Re: [Qemu-devel] [PATCH V2] Add "tee" option to qemu char device

From: Alexander Graf <agraf@suse.de>
To: Anthony Liguori <anthony@codemonkey.ws>
Cc: Chunyan Liu <cyliu@novell.com>, qemu-devel@nongnu.org
Subject: Re: [Qemu-devel] [PATCH V2] Add "tee" option to qemu char device
Date: Sun, 24 Jul 2011 11:47:59 +0200	[thread overview]
Message-ID: <E1FDE9FA-6A58-4BF4-83AE-E60B295AE334@suse.de> (raw)
In-Reply-To: <4E2B1CEB.5040606@codemonkey.ws>

On 23.07.2011, at 21:11, Anthony Liguori wrote:

> On 07/23/2011 01:31 PM, Alexander Graf wrote:
>> 
>> On 23.07.2011, at 17:23, Anthony Liguori wrote:
>> 
>>> On 07/14/2011 03:58 AM, Chunyan Liu wrote:
>>>> Add "tee" backend to char device. It could be used as follows:
>>>>     -serial tee:filepath,pty
>>>>     -chardev tee,tee_fpath=path,tee_backend=pty,,path=path,,[mux=on|off]
>>>> With "tee" option, "pty" output would be duplicated to filepath.
>>>> Related thread:
>>>> http://lists.nongnu.org/archive/html/qemu-devel/2011-07/msg00105.html
>>> 
>>> I loathe adding even more complexity to the the char layer.  Can't you do this just as well with socat?
>> 
>> I disagree. For socat we'd have to open some listening port (unix, tcg, etc) and then have socat connect to it.
> 
> And what's the problem?  Use a unix domain socket and call it a day.
> 
>> While socat is not up yet, the VM won't run. It also adds another layer of complexity (syncing of socat and qemu process) to the picture that I don't like.
> 
> These arguments all apply to any possible option.  Why not a grep target?  Why not a cut or less target?

Because they don't make sense.

>> As long as the tee target is reasonably isolated, I don't think it'd clutter the char backend.
> 
> It'll never be tested and end up becoming dead bloat code.

I'll be extensively tested in Xen code, since that's how Xen will invoke qemu.

> For uncommon use cases where there's another way to do something with no real obvious technical advantages, using existing code (and utilities) always wins vs reinventing the wheel IMHO.

I tend to agree, but this time the solution in qemu is cleaner and easier IMHO. Also, I've had issues with tee several times already. In fact, I actually tried to use tee plenty of times to log console and debug output to a file while working with Qemu and it randomly just stopped processing data. While that could be a bug somewhere in one of the random layers, it shows me how bad complexity grows if you add even more programs to the puzzle that you can't rely on.

I'm in favor of putting a tee target into qemu's char layer and I'm sure it'll become a well used target.

Alex