Re: [Qemu-devel] qemu-nbd daemonizing?

[Paolo Bonzini <pbonzini@redhat.com>]

On 01/15/2012 05:54 PM, Michael Tokarev wrote:
> On 15.01.2012 20:17, Paolo Bonzini wrote:
>> On 01/15/2012 01:59 PM, Michael Tokarev wrote:
>>> On 15.01.2012 14:51, Paolo Bonzini wrote:
>>>> On 01/14/2012 10:26 AM, Michael Tokarev wrote:
>>>>> After looking at the yesterdays issue with non-absolute
>>>>> paths for qemu-nbd arguments and daemon(3), I've a
>>>>> question.
>>>>>
>>>>> Why qemu-nbd daemonizes, and does that only when device
>>>>> argument is given (dropping -v/verbose case for now)?
>>>>>
>>>>> This raises two questions:
>>>>>
>>>>>     - shouldn't it do the same daemonizing in case of usual
>>>>>       tcp export?
>>>>
>>>> Perhaps yes, but in that case you cannot use "qemu-nbd -d" to
>>>> kill the daemonized process.
>>>
>>> Sorry?  qemu-nbd -d will connect to the socket the daemonized
>>> daemon is listening on, the same way as it is done now --
>>> nothing really changes there.
>>
>> No, "qemu-nbd -d" will connect to the /dev/nbdX device and tell
>> thekernel to disconnect. This will terminate the daemon (as long as you
>> didn't use --persistent, at least).
>
> Whatever.  In any case qemu-nbd is able to terminate currently
> running daemon, nothing changes with addition daemonizing here
> or without - that was my point.

Not "whatever", and not "in any case".  qemu-nbd is able to terminate a 
currently running daemon indirectly, by detaching /dev/nbd.  The same 
cannot be done in the TCP export case (where --persistent is used more 
often than not).  The point is that qemu-nbd only daemonizes when there 
is a simple way to terminate it externally.

Could it have been done better?  Yes.  Do we have to deal with it?  Yes.

>>> The primary question was why -c daemonizes unconditionally.
>>
>> I don't know. But I don't think daemonizing is really a feature,
>> just something historical that we have to deal with.
>
> Aha.  So can't it just go away, or be controlled by another
> option, not related to -c ?  I think it should be good.

I think it could go away, but I'm quite torn.  It does make sense.

Leaving the current historical default "as is", but also allowing 
daemonization in the TCP case makes sense too.  However, you need to 
preserve the way exit statuses are sent for initialization problems, 
even where daemonizing.

To fix this you would have to add three options: --pidfile (to terminate 
a currently running daemon), --daemonize, --no-daemonize.  Honestly, I 
don't think it's worth it, but I wouldn't refuse patches and, of course, 
improving the documentation would also be a very good idea.

>>> qemu-nbd was runnable on win32 before, so historically it
>>> was exactly the opposite.   I asked you because you mentioned
>>> it is linux-only for the first time, but indeed, at that
>>> time it wasn't compilable on win32 already.
>>
>> Fixing it would be good indeed. We can use qemu-thread for that
>> forexample. But it looks like the third commit ever to qemu-nbd.c already
>> made it non-Linux-only. I don't think there ever was a release that
>> supported qemu-nbd on Win32, right?
>
> It is kinda trivial to fix this.  But this task becomes
> completely unrealistic if that will involve another
> discussion like we currently have.

I'm pretty sure it shouldn't.  This discussion is happening because you 
haven't completely read the code, which has more nuances than you think. 
  Threads, daemonization, error messages, exit statuses, all this 
together makes it quite complex.

Paolo