-bys:
* unit test rework for test_altcloud and test_azure to simplify mocks
and make use of existing util and test_helpers functions.
|
|
Azure datasource now queries IMDS metadata service for network
configuration at link local address
http://169.254.169.254/metadata/instance?api-version=2017-12-01. The
azure metadata service presents a list of macs and allocated ip addresses
associated with this instance. Azure will now also regenerate network
configuration on every boot because it subscribes to EventType.BOOT
maintenance events as well as the 'first boot'
EventType.BOOT_NEW_INSTANCE.
For testing add azure-imds --kind to cloud-init devel net_convert tool
for debugging IMDS metadata.
Also refactor _get_data into 3 discrete methods:
- is_platform_viable: check quickly whether the datasource is
potentially compatible with the platform on which is is running
- crawl_metadata: walk all potential metadata candidates, returning a
structured dict of all metadata and userdata. Raise InvalidMetaData on
error.
- _get_data: call crawl_metadata and process results or error. Cache
instance data on class attributes: metadata, userdata_raw etc.
|
|
The Azure data source provides a method to check whether a NTFS partition
on the ephemeral disk is safe for reformatting to ext4. The method checks
to see if there are customer data files on the disk. However, mounting
the partition fails on systems that do not have the capability of
mounting NTFS. Note that in this case, it is also very unlikely that the
NTFS partition would have been used by the system (since it can't mount
it). The only case would be where an update to the system removed the
capability to mount NTFS, the likelihood of which is also very small.
This change allows the reformatting of the ephemeral disk to ext4 on
systems where mounting NTFS is not supported.
|
|
This change is for Azure VM Preprovisioning. A bug was found when after
azure VMs report ready the first time, during the time when VM is polling
indefinitely for the new ovf-env.xml from Instance Metadata Service
(IMDS), if a reboot happens, we send another report ready signal to the
fabric, which deletes the reprovisioning data on the node.
This marker file is used to fix this issue so that we will only send a
report ready signal to the fabric when no marker file is present. Then,
create a marker file so that when a reboot does occur, we check if a
marker file has been created and decide whether we would like to send the
repot ready signal.
LP: #1765214
|
|
This enables warnings produced by pylint for unused variables (W0612),
and fixes the existing errors.
|
|
Reducing timeout to 1 second as IMDS responds within a handful
of milliseconds. Also get rid of max_retries to prevent exiting
out of polling loop early due to IMDS outage / upgrade.
Reduce Azure PreProvisioning HTTP timeouts during polling to
avoid waiting an extra minute.
LP: #1752977
|
|
LP: #1754495
|
|
This change will enable azure vms to report provisioning has completed
twice, first to tell the fabric it has completed then a second time to
enable customer settings. The datasource for the second provisioning is
the Instance Metadata Service (IMDS),and the VM will poll indefinitely for
the new ovf-env.xml from IMDS.
This branch introduces EphemeralDHCPv4 which encapsulates common logic
used by both DataSourceEc2 an DataSourceAzure for temporary DHCP
interactions without side-effects.
LP: #1734991
|
|
This fixes a traceback when attempting to bounce the network after
hostname resets.
In artful and bionic ifupdown package is no longer installed in default
cloud images. As such, Azure can't use those tools to bounce the network
informing DDNS about hostname changes. This doesn't affect DDNS updates
though because systemd-networkd is now watching hostname deltas and with
default behavior to SendHostname=True over dhcp for all hostname updates
which publishes DDNS for us.
LP: #1722668
|
