> On Jan. 30, 2013, 12:28 p.m., Gordon Sim wrote: > > Is abort() the right call to make? What about close()? It looks to me like > > the 'purpose' of abort is to trigger a simulated eof call on the connection > > processing thread from e.g. the heartbeat timer thread. Since in this case > > we are already on the connection processing thread, why would close() not > > do the job (this is not the same as issuing a clean connection.close > > sequence I don't believe). > > > > One thing to remember with any change to the IO code is that different > > 'transports' (ssl, rdma) and platforms (windows) may involve different > > codepaths. > > Alan Conway wrote: > Clients do not fail-over if the connection is closed politely, which is > the objective here. > > Gordon Sim wrote: > Understood and to be clear I'm not suggesting what I would call a > 'polite' close (i.e. connection-close; connection-close-ok handshake). > However calling qpid::amqp_0_10::Connection::close() should result in the IO > layer detecting that the upper layer wants to close and doing so (see > qpid::sys::AsynchIOHandler::idle() for example). > > The behaviour of these calls is not well defined certainly. My only > concern is we seem to be modifying what abort() is supposed to do and at > least from a skim of the code it seems close() might do what is required > (certainly it should result in aio->queueWriteClose() being called in > AsynchIOHandler). > > Andrew Stitcher wrote: > It is currently true on trunk that that ssl and tcp share the code paths > at this level but rdma does not have abort() implemented at all. > > I tend to agree that overloading abort() is not entirely the correct > semantic although does seem pretty close - abort() says "abort this > connection; I think it has failed but is still connected for some reason". > The correct semantic here could be "close this connection with error". > > I wonder if it is possible to hook into the authentication in a similar > way to ACL to reject the connection there with a failure code. > > Alan Conway wrote: > I'd argue that this fixes a bug in abort() - you shouldn't be able to > write to an aborted connection. It all works nicely now, unless there's a > compelling reason to take a different approach I'd like to leave it as-is. > Semantically it seems fine, abort() means terminate the connection abruptly > without the close handshake, which is exactly what the HA code wants to do. > > Gordon Sim wrote: > I would argue that at present a key part of the role of abort() in > practice is to be called on another thread, hence the requesting of a > callback to eof on the IO processing thread which is likely the source of the > original problem (i.e. abort returning doesn't mean the eof has actually been > handled). Is this change safe if called from a thread other than the > connections IO thread? It seems to me that the right fix is to provide an > 'inline' way of doing the same thing as abort(). (I'm suggesting calling > qpid::amqp_0_10::Connection::close() _may_ be that). > > Andrew Stitcher wrote: > It is safe to call queueWriteClosed() in any thread. So I think this > change is: > > 1. Safe. > 2. Correct, in that a more robust semantic for abort() is indeed to not > transmit anything further on this connection (even if the original case for > abort() did not require this) > > I'm saying this from the POV of the IO code though, not necessarily the > application logic. > > Gordon Sim wrote: > I think the change is odd. Requesting a callback to eof if we are already > on the IO thread seems an odd thing to do. I think it would be clearer to > distinguish between closing a connection abruptly from the IO thread (the new > use case), and requesting an eof to be simulated on the IO thread from some > other thread (the original use case). Partly this is down to abort being a > bad name for the original use case in my view. > > However, if you are sure its safe (baring in mind there are no tests for > heartbeats, concurrent shutdown by both ends etc), then I don't object to it > being committed. > > Andrew Stitcher wrote: > We could add a new abortOnIOThread() call and rename the existing to > abortArbitraryThread() (or some such names) but the latter subsumes the > former after this change so why would we want to do that? > > To me closing the connection abruptly has the same meaning irrespective > of the thread it comes from. The fact that the eof machinery is reused "under > the hood" is irrelevant to my mind the call is meant to abort the connection. >
I think closing abruptly while processing IO is a different thing from requesting simulation of an eof by an arbitrary thread. The expectations are different in each case (as exemplified buy this new use case). If you were implementing the former you would not do it by requesting a callback. So, while this may work and be safe, it is not in my personal opinion particularly clear code. Then again this interface between the broker and the IO code is rather a mess in general, so if you are sure its safe, as I say I have no objection. (i.e. feel free to ship it). - Gordon ----------------------------------------------------------- This is an automatically generated e-mail. To reply, visit: https://reviews.apache.org/r/9137/#review15837 ----------------------------------------------------------- On Jan. 29, 2013, 9:49 p.m., Alan Conway wrote: > > ----------------------------------------------------------- > This is an automatically generated e-mail. To reply, visit: > https://reviews.apache.org/r/9137/ > ----------------------------------------------------------- > > (Updated Jan. 29, 2013, 9:49 p.m.) > > > Review request for qpid and Andrew-Duplicate-Accct-Inactiv Konwinski. > > > Description > ------- > > HA Fix race condition in rejecting connections. > > Sporadic failure of test_failover_python was caused by a race in rejecting > connections. There was a very small window where work could be done by a > connection after it was rejected. > > > Diffs > ----- > > /trunk/qpid/cpp/src/qpid/sys/AsynchIOHandler.cpp 1439431 > > Diff: https://reviews.apache.org/r/9137/diff/ > > > Testing > ------- > > make check, manual heartbeat test > > > Thanks, > > Alan Conway > >
