Re: [RFC] vtunerc: virtual DVB device - is it ok to NACK driver because of worrying about possible misusage?

[Andreas Oberritter <obi@linuxtv.org>]

On 02.12.2011 12:14, Mauro Carvalho Chehab wrote:
> On 01-12-2011 20:55, Andreas Oberritter wrote:
>> On 01.12.2011 21:38, Mauro Carvalho Chehab wrote:
>>> I fail to see where do you need to duplicate dvb-core. An userspace
>>> LD_PRELOAD handler that would do:
>>>
>>> int socket;
>>>
>>> int dvb_ioctl(int fd, unsigned long int request, ...)
>>> {
>>>          void *arg;
>>>          va_list ap;
>>>
>>>          va_start(ap, request);
>>>          arg = va_arg(ap, void *);
>>>          va_end(ap);
>>>
>>>      send_net_ioctl_packet(socket, request, arg);
>>> }
>>>
>>> Is probably all you need to send _any_ ioctl's to a remote machine
>>> (plus client's machine that would decode the ioctl packet and send
>>> the ioctl to the actual driver).
>>>
>>> Of course, you'll need hooks for all syscalls used (likely open, close,
>>> ioctl, read, poll).
>>>
>>> So, there's not much duplication, even if, for whatever reason, you
>>> might need to hook some specific ioctls in order to optimize the
>>> network performance.
>>
>> Mauro, we've already had that discussion last time. In order to
>> intercept ioctls of a device, the device needs to exist to begin with,
>> right? That's where vtuner comes in: It creates the virtual device.
> 
> Yes.
> 
>> For that reason your suggested approach using LD_PRELOAD won't work.
> 
> If you're referring to the device name under /dev, a daemon emulating
> a physical device could create Unix sockets under /dev/dvb.

This won't work either, as it would conflict with kernel device
management. Furthermore, programs will detect that a socket is not a
character device.

> Or (with is the right solution) bind such library into the applications
> that will be used.

You mean "such" libraries broken by design?

>> Besides that, suggesting LD_PRELOAD for something other than a hack
>> can't be taken seriously.
> 
> A Kernel pigback plugin is also a hack.

Inventing random terms doen't help. Even if you consider it a hack,
you're one of very few people to do so. So it's a better hack than using
LD_PRELOAD. Btw., since when are kernel modules called plugins?

>> I think you didn't even understand what vtuner does, after all the
>> discussion that took place.
>>
>>>>>> Of course
>>>>>> I can be wrong, I'm no big kernel hacker. So please show me the
>>>>>> way for it. BTW, even if you can find the way, then data copying
>>>>>> from userspace to the kernel and back is also necessery.
>>>>>
>>>>> See libv4l, at v4l2-utils.git (at linuxtv.org).
>>>>>
>>>>>> I really
>>>>>> don't see any advantage of you solution.
>>>>>
>>>>> And I can't see any advantage on yours ;) Putting something that
>>>>> belongs
>>>>> to userspace into kernelspace just because it is easier to re-use the
>>>>> existing code inside the kernel is not a good argument.
>>>>
>>>> It is only your POV that it should be in userspace.
>>>>
>>>> Creating additional code which not only enlarge code size by 2
>>>> but I think by 10 is really not good idea.  And it get no advantage
>>>> only disadvantages.
>>>>
>>>>>
>>>>> Don't get me wrong but if you want to submit a code to be merged
>>>>> on any existing software (being open source or not), you should be
>>>>> prepared to defend your code and justify the need for it to the
>>>>> other developers.
>>>>
>>>> Sure. I was prepared for technical disscussion, but was fully suprised
>>>> that it was not happend (ok, to be correct, few guys are exception,
>>>> like
>>>> Andreas and few others. I really appreciate it).
>>>>
>>>> So, my question was still not answered: "Can be driver NACKed only
>>>> because of worrying about possible misuse?"
>>>
>>> To answer your question: your driver were nacked because of several
>>> reasons:
>>> it is not a driver for an unsupported hardware,
>>
>> It's not a driver for supported hardware either. You named it before:
>> It's not a driver in your definition at all. It's a way to remotely
>> access digital TV tuners over a network.
> 
> Yes, this is not a driver. It is just a hack to avoid adding network
> support at the userspace applications.

One could argue about the term hack, but yes, exactly, that's what it
is. And this is a very good feature.

>>> you failed to convince
>>> people
>>> why this can't be implemented on userspace,
>>
>> Wrong. You failed to convince people why this must be implemented in
>> userspace. Even Michael Krufky, who's "only" against merging it, likes
>> the idea, because it's useful.
> 
> Sometimes, when I'm debugging a driver, I use to add several hacks inside
> the kernelspace, in order to do things that are useful on my development
> (debug printk's, dirty hacks, etc). I even have my own set of patches that
> I apply on kvm, in order to sniff PCI traffic. This doesn't mean that
> I should send all those crap upstream.

That's a nice story, but it's a completely different topic.

>> Just because something can be implemented in userspace doesn't mean that
>> it's technically superior.
> 
> True, but I didn't see anything at the submitted code or at the discussions
> showing that implementing it in kernelspace is technically superior.
> 
> What I'm seeing is what is coded there:
> 
>     http://code.google.com/p/vtuner/
> 
> The kernelspace part is just a piggyback driver, that just copies data
> from/to
> the dvb calls into another device, that sends the request back to
> userspace.

Is it a driver now or not? Why piggyback? Because it uses dvb-core? Are
all drivers using dvb-core piggyback drivers?

> A separate userspace daemon will get such results and send to the
> network stack:
>     http://code.google.com/p/vtuner/source/browse/vtuner-network.c?repo=apps
> 
> 
> This is technically inferior of letting the application just talk to vtuner
> directly via some library call.

The goal is to *transparently* add remote tuners to existing
applications. Not so hard to understand.

> Btw, applications like vdr, vlc, kaffeine and others already implement
> their
> own ways to remotelly access the DVB devices without requiring any
> kernelspace piggyback driver.

Can vdr, vlc, kaffeine use remote tuners on hosts not running vdr, vlc
or kaffeine? Should we implement networking protocols in scan, w_scan,
czap, tzap, szap, mplayer, dvbsnoop, dvbstream, etc. instead?

>>> the driver adds hooks at
>>> kernelspace
>>> that would open internal API's that several developers don't agree on
>>> exposing
>>> at userspace, as would allow non GPL license compatible drivers to
>>> re-use
>>> their work in a way they are against.
>>
>> What's left is your unreasonable GPL blah blah. So the answer to Honza's
>> question is: Yes, Mauro is nacking the driver because he's worrying
>> about possible misuse.
>>
>> Regards,
>> Andreas
>