Re: [libvirt] [PATCH 0/9] Resolve libvirtd hang on termination with connected long running client

On 19.01.2018 20:23, John Ferlan wrote: Hi, John. So you suggest a different appoarch. Instead of introducing means to help rpc threads to finish after event loop is finished (stateShutdown hook in my series) you suggest to block futher new rpc processing and then waiting running rpc calls to finish keeping event loop running for that purpose. This could lead to libvirtd never finish its termination if one of qemu processes do not respond for example. In case of long running operation such as migration finishing can take much time. On the over hand if we finish rpc threads abruptly as in case of stateShutdown hook we will deal with all possible error paths on daemon finishing. May be good approach is to abort long running operation, then give rpc threads some time to finish as you suggest and only after that finish them abruptly to handle hanging qemu etc. Also in this approach we keep event loop running for rpc threads only but there can be other threads that depend on the loop. For example if qemu is rebooted we spawn a thread that executes qemuProcessFakeReboot that sends commands to qemu (i.e. depends on event loop). We need to await such threads finishing too while keep event loop running. In approach of stateShutdown hook we finish all threads that uses qemu monitor by closing the monitor. And hack with timeout in a loop... I think of a different aproach for waiting rpc threads to finish their work. First let's make drain function only flush job queue and take some callback which will be called when all workers will be free from work (let's keep worker threads as Dan suggested). In this callback we can use same technique as in virNetDaemonSignalHandler. That is make some IO to set some flag in the event loop thread and cause virEventRunDefaultImpl to finish and then test this flag in virNetDaemonRun. Nikolay