Re: [Qemu-devel] [PATCH 1/5] monitor: screen_dump async

[Alon Levy <alevy@redhat.com>]

On Tue, Oct 25, 2011 at 10:51:30AM -0200, Luiz Capitulino wrote:
> On Tue, 25 Oct 2011 12:13:09 +0200
> Alon Levy <alevy@redhat.com> wrote:
> 
> > On Mon, Oct 24, 2011 at 10:31:48PM -0200, Luiz Capitulino wrote:
> > > On Mon, 24 Oct 2011 19:29:37 +0200
> > > Alon Levy <alevy@redhat.com> wrote:
> > > 
> > > > On Mon, Oct 24, 2011 at 01:45:16PM -0200, Luiz Capitulino wrote:
> > > > > On Mon, 24 Oct 2011 17:13:14 +0200
> > > > > Gerd Hoffmann <kraxel@redhat.com> wrote:
> > > > > 
> > > > > > On 10/24/11 14:02, Alon Levy wrote:
> > > > > > > Make screen_dump monitor command an async command to allow next for qxl
> > > > > > > to implement it as a initiating call to red_worker and completion on
> > > > > > > callback, to fix a deadlock when issueing a screendump command via
> > > > > > > libvirt while connected with a libvirt controlled spice-gtk client.
> > > > > > 
> > > > > > Approach looks reasonable to me.  Patch breaks the build though, you've
> > > > > > missed a bunch of screen_dump functions in non-x86 targets.
> > > > > 
> > > > > There are two problems actually.
> > > > > 
> > > > > The first one is that changing an existing command from synchronous
> > > > > to asynchronous is an incompatible change because asynchronous commands
> > > > > semantics is different. For an example of possible problems please
> > > > > check: https://bugzilla.redhat.com/show_bug.cgi?id=623903.
> > > > > 
> > > > > The second problem is that the existing asynchronous interface in the
> > > > > monitor is incomplete and has never been used for real. Our plan is to
> > > > > use QAPI's async support, but that has not landed in master yet and iirc
> > > > > there wasn't consensus about it. I also think it's a bit late for its
> > > > > inclusion in 1.0 (and certainly not a candidate for stable).
> > > > > 
> > > > > If all you need here is to delay sending the response, then maybe the
> > > > > current interface could work (although I honestly don't trust it and
> > > > > regret not having dropped it). Otherwise our only choice would be to
> > > > > work on getting the QAPI async support merged.
> > > > 
> > > > My problem is that the io thread keeps the global mutex during the wait,
> > > > that's why the async monitor is perfect for what I want - this is
> > > > exactly what it does.
> > > 
> > > Let's not mix internal implementation details with what we want as
> > > an external interface.
> > > 
> > > Can't you just make a vga_hw_screen_dump() specific callback?
> > > 
> > 
> > I don't understand how that would help - if the monitor command doesn't
> > return (normal sync operation) then the mutex is never dropped, and any
> > callback won't change that.
> 
> I'm trying to figure out a different solution.
> 
> Our primary motivation for making a QMP command asynchronous must be to
> give clients the ability to keep issuing commands while "slow" commands
> are running. If that's not what you want nor your primary motivation,
> then you're probably taking the wrong approach.

That sounds right, and it was what I assumed the async monitor
implementation would do, boy was I surprised to discover it doesn't do
any such thing, but what it does do is return early, allow *other* io
related events (select returns) to be handled, and keeps the serialized
only-one-command-ongoing monitor usage.

> 
> If that is what you want, then you'll have to add a new command, because
> changing from asynchronous to synchronous is an incompatible change _and_
> you shouldn't use the current interface, because it's botched (actually,
> I think I'm going to drop it right now as my last series fixed its only user).
> 

This is not what I want. I understand of course that it is what one
would design an async monitor / api to allow (it was after all what I
thought async monitor support meant).

> Using a botched interface that doesn't do what's supposed to do but happens to
> solve a bug as a side effect will very likely end in tears at some point in
> the future.
> 

Right, but it's the opposite of the current case.

> Now, I did some research and found this description of the problem:
> 
> """
> In testing my patches for 'add_client' support with SPICE, I noticed
> deadlock in the QEMU/SPICE code. It only happens if I close a SPICE
> client and then immediately reconnect within about 1 second. If I
> wait a couple of seconds before reconnecting the SPICE client I don't
> see the deadlock.
> """
> (http://lists.nongnu.org/archive/html/qemu-devel/2011-10/msg02599.html)
> 
> Is this accurate? Why does it _work_ after 1 second?
>

This is an unrelated bug, I know I said different on the thread but now
rereading the callstack it is the channel_event locking workaround - and
it is fixed by a spice-server only patch (which stops calling
channel_event from the worker thread).

> > On the other hand, thinking a bit about the reference to 623903 baloon
> > bug, I don't see a problem: the change doesn't affect the semantics for
> > any other device except qxl, which I've tested. For any other device,
> > the only difference is that instead of:
> > 
> > do_screen_dump call
> >  device specific implementation
> >  return
> > 
> > It becomes
> > 
> > do_screen_dump call
> >  device specific implementation (not qxl)
> >   callback called (always - not conditional, no one stores it except
> >    qxl)
> >  return
> 
> 
> > 
> > > > I haven't looked at QAPI async support, but I
> > > > understand it's a bit in the future.
> > > 
> > > Yes, it's not for the immediate term.
> > 
>