[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]

Re: [libvirt] [PATCH v4 05/14] qemu: Store pr runtime data in status XML



On 04/17/2018 05:07 PM, John Ferlan wrote:
> 
> 
> On 04/17/2018 10:19 AM, Michal Privoznik wrote:
>> On 04/17/2018 02:00 PM, John Ferlan wrote:
>>>
>>>
>>> On 04/16/2018 10:56 AM, Michal Privoznik wrote:
>>>> On 04/13/2018 10:57 PM, John Ferlan wrote:
>>>>>
>>>>>
>>>>> On 04/10/2018 10:58 AM, Michal Privoznik wrote:
>>>>>> Now that we generate pr-manager alias and socket path store them
>>>>>> in status XML so that they are preserved across daemon restarts.
>>>>>>
>>>>>> Signed-off-by: Michal Privoznik <mprivozn redhat com>
>>>>>> ---
>>>>>>  src/qemu/qemu_domain.c | 64 ++++++++++++++++++++++++++++++++++++++++++++++++++
>>>>>>  1 file changed, 64 insertions(+)
>>>>>>
>>>>>
>>>>> So if we were to save this information (and at this point I don't think
>>>>> we need to), then this is where we should be allocating and filling in
>>>>> the private data (e.g. not in the previous patch).
>>>>
>>>> How come? What would be left from the previous patch if private runtime
>>>> struct would be introduced only here? Or are you just suggesting
>>>> swapping these two patches?
>>>>
>>>
>>> I hope I provided enough feedback in the prior response to answer this.
>>>
>>>>>
>>>>> Again as I noted previously, save the alias when printing the domain
>>>>> active information and I think you're good.
>>>>
>>>> No, I don't want to expose info on PR helper more than is necessary. The
>>>> fact that it's a separate process should not be visible to users because
>>>> it is an implementation detail of QEMU. Other hypervisors might do this
>>>> differently. And even though it might not be visible from the patches,
>>>> using unmanaged mode is discouraged. In fact, unmanaged mode is on the
>>>> edge. If pr-helper is viewed as internal implementation the unmanaged
>>>> mode has no place in libvirt. However, qemu devels are experimenting
>>>> with systemd socket activation and for socket path must be configurable
>>>> through libvirt. So the only reason for using unmanaged PRs is systemd
>>>> socket activation.
>>>
>>> We "expose" aliases a lot in the active domain XML. Someone that's going
>>> to add a <reservations enabled='yes' managed='yes'/> to their <disk...>
>>> definition I cannot believe would be surprised to see an alias printed.
>>
>> We already don't expose some aliases. For instance, if a domain is
>> configured to use hugepages and we use memory-backend-file we don't
>> report generated aliases anywhere. Why? Because the fact we are using
>> memory-backend-file to tell qemu to use hugepages is internal
>> implementation. And users should not be concerned with that. It is the
>> same story with pr-manager and its alias. It is internal implementation
>> deatail and as such we should not expose it.
>>
> 
> Does that code save the alias in some private structure? The correlation
> being it's the libvirt <-> qemu interaction and saving it has
> implications related to what any future change may need to handle.

No it doesn't. Because that devices can't be manipulated. So the alias
is constructed just once, added to the command line and the forgot. But
with pr-manger we need to be able to manipulate it.

> 
>>>
>>> How would they know from the alias that it's a separate process? The
>>> only way to correlate the two would be to read the code and know what
>>> QEMU did to make libvirt do a little dance in order to support.
>>
>> You probably misunderstood what I meant. My idea is to expose as little
>> info back to user as possible in this case. I don't see any compelling
>> reason for user to learn the pr-manager's alias.
>>
>>>
>>> As for systemd, oh great another area to fall flat on our faces...
>>> Wasn't another reason to shorten the path w/ domain name because there
>>> was some sort of bad systemd interaction?
>>
>> Don't recall. It's not relevant.
>>
>>>
>>>>
>>>> Side note, we are not even exposing qemu's PID anywhere because not
>>>> every hypervisor we support has VMs as separate processes.
>>>>
>>>
>>> The PID though could be an unexposed domain private data, couldn't it?
>>
>> Why should we track PID of pr-helper? What do we need it for? As Peter
>> pointed out in review to my previous patches, PIDs change therefore if
>> we start pr-helper process with PID X, later when shutting down domain
>> we could find a different process under the same PID. Because pr-helper
>> might have died, released the PID and another process could have been
>> started with the same PID.
>>
> 
> True...  Of course death of pr-helper is another problem not entirely
> managed by any of this IIRC - the assumption being that if we started
> one once, then it'll run forever, but if it does die then we won't
> because we assume that once started is always started. Still, as you
> pointed out elsewhere some sort of future event should help us to
> perform a restart. Leaving libvirt to manage the qemu problem.
> 
> I guess I see pr-helper as a domain private thing ready to start/stop
> when some other disk source code comes along and says I have this thing
> and need to use the managed domain pr-helper - please add me. The domain
> pr-helper code then could say, well this is a first - something needs to
> be started too. Using return values you could know failure=-1, new=0,
> just-add=1. Keeping track of the number of disks using it in the domain
> private structure and when hotunplug causes that count to go to zero, we
> can remove the pr-helper. A hotplug or for a first time afterwards would
> need to start the pr [again] (and I suppose consider the namespace and
> cgroup implications).
> 
> It's just a different way of looking at it. I find the multiple travels
> through ndisks and the usage of storage source private as unnecessary,
> but it's part of your design model so I also understand your reluctance
> to change that.

Okay, whatever. I will send another version. I give up. It will use
static alias and not save anything into status XML. If it helps me to
get this in so be it.

Michal


[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]