[pve-devel] pvedaemon hanging because of qga retry

[Thomas Lamprecht]

On 5/21/18 3:02 PM, Alexandre DERUMIER wrote:
>>> Seems this patch does not solve the 'high load' problem at all?
> 
> I can't reproduce this high load, so I can't say.

For the high fsfreeze timeout my commit message should provide some
context:

> commit cfb7a70165199eca25f92272490c863551efcd89
> Author: Thomas Lamprecht <t.lamprecht at proxmox.com>
> Date:   Wed Nov 23 11:40:41 2016 +0100
> 
> increase timeout from guest-fsfreeze-freeze
> 
> The qmp command 'guest-fsfreeze-freeze' issues in linux a FIFREEZE
> ioctl call on all mounted guest FS.
> This ioctl call locks the filesystem and gets it into an consistent
> state. For this all caches must be synced after blocking new writes
> to the FS, which may need a relative long time, especially under high
> IO load on the backing storage.
> 
> In windows a VSS (Volume Shadow Copy Service) request_freeze will
> issued. As of the closed Windows nature the exact mechanisms cannot
> be checked but some microsoft blog posts and other forum post suggest
> that it should return fast but certain workloads can still trigger a
> long delay resulting an similar problems.
> 
> Thus try to minimize the error probability and increase the timeout
> significantly.
> We use 60 minutes as timeout as this seems a limit which should not
> get trespassed in a somewhat healthy system.
> 
> See:
> https://forum.proxmox.com/threads/22192/
> 
> see the 'freeze_super' and 'thaw_super' function in fs/super.c from
> the linux kernel tree for more details on the freeze behavior in
> Linux guests.


> My main concern is to not wait for a down daemon. (which will never response).
> 
> If we can be sure that daemon is running, with high load, simply wait for a response with a longer timeout.
> 
> 

But, AFAICT, this isn't your real concern, you propose to make a "simple"
qmp call, be it through the VSERPORT_CHANGE, or a backward compatible ping,
where we know that the time needed to answer cannot be that high, as no IO
is involved. That could be done with a relative small timeout and if that
fails we know that it doesn't makes sense to make the fsfreeze call with it
- reasonable - high timeout. If I understood correctly?

> 
> 
> ----- Mail original -----
> De: "dietmar" <dietmar at proxmox.com>
> À: "aderumier" <aderumier at odiso.com>
> Cc: "pve-devel" <pve-devel at pve.proxmox.com>
> Envoyé: Lundi 21 Mai 2018 09:56:03
> Objet: Re: [pve-devel] pvedaemon hanging because of qga retry
> 
>> I have looked at libvirt/ovirt. 
>>
>> It seem that's it's possible to detect if agent is connected, through a qmp 
>> event VSERPORT_CHANGE. 
>>
>> https://git.qemu.org/?p=qemu.git;a=commit;h=e2ae6159 
>> https://git.qemu.org/?p=qemu.git;a=blobdiff;f=docs/qmp/qmp-events.txt;h=d759d197486a3edf3b629fb11e9922ad92fb041a;hp=9d7439e3073ac63b639ce282c7466933ccb411b4;hb=032baddea36330384b3654fcbfafa74cc815471c;hpb=db52658b38fea4e54c23c9cfbced9478d368aa84 
> 
> Seems this patch does not solve the 'high load' problem at all? 
> 
> _______________________________________________
> pve-devel mailing list
> pve-devel at pve.proxmox.com
> https://pve.proxmox.com/cgi-bin/mailman/listinfo/pve-devel
>