On Mon, Sep 18, 2017 at 1:26 PM, Dr. David Alan Gilbert <dgilb...@redhat.com> wrote: > * Marc-André Lureau (marcandre.lur...@gmail.com) wrote: >> Hi >> >> On Mon, Sep 18, 2017 at 12:55 PM, Dr. David Alan Gilbert >> <dgilb...@redhat.com> wrote: >> > * Marc-André Lureau (marcandre.lur...@gmail.com) wrote: >> >> Hi >> >> >> >> On Mon, Sep 18, 2017 at 10:37 AM, Peter Xu <pet...@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 <pet...@redhat.com> wrote: >> >> >> > On Thu, Sep 14, 2017 at 07:53:15PM +0100, Dr. David Alan Gilbert >> >> >> > wrote: >> >> >> >> * Marc-André Lureau (marcandre.lur...@gmail.com) wrote: >> >> >> >> > Hi >> >> >> >> > >> >> >> >> > On Thu, Sep 14, 2017 at 9:50 AM, Peter Xu <pet...@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. > >> How would you tell it not to use old IDs? Just by writing an unwritten >> rule, because we don't want to fix the per connection client session >> handling in qemu? > > BY writing a written rule! This out of order stuff we're adding here > is a change to the interface and we can define what we require of the > client. As long as what we expect is reasonable then we might end > up with something that's simpler for both the client and qemu. As long as we don't break existing qmp clients. > And I worry this series keeps getting more and more complex for weird > edge cases. That's an interesting point-of-view. I see the point in fixing weird edge cases in qemu RPC code. More than other code we develop with weird edge cases in mind & tests, like the parsing/checking of the json schema for ex, in a similar area with the same maintainer. > Dave > >> > >> > Dave >> > >> >> > >> >> > Then, if client disconnected and reconnected, it should not have the >> >> > problem to generate duplicated id for request, since it should know >> >> > what requests it has sent already. A simplest case I can think of is, >> >> > the ID should contains the following tuple: >> >> >> >> If you assume the "same" client will recover its state, yes. >> >> >> >> > >> >> > (client name, client unique ID, request ID) >> >> > >> >> > Here "client name" can be something like "libvirt", which is the name >> >> > of client application; >> >> > >> >> > "client unique ID" can be anything generated when client starts, it >> >> > identifies a single client session, maybe a UUID. >> >> > >> >> > "request ID" can be a unsigned integer starts from zero, and increases >> >> > each time the client sends one request. >> >> >> >> This is introducing session handling, and can be done in server side >> >> only without changes in the protocol I believe. >> >> >> >> > >> >> > I believe current libvirt is using "client name" + "request ID". It's >> >> > something similar (after all I think we don't normally have >1 libvirt >> >> > to manage single QEMU, so I think it should be good enough). >> >> >> >> I am not sure we should base our protocol usage assumptions based on >> >> libvirt only, but rather on what is possible today (like queuing >> >> requests in the socket etc..). >> >> >> >> > Then even if client disconnect and reconnect, request ID won't lose, >> >> > and no duplication would happen IMHO. >> >> > >> >> >> >> >> >> > >> >> >> >> > >> >> >> >> > > >> >> >> >> > > Then it introduced the "allow-oob" parameter in QAPI schema to >> >> >> >> > > define >> >> >> >> > > commands, and "run-oob" flag to let oob-allowed command to run >> >> >> >> > > in the >> >> >> >> > > parser. >> >> >> >> > >> >> >> >> > From a protocol point of view, I find that "run-oob" distinction >> >> >> >> > per >> >> >> >> > command a bit pointless. It helps with legacy client that wouldn't >> >> >> >> > expect out-of-order replies if qemu were to run oob commands oob >> >> >> >> > by >> >> >> >> > default though. >> >> >> > >> >> >> > After all oob somehow breaks existing rules or sync execution. I >> >> >> > thought the more important goal was at least to keep the legacy >> >> >> > behaviors when adding new things, no? >> >> >> >> >> >> Of course we have to keep compatibily. What do you mean by "oob >> >> >> somehow breaks existing rules or sync execution"? oob means queuing >> >> >> and unordered reply support, so clearly this is breaking the current >> >> >> "mostly ordered" behaviour (mostly because events may still come any >> >> >> time..., and the reconnect issue discussed above). >> >> > >> >> > Yes. That's what I mean, it breaks the synchronous scemantic. But >> >> > I should definitely not call it a "break" though since old clients >> >> > will work perfectly fine with it. Sorry for the bad wording. >> >> > >> >> >> >> >> >> >> > Clients shouldn't care about how/where a command is >> >> >> >> > being queued or not. If they send a command, they want it >> >> >> >> > processed as >> >> >> >> > quickly as possible. However, it can be interesting to know if the >> >> >> >> > implementation of the command will be able to deliver oob, so that >> >> >> >> > data in the introspection could be useful. >> >> >> >> > >> >> >> >> > I would rather propose a client/server capability in >> >> >> >> > qmp_capabilities, >> >> >> >> > call it "oob": >> >> >> >> > >> >> >> >> > This capability indicates oob commands support. >> >> >> >> >> >> >> >> The problem is indicating which commands support oob as opposed to >> >> >> >> indicating whether oob is present at all. Future versions will >> >> >> >> probably make more commands oob-able and a client will want to know >> >> >> >> whether it can rely on a particular command being non-blocking. >> >> >> > >> >> >> > Yes. >> >> >> > >> >> >> > And IMHO we don't urgently need that "whether the server globally >> >> >> > supports oob" thing. Client can just know that from query-qmp-schema >> >> >> > already - there will always be the "allow-oob" new field for command >> >> >> > typed entries. IMHO that's a solid hint. >> >> >> > >> >> >> > But I don't object to return it as well in qmp_capabilities. >> >> >> >> >> >> Does it feel right that the client can specify how the command are >> >> >> processed / queued ? Isn't it preferable to leave that to the server >> >> >> to decide? Why would a client specify that? And should the server be >> >> >> expected to behave differently? What the client needs to be able is to >> >> >> match the unordered replies, and that can be stated during cap >> >> >> negotiation / qmp_capabilties. The server is expected to do a best >> >> >> effort to handle commands and their priorities. If the client needs >> >> >> several command queue, it is simpler to open several connection rather >> >> >> than trying to fit that weird priority logic in the protocol imho. >> >> > >> >> > Sorry I may have missed the point here. We were discussing about a >> >> > global hint for "oob" support, am I right? Then, could I ask what's >> >> > the "weird priority logic" you mentioned? >> >> >> >> I call per-message oob hint a kind of priority logic, since you can >> >> make the same request without oob in the same session and in parallel. >> >> >> >> >> >> >> >> > >> >> >> >> >> >> >> >> > An oob command is a regular client message request with the "id" >> >> >> >> > member mandatory, but the reply may be delivered >> >> >> >> > out of order by the server if the client supports >> >> >> >> > it too. >> >> >> >> > >> >> >> >> > If both the server and the client have the "oob" capability, the >> >> >> >> > server can handle new client requests while previous requests are >> >> >> >> > being >> >> >> >> > processed. >> >> >> >> > >> >> >> >> > If the client doesn't have the "oob" capability, it may still call >> >> >> >> > an oob command, and make multiple outstanding calls. In this case, >> >> >> >> > the commands are processed in order, so the replies will also be >> >> >> >> > in >> >> >> >> > order. The "id" member isn't mandatory in this case. >> >> >> >> > >> >> >> >> > The client should match the replies with the "id" member >> >> >> >> > associated >> >> >> >> > with the requests. >> >> >> >> > >> >> >> >> > When a client is disconnected, the pending commands are not >> >> >> >> > necessarily cancelled. But the future clients will not get >> >> >> >> > replies from >> >> >> >> > commands they didn't make (they might, however, receive >> >> >> >> > side-effects >> >> >> >> > events). >> >> >> >> >> >> >> >> What's the behaviour on the current monitor? >> >> >> > >> >> >> > Yeah I want to ask the same question, along with questioning about >> >> >> > above [1]. >> >> >> > >> >> >> > IMHO this series will not change the behaviors of these, so IMHO the >> >> >> > behaviors will be the same before/after this series. E.g., when >> >> >> > client >> >> >> > dropped right after the command is executed, I think we will still >> >> >> > execute the command, though we should encounter something odd in >> >> >> > monitor_json_emitter() somewhere when we want to respond. And it >> >> >> > will >> >> >> > happen the same after this series. >> >> >> >> >> >> I think it can get worse after your series, because you queue the >> >> >> commands, so clearly a new client can get replies from an old client >> >> >> commands. As said above, I am not convinced you can get in that >> >> >> situation with current code. >> >> > >> >> > Hmm, seems so. But would this a big problem? >> >> > >> >> > I really think the new client should just throw that response away if >> >> > it does not really know that response (from peeking at "id" field), >> >> > just like my opinion above. >> >> >> >> This is a high expectation. >> >> >> >> >> >> -- >> >> Marc-André Lureau >> > -- >> > Dr. David Alan Gilbert / dgilb...@redhat.com / Manchester, UK >> >> >> >> -- >> Marc-André Lureau > -- > Dr. David Alan Gilbert / dgilb...@redhat.com / Manchester, UK -- Marc-André Lureau