fc.qemu 1.1

Qemu VM management utilities

This package provides a utility to manage virtual machines and their life cycle in the Flying Circus. We try to keep specifics of our environment out of there, but we make a few assumptions:

• VM disks (root, swap, tmp) are stored in Ceph

• There is a script create-vm that will prepare a fresh root disk image.

The utility allows you to

• start, stop and migrate VMs between hosts

• run a daemon that enforces the policy about running VMs given by a set of config files

• resize disks.

Config format

Generic template

The Qemu config file will be generated from a template. If no template is found in /etc/qemu/qemu.vm.cfg.in, a built-in default template will be used. Refer to qemu.vm.cfg.in in the source distribution.

Per-VM configuration

Expects a config file for each VM in /etc/qemu/vm/*.cfg.

The config file format is YAML.

Format:

name: test00
parameters:
    id: 12345
    resource_group: test

    online: true
    kvm_host: bob

    disk: 5
    memory: 512
    cores: 1

    nics:
    - srv: 00-01-02-03-05-06
    - fe: 00-01-02-03-05-06

Release notes

1.1 (2019-08-30)

• Add capabilities to discover, report and select CPU models supported by the hosts and requested by guests. #114627

• Move base support from Qemu 2.7 to Qemu 4.1

• Various test fixture improvements for test performance and stability.

• Add global ‘shutdown-all’ command to improve KVM host shutdown timing.

1.0.8 (2019-08-30)

• Brownbag release was intended to be 1.1.

1.0.7 (2019-03-27)

• Suppress error warnings when processing consul snapshot requests for foreign VMs.

1.0.6 (2019-03-26)

• Explicitly fail when trying to take a snapshot of a VM that isn’t running locally. Also improve the snapshot output. (#108360)

1.0.5 (2019-02-14)

• Fix process matching we introduced in 1.0.4 to also cover long VM names. The proc.name() field under Linux only holds 16 bytes and we definitely have VM names that are much longer …

1.0.4 (2019-02-05)

• Do not - I repeat - do not kill processes, based on the PID file, if the process name does not match the expected VM process naming pattern.

  This only manifested under a number of circumstances:

  • a Ceph bug was leaving deleted VMs unable to delete their locks

  • deleted VMs were sometimes considered stale

  • the targetted PID was suddenly different VM because we recycled PIDs much faster than expected (about every hour)

1.0.3 (2019-01-30)

• Add custom daemonization wrapper for Qemu to capture stdout/stdin/exit code (reliably).

• Add helper to support migrations on Ceph Hammer w/ broken lock IDs.

• Fix minor bug when showing status and Consul is inconsistent.

1.0.2 (2019-01-08)

• Fix an edge condition: VMs that are still shutting down and then asked to migrate may leave the inmigrate job hanging until it times out. We already have opportunistic handling of config changes and this is now one of them thus triggering a regular start of affected VMs much quicker.

• Opportunistically try to re-acquire lost locks. This is needed to avoid an invalid “inconsistent” decision and allows us to repair broken locks due to a Ceph bug. This will still cause visible failures if the lock can’t be properly re-acquired.

• Clean up expectations around in-migrations and how to discover whether it makes sense to wait for an incoming migration. Use consul _and_ Ceph lock information and concentrate that decision in a single place.

1.0.1 (2018-06-04)

• Bugfix: having a staging config but no active config caused the agent to globally file and thus blocked creating new VMs.

1.0 (2018-05-23)

• Remove deprecated config file locking mechanism.

• Remove (synchronous and superfluous) memory compaction during VM startup to avoid unexpected delays that caused timeouts and delays VM startup unnecessarily.

• Improve the configuration file update process.

  This adds a staging area so that consul updates become faster and handling change events now has become asynchronous. Also, pay attention to the consul ModifyIndex here to avoid lost updates and use that for actual triggers. Includes a “settling” mechanism so that runs of ensure that see changes will look for further changes while holding the lock. This means that for every VM on a KVM host only one agent process will be running even if a flurry of updates is coming in.

• Introduce a host-specific migration lock: we want VMs to run as fast as possible and not overload the host. Having a single migration run at one point in time means there’s a single TCP stream that will compete with different other steams that run on the same network (Ceph client in our case). Also, competing streams caused some streams to establish a slow speed that wouldn’t increase after being the only one remaining.

  Further optimizations here include that hosts will retry quickly to get those locks and keep pushing migrations forward.

• Improve timeout handling for migrations.

  With the new asynchronous agent we now wait for a long time to discover our peer and have fixed missing timeout resets whenever our peer contacted us. (This was an issue with the now removed memory compaction that took longer than the timeout and would cause ‘sudden death’ on migrations.)

• Make the guest agent part of the ‘ensure’ method a bit more robust against missing guest agents.

• Improve cleanup of incoming migration services in consul. We did tend to leave old ones around. We will now try connecting to most recent services first, then try older ones and eliminate those that we can’t talk to in the process. This may leave some around for a while but the tendency should be that they will get cleaned up at some point.

0.9.8 (2018-01-30)

• Restore bare except.

0.9.7 (2018-01-30)

• When several vm-inmigrate services are found, select the one with the highest ‘ModifyIndex’ value.

0.9.6 (2018-01-30)

• Fix issue with using the ls and check commands. They do not perform regular locking as they do not use any of the critical resources and should not block nor be blocked by ongoing operations.

  Nevertheless, the cleanup code for managing Qemu QMP monitor connections accidentally triggered creating a superfluous session upon cleanup and that caused spurious blocks.

0.9.5 (2018-01-25)

• Relax version requirements for consulate (#27695).

0.9.4 (2017-12-19)

• Introduce a Nagios/Sensu compatible “check” command that performs a couple of checks. At the moment this verifies whether RAM allocations are met, both on a VM level (not exceeding guest memory + 2 times the expected overhead) and a host level (not exceeding guest memory + total overhead + 10%). This command may be expanded upon in the future to introduce more checks.

• Allow selecting the “disk cache mode” of Qemu so we can start disabling the writeback cache that keeps using way too much memory. #28840

• Extend the memory limiting feature to also verify whether the host actually has sufficient memory (in addition to sufficient bookable memory) and also include expected overhead in addition to the guest memory.

• Fix the fc-qemu global lock to use a dedicated lock file instead of (once again) using a file that might be moved around unexpectedly.

0.9.3 (2017-12-11)

• Disable watchdog: for newly started VMs, remove the device completely. For existing VMs (and future) disable the restart action. This fixes spontaneous reboots if VMs are under memory and IO pressure and fail to inform the watchdog appropriately. In general, it appears that the situations in which a watchdog will be helpful are very limited and the misbehaviour has become too much of a burden.

0.9.2 (2017-09-18)

• Fix guest agent timeout handling so that snapshots get an actual 120s timeout to allow VMs to properly flush their IO.

0.9.1 (2017-09-01)

• Remediate brownbag release: default config options were not properly included in release.

0.9 (2017-09-01)

• [feature] Allow managing host memory with a “maximum total” of memory that can be configured on VMs (-m switch). This is not based on actual RAM usage or availability but planned and configured values of currently running VMs!

  If a VM shall be inmigrated or started and the host would go beyond the ‘vm-max-total-memory’ setting with that VM then the action will fail.

• [locking] Fix locking by moving to a separate lockfile and ensuring that reentrant use of locks is handled correctly. Also, provide upgrade scenario by additionally using the existing lock files.

• [locking] Ensure we rewrite the config file only when locked to reduce attack surface of the old mechanism while still upgrading.

• [logging] Improve logging with more specific debug output.

• [logging] Log the specific commandline that each Qemu process is started with to aid debugging.

• [logging] Switch logging from UTC to local server time. This has proven much more confusing as our other on-disk logs are in server time.

• [logging] Do not crash on broken logging (i.e. disk full or STDIO missing) to avoid accidentally crashing a VM just because the controlling script runs into issues.

• [logging] Add VM name prefixes also for stack dumps and tracebacks.

• [consul] Ensure that consul event handling doesn’t fail the process when a thread fails. (Might be snake oil but doesn’t hurt.)

• [consul] Reduce consul event handling pool from 10 to 3 to reduce strain on multiple parallel migrations.

• [ceph] Properly close RBD volume references to avoid librbd crash.

• [ceph] Ensure we don’t open unnecessary duplicate RBD image handles.

• [qemu] Increase Qemu QMP timeout to 5 minutes to tolerate latency during migration cleanup as the QMP handler runs in the main thread and could be blocked for a long period.

• [migration] Improve outgoing migrations connecting to the incoming server: simplify code and reduce unnecessary waiting periods. Better error output.

• [migration] Improve live migration: skip compression, enable unlimited bandwidth, and use ephemeral ports to avoid running into TCP timing issues when retrying live migrations quickly.

• [config] Clean up default-option handling for some config options: we used two different styles of defaults. They have been unified into a single “default.conf” that gets loaded first.

0.8.11 (2017-05-29)

• Add thousands separator in logging to live migration log to allow easier optical inspection.

• Improve fsfreeze timeout handling: this can take quite a while and if we are too eager then we end up quickly in unstable states.

• Improve error and debug logging.

• Improve resilience of continuing locally after a failed migration.

0.8.10 (2017-04-12)

• Improve logging: include PID of the running process to help detect and understand potential conflicts in parallel runs.

• Try harder to catch errors and retry correctly when resetting communication with an agent.

• Add another layer to protect the guest agent by acquiring an exclusive lock on the guest agent socket file.

• Keep trying to ensure a VM is unfrozen during regular ensure calls.

0.8.9 (2017-04-07)

• Improve guest agent communication. The guest agent may be in an inconsistent state that causes it to hang. We’ve seen this happening where we froze machines and then have the agent be inconsistent.

  This now properly resets the agent connection upon a sync by sending the recommend “wrong” UTF-8 byte that guarantees to interrupt the guest agent’s JSON parser.

• Improve logging output when destroying a VM because of an inconsistent state. (#25158)

0.8.8 (2016-11-18)

• Bugfix: migrations that ran into a timeout because the remote side did not respond accidentally unlocked and cleaned up without shutting the VM down properly. This resulted in multiple instances of a single VM.

0.8.7 (2016-11-11)

• Don’t change anything if a VM is marked online but no KVM host is assigned (#23965).

• Refactor Agent.ensure() for improved reliability and readability.

• Decline to create a consistent snapshot if a VM is offline.

• Speed up tests a bit.

• Don’t spawn unqualified partprobe invocations in parallel.

• Make debug level logging a bit less verbose.

0.8.6 (2016-11-04)

• Break inconsistent Ceph locks if the host holding an old lock is sure that a VM is not running anymore (#23695).

• Migration compatibility between Qemu 2.5 and 2.7 (#23695).

• Always clean up unused resources like Consul service registrations and run files.

• Improve error reporting and logging.

0.8.5 (2016-10-31)

• Fixed a major bug with event processing: the consul event processor was using the multiprocessing.pool API incorrectly. This wasn’t caught by the tests and resulted in silent “no ops” of all event processing mechanisms.

0.8.4 (2016-10-31)

• Waiting for Qemu to shutdown gracefully was not expecting a socket error, which caused restarts to fail (cleanly). #24434

• Limit the number of parallel processed consul events to avoid overloading the host. Can be configured in the fc-qemu config file through:

  [consul]
  event-threads = <INT>

  The default is 10 threads.

• Lower the overhead of processing a consul config change event: do not activate Ceph and Qemu connections and do not perform scrubbing (ensure) when the config hasn’t changed. Ceph and Qemu connections aren’t needed in that case and scrubbing is expected to be performed in a separate task from a scheduler, not from an event handler that is supposed to only respond to changes.

• Provide a default for the binary-generation counter to allow smooth upgrades from previous versions.

0.8.3 (2016-10-23)

• Lower the guest agent timeout to help the tests complete faster and also stay responsive if the guest shouldn’t have an agent available either yet, or currently, or generally.

• Provide a Qemu config template variable that will determine the most current Qemu machine type, given a prefix to filter for.

• Provide a “binary generation counter” that is a) injected a boot (to /tmp/fc-data/qemu-binary-generation-booted) and b) updated during every “ensure” command (to /run/qemu-binary-generation-current). The guest should use a difference between those files to schedule a cold reboot (i.e. a shutdown) to restart with a fresh Qemu binary.

0.8.2 (2016-09-20)

• Switch from using multiprocessing to threaded management of multiple consul VM event handlers to reduce Python startup overhead when processing many VMs in a single location. Also remove the sleep time.

• Fix log error in consul snapshot event handling. Improve the test coverage for consul event handling.

• Remove superfluous VM name handling where a VM config file could be specified instead of a VM name. This was causing obfuscation in the code and was barely used anyway.

• Allow using the telnet command even when fc.qemu _thinks_ that the VM is not running. This command is helpful for debugging and blocking it is a useless seatbelt.

• Do not log QMP connection errors as the are extremely common and expected.

0.8.1 (2016-09-11)

• Explicitly add logging to /var/log/fc-qemu.log. Do not filter log output there: we always want all the information we can get.

0.8 (2016-09-08)

• Introduce fc-qemu telnet command: a shortcut to connect to the human monitor port without having to look up the port from a config manually.

• Switch the monitor automation from using the telnet port to using the QMP socket. This should be a _lot_ more reliable. This also fixes a previous race condition where a migration status check may intermittently fail and then break a migration unnecessarily.

• Update vagrant environment to check against Qemu 2.6.

• Revamp output formatting to use Hynek’s great structlog library.

• Limit a few more commands to specific VM states: stop only when running.

• Implement IOPS limiting either based on VM-specific ENC data, a Ceph pool default, or a global default. This limits IOPS for all disks (individually, no groups, yet) in a VM and maintains this over time.

• Rework Ceph volume unlocking: if a client owns a lock then breaking it will cause an immediate disconnect of the rbd/rados connection to avoid it sending further data updates. This can happen to us if we’re setting up the locks on an inmigration and then have to give them up again if the migration fails.

0.7.22 (2016-07-21)

• Fix bug in fc-qemu restart which causes mkfs.xfs for tmp to fail.

0.7.21 (2016-06-20)

• Moderate swap and tmp volume sizes so that they do not scale linearly for very large VMs. #21961

0.7.20 (2016-05-03)

• More logging output to help diagnosing a rare lock recovery failure (#21345).

• Remove shrink-vm. R.I.P. (#14222).

0.7.19 (2016-04-08)

• Fix a race condition: when continuously polling monitor status to determine whether a VM is running, also consider the option that the VM was about to shutdown and the monitor has gone away.

0.7.18 (2016-04-04)

• Another brownbag release: the snapshot refactoring wasn’t tested properly if snapshots actually existed.

0.7.17 (2016-04-04)

• Fix unicode/str issue: consul json decoded into unicode but librbd requires a plain string.

0.7.16 (2016-03-31)

• Fix regression in snapshot taking.

0.7.15 (2016-03-20)

• Account for different mkfs options for XFS and mkfs.ext4 (#19079).

• Improved Vagrant VM bootstrapping.

• Refactor classes in hazmat/{ceph.py,volume.py} (#19079).

• Use the “rbd_pool” ENC option to allow VM-specific selection of the RBD pool instead of deriving it from the resource group name.

• Improve success rate of recovering from failed migrations properly: certain conditions would result in only partially released locks from the target leading to inconsistent states.

0.7.14 (2016-01-21)

• Use XFS for tmp partitions (#17873).

• Drop super-floppy setup von vdc and use a proper partition table instead (#17873).

• Fix file permissions for ENC seed JSON file.

0.7.13 (2015-12-10)

• Ignore consul requests for VMs with missing configuration (#18841).

• Speed up initial NTP sync in Vagrant to avoid failing Ceph tests due to unsynced MONs.

• Refine Ceph.unlock() to remove own locks in a best-effort manner. This is needed to recover from incomplete migrations (#18771).

• Improve error handling with failed monitor connections.

0.7.12 (2015-11-11)

• Improve error handling during migration.

• Fix timeout during fsfreeze that leads to locked up VMs (#18917).

0.7.11 (2015-11-04)

• Switch aio setting in default qemu.vm.cfg to “threads”. This will keep fc-qemu compatible with future Qemu versions (#18743).

• Improve logging.

• Place initial copy of ENC data in /tmp/fc-data/enc.json.

0.7.10 (2015-08-12)

• Refactor system-wide configuration code.

• Create swap and tmp partition with proper filesystem labels (#16783).

• Fix rare race condition during tmp volume creation.

• Set filesystem labels for swap and tmp volumes (#17078).

0.7.9 (2015-08-03)

• Add “snapshot” command. Can be triggered from command line and via consul.

0.7.8 (2015-07-27)

• Improve detection of running instances.

• Broaden check for monitor connection to handle dual stack environments.

0.7.7 (2015-07-14)

• Fix migration issue: we ended up de-registering at the wrong time.

0.7.6 (2015-07-01)

• Quickfix: newer mkfs.ext4 versions need a ‘-F’ flag to overwrite filesystems (#14920).

0.7.5 (2015-07-01)

• Spawn individual VM actions usings multiprocessing. Wait until all migrations are done (#14920).

• Increase allowed migration downtime to keep migration time for busy VMs in bounds (#14920).

• Fix exception handling errors during Consul event processing (#14920).

• Give udev mapping a bit to settle.

• Improve log readability.

0.7.4 (2015-06-02)

• Rectify brown-bag release.

• Fix some unnoticed, arbitrary test failures.

0.7.3 (2015-06-02)

• Make event processing from consul fork for each VM and return the master process early to avoid blocking the consul agent.

• More logging related to migrations.

0.7.2 (2015-05-26)

• Adapt to QEMU 2.2.1: uses now stdvga by default (#15748).

0.7.1 (2015-05-20)

• Fix bug with inmigration Consul service registration (#15313).

• Change KV name name space for nodes from “vm/” to “node/” (#14920).

0.7 (2015-05-18)

• Consul service registration (#15313).

• Coordinate migration via Consul (#15313).

0.6.4 (2015-02-27)

• Tolerate “setup” as an intermediate migration status as encountered in the wild.

0.6.3 (2015-02-19)

• Improve pid file parser to deal correctly with trailing lines and empty pid files.

• Ensure that exceptions are properly logged if they occur directly after daemonizing (e.g., in Agent.__init__()) (#13867).

0.6.2 (2015-01-22)

• Relax PyYaml and psutil version requirements to accommodate to the Flying Circus managed platform.

0.6.1 (2015-01-22)

• Improve logging and error messages (#13867).

• Fix unwanted behaviour during error conditions (#13867).

0.6 (2015-01-15)

• Implement live migration. Use “inmigrate” and “outmigrate” commands to coordinate the process (#13229).

• Note that the qemu.cfg.in template has changed!

• Improve test coverage.

0.5.1 (2014-11-22)

• Bugfix: remove Ceph discard call since it seems to be unstable (#13414).

• Improve operability by reworking what is logged to fc-qemu.log.

0.5 (2014-11-21)

• Root filesystem shrink during VM start (#13414).

• Add ‘force-unlock’ action to break stale locks (e.g., after a VM host went down).

0.4.3 (2014-11-13)

• Read Qemu config file template from /etc/qemu/qemu.vm.cfg.in.

• Fix tests and documentation.

0.4.2 (2014-11-12)

• Rate limit entropy transfer from host to guest (#13751).

• Add ‘restart’ command to simplify VM restarts.

0.4.1 (2014-09-24)

• Do not require the PID to match the machine name for determining online status. This caused issues for VMs with names longer than 11 characters: http://status.flyingcircus.io/incidents/3j8wsrszlx2w

0.4 (2014-09-16)

• Allow selecting the specific command line to call for creating a VM using a config file + formatting syntax.

• Add test coverage to show that we gracefully recover from crashed VMs upon a subsequent ‘ensure’.

0.3 (2014-09-13)

• Refactor and rename to ‘fc.qemu’. Integrate most functionality that was previously placed in our init scripts and localconfig (fc.agent) utilities.

• Add a lot of test coverage.

0.2.6 (2014-08-21)

• Fix incoming VM detection for an already locked _and_ started VM.

0.2.5 (2014-08-20)

• Implement a safety-belt to prohibit migrating VMs that have not yet been started with the supported /run/kvm.*.cfg.in format.

fc.qemu development

Workstation development

Prepare Vagrant environment:

host$ hg clone https://bitbucket.org/flyingcircus/fc.qemu
host$ cd fc.qemu
host$ vagrant up
host$ vagrant ssh

Prepare virtualenv:

vm$ cd /vagrant
vm$ virtualenv --system-site-packages .
vm$ bin/pip install -r requirements.txt

Run the tests:

vm$ cd /vagrant
vm$ sudo bin/py.test

Test execution automatically updates a coverage report in the htmlcov directory.

Run end-to-end migration test:

vm$ cd /vagrant
vm$ sudo ./test-migration.sh

Real-world testing on FCIO DEV network

• Check out the source on a VM host

• Set Puppet stopper

• Create virtualenv: virtualenv –system-site-packages .

• Install software: bin/pip install -r requirements.txt

• Make symlink /usr/src/fc.qemu point to the local checkout

• Install fc-qemu package in development mode: ACCEPT_KEYWORDS=”**” emerge -1 fc-qemu