Re: [Qemu-devel] [RFC 00/15] QMP: out-of-band (OOB) execution support

[Peter Xu <peterx@redhat.com>]

On Tue, Sep 19, 2017 at 10:19:21AM +0100, Dr. David Alan Gilbert wrote:
> * Peter Xu (peterx@redhat.com) wrote:
> > On Mon, Sep 18, 2017 at 06:09:29PM +0200, Marc-André Lureau wrote:
> > > On Mon, Sep 18, 2017 at 1:26 PM, Dr. David Alan Gilbert
> > > <dgilbert@redhat.com> wrote:
> > > > * Marc-André Lureau (marcandre.lureau@gmail.com) wrote:
> > > >> Hi
> > > >>
> > > >> On Mon, Sep 18, 2017 at 12:55 PM, Dr. David Alan Gilbert
> > > >> <dgilbert@redhat.com> wrote:
> > > >> > * Marc-André Lureau (marcandre.lureau@gmail.com) wrote:
> > > >> >> Hi
> > > >> >>
> > > >> >> On Mon, Sep 18, 2017 at 10:37 AM, Peter Xu <peterx@redhat.com> wrote:
> > > >> >> > On Fri, Sep 15, 2017 at 01:14:47PM +0200, Marc-André Lureau wrote:
> > > >> >> >> Hi
> > > >> >> >>
> > > >> >> >> On Thu, Sep 14, 2017 at 9:46 PM, Peter Xu <peterx@redhat.com> wrote:
> > > >> >> >> > On Thu, Sep 14, 2017 at 07:53:15PM +0100, Dr. David Alan Gilbert wrote:
> > > >> >> >> >> * Marc-André Lureau (marcandre.lureau@gmail.com) wrote:
> > > >> >> >> >> > Hi
> > > >> >> >> >> >
> > > >> >> >> >> > On Thu, Sep 14, 2017 at 9:50 AM, Peter Xu <peterx@redhat.com> wrote:
> > > >> >> >> >> > > This series was born from this one:
> > > >> >> >> >> > >
> > > >> >> >> >> > >   https://lists.gnu.org/archive/html/qemu-devel/2017-08/msg04310.html
> > > >> >> >> >> > >
> > > >> >> >> >> > > The design comes from Markus, and also the whole-bunch-of discussions
> > > >> >> >> >> > > in previous thread.  My heartful thanks to Markus, Daniel, Dave,
> > > >> >> >> >> > > Stefan, etc. on discussing the topic (...again!), providing shiny
> > > >> >> >> >> > > ideas and suggestions.  Finally we got such a solution that seems to
> > > >> >> >> >> > > satisfy everyone.
> > > >> >> >> >> > >
> > > >> >> >> >> > > I re-started the versioning since this series is totally different
> > > >> >> >> >> > > from previous one.  Now it's version 1.
> > > >> >> >> >> > >
> > > >> >> >> >> > > In case new reviewers come along the way without reading previous
> > > >> >> >> >> > > discussions, I will try to do a summary on what this is all about.
> > > >> >> >> >> > >
> > > >> >> >> >> > > What is OOB execution?
> > > >> >> >> >> > > ======================
> > > >> >> >> >> > >
> > > >> >> >> >> > > It's the shortcut of Out-Of-Band execution, its name is given by
> > > >> >> >> >> > > Markus.  It's a way to quickly execute a QMP request.  Say, originally
> > > >> >> >> >> > > QMP is going throw these steps:
> > > >> >> >> >> > >
> > > >> >> >> >> > >       JSON Parser --> QMP Dispatcher --> Respond
> > > >> >> >> >> > >           /|\    (2)                (3)     |
> > > >> >> >> >> > >        (1) |                               \|/ (4)
> > > >> >> >> >> > >            +---------  main thread  --------+
> > > >> >> >> >> > >
> > > >> >> >> >> > > The requests are executed by the so-called QMP-dispatcher after the
> > > >> >> >> >> > > JSON is parsed.  If OOB is on, we run the command directly in the
> > > >> >> >> >> > > parser and quickly returns.
> > > >> >> >> >> >
> > > >> >> >> >> > All commands should have the "id" field mandatory in this case, else
> > > >> >> >> >> > the client will not distinguish the replies coming from the last/oob
> > > >> >> >> >> > and the previous commands.
> > > >> >> >> >> >
> > > >> >> >> >> > This should probably be enforced upfront by client capability checks,
> > > >> >> >> >> > more below.
> > > >> >> >> >
> > > >> >> >> > Hmm yes since the oob commands are actually running in async way,
> > > >> >> >> > request ID should be needed here.  However I'm not sure whether
> > > >> >> >> > enabling the whole "request ID" thing is too big for this "try to be
> > > >> >> >> > small" oob change... And IMHO it suites better to be part of the whole
> > > >> >> >> > async work (no matter which implementation we'll use).
> > > >> >> >> >
> > > >> >> >> > How about this: we make "id" mandatory for "run-oob" requests only.
> > > >> >> >> > For oob commands, they will always have ID then no ordering issue, and
> > > >> >> >> > we can do it async; for the rest of non-oob commands, we still allow
> > > >> >> >> > them to go without ID, and since they are not oob, they'll always be
> > > >> >> >> > done in order as well.  Would this work?
> > > >> >> >>
> > > >> >> >> This mixed-mode is imho more complicated to deal with than having the
> > > >> >> >> protocol enforced one way or the other, but that should work.
> > > >> >> >>
> > > >> >> >> >
> > > >> >> >> >> >
> > > >> >> >> >> > > Yeah I know in current code the parser calls dispatcher directly
> > > >> >> >> >> > > (please see handle_qmp_command()).  However it's not true again after
> > > >> >> >> >> > > this series (parser will has its own IO thread, and dispatcher will
> > > >> >> >> >> > > still be run in main thread).  So this OOB does brings something
> > > >> >> >> >> > > different.
> > > >> >> >> >> > >
> > > >> >> >> >> > > There are more details on why OOB and the difference/relationship
> > > >> >> >> >> > > between OOB, async QMP, block/general jobs, etc.. but IMHO that's
> > > >> >> >> >> > > slightly out of topic (and believe me, it's not easy for me to
> > > >> >> >> >> > > summarize that).  For more information, please refers to [1].
> > > >> >> >> >> > >
> > > >> >> >> >> > > Summary ends here.
> > > >> >> >> >> > >
> > > >> >> >> >> > > Some Implementation Details
> > > >> >> >> >> > > ===========================
> > > >> >> >> >> > >
> > > >> >> >> >> > > Again, I mentioned that the old QMP workflow is this:
> > > >> >> >> >> > >
> > > >> >> >> >> > >       JSON Parser --> QMP Dispatcher --> Respond
> > > >> >> >> >> > >           /|\    (2)                (3)     |
> > > >> >> >> >> > >        (1) |                               \|/ (4)
> > > >> >> >> >> > >            +---------  main thread  --------+
> > > >> >> >> >> > >
> > > >> >> >> >> > > What this series does is, firstly:
> > > >> >> >> >> > >
> > > >> >> >> >> > >       JSON Parser     QMP Dispatcher --> Respond
> > > >> >> >> >> > >           /|\ |           /|\       (4)     |
> > > >> >> >> >> > >            |  | (2)        | (3)            |  (5)
> > > >> >> >> >> > >        (1) |  +----->      |               \|/
> > > >> >> >> >> > >            +---------  main thread  <-------+
> > > >> >> >> >> > >
> > > >> >> >> >> > > And further:
> > > >> >> >> >> > >
> > > >> >> >> >> > >                queue/kick
> > > >> >> >> >> > >      JSON Parser ======> QMP Dispatcher --> Respond
> > > >> >> >> >> > >          /|\ |     (3)       /|\        (4)    |
> > > >> >> >> >> > >       (1) |  | (2)            |                |  (5)
> > > >> >> >> >> > >           | \|/               |               \|/
> > > >> >> >> >> > >         IO thread         main thread  <-------+
> > > >> >> >> >> >
> > > >> >> >> >> > Is the queue per monitor or per client?
> > > >> >> >> >
> > > >> >> >> > The queue is currently global. I think yes maybe at least we can do it
> > > >> >> >> > per monitor, but I am not sure whether that is urgent or can be
> > > >> >> >> > postponed.  After all now QMPRequest (please refer to patch 11) is
> > > >> >> >> > defined as (mon, id, req) tuple, so at least "id" namespace is
> > > >> >> >> > per-monitor.
> > > >> >> >> >
> > > >> >> >> >> > And is the dispatching going
> > > >> >> >> >> > to be processed even if the client is disconnected, and are new
> > > >> >> >> >> > clients going to receive the replies from previous clients
> > > >> >> >> >> > commands?
> > > >> >> >> >
> > > >> >> >> > [1]
> > > >> >> >> >
> > > >> >> >> > (will discuss together below)
> > > >> >> >> >
> > > >> >> >> >> > I
> > > >> >> >> >> > believe there should be a per-client context, so there won't be "id"
> > > >> >> >> >> > request conflicts.
> > > >> >> >> >
> > > >> >> >> > I'd say I am not familiar with this "client" idea, since after all
> > > >> >> >> > IMHO one monitor is currently designed to mostly work with a single
> > > >> >> >> > client. Say, unix sockets, telnet, all these backends are only single
> > > >> >> >> > channeled, and one monitor instance can only work with one client at a
> > > >> >> >> > time.  Then do we really need to add this client layer upon it?  IMHO
> > > >> >> >> > the user can just provide more monitors if they wants more clients
> > > >> >> >> > (and at least these clients should know the existance of the others or
> > > >> >> >> > there might be problem, otherwise user2 will fail a migration, finally
> > > >> >> >> > noticed that user1 has already triggered one), and the user should
> > > >> >> >> > manage them well.
> > > >> >> >>
> > > >> >> >> qemu should support a management layer / libvirt restart/reconnect.
> > > >> >> >> Afaik, it mostly work today. There might be a cases where libvirt can
> > > >> >> >> be confused if it receives a reply from a previous connection command,
> > > >> >> >> but due to the sync processing of the chardev, I am not sure you can
> > > >> >> >> get in this situation.  By adding "oob" commands and queuing, the
> > > >> >> >> client will have to remember which was the last "id" used, or it will
> > > >> >> >> create more conflict after a reconnect.
> > > >> >> >>
> > > >> >> >> Imho we should introduce the client/connection concept to avoid this
> > > >> >> >> confusion (unexpected reply & per client id space).
> > > >> >> >
> > > >> >> > Hmm I agree that the reconnect feature would be nice, but if so IMHO
> > > >> >> > instead of throwing responses away when client disconnect, we should
> > > >> >> > really keep them, and when the client reconnects, we queue the
> > > >> >> > responses again.
> > > >> >> >
> > > >> >> > I think we have other quite simple ways to solve the "unexpected
> > > >> >> > reply" and "per-client-id duplication" issues you have mentioned.
> > > >> >> >
> > > >> >> > Firstly, when client gets unexpected replies ("id" field not in its
> > > >> >> > own request queue), the client should just ignore that reply, which
> > > >> >> > seems natural to me.
> > > >> >>
> > > >> >> The trouble is that it may legitimately use the same "id" value for
> > > >> >> new requests. And I don't see a simple way to handle that without
> > > >> >> races.
> > > >> >
> > > >> > Under what circumstances can it reuse the same ID for new requests?
> > > >> > Can't we simply tell it not to?
> > > >>
> > > >> I don't see any restriction today in the protocol in connecting with a
> > > >> new client that may not know anything from a previous client.
> > > >
> > > > Well, it knows it's doing a reconnection.
> > > 
> > > If you assume the "same client" reconnects to the monitor, I agree.
> > > But this is a restriction of monitor usage.
> > 
> > In monitor_qmp_event(), we can empty the request queue when got
> > CHR_EVENT_CLOSED.  Would that be a solution?
> 
> What happens to commands that are in flight?

Good questioning...

I think we can track that one as well, say, provide a simple state
machine for Monitor (possibly with a lock) that can be either "idle",
"processing", "drop".

Then a normal routine to execution of command:

0. by default, monitor state "idle"
1. when dequeue the request, mark that monitor as "processing",
   execute the command
2. when reply: if still "processing", then do it; if "drop", then drop
   that reply for current command.  Here we'll reply.

Instead, if disconnect/reconnect happens:

0. by default, monitor state "idle"
1. when dequeue the request, mark that monitor as "processing",
   execute the command
2. port disconnected, in EVENT_CLOSED, we set state to "drop"
3. port reconnected, we do nothing (so the execution state keeps
   through reconnection)
4. when reply: if still "processing", then do it; if "drop", then drop
   that reply for current command.  Here we drop that reply.

But... IMHO this is too awkward only for this single "drop the last
command reply" purpose.  I would prefer to use documentation intead to
let client drop unknown responses directly if it's ok to everyone.

Thanks,

> 
> Dave
> 
> > -- 
> > Peter Xu
> --
> Dr. David Alan Gilbert / dgilbert@redhat.com / Manchester, UK

-- 
Peter Xu