[ovirt-devel] [VDSM] Correct implementation of virt-sysprep job
Nir Soffer
nsoffer at redhat.com
Wed Dec 7 14:00:34 UTC 2016
On Wed, Dec 7, 2016 at 10:17 AM, Oved Ourfali <oourfali at redhat.com> wrote:
>
>
> On Tue, Dec 6, 2016 at 11:12 PM, Adam Litke <alitke at redhat.com> wrote:
>>
>> On 06/12/16 22:06 +0200, Arik Hadas wrote:
>>>
>>> Adam,
>>
>>
>> :) You seem upset. Sorry if I touched on a nerve...
>>
>>> Just out of curiosity: when you write "v2v has promised" - what exactly
>>> do you
>>> mean? the tool? Richard Jones (the maintainer of virt-v2v)? Shahar and I
>>> that
>>> implemented the integration with virt-v2v? I'm not aware of such a
>>> promise by
>>> any of these options :)
>>
>>
>> Some history...
>>
>> Earlier this year Nir, Francesco (added), Shahar, and I began
>> discussing the similarities between what storage needed to do with
>> external commands and what was designed specifically for v2v. I am
>> not sure if you were involved in the project at that time. The plan
>> was to create common infrastructure that could be extended to fit the
>> unique needs of the verticals. The v2v code was going to be moved
>> over to the new infrastructure (see [1]) and the only thing that
>> stopped the initial patch was lack of a VMWare testing environment for
>> verification.
>>
>> At that time storage refocused on developing verbs that used the new
>> infrastructure and have been maintaining its suitability for general
>> use. Conversion of v2v -> Host Jobs is obviously a lower priority
>> item and much more difficult now due to the early missed opportunity.
>>
>>> Anyway, let's say that you were given such a promise by someone and thus
>>> consider that mechanism to be deprecated - it doesn't really matter.
>>
>>
>> I may be biased but I think my opinion does matter.
>>
>>> The current implementation doesn't well fit to this flow (it requires
>>> per-volume job, it creates leases that are not needed for template's
>>> disks,
>>> ...) and with the "next-gen API" with proper support for virt flows not
>>> even
>>> being discussed with us (and iiuc also not with the infra team) yet, I
>>> don't
>>> understand what do you suggest except for some strong, though irrelevant,
>>> statements.
>>
>>
>> If you are willing to engage in a good-faith technical discussion I am
>> sure I can help you to understand. These operations to storage demand
>> some form of locking protection. If volume leases aren't appropriate then
>> perhaps we should use the VM Leases / xleases that Nir is finishing
>> off for 4.1 now.
>>
>>> I suggest loud and clear to reuse (not to add dependencies, not to
>>> enhance, ..)
>>> an existing mechanism for a very similar flow of virt-v2v that works well
>>> and
>>> simple.
>>
>>
>> I clearly remember discussions involving infra (hello Oved), virt
>> (hola Michal), and storage where we decided that new APIs performing
>> async operations involving external commands should use the HostJobs
>> infrastructure instead of adding more information to Host Stats.
>> These were the "famous" entity polling meetings.
We discussed these issues behind close doors, not in the public mailing list,
so it is not surprising that people do not know about the agreements we had.
>>
>> Of course plans can change but I have never been looped into any such
>> discussions.
>>
>
> Well, I think that when someone builds a good infrastructure he first needs
> to talk to all consumers and make sure it fits.
> In this case it seems like most work was done to fit the storage use-case,
> and now you check whether it can fit others as well....
The jobs framework is generic and can be used for any subsystem,
there is nothing related to storage about it. But modifying disks *is*
a storage operation, even if someone from the virt team worked on it.
V2v is also storage operation - if we compare it with copying disks:
- we create a new volume that nobody is using yet
- if the operation fails, the disk must be in illegal state
- if the operation fails we delete the disks
- if the operation succeeds the volume must be legal
- we need to limit the number of operations on a host
- we need to detect the job state if the host becomes non-responsive
- we may want to fence the job if the host becomes non-responsive
in volume jobs, we can increment the volume generation and run
the same job on another host.
- we want to take a lease on storage to ensure that other hosts cannot
access the same entity, or that the job will fail if someone else is using
this entity
- we want to take a lease on storage, ensuring that a job cannot get
stuck for long time - sanlock kill the owner of a lease when storage
becomes inaccessible.
- we want to report progress
sysprep is less risky because the operation is faster, but on storage even
fast operation can get stuck for minutes.
We need to agree on a standard way to do such operations that is safe enough
and can be managed on the engine side.
> IMO it makes much more sense to use events where possible (and you've
> promised to use those as well, but I don't see you doing that...). v2v
> should use events for sure, and they have promised to do that in the past,
> instead of using the v2v jobs. The reason events weren't used originally
> with the v2v feature, was that it was too risky and the events
> infrastructure was added too late in the game.
Events are not replacing the need for managing jobs in the vdsm side.
Engine must have a way to query the current jobs before subscribing
to events from these jobs, otherwise you will loose events and engine
will never notice a completed job after network errors.
The jobs framework supports events, see
https://gerrit.ovirt.org/67118
We are waiting for review from the infra team, maybe you can
get someone to review this?
Nir
>
>
>>>
>>> Do you "promise" to implement your "next gen API" for 4.1 as an
>>> alternative?
>>
>>
>> I guess we need the design first.
>>
>>
>>> On Tue, Dec 6, 2016 at 5:04 PM, Adam Litke <alitke at redhat.com> wrote:
>>>
>>> On 05/12/16 11:17 +0200, Arik Hadas wrote:
>>>
>>>
>>>
>>> On Mon, Dec 5, 2016 at 10:05 AM, Nir Soffer <nsoffer at redhat.com>
>>> wrote:
>>>
>>> On Sun, Dec 4, 2016 at 8:50 PM, Shmuel Melamud
>>> <smelamud at redhat.com>
>>> wrote:
>>> >
>>> > Hi!
>>> >
>>> > I'm currently working on integration of virt-sysprep into
>>> oVirt.
>>> >
>>> > Usually, if user creates a template from a regular VM, and
>>> then
>>> creates
>>> new VMs from this template, these new VMs inherit all
>>> configuration
>>> of the
>>> original VM, including SSH keys, UDEV rules, MAC addresses,
>>> system
>>> ID,
>>> hostname etc. It is unfortunate, because you cannot have two
>>> network
>>> devices with the same MAC address in the same network, for
>>> example.
>>> >
>>> > To avoid this, user must clean all machine-specific
>>> configuration
>>> from
>>> the original VM before creating a template from it. You can do
>>> this
>>> manually, but there is virt-sysprep utility that does this
>>> automatically.
>>> >
>>> > Ideally, virt-sysprep should be seamlessly integrated into
>>> template
>>> creation process. But the first step is to create a simple
>>> button:
>>> user
>>> selects a VM, clicks the button and oVirt executes virt-sysprep
>>> on
>>> the VM.
>>> >
>>> > virt-sysprep works directly on VM's filesystem. It accepts
>>> list of
>>> all
>>> disks of the VM as parameters:
>>> >
>>> > virt-sysprep -a disk1.img -a disk2.img -a disk3.img
>>> >
>>> > The architecture is as follows: command on the Engine side
>>> runs a
>>> job on
>>> VDSM side and tracks its success/failure. The job on VDSM side
>>> runs
>>> virt-sysprep.
>>> >
>>> > The question is how to implement the job correctly?
>>> >
>>> > I thought about using storage jobs, but they are designed to
>>> work
>>> only
>>> with a single volume, correct?
>>>
>>> New storage verbs are volume based. This make it easy to manage
>>> them on the engine side, and will allow parallelizing volume
>>> operations
>>> on single or multiple hosts.
>>>
>>> A storage volume job is using sanlock lease on the modified
>>> volume
>>> and volume generation number. If a host running pending jobs
>>> becomes
>>> non-responsive and cannot be fenced, we can detect the state of
>>> the job, fence the job, and start the job on another host.
>>>
>>> In the SPM task, if a host becomes non-responsive and cannot be
>>> fenced, the whole setup is stuck, there is no way to perform
>>> any
>>> storage operation.
>>> > Is is possible to use them with operation that is performed
>>> on
>>> multiple
>>> volumes?
>>> > Or, alternatively, is it possible to use some kind of 'VM
>>> jobs' -
>>> that
>>> work on VM at whole?
>>>
>>> We can do:
>>>
>>> 1. Add jobs with multiple volumes leases - can make error
>>> handling
>>> very
>>> complex. How do tell a job state if you have multiple
>>> leases?
>>> which
>>> volume generation you use?
>>>
>>> 2. Use volume job using one of the volumes (the boot volume?).
>>> This
>>> does
>>> not protect the other volumes from modification but engine
>>> is
>>> responsible
>>> for this.
>>>
>>> 3. Use new "vm jobs", using a vm lease (should be available
>>> this
>>> week
>>> on master).
>>> This protects a vm during sysprep from starting the vm.
>>> We still need a generation to detect the job state, I think
>>> we
>>> can
>>> use the sanlock
>>> lease generation for this.
>>>
>>> I like the last option since sysprep is much like running a vm.
>>> > How v2v solves this problem?
>>>
>>> It does not.
>>>
>>> v2v predates storage volume jobs. It does not use volume leases
>>> and
>>> generation
>>> and does have any way to recover if a host running v2v becomes
>>> non-responsive
>>> and cannot be fenced.
>>>
>>> It also does not use the jobs framework and does not use a
>>> thread
>>> pool for
>>> v2v jobs, so it has no limit on the number of storage
>>> operations on
>>> a host.
>>>
>>>
>>> Right, but let's be fair and present the benefits of v2v-jobs as
>>> well:
>>> 1. it is the simplest "infrastructure" in terms of LOC
>>>
>>>
>>> It is also deprecated. V2V has promised to adopt the richer Host Jobs
>>> API in the future.
>>>
>>>
>>> 2. it is the most efficient mechanism in terms of interactions
>>> between
>>> the
>>> engine and VDSM (it doesn't require new verbs/call, the data is
>>> attached to
>>> VdsStats; probably the easiest mechanism to convert to events)
>>>
>>>
>>> Engine is already polling the host jobs API so I am not sure I agree
>>> with you here.
>>>
>>>
>>> 3. it is the most efficient implementation in terms of interaction
>>> with
>>> the
>>> database (no date is persisted into the database, no polling is
>>> done)
>>>
>>>
>>> Again, we're already using the Host Jobs API. We'll gain efficiency
>>> by migrating away from the old v2v API and having a single, unified
>>> approach (Host Jobs).
>>>
>>>
>>> Currently we have 3 mechanisms to report jobs:
>>> 1. VM jobs - that is currently used for live-merge. This requires
>>> the
>>> VM entity
>>> to exist in VDSM, thus not suitable for virt-sysprep.
>>>
>>>
>>> Correct, not appropriate for this application.
>>>
>>>
>>> 2. storage jobs - complicated infrastructure, targeted for
>>> recovering
>>> from
>>> failures to maintain storage consistency. Many of the things this
>>> infrastructure knows to handle is irrelevant for virt-sysprep
>>> flow, and
>>> the
>>> fact that virt-sysprep is invoked on VM rather than particular
>>> disk
>>> makes it
>>> less suitable.
>>>
>>>
>>> These are more appropriately called HostJobs and the have the
>>> following semantics:
>>> - They represent an external process running on a single host
>>> - They are not persisted. If the host or vdsm restarts, the job is
>>> aborted
>>> - They operate on entities. Currently storage is the first adopter
>>> of the infrastructure but virt was going to adopt these for the
>>> next-gen API. Entities can be volumes, storage domains, vms,
>>> network interfaces, etc.
>>> - Job status and progress is reported by the Host Jobs API. If a job
>>> is not present, then the underlying entitie(s) must be polled by
>>> engine to determine the actual state.
>>>
>>>
>>> 3. V2V jobs - no mechanism is provided to resume failed jobs, no
>>> leases, etc
>>>
>>>
>>> This is the old infra upon which Host Jobs are built. v2v has
>>> promised to move to Host Jobs in the future so we should not add new
>>> dependencies to this code.
>>>
>>>
>>> I have some arguments for using V2V-like jobs [1]:
>>> 1. creating template from vm is rarely done - if host goes
>>> unresponsive
>>> or any
>>> other failure is detected we can just remove the template and
>>> report
>>> the error
>>>
>>>
>>> We can chose this error handling with Host Jobs as well.
>>>
>>>
>>> 2. the phase of virt-sysprep is, unlike typical storage operation,
>>> short -
>>> reducing the risk of failures during the process
>>>
>>>
>>> Reduced risk of failures is never an excuse to have lax error
>>> handling. The storage flavored host jobs provide tons of utilities
>>> for making error handling standardized, easy to implement, and
>>> correct.
>>>
>>>
>>> 3. during the operation the VM is down - by locking the
>>> VM/template and
>>> its
>>> disks on the engine side, we render leases-like mechanism
>>> redundant
>>>
>>>
>>> Eventually we want to protect all operations on storage with sanlock
>>> leases. This is safer and allows for a more distributed approach to
>>> management. Again, the use of leases correctly in host jobs requires
>>> about 5 lines of code. The benefits of standardization far outweigh
>>> any perceived simplification resulting from omitting it.
>>>
>>>
>>> 4. in the worst case - the disk will not be corrupted (only some
>>> of the
>>> data
>>> might be removed).
>>>
>>>
>>> Again, the way engine chooses to handle job failures is independent of
>>> the mechanism. Let's separate that from this discussion.
>>>
>>>
>>> So I think that the mechanism for storage jobs is an over-kill for
>>> this
>>> case.
>>> We can keep it simple by generalise the V2V-job for other
>>> virt-tools
>>> jobs, like
>>> virt-sysprep.
>>>
>>>
>>> I think we ought to standardize on the Host Jobs framework where we
>>> can collaborate on unit tests, standardized locking and error
>>> handling, abort logic, etc. When v2v moves to host jobs then we will
>>> have a unified method of handling ephemeral jobs that are tied to
>>> entities.
>>>
>>> --
>>> Adam Litke
>>>
>>>
>>
>> --
>> Adam Litke
>
>
More information about the Devel
mailing list